The IMA Volumes

in Mathematics
and its Applications

Volume 148

Series Editors
Douglas N. Arnold Arnd Scheel
 Institute for Mathematics and
its Applications (IMA)
The Institute for Mathematics and its Applications was established by a
grant from the National Science Foundation to the University of Minnesota
in 1982. The primary mission of the IMA is to foster research of a truly inter-
disciplinary nature, establishing links between mathematics of the highest
caliber and important scientific and technological problems from other dis-
ciplines and industries. To this end, the IMA organizes a wide variety of pro-
grams, ranging from short intense workshops in areas of exceptional interest
and opportunity to extensive thematic programs lasting a year. IMA Volumes
are used to communicate results of these programs that we believe are of
particular value to the broader scientific community.
The full list of IMA books can be found at the Web site of the Institute
for Mathematics and its Applications:
http://www.ima.umn.edu/springer/volumes.html
Presentation materials from the IMA talks are available at
http://www.ima.umn.edu/talks/
Douglas N. Arnold, Director of the IMA

* * * * * * * * * *

IMA ANNUAL PROGRAMS

1982–1983 Statistical and Continuum Approaches to Phase Transition

1983–1984 Mathematical Models for the Economics of Decentralized
Resource Allocation
1984–1985 Continuum Physics and Partial Differential Equations
1985–1986 Stochastic Differential Equations and Their Applications
1986–1987 Scientific Computation
1987–1988 Applied Combinatorics
1988–1989 Nonlinear Waves
1989–1990 Dynamical Systems and Their Applications
1990–1991 Phase Transitions and Free Boundaries
1991–1992 Applied Linear Algebra
1992–1993 Control Theory and its Applications
1993–1994 Emerging Applications of Probability
1994–1995 Waves and Scattering
1995–1996 Mathematical Methods in Material Science
1996–1997 Mathematics of High Performance Computing

(Continued at the back)

 Michael Stillman Nobuki Takayama
Jan Verschelde
Editors

Software for Algebraic

Geometry
Editors
Michael Stillman Nobuki Takayama
Department of Mathematics Department of Mathematics
Cornell University Kobe University
Ithaca, NY 14853-4201 Rokko, Kobe 657-8501
USA Japan

Jan Verschelde
Department of Mathematics
University of Illinois
Chicago, IL 60607-7045
USA

Series Editors
Douglas N. Arnold
Arnd Scheel
Institute for Mathematics and its
Applications
University of Minnesota
Minneapolis, MN 55455
USA

ISBN: 978-0-387-78132-7 e-ISBN: 978-0-387-78133-4

DOI: 10.1007/978-0-387-78133-4

Library of Congress Control Number: 2008923357

Mathematics Subject Classification (2000): 11R09, 11Y99, 12D05, 13P10, 14P05, 14Q05, 14Q10
52A39, 52B20, 52B55, 65D18, 65F15, 65F20, 65F22, 65H10, 65H20, 65-04, 91-08

© 2008 Springer Science + Business Media, LLC

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science + Business Media, LLC, 233 Spring Street, New York,
NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in
connection with any form of information storage and retrieval, electronic adaptation, computer
software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they
are not identified as such, is not to be taken as an expression of opinion as to whether or not they are
subject to proprietary rights.

Camera-ready copy provided by the IMA.

9 8 7 6 5 4 3 2 1

springer.com
 FOREWORD

This IMA Volume in Mathematics and its AppUcations

SOFTWARE FOR A L G E B R A I C GEOMETRY

contains papers presented at a highly successful one-week workshop on the

same title. The event was an integral part of the 2006-2007 IMA Thematic
Year on "Applications of Algebraic Geometry." We are grateful to all the
participants for making this workshop a very productive and stimulating
event. Special thanks to Michael E. Stillman (Department of Mathematics,
Cornell University), Nobuki Takayama (Department of Mathematics, Kobe
University), and Jan Verschelde (Department of Mathematics, Statistics
and Computer Science, University of Illinois at Chicago) for their superb
role as workshop organizers and editors of these proceedings.
We take this opportunity to thank the National Science Foundation
for its support of the IMA.

Series Editors
Douglas N. Arnold, Director of the IMA
Arnd Scheel, Deputy Director of the IMA
 PREFACE

The workshop on "Software for Algebraic Geometry" was held in the

week from 23 to 27 October 2006, as the second workshop in the thematic
year on Applications of Algebraic Geometry at the IMA.
Algorithms in algebraic geometry go hand in hand with software pack-
ages that implement them. Together they have established the modern field
of computational algebraic geometry which has come to play a major role
in both theoretical advances and applications. Over the past fifteen years,
several excellent general purpose packages for computations in algebraic
geometry have been developed, such as CoCoA, Singular and Macaulay 2.
While these packages evolve continuously, incorporating new mathematical
advances, they both motivate and demand the creation of new mathematics
and smarter algorithms.
Surrounding the general packages, a host of specialized packages for
dedicated and focused computations have created a platform for the in-
teraction of algebraic geometry with numerous other areas of mathemat-
ics including optimization, combinatorics, polyhedral geometry, numerical
analysis and computer science. The workshop brought together a wide
array of theoreticians and practitioners interested in the development of
algorithms and software in algebraic geometry at this workshop. Such in-
teractions are essential for dramatic increases in the power and appUcability
of algorithms in the field.
There were 89 registered participants at the workshop. At four talks
a day, 20 regular 50 minutes talks were scheduled. On Monday evening,
10 posters were presented. On Wednesday and Thursday evening we had
respectively 5 and 6 software demonstrations. The Ust of featured software
packages includes Macaulay 2, SAGE, HomLab, Bertini, APAtools, PH-
Clab, PHCmaple, PHCpack, KNOPPIX/Math, D-modules for Macaulay 2,
Singular, Risa/Asir, CRACK, diffalg, RIFsimp, Gambit, FGb/RS, Co-
CoALib, 4ti2, PHoMpara, SYNAPS, DEMiCs, Magma, Kronecker, SOS-
TOOLS, Gfan, Maple 11.
The IMA systems group had installed many of these programs on the
computers at the IMA. At the poster session, the participants were given
the opportunity to install the featured software systems on their laptop.
A demonstration cluster computer of Rocketcalc was running during the
poster session and accessible to all participants during the workshop.
The evening before the workshop dinner on Tuesday started with a
problem session. Prior to this session we made a list of problem descriptions
available on the web site. The workshop ended on Friday evening with some
additional problems, discussion on the posted problems, and a presentation
of Jiawang Nie about the application to semidefinite programming to solve
systems of algebraic equations which arise from differential equations. We
vui PREFACE

are also happy that several new research projects were stimulated by this
problem session. Some results are going to appear elsewhere.
Instead of the "second chances" (usual for IMA workshops), the partic-
ipants were given the opportunity to test the software systems on Wednes-
day and Thursday evening. The evening session started with a one hour
plenary session, where each software system on demo in the evening was
briefly explained. Following this plenary session, the participants moved
to the 4th iioor of Lind Hall, to experience the software systems on the
computers in the open poster area, or in parallel, in the classroom 409.
The IMA systems group worked hard in the weeks leading up to the
workshop to install the software systems. Their effort benefited not only the
workshop participants, but all all subsequent participants to the thematic
year, as they found their workstations equipped with the latest software
tools in algebraic geometry.
The papers in this volume describe the software packages Bertini, PH-
Clab, Gfan, DEMiCs, SYNAPS, Trim, Gambit, ApaTools, and the appli-
cation of Risa/Asir to a conjecture on multiple zeta values. We thank the
participants to the workshop, the authors and the anonymous referees. We
are grateful to the editorial staff of the IMA, Patricia V. Brick and Dzung
N. Nguyen, for their dedication and care.

Michael E. Stillman
Department of Mathematics
Cornell University
http://www.math.cornell.edu/People/Faculty/stillman.html

Nobuki Takayama
Department of Mathematics
Kobe University
http://www.math.sci.kobe-u.ac.jp/ taka/

Jan Verschelde
Department of Mathematics, Statistics ajid Computer Science
University of Illinois at Chicago
http://www2.math.uic.edu/ jan/
 CONTENTS

Foreword v
Preface vii

Software for numerical algebraic geometry: A paradigm

and progress towards its implementation 1
Daniel J. Bates, Jonathan D. Hauenstein,
Andrew J. Sommese, and Charles W. Wampler II

PHClab: A MATLAB/Octave interface to PHCpack 15

Yun Guan and Jan Verschelde

Computing Grobner fans and tropical varieties in Gfan 33

Anders Nedergaard Jensen

On a conjecture for the dimension of the space of

the multiple zeta values 47
Masanobu Kaneko, Masayuki Noro,
and Ken'ichi Tsurumaki

DEMiCs: A software package for computing the mixed

volume via d5Tiamic enumeration of all mixed cells 59
Tomohiko Mizutani and Akiko Takeda

SYNAPS, a library for dedicated applications in

symbolic numeric computing 81
Bernard Mourrain, Jean-Pascal Pavone,
Philippe Trebuchet, Elias P. Tsigaridas, and Julien Wintz

Tropical impUcitization and mixed fiber polytopes Ill

Bemd Sturmfels and Josephine Yu

Towards a black-box solver for finite games:

Computing all equilibria with Gambit and PHCpack 133
Theodore L. Turocy

ApaTools: A software toolbox for approximate

polynomial algebra 149
Zhonggang Zeng

List of workshop participants 169

ix
SOFTWARE FOR NUMERICAL ALGEBRAIC GEOMETRY:
A P A R A D I G M A N D P R O G R E S S T O W A R D S ITS
IMPLEMENTATION
DANIEL J. BATES*, JONATHAN D. HAUENSTEINt,
ANDREW J. SOMMESE*, AND CHARLES W. WAMPLER n§

Abstract. Though numerical methods to find all the isolated solutions of nonllnestr
systems of multivariate polynomials go back 30 years, it is only over the last decade that
numerical methods have been devised for the computation and manipulation of algebraic
sets coming from polynomJEil systems over the complex numbers. Collectively, these
algorithms and the underlying theory have come to be known as numerical algebraic
geometry. Several software packages are capable of carrying out some of the operations
of numerical algebraic geometry, although no one package provides all such capabilities.
This paper contains an enumeration of the operations that an ideal software package
in this field would allow. The current and upcoming capabilities of Bertini, the most
recently released package in this field, are also described.

Key words. Homotopy continuation, numerical algebrjiic geometry, polynomial

systems, software, Bertini.

AMS(MOS) subject classifications. 65H10, 65H20, 65-04, 14Q99.

1. Introduction. Numerical algebraic geometry refers to the appli-

cation of numerical methods to compute the solution sets of polynomial
systems, generally over C. In particular, basic numerical algebraic geome-
try embodies probability one algorithms for computing all isolated solutions
of a polynomial system as well as the numerical irreducible decomposition
of an algebraic set, i.e., one or more points on each irreducible component
in each dimension. More recently, numerical algebraic geometry has grown
to include more advanced techniques which make use of the basic methods
in order to compute data of interest in both real-world applications and
pure algebraic geometry.

'Institute for Mathematics and its Applications, University of Minnesota, Minneapo-

lis, MN 55122 ([email protected], www.nd.edu/~dbatesl). This author was supported by
the Duncan Chair of the University of Notre Dame, the University of Notre Dame, NSF
grant DMS-0410047, the Arthur J. Schmitt Foundation, and the Institute for Mathe-
matics and its Applications in Minneapolis (IMA).
* Department of Mathematics, University of Notre Dame, Notre Dame, IN 46556
([email protected], www.nd.edu/~jhauenst). This author was supported by the Duncan
Chair of the University of Notre Dame, NSF grant DMS-0410047, and the Institute for
Mathematics and its Applications in Minneapolis (IMA).
'•Department of Mathematics, University of Notre Dame, Notre Dame, IN 46556
([email protected], www.nd.edu/~somm«e). This author was supported by the Duncan
Chair of the University of Notre Dame, the University of Notre Dame, NSF grant DMS-
0410047, and the Institute for Mathematics and its Applications in Minneapolis (IMA).
^General Motors Research and Development, Mail Code 480-106-359, 30500 Mound
Road, Warren, MI 48090 ([email protected], www.nd.edu/~cwamplel).
This author was supported % NSF grant DMS-0410047 and the Insstitute for Math-
ematics and its Applications in Minneapolis (IMA).
2 DANIEL J. BATES ET AL.

One of the key tools used in the algorithms of numerical algebraic

geometry is homotopy continuation [1, 16], a method for finding all zero-
dimensional solutions of a polynomial system. Given a polynomial system
/ : C ^ —» C" to be solved by homotopy continuation, one first forms a poly-
nomial system g that is related to / in a prescribed way but has known, or
easily computable, solutions. The systems g and / are combined to form
a homotopy, such as the linear homotopy H{x,t) — f • {1 — t) +'y • t • g
where 7 e C is randomly chosen. For a properly formed homotopy, there
are continuous solution paths leading from the solutions of g to those of
/ which may be followed using predictor-corrector methods. Singular
solutions cause numerical diflttculties, so singular endgames [17, 18, 19]
are typically employed. Zero-dimensional solving is discussed further in
Section 2.2.
Numerical algebraic geometry treats both zero-dimensional (isolated)
solutions and jwsitive dimensional solution sets. The building blocks of the
solution set of a set of equations are the irreducible components, i.e., the
algebraic subsets of the solution set that consist of connected sets of points
with neighborhoods biholomorphic to a neighborhood of a EucUdean space.
The solution set breaks up into a union of a finite number of irreducible
components, none of which is contained in the union of the remaining com-
ponents. In numerical algebraic geometry, we associate to each component
a witness set, which is the basic data structure used to numerically describe
and manipulate positive dimensional solution sets. Given a multiplicity one
irreducible component Z of a system of polynomials f{x) = 0, a witness
set consists of a triple (/, L, W), where L is a random linear space of di-
mension complementary to that of Z and where W is a set of points such
that W := Z n L. There is a slightly more involved definition for the case
of components of multiplicity greater than one as described in [24].
Many of the algorithms of numerical algebraic geometry make abun-
dant use of homotopy continuation. For example, the cascade algorithm
[20], one of the basic methods involved in computing the numerical irre-
ducible decomposition of an algebraic set, uses repeated applications of
homotopy continuation at difierent dimensions in order to produce points
on each component in each dimension. Monodromy and trace tests then
lead to the complete numerical irreducible decomposition of the solution
set of / . In the end, the user obtains from the methods of numerical alge-
braic geometry a wealth of information regarding the characteristics of the
solution set of a given polynomial system, some of which may be difficult to
procure by purely symbolic means. For good references on zero-dimensional
solving see [13, 16] and for numerical algebraic geometry see [24].
There are several software packages available to the pubUc which carry
out some of the operations of numerical algebraic geometry. However,
no one package contains all such capabilities. These packages include
HOM4PS [6], PHoM [9], POLSYS [27], PHCpack [28], and HomLab [30].
The most recently released package, Bertini [2], is under ongoing develop-
 SOFTWARE FOR NUMERICAL ALGEBRAIC GEOMETRY 3

ment by the authors. Although all software packages were developed at

different times for different reasons, they share the goal of solving polyno-
mial systems by numerical means.
The purpose of the present paper is two-fold. One purpose is to present
a paradigm for software in the field of numerical algebraic geometry. The
following section contains an elaboration on the aforementioned algorithms
and an enumeration of the various operations required for carrying out
those algorithms, broken into four levels. Implementation-specific details,
such as data structures, are omitted. The other purpose is to provide a brief
introduction to Bertini and indicate its partial fulfillment of the paradigm
of Section 2. That is the content of Section 3, which is also broken up
into four levels to mirror Section 2. The final section includes planned
extensions of the Bertini software package.
2. A paradigm for numerical algebraic geometry software. All
good software packages share several characteristics. In particular, good
software should be reliable (i.e., it provides correct output with clear signals
upon failure), as fast as possible with estimates of time remaining for large
jobs, modular for easy modification, and user-friendly. In addition, good
numerical software must be accurate and provide error estimates for all
solutions. ReHability and accuracy generally have an adverse impact on
speed, so while efficiency is important, it should not be emphasized over
finding correct answers.
Numerical accuracy may be approached in two ways. One approach is
to use a fixed level of precision and find as accurate a solution as possible,
possibly using higher levels of precision for subsequent runs to attain more
accuracy, if necessary. The other approach is to select an accuracy before
the run and adjust the precision during the run to attain that accuracy.
Either way, it has recently become generally accepted that it is important
to have available multiple levels of precision when implementing numerical
routines.
The purpose of this section is to provide an enumerated paradigm
for software specifically in the field of numerical algebraic geometry. This
detailed list is broken into four levels, beginning with very basic operations
not specific to polynomial system solving at level 0 in Section 2.1 and
moving through extensions of basic numerical algebraic geometry at level
3 in Section 2.4. The operations of each level build upon the capabilities
of the previous level. Each of the following four sections begins with a
discussion of the necessary operations of the given level and the resulting
capabilities of the software. Each section then concludes with a brief list of
the main operations to be implemented at that level. All operations should
be implemented for various levels of precision, ideally for arbitrarily high
precision.
2.1. Level 0: Basic operations. At the very core of a numerical
polynomial solver, one must of course have access to basic arithmetic both
4 DANIEL J. BATES ET AL.

of complex numbers and complex matrices. It is important to optimize the

efficiency of this arithmetic as much as possible, particularly in high preci-
sion, as most operations in numerical algebraic geometry rely heavily upon
arithmetic. In addition to basic matrix arithmetic, standard techniques
from numerical linear algebra are needed. Among the most important are
Gaussian elimination for linear solving, QR factorization for least squares
and orthogonal complements, and the SVD, or a related technique such
as the efficient method of [14], for finding numerical ranks. See [5, 26] for
general references on numerical linear algebra, while [14] provides a more
efficient method for determining the numerical rank of a matrix.
Random numbers play a key role in numerical algebraic geometry
as many statements hold generically, i.e., for almost all random choices,
thereby making the resulting algorithms hold with probability one. Any
standard random number generator will suffice, although it is best to have
the chosen random complex numbers of approximately unit modulus, for
stability. It is also important to have a consistent mechanism for extending
the precision of a randomly chosen complex number. In particular, upon
extending the precision of a random number a to make a, truncating back
to lower precision, and then again extending the precision, one should once
again obtain St.
It is of course necessary to somehow obtain the polynomials of interest
from the user, although the specific procedure for doing so is implementa-
tion-specific. To build a general solver, it is important to allow the functions
to be defined as expressions built from subexpressions. This is beneficial
not only for ease of use, but also for efficiency and numerical stability. It
is also necessary for generality to allow for homotopies that depend on
parameters, including analytic expressions as discussed in more detail in
the following section. If the user is specifying the entire homotopy, it is
also necessary to have a way for the user to provide solutions to the start
system g. Otherwise, the automatically generated start system should be
solved by the software.
Regardless of how the input data is provided, parsed, and stored, it is
at times necessary for the software to automatically homogenize the poly-
nomials provided by the user. Homogenization is simply the mechanism
for moving from a product of one or more complex spaces to a product
of complex projective spaces. This is a purely sjmnbolic operation which
should be implemented in such a way as to be easily reversed in case the
need arises. Suppose the homogenized system involves the cross product
of V projective spaces. Then v random nonhomogeneous linear equations,
one for each projective space in the product, should be appended to the
system in order to choose a patch on each complex projective space. These
V linear equations are known as the patch polynomials.
Basic path tracking, a level 1 operation, makes heavy use of both func-
tion evaluation and Jacobian evaluation. Function evaluation is straight-
forward, although it could be optimized via specialized techniques such as
 SOFTWARE FOR NUMERICAL ALGEBRAIC GEOMETRY 5

Horner's method. The Jacobian (i.e., the matrix of partial derivatives of

the polynomials) should be computed automatically by some form of auto-
matic differentiation so that the user does not need to provide it as input.
It is generally believed that the Jacobian should be computed explicitly
rather than approximated numerically, for stability.
Some of the operations at this level are common to numerical software
in general, regardless of the specific application. It should be noted that
existing libraries, such as LAPACK, provide robust implementations of
some of these operations, albeit in one level of precision.
Summary of level 0 operations:
• Complex scalar and matrix arithmetic
• Matrix operations (e.g., linear solving, numerical ranks, and or-
thogonal complements)
• Random number generation
• Parsing of input data
• Function homogenization
• Function evaluation
• Jacobian evaluation
• All numerical operations should be available in multiprecision.

2.2. Level 1: Basic tracking and zero-dimensional solving. It

is now possible to glue together the capabilities of level 0 to build algorithms
leading up to a full zero-dimensional solver, i.e., a method for computing
aU isolated solutions of / . The concept of "solving" a system could be
interpreted two ways. Given a desired accuracy e 6 R"^, let S := {s £
C ^ : | / ( s ) | < e}. This set may break up into k disconnected pieces, say
5 = 5i U . . . U 5fc. Then one interpretation of solving is to find a. z € Si
for each i. The second way to interpret "solving" / is to find, for each
s £ C^ such that / ( s ) = 0 and s has multiplicity m as a root of / , a set
{zi,...,Zm} C C ^ such that \zi — s\ < e for all i. The solutions in the
former sense of solving depend upon the scaling of the poljmomials while
those of the latter depend upon the scaling of the variable. The latter is
more widely accepted as correct and is thus the definition used throughout
this paper.
As discussed in Section 1, homotopy continuation casts the system /
in a family of polynomial systems, another of which is g. Such parame-
terized families sometimes arise naturally and are of great interest in some
applications. Other families are artificially constructed for the specific pur-
pose of solving / by continuation, in a so-called ab initio homotopy. Often
a naturally parameterized family can be used to define a homotopy that
has fewer paths to follow than an ab initio homotopy. As a result, software
in numerical algebraic geometry should address both ab initio and natu-
ral parameter homotopies. The natural parameter spaces may be complex
analytic and thus the need for evaluating complex analjrtic expressions of
parameters and complex numbers as mentioned in the previous section.
6 DANIEL J. BATES ET AL.

At its core, homotopy continuation is simply a sequence of steps ad-

vancing the path variable and updating the solution vector x to match.
Each step forward consists of a predictor step foUowed by a sequence of
corrector steps. The prediction is commonly carried out by Euler's method,
which steps along the tangent to the path. This is generally regarded as
an acceptable approach, although secant prediction can also be employed,
as can higher order methods such as Runge-Kutta. Once a prediction is
made to a new value of t, it is necessary to refine the point back towards
the solution curve. This may be accomplished by fixing t and running a
Newton-Hke corrector until the norm of the Newton residual has dropped
below a prespecified tolerance.
Naturally, there are times when steps will fail, where failure is declared
when the corrector does not converge sufficiently fast. Such step failures
need not trigger path failure. Rather, adaptive steplength should be uti-
lized. Upon step failure when using adaptive steplength, the steplength is
decreased by a prespecified factor in the hope that convergence will occur
for a smaller step. Only if progress along the path becomes excessively slow
is the path declared a failure. Conversely, if the steps are progressing well,
it is worthwhile to try to increase the steplength. More details regarding
prediction, correction, and adaptive steplength may be found in [1, 16, 24].
Path failure may occur for a number of reasons, but the presence
of a singularity, particularly at t = 0, is a common cause. There are
several sophisticated algorithms known as endgames that help to speed
convergence at t = 0 for both nonsingular and singular endpoints. These
endgames are typically employed for every path, so it is important to have
at least one implemented in any software package for numerical algebraic
geometry. Details regarding endgames may be found in [17, 18, 19].
For zero-dimensional solving, polynomial systems given by the user
could be nonsquare with n > N. Fortunately, Bertini's theorem [24] guar-
antees that a new system consisting of N generic linear combinations of the
original n poljoiomials will have among its solutions all the isolated solu-
tions of the original system, though possibly with increased multiphcity. It
may also have nonsingular isolated extraneous solutions. The extraneous
solutions are easily detected as they do not satisfy the original system.
Unless the user chooses to specify a parameter homotopy, the software
must be able to automatically produce an appropriate start system g, solve
it, and attach it to the homogeneous, square system / in order to create
the homotopy H. There are several methods for producing start systems,
although the general rule is that the computational cost increases in order
to produce start systems with fewer paths to track. Among the common
choices, total degree start systems, consisting of polynomials of the form
x^' — 1, where di is the degree of the i*'' polynomial in / , are trivial to
build and solve but have the largest number of paths to be tracked (that
being the product of the degrees of the functions). At the other end of the
spectrum, the construction and solution of sophisticated polyhedral homo-
 SOFTWARE FOR NUMERICAL ALGEBRAIC GEOMETRY 7

topies involve far more computation time but may result in far fewer paths
(the number of which is the mixed volume). It is not clear a priori which
type of start system is best-suited for an arbitrary polynomial system, so
it is important to have multiple types of start systems available.
Once all (or most) of the aforementioned operations have been imple-
mented, it is possible to compute the zero-dimensional solutions of a given
polynomial system. Two other useful tools belong at this level. First, defla-
tion [11] is a means of constructing a new polynomial system / from / such
that / has a nonsingular solution in place of a particular singular solution
of / . This makes it possible to compute singular solutions more accurately
without relying on higher precision. The major drawback of implementing
deflation is that decisions must be made about the rank of the Jacobian
matrix at the solution point before the solution point is known accurately.
The use of endgames can improve the accuracy of the solution estimate
before deflation, helping to ensure that the correct deflation sequence is
performed but adding the cost of endgame computations. Exploration of
the numerical stabihty and efiiciency of deflation and endgames is a topic
of ongoing research. The big advantage of deflation comes when dealing
with positive dimensional components of multiplicity greater than one.
The other useful tool at this level is a post-processor to manipulate
and display the solutions computed by the solver as well as any statistics
gathered during tracking. As the functionality of such a tool is application-
specific, no more details will be discussed here.
Summary of level 1 operations:
• Differential equation solving, e.g., Euler's method
• Newton's method
• Basic path tracking with adaptive steplength control
• Adaptive precision path tracking
• Squaring of systems
• Start system and homotopy generation
• Start system solving
• Pull zero-dimensional solving
• Endgames
• Deflation
• Post-processing of zero-dimensional data.

