Operator Theory

Advances and Applications

282

M. Amélia Bastos
Luís Castro
Alexei Yu. Karlovich
Editors

Operator Theory,
Functional
Analysis and
Applications
Operator Theory: Advances and Applications
Volume 282
Founded in 1979 by Israel Gohberg

Editors:
Joseph A. Ball (Blacksburg, VA, USA)
Albrecht Böttcher (Chemnitz, Germany)
Harry Dym (Rehovot, Israel)
Heinz Langer (Wien, Austria)
Christiane Tretter (Bern, Switzerland)

Associate Editors: Honorary and Advisory Editorial Board:

Vadim Adamyan (Odessa, Ukraine) Lewis A. Coburn (Buffalo, NY, USA)
Wolfgang Arendt (Ulm, Germany) Ciprian Foias (College Station, TX, USA)
B. Malcolm Brown (Cardiff, UK) J.William Helton (San Diego, CA, USA)
Raul Curto (Iowa, IA, USA) Marinus A. Kaashoek (Amsterdam, NL)
Kenneth R. Davidson (Waterloo, ON, Canada) Thomas Kailath (Stanford, CA, USA)
Fritz Gesztesy (Waco, TX, USA) Peter Lancaster (Calgary, Canada)
Pavel Kurasov (Stockholm, Sweden) Peter D. Lax (New York, NY, USA)
Vern Paulsen (Houston, TX, USA) Bernd Silbermann (Chemnitz, Germany)
Mihai Putinar (Santa Barbara, CA, USA) Harold Widom (Santa Cruz, CA, USA)
Ilya Spitkovsky (Abu Dhabi, UAE)

Subseries
Linear Operators and Linear Systems
Subseries editors:
Daniel Alpay (Orange, CA, USA)
Birgit Jacob (Wuppertal, Germany)
André C.M. Ran (Amsterdam, The Netherlands)

Subseries
Advances in Partial Differential Equations
Subseries editors:
Bert-Wolfgang Schulze (Potsdam, Germany)
Michael Demuth (Clausthal, Germany)
Jerome A. Goldstein (Memphis, TN, USA)
Nobuyuki Tose (Yokohama, Japan)
Ingo Witt (Göttingen, Germany)

More information about this series at http://www.springer.com/series/4850

M. Amélia Bastos • Luís Castro •
Alexei Yu. Karlovich
Editors

Operator Theory, Functional

Analysis and Applications
Editors
M. Amélia Bastos Luís Castro
Instituto Superior Técnico Departamento de Matemática
Universidade de Lisboa Universidade de Aveiro
Lisboa, Portugal Aveiro, Portugal

Alexei Yu. Karlovich

Faculdade de Ciências e Tecnologia
Universidade Nova de Lisboa
Lisboa, Portugal

ISSN 0255-0156 ISSN 2296-4878 (electronic)

Operator Theory: Advances and Applications
ISBN 978-3-030-51944-5 ISBN 978-3-030-51945-2 (eBook)
https://doi.org/10.1007/978-3-030-51945-2

Mathematics Subject Classification: 47-XX

© Springer Nature Switzerland AG 2021

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This book is published under the imprint Birkhäuser, www.birkhauser-science.com, by the registered
company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

This volume is dedicated to the 30th International Workshop on Operator Theory

and its Applications, IWOTA 2019, where a wide range of topics on the recent
developments in Operator Theory and Functional Analysis was presented and
discussed.
The book is composed of 30 articles covering the different scientific areas of
IWOTA 2019. Namely:
• Group representations and determinantal hypersurfaces. Dilation theory of com-
pletely positive maps and semigroups as well as the operator algebraic approach
to dilation theory. Tight and cover-to-join representations of semilattices and
inverse semigroups. Compact sequences in quasifractal algebras. Representable
and continuous functionals of Banach quasi *-algebras. Langlands reciprocity for
C ∗ -algebras.
• Tau functions and linear systems. Bishop operators, spectral properties, and
spectral invariant subspaces. The operator Jensen–Mercer inequality. Birkhoff–
James orthogonality.
• Extended Heinz and Jensen type inequalities and rearrangements. d-Modified
Riesz potentials on central Campanato spaces. K-inner functions and K-
contractions for unitarily invariant reproducing kernel functions. Products of
unbounded block functions. Riesz–Fischer maps and frames in rigged Hilbert
spaces.
• Minimality properties of Sturm–Liouville problems. Solutions of inhomoge-
neous ill-posed problems in Banach space. Quantum graph with Rashba Hamil-
tonian. Unbounded operators on the Segal–Bargmann space. Solvability of the
Caffarelli–Silvestre extension problem.
• Semiclassical elliptic pseudodifferential operators and periodic coherent states
decomposition. Time-dependent approach to the Sommerfeld solution of a
diffraction problem. A numerical approach for approximating variable-order
fractional integral operator. Inner–outer factorization of rational matrix valued
functions.

v
vi Preface

• Toeplitz plus Hankel operators and their close relatives. Fourier convolution
operators with slowly oscillating symbols. Singular integral operators with
Cauchy and Mellin kernels. Noncommutative C ∗ -algebras generated by Toeplitz
operators on the unit sphere.
• Groups of orthogonal matrices, all orbits of which generate lattices. The inverse
characteristic polynomial problem for trees.
IWOTA 2019 was held from 22nd to 26th of July 2019 at Instituto Superior
Técnico, University of Lisbon, Portugal. It was focused on the latest developments
in Functional Analysis, Operator Theory, and related fields and was organized by M.
Amélia Bastos (IST, UL), António Bravo (IST, UL), Catarina Carvalho (IST, UL),
Luís Castro (UA), Alexei Karlovich (FCT, UNL), and Helena Mascarenhas (IST,
UL).
IWOTA 2019 had 471 participants from all over the world. The program
consisted of 11 plenary lectures:
• J. Ball, Input/State/Output Linear Systems and Their Transfer Functions: From
Single-Variable to Multivariable to Free Noncommutative Function Theory;
• A. Böttcher, Lattices from Equiangular Tight Frames;
• K. Davidson, Noncommutative Choquet Theory;
• R. Exel, Statistical Mechanics on Markov Spaces with Infinitely Many States;
• H. Feichtinger, Classical Fourier Analysis and the Banach Gelfand Triple;
• R. Kaashoek, Inverting Structured Operators and Solving Related Inverse Prob-
lems;
• I. Klep, Bianalytic Maps between Matrix Convex Sets;
• L.-E. Persson, My Life with Hardy and His Inequalities;
• P. Semrl, Automorphisms of Effect Algebras;
• B. Silbermann, Invertibility Issues for Toeplitz Plus Hankel Operators;
• C. Tretter, Spectra and Essential Spectra of Non-Self-Adjoint Operators;
16 semi-plenary lectures:
• P. Ara, Separated Graphs and Dynamics;
• S. Belinschi, Analytic Transforms of Noncommutative Distributions;
• G. Blower, Linear Systems in Random Matrix Theory;
• A. Caetano, Function Spaces Techniques in Problems of Scattering by Fractal
Screens;
• A. B. Cruzeiro, On Some Stochastic Partial Differential Equations Obtained by
a Variational Approach;
• R. Duduchava, Boundary Value Problems on Hypersurfaces and -Convergence;
• P. Freitas, Spectral Determinants of Elliptic Operators: Dependence on Spatial
Dimension and Order of the Operator;
• E. Gallardo, Invariant Subspaces for Bishop Operators and Beyond;
• Yu. Karlovich, Algebras of Singular Integral Operators with Piecewise Quasi-
continuous Coefficients and Non-Smooth Shifts;
• S. Petermichl, Change of Measure;
• S. Roch, On Quasifractal Algebras;
Preface vii

• O. M. Shalit, Dilation Theory: Fresh Directions with New Applications;

• F.-O. Speck, Advances in General Wiener–Hopf Factorization;
• I. Spitkovsky, One Hundred Years of. . . Numerical Range;
• N. Vasilevski, Algebras Generated by Toeplitz Operators on the Hardy Space
H 2 (S 2n−1 );
• N. Zorboska, Toeplitz Operators on the Bergman Space with BMOp Symbols
and the Berezin Transform;
and more than 400 contributed talks distributed among 22 special sessions:
• Analysis and Algebraic Geometry for Operator Variables (organized by I. Klep
and V. Vinnikov);
• Analysis and Synthesis for Operator Algebras (organized by A. Dor-On, R. Exel,
E. Katsoulis, and D. Pitts);
• Free Analysis and Free Probability (organized by S. Belinschi, M. Popa, and
R. Speicher);
• Functional Calculus, Spectral Sets and Constants (organized by L. Kosinski,
F. Schwenninger, and M. Wojtylak);
• Gabor Analysis and Noncommutative Geometry (organized by F. Luef and
I. Nikolaev);
• Geometry of Linear Operators and Operator Algebras (organized by K. Paul and
D. Sain);
• Hypercomplex Analysis and Operator Theory (organized by D. Alpay,
F. Colombo, and U. Kähler);
• Integral Operators and Applications (organized by R. Hagger, K.-M. Perfekt, and
J. Virtanen);
• Linear Operators and Function Spaces (organized by M. C. Câmara and M. Ptak);
• Matrix Theory and Control (organized by M. Dodig and S. M. Furtado);
• Multivariable Operator Theory (organized by J. Ball and V. Bolotnikov);
• Operators of Harmonic Analysis, Related Function Spaces, and Applications —
dedicated to Lars-Erik Persson on his 75th birthday (organized by H. Rafeiro and
N. Samko);
• Operators on Reproducing Kernel Hilbert Spaces (organized by N. Vasilevski
and K. Zhu);
• Operator Theoretical Methods in Mathematical Physics (organized by L. Castro
and F.-O. Speck);
• Orders Preserving Operators on Cones and Applications (organized by M. Akian,
S. Gaubert, A. Peperko, and G. Vigeral);
• Preserver Problems in Operator Theory and Functional Analysis (organized by
F. Botelho and G. Geher);
• Random Matrix Theory (organized by H. Hedenmalm and J. Virtanen);
• Representation Theory of Algebras and Groups (organized by C. André,
S. Lopes, and A. P. Santana);
• Semigroups and Evolution Equations (organized by C. Budde and C. Seifert);
• Spectral Theory and Differential Operators (organized by A. Khrabustovskyi,
O. Post, and C. Trunk);
viii Preface

• Toeplitz Operators, Convolution Type Operators, and Operator Algebras —

dedicated to Yuri Karlovich on his 70th birthday (organized by M. A. Bastos
and A. Karlovich);
• Truncated Moment Problems (organized by M. Infusino and S. Kuhlmann).
Finally, the editors of this volume express their gratitude to the IWOTA
2019 sponsors: the Center for Functional Analysis, Linear Structures and
Applications supported by the projects CEAFEL-UID/MAT/04721/2013,
UID/MAT/04721/2019; the Center for Research and Development in Mathematics
and Applications supported by the project CIDMA-UID/MAT/04106/2019;
the Centre for Mathematics and Applications supported by the project CMA-
UID/MAT/00297/2019; and to the Rector of Lisbon University and the Portuguese
Foundation for Science and Technology.

Lisboa, Portugal M. Amélia Bastos

Aveiro, Portugal Luís Castro
Lisboa, Portugal Alexei Yu. Karlovich
May 2020
Contents

Extended Heinz and Jensen Type Inequalities and Rearrangements . . . . . . 1

Shoshana Abramovich
On Some Applications of Representable and Continuous
Functionals of Banach Quasi *-Algebras .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 15
Maria Stella Adamo
Minimality Properties of Sturm-Liouville Problems with Increasing
Affine Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 33
Yagub N. Aliyev
Scattering, Spectrum and Resonance States Completeness for a
Quantum Graph with Rashba Hamiltonian . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 51
Irina V. Blinova, Igor Y. Popov, and Maria O. Smolkina
Tau Functions Associated with Linear Systems . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 63
Gordon Blower and Samantha L. Newsham
Groups of Orthogonal Matrices All Orbits of Which
Generate Lattices.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
Albrecht Böttcher
Invertibility Issues for Toeplitz Plus Hankel Operators and Their
Close Relatives .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 113
Victor D. Didenko and Bernd Silbermann
K-Inner Functions and K-Contractions .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 157
Jörg Eschmeier and Sebastian Toth
Tight and Cover-to-Join Representations of Semilattices and
Inverse Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 183
Ruy Exel

ix
x Contents

Calkin Images of Fourier Convolution Operators with Slowly

Oscillating Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 193
C. A. Fernandes, A. Yu. Karlovich, and Yu. I. Karlovich
Inner Outer Factorization of Wide Rational Matrix Valued
Functions on the Half Plane.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 219
A. E. Frazho and A. C. M. Ran
Convergence Rates for Solutions of Inhomogeneous Ill-posed
Problems in Banach Space with Sufficiently Smooth Data .. . . . . . . . . . . . . . . . . 235
Matthew A. Fury
A Closer Look at Bishop Operators.. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 255
Eva A. Gallardo-Gutiérrez and Miguel Monsalve-López
Products of Unbounded Bloch Functions. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 283
Daniel Girela
Birkhoff–James Orthogonality and Applications: A Survey .. . . . . . . . . . . . . . . 293
Priyanka Grover and Sushil Singla
The Generalized ∂-Complex on the Segal–Bargmann Space.. . . . . . . . . . . . . . . 317
Friedrich Haslinger
The Inverse Characteristic Polynomial Problem for Trees .. . . . . . . . . . . . . . . . . 329
Charles R. Johnson and Emma Gruner
A Note on the Fredholm Theory of Singular Integral Operators
with Cauchy and Mellin Kernels, II . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353
Peter Junghanns and Robert Kaiser
A Note on Group Representations, Determinantal Hypersurfaces
and Their Quantizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 393
Igor Klep and Jurij Volčič
Algebras Generated by Toeplitz Operators on the Unit Sphere II:
Non Commutative Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 403
Maribel Loaiza and Nikolai Vasilevski
d-Modified Riesz Potentials on Central Campanato Spaces . . . . . . . . . . . . . . . . 423
Katsuo Matsuoka
On Some Consequences of the Solvability of the Caffarelli–Silvestre
Extension Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 441
Jan Meichsner and Christian Seifert
Time-Dependent Approach to Uniqueness of the Sommerfeld
Solution to a Problem of Diffraction by a Half-Plane .. . .. . . . . . . . . . . . . . . . . . . . 455
A. Merzon, P. Zhevandrov, J. E. De la Paz Méndez, and T. J. Villalba Vega
On the Operator Jensen-Mercer Inequality.. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 483
H. R. Moradi, S. Furuichi, and M. Sababheh
Contents xi

A Numerical Approach for Approximating Variable-Order

Fractional Integral Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 495
Somayeh Nemati
Langlands Reciprocity for C ∗ -Algebras . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 515
Igor V. Nikolaev
Compact Sequences in Quasifractal Algebras . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 529
Steffen Roch
Dilation Theory: A Guided Tour . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 551
Orr Moshe Shalit
Riesz-Fischer Maps, Semi-frames and Frames in Rigged
Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 625
Francesco Tschinke
Periodic Coherent States Decomposition and Quantum Dynamics
on the Flat Torus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 647
Lorenzo Zanelli
Extended Heinz and Jensen Type
Inequalities and Rearrangements

Shoshana Abramovich

Dedicated to Lars-Erik Persson on the occasion of his 75th

birthday.

Abstract In this paper we extend the well known Heinz inequality which says that
√
2 a1 a2 ≤ H (t) ≤ a1 + a2 , a1 , a2 > 0, 0 ≤ t ≤ 1, where H (t) = a1t a21−t + a11−t a2t .
We discuss the bounds of H (t) in the intervals t ∈ [1, 2] and t ∈ [2, ∞) using the
subquadracity and the superquadracity of ϕ(x) = x t , x ≥ 0respectively. Further, 
 
we extend H (t) to get results related to ni=1 Hi (t) = ni=1 ait ai+1 1−t
+ ai1−t ai+1 t ,
an+1 = a1 , ai > 0, i = 1, . . . , n, where H1 (t) = H (t). These results, obtained
by using rearrangement
 techniques, show that the minimum and the maximum
of the sum ni=1 Hi (t) for a given t, depend only on the specific arrangements
called circular alternating order rearrangement and circular symmetrical order
rearrangement of a given set (a) = (a1 , a2 , . . 
. , an ), ai > 0, i = 1, 2, . . . , n. These
lead to extended Heinz type inequalities of ni=1 Hi (t) for different intervals of
t. The results may also be considered as special cases of Jensen type inequalities
for concave, convex, subquadratic and superquadratic functions, which are also
discussed in this paper.

Keywords Rearrangements · Heinz inequality · Jensen inequality · Convexity ·

Superquadracity

Mathematics Subject Classification (2010) 26D15, 26A51, 47A50, 47A60

S. Abramovich ()
Department of Mathematics, University of Haifa, Mount Carmel, Haifa, Israel
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 1

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_1
2 S. Abramovich

1 Introduction and Basic Results

In this paper we extend the well known Heinz inequality

√
2 a1 a2 ≤ H (t) ≤ a1 + a2 , a1 , a2 > 0, 0 ≤ t ≤ 1, (1.1)

where

H (t) = a1t a21−t + a11−t a2t .

In the past few years attention has been put toward refining or reversing this
inequality. Recently, in 2019, new refinements of (1.1) have been proved in [5].
We start by discussing the bounds of H (t) in the intervals t ∈ [1, 2] and
t ∈ [2, ∞) using the subquadracity and the superquadracity of ϕ(x) = x t , x ≥ 0
respectively.
Then, we continue extending H (t) to get results related to


n n 
 
1−t
Hi (t) = ait ai+1 + ai1−t ai+1
t
, an+1 = a1 , ai > 0, i = 1, . . . , n,
i=1 i=1

where H1 (t) = H (t) and t ∈ R.

These results, obtained by using  rearrangement techniques, show that the mini-
mum and the maximum of the sum ni=1 Hi (t) for a specific t, depend only on the
specific arrangements called circular alternating order rearrangement and circular
symmetrical order rearrangement of a given set (a) = (a1 , a2 , . . . , an ), ai > 0,
i = 1, 2, . . . , n. These lead to extended Heinz type inequalities of ni=1 Hi (t) for
different intervals of t. The results may also be considered as special cases of Jensen
type inequalities for concave, convex, subquadratic and superquadratic functions,
which are also discussed in this paper.
We start with a lemma which easily leads to (1.1):
Lemma 1.1 Let ϕ : R+ → R+ be a non-negative increasing function and let
Gp (a, b) be the function
 p    a p  a 
b b 1
Gp (a, b) = a ϕ +b ϕ , p≥ , a, b > 0.
a a b b 2

Then,
    a 
1 1 b
Gp (a, b) ≥ a 2 b 2 ϕ +ϕ . (1.2)
a b
Heinz and Jensen Type Inequalities and Rearrangements 3

Proof Computing the derivative of Gp (a, b) we see that

  p    a p  a  b
∂Gp (a, b) b b
= a ϕ −b ϕ ln ≥ 0, (1.3)
∂p a a b b a

when without loss of generality we assume that b ≥ a > 0. Therefore (1.2) follows
from (1.3).
When ϕ (x) ≡ 1 we get from Lemma 1.1:
Corollary 1.2 Let a, b, t ∈ R and let the symmetric function around t = 1
2 be

H (t) = a 1−t bt + a t b1−t , a, b > 0.

Then H (t) is symmetric around t = 1

2, increasing when t ≥ 1
2, decreasing when
t < 12 and the inequality

1 1
H (t) = a 1−t bt + a t b1−t ≥ 2a 2 b 2

holds for any t ∈ R.

Also, because of the monotonicity and the symmetry of H (t) we get that for
a, b > 0,
1 1
2a 2 b 2 ≤ a 1−t bt + a t b1−t ≤ a + b, 0 < t < 1.

When n ≤ t ≤ n + 1 and n = 1, 2, . . . ,

a 1−n bn + a n b1−n ≤ a 1−t bt + a t b1−t ≤ a −n bn+1 + a n+1 b−n .

When n ≤ t ≤ n + 1 and n = −1, −2, . . . ,

a n+1 b−n + a −n bn+1 ≤ a 1−t bt + a t b1−t ≤ a n b−n+1 + a −n+1 bn .

In order to obtain more Heinz type inequalities we introduce the definitions of

superquadratic and subquadratic functions (see for instance [2, 4] and [7]) which
include the functions f (x) = x t , x ≥ 0, when t ≥ 2 and 1 ≤ t ≤ 2, respectively.
Definition 1.3 A function ϕ : [0, B) → R, 0 < B ≤ ∞ is superquadratic provided
that for all x ∈ [0, B) there exists a constant Cϕ (x) ∈ R such that the inequality

ϕ(y) ≥ ϕ(x) + Cϕ (x)(y − x) + ϕ(|y − x|)

holds for all y ∈ [0, B), (see [2, Definition 2.1], there [0, ∞) instead [0, B)).
The function ϕ is called subquadratic if −ϕ is superquadratic.
4 S. Abramovich

Proposition
1.4 Suppose that f is superquadratic. Let  0 ≤ xi < B, i = 1, . . . , n
and let x = ni=1 ai xi , where ai ≥ 0, i = 1, . . . , n and ni=1 ai = 1. Then


n 
n
ai f (xi ) − f (x) ≥ ai f (|xi − x|) . (1.4)
i=1 i=1

If f is non-negative, it is also convex and the inequality refines Jensen’s inequality.

In particular, the functions f (x) = x r , x ≥ 0, r ≥ 2 are superquadratic and convex,
and equality holds in inequality (1.4) when r = 2.
Similarly,
 suppose that f is subquadratic. Let 0 ≤ xi < B, i = 1, . . . , n and let
x = ni=1 ai xi , where ai ≥ 0, i = 1, . . . , n and ni=1 ai = 1. Then


n 
n
ai f (xi ) − f (x) ≤ ai f (|xi − x|) . (1.5)
i=1 i=1

In Theorem 1.5 we employ for t ≥ 2 the inequality satisfied by the superquadratic

function φ (x) = x t .
Theorem 1.5 Let a, b > 0 and t ≥ 2. Then
 
a t b1−t + a 1−t bt ≥ (a + b) + a 1−t + b1−t |b − a|t
    2
2|b − a| t a b2
≥ max (a + b) 1 + , + . (1.6)
a+b b a

Proof From the superquadracity of φ (x) = x t , t ≥ 2, n = 2, we get that

 a t  t
b
b +a = a t b1−t + a 1−t bt
b a
 t 
a b b a t
≥ (a + b) 1 + −1 + 1−
a+b a a+b b
 
a 1−t + b 1−t
= (a + b) 1 + |b − a|t
a+b
    2
2|b − a| t a b2
≥ max (a + b) 1 + , + . (1.7)
a+b b a

Indeed, the first inequality in (1.7) follows from the superquadracity of φ(x) = x t ,
t ≥ 2, x ≥ 0. The second inequality follows from the convexity of φ(x) = x t ,
t ≥ 1, x ≥ 0, and the monotonicity of a t b1−t + a 1−t bt .
The proof is complete.
Heinz and Jensen Type Inequalities and Rearrangements 5

It is easy to verify that there are cases by which the superquadratic property leads to
a better result in (1.6) than the monotonicity property of H (t) and vice versa.
In Theorem 1.6 we employ for 1 ≤ t ≤ 2 the inequality satisfied by the
subquadratic function φ (x) = x t .
Theorem 1.6 Let a, b > 0 and 1 < t < 2. Then

a + b ≤ a t b1−t + a 1−t bt
   a2 b2
≤ min (a + b) + a 1−t + b1−t |b − a|t , + (1.8)
b a

holds.
Proof The left hand-side of (1.8) follows from the monotonicity of a t b1−t +a 1−t bt ,
x ≥ 0, t ≥ 12 . The right hand-side follows from the subquadracity of φ (x) = x t ,
x ≥ 0, 1 ≤ t ≤ 2, and the monotonicity of a t b1−t + a 1−t bt , 1 ≤ t ≤ 2.
The proof is complete.
In the sequel we use the following lemma.
Lemma 1.7 Let ϕ be a continuous function on x ≥ 0 which is twice differentiable
on x > 0, with ϕ(0) = 0 and

lim xϕ (x) = 0.
x→0+

If ϕ is convex, then xϕ (x) − ϕ(x) ≥ 0 on x > 0.

If ϕ is concave, then xϕ x) − ϕ(x) ≤ 0 on x > 0.
Proof Assuming ϕ is convex, we have ϕ (x) ≥ 0 for x > 0, hence

(xϕ (x) − ϕ(x)) = xϕ (x) ≥ 0

for x > 0, and xϕ (x) − ϕ (x) is increasing. Also

lim (xϕ (x) − ϕ(x)) = 0,

x→0+

and therefore xϕ (x) − ϕ(x) ≥ 0 when x > 0.

Similarly we prove the concavity case.
6 S. Abramovich

2 Rearrangements and New Jensen and Heinz Type

Inequalities

In [3] the authors deal with Jensen type inequalities and rearrangements. The
following definitions appear first in [6]. Theorem 2.6 below appears in [3].
Definition 2.1 ([6]) An ordered set (x) = (x1 , . . . , xn ) of n real numbers is
arranged in symmetrical decreasing order if

x1 ≤ xn ≤ x2 ≤ · · · ≤ x[(n+2/2)]

or if

xn ≤ x1 ≤ xn−1 ≤ · · · ≤ x[(n+1/2)] .

Definition 2.2 ([6]) A circular rearrangement of an ordered set (x) is a cyclic

rearrangement of (x) or a cyclic rearrangement followed by inversion.
For example, the circular rearrangements of the ordered set (1, 2, 3, 4) are the sets

(1, 2, 3, 4) , (2, 3, 4, 1) , (3, 4, 1, 2) , (4, 1, 2, 3) ,

(4, 3, 2, 1) , (1, 4, 3, 2) , (2, 1, 4, 3) , (3, 2, 1, 4) .

Definition 2.3 ([6]) A set (x) is arranged in circular symmetrical order if one of its
circular rearrangements is symmetrically decreasing.
The alternating order of (x) as definedin Definition 2.4 below was introduced and
proved in [8] to be the minimum of ni=1 xi xi+1 under rearrangement. Here the
minimum and maximum of
n 
    
xi+1 xi
xi ϕ + xi+1 ϕ
xi xi+1
i=1

and of
n 
 
1−t
xit xi+1 + xi1−t xi+1
t

i=1

under rearrangement of (x) is obtained, which leads to Heinz and Jensen type
inequalities.
Heinz and Jensen Type Inequalities and Rearrangements 7

Definition 2.4 An ordered set (x) = (x1 , . . . , xn ) of n real numbers is arranged in

alternating order if

x1 ≤ xn−1 ≤ x3 ≤ xn−3 ≤ · · · ≤ x n+1 ≤ · · · ≤ x4 ≤ xn−2 ≤ x2 ≤ xn , (2.1)

2

or if

xn ≤ x2 ≤ xn−2 ≤ x4 ≤ · · · ≤ x n+1 ≤ · · · ≤ xn−3 ≤ x3 ≤ xn−1 ≤ x1 . (2.2)

2

Definition 2.5 A set (x) is arranged in circular alternating order if one of its
circular rearrangements is arranged in an alternating order.
Theorem 2.6 ([3]) Let F (u, v) be differentiable and symmetric real function
defined on (α, β), −∞ ≤ α < β ≤ ∞ and α ≤ u, v, w ≤ β. Assume that

∂F (v, u) ∂F (v, w)
≤
∂v ∂v
for u ≤ min{w, v}. Then, for any set (x) = (x1 , x2 , . . . , xn ), α ≤ xi ≤ β, i =
1, . . . , n given except its arrangements


n
F (xi , xi+1 ) , xn+1 = x1
i=1

is maximal if (x) is arranged in circular symmetrical order and minimal if (x) is

arranged in circular alternating order as defined above.

It is proved in [1] that the maximal arrangement of ni=1 F (xi , xi+1 ), xn+1 = x1 is
attained when (x) is arranged in circular symmetrical
 order.
An outline of the minimal arrangement of ni=1 F (xi , xi+1 ), xn+1 = x1 as stated
in Theorem 2.6 and proved in [3] is as follows:
We denote a given set of n real numbers according to their increasing order (a) =
(a1 , a2 , . . . , an−1 , an ), where a1 ≤ a2 ≤ · · · ≤ an−1 ≤ an . We start with an
arbitrary permutation of  (a) called (b) = (b1 , . . . , bn ). As F (u, v) is symmetric
and we are interested in ni=1 F (bi , bi+1 ), bn+1 = b1 , which is clearly invariant
under all circular rearrangements, we can assume that b1 = a1 . Now we go through
three permutations which bring us from (b) → (c) → (d) → (e) in which


n 
n 
n 
n
F (bi , bi+1 ) ≥ F (ci , ci+1 ) ≥ F (di , di+1 ) ≥ F (ei , ei+1 ) ,
i=1 i=1 i=1 i=1

and we make sure that the two first and two last numbers in (e) are e1 = a1 , e2 =
an−1 , en−1 = a2 , en = an which are already the two first and the two last in the
rearrangements of the alternating order of type (2.1). We realize also that when we
8 S. Abramovich

check (e2 , e3 , . . . , en−2 , en−1 ) we already have that e2 and en−1 are the largest and
the smallest numbers respectively in (e2 , e3 , . . . , en−2 , en−1 ).
Now we use the induction procedure: We assume the validity of Theorem 2.6
for the set of n − 2 numbers and show that this implies its validity for the set
of n numbers. More specifically, the n − 2 numbers if rearranged in alternating
order
n−2 of (2.2) give, according to the induction hypothesis, the smallest value of
i=2 F (ei , ei+1 ) and in the same time we get that (e1 , . . . , en ) is arranged in
alternating order too, this time according to (2.1) and therefore the proof by
induction for n numbers is obtained.  
In Theorem 2.7 we introduce the symmetric function F (u, v) = uϕ uv +vϕ uv
and show when it satisfies Theorem 2.6. Similarly, in Theorem 2.8 we introduce the
symmetric function F (u, v) = us v t + ut v s and show when it satisfies Theorem 2.6.
We denote in the sequel ( x) = ( x1 , 
x2 , . . . ,  x) = (
xn ) and ( x1 , 
x2 , . . . , 
xn )
to be the circular alternating order rearrangement and circular symmetrical order
rearrangement of a given set (x) = (x1 , x2 , . . . , xn ) ∈ Rn .
Theorem 2.7 Assume ϕ is a concave differentiable function on R+ and

lim (xϕ (x) − ϕ(x)) = 0.

x→0+

Then the inequalities

n 
    

xi+1 
xi

xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1
n 
    
xi+1 xi
≤ xi ϕ + xi+1 ϕ
xi xi+1
i=1
n 
    

xi+1 
xi
≤ 
xi ϕ +
xi+1 ϕ (2.3)

xi 
xi+1
i=1

hold where xi > 0, i = 1, 2, . . . , n and 

xn+1 = 
x1 , xn+1 = x1 , 
xn+1 = 
x1 .
Assume ϕ is a convex differentiable function on R+ and

lim (xϕ (x) − ϕ(x)) = 0.

x→0+

Then the reverse of inequalities (2.3) hold.

Proof Using Theorem 2.6 we define
v  u
F (u, v) = uϕ + vϕ .
u v
Heinz and Jensen Type Inequalities and Rearrangements 9

Then, for u ≤ v, w

∂F (u, v) v  u u u

=ϕ +ϕ − ϕ
∂v u v v v
v  w  w  w  ∂F (w, v)
≤ϕ +ϕ − ϕ = .
w v v v ∂v

Indeed, ϕ is concave, that is, ϕ is decreasing, and

v  v
ϕ ≤ ϕ .
u w
Moreover,
u u u w w w 
ϕ − ϕ ≤ϕ − ϕ
v v v v v v
according to Lemma 1.7. Hence, according to Theorem 2.6 inequalities (2.3) hold.
Similarly, we get that the reversed (2.3) hold when ϕ is convex.
Theorem 2.6 above is used in Theorem 2.8 to prove an extension of Heinz type
inequalities.
Theorem 2.8 Let F (x, y) = x s y t + x t y s .
(a) If x, y, s, t ∈ R+ , then for (x) ∈ Rn+ , the inequalities


n
 n


xit s
xi+1 +t
xi+1xis ≤ xis xi+1
t
+ xi+1
t
xis
i=1 i=1


n

≤ 
xit s
xi+1 +t
xi+1xis , (2.4)
i=1

hold, where  xn+1 =  x1 , xn+1 = x1 , xn+1 =  x1 , and (

x) is the circular
alternating order rearrangement of (x) and ( x) is the circular symmetrical
order rearrangement of (x). In particular, (2.4) holds when t + s = 1 and
0 ≤ t ≤ 1.
(b) If s ≤ 0, t ≥ 0 and xi > 0, i = 1, 2, . . . , n, the reverse of (2.4) holds. In
particular, the reverse of (2.4) holds when t + s = 1 and t ≥ 1.
Proof We use Theorem 2.6, which guarantees that (2.4) holds for F (x, y) = x s y t +
x t y s if

∂F (u, v) ∂F (w, v)
≤ when u ≤ min (v, w) .
∂v ∂v
10 S. Abramovich

Since

∂F (u, v) ∂ us v t + ut v s
= = tus v t −1 + sut v s−1 ,
∂v ∂v
we have to show that

tus v t −1 + sut v s−1 ≤ tws v t −1 + swt v s−1 when u ≤ min (v, w) .

This holds because

 
tv t −1 us − ws ≤ 0 ≤ sv s−1 wt − ut when u ≤ min (v, w) . (2.5)

This proves Case (a).

In the same way as Case (a) is proved for s, t > 0, we see that in Case (b) we get
the reverse of (2.5) and therefore we get the reverse of (2.4).
The proof is complete.
Remark 2.9 Inequalities (2.4) when 0 ≤ t ≤ 1 and t + s = 1, and the reverse of
inequalities (2.4) when t ≥ 1 and t + s = 1 follow also from Theorem 2.7 when the
function is ϕ(x) = x t , x > 0.
In Theorem 2.10 we combine the results obtained in Theorem 2.7 concerning
rearrangement with inequalities satisfied by concave functions, convex functions,
subquadratic functions and superquadratic functions to get new Jensen type inequal-
ities.
Theorem 2.10 Let xi > 0, i = 1, 2, . . . , n and let ( x) and (
x) be the circular
alternating order rearrangement and the circular symmetrical order rearrangement
respectively of (x) = (x1 , x2 , . . . , xn ). Let 
xn+1 = 
x1 , xn+1 = x1 , 
xn+1 = x1 .
Then
(a) For a concave function ϕ : R+ → R the following inequalities hold:
n 
    

xi+1 
xi

xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1
n 
    
xi+1 xi
≤ xi ϕ + xi+1 ϕ
xi xi+1
i=1
n 
    

xi+1 
xi
≤ 
xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1


n
≤ 2ϕ(1) xi . (2.6)
i=1
Heinz and Jensen Type Inequalities and Rearrangements 11

(b) For a convex and subquadratic function ϕ the inequalities


n 
n    

xi − 
xi+1 
xi − 
xi+1
2ϕ(1) 
xi + 
xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1 i=1


n    

xi+1 
xi
≥ 
xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1


n    
xi+1 xi
≥ xi ϕ + xi+1 ϕ
xi xi+1
i=1


n    

xi+1 
xi
≥ 
xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1


n
≥ 2ϕ (1) xi (2.7)
i=1

hold.
(c) For a superquadratic function ϕ the inequalities
n 
    

xi+1 
xi

xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1
n 
    
xi+1 xi
≥ xi ϕ + xi+1 ϕ
xi xi+1
i=1
n 
    

xi+1 
xi
≥ 
xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1
n 
    

xi − 
xi+1 
xi − 
xi+1
≥ 
xi ϕ +
xi+1 ϕ

xi 
xi+1
i=1


n
+2ϕ(1) 
xi (2.8)
i=1

hold.
Proof We start with the proof of Case (a). From the concavity of ϕ we get that
   
xi+1 xi
xi ϕ + xi+1 ϕ ≤ ϕ(1) (xi + xi+1 ) , i = 1, . . . , n, xn+1 = x1 .
xi xi+1
12 S. Abramovich

Therefore,
n 
     
n
xi+1 xi
xi ϕ + xi+1 ϕ ≤ 2ϕ(1) xi .
xi xi+1
i=1 i=1

Together with Theorem 2.7 we get inequalities (2.6).

Case (b): Inequality (1.5) says that for n = 2 when ϕ is subquadratic, the
inequality
   
xi+1 xi
xi ϕ + xi+1 ϕ
xi xi+1
   
xi+1 xi
≤ (xi + xi+1 ) ϕ (1) + xi ϕ 1− + xi+1 ϕ 1−
xi xi+1

holds. Therefore summing up, we get


n n 
    
xi+1 xi
2ϕ(1) xi + xi ϕ 1 − + xi+1 ϕ 1 −
xi xi+1
i=1 i=1
n 
    
xi+1 xi
≥ xi ϕ + xi+1 ϕ ,
xi xi+1
i=1

and together with Theorem 2.7 we get that inequalities (2.7) hold.
Case (c) follows similarly from (1.4) and Theorem 2.7. The proof is complete.

Theorem 2.11 combines the results related to rearrangements proved in Theorem 2.7
with the concavity of x t , x ≥ 0, 0 < t ≤ 1, convexity and subquadracity properties
of x t , x ≥ 0, for 1 ≤ t ≤ 2 and its superquadracity properties for t ≥ 2, to get what
we call: “Extended Heinz type inequalities”.
Alternatively the results of Theorem 2.11 can also be obtained from Theorem 2.8
for 0 ≤ t < 1, s = 1 − t and for t ≥ 1 and s = 1 − t, combined with the concavity
of x t , x ≥ 0, 0 < t ≤ 1, convexity and subquadracity properties of x t , x ≥ 0, for
1 ≤ t ≤ 2 and its superquadracity properties for t ≥ 2, to get the “Extended Heinz
type inequalities”.
Theorem 2.11 Let (x) ∈ Rn+ , (
x) and (
x) its circular alternating order rearrange-
ment and its circular symmetrical order rearrangement respectively, then when
Heinz and Jensen Type Inequalities and Rearrangements 13


xn+1 =  x1 , xn+1 = x1 , 
xn+1 = 
x1 , we get three cases of Extended Heinz type
inequalities:
(a) For 0 ≤ t < 1, the inequalities

n 1 1
n 
 
1−t
2 
xi 
2 2
xi+1 ≤ xit 
 xi+1 +t
xi+1xi1−t
i=1 i=1
n 
 
1−t
≤ xit xi+1 + xi+1
t
xi1−t
i=1
n 
 
1−t
≤ 
xit 
xi+1 +t
xi+1xi1−t
i=1


n
≤ 2xi
i=1

hold.
(b) For 1 ≤ t ≤ 2 and xi > 0, i = 1, 2, . . . , n, the inequalities

n n 
  n 
 
2 
xi + xi1−t + 1−t
xi+1 (| xi+1 |)t ≥
xi −  
xit 1−t
xi+1 +t
xi+1xi1−t
i=1 i=1 i=1
n 
 
≥ xi1−t xi+1
t
+ xi+1
t
xi1−t
i=1
n 
 
1−t
≥ 
xit 
xi+1 +t
xi+1xi1−t
i=1


n
≥2 xi
i=1

hold.
(c) For t ≥ 2 and xi > 0, i = 1, 2, . . . , n, the inequalities
n 
  n  
1−t

xit 
xi+1 +t
xi+1xi1−t ≥ xi1−t xi+1
t
+ xi+1
t
xi1−t
i=1 i=1
n 
 
1−t
≥ xit 
 xi+1 +t
xi+1xi1−t
i=1


n n 
 
≥2 
xi + xi1−t + 
 1−t
xi+1 xi+1 |)t
xi − 
(|
i=1 i=1

hold.
14 S. Abramovich

Proof From (2.6) in Theorem 2.10 for ϕ(x) = x t , 0 ≤ t ≤ 1 together with the left
hand-side inequality (1.1) we get that Case (a) holds.
From (2.7) in Theorem 2.10 for ϕ(x) = x t , 1 ≤ t ≤ 2 we get that Case (b) holds.
From (2.8) in Theorem 2.10 for ϕ(x) = x t , t ≥ 2 we get that Case (c) holds.
Alternatively we obtain the result of the theorem using Theorem 2.8 and the
concavity, convexity, subquadracity and superquadracity of x t , x > 0 in the relevant
intervals of t. The proof is complete.

References

1. S. Abramovich, The increase of sums and products dependent on (y1 , . . . , yn ) by rearrangement

of this set. Israel J. Math. 5, 177–181 (1967)
2. S. Abramovich, G. Jameson, G. Sinnamon, Refining Jensen’s inequality. Bull. Math. Soc. Sci.
Math. Roumanie (N.S.) 47(95)(1–2), 3–14 (2004)
3. S. Abramovich, L.-E. Persson, Rearrangements and Jensen type inequalities related to convexity,
superquadracity, strong convexity and 1-quasiconvexity. J. Math. Inequal. 14, 641–659 (2020).
4. S. Abramovich, L.-E. Persson, N. Samko, On γ -quasiconvexity, superquadracity and two sided
reversed Jensen type inequalities. Math. Inequal. Appl. 18, 615–628 (2015)
5. F. Kittaneh, M.S. Moslehian, M. Sababheh, Quadratic interpolation of the Heinz mean. Math.
Inequal. Appl. 21, 739–757 (2018)
6. A.L. Lehman, Results on rearrangements. Israel J. Math. 1, 22–28 (1963)
7. C.P. Niculescu, L.-E. Persson, Convex Functions and Their Applications - A Contemporary
Approach, 2nd edn. (Springer, Cham, 2018)
8. H. Yu, Circular rearrangement inequality. J. Math. Inequal. 12, 635–643 (2018)
On Some Applications of Representable
and Continuous Functionals of Banach
Quasi *-Algebras

Maria Stella Adamo

Abstract This survey aims to highlight some of the consequences that repre-
sentable (and continuous) functionals have in the framework of Banach quasi
*-algebras. In particular, we look at the link between the notions of *-semisimplicity
and full representability in which representable functionals are involved. Then,
we emphasize their essential role in studying *-derivations and representability
properties for the tensor product of Hilbert quasi *-algebras, a special class of
Banach quasi *-algebras.

Keywords Representable functionals · Banach and Hilbert quasi *-algebras ·

Weak derivations on Banach quasi *-algebras · Tensor product of Hilbert quasi
*-algebras

Mathematics Subject Classification (2010) Primary 46L08; Secondary 46A32,

46L57, 46L89, 47L60

1 Introduction and Preliminaries

The investigation of (locally convex) quasi *-algebras was undertaken around the
beginning of the ’80s, in the last century, to give a solution to specific problems
concerning quantum statistical mechanics and quantum field theory, that required
instead a representation of observables as unbounded operators, see, e.g., [9, 28].
They were introduced by G. Lassner in the series of papers [21] and [22] in 1988.
A particular interest has been shown for the theory of *-representations of
quasi *-algebras in a specific family of unbounded densely defined and closable
operators. In this framework, a central role is played by representable functionals,
i.e., those functionals that admit a GNS-like construction. In the process of looking

M. S. Adamo ()
Dipartimento di Matematica, Università di Roma “Tor Vergata”, Roma, Italy
e-mail: [email protected]; [email protected]

© Springer Nature Switzerland AG 2021 15

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_2
16 M. S. Adamo

at structural properties of locally convex quasi *-algebras, *-semisimplicity and

full representability are the critical notions involved, making an extensive use of
representable (and continuous) functionals [4, 7, 15, 30].
The goal of this survey is to point out some of the various connections that
representable and continuous functionals for Banach quasi *-algebras have and
examine their applications, e.g., the study of unbounded *-derivations and tensor
products. The importance of investigating these themes stem from the study of
physical phenomena. Moreover, very little is known about unbounded *-derivations
and tensor products in the framework of unbounded operator algebras. The reader
is referred to as [16–18, 31–33].
Banach quasi *-algebras constitute a particular subclass of locally convex
*-algebras. In this context, the interesting question is to understand whether repre-
sentable functionals are continuous. The reasons to get to know about the continuity
of these functionals are several. Among them, the continuity is a crucial feature for
representable functionals, since it would reflect on the sesquilinear forms and the *-
representations associated with them through the GNS-like construction (see [30]).
Furthermore, no example of representable functional that is not continuous is known
in the literature.
A positive answer to this question has been given for the space L2 (I, dλ), where
I = [0, 1] and λ is the Lebesgue measure, over continuous or essentially bounded
functions, and more in general, for commutative Hilbert quasi *-algebras under
certain conditions. In the case of Lp -spaces for p ≥ 1, we observe a discontinuous
behaviour in the quantity of representable and continuous functional when p ≥ 1
gets bigger. For p ≥ 2, the Lp -spaces turn out to be fully representable and *-
semisimple Banach quasi *-algebras and in this case these notions coincide (see
Sect. 2, [4, 7, 8, 10, 15]).
Hilbert quasi *-algebras constitute a class of *-semisimple and fully repre-
sentable quasi *-algebras. In this case, representable and continuous functionals
are in 1-1 correspondence with weakly positive and bounded elements. This result
allows us to get the representability for the tensor product of two representable
and continuous functionals in the framework of tensor product of Hilbert quasi *-
algebras (refer to [2, 3]).
*-Semisimplicity allows us to define a proper notion of *-derivation in the case of
Banach quasi *-algebras and prove a result characterizing infinitesimal generators
of one-parameter group of *-automorphisms in the Banach quasi *-algebras case,
extending results of Bratteli–Robinson for C*-algebras (see, e.g., [11]).
The survey is structured as follows. Firstly, we give some preliminaries about
(Banach) quasi *-algebras, representable (and continuous) functionals and the
GNS-construction. For these quasi *-algebras, in Sect. 2 we recall the notions
of *-semisimplicity and full representability, summing up the main results about
their link in the Banach case and the characterization we have for Hilbert quasi
*-algebras. In Sect. 3, we concentrate on the case of *-semisimple Banach quasi *-
algebras and deal with weak *-derivations and related results. In the last section, we
look at the construction of the tensor product Hilbert quasi *-algebra and look at the
On Some Applications of Representable and Continuous Functionals 17

properties of tensor products of representable functionals and the sesquilinear forms

involved in the definition of *-semisimplicity.
For the reader’s convenience, we recall some preliminary notions for future use.
Further details can be found in [7].
Definition 1.1 A quasi *-algebra (A, A0 ) (or over A0 ) is a pair consisting of a
vector space A and a *-algebra A0 contained in A as a subspace and such that
(i) the left multiplication ax and the right multiplication xa of an element a ∈ A
and x ∈ A0 are always defined and bilinear;
(ii) (xa)y = x(ay) and a(xy) = (ax)y for each x, y ∈ A0 and a ∈ A;
(iii) an involution ∗, which extends the involution of A0 , is defined in A with the
property (ax)∗ = x ∗ a ∗ and (xa)∗ = a ∗ x ∗ for all a ∈ A and x ∈ A0 .
We say that a quasi *-algebra (A, A0 ) has a unit, if there is an element in A0 ,
denoted by 1, such that a1 = a = 1a for all a ∈ A. If a unit exists, then it is always
unique.
Definition 1.2 Let (A, A0 ) be a quasi *-algebra. A linear functional ω : A → C
satisfying
(L.1) ω(x ∗ x) ≥ 0 for all x ∈ A0 ;
(L.2) ω(y ∗ a ∗ x) = ω(x ∗ ay) for all x, y ∈ A0 , a ∈ A;
(L.3) for all a ∈ A, there exists γa > 0 such that

|ω(a ∗ x)| ≤ γa ω(x ∗ x) 2 ,

1
∀x ∈ A0 ,

is called representable on the quasi *-algebra (A, A0 ).

The family of all representable functionals on the quasi *-algebra (A, A0 ) will be
denoted by R(A, A0 ).
This definition is justified by the following Theorem 1.4, proving the existence of
a GNS-like construction of a *-representation πω and a Hilbert space Hω for a
representable functional ω on A.
Let H be a Hilbert space and let D be a dense linear subspace of H. We denote
by L† (D, H) the following family of closable operators:
 
L† (D, H) = X : D → H : D(X) = D, D(X∗ ) ⊃ D .

L† (D, H) is a C-vector space with the usual sum and scalar multiplication. If we
define the involution † and partial multiplication  as

X → X† ≡ X∗ D and XY = X†∗ Y,

then L† (D, H) is a partial *-algebra defined in [7].

18 M. S. Adamo

Definition 1.3 A *-representation of a quasi *-algebra (A, A0 ) is a *-homomor-

phism π : A → L† (Dπ , Hπ ), where Dπ is a dense subspace of the Hilbert space
Hπ , with the following properties:
(i) π(a ∗ ) = π(a)† for all a ∈ A;
(ii) if a ∈ A and x ∈ A0 , then π(a) is a left multiplier of π(x) and

π(a)π(x) = π(ax).

A *-representation π is
• cyclic if π(A0 )ξ is dense in Hπ for some ξ ∈ Dπ ;
• closed if π coincides with its closure 
π defined in [30, Section 2].
If (A, A0 ) has a unit 1, then we suppose that π(1) = ID , the identity operator of D.
Theorem 1.4 ([30]) Let (A, A0 ) be a quasi *-algebra with unit 1 and let ω be a
linear functional on (A, A0 ) that satisfies the conditions (L.1)–(L.3) of Definition
1.2. Then, there exists a closed cyclic *-representation πω of (A, A0 ), with cyclic
vector ξω such that

ω(a) = πω (a)ξω |ξω  , ∀a ∈ A.

This representation is unique up to unitary equivalence.

1.1 Normed Quasi *-Algebras

Definition 1.5 A quasi *-algebra (A, A0 ) is called a normed quasi *-algebra if a

norm  ·  is defined on A with the properties
(i) a ∗  = a, ∀a ∈ A;
(ii) A0 is dense in A;
(iii) for every x ∈ A0 , the map Rx : a ∈ A → ax ∈ A is continuous in A.
The continuity of the involution implies that
(iii’) for every x ∈ A0 , the map Lx : a ∈ A → xa ∈ A is continuous in A.
Definition 1.6 If (A,  · ) is a Banach space, we say that (A, A0 ) is a Banach quasi
*-algebra.
The norm topology of A will be denoted by τn .
An important class of Banach quasi *-algebras is given by Hilbert quasi *-
algebras.
Definition 1.7 Let A0 be a *-algebra which is also a pre-Hilbert space with respect
to the inner product ·|· such that:
On Some Applications of Representable and Continuous Functionals 19

(1) the map y → xy is continuous with respect to the norm defined by the inner
product;
(2) xy|z = y|x ∗ z for all x, y, z ∈ A0 ;
(3) x|y = y ∗ |x ∗  for all x, y ∈ A0 ;
(4) A20 is total in A0 .
Such a *-algebra A0 is said to be a Hilbert algebra. If H denotes the Hilbert space
completion of A0 with respect to the inner product ·|·, then (H, A0 ) is called a
Hilbert quasi *-algebra.

2 Analogy Between *-Semisimplicity and Full

Representability

Representable functionals constitute a valid tool for investigating structural prop-

erties of Banach quasi *-algebras. Indeed, the notions of *-semisimplicity and full
representability are strongly related to these functionals. They turn out to be the
same notion when dealing with Banach quasi *-algebras that verify the condition
(P ). For further reading, see [4, 7, 8, 15, 30].
If ω ∈ R(A, A0 ), then we can associate with it the sesquilinear form ϕω defined
on A0 × A0 as

ϕω (x, y) := ω(y ∗ x), x, y ∈ A0 . (2.1)

If (A[τn ], A0 ) is a normed quasi *-algebra, we denote by Rc (A, A0 ) the subset

of R(A, A0 ) consisting of continuous functionals.
Let ω ∈ Rc (A, A0 ). Then ω is continuous on A, but ϕω is not necessarily
continuous on A0 × A0 . ϕω is said to be closable if for every sequence of elements
{xn } in A0 such that
τn
xn −
→0 and ϕω (xn − xm , xn − xm ) → 0 as n→∞ (2.2)

then ϕω (xn , xn ) → 0 as n → ∞.
By the condition (2.2), the closure of ϕω , denoted by ϕ ω , is a well-defined
sesquilinear form on D(ϕ ω ) × D(ϕ ω ) as

ϕ ω (a, a) := lim ϕω (xn , xn ),

n→∞

where D(ϕ ω ) is the following dense domain

τn
D(ϕ ω ) = {a ∈ A : ∃{xn } ⊂ A0 s.t. xn −
→ a and
ϕω (xn − xm , xn − xm ) → 0}.
20 M. S. Adamo

For a locally convex quasi *-algebra (A, A0 ), ϕ ω always exists, [15]. Neverthe-
less, it is unclear whether D(ϕ ω ) is the whole space A. We show in Proposition 2.1
below that D(ϕ ω ) is A in the case of a Banach quasi *-algebra.
Proposition 2.1 ([4]) Let (A, A0 ) be a Banach quasi *-algebra with unit 1, ω ∈
Rc (A, A0 ) and ϕω the associated sesquilinear form on A0 × A0 defined as in (2.1).
Then D(ϕ ω ) = A; hence ϕ ω is everywhere defined and bounded.
Set now
 n 

+
A0 := xk∗ xk , xk ∈ A0 , n ∈ N .
k=1

Then A+ +
0 is a wedge in A0 and we call the elements of A0 positive elements of A0 .
τn τn
As in [15], we call positive elements of A the elements of A+
0 . We set A+ := A+
0 .
Definition 2.2 A linear functional on A is positive if ω(a) ≥ 0 for every a ∈ A+ .
A family of positive linear functionals F on (A, A0 ) is called sufficient if for every
a ∈ A+ , a = 0, there exists ω ∈ F such that ω(a) > 0.
Definition 2.3 A normed quasi ∗ -algebra (A, A0 ) is called fully representable if
Rc (A, A0 ) is sufficient and D(ϕ ω ) = A for every ω in Rc (A, A0 ).
We denote by QA0 (A) the family of all sesquilinear forms  : A × A → C such
that
(i) (a, a) ≥ 0 for every a ∈ A;
(ii) (ax, y) = (x, a ∗y) for every a ∈ A, x, y ∈ A0 ;
We denote by SA0 (A) the subset of QA0 (A) consisting of all continuous sesquilinear
forms having the property that also
(iii) |(a, b)| ≤ ab, for all a, b ∈ A.
Definition 2.4 A normed quasi *-algebra (A, A0 ) is called *-semisimple if, for
every 0 = a ∈ A, there exists  ∈ SA0 (A) such that (a, a) > 0.
Proposition 2.1 is useful to show the following result, clarifying the link between
*-semisimplicity and full representability. We need first to introduce the following
condition of positivity

a ∈ A and ω(x ∗ ax) ≥ 0 ∀ω ∈ Rc (A, A0 ) and x ∈ A0 ⇒ a ∈ A+ . (P)

Theorem 2.5 ([4]) Let (A, A0 ) be a Banach quasi *-algebra with unit 1. The
following statements are equivalent.
(i) Rc (A, A0 ) is sufficient.
(ii) (A, A0 ) is fully representable.
On Some Applications of Representable and Continuous Functionals 21

If the condition (P ) holds, (i) and (ii) are equivalent to the following
(iii) (A, A0 ) is *-semisimple.
The condition (P ) is not needed to show (iii) ⇒ (ii) of Theorem 2.5.
Theorem 2.5 shows the deep connection between full representability and *-
semisimplicity for a Banach quasi *-algebra. Under the condition of positivity (P ),
the families of sesquilinear forms involved in the definitions of full representability
and *-semisimplicity can be identified.
For a Hilbert quasi *-algebra (H, A0 ), representable and continuous functionals
are in 1-1 correspondence with a certain family of elements in H.
Definition 2.6 Let (H, A0 ) be a Hilbert quasi *-algebra. An element ξ ∈ H is
called
(i) weakly positive if the operator Lξ : A0 → H defined as Lξ (x) := ξ x is positive.
(ii) bounded if the operator Lξ : A0 → H is bounded.
+ (resp.
The set of all weakly positive (resp. bounded) elements will be denoted as Hw
Hb ), see [4, 29].
Theorem 2.7 ([4]) Suppose that (H, A0 ) is a Hilbert quasi *-algebra. Then ω ∈
Rc (H, A0 ) if, and only if, there exists a unique weakly positive bounded element
η ∈ H such that

ω(ξ ) = ξ |η , ∀ξ ∈ H.

2.1 Case of Lp (I, dλ) for p ≥ 1

Consider A0 to be L∞ (I, dλ), where I is a compact interval of the real line and λ is
the Lebesgue measure. Let τn be the topology generated by the p-norm
 1
p
f p := |f |p dλ , ∀f ∈ L∞ (I, dλ),
I

for p ≥ 1. Then, the completion of L∞ (I, dλ) with respect to the  · p -norm is
given by Lp (I, dλ).
We conclude that, for p ≥ 1, (Lp (I, dλ), L∞ (I, dλ)) is a Banach quasi *-
algebra. The same conclusion holds if we consider the *-algebra of continuous
functions over I = [0, 1], denoted by C(I ).
For p ≥ 2, (Lp (I, dλ), L∞ (I, dλ)) and (Lp (I, dλ), C(I )) are fully repre-
sentable and *-semisimple Banach quasi *-algebras. For 1 ≤ p < 2, we have
Rc (Lp (I, dλ), A0 ) = {0} for both A0 = L∞ (I, dλ) or A0 = C(I ). Note that the
Banach quasi *-algebras (Lp (I, dλ), L∞ (I, dλ)) and (Lp (I, dλ), C(I )) verify the
condition (P ) for all p ≥ 1.
22 M. S. Adamo

The absence of representable and continuous functionals that we observe passing

the threshold p = 2 and the lack in the literature of an example of representable
functional that is not continuous led us to pose the following question.
Question Is every representable functional on a Banach quasi *-algebra continu-
ous?
The answer to this question is positive in the case of the Hilbert quasi *-algebras
(L2 (I, dλ), L∞ (I, dλ)) and (L2 (I, dλ), C(I )), see [4].
Theorem 2.8 ([4]) Let ω be a representable functional on the Hilbert quasi
*-algebra (L2 (I, dλ), L∞ (I, dλ)). Then there exists a unique bounded finitely
additive measure ν on I which vanish on subsets of I of zero λ-measure and a
unique bounded linear operator S : L2 (I, dλ) → L2 (I, dν) such that

ω(f ) = (Sf )dν, ∀f ∈ L2 (I, dλ).
I

The operator S satisfies the following conditions:

S(f φ) = (Sf )φ = φ(Sf ) ∀f ∈ L2 (I, dλ), φ ∈ L∞ (I, dλ);

Sφ = φ, ∀φ ∈ L∞ (I, dλ).

Thus, every representable functional ω on L2 (I, dλ), L∞ (I, dλ) is continuous.
Theorem 2.9 ([4]) Let ω be a representable functional on the Hilbert quasi *-
algebra (L2 (I, dλ), C(I )). Then there exists a unique Borel measure μ on I and
a unique bounded linear operator T : L2 (I, dλ) → L2 (I, dμ) such that

ω(f ) = (Tf )dμ, ∀f ∈ L2 (I, dλ).
I

The operator T satisfies the following conditions:

T (f φ) = (Tf )φ = φ(Tf ) ∀f ∈ L2 (I, dλ), φ ∈ C(I );

T φ = φ, ∀φ ∈ C(I ).

Thus, every representable functional ω on L2 (I, dλ), C(I ) is continuous. More-
over, μ is absolutely continuous with respect to λ.
The above theorems can be extended to the case of a commutative Hilbert quasi
*-algebra, under certain hypotheses.
Theorem 2.10 ([4]) Let (H, A0 ) be a commutative Hilbert quasi *-algebra with
unit 1. Assume that A0 [ · 0 ] is a Banach *-algebra and that there exists an
element x of A0 such that the spectrum σ (R x ) of the bounded operator R x of right
On Some Applications of Representable and Continuous Functionals 23

multiplication by x consists only of its continuous part σc (R x ). If ω is representable

on (A, A0 ), then ω is bounded.
The main tools to get these results are the Riesz-Markov Representation Theorem
for continuous functions on a compact space and the intertwining theory on Hilbert
spaces (see [26]). Unfortunately, these tools turn out to be unsuitable to the case of
a (non-commutative) Hilbert quasi *-algebra or a Banach quasi *-algebra.

3 First Application: Derivations and Their Closability

*-Derivations have been widely employed to describe the dynamics for a quantum
phenomenon. For a quantum system of finite volume V , the Hamiltonian belongs
to the local C*-algebra and implements the inner *-derivation for the dynamics.
Nevertheless, the thermodynamical limit in general fails to exist, see [7].
Under some assumptions, the limit turns out to be a weak *-derivation generating
a one parameter group of weak *-automorphisms, defined for a *-semisimple
Banach quasi *-algebra, as we will investigate in the following section. For detailed
discussion, see [5, 6, 20].
The derivation δ : A0 [ · ] → A[ · ] is densely defined. If δ is closable, then
its closure δ as a linear map is not a derivation in general.
• In this section, we only consider *-semisimple Banach quasi *-algebras, if not
otherwise specified.
For these Banach quasi *-algebras, define a weaker multiplication in A as in [29].
Definition 3.1 Let (A, A0 ) be a Banach quasi *-algebra. Let a, b ∈ A. We say that
the weak multiplication a b is well-defined if there exists a (necessarily unique)
c ∈ A such that:

ϕ(bx, a ∗y) = ϕ(cx, y), ∀ x, y ∈ A0 , ∀ ϕ ∈ SA0 (A).

In this case, we put a b := c.

Let a ∈ A. The space of right (resp. left) weak multipliers of a, i.e., the space of
all the elements b ∈ A such that a b (resp. ba) is well-defined, will be denoted as
Rw (a) (resp. Lw (a)). We indicate by Rw (A) (resp. Lw (A)) the space of universal
right (resp. left) multipliers of A, i.e., all the elements b ∈ A such that b ∈ Rw (a)
(resp. b ∈ Lw (a)) for every a ∈ A. Clearly, A0 ⊆ Rw (A) ∩ Lw (A).
Let (A, A0 ) be a Banach quasi *-algebra. For every a ∈ A, two linear operators
La and Ra are defined in the following way

La : A0 → A La (x) = ax ∀x ∈ A0 (3.1)
Ra : A0 → A Ra (x) = xa ∀x ∈ A0 . (3.2)
24 M. S. Adamo

An element a ∈ A is called bounded if the operators La and Ra defined in (3.1)

and (3.2) are  · -continuous and thus extendible to the whole space A. As for
Hilbert quasi *-algebras in Definition 2.6, the set of all bounded elements will be
denoted by Ab .
Lemma 3.2 ([5, Lemma 2.16]) If (A, A0 ) is a *-semisimple Banach quasi *-algeb-
ra with unit 1, the set Ab of bounded elements coincides with the set Rw (A)∩Lw (A).
Let θ : A → A be a linear bijection. We say that θ is a weak *-automorphism of
(A, A0 ) if
(i) θ (a ∗ ) = θ (a)∗ , for every a ∈ A;
(ii) θ (a)θ (b) is well defined if, and only if, a b is well defined and, in this case,

θ (a b) = θ (a)θ (b).

Definition 3.3 Let βt be a weak *-automorphism of A for every t ∈ R. If

(i) β0 (a) = a, ∀a ∈ A
(ii) βt +s (a) = βt (βs (a)), ∀a ∈ A
then we say that βt is a one-parameter group of weak *-automorphisms of (A, A0 ).
If τ is a topology on A and the map t → βt (a) is τ -continuous, for every a ∈ A, we
say that βt is a τ -continuous weak *-automorphism group.
In this case

βt (a) − a
D(δτ ) = a ∈ A : lim exists in A[τ ]
t →0 t

and
βt (a) − a
δτ (a) = τ − lim , a ∈ D(δτ ).
t →0 t

Then δτ is called the infinitesimal generator of the one-parameter group {βt } of

weak *-automorphisms of (A, A0 ).
Definition 3.4 ([5]) Let (A, A0 ) be a *-semisimple Banach quasi *-algebra and δ
a linear map of D(δ) into A, where D(δ) is a partial *-algebra with respect to the
weak multiplication . We say that δ is a weak *-derivation of (A, A0 ) if
(i) A0 ⊂ D(δ)
(ii) δ(x ∗ ) = δ(x)∗ , ∀x ∈ A0
(iii) if a, b ∈ D(δ) and a b is well defined, then a b ∈ D(δ) and

ϕ(δ(a b)x, y) = ϕ(bx, δ(a)∗y) + ϕ(δ(b)x, a ∗ y),

for all ϕ ∈ SA0 (A), for every x, y ∈ A0 .

On Some Applications of Representable and Continuous Functionals 25

Analogously to the case of a C*-algebra, to a uniformly bounded norm continu-

ous weak *-automorphisms group there corresponds a closed weak *-derivation that
generates the group (see [11]).
Theorem 3.5 ([5]) Let δ : D(δ) → A[ · ] be a weak *-derivation on a *-
semisimple Banach quasi *-algebra (A, A0 ). Suppose that δ is the infinitesimal
generator of a uniformly bounded, τn -continuous group of weak *-automorphisms
of (A, A0 ). Then δ is closed; its resolvent set ρ(δ) contains R \ {0} and

δ(a) − λa ≥ |λ| a, a ∈ D(δ), λ ∈ R.

Theorem 3.6 ([5]) Let δ : D(δ) ⊂ Ab → A[ · ] be a closed weak *-derivation on

a *-semisimple Banach quasi *-algebra (A, A0 ). Suppose that δ verifies the same
conditions on its spectrum of Theorem 3.5 and A0 is a core for every multiplication
operator L̂a for a ∈ A, i.e. L̂a = La . Then δ is the infinitesimal generator of a
uniformly bounded, τn -continuous group of weak *-automorphisms of (A, A0 ).
In Theorem 3.6, we assumed further conditions, for instance that the domain D(δ)
of the weak *-derivation is contained in Ab , which turns out to be satisfied in some
interesting situations such as the weak derivative in Lp -spaces.
Example 3.7 (Inner Weak *-Derivations for Unbounded Hamiltonian h) Let
(A, A0 ) be a *-semisimple Banach quasi *-algebra. Let h ∈ A be a self-adjoint
unbounded element, i.e., h = h∗ and σ (h) ⊂ R. We define the following derivation

δh : A0 → A, defined as δh (x) = i[h, x], x ∈ A0 .

Then, for every fixed t ∈ R, βt (a) = eit h a e−it h is a well-defined weak

*-automorphism of (A, A0 ) since eit h , e−it h are bounded elements in A (see
[5]) and thus (eit h a)e−it h = eit h (a e−it h ) for every a ∈ A. Moreover, βt
is a uniformly bounded norm continuous group of weak *-automorphisms. The
infinitesimal generator is given by

βt (a) − a eit h a e−it h − a

δ h (a) := lim = lim = i[ha − a h],
t →0 t t →0 t

when a is bounded.
In the following Proposition, the *-semisimplicity is automatically given by
assuming the existence of a representable and continuous functional with associated
faithful *-representation (see [1, 11]).
Proposition 3.8 ([1]) Let (A, A0 ) be a Banach quasi *-algebra with unit 1 and
let δ be a weak *-derivation of (A, A0 ) such that D(δ) = A0 . Suppose that there
exists a representable and continuous functional ω with ω(δ(x)) = 0 for x ∈ A0
and let (Hω , πω , λω ) be the GNS-construction associated with ω. Suppose that πω
26 M. S. Adamo

is a faithful *-representation of (A, A0 ). Then there exists an element H = H † of

L† (λω (A0 )) such that

πω (δ(x)) = −i[H, πω (x)], ∀x ∈ A0

and δ is closable.

4 Second Application: Tensor Product of Hilbert Quasi

*-Algebras

In this section, we recall the construction of the tensor product Hilbert quasi
*-algebra of two given Hilbert quasi *-algebras (H1 , A0 ) and (H2 , B0 ), giving
some results about the relationship between the representability properties for the
tensor product and those for the factors. For further reading on the algebraic and
topological tensor product, see [12–14, 19, 23, 24], for the tensor product Hilbert
quasi *-algebras refer to [2, 3].
For convenience, we assume that (H1 , A0 ) and (H2 , B0 ) are unital Hilbert quasi
*-algebras.
The algebraic tensor product A0 ⊗ B0 is the tensor product *-algebra of A0
and B0 , it is endowed with the canonical multiplication and involution and it is
considered as a subspace of the tensor product vector space H1 ⊗ H2 .
If H1 and H2 are endowed with the inner product ·|·1 and ·|·2 respectively,
then A0 ⊗ B0 satisfies the requirements of Definition 1.7 if we endow it with the
following well-defined inner product

 m 
n    
 
z|z h
:= xi |xj yi |yj , ∀z, z ∈ A0 ⊗ B0 , (4.1)
1 2
i=1 j =1

n m
where z = i=1 xi ⊗ yi and z = j =1 xj ⊗ yj (see [24, 25, 27]). Then, the
completion of A0 ⊗ B0 with respect to the norm  · h induced by the inner product
in (4.1) is a Hilbert quasi *-algebra. Since A0 , B0 are respectively dense in H1 , H2
and  · h is a cross-norm, i.e. x ⊗ yh = x1 y2 for all x ⊗ y ∈ A0 ⊗ B0 , the
tensor product *-algebra A0 ⊗ B0 is  · h -dense in H1 ⊗h H2 . We conclude that

h B0 ≡ H1 ⊗
A0 ⊗ h H2

and (H1 ⊗ h H2 , A0 ⊗ B0 ) is a Hilbert quasi *-algebra, see [2, 3].

For the tensor product Hilbert quasi *-algebra (H1 ⊗ h H2 , A0 ⊗B0 ) we can apply
all the known results for Banach quasi *-algebras about representability presented in
h H2 , A0 ⊗B0 ) is always a *-
Sect. 2, see also [4, 5]. In particular, we know that (H1 ⊗
On Some Applications of Representable and Continuous Functionals 27

semisimple and fully representable Hilbert quasi *-algebra, applying Theorems 2.5
and 2.7.
Employing Theorem 2.7 and Lemma 4.1 in [2], we can give an alternative proof
for Theorem 4.1. The proof will be given below after the proof of Theorem 4.2.
Theorem 4.1 ([2]) Let (H1 ⊗ h H2 , A0 ⊗ B0 ) be the tensor product Hilbert quasi *-
algebra of (H1 , A0 ) and (H2 , B0 ). Then, if ω1 , ω2 are representable and continuous
functionals on H1 and H2 respectively, then ω1 ⊗ ω2 extends to a representable and
continuous functional  on the tensor product Hilbert quasi *-algebra H1 ⊗ h H2 .
We now look at what happens to the sesquilinear forms in SA0 (H1 ) and
SB0 (H2 ).
Theorem 4.2 Let (H1 ⊗ h H2 , A0 ⊗ B0 ) be the tensor product Hilbert quasi *-
algebra of (H1 , A0 ) and (H2 , B0 ). Let φ1 ∈ SA0 (H1 ) and φ2 ∈ SB0 (H2 ). Then,
φ2 ∈ SA0 ⊗B0 (H1 ⊗
φ1 ⊗ h H2 ), i.e., it satisfies the conditions (i), (ii) and (iii) after
the Definition 2.3.
Proof Let φ1 ∈ SA0 (H1 ) and φ2 ∈ SB0 (H2 ). These sesquilinear forms are
continuous, thus by the representation theorem for bounded sesquilinear forms
over a Hilbert space, there exist unique bounded operators T1 : H1 → H1 and
T2 : H2 → H2 such that
   
φ1 (ξ, ξ ) = ξ |T1 ξ and φ2 (η, η ) = η|T2 η ,

for all ξ, ξ ∈ H1 , η, η ∈ H2 , and Ti  = φi  ≤ 1 for i = 1, 2. Moreover, T1 and

T2 are positive operators on H1 and H2 respectively.
On the pre-completion H1 ⊗h H2 , define the tensor product of φ1 and φ2 as

n
φ1 ⊗ φ2 (ζ, ζ ) := φ1 (ξi , ξj )φ2 (ηi , ηj )
i,j =1
 
for all ζ = ni=1 ξi ⊗ ηi , ζ = nj=1 ξj ⊗ ηj in H1 ⊗h H2 .
By [19], it is known that T1 ⊗ T2 extends to a bounded operator T1 ⊗ T2 on the
completion H1 ⊗  H2 such that T1 ⊗
h T2  ≤ T1 T2  ≤ 1.
We show that φ1 ⊗ φ2 is represented by T1 ⊗ T2 on H1 ⊗h H2 . Thus, φ1 ⊗ φ2
is continuous and can be extended to H1 ⊗h H2 . Its extension will be denoted by
φ2 and it corresponds to the bounded operator T1 ⊗
φ1 ⊗ T2 . Indeed, by the definition
of φ1 ⊗ φ2 , we have

n
φ1 ⊗ φ2 (ζ, ζ ) = φ1 (ξi , ξj )φ2 (ηi , ηj )
i,j =1
n 
  
= ξi |T1 ξj ηi , T2 ηj
i,j =1
 
= ζ |(T1 ⊗ T2 )ζ .
28 M. S. Adamo

To conclude the proof, we show that φ1 ⊗ φ2 is in SA0 ⊗B0 (H1 ⊗ h H2 ). Indeed,
φ2 is a positive sesquilinear form, since T1 ⊗
φ1 ⊗ T2 is a positive operator as a tensor
product of positive operators on a Hilbert space. Thus, the condition (i) is verified.
The condition (ii) can be easily verified using the corresponding properties of φ1
and φ2 . For (iii), we know that

φ2  = T1 ⊗
φ1 ⊗ T2  ≤ T1 T2  ≤ 1.

φ2 to belong to SA0 ⊗B0 (H1 ⊗

Hence, all the conditions for φ1 ⊗ h H2 ) are verified.

Using the same argument as in Theorem 2.5 (refer to [4] for a complete proof), the
sesquilinear form ϕ ω associated with a representable and continuous functional ω
in a Hilbert quasi *-algebra (H, A0 ) with unit 1 is bounded and in QA0 (H). Hence,
it is represented by a bounded operator Sω such that

ω(ξ ) = ϕ ω (ξ, 1) = ξ |Sω 1 , ξ ∈ H.

We want to show that Sω 1 is a weakly positive and bounded element. Indeed, for all
x ∈ A0 , we have
   
x|RSω 1 x = x|x(Sω 1) = x ∗ x|Sω 1 = ω(x ∗ x) ≥ 0,

thus Sω 1 is weakly positive. Moreover, Sω 1 is bounded by the condition (L.3) of

representability in Definition 1.2. Indeed, (L.3) means that for every ξ ∈ H, there
exists γξ > 0 such that
1
ω(ξ ∗ x) ≤ γξ ω(x ∗ x) 2 ,

for all x ∈ A0 . By Proposition 2.1, ϕ ω is everywhere defined and bounded, hence

ω(ξ ∗ x) = ϕ ω (x, ξ ) ≤ cξ x,

for some positive constant cξ . Hence, by the

 Riesz representation theorem, there
exists χξ ∈ H such that ω(ξ ∗ x) = x|χξ for all x ∈ A0 . Therefore, the weak
product (in the sense of [4, Definition 4.4]) ξ Sω 1 is well-defined for all ξ ∈ H.
Indeed, for x, y ∈ A0 , we have
       
ξ ∗ x|Sω 1y = ξ ∗ xy ∗ |Sω 1 = xy ∗|χξ = x|χξ y .

A similar argument shows that Sω 1ξ is well-defined for all ξ ∈ H. Recall that an
element ξ ∈ H is bounded if, and only if, Rw (ξ ) = Lw (ξ ) = H, where Rw (ξ )
(resp. Lw (ξ )) is the space of universal right (resp. left) weak multipliers of ξ (see
[4, Proposition 4.10]). Then, Sω 1 is a bounded element.
On Some Applications of Representable and Continuous Functionals 29

By Theorem 2.7, we have that Sω 1 is the weakly positive and bounded element
in H corresponding to the representable and continuous functional ω.
For what we just argued, we can give an alternative proof of Theorem 4.1.
h H2 are corresponding to
The proof tells us explicitly what the elements in H1 ⊗
ω2 for ω1 ∈ Rc (H1 , A0 ),
representable and continuous functionals of the form ω1 ⊗
ω2 ∈ Rc (H2 , B0 ).
Proof of Theorem 4.1 Let ω1 , ω2 be representable and continuous functionals on
H1 , H2 respectively. Then,
 for what  we discussed above, ω1 (ξ ) = ξ |Sω1 1H1 for
all ξ ∈ H1 and ω2 (η) = η|Sω2 1H2 for η ∈ H2 .
Looking at their tensor product on H1 ⊗h H2 , we have


n
 
ω1 ⊗ ω(ζ ) = ξi ⊗ ηi |Sω1 1H1 ⊗ Sω2 1H2 ,
i=1


for every ζ = ni=1 ξi ⊗ ηi in H1 ⊗h H2 .
Since ϕ ω1 and ϕ ω2 are in QA0 (H1 ) and QB0 (H2 ) respectively, with the
same argument of Theorem 4.2, ϕ ω1 ⊗ ϕ ω2 is continuous and belongs to
QA0 ⊗B0 (H1 ⊗ H2 ). Furthermore, it is represented by Sω1 ⊗
h Sω2 . Hence

Sω1 ⊗ Sω2 (1H1 ⊗ 1H2 ) = Sω1 1H1 ⊗ Sω2 1H2

is weakly bounded and positive. Hence, by Theorem 2.7, ω1 ⊗ ω2 is representable

and continuous. Furthermore, the continuous extension of ω1 ⊗ ω2 is representable.
Indeed, let us denote this extension as , then

(ψ) = lim ω1 ⊗ ω2 (zn )

n→+∞
 
= lim zn |Sω1 ⊗ Sω2 (1H1 ⊗ 1H2 )
n→+∞
 
= ψ|Sω1 ⊗ Sω2 (1H1 ⊗ 1H2 ) ,

H2 .
where {zn } is a sequence of elements in A0 ⊗ B0  · h -converging to ψ ∈ H1 ⊗

It would be of interest to look at the same questions in the more general

framework of Banach quasi *-algebras. This is work in progress with Maria
Fragouloupoulou. In this case, we need further assumptions on the considered cross-
norm to get some of the properties that we showed in this work, see [3].

Acknowledgments The author is grateful to the organizers of the International Workshop on

Operator Theory and its Applications 2019, especially to the Organizers of the section entitled
“Linear Operators and Function Spaces”, for this interesting and delightful conference and the
Instituto Superior Técnico of Lisbon for its hospitality. The author was financially supported by
the ERC Advanced Grant no. 669240 QUEST “Quantum Algebraic Structures and Models”.
30 M. S. Adamo

The author wishes to thank the anonymous referees for their useful suggestions that improved
the presentation of this manuscript.

References

1. M.S. Adamo, The interplay between representable functionals and derivations on Banach
quasi *-algebras, in Proceedings of the International Conference on Topological Algebras and
Their Applications – ICTAA 2018, ed. by M. Abel. Mathematics Studies (Tartu), 7 (Estonian
Mathematical Society, Tartu, 2018), pp. 48–59
2. M.S. Adamo, About tensor product of Hilbert quasi *-algebras and their representability,
(accepted for publication in the Proceedings of OT27, Theta 2020)
3. M.S. Adamo, M. Fragoulopoulou, Tensor products of normed and Banach quasi *-algebras. J.
Math. Anal. Appl. 490, 2 (2020)
4. M.S. Adamo, C. Trapani, Representable and continuous functionals on a Banach quasi ∗ -
algebra. Mediterr. J. Math. 14, 157 (2017)
5. M.S. Adamo, C. Trapani, Unbounded derivations and *-automorphisms groups of Banach
quasi *-algebras. Ann. Mat. Pura Appl. (4) 198, 1711–1729 (2019)
6. J.-P. Antoine, A. Inoue, C. Trapani, O*-dynamical systems and *-derivations of unbounded
operator algebras. Math. Nachr. 204, 5–28 (1999)
7. J.-P. Antoine, A. Inoue, C. Trapani, Partial *-Algebras and Their Operator Realizations
(Kluwer Academic, Dordrecht, 2003)
8. F. Bagarello, C. Trapani, CQ∗ -algebras: structure properties. Publ. RIMS Kyoto Univ. 32, 85–
116 (1996)
9. F. Bagarello, C. Trapani, The Heisenberg dynamics of spin systems: a quasi *-algebra
approach. J. Math. Phys. 37, 4219–4234 (1996)
10. F. Bagarello, C. Trapani, Lp -spaces as quasi *-algebras. J. Math. Anal. Appl. 197, 810–824
(1996)
11. O. Bratteli, D.W. Robinson, Unbounded derivations of C*-algebras. Commun. Math. Phys. 42,
253–268 (1975)
12. L. Chambadal, J.L. Ovaert, Algèbre Linéaire et Algèbre Tensorielle (Dunod Université, Paris,
1968)
13. A. Defant, K. Floret, Tensor Norms and Operator Ideals (North-Holland, Amsterdam, 1993)
14. M. Fragoulopoulou, Topological Algebras with Involution (North-Holland, Amsterdam, 2005)
15. M. Fragoulopoulou, C. Trapani, S. Triolo, Locally convex quasi *-algebras with sufficiently
many *-representations. J. Math. Anal. Appl. 388, 1180–1193 (2012)
16. M. Fragoulopoulou, A. Inoue, M. Weigt, Tensor products of generalized B ∗ -algebras. J. Math.
Anal. Appl. 420, 1787–1802 (2014)
17. M. Fragoulopoulou, A. Inoue, M. Weigt, Tensor products of unbounded operator algebras.
Rocky Mt. 44, 895–912 (2014)
18. W.-D. Heinrichs, Topological tensor products of unbounded operator algebras on Frèchet
domains. Publ. RIMS Kyoto Univ. 33, 241–255 (1997)
19. A.Ya. Helemskii, Lectures and Exercises on Functional Analysis (American Mathematical
Society, Providence, 2006)
20. E. Hille, R. Phillips, Functional Analysis and Semi-groups (American Mathematical Society,
Providence, 1996)
21. G. Lassner, Topological algebras and their applications in quantum statistics. Wiss. Z. KMU-
Leipzig, Math. Naturwiss. R. 30, 572–595 (1981)
22. G. Lassner, Algebras of unbounded operators and quantum dynamics. Physica A 124, 471–480
(1984)
23. K.B. Laursen, Tensor products of Banach algebras with involution. Trans. Am. Math. Soc. 136,
467–487 (1969)
On Some Applications of Representable and Continuous Functionals 31

24. A. Mallios, Topological Algebras. Selected Topics (North-Holland, Amsterdam, 1986)

25. G.J. Murphy, C*-Algebras and Operator Theory (Academic, Boston, 2014)
26. A.M. Sinclair, Automatic Continuity of Linear Operators. London Mathematical Society,
Lecture Notes Series 21 (Cambridge University Press, Cambridge, 1976)
27. M. Takesaki, Theory of Operator Algebras I (Springer, New York, 1979)
28. C. Trapani, Quasi *-algebras of operators and their applications. Rev. Math. Phys. 7, 1303–
1332 (1995)
29. C. Trapani, Bounded elements and spectrum in Banach quasi *-algebras. Stud. Math. 172,
249–273 (2006)
30. C. Trapani, *-Representations, seminorms and structure properties of normed quasi *-algebras.
Stud. Math. 186, 47–75 (2008)
31. M. Weigt, Derivations of τ -measurable operators. Oper. Theory Adv. Appl. 195, 273–286
(2009)
32. M. Weigt, I. Zarakas, Unbounded derivations of GB*-algebras. Oper. Theory Adv. Appl. 247,
69–82 (2015)
33. M. Weigt, I. Zarakas, Derivations of Fréchet nuclear GB*-algebras. Bull. Aust. Math. Soc. 92,
290–301 (2015)
Minimality Properties of Sturm-Liouville
Problems with Increasing Affine
Boundary Conditions

Yagub N. Aliyev

Abstract We consider Sturm-Liouville problems with a boundary condition lin-

early dependent on the eigenparameter. We concentrate the study on the cases where
non-real or non-simple (multiple) eigenvalues are possible. We prove that the system
of root (i.e. eigen and associated) functions of the corresponding operator, with an
arbitrary function removed, form a minimal system in L2 (0, 1), except some cases
where this system is neither complete nor minimal. The method used is based on
the determination of the explicit form of the biorthogonal system. These minimality
results can be extended to basis properties in L2 (0, 1).

Keywords Sturm-Liouville · Eigenparameter-dependent boundary conditions ·

Minimal system · Root functions

Mathematics Subject Classification (2010) Primary 34B24; Secondary 34L10

1 Introduction

Consider the following spectral problem

− y + q(x)y = λy, 0 < x < 1, (1.1)

y(0) cos β = y (0) sin β, 0 ≤ β < π, (1.2)

y(1) = (cλ + d)y (1), (1.3)

This work was completed with the support of ADA University Faculty Research and Development
Fund.

Y. N. Aliyev ()
School of IT and Engineering, ADA University, Baku, Azerbaijan
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 33

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_3
34 Y. N. Aliyev

where c, d are real constants and c > 0, λ is the spectral parameter, q(x) is a real
valued and continuous function over the interval [0, 1].
The present article is about the minimality properties in L2 (0, 1) of the system
of root functions of the boundary value problem (1.1)–(1.3).
It was proved in [2] (see also [3]) that the eigenvalues of the boundary value
problem (1.1)–(1.3) form an infinite sequence accumulating only at +∞ and only
the following cases are possible:
(a) all the eigenvalues are real and simple;
(b) all the eigenvalues are real and all, except one double, are simple;
(c) all the eigenvalues are real and all, except one triple, are simple;
(d) all the eigenvalues are simple and all, except a conjugate pair of non-real, are
real.
Let {vn }∞
n=0 be a sequence of elements from L2 (0, 1) and Vk the closure (in the
norm of L2 (0, 1)) of the linear span of {vn }∞ ∞
n=0, n=k . The system {vn }n=0 is called
minimal in L2 (0, 1) if vk ∈/ Vk for all k = 0, 1, 2, . . ..
The eigenvalues λn (n ≥ 0) will be considered to be listed according to non-
decreasing real part and repeated according to algebraic multiplicity. Asymptotics of
eigenvalues and oscillation of eigenfunctions of the boundary value problem (1.1)–
(1.3), with linear function in the boundary condition, replaced by general rational
function, were studied in paper [3]. The case c < 0 in our problem does not involve
any non-real or non-simple eigenvalues and can be found as a special case in papers
[5, 6].
The main objective of the present paper is to show that the set of root functions
with an arbitrary function removed form a minimal system in L2 (0, 1), except some
cases where this system is neither complete nor minimal. In more general form this
result was obtained in [8, 9]. The main advantage of the current method is that it does
not use any heavy machinery or extensions into more general operators defined over
L2 ⊕ C. Also the special associated functions yk+1 ∗ , y # , y # , which are called in
k+1 k+2
the paper as auxiliary associated functions and used to describe the necessary and
sufficient conditions for minimality in a nice way, are very helpful when a concrete
boundary value problem is studied. One such worked out example is given at the
end of the current paper. The mere number of possibilities that arise in the simple
case of linear dependence shows how different the minimality properties are from
corresponding properties in the cases of boundary conditions free of any eigenvalue
parameters. These minimality results can be used to prove basis properties of the
root functions in L2 (0, 1).

2 Inner Products and Norms of Eigenfunctions

The proofs in this and subsequent sections are similar to the corresponding results
in [1] so we will skip some of them. Let y(x, λ) be a non-zero solution of (1.1),
satisfying initial conditions y(0) = sin β and y (0) = sin β. Then we can write the
Minimality Properties of Sturm-Liouville Problems 35

characteristic equation as

 (λ) = y(1, λ) − (cλ + d)y (1, λ). (2.1)

By (1.3), λn is an eigenvalue of (1.1)–(1.3) if  (λn ) = 0. It is a simple

eigenvalue if  (λn ) = 0 =  (λn ). Similarly, λk is a double eigenvalue if

 (λk ) =  (λk ) = 0 =  (λk ),

and a triple eigenvalue if

 (λk ) =  (λk ) =  (λk ) = 0 =  (λk ).

We also note that y(x, λ) → y(x, λn ), uniformly, for x ∈ [0, 1], as λ → λn , and
the function yn (x) = y(x, λn ) is an eigenfunction of (1.1)–(1.3) corresponding to
eigenvalue λn .
By (1.1)–(1.3) we have,

−yn + q(x)yn = λn yn , yn (0) sin β = yn (0) cos β, yn (1) = (cλn + d)yn (1).

We denote by (·, ·) the inner product in L2 (0, 1).

Lemma 2.1 Let yn , ym be eigenfunctions corresponding to eigenvalues λn , λm
(λn = λm ). Then

(yn , ym ) = cyn (1)ym (1). (2.2)

Proof To begin we note that

d
(y(x, λ)y (x, μ) − y (x, λ)y(x, μ)) = (λ − μ)y(x, λ)y(x, μ).
dx
By integrating this identity from 0 to 1, we obtain

1
(λ − μ)(y(·, λ), y(·, μ)) = (y(x, λ)y (x, μ) − y (x, λ)y(x, μ)) . (2.3)
0

From (1.2), we obtain

y(0, λ)y (0, μ) − y (0, λ)y(0, μ) = 0. (2.4)

By (2.1),

y(1, λ)y (1, μ) − y (1, λ)y(1, μ)

= c(λ − μ)y (1, λ)y (1, μ) − y (1, λ) (μ) + y (1, μ) (λ). (2.5)
36 Y. N. Aliyev

From (2.3)–(2.5), it follows that for λ = μ,

 (μ)  (λ)
(y(·, λ), y(·, μ)) = cy (1, λ)y (1, μ) − y (1, λ) + y (1, μ) . (2.6)
λ−μ λ−μ

Note the fact that  (λn ) =  (λm ) = 0. The required equality (2.2) is now obvious
if we substitute the parameters λ, μ by λn , λm , respectively.
Remark 2.2 Since λn , λm are the eigenvalues of (1.1)–(1.3), it is also possible to
prove (2.2) by letting λ → λn (μ = λn ) and then letting μ → λm in (2.6). It is also
possible to substitute the parameters λ, μ, respectively by λn , λm at the beginning of
the proof, in (2.3). Then the proof would be much simpler. But we chose the present
proof because we will need (2.6) in the subsequent arguments.
In the remaining part of this section we will collect some simple facts about the
inner products and norms of the eigenfunctions.
Lemma 2.3 If λn is a real eigenvalue then

2
yn 22 = (yn , yn ) = cyn (1) + yn (1) (λn ). (2.7)

Corollary 2.4 If λk is a multiple eigenvalue then

2
yk 22 = (yk , yk ) = cyk (1) . (2.8)

An immediate corollary of (2.2) is the following.

Corollary 2.5 If λr is a non-real eigenvalue then

2
yr 22 = c|yr (1)| . (2.9)

For the eigenfunction yn define

2
Bn = yn 22 − c|yn (1)| . (2.10)

The following corollary of (2.7)–(2.9) will be needed (cf. [2, Theorem 4.3]).
Corollary 2.6 Bn = 0 if and only if the corresponding eigenvalue λn is real and
simple.
If λk is a multiple (double or triple) eigenvalue (λk = λk+1 ) then Bk =
yk (1)ω (λk ) = 0 and Bk+1 is not defined, so we set Bk+1 = yk (1)ω (λk )/2. If
λk is a triple eigenvalue (λk = λk+1 = λk+2 ) then Bk+1 = 0 and Bk+2 is not
defined, so we set Bk+2 = yk (1)ω (λk )/6.
We conclude this section with the following.
Minimality Properties of Sturm-Liouville Problems 37

Lemma 2.7 If λr and λs = λr are a conjugate pair of non-real eigenvalues then

2
(yr , ys ) = cyr (1) + yr (1) (λr ). (2.11)

Since in the following text  (λr ) will appear in denominators of some of the
fractions, it is useful to note here that all non-real eigenvalues of (1.1)–(1.3) are
simple and therefore  (λr ) = 0 in (2.11).

3 Inner Products and Norms of Associated Functions

In the previous section we collected some simple facts about the inner products and
norms of the eigenfunctions. But in the case of multiple eigenvalues there are also
some associated functions. In this section we will find formulae for inner products
and norms involving associated functions. These cases appear only for the real
eigenvalues so, throughout these sections we assume that all the eigenvalues and
eigenfunctions are real. In particular, we will not write complex conjugate sign that
appeared in the previous formulae.
If λk is a multiple eigenvalue (λk = λk+1 ) then for a first order associated
function yk+1 corresponding to the eigenfunction yk , following relations hold true
[7, p. 28]:

−yk+1 + q(x)yk+1 = λk yk+1 + yk ,

yk+1 (0) cos β = yk+1 (0) sin β,

yk+1 (1) = (cλk + d)yk+1 (1) + cyk (1).

If λk is a triple eigenvalue (λk = λk+1 = λk+2 ) then together with the first order
associated function yk+1 there exists a second order associated function yk+2 for
which

−yk+2 + q(x)yk+2 = λk yk+2 + yk+1 ,

yk+2 (0) cos β = yk+2 (0) sin β,

yk+2 (1) = (cλk + d)yk+2 (1) + cyk+1 (1).

The following well known properties of the associated functions play an

important role in our investigation. The functions yk+1 +Cyk and yk+2 +Dyk , where
C and D are arbitrary constants, are also associated functions of the first and second
order, respectively. Next, we observe that if we replace the associated function yk+1
by yk+1 + Cyk , then the associated function yk+2 changes to yk+2 + Cyk+1 .
38 Y. N. Aliyev

By differentiating (1.1), (1.2) and (2.1) with respect to λ we obtain

−yλ (x, λ) + q(x)yλ(x, λ) = λyλ (x, λ) + y(x, λ),

yλ (0, λ) cos β = yλ (0, λ) sin β,

 (λ) = yλ (1, λ) − (cλ + d)yλ (1, λ) − cy (1, λ),

where the subscript denotes differentiation with respect to λ.

Let λk be a multiple (double or triple) eigenvalue of (1.1)–(1.3). Since  (λk ) =
 (λk ) = 0, it follows that y(x, λ) → yk , yλ (x, λ) → ỹk+1 , uniformly with respect
to x ∈ [0, 1], as λ → λk , where ỹk+1 is one of the associated functions of the first
order, and it is obvious that ỹk+1 = yk+1 + C̃yk , for a certain constant C̃. Note that
C̃ = (ỹk+1 (1) − yk+1 (1))/yk (1).
Similarly, if we differentiate (1.1), (1.2) and (2.1) with respect to λ again we
obtain

−yλλ (x, λ) + q(x)yλλ(x, λ) = λyλλ (x, λ) + 2yλ(x, λ),

yλλ(0, λ) cos β = yλλ (0, λ) sin β,

 (λ) = yλλ (1, λ) − (cλ + d)yλλ(1, λ) − 2cyλ (1, λ).

We note again that if λk is a triple eigenvalue of (1.1)–(1.3) then  (λk ) = 0,

hence yλλ → 2ỹk+2 , uniformly with respect to x ∈ [0, 1], as λ → λk , where
ỹk+2 is one of the associated functions of the second order corresponding to the first
associated function ỹk+1 , and it is obvious that ỹk+2 = yk+2 + C̃yk+1 + D̃yk , for a
certain constant D̃. Note that D̃ = (ỹk+2 (1) − yk+2 (1) − C̃yk+1 (1))/yk (1).
Lemma 3.1 If λk is a multiple eigenvalue and λn = λk then

(yk+1 , yn ) = cyk+1 (1)yn (1). (3.1)

Proof Differentiating (2.6) with respect to λ we obtain

 (μ)  (μ)
(yλ (·, λ), y(·, μ)) =cyλ (1, λ)y (1, μ) − yλ (1, λ) + y (1, λ)
λ−μ (λ − μ)2
 (λ)  (λ)
+ y (1, μ) − y (1, μ) . (3.2)
λ−μ (λ − μ)2

Letting μ → λn (λ = λn ) and then λ → λk in (3.2) we obtain that (3.1) is true

(cf. [1]). We used the fact that the differentiation and the subsequent passage to
the limit within the integrals are meaningful because all the involved functions are
continuous with respect to both x and λ [4, Ch. 3, §4, Theorems 1, 2].
Minimality Properties of Sturm-Liouville Problems 39

The same result can be achieved if we started with the identity

d 
yk+1 yn − yk+1 yn = (λk − λn )yk+1 yn + yk yn ,
dx
which can be easily derived using the definition of yk+1 . By integrating this equality
from 0 to 1 we will obtain
 1
(λk − λn )(yk+1 , yn ) + (yk , yn ) = yk+1 yn − yk+1 yn 0
.

By using the boundary conditions for yk+1 and yn , and the fact that (yk , yn ) =
cyk (1)yn (1) we obtain (3.1) again.
Lemma 3.2 If λk is a multiple eigenvalue then

 (λk )
(yk+1 , yk ) = cyk+1 (1)yk (1) + yk (1) . (3.3)
2

Let us apply the procedure mentioned in the comments following the proof of (3.1)
to the functions yk+1 and yk . By integrating the identity
d 
yk+1 yk − yk+1 yk = yk2 ,
dx
from 0 to 1, we will obtain
 1
yk 22 = yk+1 yk − yk+1 yk 0
.

By using the boundary conditions for yk+1 and yk in the last equality we obtain (2.8)
again.
Lemma 3.3 If λk is a multiple eigenvalue then

2  (λk )  (λk )
yk+1 22 = cyk+1 (1) + 
yk+1 (1) + yk (1) . (3.4)
2 6

where 
yk+1 = yk+1 − C̃yk .
Proof Differentiating (3.2) with respect to μ we obtain

 (μ)  (μ)
(yλ (·, λ), yμ (·, μ)) = cyλ (1, λ)yμ (1, μ) − yλ (1, λ) − yλ (1, λ)
λ−μ (λ − μ)2
 (μ) 2 (μ)  (λ)
+ y (1, λ) + y (1, λ) + yμ (1, μ)
(λ − μ)2 (λ − μ)3 λ−μ
 (λ)  (λ) 2 (λ)
+ y (1, μ) − yμ (1, μ) − y (1, μ) .
(λ − μ) 2 (λ − μ) 2 (λ − μ)3
Letting μ → λk (λ = λk ) (cf. [1]) and then λ → λk we obtain (3.4).
40 Y. N. Aliyev

Lemma 3.4 If λk is a triple eigenvalue and λn = λk then

(yk+2 , yn ) = cyk+2 (1)yn (1). (3.5)

Proof Differentiating (3.2) with respect to λ, letting λ → λk (μ = λk ) we obtain

 (μ)
(ỹk+2 , y(·, μ)) =cỹk+2 (1)y (1, μ) − ỹk+2 (1)
λk − μ
 (μ)  (μ)
+ ỹk+1 (1) − yk (1) ,
(λk − μ)2 (λk − μ)3

from which (3.5) easily follows (cf. [1]).

Again, the same result can be achieved if we started with the identity

d 
yk+2 yn − yk+2 yn = (λk − λn )yk+2 yn + yk+1 yn ,
dx
which can be again easily derived using the definition of yk+2 . By integrating this
equality from 0 to 1 we will obtain
 1
(λk − λn )(yk+2 , yn ) + (yk+1 , yn ) = yk+2 yn − yk+2 yn 0
.

By using the boundary conditions for yk+2 and yn , and (3.1) we obtain a new proof
for (3.5).
Lemma 3.5 If λk is a triple eigenvalue then

 (λk )
(yk+2 , yk ) = cyk+2 (1)yk (1) + yk (1) . (3.6)
6
Again, by applying the above mentioned procedure to the functions yk+2 and yk
we obtain
d 
yk+2 yk − yk+2 yk = yk+1 yk .
dx
By integrating this equality from 0 to 1 we will obtain
 1
(yk+1 , yk ) = yk+2 yk − yk+2 yk 0
.

By using the boundary conditions for yk+2 and yk , we obtain

(yk+1 , yk ) = cyk+1 (1)yk (1),

Minimality Properties of Sturm-Liouville Problems 41

which is in perfect agreement with (3.3), because in triple eigenvalue case

 (λk ) = 0.
Lemma 3.6 If λk is a triple eigenvalue then

 (λk )  I V (λk )
(yk+2 , yk+1 ) = cyk+2 (1)yk+1 (1) + 
yk+1 (1) + yk (1) .
6 24
By applying the above mentioned procedure one more time but now to the functions
yk+2 and yk+1 we obtain

d  2
yk+2 yk+1 − yk+2 yk+1 = yk+1 − yk+2 yk .
dx
Integration of this equality from 0 to 1 will give us
 1
yk+1 22 − (yk+2 , yk ) = yk+2 yk+1 − yk+2 yk+1 0
.

By using the boundary conditions for yk+2 and yk+1 , we obtain

yk+1 22 − (yk+2 , yk ) = cyk+1 (1)2 − cyk+2 (1)yk (1),

which is again in perfect agreement with (3.4) and (3.6), because  (λk ) = 0.
Lemma 3.7 If λk is a triple eigenvalue then

2  (λk )
yk+2 22 =cyk+2 (1) + 
yk+2 (1)
6
 I V (λk )  V (λk )
+
yk+1 (1) + yk (1) ,
24 120

where 
yk+2 = yk+2 − C̃
yk+1 − D̃yk .

4 Existence of Auxiliary Associated Functions

By comparing the equalities in Sects. 2 and 3 we can see that there are some
fundamental differences between the formulae for the inner products and norms of
the eigenfunctions and the corresponding formulae for the associated functions. In
this section we will prove that it is possible to find special associated functions (we
call them auxiliary associated functions) whose properties in inner products make
them more close to the eigenfunctions than the other associated functions. In the
last section, these functions will play a crucial role in our description of minimality
properties.
42 Y. N. Aliyev

Lemma 4.1 If λk is a double eigenvalue then there exists an associated function

∗
yk+1 = yk+1 + C1 yk , where

yk (1) (λk ) + 3
yk+1 (1) (λk )
C1 = − ,
3yk (1) (λk )

such that
∗ ∗
(yk+1 , yk+1 ) = c(yk+1 ) (1)yk+1 (1). (4.1)

∗ ) (1) = 0 if and only if

Here, it should be pointed out that (yk+1

 (λk ) = 3C̃ (λk ).

Before proceeding, we also note that for λn = λk ,

∗ ∗
(yk+1 , yn ) = c(yk+1 ) (1)yn (1), (4.2)

∗ ∗  (λk )
(yk+1 , yk ) = c(yk+1 ) (1)yk (1) + yk (1) .
2
We shall now concentrate on the triple eigenvalue case. Although we will not
∗
need the function yk+1 in the triple eigenvalue case, it is still worthwhile to note that
such a function does not exist in this case. Instead, we will need other associated
#
functions of yk which will be denoted by yk+1 # .
and yk+2
Lemma 4.2 If λk is a triple eigenvalue then there exists an associated function
#
yk+1 = yk+1 + C2 yk , where

yk (1) I V (λk ) + 4
yk+1 (1) (λk )
C2 = − ,
4yk (1) (λk )

for which
#
(yk+1 , yk+2 ) = c(yk+1
#
) (1)yk+2 (1).

# ) (1) = 0 if and only if  I V (λ ) = 4C̃ (λ ).

It is worthwhile to note that (yk+1 k k
#
We now indicate some relations between yk+1 and the other root functions:

#
(yk+1 , yn ) = c(yk+1
#
) (1)yn (1), (n = k + 1, k + 2),

# #  (λk )
(yk+1 , yk+1 ) = c(yk+1 ) (1)yk+1 (1) + yk (1) .
6
Minimality Properties of Sturm-Liouville Problems 43

∗ , defined by y ∗
Note that the function yk+2 k+2 = yk+2 + C2 yk+1 , where C2 is the
same constant, also enjoys similar properties:
∗ ∗
(yk+2 , yk+1 ) = c(yk+2 ) (1)yk+1(1), (4.3)
∗ ∗
(yk+2 , yn ) = c(yk+2 ) (1)yn (1), (n = k, k + 1, k + 2), (4.4)

∗ ∗  (λk )
(yk+2 , yk ) = c(yk+2 ) (1)yk (1) + yk (1) . (4.5)
6
Lemma 4.3 If λk is a triple eigenvalue then there exists an associated function
#
yk+2 ∗
= yk+2 + D1 yk , where D1 is a constant, for which

#
(yk+2 , yk+1 ) = c(yk+2
#
) (1)yk+1(1), (4.6)
#
(yk+2 , yk+2 ) = c(yk+2
#
) (1)yk+2(1). (4.7)

Proof Note first that

∗ ∗
(yk+2 , yk+2 ) = c(yk+2 ) (1)yk+2(1) + Qk ,

where

 (λk )  I V (λk )  V (λk )

Qk =
yk+2 (1) +
yk+1 (1) + yk (1)
6 24 120
 I V 
 (λk )  (λk )
+ C2 
yk+1 (1) + yk (1) .
6 24

#
It is not difficult to check that for the function yk+2 ∗
= yk+2 + D1 yk , where

6Qk
D1 = − ,
yk (1) (λk )

both equalities (4.6) and (4.7) hold true.

# , equalities like (4.4), (4.5) are also true:
Note also that for the function yk+2

# #
(yk+2 , yn ) = c(yk+2 ) (1)yn (1), (n = k, k + 1, k + 2);

# #  (λk )
(yk+2 , yk ) = c(yk+2 ) (1)yk (1) + yk (1) .
6
# ) (1) = 0 if and only if
We remark that (yk+2
   
5 I V (λk )  I V (λk ) − 4C̃ (λk ) = 4 (λk )  V (λk ) − 20D̃ (λk ) .
44 Y. N. Aliyev

5 Minimality of the System of Root Functions

In this section we will consider all possible cases of the choice of the root function
which will be deleted from the system to obtain a minimal system. In each case we
will construct explicitly a biorthogonal system.

5.1 Case (a)

Theorem 5.1 If all the eigenvalues of (1.1)–(1.3) are real and simple then the
system

{yn } (n = 0, 1, . . . ; n = l), (5.1)

where l is a non-negative integer, is minimal in L2 (0, 1).

Proof It suffices to show the existence of a system

{un } (n = 0, 1, . . . ; n = l), (5.2)

biorthogonal to the system (5.1). Noting the relation Bn = 0 we define elements of

the system (5.2) by

1 yn (x) yn (1)
un (x) = . (5.3)
Bn yl (1) yl (x) yl (1)

It remains to see, noting (2.2), (2.7) and (2.10), that

(un , ym ) = δnm ,

where δnm (n, m = 0, 1, . . . ; n, m = l) denotes as usually, Kronecker’s symbol:

δnm = 0 if n = m and δnn = 1.

5.2 Case (b)

Theorem 5.2 If λk is a double eigenvalue then the system

{yn } (n = 0, 1, . . . ; n = k + 1),

is minimal in L2 (0, 1).

Minimality Properties of Sturm-Liouville Problems 45

Proof In this case the biorthogonal system is defined by

1 yn (x) yn (1)
un (x) = (n = k, k + 1), (5.4)
Bn yk (1) yk (x) yk (1)
1 yk+1 (x) yk+1 (1)
uk (x) = .
Bk+1 yk (1) yk (x) yk (1)
∗ ) (1) = 0 then the system
Theorem 5.3 If λk is a double eigenvalue, and if (yk+1

{yn } (n = 0, 1, . . . ; n = k), (5.5)

is minimal in L2 (0, 1).

Proof The elements of the biorthogonal system are defined as follows

1 yn (x) yn (1)
un (x) = ∗ ) (1) y ∗ (x) (y ∗ ) (1) (n = k, k + 1),
Bn (yk+1 k+1 k+1

1 yk (x) yk (1)
uk+1 (x) = ∗ ∗ (x) (y ∗ ) (1) .
y
Bk+1 (yk+1 ) (1) k+1 k+1

Before proceeding we comment on the condition (yk+1 ∗ ) (1) = 0 above. Let

∗ ∗
(yk+1 ) (1) = 0, then by (4.1), (4.2) the function yk+1 is orthogonal to all the
elements of the system (5.5). Therefore the system (5.5) is not complete in L2 (0, 1).
Theorem 5.4 If λk is a double eigenvalue then the system

{yn } (n = 0, 1, . . . ; n = l),

where l = k, k + 1 is a non-negative integer, is minimal in L2 (0, 1).

Proof The biorthogonal system is given by the formula (5.3) for n = k, k + 1, and

1 yk (x) yk (1)
uk+1 (x) = ,
Bk+1 yl (1) yl (x) yl (1)
1 ∗ (x) (y ∗ ) (1)
yk+1
uk (x) = k+1 .
Bk+1 yl (1) yl (x) yl (1)

5.3 Case (c)

Theorem 5.5 If λk is a triple eigenvalue then the system

{yn } (n = 0, 1, . . . ; n = k + 2),

is minimal in L2 (0, 1).

46 Y. N. Aliyev

Proof The biorthogonal system is given by the formula (5.4) for n = k, k+1, k+2,
and

1 yk+1 (x) yk+1 (1)

uk+1 (x) = ,
Bk+2 yk (1) yk (x) yk (1)
1 # (x) (y # ) (1)
yk+2
uk (x) = k+2 .
Bk+2 yk (1) yk (x) yk (1)
# ) (1) = 0 then the system
Theorem 5.6 If λk is a triple eigenvalue, and if (yk+1

{yn } (n = 0, 1, . . . ; n = k + 1), (5.6)

is minimal in L2 (0, 1).

Proof In this case the elements of the biorthogonal system are

1 yn (x) yn (1)
un (x) = # ) (1) y # (x) (y # ) (1)
(n = k, k + 1, k + 2),
Bn (yk+1 k+1 k+1

1 yk (x) yk (1)
uk+2 (x) = # # # ,
Bk+2 (yk+1 ) (1) yk+1 (x) (yk+1 ) (1)

1 # (x) (y # ) (1)
yk+2
uk (x) = #
k+2
# (x) (y # ) (1) .
Bk+2 (yk+1 ) (1) yk+1 k+1

In analogy with Theorem 5.3, we may show that if (yk+1 # ) (1) = 0 then the function
#
yk+1 (x) is orthogonal to all the elements of the system (5.6); hence the system (5.6)
is not complete.
# ) (1) = 0 then the system
Theorem 5.7 If λk is a triple eigenvalue, and if (yk+2

{yn } (n = 0, 1, . . . ; n = k), (5.7)

is minimal in L2 (0, 1).

Proof We define the elements of the biorthogonal system by

1 yn (x) yn (1)
un (x) = # ) # (x) (y # ) (1) (n = k, k + 1, k + 2),
Bn (yk+2 (1) yk+2 k+2

1 yk (x) yk (1)
uk+2 (x) = # # (x) (y # ) (1) ,
y
Bk+2 (yk+2 ) (1) k+2 k+2

1 yk+1 (x) yk+1 (1)

uk+1 (x) = # # (x) (y # ) (1) .
Bk+2 (yk+2 ) (1) yk+2 k+2

# ) (1) = 0 then the system (5.7) is not complete.

If (yk+2
Minimality Properties of Sturm-Liouville Problems 47

Theorem 5.8 If λk is a triple eigenvalue then the system

{yn } (n = 0, 1, . . . ; n = l),

where l = k, k + 1, k + 2 is a non-negative integer, is minimal in L2 (0, 1).

Proof The elements of the biorthogonal system can be defined by (5.3) for n =
k, k + 1, k + 2, l and

1 yk (x) yk (1)
uk+2 (x) = ,
Bk+2 yl (1) yl (x) yl (1)
1 # (x) (y # ) (1)
yk+1
uk+1 (x) = k+1 ,
Bk+2 yl (1) yl (x) yl (1)
1 # (x) (y # ) (1)
yk+2
uk (x) = k+2 .
Bk+2 yl (1) yl (x) yl (1)

5.4 Case (d)

Theorem 5.9 If λr and λs = λr are a conjugate pair of non-real eigenvalues then

each of the systems

{yn } (n = 0, 1, . . . ; n = r), (5.8)

{yn } (n = 0, 1, . . . ; n = l), (5.9)

where l = r, s is a non-negative integer, is minimal in L2 (0, 1).

Proof The biorthogonal system of (5.8) is as follows

1 yn (x) yn (1)
un (x) = (n = r, s),
Bn ys (1) ys (x) ys (1)
1 yr (x) yr (1)
us (x) = .
yr (1)ys (1) (λr ) ys (x) ys (1)

The biorthogonal system of (5.9) is defined by (5.3) for n = r, s, l and

1 ys (x) ys (1)
ur (x) = ,
ys (1)yl (1) (λs ) yl (x) yl (1)
1 yr (x) yr (1)
us (x) = .
yr (1)yl (1) (λr ) yl (x) yl (1)
48 Y. N. Aliyev

Remark 5.10 Using the method of the paper [5] these minimality results can also
be extended to basis properties. Then the sufficient conditions
∗ # #
(yk+1 ) (1) = 0, (yk+1 ) (1) = 0, (yk+2 ) (1) = 0

in Theorems 5.3, 5.6 and 5.7, respectively, will be necessary conditions, too.

6 Example

As an illustration of the above theory, we present a particular result for the following
problem
 
λ
−y = λy, 0 < x < 1, y(0) = 0, y(1) = + 1 y (1).
3

For this problem λ0 = λ1 = 0 is a double eigenvalue.

√ √The other eigenvalues
 λ2 <
λ3 < . . . are the solutions of the equation tan λ = λ λ3 + 1 . Eigenfunctions
√
are y0 = x, yn = sin λn x (n ≥ 2) and an associated function corresponding to
y0 is y1 = − 16 x 3 + Cx, where C is an arbitrary constant. We look for an auxiliary
associated function of the form y1∗ = − 16 x 3 + C x. In other words, C1 = C − C.
By (4.1),
 11

1 3

1

1

1

− x 3 + Cx − x + C x dx = − +C − +C .
0 6 6 3 2 2

From this equality we obtain that C = −C + 25

42 , so
 
1 25
(y1∗ ) (x) = − x 3 + −C + x.
6 42

Consequently, (y1∗ ) (1) = 21

2
− C. Therefore the above condition (y1∗ ) (1) = 0 in
Theorem 5.3 is equivalent to C = 21
2
. So, the system
   
1
{yn } (n = 1, 2, . . .) = sin λn x (n = 2, 3, . . .) ∪ − x 3 + Cx ,
6

from which only the eigenfunction y0 is excluded but the associated function y1 is
included, is minimal in L2 (0, 1) if C = 21
2
.
We shall now apply√a different method to this problem. Note that for this
√
problem y(x, λ) = sin√ λx . We need λ in the denominator to make sure that
λ
Minimality Properties of Sturm-Liouville Problems 49

limλ→0 y(x, λ) = y0 = x. Then

√ √
x cos λx sin λx
yλ (x, λ) = − √
2λ 2λ λ
3
and therefore ỹ1 = limλ→0 yλ (x, λ) = − x6 . Let y1 = − 16 x 3 + C. Then C̃ = −C.
√ √ √
Note also that  (λ) = sin λ − λ3 + 1 λ cos λ, then

 (0) = lim  (λ) = 0,  (0) = lim  (λ) = 0,

λ→0 λ→0
 (0) = lim  (λ) = 4/15,  (0) = lim  (λ) = −8/105.
λ→0 λ→0

As was pointed out in the comments following the proof of Lemma 4.1, the
condition (y1∗ ) (1) = 0 is equivalent to  (λk ) = 3C̃ (λk ). Since C = −C̃,
we obtain, once again, C = 21 2
.
3
If C = 21 2
then we obtain y1∗ (x) = − x6 + x
2 which is orthogonal to all the
elements of the system {yn } (n = 1, 2, . . .).

Acknowledgments Many thanks to ADA University and especially School of IT and Engineering
for their constant support.

References

1. Y.N. Aliyev, Minimality of the system of root functions of Sturm-Liouville problems with
decreasing affine boundary conditions. Colloq. Math. 109, 147–162 (2007)
2. P.A. Binding, P.J. Browne, Application of two parameter eigencurves to Sturm-Liouville
problems with eigenparameter-dependent boundary conditions. Proc. R. Soc. Edinb. 125A,
1205–1218 (1995)
3. P.A. Binding, P.J. Browne, B.A. Watson, Equivalence of inverse Sturm-Liouville problems with
boundary conditions rationally dependent on the eigenparameter. J. Math. Anal. Appl. 291, 246–
261 (2004)
4. A.P. Kartashev, B.L. Rojdestvenskiy, Ordinary Differential Equations and Foundations of
Calculus of Variations (Nauka, Moscow, 1980, in Russian)
5. N.B. Kerimov, Y.N. Aliyev, The basis property in Lp of the boundary value problem rationally
dependent on the eigenparameter. Stud. Math. 174, 201–212 (2006)
6. N.B. Kerimov, V.S. Mirzoev, On the basis properties of one spectral problem with a spectral
parameter in boundary conditions. Sib. Math. J. 44, 813–816 (2003)
7. M.A. Naimark, Linear Differential Operators, 2nd edn. (Nauka, Moscow, 1969, in Russian).
English trans. of 1st edn., Parts I, II (Ungar, New York, 1967, 1968)
8. A.A. Shkalikov, Boundary value problems for ordinary differential equations with a parameter
in the boundary conditions. Tr. Semin. Im. I. G. Petrovskogo 9, 190–229 (1983, in Russian)
9. A.A. Shkalikov, Basis properties of root functions of differential operators with spectral
parameter in the boundary conditions. Differ. Equ. 55, 631–643 (2019)
Scattering, Spectrum and Resonance
States Completeness for a Quantum
Graph with Rashba Hamiltonian

Irina V. Blinova, Igor Y. Popov, and Maria O. Smolkina

Abstract Quantum graphs consisting of a ring with two semi-infinite edges

attached to the same point of the ring is considered. We deal with the Rashba
spin-orbit Hamiltonian on the graph. A theorem concerning to completeness of
the resonance states on the ring is proved. Due to use of a functional model, the
problem reduces to factorization of the characteristic matrix-function. The result is
compared with the corresponding completeness theorem for the Schrödinger, Dirac
and Landau quantum graphs.

Keywords Spectrum · Resonance · Completeness · Spin-orbit interaction

Mathematics Subject Classification (2010) Primary 81U20; Secondary 46N50

1 Introduction

The problem of resonances was studied during a long time starting with the
famous Rayleigh work on the Helmholtz resonator. The completeness problem for
resonance states is younger. Mathematicians dealt with it during a half of century.
The problem is related to the stability of the completeness under some perturbation.
Namely, one starts with a system with purely discrete spectrum and complete system
of eigenstates, e.g., a closed resonator. Then one perturbs the system in such a way
that eigenvalues turn to resonances and eigenstates—to resonance states, e.g., the
resonator with a boundary window [2, 7, 8, 10]. Correspondingly, a natural question
appears: is the system of quasi-eigenstates complete? One of the approaches to this
problem is related to the Sz.-Nagy functional model [14, 21, 25]. Starting with work
[1], it is known that the scattering matrix is the same as the characteristic function
from the functional model. In particular, root vectors in the functional model

I. V. Blinova · I. Y. Popov () · M. O. Smolkina

ITMO University, St. Petersburg, Russia
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 51

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_4
52 I. V. Blinova et al.

correspond to resonance states in scattering theory. The completeness problem for

the system of root vectors is related to the factorization problem for the characteristic
function as a function from the Hardy space. We will use this relation to study the
completeness. Particularly, for the finite-dimensional case, this approach gives one
an effective completeness criterion [21].
The simplest model for an open resonator is based on a quantum graph, well-
developed model in quantum theory [3, 9, 18]. The corresponding completeness
problem was considered in [4, 11, 23, 24].
Spin-orbit interaction attracts great attention last time due to possible perspec-
tives of applications in nanoelectronics, particularly, in spintronics and quantum
computing [5, 6, 13, 15, 16, 19, 20]. Full spin-orbit Hamiltonian is very complicated
even for quantum graph. Usually, theoreticians consider its particular cases, Rashba
or Dresselhaus Hamiltonians. In this paper we deal with the Rashba Hamiltonian.
We investigate this operator on a quantum graph with loop imbedded in a plane in
R3 . The graph is posed in a constant magnetic field orthogonal to the graph plane.
We choose such type of the graph (a loop touching a line) because the Schrödinger,
Dirac and Landau operators on such graph lead to incompleteness of the resonance
states [12]. In some sense, this graph presents a degenerate case and any small
perturbation, e.g., point-like potential at the vertex, restores the completeness. It
is interesting that the Rashba Hamiltonian leads to the completeness.

1.1 Scattering, Functional Model and Completeness Criterion

It is convenient to consider resonances in the framework of the Lax-Phillips scat-

tering theory [17]. It is based on non-stationary problem. Let H be the Hamiltonian
for the quantum graph, i.e. the Rashba Hamiltonian on the ring and the Schrödinger
operator on the half-axes. The wave function is a two-component vector. Different
components correspond to different spin directions. Consider the Cauchy problem
for the time-dependent problem on the graph :

i h̄ut = H u,
u(x, 0) = u0 (x), x ∈ .

Below, we will use the system of units in which h̄ = 1. The standard Lax-
Phillips approach is applied to the wave (acoustic) equation. There is a close relation
between the Schrödinger and wave cases. We will describe it briefly following [17,
Section 6.4]. Namely, it is necessary to consider the operator A2 (here A is the
generator of the evolution group U (t) for the wave equation, see below):
 
H 0
A =
2
.
0 H
Resonance States Completeness 53

One can see that A2 acts as the Hamiltonian H on each component of the data
of the acoustic problem. This allows one to use the “acoustic” construction for the
Schrödinger case and, as a result, comes to the relation between the Schrödinger
(S Schr ) and the acoustic (S) scattering matrices:
√
S Schr (z) = S( z).

Let us describe briefly the acoustic case. Consider the Cauchy problem for the wave
equation

ut t = uxx ,
(1.1)
u(x, 0) = u0 (x), ut (x, 0) = u1 (x), x ∈ .

Here u, u0 , u1 are two-component vectors. Let E be the Hilbert space of four-

component functions (u0 , u1 ) on the graph with finite energy

−1
(u0 , u1 )E = 2
2
(|u0 |2 + |u1 |2 )dx.


The pair (u0 , u1 ) is called the Cauchy data. The unitary (in E) group U (t),
U (t)(u0 , u1 ) = (u(x, t), ut (x, t)), solves the problem (1.1). The unitary group
U (t)|t ∈R has two orthogonal (in E) subspaces, D− and D+ , called, correspondingly,
incoming and outgoing subspaces.
Definition 1.1 The outgoing (incoming) subspace D+ (D− ) is a subspace of E
having the following properties:
(a) U (t)D+ ⊂ D+ for t > 0; U (t)D− ⊂ D− for t < 0,
(b) ∩t >0 U (t)D+ = {0}; ∩t <0 U (t)D− = {0}
(c) ∪t <0 U (t)D+ = E, ∪t >0 U (t)D− = E.
The existence of the incoming and outgoing subspaces is related to the spectral
properties of the operator H . The property (c) from the definition is fulfilled if the
spectrum of the operator is continuous [17]. The operator H is self-adjoint. Using
the spectral expansion, one can obtain the continuous subspace Ec by removing the
discrete subspace (i.e. eigenspaces) from E. Below we will deal with H |Ec (we will
not introduce a new notation for this operator with the continuous spectrum). The
choice of incoming and outgoing subspaces is not unique. For the graph , one
can choose the subspace D+ containing four-component functions vanishing at the
ring 0 and satisfying the radiation condition on all leads (infinite edges). Due to
the radiating condition, one has only outgoing exponential at the edges. It gives
one properties (a), (b) for the subspace. As for (c), it takes place for Ec due to the
self-adjointness of the operator H :

∪t <0 U (t)D+ = Ec .
54 I. V. Blinova et al.

The definition of the subspace D− is analogous. Property (c) for D− takes the form

∪t >0 U (t)D− = Ec .

Let P± be the orthogonal projection of E onto the orthogonal complement of D± .

Consider the semigroup {Z(t)}|t ≥0 of operators on E defined by

Z(t) = P+ U (t)P− , t ≥ 0.

The square roots of the eigenvalues of the generator B of Z(t) are resonances. The
following subspace K = E  (D− ⊕ D+ ) is very important for the construction.
The operators {Z(t)}|t ≥0 map the subspace K into itself. Lax and Phillips proved
the following theorem [17].
Theorem 1.2 There is a pair of isometric maps T± : E → L2 (R, N) (the
outgoing and incoming spectral representations), N is an auxiliary space, having
the following properties:

T± U (t) = eikt T± , T± D± = H±2 (N), T− D+ = SH+2 (N),

where H±2 (N) is the Hardy space in the upper (lower) half-plane, the matrix-
function S is an inner function in C+ , and

K− = T− K = H+2  SH+2 , T− Z(t)|K = PK− eikt T− |K− .

Remark 1.3 S is known in functional analysis as characteristic function. At the

same time, it is the well-known physical object—the S-matrix.
We deal with the completeness of the system of resonance states (i.e. root vectors
of B). There is an interesting relation between the completeness problem and the
factorization of the scattering matrix. Namely, as an inner operator-function, S can
be represented in the form S = , where  is a Blaschke-Potapov product and 
is a singular inner function [14, 21, 25]. The next theorem shows this relation.
Theorem 1.4 (Completeness Criterion from [21]) The following statements are
equivalent:
1. The system of root vectors of the operator B is complete;
2. The system of root vectors of the operator B∗ is complete;
3. S is a Blaschke-Potapov product.
There is a simple criterion for the absence of the singular inner factor in the case
dim N < ∞ (in the general operator case there is no such simple criterion).
Resonance States Completeness 55

Theorem 1.5 ([21]) Let dim N < ∞. The following statements are equivalent:
1. S is a Blaschke-Potapov product;
2.

2i
lim ln |det S(k)| dk = 0, (1.2)
r→1 (k + i)2
Cr

where Cr is the image of |ζ | = r under the inverse Cayley transform.

Remark 1.6 The theorem in [21] is formulated for the unit disk. We use the Cayley
transform which maps the upper half-plane to the unit disk:
z−i
W (z) = ,
z+i
whereas the inverse Cayley transform maps the unit disk to the upper half-plane:
1+ζ
w(ζ ) = i .
1−ζ
One notable property of the Cayley transform is that it injectively maps R into
the unit circle. Another important property we are going to use is that the Cayley
transform preserves circles.
The integration curve can be parameterized as

Cr = {R(r)eit + iC(r) | t ∈ [0, 2π)}

where
1 + r2 2r
C(r) = , R(r) = .
1 − r2 1 − r2

It should be noted that R → ∞ corresponds to r → 1.

For brevity, we define

s(k) = |det S(k)| ,

and after throwing away constants which are irrelevant for convergence, we obtain
the final form of the criterion (1.2), which is convenient for us and will be used
afterwards:

2π
R(r) ln(s(R(r)eit + iC(r))) it
lim e dt = 0. (1.3)
r→1 (R(r)eit + iC(r) + i)2
0

Our main theorem is as follows.

56 I. V. Blinova et al.

Theorem 1.7 The system of resonance states is complete on L2 (0 ) ⊕ L2 (0 ).

Remark 1.8 We recall that two components correspond to two directions of spin
in respect to magnetic field. It is interesting to compare the result with the
corresponding theorems for other operators. If one deals with the same geometric
graph with the Shrödinger operator, the Dirac operator or the Landau operator (the
Schrödinger operator with a magnetic field) and the Kirchhoff coupling condition
at the vertex then there is an incompleteness. But small perturbation of the coupling
condition (delta-potential at the vertex) restores the completeness. In our case the
spin-orbit interaction plays a role of such perturbation.

2 Scattering for Rashba Hamiltonian

2.1 Rashba Operator

We consider a quantum graph shown in Fig. 1. The quantum ring of radius

a is located in the plane xy in the presence of Rashba spin-orbit interaction
and a perpendicular magnetic field B. We assume that the magnetic field B is
relatively weak. In such case the interaction between the electron spin and the field
(Zeeman term) can be treated as a perturbation and the corresponding dimensionless
Hamiltonian reads
  
∂ ωSO  2 ωSO 2
ωL
Ĥ = −i + σr − − + σz ,
∂ϕ 2 0 42 

where ϕ is the azimuthal angle of a point on the ring,  denotes the magnetic
flux encircled by the ring, 0 = h̄e , e is the electron charge, ωSO = h̄aα
is the
frequency associated with the spin-orbit interaction (with interaction constant α),

Fig. 1 Task 1 (upper row), Task 2 (lower row); directions of edges are shown
Resonance States Completeness 57

h̄2
h̄ = , m∗ is the effective mass of the electron. The radial spin operator is
2m∗ a 2
given by
 
0 e−iϕ
σr = σx cos ϕ + σy sin ϕ = iϕ .
e 0
∗
The constant in Zeeman term is as follows ωL = g4m eB
with g ∗ and m being the
effective gyromagnetic ratio and the free electron mass, respectively.

2.2 Solution on the Line

To construct the scattering matrix for the graph  we solve two scattering problems
differing in solutions on the lines.
Task 1:
     
1 ikx R↑↑ −ikx T↑↑ ikx
ψ e1 ↑ (x) = e + e , ψ e2 ↑ (x) = e .
0 R↑↓ T↑↓

Task 2:
     
e1 ↓ 0 ikx R↓↑ −ikx e2 ↓ T↓↑ ikx
ψ (x) = e + e , ψ (x) = e .
1 R↓↓ T↓↓

2.3 Solution on the Ring

For the both tasks one has the following form of the solution on the ring:
  μ μ
ψ(ϕ) = aj eKj ϕ χ (μ) (ϕ),
μ=1,2 j =1,2

where
 −i ϕ  ϕ
1 e 2 cos θ2 1 e−i 2 sin θ2
χ (1)(ϕ) = √ ϕ
, χ (2) (ϕ) = √ ϕ
.
2π ei 2 sin θ2 2π −ei 2 cos θ2

Here θ = arctan(−α),
! 2 "1/2  1/2
2m∗ E
μ
1 AC  α E
+ (−1)μ+j +1
μ
Kj = + + + , k= .
2 2π 2π 4 h̄ h̄2
58 I. V. Blinova et al.

The value of j = 1 and j = 2 corresponds to the clockwise and counterclockwise

motions of an electron through the ring, respectively. Finally,

μ
AC = −π[1 + (−1)μ 1 + α 2 ]

is the so-called Aharonov-Casher (AC) phase.

2.4 Vertex Coupling Conditions

We assume the standard coupling condition at vertex v for the spin-orbit Hamilto-
nian: the continuity of the wave functions and vanishing net spin current densities
[22]. For our graph (see Fig. 1) it has the following form:

(−1)[ej ] Dj ψj (v) = 0, ψ e1 (v) = ψ e2 (v) = · · · = ψ ej (v),
ej

where summation is over all edges ej adjacent to the vertex v, [ej ] = 0 for the
outgoing edge and [ej ] = 1 for the incoming edge. The third edge (see Fig. 1) gives
one two terms (outgoing and incoming) with
    
∂ α  0 e−iϕ
Dj = −i σr − , σ r = iϕ
∂ϕ 2 0 e 0

(in our case for the third edge, one has φ = 0 for the outgoing term and φ = 2π for
the incoming term). For straight semi-infinite edges (the first and the second edges
in Fig. 1) Dj = dxd
.
Remark 2.1 From mathematical point of view, these conditions ensure the self-
adjointness of the operator on the graph. In physical literature the conditions are
known as Griffith conditions (see, e.g., [13, 20]).
These conditions take the following form for our graph.
For Task 1:
e ↑ e ↑ e e
ψ1 1 (0) = ψ1 2 (0) = ψ1 3 (0) = ψ1 3 (2πa),

e ↑ e ↑ e e
ψ2 1 (0) = ψ2 2 (0) = ψ2 3 (0) = ψ2 3 (2πa),
     
∂ψ(0) ∂ψ(2πa) 1 T↑↑ R↑↑
−βψ(0)− +βψ(2πa)−ik +ik +ik = 0,
∂ϕ ∂ϕ 0 T↑↓ R↑↓
Resonance States Completeness 59

where
 
α 
β=i σr − ,
2 0

ψ1 and ψ2 denote, correspondingly, the first and the second component of the
vector ψ, 2πa is the length of the ring.
For Task 2:
e ↓ e ↓
ψ1 1 (0) = ψ1 2 (0) = ψ1e3 (0) = ψ1e3 (2πa),

e ↓ e ↓
ψ2 1 (0) = ψ2 2 (0) = ψ2e3 (0) = ψ2e3 (2πa),
     
∂ψ(0) ∂ψ(2πa) T 0 R↓↑
−βψ(0)− +βψ(2πa)+ik ↓↑ −ik +ik = 0.
∂ϕ ∂ϕ T↓↓ 1 R↓↓

2.5 S-Matrix

Solutions of Task 1 and Task 2 give us the entries of the S-matrix (see [13]):
⎛ ⎞
R↑↑ R↓↑ T↑↑ T↓↑
⎜R↑↓ R↓↓ T↑↓ T↓↓ ⎟
S(k) = ⎜
⎝ T↑↑
⎟
T↓↑ R↑↑ R↓↑ ⎠
T↑↓ T↓↓ R↑↓ R↓↓

where
θ θ
R↑↑ = (1) cos2 + (2) sin2 − 1,
2 2
θ θ
R↑↓ = ((1) − (2) ) sin cos , (2.1)
2 2
R↓↑ = R↑↓ ,
θ θ
R↓↓ = (1) sin2 + (2) cos2 − 1
2 2
and
θ θ
T↑↑ = τ (1) cos2 + τ (2) sin2 ,
2 2
θ θ
T↑↓ = (τ (1) − τ (2)) sin cos , (2.2)
2 2
T↓↑ = T↑↓ ,
θ θ
T↓↓ = τ (1) sin2 + τ (2) cos2 ,
2 2
60 I. V. Blinova et al.

where

4k 2 a 2 (μ) 4ikaq (μ)

(μ) = iq sin(2q (μ)π), τ (μ) = sin(2πq (μ) ),
ŷ (μ) ŷ (μ)

  

ŷ (μ) = 4ikaq (μ) sin(2q (μ)π) − 4(q (μ))2 cos (−1)μ+1 ω − 2 π
0
− 4(q (μ))2 cos(2q (μ)π),
) *
ωL 1 ω 2 E h̄2 k 2
SO
q (μ) = q 2 + (−1)μ , q= + , E= ,
 ω 2 h̄ 2m∗
*
2
ωSO g ∗ eB α
ω= 1+ ωL = , ωSO = .
2 4m h̄a

2.6 Completeness

The completeness criterion (see above) is related to the absence or presence of

an exponential in k factor in det S(k). Such a factor gives one a linear growth of
ln det S(k) for k → ∞ (i.e. r → 1) which leads to the destroying of integral
convergence to zero in (1.3), i.e. to incompleteness of resonance states. One more
possible reason for destroying of such convergence is the possibility for singularities
of ln det S(k) to be placed at the integration curve. These singularities are roots and
singularities of det S(k). Taking into account the obtained expressions for the entries
of the matrix S(k), one concludes that det S(k) is a ratio of two analytic functions.
Correspondingly, the roots have no accumulation points on the complex plane. It
means that there are not more than a finite number of singularities of ln det S(k) at
the integration curve. Moreover, these singularities are logarithmic, i.e. integrable.
Hence, these singularities do not destroy the convergence (1.3). To reveal whether
there is an exponential factor in det S(k), one can look after the behaviour of det S(k)
on the imaginary axis. If there is an exponential factor, one has an exponential
growth at one of the imaginary half-axes. One can check that all entries of the S-
matrix (see the explicit expressions (2.1) and (2.2)) have no exponential growth.
It means that one has the convergence (1.3) in the completeness criterion and,
correspondingly, the system of resonance states is complete. This finalizes the proof
of Theorem 1.7.
Resonance States Completeness 61

–5

–10

–15

–20

–25
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Fig. 2 Behaviour of the integrand of (1.3)

2.7 Numerical Results

Having explicit expression for the S-matrix, one can simply demonstrate the
behaviour of the integrand of (1.3) at infinity. It is shown in Fig. 2 for the main
part of the integrand:

ln(|det (S(k))|)
F (k) = .
|k|

Acknowledgments This work was partially financially supported by the Government of the
Russian Federation (grant 08-08), grant 16-11-10330 of Russian Science Foundation and grant
19-31-90164 of Russian Foundation for Basic Researches.

References

1. V.M. Adamyan, D.Z. Arov, On a class of scattering operators and characteristic operator-
functions of contractions. Dokl. Akad. Nauk SSSR 160, 9–12 (1965) (in Russian)
2. A. Aslanyan, L. Parnovski, D. Vassiliev, Complex resonances in acoustic waveguides. Q. J.
Mech. Appl. Math. 53, 429–447 (2000)
3. G. Berkolaiko, P. Kuchment, Introduction to Quantum Graphs (AMS, Providence, 2012)
62 I. V. Blinova et al.

4. I.V. Blinova, I.Y. Popov, Quantum graph with the Dirac operator and resonance states
completeness. Oper. Theor. Adv. Appl. 268, 111–124 (2018)
5. A. Chatterjee, I.Y. Popov, M.O. Smolkina, Persistent current in a chain of two Holstein–
Hubbard rings in the presence of Rashba spin-orbit interaction. Nanosyst. Phys. Chem. Math.
10, 50–62 (2019)
6. E. Dehghana, D.S. Khoshnoud, A.S. Naeimi, Logical spin-filtering in a triangular network of
quantum nanorings with a Rashba spin-orbit interaction. Physica B 529, 21–26 (2018)
7. P. Duclos, P. Exner, B. Meller, Open quantum dots: resonances from perturbed symmetry and
bound states in strong magnetic fields. Rep. Math. Phys. 47, 253–267 (2001)
8. J. Edward, On the resonances of the Laplacian on waveguides. J. Math. Anal. Appl. 272, 89–
116 (2002)
9. P. Exner, P. Keating, P. Kuchment, T. Sunada, A. Teplyaev (eds.), Analysis on Graphs and Its
Applications (AMS, Providence, 2008)
10. P. Exner, V. Lotoreichik, M. Tater, On resonances and bound states of Smilansky Hamiltonian.
Nanosyst. Phys. Chem. Math. 7, 789–802 (2016)
11. D.A. Gerasimov, I.Y. Popov, Completeness of resonance states for quantum graph with two
semi-infinite edges. Complex Var. Elliptic Equ. 63, 996–1010 (2018)
12. D. Gerasimov, I. Popov, I. Blinova, A. Popov, Incompleteness of resonance states for quantum
ring with two semi-infinite edges. Anal. Math. Phys. 9, 1287–1302 (2019)
13. O. Kálmán, P. Földi, M.G. Benedict, F.M. Peeters, Magnetoconductance of rectangular arrays
of quantum rings. Phys. Rev. B 78, 125306 (2008)
14. S.V. Khrushchev, N.K. Nikol’skii, B.S. Pavlov, Unconditional bases of exponentials and of
reproducing kernels, in Complex Analysis and Spectral Theory (Leningrad, 1979/1980), vol.
864. Lecture Notes in Mathematics (Springer, Berlin, 1981), pp. 214–335
15. V.K. Kozin, I.V. Iorsh, O.V. Kibis, I.A. Shelykh, Quantum ring with the Rashba spin-orbit
interaction in the regime of strong light-matter coupling. Phys. Rev. B 97, 155434 (2018)
16. V.V. Kudryashov, A.V. Baran, Rashba spin-orbit interaction in a circular quantum ring in the
presence of a magnetic field. Nonlinear Phenom. Complex Syst. Minsk 14(1), 89–95 (2011)
17. P.D. Lax, R.S. Phillips, Scattering Theory (Academic Press, New York, 1967)
18. J. Lipovsky, Quantum graphs and their resonance properties. Acta Physica Slovaca 66, 265–
363 (2016)
19. A.S. Naeimi, L. Eslami, M. Esmaeilzadeh, A wide range of energy spin-filtering in a Rashba
quantum ring using S-matrix method. J. Appl. Phys. 113, 044316 (2013)
20. A.S. Naeimi, L. Eslami, M. Esmaeilzadeh, M.R. Abolhassani, Spin transport properties in a
double quantum ring with Rashba spin-orbit interaction. J. Appl. Phys. 113, 014303 (2013)
21. N. Nikol’skii, Treatise on the Shift Operator: Spectral Function Theory (Springer, Berlin, 1986)
22. K. Pankrashkin, Localization effects in a periodic quantum graph with magnetic field and spin-
orbit interaction. J. Math. Phys. 47, 112105 (2006)
23. I.Y. Popov, A.I. Popov, Quantum dot with attached wires: Resonant states completeness. Rep.
Math. Phys. 80, 1–10 (2017)
24. I.Y. Popov, A.I. Popov, Line with attached segment as a model of Helmholtz resonator: resonant
states completeness. J. King Saud Univ. Sci. 29, 133–136 (2017)
25. B. Sz-Nagy, C. Foias, H. Bercovici, L. Kerchy, Harmonic Analysis of Operators on Hilbert
Space, 2nd edn. (Springer, Berlin, 2010)
Tau Functions Associated with Linear
Systems

Gordon Blower and Samantha L. Newsham

Abstract Let (−A, B, C) be a linear system in continuous time t > 0 with input
and output space C and state space H . The function φ(x) (t) = Ce−(t +2x)AB
determines a Hankel integral operator φ(x) on L2 ((0, ∞); C); if φ(x) is trace class,
then the Fredholm determinant τ (x) = det(I + φ(x) ) defines the tau function
of (−A, B, C). Such tau functions arise in Tracy and Widom’s theory of matrix
models, where they describe the fundamental probability distributions of random
matrix theory. Dyson considered such tau functions in the inverse spectral problem
for Schrödinger’s equation −f + uf = λf , and derived the formula for the
d2
potential u(x) = −2 dx 2 log τ (x) in the self-adjoint scattering case (Commun Math
Phys 47:171–183, 1976). This paper introduces a operator function Rx that satisfies
Lyapunov’s equation dR dx = −ARx − Rx A and τ (x) = det(I + Rx ), without
x

assumptions of self-adjointness. When −A is sectorial, and B, C are Hilbert–

Schmidt, there exists a non-commutative differential ring A of operators in H and
a differential ring homomorphism ! " : A → C[u, u , . . . ] such that u = −4!A",
which extends the multiplication rules for Hankel operators considered by Pöppe
and McKean (Cent Eur J Math 9:205–243, 2011).

Keywords Integrable systems · Fredholm determinant · Inverse scattering

Mathematics Subject Classification (2010) Primary 47B3; Secondary 5, 34B27

1 Introduction

This paper is concerned with Fredholm determinants which arise in the theory of
linear systems and their application to inverse spectral problem for Schrödinger’s
equation. For φ ∈ L2 ((0, ∞); R), the Hankel integral operator corresponding to φ

G. Blower () · S. L. Newsham

Department of Mathematics and Statistics, Lancaster University, Lancaster, UK
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 63

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_5
64 G. Blower and S. L. Newsham

is φ where
 ∞
φ f (x) = φ(x + y)f (y) dy (f ∈ L2 ((0, ∞); C).
0

Using the Laguerre system of orthogonal functions as in [30], one can express φ as
a matrix [γj +k ]∞ 2
j,k=1 on  , which has the characteristic shape of a Hankel matrix,
and one can establish criteria for the operator to be bounded on L2 ((0, ∞); C).
Megretskii et al. [25] determined the possible spectrum and spectral multiplicity
function that can arise from a bounded and self-adjoint Hankel operator. Thus
they characterized the class of bounded self-adjoint Hankel operators up to unitary
equivalence. Their method involved introducing suitable linear systems on a state
space H , and this motivated the approach of our paper.
Following earlier works by Faddeev and others in the Russian literature, Dyson
[8] considered the inverse spectral problem for Schrödinger’s equation −f + uf =
λf , for u ∈ C 2 (R; R) that decays rapidly as x → ±∞. From the asymp-
totic solutions, he introduced a scattering function φ, considered the translations
φ(x)(y) = φ(y + 2x), and established connections with eigenvalue distributions in
random matrix theory which are described in [37]. He showed that the potential can
be recovered from the scattering data by means of the formula

d2
u(x) = −2 log det(I + φ(x) ), (1.1)
dx 2
These results were developed further by Ercolani and McKean [10] and others
[13, 38, 39] to describe the inverse spectral problem for self-adjoint Schrödinger
operators on R. Grudsky and Rybkin [17] describes the inverse scattering theory of
the KdV equation in terms of Hankel and Toeplitz operators. The latter paper uses
Sarason’s algebra H ∞ + C on the unit disc to describe compact Hankel operators.
In the current paper, we use Hankel operators within the setting of linear systems in
continuous time.
Remarkably, some of the methods of inverse scattering theory do not really need
self-adjointness. However, a significant obstacle in this approach is that Hankel
operators do not have a natural product structure, so it is unclear as to how one
can fully exploit the multiplicative properties of determinants. This paper seeks
to address this issue, by realizing Hankel operators from linear systems, and then
introducing algebras of operators on state space that reflect the properties of Hankel
operators and their Fredholm determinants. As in [25], the Lyapunov differential
equation is fundamental to the development of the theory.
Definition 1.1 (Lyapunov Equation)
(i) Let H be a complex Hilbert space, known as the state space, and L(H ) the
algebra of bounded linear operators on H with the usual operator norm. Let
(e−t A )t ≥0 be a strongly continuous (C0 ) semigroup of bounded linear operators
on H such that e−t A L(H ) ≤ M for all t ≥ 0 and some M < ∞. Let D(A)
Tau Functions Associated with Linear Systems 65

be the domain of the generator −A so that D(A) is itself a Hilbert space for the
graph norm

ξ 2D(A) = ξ 2H + Aξ 2H ,

and let A† be the adjoint of A. Let R : (0, ∞) → L(H ) be a differentiable

function. The Lyapunov equation is

dRz
− = ARz + Rz A (z > 0), (1.2)
dz

where the right-hand side is to be interpreted as a bounded bilinear form on

D(A) × D(A† ). (This is a modified form of the version in [30, p. 502].)
(ii) (Operator ideals). Let L2 (H ) be the space of Hilbert–Schmidt operators on
H , and L1 (H ) be the space of trace class operators on H , so L1 (H ) = {T :
T = V W ; V , W ∈ L2 (H )} and let det be the Fredholm determinant defined on
{I + T : T ∈ L1 (H )}; see [24].
Definition 1.2
(i) (Linear system). Let H0 be a complex separable Hilbert space which serves as
the input and output spaces; let B : H0 → H and C : H → H0 be bounded
linear operators. The continuous-time linear system (−A, B, C) is

dX
= −AX + BU, Y = CX.
dt

(ii) (Scattering function). The scattering function is φ(x) = Ce−xA B, which

is a bounded and weakly continuous function φ : (0, ∞) → L(H0 ). The
terminology is justified by [10, p. 493]. In control theory, the transfer function
is the Laplace transform of φ; see [30, p. 467].
(iii) (Hankel operator). Suppose that φ ∈ L2 ((0, ∞); L(H0)). Then the corre-
sponding Hankel operator is φ on L2 ((0, ∞); H0), where
 ∞
φ f (x) = φ(x + y)f (y) dy;
0

see [28, 30] for boundedness criteria.

Definition 1.3 (Admissible Linear System) Let (−A, B, C) be a linear system as
above; suppose furthermore that the observability operator 0 : L2 ((0, ∞); H0) →
H is bounded, where
 ∞
e−sA C † f (s) ds;
†
0 f = (1.3)
0
66 G. Blower and S. L. Newsham

suppose that the controllability operator 0 : L2 ((0, ∞); H0) → H is also

bounded, where
 ∞
0 f = e−sA Bf (s) ds.
0

(i) Then (−A, B, C) is an admissible linear system. See [30, p. 469].

(ii) Suppose furthermore that 0 and 0 belong to the ideal L2 of Hilbert–Schmidt
operators. Then we say that (−A, B, C) is (2, 2)-admissible.
The scattering map associates to any (2, 2) admissible linear system (−A, B, C) the
corresponding scattering function φ(x) = Ce−xA B. The inverse scattering problem
involves recovering data about u from φ, as in (1.1). In Sect. 2 of this paper, we
analyze the existence and uniqueness problem for the Lyapunov equation, and show
that for any (2, 2) admissible linear system, the operator, as in [1, 5],
 ∞
Rx = e−t A BCe−t A dt
x

is trace class and gives the unique solution to (1.2) with the initial condition
 dR 
x
= −AR0 − R0 A = −BC. (1.4)
dx x=0

Also, Rx ∈ L1 (H ) and the Fredholm determinant satisfies

det(I + λRx ) = det(I + λφ(x) ) (x > 0, λ ∈ C). (1.5)

Definition 1.4 (Tau Function) Given a (2, 2) admissible linear system (−A, B, C),
we define

τ (x) = det(I + Rx ).

Using this general definition of τ , we can unify several results from the scattering
theory of ordinary differential equations. Under circumstances discussed in [17] and
[34], this becomes the well-known Hitota tau function of soliton theory. Such tau
functions are also strongly analogous to the tau functions introduced by Miwa et al.
[26] to describe the isomonodromy of rational differential equations and generalize
classical results on theta functions. The connection between Fredholm determinants
and rational differential equations is further described in [11] and [37]; see also [20].
The Gelfand–Levitan–Marchenko equation [12] provides the linkage between φ and
u via Rx . Consider
 ∞
T (x, y) + (x + y) + μ T (x, z)(z + y) dz = 0 (0 < x < y) (1.6)
x
Tau Functions Associated with Linear Systems 67

where T (x, y) and (x + y) are m × m matrices with scalar entries. In the context
of (−A, B, C) we assume that (x) = Ce−xA B is known and aim to find T (x, y).
In section two, we use Rx to construct solutions to the associated Gelfand–Levitan
equation (1.6), and introduce a potential

d2
u(x) = −2 log det(I + Rx ).
dx 2
In section three, we obtain a differential equation linking (x) to u(x). In examples
of interest in scattering theory, one can calculate det(I + λRx ) more easily than
the Hankel determinant of (x) directly [10], since Rx has additional properties
that originate from Lyapunov’s equation. In section four, we introduce a differential
algebra of operators on the state space, and a homomorphism to the differential
algebra C[u, u , . . . ] that is generated by the potential. In section five, we describe
the connection between this algebra and the stationary KdV hierarchy. There is a
fundamental connection between theta functions and equations of KdV and KP type;
see [27].

2 τ Functions in Terms of Lyapunov’s Equation

and the Gelfand–Levitan Equation

The following section proves existence and uniqueness of solutions of the Lyapunov
equation (1.2), in a style suggested by Peller [30, p. 503]. Peller discusses
scattering functions that produce bounded self-adjoint Hankel operators φ , and
their realization in terms of continuous time linear systems. He observes that in
some cases one needs a bounded semigroup with unbounded generator (−A). We
prove the uniqueness results for bounded and strongly continuous semigroups, then
specialize to holomorphic semigroups. The main application is to the Gelfand–
Levitan equation (1.6), and associated determinants.
Proposition 2.1 Let (e−t A )t ≥0 be a strongly continuous and weakly asymptotically
stable semigroup on a complex Hilbert space H , so e−t A f → 0 weakly as t → ∞
for all f ∈ H . Then
(i) St : R → e−t A Re−t A for t ≥ 0 defines a strongly continuous semigroup on
L1 (H ), which has generator (−L), with dense domain of definition D(L) such
that

L(R) = AR + RA (R ∈ D(L)).
68 G. Blower and S. L. Newsham

(ii) The linear operator L : D(L) → L1 (H ) is injective, and for each R0 ∈ D(L)
with L(R0 ) = X, there exists a weakly convergent integral
 ∞
R0 = e−t A Xe−t A dt. (2.1)
0

(iii) Suppose moreover that e−t0 A L(H ) < 1 for some t0 > 0. Then L : D(L) →
L1 (H ) is surjective, the integral (2.1) converges absolutely in L1 (H ) and R0
gives the unique solution to AR0 + R0 A = X.
Proof
(i) First observe that by the uniform boundedness theorem, there exists M such
that e−t A L(H ) ≤ M for all t ≥ 0, so (e−t A )t ≥0 is uniformly bounded. Also,
the adjoint semigroup (e−t A )t ≥0 is also strongly continuous and uniformly
†

bounded, so A and A† have dense domains D(A) and D(A† ) in H . Now

L1 (H ) = H ⊗H ˆ , the projective tensor product, so for all X ∈ L1 (H ), there
exists a nuclear decomposition
∞

X= Bj Cj
j =1

∞
where Bj , Cj ∈ H satisfy XL1 (H ) = j =1 Bj H Cj H . Then

∞
 ∞

−t A −t A −t A
St (X) − X = (e Bj Cj e − Bj Cj e ) + (Bj Cj e−t A − Bj Cj )
j =1 j =1

where (e−t A ) is bounded, e−t A Bj − Bj H → 0 and e−t A Cj − Cj H → 0

†

as t → 0+; so St (X) − XL1 (H ) → 0 as t → 0+; so (St )t ≥0 is strongly

continuous on L1 (H ). By the Hille–Yoshida theorem [15, p. 16], there exists
a dense linear subspace D(L) of L1 (H ) such that St (R) is differentiable at
t = 0+ for all R ∈ D(L), and (d/dt)t =0+ St (R) = −AR − RA, so the
generator is (−L), where L(R) = AR + RA.
(ii) Certainly D(L) contains D(A† )⊗D(A)ˆ ˆ . Choosing f ∈
in L1 (H ) = H ⊗H
D(A) and g ∈ D(A ), we find that
†

d  −t A     
e R0 e−t A f, g = − e−t A (AR0 + R0 A)e−t A f, g = − e−t A Xe−t A f, g
dt
a continuous function of t > 0; so integrating we obtain

   −sA −sA
 s  
R0 f, g − e R0 e f, g = e−t A Xe−t A f, g dt.
0
Tau Functions Associated with Linear Systems 69

We extend this identity to all f, g ∈ H by joint continuity; then we let s →

∞ and observe that R0 : H → H is trace class and hence is completely
continuous, hence R0 maps the weakly null family (e−sA f )s→∞ to the norm
convergent family (R0 e−sA f )s→∞ , so e−sA R0 e−sA f, g → 0 as s → ∞;
hence we have a weakly convergent improper integral

  s  
R0 f, g = lim e−t A Xe−t A f, g dt (f, g ∈ H ).
s→∞ 0

(iii) The function t → e−t A Xe−t A takes values in the separable space L1 (H )
and is weakly continuous, hence strongly measurable, by Pettis’s theorem.
By considering the spectral radius, Engel and Nagel [9] show that there exist
δ > 0 and Mδ > 0 such that e−t A L(H ) ≤ Mδ e−δt for all t ≥ 0; hence (2.1)
converges as a Bochner–Lebesgue integral with
 ∞ Mδ2
Rx L1 (H ) ≤ Mδ2 XL1 (H ) e−2δt dt ≤ XL1 (H ) e−2δx .
x 2δ

Furthermore, A is a closed linear operator and satisfies

  
s s s d −t A −t A 
A e−t A Xe−t A dt + e−t A Xe−t A dtA = − e Xe dt
x x x dt
−xA
=e Xe−xA − e−sA Xe−sA
→ e−xA Xe−xA

as s → ∞ where
 s
e−t A Xe−t A dt → Rx ;
x

so ARx + Rx A = e−xA Xe−xA for all x ≥ 0. We deduce that x → Rx is a

differentiable function from (0, ∞) to L1 (H ) and that the modified Lyapunov
equation (1.2) holds.
The hypotheses (i) and (ii) are symmetrical under the adjoint (A, R0 ) → (A† , R0† );
however, the hypothesis (iii) is rather stringent, and in many applications one only
needs existence of the integral (2.1).
Definition 2.2 ((2, 2) Admissible Linear Systems)
(i) Let H be a complex Hilbert space and let  = (−A, B, C) be a linear system
with state space H . Suppose that the integral
 ∞
e−t A BB † e−t A dt
†
Wc =
0
70 G. Blower and S. L. Newsham

converges weakly and defines a bounded linear operator on H ; then Wc is the

controllability Gramian. Suppose further that the integral
 ∞
e−t A C † Ce−t A dt
†
Wo =
0

converges weakly and defines a bounded linear operator on H ; then Wo is the

observability Gramian.
(ii) Then as in [5, p. 318] we define Rx to be the bounded linear operator on H
determined by the weakly convergent integral
 ∞
Rx = e−t A BCe−t A dt. (2.2)
x

(iii) Then  satisfying (i) is said to be balanced if Wc = Wo and ker(Wc ) = 0; see

[30, p. 499].
(iv) Also,  satisfying (i) is said to be (2, 2) admissible if Wc and Wo are trace
class, or equivalently 0 and 0 are Hilbert-Schmidt; see [5].
(v) We introduce the scattering function φ(t) = Ce−t A B and the shifted
scattering function φ(x) (t) = φ(t + 2x) for x, t > 0.
(vi) (Sectorial operator). For 0 < θ ≤ π, we introduce the sector

Sθ = {z ∈ C \ {0} : | arg z| < θ }.

A closed and densely defined linear operator −A is sectorial [9, 15] if there
exists π/2 < θ < π such that Sθ is contained in the resolvent set of −A and
|λ|(λI + A)−1 L(H ) ≤ M for all λ ∈ Sθ . Let D(A) be the domain of A and
D(A∞ ) = ∩∞ n=0 D(A ). See [15, p. 37].
n

(vii) For π/2 < δ < π, we introduce Xδ = {ζ ∈ Sδ : −ζ ∈ Sδ } which is an open

set, symmetrical about iR and bounded by lines passing through 0.
Theorem 2.3 Let (−A, B, C) be a linear system such that e−t0 A L(H ) < 1
for some t0 > 0, and that B and C are Hilbert–Schmidt operators such that
BL2 (H0 ;H ) CL2 (H ;H0 ) ≤ 1. Suppose further that −A is sectorial on Sθ for some
π/2 < θ < π.
(i) Then (−A, B, C) is (2, 2)-admissible, so the trace class operators (Rx )x>0
give the solution to Lyapunov’s equation (1.2) for x > 0 that satisfies the
initial condition (1.4), and the solution to (1.2) with (1.4) is unique.
(ii) The function τ (x) = det(I + Rx ) is differentiable for x ∈ (0, ∞).
(iii) Then Rz extends to a holomorphic function that satisfies (1.2) on Sθ−π/2 , and
Rz → 0 as z → ∞ in Sθ−ε−π/2 for all 0 < ε < θ − π/2.
Proof
(i) Since BC ∈ L1 (H ), the integrand of (2.2) takes values in L1 (H ), and we can
apply Proposition 2.1(iii) to X = BC.
Tau Functions Associated with Linear Systems 71

(ii) The Fredholm determinant R → det(I +R) is a continuous function on L1 (H ).

Also the integral
 ∞
Rx = e−t A BCe−t A dt
x

belongs to D(L) and gives a differentiable function of x > 0 with values in

L1 (H ).
(iii) By classical results of Hille [15, p. 34], (e−zA )z∈Sθ −π/2 defines an analytic
semigroup on Sθ−π/2 , bounded on Sν for all 0 < ν < θ −π/2, so we can define
Rz = e−zA R0 e−zA and obtain an analytic solution to Lyapunov’s equation. For
all 0 < ε < θ −π/2, there exists Mε such that e−zA L(H ) ≤ Mε for all z ∈ Sδ
where δ = θ − ε − π/2. Now for z ∈ Sδ/2 , we write z = x/2 + (x/2 + iy)
with x/2 + iy ∈ Sδ and use the bound

e−zA L(H ) ≤ e−xA/2 L(H ) e−(x/2+iy)AL(H )

to obtain e−zA L(H ) ≤ Mε2 e−t0 A L(H )0 , so e−zA L(H ) → 0 exponentially

x/(4t )

fast as z → ∞ in the sector Sδ/2 . Hence Rz is holomorphic and bounded on

S(θ−ε−π/2) and by (2.2), Rz → 0 as z → ∞ in S(θ−ε−π/2)/2.
Example
(i) Let ! = −d 2 /dx 2 be the usual Laplace operator which is essentially self-
adjoint and non-negative on Cc∞ (R; C) in L2 (R; C). We  introduce A =
√
I + ! which is given by the Fourier multiplier F Af (ξ ) = 1 + ξ 2 F f (ξ ).
Then (e−zA ) and (e−zA ) give bounded holomorphic semigroups on H , as in
2

Theorem 2.3, on the right half-plane {z ∈ C : #z ≥ 0}, which is the closure

of Sπ/2 . On the imaginary axis, we have unitary groups (eit A ) and (e−it A ). By
2

classical results from wave equations, we can write eit A + e−it A = 2 cos(tA)
where u(x, t) = cos(tA)f (x) for f ∈ Cc∞ (R; C) is given by
 
1  t x+t J0 ( t 2 − (x − s)2 )
u(x, t) = f (x + t) + f (x − t) + f (y)  ds,
2 2 x−t t 2 − (x − s)2

where J0 is Bessel’s function of the first kind of order zero, and u satisfies

∂ 2u ∂ 2u ∂u
− 2 = u(x, t), u(x, 0) = f (x), (x, 0) = 0.
∂x 2 ∂t ∂t

See [15, p. 121]. Note that (exp(t (iA)2j −1)) gives a unitary group on H
for j = 0, 1, 2, . . . . This can be used to deform the linear system in the
sense of Proposition 2.5(iii). Unitary deformation groups for tau functions are
considered in [26]
72 G. Blower and S. L. Newsham

(ii) In section 4 of [5], we introduced linear systems to describe Schrödinger’s

equation when the potential is smooth and localized. In [17], the authors obtain
detailed results about the corresponding Hankel operator.
Definition 2.4 (Block Hankel Operators)
(i) Say that  ∈ L(H ) is block Hankel if there exists 1 ≤ m < ∞ such that
 is unitarily equivalent to the block matrix [Aj +k−2 ]∞ 2 m
j,k=1 on  (C ) where
Aj ∈ C m×m for j = 0, 1, . . . .
(ii) Let (−A, B, C) be a (2, 2) admissible linear system with input and output
space H0 , where the dimension of H0 over C is m < ∞. Then m is the number
of outputs of the system, and systems with finite m > 1 are known as MIMO
for multiple input, multiple output, and give rise to block Hankel operators
with (x) = Ce−xA B.
(iii) The Gelfand–Levitan integral equation for (−A, B, C) as in (ii) is (1.6), where
T (x, y) and (x + y) are m × m matrices with scalar entries, and μ ∈ C. We
proceed to obtain a solution.

Proposition 2.5
(i) In the notation of Theorem 2.3, there exists x0 > 0 such that

Tμ (x, y) = −Ce−xA (I + μRx )−1 e−yA B

satisfies the integral equation (1.6) for x0 < x < y and |μ| < 1.
(ii) The determinant satisfies det(I + μRx ) = det(I + μ(x) ) and

d
μ trace Tμ (x, x) = log det(I + μRx ).
dx

(iii) Suppose that t → U (t) is a continuous function [0, 1] → L(H ) such that
U (t)A = AU (t) and U (t)L(H ) ≤ 1. Then there is a family of (2, 2)
admissible linear systems

(t) = (−A, U (t)B, CU (t)) (t ∈ [0, 1]);

the corresponding tau function τ (x, t) is continuous for (x, t) ∈ (0, ∞) ×

[0, 1].
Proof
(i) We choose x0 so large that eδx0 ≥ Mδ /2δ, then by Theorem 2.3(iii), we have
|μ|Rx L(H ) < 1 for x > x0 , so I + μRx is invertible. Substituting Tμ (x, y)
into the integral equation (1.6), we obtain

Ce−(x+y)A B − Ce−xA (I + μRx )−1 e−yA B

 ∞
− μCe−xA (I + μRx )−1 e−zA BCe−zA dze−yA B
x
Tau Functions Associated with Linear Systems 73

= Ce−(x+y)A B − Ce−xA (I + μRx )−1 e−yA B

− μCe−xA (I + μRx )−1 Rx e−yA B
= 0.

(ii) As in (1.3), the operator x : L2 (0, ∞) → H is Hilbert–Schmidt; likewise

x : L2 (0, ∞) → H is Hilbert–Schmidt; so (−A, B, C) is (2, 2)-admissible.
Hence (x) = †x x and Rx = x †x are trace class, (I + μRx ) is a
holomorphic function of x on some sector Sδ as in Theorem 2.3 and

det(I + μRx ) = det(I + μx †x ) = det(I + μ†x x ) = det(I + μ(x) ).

By the Riesz functional calculus, (I + μRx )−1 is meromorphic for x in some

Sδ . Correcting a typographic error in [5, p. 324], we rearrange terms and
calculate the derivative
 
μTμ (x, x) = −μtrace Ce−xA (I + μRx )−1 e−xA B
 
= −μtrace (I + μRx )−1 e−xA BCe−xA
 dRx 
= μtrace (I + μRx )−1
dx
d
= trace log(I + μRx ).
dx
This identity is proved for |μ| < 1 and extends by analytic continuation to the
maximal domain of Tμ (x, x).
(iii) Since A commutes with U (t), the domain D(A) is invariant under U (t), and
the multiplications B → U (t)B and C → CU (t) preserve the hypotheses of
Theorem 2.3, so (−A, U (t)B, CU (t)) is (2, 2) admissible. By commutativity,
we have τ (x, t) = det(I + U (t)Rx U (t)), which depends continuously on
(x, t).

3 The Baker–Akhiezer Function of an Admissible Linear

System

In this section, we consider the Darboux addition rule for potentials and analyze
the transformation (−A, B, C) → (−A, B, −C) and the effect on the ratios and
derivatives of τ functions. This generalizes [10, section 3.4], and allows us to
introduce a version of the Baker–Akhiezer function for a family of linear systems
with properties that are similar to the classical case, as presented in [3] and [22].
74 G. Blower and S. L. Newsham

Definition 3.1 (Baker–Akhiezer Function)

(i) Let (−A, B, C) be as in Theorem 2.3, and let

ζ = (−A, (ζ I + A)(ζ I − A)−1 B, C) (ζ ∈ C ∪ {∞} \ Spec(A))

so that ζ defines a (2, 2) admissible linear systems for ζ in an open subset

of C ∪ {∞} which includes {ζ ∈ C : −ζ ∈ Sθ } for some π/2 < θ < π. We
identify ∞ with (−A, B, C), and 0 with (−A, B, −C).
(ii) Let τζ be the tau function of ζ , and let the Baker–Akhiezer function for the
family of linear systems be

τζ (x) 
ψζ (x) = exp ζ x . (3.1)
τ∞ (x)

(iii) Let τζ∗ (x) = τζ̄ (x̄) as in Schwarz’s reflection principle, and let

ζ∗ = (−A† , C † , B † (ζ I + A† )(ζ I − A† )−1 ) (ζ ∈ C ∪ {∞} \ Spec(A† ))

so ζ → ζ∗ is an involution, and ζ∗ has tau function τ ∗ .

The following result introduces a family of solutions of Schrödinger equation
corresponding to the ζ with an addition rule in the style of Darboux.
Proposition 3.2 Let (−A, B, C) be as in Theorem 2.3.
(i) Then for −ζ ∈ Sθ , the linear system ζ is also (2, 2) admissible, and the
Baker–Akhiezer function satisfies

d2
− ψζ (x) + u∞ (x)ψζ (x) = −ζ 2 ψζ (x). (3.2)
dx 2

(ii) There exist hj ∈ C ∞ ((0, ∞); C) such that there is an asymptotic expansion
 h1 (x) h2 (x) 
ψζ (x) $ eζ x 1 + + 2
+...
ζ ζ

as ζ → ±i∞, and the expansion is uniform for x in compact subsets of (0, ∞).
Proof
(i) For all ζ ∈ C \ Spec(A), there exists x0 (ζ ) such that

(ζ I + A)(ζ I − A)−1 Rx L1 (H ) < 1

for all x > x0 (ζ ), so that τζ (x) is continuously differentiable and non-zero as a

function of x ∈ (x0 (ζ ), ∞). In particular, suppose that #ζ < 0, then −ζ ∈ Sθ
so ζ I − A is invertible. Using the R function for ζ , we write
Tau Functions Associated with Linear Systems 75


τζ (x) det I + (ζ I + A)(ζ I − A)−1 Rx
= 
τ∞ (x) det I + Rx

det I + (ζ I − A)−1 ((ζ I − A)Rx + ARx + Rx A)
= 
det I + Rx

det I + Rx + (ζ I − A)−1 (ARx + Rx A)
=  (3.3)
det I + Rx

so that when ARx + Rx A has rank one, the perturbing term (ζ I − A)−1 (ARx +
Rx A) has rank one; continuing we find

τζ (x) 
= det I + (ζ I − A)−1 e−xA BCe−xA (I + Rx )−1
τ∞ (x)

= det I + Ce−xA (I + Rx )−1 (ζ I − A)−1 e−xA B
= 1 + Ce−xA (I + Rx )−1 (ζ I − A)−1 e−xA B,

since B : C → H and C : H → C have rank one. Hence

τζ (x) 
ψζ (x) = exp ζ x
τ∞ (x)
 
= exp ζ x + Ce−xA (I + Rx )−1 (ζ I − A)−1 e−xA B exp ζ
 ∞
 
= exp ζ x − Ce−xA (I + Rx )−1 e−yA B exp ζy dy
x
 ∞
 
= exp ζ x + T (x, y) exp ζy dy.
x

Here T satisfies the Gelfand–Levitan equation, and by integrating by parts, we

see that

∂ 2T ∂ 2T
− = u(x)T (x, y)
∂x 2 ∂y 2
2
where u(x) = −2 dx d
2 log τ (x). Then by integrating by parts, we see that ψζ
satisfies Schrödinger’s equation.
The solutions of the differential equation depend analytically on ζ at those
points where the potential depends analytically on ζ ; note that ζ → τζ (x) is
holomorphic and non zero for Rx  < 1 and −ζ ∈ Sθ . Then we continue the
solutions analytically to all −ζ in the sector Sθ , on which ψζ (x) is holomorphic
as a function of ζ for x > 0.
(ii) Observe that Xθ = Sθ ∩ (−Sθ ) contains iR \ {0}. For ζ ∈ Sθ ∩ (−Sθ ), by (i)
there exist solutions ψζ (x) and ψ−ζ (x) to (3.2). In particular, ψik and ψ−ik (x)
76 G. Blower and S. L. Newsham

are solutions for k > 0. We integrate by parts repeatedly

 ∞
e−xA (ζ I − A)−1 =e−xA eζ s e−sA ds
0

e−xA Ae−xA Ak−1 e−xA

= + + · · · +
ζ ζ2 ζk
 ∞ k −xA
A e
+ eζ s e−sA ds,
0 ζk

where the integral converges by the hypothesis of Theorem 2.3. Also, (e−zA ) is
an analytic semigroup in the sector Sθ−π/2 , so D(Aj ) is a dense linear subspace
of H for all j = 1, 2, . . . and Aj e−xA ∈ L(H ) and by Cauchy’s estimates
there exists C > 0 such that Aj e−xA L(H ) ≤ Cj !/x j for all x > 0. So we can
generate an asymptotic expansion of (3.3) with terms

hj (x) = Ce−xA (I + Rx )−1 Aj −1 e−xA B

which are bounded on compact subsets of (0, ∞).

Definition 3.3 (Darboux Transforms) Let (−A, B, C) be a (2, 2) admissible
linear system with tau function τ∞ (x; μ) = det(I + μRx ). Define the Dar-
boux transform of (−A, B, C) to be (−A, B, −C) with tau function transform
τ0 (x; μ) = det(I − μRx ). Let

1 d τ∞ 1 d 
v= log , w= log τ0 τ∞ ,
μ dx τ0 μ dx
2 d2 2 d2
u∞ = − 2 2 log τ∞ , u0 = − 2 2 log τ0 .
μ dx μ dx

In the following result, we show how products and quotients of τ functions can be
linked by the Gelfand–Levitan equation for 2 × 2 matrices, and satisfy the identities
usually associated with Darboux transforms in the theory of integrable systems. See
[23].
Theorem 3.4 Let (−A, B, C) be a (2, 2)-admissible linear system with input and
output spaces C, and let φ(x) = Ce−xA B.
(i) Then there exists δ > 0 such that for all μ ∈ C such that |μ| < δ, the Gelfand-
Levitan equation (1.6) with
! " ! "
W (x, y) V (x, y) 0 φ(x + y)
T (x, y) = , (x + y) =
V (x, y) W (x, y) φ(x + y) 0
Tau Functions Associated with Linear Systems 77

has a solution such that

1 d  1 d τ∞ (x; μ)
W (x, x) = log τ∞ (x; μ)τ0 (x; μ) , V (x, x) = log
2μ dx 2μ dx τ0 (x; μ)

and
1 d
W (x, x) = −V (x, x)2 ;
2μ dx

(ii) also Toda’s equation holds in the form

τ0 τ∞ − 2τ0 τ∞ + τ0 τ∞ = 0. (3.4)

Proof
(i) Let

T∞ (x, y) = −Ce−xA (I + μRx )−1 e−yA B,

T0 (x, y) = Ce−xA (I − μRx )−1 e−yA B

and
! "
0 φ(x)
(x) = .
φ(x) 0

Now let
! "
1 T∞ + T0 T∞ − T0
T (x, y) =
2 T∞ − T0 T∞ + T0

so that
! "! "! "−1 ! −yA "! "
C 0 e−xA 0 I μRx e 0 0 B
T (x, y) = −
0 C 0 e−xA μRx I 0 e−yA B 0

hence T satisfies the Gelfand–Levitan equation (1.6).

(ii) As in Proposition 2.5,

1 d 1 d
T∞ (x, x) = log τ∞ (x), T0 (x, x) = log τ0 (x);
μ dx μ dx

hence (3.4) is equivalent to the condition

d 2 d
T0 (x, x) + μ T0 (x, x) − T∞ (x, x) + T∞ (x, x) = 0, (3.5)
dx dx
78 G. Blower and S. L. Newsham

which we now verify. The left-hand side of (3.5) equals


Ce−xA − A(I − μRx )−1 − (I − μRx )−1 μ(ARx + Rx A)(I − μRx )−1

− (I − μRx )−1 A e−xA B
 
+ Ce−xA (I − μRx )−1 + (I + μRx )−1 e−xA μBCe−xA
 
× (I − μRx )−1 + (I + μRx )−1 e−xA

+ Ce−xA A(I + μRx )−1 − (I + μRx )−1 μ(ARx + Rx A)(I + μRx )−1

+ (I + μRx )−1 A e−xA B. (3.6)

All of the terms begin with Ce−xA and end with e−xA B, and we can replace
e−xA μBCe−xA by μ(ARx + Rx A) to obtain

(3.6) = Ce−xA − 2(I − μRx )−1 A(I − μRx )−1

+ 4(I − μ2 Rx2 )−1 μ(ARx + Rx A)(I − μ2 Rx2 )−1


+ 2(I + μRx )−1 A(I + μRx )−1 e−xA B

= 0.

This proves (3.5), and one can easily check that (3.4) is equivalent to
1 dv 1 dw
u0 (x) = + v(x)2 , v(x)2 = − .
μ dx μ dx
The entries of T satisfy the pair of coupled integral equations
 ∞
0 = W (x, y) + μ V (x, s)φ(s + y) ds
x
 ∞
0 = V (x, y) + φ(x + y) + μ W (x, s)φ(s + y) ds;
x

so W satisfies
 ∞
0 = −W (x, z) + μ φ(x + y)φ(y + z) dy
x
 ∞  ∞
2
+μ W (x, s) φ(s + y)φ(y + z) dyds,
x x

which explains how μ2 φ2 enters into several determinant formulas [37].
Tau Functions Associated with Linear Systems 79

Definition 3.5 (Darboux Addition)

(i) For −ζ ∈ Sθ ∪ {0} we define the Darboux addition rule on (2, 2) admissible
linear systems by

Mζ : (−A, B, C) → (−A, (ζ I + A)(ζ I − A)−1 B, C)

and correspondingly on potentials by

u∞ → uζ = u∞ − 2(log ψζ ) .

(ii) Let Wr(ϕ, ψ) be the Wronskian of ψ, ϕ ∈ C 1 ((0, ∞); C).

Corollary 3.6 The set {Mζ , (ζ ∈ Xθ ), M0 , M∞ = I } generates a group such that
M02 = I , Mζ M−ζ = I and Mζ Mη corresponds to adding

d2
−2 log Wr(ψζ , ψη )
dx 2
to the potential.
Proof The definition is consistent with [10, p. 484]. In particular, ψ0 (x) =
d2
τ0 (x)/τ∞ (x), and u0 (x) = u∞ (x) − 2 dx 2 log ψ0 (x), which is consistent with (3.5).
For ζ1 = ζ2 , let "(x) = Wr(ψζ1 , ψζ2 )/ψζ2 , and observe that

" = ζ22 + u∞ − 2(log ψζ1 ) ".

This gives the basic composition rule for Mζ2 Mζ1 . The other statements follow from
Proposition 3.2 and Theorem 3.4. See [24].

4 The State Ring Associated with an Admissible Linear

System

Gelfand and Dikii [11] considered the algebra Au = C[u, u , u , . . . ] of complex

polynomials in a smooth potential u and its derivatives. They showed that if u
satisfies the stationary higher order KdV equations (5.1), then Au is a Noetherian
ring [2] and the associated Schrödinger equation is integrable by quadratures; see
[7]. In this section, we introduce an analogue A for an admissible linear system.
We develop a calculus for Rx which is the counterpart of Pöppe’s functional
calculus for Hankel operators from [24, 31, 32]. As we see in other papers, our
theory of state rings has wider scope for generalization.
80 G. Blower and S. L. Newsham

Definition 4.1 (Differential Rings)

(1) Let R be a ring with ideal J , and let ∂ : R → R be a derivation. Then
RJ = {r ∈ R : ∂(r) ∈ J } gives a subring of R, the ring of constants relative
to J . When R is an algebra over C and J = (0), we call R0 the constants; see
[33].
(2) (State ring of a linear system). Let (−A, B, C) be a linear system such that
A ∈ L(H ). Suppose that:
(i) S is a differential subring of C ∞ ((0, ∞); L(H ));
(ii) I, A and BC are constant elements of S;
(iii) e−xA , Rx and Fx = (I + Rx )−1 belong to S.
Then S is a state ring for (−A, B, C).
Lemma 4.2 Suppose that (−A, B, C) is a linear system with A ∈ L(H ) and that
Rx gives a solution of Lyapunov’s equation (1.2) such that I + Rx is invertible
for x > 0 with inverse Fx . Then the free associative algebra S generated by
I, R0 , A, F0 , e−xA , Rx and Fx is a state ring for (−A, B, C) on (0, ∞). For all
t > 0, there exists a ring homomorphism St : S → S given by St : G(x) →
G(x + t) such that St commutes with d/dx.
Proof We can regard S as a subring of Cb ((0, ∞), L(H ))), so the multiplication is
well defined. Then we note that BC = AR0 + R0 A belongs to S, as required. We
also note that (d/dx)e−xA = −Ae−xA and that Lyapunov’s equation (1.2) gives

d
(I + Rx )−1 = (I + Rx )−1 (ARx + Rx A)(I + Rx )−1 ,
dx
which implies

dFx
= AFx + Fx A − 2Fx AFx ,
dx
with the initial condition

AF0 + F0 A − 2F0 AF0 = F0 BCF0 .

Hence S is a differential ring.

We can map I → I , e−xA → e−(x+t )A, R0 → e−t A R0 e−t A , Rx → e−t A Rx e−t A
and Fx → (I + e−t A Rx e−t A )−1 , and thus produce a ring homomorphism G(x) →
G(x + t) which satisfies

(d/dx)St G(x) = G (x + t) = St (d/dx)G(x).

Tau Functions Associated with Linear Systems 81

Definition 4.3 (Products and Brackets)

(i) Given a state ring S for (−A, B, C), let B be any differential ring of functions
from (0, ∞) → L(H0 ). Let

A = spanC {An1 , An1 Fx An2 . . . Fx Anr : nj ∈ N}.

(ii) On S we introduce the associative product ∗ by

P ∗ Q = P (AF + FA − 2FAF )Q (P , Q ∈ S),

which is distributive over the standard addition, and the derivation ∂ : S → S

by

dP
∂P = A(I − 2F )P + + P (I − 2F )A (P ∈ S). (4.1)
dx
(iii) Let ! · " : S → B be the linear map

!Y " = Ce−xA Fx Y Fx e−xA B (Y ∈ S),

so that x → !Y " is a differentiable function (x0 , ∞) → L(H0 ).

