See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/226621630

Lecture Notes in Mathematics

Chapter · January 2008

DOI: 10.1007/978-3-540-78547-7_5

CITATIONS READS

31 9,794

7 authors, including:

Ana Carpio Maria-Luisa Rapun

Complutense University of Madrid Universidad Politécnica de Madrid
160 PUBLICATIONS 2,012 CITATIONS 65 PUBLICATIONS 851 CITATIONS

SEE PROFILE SEE PROFILE

Oliver Dorn Miguel Moscoso

The University of Manchester University Carlos III de Madrid
110 PUBLICATIONS 2,397 CITATIONS 117 PUBLICATIONS 2,470 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Maria-Luisa Rapun on 13 March 2014.

The user has requested enhancement of the downloaded file.

Lecture Notes in Mathematics 1943
Editors:
J.-M. Morel, Cachan
F. Takens, Groningen
B. Teissier, Paris
C.I.M.E. means Centro Internazionale Matematico Estivo, that is, International Mathematical Summer
Center. Conceived in the early fifties, it was born in 1954 and made welcome by the world mathematical
community where it remains in good health and spirit. Many mathematicians from all over the world
have been involved in a way or another in C.I.M.E.’s activities during the past years.

So they already know what the C.I.M.E. is all about. For the benefit of future potential users and co-
operators the main purposes and the functioning of the Centre may be summarized as follows: every
year, during the summer, Sessions (three or four as a rule) on different themes from pure and applied
mathematics are offered by application to mathematicians from all countries. Each session is generally
based on three or four main courses (24−30 hours over a period of 6-8 working days) held from
specialists of international renown, plus a certain number of seminars.

A C.I.M.E. Session, therefore, is neither a Symposium, nor just a School, but maybe a blend of both.
The aim is that of bringing to the attention of younger researchers the origins, later developments, and
perspectives of some branch of live mathematics.

The topics of the courses are generally of international resonance and the participation of the courses
cover the expertise of different countries and continents. Such combination, gave an excellent opportu-
nity to young participants to be acquainted with the most advance research in the topics of the courses
and the possibility of an interchange with the world famous specialists. The full immersion atmosphere
of the courses and the daily exchange among participants are a first building brick in the edifice of
international collaboration in mathematical research.

C.I.M.E. Director C.I.M.E. Secretary

Pietro ZECCA Elvira MASCOLO
Dipartimento di Energetica “S. Stecco” Dipartimento di Matematica
Università di Firenze Università di Firenze
Via S. Marta, 3 viale G.B. Morgagni 67/A
50139 Florence 50134 Florence
Italy Italy
e-mail: [email protected] e-mail: [email protected]

For more information see CIME’s homepage: http://www.cime.unifi.it

CIME’s activity is supported by:

– Ministero degli Affari Esteri, Direzione Generale per la Promozione e la

Cooperazione, Ufficio V
– Consiglio Nazionale delle Ricerche
– E.U. under the Training and Mobility of Researchers Programme
Luis L. Bonilla (Ed.)

Inverse Problems
and Imaging

Lectures given at the

C.I.M.E. Summer School
held in Martina Franca, Italy
September 15–21, 2002

Chapters by:
A. Carpio · O. Dorn · M. Moscoso · F. Natterer
G.C. Papanicolaou · M.L. Rapún · A. Teta

ABC
Luis L. Bonilla (Ed.) George C. Papanicolaou
Gregorio Millán Institute of Fluid Dynamics Mathematics Department
Nanoscience and Industrial Mathematics Building 380, 383V
Universidad Carlos III de Madrid Stanford University
Avda. de la Universidad 30 Stanford, CA 94305, USA
28911 Leganés, Madrid, Spain [email protected]
[email protected]
http://scala.uc3m.es Maria Luisa Rapún
Departamento de Fundamentos
Ana Carpio Matemáticos de la Tecnología Aeronáutica
Departamento de Matemática Aplicada Escuela Técnica Superior de Ingenieros
Facultad de Matemáticas Aeronáuticos, Universidad Politécnica
Universidad Complutense de Madrid de Madrid, Plaza del Cardenal Cisneros 3
Plaza de Ciencias 3 28040 Madrid, Spain
28040 Madrid, Spain [email protected]
[email protected]
www.mat.ucm.es/∼acarpio Alessandro Teta
Dipartimento di Matematica Pura
Oliver Dorn e Applicata, Università di L’Aquila
Miguel Moscoso via Vetoio – loc. Coppito
Gregorio Millán Institute of Fluid Dynamics 67100 L’Aquila, Italy
Nanoscience and Industrial Mathematics [email protected]
Universidad Carlos III de Madrid
Avda. de la Universidad 30
28911 Leganés, Madrid, Spain
[email protected]
[email protected]

Frank Natterer
Institut für Numerische und Angewandte
Mathematik, University of Münster
Einsteinstraße 62
48149 Münster
Germany
[email protected]

ISBN: 978-3-540-78545-3 e-ISBN: 978-3-540-78547-7

DOI: 10.1007/978-3-540-78547-7
Lecture Notes in Mathematics ISSN print edition: 0075-8434
ISSN electronic edition: 1617-9692
Library of Congress Control Number: 2008922185
Mathematics Subject Classification (2000): 65R32, 44A12, 92C55, 85A25, 35Q40
°c 2008 Springer-Verlag Berlin Heidelberg
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations are
liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, 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.
Cover design: WMX Design GmbH, Heidelberg
Printed on acid-free paper
9 8 7 6 5 4 3 2 1
springer.com
Preface

Nowadays, we are facing numerous important imaging problems, as for exam-

ple, the detection of anti-personal land mines in post-war remediation areas,
detection of unexploded ordnances (UXO), nondestructive testing of materials,
monitoring of industrial processes, enhancement of oil production by efficient
reservoir characterization, and the various exciting and emerging developments
in noninvasive imaging techniques for medical purposes – computerized tomo-
graphy (CT), magnetic resonance imaging (MRI), positron emission tomography
(PET) and ultrasound tomography, to mention only a few. It is broadly recog-
nized that these problems can only be solved by a joint effort of experts in
mathematical and technical sciences.
The CIME Summer School on Imaging, held in Martina Franca, Italy, from
15 to 21 September, 2002, encompassed the theory and applications of imaging in
different disciplines including, medicine, geophysics, engineering, etc. The Sum-
mer School brought together leading experts in mathematical techniques and
applications in many different fields, to present a broad and useful introduction
for non-experts and practitioners alike to many aspects of this exciting field. The
main lecturers were Simon Arridge, Frank Natterer, George C. Papanicolaou and
William Symes. Seminars on related special topics were contributed by Oliver
Dorn, Miguel Moscoso and Alessandro Teta. Among the different topics dealt
with in the school, we may cite X-ray tomography, diffusive optical tomography
with possible applications to tumor detection in medicine by optical means, elec-
tromagnetic induction tomography used in geophysics, and techniques of seismic
tomography to image the wave velocity structure of a region and to obtain infor-
mation on possible oil or gas reservoirs, etc. Furthermore, there were extensive
discussions on the mathematical bases for analyzing these methods. The math-
ematical and computational techniques that are used for imaging have many
common features and form a rapidly developing part of applied mathematics.
The present volume contains a general introduction on image reconstruction
by M. Moscoso, some of the lectures and presentations given in the Summer
School (F. Natterer, O. Dorn et al., M. Moscoso, G. Dell’Antonio et al.), and
two additional lectures on other imaging techniques by A. Carpio and M.L.
Rapún and by O. Dorn.
VI Preface

The lectures by Prof. Frank Natterer introduce the mathematical theory and
the reconstruction algorithms of computerized X-ray tomography. These lectures
give a short account of integral geometry and the Radon transform, reconstruc-
tion algorithms such as the filtered back projection algorithm, iterative methods
(for example, the Kaczmarz method) and Fourier methods. They also comment
on the three-dimensional case, which is the subject of current research. Many of
the fundamental tools and issues of computerized tomography, such as back pro-
jection, sampling, and high frequency analysis, have their counterparts in more
advanced imaging techniques for impedance, optical or ultrasound tomography
and are most easily studied in the framework of computerized tomography.
The chapter by O. Dorn, H. Bertete-Aguirre and G.C. Papanicolaou reviews
electromagnetic induction tomography, used to solve imaging problems in geo-
physical and environmental imaging applications. The focus is on realistic 3D
situations which provide serious computational challenges as well as interesting
novel mathematical problems to the practitioners. The chapter first introduces
the reader to the mathematical formulation of the underlying inverse problem;
it then describes the theory of sensitivity analysis in this application; it proposes
a nonlinear reconstruction algorithm for solving such problems efficiently; it dis-
cusses a regularization technique for stabilizing the reconstruction; and finally it
presents various numerical examples for illustrating the discussed concepts and
ideas.
The chapter by M. Moscoso presents optical imaging of biological tissue using
the polarization effects of a narrow beam of light. The biological tissue is modeled
as a continuous medium which varies randomly in space and which contains
inhomogeneities with no sharp boundaries. This differs from the more usual
point of view in which the biological tissue is modeled as a medium containing
discrete spherical particles of the same or different sizes. The propagation of
light is then described by a vector radiative transport equation which is solved
by a Monte Carlo method. A discussion on how to use polarization to improve
image reconstruction is given as well.
The chapter by A. Carpio and M.L. Rapún explains how to use topological
derivative methods to solve constrained optimization reformulations of inverse
scattering problems. This chapter gives formulas to calculate the topological
derivatives for the Helmholtz equation and for the equations of elastic waves.
Furthermore they explain and implement a practical iterative numerical scheme
to detect objects based on computing the topological derivative of a cost func-
tional associated to these equations in successive approximate domains. Many
examples of reconstruction of objects illustrate this method.
The chapter by O. Dorn deals with an inverse problem in underwater acoustic
and wireless communication. He establishes a link between the time-reversal
and adjoint methods for imaging and proposes a method for solving the inverse
problem based on iterative time-reversal experiments.
Lastly, the chapter by G. Dell’Antonio, R. Figari and A. Teta reviews the
theory of Hamiltonians with point interactions, i.e., with potentials supported on
a finite set of points. This chapter studies the mathematical basis of scattering
 Preface VII

with point scatterers, analyzes how such an idealized situation relates to short-
range potentials and discusses situations in which the strength of the potential
depends on the wave function as it has been proposed in the physics literature
on double barriers and other nanostructures. Knowing the solution of the di-
rect problem is of course a prerequisite to be able to image the scatterers from
measurements on a boundary.

Universidad Carlos III de Madrid Luis L. Bonilla

Contents

Introduction to Image Reconstruction

Miguel Moscoso . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1 Medicine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Geophysics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Industry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4 Image Reconstruction Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.1 Computerized Tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2 Diffuse Optical Tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.3 Other Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

X-ray Tomography
Frank Natterer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Integral Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.1 The Radon Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 The Ray Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 The Cone-Beam Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 The Attenuated Radon Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 Vectorial Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3 Reconstruction Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1 The Filtered Backprojection Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Iterative Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Fourier Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4 Reconstruction from Cone Beam Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
X Contents

Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging

in Isotropic and Anisotropic Media
Oliver Dorn, Hugo Bertete-Aguirre, and George C. Papanicolaou . . . . . . . . . 35
1 Introduction: Electromagnetic Imaging of the Earth . . . . . . . . . . . . . . . . . 35
2 Sensitivities and the Adjoint Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3 Maxwell’s Equations for Anisotropic Media . . . . . . . . . . . . . . . . . . . . . . . . . 41
4 The Linearized Residual Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5 Adjoint Sensitivity Analysis for Anisotropic Media . . . . . . . . . . . . . . . . . . . 45
5.1 The Adjoint Linearized Residual Operator . . . . . . . . . . . . . . . . . . . . . . 45
5.2 Decomposition of the Residual Operator and Its Adjoint . . . . . . . . . 46
5.3 A Special Case: Isotropic Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6 A Single Step Adjoint Field Inversion Scheme . . . . . . . . . . . . . . . . . . . . . . . 48
6.1 Outline of the Inversion Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.2 Efficient Calculation of the Adjoint of the Linearized
Residual Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.3 Iterative Algorithm for the Adjoint Field Inversion . . . . . . . . . . . . . . 52
7 Regularization and Smoothing with Function Spaces . . . . . . . . . . . . . . . . . 53
8 Numerical Examples for Electromagnetic Sensitivity Functions . . . . . . . . 56
8.1 Some Numerically Calculated Sensitivity Functions . . . . . . . . . . . . . . 56
8.2 A Short Discussion of Sensitivity Functions . . . . . . . . . . . . . . . . . . . . . 60
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

Polarization-Based Optical Imaging

Miguel Moscoso . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2 Radiative Transfer Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4 Monte Carlo Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.1 Analog Monte Carlo Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2 Splitting and Russian Roulette . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3 Point Detectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

Topological Derivatives for Shape Reconstruction

Ana Carpio and Maria Luisa Rapún . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2 Helmholtz Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
2.1 The Forward and Inverse Scattering Problems . . . . . . . . . . . . . . . . . . 87
2.2 Topological Derivatives Applied to Inverse Scattering . . . . . . . . . . . . 89
3 Numerical Computation of Topological Derivatives . . . . . . . . . . . . . . . . . . 92
4 An Iterative Method Based on Topological Derivatives . . . . . . . . . . . . . . . 101
5 Numerical Solution of Forward and Adjoint Problems . . . . . . . . . . . . . . . . 107
6 Explicit Expressions for the Topological Derivatives . . . . . . . . . . . . . . . . . . 112
6.1 Shape Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.2 Proof of Theorems 2.1 and 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
 Contents XI

7 Sounding Solids by Elastic Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

7.1 The Forward and Inverse Scattering Problems . . . . . . . . . . . . . . . . . . 122
7.2 Shape Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.3 Topological Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

Time-Reversal and the Adjoint Imaging Method

with an Application in Telecommunication
Oliver Dorn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
2 Time-Reversal and Communication in the Ocean . . . . . . . . . . . . . . . . . . . . 137
3 The MIMO Setup in Wireless Communication . . . . . . . . . . . . . . . . . . . . . . 139
4 Symmetric Hyperbolic Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5 Examples for Symmetric Hyperbolic Systems . . . . . . . . . . . . . . . . . . . . . . . 141
5.1 Linearized Acoustic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.2 Maxwell’s Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.3 Elastic Waves Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6 The Basic Spaces and Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
7 An Inverse Problem Arising in Communication . . . . . . . . . . . . . . . . . . . . . . 146
8 The Basic Approach for Solving the Inverse Problem
of Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
9 The Adjoint Operator A∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
10 The Acoustic Time-Reversal Mirror . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
11 The Electromagnetic Time-Reversal Mirror . . . . . . . . . . . . . . . . . . . . . . . . . 152
12 Time-Reversal and the Adjoint Operator A∗ . . . . . . . . . . . . . . . . . . . . . . . . 155
13 Iterative Time-Reversal for the Gradient Method . . . . . . . . . . . . . . . . . . . . 156
13.1 The Basic Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
13.2 The Regularized Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
14 The Minimum Norm Solution Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
14.1 The Basic Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
14.2 The Regularized Version . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
15 The Regularized Least Squares Solution Revisited . . . . . . . . . . . . . . . . . . . 161
16 Partial and Generalized Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
17 Summary and Future Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . 164
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

A Brief Review on Point Interactions

Gianfausto Dell’Antonio, Rodolfo Figari, and Alessandro Teta . . . . . . . . . . . 171
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
2 Construction of Point Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
3 Connection with Smooth Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
4 Time Dependent Point Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

List of Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

Introduction to Image Reconstruction

Miguel Moscoso

Gregorio Millán Institute of Fluid Dynamics, Nanoscience and Industrial

Mathematics, Universidad Carlos III de Madrid, Leganés, Spain
[email protected]

Summary. The problem of reconstructing images from measurements at the bound-

ary of a domain belong to the class of inverse problems. In practice, these measurements
are incomplete and inaccurate leading to ill-posed problems. This means that ‘exact’
reconstructions are usually not possible. In this Introduction the reader will find some
applications in which the main ideas about stability and resolution in image recon-
struction are discussed. We will see that although different applications or imaging
modalities work under different physical principles and map different physical parame-
ters, they all share the same mathematical foundations and the tools used to create the
images have a great deal in common. Current imaging problems deal with understand-
ing the trade off between data size, the quality of the image and the computational
tools used to create the image. In many cases, these tools represent the performance
bottleneck due to the high operational count and the memory cost.

1 Introduction

Imaging is a broad field which covers all aspects of the analysis, modification,
compression, visualization, and generation of images. It is a highly interdisci-
plinary field in which researchers from biology, medicine, engineering, computer
science, physics, and mathematics, among others, work together to provide the
best possible image. Imaging science is profoundly mathematical and challenging
from the modeling and the scientific computing point of view [1–5].
There are at least two major areas in imaging science in which applied math-
ematics has a strong impact: image processing, and image reconstruction. In
image processing the input is a (digital) image such as a photograph, while in
image reconstruction the input is a set of data. In the latter case, the data is
limited, and its poor information content is not enough to generate an image to
start with.
Image processing and analysis of information in images are methods that
become increasingly important in many technical and scientific fields. Image
processing techniques treat an image and apply numerical algorithms to either
improve the given image or to extract different features of it. Image segmentation
2 M. Moscoso

is typically used for the latter purpose. It refers to the process of partitioning an
image into multiple regions (locating objects and boundaries) in order to simplify
its representation for its further analysis. Each region shares the same properties
or characteristics such as color, intensity or texture. Many different techniques
have been applied for image segmentation. We mention here, graph partitioning
methods in which the image is modeled as a graph, level-sets methods in which
an initial shape is evolved towards the object boundary, or statistical methods
in which we view a region of an image as one realization of the various random
processes involved in the formation of that region (probability distribution func-
tions and histograms are used to estimate the characteristics of the regions). We
will not discuss the mathematics of image processing here.
Image reconstruction refers to the techniques used to create an image of the
interior of a body (or region) non-invasively, from data collected on its bound-
ary [1–5]. Image reconstruction can be seen as the solution of a mathematical
inverse problem in which the cause is inferred from the effect. As a consequence,
measurement and recording techniques designed to produce the images depend
deeply on the application we consider. In medicine, for example, very differ-
ent procedures are applied depending on the disease the physician is looking
for. Among these imaging techniques there are now in routine use, computer-
ized tomography, magnetic resonance, ultrasound, positron emission tomogra-
phy (PET), or electroencephalography (EEG), among others. They all acquire
data susceptible to be represented as maps containing positional information of
medical interest but, while each procedure operates on a different physical back-
ground principle and is sensitive to a different characteristic of the human body
(X-ray attenuation, dielectric properties, hydrogen nucleus density, reflection by
tissue interfaces, . . . ), they all share the same mathematical formulation of the
problem and similar numerical algorithms that allow the reconstruction of the
desired image.
We want to stress here, though, that it is always necessary a deep under-
standing of a mathematical model with which to interpret the measurements
and design numerical algorithms to reconstruct the desired images. Progress can
hardly be carried out without efficient and well designed numerical solvers. Imag-
ing is more than showing that an inverse problem may have a unique solution
under circumstances that are rarely satisfied. Modern imaging approaches deal
with understanding the trade off between data size, the quality of the image, the
computational complexity of the forward model used to generate the measure-
ments, and the complexity and stability of the numerical algorithm employed to
obtain the images. One neither has all the data he wants, nor can solve a very
general forward model to invert the data.
The purpose of this contribution is two fold: to give a first insight into dif-
ferent applications in which imaging has led to an important development, and
to outline the close mathematical connections between the imaging techniques
used in these applications. In Sect. 2 we review the impact that imaging has had
in medicine, geophysics and industry. In Sect. 3 we discuss the main ideas about
stability and resolution in imaging to introduce the reader to the basic notions
 Introduction to Image Reconstruction 3

and difficulties encountered in ill-posed inverse problems. Finally, in Sect. 4, we

present different image reconstruction techniques in computerized tomography
(which can be viewed as a linear inverse problem) and in diffuse optical tomog-
raphy (which can be viewed as a nonlinear inverse problem).

2 Applications
The problem of reconstructing images from boundary data is of fundamental
importance in many areas of science, engineering and industry. We will briefly
describe applications in different areas in which imaging has seen a significant
growth and innovation over the last decades. The presentation below has been
organized in three subsections: medicine, geophysics, and industry applications.
This organization is, however, somehow artificial. Seismic imaging, for example,
is described within the Geophysics subsection, but it is clear that seismic imaging
is of fundamental importance in the oil industry. Many of the imaging techniques
presented in each subsection have been used along the years, in one form or
another, within different applications as well. Magnetic resonance imaging, for
example, is described in the Medicine subsection since it is primarily used to
produce high quality images of the inside of the human body. However, magnetic
resonance is also used (in a different setting), for example, in oil exploration and
production since it can provide good estimates of permeability and fluid flow.

2.1 Medicine

X-ray tomography, ultrasound or magnetic resonance imaging have dramatically

changed the practice in medicine [2–5]. These techniques are now in routine use.
Nevertheless, they are still under current improvement, and an important part
of the scientific research is focused towards the limits these techniques can reach.
At the same time, new imaging techniques have been proposed along the last
decades with some advantages over the more traditional techniques. These new
techniques are the subject of a lot of academic research and their benefits need
to be investigated thoroughly in order to assess their validity as daily diagnostic
techniques. The development of these new technologies requires the combine
effort of many people with very distinct specialization.
Next, we mention four imaging techniques. The first two are well established
and are of daily use in our hospitals. The last two are still under development
and might have an enormous medical impact in a near future.

X-ray Tomography

One of the main revolutions which medicine has experienced during the last few
decades began in 1972 with the first clinical application of X-ray tomography or
computerized tomography [2–5]. However, the theoretical foundations underly-
ing these image reconstruction technique are due to the mathematician Johan
Radon, back in 1917.
4 M. Moscoso

Object

Source

Film
Fig. 1. An X-ray source illuminates a 2D object from a single view. Measurements,
as line integrals of the attenuation coefficient over each line, are registered on the film
plane. Observe that depth information is lost

X-ray tomography development was due to the necessity of overcoming some

limitations of the X-ray images as they were used in their time. The main draw-
back of a traditional X-ray image is that it represents a three-dimensional object
in a two-dimensional film. X-rays go through the human body, and some fraction
of the incident radiation is absorbed (or scattered) by the internal structures of
the body. The overall attenuation of the incident X-ray is proportional to the
average density of the body along the straight line joining the X-ray source with
a point of the film, in which this information is registered (Fig. 1). Since these
measurements involve line integrals the third dimension is lost. This means that
different structures that are completely separate in the body may overlap in the
final image. This is particularly troublesome in traditional X-ray images that
are read by eye by the physician.
The first experiments on medical applications were carried out by the physi-
cist Alan Cormack between 1957 and 1964. He derived a quantitative theory for
calculating the attenuation distribution of X-rays inside of the human body from
the acknowledgment of the transmitted X-rays. He actually learned of Radon’s
work much later regretting the time lost. The first practical implementation of
X-ray tomography was achieved by the engineer Godfrey Hounsfield in 1972.
Like Cormack, Hounsfield realized that three-dimensional information could be
retrieved from several two-dimensional projections, and like him, he worked with-
out the knowledge of previous work. As a historical note, we mention that the
 Introduction to Image Reconstruction 5

first applications of Radon’s theory were done by Bracewell in 1956 in the con-
text of radioastronomy, but his work had little impact and was not known in the
medical world. Both, Cormack and Hounsfield were awarded the Nobel Prize for
medicine in 1979. The new and powerful computers needed for the implementa-
tion of the algorithms, were crucial for the success of this technology.
A description of the theory of imaging in X-ray tomography can be found in
the lecture by Frank Natterer in this book. He surveys integral geometry and re-
lated imaging techniques. Computerized tomography reconstruction algorithms,
such as the filtered backprojection algorithm, Kaczmarz’s method, or Fourier
methods, are also presented.

Magnetic Resonance Imaging

While X-ray tomography uses ionizing radiation to form images, magnetic res-
onance imaging (MRI) uses non-ionizing signals. Magnetic resonance was orig-
inally named nuclear magnetic resonance, but the word nuclear was dropped
because of the negative connotations associated with the word nuclear. Neither
people nor local politicians wanted anything related to nuclear in their com-
munities. The term nuclear simply referred to the fact that all atoms have a
nucleus. The measurements are taken under powerful magnetic fields in con-
junction with radio waves. When placed in an external magnetic field, particles
with spin different than zero, such as protons, can absorb energy that is later
reemitted producing secondary radio waves. The detected signal results from the
difference between the absorbed and the emitted energies.
Magnetic resonance imaging, proposed by Richard Ernst in 1975, is proba-
bly the major competitor to X-ray tomography. In 1980 Edelstein and coworkers
demonstrated imaging of the body with his technique. Richard Ernst was re-
warded for his contributions to the development of high resolution techniques
in magnetic resonance imaging with the Nobel Prize in Chemistry. At the be-
ginning of the 90s magnetic resonance imaging started to be used for functional
imaging purposes. This new application allows the study of the function of the
various regions of the human brain. In 2003, Paul C. Lauterbur and Sir Peter
Mansfield were awarded the Nobel Prize in Medicine for their seminal discoveries
concerning the use of magnetic resonance to visualize different structures which
led to revolutionary insights into the functions of the brain and the workings of
the human body.
Besides its ability for functional imaging, magnetic resonance imaging is also
better suited for soft tissue than X-ray tomography. It can image different tis-
sue properties by variation of the scanning parameters, and tissue contrast can
be changed and enhanced in various ways to detect different features. The dis-
advantages of magnetic resonance imaging over X-ray tomography is that it
is more expensive and more time consuming. Nevertheless, the imaging tech-
niques behind this modality are the same used to reconstruct images in X-ray
tomography.
6 M. Moscoso

Diffuse Optical Tomography

Can we use light to see or image inside the body? Diffuse optical tomography in
the near-infrared is an emerging modality with potential in medical diagnosis and
biomedicine. A near-infrared light source impinges upon the surface and the light
propagates through the body (Fig. 2). The light that emerges is detected at an
array and used to infer the local optical properties (absorption and scattering) of
the illuminated tissue. The major absorbers when using near-infrared light are
water and both oxygenated and deoxygenated hemoglobin. Cancer metastasis
requires the growth of a new network of blood vessels (angiogenesis). The greater
supply of blood around the tumors give rise to an absorbing obstacle feasible to
be detected by this technique.
Another important current application of diffuse optical tomography is brain
imaging. By shining near-infrared light on the scalp, changes in neural activ-
ity can be measured because of the tight correlation existing between brain
activation and the changes in concentration of oxygenated and deoxygenated
hemoglobin (cerebral hemodynamics). This is possible due to the relationship of
the absorption spectra of oxygenated hemoglobin and deoxygenated hemoglobin
at near-infrared wavelengths. In addition, diffuse optical tomography can detect
and localize important events such as hemorrhagic strokes in the brain.
We would like to emphasize that to reconstruct quantitative images of the
absorption and scattering parameters from near-infrared light measurements one

Object

Near−infrared
source

Detectors
Fig. 2. A near-infrared source illuminates a 2D object. The photons do not travel
along straight lines. This causes blurring
 Introduction to Image Reconstruction 7

needs a mathematical model to interpret the data. The most comprehensive

model for diffuse optical tomography is the radiative transport equation

1 ∂I(x, k̂, t)
+ k̂ · ∇x I(x, k̂, t) + (Σ (s) (x) + Σ (a) (x))I(x, k̂, t) =
v ∂t

Σ (s) (x) P (k̂ · k̂ )I(x, k̂ , t) dk̂ , (1)

where v, Σ (s) , and Σ (a) are the velocity of the light in the medium, the scat-
tering coefficient, and the absorption coefficient, respectively. The fundamental
quantity of radiative transfer is the specific intensity I which depends on the
position vector x and unit direction vector k̂. The phase function P (k̂ · k̂ ) in (1)
describes the directional distribution of light that scatters in direction k̂ due to
light of unit energy density in direction k̂ . We have considered here the scalar
radiative transport equation.
The radiative transport equation (1) takes into account absorption and scat-
tering due to the inhomogeneities in tissue. Analytical solutions to this inte-
grodifferential equation are known only for relatively simple problems. For more
complicated problems numerical solutions are needed. The use of this theory
for imaging purposes in biological tissue can be found in the lecture by Miguel
Moscoso. A Monte Carlo method to solve this equation can also be found.
Before we proceed, we mention that one can use the diffusion equation to
model light propagation in tissue as an alternative to the more complicated
radiative transport equation. The diffusion equation is much easier to solve but
has limitations such that it does not describe accurately light propagation in
optically thin media nor light near sources and boundaries. Nevertheless, the
diffusion approximation to the radiative transport equation is widely used in the
literature.

Microwave Breast Imaging

Microwave imaging is being investigated as a possible alternative to mammogra-

phy. It uses electromagnetic data in the range from 300 MHz to 3 GHZ. Due to the
large contrast that exists between normal breast tissue and malignant tumors,
microwave imaging is also showing its potential to complement the mammogra-
phy to differentiate between malignant tumors and other possible lesions that
may appear in the breast. The key property of this new imaging modality is that
the contrast between malignancies and healthy tissue at microwave frequencies is
significantly greater than the small contrast exploited by X-ray mammography.
This technique is still the subject of a lot of research from the engineering
and mathematical point of views. Regarding the mathematical point of view,
a major limitation is the lack of efficient reconstruction algorithms that are
able to produce sharp images of small tumors. We recall here, that the large
contrast exhibited by the tumor give rise to a highly non-linear inverse problem.
Therefore, one-step method like the Born and Rytov approximations are not
well suited for this application. On the other hand, in this application one is not
8 M. Moscoso

so much interested on the detailed reconstruction of the spatial distribution of

the dielectric properties, but to know, in a reliable way, whether or not there is
a tumor, its location and size, and it malignancy.
To answer these questions, different coherent systems, that measure phases
(or arrival times) as well as intensities, are under investigation. (We remem-
ber here that X-ray tomography and diffuse optical tomography are incoherent
imaging modalities that only use intensities as input.) Other reconstruction algo-
rithms, based on the evolution of initial shapes that model the tumor boundary
are also under development. Shape-based approaches offer several advantages,
as for example, well defined boundaries and the incorporation of a-priori as-
sumptions regarding the general anatomical structures in the body that help to
simplify, and thereby, to stabilize the reconstruction algorithm. For the evolu-
tion law of the initial shape that model the tumor, sensitivity distributions of
the associated inverse problem are usually used.

2.2 Geophysics

Traditionally, geophysics is an interdisciplinary science in which modern applied

mathematics have had a crucial role in quantitative modeling, analysis, and
computational development. All of these issues are essential to interpret and
understand the measurements, as well as to asses the validity of the modeling.
Once again, modeling and scientific computing are the first steps to develop
efficient and robust imaging tools.
We now outline two applications of interest for imaging earth’s interior. They
involve different techniques depending on the nature and depth of the targets to
image. Seismic imaging uses information carried by elastic waves, and reflections
appear at boundaries with different impedances, while ground penetrating radar
uses electromagnetic waves and reflections appear at boundaries with different
dielectric constants.

Seismic Imaging

Seismic surveying involves sending sound waves underground and recording the
reflected signals from the earth’s interior [1]. Reflection occurs when the waves
hit a change in medium and experience a change in their velocity. This is a
simple fact. But changes in the earth’s properties take place in very different
spatial scales and are not easy to model. As a first approximation the earth is a
stratified medium in which the thickness of the layer might vary from less than
a meter to several hundreds of meters. On the other hand, tectonic forces in the
earth can bend these layers or can give rise to fractures. Besides, the interior
structure can be severely affected by the influence of other processes like the
extreme heat generated from the earth’s interior at some locations. As a result,
the seismic waves propagating through the earth will be refracted, reflected,
and diffracted from the heterogeneous medium, and the recorded signal will
show a very complicated structure. Earth scientists use these data to create a
 Introduction to Image Reconstruction 9

three-dimensional model of the subsurface, i.e., a spatial image of a property of

the earth’s interior (typically the impedance or the wave velocity).
In order to accomplish this task, one must solve an inverse scattering problem
to undo all these wave propagation effects and find the discontinuities. Migration
or back-propagation methods are the dominant imaging tools used for focusing
seismic images of subsurface structures. Seismic migration is used to transform
seismic data into seismic images that can be interpreted as geologic sections. This
technique uses an estimation of the local velocity to migrate or back-propagate
the surface data down into the earth. Thereby, seismic reflection images are
moved to their original position. Seismic migration is therefore, in some sense,
the inverse of seismic modeling that describes the forward process of propagating
waves from the sources to the scatterers, and to the receivers.
Migration of the seismic data is the most computationally expensive step in
order to obtain the images. This process is usually accomplished by the Kirchhoff
integral in which each measured data is summed into the image at the points
where the measured travel time agrees with the total propagation time (derived
from the assumed local velocity). The reflectors in the image are, therefore, built
up by constructive interference of the data, collapsing energy from reflections
back to their proper positions.

Electromagnetic Imaging

Electromagnetic data is a useful source of information for shallow subsurface

earth characterization. Ground penetrating radar is a non-destructive technique
that uses radar pulses to locate buried objects such as voids, contaminant plumes
such as tank-leaks, pipelines, or cables in the earth’s interior. It is also used to
locate, map and study buried archaeological features.
Ground penetrating radar sends a pulse from the surface. The probing elec-
tromagnetic waves are usually between the radio and microwave range of frequen-
cies. When the waves hit a buried object or a boundary with different dielectric
constants, the reflected signals are received at the surface. The mathematical
tools used to locate the object are similar to reflection seismology. The only
difference is that electromagnetic waves are used instead of elastic waves. It is
important to note that while seismic imaging can be adapted to any penetration
depth, ground penetrating radar suffers an important limitation due to dissipa-
tion. The depth range depends on the electrical conductivity of the ground and
the transmitted frequency. Typical penetration depths range from 5 to 10 m if
the electrical conductivity of the medium is low.
Another popular electromagnetic method of shallow subsurface imaging is
cross-borehole electromagnetic induction tomography. The source is a low fre-
quency magnetic field (1 KHz) generated in one vertical borehole, and the sig-
nals are received at a more distant borehole. The goal of this modality is to
image electrical conductivity variations in the earth. This technique can provide
high resolution images between wells up to 1,000 m apart. Optimum operating
frequencies, that depend on the borehole separation and ground conductivity
value, ranges between 40 and 100 KHz. Lower frequencies limit the resolution
10 M. Moscoso

but have a larger penetration range. Higher frequencies provide better resolu-
tion but limit the range of the measurements. The chapter by Oliver Dorn, Hugo
Bertete-Aguirre, and George Papanicolaou presents this modality of imaging and
a method for solving the inverse Maxwell problem in three dimensions.
More recent method involve deployment of both vertical borehole arrays and
large surface arrays.

2.3 Industry

It is apparent that imaging has transformed the development of many industries.

This is the case of some of the applications mentioned above such as seismic
imaging which is the basis of crucial exploration, development, and production
decisions in the oil industry.
We now survey two other industrial applications in which new imaging tools
specifically designed for industrial purposes have been developed.

Nondestructive Testing

The problem in which one needs to detect the presence of fissures or flaws from
measurements on the surface of a material is of much interest in industry. Nonde-
structive testing is any examination procedure performed on an object (without
destroying it) in order to detect the presence of defects such as cracks that can
alter its usefulness [3]. Nondestructive testing has become, in fact, an integral
part of almost all industrial processes in which a product failure might result in
an accident or body injury. For example, in aircraft maintenance it is important
to inspect the mechanical damage and to evaluate the amount of work necessary
to repair it. It is also important to do it in an efficient way, as the maintenance of
aircraft must be accomplished within a scheduled time. Nondestructive testing
can accomplish the inspection and can detect cracks or any other irregularities
in the airframe structure that are not visible. In other words, nondestructive
testing is the technology which provides information regarding the condition
of the material under examination once the necessary measurements have been
performed.
It is difficult to say when or who performed the first nondestructive testing.
Nondestructive tests can actually be performed in a very natural way without
employing sophisticated technologies. Listening to the sound of different metals
has been used along the centuries to determine the quality of the piece being
shaped (a sword, a ring, a bell, . . . ). However, it was in the late 1950s when
nondestructive techniques started to experience an important growth through
the development of new technologies and ideas. The mathematical study of this
inverse problem was initiated by A. Friedman and M. Vogelius in 1989.
There is a broad range of techniques that are applied nowadays depending
on the kind of measured data and the specific industrial situation under consid-
eration. Boundary data can be obtained from thermal, acoustic, elastostatic, or
electrical measurements. Lately, hybrid methods that perform different kind of
 Introduction to Image Reconstruction 11

measurements or use different technologies are under development. For example,

thermally induced acoustic waves is a new technique for characterizing aerospace
materials and structures from remote sensing. To successfully develop this tech-
nique for analyzing thermally generated images, new mathematical models to
describe the heat flow and the acoustic wave propagation induced by laser or
other thermal sources are needed.
Regardless of the kind of measurements taken on the boundary, many differ-
ent inversion approaches are used in nondestructive testing of industrial compo-
nents in order to locate any cavities, inclusions, or discontinuities. There are two
classes of methods: non-iterative methods and iterative methods. Non-iterative
methods, as the reciprocity gap technique for straight-line cracks or the factor-
ization method are fast but do not provide detail information about the crack.
On the other hand, iterative methods are computationally more expensive. They
need to solve a direct problem at each step to compute the Fréchet derivatives
of the direct operator but they usually offer very accurate results (if an initial
good estimate of the crack is used). Neural network operators can also be used
for defect detection.
The chapter by Ana Carpio and Maria Luisa Rapún considers the inverse
scattering problem for the detection of objects buried in a medium where waves
are described by the Helmholtz equation. The task is to find the locations and
shapes of the targets, given many input and output waves. Extensions to the
problem in which the objects are buried in a elastic medium is also given.

History Matching

Reservoir simulators are routinely used to predict and optimize oil production.
The accuracy of the numerical solution computed by the simulation depends
on the accuracy of the physical parameters that characterize the reservoir and,
unfortunately, these are poorly known due to the relative inaccessibility of the
reservoir to sampling. Therefore, proper characterization of the petroleum reser-
voir heterogeneity is a crucial aspect of any optimal reservoir management strat-
egy. It helps to better understand the reservoir behavior so that its performance
can be predicted and controlled with higher reliability. The history matching
problem consists in adjusting a set of parameters, such as the permeability and
porosity distributions, in order to match the data obtained with the simulator
to the actual production data in the reservoir. The input data used to solve the
inverse problem in history matching usually consist of the pressure or the flow
at the production wells.
One important difference between the history matching problem and other
applications such as X-ray tomography or magnetic resonance imaging is that in
the history matching problem the experimental setup cannot easily be changed
in order to obtain independent data. Typically, only one field experiment is
available due to the production process. A small reservoir model might have
50,000 to 100,000 grid blocks and solutions to the history matching problem
are far from being unique. As a consequence, production data rarely suffices to
characterize reservoir heterogeneity and it is, therefore, desirable to integrate all
12 M. Moscoso

other available data into the model, such as geological or seismic interpretations,
to reduce the amount of uncertainty. Lately, several researchers are making a lot
of effort to incorporate geological data, coming from sparse measurements of the
permeability, as constrains to the inverse problem. Most of the approaches are
formulated to honor the histograms generated from these data.

3 Stability
Let us consider the Hilbert spaces P and D (the parameter and data spaces,
respectively) and the linear bounded operator A : P → D. The goal of any
imaging technique is to take a set of data d ∈ D and produce an image of some
(unknown) physical quantity of interest f ∈ P . Mathematically, this is done by
solving an inverse problem represented in abstract form by

Af = d. (2)

However, the inverse problems arising from image reconstruction are usually
ill-posed in the sense of Hadamard. A problem is well-posed in the sense of
Hadamard if:
1. A solution exists.
2. The solution is unique.
3. The solution depends continuously on the given data.
Otherwise, the problem is ill-posed. A good measure of the degree of ill-posedness
is given by the singular value decomposition (SVD) of A. The faster the decay
rate of the singular values, the worse the degree of ill-posedness. If the singu-
lar values decay asymptotically at an exponential rate, the inverse problem is
severely ill-posed. This is the case, for example, of diffuse optical tomography or
cross-borehole electromagnetic induction tomography.
The most important issue in image reconstruction, from the computational
point of view, is the failure of the third property. One can deal with the first two
by approximate inverses such as the well known Moore–Penrose pseudoinverse
A† , but if the third property is not fulfilled any simple image reconstruction
algorithm will lead to instabilities and the image will not be reliable. To impose
stability on an ill-posed inverse problem we must introduce continuity in the
inverse operator A−1 by means of a regularization method [4].
We stress here that any measurement procedure leads to data perturbed by
noise. Rather than attempting to invert A directly (perfect inverses in realis-
tic problems, especially those arising in imaging reconstruction, typically never
happen), we adopt a least square approach and seek a solution f that minimizes
the cost functional
1
J(f ) = ||R(f )||22 , (3)
2
where R(f ) = A f − d is the residual operator which measures the data misfit.
 Introduction to Image Reconstruction 13

The basic idea of regularization is to replace the original ill-posed problem by

a nearby well-posed problem. There are two standard methods to accomplish reg-
ularization. The first method, Tikhonov–Phillips regularization, involves adding
the identity operator (or other simple operator), multiplied by a small factor
called the regularization parameter, to the approximate inverse of A. The choice
of an optimal, or at least good, regularization parameter is not trivial at all, and
depends on the specific imaging technique we consider. While a regularization
parameter too small do not restore the continuous dependence of the solution
upon the data, a too large one compromises the resolution which might lead to
a useless image for practical purposes. The second method, truncated SVD, dis-
card the smallest singular vectors corresponding to singular values below some
small threshold . This eliminates the impact of the noise-like data on the final
image. The regularization parameter here is .
One way to introduce regularization is to change the cost functional (3) by
incorporating an additional term. From this point of view, Tikhonov–Phillips
regularization amounts to minimize the cost functional
1
J(f ) = ||R(f )||22 + λ||Q(f )||22 (4)
2
where λ > 0 is the regularization parameter. The second term in (4) is the
penalty term through which prior information about the unknown f can be
taken into account. The prior information operators most often used in image
reconstruction are the first and second differential operators. The corresponding
implicit prior assumptions made when using these operators are that f changes
slowly or f is smooth, respectively.
Tikhonov–Phillips type regularizations, therefore, changes the cost functional
significantly which may not be desirable in some applications. In fact, this might
compromise the resolution of the final image. By adding a operator proportional
to the identity to the approximate inverse, Tikhonov–Phillips regularization ef-
fectively shifts all the singular values away from zero, thereby reducing the influ-
ence of the smallest singular values (noise-like) on the final reconstructed image
but, at the same time, reducing the resolvability of the largest ones (signal-like).
On the other hand, truncated SVD regularization just set the smallest singular
values to zero, so they are preferred when resolution is the only issue. We note,
however, that SVD is computationally highly demanding, and therefore, they
are not applied for large scale problems (as it is often the case in realistic inverse
problems).
The continue quest for methods that can improve resolution beyond the
natural or expected resolution limits is a major issue in imaging. Another possible
approach to regularization is to work with the original cost functional (3) and
introduce constraints on f requiring that the solution belongs to a smaller subset
of the parameter space P . This alternative form of regularization is presented in
the chapter by Oliver Dorn, Hugo Bertete-Aguirre, and George Papanicolaou.
Two different and interesting interpretations of this regularization approach are
also given.
14 M. Moscoso

In this same spirit, a similar strategy is to incorporate parameter space con-

straints so that the image exhibit the desirable properties. Lately, there is a lot of
interest in shape-reconstruction techniques that provide well-defined boundaries
of the objects to be imaged. Besides, these techniques incorporate an additional
intrinsic regularization (by constraining the parameter space due to prior knowl-
edge) that reduces significantly the dimensionality of the inverse problem, and
thereby stabilizing the reconstruction.

4 Image Reconstruction Techniques

Although the different imaging modalities presented in the previous section

probe different parameters or characteristics that describe the media, and work
under different physical principles, they all share the same mathematical foun-
dations, and the tools used to solve the associated inverse problem have a great
deal in common. Here, we will first consider two cases in which the applied imag-
ing techniques can be seen as different. In the last subsection we briefly mention
other problems and image reconstruction techniques that the reader can find in
the lectures of this book.

4.1 Computerized Tomography

Image reconstruction from projections is a linear inverse problem (typically un-

derdetermined) that attempts to retrieve an unknown function (the image) from
a set of lower dimensional profiles of this unknown function. In several clini-
cal imaging techniques such as computerized tomography, magnetic resonance
imaging, positron emission tomography, or single particle emission computed
tomography, these profiles are obtained by integrating the unknown function
along straight lines orthogonal to the projections. As a consequence, they all
share similar imaging algorithms. Some of this techniques, as the backprojec-
tion algorithm, the Kaczmarz’s method, or Fourier methods are outlined in the
lecture by Frank Natterer in this book.
Mathematically, the inverse problem can be formulated as follows. Let us
assume that the set of measurements


(Rf )(θ, s) = f (x, y)δ(x cos θ + y sin θ − s) dx dy (5)
R2

over s ∈ R is available for some projections θ ∈ [0, π], where δ is the Dirac’s delta
function. The inverse problem is to find f from a finite number of measurements
(Rf )(θi , sj ) = dij . If Rf is available for all θ and s (without noise), then the
inverse Radon transform solves the problem exactly. This inverse problem is,
nevertheless, ill posed.
Simple backprojection is conceptually quite easy but it does not correctly
solve this inverse problem. Simple backprojection reconstructs an image by tak-
ing each projection and smearing (or backprojecting) it along the line it was
 Introduction to Image Reconstruction 15

acquired. As a result of this simple procedure, the backprojected image is blurred.

Filtered backprojection, which is the most important reconstruction algorithm
in tomography, is a technique devised to correct the blurring encountered in
simple backprojection. Each projection is properly filtered before the backpro-
jection is done to counteract the blur. The image generated by this algorithm
agrees with the true image if an infinite number of projections are available. The
filtered backprojection can actually be viewed as a numerical implementation of
the inverse Radon transform which solves the inverse problem exactly.
The Kaczmarz’s method, also called algebraic reconstruction technique
(ART), is an iterative technique. Iterative techniques may be slow, but they are
generally used when better algorithms are not available. In fact, they were used
in the first commercial medical CT scanner in 1972. The ART algorithm uses
successive orthogonal projections to approximate solutions to systems of linear
equations. This algorithm approximates a solution by projecting a given vector
successively onto the hyperplanes defined by the individual equations. To start
the ART algorithm, all the pixels in the image are set to an initial arbitrary
value.
Finally, Fourier methods takes the two-dimensional Fourier transform of the
image and the one-dimensional Fourier transform of each of its views. It turns
out that the relationship between an image and its views is much simpler in the
frequency domain than in the spatial domain.

4.2 Diffuse Optical Tomography

Diffuse optical tomography is an example of an imaging modality that requires

the solution of a nonlinear inverse problem. This kind of imaging techniques are
still the subject of a lot of academic research and their future usefulness in areas
such as medicine are still under discussion.
Several challenges arise in diffuse optical tomography due to the multiple
scattering of light in biological tissue. Physically, multiple scattering causes se-
vere image blurring. Hence, one can not make use of direct images. Since light
does not travel along straight lines, one can not make use of the reconstruc-
tion algorithms applied in X-ray tomography neither. Rather, one must develop
methods to reconstruct images from strongly scattered light measurements. One
possible approach is to discriminate between weakly and strongly scattered pho-
tons. Various techniques, such as gating techniques, have been devised to carry
out this task. Time gating uses picosecond time-resolved techniques to select the
early arriving photons at the detector, thereby selecting those photons traveling
along straight paths. Therefore, the information contained in the detected signal
give better contrast and/or could be used in X-ray tomography based imaging al-
gorithms. Polarization gating is based on the fact that weakly scattered photons
(carrying the useful information) retains the polarization state of the incident
light while strongly scattered photons do not. In this book, the reader can find
a study by Miguel Moscoso that uses polarization effects to obtain enhanced
resolution images in biological tissue.
16 M. Moscoso

4.3 Other Problems

Image reconstruction methods using a backpropagation strategy based on adjoint

fields have been applied in diffuse optical tomography. The reader can find an
exposition of this method for the cross-borehole electromagnetic induction to-
mography application in the lecture by Oliver Dorn, Hugo Bertete-Aguirre, and
George Papanicolaou. The adjoint field technique is an iterative method that
approximately solves the inverse problem (upon linearization) by making use of
the same forward modeling code two time in each step of the algorithm: one
to compute the misfit between the ‘real’ and the simulated data, and one to
compute the sensitivity map.
Ana Carpio and Marı́a Luisa Rapún explain a promising numerical scheme for
solving inverse scattering problems by means of a topological derivative method.
The topological derivative is essentially the gradient of an objective functional in
the limit of a hole vanishing to zero. Thereby, it provides information about the
benefits of introducing a hole at a given location without the need of a further
analysis. Lately, there is a lot of interest in the use of topological derivatives for
computationally efficient methods for shape functional optimization problems.
An example of coherent imaging can be found in the lecture by Oliver Dorn.
He establishes a direct link between the time-reversal technique and the adjoint
method for imaging. Using this relationship, he derives new solution strategies
for an inverse problem in telecommunication. These strategies are all based on
iterative time-reversal experiments, which try to solve the inverse problem exper-
imentally instead of computationally. He focuses, in particular, on examples from
underwater acoustic communication and wireless communication in a Multiple-
Input Multiple-Output (MIMO) setup.

Acknowledgements

The author acknowledges support from the Spanish Ministerio de Educación y

Ciencia (grants n◦ FIS2004-03767 and FIS2007-62673) and from the European
Union (grant LSHG-CT-2003-503259).

References
1. Claerbout J.F., Fundamentals of Geophysical Data Processing, McGraw-Hill
international series in the earth and planetary sciences, 1976.
2. Epstein C.L., Introduction to the Mathematics of Medical Imaging, Prentice Hall,
2003.
3. Hellier C., Handbook of nondestructive evaluation, McGraw-Hill, 2001.
4. Kirsch A., An introduction to the mathematical theory of inverse problems, Applied
Mathematical Science vol. 120, Springer, 1996.
5. Natterer F. and Wübbeling F., Mathematical Methods in Image Reconstruction,
SIAM Monographs on Mathematical Modeling and Computation, 2001.
X-ray Tomography

Frank Natterer

Institut für Numerische und Angewandte Mathematik, University of Münster,

Einsteinstraße 62, 48149 Münster, Germany
[email protected]

Summary. We give a survey on the mathematics of computerized tomography. We

start with a short introduction to integral geometry, concentrating on inversion formu-
las, stability, and ranges. We then go over to inversion algorithms. We give a detailed
analysis of the filtered backprojection algorithm in the light of the sampling theorem.
We also describe the convergence properties of iterative algorithms. We shortly mention
Fourier based algorithms and the recent progresses made in their accurate implemen-
tation. We conclude with the basics of algorithms for cone beam scanning which is the
standard scanning mode in present days clinical practice.

1 Introduction

X-ray tomography, or computerized tomography (CT) is a technique for imaging

cross sections or slices (slice = τ oµos (greek)) of the human body. It was intro-
duced in clinical practice in the 70s of the last century and has revolutionized
clinical radiology. See Webb (1990), Natterer and Ritman (2002) for the history
of CT.
The mathematical problem behind CT is the reconstruction of a function f in
R2 from the set of its line integrals. This is a special case of integral geometry, i.e.
the reconstruction of a function from integrals over lower dimensional manifolds.
Thus, integral geometry is the backbone of the mathematical theory of CT.
CT was soon followed by other imaging modalities, such as emission tomo-
graphy (single particle emission tomography (SPECT) and positron emission
tomography (PET)), magnetic resonance imaging (MRI). CT also found appli-
cations in various branches of science and technology, e.g. in seismics, radar,
electron microscopy, and flow. Standard references are Herman (1980), Kak and
Slaney (1987), Natterer and Wübbeling (2001).
In these lectures we give an introduction into the mathematical theory and
the reconstruction algorithms of CT. We also discuss matters of resolution and
stability. Necessary prerequisites are calculus in Rn , Fourier transforms, and the
elements of sampling theory.
18 F. Natterer

Since the advent of CT many imaging techniques have come into being, some
of them being only remotely related to the straight line paradigm of CT, such
as impedance tomography, optical tomography, and ultrasound transmission to-
mography. For the study of these advanced technique a thorough understanding
of CT is at least useful. Many of the fundamental tools and issues of CT, such as
backprojection, sampling, and high frequency analysis, have their counterparts
in the more advanced techniques and are most easily in the framework of CT.
The outline of the paper is as follows. We start with a short account of the
relevant parts of integral geometry, with the Radon transform as the central tool.
Then we discuss in some detail reconstruction algorithms, in particular the nec-
essary discretizations for methods based on inversion formulas and convergence
properties of iterative algorithm. Finally we concentrate on the 3D case which
is the subject of current research.

2 Integral Geometry
In this section we introduce the relevant integral transforms, derive inversion
formulas, and study the ranges. For a thorough treatment see Gelfand, Graev,
and Vilenkin (1965), Helgason (1999), Natterer (1986).

2.1 The Radon Transform

Let f be a function in Rn . To avoid purely technical difficulties we assume f to
be smooth and of compact support, if not said otherwise.
For θ ∈ S n−1 , s ∈ R1 we define


(Rf )(θ, s) = f (x)dx . (1)
x·θ=s

R is the Radon transform. dx is the restriction of the Lebesgue measure in Rn

to x · θ = s. Other notations are


(Rf )(θ, s) = f (sθ + y)dy (2)
θ⊥

with θ⊥ the subspace of Rn orthogonal to θ, and

(Rf )(θ, s) = δ(x · θ − s)f (x)dx (3)

with δ the Dirac δ-function.

The Radon transform is closely related to the Fourier transform


−n/2
ˆ
f (ξ) = (2π) e−ix·ξ f (x)dx .
Rn

Namely, we have the so-called projection-slice theorem.

 X-ray Tomography 19

Theorem 2.1.
(Rf )∧ (θ, σ) = (2π)(n−1)/2 fˆ(σθ) .
Note that the Fourier transform on the left hand side is the 1D Fourier
transform of Rf with respect to its second variable, while the Fourier transform
on the right hand side is the nD Fourier transform of f .
For the proof of Theorem 2.1 we make use of the definition (2) of Radon
transform, yielding


(Rf )∧ (θ, σ) = (2π)−1/2 e−isσ (Rf )(θ, s)ds
R1



= (2π)−1/2 e−isσ f (sθ + y)dyds .
R1 θ⊥

Putting x = sθ + y, i.e. s = x · θ, we have

∧ −1/2
(Rf ) (θ, s) = (2π) e−iσx·θ f (x)dx
Rn

= (2π)(n−1)/2 fˆ(σθ) .
The proofs of many of the following results follow a similar pattern. So we omit
proofs unless more sophisticated tools are needed.
It is possible to extend R to a bounded operator R : L2 (Rn ) → L2 (S n−1 ×R1 ).
As such R has an adjoint R∗ : L2 (S n−1 × R1 ) → L2 (Rn ) which is easily seen
to be

(R∗ g)(x) = g(θ, x · θ)dθ . (4)
S n−1
∗
R is called the backprojection operator in the imaging literature.
We also will make use of the Hilbert transform


1 f (t)
(Hf )(s) = dt (5)
π s−t
which, in Fourier domain, is given by
(Hf )∧ (σ) = −i sgn(σ)fˆ(σ) (6)
with sgn(σ) the sign of the real number σ. Now we are ready to state and to
prove Radon’s inversion formula:

Theorem 2.2. Let g = Rf . Then

1
f= (2π)1−n R∗ H n−1 g (n−1)
2
where g (n−1) stands for the derivative of order n − 1 of g with respect to the
second argument.
20 F. Natterer

For the proof we write the Fourier inversion formula in polar coordinates and
make use of Theorem 2.1 and the evenness of g, i.e. g(θ, s) = g(−θ, −s).
Special cases of Theorem 2.2 are



1 g (θ, s)
f (x) = dsdθ (7)
4π 2 x·θ−s
S 1 R1

for n = 2 and

1
f (x) = − g  (θ, x · θ)dθ (8)
8π 2
S2
for n = 3. Note that there is a distinctive difference between (7) and (8): The
latter one is local in the sense that for the reconstruction of f at x0 only the
integrals g(θ, s) over those planes x · θ = s are needed that pass through x0 or
nearby. In contrast, (7) is not local in this sense, due to the integral over R1 .
In order to study the stability of the inversion process we introduce the
Sobolev spaces H α on Rn , S n−1 × R1 with norms


f 2H α (Rn ) = (1 + |ξ|2 )α |fˆ(ξ)|2 dξ ,
Rn



gH α (Sn−1 ×R1 ) =
2
(1 + σ 2 )α |ĝ(θ, σ)|2 dσdθ .
S n−1 R1

As in Theorem 2.1, the Fourier transform ĝ is the 1D Fourier transform of g

with respect to the second argument.

Theorem 2.3. There exist constants c, C > 0 depending only on α, n, such that
cf H α (Rn ) ≤ Rf  (S n−1 ×R1 ) ≤ Cf H α (Rn )
H α+(n−1)/2

for f ∈ H α (Rn ) with support in |x| < 1.

The conclusion is that Rf is smoother than f by the order (n − 1)/2. This
indicates that the reconstruction process, i.e. the inversion of R, is slightly ill-
posed.
Finally we study the range of R.

Theorem 2.4. For m ≥ 0 an integer let

pm (θ) = sm (Rf )(θ, s)ds .
R1

Then, pm is a homogeneous polynomial of degree m in θ.

The proof is by verification. The question wether or not the condition of
Theorem 2.4 is sufficient for a function g of θ, s being in the range of R is the
subject of the famous Helgason–Ludwig theorem.
Theorem 2.4 has applications to cases in which the data function g is not
fully specified.
 X-ray Tomography 21

2.2 The Ray Transform

For f a function in Rn , θ ∈ S n−1 , x ⊥ θ⊥ we define


(P f )(θ, x) = f (x + tθ)dt (9)
R1

P is called the ray (or X-ray) transform. The projection-slice theorem for P reads

Theorem 2.5.
(P f )∧ (θ, ξ) = (2π)1/2 fˆ(ξ), ξ ∈ θ⊥ .
The Fourier transform on the left hand side is the (n − 1)D Fourier transform
in θ⊥ , while the Fourier transform on the right hand side is the nD Fourier
transform in Rn .
The adjoint ray transform, which we call backprojection again, is given by


(P ∗ g)(x) = g(θ, Eθ x)dθ (10)
S n−1

with Eθ the orthogonal projection onto θ⊥ , i.e. Eθ x = x − (x · θ)θ. We have the

following analogue of Radon’s inversion formula:

Theorem 2.6. Let g = P f . Then

1
f= P ∗ I −1 g
2π|S n−1 |
with the Riesz potential
(I −1 g)∧ (ξ) = |ξ|ĝ(ξ)
in θ⊥ .
Theorem 2.6 is not as useful as Theorem 2.2. The reason is that for n ≥ 3 (for
n = 2 Theorems 2.2 and 2.6 are equivalent) the dimension 2(n − 1) of the data
function g is greater than the dimension n of f . Hence the problem of inverting
P is vastly overdetermined. A useful formula would compute f from the values
g(θ, ·) where θ is restricted to some set S0 ⊆ S n−1 . Under which conditions on
S0 is f uniquely determined?