2.3. Level 2: Positive-dimensional solving. The solution set Z

of a polynomial system / may have several components and these may
not all have the same dimension. Letting D := dixnZ, the irreducible
decomposition may be written as Z = U ^ Q Z J = U ^ Q Ujeii ^i,jj where
each Zij is an irreducible component of dimension i, and accordingly each
Zi is the pure i-dimensional component of Z. (Symbol Ij in the above is
just an index set for the components of dimension.)
One of the key objectives in numerical algebraic geometry is to find
a numerical irreducible decomposition of Z, which consists of sets Wij —
8 DANIEL J. BATES ET AL.

Zij n L^-i, where Ljv-j is a generic linear subspace of dimension N — i.

Wij, together with L^-i, is known as a witness set for Zij, and by abuse of
notation, Wi = Uj£l^WiJ is called a witness set for Zj. We briefly describe
the algorithms for computing the irreducible decomposition below. Full
details may be found in the references cited below or in [24].
The main steps in computing a numerical irreducible decomposi-
tion are:
• find witness supersets Wi D Wi for each dimension i,
• prune out "junk points" from Wi to extract the witness sets Wi,
and
• break Wi into distinct sets Wij, the witness sets for the irreducible
components Zij.
The witness supersets Wi are generated by the application of zero-dimen-
sional solving to find the isolated points in the slice Z n Lj^-i. All of
the Wi, for 0 < J < D, can be obtained using the cascade algorithm
[20], starting at i = D and cascading sequentially down to i = 0. The
junk points in Wi must lie on some Zj with j > i. Thus, the junk may
be removed by testing each point p € Wi for membership in a higher-
dimensional component. This can be done using continuation on slices to
see if any of the witness sets Wj, j > i connect to p as the slicing linear
space LN-I is moved continuously until it contains p. The final break up
of Wi into irreducibles is accomplished by first using monodromy, which
comes down to discovering connections between witness points as the linear
slicing space is moved around a closed loop in the associated Grassmannian
[21]. This is followed by checking if the connected groups so discovered are
complete, by means of the trace test [22]. The trace method may also be
used to complete a partial decomposition. Both monodromy and the trace
method involve specialized continuation of slices and careful bookkeeping.
Squaring is a concern for positive-dimensional solving just as it is for
zero-dimensional solving. Although much carries over, one difference is that
the size to which the system should be squared depends on the dimension
of the component, e.g., for components of dimension k, the defining system
should be randomized to a system oi N — k equations.
Given witness data for an algebraic set Z, there are three operations
of particular interest for users. First, a user might want to find many
points on a specific component. This is known as sampling and is very
closely related to monodromy, as both use the continuation of sHces to
move witness points around on the component. Second, a user might want
to know if a given point lies on an irreducible component of Z. This is
the same component membership test used in the junk removal stage of
computing an irreducible decomposition. Finally, a user might want the
accuracy of some endpoint sharpened. This is just a matter of running
Newton's method appropriately, perhaps after deflating / . Of course, as in
 SOFTWARE FOR NUMERICAL ALGEBRAIC GEOMETRY 9

the previous section, a post-processor would be appropriate, although the

exact functionality is again application-specific.
Deflation is particularly valuable as a method for multiple components,
e.g., to track intersections of a multiplicity greater than one component
with a one-parameter family of linear spaces of complementary dimension,
as is done to sample a multiple component. Roughly speaking, if Z is
a fc-dimensional component of the solution set of a system f — 0, then
the computation of the deflation oi Z f) L ioi f restricted to a generic
fe-codimensional Unear space gives rise to a "deflation" of the whole com-
ponent. At the expense of increasing the number of variables, this allows
us to numerically treat all components as multiplicity one components.
S u m m a r y of level 2 operations:
• Continuation of slices
• Monodromy
• Traces
• Squaring of systems for positive-dimensional solving
• Cascade algorithm
• Pull numerical irreducible decomposition
• Sampling
• Component membership
• Endpoint sharpening
• Post-processing of witness data
• Deflation for components.

2.4. Level 3: Extensions and applications of the basics. This

highest level consists of operations that make use of the basic numerical
algebraic geometry maneuvers described in the previous three levels. For
example, for two algebraic sets X and Y that are the solution sets of
polynomial systems / and g, respectively, suppose one has witness sets Wx
and WY for X and Y but would Uke a witness set W for X n Y. There is a
now a method [23] for computing the numerical irreducible decomposition
of such an intersection, the inclusion of which, while not essential for basic
software in numerical algebraic geometry, will be important as the field
continues to develop.
There are several other advanced algorithms that should be included
in a complete state-of-the-art implementation. Another such technique is
that of [25], which provides a way of finding exceptional sets, i.e., the sets of
points in the parameter space above which the fiber has dimension higher
than the generic fiber dimension. Fiber products play a key role in this
algorithm and therefore need to be available in the software before the
method for finding exceptional sets may be implemented. Also, in real-
world applications, real solutions are often of more interest than complex
solutions, so the extraction of real solutions from the numerical irreducible
decomposition, for example, by an extension of the method of [15], would
be very useful.
10 DANIEL J. BATES ET AL.

This is not intended to be a complete list, and it is anticipated that

many more operations could be added in the near future. One capabil-
ity, though, that is important now and will only become more essential
over time, is parallelization. Although not every algorithm of numerical
algebraic geometry is fully parallelizable, basic path tracking is easily par-
allelized so great savings can be made throughout levels 1, 2, and 3 by
doing so [10, 12, 27, 29].
Summary of level 3 operations:
• Intersection of components
• Fiber products
• Finding exceptional sets
• Extracting real solution sets from complex components
• Parallelization.
3. Bertini. Bertini [2] is a new software package under ongoing devel-
opment by the authors for computation in the field of numerical algebraic
geometry. Bertini itself has evolved from a program called Polysolve cre-
ated by Bates, C. Monico (Texas Tech University), Sommese and Wampler,
although nothing substantial remains in Bertini from Polysolve.
Bertini is written in the C programming language and makes use of
several specialized Mbraries, particularly lex and yacc for parsing and GMP
and MPFR for multiple precision support. The beta version of Bertini was
released in October 2006 to coincide with the Software for Algebraic Ge-
ometry workshop of the Institute for Mathematics and its Applications. It
is currently anticipated that Bertini 1.0 will be made available to the pub-
He sometime in 2007. Bertini is currently only available as an executable
file for 32- or 64-bit Linux and for Windows, via Cygwin. Specific instruc-
tions for using Bertini are included with the distribution and on the Bertini
website.
Among other things, Bertini is capable of producing all complex iso-
lated solutions of a given polynomial and witness sets for each positive-
dimensional irreducible component. The goal of the Bertini development
team is to eventually include in Bertini all operations described above
in Sections 2.1 through 2.4. The pmrpose of this section is to indicate
briefly which of those operations are already available in the beta version
of Bertini. The specific algorithms implemented are also described, when
appropriate. Details regarding the development plans for Bertini 1.0 and
beyond may be found in Sections 4.1 and 4.2, respectively.

3.1. Level 0. By default, Bertini uses IEEE double precision, al-

though it also allows for any fixed level of precision available in MPFR.
In particular, precision is available starting from 64 bits, increasing in 32
bit increments. Furthermore, the beta version of Bertini allows the user
to select adaptive precision for zero-dimensional solving. In adaptive pre-
cision mode, Bertini begins in double precision and increases precision as
necessary, determined by the algorithm described in [3].
 SOFTWARE FOR NUMERICAL ALGEBRAIC GEOMETRY 11

All matrix operations described in Section 2.1 have already been im-
plemented in Bertini with the exception of the QR algorithm (i.e., the QR
decomposition of a matrix), which will be implemented by the next release.
Gaussian elimination with scaled partial pivoting is used for linear solving.
The SVD algorithm is based on the description provided in [26].
Bertini uses the basic random number generator found in the C stan-
dard library. Each random number is scaled to be of the appropriate mod-
ulus and stored as an exact rational number with the denominator a power
of ten. Thus, to increase precision, all new digits are set to zero, although
this is not ideal. It would be better to increase precision using randomly
chosen digits such that subsequent truncations and extensions result in the
same additional digits. This will be changed in a future version of Bertini.
The user provides poljTiomials (either the entire homotopy or just the
target system) to Bertini in a file which is parsed using lex and yacc. The
polynomials are stored in straight-line format for easy automatic differenti-
ation and homogenization as well as efficient evaluation. Any optimization,
such as using Horner's method, is the responsibility of the user. Homoge-
nization is carried out as the polynomials are being parsed, and differenti-
ation is carried out after the entire system has been parsed, currently via
forward automatic differentiation as described in [8].
Bertini allows the user to include constants and subfunctions when
defining the target system and also the path variable and parameters when
defining a homotopy. There are various restrictions placed on the homoge-
neous structures allowed depending upon the scenario, but multiple vari-
able groups are generally supported. All coefficients, constants, and other
numbers are stored as exact rational numbers, as described for random
numbers above, and are assumed to be exact. As a result, function and Ja-
cobian evaluation are available at any level of precision. Analytic functions
of constants and parameters are not yet supported.

3.2. Level 1. All level 1 operations of Section 2.2 have been imple-
mented in Bertini with the exception of deflation. Euler's method is used as
the predictor, and Newton's method as the corrector. The fractional power
series endgame [19] is available in basic precision, fixed higher precision,
and adaptive precision.
All polynomial systems are automatically squared to the appropriate
number of polynomials and variables, although the details of the squaring
mechanism are excluded from the present paper for brevity. If the user
specifies only the target system and variable groupings, Bertini multiho-
mogenizes the system according to the groupings and attaches a compatible
m-homogeneous start system via a linear homotopy. Each polynomial in
the start system consists of a product of random linear functions matching
the homogeneous structure of the corresponding target polynomial. If the
user specifies only one variable group, the result is a total degree start sys-
tem. Bertini also has a post-processor which collects the endpoint data for
12 DANIEL J. BATES ET AL.

the run and creates a number of files, some of which are human-readable
and some of which are better suited for machine reading. The files provide
lists of singular endpoints, nonsingular endpoints, real endpoints, and so
on. Important path characteristics, such as an estimate of the error in the
endpoint, are also supplied in these files. Bertini also indicates whether any
paths failed (and why they did) and whether any path crossing occurred
prior to the endgame.
3.3. Level 2. All of the operations of level 2 described in Section 2.3
have been implemented in double precision in Bertini, with the exceptions
of endpoint sharpening and deflation. Users have the option of computing
the numerical irreducible decomposition of the polynomial system input
file, sampling a component of a system for which witness data has pre-
viously been computed, or performing component membership on such a
system. The algorithms are then carried out as described in Section 2.3
and [24].
In the case of computing the numerical irreducible decomposition,
Bertini stores all data in memory until the end of the run. At the end
of the run, a postprocessor sorts the data and creates a number of files, as
in the case of a zero-dimensional run. The structure of the algebraic set is
also provided in a table on the screen, namely the number of irreducible
components of each degree in each dimension. Additional details may be
found in the Bertini documentation available in [2].
3.4. Level 3. None of the level 3 operations described in Section 2.4
are available in the beta version of Bertini. Much of the further develop-
ment of Bertini will involve level 3 operations.
4. T h e future for B e r t i n i . It is of course impossible to predict what
will happen with either Bertini or numerical algebraic geometry in the
distant future. However, there is a short-term development plan for Bertini.
The following sections describe the development plan for Bertini 1.0 and
further anticipated developments beyond Bertini 1.0.
4 . 1 . B e r t i n i 1.0. There are several key developments currently in the
process of being implemented in order to move Bertini from the beta version
to version 1.0. The four main improvements are the inclusion of the cascade
algorithm in fixed multiple precision as well as adaptive precision, deflation
in both the zero- and positive-dimensional cases, endpoint sharpening (as
described in Section 2.2), and the QR decomposition. Other improvements
are minor, such as several small changes to the output files and an improved
post-processor. The goal is to have most of the operations of levels 0 , 1 , and
2 available in Bertini 1.0 in fixed user-specified multiprecision and adaptive
precision.
4.2. Beyond B e r t i n i 1.0. While there will certainly be a number of
changes to the core functions of Bertini after version 1.0 is released (such as
improving the way that random numbers are handled), many of the major
 SOFTWARE F O R NUMERICAL ALGEBRAIC GEOMETRY 13

changes will involve the implementation of the operations at level 3 and

new algorithms currently under development.
Other planned extensions include allowing the user to specify analytic
functions of parameters and constants, the inclusion of polyhedral meth-
ods, e.g., using MixedVol [7], and various forms of parallelization. Bertini
currently operates by setting up files with input data, calling Bertini, and
reading the results from files. It is anticipated that Bertini will eventu-
ally include a scripting language or make use of software that provides
an interactive environment. Bertini will also undergo a move from the C
programming language to C4-I- sometime after the release of version 1.0,
both for increased modularity and for easier extension. Finally, Bertini will
eventually make use of more efficient multiprecision numerical libraries as
they become available and will include interfaces to other software packages
commonly used in the mathematics community.

REFERENCES

[1] E.L. ALLGOWER AND K . GEORG, Introduction to numerical continuation methods,

Classics in Applied Mathematics, 4 5 , SIAM, Philadelphia, 2003. Reprint of
the 1990 edition (Springer-Verlag, Berlin).
[2] D . J . BATES, J . D . HAUENSTEIN, A . J . SOMMESE, AND C . W . WAMPLER, Bertini:
Software for Numerical Algebraic Geometry, available at www.nd.edu/
~sommese/bertini.
[3] , Adaptive multiprecision path tracking, to appear SIAM J. Num. Anal.
[4] D.F. DAVIDENKO, On a new method of numerical solution of systems of nonlinear
equations, Doklady Acad. Nauk SSSR (N.S.), 88:601-602, 1953.
[5] J . W . DEMMEL, Applied Numerical Linear Algebra, SIAM, Philadelphia, 1997.
[6] T . GAG AND T . Y . L I , HOM4PS, available at www.csulb.edu/~tgao.
[7] T . G A O , T.Y. L I , AND M . W U , Algorithm 846: MixedVol: A software package
for mixed-volume computation, ACM Trans. Math Software, 31(4):555-560,
2005. Software available at www.csulb.edu/~tgao.
[8] A. GRIEWANK, Evaluating derivatives. Principles and techniques of algorithmic
differentiation, FVontiers in Applied Mathematics, 19, SIAM, Philadelphia,
2000.
[9] T . GuNJi, S. K I M , M . K O J I M A , A. TAKEDA, K . FUJISAWA, AND T . MIZUTANI,
PHoM—o polyhedral homotopy continuation method for polynomial systems.
Computing, 73(l):55-77, 2004. Software available at www.is.titech.ac.jp/
~kojima.
[10] T . G U N J I , S . K I M , K . FUJISAWA, AND M . KOJIMA, PHoMpara—parallel imple-
menation of the Polyhedral Homotopy continuation Method for polynomial
systems, Computing, 77{4):387-411, 2006.
[11] A. LEYKIN, J. VERSCHELDE, AND A. ZHAO, Evaluation of Jacobian matrices for
Newton's method with deflation for isolated singularities of polynomial sys-
tems, in proceedings of SNC 2005 (International Workshop on Symbolic-
Numeric Computation, Xi'an, China, July 19-21, 2005), edited by Dongming
Wang and Lihong Zhi, pp. 19-28, 2005.
[12] A. LEYKIN, J. VERSCHELDE, AND Y . ZHUANG, Parallel Homotopy Algorithms to
Solve Polynomial Systems, in proceedings of ICMS 2006 (Second International
Congress on Mathematical Software, Castro Urdiales, Spain, September 1-3,
2006), Lecture Notes in Computer Science, 4151:225-234, 2006.
[13] T.Y. L I , Numerical solution of multivariate polynomial systems by homotopy con-
tinuation methods. Acta Numer., 6:399-436, 1997.
14 DANIEL J. BATES E T AL.

[14] T.Y. Ll AND Z. ZENG, A rank-revealing method with updating, downdating, and
applications, SIAM J. Matrix Anal. AppL, 26(4):918-946, 2005.
[15] Y. Lu, D . J . BATES, A.J. SOMMESE, AND C . W . WAMPLER, Finding all real points
of a complex curve, to appear Contemporary Mathematics.
[16] A.P. MORGAN, Solving polynomial systems using continuation for engineering and
scientific problems, Prentice-HaJl, Englewood Cliffs, N.J., 1987.
[17] A . P . MORGAN, A.J. SOMMESE, AND C.W. WAMPLER, Computing singular solutions
to nonlinear analytic systems, Numer. Math., 58(7):669-684, 1991.
[18] , Computing singular solutions to polynomial systems, Adv. Appl. Math.,
13(3):305-327, 1992.
[19] , A power series method for computing singular solutions to nonlinear ana-
lytic systems, Numer. Math., 63(3):391-409, 1992.
[20] A.J. SOMMESE AND J. VERSCHELDE, Nun,erical homotopies to compute generic
points on positive dimensional algebraic sets, J. Complexity, 16(3):572-602,
2000.
[21] A.J. SOMMESE, J. VERSCHELDE, AND C.W. WAMPLER, Using monodromy to de-
compose solution seta of polynomial systems into irreducible components, in
proceedings of Applications of algebraic geometry to coding theory, physics
and computation (Eilat, 2001), NATO Sci. Ser. II Math. Phys. Chem., 3 6 ,
2001.
[22] , Symmetric functions applied to decomposing solution sets of polynomial
systems, SIAM J. Num. Anal., 40{6):2026-2046, 2002.
[23] , Homotopies for intersecting solution components of polynomial systems,
SIAM J. Num. AnaJ., 42(4):1552-1571, 2004.
[24] A.J. SOMMESE AND C.W. WAMPLER, Numerical solution of systems of polynomials
arising in engineering and science, World Scientific, Singapore, 2005.
[25] , Exceptional sets and fiber products, to appear Foundations of Computa-
tional Mathematics.
[26] G.W. STEWART, Matrix algorithms. Vol. I. Basic decompositions, SIAM, Philadel-
phia, 1998.
[27] H.-J. Su, J. M C C A R T H Y , M . SOSONKINA, AND L . T . WATSON, Algorithm 857: POL-
SYS-GLP - A parallel general linear product homotopy code for solving poly-
nomial systems of equations, ACM Trans. Math. Software, 32(4):561-579,
2006. Software available at www.vrac.iastate.edu/~haijunsu.
[28] J. VERSCHELDE, Algorithm 795: PHCpack: a general-purpose solver for poly-
nomial systems by homotopy continuation, ACM Trans. Math. Software,
25(2):251-276, 1999. Software available at www.math.uic.edu/~jan.
[29] J. VERSCHELDE AND Y . ZHUANG, Parallel implementation of the polyhedral homo-
topy method, in proceedings of I C P P W '06 (International Conference Work-
shojjs on Parallel Processing), IEEE Computer Society, Washington, DC, 4 8 1 -
488, 2006.
[30] C.W. WAMPLER, HomLab: Homotopy Continuation Lab, available at www.nd.edu/
'^^cwamplel.
P H C L A B : A M A T L A B / O C T A V E INTERFACE TO PHCPACK*
YUN GUANt AND JAN VERSCHELDE*

A b s t r a c t . PHCpack is a software package for Polynomial Homotopy Continuation,

to numerically solve systems of polynomial equations. The executable program "phc"
produced by PHCpack has several options (the most popular one "-b" offers a blackbox
solver) and is menu driven. PHClab is a collection of scripts which call phc from within
a MATLAB or Octave session. It provides an interface to the blackbox solver for finding
isolated solutions. We executed the PHClab functions on our cluster computer using
the MPI ToolBox (MPITB) for Octave to solve a list of polynomial systems. PHClab
also interfaces to the numerical irreducible decomposition, giving access to the tools to
represent, factor, and intersect positive dimensional solution sets.

1. Introduction. Polynomial systems arise in various fields of science

and engineering, e.g.: the design of a robot arm [13] so its hands passes
through a prescribed sequence of points in space requires the solution of
a polynomial system. Homotopy continuation methods are efficient nu-
merical algorithms to approximate all isolated solutions of a polynomial
system [10]. Recently homotopies have been developed to describe positive
dimensional solution sets [20].
This paper documents an interface PHClab to use the functionality
provided by PHCpack [21] from within a MATLAB or Octave session.
The main executable program provided by PHCpack is phc, available for
downloading on a wide variety of computers and operating systems. The
program phc requires no compilation. Its most popular mode of operation
is via the blackbox option, i.e.: as phc -b input output. Recently the
program has been updated with tools to compute a numerical irreducible
decomposion [18].
The main motivation for PHClab is to make it easier to use phc by
automatic conversions of the formats for poljaiomial systems (on input)
and solutions (on output). Manual or adhoc conversions can be tedious
and lead to errors. As PHCpack has no scripting language on its own, the
second advantage of PHClab is help the user to systematically use the full
capabilities of PHCpack. As MATLAB (and its freely available counterpart
Octave) is a very popular scientific software system, PHClab will be a useful
addition to PHCpack.