For x0 ≥ 0 and 0 < φ < π, let Sδx0 be the translated sector
x
Sδ 0 = {z = x0 + w : w ∈ C \ {0}; | arg w| < δ}

and let H ∞ (Sδx0 ) the bounded holomorphic complex functions on Sδx0 . Then let
∞ = ∪ ∞ x0
H∞ x0 >0 H (Sδ ) be the algebra of complex functions which are bounded on
some translated sector Sδx0 , with the usual pointwise multiplication.
Theorem 4.4 Let (−A, B, C) be a (2, 2)-admissible linear system with H0 = C as
in Theorem 2.3, so (e−zA ) for z ∈ Sφ0 is a bounded holomorphic semigroup on H .
Let 0 = {P ∈ A : !P " = 0}.
(i) Then (A , ∗, ∂) is a differential ring with bracket !·";
(ii) there is a homomorphism of differential rings ! · " : (A , ∗, ∂) →
(H∞ ∞ , ·, d/dz);

(iii) 0 is a differential ideal in (A , ∗, ∂) such that A /0 is a commutative

differential ring, and an integral domain.
Proof
(i) We can multiply elements in S by concatenating words and taking linear
combinations. Since all words in A begin and end with A, we obtain
words of the required form, hence A is a subring of S. To differentiate
a word in A we add words in which we successively replace each Fx by
AFx + Fx A − 2Fx AFx , giving a linear combination of words of the required
82 G. Blower and S. L. Newsham

form. The basic observation is that dF /dx = AF + FA − 2FAF , so one can

check that

∂(P ∗ Q) = (∂P ) ∗ Q + P ∗ (∂Q); (4.2)

hence (S, ∗, ∂) is a differential ring with differential subring (A , ∗, ∂).

(ii) Now we verify that there is a homomorphism of differential rings
(A , ∗, ∂) → (B, ·, d/dx) given by P → !P ". From the definition of
Rx , we have ARx + Rx A = e−xA BCe−xA , and hence

Fx e−xA BCe−xA Fx = AFx + Fx A − 2Fx AFx ,

which implies
+ ,+ ,
P Q = Ce−xA Fx P Fx e−xA BCe−xA Fx QFx e−xA B
= Ce−xA Fx P (AFx + Fx A − 2Fx AFx )QFx e−xA B
+ ,
= P (AFx + Fx A − 2Fx AFx )Q
= !P ∗ Q".

Moreover, the first and last terms in !P " have derivatives

d d
Ce−xA Fx = Ce−xA Fx A(I − 2Fx ), Fx e−xA B = (I − 2Fx )AFx e−xA B,
dx dx

so the bracket operation satisfies

d + , - dP .
P = A(I − 2Fx )P + + P (I − 2Fx )A = !∂P ". (4.3)
dx dx
In this case A is possibly unbounded as an operator, so we use the holomorphic
semigroup to ensure that products (4.1) and brackets (4.2) are well defined.
We observe that A has a grading A = ⊕∞ n=1 An , where An is the span of
the elements that have total degree n when viewed as products of A and F .
For Xn ∈ An and Ym ∈ Am , we have Xn ∗ Ym ∈ An+m+2 ⊕ An+m+3 and
∂Xn ∈ An+1 ⊕ An+2 .
Also we have Ak e−zA ∈ L(H ) for all z ∈ Sφ0 and Ak e−zA L(H ) → 0
as z → ∞ in Sφ0 ; hence Rz Ak → 0 and Ak Rz → 0 in L(H ) as z → ∞
in Sφ0 . Hence there exists an increasing positive sequence (xk )∞k=0 such that
x
Ak Fz − Ak ∈ L(H ) for all z ∈ Sφk and Ak Fz − Ak → 0 in L(H ) as z → ∞ in
Sφxk . Let Xn ∈ An and consider a typical summand AFz Ak Fz . . . A in Xn ; we
replace each factor like Ak Fz by the sum of Ak (Fz − I ) and Ak where k ≤ n;
then we observe that there in an initial factor Ce−zA and a final factor e−zA B
in !Xn "; hence !Xn " determines an element of H ∞ (Sφxn ).
Tau Functions Associated with Linear Systems 83

∞ with the algebraic direct limit

We can identify H∞

∞ x +n
H∞ = lim H ∞ (Sφ0 ).
n→∞

∞ is consistently
By the principle of isolated zeros, the multiplication on H∞
∞
defined, and H∞ is an integral domain. Now each f ∈ H∞ ∞ gives f ∈
∞ x0 ∞ x0 +1 ∞ . From
H (Sφ ) so f ∈ H (Sφ ) by Cauchy’s estimates, so f ∈ H∞
(i) we deduce that

! · " : ⊕∞ ∞ ∞ xn
n=1 An → ∪n=1 H (Sφ )

is well-defined and the bracket is multiplicative with respect to ∗, and behaves

naturally with respect to differentiation.
(iii) We check that ! · " is commutative on (A , ∗, ∂), by computing

!P ∗ Q" = trace Ce−xA F P F e−xA BCe−xA F QF e−xA B

= trace Ce−xA F QF e−xA BCe−xA F P F e−xA B
= !Q ∗ P ".

Hence 0 contains all the commutators P ∗ Q − Q ∗ P , and 0 is the kernel

of the homomorphism ! · ", hence is an ideal for ∗. Also, we observe that for
all Q ∈ 0 , we have ∂Q ∈ 0 since !∂Q" = (d/dx)!Q" = 0. Hence 0
is a differential ideal which contains the commutator subspace of (A , ∗), so
A /0 is a commutative algebra. Also, ∂ determines a unique derivation ∂¯ on
A /0 by ∂Q ¯ = ∂Q + 0 for all Q ∈ A ; hence A /0 is a differential
algebra. We can identify A /0 with a subalgebra of H∞ ∞ , which is an integral

domain.
Remark 4.5 Pöppe [31] introduced a linear functional & . ' on Fredholm kernels
K(x, y) on L2 (0, ∞) by &K' = K(0, 0). In particular, let K, G, H, L be integral
operators on L2 (0, ∞) that have smooth kernels of compact support, let  = φ(x)
have kernel φ(s + t + 2x), let  = dx
d
 and G = ψ(x) be another Hankel operator;
then the trace satisfies
d
&' = − trace , (4.4)
dx
1 d
&KG' = − trace KG, (4.5)
2 dx

&(I + )−1 ' = −trace (I + )−1  , (4.6)
1
&K'&GL' = − &K( G + G )L', (4.7)
2
84 G. Blower and S. L. Newsham

where (4.7) is known as the product formula. The easiest way to prove (4.4)–(4.7)
is to observe that  G + G is the integral operator with kernel −2φ(x)(s)ψ(x) (t),
which has rank one. These ideas were subsequently revived by McKean [24], and are
implicit in some results of [37]. Our formulas (4.2) and (4.3) incorporate a similar
idea, and are the basis of the proof of Theorem 4.4. The results we obtain appear to
be more general than those of Pöppe, and extend to periodic linear systems [6].
For the remainder of this section, we let A be a n × n complex matrix with
eigenvalues λj (j = 1, . . . , m) with geometric multiplicity nj such that λj +λk = 0
for all j, k ∈ {1, . . . , m}; let K = C(e−λ1 t , . . . , e−λm t , t). Also, let B ∈ Cn×1
and C ∈ C1×n . The formula (4.9) resembles the expressions used to obtain soliton
solutions of KdV, as in [19, (14.12.11)] and [16]. In [17, (6.25)], there is a discussion
of how the scattering data evolve under the time evolution associated with the KdV
flow.
Proposition 4.6
(i) There exists a solution Rt to Lyapunov’s equation (1.2) with R0 = BC, such
that the entries of Rt belong to K, and τ (t) ∈ K;
(ii) φ ∈ K satisfies a linear differential equation with constant coefficients.
(iii) Suppose further that all the eigenvalues of A are simple. Then there exists an
invertible matrix S such that

S −1 B = (bj )nj=1 ∈ Cn×1 , CS = (cj )nj=1 ∈ C1×n

and the tau function is given by


n
bj cj e−2λj t
τ (t) =1 +
2λj
j =1

 bj bm ck cp e−(λj +λk +λm +λp )t

+ (−1)j +k+m+p + ...
(λj + λm )(λk + λp )
(j,k),(m,p):j =m;k=p

/
n
b j cj / (λj − λk )2 −2 nj=1 λj t
+ e .
2λj (λj + λk )2
j =1 1≤j <k≤n

Proof
(i) By the hypothesis, we can introduce a chain of circles C that go once round
each λj in the positive sense and have all the points −λk in their exterior. Then
by [4], the matrix

−1
R0 = (A + λI )−1 BC(A − λI )−1 dλ
2πi C
Tau Functions Associated with Linear Systems 85

gives a solution to Sylvester’s equation in the form −AR0 − R0 A = −BC. To

see this, one considers (A + λI )R0 + R0 (A − λI ) and then uses the calculus
of residues. By the Riesz functional calculus, we also have

1 −1
e−t A = λI − A e−t λ dλ;
2πi C

hence by Cauchy’s residue theorem, there exist complex polynomials pj and

qj , and integers mj ≥ 0 such that


m
e−t A = qj (t)e−t λj pj (A), (4.8)
j =1

where qj (t) is constant if the corresponding eigenvalue is simple. We let

Rt = e−t A R0 e−t A , which gives a solution to Lyapunov’s equation with
initial condition −BC. From (4.8), we see that all the entries of Rt belong
to K. By the Laplace expansion of the determinant, we see that all entries of
τ (t) = det(I + Rt ) also belong to K.
(ii) We have φ(t) = Ce−t A B ∈ K by (4.8). Also, n we introduce the characteristic
polynomial of (−A) by det(λI + A) = j =0 a j λj . Then by the Cayley–

Hamilton theorem, nj=0 aj φ (j ) (t) = 0.
(iii) There exists an invertible matrix S such that SAS −1 is the n×n diagonal matrix
D = diag(λ1 , . . . , λn ), and we observe that

bj ck e−(λj +λk )t n
Rt =
λj + λk j,k=1

satisfies
d
Rt = −[bj ck e−(λj +λk )t ]nj,k=1 , −DRt − Rt D = −[bj ck e−(λj +λk )t ]nj,k=1 ;
dt
so Rt gives a solution of the Lyapunov equation with generator −D and initial
condition given by the rank-one matrix −S −1 BCS = −[bj ck ]nj,k=1 . Hence
the tau function is given by τ (t) = det(I + Rt ) for this matrix, and there is an
expansion

bj ck e−(λj +λk )x n  bj ck e−λj x−λk x

det δj k + = det
λj + λk j,k=1 λj + λk j,k∈σ
σ ⊆{1,...,n}
(4.9)
86 G. Blower and S. L. Newsham

in which each subset σ of {1, . . . , n} of order #σ , contributes a minor indexed

by j, k ∈ σ . From the Cauchy determinant formula, we obtain the identity

bj ck e−λj x−λk x / bj cj e−2λj x / λj − λk

det = .
λj + λk j,k∈σ 2λj λj + λk
j ∈σ j,k∈σ :j =k

5 Diagonal Green’s Function and Stationary KdV Hierarchy

In this section, we obtain properties of A in terms of the brackets of odd powers of

A. Thus we obtain some sufficient conditions for some differential equations to be
integrable. Throughout this section, we suppose that the hypotheses of Theorem 4.4
are in force, so that any finite set of elements of A are holomorphic functions on a
some sector  containing (x0 , ∞) for some x0 ≥ 0. We do not generally require u to
be real valued, although in Theorem 5.4(iv) we impose this further condition so that
we can compare our results with the classical spectral theory for the Schrödinger
equation on the real line.
Definition 5.1 (Stationary KdV Hierarchy)
(i) Let f0 = 1 and f1 = (1/2)u. Then the KdV recursion formula is

d d d  d3
4 fm+1 (x) = 4f1 (x) fm (x) + 4 f1 (x)fm (x) − 3 fm (x). (5.1)
dx dx dx dx
(ii) If u satisfies fm = 0 for all m greater than or equal to some m0 , then u satisfies
the stationary KdV hierarchy and is said to be an algebro-geometric (finite
gap) potential; see [10, 11, 13, 29, 35].
(iii) Suppose that u(x) → 0 as x → ∞, and likewise for all the partial derivatives
∂  u/∂x  ; suppose further that fj (x) → 0 as x → 0 as x → ∞ for
all j = 1, 2, . . . . Then we say that the fj are homogeneous solutions of
the KdV hierarchy, and we consider cases where the system of differential
equations (5.1) has no arbitrary constants of integration.
Proposition 5.2 Let A be as in Theorem 4.4]. Then

fm = (−1)m 2!A2m−1 ", m = 1, 2, . . .

satisfies the stationary KdV hierarchy (Novikov’s equations), since

d d3 d 
4 !A2m+3 " = 3 !A2m+1 " + 8 !A" !A2m+1 "
dx dx dx
d 
+ 16!A" !A2m+1 " . (5.2)
dx
Tau Functions Associated with Linear Systems 87

Proof
(i) We have the basic identities

!A(I − 2F )A(I − 2F )X" = !A2 X" − 2!A"!X"; (5.3)

−2A(AF + FA − 2FAF ) = A(I − 2F )A(I − 2F ) − A2 (5.4)

and their mirror images. Hence

d
!A2m+1 " = !A(I − 2F )A2m+1 + A2m+1 (I − 2F )A",
dx
so

d2
!A2m+1 " = !A(I − 2F )A(I − 2F )A2m+1 + 2A(I − 2F )A2m+1 (I − 2F )A
dx 2
+ A2m+1 (I − 2F )A(I − 2F )A

− 2A(AF + AF − 2F AF )A2m+1

− 2A2m+1 (AF + F A − 2F AF )A"

= !A(I − 2F )A(I − 2F )A2m+1 + 2A(I − 2F )A2m+1 (I − 2F )A

+ A2m+1 (I − 2F )A(I − 2F )A

+ A(I − 2F )A(I − 2F )A2m+1 − A2m+3

+ A2m+1 (I − 2F )A(I − 2F )A − A2m+3 "

and by the basic identities (5.3) and (5.4)

d2
!A2m+1 " = 2!A(I − 2F )A2m+1(I − 2F )A" − 2!A2m+3 "
dx 2
+ 2!A(I − 2F )A(I − 2F )A2m+1"
+ 2!A2m+1 (I − 2F )A(I − 2F )A"
= 2!A(I − 2F )A2m+1(I − 2F )A" + 2!A2m+3 "
− 4!A2m+1 "!A" − 4!A"!A2m+1".

Now we differentiate the first summand of the final term

d
2!A(I − 2F )A2m+1 (I − 2F )A"
dx
= 2!A(I − 2F )A(I − 2F )A2m+1 (I − 2F )A"
88 G. Blower and S. L. Newsham

+ 2!A(I − 2F )A2m+1 (I − 2F )A(I − 2F )A"

− 4!A(AF + FA − 2FAF )A2m+1 (I − 2F )A"
− 4!A(I − 2F )A2m+1 (AF + FA − 2FAF )A"

= 2!A(I − 2F )A(I − 2F )A2m+1 (I − 2F )A"

+ 2!A(I − 2F )A2m+1 (I − 2F )A(I − 2F )A"
+ 2!A(I − 2F )A(I − 2F )A2m+1 (I − 2F )A"
− 2!A2m+3 (I − 2F )A"
+ 2!A(I − 2F )A2m+1 (I − 2F )A(I − 2F )A"
− 2!A(I − 2F )A2m+3 "

thus we obtain

d2
!A2m+1 " = 4!A(I − 2F )A(I − 2F )A2m+1 (I − 2F )A"
dx 2
+ 4!A(I − 2F )A2m+1 (I − 2F )A(I − 2F )A"
− 2!A(I − 2F )A2m+3 + A2m+3 (I − 2F )A"
= −8!A"!A2m+1(I − 2F )A" + 4!A2m+3 (I − 2F )A"
− 8!A"!A(I − 2F )A2m+1 "
d
+ 4!A(I − 2F )A2m+3 " − 2 !A2m+3 "
dx
= −8!A"!A(I − 2F )A2m+1 + A2m+1 (I − 2F )A"
d
+ 4!A(I − 2F )A2m+3 + A2m+3 (I − 2F )A" − 2 !A2m+3 "
dx
d d
= −8!A" !A2m+1 " + 2 !A2m+3 ";
dx dx
hence

d3 d d d  
!A 2m+1
" = −8!A" !A 2m+1
" + 4 !A 2m+3
" − 8 !A"!A 2m+1
" ;
dx 3 dx dx dx
which gives the stated result (5.2).
Tau Functions Associated with Linear Systems 89

Definition 5.3 (Diagonal Green’s Function) Let (−A, √ B, C) be as in Theo-

rem 2.3. Then the diagonal Green’s function is g0 (x; ζ )/ ζ where

g0 (x; ζ ) = (1/2) + !A(ζ I − A2 )−1 ". (5.5)

The notation g0 (x; ζ ) is chosen to indicate a generating function and also the
diagonal of a Green’s function; now in Theorem 5.4(iv) we explain the latter
connection. Let C+ = {λ ∈ C : (λ > 0}.
Theorem 5.4 Let (−A, B, C) be as in Theorem 2.3.
(i) Then g0 (x; ζ ) is bounded and continuously differentiable in x and has a unique
asymptotic expansion depending on the bracketed odd powers of A,

1 !A" !A3 " !A5 "

g0 (x; ζ ) $ + + 2 + 3 +... (ζ → −∞); (5.6)
2 ζ ζ ζ

(ii) g0 (x; ζ ) satisfies Drach’s equation

d 3 g0 dg0 du
= 4(u + ζ ) + 2 g0 (x > x0 ; −ζ > ω); (5.7)
dx 3 dx dx
(iii) there exists x1 > 0 such that
    x 
dy
ψ± (x, ζ ) = g0 (x, −ζ ) exp ∓ −ζ (5.8)
x1 2g0 (y; −ζ )

satisfies Schrödinger’s equation

−ψ± (x; ζ ) + u(x)ψ± (x, ζ ) = ζ ψ± (x; ζ ) (x > x1 , ζ > ω).

(iv) Suppose that u is a continuous real function that is bounded below, and that ψ±
from (iii) satisfy ψ+ (x; ζ ) ∈ L2 ((0, ∞); C) and ψ− (x; ζ ) ∈ L2 ((−∞, 0); C)
d2
for all ζ ∈ C+ . Then L = − dx 2 + u(x) defines an essentially self-adjoint

operator in L (R; C), and the Green’s function G(x, y; ζ ) which represents
2

(ζ I − L)−1 has a diagonal that satisfies

g0 (x; −ζ )
G(x, x; ζ ) = √ .
−ζ

Proof
(i) Let π − θ < arg λ < θ , so λ and −λ both lie in Sθ , hence ζ = λ2 satisfies
2π − 2θ < arg ζ < 2θ , so ζ lies close to (−∞, 0). Then ζ I − A2 is invertible
90 G. Blower and S. L. Newsham

and |ζ |(ζ I − A2 )−1 L(H ) ≤ M. The function

1
g0 (x; ζ ) = + Ce−xA (I + Rx )−1 A(ζ I − A2 )−1 (I + Rx )−1 e−xA B (x > 0)
2
is well defined by Theorem 2.3(iii).
To obtain the asymptotic expansion, we note that e−xA (I + Rx )−1 and (I +
Rx )e−xA involve the factor e−xA , where (e−zA ) is a holomorphic semigroup
on Sθ−π/2 . Hence A2j +1 e−xA ∈ L(H ) and by Cauchy’s estimates there exist
δ, x0 , M0 > 0 such that A2j +1 e−xA L(H ) ≤ M0 (2j + 1)! for all x ≥ x0 > 0,
and e−sA L(H ) ≤ M0 e−sδ . As in Proposition 3.2, we have an asymptotic
expansion of

e−zA (λI − A)−1 − (λI + A)−1
 ∞  ∞
= −e−zA eλs e−sA ds − e−zA e−λs e−sA ds
0 0
A A3 A2j −1 
= e−zA + + ···+
λ2 λ4 λ2j

e−zA ∞
+ A2j +1 e−sA (esλ − e−λs ) ds,
λ2j +1 0

in which all the summands are in L(H ) due to the factor e−zA for z ∈ Sθ−π/2 .
Hence
 ∞
Ce−xA (I + Rx )−1 A2j +1 e−sA (esλ − e−sλ ) ds(I + Rx )−1 e−xA B → 0 (x > 0)
0

as λ → i∞, or equivalently ζ → −∞, so

A A3 A2j −1 
g0 (x, ζ ) = Ce−xA (I + Rx )−1 + + · · · + (I + Rx )−1 e−xA B
ζ ζ2 ζj
1  1 
+ + O j +1 .
2 ζ
This gives the asymptotic series; generally, the series is not convergent since
the implied constants in the term O(ζ −(j +1)) involve (2j + 1)!.
(ii) From Proposition 5.2 we have
∞
d  !A2m+3 "
∞
d 3  !A2m+1 " d 
∞
!A2m+1 "
4 m+1
= 3 m+1
+ 8 !A"
dx ζ dx ζ dx ζ m+1
m=0 m=0 m=0
∞

d !A2m+1 "
+ 16!A" ;
dx ζ m+1
m=0

the required result follows on rearranging.

Tau Functions Associated with Linear Systems 91

Conversely, suppose that g0 as defined in (5.5) has an asymptotic expansion

with coefficients in C ∞ ((0, ∞); C) as ζ → −∞ and that g0 (x; ζ ) satis-
fies (5.7). Then the coefficients of ζ −j satisfy a recurrence relation which is
equivalent to the systems of differential equations (5.1).
The asymptotic expansion is unique in the following sense. Suppose
momentarily that t → !Ae−t A " is bounded and repeatedly differentiable on
2

(0, ∞), with M, ω > 0 such that |!Ae−t A "| ≤ Meωt for t > 0, and that there
2

is a Maclaurin expansion

!A5 "t 2
!Ae−t A " = !A" − !A3 "t +
2
− · · · + O(t k )
2!
on some neighbourhood of 0+. Then by Watson’s Lemma [36, p. 188], the
integral
 ∞
!Ae−t A "et ζ dt
2

has an asymptotic expansion as ζ → −∞, where the coefficients give the

formula (5.6).
(iii) Since (e−t A )t >0 is a contraction semigroup on H , we have D(A2 ) ⊆ D(A)
and Af 2H ≤ 2A2 f H f H for all f ∈ D(A2 ) √ by the Hardy-Littlewood-
Landau inequality [15, p. 65], so ζf + A2 f H ≥ ζ Af H for ζ > 0. We
deduce that A2 − 2A + ζ I is invertible for ζ > 9 and generally for all ζ ∈ C
such that #ζ is sufficiently large. By Proposition 5.2 and the multiplicative
property of the bracket, we have

1 + ,
= 1 + 2A(ζ I + A2 − 2A)−1 ,
2g0 (x; −ζ )

and we observe that g0 (x; −ζ ) → 1/2 as x → ∞, so there exists x1 > 0 such

that g0 (x, −ζ ) > 0 for all x > x1 and the differential equation integrates to

d 2 g0 1  dg0 2 ζ
g0 − = 2(u − ζ )g02 + . (5.9)
dx 2 2 dx 2

So one can define ψ(x; ζ ) as in (5.8), and then one verifies the differential
equation for ψ(x; ζ ) by using (5.9).
(iv) By a theorem of Weyl [18, 10.1.4], L is of limit point type at ±∞, and
there exist nontrivial solutions ψ± (x; ζ ) to −ψ± (x; ζ ) + u(x)ψ± (x; ζ ) =
ζ ψ± (x; ζ ) such that ψ+ (x; ζ ) ∈ L2 (0, ∞) and ψ− (x; ζ ) ∈ L2 (−∞, 0),
and these are unique up to constant multiples. Also the inverse operator
(−ζ I + L)−1 may be represented as an integral operator in L2 (R; C) with
92 G. Blower and S. L. Newsham

kernel G(x, y; ζ ), which has diagonal

ψ+ (x; ζ )ψ− (x; ζ )

G(x, x; ζ ) = ((ζ > 0).
Wr(ψ+ ( ; ζ ), ψ− ( ; ζ ))

√ ψ+ (x; ζ )ψ− (x; ζ ) = g0 (x; −ζ ) and

Given ψ∓ as in (iii), we can compute
their Wronskian is Wr(ψ+ , ψ− ) = −ζ , hence the result.
Remark 5.5
(i) The importance of the diagonal Green’s function is emphasized in [14].
Gesztesy and Holden [13] obtain an asymptotic expansion of the diagonal
G(x, x; ζ ) which is consistent with Theorem 5.4(i). Under conditions dis-
cussed in Theorem 5.4, we have similar asymptotics as −ζ → ∞.
(ii) Drach observed that one can start with the differential equation (5.7), and
produce the solutions (5.8); see [7]. He showed that Schrödinger’s equation
is integrable by quadratures, if and only if (5.7) can be integrated by quadra-
tures for typical values of ζ , and Brezhnev translated his results into the
modern theory of finite gap integration [7]. Having established integrability
of Schrödinger’s equation by quadratures, one can introduce the hyperelliptic
spectral curve E with g < ∞ and proceed to express the solution in terms of
the Baker–Akhiezer function. Hence one can integrate the equation and express
the solution in terms of the Riemann’s theta function on the Jacobian of E, as
in [3, 10, 13].
(iii) Kotani [21] has introduced the Baker–Akhiezer function and the τ function
via the Weyl m-function for a suitable class of potentials that included multi-
solitons and algebro-geometric potentials. There is a determinant formula for
τ corresponding to (1.5) and (3.1), and the theory develops themes from [35].
(iv) The deformation theory for rational differential equations is discussed in [20].

Acknowledgments GB thanks Henry McKean for helpful conversations. SLN thanks EPSRC for
financially supporting this research. The authors thank the referee for drawing attention to recent
literature.

References

1. T. Aktosun, F. Demontis, C. van der Mee, Exact solutions to the focusing nonlinear Schrödinger
equation. Inverse Prob. 23, 2171–2195 (2007)
2. M.F. Atiyah, I.G. Macdonald, Introduction to Commutative Algebra (Addison Wesley, Read-
ing, 1969)
3. H.F. Baker, Abelian Functions: Abel’s Theorem and the Allied Theory of Theta Functions
(Cambridge University Press, Cambridge, 1995)
4. R. Bhatia, P. Rosenthal, How and why to solve the operator equation AX − XB = Y . Bull.
Lond. Math. Soc. 29, 1–21 (1997)
5. G. Blower, Linear systems and determinantal random point fields. J. Math. Anal. Appl. 335,
311–334 (2009)
Tau Functions Associated with Linear Systems 93

6. G. Blower, On tau functions for orthogonal polynomials and matrix models. J. Phys. A 44,
285202 (2011)
7. Y.V. Brezhnev, What does integrability of finite-gap or soliton potentials mean? Philos. Trans.
R. Soc. Lond. Ser. A Math. Phys. Eng. Sci. 366, 923–945 (2008)
8. F.J. Dyson, Fredholm determinants and inverse scattering problems. Commun. Math. Phys. 47,
171–183 (1976)
9. K.-J. Engel, R. Nagel, One Parameter Semigroups for Linear Evolution Equations (Springer,
New York, 2000)
10. N. Ercolani, H.P. McKean, Geometry of KdV. IV: Abel sums, Jacobi variety and theta function
in the scattering case. Invent. Math. 99, 483–544 (1990)
11. I.M. Gelfand, L.A. Dikii, Integrable nonlinear equations and the Liouville theorem. Funct.
Anal. Appl. 13, 6–15 (1979)
12. I.M. Gelfand, B.M. Levitan, On the determination of a differential equation from its spectral
function. Izvestiya Akad. Nauk SSSR Ser. Mat. 15, 309–360 (1951)
13. F. Gesztesy, H. Holden, Soliton Equations and Their Algebro-Geometric Solutions Volume I:
(1 + 1)-Dimensional Continuous Models (Cambridge University Press, Cambridge, 2003)
14. F. Gesztesy, B. Simon, The Xi function. Acta Math. 176, 49–71 (1996)
15. J.A. Goldstein, Semigroups of Linear Operators and Applications (Oxford University Press,
Oxford, 1985)
16. S. Grudsky, A. Rybkin, On classical solutions of the KdV equation. Proc. Lond. Math. Soc.
121, 354–371 (2020)
17. S. Grudsky, A. Rybkin, Soliton theory and Hankel operators. SIAM J. Math. Anal. 47, 2283–
2323 (2015)
18. E. Hille, Lectures on Ordinary Differential Equations (Addison-Wesley, Reading, 1968)
19. V.G. Kac, Infinite Dimensional Lie Algebras (Cambridge University Press, Cambridge, 1985)
20. V. Katsnelson, D. Volok, Rational solutions of the Schlesinger system and isoprincipal
deformations of rational matrix functions. I. Oper. Theory Adv. Appl. 149, 291–348 (2004)
21. S. Kotani, Construction of KdV flow I. τ -function via Weyl function. Zh. Mat. Fiz. Anal.
Geom. 14, 297–335 (2018)
22. I.M. Krichever, The integration of nonlinear equations by the methods of algebraic geometry.
Funct. Anal. Appl. 11, 12–26 (1977)
23. V.B. Matveev, Darboux transformation and explicit solutions of the Kadomtcev–Petviaschvily
equation, depending upon functional parameters. Lett. Math. Phys. 3, 213–216 (1979)
24. H.P. McKean, Fredholm determinants. Cent. Eur. J. Math. 9, 205–243 (2011)
25. A.V. Megretskii, V.V. Peller, S.R. Treil, The inverse spectral problem for self-adjoint Hankel
operators. Acta Math. 174, 241–309 (1995)
26. T. Miwa, M. Jimbo, E. Date, Solitons: Differential Equations, Symmetries, and Infinite
Dimensional Algebras (Cambridge University Press, Cambridge, 2000)
27. M. Mulase, Cohomological structure in soliton equations and Jacobian varieties. J. Differ.
Geom. 19, 403–430 (1984)
28. N.K. Nikolski, Operators, Functions and Systems: An Easy Reading, vol. 1 (American
Mathematical Society, Providence, 2002)
29. S. Novikov, S.V. Manakov, L.P. Pitaevskii, V.F. Zakharov, Theory of Solitons, the Inverse
Scattering Method (Consultants Bureau, New York and London, 1984)
30. V.V. Peller, Hankel Operators and Their Applications (Springer, New York, 2003)
31. C. Pöppe, The Fredholm determinant method for the KdV equations. Phys. D 13, 137–160
(1984)
32. C. Pöppe, D.H. Sattinger, Fredholm determinants and the τ function for the Kadomtsev–
Petviashvili hierarchy. Publ. Res. Inst. Math. Sci. 24, 505–538 (1988)
33. M. van der Put, M.F. Singer, Galois Theory of Linear Differential Equations (Springer, Berlin,
2003)
34. A. Rybkin, The Hirota τ -function and well-posedness of the KdV equation with an arbitrary
step-like initial profile decaying on the right half line. Nonlinearity 24, 2953–2990 (2011)
94 G. Blower and S. L. Newsham

35. G. Segal, G. Wilson, Loop groups and equations of KdV type. Inst. Hautes Études Sci. Publ.
Math. 61, 5–65 (1985)
36. I.N. Sneddon, The Use of Integral Transforms (McGraw-Hill, New York, 1972)
37. C.A. Tracy, H. Widom, Fredholm determinants, differential equations and matrix models.
Commun. Math. Phys. 163, 33–72 (1994)
38. M. Trubowitz, The inverse problem for periodic potentials. Commun. Pure Appl. Math. 30,
321–337 (1977)
39. T. Zhang, S. Venakides, Periodic limit of inverse scattering. Commun. Pure Appl. Math. 46,
819–865 (1993)
Groups of Orthogonal Matrices All
Orbits of Which Generate Lattices

Albrecht Böttcher

Abstract There are infinitely many finite groups of orthogonal matrices all orbits
of which, including those of irrational vectors, span lattices, that is, discrete additive
subgroups of the underlying Euclidean space. We show that, both up to isomorphism
and up to orthogonal similarity, exactly eight of these groups are irreducible: the
two trivial groups in one dimension, the cyclic groups of orders 3, 4, 6 in two
dimensions, and the quaternion, binary dihedral, binary tetrahedral groups in four
dimensions.

Keywords Matrix groups · Orthogonal matrices · Lattices in Euclidean space ·

Tight frames

Mathematics Subject Classification (2010) Primary 20H15; Secondary 11H06,

15B10, 42C15, 52C07

1 Introduction and Main Results

Let k ≥ 1 and let G be a finite subgroup of O(k), that is, let G be a finite group
of orthogonal k × k matrices. Suppose the order of G is n ≥ 1 and write G =
{G1 , . . . , Gn }. We think of Rk as a column space. For f ∈ Rk , consider the k × n
matrix

F = G1 f G2 f . . . Gn f (1)

A. Böttcher ()
Fakultät für Mathematik, TU Chemnitz, Chemnitz, Germany
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 95

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_6
96 A. Böttcher

whose columns are the orbit of f under G. We denote by $(G, f ) ⊂ Rk the set of
all linear combinations of the columns of F with integer coefficients,

$(G, f ) = spanZ {G1 f, . . . , Gn f } = spanZ {Gf : G ∈ G}

= {F x : x ∈ Zn } = F Zn .

Inspired by Fukshansky et al. [2], we are interested in whether $(G, f ) is a lattice,

that is, a discrete additive subgroup of Rk . Since $(G, f ) is always a subgroup
of Rk , the problem is its discreteness. Clearly, whether $(G, f ) is a lattice or not
depends on both G and f . In contrast to [2], we here embark on the question whether
there are groups G for which $(G, f ) is a lattice independently of f . We call such
groups lattice generating. Thus, G is lattice generating if $(G, f ) is a lattice for
every f ∈ Rk .
Obviously, the two subgroups {1} and {1, −1} of O(1) are lattice generating. The
trivial group {I } ⊂ O(k) is also lattice generating. The interesting cases are k ≥ 2
and n ≥ 2.
For each Gj , we have

Gj $(G, f ) = Gj spanZ {Gf : G ∈ G} = spanZ {Gj Gf : G ∈ G}

= spanZ {Gf : G ∈ G} = $(G, f ).

It follows that if $(G, f ) is a nontrivial lattice for at least one f , then each Gj ∈ G
must leave this lattice invariant. This indicates that we have to search for lattice
generating groups within the crystallographic point groups.
The 10 crystallographic point groups in R2 are the rotation groups C for  =
1, 2, 3, 4, 6 and the dihedral groups D for  = 1, 2, 3, 4, 6. It is not difficult to
check that the 7 groups C1 , C2 , C3 , C4 , C6 , D1 , D2 are lattice generating, whereas
the groups D3 , D4 , D6 are not lattice generating. It can be shown that exactly half
of the 32 crystallographic point groups in R3 are lattice generating. The table lists
the 32 groups (in Schoenflies notation). The 16 groups in boldface are the lattice
generating groups, the other 16 groups are not lattice generating.

C1 , C 2 , C 3 , C 4 , C 6 ,
C1h , C2h , C3h , C4h , C6h ,
C2v , C3v , C4v , C6v ,
D2 ,D3 , D4 , D6 , D2h ,D3h , D4h , D6h , D2d , D3d ,
S2 , S4 , S6 ,
T , Td , Th , O, Oh .

Thus, taking direct sums of these 2 + 7 + 16 = 25 lattice generating groups

we obtain lots of lattice generating groups in every dimension. However, the truly
Groups of Orthogonal Matrices and Lattices 97

interesting groups are the irreducible ones. A finite group G ⊂ O(k) is said to be
irreducible if the members of the group do not share a common invariant subspace
except for {0} and all of Rk . Equivalently, G is irreducible if and only if spanR {Gf :
G ∈ G} = Rk for every f = 0, which in turn is equivalent to the requirement
that the rank of the matrix (1) is k for every f = 0. An irreducible group must in
particular have n ≥ k elements.
The direct sums of the 25 lattice generating groups we found do not yield
irreducible groups. But what about these groups themselves? It turns out that
C3 , C4 , C6 are the only irreducible lattice generating subgroups of O(2) and that
none of the 16 lattice generating subgroups of O(3) is irreducible (because each of
the latter leaves a rotation axis invariant). Tensor products behave better with regard
to irreducibility. Do we obtain more lattice generating groups in this way? Herewith
our first result.
Theorem 1.1 The tensor product of two irreducible and lattice generating groups
acting on at least two-dimensional spaces is never both irreducible and lattice
generating.
Note that the theorem is not true but becomes a triviality if one of the groups
acts on R1 : we have {1} ⊗ G = G and {1, −1} ⊗ G = −G ∪ G. Clearly, −G ∪ G is
irreducible if and only if so is G, and −G ∪ G is lattice generating if and only if G
has this property. Thus, to search for irreducible lattice generating groups we have
to proceed differently.
All unitary irreducible representations of a finite abelian group are in U(1) and
hence all irreducible representations of a finite abelian group in O(k) must have
degree k ≤ 2. Consequently, if k ≥ 3, the O(k) does not contain irreducible
finite abelian groups and thus all the more does not contain irreducible and lattice
generating abelian groups. So we are left with non-abelian groups.
Theorem 1.2 If k ≥ 2 and a subgroup of O(k) is irreducible and lattice generating,
then it is actually contained in SO(k).
In particular, irreducible Coxeter (reflection) groups are never lattice generating.
Theorem 1.3 If k ≥ 3 is odd, then SO(k) does not contain irreducible lattice
generating groups.
Eventually we have the following.
Theorem 1.4 If k ≥ 6 is even, then there are no irreducible and lattice generating
groups in SO(k). If a subgroup of SO(4) is irreducible and lattice generating, then it
is either isomorphic to the quaternion group Q8 of order 8 or to the binary dihedral
group Q12 of order 12 or to the binary tetrahedral group 2T of order 24. The
groups Q8 , Q12 , 2T have faithful irreducible and lattice generating representations
in SO(4).
The quaternion group Q8 may be given by a, b, c : a 2 = b2 = c2 = abc. The
binary dihedral group Q12 also goes under the notations Dic12 , Dic6 , or Dic3 . It has
98 A. Böttcher

the group presentation

a, b, c : a 3 = b2 = c2 = abc.

The binary tetrahedral group 2T is isomorphic to SL(2, 3), that is, to the 2 × 2
matrices over the field F3 with determinant 1. It is also isomorphic to the group of
units in the ring of Hurwitz integers. A group presentation of 2T is

a, b, c : a 2 = b3 = c3 = abc.

We will construct faithful irreducible and lattice generating representations of these

three groups in SO(4) when proving Theorem 1.4.
Two subgroups G and H of SO(k) are said to be orthogonally similar (or simply
to be equivalent) if the there is an orthogonal matrix U such that G → U GU −1
is a bijection of G onto H. From the character tables of Q8 , Q12 , 2T which
can be found in [4, 6] it follows that the faithful irreducible representations of
Q8 and Q12 in SO(4) are all orthogonally similar but that 2T has exactly two
classes of orthogonally similar faithful representations in SO(4), which are called
the quaternionic and the complex representations. We here prove the following.
Theorem 1.5 The quaternionic representations of 2T in SO(4) are lattice generat-
ing whereas the complex representations of 2T in SO(4) are not lattice generating.
Thus, not only up to isomorphism but also up to orthogonal similarity, there exist
exactly eight irreducible and lattice generating groups.
The results on groups we will use in the following are all well known and can be
found in the basic literature. All we use from representation theory is some general
facts that are in the classic [7] or in the recent text [3], for example, and two character
tables of real representations which are explicitly given in [4, 6].

2 Proofs

We denote by (·, ·) and · the usual scalar product and norm in the Euclidean space
Rk . The transpose of a matrix A is denoted by A .
Assertion (a) of the following theorem is Corollary 10.5 of [8]. The other
assertions of the theorem were established in [1] for k ∈ {2, 3} and subsequently
in [2] for general k ≥ 2. For the reader’s convenience, we include the proof from [2].
Theorem 2.1 Suppose n ≥ k ≥ 2. Let G ⊂ O(k) be a finite irreducible group of
order n, let f ∈ Rk \ {0}, and let F be the k × n matrix (1). Then the following
hold.
(a) The matrix F F is a nonzero scalar multiple of the identity matrix,

F F = γ I.
Groups of Orthogonal Matrices and Lattices 99

(b) The set $(G, f ) is a lattice if and only if the Gram matrix

F F = ((Gj f, Gk f ))nj,k=1

is a scalar multiple of a rational matrix.

(c) In case $(G, f ) is a lattice, we actually have F F ∈ γ Qn×n .
(d) If f  = 1, then $(G, f ) is a lattice if and only if F F is a rational matrix.
Proof
(a) Corollary 10.5 of [8] states that if G is a finite irreducible subgroup of O(k),
then the matrix F is a tight frame for every f ∈ Rk \ {0}, which is equivalent to
saying that F F = γ I with some nonzero real γ .
(b) Suppose μF F ∈ Qn×n for some nonzero μ ∈ R. Then μdF F ∈ Zn×n for
some positive integer d. It follows that

μ2 d 2 F x2 = μd(μdF F x, x) ∈ μdZ

for all x ∈ Zn . Hence, if x ∈ Zn , then either F x = 0 of F x2 ≥ 1/(μd). This

proves that $(G, f ) = F Zn is discrete and thus a lattice.
Conversely, suppose $(G, f ) is a lattice. We know from (a) that F F = γ I .
√
This implies that (1/ γ )F is built by the first k rows of an orthogonal matrix,
and hence the rank of F is k. Thus, the columns of F span all of Rk . Let B ∈
Rk×k be a basis matrix for $(G, f ), that is, B is invertible and $(G, f ) =
{Bx : x ∈ Zk }. It follows that there is a matrix Z ∈ Zk×n such that F = BZ.
We obtain that BZZ B = F F = γ I and hence ZZ = γ B −1 (B )−1 =
γ (B B)−1 . Consequently, F F = Z B BZ = Z (γ (ZZ )−1 )Z ∈ γ Qn×n .
(c) We just showed that if $(G, f ) is a lattice, then F F ∈ γ Qn×n .
(d) If f has norm 1, then 1 = (f, f ) is an entry of F F ∈ γ Qn×n , which implies
that γ must be rational. Thus, F F ∈ Qn×n .
Every orthogonal matrix G is of the form G = U SU with an orthogonal matrix
U and a real block-diagonal matrix S. The blocks of S are either 2 × 2 matrices of
the form
 
α −β
with α 2 + β 2 = 1, β > 0 (2)
β α

or the 1 × 1 matrices (±1). Note that since

     
α −β 01 αβ 01
= ,
β α 10 −β α 10

the restriction β > 0 can always be achieved by changing U . We say that G has
equal blocks if all blocks of S are equal to each other. The trivial cases are G =
U I U = I and G = U (−I )U = −I .
100 A. Böttcher

Lemma 2.2 If G is an irreducible and lattice generating subgroup of O(k), then

each matrix in G has equal blocks.
Proof Let G = {G1 , . . . , Gn }. From Theorem 2.1(c) we infer that

(Gj f, Gk f ) ∈ γf Q

with some γf = 0 for every f and all j, k. Taking Gk = I , we arrive at the

conclusion that (Gf, f ) ∈ γf Q for every G ∈ G and every f . It follows that
(Gf, f )/(f, f ) ∈ Q for all f = 0. Let now G = U SU . Since

(Gf, f ) (U SU f, f ) (SU f, U f )
= = ,
(f, f ) (U U f, f ) (U f, U f )

and every h = 0 is of the form U f with f = 0, we conclude that (Sh, h)/(h, h)

must be rational for every h = 0.
Thus, if G = U SU , then (Sh, h) must be rational for every h ∈ Rk of norm 1.
Suppose S has two 2 × 2 blocks
   
α −β γ −δ
B= , C= .
β α δ γ

We may without loss of generality assume that these are the first two blocks of S,
that is, S = diag(B, C, . . .). We take h = (f1 , f2 , g1 , g2 , 0, . . . , 0) ∈ Rk of norm
1 and put f = (f1 , f2 ) , g = (g1 , g2 ) . Thus, h2 = f 2 + g2 = 1. Then

(Sh, h) = (Bf, f ) + (Cg, g) = α(f12 + f22 ) + γ (g12 + g22 )

= αf 2 + γ g2 = γ + (α − γ )f 2 ,

and if α = γ , we may clearly choose f so that γ + (α − γ )f 2 is irrational.

Thus, we must have α = γ . Since β 2 = δ 2 = 1 − α 2 and β, δ > 0, it follows that
β = δ. If S has only one 2 × 2 block, S = diag(B, ±1, . . .), we get analogously
with h = (f1 , f2 , g, 0, . . .) and f12 + f22 + g 2 = 1 that

(Sh, h) = (Bf, f ) ± g 2 = α(f12 + f22 ) ± g 2 = α − (α ∓ 1)g 2 .

Since −1 < α < 1, there are g such that this is irrational. Thus, the presence of
one 1 × 1 block forces all blocks to be 1 × 1. Finally, if S has two different 1 × 1
blocks, S = diag(1, −1, . . .), then h = (f, g, 0, . . .) with f 2 + g 2 = 1 gives
(Sh, h) = f 2 − g 2 = 2f 2 − 1, which may also be irrational.
Proof of Theorems 1.2 and 1.3 Let G ⊂ O(k) be irreducible and lattice generating.
If det G = −1 for some G = U SU ∈ G, then S must contain a 1 × 1 block (−1),
so Lemma 2.2 implies that G = −I , and thus G = {I, −I }. But this group is not
Groups of Orthogonal Matrices and Lattices 101

irreducible. If k is an odd number, then S must have a 1 × 1 block (±1) and we

arrive again at the reducible groups G = {I } or G = {I, −I }.
If G ⊂ O(k) is a finite group, then each element G ∈ G is of finite order (as an
element of the group). If G = U SU has equal blocks, then G is of finite order if and
only if G ∈ {I, −I } or if the block is (2) with α ∈ {−1/2, 0, 1/2} (Niven’s theorem;
see [5, pp. 37–41]). In these cases the order of G is 1, 2, 3, 4, 6, respectively. Let us
denote the blocks with α = −1/2, 0, 1/2 by !3 , !4 , !6 , respectively.
Corollary 2.3 Let G ⊂ O(k) be irreducible and lattice generating. If −I ∈
/ G, then

G = {I, B1 , . . . , Bq } (3)

where each Bj is of order 3. If −I ∈ G, then

G = {I, −I, A1 , . . . , As , −A1 , . . . , −As , B1 , . . . , Bq , −B1 , . . . , −Bq } (4)

(s, q ≥ 0, s + q > 0) where the ±Aj are of order 4 and satisfy (±Aj )2 = −I and
−Aj = A3j , each Bj is of order 6 and satisfies Bj3 = −I , and each −Bj is of order
3 and satisfies −Bj = Bj4 .
Proof We know from Lemma 2.2 that each G ∈ G \{I } has equal blocks. This block
is (−1), !6 , !4 , or !3 . If −I ∈
/ G, then the block cannot be !6 or !4 , since then
G3 or G2 would be −I . Thus, in this case G is of the form (3) in which the matrices
Bj have the block !3 . Now suppose −I ∈ G and let G ∈ G \ {I, −I }. If the block
of G is !4 , then G is of order 4 with G2 = −I and G3 = −G, which gives us the
elements ±Aj in (4). If the block is !6 , then G is of order 6 and we have G3 = −I
and G4 = −G. These elements G are the ±Bj in (4). Matrices with the block !3
are of the order 3. The map G → −G is a bijection of every group containing I and
−I and it changes the orders 4 and 3 to the orders 4 and 6, respectively. Thus, the
number r of elements of order 3 is equal to the number q of elements of the order
6, and since the elements −B1 , . . . , −Bq have the order 3, it follows that G must be
among these elements.
Proof of Theorem 1.1 By virtue of Theorem 1.3 we may suppose that the groups
are subgroups of O(2m) and O(2k) with m, k ≥ 1. Lemma 2.2 shows that the trace
of an element G of an irreducible and lattice generating subgroup of O(2) is 
if the order of G is 6, equals 0 if G has the order 4, and is − if G is of order
3. Let G ⊂ O(2m) and H ⊂ O(2k) be irreducible and lattice generating groups.
Assume G ⊗ H is an irreducible and lattice generating subgroup of O(4mk). We
have tr (G⊗H ) = (tr G)(tr H ) for G ∈ G and H ∈ H. Thus if both G and H contain
elements G, H of orders 3 or 6, then tr (G ⊗ H ) = ±mk whereas the possible traces
of elements of G ⊗ H are ±4mk (for I, −I ), ±2mk, and 0. It follows that at least
one of the groups, say G, has no elements of the orders 3 and 6. By Corollary 2.3, G
contains a G of order 4 such that G2 = −I . If in the other group there is an element
H of order 3 or 6, then G ⊗ H has the order 12, which is impossible. Consequently,
102 A. Böttcher

again by Corollary 2.3, the other group has an H of order 4 satisfying H 2 = −I .

But then G ⊗ H satisfies (G ⊗ H )2 = (−I ) ⊗ (−I ) = I ⊗ I , which shows that
G ⊗ H has two elements of order 2. However, Corollary 2.3 tells us that this cannot
happen.
Lemma 2.4 An irreducible lattice generating group cannot contain a subgroup
isomorphic to Z23 .
Proof Let G be irreducible and lattice generating and H be a subgroup isomorphic
to Z23 . Let X be a nontrivial invariant subspace of H. Since all matrices in H are
orthogonal, X⊥ is also a nontrivial invariant subspace of H. Choose orthonormal
bases in X and X⊥ , take their union, and represent the matrices in H in this new
bases. The new matrices are composed by two diagonal blocks (of the same sizes for
all members of H), are again orthogonal and result from the original matrices by a
transformation H → W H W with an orthogonal matrix W . Such a transformation
leads to an isomorphic group, and it does not violate irreducibility or the property
of being lattice generating. By repeating this construction until all blocks are
irreducible, we eventually have an orthogonal matrix W and numbers mj ≥ 1 such
that the matrices in W HW are all of the form diag(H1 , . . . , H ) with Hj ∈ O(mj ).
The map H → O(mj ) given by H → Hj is a group homomorphism. Hence the
image of this map is a subgroup Mj of O(mj ). This subgroup is irreducible, lattice
generating, and isomorphic to a subgroup of Z23 .
Since all Mj are abelian, we conclude that mj ≤ 2 for all j . Now the fact that G
is both irreducible and lattice generating comes into play. Lemma 2.2 implies that all
matrices in G, and in particular those in H, must have equal blocks. Consequently,
if mj = 1 for one j , then mj = 1 for all j and it follows that W HW ⊂ {I, −I },
which is impossible because W HW is isomorphic to Z23 . Hence mj = 2 for all j ,
and for each H ∈ H the blocks of Hj are !3 . Let A and B be two generators of H.
We so have

W AW = diag(Uj !3 Uj ), W BW = diag(Vj !3 Vj ).

Put U = diag(Uj ). Then

A := U W AW U = diag(!3 ) =: diag(Aj ),
B := U W BW U = diag(Zj !3 Zj ) =: diag(Bj )

with Zj ∈ O(2). But if Z ∈ O(2), then Z!3 Z = !3 if det Z = 1 and Z!3 Z = !23
if det Z = −1. It follows that each block Bj is !3 or !23 . If Bj = !3 , then the j th
block of U W ABW U is Aj Bj = !3 ·!3 = !23 , while if Bj = !23 , the j th block of
U W ABW U is Aj Bj = !3 ·!23 = I . Thus, as the number of blocks is at least two,
AB does not have equal blocks if B contains two different blocks. Therefore either
B = diag(!3 ) or B = diag(!23 ). We arrive at the conclusion that H is isomorphic
Groups of Orthogonal Matrices and Lattices 103

to the group generated by !3 alone, that is, to Z3 . Since H is isomorphic to Z23 , this
is a contradiction.
The following lemma proves part of Theorem 1.4.
Lemma 2.5 If k ≥ 4 is even and G ⊂ SO(k) is an irreducible and lattice generating
group, the G is either isomorphic to the quaternion group Q8 of order 8 or to the
binary dihedral group Q12 of order 12 or to the binary tetrahedral group 2T of
order 24.
Proof If G is a finite group and a prime number p divides the order of G, then G
contains an element of order p (Cauchy’s theorem). By Corollary 2.3, the orders of
the elements of our group are 3, 4, 6, and hence the order of G must be n = 2r 3s . The
Sylow theorems imply that if G is a finite group and p is a prime power dividing
the order of G, then G has at least one subgroup of order p . Thus, if r ≥ 4, then
G contains a subgroup of order 16. The groups Q16 and Z16 have an element of
order 8 and the other 8 groups of order 16 all have at least two elements of order
2. By Corollary 2.3, this is impossible. If s ≥ 2, the G contains a subgroup of
order 9. Lemma 2.4 shows that this subgroup cannot be isomorphic to Z23 , and it is
also impossible that it is isomorphic to Z9 , which contains elements of order 9. The
remaining possible orders are {2, 3, 4, 6, 8, 12, 24}. In what follows we permanently
employ Corollary 2.3 without mentioning this each time.
n = 2, 3, 4 These groups are abelian and hence not irreducible.
n = 6 C6 is abelian and the symmetric group S3 has 3 elements of order 2.
n = 8 C8 , C4 × C2 , C23 are abelian and D4 has more than one element of order
2. The only group remaining is Q8 .
n = 12 We could rule out the abelian groups immediately, but with the case n =
24 in mind, we argue as follows: C12 contains an element of order 12 and C3 ×C22
has 3 elements of order 2. The three non-abelian groups are A4 , the dihedral
group D6 , and the binary dihedral group Q12 . The first two of them have more
than one element of order 2, so that only Q12 is left.
n = 24 We may exclude the three abelian groups. There are 12 non-abelian
groups. Except for 2T , each of the remaining 11 groups contains a subgroup
of order 12. From our arguments to settle the case n = 12 we know that, this
time with the exception of Q12 , each of these subgroups has an element of order
12 or more than one element of order 2. The only of the 11 groups having Q12
as a subgroup is Q12 × C2 . But this group has at least three elements of order 2.
Thus, 2T is the only possible group.
Lemma 2.6 If k ≥ 6 is even and a group G ⊂ O(k) is isomorphic to Q8 , Q12 , or
2T , then G is reducible.
Proof The degrees of the irreducible representations over C of the three groups
Q8 , Q12 , 2T are

Q8 : 1, 1, 1, 1, 2, Q12 : 1, 1, 1, 1, 2, 2, 2T : 1, 1, 1, 2, 2, 2, 3, (5)
104 A. Böttcher

and degrees of the irreducible representations over R of these groups are

Q8 : 1, 1, 1, 1, 4, Q12 : 1, 1, 2, 2, 4 2T : 1, 2, 3, 4, 4. (6)

The assertion of the lemma is therefore immediate from (6). The lists (5) and the
list for Q8 in (6) are well known. The lists for Q12 and 2T in (6) are explicitly on
the website [4] and in the lecture notes [6]. An alternative proof of the lemma based
solely on (5) and the list for Q8 in (6) is as follows.
If, for a finite group, the maximal degree of an irreducible representation
over C is , then the maximal degree k of an irreducible representation over
R, that is, in O(k), satisfies k ≤ 2. It follows that Q8 and Q12 have no
irreducible representations in O(k) for even k ≥ 6 and that 2T has no irreducible
representations in O(k) for even k ≥ 8. So consider 2T ⊂ O(6).
Let H be a subgroup of O(6) which is isomorphic to Q8 . We claim that
this group has a 5-dimensional invariant subspace. The degrees of the irreducible
representations of Q8 in O(k) are 1, 1, 1, 1, 4. Thus, H is reducible in dimensions
different from 1 and 4. Taking this into account, we may proceed as in the proof of
Lemma 2.4 to get an orthogonal matrix W such that each element of W HW is of the
form diag(H1 , . . . , H ) with Hj ∈ O(mj ) and mj ∈ {1, 4} for all j . Accordingly,

R6 = X1 ⊕ · · · ⊕ X , (7)

where each Xj is an mj -dimensional invariant subspace of W HW . Since m1 +

· · · + m = 6, we necessarily have mj = 1 for some j . Replacing the Xj on the
right of (7) by {0}, we get a 5-dimensional invariant subspace V of W HW , and
hence W V is a 5-dimensional subspace of H.
Suppose now G ⊂ O(6) is isomorphic to 2T . It is well known that 2T ∼
= Q8 Z3 ,
j
that is, every G ∈ G may (uniquely) be written as G = H C3 with j ∈ {0, 1, 2}, H
in a subgroup H isomorphic to Q8 , and an element C3 ∈ G of order 3. From what
was just proved we conclude that the subgroup H has a 5-dimensional invariant
subspace U := W V ⊂ R6 . The image space C3 U ⊂ R6 also has the dimension 5.
Consequently, dim(U ∩ C3 U ) ≥ 4. As C32 U is of dimension 5 as well, it follows
that U ∩ C3 U ∩ C32 U is of dimension at least 3. Take u = C3 v = C32 w = 0 from
U ∩ C3 U ∩ C32 U . For every H ∈ H we then have H u ∈ U , H C3 u = H w ∈ U , and
H C32 u = H v ∈ U . Consequently, the orbit of u does not span R6 , which proves
that G ∼= 2T cannot be irreducible in O(6).
We finally prove the second half of Theorem 1.4.
Lemma 2.7 The groups Q8 , Q12 , 2T have faithful irreducible and lattice generat-
ing representations in SO(4).
Groups of Orthogonal Matrices and Lattices 105

Proof The quaternion group Q8 is faithfully represented in SO(4) by

       
10 0 −1 0i −i 0
, , , ,
01 1 0 i0 0i
       
−1 0 01 0 −i i 0
, , ,
0 −1 −1 0 −i 0 0 −i

with
   
10 0 −1
1= , i= .
01 1 0

This group contains I, −I and six elements of order 4. This is the case s = 3 and
q = 0 in (4). For f = (a, b, c, d) ∈ R4 \ {0} the matrix

F = G1 f G2 f . . . G8 f

is
⎛ ⎞
a −c −d b −a c d −b
⎜ b −d c −a −b d −c a ⎟
F =⎜ ⎟
⎝ c a −b −d −c −a b d ⎠ .
d b a c −d −b −a −c

Multiplying F from the right by ξ = (x, y, z, w, p, q, r, s) ∈ Z8 and putting

X = x − p, Y = y − q, Z = z − r, W = w − s,

we get after some elementary computations that

⎛ ⎞⎛ ⎞
a −c −d b X
⎜ b −d c −a ⎟ ⎜ Y ⎟
Fξ = ⎜ ⎟⎜ ⎟
⎝ c a −b −d ⎠ ⎝ Z ⎠ =: Av.
d b a c W
√
The matrix A is a 2 + b 2 + c2 + d 2 times an orthogonal matrix and therefore
invertible. Thus,

F ξ 2 = Av2 ≥ mv2

with some m > 0, and since mv2 does not assume values in (0, m), it follows that
F Z8 is a lattice. Finally, the span of {Gf : G ∈ G} is the span of the columns of F .
106 A. Böttcher

The first four columns form the matrix A, and since this matrix is invertible, F has
the rank 4, which shows that the columns of F span all of R4 . Thus, G is irreducible.
The group Q12 ⊂ SO(4) is faithfully represented by the 12 matrices
   
10 −1 0
I = , −I = ,
01 0 −1
   
0 −1 01
A1 = , −A1 = ,
1 0 −1 0
   
0 −ω 0ω
A2 = , −A2 = ,
−ω2 0 ω2 0
   
0 −ω2 0 ω2
A3 = , −A3 = ,
−ω 0 ω 0
   2 
ω 0 ω 0
B = , B2 = ,
0 −ω2 0 −ω
   
−ω 0 −ω2 0
B4 = , B5 = ,
0 ω2 0ω

where
    √
10 α −β 1 3
1= , ω= , α= , β=− .
01 β α 2 2

This group contains I, −I , six elements of order 4, two elements of order 6 (B and
B 5 ), and two elements of order 3 (B 2 and B 4 ), that is, we have the situation s = 3
and q = 2 in (4).
That Q12 is irreducible and lattice generating can be proved as for Q8 . Given
f = (a, b, c, d) ∈ R4 \ {0}, let F be the matrix

F = G1 f G2 f . . . G12 f ,

where the ordering G1 , . . . , G12 is

I, A1 , A2 , A3 , −A1 , −A2 , −A3 , B, B 2 , B 3 = −I, B 4 , B 5 .

The first four columns of F are a scalar multiple of an orthogonal matrix, which
implies that F has rank 4 and hence that Q12 is irreducible. If

ξ = (g, x, y, z, p, q, r, s, t, u, v, w) ∈ Z12 ,
Groups of Orthogonal Matrices and Lattices 107

then, again after elementary calculations,

⎛ ⎞⎛ ⎞
a c d b X
⎜ b d −c −a ⎟ ⎜ Y ⎟
Fξ = ⎜ ⎟⎜ ⎟
⎝ c −a b −d ⎠ ⎝ Z ⎠ =: CV
d −b −a c W

with

X = g − u + α(s + w − t − v), Y = p − x + α(z + q − y − r),

Z = β(y + z − q − r), W = β(v + w − s − t).

The matrix C is a scalar multiple of an orthogonal matrix. It follows that

F ξ 2 = CV 2 ≥ mV 2

with some m > 0. But

V 2 =X2 + Y 2 + Z 2 + W 2
3 3
= (q + r − y − z)2 + (s + t − v − w)2
4 4
 2  2
1 1
+ g − u + (s + w − t − v) + p − x + (z + q − y − r)
2 2

is either 0 or at least 1/4, which shows that F Z12 is discrete and thus a lattice.
The group 2T is faithfully represented in SO(4) by the eight matrices
   
10 0 −1
I= , A1 = , −I, −A1 ,
01 1 0
   
0i −i 0
A2 = , A3 = , −A2 , −A3 ,
i0 0i

with
   
10 0 −1
1= , i=
01 1 0

and by the sixteen matrices

1
(±I ± A1 ± A2 ± A3 ).
2
108 A. Böttcher

We let B1 , . . . , B8 denote the latter matrices with I having the + sign and ordered
according the lexicographic order +++, ++−, +−+, +−−, . . . of the remaining
signs. The other eight matrices then are −B1 , . . . , −B8 . With these notations we are
in the situation of Corollary 2.3 with s = 3 and q = 8. The rest of the proof is as
for Q8 and Q12 . Given f = (a, b, c, d) ∈ R4 \ {0}, let F be the matrix

F = G1 f G2 f . . . G24 f

where G1 , . . . , G24 are the matrices

I, A1 , A2 , A3 , B1 , . . . , B8 , −I, −A1 , −A2 , −A3 , −B1 , . . . , −B8

in this order. The first four columns of F are the same as the first four columns we
had for Q8 . Thus, they form scalar multiple of an orthogonal matrix, which implies
that F has rank 4 and hence that G is irreducible. Let ξ ∈ Z24 be the vector

ξ =(g1 , x1 , y1 , z1 , p1 , q1 , r1 , s1 , t1 , u1 , v1 , w1 ,
g2 , x2 , y2 , z2 , p2 , q2 , r2 , s2 , t2 , u2 , v2 , w2 )

and put

g = g1 − g2 , x = x1 − x2 , . . . , w = w1 − w2 .

After elementary calculations we arrive at the equality

⎛ ⎞⎛ ⎞
a c d b X
⎜ b d −c −a ⎟ ⎜ Y ⎟
2F ξ = ⎜ ⎟⎜ ⎟
⎝ c −a b −d ⎠ ⎝ Z ⎠ =: CV
d −b −a c W

with

X = 2g + p + q + r + s + t + u + v + w,
Y = −2x − p − q − r − s + t + u + v + w,
Z = −2y − p − q + r + s − t − u + v + w,
W = 2z + p − q + r − s + t − u + v − w.

The matrix C is a scalar multiple of an orthogonal matrix. It follows that

4F ξ 2 = CV 2 ≥ mV 2

Groups of Orthogonal Matrices and Lattices 109

with some m > 0 and since V 2 is either 0 or at least 1, we see that F Z24 is
discrete and thus a lattice.
As said, Lemmas 2.5, 2.6, 2.7 prove Theorem 1.4.
Proof of Theorem 1.5 The representation of 2T we used in the proof of Theo-
rem 1.4 is the quaternionic representation. The complex representation over R can
be constructed as follows. Put
       
10 0 −1 0 i −i 0
E= , I = , J = , K=
01 1 0 i0 0 i

and let B1 , . . . , B8 be the matrices (1/2)(E ± I ± J ± K) listed according the

lexicographic order of + and −. Thus,

1 1 1
B1 = (E + I + J + K), B2 = (E + I + J − K), B2 = (E + I − J + K), . . . .
2 2 2
Then the 24 matrices ±E, ±I, ±J, ±K, ±B1 , . . . , ±B8 are a faithful irreducible
representation of 2T in U(2). We denote this matrix group simply by 2T . The seven
conjugacy classes are {E}, {−E}, {±I, ±J, ±K},

C1 := {B1 , B4 , B6 , B7 }, C2 := {B2 , B3 , B5 , B8 },
√
and −C1 , −C2 . Let ω = −1/2 + i 3/2 and let  : 2T → C be the homomorphism
sending ±E, ±I, ±J, ±K to 1, the elements of −C1 ∪ C1 to ω, and the elements
of −C2 ∪ C2 to ω2 . This is a representation of 2T in U(1). It follows that the map
τ : 2T → U(2), G → (G)G is a faithful representation of 2T . The representing
matrices are complex of the form A + iB. Replacing each such matrix by
 
A −B
,
B A

we get a faithful representation σ : 2T → O(4). This is what is called a complex

representation of 2T over R. The matrices σ (E), σ (I ), σ (J ), σ (K) are
   
E 0 I 0
σ (E) = , σ (I ) = ,
0E 0I
   
0 −W 0 −S
σ (J ) = , σ (K) =
W 0 S 0

with
   
01 −1 0
W = , S= .
10 01
110 A. Böttcher

The matrix σ (B1 ) equals

⎛ ⎞
δ −δ −γ γ √ √
⎜ −γ −γ −δ −δ ⎟ 3+1 3−1
σ (B1 ) = ⎜ ⎟
⎝ γ −γ δ −δ ⎠ , γ = , δ= .
4 4
δ δ −γ −γ

Now consider the matrix


F = σ (E)f σ (I )f σ (J )f σ (K)f σ (B1 )f . . .

with f = (a, b, c, d) ∈ R4 . (Incidentally, again the first columns of F form

an orthogonal matrix, which shows that σ (2T ) is irreducible.) Taking f equal to
(1, 0, 0, 0) we get
⎛ ⎞
1 00 0 δ ...
⎜0 1 0 0 −γ ...⎟
F =⎜
⎝0
⎟.
0 0 −1 γ ...⎠
0 01 0 δ ...

Multiplying F from the right by ξ = (x, y, z, w, p, 0, . . . , 0) ∈ Z24 we obtain

F ξ = (x + δp, y − γp, −w + γp, z + δp) .

By Kronecker’s theorem, μZ modulo 1 is dense in (0, 1) whenever μ is irrational.

Consequently, given m ∈ N, there are

ξm = (xm , ym , zm , wm , pm , 0, . . . , 0) ∈ Z24

such that F ξm  ∈ (0, 1/m), which shows that F Z24 is not discrete and proves that
σ (2T ) is not lattice generating.

3 Groups of Unitary Matrices

The topic is trivial for unitary matrices, that is, for finite subgroups of U(k). A
group G ⊂ U(k) of order n is said to be lattice generating if, with F given by (1),
$(G, f ) = F Z[i]n is a discrete additive subgroup of Ck for every f ∈ Ck .
Theorem 3.1 The only irreducible and lattice generating subgroups of U(k) are
G = {1, i, −1, −i} ⊂ U(1) and its subgroups {1} and {1, −1}.
Proof Let G ⊂ U(k) be an irreducible and lattice generating group. In the complex
case, Theorem 2.1 reads as follows. Suppose n ≥ k ≥ 1. Let G ⊂ U(k) be a finite
Groups of Orthogonal Matrices and Lattices 111

irreducible group of order n, let f ∈ Ck \ {0}, and let F be the k × n matrix (1).
Then the following hold.
(a) We have F F ∗ = γ I with some real γ > 0.
(b) The set $(G, f ) is a lattice if and only if the Gram matrix F ∗ F =
((Gj f, Gk f ))nj,k=1 belongs to μQ[i]n×n for some nonzero μ ∈ C.
(c) In case $(G, f ) is a lattice, we actually have F ∗ F ∈ γ Q[i]n×n .
(d) If f  = 1, then $(G, f ) is a lattice if and only if F ∗ F ∈ Q[i]n×n .
The reasoning of the proof of Lemma 2.2 shows that if G = U SU ∗ with a diagonal
matrix S, then S must have equal entries, that is, S = ωI with some ω ∈ T.
Consequently, G is composed of scalar multiples of the identity matrix. It follows
that G = {I, ωI, . . . , ωn−1 I } with ω = e2πi/n , which is reducible for k ≥ 2. So let
k = 1. For |f | = 1, the Gram matrix is F ∗ F = (ωk ωj )n−1 j,k=0 . Since ω is an entry
of this matrix, it must be in Q[i]. The only such roots of unity are ω = 1, −1, i.
Conversely, it is clear, that the (irreducible) subgroups of {1, i, −1, −i} are indeed
lattice generating.

Acknowledgments I sincerely thank Lenny Fukshansky, Christian Lehn, Josiah Park, and Dmytro
Shklyarov for stimulating and helpful discussions.

References

1. A. Böttcher, L. Fukshansky, Addendum to “Lattices from equiangular tight frames”. Linear

Algebra Appl. 531, 592–601 (2017)
2. L. Fukshansky, D. Needell, J. Park, Y. Xin, Lattices from tight frames and vertex transitive
graphs. Electron. J. Comb. 26(3), #P3.49, 30 (2019)
3. C. Gruson, V. Serganova, A Journey Through Representation Theory (Springer, Basel, 2018)
4. J. Montaldi, Real Representations. http://www.maths.manchester.ac.uk/~jm/wiki/
Representations/Representations
5. I.M. Niven, Irrational Numbers (Wiley, New York, 1956)
6. M. Reif, in Groups and Representations. Lecture Notes for a Course Held (2018). http://
homepages.warwick.ac.uk/~masda/MA3E1/GrRepns_2018.pdf
7. J.-P. Serre, Linear Representations of Finite Groups (Springer, New York, 1977)
8. S.F.D. Waldron, An Introduction to Finite Tight Frames (Birkhäuser, New York, 2018)
Invertibility Issues for Toeplitz Plus
Hankel Operators and Their Close
Relatives

Victor D. Didenko and Bernd Silbermann

Abstract The paper describes various approaches to the invertibility of Toeplitz

plus Hankel operators in Hardy and l p -spaces, integral and difference Wiener-
Hopf plus Hankel operators and generalized Toeplitz plus Hankel operators. Special
attention is paid to a newly developed method, which allows to establish necessary,
sufficient and also necessary and sufficient conditions of invertibility, one-sided and
generalized invertibility for wide classes of operators and derive efficient formulas
for the corresponding inverses. The work also contains a number of problems whose
solution would be of interest in both theoretical and applied contexts.

Keywords Toeplitz plus Hankel operators · Wiener-Hopf plus Hankel operators ·

Invertibility · Inverses

Mathematics Subject Classification (2010) Primary 47B35, 47B38; Secondary

47B33, 45E10

This work was supported by the Special Project on High-Performance Computing of the National
Key R&D Program of China (Grant No. 2016YFB0200604), the National Natural Science
Foundation of China (Grant No. 11731006) and the Science Challenge Project of China (Grant
No. TZ2018001).

V. D. Didenko ()
Department of Mathematics, SUSTech International Center for Mathematics, Southern University
of Science and Technology, Shenzhen, China
B. Silbermann
Technische Universität Chemnitz, Fakultät für Mathematik, Chemnitz, Germany
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 113

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_7
114 V. D. Didenko and B. Silbermann

1 Introduction

Toeplitz T (a) and Hankel H (b) operators appear in various fields of mathematics,
physics, statistical mechanics and they have been thoroughly studied [7, 42, 47].
Toeplitz plus Hankel operators T (a)+H (b) and Wiener-Hopf plus Hankel operators
W (a)+H (b) play an important role in random matrix theory [1, 6, 30] and scattering
theory [34–36, 43, 45, 46, 56]. Although Fredholm properties and index formulas
for such operators acting on different Banach and Hilbert spaces are often known—
cf. [7, 13, 38–41, 48–52], their invertibility is little studied. So far progress has
been made only in rare special cases. In this work, we want to present an approach,
which allows to treat invertibility problem for a wide classes of Toeplitz plus Hankel
operators on classic Hardy spaces and also for their close relatives: Toeplitz plus
Hankel operators on l p -spaces, generalized Toeplitz plus Hankel operators and
integral and difference Wiener-Hopf plus Hankel operators. The operators acting
on classical Hardy spaces are discussed in more details, whereas for other classes of
operators we only provide a brief overview of the corresponding results.

2 Toeplitz and Hankel Operators on Hardy Spaces

Let T := {t ∈ C : |t| = 1} be the counterclockwise oriented unit circle in the

complex plane C and let p ∈ [1, ∞]. Consider the Hardy spaces

H p = H p (T) := {f ∈ Lp (T) : fn = 0 for all n < 0},

H p = H p (T) := {f ∈ Lp (T) : fn = 0 for all n > 0},

where fn , n ∈ N are the Fourier coefficients of function f . Moreover, let I denote
the identity operator, P : Lp (T) → H p (T) the projection defined by
∞
 ∞

P : fn einθ → fn einθ
n=−∞ n=0

and Q = I − P . If p ∈ (1, ∞), the Riesz projection P is bounded and im P = H p .

Note that here we consider operators acting on the spaces Lp , H p or l p . In this
connection, let us agree that whenever p and q appears in the text, they are related
as 1/p + 1/q = 1.
On the space H p , 1 < p < ∞, any function a ∈ L∞ generates two operators—
viz. the Toeplitz operator T (a) : : f → P af and the Hankel operator H (a) : f →
P aQJf , where J : Lp → Lp is the flip operator,
 
(Jf )(t) := t −1 f t −1 , t ∈ T.
Invertibility Issues for Toeplitz Plus Hankel Operators 115

We note that

J 2 = I, J P J = Q, J QJ = P , J aJ = 
a, 
a (t) := a(1/t),

and the operators T (a) and H (b) are related to each other as follows

T (ab) = T (a)T (b) + H (a)H (

b),
(2.1)
H (ab) = T (a)H (b) + H (a)T (
b).

We now consider the invertibility of Toeplitz plus Hankel operators T (a) + H (b)
generated by L∞ -functions a and b and acting on a Hardy space H p (T). Observe
that the matrix representation of such operators in the standard basis {t n }∞n=0 of
H p (T) is

ak−j + 
T (a) + H (b) ∼ ( bk+j +1 )∞
k,j =0 .

There are a variety of approaches to the study of their invertibility and we briefly
discuss some of them.

2.1 Classical Approach: I. Gohberg, N. Krupnik

and G. Litvinchuk

Let L(X) and F (X) be, respectively, the sets of linear bounded and Fredholm
operators on the Banach space X. Besides, if A is a unital Banach algebra, then
GA stands for the group of all invertible elements in A.
Assume that a ∈ GL∞ , b ∈ L∞ and set
 
a − ba −1 d
b
V (a, b) := ,
−c a −1


where c :=  a −1 , d := b
b a −1. Writing T (a)±H (b)+Q for (T (a)±H (b))P +Q, we
consider the operator diag (T (a)+H (b)+Q, T (a)−H (b)+Q) on Lp (T)×Lp (T)
and represent it in the form
 
T (a) + H (b) + Q 0
0 T (a) − H (b) + Q
= A(T (V (a, b)) + diag (Q, Q))B (2.2)
116 V. D. Didenko and B. Silbermann

with invertible operators A and B. More precisely,

  
I 0 I I
B=  .
bI 
aI J −J

and the operator A is also known but its concrete form is not important now.
An immediate consequence of Eq. (2.2) is that both operators T (a) ± H (b) are
Fredholm if and only if the block Toeplitz operator T (V (a, b)) is Fredholm. This
representation is of restricted use because there are piecewise continuous functions
a, b such that only one of the operators T (a) ± H (b) is Fredholm. In addition, both
operators T (a) ± H (b) can be Fredholm but may have different indices. Therefore,
more efficient methods for studying Toeplitz plus Hankel operators are needed,
especially for discontinuous generating functions a and b.
Let us recall that there is a well-developed Fredholm theory for the operators
T (a) + H (b) with generating functions a, b from the set of piecewise continuous
functions P C = P C(T)—cf. [7] for operators acting on the space H 2 and [50] for
the ones on H p , p = 2. However, the defect numbers of these operators, conditions
for their invertibility, and inverse operators can be rarely determined directly from
Eq. (2.2).

2.2 Basor-Ehrhardt Approach

This approach is aimed at the study of defect numbers of T (a) + H (b) ∈ F (H p )

by means of a factorization theory. It is well-known that for b = 0, the problem can
be solved by Wiener-Hopf factorization. Since this notion is important for what
follows, we recall the definition here. Note that from now on, all operators are
considered in the spaces Lp or H p for p ∈ (1, ∞). Recall that p and q are related
as 1/p + 1/q = 1.
Definition 2.1 We say that a function a ∈ L∞ admits a Wiener-Hopf factorization
in Lp if it can be represented in the form

a = a − χm a + , (2.3)
−1 −1
where a+ ∈ H q , a+ ∈ H p , a− ∈ H p , a− ∈ H q , χm (t) := t m , m ∈ Z, the term
−1 p
a+ P (a+ ϕ) belongs to L for any ϕ from the set of all trigonometrical polynomials
P = P(T) and there is a constant cp such that

−1
||a+ P (a+ ϕ)||p ≤ cp ||ϕ||p for all ϕ ∈ P.
Invertibility Issues for Toeplitz Plus Hankel Operators 117

Theorem 2.2 (Simonenko [53]) If a ∈ L∞ , then T (a) ∈ F (H p ) if and only if

a ∈ GL∞ and admits the Wiener-Hopf factorization (2.3) in Lp . In this case

ind T (a) = −m.

This result extends to the case of matrix-valued generating functions as follows.

Theorem 2.3 If a ∈ L∞ ∞
p
N×N , then T (a) ∈ F (HN ) if and only if a ∈ GLN×N and
admits a factorization a = a− da+ , where
q −1 p −1 p q
a+ ∈ HN×N , a+ ∈ HN×N , a− ∈ H N×N , a− ∈ H N×N ,
d = diag (χk1 , χk2 , · · · , χkN ), κ1 , κ2 , · · · , κN ∈ Z,

−1 p
the term a+ P (a+ ϕ) belongs to LN for any ϕ ∈ PN and there is a constant cp such
that
−1
||a+ P (a+ ϕ)||p ≤ cp ||ϕ||p for all ϕ ∈ PN .
p
Moreover, if T (a) ∈ F (HN ), then
 
dim ker T (a) = − κj , dim coker T (a) = − κj .
κj <0 κj >0

The numbers κj , called partial indices, are uniquely defined. Moreover, in some
sense, the Wiener-Hopf factorization is unique if it exists. For example, for N = 1
the uniqueness of factorization can be ensured by the condition a− (∞) = 1. We
also note that if T (a) ∈ F (H p ) and ind T (a) ≥ 0 (ind T (a) ≤ 0) then T (a) is
right-invertible (left-invertible) and if κ := ind T (a) > 0, the functions
−1
a+ χj , j = 0, . . . , κ − 1

form a basis in the space ker T (a) and one of the right-inverses has the form
T −1 (aχκ )T (χκ ).
A comprehensive information about Wiener-Hopf factorization is provided in
[11, 44] and in books, which deal with singular integral and convolution operators
[7, 32, 32, 33]. In particular, Wiener-Hopf factorization furnishes conditions for
invertibility of related operators. However, generally there are no efficient methods
for constructing such factorizations and computing partial indices even for contin-
uous matrix-functions. Therefore, in order to study the invertibility of Toeplitz plus
Hankel operators, we have to restrict ourselves to suitable classes of generating
functions.
In the beginning of this century, Ehrhardt [28, 29] developed a factorization
theory to study invertibility for large classes of convolution operators with flips.
Toeplitz plus Hankel operators are included in this general framework. In particular,
118 V. D. Didenko and B. Silbermann

it was shown that an operator T (a) + H (b), a ∈ GL∞ , b ∈ L∞ is Fredholm if and

only if the matrix-function
 
ba −1 a − ba −1
b
V τ (a, b) =
a −1 a 
− −1 b

admits a certain type of antisymmetric factorization. Moreover, the defect numbers

of the operator T (a)+H (b) can be expressed via partial indices of this factorization.
However, it is not known how the partial indices of general matrix-functions can be
determined.
It was already mentioned that there are functions a, b ∈ L∞ such that the
operator T (a) + H (b) is Fredholm but T (a) − H (b) is not. In this case we can
use the representation (2.2) and conclude that the matrix V τ (a, b) does not admit
a Wiener-Hopf factorization. Thus, if V (a, b) admits a Wiener-Hopf factorization,
then V τ (a, b) has the antisymmetric factorization mentioned but not vice versa.
This discussion shows that one has to select a class of generating functions a
and b such that the defect numbers of operators T (a) + H (b) ∈ F (H p ) can be
determined. An important class of suitable generating functions a ∈ GL∞ (T), b ∈
L∞ (T) is given by the condition

a = b
a b. (2.4)

This class of pairs of functions first appears in [4] and [13]. Equation (2.4) was
latter called the matching condition and the corresponding duo (a, b) a matching
pair—cf. [16]. Let us note that Toeplitz plus Hankel operators of the form

T (a) ± H (a), T (a) − H (at −1 ), T (a) + H (at) (2.5)

appear in random ensembles [1, 6, 30] and in numerical methods for singular integral
equations on intervals [37]. The generating functions of the operators (2.5) clearly
satisfy the matching condition (2.4).
It is notable that the Fredholmness of the operator T (a) + H (b) implies that
a ∈ GL∞ (T). Therefore, the term b in the matching pair (a, b) is also invertible in
L∞ (T) and one can introduce another pair (c, d), called the subordinated pair for
(a, b), with the functions c and d defined by

c := a/b =  a , d := a/
b/ b = b/
a.

c = d d = 1. In what follows, any

An important property of these functions is that c
∞
function g ∈ L (T) satisfying the equation g g = 1 is referred to as a matching
function.
Let us point out that the sets of matching functions and matching pairs are quite
large. In particular, we have:
Invertibility Issues for Toeplitz Plus Hankel Operators 119

1. Let T+ := {t ∈ T : (t > 0} be the upper half-circle. If an element g0 ∈ GL∞ ,

then

g0 (t) if t ∈ T+
g(t) := −1 ,
g0 (1/t) if t ∈ T \ T+

is a matching function.
2. If g1 , g2 are matching functions, then the product g = g1 g2 is also a matching
function.
3. If g is a matching function, then for any a ∈ GL∞ the duo (a, ag) is a matching
pair.
4. Any matching pair (a, b), a ∈ GL∞ can be represented in the form (a, ag),
where g =  ab−1 is a matching function.
In this section we discuss the Basor-Ehrhardt approach to the study of defect
numbers of the operators T (a) + H (b) ∈ F (H p (T)) if a and b are piecewise
continuous functions satisfying the condition (2.4). Then we present an explicit
criterion for the Fredholmness of such operators. Recall that the circle T is
counterclockwise oriented and f ∈ P C if and only if for any t ∈ T, the one-sided
limits

f ± (t) := lim f (teiε )

ε→±0

exist. Without loss of generality we assume that a, b ∈ GP C.

Theorem 2.4 (Basor and Ehrhardt [4]) Let a, b ∈ GP C form a matching pair
and let (c, d) be the subordinated pair. The operator T (a) + H (b) is Fredholm on
the space H p if and only if the following conditions hold:
 
1 1 1 1 1 1
arg c− (1) ∈
/ + +Z , arg d− (1) ∈
/ + +Z , (2.6)
2π 2 2p 2π 2 2q
 
1 − 1 1 − 1
arg c (−1) ∈ / +Z , arg d (−1) ∈ / +Z , (2.7)
2π 2p 2π 2q
 −    −  
1 c (τ ) 1 1 (d) (τ ) 1
arg + ∈
/ +Z , arg ∈
/ +Z , ∀τ ∈ T+ . (2.8)
2π c (τ ) p 2π  +
(d) (τ ) q

The definition of d in [4] differs from the one used here. In fact, d in [4]
corresponds to d here. We keep this notation for sake of easy comparability of the
results.
Theorem 2.4 already shows the exceptional role of the endpoints +1 and −1 of
the upper semicircle T+ . To establish the index formula mentioned in [4], a more
geometric interpretation of the Fredholm conditions (2.6)–(2.8) is needed. Here
there are a few details from [4].
120 V. D. Didenko and B. Silbermann

For z1 , z2 ∈ C and θ ∈ (0, 1), we consider the open arc A(z1 , z2 , θ ) connecting
the points z1 and z2 and defined by
  
1 z − z1
A(z1 , z2 , θ ) := z ∈ C \ {z1 , z2 } : arg ∈ {θ + Z} .
2π z − z2

For θ = 1/2 this arc becomes a line segment and if z1 = z2 it is an empty set.
Assuming that a, b ∈ GP C, a a = b 
b and using the auxiliary functions c and d,
one can show that the conditions (2.6)–(2.8) mean that any of the arcs
     
1 1 1 1
A 1, c+ (1); + , A c− (τ ), c+ (τ ); , A c− (−1), 1; ,
2 2p p 2p
     
 + (1); 1 + 1 , A (d)
A 1, (d)  + (τ ); 1 , A (d)
 − (τ ), (d)  − (−1), 1; 1 ,
2 2q q 2q

where τ ∈ T+ , does not cross the origin. It is clear that one has to take into account
only the jump discontinuity points since c, d ∈ GP C and consequently c± (τ ) = 0
 (τ ) = 0.
and (d) ±

The functions c and  d satisfy the condition c c = d d = 1, so that they are

effectively defined by their values on T+ only. If we let τ run along T+ from
τ = 1 to τ = −1, the image of the function c is the curve with possible
jump discontinuities, starting at the point c+ (1) and terminating at c− (−1). We
now add the arcs A(c− (τ ), c+ (τ ); 1/p) to any discontinuity point τ of c located
on T+ . Besides, if necessary we also add the arcs A(1, c+ (1); 1/2 + 1/(2p))
and A(c− (−1), 1; 1/(2p)) connecting the endpoints c+ (1) and c− (−1) with the
point τ = 1, respectively. That way, we obtain a closed oriented curve. If the
operator T (a) + H (b) is Fredholm, the curve does not cross the origin and we
consider its winding number wind (c#,p ) ∈ Z. Similar constructions lead to the
curve  d #,q with a winding number wind (d#,q ) ∈ Z. Now we can conclude that
T (a) + H (b) ∈ F (H p ) if and only if the origin does not belong to the curve c#,p
or 
d #,q .
Theorem 2.5 (Basor and Ehrhardt [4]) Assume that a, b ∈ GP C is a matching
pair with the subordinated pair (c, d) such that the conditions (2.6)–(2.8) hold. Then
T (a) + H (b) ∈ F (H p ) with the Fredholm index

ind (T (a) + H (b)) = wind (d#,q ) − wind (c#,p ).

For what follows we need a definition.

Definition 2.6 (Basor and Ehrhardt [4]) A matching pair (a, b), a, b ∈ L∞ with
the subordinated pair (c, d) satisfies the basic factorization condition in H p if c and
d admit the factorization of the form
−1 −1
c(t) = c+ (t)t 2n c+ (t ), n ∈ Z, (2.9)
 = (d)
d(t)  −1
 + (t)t 2m (d) −1
+ (t ), m∈Z (2.10)
Invertibility Issues for Toeplitz Plus Hankel Operators 121

and
−1
(1 + t)c+ (t) ∈ H q , (1 − t)c+ (t) ∈ H p ,

(1 + t)(d)  −1
 + (t) ∈ H p , (1 − t)(d) + (t) ∈ H .
q

Note that the indices m, n are uniquely determined and the functions c+ and d+
are also unique up to a multiplicative constant. The representations (2.9), and (2.10)
 respectively.
are called antisymmetric factorization of the functions c and d,
Theorem 2.7 (Basor and Ehrhardt [4]) If (a, b), a, b ∈ P C is a matching pair
and T (a) + H (b) ∈ F (H p ), then (a, b) satisfies the basic factorization condition.
The next result allows to determine the defect numbers of the operators T (a) +
H (b) in certain situations.
Theorem 2.8 (Basor and Ehrhardt [4]) Assume that the matching pair (a, b),
a, b ∈ L∞ satisfies the basic factorization condition in H p with m, n ∈ Z. If
T (a) + H (b) ∈ F (H p ), then
⎧
⎪
⎪ 0 if n > 0, m ≤ 0,
⎪
⎪
⎪
⎨ −n if n ≤ 0, m ≤ 0,
dim ker(T (a) + H (b)) =
⎪
⎪m − n
⎪
⎪
if n ≤ 0, m > 0,
⎪
⎩
dim ker An,m if n > 0, m > 0,
⎧
⎪
⎪ 0 if m > 0, n ≤ 0,
⎪
⎪
⎪
⎨ −m if m ≤ 0, n ≤ 0,
dim ker(T (a) + H (b))∗ =
⎪
⎪ n−m if m ≤ 0, n > 0,
⎪
⎪
⎪
⎩
dim ker(An,m ) if m > 0, n > 0,

Therein, in case n > 0, m > 0,

4 5n−1,m−1
An,m := ρi−j + ρi+j i=0,j =0

and

c+ d+ b−1 ∈ L1 (T).

ρ(t) := t −m−n (1 + t)(1 + t −1 )

In particular, the Fredholm index of T (a) + H (b) is m − n.

Let us now briefly discuss the notion of the adjoint operator used in [4]. If we
identify the dual to H p with H q via the mapping g ∈ H q →< g, · >∈ (H p ) ,
 2π
1
< g, f >:= g(e−iθ )f (eiθ ) dθ,
2π 0
122 V. D. Didenko and B. Silbermann

then the adjoint operator to T (a) + H (b) has the form T (

a ) + H (b), so that

dim ker(T (a) + H (b))∗ = dim ker(T (

a ) + H (b)).

A natural question to ask is whether the Fredholmness of T (a) + H (b), a, b ∈ L∞


implies the existence of the antisymmetric factorizations of the functions c and d.
This problem has been also discussed in [4] and an example there shows that this is
not true in general.
The theory above has been used to examine the operators (2.5) and also the
operators I + H (b) with matching functions b. We are not going to discuss this
approach here. However, in what follows, we present a simple method to handle
these operators.

2.3 Classical Approach Revisited

Now we turn attention to another method based on the classical approach and
recently developed by the authors of this paper. Let us start with a special factor-
ization of the operator T (V (a, b)) for generating functions a and b constituting a
matching pair. If a ∈ GL∞ (T), b ∈ L∞ (T) satisfy the matching condition (2.4),
the matrix function V (a, b) in (2.2) has the form
 
0 d a a
V (a, b) = , c= , d= . (2.11)
a −1
−c  b 
b

It follows that the corresponding block Toeplitz operator T (V (a, b)) with the
generating matrix-function (2.11) can be represented in the form
 
0 T (d)
T (V (a, b)) =
a −1 )
−T (c) T (
   
−T (d) 0 0 −I −T (c) 0
= , (2.12)
0 I a −1 )
I T ( 0 I

with the invertible operator

 
0 −I
: H p × H p → H p × H p.
a −1 )
I T (

This representations turns out to be extremely useful in the study of Toeplitz plus
Hankel operators. To show this we start with the Coburn-Simonenko theorem. For
Toeplitz operators this theorem indicates that for any non-zero a ∈ L∞ (T) one has

min{dim ker T (a), dim coker T (a)} = 0.

Invertibility Issues for Toeplitz Plus Hankel Operators 123

It follows that Fredholm Toeplitz operators with index zero are invertible. However,
in general for block-Toeplitz and for Toeplitz plus Hankel operators, Coburn-
Simonenko theorem is not true. This causes serious difficulties when studying the
invertibility of the operators involved. Nevertheless, the following theorem holds.
Theorem 2.9 Let a ∈ GL∞ (T) and A refer to one of the operators T (a) +
H (at), T (a) − H (at −1 ), T (a) ± H (a). Then

min{dim ker A, dim coker A} = 0.

Proof For a ∈ P C(T) this result goes back to Basor and Ehrhardt [2, 3, 28] with
involved proofs. However, there is an extremely simple proof—cf. [16], based on
the representation (2.2) and valid for generating functions a ∈ GL∞ . We would
like to sketch this proof here. Thus one of the consequences of Eq. (2.2) is that
there is an isomorphism between the kernels of the operators T (V (a, b)) and
diag (T (a) + H (b), T (a) − H (b)). Let us start with the operators T (a) ± H (a).
The corresponding subordinated pairs (c, d) have the form (±1, a a −1), the third
operator in Eq. (2.12) is diag (∓I, I ), so that it does not influence the kernel and
the image of T (V (a, ±a)). Considering the two remaining operators in (2.12), we
note that the Coburn-Simonenko theorem is valid for the block Toeplitz operator
T (V (a, ±a)) and hence for T (a) ± H (a).
Consider now the operators T (a) + H (at). The duo (a, at) is a matching pair
with the subordinated pair (t −1 , d), d = aa −1 t. The operator T (t −1 ) is Fredholm
−1 −1
and ind T (t ) = 1, so that im T (t ) = H p . Besides, a direct check shows that
the function e := e(t) = 1, t ∈ T belongs to the kernels of both operator T (t −1 )
and T (a) − H (at). Assuming that dim ker T (d) > 0 and using Coburn-Simonenko
theorem for Toeplitz operators, we note that im T (d) is dense in H p . On the other
hand, the factorization (2.12) and Eq. (2.2) yield that both spaces im T (V (a, at))
and im diag (T (a) + H (at), T (a) − H (at)) are dense in H p × H p . Hence,

coker (T (a) + H (at)) = coker (T (a) − H (at) = {0}.

Passing to the case dim ker T (d) = 0, we first note that

1 = dim ker T (t −1 )) = dim ker diag ((T (a) + H (at), T (a) − H (at)),

and since the kernel of the operator T (a) − H (at) contains the function e(t) = 1, it
follows that

ker(T (a) + H (at)) = {0}.

a −1t) > 0, then coker (T (a) + H (at)) = {0}, otherwise

Thus if dim ker T (a
ker(T (a) + H (at)) = {0} and the Coburn-Simonenko theorem is proved for
the operators T (a) + H (at). The operators T (a) − H (at −1 ) can be considered
analogously [16].
124 V. D. Didenko and B. Silbermann

Theorem 2.9 can be extended in a few directions—cf. Proposition 3.9 and

Corollary 3.10 below. It is also valid for Toeplitz plus Hankel operators acting
on l p -spaces [14] and for Wiener-Hopf plus Hankel integral operators acting on
Lp (R+ )-spaces [21].

3 Kernel Representations

As was already mentioned, Eq. (2.2) is of limited use in studying the Fredholmness
of the operators T (a) + H (b). However, if the generating functions a and b satisfy
the matching condition, the representation (2.2) allows to determine defect numbers
and obtain efficient representations for the kernels and cokernels of Toeplitz plus
Hankel operators. In order to present the method, we recall relevant results for oper-
ators acting on Hardy spaces—cf. [16]. Let us start with the connections between
the kernels of Toeplitz plus/minus operators and the kernels of the corresponding
block Toeplitz operators. The following lemma is a direct consequence of Eq. (2.2)
and is valid even if a and b do not constitute a matching pair.
Lemma 3.1 If a ∈ GL∞ , b ∈ L∞ and the operators T (a) ± H (b) are considered
on the space H p , 1 < p < ∞, then the following relations hold:
• If (ϕ, ψ)T ∈ ker T (V (a, b)), then

1
a −1 ψ, ϕ + J Qcϕ − J Q
(, ")T = (ϕ − J Qcϕ + J Q a −1 ψ)T (3.1)
2
∈ ker diag (T (a) + H (b), T (a) − H (b)).

• If (, ")T ∈ ker diag (T (a) + H (b), T (a) − H (b)), then

( + ", P ( a J P ( − "))T ∈ ker T (V (a, b)).

b( + ") +  (3.2)

Moreover, the operators

E1 : ker T (V (a, b)) → ker diag (T (a) + H (b), T (a) − H (b)),

E2 : ker diag (T (a) + H (b), T (a) − H (b)) → ker T (V (a, b)),

defined, respectively, by Eqs. (3.1) and (3.2) are mutually inverses to each other.

3.1 Subordinated Operators and Kernels of T (a) + H (b)

Thus, if the kernel of the operator T (V (a, b)) is known, the kernels of both operators
T (a) + H (b) and T (a) − H (b) can be also determined. However, the kernels of the
Invertibility Issues for Toeplitz Plus Hankel Operators 125

operators T (V (a, b)) are known only for a few special classes of matrices V (a, b),
and in the case of general generating functions a, b ∈ L∞ the kernel T (V (a, b)) is
not available. The problem becomes more manageable if a and b form a matching
pair. In this case, V (a, b) is a triangular matrix—cf. (2.11) and the subordinated
functions c and d satisfy the equations

c = 1,
c d d = 1.

Moreover, it follows from Eq. (2.12) that for any function ϕ ∈ ker T (c), the vector
(ϕ, 0) belongs to the kernel of the operator T (V (a, b)) and the first assertion in
Lemma 3.1 shows that

ϕ − J QcP ϕ ∈ ker(T (a) + H (b)),

(3.3)
ϕ + J QcP ϕ ∈ ker(T (a) − H (b)).

Another remarkable fact is that both functions ϕ + J QcP ϕ and ϕ − J QcP ϕ also
belong to the kernel of the operator T (c). In order to show this, we need an auxiliary
result.
Lemma 3.2 Let g ∈ L∞ satisfy the relation g
g = 1 and f ∈ ker T (g). Then

J QgPf ∈ ker T (g) and (J QgP )2 f = f. (3.4)

Proof We only check the first relation (3.4)). Thus

T (g)(J QgPf ) = P gP J QgPf = J Q

g QgPf = J Q
g gPf − J Q
g P gPf = 0,

so that J QgPf ∈ ker T (g).

Considering the operators P±
g : ker T (g) → ker T (g),

1
P±
g := (I ± J QgP ) ,
2 ker T (g)

we observe that according to Lemma 3.2, one has (P± ±

g ) = Pg . Therefore, Pg
2 ±

are complementary projections. This property and Eqs. (3.3) lead to the following
conclusion.
Corollary 3.3 If (c, d) is the subordinated pair for the matching pair (a, b) ∈
L∞ × L∞ , then

ker T (c) = im P− +
c  im Pc ,

im P−
c ⊂ ker(T (a) + H (b)),

im P+
c ⊂ ker(T (a) − H (b)).
126 V. D. Didenko and B. Silbermann

Corollary 3.3 shows the impact of the operator T (c) on the kernels of T (a) +
H (b) and T (a) − H (b). The impact of the operator T (d) on ker(T (a) ± H (b)) is
much more involved. Thus if T (c) is right-invertible and s ∈ ker T (d), then

(Tr−1 (c)T (
a −1 )s, s) ∈ ker T (V (a, b)),

where Tr−1 (c) is one of the right-inverses for T (c).

We now can obtain the following representation of the kernel of the operator
T (V (a, b)).
Lemma 3.4 Let (a, b) be a matching pair such that the operator T (c) is invertible
from the right. Then


ker T (V (a, b)) = (c)  (d)

where
 
(c) := (ϕ, 0)T : ϕ ∈ ker T (c) ,
 
(d) := (Tr−1 (c)T (
 a −1 )s, s)T : s ∈ ker T (d) .

Taking into account this representation and using the first assertion in Lemma 3.1,
we obtain that if s ∈ ker T (d), then

2ϕ± (s) = Tr−1 (c)T (

a −1 )s ∓ J QcP Tr−1 (c)T (
a −1 )s ± J Q
a −1 s
(3.5)
∈ ker(T (a) ± H (b))

The operators ϕ± can be referred as transition operators, since they transform the
kernel of T (d) into the kernels of the operators T (a)±H (b). An important property
of these operators ϕ± is expressed by the following lemma.
Lemma 3.5 Let (c, d) be the subordinated pair for a matching pair (a, b) ∈
L∞ × L∞ . If the operator T (c) is right-invertible, then for every s ∈ ker T (d)
the following relations

(P  a J P )ϕ+ (s) = P+
bP + P d (s),

(P  a J P )ϕ− (s) = P−
bP − P d (s),

hold. Thus the transition operators ϕ+ and ϕ− are injections on the spaces im P+
d
and im P−
d , respectively.
Using Lemmas 3.1–3.5, one can obtain a complete description for the kernel of
the operator T (a) + H (b) if (a, b) is a matching pair and T (c) is right-invertible.
Invertibility Issues for Toeplitz Plus Hankel Operators 127

Proposition 3.6 (VD and BS [16]) Let (c, d) be the subordinated pair for the
matching pair (a, b) ∈ L∞ × L∞ . If the operator T (c) is right-invertible, then
the kernels of the operators T (a) ± H (b) can be represented in the form

ker(T (a) + H (b)) = ϕ+ (im P+ −

d )  im Pc ,

ker(T (a) − H (b)) = ϕ− (im P− +

d )  im Pc .

Remark 3.7 It was shown in [16] that the operators ϕ+ and ϕ− send the elements
of the spaces im P− + − +
d and im Pd into im Pc and im Pc , respectively. Therefore,

ϕ+ : im P− −
d → im Pc , ϕ− : im P+ +
d → im Pc

are well-defined linear operators. If im P− +

c = {0} (im Pc = {0}), then ϕ+ (s− ) = 0
− +
for all s− ∈ im Pd (ϕ− (s+ ) = 0 for all s+ ∈ im Pd ), which yields

a −1 )s− , s− ∈ im P−
ϕ− (s− ) = Tr−1 (c)T ( d,
 
ϕ+ (s+ ) = Tr−1 (c)T (a −1 )s+ , s+ ∈ im P+
d .

Indeed, assume for instance that s− ∈ im P−

d . Then ϕ+ (s− ) = 0 leads to

0 = 2ϕ+ (s− ) = Tr−1 (c)T (

a −1 )s− − J QcP Tr−1 (c)T (
a −1 )s− + J Q
a −1 s− .

Hence,

Tr−1 (c)T (
a −1 )s− = J QcP Tr−1 (c)T (
a −1 )s− − J Q
a −1 s− .

Thus

2ϕ− (s− ) = Tr−1 (c)T (

a −1 )s− + J QcP Tr−1 (c)T (
a −1 )s− − J Q
a −1 s−
= 2Tr−1 (c)T (
a −1 )s−

and the claim follows.

These representations of the transition operators are simpler than (3.5) and
it would be interesting to find out which conditions ensure that the restrictions
ϕ+ im P− and ϕ− im P+ become zero functions.
d d

Thus in order to obtain an efficient description of the spaces ker(T (a) + H (b))
±
and ker(T (a)−H (b)), one has to characterize the projections P± c and Pd first. Such
a characterization can be derived from the Wiener-Hopf factorization (2.3) of the
subordinated functions c and d. The Wiener-Hopf factorization of these functions
can be described in more details, which yields a very comprehensive representation
±
of the kernels of T (c), T (d) and the related projections P±
c , Pd .
128 V. D. Didenko and B. Silbermann

We first consider related constructions for a matching function g such that the
operator T (g) is Fredholm on H p with the index n. One can show that under the
condition g− (∞) = 1, the factorization (2.3) takes the form
−1
g = σ (g)g+ t −n
g+ , (3.6)

−1
where σ (g) = g+ (0) = ±1 and g− = σ (g) g+ . The term σ (g) is called
factorization signature. The finding of σ (g) is a non-trivial problem but if T (g)
is invertible and g is continuous at t = 1 or t = −1, then n = 0 and σ (g) = g(1) or
σ (g) = g(−1), respectively. For piecewise continuous functions g, the term σ (g)
can be also determined.
Notice that T (a) − H (b) can be also written as Toeplitz plus Hankel operator
T (a) + H (−b). Thereby, the duo (a, −b) is again a matching pair with the
subordinated pair (−c, −d) and for the factorization signatures we have σ (−c) =
−σ (c), σ (−d) = −σ (d). This observation shows that we can restrict ourselves
to the study of Toeplitz plus Hankel operators. Nevertheless, in some cases it is
preferable to consider the operator T (a) − H (b) too. But then the leading role still
belongs to the operator T (a)+H (b) since the notions of matching pair, subordinated
pair and factorization signature are associated with this operator.
Let g stand for the subordinated function c or d, so that g g = 1. If T (g)
is Fredholm, then the factorization (3.6) exists with a function g+ satisfying the
conditions for factorization (2.3) and we can describe the spaces im P± g . This
description depends on the evenness of the index of T (g).
Theorem 3.8 (VD and BS [16]) Assume that g is a matching function, the operator
T (g) is Fredholm, ind T (g) = n ≥ 0 and g+ is the plus factor in the Wiener-Hopf
factorization (3.6) of g in H p . Then
• For n = 2r, r ∈ N, the systems of functions
−1 r−k−1
B± (g) := {g+ (t ± σ (g)t r+k ) : k = 0, 1, · · · , r − 1},

form bases in the spaces im P± ±

g and dim ker Pg = r.
• For n = 2r + 1, r ∈ Z+ , the systems of functions
−1 r+k
B± (g) := {g+ (t ± σ (g)t r−k ) : k = 0, 1, · · · , r} \ {0},

form bases in the spaces im P± ±

g and dim ker Pg = r + (1 ± σ (g))/2.

Thus if T (c), T (d) are Fredholm and T (c) is right-invertible, Proposition 3.6
provides complete description of the spaces ker(T (a) ± H (b)). On the other hand,
if T (c) is Fredholm but not right-invertible, the representation

T (a) + H (b) = (T (at −n ) + H (bt n ))T (t n )

Invertibility Issues for Toeplitz Plus Hankel Operators 129

can be used to study ker(T (a) + H (b)). This is because for any matching pair (a, b)
the duo (at −n , bt n ) is also a matching pair with the subordinated pair (ct −2n , d). A
suitable choice of n leads to the right-invertibility of the operator T (ct −2n ) and we
consequently obtain

ker(T (a) + H (b)) = ker(T (at −n ) + H (bt n )) ∩ im T (t n ). (3.7)

The representation (3.7) has been used in [16], to derive the description of the
kernels of the operators T (a) ± H (b). It can be also employed to study one-sided
or generalized invertibility of Toeplitz plus Hankel operators and to construct the
corresponding one-sided and generalized inverses [17, 19, 22]. In the forthcoming
sections, invertibility problems will be discussed in more details. In this connection,
we note that the (familiar) adjoint operator (T (a) + H (b))∗ can be identified with
the operator T (a) + H ( b) acting on the space H q , 1/p + 1/q = 1. It is easily
seen that for any matching pair (a, b), the duo (a,  b) is also a matching pair with
the subordinated pair (d, c) and σ (c) = σ (c), σ (d) = σ (d). Therefore, cokernel
description can be determined directly from the previous results for the kernels of
Toeplitz plus Hankel operators.
In some cases the approach above allows omitting the condition of Fredholmness
of the operator T (d). We note a few results, which can be viewed as an extension of
Coburn-Simonenko Theorem 2.9.
Proposition 3.9 (VD and BS [16]) Let (a, b) ∈ L∞ × L∞ be a matching pair with
the subordinated pair (c, d), and let T (c) be a Fredholm operator. Then:
(a) If ind T (c) = 1 and σ (c) = 1, then

min{dim ker(T (a) + H (b)), dim coker (T (a) + H (b))} = 0.

(b) If ind T (c) = −1 and σ (c) = 1, then

min{dim ker(T (a) − H (b)), dim coker (T (a) − H (b))} = 0.

(c) If ind T (c) = 0, then

min{dim ker(T (a) ± H (b)), dim coker (T (a) ± H (b))} = 0.

An immediate consequence of Proposition 3.9 concerns the Toeplitz plus Hankel

operators of the form I ± H (b).
Corollary 3.10 Let b ∈ L∞ be a matching function such that T (
b) is a Fredholm
operator. Then:
(a) If ind T (
b) = 1 and σ (
b) = 1, then

min{dim ker(I + H (b)), dim coker (I + H (b))} = 0.

130 V. D. Didenko and B. Silbermann

(b) If ind T (
b) = −1 and σ (
b) = 1, then

min{dim ker(I − H (b)), dim coker (I − H (b))} = 0.

(c) If ind T (
b) = 0, then

min{dim ker(I ± H (b)), dim coker (I ± H (b))} = 0.

3.2 Kernels of T (a) + H (b) for Piecewise Continuous

Generating Functions

If more information about generating functions is available, then the kernel of

the Fredholm operator T (a) + H (b) ∈ L(H p ) can be studied under weaker
conditions. Thus for piecewise continuous functions a and b, the assumption that
the subordinated operators T (c), T (d) ∈ L(H p ) are Fredholm can be dropped. In
order to show this we need a few facts from [16].
Let A be an operator defined on all spaces Lp for 1 < p < ∞. Consider the set

AF := {p ∈ (1, ∞) such that the operator A : H p → H p is Fredholm}.

Proposition 3.11 (Šneı̆berg [54]) The set AF is open. Moreover, for each con-
nected component γ ∈ AF , the index of the operator A : Lp → Lp , p ∈ γ is
constant.
This result also holds for operators A acting on the spaces H p , 1 < p < ∞,
since any operator A : H p → H p can be identified with the operator AP + Q
acting already on Lp . For Toeplitz operators the structure of the set AF can be
characterized as follows.
Proposition 3.12 (Spitkovskiı̆ [55]) Let G be an invertible matrix–function with
entries from P C and let A := T (G). Then there is an at most countable subset
SA ⊂ (1, ∞) with the only possible accumulation points t = 1 and t = ∞ such
that AF = (1, ∞) \ SA .
Clearly, if G is piecewise continuous with only a finite number of discontinuities,
then SA is at most finite. This result can be used to describe the corresponding set
AF for Toeplitz plus Hankel operators with P C-generating functions.
Corollary 3.13 Let a, b ∈ P C and

A := diag (T (a) + H (b), T (a) − H (b)) : H p × H p → H p × H p .

Then there is at most countable subset SA ⊂ (1, ∞) with only possible accumula-
tion points t = 1 and t = ∞ such that AF = (1, ∞) \ SA .
Invertibility Issues for Toeplitz Plus Hankel Operators 131

Proof It follows directly from Proposition 3.12 since diag (T (a) + H (b), T (a) −
H (b)) is Fredholm if and only if so is the operator T (V (a, b)).
Thus if a, b ∈ P C and the operator T (a) + H (b) is Fredholm on H p , there is
an interval (p , p ) containing p such that T (a) + H (b) is Fredholm on all spaces
H r , r ∈ (p , p ) and the index of this operator does not depend on r. Moreover,
there is an interval (p, p0 ) ⊂ (p , p ), p < p0 such that T (a) − H (b) is Fredholm
on H r , r ∈ (p, p0 ) and its index does not depend on r. Recalling that for ν < s
the space H s is continuously embedded into H ν , we note that the kernel of T (a) +
H (b) : H r → H r does not depend on r ∈ (p , p ). The same is also true for
ker(T (a) − H (b)), r ∈ (p, p0 ). We want to note that both claims are based on the
following well-known result.
Lemma 3.14 (Gohberg and Feldman [31]) Let a Banach space B1 be continu-
ously and densely embedded into a Banach space B2 . Further, let A1 and A2 be
bounded Fredholm operators which respectively act on B1 and B2 and have equal
indices. If A2 is an extension of A1 , then

ker A1 = ker A2 .

Hence, the kernel of the operator T (a) + H (b) acting on the space H p coincides
with the kernel of this operator acting on the space H r for r ∈ (p, p0 ) and the
latter can be studied by the approach above. Therefore, if T (a) + H (b) ∈ L(H p ) is
Fredholm and a, b ∈ P C form a matching pair, the kernel of the operator T (a) +
H (b) can be described.

4 Generalized Invertibility, One-Sided Invertibility,

and Invertibility

Let a, b ∈ L∞ be a matching pair with the subordinated pair (c, d). The pair is
called a Fredholm matching pair (with respect to H p ) if the operators T (c), T (d) ∈
p
L(H p ) are Fredholm. We write T (a) + H (b) ∈ FT H (κ1 , κ2 ) if (a, b) is a Fredholm
matching pair with the subordinated pair (c, d) such that ind T (c) = κ1 , ind T (d) =
κ2 . It was first noted in [15] that if κ1 ≥ 0, κ2 ≥ 0 or if κ1 ≤ 0, κ2 ≤ 0, then (2.2)
and (2.12) yield one-sided invertibility of the operator T (a) + H (b). However, if
κ1 κ2 < 0, the invertibility issues become more involved. We start this section by
considering the generalized invertibility of the operators T (a) + H (b), a, b ∈ L∞ .
Set

B := PV (a, b)P + Q,

where P := diag (P , P ), Q := diag (Q, Q).

132 V. D. Didenko and B. Silbermann

Theorem 4.1 (VD and BS [17]) Assume that (a, b) is a matching pair with the
subordinated pair (c, d) and B is generalized invertible operator, which has a
generalized inverse B −1 of the form
 
AB
B −1 = + Q, (4.1)
D0

where A, B and D are operators acting in the space H p . Then the operator T (a) +
H (b) : H p → H p is also generalized invertible and one of its generalized inverses
has the form

(T (a) + H (b))−1
g = −H ( a −1))
c)(A(I − H (d)) − BH (

+ H (a −1 )D(I − H (d)) + T (a −1 ).

This result can now be used to construct generalized inverses for the operator
T (a) + H (b) in the following cases—cf. [17]:
(a) If κ1 ≥ 0 and κ2 ≥ 0, then B has generalized inverse of the form (4.1) with
A = Tr−1 (c)T (
a −1 )Tr−1 (d), B = −Tr−1 (c) and D = Tr−1 (d).
(b) If κ1 ≤ 0 and κ2 ≤ 0, then B has generalized inverse of the form (4.1) with
A = Tl−1 (c)T (
a −1 )Tl−1 (d), B = −Tl−1 (c) and D = Tl−1 (d).
(c) If κ1 ≥ 0 and κ2 ≤ 0, then B has generalized inverse of the form (4.1) with
a −1 )Tl−1 (d), B = −Tr−1 (c) and D = Tl−1 (d).
A = Tr−1 (c)T (
It is clear that in the cases (a) and (b), generalized inverses are, respectively, right
and left inverses. We also note that under conditions of assertion (a), a right inverse
of T (a) + H (b) can be written in a simpler form—cf. [19]

c))Tr−1 (c)T (
B := (I − H ( a −1 )Tr−1 (d) + H (a −1)Tr−1 (d). (4.2)

The proof of this result is straightforward—i.e. one can use the relations (2.1) to
verify that (T (a) + H (b))B = I . On the other hand, under conditions of (b), a
simpler representation of the left-inverse of T (a) + H (b) can be derived from (4.2)
by passing to the adjoint operator.
In addition to the cases considered, there is one more situation—viz.
(d) κ1 < 0, κ2 > 0.
This case is much more involved and we will deal with it later on. At the moment,
p
we focus on invertibility of operators from FT H (κ1 , κ2 ), 1 < p < ∞.
p
Theorem 4.2 (VD and BS [19]) Assume that T (a) + H (b) ∈ FT H (κ1 , κ2 ) is
invertible. Then:
(i) If κ1 ≥ κ2 or κ1 κ2 ≥ 0, then

|κ1 | ≤ 1, |κ2 | ≤ 1. (4.3)

Invertibility Issues for Toeplitz Plus Hankel Operators 133

(ii) If κ1 < 0 and κ2 > 0, then

(a) If κ1 and κ2 are even numbers, then κ2 = −κ1 .
(b) If κ1 is an odd number and κ2 is an even one, then κ2 = −κ1 + σ (c).
(c) If κ1 is an even number and κ2 is an odd one, then κ2 = −κ1 − σ (d).
(d) If κ1 and κ2 are odd numbers, then κ2 = −κ1 + σ (c) − σ (d).
Our next goal is to obtain sufficient invertibility conditions for the invertibility
p
of the operators from FT H (κ1 , κ2 ) and to provide effective representations for their
inverses. We assume first that κ1 and κ2 satisfy conditions (4.3).
p
Theorem 4.3 (VD and BS [19]) Assume that T (a) + H (b) ∈ FT H (κ1 , κ2 ) and the
indices of the subordinated operators T (c), T (d) and the factorization signatures
of c and d satisfy one of the following conditions:
(i) κ1 = 0, κ2 = 0;
(ii) κ1 = 1, κ2 = 0 and σ (c) = 1;
(iii) κ1 = 0, κ2 = 1 and σ (d) = −1;
(iv) κ1 = 1, κ2 = 1 and σ (c) = 1, σ (d) = −1;
(v) κ1 = 0, κ2 = −1 and σ (d) = 1;
(vi) κ1 = −1, κ2 = 0 and σ (c) = −1;
(vii) κ1 = −1, κ2 = −1 and σ (c) = −1, σ (d) = 1;
(viii) κ1 = 1, κ2 = −1 and σ (c) = 1, σ (d) = 1;
(ix) κ1 = −1, κ2 = 1 and σ (c) = −1, σ (d) = −1.
Then the operator T (a) + H (b) is invertible. Moreover, we have:
1. Under conditions (i)–(iv), the inverse operator has the form (4.2), where right-
inverses of T (c) or/and T (d) shall be replaced by the corresponding inverses.
2. Under conditions (v)–(vii), the inverse operator has the form

c )(Tl−1 (c)T (
(T (a) + H (b))−1 = − H ( a −1 )Tl−1 (d)(I − H (d))

+ Tl−1 (c)H (
a −1 )) + H (a −1 )Tl−1 (d)(I − H (d)) + T (a −1 ).

3. Under condition (viii), the inverse operator has the form

(T (a) + H (b))−1 = − H ( a −1 )Tl−1 (d)(I − H (d))

c )(Tr−1 (c)T (

a −1 )) + H (a −1 )Tl−1 (d)(I − H (d)) + T (a −1 ).

+ Tr−1 (c)H (

4. Under condition (ix), the inverse operator has the form

−1
(T (a) + H (b))−1 = T (t −1 )(I − c+ tQt −1 )
c))Tr−1 (t −2 c)T (
× [(I − H (t 2 a −1 t −1 )Tr−1 (d) + H (a −1 t)Tr−1 (d)],

where c+ is the plus factor in factorization (3.6) for the function c.

134 V. D. Didenko and B. Silbermann

Theorem 4.3 is, in fact, the collection of various results from [19]. On the other
hand, conditions (i)–(ix) are not necessary for the invertibility of T (a) + H (b) in
the case (4.3). Thus if κ1 = −1, κ2 = 1, then the operator T (a) + H (b) can be
invertible even if σ (c) and σ (d) do not satisfy condition (ix). This case, however,
requires special consideration and has been not yet studied.
Consider now the situation (ix) in more detail. This is a subcase of assertion (ii)
in Theorem 4.2 and a closer inspection shows substantial difference from the other
cases in Theorem 4.3. What makes it very special is the presence of the factorization
factor of c in the representation of the inverse operator. It is also worth noting that
the construction of (T (a) + H (b))−1 is more involved and requires the following
result.
Lemma 4.4 (VD and BS [19, 22]) Let C, D be operators acting on a Banach space
X. If A = CD is an invertible operator, then C and D are, respectively, right- and
left-invertible operators. Moreover, the operator D : X → im D is invertible and
if D0−1 : im D → X is the corresponding inverse, then the operator A−1 can be
represented in the form

A−1 = D0−1 P0 Cr−1 ,

where P0 is the projection from X onto im D parallel to ker C and Cr−1 is any right-
inverse of C.
Theorem 4.2(ii) provides necessary conditions for the invertibility of T (a) +
p
H (b) ∈ FT H (κ1 , κ2 ) for negative κ1 and positive κ2 . In the next section, we take a
closer look at the condition (iia). The related ideas can be seen as a model to study
invertibility in cases (iib)–(iid) of Theorem 4.2.
Now we would like to discuss a few examples.
Example 4.5 Let us consider the operator T (a) + H (b) in the case where a = b. In
this situation c(t) = 1 and d(t) = a(t)a −1 (t). Hence, H (
c) = 0, T (c) = I and if
ind T (d) = 0, then the operator T (a) + H (a) is also invertible and

(T (a) + H (a))−1 = (T (
a −1 ) + H (a −1 ))T −1 (a
a −1).

Example 4.6 Similar approach show that the operator T (a) + H (

a) is invertible if
a −1 (t) is invertible and
T (c), c(t) = a(t)

a ))−1 = (I − H (
(T (a) + H ( aa −1 ))T −1 (a
a −1)T (
a −1 ) + H (a −1 ).

Example 4.7 Consider the operator I +H (φ0 b), where b(t) b(t) = 1 and φ0 (t) = t,
φ0 (t) = −t −1 or ϕ0 (t) = ±1 for all t ∈ T. Assume that the operator T ( b) is
Fredholm. Corollary 3.10 indicates that if one of the conditions
(a) ind T (
b) = 0 and ϕ0 (t) = ±1,
(b) ind T (
b) = 0, σ (
b) = 1 and ϕ0 (t) = t,
(c) ind T (b) = 0, σ (
 b) = 1 and ϕ0 (t) = −t −1 ,
Invertibility Issues for Toeplitz Plus Hankel Operators 135

holds, then

min{dim ker(I + H (ϕ0 b)), dim coker (I + H (ϕ0b))} = 0.

Therefore, if I +H (ϕ0 b) is Fredholm with index zero, then this operator is invertible.
However, for b ∈ P C, the Fredholmness of the operators T ( b) and I + H (ϕ0b) can
be studied by Theorems 2.4 and 2.5. It is also possible to construct the inverse
operator using the corresponding results on the factorization of P C-functions.
However, instead of going into details, we would like to observe that if T (b),
b ∈ L∞ is invertible, then I + T (ϕ0 b) is also invertible under the conditions of
Theorem 4.3(i), (viii), and (ix), respectively. Moreover, the inverse operator can be
explicitly constructed.
Using a distinct method, Basor and Ehrhardt [5] also proved the invertibility
of this operator on H 2 under the condition that T ( b) : H 2 → H 2 is invertible.
2
For the H -space, the invertibility of T (b) automatically follows from that of
T (b). However, if p = 2, this is not true and the corresponding examples can be
found already among operators with piecewise continuous generating functions. It
is interesting enough that in each case ϕ0 (t) = ±1, ϕ0 (t) = t or ϕ0 (t) = −t −1 , the
inverse operator can be represented in the form

(I + H (ϕ0b))−1 = (T (
b) + H (ϕ0))−1 (I + H (ϕ0
b))(T (b) + H (ϕ0 ))−1 .

p
5 Invertibility of Operators from FT H (−2n, 2n), n ∈ N

Theorem 4.2(ii) provides necessary conditions for invertibility of the operator

p
T (a)+H (b) ∈ FT H (κ1 , κ2 ) if κ1 < 0 and κ2 > 0. There are four different situations
p
to consider. Here we focus only on the case T (a) + H (b) ∈ FT H (−2n, 2n), n ∈ N,
but the reader can handle the remaining cases using the ideas below. Let us start with
p
the simplest case T (a) + H (b) ∈ FT H (−2, 2) and let (c, d) be the subordinated
pair for the matching pair (a, b) ∈ L∞ × L∞ . In passing note that if T (a) + H (b) ∈
FT H (−2, 2), then the adjoint operator (T (a) + H (b))∗ = T (a) + H (
p
b) belongs to
q
the set FT H (−2, 2).
According to (3.6), the Wiener-Hopf factorization of the function d is

d(t) = σ (d)d+ (t) t −2 −1

d+ (t).

It is easily seen that the operator T (a) + H (b) can be represented in the form

T (a) + H (b) = (T (a1 ) + H (b1 ))T (t), (5.1)

where a1 = at −1 and b1 = tb. The duo (a1 , b1 ) = (at −1 , bt) is a matching pair
with the subordinated pair (c1 , d1 ) = (ct −2 , d). Hence, T (a1 )+H (b1 ) ∈ FT H (0, 2)
p
136 V. D. Didenko and B. Silbermann

and we note that T (c1 ) is invertible. The invertibility of T (c1 ) implies that both
projections P+ −
c1 are Pc1 are the zero operators. According to Remark 3.7, the
functions ϕ± admit the representations

ϕ± (s± ) = Tr−1 (c)T (

a −1 )s± , s± ∈ im P±
d.

so that

a1−1 )(im P±
ker(T (a1 ) ± H (b1)) = Tr−1 (c1 )T ( d ).

Further, we also note that

im P± −1
d = {νd+ (1 ± σ (d)t : ν ∈ C}

is a one-dimensional subspace of ker T (d).

By ωa,b,± we denote the functions
−1 −1
ωa,b,± (t) = T −1 (ct −2 )T (
a −1 t −1 )(d+ (t) ± σ (d)d+ (t)t), (5.2)

which respectively belong to the kernels of the operators T (a1 ) ± H (b1). It is clear
that ωa,b,± also depend on p.
Representation (5.1) shows that T (a) + H (b) has trivial kernel if and only if

ker(T (a1 ) + H (b1)) ∩ im T (t) = {0}.

It is possible only if

ω0a,b,+ = 0,


where  ω0a,b,+ is the zero Fourier coefficient of the function ωa,b,+ (t). Similar result
is valid for the operator T (a) − H (b). Note that if T (a) + H (b) belongs to the set
p
FT H (−2, 2), then so is the operator T (a) − H (b), since T (a) − H (b) = T (a) +
H (−b).
p
Theorem 5.1 (VD and BS [22]) Let T (a) + H (b) ∈ FT H (−2, 2).
(a) The operator T (a) + H (b) (T (a) − H (b)) is left-invertible if and only if
ω0a,b,+ = 0 (
 ω0a,b,− = 0).
(b) The operator T (a) + H (b) (T (a) − H (b)) is right-invertible if and only if
 
0a,b,+ = 0 (
ω ω0a,b,− = 0).
ω0a,b,+ = 0
(c) The operator T (a)+H (b) (T (a)−H (b)) is invertible if and only if 
 
and ω0a,b,+ = 0 (
ω0a,b,− = 0 and ω0a,b,− = 0).
a,b,+ a,b,−
(d) If 
ω0 = 0 and ω0 = 0, then both operators T (a) + H (b) and T (a) −
H (b) are invertible.
Invertibility Issues for Toeplitz Plus Hankel Operators 137

Let us sketch the proof of Assertion (d). It follows from the representations (2.2)
and (2.12) that

ind (T (a) + H (b)) + ind (T (a) − H (b)) = 0. (5.3)

By Assertion (a), both operators T (a) + H (b) and T (a) − H (b) are left-invertible.
Therefore, ind (T (a) + H (b)) ≤ 0, ind (T (a) − H (b)) ≤ 0 and taking into account
(5.3), we obtain that

ind (T (a) + H (b)) = ind (T (a) − H (b)) = 0,

which yields the invertibility of both operators under consideration.

p
If an operator T (a) + H (b) ∈ FT H (−2, 2) is invertible, we can construct its
inverse by using Lemma 4.4. In particular, we have.
p
Theorem 5.2 (VD and BS [22]) If T (a) + H (b) ∈ FT H (−2, 2) is invertible, then
the inverse operator can be represented in the form

(T (a) + H (b))−1

1 −1
= T (t −1 ) I − T −1 (ct −2 )T (
a −1 t −1 )d+ (t)(1 + σ (d)t)tQt −1
ω0a,b,+

 
c))T −1 (t −2 c)T (
× (I − H (t 2 a −1 t −1 )Tr−1 (d) + H (a −1t)Tr−1 (d) . (5.4)

Example 5.3 We now consider the operator A = T (a) + H (t −2 a), defined by the
function

a(t) := (1 − γ t −1 )(1 − γ t)−1 ,

where γ is a fixed number in the interval (0, 1).

The duo (a, at −2 ) is a matching pair with the subordinated pair (c, d) =
(t 2 , a −1 t 2 ) and A ∈ FT H (−2, 2). The corresponding Wiener-Hopf factorizations
p

of a and d are the same in all H p . More precisely, we have

−1
a(t) = a+ (t)
a+ (t), a+ (t) = (1 − γ t)−1 ,

d(t) = 1 · d+ (t)t −2 d+

−1
(t), d+ (t) = (1 − γ t)−2 .

Hence, σ (d) = 1 and computing the zero Fourier coefficients of the corresponding
functions (5.2), we obtain
−2 ,+ −2 ,−
0a,at
w 0a,at
=w = γ 2 − γ + 1.
138 V. D. Didenko and B. Silbermann

It is easily seen that for any γ ∈ (0, 1) these coefficients are not equal to zero, so that
by Theorem 5.1(c), the operator T (a) + H (at −2 ) is invertible. The corresponding
inverse operator, which is constructed according to the representation (5.4), has the
form

(T (a) + H (t −2 a))−1
   
1
= T (t −1 ) I − 2 P (1 − γ t −1 )(1 − γ t)(1 + t) Qt −1
γ −γ +1
    
(1 − γ t −1 )t −1 (1 − γ t)t
× T +H
1 − γt 1 − γ t −1
× T ((1 − γ t)2 )T ((1 − γ t −1 )−2 )T (t 2 ).

Consider now the invertibility of the operators T (a) + H (b) from the set
p
FT H (−2n, 2n) for n greater than 1. Thus we assume that the subordinated operator
T (c) = T (ab−1 ) and T (d) = T (ab −1) are Fredholm and

ind T (c) = −2n, ind T (d) = 2n, n > 1.

Considering the functions

−1
ωka,b (t) := T −1 (ct −2n )T (
a −1 t −n )d+ (t)(t n−k−1 + σ (d)t n+k )

for k = 0, 1, · · · , n − 1, we introduce the matrix

Wn (a, b) = (ωja,b n−1

k )k,j =0 ,

with the entries

 2π
1
ωja,b
k = ωka,b (eiθ )e−ij θ dθ, j, k = 0, 1, · · · , n − 1,
2π 0

and the terms d+ and σ (d) defined by the Wiener-Hopf factorization

d(t) = σ (d)d+ (t) t −2n −1

d+ (t),

with respect to Lp . Notice that the functions ωka,b form a basis in ϕ+ (im P+
d ), where
+
ϕ+ : im Pd → ker(T (a1 ) + H (b1 )) is defined by

ϕ+ = T −1 (ct −2n )T (
a −1 t −n )).
p
The invertibility of the operators from FT H (−2n, 2n) is described by the following
theorem.
Invertibility Issues for Toeplitz Plus Hankel Operators 139

p
Theorem 5.4 (VD and BS [22]) If T (a) + H (b) ∈ FT H (−2n, 2n), then:
(a) T (a)+H (b) is left-invertible if and only if Wn (a, b) is a non-degenerate matrix.
(b) T (a) + H (b) is right-invertible if and only if Wn (a,  b) is a non-degenerate
matrix.
(c) T (a) + H (b) is invertible if and only if Wn (a, b) and Wn (a,  b) are non-
degenerate matrices.
(d) If Wn (a, b) and Wn (a, −b) are non-degenerate matrices, then both operators
T (a) + H (b) and T (a) − H (b) are invertible.
Example 5.5 Consider operator T (a) + H (at −2n ), n ∈ N and

1 − γ t −1
a(t) = , γ ∈ (0, 1). (5.5)
1 − γt

It was shown in [22] that for any n ∈ N, the above operator is left-invertible in any
space H p , 1 < p < ∞. Moreover, since H (at −2n ) is compact and ind T (a) = 0,
the operator at hand is even invertible. The inverse of T (a) + H (at −2n ) can be
constructed in explicit form.
p
Remark 5.6 If m, n ∈ N and m = n, the set FT H (−2m, 2n) does not contain
invertible operators, but it still includes one-sided invertible operators.

6 Toeplitz Plus Hankel Operators on l p -Spaces

A substantial portion of the results presented in Sects. 2–5 can be extended to

Toeplitz plus Hankel operators acting on l p -spaces, 1 < p < ∞. Such an extension
is highly non-trivial because many tools perfectly working in classical Hardy spaces
H p , are not available for operators on l p -spaces. In particular, a big problem is the
absence of a Wiener-Hopf factorization, which plays outstanding role in the study
of Toeplitz plus Hankel operators on classical H p -spaces.

6.1 Spaces and Operators

Let l p (Z) denote the complex Banach space of all sequences ξ = (ξn )n∈N of
complex numbers with the norm
 1/p

||ξ ||p = |ξn | p
, 1 ≤ p < ∞.
n∈Z
140 V. D. Didenko and B. Silbermann

As usual, Z denotes the set of all integers. If we replace Z by the set of all non-
negative integers Z+ , we get another Banach space l p (Z+ ). It can be viewed as a
subspace of l p (Z) and we will often write l p for l p (Z+ ). By P , we now denote the
canonical projection from l p (Z) onto l p (Z+ ) and let Q := I − P . Further, let J
refer to the operator on l p (Z) defined by

(J ξ )n = ξ−n−1 , n ∈ Z.

The operators J, P and Q are connected with each other by the relations

J 2 = I, J P J = Q, J QJ = P .

For each a ∈ Lp = Lp (T), let ( ak )k∈Z be the sequence of its Fourier coefficients.
The Laurent operator L(a) associated with a ∈ L∞ (T) acts on the space l 0 (Z) of
all finitely supported sequences on Z by

(L(a)x)k := 
ak−m xm , (6.1)
m∈Z

where the sum in the right-hand side of (6.1) contains only a finite number of non-
zero terms for every k ∈ Z. We say that a is a multiplier on l p (Z) if L(a)x ∈ l p (Z)
for any x ∈ l0 (Z) and if

||L(a)|| := sup{||L(a)x||p : x ∈ l0 (Z), ||x||p = 1}

is finite. In this case, L(a) extends to a bounded linear operator on l p (Z), which is
again denoted by L(a). The set M p of all multipliers on l p (Z) is a Banach algebra
under the norm ||a||Mp := ||L(a)||—cf. [7]. It is well-known that M 2 = L∞ (T).
Moreover, every function a ∈ L∞ (T) with bounded total variation Var (a) is in M p
for every p ∈ (1, ∞) and the Stechkin inequality

||a||M p ≤ cp (||a||∞ + Var (a))

holds with a constant cp independent of a. In particular, every trigonometric

polynomial and every piecewise constant function on T are multipliers on any space
l p (Z), p ∈ (1, ∞). By Cp and P Cp we, respectively, denote the closures of algebras
of all trigonometric polynomials E and all piecewise constant functions P C in M p .
Note that C2 is just the algebra C(T) of all continuous functions on T, and P C2 is
the algebra P C(T) of all piecewise continuous functions on T. We also note that
the Wiener algebra W of the functions with absolutely converging Fourier series is
also a subalgebra of M p and

W ⊂ Cp ⊂ P Cp ⊂ P C and M p ⊂ L∞ .
Invertibility Issues for Toeplitz Plus Hankel Operators 141

For this and other properties of multiplier cf. [7]. We also recall the equation
J L(a)J = L( a ) often used in what follows.
Let a ∈ M p . The operators T (a) : l p → l p , x → P L(a)x and H (a) : l p → l p ,
x → P L(a)J x = P L(a)QJ are, respectively, called Toeplitz and Hankel operators
with generating function a. It is well-known that ||T (a)|| = ||a||M p and ||H (a)|| ≤
||a||M p for every multiplier a ∈ M p . In this section we also use the notation Tp (a)
or Hp (a) in order to underline that the corresponding Toeplitz or Hankel operator
is considered on the space l p for a fixed p ∈ (1, ∞). The action of the operators
Tp (a) and Hp (a) on the elements from l p can be written as follows
⎛ ⎞

T (a) : (ξj )j ∈Z+ → ⎝ aj −k ξk ⎠
 ,
k∈Z+ j ∈Z+
⎛ ⎞

H (a) : (ξj )j ∈Z+ → ⎝ aj +k+1 ξk ⎠
 .
k∈Z+ j ∈Z+

Let GM p denote the group of invertible elements in M p .

Lemma 6.1 (cf. Ref. [7]) Let p ∈ (1, ∞).
1. If Tp (a) is Fredholm, then a ∈ GM p .
2. If a ∈ M p , then one of the kernels of the operators Tp (a) or Tq∗ (a), 1/p + 1/q =
1 is trivial.
3. If a ∈ GM p , the operator Tp (a) is Fredholm, and ind Tp (a) = 0, then T (a) is
invertible on l p .

6.2 Kernels of a Class of Toeplitz Plus Hankel Operators

The goal of this subsection is to present a method on how to study certain problems
for Toeplitz plus Hankel operator Tp (a) + Hp (b) defined on l p via known results
obtained in Sects. 2–5. Since M p ⊂ L∞ (T), for any given elements a, b ∈ M p the
operator Tp (a) + Hp (b) ∈ L(l p ) generates the operator Ts (a) + Hs (b) ∈ L(H s ),
1 < s < ∞ in an obvious manner. We denote by Lp,q (l p ) the collection of
all Toeplitz plus Hankel operators acting on the space l p such that the following
conditions hold:
(a) a, b ∈ M p .
(b) Tp (a) + Hp (b) ∈ L(l p ) is Fredholm.
(c) Tq (a) + Hq (b) ∈ L(H q ), 1/p + 1/q = 1 is Fredholm.
(d) Both operators acting on the spaces mentioned have the same Fredholm index.
142 V. D. Didenko and B. Silbermann

The following observation is crucial for what follows. The famous Hausdorff-Young
Theorem connects the spaces l p and H q , 1/p + 1/q = 1 via Fourier transform
F (a) = (
an )n∈Z , a ∈ H q .
Theorem 6.2 (Hausdorff and Young [27])
(a) If g ∈ H p and 1 ≤ p ≤ 2, then F g ∈ l q and

||F g||q ≤ ||g||p .


(b) If ϕ = (ϕn )n∈Z ∈ l p (Z) and 1 ≤ p ≤ 2, then the series ϕn eint converges in
Lq to a function ϕ̆ and

||ϕ̆||q ≤ ||ϕ||p .

Theorem 6.2 gives rise to the following construction. Let H p , p ∈ (1, ∞) be

the set of all sequences (χn )n∈Z+ such that there exists a function h ∈ H p with
p equipped with the norm ||(gk )k∈Z+ || :=
the property F h = (χn )n∈Z+ . The set H
||g||H p , becomes a Banach space isometrically isomorphic to H p . Part (a) of the
Hausdorff-Young Theorem assures that H p is densely continuously embedded in
the space l and part (b) claims that l , 1 ≤ p ≤ 2 is continuously and densely
q p

embedded into H q .

Theorem 6.3 (VD and BS [21]) Let p ∈ (1, ∞). If Tp (a) + Hp (b) ∈ Lp,q , then
the Fourier transform F is an isomorphism between the spaces ker(Tq (a) + Hq (b))
and ker(Tp (a) + Hp (b)).
Let us give a sketch of the proof. The matrix representation [Tq (a) + Hq (b)] of
the operator Tq (a) + Hq (b) in the standard basis (t n )n∈Z+ induces a linear bounded
operator on Hq , so that

ai−j + 
[Tq (a) + Hq (b)] = ( bi+j +1 )∞
i,j =0

The operators Tq (a) + Hq (b) ∈ L(H q ) and [Tq (a) + Hq (b)] ∈ L(H q ) have the
same Fredholm properties as Tp (a) + Hp (b) ∈ L(l ) for 1 < q ≤ 2. Moreover, the
p

operator Tp (a) + Hp (b) ∈ L(l p ) is the extension of [Tq (a) + Hq (b)] with the same
index. However, in this case Lemma 3.14 then indicates that

ker[Tq (a) + Hq (b)] = ker(Tp (a) + Hp (b)), Tp (a) + Hp (b) ∈ L(l p ),

and for p ≥ 2 the assertion follows. The case 1 < p ≤ 2 can be treated analogously.
Thus the main problem in using Theorem 6.2 is whether it is known that Tp (a) +
Hp (b) ∈ Lp,q . This is a non-trivial fact but the following result holds.
Proposition 6.4 (VD and BS [21]) Let a, b ∈ P Cp , 1 < p < ∞. If the operator
Tp (a) + Hp (b) ∈ L(l p ) is Fredholm, then Tp (a) + Hp (b) ∈ Lp,q .
Invertibility Issues for Toeplitz Plus Hankel Operators 143

The proof of this proposition can be carried out using ideas from [50] and [51].
However, a simpler proof can be obtained if the generating function a and b satisfy
the matching condition a a = bb and the operators Tp (c), Tp (d) are Fredholm say
in l . Then Tp (c), Tp (d) ∈ Lp,q —cf. [7, Chapters 4 and 6], and using classical
p

approach, which also works in l p situation, one obtains the result.

Thus, it is now clear how to extend the results of Sects. 2–5 to Toeplitz plus
Hankel operators acting on spaces l p . Let us formulate just one such result without
going into much details.
Theorem 6.5 Let (a, b) ∈ P Cp × P Cp be a Fredholm matching pair with the
−1
subordinated pair (c, d), and let 
c+,j , j ∈ Z+ be the Fourier coefficients of the
−1
function c+ , where c+ is the plus factor in the Wiener-Hopf factorization (3.6) of
the function c in H q . If κ1 := ind Tp (c) > 0, κ2 := ind Tp (d) ≤ 0, then the kernel
of the operator Tp (a) + Hp (b) admits the following representation:
(a) If κ1 = 1 and σ (c) = 1, then

ker(Tp (a) + Hp (b)) = {0}.

(b) If κ1 = 1 and σ (c) = −1, then

−1
ker(Tp (a) + Hp (b)) = lin span{(
c+,j )j ∈Z+ }.

(c) If κ1 > 1 is odd, then

−1 −1
ker(Tp (a) + Hp (b)) = lin span{(
c+,j −(κ1 −1)/2−l − σ (c)
c+,j −(κ1 −1)/2+l )j ∈Z+ :

l = 0, · · · , (κ1 − 1)/2}.

(d) If κ1 is even, then

−1 −1
ker(Tp (a) + Hp (b)) = lin span{(
c+,j −κ1 /2+l+1 − σ (c)
c+,j −κ1 /2−l )j ∈Z+ :

l = 0, 1, · · · , κ1 /2 − 1}.

Remark 6.6 Sometime the study of invertibility of Toeplitz plus Hankel operators
Tp (a) + Hp (b) can be carried out even if it is not known, whether this operator
belongs to Lp,q . Thus l p -versions of Theorem 2.9, Proposition 3.9 and Corol-
lary 3.10 can be directly proved.

7 Wiener-Hopf Plus Hankel Operators

This section is devoted to Wiener-Hopf plus Hankel operators. Let χE refer to the
characteristic function of the subset E ⊂ R and let F and F −1 be the direct and
144 V. D. Didenko and B. Silbermann

inverse Fourier transforms—i.e.

 ∞  ∞
−1 1
F ϕ(ξ ) := e ϕ(x) dx, F ψ(x) :=
iξ x
e−iξ x ψ(ξ ) dξ, x ∈ R.
−∞ 2π −∞

In what follows, we identify the spaces Lp (R+ ) and Lp (R− ), 1 ≤ p ≤ ∞ with the
subspaces P (Lp (R)) and Q(Lp (R)) of the space Lp (R), where P and Q are the
projections in Lp (R) defined by Pf := χR+ f and Q := I − P , respectively.
Consider the set G of functions g having the form

g(t) = (F k)(t) + gj eiδj t , t ∈ R, (7.1)
j ∈Z

where k ∈ L(R), δj ∈ R, gj ∈ C and the series in the right-hand side

of (7.1) is absolutely convergent. Any function g ∈ G generates an operator
W 0 (a) : Lp (R) → Lp (R) and operators W (g), H (g) : Lp (R+ ) → Lp (R+ )
defined by

W 0 (g) := F −1 gF ϕ, W (g) := P W 0 (g), H (g) := P W 0 (g)QJ.

Here and throughout this section, J : Lp (R) → Lp (R) is the reflection operator
defined by J ϕ :=  ϕ and 
ϕ (t) := ϕ(−t) for any ϕ ∈ Lp (R), p ∈ [1, ∞]. Operators
W (g) and H (g) are, respectively, called Wiener-Hopf and Hankel operators on the
semi-axis R+ . It is well-known [31] that for g ∈ G, all three operators above are
bounded on any space Lp , p ∈ [1, ∞).
In particular, the operator W (g) has the form
∞
  ∞
W (g)ϕ(t) = gj Bδj ϕ(t) + k(t − s)ϕ(s) ds, t ∈ R+ ,
j =−∞ 0

where

Bδj ϕ(t) = ϕ(t − δj ) if δj ≤ 0,


0, 0 ≤ t ≤ δj
Bδj ϕ(t) = if δj > 0.
ϕ(t − δj ), t > δj

Moreover, for g = F k the operator H (a) acts as

 ∞
H (g)ϕ(t) = k(t + s)ϕ(s) ds
0
Invertibility Issues for Toeplitz Plus Hankel Operators 145

and for g = eδt , one has H (a)ϕ(t) = 0 if δ ≤ 0 and


ϕ(δ − t), 0 ≤ t ≤ δ,
H (g)ϕ(t) =
0, t > δ,

if δ > 0.
For various classes of generating functions, the Fredholm properties of operators
W (a) are well studied [7, 8, 12, 24–26, 31]. In particular, Fredholm and semi-
Fredholm Wiener-Hopf operators are one-sided invertible, and for a ∈ G there
is efficient description of the kernels and cokernels of W (a) and formulas for the
corresponding one-sided inverses.
The study of Wiener-Hopf plus Hankel operators

W(a, b) = W (a) + H (b), a, b ∈ L∞ (R). (7.2)

is much more involved. The invertibility and Fredholmness of such operators in

the space L2 is probably first considered by Lebre et al. [43]. In particular, the
invertibility of W(a, b) has been connected with the invertibility of a block Wiener-
Hopf operator W (G). Assuming that a admits canonical Wiener-Hopf factorization
in L2 and the Wiener-Hopf factorization of the matrix G is known, Lebre et al.
provided a formula for W−1 (a, b). Nevertheless, the difficulties with Wiener-Hopf
factorization of matrix G influence the efficiency of the method. For piecewise
continuous generating functions, the conditions of Fredholmness are obtained in
[51]. A different method, called equivalence after extension, has been applied to
Wiener-Hopf plus Hankel operators W(a, a) = W (a) + H (a) with generating
function a belonging to various functional classes [9, 10]. The Fredholmness,
one-sided invertibility and invertibility of such operators are equivalent to the
corresponding properties of the Wiener-Hopf operator W (a a −1 ). Therefore, known
results about the invertibility and Fredholmness of W (a −1
a ) can be retranslated to
the operator W(a, a). On the other hand, the equivalence after extension has not
been used to establish any representations of the corresponding inverses. Another
approach has been exploited in [38, 39] to study the essential spectrum and the
index of the operators I + H (b).
We now consider Wiener-Hopf plus Hankel operators (7.2) with generating
functions a, b ∈ G satisfying the matching condition (2.4), where  a (t) and 
b(t)
denote the functions a(−t) and b(−t), respectively. Thus we now assume that

a(t)a(−t) = b(t)b(−t) (7.3)

and define the subordinated functions c and d by

c(t) := a(t)b−1 (t), d(t) := a(t)

b −1 (t) = a(t)b−1 (−t), t ∈ R.
146 V. D. Didenko and B. Silbermann

The assumption (7.3) allows us to employ the method developed in Sects. 2–5 and
establish necessary and also sufficient conditions for invertibility and one-sided
invertibility of the operators under consideration and obtain efficient representations
for the corresponding inverses. Note that here we only provide the main results. For
more details the reader can consult [14, 20, 23].
Let G+ ⊂ G and G− ⊂ G denote the sets of functions, which admit holomorphic
extensions to the upper and lower half-planes, respectively. If g ∈ G is a matching
function—i.e. g g = 1, then according to [14, 20], it can be represented in the form
   t − i n
−1
g(t) = σ (g) 
g+ (t) eiνt g+ (t),
t +i
±1
where ν = ν(g) ∈ R, n = n(g) ∈ N, σ (g) = (−1)n g(0),  g+ (t) ∈ G− and

g+ (∞) = 1. This representation is unique and the numbers ν(g) and n(g) in the
representation (7.1) are defined as follows:

1 1
ν(g) := lim [arg g(t)]l−l , n(g) := [arg(1 + g −1 (t)(F k)(t)]∞
t =−∞ .
l→∞ 2l 2π

Moreover, following the considerations of Sects. 2–3, for any right-invertible oper-
ator W (g) generated by a matching function g, we can introduce complementary
projections P±g on ker W (g). More precisely, if ν < 0 or if ν = 0 and n < 0, then
W (g) is right-invertible and the projections P±
g are defined by

P±
g := (1/2)(I ± J QP W (g)P ) : ker W (g) → ker W (g).
0

In addition, we also need the translation operator ϕ + defined by the subordinated

functions c and d. Assume that W (c) is right-invertible operator. Let Wr−1 (c) be any
of its right-inverses and consider the following function:

1 −1
ϕ + = ϕ + (a, b) := (W (c)W (ã −1 ) − J QW 0 (c)P Wr−1 (c)W (ã −1 ))
2 r
1
+ J QW 0 (ã −1 ),
2
We are ready to discuss the invertibility of Wiener-Hopf plus Hankel operators
starting with necessary conditions in the case where at least one of the indices
ν1 := ν(c) or ν2 := ν(d) is not equal to zero. The situation ν1 = ν2 = 0 will
be considered later on. Let n1 and n2 denote the indices n(c) and n(d), respectively.
Theorem 7.1 (VD and BS [23]) If a, b ∈ G, the operator W (a) + H (b) is one-
sided invertible in Lp (R+ ) and at least one of the indices ν1 or ν2 is not equal to
zero, then:
(i) Either ν1 ν2 ≥ 0 or ν1 > 0 and ν2 < 0.
(ii) If ν1 = 0 and ν2 > 0, then n1 > −1 or n1 = −1 and σ (c) = −1.
(iii) If ν1 < 0 and ν2 = 0, then n2 < 1 or n2 = 1 and σ (d) = −1.
Invertibility Issues for Toeplitz Plus Hankel Operators 147

Consider now the case ν1 > 0 and ν2 < 0 in more detail. It can be specified as
follows.
p
Theorem 7.2 (VD and BS [23]) Let ν1 > 0, ν2 < 0, n1 = n2 = 0 and Nν , ν > 0
denote the set of functions f ∈ Lp (R+ ) such that f (t) = 0 for t ∈ (0, ν).
(i) If the operator W (a) + H (b) : Lp (R+ ) → Lp (R+ ), 1 < p < ∞ is invertible
from the left, then

ϕ + (P+
p
d ) ∩ Nν1 /2 = {0},

where ϕ + = ϕ + (ae−iν1 t /2, beiν1 t /2 ).

(ii) If the operator W (a) + H (b) : Lp (R+ ) → Lp (R+ ), 1 < p < ∞ is invertible
from the right, then

ϕ + (P+
p
c ) ∩ N−ν2 /2 = {0}, (7.4)

where ϕ + = ϕ + (aeiν2 t /2, b̃e−iν2 t /2 ).

Passing to the case ν1 = ν2 = 0, we note that now the indices n1 and n2 take
over.
Theorem 7.3 (VD and BS [23]) Let a, b ∈ G, ν1 = ν2 = 0 and the operator
W (a) + H (b) is invertible from the left. Then:
(i) In the case n2 ≥ n1 , the index n1 satisfies the inequality

n1 ≥ −1

and if n1 = −1, then σ (c) = −1 and n2 > n1 .

(ii) In the case n1 > n2 , the index n1 satisfies the inequality

n1 ≥ 1,

and the index n2 is either non-negative or n2 < 0 and n1 ≥ −n2 .

The necessary conditions of the right-invertibility have similar form and we refer
the reader to [23] for details. The proof of the above results is based on the analysis
of the kernel and cokernel of the operator W (a)+T (b). In particular, if W (a)+T (b)
is left-invertible and one of the corresponding conditions is not satisfied, then this
operator should have a non-zero kernel, which cannot be true.
The sufficient conditions of one-sided invertibility can be also formulated in
terms of indices ν1 , ν2 , n1 and n2 . For example, the following theorem holds.
148 V. D. Didenko and B. Silbermann

Theorem 7.4 (VD and BS [23]) Let a, b ∈ G and indices ν1 , ν2 , n1 and n2 satisfy
any of the following conditions:
(i) ν1 < 0 and ν2 < 0.
(ii) ν1 > 0, ν2 < 0, n1 = n2 = 0, operator W (a) + H (b) is normally solvable
and satisfies the condition (7.4).
(iii) ν1 < 0, ν2 = 0 and n2 < 1 or n2 = 1 and σ (d) = −1.
(iv) ν1 = 0, n1 ≤ 0 and ν2 < 0.
(v) ν1 = 0 and ν2 = 0
(a) n1 ≤ 0, n2 < 1;
(b) n1 ≤ 0, n2 = 1 and σ (d) = −1;
Then the operator W (a) + H (b) is right-invertible.
The sufficient conditions for left-invertibility of W (a) + H (b) can be obtained
by passing to the adjoint operator. Here we are not going to discuss this problem
in whole generality. However, we use assumptions, which allow to derive simple
formulas for left- or right-inverses. These conditions are not necessary for one-
sided invertibility and the corresponding inverses can be also constructed even if
the conditions above are not satisfied.
Theorem 7.5 (VD and BS [23]) Let (a, b) ∈ G × G be a matching pair. Then:
1. If W (c) and W (d) are left-invertible, the operator W (a) + H (b) is also left-
invertible and one of its left-inverses has the form

(W (a) + H (b))−1 −1 −1 −1
l =Wl (c)W (ã )Wl (d)(I − H (d̃))

+ Wl−1 (c)H (ã −1). (7.5)

2. If W (c) and W (d) are right-invertible, the operator W (a) + H (b) is also right-
invertible and one of its right-inverses has the form

(W (a) + H (b))−1 −1 −1 −1
r =(I − H (c̃))Wr (c)W (ã )Wr (d)

+ H (a −1 )Wr−1 (d). (7.6)

For g ∈ G, the corresponding one-sided inverse of the operator W (g) can

be written by using Wiener-Hopf factorization of g [31]. We also note that if
W (c) and W (d) are invertible, formulas (7.5) and (7.6) can be used to write
the inverse operator for W (a) + H (b). They also play an important role when
establishing inverse operators in a variety of situations not covered by Theorem 7.5.
The corresponding proofs run similar to considerations of Sects. 4 and 5, but there
are essential technical differences because the corresponding kernel spaces can be
infinite dimensional.
It is worth noting that the generalized invertibility of the operators W (a) + H (b)
can also be studied—cf. [23].
Invertibility Issues for Toeplitz Plus Hankel Operators 149

8 Generalized Toeplitz Plus Hankel Operators

Here we briefly discuss generalized Toeplitz plus Hankel operators. These operators
are similar to classical Toeplitz plus Hankel operators considered in Sect. 2, but the
flip operator J is replaced by another operator Jα generated by a linear fractional
shift α. It turns out that the classical approach of Sects. 2–5 can also be used but
the application of the method is not straightforward and requires solving various
specific problems. Therefore, in this section we mainly focus on the description of
the kernels and cokernels of the corresponding operators. These results lay down
foundation for the invertibility study. We also feel that the Basor-Ehrhardt method
can be realized in this situation, but we are not going to pursue this problem here.
The following construction is based on the considerations of [18]. Let S denote
the Riemann sphere. We consider a mapping α : S → S defined by

z−β
α(z) := , (8.1)
βz − 1

where β is a complex number such that |β| > 1.

Let us recall basic properties of α.
1. The mapping α : S2 → S2 is one-to-one, α(T) = T, and if D := {z ∈ C : |z| <
1} is the interior of the unite circle T and D := D ∪ T, then

α(D) = S2 \ D, α(S2 \ D) = D.

We note that α is an automorphism of the Riemann sphere and the mappings

H p → H p , h → h ◦ α and H p → H p , h → h ◦ α are well-defined
isomorphisms.
2. The mapping α : T → T changes the orientation of T, satisfies the Carleman
condition α(α(t)) = t for all t ∈ T, and possesses two fixed points—viz.

t+ = (1 + λ)/β and t− = (1 − λ)/β, (8.2)


where λ := i |β|2 − 1.
3. The mapping α admits the factorization

α(t) = α+ (t) t −1 α− (t)

with the factorization factors

t −β λt
α+ (t) = , α− (t) = .
λ βt − 1
150 V. D. Didenko and B. Silbermann

4. On the space Lp , 1 < p < ∞, the mapping α generates a bounded linear

operator Jα , called weighted shift operator and defined by

Jα ϕ(t) := t −1 α− (t)ϕ(α(t)), t ∈ T.

Further, for a ∈ L∞ let aα denote the composition of the functions a and α—i.e.

aα (t) := a(α(t)), t ∈ T.

The operators Jα , P , Q and aI are connected with each other by the equations

Jα2 = I, Jα aI = aα Jα , Jα P = QJα , Jα Q = P Jα ,

and for any n ∈ Z, one has (a n )α = (aα )n := aαn . In addition to the Toeplitz
operator T (a), any element a ∈ L∞ defines another operator Hα := P αQJα ,
called generalized Hankel operator. Generalized Hankel operators are similar to
classical Hankel operators H (a). For example, analogously to (2.1) operators Hα
are connected with Toeplitz operators by the relations

T (ab) = T (a)T (b) + Hα (a)Hα (bα ),

Hα (ab) = T (a)Hα (b) + Hα (a)T (bα ).

On the space Lp , we now consider the operators of the form T (a) + Hα (b) and call
them generalized Toeplitz plus Hankel operators generated by the functions a, b and
the shift α.
The classical approach of Sect. 2 can be also employed to describe the kernels and
cokernels of T (a) + Hα (b). To this aim we develop a suitable framework, which is
not a straightforward extension of the methods of Sect. 2. Let us now assume that
a belongs to the group of invertible elements GL∞ and the duo a, b satisfy the
condition

aaα = bbα . (8.3)

Relation (8.3) is again called the matching condition and the duo (a, b) are α-
matching functions. To each matching pair, one can assign another α-matching
pair (c, d) := (ab−1, abα−1 ) called the subordinated pair for (a, b). It is easily
seen that ccα = ddα = 1. In what follows, any element g ∈ L∞ satisfying the
relation ggα = 1 is referred to as α-matching function. The functions c and d can
also be expressed in the form c = bα aα−1 , d = bα−1 a. Besides, if (c, d) is the
subordinated pair for an α-matching pair (a, b), then (d, c) is the subordinated pair
for the matching pair (a, bα ), which defines the adjoint operator

(T (a) + Hα (b))∗ = T (a) + Hα (bα ).

Invertibility Issues for Toeplitz Plus Hankel Operators 151

Rewrite the operator Jα : Lp → Lp in the form

Jα ϕ(t) := χ −1 (t)ϕ(α(t)),

where χ(t) = t/(α− (t)), and note that if α is the shift (8.1), then:
1. χ ∈ H ∞ is an α-matching function and wind χ = 1, where wind χ denotes the
winding number of the function χ with respect to the origin.
2. The function χα ∈ H ∞ and χα (∞) = 0.
3. If a, b ∈ L∞ and n is a positive integer, then

T (a) + Hα (b) = (T (aχ −n ) + Hα (bχ n ))T (χ n ).

These properties allow us to establish the following version of the Coburn-

Simonenko Theorem for generalized Toeplitz plus Hankel operators.
Theorem 8.1 Let a ∈ GL∞ and let A denote any of the operators T (a) −
Hα (aχ −1 ), T (a) + Hα (aχ), T (a) + Hα (a), T (a) − Hα (a). Then

min{dim ker A, dim coker A} = 0.

Further development runs similar to the one presented in Sect. 2.3 and all results
are valid if the operator H (b) is replaced by Hα (b),  a by aα and b by bα . For
example, if (a, b) is an α-matching pair with the subordinated pair (c, d), then the
block Toeplitz operator T (V (a, b)) with the matrix function
 
0 d
V (a, b) = ,
−c aα−1

can be represented in the form

 
0 T (d)
T (V (a, b)) =
−T (c) T (aα−1 )
   
−T (d) 0 0 −I −T (c) 0
= .
0 I I T (aα−1 ) 0 I

Moreover, assuming that T (c) is invertible from the right and Tr−1 (c) is one of the
right inverses, we can again represent the kernel of T (V (a, b)) as the direct sum

(d),
ker T (V (a, b)) = (c)  
152 V. D. Didenko and B. Silbermann

where
 
(c) := (ϕ, 0)T : ϕ ∈ ker T (c) ,
 
(d) := (Tr−1 (c)T (aα−1 )s, s)T : s ∈ ker T (d) .


Further, we observe that the operators

1
P±
α (c) := (I ± Jα QcP ) : ker T (c) → ker T (c),
2
1
P±
α (d) := (I ± Jα QdP ) : ker T (d) → ker T (d),
2
are projections on the corresponding spaces and consider the translation operators

ϕα± : P±
α (d) → ker(T (a) ± Hα (b)),

which are defined similar to (3.5), but with J and  a −1 replaced by Jα and aα−1 ,
respectively.
In this situation, Proposition 3.6 reads as follows.
Proposition 8.2 Assume that (a, b) ∈ L∞ × L∞ is a Fredholm matching pair. If
the operator T (c) is right-invertible, then

ker(T (a) + Hα (b)) = im P− + +

α (c)  ϕα (im Pα (d)),

ker(T (a) − Hα (b)) = im P+ − −

α (c)  ϕα (im Pα (d)).

Thus there is a one-to-one correspondence between ker T (V (a, b)) and

ker diag (T (a) + Hα (b), T (a) − Hα (b)) and in order to establish the latter, we
need an efficient description of the spaces im P+ +
α (c) and im Pα (d).
Let g stand for one of the functions c or d. Then g is a matching function and
T (g) is a Fredholm Toeplitz operator. We may assume that ind T (g) > 0, since
only in this case at least one of the spaces im P+ −
α (g) or im Pα (g) is non-trivial. The
following factorization result is crucial for the description of im P±α (g).

Theorem 8.3 If g ∈ L∞ satisfies the matching condition ggα = 1 and wind g = n,

n ∈ Z, then g can be represented in the form
−1
g = ξg+ χ −n (g+ )α ,

where g+ and n occur in the Wiener-Hopf factorization

g = g− t −n g+ , g− (∞) = 1,
Invertibility Issues for Toeplitz Plus Hankel Operators 153

of the function g, whereas ξ ∈ {−1, 1} and is defined by

 n  
λ −1 1
ξ= g+ . (8.4)
β β

Definition 8.4 The number ξ in (8.4) is called the α-factorization signature, or

simply, α-signature of g and is denoted by σα (g).
The α-signature is used to describe the kernels of the operators T (a) + Hα (b)
and for certain classes of generating functions it can be defined with relative ease.
For example, assuming that the operator T (g) is Fredholm, n := ind T (g) and g is
Hölder continuous at the fixed point t+ or t− of (8.2), one can show that σα (g) =
g(t+ ) or σα (g) = g(t− )(−1)n . For piecewise continuous functions, the situation is
more complicated but still can be handled—cf. [18].
Theorem 8.5 Let g ∈ L∞ be an α-matching function such that the operator T (g) :
H p → H p is Fredholm and n := ind T (g) > 0. If g = g− t −n g+ , g− (∞) = 1 is
the corresponding Wiener–Hopf factorization of g in H p , then the following systems
of functions Bα± (g) form bases in the spaces im P±
α (g):

1. If n = 2m, m ∈ N, then
−1 m−k−1
Bα± (g) := {g+ (χ ± σα (g)χ m+k ) : k = 0, 1, · · · , m − 1},

and dim im P±
α (g) = m.
2. If n = 2m + 1, m ∈ Z+ , then
−1 m+k
Bα± (g) := {g+ (χ ± σα (g)χ m−k ) : k = 0, 1, · · · , m} \ {0},
dim im P±
α (g) = m + (1 ± σα (g))/2 ,

and the zero element belongs to one of the sets Bα+ (g) or Bα− (g)—viz. for k = 0
one of the terms χ m (1 ± σα (g)) is equal to zero.
The above considerations provide a powerful tool for the study of invertibility of
generalized Toeplitz plus Hankel operators and it should be clear that the results
obtained in Sects. 4 and 5 can be extended to this class of operators. However,
additional studies may be needed in certain situations. Nevertheless, let us mention
one of such results here.
Proposition 8.6 Assume that (a, b) ∈ L∞ × L∞ is a Fredholm matching pair and
the operators T (c) and T (d) are right-invertible. Then T (a) + Hα (b) and T (a) −
Hα (b) are also right-invertible and corresponding right inverses are given by

(T (a) ± Hα (b))−1 −1 −1 −1 −1 −1
r = (I ∓ Hα (cα ))Tr (c)T (aα )Tr (d) ± Hα (a )Tr (d).
154 V. D. Didenko and B. Silbermann

In conclusion of this section, we note the works [40, 41] where more general
operators with the shift (8.1) are considered. However, the conditions imposed on
coefficient functions are more restrictive and the results obtained are less complete.

9 Final Remarks

We considered various approaches to the study of invertibility of Toeplitz plus

Hankel operators and their close relatives. Before concluding this survey, we
would like to mention two more problems of special interest. The first one is the
construction of Wiener-Hopf factorizations for multipliers acting on the spaces
l p and also for those on Lp (R)-spaces. Some ideas on how to proceed with this
problem in the l p -context have been noted in [21].
Another problem of interest is the study of the kernels and cokernels of Wiener-
Hopf plus Hankel operators acting on Lp (R)-spaces, p = 2 and generated by
multipliers from sets more involved than the set G considered in Sect. 7, such as
piecewise continuous multipliers, for example.
In should be clear that the list of open problems in the theory of Toeplitz plus
Hankel operators is not limited to those mentioned in this work and it is up to the
interested reader to single out a one for further consideration.

Acknowledgments The authors express their sincere gratitude to anonymous referees for insight-
ful comments and suggestions that helped to improve the paper.

References

1. J. Baik, E.M. Rains, Algebraic aspects of increasing subsequences. Duke Math. J. 109, 1–65
(2001)
2. E.L. Basor, T. Ehrhardt, On a class of Toeplitz + Hankel operators. New York J. Math. 5, 1–16
(1999)
3. E.L. Basor, T. Ehrhardt, Factorization theory for a class of Toeplitz + Hankel operators. J.
Oper. Theory 51, 411–433 (2004)
4. E.L. Basor, T. Ehrhardt, Fredholm and invertibility theory for a special class of
Toeplitz + Hankel operators. J. Spectral Theory 3, 171–214 (2013)
5. E.L. Basor, T. Ehrhardt, Asymptotic formulas for determinants of a special class of
Toeplitz + Hankel matrices, in Large Truncated Toeplitz Matrices, Toeplitz Operators, and
Related Topics, vol. 259. The Albrecht Böttcher Anniversary Volume. Operator Theory:
Advances and Applications (Birkhäuser, Basel, 2017), pp. 125–154
6. E. Basor, Y. Chen, T. Ehrhardt, Painlevé V and time-dependent Jacobi polynomials. J. Phys. A
43, 015204, 25 (2010)
7. A. Böttcher, B. Silbermann, Analysis of Toeplitz Operators, 2nd edn. Springer Monographs in
Mathematics (Springer, Berlin, 2006)
8. A. Böttcher, Y.I. Karlovich, I.M. Spitkovsky, Convolution Operators and Factorization of
Almost Periodic Matrix Functions (Birkhäuser, Basel, 2002)
Invertibility Issues for Toeplitz Plus Hankel Operators 155

9. L.P. Castro, A.P. Nolasco, A semi-Fredholm theory for Wiener-Hopf-Hankel operators with
piecewise almost periodic Fourier symbols. J. Oper. Theory 62, 3–31 (2009)
10. L.P. Castro, A.S. Silva, Wiener-Hopf and Wiener-Hopf-Hankel operators with piecewise-
almost periodic symbols on weighted Lebesgue spaces. Mem. Diff. Equ. Math. Phys. 53, 39–62
(2011)
11. K. Clancey, I. Gohberg, Factorization of Matrix Functions and Singular Integral Operators
(Birkhäuser, Basel, 1981)
12. L.A. Coburn, R.G. Douglas, Translation operators on the half-line. Proc. Nat. Acad. Sci. USA
62, 1010–1013 (1969)
13. V.D. Didenko, B. Silbermann, Index calculation for Toeplitz plus Hankel operators with
piecewise quasi-continuous generating functions. Bull. London Math. Soc. 45, 633–650 (2013)
14. V.D. Didenko, B. Silbermann, The Coburn-Simonenko Theorem for some classes of Wiener–
Hopf plus Hankel operators. Publ. de l’Institut Mathématique 96(110), 85–102 (2014)
15. V.D. Didenko, B. Silbermann, Some results on the invertibility of Toeplitz plus Hankel
operators. Ann. Acad. Sci. Fenn. Math. 39, 443–461 (2014)
16. V.D. Didenko, B. Silbermann, Structure of kernels and cokernels of Toeplitz plus Hankel
operators. Integr. Equ. Oper. Theory 80, 1–31 (2014)
17. V.D. Didenko, B. Silbermann, Generalized inverses and solution of equations with Toeplitz
plus Hankel operators. Bol. Soc. Mat. Mex. 22, 645–667 (2016)
18. V.D. Didenko, B. Silbermann, Generalized Toeplitz plus Hankel operators: kernel structure
and defect numbers. Compl. Anal. Oper. Theory 10, 1351–1381 (2016)
19. V.D. Didenko, B. Silbermann, Invertibility and inverses of Toeplitz plus Hankel operators. J.
Oper. Theory 72, 293–307 (2017)
20. V.D. Didenko, B. Silbermann, Kernels of Wiener-Hopf plus Hankel operators with matching
generating functions, in Recent Trends in Operator Theory and Partial Differential Equations,
vol. 258. The Roland Duduchava Anniversary Volume. Operator Theory: Advances and
Applications (Birkhäuser, Basel, 2017), pp. 111–127
21. V.D. Didenko, B. Silbermann, Kernels of a class of Toeplitz plus Hankel operators with
piecewise continuous generating functions, in Contemporary Computational Mathematics –
A Celebration of the 80th Birthday of Ian Sloan, ed. by J. Dick, F.Y. Kuo, H. Woźniakowski
(eds). (Springer, Cham, 2018), pp. 317–337
22. V.D. Didenko, B. Silbermann, The invertibility of Toeplitz plus Hankel operators with
subordinated operators of even index. Linear Algebra Appl. 578, 425–445 (2019)
23. V.D. Didenko, B. Silbermann, Invertibility issues for a class of Wiener-Hopf plus Hankel
operators. J. Spectral Theory 11 (2021)
24. R.V. Duduchava, Wiener-Hopf integral operators with discontinuous symbols. Dokl. Akad.
Nauk SSSR 211, 277–280 (1973) (in Russian)
25. R.V. Duduchava, Integral operators of convolution type with discontinuous coefficients. Math.
Nachr. 79, 75–98 (1977)
26. R.V. Duduchava, Integral Equations with Fixed Singularities (BSB B. G. Teubner Verlagsge-
sellschaft, Leipzig, 1979)
27. R. Edwards, Fourier Series. A Modern Introduction, vol. 1. Graduate Texts in Mathematics,
vol. 85 (Springer, Berlin, 1982)
28. T. Ehrhardt, Factorization theory for Toeplitz+Hankel operators and singular integral opera-
tors with flip. Habilitation Thesis, Technische Universität Chemnitz (2004)
29. T. Ehrhardt, Invertibility theory for Toeplitz plus Hankel operators and singular integral
operators with flip. J. Funct. Anal. 208, 64–106 (2004)
30. P.J. Forrester, N.E. Frankel, Applications and generalizations of Fisher-Hartwig asymptotics.
J. Math. Phys. 45, 2003–2028 (2004)
31. I.C. Gohberg, I.A. Feldman, Convolution Equations and Projection Methods for Their Solution
(American Mathematical Society, Providence, 1974)
32. I. Gohberg, N. Krupnik, One-Dimensional Linear Singular Integral Equations. I, vol. 53.
Operator Theory: Advances and Applications (Birkhäuser Verlag, Basel, 1992)
156 V. D. Didenko and B. Silbermann

33. I. Gohberg, N. Krupnik, One-Dimensional Linear Singular Integral Equations. II, vol. 54.
Operator Theory: Advances and Applications (Birkhäuser Verlag, Basel, 1992)
34. S. Grudsky, A. Rybkin, On positive type initial profiles for the KdV equation. Proc. Am. Math.
Soc. 142, 2079–2086 (2014)
35. S. Grudsky, A. Rybkin, Soliton theory and Hankel operators. SIAM J. Math. Anal. 47, 2283–
2323 (2015)
36. S.M. Grudsky, A.V. Rybkin, On the trace-class property of Hankel operators arising in the
theory of the Korteweg-de Vries equation. Math. Notes 104, 377–394 (2018)
37. P. Junghanns, R. Kaiser, A note on Kalandiya’s method for a crack problem. Appl. Numer.
Math. 149, 52–64 (2020)
38. N.K. Karapetiants, S.G. Samko, On Fredholm properties of a class of Hankel operators. Math.
Nachr. 217, 75–103 (2000)
39. N.K. Karapetiants, S.G. Samko, Equations with Involutive Operators (Birkhäuser Boston Inc.,
Boston, 2001)
40. V.G. Kravchenko, A.B. Lebre, J.S. Rodríguez, Factorization of singular integral operators with
a Carleman shift via factorization of matrix functions: the anticommutative case. Math. Nachr.
280, 1157–1175 (2007)
41. V.G. Kravchenko, A.B. Lebre, J.S. Rodríguez, Factorization of singular integral operators with
a Carleman backward shift: the case of bounded measurable coefficients. J. Anal. Math. 107,
1–37 (2009)
42. N.Y. Krupnik, Banach Algebras with Symbol and Singular Integral Operators, vol. 26.
Operator Theory: Advances and Applications (Birkhäuser Verlag, Basel, 1987)
43. A.B. Lebre, E. Meister, F.S. Teixeira, Some results on the invertibility of Wiener-Hopf-Hankel
operators. Z. Anal. Anwend. 11, 57–76 (1992)
44. G.S. Litvinchuk, I.M. Spitkovskii, Factorization of Measurable Matrix Functions, vol. 25.
Operator Theory: Advances and Applications (Birkhäuser Verlag, Basel, 1987)
45. E. Meister, F. Penzel, F.-O. Speck, F.S. Teixeira, Two-media scattering problems in a half-
space, in Partial Differential Equations with Real Analysis. Dedicated to Robert Pertsch Gilbert
on the occasion of his 60th birthday (Longman Scientific & Technical, Harlow; John Wiley &
Sons, Inc., New York, 1992), pp. 122–146
46. E. Meister, F.-O. Speck, F.S. Teixeira, Wiener-Hopf-Hankel operators for some wedge
diffraction problems with mixed boundary conditions. J. Integral Equ. Appl. 4, 229–255 (1992)
47. V.V. Peller, Hankel Operators and Their Applications. Springer Monographs in Mathematics
(Springer, New York, 2003)
48. S.C. Power, C*-algebras generated by Hankel operators and Toeplitz operators. J. Funct. Anal.
31, 52–68 (1979)
49. S. Roch, B. Silbermann, Algebras of convolution operators and their image in the Calkin
algebra, vol. 90. Report MATH (Akademie der Wissenschaften der DDR Karl-Weierstrass-
Institut für Mathematik, Berlin, 1990)
50. S. Roch, B. Silbermann, A handy formula for the Fredholm index of Toeplitz plus Hankel
operators. Indag. Math. 23, 663–689 (2012)
51. S. Roch, P.A. Santos, B. Silbermann, Non-Commutative Gelfand Theories. A Tool-Kit for
Operator Theorists and Numerical Analysts. Universitext (Springer, London, 2011)
52. B. Silbermann, The C ∗ -algebra generated by Toeplitz and Hankel operators with piecewise
quasicontinuous symbols. Integr. Equ. Oper. Theory 10, 730–738 (1987)
53. I.B. Simonenko, Some general questions in the theory of Riemann boundary problem. Math.
USSR Izvestiya 2, 1091–1099 (1968)
54. I.J. Šneı̆berg, Spectral properties of linear operators in interpolation families of Banach spaces.
Mat. Issled. 9, 2(32), 214–229 (1974) (in Russian)
55. I.M. Spitkovskiı̆, The problem of the factorization of measurable matrix-valued functions.
Dokl. Akad. Nauk SSSR 227, 576–579 (1976) (in Russian)
56. F.S. Teixeira, Diffraction by a rectangular wedge: Wiener-Hopf-Hankel formulation. Integr.
Equ. Oper. Theory 14, 436–454 (1991)
K-Inner Functions and K-Contractions

Jörg Eschmeier and Sebastian Toth

Abstract For a large class of unitarily invariant reproducing kernel functions

K on the unit ball Bd in Cd , we characterize the K-inner functions on Bd as
functions admitting a suitable transfer function realization. We associate with each
K-contraction T ∈ L(H )d a canonical operator-valued K-inner function and extend
a uniqueness theorem of Arveson for minimal K-dilations to our setting. We thus
generalize results of Olofsson for m-hypercontractions on the unit disc and of the
first named author for m-hypercontractions on the unit ball.

Keywords K-inner functions · K-contractions · Wandering subspaces

Mathematics Subject Classification (2010) Primary 47A13; Secondary

47A20,47A45, 47A48

1 Introduction

Let Bd ⊂ Cd be the open Euclidean unit ball and let

∞

k : D → C, k(z) = an z n
n=0

be an analytic function without zeros on the unit disc D in C such that a0 = 1, an >
0 for all n ∈ N and such that
an an
0 < inf ≤ sup < ∞.
n∈N an+1 n∈N an+1

J. Eschmeier () · S. Toth

Fachrichtung Mathematik, Universität des Saarlandes, Saarbrücken, Germany
e-mail: [email protected]; [email protected]

© Springer Nature Switzerland AG 2021 157

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_8
158 J. Eschmeier and S. Toth

Since k has no zeros, the reciprocal function 1/k ∈ O(D) admits a Taylor expansion
∞

(1/k)(z) = cn z n (z ∈ D).
n=0

The reproducing kernel

K : Bd × Bd → C, K(z, w) = k(z, w)

defines an analytic functional Hilbert space HK such that the row operator
Mz : HKd → HK is bounded and has closed range [2, Theorem A.1]. Typical
examples of functional Hilbert spaces of this type on the unit ball Bd are the Drury–
Arveson space, the Dirichlet space, the Hardy space and the weighted Bergman
spaces.
Let T = (T1 , . . . , Td ) ∈ L(H )d be a commuting tuple of bounded linear
 Hilbert space H and let σT : L(H ) → L(H ) be the map
operators on a complex
defined by σT (X) = di=1 Ti XTi∗ . The tuple T is called a K-contraction if the limit

∞
 
1
(T ) = SOT− cn σTn (1H ) = SOT− lim c|α| γα T α T ∗α
K N→∞
n=0 |α|≤N

exists and defines a positive operator. Here and in the following we use the
abbreviation γα = |α|!/α! for α ∈ Nd .
If K(z, w) = 1/(1 − z, w) is the Drury–Arveson kernel, then under a
natural pureness condition the K-contractions coincide with the commuting row
contractions of class C·0 . If m is a positive integer and Km (z, w) = 1/(1 −z, w)m ,
then the pure Km -contractions are precisely the row-m-hypercontractions of class
C·0 [17, Theorem 3.49] and [12, Lemma 2].
An operator-valued analytic function W : Bd → L(E∗ , E ) with Hilbert spaces
E and E∗ is called K-inner if the map E∗ → HK (E ), x → W x, is a well-defined
isometry and

(W E∗ ) ⊥ Mzα (W E∗ ) for all α ∈ Nd \ {0}.

Here HK (E ) is the E -valued functional Hilbert space on Bd with reproducing kernel

KE : Bd × Bd → L(E ), (z, w) → K(z, w)1E . A well known more explicit
description of the functional Hilbert space HK (E ) is given by
   fα 2
HK (E ) = f = fα zα ∈ O(Bd , E ); f 2 = <∞ .
a γ
d |α| α
α∈Nd α∈N
K-Inner Functions and K-Contractions 159

It was shown by Olofsson [13] that, for d = 1 and the Bergman-type kernel

1
Km : D × D → C, Km (z, w) = (m ∈ N \ {0}),
(1 − zw)m

the Km -inner functions W : D → L(E∗ , E ) are precisely the functions of the form


m
W (z) = D + C (1 − zT ∗ )−k B,
k=1

where T ∈ L(H ) is a pure m-hypercontraction on some Hilbert space H and B ∈

L(E∗ , H ), C ∈ L(H, E ) and D ∈ L(E∗ , E ) are bounded operators satisfying the
operator equations

C ∗ C = (1/Km )(T ), D ∗ C + B ∗ !T T ∗ = 0, D ∗ D + B ∗ !T B = 1E∗ .