Theorem 2.7. Let n = 3, and assume that each great circle on S 2 meets S0 .
Then, f is uniquely determined by g(θ, ·), θ ∈ S0 .
The condition on S0 in Theorem 2.7 is called Orlov’s completeness condition.
In the light of Theorem 2.5, the proof of Theorem 2.7 is almost trivial: Let
ξ ∈ R3 \ {0}. According to Orlov’s condition there exists θ ∈ S0 on the great
circle ξ ⊥ ∩ S 2 . For this θ, ξ ∈ θ⊥ , hence
fˆ(ξ) = (2π)−1/2 ĝ(θ, ξ)
by Theorem 2.5. Thus fˆ is uniquely (and stably) determined by g on S0 .
22 F. Natterer

An obvious example for a set S0 ⊆ S 2 satisfying Orlov’s completeness con-

dition is a great circle. In that case inversion of P is simply 2D Radon inversion
on planes parallel to S0 . Orlov gave an inversion formula for S0 a spherical zone.

2.3 The Cone-Beam Transform

The relevant integral transform for fully 3D X-ray tomography is the cone beam
transform in R3

∞
(Cf )(a, θ) = f (a + tθ)dt (11)
0

where θ ∈ S . The main difference to the ray transform P is that the source
2

point a is restricted to a source curve A surrounding the object to be imaged.

Of course the following question arises: Which condition on A guarantees
that f is uniquely determined by (Cf )(a, ·)?
A first answer is: Whenever A has an accumulation point outside supp (f )
(provided f is sufficiently regular). Unfortunately, inversion under this assump-
tions is based on analytic continuation, a process that is notoriously unstable.
So this result is useless for practical reconstruction.
In order to find a stable reconstruction method (for suitable source curves A)
we make use of a relationship between C and R discovered by Grangeat (1991):

Theorem 2.8.


∂ ∂
(Rf )(θ, s) |s=a·θ = (Cf )(a, ω)dω .
∂s ∂θ
θ ⊥ ∩S 2

The notation on the right hand side needs explication: (Cf )(a, ·) is a function
on S 2 . θ is a tangent vector to S 2 for each ω ∈ S 2 ∩ θ⊥ . Thus the derivative ∂θ
∂

makes sense on S 2 ∩ θ⊥ .
Since Theorem 2.8 is the basis for present day’s work on 3D reconstruction
we sketch two proofs. The first one is completely elementary. Obviously it suffices
to put θ = e3 , the third unit vector, in which case it reads
⎡ ⎛ ⎛ ⎞⎞⎤


∂ ∂ ω
(Rf )(θ, a3 ) = ⎣ (Cf ) ⎝a, ⎝ ⎠⎠⎦ dω
∂s ∂z z
ω∈S 1 z=0

where a3 is the third component of the source point a. It is easy to verify that
right- and left-hand side both coincide with
⎛ ⎛ ⎞⎞


∂f ⎝ x
a + ⎝ ⎠⎠ dx .
∂x3 0
R2

The second proof makes use of the formula

 X-ray Tomography 23



(Cf )(a, ω)h(θ · ω)dω = (Rf )(θ, s)h(s − a · θ)ds
S n−1 R1

that holds – for suitable functions f in R2 – for h a function in R1 homogeneous

of degree 1 − n; see Hamaker et al. (1980).
Putting h = δ  yields Theorem 2.8.

Theorem 2.9. Let the source curve A satisfy the following condition: Each plane
meeting supp (f ) intersects A transversally. Then, f is uniquely (and stably)
determined by (Cf )(a, θ), a ∈ A, θ ∈ S 2 .

The condition on A in Theorem 2.9 is called the Kirillov–Tuy condition.

The proof of Theorem 2.9 starts out from Radon’s inversion formula (8) for
n = 3:

1 ∂2
f (x) = − 2 Rf (θ, s) dθ (12)
8π ∂s2 s=x·θ
S2

Now assume A satisfies the Kirillov–Tuy condition, and let x · θ = s be a plane

hitting supp (f ). Then there exists a ∈ A such that a · θ = s, and

∂ ∂
Rf (θ, s) = Rf (θ, s)
∂s s=x·θ ∂s s=a·θ


∂
= (Cf )(a, ω)dω
∂θ
θ ⊥ ∩S 2

is known by Theorem 2.8. Since A intersects the plane x · θ = s transversally

this applies also to the second derivative in (12). Hence the integral in (12) can
be computed from the known values of Cf for all planes x · θ = s hitting supp
(f ), and for the others it is zero.
Theorem 2.8 in conjunction with Radon’s inversion formula (8) is the basis
for reconstruction formulas for C with source curves A satisfying the Kirillov–
Tuy condition. The simplest source curve one can think of, a circle around the
object, does not satisfy the Kirillov–Tuy condition. For a circular source curve
an approximate inversion formula, the FDK formula, exists; see Feldkamp et al.
(1984). In medical applications the source curve is a helix.

2.4 The Attenuated Radon Transform

Let n = 2. The attenuated Radon transform of f is defined to be

(Rµ f )(θ, s) = e−(Cµ )(x,θ⊥ ) f (x)dx
x·θ=s
24 F. Natterer

where µ is another function in R2 and θ⊥ is the unit vector perpendicular to

θ such that det (θ, θ⊥ ) = 1. This transform comes up in SPECT, where f the
(sought-for) activity distribution and µ the attenuation map.
Rµ admits an inversion formula very similar to Radon’s inversion formula
for R:

Theorem 2.10. Let g = Rµ f . Then,

1 ∗
f= Re div R−µ (θe−h Heh g)
4π
where
1
h = (I + iH)Rµ ,
2


(Rµ∗ g)(x) = e−(Cµ)(x,θ⊥ ) g(θ, x · θ)dθ .
S1

This is Novikov’s inversion formula; see Novikov (2000). Note that the back-
∗
projection R−µ is applied to a vector valued function. For µ = 0 Novikov’s
formula reduces to Radon’s inversion formula (7).
There is also an extension of Theorem 2.4 to the attenuated case:

Theorem 2.11. Let k > m ≥ 0 integer and h the function from Theorem 2.10.
Then,


sm eikϕ+h(θ,s) (Rµ f )(θ, s)dθds = 0
R1 S 1
⎛ ⎞
cos ϕ
where θ = ⎝ ⎠.
sin ϕ

The proofs of Theorems 2.10 and 2.11 are quite involved. They are both
based on the following remark: Let

u(x, θ) = h(θ, x · θ) − (Cµ)(x, θ⊥ ) .

Then, u admits a Fourier series representation as

u(x, θ) = u (x)eiϕ
>0 odd

⎛ ⎞
cos ϕ
with certain functions u (x) and θ = ⎝ ⎠.
sin ϕ
Of course this amounts to saying that u(x, ·) admits an analytic continuation
from S 1 into the unit disk.
 X-ray Tomography 25

2.5 Vectorial Transforms

Now let f be a vector field. The scalar transforms P , R can be applied compo-
nentwise, so P f , Rf make sense. They give rise to the vectorial ray transform

(Pf )(θ, s) = θ · (P f )(θ, s) , x ∈ θ⊥

and the Radon normal transform

(R⊥ f )(θ, s) = θ · (Rf )(θ, s) .

Obviously, vectorial transforms can’t be invertible. In fact their null spaces are
huge.

Theorem 2.12. Pf = 0 if and only if f = ∇ψ for some ψ.

The nontrivial part of the proof makes use of Theorem 2.5: If Pf = 0, then,
for ξ ∈ θ⊥

(Pf )∧ (θ, ξ) = θ · (P f )∧ (θ, ξ) = (2π)1/2 θ · fˆ(ξ) = 0 .

Hence θ⊥ is invariant under fˆ. This is only possible if fˆ(ξ) = iψ̂(ξ)ξ with some
ψ̂(ξ). It remains to show that ψ̂ is sufficiently regular to conclude that f = ∇ψ.
A similar proof leads to

Theorem 2.13. Let n = 3. R⊥ f = 0 if and only if f = curl φ for some φ.

Inversion formulas can be derived for those parts of the solution that are
uniquely determined. We make use of the Helmholtz decomposition to write a
vector f in the form
f = fs + fi
where f s is the solenoidal part (i.e. div f s = 0 or f s = curl φ) and f i is the
irrotational part (i.e. curl f i = 0, f i = ∇ψ). Then, Theorem 2.12 states that the
solenoidal part of f is uniquely determined by Pf , while the irrotational part of
f is uniquely determined by R⊥ f .
We have the following inversion formula:

Theorem 2.14. Let g = Pf and let f be solenoidal (i.e. f i = 0). Then,

n−1
f= P ∗ I −1 g ,
2π|S n−2 |


(P ∗ g)(x) = g(θ, Eθ x)θdθ
S n−1

and I −1 is the Riesz potential already used in Theorem 2.6.

For more results see Sharafutdinov (1994).
26 F. Natterer

3 Reconstruction Algorithms
There are essentially three classes of reconstruction algorithms. The first class
is based on exact or approximate inversion formulas, such as Radon’s inversion
formula. The prime example is the filtered backprojection algorithm. It’s not only
the work horse in clinical radiology, but also the model for many algorithms in
other fields. The second class consists of iterative methods, with Kaczmarz’s
method (called algebraic reconstruction technique (ART) in tomography) as
prime example. The third class is usually called Fourier methods, even though
Fourier techniques are used also in the first class. Fourier methods are direct
implementations of the projection slice theorem (Theorems 2.1, 2.5) without any
reference to integral geometry.

3.1 The Filtered Backprojection Algorithm

This can be viewed as an implementation of Radon’s inversion formula
(Theorem 2.2). However it is more convenient to start out from the formula

Theorem 3.1. Let V = R∗ v. Then,

V ∗ f = R∗ (v ∗ g)
where the convolution on the left hand side is in Rn , i.e.


(V ∗ f )(x) = V (x − y)f (y)dy
Rn

and the convolution on the right hand side is in R1 , i.e.

(v ∗ g)(θ, s) = v(θ, s − t)g(θ, t)dt .
R1

The proof of Theorem 3.1 require not much more than the definition of R∗ .
Theorem 3.1 is turned into an (approximate) inversion formula for R by
choosing V ∼ δ. Obviously V is the result of the reconstruction procedure for
f = δ. Hence V is the point spread function, i.e. the function the algorithm
would reconstruct if the true f were the δ-function.

Theorem 3.2. Let V = R∗ v with v independent of θ. Then,

V̂ (ξ) = 2(2π)(n−1)/2 |ξ|1−n v̂(|ξ|) .
For V to be close to δ we need V̂ to be close to (2π)− 2 , hence
n

1
v̂(σ) ∼ (2π)−n+1/2 σ n−1 .
2
Let φ be a low pass filter, i.e. φ(σ) = 0 for |σ| > 1 and, for convenience, φ
even. Let Ω be the (spatial) bandwidth. By Shannon’s sampling theorem Ω
corresponds to the (spatial) resolution 2π/Ω:
 X-ray Tomography 27

Theorem 3.3. Let f be Ω-bandlimited, i.e. fˆ(ξ) = 0 for |ξ| > Ω. Then, f is
uniquely determined by f (h),  ∈ Zn , h ≤ π/Ω. Moreover, if h ≤ 2π/Ω, then


f (x)dx = hn f (h) .
Rn 

An example for a Ω-bandlimited function in R1 is f (x) = sinc(Ωx) where

sinc(x) = sin(x)/x is the so-called sinc-function. From looking at the graph of f
we conclude that bandwidth Ω corresponds to resolution 2π/Ω.
We remark that Theorem 3.3 has the following corollary:
Assume that f , g are Ω-bandlimited. Then,


f (x)g(x)dx = hn f (h)g(h)
Rn 

provided that h ≤ π/Ω. This is true since f g has bandwidth 2Ω.

We put
v̂(σ) = 2(2π)(n−1)/2 |σ|n−1 φ(σ/Ω) .
v is called the reconstruction filter. As an example we put n = 2 and take as φ
the ideal low pass, i.e. ⎧
⎨ 1 , |σ| ≤ 1 ,
φ(σ) =
⎩ 0 , |σ| > 1 .

Then,
Ω2 1  s 2
v(s) = u(Ωs) , u(s) = sinc(s) − sinc .
4π 2 2 2
Now we describe the filtered backprojection algorithm for the reconstruction of
the function f from g = Rf . We assume that f has the following properties:

f is supported in |x| ≤ ρ (13)

f is essentially Ω-bandlimited, i.e. (14)

fˆ(ξ) is negligible in some sense for |ξ| > Ω .

Since f is essentially Ω-bandlimited we conclude from Theorem 2.1 that the

same is true for g. Hence, by a loose application of the sampling theorem,

(v ∗ g)(θ, s ) = ∆s v(θ, s − sk )g(θ, sk ) (15)

k

where s = ∆s and ∆s ≤ π/Ω. Equation (15) is the filtering step in the filtered
backprojection algorithm. In the backprojection step we compute R∗ (v ∗ g) in a
discrete setting. We restrict ourselves to the case n = 2, i.e.
28 F. Natterer


(R∗ (v ∗ g))(x) = (v ∗ g)(θ, x·)dθ .
S1

In order to discretize this integral properly we make use of Theorem 3.3 again.
For this we⎛have to⎞determine the bandwidth of the function ϕ → (v ∗ g)(θ, x · θ)
cos ϕ
where θ = ⎝ ⎠.
sin ϕ

Theorem 3.4. For k an integer we have

2π

Ω
−ikϕ ik −ikψ
e (v ∗ g)(θ, x · θ)dϕ = f (y)e v̂k (σ)Jk (−σ|x − y|)dσdy
2π
0 |y|<ρ −Ω

where ψ = arg(y) and Jk is the Bessel function of order k of the first kind.

All that is needed for the proof is the integral representation

2π
i−k
Jk (s) = e−ikϕ+s cos ϕ dϕ
2π
0

of the Bessel function.

By Debye’s asymptotic relation (see Abramowitz and Stegun (1970)), Jk (s) is
negligibly small for |s| < |k| and |k| large. Hence the left hand side in Theorem 3.4
is negligible for 2Ωρ < |k|. In other words, the function ϕ → (v ∗ g)(θ, x · θ) has
essential bandwidth 2Ωρ. According to Theorem 3.3,

(R∗ (v ∗ g))(x) = ∆ϕ (v ∗ g)(θj , x · θj ) (16)

j
⎛ ⎞
cos ϕj
where θj = ⎝ ⎠, ϕj = j∆ϕ, provided that ∆ϕ ≤ π/Ωρ.
sin ϕj
The formulas (16), (17) describe the filtered backprojection algorithm – ex-
cept for an interpolation step in the second argument. It is generally acknowl-
edged that linear interpolation suffices.

3.2 Iterative Algorithms

The equations one has to solve in CT are usually of the form

Rj f = gj , j = 1, . . . , p (17)
 X-ray Tomography 29

where Rj are certain linear operators from a Hilbert space into an Hilbert space
Hj . For instance, for the 2D problem we have

(Rj f )(s) = (Rf )(θj , s) . (18)

Kaczmarz’s method solves linear systems of the kind (18) by successively pro-
jecting orthogonally onto the affine subspaces Rj f = gj of H:

fj = Pj modp fj−1 , j = 1, 2, . . . .

Here, Pj is the orthogonal projection onto Rj f = gj , i.e.

Pj f = f − Rj∗ (Rj Rj∗ )−1 (Rj f − gj ) .

For the application to tomography we replace the operator Rj Rj∗ by a simpler

one, Cj , and introduce a relaxation parameter ω. Putting

Pj f = f − Rj∗ Cj−1 (Rj f − gj ) ,

Pjω f = (1 − ω)f + ωPj f

we have the following convergence result.

Theorem 3.5. Let (17) be consistent. Assume that Cj ≤ Rj Rj∗ (i.e. Cj her-
mitian and Cj − Rj Rj∗ positive semidefinite) and 0 < ω < 2. Then, fj =
Pjωmodp fj−1 converges for f0 = 0 to the solution of (18) with minimal norm.
The condition 0 < ω < 2 is reminiscent of the SOR (successive overre-
laxation) method of numerical analysis. In fact Kaczmarz’s method can be
viewed as an SOR method applied to the linear system R∗ Rf = R∗ g where
R = R 1 × · · · × Rp .
Theorem 3.5 can be applied directly to tomographic problems. The conver-
gence behaviour depends critically on the choice of ω and of the ordering of (17).
We analyse this for the standard case (18). For convenience we assume that f is
supported in |x| < 1 and gj = 0, j = 1, . . . , p.
The following result is due to Hamaker and Solmon (1978):

Theorem 3.6. Let Cm be the linear subspace of L2 (|x| < 1) spanned by Tm (x ·

θ1 ), . . . , Tm (x · θp ), Tm the first kind Chebysheff polynomials. Then,

Pj Cm ⊆ Cm .

This means that Cm is an invariant subspace under the iteration of

Theorem 3.6. Thus we can study the convergence behaviour on each of these
finite dimensional subspaces separately. Let ρm (ω) be the spectral radius of

P ω = P1ω · · · Ppω .

ρm (ω) can be computed numerically:

30 F. Natterer
ρm (ω)
1 ω = 0.1

ω = 0.5

ω = 1.0

0 m
0 30

ρm (ω)
1 ω = 0.1

ω = 0.5

ω = 1.0

0 m
0 30

ρm (ω)
1 ω = 0.1

ω = 0.5

ω = 1.0

0 m
0 30

First we consider the solid lines. They correspond to ω = 1. For the sequential
ordering ϕj = jπ/p, j = 1, . . . , p the values of ρm (ω) are displayed in the top
figure. We see that ρm (ω) is fairly large for low order subspaces Cm , indicating
slow convergence for low frequency parts of f , i.e. for the overall features of f .
The other figures shows the corresponding values of ρm (ω) for a non-consecutive
ordering of the ϕj suggested by Herman and Meyer (1993) and for a random
ordering respectively. Both orderings yield much smaller values of ρm (ω), in
 X-ray Tomography 31

particular on the low order subspaces. The conclusion is that for practical work
one should never use the consecutive ordering, and random ordering is usually
as good as any other more sophisticated ordering.
In particular in emission tomography the use of multiplicative iterative meth-
ods is quite common. The most popular of these multiplicative algorithms is the
expectation-maximization (EM) algorithm: Let Af = g be a linear system, with
all the elements of the (not necessarily square) matrix A being non-negative.
Assume that A is normalized such that A1 = 1 where 1 stands for vectors of
suitable dimension containing only 1 s. The EM algorithm for the solution of
Af = g reads
g
f j+1 = f j A∗ , j = 1, 2, . . . ,
Af j
where division and multiplication are understood componentwise. One can show
that f j converges to the minimizer of the log likelihood function

(f ) = (gi log(Af )i − (Af )i ) .

i

There exists also a version of the EM algorithm that decomposes the system
Af = g into smaller ones, Aj f = gj , j = 1, . . . , p, as in the Kaczmarz method
(ordered subset EM, OSEM; see Hudson and Larkin (1992)). As in the Kaczmarz
method one can obtain good convergence properties by a good choice of the
ordering, and the convergence analysis is very much the same.
For more on iterative methods in CT see Censor (1981).

3.3 Fourier Methods

Fourier methods make direct use of relations such as Theorem 2.1. Let’s consider
the 2D case in which we have

fˆ(σθ) = (2π)−1/2 ĝ(θ, σ) , g = Rf . (19)

In order to implement this we first have to do a 1D discrete Fourier transform

on g for each direction θj :

ĝ(θj , σ ) = (2π)−1/2 ∆s e−iσ sk g(θj , sk )

k

where sk = k∆s and σ = ∆σ. This provides fˆ at the points σ θj which

form a grid in polar coordinates. The problem is now to do a 2D inverse Fourier
transform on this polar coordinate grid. Various methods have been developed for
doing this, in particular the gridding method and various fast Fourier transform
on non-equispaced grids; see Fourmont (1999), Beylkin (1995), Steidl (1998).
32 F. Natterer

4 Reconstruction from Cone Beam Data

One of the most urgent needs of present day’s X-ray CT is the development
of accurate and efficient algorithms for the reconstruction of f from (Cf )(a, θ)
where a is on a source curve A (typically a helix) and θ ∈ S 2 . In practice, θ is
restricted to a subset of S0 , but we ignore this additional difficulty.
We assume that the source curve A satisfies Tuy’s condition, i.e. A intersects
each plane hitting supp (f ) transversally. Let A be the curve x = a(λ), λ ∈ Λ.
Tuy’s condition implies that for each s with (Rf )(θ, s) = 0 there is λ such that
s = a(λ) · θ. In fact there may be more than one such λ, but we assume that
their number is finite. Then we may choose a function M (θ, s) such that

M (θ, λ) = 1
λ
s=a(λ)·θ

for each s. Let g(λ, θ) = (Cf )(a(λ), θ) be the data function. Put


∂
G(λ, θ) = g(λ, ω)dω .
∂θ
θ ⊥ ∩S 2

Then we have

Theorem 4.1. Let w be a function in R1 and V = R∗ w . Then,

(V ∗ f )(x) = w((x − a(λ)) · θ)G(λ, θ)|a (λ) · θ|M (θ, λ)dλdθ .
S2 Λ

For the proof we start out from the 3D case of Theorem 3.1:



(V ∗ f )(x) = w (x · θ − s)(Rf )(θ, s)dsdθ
S 2 R1



= w(x · θ − s)(Rf ) (θ, s)dsdθ .
S 2 R1

In the s integral we make the substitution s = a(λ) · θ, obtaining

(V ∗ f )(x) = w((x − a) · θ)(Rf ) (θ, a(λ) · θ)|a (λ) · θ|M (θ, λ)dλdθ ,
S2 Λ

the factor M (θ, s) being due to the fact that λ → a(λ) · θ is not one-to-one. By
Theorem 2.8,
(Rf ) (θ, a(λ) · θ) = G(λ, θ) ,
and this finishes the proof.
 X-ray Tomography 33

Theorem 4.1 is the starting point for a filtered backprojection algorithm. For
1 
w = − δ ,
8π 2
δ the 1D δ-function, we have V = δ, the 3D δ-function, hence

 
x − a(λ)
f (x) = |x − a(λ)|−2 Gw λ, dλ ,
|x − a(λ)|
Λ


1
G (λ, ω) = − 2
w
δ  (ω · θ)G(λ, θ)|a (λ) · θ|M (θ, λ)dθ, ω ∈ S 2 .
8π
S2

This is the formula of Clack and Defrise (1994) and Kudo and Saito (1994). For
more recent work see Katsevich (2002).
There exist various approximate solutions to the cone-beam reconstruction
problem, most prominently the FDK formula for a circular source curve (which
doesn’t suffice Tuy’s condition). It can be viewed as a clever adaption of the 2D
filtered backprojection algorithm to 3D; see Feldkamp, Davis and Kress (1984).
An approximate algorithm based on Orlov’s condition of Theorem 2.7 is the π
method of Danielsson et al. (1999).

References
1. M. Abramowitz and I.A. Stegun (1970): Handbook of Mathematical Functions.
Dover.
2. G. Beylkin (1995): ‘On the fast Fourier transform of functions with singularities’,
Appl. Comp. Harm. Anal. 2, 363-381.
3. Y. Censor (1981): ‘Row–action methods for huge and sparce systems and their
applications’, SIAM Review 23, 444-466.
4. R. Clack and M. Defrise (1994): ‘A Cone–Beam Reconstruction Algorithm Using
Shift–Variant Filtering and Cone–Beam Backprojection’, IEEE Trans. Med. Imag.
13, 186-195.
5. P.E. Danielsson, P. Edholm, J. Eriksson, M. Magnusson Seger, and H. Turbell
(1999): The original π-method for helical cone-beam CT, Proc. Int. Meeting on
Fully 3D-reconstruction, Egmond aan Zee, June 3-6.
6. L.A. Feldkamp, L.C. Davis, and J.W. Kress (1984): ‘Practical cone–beam algo-
rithm’, J. Opt. Soc. Amer. A 6, 612-619.
7. K. Fourmont (1999): ‘Schnelle Fourier–Transformation bei nicht–äquidistanten
Gittern und tomographische Anwendungen’, Dissertation Fachbereich Mathematik
und Informatik der Universität Münster, Münster, Germany.
8. I.M. Gel‘fand, M.I. Graev, N.Y. Vilenkin (1965): Generalized Functions, Vol. 5:
Integral Geometry and Representation Theory. Academic Press.
9. P. Grangeat (1991): ‘Mathematical framework of cone–beam reconstruction via
the first derivative of the Radon transform’ in: G.T. Herman, A.K. Louis and
F. Natterer (eds.): Lecture Notes in Mathematics 1497, 66-97.
10. C. Hamaker, K.T. Smith, D.C. Solmon and Wagner, S.L. (1980): ‘The divergent
beam X-ray transform’, Rocky Mountain J. Math. 10, 253-283.
34 F. Natterer

11. C. Hamaker and D.C. Solmon (1978): ‘The angles between the null spaces of
X-rays’, J. Math. Anal. Appl. 62, 1-23.
12. S. Helgason (1999): The Radon Transform. Second Edition. Birkhäuser, Boston.
13. G.T. Herman (1980): Image Reconstruction from Projection. The Fundamentals of
Computerized Tomography. Academic Press.
14. G.T. Herman and L. Meyer (1993): ‘Algebraic reconstruction techniques can be
made computationally efficient’, IEEE Trans. Med. Imag. 12, 600-609.
15. H.M. Hudson and R.S. Larkin (1994): ‘Accelerated EM reconstruction using or-
dered subsets of projection data’, IEEE Trans. Med. Imag. 13, 601-609.
16. A.C. Kak and M. Slaney (1987): Principles of Computerized Tomography Imaging.
IEEE Press, New York.
17. A. Katsevich (2002): ‘Theoretically exact filtered backprojection-type inversion
algorithm for spiral CT’, SIAM J. Appl. Math. 62, 2012-2026.
18. H. Kudo and T. Saito (1994): ‘Derivation and implementation of a cone–beam
reconstruction algorithm for nonplanar orbits’, IEEE Trans. Med. Imag. 13,
196-211.
19. F. Natterer (1986): The Mathematics of Computerized Tomography. John Wiley &
Sons and B.G. Teubner. Reprint SIAM 2001.
20. F. Natterer and F. Wübbeling (2001): Mathematical Methods in Image Reconstruc-
tion. SIAM, Philadelphia.
21. F. Natterer and E.L. Ritman (2002): ‘Past and Future Directions in X-Ray Com-
puted Tomography (CT)’, to appear in Int. J. Imaging Systems & Technology.
22. R.G. Novikov (2000): ‘An inversion formula for the attenuated X-ray transform’,
Preprint, Departement de Mathématique, Université de Nantes.
23. V.A. Sharafutdinov (1994): Integral geometry of Tensor Fields. VSP, Utrecht.
24. G. Steidl (1998): ‘A note on fast Fourier transforms for nonequispaced grids’, Ad-
vances in Computational Mathematics 9, 337-352.
25. S. Webb (1990): From the Watching of Shadows. Adam Hilger.
Adjoint Fields and Sensitivities
for 3D Electromagnetic Imaging in Isotropic
and Anisotropic Media

Oliver Dorn1 , Hugo Bertete-Aguirre2 , and George C. Papanicolaou3

1
Gregorio Millán Institute of Fluid Dynamics, Nanoscience and Industrial
Mathematics, Universidad Carlos III de Madrid, Avda. de la Universidad 30,
28911 Leganés, Madrid, Spain
[email protected]
2
Department of Mathematics, University of California Irvine, Irvine, CA, USA
[email protected]
3
Department of Mathematics, Building 380, 383V, Stanford University, Stanford,
CA, USA
[email protected]

Summary. In this paper we give an overview of a recently developed method for

solving an inverse Maxwell problem in environmental and geophysical imaging. Our
main focus is on low-frequency cross-borehole electromagnetic induction tomography
(EMIT), although related problems arise also in other applications in nondestructive
testing and medical imaging. In typical applications (e.g. in environmental remedia-
tion), the isotropic or anisotropic conductivity distribution in the earth needs to be
reconstructed from surface-to-borehole electromagnetic data. Our method uses a back-
propagation strategy (based on adjoint fields) for solving this inverse problem. The
method works iteratively, and can be considered as a nonlinear generalization of the
Algebraic Reconstruction Technique (ART) in X-ray tomography, or as a nonlinear
Kaczmarz-type approach. We will also propose a new regularization scheme for this
method which is based on a proper choice of the function spaces for the inversion. A de-
tailed sensitivity analysis for this problem is given, and a set of numerically calculated
sensitivity functions for homogeneous isotropic media is presented.

1 Introduction: Electromagnetic Imaging of the Earth

The solid earth and certain pore fluids (such as hydrocarbons) are compara-
tively poor electrical conductors, while other pore fluids (such as brines) are
good electrical conductors. If we wish to image the shape of conducting fluid
plumes underground (as we might, e.g. for environmental cleanup purposes),
one practical way of doing so is to make either electrical or electromagnetic
measurements between boreholes (when two or more boreholes are available)
or between borehole and surface arrays. One popular electromagnetic method of
imaging without the need for boreholes is Ground Penetrating Radar (GPR), see
36 O. Dorn et al.

for example Davis and Annan [DA89] or Fisher et al. [FMA92]. Here the probing
waves are normally in the radio to microwave range of frequencies. GPR meth-
ods for imaging conductivity use comparatively high frequencies (≥ 100 MHz) in
order to have a signal that propagates through the Earth as a wave. The main
practical difficulty with GPR imaging is that at these frequencies its depth of
penetration is usually no deeper than about 5 m into the Earth in dry soil/sand,
while maximum depths of less than 1–2 m are more typical in wet earth due to
high local wave attenuation.
ElectroMagnetic Induction Tomography (EMIT) has been investigated re-
cently as a promising new tool for imaging electrical conductivity variations in
the earth (Wilt et al. [WAMBL, WMBTL] and Buettner et al. [BB99]) with
a depth of investigation greater than Ground Penetrating Radar (GPR). The
source is a magnetic field generated by currents in wire coils. This source field
is normally produced in one borehole, while the received signals are the mea-
sured small changes in magnetic field in another, more distant borehole (Fig. 1).
The method may also be used successfully in combination with surface sources
and receivers. The goal of this procedure is to image electrical conductivity
variations in the earth, much as X-ray tomography is used to image density
variations through cross-sections of the body. However, the fields tend to dif-
fuse rather than propagate as waves in this frequency band (from about 1 to
20 kHz) and the inversion techniques are therefore very different from those for
GPR. Although field techniques have been developed and applied previously to
collection of such EM data, the algorithms for inverting the magnetic field data
to produce the desired images of electrical conductivity have not kept pace. The
current state of the art in electromagnetic data inversion (Alumbaugh and Mor-
rison [AM95a, AM95b]) is based on the Born/Rytov approximation (requiring

source air
receivers

soil

Fig. 1. The figure shows a possible experimental setup for 3D-EMIT. Three boreholes
surround the area of interest. A z-directed dipole transmitter q is deployed in one of the
boreholes, and data are gathered in the other boreholes and at the surface at positions
dn , n = 1, . . . , N . From these data, we want to recover the (isotropic or anisotropic)
conductivity and permittivity distribution in the earth
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 37

a low contrast assumption), even though it is known that conductivity vari-

ations range over several orders of magnitude and therefore require nonlinear
analysis. Other work on inversion for this problem has been presented for ex-
ample by Habashy et al. [HGS93], Newman and Alumbaugh [NA97], and Haber
et al. [HAO00]. An alternative fully nonlinear approach to solving the EMIT
imaging problem was introduced recently by Dorn et al. [DBBP1, DBBP2]. The
current paper is mainly concentrating on this inversion technique. The ultimate
goal of our work is to produce better images of electrically conducting fluids
underground by using robust inversion techniques over greater distances and
higher contrasts than were previously possible.
Although other technologies, such as Electrical Resistance Tomography or
ERT (Berryman and Kohn [BK90]; Daily et al. [DRLN]; Ramirez et al. [RD-
LOC]; Borcea et al. [BBP96]; Borcea et al. [BBP99]) can produce electrical con-
ductivity images at a useful spatial scale, the advantage of EMIT is that we can
make use of existing monitoring wells and the surface to do our imaging. Since
signals are transmitted and received inductively, we do not need to make ground
contact (no ground penetrating electrodes); the technology is therefore relatively
noninvasive. Furthermore, all conductors present can in principle be imaged us-
ing the EMIT technique, whereas ERT can only image those conductors main-
taining a continuous electrical current path between the source and/or receiver
electrodes. There is also the important potential advantage for both techniques
that multiple frequencies can be employed to improve the imaging capability.
Our inversion method is based on the so-called ‘adjoint technique’, which (for
the present application) has the very useful property that the inverse problem
can be solved approximately by making two uses of the same forward modelling
code. Using a somewhat oversimplified description of our technique, the updates
to the electrical conductivity are obtained by first making one pass through the
code using the latest best guess of the nature of the conducting medium, and then
another pass with the adjoint operator applied to the differences in computed
and measured data. Then the results of these two calculations are combined to
determine updates to the original conductivity model. The resulting procedure
is iterative and can be applied successively to parts of the data, e.g. data associ-
ated with one transmitter location can be used to update the model before other
transmitter locations are considered. This procedure has several of the same ad-
vantages as wave equation migration in reflection seismology (Claerbout [Cla76])
and is also related to recent methods in electromagnetic migration introduced
by Zhdanov et al. [ZTB96].
An important and interesting question that is commonly raised about all
electrical imaging methods in the diffusive regime (frequencies f < 1 MHz) con-
cerns the issue of resolution. Our physical ideas of resolution are normally based
on concepts such as that of Rayleigh criterion for resolving power of an aper-
ture in which two point sources of radiation are considered to be resolvable
if the maximum of a diffraction pattern of one source coincides with the first
minimum of the other (see Jenkins and White [JW57]). The spatial separation
of two objects that can be resolved then clearly depends on the ratio of the
38 O. Dorn et al.

wavelength of the radiating sources and the size of the aperture. When we are
considering electromagnetic fields in the diffusive regime, it becomes problem-
atic to define a useful resolution criterion. At very low frequencies such as those
used in Electrical Resistance/Impedance Tomography (ERT/EIT) there is no
wavelength associated with the signals and the resolution must instead be de-
fined in terms of the numbers of source/receiver electrode pairs and the size of
the region being imaged (see Borcea [Bor02] or Ramirez et al. [RDLOC]). The
mathematically analogous problem in hydrology has also been treated (see Vasco
et al. [VDL97]). For EMIT, the situation is closer to that of ERT/EIT than it is
to that of wave propagation problem like seismic, ground penetrating radar, or
optics. The wavelength does not provide a useful measure of the method’s true
resolution. A better approach in this situation is to define sensitivity functions
(see for example Spies and Habashy [SH95]). A short introduction into sensitivity
analysis for inverse problems is given in the following.

2 Sensitivities and the Adjoint Method

Sensitivity relations for inverse problems (or, more generally, parameter estima-
tion problems) have been discussed many times in the literature. For example,
a well-known reconstruction method in medical imaging is the filtered back-
projection method in X-ray tomography [Nat86], which can be interpreted as
a backprojection of filtered X-ray data of the human body along sensitivity
functions. These sensitivity functions are concentrated on straight lines con-
necting the source and receiver positions, and the backprojection takes place
uniformly over these lines. A generalization of this procedure can be found in
Diffuse Optical Tomography (DOT) [Ar99], where the straight lines are replaced
by ‘banana-shaped’ regions, which are in fact sensitivity functions distributed
over the whole imaging domain with space-dependent weights. These weights are
highest in some (banana-shaped) regions connecting the source and the receiver
positions, but have nonvanishing values also in more remote areas of the imaging
domain. For more details see for example Arridge et al. [Ar95a, Ar95b, Ar99] or
Dorn [Dor00]. Both, the limited resolution usually encountered in the inverse
problem of DOT and the severe ill-posedness are directly related to the specific
shapes of these sensitivity functions.
The situation in low-frequency electromagnetic imaging is in many ways sim-
ilar to the DOT problem. The physical propagation of signals is in both cases
diffusive, which lets us expect that similar concepts of resolution should apply.
The sensitivity functions in low-frequency EMIT are, as in DOT, distributed
over the whole imaging domain. However, from energy arguments it follows that
sensitivities far away from sources and receivers are small.
There are also some characteristic differences between DOT and EMIT.
For example, in electromagnetic cross-borehole tomography, the sources and
receivers are typically not located at the boundaries of the imaging domain,
such that sidelobes of the sensitivity functions, which occur beyond the region
of interest, have to be taken into account in the imaging process. Moreover,
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 39

the propagating fields are now vector fields rather than scalar fields. This gives
us a richer variety of sensitivity functions, depending on the component of the
field which is used for collecting the data. We will see that the shapes of the
sensitivity functions in EMIT show more variations than in DOT. For certain
source-receiver geometries, the sensitivities in the region of interest are also here
mainly concentrated on some almost ‘banana-shaped’ regions, whereas for other
source-receiver constellations these regions look more like ‘elongated doughnuts’,
with almost vanishing sensitivities close to the connecting line of source and re-
ceiver. We mention that another related geophysical application, which EMIT
and DOT are interesting to compare with, is hydrology. That situation is treated
in more details for example in Vasco et al. [VDL97].
The nonlinear Kaczmarz-type approach for solving the inversion problem
(which can be characterized as a ‘single-step adjoint-field inversion method’)
will be presented further below in this paper. It can be viewed as a nonlinear
generalization of the algebraic reconstruction technique (ART) in X-ray tomogra-
phy [Nat86,NW01]. In contrary to the filtered backprojection method mentioned
earlier, ART uses only part of the data in each step of the reconstruction. It is
an iterative technique and applies successive corrections to some initial guess
for the parameter distribution, which in X-ray tomography is the attenuation.
Each of these corrections is calculated by ‘backprojecting’ a (preprocessed or
filtered) part of the residuals along those straight lines into the imaging domain
which correspond to the incoming direction of the probing X-rays. We will show
that, in an analogous way, each individual backprojection step of our generalized
ART algorithm for EMIT takes place along sensitivity functions. An important
difference of our generalized ART scheme to the classical ART algorithm is its
nonlinearity. Whereas the sensitivities in the classical ART method are always
concentrated on the same straight lines, the shapes of our sensitivity functions
depend on the latest best guess for the parameters and therefore change with
each update. The final reconstruction of our inversion method is given as a super-
position of sensitivity functions with weights determined by the residual values
which appear during the reconstruction process. The knowledge of the general
structure of the sensitivity functions during the reconstruction process provides
us with useful information about the resolution which we can expect from the
final reconstruction.
In this paper we make use of the so-called adjoint technique for calculat-
ing sensitivity functions. In the more traditional ‘direct method’ for calculating
sensitivities, typically as many forward problems need to be solved as there
are parameters to be reconstructed. In large scale inverse problems these are
typically several thousand parameters, one for each pixel or voxel used for
the forward modelling. It has been pointed out in the literature that in such
large scale parameter estimation problems the adjoint method can be much less
computationally expensive than this ‘direct method’ for calculating sensitivi-
ties [Ar95a, Ar95b, AP98, Dor00, FO96, HAO00, MO90, MOEH, Nat96]. This is
in particular true in cases where the Green’s functions of the underlying for-
ward problem (for example Maxwell’s equations) are not known analytically,
40 O. Dorn et al.

xsc xsc

xsc xsc xd
xs xd xs

xsc xsc

direct form adjoint form

Fig. 2. Feynman diagrams visualizing direct and adjoint forms of sensitivity functions.
xs source position, xd receiver position, xsc scattering points

and have to be evaluated by employing some numerical forward modelling code.

For example, given a source and a receiver position for which we want to calcu-
late the space-dependent sensitivity functions in a 3D EMIT inversion problem,
the ‘direct method’ would require us to solve 105 –106 forward problems in order
to find one sensitivity function. Using the ‘adjoint method’ we only need to solve
one forward and one adjoint problem in order to calculate the same sensitivity
function. The Feynman diagrams in Fig. 2 visualize this relationship. Each knot
in the figure where one or more arrows originate indicates a separate run of the
forward modelling code. Therefore, in the simple situation displayed in Fig. 2,
we would have to solve six forward problems in order to calculate one sensitivity
function in the direct form, whereas the adjoint scheme only requires two runs
of our forward modelling code.
Sensitivity functions are closely related to Fréchet derivatives, which are eval-
uated at some reference parameter distribution. They describe the amount of
which the receiver measurements will change if we slightly perturb the given
reference parameter distribution at a certain position in the medium. Our lin-
earized residual operator is just the Fréchet derivative of the nonlinear residual
operator. We want to mention, however, that actually proving Fréchet differen-
tiability of the nonlinear residual operator is not an easy task, and is beyond
the scope of this overview. A possible method of proving Fréchet differentiability
for general bilinear inverse problems can be found in [DDNPS]. We also men-
tion that Fréchet differentiability for a related 3D-Maxwell problem has been
discussed recently in [AB01].
Assuming that our residual operator is differentiable, it can be shown (and
we will do that further below) that the sensitivity functions decompose this lin-
earized residual operator as well as its adjoint operator. In many applications, in
particular if the starting point is the discretized model of the forward problem,
this Fréchet derivative is called the ‘Jacobian’ or ‘sensitivity matrix’ of the un-
derlying forward operator. The sensitivity functions can then be interpreted as
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 41

the rows (or columns) of the Jacobian or sensitivity matrix. They play a major
role in basically all gradient-based inversion methods which try to minimize
the least squares measure of the misfit between given data and predicted data
corresponding to a reference medium. Therefore, the ideas presented here are
certainly applicable also beyond the single step adjoint field inversion method
which we are concentrating on in this paper.

3 Maxwell’s Equations for Anisotropic Media

The system of time-harmonic Maxwell’s Equations in three spatial dimensions

can be formulated as follows

∇ × H(x) − (iω(x) + σ(x)) E(x) = J(x), (1)

∇ × E(x) + iωµ(x) H(x) = M(x), (2)

∇ · E(x) = 0, ∇ · µH(x) = 0, (3)

where E(x) and H(x) denote the electric and magnetic field, J(x) and M(x)
are the electric and magnetic
√ source currents, ω > 0 is the angular frequency,
ω = 2πf , and i = −1. The parameters describing the underlying medium
are the electric conductivity σ(x) with dimension Siemens per metre (S m−1 ),
the dielectric permittivity (x) with dimension Farad per metre (F m−1 ), and
the magnetic permeability µ(x) with dimension Henry per metre (H m−1 ). We
assume that all these parameters are strictly positive in the physical domain
σ(x) > 0, (x) > 0, and µ(x) > 0 in R3 . Equations (3) indicate that there are
no electric or magnetic space charges in the medium.
In the case of isotropic media the physical parameters are described by (com-
plex valued) scalar functions. For anisotropic media these have to be replaced
by general 3 × 3 tensor functions. We will assume here without loss of generality
that σ(x), (x), and µ(x) are described by diagonal tensors with real positive
numbers on the diagonal and zero off-diagonal elements [DBBP2]. We get the
representations

σ(x) = diag(σ (x) (x), σ (y) (x), σ (z) (x)),

(x) = diag((x) (x), (y) (x), (z) (x)), (4)
(x) (y) (z)
µ(x) = diag(µ (x), µ (x), µ (x)).

For the following, we introduce the notation

a = iωµ, b = iω + σ, (5)

with

a(x) = diag(a(x) (x), a(y) (x), a(z) (x)),

b(x) = diag(b(x) (x), b(y) (x), b(z) (x)).
42 O. Dorn et al.

With the notation of (5) we write (1), (2) as

∇ × H(x) − b(x) E(x) = J(x), (6)

∇ × E(x) + a(x) H(x) = M(x). (7)

This is the form of Maxwell’s equations that we will use.

Equations (6), (7) describe the propagation of an electromagnetic field which
is generated by a (time-harmonic) source distribution
 
iωt J(x)
q̃(x, t) = q(x)e = eiωt ,
M(x)
where we typically will apply a magnetic dipole source of the form
 
0
q(x) = . (8)
ms δ(x − s)
Here, ms is the magnetic dipole moment of the source and s the source position.
In the following we consider (6), (7) in the unbounded domain Ω = R3 for a
given source (8). In Sect. 6, we will extend the analysis to the case when data
corresponding to many source positions are used.
We assume that (6), (7) provide a model for the propagation of the electric
and magnetic field emitted by the source q, and that we measure the magnetic
field H(dn ) for N discrete detector positions dn , n = 1, . . . , N .
In geophysical applications, it is often the case that the coefficient a does not
vary significantly in the region of interest. In this paper, we will therefore assume
that a is known in the whole domain Ω, e.g. from interpolation of well-log data.
In our numerical experiments, we will furthermore only consider the case that
a is constant everywhere, although the discussion of the paper applies to the
more general situation of space dependent parameters a(x). The coefficient b is
assumed to be unknown in the medium and has to be recovered from the data.
The inverse problem we consider here is therefore as follows.
Inverse Problem: Assume that for many source distributions qj = (Jj , Mj )T ,
j = 1, . . . , p, for example of the form (8), the corresponding magnetic fields
Hj (dn ) are measured at the receiver locations dn , n = 1 . . . , N . Find a parameter
distribution b̃ such that for the solutions (Ẽj , H̃j )T of

∇ × H̃j − b̃ Ẽj = Jj , (9)

∇ × Ẽj + a H̃j = Mj , (10)

the magnetic field values at the receiver positions coincide with the data.
If there is no noise in the data, then there may be many parameter distri-
butions b̃ for (9), (10) that satisfy the data, so we have to select one using a
regularization criterion. If there is noise, the usual case, then we need to min-
imize the sum of the squares of the differences between the data calculated by
(9), (10) and those which are physically measured.
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 43

4 The Linearized Residual Operator

We want to define now the basic function spaces which are used throughout this
paper. The main objective of this section is to introduce the inner products which
will be used in the derivation of the adjoint of the linearized residual operator
which is discussed in Sect. 5.1. For simplicity in the notation, we will assume in
the present section that we have only one source function q(x) given which is
fixed. In Sect. 6 we will come back to the situation where more than one physical
experiment are performed for gathering the data.
The basic spaces which we will need are as follows.
   
F = b = diag b(x) , b(y) , b(z) ; b(x) , b(y) , b(z) ∈ L2 (Ω) ,
 
Z = ζ = (h1 , . . . , hN ), hn ∈ C3 ,
  E(x)  
U = ; E, H ∈ [L2 (Ω)]3 ,
H(x)
  J(x)  
Y = ; J, M ∈ [L2 (Ω)]3 .
M(x)

Notice that in the case of isotropic media we have b(x) = b(y) = b(z) = b such
that we typically will identify in that situation F = L2 (Ω) with b ∈ F . It will
always be clear from the context which space we refer to, such that we omit
using different symbols for this space in the cases of isotropic and anisotropic
media. All these quantities carry a physical dimension, which has to be taken
into account when introducing the following inner products for these spaces.
   


E1 E2 2 2
, = γE E1 E2 dx + γH H1 H2 dx,
H1 H2 U Ω Ω
   


J1 J2 2 2
, = γJ J1 J2 dx + γM M1 M2 dx,
M1 M2 Y Ω Ω
N N
(i) (i)
ζ1 , ζ2 Z = 2
γH h1,n h2,n = 2
γH h1,n h2,n
n=1 n=1 i=x,y,z

with ζl = (hl,1 , . . . , hl,N ), ĥl,n := γH hl,n ∈ C3 for l = 1, 2, n = 1, . . . , N .

  (x) (y) (z)   (x) (y) (z) 
diag b1 , b1 , b1 , diag b2 , b2 , b2
F

 
(x) (x) (y) (y) (z) (z)
= γb2 b1 (x)b2 (x) + b1 (x)b2 (x) + b1 (x)b2 (x) dx.
Ω

An analogous expression holds for the inner product of F = L2 (Ω) in the case
of isotropic media. The constants γE , γH , γJ , γM and γb are introduced in
order to make the given scalar products dimensionless. More precisely, we have
γE = [E]−1 , γH = [H]−1 , γJ = [J]−1 , γM = [M]−1 , and γb = [b]−1 , where the
44 O. Dorn et al.

symbol [.] denotes the dimension of the corresponding physical quantity. Notice
that γb γE = γJ , which follows immediately from (6). We will also consider in
this paper the corresponding dimensionless quantities which we always indicate
by a roof on top of the symbol, for example b̂ = γb b for b ∈ F or ĥ = γH h for
h ∈ Z. For these quantities, we will accordingly use the inner products  , F̂ ,
 , Ẑ , . . . , which are defined as above but without the additional factors in the
integrals. The spaces F , U , Y and Z (as well as F̂ , Û , Ŷ and Ẑ) are Hilbert
spaces with these inner products.
We write (6), (7) in system form

ΛM (b) u = q, (11)

with
⎛ ⎞
−b ∇×
⎜ ⎟
⎜ ⎟
ΛM (b) = ⎜ ⎟, (12)
⎝ ⎠
∇× a
and
   
E J
u = , q = .
H M
The Maxwell Operator ΛM (b) is defined on a dense subset of the Hilbert space
U, given the parameter distribution b. When the conductivity is everywhere
positive, σ > 0, (11) has a unique solution u ∈ U for each source q ∈ Y . For a
function u ∈ U of (11) we define the measurement operator M : U → Z by


(M u)n (x) = δ(x − dn ) H(x) dx (n = 1, . . . , N ) (13)
Ω
 
E
with u = H , and with dn , n = 1, . . . , N , being the detector positions. The
measurement operator M is well defined by (13) and maps the fields u ∈ U to
a vector of complex numbers M u ∈ Z, which consists of point measurements of
the magnetic field components at the detector positions dn , n = 1, . . . , N . More
precisely, we can write
 
E T
M = ( H(d1 ) , H(d2 ) , . . . , H(dN ) ) ,
H

where the symbol T stands for transpose. For the given source q and given data
g ∈ Z we define the residual operator R by

R : F → Z, R(b) = M u(b) − g,

where u(b) is the solution of (11) with parameter b. The operator R is a nonlinear
operator from the space of parameters F to the space of measurements Z.
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 45

We will assume that the source q is given by (8) with

ms = α1 e1 + α2 e2 + α3 e3 , (14)

with α̂i := γM αi ∈ C for i = 1, 2, 3. The symbols ei , i = 1, 2, 3, will always

denote the cartesian unit vectors, for example e1 = (1, 0, 0)T . We will also often
replace i = 1, 2, 3 by i = x, y, z in the notation in order to emphasize the spatial
dimension which we refer to. We denote by E(x) = (E (x) (x), E (y) (x), E (z) (x))T
(and similarly H(x)) the solution of the forward problem
   
E 0
ΛM (b) = (15)
H ms δ(x − s)

with ΛM (b) defined as in (12), and a, b defined as in (5). In order to find a useful

representation of the linearized residual operator R of R, we perturb b, E and
H by
b → b + δb E →E+v H → H + w,
and plug this into (6), (7). When neglecting terms of higher than linear order
in δb, v and w, and applying the measurement operator M , we arrive at the
following result.

Theorem 4.1. The linearized residual operator R [b] is given by

v
R [b] δb = M , (16)
w
v
where w solves
v  
δb E
ΛM (b) = . (17)
w 0

5 Adjoint Sensitivity Analysis for Anisotropic Media

5.1 The Adjoint Linearized Residual Operator

The adjoint linearized residual operator R [b]∗ is formally defined by the identity
    

R [b] δb , ζ = δb , R [b]∗ ζ .
Z F

It is a mapping from the data space Z into the parameter space F . In the next

theorem, we will present a useful expression for the action of R [b]∗ on a given
vector ζ ∈ Z. The proof of this theorem can be found in [DBBP2].
Theorem 5.1. Let us assume that we are given ζ = (h1 , . . . , hN ) ∈ Z with
(x) (y) (z)
ĥn = (ĥn , ĥn , ĥn )T ∈ C3 for n = 1, . . . , N . Furthermore, we denote by
E(x) = (E (x), E (x), E (z) (x))T the solution of the adjoint problem
(x) (y)
46 O. Dorn et al.
  
E 0
Λ∗M (b) = −1 N  (i) , (18)
H γM n=1 i=x,y,z ĥn ei δ(x − dn )

with ⎛ ⎞
−b ∇×
Λ∗M (b) = ⎝ ⎠, (19)
∇× a

b = −iω + σ, a = −iωµ.

∗
Then R [b] ζ ∈ F is given by
    
R [b]∗ ζ (x) = γb−1 diag Ê (x) (x)Ê (x) (x) , Ê (y) (x)Ê (y) (x) , Ê (z) (x)Ê (z) (x) ,
(20)
where E is the solution of (15).

5.2 Decomposition of the Residual Operator and Its Adjoint

In this section, we will present the basic decompositions of the linearized residual
operator and its adjoint into sensitivity functions. The proofs of the theorems in
this section and in the next section can be found in [DBBP2].
(i)
Theorem 5.2. Given a source (14). We can find functions Sk (n, l; y), i =

1, 2, 3, such that the linearized residual operator R [b] can be decomposed as
 3

   3  3 
R [b]δb = γ −1 ! (i) (y)
δb
(i)
α̂k S (n, l; y) dy el . (21)
H k
n Ω
l=1 i=1 k=1

(i)
Moreover, the same functions Sk (n, l; y) also decompose the adjoint linearized

residual operator R [b]∗ in the following way
 3
  (i) N 3
∗ −1 (l) (i)
R [b] ζ (y) = γb ĥn α̂k Sk (n, l; y) . (22)
n=1 l=1 k=1

(i)
The functions Sk (n, l; y) are called anisotropic sensitivity functions.
(i)
In the following, we will first define these sensitivity functions Sk (n, l; y),
and then reformulate the above theorem in a slightly more formal way. First, we
decompose the given vector ζ = (h1 , . . . , hN ) ∈ Z as

hn = h(1) (2) (3)

n e1 + hn e2 + hn e3 (23)
(i) (i)
with ĥn := γH hn ∈ C for i = 1, 2, 3 and n = 1, . . . , N . Let (Ek , Hk )T , k =
1, 2, 3, be the solutions of the direct problems
   
Ek 0
ΛM (b) = −1 . (24)
Hk γM ek δ(x − s)
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 47

Moreover, let (En,l , Hn,l )T , l = 1, 2, 3, n = 1 . . . , N , be solutions of the adjoint

problems    
∗ En,l 0
ΛM (b) = −1 , (25)
Hn,l γM el δ(x − dn )
where Λ∗M (b) is defined as in (19). With this notation, the anisotropic sensitivity
(i)
functions Sk (n, l; y) are defined by

(i) (i) (i)

Sk (n, l; y) = Êk (y) Ên,l (y), (26)

where the fields (E, H)T are solutions of (15). Notice that these sensitivity func-
tions are defined to be dimensionless.
Now, making use of the inner products which have been introduced in Sect. 4,
we get the following equivalent formulation of Theorem 5.2.
(i)
Theorem 5.3. Given a source (14). Define the functions Sk (n, l; y) as in (26).
Then we have the following decompositions of the linearized residual operator
 
R [b] and its adjoint R [b]∗ .
Projection Step:
 3 3  
   3
R [b]δb = γ −1 ! (i) , S (i) (n, l; . )
α̂k δb el . (27)
H k
n F̂
l=1 i=1 k=1

Backprojection Step:
 
(i) 3  
∗
γb−1
(i)
R [b] ζ (y) = α̂k ζ̂ , Sk ( . , . ; y) . (28)
Ẑ
k=1


Remark 5.1. We note that (27) gives rise to the interpretation of R [b] as a ‘pro-
jection operator’. It ‘projects’ each component δb(i) of the parameter perturba-
(i)
tion δb along the sensitivity function Sk (n, l; . ) into the data space Z. Summing

up all these contributions yields the (linearized) residuals R [b]δb. Analogously,

(28) gives rise to the interpretation of R [b]∗ as a ‘backprojection operator’. It
‘backprojects’ the residual vector ζ of the data space Z along the sensitivity
(i)
functions Sk ( . , . ; y) into the parameter space F .

5.3 A Special Case: Isotropic Media

Let us consider an isotropic medium, i.e. let biso (y), δbiso (y) be given by
 
biso (y) = diag b(0) (y), b(0) (y), b(0) (y) ,
 
δbiso (y) = diag δb(0) (y), δb(0) (y), δb(0) (y) .
48 O. Dorn et al.

We can replace the diagonal tensors for describing the parameters δbiso and
biso by the scalar functions δb(0) , b(0) , which live in the reduced function space
L2 (R3 ). The isotropic sensitivity functions Sk (n, l; y) are defined as

Sk (n, l; y) = Êk (y) · Ên,l (y), (29)

where Ek (y) solves (24), E(y) solves (15), and En,l (y) solves (25). Obviously,
we have
3
(i)
Sk (n, l; y) = Sk (n, l; y). (30)
i=1
With these functions, we can formulate the analogous form of Theorem 5.2 for
isotropic media.
Theorem 5.4. Given a source (14). Let the functions Sk (n, l; y) be defined as

in (29). Then we have the following decompositions of R [b]
 3

   3
R [b]δb = γH−1
α̂k !
δb(y)S k (n, l; y) dy el . (31)
n Ω
l=1 k=1

In addition, R [b]∗ can be decomposed as
 
 N 3 3
R [b]∗ ζ (y) = γb−1 ĥ(l)
n α̂k Sk (n, l; y). (32)
n=1 l=1 k=1

Using a more formal notation, this theorem can be reformulated as follows.

Theorem 5.5. Given a source (14). Define the functions Sk (n, l; y) as in (29).
 