*This material is based upon work supported by the National Science Foundation
under Grant No. 0105739 and Grant No. 0134611.
t Department of Mathematics, Statistics, and Computer Science, University of
Illinois at Chicago, 851 South Morgan ( M / C 249), Chicago, IL 60607-7045, USA
(guanamath.ui c.edu).
* Department of Mathematics, Statistics, and Computer Science, Univer-
sity of Illinois at Chicago, 851 South Morgan (M/C 249), Chicago, IL
60607-7045, USA (jan9math.uic.edu or j a n . v e r s c h e l d e a n a - n e t . o m l . g o v ; h t t p : / /
www .math. u i c . e d u / ~ j an).
15
16 YUN GUAN AND JAN VERSCHELDE

Another feature of PHClab is the possibiUty of developing parallel code

at a high level, when using the MPI ToolBox (MPITB [2]) for Octave.
2. The design of PHClab. The first interface from a C program to
phc was written by Nobuki Takayama and is still available via the PHCpack
download web site. Via this interface, phc became part of OpenXM [11]
(see also [12]). We used the same idea to build PHCmaple [7, 8], defining a
Maple interface to PHCpack.
All that is needed to make the interface work is the executable form of
phc. PHClab is a collection of scripts written in the language of MATLAB
and Octave. These scripts call phc with the appropriate options and menu
choices.
3. Downloading and installing. PHClab was tested on Matlab 6.5
and Octave 2.1.64 on computers running Windows and Linux. On an Apple
laptop running Mac OS X version 10.3.7, we executed PHClab in Octave
2.1.57.
The most recent version of PHCpack and PHClab can be retrieved
from
h t t p : //www. math. u i c . e d u / ^ j an/download. html
which we fi-om now on call the download web site. To install and use
PHClab, execute the following steps:
1. Prom the download web site, either download the source code for
phc (a makefile is provided with the code), or select an executable
version of phc. Currently, phc is available in executable form on
Windows, workstations from IBM (running AIX 5.3) and SUN
(running SunOS 5.8), and PCs running Linux and Mac OS X 10.3.
Except for Windows (which comes just as a plain phc.exe), one
has to run gimzip followed by t a r xpf on the downloaded file.
2. The PHClab distribution is available as PHClab. t a r . gz firom the
download web site. To install PHClab in the directory /tmp, save
PHClab. t a r . gz first in /tmp, and then execute the following se-
quence of commands:
cd /tmp; mkdir PHClab; mv /tmp/PHClab. t a r . gz PHClab;
cd /tmp/PHClab; gunzip PHClab.tar.gz; t a r xpf PHClab.tar.
3. Either launch MATLAB or Octave in the directory PHClab, or
add the name of the directory which contains PHClab to the path
of MATLAB or Octave.
The first command of PHClab to be executed is set_phcpath. This
command takes one argument: the full path name of the file name which
contains the executable program phc. For example, if phc was saved in
/tmp, then a session with PHClab starts with s e t _ p h c p a t h ( ' / t m p / p h c ' ) .
4. Solving polynomial s y s t e m s . In this section we define the basic
commands to solve polynomial systems using PHClab. We first define the
input/output formats, introducing the function make_system to convert a
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 17

matrix format for a polynomial system into a sjmibolic input format to phc.
The blackbox solver of PHCpack is called by the command solve_systein.
Besides the solution vectors, the solver returns extra diagnostical informa-
tion about the quahty of each solution.
Path tracking typically starts from a generic system (without any sin-
gular solutions) to a more specific system. We use the system we first
solved by the blackbox solver as start system to solve a system with spe-
cific coefficients, using the function track. Because of our specific choice
of the coefficients, we generated a polynomial system with a double solu-
tion, i.e.: of multiphcity two. Via the function refine_sols and deflation,
we respectively refine a solution and deflate its multiplicity into a regular
problem.
The last function we introduce in this section is mixed_volume, to
compute the mixed volume for a polynomial system and (optionally) create
and solve a random coefficient start system. The mixed volume equals the
number of roots without zero components of a polynomial system with suf-
ficiently generic coefficients. The function mixed.volume calls the trans-
lated form of the code Mixed Vol [3].
4.1. I / O formats and the blackbox solver. The input to the
solver is a system of multivariate equations with complex floating-point
coefficients. For example, consider the system g{x) = 0:

9ixi,x,) = ^ 2.1x1-1.9x1=0' ^itl^* = v ^ - (4-1)

This system is encoded as a matrix, with in its rows the terms of each
polynomial. A zero row in the matrix marks the end of a polynomial
in the system. A nonzero row in the matrix represents a term as the
coefficient followed by the exponents for each variable. For example 4.7x2
is represented by the row 4 . 7 0 2. If n is the number of variables and m
the total number of terms, then the matrix encoding the system has m + n
rows and n + 1 columns.
To solve the system g{x.) = 0 using PHClab, we may execute the
following sequence of instructions:
7, t a b l e a u input for a system :
t = [1.3 2 0; 4 . 7 0 2; - 3 . 1 + 2.3*i 0 0; 0 0 0;
2.1 0 2; -1.9 1 0 ; 0 0 0];
make_system(t) % shows symbolic format of the system
s = solve_system(t); % call the blackbox solver
ns = size(s,2) % check the number of solutions
s3 = s(3) % look at the 3rd solution
Then we see the following output on screen:
ans =
' + 1.3*xl**2 + 4.7*x2**2 + (-3.1+2.3*1)'
18 YUN GUAN AND JAN VERSCHELDE

' + 2.1*x2**2 - 1 . 9 * x l '

ns =
4
s3 =

time: 1
multiplicity: 1
err: 4.0340e-16
rco: 0.1243
res: 2.7760e-16
xl: -3.9180 + 0.38761
x2: 0.0930 + 1.88511

We see the coordinates of the solution are in the last fields (displayed
by default in short format, we may see more in format long) and extra
diagnostics in the first five fields, briefly explained below.
t i m e is the end value of the continuation parameter. If this value is not
equal to one, then it means that the path tracker did not manage to
reach the end of the path. This may happens with paths diverging
to infinity or with highly singular solutions.
multiplicity is the multiplicity of the solution. A solution is regular when
the multiplicity is one. When the approximation for a solution
is not yet accurate enough, then the multiplicity might still be
reported as one, although the value for rco might be close to the
threshold.
e r r is the magnitude of the last update Newton's method made to the
solution. At singular solutions, the polynomial functions exhibit a
typical "flat" behavior. Although the residual may then be already
very small, the value for this e r r can be still large.
rco is an estimate for the inverse of a condition number of the Jacobian ma-
trix evaluated at the approximate solution. A solution is deemed
singular when this number drops below the threshold value of 10~^.
Multiple solutions are singular. The condition number C of the Ja-
cobian matrix measures the forward error, i.e.: if the coefficients
are given with D digits precision, then the error on the approxi-
mate solution can be as large as C x 1 0 " ^ .
res is the residual, or the magnitude of the polynomial system evaluated
at the approximate solution. This residual measures the backward
error: how much one should change the coefiicients of the given
system to have the computed approximation as the exact solution.
The values of the coordinates of the solutions are by default displayed in
MATLAB's (or Octave's) format short. By format long e we can see
the full length in scientific format. For the solution above, the values of
e r r , rco, and r e s indicate an excellent quality of the computed solution.
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 19

4.2. P a t h tracking from a generic t o a specific s y s t e m . The

four solutions of the system we solved are all very well conditioned, so we
may use them as start solutions to solve a system with the same coefficient
structure, but with more specific coefficients:

n-^'-^) = {4U+T-tzl withz = v^. (4.2)

Geometrically, the polynomials in the system / ( x ) = 0 respectively de-

fine an ellipse and a parabola, positioned in such a way that their real
intersection point is a double solution.
In the sequence of instructions below we use the function track, using
the new system double (the system / ( x ) = 0) as target system and the
system t we solved as start system (we called it g(x) = 0). Note that
before calling track, we must set the value of time in every solution to
zero, so s contains proper start solutions.
double = [1.0 2 0; 4.0 0 2; - 4 . 0 0 0; 0 0 0;
- 2 . 0 0 2; +1.0 1 0; - 2 . 0 0 0; 0 0 0 ] ;
make_system(double) % shows system
s ( l ) .time = 0; s ( 2 ) . t i m e = 0; '/. i n i t i a l i z e time for every
s ( 3 ) . t i m e = 0; s ( 4 ) . t i m e = 0 ; % s t a r t s o l u t i o n t o zero
sols = track(double,t,s); % c a l l t h e path t r a c k e r s
ns = s i z e ( s o l s , 2 ) % check number of s o l u t i o n s
s2 = s o l s ( 2 ) % look a t t h e 2nd s o l u t i o n
The choice of the second solution was done on purpose because this solution
needs extra processing. In general however, we have no control over the
order in which the solutions are computed, i.e.: while every run should give
the same four solutions back, the order of solutions could be permuted.
The output we see on screen of the sequence above is
ans =
'xl**2 + 4*x2**2 - 4 '
' -2*x2**2 + x l - 2 '
ns =
4
s2 = '
time: 1
multiplicity: 1
err: 4.373000000000000e-07
rco: 3.i47000000000000e-07
res: 7.235000000000000e-13
xl: 2.0000000000000006+00 - 5.048709793414480e-29i
x2: -2.493339146012010e-07 - 1.879166705634450e-07i
Recall that we constructed the equations in our second system / ( x ) = 0 so
that there is a double solution at (2,0). However, since we are not yet close
enough to the actual double solution (2,0), the magnitude of the condition
20 YUN GUAN AND JAN VERSCHELDE

number is about 10^, below the threshold of 10^, so phc does not recognize
the solution as a double root. We will next see how to get closer to the
actual solution.
4.3. Refining and reconditioning singular solutions. To refine
the solution we save in s2, we execute 10 addition Newton steps, applying
refinejsols to the second solution s2:
r2 = r e f i n e _ s o l s ( d o u b l e , s 2 , 1 . 0 e - 1 6 , 1 . 0 e - 0 8 , 1 . 0 e - 1 6 , 1 0 )
We allow 10 iterations (last parameter of refine_sols) of Newton's method,
requiring that either the magnitude of the correction vector (err) or the
residual (res) is less or equal than 10~^®, as specified respectively by the
third and fifth parameter of reflne_sols.
Below, on the output we see the estimate for the inverse condition
number has decreased, along with the value for x2:
r2 =
time: 1
multiplicity: 1
err: 3.3000000000000006-09
rco: 3.8850000000000006-09
res: 6.4279999999999996-17
xl: 2.OOOOOOOOOOOOOOOe+00 4.309100000000000e-41i
x2: -2.9990621833465416-0! - 3.017695139191104e-10i
Now that the estimate for the inverse condition number has dropped
from 10~^ to 10"^, below the threshold of 10~^, we expect this solution to
be singular. To deflate the multiplicity [9] and recondition the solution, we
execute
def.sols = deflation(double,sols);
def_sols{4,l}
and then we see on screen
ans =
time: 1
multiplicity: 2
e r r : 2.1860000000000006-07
r c o : 1.2420000000000006-01
r e s : 1.0030000000000006-13
x l : 2.0000000000000106+00 + 1.929286255918420e-14i
x2: -1.7426214785217806-14 + 8.266179457715231e-15i
lm_l_l: 3.0779398018996406-01 + 6.678691166401400e-01i
lm_l_2: -6.7375245460803006-01 - 2.946929268111410e-01i
Notice the value rco which has increased dramatically firom 3.885e-09 to
1.242e-01, as a clear indication that the solution returned by deflation is
well conditioned. Yet the multiplicity is two as a solution of the original
system. The deflation procedure has constructed an augmented system for
which the double solution of the original system is a regulax root. The
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 21

values for lm_l_l and lm_l_2 are the values of the multipliers Ai,i and Ai^2
used in the first deflation of the system. The number of multipliers used
equals the one plus the numerical rank of the given approximate solution
evaluated at the Jacobian matrix of the original system. The augmented
system is returned in def _sols{3,1}.
4.4. Mixed volumes a n d r a n d o m coefficient s y s t e m s . In order
to solve the system (4.2) we used the output of the blackbox solver on a
more general system. The blackbox solver uses polyhedral homotopies [5] to
solve a system with the same sparse structure but with random coefficients.
Such random coefficient system has exactly as many isolated solutions as its
mixed volume [10]. The function mixed_volume in PHClab gives access
to the code MixedVol [3] as it is available as translated form in PHCpack.
If we continue our session with in double the tableau input for the
system (4.2), then we can compute its mixed volume and solve a random
coefficients start system via the following sequence of commands:
[ v , g , s ] = mixed_volume(double,l); % compute mixed volume
v 7. check t h e mixed volume
US = s i z e ( s , 2 ) 7. check number of s o l u t i o n s
g '/. ramdom c o e f f i c i e n t system
The output to these command is
V = 4
ns = 4

[1,1] =
+( 9.51900029533701E-01 + 3.06408769087537E-01*i)*xl~2
+( 9.94012861166580E-01 + 1.09263131180786E-01*i)
[2.1] =
+( 6.10442645118414E-01 - 7.92060462982993E-01*i)*x2-2
+(-5.76175858933274E-01 - 8.17325748757804E-01*i)

5. Solving m a n y s y s t e m s . Using PHCpack from within a MATLAB

or Octave session provides novel opportunities to solve polynomial systems.
In this section we show how the scripting environments can help to control
the quality of the developed software. The high level parallel programming
capabilities of MPITB will speed up this process in a convenient manner.
5.1. A u t o m a t i c t e s t i n g a n d b e n c h m a r k i n g . The scripting lan-
guage of MATLAB and Octave lends itself very directly to automatically
solving many polynomial systems, as one would do for benchmarking pur-
poses.
We introduce another PHClab function: read_system which reads a
polynomial system from file. The value returned by this function can be
passed to the blackbox solver. The system on file must have the following
22 YUN GUAN AND JAN VERSCHELDE

format. On the first line we have two numbers: the number of equations
and variables. Thereafter follow the polynomials, each one is terminated
by a semicolon. For example, the system ^(x) = 0 is represented as
2 2
1.3*xl**2 + 4.7*x2**2 + (-3.1+2.3*1);
2.1*x2**2 - 1 . 9 * x l ;
Note that 1 (and I) may not be used to denote variables, as they both
represent the imaginary unit v ' ^ . Because e and E are used to denote
floating-point numbers, e and E may not by used as the start of names of
variables.
If /tmp/Demo contains the polynomial systems in the files with names
kulO, cycllcS, /tmp/Demo/fbrf lve4, /tmp/Demo/game4two, (taken from
the demonstration database^ at [21]), then the script with contents
f = {'/tmp/Demo/kulO'
'/tmp/Demo/cyclicS'
'/tmp/Demo/fbrfive4'
'/tmp/Demo/game4two'};
for k= l : s i z e ( f , 1 )
p = read_system(f{k});
t o = clock;
s = solve_system(p);
et = e t i m e ( c l o c k ( ) , t O ) ;
n = size(s,2);
f p r i n t f ('Found %d s o l u t i o n s for %s i n °/,f s e c . \ n ' ,
n,f{k},et);
end;
will produce the foBowing statistics:
Found 2 s o l u t i o n s for /tmp/Demo/kulO i n 1.819892 s e c .
Foimd 70 s o l u t i o n s for /tmp/Demo/cyclic5 i n 11.094403 s e c .
Found 36 s o l u t i o n s for /tmp/Demo/fbrfive4 i n 18.750158 s e c .
Found 9 s o l u t i o n s for /tmp/DeBo/game4two i n 1.630962 s e c .
5.2., Parallel scripting with M P I T B . MPITB for Octave [2] ex-
tends Octave environment by using DLD functions. It allows Octave users
in a computer cluster to build message-passing based parallel applications,
by the means of installing the required packages and adding MPI calls to
Octave scripts. To use MPITB for Octave, djmamically linked LAM/MPI
libraries are required. All nodes in the cluster need to be able to access the
custom-compiled Octave that supports DLD functions.
Our choice of MPITB for Octave was motivated primarily by its func-
tionality and availability through open source. In our testing environment,
the latest MPITB for Octave was compiled against LAM/MPI 7.1.2 and

^available at http://www.math.uic.edu/~jaB/demo.html.
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 23

Octave 2.1.64. To illustrate conducting parallel computation with the com-

bination of MPITB, PHClab and Octave, we used origami equations [1].
The main script is small enough to be included here:
function origami
•/.
7. mpirun -c <nprocs> octavG-2.1.64 -q —funcall origami
%
% The manager distributes the systems to the worker nodes
% using dynamic load balancing. Every node writes the
% solutions to file when its job is done and sends a
7. message to the manager asking % for the next job.
%
tic 7. start the timer
info = MPI.Init; 7. MPI startup
[info rank] = MPI_Comm_rank(MPI_COMM_WORLD);
[info nprc] = MPI_Comm_size(MPI_COMM_WORLD);
pathCLOADPATH,Vhuis/phcpack/PHClab');
set_phcpath('/huis/phcpack/PHCv2/bin/phc');
if rank == 0 'I, code for t h e manager
origamisys = e x t r a c t _ s y s ( ' a l i g n m e n t e q u a t i o n s . t x t ' ) ;
distribute_tasks(nprc,origamisys);
f p r i n t f ( ' e l a p s e d time = %.3f s \ n ' , t o c ) ;
else
worker_solves_system() ; 7. code for t h e workers
end
info = MPI_Finalize;
LAM.Clean;
quit
end
Each origami system described in [1] has 4 inhomogeneous equations
in 4 variables and other free parameters. The mixed volume of Newton
Polytopes serve as a sharp upper bound for the number of solutions of
these origami systems because of the generic parameters. The output of a
run oh our Rocketcalc cluster configuration with 13 workers is below:
prom]3t$ mpirun -c 13 o c t a v e - 2 . 1 . 6 4 -q —funcall origami
Task tallies:
nO 18 (local)
nOl 14
n02 14
n03 14
n04 11
n05 12
n06 13
n07 13
24 YUN GUAN AND JAN VERSCHELDE

n08 12
n09 15
»iO 15
nil 17
nl2 15
sum 183 (SIZE 183)
elapsed time = 371.603 s
6. A numerical irreducible decomposition. There is not (yet) a
blackbox solver in PHCpack to compute a numerical irreducible decompo-
sition. In the subsections below we describe the functions which call the
tools of phc. We start by defining how we represent positive dimensional
solution set.
6.1. W i t n e s s s e t s . To obtain a numerical representation of a posi-
tive dimensional solution set, we add as many random hyperplanes as the
expected top dimension. Extra slack variables are introduced to turn the
augmented system into a square system (i.e.: having as many equations as
unknowns) for which we may then apply the blackbox solver.
We illustrate our methods on a special Stewart-Gough platform, which
are "architecturally singular" like the so-called Griffis-Duffy platform [4],
analyzed in [6]; also see [17]. Once a witness set has been computed, the
numerical irreducible decomposition in PHCpack apphes monodromy [15]
and linear traces [16].
A witness set consists of a polynomial system and a set of solutions
which satisfy this system. The polynomial system contains the original
polynomial system augmented with hyperplanes whose coefficients are ran-
domly chosen complex numbers. The number of hyperplanes added to the
original system equals the dimension of the solution set. The number of
solutions in the witness set equals the degree of the solution set.
There are two methods to compute witness sets. The (chronologically)
first method is to work top down, starting at the top dimensional solution
component and using a cascade [14] of homotopies to compute (super)
witness sets as numerical representations of solution sets of all dimensions.
The second method works top down, processing equation by equation [19].
6.2. Top down c o m p u t a t i o n using a cascade. The input to em-
bed is a system of 8 equations^ and the number 1, which is the expected
top dimension. We solve the embedded system with solve_system and
then run cascade to look for isolated solutions.
S = read_system('gdplatB'); '/. read the system from f i l e
E = embed(S,l); '/. embed with 1 e x t r a hyperplane
s o l s = solve_system(E); 7. c a l l t h e blackbox s o l v e r
s i z e ( s o l s , 2) 7, see candidate witness #points

^Maple code to generate the equations is at http://wwu.math.uic.edu/

.^ jan/FactorBench/grif duf Ael. html.
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 25

[sw,R] = c a s c a d e ( E , s o l s ) % perforin a cascade

The blackbox solver returns 40 solutions of the embedded system,
which turns out the degree of the one dimension curve, because cascade
finds no other isolated solutions. This can be read from the output shown
on screen:
ans

40

sw =

[]
[1x40 s t r u c t ]

R=

[]
{9x1 c e l l }
The function cascade returns two arrays. The first array contains the
solutions, while the second one contains the embedded systems. A witness
set for a fc-dimensional solution is defined by the {k + l)-th entries of the
arrays returned by cascade.
The top down approach has the disadvantage that it requires the user
to enter an expected top dimension. While in many practical applications
one can guess this top dimension from the context in which the apphcation
arises, the default value - taking it as high as the number of variables minus
one - is often too expensive.
6.3. Bottom up computation: Equation-by-equation. The new
equation-by-equation solver [19] relieves the user from submitting a top
dimension and seems more flexible. A disadvantage of the solver is that its
performance depends on the order of equations. For the equation describing
our Griflas-Dufiy platform, we move the simplest equations first.
p = read_system('gdplatBa')
[sw,R] = eqnbyeqn(p)

' g0*h0+gl*hl+g2*h2+g3*h3'
' g0"2+gl*2+g2-2+g3-2-h0-2-hl-2-h2-2-h3-2'
[1x102 char]
[1x308 char]
[1x333 Char]
[1x333 char]
[1x308 char]
26 YUN GUAN AND JAN VERSCHELDE

[1x308 char]

sw =

[]
[1x40 struct]

R=

[]
{9x1 c e l l }
6.4. Factoring into irreducible components. We continue with
the output (sw,R), computed either with cascade or eqnbyeqn.
dc = decompose(R{2},sw{2,l})

ans =

40

irreducible factor 1:

ans =

1x28 struct array with fields:

time
multiplicity
err
rco
res
hO
hi
h2
h3
g3-
gl
g2
go

zzl

irreducible factor 2:

ans =
time: 1
multiplicity: 1
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 27

err: 5.103000000000000e-15
rco: 1
res: 3.598000000000000e-15
hO: -3.091000000000000e-01 - 2.563000000000000e-01i
hi: -4.439300000000000e-01 + 5.353800000000000e-01i
h2: 2.5630000000000006-01 - 3.091000000000000e-01i
h3: -5.353800000000000e-01 - 4.439300000000000e-01i
g3: 4.766100000000000e-01 + 8.732700000000000e-01i
gl: 1.164500000000000e+00 - 2.351500000000000e-01i
g2: -3.3606000000000006-01 + 4.145700000000000e-01i
gO: -3.6430000000000006-03 + 8.404300000000000e-01i
zzl: 8.1717000000000006-16 + 5.026400000000000e-16i

.. % 12 more similar linear factors not shown to save space

dc =

[1x28 s t r u c t ]
[1x1 s t r u c t ]

. . 7. 12 more s i m i l a r s t r u c t s not shown t o save space

The output of decompose shows one irreducible component of degree 28
and 12 lines.

Acknowledgments. This paper was written while the second author

was a long term visitor at the IMA in the Fall of 2006, participating to
the annual thematic program on AppUcations of Algebraic Geometry. The
first author is grateful for financial support to present this work at the
IMA software workshop, held at the IMA, 23-27 October, 2006. The first
author also thanks Javier Fernandez Baldomero, the author of MPITB,
and Michael Creel for helpful communication. We thank Bernd Sturmfels
for forwarding origami equations [1] to the IMA workshop "Software for
Algebraic Geometry" as a challenge problem.

APPENDIX

A. Alphabetic List of PHClab Functions. Below is an alphabetic

list of the functions offered by PHClab.
cascade executes a sequence of homotopies, starting at the top dimen-
sional solution set to find super witness sets. The input consists
of an embedded system (the output of embed) and its solutions
(typically obtained via solve jsystem. The output of this function
is a sequence of super witness sets. A witness set is a numerical
28 YUN GUAN AND JAN VERSCHELDE

representation for a positive dimensional solution set. The "su-

per" means that the k-th super witness set may have junk points
on solutions sets of dimension higher than k.
decompose takes a witness set on input and decomposes it into irreducible
factors. The witness set is represented by two input parameters:
an embedded system and solutions which satisfy it. The number
of solutions equal the degree of the pure dimensional solution set
represented by the witness set. On return is a sequence of witness
sets, each witness set in the sequence corresponds to one irreducible
component.
deflation reconditions isolated singular solutions. The input consists of
two parameters: a polynomial system and a sequence of approx-
imate solutions to the system. Typically these solutions are ob-
tained via the blackbox solver or as the output of the function
track. On return is a list of reconditioned solutions, along with
the augmented systems which have as regular solution the multiple
solution of the original system.
embed adds extra hyperplanes and slack variables to a system, as many
as the expected top dimension of the solution set. There are two
input parameters: a polynomial system and the number of hy-
perplanes which have to be added. Typically, this number is the
top dimension of the solution set. If nothing is known about this
top dimension, a default value for this number is the number of
variables minus one.
eqnbyeqn solves polynomial systems equation by equation. For the poly-
nomial system on input, this function returns a sequence of witness
sets. The fcth witness set in the sequence is a numerical represen-
tation of the solution set of dimension k.
make_system converts the matrix format of a system into a symbohc
format acceptable to phc. A polynomial system of N equations
in n variables, with a total of m terms, is represented by a matrix
with N+m rows and n+1 columns. Each polynomial is terminated
by a zero row in the matrix. Each row represents one term in a
poljmomial, starting with its (complex) coefficient and continuing
with the values of the exponents for each variable.
mixed_volume computes the mixed volume for a system of n equations
in n variables. There are two input parameters: the system and
a flag to indicate whether a random coefficient system must be
created and solved. If the flag on input is one, then on return is a
start system which has as many solutions as the mixed volume.
phcJilter removes from a super witness set those junk points while lie
on a higher dimensional solution set. The third and last input
parameter is a set of points to be filtered. The first two param-
eters represent a witness set, given by an embedded system and
a sequence of solutions which satisfy the embedded system. On
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 29

return are those points of the third input that do not he on the
component represented by the witness set.
read_system reads a poljoiomial system from file. There is only one input
parameter: a file name. The format of the pol5Tiomial system
on file must follow the input format of PHCpack. The function
returns an array of strings, each string in the array is a multivariate
polynomial in symbolic format.
refine_sols applies Newton's method to refine a solution. There axe six in-
put parameters: a polynomial system, an approximate solution, a
tolerance for the magnitude of the correction vector e r r , a thresh-
old for to decide whether a solution is singular (relative to rco), a
tolerance for the residual r e s , and finally a natural number with
the maximal number of Newton iterations that are allowed. On
return is an array of refined approximate solutions.
set_phcpath defines the directory where the executable version of phc
is. For example, if the program phc is in the directory /tmp,
then set_phcpath('/tmp/phc') must be executed at the start of a
PHClab session. On Windows, ' /tmp/phc' could be replaced by
' C: /Downloads/phc' if phc. exe is in the directory Downloads on
the C drive.
solvejsystem calls the blackbox solver of phc. On input is a polyno-
mial system in matrix format, see the input description for the
command make_systein. An alternative input format is the cell
array returned by read_system. The output is an array of struc-
tuies. Every element in the array contains one solution at the end
of a solution path. In addition to the values for the coordinates of
the solution, an estimate for the condition number of the solution
which leads to a measure for the forward error, while the residual
measures the backward error.
t r a c k applies numerical continuation methods for a homotopy between
start and target system, for a specified set of start solutions. The
three arguments for track are respectively the target system, the
start system and the solutions of the start system. The target and
start system must be given in matrix format. If the start solutions
are singular, then the path tracker will fail to start. The output of
t r a c k is an array of the same length as the array of start solutions,
containing the values at the end of the solution paths.