Here (1/Km )(T ) is the m-th order defect operator of T and


m−1  
m
!T = (−1)k T k T ∗k .
k+1
k=0

The single variable results of Olofsson were extended in a system theoretic

framework byBall and Bolotnikov in [6] (see also [5]) to a large class of kernels
K(z, w) = n≥0 an (zw) on the unit disc D in C. In [6] Beurling–Lax type
n

representations for Mz -invariant subspaces M ⊂ Hk (E ) in terms of suitable K-

inner function families (θk )k≥0 together with transfer function realizations for the
functions θk are obtained and operator models using K-inner characteristic function
families are developed for quite general kernels K on the unit disc.
In [10] the result of Olofsson was extended to the unit ball by showing that a
corresponding characterization holds for functions W : Bd → L(E∗ , E ) that are
Km -inner with respect to the generalized Bergman kernels

Km : Bd × Bd → C, Km (z, w) = 1/(1 − z, w)m (m ∈ N \ {0}).

In the present note we show that results of Olofsson from [13, 14] hold true for a
large class of kernels
∞

K : Bd × Bd → C, K(z, w) = an z, wn
n=0

including all complete Nevanlinna–Pick kernels and powers Kν (z, w) = 1/(1 −

z, w)ν of the Drury–Arveson kernel with positive real exponents. To prove that
each K-inner function admits a transfer function realization as described above we
160 J. Eschmeier and S. Toth

extend a uniqueness result for minimal K-dilations due to Arveson to our class of
kernels.

2 Wandering Subspaces

Let T = (T1 , . . . , Td ) ∈ L(H )d be a K-contraction, that is, a commuting tuple of

bounded linear operators on a complex Hilbert space H such that the limit
∞
 
1
(T ) = SOT− cn σTn (1H ) = SOT− lim c|α| γα T α T ∗α
K N→∞
n=0 |α|≤N

exists and defines a positive operator. A K-contraction T ∈ L(H )d is said to be

pure if


N
1
SOT− lim 1H − an σTn ( (T )) = 0.
N→∞ K
n=0

Let us define the defect operator and the defect space of a K-contraction T by

1 1
C= (T ) 2 and D = Im C.
K

We call an isometric linear map j : H → HK (E ) which intertwines the tuples T ∗ ∈

L(H )d and Mz∗ ∈ L(HK (E ))d componentwise a K-dilation of T . By definition
a K-dilation j : H → HK (E ) is minimal if the only reducing subspace of Mz ∈
L(HK (E ))d that contains the image of j is HK (E ).
Exactly as for row-m-hypercontractions of class C.0 , one can construct a
canonical K-dilation for each K-contraction.
Theorem 2.1 Let T ∈ L(H )d be a pure K-contraction. Then

j : H → HK (D), j (h) = a|α| γα CT ∗α hzα
α∈Nd

is a well-defined isometry such that j Ti∗ = Mz∗i j for i = 1, . . . , d.

For a proof, see [17, 
Theorem 2.15].
For h ∈ H and f = α∈Nd fα zα ∈ HK (D),
       
h, j ∗ f = CT ∗α h, fα = h, T α Cfα .
α∈Nd α∈Nd
K-Inner Functions and K-Contractions 161

An application of the uniform boundedness principle shows that the adjoint

j ∗ : HK (D) → H of the isometry j acts as
⎛ ⎞
 
j∗ ⎝ fα zα ⎠ = T α Cfα .
α∈Nd α∈Nd

Since j intertwines T ∗ and Mz∗ componentwise, the space

M = HK (D)  Im j ⊂ HK (D)

is invariant for Mz ∈ L(HK (D))d .

In the following we show that the wandering subspace of Mz restricted to M can
be described in terms of a suitable K-inner function. Recall that a closed subspace
W ⊂ H is called a wandering subspace for a commuting tuple S ∈ L(H )d if

W ⊥ Sα W (α ∈ Nd \ {0}).

The
6 space W is called a generating wandering subspace for S if in addition H =
(S α W; α ∈ Nd ). For each closed S-invariant subspace L ⊂ H , the space


d
WS (L) = L  Si L
i=1

is a wandering subspace for S, usually called the wandering subspace associated

with S on L. If W is a generating wandering subspace for S, then an elementary
argument shows that necessarily W = WS (H ).
In the following we write


d
W (M) = M  Mzi M
i=1

for the wandering subspace associated with the restriction of Mz to the invariant
subspace M = HK (D)  Im j . Our main tool will be the matrix operator

Mz∗ Mz = (Mz∗i Mzj )1≤i,j ≤d ∈ L(HK (D)d ).

Since the row operator Mz : HK (D)d → HK (D) has closed range [2, Theorem
A.1], the operator

Mz∗ Mz : Im Mz∗ → Im Mz∗

162 J. Eschmeier and S. Toth

is invertible. We denote its inverse by (Mz∗ Mz )−1 . In the following we make

essential use of the operators
⎛ ⎞
∞  ∞
 an 
δ : HK (D) → HK (D), δ⎝ fα zα ⎠ = f0 + fα zα
an−1
n=0 |α|=n n=1 |α|=n

and
⎛ ⎞
∞  ∞
 an+1 
! : HK (D) → HK (D), !⎝ fα zα ⎠ = fα zα .
an
n=0 |α|=n n=0 |α|=n

By definition δ and ! 7 are diagonal operators with respect to the orthogonal

∞
decomposition HK (D) = n=0 Hn (D) of HK (D) into the spaces Hn (D) of all
D-valued homogenous polynomials of degree n. Our hypotheses on the sequence
(an /an+1 ) imply that δ and ! are invertible positive operators on HK (D). An
elementary calculation shows that

δMzi = Mzi !

for i = 1, . . . , d.
Lemma 2.2 For f ∈ HK (D), we have

(Mz∗ Mz )−1 (Mz∗ f ) = Mz∗ δf = (⊕!) Mz∗ f.

In particular the row operator

δMz : HK (D)d → HK (D)

defines the trivial extension of the operator

−1
Mz Mz∗ Mz : Im Mz∗ → HK (D).

Proof Since the column operator Mz∗ annihilates the constant functions, to prove
suppose that f (0) = 0. With respect to the orthogonal
the first identity, we may 7
decomposition HK (D) = ∞ ∗
n=0 Hn (D) the operator Mz Mz acts as (Lemma 4.3 in
[11])
∞ ∞  
  an−1
Mz Mz∗ fn = fn .
an
n=0 n=1
K-Inner Functions and K-Contractions 163

Hence Mz Mz∗ δf = f and

−1 −1 
Mz∗ Mz Mz∗ f = Mz∗ Mz Mz∗ Mz Mz∗ δf = Mz∗ δf = (⊕!)Mz∗ f.

Since any two diagonal operators commute, it follows in particular that

−1 ∗ 
Mz Mz∗ Mz Mz = δ Mz Mz∗ . Thus also the second assertion follows.
The preceding proof shows in particular that the orthogonal projection of HK (D)
onto Im Mz acts as

PIm Mz = Mz (Mz∗ Mz )−1 Mz∗ = δ(Mz Mz∗ ) = PHK (D )D ,

where D ⊂ HK (D) is regarded as the closed subspace consisting of all constant

functions. As in the single-variable case we call the operator defined by Mz =
δMz ∈ L(HK (D)d , HK (D)) the Cauchy dual of the multiplication tuple Mz .
We use the operator !T ∈ L(H ) defined by

!T = j ∗ !j

to give a first description of the wandering subspace W (M) of Mz restricted to the

invariant subspace M = (Im j )⊥ .
Theorem 2.3 A function f ∈ HK (D) is an element of the wandering subspace
W (M) of M = (Im j )⊥ ∈ Lat(Mz , HK (D)) if and only if

f = f0 + Mz (j xi )di=1

for some vectors f0 ∈ D, x1 , . . . , xd ∈ H with (j xi )di=1 ∈ Mz∗ HK (D) and

Cf0 + T (!T xi )di=1 = 0.

In this case (j xi )di=1 = Mz∗ f .

Proof Note that  a function f ∈ HK (D) belongs to the wandering subspace
W (M) = M  di=1 zi M of Mz on M = Ker j ∗ ∈ Lat(Mz , HK (D)) if and
only if j ∗ f = 0 and (1HK (D ) − jj ∗ )Mz∗i f = 0 for i = 1, . . . , d. Using the
remark following Lemma 2.2, we obtain, for (xi )di=1 ∈ H d and f ∈ HK (D) with
(j xi )di=1 = Mz∗ f ,

j ∗ f = j ∗ (f (0) + δMz Mz∗ f )

= Cf (0) + j ∗ Mz (!j xi )di=1
= Cf (0) + T (j ∗ !j xi )di=1
= Cf (0) + T (!T xi )di=1 .
164 J. Eschmeier and S. Toth

Thus if f ∈ W (M), then (xi )di=1 = (j ∗ Mz∗i f )di=1 defines a tuple in H d with
(j xi )di=1 = Mz∗ f such that Cf (0) + T (!T xi )di=1 = j ∗ f = 0 and

f = f (0) + (f − f (0)) = f (0) + Mz (Mz∗ Mz )−1 Mz∗ f = f (0) + Mz (j xi )di=1 .

Conversely, if f = f0 + Mz (j xi )di=1 with f0 ∈ D, x1 , . . . , xd as in Theorem 2.3,

then using Lemma 2.2 we find that

Mz∗ f = Mz∗ Mz (Mz∗ Mz )−1 (j xi )di=1 = (j xi )di=1 .

Since j is an isometry, it follows that jj ∗ Mz∗i f = j xi = Mz∗i f for i = 1, . . . , d.

Since j ∗ f = Cf (0) + T (!T xi )di=1 = 0, we have shown that f ∈ W (M).
Lemma 2.4 Let T ∈ L(H )d be a pure K-contraction and let

f = f0 + Mz (j xi )di=1

be a representation of a function f ∈ W (M) as in Theorem 2.3. Then we have


d
f 2 = f0 2 + !T xi , xi .
i=1

Proof Since by Lemma 2.2

Im Mz = Mz (Mz∗ Mz )−1 Mz∗ HK (D) = Im Mz = HK (D)  D,

it follows that

f 2 − f0 2 = Mz (j xi )di=1 2

= (Mz∗ Mz )−1 Mz∗ f, (j xi )di=1 
= (⊕j ∗ )Mz∗ δf, (xi )di=1 
= (j ∗ !j xi )di=1 , (xi )di=1 .

Since by definition !T = j ∗ !j , the assertion follows.

Let T ∈ L(H )d be a pure K-contraction. Then !T = j ∗ !j is a positive operator
with
1 1
!T x, x = ! 2 j x2 ≥ !− 2 −2 j x2 = !−1 −1 x2
K-Inner Functions and K-Contractions 165

for all x ∈ H . Hence !T ∈ L(H ) is invertible and

(x, y) = !T x, y

defines a scalar product on H such that the induced norm  · T is equivalent to the
original norm with

! 2 x ≥ xT ≥ !− 2 −1 x

1 1

for x ∈ H . We write H̃ for H equipped with the norm  · T . Then

IT : H → H̃ , x → x

is an invertible bounded operator such that

IT∗ x, y = !T x, y (x ∈ H̃ , y ∈ H ).

Hence IT∗ x = !T x for x ∈ H̃ . Let T̃ = (T̃1 , . . . , T̃d ) : H̃ d → H be the row

operator with components T̃i = Ti ◦ IT∗ ∈ L(H̃ , H ). Then


d
T̃ T̃ ∗ = Ti (IT∗ IT )Ti∗ = σT (!T ) = σT (j ∗ !j ) = j ∗ Mz (⊕!)Mz∗ j
i=1

= j ∗ (δMz Mz∗ )j = j ∗ PHK (D )D j

and hence T̃ is a contraction. As in [13] we use its defect operators

DT˜ = (1H̃ d − T̃ ∗ T̃ )1/2 ∈ L(H̃ d ),

DT̃ ∗ = (1H − T̃ T̃ ∗ )1/2 = (j ∗ PD j )1/2 = C ∈ L(H ).

Here the identity (j ∗ PD j )1/2 = C follows from the definition of j and the
representation of j ∗ explained in the section following Theorem 2.1. We write
DT̃ = DT̃ H̃ d ⊂ H̃ d and DT̃ ∗ = DT̃ ∗ H = D for the defect spaces of T̃ . As in
the classical single-variable theory of contractions it follows that T̃ DT̃ = DT̃ ∗ T̃
and that
 
T̃ DT̃ ∗
U= : H̃ d ⊕ DT̃ ∗ → H ⊕ DT̃
DT̃ −T̃ ∗

is a well-defined unitary operator. In the following we define an analytically

parametrized family WT (z) ∈ L(D̃, D) (z ∈ B) of operators on the closed subspace

D̃ = {y ∈ DT̃ ; (⊕j IT−1 )DT̃ y ∈ Mz∗ HK (D)} ⊂ DT̃

166 J. Eschmeier and S. Toth

such that

W (M) = {WT x; x ∈ D̃},

where WT x : Bd → D acts as (WT x)(z) = WT (z)x. We equip D̃ with the norm

y = yH̃ d that it inherits as a closed subspace D̃ ⊂ H̃ d .
Lemma 2.5 Let T ∈ L(H )d be a pure K-contraction. Then a function f ∈ HK (D)
belongs to the wandering subspace W (M) of

M = (Im j )⊥ ∈ Lat(Mz , HK (D))

if and only if there is a vector y ∈ D̃ with

f = −T̃ y + Mz (⊕j IT−1 )DT̃ y.

In this case f 2 = y2 d .

H̃
Proof By Theorem 2.3 a function f ∈ HK (D) belongs to W (M) if and only if it is
of the form

f = f0 + Mz (j xi )di=1

with f0 ∈ D and x1 , . . . , xd ∈ H such that (j xi )di=1 ∈ Mz∗ HK (D) and

T̃ (IT xi )di=1 + DT̃ ∗ f0 = 0.

Then y = DT̃ (IT xi )di=1 − T̃ ∗ f0 ∈ DT̃ is a vector with

   
(IT xi ) 0
U = ,
f0 y

or equivalently, with
     
(IT xi ) 0∗ DT̃ y
=U = .
f0 y −T̃ y

But then y ∈ D̃ and f = −T̃ y + Mz (⊕j IT−1 )DT̃ y. Conversely, if f is of this form,
then using the definitions of T̃ , D̃ and the intertwining relation T̃ DT̃ = DT̃ ∗ T̃ one
can easily show that the vectors defined by

f0 = −T̃ y ∈ D and (xi )di=1 = (⊕IT−1 )DT̃ y ∈ H d

K-Inner Functions and K-Contractions 167

yield a representation f = f0 + Mz (j xi )di=1 as in Theorem 2.3. By Lemma 2.4 and

the definition of the scalar product on H̃ we find that


d 
d
f 2 = f0 2 + !T xi , xi  = T̃ y2 + IT xi 2H̃
i=1 i=1

= T̃ y2 + DT̃ y2H̃ d = y2H̃ d .

Recall that the reproducing kernel K : Bd × Bd → C is defined by K(z, w) =

k(z, w), where
∞

k : D → C, k(z) = an z n
n=0

is an analytic function with a0 = 1, an > 0 for all n such that

an an
0 < inf ≤ sup < ∞.
n an+1 n an+1

Let us suppose in addition that the limit

an
r = lim
n→∞ an+1

exists. Then r ∈ [1, ∞) is the radius of convergence of the power series defining k
and by Theorem 4.5 in [11] the Taylor spectrum of Mz ∈ L(HK (D))d is given by
√
σ (Mz ) = {z ∈ Cd ; z ≤ r}.

If T ∈ L(H )d is a pure K-contraction, then T ∗ is unitarily equivalent to a restriction

of Mz∗ and hence
√
σ (T ∗ ) ⊂ {z ∈ Cd ; z ≤ r}.

The function F : Dr (0) → C, F (z) = ∞ n
n=0 an+1 z , is analytic on the open disc
Dr (0) with radius r and center 0 and satisfies

k(z) − 1
F (z) = (z ∈ Dr (0) \ {0}).
z
d
For z ∈ Bd , let us denote by Z : H d → H, (hi )di=1 → i=1 zi hi , the row
operator induced by z. As a particular case of a much more general analytic spectral
168 J. Eschmeier and S. Toth

mapping theorem for the Taylor spectrum [9, Theorem 2.5.10] we find that


d
σ (ZT ∗ ) = { zi wi ; w ∈ σ (T ∗ )} ⊂ Dr (0)
i=1

for z ∈ Bd . Thus we can define an operator-valued function FT : Bd → L(H ),

⎛ ⎞
∞
 
FT (z) = F (ZT ) = ∗
an+1 ⎝ γα T ∗α α ⎠
z .
n=0 |α|=n

Lemma 2.6 For (xi )di=1 ∈ H d and z ∈ Bd ,

CF (ZT ∗ )Z(xi )di=1 = (δMz (j xi )di=1 )(z).

Proof For (xi )di=1 ∈ H d ,

⎛ ⎞

d ∞
 
δMz (j xi )di=1 = δMzi an ⎝ γα CT ∗α xi zα ⎠
i=1 n=0 |α|=n
⎛ ⎞
 ∞
d  
= an δ ⎝ γα CT ∗α xi zα+ei ⎠
i=1 n=0 |α|=n

 ∞
d  
= an+1 γα CT ∗α xi zα+ei ,
i=1 n=0 |α|=n

where the series converge in HK (D). Since the point evaluations are continuous on
HK (D), we obtain
 d
  ∞
  
∗α
δMz (j xi )di=1 (z) = an+1 γα CT zi xi zα
n=0 |α|=n i=1
∗
= CF (ZT )Z(xi )di=1

for all z ∈ Bd .
By Lemma 2.6 the map WT : Bd → L(D̃, D),

WT (z)(x) = −T (⊕!T IT−1 )x + CF (ZT ∗ )Z(⊕IT−1 )DT̃ x

= −T̃ x + CF (ZT ∗ )Z(⊕IT−1 )DT̃ x

defines an analytic operator-valued function.

K-Inner Functions and K-Contractions 169

Theorem 2.7 Let T ∈ L(H )d be a pure K-contraction. Then

W (M) = {WT x; x ∈ D̃}

and WT x = x for x ∈ D̃.

Proof For x ∈ D̃, Lemma 2.6 implies that

WT = −T̃ x + δMz (⊕j IT−1 )DT̃ x = −T̃ x + Mz (⊕j IT−1 )DT̃ x.

Thus the assertion follows from Lemma 2.5.

Since W (M) is a wandering subspace for Mz , the map WT : Bd → L(D̃, D)
is an operator-valued analytic function such that D̃ → HK (D), x → WT x, is an
isometry and
 
WT (D̃) ⊥ Mzα WT (D̃) for all α ∈ Nd \ {0}.

Thus WT : Bd → L(D̃, D) is a K-inner function with WT (D̃) = W (M). In

the case that Mz ∈ L(HK )d is a row contraction one can show that each K-inner
function W : Bd → L(E˜ , E ) defines a contractive multiplier

MW : Hd2 (E ) → HK , f → Wf

from the E -valued Drury–Arveson space Hd2 (E ) to HK (E˜ ) [3, Theorem 6.2].

3 K-Inner Functions

In the previous section we saw that the K-inner function WT : Bd → L(D̃, D)

associated with a pure K-contraction T ∈ L(H )d has the form

WT (z) = D + CF (ZT ∗ )ZB,

 1
∈ L(H, D), D = −T̃ ∈ L(D̃, D) and B = (⊕IT−1 )DT̃ ∈
2
where C = 1
K (T )
L(D̃, H d ). An elementary calculation using the definitions and the intertwining
relation T̃ DT̃ = DT̃ ∗ T̃ shows that the operators T , B, C, D satisfy the conditions

1
(K1) C ∗ C = (T ),
K
(K2) D ∗ C + B ∗ (⊕!T )T ∗ = 0,
170 J. Eschmeier and S. Toth

(K3) D ∗ D + B ∗ (⊕!T )B = 1D̃ ,

(K4) Im((⊕j )B) ⊂ Mz∗ HK (D).

If E is a Hilbert space and C ∈ L(H, E ) is any operator with C ∗ C = 1

K (T ), then
exactly as in the proof of Proposition 2.6 from [17] it follows that

jC : H → HK (E ), jC (x) = a|α| γα (CT ∗α x)zα
α∈Nd

is a well defined isometry that intertwines the tuples T ∗ ∈ L(H )d and Mz∗ ∈
L(HK (E ))d componentwise. As in the section following Theorem 2.1 one can show
that

jC∗ f = T α C ∗ fα
α∈Nd

for f = α∈Nd fα zα ∈ HK (E ). Hence we find that

jC∗ !jC x = jC∗ ! a|α| γα (CT ∗α x)zα
α∈Nd

= jC∗ a|α|+1 γα (CT ∗α x)zα
α∈Nd

= a|α|+1 γα (T α C ∗ CT ∗α x)
α∈Nd
 1
= a|α|+1 γα (T α (T )T ∗α x)
K
α∈Nd

for all x ∈ H . By performing the same chain of calculations with jC replaced by

the canonical K-dilation j of T from Theorem 2.1 we obtain that

jC∗ !jC = j ∗ !j = !T .

Our next aim is to show that any matrix operator

 
T∗ B
: H ⊕ E∗ → H d ⊕ E ,
C D

where T is a pure K-contraction and T , B, C, D satisfy the conditions (K1)–(K3)

with (D̃, D) replaced by (E∗ , E ) and

(K4) Im((⊕jC )B) ⊂ Mz∗ HK (E )

K-Inner Functions and K-Contractions 171

gives rise to a K-inner function W : Bd → L(E∗ , E ) defined as

W (z) = D + CF (ZT ∗ )ZB

and that, conversely, under a natural condition on the kernel K each K-inner
function is of this form.
Theorem 3.1 Let W : Bd → L(E∗ , E ) be an operator-valued function between
Hilbert spaces E∗ and E such that

W (z) = D + CF (ZT ∗ )ZB (z ∈ Bd ),

where T ∈ L(H )d is a pure K-contraction and the matrix operator

 
T∗ B
: H ⊕ E∗ → H d ⊕ E
C D

satisfies the condition (K1)–(K4). Then W is a K-inner function.

Proof The space M = HK (E )  Im jC ⊂ HK (E ) is a closed Mz -invariant
subspace. Let x ∈ E∗ be a fixed vector. By condition (K4) there is a function
f ∈ HK (E ) with (⊕jC )Bx = Mz∗ f . Exactly as in the proof of Lemma 2.6 it
follows that

CF (ZT ∗ )ZBx = δMz (⊕jC )Bx(z) = δMz Mz∗ f (z)

for all z ∈ Bd . Since δ(Mz Mz∗ ) = PIm Mz is an orthogonal projection and since
δMz = Mz (⊕!), we find that

W x2HK (E ) − Dx2 = δMz Mz∗ f, f HK (E )

= ⊕(jC∗ !jC )Bx, BxH d

= (⊕!T )Bx, BxH d
= (1E∗ − D ∗ D)x, x
= x2 − Dx2 .

Hence the map E∗ → HK (E ), x → W x, is a well-defined isometry. Using the

second part of Lemma 2.2 we obtain

Mz∗ (W x) = Mz∗ δMz Mz∗ f = Mz∗ f = (⊕jC )Bx

172 J. Eschmeier and S. Toth

and hence that PM Mz∗i (W x) = (1HK (E ) − jC jC∗ )Mz∗i (W x) = 0 for i = 1, · · · , d.

To see that W E∗ ⊂ M note that with x and f as above

jC∗ (W x) = C ∗ Dx + jC∗ (δMz Mz∗ f )

= C ∗ Dx + jC∗ (Mz (⊕!)Mz∗ f )
= C ∗ Dx + T (⊕jC∗ !jC )Bx
= C ∗ Dx + T (⊕!T )Bx
= 0.
d
Thus we have shown that W E∗ ⊂ M  i=1 zi M which implies that

W E∗ ⊥ zα (W E∗ )

for all α ∈ Nd \ {0}.

To prove that conversely each K-inner function W : Bd → L(E∗ , E ) has
the form described in Theorem 3.1 we make the additional assumption that the
multiplication tuple Mz ∈ L(HK )d is a K-contraction. This hypothesis is satisfied,
for instance, if HK is a complete Nevanlinna–Pick space such as the Drury–Arveson
space or the Dirichlet space or if K is a power

1
Kν : Bd × Bd → C, Kν (z, w) = (ν ∈ (0, ∞))
(1 − z, w)ν

of the Drury–Arveson kernel (see the discussion following Theorem 4.2). In the
proof we shall use a uniqueness result for minimal K-dilations whose proof we
postpone to Sect. 4.
Theorem 3.2 Let Mz ∈ L(HK )d be a K-contraction. If W : Bd → L(E∗ , E ) is a
K-inner function, then there exist a pure K-contraction T ∈ L(H )d and a matrix
operator
 
T∗ B
∈ L(H ⊕ E∗ , H d ⊕ E )
C D

satisfying the conditions (K1)–(K4) such that

W (z) = D + CF (ZT ∗ )ZB (z ∈ Bd ).

Proof Since W is K-inner, the space

W = W E∗ ⊂ HK (E )
K-Inner Functions and K-Contractions 173

is a generating wandering subspace for Mz ∈ L(HK (E ))d restricted to

8
S = Mzα W ⊂ HK (E ).
α∈Nd

The compression T = PH Mz |H of Mz ∈ L(HK (E ))d to the Mz∗ -invariant subspace

H = HK (E )  S is easily seen to be a pure K-contraction [17, Proposition 2.12
and Lemma 2.21]. Let R ⊂ HK (E ) be the smallest reducing subspace for Mz ∈
L(HK (E ))d that contains H . By Lemma 4.4
8
R= zα (R ∩ E ) = HK (R ∩ E ).
α∈Nd

Thus the inclusion map i : H → HK (R ∩ E ) is a minimal K-dilation for T . Let

j : H → HK (D) be the K-dilation of the pure K-contraction T ∈ L(H )d defined
in Theorem 2.1. Since also j is a minimal K-dilation for T (Corollary 4.5), by
Corollary 4.3 there is a unitary operator U : D → R ∩ E such that

i = (1HK ⊗ U )j.

Define Eˆ = E  (R ∩ E ). By construction

HK (Eˆ ) = HK (E )  HK (R ∩ E ) = HK (E )  R ⊂ S

is the largest reducing subspace for Mz ∈ L(HK (E ))d contained in S . In particular,

the space S admits the orthogonal decomposition

S = HK (Eˆ ) ⊕ (S ∩ HK (Eˆ )⊥ ) = HK (Eˆ ) ⊕ (HK (R ∩ E )  S ⊥ ).

We complete the proof by comparing the given K-inner function W : Bd →

L(E∗ , E ) with the K-inner function WT : Bd → L(D̃, D) associated with the pure
K-contraction T ∈ L(H )d . For this purpose, let us define the Mz -invariant subspace

M = HK (D)  Im j

and its wandering subspace



d
W (M) = M  zi M
i=1

as in Sect. 2. Using the identity i = (1HK ⊗ U )j one obtains that

1HK ⊗ U : M → HK (R ∩ E )  S ⊥ = HK (R ∩ E ) ∩ S
174 J. Eschmeier and S. Toth

defines a unitary operator that intertwines the restrictions of Mz to both sides

componentwise. Consequently we obtain the orthogonal decomposition

W = WMz (S ) = WMz (HK (Eˆ )) ⊕ WMz (HK (R ∩ E ) ∩ S )

= Eˆ ⊕ (1HK ⊗ U )W (M).

Let WT : Bd → L(D̃ , D) be the K-inner function associated with the pure K-

contraction T ∈ L(H )d . Then there is a matrix operator
 
T∗ B
∈ L(H ⊕ D̃, H d ⊕ D)
C D

such that

WT (z) = D + CF (ZT ∗ )ZB (z ∈ Bd )

and W (M) = {WT x; x ∈ D̃} (see the beginning of Sect. 3 and Theorem 2.7). Let
us denote by

P1 : W → Eˆ and P2 : W → (1HK ⊗ U )W (M)

the orthogonal projections. The K-inner functions W : Bd → L(E∗ , E ) and

WT : Bd → L(D̃ , D) induce unitary operators

E∗ → W , x → W x

and

D̃ → W (M) x → WT x.

We define surjective bounded linear operators by

U1 : E∗ → Eˆ , U1 x = P1 W x

and

U2 : E∗ → D̃, U2 x = x̃ if (1HK ⊗ U )WT x̃ = P2 W x.

By construction the column operator

(U1 , U2 ) : E∗ → Eˆ ⊕ D̃
K-Inner Functions and K-Contractions 175

defines an isometry such that

W (z)x = U1 x + U WT (z)U2 x = (U1 + U DU2 )x + (U C)F (ZT ∗ )Z(BU2 )x

holds for z ∈ Bd and x ∈ E∗ . To complete the proof we show that the operators

T ∈ L(H d , H ), B̃ = BU2 ∈ L(E∗ , H d ), C̃ = U C ∈ L(H, E )

and D̃ = U1 + U DU2 ∈ L(E∗ , E )

satisfy the conditions (K1)–(K4). To see this note that

1
C̃ ∗ C̃ = C ∗ U ∗ U C = C ∗ C = (T )
K
and

D̃ ∗ C̃ = U2∗ D ∗ U ∗ U C = U2∗ D ∗ C
= −U2∗ B ∗ (⊕!T ) T ∗ = −B̃ ∗ (⊕!T ) T ∗ .

To verify condition (K3) note that D̃ acts as the column operator

D̃ = (U1 , U DU2 ) : E∗ → E = Eˆ ⊕ (R ∩ E ).

Thus we obtain that

D̃ ∗ D̃ = U1∗ U1 + U2∗ D ∗ U ∗ U DU2

= U1∗ U1 + U2∗ U2 − U2∗ B ∗ (⊕!T ) BU2
= 1E∗ − B̃ ∗ (⊕!T ) B̃.

Since jC̃ = UjC , it follows that


⊕jC̃ B̃x = (⊕U )(⊕jC )B(U2 x) ∈ Mz∗ HK (E )

holds for all x ∈ E∗ . Thus the K-inner function W : Bd → L(E∗ , E ) admits a matrix
representation of the claimed form.

4 Minimal K-Dilations

Let A be a unital subalgebra of a unital C ∗ -Algebra B. A completely positive unital

map ϕ : B → L(H ) is called an A-morphism if ϕ(ax) = ϕ(a)ϕ(x) for a ∈ A and
176 J. Eschmeier and S. Toth

x ∈ B. Under the condition that B is the norm-closed linear span

B = span· {AA∗ }

Arveson proved in [1, Lemmma 8.6] that every unitary operator that intertwines
two A-morphisms ϕi : B → L(Hi ) (i = 1, 2) pointwise on A extends to a unitary
operator that intertwines the minimal Stinespring representations of ϕ1 and ϕ2 .
Straightforward modifications of the arguments given in [1] show that Arveson’s
result remains true if B is a von Neumann algebra which is the w∗ -closed linear
span
∗
B = spanw {AA∗ }

and if the A-morphisms ϕi : B → L(Hi ) (i = 1, 2) are supposed to be w∗ -

continuous.
Theorem 4.1 Let B be a von Neumann algebra and let A ⊂ B be a unital
subalgebra such that
∗
B = spanw {AA∗ }.

For i = 1, 2, let ϕi : B → L(Hi ) be a w∗ -continuous A-morphism and let

(πi , Vi , Hπi ) be the minimal Stinespring representations for ϕi . For every unitary
operator U : H1 → H2 with

U ϕ1 (a) = ϕ2 (a)U (a ∈ A),

there is a unique unitary operator W : Hπ1 → Hπ2 with W V1 = V2 U and

W π1 (x) = π2 (x)W for all x ∈ B.
Since this version of Arveson’s result follows in exactly the same way as the
original one [1, Lemma 8.6], we leave the details to the reader. As an application of
Theorem 4.1 we show that, under suitable conditions on the kernel K : Bd × Bd →
C, minimal K-dilations are uniquely determined. Recall that a commuting tuple
T ∈ L(H )d on a Hilbert space H is called essentially normal if Ti Ti∗ − Ti∗ Ti
is compact for i = 1, . . . , d. If T ∈ L(H )d is essentially normal, then by the
Fuglede–Putnam theorem also all cross commutators Ti Tj∗ −Tj∗ Ti (i, j = 1, . . . , d)
are compact. For our multiplication tuple Mz ∈ L(HK )d , essential normality is
equivalent to the condition that [11, Corollary 4.4]
 
an an−1
lim − = 0.
n→∞ an+1 an
K-Inner Functions and K-Contractions 177

Theorem 4.2 Suppose that Mz ∈ L(HK )d is an essentially normal K-contraction.

Then the von Neumann algebra generated by Mz1 , . . . , Mzd is given by
∗
W ∗ (Mz ) = spanw {Mzα Mz∗β ; α, β ∈ Nd }.
∗ ∗β
Proof Define L = spanw {Mzα Mz ; α, β ∈ Nd }. Obviously L ⊂ W ∗ (Mz ).
Since Mz is supposed to be a K-contraction,
∞

PC = τSOT − n
cn σM z
(1HK ) ∈ L .
n=0

For α, β ∈ Nd and w ∈ Bd , we obtain

Mzα PC Mz∗β (K(·, w)) = w β zα = zα ⊗ zβ (K(·, w)).

Since the multiplication on L(HK ) is separately w∗ -continuous, it follows that L

contains all compact operators

K(HK ) = span· {zα ⊗ zβ ; α, β ∈ Nd } ⊂ L .

But then the hypothesis that Mz is essentially normal implies that L ⊂ L(HK ) is
a subalgebra. Since the involution on L(HK ) is w∗ -continuous, the algebra L ⊂
L(HK ) is a von Neumann algebra and hence L = W ∗ (Mz ).
The tuple Mz ∈ L(HK )d is known to be a K-contraction if there is a natural
number p ∈ N such that cn ≥ 0 for all n ≥ p or cn ≤ 0 for all n ≥ p [7, Lemma
2.2] or [17, Proposition 2.10]. The latter condition holds, for instance, if HK is a
complete Nevanlinna–Pick space or if K is a kernel of the form

1
Kν : Bd × Bd → C, Kν (z, w) =
(1 − z, w)ν

with a positive real number ν > 0 (see [8, Lemma 2.1] and [17, Section 1.5.2] for
these results and further examples).
Let T ∈ L(H )d be a commuting tuple and let j : H → HK (E ) be a K-dilation
of T . We denote by B = W ∗ (Mz ) ⊂ L(HK ) the von Neumann algebra generated
by Mz and set A = {p(Mz ); p ∈ C[z]}. The unital C ∗ -homomorphism

π : B → L(HK (E )), X → X ⊗ 1E

together with the isometry j : H → HK (E ) is a Stinespring representation for the

completely positive map

ϕ : B → L(HK (E )), ϕ(X) = j ∗ (X ⊗ 1E )j.

178 J. Eschmeier and S. Toth

The map ϕ is an A-morphism, since

ϕ(p(Mz )X) = j ∗ (p(Mz ⊗ 1E )X ⊗ 1E )j = j ∗ p(Mz ⊗ 1E )(jj ∗ )(X ⊗ 1E )j

= ϕ(p(Mz ))ϕ(X)

for all p ∈ C[z] and X ∈ B. Standard duality theory for Banach space operators
shows that π is w∗ -continuous. Indeed, as an application of Krein–Smulian’s
theorem (Theorem IV. 6.4 in [16]) one only has to check that τw∗ −limα (Xα ⊗1E ) =
X ⊗ 1E for each norm-bounded net (Xα ) in B with τw∗ − limα Xα = X. To
complete the argument it suffices to recall that on norm-bounded sets the w∗ -
topology and the weak operator topology coincide. Thus we have shown that ϕ is
a w∗ -continuous A-morphism with Stinespring representation π. By definition the
K-dilation j : H → HK (E ) is minimal if and only if
8
π(X)(j H ) = HK (E ),
X∈W ∗ (Mz )

hence if and only if π as a Stinespring representation of ϕ is minimal.

Corollary 4.3 Suppose that Mz ∈ L(HK )d is an essentially normal K-contraction.
If ji : H → HK (Ei ) (i = 1, 2) are two minimal K-dilations of a commuting tuple
T ∈ L(H )d , then there is a unitary operator U ∈ L(E1 , E2 ) with j2 = (1HK ⊗ U )j1
Proof As before we denote by B = W ∗ (Mz ) ⊂ L(HK ) the von Neumann algebra
generated by Mz1 , . . . , Mzd ∈ L(HK ) and define A = {p(Mz ); p ∈ C[z]}. The
remarks preceding the corollary show that the maps

ϕi : B → L(H ), ϕi (X) = ji∗ (X ⊗ 1Ei )ji (i = 1, 2)

are w∗ -continuous A-morphisms with minimal Stinespring representations

πi : B → L(HK (Ei )), πi (X) = X ⊗ 1Ei (i = 1, 2).

Since

ϕi (p(Mz )) = j ∗ p(Mz ⊗ 1E )j = p(T )

for all p ∈ C[z] and i = 1, 2, Theorem 4.1 implies that there is a unitary operator
W : HK (E1 ) → HK (E2 ) with Wj1 = j2 and W (X ⊗ 1E1 ) = (X ⊗ 1E2 )W for all
X ∈ B. In particular, the unitary operator W satisfies the intertwining relations

W (Mzi ⊗ 1E1 ) = (Mzi ⊗ 1E2 )W (i = 1, . . . , d)

K-Inner Functions and K-Contractions 179

A standard characterization of multipliers on reproducing kernel Hilbert spaces [4,

Theorem 2.1] shows that there exist operator-valued functions A : Bd → L(E1 , E2 )
and B : Bd → L(E2 , E1 ) such that Wf = Af and W ∗ g = Bg for f ∈ HK (E1 ) and
g ∈ HK (E2 ) (see also [17, Proposition 4.5]). It follows that A(z)B(z) = 1E2 and
B(z)A(z) = 1E1 for z ∈ Bd . Since

K(z, w)x = (W W ∗ K(·, w)x)(z) = A(z)K(z, w)A(w)∗ x

for z, w ∈ Bd and x ∈ E2 , we find that A(z)A(w)∗ = 1E2 for z, w ∈ Bd . But then

the constant value A(z) ≡ U ∈ L(E1 , E2 ) is a unitary operator with W = 1HK ⊗ U .

We conclude this section by showing that the canonical K-dilation of a K-

contraction T ∈ L(H )d defined in Theorem 2.1 is minimal. To prepare this result
we first identify the Mz -reducing subspaces of HK (E ).
Lemma 4.4 Let M ⊂ HK (E ) be a closed linear subspace. If M is reducing for
Mz ∈ L(HK (E ))d , then PE M ⊂ M and
8
M= zα (M ∩ E ) = HK (M ∩ E ).
α∈Nd

Proof The hypothesis implies that M is reducing for the von Neumann algebra
W ∗ (Mz ) ⊂ L(HK (E )) generated by Mz1 , . . . Mzd ∈ L(HK (E )). Standard results
on von Neumann algebras (Corollary 17.6 and Proposition 24.1 in [18]) show that

PE = P9 Ker Mz∗ ∈ W ∗ (Mz ).

i


Hence PE M ⊂ M. Let f = α∈Nd fα zα ∈ HK (E ) be arbitrary. An elementary
calculation yields that

PE (Mz∗β f ) ∈ (C \ {0})fβ (β ∈ Nd ).

Hence, if f ∈ M, then fβ ∈ M ∩ E for all β ∈ Nd and the observation that

 8
f = fα zα ∈ zα (M ∩ E ) = HK (M ∩ E )
α∈Nd α∈Nd

completes the proof.

Corollary 4.5 Let T ∈ L(H )d be a pure K-contraction. Then the K-dilation

j : H → HK (D), j (x) = a|α| γα (CT ∗α x)zα
α∈Nd

defined in Theorem 2.1 is minimal.

180 J. Eschmeier and S. Toth

Proof Let Im j ⊂ M be a reducing subspace for Mz ∈ L(HK (D))d . We know from

Lemma 4.4 that
8
M= zα (M ∩ D)
α∈Nd

and that

CH = PD (Im j ) ⊂ PD (M) ⊂ M ∩ D.
6
It follows that D = CH = M ∩ D and that M = α∈Nd zα D = HK (D).
It should be interesting to compare the uniqueness result proved in this section
with a related result proved by Olofsson [15, Theorem 7.6] for single contractions
T ∈ L(H ) satisfying a slightly different K-contractivity condition. In [15] it is
shown that even in the non-pure case each dilation factors through a canonical
defined dilation of T .

References

1. W. Arveson, Subalgebras of C ∗ -algebras. III: multivariable operator theory. Acta Math. 181,
159–228 (1998)
2. W. Arveson, Quotients of standard Hilbert modules. Trans. Am. Math. Soc. 359, 6027–6055
(2007)
3. M. Bhattacharjee, J. Eschmeier, D.K. Keshari, J. Sarkar, Dilations, wandering subspaces and
inner functions. Linear Algebra Appl. 523, 263–280 (2017)
4. C. Barbian, A characterization of multiplication operators on reproducing kernel Hilbert
spaces. J. Oper. Theory 65, 235–240 (2011)
5. J.A. Ball, V. Bolotnikov, Weighted Bergman spaces: shift-invariant subspaces and
input/state/output linear systems. Integr. Equ. Oper. Theor. 76, 301–356 (2013)
6. J.A. Ball, V. Bolotnikov, Weighted Hardy spaces: shift-invariant subspaces and coinvariant
subspaces, linear systems and operator model theory. Acta Sci. Math. (Szeged) 79, 623–686
(2013)
7. Y. Chen, Quasi-wandering subspaces in a class of reproducing analytic Hilbert spaces. Proc.
Am. Math. Soc. 140, 4235–4242 (2012)
8. R. Clouâtre, M. Hartz, Multiplier algebras of complete Nevanlinna–Pick spaces: dilations,
boundary representations and hyperrigidity. J. Funct. Anal. 274, 1690–1738 (2018)
9. J. Eschmeier, M. Putinar, Spectral Decompositions and Analytic Sheaves. London Mathemati-
cal Society Monographs, New Series, vol. 10 (Clarendon Press, Oxford, 1996)
10. J. Eschmeier, Bergman inner functions and m-hypercontractions. J. Funct. Anal. 275, 73–102
(2018)
11. K. Guo, J. Hu, X. Xu, Toeplitz algebras, subnormal tuples and rigidity on reproducing
C[z1 , . . . , zd ]-modules. J. Funct. Anal. 210, 214–247 (2004)
12. V. Müller, F.-H. Vasilescu, Standard models for some commuting multioperators. Proc. Am.
Math. Soc. 117, 979–989 (1993)
13. A. Olofsson, A characteristic operator function for the class of n-hypercontractions. J. Funct.
Anal. 236, 517–545 (2006)
K-Inner Functions and K-Contractions 181

14. A. Olofsson, Operator-valued Bergman inner functions as transfer functions. Algebra Anal. 19,
146–173 (2007); St. Petersberg Math. J. 19, 603–623 (2008)
15. A. Olofsson, Parts of adjoint weighted shifts. J. Oper. Theory 74, 249–280 (2015)
16. H.H. Schaefer, Topological Vector Spaces (Macmillan, New York, 1966)
17. D. Schillo, K-contractions and perturbations of Toeplitz operators. Ph.D. thesis, Saarland
University, 2018
18. K. Zhu, An Introduction to Operator Algebras. Studies in Advanced Mathematics (CRC Press,
Boca Raton, 1993)
Tight and Cover-to-Join Representations
of Semilattices and Inverse Semigroups

Ruy Exel

Abstract We discuss the relationship between tight and cover-to-join representa-

tions of semilattices and inverse semigroups, showing that a slight extension of the
former, together with an appropriate selection of codomains, makes the two notions
equivalent. As a consequence, when constructing universal objects based on them,
one is allowed to substitute cover-to-join for tight and vice-versa.

Keywords Semilattice · Inverse semigroup · Tight representation ·

Cover-to-join · Boolean algebra · Non-degenerate representation · Universal
C*-algebra

Mathematics Subject Classification (2010) Primary 20M18, 20M30; Secondary

46L05

1 Introduction

Exactly 12 years ago, to be precise on March 7, 2007, I posted a paper on the

arXiv [3] describing the notion of tight representations of semilattices and inverse
semigroups, which turned out to have many applications and in particular proved
to be useful to give a unified perspective to a significant number of C*-algebras
containing a preferred generating set of partial isometries [1, 2, 4, 6–8, 12, 13].
The notion of tight representations (described below for the convenience of the
reader) is slightly involving as it depends on the analysis of certain pairs of finite
sets X and Y , but it becomes much simplified when X is a singleton and Y is empty
(see [4, Proposition 11.8]). In this simplified form it has been rediscovered and
used in many subsequent works (e.g. [2, 9–11]) under the name of cover-to-join
representations.

R. Exel ()
Departamento de Matemática, Universidade Federal de Santa Catarina, Florianópolis, SC, Brazil

© Springer Nature Switzerland AG 2021 183

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_9
184 R. Exel

The notion of cover-to-join representations, requiring a smaller set of conditions,

is consequently weaker and, as it turns out, strictly weaker, than the original notion
of tightness. Nevertheless, besides being easier to formulate, the notion of cover-to-
join representations has the advantage of being applicable to representations taking
values in generalized Boolean algebras, that is, Boolean algebras without a unit.
Explicitly mentioning the operation of complementation, tight representations only
make sense for unital Boolean algebras.
The goal of this note is to describe an attempt to reconcile the notions of tight
and cover-to-join representations: slightly extending the former, and adjusting for
the appropriate codomains, we show that, after all, the two notions coincide.
One of the main practical consequences of this fact is that the difference between
the two notions becomes irrelevant for the purpose of constructing universal objects
based on them, such as the completion of an inverse semigroup recently introduced
in [11]. We are moreover able to fix a slight imprecision in the proof of [2, Theorem
2.2], at least as far as its consequence that the universal C*-algebras for tight
vs. cover-to-join representations are isomorphic.

2 Generalized Boolean Algebras

We begin by recalling the well known notion of generalized Boolean algebras.

Definition 2.1 ([14, Definition 5]) A generalized Boolean algebra is a set B
equipped with binary operations ∧ and ∨, and containing an element 0, such that
for every a, b and c in B, one has that
(i)(commutativity) a ∨ b = b ∨ a, and a ∧ b = b ∧ a,
(ii)(associativity) (a ∧ b) ∧ c = a ∧ (b ∧ c),
(iii)(distributivity) a ∧ (b ∨ c) = (a ∧ b) ∨ (a ∧ c),
(iv) a ∨ 0 = a,
(v) (relative complement) if a = a ∧ b, there is an element x in B, such that
x ∨ a = b, and x ∧ a = 0,
(vi) a ∨ a = a = a ∧ a.
It follows that (ii) and (iii) also hold with ∨ and ∧ interchanged, meaning that ∨
is associative [14, Theorems 55 & 14], and that ∨ distributes over ∧ [14, Theorems
55 & 11].
When a = a ∧ b, as in (v), one writes a ≤ b. It is then easy to see that ≤ is a
partial order on B.
The element x referred to in (v) is called the relative complement of a in b, and
it is usually denoted b \ a.
Definition 2.2 (cf. [14, Theorem 56]) A generalized Boolean algebra B is called a
Boolean algebra if there exists an element 1 in B, such that a ∧ 1 = a, for every a
in B.
Tight and Cover-to-Join Representations 185

For Boolean algebras, the complement of an element a relative to 1 is often

denoted ¬a.
Recall that an ideal of a generalized Boolean algebra B is any nonempty subset
C of B which is closed under ∨, and such that

a ≤ b ∈ C ⇒ a ∈ C.

Such an ideal is evidently also closed under ∧ and under relative complements, so
it is a generalized Boolean algebra in itself.
Given any nonempty subset S of B, notice that the subset C defined by
 6 
C = a ∈ B : a ≤ z∈Z z, for some nonempty finite subset Z ⊆ S ,

is an ideal of B and it is clearly the smallest ideal containing

 S, so we shall call it
the ideal generated by S, and we shall denote it by S .

3 Tight and Cover-to-Join Representations of Semilattices

From now on let us fix a meet semilattice E (always assumed to have a zero
element).
Definition 3.1 A representation of E in a generalized Boolean algebra B is any
map π : E → B, such that
(i) π(0) = 0, and
(ii) π(x ∧ y) = π(x) ∧ π(y), for every x and y in E.
In order to spell out the definition of the notion of tight representations,
introduced in [4], let F be any subset of E. We then say that a given subset Z ⊆ F
is a cover for F , if for every nonzero x in F , there exists some z in Z, such that
z ∧ x = 0.
Furthermore, if X and Y are finite subsets of E, we let

E X,Y = {z ∈ E : z ≤ x, ∀x ∈ X, and z ⊥ y, ∀y ∈ Y }.

Definition 3.2 (cf. [4, Definition 11.6]) A representation π of E in a Boolean

algebra B is said to be tight if, for any finite subsets X and Y of E, and for any
finite cover Z for E X,Y , one has that
8 : :
π(z) = π(x) ∧ ¬ π(y). (3.1)
z∈Z x∈X y∈Y
186 R. Exel

Observe that if Y is empty and X is a singleton, say X = {x}, then

E X,Y = E {x},∅ = {z ∈ E : z ≤ x},

and if Z is a cover for this set, then (3.1) reads

8
π(z) = π(x). (3.2)
z∈Z

To check that a given representation is tight, it is not enough to verify (3.2), as

it is readily seen by considering the example in which E = {0, 1} and B is any
Boolean algebra containing an element x = 1. Indeed, the map π : E → B given
by π(0) = 0, and π(1) = x, satisfies all instances of (3.2) even thought it is not
tight. The reader might wonder if the fact that π fails to preserve the unit is playing
a part in this counter-example, but it is also easy to find examples of cover-to-join
representations of non-unital semilattices which are not tight.
Representations π satisfying (3.2) whenever Z is a cover for E {x},∅ have
been considered in [4, Proposition 11.8], and they have been called cover-to-join
representations in [2].
It is a trivial matter to prove that a cover-to-join representation satisfies (3.1)
whenever X is nonempty (see the proof of [4, Lemma 11.7]), so the question of
whether a cover-to-join representation is indeed tight rests on verifying (3.1) when
X is empty. In this case, and assuming that Z is a cover for E ∅,Y , it is easy to see
that Z ∪ Y is a cover for the whole of E. Should we be dealing with a semilattice
not admitting any finite cover, this situation will therefore never occur, that is, one
will never be required to check (3.1) for an empty set X, hence every cover-to-join
representation is automatically tight.
This has in fact already been observed in [4, Proposition 11.8], which says that
every cover-to-join representation is tight in case E does not admit any finite cover,
as we have just discussed, but also if E contains a finite set X such that
8
π(x) = 1. (3.3)
x∈X

The latter condition is useful for dealing with characters, i.e. with representations
of E in the Boolean algebra {0, 1}, because the requirement that a character
be nonzero immediately implies (3.3), so again cover-to-join suffices to prove
tightness.
On the other hand, an advantage of the notion of cover-to-join representations is
that it makes sense for representations in generalized Boolean algebras, while the
reference to the unary operation ¬ in (3.1) precludes it from being applied when the
target algebra lacks a unit, that is, for a representation into a generalized Boolean
algebra.
Again referring to the occurrence of ¬ in (3.1), observe that if X is nonempty,
then the right hand side of (3.1) lies in the ideal of B generated by the range of
Tight and Cover-to-Join Representations 187

 
π. This is because, even though ¬ π(y) is not necessarily in π(E) , this term will
appear besides π(x), for some x in X, and hence
  
π(x) ∧ ¬π(y) = π(x) \ π(x) ∧ π(y) ∈ π(E) .

This means that:

Proposition 3.3 If E is a semilattice not admitting any finite cover then, whenever
X and Y are finite
 subsets
 of E, and Z is a finite cover of E X,Y , the right hand side
of (3.1) lies in π(E) .
As a consequence we see that Definition 3.2 may be safely applied to a
representation of E in a generalized Boolean algebra, as long as E does not admit
a finite cover: despite the occurrence of ¬ in (3.1), once its right hand side is
expanded, it may always be expressed in terms of relative complements, hence
avoiding the use of the missing unary operation ¬.
We may therefore consider the following slight generalization of the notion of
tight representations:
Definition 3.4 A representation π of E in a generalized Boolean algebra B is
said to be tight if, either B is a Boolean algebra and π is tight in the sense of
Definition 3.2, or the following two conditions are verified:
(i) E admits no finite cover, and
(ii) (3.1) holds for any finite subsets X and Y of E, and for any finite cover Z for
E X,Y .
As already stressed, despite the occurrence of ¬ in (3.1), condition (ii) in
Definition 3.4 will always make sense in a generalized Boolean algebra.
So here is a result that perhaps may be used to reconcile the notions of tightness
and cover-to-join representations:
Theorem 3.5 Let π be a representation of the semilattice E in the generalized
Boolean algebra B. Then
(i) if π is tight then it is also cover-to-join,
(ii) if π is cover-to-join then there exists an ideal B of B, containing the range of
π, such that, once π is seen as a representation of E in B , one has that π is
tight.
Proof Point (i) being immediate, let us prove (ii). Under the assumption that E does
not admit any finite cover, we have that π is tight as a representation into B = B,
by Exel [4, Proposition 11.8], or rather by its obvious adaptation to generalized
Boolean algebras.
It therefore remains to prove (ii) in case E does admit a finite cover, say Z.
Setting
8
e= π(z), (3.4)
z∈Z
188 R. Exel

we claim that

π(x) ≤ e, ∀x ∈ E. (3.5)

To see this, pick x in E and notice that, since Z is a cover for E, we have in particular
that the set

{z ∧ x : z ∈ Z}

is a cover for x, so the cover-to-join property of π implies that

8 8
π(x) = π(z ∧ x) ≤ π(z) = e,
z∈Z z∈Z

proving (3.5). We therefore let

B = {a ∈ B : a ≤ e},

which is evidently an ideal of B containing the range of π by (3.5).

By (3.4) we then have that π satisfies [4, Lemma 11.7.(i)], as long as we see π
as a representation of E in B , whose unit is clearly e. The result then follows from
[4, Proposition 11.8].

4 Non-Degenerate Representations of Semilattices

The following is perhaps the most obvious adaptation of the notion of non-
degenerate representations extensively used in the theory of operator algebras [15,
Definition 9.3].
Definition 4.1 We shall say that a representation π of a semilattice E in a
generalized Boolean algebra B is 6 non-degenerate if, for every a in B, there is a
finite subset Z of E such that a ≤ z∈Z π(z). In other words, π is non-degenerate
if and only if B coincides with the ideal generated by the range of π.
Observe that, if both E and B have a unit, and if π is a unital map, then π is
evidently non-degenerate. More generally, if π satisfies (3.3), then the same is also
clearly true.
The following result says that, by adjusting the codomain of a representation, we
can always make it non-degenerate.
Proposition 4.2 Let π be a representation of E in the generalized Boolean algebra
B. Letting C be the ideal of B generated by the range of π, one has that π is a
non-degenerate representation of E in C.
Proof Obvious.
Tight and Cover-to-Join Representations 189

For non-degenerate representations we have the following streamlined version of

Theorem 3.5:
Corollary 4.3 Let π be a non-degenerate representation of the semilattice E in the
generalized Boolean algebra B. Then π is tight if and only if it is cover-to-join.
Proof The “only if” direction being trivial, we concentrate on the “if” part, so let
us assume that π is cover-to-join. By Theorem 3.5 there exists an ideal B of B,
containing the range of π, and such that π is tight as a representation in B . Such
an ideal will therefore contain the ideal generated by π(E), which coincides with B
by hypothesis. Therefore B = B, and hence π is tight as a representation into its
default codomain B.

5 Representations of Inverse Semigroups

By its very nature, the concept of a tight representation pertains to the realm of
semilattices and Boolean algebras. However, given the relevance of the study of
semilattices in the theory of inverse semigroups, tight representations have had a
strong impact on the latter.
Recall that a Boolean inverse semigroup (see [5] but please observe that this
notion is not equivalent to the homonym studied in [9] and [16]) is an inverse
semigroup whose idempotent semilattice E(S) is a Boolean algebra. In accordance
with what we have been discussing up to now, it is sensible to give the following:
Definition 5.1
(i) A generalized Boolean inverse semigroup is an inverse semigroup whose
idempotent semilattice is a generalized Boolean algebra.
(ii) (cf. [4, Definition 13.1] and [5, Proposition 6.2]) If S is any inverse semigroup1
and T is a generalized Boolean inverse semigroup, we say that a homomor-
phism π : S → T (always assumed to preserve zero) is tight if the restriction
of π to E(S) is a tight representation into E(T ), in the sense of Definition 3.4.
(iii) If π is as above, we say that π is cover-to-join if the restriction of π to E(S) is
cover-to-join.
Addressing the already mentioned slight imprecision in the proof of [2, Theorem
2.2], we then have the following version of Theorem 3.5 and Proposition 4.2:

1 All inverse semigroups in this note are required to have a zero.

190 R. Exel

Corollary 5.2 Let π be a representation of the inverse semigroup S in the

generalized Boolean inverse semigroup T . Then
(i) if π is tight then it is also cover-to-join;
(ii) if π is cover-to-join then there exists a generalized Boolean inverse sub-
semigroup T of T , containing the range of π, such that, once π is seen as
a representation of S in T , one has that π is tight;
(iii) if π is cover-to-join, and if the restriction of π to E(S) is non-degenerate, then
π is tight.
Proof The proof is essentially contained in the proofs of Theorem 3.5 and
Proposition 4.2, except maybe for the proof of (ii) under the assumption that E(S)
admits a finite cover, say Z. In this case, let e be as in (3.4) and put

T = {t ∈ T : t ∗ t ≤ e, tt ∗ ≤ e},

observing that T is clearly an inverse sub-semigroup of T , and that its idempotent

semilattice is a Boolean algebra. Given any s in S, observe that s ∗ s lies in E(S) and

π(s)∗ π(s) = π(s ∗ s) ≤ e,

where the last inequality above follows as in (3.5). By a similar reasoning one shows
that also π(s)π(s)∗ ≤ e, so we see that π(s) lies in T , and we may then think of π
as a representation of S in T . As in Theorem 3.5, one may now easily prove that π
becomes a tight representation into T .

6 Conclusion

As a consequence of the above results, when defining universal objects (such

as semigroups, algebras or C*-algebras) for a class of representations of inverse
semigroups, one may safely substitute cover-to-join for tight and vice-versa. Given
the widespread use of tight representations, there are many instances where the
above principle applies. Below we spell out one such result to concretely illustrate
our point, but similar results may be obtained as trivial reformulations of the
following:
∗ (S) be the universal C*-
Theorem 6.1 Let S be an inverse semigroup and let Ctight
algebra [4, Theorem 13.3] for tight Hilbert space representations of S [4, Definition
13.1]. Also let C ∗cover- (S) be the universal C*-algebra for cover-to-join Hilbert
to-join
space representations of S. Then
∗
Ctight (S) 2 C ∗cover- (S).
to-join
Tight and Cover-to-Join Representations 191

∗ (S) also has the universal property for cover-to-

Proof It suffices to prove that Ctight
join representations. So let

π : S → B(H )

be a cover-to-join representation of S on some Hilbert space H . Should the

idempotent semilattice of S admit no finite covers, one has that π is tight so there is
nothing to do. On the other hand, assuming that Z is a finite cover for E(S), let e be
as in (3.4).
Writing He for the range of e and letting K = He⊥ , we then obviously have that
H = He ⊕ K. It then follows from (3.5) that each π(s) decomposes as a direct sum
of operators

π(s) = π (s) ⊕ 0,

thus defining a representation π of S on He which is clearly also cover-to-join. It is

also clear that π is non-degenerate on E(S), so we have by Corollary 5.2(iii) that π
∗ (S)
is tight. Therefore the universal property provides a *-representation ϕ of Ctight
∗
on B(He ) coinciding with π on the canonical image of S within Ctight (S). It then
follows that ϕ := ϕ ⊕ 0 coincides with π on S, concluding the proof.
We therefore believe this clarifies [2, Corollaries 2.3 & 2.5].

References

1. G. Boava, G.G. de Castro, F.de L. Mortari, Inverse semigroups associated with labelled spaces
and their tight spectra. Semigroup Forum 94, 582–609 (2017)
2. A.P. Donsig, D. Milan, Joins and covers in inverse semigroups and tight C*-algebras. Bull.
Aust. Math. Soc. 90, 121–133 (2014)
3. R. Exel, Inverse Semigroups and Combinatorial C*-algebras (2007). arXiv:math/ 0703182v1
[math.OA]
4. R. Exel, Inverse semigroups and combinatorial C*-algebras. Bull. Braz. Math. Soc. (N.S.) 39,
191–313 (2008)
5. R. Exel, Tight representations of semilattices and inverse semigroups. Semigroup Forum 79,
159–182 (2009)
6. R. Exel, E. Pardo, Self-similar graphs, a unified treatment of Katsura and Nekrashevych C∗ -
algebras. Adv. Math. 306, 1046–1129 (2017)
7. R. Exel, C. Starling, Self-similar graph C ∗ -algebras and partial crossed products. J. Oper.
Theory 75, 299–317 (2016)
8. R. Exel, D. Goncalves, C. Starling, The tiling C*-algebra viewed as a tight inverse semigroup
algebra. Semigroup Forum 84, 229–240 (2012)
9. M.V. Lawson, Non-commutative Stone duality: inverse semigroups, topological groupoids and
C*-algebras. Internat. J. Algebra Comput. 22(6), 1250058, 47 (2012)
10. M.V. Lawson, D.G. Jones, Graph inverse semigroups: their characterization and completion. J.
Algebra 409, 444–473 (2014)
192 R. Exel

11. M.V. Lawson, A. Vdovina, The universal Boolean inverse semigroup presented by the abstract
Cuntz-Krieger relations. J. Noncommut. Geom. (2019, to appear). arXiv:1902.02583v3
[math.OA], February, 28, 2019.
12. C. Starling, Boundary quotients of C∗ -algebras of right LCM semigroups. J. Funct. Anal. 268,
3326–3356 (2015)
13. C. Starling, Inverse semigroups associated to subshifts. J. Algebra 463, 211–233 (2016)
14. M.H. Stone, Postulates for Boolean algebras and generalized Boolean algebras. Am. J. Math.
57, 703–732 (1935)
15. M. Takesaki, Theory of Operator Algebras. I (Springer, Heidelberg, 1979)
16. F. Wehrung, in Refinement Monoids, Equidecomposability Types, and Boolean Inverse Semi-
groups. Lecture Notes in Mathematics, vol. 2188 (Springer, Berlin, 2017)
Calkin Images of Fourier Convolution
Operators with Slowly Oscillating
Symbols

C. A. Fernandes, A. Yu. Karlovich, and Yu. I. Karlovich

3
Abstract Let  be a C ∗ -subalgebra of L∞ (R) and SOX(R) be the Banach algebra
of slowly oscillating Fourier multipliers on a Banach function space X(R). We show
that the intersection of the Calkin image of the algebra generated by the operators of
multiplication aI by functions a ∈  and the Calkin image of the algebra generated
3
by the Fourier convolution operators W 0 (b) with symbols in SOX(R) coincides with
the Calkin image of the algebra generated by the operators of multiplication by
constants.

Keywords Fourier convolution operator · Fourier multiplier · Multiplication

operator · Slowly oscillating function · Calkin algebra · Calkin image

Mathematics Subject Classification (2010) Primary 47G10, Secondary 42A45,

46E30

This work was partially supported by the Fundação para a Ciência e a Tecnologia (Portuguese
Foundation for Science and Technology) through the project UID/MAT/00297/2019 (Centro de
Matemática e Aplicações). The third author was also supported by the SEP-CONACYT Project
A1-S-8793 (México).

C. A. Fernandes () · A. Yu. Karlovich

Centro de Matemática e Aplicações, Departamento de Matemática, Faculdade de Ciências e
Tecnologia, Universidade Nova de Lisboa, Quinta da Torre, Portugal
e-mail: [email protected]; [email protected]
Yu. I. Karlovich
Centro de Investigación en Ciencias, Instituto de Investigación en Ciencias Básicas y Aplicadas,
Universidad Autónoma del Estado de Morelos, Cuernavaca, México
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 193

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_10
194 C. A. Fernandes et al.

1 Introduction

Let F : L2 (R) → L2 (R) denote the Fourier transform


(F f )(x) := f(x) := f (t)eit x dt, x ∈ R,
R

and let F −1 : L2 (R) → L2 (R) be the inverse of F ,


1
(F −1 g)(t) = g(x)e−it x dx, t ∈ R.
2π R

It is well known that the Fourier convolution operator

W 0 (a) := F −1 aF (1.1)

is bounded on the space L2 (R) for every a ∈ L∞ (R).

Let X(R) be a Banach function space and X (R) be its associate space. Their
technical definitions are postponed to Sect. 2.1. The class of Banach function spaces
is very large. It includes Lebesgue, Orlicz, Lorentz spaces, variable Lebesgue
spaces and their weighted analogues (see, e.g., [4, 6]). Let B(X(R)) denote the
Banach algebra of all bounded linear operators acting on X(R), let K(X(R)) be the
closed two-sided ideal of all compact operators in B(X(R)), and let B π (X(R)) =
B(X(R))/K(X(R)) be the Calkin algebra of the cosets Aπ := A+K(X(R)), where
A ∈ B(X(R)).
If X(R) is separable, then L2 (R) ∩ X(R) is dense in X(R) (see Lemma 2.1
below). A function a ∈ L∞ (R) is called a Fourier multiplier on X(R) if the
convolution operator W 0 (a) defined by (1.1) maps L2 (R) ∩ X(R) into X(R) and
extends to a bounded linear operator on X(R). The function a is called the symbol
of the Fourier convolution operator W 0 (a). The set MX(R) of all Fourier multipliers
on X(R) is a unital normed algebra under pointwise operations and the norm
; ;
; ;
aMX(R) := ;W 0 (a); .
B(X(R))

For a unital C ∗ -subalgebra  of the algebra L∞ (R), we consider the quotient

algebra MOπ () consisting of the cosets

[aI ]π := aI + K(X(R))

of multiplication operators by functions in :

MOπ () := {[aI ]π : a ∈ } = {aI + K(X(R)) : a ∈ }.

Calkin Images of Fourier Convolution Operators 195

For a unital Banach subalgebra " of the algebra MX(R) , we also consider the
quotient algebra COπ (") consisting of the cosets

[W 0 (b)]π := W 0 (b) + K(X(R))

of convolution operators with symbols in the algebra ":

COπ (") := {[W 0 (b)]π : b ∈ "} = {W 0 (b) + K(X(R)) : b ∈ "}.

It is easy to see that MOπ () and COπ (") are commutative unital Banach
subalgebras of the Calkin algebra B π (X(R)). It is natural to refer to the algebras
MOπ () and COπ (") as the Calkin images of the algebras

MO() = {aI : a ∈ } ⊂ B(X(R)), CO(") = {W 0 (b) : b ∈ "} ⊂ B(X(R)),

respectively. The algebras MO() and CO(") are building blocks of the algebra
of convolution type operators
 
A(, "; X(R)) = algB(X(R)) aI, W 0 (b) : a ∈ , b ∈ " ,

the smallest closed subalgebra of B(X(R)) that contains the algebras MO() and
CO(").
3
Let SO 3 be the C ∗ -algebra of slowly oscillating functions and SOX(R) be
the Banach algebra of all slowly oscillating Fourier multipliers on the space
X(R), which are defined below in Sects. 2.5–2.7. The third author proved in [22,
Lemma 4.3] in the case of Lebesgue spaces Lp (R, w), 1 < p < ∞, with
Muckenhoupt weights w ∈ Ap (R) that

MOπ (SO 3 ) ∩ COπ (SOL3 p (R,w) ) = MOπ (C), (1.2)

where

MOπ (C) := {[cI ]π : c ∈ C}. (1.3)

This result allowed him to describe the maximal ideal space of the commutative
Banach algebra

Aπ (SO 3 , SOL3 p (R,w) ; Lp (R, w)) = A(SO 3 , SOL3 p (R,w) ; Lp (R, w))/K(Lp (R, w))

(see [22, Theorem 3.1]). In turn, this description plays a crucial role in the study
of the Fredholmness of operators in more general algebras of convolution type
operators with piecewise slowly oscillating data on weighted Lebesgue space
Lp (R, w) (see [22, 24, 25]).
196 C. A. Fernandes et al.

Recall that the (non-centered) Hardy-Littlewood maximal function Mf of a

function f ∈ L1loc (R) is defined by

1
(Mf )(x) := sup |f (y)| dy,
I 4x |I | I

where the supremum is taken over all intervals I ⊂ R of finite length containing
x. The Hardy-Littlewood maximal operator M defined by the rule f → Mf is a
sublinear operator.
The aim of this paper is to extend (1.2) to the case of separable Banach function
spaces such that the Hardy-Littlewood maximal operator M is bounded on X(R)
and on its associate space X (R) and to the case of arbitrary algebras of functions
 ⊂ L∞ (R) in place of SO 3 .
The following statement extends [22, Lemma 4.3].
Theorem 1.1 (Main Result) Let X(R) be a separable Banach function space such
that the Hardy-Littlewood maximal operator M is bounded on the space X(R) and
on its associate space X (R). If  is a unital C ∗ -subalgebra of L∞ (R), then
3
MOπ () ∩ COπ (SOX(R) ) = MOπ (C), (1.4)

where MOπ (C) is defined by (1.3).

This result is one more step towards the study of Fredholm properties of
convolution type operators with discontinuous data on Banach function spaces more
general than weighted Lebesgue spaces initiated in the authors works [8–10].
One can expect, by analogy with the case of weighted Lebesgue spaces, that, for
3
instance, K(X(R)) ⊂ A(SO 3 , SOX(R) ; X(R)) and that the quotient algebra

Aπ (SO 3 , SOX(R)
3
; X(R)) = A(SO 3 , SOX(R)
3
; X(R))/K(X(R))

is commutative. It seems, however, that the proofs of both hypotheses will require
tools, which are not available in the setting of general Banach function spaces.
We plan to return to these questions in a forthcoming work, restricting ourselves
to particular Banach function spaces, like rearrangement-invariant spaces with
Muckenhoupt weights or variable Lebesgue spaces, where interpolation theorems
are available.
The paper is organized as follows. In Sect. 2, we collect necessary facts on
Banach function spaces and Fourier multipliers on them. Further, we recall the
definition of the C ∗ -algebra SO 3 of slowly oscillating functions and introduce
3
the Banach algebra of slowly oscillating Fourier multipliers SOX(R) on a Banach
function space X(R). In Sect. 3, we discuss the structure of the maximal ideal
3
spaces M(SO 3 ) and M(SOX(R) ) of the C ∗ -algebra SO 3 of slowly oscillating
3
functions and the Banach algebra SOX(R) of slowly oscillating Fourier multipliers
on a Banach function space X(R). In particular, we show that the fibers Mt (SO 3 )
Calkin Images of Fourier Convolution Operators 197

of M(SO 3 ) over the points t ∈ Ṙ := R ∪ {∞} can be identified with the fibers
Mt (SOt ), where SOt is the C ∗ -algebra of all bounded continuous functions on
Ṙ \ {t} that slowly oscillate at the point t. An analogous result is also obtained
for the fibers of the maximal ideal spaces of algebras of slowly oscillating Fourier
multipliers on a Banach function space X(R). In Sect. 4, we show that the maximal
ideal spaces of the algebras MOπ () and COπ (") are homeomorphic to the
maximal ideal spaces of the algebras  and ", respectively, where  is a unital C ∗ -
subalgebra of L∞ (R) and " is a unital Banach subalgebra of MX(R). In Sect. 5,
we recall the definition of a limit operator (see [26] for a general theory of limit
operators), as well as, a known fact about limit operators of compact operators acting
on Banach function spaces. Further, we calculate the limit operators of the Fourier
3
convolution operator W 0 (b) with a slowly oscillating symbol b ∈ SOX(R) . Finally,
gathering the above mentioned results on limit operators, we prove Theorem 1.1.

2 Preliminaries

2.1 Banach Function Spaces

The set of all Lebesgue measurable complex-valued functions on R is denoted by

M(R). Let M+ (R) be the subset of functions in M(R) whose values lie in [0, ∞].
The Lebesgue measure of a measurable set E ⊂ R is denoted by |E| and its
characteristic function is denoted by χE . Following [4, Chap. 1, Definition 1.1], a
mapping ρ : M+ (R) → [0, ∞] is called a Banach function norm if, for all functions
f, g, fn (n ∈ N) in M+ (R), for all constants a ≥ 0, and for all measurable subsets
E of R, the following properties hold:

(A1) ρ(f ) = 0 ⇔ f = 0 a.e., ρ(af ) = aρ(f ), ρ(f + g) ≤ ρ(f ) + ρ(g),

(A2) 0 ≤ g ≤ f a.e. ⇒ ρ(g) ≤ ρ(f ) (the lattice property),
(A3) 0 ≤ fn ↑ f a.e. ⇒ ρ(fn ) ↑ ρ(f ) (the Fatou property),
(A4) |E| < ∞ ⇒ ρ(χE ) < ∞,

(A5) |E| < ∞ ⇒ f (x) dx ≤ CE ρ(f )
E

with CE ∈ (0, ∞) which may depend on E and ρ but is independent of f . When

functions differing only on a set of measure zero are identified, the set X(R) of all
functions f ∈ M(R) for which ρ(|f |) < ∞ is called a Banach function space. For
each f ∈ X(R), the norm of f is defined by

f X(R) := ρ(|f |).

198 C. A. Fernandes et al.

Under the natural linear space operations and under this norm, the set X(R) becomes
a Banach space (see [4, Chap. 1, Theorems 1.4 and 1.6]). If ρ is a Banach function
norm, its associate norm ρ is defined on M+ (R) by

ρ (g) := sup f (x)g(x) dx : f ∈ M+ (R), ρ(f ) ≤ 1 , g ∈ M+ (R).
R

It is a Banach function norm itself [4, Chap. 1, Theorem 2.2]. The Banach function
space X (R) determined by the Banach function norm ρ is called the associate
space (Köthe dual) of X(R). The associate space X (R) is naturally identified with
a subspace of the (Banach) dual space [X(R)]∗ .

2.2 Density of Nice Functions in Separable Banach Function

Spaces

As usual, let C0∞ (R) denote the set of all infinitely differentiable functions with
compact support.
Lemma 2.1 ([8, Lemma 2.1] and [23, Lemma 2.12(a)]) If X(R) is a separable
Banach function space, then the sets C0∞ (R) and L2 (R) ∩ X(R) are dense in the
space X(R).
Let S(R) be the Schwartz space of rapidly decreasing smooth functions and let
S0 (R) denote the set of functions f ∈ S(R) such that their Fourier transforms F f
have compact support.
Theorem 2.2 ([10, Theorem 4]) Let X(R) be a separable Banach function space
such that the Hardy-Littlewood maximal operator M is bounded on X(R). Then the
set S0 (R) is dense in the space X(R).

2.3 Banach Algebra MX(R) of Fourier Multipliers

The following result plays an important role in this paper.

Theorem 2.3 ([21, Corollary 4.2] and [8, Theorem 2.4]) Let X(R) be a separable
Banach function space such that the Hardy-Littlewood maximal operator M is
bounded on X(R) and on its associate space X (R). If a ∈ MX(R), then

aL∞ (R) ≤ aMX(R) . (2.1)

The constant 1 on the right-hand side of (2.1) is best possible.

Calkin Images of Fourier Convolution Operators 199

Inequality (2.1) was established earlier in [18, Theorem 1] with some constant
on the right-hand side that depends on the space X(R).
Since (2.1) is available, an easy adaptation of the proof of [13, Proposition 2.5.13]
leads to the following (we refer to the proof of [18, Corollary 1] for details).
Corollary 2.4 Let X(R) be a separable Banach function space such that the
Hardy-Littlewood maximal operator M is bounded on X(R) and on its associate
space X (R). Then the set of Fourier multipliers MX(R) is a Banach algebra under
pointwise operations and the norm  · MX(R) .

2.4 Stechkin-Type Inequality

Let V (R) be the Banach algebra of all functions a : R → C with finite total
variation


n
V (a) := sup |a(ti ) − a(ti−1 )|,
i=1

where the supremum is taken over all finite partitions

−∞ < t0 < t1 < · · · < tn < +∞

of the real line R and the norm in V (R) is given by

aV = aL∞ (R) + V (a).

Theorem 2.5 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on X(R) and on its associate space
X (R). If a ∈ V (R), then the convolution operator W 0 (a) is bounded on the space
X(R) and

W 0 (a)B(X(R)) ≤ cX aV (2.2)

where cX is a positive constant depending only on X(R).

This result follows from [17, Theorem 4.3].
For Lebesgue spaces Lp (R), 1 < p < ∞, inequality (2.2) is usually called
Stechkin’s inequality, and the constant cLp is calculated explicitly:
⎧  
⎪ π
⎨ tan 2p if 1 < p ≤ 2,
cLp = SB(Lp (R)) =   (2.3)
⎪
⎩ cot π if 2 ≤ p < ∞,
2p
200 C. A. Fernandes et al.

where S is the Cauchy singular integral operator given by


1 f (t)
(Sf )(x) := lim dt.
πi ε→0 R\(x−ε,x+ε) t − x

We refer to [7, Theorem 2.11] for the proof of (2.2) in the case of Lebesgue
spaces Lp (R) with cLp = SB(Lp (R)) and to [12, Chap. 13, Theorem 1.3]
for the calculation of the norm of S given in the second equality in (2.3). For
Lebesgue spaces with Muckenhoupt weights Lp (R, w), the proof of Theorem 2.5
with cLp (w) = SB(Lp (R,w)) is contained in [5, Theorem 17.1]. Further, for variable
Lebesgue spaces Lp(·) (R), Theorem 2.5 with cLp(·) = SB(Lp(·) (R)) was obtained
in [20, Theorem 2].

2.5 Slowly Oscillating Functions

Let Ṙ = R ∪ {∞}. For a set E ⊂ Ṙ and a function f : Ṙ → C in L∞ (R), let the

oscillation of f over E be defined by

osc(f, E) := ess sup |f (s) − f (t)|.

s,t ∈E

Following [3, Section 4], [24, Section 2.1], and [25, Section 2.1], we say that a
function f ∈ L∞ (R) is slowly oscillating at a point λ ∈ Ṙ if for every r ∈ (0, 1) or,
equivalently, for some r ∈ (0, 1), one has

lim osc f, λ + ([−x, −rx] ∪ [rx, x]) = 0 if λ ∈ R,
x→0+ 
lim osc f, [−x, −rx] ∪ [rx, x] = 0 if λ = ∞.
x→+∞

For every λ ∈ Ṙ, let SOλ denote the C ∗ -subalgebra of L∞ (R) defined by
 
SOλ := f ∈ Cb (Ṙ \ {λ}) : f slowly oscillates at λ ,

where Cb (Ṙ \ {λ}) := C(Ṙ \ {λ}) ∩ L∞ (R).

Let SO 3 be the smallest C ∗ -subalgebra of L∞ (R) that contains all the C ∗ -
algebras SOλ with λ ∈ Ṙ. The functions in SO 3 are called slowly oscillating
functions.
Calkin Images of Fourier Convolution Operators 201

2.6 Banach Algebra SOλ3 of Three Times Continuously

Differentiable Slowly Oscillating Functions

For a point λ ∈ Ṙ, let C 3 (R \ {λ}) be the set of all three times continuously
differentiable functions a : R \ {λ} → C. Following [24, Section 2.4] and [25,
Section 2.3], consider the commutative Banach algebras

SOλ3 := a ∈ SOλ ∩ C 3 (R \ {λ}) : lim (Dλk a)(x) = 0, k = 1, 2, 3
x→λ

equipped with the norm

 1 ; ;
3
; k ;
aSO 3 := ;Dλ a ; ∞ ,
λ k! L (R)
k=0

where (Dλ a)(x) = (x − λ)a (x) for λ ∈ R and (Dλ a)(x) = xa (x) for λ = ∞.
Lemma 2.6 For every λ ∈ Ṙ, the set SOλ3 is dense in the C ∗ -algebra SOλ .
Proof In view of [2, Lemma 2.3], the set
 
∞
SO∞ := f ∈ SO∞ ∩ Cb∞ (R) : lim (D∞
k
f )(x) = 0, k ∈ N (2.4)
x→∞

is dense in the Banach algebra SO∞ . Here Cb∞ (R) denotes the set of all infinitely
differentiable functions f : R → C, which are bounded with all their derivatives.
Note that SO∞ ∞ can be equivalently defined by replacing C ∞ (R) in (2.4) by C ∞ (R),
b
because f ∈ SO∞ is bounded and its derivatives f (k) are bounded for all k ∈ N in
view of limx→∞ (D∞ k f )(x) = 0. Since SO ∞ ⊂ SO 3 , this completes the proof in
∞ ∞
the case λ = ∞.
If λ ∈ R, then by Karlovich and Loreto Hernández [25, Corollary 2.2], the
mapping T a = a ◦ βλ , where βλ : Ṙ → Ṙ is defined by

λx − 1
βλ (x) = , (2.5)
x+λ

is an isometric isomorphism of the algebra SOλ onto the algebra SO∞ . Hence each
function a ∈ SOλ can be approximated in the norm of SOλ by functions cn =
bn ◦ βλ−1 , where bn ∈ SO∞∞ for n ∈ N and

λy + 1
βλ−1 (y) = = x, x, y ∈ Ṙ. (2.6)
λ−y
202 C. A. Fernandes et al.

It remains to show that cn ∈ SOλ3 . Taking into account (2.5)–(2.6), we obtain for
y = βλ (x) ∈ R \ {λ} and x = βλ−1 (y) ∈ R:

  λ2 + 1
(Dλ cn )(y) =bn βλ−1 (y) = −bn (x)(x + λ), (2.7)
y−λ
  (λ2 + 1)2   λ2 + 1
(Dλ2 cn )(y) =bn βλ−1 (y) − bn βλ−1 (y)
y−λ y−λ
= − bn (x)(x + λ)(λ2 + 1) + bn (x)(x + λ), (2.8)

  (λ2 + 1)3   (λ2 + 1)2

(Dλ3 cn )(y) =bn βλ−1 (y) − 2bn βλ−1 (y)
y−λ y−λ
  λ2 + 1
+ bn βλ−1 (y)
y−λ
= − bn (x)(x + λ)(λ2 + 1)2 + 2bn (x)(x + λ)(λ2 + 1)
− bn (x)(x + λ). (2.9)

Since
k
lim (D∞ bn )(x) = 0 for k ∈ {1, 2, 3},
x→∞

we see that

lim x k bn(k)(x) = 0 for k ∈ {1, 2, 3}. (2.10)

x→∞

It follows from (2.7)–(2.10) that

lim (Dλk cn )(y) = 0 for k ∈ {1, 2, 3}.

y→λ

Hence cn ∈ SOλ3 for all n ∈ N, which completes the proof.

2.7 Slowly Oscillating Fourier Multipliers

The following result leads us to the definition of slowly oscillating Fourier multipli-
ers.
Theorem 2.7 ([19, Theorem 2.5]) Let X(R) be a separable Banach function space
such that the Hardy-Littlewood maximal operator M is bounded on X(R) and on
Calkin Images of Fourier Convolution Operators 203

its associate space X (R). If λ ∈ Ṙ and a ∈ SOλ3 , then the convolution operator
W 0 (a) is bounded on the space X(R) and

W 0 (a)B(X(R)) ≤ cX aSO 3 , (2.11)

λ

where cX is a positive constant depending only on X(R).

Let SOλ,X(R) denote the closure of SOλ3 in the norm of MX(R). Further, let
3
SOX(R) be the smallest Banach subalgebra of MX(R) that contains all the Banach
3
algebras SOλ,X(R) for λ ∈ Ṙ. The functions in SOX(R) will be called slowly
oscillating Fourier multipliers.
Lemma 2.8 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on X(R) and on its associate space
X (R). Then
3
SOX(R) ⊂ SOL3 2 (R) = SO 3 .

3
Proof The continuous embedding SOX(R) ⊂ SOL3 2 (R) (with the embedding
constant one) follows immediately from Theorem 2.3 and the definitions of the
3
Banach algebras SOX(R) and SOL3 2 (R) . It is clear that SOL3 2 (R) ⊂ SO 3 . The
embedding SO 3 ⊂ SOL3 2 (R) follows from Lemma 2.6.


3 Maximal Ideal Spaces of the Algebras SO  and SOX(R)

3.1 Extensions of Multiplicative Linear Functionals

on C ∗ -algebras

For a C ∗ -algebra (or, more generally, a Banach algebra) A with unit e and an element
a ∈ A, let spA (a) denote the spectrum of a in A. Recall that an element a of a C ∗ -
algebra A is said to be positive if it is self-adjoint and spA (a) ⊂ [0, ∞). A linear
functional φ on A is said to be a state if φ(a) ≥ 0 for all positive elements a ∈ A
and φ(e) = 1. The set of all states of A is denoted by S(A). The extreme points of
S(A) are called pure states of A (see, e.g., [15, Section 4.3]).
Following [1, p. 304], for a state φ, let

Gφ (A) := {a ∈ A : |φ(a)| = aA = 1}

and let Gφ+ (A) denote the set of all positive elements of Gφ (A). Let A and B be
C ∗ -algebras such that e ∈ B ⊂ A. Let φ be a state of B. Following [1, p. 310], we
say that A is B-compressible modulo φ if for each x ∈ A and each ε > 0 there is
b ∈ Gφ+ (B) and y ∈ B such that bxb − yA < ε.
204 C. A. Fernandes et al.

Since a nonzero linear functional on a commutative C ∗ -algebra is a pure state if

and only if it is multiplicative (see, e.g., [15, Proposition 4.4.1]), we immediately
get the following lemma from [1, Theorem 3.2].
Lemma 3.1 Let B be a C ∗ -subalgebra of a commutative C ∗ -algebra A. A nonzero
multiplicative linear functional φ on B admits a unique extension to a multiplicative
linear functional φ on A if and only if A is B-compressible modulo φ.

3.2 Family of Positive Elements

For t ∈ Ṙ and ω > 0, let ψt,ω be a real-valued function in C(Ṙ) such that 0 ≤
ψt,ω (x) ≤ 1 for all x ∈ R. Assume that for t ∈ R,

ψt,ω (s) = 1 if s ∈ (t − ω, t + ω), ψt,ω (s) = 0 if s ∈ R \ (t − 2ω, t + 2ω),

and for t = ∞,

ψ∞,ω (s) = 1 if s ∈ R \ (−2ω, 2ω), ψ∞,ω (s) = 0 if s ∈ (−ω, ω).

Let M(A) denote the maximal ideal space of a commutative Banach algebra A.
Lemma 3.2 For t ∈ Ṙ and ω > 0, the function ψt,ω is a positive element of the
C ∗ -algebras C(Ṙ), SOt , and SO 3 .
Proof Since M(C(Ṙ)) = Ṙ, it follows from the Gelfand theorem (see, e.g., [28,
Theorem 2.1.3]) that spC(Ṙ) (ψt,ω ) = [0, 1] for all t ∈ Ṙ and all ω > 0. Since
C(Ṙ) ⊂ SOt ⊂ SO 3 , we conclude that the functions ψt,ω for t ∈ Ṙ and ω > 0 are
positive elements of the C ∗ -algebras C(Ṙ), SOt , and SO 3 because their spectra in
each of these algebras coincide with [0, 1] in view of [15, Proposition 4.1.5].

3.3 Maximal Ideal Space of the C ∗ -algebra SO 

If B is a Banach subalgebra of A and λ ∈ M(B), then the set

Mλ (A) := {ξ ∈ M(A) : ξ |B = λ}

is called the fiber of M(A) over λ ∈ M(B). Hence for every Banach algebra  ⊂
L∞ (R) with M(C(Ṙ) ∩ ) = Ṙ and every t ∈ Ṙ, the fiber Mt () is the set of
all multiplicative linear functionals (characters) on  that annihilate the set {f ∈
C(Ṙ) ∩  : f (t) = 0}. As usual, for all a ∈  and all ξ ∈ M(), we put a(ξ ) :=
ξ(a). We will frequently identify the points t ∈ Ṙ with the evaluation functionals δt
Calkin Images of Fourier Convolution Operators 205

defined by

δt (f ) = f (t) for f ∈ C(Ṙ), t ∈ Ṙ.

Lemma 3.3 For every point t ∈ Ṙ, the fibers Mt (SOt ) and Mt (SO 3 ) can be
identified as sets:

Mt (SOt ) = Mt (SO 3 ). (3.1)

Proof Since C(Ṙ) ⊂ SOt ⊂ SO 3 , by the restriction of a multiplicative linear

functional defined on a bigger algebra to a smaller algebra, we have

M(SO 3 ) ⊂ M(SOt ) ⊂ M(C(Ṙ)), t ∈ Ṙ. (3.2)

Since
<
M() = Mt () for  ∈ {SO 3 , SOλ : λ ∈ Ṙ},
t ∈Ṙ

where

Mt () = {ζ ∈ M() : ζ |C(Ṙ) = δt }, t ∈ Ṙ, (3.3)

it follows from (3.2) and (3.3) that

Mt (SO 3 ) ⊂ Mt (SOt ), t ∈ R. (3.4)

Now fix t ∈ Ṙ and a multiplicative linear functional η ∈ Mt (SOt ). Let us

show that the C ∗ -algebra SO 3 is SOt -compressible modulo η. Take ε > 0. By
the definition of SO 3 , for a function x ∈ SO 3 , there are a finite set F ∈ Ṙ and a
finite set {xλ ∈ SOλ : λ ∈ F } such that
; ;
;  ;
; ;
;x − xλ ; < ε.
; ;
λ∈F L∞ (R)

If t = ∞, take ω such that

1
0<ω< min |λ − t|
2 λ∈F \{t }

and b := ψt,ω . Then



y := b xλ b (3.5)
λ∈F

is equal to zero outside the interval (t − 2ω, t + 2ω). Therefore, y ∈ SOt .

206 C. A. Fernandes et al.

If t = ∞, take ω such that

ω> max |λ|

λ∈F \{∞}

and b := ψ∞,ω . Then the function y defined by (3.5) is equal to zero on (−ω, ω)
and y ∈ SO∞ .
For t ∈ Ṙ, we have
;  ; ; ;
;  ; ;  ;
; ; ; ;
bxb − yL∞ (R) = ;b x − xλ b ; ≤ ;x − xλ ; < ε.
; ; ; ;
λ∈F L∞ (R) λ∈F L∞ (R)

Since b is a positive element of SOt in view of Lemma 3.2, we have b ∈ Gη+ (SOt ),
which completes the proof of the fact that SO 3 is SOt -compressible modulo the
multiplicative linear functional η ∈ Mt (SOt ).
In view of Lemma 3.1, there exists a unique extension η of the multiplicative
linear functional η to the whole algebra SO 3 . By the definition of the fiber
Mt (SO 3 ), we have η ∈ Mt (SO 3 ). Thus, we can identify Mt (SOt ) with a subset of
Mt (SO 3 ):

Mt (SOt ) ⊂ Mt (SO 3 ). (3.6)

Combining (3.4) and (3.6), we arrive at (3.1).

Corollary 3.4 The maximal ideal space of the commutative C ∗ -algebra SO 3 can
be identified with the set
<
Mt (SOt ).
t ∈Ṙ

3.4 Extensions of Multiplicative Linear Functionals

on Banach Algebras

The following theorem in a slightly different form is contained in [29, Theo-

rem 2.1.1] and [30, Theorem 3.10]. For the convenience of readers, we give its
proof here.
Theorem 3.5 Let A, B, C be commutative unital Banach algebras with common
unit and homomorphic imbeddings A ⊂ B ⊂ C, where A is dense in B. If for each
functional ϕ ∈ M(A) there exists a unique extension ϕ ∈ M(C), then for every
functional ψ ∈ M(B) there exists a unique extension ψ ∈ M(C).
Proof Let ψ ∈ M(B). Then ψ1 := ψ|A ∈ M(A). By the hypotheses, there exists a
unique extension ψ3 := (ψ1 ) ∈ M(C). Then ψ1 (a) = ψ(a) = ψ3 (a) for all a ∈ A.
Calkin Images of Fourier Convolution Operators 207

Let ψ2 := ψ3 |B ∈ M(B). Since A ⊂ B, it follows that

ψ(a) = ψ2 (a) for all a ∈ A. (3.7)

On the other hand, functionals ψ, ψ2 ∈ M(B) are continuous on B (see, e.g., [16,
Lemma 2.1.5]). Since A is dense in B, for every b ∈ B there exists a sequence
{an }n∈N ⊂ A such that an − bB → 0 as n → ∞. It follows from this observation
and (3.7) that for every b ∈ B,

ψ(b) = lim ψ(an ) = lim ψ2 (an ) = ψ2 (b) = ψ3 (b).

n→∞ n→∞

Thus ψ3 ∈ M(C) is an extension of ψ. This extension is unique by construction.

3.5 Maximal Ideal Space of the Banach Algebras SOt,X(R)

We start with the following refinement of [25, Lemma 3.4].

Lemma 3.6 Let t ∈ Ṙ. Then for each functional ϕ ∈ M(SOt3 ) there exists a unique
extension ϕ ∈ M(SOt ).
The density of SOt3 in the Banach algebra SOt essentially used in the proof of
[25, Lemma 3.4] is justified in Lemma 2.6. Note that the uniqueness of an extension
was not explicitly mentioned in [25, Lemma 3.4]. However, since M(SOt3 ) and
M(SOt ) are Hausdorff spaces (see, e.g., [16, Theorem 2.2.3]), the uniqueness of
an extension constructed in the proof of [25, Lemma 3.4] is a consequence of a
standard fact from general topology (see, e.g., [27, Theorem IV.2(b)]).
The following lemma is analogous to [25, Lemma 3.5].
Lemma 3.7 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on the space X(R) and on its associate
space X (R). If t ∈ Ṙ, then the maximal ideal spaces of the C ∗ -algebra SOt and
the Banach algebra SOt,X(R) can be identified as sets:

M(SOt ) = M(SOt,X(R)). (3.8)

Proof It follows from Theorem 2.3 that SOt3 ⊂ SOt,X(R) ⊂ SOt , where the
imbeddings are homomorphic. By the definition of the algebra SOt,X(R), the algebra
SOt3 is dense in SOt,X(R) with respect to the norm of MX(R). Taking into account
these observations and Lemma 3.6, we see that the commutative Banach algebras

A = SOt3 , B = SOt,X(R), C = SOt

satisfy all the conditions of Theorem 3.5. By this theorem, every multiplicative
linear functional on SOt,X(R) admits a unique extension to a multiplicative linear
208 C. A. Fernandes et al.

functional on SOt . Hence we can identify M(SOt,X(R)) with a subset of M(SOt ):

M(SOt,X(R)) ⊂ M(SOt ). (3.9)

On the other hand, since SOt,X(R) ⊂ SOt , by the restriction of a multiplicative

linear functional defined on a bigger algebra to a smaller algebra, we have

M(SOt ) ⊂ M(SOt,X(R)). (3.10)

Combining inclusions (3.9) and (3.10), we immediately arrive at (3.8).

The next lemma is analogous to Lemma 3.3.
Lemma 3.8 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on the space X(R) and on its associate
3
space X (R). Then, for every point t ∈ Ṙ, the fibers Mt (SOt,X(R)) and Mt (SOX(R) )
can be identified as sets:
3
Mt (SOt,X(R)) = Mt (SOX(R) ). (3.11)

3
Proof Since SOt,X(R) ⊂ SOX(R) for every t ∈ Ṙ, we conclude by the restriction of
a multiplicative linear functional defined on the bigger algebra to the smaller algebra
3
that M(SOX(R) ) ⊂ M(SOt,X(R)). Hence

3
Mt (SOX(R) ) ⊂ Mt (SOt,X(R)). (3.12)

On the other hand, in view of Lemma 3.7, any multiplicative linear functional ξ ∈
Mt (SOt,X(R)) admits a unique extension ξ ∈ M(SOt ). Moreover, ξ belongs to
Mt (SOt ) as well. By Lemma 3.3, the functional ξ ∈ Mt (SOt ) admits a unique
3
extension ξ ∈ Mt (SO 3 ). It is clear that the restriction of ξ to SOX(R) belongs to
3 3
Mt (SOX(R) ). Thus Mt (SOt,X(R)) can be identified with a subset of Mt (SOX(R) ):

3
Mt (SOt,X(R)) ⊂ Mt (SOX(R) ). (3.13)

Combining (3.12) and (3.13), we arrive at (3.11).


3.6 Maximal Ideal Space of the Banach Algebra SOX(R)

Now we are in a position to prove that the maximal ideal spaces of the commutative
3
Banach algebra SOX(R) and the C ∗ -algebra SO 3 can be identified as sets.
Theorem 3.9 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on the space X(R) and on its associate
Calkin Images of Fourier Convolution Operators 209

3
space X (R). Then the maximal ideal space of the Banach algebra SOX(R) can be
∗ 3
identified with the maximal ideal space of the C -algebra SO :
3
M(SOX(R) ) = M(SO 3 ).

Proof It follows from Lemmas 3.8, 3.7, and 3.3 that for every t ∈ Ṙ,
3
Mt (SOX(R) ) = Mt (SOt,X(R)) = Mt (SOt ) = Mt (SO 3 ).

Hence
< <
3 3
M(SOX(R) )= Mt (SOX(R) )= Mt (SO 3 ) = M(SO 3 ),
t ∈Ṙ t ∈Ṙ

which completes the proof.

4 Maximal Ideal Spaces of the Calkin Images of the Banach

Algebras MO() and CO()

4.1 Maximal Ideal Space of the Algebra MOπ ()

We start with the following known result [14, Theorem 2.4] (see also [9, Theo-
rem 3.1]).
Theorem 4.1 Let X(R) be a separable Banach function space and a ∈ L∞ (R).
Then the multiplication operator aI is compact on the space X(R) if and only if
a = 0 almost everywhere on R.
The next theorem says that one can identify the maximal ideal spaces of the
algebras MOπ () and  for an arbitrary unital C ∗ -subalgebra of L∞ (R).
Theorem 4.2 Let X(R) be a separable Banach function space. If  is a unital C ∗ -
subalagebra of L∞ (R), then the maximal ideal spaces of the commutative Banach
algebra MOπ () and the commutative C ∗ -algebra  are homeomorphic:

M(MOπ ()) = M().

Proof Consider the mapping F :  → MOπ () defined by F (a) = [aI ]π for
every a ∈ . It is clear that this mapping is surjective. If [aI ]π = [bI ]π for some
a, b ∈ , then (a − b)I ∈ K(X(R)). It follows from Theorem 4.1 that a = b a.e.
on R. This implies that the mapping F is injective. Thus, F :  → MOπ ()
is an algebraic isomorphism of commutative Banach algebras. It follows from
[16, Lemma 2.2.12] that the maximal ideal spaces M(MOπ ()) and M() are
homeomorphic.
210 C. A. Fernandes et al.

4.2 Maximal Ideal Space of the Algebra COπ ()

The following analogue of Theorem 4.1 for Fourier convolution operators was
obtained recently by the authors [8, Theorem 1.1].
Theorem 4.3 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on X(R) and on its associate space
X (R). Suppose that b ∈ MX(R). Then the Fourier convolution operator W 0 (a) is
compact on the space X(R) if and only if b = 0 almost everywhere on R.
The next theorem is an analogue of Theorem 4.2 for Fourier multipliers.
Theorem 4.4 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on X(R) and on its associate space
X (R). If " is a unital Banach subalagebra of MX(R), then the maximal ideal
spaces of the commutative Banach algebras COπ (") and " are homeomorphic:

M(COπ (")) = M(").

Proof The proof is analogous to the proof of Theorem 4.2. Consider the mapping
F : " → COπ (") defined by F (a) = [W 0 (a)]π for every a ∈ ". It is obvious
that this mapping is surjective. If [W 0 (a)]π = [W 0 (b)]π for some a, b ∈ ", then
W 0 (a − b) = W 0 (a) − W 0 (b) ∈ K(X(R)). By Theorem 4.3, we conclude that
a = b a.e. on R. Therefore, the mapping F is injective. Thus, F : " → COπ (") is
an algebraic isomorphism of commutative Banach algebras. In this case it follows
from [16, Lemma 2.2.12] that the maximal ideal spaces M(COπ (")) and M(") are
homeomorphic.

5 Applications of the Method of Limit Operators

5.1 Known Result about Limit Operators on Banach Function

Spaces

Let X(R) be a Banach function space. For a sequence of operators {An }n∈N ⊂
B(X(R)), let

s-lim An
n→∞

denote the strong limit of this sequence, if it exists. For λ, x ∈ R, consider the
function

eλ (x) := eiλx .
Calkin Images of Fourier Convolution Operators 211

Let T ∈ B(X(R)) and let h = {hn }n∈N be a sequence of numbers hn > 0 such
that hn → +∞ as n → ∞. The strong limit

Th := s- lim ehn T eh−1

n
I
n→∞

is called the limit operator of T related to the sequence h = {hn }n∈N , if it exists.
In our previous paper [9] we calculated the limit operators for all compact
operators.
Lemma 5.1 ([9, Lemma 3.2]) Let X(R) be a separable Banach function space and
K be a compact operator on X(R). Then for every sequence {hn }n∈N of positive
numbers satisfying hn → +∞ as n → ∞, one has

s-lim ehn Keh−1

n
I =0
n→∞

on the space X(R).

5.2 Limit Operators for Fourier Convolution Operators


with Symbols in the Algebra SOX(R)

Now we will calculate the limit operators for the Fourier convolution operator with
a slowly oscillating symbol.
Theorem 5.2 Let X(R) be a separable Banach function space such that the Hardy-
Littlewood maximal operator M is bounded on the space X(R) and on its associate
3
space X (R). If b ∈ SOX(R) , then for every ξ ∈ M∞ (SO 3 ) there exists a sequence
{hn }n∈N of positive numbers such that hn → +∞ as n → ∞ and

s-lim ehn W 0 (b)eh−1

n
I = b(ξ )I (5.1)
n→∞

on the space X(R).

Proof This statement is proved by analogy with [25, Lemma 5.1]. In view of
3
Lemma 2.8, SOX(R) ⊂ SO 3 . Therefore every ξ ∈ M∞ (SO 3 ) is a multiplicative
3
linear functional on SOX(R) , that is, b(ξ ) is well defined. By the definition of
3 3
SOX(R), if b ∈ SOX(R) , then there is a sequence

bm = bm,λ , m ∈ N,
λ∈Fm

where Fm ⊂ Ṙ are finite sets and bm,λ ∈ SOλ3 for λ ∈ Fm and all m ∈ N, such that

lim bm − bMX(R) = 0. (5.2)

m→∞
212 C. A. Fernandes et al.

By Lemma 3.3, M∞ (SO 3 ) = M∞ (SO∞ ). Fix ξ ∈ M∞ (SO 3 ) = M∞ (SO∞ ).

Assume first that the set

B∞ := {bm,∞ ∈ SO∞
3
: m ∈ N}

is not empty. Since the set B∞ is at most countable, it follows from [2, Corollary 3.3]
or [25, Proposition 3.1] that there exists a sequence {hn }n∈N such that hn → +∞
as n → ∞ and

ξ(bm,∞ ) = lim bm,∞ (hn ) for all bm,∞ ∈ B∞ . (5.3)

n→∞

As the functions bm,λ are continuous at ∞ if λ = ∞, we see that

<
ξ(bm,λ ) = bm,λ (∞) = lim bm,λ (hn ) for all λ ∈ Fm \ {∞}. (5.4)
n→∞
m∈N

Combining (5.3) and (5.4), for every m ∈ N, we get

 
ξ(bm ) = ξ(bm,λ ) = lim bm,λ (hn )
n→∞
λ∈Fm λ∈Fm

= lim bm,λ (hn ) = lim bm (hn ). (5.5)
n→∞ n→∞
λ∈Fm

If the set B∞ is empty, we can take an arbitrary sequence {hn }n∈N such that hn →
+∞ as n → ∞.
Let f ∈ S0 (R). Then, by a smooth version of Urysohn’s lemma (see, e.g., [11,
Proposition 6.5]), there is a function ψ ∈ C0∞ (R) such that 0 ≤ ψ ≤ 1, supp F f ⊂
supp ψ and ψ|supp F f = 1. Therefore, for all n ∈ N,

ehn W 0 (b)eh−1
n
f − b(ξ )f = W 0 [b(· + hn )]f − ξ(b)f

= F −1 [b(· + hn ) − ξ(b)]ψF f

and
;  ; ;4 5
; ehn W 0 (b)e−1 − b(ξ ) f ; ≤ ; b(· + hn ) − ξ(b) ψMX(R) f X(R) . (5.6)
hn X(R)

Since MX(R) is translation-invariant and ξ ∈ M∞ (SO 3 ) is a multiplicative linear

3
functional on SOX(R) , we infer for all m, n ∈ N that
;4 5 ;4 5
; b(· + hn ) − ξ(b) ψM ≤; b(· + hn ) − bm (· + hn ) ψMX(R)
X(R)
;4 5
+ ; bm (· + hn ) − ξ(bm ) ψMX(R)
Calkin Images of Fourier Convolution Operators 213

; 5
+ ;ξ(bm ) − ξ(b) ψMX(R)
≤2b − bm MX(R) ψMX(R)
;4 5
+ ; bm (· + hn ) − ξ(bm ) ψMX(R) . (5.7)

Fix ε > 0. By Theorem 2.5, ψMX(R) < ∞. It follows from (5.2) that there
exists a sufficiently large number m ∈ N (which we fix until the end of the proof)
such that

2b − bm MX(R) ψMX(R) < ε/2. (5.8)

Let
⎧
⎨ max |λ| if Fm \ {∞} = ∅,
$ := λ∈Fm \{∞}
⎩0 if Fm \ {∞} = ∅,

let K := supp ψ and

k := max{− inf K, sup K} ∈ [0, ∞).

For x ∈ K and n ∈ N, let In (x) be the segment with the endpoints hn and x + hn .
Then In (x) ⊂ [hn − k, hn + k]. Since hn → +∞ as n → ∞, there exists N1 ∈ N
such that for all n > N1 , one has

In (x) ⊂ [hn − k, hn + k] ⊂ ($, ∞).

For all n > N1 , we have

;4 5 ; ;4 5 ;
; bm (· + hn ) − ξ(bm ) ψ ; ≤ ; bm (· + hn ) − bm (hn ) ψ ;M
MX(R) X(R)

+ bm (hn ) − ξ(bm ) ψMX(R) , (5.9)

4 5
where the functions bm (· + hn ) − bm (hn ) ψ for n > N1 belong to SO∞ 3 because

they are three times continuously differentiable functions of compact support.

By (5.5), there exists N2 ∈ N such that N2 ≥ N1 and for all n > N2 ,

bm (hn ) − ξ(bm ) ψMX(R) < ε/4. (5.10)

4 5
On the other hand, since bm (· + hn ) − bm (hn ) ψ ∈ SO∞
3 for all n > N , it follows
2
from Theorem 2.7 that there exists a constant cX > 0 depending only on the space
214 C. A. Fernandes et al.

X(R) such that for all n > N2 ,

;4 5 ;
; bm (· + hn ) − bm (hn ) ψ ;
MX(R)
;4 5 ;
≤ cX ; bm (· + hn ) − bm (hn ) ψ ; 3
SO∞

 1 ; 5 ;
3
; j 4 ;
= cX ;D∞ bm (· + hn ) − bm (hn ) ψ ; ∞ . (5.11)
j! L (R)
j =0

For all j ∈ {0, 1, 2, 3}, we have

j 4 5 
D∞ bm (· + hn ) − bm (hn ) ψ
j  
 j 4 5  j −ν 
= ν
D∞ bm (· + hn ) − bm (hn ) D∞ ψ . (5.12)
ν
ν=0

It follows from the mean value theorem that


;4 5 ; x+hn
; bm (· + hn ) − bm (hn ) χK ; = sup bm (t) dt
L∞ (R)
x∈K hn
 x+hn dt
= sup tbm (t)
x∈K hn t

dt
≤ sup (D∞ bm )(t)
x∈K In (x) t
 hn +k dt
≤ sup (D∞ bm )(t)
t ∈[hn −k,hn +k] hn −k t
hn + k
≤ ln D∞ bm L∞ (R) . (5.13)
hn − k

It is easy to see that for x ∈ K,

4 5 x
D∞ bm (· + hn ) − bm (hn ) (x) = (D∞ bm )(x + hn ), (5.14)
x + hn
2
4 5
D∞ bm (· + hn ) − bm (hn ) (x)
x2 xhn
= (D 2 bm )(x + hn ) + (D∞ bm )(x + hn ), (5.15)
(x + hn )2 ∞ (x + hn )2
Calkin Images of Fourier Convolution Operators 215

and
3
4 5
D∞ bm (· + hn ) − bm (hn ) (x)
x3 3 3x 2 hn
= (D∞ bm )(x + hn ) + (D 2 bm )(x + hn )
(x + hn ) 3 (x + hn )3 ∞
xh2n − x 2 hn
+ (D∞ bm )(x + hn ). (5.16)
(x + hn )3

It follows from (5.14)–(5.16) that for all n > N2 ,

; 4 5 ; k
; D∞ bm (· + hn ) − bm (hn ) χK ; ∞ ≤ D∞ bm L∞ (R) , (5.17)
L (R) hn − k
; 5 ;
; 2 4 ;
; D∞ bm (· + hn ) − bm (hn ) χK ;
L∞ (R)

k2 ; ;
; 2 ; khn
≤ ;D∞ b m ; ∞ + D∞ bm L∞ (R) , (5.18)
(hn − k) 2 L (R) (hn − k)2

and
; 5 ;
; 3 4 ;
; D∞ bm (· + hn ) − bm (hn ) χK ;
L∞ (R)

k3 ; ; 3k 2 hn ; ;
; 3 ; ; 2 ;
≤ ;D∞ b m ; + ;D∞ b m ; ∞
(hn − k) 3 ∞
L (R) (hn − k) 3 L (R)

kh2n + k 2 hn
+ D∞ bm L∞ (R) . (5.19)
(hn − k)3

Since
; ;
; j ;
max ;D∞ ψ ; < ∞,
j ∈{0,1,2,3} L∞ (R)

it follows from (5.12)–(5.13) and (5.17)–(5.19) that for all j ∈ {0, 1, 2, 3},
; 5 ;
; j 4 ;
lim ;D∞ bm (· + hn ) − bm (hn ) ψ ; = 0. (5.20)
n→∞ L∞ (R)

We deduce from (5.11) and (5.20) that there exists N3 ∈ N such that N3 ≥ N2 and
for all n > N3 ,
;4 5 ;
; bm (· + hn ) − bm (hn ) ψ ; < ε/4. (5.21)
MX(R)
216 C. A. Fernandes et al.

Combining (5.7)–(5.10) and (5.21), we see that for every f ∈ S0 (R) and every
ε > 0 there exists N3 ∈ N such that for all n > N3 ,
;  ;
; ;
; ehn W 0 (b)eh−1
n
− b(ξ ) f; < εf X(R) ,
X(R)

whence for all f ∈ S0 (R),

;  ;
; ;
lim ; ehn W 0 (b)eh−1
n
I − b(ξ )I f; = 0.
n→∞ X(R)

Since S0 (R) is dense in X(R) (see Theorem 2.2), this equality immediately implies
(5.1) in view of [28, Lemma 1.4.1(ii)], which completes the proof.

5.3 Proof of Theorem 1.1

Since the function e0 ≡ 1 belongs to  and "SOX(R) 3 , we see that the set of all
3
constant functions is contained in  and in SOX(R) . Therefore

3
MOπ (C) ⊂ MOπ () ∩ COπ (SOX(R) ). (5.22)

3
Let Aπ ∈ MOπ () ∩ COπ (SOX(R) ). Then Aπ = [aI ]π = [W 0 (b)]π , where
3
a ∈  and b ∈ SOX(R) . Therefore, there is an operator K ∈ K(X(R)) such that

aI = W 0 (b) + K. (5.23)

By Theorem 5.2, for every ξ ∈ M∞ (SO 3 ) there exists a sequence {hn }n∈N of
positive numbers such that hn → +∞ as n → ∞ and

s-lim ehn W 0 (b)eh−1

n
I = b(ξ )I. (5.24)
n→∞

Equalities (5.23)–(5.24) and Lemma 5.1 imply that

aI = s-lim ehn (aI )eh−1

n
I = s-lim ehn (W 0 (b) + K)eh−1
n
I = b(ξ )I.
n→∞ n→∞

Hence [aI ]π = [b(ξ )I ]π ∈ MOπ (C) and

3
MOπ () ∩ COπ (SOX(R) ) ⊂ MOπ (C). (5.25)

Combining (5.22) and (5.25), we arrive at (1.4).

Calkin Images of Fourier Convolution Operators 217

Acknowledgments We would like to thank the anonymous referee for pointing out a gap in the
original version of the paper. To fill in this gap, we strengthened the hypotheses in the main result.

References

1. J. Anderson, Extensions, restrictions, and representations of states on C ∗ -algebras. Trans. Am.

Math. Soc. 249, 303–329 (1979)
2. M.A. Bastos, A. Bravo, Y.I. Karlovich, Convolution type operators with symbols generated by
slowly oscillating and piecewise continuous matrix functions. Oper. Theory Adv. Appl. 147,
151–174 (2004)
3. M.A. Bastos, C.A. Fernandes, Y.I. Karlovich, C ∗ -algebras of integral operators with piecewise
slowly oscillating coefficients and shifts acting freely. Integr. Equ. Oper. Theory 55, 19–67
(2006)
4. C. Bennett, R. Sharpley, Interpolation of Operators (Academic Press, Boston, 1988)
5. A. Böttcher, Y.I. Karlovich, I.M. Spitkovsky, Convolution Operators and Factorization of
Almost Periodic Matrix Functions (Birkhäuser, Basel, 2002)
6. D. Cruz-Uribe, A. Fiorenza, Variable Lebesgue Spaces (Birkhäuser/Springer, New York, 2013)
7. R.V. Duduchava, Integral Equations with Fixed Singularities (Teubner, Leipzig, 1979)
8. C.A. Fernandes, A.Y. Karlovich, Y.I. Karlovich, Noncompactness of Fourier convolution
operators on Banach function spaces. Ann. Funct. Anal. AFA 10, 553–561 (2019)
9. C.A. Fernandes, A.Y. Karlovich, Y.I. Karlovich, Algebra of convolution type operators with
continuous data on Banach function spaces. Banach Center Publ. 119, 157–171 (2019)
10. C.A. Fernandes, A.Y. Karlovich, Y.I. Karlovich, Fourier convolution operators with symbols
equivalent to zero at infinity on Banach function spaces, in Proceedings of ISAAC (2019, to
appear). arXiv:1909.13538 [math.FA]
11. G.B. Folland, A Guide to Advanced Real Analysis (The Mathematical Association of America,
Washington, 2009)
12. I. Gohberg, N. Krupnik, One-Dimensional Linear Singular Integral Equations, vol. II
(Birkhäuser, Basel, 1992)
13. L. Grafakos, Classical Fourier Analysis, 3rd ed. (Springer, New York, 2014)
14. H. Hudzik, R. Kumar, R. Kumar, Matrix multiplication operators on Banach function spaces.
Proc. Indian Acad. Sci. Math. Sci. 116, 71–81 (2006)
15. R.V. Kadison, J.R. Ringrose, Fundamentals of the Theory of Operator Algebras, in Elementary
Theory, vol. I, 2nd ed. (American Mathematical Society, Providence, 1997)
16. E. Kaniuth, A Course in Commutative Banach Algebras (Springer, New York, 2009)
17. A.Y. Karlovich, Maximally modulated singular integral operators and their applications to
pseudodifferential operators on Banach function spaces. Contemp. Math. 645, 165–178 (2015)
18. A.Y. Karlovich, Banach algebra of the Fourier multipliers on weighted Banach function spaces.
Concr. Oper. 2, 27–36 (2015)
19. A.Y. Karlovich, Commutators of convolution type operators on some Banach function spaces.
Ann. Funct. Anal. AFA 6, 191–205 (2015)
20. A.Y. Karlovich, The Stechkin inequality for Fourier multipliers on variable Lebesgue spaces.
Math. Inequal. Appl. 18, 1473–1481 (2015)
21. A. Karlovich, E. Shargorodsky, When does the norm of a Fourier multiplier dominate its L∞
norm? Proc. London Math. Soc. 118, 901–941 (2019)
22. Y.I. Karlovich, Algebras of convolution-type operators with piecewise slowly oscillating data
on weighted Lebesgue spaces. Mediterr. J. Math. 14, paper no. 182, 20 (2017)
23. A.Y. Karlovich, I.M. Spitkovsky, The Cauchy singular integral operator on weighted variable
Lebesgue spaces. Oper. Theory Adv. Appl. 236, 275–291 (2014)
218 C. A. Fernandes et al.

24. Y.I. Karlovich, I. Loreto Hernández, Algebras of convolution type operators with piecewise
slowly oscillating data. I: Local and structural study. Integr. Equ. Oper. Theory 74, 377–415
(2012)
25. Y.I. Karlovich, I. Loreto Hernández, On convolution type operators with piecewise slowly
oscillating data. Oper. Theory Adv. Appl. 228, 185–207 (2013)
26. V. Rabinovich, S. Roch, B. Silbermann, Limit Operators and Their Applications in Operator
Theory (Birkhäuser, Basel, 2004)
27. M. Reed, B. Simon, Methods of Modern Mathematical Physics. I: Functional Analysis
(Academic Press, New York, 1980)
28. S. Roch, P.A. Santos, B. Silbermann, Non-Commutative Gelfand Theories. A Tool-kit for
Operator Theorists and Numerical Analysts (Springer, Berlin, 2011)
29. I.B. Simonenko, Local Method in the Theory of Shift Invariant Operators and Their Envelopes
(Rostov University Press, Rostov on Don, 2007, in Russian)
30. I.B. Simonenko, C.N. Min, Local Method in the Theory of One-Dimensional Singular Integral
Equations with Piecewise Continuous Coefficients. Noetherity (Rostov University Press,
Rostov on Don, 1986, in Russian)
Inner Outer Factorization of Wide
Rational Matrix Valued Functions on the
Half Plane

A. E. Frazho and A. C. M. Ran

Abstract The main purpose of this note is to use operator methods to solve
a rational inner-outer factorization problem for wide functions. It is believed
that this will provide valuable insight into the inner-outer factorization problem.
Our approach involves Wiener–Hopf operators, Hankel operators and invariant
subspaces for the backward shift. It should be emphasized that the formulas for
the inner and outer factor are derived in a computational manner.

Keywords Inner-outer factorization · Matrix valued function · Wiener–Hopf

operators · State space representation

Mathematics Subject Classification (2010) Primary 47B35, 47A68; Secondary

30J99

1 Introduction

In this note we will use operator techniques to develop a method to compute the
inner-outer factorization for certain matrix valued rational functions defined on the
closure of the right half plane. We shall focus on the “wide” case, i.e., the case where
the matrix function has more columns than rows (or an equal number, thereby also
including the square case). The “tall” case, where the matrix function has more

A. E. Frazho
Department of Aeronautics and Astronautics, Purdue University, West Lafayette, IN, USA
e-mail: [email protected]
A. C. M. Ran ()
Department of Mathematics, Faculty of Science, Vrije Universiteit Amsterdam, Amsterdam,
The Netherlands
Research Focus: Pure and Applied Analytics, North-West University, Potchefstroom,
South Africa
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 219

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_11
220 A. E. Frazho and A. C. M. Ran

rows than columns (or an equal number, again including the square case) is well
understood, and presented in, e.g., [3, 6, 18] and elsewhere.
It should be emphasized that we obtain explicit formulas for the inner and outer
factors in terms of the matrices appearing in a state space realization of the original
rational matrix valued function. Finally, we shall always assume that the rational
matrix function is proper, i.e., it has a finite value at infinity.
To set the stage, let E, U and Y be finite dimensional, complex vector spaces and
dim Y ≤ dim U. The Hilbert space of all Lebesgue measurable square integrable
functions over [0, ∞) with values in E is denoted by L2+ (E). Throughout H ∞ (U, E)
is the set of all analytic functions G in the open right hand plane {s : #(s) > 0}
such that

G∞ = sup{G(s) : #(s) > 0} < ∞.

Recall that a function Gi is inner if Gi is a function in H ∞ (E, Y) and Gi (iω)

is almost everywhere an isometry. (In particular, dim E ≤ dim Y.) Equivalently
(see, e.g., [6, 18]), Gi in H ∞ (E, Y) is an inner function if and only if the Wiener–
Hopf operator TGi mapping L2+ (E) into L2+ (Y) is an isometry. (The Wiener–Hopf
operator is defined in (2.1) below.) A function Go is outer if Go is a function in
H ∞ (U, E) and the range of the Wiener–Hopf operator TGo is dense in L2+ (E).
Let G be a function in H ∞ (U, Y). Then G admits a unique inner-outer
factorization of the form G(s) = Gi (s)Go (s) where Gi (s) is an inner function in
H ∞ (E, Y) and Go (s) is an outer function in H ∞ (U, E) for some intermediate space
E (see [14]). Because Gi (eiω ) is almost everywhere an isometry, dim E ≤ dim Y.
Since Go is outer, Go (eiω ) is almost everywhere onto E, and thus, dim E ≤ dim U.
By unique we mean that if G(s) = Fi (s)Fo (s) is another inner-outer factorization
of G where Fi is an inner function in H ∞ (L, Y) and Fo is an outer function in
H ∞ (U, L), then there exists a constant unitary operator V mapping E onto L such
that Gi = Fi V and V Go = Fo ; see [2, 6, 7, 18–20] for further details.
Throughout we assume that U, E and Y are all finite dimensional. We say that
Gi in H ∞ (E, Y) is a square inner function if E and Y have the same dimension. So
if Gi Go is an inner-outer factorization of G where Gi is square, then without loss
of generality we can assume that E = Y.
We say that the inner-outer factorization G = Gi Go is full rank if Gi is a square
inner function in H ∞ (Y, Y) and the range of TGo equals L2+ (Y). An inner-outer
factorization G = Gi Go is full rank if and only if Gi is a square inner function and
the range of TG is closed. Finally, if G in H ∞ (U, Y) admits a full rank inner-outer
factorization, then dim Y ≤ dim U.
Here we are interested in computing the inner-outer factorization for full rank
rational functions G in H ∞ (U, Y). So throughout we assume that dim Y ≤
dim U. Computing inner-outer factorizations when G does not admit a full rank
factorization is numerically sensitive. This is already apparent in the scalar case.
For example, take an outer function G(s) = p(s) d(s) with zeros on the imaginary axis.
A slight perturbation of the coefficients of p(s) could yield an inner part or even
Inner Outer Factorization on the Half Plane 221

an invertible outer function. To avoid this we simply assume that G(s) is full rank.
Moreover, from an operator perspective if G(s) is not full rank, then one has to
invert an unbounded Wiener–Hopf operator which is difficult. Furthermore, if G
does not admit a full rank inner-outer factorization, then a small H ∞ perturbation
of G does admit such a factorization. First we will present necessary and sufficient
conditions to determine when G admits a full rank inner-outer factorization. Then
we will give a state space algorithm to compute Gi and then Go .
Inner-outer factorization for the “wide” case has received attention before. In
several papers procedures to find the inner and outer factors have been derived.
For instance, in [4] the factors were derived by considering G∗ G + ε2 I , and
deducing the outer factor from this by solving a reduced Riccati equation. In [21]
the reduced Riccati equation is replaced by a different approach, avoiding in fact an
approximation procedure. In that paper, first the inner-outer factorization for G(s)∗
is computed, and then balanced coordinates are used to further factor the inner part
of G(s)∗ . Then they compute another inner-outer factorization of the remaining part.
In [16], see also [13] a different approach is chosen, [16] uses state space methods
along with some interesting state space decompositions to derive algorithms for
the inner and outer factorization in a general rational setting. Their methods are
algebraic in nature and quite different from our operator approach. Yet another
approach is taken in [17]. In all of these papers state space formulas are given and
their approach is based on finite dimensional techniques with properties of Riccati
equations. A common method is to compute the outer factor first and then solve for
the inner factor using the outer factor. Our approach is different: we derive a formula
for the inner factor from the fact that the subspace ker TG∗ is shift invariant, and thus,
by the Beurling–Lax–Halmos theorem (see[14]) there is an inner function Gi such
that ker TG∗ i = ker TG∗ . It is this inner function which we construct first, providing
a state space formula for Gi , and then deduce the outer factor from Go = G∗i G on
the imaginary axis.
Finally, it is emphasized that the previous methods to compute the inner-outer
factorization rely mainly on state space methods. Roughly speaking, the previous
methods involve solving Riccati equations, Hamilton methods, decomposing the
state space with balanced coordinates or other decompositions, or eigenvectors and
eigenvalues for certain state space operators; see for instance [4, 16, 17, 21]. Our
approach is quite different. We solve the inner-outer factorization problem by using
operator techniques for the general formulas and then use state space methods to
write down an explicit formula in the rational case. The operator methods give us
valuable insight into the underlying framework. Once the foundation is constructed,
then one simply uses state space techniques to formulate a realization of the inner
and outer part.
222 A. E. Frazho and A. C. M. Ran

2 Preliminaries

In this paper we shall make heavy use of the results and methods from [10]. Let
 be a proper rational matrix function with values in L(U, Y) and no poles on the
imaginary axis iR. Let ϕ be the Lebesgue integrable (continuous) matrix function
on the imaginary axis determined by  via the Fourier transform, that is,
 ∞
(iω) = (∞) + e−iωt ϕ(t) dt.
−∞

The Wiener–Hopf operator T associated with  and the Hankel operator H , both
mapping L2+ (U) into L2+ (Y), are defined as follows for t ≥ 0 and f in L2+ (U) :
 t
(T f )(t) = (∞)f (t) + ϕ(t − τ )f (τ )dτ (2.1)
0
 ∞
(H f )(t) = ϕ(t + τ )f (τ )dτ. (2.2)
0

In the sequel we shall freely use the basic theory of Wiener–Hopf and Hankel
operators which can be found in Chapters XII and XIII of [11]. Note that in [11]
the Fourier transform is taken with respect to the real line instead of the imaginary
axis as is done here.
Now let G be the stable rational function in H ∞ (U, Y) given by the following
state space realization:

G(s) = D + C(sI − A)−1 B. (2.3)

Here A is stable operator acting on a finite dimensional space X and B maps U

into X , while C maps X into Y and D maps U into Y. (By stable we mean that
all the eigenvalues of A are contained in the open left hand half-plane of C.) The
realization for G in (2.3) is denoted by {A, B, C, D}. It is noted that G(s) is also
given by the following Laplace transform:
 ∞
G(s) = D + e−st CeAt Bdt (Re s ≥ 0).
0

Hence the corresponding Wiener–Hopf operator TG and the Hankel operator HG can
be expressed in terms of the matrices appearing in the realization of G as follows
 t
(TG f )(t) = Df (t) + CeA(t −τ )Bf (τ )dτ (for t ≥ 0) (2.4)
0
 ∞
(HG f )(t) = CeA(t +τ )Bf (τ )dτ (for t ≥ 0). (2.5)
0
Inner Outer Factorization on the Half Plane 223

Throughout this note we assume that G is full row rank and DD ∗ is invertible.
A realization is called controllable if ∨∞ j =0 Im A B equals the state space X , and
j

a realization is called observable if ∩∞ j =0 ker CA = {0}. It is well known from

j

systems theory that the dimension of the state space X is as small as possible
(over all possible realizations) if and only if the realization is both observable and
controllable. Such a realization is called minimal, and it will be assumed throughout
that the realization (2.3) for G is minimal.
With G we also associate the rational matrix function R given by R(s) =
G(s)G(−s)∗ . Notice that R is a proper rational matrix function with values in
L(Y, Y) (the set of bounded linear operators mapping Y to itself) and has no poles
on the imaginary axis. By TR we denote the corresponding Wiener–Hopf operator
acting on L2+ (Y). It is well-known (see, e.g., formula (24) in Section XII.2 of [11])
that

TR = TG TG∗ + HG HG∗ . (2.6)

Let P be the controllability Grammian for the pair {A, B}. In other words,
 ∞ ∗
P = eAt BB ∗ eA t dt.
0

Moreover, P is the unique solution to the Lyapunov equation

AP + P A∗ + BB ∗ = 0. (2.7)

Because the pair {A, B} is controllable P is strictly positive. Let Wc be the

controllability operator mapping L2+ (U) into the state space X defined by
 ∞
Wc u = eAt Bu(t) dt (u ∈ L2+ (U)).
0

It is noted that the controllability Grammian is given by P = Wc Wc∗ .

Let Wo be the observability operator mapping X into L2+ (Y) corresponding to
the pair {C, A} defined by

Wo x = CeAt x (x ∈ X ).

Because the pair {C, A} is observable, Wo is one to one.

Let us introduce the operator  mapping Y into X defined by

 = BD ∗ + P C ∗ .

Now consider the algebraic Riccati equation:

A∗ Q + QA + (C −  ∗ Q)∗ (DD ∗ )−1 (C −  ∗ Q) = 0. (2.8)

224 A. E. Frazho and A. C. M. Ran

Let Ao on X and Co mapping X into Y be the operators defined by

Ao = A − Co and Co = (DD ∗ )−1 (C −  ∗ Q). (2.9)

We say that Q is a stabilizing solution to the algebraic Riccati equation (2.8) if Q

is a positive solution to (2.8) and the operator Ao is stable. Finally, the stabilizing
solution is unique.
Proposition 2.1 in [10] states that TR is a strictly positive operator on L2+ (Y) if
and only if there exists a stabilizing solution to the algebraic Riccati equation (2.8).
Moreover, in this case, the stabilizing solution Q is given by

Q = Wo∗ TR−1 Wo ;

see Proposition 2.2, equation (2.11) in [10]. Because {C, A} is observable, Wo is

one to one, and thus, Q is strictly positive. Formula (2.14) in [10] states that

TR−1 Wo x = Co eAo t x (x ∈ X ).

In other words, TR−1 Wo equals the observability operator Co eAo t mapping X into
L2+ (Y) determined by the pair {Co , Ao }. Since TR−1 Wo is one to one, the pair
{Co , Ao } is observable.
The Hankel operator HG can be written as HG = Wo Wc ; see (2.5). Using P =
Wc Wc∗ , we have HG HG∗ = Wo P Wo∗ . This, with Eq. (2.6) implies that

TG TG∗ = TR − Wo P Wo∗ .

Hence ker TG∗ = ker(TR − Wo P Wo∗ ). By consulting Lemma 4.1 in [9], restated as
Lemma 4.4 at the end of this paper for the readers convenience, we have

dim ker TG∗ = dim ker(TR − Wo P Wo∗ ) = dim ker(I − QP ).

Later we will see that the McMillan degree of the inner part Gi of G equals
dim ker(I − QP ).
The next lemma characterizes the existence of a full rank inner-outer factoriza-
tion. The proof of this lemma is essentially the same as that of Lemma 3.1 in [9],
and is therefore omitted.
Lemma 2.1 Let G be a rational function in H ∞ (U, Y) where U and Y are finite
dimensional spaces satisfying dim Y ≤ dim U. Then G admits a full rank inner-
outer factorization if and only if

G(iω)G(iω)∗ ≥ &I (for some & > 0),

or equivalently, the Toeplitz operator TR is strictly positive.

Inner Outer Factorization on the Half Plane 225

3 Inner Functions

First we consider realizations of rational inner functions. Recall that if Gi is a

rational function with values in L(Y, Y), then Gi is an inner function if and only if
Gi is a function H ∞ (Y, Y) and Gi (iω) is a unitary operator for all ω.
Consider a rational function Gi in H ∞ (Y, Y). Let {Ai on Xi , Bi , Ci , Di } be a
minimal realization for Gi . In particular,

Gi (s) = Di + Ci (sI − Ai )−1 Bi .

Because Gi is a rational function in H ∞ (Y, Y), and the realization is minimal, the
state space operator Ai is stable. Let Qi be the observability Grammian for the pair
{Ci , Ai }. In other words, Qi is the unique solution to the Lyapunov equation

A∗i Qi + Qi Ai + Ci∗ Ci = 0. (3.1)

Then Gi is an inner function in H ∞ (Y, Y) if and only if

Di is a unitary operator and Qi Bi = −Ci∗ Di .

(The state space formula for an inner function is classical; for example, see [1, 3, 21]
and also Theorem 19.15 in [5].) In this case, Q−1 i is the controllability Grammian
for the pair {Ai , Bi }, that is, Q−1
i is the unique solution to the Lyapunov equation

Ai Q−1 −1 ∗ ∗
i + Qi Ai + Bi Bi = 0. (3.2)

Let Wi mapping Xi into L2+ (Y) be the observability operator determined by

{Ci , Ai on Xi } defined by Wi x = Ci eAi t x for x in Xi .
Next, we introduce some notation and present some well-known results. If  is
an inner function in H ∞ (E, Y), then H() is the subspace of L2+ (Y) defined by

H() = L2+ (Y)  T L2+ (E) = ker T∗ .

Let Sτ∗ be the backward shift on L2+ (Y) for τ ≥ 0, that is, (Sτ∗ f )(t) = f (t + τ )
for f in L2+ (Y). It is noted that H() is an invariant subspace for the backward
shift Sτ∗ for all τ ≥ 0. According to the Beurling–Lax–Halmos Theorem if H is any
invariant subspace for the backward shift Sτ∗ for all τ ≥ 0, then there exists a unique
inner function  in H ∞ (E, Y) such that H = H(). By unique we mean that if
H = H(") where " is an inner function in H ∞ (L, Y), then there exists a constant
unitary operator V from E onto L such that  = "V ; see [2], Section 3.9 for the
half plane case we use here, see also [6, 12, 15, 18–20] for further details in the unit
disc case.
226 A. E. Frazho and A. C. M. Ran

Now we present the following classical result; see Theorem 7.1 in [8], Sections
4.2 and 4.3 in [6] and Section XXVIII.7 in [12] for the disc case, and [2], Lemma
3.45 and Lemma 3.46 for the half plane case.
Lemma 3.1 Let  be an inner function in H ∞ (Y, Y) where Y is finite dimensional.
Then the Hankel operator H is a contraction. Moreover,

H() = Im H = ker T∗ .

Furthermore, the following holds.

(i) The subspace H() is finite dimensional if and only if  is rational.
(ii) The dimension of H() equals the McMillan degree of .
(iii) Let {Ai on Xi , Bi , Ci , Di } be a minimal realization for a rational inner
function . Then H() equals the range of the observability operator Wi
mapping Xi into L2+ (Y).
Let {Ci , Ai on Xi } be a finite dimensional stable observable pair. Then we can
compute operators Bi mapping Y into Xi and Di on Y such that {Ai , Bi , Ci , Di } is
a minimal realization for an inner function . In this case, H() equals the range
of Wi , the observability operator determined by {Ci , Ai }. To accomplish this, first
compute the solution Qi to the observability Lyapunov equation in (3.1). Note that
Qi is automatically strictly positive because Ai is stable and {Ci , Ai } is observable.
Then compute Bi from the Lyapunov equation (3.2). Now use Qi Bi Bi∗ Qi = Ci∗ Ci
to find a unitary Di with Qi Bi = −Ci∗ Di . We shall call {Bi , Di } the complementary
operators for {Ci , Ai }. According to Lemma 3.1, the subspace H() = Im Wi =
ker T∗ . Finally, the complementary operators {Bi , Di } are not unique. However, the
corresponding subspace H() = Im Wi they determine is unique.

4 Main Result

Let us combine several results of [10] into one theorem, which is the right half plane
analogue of Theorem 3.2 in [9] in the open unit disc case.
Theorem 4.1 Let {A on X , B, C, D} be a minimal realization for a rational
function G in H ∞ (U, Y) where dim Y ≤ dim U. Let R be the function in L∞ (Y, Y)
defined by R(iω) = G(iω)G(iω)∗ . Let P the controllability Grammian for the pair
{A, B}; see (2.7). Then the following statements are equivalent.
(i) The function G admits a full rank inner-outer factorization;
(ii) The Wiener–Hopf operator TR is strictly positive.
(iii) There exists a (unique) stabilizing solution Q to the algebraic Riccati equation
(2.8).
In this case, the solution Q is given by Q = Wo∗ TR−1 Wo and Q is strictly positive.
Moreover, the following holds.
Inner Outer Factorization on the Half Plane 227

(iv) The eigenvalues of QP are real numbers contained in the interval [0, 1].
(v) If Gi is the inner factor of G, then the dimension of H(Gi ) is given by

dim H(Gi ) = dim ker TG∗ i = dim ker TG∗ = dim ker(I − QP ),

and the McMillan degree of Gi is given by

dim H(Gi ) = dim ker(I − QP ).

In particular, the McMillan degree of Gi is less than or equal to the McMillan

degree of G.
(vi) The operator TR−1 Wo mapping X into L2+ (Y) is equal to the observability
operator generated by {Co , Ao }, that is,

TR−1 Wo x = Co eAo t x (x ∈ X ). (4.1)

Finally, because {C, A} is observable, TR−1 Wo is one to one and {Co , Ao } is a

stable observable pair.
The construction of the state space realization for the inner Gi and outer factor
Go from the realization of G is analogous to the approach presented in [9] for
the disc case. The crucial observation is that if G = Gi Go is an inner-outer
factorization, then H(Gi ) = ker TG∗ i = ker TG∗ because TGo is onto. This observation
can be used to find a realization for Gi , and then Go is computed from Go = G∗i G
on the imaginary axis. Note that this is quite the other way around as in the
tall case, where we first find the outer factor Go from the spectral factorization
G∗ G = G∗o Go , and then compute Gi from Gi = GG−1 o .
To describe a state space realization for the inner Gi and outer Go factors of G,
we first introduce an isometry U as in [9]. To be precise, let

k = dim ker(I − QP ).

Let U be any isometry mapping Ck onto ker(I − QP ). Note, such a U can be

constructed explicitly from the singular value decomposition of I − QP . The
analogue of Theorem 3.4 in [9] for the half plane now reads as follows.
Theorem 4.2 Let {A, B, C, D} be a minimal realization for a rational function G.
Moreover, assume that the algebraic Riccati equation (2.8) has a stabilizing solution
Q, and let Ao and Co be as defined in (2.9). Let Ai on Ck and Ci mapping Ck into
Y be the operators defined by

Ai = U ∗ QAo P U and Ci = Co P U.

Let Bi and Di be complementary operators to {Ci , Ai }. Then

Gi (s) = Di + Ci (sI − Ai )−1 Bi (4.2)

228 A. E. Frazho and A. C. M. Ran

is the inner factor for G. The outer factor Go for G is given by

Go (s) = Di∗ D + (Di∗ C + Bi∗ U ∗ )(sI − A)−1 B.

Finally, G is an outer function if and only if k = 0. In this case, Gi (s) = I .

It is noted that the state space realization for the outer function Go given in the
Theorem 4.2 is not necessarily minimal.
The proof is analogous to the proof in the disc case. According to Lemma 4.1 in
[9], restated for convenience as Lemma 4.4 at the end of this section, we have

H(Gi ) = ker TG∗i = ker TG∗ = ker(TR − Wo P Wo∗ ).

This readily implies that

dim H(Gi ) = dim ker(I − QP ) = k.

The following lemma is a crucial part of the proof. We shall give two proofs, one
algebraic in nature, and a second operator theoretic proof using the fact that H(Gi )
is an invariant subspace for the backward shift.
Lemma 4.3 Assume that the hypotheses of Theorem 4.2 hold. Then there exists
a unique stable operator Ai on Ck such that Ao P U = P U Ai . Finally, Ai =
U ∗ QAo P U .
Algebraic Proof First we prove the existence of an operator Ai such that
QAo P U = QP U Ai . We claim that Im QAo P U ⊂ Im QP U = Im U. Since
Im U equals ker(I − QP ), we have U = QP U . Let us show that

(I − QP )QAo P U = 0. (4.3)

Formula (3.8) of [10] states that

A∗o (Q − QP Q) + (Q − QP Q)Ao + C1∗ C1 = 0

where C1 = D ∗ Co + B ∗ Q. (The formula for C1 is not needed in our proof.) Using

this we obtain

(I − QP )QAo P U = −A∗o (Q − QP Q)P U − C1∗ C1 P U.

By employing U = QP U , we have (I − QP )QAo P U = −C1∗ C1 P U. Multiplying

both sides by U ∗ P and using
U ∗ = U ∗ P Q, we obtain

0 = U ∗ P (I − QP )QAo P U = −U ∗ P C1∗ C1 P U.
Inner Outer Factorization on the Half Plane 229

In other words, C1 P U = 0. Hence (4.3) holds, and thus, the range of QA0 P U
is contained in the kernel of I − QP . So there exists an operator Ai on Ck such
that QA0 P U = U Ai . Multiplying by U ∗ on the left yields, U ∗ QA0 P U = Ai . In
particular, Ai is uniquely determined by our choice of the isometry U .
Equation (4.3) also implies that

QAo P U = QP QAo P U = QP U Ai .

Because Q is invertible, we have

Ao P U = P U Ai .

Notice that X1 = Im P U is an invariant subspace for Ao . Since P is invertible,

Ai is similar to the stable operator Ao |X1 on X1 . Because Ao is stable, Ai is also
stable.
An Operator Theoretic Proof of Lemma 4.3 By Lemma 4.4 below with (4.1), we
see that TR−1 Wo P U is the operator from Ck onto H(Gi ) given by

TR−1 Wo P U x = Co eA0 t P U x (x ∈ Ck )
 
H(Gi ) = Im TR−1 Wo P U . (4.4)

Since H(Gi ) is an invariant subspace of the backward shift on L2+ (Y) (see [2,
Chapter 3]), we have for all t > 0 and t1 > 0 and all x ∈ Ck

Co eAo (t +t1 ) P U x ∈ H(Gi ).

This readily implies that Co eAo (t +t1 ) P U x = Co eAo t P U F (t1 )x for some linear
operator F (t1 ) on Ck . One easily checks that F satisfies the semigroup property
F (t1 + t2 ) = F (t1 )F (t2 ). Hence there exists an operator Ai on Ck such that
F (t) = eAi t . In other words,

Co eAo t eAo t1 P U = Co eAo (t +t1 ) P U = Co eAo t P U eAi t1 .

Since the observability operator Co eAo t is one to one, we have eAo t P U = P U eAi t
(for all t). It is now clear that Ai must be stable. To see this simply note that P is
invertible, U is an isometry, and Ao is stable. Therefore eAi t converges to zero as
t tends to infinity, and thus, Ai is stable. By taking the derivative and then letting
t approach 0, we see that Ao P U = P U Ai as desired. Multiplying both sides by
U ∗ Q with U = QP U , we obtain Ai = U ∗ QAo P U .
Proof of Theorem 4.2 Note that using the definition of Ci = Co P U and Ai we
have

Co eAo t P U = Ci eAi t . (4.5)

230 A. E. Frazho and A. C. M. Ran

Since U is an isometry, P is invertible and {Co , Ao } is an observable pair it follows

that {Ci , Ai }, is an observable pair. Because TR−1 Wo P U is onto H(Gi ), the subspace
{Ci eAi t x | x ∈ Ck } equals H(Gi ); see Eqs. (4) and (4.5).
Now let Qi be the solution to

A∗i Qi + Qi Ai + Ci∗ Ci = 0,

and let Bi and Di be the complementary operators for the pair {Ci , Ai }. Finally,
{Ai , Bi , Ci , Di } is the state space realization for our inner function Gi (s); see (4.2).
By construction H(Gi ) = ker TG∗ i .
By a direct calculation or consulting formula (2.12) in [10], we also have

A∗o Q + QA + Co∗ C = 0. (4.6)

The formula for the outer factor Go (s) can now be derived in the same way as the
last section of the paper [9]. Indeed, for s ∈ iR, we have

Gi (s)∗ = Di∗ − Bi∗ (sI + A∗i )−1 Ci∗ .

This yields
  
Gi (s)∗ G(s) = Di∗ − Bi∗ (sI + A∗i )−1 Ci∗ D + C(sI − A)−1 B

=Di∗ D − Bi∗ (sI + A∗i )−1 Ci∗ D + Di∗ C(sI − A)−1 B

− Bi∗ (sI + A∗i )−1 Ci∗ C(sI − A)−1 B.

Using Ci = Co P U with (4.6), we obtain

Ci∗ C = U ∗ P Co∗ C = −U ∗ P (A∗o Q + QA) = −U ∗ P A∗o Q − U ∗ P QA.

By Lemma 4.3 we have U ∗ P A∗o = A∗i U ∗ P . This and U ∗ P Q = U ∗ yields

Ci∗ C = −A∗i U ∗ P Q − U ∗ P QA = −A∗i U ∗ − U ∗ A. (4.7)

Rewrite this as

Ci∗ C = −(sI + A∗i )U ∗ + U ∗ (sI − A).

By taking the appropriate inverses, we obtain

(sI + A∗i )−1 Ci∗ C(sI − A)−1 = −U ∗ (sI − A)−1 + (sI + A∗i )−1 U ∗ .
Inner Outer Factorization on the Half Plane 231

This readily implies that

  
Gi (s)∗ G(s) = Di∗ − Bi∗ (sI + A∗i )−1 Ci∗ D + C(sI − A)−1 B

=Di∗ D − Bi∗ (sI + A∗i )−1 Ci∗ D + Di∗ C(sI − A)−1 B

+ Bi∗ U ∗ (sI − A)−1 B − Bi∗ (sI + A∗i )−1 U ∗ B
=Di∗ D + (Di∗ C + Bi∗ U ∗ )(sI − A)−1 B
− Bi∗ (sI + A∗i )−1 (U ∗ B + Ci∗ D).

Now observe that

U ∗ B + Ci∗ D = U ∗ P QB + U ∗ P Co∗ D = U ∗ P (QB + Co∗ D) = U ∗ P C1∗ ,

where C1 = B ∗ Q + D ∗ Co ; see also formula (3.3) in [10]. Now use equation (3.8)
in [10], which states that

A∗o (Q − QP Q) + (Q − QP Q)Ao + C1∗ C1 = 0.

Multiplying by U ∗ P on the left and by P U on the right, with QP U = U , yields

U ∗ P C1∗ C1 P U = 0 and thus U ∗ P C1∗ = 0.

In other words, U ∗ B + Ci∗ D = 0. It follows that for s ∈ iR, we have

Go (s) = Gi (s)∗ G(s) = Di∗ D + (Di∗ C + Bi∗ U ∗ )(sI − A)−1 B.

Clearly, this holds for all s except the eigenvalues for A.

For the tall case we refer to Theorem 17.26 in [3].
Note that Ci can be computed a bit more explicitly:

Ci = Co P U = (DD ∗ )−1 (C −  ∗ Q)P U,

where  ∗ = CP + DB ∗ . This with U = QP U , yields

(C −  ∗ Q)P U = (C − CP Q − DB ∗ Q)P U = −DB ∗ U.

In other words, Ci = −(DD ∗ )−1 DB ∗ U.

Recall that Ci∗ C = −A∗i U ∗ − U ∗ A; see (4.7). Multiplying by U on the right, we
obtain Ci∗ CU = −A∗i − U ∗ AU. In other words, Ai = −U ∗ C ∗ Ci − U ∗ A∗ U. Using
Ci = −(DD ∗ )−1 DB ∗ U , we have

Ai = −U ∗ A∗ U + U ∗ C ∗ (DD ∗ )−1 DB ∗ U
∗
= −U ∗ A − BD ∗ (DD ∗ )−1 C U.
232 A. E. Frazho and A. C. M. Ran

This leads to an alterative formula for the inner factor

 ∗ −1
Gi (s) = Di − (DD ∗ )−1 DB ∗ U sI + U ∗ A − BD ∗ (DD ∗ )−1 C U Bi .

In turn, this may be rewritten as follows:

 ∗ −1
Gi (s) = Di − (DD ∗ )−1 DB ∗ sI + U U ∗ A − BD ∗ (DD ∗ )−1 C U Bi .

The latter formula compares well with the formula for the inner factor in the tall
case as presented in [3],Theorem 17.26.
Finally, Lemma 4.1 in [9] is restated in the following Lemma.
Lemma 4.4 Let T be a strictly positive operator on a Hilbert space H and P a
strictly positive operator on a Hilbert space X . Let W be an operator mapping X
into H and set Q = W ∗ T −1 W . Then the following two assertions hold.
(i) Let X and H be the subspaces defined by

X = ker(I − QP ) and H = ker T − W P W ∗ .

Then the operators

$1 = W ∗ |H : H → X and $2 = T −1 W P |X : X → H

are both well defined and invertible. Moreover, $−1 1 = $2 . In particular, X and
H have the same dimension.
(ii) The operator T − W P W ∗ is positive if and only if P −1 − Q is positive, or
1 1
equivalently, P 2 QP 2 is a contraction. In this case, the spectrum of QP is
contained in [0, 1]. In particular, if X is finite dimensional, then the eigenvalues
for QP are contained in [0, 1].

References

1. D. Alpay, I. Gohberg, Unitary rational matrix functions, in Topics in Interpolation Theory

of Rational Matrix-valued Functions. Operator Theory: Advances and Applications, vol. 33
(Birkhäuser Verlag, Basel, 1988), pp. 175–222
2. D.Z. Arov, H. Dym, J -Contractive Matrix Valued Functions and Related Topics (Cambridge
University Press, Cambridge, 2008)
3. H. Bart, I. Gohberg, M.A. Kaashoek, A.C.M. Ran, A State Space Approach to Canonical
Factorization: Convolution Equations and Mathematical Systems. Operator Theory: Advances
and Applications, vol. 200 (Birkhäuser Verlag, Basel, 2010)
4. T. Chen, B.A. Francis, Spectral and inner-outer-factorizations of rational matrices. SIAM J.
Matrix Anal. Appl. 10, 1–17 (1989)
5. H. Dym, Linear Algebra in Action. Graduate Studies in Mathematics, vol. 78 (American
Mathematical Society, Providence, 2007)
Inner Outer Factorization on the Half Plane 233

6. A.E. Frazho, W. Bosri, An Operator Perspective on Signals and Systems. Operator Theory:
Advances and Applications, vol. 204 (Birkhäuser Verlag, Basel, 2010)
7. C. Foias, A. Frazho, The Commutant Lifting Approach to Interpolation Problems. Operator
Theory: Advances and Applications, vol. 44 (Birkhäuser Verlag, Basel, 1990)
8. C. Foias, A. Frazho, I. Gohberg, M.A. Kaashoek, Metric Constrained Interpolation, Commu-
tant Lifting and Systems. Operator Theory: Advances and Applications, vol. 100 (Birkhäuser
Verlag, Basel, 1998)
9. A.E. Frazho, A.C.M. Ran, A note on inner-outer factorization for wide matrix-valued functions,
in: Operator Theory, Analysis and the State Space Approach. Operator Theory: Advances and
Applications, vol. 271 (Birkhäuser Verlag, Basel, 2018), pp. 201–214
10. A.E. Frazho, M.A. Kaashoek, A.C.M. Ran, Rational matrix solutions of a Bezout type equation
on the half plane, in: Advances in Structured Operator Theory and Related Areas. Operator
Theory: Advances and Applications, vol. 237 (Birkhäuser Verlag, Basel, 2013), pp. 145–160
11. I. Gohberg, S. Goldberg, M.A. Kaashoek, Classes of Linear Operators, Volume I. Operator
Theory: Advances and Applications, vol. 49 (Birkhäuser Verlag, Basel, 1990)
12. I. Gohberg, S. Goldberg, M.A. Kaashoek, Classes of Linear Operators, Volume II. Operator
Theory: Advances and Applications, vol. 63 (Birkhäuser Verlag, Basel, 1993)
13. V. Ionescu, C. Oară, M. Weiss, Generalized Riccati Theory and Robust Control, A Popov
Function Approach (Wiley, Chichester, 1999)
14. P.D. Lax, Translation invariant spaces. Acta Math. 101, 163–178 (1959)
15. N.K. Nikol’skii, Treatise on the Shift Operator. Grundlehren, vol. 273 (Springer, Berlin, 1986)
16. C. Oară, A. Varga, Computation of general inner–outer and spectral factorizations. IEEE Trans.
Autom. Control 45, 2307–2325 (2000)
17. T. Reis, M. Voigt, Inner-outer factorization for differential-algebraic systems. Math. Control
Signals Syst. 30, Art. 15, 19pp. (2018)
18. M. Rosenblum, J. Rovnyak, Hardy Classes and Operator Theory (Oxford University Press,
Oxford, 1985)
19. B. Sz.-Nagy, C. Foias, Harmonic Analysis of Operators on Hilbert Space (North-Holland,
Amsterdam, 1970)
20. B. Sz.-Nagy, C. Foias, H. Bercovici, L. Kérchy, Harmonic Analysis of Operators on Hilbert
Space (Springer, New York, 2010)
21. F.-B. Yeh, L.-F. Wei, Inner-outer factorizations of right-invertible real-rational matrices. Syst.
Control Lett. 14, 31–36 (1990)
Convergence Rates for Solutions
of Inhomogeneous Ill-posed Problems
in Banach Space with Sufficiently Smooth
Data

Matthew A. Fury

Abstract We consider the inhomogeneous, ill-posed Cauchy problem

u (t) = Au + h(t), 0 < t < T , u(0) = ϕ

where −A is the infinitesimal generator of a holomorphic semigroup of angle

θ in Banach space. As in conventional regularization methods, certain auxiliary
well-posed problems and their associated C0 semigroups are applied in order to
approximate a known solution u. A key property however, that the semigroups
adhere to requisite growth orders, may fail depending on the value of the angle
θ . Our results show that an approximation of u may be still be established in such
situations as long as the data of the original problem is sufficiently smooth, i.e.
in a small enough domain. Our results include well-known examples applied in
the approach of quasi-reversibility as well as other types of approximations. The
outcomes of the paper may be applied to partial differential equations in Lp spaces,
1 < p < ∞ defined by strongly elliptic differential operators.

Keywords Ill-posed problem · Regularizing family of operators · Holomorphic

semigroup · Continuous dependence of solutions

Mathematics Subject Classification (2010) Primary 47A52; Secondary 47D06

M. A. Fury ()
Penn State Abington, Department of Mathematics, Abington, PA, USA
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 235

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_12
236 M. A. Fury

1 Introduction

In this paper, we consider the generally ill-posed, linear Cauchy problem

u (t) = Au + h(t), 0 < t < T, (1.1)

u(0) = ϕ

where A is a closed, linear operator in a Banach space X, ϕ ∈ X, and h is a function

from [0, T ] into X. Ill-posed problems, whose solutions either do not exist for ϕ
in a dense subspace of X, are not unique, or do not depend continuously upon ϕ,
appear abundantly in several fields, most notably as backward parabolic problems
 2
(cf. [23, 25, 28, 34]). For instance, let A = −! defined by !u = ni=1 ∂ u2 for
∂xi
all u ∈ X = Lp (Rn ), 1 < p < ∞, whose generalized derivatives up to order 2
are also in X. Then (1.1) becomes the classic ill-posed problem that is the backward
heat equation. Related to ill-posed problems, inverse problems are also of significant
interest with applications such as parameter identification in medical imaging and
source identification in groundwater pollution (cf. [19, 22, 30, 33, 34, 36, 39]).
Due to the instability of ill-posed problems, it is common practice to estimate
a supposed solution u(t) of (1.1) via some approximation method. For instance,
several regularization techniques have been applied to (1.1) [2, 17, 18, 24, 25, 28, 35]
which utilize the generator of a C0 semigroup of linear operators, that is a family of
bounded linear operators {E(t)}t ≥0 on X satisfying

E(0) = I, E(s + t) = E(s)E(t) for all s, t ≥ 0, and

E(t)x → x as t → 0+ for all x ∈ X;

formally, we regard E(t) as E(t) = et B where B is its generator (cf. [29,

Section 1.2]). While the literature contains many results in Hilbert space, recently
including concrete numerical experiments which validate the theory [21, 37, 38],
fewer results are available in pure Banach space. Generalizing the example A =
−!, authors have considered the case in Banach space where −A generates
a holomorphic semigroup e−t A which extends analytically into a sector of the
complex plane. The most well-known and cited approximations here are fβ (A) =
A − βA2 [20] and fβ (A) = A(I + βA)−1 [35], though each of these regularizations
depends on the sector angle θ of the semigroup e−t A (cf. [2, 5, 6, 10, 17, 18, 23, 35]).
For fβ (A) = A − βA2 , the restriction θ ∈ ( π4 , π2 ] is required in order for
fβ (A) to generate a semigroup. In this case, under appropriate conditions on h,
the approximate Cauchy problem

v (t) = fβ (A)v + h(t), 0 < t < T, (1.2)

v(0) = ϕ
Solutions of Ill-posed Problems with Sufficiently Smooth Data 237

=t
is well-posed with unique solution vβ (t) = etfβ (A) ϕ + 0 e(t −s)fβ (A) h(s)ds. The
sum −A + fβ (A) = −βA2 generates a C0 semigroup as well, and in fact its growth
order does not depend on β. Consequently, regularization is established as vβ (t)
then converges to u(t) for each t ∈ [0, T ] as β → 0 (cf. [24, Section 3.1.1]). As for
fβ (A) = A(I + βA)−1 , regardless of the angle θ , this approximation yields well-
posedness of (1.2) being a bounded operator. The auxiliary operator −A + fβ (A)
generates a C0 semigroup as well, but as shown in the literature, its growth order is
independent of β only when θ ∈ ( π4 , π2 ] (cf. [2, 10, 18, 35]).
Recently, authors have investigated possibilities where the restriction on the
angle θ of the semigroup may be relaxed. Huang and Zheng [17] show that the
approximation fβ (A) = A − βA2 may be modified by use of the fractional
power of A [3]. Here, even if 0 < θ < π4 , then both fβ (A) = A − βAσ and
−A + fβ (A) = −βAσ still generate C0 semigroups for a suitable 1 < σ < 2. Also,
a logarithmic approximation

1
fβ (A) = − ln(β + e−pT A ), 0 < β < 1, p≥1 (1.3)
pT

first introduced by Boussetila and Rebbani [4] in Hilbert space, and later modified
by Tuan and Trong [38], may be applied in Banach space (cf. [16], [6], [12]). The
approximation (1.3) has lately received significant attention since it induces an error
that is less severe than those of A − βA2 and A(I + βA)−1 , both of which satisfy

(−A + fβ (A))x ≤ βA2 x for x ∈ Dom(A2 ), (1.4)

and etfβ (A)  ≤ et C/β . Recently, the author [11] investigated two additional
approximations fβ (A) = Ae−βA and fβ (A) = (ln 2)−1 A Log(1 + e−βA ) which
satisfy these same properties, but in such a way that the calculations do not rely on
the value of the angle θ .
In this paper, we generalize the results of [11] in order to unify all four of the
approximations

A − βAσ , A(I + βA)−1 , Ae−βA , (ln 2)−1 A Log(1 + e−βA ),

and prove convergence estimates between u(t) and vβ (t) without needing to restrict
the angle θ (Proposition 2.6 and Theorem 2.7 below). For this, our method relies on
an assumed smoothness of the data ϕ and h in (1.1). For example, if ϕ ∈ Dom(eQA )
for large enough Q > T then et A ϕ behaves like e−(Q−t )A(eQA ϕ) where e−(Q−t )A
is bounded for each t. We note that such a requirement is not out of the ordinary
since, for example, a solution u(t) of (1.1) with h ≡ 0 exists if and only if u(t) ∈
Dom(et A ) for each t (cf. [17, Introduction]). Also, many of the results associated
with the logarithmic approximation (1.3) demonstrate a stricter property than (1.4),
that is (−A + fβ (A))x ≤ βeτ A x for all x ∈ Dom(eτ A ) where τ may be larger
than T (cf. [37, Definition 1], [13, Lemma 1], [12, Proposition 6]). It is notable
238 M. A. Fury

that our intention, to regularize problem (1.1) when favorable C0 semigroups are
unavailable, likens to deLaubenfels’s motivation of C-semigroups, for example
{Cet A }t ≥0 where C is a bounded, injective operator such that Cet A = et A C is a
bounded linear operator for every t ≥ 0 [7, 8].
Below, B(X) denotes the space of all bounded linear operators on X. For a linear
operator A in X, ρ(A) denotes the resolvent set of A and σ (A), the spectrum of A, is
the complement of ρ(A) in C. Also, a strong solution of (1.1) is a function u which is
differentiable almost everywhere on [0, T ] such that u ∈ L1 ((0, T ) : X), u(0) = ϕ,
and u (t) = Au(t) + h(t) almost everywhere on [0, T ] (cf. [29, Definition 4.2.8]).

2 A Unifying Condition for Well-Posed Approximation

Assume −A is the infinitesimal generator of a bounded holomorphic semigroup

e−t A of angle θ ∈ (0, π2 ] on a Banach space X with 0 ∈ ρ(A). By definition, then
e−t A extends to an analytic function e−wA defined in the open sector Sθ = {w = 0 ∈
C : | arg w| < θ } which is uniformly bounded in every sector Sθ1 with θ1 ∈ (0, θ ).
Also, we have the following equivalence in terms of resolvent operators.
Theorem 2.1 ([31, Theorem X.52]) For a closed operator A on a Banach space
X, −A is the infinitesimal generator of a bounded holomorphic semigroup of angle
θ if and only if for each θ1 ∈ (0, θ ), there exists a constant M1 > 0 such that

/ S π2 −θ1 7⇒ w ∈ ρ(A),
w∈

and
M1
(w − A)−1  ≤ for all w ∈
/ S π2 −θ1 . (2.1)
dist(w, S̄ π2 −θ1 )

We note that by a simple geometric argument, the condition (2.1) is equivalent to

M1
(w − A)−1  ≤ for all w ∈
/ S̄ π2 −θ1 (2.2)
|w|

for a possibly different constant M1 depending on θ1 .

Under our assumptions, (1.1) is generally ill-posed with instability of solutions
as described in Sect. 1. To ensure that (1.1) has a well-posed approximation, we call
upon the following result.
Theorem 2.2 ([29, Corollary 4.2.10]) For each 0 < β < 1, let fβ (A) be the
infinitesimal generator of a C0 semigroup of bounded linear operators on X. If h is
differentiable almost everywhere on [0, T ] and h ∈ L1 ((0, T ) : X), then (1.2) has
Solutions of Ill-posed Problems with Sufficiently Smooth Data 239

a unique strong solution given by

 t
vβ (t) = etfβ (A) ϕ + e(t −s)fβ (A) h(s)ds
0

for every ϕ ∈ Dom(fβ (A)).

Unless the representation fβ (A) is elementary, the operator fβ (A) will be defined
by the Dunford integral for functions fβ (w) bounded and holomorphic in a sector
of the complex plane, particularly for us in most cases

1
fβ (A) = fβ (w)(w − A)−1 dw
2πi α

where α = ∂Sα , π2 − θ < α < π2 , oriented so that Im(w) decreases as w travels

along α (cf. [9, Chapter VII] and also [15, Chapter 2]).
The relationship between A and fβ (A) will need to be scrutinized in order
to estimate known solutions of problem (1.1). Indeed, we aim to incorporate a
condition similar to (1.4) where fβ (A)x → Ax as β → 0 for x in an appropriate
domain. Ultimately, this condition establishes continuous dependence on modeling
since our final goal is to show a similar convergence for the corresponding solutions
vβ (t) and u(t). To this end, we endorse the approximation condition, Condition (A)
of Ames and Hughes [2, Definition 1].
Definition 2.3 Let −A be the infinitesimal generator of a bounded holomorphic
semigroup of angle θ ∈ (0, π2 ] on a Banach space X with 0 ∈ ρ(A), and suppose
that for each 0 < β < 1, fβ (A) is the infinitesimal generator of a C0 semigroup.
Then fβ (A) satisfies Condition (A) if there exists a positive constant p independent
of β such that Dom(fβ (A)) ⊇ Dom(A1+p ) and

(−A + fβ (A))x ≤ βA1+p x (2.3)

for all x ∈ Dom(A1+p ).

Note that in the definition of Condition (A), p need not be an integer, in which
case A1+p is defined by the fractional power of A [3] (see also [29, Section 2.6]).
Nevertheless, we still have Dom(A1+p ) ⊆ Dom(A) so that (2.3) is valid.
As we will see for our main results below (Proposition 2.6 and Theorem 2.7),
not only do we require that fβ (A) satisfy the stipulations of Condition (A), but also
the operator −A + fβ (A), with domain Dom(A) ∩ Dom(fβ (A)), must have stable
enough properties in a certain sense. Therefore, we assume a stronger condition.
Definition 2.4 Let −A be the infinitesimal generator of a bounded holomorphic
semigroup of angle θ ∈ (0, π2 ] on a Banach space X with 0 ∈ ρ(A), and suppose
240 M. A. Fury

that for each 0 < β < 1, fβ (A) is the infinitesimal generator of a C0 semigroup.
Then fβ (A) satisfies Condition (A+ ) if the following are satisfied:
(i) Dom(fβ (A)) ⊇ Dom(A2 ) and there exists a constant R independent of β such
that

(−A + fβ (A))x ≤ βRA2 x for all x ∈ Dom(A2 ),

(ii) There exist π

2 −θ <ν < π
2 and C ≥ 0 independent of β such that

Re(−w + fβ (w)) ≤ C|w| for all w in the open sector

Sν = {w = 0 ∈ C : | arg w| < ν}.

Remark 2.5 Conventionally, a more desirable condition is one similar to Condition

(A+ ) (ii) but the constant C being negative. In that case, the sum −A + fβ (A)
generates a C0 semigroup whose growth order Meωt does not depend on β, a
property which is required for standard regularization arguments [17, 18, 23]. In
fact, this is possible even if C < 0 depends on β, for instance with the example
−A + fβ (A) = −βAσ (cf. [17, Theorem 3.1] and also [10, Proposition 3.4]). In
any case, in our general framework, we may not have this desired property, for
example (as noted in Sect. 1) if fβ (A) = A(I + βA)−1 and 0 < θ < π4 .
Despite Remark 2.5, we are able to prove regularization-type calculations via
Theorem 2.7 below if we assume strong enough smoothness properties on ϕ and h.
Proposition 2.6 demonstrates an initial estimate that may be established by choosing
data in a small enough domain. As each operator e−t A , t ≥ 0 is a bounded,
injective operator (cf. [7, Lemma 3.1]), we find an appropriate Q with ϕ and Ran(h)
contained in Dom(eQA ) = Dom((e−QA )−1 ).
Proposition 2.6 Let −A be the infinitesimal generator of a bounded holomorphic
semigroup of angle θ ∈ (0, π2 ] on a Banach space X with 0 ∈ ρ(A) and assume
fβ (A) satisfies Condition (A+ ). For every 0 < & < 1 and π2 − θ < α < ν, there
exists a constant Q = Q(&, α) > T such that if x ∈ Dom(eQA ), then for each
integer n ≥ 0,

An (et A x − etfβ (A) x) ≤ βK(T − t + &)−(n+1) eQA x for 0 ≤ t ≤ T

where K is a constant depending on n and α but independent of β, &, and t.

Proof Let 0 < & < 1 and π2 − θ < α < ν where ν and C are the constants
from Condition (A+ ) (ii). Further, let Q be any constant larger than T satisfying the
inequality (Q − T ) cos α ≥ C(T + &). If x ∈ Dom(eQA ), then by Condition (A+ )
(i) and standard semigroup properties (cf. [29, Theorem 1.2.4]),

An (et A x − etfβ (A) x) = (I − etfβ (A) e−t A )An et A x
= (I − etfβ (A) e−t A )An (et A e−QA (eQA x))
Solutions of Ill-posed Problems with Sufficiently Smooth Data 241

;  t  ;
; ∂ τfβ (A) −τ A ;
;
= ;− e e A e e (e x) dτ ;
n t A −QA QA
∂τ ;
0
;  t ;
; ;
;
= ;− (−A + fβ (A))e e A e e (e x) dτ ;
τfβ (A) −τ A n t A −QA QA
;
0
 t
≤ βR A2+n eτfβ (A) e−τ A et A e−QA (eQA x) dτ. (2.4)
0

By the estimate (Q − T ) cos α ≥ C(T + &), then for w = re±iα ∈ α = ∂Sα ,

A2+n eτfβ (A) e−τ A et A e−QA (eQA x)

;  ;
; 1 ;
=;; 2πi w 2+n τ (−w+fβ (w)) −(Q−t )w
e e (w − A) −1 QA
(e x) dw ;
;
α

1
≤ |w|2+n eτ Re(−w+fβ (w)) e−(Q−t )Re(w)(w − A)−1  eQA x |dw|
2π α

1 ∞ 1+n τ Cr −(Q−t )r cos α
≤ r e e Mα eQA x dr
π 0

Mα ∞ 1+n τ Cr −(Q−T )r cos α QA
≤ r e e e x dr
π 0

Mα ∞ 1+n C(τ −(T +&))r QA
≤ r e e x dr
π 0
Mα
= (n + 1)!C −(2+n) (T − τ + &)−(2+n) eQA x (2.5)
π
where Mα is a constant due to (2.2). Returning to (2.4),

An (et A x − etfβ (A) x)

 t
Mα
≤ βR (n + 1)!C −(2+n) (T − τ + &)−(2+n) eQA x dτ
π 0

Mα (T + &)n+1 − (T − t + &)n+1 QA
= βR (n + 1)!C −(2+n) e x
π (T − t + &)n+1 (T + &)n+1
Mα
≤ βR (n + 1)!C −(2+n) (T − t + &)−(n+1) eQA x.
π

Now, we prove our main approximation theorem concerning strong solutions of

(1.1) and (1.2).
Theorem 2.7 Let −A be the infinitesimal generator of a bounded holomorphic
semigroup of angle θ ∈ (0, π2 ] on a Banach space X with 0 ∈ ρ(A) and assume
fβ (A) satisfies Condition (A+ ). Fix 0 < & < 1 and π2 − θ < α < ν, and let
242 M. A. Fury

Q = Q(&, α) > T be as in Proposition 2.6. Assume that u(t) and vβ (t) are
strong solutions of (1.1) and (1.2) respectively where h satisfies the conditions
of Theorem 2.2, and as well that initial data ϕ and Ran(h) are contained in
Dom(eQA ) with eQA h ∈ L1 ((0, T ) : X). Then there exist constants C̃ and M
each independent of β and & such that for 0 ≤ t < T ,

u(t) − vβ (t) ≤ C̃β 1−ω(t )M ω(t )& −2 N (2.6)

where
 T
N = eQA ϕ + eQA h(t)dt, (2.7)
0

and ω(ζ ) is a harmonic function which is bounded and continuous on the bent strip

$ = {t + re±iγ | 0 ≤ t ≤ T , r ≥ 0}, γ ∈ (0, θ ),

satisfying 0 = ω(0) ≤ ω(t) < ω(T ) = 1 for all 0 ≤ t < T .

Proof Fix γ ∈ (0, θ ). By the definition of the semigroup e−t A , we know that
e−wA is uniformly bounded in Sγ (cf. [31, Theorem X.52]). Following [2, 5, 26],
we extend the solutions u(t) and vβ (t) into the bent strip $ and apply Carleman’s
inequality (cf. [14, p. 346]). For ζ = t + re±iγ ∈ $, define
±iγ )A
w(ζ ) = e−(re (u(t) − vβ (t)).

Also following [1], define

∞
T  
1 ¯ 1 1
F (ζ ) = w(ζ ) + ∂w(z) + dξ dη (2.8)
π z−ζ z̄ + 1 + ζ
0 0

where z = ξ +ηe±iγ and ∂¯ denotes the Cauchy-Riemann operator (cf. [32]). Denote
 
±iγ
M = max e−(re )A  .
r≥0

By Proposition 2.6 and (2.7),

w(ζ ) ≤ M u(t) − vβ (t)

  t 
(t −s)A (t −s)fβ (A)
≤ M e ϕ − e
tA tfβ (A)
ϕ + (e −e )h(s)ds
0

≤ M βK(T − t + &)−1 eQA ϕ
Solutions of Ill-posed Problems with Sufficiently Smooth Data 243

 t 
−1
+ βK(T − (t − s) + &) e QA
h(s)ds
0
  T 
−1
≤ M βK& e ϕ +
QA
e QA
h(y)dy
0

= M βK& −1 N. (2.9)

Next, by the definition of w(ζ ), a straightforward calculation shows that

 
¯ 1 ±iγ ∂ ∂
∂w(ζ )= e w(ζ ) − w(ζ )
2i sin(±γ ) ∂t ∂η
e±iγ ±iγ 
= e−(re )A (Au(t) − fβ (A)vβ (t)) + (Au(t) − Avβ (t)) .
2i sin(±γ )

Again using Proposition 2.6 and (2.7),

Au(t) − Avβ (t)

;   t 
;
=; ;A e tA
ϕ + e (t −s)A
h(s)ds
0
  t ;
;
+ A etfβ (A) ϕ + e(t −s)fβ (A) h(s)ds ;
;
0
 t
≤ A(e tA
−e tfβ (A)
)ϕ + A(e(t −s)A − e(t −s)fβ (A) )h(s)ds
0
 t
≤ βK(T − t + &)−2 eQA ϕ + βK(T − (t − s) + &)−2 eQA h(s)ds
0

≤ βK& −2 N.

Also, by Condition (A+ ) (i), (2.5), and (2.7),

Avβ (t) − fβ (A)vβ (t) = (−A + fβ (A))vβ (t)

≤ βRA2 vβ (t)
;   t ;
; 2 tf (A) ;
;
= βR ;A e β ϕ+ e (t −s)fβ (A)
h(s)ds ;
;
0

  t 
≤ βR A2 etfβ (A) ϕ + A2 e(t −s)fβ (A) h(s)ds
0

= βR A2 etfβ (A) e−t A e−(Q−t )A(eQA ϕ)
244 M. A. Fury

 t 
(t −s)fβ (A) −(t −s)A −(Q−(t −s))A
+ e e e (e QA
h(s))ds
0

≤ βRL& −2 N

where L is a constant independent of β, &, and t. Altogether, we have shown

¯
∂w(ζ ) ≤ βC & −2 N, (2.10)

where C is a constant independent of β, &, and ζ .

If x ∗ is a member of the dual space of X with x ∗  ≤ 1, then it follows that
∗
x (F (ζ )) as defined by (2.8) is a bounded, continuous function on $ which is also
analytic on the interior of $ (cf. [1, 5]). Furthermore, there exists a constant L
independent of ζ such that
 
∗ ∗ ¯
|x F (ζ )| ≤ x  w(ζ ) + L max ∂(w(ζ )) .
ζ ∈$

By (2.9) and (2.10), together with the fact that 0 < & < 1, then

|x ∗ F (ζ )| ≤ βM& −2 Nx ∗  (2.11)

where M is a constant independent of β, &, and ζ . Hence Carleman’s Inequality

implies

|x ∗ F (t)| ≤ M(0)1−ω(t )M(T )ω(t ) , (2.12)

for 0 ≤ t ≤ T , where

M(t) = max |x ∗ F (t + re±iγ )|

r≥0

and ω is a harmonic function which is bounded and continuous on $, satisfying

0 = ω(0) ≤ ω(t) < ω(T ) = 1 for all 0 ≤ t < T (cf. [26], [14, p. 346]). Note by
(2.10),
 
¯ ±iγ )
M(0) ≤ x ∗  w(re±iγ ) + L max ∂(re
r≥0
 
±iγ
¯ ±iγ )
= x ∗  e−(re )A (ϕ − ϕ) + L max ∂(re
r≥0

¯
= L max ∂(re ±iγ ∗
)x 
r≥0

≤ L βC & −2 Nx ∗ .
Solutions of Ill-posed Problems with Sufficiently Smooth Data 245

Also, from (2.11) and the fact that 0 < β < 1,

M(T ) ≤ M& −2 Nx ∗ .

Returning to (2.12), then

|x ∗ F (t)| ≤ (L βC )1−ω(t )M ω(t ) & −2 Nx ∗ 

and taking the supremum over x ∗ ∈ X∗ with x ∗  ≤ 1, we obtain

F (t) ≤ C̃β 1−ω(t )M ω(t ) & −2 N

for 0 ≤ t ≤ T where C̃ and M are constants each independent of β and &. Then for
0 ≤ t ≤ T,

u(t) − vβ (t) = w(t)

¯
≤ F (t) + L max ∂w(t)
0≤t ≤T

≤ C̃β 1−ω(t )M ω(t ) & −2 N + L βC & −2 N

= (C̃ + L β ω(t )C M −ω(t ) )β 1−ω(t )M ω(t )& −2 N
≤ C̃β 1−ω(t )M ω(t ) & −2 N

for a possibly different constant C̃ independent of β and &.

3 Examples Illustrating Condition (A+ )

Here we outline several examples which satisfy Condition (A+ ), so that Theo-
rem 2.7 may be applied. As mentioned in Sect. 1, each of these approximations may
be applied to backward problems defined in Lp (Rn ), 1 < p < ∞. For specifics as
well as further applications, see Remark 4.2 below.
Example (1) fβ (A) = A(I + βA)−1 . Property (2.2) implies (I + βA)−1 ∈ B(X)
and
R
(I + βA)−1  = β −1 (−β −1 − A)−1  ≤ β −1 =R (3.1)
β −1

for some constant R. Then

fβ (A) = β −1 ((I + βA) − I )(I + βA)−1 = β −1 (I − (I + βA)−1 )

246 M. A. Fury

is also in B(X) being a linear combination of two bounded operators, and by

(3.1), fβ (A) ≤ (1 + R)β −1 . Therefore, fβ (A) generates a uniformly continuous
semigroup satisfying
−1
etfβ (A)  ≤ et (1+R)β

for each t ≥ 0.
Note Condition (A+ ) (i) is satisfied as

−A + A(I + βA)−1 = (−A(I + βA) + A)(I + βA)−1

= −βA2(I + βA)−1 = −β(I + βA)−1 A2

on Dom(A2 ) and by (3.1),

 − β(I + βA)−1 A2 x ≤ βRA2 x

for all x ∈ Dom(A2 ).

Next, let π2 − θ < ν < π
2. Consider for w = re±iα ∈ Sν ,

w
Re(−w + fβ (w)) = Re(−w + )
1 + βw
βw2
= −Re( )
1 + βw
 2 
βr cos(2α) ± iβr 2 sin(2α)
= −Re
(1 + βr cos α) ± iβr sin α
βr 2 cos(2α) + β 2 r 3 (cos(2α) cos α + sin(2α) sin(α))
=−
(1 + βr cos α)2 + (βr sin α)2
βr 2 cos(2α) + β 2 r 3 cos α
=− := g(r). (3.2)
1 + 2βr cos α + β 2 r 2

We claim that r − g(r) ≥ 0 for all r ≥ 0. Indeed,

r + 2βr 2 cos α + β 2 r 3 + βr 2 cos(2α) + β 2 r 3 cos α

r − g(r) =
1 + 2βr cos α + β 2 r 2
r + (2 cos α + cos(2α))βr 2 + (1 + cos α)β 2 r 3
= .
1 + 2βr cos α + β 2 r 2

Define
π
k(t) = 2 cos t + cos(2t), 0≤t ≤ .
2
Solutions of Ill-posed Problems with Sufficiently Smooth Data 247

Then k (t) = −2(sin t +sin(2t)) ≤ 0 since both sin t and sin(2t) are nonnegative on
[0, π2 ]. Consequently, k(t) is a decreasing function, and thus is never smaller than
k( π2 ) = −1. We have

r + (2 cos α + cos(2α))βr 2 + (1 + cos α)β 2 r 3

 
1 2 3
≥ r − βr + β r = r(1 − βr + β r ) = r
2 2 3
βr −2 2
+ ≥0
2 4

and so r − g(r) ≥ 0 as well. We have shown

|w| − Re(−w + fβ (w)) ≥ 0

and so Condition (A+ ) (ii) is satisfied with C = 1.