Then, the linearized residual operator R [b] and its adjoint R [b]∗ have the fol-
lowing decompositions.
Projection step:
  3 3  

R [b]δb −1
= γH ! , Sk (n, l; . )
α̂k δb el . (33)
n F̂
l=1 k=1

Backprojection step:
 
 3  
R [b]∗ ζ (y) = γb−1 α̂k ζ̂ , Sk ( . , . ; y) . (34)
Ẑ
k=1

6 A Single Step Adjoint Field Inversion Scheme

6.1 Outline of the Inversion Scheme
In this section we will give a brief outline of the adjoint field inversion scheme for
solving the 3D EMIT problem. More details can be found in [DBBP1, DBBP2].
It was already mentioned earlier that the method can be considered as a non-
linear generalization of the algebraic reconstruction technique (ART) in X-ray
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 49

tomography [Nat86], and it is also similar to the simultaneous iterative recon-

struction technique (SIRT) as presented in [DL79] or to the Kaczmarz method
for solving linear systems of equations (see also [Nat86, NW01, NW95, Dor98]).
We will assume for simplicity that the medium is isotropic. Extensions to the
case of anisotropic media are straightforward. When solving the inverse problem
which was formulated in Sect. 3, we ideally want to find a parameter b̃ such that
for all given source positions qj , j = 1, . . . , p, the residuals vanish

Rj (b̃) = 0, j = 1, . . . , p. (35)

The index j refers here to the actual source position qj which is considered by
the operator Rj (b̃). Using real data, we cannot be sure whether these equations
can be satisfied exactly. Therefore, we generalize our criterion for a solution.
Defining the least squares cost functionals
1
Jj (b) = Rj (b)2L2 (36)
2
for j = 1, . . . , p, we are searching for a joint minimizer of these cost functionals.
A standard method for finding a minimizer of (36) is to start a search with some
initial guess b(0) , and to find descent directions for the sum
p
J (b) = Jj (b)
j=1

in each step of an iterative scheme. Popular choices for descent directions are
for example the gradient direction, conjugate gradient directions, Newton- or
Quasi-Newton directions (see for example [DS96, NW99, Vog02] for details). In
our Kaczmarz-type approach, however, we consider only individual terms of (36)
in a given step. We search for descent directions only with respect to the chosen
term Jj (b), and apply it to the latest best guess. In the next step, we will then
choose a different term Jj (b) of (36) and proceed in the same way. After having
used each of the terms in Jj (b) exactly once, we have completed one sweep of
our algorithm.
This strategy of switching back and forth between different data subsets
does not only have the advantage that a correction of our latest best guess
for the parameters can be achieved more rapidly, but it also helps to avoid
getting trapped in so-called ‘local minima’ of more classical gradient-based search
algorithms. This is because the variety of search criteria is increased in our
algorithm. Instead of having just one search direction available per data set
(considering only descent directions of J (b)), we have now as many useful search
directions as we have source positions. Even if one of these search directions does
not advance our reconstruction, typically at least one of the remaining search
directions gives us an improvement of it. Notice also that the combined cost
J (b) is certainly allowed to increase for a while during the reconstruction using
the Kaczmarz-type scheme. This will for example happen when a gradient-based
inversion algorithm with respect to J (b) would get close to a local minimum,
50 O. Dorn et al.

but at least one of the individual Kaczmarz-type search directions is not affected
by this local minimum of the combined least-squares cost functional.
We will in the following derive our Kaczmarz-type update directions for re-
ducing the individual terms (36) using the formulation (35). These update direc-
tions will have the useful property that they can easily be generalized in order
to incorporate efficient regularization schemes in our algorithm.
In order to find an ‘update’ (or ‘correction’) δb for our parameter b we lin-
earize the nonlinear operator Rj (assuming that this linearized operator Rj [b]
exists and is well-defined) and write

Rj (b + δb) = Rj (b) + Rj [b]δb + O(||δb||2 ). (37)

The linearized operator Rj [b] is often called the Fréchet derivative of Rj at b.
(See for example [DDNPS] and references therein for some formal derivations
of Fréchet derivatives in different applications.) It is also closely related to the
‘sensitivity functions’ of the parameter profile with respect to the data. Using
(37) we want to look for a correction δb such that Rj (b + δb) = 0. Neglecting
terms of order O(||δb||2 ) in (37), this amounts to solving

Rj [b]δb = −Rj (b) . (38)

Certainly, due to the ill-posedness of our problem, this equation needs to be

handled with care. Treated as an ill-posed linear inverse problem, a classical
solution of (38) will be the minimum-norm solution

δbM N = −Rj [b]∗ (Rj [b]Rj [b]∗ )−1 Rj (b), (39)

where Rj [b]∗ is the adjoint operator of Rj [b] with respect to our chosen spaces
F and Z. In applications with very few data, this form has the useful property
that it avoids contributions in the solution which are in the (often non-empty)
null-space of the (linearized) forward operator Rj [b]. Using (37) it can be verified
by direct calculation that

Jj (b + κδbM N ) = Jj (b) − κRj (b)2Z + O(κδbM N 2F ) (40)

for a step-size κ > 0, such that (39) also is a descent direction of the least squares
cost functional (36).
In our application the operator C = (Rj [b]Rj [b]∗ )−1 is very ill-conditioned,
such that a regularized version needs to be used. This can be, for example,
Ĉ = (Rj [b]Rj [b]∗ + λI)−1 where λ is some regularization parameter and I is the
identity operator. Unfortunately, in practice both, C and Ĉ, are very expensive
to calculate and to apply to the residuals Rj . Typically, a direct calculation of
the operator Ĉ would require us to solve as many forward and adjoint problems
as we have independent data values.
When using a very large regularization parameter λ, the contribution of
Rj [b]Rj [b]∗ can be neglected and we end up with essentially (i.e. up to the
scaling factor λ−1 ) calculating
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 51

δb = −Rj [b]∗ Rj (b). (41)

For this update direction we have

Jj (b + κδb) = Jj (b) − κRj [b]∗ Rj (b)2F + O(κδb2F ) (42)

such that it is also a descent direction for (36).

Regardless which of the update formulas (39) or (41) we use, we will need to
apply the adjoint linearized residual operator Rj [b]∗ to some vector in the data
space. In the following we will derive and test efficient schemes for doing this
(the basic propagation–backpropagation scheme), and moreover we will derive a
new regularization scheme for this backpropagation technique.
A standard method for deriving regularization schemes is to explicitly try
to minimize a cost functional which incorporates, in addition to the usual least
squares data misfit, a Tikhonov–Phillips regularization term:
η
JT P (b) = J (b) + b2α , (43)
2
or, in a single step fashion
η
JT P,j (b) = Jj (b) + b2α (44)
2
where η > 0 is the regularization parameter and  . α indicates some norm or
semi-norm, e.g. bα = ∇bL2 [Vog02]. Using this approach, the cost functional
is changed significantly with the goal of obtaining in a stable way a global min-
imizer. We do not want to take this route, but prefer instead to keep working
with the original least-squares cost functional (36) which only involves the data
fit. We will minimize this cost functional by restricting the search to elements
of a smaller function space, which is an alternative form of regularization.
The regularization scheme will be derived and discussed in details in Sect. 7.
In the following, we will present the basic structure of our inversion method, and
we will derive practical ways of applying the adjoint linearized residual operator
Rj [b]∗ to vectors Rj in the data space Z. This will lead us to the propagation–
backpropagation technique which is considered in this paper.

6.2 Efficient Calculation of the Adjoint of the Linearized Residual

Operator

In this section we want to show how the adjoint residual operator is applied in
the case where many receiver positions are used simultaneously for a given source
and where the medium is assumed to be possibly inhomogeneous but isotropic.
A formal derivation of this result can be found in [DBBP1].
Let Ĉj denote some suitably chosen approximation to the operator Cj , which
might depend on the source index j. For the inversion, we will need some efficient
way to compute 
Rj (b)∗ ζj
52 O. Dorn et al.

for a given data vector ζj = Ĉj Rj (b) ∈ Z. Let

ζj = ( hj,1 , . . . , hj,N ) ∈ Z (45)

be the data vector, which in the given step j of the inversion scheme usually con-
sists of the (filtered) residual values of all receiver positions which correspond to
the source qj . We will in the following omit all dimensional factors (for example

γH ) to simplify the notation. Then Rj (b)∗ ζj ∈ F is given by
"  ∗ #
Rj (b) ζj (x) = Ej (x) · Ej (x), (46)

where Ej (x) is the solution of the direct problem

 
Ej
ΛM (b) = qj , (47)
Hj

and Ej (x) solves the corresponding adjoint problem

   
Ej 0
Λ∗M (b) = N , (48)
Hj n=1 hj,n δ(x − dn )

with ⎛ ⎞ ⎛ ⎞
−b ∇× −b̄ ∇×
ΛM (b) = ⎝ ⎠, Λ∗M (b) = ⎝ ⎠, (49)
∇× a ∇× ā

b = iω + σ, b̄ = −iω + σ,
(50)
a = iωµ, ā = −iωµ.
We see that we only have to solve one forward problem (47) and one adjoint

problem (48) in order to calculate Rj (b)∗ ζj in an efficient way (compare Fig. 2).

6.3 Iterative Algorithm for the Adjoint Field Inversion

We will now describe in detail the algorithm which we use in the numerical
experiments. In brief algorithmic form, the iterative scheme can be written as
(0)
bp = b(0)
sweep loop: DO i = 1, Imax
(i) (i−1)
b0 = b p
source loop: DO j = 1, p

= −Rj (bj−1 )∗ Ĉj Rj (bj−1 )
(i) (i) (i)
δbj
(i) (i) (i)
bj = bj−1 + κj δbj
END DO source loop
END DO sweep loop
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 53

Here, b(0) is some initial guess for b and Imax is the total number of sweeps.
The parameter κj is the step-size in step j, and Ĉj the chosen approximation to
Cj . For example, b(0) can be taken to be the correct layered background parame-
ter distribution. The aim will then be to detect some unknown inclusions which
are imbedded in this layered background medium, see for example [DBBP1].
Notice that in one single step of the algorithm we have to solve one forward
(i)
problem (47) to compute Rj (bj−1 ), and one adjoint problem (48) to compute
(i) (i)
δbj from a given Rj (bj−1 ).
We mention that, in the algorithm presented so far, no explicit regularization
has been applied. If there is a sufficient amount of independent data available,
this algorithm already yields good reconstructions. However, for practical ap-
plications it would be desirable to have an efficient toolbox of regularization
schemes available which can be used for stabilizing the inversion in situations
with only very few data, and in order to incorporate prior information into the
reconstructions. In the following section, we propose such an efficient and flexi-
ble regularization tool. It has been tested and evaluated numerically with great
success in a different but related geophysical application in [GKMD]. The nu-
merical evaluation of this scheme in the situation of 3D EMIT still needs to be
done, and is planned as part of our future research.

7 Regularization and Smoothing with Function Spaces

We have presented above the basic algorithm which recovers L2 functions of
parameters from given data such that the misfit in the data is minimized. This
procedure does not incorporate explicit regularization (except of the stabilizing
procedure incorporated in the operator Ĉ). In some situations, it might be nec-
essary or desirable to restrict the search for parameter functions to a smaller
subset of L2 , for example of smoothly varying functions. This might be so in
order to regularize the reconstruction algorithm, or in order to take into account
some prior information or assumptions on the solution we are looking for. For
example, the field engineer might know or assume that the parameter distribu-
tion in some region is fairly smoothly varying. Or, he might only have very few
data available for the inversion, so that he wants to select a smoothly varying
profile as a regularized form of the reconstructed parameter distribution. This
can be easily done in our framework.
Instead of looking for parameter distributions in L2 (Ω), let us assume now
that we require the parameter to be an element of the smaller subspace

H1 (Ω) := {m ∈ L2 (Ω), ∂i m ∈ L2 (Ω) for i = 1, 2, 3} .

This Sobolev space is usually equipped with the standard norm

 1/2
m1,1 := m2L2 + ∇m2L2

and the standard inner product

54 O. Dorn et al.

m1 , m2 1,1 := m1 , m2 L2 + ∇m1 , ∇m2 L2 .

For reasons explained below, we will instead prefer to work with the equivalent
norm  1/2
mα,β := αm2L2 + β∇m2L2 , α, β > 0
and its associated inner product

m1 , m2 α,β := α m1 , m2 L2 + β ∇m1 , ∇m2 L2 .

A proper choice of the weighting parameters α and β will allow us to steer the
regularization properties of our algorithm in an efficient and predictable way.
Let us denote the new parameter space H1 (Ω), when equipped with the
weighted norm  . α,β , by F̂ . When using this modified space in our algorithm,
we also have to adjust the operators acting on it, in particular the adjoint of
the linearized residual operator. This operator is now required to map from the
data space Z into F̂ . Moreover, the minimum norm solution of (38) is now taken
with respect to the weighted norm  . α,β , which clearly gives us a different
candidate. The necessary adjustments for our algorithm can be done as follows.
Denote as before by Rj [b]∗ ζ the image of ζ ∈ Z under application of the
adjoint linearized residual operator as calculated in Sect. 6.2, considered as an
operator mapping from Z into F . Denote furthermore by Rj [b]◦ ζ its image un-
der the adjoint linearized residual operator with respect to the newly defined
weighted inner product, mapping into the smaller space F̂ . With a straightfor-
ward calculation, using the definitions of the two adjoint operators
  
Rj [b]x, ζZ = x, Rj [b]∗ ζF = x, Rj [b]◦ ζF̂ , (51)

it follows that
−1
Rj [b]◦ ζ = (αI − β∆) Rj [b]∗ ζ, (52)
−1
where we supplement the inverted differential operator (αI − β∆) by the
boundary condition ∇(Rj [b]◦ ζ) · n = 0 on ∂Ω. The symbol I stands for the
identity, and ∆ stands for the Laplacian operator. Equation (52) can be easily
derived by applying Green’s formula to the right hand side equality in (51).
In practice, the ratio γ = β/α (which can be considered being a ‘regular-
ization parameter’) is an indicator for the ‘smoothing properties’ of our scheme.
The larger this ratio, the more weight is put on minimizing the derivatives of our
solution. Therefore, by properly choosing this ratio, we can steer the smoothness
properties of our final reconstruction to a certain degree. In our numerical ex-
periments, we will choose this ratio once, when starting the algorithm, and keep
it fixed during the iterations. The other free parameter, say α, will be chosen in
each individual step to scale the update properly. In our numerical experiments,
we choose α such that

Rj [b]◦ ζL2 = Rj [b]∗ ζL2

is satisfied for the current update. This possibility of scaling the updates is the
main reason for keeping the parameter α throughout the calculations instead of
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 55

simply putting it to 1 right at the beginning. When testing and comparing the
performance of different regularization parameters γ it is practically useful that
the order of magnitude of the calculated values of Rj [b]◦ ζ does not depend too
much on γ.
Notice also that the new search directions using this modified adjoint operator
are still descent directions for the least squares cost functional (36), as can be
verified easily by replacing F by F̂ in (40) and (42).
Practically, the scheme is implemented as follows:

γ is fixed regularization parameter

Define Ψ = Rj [b]∗ ζ.
Solve (I − γ∆)ϕ = Ψ, ∇ϕ · n = 0 on ∂Ω.
ϕ
Define Φ = ϕ
α with α = Ψ (such that Φ = Ψ )
Then we have (αI − β∆)Φ = Ψ, ∇Φ · n = 0 on ∂Ω, with β = αγ.
Put Rj [b]◦ ζ = Φ. (53)
We mention that applying this regularization scheme amounts to applying the
−1
postprocessing operator (αI − β∆) to the updates calculated in the previous
‘unregularized’ scheme. Therefore, the effect of the regularization is similar to
filtering the updates with a carefully designed (iteration-dependent) filtering
operator.
In the following, we want to give an interesting additional interpretation of
this regularization scheme.
Define the cost functional
κ1 κ2 κ3
Jˆ(Φ) = Φ2L2 + ∇Φ2L2 + Φ − Ψ 2L2 (54)
2 2 2

with Ψ = Rj [b]∗ ζ. Here, the third term penalizes the misfit between the unregu-

larized update direction δb = Rj [b]∗ ζ and the new candidate Φ, whereas the first
two terms penalize roughness of Φ. The gradient direction for this cost functional
is [(κ1 + κ3 )I − κ2 ∆]Φ − κ3 Ψ (where the Laplace operator is again understood
to be accompanied by the boundary condition ∇Φ · n = 0 on ∂Ω). Therefore, a
necessary condition for the minimum can be stated as
[(κ1 + κ3 )I − κ2 ∆]Φ = κ3 Ψ. (55)
Choosing κ3 = 1, κ2 = β ≥ 0 and κ1 = α − 1 ≥ 0 this amounts to calculating
Φ = (αI − β∆)−1 Ψ, (56)
which is equivalent to (52). Therefore, applying function space regularization as
described above can be interpreted as minimizing the cost functional (54) with
specifically chosen parameters κ1 , κ2 and κ3 .
As already mentioned, this regularization scheme has been applied with great
success to a different but similar inverse problem of reservoir characterization
in [GKMD].
56 O. Dorn et al.

8 Numerical Examples for Electromagnetic Sensitivity

Functions
8.1 Some Numerically Calculated Sensitivity Functions

In the numerical simulations presented here, we use a 40 × 60 × 55 staggered

grid for discretizing the physical domain of interest (see [CBB01, DBBP1] for
details). The computational grid is surrounded by 10 additional cells of PML
(perfectly matched layer) on each side. The total grid (including PML) which
is used in our computations is therefore 60 × 80 × 75 grid cells. Each of the
individual grid cells in the physical domain has a size of 3.5 × 3.5 × 3.5 m3 ,
whereas those in the PML have a size of 10 m in the direction perpendicular to
the boundary, and 3.5 m in the remaining directions. The total interior domain
which is modelled by the computational grid without the PML has therefore the
physical size 140 × 210 × 192.5 m3 , and including the PML it has the physical
size 340 × 410 × 392.5 m3 .
The sources and adjoint sources (receivers) are in all examples y or z-directed
magnetic dipole sources of the strength 1 A m2 and with frequency f = 1.0 kHz.
They are located in two boreholes which are a distance of 30 grid cells or 105 m
apart from each other. We will only consider homogeneous media in this paper
which have constant parameter distributions everywhere. The more general sit-
uation of layered media with an air/soil interface and with objects imbedded
are discussed in [DBBP1, DBBP2]. Nevertheless, we will introduce a (fictitious)
air/soil interface which separates two parts of our homogeneous medium, an up-
per part (called ‘air’) and a lower part (called ‘soil’). The separating interface is
called the ‘free surface’. In our sensitivity analysis we use two different depths
for the source and receiver locations, one being at about 80 m, and the other one
at about 7 m below the free surface.
Figures 3–9 show some examples for numerically calculated sensitivity func-
tions. In all situations, the parameter σ in the medium is 0.1 S m−1 everywhere,
and the relative permittivity r , as well as the permeability parameter µr , are
1.0. The source is a z-directed magnetic dipole transmitter which is positioned
in one borehole at a depth of 80 m below the free surface. A second borehole
containing a receiver antenna is located at the same x-coordinate, but it has a
horizontal distance in the y-coordinate to the first borehole of 105 m. The z-axis
points ‘downward’ into the earth and indicates ‘depth’. The receiving antenna
in the second borehole is located either at the same depth of 80 m below the
free surface (Figs. 3–7), or at a depth of 7 m below the free surface (Figs. 8–9). It
measures either the z-component or the y-component of the magnetic field, as it
is indicated in the individual figures. The frequency is, as in all of our numerical
experiments, f = 1 KHz. All sensitivity functions displayed here are ‘isotropic
sensitivity functions’. For more details regarding the behaviour of ‘anisotropic
sensitivity functions’, we refer the reader to [DBBP2].
Displayed in Fig. 3 is the real part of the numerically calculated sensitivity
function Sz (1, z; y), which we will denote in the following simply by S(z, z; y).
(Similarly, we will denote Sz (1, z; y) simply by S(z, y; y), ignoring the 1 in the
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 57

Fig. 3. Shown are different cross-sections through the real part of the numerically
calculated isotropic sensitivity function S(z, z; y) as a function of y for an infinitely
extended homogeneous medium. The source and the receiver are z-directed magnetic
dipoles at depth 80 m. The scaling factor is Rmax = 3.62 × 10−13

Fig. 4. Shown are the same cross-sections through the real part of the isotropic sensitiv-
ity function S(z, z; y) as in Fig. 3, but now calculated analytically using the expressions
derived in [DBBP2]. The scaling factor is again Rmax = 3.62 × 10−13 . The agreement
with Fig. 3 is very good
58 O. Dorn et al.

Fig. 5. Shown are different cross-sections through the imaginary part of the numerically
calculated isotropic sensitivity function S(z, z; y) as a function of y for an infinitely
extended homogeneous medium. The source and the receiver are z-directed magnetic
dipoles at depth 80 m. The scaling factor is Imax = 2.96 × 10−12

notation which indicates that we consider only one receiver.) In this simplified
notation, the first argument z indicates the orientation of the dipole source,
and the second argument z says that we are measuring the z-component of the
magnetic field. The argument y is the independent variable of the sensitivity
function ranging over all possible scattering positions in the medium.
We mention that all sensitivity functions in this paper are displayed on a
logarithmic scale in the following sense. Let us define the values Rmax and Imax
according to
$ $
$ $
Rmax = max $Re(S(z, z; y))$ ,
y∈Ω
$ $
$ $
Imax = max $Im(S(z, z; y))$ ,
y∈Ω

and analogous expressions for S(z, y; y). Here, the maximum is taken over the
finite number of computer evaluations of S(z, z; y) which is given by the FDFD
forward modelling code. This has the effect that all values of S(z, z; y) which
contribute to Rmax and Imax are actually finite, whereas the theoretically cal-
culated values would become singular at the source and receiver positions. With
these reference values Rmax and Imax , we define the logarithmic sensitivity func-
tions by
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 59

Fig. 6. Shown are different cross-sections through the real part of the numerically
calculated isotropic sensitivity function S(z, y; y) as a function of y for an infinitely
extended homogeneous medium. The source is a z-directed magnetic dipole at depth
80 m, and the receiver is a y-directed magnetic dipole at depth 80 m. The scaling factor
is Rmax = 3.36 × 10−13

  
Rmax
logreal(z, z; y) = −sign(Re(S(z, z; y))) 60 − 10 × log
|Re(S(z, z; y))| +
  
Imax
logimag(z, z; y) = −sign(Im(S(z, z; y))) 60 − 10 × log ,
|Im(S(z, z; y))| +

and analogous expressions hold for S(z, y; y). Here, sign(x) gives the sign (‘plus’
or ‘minus’) of the real number x, the symbol (x)+ = max{0, x} acts as the
identity for positive values of x and delivers zero for negative values. ‘Re’ means
‘real part’, and ‘Im’ means ‘imaginary part’. It is these logarithmic sensitivity
functions which are displayed in Figs. 3–9.
In Fig. 4 we have displayed the same sensitivity functions as in Fig. 3, but
now being calculated with an independent analytical expressions which has been
derived in [DBBP2]. From the agreement of the two Figs. 3 and 4 we can con-
clude that our numerically calculated sensitivity functions are indeed correct.
Additional numerical verifications have been done in [DBBP2]. Notice also that
all our sensitivity functions have been evaluated on the grid positions of our
staggered FDFD grid.
Figure 5 displays the imaginary part of the sensitivity function for the same
setup as in Figs. 3 and 4. In Figs. 6 and 7 the real and imaginary part of the
60 O. Dorn et al.

Fig. 7. Shown are different cross-sections through the imaginary part of the numerically
calculated isotropic sensitivity function S(z, y; y) as a function of y for an infinitely
extended homogeneous medium. The source is a z-directed magnetic dipole at depth
80 m, and the receiver is a y-directed magnetic dipole at depth 80 m. The scaling factor
is Imax = 2.96 × 10−12

sensitivity functions are displayed, respectively, which correspond to the situa-

tion where the transmitter is a z-directed magnetic dipole, but the receiver is a
y-directed magnetic dipole. Finally, Figs. 8 and 9 consider the situation where
both the transmitter and the receiver are z-directed magnetic dipoles, but the
transmitter is located at a depth of 80 m whereas the receiving antenna is located
in the second borehole at a depth of only 7 m.
Certainly, the idealized homogeneous sensitivity functions as presented here
only can give some indication about the true sensitivity structure in a more
realistic scenario where also the air–soil interface is taken into account, and where
a layered medium is assumed for the soil. Such a scenario has been investigated
in more details in [DBBP2].

8.2 A Short Discussion of Sensitivity Functions

We have presented in Sect. 8.1 some results of our numerical experiments for
calculating sensitivity functions. In the following, we want to indicate a few
problems which come up in typical imaging scenarios, and where the study of
sensitivity functions might be very useful in order to address these problems
successfully.
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 61

Fig. 8. Shown are different cross-sections through the real part of the numerically
calculated isotropic sensitivity function S(z, y; y) as a function of y for an infinitely
extended homogeneous medium. The source is a z-directed magnetic dipole at depth
80 m, and the receiver is a z-directed magnetic dipole at depth 7 m. The scaling factor
is Rmax = 4.00 × 10−13

First, when imaging a geophysical test site, some decision has to be made
about which degree of anisotropy should be expected in the given situation.
For a discussion of some implications of electrical anisotropy in applications
see also Weidelt [Wei95]. It can be observed in our numerical experiments (not
shown here, but see [DBBP2]) that the shapes of the x, y and z-component
sensitivity functions typically differ from each other, and certain source-receiver
constellations are very sensitive to some of these components and much less
sensitive to others. This indicates that it should in general be possible to resolve
anisotropic structures in the earth from electromagnetic data, provided that a
sufficient amount of appropriately gathered data is available. On the other hand,
it indicates that the treatment of strongly anisotropic media with an isotropic
imaging approach can lead to inaccurate reconstructions because of the related
misinterpretations of the data.
Second, in basically all imaging problems the question comes up which degree
of resolution can be expected from the reconstructions. In our situation, since
the reconstructions are essentially results of a series of backprojection steps, only
features of the medium can be recovered which are actually represented by the
sensitivity functions. This gives us some information about the resolution which
we can expect. If we want to further quantify the degree of resolution, we could
use for example the concept of a ‘resolution matrix’ which has been discussed
62 O. Dorn et al.

Fig. 9. Shown are different cross-sections through the imaginary part of the numerically
calculated isotropic sensitivity function S(z, y; y) as a function of y for an infinitely
extended homogeneous medium. The source is a z-directed magnetic dipole at depth
80 m, and the receiver is a z-directed magnetic dipole at depth 7 m. The scaling factor
is Imax = 1.22 × 10−12

in the similar situation of hydrology by Vasco et al. [VDL97]. The sensitivity

functions play a natural role in this approach, which is not very surprising since
we would intuitively expect that for example the broadening of the sensitivity
functions some distance away from the sources and receivers will give rise to
a decrease in the resolution of our imaging method at these points. A similar
observation was already mentioned above in the discussion of the broadening of
the sensitivity functions in DOT which leads to a decreased resolution compared
to X-ray tomography. See for example [Ar95a, Ar95b, Ar99, Dor00].
Third, the knowledge of the general form of the sensitivity functions can be
used in order to design experiments more efficiently. See for example Maurer
et al. [MBC00] and the references therein for a discussion of possible design
strategies for electromagnetic geophysical surveys. The Fréchet derivative plays
an important role in these strategies, and we have seen above that the Fréchet
derivative is directly related to our sensitivity functions. A possible design strat-
egy would for example be to combine sources and receivers in a way such that
the resulting sensitivity structure is focused on certain regions of interest in the
earth and is small in other areas. This typically has to be done under certain
constraints on the availability and number of boreholes, on the given source pat-
tern and receiver characteristics, and other test site specific constraints. We refer
for more details again to [MBC00] and the references given there.
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 63

Acknowledgments

Part of the work presented here has been performed during the research visit of
O.D. at UCLA, Los Angeles, in the fall of 2003 for the IPAM special program
on ‘imaging’. The support of IPAM is gratefully acknowledged.

References
[AM95a] Alumbaugh D L and Morrison H F 1995 Monitoring subsurface
changes over time with cross-well electromagnetic tomography Geophys-
ical Prospecting 43 873–902
[AM95b] Alumbaugh D L and Morrison H F 1995 Theoretical and practical con-
siderations for crosswell electromagnetic tomography assuming cylindrical
geometry Geophysics 60 846–870
[AB01] Ammari H and Bao G 2001 Analysis of the scattering map of a linearized
inverse medium problem for electromagnetic waves Inverse Problems 17
219–234
[Ar95a] Arridge S R 1995 Photon-measurement density functions. Part I: Analyt-
ical forms Applied Optics 34 (34) 7395-7409
[Ar95b] Arridge S R and Schweiger M 1995 Photon-measurement density func-
tions. Part II: Finite-element-method calculations Applied Optics 34 (31)
8026-8037
[Ar99] Arridge S (1999) Optical Tomography in Medical Imaging Inverse Prob-
lems 15 R41-R93
[AP98] Ascher U M and Petzold L R (1998) Computer Methods for Ordi-
nary Differential Equations and Differential-Algebraic Equations (SIAM:
Philadelphia)
[BK90] Berryman, J. G., and R. V. Kohn, 1990, Variational constraints for elec-
trical impedance tomography, Phys. Rev. Lett. 65, 325–328
[BBP96] Borcea L, Berryman J G, and Papanicolaou G C 1996 High Contrast Im-
pedance Tomography, Inverse Problems 12 935–958
[BBP99] Borcea L, Berryman J G, and Papanicolaou G C 1999 Matching pursuit
for imaging high contrast conductivity Inverse Problems 15 811–849
[Bor02] L. Borcea 2002 Electrical Impedance Tomography, Inverse Problems 18
R99–R136
[BB99] Buettner H M and Berryman J G 1999 An electromagnetic induction
tomography field experiment at Lost Hills, CA, in the Proceedings of
SAGEEP, Oakland, CA, March 14–18, 663–672
[CBB01] N. J. Champagne II, J. G. Berryman, and H. M Buettner, FDFD (2001) A
3D finite-difference frequency-domain code for electromagnetic induction
tomography J. Comput. Phys. 170 830–848
[Cla76] Claerbout, J. F., 1976, Fundamentals of Geophysical Data Processing: With
Applications to Petroleum Prospecting, McGraw-Hill, New York.
[DRLN] Daily, W. D., A. Ramirez, D. LaBrecque, and J. Nitao, 1992, Electrical
resistivity tomography of vadose water movement, Water Resources Res.
28, 1429–1442
[DA89] Davis, J. L., and A. P. Annan, 1989, Ground penetrating radar for high
resolution mapping of soil and rock stratigraphy, Geophysical Prospect. 37,
531–551
64 O. Dorn et al.

[DS96] Dennis J E and Schnabel R B 1996 Numerical Methods for Unconstrained

Optimization and Nonlinear Equations, (Reprint in the SIAM Classics in
Applied Mathematics Series No. 16)
[DDNPS] Dierkes T, Dorn O, Natterer F, Palamodov V and Sielschott H 2002 Fréchet
derivatives for some bilinear inverse problems, SIAM J. Appl. Math. 62 (6)
2092-2113
[DL79] Dines K A and Lytle R J 1979 Computerized geophysical tomography
Proc. IEEE 67 1065-73
[Dor98] Dorn O 1998 A transport-backtransport method for optical tomography
1998 Inverse Problems 14 1107-1130
[DBBP1] Dorn O, Bertete-Aguirre H, Berryman J G and Papanicolaou G C 1999 A
nonlinear inversion method for 3D-electromagnetic imaging using adjoint
fields Inverse Problems 15 1523-1558
[Dor00] Dorn O 2000 Scattering and absorption transport sensitivity functions for
optical tomography Optics Express 7 (13) 492–506
[DBBP2] Dorn O, Bertete-Aguirre H, Berryman J G and Papanicolaou G C 2002
Sensitivity analysis of a nonlinear inversion method for 3D electromagnetic
imaging in anisotropic media Inverse Problems 18 285-317
[FO96] Farquharson C G and Oldenburg D W 1996 Approximate sensitivities for
the electromagnetic inverse problem Geophysical Journal International 126
235-252
[FMA92] Fisher, E., G. A. McMechan, and A. P. Annan, 1992, Acquisition and
processing of wide-aperature ground penetrating radar data, Geophysics
57, 495–504.
[GKMD] Gonzáles-Rodrı́guez P, Kindelan M, Moscoso M and Dorn O 2005 His-
tory matching problem in reservoir engineering using the propagation-
backpropagation method Inverse Problems 21, 565–590
[HGS93] Habashy T M, Groom R W and Spies B R 1993 Beyond the Born and
Rytov approximations: A nonlinear approach to electromagnetic scattering
J. Geophys. Res. 98 1759–1775
[HAO00] Haber E, Ascher U and Oldenburg D 2000 On optimization techniques for
solving nonlinear inverse problems Inverse Problems 16 1263-1280
[JW57] Jenkins F A and White H E 1957 Fundamentals of Optics (McGraw-Hill:
New York) p 300
[MBC00] Maurer H, Boerner D E and Curtis A 2000 Design strategies for electro-
magnetic geophysical surveys Inverse Problems 16 1097-1117
[MO90] McGillivray P R and Oldenburg D W 1990 Methods for calculating Frechet
derivatives and sensitivities for the non-linear inverse problem: a compar-
ative study Geophysical Prospecting 38 499-524
[MOEH] McGillivray P R, Oldenburg D W, Ellis R G, and Habashy T M 1994 Cal-
culation of sensitivities for the frequency-domain electromagnetic problem
Geophysical Journal International 116 1-4
[Nat86] Natterer F 1986 The Mathematics of Computerized Tomography (Stuttgart:
Teubner)
[NW95] Natterer F and Wübbeling F 1995 A propagation-backpropagation method
for ultrasound tomography Inverse Problems 11 1225-1232
[Nat96] Natterer F (1996) Numerical Solution of Bilinear Inverse Problems
Preprints “Angewandte Mathematik und Informatik” 19/96-N Münster
[NW01] Natterer F and Wübbeling F Mathematical Methods in Image Reconstruc-
tion, Monographs on Mathematical Modeling and Computation 5, SIAM
2001
 Adjoint Fields and Sensitivities for 3D Electromagnetic Imaging 65

[NA97] Newman G A and Alumbaugh D L 1997 Three-dimensional massively par-

allel electromagnetic inversion I. Inversion Geophys. J. Int. 128 345–354
[NW99] Nocedal J and Wright S J 1999 Numerical Optimization, (Springer: New
York)
[RDLOC] Ramirez A, Daily W, LaBrecque D, Owen E and Chesnut D 1993 Moni-
toring an underground steam injection process using electrical resistance
tomography Water Resources Res. 29 73–87
[SH95] Spies B P and Habashy T M 1995 Sensitivity analysis of crosswell electro-
magnetics Geophysics 60 834–845
[VDL97] Vasco D W, Datta-Gupta A and Long J C S 1997 Resolution and uncer-
tainty in hydrologic characterization Water Resources Res. 33 379–397
[Vog02] Vogel C R Computational Methods for Inverse Problems (SIAM series
‘Frontiers in Applied Mathematics’ 2002)
[Wei95] Weidelt P 1995 Three-dimensional conductivity models: implications of
electrical anisotropy. In Oristaglio M and Spies B (ed) International
Sumposium on Three-dimensional Electromagnetics, Oct 4-6, 1995, at
Schlumberger-Doll Research, Ridgefield/Connecticut, USA, pp 1-11
[WAMBL] Wilt, M. J., D. L. Alumbaugh, H. F. Morrison, A. Becker, K. H. Lee,
and M. Deszcz-Pan, 1995a, Crosswell electromagnetic tomography: System
design considerations and field results, Geophysics 60, 871–885
[WMBTL] Wilt, M. J., H. F. Morrison, A. Becker, H.-W. Tseng, K. H. Lee, C.
Torres-Verdin, and D. Alumbaugh, 1995b, Crosshole electromagnetic to-
mography: A new technology for oil field characterization, The Leading
Edge 14, March, 1995, pp. 173–177
[ZTB96] Zhdanov, M. S., P. Traynin, and J. R. Booker, 1996, Underground imaging
by frequency-domain electromagnetic migration, Geophysics 61, 666–682
Polarization-Based Optical Imaging

Miguel Moscoso

Gregorio Millán Institute of Fluid Dynamics, Nanoscience and Industrial

Mathematics, Universidad Carlos III de Madrid, Avda. de la Universidad 30,
28911 Leganés, Madrid, Spain
[email protected]

Summary. In this study I present polarization effects resulting from the reflection
and transmission of a narrow beam of light through biological tissues. This is done
numerically with a Monte Carlo method based on a transport equation which takes into
account polarization of light. We will show both time-independent and time-dependent
computations, and we will discuss how polarization can be used in order to obtain
better images. We consider biological tissues that can be modeled as continuous media
varying randomly in space, containing inhomogeneities with no sharp boundaries.

1 Introduction

Optical imaging is proving to be a potentially useful non-invasive technique for

the detection of objects in the human body. The ultimate goal is the detection of
millimeter-sized cancerous tissue before it spreads into the surrounding healthy
tissue [1–3]. Cancerous tissue absorbs more light than healthy tissue due to
its higher blood content, giving rise to a perturbation in the measured light
intensity. The main problem in using light as a diagnostic tool is that it is strongly
scattered by normal tissue, resulting in image blurring. Various techniques have
been devised to overcome this difficulty, e.g., time gating, optical coherence,
confocal detection, etc. All these techniques distinguish between weakly and
strongly scattered photons, i.e., those scattered through small angles and those
scattered through large angles.
Recently, there has been a considerable interest in the polarization properties
of the reflected and transmitted light. Multiple scattering gives rise to diffusion
and depolarization. The effectiveness of polarization-sensitive techniques to dis-
criminate between weakly and strongly scattered photons has been demonstrated
experimentally by Schmitt et al. [4], Rowe et al. [5], and Demos et al. [6], among
others. We have reported numerical evidence for this previously using the theory
of radiative transfer taking into account polarization of light [7].
The theory of radiative transfer, employing Stokes parameters, can be used
to describe light propagation through a medium. Analytical solutions are not
known except in simple particular cases, such as plane parallel atmospheres
68 M. Moscoso

with a constant net flux [8]. When the incident laser beam is narrow no solution
is known. Therefore we have developed a Monte Carlo method to obtain the
spatial distribution of the total intensity and of the polarization components
of the transmitted and reflected beams. Other methods for solving the vector
radiative transfer equation have been explored in [9].
In [9] we presented a complete discussion of Chebyshev spectral methods
for solving radiative transfer problems. In this method, we approximated the
spatial dependence of the intensity by an expansion of Chebyshev polynomials.
This yields a coupled system of integro-differential equations for the expansion
coefficients that depend on angle and time. Next, we approximated the integral
operation on the angle variables using a Gaussian quadrature rule resulting in
a coupled system of differential equations with respect to time. Using a second
order finite-difference approximation, we discretized the time variable. We solved
the resultant system of equations with an efficient algorithm that makes Cheby-
shev spectral methods competitive with other methods for radiative transfer
equations.
Section 2 contains the theory of radiative transfer employing Stokes para-
meters. In Sect. 3 we describe our model where we represent a biological tis-
sue as a medium with weak random fluctuations of the dielectric permittivity
(r) = 0 [1 + δ(r)], where r denotes position and δ(r) is the continuous ran-
dom fractional permittivity fluctuation. In Sect. 4 we describe a Monte Carlo
method for three-dimensional problems which take into account the vector na-
ture of light. Section 5 describes our numerical calculations. Section 6 contains
our conclusions.

2 Radiative Transfer Equation

The theory of radiative transport employing Stokes parameters can be used
to describe how light propagates through an absorbing scattering medium. Both
the absorption coefficient Σa and the scattering coefficient Σs of the background
medium, can be determined from reflection and transmission measurements. Ab-
sorption of light in tissue is due to natural chromophores, such as the heme
pigment of hemoglobin, the cytochrome pigments of the respiratory chain in
mitochondria, and others [10]. Scattering in soft tissue is due to random low
contrast inhomogeneities of the refractive index, of the order of 5% [11]. Due to
interaction with the inhomogeneities, light waves with wave vector k at point r
can scatter into any direction k̂ with wave vector k, where k̂ = k/|k| . The sta-
tistical properties of the light propagating through the medium can be specified
by the Stokes vector ⎛ ⎞
I
⎜ ⎟
⎜ ⎟
⎜Q⎟
I=⎜ ⎟
⎜ ⎟ , (1)
⎜U ⎟
⎝ ⎠
V
 Polarization-Based Optical Imaging 69

er
θ
el

y
ϕ

x
Fig. 1. Coordinate system for the Stokes vector. The shaded plane φ = contant con-
taining the wavevector k is called the meridian plane. We choose el to lie in this plane
and to be orthogonal to k

where
I = El El∗ + Er Er∗  ,
Q = El El∗ − Er Er∗  ,
U = El Er∗ + Er El∗  ,
V = iEl Er∗ − Er El∗  . (2)
Here El and Er are the complex amplitudes of the electric field referred to an
orthonormal system (k̂, el , er ); see Fig. 1. I is the total intensity. The Stokes
parameter Q is the difference between the intensities transmitted through two
linear analyzers in orthogonal directions. U is defined like Q, but with reference
to linear analyzers rotated by 45◦ with respect to those of Q. V is the difference
between the intensities passing through right and left circular analyzers.
The Stokes vector I(x, k̂, t), defined for all directions k̂ at each point x and
time t, satisfies the transport equation [8]
1 ∂I(x, k̂, t)
+ k̂ · ∇x I(x, k̂, t) + (Σ (s) (x) + Σ (a) (x))I(x, k̂, t) =
v ∂t

Σ (s) (x) F (k̂ · k̂ )I(x, k̂ , t)dk̂ . (3)

Here v, Σ (s) , and Σ (a) are the velocity of the light in the medium, the scattering
coefficient, and the absorption coefficient, respectively. The 4 × 4 scattering ma-
trix F (k̂ · k̂ ) in (3) describes the probability for a photon entering a scattering
70 M. Moscoso

process with Stokes vector I(i) and direction k̂ to leave the process with Stokes
vector I(s) and direction k̂. The dot product in the argument of F means that
it depends only on the cosine of the scattering angle Θ.

3 Model
We consider biological tissue with weak random fluctuations of the dielectric
permittivity (r) = 0 [1 + δ(r)]. Then F is related to the power spectral density
R̂(q) of δ(r) by [12]:
π 4 Θ
F (Θ) = k R̂(2k sin )S(Θ) . (4)
2 2
Here ⎛ ⎞
1 + cos2 Θ cos2 Θ − 1 0 0
⎜ ⎟
⎜ 2 ⎟
1⎜
⎜
cos Θ − 1 1 + cos2 Θ 0 0 ⎟
⎟
S(Θ) = ⎜ ⎟ (5)
2⎜ 0 0 2 cos Θ 0 ⎟
⎝ ⎠
0 0 0 2 cos Θ

is the Rayleigh scattering matrix. R̂ contains statistical information about the

random inhomogeneities of the medium, which can be modeled or obtained ex-
perimentally [11].
S(Θ) in (5) is referred to the scattering plane, i.e., the plane through the
incident and scattered directions. It is more convenient to express S with respect
to the directions el , both incident and scattered, parallel to the meridian plane
as in Fig. 1. We refer to [8] for more details about the form of the scattering
matrix in this basis, and to [13] and [14] for a discussion of depolarization of
multiply scattered light.
Equation (4) is the basic expression for the scattering matrix in a weakly
fluctuating random medium, and relates the angular scattering pattern to the
statistical characteristics of the medium. The scattering pattern is the product
of the matrix S(Θ), which is a purely geometrical factor, and the scalar function
R̂(2k sin Θ/2), which involves the statistical properties of the medium. The power
spectral density function R̂ can be obtained experimentally [11]. However, in
many practical situations the spectral characteristics of the fluctuations can be
modeled by simple functions. For mathematical simplicity we consider a model
with only two parameters: the relative strength of the fluctuations, ε, and the
correlation length, l. We use an exponential form for the correlation function,

R(r) = ε2 e−r/l . (6)

In this case the power spectral density function is given by

ε2 l 3
R̂(k) = . (7)
π 2 (1 + k 2 l2 )2
 Polarization-Based Optical Imaging 71

The exponential correlation function (6) is a especial case of the von Karman
correlation function
ε2 21−κ r κ r
R(r) = ( ) Kκ ( ) , κ ∈ [0, 0.5] (8)
Γ (κ) l l

for κ = 0.5, used to model the spectral characteristics of the fluctuations in

turbulent fluids. In (8), Γ is the gamma function and Kκ is the modified Bessel
function of the second kind of order κ. The von Karman correlation function
is commonly used to model the spectral characteristics of the fluctuations in
turbulence media.
We note that the power spectral density function (7) becomes independent
of k for kl  1, so then the factor R̂ in (4) is independent of Θ. R̂(2k sin Θ/2)
is proportional to (2kl sin Θ/2)−4 for kl  1, so as kl increases, the scattering
becomes more peaked in the forward direction.
The mean value of the cosine of the scattering angle, called the anisotropy
factor, is defined by

1
π2 k4 √
g(k) = cos ΘR̂(k 2 − 2 cos Θ)[1 + cos2 Θ]d(cos Θ) . (9)
2Σs (k) −1

It is one of the parameters used to describe the optical properties of a medium.

For the correlation function (6) with kl equal to 0, 1, 2 or 5, g is equal to 0, 0.30,
0.48, or 0.66, respectively. We note that g is typically larger than these values
in biological tissues (> 0.70 for wavelengths in the visible). There, the majority
of the scattering takes place from structures within the cell, like mitochondria
(diameter of the order 0.3–0.7 µm) or lysosomes and peroxisomes (diameter of
the order 0.2–0.5 µm) [15, 16]. More complicated correlation functions than (6)
could yield larger values of g for wavelengths in the visible.
In the next section we describe a Monte Carlo Method to solve the radiative
transport equation (3).

4 Monte Carlo Method

Monte Carlo methods are based on the stochastic nature of the propagation of
light in a random medium. They simulate photon histories according to funda-
mental probability laws which depend on the scattering medium. Their main
advantage is their relative simplicity and their ability to handle complicated
geometries. They also provide exact solutions (up to statistical errors) of the ra-
diative transport equation (3). Their main drawback is that they require a large
number of photons to obtain statistical accuracy for large optical depths. The
rate of convergence is given by the central limit theorem of probability theory. It
states that for N independent random variables I i from theNsame distribution,
with finite mean m and variance σ 2 , the average IN = i I i /N is normally
distributed about m with variance σ 2 /N . This means that IN converges to the
72 M. Moscoso
√
expected value at the rate const./ N . To make it converge faster several vari-
ance reduction techniques can be used [7].
We use two basic approaches to improve the convergence [17, 18]. First, we
modify the random walk sampling to follow more particles in important regions
where the scoring contributes more to the final tally. This is especially relevant
in computing light transmission through a medium of large optical thickness. In
that case, only a small fraction of particles penetrates into the deeper regions
of the medium, producing bad statistics in the computation. Second, we modify
the scoring method to maximize the statistical information of a given random
walk. In doing this, a trajectory is no longer associated with a single particle.
The replacement of the analog simulation by a non-analog one is usually called
a variance reduction technique.
In the next subsection we describe the simplest case of the analog simula-
tion. In Sect. 4.2 we explain the geometry splitting and Russian roulette, whose
objective is to spend more time sampling particles in important regions and less
time sampling particles in unimportant regions. In Sect. 4.3 we explain the point
detector technique, which is a deterministic estimate of the flux at a point from
each collision.

4.1 Analog Monte Carlo Sampling

Consider a non-absorbing homogeneous medium with scattering coefficient Σs .

The history of each photon is initiated at time t = 0 by releasing it from the
point rin = (xin , yin , 0) with direction k̂in = (0, 0, 1) along the inward normal
to the surface of the medium, and with a specified polarization state. When the
beam is linearly polarized, it is 100% polarized parallel to the (x, z) plane, so
the Stokes vector can be written as [Iin , Qin , Uin , Vin ] = [1, 1, 0, 0]. When the
beam is circularly polarized, the Stokes vector is [Iin , Qin , Uin , Vin ] = [1, 0, 0, 1].
We will consider only linear initial polarization here.
The path length traveled by the photon before undergoing scattering is
given by
lcoll = −ln(ξ)/Σs , (10)
where ξ is a random number uniformly distributed in the interval (0, 1). The
photon’s scattering position r1 is

rf = rin + k̂in lcoll , (11)

with k̂in = (sin θin cos φin , sin θin sin φin , cos θin ). Next, the scattering direction
k̂f is determined according to the probability density function [7, 19]

1 % &
P (k̂f , k̂in , Qin , Uin ) = F̄11 + F̄12 Qin + F̄13 Uin . (12)
Σs

Here F̄ij , with i, j = 1, . . . , 4, represents the (i, j) element of the scattering ma-
trix F̄ referred to the meridian plane (plane of constant φd ) as plane of reference
 Polarization-Based Optical Imaging 73

for the Stokes vectors. See [8] for the required rotation. P depends on the di-
rection of incidence and scattering, and on the incoming polarization state. The
dependence on the polarization state is essential because linearly polarized pho-
tons cannot be scattered in the direction of oscillation of the electric field. This
property, related to the vector nature of light, is lost if the probability density
function does not depend on the polarization. Note that P does not depend on
Vin because F14 = 0 in weakly fluctuating media. P has been normalized so that
for all incident directions and polarization states


P (k̂f , k̂in , Qin , Uin )dk̂f = 1 , (13)

This equation expresses particle conservation in each scattering event. We note

that in (13), the integrals of the terms multiplied by Qin and Uin are zero.
Once the new direction k̂1 is chosen, the outgoing Stokes vector is
1
If = F̄ (k̂f , k̂in )Iin . (14)
Σs P (k̂f , k̂in , Qin , Uin )

Equation (14) does not depend on R̂(k). The normalization in (14) assures that
the intensity of the scattered photon remains equal to the incident one. From
(14) it also follows that if the incident photon is completely polarized then the
scattered photon is also completely polarized.
Next, a new path length is calculated from (10) and the tracing continues
using (11)–(14) until the photon exit the medium.
If a photon reaches the detector at an angle with the normal to the sample
surface that is less than its semiaperture, then its contribution to the quantities of
interest (intensity and polarization) are stored. If the incident beam is circularly
polarized, two quantities are computed: the total intensity I and the circular
component V . In the case of a linearly polarized incident beam, three intensities
are computed: the total intensity I, the parallel-polarized component Il , and
the cross-polarized component Ir . If a photon of intensity dI = 1 arrives at the
detector with polar direction (θd , φd ), these two last intensities can be calculated
easily from the Stokes parameters Qd and Ud of the photon. The direction of
polarization makes an angle

χd = ±1/2 arctan(Ud /Qd ) (15)

with the meridian plane. The sign is positive for transmission, and negative for
reflection. The difference in signs is due to the change of sign of U when the z
component of the wave-vector k̂ changes sign. We have chosen el × er = k̂ (see
Fig. 1). If el × er = −k̂ instead, then the signs must be changed in (15). The
contribution of the photon to the transmitted intensities is:

I = dI , (16)
Il = dI[cos2 (φd + χd ) cos2 θd + sin2 θd ] , (17)
Ir = dI[sin2 (φd + χd ) cos2 θd + sin2 θd ] . (18)
74 M. Moscoso

To compute the reflected intensities, we must change χd to −χd in (17) and (18).
(n) (n)
After N photons have been traced, and the intensities I (n) , Il , Ir , and
(n)
V of the N histories have been stored, we average to obtain the statistical
result,
N N
1 1 (n)
I¯ = I (n) , I¯l = Il , (19)
N n=1
N n=1
N N
1 1
I¯r = Ir(n) , V̄ = V (n) . (20)
N n=1
N n=1

We note from (17) and (18) that Il + Ir = I(1 + sin2 θd ). This is not equal to
I unless sin θd = 0, because if sin θd = 0 there is a component of the electric field
normal to the analyzers. This component is transmitted through them indepen-
dently of their orientation. Consequently, when measuring a bundle of light, in a
typical experimental situation, the sum I¯l + I¯r of the parallel and cross-polarized
components is not equal to the total intensity I, ¯ measured without polarization
analyzers.
We also note that the fraction of light with the original polarization, (I¯l −
¯r )/I,
I' ¯ coincides with the common definition of the degree of polarization
¯ only for small apertures. In this case, the intensity of light that
Q̄ + Ū 2 /I,
2

has maintained the incident polarization state, I¯l − I¯r , is a measure of the Stokes
parameter Q̄. In the problem that we shall consider, Ū is zero due to symmetry.

4.2 Splitting and Russian Roulette

One of the most widely used variance reduction methods is the combination of
the so called importance splitting and Russian roulette techniques. The problem
domain is divided into M subdomains with different importances I (m) (m =
0, . . . , M − 1). We will assume for simplicity that I (m) = cm where c is a chosen
positive integer. The more probable it is that a particle in a sub-domain can score
at the detector, the more important is the sub-domain. Let a particle of weight
W evolve between two collisions i and i + 1, and let us denote the importance
of the particle by Ii and Ii+1 at each collision (Ii = I (m) if at collision i the
photon is in sub-domain m). We form the importance ratio η = Ii+1 /Ii . If η = 1
the importance of the photon does not change and its transport continues in the
usual way. If η > 1 the photon has moved to a more important region and it
is split into Nsplit = η photons. The weight W of each of the split photons is
divided by Nsplit . If η < 1 it has entered into a less important region, and a
Russian roulette game is played where the particle has probability psurv = η
of surviving. A random number ξ is drawn uniformly on the interval (0, 1) and
compared with psurv . If ξ < psurv the photon survives and its weight is increased
by 1/psurv . Otherwise, it is eliminated. In doing this, however, a balance has to
be struck. Implementing splitting generally decreases the history variance but
increases the computing time per history. On the contrary, Russian roulette
 Polarization-Based Optical Imaging 75

increases the former and decreases the latter, so that more histories can be run.
The usual measure of efficiency of the Monte Carlo simulations is the figure of
merit 1/σ 2 t, where the variance σ 2 measures the inaccuracy of the result, and
t is the mean time per history. We have seen experimentally that, for example,
for a medium of optical thickness 12, the choice c = 2 and M = 5 improves
the figure of merit by a factor 10. However, for optical thickness less than 8 we
have not found the splitting/Russian roulette necessary. For deep penetration,
experience has indicated that the best splitting results are achieved keeping the
particle flux through the medium close to a flat distribution, and that there is
a broad range of parameters close to the optimal choice [18]. If splitting with
Russian roulette is the only variance reduction technique used, all particles in
the same sub-domain m have the same weight 1/I (m) , assuming that all particle
histories began with weight 1.

4.3 Point Detectors

We now consider scoring at a small detector of area dA, often called a point
detector. For each scattering event i of the n-th photon, we store the probability
(n)
that the n-th scattered photon, of weight Wi , will arrive at the detector with-
out further collisions. Let P (Θ, P)dΩ be the probability of scattering into the
solid angle dΩ about Θ, where Θ is the scattering angle directed to the detector,
and P = (Q0 , U0 ) denotes the liner polarization state of the photon. Taking into
account the attenuation suffered by particles traveling through the medium, the
probability of arriving at the detector is
(R
P (Θ, P)dΩe− 0
Σs (l)dl
, (21)

where R is the distance covered within the medium.

We recall that we are considering non-absorbing media. Absorption can be
accounted for easily by replacing the scattering coefficient Σs in (21) by the total
extinction coefficient Σt = Σs + Σa , where Σa is the absorption coefficient, and
multiplying dI by Σs /Σt at each scattering event.
If the detector of area dA has normal vector forming the angle θd with respect
to the line of flight, then dΩ = cos θd dA/R2 , and (21) becomes
(R
cos θd e− 0 Σs (l)dl
P (Θ, P) dA . (22)
R2
Since the flux is the number of particles crossing unit area per unit time, the
contribution of the i-th collision to the flux at the detector is given by

(n) (n) cos θd e−Σs R

FIi = Wi P (Θ, P) . (23)
R2
Here we have evaluated the integral for a homogeneous medium. The contri-
(n)
bution to the Stokes parameters is given by (14) times FIi . In the case of a
76 M. Moscoso

linearly polarized incident beam, the contributions to the parallel and cross-
polarized components are evaluated with expressions similar to (17) and (18),
respectively.
After N photons have been traced, the statistical averages are computed as
N Sn N Sn
1 (n) 1 (n)
F̄I = FIi , F̄l = Fli , (24)
N n=1 i=1
N n=1 i=1

N Sn N Sn
1 (n) 1 (n)
F̄r = Fri , F̄V = FV i , (25)
N n=1 i=1
N n=1 i=1

where Sn is the number of scattering events of the n-th photon.

In spite of the simplicity of the method, the point detector estimator has a
serious drawback. The expression (23) becomes infinite when a collision takes
place infinitely close to the detector, giving rise to an unbounded estimator with
infinite variance and consequently a low rate of convergence. However, if the
detector is not in the scattering medium, a collision close to the detector is
impossible and the estimator is bounded. Our point detectors are placed at a
distance d = 1 mm from the exit surface, so we replace the R in the denomi-
nator of (23) by (R + d/ cos θd ). Point detectors have been used widely in the
literature [20–23].
Since tracing photons from each collision to the detector is time consuming,
especially in the case where an object is embedded, we have also implemented a
Russian roulette game for the detector. If the collision takes place in a region of
importance I (m) , we introduce the probability to contribute to the flux equal to
I (m) /I (M −1) = cm−M +1 . If the particle wins the game its deterministic contri-
bution is multiplied by I (M −1) /I (m) and it is stored. Otherwise, the contribution
is not stored.
Finally, we note that other techniques which improve the convergence of a
Monte Carlo simulation can be combined with the procedure described above.
For example, Tinet et al. [22] use the statistical estimator concept to accelerate
the convergence for the scalar transport equation. In their scheme, they divide
the computation in two stages. The first one is the generation and storage of all
the necessary information, with a random walk similar to the one explained in
Sect. 4.1 but without polarization effects. However, they also compute at each
scattering event, the part of the light that escapes the medium without further
collisions, and they subtract it from the weight of the photon. Consequently, the
photon is not allowed to leave the medium and the energy distribution within
the medium is closer to the average energy distribution. In the second stage
they use the information generated previously to estimate, with a point detector
technique, the contribution to any quantity of interest. They report a better con-
vergence rate than the use of the point detector alone. Its procedure is also more
flexible due to the separation into two stages, which allow the post-computation
of any quantity of interest related to the problem. The main drawback is the
need to store the position and weight of the photons at all the scattering events,
which may require a large amount of memory.
 Polarization-Based Optical Imaging 77

5 Numerical Results

In our numerical experiments we consider an ultra-short laser beam impinging

normally at some point (xin , yin ) on the surface z = 0 of a sample of dimensions
L × L × Z; see Fig. 2. The laser beam is linearly polarized parallel to the (x, z)
plane. The sample consists of a uniform background medium with scattering
(s) (a)
and absorption coefficients Σbg and Σbg respectively, containing a spherical
(s) (a)
inclusion of optical coefficients Σincl and Σincl . See [23] for a good description of
the modifications needed to handle the case where a spherical object is imbedded.
The dimensions of our sample are 10 cm × 10 cm × 5 cm; see Fig. 2. The inci-
dent beam is 100% linearly polarized parallel to the (x, z) plane and enters into
the medium at time t = 0. The incident laser positions are on the x axis 2.5 mm
apart. For each laser position, two detectors are placed on the optical axis, one
at the bottom, and another at the top of the sample, to measure the trans-
mitted and reflected intensities of the light. Both detectors are at d = 1.0 cm
from the exit surface. The sample consists of a uniform background medium
(s) (a)
with optical parameters Σbg and Σbg , containing a spherical inclusion of ra-
(s) (a)
dius Rincl = 5.0 mm and optical parameters Σincl and Σincl . We assume that

receiver

incident beam
scan direction

x
y
s Z

L x
y d

z receiver
L
(s) (s)
Scattering coefficients of the background Σ bg , and of the inclusion Σ incl .

(a) (a)
Absorption coefficients of the background Σ bg , and of the inclusion Σ incl .