B . Exercises.
1. Use the blackbox solver to solve (the phc input format is on the
right):

x^ly^-lJo x-2-Fy-2-l; (B.l)

^ +^ '• ^ x-3 + y - 3 - l ;
30 YUN GUAN AND JAN VERSCHELDE

The exact solutions are (1,0), (0,1), (—1 + i^,—l — i^) and

How many solutions does solve_system return? Verify whether

the output matches the exact solutions. Use refine_sols to dis-
cover what the multiplicities of the solutions (0,1) and (1,0) are.
2. If we have to solve repeatedly a polynomial system with the same
structure, we may want to save a start system. To solve systems
with the same monomials as in (B.l), we could use

, X _ / x2 + 1.232J/2 + l.uili = 0 ,„ _,
9[^, y)-<^y3_ o.872y2 - 0.6231 + 1.032i = 0 ^^'"^^

Since the coefficients are random complex numbers (feel free to

make other random choices) all solutions of the system ^(ar, y) =0
will be regular.
(a) Solve the system g{x,y) = 0, using solve_system. Verify
that all solutions are regular.
(b) Use track to solve the system in (B.l).
Check whether you find the same solutions, eventually com-
puted in a different order.
3. The following system has multiple roots:

x2 + y - 3 = 0
a;-I-0.125y2 - 1.5 = 0 ^ '

(a) Use solve_system to find approximate roots. Can you see

which roots are multiple?
(b) Apply deflation to the approximate roots.
Observe the values of the field rco of the solutions before and
after the deflation.
(c) What is the multiplicity of each solution?
4. All adjacent minors of an indeterminate 2-by-4 matrix for a system
of 3 equations in 8 variables:

X11X22 - a;2ia;i2 = 0
a;i2a;23 - a;22Xi3 = 0 (B.4)
a;i3a;24 — a;23a;i4 = 0.

(a) Use embed to add 5 random hyperplanes.

(b) Solve the embedded system. What is the degree of this 5-
dimensional solution set?
(c) Apply decompose to factor the solution set. How many
irreducible factors do you find?
(d) Repeat the process for larger instances of this problem, for
n = 5,6, —
 PHCLAB: A MATLAB/OCTAVE INTERFACE TO PHCPACK 31

5. Consider the system

(xl - a;2)(a;i - 0.5) = 0
{xf - X3)(x2 - 0.5) = 0 (B.5)
(xiX2 - a;3)(x3 - 0.5) = 0.
Solving this system means to compute witness sets for all irre-
ducible factors.
(a) Use embed to add 1 random hyperplane.
(b) Solve the embedded system with solve_system.
Among the solutions, can you see the three witness points on
the twisted cubic?
Look for solutions with a slack variable close to zero.
(c) Apply CEiscade to find candidate isolated solutions.
(d) Use phc_filter to filter the candidate isolated solutions.
How many isolated solutions does the system have?

REFERENCES

[1] R.C. ALPERIN AND R.J. LANG. One and Two-Fold Origami Axioms 2006 4OSME
Proceedings, Pasadena, CA, A.K.Peters, 2007.
[2] J. FERNANDEZ, M . ANGUITA, E . R O S , AND J . L . BERNIER. S C E Toolboxes for the
development of high-level parallel applications. Proceedings of ICCS 2006,
Volume 3992 of Lecture Notes in Computer Science, pages 518-525, Springer-
Verlag, 2006
[3] T. GAO, T . Y . L I , AND M . W U . Algorithm 846: MixedVol: a software package for
mixed-volume computation. ACM Trans. Math. Softw., 31(4):555-560, 2005.
[4] M. GRIFFIS AND J. DUFFY. Method and apparatus for controlling geometrically
simple parallel mechanisms with distinctive connections. US Patent 5,179,525,
1993.
[5] B. HUBER AND B . STURMFELS. A polyhedral method for solving sparse polynomial
systems. Math. Comp., 64(212):1541-1555, 1995.
[6] M.L. HusTY AND A . KARGER. Self-motions of Griffis-Duiiy type parallel manipu-
lators. PTOC. 2000 IEEE Int. Conf. Robotics and Automation, CDROM, San
FVancisco, CA, April 24-28, 2000.
[7] A. LEYKIN AND J. VERSCHELDe. PHCmaple: A Maple interface to the numerical
homotopy algorithms in PHCpack. In Quoc-Nam Tran, editor. Proceedings
of the Tenth International Conference on Applications of Computer Algebra
(ACA '2004), pages 139-147, 2004.
[8] A. LEYKIN AND J. VERSCHELDE. Interfacing with the numerical homotopy algo-
rithms in PHCpack. In Nobuki Takayama and Andres Iglesias, editors. Pro-
ceedings of ICMS 2006. Volume 4 1 5 1 of Lecture Notes in Computer Science,
pages 354-360, Springer-Verlag, 2006.
[9] A. LEYKIN, J. VERSCHELDE, AND A. ZHAO. Newton's method with deflation for
isolated singularities of polynomial systems. Theoretical Computer Science
359(1-3):111-122, 2006.
[10] T.Y. L I . Numerical solution of polynomial systems by homotopy continuation
methods. In F. Cucker, editor. Handbook of Numerical Analysis. Volume XI.
Special Volume: Foundations of Computational Mathematics, pages 209-304.
North-Holland, 2003.
[11] M. MAEKAWA, M . N O R O , K . OHARA, Y . OKUTANI, N . TAKAYAMA, AND Y . TAMURA.
OpenXM - an open system to integrate mathematiccJ softwares. Available at
http://www.OpenXM.org/.
32 YUN GUAN AND JAN VERSCHELDE

[12] M. MAEKAWA, M . N O R O , K . O H A R A , N . TAKAYAMA, AND Y . TAMURA. T h e de-

sign and implementation of OpenXM-RFC 100 and 101. In K. Shirayanagi
and K. Yokoyama, editors, Computer mathematics. Proceedings of the Fifth
Asian Symposium (ASCM 2001) Matsuyama, Japan 26-28, Septem.ber 2001,
Volume 9 of Lecture Notes Series on Computing, pages 102-111. World Sci-
entific, 2001.
[13] J.M. MCCARTHY. Geometric Design of Linkages. Volume 11 of Interdisciplinary
Applied Mathematics, Springer-Verlag, 2000.
[14] A . J . SOMMESE AND J. VERSCHELDE. Numerical homotopies to compute generic
points on positive dimensional algebraic sets. J. of Complexity, 16(3):572-
602, 2000.
[15] A.J. SOMMESE, J. VERSCHELDE, AND C.W. WAMPLER. Using monodromy t o de-
compose solution sets of polynomial systems into irreducible components. In
C. Ciliberto, F . Hirzebruch, R. Miranda, and M. Teicher, editors. Application
of Algebraic Geometry to Coding Theory, Physics and Computation, pages
297-315. Kluwer Academic Publishers, 2001. Proceedings of a NATO Confer-
ence, February 25-March 1, 2001, Eilat, Israel.
[16] A.J. SoMMESE, J. VERSCHELDE, AND C.W. WAMPLER. Symmetric functions ap-
plied t o decomposing solution sets of polynomial systems. SIAM J. Numer.
Anal, 40(6):2026-2046, 2002.
[17] A.J. SOMMESE, J. VERSCHELDE, AND C.W. WAMPLER. Advances in polynomial
continuation for solving problems in kinematics. AS ME Journal of Mechanical
Design 126(2):262-268, 2004.
[18] A.J. SOMMESE, J. VERSCHELDE, AND C.W. WAMPLER. Numerical irreducible de-
composition using PHCpack. In M. Joswig and N. Takayama, editors, Algebra,
Geometry, and Software Systems, pages 109-130. Springer-Verlag, 2003.
[19] A . J . SOMMESE, J. VERSCHELDE, AND C.W. WAMPLER. Solving polynomial sys-
tems equation by equation. Accepted for publication in t h e IMA volume on
Algorithms for Algebraic Geometry.
[20] A.J. SOMMESE AND C.W. WAMPLER. The Numerical solution of systems of poly-
nomials arising in engineering and science. World Scientific Press, Singapore,
2005.
[21] J. VERSCHELDE. Algorithm 795: PHCpack: A general-purpose solver for poly-
nomial systems by homotopy continuation. ACM Trans. Math. Softw.,
25(2):251-276, 1999. Software available at h t t p : / / w w w . m a t h . u i c . e d u / ~ j a n .
 C O M P U T I N G G R O B N E R FANS A N D
TROPICAL VARIETIES IN GFAN*
ANDERS NEDERGAARD JENSENt

A b s t r a c t . The Grobner fan of an ideal in the polynomial ring in n variables is

an n-dimensional polyhedral complex and the tropical variety of the ideal is a certain
subcomplex. In this paper we describe the software Gfan for computing these fans.
Computing the Grobner fan is equivalent to computing all the reduced Grobner bases
of the ideal.

K e y w o r d s . Grobner ians, tropical geometry, Gfan, software.

1. Introduction. Gfan [13] is a software package for computing

Grobner fans and tropical varieties of polynomial ideals. The Grobner
fan of an ideal / C Q[xi,... ,Xn] is a polyhedral complex defined in [15].
For a homogeneous ideal the Grobner fan is a complete fan and the normal
fan of a polytope. Its cones are in bijection with the various initial ideals of
I. In particular, the full-dimensional cones are in bijection with the mono-
mial initial ideals and thereby also in bijection with the reduced Grobner
bases of I. In [4] the local basis change of Grobner bases was introduced.
This method allows us to go from one Grobner basis in the fan to a neigh-
boring one, giving an effective algorithm for computing the Grobner fan by
traversing its maximal cones. This algorithm is the foundation of Gfan. A
detailed description of the algorithm is given in [6].
A computation of the Grobner fan of / is useful when searching for an
initial ideal with a particular property. The computer program TiGERS
[11] is an earher implementation of the Grobner fan traversal for toric ideals.
Gfan is more general as it can compute the Grobner fan of any polynomial
ideal I C Q [ x i , . . . , a;„]. Taking the union of all reduced Grobner bases of
/ we get a universal Grobner basis, that is, a set which is a Grobner basis
with respect to any term order.
Recently the field of tropical mathematics has received much attention.
In tropical mathematics the semi-ring (M, max, -I-) is considered. Here max-
imum .takes the role of addition and plus the role of multiplication. Hence
tropical polynomials define piecewise linear functions. In tropical math-
ematics classical objects have tropical analogs. For example the tropical
variety T{I) of a polynomial ideal / C Q[xi,..., a;„] is the analog of the
usual variety of the ideal. A tropical variety is a piecewise linear object.
There are many equivalent ways of defining the tropical variety of / . We
prefer to state the definition in terms of initial ideals:

'Research partially supported by the Fticulty of Science, University of Aarhus, Dan-

ish Research Training Council (Forskeruddannelsesradet, FUR), Institute for Operations
Research ETH, the Swiss National Science Foundation Project 200021-105202, and In-
stitute for Mathematics and its Applications, University of Minnesota.
tfnstitat fiir Mathematik, Technische Universitat Berlin, D-10623, Germany.
33
34 ANDERS N. JENSEN

T{I) := {tj e R" : in^il) contains no monomials}

where in^;(/) denotes the initial ideal of / with respect to the vector uj. This
definition appeared in [17]. The fact that the tropical variety is a union of
Grobner cones ensures that the tropical variety can be given the structure
of a polyhedral complex making it a subfan of the Grobner fan. An ex-
haustive traversal algorithm for computing tropical varieties of prime ideals
was developed in [3]. The algorithm rehes on the relation to Grobner fans,
a connectivity result, constructions of tropical bases, polyhedral computa-
tions and lifting results. Gfan contains the only existing implementation
of this algorithm.
Gfan is a collection of command line tools written in C-l—1-. The li-
braries GMP [8] and cddlib [5] are used for exact arithmetic and polyhedral
computations. Gfan contains a simple implementation of Buchberger's al-
gorithm which is sufficient for the kind of enumerations Gfan does but
which is not competitive to implementations found in other algebra soft-
ware. Gfan is released under the GNU GPL license, runs on Linux and
Mac OS X, and can be compiled with newer versions of gcc.
This paper is organized as follows. The first section gives an intro-
duction to Grobner fans while illustrating how these may be computed in
Gfan. After this follows a similar section about tropical varieties. We then
discuss some of the algorithms in Gfan, give a few benchmark examples
and discuss how Gfan can interface other mathematical programs. This
paper is an extended version of [14].
2. Computing Grobner fans. The Gfan package consists of several
command Une programs. The most important one is gfan which takes
generators for an ideal / C Q [ x i , . . . , x„] and computes the complete set of
reduced Grobner bases for I.
EXAMPLE 1. Running the command gfan and typing in the input
{a-b-ab, a~2+ab} produces the list
{{b-3-2*b-2.
a-b+b'2}

{b~2-b+a,
a*b+b-a,
a~2-b+a}
i

{b-a-o.~2,
a.'3+2*a~2)}
of all marked reduced Grobner bases for the ideal generated by the input
polynomials.
By a marked Grobner basis we mean a basis where each polynomial has
a distinguished, marked term. In Gfan the marked term of a polynomial is
the first term listed. Furthermore, the marked term should be the initial
term of the polynomial with respect to the term order.
 COMPUTING GROBNER FANS AND TROPICAL VARIETIES IN GFAN 35

Typically a user will ask Gfan to manipulate files instead of reading

from the keyboard and writing to the screen. This is achieved in the shell
by redirecting the standard input and output as follows.
gfan < g e n e r a t o r s . t x t >listofbases.txt
Here the ideal is read in from the file g e n e r a t o r s . t x t and the output is
written to the file l i s t o f b a s e s . t x t .
We will explain the connection to polyhedral geometry in the following.
Recall that given a term order -< on Q [ x i , . . . , x„] and a vector uj € K" Q
we may define the term order -<^:

x" X(^ x" •» u • a; < t; • w V (u • w = ?; • w A x" ^ x") for u,v € N".

Here we use the notation x*" = x]"^ • • • xj^" for ty € N". The term order -<^
is called the refinement of ui with respect to -<. Furthermore, recall that
Buchberger's algorithm for computing a reduced Grobner basis takes as
input an ideal and a term order. If the term order is of the above form -<i^,
then diSerent u may lead to different reduced Grobner bases. The Grobner
fan of / encodes which vectors give the same marked reduced Grobner
basis. For each marked reduced Grobner basis Q we consider the region of
vectors in E " Q that will produce G with Buchberger's algorithm. To avoid
ambiguities we take the closure of this region and call it the Grobner cone
of G- The collection of all Grobner cones is called the (restricted) Grobner
fan of / . We shall give a more precise definition in the case of homogeneous
ideals later.
It is well-known (see [11]) that we have bijections between
• the marked reduced Grobner bases of I,
• the full-dimensional cones in the Grobner fan of 7, and
• the monomial initial ideals of I (with respect to term orders).
Here the bijection between the marked Grobner bases and the initial ideals
follows from uniqueness of reduced Grobner bases while the bijection be-
tween the two first sets follows from the definition of the Grobner fan and
an argument that the Grobner cones for term orders are full-dimensional;
see [19, Proposition 1.13]. The bijections are illustrated in Figure 1.
In Example 1 we computed the Grobner fan of I, although, it was
given to us in a rather inconvenient way as a collection of marked Grobner
bases. In the next example we show how to compute defining equations for
the, Grobner cones. Notice that in this paper we describe the behavior of
Gfan version 0.2.2. Future versions will be slightly different.
EXAMPLE 2. Consider the first marked reduced Grobner basis in
Example 1:

G = {^-2b'^,a-b + b'^}.

We may ask ourselves for which term orders -<^ this is a Grobner basis.
An important observation is that Buchberger's S-pair criterion only depends
36 ANDERS N. JENSEN

FIGURE 1. The Grobner fan in R?.^ of the ideal in Example 1. For each Grobner
cone its monomial initial ideal has been listed.

on the marked terms - not on the term order. Hence, any w G K?.o that
picks out the marked terms as initial monomials will define a term order
for which Q is a Grobner basis. To pick out the marked terms, ui must
satisfy:

3a;2 > 2w2 ,

wi > W2 , and
uji > 2a;2 •

These inequalities, of which wi > W2 is redundant, define the open Grobner

cone of Q. The command gfan.groehnercone will take the marked re-
duced Grobner basis as input and produce the defining inequalities while
gfan-facets will cut the set of inequalities down to the ones defining facets.
The following command line combines the programs in the shell and pro-
duces the list { (0,1), (1, -2)} of inner facet normals:
gfcin.groebnercone I gfaii_facets
We now show how to use Gfan for making drawings of Grobner fans.
EXAMPLE 3. In the case of three variables Gfan can draw a picture of
the positive orthant for us. For example:
gfan <input.txt I gfaii_render > p i c t u r e l . f i g
produces the Xfig drauiing of the Grobner fan of the ideal in i n p u t . t x t ;
see Figure 2.
Sometimes the ideals we consider have symmetry. This symmetry can
be exploited by Gfan. The advantage is that the computation is faster and
that the output may be more readable.
EXAMPLE 4. Consider the ideal I = {a^b — c,fe^c— a, (P'a — b). The
symmetric group S3 acts on the polynomial ring Q[a,b,c] by permuting
variables. It also acts on R2.O ^2/ permuting coordinates. The alternating
subgroup A3 C S3 keeps I fixed. As a result A3 acts on the set of Grobner
cones of I. We wish to exploit the A3 C ^3 symmetry. We will use the
command line gfan —symmetry I gfan-render > output.fig and the
 COMPUTING GROBNER FANS AND TROPICAL VARIETIES IN GFAN 37

FIGURE 2. The Grobner fan of the ideal {xf + 3;| + a;| - 1, i f + a;| + 0:3 — 1, xf +
ccl + 1 3 — 1) intersected with the standard 2-simplex in R^. See [19, Example 3.9].

FIGURE 3. In Example 4 Gfan only produces one Grobner cone in each orbit with
respect to the action of A3. For this reason cones are missing in the picture.

input mil be a list of generators for the ideal plus a list of permutations
generating A3. In general we will need several generators to describe a
subgroup of Sn, but in the case of A3 one generator suffices.
{a"2b-c,b"2c-a,c"2a-b}
{(1.2,0)}
The output is the Xfig-file shown in Figure 3.
Gfan outputs only one Grobner cone in each orbit. When a new
Grobner basis is computed it is checked if it is in the orbit of a previously
computed basis. In this way Gfan avoids computing the entire Grobner
fan. Checking if two Grobner bases are in the same orbit is time consuming
put cheap compared to doing the sjonmetric Grobner basis computations
needed for a complete Grobner fan traversal. The symmetry techniques
used in Gfan are similar to those used in TOPCOM [16] for computing the
regular triangulations of a point configuration up to symmetry.
Before we move on to discuss tropical varieties, we will need the formal
definition of the Grobner fan of a homogeneous ideal as it was defined by
Mora and Robbiano [15]. Recall that for w e E", irK;{I) is the initial ideal
38 ANDERS N. JENSEN

generated by the w-initial forms of elements in I and that this initial ideal
does not have to be a monomial ideal.
DEFINITION 2.1. Let I C Q[a;i,... ,x„] he a homogeneous ideal. We
define the equivalence relation ~ on K" as follows:

u ~ v <^ in„(/) = in„(/),

where u,v S K". The equivalence classes are relatively open polyhedral
cones in R". Their usual topological closures are called the Grobner cones
of I. The set of all Grobner cones form the Grobner fan of I.
A Grobner fan is indeed a polyhedral fan with nice intersection prop-
erties of its cones. See [19] or [6] for a proof. In particular it contains cones
of different dimensions and not just the full-dimensional cones we discussed
earlier. For a homogeneous ideal, Gfan lets you compute the f-vector of the
fan with the line.
gfan I gfan_fvector
The intersection of all cones in the Grobner fan of an ideal / is again a
Grobner cone. We call this Grobner cone the homogeneity space of / . It
may be computed with the command gf anJhomogeneityspace which takes
a reduced Grobner basis of the ideal / as input. The homogeneity space of
I consists of all vectors which induce gradings on Q [ x i , . . . , x„] for which I
is homogeneous. The homogeneity space is a subspace and a face of every
cone in the Grobner fan.
A note on the different definitions of Grobner fans related to Gfan
and the connection between them is in place. For a homogeneous ideal
the restricted Grobner fan which we started out by defining is gotten by
taking the common refinement of the Grobner fan defined above and the fan
consisting of the faces of the non-negative orthant. In [6] a third uniform
Grobner fan definition is given which for homogeneous ideals coincides with
the Grobner fan definition above and for non-homogeneous ideal has the
property that any cone is the closure of an equivalence class in the sense
above. Again taking the common refinement of this fan with the non-
negative orthant we get the restricted Grobner fan. While the last uniform
definition has the nicest properties, firom a practical point of view it is less
important which definition we use since for all three the maximal cones are
in bijection.
As a last feature for computing Grobner fans we should mention the
program gfan_interactive. This program lets the user waUc around in
the full-dimensional cones of the Grobner fan. At each cone the user gets to
choose a facet to walk through. This is useful when studying the Grobner
fan locally or when searching for an initial ideal with a special property.
3. Computing tropical varieties. The tropical variety of an ideal
7 C Q [ x i , . . . , Xn] is a certain subset of R", namely, it consists of all a; € R"
such that ixhj{I) does not contain a monomial. Besides considering the
 COMPUTING GROBNER FANS AND TROPICAL VARIETIES IN GFAN 39

tropical variety as a set we may also consider it as a subfan of the Grobner

fan by considering all Grobner cones whose initial ideal is monomial-free.
Since the tropical variety is a closed set, this actually defines a subfan.
There is nothing special about the fan structure that is induced by the
Grobner fan — there may be other coverings of the tropical variety with a
fewer number of cones. An example of this is the ideal in [3, Example 20]
which is generated by the 3 x 3-minors of a generic 4 x 4 Hankel matrix. Gfan
uses the structure induced by the Grobner fan for its tropical computations.
More generally one may be interested in ideals in polynomial rings over
different fields, for example over a field with a valuation being taken into
account when defining initial ideals. Most examples in practice, however,
reduce to the situation of ideals in Q[a;i,... ,x„].
Gfan can handle fom: cases of tropical varieties:
• tropical hyper surfaces,
• intersections of tropical hypersurfaces,
• tropical curves, and
• varieties defined by prime ideals.
We shall discuss these cases below.
In the special case of a principal ideal / = (/) computing the tropical
variety is straight forward. In this case the variety is called a tropical
hypersurface. It is the codimension 1 skeleton of the normal fan of the
Newton poljd^ope of / . We may compute it in the following way with Gfan.
EXAMPLE 5. If we run gfan-tropicalintersection —incidence
on the input
{a+b+c+d} we get the following as a part of the output:
Rays:
{
0: ( - 1 . 0 . 0 , 0 ) ,
1: ( 0 . - 1 . 0 , 0 ) ,
2: ( 1 , 1 , 1 , 0 ) .
3: ( 0 , 0 , - 1 , 0 ) }
Printing index list for dimension 3 cones:
i
•CO.l},
{0.2}.
{0.3},
{2,3}.
{1.2}.
{1.3}}
F-vector:
(4.6)
The first list is the set of rays in the tropical hypersurface. The sec-
ond list shows how these rays are combined to form the maximal cones
in the hypersurface. Notice that the homogeneity space of the ideal is 1-
dimensional. Hence, the rays are really 2-dimensional and the maximal
40 ANDERS N. JENSEN

cones have dimension 3. For example the first ray (—1,0,0,0) is the cone
{ s ( - l , 0 , 0 , 0 ) + i ( l , 1,1,1) where {s,t) G R>o x K}. / / the hypersurface
consists of cones of other dimensions, then these will also be listed.
The command gf a i i _ t r o p i c a l i n t e r s e c t i o n is actually more general
than illustrated above. It takes a list of polynomials, each defining a tropi-
cal hypersurface and computes the intersection of these hypersurfaces. The
result is a polyhedral fan which is the common refinement of all the input
tropical hypersurfaces. This intersection is not a tropical variety in general.
Take for example the two tropical hypersurfaces defined by the polynomials
a + b + ab^ and 6 + 1 . Their intersection is the non-negative a-axis which
is not a tropical variety since it does not satisfy the tropical zero-tension
condition at 0.
An alternative way of defining the tropical variety T(7) of an ideal
/ is as the intersection of all tropical hypersurfaces of polynomials in I.
A finite set of polynomials generating an ideal I is called a tropical basis
for / if the intersection of the hypersurfaces the polynomials define equals
T{I). Gfan contains the program gf a n _ t r o p i c a l b a s i s which will compute
a tropical basis for an ideal defining a tropical curve. Here an ideal is said
to define a tropical curve if it has dimension one larger than the dimension
of its homogeneity space. We should point out that it is important that
the ideal defines a curve. The reason for this can be found in the proof of
[3, Algorithm 5]. Basically, what is needed is that it suffices to check if a
single relative interior vector of a cone is in a tropical variety to know that
the cone is contained in the tropical variety. This is not true in general.
EXAMPLE 6. Consider the ideal I C Q[a,b,c,dl generated by {ab +
aa + CC + dd, ba + ca + db}. The homogeneity space of I has dimension 1
and the Krull dimension of Qla,b,c,d\/I is 2. Hence the ideal defines a
tropical curve and we may use gfan-tropicalbasis to compute a tropical
basis of I. This gives the following output:
i
a*b+d"2+c"2+a"2,
a*b+b*d+a*c,
a*2*b-l/2*c*d~2-l/2*c"3+b*c*d+l/2*b*c"2-l/2*b~2*d-l/2*a*c*d+a*c"2}
With gfan-tropicalintersection one can verify that the intersection
of the tropical hypersurfaces defined by the output polynomials is actually
smaller than the intersection of the hypersurfaces defined by the input poly-
nomials.
A theorem by Bieri and Groves [2] says that the tropical variety of
a d-dimensional prime ideal is pure of dimension d, which means that
all maximal cones have dimension d. Furthermore, Speyer's theorem [3,
Theorem 14] says that this pure ci-dimensional complex is connected in
codimension one. Being connected in codimension 1 means that any two
d-dimensional cones in the complex can be connected by a sequence of d-
dimensional cones with any two consecutive cones intersecting in a (d — 1)-
dimensional ridge.
 COMPUTING GROBNER FANS AND TROPICAL VARIETIES IN GFAN 41

These theorems form the basis for Gfan's methods for computing trop-
ical varieties of prime ideals. The following example shows how to combine
the programs gf a n _ t r o p i c a l s t a r t i n g c o n e and gf a n _ t r o p i c a l t r a v e r s e
to compute the tropical variety of a homogeneous prime ideal.
EXAMPLE 7. Let a,b,...,j he the 10 2x2 minors of the 2x5 matrix
Xii Xi2 Xiz Xii Xiz
a;21 X22 X23 X2A 3^25

For a suitable ordering they satisfy the Grassmann-Plucker relations:

0 = bf—ah—ce = bg — ai—de = cg—aj—df = d — bj—dh = fi — ej—gh.