Remark 3.1 As pointed out in [18, Remark 3.3], if θ > π4 , so that π2 − θ < π4 ,
then ν can be chosen less than π4 and so g(r) in (3.2) is strictly negative. In this
case, −A + fβ (A) generates a C0 semigroup whose growth order is independent
of β [18, Theorem 2.1] and therefore, sharper estimates than (2.6) may be used
to establish regularization as in the literature cited in Sect. 1. Nevertheless, in our
framework, Condition (A+ ) (ii) is still satisfied, this time with C = 0.
Example (2) fβ (A) = A −βAσ where σ and ν satisfy π2 −θ < ν < σ ν < π2 . Note,
σ is necessarily larger than 1 but its upper bound is determined by the value of θ .
For instance, if π2 − θ < π4 , then ν can be chosen less than π4 and σ = 2 suffices. In
this case, Condition (A+ ) (i) is easily satisfied as

(−A + fβ (A))x =  − βA2 x = βA2 x

for all x ∈ Dom(A2 ). Condition (A+ ) (ii) is also satisfied with C = 0 since 2ν < π
2
implies

Re(−w + fβ (w)) = Re(−βw2 ) = −βr 2 cos(2α) < −βr 2 cos(2ν) < 0

for all w = re±iα ∈ Sν .

In the case that π2 − θ ≥ π4 , we must revise our calculations. We can choose
π
1 < σ < 2ν so that σ ν < π2 but in this case σ is not an integer since ν > π4 .
Therefore, fβ (A) = A − βAσ is defined by the fractional power of A, that is

−σ 1
A = w−σ (w − A)−1 dw, Aσ = (A−σ )−1
2πi 

where  is a contour similar to α = ∂Sα but avoids both the origin and the negative
real axis [29, Section 2.6]. In fact, since 0 ∈ ρ(A), there exists δ > 0 such that the
closed disk of radius δ, centered at the origin is contained in ρ(A). Hence, we may
248 M. A. Fury

revise α to be the contour

α = 1 ∪ 2 ∪ 3 ,
1 = {reiα : r ≥ δ},
2 = {δeiφ : −α ≤ φ ≤ α},
3 = {re−iα : r ≥ δ},

oriented so that Im(w) decreases as w travels along α (cf. [17]). But 2 is bounded
and does not affect the convergence of the contour integral, so the calculations in
Proposition 2.6 remain fundamentally unchanged. Henceforth, using the properties
(cf. [29, Theorem 2.6.8, Lemma 2.6.3])

Dom(A2 ) ⊆ Dom(Aσ ) ⊆ Dom(A) for 1 < σ < 2,

Aσ1 (Aσ2 )x = Aσ1 +σ2 x = Aσ2 (Aσ1 )x for x ∈ Dom(Amax{σ1 ,σ2 ,σ1 +σ2 } ),

A−σ  ≤ κ for 0 < σ < 1,

Condition (A+ ) (i) is satisfied by

(−A + fβ (A)x) = βAσ x = βA−(2−σ ) A2 x ≤ βκA2 x

for all x ∈ Dom(A2 ). Also similar to the case σ = 2, Condition (A+ ) (ii) is satisfied
with C = 0 since σ ν < π2 implies

Re(−w + fβ (w)) = Re(−βwσ ) = −βr σ cos(σ α) < −βr σ cos(σ ν) < 0.

Finally, we point out that in either case whether σ = 2 or 1 < σ < 2, while
fβ (A) and −A+fβ (A) are unbounded operators, it may be shown that both generate
C0 semigroups. Again, if σ is not an integer, careful revisions to α must be taken
near the origin. In either case, one may show the semigroup generated by fβ (A)
satisfies
−1
β (1−σ )
etfβ (A)  ≤ P et P

for some constants P , P independent of β (cf. [17, Theorem 3.2]). Furthermore,

the semigroup generated by −A+fβ (A) has growth order independent of β (see the
earlier Remark 2.5), and so as noted in Remark 3.1, one may prove regularization
the standard way.
Example (3) fβ (A) = Ae−βA . This approximation makes direct use of the
holomorphic semigroup e−t A generated by −A. By classic semigroup theory,
fβ (A) ∈ B(X) and satisfies fβ (A) ≤ M
β where M is a constant independent of
Solutions of Ill-posed Problems with Sufficiently Smooth Data 249

β (cf. [29, Theorem 2.5.2 (d)]). Therefore, fβ (A) generates a uniformly continuous
semigroup satisfying
−1
etfβ (A)  ≤ et Mβ

for each t ≥ 0. Next, since e−t A is uniformly bounded,

 β
(−A + Ae−βA )x = (I − e−βA )Ax ≤ e−t A A2 xdt ≤ βM A2 x
0

for all x ∈ Dom(A2 ). Therefore, Condition (A+ ) (i) is satisfied. Also, for w =
re±iα ∈ Sν ,

Re(−w + fβ (w)) = Re(−w + we−βw )

= −r(cos α − e−βr cos α (cos(α − βr sin α)))
≤ re−βr cos α cos(α − βr sin α) ≤ r

and so Condition (A+ ) (ii) is satisfied with C = 1 and any π

2 −θ <ν < π
2.

Example (4) fβ (A) = (ln 2)−1 A Log(1 + e−βA ). Here we define fβ (A) by the
Dunford integral

1
fβ (A) = w Log(1 + e−βw )(w − A)−1 dw
2 ln 2 πi α

where the principal branch of the logarithm is taken [9, 27]. The calculations
for Condition (A+ ) here are involved and so are omitted, but we can refer to
[11] for a complete analysis. There, it is found that fβ (A) is a bounded operator
with fβ (A) ≤ Cβα where Cα is a constant independent of β but dependent
on α. Further, it may be shown that fβ (A) satisfies Condition (A+ ) (i) with a
constant R independent of β and also Condition (A+ ) (ii) with C = 2 lnπ
2 and any
π
2 − θ < ν < π
2 [11, Lemma 2.4 and Proposition 3.1].

4 Comments on Regularization and Applications

Traditionally, regularization arguments are made where u(t) is any solution of (1.1)
with limited assumptions on the data (cf. [17, Definition 4.1]). As is the theme
throughout this paper, our calculations rely on sufficient smoothness of the data ϕ
and h as in (2.7). We outline a regularization-type argument under the assumptions
of Theorem 2.7.
250 M. A. Fury

First, note that in each of the four examples listed in Sect. 3, with the exception
of the case 1 < σ < 2 in Example (2), fβ (A) generates a C0 semigroup satisfying
−1
etfβ (A)  ≤ P et P β for some constants P , P independent of β. Therefore, if
δ
vβ (t) is the solution of the approximate well-posed problem with perturbed data

v (t) = fβ (A)v + h(t), 0 < t < T,

v(0) = ϕδ

satisfying ϕ − ϕδ  ≤ δ, then
;  t 
; tf (A)
vβ (t) − vβδ (t) =;
; e β ϕ + e (t −s)fβ (A)
h(s)ds
0
  t ;
;
− etfβ (A) ϕδ + e(t −s)fβ (A) h(s)ds ;
;
0

=etfβ (A) (ϕ − ϕδ )
β −1
≤P eT P δ.

Now, let u(t) be a strong solution of (1.1) that satisfies the hypotheses of Theo-
rem 2.7. Choosing β = −2T P (ln δ)−1 we have β → 0 as δ → 0, and

u(t) − vβδ (t) ≤ u(t) − vβ (t) + vβ (t) − vβδ (t)

β −1
≤ C̃β 1−ω(t )M ω(t ) & −2 N + P eT P δ
√
= C̃β 1−ω(t )M ω(t ) & −2 N + P δ
→ 0 as δ → 0

for all 0 ≤ t < T . Also, the case for t = T may be addressed separately. Indeed,
similar to the calculation (2.9), we obtain

u(T ) − vβδ (T ) ≤ u(T ) − vβ (T ) + vβ (T ) − vβδ (T )

β −1
≤ βK& −1 N + P eT P δ
√
= βK& −1 N + P δ
→ 0 as δ → 0.

In the case of Sect. 3 Example (2) where 1 < σ < 2, the only difference is that
etfβ (A) satisfies
−1
β −(σ −1)
etfβ (A)  ≤ P et P .
Solutions of Ill-posed Problems with Sufficiently Smooth Data 251

Nevertheless, a similar calculation may be made if instead we choose

β = (−2T P (ln δ)−1 )σ −1 .

Remark 4.1 While convergence depends continuously on β (and δ) in our calcu-

lations, the estimate (2.6) of Theorem 2.7 may depend severely upon & which is
related to the choice for Q in the estimate (Q − T ) cos α ≥ C(T + &) as in the proof
of Proposition 2.6. For instance, suppose C = 1 and cos α = 12 (which is possible in
Examples 1 and 3). If T ≥ 1 and & = 0.5, then we may choose Q = 4T so that the
conditions ϕ and Ran(h) in Dom(e4T A ) with e4T A h ∈ L1 ((0, T ) : X) are required.
If & = 0.1, then a smaller value for Q may be chosen which alleviates the restriction
on ϕ and h. However, in this case the factor & −2 in (2.6) obviously becomes much
larger. In a different case, if C = 0 (which is possible in Example 2), then regardless
of the value of α, we may choose Q arbitrarily close to T .
Remark 4.2 While our assumptions of the paper include that −A generate a
bounded holomorphic semigroup of angle θ with 0 ∈ ρ(A) for convenience, this
is not the most general case. However, it is known that our assumptions imply that
for θ ∈ (0, θ ), there exists λ ∈ R such that −A + λ is the infinitesimal generator
of a bounded holomorphic semigroup of angle θ with 0 ∈ ρ(A − λ) (cf. [31,
Theorem X.53], [7]). Therefore, as in [10, 11, 17], we may still apply the results to
the ill-posed backward heat equation in Lp (Rn ), 1 < p < ∞ (A = −!) and more
generally to PDE’s defined by strongly elliptic differential operators of even order
(cf. [29], [17, Section 5]).

Acknowledgments The author would like to thank the editors for the proceedings of the IWOTA
2019 and the referee for their helpful suggestions.

References

1. S. Agmon, L. Nirenberg, Properties of solutions of ordinary differential equations in Banach

space. Comm. Pure Appl. Math. 16, 121–151 (1963)
2. K.A. Ames, R.J. Hughes, Structural stability for ill-posed problems in Banach space. Semi-
group Forum 70, 127–145 (2005)
3. A.V. Balakrishnan, Fractional powers of closed operators and the semigroups generated by
them. Pacific. J. Math. 10, 419–437 (1960)
4. N. Boussetila, F. Rebbani, A modified quasi-reversibility method for a class of ill-posed Cauchy
problems. Georgian Math. J. 14, 627–642 (2007)
5. B. Campbell Hetrick, R.J. Hughes, Continuous dependence on modeling for nonlinear ill-posed
problems. J. Math. Anal. Appl. 349, 420–435 (2009)
6. D. Chen, B. Hofmann, J. Zou, Regularization and convergence for ill-posed backward
evolution equations in Banach spaces. J. Differ. Eq. 265, 3533–3566 (2018)
7. R. deLaubenfels, Entire solutions of the abstract Cauchy problem. Semigroup Forum 42, 83–
105 (1991)
8. R. deLaubenfels, C-semigroups and the Cauchy problem. J. Funct. Anal. 111, 44–61 (1993)
9. N. Dunford, J. Schwartz, Linear Operators, Part I (Wiley, New York 1957)
252 M. A. Fury

10. M.A. Fury, Nonautonomous ill-posed evolution problems with strongly elliptic differential
operators. Electron. J. Differ. Eq. 2013(92), 1–25 (2013)
11. M.A. Fury, A class of well-posed approximations for ill-posed problems in Banach spaces.
Commun. Appl. Anal. 23(1), 97–14 (2019)
12. M.A. Fury, Logarithmic well-posed approximation of the backward heat equation in Banach
space. J. Math. Anal. Appl. 475, 1367–1384 (2019)
13. M. Fury, B. Campbell Hetrick, W. Huddell, Continuous dependence on modeling in Banach
space using a logarithmic approximation, in Mathematical and Computation Approaches in
Advancing Modern Science and Engineering (Springer, Cham, 2016)
14. A. Gorny, Contribution à l’étude des fonctions dérivables d’une variable réelle. Acta Math. 71,
317–358 (1993)
15. M. Haase, The Functional Calculus for Sectorial Operators (Birkäuser Verlag, Basel, 2006)
16. Y. Huang, Modified quasi-reversibility method for final value problems in Banach spaces. J.
Math. Anal. Appl. 340, 757–769 (2008)
17. Y. Huang, Q. Zheng, Regularization for ill-posed Cauchy problems associated with generators
of analytic semigroups. J. Differ. Eq. 203, 38–54 (2004)
18. Y. Huang, Q. Zheng, Regularization for a class of ill-posed Cauchy problems. Proc. Amer.
Math. Soc. 133–10, 3005–3012 (2005)
19. K. Ito, B. Jin, Inverse Problems: Tikhonov Theory and Algorithms (World Scientific, Singapore,
2014)
20. R. Lattes, J.L. Lions, The Method of Quasi-reversibility, Applications to Partial Differential
Equations (Elsevier, New York, 1969)
21. N.T. Long, A.P.N. Dinh, Approximation of a parabolic non-linear evolution equation back-
wards in time. Inverse Probl. 10, 905–914 (1994)
22. A. Lorenzi, I.I. Vrabie, An identification problem for a linear evolution equation in a Banach
space and applications. Discrete Contin. Dyn. Syst Ser. S 4, 671–691 (2011)
23. I.V. Mel’nikova, General theory of the ill-posed Cauchy problem. J. Inverse Ill-posed Probl. 3,
149–171 (1995)
24. I.V. Mel’nikova, A.I. Filinkov, Abstract Cauchy Problems: Three Approaches. Chapman &
Hall/CRC Monographs and Surveys in Pure and Applied Mathematics, vol. 120 (Chapman &
Hall, Boca Raton, 2001)
25. K. Miller, Stabilized quasi-reversibility and other nearly-best-possible methods for non-well-
posed problems, in Symposium on Non-Well-Posed Problems and Logarithmic Convexity.
Lecture Notes in Mathematics, vol. 316 (Springer, Berlin, 1973), pp. 161–176
26. K. Miller, Logarithmic convexity results for holomorphic semigroups. Pacific J. Math. 58, 549–
551 (1975)
27. V. Nollau, Über den logarithmus abgeschlossener operatoren in Banachschen Räumen. Acta
Sci. Math. 30, 161–174 (1969)
28. L.E. Payne, Improperly Posed Problems in Partial Differential Equations. CBMS Regional
Conference Series in Applied Mathematics, vol. 22 (Society for Industrial and Applied
Mathematics, Philadelphia, 1975)
29. A. Pazy, Semigroups of Linear Operators and Applications to Partial Differential Equations
(Springer, New York, 1983)
30. A.I. Prilepko, D.G. Orlovsky, I.A. Vasin, Methods for Solving Inverse Problems in Mathemat-
ical Physics (Dekker, New York, 2000)
31. M. Reed, B. Simon, Methods of Modern Mathematical Physics, Vol. II: Fourier Analysis, Self-
Adjointness (Academic, New York, 1975)
32. W. Rudin, Real and Complex Analysis, 3rd edn. (McGraw-Hill, New York, 1987)
33. O. Scherzer, M. Grasmair, H. Grossauer, M. Haltmeier, F. Lenzen, Variational Methods in
Imaging. Applied Mathematical Sciences, vol. 167 (Springer, New York, 2009)
34. T. Schuster, B. Kaltenbacher, B. Hofmann, K.S. Kazimierski, Regularization Methods in
Banach Spaces (Walter de Gruyter, Berlin, 2012)
35. R.E. Showalter, The final value problem for evolution equations. J. Math. Anal. Appl. 47, 563–
572 (1974)
Solutions of Ill-posed Problems with Sufficiently Smooth Data 253

36. T.H. Skaggs, Z.J. Kabala, Recovering the release history of a groundwater contaminant. Water
Resour. Res. 30, 71–79 (1994)
37. D.D. Trong, N.H. Tuan, Stabilized quasi-reversibility method for a class of nonlinear ill-posed
problems. Electron. J. Differ. Eq. 2008(84), 1–12 (2008)
38. N.H. Tuan, D.D. Trong, On a backward parabolic problem with local Lipschitz source. J. Math.
Anal. Appl. 414, 678–692 (2014)
39. N.H. Tuan, D.D. Trong, T.H. Thong, N.D. Minh, Identification of the pollution source of a
parabolic equation with the time-dependent heat conduction. J. Inequal. Appl. 2014, 1–15
(2014)
A Closer Look at Bishop Operators

Eva A. Gallardo-Gutiérrez and Miguel Monsalve-López

Abstract The purpose of this work is to provide a survey, essentially self-

contained, of those results mainly concerned with the study of Bishop operators,
their (local) spectral properties and spectral invariant subspaces, whenever they
do exist. Finally, we will discuss Bishop-type operators, addressing some open
questions in this context.

Keywords Bishop operators · Invariant subspace problem · Spectral subspaces

Mathematics Subject Classification (2010) 47A15, 47B37, 47B38

1 Introduction

Given an irrational number α ∈ (0, 1), the Bishop operator Tα is defined on

Lp [0, 1), 1 ≤ p ≤ ∞, by

Tα f (t) = tf ({t + α}), t ∈ [0, 1),

where { · } denotes the fractional part. Clearly every Bishop operator Tα is the
product of two simple and well-understood operators, namely the multiplication
operator Mt by the independent variable in Lp [0, 1) and the composition operator
Cτα induced by the symbol τα (t) = {t + α}. Nevertheless, the structure of Tα
is largely unknown for every irrational α ∈ (0, 1). In particular, it is unknown

E. A. Gallardo-Gutiérrez ()
Dpto. de Análisis Matemático y Matemática Aplicada, Facultad de Ciencias Matemáticas,
Universidad Complutense de Madrid, Madrid, Spain
e-mail: [email protected]
M. Monsalve-López
Instituto de Ciencias Matemáticas ICMAT (CSIC-UAM-UC3M-UCM), Madrid, Spain
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 255

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_13
256 E. A. Gallardo-Gutiérrez and M. Monsalve-López

whether Tα has non-trivial closed invariant subspaces in Lp [0, 1), 1 ≤ p < ∞,

for every irrational α ∈ (0, 1). Indeed, according to Davie [15], these examples
were suggested by Errett Bishop in the fifties as possible candidates for operators
without non-trivial closed invariant subspaces, and therefore, for counterexamples
to the Invariant Subspace Problem.
Bishop operators are particular instances of the so-called weighted translation
operators defined in a more general measurable setting as follows: if (X, F , μ) is a
non-atomic measure space, φ ∈ L∞ (X) and τ a measurable measure-preserving
mapping from X to itself which is invertible with measurable inverse, then the
weighted translation operator Wφ,τ is defined by the equation

Wφ,τ f = φ (f ◦ τ ).

The class consisting of weighted translations operators is a large one, containing

as particular examples bilateral weighted shifts and, therefore, roughly speaking,
model operators.
Weighted translation operators were firstly studied by Parrott [26] in his Ph.D.
thesis, analyzing the spectrum, numerical range and reducing subspaces of such
operators. Indeed, Parrott computed the spectrum of Bishop operators showing, in
particular, that it is the disc
 
σ (Tα ) = z ∈ C : |z| ≤ e−1 ,

independently of the irrational α ∈ (0, 1) and moreover, Tα lacks point spectrum

for every α. In 1973, Bastian [6] gave unitary invariants for some weighted
translation operators and studied properties like subnormality and hyponormality
among them. Later on, Petersen [27] showed some results on the commutant of
weighted translation operators in an attempt to get a deeper insight in the general
context. Nevertheless, many open questions are left and complete answers seem to
be far away.
As far as Bishop operators concerns, one of the most striking result was proved
by Davie [15] in 1974, who by means of a functional calculus approach, was able
to show the existence of non-trivial closed invariant subspaces in Lp [0, 1) for Tα
whenever α is a non-Liouville irrational number in (0, 1). Recall that an irrational
α is a Liouville number if for every n ∈ N there exists an irreducible rational number
pn /qn such that

pn 1
α− < n.
qn qn

Observe that by Jarník-Besicovitch Theorem (see [11, Section 5.5], for instance),
Liouville irrationals form a set of vanishing Hausdorff dimension; so for almost
every α ∈ (0, 1), Davie’s Theorem states that Tα has non-trivial closed invariant
subspaces in Lp [0, 1). More indeed, Tα has non-trivial closed hyperinvariant
A Closer Look at Bishop Operators 257

subspaces, that is, closed subspaces invariant under every linear bounded operator
in the commutant of Tα .
In the nineties, extensions strengthening Davie’s Theorem were due to Blecher
and Davie [9] and MacDonald in [22] for Bishop-type operators, that is, weighted
translation operators in Lp [0, 1) where τ = τα :

Wφ,τα f (t) = φ(t) f ({t + α}), t ∈ [0, 1).

Once again, the brick wall consisted of Liouville irrationals, and despite of the
efforts, the interesting extensions regarding the non-invertible weighted translation
operators included many weights φ but neither Liouville number.
In 2008, Flattot [18], by refining the functional calculus approach, was able to
provide a large class of irrationals α ∈ (0, 1) including some Liouville numbers for
which Tα has non-trivial closed hyperinvariant ∞ subspaces in Lp [0, 1) -for instance,
the classical example of a Liouville number n=1 10−n! .
Recently, in [12] the authors have extended the class of irrationals α ∈ (0, 1) such
that Tα has non-trivial closed hyperinvariant subspaces in Lp [0, 1) by considering
arithmetical techniques to strengthen the analysis of certain functions associated
to the functional calculus model. Indeed, for these Liouville numbers, Gallardo-
Gutiérrez and Monsalve-López [20] have recently shown the existence of non-trivial
spectral subspaces, revealing, indeed, the spectral nature of the hyperinvariant
subspaces. Moreover, those Liouville numbers α left, that is, those for which it is
open whether Tα has a non-trivial closed invariant subspace, are so extreme that the
functional calculus approach fails to produce invariant subspaces (see [12, Theorem
4.1]).
The aim of this work is to survey the recent results at this regard and show how
techniques in Operator Theory, Analytic Number Theory or Spectral Theory are
linked together to produce, when it succeeds, invariant subspaces for such a simple
family of operators as Bishop operators are. To that end, the rest of the manuscript
is organized as follows. In Sect. 2 we deal with a first approach by studying
the behaviour of the norm of the iterates Tαn (clearly, they tend to 0 by means
of Parrott’s characterization of σ (Tα )). In Sect. 3, we detail the main techniques
and approaches previously mentioned to provide invariant subspaces for Bishop
operators recalling the updated results in this context. In Sect. 4, our approach deals
with local spectral properties fulfilled by Bishop operators Tα , independently of
the irrational α ∈ (0, 1). Note that though all Bishop operators Tα share the same
spectrum, not all of them are known to possess invariant subspaces; and hence, a
deeper insight in σ (Tα ) could lead to study invariant subspaces by considering local
spectral properties like the Dunford property (Property (C)) or the Bishop property
(β). Indeed, in Sect. 5 the local spectral analysis is pushed further to provide spectral
subspaces for all those Bishop operators that, up to now, are known to have non-
trivial closed hyperinvariant subspaces. Finally, in Sect. 6, we deal with Bishop-type
operators and their properties, ending up with some open questions in this context.
258 E. A. Gallardo-Gutiérrez and M. Monsalve-López

2 A First Approach: Understanding the Behaviour of Tαn 

In this section, and as a preliminary stage, we determine explicitly the norm of the
iterates of Tα acting on Lp [0, 1) for 1 ≤ p ≤ ∞. In particular, it provides an insight
of the behaviour of Tα depending on the irrational α.
Let n be a positive integer and denote by Tαn the n-th iterate of Tα . A simple
computation shows that

Tαn f (t) = t {t + α} · · · {t + (n − 1)α} f ({t + nα}) ,

for any Lp [0, 1). Hence,

Tαn  = Sn−1 (α)

where

Sn (α) := ess sup t{t + α} · · · {t + nα}.

t ∈[0,1)

In Figs. 1 and 2, we represent the plots of Sn (α) for small values of n.

It turns out that the minimum of Sm (α) is reached at α = m+1 1
and its value is
   
1 1 n n!
Sn = lim t t+ ··· t + = .
n+1 t →1− n+1 n+1 (n + 1)n
So, upon applying Stirling Formula, one deduces

n! √
min Sn (α) = ∼ e−(n+1) 2πn.
α∈[0,1) (n + 1) n

Accordingly, the following straightforward proposition follows:

Proposition 2.1 For every irrational α ∈ [0, 1) and n a positive integer
 
n−1 
(n − 1)!
≤ Tαn  = Sn−1 (α) ≤ exp ess sup log{t + j α}
nn−1 t ∈[0,1) j =0

Observe that, by means of Birkoff Ergodic Theorem, it is possible to compute

the
⎛ ⎞
1 
n−1
lim exp ⎝ess sup log{t + j α}⎠
n t ∈[0,1) n j =0

and hence deduce that the spectral radius of Tα

r(Tα ) = lim Tαn 1/n = e−1

n

as Parrott showed in [26].

A Closer Look at Bishop Operators 259

Gallardo-Gutiérrez and Monsalve-López

1 1

0.75 0.75

0.5 0.5

0.25 0.25

0.25 0.5 0.75 1 0.25 0.5 0.75 1

(a) (b)

1 1

0.75 0.75

0.5 0.5

0.25 0.25

0.25 0.5 0.75 1 0.25 0.5 0.75 1

(c) (d)

Fig. 1 Sm (α) for m = 1, . . . , 4. (a) S1 (α) for α ∈ [0, 1). (b) S2 (α) for α ∈ [0, 1). (c) S3 (α) for
α ∈ [0, 1). (d) S4 (α) for α ∈ [0, 1)

3 Invariant Subspaces of Bishop Operators and a Theorem

of Atzmon

The first goal of this section will be giving a detailed outline of the main techniques
and approaches used to find invariant subspaces for Bishop operators. This will serve
as a context in order to, thereupon, present in more detail our results, in which we
considerably enlarge the set of known values α such that the Bishop operator Tα has
invariant subspaces on each Lp [0, 1). On the other hand, at the end of the section, we
will show that in some sense, when our approach fails to produce invariant subspaces
for Tα it is actually because the standard techniques cannot be applied anymore.
All the original results appearing in this section are included in the article [12].
260 E. A. Gallardo-Gutiérrez and M. Monsalve-López

A closer look at Bishop operators

1 1

0.75 0.75

0.5 0.5

0.25 0.25

0.25 0.5 0.75 1 0.25 0.5 0.75 1

(a) (b)

1 1

0.75 0.75

0.5 0.5

0.25 0.25

0.25 0.5 0.75 1 0.25 0.5 0.75 1

(c) (d)

Fig. 2 Sm (α) for m = 5, . . . , 8. (a) S5 (α) for α ∈ [0, 1). (b) S6 (α) for α ∈ [0, 1). (c) S7 (α) for
α ∈ [0, 1). (d) S8 (α) for α ∈ [0, 1)

3.1 Beurling Algebras and a Theorem of Atzmon

As it was aforementioned above, many authors have addressed the problem of

seeking non-trivial closed invariant subspaces for all Bishop operators [12, 15, 18].
Ignoring the differences among all such distinct approaches within the literature,
all of them rely essentially on the same idea: a functional calculus based on regular
Beurling algebras which was properly formalized by Atzmon [4] in the mid eighties.
Given a sequence ρ := (ρn )n∈Z in [1, +∞) with ρ0 = 1 and such that

ρm+n ≤ ρm ρn for every m, n ∈ Z, (3.1)

A Closer Look at Bishop Operators 261

1/n
lim ρn = 1; (3.2)
|n|→∞

we may consider its corresponding Beurling algebra Aρ consisting of all continuous

functions f : T → C with norm defined by

f ρ := |f(n)| ρn < ∞,
n∈Z

where (f(n))n∈Z denotes the sequence of Fourier coefficients of f . Endowed with

the norm  · ρ , the Beurling algebra Aρ is a semi-simple commutative Banach
algebra.
One of the most remarkable results regarding Beurling algebras Aρ is the
subsequent sufficient criterion, which seems to date back to Beurling [8], to
determine the regularity of Aρ . Recall that a function algebra A on a compact space
K is said to be regular if for all p ∈ K and all compact subset M ⊆ K with p ∈ M,
there exists f ∈ A such that f (p) = 1 and f = 0 on M.
Theorem 3.1 (Beurling [8]) Let ρ := (ρn )n∈Z be a real sequence satisfying
both (3.1) and (3.2). Then, the Banach algebra Aρ is regular whenever

 log ρn
< ∞. (3.3)
1 + n2
n∈Z

Condition (3.3) is usually known as Beurling condition and it is closely related to

the Denjoy-Carleman theorem on quasi-analytic classes [28]. Keeping the previous
result in mind, it is natural to consider the following definition:
Definition 3.2 A sequence of real numbers (ρn )n∈Z such that ρ0 = 1 and ρn ≥ 1
for all n ∈ Z, is called a Beurling sequence if
 log ρn
ρm+n ≤ ρm ρn (∀m, n ∈ Z) and < +∞.
1 + n2
n∈Z

One advantage of regularity in a function algebra is that it enables to construct

two non-zero functions whose product is identically zero. This idea, combined with
a functional calculus argument, provides a powerful method for obtaining invariant
subspaces. Such a strategy, firstly studied by Wermer [29] for invertible operators,
was lately refined by Atzmon [4] to the non-invertible case:
Theorem 3.3 (Atzmon [4]) Let T ∈ B(X) be an operator on a complex Banach
space X and suppose there exist two sequences (xn )n∈Z in X and (yn )n∈Z in X∗
such that x0 = 0, y0 = 0 and

T xn = xn+1 and T ∗ yn = yn+1 (∀n ∈ Z).

262 E. A. Gallardo-Gutiérrez and M. Monsalve-López

Suppose further that both sequences (xn )n∈Z and (yn )n∈Z are dominated by a
Beurling sequence, and there is at least a λ ∈ T at which the following functions
Gx and Gy do not both possess analytic continuation into a neighbourhood of λ:

∞
n=1 x−n z
n−1 if |z| < 1;
Gx (z) := 0
− n=−∞ x−n zn−1 if |z| > 1.

∞
n=1 y−n z
n−1 if |z| < 1;
Gy (z) := 0
− n=−∞ y−n zn−1 if |z| > 1.

Then, either T is a multiple of the identity or it has a non-trivial closed hyperinvari-

ant subspace.
As it was announced at the beginning of this subsection, Atzmon’s Theorem
turns out to be a useful machinery to produce invariant subspaces for many
Bishop operators. Nevertheless, two things are still necessary: a right choice of the
sequences (xn )n∈Z and (yn )n∈Z and a careful analysis on the growth of their norms.
All this will be pursued in the following subsection.

3.2 Invariant Subspaces of Bishop Operators

Applying a novel version of these methods, Davie [15] was the first in obtaining
positive results on the existence of invariant subspaces for some classes of Bishop
operators. In particular, he showed that whenever α ∈ (0, 1) is a non-Liouville
number, the corresponding Bishop operator Tα has a non-trivial hyperinvariant
subspace on each Lp [0, 1).
Further refinements due to Flattot [18] enlarged the class of such irrational
numbers, embracing some Liouville numbers. More specifically, using the language
of continued fractions, if (aj /qj )j ≥0 denote the convergents of α, the limit of
Flattot’s result ensures the existence of invariant subspaces for Tα on each Lp [0, 1)
whenever
1/2−ε 
log qj +1 = O qj for every ε > 0. (3.4)

Turning to Bishop operators, up to date, our approach yields the most general
result regarding the existence of invariant subspaces [12]:
Theorem 3.4 (Chamizo et al. [12]) Let α ∈ (0, 1) be any irrational and
(aj /qj )j ≥0 the convergents in its continuous fraction. If the following condition
holds
 q 
j
log qj +1 = O ; (3.5)
log3 qj

then, the Bishop operator Tα has a non-trivial closed hyperinvariant subspace in

Lp [0, 1) for every 1 ≤ p ≤ ∞.
A Closer Look at Bishop Operators 263

Again, the idea behind our address is exactly the same: a careful application of
Atzmon’s Theorem. For this, it shall be convenient to work with the operators
α := e Tα .
T

This scalar multiple of Tα , far from being arbitrary, is determined by the fact
σ (Tα ) = D. Now, without going into much details, an insight in the proof may
be the following: given any irrational α ∈ (0, 1), in the sequel, consider the real
functions


n−1

Ln (t) := 1 + log({t + j α}) , (n ∈ N).
j =0

It is plain that the functions Ln (t) play a fundamental role in the understanding of
the iterates of both Tα and Tα∗ , since they codify their behaviours via the equations

Tαn f (t) = eLn (t )f ({t + nα}),

(n ∈ N), (3.6)
α−n f ({t + nα}) = e−Ln (t )f (t),
T

and, in a similar way, for Tα∗ . Note that Eq. (3.6) makes sense for every Lp -function
despite of the fact that Tα has not a bounded inverse.
In light of (3.6), it is not hard to figure out that, in order to control the growth
of both iterates T αn f and (Tα∗ )n f , it could be a good idea to construct ad hoc a
function which kills each of the singularities arising from each summand of Ln (t).
To accomplish that, we may consider, for instance, the characteristic function of the
set
1 19 1 
Bα := <t < : t − nα > 2
for every n ∈ Z \ {0} ,
20 20 20n

where t := min {t}, 1 − {t} denotes the distance from t to the closest integer.
First, note that, since the set Bα has positive Lebesgue measure, it follows that 1Bα
does not vanish identically as an element of Lp [0, 1). On the other hand, it turns
out that |Ln (t)| can be bound properly in terms of the sequence (qj )j ≥0 for t ∈ Bα ,
when |n| is comparable to qj +1 .
We prefer not to deal with technicalities which are out of the scope of this survey.
Anyway, as explained in [12], after some arithmetical estimations, we have:
Proposition 3.5 Let α ∈ (0, 1) be any irrational and (aj /qj )j ≥0 the convergents
3/2
in its continuous fraction. Then, for every n ∈ Z with |n| ≤ qj +1 , we have
; n ; ; ; 
log 1 + ;
Tα 1Bα ;∞ + ;(Tα∗ )n 1Bα ;∞
|n| |n| + qj +1 (3.7)
9 qj + log qj + log(|n| + 1).
qj qj +1
264 E. A. Gallardo-Gutiérrez and M. Monsalve-López

Now, as a consequence of the asymptotic bound (3.7), if we impose a suitable

control on the growth of qj +1 in terms of qj , we manage to find a Beurling sequence
which dominates both (Tαn 1Bα )n∈Z and ((Tα∗ )n 1Bα )n∈Z . More specifically, it
may be seen that if condition (3.5) is satisfied, then the application of Atzmon’s
Theorem to Bishop operators can be achieved from picking the Beurling sequence
 
C |n|
ρn = exp (n ∈ Z).
log(2 + |n|)(log log(4 + |n|))2

Observe that Theorem 3.4 relaxes significantly the condition imposed on α by

Flattot (3.4), allowing the exponent 1 instead of 1/2 and quantifying the role of ε.
Surprisingly, as we shall see hereafter, Theorem 3.4 is essentially the best possible
result attainable from Atzmon’s Theorem approach and any improvement seems to
require different functional analytical techniques.
To conclude this subsection, we remark that it is possible to measure the
difference between those cases covered by Davie [15], Flattot [18] and Theorem 3.4.
For this, it suffices to consider the logarithmic Hausdorff dimension via the family
of functions (| log x|−s )s≥0 (instead of the usual (x s )s≥0 ). With such a dimension, as
a consequence of [11, Thm. 6.8], one has that the set of exceptions in Davie, Flattot
and our case have dimension ∞, 4 and 2 respectively.

3.3 The Limits of Atzmon Theorem

Apart from being a refinement of Flattot/Davie approaches, the importance of

Theorem 3.4 is evinced when it is understood in conjunction with the following
statement:
Theorem 3.6 (Chamizo et al. [12]) Let α ∈ (0, 1) be an irrational not in
  q 
j
M := α ∈ (0, 1) : log qj +1 = O
log qj

and consider Tα on Lp [0, 1) for some 1 ≤ p < ∞. Then, for every non-zero
f ∈ Lp [0, 1), we have
; n ; 
 log 1 + ;
Tα f ;p
= ∞.
1 + n2
n∈Z

Here, as an abuse of notation, for f ∈ Lp [0, 1) and n > 0, Tα−n f  denotes

the norm of the n-th backward iterate Tα−n f whenever it belongs to Lp [0, 1) or ∞,
otherwise.
Observe that this latter result shows that we cannot hope to improve much on
Theorem 3.4, since at most, we shall be able to gain a power in log qj . Hence, it
A Closer Look at Bishop Operators 265

establishes a threshold limit in the growth of the denominators of the convergents of

α for the application of Atzmon’s Theorem in order to find invariant subspaces for
Tα .
Heuristically, in order to prove Theorem 3.6, we must show that, independently
of f ∈ Lp [0, 1), one of the quantities Tαn f p or Tα−n f p is large enough for
many values of n ∈ N, inducing the series
; m ; 
 log 1 + ;
Tα f ;p
1 + m2
m∈Z

to be divergent. To do so, we consider the identity

 1 
Tαn f p + Tα−n f p = epLn (t −nα) + e−pLn (t ) |f (t)|p dt,
p p
(3.8)
0

and observe that, α ∈ M means exactly that α is pretty close to some rationals a/q.
Roughly speaking, this implies that whenever n is comparable to q and q | n, we
have a very accurate approximation between
 a
Ln (t − nα) ≈ Ln t − n = Ln (t).
q

But, in such a case, it results that the integral in (3.8) must be large unless |Ln (t)| is
small, which turns out to happen rarely.

4 Local Spectral Properties of Bishop Operators

The aim of the current section is identifying which local spectral properties are
shared by all Bishop operators Tα , independently of the irrational α ∈ (0, 1).
This apparent digression is justified by the facts presented in the preceding
sections: while all Bishop operators Tα share the same spectrum, only a few of
them are known to possess invariant subspaces; hence, might a deeper insight in
σ (Tα ) help us to deal with those cases not covered by Atzmon’s Theorem? Actually,
this naive idea of studying invariant subspaces via spectral subsets rarely succeeds;
nevertheless, as we shall argue at the end of the survey, this will not be completely
senseless in the case of Bishop operators.
All the original results presented in Sect. 4.2 may be found in [19]; on the other
hand, those appearing in Sect. 4.3 are stated in [12].
266 E. A. Gallardo-Gutiérrez and M. Monsalve-López

4.1 Some Preliminaries on Local Spectral Theory

We begin this part by recalling some preliminaries regarding local spectral theory.
In what follows, X will stand for an arbitrary complex Banach space, L(X) will
denote the class of linear operators on X and B(X) the Banach algebra of linear
bounded operators on X.
Given any operator T ∈ B(X), let σT (x) denote the local spectrum of T at
x ∈ X; i.e., the complement of the set of all λ ∈ C for which there exists an open
neighbourhood Uλ 4 λ and an analytic function f : Uλ → X such that

(T − zI )f (z) = x for every z ∈ Uλ . (4.1)

The complement of σT (x), which is denoted by ρT (x), stands for the local resolvent
of T at x ∈ X; and a function verifying (4.1) is called a local resolvent function
nearby λ ∈ C for x ∈ X. In general, the uniqueness of the local resolvent
function cannot be assumed. Thus, T satisfies the single-valued extension property
(abbreviated as SVEP) if, fixed any x ∈ X and λ ∈ ρT (x), there exists a unique
local resolvent function on a sufficiently small neighbourhood of λ. We highlight
that any operator with σp (T ) = ∅ enjoys the SVEP.
The notion of local spectrum allows us to gain a further knowledge on what con-
stitutes each part of σ (T ), this may be accomplished via local spectral manifolds:
given any subset  ⊆ C, the local spectral manifold XT () is defined as
 
XT () := x ∈ X : σT (x) ⊆  .

In general, XT () is always a T -hyperinvariant linear manifold; nevertheless, its

closeness cannot be assured, even when  ⊆ C is closed. Those operators T ∈
B(X) for which XT (F ) is norm-closed in X for every closed subset F ⊆ C are said
to satisfy the Dunford property or Property (C). We remark that Dunford property
(C) implies SVEP.
An important class of Banach space-operators devised to resemble the spectral
behaviour of normal operators is the class of decomposable operators. An operator
T ∈ B(X) is said to be decomposable if given any open cover {U, V } of C, there
exist two closed invariant subspaces Y, Z ⊆ X such that

σ (T |Y ) ⊆ U and σ (T |Z) ⊆ V ,

with X = Y + Z. Related to decomposability, two other weaker local spectral

properties arise: the Bishop’s property (β) and the decomposition property (δ). An
operator T ∈ B(X) has the property (β) if, for every open set U ⊆ C and any
sequence of analytic functions fn : U → X such that

(T − zI )fn (z) → 0 as n → ∞
A Closer Look at Bishop Operators 267

locally uniformly on U , then fn → 0 locally uniformly on U as well. We point

out that property (β) entails both property (C) and SVEP. On the other hand, an
operator T ∈ B(X) satisfies property (δ) if, for every open cover {U, V } of C we
have

X = XT (U ) + XT (V );

where, for each closed F ⊆ C, the linear manifold XT (F ) is formed by all x ∈

XT (F ) whose local resolvent function is globally defined on C \ F . It may be seen
that an operator is decomposable if and only if verifies both properties (β) and (δ).
To conclude this part, just mention that one of the most remarkable results in
local spectral theory, due to Albrecht and Eschmeier [2], claims that properties (β)
and (δ) are duals of each other. Furthermore, it happens that both properties (β) and
(δ) can be comprehended in terms of decomposability: property (β) characterizes
the restrictions of decomposable operators to invariant subspaces, while property
(δ) characterizes the quotients of decomposable operators by invariant subspaces.
Finally, we disclose that properties (β) and (δ) can be combined with the Scott
Brown techniques to produce invariant subspaces [16], extending substantially the
classical result by S. W. Brown [10] on hyponormal operators.

4.2 Power-Regularity and the Local Spectral Radius of Bishop

Operators

The programme to followed up along the subsequent pages consists on checking

which of the aforemention local spectral properties are satisfied by all Bishop
operators. In order to accomplish such a task, we shall make use of several
estimations and previous results. The simplest case turns out to be the SVEP, since
it follows immediately from the fact σp (Tα ) = ∅:
Proposition 4.1 Let α ∈ (0, 1) be any irrational and Tα acting on Lp [0, 1) for
1 ≤ p ≤ ∞. Then, Tα satisfies the SVEP.
Not surprisingly, regarding the remainder of local spectral properties, we shall
need to work significantly harder and provide ourselves with new estimations
concerning the behaviour of Bishop operators.
Given any operator T ∈ B(X), the Gelfand’s formula for the spectral radius
asserts that the sequence (T n 1/n )n≥0 is always convergent and
 
r(T ) := max |λ| : λ ∈ σ (T ) = lim T n 1/n .
n→∞

Nevertheless, if we aspire to wield a local version of Gelfand’s formula, we will need

to perform some modifications, since it turns out that (T n x1/n )n≥0 may be non-
convergent. At this regard, Atzmon [5] introduced the notion of power-regularity:
we recall that T ∈ B(X) is called power-regular if the sequence (T n x1/n )n≥1
268 E. A. Gallardo-Gutiérrez and M. Monsalve-López

is convergent for every x ∈ X. Furthermore, Atzmon proved a general criterion

showing that a wide class of Banach space-operators, including decomposable
operators, are power-regular. This forces to define the local spectral radius of T
at x ∈ X as

rT (x) := lim sup T n x1/n .

n→∞

Not surprisingly, likewise in Gelfand’s formula, the local spectral radius and the
local spectrum may be related via the inequality
 
rT (x) ≥ max |λ| : λ ∈ σT (x) ,

which can be replaced by an equality whenever T has the SVEP.

A general result due to Müller [24], asserts that given any T ∈ B(X), the equality
rT (x) = r(T ) must hold on a dense subset of X, which is indeed of second category
[14]. Our next proposition exhibits that Bishop operators are extreme examples in
this sense; more precisely, it shows that Tα is always power-regular with rTα (f ) =
r(Tα ) for every non-zero f :
Theorem 4.2 (Gallardo-Gutiérrez and Monsalve-López [19]) Let α ∈ (0, 1) be
any irrational and consider Tα acting on Lp [0, 1) for 1 ≤ p < ∞. Then, for every
non-zero f ∈ Lp [0, 1), we have

= r(Tα ) = e−1 .
1/n
lim Tαn f p
n→∞

Moreover, the same holds for Tα∗ .

We point out that the latter proposition can be proved for a considerably larger
family of weighted translation operators [19, Cor. 2.3]. On the other hand, it is worth
mentioning that its proof is mainly based on an application of ergodic theorems.
Anyway, it entails the following striking corollary:

Corollary 4.3 Let α ∈ (0, 1) be any irrational and Tα ∈ B Lp [0, 1) for 1 ≤ p <
∞. Suppose M is a non-zero closed invariant subspace for Tα , then

r(Tα |M) = e−1 .

Given any Bishop operator Tα , as a somewhat direct consequence of the

preceding results and the asymptotic bound provided by Proposition 3.5, we will
be able to discard the fulfilment of the rest of the local spectral properties. In
this subsection, we shall exclusively focus on the strongest ones: decomposability,
property (β) and property (δ):
Theorem 4.4 (Gallardo-Gutiérrez and Monsalve-López [19]) Let α ∈ (0, 1) be
any irrational and consider Tα acting on Lp [0, 1) for some 1 ≤ p ≤ ∞. Then, Tα
is not decomposable.
A Closer Look at Bishop Operators 269

Once we know Corollary 4.3, the Proof of Theorem 4.4 is pretty simple: suppose
that Tα is decomposable in Lp [0, 1); then, fixed arbitrary 0 < r < s < e−1 , we may
consider the open cover

U = D(0, s) and V = C \ D(0, r).

By means of Corollary 4.3, the unique closed invariant subspace Y ⊆ Lp [0, 1) such
that σ (Tα |Y ) ⊆ U would be Y = {0}; but this leads us to a contradiction, since
σ (Tα ) ⊆ V .
Furthermore, it is possible to obtain a slightly stronger result by combining
a similar argument to the previous one with the duality principle [2] between
properties (β) and (δ):
Theorem 4.5 (Gallardo-Gutiérrez and Monsalve-López [19]) Let α ∈ (0, 1) be
any irrational and consider Tα on Lp [0, 1) for 1 < p < ∞. Then, Tα does not
satisfy either property (β) or (δ).
As a final remark, note that one could reasonably argue that, in a very vague
sense, Proposition 4.2 and Corollary 4.3 seem to be suggesting that the most
significant part of σ (Tα ) is precisely its boundary. This intuition shall be explored
more deeply and confirmed in more concrete terms throughout the rest of the article.

4.3 Dunford Property and a Dense Local Spectral Manifold

Throughout this subsection, we shall prove that Bishop operators can neither satisfy
property (C); nevertheless, along the way, something much more stronger will be
shown. More precisely, we shall see that, unlike for subsets  ⊆ int(σ (Tα )), for
which Proposition 4.2 ensured that XTα () = {0}; the local spectral manifold
XTα (∂σ (Tα )) turns out to be dense for every irrational α. To do so, we are going
to split the proof into two parts: firstly, using (3.7), we will check that σTα (1Bα )
must be confined into ∂σ (Tα ); on the other hand, an argument with L∞ functions
will let us to elucidate the density of XTα (∂σ (Tα )).
Proposition 4.6 Let α ∈ (0, 1) be any irrational number and consider Tα acting
on Lp [0, 1) for some 1 ≤ p < ∞. Then, both local spectra

σTα (1Bα ) ⊆ ∂σ (Tα ) and σTα∗ (1Bα ) ⊆ ∂σ (Tα ).

In order to prove Proposition 4.6, consider the Lp [0, 1)-valued analytic function
given by
∞
 
gα (z) := α−n 1Bα · zn−1 .
T
n=1
270 E. A. Gallardo-Gutiérrez and M. Monsalve-López

Now, note that gα must be analytic on the open disk |z| < 1, since, by means of the
asymptotic estimate (3.7), we have
; −n ;1/n
lim sup ;
Tα 1Bα ;p ≤ 1.
n→∞

Finally, it is routine to check that (Tα − zI )gα (z) = 1Bα for every |z| < 1.
An important remark is in order: given an operator with σp (T ) = σp (T ∗ ) = ∅,
it may be seen that if Atzmon’s Theorem may be applied to the sequences

xn := T n x and yn := (T ∗ )n y (n ∈ Z)

for a pair x ∈ X and y ∈ X∗ , the corresponding local spectra σT (x) and σT ∗ (y)
must be inside ∂D. Nevertheless, the converse is not true. This evinces the main
gain of the asymptotic (3.7) with respect to the ones in the previous works [15] and
[18], since Proposition 3.5 can be applied independently of the irrational α.
Now, once we are aware of 1Bα ∈ XTα (∂σ (Tα )), we are in position to take
another step ahead and prove the density of XTα (∂σ (Tα )):
Theorem 4.7 (Chamizo et al. [12]) Let α ∈ (0, 1) be any irrational number. Then,
the local spectral manifold
  
XTα ∂σ (Tα ) = f ∈ Lp [0, 1) : σTα (f ) ⊆ ∂σ (Tα )

is norm-dense in Lp [0, 1) for each 1 ≤ p < ∞. In particular, Tα does not satify the
Dunford property (C) on Lp [0, 1) for 1 ≤ p < ∞.
In the argument of Theorem 4.7, we repeatedly exploit the fact that 1Bα is a
characteristic function. Firstly, note that standard bounds yield the inclusion

g(t)1Bα (t) : ess sup |g(t)| < ∞ ⊆ XTα (∂σ (Tα )).
t ∈[0,1)

Therefore, by the density of L∞ into Lp for each 1 ≤ p < ∞ and taking into
account that the support of 1Bα is precisely the set Bα , we have
 
f ∈ Lp [0, 1) : supp(f ) ⊆ Bα ⊆ XTα (∂σ (Tα )).

Now, since Tα 1Bα ∈ XTα (∂σ (Tα )) as well, we deduce a similar inclusion:

tg(t)1Bα ({t + α}) : ess sup |g(t)| < ∞ ⊆ XTα (∂σ (Tα )),
t ∈[0,1)

but, as the multiplication operator Mt is of dense-range, this again entails

 
f ∈ Lp [0, 1) : supp(f ) ⊆ τα−1 (Bα ) ⊆ XTα (∂σ (Tα )).
A Closer Look at Bishop Operators 271

Repeating a similar discussion with Tαn 1Bα for each n ∈ Z, we have

 
span f ∈ Lp [0, 1) : supp(f ) ⊆ ταm (Bα ) ⊆ XTα (∂σ (Tα )),
m=−N,...,N

but, since Bα has strictly positive measure and τα is ergodic, the proof is done.
As it was promised before beginning this subsection, we may clearly appreciate
that the meaning behind each part of σ (Tα ) differs significantly. In this sense, it is
plain that the boundary of σ (Tα ) stores much more information about Tα than the
interior. This idea will be raised into a new level in the following section, showing
that those already known closed invariant subspaces for Tα can be characterized as
the norm-closure of some local spectral manifolds related to ∂σ (Tα ).

5 Spectral Decompositions of Bishop Operators

It is plain from the preceding section, that we cannot expect to encounter invariant
subspaces for Tα when we consider certain local spectral manifolds. For example,
if we pick an arbitrary set  inside in C \ ∂σ (Tα ), we fall too short since, in such a
case, the associated local spectral manifold is trivial:

XTα () = {0};

on the other hand, if ∂σ (Tα ) ⊆ , we go too far since, in such case, the associated
local spectral manifold is dense:

XTα () = Lp [0, 1),

for the corresponding 1 ≤ p < ∞. These two facts restrict significantly our quest
of invariant subspaces for Tα via local spectral manifolds and indicate us that, our
unique hope could be choosing subsets  which intersect the boundary of σ (Tα ) but
not covering it entirely. Surprisingly, at least for many values of α, this new attempt
turns out to be successful, bringing about local spectral manifolds verifying

{0} = XTα () = Lp [0, 1).

Nevertheless, not everything shall be good news, since our address will depend
on a local spectral variant of Atzmon’s Theorem and the restriction imposed by
Theorem 3.6 still remains. Anyway, considering these ideas, it seems fully justified
to conjecture about the possibility of addressing the invariant subspace problem for
general Bishop operators Tα using techniques borrowed from local spectral theory.
This section is mainly based on the article [20].
272 E. A. Gallardo-Gutiérrez and M. Monsalve-López

5.1 The Spectral Meaning Behind Atzmon’s Theorem

As before, X will be any infinite-dimensional Banach space, L(X) will stand for
the class of linear operators on X while B(X) will denote the Banach algebra of
linear bounded operators on X. Such distinction between linear and linear bounded
operators will be of particular interest along the current section.
As the title announces, throughout the first half of this section, our aim is
unravelling the spectral meaning hidden behind the statement of Atzmon’s Theorem
(and related results in [3] and [7]) in order to, further on, apply it in the study of
Bishop operators.
The preliminary step before understanding properly Atzmon’s Theorem from a
spectral perspective should be the following decomposability version of Wermer’s
theorem [29] given by Colojoară and Foiaş [13]:
Theorem 5.1 (Colojoară and Foiaş [13]) Let T ∈ B(X) be an invertible operator
on a complex Banach space X with σ (T ) ⊆ T satisfying
 log T n 
< ∞. (5.1)
1 + n2
n∈Z

Then, T is decomposable. In particular, if σ (T ) is not reduced to a singleton, the

operator T has a non-trivial closed hyperinvariant subspace in X.
A scheme of the proof may be the following: firstly, note that the real sequence
ρ := (T n )n∈Z is submultiplicative (3.1), satisfies the limit-type condition (3.2)
and, as a byproduct of (5.1), defines a Beurling sequence. Therefore, its corre-
sponding Beurling algebra Aρ is regular. Thus, we may consider the continuous
algebraic homomorphism determined by

φ : Aρ → B(X)
(5.2)
eınt → T n

for every n ∈ Z. Now, given any open cover {U, V } of σ (T ), due to the regularity
of Aρ , we can find a function h ∈ Aρ such that

h ≡ 1 on (C \ V ) ∩ σ (T ) and h ≡ 0 on (C \ U ) ∩ σ (T );

but, the manner in which h has been chosen causes that the ranges of the operators
φ(h) and I − φ(h) split the spectrum of T in the following way:
 
σ T | φ(h)(X) ⊆ U and σ T | (I − φ(h))(X) ⊆ V ,

and consequently, as we intended to show, T turns out to be decomposable.

At a glance, it sounds reasonable to believe that Atzmon’s Theorem looks
like a localized version of Colojoară and Foiaş decomposability theorem. Hence,
A Closer Look at Bishop Operators 273

one may rightly conjecture that Atzmon’s Theorem should also involve a sort
of decomposition on the spectrum σ (T ); nevertheless, in this case, such spectral
decomposition shall be much more subtle and less manageable. More specifically,
when the invertibility of T is missing, we are forced to consider a decomposition
of σ (T ) only with respect to a proper linear submanifold of X. For the sake of
completeness, we explain it in more accurate terms: in the sequel, we are intended
to modify the Proof of Theorem 5.1 with the aim of fitting it into the hypothesis
required by Atzmon’s Theorem.
Let ρ := (ρn )n∈Z be an arbitrary Beurling sequence. Since T needs not to be
invertible anymore, the algebraic homomorphism (5.2) must be replaced by

φ : Aρ → L(X)
(n ∈ Z)
eınt → T n

embracing some non-bounded linear operators. Nevertheless, in order to retrieve

some appropriate spectral properties, we must restrict ourselves to work on a linear
submanifold of X in which φ can be properly controlled; this is the goal behind the
definition of the continuity core:
 
Dφ := x ∈ X : the map Aρ → X, f → φ(f ) x is bounded ,

which, trivially, must be constrained between

>
{0} ⊆ Dφ ⊆ Dom(T n ).
n∈Z

By definition, the vectors of Dφ satisfy the asymptotic bound

T n x 9 ρn , (n ∈ Z);

and, on addition, the Beurling condition

 log ρn
<∞
1 + n2
n∈Z

allows us to employ again the regularity of the Beurling algebra Aρ . Thus, for any
open cover {U, V } of σ (T ), the regularity of Aρ enables us to pick a function h ∈
Aρ such that

h ≡ 1 on (C \ V ) ∩ σ (T ) and h ≡ 0 on (C \ U ) ∩ σ (T ).

Now, given an arbitrary x ∈ Dφ , an adjustment of the previous estimations show

that the local spectra
 
σT φ(h) x ⊆ U and σT (I − φ(h)) x ⊆ V .
274 E. A. Gallardo-Gutiérrez and M. Monsalve-López

This clearly provides a local spectral decomposition of T with respect to the

elements in Dφ . We resume the above discussion in the following proposition:
Proposition 5.2 Let Aρ be a regular Beurling algebra, X a complex Banach space
and φ : Aρ → L(X) a algebra action with continuity core Dφ . Consider the
operator T := φ(eıt ) ∈ B(X). Then, for every closed F ⊆ C with int(F ) = ∅,
 
XT (F ) ⊇ φ(h)(Dφ ) : h ∈ Aρ with supp(h) ⊆ F .

Moreover, the inclusion

Dφ ⊆ XT (U1 ) + . . . + XT (Un ) (5.3)

holds for every open cover {U1 , . . . , Un } of C.

Some remarks are in order: firstly, note that Eq. (5.3) reveals us that Atzmon’s
Theorem entails a sort of weak decomposability for T with respect to the subman-
ifold Dφ . On the other hand, following the philosophy of [25], it is not hard to
extend Proposition 5.2 to a much wider class of functional calculi. Such extension,
performed using the language of Gelfand theory [20], allows us to consider arbitrary
commutative Banach algebras as the model of the operator T .
But, certainly, the reader may wisely argue that Atzmon’s Theorem granted the
existence of non-trivial closed hyperinvariant subspaces for T while Proposition 5.2
is still far from guaranteeing them. Actually, and exclusively when the continuity
core Dφ = {0}, Proposition 5.2 ensures the existence of non-zero local spectral
manifolds. So, maybe the next question must be: how long is Atzmon’s Theorem
from Proposition 5.2? The keypoint is that, when the preceding techniques can be
applied to both T and its adjoint T ∗ for two non-zero x ∈ X and y ∈ X∗ , we
manage to construct non-trivial closed invariant subspaces due to the duality relation
between local spectral manifolds [21, Prop. 2.5.1] given by

XT (F ) ⊆ ⊥ XT∗ ∗ (G),

valid whenever T has the SVEP and for every pair of disjoint closed sets F, G ⊆ C;
where, as usual, ⊥ N denotes the preannihilator of a subset in N ⊆ X∗ , i.e.
⊥
 
N := x ∈ X : ϕ(x) = 0 for every ϕ ∈ N .

As a consequence, we deduce the following local spectral version of Atzmon’s

Theorem:
Theorem 5.3 (Gallardo-Gutiérrez and Monsalve-López [20]) Let T ∈ B(X) be
an operator on a Banach space X such that σp (T ) = σp (T ∗ ) = ∅. Assume that
there exist non-zero x ∈ X and y ∈ X∗ for which
; n ; ; ;
;T x ; 9 ρn and ;(T ∗ )n y ; 9 ρn (n ∈ Z)
A Closer Look at Bishop Operators 275

for some Beurling sequence ρ := (ρn )n∈Z . Then, if σT (x)∪σT ∗ (y) is not a singleton,
for every open subset U ⊆ C such that U ∩ σT (x) = ∅ and σT ∗ (y) \ U = ∅, we
have

{0} = XT (U ) = X.

In particular, T has a non-trivial closed hyperinvariant subspace.

In the same way as for Proposition 5.2, one has that Theorem 5.3 stands as the
particular case regarding regular Beurling algebras of a much more general result
[20, Thm. 2.5], aimed to extend Atzmon’s Theorem to a wider class of models based
on arbitrary commutative Banach algebras.
Unsurprisingly, our motivation in Atzmon’s Theorem and its spectral implica-
tions are focused mainly on Bishop operators. This is the goal of the following
subsection: analyse carefully what is happening to σ (Tα ) and relate it, if possible,
to the invariant subspaces of Tα .

5.2 Local Spectral Decomposition of Bishop Operators

According to the exposition held in Sect. 4, the lack of any profitable local spectral
property seems to be an evident feature regarding Bishop operators. Nevertheless, as
it shall be discussed along the following lines with the aid of the results exposed in
the preceding subsection, this turns out to be just a kind of illusion (at least in some
cases) caused by a misguided choice of the local spectral manifolds. Of course, the
cornerstone in the development of such richer spectral theory for Bishop operators
shall be the aforementioned local spectral version of Atzmon’s Theorem.
Again, as we aim to apply Atzmon’s Theorem to Bishop operators, we recall that
the condition
 q 
j
log qj +1 = O , (5.4)
log3 qj

which shall play a prominent role in the sequel. By means of Theorem 3.4, we know
that given an irrational α ∈ (0, 1) satisfying (5.4), there exists a Beurling sequence
ρ = (ρn )n∈Z such that
; n ; ; ;
;
Tα 1Bα ; 9 ρn and ;(Tα∗ )n 1Bα ; 9 ρn .

In addition, by similarity, it may be checked that for every , m ∈ Z,

σTα (e2πıt 1Bα ) = e2πıα σTα (1Bα ),

σT∗ (e2πımt 1Bα ) = e−2πımα σT∗ (1Bα ).

α α
276 E. A. Gallardo-Gutiérrez and M. Monsalve-López

Thus, due to the irrationality of α, it is plain that provided any open subset U ⊆ C
such that U ∩T = T and T\U = ∅, we may find , m ∈ Z satisfying both conditions

U ∩ σTα (e2πıt 1Bα ) = ∅ and σTα∗ (e2πımt 1Bα ) \ U = ∅.

Therefore, as a corollary of Theorem 5.3, the above discussion leads us to the

following local spectral version of Theorem 3.4:
Theorem 5.4 (Gallardo-Gutiérrez and Monsalve-López [20]) Let α ∈ (0, 1) be
an irrational satisfying (5.4). Suppose that Tα acts on Lp [0, 1) for a fixed 1 ≤
p < ∞. Then, given any open subset U ⊆ C such that U ∩ ∂σ (Tα ) = ∅ and
∂σ (Tα ) \ U = ∅, we have

{0} = XTα (U ) = Lp [0, 1).

In particular, Tα has a non-trivial closed hyperinvariant subspace.

Although, in mere quantitative terms both Theorems 3.4 and 5.4 cover exactly
the same cases; clearly, the advantage of the latter result with respect to Theorem 3.4
is that it allows us to identify the spectral nature of the invariant subspaces involved.
In addition, an application of Proposition 5.2 in conjunction with some of the
arguments appearing in the Proof of Theorem 4.7, shows the following sort of weak
decomposability fulfilled by all Bishop operators:
Theorem 5.5 (Gallardo-Gutiérrez and Monsalve-López [20]) Let α ∈ (0, 1) be
an irrational number satisfying (5.4). Then, for every open cover {U1 , . . . , Un } of
∂σ (Tα ), the algebraic sum

XTα (U1 ) + . . . + XTα (Un )

is norm-dense in Lp [0, 1) for every 1 ≤ p < ∞.

Due to the absence of the Dunford property (C) for each Bishop operator Tα ,
the feature described within Theorem 5.5 cannot be properly considered a spectral
decomposition. Anyway, it clearly suggests that Bishop operators verify much more
interesting spectral properties than it seemed a priori and invites to study certain
weaker local spectral decompositions for Tα than the ones mentioned in Sect. 4.
Finally, we remind that in Theorem 3.6 we stated that, regarding Bishop operators
Tα , there exists a threshold limit on α from which Atzmon’s Theorem cannot be
applied anymore. In the language of Proposition 5.2, such result may be translated
saying that if α does not belong to
  q 
j
M := α ∈ (0, 1) : log qj +1 = O
log qj
A Closer Look at Bishop Operators 277

then, for every algebra action φ : Aρ → L(Lp [0, 1)) with φ(1) = I and φ(eınt ) =
αn from a regular Beurling algebra Aρ , the corresponding continuity core must be
T
Dφ = {0}.
On the other hand, it seems reasonable to expect that, for those cases Tα uncov-
ered by Atzmon’s Theorem (despite the absence of a profitable functional calculus),
their local spectral properties remain identical and rich enough to construct closed
invariant subspaces. In this sense, we suggest that one feasible route to solve the
invariant subspace problem for all Bishop operators could be understanding in depth
their local spectral manifolds XTα (U ) arising from open sets U intersecting the
boundary of σ (Tα ). To support our question in the positive, we recall that there exist
decomposable operators T ∈ B(X) whose spectral behaviour cannot be described
in terms of a functional calculus from a suitable algebra [1]. We pose it as an open
question:
Question 1 Let α ∈ (0, 1) be any irrational number and ε > 0 sufficiently small. Is
the local spectral manifold

XTα D(e−1 , ε)

non-trivial and non-dense in Lp [0, 1) for every 1 ≤ p < ∞?

6 Bishop-Type Operators

In order to complete the survey, our purpose throughout this last section is discussing
analogous results to the ones described above in a more general framework. In
particular, we shall show that many Bishop-type operators also admit a rich spectral
theory.
We begin by recalling that MacDonald in [22] was able to determine the
spectrum of many weighted translations operators. In order to state his result,
let us fix (, B, μ) a non-atomic probability space which arises from the Borel
sets of a compact metrizable space. Recall that any invertible measure-preserving
transformation τ :  →  is called ergodic for (, B, μ) if given any B ∈ B such
that τ −1 (B) = B then either μ(B) = 0 or μ(B) = 1 and, it is called uniquely
ergodic whenever it is continuous and μ is the unique -probability measure for
which τ is ergodic. As an important instance, observe that whenever  = [0, 1) is
endowed with the usual Lebesgue measure, having in mind the natural identification
with T = {e2πit : t ∈ [0, 1)}, the transformation τα (t) := {t + α} with α ∈ Q is
uniquely ergodic.
We are now in position to state MacDonald’s result. Recall that, given an arbitrary
φ ∈ L∞ [0, 1) and τ any measure-preserving transformation with measurable
inverse, the Bishop-type operator Tφ,τ is defined as

Tφ,τ f := φ · (f ◦ τ )

on Lp (, μ) for each 1 ≤ p ≤ ∞.

278 E. A. Gallardo-Gutiérrez and M. Monsalve-López

Theorem 6.1 (MacDonald [22]) Let (, B, μ, τ ) be a uniquely ergodic sys-

tem and φ ∈ L∞ (, μ) continuous μ-almost everywhere. Consider Tφ,τ ∈
L Lp (, μ) for 1 ≤ p < ∞ fixed. Then:
• If 0 ∈ ess ran(φ), the spectrum of Tφ,τ is
  
σ (Tφ,τ ) = λ ∈ C : |λ| ≤ exp log |φ| dμ .


• If 0 ∈ ess ran(φ), the spectrum of Tφ,τ is

  
σ (Tφ,τ ) = λ ∈ C : |λ| = exp log |φ| dμ .


Clearly the assumptions on φ and τ determine the spectrum of Tφ,τ , but the
general case is open and the following difficult question arises naturally:
Question 2 Determine the spectrum of weighted translation operators.
On the other hand, following Atzmon’s theorem approach, MacDonald was able
to establish the existence of non-trivial closed hyperinvariant subspaces for a large
class of Bishop-type operators Tφ,τα , for the sake of brevity denoted by Tφ,α (see
[22, Theorem 2.5, Theorem 2.6]). Consequently, in the spirit of Sect. 5, for those
cases covered by MacDonald, it is possible to obtain spectral decompositions similar
to those appearing in Theorems 5.4 and 5.5. We address that in the following
subsection.

6.1 Spectral Decomposition of Bishop-Type Operators

In order to state the results accurately, we must dwell on some minor technicalities
for a moment: given a positive real number M, we recall from [22] the class of step
functions
 
 

SM = S = rj χIj : rj ∈ R, Ij intervals and |rj | ≤ M .
j =1 j =1

In addition, let L denote the class of all real functions f ∈ L∞ [0, 1) for which there
exists γ > 0 and Kf > 0 (Kf depending exclusively on f ) such that

1
inf {f − S∞ : S ∈ SM } < Kf
M
A Closer Look at Bishop Operators 279

for each positive real M. Thus, one immediately has:

Theorem 6.2 (Gallardo-Gutiérrez and Monsalve-López [20]) Given any φ ∈
L∞ [0, 1) with log |φ| ∈ L and α ∈ (0, 1) a non-Liouville irrational number, let
Tφ,α be the induced Bishop-type operator on Lp [0, 1) for some 1 < p < ∞. Then,
given any open set U ⊂ C such that U ∩ ∂σ (Tφ,α ) = ∅ and ∂σ (Tφ,α ) \ U = ∅, we
have

{0} = XTφ,α (U ) = Lp [0, 1).

Theorem 6.3 (Gallardo-Gutiérrez, Monsalve-López [20]) Given any φ ∈

L∞ [0, 1) with log |φ| ∈ L and α ∈ (0, 1) a non-Liouville irrational number,
let Tφ,α be the induced Bishop-type operator on Lp [0, 1) for some 1 < p < ∞.
Then, for every open cover {U1 , . . . , Un } of ∂σ (Tφ,α ), the sum

XTφ,α (U1 ) + . . . + XTφ,α (Un ) (6.1)

is norm-dense in Lp [0, 1) for 1 < p < ∞.

Note that, by means of Colojoară and Foiaş decomposability theorem, somewhat
stronger conclusions hold for Theorem 6.3 whenever the operator Tφ,α is invertible.
In such a case, it happens that the operator Tφ,α turns out to be decomposable
and the algebraic sum (6.1) comprises the whole space Lp [0, 1). Moreover, further
refinements established as well by MacDonald [23], showed decomposability for
some Bishop-type operators Tφ,α whenever the condition

∞
 1 q 
j +1
log <∞
qj qj
j =0

is satisfied by the convergents (aj /qj )j ≥0 of the irrational α. Note that, as a matter
of fact, the latter condition embraces some Liouville numbers.
In view of the above discussion, as it was previously posed for Bishop operators,
it might happen that addressing of the invariant subspace problem for general
Bishop-type operators based on identifying some of their local spectral manifolds
was a fruitful approach. Similarly, we pose it as an open question:
Question 3 Let α ∈ (0, 1) be any irrational number and U ⊆ C a “sufficiently 
small” open set intersecting ∂σ (Tφ,α ). Is the local spectral manifold XTφ,α U non-
trivial and non-dense in Lp [0, 1) for every 1 ≤ p < ∞?

Acknowledgments Authors “Eva A. Gallardo-Gutiérrez and Miguel Monsalve-López” are par-

tially supported by Plan Nacional I+D grant nos. MTM2016-77710-P (Spain) and PID2019-
105979GB-I 00 (Spain) and by “Severo Ochoa Programme for Centres of Excellence in R&D”
(SEV-2015-0554). In addition, M. Monsalve-López also acknowledges support of the grant Ayudas
de la Universidad Complutense de Madrid para contratos predoctorales de personal investigador
en formación, ref. no. CT27/16.
280 E. A. Gallardo-Gutiérrez and M. Monsalve-López

References

1. E. Albrecht, On two questions of I. Colojoară and C. Foiaş. Manuscripta Math. 25, 1–15 (1978)
2. E. Albrecht, J. Eschmeier, Functional models and local spectral theory. Proc. London Math.
Soc. 75, 323–348 (1997)
3. A. Atzmon, Operators which are annihilated by analytic functions and invariant subspaces.
Acta Math. 144, 27–63 (1980)
4. A. Atzmon, On the existence of hyperinvariant subspaces. J. Operator Theor. 11, 4–40 (1984)
5. A. Atzmon, Power-regular operators. Trans. Amer. Math. Soc. 347, 3101–3109 (1995)
6. J.J. Bastian, Decomposition of weighted translation operators. Ph.D. Dissertation. Indiana
University, 1973
7. B. Beauzamy, Sous-espaces invariants de type fonctionnel dans les espaces de Banach. Acta
Math. 144, 65–82 (1980)
8. A. Beurling, Sur les intégrales de Fourier absolument convergentes et leur application à une
transformation fonctionelle, in Ninth Scandinavian Mathematical Congress (1938), pp. 345–
366
9. D.P. Blecher, A.M. Davie, Invariant subspaces for an operator on L2 () composed of a
multiplication and a translation. J. Operator Theory. 23, 115–123 (1990)
10. S.W. Brown, Hyponormal operators with thick spectra have invariant subspaces. Ann. Math.
125, 93–103 (1987)
11. Y. Bugeaud, Approximation by Algebraic Numbers (Cambridge University Press, Cambridge,
2004)
12. F. Chamizo, E.A. Gallardo-Gutiérrez, M. Monsalve-López, A. Ubis, Invariant subspaces for
Bishop operators and beyond. Adv. Math. 375, 107365 (2020)
13. I. Colojoară, C. Foiaş, Theory of Generalized Spectral Operators (CRC Press, Boca Raton,
1968)
14. J. Daneš, On local spectral radius. C̆asopis Pes̆t. Mat. 112, 177–187 (1987)
15. A.M. Davie, Invariant subspaces for Bishop’s operators. Bull. London Math. Soc. 6, 343–348
(1974)
16. J. Eschmeier, B. Prunaru, Invariant subspaces for operators with property (β) and thick
spectrum. J. Funct. Anal. 94, 196–222 (1990)
17. J. Eschmeier, B. Prunaru, Invariant subspaces and localizable spectrum. Integr. Equ. Oper.
Theory 42, 461–471 (2002)
18. A. Flattot, Hyperinvariant subspaces for Bishop-type operators. Acta. Sci. Math. 74, 689–718
(2008)
19. E. Gallardo-Gutiérrez, M. Monsalve-López, Power-regular Bishop operators and spectral
decompositions. J. Operator Theory (in press). https://doi.org/10.7900/jot.2019sep21.2256
20. E. Gallardo-Gutiérrez, M. Monsalve-López, Spectral decompositions arising from Atzmon’s
hyperinvariant subspace theorem. (under review)
21. K. Laursen, M. Neumann, An Introduction to Local Spectral Theory (Clarendon Press, Oxford,
2000)
22. G.W. MacDonald, Invariant subspaces for Bishop-type operators. J. Funct. Anal. 91, 287–311
(1990)
23. G.W. MacDonald, Decomposable weighted rotations on the unit circle. J. Operator Theory 35,
205–221 (1996)
24. V. Müller, Local spectral radius formula for operators in Banach spaces. Czechoslovak Math.
J. 38, 726–729 (1988)
25. M. Neumann, Banach algebras, decomposable convolution operators, and a spectral mapping
property, in Function Spaces. Marcel Dekker Series in Pure and Applied Mathematics, vol.
136 (Dekker, New York, 1992), pp. 307–323
26. S.K. Parrott, Weighted translation operators. ProQuest LLC, Ann Arbor, MI. Thesis (Ph.D)-
University of Michigan, 1965
A Closer Look at Bishop Operators 281

27. K. Petersen, The spectrum and commutant of a certain weighted translation operator. Math.
Scand. 37, 297–306 (1975)
28. W. Rudin, Real and Complex Analysis (Tata McGraw-Hill Education, New York, 2006)
29. J. Wermer, The existence of invariant subspaces. Duke Math. J. 19, 615–622 (1952)
Products of Unbounded Bloch Functions

Daniel Girela

Abstract We give new constructions of pair of functions (f, g), analytic in the unit
disc, with g ∈ H ∞ and f an unbounded Bloch function, such that the product g · f
is not a Bloch function.

Keywords Bloch function · Normal function · Blaschke product · Inner

function · Minimal Besov space · Analytic mean Lipschitz spaces

Mathematics Subject Classification (2010) Primary 30D45; Secondary 30H30

1 Introduction and Statements of the Results

Let D = {z ∈ C : |z| < 1} denote the open unit disc in the complex plane C. The
space of all analytic functions in D will be denoted by Hol(D).
For 0 < p ≤ ∞, the classical Hardy space H p is defined as the set of all
f ∈ Hol(D) for which

def
f H p = sup Mp (r, f ) < ∞,
0<r<1

where, for 0 < r < 1 and f ∈ Hol(D),

  2π 1/p
1
Mp (r, f ) = |f (re )| dθ
iθ p
, (0 < p < ∞);
2π 0

M∞ (r, f ) = sup |f (reiθ )|.

θ∈R

D. Girela ()
Análisis Matemático, Facultad de Ciencias, Universidad de Málaga, Málaga, Spain
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 283

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_14
284 D. Girela

We mention [7] as a general reference for the theory of Hardy spaces.

A function f ∈ Hol(D) is said to be a Bloch function if

def
f B = |f (0)| + sup (1 − |z|2 )|f (z)| < ∞.
z∈D

The space of all Bloch functions is denoted by B, it is a Banach space with the just
defined norm  · B . It is well known that

H ∞  B.

A typical example of an unbounded Bloch function is the function f defined by

1
f (z) = log , z ∈ D.
1−z

We mention [1] as a general reference for the theory of Bloch functions.

A function f which is meromorphic in D is said to be a normal function in the
sense of Lehto and Virtanen [15] if

|f (z)|
sup(1 − |z|2 ) < ∞.
z∈D 1 + |f (z)|2

For simplicity, we shall let N denote the set of all holomorphic normal functions in
D. It is clear that any Bloch function is a normal function, that is, we have B ⊂ N .
We refer to [1, 15] and [16] for the theory of normal functions. In particular, we
remark here that if f ∈ N , ξ ∈ ∂D and f has the asymptotic value L at ξ , (that is,
there exists a curve γ in D ending at ξ such that f (z) → L, as z → ξ along γ ) then
f has the non-tangential limit L at ξ .
Let us recall that if a sequence of points {an } in the unit disc satisfies the Blaschke
condition:
∞

(1 − |an |) < ∞,
n=1

the corresponding Blaschke product B is defined as

∞
/ |an | an − z
B(z) = .
an 1 − an z
n=1

Such a product is analytic in D. In fact, it is an inner function, that is, an H ∞ -

function with radial limit of absolute value 1 at almost every point of ∂D (cf. [7,
Chapter 2]).
Products of Unbounded Bloch Functions 285

If {an } is a Blaschke sequence and there exists δ > 0 such that

/ an − am
≥ δ, for all n,
1 − a n am
m=n

we say that the sequence {an } is uniformly separated and that B is an interpolating
Blaschke product. Equivalently,

B is an interpolating Blaschke product ⇔ inf (1 − |an |2 )|B (an )| > 0.

n≥1

We refer to [7, Chapter 9] and [9, Chapter VII] for the basic properties of inter-
polating Blaschke products. In particular, we recall that an exponential sequence is
uniformly separated and that the converse holds if all the ak ’s are positive.
Lappan [14, Theorem 3] proved that if B is an interpolating Blaschke product
and f is a normal analytic function in D, the product B · f need not be normal.
Lappan used this to show that N is not a vector space.
Lappan’s result is a consequence of the following easy fact: if B is an interpolat-
ing Blaschke product whose sequence of zeros is {an } and G is an analytic function
in D with G(an ) → ∞, then f = B · G is not a normal function (and hence it
is not a Bloch function either). This result has been used by several authors (see
[3, 5, 10, 11, 17, 18]) to construct distinct classes of non-normal functions.
The author and Suárez proved in [12] a result of this kind dealing with Blaschke
products with zeros in a Stolz angle but not necessarily interpolating, improving a
result of [13]. Namely, Theorem 1 of [12] is the following.
Theorem A Let B be an infinite Blaschke product whose sequence of zeros {an } is
contained in a Stolz angle with vertex at 1 and let G be analytic in D with G(z) →
∞, as z → 1. Then the function f = B · G is not a normal function.
It is natural to ask whether it is possible to prove results similar to those
described, substituting “Blaschke products” by some other classes of H ∞ -
functions. Our first result in this paper deals with the atomic singular inner function.
Theorem 1.1 Let S be the atomic singular inner function defined by
 
1+z
S(z) = exp − , z ∈ D, (1.1)
1−z

and let f be a Bloch function with

lim |f (z)| = ∞.
z→1

Then the function F defined by F (z) = S(z)f (z) is not a normal function (hence, it
is not a Bloch function).
286 D. Girela

In particular, the function F defined by F (z) = S(z) · log 1−z

1
(z ∈ D) is not
normal.
A Bloch function f satisfies
 
1
M∞ (r, f ) = O log
1−r

and, consequently,
 
1+r
|f (r)| = o exp , as r → 1− .
1−r

Thus, Theorem 1.1 follows from the following result.

Theorem 1.2 Let S be the singular inner function defined by (1.1) and let f be an
analytic function in D satisfying:
(i) limz→1 |f (z)|
 = ∞. 
(ii) |f (r)| = o exp 1+r −
1−r , as r → 1 .

Then the function F defined by F (z) = S(z)f (z) is not a normal function (hence, it
is not a Bloch function).
For g ∈ Hol(D), the multiplication operator Mg is defined by

def
Mg (f )(z) = g(z)f (z), f ∈ Hol(D), z ∈ D.

Let us recall that if X and Y are two spaces of analytic function in D and g ∈ Hol(D)
then g is said to be a multiplier from X to Y if Mg (X) ⊂ Y . The space of all
multipliers from X to Y will be denoted by M(X, Y ) and M(X) will stand for
M(X, X). Brown and Shields [4] characterized the space of multipliers of the Bloch
space M(B) as follows.
Theorem B A function g ∈ Hol(D) is a multiplier of the Bloch space if and only
if g ∈ H ∞ ∩ Blog , where Blog is the Banach space of those functions f ∈ Hol(D)
which satisfy
 
def 2
f Blog = |f (0)| + sup(1 − |z| ) log
2
|f (z)| < ∞.
z∈D 1 − |z|2

Thus, if g ∈ H ∞ \ Blog there exists a function f ∈ B \ H ∞ such that g · f ∈ / B.

It is easy to see that the analytic Lipschitz spaces $α (0 < α ≤ 1) and the mean
p
Lipschitz spaces $α (1 < p < ∞, 1/p < α ≤ 1) are contained in M(B), we refer
to [7, Chapter 5] for the definitions of these spaces. Let us simply recall here that

$11 = {f ∈ Hol(D) : f ∈ H 1 }.
Products of Unbounded Bloch Functions 287

On the other hand, Theorem 1 of [8] shows the existence of a Jordan domain  with
rectifiable boundary and 0 ∈ , and such that the conformal mapping g from D
onto  with g(0) = 0 and g (0) > 0 does not belong to Blog. For this function g we
have that g ∈ $11 but g is not a multiplier of B. Thus we have:

$11 ⊂ M(B).

In view of this and the results involving Blaschke products that we have
mentioned above, it is natural to ask the following question:
Question 1.3 Is it true that for any given f ∈ B \H ∞ there exists a function g ∈ $11
such that g · f ∈
/ B?
We shall show that the answer to this question is affirmative. Actually we shall
prove a stronger result.
We let B 1 denote the minimal Besov space which consists of those functions
f ∈ Hol(D) such that

|f (z)| dA(z) < ∞.
D

Here dA denotes the area measure on D. Alternatively, the space B 1 can be

characterized as follows (see [2]):
For f ∈ Hol(D), we have that f ∈ B 1 if and only there exist a sequence of
points {ak }∞ ∞
k=1 ⊂ D and a sequence {λk }k=0 ∈  such that
1

∞

f (z) = λ0 + λk ϕak (z), z ∈ D.
k=1

Here, for a ∈ D, ϕa : D → D denotes the Möbius transformation defined by

a−z
ϕa (z) = , z ∈ D.
1 − az

It is well known that B 1 ⊂ $11 (see [2, 6]) and then our next result implies that the
answer to question 1.3 is affirmative.
Theorem 1.4 If f ∈ B \ H ∞ then there exists g ∈ B 1 such that g · f ∈
/ B.
The proofs of Theorems 1.2 and 1.4 will be presented in Sect. 2. We close
this section noticing that throughout the paper we shall be using the convention
that C = C(p, α, q, β, . . . ) will denote a positive constant which depends only
upon the displayed parameters p, α, q, β, . . . (which often will be omitted) but
not necessarily the same at different occurrences. Moreover, for two real-valued
functions E1 , E2 we write E1  E2 , or E1  E2 , if there exists a positive constant
C independent of the arguments such that E1 ≤ CE2 , respectively E1 ≥ CE2 . If
288 D. Girela

we have E1  E2 and E1  E2 simultaneously then we say that E1 and E2 are

equivalent and we write E1 $ E2 .

2 The Proofs

2.1 Proof of Theorem 1.2

For 0 < a < 1, set a = {z ∈ D : |z − a| = 1 − a}. If z ∈ a then

1+z a
Re =
1−z 1−a

and, hence,
 
−a
|S(z)| = exp , z ∈ a .
1−a

This, together with (i), implies that

F (z) → ∞, as z → 1 along a .

Hence F has the asymptotic value ∞ at 1. On the other hand, (ii) implies that F has
the radial limit 0 at 1. Then it follows that F is not normal.

2.2 Proof of Theorem 1.4

Take f ∈ B \ H ∞ . Set

ϕ(r) = M∞ (r, f ), 0 < r < 1.

Clearly, ϕ(r) → ∞, as r → 1 and it is well known that

 
1
ϕ(r) = O log .
1−r

This implies that

(1 − r)2 ϕ(r) → 0, as r → 1. (2.1)

Products of Unbounded Bloch Functions 289

Choose a sequence of numbers {rn } ⊂ (0, 1) satisfying the following properties:

(i) {rn } is increasing.  2
1 − rn−1
(ii) (1 − rn ) ϕ(rn ) = o
2
, as n → ∞.
n
(iii) ϕ(rn ) ≥ 2ϕ(rn−1 ), for all n.
1 − rn+1
(iv) → 0, as n → ∞.
1 − rn
The existence of such a sequence is clear, bearing in mind (2.1) and the fact that
ϕ(r) → ∞, as r → 1.
Set

λk = ϕ(rk )−1/2 , k = 1, 2, . . . .

For each k, take ak ∈ D with |ak | = rk and |f (ak )| = ϕ(rk ). Using (iii), it follows
that
∞

λk < ∞. (2.2)
k=1

Define
∞

g(z) = λk ϕak (z), z ∈ D. (2.3)
k=1

Using (2.2) we see that the sum in (2.3) defines an analytic function in D which
belongs to B 1 . Set

F (z) = g(z)f (z), z ∈ D.

Since g ∈ H ∞ and f ∈ B we see that

 
1
|g(an )f (an )| = O . (2.4)
1 − |an |

On the other hand,

|g (an )f (an )|  I − I I − I I I, (2.5)

290 D. Girela

where

I = |f (an )|λn |ϕan (an )|,


n−1
1 − |ak |2
I I  |f (an )| λk ,
|1 − ak an |2
k=1
∞
 1 − |ak |2
I I I  |f (an )| λk .
|1 − ak an |2
k=n+1

Clearly,

ϕ(rn )1/2
I = |f (an )|λn |ϕan (an )| $ . (2.6)
1 − rn

Using the definitions, the facts that ϕ and the sequence {rn } are increasing, and (ii),
we obtain


n−1
1 − |ak |2
I I  |f (an )| λk
|1 − ak an |2
k=1


n−1
1 − |ak |
 ϕ(rn ) ϕ(rk )−1/2
[(1 − |ak |) + (1 − |an )]2
k=1


n−1
1
 ϕ(rn )
ϕ(rk )1/2 (1 − rk )
k=1
nϕ(rn )

1 − rn−1
ϕ(rn )1/2 n(1 − rn )
= ϕ(rn )1/2
1 − rn 1 − rn−1
 1/2 
ϕ(rn )
=o . (2.7)
1 − rn

Likewise, using the definitions, the facts that ϕ and the sequence {rn } are increasing,
(iii), and (iv), we obtain
∞
 ϕ(rk )−1/2 (1 − rk )
I I I  ϕ(rn )
[(1 − rk ) + (1 − rn )]2
k=n+1
∞
 1 − rk
 ϕ(rn ) ϕ(rk )−1/2
(1 − rn )2
k=n+1
Products of Unbounded Bloch Functions 291

∞
1 − rn+1 
 ϕ(rn ) ϕ(rk )−1/2
(1 − rn )2
k=n+1

ϕ(rn )1/2 1 − rn+1

 ·
1 − rn 1 − rn
 
ϕ(rn )1/2
=o . (2.8)
1 − rn

Using (2.5)–(2.8), and the fact that lim ϕ(rn ) = ∞, we deduce that

(1 − |an |)|g (an )f (an )| → ∞, as n → ∞.

This and (2.4) imply that F is not a Bloch function.

Acknowledgments I wish to thank the referees for their careful reading of the article and for their
suggestions to improve it.
This research is supported in part by a grant from “El Ministerio de Economía y Competitivi-
dad”, Spain (PGC2018-096166-B-I00) and by grants from la Junta de Andalucía (FQM-210 and
UMA18-FEDERJA-002).

References

1. J.M. Anderson, J. Clunie, Ch. Pommerenke, On Bloch functions and normal functions. J. Reine
Angew. Math. 270, 12–37 (1974)
2. J. Arazy, S.D. Fisher, J. Peetre, Möbius invariant function spaces. J. Reine Angew. Math. 363,
110–145 (1985)
3. O. Blasco, D. Girela, M.A. Márquez, Mean growth of the derivative of analytic functions,
bounded mean oscillation, and normal functions. Indiana Univ. Math. J. 47, 893–912 (1998)
4. L. Brown, A.L. Shields, Multipliers and cyclic vectors in the Bloch space. Michigan Math. J.
38, 141–146 (1991)
5. D.M. Campbell, Nonnormal sums and products of unbounded normal functions. II. Proc. Amer.
Math. Soc. 74, 202–203 (1979)
6. J.J. Donaire, D. Girela, D. Vukotić, On univalent functions in some Möbius invariant spaces. J.
Reine Angew. Math. 553, 43–72 (2002)
7. P.L. Duren, Theory of H p Spaces (Academic, New York, 1970; Reprint: Dover, Mineola-New
York, 2000)
8. P. Galanopoulos, D. Girela, R. Hernández, Univalent functions, VMOA and related spaces. J.
Geom. Anal. 21, 665–682 (2011)
9. J.B. Garnett, Bounded Analytic Functions (Academic, New York, 1981)
10. D. Girela, On a theorem of Privalov and normal funcions. Proc. Amer. Math. Soc. 125, 433–
442 (1997)
11. D. Girela, Mean Lipschitz spaces and bounded mean oscillation. Illinois J. Math. 41, 214–230
(1997)
12. D. Girela, D. Suárez, On Blaschke products, Bloch functions and normal functions. Rev. Mat.
Complut. 24, 49–57 (2011)
13. D. Girela, C. González, J.A. Peláez, Multiplication and division by inner functions in the space
of Bloch functions. Proc. Amer. Math. Soc. 134, 1309–1314 (2006)
292 D. Girela

14. P. Lappan, Non-normal sums and products of unbounded normal function. Michigan Math. J.
8, 187–192 (1961)
15. O. Lehto, K.I. Virtanen, Boundary behaviour and normal meromorphic functions. Acta Math.
97, 47–65 (1957)
16. Ch. Pommerenke, Univalent Functions (Vandenhoeck und Ruprecht, Göttingen, 1975)
17. S. Yamashita, A nonnormal function whose derivative has finite area integral of order 0 < p <
2. Ann. Acad. Sci. Fenn. Ser. A I Math. 4(2), 293–298 (1979)
18. S. Yamashita, A nonnormal function whose derivative is of Hardy class H p , 0 < p < 1.
Canad. Math. Bull. 23, 499–500 (1980)
Birkhoff–James Orthogonality
and Applications: A Survey

Priyanka Grover and Sushil Singla

Abstract In the last few decades, the concept of Birkhoff–James orthogonality has
been used in several applications. In this survey article, the results known on the nec-
essary and sufficient conditions for Birkhoff–James orthogonality in certain Banach
spaces are mentioned. Their applications in studying the geometry of normed
spaces are given. The connections between this concept of orthogonality, and the
Gateaux derivative and the subdifferential set of the norm function are provided.
Several interesting distance formulas can be obtained using the characterizations
of Birkhoff–James orthogonality, which are also mentioned. In the end, some new
results are obtained.

Keywords Orthogonality · Tangent hyperplane · Smooth point · Faces of unit

ball · Gateaux differentiability · Subdifferential set · State on a C ∗ -algebra ·
Cyclic representation · Norm-parallelism · Conditional expectation

Mathematics Subject Classification (2010) Primary 15A60, 41A50, 46B20,

46L05, 46L08; Secondary 46G05, 47B47

1 Introduction

Let (V , ·) be a normed space over the field R or C. For normed spaces V1 , V2 , let
B(V1 , V2 ) denotes the space of bounded linear operators from V1 to V2 endowed
with the operator norm, and let B(V ) denotes B(V , V ). Let K(V1 , V2 ) denotes the
space of compact operators from V1 to V2 . Let H be a Hilbert space over R or C.
If the underlying field is C, the inner product on H is taken to be linear in the first
coordinate and conjugate linear in the second coordinate. The notations Mn (R) and
Mn (C) stand for n × n real and complex matrices, respectively.

P. Grover () · S. Singla

Department of Mathematics, Shiv Nadar University, Dadri, Uttar Pradesh, India
e-mail: [email protected]; [email protected]

© Springer Nature Switzerland AG 2021 293

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_15
294 P. Grover and S. Singla

Normed spaces provide a natural setting for studying geometry in the context of
vector spaces. While inner product spaces capture the concept of the measure of an
angle, orthogonality of two vectors can be described without knowing the notion of
measure of angle. For example, a vector v is orthogonal to another vector u in Rn if
and only if there exists a rigid motion T fixing the origin such that the union of rays
→ −−−→ −−−→
− →
−
0u, 0T (v), 0T (u) minus the open ray 0v is the one dimensional subspace generated
by u. This description of orthogonality by just using the notion of distance in Rn
1

motivates to try and define orthogonality in normed spaces. In this approach, one can
use the intuition about orthogonality in Rn to guess the results in general normed
spaces and then prove them algebraically. This has been done in [3, 16, 44, 78].
One of the definitions for orthogonality in a normed space suggested by Roberts
[78], known as Roberts orthogonality, is defined as follows: elements u and v are
said to be (Roberts) orthogonal if v + tu = v − tu for all scalars t. In [44,
Example 2.1], it was shown that this definition has a disadvantage that there exist
normed spaces in which two elements are Roberts orthogonal implies that one of the
element has to be zero. In [44], two more inequivalent definitions of orthogonality
in normed spaces were introduced. One of them is isosceles orthogonality which
says that v is isosceles orthogonal to u if v + u = v − u. The other one
is called Pythagorean orthogonality, that is, v is Pythagorean orthogonal to u if
v2 + u2 = v − u2 . Note that if V is an inner product space, all the above
mentioned definitions are equivalent to the usual orthogonality in an inner product
space. Isosceles and Pythagorean orthogonalities have geometric intuitions for the
corresponding definitions. In Rn , two vectors are isosceles perpendicular if and only
if their sum and difference can be sides of an isosceles triangle, and two vectors are
Pythagorean perpendicular if there is a right triangle having the two vectors as legs.
In [44], it was also proved that if u and v are two elements of a normed space,
then there exist scalars a and b such that v is isosceles orthogonal to av + u (see
[44, Theorem 4.4]) and v is Pythagorean orthogonal to bv + u (see [44, Theorem
5.1]). So these definitions don’t have the above mentioned weakness of Roberts
orthogonality.
In an inner product space, the following properties are satisfied by orthogonality.
Let u, u1 , u2 , v, v1 , v2 ∈ V .
1. Symmetry: If v⊥u, then u⊥v.
2. Homogeneity: If v⊥u, then av⊥bu for all scalars a and b.
3. Right additivity: If v⊥u1 and v⊥u2 , then v⊥(u1 + u2 ).
4. Left additivity: If v1 ⊥u and v2 ⊥u, then (v1 + v2 )⊥u.
5. There exists a scalar a such that v⊥av + u. (In Rn , this corresponds to saying
that any plane containing a vector v contains a vector perpendicular to v.)
It is a natural question to study the above properties for any given definition
of orthogonality. All the above definitions clearly satisfy symmetry. James [44,
Theorem 4.7, Theorem 4.8, Theorem 5.2, Theorem 5.3] proved that if isosceles

1 We learnt this characterization of orthogonality in Rn from Amber Habib.

Birkhoff–James Orthogonality and Applications: A Survey 295

or Phythagorean orthogonality satisfy homogeneity or (left or right) additivity, then

V has to be an inner product space. These orthogonalities have been extensively
studied in [3, 44, 78].
In [16], Birkhoff defined a concept of orthogonality, of which several properties
were studied by James in [45]. An element v is said to be Birkhoff–James orthogonal
to u if v ≤ v + ku for all scalars k. The analogy in Rn is that if two lines L1
and L2 intersect at p, then L1 ⊥ L2 if and only if the distance from a point of L2
to a given point q of L1 is never less than the distance from p to q. This definition
clearly satisfies the homogeneity property. In [45, Corollary 2.2], it was shown that
this definition also satisfies (4). But it lacks symmetry, for example, in (R2 , ·max ),
where (t1 , t2 )max = max{|t1 |, |t2 |}, take v = (1, 1) and u = (1, 0) (v ⊥ u but
u ⊥ v). It is not right additive, for example, in (R2 ,  · max ), take v = (1, 1),
u1 = (1, 0) and u2 = (0, 1). It is also not left additive, for example, in (R2 , ·max ),
take v1 = (1, 1), v2 = (0, −1) and u = (1, 0).
Let W be a subspace of V . Then an element v ∈ V is said to be Birkhoff–James
orthogonal to W if v is Birkhoff–James orthogonal to w for all w ∈ W . A closely
related concept is that of a best approximation to a point in a subspace. An element
w0 ∈ W is said to be a best approximation to v in W if v − w0  ≤ v − w
for all w ∈ W . Note that w0 is a best approximation to v in W if and only if
v − w0 is Birkhoff–James orthogonal to W . These are also equivalent to saying
that dist(v, W ) := inf{v − w : w ∈ W } = v − w0 . So v is Birkhoff–James
orthogonal to W if and only if dist(v, W ) is attained at 0. Therefore the study of
these concepts go hand in hand (see [89]). This is one of the reasons that this
definition of orthogonality, even though not symmetric, is still being extensively
studied in literature. Henceforward, orthogonality will stand for Birkhoff–James
orthogonality.
Recently, a lot of work has been done in the form of applications of this concept
of orthogonality and the main goal of this survey article is to bring all the related
work under one roof. In Sect. 2, we mention the connections between orthogonality
and geometry of normed spaces. We also deal with the question as to when the
orthogonality is symmetric or (left or right) additive. This leads us to the study
of various related notions like characterizations of smooth points and extreme
points, subdifferential set, ϕ-Gateaux derivatives etc. In Sect. 3, characterizations of
orthogonality in various Banach spaces are discussed along with some applications.
In Sect. 4, these characterizations are used to obtain distance formulas in some
Banach spaces. Some of the stated results are new and will appear in more detail
in [35]. Theorems 4.5, 4.6 and 4.7 are the new results given with proofs only here.

2 Orthogonality and Geometry of Normed Spaces

A hyperplane is a closed subspace of codimension one. A connection between the

concept of orthogonality and hyperplanes is given in the next theorem. An element
v is orthogonal to a subspace W if and only if there exists a linear functional f on
296 P. Grover and S. Singla

V such that f  = 1, f (w) = 0 for all w ∈ W and f (v) = v (see [89, Theorem
1.1, Ch. I]). This is equivalent to the following.
Theorem 2.1 ([45, Theorem 2.1]) Let W be a subspace of V . Let v ∈ V . Then v
is orthogonal to W if and only if there is a hyperplane H with v orthogonal to H
and W ⊆ H .
By the Hahn–Banach theorem and Theorem 2.1, it is easy to see that any element
of a normed space is orthogonal to some hyperplane (see [45, Theorem 2.2]). The
relation between orthogonality and hyperplanes is much deeper. We first recall some
definitions. For v ∈ V , we say S ⊆ V supports the closed ball D[v, r] := {x ∈ V :
x − v ≤ r} if dist(S, D[v, r]) = 0 and S ∩ Int D[v, r] = ∅. This is also
equivalent to saying that dist(v, S) = r (see [89, Lemma 1.3, Ch. I]). Let v0 be an
element of the boundary of D[v, r]. A hyperplane H is called a support hyperplane
to D[v, r] at v0 if H passes through v0 and supports D[v, r], and it is called a
tangent hyperplane to D[v, r] at v0 if H is the only support hyperplane to D[v, r]
at v0 . A real hyperplane is a hyperplane in V , when V is considered as a real normed
space.
Theorem 2.2 ([89, Theorem 1.2, Ch. I]) Let W be a subspace of V . Let v ∈ V .
Then v is orthogonal to W if and only if there exists a support hyperplane to D[v, r]
at 0 passing through W if and only if there exists a real hyperplane which supports
the closed ball D[v, v] at 0 and passes through W .
A direct consequence follows. If W is a non-trivial subspace of V , then 0 is the
unique best approximation of v in W if and only if there exists a tangent hyperplane
to D[v, r] at 0 passing through W (see [89, Corollary 1.5, Ch. I]).
The above results are related to the questions as to when the orthogonality is (left
or right) additive or symmetric. It was shown in [45, Theorem 5.1] that orthogonality
is right additive in V if and only if for any unit vector v ∈ V , there is a tangent
hyperplane to D[v, v] at 0. There are other interesting characterizations for
(left or right) additivity of orthogonality. To state them, some more definitions are
required. A normed space V is called a strictly convex space if given any v1 , v2 ∈ V ,
whenever v1 +v2  = v1 + v2  and v2 = 0, then there exists a scalar k such that
v1 = kv2 . This is also equivalent to saying that if v1  = v2  = 1 and v1 = v2 ,
then v1 + v2  < 2. The norm · is said to be Gateaux differentiable at v if

v + hu − v
lim
h→0 h

exists for all u ∈ V .

Now we have the following characterizations for the orthogonality to be right
additive in V .
Theorem 2.3 The following statements are equivalent.
1. Orthogonality is right additive.
2. Norm is Gateaux differentiable at each nonzero point.
Birkhoff–James Orthogonality and Applications: A Survey 297

3. For v ∈ V , there exists a unique functional f of norm one on V such that f (v) =
v.
4. For v ∈ V , there is a tangent hyperplane to D[v, v] at 0.
If V is a reflexive space, then the above are also equivalent to the following
statements.
(5) Any bounded linear functional on a given subspace of V has a unique norm
preserving Hahn–Banach extension on V .
(6) The dual space V ∗ is strictly convex.
Proof Equivalence of (1) and (2) is proved in [45, Theorem 4.2] and equivalence
of (1), (3) and (4) is proved in [45, Theorem 5.1]. For a reflexive space, equivalence
of (1) and (5) is given in [45, Theorem 5.7]. Equivalence of (5) and (6) is a routine
exercise in functional analysis.
Characterization of inner product spaces of dimension three or more can be given in
terms of (left or right) additivity or symmetry of orthogonality. Birkhoff [16] gave a
necessary and sufficient condition for a normed space of dimension at least three to
be an inner product space, and examples to justify the restriction on the dimension.
James [45, Theorem 6.1] showed that a normed space of dimension at least three is
an inner product space if and only if orthogonality is right additive and symmetric
if and only if the normed space is strictly convex and orthogonality is symmetric.
Later, James improved his result and proved a much stronger theorem.
Theorem 2.4 ([46, Theorem 1, Theorem 2]) Let V be a normed space of dimen-
sion at least three. Then V is an inner product space if and only if orthogonality is
symmetric or left additive.
A characterization of orthogonality to be symmetric or left additive in a normed
space of dimension two can be found in [2]. Several other necessary and sufficient
conditions for a normed space to be an inner product space are given in [2, 44]. This
problem has also been extensively studied in [3, 89].
An element v is called a smooth point of D[0, v] if there exists a hyperplane
tangent to D[v, v] at 0. We say v is a smooth point if it is a smooth point
of D[0, v]. Equivalently, v is a smooth point if there exists a unique affine
hyperplane passing through v which supports D[0, v] at v (such an affine
hyperplane is called the affine hyperplane tangent to D[0, v] at v). A normed
space is called smooth if all its vectors are smooth points. By Theorem 2.3, we get
that orthogonality in a normed space is right additive if and only if the normed space
is smooth. We also have that v is a smooth point if and only if the norm function is
Gateaux differentiable at v:
Theorem 2.5 Let v ∈ V . The norm function is Gateaux differentiable at v if and
only if there is a unique f ∈ V ∗ such that f  = 1 and f (v) = v. In this
case, the Gateaux derivative of the norm at v is given by Re f (u) for all u ∈ V . In
addition, for u ∈ V , we have that v is orthogonal to u if and only if f (u) = 0.
298 P. Grover and S. Singla

Smooth points and this connection with Gateaux differentiability was studied in
[1, 22, 52, 53] and many interesting results can be obtained as their applications. Let
the space of continuous functions on a compact Hausdorff space X be denoted by
C(X) and let the space of bounded continuous functions on a normal space  be
denoted by Cb (). Kečkić [53, Corollary 2.2, Corollary 3.2] gave characterizations
of smooth points in C(X) and Cb (). A characterization of smooth points in B(H)
was given in [52, Corollary 3.3]. For H separable, Abatzoglou [1, Corollary 3.1]
showed that the operators in B(H) of unit norm which are also smooth points are
dense in the unit sphere of B(H). In K(H), this result was first proved by Holub [42,
Corollary 3.4]. Heinrich [40, Corollary 2.3] generalized this result for K(V1 , V2 ),
where V1 is a separable reflexive Banach space and V2 is any normed space. He
proved that the operators which attain their norm at a unique unit vector (upto scalar
multiplication) are dense in K(V1 , V2 ).
In this paragraph, H is a separable Hilbert space. Schatten [86] proved that
D[0, 1] in K(H) has no extreme points. In [42], the geometry of K(H) and its
dual B1 (H), the trace class, was studied by characterizing the smooth points and
extreme points of their closed unit balls. It was shown in [42, Corollary 3.1] that
the trace class operators of rank one and unit norm are exactly the extreme points of
D[0, 1] in B1 (H). The space B1 (H) is predual of B(H) and hence is isometrically
isomorphic to a subspace of B(H)∗ . An interesting result in [1, Corollary 3.3] is
that all the trace class operators of rank one and unit norm are also extreme points
of D[0, 1] in B(H)∗ . In [40], this study was continued to understand the geometry
of K(V1 , V2 ), B(V1 , V2 ) and the weak tensor product of V1 and V2 , where V1 and
V2 are Banach spaces. Characterizations of Gateaux differentiability and Fréchet
differentiability of the norm at an operator T in these spaces were obtained. For
Schatten classes of H, this problem was addressed in [1, Theorem 2.2, Theorem
2.3]. In [1, Theorem 3.1], another characterization of Fréchet differentiability of
the norm at T in B(H) was given, an alternative proof of which can be found in
[72, Theorem 4.6]. In [40, Corollary 2.2], a necessary and sufficient condition for
0 = T ∈ K(V1 , V2 ) to be a smooth point is obtained, where V1 is a reflexive Banach
space and V2 is any Banach space. It is shown that such a T is a smooth point if and
only if T attains its norm on the unique unit vector x0 (up to a scalar factor) and
T x0 is a smooth point. (This was proved for K(H) in [42, Theorem 3.3].) Recently,
as an application of orthogonality, it was shown in [72, Theorem 4.1, Theorem 4.2]
that this characterization also holds when V2 is any normed space (not necessarily
complete).
If T ∈ B(V1 , V2 ) attains its norm on the unique unit vector x0 (up to a scalar
factor) and T x0 is a smooth point of V2 , then T is said to satisfy Holub’s condition
(see [39]). Then Theorem 4.1 and Theorem 4.2 in [72] say that for a reflexive
Banach space V1 and any normed space V2 , smooth points of K(V1 , V2 ) are exactly
those operators which satisfy Holub’s condition. This characterization may not hold
if T is not compact (see [39, Example (a)]) or when V1 is not a reflexive space
(see [39, Example (b), Example (c)]). In the case when V1 is not a reflexive space,
usually some extra condition is needed along with Holub’s condition to characterize
smooth points. For example, Corollary 1 in [37] states that for 1 < p, r < ∞,
Birkhoff–James Orthogonality and Applications: A Survey 299

a necessary and sufficient condition for T ∈ B(l p , l r ) to be a smooth point is

that T satisfies Holub’s condition and dist(T , K(l p , l r )) < T . As an application
of orthogonality, it is proved in [60, Theorem 4.5] that for any normed spaces
V1 , V2 , if T ∈ B(V1 , V2 ) attains its norm and is a smooth point, then T satisfies
Holub’s condition and dist(T , K(V1 , V2 )) < T . The converse is true when V1 is
a reflexive Banach space and V2 is any Banach space and K(V1 , V2 ) is an M-ideal
in B(V1 , V2 ) (see [60, Theorem 4.6]). It is an open question whether or not these
extra assumptions on V1 and V2 are required. Some sufficient conditions, along with
Holub’s condition, for an operator to be smooth are also known when the underlying
field is R. If V1 is a real Banach space and V2 is a real normed space, one such
condition for smooth points in B(V1 , V2 ) is given in [72, Theorem 4.3]. When V1
and V2 are any real normed spaces, such conditions are given in [82, Theorem 3.2]
and [84, Theorem 3.4]. The extra condition which along with Holub’s condition
gives the characterization for smoothness of any non zero norm attaining operator
T ∈ B(V1 , V2 ) (for any real normed spaces V1 , V2 ) is obtained in [84, Theorem
3.3]. For further study of smooth points, we refer the readers to [36, 57, 73–75, 102].
Extreme points of D[0, 1] are important because of Krein–Milman theorem.
Along with the extreme points, the faces of D[0, 1] in any normed space have
also been of interest. (Note that the extreme points are exactly faces with a
single element.) Let Mn (R) or Mn (C) be equipped with any unitarily invariant
norm, ||| · ||| (that is, for any matrix A and U, U unitary, |||U AU ||| = |||A|||).
Then there is a unique symmetric gauge function  on Rn such that |||A||| =
((s1 (A), . . . , sn (A)), where si (A) are singular values of A arranged as s1 (A) ≥
· · · ≥ sn (A). Zi˛etak [108, Theorem 5.1] showed that a necessary and sufficient
condition for a matrix A to be an extreme point of the closed unit ball in
(Mn (R), ||| · |||) is that (s1 (A), . . . , sn (A)) is an extreme point of the closed unit ball
in (Rn , ). This result was extended to Mn (C) in [90, Theorem 1] (these results also
follow from the results in [11]). Li and Schneider [58, Proposition 4.1] characterized
the extreme points of D[0, 1] in Mn (R) and Mn (C), equipped with the dual of an
induced norm. In B(H), the extreme points of D[0, 1] are exactly the isometries
and the coisometries (see [38, p. 263]). It was proved in [95, Theorem 2.5] that
A ∈ B(H) is an isometry or a coisometry if and only if A = 1 and A is right
symmetric (for definition, see [30]). So the extreme points of D[0, 1] in B(H) are
precisely those operators which are of unit norm and are also right symmetric. There
is also a concept of a left symmetric operator, the study of which can be found in
[30, 71, 81, 96].
Theorem 2, Theorem 3 and Theorem 4 in [90] give characterizations of proper
closed faces in Mn (C), equipped with Schatten p-norms. Theorem 4.1 in [109] and
the discussion above it give a characterization of faces of D[0, 1] in (Mn (C), ||| · |||)
as follows: F is a face of (Mn (C), |||·|||) if and only if there exists A ∈ Mn (C) such
that F is a face of ∂|||A|||∗, the subdifferential set of ||| · |||∗ at A, where ||| · |||∗ is
the dual norm of ||| · |||. In a normed space V , the subdifferential set of a continuous
300 P. Grover and S. Singla

convex function g : V → R at v ∈ V is denoted by ∂g(v), and is defined as the set

of bounded linear functionals f ∈ V ∗ satisfying the below condition:

g(u) − g(v) ≥ Re f (u − v) for all u ∈ V .

It is a non-empty weak* compact convex subset of V ∗ . The below two propositions

are easy to check. We refer the readers to [33, 41] for more details.
Proposition 2.6 Let v ∈ V . Then

∂v = {f ∈ V ∗ : Re f (v) = v, f  ≤ 1}.

In particular, for A ∈ Mn (C),

∂|||A||| = {G ∈ Mn (C) : Re tr(G∗ A) = |||A|||, |||G|||∗ ≤ 1}.

Proposition 2.7 Let u, v ∈ V . Then we have

v + tu − v
lim = max{Re f (u) : f ∈ V ∗ , f  = 1, f (v) = v}.
t →0+ t

Using this, Watson [97, Theorem 4] gave a characterization of ∂||| · ||| in the space
(Mn (R), ||| · |||). Zi˛etak [109, Theorem 3.1, Theorem 3.2] improved this result and
showed the following.
Theorem 2.8 ([109, Theorem 3.1, Theorem 3.2]) For A ∈ Mn (C),
∗
: A = U U ∗ is a singular value
∂|||A||| ={U diag(d1 , . . . , dn )U

decomposition of A, si (A)di = |||A||| = ((s1 , . . . , sn )),

∗ ((d1 , . . . , dn )) = 1}.

In [98, Theorem 1], the above result was proved using a different approach. For the
operator norm  ·  on Mn (C), we have the following.
Corollary 2.9 ([98, Example 3]) For A ∈ Mn (C),

∂A = conv {uv ∗ : u = v = 1, Av = Au},

where conv S denotes the convex hull of a set S.

Along the similar lines of [97] (that is, by using Proposition 2.7), the subdifferential
set of the Ky Fan k-norms,  · (k) , on Mn (C) was obtained in [34]. In [15, 32–
34], the subdifferential set was used to obtain characterizations of orthogonality in
Mn (C), equipped with various norms.
Birkhoff–James Orthogonality and Applications: A Survey 301

Actually the right hand derivative has a deeper connection with orthogonality
as explored by Kečkić [51], where the author introduced the notion of ϕ-Gateaux
derivatives: for u, v ∈ V and ϕ ∈ [0, 2π), the ϕ-Gateaux derivative of norm at v in
the direction u is defined as
v + teιϕ u − v
Dϕ,v (u) = lim .
t →0+ t

These always exist for any two vectors u and v (see [51, Proposition 1.2]). A
characterization of orthogonality follows.
Theorem 2.10 ([51, Theorem 1.4]) Let u, v ∈ V . Then v is orthogonal to u if and
only if

inf Dϕ,v (u) ≥ 0.

0≤ϕ≤2π

In [52, Theorem 2.4], the expression for the ϕ-Gateaux derivative of the norm on
B(H) was obtained. Using the above proposition, a characterization of orthogonal-
ity in B(H) was given in [52, Corollary 3.1], which was first proved in [14] using a
completely different approach. This characterization of orthogonality and many of
its generalizations are the main content of the next section.

3 Characterizations and Applications of Orthogonality

Bhatia and Šemrl [14] gave characterizations of orthogonality in B(H) in terms of

orthogonality of vectors in the underlying Hilbert space H. These are given in the
next two theorems. An independent proof of Theorem 3.2 was also given by Paul
[70].
Theorem 3.1 ([14, Theorem 1.1]) Let A, B ∈ Mn (C). Then A is orthogonal to
B if and only if there exists a unit vector x ∈ Cn such that Ax = A and
Ax|Bx = 0.
Let · be any norm on Cn or Rn and let · be the corresponding induced norm on
Mn (C) or Mn (R), respectively. It was conjectured in [14, Remark 3.3] that a matrix
A is orthogonal to another matrix B in (Mn (C), · ) if and only if there exists a unit
vector x ∈ Cn such that Ax = A and Ax is orthogonal to Bx in (Cn , ·).
Li and Schneider [58, Example 4.3] gave an example to show that the conjecture is
false in Mn (C) as well as in Mn (R). In (Mn (C),  ·  ) (or (Mn (R),  ·  )), a matrix
A is said to satisfy BŠ property if for any matrix B, whenever A is orthogonal to B,
there exists a unit vector x such that Ax = A and Ax is orthogonal to Bx in
(Cn ,  · ) (or (Rn ,  · )) (see [80, Definition 1.1]). It was proved in [13] that
(Rn , ·) is an inner product space if and only if every A ∈ Mn (R) satisfies BŠ
property. In [80, Theorem 2.2], it was shown that if (Rn , ·) is a smooth space
302 P. Grover and S. Singla

and A ∈ Mn (R) is such that {x ∈ Rn : x = 1, Ax = A } is a countable

set with more than two points, then A does not satisfy BŠ property. Example 4.3
in [58] for Mn (R) is a special case of this. It was shown in [79, Corollary 2.1.1]
that if A ∈ Mn (R) attains its norm at exactly two points, then A satisfies BŠ
property. A generalization of this theorem can be found in [101, Theorem 3.1]. In
[79, Theorem 2.1], another sufficient condition for A to satisfy BŠ property was
given. If (Rn , ·) is a strictly convex space, then the collection of the matrices
which satisfy BŠ property are dense in Mn (R) (see [80, Theorem 2.6]).
Theorem 3.2 ([14, Remark 3.1], [70, Lemma 2]) Let H be a complex Hilbert
space. Let A, B ∈ B(H). Then A is orthogonal to B if and only if there exists a
sequence of unit vectors hn ∈ H such that Ahn  → A and Ahn |Bhn  → 0, as
n → ∞.
When H is an infinite dimensional space, one can’t expect to get a single vector h
in Theorem 3.2 such that Ah = A and Ah|Bh = 0. In fact it was proved in
[72, Theorem 3.1] that for A ∈ B(H), the following are equivalent.
(a) For B ∈ B(H), A is orthogonal to B if and only if there exists a unit vector
h ∈ H such that Ah = A and Ah|Bh = 0.
(b) There exists a finite dimensional subspace H0 of H such that
; ;
; ;
{h ∈ H : h = 1, A = Ah} = {h ∈ H0 : h = 1} and ;A|H⊥ ; < A .
0

It was noted in [14] that Theorem 3.1 is equivalent to saying that for A, B ∈
Mn (C),
 
dist(A, CB) = max Ax|y : x = y = 1 and y⊥Bx . (3.1)

It is natural to expect that in the infinite dimensional case, we should have for A, B ∈
B(H),
 
dist(A, CB) = sup Ax|y : x = y = 1 and y⊥Bx . (3.2)

This was indeed shown to be true in [14] by using the approach given in [5, p.
207]. We would like to point out that the book [5] deals with only separable spaces.
However the arguments can be modified by replacing the sequence of finite rank
operators converging pointwise to the identity operator by a net with this property.
Since the same proof as in [5, p. 207] was used in the proof of Theorem 2.4 of [96],
a similar modification is required there too.
Later, several authors have used different methods to prove Theorem 3.2. One of
these techniques was given in [7, Remark 2.2] using a different distance formula [7,
Proposition 2.1]. Another approach in [52, Corollary 3.1] uses Theorem 2.10 and
the expression for the ϕ-Gateaux derivative of the norm on B(H) (which is given
Birkhoff–James Orthogonality and Applications: A Survey 303

in [52, Theorem 2.4]). Using Theorem 2.10, Wójcik [99] extended Theorem 3.1 for
compact operators between two reflexive Banach spaces over C:
Theorem 3.3 ([99, Theorem 3.1]) Let V1 and V2 be reflexive Banach spaces over
C. Suppose A, B ∈ K(V1 , V2 ) and A = 0. Then A is orthogonal to B if and only if

min{max{Dϕ,Ay (By) : y = 1, Ay = A} : ϕ ∈ [0, 2π)} ≥ 0.

Let H1 and H2 be Hilbert spaces. In K(H1 , H2 ), the above theorem reduces to

saying that for A, B ∈ K(H1 , H2 ), A is orthogonal to B if and only if there is a unit
vector h ∈ H1 such that Ah = A and Ah|Bh = 0. But this is not always the
case with reflexive Banach spaces.
An alternate proof of Theorem 3.1 was given in [15] by first giving a character-
ization of A ≤ A + tB for all t ∈ R using Corollary 2.9, and then extend the
result to complex scalars to obtain Theorem 3.1. In [96], A ≤ A + tB for all
t ∈ R is termed as A is r-orthogonal to B, and the same characterization as in [15]
is given for r-orthogonality using a different approach. Using the same idea as in
[15, Theorem 2.7], a proof of Theorem 3.1 was given in [96, Corollary 2.2].
The technique of using the subdifferential set as done in [15] has advantages that
it gives a way to generalize Theorem 3.1 to characterize orthogonality to a subspace
of Mn (C).
Theorem 3.4 ([32, Theorem 1]) Let A ∈ Mn (C). Let m(A) denotes the multiplic-
ity of maximum singular value A of A. Let B be any (real or complex) subspace
of Mn (C). Then A is orthogonal to B if and only if there exists a density matrix P
of complex rank at most m(A) such that A∗ AP = A2 P and tr(AP B ∗ ) = 0 for
all B ∈ B.
Theorem 3.4 can be expressed in terms of states on Mn (C). Let A be a unital C ∗ -
algebra over F(= R or C) with the identity element 1A . For F = C, a state on A
is a linear functional φ on A which takes 1A to 1 and positive elements of A to
non-negative real numbers. For F = R, an additional requirement for φ to be a state
is that φ(a ∗ ) = φ(a) for all a ∈ A. Let SA denotes the set of states on A. Recently,
the authors noticed in [35] that if P is a density matrix such that tr(A∗ AP ) = A2 ,
then P is a matrix of complex rank atmost m(A) such that A∗ AP = A2 P (the
proof of this fact is along the lines of proof of Theorem 1.1 in [14]). Due to this
fact, the above theorem can be restated in terms of states on Mn (C) as follows: A is
orthogonal to B if and only if there exists φ ∈ SMn (C) such that φ(A∗ A) = A2 and
φ(AB ∗ ) = 0 for all B ∈ B. In a general complex C ∗ -algebra A, it was shown in [7,
Theorem 2.7] that an element a ∈ A is orthogonal to another element b ∈ A if and
only if there exists φ ∈ SA such that φ(a ∗ a) = a2 and φ(a ∗ b) = 0. A different
proof of this result was given in [15, Proposition 4.1]. Theorem 6.1 in [76] shows
that if B is a unital C ∗ -subalgebra of a complex C ∗ -algebra A and if a Hermitian
element a of A is orthogonal to B, then there exists φ ∈ SA such that φ(a 2 ) = a2
and φ(ab+b∗a) = 0 for all b ∈ B. Recently, the authors have extended these results
to any (real or complex) C ∗ -algebra A for any element a ∈ A and any subspace B
of A (see [35]). These are given in the next theorem.
304 P. Grover and S. Singla

If A is a complex (or real) unital C ∗ -algebra, then the triple (H, π, ξ ) denotes
a cyclic representation of A, where H is a complex (or real) Hilbert space and
π : A → B(H) is a ∗ -algebra map such that π(1A ) = 1B (H) and {π(a)ξ : a ∈ A}
is dense in B(H). For φ ∈ SA , there exists a cyclic representation (H, π, ξ ) such
that φ(a) = π(a)ξ |ξ  for all a ∈ A (see [21, p. 250], [31, Proposition 15.2]).
Theorem 3.5 ([35, Corollary 1.3]) Let a ∈ A. Let B be a subspace of A. Then the
following are equivalent.
1. a is orthogonal to B.
2. There exists φ ∈ SA such that φ(a ∗ a) = a2 and φ(a ∗ b) = 0 for all b ∈ B.
3. There exists a cyclic representation (H, π, ξ ) such that π(a)ξ  = a and
π(a)ξ |π(b)ξ  = 0 for all b ∈ B.
When A = C(X), Theorem 3.5 and Riesz Representation Theorem yield the
following theorem by Singer [89, Theorem 1.3, Ch. I].
Corollary 3.6 ([89, Theorem 1.3, Ch. I]) Let f ∈ C(X). Let B be a subspace of
C(X). Then f is orthogonal to B if and only if there exists a probability measure μ
on X such that
 
dist(a, B)2 = 2
|f | dμ and f hdμ = 0
X X

for all h ∈ B.
The condition

f 2∞ 2
= dist(a, B) = |f |2 dμ
X

is equivalent to saying that the support of μ is contained in the set {x ∈ X : |f (x)| =

f ∞ }. When B is one dimensional, this was proved in [53, Corollary 2.1] using
Theorem 2.10.
Characterizations of orthogonality have been studied in several normed spaces.
Using Theorem 2.10, a characterization of orthogonality in Cb () was obtained in
[53, Corollary 3.1]. In the Banach spaces L1 (X, ν) and c0 , Theorem 2.10 was used
to obtain such characterizations in [51, Example 1.6, Example 1.7]. For a separable
Hilbert space H, expressions for ϕ-Gateaux derivative of the norms on B1 (H)
and K(H) were given in [51, Theorem 2.1, Theorem 2.6] and were used to give
characterizations of orthogonality in these spaces in [51, Corollary 2.5, Corollary
2.8]. Using tools of subdifferential calculus, characterizations of orthogonality in
(Mn (C), ·(k) ) are given in [34, Theorem 1.1, Theorem 1.2]. A necessary condition
for orthogonality of a matrix A to a subspace in (Mn (C),  · (k) ) is given in
[34, Theorem 1.3]. Under the condition that sk (A) > 0, the same condition is
shown to be sufficient also. Using [89, Theorem 1.1, Ch. II], a characterization
of orthogonality in Mn (C), with any norm, is given in [58, Proposition 2.1] in
terms of the dual norm. Using this, orthogonality in Mn (C), with induced norms, is
Birkhoff–James Orthogonality and Applications: A Survey 305

obtained in [58, Proposition 4.2]. In Mn (C), with Schatten p-norms (1 ≤ p ≤ ∞),

characterizations of orthogonality are given in [58, Theorem 3.2, Theorem 3.3]. For
1 < p < ∞, this was also given in [14, Theorem 2.1].
Orthogonality has been characterized in more general normed spaces, namely,
Hilbert C ∗ -modules. It was shown in [7, Theorem 2.7] that in a Hilbert C ∗ -module
E over a complex unital C ∗ -algebra A, an element e1 ∈ E is orthogonal to another
element e2 ∈ E if and only if there exists φ ∈ SA such that φ(e1 |e1 ) =
e1 2 and φ(e1 |e2 ) = 0. Another proof of this result was given in [15, Theorem
4.4]. This can be generalized to obtain a characterization of orthogonality to
subspaces of Hilbert C ∗ -modules as follows.
Theorem 3.7 ([35, Theorem 3.5]) Let E be a Hilbert C ∗ -module over a unital
complex C ∗ -algebra A. Let e ∈ E . Let F be a subspace of E . Then e is orthogonal
to F if and only if there exists φ ∈ SA such that φ(e|e) = e2 and φ(e|f ) = 0
for all f ∈ F .
A proof of Theorem 3.7 can be found in [35]. Alternatively, this can also be proved
along the same lines of the proof of [7, Theorem 2.4] by finding a generalization
of the distance formula [7, Proposition 2.3] to a subspace. This extension of the
distance formula is mentioned in the next section.
We end this section with various directions of research happening around the
concept of orthogonality, where it comes into play naturally. In Hilbert C ∗ -modules,
the role of scalars is played by the elements of the underlying C ∗ -algebra. Using this
fact, a strong version of orthogonality was introduced in Hilbert C ∗ -modules in [8].
For a left Hilbert C ∗ -module, an element e1 ∈ E is said to be strong orthogonal
to another element e2 ∈ E if e1  ≤ e1 + ae2 for all a ∈ A. Clearly, if e1
is strong orthogonal to e2 , then we have e1 is orthogonal to e2 . However, strong
orthogonality is weaker than inner product orthogonality in E (see [8, Example
2.4]). Necessary and sufficient conditions are studied in [7, Theorem 3.1] and [9,
Theorem 3.5, Corollary 4.9], when any two of these three orthogonalities coincide
in a full Hilbert C ∗ -module. In [10, Theorem 2.6], it was shown that in a full
Hilbert C ∗ -module, strong orthogonality is symmetric if and only if Birkhoff–
James orthogonality is symmetric if and only if strong orthogonality coincides with
inner product orthogonality. Theorem 2.5 of [8] gives characterization of strong
orthogonality in terms of Birkhoff–James orthogonality.
Characterizations of orthogonality are also useful in finding conditions for
equality in triangle inequality in a normed space:
Proposition 3.8 ([7, Proposition 4.1]) Let V be a normed space. Let u, v ∈ V .
Then the following are equivalent.
1. u + v = u + v.
2. v is orthogonal to uv − vu.
3. u is orthogonal to uv − vu.
This can be extended to the case of arbitrarily finite families of vectors (see [7,
Remark 4.2]). As mentioned in [7, Remark 4.2], a characterization of triangle
306 P. Grover and S. Singla

equality in a pre-Hilbert C ∗ -module given in [6, Theorem 2.1] can be proved

using Theorem 3.7 and Proposition 3.8. For the study of various other equivalent
conditions for equality in triangle inequality or Pythagoras equality, the interested
reader is referred to [6, 7, 69].
In a normed linear space V , an element u is said to be norm-parallel to another
element v (denoted by u  v) if u + λv = u + v for some λ ∈ F with
|λ| = 1 [87]. In the case of inner product spaces, the norm-parallel relation is exactly
the usual vectorial parallel relation, that is, u  v if and only if u and v are linearly
dependent. Seddik [87] introduced this concept while studying elementary operators
on a standard operator algebra. Interested readers for the work on elementary
operators and orthogonality are referred to [4, 25, 26, 87, 88, 103], and also to [7,
Theorem 4.7] and [51, Section 3].
As a direct consequence of Proposition 3.8, norm-parallelism can be character-
ized using the concept of orthogonality (this characterization is also given in [68,
Theorem 2.4]). So the results on orthogonality can be used to find results on norm-
parallelism, for example, see [68, Proposition 2.19] for Mn (C) equipped with the
Schatten p-norms and [34, Remark 2] for (Mn (C),  · (k) ). The characterizations
of norm-parallelism in C(X) are given in [105], and in B1 (H) and K(H) are
given in [104]. Other results in B(V1 , V2 ) (with restrictions on V1 and V2 , and the
operators) are given in [100, 107]. Some of these results can be obtained by using
[68, Theorem 2.4] and the corresponding results on orthogonality. Some variants of
the definition of norm-parallelism have been introduced in [61, 104, 106]. Concepts
of approximate Birkhoff–James orthogonality and ε-Birkhoff orthogonality have
been studied in [19, 20, 23, 24, 43, 60, 83]. The idea to define these concepts
of approximate Birkhoff–James orthogonality and ε-Birkhoff orthogonality in a
normed space is to generalize the concept of approximate orthogonality in inner
product spaces, which is defined as v⊥ε u ⇐⇒ v|u ≤ ε v u. The latter has
been studied in [18] and [101, Section 5.2].

4 Distance Formulas and Conditional Expectations

An important connection of orthogonality with distance formulas was noted in (3.1)

and (3.2). From (3.1), we also get that for any A ∈ Mn (C),
 
dist(A, C1Mn (C) ) = max Ax|y : x = y = 1 and y⊥x .

Using this, one obtains

; ;
dist(A, C1Mn (C) ) = 2 max{;U AU ∗ − U AU ∗ ; : U, U unitary}
= 2 max{AU − U A : U unitary}
= 2 max{AT − T A : T ∈ Mn (C), T  = 1}. (4.1)
Birkhoff–James Orthogonality and Applications: A Survey 307

This was proved in [14, Theorem 1.2] and the discussion after that. The operator
δA (T ) = AT − T A on Mn (C) is called an inner derivation. So this gives that
δA  = 2 dist(A, C1Mn (C) ). This was also extended to the infinite dimensional case
in [14, Remark 3.2]. These results were first proved by Stampfli [91] using a com-
pletely different approach. Since all the derivations on B(H) are inner derivations
(see [49, Theorem 9]), the norm of any derivation on B(H) is 2 dist(A, C1B (H) ),
for some A ∈ B(H). For A, B ∈ B(H), let δA,B (T ) = AT − T B for all
T ∈ B(H). In [91, Theorem 8], an expression for the norm of the elementary
operator δA,B is given. In [91, Theorem 5], a distance formula was obtained in an
irreducible unital C ∗ -algebra as follows.
Theorem 4.1 ([91, Theorem 5]) Let H be a complex Hilbert space. Let B be an
irreducible unital C ∗ -subalgebra of B(H). Let A ∈ B. Then

2 dist(A, C1B (H) ) = sup{AT − T A : T ∈ B and T  = 1} =  δA |B  .

By the Russo–Dye theorem [17, II.3.2.15], under the assumptions of the above
theorem, we obtain

2 dist(A, C1B (H) ) = sup{AU − U A : U ∈ B and U is unitary}. (4.2)

Expressions for the norm of a derivation on von Neumann algebras can be found in
[29]. The most important fact used here is that all the derivations on von Neumann
algebras are inner. This was a conjecture by Kadison for a long time and was proved
in [85]. More on derivations on a C ∗ -algebra can be found in [47, 48, 67, 110]. A
lot of work has been done to answer the question when the range of a derivation
is orthogonal to its kernel. It was proved in [4, Theorem 1.7] that if N is a normal
operator in B(H), then the kernel of δN is orthogonal to the range of δN . In [54,
Theorem 1], it was shown that the Hilbert–Schmidt operators in the kernel of δN are
orthogonal to the Hilbert–Schmidt operators in the range of δN , in the usual Hilbert
space sense. In [59, Theorem 3.2(a)], the Schatten p-class operators in the kernel
of δN were shown to be orthogonal to the Schatten p-class operators in the range
of δN , in the Schatten p-norm. A similar result for the orthogonality in unitarily
invariant norms defined on the norm ideals of K(H) is given in [55, Theorem 1].
For related study on derivations, elementary operators and orthogonality in these
normed spaces, see [27, 50, 56, 62–66, 92–94].
Similar to (4.2), an expression for the distance of an element of a general C ∗ -
algebra from a C ∗ -subalgebra can be obtained from the below theorem of Rieffel
[76].
Theorem 4.2 ([76, Theorem 3.2]) Let A be a C ∗ -algebra. Let B be a C ∗ -
subalgebra of A which contains a bounded approximate identity for A. Let a ∈ A.
Then there exists a cyclic representation (H, π, ξ ) of A and a Hermitian as well
as a unitary operator U on H such that π(b)U = U π(b) for all b ∈ B and
dist(a, B) = 12 π(a)U − U π(a).
308 P. Grover and S. Singla

By Theorem 4.2, we obtain

2 dist(a, B) = max{ π(a)U − U π(a) : U ∈ B(H), U = U ∗ , U 2 = 1B (H),

(H, π, ξ ) is a cyclic representation of A, and
π(b)U = U π(b) for all b ∈ B}.

Looking at the last expression and (4.2), it is tempting to conjecture that if A is

a unital irreducible C ∗ -algebra, a ∈ A and B is a unital C ∗ -subalgebra of A, then

2 dist(a, B) = sup{ au − ua : u ∈ A, u is a unitary element,

and bu = ub for all b ∈ B}. (4.3)

We note that it is not possible to prove (4.3) by proceeding along the lines of the
proof of Theorem 4.1 given in [14], which uses (4.1). In particular, the following
does not hold true in Mn (C) :
 
dist(A, B) = max Ax|y : x = y = 1 and y⊥Bx for all B ∈ B .

For example, take A = 1Mn (C) and B = {X ∈ Mn (C) : tr(X) = 0}. Then 1Mn (C) is
orthogonal to B. Now if the above is true, then we would get unit vectors x, y such
that x|y = 1 and Bx|y = 0 for all B ∈ B. Let P = xy ∗. Then rank P = 1
and tr(BP ) = 0 for all B ∈ B. But tr(BP ) = 0 for all B ∈ B gives P = λ1Mn (C)
for some λ ∈ C (see [32, Remark 3]), which contradicts the fact that rank P = 1.
So this approach to prove (4.3) does not work. However it would be interesting to
know if (4.3) is true or not. This is an open question.
The above example contradicts Theorem 5.3 of [101], which says that for Hilbert
spaces H and K, if A ∈ K(H, K) and B is a finite dimensional subspace of
K(H, K), then
 
dist(A, B) = sup Ax|y : x = y = 1 and y⊥Bx for all B ∈ B .

The proof of this theorem has a gap, after invoking Theorem 5.2, in [101].
As an application of Theorem 3.5, we obtain the following distance formula.
Theorem 4.3 Let a ∈ A. Let B be a subspace of A. Suppose there is a best
approximation to a in B. Then

dist(a, B) = max π(a)ξ |η : (H, π, ξ ) is a cyclic representation of A,

η ∈ H, η = 1 and π(b)ξ |η = 0 for all b ∈ B . (4.4)

Proof Clearly RH S ≤ LH S. To prove equality, we need to find a cyclic

representation (H, π, ξ ) of A and a unit vector η ∈ H such that dist(a, B) =
π(a)ξ |η and π(b)ξ |η = 0 for all b ∈ B. Let b0 be a best approximation to
Birkhoff–James Orthogonality and Applications: A Survey 309

a in B. By Theorem 3.5, there exists φ ∈ SA such that

φ((a − b0 )∗ (a − b0 )) = a − b0 2

and φ((a − b0 )∗ b) = 0 for all b ∈ B. Now there exists a cyclic representation

(H, π, ξ ) such that φ(c) = π(c)ξ |ξ  for all c ∈ A. So π(a − b0 )ξ  = a − b0 
and π(a − b0 )ξ |π(b)ξ  = 0 for all b ∈ B. Taking η = a−b1
0
π(a − b0 )ξ , we get
the required result.
The authors have recently observed in [35] that the above theorem also holds true
without the existence of a best approximation to a in B. Notice that the right
hand side of (4.4) uses only algebraic structure of A (as cyclic representations are
defined by the algebraic structure of A). More such distance formulas using only the
algebraic structure of A are also known. When B = C1A , Williams [103, Theorem
2] proved that for a ∈ A,

dist(a, C1A )2 = max{φ(a ∗ a) − |φ(a)|2 : φ ∈ SA }. (4.5)

When A = Mn (C), another proof of (4.5) was given by Audenaert [12, Theorem
9]. Rieffel [77, Theorem 3.10] obtained (4.5), using a different method. In [77], it
was also desired to have a generalization of (4.5) with C1A replaced by a unital
C ∗ -subalgebra. For A = Mn (C), a formula in this direction was obtained in
[32, Theorem 2]. An immediate application of Theorem 3.5 gives the following
generalization of (4.5), when C1A is replaced by a subspace B of A and there is a
best approximation to a in B.
Theorem 4.4 ([35, Corollary 1.2]) Let a ∈ A. Let B be a subspace of A. Let b0
be a best approximation to a in B. Then

dist(a, B)2 = max{φ(a ∗ a) − φ(b0∗ b0 ) : φ ∈ SA , φ(a ∗ b) = φ(b0∗ b) for all b ∈ B}.

For details, see [35]. Geometric interpretations of Theorems 3.5 and 4.4 have also
been explained in [35].
Henceforward, C ∗ -algebras are assumed to be complex C ∗ -algebras. Another
distance formula, which is a generalization of [7, Proposition 2.3], is given below.
Some notations are in order. Given φ ∈ SA , let L = {c ∈ A : φ(c∗ c) = 0}, and let
a1 + L|a2 + LA/L = φ(a1∗ a2 ), for all a1 , a2 ∈ A. Then A/L is an inner product
space. For a ∈ A, let b0 be a best approximation to a in B. Let

Ma,B (φ) = sup{φ((a − b0 )∗ (a − b0 )) − |φ((a − b0 )∗ bα )|2 },
α

where the supremum is taken over all orthonormal bases {bα + L} of B/L in A/L.
310 P. Grover and S. Singla

Theorem 4.5 Let a ∈ A. Let B be a subspace of A. Let b0 be a best approximation

to a in B. Then

dist(a, B)2 = max{Ma,B (φ) : φ ∈ SA }.

Proof Clearly RH S ≤ LH S. For an orthonormal basis {bα + L} of B/L, we have


φ((a − b0 )∗ (a − b0 )) − |φ((a − b0 )∗ bα )|2 ≤ LH S.
α

And equality occurs because by Theorem 3.5, there exists φ ∈ SA such that φ((a −
b0 )∗ (a − b0 )) = dist(a, B)2 and φ((a − b0 )∗ b) = 0 for all b ∈ B.
Now along the lines of the proof of [7, Theorem 2.4] and using Theorem 4.5, we get
the next result. For a Hilbert C ∗ -module E over A and φ ∈ SA , let L = {e ∈ E :
φ(e|e) = 0}. On E /L, define an inner product as e1 +L|e2 +LE /L = φ(e1 |e2 )
for all e1 , e2 ∈ E . For e ∈ E , let f0 be a best approximation to e in F . Let

Me,F (φ) = sup{φ(e − f0 |e − f0 ) − |φ(e − f0 |fα )|2 },
α

where the supremum is taken over all orthonormal bases {fα + L} of F /L in E /L.
Theorem 4.6 Let E be a Hilbert C ∗ -module over A. Let e ∈ E . Let F be a
subspace of E . Let f0 be a best approximation to e in F . Then

dist(e, F ) = max{Me,F (φ) : φ ∈ SA }.

Rieffel [77, p. 46] had questioned to have expressions of distance formulas in terms
of conditional expectations. We end the discussion on distance formulas with our
progress in this direction. For a C ∗ -algebra A and a C ∗ -subalgebra B of A, a
conditional expectation from A to B is a completely positive map E : A → B
of unit norm such that E(b) = b, E(ba) = bE(a) and E(ab) = E(a)b, for all
a ∈ A and b ∈ B [17, p. 141]. In fact any projection E : A → B of norm
one is a conditional expectation and vice-a-versa (see [17, Theorem II.6.10.2]). An
interesting fact is that a map E : A → B is a conditional expectation if and only
if E is idempotent, positive and satisfies E(b1 ab2) = b1 E(a)b2, for all a ∈ A and
b1 , b2 ∈ B (see [17, Theorem II.6.10.3]). Thus conditional expectations from A to B
are also determined completely by the algebraic structure. A Banach space V1 is said
to be injective if for any inclusion of Banach spaces V3 ⊆ V2 , every bounded linear
mapping f0 : V3 → V1 has a linear extension f : V2 → V1 with f  = f0 . A
Banach space is injective if and only if it is isometrically isomorphic to C(X), where
X is a compact Hausdorff space in which closure of any open set is an open set (see
[28, p. 70]). For v ∈ V and W a subspace of V , let v, W  denote the subspace
generated by v and W .
Birkhoff–James Orthogonality and Applications: A Survey 311

Theorem 4.7 Let a ∈ A. Let B be a subspace of A such that B is an injective

Banach space and 1A ∈ B. Suppose there is a best approximation to a in B. Then
there exists φ ∈ SA and a projection E : A → B of norm atmost two such that
φ ◦ E = φ and dist(a, B)2 = φ(a ∗ a) − φ(E(a)∗ E(a)).
Proof Let b0 be a best approximation to a in B. We define Ẽ : a, B → B as
Ẽ(b) = b for all b ∈ B and Ẽ(a) = b0 and extend it linearly on a, B. Using
Theorem 3.5, there exists φ ∈ SA such that dist(a, B)2 = φ(a ∗ a) − φ(b0∗ b0 ) and
φ(a ∗ b) = φ(b0∗ b) for all b ∈ B. Since 1A ∈ B, we get φ(a) = φ(b0 ) = φ(E(a)).
And clearly φ(b) = φ(Ẽ(b)) for all b ∈ B. Thus φ ◦ Ẽ = φ. Since b0 is a best
approximation to a in B, a − b0  ≤ a. So b0  ≤ 2 a. Now let b ∈ B and
α ∈ C. Then Ẽ(αa + b) = αb0 + b and αb0 + b is a best approximation to αa + b
in B. Thus αb0 + b ≤ 2 αa + b. Hence Ẽ ≤ 2. Since B is injective, there
exists a linear extension E : A → B with norm same as that of Ẽ. This E is the
required projection.
For any given conditional expectation E from A to B, we can define a B-valued
inner product on A given by a1 |a2 E = E(a1∗ a2 ) (see [77]). In [35], we obtain a
lower bound for dist(a, B) as follows.
Theorem 4.8 Let a ∈ A. Let B be a C ∗ -subalgebra of A such that 1A ∈ B. Then

dist(a, B)2 ≥ sup{φ(a − E(a)|a − E(a)E ) : φ ∈ SA , E is a conditional

expectation from A onto B}. (4.6)

(Here we follow the convention that sup(∅) = −∞.)

Remark 4.9 It is worth mentioning here that if in Theorem 4.7, we take B to be
a C ∗ -algebra and we are able to find a projection of norm one, then we will get
equality in (4.6), that is,

dist(a, B)2 = sup{φ(a − E(a)|a − E(a)E ) : φ ∈ SA , E is a conditional

expectation from A onto B}.

This happens in the special case when B = C1A , because for any c ∈ A, the norm
of the best approximation of c to C1A is less than or equal to c, and thus the norm
of the projection E in Theorem 4.7 is one.

Acknowledgments The authors would like to thank Amber Habib and Ved Prakash Gupta for
many useful discussions. The authors would also like to acknowledge very helpful comments by
the referees.
The research of P. Grover is supported by INSPIRE Faculty Award IFA14-MA-52 of DST,
India, and by Early Career Research Award ECR/2018/001784 of SERB, India.
312 P. Grover and S. Singla

References

1. T.J. Abatzoglou, Norm derivatives on spaces of operators. Math. Ann. 239, 129–135 (1979)
2. J. Alonso, Some properties of Birkhoff and isosceles orthogonality in normed linear spaces,
in Inner Product Spaces and Applications. Pitman Research Notes in Mathematical Series,
vol. 376 (Longman, Harlow, 1997), pp. 1–11
3. D. Amir, Characterizations of Inner Product Spaces. Operator Theory: Advances and
Applications, vol. 20 (Birkhäuser Verlag, Basel, 1986)
4. J. Anderson, On normal derivations. Proc. Am. Math. Soc. 38, 135–140 (1973)
5. C. Apostol, L.A. Fialkow, D.A. Herrero, D. Voiculescu, Approximation of Hilbert Space
Operators II. Research Notes in Mathematics, vol. 102 (Pitman (Advanced Publishing
Program), Boston, 1984)
6. L. Arambašić, R. Rajić, On some norm equalities in pre-Hilbert C ∗ -modules. Linear Algebra
Appl. 414, 19–28 (2006)
7. L. Arambašić, R. Rajić, The Birkhoff-James orthogonality in Hilbert C ∗ -modules. Linear
Algebra Appl. 437, 1913–1929 (2012)
8. L. Arambašić, R. Rajić, A strong version of the Birkhoff-James orthogonality in Hilbert C ∗ -
modules. Ann. Funct. Anal. 5, 109–120 (2014)
9. L. Arambašić, R. Rajić, On three concepts of orthogonality in Hilbert C ∗ -modules. Linear
Multilinear Algebra 63, 1485–1500 (2015)
10. L. Arambašić, R. Rajić, On symmetry of the (strong) Birkhoff-James orthogonality in Hilbert
C ∗ -modules. Ann. Funct. Anal. 7, 17–23 (2016)
11. J. Arazy, On the geometry of the unit ball of unitary matrix spaces. Integr. Equ. Oper. Theory
4, 151–171 (1981)
12. K.M.R. Audenaert, Variance bounds, with an application to norm bounds for commutators.
Linear Algebra Appl. 432, 1126–1143 (2010)
13. C. Benítez, M. Fernández, M.L. Soriano, Orthogonality of matrices. Linear Algebra Appl.
422, 155–163 (2007)
14. R. Bhatia, P. Šemrl, Orthogonality of matrices and some distance problems. Linear Algebra
Appl. 287, 77–85 (1999)
15. T. Bhattacharyya, P. Grover, Characterization of Birkhoff-James orthogonality. J. Math. Anal.
Appl. 407, 350–358 (2013)
16. G. Birkhoff, Orthogonality in linear metric spaces. Duke Math. J. 1, 169–172 (1935)
17. B. Blackadar, Operator Algebras-Theory of C ∗ -Algebras and von Neumann Algebras
(Springer, Berlin, 2006)
18. J. Chmieliński, Linear mappings approximately preserving orthogonality. J. Math. Anal.
Appl. 304, 158–169 (2005)
19. J. Chmieliński, On an ε-Birkhoff orthogonality. J. Inequal. Pure Appl. Math. 6, 1–7 (2005)
20. J. Chmieliński, T. Stypula, P. Wójcik, Approximate orthogonality in normed spaces and its
applications. Linear Algebra Appl. 531, 305–317 (2017)
21. J.B. Conway, A Course in Functional Analysis (Springer, New York, 1990)
22. J. Diestel, Geometry of Banach Spaces. Lecture Notes in Mathematics, vol. 485 (Springer,
Berlin, 1975)
23. S.S. Dragomir, On Approximation of Continuous Linear Functionals in Normed Linear
Spaces. An. Univ. Timişoara Ser. Ştiint. Mat. 29, 51–58 (1991)
24. S.S. Dragomir, Continuous linear functionals and norm derivatives in real normed spaces.
Univ. Beograd. Publ. Elektrotehn. Fak. Ser. Mat. 3, 5–12 (1992)
25. H.-K. Du, Another generalization of Anderson’s theorem. Proc. Am. Math. Soc. 123, 2709–
2714 (1995)
26. B.P. Duggal, A remark on normal derivations. Proc. Am. Math. Soc. 126, 2047–2052 (1998)

n
27. B.P. Duggal, Range-kernel orthogonality of the elementary operator X → Ai XBi − X.
i=1
Linear Algebra Appl. 337, 79–86 (2001)
Birkhoff–James Orthogonality and Applications: A Survey 313

28. E.G. Effros, Z.-J. Ruan, Operator Spaces (Oxford University Press, New York, 2000)
29. P. Gajendragadkar, Norm of a derivation on a von Neumann algebra. Trans. Am. Math. Soc.
170, 165–170 (1972)
30. P. Ghosh, D. Sain, K. Paul, On symmetry of Birkhoff-James orthogonality of linear operators.
Adv. Oper. Theory 2, 428–434 (2017)
31. K.R. Goodearl, Notes on Real and Complex C ∗ -Algebras (Shiva, Cambridge, 1982)
32. P. Grover, Orthogonality to matrix subspaces, and a distance formula. Linear Algebra Appl.
445, 280–288 (2014)
33. P. Grover, Some problems in differential and subdifferential calculus of matrices. Ph.D.
Thesis, Indian Statistical Institute (2014)
34. P. Grover, Orthogonality of matrices in the Ky Fan k-norms. Linear Multilinear Algebra 65,
496–509 (2017)
35. P. Grover, S. Singla, Best approximations, distance formulas and orthogonality in C ∗ -
algebras. J. Ramanujan Math. Soc. (to appear)
36. R. Grza̧ślewicz, R. Younis, Smooth points of some operator spaces. Arch. Math. 57, 402–405
(1991)
37. R. Grza̧ślewicz, R. Younis, Smooth points and M-ideals. J. Math. Anal. Appl. 175, 91–95
(1993)
38. P. Halmos, A Hilbert Space Problem Book (D. Van Nostrand, Princeton, 1967)
39. J. Hennefeld, Smooth, compact operators. Proc. Am. Math. Soc. 77, 87–90 (1979)
40. S. Heinrich, The differentiability of the norm in spaces of operators. Funct. Anal. Appl. 9,
360–362 (1975)
41. J.B. Hiriart-Urruty, C. Lemarèchal, Fundamentals of Convex Analysis (Springer, Berlin, 2000)
42. J.R. Holub, On the metric geometry of ideals of operators on Hilbert space. Math. Ann. 201,
157–163 (1973)
43. T. Jahn, Orthogonality in generalized Minkowski spaces. J. Convex Anal. 26, 49–76 (2019)
44. R.C. James, Orthogonality in normed linear spaces. Duke Math. J. 12, 291–302 (1945)
45. R.C. James, Orthogonality and linear functionals in normed linear spaces. Trans. Am. Math.
Soc. 61, 265–292 (1947)
46. R.C. James, Inner products in normed linear spaces. Bull. Am. Math. Soc. 53, 559–566 (1947)
47. B. Johnson, Characterization and norms of derivations on von Neumann algebras, in Algèbres
d’Opérateurs. Lecture Notes in Mathematics, vol. 725 (Springer, Berlin, 1979), pp. 228–236
48. R.V. Kadison, Derivations of operator algebras. Ann. Math. 83, 280–293 (1966)
49. I. Kaplansky, Modules over operator algebras. Am. J. Math. 75, 839–858 (1953)
50. D.J. Kečkić, Orthogonality of the range and the kernel of some elementary operators. Proc.
Am. Math. Soc. 128, 3369–3377 (2000)
51. D.J. Kečkić, Orthogonality in S1 and S∞ spaces and normal derivations. J. Oper. Theory 51,
89–104 (2004)
52. D.J. Kečkić, Gateaux derivative of B(H ) norm. Proc. Am. Math. Soc. 133, 2061–2067 (2005)
53. D.J. Kečkić, Orthogonality and smooth points in C(K) and Cb (). Eur. Math. J. 3, 44–52
(2012)
54. F. Kittaneh, On normal derivations of Hilbert-Schmidt type. Glasg. Math. J. 29, 245–248
(1987)
55. F. Kittaneh, Normal derivations in norm ideals. Proc. Am. Math. Soc. 123, 1779–1785 (1995)
56. F. Kittaneh, Operators that are orthogonal to the range of a derivation. J. Math. Anal. Appl.
203, 868–873 (1996)
57. F. Kittaneh, R. Younis, Smooth points of certain operator spaces. Integr. Equ. Oper. Theory
13, 849–855 (1990)
58. C.K. Li, H. Schneider, Orthogonality of matrices. Linear Algebra Appl. 347, 115–122 (2002)
59. P.J. Maher, Commutator approximants. Proc. Am. Math. Soc. 115, 995–1000 (1992)
60. A. Mal, K. Paul, T.S.S.R.K. Rao, D. Sain, Approximate Birkhoff-James orthogonality and
smoothness in the space of bounded linear operators. Monatsh. Math. 190, 549–558 (2019)
61. A. Mal, D. Sain, K. Paul, On some geometric properties of operator spaces. Banach J. Math.
Anal. 13, 174–191 (2019)
314 P. Grover and S. Singla

62. A. Mazouz, On the range and the kernel of the operator X → AXB − X. Proc. Am. Math.
Soc. 127, 2105–2107 (1999)
p
63. S. Mecheri, On minimizing S − (AX − XB)p . Serdica Math. J. 26, 119–126 (2000)
64. S. Mecheri, Some versions of Anderson’s and Maher’s inequalities I. Int. J. Math. Math. Sci.
52, 3281–3297 (2003)
65. S. Mecheri, Some versions of Anderson’s and Maher’s inequalities II. Int. J. Math. Math. Sci.
53, 3355–3372 (2003)
66. S. Mecheri, M. Bounkhel, Some variants of Anderson’s inequality in C1 -classes. JIPAM. J.
Inequal. Pure Appl. Math. 4, Article 24 (2003)
67. P. Miles, Derivations on B ∗ algebras. Pac. J. Math. 14, 1359–1366 (1964)
68. M.S. Moslehian, A. Zamani, Norm-parallelism in the geometry of Hilbert C ∗ -modules. Indag.
Math. 27, 266–281 (2016)
69. R. Nakamoto, S. Takahasi, Norm equality condition in triangular inequality. Sci. Math. Jpn.
55, 463–466 (2002)
70. K. Paul, Translatable radii of an operator in the direction of another operator. Sci. Math. 2,
119–122 (1999)
71. K. Paul, A. Mal, P. Wójcik, Symmetry of Birkhoff-James orthogonality of operators defined
between infinite dimensional Banach spaces. Linear Algebra Appl. 563, 142–153 (2019)
72. K. Paul, D. Sain, P. Ghosh, Birkhoff-James orthogonality and smoothness of bounded linear
operators. Linear Algebra Appl. 506, 551–563 (2016)
73. T.S.S.R.K. Rao, Very smooth points in spaces of operators. Proc. Indian Acad. Sci. Math. Sci.
113, 53–64 (2003)
74. T.S.S.R.K. Rao, Smooth points in spaces of operators. Linear Algebra Appl. 517, 129–133
(2017)
75. T.S.S.R.K. Rao, Adjoints of operators as smooth points in spaces of compact operators. Linear
Multilinear Algebra 66, 668–670 (2018)
76. M.A. Rieffel, Leibniz seminorms and best approximation from C ∗ -subalgebras. Sci. China
Math. 54, 2259–2274 (2011)
77. M.A. Rieffel, Standard deviation is a strongly Leibniz seminorm. New York J. Math. 20,
35–56 (2014)
78. B.D. Roberts, On geometry of abstract vector spaces. Tohoku Math. J. 39, 42–59 (1934)
79. D. Sain, K. Paul, Operator norm attainment and inner product spaces. Linear Algebra Appl.
439, 2448–2452 (2013)
80. D. Sain, K. Paul, S. Hait, Operator norm attainment and Birkhoff-James orthogonality. Linear
Algebra Appl. 476, 85–97 (2015)
81. D. Sain, P. Ghosh, K. Paul, On symmetry of Birkhoff-James orthogonality of linear operators
on finite-dimensional real Banach spaces. Oper. Matrices 11, 1087–1095 (2017)
82. D. Sain, K. Paul, A. Mal, A complete characterization of Birkhoff-James orthogonality in
infinite dimensional normed space. J. Oper. Theory 80, 399–413 (2018)
83. D. Sain, K. Paul, A. Mal, On approximate Birkhoff-James orthogonality and normal cones in
a normed space. J. Convex Anal. 26, 341–351 (2019)
84. D. Sain, K. Paul, A. Mal, A. Ray, A complete characterization of smoothness in the space
of bounded linear operators. Linear Multilinear Algebra (2019). https://doi.org/10.1080/
03081087.2019.1586824
85. S. Sakai, Derivations of W ∗ -algebras. Ann. Math. 83, 273–279 (1966)
86. R. Schatten, The space of completely continuous operators on a Hilbert space. Math. Ann.
134, 47–49 (1957)
87. A. Seddik, Rank one operators and norm of elementary operators. Linear Algebra Appl. 424,
177–183 (2007)
n
88. A. Seddik, On the injective norm of Ai ⊗ Bi and characterization of normaloid operators.
i=1
Oper. Matrices 2, 67–77 (2008)
89. I. Singer, Best Approximation in Normed Linear Spaces by Elements of Linear Subspaces
(Springer, New York, 1970)
Birkhoff–James Orthogonality and Applications: A Survey 315

90. W. So, Facial structures of Schatten p-norms. Linear Multilinear Algebra 27, 207–212 (1990)
91. J. Stampfli, The norm of a derivation. Pac. J. Math. 33, 737–747 (1970)
92. A. Turnšek, Elementary operators and orthogonality. Linear Algebra Appl. 317, 207–216
(2000)
93. A. Turnšek, Orthogonality in Cp classes. Monatsh. Math. 132, 349–354 (2001)
94. A. Turnšek, Generalized Anderson’s inequality. J. Math. Anal. Appl. 263, 121–134 (2001)
95. A. Turnšek, On operators preserving James’ orthogonality. Linear Algebra Appl. 407, 189–
195 (2005)
96. A. Turnšek, A remark on orthogonality and symmetry of operators in B (H ). Linear Algebra
Appl. 535, 141–150 (2017)
97. G.A. Watson, Characterization of the subdifferential of some matrix norms. Linear Algebra
Appl. 170, 33–45 (1992)
98. G.A. Watson, On matrix approximation problems with Ky Fan k norms. Numer. Algorithms
5, 263–272 (1993)
99. P. Wójcik, Gateaux derivative of norm in K(X; Y ). Ann. Funct. Anal. 7, 678–685 (2016)
100. P. Wójcik, Norm-parallelism in classical M-ideals. Indag. Math. 28, 287–293 (2017)
101. P. Wójcik, Orthogonality of compact operators. Expo. Math. 35, 86–94 (2017)
102. W. Werner, Smooth points in some spaces of bounded operators. Integr. Equ. Oper. Theory
15, 496–502 (1992)
103. J.P. Williams, Finite operators. Proc. Am. Math. Soc. 26, 129–136 (1970)
104. A. Zamani, The operator-valued parallelism. Linear Algebra Appl. 505, 282–295 (2016)
105. A. Zamani, Characterizations of norm-parallelism in spaces of continuous functions. Bull.
Iranian Math. Soc. 45, 557–567 (2019)
106. A. Zamani, M.S. Moslehian, Exact and approximate operator parallelism. Can. Math. Bull.
58, 207–224 (2015)
107. A. Zamani, M.S. Moslehian, M.-T. Chien, H. Nakazato, Norm-parallelism and the Davis-
Wielandt radius of Hilbert space operators. Linear Multilinear Algebra 67, 2147–2158 (2019)
108. K. Zi˛etak, On the characterization of the extremal points of the unit sphere of matrices. Linear
Algebra Appl. 106, 57–75 (1988)
109. K. Zi˛etak, Subdifferentials, faces, and dual matrices. Linear Algebra Appl. 185, 125–141
(1993)
110. L. Zsidó, The norm of a derivation in a W ∗ -algebra. Proc. Am. Math. Soc. 38, 147–150 (1973)
The Generalized ∂-Complex on the
Segal–Bargmann Space

Friedrich Haslinger

Abstract We study certain densely defined unbounded operators on the Segal-

Bargmann space, related to the annihilation and creation operators of quantum
mechanics. We consider the corresponding D-complex and study properties of the
corresponding complex Laplacian ˜ D = DD ∗ + D ∗ D, where D is a differential
operator of polynomial type.

Keywords ∂-complex · Segal–Bargmann space

Mathematics Subject Classification (2010) Primary 30H20, 32A36, 32W50;

Secondary 47B38.

1 Introduction

We consider the classical Segal–Bargmann space

 
−|z|2
|u(z)|2 e−|z| dλ(z) < ∞
2 2
n
A (C , e ) = u : C −→ C entire :
n
Cn

with inner product


u(z) v(z) e−|z| dλ(z)
2
(u, v) =
Cn

and replace a single derivative with respect to zj by a differential operator of the

form pj ( ∂z∂ 1 , . . . , ∂z∂ n ), where pj is a complex polynomial on Cn (see [5, 6]). We

F. Haslinger ()
Fakultät für Mathematik, Universität Wien, Wien, Austria
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 317

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_16
318 F. Haslinger

write pj (u) for pj ( ∂z∂ 1 , . . . , ∂z∂ n )u, where u ∈ A2 (Cn , e−|z| ), and consider the
2

densely defined operators


n
Du = pj (u) dzj , (1.1)
j =1

where u ∈ A2 (Cn , e−|z| ) and pj ( ∂z∂ 1 , . . . , ∂z∂ n ) are polynomial differential opera-
2

tors with constant coefficients [3].

More generally we define

 
n
Du = pk (uJ ) dzk ∧ dzJ ,
|J |=p k=1

 2 n −|z|2 ),
where u = |J |=p uJ dzJ is a (p, 0)-form with coefficients in A (C , e
here J = (j1 , . . . , jp ) is a multiindex and dzJ = dzj1 ∧ · · · ∧ dzjp and the
summation is taken only over increasing multiindices.
It is clear that D 2 = 0 and that we have

(Du, v) = (u, D ∗ v), (1.2)

where
 
u ∈ dom(D) = u ∈ A2(p,0)(Cn , e−|z| ) : Du ∈ A2(p+1,0)(Cn , e−|z| )
2 2

and
 
n
∗
D v= pj∗ vj K dzK
|K|=p−1 j =1
 ∗
for v = |J |=p vJ dzJ and where pj (z1 , . . . , zn ) is the polynomial pj with
complex conjugate coefficients, taken as multiplication operator.
Now the corresponding D-complex has the form
D D
A2(p−1,0)(Cn , e−|z| ) −→ A2(p,0)(Cn , e−|z| ) −→ A2(p+1,0)(Cn , e−|z| ).
2 2 2

←− ←−
D∗ D∗

Similarly to the classical ∂-complex (see [3]) we consider the generalized box
operator ˜ D,p := D ∗ D + DD ∗ as a densely defined self-adjoint operator on
A2(p,0)(Cn , e−|z| ) with
2

˜ D,p ) = {f ∈ dom(D) ∩ dom(D ∗ ) : Df ∈ dom(D ∗ ), D ∗ f ∈ dom(D)},

dom(

see [4] for more details.

The Generalized ∂-Complex on the Segal–Bargmann Space 319

The (p, 0)-forms with polynomial components are dense in the space
A2(p,0)(Cn , e−|z| ). In addition we have the following.
2

Lemma 1.1 The (p, 0)-forms with polynomial components are also dense in
dom(D) ∩ dom(D ∗ ) endowed with the graph norm

u → (u2 + Du2 + D ∗ u2 )1/2 .


Proof Let u = |J |=p uJ dzJ ∈ dom(D) ∩ dom(D ∗ ) and consider the partial
sums of the Fourier series expansions of

uJ = uJ,α ϕα ,
α

where

zα 
ϕα (z) = √ and |uJ,α |2 < ∞
n
π α! α

and α! = α1 ! . . . αn !. We have that Du ∈ A2(p+1,0)(Cn , e−|z| ) and D ∗ u ∈

2

A2(p−1,0)(Cn , e−|z| ). Hence the partial sums of the Fourier series of the components
2

of Du converge to the components of Du in A2(p+1,0)(Cn , e−|z| ) and the partial

2

sums of the Fourier series of the components of D ∗ u converge to the components

of D ∗ u in A2(p−1,0)(Cn , e−|z| ).
2

2 The Basic Estimate

We want to find conditions under which ˜ D,1 has a bounded inverse. For this
purpose we have to consider the graph norm (u2 + Du2 + D ∗ u2 )1/2 on
dom(D) ∩ dom(D ∗ ). We refer to [4, Theorem 5.1] here in a slightly different
improved form.
Theorem 2.1 Suppose that there exists a constant C > 0 such that


n
u2 ≤ C ([pk , pj∗ ]uj , uk ), (2.1)
j,k=1

n
for any (1, 0)-form u = j =1 uj dzj with polynomial components. Then

u2 ≤ C(Du2 + D ∗ u2 ), (2.2)

for any u ∈ dom(D) ∩ dom(D ∗ ).

320 F. Haslinger

Proof First we have

 
n
Du = (pj (uk ) − pk (uj )) dzj ∧ dzk and D ∗ u = pj∗ uj ,
j <k j =1

hence
 
Du2 + D ∗ u2 = |pk (uj ) − pj (uk )|2 e−|z| dλ
2

Cn j <k

 
n
pj∗ uj pk∗ uk e−|z| dλ
2
+
Cn j,k=1

n 

|pk (uj )|2 e−|z| dλ
2
=
j,k=1 Cn

n 

(pj∗ uj pk∗ uk − pk (uj )pj (uk )) e−|z| dλ
2
+
j,k=1 Cn


n 
|pk (uj )|2 e−|z| dλ
2
=
j,k=1 C
n

n 

[pk , pj∗ ]uj uk e−|z| dλ,
2
+
j,k=1 C
n

where we used (1.2). Note that the expression

n 

|pk (uj )|2 e−|z| dλ
2

j,k=1 Cn

is finite, since the components uj are polynomials, and it follows that the expression
 n ∗
j,k=1 ([pk , pj ]uj , uk ) is a real number.
Now the assumption (2.1) implies that (2.2) holds for (1, 0)-forms with poly-
nomial components and, by Lemma 1.1, we obtain (2.2) for any u ∈ dom(D) ∩
dom(D ∗ ).
Remark 2.2 In Theorem 2.1 we implicitly suppose that the expression


n
([pk , pj∗ ]uj , uk )
j,k=1

is nonnegative.
The Generalized ∂-Complex on the Segal–Bargmann Space 321

First we consider the one-dimensional case. Let pm denote the polynomial differen-
tial operator

∂ ∂m
pm = a0 + a1 + · · · + am m ,
∂z ∂z
∗ denote the polynomial
with constant coefficients a0 , a1 , . . . , am ∈ C, and let pm
∗
pm (z) = a 0 + a 1 z + · · · + a m zm ,

with the complex conjugate coefficients a 0 , a 1 , . . . , a m ∈ C.

We consider the densely defined operator

Du = pm (u) dz,

where u ∈ A2 (C, e−|z| ) and pm (u) dz is considered as a (1, 0)-form.

2

It is clear that D 2 = 0, as all (2, 0)-forms are zero if n = 1, and that we have

(Du, v) = (u, D ∗ v),

where u ∈ dom(D) = {u ∈ A2 (C, e−|z| ) : Du ∈ A2(1,0)(C, e−|z| )} and

2 2

D ∗ v dz = pm
∗
v.

In the sequel we consider the generalized box operator

˜ D,1 := DD ∗


as a densely defined self-adjoint operator on A2(1,0)(C, e−|z| ) with

2

˜ D,1 ) = {f ∈ dom(D ∗ ) : D ∗ f ∈ dom(D)}.

dom(

Lemma 2.3 Let u be an arbitrary polynomial. Then


∗
]u(z) u(z) e−|z| dλ(z)
2
[pm , pm
C


m  m  

2
k
e−|z| dλ(z).
2
= ! ak u(k−)(z) (2.3)
C k= 
=1

Proof On the right hand side of (2.3) we have the integrand


m m   
 k j
! ak a j u(k−) u(j −) . (2.4)
 
=1 k,j =
322 F. Haslinger

For the left hand side of (2.3) we first compute

 ∂k       
m m m m
∗
[pm , pm ]u = ak k
a j zj u − a j zj aj u(j ) .
∂z
k=0 j =0 j =0 j =0

We use the Leibniz rule to get

  
∂k      
m m m j
∂k j k j (k−) j −
aj z u =
j
a j k (z u) = aj ! u z ,
∂zk ∂z  
j =0 j =0 j =0 =0

k
notice that  = 0, in the case k < . Hence we obtain


m 
j   
∗ k j (k−) j −
[pm , pm ]u = ak a j ! u z .
 
j,k=1 =1

After integration we obtain


∗
]u(z) u(z) e−|z| dλ(z)
2
[pm , pm
C


m 
j    
k j
u(k−)u(j −) e−|z| dλ(z).
2
= ak a j ! (2.5)
  C
j,k=1 =1

Now it is easy to show that integration of (2.4) coincides with (2.5) and we are done.

Lemma 2.4 There exists a constant C > 0 such that, for each u ∈ dom(D ∗ ),

u ≤ CD ∗ u.

Proof First let u be a polynomial and note that the last term in (2.3) equals

|am u(z)|2 e−|z| dλ(z) = m! |am |2 u2
2
m!
C

and all the other terms are non-negative, see Lemma 2.3. Now we get that

1
D ∗ u2 = (pm
∗ ∗
u, pm ∗
u) = ([pm , pm ]u, u) + (pm (u), pm (u)) ≥ u2 ,
C
where
1
C= ,
m! |am |2
The Generalized ∂-Complex on the Segal–Bargmann Space 323

if we suppose that am = 0, and we are done. Finally apply Lemma 1.1 to obtain the
desired result.
˜ D,1 = DD ∗ has a bounded inverse
Theorem 2.5 Let D be as in (1.1). Then 

˜ D,1 ).
ÑD,1 : A2(1,0)(C, e−|z| ) −→ dom(
2

If α ∈ A2(1,0)(C, e−|z| ), then u0 = D ∗ ÑD,1 α is the canonical solution of Du = α,

2

this means that Du0 = α and u0 ∈ (kerD)⊥ = imD ∗ , and

D ∗ ÑD,1 α ≤ Cα

for some constant C > 0 independent of α.

Proof Using Lemma 2.4 we obtain that

˜ D,1 : dom(
 ˜ D,1 ) −→ A2 (C, e−|z|2 )
(1,0)

is bijective and has the bounded inverse ÑD,1 , see [4, Theorem 5.1]. The rest follows
from [4, Theorem 5.2].

3 Commutators

Let Aj and Bj , j = 1, . . . , n be operators satisfying

[Aj , Ak ] = [Bj , Bk ] = [Aj , Bk ] = 0, j = k

and

[Aj , Bj ] = I, j = 1, . . . , n.

Let P and Q be polynomials of n variables and write A = (A1 , . . . , An ) and B =

(B1 , . . . , Bn ). Then
 1
Q(A)P (B) = P (α) (B)Q(α) (A), (3.1)
α!
|α|≥0

where α = (α1 , . . . , αn ) are multiindices and |α| = α1 + · · · + αn and α! =

α1 ! . . . αn !, see [7, 8].
The assumptions are satisfied, if one takes Aj = ∂z∂ j and Bj = zj the
multiplication operator. The inspiration for this comes from quantum mechanics,
where the annihilation operator Aj can be represented by the differentiation with
324 F. Haslinger

respect to zj on A2 (Cn , e−|z| ) and its adjoint, the creation operator Bj , by the
2

multiplication by zj , both operators being unbounded densely defined (see [1, 2]).
One can show that A2 (Cn , e−|z| ) with this action of the Bj and Aj is an irreducible
2

representation M of the Heisenberg group; by the Stone-von Neumann theorem it

is the only one up to unitary equivalence. Physically M can be thought of as the
Hilbert space of a harmonic oscillator with n degrees of freedom and Hamiltonian
operator


n
1
H = (Aj Bj + Bj Aj ).
2
j =1

Remark 3.1 If we apply (3.1) for the one-dimensional case of Lemma 2.3, we get
 m 
 2
∗
]u(z) u(z) e−|z| dλ(z) = e−|z| dλ(z),
2 2
[pm , pm ()
pm u(z)
C =1 C

which coincides with (2.3).

In the following we consider C2 and choose p1 and p2 to be polynomials of degree
2 in 2 variables.
Theorem 3.2 Let p1 , p2 be polynomials of degree 2. Suppose that

p2(e1 )∗ p1(e1 ) = ± p1(e2 )∗ p2(e2 ) , p1(e1 )∗ p2(e1 ) = ± p2(e2 )∗ p1(e2 ) , (3.2)

where (e1 ) and (e2 ) denote the derivatives with respect to z1 and z2 respectively. In
addition suppose that for all derivatives (α) of order 2 we have

(α)∗ (α)
pj pk = δj,k cj,α j, k = 1, 2, (3.3)

where
 1  1
C1 = c1,α > 0 and C2 = c2,α > 0.
α! α!
|α|=2 |α|=2

Then

1 
2
u2 ≤ ([pk , pj∗ ]uj , uk ), (3.4)
min(C1 , C2 )
j,k=1

2
for any (1, 0)-form u = j =1 uj dzj with polynomial components.
The Generalized ∂-Complex on the Segal–Bargmann Space 325

Proof Using (3.1) we obtain

 1 (α)∗ (α)
([pk , pj∗ ]uj , uk ) = (p pk uj , uk ).
α! j
|α|≥1

Now we use (3.2) and get for the first order derivatives


2
(e1 )∗ (e1 ) (e2 )∗ (e2 )
[(pj pk uj , uk ) + (pj pk uj , uk )]
j,k=1
(e )∗ (e ) (e )∗ (e )
=(p1 1 p1 1 u1 , u1 ) + (p2 1 p2 1 u2 , u2 )
(e )∗ (e ) (e )∗ (e )
± (p2 2 p1 2 u1 , u2 ) ± (p1 2 p2 2 u2 , u1 )
(e )∗ (e ) (e )∗ (e )
+ (p1 2 p1 2 u1 , u1 ) + (p2 2 p2 2 u2 , u2 )
(e )∗ (e ) (e )∗ (e )
± (p2 1 p1 1 u1 , u2 ) ± (p1 1 p2 1 u2 , u1 )
(e ) (e ) (e ) (e )
=(p1 1 u1 ± p2 1 u2 , p1 1 u1 ± p2 1 u2 )
(e ) (e ) (e ) (e )
+ (p1 2 u1 ± p2 2 u2 , p1 2 u1 ± p2 2 u2 )
(e ) (e ) (e ) (e )
=p1 1 u1 ± p2 1 u2 2 + p1 2 u1 ± p2 2 u2 2 .

For the second order derivatives we obtain


2  1 (α)∗ (α)
(p pk uj , uk ) = C1 u1 2 + C2 u2 2 .
α! j
j,k=1 |α|=2

Hence we get


2
([pk , pj∗ ]uj , uk ) =C1 u1 2 + C2 u2 2
j,k=1
(e ) (e ) (e ) (e )
+ p1 1 u1 ± p2 1 u2 2 + p1 2 u1 ± p2 2 u2 2 ,

which gives (3.4).

In a similar way one shows the following.
Theorem 3.3 Let p1 , p2 be polynomials of degree 2. Suppose that

(e )∗ (e ) (e )∗ (e ) (e )∗ (e ) (e )∗ (e )
p2 1 p1 1 = ± p1 1 p2 1 , p1 2 p2 2 = ± p2 2 p1 2 , (3.5)
326 F. Haslinger

where (e1 ) and (e2 ) denote the derivatives with respect to z1 and z2 respectively. In
addition suppose that for all derivatives (α) of order 2 we have

(α)∗ (α)
pj pk = δj,k cj,α j, k = 1, 2, (3.6)

where
 1  1
C1 = c1,α > 0 and C2 = c2,α > 0.
α! α!
|α|=2 |α|=2

Then

1 
2
u2 ≤ ([pk , pj∗ ]uj , uk ),
min(C1 , C2 )
j,k=1

2
for any (1, 0)-form u = j =1 uj dzj with polynomial components.
Finally we exhibit some examples, where conditions (3.2) and (3.3), or (3.5)
and
n (3.6) are∗ checked. Examples (a) and (c) are taken from [4], where
j,k=1 ([pk , pj ]uj , uk ) was directly computed.
Example
∂2 ∂2 ∂2
(a) We take p1 = and p2 = + . Then p1∗ (z) = z1 z2 and p2∗ (z) =
∂z1 ∂z2 ∂z12 ∂z22
z12 + z22 and we see that (3.2) and (3.3) are satisfied:

p2(e1 )∗ p1(e1 ) = p1(e2 )∗ p2(e2 ) , p1(e1 )∗ p2(e1 ) = p2(e2 )∗ p1(e2 ) ,

and we obtain


2
([pk , pj∗ ]uj , uk )
j,k=1
 
2 2
∂u1 ∂u2 ∂u1 ∂u2
e−|z| dλ,
2
= |u1 |2 + 4|u2 |2 + +2 + +2
C2 ∂z1 ∂z2 ∂z2 ∂z1

2
for u = j =1 uj dzj with polynomial components.
2 ∂2 ∂2
(b) Taking p1 = i ∂z∂1 ∂z2 and p2 = ∂z 2 + ∂z2 we have that p1∗ (z) = −iz1 z2 and
1 2
p2∗ (z) = z12 + z22 and that (3.2) and (3.3) are satisfied:

(e )∗ (e ) (e )∗ (e ) (e )∗ (e ) (e )∗ (e )
p2 1 p1 1 = −p1 2 p2 2 , p1 1 p2 1 = −p2 2 p1 2 ,
The Generalized ∂-Complex on the Segal–Bargmann Space 327

and we obtain


2
([pk , pj∗ ]uj , uk )
j,k=1
 
2 2
∂u1 ∂u2 ∂u1 ∂u2
e−|z| dλ,
2
= |u1 |2 + 4|u2 |2 + + 2i + + 2i
C2 ∂z1 ∂z2 ∂z2 ∂z1

2
for u = j =1 uj dzj with polynomial components.
∂2
(c) Let pk = ,k = 1, 2. Then pj∗ (z) = zj2 , j = 1, 2 and we see that (3.5)
∂zk2
and (3.6) are satisfied and we have


2 
2 
2
∂uj
([pk , pj∗ ]uj , uk ) = (2δj,k uj , uk ) + (4δj k zj , uk )
∂zk
j,k=1 j,k=1 j,k=1

2 ;

;
; ∂uj ;2
= 2u + 4
2 ; ;
; ∂zj ; .
j =1

∂2 ∂2
(d) For p1 = + ∂
and p2 = ∂
+ we have p1∗ (z) = z12 + z2 and p2∗ (z) =
∂z12 ∂z2 ∂z1 ∂z22
z1 + z22 and we see that (3.2) and (3.5) are not satisfied. In particular,
; ; ; ;

2
; ∂u1 ;2 ; ∂u2 ;2
∗ ;
([pk , pj ]uj , uk ) =3(u1  + u2  ) + 4 ;
2 2 ; ;
+; ;
∂z1 ; ∂z2 ;
j,k=1

+ 2(u1 , z2 u2 ) + 2(z1 u1 , u2 ) + 2(u2 , z1 u1 ) + 2(z2 u2 , u1 )

2
for u = j =1 uj dzj with polynomial components.