Fig. 2. The setup used in the simulations. The dimensions of the sample are L×L×Z. It
consists of a uniform medium in which a spherical object of radius 5.0 mm, with different
optical properties, is imbedded. The center of the object is at 1.6 cm from the input
surface. The distance between two consecutive incident beam positions is 2.5 mm. The
detectors are placed at a distance d = 1.0 cm from the exit surface. The semiaperture
angle of the each detector is 90◦ . A complete scan consists of 17 measurements
78 M. Moscoso

Itot and Il Itot and Ir

1 1
Relative Amplitude
0.95 0.95

0.9 0.9

0.85 0.85

0.8 (a) 0.8 (b)

0.75 0.75
−2 −1 0 1 2 −2 −1 0 1 2

Itot and (Il – Ir) Itot and (Il – Ir)/Itot

1 1
Relative Amplitude

0.95 0.95

0.9 0.9

0.85 0.85

0.8 (c) 0.8 (d)

0.75 0.75
−2 −1 0 1 2 −2 −1 0 1 2
Distance (cm) Distance (cm)

Fig. 3. Reflection. Σbg = Σincl = 1 cm−1 , Σbg = 0.1 cm−1 , Σincl = 100 cm−1 . The
(s) (s) (a) (a)

solid lines show the total intensity I. The dashed lines show (a) the parallel-polarized
component Il , (b) the cross-polarized component Ir , (c) the polarization-difference
Il − Ir , and (d) the degree of polarization (Il − Ir )/I

the scattering matrix S is the same for inclusion and background. However the
scattering coefficients and the absorption coefficients may differ.
Figures 3 and 4 show numerical results for the time-integrated backscattered
(s) (s)
and transmitted light, respectively. The optical parameters are Σbg = Σincl =
1 cm−1 , Σbg = 0.1 cm−1 , Σincl = 100 cm−1 . The solid lines show the total
(a) (a)

intensity I, and the dashed lines show: (a) the parallel-polarized component
Il , (b) the cross-polarized component Ir , (c) the polarization-difference Il − Ir ,
and (d) the degree of polarization (Il − Ir )/I. We observe in Fig. 3 that the
backscattered cross-polarized component Ir gives the best image of the absorbing
object. The reason for the enhanced resolution is that this image is formed
by photons that probe deeper into the medium. However, the best image in
transillumination is obtained by subtracting the cross-polarized component Ir
from the parallel polarized component Il . In the difference Il − Ir , only ballistic
photons and weakly scattered photons, which keep their initial polarization, are
retained [4, 7].
As a second example (Figs. 5 and 6) we consider a scattering object with
Σincl = 10Σbg = 10 cm−1 and Σincl = Σbg = 0.1 cm−1 . The curves in the
(s) (s) (a) (a)

figures have the same meaning as in Figs. 3 and 4, and the time-integrated
backscattered and transmitted intensity is also shown. We obtain similar
 Polarization-Based Optical Imaging 79

Itot and Il Itot and Ir

1 1

Relative Amplitude 0.8 0.8

0.6 0.6

0.4 0.4

0.2 (a) 0.2 (b)

0 0
−2 −1 0 1 2 −2 −1 0 1 2

Itot and (Il – Ir) Itot and (Il – Ir)/Itot

1 1
Relative Amplitude

0.8 0.8

0.6 0.6

0.4 0.4

0.2 (c) 0.2 (d)

0 0
−2 −1 0 1 2 −2 −1 0 1 2
Distance (cm) Distance (cm)

Fig. 4. Transmission. Σbg = Σincl = 1 cm−1 , Σbg = 0.1 cm−1 , Σincl = 100 cm−1 . The
(s) (s) (a) (a)

solid lines show the total intensity I. The dashed lines show (a) the parallel-polarized
component Il , (b) the cross-polarized component Ir , (c) the polarization-difference
Il − Ir , and (d) the degree of polarization (Il − Ir )/I
Itot and Il Itot and Ir
1 1
Relative Amplitude

0.95 0.95

0.9 0.9

0.85 0.85

0.8 0.8

0.75 (a) 0.75 (b)

0.7 0.7
−2 −1 0 1 2 −2 −1 0 1 2

Itot and (Il − Ir) Itot and (Il − Ir)/Itot

1 1
Relative Amplitude

0.95 0.95

0.9 0.9

0.85 0.85

0.8 0.8

0.75 (c) 0.75 (d)

0.7 0.7
−2 −1 0 1 2 −2 −1 0 1 2
Distance (cm) Distance (cm)

Fig. 5. Reflection. Σincl = 10Σbg = 10 cm−1 , Σincl = Σbg = 0.1 cm−1 . The solid lines
(s) (s) (a) (a)

show the total intensity I. The dashed lines show (a) the parallel-polarized component
Il , (b) the cross-polarized component Ir , (c) the polarization-difference Il − Ir , and
(d) the degree of polarization (Il − Ir )/I
80 M. Moscoso

Itot and Il Itot and Ir

1 1
Relative Amplitude
0.8 0.8

0.6 0.6

0.4 0.4

0.2 (a) 0.2 (b)

0 0
−2 −1 0 1 2 −2 −1 0 1 2

Itot and (Il − Ir) Itot and (Il − Ir)/Itot

1 1
Relative Amplitude

0.8 0.8

0.6 0.6

0.4 0.4

0.2 (c) 0.2 (d)

0 0
−2 −1 0 1 2 −2 −1 0 1 2
Distance (cm) Distance (cm)

Fig. 6. Transmission. Σincl = 10Σbg = 10 cm−1 , Σincl = Σbg = 0.1 cm−1 . The
(s) (s) (a) (a)

solid lines show the total intensity I. The dashed lines show (a) the parallel-polarized
component Il , (b) the cross-polarized component Ir , (c) the polarization-difference
Il − Ir , and (d) the degree of polarization (Il − Ir )/I

conclusions as in the previous example. If we measure reflected light from the

medium then the best resolution of the scattering abnormality is obtained with
the cross-polarized component Ir . If we measure the transmitted light through
the medium, then the polarization-difference Il − Ir achieves the best resolution.
(s) (s) (s)
We have seen that increasing Σbg and keeping Σincl = 10Σbg increases
the benefit of the transmitted polarization-difference Il − Ir compared to the
total transmitted intensity I. However, it decreases the benefit of the reflected
cross-polarized component Ir .
As the last example we show in Fig. 7 the numerical results for the time-
dependent backscattered intensity. Figure 7 shows (a) the time profile of the total
intensity, (b) the co-polarized component, (c) the cross-polarized component,
and (d) the polarization-difference, at two different scan positions. The solid
lines correspond to the signal received at the scan position (x, y) = (0, 0); see
Fig. 2. The dashed-lines are the response received at the position furthest from
(s) (s)
the target (a scattering sphere with Σincl = 10Σbg ). We observe that all the
figures show a perturbation in the intensity due to the presence of the object.
However, the differences between the cross-polarized time profile and the others
are apparent. We clearly see that a very small amount of light that scatters
just beneath the surface z = 0 is present in the signal. The presence of an
object that scatters light more makes the light depolarize faster. Therefore more
 Polarization-Based Optical Imaging 81

1 1

Co−polarized Intensity
(a) (b)
0.8 0.8
Total Intensity

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 0 5 10 15
Time Time
1 1
Cross−polarized Intensity

Polarization−Difference
(c) (d)
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 5 10 15 0 5 10 15
Time Time

Fig. 7. Time trace of the backscattered light response. Σincl = 10Σbg = 15 cm−1 ,
(s) (s)

(a) (a) −1
Σincl = Σbg
= 0.1 cm , L = 16 cm, Z = 5 cm, s = 1.6 cm. (a) Normalized total
intensity, (b) normalized co-polarized component, (c) normalized cross-polarized com-
ponent, and (d) normalized polarization-difference. The maxima in (b), (c), and (d) are
1.01, 20.36 and 1.02 times smaller than the maximum in (a). Solid curves correspond
to the signal received at the scan position (x, y) = (0, 0), just above the center of the
sphere. Dashed curves are the signal received at the scan position, (x, y) = (2.8, 0) cm,
furthest from the sphere

photons that preserve the initial polarization are present in the cross-polarized
image. The fact that the cross-polarized intensity is composed of photons that
have propagated deeper into the medium may be very useful for reconstruction
of images when one is not interested in details at and near the surface.

6 Conclusions

We have shown that simple polarization imaging techniques can be used to en-
hance the resolution of objects hidden in scattering media. Polarization imaging
is based on the fact that the polarization of the initial beam is preserved over
large optical distances. For backscattered light, we find that the cross-polarized
component shows the best resolution, while for transmitted light, images ob-
tained by subtraction of the cross-polarized component from the co-polarized
component are better. Differences between the time-dependent backscattered
signal have also been presented.
82 M. Moscoso

Acknowledgements

The author acknowledges support from the Spanish Ministerio de Educación y

Ciencia (grants no. FIS2004-03767 and FIS2007-62673) and from the European
Union (grant LSHG-CT-2003-503259).

References
1. S. R. Arridge and J. C. Hebden, “Optical Imaging in Medicine: II. Modelling and
reconstruction,” Phys. Med. Biol. 42, pp. 841–853, 1997.
2. P. J. Michielsen K., DeRaedt H. and G. N, “Computer simulation of time-resolved
optical imaging of objects hidden in turbid media,” Phys. Reports 304, pp. 90–144,
1998.
3. S. L. Jacques, J. R. Roman, and K. Lee, “Imaging superficial tissues with polarized
light,” Laser. Surg. Med. 26(2), pp. 119–129, 2000.
4. J. M. Schmitt, A. H. Gandjbakhche, and R. F. Bonner, “Use of polarized light to
discriminate short-path photons in a multiply scattering medium,” Appl. Opt. 31,
pp. 6535–6546, 1992.
5. M. Rowe, E. Pungh, J. Tyo, and N. Engheta, “Polarization-difference imaging:
a biologically inspired technique for observation through scattering media,” Opt.
Lett. 20, pp. 608–610, 1995.
6. S. Demos and R. Alfano, “Temporal gating in highly scattering media by the degree
of optical polarization,” Opt. Lett. 21, pp. 161–163, 1996.
7. M. Moscoso, J. Keller, and G. Papanicolaou, “Depolarization and blurring of op-
tical images by biological tissues,” J. Opt. Soc. Am. A 18, No. 4, pp. 948–960,
2001.
8. S. Chandrasekhar, Radiative Transfer, Oxford University Press, Cambridge, 1960.
9. A. Kim and M. Moscoso, “Chebyshev spectral methods for radiative transfer,”
SIAM J. Sci. Comput. 23, No. 6, pp. 2075–2095, 2002.
10. B. C. Wilson and S. Jacques, “Optical reflectance and transmittance of tissues:
Principles and applications,” IEEE Quant. Electron. 26, pp. 2186–2199, 1990.
11. J. M. Schmitt and G. Kumar, “Turbulent nature of refractive-index variations in
biological tissue,” Opt. Lett. 21, pp. 1310–1312, 1996.
12. L. Ryzhik, G. Papanicolaou, and J. B. Keller, “Transport equations for elastic and
other waves in random media,” Wave Motion 24, pp. 327–370, 1996.
13. A. Kim and M. Moscoso, “Influence of the relative refractive index on depolariza-
tion of multiply scattered waves,” Physical Review E 64, p. 026612, 2001.
14. A. Kim and M. Moscoso, “Backscattering of circularly polarized pulses,” Opt.
Letters. 27, pp. 1589–1591, 2002.
15. B. Beauvoit, T. Kitai, and B. Chance, “Contribution to the mitochondrial com-
partment to the optical properties of the rat liver: a theoretical and practical
approach,” Biophys. J. 67, pp. 2501–2510, 1994.
16. J. Mourant, J. Freyer, A. Hielscher, A. Eick, D. Shen, and T. Johnson, “Mecha-
nisms of light scattering from biological cells relevant to noninvasive optical-tissue
diagnostics,” Appl. Opt. 37, pp. 3586–3593, 1998.
17. M. H. Kalos and P. A. Whitlock, Monte Carlo Methods, New York : J. Wiley Sons,
1986.
18. E. E. Lewis and W. F. Miller Jr., Computational Methods of Neutron Transport,
John Wiley and sons, New York, 1984.
 Polarization-Based Optical Imaging 83

19. G. Bal, G. Papanicolaou, and L. Ryzhik, “Probabilistic Theory of Transport

Processes with Polarization,” SIAM Appl. Math. 60, pp. 1639–1666, 2000.
20. M. Kalos, “On the estimation of flux at a point by Monte Carlo,” Nucl. Sci. Eng.
16, pp. 111–117, 1963.
21. P. Bruscaglioni, G. Zaccanti, and Q. Wei, “Transmission of a pulse polarized light
beam through thick turbid media: numerical results,” Appl. Opt. 32, pp. 6142–
6150, 1993.
22. E. Tinet, S. Avrillier, and M. Tualle, “Fast semianalytical monte carlo simulation
for time-resolved light propagation in turbid media,” J. Opt. Soc. Am. A 13,
pp. 1903–1915, 1996.
23. S. Chatigny, M. Morin, D. Asselin, Y. Painchaud, and P. Beaudry, “Hybrid monte
carlo for photon transport through optically thick scattering media.,” Appl. Opt.
38, pp. 6075–6086, 1999.
Topological Derivatives for Shape
Reconstruction

Ana Carpio1 and Maria Luisa Rapún2

1
Departamento de Matemática Aplicada, Facultad de Matemáticas, Universidad
Complutense de Madrid, Plaza de Ciencias 3, 28040 Madrid, Spain
ana [email protected]
2
Departamento de Fundamentos Matemáticos de la Tecnologı́a Aeronáutica, Escuela
Técnica Superior de Ingenieros Aeronáuticos, Universidad Politécnica de Madrid,
Plaza del Cardenal Cisneros 3, 28040 Madrid, Spain
[email protected]

Summary. Topological derivative methods are used to solve constrained optimiza-

tion reformulations of inverse scattering problems. The constraints take the form of
Helmholtz or elasticity problems with different boundary conditions at the interface
between the surrounding medium and the scatterers. Formulae for the topological deriv-
atives are found by first computing shape derivatives and then performing suitable
asymptotic expansions in domains with vanishing holes. We discuss integral methods
for the numerical approximation of the scatterers using topological derivatives and im-
plement a fast iterative procedure to improve the description of their number, size,
location and shape.

1 Introduction
A huge number of applications lead to inverse scattering problems whose goal is
to detect objects buried in a medium [2, 11, 17]. Such is the case of ultrasound
in medicine, X-ray diffraction to retrieve information about the DNA, reflection
of seismic waves in oil prospecting or crack detection in structural mechanics. A
standard procedure to find information about the obstacles consists in emitting
a certain type of radiation which interacts with the objects and the surrounding
medium and is then measured at the detector locations. The total field, con-
sisting of emitted, scattered and transmitted waves, solves a partial differential
equation (Maxwell’s equations, radiative transfer equation, equations of elastic-
ity, Helmholtz equation. . . ) with boundary conditions at the interface between
the scatterers and the medium. The original inverse problem is the following:
knowing the emitted waves and the measured patterns, find the obstacles for
which the solution of the corresponding boundary value problem agrees with the
measured values at the detector locations. There is a large literature on how
to solve it, see [8, 11] for classical results and [2, 17] for more recent advances.
86 A. Carpio and M.L. Rapún

Some of the methods proposed to solve the nonlinear inverse problem are based
on linear approximations such as the Born approximation [14], the Kirchhoff or
physical optics approximation [36] or backpropagation principles [41]. The linear
sampling method amounts to solving an equation for the contour of the unknown
domain or its representative coefficients [9, 12].
Here, we will follow the variational approach. The original formulation is
too restrictive [8, 30] and weaker formulations have been proposed in the vari-
ational setting [18, 29], leading to optimization problems with constraints given
by boundary value problems for partial differential equations. Most optimization
strategies rely on gradient principles, see for instance [33]. They deform an initial
guess for the contours of the obstacles through an iterative process in such a way
that the derivative of the functional to be minimized decreases. The techniques
differ in the type of deformation, the representation of the contours or the way
to initialize the process. The first attempts relied on classical shape optimiza-
tion using a small perturbation of the identity [27, 31, 38, 44, 52]. However, this
strategy does not allow for topological changes in the obstacle, i.e., the number
of components has to be known from the beginning. This problem was solved
by introducing deformations inspired in level-set methods [37, 50]. In this way,
contours may be created or destroyed during the iterative process. Nevertheless,
level-set based iterative methods may be rather slow unless a good initial guess
of the scatterers is available.
Topological derivative methods arise as an efficient alternative to solve inverse
scattering problems [53]. There is a growing literature on the subject, see for
instance [2, 18, 26] for scattering problems in acoustics, [21, 25] for the elasticity
case and [17, 39, 49] for problems involving electromagnetic waves. In scattering
problems, the topological derivative of the cost functional under study provides
a good first guess of the number and location of obstacles, as shown in [18, 25].
This guess may be improved using iterative schemes entirely based on topological
derivatives, as discussed in [4,21]. Hybrid level set-topological derivative methods
have been suggested in [3].
The paper is organized as follows. Section 2 discusses topological derivative
methods when the interaction between the scatterers, the medium and the in-
cident radiation is described by Helmholtz problems. We describe the forward
scattering model and reformulate the inverse problem as a constrained optimiza-
tion problem. Then, we recall the expressions for the topological derivatives of
the resulting functionals with different boundary conditions. Reconstructions of
objects obtained using topological derivatives with synthetic data are shown in
Sect. 3. Good approximations for the number, location and size of the scatterers
are found. Section 4 presents an iterative scheme which improves the description
of the shape of the obstacles in a few iterations. Section 5 gives some details
on integral methods for the numerical approximation of topological derivatives.
In Sect. 6, the analytic procedure to compute explicit expressions for the topo-
logical derivatives in Helmholtz problems is explained. Extensions to detection
of objects by means of acoustic waves in elastodynamic problems are presented
in Sect. 7. Finally, conclusions and directions for future work are discussed in
Sect. 8.
 Topological Derivatives for Shape Reconstruction 87

2 Helmholtz Equations

2.1 The Forward and Inverse Scattering Problems

Let us assume we have a certain medium where a finite number of obstacles

Ωi,j , j = 1, . . . , d are buried (the subscript i stands for interior), as in Fig. 1. A
classical procedure to locate the unknown obstacles, consists in illuminating the
medium with a particular type of wave (electromagnetic, acoustic, pressure. . . ).
The incident radiation interacts with the medium and the obstacles, is reflected
and then measured at a set of receptors Γmeas , placed far enough from the
objects. Knowing the emitted and measured wavefields, it might be possible to
find the obstacles. Many detecting devices follow this principle: radars, sonars,
lidars, scanners. . . .
The total field u = uinc + usc formed by the incident and the scattered wave
satisfies a boundary problem for some partial differential equation (Maxwell’s
equations, elasticity equations . . . ). This section develops techniques to detect
objects when the governing partial differential equation is a Helmholtz equation.
This happens in simple geometries for either electromagnetic (TM, TE polarized
waves) or acoustic waves [17, 18, 24].
Let us make precise the setting we will be working in. The recepting set Γmeas
may be a curve in simple tests, or a number of discrete points at which receptors
are located in more realistic situations. To simplify, we take the medium to be R2
and the incident radiation to be a plane wave uinc (x) = exp(ı λe x · d) with wave
number λe and propagation direction d, |d| = 1. The obstacle Ωi ⊂ R2 is an open
bounded set with smooth boundary Γ := ∂Ωi but has no assumed connectivity.
There may be an unknown number of isolated components: Ωi = ∪dj=1 Ωi,j in
which Ωi,j are open connected bounded sets satisfying Ω i,l ∩ Ω i,j = ∅ for l = j.
When the incident wave penetrates the scatterer (‘penetrable’ obstacle), part
of the wave is reflected and the rest is transmitted inside. Transmission conditions
are imposed at the interface between the obstacle and the surrounding medium.

ΓR
Γmeas

Ωi,1n
Ωi,2 n

Ωi,3n

Ωe

d
Fig. 1. Geometry of the problem: Ωi = j=1 Ωi,j and Ωe = R2 \ Ω i
88 A. Carpio and M.L. Rapún

The total wavefield u = uinc + usc in Ωe = R2 \ Ω i (the subscript e stands for

exterior) and the transmitted wave field u = utr in Ωi solve
⎧
⎪
⎪ 2
⎪ ∆u + λe u = 0,
⎪
in Ωe ,
⎪
⎪
⎪
⎪
⎪
⎪ αi ∆u + λ2i u = 0, in Ωi ,
⎪
⎪
⎨
u− − u+ = 0, on Γ , (1)
⎪
⎪
⎪
⎪
⎪
⎪ αi ∂n u− − ∂n u+ = 0,
⎪
⎪ on Γ ,
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂r (u − uinc ) − ıλe (u − uinc )) = 0, r = |x|.
r→∞

We select the unit normal vector n pointing outside Ωe . ∂n stands for the normal
derivative at the interface and ∂r is the radial derivative. We denote by u+ and
u− the limits of u from the exterior and interior of Ωi respectively. To simplify,
the parameters λe , λi , αi are taken to be constant, real and positive. With this
assumption, we can impose the standard Sommerfeld radiation condition on the
propagation of the scattered field usc = u − uinc at infinity, which implies that
only outgoing waves are allowed.
General conditions on the parameters λe , λi , αi ensuring existence and
uniqueness of a solution u ∈ Hloc 1
(R2 ) for this problem can be found for in-
stance in [13, 32, 34, 54, 55].
Alternatively, we might reformulate (1) as a boundary value problem for the
scattered and the transmitted fields with transmission conditions utr −usc = uinc
and αi ∂n utr − ∂n usc = ∂n uinc . This is usually done for numerical purposes.
However, (1) simplifies the later computation of shape and topological derivatives
for this problem. The parameter αi is not scaled out in the second equation for
the same reason.
When the scatterer is opaque to the incident radiation (‘sound hard’ obsta-
cle), there is no transmitted wave and a Neumann boundary condition is imposed
at the interface:
⎧
⎪
⎪ ∆u + λ2e u = 0, in Ωe ,
⎪
⎪
⎨
∂n u = 0, on Γ , (2)
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂ (u − u ) − ıλ (u − u )) = 0, r = |x|.
r inc e inc
r→∞

The forward problem consists in solving (1) (resp. (2)) when the obstacles
Ωi and the incident field uinc are known. Evaluating the solution at Γmeas , we
recover the measured pattern: u|Γmeas = umeas (d). In practical experiments, the
total field u is known on the set of receptors Γmeas for different directions of
the incident wave dj . The inverse problem consists in finding the shape and
structure of the obstacle Ωi such that the solution of the forward problem (1)
(resp. (2)) equals the measured values umeas (dj ) at the receptors.
 Topological Derivatives for Shape Reconstruction 89

This inverse problem is nonlinear with respect to Ωi and strongly ill-posed [8].
Given arbitrary data umeas , an associated scatterer Ωi may not exist. When it
exists, it may not depend continuously on umeas . It is well-known that the ob-
stacles are uniquely determined by the far-field pattern of the scattered wave for
all incident directions and one fixed wave number λe (see the relevant paper [30]
and the latter work [23] where the proofs in [30] for the transmission problem
were simplified). Therefore, by an analyticity argument (see [11]), if Γmeas is a
circumference, then the values of the total wave on Γmeas for all incident waves
determine uniquely the obstacles. The question of uniqueness without any a pri-
ori knowledge about the location of the obstacles for just one (or a finite number
of) incident plane waves is still an open problem.
Different strategies to regularize the inverse problem have been proposed in
the literature. We resort here to the variational approach: look for a domain Ωi
which minimizes an error in some sense. This leads to a constrained optimization
problem: minimize


1
J(Ωi ) := |u − umeas |2 dl, (3)
2 Γmeas

where u is the solution of the forward problem (1) (resp. (2)) and umeas the
measured total field on Γmeas . This functional depends on the design variable
Ωi through the boundary value problem, which acts as a constraint. When sev-
eral measurements corresponding to different directions of illumination dj are
available, the optimization problem becomes: minimize
N

1
J(Ωi ) := |uj − ujmeas |2 dl, (4)
2 j=1 Γmeas

in which uj are the solutions of N forward problems with incident waves

ujinc (x) = exp(ı λe x · dj ). We have now N constraints.

2.2 Topological Derivatives Applied to Inverse Scattering

Different notions of derivative have been introduced for shape functionals. The
topological derivative of a functional measures its sensitivity when infinitesi-
mal holes are removed from a domain. When applied to functionals associated
to inverse scattering problems such as (3), the topological derivative provides
information on the location of the scatterers: Regions where the topological
derivative takes on large negative values will be identified with places where an
object should be located.
The standard formal definition is the following. Let us consider a small ball
Bε (x) = B(x, ε), x ∈ R ⊂ R2 , and the domain Rε := R\B ε (x). The topological
derivative of J (R) is a scalar function of x ∈ R defined as

J (Rε ) − J (R)
DT (x, R) := lim . (5)
ε→0 V(ε)
90 A. Carpio and M.L. Rapún

The scalar function V(ε) is chosen in such a way that the limit (5) exists, is
finite and does not vanish. V(ε) is usually related to the measure of the ball. In
our case, V(ε) = −πε2 . In general, it will be a decreasing and negative function,
satisfying limε→0 V(ε) = 0. The value of the limit (5) depends on the partial
differential equation we use as a constraint and on the boundary conditions we
impose on the boundary of the hole. This limit also depends on the kind of hole
we are removing. It would change slightly for shapes different from a ball.
The relationship

J (Rε ) = J (R) + V(ε) DT (x, R) + o(V(ε)),

allows to establish a link with the shape derivative (see [18]), which is useful to
obtain explicit expressions. The shape derivative of a functional J (R) along a
vector field V(z) is defined as
$
d $
DJ (R) · V := J (ϕτ (R))$$ ,
dτ τ =0

where ϕτ are deformations along the field V:

ϕτ (z) := z + τ V(z) =: zτ , z ∈ R2 .

To compute topological derivatives, we select a smooth vector field of the form

V(z) = Vn n(z), z ∈ ∂Bε (x),

with constant Vn < 0 and extend V to R2 in such a way that it vanishes out of
a neighborhood of ∂Bε . Then,
 $ 
1 d $
DT (x, R) = lim J (ϕτ (R ))
ε $
$ , x ∈ R, (6)
ε→0 V  (ε)|Vn | dτ
τ =0

where V  (ε) is the derivative of the function V(ε).

The value of the topological derivative depends on the set R. When there is
no a priori information on the scatterers, one computes the topological derivative
of the functional (3) setting R = R2 and Ωi = ∅ to obtain a first guess of the
number and location of the obstacles. As we will show later, a first approximation
to the obstacles is found by selecting a domain Ωi1 where DT (x, R2 ) is smaller
1
than a certain large negative value. Setting now R = R2 \ Ω i , we may compute
1
DT (x, R2 \ Ω i ) and select a new approximation Ωi2 . Iterating this procedure,
we get a sequence of approximations Ωik . At each step, we remove from R2 a
set Ωik in which the topological derivative is large and negative, the value of
k
DT (x, R2 \ Ω i ) becomes smaller and Ωik gets closer to the true scatterer Ωi .
Using (6), we find explicit expressions for the required topological derivatives
of (3). We collect the main results about Helmholtz transmission and Neumann
problems in two theorems whose proofs are rather technical and will be post-
poned till Sect. 6. Let us first consider the topological derivative in the whole
space R2 .
 Topological Derivatives for Shape Reconstruction 91

Theorem 2.1. Depending on the boundary conditions at the interface between

the medium and the obstacles, the topological derivative of the cost functional
(3) in R = R2 takes the following form:
• General transmission problem (penetrable obstacle)

2(1 − αi )
DT (x, R2 ) = Re ∇u(x)∇p(x) + (λ2i − λ2e ) u(x)p(x) , (7)
1 + αi

• Transmission problem with αi = 1

" #
DT (x, R2 ) = Re (λ2i − λ2e ) u(x) p(x) , (8)

• Neumann problem (rigid obstacle)

" #
DT (x, R2 ) = Re 2∇u(x)∇p(x) − λ2e u(x)p(x) , (9)

for x ∈ R2 . In all cases, u and p solve forward and adjoint problems with Ωi = ∅.
The solution u of the forward problem is the incident wave uinc (x). The adjoint
state p solves
⎧
⎪
⎨ ∆p + λ2e p = (umeas − u)δΓmeas , in R2 ,
(10)
⎪
⎩ lim r1/2 (∂r p + ıλe p) = 0,
r→∞

where δmeas is the Dirac delta function on the sampling interface Γmeas .

Theorem 2.1 holds regardless of the structure of the incident wave. It may
be a plane wave uinc (x) = exp(ı λe d · x), or a different type of source, as in
Sect. 3. Notice that the forward and adjoint solutions needed for the computation
of topological derivatives are independent of the boundary conditions at the
interface. However, they affect the final expression when calculating the limit
ε → 0.
Let us see how (7)–(9) change when we compute the topological derivative
of (3) in a domain with a hole R = R2 \ Ω, Ω being an open bounded set, not
necessarily connected. The new expression is formally identical, but the adjoint
and forward problems are solved in Ωe = R2 \ Ω and Ωi = Ω, with transmission
or Neumann boundary conditions at the interface Γ = ∂Ω. Now, the boundary
conditions not only determine the limit value but also affect the forward and
adjoint fields u and p.

Theorem 2.2. Depending on the boundary conditions at the interface, the topo-
logical derivative of the cost functional (3) in R = R2 \ Ω is given by
• General transmission problem

2(1 − αi )
DT (x, R2 \ Ω) = Re ∇u(x)∇p(x) + (λ2i − λ2e ) u(x)p(x) , (11)
1 + αi
92 A. Carpio and M.L. Rapún

where u solves the forward transmission problem (1) with Ωi = Ω and p

solves the adjoint transmission problem
⎧
⎪
⎪ ∆p + λ2e p = (umeas − u) δΓmeas , in Ωe ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ αi ∆p + λ2i p = 0, in Ωi ,
⎪
⎪
⎨
p− − p+ = 0, on Γ , (12)
⎪
⎪
⎪
⎪
⎪
⎪ −
⎪ αi ∂n p − ∂n p = 0,
+
⎪ on Γ ,
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂r p + ıλe p) = 0,
r→∞

• Transmission problem with αi = 1

" #
DT (x, R2 \ Ω) = Re (λ2i − λ2e ) u(x) p(x) , (13)

where u and p solve the forward and adjoint transmission problems (1) and
(12) with Ωi = Ω and αi = 1,
• Neumann problem
" #
DT (x, R2 \ Ω) = Re 2∇u(x)∇p(x) − λ2e u(x)p(x) , (14)

where u solves the Neumann problem (2) with Ωi = Ω and p solves the adjoint
Neumann problem
⎧
⎪
⎪ ∆p + λe p = (umeas − u) δΓmeas ,
2
⎪ in Ωe ,
⎪
⎨
∂n p = 0, on Γ , (15)
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂ p + ıλ p) = 0,
r e
r→∞

for x ∈ R2 \ Ω.
The Neumann case is recovered taking the limit αi → 0, as expected [47].
Analogous formulae hold in three dimensions. The gradient term takes the
form ∇uP∇p, where P depends on the solutions of auxiliary transmission prob-
lems in a fixed ball, which change with dimension, see [24]. Inhomogeneous prob-
lems with variable parameters can be handled in a similar way, see [4].
The next sections present several tests illustrating the ability of topological
derivative based methods to locate scatterers.

3 Numerical Computation of Topological Derivatives

In this section, we reconstruct scatterers buried in a medium by schemes based
on the computation of topological derivatives. Several numerical experiments
illustrate the main advantages and limitations of the method. We explain the
 Topological Derivatives for Shape Reconstruction 93

technique in the model case of the Helmholtz transmission problem discussed in

Sect. 2. The method can be extended to elasticity problems using the results in
Sect. 7.
The idea is to plot DT (x, R2 ) on a grid of points to obtain a map showing
the regions where DT (x, R2 ) is large or small. Domains where the topological
derivative takes large negative values indicate the possible location of an object.
Notice that the true scatterers enter the adjoint problem through the measured
data umeas . Our tests here use artificial data generated by solving the forward
transmission problem for Ωi and evaluating its solution at Γmeas . Alternatively,
real scattering data might be used.
In practice, one generates one or several different incident fields and measures
the resulting total field not on a whole interface Γmeas but on a finite set of
receptors {x1 , . . . , xM }. We will only consider two kinds of incident fields: plane
waves, which are modeled by the functions
uinc (x) := exp(ıλe x · d), |d| = 1, (16)
where d is the propagation direction, and point sources, which are described by
the functions
⎧
⎪ ı (1)
⎪
⎨ 4 H0 (λe |x − z|), in 2D,
uinc (x) := (17)
⎪
⎪ exp(ıλe |x − z|)
⎩ , in 3D,
4π|x − z|
(1)
where z is the focus and H0 is the Hankel function of the first kind and order
zero.
Our goal is to minimize the functional (compare with (3))
M
1
J(Ωi ) := |u(xk ) − umeas (xk )|2 , (18)
2
k=1

where umeas (xk ) is the measured value of the total field at the k-th receptor xk
and u is the solution to the Helmholtz transmission problem associated to the
domain Ωi for an incident wave of the form (16) or (17). It is common in practice
to know the total fields for several incident waves, that is, for incident fields
corresponding to different propagation directions d1 , . . . , dN , or for different
point sources located at z1 , . . . , zN . In this case, we minimize the cost functional
(which is the discrete version of (4))
N M
1
J(Ωi ) := |uj (xk ) − ujmeas (xk )|2 , (19)
2 j=1 k=1

where uj is the solution to the forward transmission problem for the planar
incident field in the direction dj or the field generated at the source point zj ,
and ujmeas (xk ) is the measured total field at the observation point xk .
In this section, we focus on the practical situation in which one does not have
any information about the number of obstacles or its location and therefore starts
94 A. Carpio and M.L. Rapún

by assuming that no obstacles exist. We set Ωi = ∅ and Ωe = Rn , with n = 2

or n = 3. Then, the solution to the direct problem is simply the incident wave,
given by the explicit formulae (16) or (17). The adjoint problem for p is:
⎧
⎪
⎪
M
⎪
⎨ ∆p + λ2e p = (umeas (xk ) − u(xk ))δxk ,
k=1 (20)
⎪
⎪
⎪
⎩ lim r(n−1)/2 (∂r p − ıλe p) = 0.
r→∞

An explicit formula for p is obtained using the outgoing fundamental solutions

of the Helmholtz equation ∆u + λ2 u = 0
⎧ı
⎪ (1)
⎨ 4 H0 (λ|x − y|),
⎪ in 2D,
φλ (x, y) := (21)
⎪
⎪ exp(ıλ|x − y|)
⎩ , in 3D,
4π|x − y|
which satisfy the Sommerfeld radiation condition at infinity and ∆u + λ2 u = δy .
Then, the solution of the adjoint problem (20) is
M
p(x) = (umeas (xk ) − u(xk )) φλe (x, xk ). (22)
k=1

We are ready now to compute the topological derivative of the cost functionals.
In the 2D case, the topological derivative of (18) is
2(1 − αi )
DT (x, R2 ) = Re ∇u(x)∇p(x) + (λ2i − λ2e )u(x)p(x) ,
1 + αi
formally identical to the result in Theorem 2.1, but now the adjoint field is given
by (22). The topological derivative of the cost functional (19) is
N
2(1 − αi )
DT (x, R2 ) = Re ∇uj (x)∇pj (x) + (λ2i − λ2e )uj (x)pj (x) , (23)
j=1
1 + αi

with pj given by (22) with u and umeas replaced by uj and ujmeas , respectively.
In 3D, the factor 2(1 − αi )/(1 + αi ) should be replaced by 3(1 − αi )/(2 + αi ) in
the case of spherical holes (see [4, 24]).
If we had a priori information of the number, size and location of the ob-
stacles, we might set Ωi = ∅. In this case, u is not just the incident wave but
the solution to the Helmholtz transmission problem for Ωi . The adjoint problem
is also slightly modified, and again a transmission problem has to be solved, as
stated in Theorem 2.2. This case can be seen as the second step in an iterative
method based on topological derivatives and will be studied in Sect. 4.
Let us investigate the performance of the method when no a priori infor-
mation is available (that is, starting with Ωi = ∅) for scattering problems of
time-harmonic acoustic waves. A time-harmonic incident field has the form
 Topological Derivatives for Shape Reconstruction 95

Uinc (x, t) = Re(uinc (x)e−ıωt ),

for a frequency ω > 0. The complex amplitude of the incident field, uinc (x),
will be a planar wave, given by (16) or a point source, given by (17). Asymp-
totically, it generates a time-harmonic response U (x, t) = Re(u(x)e−ıωt ). The
space-dependent amplitudes of the scattered wave u in Ωe and the transmitted
wave u = utr in Ωi solve the Helmholtz transmission problem (1) with wave
√
numbers λk = ωρk , k = i, e, where ρk is the density of the material occupying
the domain Ωk .
We start by describing a simple 2D example in a material with a single
obstacle inside. The experiment consists in measuring the total field on sampling
points located on a circumference, for planar incident fields of the form (16) with
directions dj := (cos θj , sin θj ) . In our numerical experiments, the propagation
angles θj are always uniformly distributed on the whole interval [0, 2π) or on a
general interval [β1 , β2 ] ⊂ [0, 2π). We take αi = 1, and densities ρe , ρi satisfying
*
λe ρe
= = 5.
λi ρi
The composite material is excited at different frequencies ω corresponding to
wave numbers λe and λi satisfying the ratio above.
In Fig. 2 we represent the topological derivative (23) obtained for measure-
ments of the total wavefield at the 14 sampling points marked with the symbol

(a) (b)
3 3
−1 −4
2 2 −6
−2
−8
1 −3 1
−10
0 −4 0 −12
−14
−1 −5 −1
−16
−6 −18
−2 −2
−7 −20
−3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

(c) (d)
3 40 3 60

30
2 2 40
20
1 1 20
10
0 0 0
0

−1 −10 −1 −20

−2 −20 −2 −40

−30
−3 −3 −60
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

Fig. 2. Topological derivative with data from 24 incident directions uniformly distrib-
uted in [0,2π) (every 15◦ ) at the 14 receptors (‘×’). (a) λe = 1.25, λi = 0.25, αi = 1
(b) λe = 2.5, λi = 0.5, αi = 1 (c) λe = 5, λi = 1, αi = 1 (d) λe = 10, λi = 2, αi = 1
96 A. Carpio and M.L. Rapún

‘×’ for 24 incident planar waves with angles in [0, 2π). Incident fields were gen-
erated at different frequencies ω, increasing its values from (a) to (d). In the
four simulations we detect the presence of just one obstacle. Moreover, we guess
its location and approximate size. For some of the frequencies, the approximate
shape of the obstacle is also quite well recovered. Low and high frequencies pro-
vide different information about the obstacle. At low frequencies (see Fig. 2a,b),
the lowest values of the topological derivative are reached inside the obstacle,
located in a region of circular or elliptical shape. This allows for good recon-
structions of the location and size of the obstacle, but gives little information
about its shape. On the other hand, for high frequencies (see Fig. 2c,d), the
lowest values of the topological derivative are concentrated in a small annular
region about the boundary. Increasing the frequency, the annular region becomes
thinner. With this information, one can predict more accurately the exact shape
of the obstacle. For the same amount of data, the topological derivative is less
smooth at high frequencies (in the sense that it shows strong oscillations) than
at low frequencies. This fact makes probably more difficult to distinguish the
obstacle.
Considering a larger number of receptors and/or a larger number of incident
waves, the reconstructions for the low frequencies improve only a bit, but the
quality for the higher ones is considerably better, as can be seen in Fig. 3.

(a) (b)
3 3 −40
−10
−60
2 −20 2
−80
−30 −100
1 1
−40 −120
0 0 −140
−50
−160
−1 −60 −1 −180
−70 −200
−2 −2 −220
−80
−240
−3 −3
−90
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
(c) (d)
3 3 600
400
2 300 2 400

200 1 200
1
100
0
0 0
0
−200
−1 −100 −1
−200 −400
−2 −2
−300 −600
−3 −400 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

Fig. 3. Topological derivative with data from 72 incident directions uniformly distrib-
uted in [0,2π) (every 5◦ ) at the 60 receptors (‘×’). (a) λe = 1.25, λi = 0.25, αi = 1 (b)
λe = 2.5, λi = 0.5, αi = 1 (c) λe = 5, λi = 1, αi = 1 (d) λe = 10, λi = 2, αi = 1
 Topological Derivatives for Shape Reconstruction 97

(a) (b)
3 3
40
−10
2 2
20
−15
1 1
−20 0
0 0
−25 −20
−1 −1
−30 −40
−2 −2
−35 −60
−3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
(c) (d)
3 3

−10 40
2 2

−15 20
1 1

0 −20 0 0

−1 −25 −1 −20

−2 −30 −2 −40

−3 −35 −3 −60
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

Fig. 4. Topological derivative with data from 24 incident directions uniformly distrib-
uted in [0,2π) (every 15◦ ) at the 24 receptors (‘×’). (a, c) λe = 2.5, λi = 0.5, αi = 1
(b, d) λe = 5, λi = 1, αi = 1

Let us analyze the performance of the method when dealing with more com-
plex geometries. We show results for composite materials with two and three
inclusions in Fig. 4a,b and Fig. 4c,d respectively. The topological derivative is
computed for a low frequency in Fig. 4a,c and for a high one in Fig. 4b,d. The
quality of the reconstruction of the boundaries is acceptable near the observa-
tion points. The reconstruction of other parts of the boundaries is made more
difficult by the influence of each obstacle on the others. Notice that the over-
all reconstruction worsens in comparison with the case of a single obstacle. In
Fig. 4a,b we seem to detect three inclusions instead of two. The small defect is
almost undistinguishable in Fig. 4c,d. Furthermore, at low frequencies neither
shapes nor locations are well recovered.
In the previous examples we considered planar incident waves. We explore
now the choice of point sources. We repeat the experiment for the configuration
with two obstacles, keeping the same frequencies and receptors as in Fig. 4a,b,
but working now with incident fields generated at 24 uniformly distributed points
on the circumference. The location of the source points is represented in Fig. 5
with a ‘•’ symbol. The quality of the reconstruction is similar for point sources
and planar waves.
98 A. Carpio and M.L. Rapún

(a) (b)
3 −0.04 3
0.1
2 −0.06 2
−0.08 0.05
1 1
−0.1 0
0 −0.12 0
−0.05
−0.14
−1 −1
−0.16 −0.1
−2 −2
−0.18 −0.15
−3 −0.2 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

Fig. 5. Topological derivative with data from 24 incident fields generated at the source
points marked with ‘•’ and measured at the 24 receptors marked with ‘×’. (a) λe =
2.5, λi = 0.5, αi = 1 (b) λe = 5, λi = 1, αi = 1

We describe now numerical reconstructions when the observation points and

the sources are distributed over an angle of limited aperture. As one can expect,
the illuminated regions will allow for better reconstructions.
The experiments in Fig. 6 have been carried out for the configuration with
two obstacles at low frequencies corresponding to wavenumbers λe = 2.5 and
λi = 0.5. In Fig. 6a,b, sampling and source points are located on the same half
of a circumference. Only illuminated regions provide reliable information. We
recover the upper part of the obstacles in Fig. 6a and their lower part in Fig. 6b.
In Fig. 6c,d source and observation points are located in complementary half
circumferences, covering the whole circumference. In contrast with the results
shown in Fig. 5, we predict the occurrence of only two obstacles, but their ap-
proximate sizes and shapes are poorly reproduced.
We have also investigated what happens when the source and sampling points
are located on a half circumference close to one obstacle but far from the other,
see Fig. 6e. As one can expect, we clearly distinguish the presence of the nearest
obstacle to the sources and receptors. Its size and shape is quite well recovered
although its location is displaced to the left (compare with the reconstruction
in Fig. 5 where this obstacle seems to be split into two different ones, but the
reconstruction of its location is more accurate). The object located in the shadow
region is completely ignored by the topological derivative. Finally, in Fig. 6f, data
are sampled in a quarter of the circumference and only the inferior part of the
illuminated obstacle is detected.
Our numerical experiments are restricted here to Helmholtz transmission
problems with αi = 1 because formula (8) is slightly simpler to implement.
Numerical computations for general transmission problems with αi = 1 yield
analogous conclusions. The interested reader can find some examples in [4]. For
tests with the Neumann problem we refer to [18].
We end this section by inspecting the 3D case. Our goal is to recover the
shape, size and location of a sphere when the total wavefield is known on sam-
pling points that are located in each of the two configurations shown in Fig. 7.
 Topological Derivatives for Shape Reconstruction 99

(a) (b)
3 3 0.04
0.04
2 2 0.02
0.02
1 1 0
0

0 0 −0.02
−0.02

−0.04 −0.04
−1 −1
−0.06 −0.06
−2 −2
−0.08 −0.08
−3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
(c) (d)
3 −0.01 3 −0.01
−0.02
2 2 −0.02
−0.03
1 1 −0.03
−0.04
−0.04
0 0
−0.05
−0.05
−1 −0.06 −1
−0.06
−2 −0.07 −2
−0.07
−0.08
−3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3
(e) (f)
3 0.05 3 0.02
0.015
2 2 0.01
0.005
1 0 1
0
0 0 −0.005
−0.01
−1 −0.05 −1 −0.015
−0.02
−2 −2
−0.025
−3 −0.1 −3 −0.03
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

Fig. 6. Topological derivative when the incident fields are generated at the source
points marked with ‘•’ and the total field is measured at the receptors marked with
‘×’. λe = 2.5, λi = 0.5, αi = 1
(a) (b)

3 3
2 2
1 1
0 0
−1 −1
−2 −2
−3 −3
2 2
2 2
0 0
0 0
−2 −2 −2 −2

Fig. 7. Location of the observation points

100 A. Carpio and M.L. Rapún

(a) (b)

1 1
–5
0.8 0.8 –50

0.6 0.6
–10
0.4 0.4 –100
0.2 0.2
–15
0 0
–0.2 –150
–0.2
–20
–0.4 –0.4
–0.6 –0.6 –200
–25
–0.8 –0.8
–1 –0.1
1 0.5 1 –30 1 0.5 1 –250
0 –0.5 –1 –0.5 0 0.5 0 –0.5 –1 –0.5 0 0.5

Fig. 8. Topological derivative with data from 22 incident waves for directions
(cos θj , sin θj , 0) . The angles θj are uniformly distributed in [0, 2π). The receptors
are plotted in Fig. 7a. (a) λe = 2, λi = 1, αi = 1, (b) λe = 4, λi = 2, αi = 1

For the reconstructions in Fig. 8, we have considered 21 planar incident waves

with directions
dj := (cos θj , sin θj , 0)
and angles θj uniformly distributed in [0, 2π). Notice that all these directions
are contained in the plane {(x, y, z) ∈ R3 , z = 0}. The sampling points are
represented with a ‘•’ symbol in Fig. 7a. All of them belong to the cylinder
{(x, y, z) ∈ R3 , x2 + y 2 = 9}. This particular choice of incident fields and
sampling points produces acceptable reconstructions in the x and y coordinates,
but not in the z coordinate. More than a sphere, the obstacle seems to be an
ellipsoid with a larger semiaxis in the z-direction than in the x and y directions.
This feature is specially stressed at low frequencies, as shown in Fig. 8a.
In order to improve the results, we consider now the observation points rep-
resented in Fig. 7b, which belong to one of the following cylinders, {(x, y, z) ∈
R3 , x2 + y 2 = 9}, {(x, y, z) ∈ R3 , x2 + z 2 = 9}, {(x, y, z) ∈ R3 , y 2 + z 2 = 9}.
We have tried planar incident waves in the directions

(cos θj , sin θj , 0) , (cos θj , 0, sin θj ) , (0, cos θj , sin θj ) ,

for seven angles θj uniformly distributed in [0, 2π) (the total amount of incident
waves is therefore 21, as in the previous example). The resulting topological
derivatives for the same low and high frequencies are shown in Fig. 9. In this
simple geometrical configuration with a single spherical object, both low and
high frequencies provide accurate reconstructions when choosing a suitable dis-
tribution of sampling points and incident directions.
As in the 2D case, when we consider several obstacles and more complex
shapes, the influence of one obstacle on the others as well as the appearance of
illuminated and shadow regions result in poorer reconstruction quality.
In conclusion, for simple geometries and configurations, the method produces
quite satisfactory approximations of the number, shape, size and location of the
obstacles. If we consider more complicated problems, then, the method provides
 Topological Derivatives for Shape Reconstruction 101

(a) (b)
0

1 –5 1
–50
0.8 –10 0.8
0.6 0.6 –100
–15
0.4 0.4
0.2 –20 0.2 –150
0 0
–25 –200
–0.2 –0.2
–0.4 –30 –0.4
–250
–0.6 –0.6
–35
–0.8 –0.8 –300
–40 –1
–1
1 0.5 1 1 0.5 1 –350
0 –0.5 –0.5 0 0.5 –45 0 –0.5 0 0.5
–1 –1 –0.5

Fig. 9. Topological derivative with data from seven incident waves with directions
(cos θj , sin θj , 0) , seven incident waves with directions (cos θj , 0, sin θj ) and seven in-
cident waves with directions (0, cos θj , sin θj ) . The angles θj are uniformly distributed
in [0, 2π). The receptors are plotted in Fig. 7b. (a) λe = 2, λi = 1, αi = 1, (b) λe = 4,
λi = 2, αi = 1

at least a good first guess for more sophisticated iterative methods which can
circumvent the problems observed in the previous examples.

4 An Iterative Method Based on Topological Derivatives

There are several ways of improving the first guesses obtained by the method
described in the previous section. One alternative is to make small deformations
of the initial guess to reduce the cost functional, as done in [27]. However, this
has the disadvantage of not allowing for topological changes in the obstacles,
that is, the number of contours has to be known from the beginning. We are
more interested in numerical methods with the ability of creating new obstacles,
merging two or more contours or even destroying existing ones. A possibility is
to use the approximation of the obstacle provided by the topological derivative
combined with an optimization scheme based in level-set methods to improve
the reconstruction, as done in [37] (see also [17, 50]). Another option is to intro-
duce the topological derivative in the Hamilton–Jacobi equation controlling the
evolution of the level-sets, as proposed in [3].
In this section we will develop an efficient and fast numerical method based
on the computation of successive topological derivatives. This approach was
suggested in [21] for elasticity problems with Dirichlet and Neumann boundary
conditions. We restrict to 2D experiments to reduce the computational cost in
the tests, but the method applies without further difficulties (in principle) to 3D
models. Again, the model problem will be the Helmholtz transmission problem,
but the theory extends to other equations.
The idea is quite simple. We start by computing the topological derivative
as in the previous section, that is, taking Ωi = ∅. Then, we look at the values of
the topological derivative that fall below a certain negative threshold and create
the initial guess Ω1 of Ωi . In the next iteration, we compute the topological
102 A. Carpio and M.L. Rapún

derivative in a new setting, with R2 \ Ω 1 playing the role of R2 . Then, we deter-

mine the points where the lowest values of the topological derivative are reached.
If one of those points is close to any of the existing obstacles, we conclude that
it belongs to that obstacle. If they are far enough from the existing obstacles, a
new obstacle is created. Finally, if some of those points are near two obstacles,
then we also consider the possibility of merging them. Once the configuration is
updated, we repeat the procedure in the next step. As we will see in the numer-
ical experiments, the topological derivative also contains information about the
points that should not be included in an obstacle, although we had decided to
include them in a previous step.
There are two questions to answer now. The first one is how to compute the
topological derivative when Ωi = ∅ and the second is how to characterize the
obstacles at each step in a suitable way for the numerical computations.
We will devote Sect. 5 to the first issue. The forward and the adjoint problems
for Ωi = ∅ are solved by a fully discrete version of a Galerkin method with
trigonometric polynomials applied to the system of boundary integral equations
that is obtained when using Brakhage–Werner potentials.
Let us address the second question. Typically, boundary element discretiza-
tions are based on the hypothesis that one can describe the boundary Γ by
a (smooth) parametrization of the boundary. We describe below a numerical
method for the practical computation of such parametrization in the 2D case
proposed in [4]. This is essential to determine the first guess Ω1 knowing the
topological derivative when Ωi = ∅, or in general, to construct the approximation
Ωk+1 from the lowest values reached by the topological derivative when Ωi = Ωk .
First of all, by simple inspection we determine the number of components of
Ωi , that is, the number d such that Ωi = ∪dj=1 Ωi,j . The idea now is to represent
each obstacle Ωi,j using a 1-periodic function yj : R → Γj of the form

yj (t) := (c1j + rj (t) cos(2πt), c2j + rj (t) sin(2πt)),

where rj (t) : R → R is also a 1-periodic function. The domains that admit this
type of parametrization are usually said to be star-shaped.
The location of the obstacles and their centers (c1j , c2j ) is decided by simple
inspection. Then, we solve a least squares problem for each obstacle to find an
approximation for the corresponding function rj (t) of the form
K K
rj (t) ≈ a0 + ak cos(2πkt) + bk sin(2πkt), a0 , ak , bk ∈ R.
k=1 k=1

The choice of K depends on the shape of the obstacle. Our numerical experiments
suggest that, at low frequencies, the lowest values of the topological derivative are
reached inside a circular or elliptical like region and therefore we simply choose
K = 1. However, for higher frequencies, the lowest values of the topological
derivative are located in annular regions enclosing the boundaries. Depending
on the shape of the obstacle, one can decide that K = 1 is not enough and
increase its value. For the numerical experiments that we will present below, we
have taken in all cases K = 1 or K = 2 to compute initial guesses and values
 Topological Derivatives for Shape Reconstruction 103

between K = 1 and K = 5 for the subsequent iterations. One could also look
for 1-periodic spline functions rj (t), which are more flexible than trigonometric
polynomials, to approximate difficult points if the original boundary Γj is smooth
but not C ∞ .
In three dimensions we can proceed in a similar way, looking for a parame-
trization in polar coordinates. Spherical harmonics play then the role of trigono-
metric polynomials.
We reproduce now the tests in Sect. 3 for the same geometries with one,
two and three scatterers. After a few iterations, the description of the obstacles
improves in a significant way.
First, we apply the iterative procedure in the configuration with a single scat-
terer. Figure 10 illustrates the performance of the method at low frequencies.
Figure 10a is the equivalent of Fig. 2b but omitting the location of the observation
points. The first guess, denoted by Ω1 , is superimposed on the representation of

(c)
(a) (b) −1
1.5
1.5 −4 1.5 −2
−2
1
1 −6 1 −3
−4
−8 0.5 −4
0.5 0.5
−10
−6 Ω
Ω1 0
2 −5
0 −12 0
−6
−14 −8 −0.5
−0.5 −0.5
−16 −7
−1 −1 −10 −1
−18 −8
−20 −1.5 −9
−1.5 −1.5 −12
−22
−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5

(d) (e) (f )
−1 −0.5
1.5 1.5 −1 1.5
−1
−2
1 1 1 −1.5
−2
−3 −2
0.5 0.5 0.5
−3 −2.5
Ω3 −4
Ω4 Ω5
0 0 0 −3
−5 −4 −3.5
−0.5 −0.5 −0.5 −4
−6
−5 −4.5
−1 −7 −1 −1
−5
−8 −6
−1.5 −1.5 −1.5 −5.5

−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5

(g) (h) (i)

−0.5 −0.5
1.5 1.5 1.5 −0.5
−1 −1
1 1 1 −1
−1.5
−1.5
0.5 −2 −1.5
0.5 −2 0.5
Ω6 −2.5 Ω7 Ω8
−2.5 −2
0 0 0
−3
−3 −2.5
−0.5 −3.5 −0.5 −0.5
−3.5
−4 −3
−1 −1 −4 −1
−4.5
−3.5
−1.5 −1.5 −4.5 −1.5
−5

−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5

( j) (k) (l)
1.5 −0.5 1.5 1.5
−0.5
−0.5
1 1 1
−1 −1
0.5 0.5 0.5 −1
−1.5
Ω 10 −1.5 Ω 11
Ω9
0 0 0
−2 −1.5
−2
−0.5 −0.5 −0.5
−2.5
−2.5 −2
−1 −1 −1
−3
−1.5 −3 −2.5
−1.5 −1.5
−3.5
−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5

Fig. 10. Topological derivative and first 11 iterations. Parameter values are λe = 2.5,
λi = 0.5 and αi = 1
104 A. Carpio and M.L. Rapún

(a) (b) (c)

40 15
1.5 1.5 1.5
30 10 5
1 1 1
20 5 0
0.5 0.5
0.5 0
10
Ω2 −5
Ω1 −5 0
0 0
0
−10 −10
−0.5 −0.5
−10 −0.5
−15 −15
−1 −20 −1
−1 −20
−20
−1.5 −30 −25 −1.5
−1.5

−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5

(d) (e) (f)

6
1.5 1.5 1.5
5 4 2

1 1 2 1
0
0 0
0.5 0.5 0.5
−2
Ω3 Ω4 −2 Ω5
0 −5 0 0
−4 −4
−0.5 −0.5 −6 −0.5
−10 −6
−1 −1 −8 −1
−10 −8
−1.5 −15 −1.5 −1.5
−12
−1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5 −1.5 −1 −0.5 0 0.5 1 1.5

Fig. 11. Topological derivative and first five iterations. Parameter values are λe = 5,
λi = 1 and αi = 1

the values attained by the topological derivative when Ωi = Ω1 given in Fig. 10b.
After the first step, the magnitude of the topological derivative is almost divided
by two. In successive iterations, the order of magnitude of the updated topolog-
ical derivatives decays progressively. Their values are represented in Fig. 10c–l
for the guesses Ωi = Ωk , k = 1, . . . , 11. Ωk is the reconstruction of the obstacle
at the k-th step. Notice that the topological derivative at this step is defined
in R2 \ Ω k and Ωk appears as a solid region. The location of the obstacle is
determined in the first guess. Its approximate size and shape are satisfactorily
described after a few iterations. In 11 steps, the description of the obstacle is
almost exact.
High frequencies allow for a faster reconstruction of the obstacle in this par-
ticular geometry, as can be seen in Fig. 11. Figure 11a is the equivalent of Fig. 2c.
Notice that each approximation Ωk+1 contains the previous approximation Ωk
plus new points where the updated topological derivative falls below a threshold
−Ck , Ck > 0. A first trial value for the threshold is proposed by inspection.
We then determine Ωk+1 and update the topological derivative. In this sequence
of approximations, we observe the appearance of regions close to the boundary
where the topological derivative takes larger values (light colors). This indicates
that our current approximation may contain spurious points which do not be-
long to the obstacle. We can observe this phenomenon in Fig. 11f. In practice,
before accepting a threshold Ck , we check that the updated derivative does not
show this anomaly. Otherwise, we increase Ck .
For more complex geometries involving several scatterers, the oscillatory be-
havior of topological derivatives at high frequencies may produce patterns which
are difficult to interpret without a priori information. For this reason, we apply
the iterative procedure at low frequencies, although we know that convergence
may be slower.
 Topological Derivatives for Shape Reconstruction 105

(a) (b) (c)

2 2 –2 2 –2
1.5 −10 –4 1.5
1.5 –4
1 –6
−15 1 1 –6
Ω1 –8 Ω 2
0.5 0.5 0.5 –8
−20 –10
0 Ω1 Ω –10
0 –12 0 2
−25 –12
−0.5 –14 –0.5
–0.5
Ω1 Ω 2
−30 –16 –14
−1 –1 –1
–18 –16
−1.5 −35 –1.5 –1.5
–20
–18
−2 –2 –2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2

(d) (e) (f)

2 –2 2 2
–2
–2
1.5 –4 1.5 1.5
–4
1 1 –4
–6 1
Ω3 Ω4 Ω5
–6
0.5 –8 0.5 0.5 –6
Ω3 Ω4 Ω5
0 –10 0 –8 0 –8
–0.5 –12 –0.5 –0.5
Ω3 Ω4 –10 Ω5 –10
–1 –1 –1
–14
–12 –12
–1.5 –1.5 –1.5
–16
–2 –2 –14 –2 –14
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2

(g) (h) (i)

2 2 2
–1 –1 1
1.5 1.5 1.5
–2 –2 0
1 –3 1 1
–3 –1
Ω6 Ω7 Ω8
0.5 –4 0.5 0.5
–4 –2
Ω6 –5 0 Ω7 Ω8
0 0 –3
–5
–0.5 –6 –0.5 –0.5 –4
Ω6 Ω7 –6 Ω8
–7 –1 –5
–1 –1
–7
–8 –6
–1.5 –1.5 –1.5
–8
–9 –7
–2 –2 –2
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2

Fig. 12. Topological derivative and first eight iterations. Parameter values are λe = 2.5,
λi = 0.5 and αi = 1

The results for the geometry with two scatterers, are presented in Fig. 12.
The initial guess for the iterative scheme is computed using Fig. 4a, which seems
to indicate the presence of three obstacles, one of them misplaced. After few
iterations we find a good approximation of the real configuration. Figure 12
shows the first eight iterations. After three more iterations, the two obstacles on
the left merge and the obstacle on the right is almost recovered. As expected, the
biggest discrepancies are located in the region that is farther from the observation
points and where the influence of one obstacle on the other is stronger.
Let us now consider the geometry with three scatterers. The first guess com-
puted from Fig. 4c detects only the two largest obstacles, ignoring the small one.
For computational simplicity, if a point is included in an obstacle, it remains in-
side for subsequent iterations. In the same way, once we have created an obstacle,
we do not destroy it. It can only be merged with a neighboring one. Therefore, we
are cautious when creating a new obstacle and, although Figs. 13b,c are pointing
out the presence of the third obstacle, we have waited until the third iteration
to create it. Again, a few iterations provide a good approximation to the true
configuration.
106 A. Carpio and M.L. Rapún

(a) (b) (c)

2 2 2
–5 –4
1.5 1.5 1.5
–10 –6
1 1 –10 1 –8
Ω1 Ω2
–15
0.5 –10
0.5 0.5
Ω2 –12
0
–20 0 Ω1 –15 0
–14
–0.5 –25 –0.5 –0.5 –16
–20
–1 –1 –1 –18
–30
–20
–1.5 –1.5 –1.5
–25 –22
–35 –2
–2 –2
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2

(d) (e) (f)

2 2 2 –2
–4 1.5 1.5 –3
1.5 –4
–6 –4
1 1 –6 1
Ω3 Ω4 Ω5 –5
0.5 –8 0.5 0.5
–8 –6
Ω3 Ω4 Ω5
0 –10 0 0 –7
–10
–0.5 –12 –0.5 –0.5 –8
Ω3 Ω4 –12 Ω5 –9
–1 –14 –1 –1
–10
–14
–1.5 –16 –1.5 –1.5
–11
–2 –2 –16 –2
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2

(g) (h) (i)

2 2 2
–2 –2 –1
1.5 –3 1.5 1.5
–3
–2
1 –4 1 1
–4
Ω6 –5 Ω7 Ω8 –3
0.5 0.5 0.5
–5
Ω6 –6 Ω7 Ω8 –4
0 0 0
–7 –6
–0.5 –0.5 –0.5 –5
–8 –7
Ω6 Ω7 Ω8
–1 –9 –1 –8 –1 –6

–1.5 –10 –1.5 –1.5

–9 –7
–2 –11 –2 –2
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 –0.5 0 0.5 1 1.5 2

Fig. 13. Topological derivative and first eight iterations. Parameter values λe = 2.5,
λi = 0.5 and αi = 1

The iterative procedure implemented in this section generates an increasing

sequence of approximations Ωk ⊂ Ωk+1 (up to small deformations due to solving
a problem in a least squares sense). This restriction comes from definition (5).
If R = R2 \ Ω k , (5) defines the topological derivative only for points x ∈ R.
The guess Ωk can only be corrected by adding exterior points for which the
topological derivative falls below a negative threshold. This restriction may be
eliminated by extending the definition of topological derivative to x ∈ / R.
Notice that for the particular type of functional we consider here, we can
remove balls both in R = R2 \ Ω and in Ω. We may define the topological
derivative for x ∈ R2 as follows
Jε (R) − J (R)
DT (x, R) := lim (24)
ε→0 V(ε)
with J (R) = J(Ω) and Jε (R) defined as J(Ω), but removing the ball Bε (x). In
our case,

1
Jε (R) = Jε (Ω) := |uε − umeas |2 dl,
2 Γmeas
where uε solves (1) with
 Topological Derivatives for Shape Reconstruction 107

Ωi = Ω \ B ε (x), Ωe = R2 \ Ω i , if x ∈ Ω,
Ωi = Ω ∪ Bε (x), Ωe = R2 \ Ω i , if x ∈
/ Ω.
In this way, we might implement an iterative procedure in which Ωk+1 = {x ∈
R2 | DT (x, R2 \ Ω k ) < −Ck+1 }. At each iteration, points can be added or re-
moved, see [4] for a comparison of both strategies. This would allow to make
holes inside an obstacle, for instance.