The ideal generated by these relations is called the Grassmann-Plucker ideal

I C Q[a, • • .,j]. The ideal is prime and has dimension 7. Hence the tropical
variety of I is pure of dimension 7. To exploit the connectivity of the trop-
ical variety we first need to compute a single 7-dimensional cone ofT{I).
In the worst case this may be as difficult as computing T{I) itself. The
program gfan^tropicalstartingcone has a heuristic method for guessing
a starting cone. The idea is to build up the cone one dimension at a time
as explained in [3, Algorithm 9]. In each step a ray in the tropical variety
outside the homogeneity space is needed. To get this ray Gfan computes
extreme rays of the Grobner cone of the input Grobner basis and checks
if the rays are in the tropical variety. The heuristics work reasonable well
but may fail. In that case it is usually worthwhile to try a different input
Grobner basis for the program. For our example the program produces the
following output:
-Cf * i - e * j ,
d*h-c*i,
d*f+a*j,
d*e+a*i,
c*e+a*h)-
{f*i-g*h-e*j,
d*h-c*i+b*j,
d*f-c*g+a*j,
d*e-b*g+a*i,
c*e-b*f+a*h>
This is a pair of reduced Grobner bases that Gfan uses for representing the
d-dimensional starting Grobner cone. For n-dimensional Grobner cones
only a single Grobner basis is needed as we saw in Section 2. For lower-
dimensional cones the pair consists of a Grobner basis for the initial ideal
iiiuil) and a Grobner basis for I itself with respect to a term order -<u,
where u) is in the relative interior of the Grobner cone to be represented.
Having computed the starting cone we may run
gfari-tropicaltraverse which traverses the 7-dimensional cones and
produces an output in a format similar to the one in Example 5. The
42 ANDERS N. JENSEN

FIGURE 4. The tropical variety of the ideal in Example 7 drawn projectively.

homogeneity space of I has dimension 5 and all cones in the tropical variety
contains this subspace. Besides the homogeneity space the tropical variety
consists of 10 6-dimensional cones and 15 7-dimensional cones. Modulo
the homogeneity space the tropical variety is 2-dimensional. Projectively
we may draw its combinatorics as the Petersen graph; see Figure 4- The
homogeneity space is the center of the projection.
If the ideal is equi-dimensional, but not prime it may be that
the tropical variety is not connected in codimension 1. In that case
gf a n - t r o p i c a l t r a v e r s e computes a connected component. Just as it is
the case for Grobner fans Gfan is able to enumerate the cones of a tropical
variety up to symmetry.
For gf a n . t r o p i c a l t r a v e r s e to compute the tropical variety of an
ideal the ideal must be homogeneous and defined over the rationals. How-
ever most tropical papers consider the case where / is defined in the poly-
nomial ring over the Puiseux series field C{{i}}. In this case the valuation
of the coefficients is taken into account when defining initial forms and the
tropical variety is a polyhedral complex which need not consist of cones.
Thus, a priori, Gfan seems too restrictive. Notice however,
• the tropical variety of the homogenization of an ideal I equals, as
A set, the tropical variety of / except that it lives in a vector space
of dimension one higher; see [3, Lemma 4].
• the tropical variety of an ideal / C C{{f}}[a;i,...,a;„] generated
by elements in the polynomial ring over the rational functions
in t, C{t)[xi,... ,Xn], may be computed by considering the ideal
/ n C[i, xi,..., Xn] in C[t, xi,..., x„] and intersecting its tropical
variety with the t = 1 hyperplane, see [3, Lemma 1].
• if an ideal I is generated by elements in K[xi,.. .,Xn] where
K is an algebraic field extension of Q with an algebraic root a
then T{I) can be computed as the tropical variety of an ideal in
Q[a, xi,..., Xn] where the minimal polynomial of a has been added
as a generator, see [12, Lemma 3.12].
 COMPUTING GROBNER FANS AND TROPICAL VARIETIES IN GFAN 43

In addition Gfan can do computations in the polynomial ring

Z/pZ[a;i,.. .,Xn] where p is a prime. Hence, Gfan has the theoretic ability
to compute the tropical variety of most ideals we encounter in practice.

4. Algorithms. We briefly summarize the most important algorithms

implemented in Gfan. For the Grobner fan part we have the following
algorithms:
• a simple implementation of Buchberger's algorithm,
• an algorithm for finding facets of maximal Grobner cones,
• an implementation of the local basis change procedure of [4], and
• two algorithms for doing global enumerations of the maximal cones
in the Grobner fan.
Finding fax;ets of Grobner cones amounts to solving linear programming
problems. For this Gfan uses cddlib [5] which does its computations in
exact arithmetic.
The local basis change algorithm takes as input a Grobner basis Q
and a facet F of its Grobner cone. The output is the Grobner basis of the
cone on the other side of F. The algorithm consists of two steps. Step 1
is to compute a certain Grobner basis of the initial ideal int^(7) where u)
is in the relative interior of F. This Grobner basis computation turns out
to be relatively cheap. The reason for this is that ini^(7) is homogeneous
with respect to any grading given by a vector in F. In fact, the Newton
polytopes of the polynomials that need to be manipulated are all parallel
line segments. The second step is to lift the Grobner basis of ini^(/) to the
right Grobner basis of / . In Gfan turning the lifted basis into a reduced
basis is often the most time consuming part of the whole local basis change
process.
For the global enumeration of the Grobner fan two different approaches
are implemented. The straight forward approach is to make an exhaustive
enumeration of all maximal Grobner cones. This method has the advan-
tage that it is easy to exploit symmetry. The other approach is to apply
the reverse search technique [1]. The advantage is that reverse search is
memory-less in the sense that it does not have to store the set of already
computed cones. Saving memory is important if Grobner fans with millions
of cones need to be enumerated. In [6] a reverse search rule which works
even in the non-homogeneous case was presented. Unfortunately, it seems
that reverse search is not compatible with enumeration up to symmetry.
The algorithms for computing Grobner fans are described in detail in [6].
The tropical part of Gfan has implementations of the following
algorithms
• algorithms for handling polyhedral cones and turning them into
unique form,
• an algorithm for constructing the codimension one skeleton of the
normal fan of a polytope,
44 ANDERS N. JENSEN

• an algorithm for computing common refinements (intersections) of

tropical varieties,
• an algorithm for computing tropical bases of curves, and
• an algorithm for traversing a connected component of a tropical
variety.
Again, the basic polyhedral computations are handled by cddlib. Com-
puting refinements of tropical hypersurfaces is equivalent to computing the
mixed faces of a Minkowski sum. Computation of mixed faces is a well
studied problem as it can be used for mixed volume computations and
for the first step in the polyhedral homotopy method for solving polyno-
mial system numerically. In these cases one is usually only interested in
the mixed facets and has the option of applying a generic Ufting to make
things generic. However, for Gfan the exact combinatorics are needed. This
means that Gfan computes mixed faces of every dimension.
The algorithm for computing tropical bases of curves was described
in [3, Algorithm 5]. It relies on the intersection of tropical hypersurfaces
and a Grobner basis computation with respect to an order that depends
on the polyhedral computations. This Grobner basis step can be difficult
and thereby it contrasts the easy Grobner basis computations needed in
the traversal of the Grobner fan.
The global traversal is an exhaustive search where each local step
amounts to a tropical curve computation; see [3, Algorithm 7]. Here the
enumeration can again be performed up to symmetry. We should mention
that in Gfan symmetry is never exploited in Grobner basis computations
but only in the global enumeration of fans.
5. A few benchmarks. While Grobner basis computations are dou-
bly exponential in the worst case the Buchberger step of the local Grobner
basis change procedure is relatively easy in practice due to the homogene-
ity and other properties of the ideals in question. However, sometimes the
Grobner fan is simply too big to be computed or the conversion steps take
too much time. Here is a few examples of what can be computed.
In Figure 2 the 360 maximal cones were computed in 58 seconds using
5 megabytes of memory. It is easy to find four variable ideals where the
Grobner'fan is too big to be traversed or the local steps are too time con-
suming. For some classes of ideals the Grobner fan enumeration algorithm
can handle much larger examples.
EXAMPLE 8. Let I be the ideal generated by the 3x3 minors of the 4^4
matrix

( Xii XX2 X\z Xu \

X21 a;22 a;23 2:24
3:31 2:32 X33 0:34
\ a;4i a;42 2:43 xu )

in the polynomial ring of 16 variables.

 COMPUTING GROBNER FANS AND TROPICAL VARIETIES IN GFAN 45

The Grobner fan consists of 163032 full-dimensional cones. The group

S4 X 5^4 of order 576 acts on the polynomial ring by permuting rows and
columns of the matrix. Under this action I is fixed as an ideal. The 163032
marked reduced Grobner bases can be computed up to symmetry in 7 min-
utes using 9 megabytes of memory. There are only 289 orbits to consider.
Without exploiting symmetry the computation takes I4 hours.
It is not only the Grobner fan computation which has bad complexity.
In [22] several decision problems related to tropical varieties are shown
to be NP-haxd. The tropical variety T(7) in the example above is a 12-
dimensional sub-complex of the Grobner fan with 936 maximal cones. Its
f-vector is (1,50,360,1128,1680,936), meaning that there is one cone of
dimension 7 which is the homogeneity space. Traversing the maximal cones
of T{I) while exploiting symmetry takes 2 minutes.
EXAMPLE 9. Consider the 20 3x3 minors of a 3x6 matrix
xn x\2 a;i3 xi4 iis xie
a;21 X22 X23, X24 2:25 3:26
X31 X32 X33 X34 X35 X36

The relations on these generate the Grassmann-Pliicker ideal Is^ which

is an ideal in a polynomial ring urith 20 variables. Its tropical variety is
pure of dimension 10 and has a 6-dimensional homogeneity space. It con-
sists of 1035 maximal cones which were computed in 3-4 hours exploiting
symmetry. The cones come in 102 orbits.
6. Interfacing other programs and future plans. Interfacing dif-
ferent pieces of mathematical software is a non-trivial task. For Gfan one
difficulty is that on one hand it is a software system for commutative al-
gebra and on the other hand a software system for polyhedral geometry.
Thus no single interface suffices.
Thanks to the developers of Macaulay 2 [9] and SAGE [18] it is now
possible to invoke some basic Gfan functions from these programs. For
SAGE one additional advantage of this is that Gfan is part of the SAGE
distribution and therefore does not need a separate installation. Due to the
command line interface of Gfan it is possible, but cumbersome, to invoke it
from other mathematical software such as Singular [10] through a system()
call. This approach was chosen in the Singular implementation of the main
algorithm in [12].
For the polyhedral part the upcoming Gfan version 0.3 will output the
computed tropical varieties in a PoljTnake [7] compatible format. Since
Polymake does not handle polyhedral fans in its current version this has
limited use at the moment. However, it is our hope that the Polymake
format will serve as a standard format to be used by other tropical variety
software such as Trim [20].
Internally Gfan is linked to GMP [8] and cddlib [5] for exact arithmetic
and polyhedral computations. As pointed out in the discussion about per-
formance for tropical varieties the Grobner basis computation is one weak
46 ANDERS N. JENSEN

point of Gfan. One solution is to link Gfan to a commutative algebra

software library. CoCoALib [21] is currently under consideration.

REFERENCES

[1] D. AVIS AND K . FUKUDA, A basis enumeration algorithm for convex hulls and
vertex enumeration of arrangements and polyhedra., Discrete Computational
Geometry, 8 (1992), pp. 295-313.
[2] R. BlERi AND J. GROVES, The geometry of the set of characters induced by
valuations, J. reine und angewajidte Mathematik, 3 4 7 (1984), pp. 168-195.
[3] T . BOGART, A. JENSEN, R . THOMAS, D . S P E Y E R , AND B . STURMFELS, Computing
tropical varieties., J. Symb. Comput., 4 2 (2007), pp. 54-73.
[4] S. COLLART, M. KALKBRENER, AND D. MALL, Converting bases with the Grobner
walk., J. Symb. Comput., 2 4 (1997), pp. 465-469.
[5] K. FUKUDA, cddlib reference manual, cddlib Version 094b, Swiss Federal Institute
of Technology, Lausanne and Zurich, Switzerland, 2005. ( h t t p : / / w w w . i f o r .
math. e t h z . c h / ~ f ukuda/cdd_home/cdd. html.)
[6] K. FUKUDA, A. JENSEN, AND R . THOMAS, Computing Grobner fans., Mathematics
of Computation, 76 (2007), pp. 2189-2212.
[7] E. GAWRILOW AND M . JOSWIG, polymake: a framework for analyzing convex
polytopes, in Polytopes — Combinatorics and Computation, G. Kalai and
G.M. Ziegler, eds., Birkhauser, 2000, pp. 43-74.
[8] T. GRANLUND AND E T AL., GNU multiple precision arithmetic library 4-l-^,
December 2002. (http://swox.com/gmp/.)
[9] D.R. GRAYSON AND M . E . STILLMAN, Macaulay 2, a software system for research
in algebraic geometry, 2007. (http://www.math.uiuc.edu/Macaulay2/.)
[10] G.-M. GREUEL, G . P P I S T E R , AND H . SCHONEMANN, SINGULAR 2.0.5, A Computer
Algebra System for Polynomial Computations, Centre for Computer Algebra,
University of Kaiserslautern, 2CK)4. ( h t t p : / / w w w . s i n g u l a r . u n i - k l . d e . )
[11] B. HUBER AND R . R . THOMAS, Computing Grobner fans of toric ideals., Experi-
mental Mathematics, 9 (2000), pp. 321-331.
[12] A. JENSEN, H . MARKWIG, AND T . MARKWIG, An algorithm for lifting points in a
tropical variety, 2007. Preprint, math.AG/0705.2441.
[13] A.N. JENSEN, Gfan, a software system for Grobner fans, 2006. (http://www.
math.tu-berlin.de/~jensen/software/gfan/gfan.html.)
[14] A.N. JENSEN, A presentation of the Gfan software., in Mathematical Software -
ICMS 2006, Iglesias and Takayama, eds., Springer, 2006, pp. 222-224.
[15] T . MORA AND L . ROBBIANO, The Grobner fan of an ideal, J. Symb. Comput.,
6 (1988), pp. 183-208.
[16] J. RAMBAU, TOPCOM: Triangulations of point configurations and oriented
matroids, ZIB report, 02-17 (2002).
[17] D. SPEYER AND B . STURMFELS, The tropical Grassmannian, Adv. Geom., 4 (2004),
pp. 389-411.
[18] W. STEIN, Sage: Software for algebra and geometry experimentation., 2007.
(http://www.sagemath.org/,http://sage.scipy.org/.)
[19] B . STURMFELS, Grobner bases and Convex Polytopes, Vol. 8 of University Lecture
Series, American Mathematical Society, 1996.
[20] B . STURMFELS AND J. Yu, TYopical implicitization and mixed fiber polytopes, IMA
Volume 148: Software for Algebraic Geometry; Editors: Michael E. Stillman,
Nobuki Takayama, and Jan Verschelde; Publisher: Springer Science-t-Business
Media, 2008.
[21] T H E C O C O A TEAM, The CoCoA Project, 2007. ( h t t p : / / c o c o a . d i m a . u n i g e . i t / . )
[22] T . THEOBALD, On the frontiers of polynomial computations in tropical geometry,
J. Symbolic Comput., 4 1 (2006), pp. 1360-1375.
 O N A C O N J E C T U R E FOR T H E D I M E N S I O N OF
T H E SPACE OF T H E MULTIPLE ZETA VALUES
MASANOBU KANEKO*, MASAYUKI NORQt, AND KEN'ICHI TSURUMAKI*

A b s t r a c t . Since Euler, values of various zeta functions have long attracted a lot
of mathematicians. In computer algebra community, Apery's proof of the irrationality
of C(3) is well known. In this paper, we are concerned with the "multiple zeta value
(MZV)". More than fifteen years ago, D. Zagier gave a conjecture on MZVs based on
numerical computations on PARI. Since then there have been various derived conjectures
and two kinds of efforts for attacking them: one is a mathematical proof and another
one is a computational experiment to get more confidence to verify a conjecture. We
have checked one of these conjectures up to weight k = 20, which will be explained later,
with Risa/Asir function for non-commutative polynomials and special parallel programs
of linear algebra designed for this purpose.

K e y words. Multiple zeta value, double shuffle relation, symbolic computation,

parallel computation.

A M S ( M O S ) subject classifications. Primary 14G10, 11M06, 11Y99, 68W30.

1. Introduction. The multiple zeta value (MZV) is a real number

defined by the convergent series

C(k) = C(fci,fc2,...,M= E -^r^ fc;:' (i-i)

where k = (fci, ^ 2 , . . . , fcn) is an index set of positive integers with fci > 1
(which ensures the convergence). In recent years, the MZVs have been
appeared in various areas of mathematics and physics and aroused stim-
ulating interest among researchers. In particular, it has become apparent
that the structures of (linear or algebraic) relations over the rationals Q
among MZVs reflect properties or structures of various, seemingly unre-
lated mathematical objects. We refer the readers to Zagier's pioneer work
[13] and [2], [12], [8] together with the references therein for more on the
subjeet. In the present paper, we discuss experiments concerning a certain
conjecture on the linear relations among MZVs.
For each integer A; > 2, let -Zfe be the Q-vector space spanned by the
MZVs C(^i)• • •)^n) whose weight == ki + ••• + kn is equal to k. In [13],
Don Zagier gave a remarkable conjectural formula for dimq Zk'-

dimq Zk = dk,

' G r a d u a t e School of Mathematics, Kyushu University, 33, Pukuoka 812-8581, Japan.

^Department of Mathematics, Graduate School of Science, Kobe University, 1-1,
Rokkodai, Nada-ku, Kobe 657-8501, Japan.
^•Oracle Corporation, Japan.
47
48 MASANOBU KANEKO, MASAYUKI NORO, AND KEN'ICHI TSURUMAKI

where the number dk is determined by the Fibonacci-like recurrence

d2 = d3 = di = 1, dk = dk-2 + dk-3 {k > 5).

The total number of index sets of weight k is 2*^^^, which is much bigger
than dk « 0.4115 • • • x (1.3247 •••)''• For instance, 2^0-2 = 262144 whereas
^20 = 114. Hence we expect many linear relations among MZVs of given
weight (note that any relations known so far are the relations among MZVs
of same weight). One of the recent major progress in the theory is that
Goncharov [5] and Terasoma [11] independently proved that the number
dk gives an upper bound of dimg Zk'.
THEOREM 1.1 ([5],[11]). The inequality dimq Zk < dk holds for all k.
However, their proofs (both relying on the theory of mixed Tate motif)
do not give any explicit set of hnear relations to reduce the number of gen-
erators to the upper bound, and the question as to what sorts of relations
are needed for that is still unanswered. Concerning to this question, there
is a conjecture in which we are mainly interested:
CONJECTURE 1.1 ([7]). The extended double shuffle relations suffice
to give all linear relations among MZVs.
A stronger version of this conjecture was proposed by H.N. Minh,
M. Petitot et al. in [9] and they verified it (in the sense that their proposed
relations suffice to reduce dimQ-Zjt to dk) up to weight 16 (private com-
munication). Also, Espie, Novelli and Racinet [3] verified the conjecture
up to weight 19 (but modulo powers of TT^ at even weights), in a different
context of certain Lie algebra closely related to the "Drinfel'd associator".
Our main objective is to verify (a still stronger version of) this conjecture
up to weight k = 20. In the next section we explain what this conjecture
exactly means and introduce an algebraic setup to study the conjecture.
Using this setup, we can implement tools for generating relations among
MZVs systematically. This will be carried out in Section 3. In Section 4,
we will verify the conjectiure up to fc = 20 by using the algebraic tools and
special parallel programs of linear algebra. This work follows an experi-
mental computation by Minh and Petitot. We also give a new conjecture
in Section 4, which is a refined version of Conjecture 1.1.
2. The algebraic formulation. The MZV can be given not only as
a sum (1.1) but also as an integral

C(A;i,fc2,...,fc„) = /••'/ ^i{ti)i^2it2)---i^k{tk), (2.1)

l>tl>t2>->tfc>0

where k = ki + k^ + • • • + kn is the weight and uJi{t) = dt/{l — t) if

i € {ki,ki + ^2,.. .,ki + k2 + ••• + kn} and Wj(i) = dt/t otherwise. From
each of these representations one finds that the product of two MZVs is a
Z-linear combination of MZVs. The first example is
 THE DIMENSION OF THE SPACE OF THE MULTIPLE ZETA VALUES 49

C(2)2 =
\m>0 / \n>0 / m,n>0

E + E -- E
\TM>n>0 n>7n>0
^

m=n>Oj
2C(2,2) + C(4)
and

dtidt2 ff dt[dt'2
C(2)2 =
/ / tl{l-t2) 11 w^,
dtidt2dt'idt'2
JJJj fl(l-t2K(l-*2)
l>ti>t2>0
i>'2 =
dsids^ds^dsi
4
JjJj SlS2(l - S3)(l - S4)
1>S1>S2>S3>S4>0
S3>S4>0
f fff dsids2ds3ds4
+2
Sl(l -52)53(1 -Si)
1>S1>S2>S3>S4>0
= 4C(3,1) + 2C(2,2),
and this gives C(4) = 4^(3,1). The point here is that the two expressions
obtained are always different, and thus their equaUty gives a collection of
linear relations among MZVs which we call the finite double shuffle rela-
tions (FDS). Moreover, one can extend the finite double shufHe relations
by taking divergent sums and integrals into account together with a certain
regularization procedure. We call these generalized relations the extended
double shuffle relations (EDS), and the conjecture is that the EDS suffices
to give all linear relations among MZVs (Conjecture 1.1).
The two multiplication rules mentioned above are described in a purely
algebraic manner, as given in Hoffman [6]. Let Sj = Q{x,y) be the non-
commutative polynomial algebra over the rationals in two indeterminates
X and y, and Sj^ and f)° its subalgebras Q+S)y and Q + xSjy, respectively.
Let Z : i j " —+ R be the Q-linear map ("evaluation map") which assigns to
each word (monomial) U1U2 •• -Uk {ui E {x,y}) in Sj^ the multiple integral

W«l itl)^U2 (^2) •••'^Uk (*fc)

l>tl>t2>-••>tfc>0

where ujx{t) = dt/t, ujy(t) = dt/{l — t). We set Z{1) = 1. Since the word
uiU2-•-Wfe isini5°, we always have a;„i(i) = dt/t and Uuk{t) = dt/i^—t),
so the integral converges. By the integral representation (2.1), we have
Z{x^'-^yx''-'-'^y •x^"-''y) = C,{kx,k2,...,kn).
50 MASANOBU KANEKO, MASAYUKI NORO, AND KEN'ICHI TSURUMAKI

The weight A; = fci + ^2 -I 1-fc„of ^(^1,^2) • • •, ^n) is the total degree of

the corresponding monomial x^^~^yx''''^^y • • • x'^'^~^y.
Let Zk '•= x''~^y, which corresponds under Z to the Riemann zeta
value C(fc)- Then ^^ is freely generated by Zk (fc = 1,2,3,...). We define
the harmonic product * on Sj^ inductively by

l*W = W*l=W, ZkWi*ZlW2 = Zkiwi* Z1W2) + Zl{ZkWi*W2) + Zk+l{wi*W2),

for all k, I > 1 and any words w, wi, W2 S i 5 \ and then extending by
Q-bilinearity. Equipped with this product, f)^ becomes a commutative
algebra ([6]) and i3° a subalgebra. The first multiplication law of MZVs
asserts that the evaluation map Z : i3° —> R is an algebra homomorphism
with respect to the multiphcation *, i.e.,

Z{WI*W2) = Z{WI)Z{W2) {wi,W2&^°). (2.2)

For instance, the harmonic product Zk*zi = ZkZi4-ziZk + Zk+i corresponds
to the identity C.{k)C,{l) = C(fc, 0 + C.(l, k) + C,{k + /)•
The other commutative product m, referred to as the shuffle product,
corresponding to the product of two integrals in (2.1), is defined on all of
Sj inductively by setting

Imw = wml = w, uwimvw2 — u{wi\avw2) + v(uw\m.W2).,

for any words ly, wi, UJ2 € -Q and u, v € {x, y}, and again extending by Q-
bilinearity. The character 'm' is the Cyrillic sha, standing for shuffle. This
product gives ^ the structure of a commutative Q-algebra ([10]) which we
denote by ^ui • Obviously the subspaces Sj^ and Sj° become subalgebras of
Sjui, denoted by Sjj^ and Sj^ respectively. By the standard shuffle product
identity of iterated integrals, the evaluation map Z is again an algebra
homomorphism for the multiplication in:

Z{wimw2) = Z{wi)Z{w2) {wi,W2 e i3°). (2.3)

By equating (2.2) and (2.3), we get the finite double shuffle relations
(FDS) of MZV:

Z{wimw2 — Wi * W'2) = 0 {wi,W2 € S)°)- (2.4)

These relations, however, do not sufiice to obtain all relations. For instance,
there are two MZVs of weight 3, C(3) and ^(2,1). As Euler already showed
in [4], these two values are actually equal, and therefore dimqZs = 1.
But the least weight of FDS is 4 so that we cannot deduce the relation
C(3) = C(2,1) by the FDS. In order to obtain more relations, we introduce
a "regularization" procedure which amounts to incorporate the divergent
MZVs into the picture. In the following we only give a minimum of what we
need to formulate the extended double shuffle relations, and the interested
readers are invited to consult the paper [7].
 THE DIMENSION OF THE SPACE OF THE MULTIPLE ZETA VALUES 51

It is known (cf. [10]) that the commutative algebra i^ii is isomorphic

to the polynomial algebra over ijfj, in y:

^i-^mb)- (2.5)
Let regm ("regularization map") be the map from S)^ to S)'^ obtained by
taking the "constant term" with respect to y under the isomorphism (2.5):

regu Sji^w = ^WiUiy"'' >-^woeSji

i=0
Note that the map reg^, is the identity if restricted to the subspace S)^.
We then have the following theorem. For a proof, we refer the reader to [7].
THEOREM 2.1 (extended double shuiSe relations (EDS), [7]). For any
Wi € Sj^ and WQ € Sj'^, we have

Z(reg„j {wiuiwo -wi* WQ)) = 0.