Acknowledgments The author thanks the referees for several useful suggestions.
This project was partially supported by the Austrian Science Fund (FWF) project P28154.

References

1. L.D. Fadeev, O.A. Yakubovskii, Lectures on Quantum Mechanics for Mathematics Students.
Student Mathematical Library, vol. 47 (American Mathematical Society, Providence, 2009)
2. G.B. Folland, Harmonic Analysis in Phase Space. Annals of Mathematics Studies, vol. 122
(Princeton University Press, Princeton, 1989)
3. F. Haslinger, The ∂-Neumann Problem and Schrödinger Operators. de Gruyter Expositions in
Mathematics, vol. 59 (Walter de Gruyter, Berlin, 2014)
328 F. Haslinger

4. F. Haslinger, The ∂-complex on the Segal-Bargmann space. Ann. Polon. Mat. 123, 295–317
(2019)
5. D.J. Newman, H.S. Shapiro, Certain Hilbert spaces of entire functions. Bull. Am. Math. Soc.
72, 971–977 (1966)
6. D.J. Newman, H.S. Shapiro, Fischer spaces of entire functions, in Entire Functions and Related
Parts of Analysis. Proceedings of Symposium in Pure Mathematics La Jolla, 1966 (American
Mathematical Society, Providence, 1968), pp. 360–369
7. D.G. Quillen, On the representation of Hermitian forms as sums of squares. Invent. Math. 5,
237–242 (1968)
8. F. Treves, Linear Partial Differential Equations with Constant Coefficients (Gordon and Breach,
New York, 1966)
The Inverse Characteristic Polynomial
Problem for Trees

Charles R. Johnson and Emma Gruner

Abstract It is known that any real polynomial is attained as the characteristic

polynomial of a real combinatorially symmetric matrix, whose graph is either a
path or a star. We conjecture that the same is true for any tree (this is so for complex
characteristic polynomials and complex matrices). Here, we constructively prove
a very large portion of this conjecture by a method that mates the graph of the
polynomial with a notion of balance of the tree relative to a choice of root for
the tree. Included is the first constructive proof for the case of the path, as well
as the case of any tree on fewer than ten vertices. It also includes the known case
of polynomials with distinct real roots for any tree (in a new way). This work is
motivated by, and lies in contrast to, the considerable study of possible multiplicity
lists for the eigenvalues of real symmetric matrices, whose graph is a tree.

Keywords Inverse eigenvalue problem · Characteristic polynomial · Partial

fractions · Graph theory · Matrix theory

Mathematics Subject Classification (2010) Primary 15A29, 05C50; Secondary

05C05

1 Introduction

During the last 20+ years, a theory of the possible multiplicities of the eigenvalues,
of Hermitian matrices with a given graph, has developed. This is the most developed,
and most interesting, in the case of trees, but much is now known for general graphs.

C. R. Johnson ()
The College of William and Mary, Williamsburg, VA, USA
e-mail: [email protected]
E. Gruner
Gettysburg College, Gettysburg, PA, USA
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 329

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_17
330 C. R. Johnson and E. Gruner

Recently the theory has expanded to geometric multiplicities for the eigenvalues of
a general (combinatorially symmetric) matrix; there are strong similarities and some
notable differences. Though ongoing, much of this work is summarized in the recent
book [6].
The question then naturally arises: “What about algebraic multiplicities of the
eigenvalues of a matrix with a given graph?” Here, the field makes a big difference.
In the case of the complex numbers, there is a complete answer because of the
“additive inverse eigenvalue problem.” Recall that the diagonal entries of a matrix
are independent of its undirected graph. Over the complex numbers, for a given
n-by-n matrix A and given desired complex numbers λ1 , λ2 , . . . , λn , there is a
diagonal matrix D such that A + D has eigenvalues λ1 , λ2 , . . . , λn [3].
This is not true over the reals. There it is natural to ask which real polynomials
occur for real matrices with a given undirected graph. There is good reason to guess
that any real polynomial may occur when the graph is connected. But, this is not
easy to prove. Initially, it is natural to consider trees. Recently, it has been shown [2]
that for the path (tridiagonal matrices), any real polynomial can occur. But, the proof
is existential. It is shown that a real matrix may be constructed for a given monic,
degree n real polynomial when there is a degree n − 1 polynomial with a certain
relationship to the former. Existence may be shown, but construction is difficult.
For the star (all vertices pendent from a single vertex), it has also recently been
shown [5] that any real polynomial occurs over R. The proof is constructive and
much more is shown, just for the star. We benefit here from the technology therein,
although the technology does not easily transfer to other trees.
Here, we consider general trees and real matrices; for a tree T , R(T ) denotes
the set of real matrices whose graph is T (not necessarily symmetric, but combi-
natorially symmetric, and no restriction besides reality on the diagonal entries). A
method is developed that, for each tree, realizes many characteristic polynomials in
R(T ), and for many trees, realizes all real characteristic polynomials. The method
is constructive and gives all real polynomials for the path (the first time that has
occurred), as well as many other trees. The smallest tree for which not all real
polynomials are realized has ten vertices, and this is not because of the number
of vertices, but because of a notion of balance of the tree.

2 Preliminaries

We assume that the reader is familiar with the basic terminology from linear algebra
and graph theory: see [4] and [6]. However, we shall briefly elaborate on the concept
of the graph of a matrix discussed in the introduction.
Given an n-by-n matrix A = (aij ) over a field F, we let G (A) denote the graph
of A. This graph consists of vertices {1, 2, . . . , n} with an edge {i, j }, i = j , if and
only if aij = 0. For G (A) to be undirected, A must be combinatorially symmetric;
that is, aij = 0 if and only if aj i = 0. This paper will only be concerned with real
matrices; given an undirected graph G, we let R(G) denote all real matrices whose
The Inverse Characteristic Polynomial Problem for Trees 331

graph is G. If v is a vertex of G (A), then we let A(v) denote the principal submatrix
of A resulting from the deletion of the row and column corresponding to v.
The matrices we construct will all be of a specific form. Suppose we have a tree
T on n vertices, and let v be any vertex with degree s. If s ≥ 1, let u1 , u2 , . . . , us be
the neighboring vertices of v. Additionally, let T1 , T2 , . . . , Ts be the branches of T
at v, in which each Ti contains the vertex ui , and let 1 , 2 , . . . , s be the respective
number of vertices in T1 , T2 , . . . , Ts . From this point forward, we will refer to the
number of vertices in a tree or a branch of a tree as its weight.
We say that A ∈ R(T ) is centered at v if it is 1-by-1, or if it has the following
form:
⎡ ⎤
b v1 · · · vs
⎢e1,1 A1 0 0⎥
⎢ ⎥
A=⎢ . .. ⎥ .
⎣ .. 0 . . . . ⎦
e1,s 0 · · · As

In this form, b ∈ R, e1,i = e1 ∈ Ri , and vi = (ai e1,i )T , with ai ∈ R \ {0}. (Note
that e1 represents the basic unit vector with a 1 in the first entry and 0’s elsewhere.)
Each Ai is a matrix in R(Ti ) that is centered at ui .
Define A to be pseudosymmetric if it is 1-by-1, or if it has the above form and
each Ai is diagonally similar to a symmetric matrix. We will usually be constructing
matrices that have this property.
The neighbors formula [6] provides a convenient recursive representation of the
characteristic polynomial for a matrix in this form. Let pA denote the characteristic
polynomial of the matrix A. If A = [b], then pA = (x − b). Otherwise, we have:

/
s 
s /
s
pA = (x − b) pAi − ai pAi (ui ) pAj
i=1 i=1 j =1,j =i


s /
s
= (x − b)pA(v) − ai pAi (ui ) pAj . (1)
i=1 j =1,j =i

We define the empty product, and the characteristic polynomial of the empty
matrix, to be 1.
Given any vertex v of T and any matrix A ∈ R(T ), there is a diagonal and/or
permutation similarity that takes A to a matrix centered at v. Since the characteristic
polynomial is similarity invariant, if we are trying to realize a particular polynomial
for a particular tree, we can take our matrix to be centered at whatever vertex is most
convenient for us.
Finally, we will provide an overview of the algebraic technique known as the
partial fraction decomposition, or PFD.
The following theorem was adapted from a more general statement in [1]. The
proof is straightforward and is thus omitted.
332 C. R. Johnson and E. Gruner

E
Theorem 2.1 ([1]) Let g(x) = ki=1 (x − γi ), in which γ1 , γ2 , . . . , γk are real
E
numbers such that γ1 > γ2 > · · · > γk . Let Qi (x) = kj =1,j =i (x − γj ). Finally,
let f (x) be a real polynomial of degree less than k. Then we have

f (x) a1 a2 ak
= + + ···+
g(x) x − γ1 x − γ2 x − γk

in which each ai is a real number determined uniquely by the expression ai =

f (γi )/Qi (γi ).
Also of interest to us is the following lemma. The case in which the polynomials
f and g are both monic is well-known, but the proof is easily generalized.
Lemma 2.2 Let g(x), Qi (x) and f (x) be defined as in Theorem 2.1. The poly-
nomial f (x) has exactly k − 1 distinct real roots that strictly interlace the roots
of g(x) if and only if the coefficients of the partial fraction decomposition fg(x)
(x)
,
denoted a1 , a2 , . . . , ak , are either all positive or all negative.
Proof Consider the coefficients ai and ai+1 for i ∈ {1, 2, . . . , k − 1}. By the
previous theorem, we have ai = f (γi )/Qi (γi ) and ai+1 = f (γi+1 )/Qi+1 (γi+1 ).
Observe that Qi (γi ) and Qi+1 (γi+1 ) will always be nonzero and of opposite signs;
therefore, ai and ai+1 will be nonzero and of the same sign if and only if f (γi ) and
f (γi+1 ) are nonzero and of opposite signs. This will occur if and only if neither γi
nor γi+1 are roots of f , but f has an odd number of roots between γi and γi+1 .
Therefore, if f has k − 1 distinct real roots that strictly interlace the numbers
γ1 , γ2 , . . . , γk , then the ai ’s must either be all positive or all negative. Conversely, if
the ai ’s are all of the same sign, then f must have at least one real root lying strictly
between each consecutive pair of γi ’s. This means that f has at least k − 1 distinct
real roots; however, since deg f (x) < k, f can have at most k − 1 roots, and the
result follows.
Finally, we provide the following result, which combines the techniques of partial
fractions and the polynomial division algorithm.
Lemma 2.3 Let p be a real monic polynomial of degree n, and let g(x) =
E n−1
i=1 (x − γi ), in which γ1 , γ2 , . . . , γn−1 are real
E
numbers such that γ1 > γ2 >
· · · > γn−1 and p(γi ) = 0 for all i. Let Qi (x) = n−1 j =1,j =i (x − γj ). Then we have

p(x) a1 a2 an−1
= (x − b) − − − ··· − ,
g(x) x − γ1 x − γ2 x − γn−1

with b ∈ R and ai = −p(γi )/Qi (γi ). Furthermore, ai and ai+1 are of the same
sign for any i ∈ {1, 2, . . . , n − 2} if and only if p has an odd number of roots lying
strictly between γi and γi+1 .
Proof By the division algorithm for polynomials, we can write

p(x) = q(x)g(x) − r(x),

The Inverse Characteristic Polynomial Problem for Trees 333

in which q(x) and r(x) are real polynomials and deg r(x) < deg g(x) = n −
1. (While the division algorithm normally allows for the possibility of a zero
remainder, this cannot happen in our case because p and g are relatively prime.)
Since p and g are both monic and they differ by only one degree, we know that
q(x) = x − b for some unique real number b. So we may write

p(x) (x − b)g(x) − r(x) r(x)

= = (x − b) − .
g(x) g(x) g(x)

Since deg r(x) < deg g(x), we can apply Theorem 2.1 to write

r(x) a1 a2 an−1
= + + ···+ ,
g(x) x − γ1 x − γ2 x − γn−1

with ai = r(γi )/Qi (γi ). Since each γi is a root of g, we have r(γi ) = −p(γi ) for
all i ∈ {1, 2, . . . , n − 1}, so we can just as well write ai = −p(γi )/Qi (γi ). The last
statement of the lemma follows from a proof similar to that of Lemma 2.2.

3 Main Results

The ease of constructing a matrix for a particular polynomial and tree hinges
largely on the possible signs of the coefficients in a partial fraction decomposition.
Therefore, we would like to find a way to classify a polynomial according to this
characteristic.
Suppose we have a monic real polynomial p of degree n. As x ranges from −∞
to +∞, p(x) undergoes n − d strict sign changes, in which d is an even integer at
least 0 and at most n (if n is even) or n − 1 (if n is odd). We shall call this integer d
the root deficiency of p.
The root deficiency of p is directly related to its complete factorization over the
real polynomials. Suppose we have
t
p = w1s1 w2s2 · · · wαsα q1t1 q2t2 · · · qββ ,

in which w1 , . . . , wα are distinct linear polynomials, q1 , . . . , qβ are distinct irre-

ducible quadratic polynomials, and each si and ti is a positive integer. (The fact that
such a representation of p exists, and is unique, is a consequence of the fundamental
theorem of algebra.) Define ρ(si ) as follows:

si , if si is even
ρ(si ) =
si − 1, if si is odd.
α β
We then have d = i=1 ρ(si ) + j =1 2tj .
334 C. R. Johnson and E. Gruner

Observe that d = 0 if and only if p has n distinct linear factors. On the other
hand, d has maximal value if and only if 1) p is composed entirely of irreducible
quadratics, or 2) at most one si is odd.
Let us also define a signed root to be a root at which p undergoes a strict sign
change. These are precisely the roots whose corresponding linear terms have odd
exponents, and there are n − d of them.
Example 3.1 The polynomial p1 (x) = x 3 + x 2 − 4x − 4 = (x + 1)(x − 2)(x + 2)
has root deficiency 0. The signed roots of p1 are −1, 2, and −2.
Example 3.2 The polynomial p2 (x) = x 2 (x −5)3 (x +1)(x 2 +1) has root deficiency
6. The signed roots of p2 are 5 and −1.
Let p be a monic real polynomial of degree n, and let γ1 > γ2 > · · · > γn−1
be real numbers with p(γi ) = 0 for each i ∈ {1, 2, . . . , n − 1}. Suppose
E we apply
the combined division algorithm/PFD procedure to p and g = n−1 i=1 (x − γi ) as
in Lemma 2.3. The potential positive/negative breakdown of the ai ’s we obtain is
connected to the root deficiency of p.
Proposition 3.3 Let p be a monic real polynomial of degree n and root deficiency
d. Let k1 ≥ k2 be nonnegative integers with k1 + k2 = n − 1. We can choose real
p(x)
numbers γ1 > γ2 > · · · > γn−1 such that the representation of En−1 shown in
i=1 (x−γi )
Lemma 2.3 yields k1 negative ai ’s and k2 positive ai ’s if and only if k2 ≥ (d − 2)/2.
Proof For the “only if” direction, assume that we have a selection of (n − 1) γi ’s
that yields k1 negative ai ’s and k2 positive ai ’s for a particular polynomial p, with
k1 ≥ k2 . We shall prove that d must be at most 2k2 +2, implying that k2 ≥ (d −2)/2.
Let z = k1 − k2 . We claim that we must have at least z − 1 distinct consecutive
pairs of negative ai ’s. We consider two consecutive pairs distinct if they differ in at
least one element: for example, the set {a1 , a2 , a3 } consists of two consecutive pairs.
If we ignore all the positive ai ’s, then we have k1 negative ai ’s, which consists
of k1 − 1 distinct consecutive pairs. Each positive ai that we introduce can break
up at most one of these consecutive pairs (none if it is terminal or adjacent to an
already introduced ai ). When we have introduced all k2 of these ai ’s, we are left
with at least (k1 − 1) − k2 = z − 1 consecutive pairs of negative ai ’s. Since each
consecutive pair of same-sign ai ’s necessitates at least one signed root of p between
the corresponding γi ’s (see Lemma 2.3), p must have at least z − 1 signed roots.
Recall that p has n−d distinct signed roots, so z−1 ≤ n−d. Substituting k1 −k2
for z and k1 + k2 + 1 for n, we determine that d ≤ 2k2 + 2, or k2 ≥ (d − 2)/2, as
claimed.
For the “if” direction, suppose we have k1 and k2 in mind such that k1 +k2 = n−1
and k2 ≥ (d − 2)/2. When we choose γi ’s in the following steps, we assume that
no γi is a root of p, signed or “unsigned.”
Let R1 , R2 , . . . , Rn−d be the distinct signed roots of p, with R1 > R2 > · · · >
Rn−d . If n − d > 0, then take R = R1 ; otherwise, take R to be any real number.
If k2 = k1 = (n − 1)/2, then choose all γi ’s to be larger than R. Then
by Lemma2.3, the ai ’s will alternate in sign, and with n − 1 being even we
The Inverse Characteristic Polynomial Problem for Trees 335

will have equal numbers of positive and negative ai ’s. Otherwise, choose only
γ1 , γ2 , . . . , γ2k2 +1 to be larger than R. This choice will ensure that the correspond-
ing ai ’s alternate in sign. Note that a1 = −p(γ 1)
Q1 (γ1 ) , as defined in Lemma 2.3; since
both p(γ1 ) and Q1 (γ1 ) are positive, a1 will be negative, as will a3 , a5 , . . . , a2k2 +1 .
Likewise, a2 , a4 , . . . , a2k2 will be positive. In total, we will have k2 +1 negative ai ’s
and k2 positive ai ’s.
Since we now have all the positive ai ’s we want, we’d like the remaining ai ’s, of
which there are (k1 + k2 ) − (2k2 + 1) = k1 − k2 − 1 = z − 1, to all be negative.
But since (d − 2)/2 ≤ k2 , we have that z − 1 ≤ n − d. Therefore, we can choose
the remaining γi ’s such that

R1 > γ2k2 +2 > R2 > γ2k2 +3 > R3 > · · · > Rz−1 > γn−1 .

Since exactly one signed root occurs between γi and γi+1 for i ∈ {2k2 + 1, 2k2 +
2, . . . , n−2}, by Lemma 2.3 a sign change will not occur between the corresponding
ai and ai+1 , so the remaining ai ’s will all be negative, as desired.
Arguably the most relevant previous result to our work is the following theorem,
first proved by Leal-Duarte in [7]. We present a slightly different proof, but the basic
argument is the same.
Theorem 3.4 ([7]) Let T be a tree on n vertices, and let v be any vertex of T .
Suppose p and g are monic polynomials of degrees n and n − 1 respectively,
that both polynomials have all distinct, real roots, and that the roots of g strictly
interlace the roots of p. Then there exists a matrix A ∈ R(T ), centered at v, such
that A has characteristic polynomial p and A(v) has characteristic polynomial g.
Furthermore, this matrix is diagonally similar to a symmetric matrix.
Proof We proceed by induction on n. Suppose first that n = 1; then g = 1, p =
(x − b) for some real number b, and A = [b] is the matrix we seek. Note that A
is diagonally similar to a symmetric matrix: namely, itself, via the similarity matrix
D = [1]. Now suppose n > 1, and the claim holds for t with n > t ≥ 1. Then
v must have degree s ≥ 1; let u1 , u2 , . . . , us be the neighboring vertices of v, in
which each ui is E
contained in the branch Ti with weight i .
Suppose g = n−1 i=1 (x − μi ) for μ1 > μ2 > · · · > μn−1 , and write
! "
p(x) a1 a2 an−1
= (x − b) − + + ···+ ,
g(x) x − μ1 x − μ2 x − μn−1

as in Lemma 2.3.
Since p and g have strictly interlacing roots, we can use the last statement of
Lemma 2.3 to conclude that all ai ’s are of the same sign. In this case, they are all
positive, since a1 = −p(μ1 )/Q1 (μ1 ) is positive.
Suppose we partition the n − 1 terms of the PFD into s partial sums, with the
ith partial sum containing i terms of the original expression—the exact grouping
336 C. R. Johnson and E. Gruner

does not matter here. Then we combine the terms within each partition into a single
rational expression hi (x)/Pi (x), with deg Pi (x) = i and deg hi (x) < i . We now
have:
! "
p(x) h1 (x) h2 (x) hs (x)
= (x − b) − + + ··· + .
g(x) P1 (x) P2 (x) Ps (x)

By the uniqueness of the partial fraction decomposition, we know that if we

performed a PFD of any individual expression hi (x)/Pi (x), we would simply
regenerate i terms of our original PFD. We already established that these terms
all have positive numerators; therefore by Lemma 2.2 each hi (x) has i − 1 real
roots that strictly interlace those of Pi (x).
Furthermore, we can write each hi (x) as αi ĥi (x), in which αi is the leading
coefficient of hi (x) and ĥi (x) is the monic polynomial obtained by dividing hi (x)
by αi . Observe that αi is just the sum of the numerators of the partial fraction terms
that were consolidated to yield hi (x)/Pi (x); since these numerators are all positive,
so is αi .
Also observe that ĥi has the same roots as hi , which, as just discussed, strictly
interlace those of Pi . Since deg Pi (x) = i < n, by the inductive hypothesis
there exists a matrix Ai ∈ R(Ti ), centered at ui , such that Ai has characteristic
polynomial Pi and Ai (ui ) has characteristic polynomial ĥi . Construct such an Ai
for each pair of ĥi and Pi .
Now, let
⎡ ⎤
b (α1 e1,1 )T · · · (αs e1,s )T
⎢e1,1 A1 0 0 ⎥
⎢ ⎥
A=⎢ . .. .. ⎥,
⎣ .. 0 . . ⎦
e1,s 0 ··· As

which is a matrix in R(T ) and centered at v.

By the neighbors formula (1), we have:

/
s 
s /
s
pA = (x − b) pAi − ai pAi (ui ) pAj
i=1 i=1 j =1,j =i


s /
s
= (x − b)P1 (x) · · · Ps (x) − αi ĥi (x) Pj (x)
i=1 j =1,j =i


s /
s
= (x − b)g(x) − αi ĥi (x) Pj (x) = p(x).
i=1 j =1,j =i
The Inverse Characteristic Polynomial Problem for Trees 337

Furthermore, the matrix A(v) is equal to

⎡ ⎤
A1 0 · · · 0
⎢ 0 A2 0 0 ⎥
⎢ ⎥
A(v) = ⎢ . .. ⎥,
⎣ .. 0 . . . . ⎦
0 0 ··· As

which has characteristic polynomial

/
s
pA(v) = pAi = P1 (x) · · · Ps (x) = g(x),
i=1

as claimed.
By our inductive hypothesis, we have that for each i ∈ {1, 2, . . . , s}, Ai =
Di−1 Bi Di , in which Di is a diagonal matrix of appropriate size and Bi is symmetric.
Since each αi is positive, we can construct the diagonal matrix
⎡ ⎤
1 0 0 ··· 0
⎢0 √α1 D1 0 ··· 0 ⎥
⎢ √ ⎥
⎢0 ··· ⎥
D=⎢ 0 α 2 D2 0 ⎥.
⎢. . . .. .. ⎥
⎣ .. .. .. . . ⎦
√
0 0 0 ··· αs Ds

Note that as long as D1 , D2 , . . . , Ds are constructed in this same inductive

manner, they will each have the (1, 1) entry equal to 1. Therefore, one can verify
that
⎡ √ √ ⎤
b ( α1 e1,1 )T · · · ( αs e1,s )T
√
⎢ α1 e1,1 ⎥
⎢ B1 0 0 ⎥
DAD −1 = ⎢ .. .. .. ⎥,
⎣ . 0 . . ⎦
√
αs e1,s 0 ··· Bs

which is symmetric. Therefore, our constructed matrix A is diagonally similar to a

symmetric matrix, as claimed.
We note that the original theorem and proof in [7] produce a matrix A that is
itself symmetric, instead of just symmetric by similarity. However, we amended it
to keep consistency with the canonical form we use.
We are now ready to prove our main result.
Theorem 3.5 Let p be a monic real polynomial of degree n and root deficiency
d, and let T be a tree on n vertices. Let k = n − 1 and  = (d − 2)/2. Then p
has a pseudosymmetric realization A ∈ R(T ) if there exists a vertex v of T with
branches T1 , . . . , Ts with respective weights 1 , . . . , s such that some partial sum
of {1 , . . . , s } ∈ [, k − ].
338 C. R. Johnson and E. Gruner

Proof Let v be a vertex of T that meets the above specifications. Without loss of
generality, assume that 1 + 2 + · · · + t = C, with t ≤ s and C ∈ [, k − ]. (Note
that since d ≤ n, we know  < k/2 and therefore the range [, k−] contains at least
two integers.) Since k − C would also be in [, k − ], we can also assume without
loss of generality that CE≤ k − C. By Proposition 3.3, we can choose γ1 , γ2 , . . . , γk
such that, if we let g = ki=1 (x − γi ) and represent p/g as in Lemma 2.3, we obtain
C positive coefficients and k − C negative coefficients among the ai ’s.
Relabeling the γi ’s and ai ’s if necessary, write

p(x)
=(x − b)
g(x)
! "
a1 a2 aC aC+1 ak
− + + ··· + + + ··· + ,
x − γ1 x − γ2 x − γC x − γC+1 x − γk

such that a1 , . . . , aC are positive and aC+1 , . . . , ak are negative.

Now the proof resembles that of the last theorem. We can partition the first C
terms of the PFD into t partial sums, and then the last k − C terms into s − t
partial sums, with the ith partial sum containing i terms of the original expression.
Then we combine the terms within each partition into a single rational expression
hi (x)/Pi (x), with deg Pi (x) = i and deg hi (x) < i . We now have:
! "
p(x) h1 (x) hs (x)
= (x − b) − + ··· + .
g(x) P1 (x) Ps (x)

If we perform a PFD of any individual expression hi (x)/Pi (x), we would, again,

simply regenerate i terms of our original PFD. Because of how we combined the
terms originally, the terms we regenerate will either have all positive numerators (if
i ≤ t) or all negative numerators (if i > t). Either way, by Lemma 2.2 each hi (x)
has i − 1 real roots that strictly interlace those of Pi (x).
We can again write each hi (x) as αi ĥi (x), in which αi is the leading coefficient of
hi (x) and ĥi (x) is the monic polynomial obtained by dividing hi (x) by αi . Again,
ĥi has the same roots as hi , which strictly interlace those of Pi ; by Theorem 3.4
there exists a matrix Ai ∈ R(Ti ), centered at ui , such that Ai has characteristic
polynomial Pi , Ai (ui ) has characteristic polynomial ĥi , and Ai is diagonally similar
to a symmetric matrix. Construct such an Ai for each pair of ĥi and Pi .
Now, let
⎡ ⎤
b (α1 e1,1 )T · · · (αs e1,s )T
⎢e1,1 A1 0 0 ⎥
⎢ ⎥
A=⎢ . .. .. ⎥.
⎣ .. 0 . . ⎦
e1,s 0 ··· As

Note that by construction, we have A ∈ R(T ) and it is centered at v.

The Inverse Characteristic Polynomial Problem for Trees 339

By the neighbors formula, we have:

/
s 
s /
s
pA = (x − b) pAi − ai pAi (ui ) pAj
i=1 i=1 j =1,j =i


s /
s
= (x − b)P1 (x) · · · Ps (x) − αi ĥi (x) Pj (x)
i=1 j =1,j =i


s /
s
= (x − b)g(x) − αi ĥi (x) Pj (x) = p(x).
i=1 j =1,j =i

Since each Ai is diagonally similar to a symmetric matrix, A is pseudosymmetric,

as claimed.

4 Constructive Examples

We present two examples that demonstrate the use of this algorithm. The first
example uses a polynomial with real, distinct roots, which can be realized using
the method presented in Theorem 3.4. The second example uses a polynomial with
repeated and complex roots, which can be realized using the method presented in
Theorem 3.5.
Example 4.1 Let p1 be the polynomial

p1 (x) = (x + 3)(x + 2)(x + 1)(x)(x − 1)(x − 2)(x − 3)(x − 4),

and let T be the following 8-vertex tree:

We wish to find a matrix A ∈ R(T ) with characteristic polynomial p1 .

(Note: in the following calculations, all real numbers are rounded to the nearest
thousandth.)
In order to use the algorithm presented in Theorem 3.4, we first need to choose a
“central” vertex for our matrix A and a characteristic polynomial for A(v).
340 C. R. Johnson and E. Gruner

Fig. 1 Our tree T with

highlighted vertex v

Since there are no restrictions on the central vertex in the real, distinct root case,
let v be the far left vertex indicated in Fig. 1.
We need our characteristic polynomial for A(v) to be of degree 7 and have real,
distinct roots that strictly interlace those of p1 . Any such polynomial will suffice, so
let us choose

g(x) = (x + 2.5)(x + 1.5)(x + 0.5)(x − 0.5)(x − 1.5)(x − 2.5)(x − 3.5).

Now, by Lemma 2.3 we can write

 
p1 0.733 1.015 1.154
=(x − 0.5) − + +
g x + 2.5 x + 1.5 x + 0.5
 
1.196 1.154 1.015 0.733
− + + + .
x − 0.5 x − 1.5 x − 2.5 x − 3.5

Observe that all the numerators in the PFD are positive, as the algorithm claims.
In this case T has only one branch at v - call it T1 - so we may combine all partial
fraction terms into a single rational expression:
p1
= (x − 0.5)
g
7x 6 − 21x 5 − 48.125x 4 + 131.25x 3 + 60.183x 2 − 129.938x − 6.152
− .
g

By pulling out the leading coefficient of the PFD numerator, and factoring the
resulting monic polynomial (call it ĥ1 ), we have:
p1
= (x − 0.5)
g
7(x + 2.261)(x + 1.144)(x + 0.046)(x − 1.046)(x − 2.144)(x − 3.261)
− .
(x + 2.5)(x + 1.5)(x + 0.5)(x − 0.5)(x − 1.5)(x − 2.5)(x − 3.5)

Observe that ĥ1 is of degree 6 and has real, distinct roots that strictly interlace
the roots of g. Letting u denote the sole neighboring vertex of v, we can construct a
The Inverse Characteristic Polynomial Problem for Trees 341

matrix A1 in R(T1 ), centered at u, such that A1 has characteristic polynomial g and

A1 (u) has characteristic polynomial ĥ1 . Our desired matrix A will be of the form:
! "
0.5 (7e1 )T
A= .
e1 A 1

One can verify with the neighbors formula (1) that A has characteristic polyno-
mial p1 .
Now, we can follow the same process to construct A1 . Dividing g by ĥ1 as in
Lemma 2.3 yields:
 
g 0.458 0.588 0.641
=(x − 0.5) − + +
ĥ1 x + 2.261 x + 1.144 x + 0.046
 
0.641 0.588 0.458
− + + .
x − 1.046 x − 2.144 x − 3.261

Now, T1 has two branches at u, with weights 5 and 1; call these branches S1 and
S2 . We would like to combine the new PFD into two rational expressions such that
the denominators have degrees 5 and 1. There are several ways we could do this, but
let’s consolidate the first five terms of the PFD and leave the last term alone.
g
=(x − 0.5)
hˆ1
2.917(x + 1.941)(x + 0.687)(x − 0.535)(x − 1.767)
−
(x + 2.261)(x + 1.144)(x + 0.046)(x − 1.046)(x − 2.144)
0.458
− .
x − 3.261

Let w1 and w2 denote the neighbors of v contained in S1 and S2 respectively.

Let 2.917
P2
ĥ2
denote the first rational expression above. Again, since P2 and ĥ2 have
strictly interlacing roots, we can construct a matrix B in R(S1 ) centered at w1 such
that B has characteristic polynomial P2 and B(w1 ) has characteristic polynomial
0.458
ĥ2 . As for the last rational expression, x−3.261 , the matrix [3.621] ∈ R(S2 ) has
characteristic polynomial x − 3.261, so we have reached the base case of our
inductive algorithm.
Thus, our matrix A1 will be of the form:
⎡ ⎤
0.5 (2.917e1)T 0.458
A 1 = ⎣ e1 B 0 ⎦.
1 0 3.261
342 C. R. Johnson and E. Gruner

By continuing this process, we can calculate

⎡ ⎤
0.066 1.027 0 1.204 0
⎢ 1 −0.082 3.438 0 ⎥
⎢ 0 ⎥
⎢ ⎥
B=⎢ 0 1 −0.093 0 0 ⎥,
⎢ ⎥
⎣ 1 0 0 −0.076 0.373 ⎦
0 0 0 1 −0.077

which, by performing the appropriate substitutions, yields

⎡ ⎤
0.5 7 0 0 0 0 0 0
⎢ 1 0.5 2.917 0 0 0 0 0.458⎥
⎢ ⎥
⎢ 0 1 0.066 1.027 0 ⎥
⎢ 0 1.204 0 ⎥
⎢ ⎥
⎢0 0 1 −0.082 3.438 0 0 0 ⎥
A=⎢ ⎥.
⎢0 0 0 1 −0.093 0 0 0 ⎥
⎢ ⎥
⎢0 0 1 0 0 −0.076 0.373 0 ⎥
⎢ ⎥
⎣0 0 0 0 0 1 −0.077 0 ⎦
0 1 0 0 0 0 0 3.261

While we will not perform the calculation here, one can use a similar inductive
process to confirm that A is diagonally similar to a symmetric matrix, since all off-
diagonal entries are positive.
Example 4.2 Let T be the same 8-vertex tree used in the previous example, and let
p2 be the following polynomial:

p2 (x) = x 2 (x − 5)3 (x + 1)(x 2 + 1).

We wish to find a matrix A ∈ R(T ) with characteristic polynomial p2 .

(In the following calculations, real numbers greater than 10 in absolute value
are rounded to the nearest tenth. Numbers with magnitude between 1 and 10 are
rounded to the nearest 100th, and numbers with magnitude less than 1 are rounded
to the nearest 1000th.)
The polynomial p2 has root deficiency 6, so let  = (6 − 2)/2 = 2. We wish to
find a vertex v of T such that some partial sum of the branch weights of T at v lies
in the range [2, 5]. Fortunately, the highlighted vertex v indicated in Fig. 2 suffices,
since it has a branch of weight 2.
We will center our matrix A at v. Now, we need to construct an appropriate
degree 7 polynomial g to be the characteristic polynomial of A(v).
If γ1 > γ2 > · · · > γ7 are the roots of g, and 5 and −1 the signed roots of p2 ,
we would like to choose γ1 , γ2 , . . . , γ5 to be greater than 5, γ6 to be between 5 and
−1, and γ7 to be less than −1. This will ensure that the PFD of pg2 will produce five
The Inverse Characteristic Polynomial Problem for Trees 343

Fig. 2 Our tree T with

highlighted vertex v

negative and two positive numerators, as shown in the proof of Proposition 3.3. So
let

g(x) = (x − 8)(x − 7.5)(x − 7)(x − 6.5)(x − 6)(x − 1)(x + 2).

Now we can write

 
p2 −9627.4 1847.0 −11614.8 2638.4
=(x + 20) − + + +
g x−8 x − 7.5 x−7 x − 6.5
 
−155.4 −0.011 −0.039
− + + .
x−6 x−1 x+2

We see that T has three branches at v, which have weights 3, 2, and 2. Call these
branches T1 , T2 , and T3 respectively, and suppose they each contain the respective
neighbors u1 , u2 , and u3 of v.
We would like to consolidate the PFD terms into three rational expressions such
that the denominators have degrees 3, 2, and 2. However, in order to apply our
algorithm, we must only consolidate terms whose numerators have the same sign.
So let’s combine the first, third, and fifth terms into one expression, the second and
fourth terms into another, and the last two terms into our final expression.
 
p2 −21397.6(x − 7.55)(x − 6.01) 21108.4(x − 6.63)
=(x + 20) − +
g (x − 8)(x − 7)(x − 6) (x − 7.5)(x − 6.5)
 
−0.051(x + 1.33)
−
(x − 1)(x + 2)

−21397.6ĥ1 21108.4ĥ2 −0.051ĥ3
=(x + 20) − + + .
P1 P2 P3

Since each ĥi and Pi have strictly interlacing roots, we can construct a matrix
Ai ∈ R(Ti ), centered at ui , such that Ai has characteristic polynomial Pi , Ai (ui )
344 C. R. Johnson and E. Gruner

has characteristic polynomial ĥi , and Ai is diagonally similar to a symmetric matrix.

Our desired matrix A will have the form
⎡ ⎤
−20 (−21397.6e1)T (21108.4e1)T (−0.051e1)T
⎢ e1 A1 0 0 ⎥
A1 = ⎢
⎣ e1
⎥.
⎦
0 A2 0
e1 0 0 A3

One can use the neighbors formula to verify that A does indeed have character-
istic polynomial p2 .
Finally, we can construct A1 , A2 , and A3 using the algorithm in Theorem 3.4 to
yield the matrix
⎡ ⎤
−20 −21397.6 0 0 21108.4 0 −0.051 0
⎢ 1 7.44 0.249 0.012 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 0 ⎥
⎢ 1 7.55 0 0 0 0 ⎥
⎢ ⎥
⎢ 0 1 0 6.01 0 0 0 0 ⎥
A=⎢ ⎥.
⎢ 1 0 0 0 7.38 0.109 0 0 ⎥
⎢ ⎥
⎢ 0 0 0 0 1 6.63 0 0 ⎥
⎢ ⎥
⎣ 1 0 0 0 0 0 0.327 1.57 ⎦
0 0 0 0 0 0 1 −1.33

5 Corollaries, Applications, and Special Cases

To help narrow the scope of Theorem 3.5, and to pave the way for useful corollaries,
we present a few entirely graph-theoretic results. These theorems will aid us in
choosing the most appropriate vertex of the graph on which to center our constructed
matrix.
Unless otherwise specified, let T be an arbitrary tree on n vertices. We let k =
n − 1,  be a nonnegative integer with  < k/2, and m = k − 2 + 1, the number of
integers in the range [, k − ]. Observe that since  < k/2, we must have m ≥ 2.
With respect to , we say that a vertex v of T is ideal if the largest branch of T
at v has weight k −  or smaller.
If T is our tree and p is our desired polynomial with root deficiency d ≥ 2, then
the only vertices upon which we could center our constructed matrix A are the ideal
vertices of T with respect to  = (d − 2)/2. If a vertex v is not ideal - that is, if T
has a branch at v with weight greater than k −  - then any partial sum of the branch
weights of T at v will either be smaller than  or larger than k − .
Proposition 5.1 T has at least one ideal vertex with respect to .
Proof Let v be any vertex of T , and let M be the weight of the largest branch of T
at v. We proceed by induction on M.
The Inverse Characteristic Polynomial Problem for Trees 345

If M ≤ k − , then v is an ideal vertex, and we are done. So suppose M > k − ,

and that the claim holds for all maximum branch weights c with M > c ≥ k −. Let
T1 , T2 , . . . , Ts be the branches of T at v, in which each Ti has weight i and contains
ui , a neighboring vertex of v. Without loss of generality, assume that 1 = M. We
must have 1 + 2 + · · · + s = k, so if 1 > k − , then 2 + · · · + s < .
Now let S1 , S2 , . . . , St be the branches of T at u1 , and assume S1 is the branch that
contains v. By the previous statement, S1 can have weight at most . Furthermore,
the weights of S2 , . . . , St sum to M − 1, so no individual Si can have weight greater
than M − 1. This includes S1 ; since  < k/2, we have  < k −  < M. The choice
of our starting vertex was arbitrary, so by our inductive hypothesis, T has an ideal
vertex.
Note that the proof of the above theorem not only confirms the existence of an
ideal vertex for any tree T , but also provides guidance for how to locate it. We also
have the following useful lemma regarding the positions of the ideal vertices:
Lemma 5.2 The ideal vertices of T with respect to , along with the edges
connecting them, form a subtree of T .
Proof The case when T only has one ideal vertex is trivial, so we assume that v and
w are two distinct ideal vertices of T . Since T is a tree, there exists a unique path on
T connecting v and w; it suffices to prove that every vertex along this path is ideal.
If v and w are the only vertices on this path (i.e. if v and w are neighbors), then
the case is again trivial. So assume that the path contains a vertex u distinct from
v and w. Let Tv and Tw be the (necessarily distinct) branches of T at u containing
the vertices v and w respectively, and let T1 , . . . , Ts be the remaining branches of
T at u. The vertex u is contained in some branch S with respect to v, and since v
is ideal, S has weight k −  or smaller. Therefore, the branches of T at u that do
not contain v can have weights that sum to at most k −  − 1, implying that Tw and
T1 , . . . , Ts all have weight less than k − . However, since w is also ideal, we can
apply a symmetric argument and conclude that Tv has weight less than k −  as well.
Since all branches of T at u have weight less than k − , u must be ideal (Fig. 3).
Now suppose that, for a given , we have an ideal vertex v for a tree T . If any
branch of T at v has its weight in the interval [, k − ], then by Theorem 3.5 we
can construct a pseudosymmetric matrix A for any polynomial of root deficiency
d = 2 + 2. We need only be concerned if all branches of T at v have weight
less than . In that case, we call v deficient with respect to . Not all trees have a
deficient vertex. In fact, if a tree has more than one ideal vertex, then we can say
with certainty that none of them are deficient.
Proposition 5.3 If T has a deficient vertex v with respect to , then v is the only
ideal vertex of T with respect to .
Proof Let v be a deficient vertex of T , and let u1 , u2 , . . . , us be the neighboring
vertices of v, in which ui is contained in the branch Ti with weight i . By
Lemma 5.2, the ideal vertices of T form a subtree, so if T contained an additional
ideal vertex, then at least one ui would have to be ideal. However, by assumption
346 C. R. Johnson and E. Gruner

Fig. 3 Images (a), (b), and (c) denote the ideal vertices of this 13-vertex tree with respect to 5, 4,
and 3 respectively


i <  for each i, which also means that sj =1,j =i j > k −  for each i. Thus for
any ui , the branch containing v would have weight greater than k − , so ui cannot
be ideal.
Therefore, if our desired polynomial p has root deficiency d, and our tree T has
a deficient vertex v with respect to  = (d −2)/2, then any pseudosymmetric matrix
A would have to be centered at v. Fortunately, even in this case, we can still usually
come up with an appropriate partial sum. We will present some sufficient conditions
for that later.
We will conclude this graph theory discussion with a theorem that places an
upper bound on the number of ideal vertices a tree can have with respect to a given
.
This may not be the most useful result from a theoretical perspective—we only
need to find one appropriate ideal vertex to know that a matrix can be constructed—
but from a computational standpoint certain vertices may be better for centering
than others. An upper bound—along with the fact that all ideal vertices must be
connected—allows the algorithm user to know when he has found all possible ideal
vertices, allowing him to make the best judgment on which one to use.
First, we introduce one more term and lemma. We call an ideal vertex v of T
ideal pendent (with respect to ) if it is a pendent vertex on the subtree of ideal
vertices with respect to . It is well known that a tree on two or more vertices must
have two or more of those vertices pendent. So as long as T has more than one ideal
vertex, we are guaranteed at least two pendent ideal vertices.
The Inverse Characteristic Polynomial Problem for Trees 347

Lemma 5.4 Suppose v is an ideal pendent vertex of T , and let T1 , . . . , Ts be the

branches of T at v with respective lengths 1 , . . . , s . Without loss of generality,
assume that T1 is the only branch of T at v that contains an ideal vertex. Then
2 + 3 + · · · + s ≥ .
Proof If 2 + 3 + · · · + s < , then 1 > k −  and v would not be an ideal
vertex.
Now, we present and prove our upper-bound result:
Proposition 5.5 T has at most m = k − 2 + 1 ideal vertices with respect to .
Proof We proceed by induction on . If  = 0, then m = n, and the claim holds.
Now assume  > 0, and that the claim holds for s with  > s ≥ 0. Suppose
T has α ideal vertices with respect to  − 1 and β ideal vertices with respect to .
Note that β ≤ α; any vertex that is ideal with respect to  will also be ideal with
respect to  − 1. By our inductive hypothesis, α ≤ k − 2( − 1) + 1 = m + 2. If
α ≤ m, then the claim holds automatically. So we only need to consider the cases
when α = m + 1 or α = m + 2. Note that in either case, α > 2 since m ≥ 2.
In the case where α = m + 1, we would have n − (m + 1) = 2 − 1 non-ideal
vertices with respect to −1. However, recall that at least two of our α ideal vertices
are ideal pendent, and by Lemma 5.4 they each have branches consisting entirely
of non-ideal vertices whose weights sum to  − 1 or greater. (These branches are
distinct since our graph is minimally connected.) If both of these sums were greater
than  − 1, then the total number of non-ideal vertices would be 2 or greater, a
contradiction. Therefore, at least one of our ideal pendent vertices has its “non-
ideal” branch weights summing to exactly  − 1, implying that the branch that does
contain ideal vertices has weight k−+1. Therefore, when we change our parameter
from  − 1 to , this particular vertex will no longer be ideal. So β ≤ α − 1 = m, as
desired.
The case where α = m + 2 is similar. In that case, both of our pendent ideal
vertices with respect to  − 1 lose their ideal status upon changing the parameter to
. We still find that β ≤ m, as desired (Fig. 4).
We can now combine our main theorem with the previous results on ideal vertices
to draw some useful corollaries, targeting specific types of trees or polynomials that
can be realized with our algorithm.
For the remainder of this section, we will let T be an arbitrary tree on n vertices,
and p be an arbitrary polynomial of degree n and root deficiency d ≥ 2. (The case
when d = 0 merits no further study due to Theorem 3.4). We let  = (d − 2)/2,
k = n − 1, and m = k − 2 + 1, the number of integers in the range [, k − ].

Fig. 4 With respect to  = 4,

this 10-vertex tree has exactly
m = 2 ideal vertices
348 C. R. Johnson and E. Gruner

Furthermore, observe that since d can be at most n, we always have (d − 2)/2 =

 < (n − 1)/2 = k/2.
The first result we already mentioned, but we state it again formally.
Theorem 5.6 If T has an ideal vertex v with respect to  that is not deficient, then
p has a pseudosymmetric realization A ∈ R(T ).
Proof By the definition of ideal vertex, T can have no branches at v with weight
larger than k − . Since v is not deficient, T must have at least one branch at v with
weight  or greater. The result follows from Theorem 3.5.
As a natural follow-up to this corollary, we have:
Corollary 5.7 If T has more than one ideal vertex with respect to , then p has a
pseudosymmetric realization A ∈ R(T ).
Proof By Proposition 5.3, if T has more than one ideal vertex, then none of them
can be deficient. See Theorem 5.6.
Now the only case of interest is when T has a single, deficient ideal vertex v with
respect to the appropriate . The next result covers a large number of cases:
Proposition 5.8 Suppose T has a deficient ideal vertex v with respect to . The
polynomial p has a pseudosymmetric realization A ∈ R(T ) if at most two branches
of T at v have weight greater than m.
Proof Let T1 , . . . , Ts be the branches of T at v with respective weights 1 , . . . , s .
By definition of deficient vertex, each i < . Without loss of generality, assume
that 2 , 3 , . . . , s−1 are all less than or equal to m. Let ki denote the partial sum
1 + 2 + · · · + i ; clearly k1 < k2 < · · · < ks . Note that k1 = 1 < , but
ks−1 = k − s > k − .
If there existed no ki ∈ [, k − ], then for some i ∈ [1, s − 2] we would have
ki+1 − ki > (k − ) −  + 1 = m. But ki+1 − ki = i+1 , and by assumption
2 , . . . , s−1 are all less than or equal to m; a contradiction. Therefore, there must
exist some partial sum ki ∈ [, k − ], which by Theorem 3.5 guarantees that p has
a pseudosymmetric realization in R(T ).
This last result paves the way for a series of increasingly specific corollaries.
The first one concerns linear trees. A linear tree T is a tree whose high-degree
vertices all occur along a single induced path. (We define a high-degree vertex to be
a vertex with degree at least three.) Furthermore, if P is the longest induced path
that includes all of these vertices, then the depth of T is the maximum distance (in
number of vertices) a vertex can be from this path P .
Proposition 5.9 Suppose T is a linear tree with depth at most m. Then the
polynomial p has a pseudosymmetric realization A ∈ R(T ).
Proof Let v be an ideal vertex of T . If v is not deficient, then the result follows
from Theorem 5.6. If v is deficient, then it must be high-degree. Since  < k/2, if T
The Inverse Characteristic Polynomial Problem for Trees 349

Fig. 5 A linear tree on 12 vertices with depth 2, with one deficient ideal vertex v with respect to
 = 5. Note that the branch weights 4 and 1 sum to 5 ∈ [5, 6] = [, k − ]

had two or fewer branches at v, each with weight less than , then the sum of their
weights would be less than k, a contradiction.
Since T is a linear tree, at most two branches of T at v can contain additional
high-degree vertices, which means the remaining branches are all paths of length m
or smaller. The result follows from Proposition 5.8.
The previous result yields the following specific case:
Corollary 5.10 Suppose T is a linear tree with depth at most 2. Then p has a
pseudosymmetric realization A ∈ R(T ).
Proof Since  < k/2, we know that  is strictly less than k−, so the range [, k−]
must contain at least two integers. Therefore, m = k − 2 + 1 ≥ 2. See the previous
result (Fig. 5).
This corollary includes the case of paths, which can be considered linear trees
with depth 0. While the existence of arbitrary polynomial realizations for paths had
already been proven in the real number case, this is the first proof that uses an
explicit, constructive argument.
The next few results also follow from Proposition 5.8.
Proposition 5.11 If m ≥  − 1, or, equivalently, if n ≥ (3/2)(d − 2) − 1, then p
has a pseudosymmetric realization A ∈ R(T ).
Proof Let v be an ideal vertex of T . If v is not deficient, then the result follows from
Theorem 5.6. If v is deficient, then by definition all branches of T at v have weight
less than , which in this case also means they have weight less than or equal to m.
The result now follows from Proposition 5.8 (Fig. 6).

Fig. 6 A tree on 15 vertices,

with one deficient ideal vertex
v with respect to  = 5. Note
that m = 14 − 10 + 1 = 5 ≥
 − 1 = 4, and we have
branch weights 4 and 2 that
sum to 6 ∈ [5, 9] = [, k − ]
350 C. R. Johnson and E. Gruner

As direct consequence of Proposition 5.11, we have:

Corollary 5.12 Suppose n < 10. Then p has a pseudosymmetric realization A ∈
R(T ).
Proof By the previous result, p is realizable for any tree if its degree n and root
deficiency d satisfy the inequality n ≥ (3/2)(d − 2) − 1. If n < 10, then d can be
either 0, 2, 4, 6, or 8. In each of these cases, substituting d into the right-hand side
of the inequality yields a value less than or equal to d. Since n must be at least d,
the inequality is satisfied.
Note that the inequality n ≥ (3/2)(d − 2) − 1 is equivalent to d ≤ (2/3)(n +
1) + 2. This form is perhaps more useful, as it allows one to see the minimum
number of signed roots needed for a degree n polynomial in order to guarantee a
pseudosymmetric realization for every tree.
The last corollary we will state does not directly follow from any of the previous
ones, but is an interesting observation nonetheless.
Corollary 5.13 Suppose n is even, and T is composed of two distinct subtrees
T1 and T2 , each with weight n/2, that are joined by a single edge. Then p has a
pseudosymmetric realization A ∈ R(T ).
Proof Let v be one of the two vertices on the edge that connects T1 and T2 . Then T
will have one branch at v with weight n/2 (equal to either T1 or T2 ), and additional
branches whose weights sum to n/2 − 1. The root deficiency d of p can be at most
n; therefore,  = (d − 2)/2 will be at most (n − 2)/2 = n/2 − 1, and k −  will be
at least n/2. Therefore, the branch partial sum n/2 − 1 is guaranteed to fall in the
range [, k − ], and the result follows from Theorem 3.5 (Fig. 7).
We emphasize that none of these corollaries are “if and only if” statements; a
polynomial and tree pair may fall under none of the categories described by the
corollaries yet still have a pseudosymmetric realization. Consider the polynomial
p(x) = x 14 and the generalized star S shown in Fig. 8.
Since p has root deficiency 14, we have  = 6, [, k − ] = [6, 7], and m = 2.
Here, S has a single deficient ideal vertex v, and T has four branches at v with
weights 5, 4, 3, and 1. We can realize p with a pseudosymmetric matrix A ∈ R(S)
because we have 5 + 1 = 6 ∈ [, k − ]. However, we have m <  − 1 (violating the

Fig. 7 A 20-vertex tree

composed of two 10-vertex
subtrees joined by a single
edge. With respect to  = 9,
we have an ideal vertex with a
branch of length 10, and
10 ∈ [9, 10] = [, k − ]
The Inverse Characteristic Polynomial Problem for Trees 351

Fig. 8 A generalized star

with arm lengths 5, 4, 3,
and 1

conditions of Proposition 5.11), S has depth greater than m when viewed as a linear
tree (violating the conditions of Proposition 5.9), and we have three branches with
weight greater than m = 2 (violating the conditions of Proposition 5.8).

Acknowledgement This work was supported by the 2019 National Science Foundation grant
DMS #1757603.

References

1. W. Adkins, M. Davidson, Synthetic partial fraction decompositions. Math. Mag. 81(1), 16–26
(2008)
2. R.S. Cuestas-Santos, C.R. Johnson, Spectra of tridiagonal matrices over a field (2018). Preprint.
arXiv:1807.08877
3. S. Friedland, Inverse eigenvalue problems. Linear Algebra Appl. 17, 15–51 (1977)
4. R.A. Horn, C.R. Johnson, Matrix Analysis, 2nd edn. (Cambridge University Press, New York,
2013)
5. C.R. Johnson, A. Leal-Duarte, Complete spectral theory for matrices over a field whose graph
is a star. Manuscript
6. C.R. Johnson, C.M. Saiago, Eigenvalues, Multiplicities, and Graphs (Cambridge University
Press, New York, 2018)
7. A. Leal-Duarte, Construction of acyclic matrices from spectral data. Linear Algebra Appl. 113,
173–182 (1989)
A Note on the Fredholm Theory
of Singular Integral Operators
with Cauchy and Mellin Kernels, II

Peter Junghanns and Robert Kaiser

Abstract Necessary and sufficient conditions for the Fredholmness of a class of

singular integral operators in weighted Lp -spaces on the interval (0, 1) of the
real line are formulated under weaker conditions than in Junghanns and Kaiser
(Oper Theory Adv Appl 271:291–325, 2018). Moreover, results on the one-sided
invertibility of the operators under consideration are proved.

Keywords Singular integral operators · Cauchy kernel · Mellin kernel ·

Fredholm theory · One-sided invertibility

Mathematics Subject Classification (2010) Primary 45E05; Secondary 45E10

1 Introduction

In this paper we consider linear integral operators, which are made up by multipli-
cation operators, the Cauchy singular integral operator and Mellin type operators.
More precisely, the operators under consideration are given by
 1  1  
b(x) u(y) dy x u(y) dy
(Au)(x) :=a(x)u(x) + + c+ (x) k+
πi 0 y−x 0 y y
(1.1)
  
1 1−x u(y) dy
+ c− (x) k− , 0 < x < 1,
0 1−y 1−y

P. Junghanns ()
Chemnitz University of Technology, Faculty of Mathematics, Chemnitz, Germany
e-mail: [email protected]
R. Kaiser
TU Bergakademie Freiberg, Faculty of Mathematics and Computer Science, Freiberg, Germany
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 353

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_18
354 P. Junghanns and R. Kaiser

p p
and are considered in weighted Lp -spaces Lρ,σ := Lv ρ,σ (0, 1), where a, b, c± :
[0, 1] → C are piecewise continuous functions,

v ρ,σ (x) = x ρ (1 + x)σ , x ∈ (0, 1) , ρ, σ > −1 ,

is a classical Jacobi weight, and the Mellin transforms of the functions k± :

(0, ∞) → C are supposed to be continuous.
The paper continues the investigations in [5] and extends the results given there
by using techniques from [2] and properties of Fourier convolution operators in
weighted Lp -spaces proved in [1] and [8]. To be more precise, in [5] the main
theorem formulates necessary and sufficient conditions for the Fredholmness of the
operator A as well as an index formula based on the winding number of a certain
closed curve in C associated with A. In the present paper we will state and prove
this theorem again but under weaker conditions regarding the functions k± . In order
to do this, we basically use [1, Theorem 5.7] (cf. also [8, Theorem 1.3 and Theorem
3.2]).
Additionally to the results proved in [5], we investigate the regularity of solutions
of equation Au = f and prove results concerning the one-sided invertibility of the
p
operator A in Lρ,σ .
The paper is structured as follows. In Sect. 2, we define weighted Lp -spaces
on the real line and on the half line, collect useful properties of the Fourier
transformation, and introduce specific classes of convolution operators. In the
following Sect. 3 we are going to state results regarding the boundedness of
p
operator A in the weighted Lp -spaces Lρ,σ . To achieve this, we introduce the
Mellin transformation at first and recall useful results from [5]. Section 4 starts
with a result, which connects operators of form (1.1) and convolution operators
introduced in Sect. 2, followed by the above mentioned Fredholm theorem, first in
case a, b, c± are constants and later for piecewise continuous functions a, b, c± .
In Sect. 5, we investigate the smoothness of solutions of equation Au = f as
well as their asymptotic behaviour near the boundary point 1. Here beside the case
c− (x) ≡ 0, we distinguish between the case of b(x) being the zero function, that
means, that the Cauchy singular integral operator does not appear, and the case of
both functions c± (x) being identically zero, i.e., the Mellin-type operators Bk± do
not occur. Results and techniques from [2, Section 6] will be used and extended.
p
The final Sect. 6 deals with the one-sided invertibility of A in Lρ,σ in case c+ or c−
vanish.

2 Preliminaries

In this section we first collect some properties of Fourier convolution operators.

With that, we are going to introduce a specific class of convolution operators.
Fredholm Theory of Singular Integral Operators 355

Subsequently, results concerning the one-sided invertibility of such operators are

formulated.
For 1 < p < ∞ and −1 < σ < p − 1, let us introduce the weighted Lp -space
p
Lσ (R) defined by the norm
 ∞ 1
|t|σ p
uσ,p,R = |u(t)|p dt . (2.1)
−∞ (1 + |t|)σ
p
We are going to use the abbreviations Lp (R) := L0 (R) and up,R := u0,p,R .
For u ∈ L1 (R), the well-known Fourier transform is defined by
 ∞
(F u) (η) := e−iηt u(t) dt, η ∈ R.
−∞

For 0 < R < ∞ and u ∈ L1 (−R, R), set

 R
(FR u) (η) = e−iηt u(t) dt, η ∈ R, (2.2)
−R

and

 1 R
FR− u (t) = eiηt u(η) dη, t ∈ R.
2π −R

Let 1 < p ≤ 2, p1 + q1 = 1 and u ∈ Lp (R). Then (cf. [9, p. 96, Theor. 74]) FR u
u and FR−
converges for R → ∞ in Lq (R) to a function  u converges in Lp (R) to u.
Moreover, we have
1

uq,R ≤ (2π) q up,R .

Obviously, if u ∈ L1 (R) ∩ Lp (R), then, due to (2.2), (F u)(η) = 

u(η) for almost
every η ∈ R. That is why, we will call the operator

F : Lp (R) → Lq (R), u → 
u

the Fourier transformation and the function F u the Fourier transform of u ∈ Lp (R).
Of course, F − : Lp (R) → Lq (R) is also well defined, where for u ∈ Lp (R),
we set F − u = limR→∞ FR− u and take the limit in the Lq (R)-sense. Note that,
F v and F − v are also defined for v = F − u and v = F u, respectively, where
F v := limR→∞ FR v as well as F − v := limR→∞ FR− v and the limits are taken in
the Lp (R)-sense. Moreover, with these definitions we have

F −F u = F F −u = u ∀ u ∈ Lp (R), 1 < p ≤ 2.
356 P. Junghanns and R. Kaiser

In case p = 2, it’s well-known that F : L2 (R) → L2 (R) is an isomorphism, where

1 ∗
F = F −1 = F − . (2.3)
2π
Let f, g : R → C be measurable functions. We assume that there is a set N ⊂ R of
measure zero such that f (t − ·)g(·) is integrable for all t ∈ R\N. In this case, we
define the convolution f ∗ g : R → C of these functions by
⎧ ∞
⎨ f (t − s)g(s) ds : t ∈ R\N
(f ∗ g)(t) := −∞
⎩
0 : t ∈ N.

If p ∈ (1, ∞), 1
p + 1
q = 1, f ∈ Lp (R), and g ∈ Lq (R), then

(f ∗ g)(t) ≤ f p,R gq,R , t ∈ R.

This relation is generalized by the well known Young’s inequality for convolutions

f ∗ gr,R ≤ f p,R gq,R , (2.4)

which is true for f ∈ Lp (R), g ∈ Lq (R) with 1 ≤ p, q ≤ ∞ and p1 + q1 ≥ 1 as well

 
as p1 + q1 = 1 + 1r ∞1
:= 0 . In particular, under these assumptions, f ∗ g ∈ Lr (R).
If
1 1
1 < p, q ≤ 2 and + >1
p q

hold true, then we have the convolution theorem

F (f ∗ g) = F f · F g and F − (f ∗ g) = 2π F − f · F − g

as well as

f ∗ g = F − (F f · F g) and f ∗ g = 2π F (F − f · F − g) (2.5)

(cf. [9, p. 106, Theor. 78]).

By SR we denote the Cauchy singular integral operator
 ∞
1 f (t)
(SR f )(s) = dt, s ∈ R,
πi −∞ t −s
Fredholm Theory of Singular Integral Operators 357

where the integral is considered as a Cauchy principal value integral. It is well

p p
known (see, for example, [3, Sections 1.2, 1.5]) that SR : Lσ (R) → Lσ (R) is a
linear and bounded operator, since 1 < p < ∞ and −1 < σ < p − 1.
Let 1 < p ≤ 2, γ ∈ R, and bγ (t) = eiγ t . Then (cf. [2, Lemma 1.35])

bγ−1 SR bγ u = −F sgn(· − γ )F − u ∀ u ∈ Lp (R), (2.6)

where sgn : R → {−1, 0, 1} denotes the sign function.

For p ∈ (1, ∞) and ρ ∈ (−1, p − 1), let us introduce the weighted Lp -space

Lρ (R+ ) defined by the norm
p

 ∞ 1
p
uρ,p,R+ ,∼ = |u(t)| t dt
p ρ
.
0

By SR+ we denote the Cauchy singular integral operator

 ∞
1 f (t)
(SR+ f )(s) = dt, s ∈ R+ ,
πi 0 t −s

where again the integral is considered as a Cauchy principal value integral. The
operator SR+ is a linear and bounded operator in  Lρ (R+ ) (see, for example, [3,
p

Sections 1.2, 1.5]). For ξ = 1+ρ

p , we introduce the mapping

ξ : 
Z Lpρ (R+ ) → Lp (R), f → e−ξ · f (e−· ),
 
which is an isometric isomorphism with Z −1 f (y) = y −ξ f (− ln y). Now, a
ξ
consequence of relation (2.6) is the formula

ξ SR+ Z
Z −1 u = F aF − u ∀u ∈ Lp (R), (2.7)
ξ

where a(t) = −i cot(πξ − iπt), t ∈ R, and 1 < p ≤ 2.

For every a ∈ L∞ (R) the operator

Wa0 : L2 (R) → L2 (R), u → F aF −1 u

is well defined, linear, and bounded (cf. (2.3)). In particular, for bγ (t) = eiγ t , γ ∈ R,
and aγ (t) = −sgn(t − γ ), due to (2.6), we have

Wa0γ = bγ−1 SR bγ I. (2.8)

358 P. Junghanns and R. Kaiser

Let p ∈ (1, ∞), −1 < σ < p − 1, and a ∈ L∞ (R). If there is a finite constant
M = M(a, p, σ ) such that
; ;
; 0 ;
;Wa u; ≤ M uσ,p,R ∀ u ∈ L2 (R) ∩ Lpσ (R),
σ,p,R

p
then the uniquely defined linear and continuous extension of Wa0 onto Lσ (R) is
again denoted by Wa0 : Lσ (R) → Lσ (R) and a ∈ L∞ (R) is called a (p, σ )-
p p

multiplier. Let us denote the set of all (p, σ )-multipliers by Mp,σ . In view of
p
the above example given by (2.8) and the fact that SR ∈ L(Lσ (R)), we have
p
Waγ ∈ L(Lσ (R)) for every γ ∈ R. Since, for the characteristic function χ(a,b]
0

of the interval (a, b] ⊂ R, we have the representation χ(a,b](t) = sgn(t − a) −

sgn(t − b), the set of all finite linear combinations of characteristic functions of
(also unbounded) intervals belongs to Mp,σ .
As usual, by V1 (R) we denote the set of all functions a : R → C with bounded
total variation and by var(a) the total variation of a ∈ V1 (R). It turns out (see [8,
Theorem 1.3]) V1 (R) ⊂ Mp,σ and that, for all a ∈ V1 (R), the inequality

aMp,σ ≤ Cp,σ SR L  4a + var(a)

5
(2.9)
p
Lσ (R) ∞

is true, where the constant Cp,σ does only depend on p and σ and not on a ∈ V1 .
In what follows, R = R ∪ {±∞} and Ṙ denote the two-point compactification
and the one-point compactification of R, respectively. By PC(R) we refer to the
algebra of all piecewise continuous functions a : R → C (i.e., f ∈ PC(R) if and
only if all limits a(t ±0), t ∈ Ṙ, where a(∞±0) := a(∓∞) = limt →∓∞ a(t), exist
and are finite). Moreover, by PC  p,σ we denote the completion of the; set ;PC(R) ∩
V1 (R) in the Banach algebra Mp,σ , ·Mp,σ , where aMp,σ = ;Wa0 ;L(Lp (R))
; ; σ
(cf. [1, p. 254]). Since a∞ = ;Wa0 ;2,R (cf.[2, Proposition 2.3]), we have PC2,0 =
(PC(R), ·∞ ). Note that PCp,σ ⊂ PC(R) (cf. [1]).
We introduce the space Lσ (R+ ) as the compression of Lσ (R) to the positive
p p
+ p +
half line R . Thus, Lσ (R ) may be identified with the image of the projection
p p
P : Lσ (R) → Lσ (R), which is defined by

u(t) : t ∈ R+ ,
(Pu)(t) =
0 : t ∈ R \ R+ .

Due to (2.1), the norm in Lσ (R+ ) is given by

p

 ∞ 1
tσ p
uσ,p,R+ = |u(t)| p
dt ,
0 (1 + t)σ
Fredholm Theory of Singular Integral Operators 359

where we will use the abbreviation up,R+ := u0,p,R+ . It is easily seen that

 ∞ 1
p
−t σ
up,σ,∼ := |u(t)| (1 − e ) dt
p
0

defines an equivalent norm on Lσ (R+ ), i.e., there are positive constants C1 and C2
p

such that

C1 up,σ,∼ ≤ up,σ ≤ C2 up,σ,∼ ∀ u ∈ Lpσ (R+ ) .

For a ∈ Mp,σ , the so-called Wiener-Hopf integral operator Wa ∈ L(Lσ (R+ )) is

p

defined as the restriction of PWa0 P to the space Lσ (R+ ). For 1 < p < ∞, note
p

that, w.r.t. the dual product


u, vR := u(t)v(t) dt
R

p q p
the dual space of Lσ (R) is equal to Lμ (R) with μ = (1 − q)σ and q = p−1 .
∗
Analogously, Lσ (R+ ) = Lμ (R+ ).
p q

Lemma 2.1 Let 1 < p < ∞ and p1 + q1 = 1. Then Mp,σ = Mq,μ . Moreover, if
∗
a ∈ Mp,σ , then Wa0 = Wa0 as well as (Wa )∗ = Wa .
p p
Proof Define R : Lσ (R) → Lσ (R) by (Ru)(t) = u(−t). Then, RFu = F Ru =
2πF − u for u ∈ Lp (R) and 1 < p ≤ 2. Let a ∈ Mp,σ and p ∈ (1, ∞). Taking into
p q
account (2.3), for u ∈ L2 (R) ∩ Lσ (R) and v ∈ L2 (R) ∩ Lμ (R), we have
     
Wa0 u, v = u, F aF −1 v = u, Wa0 v .
R R R

p 
Since Wa0 is continuous on L2 (R) ∩ Lσ (R), .p,σ,R , the operator Wa0 is contin-
q 
uous on L2 (R) ∩ Lμ (R), .q,μ,R , which means a ∈ Mq,μ . Using

RF v = F v and F −1 v = (2π)−1 F v = F −1 Rv on L2 (R),

we get

RF aF −1 v = F aF −1 v = F aF −1 Rv ∀ v ∈ L2 (R) ∩ Lqμ (R),

which implies a ∈ Mq,μ . Hence, due to symmetry reasons, Mp,σ = Mq,μ . By the
∗ ∗
way Wa0 = Wa0 and, consequently, (Wa )∗ = P Wa0 P = Wa , since P ∗ = P.
360 P. Junghanns and R. Kaiser

Let us finish this section with two lemmata on the triviality of the nullspace of a
Wiener-Hopf integral operator or of its adjoint.
Lemma 2.2 ([2], Prop. 2.8) Let us assume that 1 < p < ∞, a ∈ Mp,0 , and
a = 0 almost everywhere on R. Then the nullspace of Wa : Lp (R+ ) → Lp (R+ ) or
the nullspace of the adjoint operator Wa : Lq (R+ ) → Lq (R+ ) are trivial, where
p + q = 1.
1 1

The following lemma generalizes the previous one to the case of weighted spaces
using a stronger assumption on a(t). Here, by Cp,σ we refer to the closure of the
set of piecewise
  functions on Ṙ (with finitely many jumps) in the Banach
constant
algebra Mp,σ , .Mp,σ .

Lemma 2.3 ([8], Prop. 1.8) If a ∈ Cp,σ and inft ∈R |a(t)| > 0, then the
homogeneous equation Wa u =  in the space Lσ (R+ ) or the adjoint equation
p

Wa v =  in the space L(1−q)σ (R+ ), p1 + q1 = 1, have only the trivial solution.

q

3 Boundedness

In this section we introduce the Mellin transformation and the Mellin type operators
as well as the Cauchy singular integral operator on the finite interval (0, 1).
Useful properties of the Mellin transformation as well as a relation between the
convolution operators from the previous section and the Mellin type operators are
given. Furthermore, we formulate conditions for the Mellin operators to be bounded
in weighted Lp -spaces.
For z ∈ C and a measurable function f : (0, ∞) → C, for which t z−1 f (t)
is integrable on each compact subinterval of (0, ∞), the Mellin transform f(z) is
defined as
 R
f(z) = lim t z−1 f (t) dt,
R→∞ R −1

if this limit exists. For f ∈ 

L1ξ −1 (R+ ) and η ∈ R, we have
 ∞  ∞
f(ξ + iη) = t ξ +iη−1 f (t) dt = e−(ξ +iη)s f (e−s ) ds
0 −∞ (3.1)
 
ξ f (η) = 2π F − Z
= FZ ξ f (−η).

Let ξ := {z ∈ C : #z = ξ } and

C0 (ξ ) = u ∈ C(ξ ) : lim u(ξ + iη) = 0 .
|η|→∞
Fredholm Theory of Singular Integral Operators 361

Since, for u ∈ L1 (R), F u ∈ C(R) and lim|η|→∞ (F u)(η) = 0 (due to the Riemann-
Lebesgue theorem), by (3.1) we see that, for ξ ∈ R, the Mellin transformation

Mξ : 
L1ξ −1 (R+ ) → C0 (ξ ), f → f

is well defined. Moreover, if f(ξ + i·) ∈ L1 (R) and f ∈ C(0, ∞), then (cf. [4,
Lemma 2.8])
 ∞
1
f (t) = t −ξ −iη f(ξ + iη) dη , 0 < x < ∞. (3.2)
2π −∞

By Lp (ξ ) we denote the Lp -space given by the norm

 ∞ 1
p
gξ ,p = |g(ξ + iη)| dη
p
.
−∞

Taking into account relation (3.1), we are able to define the Mellin transform for
f ∈Lpξ −1 (R+ ) and 1 < p ≤ 2 by
p

 eR
f(ξ + i·) := F Z ξ f = lim
ξ f = lim FR Z t ξ +i ·−1 f (t) dt, (3.3)
R→∞ R→∞ e−R

where the limit is taken in the Lq (ξ )-sense and where 1

p + 1
q = 1.

Corollary 3.1 Let 1 < p ≤ 2 and 1

p + q1 . For ξ ∈ R, the Mellin transformation

Mξ : 
Lpξ −1 (R+ ) → Lq (ξ ), f → f,
p

defined by (3.3), is a linear and bounded operator.

For −∞ < α < β < ∞, by α,β we refer to the strip

α,β = {z ∈ C : α < #z < β}

of the complex plane. The following lemma modifies [4, Cor. 2.9].
Lemma 3.2 Let 1 < p ≤ 2, p1 + q1 = 1, and α, β ∈ R with α < β. Moreover, let
k ∈ Lpξ −1 (R+ )∩C(R+ ) for every ξ ∈ (α, β). By Corollary 3.1 we have 
p
k ∈ Lq (ξ )
for all ξ ∈ (α, β). If, additionally, 
k is holomorphic in the strip α,β satisfying
 
Mα0 ,β0 := sup (1 + |z|)1+δ 
k (z) : z ∈ α0 ,β0 < ∞ (3.4)
362 P. Junghanns and R. Kaiser

for all intervals [α0 , β0 ] ⊂ (α, β) and some δ = δ(α0 , β0 ) > 0, then there hold
 ∞
tμ 4 5
k(t) = t −ξ −iη k(ξ − μ + iη) − 
k(ξ + μ + iη) dη (3.5)
2π(1 − t 2μ ) −∞

for all ξ ∈ (α, β), x ∈ R+ , and 0 < μ < min {ξ − α, β − ξ }. Moreover, 

k(ξ − μ +
i·) − 
k(ξ + μ + i·) ∈ L1 (R) and
 ∞ 4 5

k(ξ − μ + iη) − 
k(ξ + μ + iη) dη = 0 . (3.6)
−∞

Proof For α < ξ ± μ < β, η ∈ R, and 0 < x < ∞, we have

 μ

k(ξ + μ + iη) − 
k(ξ − μ + iη) = 
k (ξ + s + iη) ds ,
−μ

such that, due to (3.4),

 ∞

k(ξ + μ + iη) − 
k(ξ − μ + iη) dη
−∞
 ∞  μ
≤ 
k (ξ + s + iη) ds dη
−∞ −μ
 ∞  μ ds dη
≤ Mξ −μ,ξ +μ <∞
−∞ −μ (1 + |ξ + s + iη|)1+δ

Hence, the continuous function g : (0, ∞), x → k(x)x −μ − k(x)x μ , belongs to

p
Lpξ −1 (by assumption) and its Mellin transform

g (ξ + i·) = 
 k(ξ − μ + i·) − 
k(ξ + μ + i·)

to L1 (R). Relation (3.2) yields (3.5), and relation (3.6) is a consequence of g(1) = 0.

Corollary 3.3 Under the assumptions of Lemma 3.2, we have

lim t α+ε k(t) = 0 and lim t β−ε k(t) = 0

t →+0 t →∞

for every ε > 0.

Proof For given ε > 0, choose ξ ∈ (α, β) and μ > 0 such that 0 < μ < ξ − α <
β − ξ and ξ − α − μ < ε. Then, for 0 < t < 12 , by taking into account (3.5)
and (3.6),
; ;
t α+ε |k(t)| ≤ const t α+ε+μ−ξ ; k(ξ + μ + i·);L1 (R) → 0
k(ξ − μ + i·) − 
Fredholm Theory of Singular Integral Operators 363

if t → +0. Analogously, we can choose ξ ∈ (α, β) and μ > 0 such that 0 < μ <
β − ξ < ξ − α and β − ξ − μ < ε. Again by (3.5) and (3.6) we get, for t > 2,
; ;
t β−ε |k(t)| ≤ const t β−ε−μ−ξ ; k(ξ + μ + i·);L1 (R) → 0
k(ξ − μ + i·) − 

if t → ∞.
The following Lemma modifies Lemma 3.10 in [5].
Lemma 3.4 Let p ∈ [1, ∞), −∞ < α < β < ∞, and

Lpα (R+ ) ∩ 
k ∈ Lβ (R+ ),
p

Then we have k ∈ 
L11+ρ (R+ ) for all ρ ∈ (α, β).
p −1

Proof Let α < ρ < β, 1 < p < ∞, and 1

p + 1
q = 1. Then, due to Hölder’s
inequality,
 ∞ 1+ρ
 1 1+ρ−α
 ∞ 1+ρ−β β
p −1 −1 α
t |f (t)| dt = t p |f (t)| t dt +p t p |f (t)| t p dt
0 0 1

 1

1+ρ−α
  q1
−1 q
≤ t p
dt f α,p,R+ ,∼
0

 ∞  1+ρ−β   q1
−1 q
+ t p
dt f β,p,R+ ,∼ ,
1

where
   
1+ρ −α 1+ρ −β
− 1 q > −1 and − 1 q < −1.
p p

In case p = 1, we simply have

 ∞  1  ∞
t ρ |f (t)| dt ≤ t α |f (t)| dt + t β |f (t)| dt
0 0 1

≤ f α,1,R+ ,∼ + f β,1,R+,∼ ,

and the corollary is proved.

Corollary 3.5 Let k ∈ Lpα−1 (R+ ) ∩ 
Lpβ−1 (R+ ) for some p ∈ [1, ∞) and some
p p

real numbers α, β with α < β. Then the Mellin transform 

k is holomorphic in the
strip α,β .
364 P. Junghanns and R. Kaiser

Proof By applying Lemma 3.4 we infer k ∈ 

L1ξ −1 (R+ ) for all ξ ∈ (α, β), and [5,
Lemma 2.14] yields the assertion.
p p
Let p ∈ [1, ∞), ρ, σ ∈ R. We denote by Lρ,σ := Lρ,σ (0, 1) the weighted
Lp -space equipped with the norm
 1 1/p
f ρ,σ,p := |f (x)|p υ ρ,σ (x) dx , υ ρ,σ (x) = x ρ (1 − x)σ .
0

By S we denote the Cauchy singular integral operator given by

 1
1 f (y)
(Sf )(x) = dy, x ∈ (0, 1),
πi 0 y−x

where the integral is considered as a Cauchy principal value one. It is well known
p p
that S : Lρ,σ → Lρ,σ is a linear and bounded operator if and only if 1 < p < ∞
and −1 < ρ, σ < p − 1 (see, for example, [3, Sections 1.2, 1.5]). For a measurable
function k : R+ → C we define the Mellin operator Bk by
 1  
x f (y)
(Bk f )(x) = k dy, x ∈ (0, 1).
0 y y
p p
Recall that, for p ∈ [1, ∞) and ρ ∈ R, the integral operator Bk : Lρ,0 → Lρ,0 is
bounded, if k ∈ L11+ρ (R+ ) (cf. [5, Lemma 3.7]).
p −1
1+ρ
Let p ∈ [1, ∞), ρ, σ ∈ R and ξ = p . We introduce the mapping

Zξ : Lpρ,σ → Lpσ (R+ ), f → e−ξ · f (e−· ),

 
which is a continuous isomorphism with Zξ−1 f (x) = x −ξ f (− ln x).

Lemma 3.6 For p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1), and a(t) = −i cot(πξ − iπt),
the relation Zξ SZξ−1 = Wa holds true in Lσ (R+ ), where ξ = 1+ρ
p
p .

Proof From (2.7) and the boundedness of SR+ : Lσ (R+ ) → Lσ (R+ ), we infer
p p

; ;
; ;
;F aF −1 u; ≤ SR+ L p
Lσ (R+ )
 u
σ,p,R
σ,p,R

for all u ∈ L2 (R) ∩ Lσ (R). Hence, a ∈ Mp,σ . On the other hand, if u ∈ Lσ (R+ )
p p

and   then
u = Pu,
  
e−ξ t ∞ y −ξ 
u(− ln y) dy
Zξ SZξ−1 u (t) = , t ∈ R+ ,
πi 0 y − e−t
Fredholm Theory of Singular Integral Operators 365

i.e., Zξ SZξ−1 = P Z −1 P

ξ SR+ Z  = PWa0 P,
 where we again used relation (2.7).
ξ
This proves the lemma.
Lemma 3.7 Let 1 < p ≤ 2, ρ ∈ R, and ξ = and set a(t) = 
1+ρ
k(ξ − it). We
p ,

assume that k ∈  +
q p
Lqξ −1 (R ) is satisfied for some q ∈ 1, p−1 ∩ [1, 2]. Then we
have
 1  
x g(y) 
k dy = x −ξ PF aF − PZ ξ g (− ln x) (3.7)
0 y y
p
for all g ∈ Lρ,0 and for almost all x ∈ (0, 1).
 
Proof Let κ(t) := Z ξ k (t) = e−ξ t k e−t , t ∈ R. Then κ ∈ Lq (R) and, due
to (2.4),
 ∞
1
Kf := κ ∗ f = κ(· − s)f (s) ds ∈ Lr (R), r = 1 ,
−∞ p + q −1
1

for all f ∈ Lp (R). Moreover, in virtue of (2.5),


Kf = 2πF F − κ · F − f .

Hence Kf = F aF − f for all f ∈ Lp (R), where (cf. (3.1))

 
ξ k (t) = 
a(t) = 2π F − κ (t) = 2π F − Z k(ξ − it).

On the other hand we have, for f ∈ Lp (R+ ) and for almost all x ∈ (0, 1),
 ∞
 
 (− ln x) =
x −ξ PKPf k es x eξ s f (s) ds
0
 1  
x dy
= k y −ξ f (− ln y)
0 y y
  
1 x  dy
= k Zξ−1 f (y) ,
0 y y
p
and the lemma is proved, since every g ∈ Lρ,0 can be represented in the form
g = Zξ−1 f with f ∈ Lp (R).
Lemma 3.8 ([5], Prop. 3.13) Let p ∈ (1, ∞), ρ ∈ (−1, p − 1), and k ∈ C(R+ ).
Moreover, we assume that there are real numbers α, β with α < β such that 1+ρ
p ∈
(α, β) and such that

lim t α k(t) = 0 and lim t β k(t) = 0.

t →+0 t →∞
366 P. Junghanns and R. Kaiser

p p
Then, for all σ ∈ (−1, p − 1), the integral operator Bk : Lρ,σ → Lρ,σ is bounded.
Corollary 3.9 Let p ∈ (1, ∞), ρ ∈ (−1, p − 1), and k ∈ C(R+ ). Furthermore, we
assume that there are real numbers α, β with α < β such that ξ := 1+ρ p ∈ (α, β)
and k ∈ Lqη−1 (R+ ) for some q ∈ (1, 2] and all η ∈ (α, β). Moreover, we suppose
q

that the Mellin transform k is holomorphic in the strip α,β and fulfils (3.4). Then
p p
the integral operator Bk : Lρ,σ → Lρ,σ is bounded for all σ ∈ (−1, p − 1).
Proof Choose ε > 0 such that α + ε < ξ < β − ε. Corollary 3.3 yields

lim x α+ε k(x) = lim x β−ε k(x) = 0 ,

x→+0 x→∞

and by Lemma 3.8 we get the assertion.

Lemma 3.10 Let p ∈ (1, ∞), ρ ∈ (−1, p − 1), and ξ = 1+ρ p . We assume that

k∈Lqξ −1 (R+ ) for some q ∈ 1, p−1 ∩ [1, 2] and 
q p
k(ξ + i·) ∈ V1 (R). Then the
p p
operator Bk is bounded on the space Lρ,0 ∩ Lρ,σ equipped with the norm ·ρ,σ,p .
p
Thus, the operator possesses a unique extension to the space Lρ,σ .
Proof Let a(t) = k(ξ − it). Relation (2.9) delivers a ∈ Mp,σ , which means that the
operator Wa : Lσ (R+ ) → Lσ (R+ ) is well-defined and bounded. Thus Zξ−1 Wa Zξ :
p p
p p
Lρ,σ → Lρ,σ is bounded, too. Now from (3.7) follows

Bk f = Zξ−1 Wa Zξ f,
p
f ∈ Lρ,0 ∩ Lpρ,σ ,

which completes the proof.

4 Fredholm Properties

Here we derive necessary and sufficient conditions for the Fredholmness of the
operators of the form (1.1) and also an index formula based on the winding number
of a certain closed curve in C associated with such operators.
Lemma 4.1 Let p ∈ (1, ∞), p1 + q1 = 1 and ρ, σ ∈ (−1, p − 1) as well as
k ∈ C(R+ ). For A ∈ L Lρ,σ ) we define the adjoint operator by
p

 1  1
q
(Au)(x)v(x) dx = u(x)(A∗ v)(x) dx, ∀ u ∈ Lpρ,σ , ∀ v ∈ Lρ ,σ
0 0

p
with ρ = (1 − q)ρ and σ = (1 − q)σ . If Bk ∈ L Lρ,σ ) then

(S + Bk )∗ = S + Bk1 , k1 (t) = k(t −1 )t −1 .

 Fredholm Theory of Singular Integral Operators 367

Proof The proof is straightforward if one takes into account the commutation
formula for Cauchy principal value and usual integrals (cf.[6, Chapter II, Prop. 4.4]).

Lpα−1 (R+ ) ∩ 
We recall that the assumption k ∈  Lpβ−1 (R+ ) implies that the Mellin
p p

transform k : α,β → C is holomorphic (cf. Corollary 3.5), and we formulate

the following condition for a function k ∈ C(R+ ) and 1 < p < ∞ as well as
−1 < ρ < p − 1:
(A) There exist real numbers α and β with α < β such that ξ = 1+ρ p ∈ (α, β) and
k∈Lrα−1 (R ) ∩ 
r + r +
Lrβ−1 (R ) for some r ∈ (1, 2] as well as (3.4) is satisfied.
In contrast to [5] we replaced condition
 
sup (1 + |z|)1+ 
k ()(z) : z ∈ α,β < ∞,  = 0, 1, 2, . . .

(cf. [5, equation (3.11)]) by (3.4), which is a weaker condition regarding the Mellin
transform k.
Lemma 4.2 Let a, b ∈ C, p ∈ (1, ∞), ρ ∈ (−1, p−1), p1 + q1 = 1, and k ∈ C(R+ )
satisfy condition (A). Then, for all σ ∈ (−1, p − 1), we have the representations

Zξ (aI + bS + Bk )Zξ−1 = Wa (4.1)

and
−1
Z1−ξ (aI + bS + Bk )∗ Z1−ξ = Wa = (Wa )∗ , (4.2)

in the spaces Lσ (R+ ) and L(1−q)σ (R+ ), respectively, where

p q

a(t) = a − bi cot π(ξ − it) + 

k(ξ − it), t ∈ R.

map Zξ : Lρ,σ → Lσ (R+ ) is a continuous isomorphism and

p p
Proof Since the
p 
Bk ∈ L Lρ,σ for all σ ∈ (−1, p − 1) (cf. Corollary 3.9), Lemma 3.7 delivers
immediately the relation

 ξ Bk Z −1 f = PF
PZ 
k(ξ − i·)F −1 Pf, f ∈ L2 (R+ ) ∩ Lpσ (R+ ).
ξ

k(ξ −i·) ∈ Mp,σ (cf. (2.9)). This yields Zξ Bk Zξ−1 = Wb

Due to (3.4), we have that 
with b(t) = k(ξ − it). Together with Lemma 3.6 we get (4.1). Relation (4.2) is now
−1
a consequence of Zξ∗ = Z1−ξ and Lemma 2.1.
Lemma 4.3 ([2], Corollary 1.19) Let p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1), and a, b ∈
p
C. Moreover, let ξ = 1+ρ
p , η = p . Then the operator A = aI + bS : Lρ,σ →
1+σ
368 P. Junghanns and R. Kaiser

p
Lρ,σ is Fredholm if and only if a ± b = 0 and there exists an integer κ satisfying

− ξ < #ν < 1 − ξ and η − 1 < #ν + κ < η , (4.3)

where
a+b
= e2πiν . (4.4)
a−b

(Note that κ can only take the values 0, −1, and 1). In this case, we have ind A = κ
and the operator A is invertible, invertible from the left, or invertible from the right,
if the index of A is zero, −1, or 1, respectively. The corresponding (one-sided)
inverse is given by
  1  ν   
  1 b x 1 − y ν+κ f (y)
A(−1) f (x) = af (x) − dy ,
a 2 − b2 πi 0 y 1−x y−x

−1 4 5
i.e., A(−1) = a 2 − b 2 aI − bv ν,−ν−κ Sv −ν,ν+κ I .
a−b = e
Moreover, for a+b 2πiμ , μ = 0, and 0 ≤ #μ < 1,

 1
b v μ−1,−μ (y) dy
av μ−1,−μ
(x) + = 0, 0 < x < 1, (4.5)
πi 0 y−x
p
i.e., in case κ = 1, we have μ = 1+ν and the nullspace of the operator A : Lρ,σ →
p
Lρ,σ is spanned by v ν,−ν−1 .
Corollary 4.4 In (4.4) we can choose ν in such a way that 0 ≤ #ν < 1. Then,

a−b
= e2πi(1−ν)
a+b

and, due to (4.5),


b 1 v −ν,ν−1 (y) dy
a v −ν,ν−1 (x) − = 0, 0 < x < 1. (4.6)
πi 0 y−x

Moreover,

b 1 v −ν,ν (y) dy
a v −ν,ν (x) − = γ0 , 0 < x < 1, (4.7)
πi 0 y−x

and
 1
1−ν,ν b v 1−ν,ν (y) dy
av (x) − = δ0 + δ1 x , 0 < x < 1, (4.8)
πi 0 y−x
Fredholm Theory of Singular Integral Operators 369

with certain constants γ0 , δ0 , δ1 ∈ C, where (4.7) remains true also in case of −1 <
#ν < 0.
Proof It remains to prove relations (4.7) and (4.8). By using (4.6), we get

b 1 v −ν,ν (y) dy
a v −ν,ν (x) −
πi 0 y−x
!  1 −ν,ν−1 "  1
b v (y) dy b
= (1 − x) a v −ν,ν−1 (x) − + υ −ν,ν−1 (y) dy
πi 0 y−x πi 0
 1
b
= υ −ν,ν−1 (y) dy , 0 < x < 1,
πi 0

and (4.7) is proved. Analogously,

 1
b v 1−ν,ν (y) dy
a v 1−ν,ν (x) −
πi 0 y−x
!  1 −ν,ν−1 "
−ν,ν−1 b v (y) dy
= x(1 − x) a v (x) −
πi 0 y−x
 1
b
+ (y + x − 1)υ −ν,ν−1 (y) dy
πi 0
 1
b
= (y + x − 1)υ −ν,ν−1 (y) dy , 0 < x < 1,
πi 0

which proves (4.8). To prove (4.7) in case −1 < #ν < 0, we use a−b
a+b = e2πi(−ν)
and get from (4.5) the relation

−ν−1,ν b 1 υ −ν−1,ν (y) dy
aυ (x) − = 0, 0 < x < 1,
πi 0 y−x

and analogously as above we conclude (4.7) also in this case.

Definition 4.5 For p ∈ (1, ∞) and σ ∈ (−1, p − 1), we define the function σ∞ :
Ṙ → R by
 1
p : t ∈R
σ∞ (t) := 1+σ
p : t = ∞.
370 P. Junghanns and R. Kaiser

Moreover, for a ∈ PCp,σ , we define the function ap,σ : Ṙ × R → C by

14 5
ap,σ (t, x) = a(t + 0) + a(t − 0)
2
(4.9)
i4 5 4 5
− a(t + 0) − a(t − 0) cot π σ∞ (t) − ix .
2
Then, the image of ap,σ defines a closed curve in the complex plane, which we
denote by ap,σ . Under the assumption
 
inf |ap,σ (t, x)| : t ∈ R, x ∈ R > 0 (4.10)

the winding number wind ap,σ is well defined, where the orientation of ap,σ is
given by its inherent parametrization.
Lemma 4.6 ([1], Theorem 5.7, cf. also [8], Theorem 1.2) Let p ∈ (1, ∞), σ ∈
(−1, p − 1) and a ∈ PCp,σ . Then Wa is a Fredholm operator on Lσ (R+ ) if
p

and only if (4.10)

 is satisfied. In this case, the Fredholm index of Wa is equal to
− wind ap,σ .
1+ρ
In the following, let ξ = p for p ∈ (1, ∞) and ρ ∈ (−1, p − 1).

Corollary 4.7 Let a, b ∈ C, p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1), and k ∈ C(R+ )

satisfy condition (A). Then the integral operator A := aI + bS + Bk is Fredholm
p
in the space Lρ,σ if and only if the curve
 
A := a − bi cot π(ξ − it) + 
k(ξ − it) : t ∈ R
   (4.11)
1+σ
∪ a + bi cot π − it : t ∈ R
p
p
does not run through the zero point. In this case, the Fredholm index of A : Lρ,σ →
p
Lρ,σ is equal to negative winding number of the curve A , where the orientation is
given by its inherent parametrization.
Proof From Lemma 4.2 follows

A = Zξ−1 Wa Zξ in Lpρ,σ ,

where a(t) = a − bi cot π(ξ − it) +  k(ξ − it), t ∈ R. Hence, a(∞ ± 0) = a ∓ b,

such that, due to (4.9),
⎧
⎨ a(t) : t ∈ R, x ∈ R,
ap,σ (t, x) =  
⎩ a + ib cot 1+σ − ix : t = ∞, x ∈ R .
p
Fredholm Theory of Singular Integral Operators 371

Consequently, the curves A and ap,σ coincide. Since furthermore

a ∈ V1 (R) ∩ PC(R) ⊂ PCp,σ ,

Lemma 4.6 delivers the assertion.

Using Lemma 3.10, we can prove in the same way the following corollary.
Corollary 4.8 Let p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1) and a, b ∈ C. We assume that
the function k belongs to 
Lqξ −1 (R+ ) for some q ∈ 1, p−1
q p
∩ [1, 2] and fulfils

k(ξ + i·) ∈ V1 (R). Moreover, for t ∈ R, we set

a(t) := a − biπ cot(ξ − it) + 

k(ξ − it).

Then

Zξ−1 Wa Zξ u = (aI + bS + Bk )u
p
∀u ∈ Lρ,0 ∩ Lpρ,σ

and Zξ−1 Wa Zξ is Fredholm on the space Lρ,σ if and only if the curve (4.11) does
p

not run through the zero point. In this case, the Fredholm index of Zξ−1 Wa Zξ :
p p
Lρ,σ → Lρ,σ is equal to the negative winding number of this curve, where the
orientation is given by its inherent parametrization.
4 5n 4 5n
For a = aj j =1 ∈ PC(R)n and b = bj j =1 ∈ PCnp,σ , now we consider
operators of the form


n
"a,b := aj Wb0j . (4.12)
j =1

Moreover, let


n 
n
a± (t) := aj (t)bj (±∞) and b± (t) := aj (±∞)bj (t).
j =1 j =1

Definition 4.9 Let n ∈ N and aj ∈ PC(R), bj ∈ PCp,σ for j = 1, · · · , n. Under

the assumptions

inf {|a+ (t)| : t ∈ R} > 0 and inf {|b− (t)| : t ∈ R} > 0 , (4.13)
372 P. Junghanns and R. Kaiser

we define the functions Ap,σ : R × R → C and Bp,σ : R × R → C by

! "
1 a− (t + 0) a− (t − 0)
Ap,σ (t, x) := +
2 a+ (t + 0) a+ (t − 0)
! "
i a− (t + 0) a− (t − 0)
− − cot π [σ0 (t) − ix] , t ∈ R,
2 a+ (t + 0) a+ (t − 0)

a− (±∞)
Ap,σ (±∞, x) := ,
a+ (±∞)

and
! "
1 b+ (t + 0) b+ (t − 0)
Bp,σ (t, x) := +
2 b− (t + 0) b− (t − 0)
! "  
i b+ (t + 0) b+ (t − 0) 1
− − cot π − ix , t ∈ R,
2 b− (t + 0) b− (t − 0) p

b+ (±∞)
Bp,σ (±∞, x) := ,
b− (±∞)

respectively, where
 1
p : t ∈ R\{0}
σ0 (t) := 1+σ
p : t = 0.

If the conditions
 
inf |Ap,σ (t, x)| : t, ∈ R, x ∈ R > 0 (4.14)

and
 
inf |Bp,σ (t, x)| : t ∈ R, x ∈ R > 0 (4.15)

are satisfied, we define κp,σ as the increment of the function

1 4 5
argAp,σ (t, x) + argBp,σ (t, x) ,
2π
a− b+
where t runs over R, and at the points of discontinuity of a+ or b− , the variable x
runs over R.
Fredholm Theory of Singular Integral Operators 373

Lemma 4.10 ([1], Theorem 5.7, cf. also [8], Theorem 3.2) Let n ∈ N, p ∈
(1, ∞), σ ∈ (−1, p − 1) and aj ∈ PC(R), bj ∈ PCp,σ for j = 1, · · · , n. Then
p
"a,b is a Fredholm operator on Lσ (R) if and only if (4.13), (4.14), and (4.15) hold
p p
true. In this case, the Fredholm index of "a,b : Lσ (R) → Lσ (R) is equal to −κp,σ .
A function a : [0, 1] → C is called piecewise continuous, if it is continuous
at 0 and 1, if the one-sided limits a(x ± 0) exist for all x ∈ (0, 1), and if at least
one of them coincides with a(x). The set of all piecewise continuous functions a :
[0, 1] → C having only a finite number of jumps is denoted by PC[0, 1].
Corollary 4.11 Let p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1) and a, b ∈ PC[0, 1], and
assume that k ∈ C(R+ ) satisfies condition (A). Then the integral operator A :=
p
aI + bS + Bk is Fredholm on the space Lρ,σ if and only if a(x ± 0) + b(x ± 0) = 0
for all x ∈ (0, 1), if a(x) + b(x) = 0 for x ∈ {0, 1}, and if the closed curve
d
A := 
0d ∪ 1d ∪ 1d ∪ . . . ∪ N
d
∪ Nd ∪ N+1
d
∪
1d

does not contain the point 0. Here N stands for the number of discontinuity points
xj , j = 1, . . . , N, of the function

a(x) − b(x)
d(x) =
a(x) + b(x)

chosen in such way that x0 := 0 < x1 < · · · < xN < xN+1 := 1. Using these xj ,
the curves jd , j = 1, . . . , N + 1 and jd , j = 1, . . . , N are given by
 
jd := d(x) : xj −1 < x < xj

and

14 5
jd := d(xj + 0) + d(xj − 0)
2
 
i4 5 1
+ d(xj + 0) − d(xj − 0) cot π + it :t ∈R ,
2 p

respectively. The curves 

jd , j ∈ {0, 1}, connecting the point 1 with one of the
d d
endpoints of 1 and N+1 , respectively, are given by the formulas

 a(0) − b(0)i cot π(ξ + it) + 
k(ξ + it)
0d := :t ∈R
a(0) + b(0)
374 P. Junghanns and R. Kaiser

and
  
 14 5 i4 5 1+σ
1d := 1 + d(1) + 1 − d(1) cot π + it :t∈R .
2 2 p

In this case, the Fredholm index of A is equal to the winding number of the curve
A , where the orientation of A is due to the above given parametrization.
Proof Due to Lemma 4.2, we have the representation

A = Zξ−1 
aI + 
bWa + Wb Zξ in Lpρ,σ

with a(t) = −i cot π(ξ − it), b(t) =  a (t) = a(e−t ), 

k(ξ − it) and  b(t) = b(e−t ).
aI + 
It is well known, that the Fredholm properties of the operator  bWa + Wb in
Lσ (R+ ) are equivalent to the Fredholm properties of
p

 = a1 I + b1 Wa0 + χ+ W 0 + χ− I
A in Lpσ (R),
b

where
 
a (t) : t ∈ R+ ,
 
b(t) : t ∈ R+ ,
a1 (t) = b1 (t) =
0 : otherwise, 0 : otherwise,

χ− (t) = 1 − χ+ (t), and


1 : t ∈ R+ ,
χ+ (t) =
0 : otherwise.

 is of the form (4.12), we can make use of Lemma 4.10. We have

Since A

a+ (t) = a1 (t) + b1 (t) + χ− (t), a− (t) = a1 (t) − b1 (t) + χ− (t)

and

b+ (t) = a(0) − b(0)i cot π(ξ − it) + 

k(ξ − it), b− (t) = 1.

Due to our assumptions, condition (4.13) is fulfilled, and hence we have, for 0 <
t < ∞,
! "
1 a (t + 0) − 
b(t + 0) a (t − 0) − 
b(t − 0)
Ap,σ (t, x) = +
2 a (t + 0) + 
b(t + 0) a (t − 0) + 
b(t − 0)
! "  
i a (t + 0) − 
b(t + 0) a (t − 0) − 
b(t − 0) 1
− − cot π − ix .
2 a (t + 0) + 
b(t + 0) a (t − 0) + 
b(t − 0) p
Fredholm Theory of Singular Integral Operators 375

Moreover,
! " ! "  
1 a(1) − b(1) i a(1) − b(1) 1+σ
Ap,σ (0, x) = +1 − − 1 cot π − ix ,
2 a(1) + b(1) 2 a(1) + b(1) p

a(0) − b(0)
Ap,σ (+∞, x) = ,
a(0) + b(0)

and, for −∞ < t < 0,

Ap,σ (t, x) = Ap,σ (−∞, x) = 1 .

Finally,

Bp,σ (±∞, x) = a(0) ± b(0)

and, for t ∈ R,

Bp,σ (t, x) = a(0) − b(0)i cot π(ξ − it) + 

k(ξ − it)

 is Fredholm on Lσ (R) if and only if

Applying Lemma 4.10, we get, that A
p

 
a (t) + 
inf | b(t)| : t ∈ R+ > 0,
 
inf |a(0) − b(0)i cot π(ξ − it) + 
k(ξ − it)| : t ∈ R > 0,

and
 
inf |Ap,σ (t, x)| : t, ∈ R, x ∈ R > 0.

But this is obviously equivalent to our assumptions. In this case, the Fredholm index
 is equal to −κp,σ , where κp,σ is defined as the increment of the function
of A

1 Bp,σ (t, x)
arg Ap,σ (t, x) + arg ,
2π a(0) + b(0)
 
where t runs over R, and at the points of discontinuity of a −b /a +b , the
variable x runs over R. But κp,σ is equal to the negative winding number of the
curve A . This completes the proof.
p p
We define the operator R : Lρ,σ → Lσ,ρ by (Rf )(x) = f (1 − x). Moreover,
we set

A := aI + bS + c+ Bk+ + c− RBk− R, (4.16)

where a, b, c± ∈ L∞ (0, 1) and k± ∈ C(R+ ).

376 P. Junghanns and R. Kaiser

Theorem 4.12 Let p ∈ (1, ∞), σ± ∈ (−1, p − 1) and a, b, c± ∈ PC[0, 1]. If the
functions k± ∈ C(R+ ) satisfy condition (A) for ξ = ξ± := 1+σ ±
p , then the operator
p
A defined in (4.16) is Fredholm on Lσ+ ,σ− if and only if a(x ± 0) − b(x ± 0) = 0
for all x ∈ (0, 1), if a(x) − b(x) = 0 for x ∈ {0, 1}, and if the closed curve
c
A := 
0c ∪ 1c ∪ 1c ∪ . . . ∪ N
c
∪ Nc ∪ N+1
c
∪
1c

does not contain the point 0. Here N stands for the number of discontinuity points
xj , j = 1, . . . , N, of the function

a(x) + b(x)
c(x) =
a(x) − b(x)

chosen in such way that x0 := 0 < x1 < · · · < xN < xN+1 := 1. Using these xj ,
the curves jc , j = 1, . . . , N + 1 and jc , j = 1, . . . , N are given by
 
jc := c(x) : xj −1 < x < xj

and

14 5
jc := c(xj + 0) + c(xj − 0)
2
 
i4 5 1
− c(xj + 0) − c(xj − 0) cot π − it :t ∈R .
2 p

The curves j , j ∈ {0, 1} connecting the point 1 with one of the endpoints of 1c
c
and N+1 , respectively, are given by the formulas

a(0) − b(0)i cot π(ξ+ − it) + c+(0)
k+ (ξ+ − it)

0c := :t ∈R
a(0) − b(0)

and

a(1) + b(1)i cot π(ξ− − it) + c− (1)
k− (ξ− − it)

1c := :t ∈R .
a(1) − b(1)

In this case, the Fredholm index of A is equal to the negative winding number of the
c c
curve A , where the orientation of A is due to the above given parametrization.
Proof At first, we consider the case c+ = 1 and c− = 0. Setting ρ = σ+ , σ = σ− ,
ξ = ξ+ = 1+ρp , and ξ− = p . Having regard to Corollary 4.11, we have to show
1+σ

0 ∈ A
c
is equivalent to 0 ∈ Ad c
and that in this case wind A = −wind A d
.
Fredholm Theory of Singular Integral Operators 377

But, this is seen by the relations

  
1
jd = : xj −1 < x < xj = z−1 : z ∈ jc ,
c(x)

14 5
jd = d(xj + 0)d(xj − 0) c(xj + 0) + c(xj − 0)
2
 
i4 5 1
− c(xj + 0) − c(xj − 0) cot π + it : t ∈ R ,
2 p

 a(0) − b(0) a(0 − b(0)i cot π(ξ + it) +  k(ξ + it)
0d = :t∈R ,
a(0) + b(0) a(0) − b(0)
  
1 i 1+σ

1d = d(1) [1 + c(1)] − [1 − c(1)] cot π + it : t ∈ R ,
2 2 p

if one compares the increments of the argument along the respective pieces of the
c d
curves A and A .

5 Regularity Properties of Solutions of Eq. (1.1)

This section deals with some specific properties of the Mellin operators and the
smoothness of the solutions of the equation

au + bSu + cBk u = f in Lpρ,σ .

In addition, their asymptotic behaviour near the end point 1 is investigated. Let
n ∈ N0 . For k ∈ Cn (R+ ) and f : (0, 1) → C we define the operators ∂n Bk by
 1  
x f (y)
(∂n Bk f )(x) := k (n) dy.
0 y y n+1

Lemma 5.1 Let p ∈ (1, ∞), σ < p−1, ρ ∈ R, and n ∈ N0 as well as k ∈ Cn (R+ ).
We assume that there is a β > 1+ρ
p such that the functions
 
 x
k := [a, 1] × [0, 1] → C, (x, y) → y −(β+) k () (5.1)
y
378 P. Junghanns and R. Kaiser

p
are continuous for all a ∈ (0, 1) and all  ∈ {0, 1, . . . , n}. Then, for every f ∈ Lρ,σ ,
the function Bk f is n-times continuously differentiable on (0, 1], where

d
(Bk f )(x) = (∂ Bk f )(x), x ∈ (0, 1] (5.2)
dx 
holds true.
Proof Let ε > 0. Then, there exists a δ = δ(ε) > 0, such that

(∂ Bk f )(x) − (∂ Bk f )(x )

 1
≤ k (x, y) − 
 k (x , y) y β−1 |f (y)| dy
0
 1
β−1− pρ ,− σp ρ σ
,p
≤ε υ (y) υ p (y) |f (y)| dy
0

 1  
p
 p−1
p
β−1− pρ p−1 σ
,− p−1
≤ε υ (y) dy f ρ,σ,p
0

≤ const · ε

 all x, x ∈ [a, 1], which satisfy |x − x | < δ. Here we took into account that
for
β − 1 − pρ p−1 p
> −1 is equivalent to β > 1+ρ p and that σ < p − 1. Hence
∂ Bk f is continuous on (0, 1]. Analogously, one can show that, for x ∈ [a, 1], the
relations

∂ Bk f (x) ≤ const f ρ,σ,p , (5.3)

are true, where the constant does only depend on a ∈ (0, 1). The differentiability of
Bk f follows now from
 x  x  1  
ξ f (y)
(∂1 Bk f )(ξ ) dξ = k dy dξ
c c 0 y y2
 x 1    x ε  
ξ f (y) ξ f (y)
= lim k 2
dy dξ + lim k dy dξ
ε→0 c ε y y ε→0 c 0 y y2
 1  x    x ε  
ξ f (y) ξ f (y)
= lim k 2
dξ dy + lim k dy dξ
ε→0 ε c y y ε→0 c 0 y y2
 1 !     "  x ε  
x f (y) c f (y) ξ f (y)
= lim k −k dy + lim k dy dξ,
ε→0 ε y y y y ε→0 c 0 y y2
Fredholm Theory of Singular Integral Operators 379

where x, c ∈ (0, 1]. With regard to (5.3), we can apply Lebesgue’s dominated
convergence theorem and get
 x  1!  
x f (y)
 
c f (y)
"
(∂1 Bk f )(ξ ) dξ = k −k dy = (Bk f )(x) + const.
c 0 y y y y

d
Hence, dx (Bk f )(x) = (∂1 Bk f )(x), x ∈ (0, 1]. The general case follows now by
induction.
Remark 5.2 Note, that the function  k defined in (5.1) is continuous for all a ∈
(0, 1) if the limit lim t β+ k () (t) exists and is finite.
t →∞
Let ψ, ζ ≥ 0. By BCψ,ζ = BCψ,ζ (0, 1) we denote the set of all continuous
functions f : (0, 1) → C, for which the function υ ψ,ζ f is bounded on (0, 1). If we
introduce the norm
 
f ψ,ζ,∞ = sup υ ψ,ζ (x)|f (x)| : 0 < x < 1 ,

then BCψ,ζ becomes a Banach space. Moreover by Cψ,ζ = Cψ,ζ (0, 1) we denote
the set of all continuous functions f : (0, 1) → C, for which the limits

lim x ψ f (x) and lim (1 − x)ζ f (x)

x→0 x→1

exist and if these limits are equal to zero if ψ > 0 or ζ > 0, respectively. The space
Cψ,ζ is a closed subspace of BCψ,ζ and, consequently, also a Banach space.
Lemma 5.3 Let p ∈ (1, ∞), 1
p + 1
q = 1, σ < p − 1, and ρ ∈ R, as well as
k∈ C(R+ ). Furthermore, we assume that there is a β > 1+ρ
p such that the function
 
 x
k := [a, 1] × [0, 1] → C, (x, y) → y −β k
y
is continuous for all a ∈ (0, 1). If
 ∞ 
ρ

+1 q−2
|k(t)|q t p dt < ∞
0

and if there are numbers χ ∈ R, t0 > 0, and c0 > 0 such that

|k(t)| ≤ c0 t −χ , 0 < t < t0 , (5.4)

then
⎧   
⎨ L Lpρ,σ , BCψ,0 : max 0, 1+ρ
p , χ ≤ψ,
Bk ∈   
⎩ L Lp , C 1+ρ
ρ,σ ψ,0 : max 0, p , χ < ψ .
380 P. Junghanns and R. Kaiser

p
Proof First we note that, due to Lemma 5.1, for f ∈ Lρ,σ , the function (Bk f ) (x)
is continuous on (0, 1]. Moreover, for x ∈ (0, 1), using Hölder’s inequality we can
estimate
 1  
x f (y) dy
v ψ,0 (x) (Bk f ) (x) = x ψ k
0 y y
   − ρq σq 1
p ,− p
1 q q
x v (y) dy
≤x ψ
k f ρ,σ ,p
0 y yq

=: N(x) f ρ,σ ,p ,

where, with the help of the substitution t = xy ,

  x − ρq −q 
∞ x − p dt
σq
p
N(x)q = x ψq+1 |k(t)|q 1−
x t t t2
  ∞  x − ρq −q  x − p dt
2x σq
p
= x ψq+1 + |k(t)|q 1−
x 2x t t t2

=: N1 (x) + N2 (x) .
 t0 
Setting M0 = max |k(t)| : 2 ≤ t ≤ 2 and taking (5.4) into account, we get
⎧   x − ρq −q  x − p
2x σq
⎪
⎪ q p
t −χq−2
1− dt : 2x < t0 ,
⎨ c0
t t
N1 (x) ≤ x ψq+1 x
 2x  − ρq −q 
⎪ x − p
σq
⎪
⎩ Mq −2 x p
0 t 1 − dt : t0 ≤ 2x ,
x t t
⎧  2
⎪
⎪ q (ψ−χ)q ρq σq
p +q−2+ p −χq (s − 1)
− σpq
ds : 2x < t0 ,
⎨ c0 x s
t =xs
= 1 2
⎪
⎪ ρq σq
− σpq
⎩ q
M0 x ψq s p +q−2+ p (s − 1) ds : t0 ≤ 2x ,
1

and
   ∞

ρ

p +1
σq
ψq+1− ρq
p −q
q−2
N2 (x) ≤ max 1, 2 p x |k(t)| t q
dt .
0

σq ρq
Since p = σ
p−1 < 1 and since ψq + 1 − p − q ≥ (>) 0 is equivalent to
1+ρ
ψ ≥ (>) p , we get the assertion.
Fredholm Theory of Singular Integral Operators 381

Writing
 1  
− x f (y) dy
(∂ Bk f ) (x) = x k
0 y y

with k (t) = k ()(t)t  , from Lemma 5.3 we get

Corollary 5.4 Let p ∈ (1, ∞), 1
p + 1
q = 1, σ < p − 1, and ρ ∈ R,  ∈ N0 , as
well as k ∈ C (R+ ).Furthermore, we assume that there is a β > 1+ρ
p such that the

function k defined in (5.1) is continuous for all a ∈ (0, 1). If
 ∞ 
ρ

q p +1 q+q−2
|k (t)| t
()
dt < ∞
0

and if there are numbers χ ∈ R, t0 > 0, and c0 > 0 such that

|k ()(t)| ≤ c0 t −χ− , 0 < t < t0 ,

then
⎧   
⎨ L Lpρ,σ , BCψ,0 :  + max 0, 1+ρ
p , χ ≤ψ,
∂ Bk ∈   
⎩ L L ,C
p 1+ρ
ρ,σ ψ,0 :  + max 0, p , χ < ψ .

Corollary 5.5 Let p ∈ (1, ∞), σ < p − 1, ρ ∈ R, and  ∈ N0 , k ∈ C (R+ ). If

there are numbers α, β ∈ R such that α < 1+ρ
p < β, that t
α+ k () (t) is bounded

for t → +0, and that the finite limit lim t β+ k ()(t) exists, then
t →∞

⎧   
⎨ L Lpρ,σ , BCψ,0 :  + max 0, 1+ρ
p ≤ψ,
∂ Bk ∈   
⎩ L L ,C
p 1+ρ
ρ,σ ψ,0 :  + max 0, p <ψ.

ProofTaking intoaccount Remark 5.2, we can apply Corollary 5.4 with χ = α,

since p1 + q1 = 1

 ∞

ρ

p +1 q+q−2
|k ()(t)|q t dt
0
 1 ρ   ∞ ρ 
p +1 q−αq−2 p +1 q−βq−2
≤ c1 t dt + c2 t dt
0 1

and
   
ρ 1+ρ 1
+ 1 q − αq − 2 = −α− q > −1
p p q
382 P. Junghanns and R. Kaiser

as well as
   
ρ 1+ρ 1
+ 1 q − βq − 2 = −β − q < −1 .
p p q

Corollary 5.6 Let a, c, f ∈ Cn (0, 1] and let the assumptions of Lemma 5.1 be
p
fulfilled. If the operator equation au + cBk u = f has a solution u ∈ Lρ,σ and if
a(x) = 0 for all x ∈ (0, 1], then u ∈ C (0, 1].
n

Proof With the help of Lemma 5.1, we get u = a −1 (f − cBk )u ∈ Cn (0, 1].
Corollary 5.7 Let a, c, f ∈ Cn [0, 1] and a(x) = 0 for all x ∈ [0, 1]. If the
conditions of Corollary 5.5 are satisfied and if the equation au + cBk u = f has
p
a solution u ∈ Lρ,σ , then

u() ∈ C 1+ρ ++ε,0 for  ∈ {0, . . . , n} and all ε > 0.

p

Proof Due to the assumptions, Corollary 5.5 in combination with (5.2) deliver
4 5()
u() = a −1 (f − cBk )u ∈ C 1+ρ ++ε,0
p

for  ∈ {0, . . . , n} and all ε > 0.

Let 0 < γ < 1, m ∈ N0 , and −∞ < a < b < ∞. We denote by Cm,γ [a, b] the set
of all m times differentiable functions f : [a, b] → C, for which
 
|f (m) (x) − f (m) (y)|
sup : a ≤ x, y ≤ b, x = y
|x − y|γ

is finite. For c ∈ [a, b], m ∈ N0 , and 0 < γ < 1, we set

m,γ  
C[a,b] (c) = f : ∃ε > 0 with f ∈ Cm,γ [c − ε, c + ε] ∩ [a, b]

Moreover , let
< < m,γ
Cm,0 [a, b] = Cm,γ [a, b] and Cm,0
[a,b] (c) = C[a,b] (c) .
0<γ <1 0<γ <1

For γ ∈ [0, ∞) and −∞ < a < b < ∞, the function class Hγ (a, b) is defined as
the set of all functions f : (a, b) → C belonging to C[γ ],γ −[γ ] [c, d] for all intervals
[c, d] ⊂ (a, b), where [γ ] is the integer, which fulfills γ − 1 < [γ ] ≤ γ . In the
same manner we define the classes Hγ (a, b], Hγ [a, b), and Hγ [a, b]. Of course,
Hγ [a, b] = C[γ ],γ −[γ ] [a, b].
Lemma 5.8 Let p ∈ (1, ∞), σ < p − 1, n ∈ N0 and ρ, β ∈ R with
1+ρ
p < β. For k ∈ C(R+ ), assume that t β k(t) is bounded for t → ∞ (i.e.,
Fredholm Theory of Singular Integral Operators 383

 
ma := sup t β k(t) : a ≤ t < ∞ < ∞ for all a > 0) and that
 
s β k(s) − t β k(t)
Ma := sup : s, t ∈ [a, ∞), s = t <∞
|s − t|γ

1+ρ
for some γ ∈ (0, 1) satisfying p +γ < β and for all a > 0. Then Bk f ∈ Hγ (0, 1]
p
for all f ∈ Lρ,σ .
p
Proof Let f ∈ Lρ,σ and 0 < a ≤ x1 < x2 ≤ 1. Then

|(Bk f )(x1 ) − (Bk f )(x2 )|

|x1 − x2 |γ
     β  −β
 1 k x1 − k x2 x1 x1
y y y y |f (y)| dy
≤ x1 x2 γ
0 |y − y| y 1+γ
 β    β    −β
x1
 k xy1 − xy2 k xy2 x1
1 y y |f (y)|
≤ dy
0 | xy1 − x2 γ
y | y 1+γ
  ! β  β "
x2 x2 x1
 k −
1 y y y |f (y)| dy
+  β
y 1+γ
0 | xy1 − x2 γ
y |
x1
y

−β |f (y)| dy
1
≤ x1 Ma 1+γ −β
0 y
 β−γ
 1 x β  
x2
x1 − 1 x2 |f (y)| dy
2
+ β γ
k
x2 x1 0 y y y 1−β
  β−γ  1
1 |f (y)| dy |f (y)| dy
≤ a −β Ma + a −β−γ
a −1
− 1 m a
0 y 1+γ −β 0 y 1−β
⎡
 1    q1
−1−γ +β− pρ q.− p−1
σ
≤ const ⎣ υ (y) dy
0

⎤
 1
   q1
−1+β− pρ q,− p−1
σ
+ υ (y) dy ⎦ f ρ,σ,p
0

≤ const f ρ,σ,p ,
384 P. Junghanns and R. Kaiser

where
 the constant does not depend on x1 and x2 and where we took into account
that −1 − γ + β − pρ q > −1 is equivalent to 1+ρ p + γ < β.

Lemma 5.9 ([7], Theorem in §19) Let 0 ≤ a < b ≤ 1 and v ∈ C0,γ [a, b] for
some γ ∈ [0, 1). Then we have

Sχ[a,b] v ∈ Hγ (a, b),

where χ[a,b] is the characteristic function of the interval [a, b]. Moreover, if v(a) =
0 or v(b) = 0 are satisfied, then we even get

Sχ[a,b] v ∈ Hγ [a, b) and Sχ[a,b] v ∈ Hγ (a, b],

respectively.
Corollary 5.10 Let 0 ≤ a < b ≤ 1 and v ∈ Hγ (a, b) for some γ ∈ [0, ∞) as well
as χ[a,b] v ∈ L10,0 . Then Sχ[a,b] v ∈ Hγ (a, b).
Proof Without loss of generality, we can assume that a = 0 and b = 1. Let [c, d] ⊂
(0, 1) and choose ε > 0 such that ε < c and d < 1 − ε. Write

Sv = I1 + I2 + I3

with
 ε  1−ε
1 v(y) dy 1 v(y) dy
I1 (x) = , I2 (x) = ,
πi 0 y−x πi ε y−x

and
 1
1 v(y) dy
I3 (x) = .
πi 1−ε y−x

Then I1 , I3 ∈ C∞ [c, d].

Now we consider I2 . If γ ∈ [0, 1), then, by our assumptions, v ∈ C0,γ [ε, 1 − ε],
and Lemma 5.9 delivers I2 ∈ C0,γ [c, d]. Assume that

I2 ∈ C[γ ],γ −[γ ] [c, d] if γ ∈ [0, n) (5.5)

Fredholm Theory of Singular Integral Operators 385

for some n ∈ N. We show that then I2 ∈ C[γ ],γ −[γ ] [c, d] also holds if γ ∈ [n, n+1).
Indeed, in that case v ∈ Hγ −1 (0, 1) and, by [6, Lemma 6.1],

v(ε) v(1 − ε) 1−ε v (y) dy
I2 (x) = − − , ε < x < 1−ε.
ε−x 1−ε−x ε y−x

By our assumption (5.5),

 1−ε v (y) dy
∈ C[γ ]−1,γ −[γ ] [c, d] ,
ε y−·

which implies I2 ∈ C[γ ],γ −[γ ] [c, d]. The corollary is proved.
The following corollary can be proved analogously.

Corollary 5.11 Let 0 ≤ a < b ≤ 1, c ∈ {a, b} and v ∈ Hγ (a, b) ∪ {c} with
v(c) = v (c) = . . . = v [γ ] (c) = 0 for some γ ∈ [0, ∞) as well as χ[a,b] v ∈ L10,0 .
Then

Sχ[a,b] v ∈ Hγ (a, b) ∪ {c} .

Lemma 5.12 ([7], §29, 3o , 4o , §32) Let 0 ≤ α < 1, β ∈ R, and μ = α + iβ as

well as u ∈ H0 (0, 1) ∩ L10,0 . Moreover, assume that there are ε, δ ∈ (0, 1) such that,
for 1 − ε < x < 1,

u∗ (x)
u(x) = and u∗ ∈ C0,δ
[0,1] (1) ,
(1 − x)μ

where, for x ∈ (0, 1),

(1 − x)μ = (1 − x)α eiβ ln(1−x) .

Then, for x ∈ (1 − ε, 1), we have

⎧ ∗
⎪
⎪
u (1)
ln(1 − x) + u∗∗ (x) : μ = 0 ,
⎨
πi
(Su)(x) =
⎪
⎪ i cot(πμ) ∗
⎩ u (1) + u∗∗ (x) : μ = 0,
(1 − x)μ
386 P. Junghanns and R. Kaiser

where, in case α = 0,

u∗∗ ∈ C0,δ
[0,1] (1)

and, in case 0 < α < 1,

υ 0,α0 u∗∗ ∈ C0,λ

[0,1] (1)

with υ 0,α0 u∗∗ (1) = 0 for

max {0, α − δ} < α0 < α and 0 < λ < α0 − max {0, α − δ} .

Remark 5.13 If (4.3) holds for some κ ∈ {−1, 0, 1}, then

v −ν,ν+κ Lpρ,σ ⊂ L10,0 .

Proof This follows immediately from

 1  1  p1
−ν,ν+κ
|(v f )(x)| dx ≤ |f (x)| v p ρ,σ
(x) dx
0 0
 1  
p
 
p
 p−1
p
−#ν− pρ p−1 , #ν+κ− σp p−1
× v (x) dx <∞
0

p
for f ∈ Lρ,σ .
Remark 5.14 ([7], §7) For α > 0 and β ∈ R, the function x → (1−x)α+iβ belongs
to C0,α 0,δ
[0,1] (1). Hence, if f ∈ C[0,1] (1) for some δ ∈ [0, 1), then

g ∈ Cmin{α,δ}
[0,1] (1) ,

where g(x) = (1 − x)α+iβ f (x).

Proposition 5.15 Let p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1), 0 ≤ γ < ∞, and a ∈
C, b ∈ C \ {0} with a ± b = 0. Moreover, assume that there is a ν ∈ C and a
p
κ ∈ {0, ±1} such that (4.3) is fulfilled. If u ∈ Lρ,σ and

f = au + bSu ∈ Hγ (0, 1) ∩ C0,δ

[0,1] (1)
Fredholm Theory of Singular Integral Operators 387

for some δ ∈ (0, 1), then

⎧ ⎧ ⎫
⎪
⎪ ⎪
⎪ u ∈ Hγ (0, 1) ∩ C0,δ1
[0,1] (1) , u(1) = 0 , ⎪
⎪
⎪
⎪ ⎨ ⎬
⎪
⎪ : κ = −1 ,
⎪
⎪ δ1 = min {1 − #ν, δ} − ε
⎪
⎪ ⎪
⎪  ⎪
⎪
⎪
⎪ ⎩ ⎭
⎪
⎪ ∀ ε ∈ 0, min {1 − #ν, δ}
⎪
⎪ ⎧ ⎫
⎪
⎪
⎪
⎪ ⎪
⎪ u = u1 + f (1)p0 υ ν,−ν , ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎨ u1 ∈ Hγ (0, 1) ∩ C 1 (1) , u1 (1) = 0 , ⎪ ⎬
⎪
⎪
0,δ
⎪
⎪ [0,1]
: κ = 0 , −1 < #ν < 0 ,
⎪
⎪ ⎪ ⎪
⎪
⎪ ⎪
⎪ δ1 = min {−#ν, δ} − ε ⎪
⎪
⎪
⎪ ⎪
⎪ ⎪
⎪
⎪
⎪ ⎩  ⎭
⎨ ∀ ε ∈ 0, min {−#ν, δ}
⎧ ⎫
⎪
⎪ ⎨ u = u1 + f (1)p0 υ ν,−ν , ⎬
⎪
⎪
⎪
⎪   : κ = 0 , #ν = 0 ,
⎪
⎪ ⎩ u1 ∈ Hγ (0, 1) ∩ υ 0,−ν C0,δ (1) ⎭
⎪
⎪ [0,1]
⎪
⎪
⎪
⎪ ⎧ ⎫
⎪
⎪
⎪
⎪ ⎨ u = u1 + f (1)p0 υ ν,−ν , ⎬
⎪
⎪   : κ = 0 , 0 < #ν < 1 ,
⎪
⎪
⎪
⎪ ⎩ u1 ∈ Hγ (0, 1) ∩ υ 0,−ν C0,min{#ν,δ} (1) ⎭
⎪
⎪ [0,1]
⎪
⎪ ⎧ ⎫
⎪
⎪ 4 5 ν,−ν−1
⎪
⎪ ⎨ u = u + p + f (1)p υ , ⎬
⎪
⎪
1 0 1
⎪  
⎩ ⎩ u ∈ Hγ (0, 1) ∩ υ 0,−ν−1 C0,min{#ν+1,δ} (1) ⎭ :
⎪ κ = 1,
1 [0,1]

with certain polynomials pj (x) of degree less or equal than j .

p p
Proof Since S maps Lρ,σ into itself, we have f ∈ Lρ,σ . First, we consider the case
κ = −1. Hence, 0 < η < #ν < 1 − ξ < 1 (cf. (4.3)). Due to Lemma 4.3, the
p p
operator aI + bS : Lρ,σ → Lρ,σ is left-sided invertible and
 −1
u = a 2 − b2 af − bv ν,1−ν Sv −ν,ν−1 f , (5.6)

where v −ν,ν−1 f ∈ L10,0 in view of Remark 5.13. Moreover, from (5.6) and (4.6) we
deliver
  
1 bυ ν,1−ν (x) 1 y −ν [f (y) − f (1)] dy
u(x) = 2 a[f (x) − f (1)] − .
a − b2 πi 0 (1 − y)1−ν y−x

In case δ > 1 − #ν, Lemma 5.12 delivers that the integral in the last equation
belongs to υ 0,α0 C0,λ
[0,1] (1) for 0 < α0 < 1 − #ν and 0 < λ < α0 . In particular,
u(1) = 0. Choosing α0 = 1 − #ν − ε2 and λ = 1 − #ν − ε, we get, taking into
account Remark 5.14, that u ∈ C0,δ [0,1] (1) for δ1 = 1−#ν −ε and all ε ∈ (0, 1−#ν).
1

Analogously, in case δ ≤ 1 − #ν, we obtain u ∈ C0,δ [0,1] (1) for δ1 = δ − ε and all
1

ε ∈ (0, δ). It remains to apply Corollary 5.10 to conclude u ∈ Hγ (0, 1).

388 P. Junghanns and R. Kaiser

Now, let us consider the case κ = 0. Due to Lemma 4.3, the operator aI + bS :
p p
Lρ,σ → Lρ,σ is invertible and
 −1 4 5
u = a 2 − b2 af − bv ν,−ν Sv −ν,ν f ,

where −1 < #ν < 1. From (4.7) we get

  
1 bυ ν,−ν (x) 1 y −ν [f (y) − f (1)] dy
u(x) = a[f (x) − f (1)] −
a 2 − b2 πi 0 (1 − y)−ν y−x

+ γ0 f (1)v ν,−ν (x) .

with a constant γ0 ∈ C. In case −1 < #ν ≤ 0, we get the assertion in the same

manner as in the previous case, i.e., u = u1 + γ0 v ν,−ν f (1) with u1 ∈ Hγ (0, 1) ∩
C0,δ
[0,1] (1) and u1 (1) = 0. In case 0 < #ν < 1, we conclude
1

 
0,min{#ν,δ}
u1 ∈ Hγ (0, 1) ∩ υ 0,−ν C[0,1] (1)

by using Corollary 5.11 and Remark 5.14 together with Corollary 5.10. Note that
ν = 0 due to b = 0. If #ν = 0, then we apply Lemma 5.12 in case α = 0.
p p
In case κ = 1, due to Lemma 4.3, the operator aI + bS : Lρ,σ → Lρ,σ is
invertible from the right and its null space is spanned by υ ν,−ν−1 , such that
 −1
u = a 2 − b2 af − bυ ν,−ν−1 Sυ −ν,ν+1 f + γ0 υ ν,−ν−1

with some constant γ0 ∈ C. Using (4.8) for 1 + ν instead of ν, we deliver

1 
u(x) = a[f (x) − f (1)]
a 2 − b2

bυ ν,−ν−1(x) 1 y −ν (1 − y)ν+1 [f (y) − 1] 
− dy
πi 0 y−x
4 5
+ γ0 + f (1)(δ0 + δ1 x) υ ν,−ν−1 (x) .

Now, we can again make use of Corollary 5.11 and Remark 5.14 together with
Corollary 5.10 to get the assertion also for this case.
For ρ, σ > −1, we define the space
>

L∞
ρ,σ = Lpρ,σ .
1<p<∞
Fredholm Theory of Singular Integral Operators 389

Corollary 5.16 Let ρ, σ > −1, 0 ≤ γ < ∞, and a ∈ C, b ∈ C\{0} with a±b = 0.
If u ∈ 
L∞
ρ,σ and

f = au + bSu ∈ Hγ (0, 1) ∩ C0,δ

[0,1] (1) ,

then

u ∈ Hγ (0, 1) ∩ C0,δ
[0,1] (1) and u(1) = 0 ,
1

where δ1 = min {1 − #ν, δ} − ε and a+b

= e2πiν , 0 ≤ #ν < 1, and where
 a−b
ε ∈ 0, min {1 − #ν, δ} is arbitrary.
p
Proof Since, for sufficiently large p, the operator S maps Lρ,σ into itself, we have
f ∈ L∞ρ,σ . Hence, in Lemma 4.3 we can choose 0 ≤ #ν < 1 and p > 1 sufficiently
large such that (4.3) is satisfied for κ = −1. It remains to apply Proposition 5.15.

Remark 5.17 Let p, ρ, σ , a, b, ν, and κ fulfil the conditions of Proposition 5.15, let
c ∈ Hγ (0, 1], and let k(t) satisfy the conditions of Lemma 5.8 for some γ ∈ (0, 1).
If u ∈ 
p
Lρ,σ and

f = au + bSu + cBk u ∈ Hγ (0, 1) ∩ C0,δ

[0,1] (1)

for some δ ∈ (0, 1), then

0,min{γ ,δ}
au + bSu = f − cBk u ∈ Hγ (0, 1) ∩ C[0,1] ,

and we can apply Proposition 5.15 with min {γ , δ} instead of δ to deduce a regularity
property for u.
Analogously, using Corollary 5.16, we get the following corollary.
Corollary 5.18 Let ρ, σ , a, and b satisfy the conditions of Corollary 5.16, let c ∈
Hγ (0, 1], and let k(t) satisfy the conditions of Lemma 5.8 for some γ ∈ (0, 1). If
u∈ L∞ρ,σ and

f = au + bSu + cBk u ∈ Hγ (0, 1) ∩ C0,δ

[0,1] (1)

for some δ ∈ (0, 1), then

u ∈ Hγ (0, 1) ∩ C0,δ1
[0,1] (1)

and u(1) = 0, where δ1 = min {1 − #ν, γ , δ} and where ν ∈ C is defined as in

Corollary 5.16.
390 P. Junghanns and R. Kaiser

6 One-Sided Invertibility

In this section, we are going to investigate the one-sided invertibility of integral

p
operators of the form aI + bS + Bk in the space Lρ,σ . Under certain conditions
regarding the numbers ρ, σ, p, the coefficients a, b, as well as the kernel function
k we prove that the homogeneous equation (aI + bS + cBk )u = 0 in the space
Lρ,σ or the adjoint equation (aI + bS + cBk )∗ v = 0 in the space L(1−q)ρ,(1−q)σ ,
p q

1
p + q = 1, have only the
1
trivial solution. In the remaining part of this section, let
1+ρ
ξ= p and p + q = 1.
1 1

Proposition 6.1 Let p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1), a ∈ C \ {0}, and k ∈ C(R+ )
satisfy condition (A). Then the homogeneous equations (aI +Bk )u = 0 in the space
Lρ,σ or (aI + Bk )∗ v = 0 in the space L(1−q)ρ,(1−q)σ have only the trivial solution.
p q

p q
Proof Let u ∈ Lρ,σ , v ∈ L(1−q)ρ,(1−q)σ and

(aI + Bk )∗ v
Lemma 4.1
(aI + Bk )u = 0 , = (aI + Bk1 )v = 0 , (6.1)

where k1 (t) = k(t −1 )t −1 . Due to Corollary 3.3, we can choose α0 , β0 ∈ R, such

that α < α0 < ξ < β0 < β and

lim t α0 k(t) = 0 and lim t β0 k(t) = 0 .

t →+0 t →∞

This implies

lim t 1−β0 k1 (t) = 0 and lim t 1−α0 k1 (t) = 0.

t →0 t →∞

1+(1−q)ρ
Since α0 < 1+ρ p is equivalent to q < 1 − α0 , we get, by using Remark 5.2
p q
together with Corollary 5.6 (both in case n = 0), that u ∈ Lρ,0 and v ∈ L(1−q)ρ,0 .
p
By Lemma 3.8, Bk ∈ L Lρ,σ ) is true for all σ ∈ (−1, p−1). Hence we can consider
p q
the equations in (6.1) in the spaces Lρ,0 and L(1−q)ρ,0 , respectively. Because of the
relations (4.1) and (4.2), it only remains to apply Lemma 2.2.
The relations (4.1) and (4.2) together with Lemma 2.2 also immediately deliver
the following proposition.
Proposition 6.2 Let p ∈ (1, ∞), ρ ∈ (−1, p − 1), a ∈ C, b ∈ C \ {0}, and
k ∈ C(R+ ) satisfy condition (A). Then the homogeneous equations

(aI + bS + Bk )u = 0
Fredholm Theory of Singular Integral Operators 391

p
in the space Lρ,0 or

(aI + bS + Bk )∗ v = 0
q
in the space L(1−q)ρ,0 have only the trivial solution.
Proposition 6.3 Let p ∈ (1, ∞), ρ, σ ∈ (−1, p − 1), a ∈ C, b ∈ C \ {0}, and
k ∈ C(R+ ) satisfy condition (A). Moreover, we assume that k(t) fulfils
 
s β k(s) − t β k(t)
Mβ,a,γ1 (k) := sup : s, t ∈ [a, ∞) , s = t <∞
|s − t|γ1

and

|s α k(s) − t α k(t)|
Nα,a,γ2 (k) := sup : s, t ∈ (0, a] , s = t < ∞ (6.2)
|s − t|γ2

for all a > 0 and for some γi ∈ (0, 1), i ∈ {1, 2}, with ξ + γ1 < β and α < ξ − γ2 .
p
Let aI + bS be invertible in Lρ,σ , i.e., there is a ν ∈ C satisfying (4.3) and (4.4)
for κ = 0. Moreover, we assume that

1 1
#ν < and − #ν < . (6.3)
p q

Then the homogeneous equations

(aI + bS + Bk )u = 0
p
in the space Lρ,σ or

(aI + bS + Bk )∗ v = 0
q
in the space L(1−q)ρ,(1−q)σ have only the trivial solution.
p q
Proof Let u ∈ Lρ,σ , v ∈ L(1−q)ρ,(1−q)σ , and

(aI + bS + Bk )u = 0 , (aI + bS + Bk )∗ v
Lemma 4.1
= (aI + bS + Bk1 )v = 0 ,

where k1 (t) = k(t −1 )t −1 . Since the function k(t) fulfils (6.2), we get that k1 (t)
satisfies M1−α,a,γ2 (k1 ) < ∞ for all a > 0. Thus, Lemma 5.8 delivers Bk u ∈
Hγ1 (0, 1] and Bk1 v ∈ Hγ2 (0, 1]. From Proposition 5.15 (in case κ = 0), we derive
p q
u ∈ Lρ,0 and v ∈ L(1−q)ρ,0 , where we took into account assumption (6.3). Thus, it
only remains to apply (4.1) in combination with Lemma 2.2.
392 P. Junghanns and R. Kaiser

Proposition 6.4 Let p ∈ (1, ∞), ρ ∈ (−1, p − 1), a, b ∈ C, and k ∈ C(R+ )

satisfy condition (A). If
 
inf a − bi cot(πξ + it) + 
k(ξ + it) : t ∈ R > 0 ,

then the homogeneous equation

(aI + bS + Bk )u = 0
p
in the space Lρ,σ or the adjoint equation

(aI + bS + Bk )∗ v = 0
q
in the space L(1−q)ρ,(1−q)σ have only the trivial solution.
Proof Relation (4.1) in combination with Lemma 2.3 immediately delivers the
assertion.

References

1. A. Böttcher, I.M. Spitkovsky, Pseudodifferential operators with heavy spectrum. Integr. Equ.
Oper. Theory 19, 251–269 (1994)
2. R. Duduchava, Integral Equations with Fixed Singularities (BSB B. G. Teubner Verlagsge-
sellschaft, Leipzig, 1979)
3. I. Gohberg, N. Krupnik, One-Dimensional Linear Singular Integral Equations. Vol I: Introduc-
tion. Operator Theory: Advances and Applications, vol. 53 (Birkhäuser Verlag, Basel, 1992)
4. P. Junghanns, R. Kaiser, A numerical approach for a special crack problem. Dolomites Res.
Notes Approx. 10, 56–67 (2017)
5. P. Junghanns, R. Kaiser, A note on the Fredholm theory of singular integral operators with
Cauchy and Mellin kernels, in In Operator Theory, Analysis and the State Space Approach. In
Honor of Rien Kaashoek. Operator Theory : Advances and Applications, vol. 271 (Birkhäuser,
Basel, 2018), 291–325
6. S.G. Mikhlin, S. Prössdorf, Singular Integral Operators (Akademie-Verlag, Berlin, 1986)
7. N.I. Muskhelishvili, Singular Integral Equations. Boundary Problems of Function Theory and
Their Application to Mathematical Physics (P. Noordhoff N. V., Groningen, 1953)
8. R. Schneider, Integral equations with piecewise continuous coefficients in Lp -spaces with
weight. J. Integr. Equ. 9, 135–152 (1985)
9. E.C. Titchmarsh, Introduction to the Theory of Fourier Integrals, 3rd edn. (Chelsea Publishing
Co., New York, 1986)
A Note on Group Representations,
Determinantal Hypersurfaces
and Their Quantizations

Igor Klep and Jurij Volčič

Abstract Recently, there have been exciting developments on the interplay

between representation theory of finite groups and determinantal hypersurfaces. For
example, a finite Coxeter group is determined by the determinantal hypersurface
described by its natural generators under the regular representation. This
short note solves three problems about extending this result in the negative.
On the affirmative side, it is shown that a quantization of a determinantal
hypersurface, the so-called free locus, correlates well with representation theory. If
A1 , . . . , A ∈ GLd (C) generate a finite group G, then the family of hypersurfaces
{X ∈ Mn (C)d : det(I + A1 ⊗ X1 + · · · + A ⊗ X ) = 0} for n ∈ N determines G
up to isomorphism.

Keywords Linear pencil · Group representation · Determinantal hypersurface ·

Free locus

Mathematics Subject Classification (2010) Primary 20C15, 15A22; Secondary

47A56, 14J70

The first author was supported by the Slovenian Research Agency grants J1-8132, N1-0057, P1-
0222, and partially supported by the Marsden Fund Council of the Royal Society of New Zealand.

I. Klep
Department of Mathematics, Faculty of Mathematics and Physics, University of Ljubljana,
Ljubljana, Slovenia
e-mail: [email protected]
J. Volčič ()
Department of Mathematics, Texas A&M University, College Station, TX, USA
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 393

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_19
394 I. Klep and J. Volčič

1 Introduction

To A0 , . . . , A ∈ Md (C) one assigns the determinantal hypersurface

{[ξ0 : · · · : ξ ] ∈ CP : det(ξ0 A0 + · · · + ξ A ) = 0}. (1.1)

This is a classical object in algebraic geometry [1, 6, 10, 11], where a key question
asks which hypersurfaces admit determinantal representations. When Aj are real
symmetric matrices, determinantal hypersurfaces pertain to hyperbolic and stable
polynomials [2, 15, 18, 23, 24]. The geometry of the hypersurface (1.1) is also
explored in multivariate operator theory [3, 4, 26]. If Aj are bounded operators
on a Hilbert space and the determinant in (1.1) is replaced with the condition that
ξ0 A0 + · · · + ξ A is not invertible, then (1.1) is known as the projective joint
spectrum of A0 , . . . , A (cf. Taylor spectrum [22] for ensembles of commuting
operators).
Through the work of Frobenius [13] and Dedekind [7] on group determinants
(see also [9]), determinantal hypersurfaces also pertain to representation theory.
Several fascinating developments in this direction [5, 14, 21] have been recently
made. This note addresses certain limitations for extensions of these results.
Let G be a finitely generated group. If T = (g1 , . . . , g ) is a finite sequence of
generators for G and ρ : G → GLd (C) is a representation of G, then denote
 
Z1 (T , ρ) = ξ ∈ C : det (Id + ξ1 ρ(g1 ) + · · · + ξ ρ(g )) = 0 . (1.2)

It is natural to ask what kind of information the affine hypersurface Z1 (T , ρ)

carries about ρ and G. For example, if G1 , G2 are finite groups with left regular
representations λ1 , λ2 , then Z1 (G1 \{1}, λ1 ) = Z1 (G2 \{1}, λ2 ) implies that G1 , G2
are isomorphic [12]. However, one is typically interested in smaller generating sets
or in finitely generated groups which are not necessarily finite. In [14], the authors
computed the joint spectrum for the infinite dihedral group

D∞ = a, t | a 2 = t 2 = 1

with respect to the generating set (1, a, t), and analyzed its properties through
the representation theory of D∞ . Determinantal hypersurfaces also have a strong
connection with representation theory in the case of finite Coxeter groups [5]. A
Coxeter group is a finitely generated group on generators g1 , . . . , g satisfying

(gi gj )mij = 1

where mii = 1 and mij ≥ 2 for i = j . In [5] the authors first showed that if G is a
finite Coxeter group, λ is its left regular representation, and T = (g1 , . . . , g ) are the
generators as above, then Z1 (T , λ) determines G up to isomorphism. Furthermore,
Group Representations, Determinantal Hypersurfaces and Quantizations 395

if G is not of exceptional type (in the Coxeter diagram sense) and ρ is an arbitrary
finite-dimensional representation of G, then Z1 (T , ρ) determines ρ.
These theorems were presented during the Multivariable Spectral Theory and
Representation Theory workshop at the Banff International Research Station in
April 2019. Several problems about extending these results beyond Coxeter groups
were posed by the speakers; among them were the following.
Questions 1.1 Let G be a finite group, T a fixed generating set for G, and ρ1 , ρ2
irreducible complex representations of G.
(1) Is Z1 (T , ρ1 ) a reduced and irreducible hypersurface?
(2) If Z1 (T , ρ1 ) = Z1 (T , ρ2 ), are ρ1 and ρ2 equivalent?
As usual, ρ1 : G → GLd1 (C) and ρ2 : G → GLd2 (C) are equivalent if d1 = d2
and ρ2 = Pρ2 P −1 for some P ∈ GLd1 (C). A representation ρ1 is irreducible if
its image does not admit a nontrivial common invariant subspace; equivalently, it
generates Md1 (C) as a C-algebra by Burnside’s theorem [17, Corollary 1.17]. The
hypersurface Z1 (T , ρ1 ) is reduced and irreducible (in the scheme-theoretic sense)
if its defining determinant in (1.2) is an irreducible polynomial. The main result of
this note is the following.
Theorem 1.2 Questions 1.1(1) and (2) have negative answers in general.
See Sects. 2.1 and 2.2 for concrete examples. On a more positive side, in
Sect. 3 we show that representation theory aligns well with a quantization of
the determinantal hypersurface, the free locus; see Theorem 3.1. Furthermore,
Proposition 3.4 determines whether a free locus arises from a representation of
a finite group, and Proposition 3.7 characterizes finite abelian groups from the
perspective of determinantal hypersurfaces. We conclude this note with an open
question.

2 Representations Versus Determinants

In this section we give negative answers to Questions 1.1. The representations were
found with the help of the computer algebra system GAP and the online repository
ATLAS of Finite Group Representations. Verifying equivalence and irreducibility
of representations was sometimes done symbolically with the computing system
Mathematica.

2.1 Irreducible Representation with Reducible Determinant

The alternating group G = A6 admits a presentation

 
G = g1 , g2 | g12 , g24 , (g1 g2 )5 , (g1 g22 )5 .
396 I. Klep and J. Volčič

Let
⎛ ⎞
1 0 0 0 0 0 0 0 0
⎜ 0 0 1 0 0 0 0 0 0⎟
⎜ ⎟
⎜ 0 1 0 0 0 0 0 0 0⎟
⎜ ⎟
⎜ 0 0 0 0 1 0 0 0 0⎟
⎜ ⎟
⎜ ⎟
A1 = ⎜ 0 0 0 1 0 0 0 0 0 ⎟
⎜ ⎟
⎜ 0 0 0 0 0 0 0 1 0⎟
⎜ ⎟
⎜ 0 0 0 0 0 0 1 0 0⎟
⎜ ⎟
⎝ 0 0 0 0 0 1 0 0 0⎠
−1 −1 −1 −1 −1 −1 −1 −1 −1

and
⎛ ⎞
010 00 000 0
⎜0 0 0 10 000 0⎟
⎜ ⎟
⎜0 0 0 0⎟
⎜ 01 000 ⎟
⎜0 0 0 0⎟
⎜ 00 100 ⎟
⎜ ⎟
A2 = ⎜ 0 0 0 00 010 0⎟.
⎜ ⎟
⎜1 0 0 00 000 0⎟
⎜ ⎟
⎜0 0 0 00 000 1⎟
⎜ ⎟
⎝0 0 0 00 001 0⎠
001 00 000 0

Then ρ(g1 ) = A1 and ρ(g2 ) = A2 determines a faithful and irreducible

representation ρ : G → GL9 (C). Indeed, we can directly check that

A21 = A42 = (A1 A2 )5 = (A1 A22 )5 = I,

so ρ is a representation of G, and is moreover faithful since it is nontrivial and

G is simple. Furthermore, all the possible products of A1 and A2 with at most 8
factors span the whole M9 (C), so ρ is irreducible. However, we claim that the curve
Z1 ((g1 , g2 ), ρ) in C2 is not irreducible. We can compute the determinant of I +
x1 ρ(g1 ) + x2 ρ(g2 ),
⎛ ⎞
1 + x1 x2 0 0 0 0 0 0 0
⎜ 0 1 x1 x2 0 0 0 0 0⎟
⎜ ⎟
⎜ 0⎟
⎜ 0 x1 1 0 x2 0 0 0 ⎟
⎜ 0⎟
⎜ 0 0 0 1 x1 x2 0 0 ⎟
⎜ ⎟
det ⎜ 0 0 0 x1 1 0 x2 0 0⎟
⎜ ⎟
⎜ x2 0 0 0 0 1 0 x1 0⎟
⎜ ⎟
⎜ 0 0 0 0 0 0 1 + x1 0 x2 ⎟
⎜ ⎟
⎝ 0 0 0 0 0 x1 0 1 + x2 0⎠
−x1 −x1 x2 − x1 −x1 −x1 −x1 −x1 −x1 1 − x1
Group Representations, Determinantal Hypersurfaces and Quantizations 397

by cofactor expansion along the rows. The reader will have no difficulty verifying
that det(I + x1 ρ(g1 ) + x2 ρ(g2 )) equals

1 + x1 − 4x12 − 4x13 + 6x14 + 6x15 − 4x16 − 4x17 + x18 + x19 + x2 + 2x1x2 − 2x12 x2
− 6x13x2 + 6x15 x2 + 2x16 x2 − 2x17 x2 − x18 x2 + x12 x22 + x13 x22 − 2x14 x22 − 2x15x22
+ x16 x22 + x17 x22 − x12 x23 + 2x14x23 − x16 x23 − 2x24 + x12 x24 − x13 x24 + x14 x24 + x15 x24
− 2x25 − 2x1 x25 − x12 x25 + x14 x25 − x12 x26 − x13 x26 + x12 x27 + x28 − x1 x28 + x29

which is the product of the following two irreducible polynomials:

1 + 2x1 − 2x13 − x14 + x1 x2 + 2x12 x2 + x13 x2 − x1 x22 − x12 x22 + x1 x23 − x24 ,
1 − x1 − 2x12 + 2x13 + x14 − x15 + x2 − x1 x2 − x12 x2 + x13 x2 − x24 − x25 .

Some of the subsequent examples are presented in a more terse way to maintain the
focus on their intent.
Note that the above irreducible representation of A6 has dimension 9, which
is not the minimum among nontrivial complex representations of A6 ; namely,
A6 admits a representation σ of minimal dimension 5, and Z1 ((g1 , g2 ), σ ) is
irreducible. One might thus be tempted to suggest that for a group G generated by
a finite set T and its (irreducible) representation σ of minimal dimension, Z1 (T , σ )
is irreducible. However, even this weaker conjecture fails. The counterexample is
given by the Janko group J2 ,
 
J2 = g1 , g2 | g12 , g23 , (g1 g2 )7 , (g1 g2 g1−1 g2−1 )12 , (g1 g2 (g1 g2 g1 g2−1 )2 )6 .

This sporadic simple group of order 604800 admits two non-isomorphic complex
representations σ1 , σ2 of minimal dimension 14, courtesy of ATLAS of Finite Group
Representations. As in the previous example (albeit with slightly longer calcula-
tions), one can explicitly check that the curve Z1 ((g1 , g2 ), σ1 ) = Z1 ((g1 , g2 ), σ2 )
has two irreducible components.

2.2 Non-equivalent Representations with the Same

Determinant

The classical group G = GL2 (Z/3Z) admits the presentation

 
G = g1 , g2 , g3 | g12 , (g1 g2−1 )2 , (g1 g3−1 )2 , g22 g3 g2−1 g3 , g2 g32 g2 g3−1 .
398 I. Klep and J. Volčič

Let A1 , A2 , A3 be the matrices

 1 1
− √1 − 12 − i
2 2 + i
2
√1
2 − i
2
√i
2 , 2 , 2 .
− 12 + i
2
√1 − √1 1
− i √i 1
+ i
2 2 2 2 2 2 2

There are faithful irreducible unitary representations ρ+ , ρ− : G → GL2 (C) given

by

ρ± (g1 ) = ±A1 , ρ± (g2 ) = A2 , ρ± (g3 ) = A3 .

It is easy to check that ρ+ and ρ− are not equivalent. On the other hand,

Z1 ((g1 , g2 , g3 ), ρ± ) = {(ξ1 , ξ2 , ξ3 ) : 1 − ξ12 + ξ2 + ξ22 + ξ3 + ξ32 = 0}.

3 Free Locus Perspective

In this section we will see how representations of a finitely generated group are
determined by a noncommutative relaxation of (1.2). To A ∈ Md (C) we associate
the monic matrix pencil LA = I + A1 x1 + · · · + A x of size d in freely
noncommuting variables x = (x1 , . . . , x ). Thus L is an affine matrix over the
free algebra C<x>. At a matrix point X ∈ Mn (C) it evaluates as

LA (X) = Idn + A1 ⊗ X1 + · · · + A ⊗ X ∈ Mdn (C).

The free locus [19] of LA is the disjoint union of determinantal hypersurfaces

I  
Z (LA ) = Zn (LA ), Zn (LA ) = X ∈ Mn (C) : det LA (X) = 0 .
n∈N

Given a group G generated by T = (g1 , . . . , gn ) and a complex representation

ρ : G → GLd (C), we write

Z (T , ρ) = Z Lρ(g1 ),...,ρ(g ) . (3.1)

By the definition of the free locus we see that (3.1) is indeed a quantization of (1.2).
The existing results on free loci [16, 19] readily apply to group representations.
Theorem 3.1 For i = 1, 2 let Gi be a group generated by a finite sequence Ti and
let ρi be a complex representation of Gi . Assume |T1 | = |T2 |.
(1) If ρi is irreducible, then there exists n0 ∈ N such that Zn (T1 , ρ1 ) is a reduced
and irreducible hypersurface for all n ≥ n0 .
Group Representations, Determinantal Hypersurfaces and Quantizations 399

(2) If ρ1 and ρ2 are irreducible, then Z (T1 , ρ1 ) = Z (T2 , ρ2 ) if and only if

G1 / ker ρ1 ∼= G2 / ker ρ2 and ρ1 , ρ2 are equivalent.
(3) For i = 1, 2 assume that Gi is finite and ρi is a faithful representation. Then
Z (T1 , ρ1 ) = Z (T2 , ρ2 ) if and only if G1 ∼
= G2 via an isomorphism mapping
T1 to T2 .
Proof
(1) A consequence of [16, Theorem 3.4].
(2) A consequence of [19, Theorem 3.11].
(3) Let Ti be the C-algebra generated by Ti . Since Gi is finite, its group algebra
CGi is semisimple by Maschke’s theorem [17, Theorem 1.9]. Since Ti is a
quotient of CGi , it is also semisimple. Then Z (T1 , ρ1 ) = Z (T2 , ρ2 ) if and
only if T1 → T2 induces an algebra isomorphism T1 → T2 by [19, Corollary
3.8]. This isomorphism then restricts to the group isomorphism G1 → G2 .
Remark 3.2 There is a deterministic bound on n0 in Theorem 3.1(1) that is
exponential in |T1 | and the dimension of ρ1 by [16, Remark 3.5] (the bound is likely
not optimal). Similarly, to verify Z (T1 , ρ1 ) = Z (T2 , ρ2 ) of Theorem 3.1(2,3), it
suffices to check Zn (T1 , ρ1 ) = Zn (T2 , ρ2 ) for a fixed large enough n, exponential
in |Ti | and the dimension of ρi by [19, Remark 3.7].
Free loci are defined for monic pencils with arbitrary matrix coefficients; we now
describe how the geometry of the free locus Z (LA ) detects whether the coefficients
A1 , . . . , A generate a finite group. See also [8] for an efficient algorithm that
determines finiteness of a finitely generated linear group.
Definition 3.3 Let , n ∈ N. Let C ∈ GLn (Z) be the permutation matrix
corresponding to the cycle (1 2 · · · n). If {1, . . . , n} = S1 · · · S and Pj is
the orthogonal projection onto span{ek : k ∈ Sj }, then the matrix point

X = (P1 C, · · · , P C) ∈ Mn (Z)

is called a cycle partition. For given , n we thus have n cycle partitions.

Let μ∞ ⊂ C \ {0} be the group of all roots of unity. The next proposition shows
that if A1 , . . . , A generate a finite group, then Z (LA ) intersects complex lines
through cycle partitions only in points from μ∞ .
Proposition 3.4 Let A ∈ Md (C) . Then A1 , . . . , A generate a finite group if and
only if the following hold:
(i) there is a positive definite P ∈ Md (C) such that A∗j P Aj = P for all j ;
(ii) for every cycle partition X and t ∈ C,

tX ∈ Z (LA ) 7⇒ t ∈ μ∞ .

Proof (⇐) Every Aj is invertible by (i). Let G be a group generated by A1 , . . . , A .

Also by (i), G is a subgroup of the unitary group in GLd (C) with respect to the
400 I. Klep and J. Volčič

inner product u, v = u∗ P v. Hence every element of G is diagonalizable. By [25,

Corollary 4.9], a finitely generated subgroup of GLd (C) is finite if and only if it
is periodic (or torsion; i.e., every element has finite order). Since a diagonalizable
matrix has a finite order if and only if all its eigenvalues lie in μ∞ , it suffices to
verify that eigenvalues of every element of G lie in μ∞ .
To (i1 , . . . , in ) ∈ {1, . . . , }n we associate the cycle partition X ∈ Mn (Z) by
choosing Sj = {ek : ik = j }. We claim that tX ∈ Z (LA ) if and only if (−t)n is an
eigenvalue of Ai1 · · · Ain . Indeed, using Schur complements it is easy to check that
 
I tAi1
det(I − (−1) t Ai1 · · · Ain ) = det
n n
(−1)n t n−1 Ai2 · · · Ain I
⎛ ⎞
I tAi1 0
= det ⎝ 0 I tAi2 ⎠
−(−1) tn n−2 Ai3 · · · Ain 0 I
= ···
⎛ ⎞
I
tAi1
⎜ .. .. ⎟
⎜ . . ⎟
= det ⎜ ⎟
⎝ I tAin−1 ⎠
tAin I
= det LA (tX).

Thus the matrix Ai1 · · · Ain has finite order if and only if tX ∈ Z (LA ) implies
t ∈ μ∞ , which holds by (ii).
(⇒) If A1 , . . . , A generate a finite group G, then Cd admits a G-invariant inner
product

u, v = (gu)∗ (gv).
g∈G

If P is the positive definite matrix satisfying u, v = u∗ P v, then (i) holds.

Furthermore, the proof of (ii) is already given in the previous paragraph.
Remark 3.5 If additional information about A1 , . . . , A is given, say that their
entries generate a number field (finite extension of Q), then the size of the cycle
partitions, which have to be tested in Proposition 3.4, can be bounded using Schur’s
theorem on orders of finite matrix groups [17, Theorem 14.19].
Remark 3.6 Let p ∈ N be prime. If μ∞ in Proposition 3.4 is replaced by the group
of power-of-p roots of unity, one obtains a free locus characterization of matrix
tuples that generate a finite p-group.
Group Representations, Determinantal Hypersurfaces and Quantizations 401

We also show how the free locus certifies whether its defining coefficients
generate a finite abelian group. The degree of an affine variety of codimension m
is the number of intersection points of the variety with m hyperplanes in general
position; in the case of a hypersurface, it is simply the degree of its square-free
defining polynomial.
Proposition 3.7 Let G be a finite group generated by A1 , . . . , A ∈ Md (C). Then
G is abelian if and only if the irreducible components of Zn (LA ) have degree n for
all n ∈ N.
Proof Let A be the C-algebra generated by A1 , . . . , A . As in the proof of
Theorem 3.1(3) we see that A is semisimple. After a basis change (which does
not affect the structure of G or Z (LA )) we can thus assume that

Aj = A(1) (s)
j ⊕ · · · ⊕ Aj

(k) (k)
where A1 , . . . , A ∈ Mdk (C) determine an irreducible representation of G for
every k = 1, . . . , s. For X ∈ Mn (C)d let us view det LA(k) (X) as a polynomial
in the entries of X. If dk = 1, then det LA(k) (X) is up to an affine change of
coordinates equal to the determinant of a generic n × n matrix, and hence an
irreducible polynomial of degree n. On the other hand, if dk > 1, then det LA(k) (X)
is a polynomial of degree dk n > n for all n, and irreducible for all large enough n
by [16, Theorem 3.4]. Since G is abelian if and only if d1 = · · · = ds = 1, and

Zn (LA ) = Zn (L(1) (s)

A ) ∪ · · · ∪ Zn (LA ),

it follows that G is abelian if and only if the irreducible components of Zn (LA ) are
hypersurfaces of degree n.
Remark 3.8 If  = 2 and A1 , A2 are hermitian, then Z1 (LA ) alone determines
whether G is abelian, cf. [20].
The last two propositions offer some directions for future research. Theorem 3.1
implies that the linear group G generated by a tuple A is determined by Z (LA ).
It would be interesting to know which properties of G can be deduced from the
geometry of Z (LA ). For example, intersections of Z (LA ) with certain lines and
hyperplanes determine whether G is finite or abelian. An open problem is how to
decide whether a finite group G is nilpotent/solvable/simple (or any other group-
theoretic property) by considering the geometry of the hypersurfaces Zn (LA ).

Acknowledgments The authors thank Banff International Research Station for the hospitality
during the Multivariable Spectral Theory and Representation Theory workshop, and participants
for sharing their ideas.
402 I. Klep and J. Volčič

References

1. A. Beauville, Determinantal hypersurfaces. Michigan Math. J. 48, 39–64 (2000)

2. P. Brändén, Obstructions to determinantal representability. Adv. Math. 226, 1202–1212 (2011)
3. P. Cade, R. Yang, Projective spectrum and cyclic cohomology. J. Funct. Anal. 265, 1916–1933
(2013)
4. I. Chagouel, M.I. Stessin, K. Zhu, Geometric spectral theory for compact operators. Trans. Am.
Math. Soc. 368, 1559–1582 (2016)
5. Ž. Čučković, M.I. Stessin, A.B. Tchernev, Determinantal hypersurfaces and representations of
Coxeter groups. Preprint. arXiv:1810.12893
6. C. de Concini, D. Eisenbud, C. Procesi, Young diagrams and determinantal varieties. Invent.
Math. 56, 129–165 (1980)
7. R. Dedekind, Gesammelte mathematische Werke. Band II (Chelsea Publishing Co., New York,
1968)
8. A.S. Detinko, D.L. Flannery, E.A. O’Brien, Recognizing finite matrix groups over infinite
fields. J. Symb. Comput. 50, 100–109 (2013)
9. L.E. Dickson, An elementary exposition of Frobenius’s theory of group-characters and group-
determinants. Ann. Math. 4, 25–49 (1902)
10. L.E. Dickson, Determination of all general homogeneous polynomials expressible as determi-
nants with linear elements. Trans. Am. Math. Soc. 22, 167–179 (1921)
11. I. Dolgachev, Classical Algebraic Geometry: a Modern View (Cambridge University Press,
Cambridge, 2012)
12. E. Formanek, D. Sibley, The group determinant determines the group. Proc. Am. Math. Soc.
112, 649–656 (1991)
13. F.G. Frobenius, Gesammelte Abhandlungen. Bände I–III (Springer, Berlin/New York, 1968)
14. R.I. Grigorchuk, R. Yang, Joint spectrum and the infinite dihedral group. Proc. Steklov Inst.
Math. 297, 145–178 (2017)
15. J.W. Helton, V. Vinnikov, Linear matrix inequality representation of sets. Commun. Pure Appl.
Math. 60, 654–674 (2007)
16. J.W. Helton, I. Klep, J. Volčič, Geometry of free loci and factorization of noncommutative
polynomials. Adv. Math. 331, 589–626 (2018)
17. I.M. Isaacs, Character Theory of Finite Groups. Pure and Applied Mathematics, vol. 69
(Academic, New York/London, 1976)
18. D. Kerner, V. Vinnikov, Determinantal representations of singular hypersurfaces in Pn . Adv.
Math. 231, 1619–1654 (2012)
19. I. Klep, J. Volčič, Free loci of matrix pencils and domains of noncommutative rational
functions. Comment. Math. Helv. 92, 105–130 (2017)
20. T.S. Motzkin, O. Taussky, Pairs of matrices with property L. Trans. Am. Math. Soc. 73, 108–
114 (1952)
21. M.I. Stessin, A.B. Tchernev, Spectral algebraic curves and decomposable operator tuples. J.
Oper. Theory 82, 75–113 (2019)
22. J.L. Taylor, A joint spectrum for several commuting operators. J. Funct. Anal. 6, 172–191
(1970)
23. J. Volčič, Stable noncommutative polynomials and their determinantal representations. SIAM
J. Appl. Algebra Geom. 3, 152–171 (2019)
24. D.G. Wagner, Multivariate stable polynomials: theory and applications. Bull. Am. Math. Soc.
(N.S.) 48, 53–84 (2011)
25. B.A.F. Wehrfritz, Infinite Linear Groups. Ergebnisse der Matematik und ihrer Grenzgebiete,
vol. 76 (Springer, New York/Heidelberg, 1973)
26. R. Yang, Projective spectrum in Banach algebras. J. Topol. Anal. 1, 289–306 (2009)
Algebras Generated by Toeplitz
Operators on the Unit Sphere II: Non
Commutative Case

Maribel Loaiza and Nikolai Vasilevski

Abstract In Loaiza and Vasilevski (Commutative algebras generated by Toeplitz

operators on the unit sphere. Intgr. Equ. Oper. Theory, v. 92, 25 (2020)), we
represented the Hardy space H 2 (S 2n−1 ), with n ≥ 2, as a direct sum of weighted
Bergman spaces A2p (Bn−1 ), with p ∈ Z+ , This permitted us to represent
Toeplitz operators, whose symbols are invariant under certain T-action, acting on
H 2 (S 2n−1 ), as direct sums of Toeplitz operators, acting on corresponding Bergman
spaces. As a benefit we were able to use already known results on Toeplitz operators
on Bergman spaces, and a wide variety of commutative algebras, generated by
Toeplitz operators on H 2 (S 2n−1 ), was described.
Following the same approach, in the present paper we pass to the non-
commutative case, presenting the detailed description of two non-commutative
C ∗ -algebras generated by Toeplitz operators on the Hardy space H 2 (S 2n−1 ).

1 Introduction

The paper continues the study of algebras generated by Toeplitz operators, acting on
the multidimensional Hardy space H 2 (S 2n−1 ), which was started in [12]. Recall in
this connection the principal difference in algebraic properties of Toeplitz operators
acting on the one-dimensional Hardy H 2 (S 1 ) and Bergman A2 (D) spaces.
The classical result by Brown and Halmos [9] implies that there is no nontrivial
commutative C ∗ -algebra generated by Toeplitz operators acting on the Hardy space
H 2 (S 1 ), while there are only two commutative Banach algebras. One of them
is generated by Toeplitz operators with analytic symbols, and the other one is
generated by Toeplitz operators with conjugate analytic symbols. Of course, such
two algebras remain to be commutative for H 2 (S 2n−1 ), for all n > 1.

M. Loaiza · N. Vasilevski ()

Departamento de Matemáticas, CINVESTAV, México, Mexico
e-mail: [email protected]; [email protected]

© Springer Nature Switzerland AG 2021 403

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_20
404 M. Loaiza and N. Vasilevski

At the same time, at the turn of this century, it was observed [15, 16] (see also
[17]) that there are many nontrivial commutative C ∗ -algebras, generated by Toeplitz
operators acting on the Bergman space A2 (D) over the unit disk D ⊂ C. These
results were extended then to the case of Toeplitz operators acting on weighted
Bergman spaces on the unit ball Bn , see [13]. Further on many nontrivial Banach
algebras generated by Toeplitz operators, that are commutative on each standard
weighted Bergman space over Bn , were discovered and studied.
In this connection a rather challenging question appeared: what is the situation
with both C ∗ and Banach algebras generated by Toeplitz operators on the multidi-
mensional Hardy space H 2 (S 2n−1 ). A first step in this direction was made in [1]
(see as well [2]), where Z. Akkar, following all the reasonings of [13], described the
commutative C ∗ -algebras generated by Toeplitz operators on H 2 (S 2n−1 ).
In [12], we developed an alternative approach to the problem. Therein we
represented the Hardy space H 2 (S 2n−1 ) as a direct sum of weighted Bergman
spaces A2p (Bn−1 ), with p ∈ Z+ , which permitted us to represent Toeplitz operators,
acting on H 2 (S 2n−1 ), as direct sums of Toeplitz operators, acting on corresponding
Bergman spaces. The benefit of such an approach is that we can use now all the
results on Toeplitz operators on Bergman spaces in their full power. In particular, we
showed in [12] how to recover the results of [1] just as simple and straightforward
corollaries of [13], and explained how to unhide and describe a wide variety
of nontrivial commutative Banach algebras generated by Toeplitz operators on
H 2 (S 2n−1 ).
Following the same approach, in the present paper we pass to the description
of non-commutative algebras, and present the detailed description of two non-
commutative C ∗ -algebras generated by Toeplitz operators, acting on the Hardy
space H 2 (S 2n−1 ), based on the already obtained results [4, 6–8] for Toeplitz
operators, acting on the Bergman space.

2 Preliminaries: Bergman and Hardy Spaces, Toeplitz

Operators

We recall here the results on the representation of the multidimensional Hardy space
in terms of the Bergman spaces as well as the corresponding representation of
Toeplitz operators, acting on the Hardy space, in terms of Toeplitz operators, acting
on Bergman spaces. All proofs and details can be found in [12].
Denote by

Bn = {z = (z1 , z2 , . . . , zn ) ∈ Cn : |z|2 := |z1 |2 + · · · + |zn |2 < 1}

the unit ball in Cn and by

S 2n−1 = {z = (z1 , z2 , . . . , zn ) ∈ Cn : |z1 |2 + · · · + |zn |2 = 1}

Toeplitz Operators on the Unit Sphere 405

the unit sphere, being the boundary of Bn . The following standard notations will
be used throughout the paper. For α = (α1 , . . . , αn ) ∈ Zn+ , where Z+ =
{0, 1, 2, . . .} ⊂ Z,

zα = z1α1 . . . znαn ,
α! = α1 ! . . . αn !,
|α| = α1 + · · · + αn .

The standard Lebesgue measure in Cn is denoted by dV (z), i.e.

dV (z) = dx1 dy1 . . . dxn dyn ,

where z = (z1 , z2 , . . . , zn ) and zk = xk + iyk , k = 1, . . . , n. For each λ ∈ (−1, ∞)

introduce the standard normalized weighted measure

(n + λ + 1)  λ
dvλ (z) = 1 − |z| 2
dV (z).
π n (λ + 1)

The weighted Bergman space A2λ (Bn ) is the closed subspace of L2 (Bn , dvλ ) that
consists of all analytic functions. The Hardy space H 2 (S 2n−1 ) is defined as the set
of all holomorphic functions f , defined in Bn , such that

f 22 := sup |f (rζ )|2 dσ (ζ ) < ∞,
0<r<1 S 2n−1

where dσ denotes the normalized measure for S 2n−1 . This space can be defined,
alternatively, as the closed subspace of L2 (S 2n−1 , dσ ) that consists of all functions
f satisfying the tangential Cauchy-Riemann equations:
 
∂ ∂
Lk,j f = zk − zj f = 0, 1 ≤ k < j ≤ n. (2.1)
∂zj ∂zk

Observe that f ∈ L2 (S 2n−1 , dσ ) if and only if

J
f(z , tn ) := f (z , 1 − |z |2 tn ) ∈ L2 (Bn−1 , dv(z )) ⊗ L2 (S 1 , dμ(tn )),

where
J
(n − 1)!
z = (z1 , . . . , zn−1 ), zn = 1 − |z |2 tn , dv(z ) = dV (z ),
π n−1

dV (z ) = dx1 dy1 · · · dxn−1 dyn−1 ,

406 M. Loaiza and N. Vasilevski

1 dtn
and dμ(tn ) = 2π 1
itn is the normalized arc-length measure on S .
Equations (2.1) are not independent. The equations that are independent and
equivalent to (2.1) have the form
! "
∂ 1 zk ∂ 
Dk (f) = + tn f = 0, k = 1, 2, . . . , n − 1.
∂zk 2 1 − |z |2 ∂tn

Recall that the discrete Fourier transform F : L2 (S 1 , dμ(tn )) → 2 (Z), has the
form

−p
F : g → cn = g(tn )tn dμ(tn ) .
S1 p∈Z

Introduce the unitary operator U1 = I ⊗ F , being an isometric isomorphism of

L2 (S 2n−1 , dσ ) ∼
= L2 (Bn−1 , dv(z )) ⊗ L2 (S 1 , dμ(tn ))

onto

L2 (Bn−1 , dv(z )) ⊗ 2 (Z) = 2 (Z, L2 (Bn−1 , dv(z )))

and acting as follows

 
U1 : f(z , tn ) −→  cp (z ) p∈Z ,

where

−p

cp (z ) = f(z , tn )tn dμ(tn ).
S1

Observe that
K
2 (Z, L2 (Bn−1 , dv(z ))) = L2 (Bn−1 , dv(z )).
p∈Z

For each p ∈ Z introduce then the unitary operator

up : L2 (Bn−1 , dv(z )) → L2 (Bn−1 , dvp (z )),

(n+|p|−1)!
where dvp (z ) = π n−1 |p|!
(1 − |z |2 )p dV (z ), which acts as follows
*
(n − 1)!|p|! p
up : 
cp (z ) −→ cp (z ) = cp (z )(1 − |z |2 )− 2 .

(n + |p| − 1)!
Toeplitz Operators on the Unit Sphere 407

Finally, introduce the unitary operator

K K K
U2 = up : L2 (Bn−1 , dv(z )) −→ L2 (Bn−1 , dvp (z )).
p∈Z p∈Z p∈Z

Theorem 2.1 ([12, Theorem 2.1]) The unitary operator U = U2 U1 is an isometric

isomorphism of

L2 (S 2n−1 , dσ ) ∼
= L2 (Bn−1 , dv(z )) ⊗ L2 (S 1 , dμ(tn ))

onto
K
L2 (Bn−1 , dvp (z ))
p∈Z

7
under which H 2 (S 2n−1 ) is mapped onto p∈Z+ A2p (Bn−1 ).
For each p ∈ Z+ , we denote by Bp the Bergman orthogonal projection from
L2 (Bn−1 , dvp (z )) onto the weighted Bergman space A2p (Bn−1 ), and let PS 2n−1 be
the Szegő projection from L2 (S 2n−1 , dσ ) onto the Hardy space H 2 (S 2n−1 ).
Corollary 2.2 ([12, Corollary 2.2]) The Szegő projection and the weighted
Bergman projections are connected as follows
K
U PS 2n−1 U −1 = Bp .
p∈Z+

Let z = (z , zn ) be a point  in the unit sphere S

2n−1 in Cn , where z =

(z1 , . . . , zn−1 ) and write zn = 1 − |z |2 tn , where tn ∈ S 1 . Consider a function

a(z , zn ) defined in the unit sphere S 2n−1 . The Toeplitz operator Ta , acting on the
Hardy space H 2 (S 2n−1 ), is defined as follows

Ta f = PS 2n−1 (af ).

The crucial role of the Fourier transform F in establishing of Theorem 2.1

suggests us to consider the symbols a of the form

a = a(z1 , . . . , zn−1 , |zn |) = a(z , |zn |), (2.2)

i.e., the symbols that do not depend on tn , commuting thus with U1 = I ⊗ F .

Given the above a ∈ L∞ (S 2n−1 ), introduce the associated function

a = a(z1 , . . . , zn−1 ) = a(z )

J J
= a(z1 , . . . , zn−1 , 1 − |z1 |2 − . . . − |zn−1 |2 ) = a(z , 1 − |z |2 ), (2.3)
408 M. Loaiza and N. Vasilevski

where (z1 , . . . , zn−1 ) = z ∈ Bn−1 . Note that each function a ∈ L∞ (Bn−1 ) defines
in its turn the function a of the form (2.2) by a(z , |zn |) = a(z ).
p
We denote then by Ta the Toeplitz operator, with symbol a acting on the
weighted Bergman space A2p (Bn−1 ), with p ∈ Z+ .
Theorem 2.3 ([12, Theorem 3.1]) Given a bounded measurable symbol
a(z , |zn |), defined in S 2n−1 . Under the isomorphism U = U2 U1 , the Toeplitz
operator 7 Ta , acting on the Hardy space H 2 (S 2n−1 ), is unitarily equivalent to the
p
operator p∈Z+ Ta , acting on
K
A2p (Bn−1 ),
p∈Z+

where a = a(z ) is of the form (2.3).

In what follows we will consider two classes (algebras) S of functions a of the
form (2.2) as well as the corresponding classes (algebras) S of functions a defined
by (2.3) for a ∈ S. Of course, symmetrically, each class S of functions a = a(z ),
z ∈ Bn−1 defines the class S of functions a of the form (2.2), connected with a ∈ S
by (2.3).
Given a class S, we denote by T (S) the closed unital algebra generated by all
Toeplitz operators T a , with a ∈ S, acting on the Hardy space H 2 (S 2n−1 ). For
a corresponding class S, we denote by Tp (S) the closed unital algebra generated
p
by all Toeplitz operators Ta , with a ∈ S, acting on the weighted Bergman space
Ap (B ), with p ∈ Z+ .
2 n−1

Theorems 2.1 and 2.3 state then that for each generator T a of the algebra
T (S), the operator U T a U −1 leaves invariant each subspace in the direct sum
decomposition
K
U (H 2 (S 2n−1 )) = A2p (Bn−1 )
p∈Z+

and
K
U T a U −1 =
p
Ta . (2.4)
p∈Z+

Moreover, for each operator T ∈ T (S) there is a unique sequence of operators

T p ∈ Tp (S), p ∈ Z+ , such that
K
U T U −1 = T p.
p∈Z+
Toeplitz Operators on the Unit Sphere 409

In what follows we will abbreviate the above equation as

K
T $ T p. (2.5)
p∈Z+

3 Compact Semi-commutator C ∗ -Algebra

We consider the case when S = VO∂ (Bn−1 ) with the corresponding class S VO ⊂
L∞ (S 2n−1 ). Recall [8], that a bounded continuous function a belongs to VO∂ (Bn−1 )
(has a vanishing oscillation at the boundary) if
4 5
lim sup {|a(z ) − a(w )| : β(z , w ) ≤ 1} = 0,
z →∂Bn−1

here β(·, ·) is the Bergman metric in Bn−1 , and that VO∂ (Bn−1 ) is a norm closed
C ∗ -subalgebra of L∞ (Bn−1 ).
Our choice of S = VO∂ (Bn−1 ) is motivated by the following observation. C.
Berger, L. Coburn, and K. Zhu described in [8] the largest C ∗ -subalgebra Q of
L∞ (Bn−1 ) which possesses the compact semi-commutator property, i.e.,
p p p p p
[Ta , Tb ) := Ta Tb − Tab is compact for every a, b ∈ Q.

n−1 n−1
Then, C(B ) ⊂ VO∂ (Bn−1 ) ⊂ Q and the algebra Tp (C(B )) contains all
compact operators of A2p (Bn−1 ), which, in particular, implies that the algebra
Tp (VO∂ (Bn−1 )) coincides with Tp (Q). At the same time it is more pleasant to
deal with (bounded uniformly continuous) symbols from VO∂ (Bn−1 ) than with
(generically discontinuous) symbols from Q.
Note, in this connection, that the C ∗ -algebra Tp (VO∂ (Bn−1 )) = Tp (Q) is
irreducible, contains the ideal Kp of all compact operators, and each its element
p
admits the representation T p = Ta + K p , with a ∈ VO∂ (Bn−1 ) and compact K p .
We recall as well that the Berezin transform Bp : L(A2p (Bn−1 )) → L∞ (Bn−1 ) is
 p p p
given by Bp [A](z ) := Akz , kz p , where kz denotes the normalized reproducing
kernel function of A2p (Bn−1 ) defined by

n+p
p (1 − |z |2 ) 2
kz (w )= , z , w ∈ Bn−1 .
(1 − w · z )n+p
p
For the special case of a Toeplitz operator Ta ,
p  p p
Bp [Ta ](z ) = Bp [a](z ) = a kz , kz p .
410 M. Loaiza and N. Vasilevski

The aim of this section is to study the algebra T (S VO ) generated by all Toeplitz
operators T a acting on H 2 (S 2n−1 ) and having symbols a ∈ S VO , associated to
S = VO∂ (Bn−1 ).
7
Lemma 3.1 For each T $ p∈Z+ T p we have

T  = sup T p ,
p∈Z+

while for the generating operators we have

T a  = a(z )∞ .