5 Numerical Solution of Forward and Adjoint Problems

Section 4 shows that efficient procedures to solve adjoint and forward problems
are required if we want to implement topological derivative methods for inverse
scattering problems. Techniques based in series expansions, finite elements or
the method of moments are discussed in [18, 19, 37]. Inhomogeneous problems
can be handled by coupled BEM–FEM methods, see [4]. For the models with
constant parameters we are considering, integral methods are very competitive
and can be easily adapted to 3D tests. We explain here the general setting for
Helmholtz problems. The techniques can be extended to other boundary value
problems by modifying the integral equations and fundamental solutions.
When Ωi = ∅, the forward and adjoint fields that appear in the formula of the
topological derivative solve the Helmholtz transmission problem (1) and (12). We
assume that Ωi has d disjoint and connected components, that is, Ωi = ∪dj=1 Ωi,j ,
with Ω i,j ∩ Ω i,k for j = k, and denote Γj := ∂Ωi,j . The theoretical and practical
study of boundary element methods for this type of problems has attracted a lot
of attention in the past decades. Different formulations using integral equations
appear in [10, 13, 32, 55], with an emphasis on existence and uniqueness results.
For a detailed description of the use of single and double layer potentials we refer
to [10]. The recent papers [15, 45, 46, 48] deal with different boundary integral
formulations and analyze a wide class of numerical methods.
In particular, a simple quadrature scheme based on the ideas of qualocation
methods [6,16,51] that requires no implementation effort at all is studied in [15]
(see also [48]). For the sake of brevity, we just describe in detail this method. This
approach fails when either −λ +2 or −λ2 are Dirichlet eigenvalues of the Laplacian
i e
in any of the interior domains Ωi,j . In these cases, the scheme detects the prox-
imity of the eigenvalues by means of a drastic increase in the condition number.
At that point, the method must be changed. For instance, a formulation based
on Brakhage–Werner potentials (a linear combination of a single and a double
layer potential) can be used. The corresponding system of boundary integral
equations includes hypersingular operators and the development and analysis of
simple quadrature methods for these operators is still a challenging issue. We
resort to full discretization using trigonometric polynomials, following [46]. This
is the method we implement in our numerical computations to avoid possible
eigenvalues. However, it is much more involved that the one we present here. For
this reason, we will just give some remarks and references about it at the end of
this section.
108 A. Carpio and M.L. Rapún

Let us describe the simple quadrature scheme for the forward and adjoint
transmission problems in Sect. 4. First of all, we rewrite the transmission problem
in terms of the scattered and the transmitted waves, which is more suitable for
numerical purposes. The function
⎧
⎪
⎨ u − uinc , in Ωe ,
v :=
⎪
⎩ u, in Ωi = ∪dj=1 Ωi,j ,

solves a problem equivalent to (1)

⎧
⎪
⎪ 2
⎪ ∆v + λe v = 0,
⎪
in Ωe ,
⎪
⎪
⎪
⎪ +2 v = 0,
⎪
⎪ ∆v + λ in Ωi,j , j = 1, . . . , d,
⎪
⎪ i
⎨
v − − v + = uinc , on Γj , j = 1, . . . , d, (25)
⎪
⎪
⎪
⎪
⎪
⎪ αi ∂n v − − ∂n v + = ∂n uinc ,
⎪
⎪ on Γj , j = 1, . . . , d,
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂r v − ıλe v) = 0,
r→∞

with
+i := λi /√αi .
λ
The incident field appears in the transmission conditions instead of the radiation
condition at infinity. We have switched to the standard notation in the boundary
and finite element literature where n typically stands for the exterior normal
vector pointing outside Ωi . In the literature on topological derivatives n denotes
the unit normal vector pointing inside the bounded domain Ωi . We use this
notation in Sects. 2, 6 and 7, so that the formulae agree with the ones in that
literature. We only switch to the new notation in this section.
The adjoint field p solves
⎧ M
⎪
⎪
⎪
⎪ ∆p + λ 2
p = (umeas (xk ) − u(xk )) δxk , in Ωe ,
⎪
⎪
e
⎪
⎪ k=1
⎪
⎪
⎪
⎪ ∆p + λ+2 p = 0, in Ωi,j , j = 1, . . . , d,
⎪
⎨ i

⎪
⎪ p− − p+ = 0, on Γj , j = 1, . . . , d,
⎪
⎪
⎪
⎪
⎪
⎪ α ∂ p− − ∂n p+ = 0, on Γj , j = 1, . . . , d,
⎪
⎪ i n
⎪
⎪
⎪
⎩ lim r1/2 (∂ p − ıλ p) = 0.
r e
r→∞

We decompose p as p = p1 + p2 with
⎧ N
⎪
⎪
⎨ (umeas (xk ) − u(xk )) φλe (x, xk ), in Ωe ,
p1 := k=1
⎪
⎪
⎩
0, in Ωi,j , j = 1, . . . , d,
 Topological Derivatives for Shape Reconstruction 109

φλe being the fundamental solution of the Helmholtz equation (see definition
(21)). Recalling the results at the beginning of Sect. 3, the adjoint problem re-
duces to solving the following transmission problem for p2 :
⎧
⎪
⎪ 2
⎪ ∆p2 + λe p2 = 0,
⎪
in Ωe ,
⎪
⎪
⎪
⎪ +2 p2 = 0,
⎪
⎪ ∆p2 + λ in Ωi,j , j = 1, . . . , d,
⎪
⎪ i
⎨
p−2 − p2 = p1 ,
+ +
on Γj , j = 1, . . . , d, (26)
⎪
⎪
⎪
⎪
⎪
⎪ αi ∂n p−
2 − ∂ n p2 = ∂n p1 ,
+ +
⎪
⎪ on Γj , j = 1, . . . , d,
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂r p2 − ıλe p2 ) = 0.
r→∞

Notice that problems (25) and (26) are completely analogous. The only dif-
ference comes from the right hand sides. Therefore, to compute numerically the
topological derivative one just has to assemble one matrix and solve two prob-
lems with different right hand sides. We denote the right hand sides by fj and
gj . More precisely,

fj = uinc |Γj , gj = ∂n uinc |Γj , for the forward problem,

1 |Γj ,
fj = p+ 1 |Γj ,
gj = ∂n p+ for the adjoint problem.

In case data for several incident waves are available, a couple of problems of
the form (25) and (26) has to be solved for each of them. After discretization,
all of them will share the same matrix and only the source terms change.
Let us assume that yk : R → Γk is a smooth 1-periodic parametrization of
Γk . We define the functions
λ
Vkj (s, t) := φλ (yk (s), yj (t)),
λ
Jkj (s, t) := |yk (s)|∂nk (s) φλ (yk (s), yj (t)),

where nk (s) is the outward normal vector of Γk at the point yk (s). For the right
hand sides, we set

fk (s) := fk (yk (s)), gk (s) := |yk (s)|gk (yk (s)).

The method is as follows. We select a discretization parameter n ∈ N and define

the meshes

tp := p/n, tp+ε := (p + ε)/n, p = 1, . . . , n,

for 0 = ε ∈ (−1/2, 1/2). We take ε = 0 to avoid the singularity of the Hankel

function and ε = ±1/2 for stability questions. Then, we solve the following linear
problem: find
110 A. Carpio and M.L. Rapún

ϕiq,j , ϕeq,j ∈ C, q = 1, . . . , n, j = 1, . . . , d,

such that for p = 1, . . . , n and k = 1, . . . , d,

n d n
+
λi
Vkk (tp+ε , tq )ϕiq,k − λe
Vkj (tp+ε , tq )ϕeq,j
q=1 j=1 q=1

= fk (tp+ε ),
p = 1, . . . , n, k = 1, . . . , d, (27)
 ⎛ ⎞
n d n
1 i 1 + 1 1
αi ϕ + J λi (tp , tq )ϕiq,k + ⎝ ϕep,k + J λe (tp , tq )ϕeq,k ⎠
2 p,k n q=1 kk 2 n j=1 q=1 kj

1
= gk (tp ), p = 1, . . . , n, k = 1, . . . , d. (28)
n
Finally, the approximated solution to the transmission problem is computed as
⎧
⎪
⎪
d n
⎪
⎪ φλe (z, yj (tq ))ϕeq,j , z ∈ Ωe ,
⎪
⎨
ε j=1 q=1
un (z) := (29)
⎪
⎪
n
⎪
⎪ i
φλ+ i (z, yj (tq ))ϕq,j , z ∈ Ωi,j , j = 1, . . . , d.
⎪
⎩
q=1

This method is a discrete version of a boundary integral formulation of the

original Helmholtz transmission problem. We look for a function of the form
⎧ d
⎪
⎪
⎨ Sjλe ϕej , in Ωe ,
u= j=1 (30)
⎪
⎪
⎩ +
Sjλi ϕij in Ωi,j , j = 1, . . . , d,

where the single layer potentials Sjλ are defined by

1
Sjλ ϕ := φλ ( · , yj (t))ϕ(t)dt, (31)
0

and the densities ϕej , ϕij have to be determined. These densities are approximated
then by Dirac delta distributions on the points yj (tq ). The first set of equations
(27) is obtained when testing the transmission conditions

u− − u+ = fk , on Γk , k = 1, . . . , d,

at the observation points yk (tp+ε ), (or equivalently, we are testing the equations
with Dirac delta distributions on the points yk (tp+ε )). The second set of equa-
tions (28) is equivalent to the application of the classical Nyström method for
the second transmission conditions
 Topological Derivatives for Shape Reconstruction 111

α∂n u− − ∂n u+ = gk , on Γk , k = 1, . . . , d,

which in terms of the densities are integral equations of the second kind. For
further details we refer to [15].
For any 0 = ε ∈ (−1/2, 1/2), we have a first order method. For the particular
choices ε = ±1/6, the method has quadratic convergence order, that is,

|u(z) − u±1/6
n (z)| ≤ Cz (1/n)2

where Cz > 0 is a constant that depends on z but not on n. Moreover, if we

keep the n × d equations (28) unchanged and we replace
λ
Vkj (tp+ε , tq ), +i )
(for λ = λe and λ = λ

in (27) by the linear combination

5 λ λ
 1 λ λ

Vkj (tp−1/6 , tq ) + Vkj (tp+1/6 , tq ) + Vkj (tp−5/6 , tq ) + Vkj (tp+5/6 , tq )
6 6
as well as the value fk (tp+ε ) at the right hand side by
5  1 
fk (tp−1/6 ) + fk (tp+1/6 ) + fk (tp−5/6 ) + fk (tp+5/6 )
6 6
we obtain a new linear system of equations to compute the unknown values
ϕiq,j , ϕeq,j ∈ C, q = 1, . . . , n, j = 1, . . . , d. If we define the discrete solution uεn as
in (29), then the method has cubic convergence order,

|u(z) − uεn (z)| ≤ Cz (1/n)3 .

This strategy was found using the ideas of qualocation methods, a variation of
collocation methods that improves the order of convergence (see [6, 16, 51]).
This method is specially well suited for our purposes because moving the
boundary requires almost no additional computational effort in recomputing the
elements in the matrix of the linear system of equations. As we have already
mentioned, this technique cannot be applied in the singular cases associated
with interior Dirichlet eigenvalues of the Laplace operator. To overcome this
problem, one can represent the solution of the transmission problem as
⎧ d
⎪
⎪
⎨ (Sjλe − ıηDjλe )ϕej , in Ωe ,
u := j=1 (32)
⎪
⎪
⎩ (S λi − ıηDλi )ϕi ,
+ +
in Ω , j = 1, . . . , d,
j j j i,j

where η > 0 is a fixed parameter, Sjλ are the single layer potentials that were
introduced in (31) and the double layer potentials Djλ are defined by

1
Djλ := |yj (t)| ∂nj (t) φλ ( · , yj (t))ϕ(t)dt.
0
112 A. Carpio and M.L. Rapún

Linear combinations of single and double layer potentials of the form S − ıηD
with η > 0 are usually called mixed or Brakhage–Werner potentials. The sys-
tem of integral equations found by imposing the transmission conditions at the
interfaces Γk is more complicated than the system obtained for the simpler rep-
resentation (30). Now, we deal with hypersingular operators. The equivalence of
this formulation to the original problem was shown in [46]. We also refer to that
work for a complete analysis of convergence of a wide class of Petrov–Galerkin
methods. In the numerical tests presented in Sect. 4, we used a fully discrete
Galerkin method with trigonometric polynomials that has superalgebraic con-
vergence order in terms of the discretization parameter, avoiding therefore the
occurrence of eigenvalues.
In 3D both integral representations (30) and (32) remain valid using the
corresponding fundamental solution in 3D given by (21). The resultant systems
of integral equations can be solved by using Petrov–Galerkin methods, as indi-
cated in [45,46]. For the implementation of fully discrete Galerkin methods with
trigonometric polynomials we refer to [40]. Discretizations of Petrov–Galerkin
methods for spline functions are studied in [45].

6 Explicit Expressions for the Topological Derivatives

Different strategies for the practical computation of topological derivatives are

found in the literature. For instance, in [18], a useful connection with the shape
derivative is given. Integral expressions in terms of Green functions of Helmholtz
equations are exploited in [24]. Our procedure to obtain analytical expressions
for topological derivatives combines both of them. First, we calculate the shape
derivative of the functionals involved. Then, we perform asymptotic expansions
of the solutions of Helmholtz equations with vanishing holes. Such expansions
are performed directly on the partial differential equation formulation.

6.1 Shape Derivative

The shape derivative of the functional J (R) = J(Ωi ) in the direction of a smooth
vector field V(x) is defined as
$
d $
DJ(Ωi ) · V := J(φτ (Ωi ))$$ ,
dτ τ =0

where φτ (x) := x + τ V(x), x ∈ R2 and V represents the direction of change of

the domain.
For practical computations, a more precise formula in terms of solutions of
two auxiliary direct and adjoint problems is used. When V satisfies V = Vn n
on ∂Ωi and V = 0 out of a small neighborhood of ∂Ωi , we have the following
result.
 Topological Derivatives for Shape Reconstruction 113

Theorem 6.1. The shape derivative of the functional J defined in (3) is given by


DJ(Ωi ) · V = Re (−αi ∇u− ∇p− + λ2i u− p− + 2αi ∂n u− ∂n p− )Vn dl
Γ


− (−∇u+ ∇p+ + λ2e u+ p+ + 2∂n u+ ∂n p+ )Vn dl , (33)
Γ

where Vn = V·n, u is the solution of the forward problem and p is the continuous
solution of an adjoint problem. For particular choices of the boundary conditions
at the interface between the medium and the obstacles, (33) takes the following
simpler forms:
• General transmission problem (penetrable obstacle):


 
DJ(Ωi ) · V = Re αi (1 − αi )∂n u− ∂n p− + (1 − αi )∂t u− ∂t p− Vn dl
Γ


+ (λ2i − λ2e ) u p Vn dl . (34)
Γ

where u solves (1) and p solves (12).

• Transmission problem with αi = 1


DJ(Ωi ) · V = Re (λ2i − λ2e ) u p Vn dl , (35)
Γ

where u and p solve the forward and adjoint transmission problems (1) and
(12) with αi = 1.
• Neumann problem (sound hard obstacle):


DJ(Ωi ) · V = Re (∇u∇p − λ2e up)Vn dl , (36)
Γ

where u and p solve the exterior homogeneous Neumann problems (2)

and (15).

This result is independent of the value of V outside Γ . Notice that (33) is

rewritten in terms only of the interior values of the solutions of the forward and
adjoint problems by using the transmission boundary conditions and splitting
the gradients in tangential and normal components:

∇u = (∇u · t)t + (∇u · n)n = ∂t u t + ∂n u n, on Γ .

The continuity of u across Γ implies the continuity of the tangential derivatives:

∂t u− = ∂t u+ . When αi = 1, the normal derivatives are also continuous and the
continuity of the gradients at the interface follows.
114 A. Carpio and M.L. Rapún

Formula (33) would extend to other boundary conditions by simply changing

the conditions at the interface between the obstacle and the surrounding medium
in the forward and adjoint problems.
Here, we only prove (33) for the generalized transmission problem. Related re-
sults are found in [4] for transmission problems in heterogeneous media in [17, 37]
for the transmission problem with αi and in [19] for the Neumann problem.
Proof. We adapt the strategy introduced in [19] to derive (36) for the Neu-
mann problem.
The idea is the following. First, we give a variational formulation of the
boundary value problem in a bounded domain using transparent boundary con-
ditions. Next, we deform the domain Ωi along the vector field V and compute
the shape functional in the deformed domains. Then, we differentiate the trans-
formed functionals with respect to the control parameter τ . The resulting ex-
pression involves the derivative of the solutions uτ in the deformed domains.
The computation of this derivative is avoided establishing a relationship with
the derivatives of modified functionals, in which a free state can be selected to
eliminate dudτ . This adjoint state solves the so-called adjoint problem. Integrating
τ

by parts the resulting expression involving adjoint solutions, we find the desired
formula for the shape derivative. The proof is organized in four steps.
Step 1: Variational formulation of the forward problem. First, we replace
the exterior problem (1) by an equivalent boundary value problem posed in
a bounded domain. Let us introduce a circle ΓR which encloses the obstacles
as in Fig. 1. The Dirichlet-to-Neumann (also called Steklov–Poincaré) operator
associates to any Dirichlet data on ΓR the normal derivative of the solution of
the exterior Dirichlet problem:

L : H 1/2 (ΓR ) −→ H −1/2 (ΓR )

f −→ ∂n w

where w ∈ Hloc
1
(R2 \ B R ), BR := B(0, R), is the unique solution of
⎧
⎪
⎪ ∆w + λ2e w = 0, in R2 \ B R ,
⎪
⎪
⎨
w = f, on ΓR ,
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂ w − ıλ w) = 0.
r e
r→∞

1
Hloc (R2 \B R ) denotes the usual Sobolev space and H 1/2 (ΓR ) and H −1/2 (ΓR ) are
the standard trace spaces. A detailed representation of the Dirichlet-to-Neumann
map in terms of Hankel functions is given in [19, 28]. Instead of the exterior
transmission problem (1) one can study an equivalent boundary value problem
in BR with a non-reflecting boundary condition on ΓR :
 Topological Derivatives for Shape Reconstruction 115
⎧
⎪
⎪ 2
in Ωe := BR \ Ω i ,
⎪ ∆u + λe u = 0,
⎪
⎪
⎪
⎪
⎪
⎪
⎪ αi ∆u + λ2i u = 0, in Ωi ,
⎪
⎪
⎨
u− − u+ = 0, on Γ , (37)
⎪
⎪
⎪
⎪
⎪
⎪ αi ∂n u− − ∂n u+ = 0,
⎪
⎪ on Γ ,
⎪
⎪
⎪
⎪
⎩ ∂n (u − uinc ) = L(u − uinc ), on ΓR .

The solution u of (37) also solves the variational equation

⎧
⎪
⎨ u ∈ H 1 (BR ),
(38)
⎪
⎩ b(Ωi ; u, v) = (v), ∀v ∈ H 1 (BR ),

where



b(Ωi ; u, v) := (∇u∇v − λ2e uv)dz + (αi ∇u∇v − λ2i uv)dz
Ωe Ωi


− Lu v dl, ∀u, v ∈ H 1 (BR ),
ΓR


(v) := (∂n uinc − Luinc ) v dl, ∀v ∈ H 1 (BR ).
ΓR

Step 2: Transformed functionals. Since V decreases rapidly to zero away from

Γ , we have φτ (ΓR ) = ΓR and φτ (Γmeas ) = Γmeas . Set

Ωi,τ := φτ (Ωi ), Ωe,τ := φτ (Ωe ) = BR \ Ω i,τ .

The variational reformulations (38) of the original Helmholtz problem in the

deformed domains take the form
⎧
⎪
⎨ uτ ∈ H 1 (BR ),
(39)
⎪
⎩ b(Ω ; u , v) = (v),
i,τ τ ∀v ∈ H (BR ),
1

with



b(Ωi,τ ; u, v) := (∇zτ u∇zτ v − λ2e uv)dzτ + (αi ∇zτ u∇zτ v − λ2i uv)dzτ

Ωe,τ Ωi,τ


− Lu vdl, ∀u, v ∈ H 1 (BR ).
ΓR

The cost functional in the transformed domains is

116 A. Carpio and M.L. Rapún


1
J(Ωi,τ ) = |uτ − umeas |2 dl, (40)
2 Γmeas

where uτ solves (39). Solving (39) with τ = 0 we recover the solution u = u0

of (38).
Step 3: Adjoint states. Differentiating (40) with respect to τ we obtain
$

d $
J(Ωi,τ )$$ = Re (u − umeas )u̇ dl , (41)
dτ τ =0 Γmeas

d
where u̇ = dτ uτ |τ =0 . To avoid computing u̇ we introduce an adjoint problem.
Let us define a family of modified functionals

L(Ωi,τ ; v, p) = J(Ωi,τ ) + Re[b(Ωi,τ ; v, p) − (p)], ∀v, p ∈ H 1 (BR ),

where p plays the role of a Lagrangian multiplier. Setting v = uτ , we obtain

L(Ωi,τ ; uτ , p) = J(Ωi,τ ), ∀p ∈ H 1 (BR ). (42)

The shape derivative of J in the direction V is then

$ $
d $ d $
DJ(Ωi ) · V = $
L(Ωi,τ ; uτ , p)$ = Re b(Ωi,τ ; u, p)$$
dτ τ =0 dτ τ =0


+Re b(Ωi ; u̇, p) + Re (u − umeas )u̇ dl , (43)
Γmeas

thanks to (41) and (42). This is true for any p ∈ H 1 (ΩR ). If p solves
⎧
⎪
⎨ p ∈ H (BR ),
1

(44)
⎪
⎩ b(Ωi ; v, p) = (umeas − u) v dl, ∀v ∈ H 1 (BR ),
Γmeas

the derivative u̇ is not needed and the shape derivative DJ(Ω) · V is given by
the first term in (43). We select the solution p of (44) as an adjoint state. The
adjoint problem (44) is equivalent to
⎧
⎪
⎨ p ∈ H (BR ),
1

(45)
⎪
⎩ b(Ωi ; p, v) = (umeas − u) v dl, ∀v ∈ H 1 (BR ),
Γmeas

which has the same structure as the direct problem (38) and p is then a solution
of (12). A property of the Dirichlet-to-Neumann operator is essential in this
argument: its adjoint operator satisfies L∗ (u) = L(u).
Step 4: Shape derivative. Let us take p to be the adjoint state and compute
the first term in (43). The change of variables from the deformed to the original
variables is governed by the identities
 Topological Derivatives for Shape Reconstruction 117

∇zτ u = Fτ− ∇u, dzτ = det Fτ dz,

and
$ $ $
d $ d $ d $
Fτ $$ = ∇V, (∇zτ u)$$ = −∇V ∇u, dzτ $$ = ∇ · Vdz,
dτ τ =0 dτ τ =0 dτ τ =0

where Fτ := ∇φτ = I + τ ∇V is the deformation gradient. For u, p ∈ H 1 (BR ),

$
d $
b(Ωi,τ ; u, p)$$ =
dτ τ =0



 
(∇u · ∇p − λ2e up) ∇ · V dz − (∇V + ∇V )∇u · ∇p dz +
Ωe Ωe



 
(αi ∇u · ∇p − λ2i up) ∇ · V dz − αi (∇V + ∇V )∇u · ∇p dz. (46)
Ωi Ωi

We may further simplify this expression integrating by parts:

(αi ∇u · ∇p − λ2i up) ∇ · V dz =
Ωi



λ2i u− p− V · n dl − αi (∇u− · ∇p− )V · n dl +
Γ

  Γ
 2 
∂u ∂p ∂ u ∂p ∂ 2 p ∂u
λ2i p+u v dz − αi v + v dz (47)
Ωi ∂x ∂x Ωi ∂xj ∂x ∂xj ∂xj ∂x ∂xj

and


 
− (∇V + ∇V )∇u · ∇p dz =
Ωi


 
∂ 2 u ∂p ∂ 2 p ∂u ∂ 2 u ∂p ∂ 2 p ∂u
2 v + v dz + v + v dz
Ωi ∂xj ∂x ∂x2j ∂x Ωi ∂xj ∂x ∂xj ∂xj ∂x ∂xj

 
∂u− ∂p− ∂p− ∂u−
+ v nj + v nj dz, (48)
Γ ∂xj ∂x ∂xj ∂x

where V = (v1 , v2 ) , n = (n1 , n2 ) and summation over repeated indexes is

understood. Recall that n points inside Ωi .
Let us now add the contributions from (47) and from (48) multiplied by αi .
Part of the integrals over Ωi cancel and the rest vanishes



(αi ∆u + λi u) ∇p · V dz +
2
(αi ∆p + λ2i p) ∇u · V dz = 0
Ωi Ωi

because u and p satisfy Helmholtz equations in Ωi . The sum of the integrals over
Γ is
118 A. Carpio and M.L. Rapún




−αi (∇u− · ∇p− )Vn dl + λ2i u− p− Vn dl + 2αi (∇u− · n)(∇p− · n)Vn dl.
Γ Γ Γ

using V = Vn n.
We reproduce the same computations for Ωe . All the integrals over ΓR are
identically zero, since V = 0 on ΓR . Recall that n is now the outward normal
vector. Integrating by parts we obtain



 
(∇u · ∇p − λ2e up) ∇ · V dz − (∇V + ∇V )∇u · ∇p dz
Ωe Ωe



= (∆u + λ2e u)∇p · V dz + (∆p + λ2e p)∇u · V dz +
Ωe Ωe




(∇u ∇p )Vn dl −
+ +
λ2e u p Vn dl − 2
+ +
(∇u+ · n)(∇p+ · n)Vn dl. (49)
Γ Γ Γ

Notice that both u and p solve the Helmholtz equations

∆u + λ2e u = 0, ∆p + λ2e p = (umeas − u) δΓmeas , in Ωe .

Due to these identities and the fact that V = 0 on Γmeas , the two integrals over
Ωe vanish.
Adding all the integrals on Γ generated integrating by parts on Ωi and Ωe ,
we finally find (33). 2

6.2 Proof of Theorems 2.1 and 2.2

Theorem 2.1.- Transmission problem without holes. Let us first calculate the
topological derivative of the cost functional (3) for the transmission problem
when R = R2 . Knowing the formula of the shape derivative for the transmission
problem, we use the identities (6) and (34) to obtain


−1
DT (x, R2 ) = lim  Re αi (1 − αi )∂n u− −
ε ∂n pε (50)
ε→0 V (ε) Γε


− −
+ (1 − αi )∂t uε ∂t pε dl + (λ2i − λ2e )uε pε dl .
Γε

Recall that Vn is constant and negative. Here, uε and pε solve the forward and
adjoint problems when Ωi = Bε (x) and Γ = Γε = ∂Bε (x). A value for this limit
is computed performing an asymptotic expansion of these solutions and their
gradients at Γε as in [4].
The asymptotic behavior of uε and pε is obtained expressing them as correc-
tions of u and p

uε (z) = u(z)χR2 \B ε (z) + vε (z), pε (z) = p(z)χR2 \B ε (z) + qε (z),

and expanding the remainders in powers of ε.

 Topological Derivatives for Shape Reconstruction 119

Let us denote by B the unit ball. Changing variables ξ := (z − x)/ε, the

correction vε (ξ) satisfies
⎧
⎪
⎪ ∆ξ vε + ε2 λ2e vε = 0, in R2 \ B,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ αi ∆ξ vε + ε2 λ2i vε = 0, in B,
⎪
⎪
⎪
⎨
vε− − vε+ = u(z) = u(x) + ε ξ · ∇u(x) + O(ε2 ), on Γ , (51)
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ αi n(ξ) · ∇ξ vε− − n(ξ) · ∇ξ vε+ = εn(ξ) · ∇u(x) + O(ε2 ), on Γ ,
⎪
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂ v − ıελ v ) = 0,
r ε e ε r = |ξ|.
r→∞

Let us expand now vε (ξ) in powers of ε : vε (ξ) = v (1) (ξ) + εv (2) (ξ) + O(ε2 ).
The leading terms of the expansion solve:
⎧
⎪
⎪ ∆ξ v (1) = 0, in R2 \ B and B,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨ v (1)− − v (1)+ = u(x), on Γ ,
(52)
⎪
⎪
⎪ αi n(ξ) · ∇ξ v
⎪
(1)−
− n(ξ) · ∇ξ v (1)+
= 0, on Γ ,
⎪
⎪
⎪
⎪
⎪
⎩ lim r1/2 ∂r v (1) = 0,
r→∞

and ⎧
⎪
⎪ ∆ξ v (2) = 0, in R2 \ B and B,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨ v (2)− − v (2)+ = ξ · ∇u(x), on Γ ,
(53)
⎪
⎪
⎪
⎪ α n(ξ) · ∇ξ v (2)− − n(ξ) · ∇ξ v (2)+ = n(ξ) · ∇u(x), on Γ ,
⎪ i
⎪
⎪
⎪
⎪
⎩  
limr→∞ r1/2 ∂r v (2) − ıλe v (1) = 0.

By inspection, v (1) (ξ) = u(x)χB (ξ). The second term, v (2) (ξ), can be found
working in polar coordinates:
 
2 1 − αi ξ
v (ξ) = ∇u(x) ·
(2)
ξ χB (ξ) + χ 2 (ξ) = ∇u(x) · g(ξ).
1 + αi 1 + αi |ξ|2 R \B

Thus,

uε (z) = u(z)χR2 \B ε (z) + u(x)χBε (z) + ε∇u(x) · g(ξ) + O(ε2 ), z ∈ Γε .

Performing a similar expansion for pε (z), we get

pε (z) = p(z)χR2 \B ε (z) + p(x)χBε (z) + ε∇p(x) · g(ξ) + O(ε2 ), z ∈ Γε .

120 A. Carpio and M.L. Rapún

This implies that

uε (z) → u(x), pε (z) → p(x), as ε → 0,

uniformly when |z − x| = ε. Expanding the derivatives we find

∂u−
ε ∂v (2)− ∂v (2)− 2 ∂u
(z) = ε (ξ) + O(ε) = (ξ) + O(ε) = (x) + O(ε),
∂zj ∂zj ∂ξj 1 + αi ∂zj
and a similar identity for pε . Therefore,
2 2
∇u−
ε (z) → ∇u(x), ∇p−
ε (z) → ∇p(x), as ε → 0.
1 + αi 1 + αi
Let us take limits in the three integrals appearing in (50). For the integral
of the normal components, we write the unit normal as a function of the angle
(n1 , n2 ) = (cos θ, sin θ). Then,


2π
4 ∂uε ∂pε
(∇u−
ε · n)(∇p −
ε · n)dl z ≈ ε (x) (x) nj n dθ
Γε (1 + αi )2 ∂zj ∂z 0

4π
= ε ∇u(x)∇p(x),
(1 + αi )2
( 2π
plus a remainder of higher order in ε, where we have used that 0 n1 n2 dθ =
( 2π
0 and 0 n2j dθ = π for j = 1, 2. To compute the integral of the tangential
components, we write the tangent vector as (t1 , t2 ) = (− sin θ, cos θ). Exactly
the same value is found. Finally, the third integral is


uε pε dlz ≈ 2επ u(x)p(x).
Γε

Taking into account that V  (ε) = −2πε, (7) follows. Formula (8) is a particular
case with αi = 1.
Theorem 2.1.- Neumann problem without holes. The analytical expression of
the topological derivative for Neumann problems (9) follows by combining (6)
and (36). Then, we perform an asymptotic expansion of uε and pε , which now
solve Neumann problems. We find that
ξ
uε (z) = u(z) + ε∇u(x) · + O(ε2 ).
|ξ|2

Thus, uε (z) = u(x) + O(ε) and

∂uε ∂u ∂ ξ
(z) = (x) + ∇u(x) · + O(ε)
∂zj ∂zj ∂ξj |ξ|2

as ε → 0, uniformly when |z − x| = ε. Similar expressions hold for pε and ∇pε .

Computing the limit, we obtain (9).
 Topological Derivatives for Shape Reconstruction 121

Theorem 2.2.- Transmission problem with holes. Let us calculate now the
topological derivative of the cost functional (3) in a domain with a hole R =
R2 \ Ω, Ω being an open bounded set, not necessarily connected. Formula (11)
follows by slightly modifying the procedure we have used to compute (7) in R2
for the transmission problem.
Now, uε (z) = u(z)χR2 \B (z) + vε (z), where u solves (1) with Ωi = Ω. Chang-
ing variables, vε (ξ) satisfies
⎧
⎪
⎪ ∆ξ vε + ε2 λ2e vε = 0, in R2 \ (B ∪ Ω ε ),
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ 2 2
in B ∪ Ωε ,
⎪ αi ∆ξ vε + ε λi vε = 0,
⎪
⎪
⎪
⎪
⎪
⎪
⎪ vε− − vε+ = u(z) = u(x) + ε ξ · ∇u(x) + O(ε2 ), on Γ ,
⎪
⎪
⎪
⎨
αi n(ξ) · ∇ξ vε− − n(ξ) · ∇ξ vε+ = εn(ξ) · ∇u(x) + O(ε2 ), on Γ ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪ vε− − vε+ = 0,
⎪ on ∂Ωε ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ αi n(ξ) · ∇ξ vε− − n(ξ) · ∇ξ vε+ = 0, on ∂Ωε ,
⎪
⎪
⎪
⎪
⎪
⎩ lim r1/2 (∂ v − ıελ v ) = 0,
r ε e ε r = |ξ|,
r→∞
(54)
with Ωε := (Ω − x)/ε.
Expanding vε (ξ) in powers of ε, the leading terms v (1) (ξ), v (2) (ξ) solve again
(52) and (53), respectively, and (11) follows. We just have to check that the
presence of Ωε does not provide corrections to the orders zero and one. Let us
consider the boundary value problems:
⎧
⎪
⎪ ∆ξ v (1) = 0, in R2 \ (B ∪ Ω ε ) and B ∪ Ωε ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ (1)−
− v (1)+ = u(x),
⎪v
⎪
on Γ ,
⎪
⎪
⎪
⎪
⎪
⎨ αi n(ξ) · ∇ξ v (1)− − n(ξ) · ∇ξ v (1)+ = 0, on Γ ,
(55)
⎪
⎪
⎪
⎪ v (1)−
− v (1)+
= 0, on ∂Ω ,
⎪
⎪
ε
⎪
⎪
⎪
⎪
⎪
⎪ αi n(ξ) · ∇ξ v
(1)−
− n(ξ) · ∇ξ v (1)+ = 0, on ∂Ωε ,
⎪
⎪
⎪
⎪
⎪
⎩ lim r1/2 ∂ v (1) = 0,
r
r→∞

and
122 A. Carpio and M.L. Rapún
⎧
⎪
⎪ ∆ξ v (2) = 0, in R2 \ (B ∪ Ω ε ) and B ∪ Ωε ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ v (2)− − v (2)+ = ξ · ∇u(x), on Γ ,
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨ αi n(ξ) · ∇ξ v (2)− − n(ξ) · ∇ξ v (2)+ = n(ξ) · ∇u(x), on Γ ,
(56)
⎪ (2)−
⎪
⎪
⎪ v − v (2)+
= 0, on ∂Ω ,
⎪
⎪
ε
⎪
⎪
⎪
⎪
⎪
⎪ αi n(ξ) · ∇ξ vε
(2)−
− n(ξ) · ∇ξ v (2)+ = 0, on ∂Ωε ,
⎪
⎪
⎪
⎪
⎪
⎩ lim  (2) 
r→∞ r
1/2
∂r v − ıλe v (1) = 0.

Then, v (1) (ξ) = u(x)χB (ξ) is still a solution of (55). Since αi = 0,

 
2 1 − αi ξ
v (ξ) = ∇u(x) ·
(2)
ξ χB (ξ) + χ 2 (ξ) ,
1 + αi 1 + αi |ξ|2 R \B

solves (56) with an error of order ε2 . Thus, the correction coming from Ωε
appears at orders higher than ε.
A similar argument works for the Neumann problem.

7 Sounding Solids by Elastic Waves

The ideas and techniques developed in the previous sections for scattering prob-
lems involving waves governed by Helmholtz equations can be extended to more
general situations. In this section, we consider the problem of detecting solids
buried in an elastic medium by means of acoustic waves. We only describe the
procedure to compute explicit formulae for the topological derivatives in this new
setting. Once these formulae are known, the numerical approximation procedures
described in Sects. 3–5 apply using the fundamental solutions and integral oper-
ators of elasticity theory.

7.1 The Forward and Inverse Scattering Problems

The general setting is similar to that in Sect. 2. The obstacle Ωi ⊂ R2 is an open

bounded set with smooth boundary Γ := ∂Ωi but has no assumed connectivity.
The scattering pattern is measured at Γmeas far enough from the scatterers, as
in Fig. 1.
Consider the 2D elastodynamic problem in a solid, where the elastic fields
depend only on the coordinates z1 , z2 and time. The Navier equations of motion
are given by
∂ 2 Uj ∂ 2 Um
ρ 2 − cjαmβ = 0,
∂t ∂zα ∂zβ
 Topological Derivatives for Shape Reconstruction 123

where cjαmβ are the stiffness tensor components, ρ is the mass density and z
and t are the 2D position vector and time, respectively. Subscripts range from 1
to 2. We adopt the convention that a repeated index is summed over its range.
Regardless of particular material symmetries, the elastic moduli satisfy the
following symmetry restrictions:

cjαmβ = cmβjα = cαjβm ,

together with the coercivity inequality:

2
ξjα cjαmβ ξ mβ ≥ C |ξjα |2
j,α=1

for some constant C > 0 and every complex ξjα such that ξjα = ξαj . The elastic
constants cjαmβ for a cubic crystal are:

cjαmβ = c12 δjα δmβ + c44 (δjm δαβ + δjβ δαm ) − Hδjαmβ (57)

where H is the anisotropy factor H = 2c44 + c12 − c11 . δjm stands for the
Kronecker delta. In the isotropic case, c12 = λ, c44 = µ and c11 = λ + 2µ.
We are interested in time harmonic solutions, in which the elastic displace-
ment field can be written in the form Uj (z, t) = Re [uj (z)e−ıωt ] for a given
frequency ω > 0. Therefore, the components uj , j = 1, 2 solve

∂ 2 um
ρω 2 uj + cjαmβ = 0. (58)
∂zα ∂zβ

In the scattering process, a time harmonic vector plane wave of unit ampli-
tude uinc illuminates an object with section Ωi embedded in an infinite elastic
medium. This process generates two additional time harmonic fields: the scat-
tered vector field usc , defined outside the obstacle and propagating outwards,
together with the transmitted vector field utr , defined inside the obstacle and
trapped in it.
Dropping the time harmonic dependence, which is implicit in (58), the inci-
dent waves have the form uinc (z) = u0 eık·z , with z = (z1 , z2 ) and k = (k1 , k2 ) =
kd, where k is the wave number and the unitary vector d is the direction of
propagation. The wave vector k, the frequency ω and u0 are related by:

(ρω 2 δjm − cjαmβ kα kβ )u0,m = 0, j = 1, 2. (59)

A non-zero solution exists only if the determinant of this matrix is zero:

%  &
det ρω 2 δjm − cjαmβ kα kβ jm = 0. (60)

Note that the frequency ω depends on the wave vector k through the dispersion
relation (60). Substituting each of its two roots ω 2 = ωj2 (k) in system (59), we
find the directions of displacement u0 in these waves (directions of polarization).
124 A. Carpio and M.L. Rapún

The equations are homogeneous, they do not determine their magnitude. We

normalize them by imposing |u0 | = 1. We may find two different directions
of polarization for the same wave vector, which are perpendicular to each other
since the matrix tensor (cjαmβ kα kβ )jm is symmetrical. They are eigenvectors for
different eigenvalues. Neither of these directions is in general purely longitudinal
or purely transverse to k.
For isotropic materials the dependence of ω on k simplifies to direct
' propor-
tionality to its magnitude k. The branch ω = cp kp , with cp = (λ + 2µ)/ρ,
corresponds
' to longitudinally polarized waves. The branch ω = cs ks , with
cs = µ/ρ, corresponds to transversely polarized waves. In this case, incident
waves take the form uinc (z) = u0 eıkd·z where the wave number k is either kp or
ks and the direction of polarization u0 is either d or is orthogonal to d (obtained
by rotating d anticlockwise π/2).
The interaction between the scatterer, the medium and the incident radiation
is described by the following transmission model where the total wavefield
⎧
⎨ u + u , in Ω := R2 \ Ω ,
inc sc e i
u=
⎩u , in Ω ,
tr i

satisfies:
⎧
⎪ ∂ 2 um
⎪
⎪ ρi 2
ω u + ci
= 0, j = 1, 2, in Ωi ,
⎪
⎪
j jαmβ
∂zα ∂zβ
⎪
⎪
⎪
⎪
⎪
⎪ ∂ 2 um
⎪
⎨ ρe ω 2 uj + cejαmβ = 0, j = 1, 2, in Ωe ,
∂zα ∂zβ
(61)
⎪
⎪
⎪ u− − u+ = 0,
⎪
⎪
⎪ on Γ := ∂Ωi ,
⎪
⎪
⎪
⎪
⎪ i
⎪ ∂u− ∂u+
⎩ cjαmβ m nα − cejαmβ m nα = 0, j = 1, 2, on Γ ,
∂zβ ∂zβ

supplemented by radiation conditions at infinity on the scattered field usc =

u − uinc implying that only outgoing waves are allowed. The dependence of
densities and elastic constants on the domain is indicated by the superscripts i
and e. We have kept the notation introduced in Sect. 2: n is the outward unit
normal vector pointing outside Ωe , and u+ and u− denote the limits of u from
the exterior and interior of Ωi .
The radiation condition is a decay condition at infinity which selects the
correct fundamental solution and implies uniqueness. For isotropic materials,
any solution of (58) can be decomposed as u = up + us into a compressional
part up (or longitudinal wave, satisfying rot(up ) = 0) and a shear part us (or
transverse wave, satisfying div(us ) = 0). They satisfy the vector Helmholtz
equation with wave numbers kp and ks , respectively. The Kupradze radiation
condition for isotropic materials reads:

lim r1/2 (∂r (uc − ucinc ) − ıkc (uc − ucinc )) = 0, c = p, s r = |x|. (62)
r→∞
 Topological Derivatives for Shape Reconstruction 125

Solutions satisfying this condition are called radiating solutions of the isotropic
2D Navier equations, see the reviews [22, 35] and references therein. Radiation
conditions for anisotropic materials are discussed in [22].
For rigid obstacles, there is no transmitted field. Then, the transmission
problem is replaced by a Neumann problem:
⎧
⎪ ∂ 2 um
⎪
⎪
e 2 e
⎨ ρ ω uj + cjαmβ ∂zα ∂zβ = 0, j = 1, 2, in Ωe ,
(63)
⎪
⎪ ∂um
⎪ e
⎩ cjαmβ nα = 0, j = 1, 2, on Γ ,
∂zβ

with a ‘sound hard’ boundary condition at the obstacle boundary plus the radi-
ation condition on u − uinc at infinity. For simplicity, in this section we will work
only with the Neumann problem and we will adopt the simpler notation ρ ≡ ρe
and cjαmβ ≡ cejαmβ . Similar, but more involved computations can be carried out
for the transmission problem (see [5]).
The forward problem consists in computing the solution u of (63) at the
measurement curve Γmeas knowing the scatterers Ωi and the incident wave uinc .
Existence and uniqueness results for problems (61) and (63) can be found in
[1, 7, 20, 25, 35, 42, 43].
As before, the inverse problem consists in finding the shape and structure
of the obstacle Ωi such that the solution of the forward Neumann problem (63)
equals the measured values umeas at the receptors, knowing the incident field.
A less restrictive formulation is given by the following constrained optimization
problem: minimize

1
J(Ωi ) := |u − umeas |2 (64)
2 Γmeas
where u is the solution of the forward Neumann problem (63) and umeas the
total field measured on Γmeas .

7.2 Shape Derivatives

The computation of the topological derivative of the shape functional (64) follows
the same steps as in Sect. 6. Let us first compute the shape derivative of (64).
Then, we will use its expression to find an explicit formula for the topological
derivative.

Theorem 7.1. Keeping the notations of the previous sections, the shape deriv-
ative of the functional J defined in (64) is given by

 
∂uj ∂pm
DJ(Ω) · V = Re cjαmβ − ρω 2 up Vn dl , (65)
Γ ∂zα ∂zβ

where p is a continuous solution of the adjoint problem

126 A. Carpio and M.L. Rapún
⎧
⎪
⎪ ∂ 2 pm
⎪
⎨ ρω 2
p j + cjαmβ = (umeas − u)j δΓmeas , j = 1, 2, in Ωe ,
∂z ∂z
α β
(66)
⎪
⎪ ∂pm
⎪
⎩ cjαmβ nα = 0, j = 1, 2, on Γ ,
∂zβ
together with a conjugated radiation condition at infinity for p.
Proof. As in the Helmholtz case, we prove the result in four steps, following
the ideas in [5].
Step 1: Variational formulation of the forward problem in a bounded do-
main. First, we replace the unbounded Neumann problem by an equivalent prob-
lem posed in a finite domain using Steklov–Poincaré operators as non-reflecting
boundary conditions in the artificial boundary.
We introduce a circle ΓR which encloses the obstacles and define the Steklov–
Poincaré operator on ΓR , see Fig. 1. This operator associates to any Dirichlet
data on ΓR the elastic conormal derivative of the solution of the exterior Dirichlet
problem in R2 \ B R , BR := B(0, R):

L : (H 1/2 (ΓR ))2 −→ (H −1/2 (ΓR ))2

 
∂wm
g −→ ∂n w = cjαmβ nα
∂zβ j

where w ∈ (H 1 (BR ))2 is the unique radiating solution of

⎧
⎪
⎪ ∂ 2 um
⎨ ρω 2 uj + cjαmβ = 0, j = 1, 2 in R2 \ B R ,
∂zα ∂zβ
⎪
⎪
⎩ w = g, on ΓR .

For a detailed representation of the Steklov–Poincaré operator in terms of inte-

gral kernels, we refer to [20] and references therein.
Then, we may replace (63) by an equivalent boundary value problem in BR
with a transparent boundary condition on the artificial boundary ΓR :
⎧ 2
⎪
⎪ ρω 2 uj + cjαmβ ∂ um = 0, j = 1, 2,
⎪
⎪ in Ωe := BR \ Ω i ,
⎪
⎪ ∂z ∂z
⎪
⎪
α β
⎨ ∂um
cjαmβ nα = 0, j = 1, 2, on Γ ,
⎪
⎪ ∂zβ
⎪
⎪
⎪
⎪ ∂(u − uinc )m
⎪
⎪
⎩ cjαmβ nα = L(u − uinc )j , j = 1, 2, on ΓR .
∂zβ
(67)
Problem (67) admits the variational formulation
⎧
⎪
⎨ u ∈ (H 1 (BR ))2 ,
(68)
⎪
⎩ b(Ω ; u, v) = (v),
i ∀v ∈ (H (BR )) ,
1 2
 Topological Derivatives for Shape Reconstruction 127

where



∂uj ∂v m
b(Ωi ; u, v) := (cjαmβ − ρω 2 uv) dz − Lu v dl,
Ωe ∂zα ∂zβ ΓR

∀u, v ∈ (H 1 (BR ))2 ,

∂uinc,m
(v) := (cjαmβ nα v j − Luinc v) dl, ∀v ∈ (H 1 (BR ))2 .
ΓR ∂zβ
Step 2: Transformed problems. Let uτ be the solution of the variational for-
mulations in the transformed domains:
⎧
⎪
⎨ uτ ∈ (H 1 (BR ))2 ,
(69)
⎪
⎩ b(Ω ; u , v) = (v),
i,τ τ ∀v ∈ (H (BR )) ,
1 2

with


∂uj ∂v m
b(Ωi,τ ; u, v) := (cjαmβ − ρω 2 uv) dzτ

Ωe,τ ∂zτ,α ∂zτ,β


− Lu v dlτ , ∀u, v ∈ (H 1 (BR ))2 .
ΓR

When τ = 0, u0 = u is the solution to problem (68). The transformed function-

als are

1
J(Ωi,τ ) = |uτ − umeas |2 dl. (70)
2 Γmeas
Step 3: Adjoint problem. Differentiating (70) with respect to τ we obtain
$

d $
J(Ωi,τ )$$ = Re (u − umeas )u̇ , (71)
dτ τ =0 Γmeas
$
d
where u̇ = dτ uτ $τ =0 . We define the Lagrangian functional

L(Ωi,τ ; v, p) = J(Ωi,τ ) + Re[b(Ωi,τ ; v, p) − (p)], ∀v, p ∈ (H 1 (BR ))2 .

When v = uτ solves (69),

L(Ωi,τ ; uτ , p) = J(Ωi,τ ), ∀p ∈ (H 1 (BR ))2 . (72)

Using (71)–(72), the shape derivative of J in the direction V is

$ $
d $ d $
DJ(Ωi ) · V = $
L(Ωi,τ ; uτ , p)$ = Re b(Ωi,τ ; u, p)$$
dτ τ =0 dτ τ =0


+ Re b(Ωi ; u̇, p) + Re (u − umeas )u̇ . (73)
Γmeas
128 A. Carpio and M.L. Rapún

Choosing p to be the solution of

⎧
⎪
⎪ p ∈ (H 1 (BR ))2 ,
⎨

(74)
⎪
⎪ (umeas − u) v, ∀v ∈ (H 1 (BR ))2 ,
⎩ b(Ωi ; v, p) =
Γmeas

the derivative u̇ is not needed and the shape derivative DJ(Ω) · V is given by
the first term in (73). Problem (74) is the adjoint problem of (68).
The definition of Steklov–Poincaré operator implies that L∗ (p) = L(p).
Thus, the adjoint problem is equivalent to:
⎧
⎪
⎪
⎨ p ∈ (H (BR )) ,
1 2

(75)
⎪
⎪ (umeas − u) v, ∀v ∈ (H 1 (BR ))2 .
⎩ b(Ωi ; p, v) =
Γmeas

Notice that p is then a solution of (66).