When wi £ Sj°, the above relation reduces to the finite double shuffle
relation and so the EDS contains the FDS. If we take wi =y asa. particular
case, it can be shown that the element ymwo — y*wo is always in S)° and
hence we obtain the relation (without the regularization)

Z{ymwo -y*wo) =0 {WQ £ Sj°). (2.6)

If we substitute x''^~^yx''^~^y • • • x^"~^y for WQ in this relation and expand

it by using the definitions of m and *, we readily obtain the relation known
as Hoffman's relation:
n
/ ^ Cv'^l' • • • > "-i—Ij "-i ^" Ij i^i+lt • • • ) •^n)
i=i ki-2 (2.7)
= ^ Y^Ciki,---,ki-i,ki-j,j + l,ki+i,...,kn).
l<i<n,fe(>2 j=0
3. I m p l e m e n t a t i o n of t h e ring of bivariate n o n - c o m m u t a t i v e
polynomials.
3.1. Prototjrping. The algebra Q(x,y) is a non-commutative poly-
nomial ring. There are several computer algebra systems supporting non-
commutative polynomial ring and we have tried some of them. For example
Mathematica can calculate non-commutative polynomials, however it did
not behave as we hoped and it was not efficient. Next we tried represent-
ing a non-commutative monomial as a list and wrote a short program to
implement it. But it was not efficient because a fist representation needs
many links by pointers and even a simple monomial operation requires a
considerable number of memory operations. Then we tried representing
a monomial by a character string. It may seem more naive than the list
52 MASANOBU KANEKO, MASAYUKI NORO, AND KEN'ICHI TSURUMAKI

representation, but in fact it can be efficiently realized because the product

of two monomials can be computed by the concatenation of two charac-
ter strings, and the comparison can be done by comparing the character
strings according to a specific ordering, say, the lexicographic ordering. By
using this method, we implemented the operations in Q{x,y) by the user
language of Risa/Asir. The program amounts to only 160 lines.

3.2. I m p l e m e n t a t i o n as a built-in d a t a t y p e . Based on the proto-

typing, we decided to implement Q(x, y) as a built-in data type in Risa/Asir
for our large scale experiments. In this implementation, a monomial is rep-
resented as a bit sequence. All the fundamental operations in Q{x,y),
including the shuffle and harmonic products are implemented as built-in
functions written in C. Preliminary experiments show that they are more
efficient than the prototype, nevertheless the speed-up is about a factor of
10 and the prototype by general facilities is proved to be efficient enough
for small experiments.
An element in Q{x,y) is converted from a QUOTE object, which is
also a built-in data type for expressing general non-commutative objects.
A QUOTE object holds a tree structure converted from an input expression.
It preserves the order of products and we can convert it to an element
in Q(a;,t/) by calling a built-in function qt_tojabp(). After computing
polynomials giving EDS relations of a fixed weight k, we convert them to
integer vectors via the monomial basis consisting of all weight k monomi-
als in ^*'. In order to make the conversion efficient, we provide a func-
tion to compute the index of a monomial in the monomial basis. Note
that the index can be computed easily if we apply the lexicographic or-
dering, http://www.math.kobe-u.ac.jp/OpenXM/Math/MZV/mzvjlsr. r r
is a simple Risa/Asir program to compute a set of generators of Zk and to
represent remaining MZVs of weight k as linear combinations of the gen-
erators. The function dsrjnatrixCk) returns a matrix constructed from
a specific subset of EDS relations of weight k, which will be explained in
Section 4.1. Each row of the matrix gives the coefficients of MZVs in an
EDS relation, where MZVs are indexed according to the lexicographic or-
dering. F I G . 1 is the output of dsr-matrix(7). For example, the first row
in F I G . 1 shows a relation

C(7) - C(6,1) - C(5,2) - C(4,3) - C(3,4) - C(2,5) = 0.

The function dsr_basis(k) returns a list [ [ M i , P i ] , . . . , [Mj,Pj],...]. In

this output Mi and P, are given as elements of Q{x,y). Each pair [M,, Pj]
means Z{Mi) — Z{Pi), which represents an MZV Z{Mi) as a linear com-
bination of the generators of Zk computed from the input matrix given
by dsr_matrix(k). The function nbp_to_mzv(F) converts a polynomial F
in Q(x, 2/) to a Unear combination of MZVs. In the output of this func-
tion, an MZV is given as a list [k\,k2, • • ] which represents C(^1J ^2, • • •)•
 THE DIMENSION OF THE SPACE OF THE MULTIPLE ZETA VALUES 53

/ 1 -1 -1 0 -1 0
0 1 1 -1 0 -1
D
0
0
0
-1
0
0
-1
0
0
0
0
0
0
0
0
0
0
0
0
-1
0
0
-1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0 "\
0
0 0 1 0 1 -1 -1 0 0 0 -1 0 0 0 0 0 0 0 -1 D 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 1 0 1 1 -1 0 0 0 -1 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 1 -1 -1 0 -1 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0 0 1 1 -1 0 -1 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 1 0 0 0 1 0 1 -1 -1 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1 0 0 0 1 0 1 1 -1 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 —1-1 0 -1 0 0 0 -1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 1 -1 0 -1 0 0 0 -1 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 1 -1 -1 0 0 0 -1 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 1 1 -1 0 0 0 -1 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 1 -1 -1 0 -1 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 1 1 -1 0 -1 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 1 0 1 -1 -1 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 1 0 0 0 1 0 1 1 -1
-I 10 4 0 3 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 -1 0 12 -1 4 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 -1 0 0 6 6 0 —1 4 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 -1 0 0 0 12 0 -1 0 4 -1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 -1 0 0 0 0 4 2 0 4 0 0 0 -1 6 2 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 -1 0 0 0 0 0 6 0 4 0 0 0 -1 0 6 -1 1 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 -1 0 0 0 0 0 0 4 6 0 0 0 -1 0 0 2 1 0 -1 4 0 0 0 0 0 0
0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 10 0 0 0 -1 0 0 0 3 0 -1 0 2 -I 1 0 0
-1 20 10 0 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 -1 0 18 0 9 3 0 -1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 -1 0 0 6 6 0 0 8 4 0 0 0 0 0 -1 6 2 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 -1 0 0 0 12 0 0 0 6 0 4 2 0 0 -1 0 2 0 2 1 0 -1 1 1 0 0 0 0 0
0 -1 0 12 0 6 3 0 0 4 2 0 2 0 0 0 -1 1 1 0 1 0 0 0 0 0 0 0 0 0 0 0
0 0 -1 -2 0 0 0 18 0 -1 -1 7 0 2 0 0 0 -1 -1 1 0 1 0 0 0 0 0 0 0 0 0 0
0 0 0 0 -1 — 1-1 0 0 0 0 6 0 8 6 0 0 -1 -1 6 -2 2 1 0 0 0 0 0 0 0 0 0
*^ 0 0 0 0 0 -1 -1 -3 0 0 0 0 0 0 0 20 0 0 -1 -3 0 -2 -2 4 0 -1 -1 1 0 0 0 0 /

FIG. 1. The output o/dsrjnatrix(7).

The following example shows that Ze is generated by

{Zixyyyxy), Z{xyyyyy)} = {C(2,1,1,2), C(2,1,1,1,1)}.
For example, the first element of the output means
1 13
C(2,1,2,1) = - - C ( 2 , 1 , 1 , 2 ) + - C ( 2 , 1 , 1 , 1 , 1 ) .
Note that the number of generators coincides with d^ = 2.
[0] l o a d ( " . / m z v . d s r . r r " ) $
[7] m a p ( p r i n t , d s r _ b a s i s ( 6 ) ) $
[(I)*xyyxyy,(-l/2)*x3ryyxy+(13/24)*xyy3ryy]
[(l)*xyyxxy,(-l)*xyyyxy+(61/48)*xyyyyy]
[(1)*xyxyyy,(-1)*xyyyxy+(3/4)*xyyyyy]
[(l)*xyxyxy, (3/16)*x3ryyyy]
[(l)*xyxxyy,(3/2)*xyyyxy+(-11/12)*xyyyyy]
[(1)*xyxxxy,(1)*xyyyxy3
[(1)*xxyyyy,(1/2)*xyyyxy+(-7/24)*xyyyyy]
[(1)*xxyyxy,(3/2)*xyyyxy+(-11/12)*xyyyyy]
[(1)*xxyxyy,(-3)*xyyyxy+(97/48)*xyyyyy]
[(l)*xxyxxy,(-1/2)*xyyyxy+(13/24)•xyyyyy]
[(l)*xxxyyy,(l)*xyyyxy+(-31/48)*xyyyyy]
C(1)*xxxyxy,(-1)*xyyyxy+(3/4)*xyyyyy]
[(1)*xxxxyy,(1/2)*xyyyxy+(-7/24)*xyyyyy]
[(1)*xxxxxy,(1)*xy3^yy]
54 MASANOBU KANEKO, MASAYUKI NORO, AND KEN'ICHI TSURUMAKI

4. E x p e r i m e n t a l results and a new conjecture. In this section,

we numerically verify Conjecture 1.1 up to A; = 20, which seems to be
a world record. Exactly speaking we show that the set of EDS relations
reduces the upper bound of dimq 2 ^ to dk- That is, the verification is to
check dimq DS'^''^ > 2*^"^ — dk, where DS^''^ denotes the Q-vector space
spanned by all EDS relations of weight k. There are several conjectures on
the set of relations sufficient for reducing the dimension to dk- Minh, Jacob,
Oussous and Petitot [9] conjecture that Hoffman's relations and the FDS
suffice and they checked it up to fc = 16. We shall check (a stronger version
of) their conjecture up to fc = 20. We choose the weight fc = 20 as our final
target because the verification of the case is hard with respect to both the
time and space complexity but it is expected that it is executable on an
ordinary computing environment, if we apply well-known techniques such
as elimination by sparse rows to reduce the size of a matrix, computation
over a finite field and parallel Gaussian elimination. Detailed explanation
of these techniques will be given in Section 4.2, 4.3 and 4.4 respectively.

4.1. Generation of the double shuflHe relations. We denote the

sets of polynomials giving all FDS and Hoffman's relations of weight k by
Lfc/2J
FC^) and F'*^) respectively F^*^) is given by F^''^ = [J FI''\ where

F^''^ = {wimw2 -•wi*w2\wie Dyi°,w2 G scrt2_J

and 3H° denotes the set of all monomials of weight i in S)^. Denoting the
cardinality of a set S by l^l, \m^\ = 2'-^ implies \FI''^\ = 2'="'* for i < k/2.
If k is even, |F^J^| = 2-/^-\2^J^-'^+i) ^ ^^^ |^(fc)| ^ ^k-a ^hus we have

| F « | + | ^ « | > (L^J - 2)2'=-^ + 2'=-^. (4.1)

In particular we have |F(*')|-|-|F(*')| > 2*^~^ for A; > 8. The verification of the
conjecture is reduced to the rank computation of a matrix M^*^) constructed
firom the coefficients of the relations. The inequality (4.1) means that the
number of rows of M^''^ is greater than that of columns and their ratio
increases if A; becomes large. Our purpose is to show rank(M(*^) > 2*'""^—d/t
and it is sufficient to show this inequality for a sub-matrix of M^''\ Our
experiments for small k indicate that

suffices to attain the lower bound of the rank and we are led to a new
conjecture:
CONJECTURE 4.1. R^''^ suffices to reduce the upper bound of dimQ Zk
to dk.
 T H E DIMENSION O F T H E SPACE O F T H E MULTIPLE ZETA VALUES 55

li k >7, \R^''^\ is equal to 2*^"^ and the matrix constructed from jR^*^^
is a square matrix. As dk is much smaller than 2'^"^, R^''^ is practically
optimal for our experimental verification. The FDS relations are generated
by shuffle and harmonic products implemented in Risa/Asir. Hoffman's
relations can be generated either by (2.6) or by the explicit representation
(2.7). F I G . 1 actually shows the matrix constructed from i?'^^. Its upper-
half part comes from Hoffman's relations and the lower-half part comes
from the FDS relations. The matrix data for fc < 20 are available from
http://www.math.kobe-u.ac.jp/OpenXM/Math/MZV/matdata. For each
k, E^''\ F^''^ and F^ are converted to matrices and stored as files mhA;,
mf dsfc_2 and mf dsA;_3 respectively. These files are written according to the
following format:

( r C) (/i ••• Ir) ( j l . i a i ) ••• ( j l . i i ' ^ i i ) ••• ( > , l « s + l ) ••• Ur,lr0.3+lr)

where (r, c) denotes the size of the matrix, li denotes the number of non-
zero elements in the i-th row and {ji,k,o,t) denotes a non-zero element at at
{hji,k)- That is, non-zero elements are stored in row-major order. In the
file, all numbers are four-byte integers and they are represented according
to the network bj^e order.
4.2. Preprocessing by Hoffman's relations. If we convert R^''^
itself to a dense matrix, then we need huge memory for a large k. Fortu-
nately the matrix contains a large number of sparse rows coming from E^''^
and we can apply preprocessing to eUminate many matrix entries with a
small cost. The coefficients of / € E^''^ are 1 or —1 and the number of
terms in / isfc— 1. Furthermore, under the lexicographic ordering we have
the following (cf. FiG. 1):
PROPOSITION 4.1. For any xxw e QH^ there uniquely exists f € £'('')
such that the leading term of f is xxw.
That is, the left-half of the 2*^"^ x 2*^""^ sub-matrix obtained from
E^'^^ is already upper triangular with unit diagonals. Thus we can easily
eliminate the left-half of the sub-matrix obtained from F2 and F^ ' by
using E'^*'. Note that this is a sparse elimination and it can be done
efficiently. After this operation, we have the lower-right sub-matrix which
is denoted by 5^*=). We set Nk = 2*^"^. S^''^ is an Nk x Nk matrix and
what we have to show is rank(S^*^^) > Nk — dk-

4.3. Rank computation over finite fields. The coefficients of the

polynomials in FJ and F3 ' are within one machine word for fc < 20.
However, the result of the whole Gaussian elimination will have larger en-
tries and it will be hard to execute the rank computation over the rationals.
For any prime p we have rank(5^'^^) > rank(5^*^^ modp). Therefore it is
sufficient for our purpose to show rank(5^*^^ modp) > Nk — dk for some
prime p. In our experiments we use a two-byte prime p = 31991 for com-
puting rank(5'*^' mod p). TABLE 1 shows the sizes of memory required for
56 MASANOBU KANEKO, MASAYUKI NORO, AND KEN'ICHI TSURUMAKI

TABLE 1
Required size of memory for storing S^'^^ mod p.

k 16 17 18 19 20
size 128MB 512MB 2GB 8GB 32GB

storing 5^*°^ mod p. If a machine is not equipped with sufficient amount

of memory, then a further preprocessing may be necessary. By examining
5^*^ for fc < 20, we find that there are A^fc/4 x Nk sparse sub-matrix whose
left-quarter part is upper triangular. We can apply a sparse elimination by
this sub-matrix, which results in a 3/4 • Nk x 3/4 • Nk matrix. For example,
32GB of memory is required to store 5^^°^ mod p. If the second prepro-
cessing is not apphed then it is practically hard to execute the computation
on a machine with just 32GB of memory, which is one of our computing
environments. If we apply the preprocessing, then the required size is re-
duced to 18GB and we can efficiently compute the rank on that machine.
We note that there are probabilistic Wiedemann-Krylov type algorithms to
compute the rank of a matrix over finite fields. In general these are more
efficient than Gaussian elimination in sparse cases and it is interesting to
compare their performances in our cases.

4.4. Parallel computation by MPI. Even if we apply the prepro-

cessing to reduce the size of the matrices, they are still huge and it is
necessary to apply parallel computation from a practical point of view. We
wrote two parallel programs by MPI to execute Gaussian elimination over
finite fields. The implemented method is a simplification of the one used in
ScaLAPACK [1]. That is, the whole matrix is decomposed according to 1-
or 2-dimensional cyclic distribution algorithm and a non-blocking Gaussian
elimination is executed in parallel.
http://www.math.kobe-u.ac.jp/OpenXM/Math/MZV/gauss.c is a C code
for computing the rank of a matrix over a small finite field by using MPI.
Each processor element reads the same input files containing fragments of
the whole matrix, and stores a subset of rows of the input matrix into its
local memory. At each step of eliminating a column, informations of the
rows in the local matrices are shared among all processor elements to de-
termine the pivot row. Then the selected pivot row is broadcasted to all
processor elements and the elimination is done locally in each processor
element. Both the algorithm and the implementation are not so optimized
but the performance is satisfactory. We used several computing environ-
ments: (1) a cluster of heterogeneous hnux PCs, (2) a cluster of large SMP
Sparc/Solaris machines, and (3) an SMP linux PC. The last one is an 8
CPU SMP machine with two quad-core Intel X5355/2.66GHz CPUs and
32GB of memory. TABLE 2 shows various statistics in the last environment
up to fc = 20. In the table, "Generation", "Preprocessing" and "Elimina-
tion" show the elapsed time for generating all relations, preprocessing by
 THE DIMENSION O F T H E SPACE O F T H E MULTIPLE ZETA VALUES 57

TABLE 2
Statistics of the rank (imputation in the environment (3).

k 16 17 18 19 20
dk 37 49 65 86 114
Nk 8192 16384 32768 65536 131072
rank(5''''^ mod p) 8155 16335 32703 65450 130958
Nk - rank(S'('=) mod p) 37 49 65 86 114
Generation (ICPU) 22sec 85sec 4.5min llmin 30min
Preprocessing (ICPU) 5sec 19sec l.Smin 9min 57min
Elimination (8CPU) 2min 13min 1.3hour 9hour 67hour
Total memory 72MB 288MB 1.2GB 4.6GB 18GB

the sparse elimination and Gaussian elimination respectively. "Total mem-

ory" shows the size to store the 3/4 • Nk x 3/4 • Nk matrix. The table shows
that rank(5'''^ mod 31991) = Nk — dk np to k — 20, thus the conjecture
is verified up to fc = 20. The table also tells us that the preprocessing is
almost negUgible compared with the final Gaussian elimination. We note
that we also tried the same computations in the second environment with
another parallel program and p = 16381, and that we obtained the same
results of the ranks.

Acknowledgements. Prof. Nobuki Takayama gave us sincere en-

couragements and valuable comments on this work. Discussion with Dr.
Naoyuki Shinohara was useful to implement an efficient representation of
a non-commutative monomial in Risa/Asir. Prof. Kinji Kimura and Prof.
Kazuhiro Yokoyama were interested in our work and provided several com-
puting environments for our experiments.

REFERENCES

[1] L.S. BLACKFORD et al., ScaLAPACK Users' Guide, SIAM (1997), h t t p : / /

www.netlib.org/scalapack/slug/.
[2] P. CARTIER, Fonctions polylogarithmes, nombres polyzetas et groupes pro-
unipotents, Seminaire Bourbaki, 53 erne annee, 2000-2001, 8 8 5 (2001).
[3] M ESPIE, J-C. NovELLi, AND G. RACINET, Formal computations about multi-
ple zeta values, in "From Combinatorics to Dynamical Systems" (Strasbourg,
2002), IRMA Lect. Math. Theor. Phys. 3, F. Fauvet and C. Mitschi (eds.), de
Gruyter, Berlin (2003), 1-16.
[4] L EULER, Meditationes circa singulare serierum genus, Novi Comm. Acad. Sci.
Petropol, 2 0 (1775), 140-186, reprinted in Opera Omnia ser. I, Vol. 15, B.G.
Teubner, Berlin (1927), 217-267.
[5] A.B. GONCHAROV, Periods and mixed motives, preprint (2002).
[6] M. HOFFMAN, The algebra of multiple harmonic series, J. Algebra, 194 (1997),
477-495.
K. IHARA, M . KANEKO AND D . ZAGIER, Derivation and double shuffle relations
for multiple zeta values, Compositio Math., 142-02 (2006), 307-338.
5 8 MASANOBU KANEKO, MASAYUKI NORO, AND KEN'ICHI TSURUMAKI

[8] M. KANEKO, Multiple zeta values (translation of Sugaku, 54 (2002), no. 4,

404-415), Sugaku Expositions, 18 (2005), no. 2, 221-232.
[9] H.N. MiNH, G. JACOB, N . E . OUSSOUS AND M . P E T I T O T , Aspects combinatoires
des polylogarithmes et des sommes d'Euler-Zagier, J. Electr. Sem. Lothar.
Combin., 4 3 (2000), Art. B43e, pp. 29.
[10] C. REUTENAUER, Free Lie Algebras, Oxford Science Publications, 1993.
[11] T .TERASOMA, Mixed Tate motives and multiple zeta values. Invent. Math., 149
(2002), 339-369.
[12] M. WALDSCHMIDT, Valeurs zeta multiples: une introduction, J. Theor. Nombres
Bordeaux, 12 (2000), 581-595.
[13] D. ZAGIER, Values of zeta functions and their applications, in ECM volume,
Progress in Math., 1 2 0 (1994), 497-512.
 DEMICS: A SOFTWARE PACKAGE FOR C O M P U T I N G
T H E M I X E D VOLUME VIA D Y N A M I C E N U M E R A T I O N
OF ALL M I X E D CELLS
TOMOHIKO MIZUTANI* AND AKIKO TAKEDA*t

A b s t r a c t . DEMiCs is a software package written in C + + for computing the mixed

volume of the Newton polytopes of a generaJ semi-mixed polynomial system through
dynamic enumeration of all mixed cells. The underlying mixed cells play an essential
role for computing all isolated zeros of a polynomial system by polyhedral homotopy
continuation method. A notable feature of DBMiCs is in the construction of a dy-
namic enumeration tree for finding all mixed cells. The dynamic enumeration method,
proposed by Mizutani, Takeda and Kojima for fully mixed polynomial systems, is ex-
tended to semi-mixed systems ajid incorporated in the package. Numerical results
show that DEMiCs significantly is faster than existing software packages for semi-mixed
polynomial systems with many distinct supports. The software package DEMiCs is
available at h t t p : / / H w w . i s . t i t e c h . a c . j p / ~ m i z u t a n 8 / D E M i C s / .

K e y words, mixed volume, mixed cell, polyhedral homotopy, polynomial system,

semi-mixed structure, dynamic enumeration.

A M S ( M O S ) subject classiflcations. Primary 68N30, 52A39, 90C05.

1. Introduction. The polyhedral homotopy continuation naethod,

proposed by Huber and Sturmfels [15] and Verschelde et al. [24], is a pow-
erful and rehable numerical method [7, 13, 14, 16, 17, 25] for computing
all isolated zeros of a polynomial system f{x) = {fi{x), /aCa;),..., /n(x))
in the variables x = (xi,X2,... ,a;„) e C". In order to build a family of
homotopy functions, this method needs all mixed cells in a fine mixed sub-
division of the Newton polytopes of a polynomial system f{x) (See [15]
and [17]). It uses the mixed cells to construct homotopy functions between
start systems, which sire polynomial systems whose zeros can be computed
easily, and target system f{x). The mixed volume, which can be obtained
from the volumes of all mixed cells, is also the total number of solutions for
the start systems. It is also an upper bound for the total number of isolated
zeros in (C \ {0})" of f{x), as guaranteed by Bernshtein's theorem [1].
T-he aim of this paper is to introduce the software package DEMiCs.
The package uses djmamic enumeration to compute all mixed cells and the
mixed volume of the support of a general semi-mbced polynomial system,
including fully mixed and unmixed systems as special cases. The dynamic
enumeration method was originally developed for a fully mixed system in
[19], and it is significantly faster than other software packages [8,10, 12,18,
25, 22] for enumerating all mixed cells for large-scale semi-mixed systems

'Department of Mathematical and Computing Sciences, Tokyo Institute of Technol-

ogy, 2-12-1 Oh-Okayama, Meguro-ku, Tokyo 152-8552, Japan ([email protected]).
t ([email protected]).