Proof The first equality easily follows by standard arguments from the direct sum
p
representation of the operator T . The second follows from Ta  ≤ a(z )∞ and
[4, Proposition 4.4], implying
p p
a(z )∞ =  lim Bp Ta ∞ ≤ sup Ta  = T a .
p→∞ p∈Z

Theorem 3.2 Each operator T ∈ T (S VO ), in the direct sum decomposition (2.5),

admits the unique representation
K p
T $ (Ta + K p ), (3.1)
p∈Z+

where a ∈ VO∂ (Bn−1 ), each K p is compact, and K p  → 0 as p → ∞.

Proof Follows from the proof of [7, Theorem 4.3] under λ = 0 and  = m = 1.
For the sake of completeness we sketch it here. Let D be the dense subalgebra of
T (S VO ) consisting of all finite sums of finite products of the generators of T (S VO ).
By induction,
7 we only need to prove the assertion for a product of just two operators
p p
Ta Tb $ p∈Z+ Ta Tb . Since VO∂ (Bn−1 ) possesses the semicommutator property,
we have that, for each p ∈ Z+ , the semicommutator
p p p
Ta Tb − Tab =: K p

is compact. Then [6, Theorem 3.8] implies that K p  → 0 when p → ∞. Thus

K p
Ta T b $ Tab + K p ,
p∈Z+

with K p  → 0 when p → ∞.
Toeplitz Operators on the Unit Sphere 411

Given now T ∈ T (S VO ), there exists a sequence {Tm }m∈N of elements in D that

converges in norm to T , where each Tm has the form
K p p
Tm $ Tam + Km ,
p∈Z+

p
with Km  → 0 when p → ∞. Using Lemma 3.1, the convergence of the sequence
p
{Tm }m∈N and that Km  → 0 when p → ∞, it is easy to prove that the sequence
{am }m∈N is convergent. Denote by a the limit of the last sequence. Then, uniformly
p p
on p, {Tam }m∈N converges to Ta . This fact implies that the sequence of compact
p
operators {Km }m∈N converges to a compact operator K p . Then,
K p
T $ Ta + K p .
p∈Z+

The standard ε3 -trick implies that K p  → 0 when p → ∞.

To prove its uniqueness of the representation (3.1), we assume that
K p p
K p p
T $ Ta + K1 = Tb + K2 ,
p∈Z+ p∈Z+

p p
where a, b ∈ VO∂ (Bn−1 ), K1 , K2 are compact operators, for each p ∈ Z+ , and
p p
where limp→∞ K1  = limp→∞ K2  = 0. Then,
p p p
Ta−b + K1 − K2 = 0, for each p ∈ Z+ . (3.2)

Last equation and Lemma 3.1 imply that

p p p p
a − b∞ = lim Ta−b  = lim Ta−b + K1 − K2  = 0.
p→∞ p→∞

p p
Then a = b and, from (3.2), K1 = K2 .
Corollary 3.3 The algebra S VO possesses the compact semi-commutator property,
i.e.,

[T a , T b ) is compact for every a, b ∈ S VO .

Each operator T ∈ T (S VO ) admits a representation

T = T a + K,

where a ∈ S VO , K ∈ T (S VO ) ∩ K(H 2 (S 2n−1 )), and where K(H 2 (S 2n−1 )) is the

ideal of all compact operators in L(H 2 (S 2n−1 )).
412 M. Loaiza and N. Vasilevski

For S = L∞ (Bn−1 ), let S ∞ be the corresponding class of symbols defined on

2n−1 . We conjecture that the algebra S ∗
S VO is the largest C -subalgebra among those
of S ⊂ S ∞ that possesses the compact semi-commutator property, i.e.,

[T a , T b ) is compact for every a, b ∈ S ⊂ S ∞ if and only if S ⊆ S VO .

We list now a family of infinite dimensional irreducible representations of the

C ∗ -algebra T (S VO ).
Proposition 3.4 For each p ∈ Z+ the mapping

ιp : T (S VO ) −→ Tp (VO∂ (Bn−1 ))
K p
T $ (Tak + Kk ) −→ Ta + K p ,
k∈Z+

is an irreducible representation of the C ∗ -algebra T (S VO ).

For different p ∈ Z+ the representations ιp are not unitarily equivalent.
Proof Note that the C ∗ -algebra, generated by the images under (2.4) of the
restrictions of the elements of T (S VO ) onto its invariant subspaces, coincides with
the algebra Tp (VO∂ (Bn−1 )). This is an irreducible C ∗ -algebra containing the ideal
of all compact operators on A2p (Bn−1 ). That is the representation ιp is irreducible.
To prove that, for p = q, the representations ιp , ιq are not unitarily equivalent it
suffices to consider the function c(z , zn ) = 1−|z |2 , whose associated function (2.3)
is given by c(z ) = 1 − |z |2 . Then,
K
Tc $ Tck ,
k∈Z+

p
and ιp (Tc ) = Tc , which is a diagonal operator with respect to the standard basis of
A2p (Bn−1 ). By [13, Theorem 10.1] and [3],

p p+1 q +1 q
T1−|z |2  = = = T1−|z |2 ,
n+p n+q

implying thus that the representations ιp , ιq are not unitarily equivalent.

It is easy to check (see [8] for the unweighted case) the following isomorphism

Tp (VO∂ (Bn−1 )/K(A2p (Bn−1 ) ∼

= VO∂ (Bn−1 )/C0 (Bn−1 )
∼
= C(M(VO∂ (Bn−1 )) \ Bn−1 ),
Toeplitz Operators on the Unit Sphere 413

which is defined by

Ta + K(A2p (Bn−1 )) −→ a + C0 (Bn−1 ) ∼

p
= a|M(VO∂ (Bn−1 ))\Bn−1 ,

where M(VO∂ (Bn−1 )) denotes the maximal ideal space of VO∂ (Bn−1 ).
This leads to the following family of one-dimensional irreducible representa-
tions:
Corollary 3.5 For every (η, p) ∈ M(VO∂ (Bn−1 )) \ Bn−1 × Z+ the map
K ιp p
πη,p : T $ (Tak + K k ) −→ Ta + K p −→ a(η)
k∈Z+

defines a one-dimensional irreducible representation. Moreover, πη1 ,p1 and πη2 ,p2
are unitarily equivalent if and only if η1 = η2 .
Further we have the following variant of [5, Lemma 4.8] and [7, Lemma 5.4].
Lemma 3.6 The mapping

ν : T (S VO ) −→ VO∂ (Bn−1 ) = C(M(VO∂ (Bn−1 ))),

defined by
K
ν : T $ T p −→ lim Bp (T p ),
p→∞
p∈Z+

where Bp is the Berezin transform, is a continuous ∗-homomorphism of the C ∗ -

algebra T (S VO ) onto C(M(VO∂ (Bn−1 ))).
This result permits us (as in [5, Remark7
4.10]) to recover the data of the unique
representation (3.1) for each operator T $ p∈Z+ T p ∈ T (S VO ):

p
a = ν(T ) = lim Bp (T p ) ∈ VO∂ (Bn−1 ) and K p = T p − Ta .
p→∞

Corollary 3.7 For each η ∈ M(VO∂ (Bn−1 )), the mapping νη : T (S VO ) −→ C,

defined by

νη : T −→ ν(T ) = a −→ a(η) ∈ C,

is a one-dimensional representation of the C ∗ -algebra T (S VO ).

We show now that we have listed all (up to the unitary equivalence) irreducible
representations of the algebra T (S VO ).
414 M. Loaiza and N. Vasilevski

Introduce the set

KVO := {K ∈ T (S VO ) : ν(K) = 0}.

In representation (3.1), these are exactly the operators of the form

K
K$ Kp (3.3)
p∈Z+

with K p compact and K p  → 0 as p → ∞.

Lemma 3.8 We have that T (S VO ) ∩ K(H 2 (S 2n−1 )) = KVO .
Proof Note that each element K ∈ KVO is a compact operator. Indeed, using the
representation (3.3) for K we have that
; ; ; ;
;K K ; ;K ;
; ; ; ;
; − p;
= ; p;
; = sup K  → 0, when k → ∞.
p p
; K K ; ; K
;p∈Z+ p≤k ; ; p>k ; p>k

7
Then p∈Z+ K p is compact
7 and, as a consequence, K 7 is compact. Consider now
a compact operator T $ p∈Z+ T p in T (S VO ). Since p∈Z+ T p is compact we
have that

lim T p  = 0.
p→∞

Then

|ν(T )| = lim |Bp (T p )| ≤ lim T p  = 0,

p→∞ p→∞

then T ∈ KVO .
This implies that the Calkin algebra

T (S VO )/(T (S VO ) ∩ K(H 2(S 2n−1 )))

is isomorphic to VO∂ (Bn−1 ) via the induced mapping

 
ν̂ : T + T (S VO ) ∩ K(H 2 (S 2n−1 )) −→ lim Bp (T p ).
p→∞

Theorem 3.9 The following list of irreducible representations of the C ∗ -algebra

T (S VO ) is complete up to unitary equivalence:
7 p
• ιp : T $ p∈Z+ (Tak + K k ) → Ta + K p , for p ∈ Z+ ,
• νη : T −→ ν(T ) = c −→ c(η) ∈ C, for η ∈ M(VO∂ (Bn−1 )).
Moreover, the above representations are pairwise not unitarily equivalent.
Toeplitz Operators on the Unit Sphere 415

Proof The proof of this theorem is exactly the same of [7, Theorem 5.7]. We
include it here for the sake of completeness. According to [11, Proposition 2.11.2]
each irreducible representation of T (S VO ) is either induced by an irreducible
representation of

T (S VO )/(T (S VO ) ∩ K(H 2(S 2n−1 ))) (3.4)

or is an extension of an irreducible representation of T (S VO ) ∩ K(H 2 (S 2n−1 )).

In the first case, the algebra (3.4) is isomorphic to VO∂ (Bn−1 ) and then, its
irreducible representations are exactly the representations νη . On the other hand,
by Lemma 3.8, T (S VO ) ∩ K(H 2 (S 2n−1 )) equals KVO . Then, each irreducible
representation of KVO , restricted to the different levels p, is either 0 or an irreducible
representation of K(A2p (Bn−1 )). In the second case, there is only one option: the
identity representation, which extends naturally to the identical representation of
Tp (VO∂ (Bn−1 ), and generates thus the representation ιp .
As a corollary of Lemmas 3.6 and 3.8, we have also the following result on the
Fredholmness of an operator T ∈ T (S VO ).
7 p
Corollary 3.10 An operator T $ p∈Z+ (Ta + K ) ∈ T (S VO ) is Fredholm if
p

and only if a(η) = 0 for all η ∈ M(VO∂ (Bn−1 )). In particular,

ess-sp T = a(M(VO∂ (Bn−1 ))).

Using Corollary 3.3 we give the following extended version of Corollary 3.10.
Proposition 3.11 The Calkin algebra
 (S VO ) = T (S VO )/(T (S VO ) ∩ K(H 2 (S 2n−1 )))
T

is isomorphic to VO∂ (Bn−1 ). The corresponding homomorphism

 (S VO ) ∼
π : T (S VO ) −→ T = VO∂ (Bn−1 )

is given by

π : T = T a + K −→ a(η),

where a ∈ S VO and a ∈ VO∂ (Bn−1 ) is connected with a by (2.3).

The Toeplitz operator T a , with symbol a ∈ S VO , is compact if and only if a ≡ 0
on S 2n−1 , being thus a zero operator.
The essential spectrum of each operator T = T a + K of the algebra T (S VO )
and the spectral radius of T a are given by

ess-sp T = clos (Range a) = clos (Range a) ,

r(T a ) = sup |a(z)| = sup |a(z )| = T a .
z∈S 2n−1 z ∈Bn−1
416 M. Loaiza and N. Vasilevski

Each common invariant subspace for all the operators from T (S VO ) is of the
form
⎛ ⎞
K
U −1 ⎝ A2p (Bn−1 )⎠ ,
p∈N

where N is a subset of Z+ , and the operator U is defined in Theorem 2.1.

An operator T = T a + K ∈ T (S VO ) is Fredholm if and only if there is δ =
δ(a) > 0 such that |a(z)| ≥ δ for all z ∈ S 2n−1 or, equivalently, |a(z )| ≥ δ for all
z ∈ Bn−1 .
It is instructive to give the following.
Remark 3.12 In the seminal paper [10] L. Coburn gave a detailed description of the
C ∗ -algebra T (C(S 2n−1 )) generated by all Toeplitz operators T a with continuous
symbols a ∈ C(S 2n−1 ).
The largest possible C ∗ -algebra that inherits the nice main properties of
T (C(S 2n−1 )) (irreducibility, form of a generic element of the algebra: T a + K,
essential spectrum formula: clos (Range a), etc) has to be generated by Toeplitz
operators with symbols from the largest C ∗ -subalgebra, say Qn , of L∞ (S 2n−1 )
possessing the compact semi-commutator property.
The results of this section show that S VO ⊂ Qn . Note that functions from
S VO are not continuous in general, they possess quite sophisticated (of vanishing
oscillation type) discontinuities at S 2n−1 ∩ {(z , zn ) : zn = 0}. Further, the algebra
S VO (apart of being a subalgebra of Qn ) consists of functions that are invariant
under the one-dimensional torus action on S 2n−1 :

τ ∈ T : (z , zn ) ∈ S 2n−1 −→ (z , τ zn ) ∈ S 2n−1 .

This T-invariance implies that our C ∗ -algebra T (S VO ) (contrary to the algebra

T (Qn )) is not anymore irreducible, and we have described all its irreducible
representations, as well as, all common invariant subspaces for its elements.
We discuss now briefly the Fredholm index formula for a matrix-valued sym-
bol situation. As shown in [19, Section 2], the index calculation for Fredholm
operators with VO∂ symbols can be reduced to those with continuous up to the
boundary symbols in the following way. Given a matrix-valued function a ∈
Matq (VO∂ (Bn−1 )) := VO∂ (Bn−1 ) ⊗ Matq (C), for each s ∈ (0, 1), we define

a(rζ ), if 0 ≤ r ≤ s,
as (rζ ) =
a(sζ ), if s < r ≤ 1,

n−1
here r = |z | and ζ ∈ S 2n−3 . Then, obviously, as (rζ ) ∈ Matq (C(B )), for all
s ∈ (0, 1).
Toeplitz Operators on the Unit Sphere 417

The matrix-valued version of Corollary 3.10 is as follows.

Corollary 3.13 Given a ∈ Matq (VO∂ (Bn−1 )), the operator
K p
T $ (Ta + K p ) ∈ T (S VO ) ⊗ Matq (C)
p∈Z+

is Fredholm if and only if the matrix a(η) is invertible for all

η ∈ M(VO∂ (Bn−1 )).

In particular,
 
ess-sp T = λ : λ ∈ sp{a(η)}, η ∈ M(VO∂ (Bn−1 )) .

In case of being Fredholm,

⎡ ⎤
K
Ind T = Ind ⎣ + K p )⎦ = 0.
p
(Ta
p∈Z+

Proof Only the index formula needs to be justified. Let T be Fredholm, then the
matrix-valued function a is invertible, and
 p
Ind T = Ind (Ta + K p ).
p∈Z+

For each p ∈ Z+ , Theorem 2.6 of [19], ensures that there is sρ,0 ∈ (0, 1) such that
p
for every s ∈ (sρ,0 , 1), each operator Tas + K p is Fredholm, and
p p
Ind (Ta + K p ) = Ind (Tas + K p ).

n−1
The matrix-valued function as is continuous on the closed unit ball B . The ball
n−1
B is retractable to a point, thus the matrix-valued function as is homotopic
n−1
to a constant matrix, say ap , in a class of invertible continuous on B matrix-
p
functions. This implies that the operator Tas + K is homotopic to the scalar-matrix
p

multiplication operator ap I . Thus, for each p ∈ Z+ ,

p
Ind (Ta + K p ) = Ind ap I = 0.

p
Note that, starting from some p0 , all operators Ta + K p are invertible, so that
both ker T and coker T are finite dimensional (see [5, p. 730] for details).
One, of course, easily makes a version of Proposition 3.11 for matrix-valued
symbols a ∈ Matq (VO∂ (Bn−1 )).
418 M. Loaiza and N. Vasilevski

4 Full Toeplitz Algebra in Levels

In this section we consider the case when, for each p ∈ Z+ , the C ∗ -algebra
consisting of all operators T p of decomposition (2.5) coincides with the full
Toeplitz algebra, i.e., with the algebra Tp (L∞ (Bn−1 )), which is generated by all
p
Toeplitz operators Ta , with a ∈ L∞ (Bn−1 ), acting on the weighted Bergman space
Ap (B ).
2 n−1

We introduce as well the C ∗ -algebra BUC(Bn−1 ) of functions a(z ), z ∈

B , which are uniformly continuous in Bn−1 with respect to the Bergman
n−1
p
metric β(·, ·). Note (see [14, Theorem 7.3]) that each operator Ta , with a ∈
L∞ (Bn−1 ) can be approximated in norm by Toeplitz operators with symbols
from BUC(Bn−1 ). Moreover, the results of [18] imply that Toeplitz operators with
symbols from BUC(Bn−1 ) form a dense set in Tp (L∞ (Bn−1 )). That is, in particular,
Tp (BUC(Bn−1 )) = Tp (L∞ (Bn−1 )).
The above results suggest us to choose S = BUC(Bn−1 ). Let S BUC be the
corresponding class of functions on S 2n−1 , connected with functions from S =
BUC(Bn−1 ) by (2.3).
We describe briefly the C ∗ -algebra T (S BUC ), which is generated by all Toeplitz
operators T a , with a ∈ S BUC , acting on the Hardy space H 2 (S 2n−1 ). Doing this we
follow the lines of [7, Section 4], where all details can be found.
We mention first that each operator T a ∈ T (S BUC ) admits the representation
K p p
Ta $ Ta , with Ta ∈ Tp (BUC(Bn−1 )).
p∈Z+

Similarly to Lemma 3.1 we have

T  = sup T p ,
p∈Z+

while for the generating operators we have

T a  = a(z )∞ .

where the last equality follows from [4, Proposition 4.4].

The proof of the next theorem literally follows the proof of Theorem 3.2 or [7,
Theorem 4.3], and thus will be omitted.
Theorem 4.1 Each operator T a ∈ T (S BUC ), in the direct sum decomposi-
tion (2.5), admits the unique representation
K p
T $ (Ta + N p ), (4.1)
p∈Z+
Toeplitz Operators on the Unit Sphere 419

where a ∈ BUC(Bn−1 ), N p belongs to the semi-commutator ideal of the algebra

Tp (BUC(Bn−1 )), and N p  → 0 as p → ∞.
Corollary 4.2 The semi-commutator ideal N (S BUC ) of T (S BUC ) consists of all
its elements
K
N$ Np,
p∈Z+

coming from the representation (4.1).

Denote by Tp (BUC(Bn−1 )) the algebra which consists of all operators T p =
p
Ta + N p that appear on the pth level of decomposition (2.5), or, which is the same,
which consists of all operators being the restrictions of elements (4.1) of T (S BUC )
onto its invariant subspace A2p (Bn−1 ).
Lemma 4.3 For each p ∈ Z+ , the algebra Tp (BUC(Bn−1 )) coincides with the
algebra Tp (BUC(Bn−1 )) = Tp (L∞ (Bn−1 )).
Proof It is easy to see that the mapping

ιp : T (S BUC ) −→ L(A2p (Bn−1 ))

K
T $ T k −→ T p
k∈Z+

is a morphism (representation) of the algebra T (S BUC ). Besides, its image is

p
Tp (BUC(Bn−1 )). Each operator Ta with a ∈ BUC(Bn−1 ) obviously belongs to
Tp (BUC(Bn−1 )). As it was already mentioned, the set of all such operators is norm
dense in Tp (L∞ (Bn−1 )), and the algebra Tp (BUC(Bn−1 )), being the image of a
representation, is norm closed.
It is instructive to mention that, in spite of the fact that for each p ∈ Z+ , the
p
algebra Tp (BUC(Bn−1 )) = Tp (BUC(Bn−1 )) contains each Toeplitz operator Ta
with a ∈ L∞ (B ), the algebra T (S BUC ) does not contain any Toeplitz operator
n−1

T a with a from S ∞ \ S BUC . That is, the algebra T (S BUC ) is a proper subalgebra
of T (S ∞ ). This can be justified literally following the arguments of the proof of [7,
Lemma 4.5].
We describe now the irreducible representations of the C ∗ -algebra
T (S BUC ). First, for each p ∈ Z+ , the representation

ιp : T (S BUC ) −→ L(A2p (Bn−1 )) (4.2)

K p
T $ (Tak + N p ) −→ Ta + N p ,
k∈Z+
420 M. Loaiza and N. Vasilevski

whose image is Tp (BUC(Bn−1 )), is irreducible (as the algebra Tp (BUC(Bn−1 ))

contains the whole ideal of compact operators). The same reasoning as in case of
T (S VO ) implies that for different p ∈ Z+ these representations are not unitary
equivalent.
Similarly to Lemma 3.6 and [7, Lemma 4.8], we have that the mapping

ν : T (S BUC ) −→ BUC(Bn−1 ) = C(M(BUC(Bn−1 ))),

defined by
K
ν : T $ T p −→ lim Bp (T p ),
p→∞
p∈Z+

where Bp is the Berezin transform, is a continuous ∗-homomorphism of the C ∗ -

algebra T (S BUC ) onto C(M(BUC(Bn−1 ))). Here M(BUC(Bn−1 )) denotes the
compact of maximal ideals of the algebra7 BUC(Bn−1 ).
As in [7, Remark 4.10], given T $ p∈Z+ T
p ∈ T (S
BUC ), the above result
permits us to recover the data of its unique representation (4.1):
p
a = ν(T λ ) = lim Bp (T p ) ∈ BUC(Bn−1 ) and N p = T p − Ta .
p→∞

Then we have
Corollary 4.4 For each η ∈ M(BUC(Bn−1 )), the mapping

νη : T (S BUC ) −→ C,

defined by

νη : T −→ ν(T ) = a −→ a(η) ∈ C, (4.3)

is a one-dimensional representation of the C ∗ -algebra T (S BUC ).

We gather all so far obtained information on the C ∗ -algebra T (S BUC ) in the
following proposition.
 (S BUC ) = T (S BUC )/N (S BUC ) of
Proposition 4.5 The quotient algebra T
T (S BUC ) by its semi-commutator ideal is isomorphic to BUC(Bn−1 ). The
corresponding homomorphism

π : T (S VO ) −→ T
  (S VO ) ∼
= BUC(Bn−1 )
Toeplitz Operators on the Unit Sphere 421

is given by
K p

π: T $ (Ta + N p ) −→ a(η),
p∈Z+

where a ∈ S BUC and a ∈ BUC(Bn−1 ) is connected with a by (2.3).

The Toeplitz operator T a , with symbol a ∈ S BUC , is compact if and only if a ≡ 0
on S 2n−1 , being thus a zero operator. 7
p
The spectrum of each operator T = p∈Z+ (Ta + N p ) of the algebra T (S BUC )
contains clos (Range a), and the spectral radius of T a is given by

r(T a ) = sup |a(z)| = sup |a(z )| = T a .

z∈S 2n−1 z ∈Bn−1

The C ∗ -algebra T (S BUC ) is reducible, its irreducible representations are given

by (4.2) and (4.3). Each common invariant subspace for all the operators from
T (S BUC ) is of the form
⎛ ⎞
K
U −1 ⎝ A2p (Bn−1 )⎠ ,
p∈N

where N is a subset of Z+ , and the operator U is defined in Theorem 2.1.

Acknowledgement This work was partially supported by CONACYT Project 238630, México.

References

1. Z. Akkar, Zur Spektraltheorie von Toeplitzoperatoren auf dem Hardyraum H 2 (Bn ). Ph.D.
Dissertation, Universität des Saarlandes, 2012
2. Z. Akkar, E. Albrecht, Spectral properties of Toeplitz operators on the unit ball and on the unit
sphere, in The Varied Landscape of Operator Theory. Theta Series in Advanced Mathematics,
vol. 17 (Theta, Bucharest, 2014), pp. 1–22
3. H. Alzer, Inequalities for the Beta function of n variables. ANZIAM J. 44, 609–623 (2003)
4. W. Bauer, L.A. Coburn, Heat flow, weighted Bergman spaces and real analytic Lipschitz
approximation. J. Reine Angew. Math. 703, 225–246 (2015)
5. W. Bauer, N. Vasilevski, On algebras generated by Toeplitz operators and their representations.
J. Funct. Anal. 272, 705–737 (2017)
6. W. Bauer, R. Hagger, N. Vasilevski, Uniform continuity and quantization on bounded
symmetric domains. J. Lond. Math. Soc. 96, 345–366 (2017)
7. W. Bauer, R. Hagger, N. Vasilevski, Algebras of Toeplitz operators on the n-dimensional unit
ball. Complex Anal. Oper. Theory 13, 493–524 (2019)
8. C.A. Berger, L.A. Coburn, K.H. Zhu, Function theory on Cartan domains and the Berezin-
Toeplitz symbol calculus. Amer. J. Math. 110, 921–953 (1988)
9. A. Brown, P.R. Halmos, Algebraic properties of Toeplitz operators. J. Reine Angew. Math.
213, 89–102 (1964)
422 M. Loaiza and N. Vasilevski

10. L. Coburn, Singular integral operators and Toeplitz operators on odd spheres. Indiana Univ.
Math. J. 23, 433–439 (1973)
11. J. Dixmier, Les C ∗ -algebres et leurs représensations (Gauthier-Villars, Paris 1964)
12. M. Loaiza, N. Vasilevski, Commutative algebras generated by Toeplitz operators on the unit
sphere. Intgr. Equ. Oper. Theory 92, 25 (2020). https://doi.org/10.1007/s00020-020-02580-x
13. R. Quiroga-Barranco, N.L. Vasilevski, Commutative algebras of Toeplitz operators on the unit
ball I: Bargmann type transforms and spectral representations of Toeplitz operators. Integr.
Equ. Oper. Theory 59, 379–419 (2007)
14. D. Suárez, The essential norm of operators in the Toeplitz algebra Ap (Bn ). Indiana Univ. Math.
J. 56, 2185–2232 (2007)
15. N.L. Vasilevski, Toeplitz operators on the Bergman spaces: inside-the-domain effects. Con-
temp. Math. 289, 79–146 (2001)
16. N.L. Vasilevski, Bergman space structure, commutative algebras of Toeplitz operators and
hyperbolic geometry. Integr. Equ. Oper. Theory 46, 235–251 (2003)
17. N.L. Vasilevski, Commutative Algebras of Toeplitz Operators on the Bergman Space. Operator
Theory: Advances and Applications, vol. 185 (Birkhäuser Verlag, Boston, 2008)
18. J. Xia, Localization and the Toeplitz algebra on the Bergman space. J. Funct. Anal. 269, 781–
814 (2015)
19. J. Xia, D. Zheng, Toeplitz operators and Toeplitz algebra with symbols of vanishing oscillation.
J. Operator Theory 76, 107–131 (2016)
d-Modified Riesz Potentials on Central
Campanato Spaces

Katsuo Matsuoka

Dedicated to Professor Lars-Erik Persson in celebration of

his 75th birthday

Abstract Recently, we defined the d-modified Riesz potentials I˜α,d and proved
several estimates of boundedness of I˜α,d on the central Morrey spaces B p,λ (Rn ),
(d)
using the central Campanato spaces $p,λ (Rn ), the generalized σ -Lipschitz spaces
(d)
Lipβ,σ (Rn ) and so on. In this paper, we will consider the results of the boundedness
for I˜α,d on the λ-central mean oscillation spaces CMOp,λ (Rn ).

Keywords Central Morrey space · λ-central mean oscillation space · Weak

λ-central mean oscillation space · σ -Lipschitz space · σ -BMO space · Central
Campanato space · Weak central Campanato space · Generalized σ -Lipschitz
space · Generalized σ -BMO space · Riesz potential · Modified Riesz potential ·
d-modified Riesz potential

Mathematics Subject Classification (2010) Primary 42B35; Secondary 26A33,

46E30, 46E35

This work was supported by Grant-in-Aid for Scientific Research (C) (Grant No. 17K05306),
Japan Society for the Promotion of Science.

K. Matsuoka ()
College of Economics, Nihon University, Misaki-cho, Kanda, Chiyoda-ku, Tokyo, Japan
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 423

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_21
424 K. Matsuoka

1 Introduction

Let Iα and I˜α , 0 < α < n, be the Riesz potential and the modified Riesz potential,
respectively, which are defined for f ∈ L1loc (Rn ) by

f (y)
Iα f (x) = dy
Rn |x − y|n−α

and
  
1 1 − χQ1 (y)
I˜α f (x) = f (y) − dy
R n |x − y| n−α |y|n−α

(as for the notation χQ1 , see Sect. 2). Then, the following boundedness results on
Lp (Rn ) are well-known : For 0 < α < n, 1 ≤ p < n/α and 1/q = 1/p − α/n,
(i) Iα : Lp (Rn ) → Lq (Rn ), 1 < p < n/α (see [7, 25]);
(ii) Iα : L1 (Rn ) → W Lq (Rn ), p = 1 (see [26]).
And for 0 < α < n, n/α ≤ p < ∞ and β = α − n/p < 1,
(iii) I˜α : Lp (Rn ) → Lipβ (Rn ), n/α < p < ∞ (cf. [23]);
(iv) I˜α : Ln/α (Rn ) → BMO(Rn ), p = n/α, i.e., β = 0 (cf. [23, 26]).
Here (i) is the so-called Hardy–Littlewood–Sobolev theorem.
After that, in [4], for Iα on the (non-homogeneous) central Morrey space
B p,λ (Rn ), i.e.,
p
B p,λ (Rn ) = {f ∈ Lloc (Rn ) : f B p,λ < ∞},

where 1 ≤ p < ∞, λ ∈ R and

  1/p
1 1
f B p,λ = sup |f (y)|p dy
r≥1 rλ |Qr | Qr

−n/p−λ
(see [1, 4]; cf. [6]), which is the (non-homogeneous) Herz space Kp,∞ (Rn ) (cf.
[3, 8]), the following boundedness result was obtained : For 0 < α < n, 1 < p <
n/α, −n/p + α ≤ μ = λ + α < 0 and 1/q = 1/p − α/n,
(i’-1) Iα : B p,λ (Rn ) → B q,μ (Rn ).
On the other hand, in [11, 19] (cf. [10]), from the Bσ -Morrey–Campanato
estimates for Iα and I˜α , the following boundedness results on B p,λ (Rn ) were
obtained as the corollaries : For 0 < α < n, −n + α ≤ μ = λ + α < 0 and
1/q = 1 − α/n,
(ii’-1) Iα : B 1,λ (Rn ) → W B q,μ (Rn ).
d-Modified Riesz Potentials on Central Campanato Spaces 425

Here W B p,λ (Rn ) is a weak central Morrey space, i.e.,

p
W B p,λ (Rn ) = {f ∈ Lloc (Rn ) : f W B p,λ < ∞},

where 1 ≤ p < ∞, λ ∈ R and

 1/p
1 1
f W B p,λ = sup sup t p |{y ∈ Qr : |f (y)| > t}|
r≥1 rλ |Qr | t >0

(see [11]). Also for 0 < α < n, 1 ≤ p < n/α, −n/p + α ≤ μ = λ + α < 1 and
1/q = 1/p − α/n,
(i’-2) I˜α : B p,λ (Rn ) → CMOq,μ (Rn ), 1 < p < n/α;
(ii’-2) I˜α : B 1,λ (Rn ) → W CMOq,μ (Rn ), p = 1.
And for 0 < α < n, n/α ≤ p < ∞, −n/p + α ≤ λ + α < 1 and β = α − n/p,
(iii’) I˜α : B p,λ (Rn ) → Lipβ,λ+n/p (Rn ), n/α < p < ∞;
(iv’) I˜α : B n/α,λ (Rn ) → BMOλ+n/p (Rn ), p = n/α, i.e., β = 0.
As for the spaces CMOp,λ (Rn ), W CMOp,λ (Rn ) and Lipβ,σ (Rn ), BMOσ (Rn ), refer
to Definitions 2.1, 2.2 and 2.11. Furthermore, the following boundedness results on
CMOp,λ (Rn ) were also gotten excepting (ii”) (as for (ii”), see Corollary 3.5): For
0 < α < 1, 1 ≤ p < n/α, −n/p + α ≤ λ + α = μ < 1 and 1/q = 1/p − α/n (see
[19]; cf. [16]),
(i”) I˜α : CMOp,λ (Rn ) → CMOq,μ (Rn ), 1 < p < n/α;
(ii”) I˜α : CMO1,λ (Rn ) → W CMOq,μ (Rn ), p = 1.
And for 0 < α < 1, n/α ≤ p < ∞, −n/p + α ≤ μ = λ + α < 1 and β =
α − n/p,
(iii”) I˜α : CMOp,λ (Rn ) → Lipβ,λ+n/p (Rn ), n/α < p < ∞;
(iv”) I˜α : CMOn/α,λ (Rn ) → BMOλ+n/p (Rn ), p = n/α, i.e., β = 0.
Recentry, for the whole of μ = λ + α such that 1 ≤ μ < ∞, we
extended the boundedness of I˜α for B p,λ (Rn ). In order to do so, in [17, 18],
we introduced the central Campanato spaces $(d) n
q,μ (R ) and the generalized σ -Lip
(d)
spaces Lipβ,σ (Rn ) (see Definitions 2.3 and 2.11), and also we defined the “ higher-
degree ” modification of Iα , i.e., the d-modified Riesz potentials I˜α,d , 0 < α < n
and d ∈ N ∪ {0}. Then, the following boundedness results were shown : For
0 < α < n, 1 < p < n/α, d ∈ N ∪ {0}, −n/p + α ≤ μ = λ + α < d + 1
and 1/q = 1/p − α/n,
(I) I˜α,d : B p,λ (Rn ) → $q,μ (Rn ), 1 < p < n/α;
(d)

(II) I˜α,d : B 1,λ (Rn ) → W $(d) n

q,μ (R ), p = 1.
426 K. Matsuoka

And for 0 < α < n, n/α ≤ p < ∞, d ∈ N ∪ {0}, −n/p + α + d ≤ λ + α < d + 1

and β = α − n/p,
(III) I˜α,d : B p,λ (Rn ) → Lipβ,λ+n/p (Rn ), n/α < p < ∞;
(d)

(IV) I˜α,d : B n/α,λ (Rn ) → BMO(d) n

λ+n/p (R ), p = n/α, i.e., β = 0.
In this paper, therefore, we will investigate the extension of the above bounded-
ness results of I˜α,d on B p,λ (Rn ), i.e., the boundedness of I˜α,d for CMOp,λ (Rn ). As
a by-product, we can obtain the estimates of I˜α for B p,λ (Rn ).
We note that the same results in this paper still hold for the homogeneous versions
of the function spaces.

2 Central Campanato Spaces and Generalized σ -Lip Spaces

We start by explaining the notation used in the present paper. The symbol A  B
stands for A ≤ CB for some constant C > 0. If A  B and B  A, we then write
A ∼ B. For r > 0, by Qr , we mean the following: Qr = {y ∈ Rn : |y| < r} or
Qr = {y = (y1 , y2 , · · · , yn ) ∈ Rn : max1≤i≤n |yi | < r}. Further, for x ∈ Rn , we
set Q(x, r) = x + Qr = {x + y : y ∈ Qr }. For a measurable set G ⊂ Rn , we
denote by |G| and χG the Lebesgue measure of G and the characteristic function
of G, respectively. And also, for a function f ∈ L1loc (Rn ) and a measurable set
G ⊂ Rn with 0 < |G| < ∞, let
 
1
fG = − f (y) dy = f (y) dy,
G |G| G

and let N0 = N ∪ {0}.

First of all, we define the λ-central mean oscillation (λ-CMO) spaces
CMOp,λ (Rn ) and the weak λ-CMO spaces W CMOp,λ (Rn ) (see [1, 15, 16]; cf.
[9, 10]) and introduce the central Campanato spaces $(d) n
p,λ (R ) and the weak central
(d)
Campanato spaces W $p,λ (Rn ) (see [6, 12, 17]; cf. [22]).
Definition 2.1 For 1 ≤ p < ∞ and λ ∈ R,
p
CMOp,λ (Rn ) = {f ∈ Lloc (Rn ) : f CMOp,λ < ∞},

where
 1/p
1
f CMOp,λ = sup λ − |f (y) − fQr | dy
p
.
r≥1 r Qr
d-Modified Riesz Potentials on Central Campanato Spaces 427

Particularly

CMOp,0 (Rn ) = CMOp (Rn ).

Definition 2.2 For 1 ≤ p < ∞ and λ ∈ R,

p
W CMOp,λ (Rn ) = {f ∈ Lloc (Rn ) : f W CMOp,λ < ∞},

where
 1/p
1 1
f W CMOp,λ = sup sup t p |{y ∈ Qr : |f (y) − fQr | > t}| ,
r≥1 rλ |Qr | t >0

Particularly

W CMOp,0 (Rn ) = W CMOp (Rn ).

Here remark that CMOp (Rn ) and W CMOp (Rn ), so-called the central mean
oscillation (CMO) space and the weak CMO space, are defined by
p
CMOp (Rn ) = {f ∈ Lloc (Rn ) : f CMOp < ∞},

where
 1/p
f CMOp = sup − |f (y) − fQr |p dy ,
r≥1 Qr

and
p
W CMOp (Rn ) = {f ∈ Lloc (Rn ) : f W CMOp < ∞},

where
 1/p
1
f W CMOp = sup sup t p |{y ∈ Qr : |f (y) − fQr | > t}| ,
r≥1 |Qr | t >0

respectively (see [2, 5]).

Definition 2.3 For 1 ≤ p < ∞, d ∈ N0 and −n/p ≤ λ < d + 1, a function
p
f ∈ Lloc (Rn ) will be said to belong to the central Campanato space $(d) n
p,λ (R ) if
and only if for every r ≥ 1, there is a polynomial Prd f of degree at most d such that
 1/p
1
f $(d) = sup λ − |f (y) − Pr f (y)| dy
d p
< ∞.
p,λ r≥1 r Qr
428 K. Matsuoka

Particularly

(0)
$p,λ (Rn ) = CMOp,λ (Rn ).

Definition 2.4 For 1 ≤ p < ∞, d ∈ N0 and −n/p ≤ λ < d + 1, a function f ∈

p
Lloc (Rn ) will be said to belong to the weak central Campanato space W $p,λ (Rn )
if and only if for every r ≥ 1, there is a polynomial Prd f of degree at most d such
that
 1/p
1 1
f W $p,λ = sup λ sup t p |{y ∈ Qr : |f (y) − Prd f (y)| > t}| < ∞.
r≥1 r |Qr | t >0

Particularly

W $(0)
p,λ (R ) = W CMO
n p,λ
(Rn ).

Here we note that identifying functions which differ by a polynomial of degree at

most d, a.e., we see that $(d) n (d) n
p,λ (R ) and W $p,λ (R ) are Banach and quasi-Banach
spaces, respectively.
Remark 2.5 (Remark 6.2 of [22]) For 1 ≤ p < ∞, d ∈ N0 , −n/p ≤ λ < d + 1
p
and f ∈ Lloc (Rn ), we have
 1/p
1
f $(d) ∼ sup inf λ
− |f (y) − P (y)| dy
p
p,λ r≥1 P ∈P (R ) r
d n
Qr

and
 1/p
1 1
f W $p,λ ∼ sup inf sup t p |{y ∈ Qr : |f (y) − P (y)| > t}| ,
r≥1 P ∈P (R ) r
d n λ |Qr | t >0

where P d (Rn ) is the set of all polynomials having degree at most d.

Next we define the generalized σ -Lipschitz spaces Lip(d) n
β,σ (R ).
Definition 2.6 (Definition 8.1 of [22]; cf. [12, 18]) Let U = Rn or U = Qr with
r > 0. For d ∈ N0 and 0 ≤ β ≤ 1, a continuous function f will be said to belong
to the generalized Lipschitz (Lip) space on U , i.e., Lip(d)
β (U ) if and only if

1
f Lip(d) (U ) = sup |Ad+1
h f (x)| < ∞,
β x,x+h∈U,h=0 |h| β
d-Modified Riesz Potentials on Central Campanato Spaces 429

where Akh is a difference operator, which is defined inductively by

A0h f = f, A1h f = Ah f = f (· + h) − f (·),

Akh f = Ak−1 k−1

h f (· + h) − Ah f (·), k = 2, 3, · · · .

In particular, we define

(d)
BMO(d)(U ) = Lip0 (U ),

which we call the generalized BMO space on U .

Remark 2.7 Let U = Rn or U = Qr with r > 0. For 0 < β < 1, d ∈ N0 and
β = 1, d ∈ N, the spaces Lip(d) (0)
β (U ) coincide with Lipβ (U ) = Lipβ (U ) (see
Remark 2.12) and Lip(1)
1 (U ), respectively, which are the so-called Nikol’skiı̌ spaces
(cf. Remark 2.4 of [18, 24]).
Remark 2.8 (Theorem 8.3 of [22]) Let U = Rn or U = Qr with r > 0. For
p
1 ≤ p < ∞, d ∈ N0 , 0 ≤ β ≤ 1 and f ∈ Lloc (Rn ), we have

f Lip(d) (U ) ∼ f L(d) (U ) ,
β p,β

where L(d)
p,β (U ) is the Campanato space on U as defined below.
Definition 2.9 Let U = Rn or U = Qr with r > 0. For 1 ≤ p < ∞, d ∈ N0
p
and −n/p ≤ λ < d + 1, a function f ∈ Lloc (U ) will be said to belong to the
(d)
Campanato space on U , i.e., Lp,λ (U ) if and only if for every Q(x, s) ⊂ U , there is
d
a polynomial PQ(x,s) f of degree at most d such that

 1/p
1
f L(d) (U ) = sup − |f (y) − PQ(x,s)
d
f (y)|p dy < ∞.
p,λ Q(x,s)⊂U sλ Q(x,s)

Particularly, when d = 0, we use the well-known notation

Lp,λ (U ) = L(0)
p,λ (U ).

Remark 2.10 (Remark 6.2 of [22]) Let U = Rn or U = Qr with r > 0. For

p
1 ≤ p < ∞, d ∈ N0 , −n/p ≤ λ < d + 1 and f ∈ Lloc (U ), we have
 1/p
1
f L(d) (U ) ∼ sup inf λ
− |f (y) − P (y)|p dy .
p,λ Q(x,s)⊂U P ∈P (U ) s
d
Q(x,s)
430 K. Matsuoka

Definition 2.11 (Definition 2.8 of [18]; cf. Definition 11 of [12]) For d ∈ N0 ,

0 ≤ β ≤ 1 and 0 ≤ σ < ∞, the continuous function f will be said to belong to the
generalized σ -Lipschitz (σ -Lip) space, i.e., Lip(d) n
β,σ (R ) if and only if

1
f Lip(d) = sup f Lip(d) (Q ) < ∞.
β,σ r≥1 rσ β r

In particular, we define
(d)
σ (R ) = Lip0,σ (R )
BMO(d) n n

and
(0)
Lipβ,σ (Rn ) = Lipβ,σ (Rn ), BMOσ (Rn ) = BMO(0) n
σ (R ),

which we call the generalized σ -BMO space and the σ -Lip space, the σ -BMO
space, respectively.
Identifying functions which differ by a polynomial of degree at most d, a.e., we
(d) (d)
see that Lipβ (Rn ) and Lipβ,σ (Rn ) are Banach spaces (see [12] and [22]).
Remark 2.12 ([17]; cf. [10]) We note that particularly

Lip(d) (d) n (d) n

β,0 (R ) = Lipβ (R ), BMO0 (R ) = BMO (R ),
n (d) n

and
(0)
Lipβ (Rn ) = Lipβ (Rn ), BMO(0) (Rn ) = BMO(Rn ).

3 d-Modified Riesz Potentials

Now, as we stated in Sect. 1, under the condition μ = λ + α ≥ 1 we consider

the boundedness of modified Riesz potentials I˜α on CMOp,λ (Rn ). Then, we use the
definition of “ higher-degree ” modification of Riesz potentials Iα , i.e., the following
definition of d-modified Riesz potentials I˜α,d , 0 < α < n and d ∈ N0 (see [13, 14,
20]; cf. Definition 3.1 of [17]).
Definition 3.1 For 0 < α < n and d ∈ N0 , we define the modified Riesz potential
of order α and degree d, i.e., d-modified Riesz potential I˜α,d , as follows : For any
f ∈ L1loc (Rn ) and x ∈ Rn ,

I˜α,d f (x)
⎧ ⎛ ⎞ ⎫
 ⎨  xl ⎬
= f (y) Kα (x − y) − ⎝ (D l Kα )(−y)⎠ (1 − χQ1 (y)) dy,
Rn ⎩ l! ⎭
{l:|l|≤d}
d-Modified Riesz Potentials on Central Campanato Spaces 431

where for any x ∈ Rn \ {0},

1
Kα (x) =
|x|n−α

and for x = (x1 , x2 , · · · , xn ) ∈ Rn and l = (l1 , l2 , · · · , ln ) ∈ Nn0 , |l| = l1 + l2 +

· · ·+ ln , x l = x1l1 x2l2 · · · xnln , l! = l1 ! l2 ! · · · ln ! and D l is the partial derivative of order
l, i.e.,

D l = (∂/∂x1 )l1 (∂/∂x2 )l2 · · · (∂/∂xn )ln .

Note that in particular

I˜α,0 = I˜α

and that I˜α,d (|f |) ≡ ∞ on Rn , if


|f (y)|
dy < ∞
Rn (1 + |y|)n−α+d+1

(cf. [21]). If Iα f is well-defined, then I˜α,d f is also well-defined and Iα f − I˜α,d f is

a polynomial of degree at most d.
Then our first results for a d-modified Riesz potential I˜α,d are the following
strong and weak estimates on CMOp,λ (Rn ), where 1 ≤ p < n/α.
Theorem 3.2 Let 0 < α < 1, 1 < p < n/α, d ∈ N0 , −n/p + α ≤ μ = λ + α <
d + 1 and q = pn/(n − pα), i.e., 1/q = 1/p − α/n. Then I˜α,d is bounded from
CMOp,λ (Rn ) to $(d) n
q,μ (R ), that is, there exists a constant C > 0 such that

I˜α,d f $(d) ≤ Cf CMOp,λ , f ∈ CMOp,λ (Rn ).

q,μ

Theorem 3.3 Let 0 < α < 1, d ∈ N0 , −n + α ≤ μ = λ + α < d + 1 and

q = n/(n − α), i.e., 1/q = 1 − α/n. Then I˜α,d is bounded from CMO1,λ (Rn ) to
(d)
W $q,μ (Rn ), that is, there exists a constant C > 0 such that

I˜α,d f W $(d) ≤ Cf CMO1,λ , f ∈ CMO1,λ (Rn ).

q,μ

In the above theorems, if d = 0, then we can get the following strong and weak
estimates for a modified Riesz potential I˜α,d on CMOp,λ (Rn ).
Corollary 3.4 (Corollary 2.7 of [19]; cf. [17]) Let 0 < α < 1, 1 < p < n/α,
−n/p + α ≤ μ = λ + α < 1 and q = pn/(n − pα), i.e., 1/q = 1/p − α/n. Then
432 K. Matsuoka

I˜α is bounded from CMOp,λ (Rn ) to CMOq,μ (Rn ), that is, there exists a constant
C > 0 such that

I˜α f CMOq,μ ≤ Cf CMOp,λ , f ∈ CMOp,λ (Rn ).

Corollary 3.5 Let 0 < α < 1, −n + α ≤ μ = λ + α < 1 and q = n/(n − α), i.e.,
1/q = 1 − α/n. Then I˜α is bounded from CMO1,λ (Rn ) to W CMOq,μ (Rn ), that is,
there exists a constant C > 0 such that

I˜α f W CMOq,μ ≤ Cf CMO1,λ , f ∈ CMO1,λ (Rn ).

Next, for a d-modified Riesz potential I˜α,d , we prove our second result, i.e., the
following estimates on CMOp,λ (Rn ), where n/α ≤ p < ∞.
Theorem 3.6 Let 0 < α < 1, n/α ≤ p < ∞, d ∈ N0 , −n/p + α + d ≤ λ + α <
d +1 and β = α −n/p. Then I˜α,d is bounded from CMOp,λ (Rn ) to Lip(d) n
β,λ+n/p (R ),
that is, there exists a constant C > 0 such that
(i) when n/α < p < ∞,

I˜α,d f Lip(d) ≤ Cf CMOp,λ , f ∈ CMOp,λ (Rn );

β,λ+n/p

(ii) when p = n/α, i.e., β = 0,

I˜α,d f BMO(d) ≤ Cf CMOn/α,λ , f ∈ CMOn/α,λ (Rn ).

λ+α

In Theorem 3.6, if d = 0, then we can obtain the following estimates for a

modified Riesz potential I˜α on CMOp,λ (Rn ).
Corollary 3.7 Let 0 < α < 1, n/α ≤ p < ∞, −n/p + α ≤ λ + α < 1 and
β = α − n/p. Then I˜α,d is bounded from CMOp,λ (Rn ) to Lipβ,λ+n/p (Rn ), that is,
there exists a constant C > 0 such that
(i) when n/α < p < ∞,

I˜α,d f Lipβ,λ+n/p ≤ Cf CMOp,λ , f ∈ CMOp,λ (Rn );

(ii) when p = n/α, i.e., β = 0,

I˜α,d f BMOλ+α ≤ Cf CMOn/α,λ , f ∈ CMOn/α,λ (Rn ).

d-Modified Riesz Potentials on Central Campanato Spaces 433

4 Proofs of Theorems

First of all, we state the following well-definedness of I˜α,d for CMOp,λ (Rn ), which
is shown by the same argument as in the proof of Theorem 3.6 of [16].
Lemma 4.1 Let 0 < α < 1, 1 ≤ p < ∞, d ∈ N0 and −n/p + α ≤ λ + α < d + 1.
Then for f ∈ CMOp,λ (Rn ), I˜α,d f is well-defined.
Proof Let f ∈ CMOp,λ (Rn ), r ≥ 1 and x ∈ Qr , and let

I˜α,d f (x) = I˜α,d (f˜χQ2r )(x) + I˜α,d (f˜(1 − χQ2r ))(x) + fQ4r I˜α,d 1(x)
 xl 
˜
= Iα (f χQ2r )(x) − f˜(y)(D l Kα )(−y) dy
l! Q2r \Q1
{l:|l|≤d}
⎛ ⎞
  xl
+ f˜(y) ⎝Kα (x − y) − (D l Kα )(−y)⎠ dy
Rn \Q2r l!
{l:|l|≤d}

+ fQ4r I˜α,d 1(x), (4.1)

where f˜ = f − fQ4r . Then, since f˜χQ2r ∈ Lp (Rn ), the first term is well-defined.
The second term is also well-defined, because (D l Kα )(χQ2r − χQ1 ) ∈ Lp (Rn ).
Here we note that the second term is a polynomial of degree at most d. For the
third term, the integral converges absolutely in virtue of Lemmas
= 4.2 and 4.4, which
are shown later, and so the present term is well-defined. As Rn \Q1 (D l Kα )(−y)dy
converges absolutely under the assumption 0 < α < 1, I˜α,d 1 ∈ P d (Qr ), and then
the forth term is well-defined.
Further, if we let for 1 ≤ s < r,

fQ4r − fQ4s = (f − fQ4s )χQ2s − (f − fQ4r )χQ2r

+ (f − fQ4s )(1 − χQ2s ) − (f − fQ4r )(1 − χQ2r ),

then it follows that for x ∈ Qs ⊂ Qr ,

0 = I˜α,d (fQ4r − fQ4s )(x)

= I˜α,d ((f − fQ4s )χQ2s )(x) − I˜α,d ((f − fQ4r )χQ2r )(x)
+ I˜α,d ((f − fQ4s )(1 − χQ2s ))(x) − I˜α,d ((f − fQ4r )(1 − χQ2r ))(x).

Therefore, I˜α,d f is independent of Qr containing x, and so I˜α,d f is well-defined

on Rn .
In order to prove Theorems 3.2, 3.3 and 3.6, it is necessary to use the following
three lemmas.
434 K. Matsuoka

Lemma 4.2 (Lemma 7.3 of [21]) Let x ∈ Rn , 0 < α < n and d ∈ N0 . If y ∈

Rn \ Q2|x| , then

 xl |x|d+1
Kα (x − y) − (D l Kα )(−y) ≤ C n−α+d+1 . (4.2)
l! |y|
{l:|l|≤d}

Here note that let Q2|x| = {0}, when x = 0.

Lemma 4.3 (Lemma 4.2 of [18]) Let x ∈ Rn , 0 < α < n and d ∈ N0 . If y ∈
Rn \ Q2|x| , then
⎛ ⎞
 xl |h|d+1
Ad+1 ⎝Kα (x − y) − (D l Kα )(−y)⎠ ≤ C n−α+d+1 . (4.3)
h
l! |y|
{l:|l|≤d}

Lemma 4.4 (Lemma 4.1 of [16]; cf. Lemma 4.2 of [19]) Let 1 ≤ p < ∞ and
λ ∈ R. If β < 0 and β + λ < 0, then there exists a positive constant C such that

|f (y) − fQ2r |
dy ≤ Cr β+λ f CMOp,λ
Rn \Q r
|y|n−β
for all f ∈ CMOp,λ (Rn ) and r ≥ 1.

Proof of Theorem 3.2 Let f ∈ CMOp,λ (Rn ), r ≥ 1 and x ∈ Qr . As I˜α,d f is

well-defined by Lemma 4.1, hence we prove only that

I˜α,d f $(d)  f CMOp,λ ,

q,μ

by the same argument as in the proof of Theorem 3.1 in [17].

Now, in (4.1), let us put

R̃rd f (x) = Rrd f˜(x) + fQ4r I˜α,d 1(x) ∈ P d (Qr ),

where
 xl 
Rrd f˜(x) = − f˜(y)(D l Kα )(−y) dy,
l! Q2r \Q1
{l:|l|≤d}

and
⎛ ⎞
  xl
J˜α,d,r f (x) = f˜(y) ⎝Kα (x − y) − (D l Kα )(−y)⎠ dy.
Rn \Q2r l!
{l:|l|≤d}
d-Modified Riesz Potentials on Central Campanato Spaces 435

Then we have
 1/q
|I˜α,d f (x) − R̃rd f (x)|q dx
Qr
 1/q  1/q
≤ |Iα (f˜χQ2r )(x)|q dx + |J˜α,d,r f (x)|q dx
Qr Qr

=: I1 + I2 . (4.4)

First we estimate I1 . By applying the strong (p, q) boundedness of Iα , we have

I1 ≤ Iα (f˜χQ2r )Lq  f˜χQ2r Lp  r λ |Q2r |1/p f˜B p,λ

∼ r λ+n/p f CMOp,λ = r μ+n/q f CMOp,λ .

Next, in order to estimate I2 , using (4.2), it follows that for x ∈ Qr and y ∈

Rn \ Q2r ,

 xl |x|d+1 r d+1
Kα (x − y) − (D l Kα )(−y)  ≤ .
l! |y|n−α+d+1 |y|n−α+d+1
{l:|l|≤d}

Consequently, we obtain by Lemma 4.4 and the assumptions 0 < α < 1 and μ+σ <
d + 1,

|f˜(y)|
|J˜α,d,r f (x)|  r d+1 dy  r λ+α f CMOp,λ
Rn \Q2r |y|n−α+d+1
= r μ f CMOp,λ , (4.5)

which shows that J˜α,d,r f is well-defined for all x ∈ Qr . Thus we get

I2 = J˜α,d,r f Lq (Qr )  r μ f CMOp,λ · |Qr |1/q ∼ r μ+n/q f CMOp,λ ,

and then
 1/q
1 q
I˜α,d f $(d)  sup μ − I˜α,d f (y) − R̃rd f (y) dy
q,μ
r≥1 r Qr
 1/q
1 1
 sup · r μ+n/q f CMOp,λ ∼ f CMOp,λ .
r≥1 rμ |Qr |

This concludes the proof.

436 K. Matsuoka

Proof of Theorem 3.3 This proof is similar to that of Theorem 3.6 of [16]. Hence,
in the same way as (4.4), we have that for f ∈ CMO1,λ (Rn ), r ≥ 1 and x ∈ Qr ,
 
sup(2t)q x ∈ Qr : |I˜α,d f (x) − R̃rd f (x)| > 2t
t >0
  
≤ 2q sup t q x ∈ Qr : |Iα (f˜χQ2r )(x)| > t
t >0

+ sup t q |{x ∈ Qr : |J˜α,d,r f (x)| > t}|

t >0

=: 2 (I3 + I4 ) .
q

Then it follows from the weak (1, q) boundedness of Iα that

1/q
I3  r λ+n f CMO1,λ = r μ+n/q f CMO1,λ ,

and by applying (4.5) with p = 1,

1/q
I4  r μ+n/q f CMO1,λ .

Thus, we obtain

I˜α,d f W $(d)  f CMO1,λ ,

q,μ

which completes the proof.

Proof of Theorem 3.6 Let f ∈ CMOp,λ (Rn ), r ≥ 1 and x ∈ Qr . Similarly to the
proof of Theorem 3.2, we prove only that for n/α ≤ p < ∞,

I˜α,d f Lip(d) (Q )  r λ+n/p f CMOp,λ ,

β r

following similar arguments to the proof of Theorem 3.2 in [18].

Now, using the notation in the proof of Theorem 3.2, it follows from
Remarks 2.8, 2.10 and (4.1) with

 
xl
Rrd f˜(x) = − f˜(y)(D l Kα )(−y) dy
l! Q2r \Q1
{l:1≤|l|≤d}
d-Modified Riesz Potentials on Central Campanato Spaces 437

that

I˜α,d f Lip(d) (Q ) ∼ I˜α,d f L(d) (Q

β r 1,β r)
 
1 ˜α,d f (y) − P (y)| dy
∼ sup infβ
− | I
Q(x,s)⊂Qr P ∈P d (Qr ) s Q(x,s)
 
1 ˜α,d f (y) − R̃rd f (y)| dy
≤ sup β
− | I
Q(x,s)⊂Qr s Q(x,s)

≤ I˜α (f˜χQ2r )Lipβ (Qr ) + J˜α,d,r f L(d) (Q

1,β r)

=: I5 + I6 .

We firstly estimate I5 . When n/α < p < ∞, which implies 0 < β < 1, if we
apply the (Lp , Lipβ ) boundedness of I˜α , then we get

I5 ≤ I˜α (f˜χQ2r )Lipβ  f˜χQ2r Lp  r λ+n/p f˜B p,λ

= r λ+n/p f CMOp,λ .

Similarly, when p = n/α, by using the (Ln/α , BMO) boundedness of I˜α , we obtain

I5 ≤ I˜α (f˜χQ2r )BMO  r λ+α f CMOn/α,λ .

Next in order to estimate I6 , since by Remark 2.8,

1
I6 ∼ J˜α,d,r f Lip(d) (Q ) = sup |Ad+1 ˜
h Jα,d,r f (x)|,
β r
x,x+h∈Qr ,h=0 |h| β

we estimate Ad+1 ˜
h Jα,d,r f (x). To do so, if we use (4.3), Lemma 4.4 and the
assumptions 0 < α < 1 and λ + α < d + 1, then we have for x ∈ Qr and
y ∈ Rn \ Q2r ,

|Ad+1 ˜
h Jα,d,r f (x)|
⎧ ⎛ ⎞⎫
 ⎨  xl ⎬
= f˜(y) Ad+1 ⎝Kα (x − y) − (D l
Kα )(−y) ⎠ dy
Rn \Q2r ⎩ h l! ⎭
{l:|l|≤d}

|f˜(y)|
 |h|d+1 dy  |h|d+1 r α−d−1+λf CMOp,λ .
Rn \Q2r |y|n−α+d+1
438 K. Matsuoka

Hence
1
I6  sup · |h|d+1 r α−d−1+λf CMOp,λ
x,x+h∈Qr ,h=0 |h| β

≤ r λ+α−β f CMOp,λ = r λ+n/p f CMOp,λ .

Thus it follows that for n/α ≤ p < ∞,

1
I˜α,d f Lip(d) = sup I˜α,d f Lip(d) (Q )  f CMOp,λ .
β,λ+n/p r≥1 r λ+n/p β r

This shows the conclusion.

Acknowledgments The author would like to express his deep gratitude to the anonymous referees
for their careful reading and fruitful comments.

References

1. J. Alvarez, M. Guzmán-Partida, J. Lakey, Spaces of bounded λ-central mean oscillation,

Morrey spaces, and λ-central Carleson measures. Collect. Math. 51, 1–47 (2000)
2. Y. Chen, K. Lau, Some new classes of Hardy spaces. J. Funct. Anal. 84, 255–278 (1989)
3. H. Feichtinger, An elementary approach to Wiener’s third Tauberian theorem on Euclidean
n-space, in Proceedings, Conference at Cortona 1984, Symposia Mathematica, vol. 29
(Academic Press, New York, 1987), pp. 267–301
4. Z. Fu, Y. Lin, S. Lu, λ-central BMO estimates for commutators of singular integral operators
with rough kernels. Acta Math. Sin. (Engl. Ser.) 24, 373–386 (2008)
5. J. García-Cuerva, Hardy spaces and Beurling algebras. J. Lond. Math. Soc. 39, 499–513 (1989)
6. J. García-Cuerva, M.J.L. Herrero, A theory of Hardy spaces associated to the Herz spaces.
Proc. Lon. Math. Soc. 69, 605–628 (1994)
7. G.H. Hardy, J.E. Littlewood, Some properties of fractional integrals. I. Math. Z. 27, 565–606
(1928); II, ibid. 34, 403–439 (1932)
8. C. Herz, Lipschitz spaces and Bernstein’s theorem on absolutely convergent Fourier trans-
forms. J. Math. Mech. 18, 283–324 (1968)
9. Y. Komori-Furuya, K. Matsuoka, Some weak-type estimates for singular integral operators on
CMO spaces. Hokkaido Math. J. 39, 115–126 (2010)
10. Y. Komori-Furuya, K. Matsuoka, Strong and weak estimates for fractional integral operators
on some Herz-type function spaces, in Proceedings of the Maratea Conference FAAT 2009,
Rendiconti del Circolo Mathematico di Palermo, Serie II, Supplementary, vol. 82 (2010), pp.
375–385
11. Y. Komori-Furuya, K. Matsuoka, E. Nakai, Y. Sawano, Integral operators on Bσ -Morrey–
Campanato spaces. Rev. Mat. Complut. 26, 1–32 (2013)
12. Y. Komori-Furuya, K. Matsuoka, E. Nakai, Y. Sawano, Applications of Littlewood–Paley
theory for Ḃσ -Morrey spaces to the boundedness of integral operators. J. Funct. Spaces Appl.
2013, 859402 (2013)
13. T. Kurokawa, Riesz potentials, higher Riesz transforms and Beppo Levi spaces. Hiroshima
Math. J. 18, 541–597 (1988)
d-Modified Riesz Potentials on Central Campanato Spaces 439

14. T. Kurokawa, Weighted norm inequalities for Riesz potentials. Jpn. J. Math. 14, 261–274
(1988)
15. S. Lu, D. Yang, The central BMO spaces and Littlewood–Paley operators. Approx. Theory
Appl. (N.S.) 11, 72–94 (1995)
16. K. Matsuoka, Bσ -Morrey–Campanato estimates and some estimates for singular integrals on
central Morrey spaces and λ-CMO spaces, in Banach and Function Spaces IV (Kitakyushu
2012) (Yokohama Publishers, Yokohama, 2014), pp. 325–335
17. K. Matsuoka, Generalized fractional integrals on central Morrey spaces and generalized λ-
CMO spaces, in Function Spaces X, Banach Center Publications, vol. 102 (Institute of
Mathematics, Polish Academy of Sciences, Warsawa, 2014), pp. 181–188
18. K. Matsuoka, Generalized fractional integrals on central Morrey spaces and generalized σ -
Lipschitz spaces, in Current Trends in Analysis and its Applications: Proceedings of the
9th ISAAC Congress, Kraków 2013. Springer Proceedings in Mathematics and Statistics
(Birkhäuser, Basel, 2015), pp. 179–189
19. K. Matsuoka, E. Nakai, Fractional integral operators on B p,λ with Morrey–Campanato norms,
in Function Spaces IX, Banach Center Publishers, vol. 92 (Institute of Mathematics, Polish
Academy of Sciences, Warsawa, 2011), pp. 249–264
20. Y. Mizuta, On the behaviour at infinity of superharmonic functions. J. Lond. Math. Soc. 27,
97–105 (1983)
21. Y. Mizuta, Potential Theory in Euclidean Spaces (Gakkōtosho, Tokyo, 1996)
22. E. Nakai, Y. Sawano, Hardy spaces with variable exponents and generalized Campanato spaces.
J. Funct. Anal. 262, 3665–3748 (2012)
23. J. Peetre, On the theory of Lp,λ spaces. J. Funct. Anal. 4, 71–87 (1969)
24. A. Pietsch, History of Banach Spaces and Linear Operators (Birkhäuser, Boston, 2007)
25. S.L. Sobolev, On a theorem in functional analysis. Mat. Sbornik 4, 471–497 (1938) (in Russian)
26. A. Zygmund, On a theorem of Marcinkiewicz concerning interpolation of operations. J. Math.
Pures Appl. 35, 223–248 (1956)
On Some Consequences of the Solvability
of the Caffarelli–Silvestre Extension
Problem

Jan Meichsner and Christian Seifert

Abstract We consider the Caffarelli–Silvestre extension problem, i.e., a Bessel

type ODE in a Banach space X with a closed and typically unbounded operator A as
right-hand side and point out a couple of consequences arising from the assumption
of the well-posedness of the problem. In the end a conjecture is stated concerning
the implications of analyticity of the solution of the extension problem.

Keywords Fractional powers · Non-negative operator · Dirichlet-to-Neumann

operator

Mathematics Subject Classification (2010) Primary 47A05; Secondary 47D06,

47A60

1 Introduction

Since the well-known Caffarelli–Silvestre paper [4] from 2007 various authors
considered an abstract version of the there presented problem by having a look on
the abstract (incomplete) ODE problem

1 − 2α
u (t) + u (t) = Au(t) (t > 0), u(0) = x (1.1)
t

J. Meichsner ()
Technische Universität Hamburg, Institut für Mathematik, Hamburg, Germany
e-mail: [email protected]
C. Seifert
Technische Universität Hamburg, Institut für Mathematik, Hamburg, Germany
Technische Universität Clausthal, Institut für Mathematik, Clausthal-Zellerfeld, Germany
e-mail: [email protected]; [email protected]

© Springer Nature Switzerland AG 2021 441

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_22
442 J. Meichsner and C. Seifert

in some Hilbert or more general Banach space X, see [2, 5, 8, 9]. The abstract
setting made it necessary to introduce new tools and viewpoints completely different
from the techniques used in [4] for the rather concrete problem considered there.
This development was first started by Stinga–Torrea in [9] by the introduction
of explicit formulas for solutions of (1.1) and further continued by Galé–Miana–
Stinga in [5]. Typically, the operator A appearing on the right-hand side of (1.1)
is assumed to be sectorial (with or without dense domain). The basic question
of interest is to establish existence and uniqueness results. For the special choice
of α = 1/2 the singular first order term in (1.1) vanishes and the problem
reduces to an incomplete second order Cauchy problem as already considered by
Balakrishnan in [3, Theorem 6.1], where the well-posedness for a sectorial operator
A is shown under the additional assumption of a globally bounded solution u
of (1.1). Concerning generalisations of Balakrishnan’s result, in [9] the authors
considered the situation for a self-adjoint, positive, so in particular sectorial,
operator on a Hilbert space. The authors derived two explicit formulas for a solution
to (1.1) including the Stinga-Torrea formula, a Poisson formula for the solution.
The uniqueness of the solution is also shown under the assumption of A having
a pure point spectrum. The corresponding Banach space setting was considered
in [5] where the authors constructed explicitly solutions for both, generators of
tempered integrated semigroups and tempered integrated cosine families. These
results particularly cover generators of bounded C0 -semigroups. Combined with
the uniqueness result from [8], (1.1) is a well-posed problem and this fact actually
extends to arbitrary sectorial right-hand sides. Given an initial condition for the
first derivative one can also tackle the problem by cosine functions, see e.g. [1,
Chapter 3]. The connection between the various results was explored a bit in [8],
where one can see that the boundedness assumption appearing in Balakrishnan’s
paper ‘couples’ the two initial values and thus is suitable to replace one of them.
In [6], the authors consider the problem from a slightly different point of view.
They assume the ODE (1.1) to be given for α = 1/2 but merely assume A to be
closed with non-empty resolvent set. As a consequence of the well-posedness of
the problem, the existence of a generator B of a C0 -semigroup is deduced which
satisfies B 2 = A. Further, analyticity of √ the solution u of (1.1) on a sector is shown
to imply sectoriality of A and then B = A in the functional calculus sense.
This is the starting point of this paper. More precisely, we attempt to state the
right setting for proving similar results as explained above in the more general case
α ∈ (0, 1). Making use of similar strategies as in [6, Theorem 6.3.3], generalisations
of the first result on factorizing A (as B 2 ) can be obtained; cf. Corollary 2.13. For the
second result that analyticity of the solution u of (1.1) on a sector yields sectoriality
of the operator A, we formulate a conjecture.
Consequences of Solvability of the Extension Problem 443

2 The Solution Operator to the Caffarelli–Silvestre

Extension Problem

In this paper X will always denote a complex Banach space with norm ·. Further
let A be a closed linear operator in X with non-empty resolvent set ρ(A) and with
dense domain D(A). We will interpret D(A) as a Banach space equipped with the
graph norm

xD(A) := x + Ax (x ∈ D(A)).

Moreover, let α ∈ (0, 1) be a parameter. 

For x ∈ X let Pα (x) be the problem of finding u ∈ Cb [0, ∞); D(A) ∩

C 2 (0, ∞); X such that u(0) = x and

1 − 2α
u (t) + u (t) = Au(t) (t > 0).
t
For the whole paper, we may assume as a standing assumption that Pα (x) has a
unique solution, denoted by ux , for every x ∈ D(A). This fact will not be mentioned
explicitly anymore but the reader should be always aware of it.
Remark 2.1 As noted in the introduction, there are various cases where Pα (x) has
a unique solution for all x ∈ D(A), e.g., when A is the generator of a bounded C0 -
semigroup ([5, Theorem 2.1]) or when A is the generator of a tempered integrated
cosine family ([5, Theorem 1.3] combined with [8, Theorem 5.8]).
For x ∈ D(A) and t ≥ 0 we may then define U x(t) := ux (t).

Lemma 2.2 We have U ∈ L D(A), Cb [0, ∞); D(A) .
Proof The proof essentially works as the one for the special case α = 1/2 as
presented in [6, Theorem 6.3.3] with the help of the closed graph theorem. So let
(xn ) be a sequence in D(A) which converges w.r.t. ·D(A) towards  x ∈ D(A).
Further, assume that (U xn ) is convergent in Cb [0, ∞); D(A) towards some
element v. Let n, m ∈ N. From the fact that U xn and U xm solve Pα (xn ) and Pα (xm ),
respectively, we get (for t ≥ 0)
; ;
; 2α−1 d 1−2α d 2α−1 d 1−2α d
;
;t t U xn (t) − t t U xm (t);
; dt dt dt dt ;

= AU xn (t) − AU xm (t) ≤ U xn − U xm Cb ([0,∞);D(A)) ,


which implies the existenceof some function u2 ∈ Cb [0, ∞); X such that
 
t → t 2α−1 dtd t 1−2α dtd U xn (t) converges in Cb [0, ∞); X towards u2 . Multiply-
ing the ODE (at t1 > 0) for the initial condition xk by t11−2α , and integrating the
444 J. Meichsner and C. Seifert

resulting term from t1 to t2 > 0 results in

t2
d d
t11−2α U xk (t1 ) − t21−2α U xk (t2 ) = s 1−2α AU xk (s)ds.
dt dt
t1

By the boundedness of s →  AU xk (s) and the integrability of s → s 1−2α we can

send t1 to 0. Hence, for k ∈ N, we can define
d
yk := lim t 1−2α U xk (t).
t →0+ dt
Let t > 0. Multiplying this time the difference of the ODE’s (at s > 0) for initial
conditions xn and xm by s 1−2α , and integrating the term from 0 to t results in

d d
t 1−2α U xn (t) − t 1−2α U xm (t)
dt dt
t
= yn − ym + s 1−2α A(U xn − U xm )(s)ds. (2.1)
0

Now, we multiply (2.1) by t 2α−1 , integrate again, take norms and solve for
yn − ym . Then we obtain

t 2α
yn − ym 
2α
; ;
; t s ;
; ;
;
= ;U (xn − xm )(t) − (xn − xm ) − s 2α−1
r 1−2α
AU (xn − xm )(r)drds ;
;
; ;
0 0
 
t2
≤ 2+ U xn − U xm Cb ([0,∞);D(A)) .
4 − 4α

Hence, (yn ) is a Cauchy sequence in X.

By (2.1), we estimate
; ;
; 1−2α d 1−2α d
;
;t U x (t) − t U x (t) ;
; dt
n
dt
m ;
; ;
; t ;
; ;
;
= ; yn − ym + s 1−2α
A(U xn − U xm )(s)ds ;
;
; ;
0

t 2−2α
≤ yn − ym  + U xn − U xm Cb ([0,∞);D(A)) .
2 − 2α
Consequences of Solvability of the Extension Problem 445


Therefore, there exists u1 ∈ C (0, ∞); X such that

d  
t → t 1−2αU xn (t) → t → t 1−2α u1 (t)
dt

uniformly on compacts in [0, ∞). Thus, v ∈ C 1 (0, ∞); X .
Furthermore, for t > 0 we have
; 2 ;
;d d2 ;
; U x (t) − U x (t) ;
; dt 2 n dt 2
m ;
; ;
; 2α−1 d 1−2α d 2α−1 d 1−2α d
;
≤; ; t t U x n (t) − t t U x m (t) ;
;
dt dt dt dt
; ;
|1 − 2α| ; ; d U xn (t) − d U xm (t);
;
+ ; ;
t dt dt
→ 0,

uniformly for t in a compact set K ⊂ (0, ∞). Therefore, v ∈ C 2 (0, ∞); X . Since

v(0) = lim (U xn )(0) = lim xn = x,

n→∞ n→∞

 v = U x. Now, the closed graph

by uniqueness of the solution of Pα (x) we obtain
theorem yields U ∈ L D(A), Cb [0, ∞); D(A) .
As a corollary of the previous lemma we obtain continuity properties of certain
derivatives of U x.
Corollary 2.3 For x ∈ D(A) we have
 d  
t → t 1−2α U x(t) ∈ C [0, ∞); X
dt
and
 d d  
t → t 2α−1 t 1−2α U x(t) ∈ Cb [0, ∞); X .
dt dt
Let λ ∈ ρ(A). As a consequence of the assumed uniqueness one can prove the
equations U (λ − A)−1 x = (λ − A)−1 U x for x ∈ D(A) and U Ax = AU x for
x ∈ D(A2 ).
446 J. Meichsner and C. Seifert


For x ∈ X let PLα (x) be the problem of finding u ∈ Cb [0, ∞); X with u(0) = x

and such that for all φ ∈ Cc∞ (0, ∞) we have

∞ ∞ ∞
d 1−2α d 2α−1
u(t)φ(t)dt ∈ D(A) and u(t) t t φ(t)dt = A u(t)φ(t)dt.
dt dt
0 0 0

Lα (x) as a weak version of Pα (x).

We can interpret P
Lemma 2.4 The mapping U interpreted
 as a function from D(A) ⊆ X to
Cb [0, ∞); D(A) ⊆ Cb [0, ∞); X admits a continuous extension  to all of X,
again denoted by U . For x ∈ X the function U x ∈ Cb [0, ∞); X is the unique
Lα (x).
solution of P
Proof Let x ∈ D(A). By Lemma 2.2 there exist C1 ≥ 0 such that

U xCb ([0,∞);X)
; ;
; ;
= ;(λ − A)U (λ − A)−1 x ;
Cb ([0,∞);X)
; ; ; ; 
; ; ; ;
≤ (|λ| + 1) ;U (λ − A)−1 x ; + ;AU (λ − A)−1 x ;
Cb ([0,∞);X) Cb ([0,∞);X)
; ;
; ;
≤ C1 ;(λ − A)−1 x ; ≤ C2 x
D (A)

for some C2 > 0. Hence, by density of D(A) in X, U can be continuously extended

to the whole of X. We denote the extension again by U . For given x ∈ X let (xn )
be a sequence in D(A) convergent
 towards x in X and φ ∈ Cc∞ (0, ∞) . Then
U xn → U x in Cb [0, ∞); X . Hence,

∞ ∞
d d d 1−2α d 2α−1
U x(t) t 1−2α t 2α−1 φ(t)dt = lim U xn (t) t t φ(t)dt
dt dt n→∞ dt dt
0 0
∞
d 1−2α d
= lim t 2α−1 t U xn (t)φ(t)dt
n→∞ dt dt
0
∞
= lim AU xn (t)φ(t)dt
n→∞
0
∞
= lim A U xn (t)φ(t)dt,
n→∞
0
Consequences of Solvability of the Extension Problem 447

by Hille’s theorem. Since

∞ ∞
U xn (t)φ(t)dt → U x(t)φ(t)dt,
0 0

by closedness of A we observe

∞
U x(t)φ(t)dt ∈ D(A)
0

and
∞ ∞
d d
U x(t) t 1−2α t 2α−1 φ(t)dt = A U x(t)φ(t)dt.
dt dt
0 0

Also

U x(0) = lim U xn (0) = lim xn = x.

n→∞ n→∞

Lα (x) as claimed.
So U x is a solution to P
Lα (x) is unique. Indeed, if u is a solution to P
Note that the solution of P Lα (0) then
v := (λ − A)−1 u is a solution to Pα (0), hence v = 0 and, therefore, u = 0.
The uniqueness of a solution to P Lα (x) yields the extension of the already found
relationships between U , A and its resolvent (λ−A)−1 to U (λ−A)−1 = (λ−A)−1 U
and AU x = U Ax for x ∈ D(A).
Remark 2.5 In view of Lemma 2.4, the authors came to the conclusion that defining
a solution to the considered problem in [8] would have been better in the way as in
Lα (x).
P
Next we show that the smoothness of u and its scaled derivatives tells us precisely
about the smoothness of the initial datum x.

Definition 2.6 We define the operator Dα in Cb [0, ∞); X by
  
D(Dα ) := u ∈ Cb [0, ∞); X ∩ C 1 (0, ∞); X |
 
t → t 1−2α u (t) ∈ Cb [0, ∞); X ,

Dα u := t → t 1−2α u (t) .

The operator Dα acts as a scaled first derivative. Note that Dα is closed.

448 J. Meichsner and C. Seifert

Lemma 2.7 Let k ∈ N0 , x ∈ X. Then x ∈ D(Ak ) if and only if


U x ∈ D (D1−α Dα )k .

Proof For k = 0 there is nothing to prove. Therefore, let k ≥ 1 and x ∈ D(Ak ).

Consider first k = 1. Then U x ∈ D(D1−α Dα ) by Corollary 2.3. If k ≥ 2 we note
that

D1−α Dα U x = AU x = U Ax

by (1.1) and commutativity of A and U on D(A). Since Ax ∈ D(A), arguing

inductively shows U x ∈ D (D1−α Dα )k and

(D1−α Dα )k U x = U Ak x.

Conversely, let k ≥ 1 and U x ∈ D (D1−α Dα )k . Consider again
 first the case
k = 1 and choose for s > 0 a sequence (φk ) in Cc∞ (0, ∞) which converges

Lα (x). Hence,
towards δs in Cc (0, ∞) . The function U x solves P

∞ ∞
D1−α Dα U x(t)φk (t)dt = A U x(t)φk (t)dt
0 0

which, by closedness of A, gives D1−α Dα U x(s) = AU x(s) as k → ∞. Further,

U x(s) → x as s → 0+. Using the closedness of A again yields x ∈ D(A) and
D1−α Dα U x = AU x = U Ax. Now one can argue again inductively.
Let us define Bα : D(A) → X by

d
Bα x := lim t 1−2α U x(t).
t →0+ dt

Note that Bα is the ‘generalized Dirichlet-to-Neumann operator’ associated to the

problem Pα (x). Since (λ − A)−1 U = U (λ − A)−1 , we obtain

Bα (λ − A)−1 x = (λ − A)−1 Bα x

for x ∈ D(A). From the proof of Lemma 2.2 one can see the existence of a constant
C > 0, depending on a fixed parameter T ∈ [0, ∞) (let us take T = 1, say), such
that Bα x ≤ C xD(A) . Commutativity with the resolvent of A and continuity
with respect to the graph norm imply that Bα is closable as an operator in X with
domain D(A) ⊆ X. To see this, consider (xn ) in D(A) convergent towards 0 in X
Consequences of Solvability of the Extension Problem 449

and such that Bα xn → y in X. Then

; ; ; ; ; ;
; ; ; ; ; ;
;(λ − A)−1 y ; = lim ;Bα (λ − A)−1 xn ; ≤ C · lim ;(λ − A)−1 xn ; = 0,
n→∞ n→∞ D (A)

which shows the closability since the above inequality implies y = 0. Let us denote
the closure of Bα in X again by Bα . The next statement will be an auxiliary lemma
for complex-valued functions.
 
Lemma 2.8 Let r ∈ [0, 1), u ∈ Cb [0, ∞) ∩ C 2 (0, ∞) such that t →
   
t r dtd t −r u (t) ∈ Cb [0, ∞) . Then also t → t −r u (t) ∈ Cb [0, ∞) .
Proof For r = 0 this is a standard result. The integrability of u near t = 0 implies
the continuity of u at t = 0 while the boundedness of u for large values of t
follows from Taylor’s theorem. For the general case consider the function v given
1   
by v(s) := u s 1+r . By assumption v ∈ Cb [0, ∞) ∩ C 2 (0, ∞) and with the
1
relationship t := s 1+r we can write v (s) = u (t)t −r and

d −r
v (s) = t −r t u (t).
dt
The claim follows now by applying the result for r = 0 to the function v.
Corollary 2.9 Let x ∈ D(A) and α ∈ [1/2, 1). Then

d  
t → t 1−2α U x(t) ∈ Cb [0, ∞); X .
dt

Proof Let r := 2α − 1 and x ∈ X . Then Lemma 2.8 applied to t →

x , U x(t)X ×X yields boundedness of this function. Hence, U x is weakly bounded
and therefore bounded.
Let us now assume that, additionally to unique solvability of Pα (x) for x ∈ D(A),
the ‘conjugated problem’ P1−α (x) is also uniquely solvable for all x ∈ D(A). We
adapt all so far introduced notations by simply changing α to 1 − α and specify
in the following the solution operators by indices, i.e., Uα x will be the solution to
Pα (x) while U1−α x solves P1−α (x). With this notation we can now state and prove
the main theorem of this paper.
Theorem 2.10 Let x ∈ D(A2 ). Then t 1−2α dtd Uα x(t) = U1−α Bα x(t) for all t ∈
[0, ∞).
For t = 0 we interpret the left-hand side of the equation in Theorem 2.10 as the
limit from the right.
450 J. Meichsner and C. Seifert

Proof For the proof we will distinguish two cases for α.

(i) Let us begin with α ∈ [1/2, 1). Define

d 
v := t → t 1−2α
Uα x(t) .
dt

By Corollary 2.9 we have v ∈ Cb [0, ∞); X and by the definition of Bα we
have v(0) = Bα x. Let us finally check that v solves the ODE for P1−α (Bα x).
Then v = U1−α Bα x. Let t > 0. Then

d d d
t 2α−1 v(t) = t 2α−1 t 1−2α Uα x(t) = AUα x(t) = Uα Ax(t)
dt dt dt
and
d 2α−1 d d
t 1−2α t v(t) = t 1−2α Uα Ax(t) = Av(t),
dt dt dt
where we used the closedness of A for the last equality.
(ii) Let us turn to the case α ∈ (0, 1/2). In this case, consider the function ṽ given
by

ṽ(t) := t 2α−1 v (t) = Uα Ax(t) (t ≥ 0),

 
which belongs to Cb [0, ∞); D(A) ⊆ Cb [0, ∞); X and which solves
Lα (Ax). Further let w be defined by
Pα (Ax). In particular, ṽ solves P

w(t) := U1−α Bα x(t) (t ≥ 0).

Let λ ∈ ρ(A). Since x ∈ D(A2 ) there exists y ∈ D(A) such that x = (λ −

A)−1 y. It follows that

Bα x = Bα (λ − A)−1 y = (λ − A)−1 Bα y ∈ D(A).

This implies that

 
w̃ := t → t 2α−1 w (t) ∈ Cb [0, ∞); X ,

where for the boundedness of w̃ we used Lemma 2.8 with r := 1 − 2α. Note
that w̃(0) = B1−α Bα x. If we apply (i) to 1 − α instead of α we find

d 2α−1 d d
t 1−2α t U1−α x(t) = U1−α Ax(t) = t 1−2α Uα B1−α x(t).
dt dt dt
Consequences of Solvability of the Extension Problem 451

Sending t → 0+ we conclude Ax = Bα B1−α x. Let us show that the

operators Bα and B1−α commute on D(A2 ). Note that for y ∈ D(A) we have
B1−α (λ − A)−1 y = (λ − A)−1 B1−α y (by uniqueness of solutions). Hence,
we get B1−α Ax = AB1−α x for x ∈ D(A2 ). This implies that the bounded
operator B1−α (λ − A)−1 commutes on D(A) with A and this in turn implies
the commutativity with Bα . So finally

Bα B1−α (λ − A)−1 y = B1−α (λ − A)−1 Bα y

= B1−α Bα (λ − A)−1 y = B1−α Bα x.

L
 w̃(0) = Ax. Finally, we show that w̃ solves Pα (Ax).
Therefore, we observe
Let φ ∈ Cc∞ (0, ∞) . Since w solves the ODE in (1.1) we obtain

∞ ∞
d d d 2α−1
w̃(t) t 1−2α t 2α−1 φ(t)dt = − AU1−α Bα x(t) t φ(t)dt.
dt dt dt
0 0

If we apply Hille’s theorem to the operator A it follows that

∞ ∞
d 2α−1
w̃(t)φ(t)dt = − U1−α Bα x(t) t φ(t)dt ∈ D(A)
dt
0 0

and
∞ ∞
d
− AU1−α Bα x(t) t 2α−1 φ(t)dt = A w̃(t)φ(t)dt.
dt
0 0

Lα (Ax), and therefore

We conclude w̃ = ṽ by uniqueness of the solution to P
w = v . Since w(0) = v(0) it follows that w = v.
Remark 2.11 On the first glance it seems awkward that the existence of a solution
to Pα (x) may not imply the existence of a solution to P1−α (x), but meanwhile
the authors are convinced that this is may indeed be the case, although a concrete
example for this is missing. Note that in the case α = 1/2 both problems are
identical which is why in this case only solvability of one problem is needed to
prove the result.
As a corollary to Theorem 2.10 we obtain a result similar to the one of
Lemma 2.7.
Corollary 2.12 Let x ∈ X. Then x ∈ D(Bα ) if and only if Uα x ∈ D(Dα ). In either
case, Dα Uα x = U1−α Bα x.
452 J. Meichsner and C. Seifert

Proof Let x ∈ D(Bα ). Then there exists a sequence (xn ) in D(A) which is
convergent towards x in X such that Bα xn → Bα x. By Lemma 2.4, we have
Uα xn → Uα x and by Theorem 2.10 we obtain

Dα Uα xn = U1−α Bα xn → U1−α Bα x.

Since Dα is closed we get Uα x ∈ D(Dα ) and Dα Uα x = U1−α Bα x.

Conversely, let Uα x ∈ D(Dα ). Choose a sequence (xn ) in D(A2 ) such that xn →
x in X (since A is densely defined with non-empty resolvent set all its powers Ak are
densely defined as well). Let s > 0. Choose a delta sequence (φk ) in Cc∞ (0, ∞)

which converges towards δs in C (0, ∞) , i.e.,

∞

∀f ∈ C (0, ∞) : lim f (t)tφk (t)dt = f (s).
k→∞
0

By Theorem 2.10, for t ≥ 0 we have

d
t 1−2α Uα xn (t) = U1−α Bα xn (t) = Bα U1−α xn (t).
dt
Using the closedness of Bα , by Hille’s theorem we obtain

∞ ∞
1−2α d
t Uα (t)xn φk (t)dt = Bα U1−α (t)xn φk (t)dt.
dt
0 0

Sending n → ∞ we get

∞ ∞
1−2α d
t Uα (t)x φk (t)dt = Bα U1−α (t)x φk (t)dt.
dt
0 0

Now, the limit k → ∞ yields by closedness of Bα that

d
s 1−2α Uα (s)x = Bα U1−α x(s)
ds
Sending s → 0+ and using a last time the closedness of Bα finally yields x ∈
D(Bα ).
Corollary 2.13 We have Bα B1−α = B1−α Bα = A.
Proof Note, that we already showed Bα B1−α x = B1−α Bα x = Ax for x ∈ D(A2 ).
So, let now x ∈ D(A). Then x ∈ D(Bα ), essentially by definition and the fact that
Consequences of Solvability of the Extension Problem 453

Dα Uα x = U1−α Bα x by Corollary 2.12. But since x ∈ D(A) we have Dα Uα x =

U1−α Bα x ∈ D(D1−α ) so applying Corollary 2.12 once more with 1 − α instead of
α we obtain Bα x ∈ D(B1−α ) and

Uα B1−α Bα x = D1−α Dα Uα x = AUα x = Uα Ax


in Cb [0, ∞); X . Evaluating at 0 yields B1−α Bα x = Ax. Conversely, let x ∈
D(B1−α Bα ). By two applications of Corollary 2.12 we then obtain that Uα x ∈
D(D1−α Dα ). By Lemma 2.7, this implies x ∈ D(A). The equality B1−α Bα x = Ax
is now verified in the same manner as before. The missing part of the Corollary
follows by exchanging α and 1 − α.
We end this paper by a conjecture, formulated as an open problem. Assume that
there exists an open sector Sφ ⊆ C of (half-)opening angle φ ∈ (0, π/2) around
the positive real axis such that for every δ ∈ (0, φ) the function Uα x is analytic on
Sδ and continuous and bounded on Sδ for all x ∈ D(A) w.r.t. the norm ·D(A) .
Note that this implies that the same holds for U1−α x. Actually, it is well known
that this even
 implies analyticity of the mappings Uα and U1−α w.r.t. the topology
of L D(A) . We conjecture that A is then sectorial of angle ω := π − 2φ, and
Bα = Aα and B1−α = A1−α . Put differently, does analyticity of the solutions of
Pα (x) for all x ∈ D(A) yield sectoriality of the operator A? An affirmative answer
to this question would yield the converse of [7, Theorem 2.2] which states that
sectoriality of A yields analyticity of the solution of Pα (x).

References

1. W. Arendt, C.J.K. Batty, M. Hieber, F. Neubrander, Vector-valued Laplace Transforms and

Cauchy Problems, 2nd edn. (Birkhäuser, Basel, 2011)
2. W. Arendt, A.F.M. ter Elst, M. Warma, Fractional powers of sectorial operators via the Dirichlet-
to-Neumann operator. Comm. Partial Differential Equations 43, 1–24 (2018)
3. A.V. Balakrishnan, Fractional powers of closed operators and the semigroups generated by them.
Pacific J. Math. 10, 419–437 (1960)
4. L. Caffarelli, L. Silvestre, An extension problem related to the fractional Laplacian. Comm.
Partial Differential Equations 32(8), 1245–1260 (2007)
5. J.E. Galé, P.J. Miana, P.R. Stinga, Extension problem and fractional operators: semigroups and
wave equations. J. Evol. Equ. 13, 343–368 (2013)
6. C. Martinez, M. Sanz, The Theory of Fractional Powers of Operators. North-Holland Mathe-
matics Studies, vol. 187 (Elsevier Science, Amsterdam, 2001)
7. J. Meichsner, C. Seifert, A Note on the harmonic extension approach to fractional powers of
non-densely defined operators. Proc. Appl. Math. Mech. 19, e201900296 (2019)
8. J. Meichsner, C. Seifert, On the harmonic extension approach to fractional powers in Banach
spaces. Arxiv preprint (2019). https://arxiv.org/abs/1905.06779
9. P.R. Stinga, J.L. Torrea, Extension problem and Harnack’s inequality for some fractional
operators. Comm. Partial Differential Equations 35, 2092–2122 (2010)
Time-Dependent Approach to Uniqueness
of the Sommerfeld Solution to a Problem
of Diffraction by a Half-Plane

A. Merzon, P. Zhevandrov, J. E. De la Paz Méndez, and T. J. Villalba Vega

Abstract We consider the Sommerfeld problem of diffraction by an opaque

half-plane interpreting it as the limiting case as t → ∞ of the corresponding
non-stationary diffraction problem. We prove that the Sommerfeld formula for the
solution is the limiting amplitude of the solution of this non-stationary problem
which belongs to a certain functional class and is unique in it. For the proof of
the uniqueness of solution to the non-stationary problem we reduce this problem,
after the Fourier–Laplace transform in t, to a stationary diffraction problem with a
complex wave number. This permits us to use the proof of the uniqueness in the
Sobolev space H 1 as in (Castro and Kapanadze, J Math Anal Appl 421(2):1295–
1314, 2015). Thus we avoid imposing the radiation condition from the beginning
and instead obtain it in a natural way.

Keywords Diffraction · Uniqueness

Mathematics Subject Classification (2010) Primary 35Q60; Secondary 78A45

Supported by CONACYT and CIC (UMSNH).

A. Merzon ()
Instituto de Física y Matemáticas, Universidad Michoacana de San Nicolás de Hidalgo, Morelia,
Michoacán, México
P. Zhevandrov
Facultad de Ciencias Físico-Matemáticas, Universidad Michoacana de San Nicolás de Hidalgo,
Morelia, Michoacán, México
J. E. De la Paz Méndez · T. J. Villalba Vega
Universidad Autonoma de Guerrero, Chilpancingo, Guerrero, México

© Springer Nature Switzerland AG 2021 455

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_23
456 A. Merzon et al.

1 Introduction

The main goal of this paper is to prove the uniqueness of a solution to the Som-
merfeld half-plane problem [23, 32, 33] with a real wave number, proceeding from
the uniqueness of the corresponding time-dependent problem in a certain functional
class. The existence and uniqueness of solutions to this problem was considered in
many papers, for example in [8, 12, 25]. However, in our opinion, the problem of
uniqueness is still not solved in a satisfactory form from the point of view of the
boundary value problems (BVPs). The fact is that this problem is a homogeneous
BVP boundary value problem which admits various nontrivial solutions. Usually the
“correct” solutions are chosen by physical reasoning [23, 25, 32, 33], for example,
using the Sommerfeld radiation conditions and regularity conditions at the edge.
The question is: from where do the radiation and regularity conditions arise,
from the mathematical point of view?
Our goal is to show that they arise automatically from the non-stationary
problem. This means the following: we prove that the Sommerfeld solution is a
limiting amplitude of a solution to the corresponding non-stationary problem which
is unique in an appropriate functional class. Since the Sommerfeld solution, as is
well-known, satisfies the radiation and regularity conditions, our limiting amplitude
also satisfies them. Of course, the limiting amplitude principle (LAP) is very well-
known for the diffraction by smooth obstacles, see e.g. [28, 29], but we are unaware
of its rigorous proof in the case of diffraction by a half-plane.
The literature devoted to diffraction by wedges including the Sommerfeld
problem is enormous (see e.g. the review in [20]), and we will only indicate some
papers where the uniqueness is treated. In paper [25] a uniqueness theorem was
proven for the Helmholtz equation (! + 1)u = 0 in two-dimensional regions D of
half-plane type. These regions can have a finite number of bounded obstacles with
singularities on their boundaries. In particular, the uniqueness of solution u to the
Sommerfeld problem was proven by means of the decomposition of the solution into
the sum u = g + h, where g describes the geometrical optics incoming and reflected
waves and h satisfies the Sommerfeld radiation condition (clearly, u should also
satisfy the regularity conditions at the edge).
In paper [8] exact conditions were found for the uniqueness in the case of
complex wave number. The problem was considered in Sobolev spaces for a wide
class of generalized incident waves, and for DD and NN boundary conditions. In
paper [12] the same problem was considered also for the complex wave number and
for DN boundary conditions. In both papers the Wiener-Hopf method has been used.
Time-dependent scattering by wedges was considered in many papers although their
number is not so large as the number of papers devoted to the stationary scattering
by wedges. We indicate here the following papers: [1–4, 13, 14, 24, 26–31]. The
detailed description of these papers is given in [19].
In [6, 7, 10, 17–20, 22], the diffraction by a wedge of magnitude φ (which can be
a half-plane in the case φ = 0 as in [20]) with real wavenumber was considered as
a stationary problem which is the “limiting case” of a non-stationary one. More
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 457

precisely, we seek the solutions of the classical diffraction problems as limiting

amplitudes of solutions to corresponding non-stationary problems, which are unique
in some appropriate functional class. We also, like in [25], decomposed the solution
of non-stationary problem separating a “bad” incident wave, so that the other part
of solution belongs to a certain appropriate functional class. Thus we avoided the
a priori use of the radiation and regularity conditions and instead obtained them in
a natural way. In papers [10, 17, 19] we considered the time-dependent scattering
with DD, DN and NN boundary conditions and proved the uniqueness of solution
in an appropriate functional class. But these results were obtained only for φ = 0
because in the proof of uniqueness we used the Method of Complex Characteristics
[15, 16, 21] which “works” only for φ = 0.
For φ = 0 we need to use other methods, namely, the reduction of the
uniqueness problem for the stationary diffraction to the uniqueness problem for the
corresponding non-stationary diffraction, which, in turn, is reduced to the proof of
uniqueness of solution of the stationary problem but with a complex wavenumber,
see e.g. [5].
Note that in [18] we proved the LAP for φ = 0 and for the DD boundary
conditions. Similar results for the NN and DN boundary conditions were obtained
in [6, 7, 10]. A generalization of these results to the case of generalized incident
wave (cf. [8]) was given in [19]. This approach (stationary diffraction as the limit
of time-dependent one) permits us to justify all the classical explicit formulas
[13, 14, 20, 28–31] and to prove their coincidence with the explicit formulas given
in [17, 19, 22]. In other words, all the classical formulas are limiting amplitudes
of solutions to non-stationary problems as t → ∞. For the Sommerfeld problem,
this was proven in [20], except for the proof of the uniqueness of the solution to
the non-stationary problem in an appropriate class. This paper makes up for this
omission.
Our plan is as follows. The non-stationary diffraction problem is reduced by
means of the Fourier–Laplace transform with respect to time t to a stationary one
with a complex wave number. For this problem the uniqueness theorems can be
proven more easily in Sobolev classes (see an important paper [5]) and do not
use the radiation conditions. Then we prove that the Fourier–Laplace transforms of
solutions to non-stationary diffraction half-plane problem, whose amplitude tends
to the Sommerfeld solution, also belong to a Sobolev space for a rather wide class
of incident waves. This permits us to reduce the problem to the case of [5].
Let us pass to the problem setting. We consider the two-dimensional time-
dependent scattering of a plane wave by the half-plane
 
W 0 := (x1 , x2 ) ∈ R2 : x2 = 0, x1 ≥ 0 .

(Obviously, W 0 is a half-line in R2 , but if one recalls that the initial problem is

three-dimensional, W 0 becomes a half-plane; the third coordinate is suppressed in
458 A. Merzon et al.

all what follows.) The non-stationary incident plane wave in the absence of obstacles
reads

ui (x, t) = e−iω0 (t −n·x)f (t − n · x), x ∈ R2 , t ∈ R, (1.1)

where

ω0 > 0, n = (n1 , n2 ) = (cos(π + α), sin(π + α)), (1.2)

and f is “a profile function”, such that f ∈ L1loc (R), and

f (s) = 0, s < 0, sup(1 + |s|)p |f (s)| < ∞ for some p ∈ R, lim f (s) = 1.
s→+∞
(1.3)

Remark 1.1 Obviously, these functions satisfy the D’Alembert equation

ui (x, t) = 0 in the sense of distributions.
For definiteness, we assume that
π
< α < π. (1.4)
2

In this case the front of the incident wave ui reaches the half-plane W 0 for the first
time at the moment t = 0 and at this moment the reflected wave ur (x, t) is born
(see Fig. 1). Thus

ur (x, t) ≡ 0, t < 0.

Note that for t → ∞ the limiting amplitude of ui is exactly equal to the Sommerfeld
incident wave [33] by (1.3), cf. also (2.1) below.
The time-dependent scattering with the Dirichlet boundary conditions is
described by the mixed problem

u(x, t) := (∂t2 − !)u(x, t) = 0, x ∈ Q
t ∈ R, (1.5)
u(x1 , ±0, t) = 0, x1 > 0

where Q := R2 \ W 0 . The “initial condition” reads

u(x, t) = ui (x, t), x ∈ Q, t < 0, (1.6)

where ui is the incident plane wave (1.1).

Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 459

Fig. 1 Time-dependent diffraction by a half-plane

Introduce the non-stationary “scattered” wave us as the difference between u and

ui ,

us (x, t) := u(x, t) − ui (x, t), x ∈ Q, t ∈ R. (1.7)

Since ui (x, t) = 0, (x, t) ∈ Q × R, we get from (1.6), (1.5) that

us (x, t) = 0, (x, t) ∈ Q × R, (1.8)

us (x, t) = 0, x ∈ Q, t < 0, (1.9)
us (x1 , ±0, t) = −ui (x1 , 0, t), x1 > 0, t > 0. (1.10)

Denote

ϕ± := π ± α. (1.11)

Everywhere below we assume that

x1 = r cos ϕ, x2 = r sin ϕ, 0 ≤ ϕ < 2π. (1.12)

Let us define the nonstationary incident wave in the presence of the obstacle W 0 ,
which is the opaque screen,

ui (ρ, ϕ, t), 0 < ϕ < ϕ+ ,
u0i (ρ, ϕ, t) := (1.13)
0, ϕ+ < ϕ < 2π.
460 A. Merzon et al.

Remark 1.2 The function us has no physical sense, since ui = u0i . The wave us
coincides with the scattered wave u0s := u − u0i in the zone {(ρ, ϕ) : 0 < ϕ < ϕ+ },
but in the zone {(ρ, ϕ) : ϕ+ < ϕ ≤ 2π} we have u0s = us + ui .
The goal of the paper is to prove that the Sommerfeld solution of half-plane
diffraction problem is the limiting amplitude of the solution to time-dependent
problem (1.5), (1.6) (with any f satisfying (1.3)) and this solution is unique in an
appropriate functional class.
The paper is organized as follows. In Sect. 2 we recall the Sommerfeld solution.
In Sect. 3 we reduce the time-dependent diffraction problem to a “stationary” one
and define a functional class of solutions. In Sect. 4 we give an explicit formula
for the solution of time-dependent problem and prove that the Sommerfeld solution
is its limiting amplitude. In Sect. 5 we prove that the solution belongs to a certain
functional class. Finally, in Sect. 6 we prove the uniqueness.

2 Sommerfeld’s Diffraction

Let us recall the Sommerfeld solution [23, 33]. The stationary incident wave (rather,
the incident wave limiting amplitude) in the presence of the obstacle is

e−iω0 ρ cos(ϕ−α) , ϕ ∈ (0, ϕ+ ),
A0i (ρ, ϕ) = (2.1)
0, ϕ ∈ (ϕ+ , 2π).

We denote this incident wave as A0i since it is the limiting amplitude of the non-
stationary incident wave u0i given by (1.13):

A0i (ρ, ϕ) = lim eiω0 t u0i (x, t),

t →∞

in view of formula (1.1), see Remark 1.2. The Sommerfeld half-plane diffraction
problem can be formulated as follows: find a function A(x), x ∈ Q, such that

(! + ω02 )A(x) = 0, x ∈ Q,
(2.2)
A(x1 , ±0) = 0, x1 > 0,

A(x) = A0i (x) + Ar (x) + Ad (x), x ∈ Q, (2.3)

where Ar (x) is the reflected wave,


−e−iω0 ρ cos(ϕ+α) , ϕ ∈ (0, ϕ− ),
Ar (x) = (2.4)
0, ϕ ∈ (ϕ− , 2π),
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 461

and Ad (x) is the wave diffracted by the edge,

Ad (x) → 0, |x| → ∞. (2.5)

A. Sommerfeld [33] found the solution of this problem in the form


1
A(ρ, ϕ) = ζ(γ , ϕ)e−iωρ cos γ dγ , ρ ≥ 0, ϕ ∈ [0, 2π],
4π
C

where
 −1  −1
ζ (γ , ϕ) := 1 − ei(−γ +ϕ−α)/2 − 1 − ei(−γ +ϕ+α)/2 , γ ∈C (2.6)

and C is the Sommerfeld contour (see [20, formula (1.1) and Fig. 3]).
In the rest of the paper we prove that this solution is the limiting amplitude of the
solution of time-dependent problem (1.5) and is unique in an appropriate functional
class.
The Sommerfeld diffraction problem can also be considered for NN and DN
half-plane. The corresponding formulas for the solution can be found in [19].
Sommerfeld obtained his solution using an original method of solutions of the
Helmholtz equation on a Riemann surface. Note that a similar approach was used
for the diffraction by a wedge of rational angle [9], where well-posedness in suitable
Sobolev space was proved.

3 Reduction to a “Stationary” Problem: Fourier–Laplace

Transform

Let 
h(ω), ω ∈ C+ , denote the Fourier–Laplace transform Ft →ω of h(t),

∞

h(ω) = Ft →ω [h(t)] = eiωt h(t) dt, h ∈ L1 (R+ ); (3.1)
0

Ft →ω is extended by continuity to S (R+ ). Assuming that us (x, t) belongs to

S (R2 × R+ ) (see (1.9) and Definition 3.1), we apply this transform to system (1.8)–
(1.10), and obtain

us (x, ω) = 0,
(! + ω2 ) x ∈ Q,
ω ∈ C+ . (3.2)

us (x1 , ±0, ω) = −
ui (x1 , ±0, ω), x1 > 0
462 A. Merzon et al.

Let us calculate 
ui (x, ω). Changing the variable t − n · x = τ , and using the fact
that suppf ⊂ R+ we obtain from (1.1) and (1.2) that

ui (x, ω) = eiωn·x f(ω − ω0 ).

 (3.3)

Hence,

ui (x1 , 0, ω) = eiωn1 x1 f(ω − ω0 ),

 x1 > 0,

and the boundary condition in (3.2) is 

us (x1 , 0, ω) = −g(ω)eiωn1 x1 . Therefore we
come to the following family of BVPs depending on ω ∈ C+ : find  us (x, ω) such
that

(! + ω2 )us (x, ω) = 0, x ∈ Q,
(3.4)

us (x1 , ±0, ω) = −g(ω)eiωn1 x1 , x1 > 0.

We are going to prove the existence and uniqueness of solution to problem (1.5),
(1.6) such that us given by (1.7) belongs to the space M, which is defined as follows:
Definition 3.1 M is the space of functions u(x, t) ∈ S (R2 × R+ ) such that its
u(x, ω) is a holomorphic function on ω ∈ C+ with
Fourier–Laplace transform 
2
values in C (Q) and

u(·, ·, ω) ∈ H 1 (Q)
 (3.5)

for any ω ∈ C+ .
Remark 3.2 We use the classical definition [11] of the space H 1 (Q) as the
completion of the space of smooth functions on Q with respect to the corresponding
norm. This definition does not coincide with the frequently used definition of
H 1 (Q) as the space restrictions of distributions from H 1 (R) to Q. In our case these
definitions lead to different spaces; in particular, the latter definition does not allow
for functions which are discontinuous across W 0 . In [34], another space allowing for
the same class of functions was introduced; the proof of uniqueness of the solution
to our problem in that space is an open question.
Remark 3.3 Note that ui (x, t) R2 ×R+
∈
/ M, where for ϕ ∈ D(R2 ),

  
ui (x, t) R2 ×R+
, ϕ := u(x, t)ϕ(x, t) dx dt.
R2 ×R+

In fact, eiωn·x = eω2 ρ cos(ϕ−α) and, for α − π/2 < ϕ < α + π/2, ω ∈ C+ it grows
exponentially as ρ → ∞, and hence does not satisfy (3.5); because of this we use
system (1.8)–(1.10) instead of (1.5) (they are equivalent by (1.6)) since (1.8)–(1.10)
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 463

involves only the values of ui on the boundary and the latter possess the Fourier–
Laplace transforms which do not grow exponentially.
Remark 3.4 Since for a (weak) solution of the Helmholtz equation us ∈ H 1 (Q)
the Dirichlet and Neumann data exist in the trace sense and in the distributional
sense, respectively (see, e.g., [5]), problem (3.4) is well-posed. Hence, problem
(1.8)–(1.10) is well-posed too.

4 Connection Between the Non-stationary Diffraction

Problem (1.5) and (1.6) and the Sommerfeld Half-Plane
Problem

In paper [20] we solved problem (1.5) and (1.6). Let us recall the corresponding
construction. First we define the non-stationary reflected wave [20, formula (26)]:

−e−iω0 (t −n·x) f (t − n · x), ϕ ∈ (0, ϕ− )
ur (x, t) = t ≥ 0, (4.1)
0, ϕ ∈ (ϕ− , 2π)

where n := (n1 , −n2 ) = (− cos α, sin α) (see Fig. 1).

Note that its limiting amplitude coincides with (2.4) similarly to the incident
wave.
Second, we define the non-stationary diffracted wave (cf. [20, formula (31) for
φ = 0]). Let

Z(β, ϕ) := Z(β + 2πi − iϕ), (4.2)

and

i
ud (ρ, ϕ, t) = Z(β, ϕ)F (t − ρ cosh β) dβ, (4.3)
8π
R

where ϕ ∈ (0, 2π), ϕ = ϕ± ; t ≥ 0,

F (s) = f (s)e−iω0 s , (4.4)

 iπ   5iπ 
Z(z) = −U − +z +U − +z , (4.5)
2 2
 π   π  1
U (ζ ) = coth q(ζ − i + iα) − coth q(ζ − i − iα) , q= (4.6)
2 2 4
for the Dirichlet boundary conditions. Below in Lemma 8.1 we give the necessary
properties of the function Z, from which the convergence of integral (4.3) follows.
464 A. Merzon et al.

Obviously, the condition supp F ⊂ [0, ∞) (see (3.1)) implies that supp ud (·, ·, t) ⊂
[0, +∞).
Remark 4.1 The function U (γ + ϕ) essentially coincides with the Sommerfeld
kernel (2.6). This is for a reason. In paper [17] it was proven that the solution to the
corresponding time-dependent diffraction problem by an arbitrary angle φ ∈ (0, π]
belonging to a certain class similar to M necessarily has the form of the Sommerfeld
type integral with the Sommerfeld type kernel.
Finally, we proved [20, Th. 3.2, Th 4.1] the following.
Theorem 4.2
(i) For f ∈ L1loc (R) the function

u(ρ, ϕ, t) := u0i (ρ, ϕ, t) + ur (ρ, ϕ, t) + ud (ρ, ϕ, t), ϕ = ϕ± (4.7)

belongs to L1loc (Q × R+ ). It is continuous up to ∂Q × R and satisfies the

boundary and initial conditions (1.5), (1.6). The D’Alembert equation in (1.5)
holds in the sense of distributions.
(ii) The LAP holds for Sommerfeld’s diffraction by a half-plane:

lim eiω0 t u(ρ, ϕ, t) = A(ρ, ϕ), ϕ = ϕ±

t →∞

(the limit here and everywhere else is pointwise).

Since the main object of our consideration will be the “scattered” wave us (x, t)
given by (1.7), we clarify the connection between us and the Sommerfeld solution
A.
Corollary 4.3 Define Ai (x) = e−iω0 ρ cos(ϕ+α) , which is the limiting amplitude of
ui (x, t) given by (1.1). The limiting amplitude of us (x, t) is the function

As (x) = A(x) − Ai (x), (4.8)

i.e. limt →∞ eiω0 t us (x, t) = As (x).

Proof The statement follows from (1.7).
Remark 4.4 The function As is the limiting amplitude of the scattered non-
stationary wave us (x, t) and As satisfies the following nonhomogeneous BVP:

(! + ω02 )As (x) = 0, x ∈ Q,
(4.9)
As (x1 , ±0) = −Ai (x1 , 0), x1 > 0.

This BVP (as well as (2.2)) is ill-posed since the homogeneous problem admits
many solutions (i.e., the solution is nonunique).
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 465

Remark 4.5 As can be decomposed similarly to (2.3). Namely, by (4.8) and (2.3),
we have

As = A0i + Ar (x) + Ad (x) − Ai (x) = Ar (x) + Ad (x) − A1i (x), (4.10)

where A1i (x) = Ai (x) − A0i (x). Obviously, problems (4.9), (4.10) and (2.2), (2.3)
with condition (2.5) are equivalent, but the first problem is more convenient as we
will see later.

5 Solution of the “Stationary” Problem

In this section we will obtain an explicit formula for the solution of (3.4) and prove
that it belongs to H 1 (Q) for all ω ∈ C+ .
Let Z(β, ϕ) be given by (4.2). First, we will need the Fourier–Laplace transforms
of the reflected and diffracted waves (4.1), (4.3).
Lemma 5.1 The Fourier–Laplace transforms of ur and ud are

−f(ω − ω0 )e−iωρ cos(ϕ+α) , ϕ ∈ (0, ϕ− ),

ur (x, ω) = (5.1)
0, ϕ ∈ (ϕ− , 2π),

i 

ud (ρ, ϕ, ω) = f (ω−ω0 ) Z(β, ϕ) eiωρ cosh β dβ, ω ∈ C+ , ϕ = ϕ± . (5.2)
8π
R

Proof From (4.1) we have


−Ft →ω e−iω0 (t −n·x) f (t − n · x) , ϕ ∈ (0, ϕ− ),
ur (x, ω) =

0, ϕ ∈ (ϕ− , 2π).

Further,

∞
−iω0 (t −n·x)
−Ft →ω e f (t − n · x) = −e iω0 (n·x)
ei(ω−ω0 )t f (t − n · x) dt.
0

Changing the variable t − n · x = τ , we obtain

∞
iω n·x

ur (x, ω) = −e ei(ω−ω0 )τ f (τ ) dτ, ϕ ∈ (0, ϕ− ).
−n·x
466 A. Merzon et al.

Moreover, by (4.1),

−n · x = ρ cos(ϕ − α) ≤ c < 0, ϕ ∈ (0, ϕ− ),

since π/2 < α < ϕ + α < π by (1.4) and (1.11). Hence, we obtain (5.1), since
suppf ⊂ R+ . The second formula in (5.1) follows from definition (4.1) of ur .
Let us prove (5.2). Everywhere below we put ω = ω1 + iω2 , ω1,2 ∈ R, ω2 > 0,
for ω ∈ C+ . By Lemma 8.1(i), (1.3) and (4.4) we have

eiωt Z(β, ϕ)F (t − ρ cosh β) ≤ Ce−ω2 t e−β/2 (1 + t)−p , ρ < 0, ϕ = ϕ± , β ∈ R.

Hence, by the Fubini Theorem there exists the Fourier–Laplace transform of

ud (·, ·, t) and

i
ud (ρ, ϕ, ω) = Z(β, ϕ)Ft →ω F (t − ρ cosh β) dβ, ϕ = ϕ± . (5.3)
8π
R

We have
∞
G(ρ, β, ω) := Ft →ω F (t − ρ cosh β) = eiωt F (t − ρ cosh β) dt, ω ∈ C+ .
0

Making the change of the variable τ = t − ρ cosh β in the last integral and
using the fact that supp F ⊂ [0, ∞) and F (ω) = f(ω − ω0 ) by (4.4), we get
G(ρ, β, ω) = e iωρ cosh β 
f (ω − ω0 ). Substituting this expression into (5.3) we obtain
(5.2). Lemma 5.1 is proven.

5.1 Estimates for 

u r , ∂ρ 
u r , ∂ϕ 
ur

Lemma 5.2 For any ω ∈ C, there exist C(ω), c(ω) > 0, such that both functions

ur and ∂ρ 
ur admit the same estimate

ur (ρ, ϕ, ω) ≤ C(ω)e−c(ω)ρ

ρ > 0, ϕ ∈ (0, 2π), ϕ = ϕ± . (5.4)
ur (ρ, ϕ, ω) ≤ C(ω)e−c(ω)ρ
∂ρ 

and ∂ϕ 
ur (ρ, ϕ, ω) admits the estimate

ur (ρ, ϕ, ω)| ≤ C(ω)ρ e−c(ω)ρ ,

|∂ϕ ρ > 0. (5.5)
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 467

Proof By (1.4) there exits c(ω) > 0 such that

e−iωρ cos(ϕ+α) = eω2 ρ cos(ϕ+α) ≤ e−c(ω)ρ , 0 < ϕ < ϕ−

by (1.4). Therefore (5.4) holds for  ur . Hence, differentiating (5.1) we obtain (5.4)
for ∂ρ 
ur and (5.5) for ∂ϕ 
ur , for ϕ = ϕ− .

5.2 Estimates for 

ud

Proposition 5.3 There exist C(ω), c(ω) > 0 such that the function 
ud , and ∂ρ 
ud ,
∂ϕ
ud admit the estimates

ud (ρ, ϕ, ω) ≤ C(ω)e−c(ω)ρ ,


ud (ρ, ϕ, ω) ≤ C(ω)e−c(ω)ρ (1 + ρ −1/2 ),

∂ρ  (5.6)

ud (ρ, ϕ, ω) ≤ C(ω)e−c(ω)ρ ρ(1 + ρ −1/2 )

∂ϕ

for ρ > 0, ϕ ∈ (0, 2π), ϕ = ϕ± .

Proof
(I) By (5.2), in order to prove (5.6) for 
ud it suffices to prove that

|A(ρ, ϕ, ω)| ≤ C(ω)e−c(ω)ρ , (5.7)

where

A(ρ, ϕ, ω) := Z(β, ϕ)eiωρ cosh β dβ, ϕ = ϕ± . (5.8)
R

Represent A as A = A1 + A2 , where

1
A1 (ρ, ϕ, ω) := Z(β, ϕ)eiωρ cosh β dβ
−1 ϕ ∈ (0, 2π), ϕ = ϕ± .
A2 (ρ, ϕ, ω) := Z(β, ϕ)eiωρ cosh β dβ
|β|≥1
(5.9)
468 A. Merzon et al.

The estimate (5.7) for A2 follows from (8.1) (see Appendix 1). It remains to
prove the same estimate for the function A1 . Let

ε± := ϕ± − ϕ. (5.10)

Representing A1 as

1
A1 (ρ, ϕ, ω) = −4K0 (ρ, w, ε+ ) + 4K0 (ρ, w, ε− ) + Ž(β, ϕ)eiωρ cosh β dβ,
−1

where K0 is defined by (8.7), we obtain (5.7) for A1 from Lemma 8.2 (i) and
(8.3).
(II) Let us prove (5.6) for ∂ρ 
ud . By (5.2) it suffices to prove that

|B(ρ, ϕ, ω)| ≤ C(ω)e−c(ω)ρ (1 + ρ 1/2 ), ϕ = ϕ± , (5.11)

where

B(ρ, ϕ, ω) := Z(β, ϕ) cosh β eiωρ cosh β dβ.
R

Represent B as B1 + B2 , where B1,2 (ρ, ϕ, ω) are defined similarly to (5.9),

1
B1 (ρ, ϕ, ω) := Z(β, ϕ) cosh βeiωρ cosh β dβ,
−1

B2 (ρ, ϕ, ω) := Z(β, ϕ) cosh βeiωρ cosh β dβ, ϕ = ϕ± .
|β|≥1

From (8.1) for Z we have

∞
1
eβ/2 e− 2 ω2 ρe dβ.
β
|B2 (ρ, ϕ, ω)| ≤ C1
1

Making the change of the variable ξ := ρeβ , we get

⎧ −1/2 , ρ ≤ 1,
⎪
⎪ C1 (ω)ρ
⎪
⎨ ∞
 −ω2 ξ/2
|B2 (ρ, ϕ, ω)| ≤ e
⎪
⎪ dξ, ρ ≥ 1.
⎪
⎩ ξ 1/2
ρ
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 469

Since for ρ ≥ 1,
∞
e−ω2 ξ/2 2 −ω2 ρ/2
1/2
dξ ≤ e ,
ξ ω2
ρ

Equation (5.11) is proved for B2 .

It remains to prove estimate (5.11) for B1 . Using (8.2) and (8.8) we write

1
B1 (ρ, ϕ, ω) = −4K1 (ρ, ω, ε+ ) + 4K1 (ρ, ω, ε− ) + Ž (β, ϕ) · cos β e iωρ cosh β dβ.
−1

Hence, B1 satisfies (5.7) (and, therefore, (5.11)) by Lemma 8.2 (i) and (8.3).
(III) Let us prove (5.6) for ∂ϕ 
ud . By (5.2) it suffices to prove this estimate for ∂ϕ A,
where A is given by (5.8). From (9.3) we have

∂ϕ A(ρ, ϕ, ω) = −ωρA3 (ρ, ϕ, ω),


(5.12)
A3 (ρ, ϕ, ω) = Z(β, ϕ) sinh β eiωρ cosh β dβ, ϕ = ϕ± .
R

Similarly to the proof of estimate (5.11) for B, we obtain the same estimate
for A3 , so, by (5.12), the estimate (5.6) follows. Proposition 5.3 is proven.

Now define the function

u0s (ρ, ϕ, t) = u(ρ, ϕ, t) − u0i (ρ, ϕ, t), ϕ = ϕ+ , t > 0, (5.13)

where u0i is given by (1.13). Then by (4.7),

u0s (ρ, ϕ, t) = ur (ρ, ϕ, t) + ud (ρ, ϕ, t), ϕ = ϕ± , t > 0, (5.14)

where ur is given by (4.1) and ud is given by (4.3).

Corollary 5.4 Let  u0s (ρ, ϕ, ω) be the Fourier–Laplace transform of the function
us (ρ, ϕ, t). Then the functions 
0 u0s , ∂ρ 
u0s and ∂ϕ
u0s satisfy (5.6).
Proof From (5.14) we have

u0s (ρ, ϕ, ω) = 
 ur (ρ, ϕ, ω) + 
ud (ρ, ϕ, ω), ϕ = ϕ± , ω ∈ C+ , (5.15)

ur and 
where  ud are defined by (5.1) and (5.2), respectively. Hence the statement
follows from Lemma 5.2 and Proposition 5.3.
470 A. Merzon et al.

5.3 Estimates for 

us (x, ω)

To estimate 
us it is convenient to introduce one more “part” u1i of the non-stationary
incident wave ui , namely the difference between ui and u0i .
From (1.7) and (5.13) it follows that

us (ρ, ϕ, t) = u0s (ρ, ϕ, t) − u1i (ρ, ϕ, t), ϕ = ϕ± (5.16)

where u1i (ρ, ϕ, t) := ui (ρ, ϕ, t) − u0i (ρ, ϕ, t). From (1.1) and (1.13) it follows that

0, 0 < ϕ < ϕ+ ,
u1i (ρ, ϕ, t) = (5.17)
−ui (ρ, ϕ, t), ϕ+ < ϕ < 2π.

By (3.3),

0, 0 < ϕ < ϕ+ ,

u1i (ρ, ϕ, ω) = (5.18)
−f(ω − ω0 ) eiωn·x , ϕ+ < ϕ < 2π.

u1i , ∂ρ 
Lemma 5.5 There exist C(ω), c(ω) > 0 such that  u1i satisfy (5.4) and ∂ϕ 
u1i
satisfies (5.5) for ϕ ∈ (0, 2π), ϕ = ϕ± .
Proof By (3.3) it suffices to prove the statement for eiωn·x when ϕ ∈ (ϕ+ , 2π).
Since |eiωn·x | = eω2 ρ cos(ϕ−α) , ϕ ∈ (ϕ+ , 2π) we have

∂ρ eω2 ρ cos(ϕ−α) = ω2 cos(ϕ − α)eω2 ρ cos(ϕ−α) ,

(5.19)
∂ϕ eω2 ρ cos(ϕ−α) = −ω2 ρ sin(ϕ − α)eω2 ρ cos(ϕ−α) ,

and for ϕ ∈ (ϕ+ , 2π), we have |eω2 ρ cos(ϕ−α) | ≤ e−cω2 ρ , c > 0, ϕ ∈ (ϕ+ , 2π),
because cos(ϕ − α) ≤ −c < 0 by (1.4). Hence the statement follows from (5.19).

us , ∂ρ 
Corollary 5.6 The functions  us and ∂ϕ 
us satisfy (5.6) for ϕ ∈ (0, 2π), ϕ =
ϕ± .
Proof From (5.16) it follows that


us (ρ, ϕ, ω) = 
u0s (ρ, ϕ, ω) − 
u1i (ρ, ϕ, ω). (5.20)

Thus the statement follows from Corollary 5.4 and Lemma 5.5.
It is possible to get rid of the restriction ϕ = ϕ± in Corollary 5.6.
Let l± = {(ρ, ϕ) : ρ > 0, ϕ = ϕ± }.
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 471

Proposition 5.7 The functions  us (·, ·, ω), ∂ρ 

us (·, ·, ω) and ∂ϕ
us (·, ·, ω) belong to
C 2 (Q), and satisfy (5.6) in Q (including l+ ∪ l− ), and

(! + ω2 )
us (ρ, ϕ, ω) = 0, (ρ, ϕ) ∈ Q, ω ∈ C+ . (5.21)

Proof The function  us (ρ, ϕ, ω) satisfies (5.21) in Q \ (l+ ∪ l− ). This follows

directly from the explicit formulas (5.20). In fact, (5.20) and (5.15) imply


us = 
ur + 
ud − 
u1i . (5.22)

The function ur satisfies (5.21) for ϕ = ϕ± ,  u1i satisfies (5.21) for ϕ = ϕ± by
(5.17) and (3.3) and ud satisfies (5.21) for ϕ = ϕ± by (5.2), see Appendix 2. It
remains only to prove that us ∈ C 2 (Q), because this will mean that (5.6) holds by
Corollary 5.6 (and continuity) and (5.21) holds in Q including l± .
Let us prove this for ϕ close to ϕ− . The case of ϕ close to ϕ+ is analyzed
similarly.
Let h(s) be defined in (C \ R) ∩ B(s ∗ ), where B(s ∗ ) is a neighborhood of s ∗ ∈ R.
Define the jump of h at the point s ∗ as

J (h, s ∗ ) := lim h(s ∗ + iε) − lim h(s ∗ − iε).

ε→0+ ε→0+

We have J (ur (ρ, ϕ, ω), ϕ− ) = f(ω − ω0 )e−iωρ by (5.1).

Similarly,
 
ur (ρ, ϕ, ω), ϕ− = 0,
J ∂ϕ ur (ρ, ϕ, ω), ϕ− = −f(ω − ω0 )(iωρ)eiωρ .
J ∂ϕϕ 

From (5.2), (5.10), (8.2), and (8.3) we have

 1 ε− =−0
i  4
J (
ud (ρ, ϕ, ω), ϕ− ) = f (ω − ω0 ) eiωρ cosh β dβ
8π −1 β + iε ε− =+0

= −J (
ur (ρ, ϕ, ω), ϕ− ) . (5.23)

Further, by (8.4),
 
ud (ρ, ϕ, ω), ϕ− = 0 = −J ∂ϕ
J ∂ϕ ur (ρ, ϕ, ω), ϕ− .

Finally, consider

M := J ∂ϕϕ 
ud (ρ, ϕ, ω), ϕ− .
472 A. Merzon et al.

Similarly to (5.23), expanding eiωρ cos β in the Taylor series in β (at 0) and noting
= β k dβ
that all the terms (β+iε )3
, k = 2, are continuous, we obtain
−

1 ε− =−0
i eiωρ cos β
M = − f(ω − ω0 ) dβ
π (β + iε− )3 ε− =+0
−1

1 ε− =−0
−i f(ω − ω0 )(iωρ)eiωρ β2
= dβ .
2π (β + iε− )3 ε− =+0
−1

Hence,

M = f(ω − ω0 )(iωρ)eiωρ = −J (
ur (ρ, ϕ, ω), ϕ− ) .

uii (ρ, ϕ, ω) is smooth on l− by (5.18), we obtain from (5.22) that 

Since  us ∈ C 2 (l− ).
Similarly using (5.1), (5.17) and (1.1) we obtain  us ∈ C (l+ ). So 
2 us ∈ C 2 (Q).
Proposition 5.7 is proven.
Corollary 5.8
us (·, ·, ω) belongs to the space H 1 (Q) for any ω ∈ C+ .
(i) The function 
(ii) The function us (x, t) ∈ M.
Proof
(i) Everywhere below x = (ρ, ϕ) ∈ Q \ (l1 ∪ l2 ). It suffices to prove that

us (·, ·, ω), ∂xk us (·, ·, ω) ∈ L2 (Q), k = 1, 2, ω ∈ C+ . (5.24)

us (x, ω) satisfies (5.6). Hence, 

First, by Proposition 5.7,  us (·, ω) ∈ L2 (Q) for
any ω ∈ C+ . Further, using (1.12), we have

| sin ϕ|2
|∂x1 us (·, ·, ω)|2 ≤ | cos ϕ|2 |∂ρ us (·, ·, ω)|2 + |∂ϕ us (·, ·, ω)|2 .
ρ2

Hence, by Proposition 5.7,

 
1
|∂x1 us (·, ·, ω)|2 ≤ C(ω)e−2c(ω) 1 + .
ρ

This implies that ∂x1 us ∈ L2 (Q), since c(ω) > 0. Similarly, ∂x2 us (·, ·, ω) ∈
L2 (Q). (5.24) is proven.
(ii) The statement follows from Definition 3.1.
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 473

6 Uniqueness

In Sect. 5 we proved the existence of solution to (1.8)–(1.10) belonging to M. In

this section prove the uniqueness of this solution in the same space.
Recall that we understand the uniqueness of the time-dependent Sommerfeld
problem (1.5)–(1.6) as the uniqueness of the solution us given by (1.7) of the mixed
problem (1.8)–(1.10) in the space M.
The following theorem is the main result of the paper.
Theorem 6.1
(i) Problem (1.8)–(1.10) admits a solution belonging to the space M. Its limiting
amplitude exists and is the solution of problem (4.9). The connection between
this limiting amplitude and the Sommerfeld solution is given by (4.8).
(ii) Problem (1.8)–(1.10) admits a unique solution in the space M.
Proof The statements contained in item (i) follow from Corollary 5.8, Corollary 4.3,
and Remark 4.4.
(ii) Let us prove the uniqueness. We follow closely the proof of Theorem 2.1
from [5]. Suppose that there exist two solutions us (x, t) and vs (x, t) of system
(1.8)–(1.10) belonging to M. Consider ws (x, t) := us (x, t) − vs (x, t).
Then w s (·, ·, ω) = us (·, ·, ω) − 
vs (·, ·, ω), where us ,  s )
vs (and, therefore, w
satisfy all the conditions of Proposition 5.7 and w s |W 0 = 0 by (3.4).
Let us prove that w s (·, ·, ω) ≡ 0. Let R be a sufficiently large positive number
and B(R) be the open disk centered at the origin with radius R. Set QR := Q ∩
B(R). Note that QR has a piecewise smooth boundary SR and denote by n(x) the
outward unit normal vector at the non-singular points x ∈ SR .
The first Green identity for ws (ρ, ϕ, ·) and its complex conjugate w s in the
domain QR , together with zero boundary conditions on SR , yield
     
|∇ w
s | − ω |
2
ws |
2 2
dx = s · ws dSR .
∂n w
QR ∂B(R)∩Q

From the real and imaginary parts of the last identity, we obtain
 
2  
s |2 + Im ω |
|∇ w ws |2 dx = Re s w
∂n w s dSR (6.1)
QR ∂B(R)∩Q

for Re ω = 0 and
 
   
− 2 Re ω Im ω |
ws |2 dx = Im ∂n ws ws dSR (6.2)
QR ∂B(R)∩Q
474 A. Merzon et al.

for Re ω = 0. Recall that we consider the case Im k = 0. Now, note that since

ωs ∈ H 1 (Q), there exist a monotonic sequence of positive numbers {Rj } such that
Rj → ∞ as j → ∞ and

lim s
∂n w s dSRj = 0.
w (6.3)
j →∞
∂B(Rj )∩Q

Indeed, in polar coordinates (ρ, ϕ), we have that the integrals

⎛ ⎞ ⎛ ⎞
∞ 2π ∞ 2π
⎝R ws (ρ, ϕ)|2 dϕ ⎠ dR
| and ⎝R s (ρ, ϕ)|2 dϕ ⎠ dR
|∂n w
0 0 0 0

are finite. This fact, in particular, implies that there exist a monotonic sequence of
positive numbers Rj such that Rj → ∞ as j → ∞ and

2π 2π
|
ws (Rj , ϕ)| dϕ = 2
o(Rj−1 ), s (Rj , ϕ)|2 dϕ = o(Rj−1 ) as j → ∞.
|∂n w
0 0

Further, applying the Cauchy-Schwarz inequality for every Rj , we get

2π 2π
∂n w ws (Ri , ϕ)dϕ ≤
s (Ri , ϕ) |∂n w
s (Ri , ϕ)
ws (Ri , ϕ)|dϕ
0 0
⎛ 2π ⎞1/2 ⎛ 2π ⎞1/2
 
s (Ri , ϕ)|2 dϕ ⎠ ⎝ |
≤ ⎝ |∂n w ws (Ri , ϕ)|2 dϕ ⎠
0 0

= o(Rj−1 ) as j → ∞,

and therefore we obtain (6.3).

Since the expressions under the integral sign in the left hand sides of equalities
(6.1) and (6.2) are non-negative, we have that these integrals are monotonic with
respect to R. This observation together with (6.3) implies
 
2
2 2
2
|∇ w
s | + Im ω |
ws | dϕ = lim s |2 + Im ω |
|∇ w ws |2 dϕ = 0
R→∞
Q QR
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 475

Fig. 2 Uniqueness

for Re ω = 0 and
 
|
ws | dϕ = lim
2
|
ws |2 dϕ = 0
R→∞
Q QR

for Re ω = 0. Thus, it follows from the last two identities that w

s = 0 in Q (Fig. 2).

7 Conclusion

We proved that the Sommerfeld solution to the half-plane diffraction problem

for a wide class of incident waves is the limiting amplitude of the solution of
the corresponding time-dependent problem in a functional class of generalized
solutions. The solution of the time-dependent problem is shown to be unique in
this class. It is also shown that the limiting amplitude automatically satisfies the
Sommerfeld radiation condition and the regularity condition at the edge.

Acknowledgments The authors are grateful to CONACYT (México) and CIC (UMSNH) for
partial financial support. We are also grateful to anonymous referees for valuable comments.

8 Appendix 1

Lemma 8.1
(i) The functions Z (given by (4.2)) and ∂ϕ Z admit uniform with respect to ϕ ∈
[0, 2π] estimates

|Z(β, ϕ)| ≤ Ce−|β|/2 , |∂ϕ Z(β, ϕ)| ≤ Ce−|β|/2 , |β| ≥ 1. (8.1)

(ii) The function Z admits the representation

4 4
Z(β, ϕ) = − + + Ž(β, ϕ), ε± = 0 (8.2)
β + iε+ β + iε−
476 A. Merzon et al.

with

Ž(β, ϕ) ∈ C ∞ (R × [0, 2π]), |Ž(β, ϕ)| ≤ C, β ∈ R × [0, 2π]. (8.3)

(iii) The function ∂ϕ Z admits the representation

4i 4i
∂ϕ Z = − + + Ž1 (β, ϕ), ε± = 0, (8.4)
(β + iε+ ) 2 (β + iε− )2

with

Ž1 (β, ϕ) ∈ C ∞ (R×[0, 2π]), |Ž1 (β, ϕ)| ≤ C, β ∈ R×[0, 2π]. (8.5)

Proof
(i) For a = im, b = in, we have

− sinh(α/2)
coth a − coth b = .
sinh(b) sinh(a)

Hence for m = −π/8 + a/4 and n = −π/8 − a/4 we obtain the estimate (8.1)
for U (ζ ) given by (4.6) with respect to ζ . So (8.1) for Z follows from (4.5)
and (4.2).
(ii) From (4.5) and (4.6) it follows that the function Z admits the representation

Z(β, ϕ) = Z+ (β, ϕ) + Z− (β, ϕ) + Z + (β, ϕ) + Z − (β, ϕ),

where
 
β + i(ϕ± − ϕ)
Z± (β, ϕ) = ± coth ,
4 
 (8.6)
β − i ϕ± + ϕ
Z ± (β, ϕ) = ± coth .
4

Further, since | coth z − 1/z| ≤ C, |Im z| ≤ π, z = 0, we have

4
Z± (β, ϕ) = ± + Ž± (β, ϕ), ϕ = ϕ± ,
β + iε±

and

Ž± (β, ϕ) ∈ C ∞ (R × [0, 2π]), |Ž± (β, ϕ)| ≤ C, (β, ϕ) ∈ R × [0, 2π].
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 477

Finally, by (1.4),

Z ± (β, ϕ) ∈ C ∞ (R × [0, 2π]), |Z ± (β, ϕ)| ≤ C, (β, ϕ) ∈ R × [0, 2π].

Therefore, (8.2) and (8.3) are proven.

(iii) From (8.2) and (5.10) we get (8.4). Finally, by (8.6),

∂ϕ Z ± (β, ϕ) ∈ C ∞ (R × [0, 2π ]), |∂ϕ Z ± (β, ϕ)| ≤ C, (β, ϕ) ∈ R × [0, 2π ].

Moreover, since

∂ϕ Z± (β, ϕ) ± [4i/(β + ε± )2 ] ∈ C ∞ ([R × [0, 2π]),

and is bounded in the same region, (8.5) holds.

For ε, β ∈ R, ε = 0, ρ > 0, ω ∈ C+ , let

1
eiωρ cosh β
K0 (β, ρ, ω, ε) := , K0 (ρ, ω, ε) := K(β, ρ, ω, ε) dβ, (8.7)
β + iε
−1

1
K1 (β, ρ, ω, ε) := cosh β · eiωρ cosh β , K1 (ρ, ω, ε):= K1 (β, ρ, ω, ε) dβ,
−1
(8.8)
1
eiωρ cosh β
K2 (β, ρ, ϕ, ε) := , K2 (ρ, ω, ε) := K2 (β, ρ, ω, ε) dβ dβ.
(β + iε)2
−1

Lemma 8.2 There exist C(ω) > 0, c(ω) > 0 such that the functions K0 , K1 , and
K2 satisfy the estimates

|K0,1,2 (ρ, ω, ε)| ≤ C(ω)e−c(ω)ρ , ρ > 0, ϕ ∈ (0, 2π), ε = 0. (8.9)

Proof It suffices to prove (8.9) for 0 < ε < ε0 , since the functions K0 , K1 , K2 are
odd with respect to ε, and for ε ≥ ε0 > 0 they satisfy the estimate

1
K0,1,2 (β, ρ, ω, ε) ≤ C(ε0 ) e−ω2 ρ dβ ≤ 2C(ε0)e−ω2 ρ .
−1
478 A. Merzon et al.

(I) Let us prove (8.9) for K0 . Let

cosh β := 1 + h(β), β ∈ C. (8.10)

Define ε0 = ε0 (ω) such that

1 ω2
|h(β)| < , |ω1 ||h(β)| ≤ for |β| ≤ 2ε0 := r, (8.11)
4 4
and define the contour

γr := {β = reiθ , −π < θ < 0}. (8.12)

Then we have by the Cauchy Theorem

K0 (ρ, ω, ε) = I1 (ρ, ω, ε) + I2 (ρ, ω, ε) − 2πi Resβ=−iε K0 (β, ρ, ω, ε),

where

  −r 1 
I1 (ρ, ω, ε) = K0 (β, ρ, ω, ε) dβ, I2 (ρ, ω, ε) = + K0 (β, ρ, ω, ε) dβ
γr −1 r

and 0 < ε < ε0 . First,

| Resβ=−iε K0 (β, ρ, ω, ε)| = e−ω2 ρ cos ε ≤ e− 2 ω2 ρ ,

1
0 < ε < ε0 ,
(8.13)
by (8.11). Further, from (8.10) we have
 
 e−ω2 ρ 1+h(β)
eiω1 ρ 1+h(β)

I1 (ρ, ω, ε) ≤ |dβ|
|β + iε|
γr

1 −ω2 ρ
≤ e e−ω2 ρ h(β)+iω1 ρ h(β) dβ , (8.14)
ε0
γr

since for β ∈ γr we have |β + iε| ≥ |β| − ε = 2ε0 − ε > ε0 , see Fig. 3.

Let h(β) := h1 (β) + ih2 (β). Then

1 −ω2 ρ
|I1 (ρ, ω, ε)| ≤ e eω2 ρ |h1 (β)| e|ω1 |ρ |h2 (β)| dβ ≤ 2πe−ω2 ρ/2 ,
ε0
γr
(8.15)
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 479

Fig. 3 Contour γr

by (8.11). Finally,

e−ω2 ρ cosh β+iω1 ρ cosh β e−ω2 ρ
|I2 (ρ, ω, ε)| ≤ dβ ≤ ,
β + iε 2ε0 (ω)
[−1,−r]∪[r,1]
(8.16)

since |β + iε| ≥ 2ε0 , β ∈ [−1, −r] ∪ [r, 1]. From (8.14)–(8.16), we obtain
(8.9) for K0 .
(II) Let us prove (8.9) for K1 . Let h(β), ε0 (ω), γr be defined by (8.10)–(8.12).
Then we have by the Cauchy Theorem

K1 (ρ, ω, ε) := K1 (β, ρ, ω, ε) dβ
γr ∪[−1,r]∪[r,1]

− 2πi Resβ=−iε K1 (β, ρ, ω, ε), 0 < ε < ε0 . (8.17)

First, similarly to (8.13), we obtain

ω2 ρ
Resβ=−iε K1 (β, ρ, ω, ε) ≤ |ω|e− 2 ,

by (8.11). Further, by (8.11) similarly to the proof of (8.14), (8.15), and using
(8.10), we get
 
|ω| 5 −ω2 ρ
K1 (β, ρ, ω, ε) dβ ≤ · e |e−ω2 ρ h(β) eiω1 ρ h(β) | |dβ|
ε0 4
γr γr
ω ρ
− 22
≤ C(ω)e . (8.18)
480 A. Merzon et al.

Finally, similarly to the proof of (8.16) we get the estimate


K1 (β, ρ, ω, ε) dβ ≤ C(ω) e−ω2 ρ . (8.19)
[−1,−r]∪[r,1]

From (8.17)–(8.19), we obtain (8.9) for K1 .

(III) Estimate (8.9) for K2 is proved similarly to the same estimate for K0 and K1
with obvious changes. Lemma 8.2 is proven.

9 Appendix 2

Lemma 9.1 We have

 
! + ω2 ud (ρ, ϕ, ω) = 0, ϕ = ϕ± , ω ∈ C+ . (9.1)

Proof By (5.2) it suffices to prove (9.1) for


Ad (ρ, ϕ, ω) := Z(β, ϕ)eiωρ cosh β dβ. (9.2)
R

Since ω ∈ C+ the integral (9.2) converges after differentiation with respect to ρ and
ϕ. We have

∂ρ Ad (ρ, ϕ, ω) = (iω) Z(β, ϕ) cosh β eiωρ cosh β dβ,
R

∂ρ2 Ad (ρ, ϕ, ω) = −ω2 Z(β, ϕ) cosh2 β eiωρ cosh β dβ.
R

Integrating by parts, we have by (4.2) and (8.1)

  
∂ϕ Ad (ρ, ϕ, ω) = ∂ϕ Z 0 (β + 2πi − iϕ) eiωρ cosh β dβ
R

= −ωρ Z(β, ϕ) sinh β eiωρ cosh β dβ, ϕ = ϕ± . (9.3)
R
Time-Dependent Approach to Uniqueness of the Sommerfeld Solution 481

Hence, similarly to (9.3)


2
∂ϕϕ Ad (ρ, ϕ, ω) = −iωρ Z(β, ϕ) cosh β + iωρ sinh2 β eiωρ cosh β dβ,
R

and
1
(! + ω2 )ud (ρ, ϕ, ω) =∂ρ2 Ad (ρ, ϕ, ω) + ∂ρ Ad (ρ, ϕω)
ρ
1 2
+ ∂ Ad (ρ, ϕ, ω) + ω2 Ad (ρ, ϕ, ω) = 0.
ρ2 ϕ

References

1. J. Bernard, Progresses on the diffraction by a wedge: transient solution for line source
illumination, single face contribution to scattered field, anew consequence of reciprocity on
the spectral function. Rev. Tech. Thomson 25(4), 1209–1220 (1993)
2. J. Bernard, On the time domain scattering by a passive classical frequency dependent-shaped
region in a lossy dispersive medium. Ann. Telecommun. 49(11–12), 673–683 (1994)
3. J. Bernard, G. Pelosi, G. Manara, A. Freni, Time domains scattering by an impedance wedge
for skew incidence, in Proceeding of the Conference ICEAA (1991), pp. 11–14
4. V. Borovikov, Diffraction at Polygons and Polyhedrons (Nauka, Moscow, 1996)
5. L.P. Castro, D. Kapanadze, Wave diffraction by wedges having arbitrary aperture angle. J.
Math. Anal. Appl. 421(2), 1295–1314 (2015)
6. J. De la Paz Méndez, A. Merzon, DN-Scattering of a plane wave by wedges. Math. Methods
Appl. Sci. 34(15), 1843–1872 (2011)
7. J. De la Paz Méndez, A. Merzon, Scattering of a plane wave by “hard-soft” wedges, in
Recent Progress in Operator Theory and Its Applications. Operator Theory: Advances and
Applications, vol. 220 (2012), pp. 207–227
8. A. Dos Santos, F. Teixeira, The Sommerfeld problem revisited: solution spaces and the edges
conditions. Math. Anal. Appl. 143, 341–357 (1989)
9. T. Ehrhardt, A. Nolasco, F.-O. Speck, A Riemann surface approach for diffraction from rational
wedges. Oper. Matrices 8(2), 301–355 (2014)
10. A. Esquivel Navarrete, A. Merzon, An explicit formula for the nonstationary diffracted wave
scattered on a NN-wedge. Acta Appl. Math. 136(1), 119–145 (2015)
11. P. Grisvard, Elliptic Problems in Nonsmooth Domains (Pitman Publishing, London, 1985)
12. A. Heins, The Sommerfeld half-plane problem revisited I: the solution of a pair of complex
Wiener–Hopf integral equations. Math. Methods Appl. Sci. 4, 74–90 (1982)
13. I. Kay, The diffraction of an arbitrary pulse by a wedge. Commun. Pure Appl. Math. 6, 521–546
(1953)
14. J. Keller, A. Blank, Diffraction and reflection of pulses by wedges and corners. Commun. Pure
Appl. Math. 4(1), 75–95 (1951)
15. A. Komech, Elliptic boundary value problems on manifolds with piecewise smooth boundary.
Math. USSR-Sb. 21(1), 91–135 (1973)
16. A. Komech, Elliptic differential equations with constant coefficients in a cone. Mosc. Univ.
Math. Bull. 29(2), 140–145 (1974)
482 A. Merzon et al.

17. A. Komech, N. Mauser, A. Merzon, On Sommerfeld representation and uniqueness in

scattering by wedges. Math. Methods Appl. Sci. 28(2), 147–183 (2005)
18. A. Komech, A. Merzon, Limiting amplitude principle in the scattering by wedges. Math.
Methods Appl. Sci. 29, 1147–1185 (2006)
19. A. Komech, A. Merzon, J. De la Paz Mendez, Time-dependent scattering of generalized plane
waves by wedges. Math. Methods Appl. Sci. 38, 4774–4785 (2015)
20. A. Komech, A. Merzon, A. Esquivel Navarrete, J. De La Paz Méndez, T. Villalba Vega,
Sommerfeld’s solution as the limiting amplitude and asymptotics for narrow wedges. Math.
Methods Appl. Sci. 42(15), 4957–4970 (2019)
21. A. Komech, A. Merzon, P. Zhevandrov, A method of complex characteristics for elliptic
problems in angles and its applications. Am. Math. Soc. Transl. 206(2), 125–159 (2002)
22. A. Merzon, A. Komech, J. De la Paz Méndez, T. Villalba Vega, On the Keller–Blank solution to
the scattering problem of pulses by wedges. Math. Methods Appl. Sci. 38, 2035–2040 (2015)
23. R. Nagem, M. Zampolli, G. Sandri, A. Sommerfeld (eds.), in Mathematical Theory of
Diffraction. Progress in Mathematical Physics, vol. 35 (Birkhäuser, Boston, 2004)
24. F. Oberhettinger, On the diffraction and reflection of waves and pulses by wedges and corners.
J. Res. Natl. Bur. Stand. 61(2), 343–365 (1958)
25. A. Peters, J. Stoker, A uniqueness and a new solution for Sommerfeld’s and other diffraction
problems. Commun. Pure Appl. Math. 7(3), 565–585 (1954)
26. K. Rottbrand, Time-dependent plane wave diffraction by a half-plane: explicit solution for
Rawlins’ mixed initial boundary value problem. Z. Angew. Math. Mech. 78(5), 321–335 (1998)
27. K. Rottbrand, Exact solution for time-dependent diffraction of plane waves by semi-infinite
soft/hard wedges and half-planes. Z. Angew. Math. Mech. 79, 763–774 (1999)
28. V. Smirnov, S. Sobolev, Sur une méthode nouvelle dans le probléme plan des vibrations
élastiques. Trudy Seismol. Inst. Acad. Nauk SSSR 20, 1–37 (1932)
29. S. Sobolev, Theory of diffraction of plane waves. Trudy Seismol. Inst. Acad. Nauk SSSR 41(1),
75–95 (1934)
30. S. Sobolev, General theory of diffraction of waves on Riemann surfaces. Tr. Fiz.-Mat. Inst.
Steklova 9, 39–105 (1935) [Russian]. English translation: S.L. Sobolev, General theory of
diffraction of waves on Riemann surfaces, in Selected Works of S.L. Sobolev, vol. I (Springer,
New York, 2006), pp. 201–262
31. S. Sobolev, Some questions in the theory of propagations of oscillations Chap. XII, in
Differential and Integral Equations of Mathematical Physics, ed. by F. Frank, P. Mizes
(Moscow, Leningrad, 1937), pp. 468–617 [Russian]
32. A. Sommerfeld, Mathematische theorie der diffraction. Math. Ann. 47, 317–374 (1896)
33. A. Sommerfeld, Optics (Lectures on Theoretical Physics), vol. 4 (Academic Press, New York,
1954)
34. F.-O. Speck, From Sommerfeld diffraction problems to operator factorisation. Constr. Math.
Anal. 2, 183–216 (2019)
On the Operator Jensen-Mercer
Inequality

H. R. Moradi, S. Furuichi, and M. Sababheh

Abstract Mercer’s inequality for convex functions is a variant of Jensen’s inequal-

ity, with an operator version that is still valid without operator convexity. This
paper is two folded. First, we present a Mercer-type inequality for operators
without assuming convexity nor operator convexity. Yet, this form refines the known
inequalities in the literature. Second, we present a log-convex version for operators.
We then use these results to refine some inequalities related to quasi-arithmetic
means of Mercer’s type for operators.

Keywords Jensen-Mercer operator inequality · Log-convex functions · Operator

quasi-arithmetic mean

Mathematics Subject Classification (2010) Primary 47A63; Secondary 47A64,

46L05, 47A60

The author (S.F.) was partially supported by JSPS KAKENHI Grant Number 16K05257.

H. R. Moradi
Young Researchers and Elite Club, Mashhad Branch, Islamic Azad University, Mashhad, Iran
e-mail: [email protected]
S. Furuichi ()
Department of Information Science, College of Humanities and Sciences, Nihon University,
Setagaya-ku, Tokyo, Japan
e-mail: [email protected]
M. Sababheh
Department of Basic Sciences, Princess Sumaya University for Technology, Amman, Jordan
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 483

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_24
484 H. R. Moradi et al.

1 Introduction

Recall that a function f : I ⊆ R → R is said to be convex on the interval I , if it

satisfies the Jensen inequality
 n
 
n
f wi xi ≤ wi f (xi ), (1.1)
i=1 i=1

for all choices of positive scalars w1 , . . . , wn with ni=1 wi = 1 and xi ∈ I . It is
well known that this general form is equivalent to the same inequality when n = 2.
In 2003, Mercer found a variant of (1.1), which reads as follows.
Theorem 1.1 ([7, Theorem 1.2]) If f is a convex function on [m, M], then


n 
n
f M +m− wi xi ≤ f (M) + f (m) − wi f (xi ), (1.2)
i=1 i=1
n
for all xi ∈ [m, M] and all wi ∈ [0, 1] (i = 1, . . . , n) with i=1 wi = 1.
There are many versions, variants and generalizations for the inequality (1.2);
see for example [1, 2, 9].
It is customary in the field of Mathematical inequalities to extend scalar
inequalities, like (1.1) and (1.2), to operators on Hilbert spaces. For this end, we
adopt the following notations. Let H and K be Hilbert spaces, B (H ) and B (K )
be the C ∗ -algebras of all bounded operators on the appropriate Hilbert space. An
operator A ∈ H is called self-adjoint if A = A∗ , where A∗ denotes the adjoint
operator of A. If A ∈ H , the notation A ≥ 0 will be used to declare that A is
positive, in the sense that Ax, x ≥ 0 for all x ∈ H . If Ax, x > 0 for all non
zero x ∈ H , we write A > 0, and we say than that A is positive definite. On the
class of self-adjoint operators, the ≤ partial order relation is well known, where we
write A ≤ B if B − A ≥ 0, when A, B are self-adjoint.
In studying operator inequalities, the notion of spectrum cannot be avoided. If
A ∈ H , the spectrum of A is defined by

σ (A) = {λ ∈ C : A − λ1H is not invertible},

where 1H denotes the identity operator on H . Finally, in these terminologies, a

linear map  : B (H ) → B (K ) is said to be positive if  (A) ≥ 0 whenever
A ≥ 0 and  is called unital if  (1H ) = 1K .
Recall that a continuous function f : I → R is said to be operator convex if
 
A+B f (A) + f (B)
f ≤
2 2
On the Operator Jensen-Mercer Inequality 485

for all self-adjoint A, B ∈ B (H ) and σ (A), σ (B) ⊂ I. This is equivalent to the

Jensen operator inequality, valid for the self-adjoint operators Ai whose spectra are
in the interval I ,
 n
 n n
f wi Ai ≤ wi f (Ai ), wi > 0, wi = 1. (1.3)
i=1 i=1 i=1

It is evident that a convex function is not necessarily operator convex, and the
function f (x) = x 4 provides such an example. Thus, a convex function does not
necessarily satisfy the operator Jensen inequality (1.3). However, it turns out that
a convex function satisfies the following operator version of the Mercer inequality
(1.2).
Theorem 1.2 ([5, Theorem 1]) Let A1 , . . . , An ∈ B (H ) be self-adjoint operators
in [m, M] and let 1 , . . . , n : B (H ) → B (K ) be positive linear
with spectra
maps with ni=1 i (1H ) = 1K . If f : [m, M] ⊆ R → R is a convex function,
then

 n
f (M + m) 1K − i (Ai )
i=1


n
≤ (f (M) + f (m)) 1K − i (f (Ai )). (1.4)
i=1

Further, in the same reference, the following series of inequalities was proved


n
f (M + m)1K − i (Ai )
i=1

≤ (f (M) + f (m)) 1K
n 
i=1 i(Ai ) − M1K m1K − ni=1 i (Ai )
+ f (m) + f (M)
M−m M −m

n
≤ (f (M) + f (m))1K − i (f (Ai )).
i=1

Later, related and analogous results have been established in [3, 4, 6].
Our main goal of this article is to present a refinement of the operator inequality
(1.4) without using convexity of f . Rather, using the idea by Mićić et al. [8], we
assume a boundedness condition on f . Then a discussion of log-convex version of
Mercer’s operator inequality will be presented.
486 H. R. Moradi et al.

2 Main Results

In this section we present our main results in two parts. In the first part, we discuss
the twice differentiable case, then we discuss the log-convex case.

2.1 Twice Differentiable Functions

We begin with the non-convex version of Theorem 1.2. We use the following symbol
in this paper.
(i) Asa = (A1 , . . . , An ), where Ai ∈ B (H ) are self-adjoint operators with
σ (Ai ) ⊆ [m, M] for some scalars 0 < m < M.
(ii) + = (1 , . . . , n ), where i : B (H ) → B (K ) are positive linear maps.
Theorem 2.1 Let A1 , . . . , An ∈ B (H ) be self-adjoint operators with spectra in
n M] and let 1 , . . . , n : B (H ) → B (K ) be positive linear maps with
[m,
i=1 i (1H ) = 1K . If f : [m, M] ⊆ R → R is a continuous twice differentiable
function such that α ≤ f ≤ β with α, β ∈ R, then


n
(f (M) + f (m)) 1K − i (f (Ai )) − βJ (m, M, Asa , + )
i=1


n
≤f (M + m)1K − i (Ai ) (2.1)
i=1


n
≤ (f (M) + f (m)) 1K − i (f (Ai )) − αJ (m, M, Asa , + ), (2.2)
i=1

where


n
J (m, M, Asa , + ) := (M + m) i (Ai ) − Mm1K
i=1
⎛ ⎞
1⎝ 
n 2

n  
− i (Ai ) + i A2i ⎠ ≥ 0.
2
i=1 i=1

Proof Notice that for any convex function f and m ≤ t ≤ M, we have

 
M −t t −m
f (t) = f m+ M ≤ Lf (t) , (2.3)
M −m M−m
On the Operator Jensen-Mercer Inequality 487

where
M −t t −m
Lf (t) := f (m) + f (M). (2.4)
M −m M −m

Letting
α 2
gα (t) := f (t) − t (m ≤ t ≤ M),
2

we observe that g is convex noting the assumption α ≤ f . Applying (2.3) to the

function g, we have g(t) ≤ Lg (t), which leads to

α 
f (t) ≤ Lf (t) − (M + m) t − Mm − t 2 . (2.5)
2
Since m ≤ M + m − t ≤ M, we can replace t in (2.5) with M + m − t, to get
α 
f (M + m − t) ≤ L0 (t) − (M + m) t − Mm − t 2 ,
2
where

L0 (t) := L(M + m − t) = f (M) + f (m) − Lf (t).

Using functional calculus for the operator


n
m1K ≤ i (Ai ) ≤ M1K ,
i=1

we infer that


n
f (M + m)1K − i (Ai )
i=1
 n

≤ L0 i (Ai )
i=1
⎧  n ⎫
α⎨
2⎬

n 
− (M + m) i (Ai ) − Mm1K − i (Ai ) . (2.6)
2⎩ ⎭
i=1 i=1

On the other hand, by applying functional calculus for the operator

m1H ≤ Ai ≤ M1H
488 H. R. Moradi et al.

in (2.5), we get
α 
f (Ai ) ≤ Lf (Ai ) − (M + m) Ai − Mm1H − A2i .
2
Applying the positive linear maps i and adding in the last inequality yield


n
i (f (Ai ))
i=1


n
≤ L0 i (Ai )
i=1
 
α 
n 
n
− (M + m) i (Ai ) − Mm1K − i (A2i ) . (2.7)
2
i=1 i=1

Combining the two inequalities (2.6) and (2.7), we get (2.2).

Finally we give the proof of J (m, M, i , Ai ) ≥ 0. Since

m1H ≤ Ai ≤ M1H ,

we have

(M1H − Ai )(Ai − m1H ) ≥ 0

which implies

(M + m)Ai − mM1H − A2i ≥ 0.

Thus we have

(M + m)(Ai ) − mM(1H ) − (A2i ) ≥ 0.

Taking a summation on i = 1, · · · , n of this inequality with taking an account for

 n
i=1 i (1H ) = 1K , we obtain


n 
n
(M + m) i (Ai ) − Mm1K − (A2i ) ≥ 0. (2.8)
i=1 i=1
n
Further, noting that m ≤ i=1 i (Ai ) ≤ M, we also have
 2

n 
n
(M + m) i (Ai ) − mM1K − i (Ai ) ≥ 0. (2.9)
i=1 i=1
On the Operator Jensen-Mercer Inequality 489

Adding (2.8) and (2.9) and dividing by 2, we obtain J (m, M, Asa , + ) ≥ 0.

The inequality (2.1) follows similarly by taking into account that

β 
Lf (t) − (M + m)t − Mm − t 2 ≤ f (t), m ≤ t ≤ M.
2
The details are left to the reader. This completes the proof.
In the following example, we present the advantage of using twice differentiable
functions in Theorem 2.1.
Example 2.2 Let f (t) = sin t (0 ≤ t ≤ 2π),
π 
0
A= 4
π
0 2

and  (A) = 12 T r [A]. Actually the function f (t) = sin t is concave on [0, π].
Letting m = π4 and M = π2 , we obtain

0.9238 ≈ f ((M + m) − (A)) ≮ f (M) + f (m) − (f (A)) ≈ 0.8535.

That is, (1.4) may fail without the convexity assumption. However, by considering
the weaker assumptions assumed in Theorem 2.1, we get

0.9238 ≈f ((M + m) − (A))

f (M) + f (m) − (f (A))

1 
− α (M + m)(A) − Mm − (A)2 + (A2 )
2
≈0.9306,

since f (t) = − sin t which gives α = −1.

To better understand the relation between Theorems 1.2 and 2.1, we present
the following remark, where we clarify how the first theorem is retrieved from the
second.
Remark 2.3 The inequality (2.2) in Theorem 2.1 with an assumption on a twice
differentiable function f such that α ≤ f ≤ β for α, β ∈ R gives a better upper
bound of

n
f (M + m) 1K − i (Ai )
i=1
490 H. R. Moradi et al.

than that in (1.4), since J (m, M, Asa , + ) ≥ 0, if we take α ≥ 0. Additionally to

this result, we obtained a reverse type inequality (2.1) which gives a lower bound of


n
f (M + m) 1K − i (Ai ) .
i=1

2.2 Log-Convex Functions

We conclude this section by presenting Mercer-type operator inequalities for log-

convex functions. Recall that a positive function defined on an interval I (or, more
generally, on a convex subset of some vector space) is called log-convex if log f (x)
is a convex function of x. We observe that such functions satisfy the elementary
inequality

f ((1 − v) a + vb) ≤ [f (a)]1−v [f (b)]v , 0≤v≤1

for any a, b ∈ I . f is called log-concave if the inequality above is reversed (that is,
when f1 is log-convex). By virtue of the arithmetic-geometric mean inequality, we
have

f ((1 − v)a + vb) ≤ [f (a)]1−v [f (b)]v ≤ (1 − v)f (a) + vf (b), (2.10)

which implies convexity of log-convex functions. This double inequality is of

special interest since (2.10) can be written as
M−t t−m
f (t) ≤ [f (m)] M−m [f (M)] M−m ≤ Lf (t) , m≤t ≤M (2.11)

where Lf (t) is as in (2.4).

Manipulating the inequality (2.11), we have the following extension of Theo-
rem 1.2 to the context of log-convex functions. The proof is left to the reader.
Theorem 2.4 Let all the assumptions of Theorem 1.2 hold except that f :
[m, M] → (0, ∞) is log-convex. Then


n
f (M + m)1K − i (Ai )
i=1
n 
 (A )−m1K M1K − n i (Ai )
i=1 i i i=1
≤ [f (m)] M−m [f (M)] M−m


n
≤ (f (M) + f (m)) 1K − i (f (Ai )).
i=1
On the Operator Jensen-Mercer Inequality 491

3 Applications

In this section, we present some applications of the main results that we have shown
so far. First, we review and introduce the notations.
(i) A+ = (A1 , . . . , An ), where Ai ∈ B (H ) are positive invertible operators with
σ (Ai ) ⊆ [m, M] for some scalars 0 < m < M.
(ii) + = (1 , . . . , n ), where i : B (H ) → B (K ) are positive linear maps.
(iii) C ([m, M]) is the set of all real valued continuous functions on an interval
[m, M].
We also need to remind the reader that a function f ∈ C([m, M]) is called operator
monotone increasing (or operator increasing for short) if f (A) ≤ f (B) whenever
A, B are self-adjoint operators with spectra in [m, M] and such that A ≤ B. That
is, when f preserves the order of self-adjoint operator. A function f ∈ C([m, M])
is said to be operator decreasing if −f is operator monotone.
The so called operator quasi-arithmetic mean of Mercer’s type was defined in [5]
as follows:

 n
Mϕ A+ , + := ϕ −1 (ϕ (M) + ϕ (m)) 1K − i (ϕ (Ai )) .
i=1

In this reference, the following result was shown.

Theorem 3.1 Let ϕ, ψ ∈ C ([m, M]) be two strictly monotonic functions.
(i) If either ψ ◦ ϕ −1 is convex and ψ −1 is operator increasing, or ψ ◦ ϕ −1 is
concave and ψ −1 is operator decreasing, then
 
M ψ A+ , + .
ϕ A+ , + ≤ M (3.1)

(ii) If either ψ ◦ ϕ −1 is concave and ψ −1 is operator increasing, or ψ ◦ ϕ −1 is

convex and ψ −1 is operator decreasing, then the inequality in (3.1) is reversed.
By virtue of Theorem 2.1, we have the following extension of this result.
Theorem 3.2 Let ϕ, ψ ∈ C ([m, M]) be two strictly monotonic functions and ψ ◦
ϕ −1 is twice differentiable function.

(i) If α ≤ ψ ◦ ϕ −1 with α ∈ R and ψ −1 is operator monotone, then

ϕ A+ , +
M
  
≤ ψ −1 ψ M ψ A+ , + − αK m, M, ϕ, A+ , + , (3.2)
492 H. R. Moradi et al.

where

K m, M, ϕ, A+ , +

n
:= (ϕ(M) + ϕ(m)) i (ϕ(Ai )) − ϕ(M)ϕ(m)1K
i=1
⎛ ⎞
1⎝ 
n 2

n  
− i (ϕ(Ai )) + i ϕ(Ai )2 ⎠ .
2
i=1 i=1


(ii) If ψ ◦ ϕ −1 ≤ β with β ∈ R and ψ −1 is operator monotone, then the reverse
inequality is valid in (3.2) with β instead of α.
Proof Let f = ψ ◦ ϕ −1 in (2.2) and replace Ai , m and M with ϕ (Ai ), ϕ (m) and
ϕ (M) respectively. This implies
  
ϕ A+ , +
ψ M ψ A+ , +
≤ψ M − αK m, M, ϕ, A+ , + .

Since ψ −1 is operator monotone, the first conclusion follows immediately. The other
case follows in a similar manner from (2.1).
Similarly, Theorem 2.4 implies the following version.
Theorem 3.3 Let ϕ, ψ ∈ C([m, M]) be two strictly monotonic functions. If ψ ◦ϕ −1
is log-convex function and ψ −1 is operator increasing, then

ϕ (A+ , + )
M
 n 
ϕ(M)1K − n
i=1 i (ϕ(Ai ))−ϕ(m)1K i=1 i (ϕ(Ai ))
−1
≤ψ [ψ(m)] ϕ(M)−ϕ(m) [ψ(M)] ϕ(M)−ϕ(m)


ψ A+ , + .
≤M

Remark 3.4 By choosing appropriate functions ϕ and ψ, and making suitable

substitutions, the above results imply some improvements of certain inequalities
governing operator power mean of Mercer’s type. We leave the details of this idea
to the interested reader as an application of our main results.
In the end of the article, we show the example such that there is no relationship
between inequalities in Theorems 3.2 and 3.3. Here, we restrict ourselves to the
power function f (t) = t p with p < 0.
On the Operator Jensen-Mercer Inequality 493

Example 3.5 It is sufficient to compare (2.5) and the first inequality of (2.11). We
take m = 1 and M = 3. Setting

M −t p t − m p p(p − 1)M p−2  

g(t) = m + M − (M + m)t − Mm − t 2
M −m M −m 2
 M−t t−m
p
− m M−m M M−m .

Some calculations show that g(2) ≈ −0.0052909 when p = −0.2, while g(2) ≈
0.0522794 when p = −1. We thus conclude that there is no ordering between the
RHS of inequality in (2.5) and the RHS of first inequality of (2.11).

Acknowledgments The authors would like to thank the referees for their careful and insightful
comments to improve our manuscript. The authors are also grateful to Dr. Trung Hoa Dinh for
fruitful discussion and revising the manuscript.

References

1. S. Abramovich, J. Barić, J. Pečarić, A variant of Jessen’s inequality of Mercer’s type for

superquadratic functions. J. Inequal. Pure Appl. Math. 9(3) (2008), Article 62
2. E. Anjidani, M.R. Changalvaiy, Reverse Jensen-Mercer type operator inequalities. Electron J.
Linear Algebra, 31, 87–99 (2016)
3. J. Barić, A. Matković, J. Pečarić, A variant of the Jensen-Mercer operator inequality for
superquadratic functions. Math. Comput. Modelling 51, 1230–1239 (2010)
4. S. Ivelić, A. Matković, J. Pečarić, On a Jensen-Mercer operator inequality. Banach J. Math.
Anal. 5, 19–28 (2011)
5. A. Matković, J. Pečarić, I. Perić, A variant of Jensen’s inequality of Mercer’s type for operators
with applications. Linear Algebra Appl. 418, 551–564 (2006)
6. A. Matković, J. Pečarić, I. Perić, Refinements of Jensen’s inequality of Mercer’s type for
operator convex functions. Math. Ineq. Appl. 11, 113–126 (2008)
7. A.McD. Mercer, A variant of Jensen’s inequality. J. Inequal. Pure Appl. Math. 4(4) (2003),
Article 73
8. J. Mićić, H.R. Moradi, S. Furuichi, Choi-Davis-Jensen’s inequality without convexity. J. Math.
Inequal. 12, 1075–1085 (2018)
9. H.R. Moradi, M.E. Omidvar, M. Adil Khan, K. Nikodem, Around Jensen’s inequality for
strongly convex functions. Aequationes Math. 92, 25–37 (2018)
A Numerical Approach for
Approximating Variable-Order
Fractional Integral Operator

Somayeh Nemati

Abstract In this work, we propose a numerical method to find an approximation of

the variable-order integral of a given function using a generalization of the modified
hat functions. First, an operational matrix of the basis functions corresponding to
the variable-order integral operator is introduced. Then, using this matrix and an
approximation of the given function, we find an approximation of the variable-order
integral operator of the function. An error estimate is proved. Two test examples are
included to show the efficiency and accuracy of our new technique. Finally, this new
technique is used to solve the variable-order differential equations numerically and
some illustrative problems are provided to validate the applicability and accuracy of
this new scheme.

Keywords Variable-order integral operator · Modified hat functions ·

Operational matrix

Mathematics Subject Classification (2010) 34A08; 65M70

1 Introduction

A recent generalization of the theory of fractional calculus is to let the fractional

order of the derivatives to be of variable order. In [16], authors have investigated
operators when the order of fractional derivative is a variable on time. The non-local
properties of systems are more visible with variable-order fractional calculus, and

S. Nemati ()
Department of Applied Mathematics, Faculty of Mathematical Sciences, University of
Mazandaran, Babolsar, Iran
Center for Research and Development in Mathematics and Applications (CIDMA),
Department of Mathematics, University of Aveiro, Aveiro, Portugal
e-mail: [email protected]; [email protected]

© Springer Nature Switzerland AG 2021 495

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_25
496 S. Nemati

many real world phenomena in physics, mechanics, control, and signal processing
have been described by this approach [2, 13–15].
Hat functions are defined on the interval [0, 1] and are continuous with shape hats
[17]. In the last few years, a modification of hat functions has been introduced and
used for solving some different kinds of problems. For instance: two-dimensional
linear Fredholm integral equations [5], integral equations of Stratonovich–Volterra
[6] and Volterra–Fredholm type [7], fractional integro-differential equations [8],
fractional pantograph differential equations [9], fractional optimal control problems
[10] and systems of fractional differential equations [11]. These functions are
defined on the interval [0, T ], with T > 0. In this paper, we consider a general
definition of these basis functions on [a, b], with b > a, called generalized modified
hat functions (GMHFs). These basis functions are utilized to present a numerical
method to approximate the variable-order fractional integral of a given function. For
y : [a, b] → R and a real function α(t) such that 0 < α(t) ≤ 1 for all t ∈ [a, b],
the left Riemann–Liouville fractional integral operator of order α(t) of y is defined
by (see, e.g., [1])
 t
α(t ) 1
a It y(t) = (t − s)α(t )−1y(s)ds, t > a,
(α(t)) a

where  is the Euler gamma function.

The organization of this paper is as follows. In Sect. 2, we give a new general
definition of the modified hat functions and introduce some of their main properties.
Section 3 is devoted to introducing a numerical method based on the GMHFs
to compute the variable-order fractional integral of a given function. The error
analysis of this new approximation is provided and the integral of two test functions
are computed to illustrate the accuracy of the method. An application of this new
technique for solving variable-order fractional differential equations is proposed in
Sect. 4. Finally, we conclude the paper in Sect. 5.

2 Definition and Function Approximation

In this section, we introduce a generalization of the modified hat functions and

present some of their main properties. To this aim, the interval [a, b] is divided
into m subintervals [a + ih, a + (i + 1)h], i = 0, 1, 2, . . . , m − 1, of equal lengths
h, where h = b−am and m ≥ 2 is an even integer number. A (m + 1)-set of GMHFs
are defined on the interval [a, b] as follows
⎧
⎪
⎨ 2h2 (t − (a + h)) (t − (a + 2h)) ,
1

ψ0 (t) = if a ≤ t ≤ a + 2h, (2.1)

⎪
⎩ 0, otherwise,
A Numerical Approach for Approximating Variable-Order Integral 497

if i is odd and 1 ≤ i ≤ m − 1:
⎧
⎪ −1
⎨ h2 (t − (a + (i − 1)h)) (t − (a + (i + 1)h)) ,
ψi (t) = if a + (i − 1)h ≤ t ≤ a + (i + 1)h, (2.2)
⎪
⎩ 0, otherwise,

if i is even and 2 ≤ i ≤ m − 2:
⎧ 1
⎪
⎪ (t − (a + (i − 1)h)) (t − (a + (i − 2)h)) ,
⎪
⎪ 2h2
⎪
⎨ it a + (i − 2)h ≤ t ≤ a + ih,
ψi (t) = 2h1 2 (t − (a + (i + 1)h)) (t − (a + (i + 2)h)) , (2.3)
⎪
⎪
⎪
⎪ if a + ih ≤ t ≤ a + (i + 2)h,
⎪
⎩
0, otherwise,

and
⎧
⎪
⎨ 2h2 (t − (b − h)) (t − (b − 2h)) ,
1

ψm (t) = if b − 2h ≤ t ≤ b, (2.4)
⎪
⎩ 0, otherwise.

The GMHFs build a set of linearly independent functions in the space L2 [a, b].
Figure 1 displays a set of the GMHFs obtained by m = 4 and defined on the interval
[−1, 1].
The following properties can be easily investigated using the definition of the
GMHFs:
 
m
1, i = j,
ψi (a + j h) = ψi (t) = 1, t ∈ [a, b],
0, i = j,
i=0


0, if i is even and |i − j | ≥ 3,
ψi (t)ψj (t) =
0, if i is odd and |i − j | ≥ 2.

Any arbitrary function y ∈ L2 [a, b] may be approximated in terms of the

GMHFs as [5]


m
y(t) 2 ym (t) = yi ψi (t) = Y T "(t), (2.5)
i=0

where

"(t) = [ψ0 (t), ψ1 (t), . . . , ψm (t)]T , (2.6)

 498

Fig. 1 Plot of the GMHFs with m = 4 on [−1, 1]

S. Nemati
A Numerical Approach for Approximating Variable-Order Integral 499

and

Y = [y0 , y1 , . . . , ym ]T ,

with

yi = y(a + ih). (2.7)

We have the following theorem for the error of the function approximation using
the GMHFs.
Theorem 2.1 If a function y ∈ C 3 ([a, b]) is approximated by the family of first
(m + 1) GMHFs as defined by (2.1)–(2.4), then

|y(t) − ym (t)| = O(h3 ),

where h = b−a
m , and ym is defined by (2.5).
Proof It can be proved in a same way as done in [8, Theorem 4.1].

3 Numerical Method

In this section, we propose a numerical method based on the GMHFs for computing
the left Riemann–Liouville variable-order integral of a given function.

3.1 Variable-Order Integration Rule

In order to introduce the method, we first obtain an operational matrix of variable-

order integration for the GMHFs basis vector. To do this, we apply the Riemann–
Liouville fractional integral operator of order 0 < α(t) ≤ 1, for all t ∈ [a, b], to the
function ψi (t), and obtain
 t
α(t ) 1
a It ψi (t) = (t − s)α(t )−1ψi (s)ds, i = 0, 1, . . . , m.
(α(t)) a

Then, we may expand these functions using the GMHFs as follows:

α(t )

m
a It ψi (t) 2 α
a Pij ψj (t), i = 0, 1, . . . , m,
j =0
500 S. Nemati

α(t )
wherein according to (2.7), a Pijα is the value of a It ψi (t) at a + j h. That is
 a+j h
1
α
a Pij = (a + j h − s)α(a+j h)−1 ψi (s)ds, i, j = 0, 1, . . . , m.
(α(a + j h)) a
(3.1)

By substituting the definitions of the functions ψi (t) given by (2.1)–(2.4), we

compute the integral part of (3.1). For i = 0, we obtain
α
a P00 = 0, (3.2)
hα(a+h)
α
a P01 = [α(a + h) (3 + 2α(a + h))] , (3.3)
2(α(a + h) + 3)

and for j > 1,

hα(a+j h)
α
a P0j =
2(α(a + j h) + 3)

× j α(a+j h)+1 (2j − 6 − 3α(a + j h))

+ 2j α(a+j h) (1 + α(a + j h)) (2 + α(a + j h))

−(j − 2)α(a+j h)+1 (2j − 2 + α(a + j h)) . (3.4)

If i is odd, we get for j < i,

α
a Pij = 0, (3.5)

2hα(a+ih)
α
a Pii = (1 + α(a + ih)), (3.6)
(α(a + ih) + 3)

and for j > i,

2hα(a+j h)
α
a Pij =
(α(a + j h) + 3)

× (j − i − 1)α(a+j h)+1(j − i + 1 + α(a + j h))

−(j − i + 1)α(a+j h)+1(j − i − 1 − α(a + j h)) . (3.7)

A Numerical Approach for Approximating Variable-Order Integral 501

Finally, if i is even, we have for j < i − 1,

α
a Pij =0, (3.8)

hα(a+(i−1)h)
α
a Pi,i−1 = [−α(a + (i − 1)h)] , (3.9)
2(α(a + (i − 1)h) + 3)
hα(a+ih)
α
a Pii = 2α(a+ih)+1(2 − α(a + ih)) , (3.10)
2(α(a + ih) + 3)
hα(a+(i+1)h)
α
a Pi,i+1 =
2(α(a + (i + 1)h) + 3)

× 3α(a+(i+1)h)+1(4 − α(a + 8i + 1)h))

− 6(2 + α(a + (i + 1)h)) , (3.11)

and for j > i + 1,

hα(a+j h)
α
a Pij =
2(α(a + j h) + 3)

× (j − i + 2)α(a+j h)+1(2j − 2i + 2 − α(a + j h))

− 6(j − i)α(a+j h)+1(2 + α(a + j h))

− (j − i − 2)α(a+j h)+1(2j − 2i − 2 + α(a + j h)) . (3.12)

As a result, we can write

α(t )
a It "(t) 2 Paα(t ) "(t), (3.13)

α(t )
where Pa is an upper Hessenberg matrix of dimension (m + 1) × (m + 1)
called operational matrix of variable-order integration of order α(t) in the Riemann-
Liouville sense, and is given using (3.2)–(3.12) as follows:
⎡ α Pα α α α α ⎤
0 a P01 a 02 a P03 a P04 ... a P0(m−1) a P0m
⎢0 Pα Pα α α α α ⎥
⎢ a 11 a 12 a P13 a P14 ... a P1(m−1) a P1m ⎥
⎢ ⎥
⎢ 0 a P21
α Pα
a 22
α
a P23
α
a P24 ... α
a P2(m−1)
α
a P2m ⎥
⎢ ⎥
⎢0 0 0 α
a P33
α
a P34 ... α
a P3(m−1)
α
a P3m ⎥
⎢ ⎥
Paα(t ) = ⎢0 0 α α α α ⎥. (3.14)
⎢ 0 a P43 a P44 ... a P4(m−1) a P4m ⎥
⎢. . ⎥
⎢. . .. .. .. .. .. ⎥
⎢. . . . . . . ⎥
⎢ ⎥
⎣0 0 0 0 0 α
. . . a P(m−1)(m−1) a P(m−1)m ⎦
α

0 0 0 0 0 α
. . . a Pm(m−1) α
a Pmm
502 S. Nemati

In order to compute the left Riemann–Liouville variable-order integral operator of

an arbitrary function y ∈ L2 [a, b], first, we approximate this function using the
GMHFs as (2.5). Then, using (3.13), we have

α(t ) α(t ) α(t )

a It y(t) 2 a It ym (t) = Y T a It "(t) 2 Y T Paα(t ) "(t).