Step 4: Shape derivative. Let us calculate the first term in (73). For all u, p ∈
(H 1 (BR ))2
$

d $ ∂uj ∂pm
b(Ωi,τ ; u, p)$$ = (cjαmβ − ρω 2 up) ∇ · V dz
dτ τ =0 Ωe ∂zα ∂zβ



∂Vγ ∂uj ∂pm ∂Vγ ∂uj ∂pm
− cjαmβ dz − cjαmβ dz.
Ωe ∂z α ∂z γ ∂z β 
Ωe ∂zβ ∂zα ∂zγ

Integrating by parts, taking into account that n is the outward normal vector
to Ωe and that V vanishes on ΓR
$

d $ ∂uj ∂pm
$
b(Ωi,τ ; u, p)$ = (cjαmβ − ρω 2 up) Vn dl
dτ τ =0 Γ ∂zα ∂zβ


∂ 2 uj ∂pm ∂uj ∂ 2 pm ∂u ∂p
+ cjαmβ ( + ) − ρω 2 ( p+u ) Vγ dz
Ωe ∂z α ∂z γ ∂z β ∂z α ∂z β ∂z γ ∂z γ ∂z γ

∂ 2 uj ∂pm ∂uj ∂ 2 pm ∂ 2 uj ∂pm ∂uj ∂ 2 pm
− cjαmβ ( + + + ) Vγ dz
Ωe ∂zα ∂zγ ∂zβ ∂zγ ∂zβ ∂zα ∂zα ∂zβ ∂zγ ∂zα ∂zβ ∂zγ


∂u− −
j ∂pm ∂u− −
j ∂pm
+ cjαmβ (nα + nβ )Vn nγ dl. (76)
Γ ∂zγ ∂zβ ∂zα ∂zγ

The sum of all the integrals over Ωe that appear in (76) is


∂ 2 uj ∂p ∂ 2 pm ∂uj
− (cjαmβ +ρω 2 um ) m + (cjαmβ +ρω 2 pj ) Vγ dz = 0

Ωe ∂z α ∂z β ∂z γ ∂z β ∂z α ∂zγ
 Topological Derivatives for Shape Reconstruction 129

because
∂ 2 um ∂ 2 pm
ρω 2 uj + cjαmβ = 0, ρω 2 pj + cjαmβ = (umeas − u)j δΓmeas , in Ωe
∂zα ∂zβ ∂zα ∂zβ
and cjαmβ = cmβjα . The field V = 0 on Γmeas , thus all the integrals over Γmeas
are identically zero.
Adding the integrals on Γ we find

 
∂uj ∂pm
cjαmβ − ρω 2 up Vn dl
Γ ∂zα ∂zβ

 
∂uj ∂pm ∂uj ∂pm
+ cjαmβ nα + nβ Vn nγ dl.
Γ ∂zγ ∂zβ ∂zα ∂zγ
Using the homogeneous Neumann condition, the second integral vanishes and
(65) follows. 2

7.3 Topological Derivative

Once the shape derivative (65) is known, we use (6) to obtain

 
−1 ∂uεj ∂pεm
DT (x, R ) = lim 
2
Re cjαmβ − ρω u p ) dl .
2 ε ε
(77)
ε→0 V (ε) Γε ∂zα ∂zβ
Here, uε and pε are the solutions of the forward and adjoint Neumann problems
(63) and (66) when Ωe = R2 \ B ε (x) and Γ = Γε = ∂Bε (x). As in the case of
Helmholtz transmission problems, an asymptotic analysis of these solutions and
their gradients at Γε will help us to obtain the final expression for the topological
derivative.
Theorem 7.2. The topological derivative of the cost functional (64) is given by
∂uj ∂p
DT (x, R2 ) = Re ajαmβ (x) m (x) − ρω 2 u(x)p(x) (78)
∂zα ∂zβ
where u and p solve the forward and adjoint Neumann problems with Ωi = ∅.
The polarization tensor ajαmβ is defined by

 
∂θrpq ∂θrp q
ajαmβ = cjαmβ + crsr s cpqjα cp q mβ dξ (79)
|ξ |=1 ∂ξs ∂ξs


∂θrpq
+ (cjαrs cpqmβ + crsmβ cpqjα ) dξ, (80)
|ξ |=1 ∂ξs

pq pq
where w(ξ) = cpqrs ∂u
∂zs (x)θ (ξ), and θ
r
solve the problems
⎧
⎪
⎪ ∂ 2 θm
pq
⎪
⎨ cjαmβ ∂ξ ∂ξ = 0, j = 1, 2, in R2 \ B(0; 1),
α β
(81)
⎪
⎪ ∂θpq
⎪ pq
⎩ −cjαmβ m nα = Mjα nα , j = 1, 2, on Γ ,
∂ξβ
130 A. Carpio and M.L. Rapún

together with a radiation condition imposing decay at infinity. The matrices M pq

are defined by
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
1 0 0 0 0 1/2
M 11 = ⎝ ⎠ , M 22 = ⎝ ⎠ , M 12 = ⎝ ⎠ = M 21 .
00 01 1/2 0

In the 3D isotropic case, an explicit formula in terms of the elastic constants is

given in [25] using an explicit expression for w. The 3D isotropic transmission
problem was studied in [26].
Proof. We expand uε (z) = u(z)χR2 \B ε (z) + wε (z), where u = uinc is the so-
lution of the forward problem without obstacle (Ωi = ∅) and wε is the radiating
solution of
⎧
⎪
⎪ ∂ 2 wm
ε
⎪
⎨ ρω 2 ε
wj + cjαmβ = 0, j = 1, 2, in R2 \ B ε ,
∂zα ∂zβ
(82)
⎪
⎪ ∂w ε
∂u
⎪ −cjαmβ
⎩ m
nα = cjαmβ
m
nα , j = 1, 2, on Γε ,
∂zβ ∂zβ

with Neumann data given by u. The change of variable ξ = (z − x)/ε transforms

this problem into:
⎧
⎪
⎪ ∂ 2 wm
ε
⎪
⎨ ε2
ρω 2 ε
wj + c jαmβ = 0, j = 1, 2, in R2 \ B,
∂ξα ∂ξβ
(83)
⎪
⎪ ∂wmε
∂um
⎪
⎩ −cjαmβ nα = εcjαmβ nα , j = 1, 2, on Γ ,
∂ξβ ∂zβ
where B is the unit ball. As ε → 0, we expand
∂um (z) ∂um (x)
= + O(ε)
∂zβ ∂zβ

and write wε = w(1) + εw(2) + O(ε2 ). The leading term, w(1) is the radiating
solution of ⎧ (1)
⎪
⎪ ∂ 2 wm
⎪ cjαmβ
⎪ = 0, in R2 \ B,
⎨ ∂ξα ∂ξβ
(84)
⎪
⎪ (1)
⎪
⎪
∂wm
⎩ −cjαmβ nα = 0, on Γ .
∂ξβ

The solution of this problem is w(1) (ξ) = 0. The second term, w(2) , is the
radiating solution of
⎧
⎪
⎪
(2)
∂ 2 wm
⎪
⎪ in R2 \ B,
⎨ jαmβ ∂ξα ∂ξβ = 0,
c
(85)
⎪
⎪ (2)
⎪
⎪ ∂w m ∂u m
⎩ −cjαmβ nα = cjαmβ (x)nα , on Γ .
∂ξβ ∂zβ
 Topological Derivatives for Shape Reconstruction 131

Notice that, since the interface Γ is the unit circle, n = ξ/|ξ| = ξ. The function
w(2) can be written in terms of the elementary solutions of (81) as w(2) (ξ) =
pq
cpqrs ∂u
∂zs (x)θ (ξ). Our expansion yields:
r

uε (z) = u(x) + O(ε),

∂uεm ∂um ∂ur ∂θ pq
(z) = (x) + cpqrs (x) m (ξ) + O(ε), m, β = 1, 2,
∂zβ ∂zβ ∂zs ∂ξβ
and similar expressions for p. Recalling that V  (ε) = −2πε, we find (78). 2

8 Conclusions
Topological derivative methods are a powerful tool to devise efficient numerical
schemes for solving inverse scattering problems. In problems where the incident
radiation on the obstacles is governed by Helmholtz or elasticity equations, we
have computed expressions for the topological derivatives of appropriate con-
strained optimization regularizations. Topological derivative based numerical
strategies to approximate the scatterers have been discussed. We have performed
a number of tests in 2D and 3D illustrating the advantages and disadvantages of
the method. In general, fairly good approximations of the number, size, location
and shape of the scatterers are obtained at a low computational cost.
Many real situations require not only information about the scatterers, but
also about constitutive parameters that we assume here to be known. An ex-
tension of topological derivative methods to provide information on both the
obstacles and their parameters is proposed in [4].

Acknowledgements
This research has been supported by Grants MAT2005-05730-C02-02 and MTM-
2004-01905 of the Spanish MEC, BSCH/UCM PR27/05-13939 and CM-910143.

References
1. Anagnostopoulos KA and Charalambopoulos A 2006 The linear sampling method
for the transmission problem in two-dimensional anisotropic elasticity Inverse prob-
lems 22 553–577
2. Bonnet M and Constantinescu A 2005 Inverse problems in elasticity Inverse prob-
lems 21 R1-R50
3. Burger M, Hackl B and Ring W 2004 Incorporating topological derivatives into
level set methods J. Comp. Phys. 194 344–362
4. Carpio A and Rapun ML, Solving inhomogeneous inverse problems by topological
derivative methods, submitted, 2007
5. Carpio A and Rapun ML, Shape reconstruction in anisotropic elastic media. In
preparation
6. Chandler GA and Sloan IH 1990. Spline qualocation methods for boundary integral
equations. Numer. Math. 58, 537–567
132 A. Carpio and M.L. Rapún

7. Charalambopoulos A 2002 On the interior transmission problem in nondissipative,

homogeneous, anisotropic elasticity J. Elast. 67 149–170
8. Colton D 1984 The inverse scattering problem for time-harmonic acoustic waves,
SIAM Review 26 323–350.
9. Colton D, Gieberman K and Monk P 2000 A regularized sampling method for
solving three dimensional inverse scattering problems SIAM J. Sci. Comput. 21
2316–2330
10. Colton D and Kress R 1983 Integral equation methods in scattering theory John
Wiley & Sons. New York.
11. Colton D and Kress R 1992 Inverse acoustic and electromagnetic scattering theory
Springer Berlin.
12. Colton D and Kirsch A 1996 A simple method for solving inverse scattering prob-
lems in the resonance region Inverse problems 12 383–393
13. Costabel M and Stephan E 1985 A direct boundary integral equation method for
transmission problems J. Math. Anal. Appl. 106 367–413
14. Devaney AJ 1984 Geophysical diffraction tomography, IEEE Trans. Geosci. Re-
mote Sens. 22 3–13
15. Domı́nguez V, Rapún ML and Sayas FJ 2007 Dirac delta methods for Helmholtz
transmission problems. To appear in Adv. Comput. Math.
16. Domı́nguez V and Sayas FJ 2003. An asymptotic series approach to qualocation
methods. J. Integral Equations Appl. 15, 113–151
17. Dorn O and Lesselier D 2006 Level set methods for inverse scattering Inverse
Problems 22 R67–R131
18. Feijoo GR 2004 A new method in inverse scattering based on the topological deriv-
ative Inverse Problems 20 1819–1840
19. Feijoo GR, Oberai AA and Pinsky PM 2004 An application of shape optimization
in the solution of inverse acoustic scattering problems Inverse problems 20 199–228
20. Gachechiladze A and Natroshvili D 2001 Boundary variational inequality approach
in the anisotropic elasticity for the Signorini problem Georgian Math. J. 8 469-492
21. Garreau S, Guillaume P and Masmoudi M 2001 The topological asymptotic for
PDE systems: the elasticity case SIAM J. Control Optim. 39 1756–1778
22. Gegelia T and Jentsch L 1994 Potential methods in continuum mechanics Georgian
Math. J. 599-640
23. Gerlach T and Kress R 1996 Uniqueness in inverse obstacle scattering with con-
ductive boundary condition. Inverse Problems 12 619–625
24. Guzina BB and Bonnet M 2006 Small-inclusion asymptotic of misfit functionals
for inverse problems in acoustics Inverse Problems 22 1761–1785
25. Guzina BB, Bonnet M 2004 Topological derivative for the inverse scattering of
elastic waves Q. Jl. Mech. Appl. Math. 57 161–179
26. Guzina BB, Chikichev I 2007 From imaging to material identification: A general-
ized concept of topological sensitivity J. Mech. Phys. Sol. 55 245–279
27. Hettlich F 1995 Fréchet derivatives in inverse obstacle scattering Inverse problems
11 371–382
28. Keller JB and Givoli D 1989 Exact non-reflecting boundary conditions J. Comput.
Phys. 82 172–192
29. Kirsch A, Kress R, Monk P and Zinn A 1988 Two methods for solving the inverse
acoustic scattering problem Inverse problems 4 749–770
30. Kirsch A and Kress R 1993 Uniqueness in inverse obstacle scattering Inverse Prob-
lems 9 285–299
31. Kirsch A 1993 The domain derivative and two applications in inverse scattering
theory Inverse Problems 9 81–93
 Topological Derivatives for Shape Reconstruction 133

32. Kleinman RE and Martin P 1988 On single integral equations for the transmission
problem of acoustics SIAM J. Appl. Math 48 307–325
33. Kleinman RE and van der Berg PM 1992 A modified gradient method for two
dimensional problems in tomography J. Comput. Appl. Math. 42 17–35
34. Kress R and Roach GF 1978 Transmission problems for the Helmholtz equation
J. Math. Phys. 19 1433–1437
35. Kupradze VD, Gegelia TG, Basheleuishvili MO and Burchauladze TV, Three di-
mensional problems of the mathematical theory of elasticity and thermoelasticity,
North-Holland Ser. Appl. Math. Mech. 25, North-Holland, Amsterdam, 1979.
36. Liseno A and Pierri R 2004 Imaging of voids by means of a physical optics based
shape reconstruction algorithm J. Opt. Soc. Am. A 21 968–974
37. Litman A, Lesselier D and Santosa F 1998 Reconstruction of a two-dimensional
binary obstacle by controlled evolution of a level set Inverse problems 14 685–706
38. Masmoudi M 1987 Outils pour la conception optimale des formes Thèse d’Etat en
Sciences Mathématiques, Université de Nice
39. Masmoudi M, Pommier J and Samet B 2005 The topological asymptotic expansion
for the Maxwell equations and some applications Inverse Problems 21 547–564
40. Meddahi S and Sayas FJ 2005 Analysis of a new BEM–FEM coupling for two
dimensional fluid–solid iteraction Num. Methods Part. Diff. Eq. 21 1017–1154
41. Natterer F and Wubbeling F 1995 A propagation backpropagation method for
ultrasound tomography Inverse problems 11 1225-1232
42. Natroshvili D 1995 Mixed interface problems for anisotropic elastic bodies Georgian
Math. J. 2 631-652
43. Natroshvili D 1996 Two-dimensional steady state oscillation problems of
anisotropic elasticity Georgian Math. J. 3 239-262
44. Potthast R 1996 Fréchet differentiability of the solution to the acoustic Neumann
scattering problem with respect to the domain J. Inverse Ill-Posed Problems 4
67–84
45. Rapún ML and Sayas FJ 2006 Boundary integral approximation of a heat diffusion
problem in time–harmonic regime Numer. Algorithms 41 127–160
46. Rapún ML and Sayas FJ 2006 Indirect methods with Brakhage–Werner potentials
for Helmholtz transmission problems. In Numerical Mathematics and advanced
applications. ENUMATH 2005. Springer 1146–1154
47. Rapún ML and Sayas FJ 2007 Exterior Dirichlet and Neumann problems for the
Hemholtz equation as limits of transmission problems. Submitted.
48. Rapún ML and Sayas FJ 2007 Boundary element simulation of thermal waves.
Arch. Comput. Methods. Engrg 14 3–46
49. Samet B, Amstutz S and Masmoudi M 2003 The topological asymptotic for the
Helmholtz equation SIAM J. Control Optimim 42 1523–1544
50. Santosa F 1996 A level set approach for inverse problems involving obstacles
ESAIM Control, Optim. Calculus Variations 1 17–33
51. Sloan IH 2000. Qualocation. J. Comput. Appl. Math. 125, 461–478
52. Sokolowski J and Zolésio JP 1992 Introduction to shape optimization. Shape sen-
sitivity analysis (Heidelberg: Springer)
53. Sokowloski J and Zochowski A 1999 On the topological derivative in shape opti-
mization SIAM J. Control Optim. 37 1251–1272
54. Torres RH and Welland GV 1993 The Helmholtz equation and transmission prob-
lems with Lipschitz interfaces Indiana Univ. Math. J. 42 1457–1485
55. von Petersdorff T 1989 Boundary integral equations for mixed Dirichlet, Neumann
and transmission problems. Math. Methods Appl. Sci. 11 185–213
Time-Reversal and the Adjoint Imaging Method
with an Application in Telecommunication

Oliver Dorn

Gregorio Millán Institute of Fluid Dynamics, Nanoscience and Industrial

Mathematics, Universidad Carlos III de Madrid, Leganés, Spain
[email protected]

Summary. These lecture notes provide a mathematical treatment of time-reversal ex-

periments with a special emphasis on telecommunication. A direct link is established
between time-reversal experiments and the adjoint imaging method. Based on this rela-
tionship, several iterative schemes are proposed for optimizing MIMO (multiple-input
multiple-output) time-reversal systems in underwater acoustic and in wireless com-
munication systems. Whereas in typical imaging applications these iterative schemes
require the repeated solution of forward problems in a computer, the analogue in time-
reversal communication schemes consists of a small number of physical time-reversal
experiments and does not require exact knowledge of the environment in which the
communication system operates. The discussion is put in the general framework of
wave propagation by symmetric hyperbolic systems, with detailed discussions of the
linear acoustic system for underwater communication and of the time-dependent sys-
tem of Maxwell’s equations for telecommunication. Moreover, in its general form as
treated here, the theory will also apply for several other models of wave-propagation,
such as for example linear elastic waves.

1 Introduction
Time-reversal techniques have attracted great attention recently due to the large
variety of interesting potential applications. The basic idea of time-reversal (often
also referred to as ‘phase-conjugation’ in the frequency domain) can be roughly
described as follows. A localized source emits a short pulse of acoustic, electro-
magnetic or elastic energy which propagates through a richly scattering envi-
ronment. A receiver array records the arriving waves, typically as a long and
complicated signal due to the complexity of the environment, and stores the
time-series of measured signals in memory. After a short while, the receiver ar-
ray sends the recorded signals time-reversed (i.e. first in – last out) back into
the same medium. Due to the time-reversibility of the wave fields, the emitted
energy backpropagates through the same environment, practically retracing the
paths of the original signal, and refocuses, again as a short pulse, on the loca-
tion where the source emitted the original pulse. This, certainly, is a slightly
oversimplified description of the rather complex physics which is involved in the
136 O. Dorn

real time-reversal experiment. In practice, the quality of the refocused pulse de-
pends on many parameters, as for example the randomness of the background
medium, the size and location of the receiver array, temporal fluctuations of
the environment, etc. Surprisingly, the quality of the refocused signal increases
with increasing complexity of the background environment. This was observed
experimentally by M. Fink and his group at the Laboratoire Ondes et Acous-
tique at Université Paris VII in Paris by a series of laboratory experiments (see
for example [17, 20, 21]), and by W. A. Kuperman and co-workers at the Scripps
Institution of Oceanography at University of California, San Diego, by a series of
experiments performed between 1996 and 2000 in a shallow ocean environment
(see for example [9, 18, 26]).
The list of possible applications of this time-reversal technique is long. In an
iterated fashion, resembling the power method for finding maximal eigenvalues
of a square matrix, the time-reversal method can be applied for focusing energy
created by an ultrasound transducer array on strongly scattering objects in the
region of interest. This can be used for example in lithotripsy for localizing and
destroying gall-stones in an automatic way, or more generally in the application
of medical imaging problems. Detection of cracks in the aeronautic industry, or
of submarines in the ocean are other examples. See for example [39, 40], and for
related work [7,12]. The general idea of time-reversibility, and its use in imaging
and detection, is certainly not that new. Looking into the literature for example
of seismic imaging, the application of this basic idea can be found in a classical
and very successful imaging strategy for detecting scattering interfaces in the
Earth, the so-called ‘migration’ technique [4,8]. However, the systematic use and
investigation of the time-reversal phenomenon and its experimental realizations
started more recently, and has been carried out during the last 10–15 years or
so by different research groups. See for example [17, 18, 20, 21, 26, 31, 46, 49] for
experimental demonstrations, and [1–3,5,6,9,15,16,23,28,30,38,39,43–45,47,50]
for theoretical and numerical approaches.
One very young and promising application of time-reversal is communication.
In these lecture notes we will mainly concentrate on that application, although
the general results should carry over also to other applications as mentioned
above.
The text is organized as follows. In Sect. 2 we give a very short introduction
into time-reversal in the ocean, with a special emphasis on underwater sound
communication. Wireless communication in a MIMO setup, our second main
application in these notes, is briefly presented in Sect. 3. In Sect. 4 we discuss
symmetric hyperbolic systems in the form needed here, and examples of such
systems are given in Sect. 5. In Sect. 6, the basic spaces and operators necessary
for our mathematical treatment are introduced. The inverse problem in com-
munication, which we are focusing on in these notes, is then defined in Sect. 7.
In Sect. 8, we derive the basic iterative scheme for solving this inverse problem.
Section 9 gives practical expressions for calculating the adjoint communication
operator, which plays a key role in the iterative time-reversal schemes. In Sect. 10
the acoustic time-reversal mirror is defined, which will provide the link between
 Time-Reversal and the Adjoint Imaging Method in Communication 137

the ‘acoustic time-reversal experiment’ and the adjoint communication operator.

The analogous results for the electromagnetic time-reversal mirror are discussed
in Sect. 11. Section 12 combines the results of these two sections, and explic-
itly provides the link between time-reversal and the adjoint imaging method.
Sections 13, 14, and 15 propose then several different iterative time-reversal
schemes for solving the inverse problem of communication, using this key re-
lationship between time-reversal and the adjoint communication operator. The
practically important issue of partial measurements (and generalized measure-
ments) is treated in Sect. 16. Finally, Sect. 17 summarizes the results of these
lecture notes, and points out some interesting future research directions.

2 Time-Reversal and Communication in the Ocean

The ocean is a complex wave-guide for sound [10]. In addition to scattering effects
at the top and the bottom of the ocean, also its temperature profile and the cor-
responding refractive effects contribute to this wave-guiding property and allow
acoustic energy to travel large distances. Typically, this propagation has a very
complicated behaviour. For example, multipathing occurs if source and receiver
of sound waves are far away from each other, since due to scattering and refrac-
tion there are many possible paths on which acoustic energy can travel between
them. Surface waves and air bubbles at the top of the ocean, sound propagation
through the rocks and sedimentary layers at the bottom of the ocean, and other
effects further contribute to the complexity of sound propagation in the ocean.
When a source (e.g. the base station of a communication system in the ocean)
emits a short signal at a given location, the receiver (e.g. a user of this commu-
nication system) some distance away from the source typically receives a long
and complicated signal due to the various influences of the ocean to this signal
along the different connecting paths. If the base station wants to communicate
with the user by sending a series of short signals, this complex response of the
ocean to the signal needs to be resolved and taken into account.
In a classical communication system, the base station which wants to com-
municate with a user broadcasts a series of short signals (e.g. a series of ‘zeros’
and ‘ones’) into the environment. The hope is that the user will receive this
message as a similar series of temporally well-resolved short signals which can
be easily identified and decoded. However, this almost never occurs in a com-
plex environment, due to the multipathing and the resulting delay-spread of the
emitted signals. Typically, when the base station broadcasts a series of short
signals into such an environment, intersymbol interference occurs at the user
position due to the temporal overlap of the multipath contributions of these
signals. In order to recover the individual signals, a significant amount of signal
processing is necessary at the user side, and, most importantly, the user needs
to have some knowledge of the propagation behaviour of the signals in this en-
vironment (i.e. he needs to know the ‘channel’). Intersymbol interference can
in principle be avoided by adding a sufficient delay between individual signals
emitted at the base station which takes into account the delay-spread in the
138 O. Dorn

medium. That, however, slows down the communication, and reduces the capac-
ity of the environment as a communication system. An additional drawback of
simply broadcasting communication signals from the base station is the obvious
lack of interception security. A different user of the system who also knows the
propagation behaviour of signals in the environment, can equally well resolve
the series of signals and decode them.
Several approaches have been suggested to circumvent the above mentioned
drawbacks of communication in multiple-scattering environments. Some very
promising techniques are based on the time-reversibility property of propagat-
ing wave-fields [18, 19, 29, 32, 34, 35, 37, 44]. The basic idea is as follows. The
user who wants to communicate with the base station, starts the communica-
tion process by sending a short pilot signal through the environment. The base
station receives this signal as a long and complex signal due to multipathing.
It time-reverses the received signal and sends it back into the environment.
The backpropagating waves will produce a complicated wave-field everywhere
due to the many interfering parts of the emitted signal. However, due to the
time-reversibility of the wave-fields, one expects that the interference will be
constructive at the position of the user who sent the original pilot signal, and
mainly destructive at all other positions. Therefore, the user will receive a short
signal very similar to (ideally, a time-reversed replica of) the originally sent pilot
signal. All other users who might be in the environment at the same time will
only receive noise speckle due to incoherently interfering contributions of the
backpropagating field. If the base station sends the individual elements (‘ones’
and ‘zeros’) of the intended message in a phase-encoded form as a long overlap-
ping string of signals, the superposition principle will ensure that, at the user
position, this string of signals will appear as a series of short well-separated sig-
nals, each resembling some phase-shifted (and time-reversed) form of the pilot
signal.
In order to find out whether this theoretically predicted scenario actually
takes place in a real multiple-scattering environment like the ocean, Kuperman
et al. have performed a series of four experiments in a shallow ocean environment
between 1996 and 2000, essentially following the above described scenario. The
experiments have been performed at a Mediterranean location close to the Italian
coast. (A similar setup was also used in an experiment performed in 2001 off the
coast of New England which has been reported in Yang [49].) A schematic view
of these experiments is shown in Fig. 1. The single ‘user’ is replaced here by a
‘probe source’, and the ‘source-receive array’ (SRA) plays the role of the ‘base
station’. An additional vertical receive array (VRA) was deployed at the position
of the probe source in order to measure the temporal and spatial spread of the
backpropagating fields in the neighbourhood of the probe source location. In this
shallow environment (depths of about 100–200 m, and distances between 10 and
30 km) the multipathing of the waves is mostly caused by multiple reflections
at the surface and the bottom of the ocean. The results of the experiments
have been reported in [18, 26, 31]. They show that in fact a strong spatial and
temporal focusing of the backpropagating waves occurs at the source position.
 Time-Reversal and the Adjoint Imaging Method in Communication 139

Source Probe Vertical

Receive Source
Receive
Array ~ 80 m Array
(SRA)
(VRA)

~ 130 m

~ 10 km

Fig. 1. The communication problem in underwater acoustics

A theoretical explanation of the temporal and spatial refocusing of time-reversed

waves in random environments has been given in Blomgren et al. [5].

3 The MIMO Setup in Wireless Communication

The underwater acoustics scenario described above directly carries over to situa-
tions which might be more familiar to most of us, namely to the more and more
popular wireless communication networks using mainly electromagnetic waves
in the microwave regime. Starting from the everyday use of cell-phones, ranging
to small wireless-operating local area networks (LAN) for computer systems or
for private enterprise communication systems, exactly the same problems arise
as in underwater communication. The typically employed microwaves of a wave-
length at about 10–30 cm are heavily scattered by environmental objects like
cars, fences, trees, doors, furniture, etc. This causes a very complicated multi-
path structure of the signals received by users of such a communication system.
Since bandwidths are limited and increasingly expensive, a need for more and
more efficient communication systems is imminent. Recently, the idea of a so-
called multiple-input multiple-output (MIMO) communication system has been
introduced with the potential to increase the capacity and efficiency of wireless
communication systems [22]. The idea is to replace a single antenna at the base
station which is responsible for multiple users, or even a system where one user
communicates with just one dedicated base antenna, by a more general system
where an array of multiple antennas at the base station is interacting simultane-
ously and in a complex way with multiple users. A schematic description of such
a MIMO system (with seven base antennas and seven users) is given in Fig. 2.
See for example [24, 42] for recent overviews on MIMO technology.
Time-reversal techniques are likely to play also here a key role in improving
communication procedures and for optimizing the use of the limited resources
140 O. Dorn

users base antennas

complex
pilot signal α(t) received signals
U1 multiple
A1

U3 A4

A7
U5
scattering

U7
environment

Fig. 2. Schematic view of a typical MIMO setup in wireless communication

(especially bandwidth) which are available for this technology [19, 29, 32, 37].
One big advantage of time-reversal techniques is that they are automatically
adapted to the complex environment and that they can be very fast since they
do not require heavy signal processing at the receiver or user side. In [34, 35],
an iterated time-reversal scheme for the optimal refocusing of signals in such
a MIMO communication system was proposed, which we will describe in more
details in Sect. 13.
We will establish a direct link between the time-reversal technique and so-
lution strategies for inverse problems. As an application of this relationship, we
will derive iterative time-reversal schemes for the optimization of wireless or un-
derwater acoustic MIMO communication systems. The derivation is performed
completely in time-domain, for very general first order symmetric hyperbolic sys-
tems describing wave propagation phenomena in a complex environment. One
of the schemes which we derive, in a certain sense the ‘basic one’, will turn out
to be practically equivalent to the scheme introduced in [34, 35], although the
derivation uses different tools. Therefore, we provide a new interpretation of that
scheme. The other schemes which we introduce are new in this application, and
can be considered as either generalizations of the basic scheme, or as independent
alternatives to that scheme. Each of them addresses slightly different objectives
and has its own very specific characteristics.

4 Symmetric Hyperbolic Systems

We treat wave propagation in communication systems in the general framework
of symmetric hyperbolic systems of the form
3
∂u ∂u
Γ (x) + Di + Φ(x)u = q (1)
∂t i=1
∂xi

u(x, 0) = 0. (2)
 Time-Reversal and the Adjoint Imaging Method in Communication 141

Here, u(x, t) and q(x, t) are real-valued time-dependent N -vectors, x ∈ R3 , and

t ∈ [0, T ]. Γ (x) is a real, symmetric, uniformly positive definite N × N -matrix,
i.e. Γ (x) ≥  for some  > 0. Moreover, Φ(x) is a symmetric positive semi-definite
N × N matrix, i.e. Φ(x) ≥ 0. It models possible energy loss through dissipation
in the medium. The Di are real, symmetric and independent of (x, t). We will
also use the short notation
3
∂
Λ := Di . (3)
i=1
∂xi

In addition to the above mentioned assumptions on the coefficients Γ (x) and

Φ(x), we will assume throughout this text that all quantities Γ (x), Φ(x), u(x, t)
and q(x, t) are ‘sufficiently regular’ in order to safely apply for example integra-
tion by parts and Green’s formulas. For details see for example [11, 33]. We will
assume that no energy reaches the boundaries ∂Ω during the time [0, T ], such
that we will always have
u(x, t) = 0 on ∂Ω × [0, T ]. (4)
The energy density E(x, t) is defined by
1
E(x, t) = Γ (x)u(x, t), u(x, t)N
2
N
1
= Γmn (x)um (x, t)un (x, t).
2 m,n=1

The total energy Ê(t) in Ω at a given time t is therefore

1
Ê(t) = Γ (x)u(x, t), u(x, t)N dx.
2 Ω
The flux F(x, t) is given by
1, i -
Fi (x, t) = D u(x, t), u(x, t) N , i = 1, 2, 3.
2

5 Examples for Symmetric Hyperbolic Systems

In the following, we want to give some examples for symmetric hyperbolic sys-
tems as defined above. Of special interest for communication are the system of
the linearized acoustic equations and the system of Maxwell’s equations. We will
discuss these two examples in detail in these lecture notes. Another important
example for symmetric hyperbolic systems is the system of elastic waves equa-
tions, which we will however leave out in our discussion for the sake of brevity. We
only mention that all the results derived here apply without restrictions also to
linear elastic waves. Elastic wave propagation becomes important for example
in ocean acoustic communication models which incorporate wave propagation
through the sedimentary and rock layers at the bottom of the ocean.
142 O. Dorn

5.1 Linearized Acoustic Equations

As a model for underwater sound propagation, we consider the following lin-

earized form of the acoustic equations in an isotropic medium
∂v
ρ(x) + grad p = qv (5)
∂t
∂p
κ(x) + div v = qp (6)
∂t
p(x, 0) = 0, v(x, 0) = 0. (7)
Here, v is the velocity, p the pressure, ρ the density, and κ the compressibility.
We have N = 4, u = (v, p)T and q = (qv , qp )T (where ‘T ’ as a superscript
always means ‘transpose’). Moreover, we have

Γ (x) = diag(ρ(x), ρ(x), ρ(x), κ(x)) and Φ(x) = 0.

With the notation ϕ = (∂1 , ∂2 , ∂3 )T , we can write Λ as

⎛ ⎞
0 ϕ
Λ = ⎝ ⎠.
ϕT 0

The operators Di , i = 1, 2, 3, can be recovered from Λ by putting

⎧
⎨ 1 where Λ
i m,n = ∂i ,
Dm,n =
⎩ 0 elsewhere

The energy density E(x, t) is given by

1 
E(x, t) = ρ(x)|v(x, t)|2 + κ(x)p2 (x, t) ,
2
and the energy flux F(x, t) is

F(x, t) = p(x, t)v(x, t).

We mention that the dissipative case Φ(x) = 0 can be treated as well in our
framework, and yields analogous results to those presented here.

5.2 Maxwell’s Equations

As a second example, we will consider Maxwell’s equations for an anisotropic

medium with some energy loss due to the inherent conductivity. This can for
example model wireless communication in a complex environment.
∂E
(x) − curl H + σ(x)E = qE (8)
∂t
∂H
µ(x) + curl E = qH (9)
∂t
 Time-Reversal and the Adjoint Imaging Method in Communication 143

E(x, 0) = 0, H(x, 0) = 0. (10)

We have N = 6, and u = (E, H)T , q = (qE , qH )T . Moreover,

Γ (x) = diag((x), µ(x))

Φ(x) = diag(σ(x), 0).

Here,  and µ are symmetric positive definite 3 × 3 matrices modelling the

anisotropic permittivity and permeability distribution in the medium, and σ
is a symmetric positive semi-definite 3 × 3 matrix which models the anisotropic
conductivity distribution. In wireless communication, this form can model for
example dissipation by conductive trees, conductive wires, rainfall, pipes, etc.
The operator Λ can be written in block form as
⎛ ⎞
0 −Ξ
Λ = ⎝ ⎠,
Ξ 0

with ⎛ ⎞
0 −∂3 ∂2
⎜ ⎟
⎜ ⎟
Ξ = ⎜ ∂3 0 −∂1 ⎟ .
⎝ ⎠
−∂2 ∂1 0

The operators Di , i = 1, 2, 3, can be recovered from Λ by putting

⎧
⎪
⎪
⎪
⎨
1 where Λm,n = ∂i ,
i
Dm,n = −1 where Λm,n = −∂i ,
⎪
⎪
⎪
⎩ 0 elsewhere

The energy density E(x, t) is given by

1 
E(x, t) = (x)|E(x, t)|2 + µ(x)|H(x, t)|2 .
2
The energy flux F(x, t) is described by the Poynting vector

F(x, t) = E(x, t) × H(x, t).

5.3 Elastic Waves Equations

As already mentioned, also elastic waves can be treated in the general framework
of symmetric hyperbolic systems. For more details we refer to [33, 41].
144 O. Dorn

6 The Basic Spaces and Operators

For our mathematical treatment of time-reversal we introduce the following
spaces and inner products. We assume that we have J users Uj , j = 1, . . . , J
in our system, and in addition a base station which consists of K antennas Ak ,
k = 1, . . . , K. Each user and each antenna at the base station can receive, process
and emit signals which we denote by sj (t) for a given user Uj , j = 1, . . . , J, and
by rk (t) for a given base antenna Ak , k = 1, . . . , K. Each of these signals con-
sists of a time-dependent N -vector indicating measured or processed signals of
time-length T . In our analysis we will often have to consider functions defined
on the time interval [T, 2T ] instead of [0, T ]. For simplicity, we will use the same
notation for the function spaces defined on [T, 2T ] as we use for those defined
on [0, T ]. It will always be obvious which space we refer to in a given situation.
Lumping together all signals at the users on the one hand, and all signals at
the base station on the other hand, yields the two fundamental quantities

s = (s1 , . . . , sJ ) ∈ Ẑ
r = (r1 , . . . , rK ) ∈ Z

with  J  K
Ẑ = L2 ([0, T ])N , Z = L2 ([0, T ])N .
The two signal spaces Z and Ẑ introduced above are equipped with the inner
products
  J
 
(1) (2) (1) (2)
s ,s = sj (t), sj (t) dt
Ẑ [0,T ] N
j=1
  K
 
(1) (2) (1) (2)
r ,r = rk (t), rk (t) dt.
Z [0,T ] N
k=1

The corresponding norms are

s2Ẑ = s, sẐ , r2Z = r, rZ

Each user Uj and each antenna Ak at the base station can send a given
signal sj (t) or rk (t), respectively. This gives rise to a source distribution q̂j (x, t),
j = 1, . . . , J, or qk (x, t), k = 1, . . . , K, respectively. Here and in the following we
will use in our notation the following convention. If one symbol appears in both
forms, with and without a ‘hat’ (ˆ) on top of this symbol, then all quantities with
the ‘hat’ symbol are related to the users, and those without the ‘hat’ symbol to
the antennas at the base station.
Each of the sources created by a user or by a base antenna will appear on
the right hand side of (1) as a mathematical source function and gives rise to a
corresponding wave field which satisfies (1), (2). When solving the system (1),
(2), typically certain Sobolev spaces need to be employed for the appropriate
description of the underlying function spaces (see for example [11]). For our
 Time-Reversal and the Adjoint Imaging Method in Communication 145

purposes, however, it will be sufficient to assume that both, source functions

and wave fields, are members of the following canonical function space U which
is defined as

U = {u ∈ L2 (Ω × [0, T ])N , u = 0 on ∂Ω × [0, T ] , uU < ∞},

and which we have equipped with the usual energy inner product



u(t), v(t)U = Γ (x)u(x, t), v(x, t)N dxdt,
[0,T ] Ω

and the corresponding energy norm

u2U = u , uU .

Also here, in order to simplify the notation, we will use the same space when
considering functions in the shifted time interval [T, 2T ] instead of [0, T ].
Typically, when a user or an antenna at the base station transforms a signal
into a source distribution, it is done according to a very specific antenna charac-
teristic which takes into account the spatial extension of the user or the antenna.
We will model this characteristic at the user by the functions γ̂j (x), j = 1, . . . , J,
and for base antennas by the functions γk (x), k = 1, . . . , K. With these func-
tions, we can introduce the linear ‘source operators’ Q̂ and Q mapping signals
s at the set of users and r at the set of base antennas into the corresponding
source distributions q̂(x, t) and q(x, t), respectively. They are given as
J
Q̂ : Ẑ → U, q̂(x, t) = Q̂s = γ̂j (x)sj (t),
j=1
K
Q : Z → U, q(x, t) = Qr = γk (x)rk (t).
k=1

We will assume that the functions γ̂j (x) are supported on a small neighbour-
hood V̂j of the user location d̂j , and that the functions γk (x) are supported on
a small neighbourhood Vk of the antenna location dk . Moreover, all these neigh-
bourhoods are strictly disjoint to each other. For example, the functions γ̂j (x)
could be assumed to be L2 -approximations of the Dirac delta measure δ(x − d̂j )
concentrated at the user locations d̂j , and the functions γk (x) could be assumed
to be L2 -approximations of the Dirac delta measure δ(x − dk ) concentrated at
the antenna locations dk .
Both, users and base antennas can also record incoming fields u ∈ U and
transform the recorded information into signals. Also here, this is usually done
according to very specific antenna characteristics of each user and each base
antenna. For simplicity (and without loss of generality), we will assume that
the antenna characteristic of a user or base antenna for receiving signals is the
same as for transmitting signals, namely γ̂j (x) for the user and γk (x) for a base
146 O. Dorn

antenna. (The case of more general source and measurement operators is dis-
cussed in Sect. 16.) With this, we can define the linear ‘measurement operators’
M̂ : U → Ẑ and M : U → Z, respectively, which transform incoming fields
into measured signals, by


sj (t) = (M̂ u)j = γ̂j (x)u(x, t)dx, (j = 1, . . . , J)

Ω
rk (t) = (M u)k = γk (x)u(x, t)dx, (k = 1, . . . , K)
Ω

Finally, we define the linear operator F mapping sources q to states u by

F : U → U, F q = u,

where u solves the problem (1), (2). As already mentioned, we assume that the
domain Ω is chosen sufficiently large and that the boundary ∂Ω is sufficiently
far away from the users and base antennas, such that there is no energy reaching
the boundary in the time interval [0, T ] (or [T, 2T ]) due to the finite speed of
signal propagation. Therefore, the operator F is well-defined.
Formally, we can now introduce the two linear communication operators A
and B. They are defined as

A : Z → Ẑ, Ar = M̂ F Qr,
B : Ẑ → Z, Bs = M F Q̂s.

The operator A models the following situation. The base station emits the signal
r(t) which propagates through the complex environment. The users measure the
arriving wave fields and transform them into measurement signals. The measured
signals at the set of all users is s(t) = Ar(t). The operator B describes exactly
the reversed situation. All users emit together the set of signals s(t), which
propagate through the given complex environment and are received by the base
station. The corresponding set of measured signals at all antennas of the base
station is just r(t) = Bs(t). No time-reversal is involved so far.

7 An Inverse Problem Arising in Communication

In the following, we outline a typical problem arising in communication, which

gives rise to a mathematically well-defined inverse problem.
A specified user of the system, say U1 , defines a (typically but not necessarily
short) pilot signal α(t) with the goal to use it as a template for receiving the
information from the base station. The base station wants to emit a signal r̃(t)
which, after having travelled through the complex environment and arriving at
the user U1 , matches this pilot signal as closely as possible. Neither the base
station nor any other user except of U1 are required (or expected) to know the
correct form of the pilot signal α(t) for this problem. As an additional constraint,
 Time-Reversal and the Adjoint Imaging Method in Communication 147

the base station wants that at the other users Uj , j > 1, as little energy as pos-
sible arrives when communicating with the specified user U1 . This is also in the
interest of the other users, who want to use a different ‘channel’ for communicat-
ing at the same time with the base antenna, and want to minimize interference
with the communication initiated by user U1 . The complex environment itself
in which the communication takes place (i.e. the ‘channel’) is assumed to be
unknown to all users and to the base station.
In order to arrive at a mathematical description of this problem, we define
the ‘ideal signal’ s̃(t) received by all users as

s̃(t) = (α(t), 0, . . . , 0)T . (11)

Each user only knows its own component of this signal, and the base antenna
does not need to know any component of this ideal signal at all.
Definition 7.1. The inverse problem of communication: In the terminol-
ogy of inverse problems, the above described scenario defines an inverse source
problem, which we call for the purpose of these lecture notes the ‘inverse problem
of communication’. The goal is to find a ‘source distribution’ r̃(t) at the base
station which satisfies the ‘data’ s̃(t) at the users:

Ar̃ = s̃. (12)

The ‘state equation’ relating sources to data is given by the symmetric hyperbolic
system (1), (2).
Remark 7.1. Notice that the basic operator A in (12) is unknown to the users
and the base station since they typically do not know the complicated medium
in which the waves propagate. If the operator A (together with s̃) would be
known at the base station by some means, the inverse source problem formu-
lated above could be solved using classical inverse problems techniques, which
would be computationally expensive but in principle doable. In the given situ-
ation, the user and the base station are able to do physical experiments, which
amounts to ‘applying’ the communication operator A to a given signal. Deter-
mining the operator A explicitly by applying it to a set of basis functions of
Z would be possible, but again it would be too expensive. We will show in the
following that, nevertheless, many of the classical solution schemes known from
inverse problems theory can be applied in this situation even without knowing
the operator A explicitly. The basic tool which we will use is a series of time-
reversal experiments, applied to carefully designed signals at the users and the
base station.
Remark 7.2. A practical template for an iterative scheme for finding an optimal
signal at the base station can be envisioned as follows. User U1 starts the com-
(0)
munication process by emitting an initial signal s1 (t) into the complex environ-
ment. This signal, after having propagated through the complex environment,
finally arrives at the base station and is received there usually as a relatively long
and complicated signal due to the multiple scattering events it experienced on its
148 O. Dorn

way. When the base station receives such a signal, it processes it and sends a new
signal r(1) (t) back through the environment which is received by all users. After
receiving this signal, all users become active. The user U1 compares the received
signal with the pilot signal. If the match is not good enough, the received signal
is processed by this user in order to optimize the match with the pilot signal
when receiving the next iterate from the base station. All other users identify
the received signal as unwanted noise, and process it with the goal to receive
in the next iterate from the base station a signal with lower amplitude, such
that it does not interfere with their own communications. All users send now
their processed signals, which together define s(1) (t), back to the base station.
The base station receives them all simultaneously, again usually as a long and
complicated signal, processes this signal and sends a new signal r(2) (t) back into
the environment which ideally will match the desired signals at all users better
than the previously emitted signal r(1) (t). This iteration stops when all users
are satisfied, i.e. when user U1 receives a signal which is sufficiently close to the
pilot signal α(t), and the energy or amplitude of the signals arriving at the other
users has decreased enough in order not to disturb their own communications.
After this learning process of the channel has been completed, the user U1 can
now start communicating safely with the base antenna using the chosen pilot
signal α(t) for decoding the received signals.

Similar schemes have been suggested in [18, 19, 29, 32, 37, 44] in a single-step
fashion, and in [34, 35] performing multiple steps of the iteration. The main
questions to be answered are certainly which signals each user and each base
antenna needs to emit in each step, how these signals need to be processed,
at which stage this iteration should be terminated, and which optimal solution
this scheme is expected to converge to. One of the main objectives of these
lecture notes is to provide a theoretical framework for answering these questions
by combining basic concepts of inverse problems theory with experimental time-
reversal techniques.

8 The Basic Approach for Solving the Inverse Problem

of Communication
A standard approach for solving problem (12) in the situation of noisy data is
to look for the least-squares solution

rLS = Minr Ar − s̃2Ẑ (13)

In order to practically find a solution of (13), we introduce the cost functional

1 
J (r) = Ar − s̃ , Ar − s̃ (14)
2 Ẑ

In the basic approach we propose to use the gradient method for finding the
minimum of (14). In this method, in each iteration a correction δr is sought for
 Time-Reversal and the Adjoint Imaging Method in Communication 149

a guess r which points into the negative gradient direction −A∗ (Ar − s̃) of the
cost functional (14). In other words, starting with the initial guess r(0) = 0, the
iteration of the gradient method goes as follows:

r(0) = 0
(15)
r(n+1) = r(n) − β (n) A∗ (Ar(n) − s̃),

where β (n) is the step-size at iteration number n. Notice that the signals r(n) are
measured at the base station, whereas the difference Ar(n) − s̃ is determined at
the users. In particular, s̃ is the pilot signal only known by the user who defined
it, combined with zero signals at the remaining users. Ar(n) is the signal received
by the users at the n-th iteration step.

9 The Adjoint Operator A∗

We see from (15) that we will have to apply the adjoint operator A∗ repeatedly
when implementing the gradient method. In this section we provide practically
useful expressions for applying this operator to a given element of Ẑ.
First we mention that obviously A∗ = Q∗ F ∗ M̂ ∗ due to the definition A =
M̂ F Q.
Theorem 9.1. We have

M ∗ = Γ −1 Q, M̂ ∗ = Γ −1 Q̂,
(16)
Q∗ = M Γ, Q̂∗ = M̂ Γ.

Proof: The proof is given in Appendix A.

Next, we want to find an expression for F ∗ v, v ∈ U , where F ∗ is the adjoint
of the operator F .
Theorem 9.2. Let z be the solution of the adjoint symmetric hyperbolic system
3
∂z ∂z
−Γ (x) − Di + Φ(x)z = Γ (x)v(x, t), (17)
∂t i=1
∂xi

z(x, T ) = 0, (18)
z(x, t) = 0 on ∂Ω × [0, T ]. (19)
Then
F ∗ v = Γ −1 (x)z(x, t). (20)
Proof: The proof is given in Appendix B.
Remark 9.1. This procedural characterization of the adjoint operator is often
used in solution strategies of large scale inverse problems, where it naturally leads
to so-called ‘backpropagation strategies’. See for example [13, 14, 25, 27, 36, 48]
and the references given there.
150 O. Dorn

Remark 9.2. Notice that in the adjoint system (17)–(19) ‘final value conditions’
are given at t = T in contrast to (1)–(4) where ‘initial value conditions’ are
prescribed at t = 0. This corresponds to the fact that time is running backward
in (17)–(19) and forward in (1)–(4).

10 The Acoustic Time-Reversal Mirror

In the following we want to define an operator Sa such that F ∗ = Γ −1 Sa F Sa Γ

holds. We will call this operator Sa the acoustic time-reversal operator. We will
also define the acoustic time-reversal mirrors Ta and T̂a , which act on the signals
instead of the sources or fields.
We consider the acoustic system
∂vf
ρ + gradpf (x, t) = qv , (21)
∂t
∂pf
κ + divvf (x, t) = qp , (22)
∂t
vf (x, 0) = 0, pf (x, 0) = 0 in Ω (23)
with t ∈ [0, T ] and zero boundary conditions at ∂Ω ×[0, T ]. We want to calculate
the action of the adjoint operator F ∗ on a vector (φ, ψ)T ∈ U .

Theorem 10.1. Let (φ, ψ)T ∈ U and let (va , pa )T be the solution of the adjoint
system
∂va
−ρ − gradpa (x, t) = ρ(x)φ(x, t) (24)
∂t
∂pa
−κ − divva (x, t) = κ(x)ψ(x, t) (25)
∂t
va (x, T ) = 0, pa (x, T ) = 0 in Ω, (26)
with t ∈ [0, T ] and zero boundary conditions at ∂Ω × [0, T ]. Then we have
   −1 
∗ φ ρ va (x, t)
F = . (27)
ψ κ−1 pa (x, t)

Proof: This theorem is just an application of Theorem 9.2 to the acoustic sym-
metric hyperbolic system. For the convenience of the reader, we will give a direct
proof as well in Appendix C.

Definition 10.1. We define the acoustic time-reversal operator Sa by putting

for all q = (qv , qp )T ∈ U (and similarly for all u = (v, p)T ∈ U )
 
  −qv (x, 2T − t)
Sa q (x, t) = (28)
qp (x, 2T − t)
 Time-Reversal and the Adjoint Imaging Method in Communication 151

Definition 10.2. We define the acoustic time-reversal mirrors Ta and T̂a by

putting for all r = (rv , rp )T ∈ Z and all s = (sv , sp )T ∈ Ẑ
   
  −rv (2T − t)   −sv (2T − t)
Ta r (t) = , T̂a s (t) = (29)
rp (2T − t) sp (2T − t)
The following lemma is easy to verify.
Lemma 10.1. We have the following commutations

M Sa = Ta M, Sa Q = QTa ,
(30)
M̂ Sa = T̂a M̂ , Sa Q̂ = Q̂T̂a .

Theorem 10.2. For (φ, ψ)T ∈ U we have

   
−1 φ ∗ φ
Γ Sa F Sa Γ = F (31)
ψ ψ
Proof: For the proof it is convenient to make the following definition.
Definition 10.3. Acoustic time-reversal experiment: For given (φ, ψ)T ∈
U define qv (x, t) = ρφ and qp (x, t) = κψ and perform the following physical
experiment
∂
ρ vtr (x, s) + gradptr (x, s) = −qv (x, 2T − s) (32)
∂s
∂
κ ptr (x, s) + divvtr (x, s) = qp (x, 2T − s) (33)
∂s
vtr (x, T ) = 0, ptr (x, T ) = 0 in Ω, (34)
with s ∈ [T, 2T ] and zero boundary conditions. Doing this experiment means to
process the data in the following way: Time-reverse all data (φ, ψ)T according to
t → 2T − s, t ∈ [0, T ], and, in addition, reverse the directions of the velocities
φ → −φ. We call this experiment the ‘acoustic time-reversal experiment’. Notice
that the time is running forward in this experiment.
The solution (vtr , ptr )T of this experiment can obviously be represented by
   
vtr φ
= F Sa Γ . (35)
ptr ψ
In order to show that the so defined time-reversal experiment is correctly mod-
elled by the adjoint system derived above, we make the following change in
variables:
τ = 2T − s, v̂tr = −vtr , p̂tr = ptr , (36)
which just corresponds to the application of the operator Sa to (vtr , ptr )T . We
have τ ∈ [0, T ]. In these variables the time-reversal system (32)–(34) gets the
form
152 O. Dorn

∂
−ρ v̂tr (x, τ ) − gradp̂tr (x, τ ) = qv (x, τ ) (37)
∂τ
∂
−κ p̂tr (x, τ ) − divv̂tr (x, τ ) = qp (x, τ ) (38)
∂τ
v̂tr (x, T ) = 0, p̂tr (x, T ) = 0 in Ω. (39)
Taking into account the definition of qv and qp , we see that
   
v̂tr va
=
p̂tr pa

where va and pa solve the adjoint system (24)–(26). Therefore, according to

Theorem 10.1:    
v̂tr φ
= ΓF∗ .
p̂tr ψ
Since we have with (35), (36) also
   
v̂tr φ
= Sa F Sa Γ ,
p̂tr ψ
the theorem is proven.

11 The Electromagnetic Time-Reversal Mirror

In the following we want to define an operator Se such that F ∗ = Γ −1 Se F Se Γ
holds. We will call this operator Se the electromagnetic time-reversal operator.
We will also define the electromagnetic time-reversal mirrors Te and T̂e , which
act on the signals instead of the sources or fields.
We consider Maxwell’s equations
∂Ef
 − curl Hf + σEf = qE (40)
∂t
∂Hf
µ + curl Ef = qH (41)
∂t
Ef (x, 0) = 0, Hf (x, 0) = 0 (42)
with t ∈ [0, T ] and zero boundary conditions. We want to calculate the action of
the adjoint operator F ∗ on a vector (φ, ψ)T ∈ U .

Theorem 11.1. Let (φ, ψ)T ∈ U and let (Ea , Ha )T be the solution of the adjoint
system
∂Ea
− + curlHa (x, t) + σEa = (x)φ(x, t) (43)
∂t
∂Ha
−µ − curlEa (x, t) = µ(x)ψ(x, t) (44)
∂t
 Time-Reversal and the Adjoint Imaging Method in Communication 153

Ea (x, T ) = 0, Ha (x, T ) = 0 in Ω (45)

with t ∈ [0, T ] and zero boundary conditions. Then we have
   −1 
∗ φ  Ea (x, t)
F = . (46)
ψ µ−1 Ha (x, t)
Proof: This theorem is just an application of Theorem 9.2 to the electro-
magnetic symmetric hyperbolic system. Again, we will give a direct proof in
Appendix D as well.
Definition 11.1. We define the electromagnetic time-reversal operator Se
by putting for all q = (qE , qH )T ∈ U (and similarly for all u = (E, H)T ∈ U )
 
  −qE (x, 2T − t)
Se q (x, t) = (47)
qH (x, 2T − t)
Definition 11.2. We define the electromagnetic time-reversal mirrors Te
and T̂e by putting for all r = (rE , rH )T ∈ Z and for all s = (sE , sH )T ∈ Ẑ
   
  −rE (2T − t)   −sE (2T − t)
Te r (t) = , T̂e s (t) = (48)
rH (2T − t) sH (2T − t)
The following lemma is easy to verify.
Lemma 11.1. We have the following commutations

M Se = Te M, Se Q = QTe ,
(49)
M̂ Se = T̂e M̂ , Se Q̂ = Q̂T̂e .

Theorem 11.2. For (φ, ψ)T ∈ U we have

   
−1 φ ∗ φ
Γ Se F Se Γ = F (50)
ψ ψ
Proof: For the proof it is convenient to make the following definition.
Definition 11.3. Electromagnetic time-reversal experiment: For a given
vector (φ, ψ)T ∈ U define qE (x, t) = φ and qH (x, t) = µψ and perform the
physical experiment
∂
 Etr (x, s) − curlHtr (x, s) + σEtr (x, s) = −qE (x, 2T − s) (51)
∂s
∂
µ Htr (x, s) + curlEtr (x, s) = qH (x, 2T − s) (52)
∂s
Etr (x, T ) = 0, Htr (x, T ) = 0 in Ω, (53)
with s ∈ [T, 2T ] and zero boundary conditions. Doing this experiment means
to process the data in the following way: Time-reverse all data according to
t → 2T − s, t ∈ [0, T ], and, in addition, reverse the directions of the electric
field component by φ → −φ. We call this experiment the ‘electromagnetic time-
reversal experiment’. Notice that the time is running forward in this experiment.
154 O. Dorn

The solution (Etr , Htr )T of this experiment can obviously be represented by

   
Etr φ
= F Se Γ . (54)
Htr ψ
In order to show that the so defined time-reversal experiment is correctly mod-
elled by the adjoint system derived above, we make the following change in
variables:
τ = 2T − s, Êtr = −Etr , Ĥtr = Htr , (55)
which just corresponds to the application of the operator Se to (Etr , Htr )T . We
have τ ∈ [0, T ]. In these variables the time-reversal system (51)–(53) gets the
form
∂
− Êtr (x, τ ) + curlĤtr (x, τ ) + σ Êtr (x, s) = qE (x, τ ), (56)
∂τ
∂
−µ Ĥtr (x, τ ) − curlÊtr (x, τ ) = qH (x, τ ) (57)
∂τ
Êtr (x, T ) = 0, Ĥtr (x, T ) = 0 in Ω. (58)
Taking into account the definition of qE and qH , we see that
  
Êtr Ea
=
Ĥtr Ha

where Ee and He solve the adjoint system (43)–(45). Therefore, according to

Theorem 11.1:   
Êtr ∗ φ
= ΓF .
Ĥtr ψ
Since we have with (54), (55) also
  
Êtr φ
= Se F Se Γ ,
Ĥtr ψ

the theorem is proven.

Remark 11.1. For electromagnetic waves, there is formally an alternative way to

define the electromagnetic time-reversal operator, namely putting for all q =
(qE , qH )T ∈ U  
  qE (x, 2T − t)
Se q (x, t) = ,
−qH (x, 2T − t)
accompanied by the analogous definitions for the electromagnetic time-reversal
mirrors. With these alternative definitions, Theorem 11.2 holds true as well, with
only very few changes in the proof. Which form to use depends mainly on the
preferred form for modelling applied antenna signals in the given antenna system.
The first formulation directly works with applied electric currents, whereas the
second form is useful for example for magnetic dipole sources.
 Time-Reversal and the Adjoint Imaging Method in Communication 155

12 Time-Reversal and the Adjoint Operator A∗

Define S = Sa , T = Ta , T̂ = T̂a for the acoustic case, and S = Se , T = Te ,

T̂ = T̂e for the electromagnetic case. We call S the time-reversal operator and
T , T̂ the time-reversal mirrors. We combine the results of Lemma 10.1 and
Lemma 11.1 into the following lemma.
Lemma 12.1. We have the following commutations

M S = T M, SQ = QT ,
(59)
M̂ S = T̂ M̂ , S Q̂ = Q̂T̂ .

Moreover, combining Theorem 10.2 and Theorem 11.2 we get

Theorem 12.1. For (φ, ψ)T ∈ U we have
   
−1 φ ∗ φ
Γ SF SΓ = F . (60)
ψ ψ

With this, we can prove the following theorem which provides the fundamental
link between time-reversal and inverse problems.
Theorem 12.2. We have
A∗ = T B T̂ . (61)
Proof: Recall that the adjoint operator A∗ can be decomposed as A∗ =
Q∗ F ∗ M̂ ∗ . With Theorem 9.1, Theorem 12.1, and Lemma 12.1, it follows there-
fore that

A∗ = M Γ Γ −1 SF SΓ Γ −1 Q̂
= M SF S Q̂
= T M F Q̂T̂
= T B T̂ ,

which proves the theorem.

Remark 12.1. The above theorem provides a direct link between the adjoint op-
erator A∗ , which plays a central role in the theory of inverse problems, and a
physical experiment modelled by B. The expression T B T̂ defines a ‘time-reversal
experiment’. We will demonstrate in the following sections how we can make use
of this relationship in order to solve the inverse problem of communication by a
series of physical time-reversal experiments.

Remark 12.2. We mention that the above results hold as well for elastic waves
with a suitable definition of the elastic time-reversal mirrors. We leave out the
details for brevity.
156 O. Dorn

13 Iterative Time-Reversal for the Gradient Method

13.1 The Basic Version

The results achieved above give rise to the following experimental procedure for
applying the gradient method (15) to the inverse problem of communication as
formulated in Sects. 7 and 8. First, the pilot signal s̃(t) is defined by user U1 as
described in (11). Moreover, we assume that the first guess r(0) (t) at the base
station is chosen to be zero. Then, using Theorem 12.2, we can write the gradient
method (15) in the equivalent form

r(0) = 0
(62)
r(n+1) = r(n) + β (n) T B T̂ (s̃ − Ar(n) ),

or, expanding it,

r(0) = 0
s(n) = s̃ − Ar(n) (63)
r(n+1) = r(n) + β (n) T B T̂ s(n)
In a more detailed form, we arrive at the following experimental procedure for
implementing the gradient method, where we fix in this description β (n) = 1 for
all iteration numbers n for simplicity:
1. The user U1 chooses a pilot signal α(t) to be used for communicating with the
base station. The objective signal at all users is then s̃(t) = (α(t), 0, . . . , 0)T .
The initial guess r(0) (t) at the base station is defined to be zero, such that
s(0) = s̃.
2. The user U1 initiates the communication by sending the time-reversed pilot
signal into the environment. This signal is T̂ s(0) (t). All other users are quiet.
3. The base station receives the pilot signal as B T̂ s(0) (t). It time-reverses this
signal and sends this time-reversed form, namely r(1) (t) = T B T̂ s(0) (t), back
into the medium.
4. The new signal arrives at all users as Ar(1) (t). All users compare the received
signals with their components of the objective signal s̃(t). They take the
difference s(1) (t) = s̃(t) − Ar(1) (t), and time-reverse it. They send this new
signal T̂ s(1) (t) back into the medium.
5. The base station receives this new signal, time-reverses it, adds it to the
previous signal r(1) (t), and sends the sum back into the medium as r(2) (t).
6. This iteration is continued until all users are satisfied with the match between
the received signal Ar(n) (t) and the objective signal s̃(t) at some iteration
number n. Alternatively, a fixed iteration number n can be specified a-priori
for stopping the iteration.
Needless to say that, in practical implementations, the laboratory time needs to
be reset to zero after each time-reversal step.
 Time-Reversal and the Adjoint Imaging Method in Communication 157

Remark 13.1. The experimental procedure which is described above is practi-

cally equivalent to the experimental procedure which was suggested and experi-
mentally verified in [34, 35]. Therefore, our basic scheme provides an alternative
derivation and interpretation of this experimental procedure.