59
60 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

with many distinct support sets. It also opens the door to computing all
zeros of larger polynomial systems by polyhedral homotopy.
The mixed cells are enumerated by using a tree on which each node
is provided with a linear inequality system. An enumeration tree has the
following properties:

(i) A leaf node corresponds to a mixed cell if and only if the linear
inequaUty system attached to the leaf node is feasible.
(ii) Each node shares a common system with the child nodes, so that
if the node is infeasible, so are all of its descendant nodes.

There are various ways of constructing such enumeration trees. The static
enumeration method [9, 10, 18, 25] fixes the structure of a tree before
finding mixed cells. However, it is ideal that most of the child nodes are
infeasible and pruned when branching at each node is carried out. To
pursue this idea, the paper [19] proposed the dynamic enumeration method
for a fully mixed system. This method chooses a suitable child node set
among all the candidates generated from a parent node, so that the total
number of nodes of a tree is expected to be small. In this situation, it
is essential for computational efiiciency to utilize information from each
node to be generated during the construction of a tree. In this paper, we
explain how the dynamic enumeration method is extended to a semi-mixed
polynomial system. Recently, we noticed that Emiris and Canny [8] also
enumerate all mixed cells dynamically for fully mixed systems, and build
the dynamic enumeration tree implicitly for this purpose. Both methods
use a linear programming (LP) problem as a means to enumeration mixed
cells. The main differences between the two methods are as follows.

(i) Formulation for finding mixed cells; Our formulation is well suited
for utilizing rich information which can be obtained by solving the
LP problems at upper level of a tree.
(ii) Choice of a child node set at each node: Our method efficiently
chooses a suitable child node set in all the candidates generated
^firom a parent node, using information obtained from the LP prob-
lems to be solved during the execution of enumeration. Emiris and
Canny's method select the best child node set in all the candidates
without taking account of such information, and hence, it is more
computationally expensive.

In fact, the numerical results in [10, 18, 22] show that Emiris and Canny's
dynamic enumeration strategy has a speed disadvantage. In Subsection
3.3, we precisely describe how the two methods construct a dynamic enu-
meration tree.
There are several related software packages for computing the mixed
volume through enumeration of all mixed cells: H0M4PS [11], MixedVol
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 61

[10, 12], MVLP [8], PHCpack [25], PHoM [22], mvol [18], etc. H0M4PS,
PHCpack and PHoM, written in FORTRAN, Ada and C + + respectively,
are software packages for computing all isolated zeros of a polynomial sys-
tem by polyhedral homotopy. These packages each contain a module for
enumeration of mixed cells. In particular, PHCpack is the most popular
of these software packages. The C package MVLP uses a dynamic enu-
meration for finding all mixed cells. The C + + package Mixed Vol, which
employs the static enumeration method, specializes in the mixed volume
computation. MixedVol performs better than all other software packages
in terms of computational efficiency and memory usage, as reported in the
papers [10, 12]. A feature of the package is that all mixed cells can be gen-
erated efiiciently for a semi-mixed system by taking account of the special
structure of a semi-mixed support.
Numerical results show that DEMiCs can find all mixed cells much
faster than the other software packages [8, 10, 12, 18, 25, 22] not only
for fully mixed poljmomial systems but also for semi-mixed polynomial
systems. Although DEMiCs has a speed advantage over the other packages
for large-scale fully mixed systems and semi-mixed systems with many
distinct support sets, it does not always excel. Indeed, for unmixed systems
and semi-mixed systems with a few distinct supports, DEMiCs is slower
than MixedVol [12] because of its computation overhead associated with
the dynamic branching rule.
This paper is organized as follows. Section 2 and 3 describe the tech-
nical details of our method. Section 2 outlines the dynamic enumeration
method for a general semi-mixed polynomial system, and Section 3 ex-
plains how to check the feasibility of each node and how to construct an
enumeration tree when a polynomial system is a semi-mixed type. Section
4 describes the usage of DEMiCs. Section 5 compares the performance
of DEMiCs with those of the other software packages on artificial semi-
mixed polynomial systems and well-known large-scale benchmark systems.
Section 6 is devoted to concluding remarks.
2. Dynamic enumeration algorithm for a semii-mixed system.
2.1. Preliminciries. In this paper, we represent each component
polynomial fi{x) in a polynomial system f{x) = {fi{x), f2{x),..., fn{x))
in a; e C" as

fiix) = J2 ^(«) ^"'

using a nonempty finite subset Ai of Z" and nonzero Ci{a) € dor a G Ai.
Here Z" denotes the set of nonnegative integer vectors in R", R and C are
the sets of real and complex numbers, respectively, and
for a = ( a i , 0 2 . . .,a„) £ Z " . We call the set Ai the support of fi{x),
and A = {Ai,A2,- • -jAn) the support of f{x). The convex hull Vi :==
conv(.4i) is called the Newton polytope of / j . Define N = {1,2, . . . , n } .
62 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

The following describes the definition of the mixed volume of the n-tuple
Newton polytopes Vi in M". For all positive number Ai, A2,..., A„, it is
known that the n-dimensional volume of the Minkowski sum

AlPl + A27'2 + • • • + KVn =- {A1P1+A2P2 + • • • + KPn • Pi & Vi, i S AT}

is given by a homogeneous poljoiomial of degree n in Aj (i € N). The

mixed volume for A is defined to be the coeflacient of A1A2 • • • A„ in the
polynomial.
Some support sets in ^ = {Ai : i G N) oi f{x) may be equal to
each other. We suppose that the polynomial system has exactly m (< n)
distinct support sets <Si, ^ 2 , . . . , Sm among Ai,A2,-- • ,An such that

Si := Aj^ = Aj^ for every j i , J2 eh (i € M), (2.1)

where we define M = { 1 , 2 , . . . , m}. Obviously, the subset li of N satis-

fies UigM-^i = N and 7jj fl /ij = 0 for every ii,i2 € M. A polynomial
system with the support S = {Si : i e M) is called semi-mixed. The
system is called unmixed when m = 1, and fully mixed when m = n, and
it is called semi-mixed of type {ki,k2,---, km) if and only if Si occurs ex-
actly ki times in Ai,A2,--- ,An, i-e., #Ii = ki. In this paper, we treat a
semi-mixed polynomial system f{x) of type (^1,^2, • • •, km) with support
S = {Si,S2, • • • ,Sm), and assume that each support set Si consists of rj
elements.
For a support S — {Si : i € M) of a semi-mixed system described
above, we use the notation Qi — conv(<St) for the Newton polytope of each
polynomial. The mixed volume for S = {Si : i £ M) of a semi-mixed type
can be computed by finding all mixed cells in a fine m,ixed subdivision of
the Minkowski sum Q = Qi -h Q2 H 1- Qm- These are essentially piece
poljd;opes, called cells, in a polyhedral subdivision of Q. The reader may
want to refer to the definition of a fine mixed subdivision in [6, 15]. It is
known that each cell in a fine mixed subdivision can be represented as the
Minkowski sum of simplices Rj := conv(Ci) + conv(C2) -I- h conv(C4)
for C = (Cj, C l , . . . , C4) where each Cf is the subset of Si and its convex
hull conv(C^) is a simplex of dimension # C ^ — 1. In particular, when a
polynomial system is a semi-mixed system of type (^1,^2, • • •, km), we call
a cell Rj that is described as a Minkowski sum of each conv(C|), a mixed
ce/Zif dim(conv(C/)) = ki for every i G M. It is shown in [15, Theorem 2.4.]
that the mixed volume for <S is obtained from the volumes of all mixed cells
in a fine mixed subdivision of Q. This means that only distinct support sets
contribute the mixed volume computation. A semi-mixed system certainly
can be treated as a fully mixed type, and the mixed volume can be obtained
by summing the volumes of all mixed cells in a fine mixed subdivision of
V = Vi-\-V2-\ h Vn- However, the numerical results in [10] show that
the computational time for finding all mixed cells can be reduced if we focus
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 63

on the only distinct support sets for a semi-mixed system. Therefore, it is

important for computational efHciency to utilize a semi-mixed structure.
The papers [2, 15] describe how to construct a fine mixed subdivision
of Q by using a real-valued function w, : <S, —*• K. The function w, lifts
Si to

Let Q denote the Minkowski sum Q = Q1-I-Q2-I l-Qm for Qi = conv(5j).

For the subset Ci of Si, we will use the notation C = (Ci, C2, • • •, Cm)-
Suppose that the function Wj gives a random number so that the value
u!i(a) is sufficiently generic for every i £ M and every a £ Si. Then, the
projection R""''^ -^ R" of the set of lower facets of Q gives a fine mixed
subdivision of Q. In this paper, we call such a function Wj a generic lifting.
Li and Li [18] proposed an efficient algorithm for finding lower facets of
Q via an enumeration tree. Recently, for a fully mixed polynomial system,
the paper [19] improved their algorithm by replacing a static enumeration
tree of [9, 10, 18, 22] with a dynamic enumeration tree. In this paper, a
dynamic enumeration method is apphed to find all mixed cells in a fine
mixed subdivision for a semi-mixed system, including a fully mixed and
unmixed type.
2.2. Algorithm. We briefiy describe the dynamic enumeration algo-
rithm of [19] and apply it to a semi-mixed polynomial system. For every
L C M = { 1 , 2 , . . . , m}, we define

rtfT\ \ /-^ i n n n \ . ^i Q Si, jfCi = ki + 1 [i £ L) 1

0(L) = | C = ( C i , C 2 , . . . , C ™ ) . c^=%{j^L) /
n = \JLQM^{L).

The set Q represents the collection of all nodes in an enumeration tree. The
tree has a root node 0*" £ 0(0) := {0"*} and the leaf nodes f2(M) C 0 .
A node at the tth level corresponds to the element in U L C M , # L = ^ 0 ( L ) .
Let L{C) = {i G M : Ci 7^ 0} for any C £ Q,{L) (L C 'M). This
definition is used for extracting every index of nonempty sets d from C =
{Ci,C2,...,Cm) G 0(L), i.e. L{C) = L ioi C £ 0(L) {L C M). Each
node C = {d : i £ L) £ Q,{L) with L Q M has a linear inequality system
AC):

(_ {ai,a) <{a,a), va £ Si\Ci

where

a - . ^
64 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

(•, •) stands for the usual inner product in Euclidean space. Note that Si is
obtained by applying a generic lifting Ui to Si, and C, is the subset of Si.
Li and Li showed in [18] that any mixed cell in a fine mixed subdivision of
S = {Si : i G M) is in one-to-one correspondence to C = (Ci, C2,...,Cm)
with Ci C Si and # C i = ki + 1 for each i G M such that the linear
inequahty system X{C) is feasible. We say that C € fl is a feasible node
when J ( C ) is feasible. Let

fi* = {C e 0 ( M ) : C is feasible}.

Then we can easily see from (2.2) that Cl* consists of every mixed cell in a
fine mixed subdivision.
For C & Q and L C M, we use the notation CL for CL = (Ci : i €
L). Regarding the root node 0"* € J^ as a feasible node, we construct an
enumeration tree according to Algorithm 2.1 of [19]. Namely, for a node
C € Q{L) with the proper subset L C. M and t & M\ L{C) we generate
the child node set W{C,t) of C

W{C,t) = { C € fi(L(C) U {t}) : CL(C) = CL^C)} •

Starting from the root node 0"^ e Q, we choose t from M \ L{C) at each
node C e Q{L) with L C. M and create child nodes of C until # L = m — 1
based on the algorithm. Thus, fi* coincides with the set of the feasible
leaf nodes C G Q,{M). If we check the feasibility of all leaf nodes, all
mixed cells in a fine mixed subdivision can be obtained. Note that this
algorithm produces various types of trees depending on a choice of an index
t £ M\L{C) at each node C G Q(L) with L C M. A static enumeration
tree is constructed in the previous works [9, 10, 18, 22], which specify how
to choose an index t G M\ L{C) at each node C G Q{L) with L C M
before the building of a tree. In contrast to this, the paper [19] develops a
dynamic enumeration tree by choosing a suitable index t from M \ L{C)
at each node C G f2(X) with L C M.
We can enumerate feasible leaf nodes of a tree efficiently if the fol-
lowing property is taken into account. The feasible region of the linear
inequality system I{C) attached to a node C contains that of J ( C ) , which
corresponds to a child node C of C . That is, we can say that if a parent
node C is infeasible, then all child nodes C G W{C,t) {t e M\L{C)) are
infeasible. If a node is detected to be infeasible, we can prune a subtree
having the node as the root node because there are no mixed cells in the
subtree. Therefore, by replacing W{C, t) in Algorithm 2.1 of [19] with

W*{C,t) = { C G W{C,t) : C is feasible} C W{C,t),

every mixed cell can be found as the feasible leaf nodes in the tree. Taking
account of this property, we search for all nodes in fi* according to the
dynamic enumeration algorithm stated below. We will use the notation
At {I G {0} U M) for the set of feasible nodes.
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 65

Algorithm 1 Dynamic enumeration algorithm.

Input: A support S = (<Si,<S2, • • •,<5m)-

Output: All mixed cells in a fine mixed subdivision.
Ai <r-$ for all i e M.
Ao ^ 0"" and £ «- 0.

while £ < m do
for allC £Ai do
Choose t from M \ L{C) : (A)
Ai+i^Ai+x\jW*{C,t) :(B)

end for

end while

Once this algorithm terminates, A^ contains all nodes in Q.*, i.e, ev-
ery mixed cell can be stored in Am- We introduce this algorithm, which
constructs a tree in breadth-first order, for simplicity of description; it is
essentially the same as Algorithm 2.2 described in [19], which is in depth-
first order. Hence, we can obtain the same result from these two algorithms
though there is the diflFerence in the search order for feasible leaf nodes. The
software package DEMiCs employs the depth-first order, which is similar
to Algorithm 2.2 in [19], to save memory.
The following two issues have a major effect on the computational
efficiency of the dynamic enumeration algorithm for semi-mixed systems.

(i) How we choose an index t from M \ L{C) in (A),

(ii) How we construct W*{C,t) in (B).

As for (i), in the static enumeration method proposed in the previous

works [9, 10, 18, 22], we set up a permutation vr of M before starting the
algorithm, and choose the index t such as t = Tr{i+1) £ M\L{C). Hence,
a choice of an index t at (A) is determined by a permutation TT. Here
we employ a refinement of the dynamic enumeration method from [19].
Suppose that C is a feasible node in A^ with i < m. Then, we consider
a choice of an index t from M \ L{C) so that many child nodes of C are
expected to be infeasible. Ideally, we would like to choose the index t such
that the size of W*{C,t) is the smallest among t £ M \ L{C). Although
this strategy is employed by Emiris and Canny's method [8], it is costly
because we need to construct W*{C,t) for each t & M\ L(C). Therefore,
instead of W*{C,t), we consider another set which can be constructed
easily. In Subsection 3.3, we explain how to construct this set.
66 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

As for (ii), it may not be an easy task to find all feasible nodes in
W{C, t) because the size of W{C, t) is not small. So we embed the con-
struction process for W*{C,t) in a tree, and prune worthless subtrees in
order to reduce computational effort. In Subsection 3.1 we discuss the de-
tails of this procedure, and show the formulation of the feasibility check of
a node in Subsection 3.2.
3. Feasibility check for a semi-mixed system.
3.1. Ttee structure for construction of W*{C,t) in ( B ) . We
now explain a tree structure for finding all elements in W*{C,t) efficiently.
Suppose that an index t is chosen from M\L{C) for a feasible node C £ Ae
with ^ < m in (B) of Algorithm 1. Then, we would like to construct
W*{C,t) C W{C,t). For a nonnegative integer k, let

U = {Ui,U2,...,Um): UtCSt,#Ut =k
Ui = ${i^ L{C) U {t})

Note that r(^t + 1; C, t) = W{C, t).

For a feasible node C £ At with ^ < m in the dynamic enumeration
algorithm, we build a tree T for constructing W*{C, t). The tree structure
is outlined as follows. Let Kt = {'d,l,...,kt + 1\. The set

r:=Ufegif,r(A;;C,t)

serves as the collection of all nodes in the tree. The tree has C G
r ( 0 ; C , i ) = {C} as the root node, and U £ T{k;C,t) with #Ut = A;
as a node at the A;th level. Each node U G r(fc; C, t) has a linear inequality
system J{U) in a variable vector a G K""*"^. Note that each system I{U)
with U S T{kt + l;C,t) is identical to the hnear inequahty system X{C)
at a node C € W(C,t). Therefore, W*{C,t) can be obtained by checking
the feasibility of any node U G T(kt + l;C,t) which corresponds to each
leaf node at the {kt + l)th level.
We now describe more precisely the tree structure for building
W*{C,t). For Ui C St, we choose a function mt : 5t —» Z which
gives the maximum number i among the indices of Oj G Ut- Namely,
mt{Ui) = max{i e Z : Ui € Ut} for Ut C Sf Let T = {V,E) be a rooted
tree, which describes the relation among elements of a node set F. Recall
that St has rt elements. For a node U G r{k;C,t), we generate the child
node set

ZiU;C,t)
^ '
=[ lueTi#Ut
^
+' l;C,t):
' '
. ^* " ^* m?i'• ^j
for every I, mt{Ut)<t<rt
1
and construct T = iV,E) with V — Ufcei<r ^fe ^•iid E = UfceiCt ^k, based
on Algorithm 2. The root node of T, constructed by the algorithm, corre-
sponds to Vo = {C}, where C is one of the elements in A^ generated by
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 67

Algorithm 2 Construction of the tree T = iV,E).

Input A feasible node C £ Ag, and an index t £ M\L{C).

Output T = {V, E) with V = Ufcex, ^fc and E = Ufeg^e ^k-
Vo^C, JEJo ^ 0 and k ^ 0.
Vi*-9 and Ei^<D for all ieKt\ {0}.

while k < kt + 1 do
for all U eVk : {\\) do
if mt(l7t) < rt then

E f c + i ^ £ ; f c + i U { ( C / , i / ) ) € l 4 x Vk+i:UGZ{U;C. ,*)}
end if
end for
k^k + 1.
end while

.{tti} {02} {as} {a4}

/ \ / \ I
{01,02^ {01,03} (01,04} {02,03} {02,04} {03,04}
/ \ I I
{oi, 02, a3}{ai, 02, a4]{ai, 03,04} {02,03,04}

FIG. 1. Tree structure generated by Algorithm 2.

Algorithm 1. When this algorithm terminates, Vk stores every element in

T{k\ C,t) for all fc G Kt. Hence, W{C,t), which is generated by a feasible
C & Ai and an index t € M\L{C) in Algorithm 1, is equal to Vkt+i, which
has each leaf node at the {kt + l)th level of the tree T.
EXAMPLE 1. We consider the distinct support sets 81,82 {8i Q Z^) of
the semi-mixed system f{x) {x € C^) of type (1,2) such that | 5 i | = 3 and
\82\ = 4. We execute Algorithm 1 for the support 8 = {81,82)- Suppose
that the algorithm produces a nonempty set A\ when t — 1 is chosen in
(A) at the first iteration. Figure 1 shows the shape of the tree generated by
Algorithm 2 for input data C £ Ai. Here, each label of nodes represents
U2C820fU = {Ui,U2) e T{k; C,2){0<k< 3).
68 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

For a node U &Vk and child node U G V^+i of U, the feasible region
oi X{U) contains that oi X{U). Hence, if the node t / € 14 is infeasible,
every child node U £ Vk+i of U is infeasible. Therefore, the subtree with
root node U can be pruned, since it does not contain any feasible leaf
nodes at the {kt + l)th level. Accordingly, even if we replace Vk at (tj) of
Algorithm 2 by

VC = {UeVk:U is feasible},

we can find every feasible leaf node at the {kt + l)th level in V^*+i, and
thus obtain W"*(C,i)-
After building T = iV,E), we enumerate all feasible nodes at the
(A;t + l)th level in T and construct V^*+i- First, we initialize V^* as VQ' = Vo-
Next, we repeat the following procedure until k = kt- Suppose that V^ is
constructed. Then, we check the feasibility of every child node U oiU e V^
and store the feasible nodes in Vj^*^.i- As a result, W*{C,t) is obtained as
^k +1- Note that here we use Algorithm 2 having a breadth-first order for
the simplicity of description and that DEMiCs has a depth-first order to
save memory in constructing W*{C,t).
3.2. Formulation of the feasibility check. Suppose that a tree
T = {V,E) with V = \JkeKt ^k and E = Ufcei^:, ^fc is generated by Algo-
rithm 2 for a feasible node C € Ae and an index t e M \ L{C). Then,
we need to check the feasibility of each node f/ S Vfe in order to construct
V^. An LP problem can be formulated for the feasibility check of a node
C/ G Vjt. Let 7 denote a specific n-dimensional real vector, and furthermore
let 7 ^ = (7^,0). To determine whether a node C/ G 14 is feasible or not,
we solve the following problem in a variable vector a G R""*"^:

P ( t / ) : I maximize (7,0:) subject to I{U).

Let Oj be an element in Ui for any i G L{U). The dual problem in a

variable vector a; G R , where d = Yli^Liu)i'''i ~ 1)) is written as

minimize ^{x;U)
subject to ^{x; U) — 7 ,
D(C/):
—00 <Xb < +00 b eUi\ {tti}
Xb'>0 b'eSi\Ui (ieL{U)).

Here, the linear functions ^{x; U) and ^ ( x ; t/) in x G R'' are defined as
follows:

^{x;U)= ^ ^ {uji{a) -uji{ai)) Xa

ieL{U) aeSi\{ai}
and *(a;; U) = \] V] i^i ~ ^) ^a,
ieL(U) a€Si\{ai}
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 69

where

x = ixa: a £Si\ {ai}, i G L{U)) GM.'^.

Any real vector 7 € R" can be chosen. If 7 is fixed so that D(t/)

is feasible, the duality theorem holds for this primal-dual pair. P{U) is
feasible if and only if D{U) is bounded below, and P{U) is infeasible if and
only if D(l/) is unbounded. Hence, the feasibility of P{U) can be revealed
if the boundedness of D(C7) is detected. The following describes how we
set up 7. Recall that t / € Vjt is a node at the fcth level in T = {V,E)
generated by Algorithm 2, and C G Vo is the root node of T. We note that
the feasible region of D(C) is included in that of D(l/) for any C/ € 14.
Consider a right-hand vector 7 of D(t/) for any C/ s V^ as

7 = *(x;C)
using an arbitrary nonnegative vector x eM.'^. Then, I>{U) becomes feasi-
ble for each C/ e Vfe.
Furthermore, we can easily find an initial feasible solution for the dual
problem D(C7) for any t7 G V^ by using an optimal solution of D(C).
Recall that the root node C G Vo was revealed to be feasible. That is, we
solved D(C) and obtained an optimal solution x* G R** of D(C), where
d = I]i€L(C)(^' ~ !)• ^^^ *"y ^ ^ ^fc' *^^ vector

( P* ) G R ^ where d= Yl (^^ ' !)• (^-l)

becomes a feasible solution of D(f/). In addition, if D(l/) is bounded below

for U eVk with k < kt+1, an optimal solution of D(t/) is a feasible solution
of D{U) for any child node fj G 14^.1 of U since the feasible region of T>{U)
is included in that of D(t/). The simplex method is suitable for solving
these dual problems with the common structiu-e. Assume that a node
U e Vk {k < hi + 1) is feasible and generates the child nodes U G Vk+i
of U. The simplex method usually does not require many iterations for
solving D(C7) when we reuse an optimal solution of D(U) as an initial
solution. In terms of problem size, the dual problems are superior to the
primal ones as stated in [19]. Therefore, we employ the dual problems to
check the feasibility of a node and solve these problems using the simplex
method.
3.3. Choice of t € M \ L{C) in (A). In Algorithm 1, we want
to detect an index t from M \ L{C) for a node C £ At with ^ < m, so
that a large portion of the child nodes are infeasible. The best way to
achieve this is to find an index t among t £ M\ L{C) such that the size
of W*{Cjt) attains the minimum size. Emiris and Canny's method [8]
employs this strategy and selects the best support set from all candidates.
70 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

However, this approach might be unrealistic because it requires a large

computation to construct W*{C,t) for every t G M \ L{C). In fact, the
numerical results in [10, 18, 22] show that MVLP, in which Emiris and
Canny's method is implemented, is slower than MixedVol [10], PHoM [22]
and mvol [18]. Remember that each element in W{C,t) is represented as
the leaf nodes C7 G Vfc^ of a tree produced by Algorithm 2, and the leaf
nodes are infeasible if the ancestor node of these leaf nodes is infeasible.
To restrict the number of feasibility checks, we therefore focus on finding
feasible nodes in Vi, which is the set of every node at the first level of the
tree and is identical to r ( l ; C,t). Obviously, the set

W^{C,t) = {C e r ( l ; C , t ) : C is feasible}

satisfies

W*{C,t) C W^{C,t) C W{C,t).