3.2 Error Bound

The aim of this section is to give an error bound for the numerical estimation of the
left Riemann–Liouville variable-order integral operator of an arbitrary function in
the sense of L2 -norm. To this aim, we consider the following L2 -norm for a function
f ∈ L2 [a, b]:

 b  12
2
f 2 = |f (t)| dt .
a

With the help of Theorem 2.1 we obtain the following result.

Theorem 3.1 Assume that y ∈ C 3 ([a, b]). Suppose that the left Riemann–Liouville
α(t )
variable-order integral operator of y is approximated by Y T Pa "(t) where "(t)
α(t )
and Pa are given by (2.6) and (3.14), respectively, and the elements of Y are
yi = y(a + ih), i = 0, 1, . . . , m, with h = b−a
m . Then,
; ;
; α(t ) ;
;a It y − Y T Paα(t ) " ; = O(h3 ).
2

Proof By assuming 0 < α(t) ≤ 1, for all t ∈ [a, b], we consider the following
α(t )
definition of the norm for the operator a It ,

a Itα(t )2 = sup a Itα(t )f 2 .

f 2 =1

We show that this operator is bounded. Using Schwarz’s inequality, we obtain

; ; ;  t ;
; α(t ) ; ; 1 ;
;a It f ; = ;; (t − s) α(t )−1
f (s)ds ;
2 (α(t)) a ;
2
;  t ;
; 1 ;
≤ f 2 ;
; (α(t)) (t − s)α(t )−1ds ; ;
a 2
; ;
; (t − a)α(t ) ;
; ;
=; ; ,
; (α(t) + 1) ;
2
A Numerical Approach for Approximating Variable-Order Integral 503

where we have used f 2 = 1. Since we have (t) > 0.8 for all t > 0, we can
write (α(t) + 1) > 45 . Therefore, we get
; ;
; (t − a)α(t ) ; 5; ;
; ; ; ;
; ; < ;(t − a)α(t ); ,
; (α(t) + 1) ; 4 2
2

which together with 0 < α(t) ≤ 1 help us to have one of the following statements:
1. If 0 < t − a < 1, then
; ;
; (t − a)α(t ) ;
; ; 5 5 1
; ; < 12 = (b − a) 2 .
; (α(t) + 1) ; 4 4
2

2. If 1 ≤ t − a ≤ b − a, then
; ;
; (t − a)α(t ) ;
; ; 5 5 3
; ; < t − a2 = √ (b − a) 2 .
; (α(t) + 1) ; 4 4 3
2

Hence there is a positive constant dependent on a and b so that

a Itα(t ) 2 < C.

α(t )
We use the boundedness property of a It to continue the proof. Using
Theorem 2.1, we have

y − ym 2 = O(h3 ) (3.15)

and
; ;
; T α(t ) ;
;Y a It " − Y T Paα(t ) " ; = O(h3 ). (3.16)
2

Finally, using (3.15) and (3.16), we obtain

; ; ; ;
; α(t ) ; ; α(t ) α(t ) α(t ) ;
;a It y − Y T Paα(t ) " ; = ;a It y − a It ym + a It ym − Y T Paα(t ) " ;
2 2
; ; ; ;
; α(t ) α(t ) ; ; α(t ) T α(t ) ;
≤ ;a It y − a It ym ; + ;a It ym − Y Pa " ;
2 2
; ; ; ;
; α(t ) ; ; T α(t ) ;
≤ ;a It ; y − ym 2 + ;Y a It " − Y T Paα(t ) " ;
2 2

= O(h ),
3

which completes the proof.

504 S. Nemati

3.3 Test Examples

In this section, in order to illustrate the accuracy and applicability of the proposed
method, we consider two functions and employ the method for computing the
left Riemann–Liouville variable-order integral of these functions. We note that the
method was carried out using Mathematica 12. To show the accuracy, the l 2 norm
of the error and the convergence order are defined, respectively, by
 1
1 
m 2
Em = (J (ti ) − Jm (ti ))2 , &m = log2 (Em /E2m ) ,
m
i=1

where ti = a + ih. Furthermore, in computation of Em , we have

J (t) = a Itα(t )y(t), Jm (t) = Y T Paα(t ) "(t)

as the exact and approximate integrals of the function y, respectively.

Example 3.1 Suppose that y(t) = sin(t), α(t) = t, a = 0 and b = 1. The exact left
Riemann–Liouville variable-order integral of y is
 
tt t 3 t t2
J (t) = 1 F2 1; 1 + , + ; − ,
(1 + t)(t) 2 2 2 4

where p Fq is the generalized hypergeometric function defined by

∞
  (a1 )k . . . (ap )k zk
p Fq a1 , . . . , ap ; b1 , . . . , bq ; z = ,
(b1 )k . . . (bq )k k!
k=0

and (a)k = a(a + 1) . . . (a + k − 1) is Pochhammer symbol.

Example 3.2 Consider y(t) = et , α(t) = sin(t), a = 1 and b = 3. In this case, we
have
 
(sin(t), t − 1)
J (t) = et 1 − ,
(sin(t))

where (c, t) is the incomplete gamma function defined by

 ∞
(β, t) = s β−1 e−s ds.
t

We have employed the present method to obtain approximations of the left

Riemann–Liouville variable-order integral operator by considering the information
given in Examples 3.1 and 3.2. Numerical results are displayed in Table 1 and Fig. 2.
A Numerical Approach for Approximating Variable-Order Integral 505

Table 1 Numerical results Example 3.1 Example 3.2

for the l 2 norm of the error
m Em &m Em &m
and the convergence order
2 1.99e−3 3.63 2.23e−1 3.42
4 1.61e−4 3.45 2.08e−2 3.46
8 1.47e−5 3.36 1.89e−3 3.39
16 1.43e−6 3.30 1.80e−4 3.35
32 1.45e−7 3.26 1.76e−5 3.33
64 1.51e−8 3.24 1.75e−6 3.31
128 1.60e−9 3.22 1.77e−7 3.28
256 1.72e−10 3.18 1.82e−8 2.51
512 1.90e−11 – 3.19e−9 –

Fig. 2 The l 2 norm of the error for some selected values of m in logarithmic scale

In Table 1, the l 2 norm of the error and the convergence order are presented which
confirm the O(h3 ) accuracy of this method. Furthermore, in Fig. 2, the results for
the l 2 norm of the error are plotted in a logarithmic scale.
506 S. Nemati

4 Application to Variable-Order Fractional Differential

Equations

In this section, a numerical method based on the use of operational matrix of

variable-order integration for the GMHFs basis vector is introduced for solving
variable-order fractional differential equations. Moreover, some examples are pro-
vided to show the accuracy of this method.

4.1 Method of Solution

Consider the variable-order fractional initial value problem


C D α(t ) y(t) = f (t, y(t)), 0 < α(t) ≤ 1, a ≤ t ≤ b,
a t (4.1)
y(a) = y0 ,

where y is the unknown function, f : [a, b] × R −→ R is a given function, y0 is a

α(t )
real constant and C
a Dt denotes the left Caputo fractional derivative of order α(t)
defined by Almeida et al. [1]:
 t
C α(t ) 1
a Dt y(t) = (t − s)−α(t )y (s)ds, t > a.
(1 − α(t)) a

Let y(t) ∈ C 1 [a, b] be the exact solution of (4.1), then an approximation of y (t)
based on the GMHFs can be considered as follows:

y (t) = C T "(t), (4.2)

where C is a vector with unknown parameters ci , i = 0, 1, . . . , m, and "(t) is given

by (2.6). Therefore, an approximation of y can be given using (3.13), (4.2) and the
initial condition as
 t
y(t) = y (s)ds + y(a) 2 C T a It1 "(t) + y0
a

2 (C T Pa1 + YaT )"(t) = Y T "(t), (4.3)

where

Ya = [y0 , y0 , . . . , y0 ]T , Y = (C T Pa1 + YaT )T = [y0 , y1 , . . . , ym ]T .

In a similar way, we have

C α(t ) 1−α(t )
a Dt y(t) = a It y (t) 2 C T Pa1−α(t ) "(t). (4.4)
A Numerical Approach for Approximating Variable-Order Integral 507

On the other hand, an approximation of the function f : ([a, b] × R) → R based

on the GMHFs is given by


n 
n
f (t, y(t)) 2 f (ti , y(ti ))ψi (t) = f (ti , yi )ψi (t) = F (, Y )"(t), (4.5)
i=0 i=0

with ti = a + ih, and

F (, Y ) = [f (a, y0 ), f (a + h, y1 ), . . . , f (b, ym )] .

Now, by substituting (4.4) and (4.5) into (4.1), we get

C T Pa1−α(t ) = F (, Y ), (4.6)

which is a system of nonlinear algebraic equations in terms of the unknown

parameters of the vector C. The first equation of the system (4.6) is a degenerate
equation. By letting α(t) → 1, we impose the condition y (a) = f (a, y0) for its
solvability. With this condition, we obtain c0 = f (a, y0 ). In order to determine the
remaining unknown entries of C, we substitute c0 = f (a, y0 ) into equations of
(4.6). By solving the resulting system, an approximation of the function y is given
by (4.3).

4.2 Numerical Simulations

Here, we employ the proposed method for solving three examples to demonstrate
the applicability and accuracy of our new method.
Example 4.3 As the first example, consider the problem (4.1) with 0 ≤ t ≤ 1 and
[4]

(4) (3)
f (t, y) = t 3 + t 3−α(t ) + t 2 + t 2−α(t ) − y.
(4 − α(t)) (3 − α(t))

The exact solution of this problem is y(t) = t 3 + t 2 . By considering α(t) =

sin t, this problem has been solved by different values of m. Plots of the error
for m = 4, 8, 16, 32 are displayed in Fig. 3. In [4], this problem has been
solved by considering different fixed values of α ∈ (0, 1) using a high order
numerical methods based on a second-degree compound quadrature formula and
the convergence order of the error at t = 1 has been reported. Since the proposed
method in this paper gives the exact solution at t = 1, we report the L2 -norm of the
error with different m for α = 0.75, and compare the convergence order with the
method of [4] in Table 2. Also, we can see in this table the CPU times (in seconds),
which have been obtained on a 2.5 GHz Core i7 personal computer with 16 GB of
 508

Fig. 3 Plots of the error with m = 4, 8, 16, 32 for Example 4.3 with α(t) = sin t
S. Nemati
A Numerical Approach for Approximating Variable-Order Integral 509

Table 2 Numerical results for the L2 -error and convergence order for Example 4.3 with α = 0.75
Present method Method of [4]
m L2 -error Convergence order CPU time Convergence order
10 2.76e−4 3.00 0.000 1.79
20 3.45e−5 3.00 0.016 2.07
40 4.31e−6 3.00 0.016 2.17
80 5.39e−7 2.99 0.062 2.21
160 6.74e−8 3.00 0.234 2.23
320 8.42e−9 3.00 0.906 2.24
640 1.05e−9 – 2.828 –

RAM using Mathematica 12. The Mathematica function FindRoot was used for
solving the resulting systems.

Example 4.4 Consider the problem (4.1) with 0 ≤ t ≤ 1 and

  7
 t 2 −α(t )
9
2
f (t, y) =   + t 7 sin(t) − y 2 sin(t).
 92 − α(t)

The exact solution of this problem is t 7/2 . By setting α(t) = 1 − 0.5 exp(−t), the
numerical results of employing the proposed method for solving this example are
given in Fig. 4 and Table 3.
Example 4.5 As the last example, consider the following Riccati fractional differ-
ential equation [3, 12]

C α(t )
0 Dt y(t) = 2y(t) − y 2 (t) + 1, 0 < α ≤ 1, 0 ≤ t ≤ 1,

with initial condition y(0) = 0. When α(t) = 1, the exact solution is

 √
√ √ 1 2−1
y(t) = 1 + 2 tanh 2t + ln √ .
2 2+1

By taking α(t) = 1, a comparison between the numerical results obtained by the

present method with m = 50, the modified homotopy perturbation method [12]
using the fourth-order term and Chebyshev wavelet method [3] using m̂ = 192 is
provided at some selected points in Table 4. Furthermore, the numerical results for
y(t) with α = 0.65, 0.75, 0.85, 0.95 and m = 10 together with the exact solution
for α = 1 are plotted in Fig. 5. As it could be expected, when α is close to 1, the
numerical solution is close to the exact solution of the case α = 1.
 510

Fig. 4 Plots of the error with m = 4, 8, 16, 32 for Example 4.4 with α(t) = 1 − 0.5 exp(−t)
S. Nemati
A Numerical Approach for Approximating Variable-Order Integral 511

Table 3 Numerical results m L2 -error Convergence order CPU time

for Example 4.4 with
α(t) = 1 − 0.5 exp(−t) 2 5.53e−2 3.04 0.000
4 6.72e−3 3.01 0.000
8 8.34e−4 3.00 0.016
16 1.04e−4 3.00 0.031
32 1.30e−5 3.00 0.078
64 1.63e−6 3.00 0.687
128 2.04e−7 – 3.953

Table 4 Numerical results for Example 4.5 with α(t) = 1

t Method of [12] Method of [3] Present method Exact solution
0 0 0.000001 0 0
0.1 0.110294 0.110311 0.110295 0.110295
0.2 0.241965 0.241995 0.241977 0.241977
0.3 0.395106 0.395123 0.395105 0.395105
0.4 0.568115 0.567829 0.567812 0.567812
0.5 0.757564 0.756029 0.756014 0.756014
0.6 0.958259 0.953576 0.953566 0.953566
0.7 1.163459 1.152955 1.152949 1.152949
0.8 1.365240 1.346365 1.346364 1.346364
0.9 1.554960 1.526909 1.526911 1.526911
1 1.723810 1.689494 1.689498 1.689498

Fig. 5 Numerical solutions with m = 10 and different values of α together with the exact solution
with α = 1 for Example 4.5
512 S. Nemati

5 Concluding Remarks

A new numerical technique has been introduced for computing the left Riemann–
Liouville variable-order integral of a given function. A generalized class of the
modified hat functions (GMHFs) has been considered and used to suggest our new
scheme. The given function is easily expanded using the GMHFs. An operational
matrix of variable-order integral of the basis vector is computed and used to
approximate the integral of the function under consideration. The convergence
order of our numerical method is proved and confirmed by the results of two
illustrative examples. Finally, this new technique is employed to solve variable-
order differential equations and the numerical results demonstrate the efficiency of
the method.

Acknowledgement The author is grateful to two anonymous referees for several positive and
constructive comments, which helped her to improve the manuscript.

References

1. R. Almeida, D. Tavares, D.F.M. Torres, The Variable-Order Fractional Calculus of Variations

(Springer, Berlin, 2019)
2. C.F.M. Coimbra, C.M. Soon, M.H. Kobayashi, The variable viscoelasticity operator. Ann.
Phys. 14, 378–389 (2005)
3. Y. Li, Solving a nonlinear fractional differential equation using Chebyshev wavelets. Commun.
Nonlinear Sci. Numer. Simul. 15, 2284–2292 (2010)
4. Z. Li, Y. Yan, N.J. Ford, Error estimates of a high order numerical method for solving linear
fractional differential equations. Appl. Numer. Math. 114, 201–220 (2017)
5. F. Mirzaee, E. Hadadiyan, Numerical solution of linear Fredholm integral equations via two
dimensional modification of hat functions. Appl. Math. Comput. 250, 805–816 (2015)
6. F. Mirzaee, E. Hadadiyan, Approximation solution of nonlinear Stratonovich Volterra integral
equations by applying modification of hat functions. J. Comput. Appl. Math. 302, 272–284
(2016)
7. F. Mirzaee, E. Hadadiyan, Numerical solution of Volterra–Fredholm integral equations via
modification of hat functions. Appl. Math. Comput. 280, 110–123 (2016)
8. S. Nemati, P.M. Lima, Numerical solution of nonlinear fractional integro-differential equations
with weakly singular kernels via a modification of hat functions. Appl. Math. Comput. 327,
79–92 (2018)
9. S. Nemati, P. Lima, S. Sedaghat, An effective numerical method for solving fractional
pantograph differential equations using modification of hat functions. Appl. Numer. Math. 131,
174–189 (2018)
10. S. Nemati, P. Lima, D.F.M Torres, A numerical approach for solving fractional optimal control
problems using modified hat functions. Commun. Nonlinear Sci. Numer. Simulat. 78, 104849
(2019)
11. S. Nemati, D.F.M Torres, A new spectral method based on two classes of hat functions for
solving systems of fractional differential equations and an application to respiratory syncytial
virus infection. Soft Comput. (2020). https://doi.org/10.1007/s00500-019-04645-5
12. Z. Odibat, S. Momani, Modified homotopy perturbation method: application to quadratic
Riccati differential equation of fractional order. Chaos Solitons Fractals 36, 167–174 (2008)
A Numerical Approach for Approximating Variable-Order Integral 513

13. T. Odzijewicz, A.B. Malinowska, D.F.M. Torres, Fractional variational calculus of variable
order, in Advances in Harmonic Analysis and Operator Theory. Operator Theory: Advances
and Applications, vol. 229 (Birkhäuser, Basel, 2013)
14. P.W. Ostalczyk, P. Duch, D.W. Brzeziński, D. Sankowski, Order functions selection in the
variable, fractional-order PID controller, in Advances in Modelling and Control of Non-integer-
Order Systems. Lecture Notes in Electrical Engineering, vol. 320 (2015), pp. 159–170.
15. M.R. Rapaić, A. Pisano, Variable-order fractional operators for adaptive order and parameter
estimation. IEEE Trans. Autom. Control 59(3), 798–803 (2014)
16. S.G. Samko, B. Ross, Integration and differentiation to a variable fractional order. Integral
Transform. Spec. Funct. 1, 277–300 (1993)
17. M.P. Tripathi, V.K. Baranwal, R.K. Pandey, O.P. Singh, A new numerical algorithm to solve
fractional differential equations based on operational matrix of generalized hat functions.
Commun. Nonlinear Sci. Numer. Simul. 18, 1327–1340 (2013)
Langlands Reciprocity for C ∗-Algebras

Igor V. Nikolaev

Abstract We introduce a C ∗ -algebra AV of a variety V over the number field K

and a C ∗ -algebra AG of a reductive group G over the ring of adeles of K. Using
Pimsner’s Theorem, we construct an embedding AV +→ AG , where V is a G-
coherent variety, e.g. the Shimura variety of G. The embedding is an analog of the
Langlands reciprocity for C ∗ -algebras. It follows from the K-theory of the inclusion
AV ⊂ AG that the Hasse-Weil L-function of V is a product of the automorphic L-
functions corresponding to irreducible representations of the group G.

Keywords Langlands program · Serre C ∗ -algebra

Mathematics Subject Classification (2010) Primary 11F70; Secondary 46L85

1 Introduction

The Langlands conjectures say that all zeta functions are automorphic [9]. In
this note we study (one of) the conjectures in terms of the C ∗ -algebras [5].
Namely, denote by G(AK ) a reductive group G over the ring of adeles AK of a
number field K and by G(K) its discrete subgroup over K. The Banach algebra
L1 (G(K)\G(AK )) consists of the integrable complex-valued functions endowed
with the operator norm. The addition of functions f1 , f2 ∈ L1 (G(K)\G(AK )) is
defined pointwise and the multiplication is given by the convolution product:

(f1 ∗ f2 )(g) = f1 (gh−1 )f2 (h)dh. (1.1)
G(K)\G(AK )

I. V. Nikolaev ()
Department of Mathematics, St. John’s University, New York, NY, USA

© Springer Nature Switzerland AG 2021 515

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_26
516 I. V. Nikolaev

Consider the enveloping C ∗ -algebra, AG , of the algebra L1 (G(K)\G(AK )); we

refer the reader to [5, Section 13.9] for details of this construction. The algebra AG
encodes all unitary irreducible representations of the locally compact group G(AK )
induced by G(K). Such representations are related to the automorphic cusp forms
and non-abelian class field theory [8]. The algebra AG has an amazingly simple
structure. Namely, let us assume G ∼ = GLn . Then AG is a stationary AF-algebra,
see Lemma 3.1; such an algebra is defined by a positive integer matrix B ∈ SLn (Z)
[2] and its K-theory is well understood [6].
Let V be a complex projective variety. For an automorphism σ : V → V
and an invertible sheaf L of the linear forms on V , one can construct a twisted
homogeneous coordinate ring B(V , L, σ ) of the variety V , i.e. a non-commutative
r ing such that:

Mod (B(V , L, σ )) / T ors ∼

= Coh (V ),

where Mod is the category of graded left modules over the graded ring B(V , L, σ ),
T ors the full subcategory of Mod of the torsion modules and Coh the category of
quasi-coherent sheaves on the variety V [18, p. 180]. The norm-closure of a self-
adjoint representation of the ring B(V , L, σ ) by the linear operators on a Hilbert
space is called the Serre C ∗ -algebra of V [10]. In what follows, we shall focus on
the case when V is defined over a number field K, i.e. V is an arithmetic variety.
The corresponding Serre C ∗ -algebra is denoted by AV . The Hasse-Weil L-function
of V was calculated in [10] in terms of the K-theory of algebra AV .
It is known that the Langlands philosophy does not distinguish between the
arithmetic and automorphic objects [9]. Therefore one can expect a regular map
between the C ∗ -algebras AV and AG , provided V is a G-coherent variety, see
Definition 1.1. We prove that such a map is an embedding AV +→ AG . To give
an exact statement, we shall need the following notions. The i-th trace cohomology
{Htir (V ) | 0 ≤ i ≤ 2 dimC V } of an arithmetic variety V is an additive abelian
subgroup of R obtained from a canonical trace on the Serre C ∗ -algebra of V [10].
Likewise, the group K0 (AG ) of the stationary AF-algebra AG is an additive abelian
subgroup of R [6, Chapter 6].
Definition 1.1 The arithmetic variety V is called G-coherent, if

Htir (V ) ⊆ K0 (AG ) for all 0 ≤ i ≤ 2 dimC V . (1.2)

Remark 1.2 If V ∼ = Sh (G, X) is the Shimura variety corresponding to the Shimura

datum (G, X) [4], then V is a G-coherent variety. This remark follows from an
adaption of the argument for the Shimura curves considered in Sect. 4. To put it
simple, the arithmetic variety V is G-coherent if for all 0 ≤ i ≤ 2 dimC V the
number fields ki := Htir (V ) ⊗ Q are subfields of (or coincide with) a number
field K := K0 (AG ) ⊗ Q. A quick example are elliptic curves with complex
multiplication, see Proposition 4.2.
Langlands Reciprocity for C ∗ -Algebras 517

Theorem 1.3 There exists a canonical embedding AV +→ AG , where V is a G-

coherent variety.
Remark 1.4 Theorem 1.3 can be viewed as an analog of the Langlands reciprocity
for C ∗ -algebras. In other words, the coordinate ring AV of a G-coherent variety V
is a sub-algebra of the algebra AG .
An application of Theorem 1.3 is as follows. Recall that to each arithmetic variety V
one can attach the Hasse-Weil (motivic) L-function. Likewise, to each irreducible
representation of the group G(AK ) one can attach an automorphic (standard) L-
function, see [8] and [9]. Theorem 1.3 implies one of the conjectures of [9].
Corollary 1.5 The Hasse-Weil L-function of a G-coherent variety V is a product
of the automorphic L-functions.
The paper is organized as follows. The definitions and preliminary results can be
found in Sect. 2. Theorem 1.3 and Corollary 1.5 are proved in Sect. 3. An example
is constructed in Sect. 4.

2 Preliminaries

This section is a brief account of preliminary facts involved in our paper; we refer
the reader to [2, 5, 9, 18].

2.1 AF-Algebras

A C ∗ -algebra is an algebra A over C with a norm a → a and an involution

a → a ∗ such that it is complete with respect to the norm and ab ≤ a b and
a ∗ a = a 2  for all a, b ∈ A. Any commutative C ∗ -algebra is isomorphic to the
algebra C0 (X) of continuous complex-valued functions on some locally compact
Hausdorff space X; otherwise, A represents a noncommutative topological space.
An AF-algebra (Approximately Finite C ∗ -algebra) is defined to be the norm
closure of an ascending sequence of finite dimensional C ∗ -algebras Mn , where
Mn is the C ∗ -algebra of the n × n matrices with entries in C. Here the index
n = (n1 , . . . , nk ) represents the semi-simple matrix algebra Mn = Mn1 ⊕· · ·⊕Mnk .
The ascending sequence mentioned above can be written as
ϕ1 ϕ2
M1 −→ M2 −→ . . . ,

where Mi are the finite dimensional C ∗ -algebras and ϕi the homomorphisms

between such algebras. If ϕi = Const, then the AF-algebra A is called stationary;
such an algebra defines and is defined by a shift automorphism σϕ : A → A
518 I. V. Nikolaev

corresponding to a map i → i + 1 on ϕi [6, p. 37]. The homomorphisms

ϕi can be arranged into a graph as follows. Let Mi = Mi1 ⊕ · · · ⊕ Mik and
Mi = Mi1 ⊕ · · · ⊕ Mik be the semi-simple C ∗ -algebras and ϕi : Mi → Mi the
homomorphism. One has two sets of vertices Vi1 , . . . , Vik and Vi1 , . . . , Vik joined
by brs edges whenever the summand Mir contains brs copies of the summand Mis
under the embedding ϕi . As i varies, one obtains an infinite graph called the Bratteli
diagram of the AF-algebra. The matrix B = (brs ) is known as a partial multiplicity
matrix; an infinite sequence of Bi defines a unique AF-algebra.
For a unital C ∗ -algebra A, let V (A) be the union (over n) of projections in
the n × n matrix C ∗ -algebra with entries in A; projections p, q ∈ V (A) are
equivalent if there exists a partial isometry u such that p = u∗ u and q = uu∗ .
The equivalence class of projection p is denoted by [p]; the equivalence classes of
orthogonal projections can be made to a semigroup by putting [p] + [q] = [p + q].
The Grothendieck completion of this semigroup to an abelian group is called the
K0 -group of the algebra A. The functor A → K0 (A) maps the category of unital
C ∗ -algebras into the category of abelian groups, so that projections in the algebra A
correspond to a positive cone K0+ ⊂ K0 (A) and the unit element 1 ∈ A corresponds
to an order unit u ∈ K0 (A). The ordered abelian group (K0 , K0+ , u) with an order
unit is called a dimension group; an order-isomorphism class of the latter we denote
by (G, G+ ).
If A is an AF-algebra, then its dimension group (K0 (A), K0+ (A), u) is a
complete isomorphism invariant of algebra A [7]. The order-isomorphism class
(K0 (A), K0+ (A)) is an invariant of the Morita equivalence of algebra A, i.e. an
isomorphism class in the category of finitely generated projective modules over A.

2.2 Trace Cohomology

Let V be an n-dimensional complex projective variety endowed with an automor-

phism σ : V → V and denote by B(V , L, σ ) its twisted homogeneous coordinate
ring, see [18]. Let R be a commutative graded ring, such that V = Spec (R). Denote
by R[t, t −1 ; σ ] the ring of skew Laurent polynomials defined by the commutation
relation bσ t = tb for all b ∈ R, where bσ is the image of b under automorphism σ .
It is known, that R[t, t −1 ; σ ] ∼
= B(V , L, σ ).
Let H be a Hilbert space and B(H) the algebra of all bounded linear operators on
H. For a ring of skew Laurent polynomials R[t, t −1 ; σ ], consider a homomorphism:

ρ : R[t, t −1 ; σ ] −→ B(H). (2.1)

Recall that B(H) is endowed with a ∗-involution; the involution comes from the
scalar product on the Hilbert space H. We shall call representation (2.1) ∗-coherent,
if (i) ρ(t) and ρ(t −1 ) are unitary operators, such that ρ ∗ (t) = ρ(t −1 ) and (ii) for
all b ∈ R it holds (ρ ∗ (b))σ (ρ) = ρ ∗ (bσ ), where σ (ρ) is an automorphism of ρ(R)
induced by σ . Whenever B = R[t, t −1 ; σ ] admits a ∗-coherent representation, ρ(B)
Langlands Reciprocity for C ∗ -Algebras 519

is a ∗-algebra; the norm closure of ρ(B) is a C ∗ -algebra [5]. We shall denote it by

AV and refer to AV as the Serre C ∗ -algebra of variety V .
Let K be the C ∗ -algebra of all compact operators on H. We shall write τ :
AV ⊗ K → R to denote the canonical normalized trace on AV ⊗ K, i.e. a
positive linear functional of norm 1 such that τ (yx) = τ (xy) for all x, y ∈
AV ⊗ K, see [1, p. 31]. Denote by C(V ) the C ∗ -algebra of complex-valued
functions on the Hausdorff space V . Because AV is a crossed product C ∗ -algebra
of the form AV ∼ = C(V )  Z [12, Lemma 5.3.2], one can use the Pimsner-
Voiculescu six term exact sequence for the crossed products, see e.g. [1, p. 83]
for the details. Thus one gets the short exact sequence of the algebraic K-groups:
i∗
0 → K0 (C(V )) → K0 (AV ) → K1 (C(V )) → 0, where the map i∗ is induced
by the natural embedding of C(V ) into AV . We have K0 (C(V )) ∼ = K 0 (V ) and
∼ −1 −1
K1 (C(V )) = K (V ), where K and K are the topological K-groups of V , see
0

[1, p.80]. By the Chern character formula, one gets K 0 (V ) ⊗ Q ∼ = H even (V ; Q)

−1 ∼
and K (V ) ⊗ Q = H (V ; Q), where H
odd even odd
(H ) is the direct sum of even
(odd, resp.) cohomology groups of V . Notice that K0 (AV ⊗ K) ∼ = K0 (AV ) because
of a stability of the K0 -group with respect to tensor products by the algebra K, see
e.g. [1, p. 32]. One gets the commutative diagram in Fig. 1, where τ∗ denotes a
homomorphism induced on K0 by the canonical trace τ on the C ∗ -algebra AV ⊗ K.
Since H even (V ) := ⊕ni=0 H 2i (V ) and H odd (V ) := ⊕ni=1 H 2i−1 (V ), one gets for
each 0 ≤ i ≤ 2n an injective homomorphism τ∗ : H i (V ) −→ R.
Definition 2.1 By an i-th trace cohomology group Htir (V ) of variety V one
understands the abelian subgroup of R defined by the map τ∗ .

2.3 Langlands Reciprocity

Let V be an n-dimensional complex projective variety over a number field K;

consider its reduction V (Fp ) modulo the prime ideal P ⊂ K over a non-ramified
prime p. Recall that the Weil zeta function is defined as:
 ∞
 tr
Zp (t) = exp |V (Fpr )| , r ∈ C, (2.2)
r
r=1

i∗
H even (V ) ⊗ Q K0 (AV ⊗ K) ⊗ Q H odd (V ) ⊗ Q

τ∗

Fig. 1 The trace cohomology

520 I. V. Nikolaev

where |V (Fpr )| is the number of points of variety V (Fpr ) defined over the field with
pr elements. It is known that:

P1 (t) . . . P2n−1 (t)

Zp (t) = , (2.3)
P0 (t) . . . P2n (t)

where P0 (t) = 1 − t, P2n = 1 − pn t and each Pi (t) forE1 ≤ i ≤ 2n − 1 is

a polynomial with integer coefficients, such that Pi (t) = (1 − αij t) for some
i
algebraic integers αij of the absolute value p 2 . Consider an infinite product:

/ L1 (s, V ) . . . L2n−1 (s, V )

L(s, V ) := Zp (p−s ) = , (2.4)
p
L0 (s, V ) . . . L2n (s, V )

E
where Li (s, V ) = p Pi (p−s ); the L(s, V ) is called the Hasse-Weil (or motivic)
L-function of V .
On the other hand, if K is a number fieldE then the adele ring AK of K is a
locally compact subring of the direct product Kv taken over all places v of K;
the AK is endowed with a canonical topology. Consider a reductive group G(AK )
over AK ; the latter is a topological group with a canonical discrete subgroup G(K).
Denote by L2 (G(K)\G(AK )) the Hilbert space of all square-integrable complex-
valued functions on the homogeneous space G(K) \ G(AK ) and consider the right
regular representation R of the locally compact group G(AK ) by linear operators
on the space L2 (G(K)\G(AK )) given by formula (1.1). It is well known, that each
irreducible component π of the unitary representation R can be written in the form
π = ⊗πv , where v are all unramified places of K. Using the spherical functions,
one gets an injection πv → [Av ], where [Av ] is a conjugacy class of matrices in the
group GLn (C). The automorphic L-function is given by the formula:
/ 4 5−1
L(s, π) = det In − [Av ](Nv)−s , s ∈ C, (2.5)
v

where Nv is the norm of place v; we refer the reader to [9, p. 170] and [8, p. 201]
for details of this construction.
The following conjecture relates the Hasse-Weil and automorphic L-functions.
Conjecture 2.2 ([9]) For each 0 ≤ i ≤ 2n there exists an irreducible representa-
tion πi of the group G(AK ), such that Li (s, V ) ≡ L(s, πi ).
Langlands Reciprocity for C ∗ -Algebras 521

3 Proofs

3.1 Proof of Theorem 1.3

We shall split the proof in two lemmas.

Lemma 3.1 The algebra AG is isomorphic to a stationary AF-algebra.
Proof Let A× K be the idele group, i.e. a group of invertible elements of the adele ring
AK . Denote by Gal (K ab |K) the Galois group of the maximal abelian extension
K ab of the number field K. The Artin reciprocity says that there exists a continuous
isomorphism:

K × \A×
K /CK −→ Gal (K |K),
ab
(3.1)

where CK is the closure of the image in K × \A× K of the identity connected

component of the archimedean part K ∞ of the A× K.
Recall that Gal (K ab |K) is a profinite abelian group, i.e. a topological group
isomorphic to the inverse limit of finite abelian groups. It follows from the Artin
reciprocity (3.1), that K × \A×
K /CK is also a profinite abelian group. Since every
k
finite abelian group is a product of the cyclic groups Z/pi i Z, we can write the
× ×
group K \AK /CK in the form:

lm 
/ 
K × \A× ∼ Z/piki Z ,
K /CK = lim
← −
(3.2)
i=1

k
where m → ∞. Notice that the cyclic group Z/pi i Z can be embedded into the
k k
finite field Fqi , where qi = pi i . Thus the group GLn (Z/pi i Z) is correctly defined
and from (3.2) one gets an isomorphism

/
lm
GLn (K × \A× ∼
K /CK ) = lim
← −
GLn (Fqi ), (3.3)
i=1

E j
where GLn (Fqi ) is a finite group of order n−1 j =0 (qi − qi ) and such a group is no
n

longer abelian. In particular, it follows from (3.3) that the GLn (K × \A× K /CK ) is a
profinite group.
(i) Let us show that the group GLn (K × \A× K /CK ) being profinite implies that the
AG is an AF-algebra. Indeed, if G is a finite group then the group algebra C[G]
has the form

C[G] ∼
= Mn1 (C) ⊕ · · · ⊕ Mnh (C),
522 I. V. Nikolaev

where ni are degrees of the irreducible representations of G and h is the total

number of such representations [16, Proposition 10]. In view of (3.3), we have

GLn (K × \A× ∼
K /CK ) = lim
←
G,
− i
(3.4)

where Gi is a finite group. Consider a group algebra

C[Gi ] ∼
= Mn(i)
1
(C) ⊕ · · · ⊕ Mn(i)
h
(C) (3.5)

corresponding to Gi . Notice that the C[Gi ] is a finite-dimensional C ∗ -algebra.

The inverse limit (3.4) defines an ascending sequence of the finite-dimensional
C ∗ -algebras of the form

lim Mn(i) (C) ⊕ · · · ⊕ Mn(i) (C). (3.6)

← − 1 h

Since AG is the norm closure of the group algebra C[GLn (K × \A× ∼

K )] =
C[GLn (K × )\GLn (A× K )] [5, Section 13.9], we conclude that there exists a C ∗-

homomorphism h : AG → AG , where AG is an AF-algebra defined by the

limit (3.6). To calculate the kernel of h, recall that CK ∼
= lim Ui , where Ui are
×
←−
×
open subgroups of the group K \AK . We repeat the construction of (3.4)–(3.6)
and obtain an AF-algebra AU . One gets an exact sequence of the C ∗ -algebras
1 → AU → AG → AG → 1. In other words, the AG is an extension of the
AF-algebra AU by the AF-algebra AG . But any such an extension must be an
AF-algebra itself [3]. Item (i) is proved.
(ii) It remains to prove that the AG is a stationary AF-algebra. Indeed, denote by
F rq the Frobenius map, i.e. an endomorphism of the finite field Fq acting
by the formula x → x q . The map F rqi induces an automorphism of the
group GLn (Fqi ). Using formula (3.3), one gets an automorphism of the group
GLn (AK ) and the corresponding group algebra C[GLn (AK )]. Taking the norm
closure of the algebra C[GLn (AK )], we conclude that there exists a non-
trivial automorphism φ of the AF-algebra AG . But the AF-algebra admits an
automorphism φ = ± I d if and only if it is a stationary AF-algebra [6, p. 37].
Thus the algebra AG is a stationary AF-algebra. Lemma 3.1 is proved.
Remark 3.2 It follows from formula (3.4) that the AF-algebra AG is determined by
a partial multiplicity matrix B of rank n, i.e. B ∈ SLn (Z). Consider an isomorphism

AG  Z ∼
= OB ⊗ K,

where the crossed product is taken by the shift automorphism of AG , OB is

the Cuntz-Krieger algebra defined by matrix B and K is the C ∗ -algebra of
compact operators [1, Exercise 10.11.9]. Consider a continuous group of modular
automorphisms {σ t : OB → OB | t ∈ R} acting on the generators s1 , . . . , sn of
the algebra OB by the formula sk → eit sk . Then a pull back of σ t corresponds
Langlands Reciprocity for C ∗ -Algebras 523

to the action of continuous symmetry group GLn (AK ) on the homogeneous space
GLn (K)\GLn (AK ). This observation can be applied to prove Weil’s conjecture on
the Tamagawa numbers.
Lemma 3.3 The algebra AV embeds into the AF-algebra AG , where V is a G-
coherent variety.
Proof We shall use the Pimsner’s Theorem [13, Theorem 7] about an embedding
of the crossed product algebra AV into an AF-algebra. It will develop that the G-
coherence of V implies that the AF-algebra is Morita equivalent to the algebra AG
of Lemma 3.1. We pass to a detailed argument.
Let V be a complex projective variety. Following [13] we shall think of V as
a compact metrizable topological space X. Recall that for a homeomorphism ϕ :
X → X the point x ∈ X is called non-wandering if for each neighborhood U of x
and every N > 0 there exists n > N such that

ϕ n (U ) ∩ U = ∅.

(In other words, the point x does not “wander” too far from its initial position in the
space X.) If each point x ∈ X is a non-wandering point, then the homeomorphism
ϕ is called non-wandering.
Let σ : V → V be an automorphism of finite order of the G-coherent variety V ,
such that the representation (2.1) is ∗-coherent. Then the crossed product

AV = C(V ) σ Z

is the Serre C ∗ -algebra of V . Since σ is of a finite order, it is a non-wandering home-

omorphism of X. In particular, the σ is a pseudo-non-wandering homeomorphism
[13, Definition 2]. Then there exists a unital (dense) embedding

AV +→ A, (3.7)

where A is an AF-algebra defined by the homeomorphism ϕ [13, Theorem 7].

Let us show that the algebra A is Morita equivalent to the AF-algebra AG .
Indeed, the embedding (3.7) induces an injective homomorphism of the K0 -groups

K0 (AV ) +→ K0 (A). (3.8)

As explained in Sect. 2.2, the map (3.8) defines an inclusion

Htir (V ) ⊆ K0 (A). (3.9)

On the other hand, the trace cohomology of a G-coherent variety V must satisfy an
inclusion

Htir (V ) ⊆ K0 (AG ). (3.10)

524 I. V. Nikolaev

Let b∗ = max0≤i≤2n bi be the maximal Betti number of variety V . Then in

formulas (3.9) and (3.10) the inclusion is an isomorphism, i.e. Ht∗r (V ) ∼
= K0 (A)
and Ht∗r (V ) ∼
= K0 (AG ). One concludes that

K0 (A) ∼
= K0 (AG ). (3.11)

In other words, the AF-algebras A and AG are Morita equivalent. The embedding
AV +→ AG follows from formulas (3.7) and (3.11). Lemma 3.3 is proved.
Theorem 1.3 follows from Lemma 3.3.

3.2 Proof of Corollary 1.5

Corollary 1.5 follows from an observation that the Frobenius action σ (F rpi ) :
Htir (V ) → Htir (V ) extends to a Hecke operator Tp : K0 (AG ) → K0 (AG ),
whenever Htir (AV ) ⊆ K0 (AG ). Let us pass to a detailed argument.
Recall that the Frobenius map on the i-th trace cohomology of variety V is given
by an integer matrix σ (F rpi ) ∈ GLbi (Z), where bi is the i-th Betti number of V ;
moreover,


2n
|V (Fp )| = (−1)i tr σ (F rpi ), (3.12)
i=0

where V (Fp ) is the reduction of V modulo a good prime p [10]. (Notice that
(3.12) is sufficient to calculate the Hasse-Weil L-function L(s, V ) of variety V via
Eq. (2.2); hence the map σ (F rpi ) : Htir (V ) → Htir (V ) is motivic.)
Definition 3.4 Denote by Tpi an endomorphism of K0 (AG ), such that the diagram
in Fig. 2 is commutative, where ι is the embedding (1.2). By Hi we understand the
algebra over Z generated by the Tpi ∈ End (K0 (AG )), where p runs through all but
a finite set of primes.

Fig. 2 Hecke operator Tpi σ(F rpi )

Htr
i
(V ) - Htr
i
(V )

ι ι
? Tpi ?
K0 (AG ) - K0 (AG )
Langlands Reciprocity for C ∗ -Algebras 525

Remark 3.5 The algebra Hi is commutative. Indeed, the endomorphisms Tpi cor-
respond to multiplication of the group K0 (AG ) by the real numbers; the latter
commute with each other. We shall call the {Hi | 0 ≤ i ≤ 2n} an i-th Hecke algebra.
Lemma 3.6 The algebra Hi defines an irreducible representations πi of the group
G(AK ).
Proof Let f ∈ L2 (G(K)\G(AK )) be an eigenfunction of the Hecke operators
Tpi ; in other words, the Fourier coefficients cp of the function f coincide with the
eigenvalues of the Hecke operators Tp up to a scalar multiple. Such an eigenfunction
is defined uniquely by the algebra Hi .
Let Lf ⊂ L2 (G(K)\G(AK )) be a subspace generated by the right translates
of f by the elements of the locally compact group G(AK ). It is immediate (see
e.g. [8, Example on p. 197]), that Lf is an irreducible subspace of the space
L2 (G(K)\G(AK )); therefore it gives rise to an irreducible representation πi of the
locally compact group G(AK ). Lemma 3.6 follows.
Lemma 3.7 L(s, πi ) ≡ Li (s, V ).
Proof Recall that the function Li (s, V ) can be written as
/ −1
Li (s, V ) = det In − σ (F rpi )p−s , (3.13)
p

where σ (F rpi ) ∈ GLbi (Z) is a matrix form of the action of F rpi on the trace
cohomology Htir (V ).
On the other hand, from (2.5) one gets
/ −1
L(s, πi ) = det In − [Aip ]p−s , (3.14)
p

where [Aip ] ⊂ GLn (C) is a conjugacy class of matrices corresponding to the

irreducible representation πi of the group G(AK ). As explained, for such a
representation we have an inclusion Tpi ∈ [Aip ]. But the action of the Hecke operator
Tpi is an extension of the action of σ (F rpi ) on Htir (V ), see Fig. 2. Therefore

σ (F rpi ) = [Aip ] (3.15)

for all but a finite set of primes p. Comparing formulas (3.13)–(3.15), we get that
L(s, πi ) ≡ Li (s, V ). Lemma 3.7 follows.
Corollary 1.5 follows from Lemma 3.7 and formula (2.4).
526 I. V. Nikolaev

4 Example

We shall illustrate Theorem 1.3 and Corollary 1.5 for the group

G∼
= GL2 (AK ),
√
where K = Q( D) is a real quadratic field.
Proposition 4.1 K0 (AG ) ∼
= Z + Zω, where
 √
1+ D
, if D ≡ 1 mod 4,
ω= √ 2 (4.1)
D, if D ≡ 2, 3 mod 4.

Proof By Lemma 3.1 and Remark 3.2, the AG is a stationary AF-algebra given by
partial multiplicity matrix B ∈ SL2 (Z). In particular, K0 (AG ) ∼
= Z + Zω, where
ω ∈ Q(λB ), where λB is the Perron-Frobenius eigenvalue of matrix B. Moreover,
by the construction End (K) ∼ = End (K0 (AG )), where End is the endomorphism
ring of the corresponding Z-module. But End (K) ∼ = OK , where OK is the ring
of integers of K. Thus, λB ∈ K and ω is given by formula (4.1). Proposition 4.1
follows.
Proposition 4.2 Let ECM ∼ = C/Ok be an elliptic curve with complex
√ multiplication
by the ring of integers of the imaginary quadratic field k = Q( −D). Then ECM is
a G-coherent variety of the group G ∼ = GL2 (AK ).
Proof The noncommutative torus Aθ is a C ∗ -algebra generated by the unitary
operators u and v satisfying the commutation relation vu = e2πiθ uv for a constant
θ ∈ R [15]. The Serre C ∗ -algebra of an elliptic curve Eτ ∼ = C/(Z + Zτ ) is
isomorphic to Aθ for any {τ | I m τ > 0}, see [12, Theorem 1.3.1]. In particular
[11], if τ ∈ Ok then

Ht0r (ECM ) = Ht2r (ECM ) ∼
= Z,
(4.2)
Ht1r (ECM ) ∼
= Z + Zω.

Comparing formulas (4.1) and (4.2), one concludes that

Htir (ECM ) ⊆ K0 (AG ),

i.e. the ECM is a G-coherent variety of the group G ∼

= GL2 (AK ). Proposition 4.2 is
proved.
Remark 4.3 The embedding of Aθ into an AF-algebra was initially constructed in
[14].
Proposition 4.4 L(s, ECM ) ≡ L(s,L(s, π1 )
π0 )L(s, π2 ) , where πi are irreducible representa-
tions of the locally compact group GL2 (AK ).
Langlands Reciprocity for C ∗ -Algebras 527

Proof The Hasse-Weil L-function of the ECM has the form:

E −1
p det (I2 − σ (F rp1 )p−s
L(s, ECM ) = , s ∈ C, (4.3)
ζ(s)ζ(s − 1)

where ζ (s) is the Riemann zeta function and the product is taken over the set of
good primes; we refer the reader to formula (3.13). It is immediate that

L(s, π0 ) = ζ(s),
L(s, π2 ) = ζ(s − 1),

where L(s, π0 ) and L(s, π2 ) are the automorphic L-functions corresponding to

the irreducible representations π0 and π2 of the group GL2 (AK ). An irreducible
representation π1 gives rise to an automorphic L-function
/ −1
L(s, π1 ) = det I2 − [A1p ]p−s .
p

But formula (3.15) says that [A1p ] = σ (F rp1 ) and therefore the numerator of (4.3)
coincides with the L(s, π1 ). Proposition 4.4 is proved.
Remark 4.5 Proposition 4.4 can be proved in terms of the Grössencharacters [17,
Chapter II, §10].

Acknowledgments I thank the referees for their interest and helpful comments on the draft of this
paper.

References

1. B. Blackadar, K-Theory for Operator Algebras (MSRI Publications, Springer, 1986)

2. O. Bratteli, Inductive limits of finite dimensional C ∗ -algebras. Trans. Amer. Math. Soc. 171,
195–234 (1972)
3. L.G. Brown, Extensions of AF algebras: the projection lifting problem, in Operator Algebras
and Applications, Proceedings of Symposia in Pure Mathematics, vol. 38 (1982), pp. 175–176
4. P. Deligne, Travaux de Shimura, vol. 244. Séminaire Bourbaki, Lecture Notes in Mathematics
(Springer, Berlin, 1971), pp. 123–165
5. J. Dixmier, C ∗ -Algebras (North-Holland Publishing Company, Amsterdam, 1977)
6. E.G. Effros, Dimensions and C ∗ -algebras, in Board of the Mathematical Sciences, Regional
Conference Series in Mathematics, vol. 46 (AMS, Providence, 1981)
7. G.A. Elliott, On the classification of inductive limits of sequences of semisimple finite-
dimensional algebras. J. Algebra 38, 29–44 (1976)
8. S. Gelbart, An elementary introduction to the Langlands program. Bull. Amer. Math. Soc. 10,
177–219 (1984)
528 I. V. Nikolaev

9. R.P. Langlands, L-functions and automorphic representations, in Proceedings of the ICM 1978,
Helsinki (1978), pp. 165–175
10. I.V. Nikolaev, On traces of Frobenius endomorphisms. Finite Fields Appl. 25, 270–279 (2014)
11. I.V. Nikolaev, On a symmetry of complex and real multiplication. Hokkaido Math. J. 45, 43–51
(2016)
12. I.V. Nikolaev, Noncommutative Geometry. De Gruyter Studies in Mathematics, vol. 66 (De
Gruyter, Berlin, 2017)
13. M.V. Pimsner, Embedding some transformation group C ∗ -algebras into AF-algebras. Ergodic
Theory Dyn. Syst. 3, 613–626 (1983)
14. M.V. Pimsner, D.V. Voiculescu, Imbedding the irrational rotation C ∗ -algebra into an AF-
algebra. J. Oper. Theory 4, 201–210 (1980)
15. M.A. Rieffel, Non-commutative tori – a case study of non-commutative differentiable mani-
folds. Contemp. Math. 105, 191–211 (1990)
16. J.-P. Serre, Représentations Linéaires des Groupes Finis (Hermann, Paris, 1967)
17. J.H. Silverman, Advanced Topics in the Arithmetic of Elliptic Curves. GTM, vol. 151 (Springer,
Berlin, 1994)
18. J.T. Stafford, M. van den Bergh, Noncommutative curves and noncommutative surfaces. Bull.
Amer. Math. Soc. 38, 171–216 (2001)
Compact Sequences in Quasifractal
Algebras

Steffen Roch

Abstract The paper is devoted to the study of compact sequences in quasifractal

algebras. We are particularly interested in the relations between the essential ranks
of fractally restricted sequences (An )|M and the essential rank of the full sequence.
We will also ask whether the dependence of the essential ranks of the fractally
restricted sequences (An )|M on (a coset of) M is continuous, which requires to
provide the fractal variety of the algebra with a suitable topology.

Keywords Fractal restriction · Quasifractal algebras · Continuous fields ·

Continuous trace algebras

Mathematics Subject Classification (2010) Primary 47N40; Secondary 65J10,

46L99

1 Introduction

Let H be a Hilbert space and (Pn )n∈N a sequence of orthogonal projections of finite
rank which converges strongly to the identity operator on H . Let F denote the set
of all bounded sequences (An )n≥1 of operators An ∈ L(im Pn ) and G the set of all
sequences (An ) ∈ F with An  → 0. Provided with the operations

(An ) + (Bn ) := (An + Bn ), (An )(Bn ) := (An Bn ), (An )∗ := (A∗n )

and the norm (An ) := sup An , F becomes a unital C ∗ -algebra and G a closed
ideal of F . Let δ(n) denote the rank of Pn . Then one can identify L(im Pn ) with
the C ∗ -algebra Cn := Mδ(n) (C) of the complex δ(n) × δ(n) matrices, and F and G
can be identified with the direct product and the direct sum of the sequence (Cn ),

S. Roch ()
Department of Mathematics, Technical University of Darmstadt, Darmstadt, Germany
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 529

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_27
530 S. Roch

respectively. In what follows we will think of F and G in this way (although some
results hold for products and sums of sequences of general C ∗ -algebras as well).
The quotient algebra F /G plays a significant role in numerical analysis, which
stems from the observation that asymptotic properties of the sequence (An ) ∈ F
can often be rephrased as a property of the coset (An ) + G. To mention only two
examples, it is

(An ) + GF /G = lim sup An 

n→∞

for every sequence (An ) ∈ F , and the coset (An )+G is invertible in F /G if and only
if the An are invertible for all sufficiently large n and if the norms of their inverses
are uniformly bounded, which is equivalent to saying that (An ) is a stable sequence.
Fractal and quasifractal algebras The correspondence between asymptotic prop-
erties of the sequence (An ) ∈ F and properties of the coset (An ) + G is particularly
close when the sequence (An ) belongs to a fractal subalgebra of F . For example, if
(An ) is a sequence in a fractal subalgebra of F , then the limit lim An  exists, and
the sequence (An ) is stable if one of its (infinite) subsequences is stable.
The perhaps simplest way to define fractal algebras is the following (which is
equivalent to the original definition in [15]):
A C ∗ -subalgebra A of F is fractal if every partial zero sequence in A is a zero sequence,
i.e., if every sequence (An ) ∈ A with lim inf An  = 0 satisfies lim An  = 0.

Not every subalgebra of F is fractal, but it turns out that every separable
subalgebra of F has a fractal restriction. Here is the precise statement of the fractal
restriction theorem from [9] (see [12] for a shorter proof):
If A is a separable C ∗ -subalgebra of F , then there is an infinite subset M of N such that the
algebra A|M of all restricted sequences (An )n∈M is a fractal subalgebra of F |M .

Since every restriction of a separable subalgebra of F is separable again, the

separable subalgebras of F are quasifractal in the sense of the following definition:
A C ∗ -subalgebra A of F is called quasifractal if every restriction of A has a fractal
restriction.

We denote by fr A the set of all infinite subsets M of N for which A|M is a

fractal algebra. Two elements M1 , M2 of fr A are called equivalent, symbolically
written as M1 ∼ M2 , if M1 ∪ M2 ∈ fr A. The relation ∼ is an equivalence relation
(Lemma 4.2 in [13]). We denote the set of all equivalence classes of this relation by
(fr A)∼ and call (fr A)∼ the fractal variety of the algebra A. For example, if A is
fractal, then fr A consists of all infinite subsets of N and (fr A)∼ is a singleton. A
standard example for a fractal algebra is the algebra of the finite sections method
for Toeplitz operators with continuous generating function (see [2] for Toeplitz
operators and their finite sections and Corollary 1.10 in [5] for the fractality result).
Examples for quasifractal algebras can be found in [13].
Compact Sequences in Quasifractal Algebras 531

Compact sequences We will need several notions of the compactness and of the
rank of an element in a general C ∗ -algebra and, in particular, of a sequence in F .
These notions can be subsumed under the following simple scheme.
Let A be a unital C ∗ -algebra and J a non-empty self-adjoint subset of A such
that AJ ⊆ J and J A ⊆ J . We call the subsets with these properties the semi-
ideals of A. Clearly, 0 ∈ J . Every semi-ideal J induces a rank function on A, i.e.,
a function r : A → N ∪ {∞} which satisfies
(a) r(a) = 0 if and only if a = 0,
(b) r(a + b) ≤ r(a) + r(b),
(c) r(ab) ≤ min{r(a), r(b)},
(d) r(a) = r(a ∗ )
for all a, b ∈ A, as follows: Set r(0) := 0. If a nonzero a ∈ A is a finite sum of
elements of J , then r(a) is the smallest positive integer such that a can be written
as a sum of r(a) elements from J . Finally, r(a) := ∞ if a is not a finite sum of
elements of J . The closure of the set of the elements with finite rank is a closed
ideal of A, called the ideal of the compact elements (relative to J ). We will reify
this scheme in several settings:
1. Let A be a unital C ∗ -algebra and J the set of all elements k ∈ A with the
property that for every a ∈ A there is a number α ∈ C such that kak = αk.
Then J is a semi-ideal; the associated rank function is called the algebraic rank
and denoted by alg rank a, and the associated ideal of the compact elements is
denoted by C(A).
2. For A = F , consider the set J of all sequences (Kn ) ∈ F such that rank Kn ≤
1 for all n ∈ N. Then J is a semi-ideal of F ; the associated rank function
is called the sequential rank and denoted by seq rank (An ), and the associated
ideal of the compact elements is denoted by K. The elements of K are called the
compact sequences. It is not hard to show that G ⊆ K and that seq rank (An ) =
supn rank An for every sequence (An ) ∈ F .
3. If A is a C ∗ -subalgebra of F and J is the set of all sequences (Kn ) in A such
that rank Kn ≤ 1 for all n ∈ N, then we denote the corresponding ideal of the
compact sequences by K(A). Clearly, K = K(F ).
4. Let J be the set of all cosets (Kn ) + G of sequences (Kn ) ∈ F with
seq rank (Kn ) ≤ 1. Then J is a semi-ideal of F /G and the corresponding ideal
of the compact elements is nothing but K/G. If r denotes the rank function
associated with J then we call ess rank (An ) := r((An )+G) the essential rank of
(An ) ∈ F . In particular, the sequences of essential rank 0 are just the sequences in
G. The essential rank of a sequence (An ) ∈ F can be characterized as the smallest
integer r ≥ 0 such that (An ) can be written as (Gn ) + (Kn ) with (Gn ) ∈ G and
supn rank Kn = r. The advantage of the essential rank of a sequence (An ) over
its sequential rank is that it depends on the coset of (An ) modulo G only.
See Sections 4.1–4.4 in [11] for some basic facts on compact sequences.
532 S. Roch

Compact Sequences in Fractal Algebras Let A be a unital and fractal C ∗ -algebra

of F . A crucial result ([10]) on compact sequences in fractal algebras is that then

(A ∩ K)/G = C A/G .

Consequently, then (A ∩ K)/G is a dual algebra, i.e., it is ∗ -isomorphic to a direct

sum of ideals K(Ht ) of compact operators on a Hilbert space Ht . There is also a
relation between the algebraic rank of the coset (An ) + G and the essential rank of
(An ) when (An ) is a compact sequence in a fractal algebra A. The general form of
this relation involves the local weights of A which we are not going to introduce
here. This relation takes a particular simple form when A has local weight 1, in
which case

ess rank (An ) = alg rank (An ) + G (1.1)

for every sequence (An ) ∈ A of finite essential rank. One can show that a fractal
algebra A has local weight 1 if and only if

A ∩ K = K(A). (1.2)

We therefore refer to (1.2) as the local weight 1 condition in what follows.

The contents of this paper The goal of this paper is to study compact sequences in
quasifractal algebras. We are particularly interested in the relations of the essential
ranks of the restricted sequences (An )|M with M ∈ fr A to the essential rank of (An )
on the one side and to the algebraic ranks of the cosets (An )|M + G|M on the other
side. We will also ask whether the dependence of the essential ranks of (An )|M on
M∼ is continuous with respect to the topology on (fr A)∼ introduced in [13]. We
will embed the latter question into a broader context and discuss to what extend
(A ∩ K)/G can be viewed of as an algebra with continuous trace.

2 Restrictions of Compact Sequences

In what follows we will often use boldface letters to denote sequences in F . We

start with picking up the first question raised at the end of the introduction: to
what extend do the compactness properties of the fractal restrictions of a sequence
determine the compactness properties of the full sequence? Clearly, if K ∈ F is
a compact sequence and M an infinite subset of N, then the restriction K|M is a
compact sequence in F |M , and

ess rank (K|M ) ≤ ess rank K. (2.1)

Compact Sequences in Quasifractal Algebras 533

Trivially, the converse is also true: If every restriction of a sequence K is compact,

then K is compact. Remarkably, already the fractal restrictions of K will do the job.
Proposition 2.1 Let A be a quasifractal C ∗ -subalgebra of F and K ∈ A. Then
(a) K is compact if and only if every restriction K|M with M ∈ fr A is compact.
(b) if K is of finite essential rank,

ess rank K = max ess rank (K|M ). (2.2)

M∈fr A

Proof We shall employ a criterion from [11] which characterizes the compactness
of a sequence (Kn ) in terms of the singular values of the matrices Kn . Thus, let
1 (A) ≥ . . . ≥ n (A) ≥ 0 denote the singular values of a matrix A ∈ Mn (C).
If K = (Kn ) is a compact sequence, then every (fractal or not) restriction of K is
compact. If K fails to be compact, then the negation of property (a) in Theorem 4.5
in [11] yields a positive constant C and strictly increasing sequences (nr ) and (kr )
with nr ≥ kr such that

kr (Knr ) ≥ C for all r.

The same theorem then implies that no restriction of the subsequence (Knr ) of (Kn )
is compact. This proves assertion (a).
Let now K be a sequence of finite essential rank r. The estimate ≥ in (2.2) comes
from (2.1). For the reverse estimate, we have to show that there is a fractal restriction
K|M of K of essential rank r. Let C := lim sup r (Kn ). By Corollary 4.6 in [11],
C is positive, and the set

M := {n ∈ N : r (Kn ) ≥ C/2}

is infinite. Since A is quasifractal, there is an infinite subset M of M such that K|M

is fractal. Moreover, by construction,

lim sup r (Km ) ≥ C/2 > 0 and lim r+1 (Km ) = 0.

M4m→∞ M4m→∞

Thus, K|M is a sequence of finite essential rank r.

It would be desirable to replace the essential ranks of the restrictions K|M on the
right-hand side of (2.2) by a more intrinsic quantity such as the algebraic rank of
the coset K|M + G|M . Since, in general, the equality ess rank K = alg rank (K + G)
does not even hold for a sequence K in a fractal algebra, we will need additional
conditions to guarantee the equality

ess rank K|M = alg rank (K|M + G|M ) (2.3)

for all M ∈ fr A.
534 S. Roch

There are several conditions which ensure the equality (2.3). Our goal is to show
(2.3) for sequences K which are stably regularizable. Recall that an element a of a
C ∗ -algebra A is Moore-Penrose invertible if there is a b ∈ A such that aba = a
and bab = b and such that ab and ba are self-adjoint. The element b is uniquely
determined by these conditions; it is called the Moore-Penrose inverse of a and
denoted by a † . All we need on Moore-Penrose invertibility is in Section 2.2.1 in
[5]. A sequence K ∈ F is then called stably regularizable if it is the sum of a
Moore-Penrose invertible sequence and a sequence in G. Stable regularizability of
a sequence K is equivalent to the Moore-Penrose invertibility of its coset K + G in
F /G.
So let K = (Kn ) ∈ A be a sequence of finite essential rank r which is stably
regularizable. Then, by definition, there are sequences (Ln ) ∈ F and (Gn ) ∈ G
such that (Kn ) = (Ln ) + (Gn ) and (Ln ) is Moore-Penrose invertible in F (and, if
G ⊆ A, even in A by inverse closedness). In particular, every Ln is Moore-Penrose
invertible and sup L†n  < ∞, such that (Ln )† = (L†n ). Clearly, (Ln ) is a compact
sequence and

ess rank (Kn ) = ess rank (Ln ).

To continue we need two simple lemmas.

Lemma 2.2 Let A be a C ∗ -algebra, J a semi-ideal in A and r the associated rank
function. If k ∈ A is Moore-Penrose invertible in F and of finite essential rank, then

r(k) = r(k ∗ ) = r(k † ) = r(k † k) = r(k ∗ k).

Proof From k = kk † k and the properties of the rank function we conclude that

r(k) ≤ r(k † k) ≤ min{r(k), r(k † )},

which implies that r(k) = r(k † k) and r(k) ≤ r(k † ). Since (k † )† = k, we obtain
r(k) = r(k † ). Further we infer from Theorem 2.15 in [5] that k † k = (k ∗ k+q)−1 k ∗ k,
where q is the Moore-Penrose projection of k. Hence,

r(k) = r(k † k) ≤ r(k ∗ k) ≤ r(k)

which finally implies that r(k) = r(k ∗ k).

Lemma 2.3 Let A be a C ∗ -algebra, p ∈ A a projection, and k ∈ pAp. Then

alg rankpAp k = alg rankA k.

Proof Let k ∈ pAp be of algebraic rank 1 in pAp and a ∈ A. Then k = pkp, and
from kak = pkp pap pkp we conclude that k has algebraic rank 1 in A.
Compact Sequences in Quasifractal Algebras 535

Set r := alg rankpAp k and s := alg rankA k. Thus, k is a sum of r elements of

algebraic rank 1 in pAp. As we have just seen, these r elements are of algebraic
rank 1 in A. Hence, s ≤ r.
Conversely, let k be the sum k1 + . . . + ks of elements ki of algebraic rank 1 in
A. Then, since k ∈ pAp, the element k = pkp is the sum pk1 p + . . . + pkm p of
m elements on algebraic rank at most 1 in pAp. Hence, r ≤ s.
From Lemma 2.2 we infer that P = (n ) := (Ln )† (Ln ) is a projection with

ess rank (n ) = ess rank (Ln ) = ess rank (Kn ) = r.

Proposition 2.4 Let A be a quasifractal C ∗ -subalgebra of F which contains G

and satisfies the local weight 1 condition K(A) = A ∩ K. Further, let P ∈ A be a
projection of finite essential rank and M ∈ fr A. Then

K (PAP)|M = (PAP)|M ∩ K|M . (2.4)

Proof Let ess rank P = r. Since P ∈ K ∩A = K(A), there are sequences (Kni ) ∈ A
of spatial rank one and (Gn ) ∈ G such that

P = (Kn1 ) + . . . + (Knr ) + (Gn ). (2.5)

We may even assume that (Kni ) ∈ PAP (otherwise multiply (2.5) from both sides
by P to get a decomposition of P into r sequences P(Kni )P ∈ PAP of essential rank
one and a sequence in PGP). Let M ∈ fr A and A ∈ PAP. Then
  
A|M = (PA)|M = (Kn1 ) + . . . + (Knr ) + (Gn ) A |M
  
= (Kn1 )A |M + . . . + (Knr )A |M + (Gn )A |M

with sequences (Kn1 )A |M of spatial rank one in (PAP)|M , which implies that

(PAP)|M ⊆ K (PAP)|M .

Since P is compact, we also have (PAP)|M ⊆ K|M , which results in the inclusion
⊇ in (2.4). The reverse inclusion is evident.
Thus, if the algebra A satisfies the local weight one condition, then so does the
algebra (PAP)|M . Since M ∈ fr A, the latter algebra is also fractal, and we conclude
from (1.1) that

ess rank K|M = alg rank(PAP)|M /G |M K|M + G|M (2.6)

for all sequences K ∈ PAP. To get rid of the restriction to the algebra PAP on the
right-hand side of (2.6), we employ Lemma 2.3 which states in the present context
536 S. Roch

that
 
alg rank(PAP)|M /G |M K|M + G|M = alg rankA|M /G |M K|M + G|M .

The following theorem summarizes what we obtained.

Theorem 2.5 Let A be a quasifractal C ∗ -subalgebra of F which contains G and
satisfies K(A) = A ∩ K. Further, let P ∈ A be a projection of finite essential rank.
Then, for all K ∈ PAP,

ess rank K = max ess rank K|M

M∈fr A

and, for all M ∈ fr A,


ess rank K|M = alg rankA|M /G |M K|M + G|M .

Remark 2.6 Let K be a sequence of finite essential rank in a quasifractal algebra

A. Does the essential rank of a restriction K|M with M ∈ fr A depend on the coset
M∼ only? In general the answer is NO, even when A is a fractal algebra (examples
can be easily found among sequences of projections). The crucial ingredient for an
affirmative answer is the local weight 1 condition.
Let A and P be as in Theorem 2.5, and let K = (Kn ) ∈ PAP be a sequence of
finite essential rank. Further, let M1 , M2 ∈ fr A and M1 ∼ M2 . Then the algebra
(PAP)|M1 ∪M2 is fractal and has local weight 1 by equality (2.4) in Proposition 2.4.
Now it follows as in Theorem 4.4 in [10] and its corollaries that the matrices Kn
have the same spatial rank for all sufficiently large n ∈ M1 ∪ M2 . (Just to mention
the point: The cited theorem deals with Fredholm sequences, i.e. with sequences
that are invertible modulo K. As in operator theory, where there is a close relation
between Fredholm operators and their kernel dimension and compact operators and
their rank, there is a close relation between the αn -numbers dealt with in Theorem
4.4 in [10] and the ranks of Kn .) Hence, ess rank K|M1 = ess rank K|M2 .

3 Algebras with Continuous Trace

There are several (equivalent) definitions of continuous trace algebras in the

literature; see, e.g. [7] and [8, Definition 5.13]. (The latter reference is an excellent
introduction into the field.) Here is Fell’s original definition from [4, IV.4.1].
Definition 3.1 A C ∗ -algebra A is called an algebra with continuous trace if
(a) A is liminal,
(b) Prim A is a Hausdorff space (with respect to the hull-kernel topology),
(c) for every L0 ∈ Prim A, there are a neighborhood U of L0 in Prim A and an
a ∈ A such that a + L is a projection of rank 1 for all L ∈ U .
Compact Sequences in Quasifractal Algebras 537

The rank condition in (c) has to be understood as follows: Since A is liminal,

there is, for every L ∈ Prim A, a unique (up to unitary equivalence) irreducible
representation (H, π) of A with ker π = L. Then the requirement is that π(a) is
a projection on H with (spatial) rank 1. Condition (c) is also referred to as Fell’s
condition.
Example
(a) The following examples taken from [8] are instructive. Of the ‘dimension-drop-
algebras’

C1 := {f ∈ C([0, 1], M2 (C)) : f (0) = diag (α, α) with α ∈ C},

C2 := {f ∈ C([0, 1], M2 (C)) : f (0) = diag (α, β) with α, β ∈ C},
C3 := {f ∈ C([0, 1], M2 (C)) : f (0) = diag (α, 0) with α ∈ C},

C3 is the only one with continuous trace. For C1 , the identity matrix diag (1, 1)
is a projection of algebraic rank one at 0, but there is no rank one projection
close to it in a neighborhood of 0, whereas for C2 , the Hausdorff property is
violated.
(b) Let (sn ) be a dense subsequence of [0, 1]. For i ∈ {1, 2, 3}, let S(Ci ) stand
for the smallest closed subalgebra of F which contains the ideal G and all
sequences (f (sn )) with f ∈ Ci where Ci is as in (a). (Here we 
assume

that
δ ≥ 2 and identify a 2 × 2-matrix A with the δ(n) × δ(n)-matrix A0 00 .) Each of
the algebras S(Ci ) is quasifractal, S(Ci ) ⊆ K, and the mapping

S(Ci )/G → Ci , (f (sn )) + G → f

is a ∗ -isomorphism from S(Ci )/G onto Ci . In particular, Prim (S(Ci )/G) ∼ =

Prim Ci . Thus, also assumptions such as compactness and quasifractality cannot
change the picture sketched in part (a): among the algebras S(Ci )/G, only
S(C2 )/G i s an algebra with continuous trace.
Let us now go back to the question from the introduction whether (A ∩ K)/G
is an algebra with continuous trace when A is quasifractal. One can show that the
ideal K is a liminal algebra. Since C ∗ -subalgebras and quotients of liminal C ∗ -
algebras are liminal (see [3], 4.2.4, 4.3.4 and 4.3.5), condition (a) in Definition
3.1 is satisfied for every algebra (A ∩ K)/G. On the other side, the Examples 3
show that we cannot expect the Hausdorff property of Prim (A ∩ K)/G in condition
(b) to hold, even for very simple examples of quasifractal algebras. So our focus
will be on Fell’s condition (c). The following three sections will prepare a natural
additional condition for the ideal of the compact sequences in A which will allow
us to overcome the problems with Fell’s condition (b).
538 S. Roch

4 A Topology on (fr A)∼

Here we recall from [13] the definition of a topology on (fr A)∼ which makes
(fr A)∼ to a compact Hausdorff space. For A as C ∗ -subalgebra of F , let Lmin (A)
denote the smallest closed complex subalgebra1 of l ∞ := l ∞ (N) which contains all
sequences (An ) where (An ) ∈ A. Clearly, Lmin (A) is a commutative C ∗ -algebra,
which is unital if A is unital.
For a C ∗ -subalgebra L of l ∞ , we let cr L stand for the set of all infinite subsets
M of N such that all sequences in the restriction L|M converge. The algebra L is
called quasiconvergent if every infinite subset of N has an infinite subset in cr L.
Then, for every C ∗ -subalgebra A of F , fr A = cr Lmin (A), and A is quasifractal if
and only if Lmin (A) is quasiconvergent.
Let L be a unital C ∗ -subalgebra of l ∞ . Then, for every M ∈ cr L, the mapping

ϕM : L → C, a → lim (a|M ) (4.1)

is a character (i.e., a non-zero continuous linear functional) on L. Since L ∩ c0 is in

the kernel of the mapping (4.1), the quotient mapping

ϕM : L/(L ∩ c0 ) → C, a + (L ∩ c0 ) → lim (a|M)

is well defined, and this mapping is a character of L/(L ∩ c0 ).

Proposition 4.1 (Proposition 4.5 in [13]) Let L be a unital and quasiconvergent
C ∗ -subalgebra of l ∞ . Then the set {ϕM : M ∈ cr L} is strictly spectral for L/(L ∩
c0 ), i.e., if b ∈ L/(L ∩ c0 ) and ϕM (b) is invertible for all M ∈ cr L, then b is
invertible.
In order to conclude that {ϕM : M ∈ cr L} is all of the maximal ideal space
Max L/(L ∩ c0 ) of L/(L ∩ c0 ) we need a further property of L: separability,2 and
in order to make the mapping M → ϕM injective, we introduce an equivalence
relation ∼ on cr L by calling M1 , M2 of cr L equivalent if M1 ∪ M2 ∈ cr L. We
denote the equivalence class of M ∈ cr L by M∼ and write (cr L)∼ for the set of all
equivalence classes.
Theorem 4.2 (Corollary 4.7 in [13]) Let L be a unital, separable and quasicon-
vergent C ∗ -subalgebra of l ∞ . Then the mapping M∼ → ϕM is a bijection from
(cr L)∼ onto Max (L/(L ∩ c0 )).
If now A is a unital and quasifractal C ∗ -subalgebra of F , then Lmin (A) is a unital
and quasiconvergent C ∗ -subalgebra of l ∞ .

1 The algebra Lmin (A) is denoted by L(A) in [13].

2 Actually, one only needs that Max L/(L ∩ c0 ) is first countable; see [6].
Compact Sequences in Quasifractal Algebras 539

Corollary 4.3 Let A be a unital and quasifractal C ∗ -subalgebra of F such that

Lmin (A) is separable. Then the mapping M∼ → ϕM is a bijection from (fr A)∼
onto Max (Lmin (A)/(Lmin (A) ∩ c0 )).
Under the conditions of the corollary, one thus can transfer the Gelfand topology
from Max (Lmin (A)/(Lmin (A) ∩ c0 )) onto (fr A)∼ making the latter a compact
Hausdorff space.
Besides Lmin (A), there is another way to associate a commutative C ∗ -algebra
with a quasifractal algebra A which induces a Hausdorff topology on (fr A)∼ .
Let Lmax (A) stand for the set of all sequences (αn ) ∈ l ∞ with the property
that the restricted sequence (αn )|M converges for every M ∈ fr A. Clearly,
Lmin (A) ⊆ Lmax (A) (hence the notation). Moreover, the algebra Lmax (A) is unital
and quasiconvergent whenever A is quasifractal, and fr A = cr Lmax (A). So one can
also generate a topology on (fr A)∼ using Lmax (A). It turns out that these topologies
coincide in important cases. For example, using ideas from the proof of Theorem 4.2
= Corollary 4.7 in [13], one can show that Lmin (A) = Lmax (A) whenever A is a
unital, separable and quasifractal C ∗ -subalgebra of F .
On the other side, there are contexts in which the use of Lmax (A) is of advantage.
For example, it is often useful to have Lmax (A) as a subalgebra inside A (in the
sense that a sequence (αn ) ∈ l ∞ is identified with the sequence (αn Pn ) ∈ F ). We
call A an Lmax -algebra in this case. The requirement to be an Lmax -algebra can be
easily fulfilled: For a C ∗ -subalgebra A of F , let Amax stand for the smallest closed
C ∗ -subalgebra of F which contains A and Lmax (A). It is elementary to check that

Lmax (Amax ) = Lmax (A),

thus Amax is an Lmax -algebra containing A.

Theorem 4.4 Let A be a unital C ∗ -subalgebra of F with G ∈ A, and assume that
A is an Lmax -algebra. Then

Lmin (A) = Lmax (A) = A ∩ l ∞ . (4.2)

Proof We first show that Lmin (L) = L for every C ∗ -subalgebra L of l ∞ . If (αn ) ∈
L, then (|αn |) = ((αn )(αn )∗ )1/2 ∈ L; hence, Lmin (L) ⊆ L. Conversely, every
sequence in L can be written as a linear combination of four non-negative sequences.
Since (βn ) = (|βn |) ∈ Lmin (L) for each of these sequences, the reverse inclusion
L ⊆ Lmin (L) follows.
Now to the proof of (4.2). The inclusion Lmin (A) ⊆ Lmax (A) holds trivially, as
already mentioned. To get the reverse inclusion, note that

Lmax (A) = Lmin (Lmax (A)) ⊆ Lmin (A),

where the equality comes from the remark at the beginning of the proof and the
inclusion holds because Lmax (A) ⊆ A and the functor Lmin is increasing. This
settles the first equality in (4.2).
540 S. Roch

For the second equality, set L := A ∩ l ∞ , considered as a subalgebra of l ∞ .

Evidently, Lmax (A) ⊆ L. For the reverse inclusion, note that L ⊆ Lmax (L), again
by the above remark. Since L ⊆ A and the functor Lmax is increasing, the assertion
follows.
The following section shows another instance where Lmax (A) naturally occurs.

5 Sequences with Sequential Rank 1

Let H be a Hilbert space and K ∈ L(H ) an operator with rank K ≤ 1. Then, for
every A ∈ L(H ), there is an α ∈ C such that KAK = αK. The number α is
uniquely determined and satisfies |α| ≤ A K whenever K = 0. If K = 0, then
KAK = αK holds for every α ∈ C, whereas the condition |α| ≤ A K only
holds when α = 0. Thus, both conditions together determine α uniquely.
Let now A be a C ∗ -subalgebra of F and let (Kn ) ∈ A be a sequence with
maxn rank Kn ≤ 1. Then, as we have just seen, for every sequence (An ) ∈ A, there
is a unique sequence (αn ) ∈ l ∞ such that

Kn An Kn = αn Kn and |αn | ≤ An  Kn  for every n ∈ N. (5.1)

Theorem 5.1 Let A be a unital C ∗ -subalgebra of F with G ⊆ A and let (An ) ∈ A.

Then the sequence (αn ) determined by (5.1) belongs to Lmax (A).
Proof It is helpful to define the support of a sequence (An ) ∈ A by

supp (An ) := {M ∈ fr A : (An )|M ∈ G|M }.

If M ∈ supp (An ), then the restriction A|M is fractal and (An )|M is not a partial zero
sequence. Thus,

M ∈ supp (An ) if and only if lim An  > 0. (5.2)

n∈M

Note also that if one of two equivalent sets in fr A belongs to supp (An ), then so does
the other. In this sense, the support of a sequence is compatible with the equivalence
relation ∼.
Let now (An ), (Kn ) ∈ A with maxn rank Kn ≤ 1, and let (αn ) be the sequence
uniquely determined by (5.1). Further let M ∈ fr A. We distinguish two cases.
Case A M ∈ supp (An ). Then (An )|M ∈ G|M by (5.2). The second condition in
(5.1) then implies that (αn )|M is a zero sequence; hence, convergent.
Case B M ∈ supp (An ). Suppose (αn )|M is not convergent. Then, since (αn ) is a
bounded sequence, there are complex numbers α = β and disjoint infinite subsets
Mα , Mβ of M such that αn → α on Mα and αn → β on Mβ .
Compact Sequences in Quasifractal Algebras 541

Consider the sequence (αn Kn − αKn ) which is equal to (Kn An Kn ) − α(Kn )

and, hence, belongs to A. Its restriction onto Mα is a zero sequence, whereas its
restriction to Mβ is not. The latter follows from the estimate

αn Kn − αKn  ≥ |α − β| Kn  − |β − αn | Kn  (5.3)

and the fact that the sequence (Kn )|M has a positive limit, k, by (5.2), which
implies that the right hand side of (5.3) converges to k|α − β| > 0.
Hence, (αn Kn −αKn )|M is a partial zero sequence but not a zero sequence, which
contradicts the fractality of A|M . This contradiction shows that the sequence (αn )|M
converges for every M ∈ fr A, whence (αn ) ∈ Lmax (A).
Let A be a unital C ∗ -algebra and C a unital C ∗ -subalgebra in the center of
A, i.e., the elements of C commute which each element of A. Then C is a unital
commutative C ∗ -subalgebra of A. Let J stand for the set of all elements k ∈ A
with the following property: For every a ∈ A there is a c ∈ C such that kak = ck.
Then J forms a semi-ideal of A. The associated rank function is called the C-rank,
and the associated ideal of the C-compact elements is denoted by JC (A). In case C
is the complete center of A, the C-rank is also called the central rank and denoted
by cen rank a, and the associated ideal of the centrally compact elements is denoted
by Jcen (A). Note also that JC (A) = C(A), with the notation introduced in the
introduction.
Theorem 5.2 Let A be a unital C ∗ -subalgebra of F with G ⊆ A, and assume A is
an Lmax -algebra. Then

K(A) = Jcen (A).

The inclusion ⊇ holds without the Lmax -assumption.

Proof Let (Kn ) be a sequence of central rank 1 in A. Since G ⊆ A by assumption,
the center of A is a subalgebra of l ∞ , and rank Kn ≤ 1 for every n ∈ N. Hence,
(Kn ) ∈ K(A) and Jcen (A) ⊆ K(A).
Conversely, let (Kn ) ∈ A be a sequence with maxn rank Kn ≤ 1. By (5.1), for
every (An ) ∈ A, there is a unique sequence (αn ) ∈ l ∞ such that Kn An Kn = αn Kn
for every n. The sequence (αn ) belongs to Lmax (A) by Theorem 5.1. Since A is an
Lmax -algebra, the sequence (αn In ) lies in A and, hence in the center of A. Thus,
(Kn ) is of central rank 1, whence K(A) ⊆ Jcen (A).

6 Fell’s Condition: Transition to C -Compact Elements

The theory of compact sequences in fractal algebras A is most satisfying when the
algebra A is of local weight 1, which happens if and only if A ∩ K = K(A). It is
only natural to impose an analogous condition when studying compact sequences in
542 S. Roch

quasifractal algebras. Thus, in what follows, we let A be a unital quasifractal C ∗ -

subalgebra of F with G ⊆ A which is an Lmax -algebra, and we suppose further that
Lmax (A) is separable and that

A ∩ K = K(A).

Then, by Theorem 5.2, A ∩ K = K(A) = Jcen (A) where Cen A = Lmax (A) =
A ∩ l ∞ (N) by Theorem 4.4.
We are going to localize the algebra B := A/G over its central subalgebra C :=
(Lmax (A) + G)/G ∼ = Lmax (A)/c0 via the local principle by Allan/Douglas (see
Theorems 2.2.2 and 2.2.11 in [14]). By Theorem 4.2,

Max C = Max (Lmax (A)/c0 ) ∼

= (fr A)∼ .

Let Ix stand for the smallest closed ideal of B which contains the maximal ideal x
of C and let x denote the quotient map B → B/Ix . Then the local principle states
(among other things) that an element b ∈ B is invertible in B if and only if x (b) is
invertible in B/Ix for every x ∈ Max B and that the function

Max C → R+ , x → x (b) (6.1)

is upper semi-continuous for every b ∈ B. In the present context on can say more:
If x = M∼ ∈ (fr A)∼ then, by Theorem 4.9 in [13] (note that Lmax (A) = Lmin (A)
by Theorem 4.4), the local algebra A/Ix is ∗ -isomorphic to the fiber A(M∼ ), and
the function (6.1) is continuous for every b ∈ B. (Roughly speaking, the continuity
results from including Lmax (A) = Lmin (A) into A.)
If (Kn ) ∈ A is a sequence of central rank one, then its coset k := (An ) + G
has the property that, for every b ∈ B, there is a c ∈ C such that kbk = ck.
Thus, k ∈ B has C-rank 1. The ideal Jcen (A)/G of B we are interested in will be
denoted by J . As we have just seen, J is generated by certain elements of C-rank
1. Hence, J ⊆ JC (B). Note also that C is a subalgebra of the center of B, hence
JC (B) ⊆ Jcen (B). We cannot claim that equality holds in either of these inclusions.

7 Fell’s Condition for C -Compact Elements

In this section, we are going to show that Fell’s condition is satisfied for the algebras
J and JC (B). We prepare the next steps with a few lemmas. Since now C stands for
a central subalgebra, we adopt the notation introduced in the previous section and
write JC (A) instead of C(A) for the closed ideal of a C ∗ -algebra A generated by
its elements of C-rank = algebraic rank 1 to avoid confusion.
Compact Sequences in Quasifractal Algebras 543

Lemma 7.1 Let B be a unital C ∗ -algebra, C a central C ∗ -subalgebra of B which

contains the unit element, x ∈ Max C, and J ⊆ JC (B) a closed ideal of B which is
generated by elements of C-rank one. Then
(a) if k ∈ B is of C-rank one, then x (k) is of algebraic rank ≤ 1 in B/Ix ,
(b) x (J ) ⊆ JC (B/Ix ),
(c) x (J ) is a dual algebra.
Proof (a) For every b ∈ B, there is a c ∈ C such that kbk = ck. Thus,
x (k)x (b)x (k) = x (c)x (k). Since x (c) is the value of the Gelfand
transform of c at x, hence a complex number, the assertion follows. Since J is
generated as an ideal by sequences of C-rank 1, (b) is a consequence of (a).
Finally, we know from [10] that JC (A) is a dual algebra for every C ∗ -algebra
A, and it is well known (see, e.g., [1]) that dual algebras are ∗ -isomorphic to C ∗ -
subalgebras of K(H ) for some Hilbert space H and that, hence, C ∗ -subalgebras of
dual algebras are dual again. Therefore x (J ), which is a C ∗ -subalgebra of a dual
algebra by (b), is dual.
Lemma 7.2
(a) Let B be a C ∗ -algebra with identity e, J a proper closed ideal of B, j ∈ J self-
adjoint, and f a continuous function on σ (j ) with f (0) = 0. Then f (j ) ∈ J .
(b) In addition, let C be a central C ∗ -subalgebra of B which contains e, let j have
C-rank one, and assume that f is continuously differentiable on some interval
containing σ (j ) and that f (0) = f (0) = 0. Then f (j ) has C-rank ≤ 1.
Proof We will prove part (b) only. Let [a, b] be a closed interval which contains
σ (j ). Since J is proper, 0 ∈ σ (j ) ⊆ [a, b]. By the Weierstrass’ approximation
theorem, there is a sequence of polynomials pn such that pn → f and pn → f
uniformly on [a, b] and pn (0) = pn (0) = 0. Define polynomials qn such that
pn (x) = xqn (x) for all x.
Since f is differentiable and f (0) = 0, the function g defined by g(x) :=
f (x)/x if x = 0 and g(0) := 0 is continuous on [a, b] and differentiable on (a, 0)
and (0, b). Let x ∈ (0, b]. Then, by the mean value theorem, there is a ξ ∈ (0, x)
such that
pn (x) − f (x)
qn (x) − g(x) = = pn (ξ ) − f (ξ ),
x

hence |qn (x) − g(x)| ≤ pn − f ∞ . This estimate holds for x ∈ [a, 0) and x = 0
as well. Thus,

qn − g∞ ≤ pn − f ∞ ,

implying that the qn converge uniformly to g. By the functional calculus again,

pn (j ) → f (j ) and pn (j ) = j qn (j ) → jg(j ) in the norm of B. Hence, f (j ) =
jg(j ), and f (j ) has C-rank ≤ 1 because j has C-rank 1.
544 S. Roch

The next results state that projections of algebraic rank one in x (J ) can be
lifted both to elements of C-rank one in J and to local projections.
Proposition 7.3 Let B be a unital C ∗ -algebra, C a central C ∗ -subalgebra of B
which contains the unit element, and J ⊆ JC (B) a closed ideal of B which is
generated by elements of C-rank one. Let a ∈ J , x ∈ Max C, and suppose that
x (a) is a projection of algebraic rank one in B/Ix . Then there is an element k ∈ J
of C-rank one such that x (k) = x (a).
Proof Set p := x (a). Being of algebraic rank one, p = 0. Since a ∈ J , a is a
limit of sums of elements rin ∈ J with C-rank one:


n
a = lim rin .
n→∞
i=1

Then


n
p = x (a) = lim x (rin )
n→∞
i=1

and, because p is a projection,


n
p = lim px (rin )p. (7.1)
n→∞
i=1

Since p is of algebraic rank one, there are numbers λin ∈ C such that px (rin )p =
λin p. Not all of these numbers can be zero (otherwise (7.1) would imply that p = 0).
Let λ := λi0 n0 = 0. Then, with r := ri0 n0 , px (r)p = λp, whence

1 1
p= px (r)p = x (a)x (r)x (a) = x (λ−1 ara).
λ λ

Then k := λ−1 ara belongs to J and has the desired properties.

Proposition 7.4 Let B be a unital C ∗ -algebra, C a central C ∗ -subalgebra of B
which contains the unit element, and J a closed ideal of B. For x ∈ Max C, let
p ∈ x (J ) be a projection. Then there are an open neighborhood U of x and a
positive element a ∈ J such that x (a) = p and the y (a) is a projection for
every y ∈ U . Moreover, if the function

Max C 4 y → y (j ) (7.2)

is continuous for every j ∈ J , then U can be chosen such that y (a) = 0 for all
y ∈ U.
Compact Sequences in Quasifractal Algebras 545

Proof Let b ∈ J be such that x (b) = p. Without loss we may assume that b is
positive (otherwise first replace b by (b + b∗ )/2 and then by max(0, b), using the
continuous functional calculus and Lemma 7.2 (a)). From

0 = p − p2  = x (b) − x (b)2 

and the upper semi-continuity of the function (7.2) we conclude that, given ε > 0,
there is an open neighborhood Uε of x such that

y (b) − y (b)2  ≤ ε − ε2 for all y ∈ Uε .

Since |s − s 2 | ≤ ε − ε2 for s ∈ R implies s ∈ [−ε, ε] ∪ [1 − ε, 1 + ε], we conclude

that

σ (y (b)) ⊆ [−ε, ε] ∪ [1 − ε, 1 + ε] for all y ∈ Uε .

Let f be a bounded continuous function on R which is 0 on [−ε, ε] and 1 on

[1 − ε, 1 + ε], and set a := f (b). Then a ∈ J by Lemma 7.2 (a), and f (y (b)) =
y (f (b)) = y (a). Thus, y (a) is a projection for all y ∈ Uε , and x (a) =
f (x (b)) = f (p) = p (note that f is the identical mapping on σ (p) ⊆ {0, 1}).
Up to here we only used the upper semi-continuity of the functions (7.2). If these
functions are continuous, then x (a) = p = 1 implies y (a) ≥ 1/2 for all
y in an open neighborhood U ⊆ Uε of x, thus giving the final assertion.
Theorem 7.5 Let B be a unital C ∗ -algebra, C a central C ∗ -subalgebra of B which
contains the unit element, and J ⊆ JC (B) a closed ideal of B which is generated
by elements of C-rank 1. We assume that the function (7.2) is continuous for every
j ∈ J . Finally, for x ∈ Max C, let p ∈ x (J ) be a projection of algebraic rank
1. Then there are an open neighborhood U of x and a positive element a ∈ J with
C-rank 1 such that x (a) = p and that y (a) is a projection of algebraic rank 1
for all y ∈ U .
Proof Let p ∈ x (J ) be a projection of algebraic rank 1. By Proposition 7.3, there
is an element k ∈ J of C-rank 1 such that x (k) = p. Further, as in the proof
of Proposition 7.4, we find a positive element j ∈ J such that x (j ) = p. Then
b := k ∗ j k is a positive element of J with x (b) = p, and b has C-rank 1 (note that
b = k ∗ j k = 0 because p = 0).
We proceed as in the proof of Proposition 7.4, but now with the function f
defined there being chosen continuously differentiable. Then a := f (b) belongs
to J by Lemma 7.2 (a) and has C-rank ≤ 1 by Lemma 7.2 (b). It follows as in
the proof of Proposition 7.4 that x (a) = p and that y (a) is a projection for all
y in a certain neighborhood U of x, and this projection has algebraic rank ≤ 1 by
Lemma 7.1 (a).
Finally, since p = 0, a has C-rank 1, and the continuity of the functions (7.2)
implies that y (a) is a projection of algebraic rank 1 for all y in a certain open
neighborhood of x (possibly smaller than U ).
546 S. Roch

One may consider Theorem 7.5 as stating Fell’s condition for J with respect to
the base space Max C. To get from here Fell’s condition for J with respect to the
primitive ideal space Prim J , we still need a ‘transfer lemma’ which relates the two
settings. For, we need some facts on primitive ideal spaces. First, the mapping

ϕ : Prim A → Max C, L → L ∩ C

is onto, and it is continuous with respect to the hull-kernel topology on Prim A (and
it is open if and only if the function Max C 4 y → y (j ) is continuous for every
a ∈ A) (Section C1 in [16]) and, second, the mapping

λJ : PrimJ A = {L ∈ Prim A : J ⊆ L} → Prim J , L → L ∩ J

is a homeomorphism (Proposition A27 in [8]). Define

ϕJ : Prim J → Max C, ϕJ := ϕ ◦ λ−1

J .

Let L0 ∈ Prim J and x0 := ϕJ (L0 ). Then x0 = ϕ(χJ−1 (L0 )) = χJ−1 (L0 ) ∩ C;

hence, x0 ⊆ χJ−1 (L0 ). Since χJ−1 (L0 ) is a closed ideal of A, this implies Ix0 ⊆
χJ−1 (L0 ), from which we conclude that

Ix0 ∩ J ⊆ λ−1 −1
J (L0 ) ∩ J = λJ (λJ (L0 )) = L0 .

Summarizing, we obtain

Ix0 ∩ J ⊆ L0 for x0 = ϕJ (L0 ). (7.3)

Lemma 7.6 Let B, C and J ⊆ JC (B) be as in Theorem 7.5.

(a) Let L0 ∈ Prim J and p a projection of algebraic rank 1 in J /L0 . Set x0 :=
ϕJ (L0 ). Then there is a a ∈ J of C-rank 1 such that a + L0 = p and a + Ix0
is a projection of algebraic rank 1 in x0 (J ).
(b) Let x0 ∈ Max C and a ∈ J of C-rank 1 such that a + Ix0 is a projection of
algebraic rank 1 in x0 (J ). Then there is an L0 ∈ Prim J with ϕJ (L0 ) = x0
such that a + L0 is a projection of algebraic rank 1 in J /L0 . This L0 is unique.
Proof
(a) Given a projection p of algebraic rank 1 in J /L0 , choose a positive b ∈ J
such that b + L0 = p. Since J is generated by sequences of C-rank 1, there are
elements jkn ∈ J of C-rank 1 and numbers kn ∈ N such that


kn
b = lim jkn .
n→∞
k=1
Compact Sequences in Quasifractal Algebras 547

Passing to cosets modulo L0 and multiplying by the projection b + L0 = p

yields


kn
b + L0 = lim (b + L0 ) (jkn + L0 ) (b + L0 ). (7.4)
n→∞
k=1

Since b + L0 = p is of algebraic rank 1, there are αkn ∈ C such that

(b + L0 ) (jkn + L0 ) (b + L0 ) = αkn (b + L0 ).

By (7.4) and since p = 0, at least one of the αkn is not zero; say αk n . Then
j := jk n /αk n is in J , is of C-rank 1, and satisfies

(b + L0 ) (j + L0 ) (b + L0 ) = b + L0 .

We may moreover assume that j ≥ 0 (otherwise replace j by j ∗ bj ). Then

a := bj b ∈ J is positive, of C-rank 1, and a + L0 = b + L0 = p.
Since a has C-rank 1, the coset a +Ix0 has algebraic rank ≤ 1. Actually, it has
algebraic rank 1 since a +Ix0 = 0 would imply p = a +L0 = 0 by (7.3), which
is impossible. Since b = 0 is positive and has C-rank 1, x0 (b/x0 (b)) is a
projection of algebraic rank 1. Thus,
2
b/x0 (b) − b/x0 (b) ∈ Ix0

whence, by (7.3),
2
b/x0 (b) − b/x0 (b) ∈ L0 .

Thus, b/x0 (b)+L0 is a projection. Since already b+L0 = p is a projection,

we conclude that x0 (b) = 1, i.e., b + L0 is a projection.
(b) Since a 2 − a, a ∗ − a ∈ Ix0 , we conclude from (7.3) that a 2 − a, a ∗ − a ∈ L for
each L ∈ Prim J with ϕJ (L) = x0 . Thus a + L is a projection for each such L.
We show next that a + L has algebraic rank ≤ 1. Given k ∈ A, there is a c ∈ C
such that aka = ca. Then aka − c(x0)k ∈ Ix0 , whence aka − c(x0)k ∈ L, again
by (7.3).
Further, x0 (J ) is a dual ideal which therefore is generated by projections
of algebraic rank 1, i.e., by the a + L. So there is at least one L0 ∈ Prim J
with ϕJ (L0 ) = x0 such that a + L0 is a projection of algebraic rank 1. Finally,
this L0 is unique: Each projection of algebraic rank 1 in x0 (J ) sits in a unique
elementary component of x0 (J ), and these components are in a one-to-one
correspondence with the primitive ideals L of J with ϕJ (L) = x0 .
548 S. Roch

The following theorem settles Fell’s condition (= condition (c) in Definition 3.1)
for ideals J which are generated by elements of C-rank 1.
Theorem 7.7 Let B be a unital C ∗ -algebra, C a central C ∗ -subalgebra of B which
contains the unit element, and J ⊆ JC (B) a closed ideal of B which is generated
by elements of C-rank 1. Assume that the function (7.2) is continuous for every
j ∈ J . Then J satisfies Fell’s condition, i.e, for every L0 ∈ Prim J , there are a
neighborhood U of L0 in Prim J and an a ∈ J such that a + L is a projection of
algebraic rank 1 for all L ∈ U .
Proof Let L0 ∈ Prim J and x0 := ϕJ (L0 ). Let p ∈ J /L0 be a projection of
algebraic rank 1. Then, by part (a) of the Transfer Lemma 7.6, there is a a ∈ J of
C-rank 1 such that a + L0 = p and a + Ix0 is a projection of algebraic rank one
in x0 (J ). By Theorem 7.5, we can choose the element a positive and of C-rank
1 and such that x (a) is a projection of algebraic rank one for all x in an open
neighborhood V ⊆ Max C of x0 . Part (b) of the transfer lemma then implies that,
for every x ∈ V , there is a unique Lx ∈ Prim J such that ϕJ (Lx ) = x and that
a + Lx is a projection of algebraic rank 1 in J /Lx .
−1
Since ϕJ is continuous, U1 := ϕJ (V ) is open in Prim J . Further, by Lemma
A.30 in [8], U2 := {L ∈ Prim J : a + L > 1/2} is open (note that Spec J and
Prim J are naturally homeomorphic). Thus, U := U1 ∩ U2 is open in Prim J . The
assertion will follow once we have shown that U = {Lx : x ∈ V }. The inclusion ⊇
follows by the definition of U .
The reverse inclusion ⊆ is a consequence of the transfer lemma again. Indeed, let
L ∈ U . Since L ∈ U1 , ϕJ (L) = x0 ∈ V . By Lemma 7.6, a + L is a projection of
algebraic rank ≤ 1, implying that the norm of a + L is either 0 or 1. Since L ∈ U2 ,
we conclude that a + L = 1. Hence, a + L is a projection of rank 1, whence
L = Lx0 by the uniqueness assertion in the transfer lemma.

8 Continuity of Local Ranks

If J is as in Theorem 7.7 and x ∈ Max C, then J /(Ix ∩ J ) is a dual algebra. Hence,

every projection in J /(Ix ∩ J ) has a finite algebraic rank. We are interested in the
dependence of these ‘local ranks’ on x.
Theorem 8.1 Let B be a unital C ∗ -algebra, C a central C ∗ -subalgebra of B which
contains the unit element, and J ⊆ JC (B) a closed ideal of B which is generated
by elements of C-rank one. Assume that the function (7.2) is continuous for every
j ∈ J . Further let a ∈ J be such that a + (Ix ∩ J ) is a projection for all x in an
open set U ⊆ Max C. Then the function

Max C → Z+ , x → alg rank (a + (Ix ∩ J ))

is continuous on U .
Compact Sequences in Quasifractal Algebras 549

Proof Let x0 ∈ U and s := alg rank (a + (Ix0 ∩ J )). Then there are finitely many
elementary components E1 , . . . , Er of the dual algebra J /(Ix0 ∩ J ) and s mutually
orthogonal projections πji ∈ Ei , j = 1, . . . , ni , of algebraic rank 1 such that


r 
ni
a + (Ix0 ∩ J ) = πji .
i=1 j =1

In each component Ei , we fix a projection βi of algebraic rank 1. Since Ei is

elementary, there are partial isometries αji ∈ Ei such that

(αji )∗ αji = πji and αji (αji )∗ = βi for j = 1, . . . , ni .

Now we continue the locally defined elements πji , αji and βi to elements of J .
By Theorem 7.5, there are positive elements bi ∈ J of central rank 1 such that
bi + (Ix0 ∩ J ) = βi and bi + (Ix ∩ J ) is a projection of algebraic rank 1 for all x
in a neighborhood Wi of x0 .
Next, it follows as in Lemmas 3.1 and 3.2 of [4], there are elements pji , aji , bi ∈
J and a neighborhood V ⊆ ∩i Wi of x0 such that

pji + (Ix0 ∩ J ) = πji , aji + (Ix0 ∩ J ) = αji , bi + (Ix0 ∩ J ) = βi

and that, for all x ∈ V , bi + (Ix ∩ J ) is a projection and the pji + (Ix ∩ J ) are
mutually orthogonal projections with

(aji )∗ aji +(Ix ∩J ) = pji +(Ix ∩J ), aji (aji )∗ +(Ix ∩J ) = bi +(Ix ∩J ). (8.1)

Since bi + (Ix0 ∩ J ) = bi + (Ix0 ∩ J ) = βi and the function (7.2) is upper semi-

continuous, we may assume that

(bi + (Ix ∩ J )) − (bi + (Ix ∩ J )) < 1 for all x ∈ V

(otherwise we lessen V accordingly). Since bi +(Ix ∩J ) is a projection of algebraic

rank 1, we conclude via a Neumann series argument that bi +(Ix ∩J ) is of algebraic
rank 1, too. But then (8.1) implies that pji + (Ix ∩ J ) is of algebraic rank 1 and that
  i
p := ri=1 nj =1 pji + (Ix ∩ J ) is a projection of algebraic rank s for all x in the
neighborhood V of x0 .

References

1. M.C.F. Berglund, Ideal C ∗ -algebras. Duke Math. J. 40, 241–257 (1973)

2. A. Böttcher, B. Silbermann, Introduction to Large Truncated Toeplitz Matrices (Springer,
Berlin, 1999)
3. J. Dixmier, C ∗ -Algebras (North Holland Publishing Company, Amsterdam, 1982)
4. J.M.G. Fell, The structure of algebras of operator fields. Acta Math. 106, 233–280 (1961)
550 S. Roch

5. R. Hagen, S. Roch, B. Silbermann, C ∗ -Algebras and Numerical Analysis (Dekker, New York,
2001)
6. V. Nistor, N. Prudhon, Exhaustive families of representations and spectra of pseudodifferential
operators. J. Oper. Theory 78, 247–279 (2017)
7. G.K. Pedersen, C ∗ -Algebras and Their Automorphism Groups (Academic, London, 1979)
8. I. Raeburn, D.P. Williams, Morita Equivalence and Continuous Trace C ∗ -Algebras (American
Mathematical Society, Providence, 1998)
9. S. Roch, Algebras of approximation sequences: Fractality, in in Problems and Methods in
Mathematical Physics. Operator Theory: Advances and Applications, vol. 121 (Birkhäuser,
Basel, 2001), pp. 471–497
10. S. Roch, Algebras of approximation sequences: fredholm theory in fractal algebras. Studia
Math. 150, 53–77 (2002)
11. S. Roch, Finite Sections of Band-Dominated Operators, vol. 191. (Memoirs of the American
Mathematical Societ, Providence, 2008), p. 895
12. S. Roch, Extension-restriction theorems for algebras of approximation sequences, in Operator
Theory, Operator Algebras, ans Matrix Theory. Operator Theory: Advances and Applications,
vol. 267 (Birkhäuser, Basel 2018), pp. 261–284
13. S. Roch, Beyond fractality: Piecewise fractal and quasifractal algebras, in The Diversity and
Beauty of Applied Operator Theory. Operator Theory: Advances and Applications, vol. 268
(Birkhäuser, Basel 2018), pp. 413–428
14. S. Roch, P.A. Santos, B. Silbermann, Non-commutative Gelfand Theories. A Tool-kit for
Operator Theorists and Numerical Analysts (Universitext, Springer, London, 2011)
15. S. Roch, B. Silbermann, C ∗ -algebra techniques in numerical analysis. J. Oper. Theory 35,
241–280 (1996)
16. D. Williams, Crossed Products of C ∗ -Algebras (American Mathematical Society, Providence,
2007)
Dilation Theory: A Guided Tour

Orr Moshe Shalit

Abstract Dilation theory is a paradigm for studying operators by way of exhibiting

an operator as a compression of another operator which is in some sense well
behaved. For example, every contraction can be dilated to (i.e., is a compression
of) a unitary operator, and on this simple fact a penetrating theory of non-normal
operators has been developed. In the first part of this survey, I will leisurely review
key classical results on dilation theory for a single operator or for several commuting
operators, and sample applications of dilation theory in operator theory and in
function theory. Then, in the second part, I will give a rapid account of a plethora
of variants of dilation theory and their applications. In particular, I will discuss
dilation theory of completely positive maps and semigroups, as well as the operator
algebraic approach to dilation theory. In the last part, I will present relatively new
dilation problems in the noncommutative setting which are related to the study
of matrix convex sets and operator systems, and are motivated by applications in
control theory. These problems include dilating tuples of noncommuting operators
to tuples of commuting normal operators with a specified joint spectrum. I will also
describe the recently studied problem of determining the optimal constant c = cθ,θ ,
such that every pair of unitaries U, V satisfying V U = eiθ U V can be dilated to a
pair of cU , cV , where U , V are unitaries that satisfy the commutation relation
V U = eiθ U V . The solution of this problem gives rise to a new and surprising
application of dilation theory to the continuity of the spectrum of the almost Mathieu
operator from mathematical physics.

Keywords Dilations · Isometric dilation · Unitary dilation · Matrix convex

sets · q-commuting unitaries · Completely positive maps · CP-semigroups

Mathematics Subject Classification (2010) 47A20, 46L07, 46L55, 47A13,

47B32, 47L25

O. M. Shalit ()
Faculty of Mathematics, Technion - Israel Institute of Technology, Haifa, Israel
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 551

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_28
552 O. M. Shalit

Dilation theory is a collection of results, tools, techniques, tricks, and points of

view in operator theory and operator algebras, that fall under the unifying idea that
one can learn a lot about an operator (or family of operators, or a map, etc.) by
viewing it as “a part of” another, well understood operator. This survey on dilation
theory consists of three parts. The first part is a stand-alone exposition aimed at
giving an idea of what dilation theory is about by describing several representative
results and applications that are, in my opinion, particularly interesting. The climax
of the first part is in Sect. 4, where as an application of the material in the first three
sections, we see how to prove the Pick interpolation theorem using the commutant
lifting theorem. Anyone who took a course in operator theory can read Part 1.
Out of the theory described in the first part, several different research directions
have developed. The second part of this survey is an attempt to give a quick account
of some of these directions. In particular, we will cover Stinespring’s dilation
theorem, and the operator algebraic approach to dilation theory that was invented
by Arveson. This survey up to Sect. 7 contains what everyone working in dilation
theory and/or nonselfadjoint operator algebras should know. I will also cover a
part of the dilation theory of CP-semigroups, and take the opportunity to report
on my work with Michael Skeide, which provides our current general outlook on
the subject.
In the third and last part I will survey some recent dilation results in the
noncommutative setting, in particular those that have been motivated by the study of
matrix convex sets. Then I will focus on my recent joint work with Malte Gerhold,
where we study the problem of dilating q-commuting unitaries. Experts on dilation
theory can read the last three sections in this survey independently.
I made an effort to include in this survey many applications of dilation theory.
The theory is interesting and elegant in itself, but the applications give it its vitality.
I believe that anyone, including experts in dilation theory, will be able to find in this
survey an interesting application which they have not seen before.
Some results are proved and others are not. For some results, only an idea of
the proof is given. The guiding principle is to include proofs that somehow together
convey the essence or philosphy of the field, so that the reader will be able to get
the core of the theory from this survey, and then be able to follow the references for
more.
As for giving references: this issue has given me a lot of headaches. On the one
hand, I would like to give a historically precise picture, and to give credit where
credit is due. On the other hand, making too big of a fuss about this might result in
an unreadable report, that looks more like a legal document than the inviting survey
that I want this to be. Some results have been rediscovered and refined several times
before reaching their final form. Who should I cite? My solution was to always
prefer the benefit of the reader. For “classical” results, I am very happy to point
the reader to an excellent textbook or monograph, that contains a proof, as well as
detailed references and sometimes also historical remarks. I attach a specific paper
to a theorem only when it is clear-cut and useful to do so. In the case of recent
results, I sometimes give all relevant references and an account of the historical
development, since this appears nowhere else.
Dilation Theory: A Guided Tour 553

There are other ways to present dilation theory, and by the end of the first section
the reader will find references to several alternative sources. Either because of my
ignorance, or because I had to make choices, some things were left out. I have not
been able to cover all topics that could fall under the title, nor did I do full justice to
the topics covered. After all this is just a survey, and that is the inevitable nature of
the genre.

Part 1: An Exposition of Classical Dilation Theory

1 The Concept of Dilations

The purpose of this introductory section is to present the notion of dilation, and to
give a first indication that this notion is interesting and can be useful.
Let H be a Hilbert space, and T ∈ B(H) be a contraction, that is, T is an
operator
√ such that T  ≤ 1. Then I − T ∗ T ≥ 0, and so we can define DT =
I − T ∗ T . Halmos [72] observed that the simple construction
 
T DT ∗
U=
DT −T ∗

gives rise to a unitary operator on H ⊕ H. Thus, every contraction T is a

compression of a unitary U , meaning that

T = PH U H , (1.1)

where PH denotes the orthogonal projection of H ⊕ H onto H ⊕ {0} ∼ = H. In this

situation we say that U is a dilation of T , and below we shall write T ≺ U to
abbreviate that U is a dilation of T .
This idea can be pushed further. Let K = HN+1 = H⊕· · ·⊕H be the (N +1)th-
fold direct sum of H with itself, and consider the following (N + 1) × (N + 1)
operator matrix
⎛ ⎞
T 0 0 · · · 0 DT ∗
⎜D 0 · · · 0 −T ∗ ⎟
⎜ T 0 ⎟
⎜ ⎟
⎜ 0 I 0 ··· 0 0 ⎟
U =⎜
⎜ 0 0 I 0 ⎟.
⎟ (1.2)
⎜ ⎟
⎜ .. .. .. .. ⎟
⎝ . . . . ⎠
0 0 ··· 0 I 0
554 O. M. Shalit

Egerváry [58] observed that U is unitary on B(K), and moreover, that

 k 
T ∗
Uk = , k = 1, 2, . . . , N, (1.3)
∗ ∗

in other words, if we identify H with the first summand of K, then

p(T ) = PH p(U ) H (1.4)

for every polynomial p of degree at most N. Such a dilation was called an N-

dilation in [95]. Thus, an operator U satisfying (1.1) might be referred to as a 1-
dilation, however, the recent ubiquity of 1-dilations has led me to refer to it simply
as a dilation.
We see that, in a sense, every contraction is a “part of” a unitary operator.
Unitaries are a very well understood class of operators, and contractions are as
general a class as one can hope to study. Can we learn anything interesting from
the dilation picture?
Theorem 1.1 (von Neumann’s Inequality [168]) Let T be a contraction on some
Hilbert space H. Then, for every polynomial p ∈ C[z],

p(T ) ≤ sup |p(z)|. (1.5)

|z|=1

Proof Suppose that the degree of p is N. Construct U on K = H ⊕ · · · ⊕ H (direct

sum N + 1 times) as in (1.2). Using (1.4), we find that

p(T ) = PH p(U ) H  ≤ p(U ) = sup |p(z)|,

z∈σ (U )

by the spectral theorem, where σ (U ) denotes the spectrum of U . Since for every
unitary U , the spectrum σ (U ) is contained in the unit circle T = {z ∈ C : |z| = 1},
the proof is complete. 
Remark 1.2 The above proof is a minor simplification of the proof of von Neu-
mann’s inequality due to Sz.-Nagy [163], which uses the existence of a unitary
power dilation (see the next section). The simplification becomes significant when
dim H < ∞, because then U is a unitary which acts on a finite dimensional space,
and the spectral theorem is then a truly elementary matter. Note that even the case
dim H = 1 is not entirely trivial: in this case von Neumann’s inequality is basically
the maximum modulus principle in the unit disc; as was observed in [145], this
fundamental results can be proved using linear algebra!
Remark  1.3 A matrix valued polynomial is a function z → p(z) ∈ Mn where
p(z) = N k=0 Ak z , and A1 , . . . , Ak ∈ Mn = Mn (C) (the n × n matrices over C).
k

Equivalently, a matrix valued polynomial is an n× n matrix of polynomials p =

(pij ), where the ij th entry is given by pij (z) = Nk=0 (Ak )ij z . If T ∈ B(H), then
k
Dilation Theory: A Guided Tour 555


we may evaluate a matrix valued polynomial p at T by setting p(T ) = N k=0 Ak ⊗
T k ∈ Mn ⊗ B(H), or, equivalently, p(T ) is the n × n matrix over B(H) with ij th
entry equal to pij (T ) (the operator p(T ) acts on the direct sum of H with itself n
times). It is not hard to see that if p is a matrix valued polynomial with values in Mn
and T is a contraction, then the inequality (1.5) still holds but with |p(z)| replaced
by p(z)Mn : first one notes that it holds for unitary operators, and then one obtains
it for a general contraction by the dilation argument that we gave.
The construction (1.2) together with Theorem 1.1 illustrate what dilation theory
is about and how it can be used: every object in a general class of objects (here,
contractions) is shown to be “a part of” an object in a smaller, better behaved class
(here, unitaries); the objects in the better behaved class are well understood (here,
by the spectral theorem), and thus the objects in the general class inherit some
good properties. The example we have just seen is an excellent one, since proving
von Neumann’s inequality for non-normal contractions is not trivial. The simple
construction (1.2) and its multivariable generalizations have other applications, for
example they lead to concrete cubature formulas and operator valued cubature
formulas [95, Section 4.3]. In Sect. 4 we will examine in detail a deeper application
of dilation theory—an operator theoretic proof of Pick’s interpolation theorem.
Additional applications are scattered throughout this survey.
Before continuing I wish to emphasize that “dilation theory” and even the word
“dilation” itself mean different things to different people. As we shall see further
down the survey, the definition changes, as do the goals and the applications. Besides
the expository essay [95], the subject is presented nicely in the surveys [8] and [13],
and certain aspects are covered in books, e.g., [2, 118], and [121] (the forthcoming
book [26] will surely be valuable when it appears). Finally, the monograph [23] is
an indispensable reference for anyone who is seriously interested in dilation theory
of contractions.

2 Classical Dilation Theory of a Single Contraction

2.1 Dilations of a Single Contraction

It is quite natural to ask whether one can modify the construction (1.2) so that (1.3)
holds for all k ∈ N, and not just up to some power N. We will soon see that the
answer is affirmative. Let us say that an operator U ∈ B(K) is a power dilation
of T ∈ B(H), if H is a subspace of K and if T k = PH U k H for all k ∈ N =
{0, 1, 2, . . .}. The reader should be warned that this is not the standard terminology
used, in the older literature one usually finds the word dilation used to describe what
we just called a power dilation, whereas the concept of N-dilation does not appear
much (while in the older-older literature one can again find power dilation).
In this and the next sections I will present what I and many others refer to as
classical dilation theory. By and large, this means the theory that has been pushed
556 O. M. Shalit

and organized by Sz.-Nagy and Foias (though there are many other contributors),
and appears in the first chapter of [23]. The book [23] is the chief reference for
classical dilation theory. The proofs of most of the results in this and the next section,
as well as references, further comments and historical remarks can be found there.
Theorem 2.1 (Sz.-Nagy’s Unitary Dilation Theorem [163]) Let T be a contrac-
tion on a Hilbert space H. Then there exists a Hilbert space K containing H and a
unitary U on K, such that

T k = PH U k H , for all k = 0, 1, 2, . . . (2.1)

Moreover, K can be chosen to be minimal in the sense that the smallest reducing
subspace for U that contains H is K. If Ui ∈ B(Ki ) (i = 1, 2) are two minimal
unitary dilations, then there exists a unitary W : K1 → K2 acting as the identity on
H such that U2 W = W U1 .
One can give a direct proof of existence of the unitary dilation by just writing
down an infinite operator matrix U acting on 2 (Z, H) = ⊕n∈Z H, similarly to
(1.2).
6 A nminimal dilation is then obtained by restricting U to the reducing nsubspace
n∈Z U H (the notation means the closure of the span of the subspaces U H). The
uniqueness of the minimal unitary dilation is then a routine matter. We will follow
a different path, that requires us to introduce another very important notion: the
minimal isometric dilation.
Before presenting the isometric dilation theorem, it is natural to ask whether it
is expected to be of any use. A unitary dilation can be useful because unitaries
are “completely understood” thanks to the spectral theorem. Are isometries well
understood? The following theorem shows that, in a sense, isometries are indeed
very well understood.
For a Hilbert space G, we write 2 (N, G) for the direct sum ⊕n∈N G. The
unilateral shift of multiplicity dim G (or simply the shift) is the operator S :
2 (N, G) → 2 (N, G) given by

S(g0 , g1 , g2 , . . .) = (0, g0 , g1 , . . .).

The space G is called the multiplicity space. Clearly, the shift is an isometry.
Similarly, the bilateral shift on 2 (Z, G) is defined to be the operator

U (. . . , g−2 , g−1 , g0 , g1 , g2 , . . .) = (. . . g−3 , g−2 , g−1 , g0 , g1 , . . .)

where we indicate with a box the element at the 0th summand of 2 (Z, G).
Theorem 2.2 (Wold Decomposition) Let V be an isometry on a Hilbert space H.
Then there exists a (unique) direct sum decomposition H = Hs ⊕ Hu such that
Hs and Hu are reducing for V , and such that V Hs is unitarily equivalent to a
unilateral shift and V Hu is unitary.
Dilation Theory: A Guided Tour 557

For the proof, the reader has no choice but to define Hu = ∩n≥0 V n H. Once one
shows that Hu is reducing, it remains to show that V H⊥ is a unilateral shift. Hint:
u
the multiplicity space is H  V H, (this suggestive notation is commonly used in
the theory; it means “the orthogonal complement of V H inside H”, which is in this
case just (V H)⊥ ).
Theorem 2.3 (Sz.-Nagy’s Isometric Dilation Theorem) Let T be a contraction
on a Hilbert space H. Then there exists a Hilbert space K containing H and an
isometry V on K, such that

T∗ = V∗ H (2.2)

and in particular, V is a power dilation of T . Moreover, K can be chosen to be

minimal in the sense that the minimal invariant subspace for V that contains H is
K. If Vi ∈ B(Ki ) (i = 1, 2) are two minimal isometric dilations, then there exists a
unitary W : K1 → K2 acting as the identity on H such that V2 W = W V1 .
Proof Set DT = (I − T ∗ T )1/2 —this is the so called defect operator, which
measures how far T is from being an isometry—and let D = DT (H). Construct
K = H ⊕ D ⊕ D ⊕ . . ., in which H is identified as the first summand. Now we
define, with respect to the above decomposition of K, the block operator matrix
⎛ ⎞
T 0 0 0 0 ···
⎜D · · ·⎟
⎜ T 0 0 0 0 ⎟
⎜ ⎟
⎜ 0 ID 0 0 0 · · ·⎟
V =⎜
⎜ 0 0 ID 0 0
⎟
⎟
⎜ ⎟
⎜ 0 0 0 ID 0 ⎟
⎝ ⎠
.. .. .. ..
. . . .

which6is readily seen to satisfy V ∗ H = T ∗ . That V is an isometry, and the fact that
K = n∈N V n H, can be proved directly and without any pain.
The uniqueness is routine, but let’s walk through it for once. If Vi ∈ B(Ki ) are
two minimal isometric dilations, then we can define a map W on span{V1n h : n ∈
N, h ∈ H} ⊆ K1 by first prescribing

W V1n h = V2n h ∈ K2 .

This map preserves inner products: assuming that m ≤ n, and using the fact that Vi
is an isometric dilation of T , we see

W V1m g, W V1n h = V2m g, V2n h = g, V2n−m h = g, T n−m h = V1m g, V1n h.
558 O. M. Shalit

The map W therefore well defines an isometry from the dense subspace span{V1n h :
n ∈ N, h ∈ H} ⊆ K1 onto the dense subspace span{V2n h : n ∈ N, h ∈ H} ⊆ K2 ,
and therefore extends to a unitary which, by definition, intertwines V1 and V2 . 
∗
Note that the minimal isometric dilation is actually a coextension: T ∗ = V H .
A coextension is always a power dilation, so T n = PH V n H for all n ∈ N,
but, of course, the converse is not true (see Theorem 3.10 and (3.6) below for the
general form of a power dilation). Note also that the minimality requirement is more
stringent than the minimality required from the minimal unitary dilation. The above
proof of existence together with the uniqueness assertion actually show that every
isometry which is a power dilation of T and is minimal in the sense of the theorem,
is in fact a coextension.
Because the minimal isometric dilation V is actually a coextension of T , the
adjoint V ∗ is a coisometric extension of T ∗ . Some people prefer to speak about
coisometric extensions instead of isometric coextensions.
Once the existence of an isometric dilation is known, the existence of a
unitary dilation follows immediately, by the Wold decomposition. Indeed, given a
contraction T , we can dilate it to an isometry V . Since V ∼ = S ⊕ Vu , where S is a
unilateral shift and Vu is a unitary, we can define a unitary dilation of T by U ⊕ Vu ,
where U is the bilateral shift of the same multiplicity as S. This proves the existence
part of Theorem 2.1; the uniqueness of the minimal unitary dilation is proved as for
the minimal isometric one.

2.2 A Glimpse at Some Applications of Single Operator

Dilation Theory

The minimal unitary dilation of a single contraction can serve as the basis of the
development of operator theory for non-normal operators. This idea was developed
to a high degree by Sz.-Nagy and Foias and others; see the monograph [23]
([23] also contains references to alternative approaches to non-normal operators,
in particular the theories of de Branges–Rovnyak, Lax–Phillips, and Livsič and his
school). The minimal unitary dilation can be used to define a refined functional
calculus on contractions, it can be employed to analyze one-parameter semigroups
of operators, it provides a “functional model” by which to analyze contractions and
by which they can be classified, and it has led to considerable progress in the study
of invariant subspaces.
To sketch just one of the above applications, let us briefly consider the functional
calculus (the following discussion might be a bit difficult for readers with little
background in function theory and measure theory; they may skip to the beginning
of Sect. 3 without much loss). We let H ∞ = H ∞ (D) denote the algebra of bounded
analytic functions on the open unit disc D. Given an operator T ∈ B(H), we wish
to define a functional calculus f → f (T ) for all f ∈ H ∞ . If the spectrum of T
is contained in D, then we can apply the holomorphic functional calculus to T to
Dilation Theory: A Guided Tour 559

define a homomorphism f → f (T ) from the algebra O(D) of analytic functions

on D into B(H). In fact, if f ∈ O(D) and σ (T ) ⊂ D, then we can simply plug T
into the power series of f . Thus, in this case we know how to define f (T ) for all
f ∈ H ∞.
Now, suppose that T is a contraction, but that σ (T ) is not contained in the open
disc D. Given a bounded analytic function f ∈ H ∞ , how can we define f (T )? Note
that the holomorphic functional calculus cannot be used, because σ (T ) contains
points on the circle T = ∂D, while not every f ∈ H ∞ extends to a holomorphic
function on a neighborhood of the closed disc.
The first case that we can treat easily is the case when f belongs to the disc
algebra A(D) ⊂ H ∞ , which is the algebra of all bounded analytic functions on the
open unit disc that extend continuously to the closure D. This case can be handled
using von Neumann’s inequality (Theorem 1.5), which is an immediate consequence
of Theorem 2.1. Indeed, it is not very hard to show that A(D) is the closure of the
polynomials with respect to the supremum norm f ∞ = sup|z|≤1 |f (z)|. If pn is a
sequence of polynomials that converges uniformly on D to f , and T is a contraction,
then von Neumann’s inequality implies that pn (T ) is a Cauchy sequence, so we can
define f (T ) to be limn pn (T ). It is not hard to show that the functional calculus
A(D) 4 f → f (T ) has all the properties one wishes for: it is a homomorphism
extending the polynomial functional calculus, it is continuous, and it agrees with
the continuous functional calculus if T is normal.
Defining a functional calculus for H ∞ is a more delicate matter, but here again
the unitary dilation leads to a resolution. The rough idea is that we can look at
the minimal unitary dilation U of T , and use spectral theory to analyze what can
be done for U . If the spectral measure of U is absolutely continuous with respect
to Lebesgue measure on the unit circle, then it turns out that we can define f (U )
for all f ∈ H ∞ , and then we can simply define f (T ) to be the compression of
f (U ) to H. In this case the functional calculus f → f (T ) is a homomorphism that
extends the polynomial and holomorphic functional calculi, it is continuous in the
appropriate sense, and it agrees with the Borel functional calculus when T is normal.
(Of course, this only becomes useful if one can find conditions that guarantee that
the minimal unitary dilation of T has absolutely continuous spectral measure. A
contraction T is said to be completely nonunitary (c.n.u.) if is has no reducing
subspace M such that the restriction T M is unitary. Every contraction splits as a
direct sum T = T0 ⊕T1 , where T0 is unitary and T1 is c.n.u. Sz.-Nagy and Foias have
shown that if T is c.n.u., then the spectral measure of its minimal unitary dilation is
absolutely continuous.)
If the spectral measure of U is not absolutely continuous with respect to
Lebesgue measure, then there exits a subalgebra HU∞ of H ∞ for which there exists
a functional calculus f → f (U ), and then one can compress to get f (T ); it was
shown that HU∞ is precisely the subalgebra of functions in H ∞ on which f → f (T )
is a well defined homomorphism. See [23, Chapter III] for precise details.
There are two interesting aspects to note. First, an intrinsic property of the
minimal dilation—the absolute continuity of its spectral measure—provides us with
560 O. M. Shalit

nontrivial information about T (whether or not it has an H ∞ -functional calculus).

The second interesting aspect is that there do exist interesting functions f ∈ H ∞
that are not holomorphic on a neighborhood of the closed disc (and are not even
continuous up to the boundary) for which we would like to evaluate f (T ). This
technical tool has real applications. See [23, Section III.8], for example.

3 Classical Dilation Theory of Commuting Contractions

3.1 Dilations of Several Commuting Contractions

The manifold applications of the unitary dilation of a contraction on a Hilbert

space motivated the question (which is appealing and natural in itself, we must
admit) whether the theory can be extended in a sensible manner to families of
operators. The basic problem is: given commuting contractions T1 , . . . , Td ∈ B(H),
to determine whether there exist commuting isometries/unitaries U1 , . . . , Ud on a
larger Hilbert space K ⊇ H such that

T1n1 · · · Tdnd = PH U1n1 · · · Udnd H (3.1)

for all (n1 , . . . , nd ) ∈ Nd . Such a family U1 , . . . , Ud is said to be an isomet-

ric/unitary dilation (I warned you that the word is used differently in different
situations!).
Clearly, it would be nice to have a unitary dilation, because commuting unitaries
are completely understood by spectral theory. On the other hand, isometric dilations
might be easier to construct. Luckily, we can have the best of both worlds, according
to the following theorem of Itô and Brehmer (see [23, Proposition I.6.2]).
Theorem 3.1 Every family of commuting isometries has a commuting unitary
extension.
Proof The main idea of the proof is that given commuting isometries V1 , . . . , Vd on
a Hilbert space H, one may extend them to commuting isometries W1 , . . . , Wd such
that (a) W1 is a unitary, and (b) if Vi is unitary then Wi is a unitary. Given that this
is possible, one my repeat the above process d times to obtain a unitary extension.
The details are left to the reader. 
In particular, every family of commuting isometries has a commuting unitary
dilation. Thus, a family of commuting contractions T1 , . . . , Td ∈ B(H) has an
isometric dilation if and only if it has a unitary dilation.
Theorem 3.2 (Andô’s Isometric Dilation Theorem [9]) Let T1 , T2 be two com-
muting contractions on a Hilbert space H. Then there exists a Hilbert space K ⊇ H
and two commuting isometries V1 , V2 on K such that

T1n1 T2n2 = PH V1n2 V2n2 H for all n1 , n2 ∈ N. (3.2)

Dilation Theory: A Guided Tour 561

In fact, V1 , V2 can be chosen such that Vi∗ H = Ti∗ for i = 1, 2.

In other words, every pair of contractions has an isometric dilation, and in fact, an
isometric coextension.
Proof The proof begins similarly to the proof of Theorem 2.3: we define the Hilbert
space K = ⊕n∈N H = H ⊕ H ⊕ · · · , and we define two isometries W1 , W2 by

Wi (h0 , h1 , h2 , . . .) = (Ti h0 , DTi h0 , 0, h1 , h2 , . . .)

for i = 1, 2. These are clearly isometric coextensions, but they do not commute:

Wj Wi (h0 , h1 , h2 , . . .) = Wj (Ti h0 , DTi h0 , 0, h1 , h2 , . . .)

= (Tj Ti h0 , DTj Ti h0 , 0, DTi h0 , 0, h1 , h2 , . . .).

Of course, in the zeroth entry we have equality T1 T2 h0 = T2 T1 h0 and from the

fifth entry on we have (h1 , h2 , . . .) = (h1 , h2 , . . .). The problem is that usually
DT1 T2 h0 = DT2 T1 h0 and DT1 h0 = DT2 h0 . However ,

DT1 T2 h0 2 + DT2 h0 2 = T2∗ (I − T1∗ T1 )T2 h0 , h0  + (I − T2∗ T2 )h0 , h0 

= (I − T2∗ T1∗ T1 T2 )h0 , h0 
= DT2 T1 h0 2 + DT1 h0 2 ,

and this allows us to define a unitary operator U0 : G := H ⊕ H ⊕ H ⊕ H → G that

satisfies

U0 (DT1 T2 h0 , 0, DT2 h0 , 0) = (DT2 T1 h0 , 0, DT1 h0 , 0).

Regrouping K = H ⊕ G ⊕ G ⊕ · · · , we put U = IH ⊕ U0 ⊕ U0 ⊕ · · · , and now

we define V1 = U W1 and V2 = W2 U −1 . The isometries V1 and V2 are isometric
coextensions—multiplying by U and U −1 did not spoil this property of W1 , W2 .
The upshot is that V1 and V2 commute; we leave this for the reader to check. 
As a consequence (by Theorem 3.1),
Theorem 3.3 (Andô’s Unitary Dilation Theorem [9]) Every pair of contractions
has a unitary dilation.
One can also get minimal dilations, but it turns out that in the multivariable
setting minimal dilations are not unique, so they are not canonical and don’t play a
prominent role. Once the existence of the unitary dilation is known, the following
two-variable version of von Neumann’s inequality follows just as in the proof of
Theorem 1.1.
562 O. M. Shalit

Theorem 3.4 Let T1 , T2 be two commuting contractions on a Hilbert space H.

Then for every complex two-variable polynomial p,

p(T ) ≤ sup |p(z)|.

z∈T2

Here and below we use the shorthand notation p(T ) = p(T1 , . . . , Td ) whenever
p is a polynomial in d variables and T = (T1 , . . . , Td ) is a d-tuple of operators. The
proof of the above theorem (which is implicit in the lines preceding it) gives rise to
an interesting principle: whenever we have a unitary or a normal dilation then we
have a von Neumann type inequality. This principle can be used in reverse, to show
that for three or more commuting contractions there might be no unitary dilation, in
general.
Example 3.5 There exist three contractions T1 , T2 , T3 on a Hilbert space H and a
complex polynomial p such that

p(T ) > p∞ := sup |p(z)|.

z∈T3

Consequently, T1 , T2 , T3 have no unitary, and hence also no isometric, dilation.

There are several concrete examples. The easiest to explain, in my opinion, is the
one presented by Crabb and Davie [39]. One takes a Hilbert space H of dimension
8 with orthonormal basis e, f1 , f2 , f3 , g1 , g2 , g3 , h, and defines partial isometries
T1 , T2 , T3 by

Ti e = fi
Ti fi = −gi
Ti fj = gk , k = i, j
Ti gj = δij h
Ti h = 0

for i, j, k = 1, 2, 3. Obviously, these are contractions, and checking that Ti Tj =

Tj Ti is probably easier than you guess. Now let p(z1 , z2 , z3 ) = z1 z2 z3 −z13 −z23 −z33 .
Directly evaluating we see that p(T1 , T2 , T3 )e = 4h, so that p(T1 , T2 , T3 ) ≥ 4.
On the other hand, it is elementary to show that |p(z)| < 4 for all z ∈ T3 , so by
compactness of T3 we get p∞ < 4, as required.
Remark 3.6 At more or less the same time that the above example appeared, Kaijser
and Varopoulos discovered three 5 × 5 commuting contractive matrices that do not
satisfy von Neumann’s inequality [165]. On the other hand, it was known (see [56,
p. 21]) that von Neumann’s inequality holds for any d-tuple of 2 × 2 matrices, in
fact, every such d-tuple has a commuting unitary dilation. It was therefore begged
of operator theorists to decide whether or not von Neumann’s inequality holds for
Dilation Theory: A Guided Tour 563

3-tuples of 3 × 3 and 4 × 4 commuting contractive matrices. Holbrook found a

4 × 4 counter example in 2001 [78], and the question of whether von Neumann’s
inequality holds for three 3 × 3 contractions remained outrageously open until
finally, only very recently, Knese [87] showed how results of Kosiński on the three
point Pick interpolation problem in the polydisc [88] imply that in the 3 × 3 case the
inequality holds (it is still an open problem whether or not every three commuting
3 × 3 contractions have a commuting unitary dilation; the case of four 3 × 3
contractions was settled negatively in [35]).
It is interesting to note that the first example of three contractions that do
not admit a unitary dilation did not involve a violation of a von Neumann type
inequality. Parrott [110] showed that if U and V are two noncommuting unitaries,
then the operators
     
00 0 0 0 0
T1 = , T2 = , T3 = (3.3)
I 0 U 0 V 0

are three commuting contractions that have no isometric dilation. However, these
operators can be shown to satisfy von Neumann’s inequality.
What is it exactly that lies behind this dramatic difference between d = 2 and
d = 3? Some people consider this to be an intriguing mystery, and there has been
effort made into trying to understand which d-tuples are the ones that admit a unitary
dilation (see, e.g., [160] and the references therein), or at least finding sufficient
conditions for the existence of a nice dilation.
A particularly nice notion of dilation is that of regular dilation. For a d-tuple T =
(T1 , . . . , Td ) ∈ B(H)d and n = (n1 , . . . , nd ) ∈ Nd , we write T n = T1n1 · · · Tdnd . If
n = (n1 , . . . , nd ) ∈ Zd , then we define

T (n) = (T n− )∗ T n+

where n+ = (max{n1 , 0}, . . . , max{nd , 0}) and n− = n+ − n. For a commuting

unitary tuple U and n ∈ Zd , we have U (n) = U n := U1n1 · · · Udnd . Now, if K
contains H and U = (U1 , . . . , Ud ) ∈ B(K)d a d-tuple of unitaries, say that U is a
regular dilation of T if

T (n) = PH U n H for all n ∈ Zd . (3.4)

Note that a unitary (power) dilation of a single contraction is automatically a regular

dilation, because applying the adjoint to (1.4) gives T (k) = PH U k H for all k ∈ Z.
However, a given unitary dilation of a pair of contractions need not satisfy (3.4),
and in fact there are pairs of commuting contractions that have no regular unitary
dilation.
In contrast with the situation of unitary dilations, the tuples of contractions that
admit a regular unitary dilation can be completely characterized.
564 O. M. Shalit

Theorem 3.7 (Regular Unitary Dilation) A d-tuple T = (T1 , . . . , Td ) of com-

muting contractions on a Hilbert space has a regular unitary dilation if and only
if,

(−1)m Ti∗1 · · · Ti∗m Ti1 · · · Tim ≥ 0 , for all S ⊆ {1, . . . , d}. (3.5)
{i1 ,...,im }⊆S

The conditions (3.5) are sometimes called Brehmer’s conditions. For the proof,
one shows that the function n → T (n) is a positive definite function on the group
Zd (see Section I.9 in [23]), and uses the fact that every positive definite function on
a group has a unitary dilation [23, Section I.7].
Corollary 3.8 The following are sufficient conditions for a d-tuple T =
(T1 , . . . , Td ) of commuting contractions on a Hilbert space to have a regular
unitary dilation:

1. di=1 Ti 2 ≤ 1.
2. T1 , . . . , Td are all isometries.
3. T1 , . . . , Td doubly commute, in the sense that Ti Tj∗ = Tj∗ Ti for all i = j (in
addition to Ti Tj = Tj Ti for all i, j ).
Proof It is not hard to show that the conditions listed in the corollary are sufficient
for Brehmer’s conditions (3.5) to hold. 

3.2 Commutant Lifting

We return to the case of two commuting contractions. The following innocuous

looking theorem, called the commutant lifting theorem, has deep applications (as we
shall see in Sect. 4) and is the prototype for numerous generalizations. It originated
in the work of Sarason [137], was refined by Sz.-Nagy and Foias (see [23]), and has
become a really big deal (see [63]).
Theorem 3.9 (Commutant Lifting Theorem) Let A be a contraction on a Hilbert
space H, and let V ∈ B(K) be the minimal isometric dilation of A. For every
contraction B that commutes with A, there exists an operator R ∈ B(K) such that
1. R commutes with V ,
2. B = R ∗ H ,
3. R ≤ 1.
In other words, every operator commuting with A can be “lifted” to an operator
commuting with its minimal dilation, without increasing its norm.
Proof Let U, W ∈ B(L) be the commuting isometric coextension of A, B, where
L is a Hilbert space that contains H (the coextension exists by Andô’s isometric
Dilation Theory: A Guided Tour 565

dilation theorem, Theorem 3.2). The restriction of the isometry U to the subspace
6
n∈N U H is clearly
n

1. an isometry,
2. a dilation of A,
3. a minimal dilation,
and6therefore (by uniqueness of the minimal isometric dilation), the restriction of U
to n∈N U n H is 6 unitarily equivalent to the minimal isometric dilation V on H, so
we identify K = n∈N U n H and V = U H . It follows (either from our knowledge
on the minimal dilation, or simply from the fact that U is a coextension) that V is a
coextension of A. With respect to the decomposition L = K ⊕ K⊥ ,
   
V X RQ
U= , W =
0 Z P N

It is evident that R ≤ 1 and that R is a coextension of B. We wish to show that

RV = V R.
From U W = W U we find that V R + XP = RV . Thus, the proof will be
complete if we show that X = 0. Equivalently, we have to show that K is invariant
under U ∗ . To see this, consider U ∗ U n h for some h ∈ H and n ∈ N. If n ≥ 1 we get
U n−1 h which is in K. If n = 0 then we get U ∗ h = A∗ h ∈ H ⊆ K, because U is a
coextension of A. That completes the proof. 

3.3 Dilations of Semigroups and Semi-Invariant Subspaces

Above we treated the case of a single operator or a tuple of commuting operators.

However, dilation theory can also be developed, or at least examined, in the context
of operator semigroups.
Let T = {Ts }s∈S ⊂ B(H) be a family of operators parametrized by a semigroup
S with unit e. Then T is a said to be a semigroup of operators over S if
1. Te = I ,
2. Tst = Ts Tt for all s, t ∈ S.
If S is a topological semigroup, then one usually requires the semigroup T to be
continuous in some sense. A semigroup V = {Vs }s∈S ⊂ B(K) is said to be a
dilation of T if K ⊃ H and if

Ts = PH Vs H , for all s ∈ S.

Note that Sz.-Nagy’s unitary dilation theorem can be rephrased by saying that
every semigroup of contractions over S = N has a unitary dilation, in the above
sense. Similarly, there are notions of extension and coextension of a semigroup
of operators. Some positive results have been obtained for various semigroups. Sz.-
566 O. M. Shalit

Nagy proved that every semigroup T = {Tt }t ∈R+ of contractions that is point-strong
continuous (in the sense that t → Tt h is continuous for all h ∈ H) has isometric and
unitary dilations, which are also point-strong continuous (see [23, Section I.10]).
This result was extended to the two parameter case by Słociński [151] and Ptak
[130]; the latter also obtained the existence of regular dilations for certain types
of multi-parameter semigroups. Douglas proved that every commutative semigroup
of isometries has a unitary extension [52]. Letting the commutative semigroup be
S = Nd , we recover Theorem 3.1. Douglas’s result was generalized by Laca to
semigroups of isometries parametrized by an Ore semigroup [92], and in fact to
“twisted” representations.
A result that somewhat sheds light on the question, which tuples of operators
have a unitary dilation and which don’t, is due to Opela. If T = {Ti }i∈I ⊂ B(H)
is a family of operators, we say that T commutes according to the graph G with
vertex set I , if Ti Tj = Tj Ti whenever {i, j } is an edge in the (undirected) graph G.
We can consider T as a semigroup parameterized by a certain quotient of the free
semigroup over I . Opela proved the following compelling result: given a graph G,
every family T = {Ti }i∈I of contractions commuting according to G has a unitary
dilation that commutes according to G, if and only if G contains no cycles [108].
It is interesting that in the general setting of semigroups of operators, one can say
something about the structure of dilations.
Theorem 3.10 (Sarason’s Lemma [136]) Let V = {Vs }s∈S ⊂ B(K) be a
semigroup of operators over a semigroup with unit S, and let H be a subspace
of K. Then the family T = {Ts := PH Vs H }s∈S is a semigroup over S if and only
if there exist two subspaces M ⊆ N ⊆ K, invariant under Vs for all s, such that
H = N  M := N ∩ M⊥ .
Proof The sufficiency of the condition is easy to see, if one writes the elements of
the semigroup V as 3 × 3 block operator matrices with respect to the decomposition
K = M ⊕ H ⊕ N ⊥.
For the converse, one has no choice but to define N = ∨s∈S Vs H (clearly an
invariant space containing H), and then it remains to prove that M := N  H is
invariant for V , or—what is the same—that PH Vt M = 0 for all t ∈ S. Fixing t, we
know that for all s,

PH Vt PH Vs PH = Tt Ts = Tt s = PH Vt s PH = PH Vt Vs PH .

It follows that PH Vt PH = PH Vt on ∨s Vs H = N . In particular, PH Vt M =

PH Vt PH M = 0 (since M ⊥ H), as required. 
The theorem describes how a general dilation looks like. A subspace H as above,
which is the difference of two invariant subspaces, is said to be semi-invariant for
the family V . In the extreme case where M = {0}, the space H = N is just an
invariant subspace for V , and V is an extension of T . In the other extreme case
when N = K, the space H is a coinvariant subspace for V , and V is a coextension
of T . In general, the situation is more complicated, but still enjoys some structure.
Dilation Theory: A Guided Tour 567

In the special S = N, case Sarason’s lemma implies that V is a (power) dilation of

an operator T if and only if it has the following block form:
⎛ ⎞
∗ ∗ ∗
⎝
V = 0T ∗⎠ . (3.6)
0 0 ∗

Sarason’s lemma is interesting and useful also in the case of dilations of a single
contraction.
Remark 3.11 Up to this point in the survey, rather than attempting to present a
general framework that encapsulates as much of the theory as possible, I chose
to sew the different parts together with a thin thread. There are, of course, also
“high level” approaches. In Sect. 7 we will see how the theory fits in the framework
of operator algebras, which is one unifying viewpoint (see also [41, 118, 121]).
There are other viewpoints. A notable one is due to Sz.-Nagy—very soon after
he proved his unitary dilation theorem for a single contraction, he found a far-
reaching generalization in terms of dilations of positive functions on ∗-semigroups;
see [164], which contains a theorem from which a multitude of dilation theorems
can be deduced (see also [162] for a more recent discussion with some perspective).
Another brief but high level look on dilation theory can be found in Arveson’s survey
[13].

4 An Application: Pick Interpolation

The purpose of this section is to illustrate how classical dilation theory can be
applied in a nontrivial way to prove theorems in complex function theory. The
example we choose is classical—Pick’s interpolation theorem—and originates in
the work of Sarason [137]. Sarason’s idea to use commutant lifting to solve the
Pick interpolation problem works for a variety of other interpolation problems as
well, including Carathéodory interpolation, matrix valued interpolation, and mixed
problems. It can also be applied in different function spaces and multivariable
settings. Here we will focus on the simplest case. Good references for operator
theoretic methods and interpolation are [2] and [63], and the reader is referred to
these sources for details and references.

4.1 The Problem

Recall that H ∞ denotes the algebra of bounded analytic functions on the unit disc
D = {z ∈ C : |z| < 1}. For f ∈ H ∞ we define

f ∞ = sup |f (z)|.
z∈D
568 O. M. Shalit

This norm turns H ∞ into a Banach algebra.

The Pick interpolation problem is the following: given n points z1 , . . . , zn in the
unit disc and n target points w1 , . . . , wn ∈ C, determine whether or not there exists
a function f ∈ H ∞ such that

f (zi ) = wi , for all i = 1, . . . , n, (4.1)

and

f ∞ ≤ 1. (4.2)

It is common knowledge that one can always find a polynomial (unique, if we take
it to be of degree less than or equal to n − 1) that interpolates the data, in the sense
that (4.1) holds. The whole point is that we require (4.2) to hold as well. Clearly,
this problem is closely related to the problem of finding the H ∞ function of minimal
norm that interpolates the points.
Recall that an n × n matrix A = (aij )ni,j =1 is said to be positive semidefinite if
for every v = (v1 , . . . , vn )T ∈ Cn


n
Av, v = aij vj v i ≥ 0.
i,j =1

If A is positive semidefinite, then we write A ≥ 0.

Theorem 4.1 (Pick’s Interpolation Theorem) Given points z1 , . . . , zn and
w1 , . . . , wn as above, there exists a function f ∈ H ∞ satisfying (4.1)-(4.2), if
and only if the following matrix inequality holds:
 n
1 − wi wj
≥ 0. (4.3)
1 − zi zj i,j =1

The n × n matrix on the left hand side of (4.3) is called the Pick matrix. What
is remarkable about this theorem is that it gives an explicit and practical necessary
and sufficient condition for the solvability of the interpolation problem: that the Pick
matrix be positive semidefinite.
At this point it is not entirely clear how this problem is related to operator theory
on Hilbert spaces, since there is currently no Hilbert space in sight. To relate this
problem to operator theory we will represent H ∞ as an operator algebra. The space
on which H ∞ acts is an interesting object in itself, and to this space we devote the
next subsection. Some important properties of H ∞ functions as operators will be
studied in Sect. 4.3, and then, in Sect. 4.4, we will prove Theorem 4.1.
Dilation Theory: A Guided Tour 569

4.2 The Hilbert Function Space H 2


The Hardy space H 2 = H 2 (D)is the space of analytic functions h(z) = ∞n=0 an z
n

on the unit disc D that satisfy |an | < ∞. It is not hard to see that H is a linear
2 2

space, and that

   
an z n , bn z n = an b n

is an inner product which makes H 2 into a Hilbert space, with norm


h2H 2 = |an |2 .

In fact, after noting that every (an )∞

n=0 ∈  :=  (N, C) gives rise to a power series
2 2

that converges (at least) in D, it is evident that the map

∞

(an )∞
n=0 → an z n
n=0

is a unitary isomorphism of 2 onto H 2 (D), so the Hardy space is a Hilbert space,

for free. The utility of representing a Hilbert space in this way will speak for itself
soon.
For w ∈ D, consider the element kw ∈ H 2 given by
∞
 1
kw (z) = w n zn = .
1 − zw
n=0

Then for h(z) = an zn , we calculate
   
h, kw  = an z n , w n zn = an wn = h(w).

We learn that the linear functional h → h(w) is a bounded functional, and that the
element of H 2 that implements this functional is kw . The functions kw are called
kernel functions, and the function k : D × D → C given by k(z, w) = kw (z)
is called the reproducing kernel of H 2 . The fact that point evaluation in H 2 is a
bounded linear functional lies at the root of a deep connection between function
theory on the one hand, and operator theory, on the other.
The property of H 2 observed in the last paragraph is so useful and important that
it is worth a general definition. A Hilbert space H ⊆ CX consisting of functions
on a set X, in which point evaluation h → h(x) is bounded for all x ∈ X, is said
to be a Hilbert function space or a reproducing kernel Hilbert space. See [119]
for a general introduction to this subject, and [2] for an introduction geared towards
570 O. M. Shalit

Pick interpolation (for readers that are in a hurry, Chapter 6 in [147] contains an
elementary introduction to H 2 as a Hilbert function space). If H is a Hilbert function
space on X, then by the Riesz representation theorem, for every x ∈ X there is an
element kx ∈ H such that h(x) = h, kx  for all h ∈ H, and one may define the
reproducing kernel of H by k(x, y) = ky (x).
The multiplier algebra of a Hilbert function space H on a set X is defined to be

Mult(H) = {f : X → C : f h ∈ H for all h ∈ H}.

Every f ∈ Mult(H) gives rise to a linear multiplication operator Mf : H → H

that acts as Mf h = f h, for all h ∈ H. By the closed graph theorem, multiplication
operators are bounded. The following characterization of multiplication operators is
key to some applications.
Proposition 4.2 Let H be a Hilbert function space on a set X. If f ∈ Mult(H),
then Mf∗ kx = f (x)kx for all x ∈ X. Conversely, if T ∈ B(H) is such that for all
x ∈ X there is some λx ∈ C such that T kx = λx kx , then there exists f ∈ Mult(H)
such that T = Mf∗ .
Proof For all h ∈ H and x ∈ X,

h, Mf∗ kx  = f h, kx  = f (x)h(x) = f (x)h, kx  = h, f (x)kx ,

so Mf∗ kx = f (x)kx . The converse is similar. 

Corollary 4.3 Every f ∈ Mult(H) is a bounded function, and

sup |f (x)| ≤ Mf .

x∈X

Proposition 4.4 Mult(H 2 ) = H ∞ and Mf  = f ∞ for every multiplier.

Proof Since 1 ∈ H 2 , every multiplier f = Mf 1 is in H 2 . In particular, every
multiplier is an analytic function. By the above corollary, Mult(H 2 ) ⊆ H ∞ , and
f ∞ ≤ Mf  for every multiplier f .

Conversely, if p(z) = N n
n=0 an z is a polynomial, then it is straightforward to
check that
 2π
1
p2H 2 = |p(eit )|2 dt.
2π 0

An approximation argument then gives

 2π
1
h2H 2 = lim |h(reit )|2 dt
rD1 2π 0

for all h ∈ H 2 . This formula for the norm in H 2 implies that H ∞ ⊆ Mult(H 2 ), and
that Mf  ≤ f ∞ . 
Dilation Theory: A Guided Tour 571

We will henceforth identify f with Mf , and we will think of H ∞ as a subalgebra

of B(H 2 ).

4.3 The Shift Mz

We learned that every bounded analytic function f ∈ H ∞ defines a bounded

multiplication operator Mf : H 2 → H 2 , but there is one that stands out as the most
important. If we abuse notation a bit and denote the identity function id : D → D
simply as z, then we obtain the multiplier Mz , defined by

(Mz h)(z) = zh(z).

It is quite clear that Mz is an isometry, and in fact it is unitarily equivalent to the

unilateral shift of multiplicity one on 2 (defined before Theorem 2.2). We will
collect a couple of important results regarding this operator, before getting back to
the proof of Pick’s theorem.
Recall that the commutant of a set of operators S ⊂ B(H) is the algebra

S = {T ∈ B(H) : ST = T S for all S ∈ S}.

Proposition 4.5 {Mz } = (H ∞ ) = H ∞ .

Proof Clearly H ∞ ⊆ (H ∞ ) ⊆ {Mz } . Now supposenthat T ∈ {Mz } . We claim
that T = Mf for f = T 1. Indeed, if p(z) = N
n=0 an z is a polynomial, then


N 
N
Tp = T an Mzn 1 = an Mzn T 1 = Mp f = fp.
n=0 n=0

An easy approximation argument would show that T = Mf , if we knew that f ∈

H ∞ ; but we still don’t. To finesse this subtlety, we find for an arbitrary h ∈ H 2 a
sequence of polynomials pn converging in norm h, and evaluate at all points w ∈ D,
to obtain:
n→∞
f (w)pn (w) = (Tpn )(w) −−−→ (T h)(w),

while f (w)pn (w) → f (w)h(w), on the other hand. This means that T h = f h for
all h and therefore f ∈ Mult(H 2 ) = H ∞ , as required. 
Let z1 , . . . , zn ∈ D. It is not hard to see that kz1 , . . . , kzn are linearly independent.
Let G = span{kz1 , . . . , kzn }, and let A = PG Mz G . By Proposition 4.2, G is
coinvariant for Mz , i.e., Mz∗ G ⊆ G, and A∗ = Mz∗ G is the diagonal operator given
by

A∗ : kzi → zi kzi . (4.4)

572 O. M. Shalit

We claim that Mz is the minimal isometric dilation of A. Well, it’s clearly an

isometric dilation, we just need to show that it is minimal. But kzi (z) = 1−zz 1
,
6 i
so kzi − zi Mz kzi = 1 ∈ n∈N Mz G. It follows that all the polynomials are in
n
6 6
n∈N Mz G, whence H = n∈N Mz G.
n 2 n

More generally, if we have a multiplier f , and we define B = PG Mf G , then we

have that B ∗ = Mf∗ G and that

B ∗ : kzi → f (zi )kzi . (4.5)

4.4 Proof of Pick’s Interpolation Theorem

We can now prove Theorem 4.1. We first show that (4.3) is a necessary condition.
Suppose that f ∈ H ∞ satisfies f (zi ) = wi for all i = 1, . . . , n and that f ∞ ≤ 1.
Define B = PG Mf G , where G = span{kz1 , . . . , kzn } as in the previous subsection.
Then, by (4.5) B ∗ is the diagonal operator given by B ∗ kzi = wi kzi . Since Mf  ≤
1, also B ∗  ≤ 1, thus for all α1 , . . . , αn ∈ C,
; n ;2 ; ;2 ; n ;2 ; n ;2
; ; ;  n ; ; ; ; ;
; ; ; ∗ ; ; ; ; ;
0≤; αi kzi ; − ;B αi kzi ; = ; αi kzi ; − ; wi αi kzi ;
; ; ; ; ; ; ; ;
i=1 i=1 i=1 i=1


n
  
= αj 1 − wj wi αi kzj , kzi
i,j =1


n  
1 − wi wj
= αj αi .
1 − zi zj
i,j =1

That is, the Pick matrix is positive semidefinite, and (4.3) holds.
Conversely, suppose that (4.3) holds. Define a diagonal operator D : G → G
by Dkzi = wi kzi for i = 1, . . . , n, and let B = D ∗ . Then the above computation
can be rearranged to show that B = B ∗  ≤ 1. Now, the diagonal operator B ∗
clearly commutes with the diagonal operator A∗ = Mz∗ G, so B commutes with
A. Since Mz is the minimal isometric dilation of A, the commutant lifting theorem
(Theorem 3.9) implies that B has a coextension to an operator T that commutes
with Mz and has T  ≤ 1. By Proposition 4.5, T = Mf for some f ∈ H ∞ , and
by Proposition 4.2, f (zi ) = wi for all i = 1, . . . , n. Since f ∞ = T  ≤ 1, the
proof is complete.
Dilation Theory: A Guided Tour 573

5 Spectral Sets, Complete Spectral Sets, and Normal

Dilations

Classical dilation theory does not end with dilating commuting contractions to
commuting unitaries. Let us say that a d-tuple N = (N1 , . . . , Nd ) is a normal
tuple if N1 , . . . , Nd are all normal operators and, in addition, they all commute with
one another. Recall that the joint spectrum σ (N) of a normal tuple is the set

σ (N) = {(ρ(N1 ), . . . , ρ(Nd )) : ρ ∈ M(C ∗ (N))} ⊂ Cd ,

where M(C ∗ (N)) is the space of all nonzero complex homomorphisms from the
unital C*-algebra C ∗ (N) generated by N to C. If N acts on a finite dimensional
space, then the joint spectrum is the set of joint eigenvalues, belonging to an
orthogonal set of joint eigenvectors that simultaneously diagonalize N1 , . . . , Nd .
A commuting tuple of unitaries U = (U1 , . . . , Ud ) is the same thing as a normal
tuple with joint spectrum contained in the torus Td . Since normal tuples are in a
sense “completely understood”, it is natural to ask which operator tuples T have a
normal dilation N (where the definition of dilation is as in (3.1)) with the spectrum
σ (N) prescribed to be contained in some set X ⊂ Cd .
Suppose that T = (T1 , . . . , Td ) is a commuting tuple of operators and that N =
(N1 , . . . , Nd ) is a normal dilation with σ (N) = X ⊂ Cd . Then we immediately
find that

p(T ) ≤ p(N) = sup |p(z)|

z∈X

for every polynomial p in d variables. In fact, it is not too hard to see that the above
inequality persists when p is taken to be a rational function that is regular on X. This
motivates the following definition: a subset X ⊆ Cd is said to be a K-spectral set
for T if X contains the joint spectrum σ (T ) of T , and if for every rational function
f that is regular on X,

f (T ) ≤ Kf X,∞ , (5.1)

where f X,∞ = supz∈X |f (z)|. If X is a K-spectral set for T with K = 1, then it

is simply said to be a spectral set for T .
I do not wish to define the joint spectrum of a non-normal commuting tuple, nor
to go into how to evaluate a rational function on a tuple of operators, so I will be
somewhat sloppy in what follows (see [11, Section 1.1]; for a textbook treatment, I
recommend also [118]). Two simplifying comments are in order:
1. In the case d = 1, i.e., just one operator T , the spectrum σ (T ) is the usual
spectrum, and the evaluation f (T ) of a rational function on T can be done
naturally, and this is the same as using the holomorphic functional calculus.
574 O. M. Shalit

2. One may also discuss polynomial spectral sets, in which (5.1) is required to
hold only for polynomials [38]. If X is polynomially convex (and in particular, if
X is convex), then considering polynomials instead of rational functions leads to
the same notion.
Thus, with the terminology introduced above, we can rephrase Theorem 3.4 by
2
saying that the bidisc D is a spectral set for every pair T = (T1 , T2 ) of commuting
contractions, and Example 3.5 shows that there exists three commuting contractions
3
for which the tridisc D is not a spectral set.
The notion of a spectral set for a single operator is due to von Neumann [168].
A nice presentation of von Neumann’s theory can be found in Sections 153–155
of [133]. The reader is referred to [19] for a rather recent survey with a certain
emphasis on the single variable case. To give just a specimen of the kind of result
that one can encounter, which is quite of a different nature than what I am covering
in this survey, let me mention the result of Crouzeix [40], which says that for every
T ∈ B(H), the numerical range W (T ) := {T h, h : h = 1} of T is a K-spectral
set for some K ≥ 2 (it is easy to see that one cannot have a constant smaller than 2;
Crouzeix conjectured that K = 2, and this conjecture is still open at the time of me
writing this survey).
It is plain to see that if T has a commuting normal dilation N with spectrum
σ (N) ⊆ X, then X is a polynomial spectral set for T , and it is true that in fact
X is a spectral set. It is natural to ask whether the converse implication holds, that
is, whether the assumption that a set X is a spectral set for a tuple T implies that
there exists a normal dilation with spectrum constrained to X (or even to the Shilov
boundary ∂X). There are cases when this is true (see [19]), but in general the answer
is no. For example, we already mentioned that Parrott’s example [110] of three
commuting contractions that have no unitary dilation (hence also no normal dilation
3
with spectrum contained in D ) does not involve a violation of von Neumann’s
3
inequality, in other words the tuple T from (3.3) has D as a spectral set but has no
unitary dilation.
The situation was clarified by Arveson’s work [11], where the notion of complete
spectral set was introduced. To explain this notion, we need matrix valued polyno-
mials and rational functions. Matrix valued polynomials in several commuting (or
noncommuting) variables, and the prescriptions for evaluating them at d-tuples of
commuting (or noncommuting) operators, are defined in a similar manner to their
definition in the one variable case in Remark 1.3. Once one knows how to evaluate
a rational function in several variables at a commuting tuple, the passage to matrix
valued rational functions is done similarly.
Given a tuple T ∈ B(H)d of commuting contractions, we say that a set X ⊂ Cd
is a complete K-spectral set for T , if σ (T ) ⊆ X and if for every matrix valued
rational function f that is regular on X, (5.1) holds, where now for an n × n matrix
valued rational function f X,∞ = supz∈X f (z)Mn . If X is a complete K-spectral
set for T with K = 1, then it is simply said to be a complete spectral set for T .
Dilation Theory: A Guided Tour 575

Theorem 5.1 (Arveson’s Dilation Theorem [11]) Let T = (T1 , . . . , Td ) be a

tuple of commuting operators on a Hilbert space H. Let X ⊂ Cd be a compact set
and let ∂X be the Shilov boundary of X with respect to the algebra rat (X) ⊆ C(X)
of rational functions that are regular on X. Then X is a complete spectral set for
T if and only if only if there exists a normal tuple N = (N1 , . . . , Nd ) acting on a
Hilbert space K ⊇ H, such that σ (N) ⊆ ∂X and for every matrix valued rational
function f that is regular on X,

f (T ) = PH f (N) H .

Putting Arveson’s dilation theorem together with some comments made above,
3
we see that D is a spectral set for the triple T from (3.3), but it is not a
complete spectral set. On the other hand, we know that for every pair of commuting
2
contractions T = (T1 , T2 ), the bidsic D is a complete spectral set. Agler and
McCarthy proved a sharper result: if T = (T1 , T2 ) acts on a finite dimensional
space, and T1 , , T2  < 1, then there exists a one dimensional complex algebraic
subvariety V ⊆ D2 (in fact, a so-called distinguished variety, which means that
V ∩ ∂(D2 ) = V ∩ T2 ), such that V is a complete spectral set for T [3].
If X ⊂ C is a spectral set for an operator T , one may ask whether or not it is
a complete spectral set. We close this section by mentioning some notable results
in this direction. It is known that if X ⊂ C is a compact spectral set for T such
that rat (X) + rat (X) is dense in C(∂X), then X is a complete spectral set, and T
has a normal dilation with spectrum in ∂X. The condition is satisfied, for example,
when X is the closure of a bounded and simply connected open set (this result is due
to Berger, Foias and Lebow (independently); see [118, Theorem 4.4]). The same is
true if X is an annulus (Agler [1]), but false if X is triply connected (Agler et al. [5]
and Dritschel and McCullough [54]).
If a pair of commuting operators T = (T1 , T2 ) has the symmetrized bidisc  :=
{(z1 + z2 , z1 z2 ) : z1 , z2 ∈ D} as a spectral set, then in fact  is a complete spectral
set for T (Agler and Young [4]). Pairs of operators having  as a spectral set have
a well developed model theory (see, e.g., Sarkar [139] and the references therein).
Building on earlier work of Bhattacharyya et al. [29], and inspired by Agler and
McCarthy’s distinguished varieties result mentioned above, Pal and Shalit showed
that if  is a spectral set for a pair T = (T1 , T2 ) of commuting operators acting on a
finite dimensional space, then there exists a distinguished one dimensional algebraic
variety V ⊆  which is a complete spectral set for T [109].
576 O. M. Shalit

Part 2: A Rapid Overview of Dilation Theories

6 Additional Results and Generalizations of Dilation Theory

6.1 Some Further Remarks on N -Dilations

The notion of a 1-dilation of a single operator, which is usually referred to simply

as dilation, has appeared through the years and found applications in operator
theory; see e.g. [22, 36, 72] (the reader should be warned that the terminology is
not universally accepted; for example, as we already mentioned, a power dilation
is usually simply referred to as dilation. Even more confusingly, in [22], a unitary
N-dilation of T means what we call here a unitary 1-dilation of T that acts on
H ⊕ CN ).
Egerváry’s simple construction (1.2) of an N-dilation, and with it the concept of
N-dilations, have been largely forgotten until [95] seemed to revive some interest
in it (see also [106]). The motivation was that the well-known Sz.-Nagy unitary
(power) dilation of a contraction T (given by Theorem 2.1) always acts on an
infinite dimensional space whenever T is nonunitary, even if T acts on a finite
dimensional space. Arguably, one cannot say that an infinite dimensional object
is better understood than a matrix. That’s what led to the rediscovery of (1.2)
and thence to the dilation-theoretic proof of von Neumann’s inequality that we
presented, which has the conceptual advantage of never leaving the realm of finite
dimensional spaces, in the case where T acts on a finite dimensional space to begin
with.
Let T = (T1 , . . . , Td ) be a d-tuple of commuting operators acting on a Hilbert
space H, and let U = (U1 , . . . , Ud ) be a d-tuple of commuting operators acting on
a Hilbert space K ⊇ H. We say that U is a an N-dilation of T if

p(T ) = PH p(U ) H

for every polynomial in d complex variables of degree less than or equal to N. We

say that this dilation is a unitary/normal dilation if every Ui (i = 1, . . . , d) is
unitary/normal. The construction (1.2) shows that every contraction has a unitary
N-dilation acting on HN+1 . In particular, it shows that every contraction acting on
a finite dimensional space has a unitary N-dilation acting on a finite dimensional
space, for all N.
Curiously, it appears that the proof of Theorem 3.2 cannot be modified to show
that every pair of commuting contractions on a finite dimensional space has a
commuting unitary N-dilation on a finite dimensional space, for all N. It was
shown by McCarthy and Shalit that indeed such a finitary version of Andô’s dilation
theorem holds [99]. Interestingly, the proof made use of Andô’s dilation theorem.
So, if one uses this finitary dilation theorem to prove von Neumann’s inequality
for pairs of matrices, one does not truly avoid infinite dimensional spaces. It is an
Dilation Theory: A Guided Tour 577

open problem to come up with an explicit construction of a unitary N-dilation for

commuting matrices.
In fact, in [99] it was also proved that a d-tuple of contractions acting on a finite
dimensional space has a unitary dilation if and only if for all N it has a unitary
N-dilation acting on a finite dimensional space. Likewise, it was shown that for
such a tuple, the existence of a regular unitary dilation is equivalent to the existence
of a regular unitary N-dilation (you can guess what that means) acting on a finite
dimensional space, for all N. Additional finitary dilation results appeared, first in the
setting of normal dilations of commuting tuples [38], and then in the setting of 1-
dilations of noncommuting operators [46, Section 7.1]. A similar phenomenon was
also observed in [69]. At last, Hartz and Lupini found a finite dimensional version
of Stinespring’s dilation theorem (see Sect. 7.1), which provides a general principle
by which one can deduce finite dimensional dilation theorems from their infinite
dimensional counterparts [73].
It is interesting to note that N-dilations found an application in simulating open
quantum systems on a quantum computer [79], and they also appeared in the context
of quantum information theory [94]. The notion of N-dilations also appeared in the
dilation theory in general Banach spaces (about which will say a few words below),
see [62].

6.2 Models

Another direction in which dilation theory for commuting d-tuples has been
developed is that of operator models. Roughly, the idea is that certain classes of
d-tuples of operators can be exhibited as the compressions of a particular “model”
d-tuple of operators. We will demonstrate this with a representative example; for a
broader point of view see [105], Chapter 14 in [2], or the surveys [8] and [138].
Our example is the d-shift on the Drury-Arveson space Hd2 [12, 55] (see also the
survey [146]).
 For a fixed d, we let Hd2 denote the space of all analytic functions
f (z) = α cα zα on the unit ball Bd such that (with standard multi-index notation)

 α!
f 2H 2 := |cα |2 < ∞.
d
α
|α|!

This norm turns the space Hd2 into a Hilbert space of analytic functions on Bd , such
that point evaluation is bounded. In fact, Hd2 is the reproducing kernel Hilbert space
determined by the kernel k(z, w) = 1−z,w1
. Some readers might jump to their feet
and object that this space is nothing but the good old symmetric Fock space, but it
is fruitful and enlightening to consider it as a space of analytic functions (so please,
sit down).
For the record, let the reader know that the possibility d = ∞ is allowed, but we
do not dwell upon this point.
578 O. M. Shalit

On Hd2 there is a d-tuple of operators S = (S1 , . . . , Sd ), called the d-shift, and

defined by

Si f (z) = zi f (z) , i = 1, . . . , d,

where z = (z1 , . . . , zd ) is the complex variable, and so Si is multiplication by the

ith coordinate function zi . The tuple S is plainly a commuting tuple: Si Sj = Sj Si
(multiplication
 of functions is commutative). A short combinatorial exercise shows
that Si Si∗ is equal to the orthogonal projection onto∗ the orthogonal complement
of the constant functions, and in particular Si Si ≤ I . Thus S is a row
contraction, meaning that the row operator [S1 S2 · · · Sd ] : Hd2 ⊕ · · · ⊕ Hd2 → Hd2
is a contraction. Another calculation reveals that S is pure, in the sense that
 α α ∗ n→∞
|α|=n S (S ) −−−→ 0 in the strong operator topology.
The remarkable fact is that Hd2 is a universal model for pure commuting row
contractions. I will now explain what these words mean. If G is a Hilbert space,
we can consider the space Hd2 ⊗ G (which can be considered as a Hilbert space of
analytic G-valued functions), and the d-shift promotes to a shift S ⊗ IG on Hd2 ⊗ G,
which is called a multiple of the d-shift. A subspace M ⊆ Hd2 ⊗ G is said to be
coinvariant if it is invariant for Si∗ ⊗ IG for all i = 1, . . . , d.
Theorem 6.1 (Universality of the d-Shift) Let T = (T1 , . . . , Td ) ∈ B(H)d be
a pure, commuting row contraction. Then there exists a Hilbert space G and a
coinvariant subspace M ⊆ Hd2 ⊗ G such that T is unitarily equivalent to the
compression of S ⊗ IG to M.
Thus, every row contraction T is unitarily equivalent to the corestriction of a
multiple of the d-shift to a coinvariant subspace. In particular, for every polynomial
p in d variables,
;  ;
p(T ) = ;PM p(S) ⊗ IG M ; ≤ p(S) , (6.1)

and this inequality replaces von Neumann’s inequality in this setting (and this was
Drury’s motivation [55]). It can be shown [55] (see also [12, 44]) that there exists
no constant C such that P (S) ≤ C supz∈Bd |p(z)|, and in particular, commuting
row contractions in general do not have normal dilations with spectrum contained
in Bd .

6.3 Dilation Theory for Noncommutative Operator Tuples

Dilation theory also plays a role in the analysis of tuples of noncommuting

operators.
 Recall that a row contraction is a tuple T = (T1 , . . . , Td ) such that
Ti Ti∗ ≤ I (as in Sect. 6.2, we allow, but do not belabor, the case d = ∞, in which
case the sum is understood in the strong-operator topology sense). A row isometry
Dilation Theory: A Guided Tour 579

is a tuple V = (V1 , . . . , Vd ) such that Vi∗ Vj = δij I , for all i, j . Thus, the operators
V1 , . . . , Vd are all isometries which have mutually orthogonal ranges, and this is
equivalent to the condition that the row operator [V1 V2 · · · Vd ] is an isometry. The
Sz.-Nagy isometric dilation theorem extends to the setting of (noncommuting) row
contractions. The following theorem is due to Frazho [64] (the case d = 2), Bunce
[33] (the case d ∈ N∪{∞}) and Popescu [123] (who proved the existence of dilation
in the case d ∈ N ∪ {∞}, and later developed a far reaching generalization of Sz-
Nagy’s and Foias’s theory for noncommuting tuples and more).
Theorem 6.2 (Row Isometric Dilation of Row Contractions) Let T ∈ B(H)d
be a row contraction. Then there exists a Hilbert space K containing H and a row
isometry V = (V1 , . . . , Vd ) ∈ B(K)d such that Vi∗ H = Ti∗ for all i.
There is also a very closely related dilation result, that shows that the shift
L = (L1 , . . . , Ld ) on the full Fock space is a universal model for pure row
contractions, which reads similarly to Theorem 6.1, with the free shift L replacing
the commutative shift S. Correspondingly, there is a von Neumann type inequality
p(T ) ≤ p(L) which holds for every row contraction T and every polynomial
p in noncommuting variables [123, 124].
Popescu has a large body of work in which this dilation/model theory is
developed, applied, and generalized. In particular, the theory can be modified to
accommodate tuples satisfying certain polynomial relations [127] (see also [150,
Section 8]) or tuples in certain noncommutative polydomains [128].
The isometric dilation of a row contractions lies at the heart of the free functional
calculus for row contractions (see, e.g., [126]), and is important for understanding
the algebraic structure of noncommutative Hardy algebras (also called analytic
Toeplitz algebras, see [45]), as well as for the study and classification of algebras of
bounded nc analytic functions on the nc unit ball and its subvarieties [134, 135].

6.4 Dilations in Banach Spaces

Until now, we have only considered operators on Hilbert spaces. But there are
other kinds of interesting spaces, and the concept of dilations has appeared and
been used in various settings. In the setting of Banach spaces, one may hope to
dilate a contraction to an invertible isometry (that is, a surjective isometry); more
generally one may wish to dilate a semigroup of operators to a group representation.
Results along these lines, including a direct analogoue of Sz.-Nagy’s unitary dilation
theorem, were obtained by Stroescu; see [161].
However, Banach spaces form a huge class of spaces, and the dilation theory in
the context of general Banach spaces contains the additional aspect that one might
like to ensure that the dilating space shares some properties with the original space.
For example, if T is a contraction on an Lp -space, one might wish to dilate to an
invertible isometry acting on an Lp -space. Moreover, if T is positive, in the sense
that Tf ≥ 0 (almost everywhere) whenever f ≥ 0 (almost everywhere), then one
580 O. M. Shalit

might hope to dilate to a positive invertible isometry. The following theorem is an

example of the kind of result one can look for.
Theorem 6.3 (Akcoglu–Sucheston [6]) Let T : X → X be a positive contraction
on an Lp -space X = Lp (μ) (1 ≤ p < ∞). Then there exists another Lp -space
Y = Lp (ν), a positive invertible isometry U : Y → Y , a positive isometry J : X →
Y , and a positive projection Q : Y → Y such that

J T n = QU n J , for all n ∈ N.

Note that even in the case p = 2, this is not exactly Sz.-Nagy’s dilation theorem:
the assumptions are stronger, but so is the conclusion. For a modern approach to
dilations in Banach spaces, generalizations, and also an overview of the history
of the theory and its applications, see [62]. Operator algebras are another class of
spaces in which dilation theory was developed and applied; we will discuss this
setting in Sects. 7 and 8 below.

6.5 Dilations of Representations of C*-Correspondences

A Hilbert C*-module is a complex linear space E which is a right module over a

C*-algebra A, which carries an “A-valued inner product” ·, · : E × E → A, that
satisfies the following conditions:
1. x, x ≥ 0 for all x ∈ E,
2. x, ya = x, ya for all x, y ∈ E and a ∈ A,
3. x, y = y, x∗ for all x, y ∈ E,
4. x, αy + βz = αx, y + βx, z for all x, y, z ∈ E and α, β ∈ C,
5. x := x, x1/2 is a norm on E which makes E into a Banach space.
The notion was introduced by Kaplansky [83] for the case where the C*-algebra A
is commutative, and then developed further by Paschke [112] and Rieffel [131] for
general C*-algebras. It is now a standard tool in some fields in operator algebras;
see [93] or Part I of [154] for an introduction.
Hilbert modules evolved into a more refined notion, called Hilbert correspon-
dences, that involves a left action. A linear operator T : E → E is said to be
adjointable if there exists a linear operator S : E → E so that T x, y = x, Sy
for all x, y ∈ E. One can show that every adjointable operator is a bounded right
module map, but the converse is not true. The set of all adjointable operators on
a Hilbert C*-correspondence E is denoted B a (E) or L(E); it is a C*-algebra. A
Hilbert C*-correspondence from A to B is a Hilbert B-module E which also
carries a left action of A by adjointable operators. If A = B then we say C*-
correspondence over A.
Given a Hilbert C*-correspondence E over the C*-algebra A, a covariant
representation of E on a Hilbert space H is a pair (T , σ ) where T is linear map
Dilation Theory: A Guided Tour 581

T : E → B(H) and σ : A → B(H) is a nondegenerate ∗-representation such

that T (a · x · b) = σ (a)T (x)σ (b) for all a, b ∈ A and x ∈ E. A covariant
representation is said to be contractive/completely contractive/bounded, etc., if
T is contractive/completely contractive/bounded, etc; it is said to be isometric if
T (x)∗ T (y) = σ (x, y) for all x, y ∈ E.
Muhly and Solel proved that every completely contractive covariant representa-
tion (T , σ ) of E on H has an isometric dilation (V , π) of E on K ⊇ H, [100,
Theorem 3.3]. By this, we mean an isometric covariant representation (V , π) on a
Hilbert space K that contains H, such that
1. H is reducing for π, and PH π(a) H = σ (a) for all a ∈ A,
2. PH V (x) H = T (x) for all x ∈ E,
3. PH V (x) H⊥ = 0 for all x ∈ E.
Moreover, they proved that such an isometric dilation can be chosen to be minimal
in a certain sense, and that the minimal isometric dilation is unique up to unitary
equivalence (the third condition that an isometric dilation is required to satisfy
looks more like something that should be called a coextension, it is actually a
consequence of minimality; sometimes it is not required). The isometric dilation
theorem was used in [100] to analyze the representation theory of the tensor
algebra T+ (E), which is a particular nonselfadjoint operator algebra, formed from
the C*-correspondence in a way which we shall not go into. This has shed light
on problems regarding an enormous class of operator algebras. Remarkably, Muhly
and Solel’s minimal isometric dilation enjoys also a commutant lifting theorem, and
this, in turn, can lead to a Nevanlinna-Pick type interpolation theorem for so-called
noncommutative Hardy algebras, with a proof reminiscent to the one we gave in
Sect. 4 (see [103]).
This dilation theorem of Muhly and Solel is a far reaching generalization of Sz.-
Nagy’s isometric dilation theorem (Theorem 2.3). In fact, the latter is obtained from
the simplest case E = A = C of Muhly and Solel’s theorem. The row-isometric
dilation of a row contraction (Theorem 6.2) is obtained as the “second simplest” case
E = Cd and A = C. Muhly and Solel’s isometric dilation theorem also reduces to
some known dilation results in the context of crossed product and semi-crossed
product operator algebras, as well in graph C*-algebras.
On the other hand, Andô’s theorem, for example, is not a special case of Muhly
and Solel’s isometric dilation theorem—a single C*-correspondence is not sufficient
to encode a pair of commuting contractions. The missing ingredient is the notion of
product systems. A product system over a monoid (i.e., a semigroup with unit e) S
is a family E = {Es }s∈S of C*-correspondences over a C*-algebra A, such that for
every s, t ∈ S there exists an isomorphism of correspondences (i.e., an adjointable
surjective isometry which is a bimodule map) us,t : Es E Et → Est such that the
multiplication xs yt := us,t (xs E yt ) is associative

(wr xs )yt = wr (xs yt ).