Remark 13.2. We mention that several refinements of this scheme are possible
and straightforward. For example, a weighted inner product can be introduced
for the user signal space Ẑ which puts different preferences on the satisfaction of
the user objectives during the iterative optimization process. For example, if the
‘importance’ of suppressing interferences with other users is valued higher than
to get an optimal signal quality at the specified user U1 , a higher weight can be
put into the inner product at those users which did not start the communication
process. A user who does not care about these interferences, simply puts a very
small weight into his component of the inner product of Ẑ.

Remark 13.3. Notice that there is no mechanism directly built into this proce-
dure which prevents the energy emitted by the base antenna to increase more
than the communication system can support. For example, if the subspace of
signals
Z0 := {r(t) : Ar = 0}
is not empty, then it might happen that during the iteration described above (e.g.
due to noise) an increasing amount of energy is put into signals emitted by the
base station which are in this subspace and which all produce zero contributions
to the measurements at all users. More generally, elements of the subspace of
signals
Zε := {r(t) : ArẐ < εrZ },
for a very small threshold 0 <  << 1, might cause problems during the iteration
if the pilot signal s̃(t) chosen by the user has contributions in the subspace
AZε (i.e. in the space of all s = Ar with r ∈ Zε ). This is so because in the
effort of decreasing the mismatch between Ar and s̃(t), the base antenna might
need to put signals with high energy into the system in order to get only small
improvements in the signal match at the user side. Since the environment (and
therefore the operator A) is unknown a-priori, it is difficult to avoid the existence
of such contributions in the pilot signal.

One possible way to prevent the energy emitted by the base station to increase
artificially would be to project the signals r(n) (t) onto the orthogonal comple-
ments of the subspaces Z0 or Zε (if they are known or can be constructed by
some means) prior to their emission. Alternatively, the iteration can be stopped
at an early stage before these unwanted contributions start to build up. (This
in fact has been suggested in [34, 35]).
In the following subsection we introduce an alternative way of ensuring that
the energy emitted by the base station stays reasonably bounded in the effort of
fitting the pilot signal at the users.
158 O. Dorn

13.2 The Regularized Version

Consider the regularized problem

 
rLSr = Minr Ar − s̃2Ẑ + λr2Z (64)

with some suitably chosen regularization parameter λ > 0. In this problem

formulation a trade-off is sought between a signal fit at the user side and a
minimized energy emission at the base station. The trade-off parameter is the
regularization parameter λ. Instead of (14) we need to consider now

1  λ 
J˜(r) = Ar − s̃ , Ar − s̃ + r, r (65)
2 Ẑ 2 Z

The negative gradient direction is now given by −A∗ (Ar − s̃) − λr, such that the
regularized iteration reads:

r(0) = 0
  (66)
r(n+1) = r(n) + β (n) T B T̂ (s̃ − Ar(n) ) − λr(n) ,

where we have replaced A∗ by T B T̂ . The time-reversal iteration can be expanded

into the following practical scheme

r(0) = 0
s(n) = s̃ − Ar(n) (67)
r(n+1) = r(n) + β (n) T B T̂ s(n) − β (n) λr(n) .

Comparing with (63), we see that the adaptations which need to be applied in
the practical implementation for stabilizing the basic algorithm can easily be
done.

14 The Minimum Norm Solution Approach

14.1 The Basic Version

In this section we want to propose an alternative scheme for solving the inverse
problem of communication. As mentioned above, a major drawback of the ba-
sic approach (13) is that the energy emitted by the base station is not limited
explicitly when solving the optimization problem. The regularized version pre-
sented above alleviates this problem. However, we want to mention here that,
under certain assumptions, there is an alternative scheme which can be em-
ployed instead and which has an energy constraint directly built in. Under the
formal assumption that there exists at least one (and presumably more than one)
 Time-Reversal and the Adjoint Imaging Method in Communication 159

solution of the inverse problem at hand (i.e. the ‘formally underdetermined

case’), we can look for the minimum norm solution

Minr rZ subject to Ar = s̃. (68)

In Hilbert spaces this solution has an explicit form. It is

rM N = A∗ (AA∗ )−1 s̃. (69)

Here, the operator (AA∗ )−1 acts as a filter on the pilot signal s̃. Instead of
sending the pilot signal to the base station, the users send the filtered version
of it. Certainly, a method must be found in order to apply the filter (AA∗ )−1
to the pilot signal. One possibility of doing so would be to try to determine the
operator AA∗ explicitly by a series of time-reversal experiments on some set of
basis functions of Ẑ, and then invert this operator numerically. However, this
might not be practical in many situations. It certainly would be slow and it
would involve a significant amount of signal-processing. Therefore, we propose
an alternative procedure. First, we notice that there is no need to determine the
whole operator (AA∗ )−1 , but that we only have to apply it to one specific signal,
namely s̃. Let us introduce the short notation

C = AA∗ .

In this notation, we are looking for a signal ŝ ∈ Ẑ such that Cŝ = s̃. We propose
to solve this equation in the least squares sense:

ŝ = Mins Cs − s̃2Ẑ . (70)

Moreover, as a suitable method for practically finding this solution, we want

to use the gradient method. Starting with the initial guess s0 = 0, the gradient
method reads
s(0) = 0
(71)
s(n+1) = s(n) − β (n) C ∗ (Cs(n) − s̃),

where C ∗ is the adjoint operator to C and β (n) is again some step-size. Expanding
this expression, and taking into account C ∗ = C and A∗ = T B T̂ , we arrive at

s(0) = 0
  (72)
s(n+1) = s(n) + β (n) AT B T̂ s̃ − AT B T̂ s(n) .

In the practical implementation, we arrive at the following iterative scheme:

160 O. Dorn

Initialize gradient iteration: s(0) = 0

for n = 0, 1, 2, . . . do:
1
r(n+ 2 ) = T B T̂ s(n)
1 1
s(n+ 2 ) = s̃ − Ar(n+ 2 )
1
(73)
r(n+1) = T B T̂ s(n+ 2 )
s(n+1) = s(n) + β (n) Ar(n+1)
Terminate iteration at step n̂: ŝ = s(n̂)
Final result: rM N = T B T̂ ŝ.

This iteration can be implemented by a series of time-reversal experiments, with-

out the need of heavy signal-processing. The final step of the above algorithm
amounts to applying A∗ to the result of the gradient iteration for calculating
ŝ = (AA∗ )−1 s̃, which yields then rM N . This will then be the signal to be applied
by the base station during the communication process with the user U1 .

14.2 The Regularized Version

In some situations it might be expected that the operator C is ill-conditioned,

such that its inversion might cause instabilities, in particular when noisy signals
are involved. For those situations, a regularized form of the minimum norm
solution is available, namely

rM N r = A∗ (AA∗ + λIẐ )−1 s̃ (74)

where IẐ denotes the identity operator in Ẑ and λ > 0 is some suitably chosen
regularization parameter. The necessary adjustments in the gradient iteration
for applying (AA∗ + λIẐ )−1 to s̃ are easily done. We only mention here the
resulting procedure for the implementation of this gradient method by a series
of time-reversal experiments:

Initialize gradient iteration: s(0) = 0

for n = 0, 1, 2, . . . do:
1
r(n+ 2 ) = T B T̂ s(n)
1 1
s(n+ 2 ) = s̃ − Ar(n+ 2 ) − λs(n)
1
(75)
r(n+1) = T B T̂ s(n+ 2 )
1
s(n+1) = s(n) + β (n) Ar(n+1) + β (n) λs(n+ 2 ) .
Terminate iteration at step n̂: ŝ = s(n̂)
Final result: rM N r = T B T̂ ŝ.
 Time-Reversal and the Adjoint Imaging Method in Communication 161

Again, the last step shown above is a final application of A∗ to the result of
the gradient iteration for calculating ŝ = (AA∗ + λIẐ )−1 s̃, which yields then
rM N r . This will then be the signal to be applied by the base station during the
communication process with the user U1 .

15 The Regularized Least Squares Solution Revisited

We have introduced above the regularized least squares solution of the inverse
problem of communication, namely
 
rLSr = Minr Ar − s̃2Ẑ + λr2Z (76)

with λ > 0 being the regularization parameter. In Hilbert spaces, the solution
of (76) has an explicit form. It is

rLSr = (A∗ A + λIZ )−1 A∗ s̃, (77)

where IZ is the identity operator in Z. It is therefore tempting to try to imple-

ment also this direct form as a series of time-reversal experiments and compare
its performance with the gradient method as it was described above. As our last
strategy which we present in these lecture notes we want to show that such an
alternative direct implementation of (76) is in fact possible.
Notice that in (76) the filtering operator (A∗ A+λIZ )−1 is applied at the base
station, in contrast to the previous case where the user signal was filtered by
the operator (AA∗ + λIẐ )−1 . Analogously to the previous case, we need to find
a practical way to apply this filter to a signal at the base station. We propose
again to solve the equation

(A∗ A + λIZ )r̂ = r̃ (78)

in the least squares sense, where r̃ = A∗ s̃. Defining

C = A∗ A + λIZ ,

and using C ∗ = C and A∗ = T B T̂ , we arrive at the following gradient iteration

for solving problem (78):

r(0) = 0
  (79)
r(n+1) = r(n) + β (n) (T B T̂ A + λIZ ) r̃ − (T B T̂ A + λIZ )r(n) .

This gives rise to the following practical implementation by a series of time-

reversal experiments:
162 O. Dorn

User sends pilot signal to base station: r̃ = T T̂ s̃

Initialize gradient iteration: r(0) = 0
for n = 0, 1, 2, . . . do:
1
s(n+ 2 ) = Ar(n)
1 1
r(n+ 2 ) = r̃ − T B T̂ s(n+ 2 ) − λr(n)
1
s(n+1) = Ar(n+ 2 )
1
r(n+1) = r(n) + β (n) T B T̂ s(n+1) + β (n) λr(n+ 2 ) .
Terminate iteration at step n̂: r̂ = r(n̂)
Final result: rLSr = r̂.
(80)
rLSr will then be the signal to be applied by the base station during the com-
munication process with the user U1 .

16 Partial and Generalized Measurements

In many practical applications, only partial measurements of the whole wave-field
are available. For example, in ocean acoustics often only pressure is measured,
whereas the velocity field is not part of the measurement process. Similarly, in
wireless communication only one or two components of the electric field might
be measured simultaneously, but the remaining electric components and all mag-
netic components are missing. We want to demonstrate in this section that all
results presented above are valid also in this situation of partial measurements,
with the suitable adaptations.
Mathematically, the measurement operator needs to be adapted for the sit-
uation of partial measurements. Let us concentrate here on the special situation
that only one component uν (ν ∈ {1, 2, 3, ...}) of the incoming wave field u is
measured by the users and the base station. All other possible situations will
then just be combinations of this particular case. It might also occur the situa-
tion that users can measure a different partial set of components than the base
station. That case also follows directly from this canonical situation.
We introduce the new signal space at the base station Y = (L2 [0, T ])K and
the corresponding ‘signal projection operator’ Pν by putting

Pν : Z → Y, Pν (r1 (t), . . . , rK (t))T = rν (t).

We see immediately that its adjoint Pν∗ is given by

Pν∗ : Y → Z, Pν∗ rν (t) = (0, . . . , 0, rν (t), 0, . . . , 0)T

where rν (t) appears on the right hand side at the ν-th position. Our new mea-
surement operator Mν , and the new source operator Qν , are then defined by
 Time-Reversal and the Adjoint Imaging Method in Communication 163

Mν : U → Y, Mν u = Pν M u
(81)
Qν : Y → U, Qν rν = QPν∗ rν .

Analogous definitions are done for Ŷ , P̂ν , M̂ν and Q̂ν at the users.
Obviously, we will have to replace now in the above derivation of the iter-
ative time-reversal procedure all measurement operators M by Mν (and M̂ by
M̂ν ) and all source operators Q by Qν (and Q̂ by Q̂ν ). In particular, the new
‘communication operators’ are now given by

Aν : Y → Ŷ , Aν rν = M̂ν F Qν rν ,
(82)
Bν : Ŷ → Y, Bν sν = Mν F Q̂ν sν .

In the following two theorems we show that the main results presented so far
carry over to these newly defined operators.

Theorem 16.1. We have

Mν∗ = Γ −1 Qν , M̂ν∗ = Γ −1 Q̂ν ,

(83)
Q∗ν = Mν Γ, Q̂∗ν = M̂ν Γ.

Proof: The proof is an easy exercise using (81) and Theorem 9.1.

Theorem 16.2. It is
A∗ν = T Bν T̂ . (84)

Proof: The proof is now identical to the proof of Theorem 12.2, using
Theorem 16.1 instead of Theorem 9.1.

Remark 16.1. In fact, it is easy to verify that all results of these notes remain
valid for arbitrarily defined linear measurement operators

MU : U → ZU , MA : U → ZA ,

where ZU and ZA are any meaningful signal spaces at the users and the base
antennas, respectively. The only requirement is that it is experimentally possible
to apply signals according to the source operators defined by

QU : ZU → U, QU = Γ MU∗
∗
QA : ZA → U, QA = Γ MA

where MU∗ and MA ∗

are the formal adjoint operators to MU and MA with respect
to the chosen signal spaces ZU and ZA . In addition, the measurement and source
operators as defined above are required to satisfy the commutation relations
as stated in Lemma 12.1. Under these assumptions, we define the generalized
communication operators à and B̃ by
164 O. Dorn

à : ZA → ZU , Ãr = MU F QA r,
B̃ : ZU → ZA , B̃s = MA F QU s.

Now the proof to Theorem 12.2 directly carries over to this generalized situation,
such that we have also here
Ã∗ = T B̃ T̂ .
This yields iterative time-reversal schemes completely analogous to those pre-
sented above.

17 Summary and Future Research Directions

We have derived in these lecture notes a direct link between the time-reversal
technique and the adjoint method for imaging. Using this relationship, we have
constructed several iterative time-reversal schemes for solving an inverse prob-
lem which arises in ocean acoustic and wireless communication. Each of these
schemes can be realized physically as a series of time-reversal experiments, with-
out the use of heavy signal processing or computations. One of the schemes which
we have derived (and which we call the ‘basic scheme’), is practically equivalent
to a technique introduced earlier in [34, 35] using different tools. Therefore, we
have given an alternative theoretical derivation of that technique, with a differ-
ent mathematical interpretation. The other schemes which we have introduced
are new in this application. They represent either generalizations of the basic
scheme, or alternatives which follow different objectives.
Many questions related to these and similar iterative time-reversal ap-
proaches for telecommunication are still open. The experimental implementa-
tion has been investigated so far only for one of these techniques in [34, 35], for
the situation of underwater sound propagation. A thorough experimental (or,
alternatively, numerical) verification of the other schemes is necessary for their
practical evaluation. An interesting and practically important problem is the
derivation of quantitative estimates for the expected focusing quality of each of
these schemes, for example following the ideas of the work performed for a sin-
gle step in [5]. Certainly, it is expected that these estimates will again strongly
depend on the randomness of the medium, on the geometry and distribution
of users and base antennas, and on technical constraints as for example partial
measurements. Also, different types of noise in the communication system need
to be taken into account. The performance in a time-varying environment is an-
other interesting issue of practical importance. All schemes presented here can be
adapted in principle to a dynamic environment by re-adjusting the constructed
optimal signals periodically. Practical ways of doing so need to be explored.

Acknowledgments

These lecture notes have been written during the stay of the author at L2S-
Supélec, France, in the spring of 2004, and are inspired by many discussions
 Time-Reversal and the Adjoint Imaging Method in Communication 165

with G. Papanicolaou on time reversal and imaging during the Mathematical

Geophysics Summer Schools at Stanford University in 1998–2002 and during
the CIME program on ‘imaging’ in Martina Franca in the fall of 2002. The
author thanks Dominique Lesselier for making the research stay at Supélec pos-
sible. He thanks George Papanicolaou, Hongkai Zhao and Knut Solna for many
exciting and useful discussions on time-reversal. He thanks Mathias Fink, Claire
Prada, Gabriel Montaldo, Francois Vignon, and the group at Laboratoire On-
des et Acoustique in Paris for pointing out to him their recent work on itera-
tive communication schemes, which stimulated part of the work presented here.
The author also thanks Bill Kuperman and the group at Scripps Institution of
Oceanography in San Diego for useful discussions. Financial support from the
Office of Naval Research (under grant N00014-02-1-0090) and the CNRS (under
grant No. 8011618) is gratefully acknowledged.

Appendix A: Proof of Theorem 9.1

For arbitrary r ∈ Z and u ∈ U we have
K


M u, rZ = γk (x)u(x, t)dx , r(t) dt
k=1 [0,T ] Ω N


. K
/
−1
= Γ (x)u(x, t) , Γ (x) γk (x)r(t) dxdt.
[0,T ] Ω k=1 N

Therefore,
M u, rZ = u, M ∗ rU
with
K
M ∗ r = Γ −1 (x) γk (x)r(t) = Γ −1 (x)Qr.
k=1

Doing the analogous calculation for M̂ ∗ we get

M ∗ = Γ −1 Q, M̂ ∗ = Γ −1 Q̂. (85)

Taking the adjoint of (85) we see that

Q∗ = M Γ, Q̂∗ = M̂ Γ

holds as well. This proves the theorem.

Appendix B: Proof of Theorem 9.2

We have the following version of Green’s formula
166 O. Dorn


. 3
/
∂u ∂u i
Γ (x) + D + Φ(x)u , z dxdt (86)
[0,T ] Ω ∂t i=1
∂xi
N





+ Γ (x)u(x, t), v(x, t)N dxdt + q(x, t) , z(x, t)N dxdt
[0,T ] Ω [0,T ] Ω


. 3
/
∂z ∂z
= u , −Γ (x) − D + Φ(x)z i
dxdt
[0,T ] Ω ∂t i=1
∂xi
N





+ Γ (x)u(x, t), v(x, t)N dxdt + q(x, t) , z(x, t)N dxdt
[0,T ] Ω [0,T ] Ω



+ Γ (x)u(x, T ), z(x, T )N dx − Γ (x)u(x, 0), z(x, 0)N dx
Ω Ω


3
, -
+ Di u, z N νi (x) dσdt,
[0,T ] ∂Ω i=1

where n(x) = (ν1 (x), ν2 (x), ν3 (x)) is the outward normal at ∂Ω in the point x.
Notice that we have augmented Green’s formula in (86) by some terms which
appear in identical form on the left hand side and on the right hand side.
We will assume here that the boundary is far away from the sources and
receivers and that no energy enters Ω from the outside, such that during the
time interval of interest [0, T ] all fields along this boundary are identically zero.
This is expressed by the boundary conditions given in (4) and (19). Let u(x, t) be
a solution of (1), (2), (4), and z(x, t) a solution of (17)–(19). Then the first term
on the left hand side of (86) and the third term on the right hand side cancel
each other because of (1). The second term on the left hand side and the first
term on the right hand side cancel each other because of (17). The (t = T )-term
and the (t = 0) term vanish due to (18) and (2), respectively, and the boundary
integral vanishes because of (4) and (19). The remaining terms (i.e. the third
term on the left hand side and the second term on the right hand side) can be
written as
F q, vU = q, F ∗ vU ,
with F ∗ v = Γ −1 (x)z(x, t) as defined in (20).

Appendix C: Direct Proof of Theorem 10.1

We prove the theorem by using Greens formula:

T

∂vf ∂pf
ρ va + gradpf va + κ pa + divvf pa dxdt (87)
0 Ω ∂t ∂t

T

T

+ [ρvf φ + κpf ψ] dxdt + [qv va + qp pa ] dxdt
0 Ω 0 Ω
 Time-Reversal and the Adjoint Imaging Method in Communication 167

T

∂va ∂pa
= − pf divva − pf κ
−vf ρ − vf gradpa dxdt
0 Ω ∂t ∂t

T

T

+ [ρvf φ + κpf ψ] dxdt + [qv va + qp pa ] dxdt
0 Ω 0 Ω

T

T

+ (va · n)pf dσdt + (vf · n)pa dσdt
0 ∂Ω 0 ∂Ω



" # " #
+ ρ (vf va )(x, T )−(vf va )(x, 0) dx + κ (pf pa )(x, T )−(pf pa )(x, 0) dx.
Ω Ω

This equation has the form (86). Notice that we have augmented Green’s formula
in (87), as already shown in (86), by some terms which appear in identical form
on the left hand side and on the right hand side.
The first term on the left hand side of (87) and the third term on the right
hand side cancel each other due to (21), (22). The second term on the left
hand side and the first term on the right hand side cancel each other because
of (24), (25). The (t = T )-terms and the (t = 0)-terms vanish due to (26),
(23), respectively, and the boundary terms vanish because of zero boundary
conditions. We are left over with the equation

T

T

[ρvf φ + κpf ψ] dxdt = [qv va + qp pa ] dxdt
0 Ω 0 Ω

Defining F ∗ by (27), this can be written as

        
qv φ qv φ
F , = , F∗ .
qp ψ U qp ψ U

Therefore, F ∗ is in fact the adjoint of F , and the theorem is proven.

Appendix D: Direct Proof of Theorem 11.1

We prove the theorem by using Green’s formula:

T

∂Ef ∂Hf
 Ea − curlHf Ea + σEf Ea + µ Ha + curlEf Ha dxdt
0 Ω ∂t ∂t

T

T

+ [Ef φ + µHf ψ] dxdt + [qE Ea + qH Ha ] dxdt
0 Ω 0 Ω

T

∂Ea ∂Ha
= −Ef  − Hf curlEa + Ef σEa − Hf µ + Ef curlHa dxdt
0 Ω ∂t ∂t

T

T

+ [Ef φ + µHf ψ] dxdt + [qE Ea + qH Ha ] dxdt
0 Ω 0 Ω
168 O. Dorn

T

T

+ Ea × Hf · n dσdt + Ef × Ha · n dσdt (88)
0 ∂Ω 0 ∂Ω



" # " #
+  (Ef Ea )(T ) − (Ef Ea )(0) dx + µ (Hf Ha )(T ) − (Hf Ha )(0) dx
Ω Ω
This equation has the form (86). Notice that we have augmented Green’s formula
in (88), as already shown in (86), by some terms which appear in identical form
on the left hand side and on the right hand side.
The first term on the left hand side of (88) and the third term on the right
hand side cancel each other because of (40) and (41). The second term on the
left hand side and the first term on the right hand side cancel each other because
of (43), (44). The (t = 0)-terms and the (t = T )-terms vanish due to (42) and
(45). The boundary terms vanish because of zero boundary conditions. We are
left over with the equation

T

T

[Ef φ + µHf ψ] dxdt = [qE Ea + qH Ha ] dxdt.
0 Ω 0 Ω
∗
Defining F by (46), this can be written as
        
qE φ qE φ
F , = , F∗ .
qH ψ U qH ψ U

Therefore, F ∗ is in fact the adjoint of F , and the theorem is proven.

References
1. Bal G and Ryzhik L 2003 Time reversal and refocusing in random media SIAM
J. Appl. Math. 63 1475–98
2. Bardos C and Fink M 2002 Mathematical foundations of the time reversal mirror
Asymptotic Analysis 29 157–182
3. Berryman J G, Borcea L, Papanicolaou G C and Tsogka C 2002 Statistically stable
ultrasonic imaging in random media Journal of the Acoustic Society of America
112 1509–1522
4. Bleistein N, Cohen J K, and Cohen J W 2001 Mathematics of multidimensional
seismic imaging, migration, and inversion (Springer: New York)
5. Blomgren P, Papanicolaou G, and Zhao H 2002 Super-Resolution in Time-Reversal
Acoustics, Journal of the Acoustical Society of America 111 238–248
6. Borcea L, Papanicolaou G and Tsogka C 2003 Theory and applications of time
reversal and interferometric imaging Inverse Problems 19 5139–5164
7. Cheney M, Isaacson D and Lassas M 2001 Optimal Acoustic Measurements SIAM
J Appl Math, 61 1628–1647
8. Claerbout J 1985 Fundamentals of geophysical data processing: with applications
to petroleum prospecting, (CA: Blackwell Scientific Publications, Palo Alto)
9. Collins M D and Kuperman W A 1991 Focalization: Environmental focusing and
source localization J. Acoust. Soc. Am. 90 (3) 1410–1422
10. Collins M D and Kuperman W A 1994 Inverse problems in ocean acoustics Inverse
Problems 10 1023–1040
 Time-Reversal and the Adjoint Imaging Method in Communication 169

11. Dautray R and Lions J L Mathematical Analysis and Numerical Methods for Sci-
ence and Technology Vol. 1–6, Springer.
12. Devaney A 2004 Superresolution processing of multi-static data using time reversal
and MUSIC to appear in J. Acoust. Soc. Am.
13. Dorn O 1998 A transport-backtransport method for optical tomography Inverse
Problems 14 1107–1130
14. Dorn O, Bertete-Aguirre H, Berryman J G and Papanicolaou G C 1999 A nonlin-
ear inversion method for 3D electromagnetic imaging using adjoint fields Inverse
Problems 15 1523-1558
15. Dowling D R and Jackson D R 1992 Narrow band performance of phase conjugate
arrays in dynamic random media J. Acoust. Soc. Am. 91 3257-3277
16. Dowling D R 1993 Acoustic pulse compression using passive phase-conjugate
processing J. Acoust. Soc. Am. 95 (3) 1450–1458
17. Draeger C and Fink M 1997 One-Channel Time Reversal of Elastic Waves in a
Chaotic 2D-Silicon Cavity Physical Review Letters 79 (3) 407–410
18. Edelmann G F, Akal T, Hodgkiss W S, Kim S, Kuperman W A and Chung H C
2002 An initial demonstration of underwater acoustic communication using time
reversal IEEE J. Ocean. Eng. 27 (3) 602–609
19. Emami S M, Hansen J, Kim A D, Papanicolaou G, Paulraj A J, Cheung D and
Prettie C 2004 Predicted time reversal performance in wireless communications
using channel measurements submitted
20. Fink M 1997 Time reversed acoustics Phys. Today 50 34-40
21. Fink M and Prada C 2001 Acoustic time-reversal mirrors Inverse Problems 17
R1–R38
22. Foschini G J 1996 Layered space-time architecture for wireless communications in
a fading environment Bell Labs Technical Journal 1 (2) 41–59
23. Fouque J P, Garnier J and Nachbin A 2004 Time reversal for dispersive waves in
random media to appear in SIAM J. Appl. Math.
24. Gesbert D, Shafi M, Shiu D, Smith P J and Naguib A 2003 From Theory to Prac-
tice: An Overview of MIMO Space-Time Coded Wireless Systems IEEE J. Select.
Areas Commun. 21 (3) (tutorial paper) 281–301
25. Haber E, Ascher U and Oldenburg D 2000 On optimization techniques for solving
nonlinear inverse problems Inverse Problems 16 1263–1280
26. Hodgkiss W S, Song H C, Kuperman W A, Akal T, Ferla C and Jackson D R 1999
A long range and variable focus phase conjugation experiment in shallow water
J. Acoust. Soc. Am. 105 1597–1604
27. Hursky P, Porter M B, Cornuelle B D, Hodgkiss W S, and Kuperman W A 2004
Adjoint modeling for acoustic inversion J. Acoust. Soc. Am. 115 (2) 607–619
28. Jackson D R and Dowling D R 1991 Phase conjugation in underwater acoustics J.
Acoust. Soc. Am. 89 (1) 171–181
29. Kim A D, Blomgren P and Papanicolaou G 2004 Spatial focusing and intersymbol
interference in time reversal submitted
30. Klibanov M V and Timonov A 2003 On the mathematical treatment of time re-
versal Inverse Problems 19 1299–1318
31. Kuperman W A, Hodgkiss W S, Song H C, Akal T, Ferla C, and Jackson D R
1998 Phase conjugation in the ocean: Experimental demonstration of an acoustic
time reversal mirror J. Acoust. Soc. Am. 103 25-40
32. Kyritsi P, Papanicolaou G, Eggers P and Oprea A 2004 MISO time reversal and
delay spread compression for FWA channels at 5GHz preprint
33. Marsden J E and Hughes T J R Mathematical Foundations of Elasticity (Prentice
Hall, 1968)
170 O. Dorn

34. Montaldo G, Lerosey G, Derode A, Tourin A, de Rosny J and Fink M 2004

Telecommunication in a disordered environment with iterative time reversal Waves
Random Media 14 287–302
35. Montaldo G, Tanter M and Fink M 2004 Real time inverse filter focusing through
iterative time reversal to appear in J. Acoust. Soc. Am. 115 (2)
36. Natterer F and Wübbeling F 1995 A propagation-backpropagation method for
ultrasound tomography Inverse Problems 11 1225-1232
37. Oestges C, Kim A D, Papanicolaou G and Paulraj A J 2004 Characterization
of space-time focusing in time reversal random fields submitted to IEEE Trans.
Antennas Prop.
38. Papanicolaou G, Ryzhik L and Solna K 2004 Statistical stability in time-reversal
SIAM J. Appl. Math. 64 1133–1155
39. Prada C, Manneville S, Spoliansky D and Fink M 1996 Decomposition of the time
reversal operator: Detection and selective focusing of two scatterers J. Acoust. Soc.
Am. 96 (4) 2067–2076
40. Prada C, Kerbrat E, Cassereau D and Fink M 2002 Time reversal techniques in
ultrasonic nondestructive testing of scattering media Inverse Problems 18 1761–
1773
41. Ryzhik L, Papanicolaou G and Keller J B 1996 Transport Equations for Elastic
Waves and Other Waves in Random Media Wave Motion 24 327–370
42. Simon H S, Moustakas A L, Stoytchev M and Safar H 2001 Com-
munication in a disordered world Physics Today 54 (9) (online at
http://www.physicstoday.org/pt/vol-54/iss-9/p38.html)
43. Snieder R K and Scales J A 1998 Time-reversed imaging as a diagnostic of wave
and particle chaos Phys. Rev. E 58 (5) 5668–5675
44. Smith K B, Abrantes A A M and Larraza A 2003 Examination of time-reversal
acoustics in shallow water and applications to non-coherent underwater communi-
cations J. Acoust. Soc. Am. 113 (6) 3095–110
45. Solna K 2002 Focusing of time-reversed reflections Waves Random Media 12 1–21
46. Song H, Kuperman W A, Hodgkiss W S, Akal T and Guerrini P 2003 Demon-
stration of a High-Frequency Acoustic Barrier With a Time-Reversal Mirror IEEE
J. Acoust. Eng. 28 (2) 246–249
47. Tsogka C and Papanicolaou G C 2002 Time reversal through a solid-liquid interface
and super-resolution Inverse Problems 18 1639-1657
48. Vögeler M 2003 Reconstruction of the three-dimensional refractive index in elec-
tromagnetic scattering by using a propagation-backpropagation method Inverse
Problems 19 739–753
49. Yang T C 2003 Temporal Resolutions of Time-Reversal and Passive-Phase Con-
jugation for Underwater Acoustic Communications IEEE J. Ocean. Eng. 28 (2)
229–245
50. Zhao H 2004 Analysis of the Response Matrix for an Extended Target SIAM
J. Appl. Math. 64 (3) 725–745
A Brief Review on Point Interactions

Gianfausto Dell’Antonio1 , Rodolfo Figari2 , and Alessandro Teta3

1
Dipartimento di Matematica, Università di Roma “La Sapienza”, Rome, Italy
and Laboratorio Interdisciplinare SISSA-ISAS, Trieste, Italy
[email protected]
2
Dipartimento di Scienze Fisiche, Universitá di Napoli, Naples, Italy and Istituto
Nazionale di Fisica Nucleare, Sezione di Napoli, Naples, Italy
[email protected]
3
Dipartimento di Matematica Pura e Applicata, Universitá di L’Aquila, via
Vetoio – loc. Coppito, L’Aquila, Italy
[email protected]

Summary. We review properties and applications of point interaction Hamiltonians.

This class of operators is first defined following a classical presentation and then gen-
eralized to cases in which some dynamical and/or geometrical parameters are varying
with time. We recall their relations with smooth short range potentials.

1 Introduction

In this lecture we shall review some basic facts on the so-called point interactions,
i.e. perturbations of the free Laplacian in Rd , d = 1, 2, 3, supported by a finite set
of points (for a comprehensive treatment we refer to the monograph [AGH-KH]).
At a formal level the operator can be written as
n
“ H = −∆ + αj δyj ” (1)
j=1

where αj ∈ R and yj ∈ Rd denote respectively the strength and the position of

the j-th point interaction.
We shall be concerned with selfadjoint realizations in L2 (Rd ) of the formal
expression (1).
Historically such kind of interactions, also called δ-interactions or zero-
range interactions or Fermi pseudo-potentials, have been introduced for the
Schrödinger equations in the early days of quantum mechanics.
Nevertheless we want to emphasize that they can be equally well introduced
in any evolution or stationary problem involving a selfadjoint perturbation of the
free Laplacian (or of a regular elliptic operator) in L2 (Rd ) and for hyperbolic
evolution equations.
172 G. Dell’Antonio et al.

In particular one can also consider applications to the heat and the wave
equations.
Physical motivations for the introduction of point interactions were given
by Fermi ([F]) in the analysis of scattering of slow neutrons from a target of
condensed matter.
In fact, for a sufficiently slow neutron, the wavelength is much larger than
the effective range of the nuclear force acting between the neutron and each
nucleus of the target and, on the other hand, such force is extremely intense.
As a consequence it appears reasonable to modelize the interaction between the
neutron and the nucleus by a zero-range potential placed at the position of the
nucleus.
As a first approximation one considers each nucleus in a fixed position yj and
this leads to the basic model Hamiltonian (1) and to the corresponding linear
evolution for the wave function of the neutrons solving the Schrödinger equation.
The main interest of point interactions consists in the fact that they define
simple but non trivial models of short range interactions which are, in fact,
explicitly solvable.
More precisely this means that detailed information about the spectrum and
the eigenfunctions of (1) are available and then all the physical relevant quantities
related to the specific problem analyzed can be explicitly computed.
This fact makes point interactions especially useful in applications where
they can be considered as a first step in a more detailed analysis.
More refined evolution models can be obtained considering the strength
and/or the position of the point interactions as given functions of time (lin-
ear non-autonomous evolution problems) or the strength as a given function of
the solution itself (non linear evolution problems).
From a mathematical point of view, the first step is to give a satisfactory
definition of the formal operator (1) as a selfadjoint operator in L2 (Rd ) corre-
sponding to the intuitive idea of point interactions.
One can start from the reasonable consideration that any rigorous counter-
part of (1) must satisfy Hu = −∆u for any u ∈ C0∞ (Rd \ {y1 , . . . , yn }).
This suggests to define the following restriction of the free Laplacian

Ĥ = −∆, D(Ĥ) = C0∞ (Rd \ {y1 , . . . , yn }) (2)

It is not hard to see that the operator (2) is symmetric but not selfadjoint
in L2 (Rd ); moreover one selfadjoint extension of (2) is trivial, i.e. it corresponds
to the free Laplacian H0 = −∆, D(H0 ) = H 2 (Rd ).
This naturally leads to the following definition.

Definition 1. We call Laplacian with point interactions placed at y1 , . . . , yn ∈ Rd

any non trivial selfadjoint extension of (2).

Many different mathematical techniques have been used to construct and

classify such class of operators.
 A Brief Review on Point Interactions 173

A general approach is based on the theory of selfadjoint extensions of sym-

metric operators, developed by von Neumann and Krein, but also approxima-
tion procedures in the resolvent sense, non-standard analysis or the theory of
quadratic forms can be used. We shall outline the construction based on the
theory of selfadjoint extensions.
The paper is organized as follows.
In Sect. 2 we consider the case of one or many interaction centers in d = 3
and d = 1. Few details will be given, mainly referring to the particularly simple
case of a single point interaction in dimension d = 3. (We shall omit to present
the d = 2 case which, in many respects, is analogous to the three dimensional
case.)
In Sect. 3 we consider the connections between point interaction Hamiltonians
and Schrödinger operators with smooth potentials.
In Sect. 4 we introduce time dependent and non linear point interactions and
we discuss possible applications of such kind of operators.

2 Construction of Point Interactions

Let us denote with (·, ·) the inner product in L2 (Rd ). As a consequence of the
definition (2), for any function ψ ∈ H 2 (R3 ) and φ ∈ D(Ĥ), we have

(ψ, −∆φ) = (−∆ψ, φ)

and
|(−∆ψ, φ)| ≤ ψH 2 φL2
implying that all functions in H 2 (R3 ) belong to the domain of the operator Ĥ ∗
adjoint to Ĥ.
On the other hand for any z ∈ C \ R+ we have

(Gzi , Ĥφ) = (zGzi , φ) φ ∈ D(Ĥ), i = 1, . . . , n (3)

where Gzi denotes the inverse Fourier transform of

eik·yi
G̃zi (k) =
k2 − z
In fact the Gzi satisfy, in the sense of distributions, (−∆ − z)Gzi = δyi . Being
the δyi together with their derivatives the only distributions supported by the
discrete set {y1 , . . . , yn } the Gzi and their derivatives are the only distributions
satisfying (3).
We conclude that the eigenspace Nz ⊂ L2 (Rd ) relative to the eigenvalue z
of the adjoint Ĥ ∗ of Ĥ contains all the Gzi and their derivatives as long as they
belong to L2 (Rd ).
174 G. Dell’Antonio et al.

Notice that for z ∈ C \ R+ :

– In d = 1 the Gzi and their first derivatives belong to L2 (R).

– In d = 2, 3 the Gzi belong to L2 (Rd ) but their derivatives do not.
– In d > 3 no function in L2 (Rd ) satisfies (3).
Summarizing the considerations made before we see that, for a finite number
of point interactions, the subspaces Nz have dimension which is constant as z
varies over the complex plane, outside the positive real axis and it is equal to 2n
in d = 1, to n in d = 2, 3 and to 0 for d > 3; Moreover the domain of the adjoint
operator Ĥ ∗ contains H 2 (Rd ) and all the subspaces Nz for z in the complex
plane outside the positive real axis.
Notice that the difference Gzi 1 − Gzi 2 for z1 , z2 ∈ C \ R+ belongs to H 2 (Rd ).
On one side this means that the linear span D generated by linear combinations
of functions f ∈ H 2 (Rd ) and g ∈ Nẑ , for some ẑ, includes all linear combinations
of the same kind with g ∈ Nz for any z ∈ C \ R+ .
On the other side any function in D can be alternatively expressed as a linear
combination of functions in Nz1 , Nz2 and H 2 (R3 \ {y1 , . . . , yn }) for any choice
of z1 , z2 ∈ C \ R+ . In fact the value taken at yi by the regular part of a linear
combination α1 Gzi 1 + α2 Gzi 2 , i.e.
 
α1 + α2
lim α1 Gzi 1 (x) + α2 Gzi 2 (x) −
x→yi 4π|x − yi |

can assume any complex value as α1 and α2 vary.

It is not hard to prove that in fact D is the entire domain of Ĥ ∗ . This result
is a particular case of a general decomposition formula (often quoted as von
Neumann formula) stating that if A is a densely defined symmetric operator in
a separable Hilbert space than the domain of its adjoint operator is completely
characterized as follows:
for all z such that z > 0, any vector f in the domain of A∗ has a unique
decomposition
f = f0 + f z + f z̄ (4)
where f0 belongs to the domain of the closure of A and f z (resp. f z̄ ) belongs to
the eigenspace of A∗ relative to the eigenvalue z (resp. z̄). The action of A∗ on
f is obviously given by
A∗ f = Af0 + zf z + z̄f z̄ (5)
As we mentioned before decomposition (4) can be easily proved directly in our
specific case. The proof in the general case can be found, e.g. in [AG].
An immediate application of the above result to our case is that there are no
extensions of Ĥ different from the Laplacian in dimension d > 3.
Formula (5) suggests the strategy one has to adopt in order to construct a
selfadjoint extension of Ĥ. Any such extension must correspond to a choice of a
subspace of the linear span of functions in Nz and in Nz̄ on which Ĥ ∗ acts as a
selfadjoint operator (symmetric as long as the dimension of Nz is finite).
 A Brief Review on Point Interactions 175

Let us follow the details of this construction in the case of one point interac-
tion in d = 3 placed at the point y ∈ R3 . We recall that for d = 3
√
exp i z|x|
Gz (x) =
4π|x|

From (4) and (5) we have in this case

0 1
D(Ĥ ∗ ) = f ∈ L2 (R3 )|f = f0 + βGz (· − y) + γGz̄ (· − y)

where f0 ∈ H 2 (R3 \ {y}) and

Ĥ ∗ f = −∆f0 + βzGz (· − y) + γ z̄Gz̄ (· − y)

A direct computation gives

 
[βGz (· − y) + γGz̄ (· − y), Ĥ ∗ [βGz (· − y) + γGz̄ (· − y)] = (6)
   
= |β|2 z + |γ|2 z̄ Gz 2 + 2 γ̄β(Gz̄ , Gz )

showing that Ĥ ∗ acts as a symmetric operator on linear combinations of Gz

and Gz̄ if and only if |β| = |γ|. We conclude that there are infinitely many
selfadjoint extension of Ĥ (equivalently selfadjoint restrictions of Ĥ ∗ ) defined in
the following way
0 1
D(Hφ,y ) = f ∈ L2 (R3 )|f = f0 + βGz (· − y) + βeiφ Gz̄ (· − y)

Hφ,y f = −∆f0 + βzGz + βeiφ z̄Gz̄ (7)

An alternative description of the family of selfadjoint extensions Hφ,y is obtained
in the following way: take λ positive and define Gλ = Gz=−λ . Notice that,
around x = y

Gz (x − y) + eiφ Gz̄ (x − y) − (1 + eiφ )Gλ (x − y) =

√ √ √
i z iφ i z̄ λ
= +e + (1 + eiφ ) + f0
4π 4π √ 4π
λ
= (1 + eiφ )(α + ) + f0 (8)
4π
where f0 is a regular function taking value zero in x = y and
√ √
 z φ  z
α= tan −
4π 2 4π
Formula (8) allows to characterize functions in the domains of the different
selfadjoint extensions as a relation connecting the behavior of the functions at
the singularity (which is β(1 + eiφ )/4π|x − y| in our case) and the value taken
at the same point by their regular part. More precisely (8) implies that for any
α ∈ R there is a selfadjoint extension of Ĥ defined in the following way
176 G. Dell’Antonio et al.

D(Hα,y ) =
 
Φλ (0)
= f ∈ L2 (R3 )|f = Φλ + qGλ (· − y), Φλ ∈ H 2 (R3 ), q= √
α+ λ/4π
(9)

Taking into account that

 
−∆ Gz (x − y) + eiφ Gz̄ (x − y) − (1 + eiφ )Gλ (x − y) =
= zGz (x − y) + eiφ z̄Gz̄ (x − y) + (1 + eiφ )λGλ (x − y)

the action of Hα,y on D(Hα,y ) is easily found to be

(Hα,y + λ)f = (−∆ + λ)Φλ (10)

Characterization (9) of the domain implies that the action of (Hα,y + λ)−1 does
not differ from the action of (−∆ + λ)−1 on functions of H 2 (R3 \ {y}) (being
q = 0 in this case).
On the other hand

((Hα,y + λ)−1 − (−∆ + λ)−1 )Gλ , (Ĥ + λ)g) = 0

for any g in the set C0∞ (R3 \ {y}), which is dense in L2 (R3 ). As a consequence,
the function ((Hα,y + λ)−1 − (−∆ + λ)−1 )Gλ belongs to N−λ , which in turn
means that it is proportional to Gλ itself.
The action of the two resolvents then differ only on the one dimensional
subspace generated by Gλ . Being D(Hα,y ) in (9) the range of (Hα,y + λ)−1 one
finally obtains
1
[(Hα,y + λ)−1 g](x) = (Gλ g)(x) + √ Gλ (x − y)(Gλ g)(y) (11)
α+ λ/4π
Notice that the free Laplacian is obtained formally in the limit α → ∞ showing
that one cannot consider α as a coupling constant. Its physical meaning becomes
clear studying scattering theory for point interaction Hamiltonians. In this set-
ting one proves that α is the inverse of the scattering length associated to the
operator Hα,y .
From the explicit expression of the resolvent (11) one can easily deduce the
spectral properties of Hα,y :
– The spectrum of Hα,y is

σ(Hα,y ) = [0, ∞) α ≥ 0
0 1
σ(Hα,y ) = −16π 2 α2 ∪ [0, ∞) α < 0
where the continuous part [0, ∞) is purely absolutely continuous.
– For α < 0 the only eigenvalue is simple and the corresponding normalized
eigenfunction is
√ exp(4πα|x − y|)
ψα (x) = −2α .
|x − y|
 A Brief Review on Point Interactions 177

– For any α ∈ R, corresponding to each positive energy E in the continuous

spectrum there are infinitely many generalized eigenfunction
 
1 exp iky exp(i|k||x − y|)
ψα (k, x) = exp(ik · x) − (12)
(2π)3/2 α − i|k|/(4π) |x − y|

with |k|2 = E.
The explicit form of the spectral decomposition of Hα,y , in terms of the
eigenfunction (in the case α < 0) and of the generalized eigenfunctions, allows
to write the solution of the Schrödinger equation
∂ψt
i = Hα,y ψt , (13)
∂t
corresponding to any initial state in L2 (R3 ), as an integral over the spectral
measure. In fact the integral kernel, in configuration space, for the propagator
of Hα,y can be explicitly computed ([ST]). The same procedure allows to write
the kernel of the semigroup generated by Hα,y ([ABD]).
An implicit characterization of the propagator will be useful in the next
section in order to define time dependent point interactions. A formal inverse
Laplace transform of (11) suggests for the solution of the Schrödinger equation
(13) the following formula showing a free-propagation contribution and a term
representing spherical waves generated at the interaction center

t
ψt (x) = (U (t)ψ0 ) (x) + i ds U (t − s, |x − y|) q(s) (14)
0

where U (t) is the propagator of the free unitary group defined by the kernel
|x−x |2
  ei 4t
U (t; x − x ) = e i∆t
(x − x ) =
(4πit)3/2
It is easily checked that (14) is the solution of the Schrödinger equation (13)
corresponding to an initial condition ψ0 ∈ D(Hα,y ) if the function q(t) satisfies
the Volterra integral equation
√
t √
t (U (s)ψ0 ) (y)
q(s)
q(t) + 4 iπα ds √ = 4 iπ ds √ (15)
0 t−s 0 t−s
More precisely one can prove that if q(t) is the unique solution of (15) then (14)
defines a function ψt which for any t ≥ 0 belongs to D(Hα,y ) (q(t) being the
coefficient of the singular part of ψt ) and satisfies (13).
We want to conclude the list of properties of the family of Hamiltonians Hα,y
recalling the expression of the associated quadratic forms.
Quadratic forms are often taken as an alternative way to define point inter-
actions, a way often preferred to the one used above because of their immediate
connection with quantities of physical interest. On D(Hα,y ) the quadratic form
uniquely associated to the operator Hα,y is
178 G. Dell’Antonio et al.

Fα,y (u) =

 √
" # λ
= dx |∇(u − qGλ (· − y))| + λ|(u − qGλ (· − y))| − λ|u| + α +
2 2 2
|q|2
R 3 4π
(16)
It is bounded below and closable and defines therefore a Dirichlet form. The
domain of its closure is

D(Fα,y ) = {u ∈ L2 (R3 ) | u = φλ + qGλ (· − y), φλ ∈ H 1 (R3 ), q ∈ C } (17)

Notice that the form domain is significantly larger than H 1 (R3 ). In particular
the form Fα,y is not a small perturbation of the one defined by the Laplacian.
We shall now go back to the n centers case in R3 and summarize the main
results about the selfadjoint extensions of the operator defined in (2).
As it was mentioned at the end of Sect. 1 the dimension of the eigenspaces Nz
relative to the eigenvalue z of the operator adjoint to Ĥ is n for any z ∈ C \ R+
in the case of n centers. Consider any unitary operator V from Nz to Nz̄ and
denote with NzV the subspace of the linear span of functions in Nz and in Nz̄
generated by the linear combinations of the type fz + V fz with fz ∈ Nz . A
computation identical to (7) shows that Ĥ ∗ acts as a symmetric operator only
if its action is restricted to any NzV .
As a consequence in this case selfadjoint extensions of Ĥ are uniquely associ-
ated with the n2 dimensional family of (complex) matrices unitarily connecting
Nz and Nz̄ .
In the literature much attention was given to a particular n dimensional
subfamily of selfadjoint extensions called local. In fact, in analogy with the one
center case, functions in the domain of the n dimensional subfamily of selfadjoint
extensions, corresponding to diagonal unitary matrices V , can be alternatively
characterized by a specific behavior around each interaction center. We briefly
recall the properties of the operators in this subfamily (for an introduction to
non local point interactions see [DG]).
For any α = {α1 , . . . , αn } with αi ∈ R, i = 1, . . . , n and y = {y1 , . . . ,
yn }, yi ∈ R3 , i = 1, . . . , n the operator Hα,y defined by

2 n
D(Hα,y ) = u ∈ L2 (R3 ) | u = φλ + qk Gλ (· − yk )
k=1
n 3
φλ ∈ H (R ),
2 3
φλ (yj ) = [Γα,y (λ)]jk qk , j = 1, ..., n (18)
k=1

(Hα,y + λ)u = (−∆ + λ)φλ (19)

where  √
λ
[Γα,y (λ)]jk = αj + δjk − Gλ (yj − yk )(1 − δjk ) (20)
4π
 A Brief Review on Point Interactions 179

is the selfadjoint extension of Ĥ referred to as the (local) point interaction

Hamiltonian with n centers located in y . In fact for any smooth function
u ∈ D(Hα,y ) vanishing at y1 , ..., yn one has q = 0 and then, from (19),
Hα,y u = −∆u.
At each point yj the elements of the domain satisfy a boundary condition
expressed by the last equality in (18). If we define rj = |x − yj | it is easy to see
that the boundary condition can be equivalently written

∂(rj u)
lim − 4παj (rj u) = 0, j = 1, ..., n (21)
rj →0 ∂rj

This explains the term “local” by which this class of extensions is known; indeed
due to the special form of the matrix (20) the generalized boundary conditions
(21) is placed at each point separately. Any other choice for the symmetric
invertible matrix Γα,y (λ) leads to an extension which is still local in the ter-
minology commonly employed for differential operators (i.e. operators which do
not change the support). The extensions obtained by (20) should be perhaps
termed “strongly local” or “local” with respect to the boundary conditions.
In order to find the explicit expression for the resolvent it is sufficient
to observe that, for any given f ∈ L2 (R3 ), the solution u of the equation
(Hα,y + λ)u = f must be written in the form u = Gλ f + qGλ (· − y) (see
(19)), where the charges q are determined imposing the boundary conditions in
(18) (or equivalently (21)). The result of the easy computation is
n
(Hα,y + λ)−1 = Gλ + [Γα,y (λ)]−1
jk Gλ (· − yj )Gλ (· − yk ) (22)
j,k=1

where λ > 0 is chosen sufficiently large so that the matrix (20) is invertible.
From the analysis of (22) one can derive that the continuous spectrum of
Hα,y is purely absolutely continuous and coincides with the positive real axis.
The point spectrum consists of (at most) n non positive eigenvalues given by
the possible solutions E ≥ 0 of the equation det [Γα,y (−E)] = 0.
Proper and generalized eigenfunctions can be explicitly computed.
Notice that the main tool used in the construction of the selfadjoint exten-
sions Hα,y was the specific behavior of the Green’s function of the Laplacian
at the singularity. Therefore the approach we described above can be adapted
to treat stationary or evolution problems related to perturbations supported by
points of any elliptic operator of the general form
d
∂ ∂
LV = (i + Ak )akj (x)(i + Aj ) + V (x)
∂xk ∂xj
k,j=1

under suitable assumptions on the coefficients and on the potential V.

According to the notation introduced above LVα,y would denote the pertur-
bation of LV supported by the points yi with strength αi . The usefulness of the
180 G. Dell’Antonio et al.

definition in the general case is linked with the possibility of finding an explicit
expression for the Green’s function of the operator LV (as in the case of the
harmonic oscillator).
It is easy to check an additivity property with respect to the potential while
on the other hand, in dimension 2 and 3, point interactions are not additive in
the sense that  
LVα,y = LVα∪β,y∪z
β,z

As we mentioned in the introduction to this section the family of selfadjoint

extensions of Ĥ has a richer structure in d = 1, due to the fact that the subspaces
Nz contain also the first derivatives of the Green’s functions of the Laplacian
centered in the interaction centers. Referring for simplicity to the case of a single
point interaction the dimension of Nz z ∈ C\R+ is equal to two. As a consequence
there is a four (real) parameter family of selfadjoint extensions of Ĥ.
Without going into details of the construction (see [AGH-KH] for a complete
treatment) we recall here the properties of two particular one parameter sub-
families of selfadjoint extensions referred to as δ and δ  interactions, for reasons
that will become clear from their definitions.
The one center (placed in y ∈ R) delta interaction Hamiltonian Hα,y is
defined as follows
2 3
dφ + dφ −
D(Hα,y ) = φ ∈ H (R) ∩ H (R \ {y}) |
1 2
(y ) − (y ) = αφ(y), α ∈ R
dx dx
(23)
and Hα,y acts on a function in its domain as −d2 /dx2 in each point x = y.
Before discussing the properties of the operators Hα,y , in order to stress their
relation with the one dimensional delta function, we introduce the associated
quadratic forms

D(Fα,y ) = H 1 (R) (24)

Fα,y (u) = dx|∇u|2 + α|u(y)|2 (25)
R

The form domain is then the same of the form domain of the Laplacian and in
fact standard Sobolev inequalities show that Fα,y is a small perturbation of the
form associated to the Laplacian (see ([T]) for details). In terms of quadratic
forms the generalization to n centers placed in {y1 , . . . , yn } ≡ y of strength
{α1 , . . . , αn } ≡ α is immediate

D(Fα,y ) = H 1 (R) (26)

n
Fα,y (u) = dx|∇u|2 + αj |u(yj )|2 (27)
R j=1

Formula (27) shows that, differently from the higher dimensional cases, in
d = 1 point interactions are genuinely additive and that the dynamical parameters
αi play the role of coupling constants.
 A Brief Review on Point Interactions 181

The explicit form of the integral kernel of the resolvent can be computed; in
the single center case one has
√  √ √ √ 
−1  e− λ|x−x | 2α λ e− λ|x−y| e− λ|x −y|
(Hα,y + λ) (x, x ) = √ − √ √ √
2 λ α+2 λ 2 λ 2 λ
Analyzing the singularities of the resolvent one can easily see that the spec-
trum σ(Hα,y ) of Hα,y has the properties

σ(Hα,y ) = [0, +∞), α≥0

2 2
3
α
σ(Hα,y = − ∪ [0, +∞), α<0
4
[0, ∞) being purely absolutely continuous.
2
In the case α < 0 the only eigenvalue − α4 is non degenerate and the corre-
sponding normalized eigenfunction is
*
α α
Ψα (x) = − e 2 |x−y|
2
We refer to [AGH-KH] for the explicit form of the generalized eigenfunctions.
Their knowledge allows to compute any bounded function of the operator Hα,y ;
in particular its propagator can be written in closed form ([S]). We want to
remind the reader that an array of delta interactions in one dimension were used
by Kronig and Penney in 1931 ([KP]) to construct a model of the dynamics of
an electron inside a crystal. That model remains, after many years, one of the
few completely solvable periodic quantum interaction.
In d = 1 the implicit equation for the solution of the Schrödinger equation,
corresponding to (14) in d = 3, can be easily obtained from the formula of the
resolvent operator

t
ψt (x) = (U0 (t)ψ0 ) (x) − iα ds U0 (t − s, |x − y|) ψs (y) (28)
0

where U0 is the free propagator in dimension one.

From (28) one sees that the solution is completely determined by the free
evolution of ψ0 and by ψt (y), which in turn, as a function of time, must satisfy
the integral equation

t
ψt (y) = (U0 (t)ψ0 ) (y) − iα ds U0 (t − s, 0) ψs (y) (29)
0

It is easy to check that in fact (28) and (29) define a unitary flow in L2 (R) whose
generator is Hα,y and therefore they may be used as a definition of Hα,y .
We want briefly to mention the other subfamily of selfadjoint extensions of
Ĥ in d = 1 known as δ  interactions.

For each β ∈ R and y ∈ R the operator Hβ,y can be defined via the action
on a function f ∈ L (R) of its resolvent in the following way
2
182 G. Dell’Antonio et al.

 2βλ
(Hβ,y + λ)−1 )f = (−∆ + λ)−1 f + √ Gλ (· − y)(Gλ f )(y) (30)
2+β λ

where Gλ is the derivative of the Green’s function for z = −λ, λ > 0

1 √
Gλ = − e− λ|x−y| x > y
2
1 −√λ|x−y|
= e x<y
2
and λ = (2/β)2 if β < 0.

The spectrum of Hβ,y is easily found to be

σ(Hβ,y ) = [0, +∞), β≥0
2 3
 4
σ(Hβ,y ) = − 2 ∪ [0, +∞), β<0
β

For further details and properties on this kind of operators see [AGH-KH].

3 Connection with Smooth Interactions

It is natural to think of a point interaction as a limit of rescaled short range

potentials. On the other hand it is conceivable that any Hamiltonian with a
smooth potential can be approximated by an array of a large number of point
interactions of suitable strengths.
Let us discuss the former aspect first.
In dimension one this intuitive picture can be made rigorous under very
general conditions. For any  > 0, consider the following approximating operators
n
1
H = −∆ + Vj (−1 (x − yj )) (31)
 j=1

where the potentials Vj can be chosen in L1 (R). Then for any λ > 0 sufficiently
large one has
lim (H + λ)−1 − (Hα,y + λ)−1  = 0 (32)
→0
(
with αj = R dxVj (x) (see [AGH-KH]).
In dimension three the situation is more delicate and the class of approxi-
mating potentials must be properly chosen.
In particular one has to restrict to potentials V such that −∆ + V has a
so-called zero-energy resonance, i.e. there exists a solution ψ ∈ L2loc (R3 ), with
∇ψ ∈ L2 (R3 ), of the equation (−∆ + V )ψ = 0. We call such a solution a zero-
energy resonance function.
 A Brief Review on Point Interactions 183

To simplify the notation, let us fix n = 1 and consider V ∈ L2 (R3 ), with

(1 + | · |)V ∈ L1 (R3 ); moreover assume that there is a unique ξ ∈ L2 (R3 )
such that
uG0 vξ = −ξ, (v, ξ) = 0 (33)
where u(x) = |V (x)|1/2 sgn[V (x)], v(x) = |V (x)|1/2 . Condition (33) implies that
ψ(x) = (G0 vξ)(x) is a zero-energy resonance function.
Now consider the following sequence of approximating operators
1 + µ
H = −∆ + V (−1 (x − y)), µ∈R (34)
2
Under the above assumptions and for any λ > 0 sufficiently large one can prove

lim (H + λ)−1 − (Hα,y + λ)−1  = 0 (35)

→0

where α = −µ|(v, ξ)|−2 (see [AGH-KH]).