582 O. M. Shalit

(Here Es E Et , denotes the internal (or interior) tensor product of Es and Et ,

sometimes also denoted Es ⊗ Et ; see [93, Chapter 4].) A covariant representation
of a product system E = {Es }s∈S on H is a family T = {Ts }s∈S such that for all
s ∈ S, the pair (Ts , Te ) is a covariant representation of Es on H, which satisfies in
addition

Tst (xs E yt ) = Ts (xs )Tt (yt )

for all s, t ∈ S and all xs ∈ Es and yt ∈ Et . An isometric representation of E

on K is a covariant representation V = {Vs }s∈S of E such that for all s ∈ S, the
pair (Vs , Ve ) is an isometric representation. One then says that V is an isometric
dilation of T if
1. H is reducing for π, and PH Te (a) H = Ve (a) for all a ∈ A,
2. PH V (x) H = T (x) for all x ∈ Es .
The theory of isometric dilations of completely contractive representations of
product systems, is analogous to the theory of isometric dilations of semigroups of
contractions. Moreover, some of the proofs rely on the same ideas and approaches,
albeit at a technical sophistication level that is one order of magnitude higher. In fact,
several results (but not all) can be reduced to the case of operator semigroups (see
[143]). We will see in Sect. 8 that the dilation theory of covariant representations is
important for the dilation theory of CP-semigroups.
Here are some sample results. Solel proved a version of Andô’s theorem in this
setting: every completely contractive covariant representation of a product system
over N2 has an isometric dilation [157, Theorem 4.4]. Solel also proved an analogue
of Theorem 3.7 (regular dilations) for product systems over Nd using a direct proof
[158] (see also [152, 153]). Shalit later found another proof by reducing to the case
of operator semigroups [143]. The method of [143] was later used in [141, Section
5] to prove a counterpart to Theorem 3.1 (see also [66]). Vernik generalized Opela’s
result on dilations of contractions commuting according to a graph (see Sect. 3.3) to
the setting of product system representations [166]. All of the above results reduce
to their counterparts that we discussed in earlier sections, when one considers the
special case A = Es = C for all s ∈ S, where S is the appropriate monoid.

7 The Operator Algebraic Perspective

The operator algebraic outlook on dilation theory began with Arveson’s visionary
papers [10, 11]. Arveson sought to develop a systematic study of nonselfadjoint
operator algebras, which is based on studying the relations between an operator
algebra and the C*-algebras that it generates. From the outset, the approach was
general and powerful enough to cover also certain operator spaces. On the one hand,
this approach opened the door by which operator algebraic techniques entered into
Dilation Theory: A Guided Tour 583

operator theory: these techniques have shed light on classical dilation theory, and
they also created a powerful framework by which new dilation results could be
obtained. On the other hand, the general philosophy of dilation theory found its way
into operator algebras, and has led to remarkable developments.
In this section I will present Stinespring’s dilation theorem, and how Arveson’s
extension theorem and his notion of C*-dilation have made Stinespring’s theorem
into a “dilation machine” that produces and explains dilation results in operator
theory. Then I will briefly discuss how dilation theory is related to the notions of
boundary representations and the C*-envelope, which lie at the heart of the above
mentioned analysis of the relationship between on operator algebra/space and the
C*-algebras that it generates.
I will not attempt to cover all the manifold ways in which dilation theory appears
in the theory of operator algebras, and I’ll just mention a few (of my favorite) recent
examples: [51, 82, 84, 85]. The reader is referred to the survey [48] or the paper [41]
in order to get an idea of the role it plays, in particular in operator algebras related
to dynamical systems and semicrossed products.

7.1 Completely Positive Maps and Stinespring’s Theorem

An operator space is a subspace M ⊆ B(H) of the bounded operators on some

Hilbert space H. We say that M is unital if 1 = IH ∈ M. If M is a subalgebra
of B(H), then it is called an operator algebra (note that operator algebras are not
assumed to be closed under the adjoint). A unital operator space M is said to be an
operator system if it is closed under the adjoint operation. Since every C*-algebra
can be represented faithfully on a Hilbert space, we can consider a subspace of a
C*-algebra as an operator space (and likewise for unital operator spaces, operator
algebras and operator systems). C*-algebras are operator algebras, and unital C*-
algebras are operator systems, of course.
An operator space M ⊆ B(H) inherits from B(H) a norm and a notion of
positivity: an element a ∈ M is said to be positive, if it is positive as an operator on
H, i.e., ah, h ≥ 0 for all h ∈ H. Operator systems are spanned by their positive
elements, indeed, if a ∈ M then its real and imaginary parts are also in M, and
if a is selfadjoint then 12 (a · 1 ± a) ≥ 0 and the difference of these two positive
elements is a.
As a consequence, it makes sense to speak of positive maps. If M and N are
operator systems, a linear map φ : M → N is said to be positive if it takes
positive elements to positive elements. The matrix spaces Mn (M) ⊆ Mn (B(H)) =
B(Hn ) and Mn (N ) are also operator systems, and then φ induces a linear map
φ (n) : Mn (M) → Mn (N )

φ (n) = φ ⊗ idMn : Mn (M) = M ⊗ Mn → Mn (N ) = N ⊗ Mn

584 O. M. Shalit

acting elementwise as

φ (n) : (aij )ni,j =1 → (φ(aij ))ni,j =1 ∈ Mn (N ).

The map φ is said be completely positive (or CP for short) if φ (n) is positive for
all n. Likewise, φ is said to be completely contractive (or CC for short) if φ (n) is
contractive for all n. A map is UCP if it is a unital CP map, and UCC if it is a unital
CC map.
Completely positive maps were introduced by Stinespring [159], but it was
Arveson who observed how important they are and opened the door to their
becoming an indispensable tool in operator theory and operator algebras [10]. There
are several excellent sources to learn about operator spaces/systems and completely
positive (and bounded) maps; see for example [118] and [122].
Completely positive maps arise also in mathematical physics in a natural way
[89]. The evolution of an open quantum system is described by a semigroup of
completely positive maps [49], and noisy channels in quantum information theory
are modelled as trace preserving completely positive maps [107]. In quantum
probability [111], semigroups of unit preserving completely positive maps play the
role of Markov semigroups.
The simplest examples of completely positive maps are ∗-homomorphisms
between C*-algebras. Next, a map of the form B(K) 4 T → V ∗ T V ∈ B(H), where
V is some fixed operator in B(H, K), is readily seen to be completely positive.
Since compositions of CP maps are evidently CP, we see that whenever A is a
C*-algebra, π : A → B(K) is a ∗-homomorphism, and V ∈ B(H, K), then the
map a → V ∗ π(a)V is a CP map. The following fundamental theorem shows that
essentially all CP maps on C*-algebras are of this form.
Theorem 7.1 (Stinespring’s Theorem [159]) Let A be a unital C*-algebra and
let φ : A → B(H) be a CP map. Then there exists a Hilbert space K, an operator
V ∈ B(H, K), and a ∗-representation π : A → B(K) such that

φ(a) = V ∗ π(a)V , for all a ∈ A.

The tuple (π, K, V ) can be chosen such that K = [π(A)H]—the smallest closed
subspace containing π(a)h for all a ∈ A and h ∈ H—and in this case the triple
(π, K, V ) is unique up to unitary equivalence.
Proof On the algebraic tensor product A ⊗ H, we define a semi-inner product by
setting a ⊗ g, b ⊗ h = g, φ(a ∗ b)hH and extending sesquilinearly (the complete
positivity guarantees that this is a positive semidefinite form). Quotienting out the
kernel and then completing gives rise to the Hilbert space K. The image of all the
elementary tensors b ⊗ h ∈ A ⊗ H in K form a total set, and we continue to
denote these images as b ⊗ h. One needs to check that for every a ∈ A, the map
π(a) : b ⊗ h → ab ⊗ h extends to a well defined, bounded linear operator on K.
Once this is done, it is easy to verify that the map a → π(a) is a ∗-homomorphism.
To recover φ, we define V : H → K by V (h) = 1 ⊗ h, and then all that remains
to do is to compute V ∗ (a ⊗h), g = a ⊗h, V (g) = a ⊗h, 1⊗g = h, φ(a ∗ )g,
Dilation Theory: A Guided Tour 585

so V ∗ (a ⊗ h) = φ(a)h, and thus

V ∗ π(a)V h = V ∗ (a ⊗ h) = φ(a)h,

as required. If [π(A)H]  K, then we replace K with [π(A)H], and obtain a

minimal representation. The uniqueness is a standard matter, and is left to the reader.

If K = [π(A)H], then (π, K, V ) (or just π sometimes) is called the minimal
Stinespring representation of φ.
Remark 7.2 If φ is unital, then 1 = φ(1) = V ∗ π(1)V = V ∗ V , so V is an isometry.
In this case it is convenient to identify H with V H ⊆ K, and the Stinespring
representation manifests itself as a dilation

φ(a) = PH π(a) H .

In this situation, the minimal Stinespring representation is referred to as the minimal

Stinespring dilation of φ.

7.2 Arveson’s Extension Theorem and C*-Dilations

The utility of completely positive maps comes from the following extension theorem
of Arveson. For a proof, see Arveson’s paper or Paulsen’s book [118, Chapter 7].
Theorem 7.3 (Arveson’s Extension Theorem [10]) Let M be an operator system
in a C*-algebra A, and let φ : M → B(H) be a CP map. Then there exists a CP
map φ̂ : A → B(H) such that φ̂ = φ and which extends φ, i.e., φ̂(a) = φ(a)
for all a ∈ M.
We will now see how the combination of Stinespring’s dilation theorem and
Arveson’s extension theorem serve as kind of all purpose “dilation machine”, that
produces dilation theorems in varied settings.
Let 1 ∈ M ⊆ A be a unital operator space. A linear map φ : M → B(H) is
said to have a C*-dilation to A if there exists a ∗-representation π : A → B(K),
K ⊇ H, such that

φ(a) = PH π(a) H , for all a ∈ M.

Arveson showed that every UCP map is UCC, and that, conversely, every UCC
map as above extends to a UCP map  φ : ML := M + M∗ → B(H) given by
 ∗ ∗
φ (a + b ) = φ(a) + φ(b) . Combining this basic fact with Theorems 7.1 and 7.3
we obtain the following versatile dilation theorem.
Theorem 7.4 Every UCC or UCP map has a C*-dilation.
586 O. M. Shalit

Arveson’s dilation theorem1 (Theorem 5.1) follows from the above theorem,
once one carefully works through the delicate issues of joint spectrum, Shilov
boundary and functional calculus (see [11]). We shall illustrate the use of the dilation
machine by proving Arveson’s dilation theorem for the simple but representative
d
example of the polydisc D .
Theorem 7.5 (Arveson’s Dilation Theorem for the Polydisc) A d-tuple of com-
muting contractions T = (T1 , . . . , Td ) has a unitary dilation if and only if the
d
polydisc D is a complete spectral set for T .
Proof To relate the statement of the theorem to the language of Sect. 5, we note that
d
the Shilov boundary of D is just the torus (∂D)d = Td , and therefore a unitary
dilation is nothing but a normal dilation with joint spectrum contained in Td . Recall
d
that D being a complete spectral set is equivalent to that

p(T ) ≤ p∞ := sup p(z) (7.1)

d
z∈D

d
for every matrix valued polynomial p (since D is convex it suffices to consider
matrix valued polynomials, and there is no need to worry about matrix valued
rational functions).
If U = (U1 , . . . , Ud ) is a tuple of commuting unitaries and p is a matrix valued
polynomial, then, using the spectral theorem, it is not hard to see that p(U ) =
supz∈σ (U ) p(z) ≤ p∞ . Now, if U is a dilation of T then p(T ) ≤ p(U ),
and so the inequality (7.1) holds. That was the easy direction.
d
Conversely, suppose that D is a complete spectral set for a commuting tuple
T ∈ B(H)d , that is, suppose that (7.1) holds for every matrix valued polynomial p.
Let M = C[z1 , . . . , zd ] be the space of polynomials in d variables, considered as a
unital subspace of the C*-algebra C(Td ), equipped with the usual supremum norm.
It is useful to note that

sup p(z) = sup p(z),

d z∈Td
z∈D

d
by applying the maximum modulus principle in several variables. The fact that D
is a complete spectral set for T implies that the unital map φ : M → B(H) given
by φ(p) = p(T ), is completely contractive. By Theorem 7.4, φ has a C*-dilation
π : C(Td ) → B(K), such that

p(T ) = φ(p) = PH π(p) H , p ∈ M.

1 The reader should be aware that Theorem 7.4 is sometimes referred to as Arveson’s dilation

theorem, whereas I used this name already for the more specific Theorem 5.1.
Dilation Theory: A Guided Tour 587

Now, π is a ∗-representation, and the coordinate functions z1 , . . . , zd ∈ C(Td ) are

all unitary, so Ui = π(zi ), i = 1, . . . , d, are commuting unitaries. Since π(p) =
p(π(z1 ), . . . , π(zd )), we find that

p(T ) = PH p(U ) H

for all p ∈ M, that is, U is a unitary dilation of T , as required. 

Following Arveson, the above method has been used extensively for proving the
existence of dilations in certain situations. The burden is then shifted from the task
of constructing a dilation, to that of showing that certain naturally defined maps are
UCP or UCC. In other words, by proving an inequality one obtains an existence
proof—a good bargain from which analysts have profited for a century. Of course,
“good bargain” does not mean that we cheat, we still need to prove something. Let
me give an example of how this works.
Example 7.6 We now prove that for every contraction T ∈ B(H), the map  :
C[z] → B(H) given by (p) = p(T ) is UCC. Combining this with Arveson’s
dilation theorem for the disc (the case d = 1 in Theorem 7.5), we obtain a genuinely
new proof of Sz.-Nagy’s unitary dilation theorem (Theorem 2.1). The reader should
be able to adapt the details of this proof to get a proof that every row contraction
has a row-isometric dilation, and that every pure commuting row contraction can be
modelled by the d-shift (for hints, see the introduction of [125] or [8], respectively).
Let S be the unilateral shift on 2 , and let en denote the nth standard basis vector
in 2 . If T is a contraction and r ∈ (0, 1), we let DrT = (I − r 2 T T ∗ )1/2, and define
Kr (T ) : H → 2 ⊗ H by
 
Kr (T )h = en ⊗ r n DrT T n∗ h ,
n

for all h ∈ H. We compute:


Kr (T )∗ Kr (T )h = r 2n T n DrT
2
T n∗ h

= r 2n T n (I − r 2 T T ∗ )T n∗ h
 
= r 2n T n T n∗ h − r 2(n+1)T n+1 T (n+1)∗ h = h

so that Kr (T ) is an isometry. On C ∗ (S) we define a UCP map

φr (a) = Kr (T )∗ (a ⊗ I )Kr (T ).
588 O. M. Shalit

We compute that

φr (S) = Kr (T )∗ (S ⊗ I )Kr (T )h
 
= Kr (T )∗ en+1 ⊗ r n DrT T n∗ h

= r 2n+1 T n+1 DrT
2
T n∗ h = rT h.

Likewise, φr (S n ) = r n T n for all n ∈ N.

Now we define a UCP map  := limrD1 φr . Then (S n ) = T n for all n.
We see that the map p(S) → p(T ) is UCC. But S is unitarily equivalent to the
multiplication operator Mz on H 2 (see Sect. 4.3). Thus, for every matrix valued
polynomial p

p(T ) ≤ p(S) = p(Mz ) = sup p(z),

|z|=1

so D is a complete spectral set for T . By Theorem 7.5, T has a unitary dilation.

7.3 Boundary Representations and the C*-Envelope

The ideas in this section are best motivated by the following classical example.
Example 7.7 Consider the disc algebra A(D), which is equal to the closure of the
polynomials with respect to the norm p∞ = supz∈D |p(z)|. The disc algebra
is an operator algebra, being a subalgebra of the C*-algebra C(D) of continuous
functions on the disc D. Moreover, C ∗ (A(D)) = C(D), that is, the C*-subalgebra
generated by A(D) ⊆ C(D) is equal to C(D). However, C(D) is not determined
uniquely by being “the C*-algebra generated by the disc algebra”. In fact, by the
maximum modulus principle, A(D) is also isometrically isomorphic to the closed
subalgebra of C(T) generated by all polynomials, and the C*-subalgebra of C(T)
generated by the polynomials is equal to C(T).
Now, C(T) is the quotient of C(D) by the ideal of all continuous functions
vanishing on the circle T. If π : C(D) → C(T) denotes the quotient map, then
π(f ) = f T , and we note, using the maximum principle again, that π is isometric
on A(D). It turns out that T is the minimal subset E ⊆ D such that the map
f → f E is isometric on A(D).
The above phenomenon arises in all uniform algebras, that is, in all unital
subalgebras A of C(X) that separate the points of X, where X is some compact
Hausdorff space. For every such algebra there exists a set ∂A ⊆ X—called the
Shilov boundary of A—which is the unique minimal closed subset E ⊆ X such
that f → f E is isometric (see [68] for the theory of uniform algebras).
Dilation Theory: A Guided Tour 589

In the spirit of noncommutative analysis, Arveson sought to generalize the Shilov

boundary to the case where the commutative C*-algebra C(X) is replaced by a
noncommutative C*-algebra B = C ∗ (A) generated by a unital operator algebra A.
An ideal I F B is said to be a boundary ideal for A in B, if the restriction of the
quotient map π : B → B/I to A is completely isometric. The Shilov ideal of A in
B is the unique largest boundary ideal for A in B. If J is the Shilov ideal of A in B,
then C*-envelope of A is defined to be the C*-algebra Ce∗ (A) = B/J . (The above
notions were introduced in [10, 11], but it took some time until the terminology
settled down. A good place for the beginner to start learning this stuff is [118].)
Example 7.8 If B = C(X) is a commutative C*-algebra generated by the uniform
algebra A, then the Shilov ideal of A is just the ideal I∂A of functions vanishing on
the Shilov boundary ∂A . In this case Ce∗ (A) = C(∂A ).
The C*-envelope Ce∗ (A) = B/J = π(B) has the following universal property:
if i : A → B is a completely isometric homomorphism such that B = C ∗ (i(A)),
then there exists a unique surjective ∗-homomorphism ρ : B → Ce∗ (A) such that
π(a) = ρ(i(a)) for all a ∈ A. It follows that the C*-envelope depends only on
the structure of A as an operator algebra, not on the concrete realization A ⊆ B
with which we started. Thus, if Ai ⊆ Bi = C ∗ (Ai ) for i = 1, 2 have trivial Shilov
ideals, then every completely isometric homomorphism φ : A1 → A2 extends to
∗-isomorphism ρ : B1 → B2 .
In fact, the algebraic structure is not essential here, and the above notions also
make sense for unital operator spaces. For some purposes, it is most convenient to
work with operator systems, and focus is then shifted to this case (as in the next
subsection). For example, the Shilov ideal of an operator system S ⊆ B = C ∗ (S)
is the largest ideal I F B such that the quotient map π : B → B/I restricts to a
complete isometry on S, etc.
How does one find the Shilov ideal? Let us return to Example 7.7 (recalling also
Example 7.8). If A(D) is given as a subalgebra of C(D), how could one characterize
its Shilov boundary? A little bit of function theory shows that a point z ∈ D is in
the unit circle T if and only if it has a unique representing measure. Recall that
when we have a uniform algebra A ⊆ C(X), a probability measure μ is said to be a
representing measure for x if

f (x) = f dμ
X

for all f ∈ A. For example, the Lebesgue measure on the circle is a representing
measure for the point 0, because
 2π
1
f (0) = f (eit )dt
2π 0

for every f ∈ A(D), as an easy calculation shows. Every point can be represented
by the delta measure δz ; the points on the circle are singled out by being those with
590 O. M. Shalit

a unique representing measure, that is, they can be represented only by the delta
measure. In the general case of a uniform algebra A ⊆ C(X), the points in X that
have a unique representing measure are referred to as the Choquet boundary of
A. It is not hard to show that the Choquet boundary of A(D) is T. In general, the
Choquet boundary of a uniform algebra is dense in the Shilov boundary.
Let return to the noncommutative case, so let A ⊆ B = C ∗ (A) be again a
unital operator algebra generating a C*-algebra. Point evaluations correspond to the
irreducible representations of a commutative C*-algebra, and probability measures
correspond to states, that is, positive maps into C. With this in mind, the reader will
hopefully agree that the following generalization is potentially useful: an irreducible
representation π : B → B(H) is said to be a boundary representation if the only
UCP map  : B → B(H) that extends π A is π itself.
Arveson proved in [10] that if an operator algebra A ⊆ B = C ∗ (A) has
sufficiently many boundary representations, in the sense that

A = sup{π (n) (A) : π : B → B(Hπ ) is a boundary representation}

for all A ∈ Mn (A), then the Shilov ideal exists, and is equal to the intersection
of all boundary ideals. For some important operator algebras, the existence of
sufficiently many boundary representations was obtained (see also [11]), but
the problem of existence of boundary representations in general remained open
almost 45 years.2 Following a sequence of important developments [15, 53, 101],
Davidson and Kennedy proved that every operator system has sufficiently many
boundary representations [42]. Their proof implies that every unital operator space,
and in particular every unital operator algebra, has sufficiently many boundary
representations as well.

7.4 Boundary Representations and Dilations

It is interesting that the solution to the existence problem of boundary representa-

tions was obtained through dilations. Davidson and Kennedy worked in the setting
of operator systems, so let us follow them in this subsection.
If S ⊆ B = C ∗ (S) is an operator system inside the C*-algebra that it generates,
an irreducible representation π : B → B(H) is a boundary representation if the only
UCP map  : B → B(H) that extends π S is π itself. This leads to the following
definition: a UCP map φ : S → B(H) is said to have the unique extension
property if there exists a unique UCP map  : B → B(H) that extends φ, and,
moreover, this  is a ∗-representation. Thus, an irreducible representation π is a

2 The existence of the C*-envelope was obtained much earlier, without making use of boundary

representations; see [118].

Dilation Theory: A Guided Tour 591

boundary representation if and only if the restriction π S has the unique extension
property.
Now, the unique extension property nicely captures the idea that S is in some
sense rigid in B, but it is hard to verify it in practice. The following notion is more
wieldy. A UCP map ψ : S → B(K) is said to be a dilation of a UCP map φ : S →
B(H) if φ(a) = PH ψ(a) H for all a ∈ S. The dilation ψ is said to be a trivial
dilation if H is reducing for φ(S), that is, ψ = φ ⊕ ρ for some UCP map ρ. A UCP
map φ is said to be maximal if it has only trivial dilations.
Penetrating observations of Muhly and Solel [101], and consequently Dritschel
and McCullough [53], can be reformulated as the following theorem. The beauty
is that the notion of maximality is intrinsic to the operator system S, and does not
take the containing C*-algebra B into account (similar reformulations exist for the
categories of unital operator spaces and operator algebras).
Theorem 7.9 A UCP map φ : S → B(H) has the unique extension property if and
only if it is maximal.
Following Dritschel and McCullough’s proof of the existence of the C*-envelope
[53] and Arveson’s consequent work [15], Davidson and Kennedy proved the
following theorem (as above, similar reformulations exist for the categories of unital
operator spaces and operator algebras).
Theorem 7.10 Every UCP map can be dilated to a maximal UCP map, and every
pure UCP map can be dilated to a pure maximal UCP map.
Davidson and Kennedy proved that pureness guarantees that the ∗-representation,
which is the unique UCP extension of the maximal dilation, is in fact irreducible.
Moreover, they showed that pure UCP maps completely norm an operator space.
Thus, by dilating sufficiently many pure UCP maps, and making use of the above
theorems, they concluded that there exist sufficiently many boundary representa-
tions [42].
Example 7.11 Let us see what are the maximal dilations in the case of the disc
algebra A(D) ⊆ C(D) (we switch back from the category of operator systems to
the category of unital operator algebras). A representation π of C(D) is determined
uniquely by a normal operator N with spectrum in D by the relation N = π(z). A
UCC representation φ : A(D) → B(H) is determined uniquely by the image of the
coordinate function z, which is a contraction T = φ(z) ∈ B(H). Conversely, by the
A(D) functional calculus (see Sect. 2.2), every contraction T ∈ B(H) gives rise to a
UCC homomorphism of A(D) into B(H). In this context, a dilation of a UCC map
φ into B(H) is simply a representation ρ : A(D) → B(K) such that

f (T ) = φ(f ) = PH ρ(f ) H = PH f (V ) H ,

for all f ∈ A(D), where V = ρ(z) ∈ B(K). So ρ is a dilation of φ if and only if

V = ρ(z) is a (power) dilation of T = φ(z) in the sense of Sect. 2.
592 O. M. Shalit

With the above notation, it is not hard to see the following two equivalent
statements: (i) a dilation ρ : A(D) → B(K) is maximal if and only if V is a
unitary, and (ii) a representation π : C(D) → B(K) is such that ρ = π A(D) has
the unique extension property if and only if V is a unitary. The fact that every UCC
representation of A(D) has a maximal dilation, is equivalent to the fact that every
contraction has a unitary dilation.
What are the boundary representations of the disc algebra? The irreducible
representations of C(D) are just point evaluations δz for z ∈ D. The boundary
representations are those point evaluations δz whose restriction to A(D) have a
unique extension to a UCP map C(D) → C. But UCP maps into the scalars are
just states, and states of C(D) are given by probability measures. Hence, boundary
representations are point evaluations δz such that z has a unique representing
measure, so that z ∈ T.
The Shilov ideal can be obtained as the intersection of the kernels of the boundary
representations, and so it is the ideal of functions vanishing on T. The C*-envelope
is the quotient of C(D) by this ideal, thus it is C(T), as we noted before.
Note that to find the boundary representations of the disc algebra we did not
need to invoke the machinery of maximal dilations. In the commutative case, the
existence of sufficiently many boundary representations is no mystery: all of them
are obtained as extensions of evaluation at a boundary point. The machinery of max-
imal dilations allows us to find the boundary representations in the noncommutative
case, where there are no function theoretic tools at our disposal.

8 Dilations of Completely Positive Semigroups

The dilation theory of semigroups of completely positive maps can be considered

as a kind of “quantization” of classical isometric dilation theory of contractions on
a Hilbert space. The original motivation comes from mathematical physics [49, 60].
The theory is also very interesting and appealing in itself, having connections and
analogies (and also surprising differences) with classical dilation theory. Studying
dilations of CP-semigroups has led to the discovery of results and some structures
that are interesting in themselves.
In this section, I will only briefly review some results in dilation theory of CP-
semigroups from the last two decades, of the kind that I am interested in. There are
formidable subtleties and technicalities that I will either ignore, or only gently hint
at. For a comprehensive and up-to-date account, including many references (also
to other kinds of dilations), see [149]. All of the facts that we state without proof
or reference below have either a proof or a reference in [149]. The reader is also
referred to the monographs [14] and [111] for different takes on quantum dynamics
and quantum probability.
Dilation Theory: A Guided Tour 593

8.1 CP-Semigroups, E-Semigroups and Dilations

Let S be a commutative monoid. By a CP-semigroup we mean family  =

{s }s∈S of contractive CP-maps on a unital C*-algebra B such that
1. 0 = idB ,
2. s ◦ t = s+t for all s, t ∈ S.
If S carries some topology, then we usually require s → s to be continuous
in some sense. A CP-semigroup  is said to be a Markov semigroup (or a
CP0 -semigroup) if every s is a unital map. A CP-semigroup  is called an E-
semigroup if every element s is a ∗-endomorphism. Finally, an E0 -semigroup is
a Markov semigroup which is also an E-semigroup.
One-parameter semigroups of ∗-automorphisms model the time evolution in a
closed (or a reversible) quantum mechanical system, and one-parameter Markov
semigroups model the time evolution in an open (or irreversible) quantum mechan-
ical system [49].
The prototypical example of a CP-semigroup is given by

s (b) = Ts bTs∗ , b ∈ B (8.1)

where T = {Ts }s∈S is a semigroup of contractions in B. We call such a semigroup

elementary. Of course, not all CP-semigroups are elementary.
For us, a dilation of a CP-semigroup is a triplet (A, α, p), where A is a C*-
algebra, p ∈ A is a projection such that B = pAp, and α = {αs }s∈S is an E-
semigroup on A, such that

s (b) = pαs (b)p

for all b ∈ B and s ∈ S. A strong dilation is a dilation in which the stronger

condition

s (pap) = pαs (a)p

holds for all a ∈ A and s ∈ S. It is a fact, not hard to show, that if  is a Markov
semigroup then every dilation is strong. Examples show that this is not true for
general CP-semigroups. Sometimes, to lighten the terminology a bit, we just say
that α is a (strong) dilation.
Remark 8.1 It is worth pausing to emphasize that the dilation defined above is
entirely different from Stinespring’s dilation: the Stinespring dilations of s and s
cannot be composed. The reader should also be aware that there are other notions
of dilations, for example in which the “small” algebra B is embedded as a unital
subalgebra of the “large” algebra A (see, e.g., [67] or [167] and the references
therein), or where additional restrictions are imposed (see, e.g. [91], and the papers
that cite it).
594 O. M. Shalit

The most important is the one-parameter case, where S = N or R+ . The

following result was proved first by Bhat in the case B = B(H) [24] (slightly later
SeLegue gave a different proof in the case B = B(H) [140]), then it was proved by
Bhat and Skeide for general unital C*-algebras [28], and then by Muhly and Solel
in the case of unital semigroups on von Neumann algebras [102] (slightly later this
case was also proved by Arveson [14, Chapter 8]).
Theorem 8.2 Every CP-semigroup  = {s }s∈S on B over S = N or S = R+
has a strong dilation (A, α, p).
Moreover, if  is unital then α can be chosen unital; if B is a von Neumann
algebra and  is normal, then A can be taken to be a von Neumann algebra and α
normal; and if further S = R+ and  is point weak-∗ continuous, in the sense that
t → ρ(t (b)) is continuous for all ρ ∈ B∗ , then α can also be chosen to be point
weak-∗ continuous.
Proof Let us illustrate the proof for the case where B = B(H), S = N, θ is a normal
contractive CP map on B(H),and n = θ n for all n. Then it is well known that θ
must have the form θ (b) = i Ti bTi∗ for a row contraction T = (Ti ) (see, e.g.,
[89, Theorem 1]). By Theorem 6.2, T has a row isometric coextension V = (Vi ) on
a Hilbert space K. Letting A = B(K), p = PH and

α(a) = Vi aVi∗
i

we obtain a strong dilation (as the reader will easily verify). 

The above proof suggests that there might be strong connections between
operator dilation theory and the dilation theory of CP-semigroups. This is true,
but there are some subtleties. Consider an elementary CP-semigroup (8.1) acting
on B = B(H), where T is a semigroup of contractions on H. If V = {Vs }s∈S
is an isometric dilation of T , then α(a) = Vs aVs∗ is a dilation of . If V is a
coextension, then α is a strong dilation. Thus, we know that we can find (strong)
dilations for elementary semigroups when the semigroup is such that isometric
dilations (coextensions) exist for every contractive semigroup; for example, when
S = N, N2 , R+ , R2+ . Moreover, we expect that we won’t always be able to dilate
CP-semigroups over monoids for which isometric dilations don’t always exist (for
example S = N3 ). This analogy based intuition is almost correct, and usually
helpful.
Remark 8.3 As in classical dilation theory, there is also a notion of minimal dila-
tion. However, it turns out that there are several reasonable notions of minimality.
In the setting of normal continuous semigroups on von Neumann algebras, the most
natural notions of minimality turn out to be equivalent in the one-parameter case,
but in the multi-parameter case they are not equivalent. See [14, Chapter 8] and
[149, Section 21] for more on this subject.
Dilation Theory: A Guided Tour 595

Theorem 8.2 has the following interesting interpretation. A one-parameter CP-

semigroup models the time evolution in an open quantum dynamical system, and
a one-parameter automorphism semigroup models time evolution in a closed one.
In many cases, E-semigroups can be extended to automorphism semigroups, and
so Theorem 8.2 can be interpreted as saying that every open quantum dynamical
system can be embedded in a closed (reversible) one (this interpretation was the
theoretical motivation for the first dilation theorems, see [49, 60]).

8.2 Main Approaches and Results

There are two general approaches by which strong dilations of CP-semigroups can
be constructed.
The Muhly–Solel Approach One approach, due to Muhly and Solel [102], seeks
to represent a CP-semigroup in a form similar to (8.1), and then to import ideas
from classical dilation theory. To give little more detail, if  = {s }s∈S is a
CP-semigroup on a von Neumann algebra B ⊆ B(H), then one tries to find a
product system E E = {Es }s∈S of B -correspondences over S (see Sect. 6.5) and
a completely contractive covariant representation T = {Ts }s∈S such that

s (b) = Ts (idEs E b)Ts∗ (8.2)

for every s ∈ S and b ∈ B. Here, Ts : Es EH → H is given by Ts (x Eh) = Ts (x)h.

This form is reminiscent of (8.1), and it is begging to try to dilate  by constructing
an isometric dilation V for T , and then defining

s (idEs E a)V
αs (a) = V s∗

for a ∈ A := V0 (B ) . This is a direct generalization of the apporach to

dilating elementary semigroups discussed in the paragraph following the proof of
Theorem 8.2, and in fact it also generalizes the proof we gave for that theorem. This
approach was used successfully to construct and analyze dilations in the discrete
and continuous one-parameter cases (Muhly and Solel, [102, 104]), in the discrete
two-parameter case (Solel, [157]; this case was solved earlier by Bhat for B = B(H)
[25]), and in the “strongly commuting” two-parameter case (Shalit, [141, 142, 144]).
However, it turns out that finding a product system and representation giving
back  as in (8.2) is not always possible, and one needs a new notion to proceed.
A subproduct system is a family E  = {Es }s∈S of C*-correspondences such that,
roughly, Es+t ⊆ Es E Et (up to certain identifications that iterate associatively).
Following earlier works [14, 102], it was shown that for every CP-semigroup
there is a subproduct system and a representation, called the Arveson–Stinespring
subproduct system and representation, satisfying (8.2) (Shalit and Solel, [150]).
Subproduct systems have appeared implicitly in the theory in several places, and in
596 O. M. Shalit

[150] they were finally formally introduced (at the same time, subproduct systems
of Hilbert spaces were introduced in Bhat and Mukherjee’s paper [27]).
The approach to dilations introduced in [150] consists of two parts: first, embed
the Arveson–Stinespring subproduct system associated with a CP-semigroup into
a product system, and then dilate the representation to an isometric dilation. This
approach was used to find necessary and sufficient conditions for the existence of
dilations. In particular, it was used to prove that a Markov semigroup has a (certain
kind of) minimal dilation if and only if the Arveson–Stinespring subproduct system
can be embedded into a product system. Moreover, the framework was used to show
that there exist CP-semigroups over N3 that have no minimal strong dilations, as
was suggested from experience with classical dilation theory. Vernik later used these
methods to prove an analogue of Opela’s theorem (see 3.3) for completely positive
maps commuting according to a graph [166].
The reader is referred to [150] for more details. The main drawback of that
approach is that it works only for CP-semigroups of normal maps on von Neumann
algebras.
The Bhat–Skeide Approach The second main approach to dilations of CP-
semigroups is due to Bhat and Skeide [28]. It has several advantages, one of
which is that it works for semigroups on unital C*-algebras (rather than von
Neumann algebras). The Bhat–Skeide approach is based on a fundamental and
useful representation theorem for CP maps called Paschke’s GNS representation
[112], which I will now describe.
For every CP map φ : A → B between two unital C*-algebras, there exists a
C*-correspondence E from A to B (see Sect. 6.5) and a vector ξ ∈ E such that
φ(a) = ξ, aξ  for all a ∈ A. The existence of such a representation follows from
a construction: one defines E to be the completion of the algebraic tensor product
A ⊗ B with respect to the B-valued inner product

a ⊗ b, a ⊗ b  = b∗ φ(a ∗ a )b ,

equipped with the natural left and right actions. Letting ξ = 1A ⊗1B , it is immediate
that φ(a) = ξ, aξ  for all a ∈ A. Moreover, ξ is cyclic, in the sense that it
generates E as a C*-correspondence. The pair (E, ξ ) is referred to as the GNS
representation of φ. The GNS representation is unique in the sense that whenever
F is a C*-correspondence from A to B and η ∈ F is a cyclic vector such that
φ(a) = η, aη for all a ∈ A, then there is an isomorphism of C*-correspondences
from E onto F that maps ξ to η.
In the Bhat–Skeide approach to dilations, the idea is to find a product system
F E = {Fs }s∈S of B-correspondences and a unit, i.e., a family ξ E = {ξs ∈ Fs }s∈S
satisfying ξs+t = ξs E ξt , such that  is recovered as

s (b) = ξs , bξs  (8.3)

Dilation Theory: A Guided Tour 597

for all s ∈ S and all b ∈ B. If  is a Markov semigroup, the dilation is obtained via
a direct limit construction. For non Markov semigroups, a dilation can be obtained
via a unitalization procedure. In [28], dilations were constructed this way in the
continuous and discrete one-parameter cases. This strategy bypasses product system
representations, but, interestingly, it can also be used to prove the existence of an
isometric dilation for any completely contractive covariant representation of a one-
parameter product system [155].
Again, it turns out that constructing such a product system is not always possible.
However, if one lets (Fs , ξs ) be the GNS representation of the CP map s , then it
is not hard to see that F  = {Fs }s∈S is a subproduct system (called the GNS
subproduct system) and ξ E = {ξs } is a unit.
The above observation was used by Shalit and Skeide to study the existence
of dilations of CP-semigroups in a very general setting [149]. If one can embed
the GNS subproduct system into a product system, then one has (8.3), and can
invoke the Bhat–Skeide approach to obtain a dilation. The paper [149] develops
this framework to give a unified treatment of dilation theory of CP-semigroups over
a large class of monoids S, including noncommutative ones. One of the main results
in [149], is the following, which generalizes a result obtained earlier in [150].
Theorem 8.4 A Markov semigroup over an Ore monoid admits a full strict dilation
if and only if its GNS subproduct system embeds into a product system.
This theorem essentially enables to recover almost all the other known dilation
theorems and counter examples. It is used in [149] to show that every Markov
semigroup over S = N2 on a von Neumann algebra has a unital dilation, and also
that for certain multi-parameter semigroups (the so called quantized convolution
semigroups) there is always a dilation. The theorem is also used in the converse
direction, to construct a large class of examples that have no dilation whatsoever.
In the setting of normal semigroups on von Neumann algebras, the Bhat–Skeide
and Muhly–Solel approaches to dilations are connected to each other by a functor
called commutant; see [149, Appendix A(iv)] for details.
New Phenomena We noted above some similarities between the theory of isomet-
ric dilations of contractions, and the dilation theory of CP-semigroups. In particular,
in both theories, there always exists a dilation when the semigroup is parameterized
by S = N, N2 , R+ , and the results support the possibility that this is true for S = R2+
as well. Moreover, in both settings, there exist semigroups over S = N3 for which
there is no dilation.
But there are also some surprises. By  Corollary 3.8, if a tuple of commuting
contractions T = (T1 , . . . , Td ) satisfies di=1 Ti 2 ≤ 1, then T has a regular
unitary dilation. Therefore,  one might think that if a commuting d-tuple of CP maps
1 , . . . , d are such that di=1 i  is sufficiently small, then this tuple has a
dilation. This is false, at least in a certain sense (that is, if one requires a strong and
minimal dilation); see [150, Section 5.3].
Moreover, by Corollary 3.8, a tuple of commuting isometries always has a unitary
dilation, and it follows that every tuple of commuting coisometries has an isometric
598 O. M. Shalit

(in fact, unitary) coextension. In (8.1) coisometries correspond to unital maps.

Hence, one might expect that commuting unital CP maps always have dilations.
Again, this is false [149, Section 18] (see also [148]). The reason for the failure
of these expectations is that not every subproduct system over Nd (when d ≥ 3)
can be embedded into a product system (in fact, there are subproduct systems that
cannot even be embedded in a superproduct system). However, the product system
of an elementary CP-semigroup is a trivial product system, so the obstruction to
embeddability does not arise in that case.

8.3 Applications of Dilations of CP-Semigroups

Besides the interesting interpretation of dilations given at the end of Sect. 8.1,
dilations have some deep applications in noncommutative dynamics. In this section,
we will use the term CP-semigroup to mean a one-parameter semigroup  =
{s }s∈S (where S = N or S = R+ ) of normal CP maps acting on a von Neumann
algebra B. In case of continuous time (i.e., S = R+ ), we will also assume that 
is point weak-∗ continuous, in the sense that t → ρ(t (b)) is continuous for every
ρ ∈ B∗ . We will use the same convention for E- or E0 -semigroups.
The Noncommutative Poisson Boundary Let  be a normal UCP map on a von
Neumann algebra B. Then one can show that the fixed point set {b ∈ B : (b) = b}
is an operator system, and moreover that it is the image of a completely positive
projection E : B → B. Hence, the Choi-Effros product x ◦ y = E(xy) turns the
fixed point set into a von Neumann algebra H ∞ (B, ), called the noncommutative
Poisson boundary of . The projection E and the concrete structure on H ∞ (B, )
are hard to get a grip with.
Arveson observed that if (A, α, p) is the minimal dilation of , and if Aα is the
fixed point algebra of α, then the compression a → pap is a unital, completely
positive order isomorphism between Aα and H ∞ (B, ). Hence Aα is a concrete
realization of the noncommutative Poisson boundary. See the survey [81] for details
(it has been observed that this result holds true also for dilations of abelian CP-
semigroups [129]).
Continuity of CP-Semigroups Recall that one-parameter CP-semigroups on a von
Neumann algebra B ⊆ B(H) are assumed to be point weak-∗ continuous. Since
CP-semigroups are bounded, this condition is equivalent to point weak-operator
continuity, i.e., that t → t (b)g, h is continuous for all b ∈ B and all g, h ∈ H.
Another natural kind of continuity to consider is point strong-operator continuity,
which means that t → t (b)h is continuous in norm for all b ∈ B and h ∈ H.
For brevity, below we shall say a semigroup is weakly continuous if it is point
weak-operator continuous, and strongly continuous if it is point strong-operator
continuous.
Dilation Theory: A Guided Tour 599

Strong continuity is in some ways easier to work with and hence it is desirable,
but it is natural to use the weak-∗ topology, because it is independent of the
representation of the von Neumann algebra. Happily, it turns out that weak (and
hence point weak-∗) continuity implies strong continuity.
One possible approach to prove the above statement is via dilation theory. First,
one notices that the implication is easy for E-semigroups. Indeed, if α is an E-
semigroup on A ⊆ B(K), then

αt (a)k − αs (a)k2 = αt (a ∗ a)k, k + αs (a ∗ a)k, k − 2 Reαt (a)k, αs (a)k.

Assuming that s tends to t, the expression on the right hand side tends to zero if
α is weakly continuous. Now, if  is a weakly continuous CP-semigroup, then its
dilation α given by Theorem 8.2 is also weakly continuous. By the above argument,
α is strongly continuous, and this continuity is obviously inherited by t (·) =
pαt (·)p. Hence, by dilation theory, weak continuity implies strong continuity.
The above argument is a half cheat, because, for a long time, the known proofs
that started from assuming weak continuity and ended with a weakly continuous
dilation, actually assumed implicitly, somewhere along the way, that CP-semigroups
are strongly continuous [14, 28, 102]. This gap was pointed out and fixed by
Markiewicz and Shalit [98], who proved directly that a weakly continuous CP-
semigroup is strongly continuous. Later, Skeide proved that the minimal dilation of a
weakly continuous semigroup of CP maps is strongly continuous, independently of
[98], thereby recovering the result “weakly continuous ⇒ strongly continuous” with
a proof that truly goes through the construction of a dilation; see [156, Appendix
A.2].
Existence of E0 -Semigroups As we have seen above, dilations can be used to
study CP-semigroups. We will now see an example, where dilations are used in
the theory of E0 -semigroups.
The fundamental classification theory E0 -semigroups on B = B(H) was
developed Arveson, Powers, and others about two decades ago; see the monograph
[14] for the theory, and in particular for the results stated below (the classification
theory of E0 -semigroups on arbitrary C* or von Neumann algebras is due to Skeide;
see [156]). For such E0 -semigroups there exists a crude grouping into type I, type
II, and type III semigroups. However, it is not at all obvious that there exist any E0 -
semigroups of every type. Given a semigroup of isometries on a Hilbert space H ,
one may use second quantization to construct E0 -semigroups on the symmetric and
anti-symmetric Fock spaces over H , called the CCR and CAR flows, respectively.
CAR and CCR flows are classified in term of their index. These E0 -semigroups are
of type I, and, conversely, every type I E0 -semigroup is cocycle conjugate to a CCR
flow, which is, in turn, conjugate to a CAR flow.
It is much more difficult to construct an E0 -semigroup that is not type I. How does
one construct a non-trivial E0 -semigroup? Theorem 8.2 provides a possible way:
construct a Markov semigroup, and then take its minimal dilation. This procedure
600 O. M. Shalit

has been applied successfully to provide examples of non type I E0 -semigroups,

even with prescribed index; see [14, 81].

Part 3: Recent Results in the Dilation Theory

of Noncommuting Operators

9 Matrix Convexity and Dilations

In recent years, dilation theory has found a new role in operator theory, through
the framework of matrix convexity. In this section I will quickly introduce matrix
convex sets in general, special examples, minimal and maximal matrix convex over
a given convex set, and the connection to dilation theory. Then I will survey the
connection to the UCP interpolation problem, some dilation results, and finally an
application to spectrahedral inclusion problems.

9.1 Matrix Convex Sets

Fix d ∈ N. The “noncommutative universe” Md is the set of all d-tuples of n × n

matrices, of all sizes n, that is,
∞
I
M =d
Mnd .
n=1

Sometimes it is useful to restrict attention to the subset Mdsa , which consists

of all tuples of selfadjoint matrices. We will refer to a subset S ⊆ Md as a
noncommutative (nc) set, and we will denote by Sn or S(n) the nth level of S,
 mean Sn = S(n) := S ∩ Mn . Let us endow tuples with the row norm
by which we d

A :=  i Ai A∗i ; this induces a metric on B(H)d for every d, and in particular
on Mnd for every n. We will say that a nc set S is closed if Sn is closed in Mnd for all
n. We will say that S is bounded if there exists some C > 0 such that X ≤ C for
all X ∈ S.
For a tuple X = (X1 , . . . , Xn ) ∈ Mnd and a linear map φ : Mn → Mk , we write
φ(X) = (φ(X1 ), . . . , φ(Xd )) ∈ Mkd . In particular, if A and B are n × k matrices,
then we write A∗ XB = (A∗ X1 B, . . . , A∗ Xd B). Another operation that we can
preform on tuples is the direct sum, that is, if X ∈ Mm d and Y ∈ M d , then we let
n
X ⊕ Y = (X1 ⊕ Y1 , . . . , Xd ⊕ Yd ) ∈ Mm+n .d
Dilation Theory: A Guided Tour 601

A matrix convex set S is a nc set S = ∞ S which is invariant under direct

n=1 n
sums and the application of UCP maps:

X ∈ Sm , Y ∈ Sn 7⇒ X ⊕ Y ∈ Sm+n

and

X ∈ Sn and φ ∈ UCP(Mn , Mk ) 7⇒ φ(X) ∈ Sk .

It is not hard to check that a nc set S ⊆ Md is matrix convex if and only if it

is closed under matrix convex combinations in the following  sense: whenever
X(j ) ∈ Snj and Vj ∈ Mnj ,n for j = 1, . . . , k are such that kj =1 Vj∗ Vj = In ,

then Vj∗ X(j ) Vj ∈ Sn .
Remark 9.1 The above notion of matrix convexity is due to Effros and Winkler [57].
Other variants appeared before and after. A very general take on matrix convexity
that I will not discuss here has recently been initiated by Davidson and Kennedy
[43]. I will follow a more pedestrian point of view, in the spirit of [47] (note: the
arxiv version [47] is a corrected version of the published version [46]. The latter
contains several incorrect statements in Section 6, which result from a missing
hypothesis; the problem and its solution are explained in [47]). We refer to the
first four chapters of [47] for explanations and/or references to the some of the
facts that will be mentioned below without proof. The papers [61, 113, 115] make a
connection between the geometry of matrix convex sets, in particular various kinds
of extreme points, and dilation theory. For a comprehensive and up-to-date account
of matrix convex sets the reader can consult [90].
Example 9.2 Let A ∈ B(H)d . The matrix range of A is the nc set W(A) =
∞ W (A) given by
n=1 n

Wn (A) = {φ(A) : φ : B(H) → Mn is UCP} .

The matrix range is a closed and bounded matrix convex set. Conversely, every
closed and bounded matrix convex set is the matrix range of some operator tuple.
If d = 1, then the first level W1 (A) of the matrix range of an operator A coincides
with the closure of the numerical range W (A)

W (A) = {Ah, h : h = 1} .

We note, however, that for d ≥ 2, the first level W1 (A) does not, in general, coincide
with the closure of what is sometimes referred to as the joint numerical range of a
tuple [96].
Matrix ranges of single operators were introduced by Arveson [11], and have
been picked up again rather recently. The matrix range of an operator tuple A
is a complete invariant of the operator system generated by A, and—as we shall
602 O. M. Shalit

see below—it is useful when considering interpolation problems for UCP maps.
Moreover, in the case of a fully compressed tuple A of compact operators or normal
operators, the matrix range determines A up to unitary equivalence [115]. The
importance of matrix ranges has led to the investigation of random matrix ranges,
see [70].
Example 9.3 Let A ∈ B(H)d . The free spectrahedron determined by A is the nc
set DA = ∞n=1 DA (n) given by
⎧ ⎫
⎨ 
d ⎬
DA (n) = X ∈ Mnd : Re Xj ⊗ Aj ≤ I .
⎩ ⎭
j =1

A free spectrahedron is always a closed matrix convex set, that contains the origin
in its interior. Conversely, every closed matrix convex set with 0 in its interior is a
free spectrahedron. In some contexts it is more natural to work with just selfadjoint
matrices. For A ∈ B(H)dsa one defines
⎧ ⎫
⎨ 
d ⎬
DA
sa
= X ∈ Mdsa : Xj ⊗ Aj ≤ I .
⎩ ⎭
j =1

The first level DA (1) is called a spectrahedron. Most authors use the word
spectrahedron to describe only sets of the form DA (1) where A is a tuple of
matrices; and likewise for the term free spectrahedron. This distinction is important
for applications of the theory, since spectrahedra determined by tuples of matrices
form a class of reasonably tractable convex sets that arise in applications, and not
every convex set with 0 in its interior can be represented as DA (1) for a tuple A
acting on a finite dimensional space.
For a matrix convex set S ⊆ Md we define its polar dual to be
   
S ◦ = X ∈ Md : Re Xj ⊗ Aj ≤ I for all A ∈ S .

If S ⊆ Mdsa , then it is more convenient to use the following variant

  
S • = X ∈ Mdsa : Xj ⊗ Aj ≤ I for all A ∈ S .

By the Effros–Winkler Hahn–Banach type separation theorem [57], S ◦◦ = S

whenever S is a matrix convex set containing 0 (if 0 ∈ / S, then S ◦◦ is equal to
the matrix convex hull of S and 0). It is not hard to see that DA = W(A)◦ , and that
when 0 ∈ W(A), we also have W(A) = DA ◦.

Example 9.4 Another natural and important way in which matrix convex sets arise,
is as positivity cones in operator systems. In [65] it was observed that a finite
dimensional abstract operator system M (see [118, Chapter 13]) generated by d
Dilation Theory: A Guided Tour 603

linearly independent elements A1 , . . . , Ad ∈ M, corresponds to a matrix convex

set C ⊆ Mdsa where every Cn is the cone in (Mnd )sa consisting of the matrix tuples
X = (X1 , . . . , Xd ) such that Xj ⊗Aj is positive in Mn (M). Such matrix convex
sets can be described by a slight modification of the notion of free spectrahedron:
  
C = X ∈ Mdsa : Xj ⊗ Aj ≥ 0 .

9.2 The UCP Interpolation Problem

Suppose we are given two d-tuples of operators A = (A1 , . . . , Ad ) ∈ B(H)d and

B = (B1 , . . . , Bd ) ∈ B(K)d . A very natural question to ask is whether there exists
a completely positive map φ : B(H) → B(K) such that φ(Ai ) = Bi . This is the CP
interpolation problem. In the realm of operator algebras it is sometimes more useful
to ask about the existence of a UCP map that interpolates between the operators,
and in quantum information theory it makes sense to ask whether there exists a
completely positive trace preserving (CPTP) interpolating map. A model result is
the following.
Theorem 9.5 Let A ∈ B(H )d and B ∈ B(K)d be d-tuples of operators.
1. There exists a UCP map φ : B(H) → B(K) such that φ(Ai ) = Bi for all
i = 1, . . . , d if and only if W(B) ⊆ W(A).
2. There exists a unital completely isometric map φ : B(H) → B(K) such that
φ(Ai ) = Bi for all i = 1, . . . , d if and only if W(B) = W(A).
This result was obtained by Davidson, Dor-On, Shalit and Solel in [47, Theorem
5.1]. An earlier result was obtained by Helton, Klep and McCullough in the case
where H and K are finite dimensional, and the condition W(B) ⊆ W(A) is
replaced by the dual condition DA ⊆ DB , under the blanket assumption that DA
is bounded [74] (see a somewhat different approach in [7]). Later, Zalar showed that
the condition DA ⊆ DB is equivalent to the existence of an interpolating UCP map
without the assumption that DA is bounded, and also in the case of operators on an
infinite dimensional space [171]. Variants of the above theorem were of interest to
mathematical physicists for some time, see the references in the above papers.
From Theorem 9.5 one can deduce also necessary and sufficient conditions
for the existence of contractive CP (CCP) or completely contractive (CC) maps
sending one family of operators to another, as well as approximate versions (see [47,
Section 5]). The theorem also leads to more effective conditions under additional
assumptions, for example when dealing with normal tuples (recall, that A =
(A1 , . . . , Ad ) is said to be normal if Ai is normal and Ai Aj = Aj Ai for all i, j ).
Corollary 9.6 Let A ∈ B(H )d and B ∈ B(K)d be two normal d-tuples of
operators. Then there exists a UCP map φ : B(H) → B(K) such that

φ(Ai ) = Bi , for all i = 1, . . . , d,

604 O. M. Shalit

if and only if

σ (B) ⊆ conv σ (A).

This result was first obtained by Li and Poon [97], in the special case where A
and B each consist of commuting selfadjoint matrices. It was later recovered in [47],
in the above generality, as a consequence of Theorem 9.6 together with the fact that
for a normal tuple N, the matrix range W(N) is the minimal matrix convex set that
contains the joint spectrum σ (N) in its first level (see [47, Corollary 4.4]). The next
section is dedicated to explaining what are the minimal and maximal matrix convex
sets over a convex set, and how these notions are related to dilation theory.

9.3 Minimal and Maximal Matrix Convex Sets

Every level Sn of a matrix convex set S is a convex subset of Mnd . In particular,

the first level S1 is a convex subset of Cd . Conversely, given a convex set K ⊆ Cd
(or K ⊆ Rd ), we may ask whether there exists a matrix convex set S ⊆ Md (or
S ⊆ Mdsa ) such that S1 = K. The next question to ask is, to what extent does the
first level S1 = K determine the matrix convex set S?
In order to approach the above questions, and also as part of a general effort
to understand inclusions between matrix convex sets (motivated by results as
Theorem 9.5), notions of minimal and maximal matrix convex sets have been
introduced by various authors [46, 65, 75]. These are very closely related (via
Example 9.4) to the notion of minimal and maximal operator systems that was
introduced earlier [120].
For brevity, we shall work in the selfadjoint setting. Let K ⊆ Rd be a convex set.
By the Hahn–Banach theorem, K can be expressed as the intersection of a family
of half spaces:

K = {x ∈ Rd : fi (x) ≤ ci for all i ∈ I}

where {fi }i∈I is 

a family of linear functionals and {ci }i∈I is a family of scalars.
Writing fi (x) = j aji xj , we define

Wnmax (K) = {X ∈ (Mnd )sa : aji Xj ≤ ci In for all i ∈ I}
j

and W max (K) = ∞ n=1 Wn (K). In other words, W

max max (K) is the nc set determined

by the linear inequalities that determine K. It is clear that W max (K) is matrix
convex, and a moment’s thought reveals that it contains every matrix convex set
that has K as its first level.
Dilation Theory: A Guided Tour 605

That settles the question, of whether or not there exists a matrix convex set with
first level equal to K. It follows, that there has to exist a minimal matrix convex set
that has K as its first level—simply intersect over all such matrix convex sets. There
is a useful description of this minimal matrix convex set. We define
 
W min (K) = X ∈ Mdsa : ∃ normal T with σ (T ) ⊆ K s.t. X ≺ T . (9.1)

Recall that X ≺ T means that T is a dilation of X.

W min (K) is clearly invariant under direct sums. To see that it is invariant also
under the application of UCP maps, one may use Stinespring’s theorem as follows.
If X ≺ T , T is normal, and φ is UCP, then the map T → X → φ(X) is UCP.
By Stinespring’s theorem there is a ∗-representation π such that φ(X) ≺ π(T ),
and π(T ) is a normal tuple with σ (π(T )) ⊆ σ (T ) (alternatively, one may use the
dilation guaranteed by Theorem 8.2).
We see that the set defined in (9.1) is matrix convex. On the other hand, any
matrix convex set containing K in the first level must contain all unitary conjugates
of tuples formed from direct sums of points in K, as well as their compressions,
therefore the minimal matrix convex set over K contains all X that have a normal
dilation T acting on a finite dimensional space such that σ (T ) ⊆ K. But for X ∈
Mdsa , the existence of a normal dilation X ≺ T with σ (T ) ⊆ K implies the existence
of a normal dilation acting on a finite dimensional space (see [46, Theorem 7.1]),
thus the nc set W min (K) that we defined above is indeed the minimal matrix convex
set over K.
Example 9.7 Let D be the closed unit disc in C. Let us compute W min (D) and
W max (D). We can consider D as a subset of R2 , and pass to the selfadjoint setting
(and back) by identifying T = Re T + i Im T ∈ M1 with the selfadjoint tuple
(Re T , Im R) ∈ M2sa . The minimal matrix convex set is just

W min (D) = {X ∈ M1 : X ≤ 1},

because by Theorem 2.1, every contraction has a unitary dilation. Since the set of
real linear inequalities determining the disc is
 
D = {z ∈ C : Re eiθ z ≤ 1 for all θ ∈ R},

it follows that
 
W max (D) = {X : Re eiθ X ≤ I for all θ ∈ R},

which equals the set of all matrices with numerical range contained in the disc.
606 O. M. Shalit

Given a convex set K ⊆ Cd , Passer, Shalit and Solel introduced a constant θ (K)
that quantifies the difference between the minimal and maximal matrix convex sets
over K [116, Section 3]. For two convex sets K, L, we define

θ (K, L) = inf{C : W max (K) ⊆ CW min (L)},

and θ (K) = θ (K, K). Note that CW min (L) = W min (CL).
Remark 9.8 In the theory of operator spaces, there are the notions of minimal and
maximal operator spaces over a normed space V , and there is a constant α(V )
that quantifies the difference between the minimal and maximal operator space
structures [117] (see also [118, Chapter 14] and [122, Chapter 3]). These notions
are analogous to the above notions of minimal and maximal matrix convex sets, but
one should not confuse them.
By the characterization of the minimal and maximal matrix convex sets, the
inclusion W max (K) ⊆ W min (L) is a very general kind of dilation result: it means
that every d-tuple X satisfying the linear inequalities defining K, has a normal
dilation X ≺ N such that σ (N) ⊆ L. Let us now review a few results obtained
regarding this dilation problem.
Theorem 9.9 (Theorem 6.9, [116]) For p ∈ [1, ∞], let Bp,d denote the unit ball
in Rd with respect to the p norm, and let Bp,d (C) denote the unit ball in Cd with
respect to the p norm. Then

θ (Bp,d ) = d 1−|1/2−1/p|

and

θ (Bp,d (C)) = 2d 1−|1/2−1/p|.

√ (sharp) inclusions W (K) ⊆ W (L). Interestingly,

See [116] for many other max min

the fact that θ (B1,d ) = d has implications in quantum information theory—it

allows to find a quantitative measure of how much noise one needs to add to a d-
tuple of quantum effects to guarantee that they become jointly measurable; see [30].
The case d = 2 in the above theorem was first obtained in [76, Section 14] and
[47, Section 7] using other methods. It also follows from the following result.
Theorem 9.10 (Theorem 5.8, [65]) Let K ⊆ Rd be a symmetric convex set, i.e.
K = −K. Then

W max (K) ⊆ dW min (K).

The above result was originally proved by Fritz, Netzer and Thom [65] for cones
with a symmetric base; to pass between the language of convex bodies and that of
cones, one may use the gadget developed in [116, Section 7]. In [116, Theorem 4.5]
Dilation Theory: A Guided Tour 607

it was observed that Theorem 9.10 is also a consequence of the methods of [47,
Section 7] together with some classical results in convex geometry.
Already in [57, Lemma 3.1] it was observed that there is only one matrix convex
S with S1 = [a, b] ⊂ R, namely the matrix interval given by Sn = {X ∈ (Mn )sa :
aIn ≤ X ≤ bIn }. Said differently, W max ([a, b]) = W min ([a, b]). It is natural to ask
whether there exists any other convex body (i.e., a compact convex set) K with the
property that W min (K) = W max (K).
Theorem 9.11 Let K ⊆ Rd be a convex body. Then W max (K) = W min (K) if and
only if K is a simplex, that is, if K is the convex hull of a set of affinely independent
points. In fact, W2max (K) = W2min (K) already implies that K is a simplex.
The result that the equality W max (K) = W min (K) is equivalent to K being
a simplex was first obtained by Fritz, Netzer and Thom [65, Corollary 5.3] for
polyhedral cones. In [116, Theorem 4.1] it was proved for general convex bodies,
and it was also shown that one does not need to check equality Wnmax (K) =
Wnmin (K) for all n in order to deduce that K is a simplex—it suffices to check
this for some n ≥ 2d−1 . For simplex pointed convex bodies, it was shown that
W2max (K) = W2min (K) already implies that K is a simplex [116, Theorem 8.8].
Huber and Netzer later obtained this for all polyhedral cones [80], and finally
Aubrun, Lami, Palazuelos and Plavala proved the result for all cones [16, Corollary
2].
Remark 9.12 The minimal matrix convex set over a “commutative” convex set
K ⊆ Rd can be considered as the matrix convex hull of K. There are some variations
on this theme. Helton, Klep and McCullough studied the matrix convex hull of free
semialgebraic sets [75]. Instead of W min (K) and W max (K), which are the minimal
and maximal matrix convex sets with prescribed first level, one can also discuss the
minimal and maximal matrix convex sets with a prescribed kth level (see [90], or
[169, 170] for the version of this notion in the framework of operator systems).
In the recent paper [114], Passer and Paulsen define, given a matrix convex set
S, the minimal and maximal matrix convex sets W min-k (S) and W max-k (S) such
that Wkmin-k (S) = Wkmax-k (S) = Sk , and they utilize quantitative measures of
discrepancy between W min-k (S), W max-k (S) and S to glean information on the
operator system corresponding to S; unfortunately, these results are beyond the
scope of this survey. The paper [114] also ties together some of the earlier work
in this direction, so it is a good place to start if one is interested in this problem.

9.4 Further Dilation Results

There are many other interesting dilation results in [46, 50, 65, 76, 113, 116]. In this
section I will review a few more.
608 O. M. Shalit

Problem 9.13 Fix d ∈ N. What is the smallest constant Cd such that for every
d-tuple of contractions A, there exists a d-tuple of commuting normal operators B,
such that A ≺ B holds with Bi  ≤ Cd for all i?
√
First, we note that the sharp dilation constant θ (B∞,d ) = d obtained in
Theorem 9.10 implies the following result, which is a solution to Problem 9.13
in the selfadjoint setting.
Theorem 9.14 (Theorem 6.7, [116]) For every d-tuple A = (A1 , . . . , Ad ) of
selfadjoint contractions, there √ exists a d-tuple of commuting selfadjoints N =
(N
√ 1 , . . . , Nd ) with Ni  ≤ d for i = 1, . . . , d, such that A ≺ N. Moreover,
d is the optimal constant for selfadjoints.
It is interesting to note that one of the proofs of the above theorem goes through
a concrete construction of the dilation. The nonselfadjoint version of Problem 9.13
is more difficult, and it does not correspond to an inclusion problem of some W max
in some W min . The best general result in the nonselfadjoint case is the following
theorem obtained by Passer.
Theorem 9.15 (Theorem 4.4, [113]) For every d-tuple A = (A1 , . . . , Ad ) of con-
√ exists a d-tuple of commuting normal operators N = (N1 , . . . , Nd )
tractions, there
with Ni  ≤ 2d for i = 1, . . . , d, such that A ≺ N.
Thus
√ √
d ≤ Cd ≤ 2d.

In the next section we will improve the lower bound in the case d = 2.
Helton, Klep, McCullough and Schweighofer obtained a remarkable result,
which is analogous to Theorem 9.14, but in which the dilation constant is indepen-
dent of the number of operators d [76]. Following Ben-Tal and Nemirovski [21],
Helton et al. defined a constant ϑ(n) as follows:
 
1 
n 
n
= min ai xi2 dμ(x) : |ai | = 1
ϑ(n) ∂Bn i=1 i=1

where μ is the uniform probability measure on the unit sphere ∂Bn ⊂ Rn .

Theorem 9.16 (Theorem 1.1, [76]) Fix n and a real n-dimensional Hilbert space
H. Let F ⊆ B(H)sa be a family of selfadjoint contractions. Then there exists a real
Hilbert space K, an isometry V : H → K, and a commuting family C in the unit
ball of B(K)sa such that for every contraction A ∈ F , there exists N ∈ C such that

1
A = V ∗ NV .
ϑ(n)
Dilation Theory: A Guided Tour 609

Moreover, ϑ(n) is the smallest constant such that the above holds for all finite sets
of contractive selfadjoints F ⊆ B(H)sa .
Note the difference from Theorem 9.14: the dimension of matrices is fixed at
n × n, but the number of matrices being simultaneously dilated is not fixed. In other
words, the constant ϑ(n) depends only on the size of the matrices being dilated (in
fact, it is shown that n can be replaced with the maximal rank of the matrices being
dilated). It is also shown that
√
πn
ϑ(n) ∼ .
2
In the next subsection I will explain the motivation for obtaining this result.

9.5 An Application: Matricial Relaxation of Spectrahedral

Inclusion Problems

Any dilation result, such as Theorem 9.14 or 9.16, leads to a von Neumann type
inequality. For example, if A is a d-tuple of selfadjoint contractions, then by
Theorem 9.14, for every matrix valued polynomial p of degree at most one, we
have the following inequality:
 √ √ 
p(A) ≤ sup p(z) : z ∈ [− d, d]d .

This result is by no means trivial, but it is the kind of application of dilation theory
that we have already seen above several times.
We will now see a deep application of Helton, Klep, McCullough and
Schweighofer’s theorem (Theorem 9.16) that is of a different nature from the
applications that we have seen hitherto, and is the main motivation for the
extraordinary paper [76]. The application builds on earlier work of Ben-Tal and
Nemirovski [21] in control theory and optimization, related to what is sometimes
called the matrix cube problem. I will give a brief account; the reader who seeks a
deeper understanding should start with the introductions of [21] and [76].
In the analysis of a linear controlled dynamical system (as in [21]), one is led to
the problem of deciding whether the cube [−1, 1]d is contained in the spectrahedron
DAsa (1), for a given a d-tuple of selfadjoint n × n matrices A , . . . , A ; this is
1 d
called the matrix cube problem. More generally, given another d-tuple of selfadjoint
matrices B1 , . . . , Bd , it is of practical interest to solve the spectrahedral inclusion
problem, that is, to be able to decide whether

DBsa (1) ⊆ DA
sa
(1).
610 O. M. Shalit

Note that the matrix cube problem is a special case of the spectrahedral inclusion
problem, since [−1, 1]d = DC sa (1) for the d-tuple of 2d × 2d diagonal matri-

ces C1 = diag(1, −1, 0, . . . , 0), C2 = diag(0, 0, 1, −1, 0, . . . , 0), . . . , Cd =

(0, . . . , 0, 1, −1). The free spectrahedron DC sa determined by C is nothing but the

nc set consisting of all d-tuples of selfadjoint contractions.

The problem of deciding whether one spectrahedron is contained in another is a
hard problem. In fact, deciding whether or not [−1, 1]d ⊆ DA sa (1) has been shown

to be NP hard (note that the naive solution of checking whether all the vertices of the
cube are in DA sa (1) requires one to test the positive semidefiniteness of 2d matrices).

However, Ben-Tal and Nemirovski introduced a tractable relaxation of this problem

[21]. In [74], Helton, Klep and McCullough showed that the relaxation from [21] is
equivalent to the free relaxation DC sa ⊆ D sa , and the subsequent work in [76] gives
A
a full understanding of this relaxation, including sharp estimates of the error bound.
Let’s take a step back. Fix two d-tuples of selfadjoint matrices A and B. We
mentioned that the problem of determining whether DBsa (1) ⊆ DA sa (1) is hard. In

[74], it was observed that the free relaxation, that is, the problem DBsa ⊆ DA sa

is tractable. Indeed, as explained after Theorem 9.5, the inclusion DB ⊆ DA is sa sa

equivalent to the UCP interpolation problem, that is, to the existence of a UCP map
sending Bi to Ai for all i = 1, . . . , d [74, Theorem 3.5]. Now, the UCP interpolation
problem can be shown to be equivalent to the solution of a certain semidefinite
program [74, Section 4]. In practice, there are numerical software packages that
can solve such problems efficiently.
So we see that instead of solving the matrix cube problem [−1, 1]d ⊆ DA sa (1),

one can solve the free relaxation DC ⊆ DA . Now, the whole point of the sharp
sa sa

results in [76] is that they give a tight estimate of how well the tractable free
relaxation approximates the hard matrix cube problem. To explain this, we need
the following lemma.
Lemma 9.17 Suppose that A is a d-tuple of selfadjoint n × n matrices. Then,

[−1, 1]d ⊆ DA
sa
(1) ⇒ DC
sa
⊆ ϑ(n)DA
sa
.

Proof Suppose that [−1, 1]d ⊆ DA sa (1). If X ∈ D sa (n), then by Theorem 9.16,
C
X ≺ ϑ(n)N, where N is a normal tuples and σ (N) ⊆ [−1, 1]d ⊆ DA sa (1). So

 
Xj ⊗ Aj ≺ ϑ(n) Nj ⊗ Aj ≤ ϑ(n)I,

where the last inequality follows easily by the spectral theorem and the assumption
[−1, 1]d ⊆ DA sa (1). 
Finally, we can now understand how to give an approximate solution to the
matrix cube problem. Simply, one tests whether DC sa ⊆ ϑ(n)D sa , which is a
A
tractable problem. If the inclusion holds, then it holds at every level and in particular
[−1, 1]d ⊆ ϑ(n)DA sa (1). If not, then, using the lemma, we conclude that [−1, 1]d 

DA (1). Thus, we are able to determine the containment of [−1, 1]d in DA

sa sa (1), up to
Dilation Theory: A Guided Tour 611

a multiplicative error of ϑ(n), which is known to high precision, and independent

of d.

10 Dilation of q-Commuting Unitaries

This section is dedicated to presenting the results Gerhold and Shalit from [69], on
dilations of q-commuting unitaries.
Let θ ∈ R and write q = eiθ . If u and v are two unitaries that satisfy
vu = quv, then we say that u and v are q-commuting. We denote by Aθ the
universal C*-algebra generated by a pair of q-commuting unitaries, and we call Aθ
θ
a rational/irrational rotation C*-algebra if 2π is rational/irrational respectively.
We shall write uθ , vθ for the generators of Aθ . The rotation C*-algebras have been
of widespread interest ever since they were introduced by Rieffel [132]. A good
reference for this subject is Boca’s book [31].
In an attempt to make some progress in our understanding of the general constant
Cd from Problem 9.13, Malte Gerhold and I studied a certain refinement of that
problem which is of independent interest. Instead of dilating arbitrary tuples of
contractions, we considered the task of dilating pairs of unitaries u, v that satisfy
the q-commutation relation vu = quv, and studied the dependence of the dilation
constant on the parameter q. In the context of Problem 9.13, it is worth noting that,
by a result of Buske and Peters [34] (see also [86]), every pair of q-commuting
contractions has a q-commuting unitary power dilation; therefore, this work has
implications to all pairs of q-commuting operators. Surprisingly, our dilation results
also have implications for the continuity of the norm and the spectrum of the almost
Mathieu operator from mathematical physics (this application will be discussed in
the final section).
For every θ ∈ R we define the optimal dilation constant

cθ := inf{c > 1 | (uθ , vθ ) ≺ c(U, V ) where U, V are commuting unitaries}.

We note that the infimum is actually a minimum, and that it is equal to the infimum
of the constants c that satisfy: for every q-commuting pair of unitaries U, V there
exists a commuting normal dilation M, N such that M, N ≤ c (see [69,
Proposition 2.3]). Thus, cθ is a lower bound for the constant C2 from Problem 9.13.

10.1 Continuity of the Dilation Scale

Theorem 10.1 (Theorem 3.2, [69]) Let θ, θ ∈ R, set q = eiθ , q = eiθ , and put
c = e 4 |θ−θ | . Then for any pair of q-commuting unitaries U, V there exists a pair
1

of q -commuting unitaries U , V such that cU , cV dilates U, V .

612 O. M. Shalit

Proof The proof makes use of the Weyl operators on symmetric Fock space (see
[111, Section 20]). For a Hilbert space H let H ⊗s k be the k-fold symmetric tensor
product of H , and let
∞
K
(H ) := H ⊗s k
k=0

be the symmetric Fock space over H . The exponential vectors

∞
1
e(x) := √ x ⊗k , x ∈ H,
k=0
k!

form a linearly independent and total subset of (H ). For z ∈ H we define the Weyl
unitary W (z) ∈ B((H )) which is determined by
 
z2
W (z)e(x) = e(z + x) exp − − x, z
2

for all exponential vectors e(x).

Consider Hilbert spaces H ⊂ K with p the projection onto H , and the symmetric
Fock spaces (H ) ⊂ (K) with P the projection onto (H ). We write p⊥ for the
projection onto the orthogonal complement H ⊥ . Note that for exponential vectors
we have P e(x) = e(px). For every y, z ∈ K, the Weyl unitaries W (y), W (z)
satisfy:
1. W (z) and W (y) commute up to the phase factor e2i Imy,z .
p⊥ z2
2. P W (z) (H )
= e− 2 W (pz), so it is a scalar multiple of a unitary on (H ).
3. P W (z) (H )
and P W (y) (H ) commute up to a phase factor e2i Impy,z .
In [69] it is shown that, assuming without loss that θ > θ , things can be arranged
so that there are two linearly independent vectors z, y so that pz and py are linearly
independent, and such that
1. p⊥ y = −ip⊥ z,
2. θ = 2 Imy, z,
3. θ = 2 Impy, z.
Then we get q -commutation of W (z) and W (y), q-commutation of the operators
P W (z) (H ) and P W (y) (H ) , and

θ − θ = −2 Imp⊥ y, z = 2p⊥ z2 = 2p⊥ y2 ,

Dilation Theory: A Guided Tour 613

so
; ; ; ; ⊥ 2
; ; ; ; − p 2y |θ −θ |
;P W (z) (H ) ;
= ;P W (y) (H ) ; = e = e− 4 .

Now if we put

|θ −θ | |θ −θ |
U =e 4 P W (z) (H )
, V =e 4 P W (y) (H )
,

and

U = W (z) , V = W (y)

then we get the statement for this particular q-commuting pair U, V . Since the Weyl
unitaries give rise to a universal representation of Aθ , the general result follows (see
[69, Proposition 2.3]). 
From the above result we obtained continuity of the dilation scale.
Corollary 10.2 (Corollary 3.4, [69]) The optimal dilation scale cθ depends Lips-
chitz continuously on θ . More precisely, for all θ, θ ∈ R we have

|cθ − cθ | ≤ 0.39 θ − θ .

10.2 The Optimal Dilation Scale

The main result of [69] is the following theorem.

Theorem 10.3 (Theorems 6.3 and 6.4, [69]) Let θ, θ ∈ R, q = eiθ , q = eiθ ,
and put γ = θ − θ . The smallest constant cθ,θ such that every pair of q-commuting
unitaries can be dilated to cθ,θ times a pair of q -commuting unitaries is given by

4
cθ,θ = .
uγ + u∗γ + vγ + vγ∗ 

In particular, for every θ ∈ R,

4
cθ = .
uθ + u∗θ + vθ + vθ∗ 

Proof Since it is a nice construction that we have not yet seen, let us show just that
4
the value of cθ,θ is no bigger than uγ +u∗ +v ∗ ; for the optimality of the dilation
γ +vγ 
γ
constant we refer the reader to [69] (the formula for cθ = cθ,0 follows, since it is
not hard to see that cθ = c−θ ).
614 O. M. Shalit

Represent C ∗ (U, V ) concretely on a Hilbert space H. Let uγ , vγ be the universal

generators of Aγ and put hγ := uγ + u∗γ + vγ + vγ∗ . We claim that there exists a
h 
state ϕ on Aγ such that |ϕ(uγ )| = |ϕ(vγ )| = 4γ (for the existence of such a state,
see [69]). Assuming the existence of such a state, we define

π(uγ ) π(vγ )
U =U⊗ , V =V ⊗ .
ϕ(uγ ) ϕ(vγ )

on K = H⊗L, where π : Aγ → B(L) is the GNS representation of ϕ. These are q -

commuting scalar multiples of unitaries, and they have norm h4γ  . By construction,
there exists a unit vector x ∈ L such that ϕ(a) = π(a)x, x for all a ∈ Aγ .
Consider the isometry W : H → H ⊗ L defined by

Wh = h ⊗ x , h ∈ H.

Then
1
W ∗U W = π(uγ )x, xU = U
ϕ(uγ )

and
1
W ∗V W = π(vγ )x, xV = V ,
ϕ(vγ )

and the proof of the existence of a dilation is complete. 

The operator hθ = uθ + u∗θ + vθ + vθ∗
is called the almost Mathieu operator, and
it has been intensively studied by mathematical physicists, before and especially
after Hofstadter’s influential paper [77] (we will return to it in the next section).
However, the precise behaviour of the norm hθ  as a function of θ is still not
completely understood. We believe that the most detailed analysis is contained in
the paper [32].
In [69, Section 7] we obtained numerical values for cθ = 4/hθ  for various θ .
We calculated by hand c 4 π ≈ 1.5279, allowing us to push the lower bound C2 ≥
5
1.41... to C2 ≥ 1.52. We also made some numerical computations, which lead to an
improved estimate C2 ≥ maxθ cθ ≥ 1.5437. The latter value is an √ approximation
of the constant cθs attained at the silver mean θs = 2π = 2π( 2 − 1) (where
√ γs
γs = 2 + 1 is the silver ratio) which we conjecture to be the angle where the
maximum is attained. However, we do not expect that the maximal value of cθ will
give a tight lower approximation for C2 . Determining the value of C2 remains an
open problem.
Dilation Theory: A Guided Tour 615

10.3 An Application: Continuity of the Spectrum of Almost

Mathieu Operators

The almost Mathieu operator hθ = uθ + u∗θ + vθ + vθ∗ , which appears in the formula
cθ = h4θ  , arises as the Hamiltonian in a certain mathematical model describing
an electron in a lattice under the influence of a magnetic field; see Hofstadter
[77]. This operator has been keeping mathematicians and physicists busy for more
than a generation. Hofstadter’s paper included a picture that depicts the spectrum
(computed numerically) of hθ for various values of θ , famously known as the
Hofstadter butterfly (please go ahead and google it). From observing the Hofstadter
butterfly, one is led to making several conjectures.
First and foremost, it appears that the spectrum of hθ varies continuously with θ ;
since θ is a physical parameter of the system studied, and the spectrum is supposed
to describe possible energy levels, any other possibility is unreasonable. There are
other natural conjectures to make, suggested just by looking at the picture. The most
famous one is perhaps what Barry Simon dubbed as the Ten Martini Problem, which
asks whether the spectrum is a Cantor set for irrational angles. This problem was
settled (in greater generality) by Avila and Jitomirskaya (see [17] for the conclusive
work as well as for references to earlier work).
The continuity of the spectrum σ (hθ ) is a delicate problem that attracted a lot of
attention. For example, in [37] Choi, Elliott, and Yui showed that the spectrum σ (hθ )
of hθ depends Hölder continuously (in the Hausdorff metric) on θ , with Hölder
exponent 1/3. This was soon improved by Avron, Mouche, and Simon to Hölder
continuity with exponent 1/2 [18]. The 1/2-Hölder continuity of the spectrum also
follows from a result of Haagerup and Rørdam, who showed that there exist 1/2-
Hölder norm continuous paths θ → uθ ∈ B(H), θ → vθ ∈ B(H) [71, Corollary
5.5].
As an application of our dilation techniques, we are able to recover the best
possible continuity result regarding the spectrum of the operator hθ . This result
is not new, but our proof is new and simple, and I believe that it is a beautiful and
exciting application of dilation theory with which to close this survey. The following
theorem also implies that the rotation C*-algebras form a continuous field of C*-
algebras, a result due to Elliott [59]. Our dilation methods can also be used to recover
the result of Bellisard [20], that the norm of hθ is a Lipschitz continuous function
of θ .
Theorem 10.4 Let p be a selfadjoint ∗-polynomial in two noncommuting variables.
Then the spectrum σ (p(uθ , vθ )) of p(uθ , vθ ) is 12 -Hölder continuous in θ with
respect to the Hausdorff distance for compact subsets of R.
Proof Let us present the idea of the proof for the most important case

p(uθ , vθ ) = hθ = uθ + u∗θ + vθ + vθ∗ ,

616 O. M. Shalit

without going into the details of Hölder continuity. The idea is that, due to
Theorem 10.1, when θ ≈ θ we have the dilation (uθ , vθ ) ≺ (cuθ , cvθ ) with
1
c = e 4 |θ−θ | ≈ 1. Thus,
 
uθ x
cuθ =
y z

and so x and y must be small, to be precise,


x, y ≤ c2 − 1 ≈ 0.

A similar estimate holds for the off diagonal block of cvθ which dilates vθ . By a
basic lemma in operator theory, for any selfadjoint operators a and b, the Hausdorff
distance between their spectra is bounded as follows:

d(σ (a), σ (b)) ≤ a − b.

We have hθ = uθ + u∗θ + vθ + vθ∗ , and so

   
hθ ∗ hθ 0
chθ = ≈ ,
∗ ∗ 0 ∗

because the off diagonal blocks have small norm, and therefore
 
hθ 0
σ (hθ ) ⊆ σ ≈ σ (chθ ) ≈ σ (hθ ).
0 ∗

In the same way one shows that σ (hθ ) is approximately contained in σ (hθ ), and
therefore the Hausdorff distance between the spectra is small. 
It is interesting to note that the above proof generalizes very easily to higher
dimensional noncommutative tori. Determining the precise dilation scales for higher
dimensional noncommutative tori remains an open problem.

Acknowledgments This survey paper grew out of the talk that I gave at the International
Workshop on Operator Theory and its Applications (IWOTA) that took place in the Instituto
Superior Técnico, Lisbon, Portugal, in July 2019. I am grateful to the organizers of IWOTA 2019
for inviting me to speak in this incredibly successful workshop, and especially to Amélia Bastos,
for inviting me to contribute to these proceedings. I used a preliminary version of this survey as
lecture notes for a mini-course that I gave in the workshop Noncommutative Geometry and its
Applications, which took place in January 2020, in NISER, Bhubaneswar, India. I am grateful to
the organizers Bata Krishna Das, Sutanu Roy and Jaydeb Sarkar, for the wonderful hospitality and
the opportunity to speak and organize my thoughts on dilation theory. I also owe thanks to Michael
Skeide and to Fanciszek Szafraniec, for helpful feedback on preliminary versions. Finally, I wish
to thank an anonymous referee for several useful comments and corrections.
This project was partially supported by ISF Grant no. 195/16.
Dilation Theory: A Guided Tour 617

References

1. J. Agler, Rational dilation on an annulus. Ann. Math. 121, 537–563 (1985)

2. J. Agler, J.E. McCarthy, Pick Interpolation and Hilbert Function Spaces. Graduate Studies in
Mathematics, vol. 44 (American Mathematical Society, Providence, 2002)
3. J. Agler, J.E. McCarthy, Distinguished varieties. Acta Math. 194, 133–153 (2005)
4. J. Agler, N.J. Young, Operators having the symmetrized bidisc as a spectral set. Proc. Edinb.
Math. Soc. 43, 195–210 (2000)
5. J. Agler, J. Harland, B.J. Raphael, Classical Function Theory, Operator Dilation Theory
and Machine Computation on Multiply-Connected Domains. Memoirs of the American
Mathematical Society (American Mathematical Society, Providence, 2008)
6. M.A. Akcoglu, L. Sucheston, Dilations of positive contractions on Lp spaces. Can. Math.
Bull. 20, 285–292 (1977)
7. C.G. Ambrozie, A. Gheondea, An interpolation problem for completely positive maps on
matrix algebras: solvability and parametrization. Linear Multilinear Algebra 63, 826–851
(2015)
8. C. Ambrozie, V. Muller, Commutative dilation theory, in Operator Theory, ed. by D. Alpay
(Springer, Berlin, 2014)
9. T. Andô, On a pair of commutative contractions. Acta Sci. Math. 24, 88–90 (1963)
10. W.B. Arveson, Subalgebras of C ∗ -algebras. Acta Math. 123, 141–224 (1969)
11. W.B. Arveson, Subalgebras of C∗ -algebras II. Acta Math. 128, 271–308 (1972)
12. W.B. Arveson, Subalgebras of C ∗ -algebras III: multivariable operator theory. Acta Math. 181,
159–228 (1998)
13. W.B. Arveson, Dilation theory yesterday and today, in A Glimpse of Hilbert Space Operators:
Paul R. Halmos in Memoriam, ed. by S. Axler, P. Rosenthal, D. Sarason (Birkhäuser, Basel,
2010)
14. W.B. Arveson, Non-commutative Dynamics and E-semigroups. Springer Monographs in
Mathematics (Springer, Berlin, 2003)
15. W.B. Arveson, The noncommutative Choquet boundary. J. Am. Math. Soc. 21, 1065–1084
(2008)
16. G. Aubrun, L. Lami, C. Palazuelos, M. Plavala, Entangleability of cones (2020).
arXiv:1911.09663
17. A. Avila, S. Jitomirskaya, The ten martini problem. Ann. Math. 170, 303–342 (2009)
18. J. Avron, P.H.M.v. Mouche, B. Simon, On the measure of the spectrum for the almost Mathieu
operator. Commun. Math. Phys. 132, 103–118 (1990)
19. C. Badea, B. Beckermann, Spectral sets, in Handbook of Linear Algebra, ed. by L. Hogben
(Chapman and Hall/CRC, Boca Raton, 2014)
20. J. Bellisard, Lipshitz continuity of gap boundaries for Hofstadter-like spectra. Commun.
Math. Phys. 160, 599–613 (1994)
21. A. Ben-Tal, A. Nemirovski, On tractable approximations of uncertain linear matrix inequali-
ties affected by interval uncertainty. SIAM J. Optim. 12, 811–833 (2002)
22. H. Bercovici, D. Timotin, The numerical range of a contraction with finite defect numbers. J.
Math. Anal. Appl. 417, 42–56 (2014)
23. H. Bercovici, C. Foias, L. Kerchy, B. Sz.-Nagy, Harmonic Analysis of Operators on Hilbert
Space. Universitext (Springer, Berlin, 2010)
24. B.V.R. Bhat, An index theory for quantum dynamical semigroups. Trans. Am. Math. Soc.
348, 561–583 (1996)
25. B.V.R. Bhat, A generalized intertwining lifting theorem, in Operator Algebras and Appli-
cations, II, Waterloo, ON, 1994-1995, Fields Institute Communications, vol. 20 (American
Mathemtical Society, Providence, 1998), pp. 1–10
26. B.V.R. Bhat, T. Bhattacharyya, Dilations, completely positive maps and geometry (forthcom-
ing book)
618 O. M. Shalit

27. B.V.R. Bhat, M. Mukherjee, Inclusion systems and amalgamated products of product systems.
Infin. Dimens. Anal. Quant. Probab. Relat. Top. 13, 1–26 (2010)
28. B.V.R. Bhat, M. Skeide, Tensor product systems of Hilbert modules and dilations of
completely positive semigroups. Infin. Dimens. Anal. Quant. Probab. Relat. Top. 3, 519–575
(2000)
29. T. Bhattacharyya, S. Pal, S. Shyam Roy, Dilations of - contractions by solving operator
equations. Adv. Math. 230, 577–606 (2012)
30. A. Bluhm, I. Nechita, Joint measurability of quantum effects and the matrix diamond. J. Math.
Phys. 59, 112202 (2018)
31. F.P. Boca, Rotation C*-Algebras and Almost Mathieu Operators (The Theta Foundation,
Bucharest, 2001)
32. F.P. Boca, A. Zaharescu, Norm estimates of almost Mathieu operators. J. Funct. Anal. 220,
76–96 (2005)
33. J.W. Bunce, Models for n-tuples of noncommuting operators. J. Funct. Anal. 57, 21–30 (1984)
34. D.R. Buske, J.R. Peters, Semicrossed products of the disk algebra: contractive representations
and maximal ideals. Pac. J. Math. 185, 97–113 (1998)
35. M.D. Choi, K.R. Davidson, A 3 × 3 dilation counterexample. Bull. Lond. Math. Soc. 45,
511–519 (2013)
36. M.D. Choi, C.K. Li, Constrained unitary dilations and numerical ranges. J. Operator Theory
46, 435–447 (2001)
37. M.-D. Choi, G.A. Elliott, N. Yui, Gauss polynomials and the rotation algebra. Invent. Math.
99, 225–246 (1990)
38. D. Cohen, Dilations of matrices. Thesis (M.Sc.), Ben-Gurion University (2015).
arXiv:1503.07334
39. M.J. Crabb, A.M. Davie, Von Neumann’s inequality for hilbert space operators. Bull. Lond.
Math. Soc. 7, 49–50 (1975)
40. M. Crouzeix, Numerical range and functional calculus in Hilbert space. J. Funct. Anal. 244,
668–990 (2007)
41. K.R. Davidson, E.G. Katsoulis, Dilation theory, commutant lifting and semicrossed products.
Doc. Math. 16, 781–868 (2011)
42. K.R. Davidson, M. Kennedy, The Choquet boundary of an operator system. Duke Math. J.
164, 2989–3004 (2015)
43. K.R. Davidson, M. Kennedy, Noncommutative Choquet theory. arXiv:1905.08436
44. K.R. Davidson, D.R. Pitts, Nevanlinna-Pick interpolation for non-commutative analytic
Toeplitz algebras. Integr. Equ. Operator Theory 31, 321–337 (1998)
45. K.R. Davidson, D.R. Pitts, The algebraic structure of non-commutative analytic Toeplitz
algebras. Math. Ann. 311, 275–303 (1998)
46. K.R. Davidson, A. Dor-On, O.M. Shalit, B. Solel, Dilations, inclusions of matrix convex sets,
and completely positive maps. Int. Math. Res. Not. 2017, 4069–4130 (2017)
47. K.R. Davidson, A. Dor-On, O.M. Shalit, B. Solel, Dilations, inclusions of matrix convex sets,
and completely positive maps (2018). arXiv:1601.07993v3 [math.OA]
48. K.R. Davidson, A.H. Fuller, E.T.A. Kakariadis, Semicrossed products of operator algebras: a
survey. New York J. Math. 24a, 56–86 (2018)
49. E.B. Davies, Quantum Theory of Open Systems (Academic, Cambridge, 1976)
50. A. Dor-On, Techniques in operator algebras: classification, dilation and non-commutative
boundary theory. Thesis (Ph.D.) University of Waterloo (2017)
51. A. Dor-On, G. Salomon, Full Cuntz-Krieger dilations via non-commutative boundaries. J.
Lond. Math. Soc. 98, 416–438 (2018)
52. R. Douglas, On extending commutative ssemigroupsemigroups of isometries. Bull. Lond.
Math. Soc. 1, 157–159 (1969)
53. M. Dritschel, S. McCullough, Boundary representations for families of representations of
operator algebras and spaces. J. Operator Theory 53, 159–167 (2005)
54. M.A. Dritschel, S. McCullough, The failure of rational dilation on a triply connected domain.
J. Am. Math. Soc. 18, 873–918 (2005)
Dilation Theory: A Guided Tour 619

55. S.W. Drury, A generalization of von Neumann’s inequality to the complex ball. Proc. Am.
Math. Soc. 68, 300–304 (1978)
56. S.W. Drury, Remarks on von Neumann’s Inequality. Lecture Notes in Mathematics, vol. 995
(Springer, Berlin, 1983)
57. E.G. Effros, S. Winkler, Matrix convexity: operator analogues of the bipolar and Hahn-
Banach theorems. J. Funct. Anal. 144, 117–152 (1997)
58. E. Egerváry, On the contractive linear transformations of n-dimensional vector space. Acta
Sci. Math. Szeged 15, 178–182 (1954)
59. G.A. Elliott, Gaps in the spectrum of an almost periodic Schrödinger operator. C.R. Math.
Rep. Acad. Sci. Can. 4, 255–299 (1982)
60. D.E. Evans, J.T. Lewis, Dilations of dynamical semi-groups. Commun. Math. Phys. 50, 219–
227 (1976)
61. E. Evert, J.W. Helton, I. Klep, S. McCullough, Extreme points of matrix convex sets, free
spectrahedra, and dilation theory. J. Geom. Anal. 28, 1373–1408 (2018)
62. S. Fackler, J. Glück, A toolkit for constructing dilations on Banach spaces. Proc. Lond. Math.
Soc. 118, 416–440 (2019)
63. C. Foias, A.E. Frazho, The Commutant Lifting Approach to Interpolation Problems
(Birkhäuser, Basel, 1990)
64. A.E. Frazho, Models for noncommuting operators. J. Funct. Anal. 48, 1–11 (1982)
65. T. Fritz, T. Netzer, A. Thom, Spectrahedral containment and operator systems with finite-
dimensional realization. SIAM J. Appl. Algebra Geom. 1, 556–574 (2017)
66. A.H. Fuller, Finitely correlated representations of product systems of C*-correspondences
over Nk . J. Funct. Anal. 260, 574–611 (2011)
67. D.J. Gaebler, Continuous unital dilations of completely positive semigroups. J. Funct. Anal.
269, 998–1027 (2015)
68. T.W. Gamelin, Uniform Algebras, vol. 311 (American Mathematical Society, Providence,
2005)
69. M. Gerhold, O.M. Shalit, Dilations of q-commuting unitaries (2019). arXiv:1902.10362
70. M. Gerhold, O.M. Shalit, On the matrix range of random matrices (2020). arXiv:1911.12102
71. U. Haagerup, M. Rørdam, Perturbations of the rotation C*-algebras and of the Heisenberg
commutation relations. Duke Math. J. 77, 227–256 (1995)
72. P.R. Halmos, Normal dilations and extensions of operators. Summa Brasil. Math. 2, 125–134
(1950)
73. M. Hartz, M. Lupini, Dilation theory in finite dimensions and matrix convexity (2019).
arXiv:1910.03549
74. J.W. Helton, I. Klep, S. McCullough, The matricial relaxation of a linear matrix inequality.
Math. Program. 138, 401–445 (2013)
75. J.W. Helton, I. Klep, S. McCullough, Matrix convex hulls of free semialgebraic sets. Trans.
Am. Math. Soc. 368, 3105–3139 (2016)
76. J.W. Helton, I. Klep, S. McCullough, M. Schweighofer, Dilations, Linear Matrix Inequalities,
the Matrix Cube Problem and Beta Distributions. Memoirs of the American Mathematical
Society (American Mathematical Society, Providence, 2019)
77. D.R. Hofstadter, Energy levels and wave functions of Bloch electrons in rational and irrational
magnetic fields. Phys. Rev. B. 14, 2239–2249 (1976)
78. J.A. Holbrook, Schur norms and the multivariate von Neumann inequality, in Operator
Theory: Advances and Applications, vol. 127 (Birkhäuser, Basel, 2001)
79. Z. Hu, R. Xia, S. Kais, A quantum algorithm for evolving open quantum dynamics on
quantum computing devices (2019). arXiv:1904.00910
80. B. Huber, T. Netzer, A note on non-commutative polytopes and polyhedra (2019).
arXiv:1809.00476
81. M. Izumi, E0 -semigroups: around and beyond Arveson’s work. J. Oper. Theory 68, 335–363
(2012)
82. E.T.A. Kakariadis, O.M. Shalit, Operator algebras of monomial ideals in noncommuting
variables. J. Math. Anal. Appl. 472, 738–813 (2019)
620 O. M. Shalit

83. I. Kaplansky, Modules over operator algebras. Am. J. Math. 75, 839–858 (1953)
84. E. Katsoulis, C. Ramsey, Crossed Products of Operator Algebras. Memoirs of the American
Mathematical Society (American Mathematical Society, Providence, 2019)
85. M. Kennedy, O.M. Shalit, Essential normality, essential norms and hyperrigidity. J. Funct.
Anal. 270, 2812–2815 (2016)
86. D. Keshari, N. Mallick, q-commuting dilation. Proc. Am. Math. Soc. 147, 655–669 (2019)
87. G. Knese, The von Neumann inequality for 3 × 3 matrices. Bull. Lond. Math. Soc. 48, 53–57
(2016)
88. Ł. Kosiński, Three-point Nevanlinna-Pick problem in the polydisc. Proc. Lond. Math. Soc.
111, 887–910 (2015)
89. K. Kraus, States, Effects, and Operations: Fundamental Notions of Quantum Theory. Lecture
Notes in Physics, vol. 190 (Springer, Berlin, 1983)
90. T.-L. Kriel, An introduction to matrix convex sets and free spectrahedra (2016).
arXiv:1611.03103v6
91. B. Kümmerer, Markov dilations and W*-algberas. J. Funct. Anal. 63, 139–177 (1985)
92. M. Laca, From endomorphisms to automorphisms and back: dilations and full corners. J.
Lond. Math. Soc. 61, 893–904 (2000)
93. E.C. Lance, Hilbert C*-modules: A Toolkit for Operator Algebraists, vol. 210 (Cambridge
University Press, Cambridge, 1995)
94. J. Levick, R.T.W. Martin, Matrix N-dilations of quantum channels. Oper. Matrices 12, 977–
995 (2018)
95. E. Levy, O.M. Shalit, Dilation theory in finite dimensions: the possible, the impossible and
the unknown. Rocky Mountain J. Math. 44, 203–221 (2014)
96. C.K. Li, Y.T. Poon, Convexity of the joint numerical range. SIAM J. Matrix Anal. Appl. 21,
668–678 (2000)
97. C.K. Li, Y.T. Poon, Interpolation by completely positive maps. Linear Multilinear Algebra
59, 1159–1170 (2011)
98. D. Markiewicz, O.M. Shalit, Continuity of CP-semigroups in the point-strong topology. J.
Operator Theory 64, 149–154 (2010)
99. J.E. McCarthy, O.M. Shalit, Unitary N-dilations for tuples of commuting matrices. Proc. Am.
Math. Soc. 14, 563–571 (2013)
100. P.S. Muhly, B. Solel, Tensor algebras over C ∗ -correspondences: representations, dilations,
and C ∗ -envelopes. J. Funct. Anal. 158, 389–457 (1998)
101. P.S. Muhly, B. Solel, An algebraic characterization of boundary representations, in Operator
Theory: Advances and Applications, vol. 104 (Birkhäuser, Basel, 1998), pp. 189–196
102. P.S. Muhly, B. Solel, Quantum Markov processes (correspondences and dilations). Int. J.
Math. 13, 863–906 (2002)
103. P.S. Muhly, B. Solel, Hardy algebras, W*-correspondences and interpolation theory. Math.
Ann. 330, 353–415 (2004)
104. P. Muhly, B. Solel, Quantum Markov semigroups (product systems and subordination). Int. J.
Math. 18, 633–669 (2007)
105. V. Müller, F-H. Vasilescu, Standard models for some commuting multioperators. Proc. Am.
Math. Soc. 117, 979–989 (1993)
106. B. Nagy, On contractions in Hilbert space. Acta Sci. Math 15, 87–92 (2013)
107. I.L. Nielsen, M.A. Chuang, Quantum Computation and Quantum Information (Cambridge
University Press, Cambridge, 2000)
108. D. Opela, A generalization of Andô’s theorem and Parrott’s example. Proc. Am Math. Soc.
134, 2703–2710 (2006)
109. S. Pal, O.M. Shalit, Spectral sets and distinguished varieties in the symmetrized bidisc. J.
Funct. Anal. 266, 5779–5800 (2014)
110. S. Parrott, Unitary dilations for commuting contractions. Pac. J. Math. 34, 481–490 (1970)
111. K.R. Parthasarathy, An Introduction to Quantum Stochastic Calculus. Monographs in Mathe-
matics, vol. 85 (Birkhäuser, Basel, 2012)
112. W.L. Paschke, Inner product modules over B ∗ -algebras. Trans. Am. Math. Soc. 182, 443–468
(1973)
Dilation Theory: A Guided Tour 621

113. B. Passer, Shape, scale, and minimality of matrix ranges. Trans. Am. Math. Soc. 372, 1451–
1484 (2019)
114. B. Passer, V.I. Paulsen, Matrix range characterization of operator system properties (preprint).
arXiv:1912.06279
115. B. Passer, O.M. Shalit, Compressions of compact tuples. Linear Algebra Appl. 564, 264–283
(2019)
116. B. Passer, O.M. Shalit, B. Solel, Minimal and maximal matrix convex sets. J. Funct. Anal.
274, 3197–3253 (2018)
117. V.I. Paulsen, Representations of function algebras, abstract operator spaces, and Banach space
geometry. J. Funct. Anal. 109(1), 113–129 (1992)
118. V.I. Paulsen, Completely Bounded Maps and Operator Algebras (Cambridge University Press,
Cambridge, 2002)
119. V.I. Paulsen, M. Raghupathi, An Introduction to the Theory of Reproducing Kernel Hilbert
Spaces. Cambridge Studies in Advanced Mathematics, vol. 152 (Cambridge University Press,
Cambridge, 2016)
120. V.I. Paulsen, I.G. Todorov, M. Tomforde, Operator system structures on ordered spaces. Proc.
Lond. Math. Soc. 102, 25–49 (2011)
121. G. Pisier, Similarity Problems and Completely Bounded Maps. Lecture Notes of Mathematics,
vol. 1618 (Springer, Berlin, 1996)
122. G. Pisier, Introduction to Operator Space Theory, vol. 294 (Cambridge University Press,
Cambridge, 2003)
123. G. Popescu, Isometric dilations for infinite sequences of noncommuting operators. Trans. Am.
Math. Soc. 316, 523–536 (1989)
124. G. Popescu, Von Neumann inequality for (B(H)n )1 . Math. Scand. 68, 292–304 (1991)
125. G. Popescu, Poisson transforms on some C*-algebras generated by isometries. J. Funct. Anal.
161, 27–61 (1999)
126. G. Popescu, Free holomorphic functions on the unit ball of B(H)n . J. Funct. Anal. 241, 268–
333 (2006)
127. G. Popescu, Operator theory on noncommutative varieties. Ind. Univ. Math. J. 56, 389–442
(2006)
128. G. Popescu, Berezin transforms on noncommutative polydomains. Trans. Am. Math. Soc.
368, 4357–4416 (2016)
129. B. Prunaru, Lifting fixed points of completely positive semigroups. Integr. Equ. Oper. Theory
72, 219–222 (2012)
130. M. Ptak, Unitary dilations of multiparameter semigroups of operators. Ann. Polon. Math. 45,
237–243 (1985)
131. M. Rieffel, Induced representations of C*-algebras. Adv. Math. 13, 176–257 (1974)
132. M. Rieffel, C*-algebras associated with irrational rotations. Pac. J. Math. 93, 415–429 (1981)
133. F. Riesz, B. Sz.-Nagy, Functional Analysis (Dover, New York, 1990) (first published in 1955)
134. G. Salomon, O.M. Shalit, E. Shamovich, Algebras of bounded noncommutative analytic
functions on subvarieties of the noncommutative unit ball. Trans. Am. Math. Soc. 370, 8639–
8690 (2018)
135. G. Salomon, O.M. Shalit, E. Shamovich, Algebras of noncommutative functions on sub-
varieties of the noncommutative ball: the bounded and completely bounded isomorphism
problem. J. Funct. Anal. (2019). arXiv:1806.00410
136. D. Sarason, On spectral sets having connected complement. Acta Sci. Math. 26, 289–299
(1965)
137. D. Sarason, Generalized interpolation in H ∞ . Trans. Am. Math. Soc. 127, 179–203 (1967)
138. J. Sarkar, Applications of Hilbert module approach to multivariable operator theory, in
Operator Theory, ed. by D. Alpay (Springer, Berlin, 2014)
139. J. Sarkar, Operator theory on symmetrized bidisc. Ind. Univ. Math. J. 64, 847–873 (2015)
140. D. SeLegue, Minimal Dilations of CP maps and C*-extension of the Szegö limit theorem.
Ph.D Dissertation, University of California, Berkeley (1997)
622 O. M. Shalit

141. O.M. Shalit, E0 -dilation of strongly commuting CP0 -semigroups. J. Funct. Anal. 255, 46–89
(2008)
142. O.M. Shalit, What type of dynamics arise in E0 -dilations of commuting Quantum Markov
Semigroups?. Infin. Dimens. Anal. Quant. Probab. Relat. Top. 11, 393–403 (2008)
143. O.M. Shalit, Representing a product system representation as a contractive semigroup and
applications to regular isometric dilations. Can. Math. Bull. 53, 550–563 (2010)
144. O.M. Shalit, E-dilations of strongly commuting CP-semigroups (the nonunital case). Houston
J. Math. 35(1), 203–232 (2011)
145. O.M. Shalit, A sneaky proof of the maximum modulus principle. Am. Math. Month. 120(4),
359–362 (2013)
146. O.M. Shalit, Operator theory and function theory in Drury-Arveson space and its quotients,
in Operator Theory, ed. by D. Alpay (Springer, Berlin, 2014)
147. O.M. Shalit, A First Course in Functional Analysis (Chapman and Hall/CRC, Boca Raton,
2017)
148. O.M. Shalit, M. Skeide, Three commuting, unital, completely positive maps that have no
minimal dilation. Integr. Equ. Oper. Theory 71, 55–63 (2011)
149. O.M. Shalit, M. Skeide, CP-semigroups and dilations, subproduct systems and superproduct
systems: the multi-parameter case and beyond** (in preparation)
150. O.M. Shalit, B. Solel, Subproduct systems. Doc. Math. 14, 801–868 (2009)
151. M. Słociński, Unitary dilation of two-parameter semi-groups of contractions. Bull. Acad.
Polon. Sci. Ser. Sci. Math. Astronom. Phys. 22, 1011–1014 (1974)
152. A. Skalski, On isometric dilations of product systems of C*-correspondences and applications
to families of contractions associated to higher-rank graphs. Ind. Univ. Math. J. 58, 2227–
2252 (2009)
153. A. Skalski, J. Zacharias, Wold decomposition for representations of product systems of C*-
correspondences. Int. J. Math. 19, 455–479 (2008)
154. M. Skeide, Hilbert modules and applications in quantum probability. Habilitationsschrift,
Cottbus (2001). http://web.unimol.it/skeide/
155. M. Skeide, Isometric dilations of representations of product systems via commutants. Int. J.
Math. 19, 521–539 (2008)
156. M. Skeide, Classification of E0 -semigroups by product systems. Mem. Am. Math. Soc. 240,
1137 (2016)
157. B. Solel, Representations of product systems over semigroups an dilations of commuting CP
maps. J. Funct. Anal. 235, 593–618 (2006)
158. B. Solel, Regular dilations of representations of product systems. Math. Proc. R. Ir. Acad.
108, 89–110 (2008)
159. W.F. Stinespring, Positive functions on C*-algebras. Proc. Am. Math. Soc. 6, 211–216 (1955)
160. J. Stochel, F.H. Szafraniec, Unitary dilation of several contractions, in Operator Theory:
Advances and Applications, vol. 127 (Birkhäuser, Basel, 2001), pp. 585–598
161. E. Stroescu, Isometric dilations of contractions on Banach spaces. Pac. J. Math. 47, 257–262
(1973)
162. F.H. Szafraniec, Murphy’s Positive definite kernels and Hilbert C*-modules sreorganized.
Banach Center Publ. 89, 275–295 (2010)
163. B. Sz.-Nagy, Sur les contractions de l’espace de Hilbert. Acta Sci. Math. 15, 87–92 (1953)
164. B. Sz.-Nagy, Extensions of linear transformations in Hilbert space which extend beyond this
space. Appendix to F. Riesz and B. Sz.-Nagy, Functional Analysis, Ungar, New York, 1960.
Translation of “Prolongements des transformations de l’espace de Hilbert qui sortent de cet
espace”, Budapest (1955)
165. N.T. Varopoulos, On an inequality of von Neumann and an application of the metric theory
of tensor products to operators theory. J. Funct. Anal. 16, 83–100 (1974)
166. A. Vernik, Dilations of CP-maps commuting according to a graph. Houston J. Math. 42,
1291–1329 (2016)
167. F. vom Ende, G. Dirr, Unitary dilations of discrete-time quantum-dynamical semigroups. J.
Math. Phys. 60, 1–17 (2019)
Dilation Theory: A Guided Tour 623

168. J. von Neumann, Eine Spektraltheorie für allgemeine Operatoren eines unitären Raumes.
Math. Nachr. 4, 258–281 (1951)
169. B. Xhabli, Universal operator system structures on ordered spaces and their applications.
Thesis (Ph.D.) University of Houston (2009)
170. B. Xhabli, The super operator system structures and their applications in quantum entangle-
ment theory. J. Funct. Anal. 262, 1466–1497 (2012)
171. A. Zalar, Operator positivestellensatze for noncommutative polynomials positive on matrix
convex sets. J. Math. Anal. Appl. 445, 32–80 (2017)
Riesz-Fischer Maps, Semi-frames
and Frames in Rigged Hilbert Spaces

Francesco Tschinke

Abstract In this note we present a review, some considerations and new results
about maps with values in a distribution space and domain in a σ -finite measure
space X. Namely, this is a survey about Bessel maps, frames and bases (in particular
Riesz and Gel’fand bases) in a distribution space. In this setting, the Riesz-Fischer
maps and semi-frames are defined and new results about them are obtained. Some
examples in tempered distributions space are examined.

Keywords Frames · Bases · Distributions · Rigged Hilbert space

Mathematics Subject Classification (2010) Primary 42C15; Secondary 47A70,

46F05

1 Introduction

Given a Hilbert space H with inner product ·|· and norm ·, a frame is a sequence
of vectors {fn } in H if there exist A, B > 0 such that:
∞

Af 2 ≤ |f |fn |2 ≤ Bf 2 , ∀f ∈ H.
k=1

As known, this notion generalizes orthonormal bases, and has reached an increasing
level of popularity in many fields of interests, such as signal theory, image
processing, etc., but it is also an important tool in pure mathematics: in fact it plays
key roles in wavelet theory, time-frequency analysis, the theory of shift-invariant
spaces, sampling theory and many other areas (see [10, 11, 19, 20, 24]).

F. Tschinke ()
Università di Palermo, Palermo, Italy
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 625

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_29
626 F. Tschinke

A generalization of frame, the continuous frame, was proposed by Kaiser [24]

and by Alí et al. [1, 2]: if (X, μ) is a measure space with μ as σ -finite positive
measure, a function F : x ∈ X → Fx ∈ H is a continuous frame with respect to
(X, μ) if:
(i) F is weakly measurable, i.e. the map x ∈ X → f |Fx  ∈ C is μ-measurable
for all f ∈ H;
(ii) there exist A, B > 0 such that, for all f ∈ H:

Af 2 ≤ | f |Fx  |2 dμ ≤ Bf 2 , ∀f ∈ H.
X

Today, the notion of continuous frames in Hilbert spaces and their link with the
theory of coherent states is well-known in the literature.
With the collaboration of C. Trapani and T. Triolo [29], the author introduced
bases and frames in distributional spaces. To illustrate the motivations for this study,
we have to consider the rigged Hilbert space (or Gel’fand triple) [16, 17]: that is, if
H is a Hilbert space, the triple:

D[t] ⊂ H ⊂ D× [t × ],

where D[t] is a dense subspace of H endowed with a locally convex topology

t stronger than the Hilbert norm and D× [t × ] is the conjugate dual space of D
endowed with the strong dual topology t × . If D is reflexive, the inclusions are dense
and continuous.
In this setting, let us consider the generalized eigenvectors of an operator, i.e.
eigenvectors that do not belong to H. More precisely: if A is an essentially self-
adjoint operator in D which maps D[t] into D[t] continuously, then A has a
continuous extension  given by its adjoint, (i.e.  = A† ) from D× into itself.
A generalized eigenvector of A, with eigenvalue λ ∈ C, is an eigenvector of Â; that
is, a conjugate linear functional ωλ ∈ D× such that:

Af |ωλ  = λ f |ωλ  , ∀f ∈ D.

The above equality can be read as Âωλ = A† ωλ = λωλ .

A simple and explicative example is given by the derivative operator: A :=
d
i dx : S(R) → S(R) where S(R) is the Schwartz space (i.e. infinitely differentiable
rapidly decreasing functions). The rigged Hilbert space is:

S(R) ⊂ L2 (R) ⊂ S × (R), (1.1)

where the set S × (R) is known as space of tempered distributions. Then ωλ (x) =
√1 e −iλx -that does not belong to L2 (R)—is a generalized eigenvector of A with λ
2π
as eigenvalue.
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 627

Each function ωλ can be viewed as a regular distribution of S × (R) through the

following integral representation:
 
1
φ|ωλ  = φ(x)ωλ (x)dx = √ φ(x)eiλx dx := φ̌(λ)
R 2π R

and the linear functional φ → φ̌(λ) defined on S(R) is continuous. Furthermore by

the Fourier-Plancherel theorem, one has:

φ̌22 = | φ|ωλ  |2 dx = φ22 .
R

With a limiting procedure, the Fourier transform can be extended to L2 (R). Since
a function f ∈ L2 (R) defines a regular tempered distribution, we have, for all
φ ∈ S(R):
    
1
φ|f  : = f (x)φ(x)dx = √ fˆ(λ)eiλx dλ φ(x)dx
R R 2π R
 
= fˆ(λ)φ̌(λ)dλ = fˆ(λ)φ|ωλ  dλ.
R R

That is:

f = fˆ(λ)ωλ dλ. (1.2)
R

in weak sense. The family {ωλ ; λ ∈ R} of the previous example can be considered
as the range of a weakly measurable function ω : R → S × (R) which allows a
representation as in (1.2) of any f ∈ L2 (R) in terms of generalized eigenvectors of
A. This is an example of a distribution basis. More precisely, since the Fourier-
Plancherel theorem corresponds to the Parseval identity, this is an example of
Gel’fand distribution basis [29, Subsec. 3.4], that plays, in S × (R), the role of an
orthonormal basis in a Hilbert space.
The example above is a particular case of the Gel’fand-Maurin theorem (see
[16, 18] for details), which states that, if D is a domain in a Hilbert space H which
is a nuclear space under a certain topology τ , and A is an essentially self-adjoint
operator on D which maps D[t] into D[t] continuously, then A admits a complete
set of generalized eigenvectors.
If σ (A) is the spectrum of the closure of the operator A, the completeness of the
set {ωλ ; λ ∈ σ (A)} is understood in the sense that the Parseval identity holds, that
is:
 1/2
f  = | f |ωλ  |2 dλ , ∀f ∈ D. (1.3)
σ (A)
628 F. Tschinke

To each λ there corresponds the subspace Dλ× ⊂ D× of all generalized eigenvectors

whose eigenvalue is λ. For all f ∈ D it is possible to define a linear functional f˜λ on
Dλ× by f˜λ (ωλ ) := ωλ |f  for all ωλ ∈ Dλ× . Following [16, 17], the correspondence
D → Dλ×× defined by f → f˜λ is called the spectral decomposition of the element
f corresponding to A. If f˜λ ≡ 0 implies f = 0 (i.e. the map f → f˜λ is injective)
then A is said to have a complete system of generalized eigenvectors.
The completeness and condition (1.3) may be interpreted as a kind of orthonor-
mality of the ωλ ’s: the family {ωλ }λ∈σ (A) in [29] is called a Gel’fand basis.
Another meaningful situation comes from quantum mechanics. Let us consider
the rigged Hilbert space (1.1) corresponding to the one-dimensional case. The
Hamiltonian operator H is an essentially self-adjoint operator on S(R), with self-
adjoint extension H on the domain D(H ), dense in L2 (R). According to the spectral
expansion theorem in the case of non-degenerate spectrum, for all f ∈ L2 (R), the
following decomposition holds:
 
f = c n un + c(α)uα dμ(α).
n∈J σc

The set {un }n∈J , J ⊂ N, is an orthonormal system of eigenvectors of H ; the measure

μ is a continuous measure on σc ⊂ R and {uα }α∈σc are generalized eigenvectors of
H in S × (R). This decomposition is unique. Furthermore:
 
f 2 = |cn |2 + |c(α)|2 dμ(α).
n∈J σc

The subset σc , corresponding to the continuous spectrum, is a union of intervals of

R, i.e. the index α is continuous. The generalized eigenvectors uα are distributions:
they do not belong to L2 (R), therefore the “orthonormality” between generalized
eigenvectors is not defined. Nevertheless, it is often denoted by the physicists with
the Dirac delta: “ uα |uα ”=δα−α .
Frames, semi-frames, Riesz bases, etc. are families of sequences that generalize
orthonormal bases in Hilbert space maintaining the possibility to reconstruct vectors
of the space as superposition of more ‘elementary’ vectors renouncing often to the
uniqueness of the representation, but gaining in versatility.
In this sense, they have been considered in literature in various spaces of
functions and distributions: see for example the following (not exhaustive) list:
[4, 8, 12, 14, 15, 25, 26].
It is remarkable that in a separable Hilbert space, orthonormal bases and Riesz
bases are countable and notions corresponding to Riesz basis have been formulated
in the continuous setting, but it is known that they exist only if the space given by
the index set is discrete [7, 22, 23]. On the other hand, in the distributions and rigged
Hilbert space setting the corresponding objects can be continuous.
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 629

Revisiting some results of [29] about Bessel maps, frames and (Gel’fand and
Riesz) bases in distribution set-up, in this paper the notions of Riesz-Fischer map
and of semi-frames in a space of distributions are proposed.
After some preliminaries and notations (Sect. 2), in Sect. 3 distribution Bessel
maps are considered and the notion of distribution Riesz-Fischer maps is proposed,
showing some new results about them (such as bounds and duality properties).
Since distribution Bessel maps are not, in general, bounded by a Hilbert norm,
we consider appropriate to define in Sect. 4 the distribution semi-frames, notion
already introduced in a Hilbert space by J.-P. Antoine and P. Balasz [3]. Finally,
distribution frames, distribution bases, Gel’fand and Riesz bases, considered in [29],
are revisited in Sect. 5 with some additional examples.

2 Preliminary Definitions and Facts

2.1 Rigged Hilbert Space

Let D be a dense subspace of a Hilbert space H endowed with a locally convex

topology t finer than the topology induced by a Hilbert norm. Denote as D×
the vector space of all continuous conjugate linear functionals on D[t], i.e., the
conjugate dual of D[t], endowed with the strong dual topology t × = β(D× , D),
which can be defined by the seminorms:

qM (F ) = sup | F |g |, F ∈ D× ,
g∈M

where M is a bounded subset of D[t]. In this way, a rigged Hilbert space is defined
in a standard fashion:

D[t] +→ H +→ D× [t × ], (2.1)

where +→ denotes a continuous injection. Since the Hilbert space H can be

identified with a subspace of D× [t × ], we will systematically read (2.1) as a chain
of topological inclusions: D[t] ⊂ H ⊂ D× [t × ]. These identifications imply that
the sesquilinear form B(·, ·) which puts D and D× in duality is an extension of
the inner product of H; i.e. B(ξ, η) = ξ |η, for every ξ, η ∈ D (to simplify
notations we adopt the symbol ·|· for both of them) and also the embedding map
ID,D× : D → D× can be taken to act on D as ID,D× f = f for every f ∈ D. For
more insights, besides to [16, 17], see also [21]. In this paper, if is not otherwise
specified, we will work with a rigged Hilbert space D[t] +→ H +→ D× [t × ] with
D[t] reflexive, in this way the embedding +→ is continuous and dense.
630 F. Tschinke

2.2 The Space L(D , D × )

If D[t] ⊂ H ⊂ D× [t × ] is a rigged Hilbert space, let us denote by L(D, D× )

the vector space of all continuous linear maps from D[t] into D× [t × ]. If D[t] is
barreled (e.g., reflexive), an involution X → X† can be introduced in L(D, D× ) by
the identity:
 
X† η|ξ = Xξ |η, ∀ξ, η ∈ D.

Hence, in this case, L(D, D× ) is a † -invariant vector space. We also denote by L(D)
the algebra of all continuous linear operators Y : D[t] → D[t] and by L(D× )
the algebra of all continuous linear operators Z : D× [t × ] → D× [t × ]. If D[t] is
reflexive, for every Y ∈ L(D) there exists a unique operator Y × ∈ L(D× ), the
adjoint of Y , such that
 
|Y g = Y × |g , ∀ ∈ D× , g ∈ D.

In similar way an operator Z ∈ L(D× ) has an adjoint Z × ∈ L(D) such that

(Z × )× = Z. In the monograph [5] the topic is treated more deeply.

2.3 Weakly Measurable Maps

In this paper a weakly measurable map is considered as a subset of D× : if (X, μ)

is a measure space with μ a σ -finite positive measure, ω : x ∈ X → ωx ∈ D× is
a weakly measurable map if, for every f ∈ D, the complex valued function x →
f |ωx  is μ-measurable. Since the form which puts D and D× in conjugate duality
is an extension of the inner product of H, we write f |ωx  for ωx |f , f ∈ D. If
not otherwise specified, throughout the paper we will work with a measure space
(X, μ) above described.
Definition 2.1 Let D[t] be a locally convex space, D× its conjugate dual and ω :
x ∈ X → ωx ∈ D× a weakly measurable map, then:
(i) ω is total if, f ∈ D and f |ωx  = 0 μ-a.e. x ∈ X implies f = 0;
(ii) ω
= is μ-independent if the unique measurable function ξ on (X, μ) such that, if
X ξ(x) g|ωx  dμ = 0 for every g ∈ D, then ξ(x) = 0 μ-a.e.
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 631

3 Bessel and Riesz-Fischer Distribution Maps

3.1 Bessel Distribution Maps

Definition 3.1 Let D[t] be a locally convex space. A weakly measurable map ω is
a Bessel distribution map (briefly: Bessel map) if for every f ∈ D,

| f |ωx  |2 dμ < ∞.
X

The following Proposition is the analogue of Proposition 2 and Theorem 3 in [30,

Section 2, Chapter 4].
Proposition 3.2 ([29, Proposition 3.1]) If D[t] a Fréchet space, and ω : x ∈ X →
ωx ∈ D× a weakly measurable map. The following statements are equivalent:
(i) ω is a Bessel map;
(ii) there exists a continuous seminorm p on D[t] such that:
 1/2
| f |ωx  | dμ 2
≤ p(f ), ∀f ∈ D;
X

(iii) for every bounded subset M of D there exists CM > 0 such that:

sup ξ(x) ωx |f  dμ ≤ CM ξ 2 , ∀ξ ∈ L2 (X, μ). (3.1)
f ∈M X

We have also the following:

Lemma 3.3 ([29, Lemma 3.4]) If D is a Fréchet space and ω a Bessel distribution
map, then:

f |ωx  ωx dμ
X

converges for every f ∈ D to an element of D× . Moreover, the map


D 4 f → f |ωx  ωx dμ ∈ D×
X

is continuous.
Let ω be a Bessel map: the previous lemma allows to define on D × D the
sesquilinear form :

(f, g) = f |ωx  ωx |g dμ.
X
632 F. Tschinke

By Proposition 3.2, one has for all f, g ∈ D,


|(f, g)| = f |ωx  g|ωx  dμ ≤  f |ωx  2  g|ωx  2 ≤ p(f )p(g).
X

This means that  is jointly continuous on D[t]. Hence there exists an operator
Sω ∈ L(D, D× ), with Sω = Sω† , Sω ≥ 0, such that:

(f, g) = Sω f |g = f |ωx  ωx |g dμ, ∀f, g ∈ D (3.2)
X

that is,

Sω f = f |ωx  ωx dμ, ∀f ∈ D.
X

Since ω is a Bessel distribution map and ξ ∈ L2 (X, μ), we put for all g ∈ D:

$ξω (g) := ξ(x) ωx |g dμ. (3.3)
X

Then $ω is a continuous conjugate linear functional on D, i.e. $ω ∈ D× . We write:

ξ ξ


$ξω := ξ(x)ωx dμ
X

in weak sense. Therefore we can define a linear map Tω : L2 (X, μ) → D× [t × ],

which will be called the synthesis operator, by:

Tω : ξ → $ξω .

By (3.1), it follows that Tω is continuous from L2 (X, μ), endowed with its natural
norm, into D× [t × ]. Hence, it possesses a continuous adjoint Tω× : D[t] →
L2 (X, μ), which is called the analysis operator, acting as follows:

Tω× : f ∈ D[t] → ξf ∈ L2 (X, μ), where ξf (x) = f |ωx  , x ∈ X.

One has that Sω = Tω Tω× .

Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 633

3.2 Riesz-Fischer Distribution Map

Definition 3.4 Let D[t] be a locally convex space. A weakly measurable map ω :
x ∈ X → ωx ∈ D× is called a Riesz-Fischer distribution map (briefly: Riesz-
Fischer map) if, for every h ∈ L2 (X, μ), there exists f ∈ D such that:

f |ωx  = h(x) μ-a.e. (3.4)

In this case, we say that f is a solution of equation f |ωx  = h(x).

Clearly, if f1 and f2 are solutions of (3.4), then

f1 − f2 ∈ ω⊥ := {g ∈ D : g|ωx  = 0, μ − a.e.}.

If ω is total, the solution is unique. We prove the following:

Lemma 3.5 Let D be a reflexive locally convex space, h(x) be a measurable
function and x ∈ X → ωx ∈ D× [t × ] a weakly measurable map. Then the equation:

f |ωx  = h(x)

admits a solution f ∈ D if, and only if, there exists a bounded subset M of D such
that

|h(x)| ≤ sup | f |ωx  | μ-a.e.

f ∈M

Proof Necessity is obvious. Conversely, let x ∈ X be a point such that f |ωx  =

h(x) = 0. Let us consider the subspace Vx of D× given by Vx := {αωx }α∈C, and
let us define the functional μ on Vx by: μ(αωx ) := αh(x). We have that

|μ(αωx )| = |αh(x)| ≤ |α| sup | f |ωx  | = sup | f |αωx  |,

f ∈M f ∈M

in other words:

|μ(Fx )| ≤ sup | f |Fx  |

f ∈M

for all Fx ∈ Vx . By the Hahn-Banach theorem, there exists an extension μ̃ to D×

such that

|μ̃(F )| ≤ sup | f |F  |,
f ∈M
634 F. Tschinke

 
for every F ∈ D× . Since D is reflexive, there exists f¯ ∈ D such that μ̃(F ) = f¯|F .
The statement follows from the fact that μ(ωx ) = h(x).
If M is a subspace of D and the topology of D is generated by the family of
seminorms {pα }α∈I , then the topology on the quotient space D/M is defined, as
usual, by the seminorms {p̃α }α∈I , where

p̃α (f˜) := inf{pα (g) : g ∈ f + M}.

The following proposition can be compared to the case of Riesz-Fischer sequences:

see [30, Chapter 4, Section 2, Proposition 2].
Proposition 3.6 Assume that D[t] is a Fréchet space. If the map ω : x ∈ X →
ωx ∈ D× is a Riesz-Fischer map, then for every continuous seminorm p on D, there
exists a constant C > 0 such that, for every solution f of (3.4),

p̃(f˜) := inf{p(g) : g ∈ f + ω⊥ } ≤ C f |ωx  2 .

Proof Since ω⊥ is closed, it follows that the quotient D/ω⊥ := Dω⊥ is a Fréchet
space. If f ∈ D, we put f˜ := f + ω⊥ . Let h ∈ L2 (X, μ) and f a solution of
(3.4) corresponding to h; then, we can define an operator S : L2 (X, μ) → Dω⊥ by
h → f˜. Let us consider a sequence hn ∈ L2 (X, μ) such that hn → 0 and, for each
n ∈ N, let fn be a corresponding solution of (3.4). One has that
 
|hn (x)|2dμ → 0, i.e. | fn |ωx  |2 dμ → 0.
X X

This  implies that fn |ωx  → 0 in measure, so there exists a subsequence such
that fnk |ωx → 0 a.e. (see [13]). On the other hand, if Shn = f˜n is a sequence
convergent to f˜ in Dω⊥ w.r. to the quotient topology defined by the seminorms
p̃(·), it follows that the sequence is convergent in the weak topology of Dω⊥ , i.e.:
   
f˜n |F̃ → f˜|F̃ ∀F̃ ∈ Dω×⊥ .

Let us consider the canonical surjection ρ : D → Dω⊥ , ρ : f → f˜ = f + ω⊥ .

Its transpose map (adjoint) ρ † : Dω×⊥ → D× is injective (see [21], p. 263) and
ρ † [Dω×⊥ ] = ω⊥⊥ . Then ρ † : Dω×⊥ → ω⊥⊥ is invertible. Hence,
     
f˜n |F̃ = ρ(fn )|(ρ † )−1 (F ) = fn |ρ † ((ρ † )−1 (F )) = fn |F  , ∀F ∈ ω⊥⊥ .

Thus, if f˜n → f˜ in the topology of Dω⊥ , then fn |F  → f |F , for all F ∈ ω⊥⊥ ,
and, in particular, since ω ⊂ ω⊥⊥ , one has fn |ωx  → f |ωx . Since fn |ωx  has
a subsequence convergent to 0, one has f ∈ ω⊥ . From the closed graph theorem, it
follows that the map S is continuous, i.e. for all continuous seminorms p̃ on Dω⊥
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 635

there exists C > 0 such that: p̃(Sh) ≤ Ch2 , for all h ∈ L2 (X, μ). The statement
follows from the definition of Riesz-Fischer map.
Corollary 3.7 Assume that D[t] is a Fréchet space. If the map ω : x ∈ X → ωx ∈
D× is a total Riesz-Fischer map, then for every continuous seminorm p on D, there
exists a constant C > 0 such that, for the solution f of (3.4),

p(f ) ≤ C f |ωx  2 .

Remark 3.8 For an arbitrary weakly measurable map ω, we define the subset of
D[t]: D(Vω ) := {f ∈ D : f |ωx  ∈ L2 (X, μ)} and the analysis operator Vω :
f ∈ D(Vω ) → f |ωx  ∈ L2 (X, μ). Clearly, ω is a Riesz-Fischer map if and only
if Vω : D(Vω ) → L2 (X, μ) is surjective. If ω is total, it is injective too, so Vω is
invertible. A consequence of Corollary 3.7 is that Vω−1 : L2 (X, μ) → D(Vω ) is
continuous.

3.3 Duality

Definition 3.9 Let D +→ H +→ D× be a rigged Hilbert space and ω a weakly

measurable map. We call dual map of ω, if it exists, a weakly measurable map θ
such that for all f, g ∈ D:

f |θx  ωx |g dμ < ∞
X

and

f |g = f |θx  ωx |g dμ, ∀f, g ∈ D.
X

Proposition 3.10 Suppose that ω is a Riesz-Fischer map. Then the map θ is a

Bessel map.
 
Proof For all h ∈ L2 (X, μ) there exists f¯ ∈ D such that f¯|ωx = h(x) μ-a.e.
Since θ is a dual map, one has that:

h(x) θx |g dμ < ∞
X

for all h ∈ L2 (X, μ). It follows that θx |g ∈ L2 (X, μ) (see [28, Chapter 6,
Exercise 4]).
636 F. Tschinke

Proposition 3.11 Let D be reflexive and let ω be a μ-independent Bessel map.

Furthermore, suppose that for all h ∈ L2 (X, μ) there exists a bounded subset M ⊂
D such that:

h(x) ωx |g dμ ≤ sup | f |g |, ∀g ∈ D,
X f ∈M

then the dual map θ is a Riesz-Fischer map.

Proof Since h ∈ L2 (X, μ), and since ω is a Bessel map, one has:

h(x) ωx |g dμ < ∞.
X

Let us consider

g= g|ωx  θx dμ
X

as element of D× . We define the following functional on D (as subspace of D× ):


μ(g) := h(x) ωx |g dμ.
X

By hypothesis, one has:

|μ(g)| ≤ sup | f |g |.

f ∈M

By the Hahn-Banach theorem, there exists an extension μ̃ to D× such that:

|μ̃(G)| ≤ sup | f |G |, ∀G ∈ D× .

f ∈M

 
Since D is reflexive, there exists f˜ ∈ D×× = D such that μ̃(G) = f˜|G . In
particular
  
f˜|g = h(x) ωx |g dμ.
X

Since θ is dual of ω, we have too:

    
f˜|g = f˜|θx ωx |g dμ.
X
 
But ω is μ-independent, then it follows that h(x) = f˜|θx μ-a.e.
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 637

4 Semi-frames and Frames

4.1 Distribution Semi-frames

Definition 4.1 Given a rigged Hilbert space D +→ H +→ D× , a Bessel map ω is a

distribution upper semi-frame if it is complete (total) and if there exists B > 0:

0< | f |ωx  |2 dμ ≤ Bf 2 , ∀f ∈ D, f = 0.
X

Since the injection D +→ H is continuous, it follows that there exists a

continuous seminorm p on D such that f  ≤ p(f ) for all f ∈ D. If ξ ∈ L2 (X, μ),
ξ
then the continuous conjugate functional $ω on D defined in (3.3) is bounded in
ξ
D[ · ]; it follows that it has a bounded extension $̃ω to H, defined, as usual, by a
limiting procedure. Therefore, there exists a unique vector hξ ∈ H such that:
 
$̃ξω (g) = hξ |g , ∀g ∈ H.

This implies that the synthesis operator Tω takes values in H, it is bounded and
Tω  ≤ B 1/2 ; its hilbertian adjoint Cω := Tω∗ extends the analysis operator Tω× .
The action of Cω can be easily described: if g ∈ H and {gn } is a sequence
of elements of D, norm converging to g, then the sequence {ηn }, where ηn (x) =
gn |ωx , is convergent in L2 (X, μ). Put η = limn→∞ ηn . Then,
 
Tω ξ |g = lim ξ(x) ωx |gn  dμ = ξ(x)η(x)dμ.
n→∞ X X

Hence Tω∗ g = η.
The function η ∈ L2 (X, μ) depends linearly on g, for each x ∈ X. Thus we can
define a linear functional ω̌x by
 
g|ω̌x = lim gn |ωx  , g ∈ H; gn → g. (4.1)
n→∞

Of course, for each x ∈ X, ω̌x extends ωx ; however ω̌x need not be continuous, as
a functional on H. We conclude that:
 
Tω∗ : g → g|ω̌x ∈ L2 (X, μ).

Moreover, in this case, the sesquilinear form  in (3.2), which is well defined on
D × D, is bounded with respect to  ·  and possesses a bounded extension  ˆ to H.
Hence there exists a bounded operator Ŝω in H, such that:
 
ˆ
(f, g) = Ŝω f |g , ∀f, g ∈ H. (4.2)
638 F. Tschinke

Since
  
Ŝω f |g = f |ωx  ωx |g dμ, ∀f, g ∈ D,
X

Ŝω extends the frame operator Sω and Sω : D → H. It is easily seen that Ŝω = Ŝω∗
and Ŝω = Tω Tω∗ . By definition, we have:

0 < Ŝω f  ≤ Bf , ∀f ∈ H, f = 0.

Then Ŝω is bounded, self-adjoint and injective too. This means that Ran Sω is dense
in H, and Ŝω−1 is densely defined. If ω is not a frame, Ŝω−1 is an unbounded, self-
adjoint operator (see [3]).
Remark 4.2 If {ωx }x∈X is an upper semi-frame, then there exists a continuous
seminorm p on D such that  f |ωx  2 ≤ p(f ) for all f ∈ D. In fact, the injection
D +→ H is continuous, i.e. f  ≤ p(f ) for all f ∈ D. The converse is not true:
let us consider the rigged Hilbert space S(R) +→ L2 (R) +→ S × (R); the system of
derivative of Dirac’s deltas {δx }x∈R is total. Since S(R) is a Fréchet space, (ii) of
Proposition 3.2 it holds. However {δx }x∈R is not a distribution upper semi-frame; in
fact:

 
| φ|δx |2 dx = φ 22 ∀φ ∈ S(R),
R

d
but the derivative operator dx : S(R) → L2 (R) is unbounded (clearly with respect
to the topology of the Hilbert norm).
Remark 4.3 In [29] it is defined the notion of bounded Bessel map, that is a Bessel
map in rigged Hilbert space such that:

| f |ωx  |2 dμ ≤ Bf 2 , ∀f ∈ D.
X

It is a more general notion than upper bounded semi-frame. In fact, we can consider,
as example, the distribution ωx := ηK (x)δx where ηK (x) is a C ∞ -function with
compact support K and M := maxx∈K |ηK (x)|:
 
| φ|ωx  | dx =
2
| φ|ηK (x)δx  |2 dx
R R
 
= |ηK (x)φ(x)|2 dx ≤ M 2 |φ(x)|2 dx ≤ M 2 φ22 .
R K

Therefore ω is a bounded Bessel map, but it is not total, then it is not an upper
semi-frame.
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 639

Definition 4.4 Given a rigged Hilbert space D +→ H +→ D× , a Bessel map ω is a

distribution lower semi-frame if there exists A > 0 such that:

Af 2 ≤ | f |ωx  |2 dμ, ∀f ∈ D.
X

By definition, it follows that ω is total. If D is a Fréchet space, by Proposition 3.2

one has Sω ∈ L(D, D× ) and, if ω is not a frame, Sω is unbounded. Furthermore, Sω
is injective, and Sω−1 is bounded.
Example Let us consider the space OM , known (see [27]) as the set of infinitely
differentiable functions on R that are polynomially bounded together with their
derivatives. Let us consider g(x) ∈ OM such that 0 < m < |g(x)|. If we define
ωx := g(x)δx , then {ωx }x∈R is a distribution lower semi-frame with A = m2 .
The proof of the following lemma is analogous to that of [3, Lemma 2.5]:
Lemma 4.5 Let ω be an upper semi-frame with upper frame bound M and θ a total
family dual to ω. Then θ is a lower semi-frame, with lower frame bound M −1 .

4.2 Distribution Frames

This section is devoted to distribution frames, with main results already shown in
[29].
Definition 4.6 ([29, Definition 3.6]) Let D[t] ⊂ H ⊂ D× [t × ] be a rigged
Hilbert space, with D[t] a reflexive space and ω a Bessel map. We say that ω is
a distribution frame if there exist A, B > 0 such that:

Af 2 ≤ | f |ωx  |2 dμ ≤ Bf 2 , ∀f ∈ D.
X

A distribution frame ω is clearly, in particular, an upper bounded semi-frame.

Thus, we can consider the operator Ŝω defined in (4.2). It is easily seen that, in this
case,

Af  ≤ Ŝω f  ≤ Bf , ∀f ∈ H.

This inequality, together with the fact that Ŝω is symmetric, implies that Ŝω has a
bounded inverse Ŝω−1 everywhere defined in H.
Remark 4.7 It is worth noticing that the fact that ω and Sω extend to H does not
mean that ω a frame in the Hilbert space H, because we do not know if the extension
of Sω has the form of (3.2) with f, g ∈ H.
To conclude this section, we recall a list of properties of frames proved in [29].
640 F. Tschinke

Lemma 4.8 ([29, Lemma 3.8]) Let ω be a distribution frame. Then, there exists
Rω ∈ L(D) such that Sω Rω f = Rω× Sω f = f , for every f ∈ D.
As a consequence, the reconstruction formulas for distribution frames hold for
all f ∈ D:

f = Rω× Sω f = f |ωx  Rω× ωx dμ;
X

f = Sω Rω f = Rω f |ωx  ωx dμ.
X

These representations have to be interpreted in the weak sense.

Remark 4.9 The operator Rω acts as an inverse of Sω . On the other hand the
operator Ŝω has a bounded inverse Ŝω−1 everywhere defined in H. It results that
[29, Remark 3.7]: Ŝω−1 D ⊂ D and Rω = Ŝω−1 D .
There exists the dual frame:
Proposition 4.10 ([29, Lemma 3.10]) Let ω be a distribution frame. Then there
exists a weakly measurable function θ such that:

f |g = f |θx  ωx |g dμ, ∀f, g ∈ D.
X

Where θx := Rω× ωx . The frame operator Sθ for θ is well defined and we have:
Sθ = ID,D× Rω .
The distribution function θ , constructed in Proposition 4.10, is also a distribution
frame, called the canonical dual frame of ω. Indeed, it results that [29]:

B −1 f 2 ≤ Sθ f |f  ≤ A−1 f 2 , ∀f ∈ D.

4.3 Parseval Distribution Frames

Definition 4.11 If ω is a distribution frame, then we say that:

(a) ω is a tight distribution frame if we can choose A = B as frame bounds. In this
case, we usually refer to A as a frame bound for ω;
(b) ω is a Parseval distribution frame if A = B = 1 are frame bounds.
More explicitly a weakly measurable distribution function ω is called a Parseval
distribution frame if [29, Definition 3.13]:

| f |ωx  |2 dμ = f 2 , f ∈ D.
X
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 641

It is clear that a Parseval distribution frame is a frame in the sense of Definition 4.6
with Sω = ID , the identity operator of D.
Lemma 4.12 ([29, Lemma 3.14]) Let D ⊂ H ⊂ D× be a rigged Hilbert space
and ω : x ∈ X → ωx ∈ D× a weakly measurable map. The following statements
are equivalent.
(i) ω is a Parseval
= distribution frame;
(ii) f |g== X f |ωx  ωx |g dμ, ∀f, g ∈ D;
(iii) f = X f |ωx  ωx dμ, the integral on the r.h.s. is understood as a continuous
conjugate linear functional on D, that is an element of D× .
The representation in (iii) of Lemma 4.12 is not necessarily unique.

5 Distribution Basis

Definition 5.1 ([29, Definition 2.3]) Let D[t] be a locally convex space, D× its
conjugate dual and ω : x ∈ X → ωx ∈ D× a weakly measurable map. Then ω
is a distribution basis for D if, for every f ∈ D, there exists a unique measurable
function ξf such that:

f |g = ξf (x) ωx |g dμ, ∀f, g ∈ D
X

and, for every x ∈ X, the linear functional f ∈ D → ξf (x) ∈ C is continuous in

D[t].
The above formula can be represented by:

f = ξf (x)ωx dμ
X

in weak sense.
Remark 5.2 Clearly, if ω is a distribution basis, then it is μ-independent. Further-
more, since f ∈ D → ξf (x) ∈ C continuously, there exists a unique weakly
μ-measurable map θ : X → D× such that: ξf (x) = f |θx  for every f ∈ D. We
call θ dual map of ω. If θ is μ-independent, then it is a distribution basis too.

5.1 Gel’fand Distribution Basis

The Gel’fand distribution basis, introduced in [29], is a good substitute for the notion
of an orthonormal basis which is meaningless in the present framework.
642 F. Tschinke

Definition 5.3 A weakly measurable map ζ is Gel’fand distribution basis if it is a

μ-independent Parseval distribution frame.
By definition and Lemma 4.12, this means that, for every f ∈ D there exists a
unique function ξf ∈ L2 (X, μ) such that:

f = ξf (x)ζx dμ (5.1)
X
=
with ξf (x) = f |ζx  μ-a.e. Furthermore f 2 = X | f |ζx  |2 dμ and ζ is total
too.
For every x ∈ X, the map f ∈ H → ξf (x) ∈ C defines as in (4.1) a linear
functional ζˇx on H, then for all f ∈ H:
  
f = f |ζˇx ζx dμ.
X

We have the following characterization result [29]:

Proposition 5.4 ([29, Proposition 3.15]) Let D ⊂ H ⊂ D× be a rigged Hilbert
space and let ζ : x ∈ X → ζx ∈ D× be a Bessel distribution map. Then the
following statements are equivalent.
(a) ζ is a Gel’fand distribution basis.
(b) The synthesis operator Tζ is an isometry of L2 (X, μ) onto H.
Example ([29, Example 3.17]) Given the rigged Hilbert space:

S(R) +→ L2 (R) +→ S × (R),

for x ∈ R the function ζx (y) = √1 e−ixy , defines a (regular) tempered distribution:

2π
in fact, denoting as usual by ĝ, ǧ, respectively, the Fourier transform and the inverse
Fourier transform of g ∈ L2 (R), one has that:

1
S(R) 4 φ → φ|ζx  = √ φ(y)e−ixy dy = φ̂(x) ∈ C.
2π R

For all x ∈ R the set of functions ζ := {ζx (y)}x∈R is a Gel’fand distribution basis,
because the synthesis operator Tζ : L2 (R) → L2 (R) defined by:

1
(Tζ ξ )(x) = √ ξ(y)e−ixy dy = ξ̂ (x), ∀ξ ∈ L2 (R)
2π R

is an isometry onto L2 (R) by Plancherel theorem. The analysis operator is: Tζ∗ f =
fˇ, for all f ∈ L2 (R).
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 643

Example ([29, Example 3.18]) Let us consider again S(R) +→ L2 (R) +→ S × (R).
For x ∈ R, let us consider the Dirac delta δx : S(R) → C, φ → φ|δx  := φ(x).
The set of Dirac deltas δ := {δx }x∈R is a Gel’fand distribution basis. In fact, the
Parseval identity holds:
 
| δx |φ |2 dx = |φ(x)|2 dx = φ22 , ∀φ ∈ S(R).
R R

The synthesis operator: Tδ : L2 (R) → L2 (R) is:

 
Tδ ξ |φ = ξ(x) δx |φ dx = ξ(x)φ(x)dx = ξ |φ , ∀φ ∈ S(R).
R R

Then Tδ ξ = ξ for all ξ ∈ L2 (R). Since Tδ is an identity, it is an isometry onto

L2 (R).

5.2 Riesz Distribution Basis

Proposition 5.4 and (5.1) suggest a more general class of bases that will play the
same role as Riesz bases in the ordinary Hilbert space framework.
Definition 5.5 Let D ⊂ H ⊂ D× be a rigged Hilbert space. A weakly measurable
map ω : x ∈ X → ωx ∈ D× is a Riesz distribution basis if ω is a μ-independent
distribution frame.
One has the following:
Proposition 5.6 ([29, Proposition 3.19]) Let D ⊂ H ⊂ D× be a rigged Hilbert
space and let ω : x ∈ X → ωx ∈ D× be a Bessel distribution map. Then the
following statements are equivalent:
(a) ω is a Riesz distribution basis;
(b) If ζ is a Gel’fand distribution basis, then the operator W defined, for f ∈ H,
by:
 
f = ξf (x)ζx dμ → Wf = ξf (x)ωx dμ
X X

is continuous and has bounded inverse;

(c) the synthesis operator Tω is a topological isomorphism of L2 (X, μ) onto H.
Proposition 5.7 If ω is a Riesz distribution basis then ω possesses a unique dual
frame θ which is also a Riesz distribution basis.
644 F. Tschinke

Example Let us consider f ∈ C ∞ (R): 0 < m < |f (x)| < M. Let us define
ωx := f (x)δx : then {ωx }x∈R is a distribution frame, in fact:
 
| ωx |φ |2 dx = |f (x)φ(x)|2 dx ≤ M 2 φ22 , ∀φ ∈ S(R),
R R

and

2
m φ22 ≤ |f (x)φ(x)|2 dx ≤ M 2 φ22 , ∀φ ∈ S(R).
R

Furthermore, {ωx }x∈R is μ-independent. In fact, putting:


ξ(x) ωx |g dx = 0, ∀g ∈ S(R),
R

one has:
 
ξ(x) ωx |g dx = ξ(x)f (x) δx |gdx = 0, ∀g ∈ S(R).
R R

Since {δx }x∈R is μ-independent, it follows that ξ(x)f (x) = 0 a.e., then ξ(x) = 0
a.e.. By definition, {ωx }x∈R is a Riesz distribution basis.

6 Concluding Remarks

In a Hilbert space, frames, semi-frames, Bessel, Riesz-Fischer sequences, and

Riesz bases are related through the action of a linear operator on elements of an
orthonormal basis (see also [29, Remark 3.22]). On the other hand, in literature
some studies on bounds (upper and lower) of these sequences have been already
considered and their links with the linear operators related to them have been studied
(see [3, 6, 9]). For that, it is desirable to continue an analogous study in rigged
Hilbert spaces by considering linear operators in L(D, D× ).

Acknowledgement The author would like to thank R. Corso for some valuable comments and
suggestions.

References

1. S.T. Ali, J.P. Antoine, J.P. Gazeau, Continuous frames in Hilbert spaces. Ann. Phys. 222, 1–37
(1993)
2. S.T. Ali, J.P. Antoine, J.P. Gazeau, Coherent States, Wavelets and Their Generalizations, 2nd
edn. (Springer, Berlin, 2014)
Riesz-Fischer Maps, Semi-frames and Frames in Rigged Hilbert Spaces 645

3. J.-P. Antoine, P. Balazs, Frames and semi-frames. J. Phys. A Math. Theor. 44, 205201 (2011)
4. J.-P. Antoine, C. Trapani, Reproducing pairs of measurable functions and partial inner product
spaces. Adv. Oper. Theory 2, 126–146 (2017)
5. J.-P. Antoine, A. Inoue, C. Trapani, Partial *-Algebras and Their Operator Realizations
(Kluwer, Dordrecht, 2002)
6. P. Balazs, D.T. Stoeva, J.-P. Antoine, Classification of general sequences by frame-related
operators. Sampl. Theory Signal Image Process. 10, 151–170 (2011)
7. P. Balasz, M. Speckbacher, Frames, their relatives and reproducing kernel Hilbert spaces. J.
Phys. A Math. Theor. 53, 015204 (2020)
8. G. Bellomonte, C. Trapani, Riesz-like bases in rigged Hilbert spaces. Zeitschr. Anal. Anwen.
35, 243–265 (2016)
9. P. Casazza, O. Christensen, S. Li, A. Lindner, Riesz-Fischer sequences and lower frame
bounds. Zeitschr. Anal. Anwen. 21, 305–314 (2002)
10. O. Christensen, Frames and Bases: An Introductory Course (Birkhäuser, Boston, 2008)
11. O. Christensen, An Introduction to Frames and Riesz Bases (Birkhaäuser, Boston, 2016)
12. E. Cordero, H. Feichtinger, F. Luef, Banach Gelfand triples for Gabor analysis, in Pseudod-
ifferential Operators. Lecture Notes in Mathematics, vol. 1949 (Springer, Berlin, 2008), pp.
1–33
13. G. de Barra, Measure Theory and Integration (New Age International (P) limited Publishers,
Darya Ganj, 1981)
14. H.G. Feichtinger, K. Gröchenig, Gabor frames and time-frequency analysis of Distributions. J.
Funct. Anal. 146, 464–495 (1997)
15. H.G. Feichtinger, G. Zimmermann, A Banach space of test functions for Gabor analysis, in
Gabor Analysis and Algorithms: Theory and Applications (Birkhäuser, Boston, 1998)
16. I.M. Gel’fand, N.Ya. Vilenkin, Generalized Functions, vol. IV (Academic, New York, 1964)
17. I.M. Gel’fand, G.E. Shilov, E. Saletan, Generalized Functions, vol. III (Academic, New York,
1967)
18. G.G. Gould, The spectral representation of normal operators on a rigged Hilbert space. J.
London Math. Soc. 43, 745–754 (1968)
19. K. Gröchenig, Foundations of Time-Frequency Analysis (Birkhäuser, Basel, 2001)
20. C. Heil, A Basis Theory Primer. Expanded Edition (Birkhäuser/Springer, New York, 2011)
21. J. Horvath, Topological Vector Spaces and Distributions (Addison-Wesley, Boston, 1966)
22. H. Hosseini Giv, M. Radjabalipour, On the structure and properties of lower bounded analytic
frames. Iran. J. Sci. Technol. 37, 227–230 (2013)
23. M.S. Jakobsen, J. Lemvig, Density and duality theorems for regular Gabor frames. J. Funct.
Anal. 270, 229–263 (2016)
24. G. Kaiser, A Friendly Guide to Wavelets (Birkhäuser, Boston, 1994)
25. G. Kyriazis, P. Petrushev, On the construction of frames for spaces of distributions. J. Funct.
Anal. 257, 2159–2187 (2009)
26. S. Pilipovic, D.T. Stoeva, Fréchet frames, general definition and expansion. Anal. Appl. 12,
195–208 (2014)
27. M. Reed, B. Simon, Methods of Modern Mathematical Physics, vols. I and II (Academic, New
York, 1980)
28. W. Rudin, Real and Complex analysis (McGraw-Hill, New York, 1987)
29. C. Trapani, S. Triolo, F. Tschinke, Distribution frames and bases. J. Fourier Anal. and Appl.
25, 2109–2140 (2019)
30. R.M. Young, An Introduction to Nonharmonic Fourier Series, 2nd edn. (Academic, Cambridge,
2001)
Periodic Coherent States Decomposition
and Quantum Dynamics on the Flat
Torus

Lorenzo Zanelli

Abstract We provide a result on the coherent states decomposition for functions

in L2 (Tn ) where Tn := (R/2πZ)n . We study such a decomposition with respect to
the quantum dynamics related to semiclassical elliptic Pseudodifferential operators,
and we prove a related invariance result.

Keywords Coherent states · Toroidal Pdo · Quantum dynamics

Mathematics Subject Classification (2010) 81R30, 58J40, 58C40

1 Introduction

Let us introduce the usual class of semiclassical coherent states on Rn :

|x−y|2
φ(x,ξ )(y) := αh e h (x−y)·ξ e−
i
2h , (x, ξ ) ∈ R2n , y ∈ Rn , 0 < h ≤ 1 (1.1)

with the L2 (Rn )—normalization constant αh := 2− 2 (πh)−3n/4 , and where h is a

n

‘semiclassical parameter’. For any ψ ∈ S (Rn ) the coherent state decomposition

reads, in the distributional sense, as
  
ψ(x0 ) = 0
φ(x,ξ (x
) 0 ) φ(x,ξ ) (y)ψ(y) dy dx dξ (1.2)
R2n Rn

as shown for example in [10, Proposition 3.1.6].

We now observe that for the flat torus Tn := (R/2πZ)n the well known inclusion
L (Tn ) ⊂ S (Rn ) implies that distributional equality (1.2) makes sense also for
2

functions in L2 (Tn ).

L. Zanelli ()
Department of Mathematics “Tullio Levi-Civita”, University of Padova, Padova, Italy
e-mail: [email protected]

© Springer Nature Switzerland AG 2021 647

M. A. Bastos et al. (eds.), Operator Theory, Functional Analysis and Applications,
Operator Theory: Advances and Applications 282,
https://doi.org/10.1007/978-3-030-51945-2_30
648 L. Zanelli

The first aim of our paper is to prove the decomposition of any ϕ ∈ L2 (Tn ) with
respect to the family of periodic coherent states  given by the periodization of
(1.1). In view of this target, we recall that the periodization operator

(φ)(y) := φ(y − 2πk)
k∈Zn

maps S(Rn ) into C ∞ (Tn ), as is shown for example in [14, Theorem 6.2]. Thus, we
can define for all 0 < h ≤ 1,

(x,ξ ) (y) := φ(x,ξ )(y − 2πk), (x, ξ ) ∈ Tn × h Zn , y ∈ Tn . (1.3)
k∈Zn

Notice that the family of coherent states in (1.3) is well posed also for ξ ∈ Rn
and the related phase space is Tn × Rn . However, our target is to show that the
decomposition of periodic functions can be done with respect to the minimal set of
coherent states in (1.3) for ξ ∈ h Zn ⊂ Rn . Furthermore, we notice that the phase
space Tn × h Zn is necessary in order to deal with a well defined setting of toroidal
Weyl operators acting on L2 (Tn ) and more in general with semiclassical toroidal
Pseudodifferential operators (see Sect. 2).
The first result of the paper is the following.
Theorem 1.1 Let ϕh ∈ C ∞ (Tn ) be such that !x ϕh L2 ≤ c h−M for some c > 0,
M ∈ N, ϕh L2 = 1 with 0 < h ≤ 1, h−1 ∈ N and let (x,ξ ) be as in (1.3). Then

 
ϕh = (x,ξ ) , ϕh L2 0(x,ξ ) dx + OL2 (h∞ ). (1.4)
ξ ∈h Zn Tn

Moreover, there exists f (h) > 0 depending on ϕh such that

 
ϕh = (x,ξ ), ϕh L2 0(x,ξ ) dx + OL2 (h∞ ). (1.5)
ξ ∈h Zn , |ξ |≤f (h) Tn

The following inclusion involving the set of frequencies ξ ∈ h Zn ⊂ Rn allows

to consider decomposition (1.4) minimal with respect to (1.2). The above result
shows also that the sum over the frequencies can be taken in the bounded region
|ξ | ≤ f (h), i.e. we can consider a finite sum by taking into account an O(h∞ )
remainder in L2 (Tn ).
An analogous result of (1.4) in the two dimensional setting is shown in [4,
Proposition 60] by the use of a different periodization operator. Same construction
of coherent states as in [4] for T2 is used in [2, 6] for the study of quantum cat maps
and equipartition of the eigenfunctions of quantized ergodic maps. In the paper
[8], covariant integral quantization using coherent states for semi-direct product
groups is implemented for the motion of a particle on the circle and in particular
the resolution of the identity formula is proved. Another class of coherent states
Periodic Coherent States Decomposition and Quantum Dynamics on the Flat Torus 649

on the torus are defined also in [7], with a related resolution of the identity, in the
understanding of the Quantum Hall effect. We also recall [5] where coherent states
and Bargmann Transform are studied on L2 (Sn ). The literature on coherent states
are quite rich, and thus we address the reader to [1].
We now devote our attention to the periodic coherent states decomposition for
eigenfunctions of elliptic semiclassical toroidal Pseudodifferential operators (see
Sect. 2). We will see that the formula (1.4) can be reduced in view of a phase-space
localization of eigenfunctions.
This is the content of the second main result of the paper.
Theorem 1.2 Let Oph (b) be an elliptic semiclassical "do as in (2.1) and h−1 ∈ N.
Let E ∈ R, and let ψh ∈ C ∞ (Tn ) be such that ψh L2 = 1 and !x ψh L2 ≤
c h−M , and which is eigenfunction of the eigenvalue problem on Tn given by

Oph (b)ψh = Eh ψh

where Eh ≤ E for any 0 < h ≤ 1. Then, there exists g(h, E) ∈ R+ such that
 
ψh = (x,ξ ), ψh L2 0(x,ξ ) dx + OL2 (h∞ ). (1.6)
ξ ∈h Zn , |ξ |≤g(h,E) Tn

We notice that for the operators −h2 !x +V (x) all the eigenfunctions with eigen-
values Eh ≤ E fulfill !x ψh L2 ≤ c h−2 . In particular, we have the asymptotics
g(h, E) → +∞ as h → 0+ . We also underline that the function g(E, h) and the
estimate on remainder OL2 (h∞ ) do not depend on the particular choice of ψh . This
implies that all these eigenfunctions take the form
(1.6) and therefore also any finite
linear combination of eigenfunctions of kind 1≤α≤N cα ψh,α where |cα | ≤ 1.
We remind that Weyl Law on the number N (h) of eigenvalues Eh,α ≤ E (with
their multiplicity) for semiclassical elliptic operators (see for example [9]) reads
N (h) 2 (2πh)−n (vol(U (E)) + O(1)).
The proof of the above result is mainly based on a uniform estimate for our
toroidal version of the Fourier-Bros-Iagolnitzer (FBI) transform

(T ψh )(x, ξ ) := (x,ξ ), ψh L2

on the unbounded region given by all x ∈ Tn and ξ ∈ hZn such that |ξ | > g(h, E).
The FBI transform on any compact manifold has already been defined and studied
in the literature, see for example [16].
We remind that, in the euclidean setting of R2n , the function

Th (ψh )(x, ξ ) := φ(x,ξ ), ψh L2 (Rn )

is the usual version of the FBI transform, which is well posed for any ψh ∈ S (Rn ).
This is used to study the phase space localization by the Microsupport of ψh (see for
example [10]), namely MS(ψh ) the complement of the set of points (x0 , ξ0 ) such
650 L. Zanelli

that Th (ψh )(x, ξ ) 2 O(e−δ/ h ) uniformly in a neighborhood of (x0 , ξ0 ). In the case

of the weaker estimate Th (ψh )(x, ξ ) 2 O(h∞ ) one can define the semiclassical
Wave Front Set W F (ψh ). Is well known (see [10]) that the Microsupport (or the
semiclassical Wave Front Set) of eigenfunctions for elliptic operators is localized in
the sublevel sets

U (E) := {(x, ξ ) ∈ R2n | b(x, ξ ) ≤ E},

i.e. MS(ψh ) ⊆ U (E). The well posedness of W F (ψh ) and MS(ψh ) in the periodic
setting can be seen starting from the euclidean setting and thanks to distributional
inclusion L2 (Tn ) ⊂ S (Rn ), (see for example section 3.1 of [3]). The semiclassical
study in the phase space for eigenfunctions in the periodic setting has also been
studied in [18] with respect to weak KAM theory.
In our Theorem 1.2 we are interested to show another kind of semiclassical
localization, namely to localize the bounded region

(E, h) := {(x, ξ ) ∈ Tn × Rn | x ∈ Tn , |ξ | ≤ g(E, h)}

which will be bigger than MS(ψh ), h—dependent and such that the coherent state
decomposition of ψh can be done up to a remainder OL2 (h∞ ).
We now focus our attention to the decompositon (1.6) under the time evolution.
Theorem 1.3 Let ϕh ∈ C ∞ (Tn ), L2 —normalized such that

ϕh = cj ψh,j (1.7)
1≤j ≤J (h)

where ψh,j are given in Theorem 1.2 and J (h) ≤ J0 h−Q for some J0 , Q > 0. Let
Oph (b) be an elliptic semiclassical "do as in (2.1) and

Uh (t) := exp{(−iOph (b)t)/ h}.

Then, there exists (h) > 0 such that for any t ∈ R,

 
Uh (t)ϕh = (x,ξ ), Uh (t)ϕh L2 0(x,ξ ) dx + OL2 (h∞ ). (1.8)
ξ ∈h Zn , |ξ |≤(h) Tn

The equality (1.8) shows that time evolution under the L2 —unitary map Uh (t)
does not change such a decomposition, since (h) does not depend on time.
The function (h) is not necessarily the same as the function f (h) contained in
Theorem 1.1 but we have that (h) ≥ f (h). In other words, this quantum dynamics
preserves the coherent state decomposition (1.5). The same result holds for any
eigenfunctions in Theorem 1.2 since in this case Uh (t)ψh = exp{(−iEh t)/ h}ψh .
Notice that here we can assume that Q > n, namely the linear combination
Periodic Coherent States Decomposition and Quantum Dynamics on the Flat Torus 651

(1.7) can be done with more eigenfunctions than the ones that have eigenvalues
Eh ≤ E with fixed energy E > 0. Notice also that we have (x,ξ ) , Uh (t)ϕh L2 =
Uh (−t)(x,ξ ), ϕh L2 for any t ∈ R and that the time evolution of the periodization
of coherent states has been used in [17] in the context of optimal transport theory.

2 Semiclassical Toroidal Pseudodifferential Operators

Let us define the flat torus Tn := (R/2πZ)n and introduce the class of symbols b ∈
m
Sρ,δ (Tn × Rn ), m ∈ R, 0 ≤ δ, ρ ≤ 1, given by functions in C ∞ (Tn × Rn ; R) which
are 2π-periodic in each variable xj , 1 ≤ j ≤ n and for which for all α, β ∈ Zn+
there exists Cαβ > 0 such that for all (x, ξ ) ∈ Tn × Rn ,

|∂xβ ∂ξα b(x, ξ )| ≤ Cαβm ξ m−ρ|α|+δ|β|

where ξ  := (1+|ξ |2)1/2 . In particular, the set S1,0

m
(Tn ×Rn ) is denoted by S m (Tn ×
R ).
n

We introduce the semiclassical toroidal Pseudodifferential Operators by the

following.
Definition 2.1 Let ψ ∈ C ∞ (Tn ; C) and 0 < h ≤ 1,

Oph (b)ψ(x) := (2π)−n eix−y,κ b(x, hκ)ψ(y) dy. (2.1)
κ∈Zn T
n

This is the semiclassical version (see [12, 13]) of the quantization by Pseudod-
ifferential Operators on the torus developed in [14] and [15]. See also [11] for the
notion of vector valued Pseudodifferential Operators on the torus.
We now notice that we have a map Oph (b) : C ∞ (Tn ) −→ D (Tn ). Indeed,
remind that u ∈ D (Tn ) are the linear maps  u : C ∞ (Tn ) −→ C such that there
exist C > 0 and k ∈ N, for which |u(φ)| ≤ C |α|≤k ∂xα φ∞ for all φ ∈ C ∞ (Tn ).
Given a symbol b ∈ S m (Tn × Rn ), the toroidal Weyl quantization reads (see
[12, 13])
  h 
−n
Oph (b)ψ(x) := (2π)
w
eix−y,κ b y, κ ψ(2y − x) dy.
n T
n 2
κ∈Z

In particular, it holds

h (b)ψ(x) = (Oph (σ ) ◦ Tx ψ)(x)

Opw
652 L. Zanelli

where Tx : C ∞ (Tn ) → C ∞ (Tn ) defined as (Tx ψ)(y) := ψ(2y − x) is linear,

invertible and L2 -norm preserving, and σ is a suitable toroidal symbol related to b,
i.e.
 1
σ ∼ Aα D (α) b(y, hξ/2) ,
α! ξ y y=x
α≥0

where Aξj f (ξ + ej ) − f (ξ ) is the difference operator (see [14, Theorem 4.2]).

The typical example is given by
 1 
Oph (H ) = − h2 !x + V (x) ψ(x)
2
 1 
−n
= (2π) eix−y,κ |hκ|2 + V (x) ψ(y) dy
n T
n 2
κ∈Z

namely the related symbol is the mechanical type Hamiltonian H (x, ξ ) = 12 |ξ |2 +

V (x). Also in the case of the Weyl operators we have
1
− h2 !x + V (x) = Opw
h (H )
2
for the same symbol (see for example [13]).
In our paper we are interested in uniform elliptic operators, namely such that the
symbol b ∈ S m (Tn × Rn ) fulfills for some constants C, c > 0 the lower bound

|b(x, ξ )| ≥ C ξ m

for any x ∈ Tn and |ξ | ≥ c. This property guarantees bounded sublevels sets for b
and discrete spectrum for the operator Oph (b) for any fixed 0 < h ≤ 1. As we see
in Theorem 1.2, this assumption permits also to prove the semiclassical localization
of all the eigenfunctions within these sublevels sets, and this localization can be
studied by our semiclassical coherent states (1.3).

3 Proofs of the Main Results

3.1 Proof of Theorem 1.1


We remind that (x,ξ ) (y) := (φ(x,ξ ))(y) and (φ)(y) := k∈Zn φ(y − 2πk).
Thus,
 
(x+2πβ,ξ )(y) = φ(x+2πβ,ξ )(y − 2πk) = φ(x,ξ )(y − 2πk − 2πβ)
k∈Zn k∈Zn

= (x,ξ )(y).
Periodic Coherent States Decomposition and Quantum Dynamics on the Flat Torus 653

We mainly adapt, in our toroidal setting, the proof of [10, Proposition 3.1.6] written
for the euclidean setting. Thus, we define the operator T 0 on functions " ∈ L2 (Tn ×
hZn ) as
 
(T ")(y) :=
0
"(x, ξ )0(x,ξ )(y) dx.
ξ ∈h Zn Tn

It can be easily seen that T 0 equals the adjoint of the operator (T ψ)(x, ξ ) :=
(x,ξ ), ψL2 (Tn ) , i.e.

T 0 ", ψL2 (Tn ) = ", T ψL2 (Tn ×hZn ) .

Thus, for all ψ1 , ψ2 ∈ C ∞ (Tn ) ⊂ L2 (Tn ) we have

T 0 ◦ T ψ1 , ψ2 L2 (Tn ) = T ψ1 , T ψ2 L2 (Tn ×hZn ) .

It remains to prove that

T ψ1 , T ψ2 L2 (Tn ×hZn ) = ψ1 , ψ2 L2 (Tn ) + O(h∞ ) (3.1)

which implies

T 0 ◦ T = Id mod O(h∞ ) (3.2)

on L2 (Tn ), and equality (3.2) is exactly the statement (1.4).

In order to prove (3.1), we recall that the periodization operator  can be
rewritten in the form (see [14, Theorem 6.2]):
 
(φ) = FT−1
n FR n φ . (3.3)
Zn

where FT−1
n stands for the inverse toroidal Fourier Transform, and FRn is the usual
euclidean version. In view of (3.3) it follows

(T ψ)(x, ξ ) := (x,ξ ), ψL2 (Tn ) = FRn φx,ξ |Zn , FTn ψL2 (Zn )

=  (k),
φx,ξ (k)0 ψ
k∈Zn

x,ξ (k) := FRn φx,ξ (k) and ψ

where φ (k) := FTn ψ(k). Thus,

T ψ1 , T ψ2 L2 (Tn ×hZn )

   0   
=  (k)
φx,ξ (k)0 ψ  (μ) dx.
φx,ξ (μ)0 ψ
ξ ∈h Zn T
n
k∈Zn μ∈Zn
654 L. Zanelli

We can rewrite this equality, in the distributional sense, as

  
T ψ1 , T ψ2 L2 (Tn ×hZn ) = 1 (k)0 ψ
ψ 2 (μ) x,ξ (k)
φ φx,ξ (μ)0 dx
k,μ∈Zn ξ ∈h Zn Q

where Q := [0, 2π]n and ψ1 , ψ2 ∈ C ∞ (Tn ). Now let ξ = hα with α ∈ Zn , so that

 
T ψ1 , T ψ2 L2 (Tn ×hZn ) = 1 (k)0 ψ
ψ 2 (μ) φx,hα (k)
 φx,hα (μ)0 dx
k,μ∈Zn α∈Zn Q

By using the explicit form of 

φx,hα and the condition h−1 ∈ N, a direct computation
shows that
  
T ψ1 , T ψ2 L2 (Tn ) = 1 (k)0 ψ
ψ 2 (μ) eiα(k−μ) + O(h∞ ) (3.4)
k,μ∈Zn α∈Zn

where O(h∞ ) does not depend on the functions ψ1 , ψ2 . We now use the assumption
!x ψL2 ≤ c h−M for some fixed c, M > 0 so that Fourier components fulfill
k | ≤ |k|−2 (2π)n/2 c h−M , and ψL2 = 1 gives |ψ
|ψ 0 | ≤ (2π)n/2 . Consequently,
 
1 (k)| ≤ (2π)n/2 + (2π)n/2 c
|ψ |k|−2 h−M , (3.5)
k∈Zn k∈Zn \{0}

and
  
1 (k)0 ψ
ψ 2 (μ) ≤ 1 (k)|
|ψ 2 (μ)|.
|ψ (3.6)
k,μ∈Zn k∈Zn μ∈Zn

To conclude, since

δ(k − μ) = eiα(k−μ),
α∈Zn

we get

T ψ1 , T ψ2 L2 (Tn ×hZn ) = 1 (k)0 ψ
ψ 2 (k) + O(h∞ )
k∈Zn

= ψ1 , ψ2 L2 (Tn ) + O(h∞ ). (3.7)

The estimates (3.5)–(3.6) together with (3.4) ensure that the remainder in (3.7) has
order O(h∞ ).
Periodic Coherent States Decomposition and Quantum Dynamics on the Flat Torus 655

In order to prove (1.5), we observe that

 
ϕh = (x,ξ ), ϕh L2 0(x,ξ ) dx + O(h∞ ).
ξ ∈h Zn Tn

is given by an L2 -convergent series. Thus, for any fixed ϕh we can say that there
exists f (h) > 0 such that
 
ϕh = (x,ξ ) , ϕh L2 0(x,ξ ) dx + O(h∞ ).
ξ ∈h Zn ,|ξ |<f (h) T
n

3.2 Proof of Theorem 1.2

We apply the statement of Theorem 1.1, for a set of linearly independent eigen-
functions ψh,i generating all the eigenspaces linked to eigenvalues Eh ≤ E and
fi (h) > 0 given by Theorem 1.1:
 
ψh,i = (x,ξ ) , ψh,i L2 0(x,ξ ) dx + Rh,i
ξ ∈h Zn ,|ξ |<f Tn
i (h)

where Rh,i L2 = O(h∞ ).

Moreover, we recall that the Weyl Law on the number N (h) of eigenvalues
Eh ≤ E (counted with their multiplicity) for semiclassical elliptic operators (see
for example [9]) reads

N (E, h) 2 (2πh)−n (vol(U (E)) + O(1)).

We define:

g(E, h) := max fi (h).

1≤i≤N (E,h)

Since
 any eigenfunction ψh linked to Eh ≤ E will be written as ψh =
i ψh,i , ψh ψh,i then the linearity of decomposition (1.4) ensures also the decom-
position (1.6) for such ψh . Namely,
 
ψh = (x,ξ ) , ψh L2 0(x,ξ ) dx + Rh
ξ ∈h Zn , |ξ |≤g(h,E) Tn
656 L. Zanelli


where Rh = 1≤i≤N (E,h) Ri,h . To conclude:

Rh L2 ≤ Ri,h L2 ≤ N (E, h) max Ri,h L2
1≤i≤N (E,h)
1≤i≤N (E,h)

= N (E, h) · O(h∞ ) = O(h∞ ).

3.3 Proof of Theorem 1.3

We assume that ϕh ∈ C ∞ (Tn ) is L2 -normalized and


ϕh = cj ψh,j
1≤j ≤J (h)

where the L2 -normalized eigenfunctions ψh,j of Oph (b) are given in Theorem 1.2
and we assume J (h) ≤ J0 h−Q for some J0 , Q > 0 that are independent on 0 <
h ≤ 1.
Define

(h) := max fj (h)

1≤j ≤J (h)

where fi (h) are associated to the functions ψh,i and given by Theorem 1.1.
We now observe that if Uh (t) := exp{(−iOph (b)t)/ h} then

cj e− h Ej,h ψh,j
i
U (t)ϕh =
1≤j ≤J (h)

for any t ∈ R.
We can now apply the decomposition formula (1.4) with the condition on the
frequencies |ξ | ≤ (h) and for the wave function U (t)ϕh and get the expected
result, namely
  
Uh (t)ϕh = (x,ξ ), Uh (t)ϕh L2 0(x,ξ ) dx + Rj,h
ξ ∈h Zn , |ξ |≤(h) Tn 1≤j ≤J

for any t ∈ R. The remainder Rh := 1≤j ≤J Rj,h can be estimated as in the
previous Theorem, namely

Rh L2 ≤ Rj,h L2 ≤ J0 h−Q max Rj,h L2
1≤j ≤J
1≤i≤J

= J0 h−Q · O(h∞ ) = O(h∞ ).

Periodic Coherent States Decomposition and Quantum Dynamics on the Flat Torus 657

References

1. J.P. Antoine, F. Bagarello, J.P. Gazeau (eds.), Coherent States and Their Applications. A
Contemporary Panorama. Springer Proceedings in Physics, vol. 205 (Springer, Berlin, 2018)
2. A. Bouzouina, S. De Bievre, Equipartition of the eigenfunctions of quantized ergodic maps on
the torus. Commun. Math. Phys. 178, 83–105 (1996)
3. F. Cardin, L. Zanelli, The geometry of the semiclassical wave front set for Schrödinger
eigenfunctions on the torus. Math. Phys. Anal. Geom. 20(2), Art. 10, 20 (2017)
4. M. Combescure, D. Robert, Coherent States and Applications in Mathematical Physics
(Springer, Berlin, 2012)
5. E. Diaz-Ortiz, C. Villegas-Blas, On a Bargmann transform and coherent states for the n-sphere.
J. Math. Phys. 53(6), 062103, 25 (2012)
6. F. Faure, S. Nonnenmacher, S. De Bievre, Scarred eigenstates for quantum cat maps of minimal
periods. Commun. Math. Phys. 239, 449–492 (2003)
7. M. Fremling, Coherent state wave functions on a torus with a constant magnetic field. J. Phys.
A 46(27), 275302, 23 (2013)
8. R. Fresneda, J.P. Gazeau, D. Noguera, Quantum localisation on the circle. J. Math. Phys. 59(5),
052105, 19 (2018)
9. V. Guillemin, S. Sternberg, Semi-Classical Analysis (International Press, Boston, 2013)
10. A. Martinez, Introduction to Semiclassical and Microlocal Analysis (Springer, New York,
2002)
11. B.B. Martinez, R. Denk, J.H. Monzón, T. Nau, Generation of semigroups for vector-valued
pseudodifferential operators on the torus. J. Fourier Anal. Appl. 22, 823–853 (2016)
12. A. Parmeggiani, L. Zanelli, Wigner measures supported on weak KAM tori. J. Anal. Math.
123, 107–137 (2014)
13. T. Paul, L. Zanelli, On the dynamics of WKB wave functions whose phase are weak KAM
solutions of H-J equation. J. Fourier Anal. Appl. 20, 1291–1327 (2014)
14. M. Ruzhansky, V. Turunen, Quantization of pseudo-differential operators on the torus. J.
Fourier Anal. Appl. 16, 943–982 (2010)
15. M. Ruzhansky, V. Turunen, Pseudo-Differential Operators and Symmetries: Background
Analysis and Advanced Topics (Birkhäuser, Basel, 2010)
16. J. Wunsch, M. Zworski, The FBI transform on compact C ∞ -manifolds. Trans. Amer. Math.
Soc. 353, 1151–1167 (2001)
17. L. Zanelli, On the optimal transport of semiclassical measures. Appl. Math. Optim. 74, 325–
342 (2016)
18. L. Zanelli, Schrödinger spectra and the effective Hamiltonian of the weak KAM theory on the
flat torus. J. Math. Phys. 57(8), 081507, 12 (2016)