It is clear from the rescaling of the potential in (34) that a point interaction in
dimension three cannot be considered as a true δ-function. Notice that for µ = 0
one gets α = 0, which does not correspond to the free Laplacian (see (22)).
It should be emphasized that if the resonance condition (33) is not fulfilled
then in the limit  → 0 one obtains a trivial result, i.e. the free Laplacian.
The approximation of a point interaction in dimension three can be made
more transparent if one considers the connection with the following boundary
value problem

(−∆ + λ)u = f in R3 \ S (36)

∂u ∂u
− = γ u on S (37)
∂n+ ∂n−

where λ > 0, f ∈ L2 (R3 ), γ ∈ R and S is the sphere of radius  and center y.

The solution of (36), (37) is nothing but (Hγ ,S +λ)−1 f , where Hγ ,S denotes
the Laplacian in R3 with a δ-shell interaction of strength γ supported by the
sphere S .
For  → 0 the sphere shrinks to the point y; assuming
1
γ = − + 4πα (38)

and λ > 0 is sufficiently large one can show that ([FT])

lim (Hγ ,S + λ)−1 − (Hα,y + λ)−1  = 0 (39)

→0

Let us now go back to the second problem mentioned at the beginning of this
section concerning a procedure allowing to approximate (in a way which should
be suitable for applications) a Schrödinger operator with smooth potential via
point interaction Hamiltonians.
The answer is obviously positive in dimension d = 1 for the particular self-
adjoint Hamiltonians we referred to as delta interaction Hamiltonians. In that
184 G. Dell’Antonio et al.

case point interactions are additive and, at least in the sense of quadratic forms,
they correspond to a potential which is a sum of delta functions. It is easy to
prove that any approximation of a potential function through linear combina-
tions of delta measures turns out to be an efficient approximation scheme for
the corresponding operators ([BFT]).
As we mentioned the situation is fairly different in dimension d = 2 and
d = 3. In these cases point interaction Hamiltonians for n centers are in no sense
sum of an unperturbed operator and a potential, as we remarked at the end of
the last section. Moreover point interactions in d = 2 and d = 3 are in no sense
connected to delta measures as we discussed in the first part of this section.
Nevertheless, with some peculiar differences with respect to the one dimen-
sional case, it is possible to work out an approximation procedure that we out-
line here restricting again our attention to the case d = 3. Let us denote with
(N ) (N )
Y N = {y1 , . . . , yN } a configuration of N points in R3 and consider a (prod-
uct) probability measure [ρ(y)dy]⊗N on such configuration space. For simplicity
let us assume that the density is continuous. Let α(y) be a real function on
R3 continuous and bounded away from zero outside a set of Lebesgue measure
(N ) (N )
zero and define αN ≡ {α(y1 ) . . . , α(yN )}. The following theorem was proved
in [FHT]
On a set of configurations of measure increasing to 1 as N tends to infinity,
for any λ positive, large enough
 −1  ρ −1
s − lim HN αN ,Y N + λ) = −∆ − + λ (40)
N →∞ α
We want to mention an immediate consequence of the result stated above. Notice
first that any smooth integrable potential V can be written in the form ρ/α
taking
|V (y)| (sgnV )(y)
ρ(y) = ( α= (
R3
|V (y)|dy R3
|V (y)|dy
We conclude that any Schrödinger operator with a smooth integrable potential
can be approximated by point interaction Hamiltonians.
As a final remark we want to stress that (40) shows that positive potentials
are obtained as limits of negative alphas and vice versa. This is a consequence
of the fact that even if all the αi ’s are positive the corresponding n centers point
interaction Hamiltonian can have negative eigenvalues if some, or all, the points
are very close to each other. This depends in turn on the fact that the non
diagonal terms in the matrix (20) can dominate the diagonal terms if there are
couples of very close points.
On the other hand if all the αi ’s are negative the rescaled Hamiltonian shows
N negative eigenvalues all tending to infinity for large N . In this second case
there is no uniform bound from below for the eiganvalues of the operators in the
approximating sequence and in particular no convergence of the corresponding
semigroups can take place.
 A Brief Review on Point Interactions 185

4 Time Dependent Point Interactions

In this last section we shall mention some more recent results obtained for time
dependent (linear or non linear) point interactions.
In order to avoid unnecessary technical difficulties we are going to outline
only the case of a single interaction center in d = 1 and d = 3.
We start with the d = 1 case noticing that the procedure of rephrasing the
Schrödinger evolution problem via formulae (28), (29) is directly generalizable
to the case of y and α varying with time. In fact, under suitable conditions on
the regularity of y(t), α(t) and on the initial condition it is possible to prove that
for each s ∈ R the map ψs → ψt induced by

t
ψt (x) = (U0 (t − s)ψs ) (x) − i dτ α(τ ) U0 (t − τ, |x − y(τ )|) ψτ (y(τ )) (41)
s

where ψt (y(t)) satisfies

t
ψt (y(t)) = (U0 (t − s)ψs ) (y(t)) − i dτ α(τ ) U0 (t − τ, |y(t) − y(τ )|) ψτ (y(τ ))
s
(42)
defines a group of unitary transformations V (s, t),with V (s, s) = 1, which solves
the Schrödinger equation
∂V (s, t)ψs
i = Hα(t),y(t) V (s, t)ψs (43)
∂t
Notice that in dimension one the form domain of the Hamiltonians Hα(t),y(t) does
not depend on time and coincide with H 1 (R). This permits to apply the general
abstract theory about the solution of non autonomous evolution problems (see,
e.g. [Si]).
In d = 3 the problem consists in generalizing the representation formulae
(14), (15) to the case in which α or the position of the interaction center are
varying in time.
Notice that the form domain in (16) changes if the position of the point y
varies. As a consequence the Schrödinger problem corresponding to a moving
point interaction is more complicate than in dimension one. The problem in the
case where the strength parameter α changes in time (in which case the operator
domain is varying but not the form domain) was solved in [SY]. In the following
theorem the results in the general case when strength and position change in
time are summarized.
If y(t) is regular curve in R3 , α(t) a smooth function in R and f ∈
∞
C0 (R3 \ {y(s)}) then there exists a unique ψs (t) ∈ D(Fα(t),y(t) ), t ∈ R, such
that ∂ψs (t)/∂t is in the dual (with respect to L2 (R3 )) of D(Fα(t),y(t) ) and

∂ψs (t)
i < v(t), > = Bα(t),y(t) (v(t), ψs (t)) ∀v(t) ∈ D(Fα(t),y(t) )
∂t
ψs (s) = f
(44)
186 G. Dell’Antonio et al.

where Bα(t),y(t) is the bilinear form corresponding to the quadratic form

Fα(t),y(t) .
Moreover ψs (t) has the following representation

t
ψs (t) = U (t − s)f + i dτ U (t − τ ; · − y(τ ))q(τ ) (45)
s

where the charges q(t) satisfy the Volterra integral equation

√

t √

4 π t q(τ ) 4 π t (U0 (τ − s)f )(y(τ ))
q(t)+ √ dτ α(τ ) √ + dτ q(τ )C(t, τ ) = √ dτ √
−i s t−τ s −i s t−τ
(46)
where


1 1 1
C(t, τ ) = − dz ' A(τ + (t − τ )z, τ ) + Ḃ(τ + (t − τ )z, τ )
π 0 (1 − z)z

B(τ + (t − τ )z, τ ) − 1
+
2(t − τ )z

w(t,τ )
(y(t) − y(τ )) · ẏ(τ ) 1 2
A(t, τ ) = i dz z 2 eiz
2(t − τ ) 3
w (t, τ ) 0

w(t,τ )
1 2
B(t, τ ) = dz eiz
w(t, τ ) 0
|y(t) − y(τ )|
w(t, τ ) = √
2 t−τ
and Ḃ(t, τ ) denotes the derivative with respect to the second argument.
Similar results in the n centers case with milder requests of regularity for
the αi (t) and yi (t) can be proved for the corresponding parabolic problem. For
details see [DFT1] and [DFT2].
We want finally consider a class of nonlinear evolution problems with point
interaction concentrated in a single point (which for sake of simplicity will be
taken to be the origin). The pointlike interaction models reproduce the basic
features of the standard nonlinear Schrödinger equation in a simpler and more
tractable context. Models of this kind, in space dimension one, appeared al-
ready in the physical literature ([J-LPS], [MA], [BKB]) to describe the resonant
tunneling of electrons through a double barrier heterostructure.
Connections between standard nonlinear Schrödinger equation and point in-
teraction Hamiltonians are more transparent in d = 1.
As an example, the one dimensional nonlinear Schrödinger equation with
critical power nonlinearity is
∂ξ ∂2ξ
i (t, x) = − 2 (t, x) − |ξ(t, x)|4 ξ(t, x)
∂t ∂ξ
whereas the one dimensional Schrödinger equation with critical nonlinear delta
interaction is given by
 A Brief Review on Point Interactions 187

∂η ∂2η
i (t, x) = − 2 (t, x) − |η(t, 0)|2 δ0 η(t, x) (47)
∂t ∂ξ

Since for a class of blow-up solutions ξ(t) of the first equation it has been shown
that |ξ(t)|2 → δ0 , it appears that, for this family of solutions, the second equation
approximates the first one.
Unfortunately, this striking analogy does not hold in higher dimensions.
In the following we are going to consider the three dimensional case. For
details on the one dimensional case see [AT1], [AT2].
Notice that even in the linear three dimensional case the solution of the
evolution problem exhibits a singularity where the interaction is placed. This
means that the nonlinear interaction cannot be introduced as in d = 1. The value
of the function at the origin should be replaced in this case by the coefficient
q(t) of the singular part. This suggests the following formulation of the evolution
problem (see (15))

t
ψt (x) = (U (t)ψ0 )(x) + i dsU (t − s, |x|)q(s)
0
√
t α(s)q(s) √
t (U (s)ψ0 )(0)
q(t) + 4 i π ds √ =4 iπ ds √
0 t−s 0 t−s
α(t) = γ|q(t)|2σ , γ ∈ R, σ > 0
(48)

The natural space where one looks for the solution of a standard nonlinear
Schrödinger equation is H 1 , i.e. the form domain of the generator of the linear
dynamics obtained setting to zero the exponent of the nonlinear term. This
suggests to look for the solution of (48) in the form domain D(Fα,0 ) of Hα,0 . It
is easy to check that the following characterization of D(Fα,0 ) is equivalent to
the one given in (17)

D(Fα,0 ) = {u ∈ L2 (R3 ) | u = φλ + qGλ , φλ ∈ H 1 (R3 ), q ∈ C}

One can prove the following result ([ADFT1])

Let ψ0 = φλ0 + qGλ ∈ D(Fα,0 ), with φλ0 ∈ H 2 (R3 ) then there exists t̄ > 0 s.t.
problem (48) has a unique solution ψt ∈ D(Fα,0 ), t ∈ [0, t̄). Moreover if either
σ < 1 or γ > 0 the solution is global in time.
Global existence for repulsive interactions (i.e. γ > 0) is a direct consequences
of the conservation of the L2 -norm and of the energy
γ
E(ψt ) = ∇φt 2 + |q(t)|2σ+2
σ+1
where ψt = φt + q(t)Gλ .
One expects global existence also for weakly attractive interactions. The proof
is not trivial because one is not working in H 1 and standard Sobolev inequalities
are not available.
188 G. Dell’Antonio et al.

Consider finally the problem of the existence of blow-up solutions. A solution

ψt of problem (48) will be said to be a blow-up solution if there exists a t0 < ∞
such that
lim ∇φt  = ∞
t→t0

or equivalently
lim |q(t)| = ∞
t→t0

With this definition of blow-up solutions, exploiting standard techniques in the

theory of nonlinear (Schrödinger) equations it is possible to prove the following
result ([ADFT2])
Let γ < 0, σ ≥ 1 and ψ0 = φλ0 + q0 Gλ , with φλ0 ∈ H 2 (R3 ), |x|ψ0 ∈ L2 (R3 ),
E(ψ0 ) < 0.
Then the solution of problem (48) blows up in both directions of time.

References
[ABD] Albeverio S., Brzeźniak Z., Da̧browski L., Fundamental Solution of the
Heat and Schrödinger Equations with Point Interaction, J. Func. Anal.,
130, 220–254, 1995.
[ADFT1] Adami R., Dell’Antonio G., Figari R., Teta A., The Cauchy Problem for the
Schrödinger Equation in Dimension Three with Concentrated Nonlinearity,
Ann. Inst. H. Poincare’ Non Lineaire, 20, no. 3, 477-500, 2003.
[ADFT2] Adami R., Dell’Antonio G., Figari R., Teta A., Blow Up Solutions for
the Schrödinger Equation with a Concentrated Nonlinearity in Dimension
Three, Ann. Inst. H. Poincare’ Non Lineaire, 21, no. 1, 121-137, 2004.
[AG] Akhiezer N.I., Glazman I.M., Theory of Linear Operators in Hilbert Space,
Vol 2, Pitman, Boston-London-Melbourne, 1981.
[AGH-KH] Albeverio S., Gesztesy F., Högh-Krohn R., Holden H., Solvable Models in
Quantum Mechanics, Springer-Verlag, New York, 1988.
[AT1] Adami R., Teta A., A Simple Model of Concentrated Nonlinearity, in Math-
ematical Results in Quantum Mechanics, Dittrich J., Exner P.,Tater M.
eds., Birkhäuser, 1999.
[AT2] Adami R., Teta A., A Class of Nonlinear Schrödinger Equation with Con-
centrated Nonlinearity, J. Func. An., 180,148-175, 2001.
[BFT] Brasche J.F., Figari R., Teta A., Singular Schrödinger Operator as Limits
of Point Interaction Hamiltonians, Potential Analysis, 8, 163-178, 1998.
[BKB] Bulashenko O.M., Kochelap V.A., Bonilla L.L., Coherent Patterns and
Self-Induced Diffraction of Electrons on a Thin Nonlinear Layer, Phys.
Rev. B, 54, 3, 1996.
[DFT1] Dell’Antonio G.F., Figari R., Teta A., Diffusion of a Particle in Presence
of N Moving Point Sources, Ann. Inst. H. Poincare’ A., 69, 413-424, 1998.
[DFT2] Dell’Antonio G.F., Figari R., Teta A., Schrödinger Equation with Moving
Point Interactions in Three Dimensions, in Stochastic Processes, Physics
and Geometry: New Interplays, vol. I, Gesztesy F., Holden H., Jost J.,
Paycha S., Röckner M., Scarlatti S. eds., A.M.S., Providence, 2000.
[DG] Da̧browski L., Grosse H., On nonlocal point interactions in one, two, and
three dimensions, J. Math. Phys., 26, 2777-2780, 1985.
 A Brief Review on Point Interactions 189

[F] Fermi E., Sul moto dei neutroni nelle sostanze idrogenate, Ricerca Sci-
entifica 7, 13-52, 1936 (Engl. Trans. in E. Fermi Collected Papers Vol.1,
University of Chicago, Chicago-London, 1962).
[FHT] Figari R., Holden H., Teta A., A law of large numbers and a central limit
theorem for the Schrödinger operator with zero range potentials, J. Stat.
Phys., 51, 205-214, 1988.
[FT] Figari R., Teta A., A boundary value problem of mixed type on perforated
domains Asymptotic Analysis, 6, 271-284, 1993.
[J-LPS] Jona-Lasinio G., Presilla C., Sjöstrand J., On Schrödinger Equations with
Concentrated Nonlinearities, Ann. Phys., 240, 1-21, 1995.
[KP] Kronig R. de L., Penney W. G., Quantum Mechanics of Electrons in Crystal
Lattices, Proc. Roy. Soc. (London), 130A, 499-513, 1931.
[MA] Malomed B., Azbel M., Modulational Instability of a Wave Scattered by a
Nonlinear Center, Phys. Rev. B, 47, 16, 1993.
[S] Schulman L.S., Application of the propagator for the delta function po-
tential, in Path integrals from mev to Mev, Gutzwiller M.C., Iuomata A.,
Klauder J.K., Streit L. eds., World Scientific, Singapore, 1986.
[ST] Scarlatti S., Teta A., Derivation of the Time-dependent Propagator for the
Three Dimensional Schródinger Equation with One Point Interaction, J.
Phys. A, 23, 1033-1035, 1990.
[Si] Simon B., Quantum Mechanics for Hamiltonians Defined as Quadratic
Forms, Princeton University Press, 1971.
[SY] Sayapova M.R., Yafaev D.R., The evolution operator for time-dependent
potentials of zero radius, Proc. Stek. Inst. Math., 2, 173-180, 1984.
[T] Teta A., Quadratic Forms for Singular Perturbations of the Laplacian, Pub-
lications of the R.I.M.S., Kyoto University, 26, 803-817, 1990.
List of Participants

1. Antonelli Laura 10. Cuoghi Paola

University of Napoli, Italy University of Modena, Italy
[email protected] [email protected]
2. Arridge Simon R. 11. Dell’Antonio Gianfausto
University College London University of Roma1, Italy
[email protected] [email protected]
3. Bandiera Francesco
University of Lecce, Italy 12. Di Cristo Michele
[email protected] University of Milano, Italy
[email protected]
4. Berowski Przemyslaw
Warsaw University, Polonia 13. Dorn Oliver
[email protected] University of Carlos III de Madrid,
5. Bonilla Luis Spain
University of Carlos III de Madrid, [email protected]
Spain (lecturer)
(editor) 14. Dubovskii Pavel
6. Canfora Michela Russian Academy of Sciences,
University of Firenze, Italy Russia
[email protected] [email protected]
7. Carpio Anna 15. Eelbode David
University of Complutense de University of Ghent, Belgium
Madrid, Spain [email protected]
ana [email protected]
8. Cebrian Elena 16. Galletti Ardelio
University of Burgos, Spain University of Napoli, Italy
[email protected] [email protected]
9. Chukalina Marina 17. Gonzalez Pedro
Institute of Microelectronics, University of Carlos III de Madrid,
Moscow, Russia Spain
[email protected] [email protected]
192 List of Participants

18. Gosse Laurent 31. Natterer Frank

IAMI-CNR, Bari, Italy Westfälische Wilhelms-University
[email protected] Münster, Germany
19. Hristova Veronica [email protected]
University of Sofia, Bulgaria (lecturer)
veronica [email protected] 32. Orlandi Enza
20. Landi Germana University of Roma3, Italy
University of Bologna, Italy [email protected]
[email protected] 33. Papanicolaou George
21. Leitao Antonio Stanford University, USA
University of Santa Catarina, [email protected]
Brasil 34. Rebolledo Aldo Franco
[email protected] University of del Valle, Colombia
22. Luquin Brigitte [email protected]
University Pierre et M. Curie, 35. Rey Fernando Gonzalo
France University of Buenos Aires,
[email protected] Argentina
23. Makarenkov Oleg [email protected]
University of Voronezh, Russia
36. Salani Claudia
[email protected]
University of Milano, Italy
24. Makrakis George
[email protected]
University of Creta, Greece
37. Sanchez Oscar
[email protected]
University of Granada, Spain
25. Mallaina Eduardo Fed
[email protected]
University of Buenos Aires,
Argentina 38. Skaug Christian
[email protected] CNR, Bari, Italy
26. Mariani Francesca [email protected]
University of Firenze, Italy 39. Soleimani Manuchehr
[email protected] University of Manhester, UK
27. Marino Zelda [email protected]
University of Napoli, Italy 40. Stasiak Magdalena
zelda [email protected] University of Lodz, Polonia
28. Massone Anna Maria [email protected]
INFM Genova, Italy 41. Symes William S.
[email protected] Rice University, USA
29. Mastronardi Nicola [email protected]
CNR, Bari, Italy 42. Teta Alessandro
[email protected] University of L’ Aquila, Italy
30. Moscoso Miguel [email protected]
University Carlos III de Madrid, (lecturer)
Spain 43. Zama Fabiana
[email protected] University of Bologna, Italy
(lecturer) [email protected]
LIST OF C.I.M.E. SEMINARS

Published by C.I.M.E
1954 1. Analisi funzionale
2. Quadratura delle superficie e questioni connesse
3. Equazioni differenziali non lineari

1955 4. Teorema di Riemann-Roch e questioni connesse

5. Teoria dei numeri
6. Topologia
7. Teorie non linearizzate in elasticità, idrodinamica, aerodinamic
8. Geometria proiettivo-differenziale

1956 9. Equazioni alle derivate parziali a caratteristiche reali

10. Propagazione delle onde elettromagnetiche automorfe
11. Teoria della funzioni di più variabili complesse e delle funzioni

1957 12. Geometria aritmetica e algebrica (2 vol.)

13. Integrali singolari e questioni connesse
14. Teoria della turbolenza (2 vol.)

1958 15. Vedute e problemi attuali in relatività generale

16. Problemi di geometria differenziale in grande
17. Il principio di minimo e le sue applicazioni alle equazioni funzionali

1959 18. Induzione e statistica

19. Teoria algebrica dei meccanismi automatici (2 vol.)
20. Gruppi, anelli di Lie e teoria della coomologia

1960 21. Sistemi dinamici e teoremi ergodici

22. Forme differenziali e loro integrali

1961 23. Geometria del calcolo delle variazioni (2 vol.)

24. Teoria delle distribuzioni
25. Onde superficiali

1962 26. Topologia differenziale

27. Autovalori e autosoluzioni
28. Magnetofluidodinamica

1963 29. Equazioni differenziali astratte

30. Funzioni e varietà complesse
31. Proprietà di media e teoremi di confronto in Fisica Matematica

1964 32. Relatività generale

33. Dinamica dei gas rarefatti
34. Alcune questioni di analisi numerica
35. Equazioni differenziali non lineari

1965 36. Non-linear continuum theories

37. Some aspects of ring theory
38. Mathematical optimization in economics
Published by Ed. Cremonese, Firenze
1966 39. Calculus of variations
40. Economia matematica
41. Classi caratteristiche e questioni connesse
42. Some aspects of diffusion theory

1967 43. Modern questions of celestial mechanics

44. Numerical analysis of partial differential equations
45. Geometry of homogeneous bounded domains

1968 46. Controllability and observability

47. Pseudo-differential operators
48. Aspects of mathematical logic

1969 49. Potential theory

50. Non-linear continuum theories in mechanics and physics and their applications
51. Questions of algebraic varieties

1970 52. Relativistic fluid dynamics

53. Theory of group representations and Fourier analysis
54. Functional equations and inequalities
55. Problems in non-linear analysis

1971 56. Stereodynamics

57. Constructive aspects of functional analysis (2 vol.)
58. Categories and commutative algebra

1972 59. Non-linear mechanics

60. Finite geometric structures and their applications
61. Geometric measure theory and minimal surfaces

1973 62. Complex analysis

63. New variational techniques in mathematical physics
64. Spectral analysis

1974 65. Stability problems

66. Singularities of analytic spaces
67. Eigenvalues of non linear problems

1975 68. Theoretical computer sciences

69. Model theory and applications
70. Differential operators and manifolds

Published by Ed. Liguori, Napoli

1976 71. Statistical Mechanics
72. Hyperbolicity
73. Differential topology

1977 74. Materials with memory

75. Pseudodifferential operators with applications
76. Algebraic surfaces

Published by Ed. Liguori, Napoli & Birkhäuser

1978 77. Stochastic differential equations
78. Dynamical systems

1979 79. Recursion theory and computational complexity

80. Mathematics of biology

1980 81. Wave propagation

82. Harmonic analysis and group representations
83. Matroid theory and its applications
Published by Springer-Verlag
1981 84. Kinetic Theories and the Boltzmann Equation (LNM 1048)
85. Algebraic Threefolds (LNM 947)
86. Nonlinear Filtering and Stochastic Control (LNM 972)

1982 87. Invariant Theory (LNM 996)

88. Thermodynamics and Constitutive Equations (LNP 228)
89. Fluid Dynamics (LNM 1047)

1983 90. Complete Intersections (LNM 1092)

91. Bifurcation Theory and Applications (LNM 1057)
92. Numerical Methods in Fluid Dynamics (LNM 1127)

1984 93. Harmonic Mappings and Minimal Immersions (LNM 1161)

94. Schrödinger Operators (LNM 1159)
95. Buildings and the Geometry of Diagrams (LNM 1181)

1985 96. Probability and Analysis (LNM 1206)

97. Some Problems in Nonlinear Diffusion (LNM 1224)
98. Theory of Moduli (LNM 1337)

1986 99. Inverse Problems (LNM 1225)

100. Mathematical Economics (LNM 1330)
101. Combinatorial Optimization (LNM 1403)

1987 102. Relativistic Fluid Dynamics (LNM 1385)

103. Topics in Calculus of Variations (LNM 1365)

1988 104. Logic and Computer Science (LNM 1429)

105. Global Geometry and Mathematical Physics (LNM 1451)

1989 106. Methods of nonconvex analysis (LNM 1446)

107. Microlocal Analysis and Applications (LNM 1495)

1990 108. Geometric Topology: Recent Developments (LNM 1504)

109. H∞ Control Theory (LNM 1496)
110. Mathematical Modelling of Industrial Processes (LNM 1521)

1991 111. Topological Methods for Ordinary Differential Equations (LNM 1537)
112. Arithmetic Algebraic Geometry (LNM 1553)
113. Transition to Chaos in Classical and Quantum Mechanics (LNM 1589)

1992 114. Dirichlet Forms (LNM 1563)

115. D-Modules, Representation Theory, and Quantum Groups (LNM 1565)
116. Nonequilibrium Problems in Many-Particle Systems (LNM 1551)

1993 117. Integrable Systems and Quantum Groups (LNM 1620)

118. Algebraic Cycles and Hodge Theory (LNM 1594)
119. Phase Transitions and Hysteresis (LNM 1584)

1994 120. Recent Mathematical Methods in Nonlinear Wave Propagation (LNM 1640)
121. Dynamical Systems (LNM 1609)
122. Transcendental Methods in Algebraic Geometry (LNM 1646)

1995 123. Probabilistic Models for Nonlinear PDE’s (LNM 1627)

124. Viscosity Solutions and Applications (LNM 1660)
125. Vector Bundles on Curves. New Directions (LNM 1649)

1996 126. Integral Geometry, Radon Transforms and Complex Analysis (LNM 1684)
127. Calculus of Variations and Geometric Evolution Problems (LNM 1713)
128. Financial Mathematics (LNM 1656)
1997 129. Mathematics Inspired by Biology (LNM 1714)
130. Advanced Numerical Approximation of Nonlinear Hyperbolic (LNM 1697)
Equations
131. Arithmetic Theory of Elliptic Curves (LNM 1716)
132. Quantum Cohomology (LNM 1776)

1998 133. Optimal Shape Design (LNM 1740)

134. Dynamical Systems and Small Divisors (LNM 1784)
135. Mathematical Problems in Semiconductor Physics (LNM 1823)
136. Stochastic PDE’s and Kolmogorov Equations in Infinite Dimension (LNM 1715)
137. Filtration in Porous Media and Industrial Applications (LNM 1734)

1999 138. Computational Mathematics driven by Industrial Applications (LNM 1739)

139. Iwahori-Hecke Algebras and Representation Theory (LNM 1804)
140. Hamiltonian Dynamics - Theory and Applications (LNM 1861)
141. Global Theory of Minimal Surfaces in Flat Spaces (LNM 1775)
142. Direct and Inverse Methods in Solving Nonlinear Evolution (LNP 632)
Equations

2000 143. Dynamical Systems (LNM 1822)

144. Diophantine Approximation (LNM 1819)
145. Mathematical Aspects of Evolving Interfaces (LNM 1812)
146. Mathematical Methods for Protein Structure (LNCS 2666)
147. Noncommutative Geometry (LNM 1831)

2001 148. Topological Fluid Mechanics to appear

149. Spatial Stochastic Processes (LNM 1802)
150. Optimal Transportation and Applications (LNM 1813)
151. Multiscale Problems and Methods in Numerical Simulations (LNM 1825)

2002 152. Real Methods in Complex and CR Geometry (LNM 1848)

153. Analytic Number Theory (LNM 1891)
154. Inverse Problems and Imaging (LNM 1943)

2003 155. Stochastic Methods in Finance (LNM 1856)

156. Hyperbolic Systems of Balance Laws (LNM 1911)
157. Symplectic 4-Manifolds and Algebraic Surfaces (LNM 1938)
158. Mathematical Foundation of Turbulent Viscous Flows (LNM 1871)

2004 159. Representation Theory and Complex Analysis (LNM 1931)

160. Nonlinear and Optimal Control Theory (LNM 1932)
161. Stochastic Geometry (LNM 1892)

2005 162. Enumerative Invariants in Algebraic Geometry and String Theory to appear
163. Calculus of Variations and Non-linear Partial Differential Equations (LNM 1927)
164. SPDE in Hydrodynamic. Recent Progress and Prospects (LNM 1942)

2006 165. Pseudo-Differential Operators, Quantization and Signals to appear

166. Mixed Finite Elements, Compatibility Conditions, and Applications (LNM 1939)
167. Multiscale Problems in the Life Sciences. From Microscopic (LNM 1940)
to Macroscopic
168. Quantum Transport: Modelling, Analysis and Asymptotics to appear

2007 169. Geometric Analysis and Partial Differential Equations to appear

170. Nonlinear Optimization to appear
171. Arithmetic Geometry to appear

2008 172. Nonlinear Partial Differential Equations and Applications announced

173. Holomorphic Dynamical Systems announced
174. Level Set and PDE based Reconstruction Methods: announced
Applications to Inverse Problems and Image Processing
175. Mathematical models in the manufacturing of glass, polymers announced
and textiles
Lecture Notes in Mathematics
For information about earlier volumes
please contact your bookseller or Springer
LNM Online archive: springerlink.com

Vol. 1755: J. Azéma, M. Émery, M. Ledoux, M. Yor Vol. 1780: J. H. Bruinier, Borcherds Products on O(2,1)
(Eds.), Séminaire de Probabilités XXXV (2001) and Chern Classes of Heegner Divisors (2002)
Vol. 1756: P. E. Zhidkov, Korteweg de Vries and Nonlin- Vol. 1781: E. Bolthausen, E. Perkins, A. van der Vaart,
ear Schrödinger Equations: Qualitative Theory (2001) Lectures on Probability Theory and Statistics. Ecole d’
Vol. 1757: R. R. Phelps, Lectures on Choquet’s Theorem Eté de Probabilités de Saint-Flour XXIX-1999. Editor:
(2001) P. Bernard (2002)
Vol. 1758: N. Monod, Continuous Bounded Cohomology Vol. 1782: C.-H. Chu, A. T.-M. Lau, Harmonic Functions
of Locally Compact Groups (2001) on Groups and Fourier Algebras (2002)
Vol. 1759: Y. Abe, K. Kopfermann, Toroidal Groups Vol. 1783: L. Grüne, Asymptotic Behavior of Dynamical
(2001) and Control Systems under Perturbation and Discretiza-
Vol. 1760: D. Filipović, Consistency Problems for Heath- tion (2002)
Jarrow-Morton Interest Rate Models (2001) Vol. 1784: L. H. Eliasson, S. B. Kuksin, S. Marmi, J.-C.
Vol. 1761: C. Adelmann, The Decomposition of Primes in Yoccoz, Dynamical Systems and Small Divisors. Cetraro,
Torsion Point Fields (2001) Italy 1998. Editors: S. Marmi, J.-C. Yoccoz (2002)
Vol. 1762: S. Cerrai, Second Order PDE’s in Finite and Vol. 1785: J. Arias de Reyna, Pointwise Convergence of
Infinite Dimension (2001) Fourier Series (2002)
Vol. 1763: J.-L. Loday, A. Frabetti, F. Chapoton, F. Goi- Vol. 1786: S. D. Cutkosky, Monomialization of Mor-
chot, Dialgebras and Related Operads (2001) phisms from 3-Folds to Surfaces (2002)
Vol. 1764: A. Cannas da Silva, Lectures on Symplectic Vol. 1787: S. Caenepeel, G. Militaru, S. Zhu, Frobenius
Geometry (2001) and Separable Functors for Generalized Module Cate-
Vol. 1765: T. Kerler, V. V. Lyubashenko, Non-Semisimple gories and Nonlinear Equations (2002)
Topological Quantum Field Theories for 3-Manifolds with Vol. 1788: A. Vasil’ev, Moduli of Families of Curves for
Corners (2001) Conformal and Quasiconformal Mappings (2002)
Vol. 1766: H. Hennion, L. Hervé, Limit Theorems for Vol. 1789: Y. Sommerhäuser, Yetter-Drinfel’d Hopf alge-
Markov Chains and Stochastic Properties of Dynamical bras over groups of prime order (2002)
Systems by Quasi-Compactness (2001)
Vol. 1790: X. Zhan, Matrix Inequalities (2002)
Vol. 1767: J. Xiao, Holomorphic Q Classes (2001)
Vol. 1791: M. Knebusch, D. Zhang, Manis Valuations
Vol. 1768: M. J. Pflaum, Analytic and Geometric Study of
and Prüfer Extensions I: A new Chapter in Commutative
Stratified Spaces (2001)
Algebra (2002)
Vol. 1769: M. Alberich-Carramiñana, Geometry of the
Vol. 1792: D. D. Ang, R. Gorenflo, V. K. Le, D. D. Trong,
Plane Cremona Maps (2002)
Moment Theory and Some Inverse Problems in Potential
Vol. 1770: H. Gluesing-Luerssen, Linear Delay-
Theory and Heat Conduction (2002)
Differential Systems with Commensurate Delays: An
Vol. 1793: J. Cortés Monforte, Geometric, Control and
Algebraic Approach (2002)
Numerical Aspects of Nonholonomic Systems (2002)
Vol. 1771: M. Émery, M. Yor (Eds.), Séminaire de Prob-
abilités 1967-1980. A Selection in Martingale Theory Vol. 1794: N. Pytheas Fogg, Substitution in Dynamics,
(2002) Arithmetics and Combinatorics. Editors: V. Berthé, S. Fer-
enczi, C. Mauduit, A. Siegel (2002)
Vol. 1772: F. Burstall, D. Ferus, K. Leschke, F. Pedit,
U. Pinkall, Conformal Geometry of Surfaces in S4 (2002) Vol. 1795: H. Li, Filtered-Graded Transfer in Using Non-
Vol. 1773: Z. Arad, M. Muzychuk, Standard Integral commutative Gröbner Bases (2002)
Table Algebras Generated by a Non-real Element of Small Vol. 1796: J.M. Melenk, hp-Finite Element Methods for
Degree (2002) Singular Perturbations (2002)
Vol. 1774: V. Runde, Lectures on Amenability (2002) Vol. 1797: B. Schmidt, Characters and Cyclotomic Fields
Vol. 1775: W. H. Meeks, A. Ros, H. Rosenberg, The in Finite Geometry (2002)
Global Theory of Minimal Surfaces in Flat Spaces. Vol. 1798: W.M. Oliva, Geometric Mechanics (2002)
Martina Franca 1999. Editor: G. P. Pirola (2002) Vol. 1799: H. Pajot, Analytic Capacity, Rectifiability,
Vol. 1776: K. Behrend, C. Gomez, V. Tarasov, G. Tian, Menger Curvature and the Cauchy Integral (2002)
Quantum Comohology. Cetraro 1997. Editors: P. de Bar- Vol. 1800: O. Gabber, L. Ramero, Almost Ring Theory
tolomeis, B. Dubrovin, C. Reina (2002) (2003)
Vol. 1777: E. García-Río, D. N. Kupeli, R. Vázquez- Vol. 1801: J. Azéma, M. Émery, M. Ledoux, M. Yor
Lorenzo, Osserman Manifolds in Semi-Riemannian (Eds.), Séminaire de Probabilités XXXVI (2003)
Geometry (2002) Vol. 1802: V. Capasso, E. Merzbach, B. G. Ivanoff,
Vol. 1778: H. Kiechle, Theory of K-Loops (2002) M. Dozzi, R. Dalang, T. Mountford, Topics in Spatial
Vol. 1779: I. Chueshov, Monotone Random Systems Stochastic Processes. Martina Franca, Italy 2001. Editor:
(2002) E. Merzbach (2003)
Vol. 1803: G. Dolzmann, Variational Methods for Crys- Vol. 1827: K. M. Pilgrim, Combinations of Complex
talline Microstructure – Analysis and Computation (2003) Dynamical Systems. IX, 118 p, 2003.
Vol. 1804: I. Cherednik, Ya. Markov, R. Howe, G. Lusztig, Vol. 1828: D. J. Green, Gröbner Bases and the Computa-
Iwahori-Hecke Algebras and their Representation Theory. tion of Group Cohomology. XII, 138 p, 2003.
Martina Franca, Italy 1999. Editors: V. Baldoni, D. Bar- Vol. 1829: E. Altman, B. Gaujal, A. Hordijk, Discrete-
basch (2003) Event Control of Stochastic Networks: Multimodularity
Vol. 1805: F. Cao, Geometric Curve Evolution and Image and Regularity. XIV, 313 p, 2003.
Processing (2003) Vol. 1830: M. I. Gil’, Operator Functions and Localization
Vol. 1806: H. Broer, I. Hoveijn. G. Lunther, G. Vegter, of Spectra. XIV, 256 p, 2003.
Bifurcations in Hamiltonian Systems. Computing Singu- Vol. 1831: A. Connes, J. Cuntz, E. Guentner, N. Hig-
larities by Gröbner Bases (2003) son, J. E. Kaminker, Noncommutative Geometry, Martina
Vol. 1807: V. D. Milman, G. Schechtman (Eds.), Geomet- Franca, Italy 2002. Editors: S. Doplicher, L. Longo (2004)
ric Aspects of Functional Analysis. Israel Seminar 2000- Vol. 1832: J. Azéma, M. Émery, M. Ledoux, M. Yor
2002 (2003) (Eds.), Séminaire de Probabilités XXXVII (2003)
Vol. 1808: W. Schindler, Measures with Symmetry Prop- Vol. 1833: D.-Q. Jiang, M. Qian, M.-P. Qian, Mathemati-
erties (2003) cal Theory of Nonequilibrium Steady States. On the Fron-
Vol. 1809: O. Steinbach, Stability Estimates for Hybrid tier of Probability and Dynamical Systems. IX, 280 p,
Coupled Domain Decomposition Methods (2003) 2004.
Vol. 1810: J. Wengenroth, Derived Functors in Functional Vol. 1834: Yo. Yomdin, G. Comte, Tame Geometry with
Analysis (2003) Application in Smooth Analysis. VIII, 186 p, 2004.
Vol. 1811: J. Stevens, Deformations of Singularities Vol. 1835: O.T. Izhboldin, B. Kahn, N.A. Karpenko,
(2003) A. Vishik, Geometric Methods in the Algebraic Theory
Vol. 1812: L. Ambrosio, K. Deckelnick, G. Dziuk, of Quadratic Forms. Summer School, Lens, 2000. Editor:
M. Mimura, V. A. Solonnikov, H. M. Soner, Mathematical J.-P. Tignol (2004)
Aspects of Evolving Interfaces. Madeira, Funchal, Portu- Vol. 1836: C. Nǎstǎsescu, F. Van Oystaeyen, Methods of
gal 2000. Editors: P. Colli, J. F. Rodrigues (2003) Graded Rings. XIII, 304 p, 2004.
Vol. 1813: L. Ambrosio, L. A. Caffarelli, Y. Brenier, Vol. 1837: S. Tavaré, O. Zeitouni, Lectures on Probabil-
G. Buttazzo, C. Villani, Optimal Transportation and its ity Theory and Statistics. Ecole d’Eté de Probabilités de
Applications. Martina Franca, Italy 2001. Editors: L. A. Saint-Flour XXXI-2001. Editor: J. Picard (2004)
Caffarelli, S. Salsa (2003) Vol. 1838: A.J. Ganesh, N.W. O’Connell, D.J. Wischik,
Vol. 1814: P. Bank, F. Baudoin, H. Föllmer, L.C.G. Big Queues. XII, 254 p, 2004.
Rogers, M. Soner, N. Touzi, Paris-Princeton Lectures on Vol. 1839: R. Gohm, Noncommutative Stationary
Mathematical Finance 2002 (2003) Processes. VIII, 170 p, 2004.
Vol. 1815: A. M. Vershik (Ed.), Asymptotic Com- Vol. 1840: B. Tsirelson, W. Werner, Lectures on Probabil-
binatorics with Applications to Mathematical Physics. ity Theory and Statistics. Ecole d’Eté de Probabilités de
St. Petersburg, Russia 2001 (2003) Saint-Flour XXXII-2002. Editor: J. Picard (2004)
Vol. 1816: S. Albeverio, W. Schachermayer, M. Tala- Vol. 1841: W. Reichel, Uniqueness Theorems for Vari-
grand, Lectures on Probability Theory and Statistics. ational Problems by the Method of Transformation
Ecole d’Eté de Probabilités de Saint-Flour XXX-2000. Groups (2004)
Editor: P. Bernard (2003) Vol. 1842: T. Johnsen, A. L. Knutsen, K3 Projective Mod-
Vol. 1817: E. Koelink, W. Van Assche (Eds.), Orthogonal els in Scrolls (2004)
Polynomials and Special Functions. Leuven 2002 (2003) Vol. 1843: B. Jefferies, Spectral Properties of Noncom-
Vol. 1818: M. Bildhauer, Convex Variational Problems muting Operators (2004)
with Linear, nearly Linear and/or Anisotropic Growth Vol. 1844: K.F. Siburg, The Principle of Least Action in
Conditions (2003) Geometry and Dynamics (2004)
Vol. 1819: D. Masser, Yu. V. Nesterenko, H. P. Schlick- Vol. 1845: Min Ho Lee, Mixed Automorphic Forms, Torus
ewei, W. M. Schmidt, M. Waldschmidt, Diophantine Bundles, and Jacobi Forms (2004)
Approximation. Cetraro, Italy 2000. Editors: F. Amoroso, Vol. 1846: H. Ammari, H. Kang, Reconstruction of Small
U. Zannier (2003) Inhomogeneities from Boundary Measurements (2004)
Vol. 1820: F. Hiai, H. Kosaki, Means of Hilbert Space Vol. 1847: T.R. Bielecki, T. Björk, M. Jeanblanc, M.
Operators (2003) Rutkowski, J.A. Scheinkman, W. Xiong, Paris-Princeton
Vol. 1821: S. Teufel, Adiabatic Perturbation Theory in Lectures on Mathematical Finance 2003 (2004)
Quantum Dynamics (2003) Vol. 1848: M. Abate, J. E. Fornaess, X. Huang, J. P. Rosay,
Vol. 1822: S.-N. Chow, R. Conti, R. Johnson, J. Mallet- A. Tumanov, Real Methods in Complex and CR Geom-
Paret, R. Nussbaum, Dynamical Systems. Cetraro, Italy etry, Martina Franca, Italy 2002. Editors: D. Zaitsev, G.
2000. Editors: J. W. Macki, P. Zecca (2003) Zampieri (2004)
Vol. 1823: A. M. Anile, W. Allegretto, C. Ring- Vol. 1849: Martin L. Brown, Heegner Modules and Ellip-
hofer, Mathematical Problems in Semiconductor Physics. tic Curves (2004)
Cetraro, Italy 1998. Editor: A. M. Anile (2003) Vol. 1850: V. D. Milman, G. Schechtman (Eds.), Geomet-
Vol. 1824: J. A. Navarro González, J. B. Sancho de Salas, ric Aspects of Functional Analysis. Israel Seminar 2002-
C ∞ – Differentiable Spaces (2003) 2003 (2004)
Vol. 1825: J. H. Bramble, A. Cohen, W. Dahmen, Mul- Vol. 1851: O. Catoni, Statistical Learning Theory and
tiscale Problems and Methods in Numerical Simulations, Stochastic Optimization (2004)
Martina Franca, Italy 2001. Editor: C. Canuto (2003) Vol. 1852: A.S. Kechris, B.D. Miller, Topics in Orbit
Vol. 1826: K. Dohmen, Improved Bonferroni Inequal- Equivalence (2004)
ities via Abstract Tubes. Inequalities and Identities of Vol. 1853: Ch. Favre, M. Jonsson, The Valuative Tree
Inclusion-Exclusion Type. VIII, 113 p, 2003. (2004)
Vol. 1854: O. Saeki, Topology of Singular Fibers of Dif- Vol. 1878: R. Cerf, The Wulff Crystal in Ising and Percol-
ferential Maps (2004) ation Models, Ecole d’Eté de Probabilités de Saint-Flour
Vol. 1855: G. Da Prato, P.C. Kunstmann, I. Lasiecka, XXXIV-2004. Editor: Jean Picard (2006)
A. Lunardi, R. Schnaubelt, L. Weis, Functional Analytic Vol. 1879: G. Slade, The Lace Expansion and its Applica-
Methods for Evolution Equations. Editors: M. Iannelli, tions, Ecole d’Eté de Probabilités de Saint-Flour XXXIV-
R. Nagel, S. Piazzera (2004) 2004. Editor: Jean Picard (2006)
Vol. 1856: K. Back, T.R. Bielecki, C. Hipp, S. Peng, Vol. 1880: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
W. Schachermayer, Stochastic Methods in Finance, Bres- Systems I, The Hamiltonian Approach (2006)
sanone/Brixen, Italy, 2003. Editors: M. Fritelli, W. Rung- Vol. 1881: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
galdier (2004) Systems II, The Markovian Approach (2006)
Vol. 1857: M. Émery, M. Ledoux, M. Yor (Eds.), Sémi- Vol. 1882: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
naire de Probabilités XXXVIII (2005) Systems III, Recent Developments (2006)
Vol. 1858: A.S. Cherny, H.-J. Engelbert, Singular Stochas- Vol. 1883: W. Van Assche, F. Marcellàn (Eds.), Orthogo-
tic Differential Equations (2005) nal Polynomials and Special Functions, Computation and
Application (2006)
Vol. 1859: E. Letellier, Fourier Transforms of Invariant
Functions on Finite Reductive Lie Algebras (2005) Vol. 1884: N. Hayashi, E.I. Kaikina, P.I. Naumkin,
I.A. Shishmarev, Asymptotics for Dissipative Nonlinear
Vol. 1860: A. Borisyuk, G.B. Ermentrout, A. Friedman,
Equations (2006)
D. Terman, Tutorials in Mathematical Biosciences I.
Vol. 1885: A. Telcs, The Art of Random Walks (2006)
Mathematical Neurosciences (2005)
Vol. 1886: S. Takamura, Splitting Deformations of Dege-
Vol. 1861: G. Benettin, J. Henrard, S. Kuksin, Hamil- nerations of Complex Curves (2006)
tonian Dynamics – Theory and Applications, Cetraro, Vol. 1887: K. Habermann, L. Habermann, Introduction to
Italy, 1999. Editor: A. Giorgilli (2005) Symplectic Dirac Operators (2006)
Vol. 1862: B. Helffer, F. Nier, Hypoelliptic Estimates and Vol. 1888: J. van der Hoeven, Transseries and Real Differ-
Spectral Theory for Fokker-Planck Operators and Witten ential Algebra (2006)
Laplacians (2005) Vol. 1889: G. Osipenko, Dynamical Systems, Graphs, and
Vol. 1863: H. Führ, Abstract Harmonic Analysis of Con- Algorithms (2006)
tinuous Wavelet Transforms (2005) Vol. 1890: M. Bunge, J. Funk, Singular Coverings of
Vol. 1864: K. Efstathiou, Metamorphoses of Hamiltonian Toposes (2006)
Systems with Symmetries (2005) Vol. 1891: J.B. Friedlander, D.R. Heath-Brown,
Vol. 1865: D. Applebaum, B.V. R. Bhat, J. Kustermans, H. Iwaniec, J. Kaczorowski, Analytic Number Theory,
J. M. Lindsay, Quantum Independent Increment Processes Cetraro, Italy, 2002. Editors: A. Perelli, C. Viola (2006)
I. From Classical Probability to Quantum Stochastic Cal- Vol. 1892: A. Baddeley, I. Bárány, R. Schneider, W. Weil,
culus. Editors: M. Schürmann, U. Franz (2005) Stochastic Geometry, Martina Franca, Italy, 2004. Editor:
Vol. 1866: O.E. Barndorff-Nielsen, U. Franz, R. Gohm, W. Weil (2007)
B. Kümmerer, S. Thorbjønsen, Quantum Independent Vol. 1893: H. Hanßmann, Local and Semi-Local Bifur-
Increment Processes II. Structure of Quantum Lévy cations in Hamiltonian Dynamical Systems, Results and
Processes, Classical Probability, and Physics. Editors: M. Examples (2007)
Schürmann, U. Franz, (2005) Vol. 1894: C.W. Groetsch, Stable Approximate Evaluation
Vol. 1867: J. Sneyd (Ed.), Tutorials in Mathematical Bio- of Unbounded Operators (2007)
sciences II. Mathematical Modeling of Calcium Dynamics Vol. 1895: L. Molnár, Selected Preserver Problems on
and Signal Transduction. (2005) Algebraic Structures of Linear Operators and on Function
Vol. 1868: J. Jorgenson, S. Lang, Posn (R) and Eisenstein Spaces (2007)
Series. (2005) Vol. 1896: P. Massart, Concentration Inequalities and
Vol. 1869: A. Dembo, T. Funaki, Lectures on Probabil- Model Selection, Ecole d’Été de Probabilités de Saint-
ity Theory and Statistics. Ecole d’Eté de Probabilités de Flour XXXIII-2003. Editor: J. Picard (2007)
Saint-Flour XXXIII-2003. Editor: J. Picard (2005) Vol. 1897: R. Doney, Fluctuation Theory for Lévy
Vol. 1870: V.I. Gurariy, W. Lusky, Geometry of Müntz Processes, Ecole d’Été de Probabilités de Saint-Flour
Spaces and Related Questions. (2005) XXXV-2005. Editor: J. Picard (2007)
Vol. 1898: H.R. Beyer, Beyond Partial Differential Equa-
Vol. 1871: P. Constantin, G. Gallavotti, A.V. Kazhikhov,
tions, On linear and Quasi-Linear Abstract Hyperbolic
Y. Meyer, S. Ukai, Mathematical Foundation of Turbu-
Evolution Equations (2007)
lent Viscous Flows, Martina Franca, Italy, 2003. Editors:
Vol. 1899: Séminaire de Probabilités XL. Editors:
M. Cannone, T. Miyakawa (2006)
C. Donati-Martin, M. Émery, A. Rouault, C. Stricker
Vol. 1872: A. Friedman (Ed.), Tutorials in Mathemati- (2007)
cal Biosciences III. Cell Cycle, Proliferation, and Cancer
Vol. 1900: E. Bolthausen, A. Bovier (Eds.), Spin Glasses
(2006)
(2007)
Vol. 1873: R. Mansuy, M. Yor, Random Times and En- Vol. 1901: O. Wittenberg, Intersections de deux
largements of Filtrations in a Brownian Setting (2006) quadriques et pinceaux de courbes de genre 1, Inter-
Vol. 1874: M. Yor, M. Émery (Eds.), In Memoriam Paul- sections of Two Quadrics and Pencils of Curves of Genus
André Meyer - Séminaire de Probabilités XXXIX (2006) 1 (2007)
Vol. 1875: J. Pitman, Combinatorial Stochastic Processes. Vol. 1902: A. Isaev, Lectures on the Automorphism
Ecole d’Eté de Probabilités de Saint-Flour XXXII-2002. Groups of Kobayashi-Hyperbolic Manifolds (2007)
Editor: J. Picard (2006) Vol. 1903: G. Kresin, V. Maz’ya, Sharp Real-Part Theo-
Vol. 1876: H. Herrlich, Axiom of Choice (2006) rems (2007)
Vol. 1877: J. Steuding, Value Distributions of L-Functions Vol. 1904: P. Giesl, Construction of Global Lyapunov
(2007) Functions Using Radial Basis Functions (2007)
 Vol. 1905: C. Prévôt, M. Röckner, A Concise Course on Vol. 1931: M. Cowling, E. Frenkel, M. Kashiwara,
Stochastic Partial Differential Equations (2007) A. Valette, D.A. Vogan, Jr., N.R. Wallach, Representation
Vol. 1906: T. Schuster, The Method of Approximate Theory and Complex Analysis. Venice, Italy 2004.
Inverse: Theory and Applications (2007) Editors: E.C. Tarabusi, A. D’Agnolo, M. Picardello
Vol. 1907: M. Rasmussen, Attractivity and Bifurcation for (2008)
Nonautonomous Dynamical Systems (2007) Vol. 1932: A.A. Agrachev, A.S. Morse, E.D. Sontag,
Vol. 1908: T.J. Lyons, M. Caruana, T. Lévy, Differential H.J. Sussmann, V.I. Utkin, Nonlinear and Optimal Control
Equations Driven by Rough Paths, Ecole d’Été de Proba- Theory. Cetraro, Italy 2004. Editors: P. Nistri, G. Stefani
bilités de Saint-Flour XXXIV-2004 (2007) (2008)
Vol. 1909: H. Akiyoshi, M. Sakuma, M. Wada, Vol. 1933: M. Petkovic, Point Estimation of Root Finding
Y. Yamashita, Punctured Torus Groups and 2-Bridge Knot Methods (2008)
Groups (I) (2007) Vol. 1934: C. Donati-Martin, M. Émery, A. Rouault,
Vol. 1910: V.D. Milman, G. Schechtman (Eds.), Geo- C. Stricker (Eds.), Séminaire de Probabilités XLI (2008)
metric Aspects of Functional Analysis. Israel Seminar Vol. 1935: A. Unterberger, Alternative Pseudodifferential
2004-2005 (2007) Analysis (2008)
Vol. 1911: A. Bressan, D. Serre, M. Williams, Vol. 1936: P. Magal, S. Ruan (Eds.), Structured Population
K. Zumbrun, Hyperbolic Systems of Balance Laws. Models in Biology and Epidemiology (2008)
Cetraro, Italy 2003. Editor: P. Marcati (2007) Vol. 1937: G. Capriz, P. Giovine, P.M. Mariano (Eds.),
Vol. 1912: V. Berinde, Iterative Approximation of Fixed Mathematical Models of Granular Matter (2008)
Points (2007) Vol. 1938: D. Auroux, F. Catanese, M. Manetti, P. Seidel,
Vol. 1913: J.E. Marsden, G. Misiołek, J.-P. Ortega, B. Siebert, I. Smith, G. Tian, Symplectic 4-Manifolds
M. Perlmutter, T.S. Ratiu, Hamiltonian Reduction by and Algebraic Surfaces. Cetraro, Italy 2003. Editors:
Stages (2007) F. Catanese, G. Tian (2008)
Vol. 1939: D. Boffi, F. Brezzi, L. Demkowicz, R.G. Durán,
Vol. 1914: G. Kutyniok, Affine Density in Wavelet
R.S. Falk, M. Fortin, Mixed Finite Elements, Compati-
Analysis (2007)
bility Conditions, and Applications. Cetraro, Italy 2006.
Vol. 1915: T. Bıyıkoǧlu, J. Leydold, P.F. Stadler, Laplacian
Editors: D. Boffi, L. Gastaldi (2008)
Eigenvectors of Graphs. Perron-Frobenius and Faber-
Vol. 1940: J. Banasiak, V. Capasso, M.A.J. Chaplain,
Krahn Type Theorems (2007)
M. Lachowicz, J. Miȩkisz, Multiscale Problems in the Life
Vol. 1916: C. Villani, F. Rezakhanlou, Entropy Methods Sciences. From Microscopic to Macroscopic. Bȩdlewo,
for the Boltzmann Equation. Editors: F. Golse, S. Olla Poland 2006. Editors: V. Capasso, M. Lachowicz (2008)
(2008) Vol. 1941: S.M.J. Haran, Arithmetical Investigations.
Vol. 1917: I. Veselić, Existence and Regularity Properties Representation Theory, Orthogonal Polynomials, and
of the Integrated Density of States of Random Schrödinger Quantum Interpolations (2008)
(2008) Vol. 1942: S. Albeverio, F. Flandoli, Y.G. Sinai, SPDE in
Vol. 1918: B. Roberts, R. Schmidt, Local Newforms for Hydrodynamic. Recent Progress and Prospects. Cetraro,
GSp(4) (2007) Italy 2005. Editors: G. Da Prato, M. Röckner (2008)
Vol. 1919: R.A. Carmona, I. Ekeland, A. Kohatsu- Vol. 1943: L.L. Bonilla (Ed.), Inverse Problems and Imag-
Higa, J.-M. Lasry, P.-L. Lions, H. Pham, E. Taflin, ing. Martina Franca, Italy 2002 (2008)
Paris-Princeton Lectures on Mathematical Finance 2004. Vol. 1944: A. Di Bartolo, G. Falcone, P. Plaumann,
Editors: R.A. Carmona, E. Çinlar, I. Ekeland, E. Jouini, K. Strambach, Algebraic Groups and Lie Groups with
J.A. Scheinkman, N. Touzi (2007) Few Factors (2008)
Vol. 1920: S.N. Evans, Probability and Real Trees. Ecole
d’Été de Probabilités de Saint-Flour XXXV-2005 (2008)
Vol. 1921: J.P. Tian, Evolution Algebras and their Appli-
cations (2008)
Vol. 1922: A. Friedman (Ed.), Tutorials in Mathematical
BioSciences IV. Evolution and Ecology (2008)
Vol. 1923: J.P.N. Bishwal, Parameter Estimation in
Stochastic Differential Equations (2008)
Recent Reprints and New Editions
Vol. 1924: M. Wilson, Littlewood-Paley Theory and Vol. 1702: J. Ma, J. Yong, Forward-Backward Stochas-
Exponential-Square Integrability (2008) tic Differential Equations and their Applications. 1999 –
Vol. 1925: M. du Sautoy, L. Woodward, Zeta Functions of Corr. 3rd printing (2007)
Groups and Rings (2008) Vol. 830: J.A. Green, Polynomial Representations of GLn ,
Vol. 1926: L. Barreira, V. Claudia, Stability of Nonauto- with an Appendix on Schensted Correspondence and Lit-
nomous Differential Equations (2008) telmann Paths by K. Erdmann, J.A. Green and M. Schoker
Vol. 1927: L. Ambrosio, L. Caffarelli, M.G. Crandall, 1980 – 2nd corr. and augmented edition (2007)
L.C. Evans, N. Fusco, Calculus of Variations and Non- Vol. 1693: S. Simons, From Hahn-Banach to Monotonic-
Linear Partial Differential Equations. Cetraro, Italy 2005. ity (Minimax and Monotonicity 1998) – 2nd exp. edition
Editors: B. Dacorogna, P. Marcellini (2008) (2008)
Vol. 1928: J. Jonsson, Simplicial Complexes of Graphs Vol. 470: R.E. Bowen, Equilibrium States and the Ergodic
(2008) Theory of Anosov Diffeomorphisms. With a preface by
Vol. 1929: Y. Mishura, Stochastic Calculus for Fractional D. Ruelle. Edited by J.-R. Chazottes. 1975 – 2nd rev.
Brownian Motion and Related Processes (2008) edition (2008)
Vol. 1930: J.M. Urbano, The Method of Intrinsic Scaling. Vol. 523: S.A. Albeverio, R.J. Høegh-Krohn,
A Systematic Approach to Regularity for Degenerate and S. Mazzucchi, Mathematical Theory of Feynman Path
Singular PDEs (2008) Integral. 1976 – 2nd corr. and enlarged edition (2008)

View publication stats