Using a feasible solution Xmu of D(t/) {U e r ( l ; C , t)) which can be

easily obtained from an optimal solution of D(C), our method estimates
the feasibility of each node in r ( l ; C , t ) , and constructs Wi{C,t,Xinit)
satisfying

w^iCt) c w;:{c,t,xinit) c w{c,t).

The definition of Wi{C,t,Xinit) is described as follows. To find fea-
sible nodes U in r ( l ; C , t), we deal with the dual problem D(C/) instead
of P(C/), and check the boundedness of this problem. As stated in the
previous subsection, we can obtain a feasible solution Xmu of D([/) for
any U G r ( l ; C , t ) by (3.1) with the use of an optimal solution a;* of
D(C). Prom this initial solution, we start the pivoting process of the sim-
plex method to check the boundedness of D(C/). The simplex method easily
solves D(£7) with Xinu as the initial solution, because the structure of these
problems D(C) and D(t/) are similar to each other. Indeed, in numerical
experiments, we see that the simplex method requires only a few iterations
for D(t/-) starting fi-om Xmit- Accordingly, we can expect that a feasible
solution Xinit of D(t/) has an unbounded direction when this problem is
unbounded. Thereby, we test whether the feasible solution Xmu of D(C/)
has an unbounded direction instead of solving this problem by the simplex
method. At (A) of Algorithm 1, we construct

Ti;-*/.^ i \ f/s n/1 /-. JS there is no unbounded direction 1

W, {C,t,Xinit) = ^Ce nUCt) : ^jj^^^^y^g f^^^ ^.^.^ i^ j5(g) I

We need to explain the phrase "unbounded direction emanating from

Xinit in D(C)" used in the definition of Wi{C,t,Xinit)- Because a free
variable Xb satisfying —oo < Xb < -l-oo on D(t7) can be represented by
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 71

FIG. 2. Direction vectors emanating from v.

Xb = xj, — x'^ using nonnegative variables x^, x'^, for simplicity, we consider
the standard LP problem
minimize (c, x)
subject to Ax = b, x >0
where a coefficient matrix A G R"**^", cost vector c G K" and constant
vector 6 G R"* are given, and a; G R" is a variable vector. Suppose that
this problem is feasible, i.e. the feasible region, which forms a polytope,
is nonempty. Then, for a vertex v of the polytope, the simplex method
computes an adjacent vertex v = v + 9d {6 > 0) oi v using a direction
vector d G R", which emanates from v. Fig. 2 illustrates the direction
vectors incident to u. A direction vector d emanating from v is said to
be unbounded if we can increase the value of 6 up to +oo while satisfying
the following: (i) v = v + 6d satisfies all constraints, and (ii) the cost
(c, v) decreases from {c, v). If the LP problem is unbounded, it has an
unbounded direction emanating from some vertex v. In (A) of Algorithm 1,
our dynamic enumeration method chooses an index t such that the size of
Wi{C,t,Xinit) is the smallest among f G M \ L{C). Although we need
a practical procedure to test whether Xinu of D(t7) has an unbounded
direction, a detailed description can be found in [19, Section 3.2].
In (A) of Algorithm 1, our method uses W{{C, t, Xmu) whereas Emiris
and Canny's method uses W*{C, t). There may be a diiference between the
sizes oiW*{C,t,Xinit) and W*{C,t) for some t G M\L{C). However, the
numerical results in Section 5 show that our method sufficiently reduces the
computational effort for finding all mixed cells in a fine mixed subdivision.
4. Usage of D E M i C s . After unpacking the software package
DEMiCs, the user will see SRC and polySys, which include source and
sample files, in the main directory. The make file exists in the directory
SRC. The executable ffie "demies" is generated by executing the following
command in the directory.
72 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

make a l l

The input file requires information regarding the support of a poly-

nomial system: the dimension of the system, the number of the distinct
support sets, the cardinahty, multipUcity and elements of each support set.
For example, consider the support sets for a semi-mixed system of type
(2,1,1) as follows:
_ . _ . _ / (1,0,0,0), (0,1,0,0), (0,0,1,0), (0,0,0,1),
oi . - >ii_^2- \ (0,0,0,0), (1,1,1,1)
{iQ,ioi {Q^iio), (0,0,2,0), (0,0,0,2),
52 := ^3 =
(0,0,0,0), (2,2,2,2)
(3,0,0,0), (0,3,0,0), (0,0,3,0), (0,0,0,3),
^^'- -^4- ^ (0,0,0,0), (3,3,3,3)

The input file for S = {81,52,83) is written in the following format:

# The dimension or t h e number of v a r i a b l e s

Dim = 4

# The number of t h e d i s t i n c t support s e t s

Support = 3

# The number of elements i n each support s e t

Elem = 6 6 6

# The m u l t i p l i c i t y of each support s e t

Type = 2 1 1

# The elements of t h e 1st support s e t

10 0 0
0 10 0
0 0 10
0 0 0 1
0 0 0 0
1111

# The elements of the 2nd support set

2 0 0 0
0 2 0 0
0 0 2 0
0 0 0 2
0 0 0 0
2 2 2 2
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 73

# The elements of t h e 3rd support s e t

3 0 0 0
0 3 0 0
0 0 3 0
0 0 0 3
0 0 0 0
3 3 3 3

The directory polySys also contains some sample files describing the support
sets of several benchmark polynomial systems.
The above input file is placed in SRC as "poly.dat". To compute the
mixed volume via a fine mixed subdivision, we simply execute

demies poly.dat

in SRC, in which the executable file "demies" and the input file "poly.dat"
exist. The software package then displays the total number of mixed cells,
the value of the mixed volume and cpu time on the screen.

# Mixed C e l l s : 4
Mixed Volume: 24

CPU time: 0 s

Furthermore, we can select three options "-c", "-s" and "-cs" when
running the program. The option "-c" ofifers information about each mixed
cell C = (Ciii e M) e n ( M ) . After executing the command

demies -e p o l y . d a t

the following information is displayed on the screen

# 1 : 1 : ( 1 2 6 ) 2 : ( 1 5 ) 3 : ( 5 3 )
Volume: 6

# 2 : 1 : ( 4 1 6 ) 2 : ( 1 5 ) 3 : ( 3 5 )
Volume: 6

# 3 : 1 : ( 4 2 6 ) 3 : ( 3 4 ) 2 : ( 6 5 )
Volume: 6

# 4 : 1 : ( 4 2 6 ) 3 : ( 4 5 ) 2 : ( 5 1 )
Volume: 6

# Mixed C e l l s : 4
74 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

Mixed Volume: 24

CPU time: 0.01 s

On the first line with " # 1", "1 : (1 2 6)" means the subset Ci =
{ail, CH2, oie} of Si. That is, the nunaber in front of a colon corresponds
to the index of the support set. We thus know that one of mixed cells
C = (Ci,C2,C3) e 5 i X 52 X ^3 consists of

Ci = { a i i , a i 2 , a i 6 } , C2 = {021,025} and C3 = {035,033}

where Uij is an element of Si. "Volume" on the next line represents the
volume of the mixed cell C = (Ci, C2, C3). Note that the mixed volume is
obtained from the summation of volumes of four mixed cells for the specific
lifting values u}i{a) (a € Si). On a line with " # " , the sequence of indices
i for the subset d of each support set St indicates the order of an index t
chosen from M \ L{C) in Algorithm 1. For example, the hue with " # 3"
shows that the support sets <Si, ^3 and S2 are chosen in this order.
The software package needs a seed number to generate a random num-
ber for each lifting value a;j(a) (a S Si). If no option is selected as stated
in the above, the seed number is automatically set to " 1 " . The option "-s"
is useful in the case where we change the seed number to generate different
lifting values for each execution. As an example, when "6" is chosen as the
seed number for the input data "poly.dat", we type the command

demies - s p o l y . d a t 6

Given the seed number "2", the "-cs" option is used as follows to get
detailed information about mixed cells.

demies - e s p o l y . d a t 2

5. N u m e r i c a l results. This software package has been tested on a

large variety of polynomial systems including unmixed, semi-mixed and
fully mixed types. The papers [10, 12] report the superiority of MixedVol
in computational time for these three types of the systems over the ex-
isting software packages: H0M4PS [11], MVLP [8], PHCpack [25], PHoM
[22] and mvol [18]. Therefore, we compare DEMiCs with MixedVol for each
type of polynomial systems in terms of the computational time. Note that
MixedVol employs the static enumeration method, while DEMiCs adopts
the dynamic enumeration method. All numerical experiments were exe-
cuted on a 2.4GHz Opteron 850 with 8GB memory, running Linux.
First, we observe how the computational time of DEMiCs and Mixed-
Vol varies depending on m, which is the number of distinct support sets
S\,S2,. • • ,Sm of the semi-mixed polynomial systems f{x) in cc € R". In
 DEMICS: SOFTWARE FOR COMPUTING THE MIXED VOLUME 75

numerical experiments, we deal with artificial semi-mixed systems, which

are created to investigate the feature of DEMiCs. Each support set Si of
the semi-mixed systems is given as follows. We choose the subset T of Z"
with # T = 2n such as

, (1,0,...,0,0), (0,1,...,0,0),---,(0,0,...,1,0),(0,0,...,0,0) \
^ ('^1 ,"0 , . . . , 0", 1^') , (0,
'M,...,0,l),---,(0,0,...,1,1),(0,0,...,0,1) /•

Note that the convex hull of T is the n-dimensional prism with a sim-
plex basis, and the n-dimensional volume is r ^ r w - Let e^ denote the
n-dimensional ith unit vector. For T C Z " , we consider the transition of
T by the direction vector Cj as Si. Namely,

Si := Bi + T = {a + Si : a e T} for each i € M.

Assume that each support set Si has the multiplicity n/m e Z for the
dimension n and the number of distinct support sets m. That is, we deal
with semi-mixed polynomial systems f(x) of type {n/m,n/m,...,n/m).
Here, the mixed volume for the support S = {Si,S2, •.. ,Sm) of f{x) is
calculated as n! x . ^^w = n because the n-dimensional volume of conv(iSi)
and conv(T) is equal to each other.
To demonstrate the performance of DEMiCs and MixedVol, we choose
two different values n = 18,24 for the dimension of elements in Si C Z " ,
and change the number of distinct supports sets m in response to each
dimension n. Table 1 and 2, which are in the case of n = 18,24 respectively,
summarize the cpu time of DEMiCs and MixedVol for each system. We
performed 10 times numerical experiments for each system by choosing
the different lifting values in DEMiCs and MixedVol. The cpu time listed
in Table 1 and 2 is the average for each trial. The column "#Supp."
means the number of distinct support sets m, and "Ratio" indicates the
ratio between the cpu time of DEMiCs and MixedVol. The symbol "-"
means that the software package has not been applied to the corresponding
system. Prom these tables, we see that DEMiCs is superior to MixedVol in
the computational time if the number of the distinct support sets is large.
To the contrary, if the number of the distinct support sets is small, we
may not expect the advantage of a dynamic enumeration method. One of
the main reasons is that there is not great difference between the structure
of the djmamic and static enumeration trees. Moreover, DEMiCs needs
more computational tasks involved in choosing an index t from M \ L{C)
at Algorithm 1. Therefore, DEMiCs takes more computational time than
MixedVol for semi-mixed systems with a few distinct supports.
Second, we consider the following benchmark polynomial systems, in-
cluding unmixed, semi-mixed and fully mixed systems. The PRS-10 and
RRS-12 systems, which are arising from kinematic problems in [21], have 12
polynomials with 12 variables and 11 polynomials with 11 variables, respec-
tively. The PRS-10 system is a semi-mixed system of type (1,1,1,9), and
76 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

TABLE 1
n = 18 (The mixed volume of all systems is 18).

# Supp. (m) DEMiCs MixedVol Ratio

m—1 0.052s 0.045s 0.87

m = 3 8.893s 4.969s 0.56
m = 6 8.715s 52.482s 6.02
m = 9 15.453s 3m40.422s 14.26
m = 18 lm7.927s Ihl3m36.590s 65.02

TABLE 2
n = 24 (The mixed vohrnie of all systems is 24).

# Supp. (m) DEMiCs MixedVol Ratio

771 = 1 0.599s 0.341s 0.57

m = 3 llm42.896s 5m32.361s 0.47
m = 6 4m20.044s Ih21ml3.110s 18.74
m = 8 4m31.044s 4h3m41.700s 53.95
m = 12 9m9.827s 23h43m40.700s 155.36
m = 24 lh40ml 1.080s -

the first three support sets have 4 elements, the last 100 elements. Also, the
RRS-12 system is an unmixed system of type (11), and the support set has
224 elements. The cycUc-n[3], chandra-n[5] and katsura-n[4] systems are
fully mixed systems, and size-expandable systems by the number n. The
katsura-n systems consist of (n +1) pol5niomials with (n+l) variables, and
the others n polynomials with n variables. The detailed description of the
systems can be found in the web site [23]. We changed the lifting values 10
times for each system, and executed numerical experiments. In Table 3, we
list the comparison of the average cpu time of DEMiCs and MixedVol. The
column ">i V" presents the mixed volume for the support of corresponding
systems, and "Ratio" is the ratio between the cpu time of these software
packages. The numerical results for the cyclic-n, chandra-n and katsura-n
systems in Table 3 show that DEMiCs improves the cpu time for finding
all mixed cells dramatically when we address the polynomial systems with
many distinct support sets. However, the numerical results on the PRS-10
and RRS-12 systems imply that it may be difficult for DEMiCs to deal
with the unmixed and semi-mixed system which has only a few distinct
support sets, compared with MixedVol.
Finally, we consider the large-scale cyclic-n, chandra-n and katsura-n
polynomial systems. Numerical experiments were carried out 5 times for
each system associated with the different lifting values. Table 4 exhibits the
 DEMICS: SOFTWARE FOR COMPUTING T H E MIXED VOLUME 77
TABLE 3
CPU time for the benchmark system,s.

System Size (n) A^V DEMiCs MixedVol Ratio

PRS-10 142,814 14.3s 4.0s 0.28

RRS-12 226,512 29.9s 0.7s 0.02

Cyclic-n n = 12 500,352 lmll.6s 4m43.0s 3.95

n = 13 2,704,156 llm0.3s 49m57.4s 4.54
n = 14 8,795,976 Ih27m27.3s 7hl4m24.1s 4.97

Chandra-n n = 17 65,536 lm9.4s 33ml3.4s 28.70

n = 18 131,072 3ml0.5s 2hl4ml5.3s 42.29
n = 19 262,144 9m21.1s 8hl9m6.3s 53.38

Katsura-n n = 12 4,096 46.9s 14m3.5s 17.98

n = 13 8,192 5m8.1s Ih21ml9.4s 15.84
n = 14 16,384 24ml7.2s 7h54m29.4s 19.54

TABLE 4
A large size of cyclic-n, chandra-n and katsura-n systems.

System Size (n) MV CPU time

Cyclic-n n = 15 35,243,520 13h33m53.8s

n = 16 135,555,072 110h21m40.5s

Chandra-n n = 20 524,288 29m50.8s

n = 21 1,048,576 Ihl5ml9.7s
n = 22 2,097,152 3h5m48.2s
n = 23 4,194,304 Ilh24m4.6s
n = 24 8,388,608 30hlm33.6s

Katsura-n n = 15 32,768 Ih30m54.7s

n = 16 65,536 9h49m20.8s
n = 17 131,072 46h37m35.8s

average cpu time of DEMiCs for each system. As compared with numerical
results in Table 2 of [19], the computational time of DEMiCs is less than
that of the program developed in [19] as the size of the systems becomes
larger. It could be due to improvement on how to use memory space in
DEMiCs.
78 TOMOHIKO MIZUTANI AND AKIKO TAKEDA

6. Concluding remarks. In this paper, we introduced the software

package DEMiCs. Using the dynamic enumeration method, this package
computes the mixed volume, which can be obtained from the volumes of
all mixed cells, for the support of a semi-mixed polynomial system. The
dynamic enumeration method, which was developed for a fully-mixed type
in [19], can be extended to a semi-mixed type naturally. Numerical results
show that this package significantly is faster than existing ones for large-
scale semi-mixed systems with many distinct support sets. We confirm that
it is important to take account of how we construct an enumeration tree
for finding mixed cells. Prom numerical results, we recognize that DEMiCs
needs more computational time than MixedVol for an unmixed system and
a semi-mixed system with a few distinct support sets. It appears that
the dynamic enumeration method does not have a beneficial eflFect on such
systems, because the structure of the dynamic tree is almost the same as
that of the static tree. Although we proposed one strategy for constructing
an enumeration tree dynamically, there may exist other strategies, which
can improve our dynamic enumeration. This is our future work. Finding
mixed cells plays a crucial role in the polyhedral homotopy method. We
expect that DEMiCs opens the way for computing all isolated zeros of
large-scale polynomial systems by polyhedral homotopy method.

Acknowledgments. The author is grateful to the anonymous referees

for their valuable comments.

REFERENCES

[1] D.N. BERNSHTEIN, The number of roots of a system of equations, Funct. Anal.
Appl. 9 (1975), pp. 183-185.
[2] U. BlTKE, Mixed volumes of polytopes, Archiv der Mathematik 58 (1992),
pp. 388-391.
[3] G. BJORK AND R . F R O B E R G , A faster way to count the solutions of inhomoge-
neous system,s of algebraic equations, J. Symbolic Comput. 12(3) (1991),
pp. 329-336.
[4] W. BoEGE, R. GEBAUER, AND H . KREDEL, Some examples for solving systems of
algebraic equaiiona by calculating Groebner bases, J. Symbolic Comput. 2(1)
11986), pp. 83-98.
[5] S. CHANDRASEKHAR, Radiative Transfer, Dover, NY, 1960.
[6] D.A. Cox, J. LrrTLE AND D . O ' S H E A , Using Algebraic Geometry, Springer-Verlag,
New York, 2nd edition, 2004.
[7] Y. DAI, S . K I M , AND M . KOJIMA, Computing all nonsingular solutions of cyclic-
n polynomial using polyhedral homotopy continuation methods, J. Comput.
Appl. Math. 152 (2003), pp. 83-97.
f8] I.Z. EMIRIS AND J . F . CANNY, Efficient incremental algorithms for the sparse resul-
tant and the mixed volume, J. Symbolic Comput. 20(2) (1995), pp. 117-149.
Softwaie available at http://cgi.di.uoa.gr/~emirjs/index-eng.html.
[9] T. G A O AND T . Y . L I , Mixed volume computation via linear programming,
Taiwanese J. Math. 4 (2000), pp. 599-619.
[10] , Mixed volume computation for semi-mixed systems, Discrete and Comput.
Geom. 29(2) (2003), pp. 257-277.
 DEMICS: SOFTWARE F O R COMPUTING T H E MIXED VOLUME 79

[11] , T h e software package HOM4PS is available at http://www.

mth.msu.edu/~li/.
[12] T. G A O , T . Y . L I , AND M . W U , Algorithm 846: MixedVol: A Software Package
for Mixed Volume Computation, ACM Trans. Math. Software 31(4) (2005),
pp. 555 - 560. Software available at http://www.csulb.edu/~tgEio/.
[13] T . G U N J I , S . K I M , M . K O J I M A , A. TAKEDA, K . FUJISAWA, AND T . MIZUTANI, PHOM
- a Polyhedral Homotopy Continuation Method. Computing 73(1) (2004),
pp. 57-77.
[14] T . GUNJI, S . K I M , K . FUJISAWA, AND M . KOJIMA, PHoMpara - Parallel imple-
mentation of the polyhedral homotopy continuation method, Computing 77(4)
(2006), pp. 387-411.
[15] B. HuBER AND B. STURMFELS, A Polyhedral method for solving sparse polynomial
systems. Math. Comp. 6 4 (1995), pp. 1541-1555.
[16] S. K I M AND M . KOJIMA, Numerical Stability of Path Tracing in Polyhedral
Homotopy Continuation Methods, Computing 7 3 (2004), pp. 329-348.
[17] T.Y. L I , Solving polynomial systems by polyhedral homotopies, Taiwanese J. Math.
3 (1999), pp. 251-279.
[18] T . Y . L I AND X. Ll, Finding Mixed Cells in the Mixed Volume Computa-
tion, Found. Comput. Math. 1 (2001), pp. 161-181. Software available at
http://www.math.msu.edu/~li/.
[19] T . MIZUTANI, A. TAKEDA, AND M . KOJIMA, Dynamic Enumeration of All Mixed
Cells, Discrete Comput. Geom. 37(3), (2007), pp. 351-367.
[20] A. MORGAN, Solving polynomial systems using continuation for engineering and
scientific problems, Pentice-Hall, New Jersey, 1987.
[21] H.-J. Su, J.M. MCCARTHY, AND L . T . WATSON, Generalized Linear Product Ho-
motopy Algorithms and the Computation of Reachable Surfaces, ASME J.
Comput. Inf. Sci. Eng. 4(3) (2004), pp. 226-234.
[22] A. TAKEDA, M . K O J I M A , AND K . FUJISAWA, Enumeration of all solutions of a
combinatorial linear inequality system arising from the polyhedral homotopy
continuation method, J. Oper. S o c , Japan 4 5 (2002), pp. 64-82. Software
available at http://www.is.titech.ac.jp/~kojima/index.html.
[23] J. VERSCHELDE, T h e database of polynomial systems is in his web site:
http://www.math.uic.edu/~jan/.
[24] J. VERSCHELDE, P . VERLINDEN, AND R . COOLS, Homotopies exploiting Newton
polytopes for solving sparse polynomial systems, SIAM J. Numer. Anal. 3 1
(1994), 915-930.
[25] J. VERSCHELDE, Algorithm 795: PHCPACK: A general-purpose solver for polyno-
mial systems by homotopy continuation, ACM Trans. Math. Softw. 2 5 (1999),
pp. 251-276. Software available at http://www.math.uic.edu/~jan/.
 S Y N A P S : A LIBRARY FOR D E D I C A T E D APPLICATIONS
IN SYMBOLIC N U M E R I C C O M P U T I N G
BERNARD MOURRAIN*, JEAN-PASCAL PAVONE*,
PHILIPPE T R E B U C H E T t , ELIAS P. TSIGARIDAS*, AND JULIEN WINTZ*

A b s t r a c t . We present an overview of the open source library SYNAPS. We describe

some of the representative algorithms of the library and illustrate them on some explicit
computations, such as solving polynomials and computing geometric information on
implicit curves and surfaces. Moreover, we describe the design and the techniques we
have developed in order to handle a hierarchy of generic and specialized data-structures
and routines, based on a view mechanism. This allows us to construct dedicated plugins,
which can be loaded easily in an external tool. Finally, we show how the design of the
library allows us to embed the algebraic operations, as a dedicated plugin, into the
external geometric modeler AXEL.

K e y words. Symbolic-numeric computation, software.

A M S ( M O S ) subject classifications. Primary 68W30, 68W25, 65D17.

The aim of this paper is to give an overview of the software library

SYNAPS^. It is an open source project, the objective of which is to provide
a coherent and efficient library for symbolic and numeric computations.
It implements data-structures and classes for manipulating basic algebraic
objects, such as (dense, sparse, structured) vectors, matrices, univariate
and multivariate polynomials. It also provides fundamental methods such
as algebraic number manipulation tools, different types of univariate and
multivariate polynomial root solvers, resultant and gcd computations, etc.
The main motivation behind this project, is the need to combine symbohc
and numeric computations, which is ubiquitous in many problems. Starting
with an exact description of the equations, in most cases, we will eventually
have to compute an approximation of the solutions. Even more, in many
problems, the coefficients of the equations may only be known with some
inaccuracy (due, for instance, to measurement errors). In these cases, we
are not dealing with a solely system but with a neighborhood of an exact
system and we have to take into account the continuity of the solutions
with respect to the input coefficients. This leads to new, interesting and
challenging questions both from a theoretical and a practical point of view,
that lie in the frontier between Algebra and Analysis and witnesses the
emergence of new investigations. In order to develop efficient implemen-
tations for such problems we have to combine algorithms from numeric
and symbolic computation and to develop and manipulate data structures

*GALAAD, Inria, B P 93, 06902 Sophia-Antipolis, Prance.

tSPIRAL, L1P6, 8 rue du capitaine Scott 75015 Paris, France.
* Department of Informatics & Telecommunications, National Kapodistrian
University of Athens, Panepistimiopolis 15784, Greece.
^ http://synaps.inria.txl
81
82 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

that are on one hand generic and on the other are easily tuned to specific
problems. Moreover, the reusability of external or third-party libraries,
such as LAPACK (Fortran library for numerical linear algebra), GMP (C li-
brary for extended arithmetic) has to be considered carefully. Specialized
routines provided by these external tools have to coexist with generic im-
plementation. Therefore, the software should be designed so that it can
connect, in an automatic and invisible to end-user way, the appropriate
implementation with the needed operation.
In this paper, we first describe representative algorithms available in
the library, and illustrate them by some explicit computations. We begin
with a description of the solvers of polynomial equations. These tools are
used as bla^k boxes in geometric computations on implicit curves and sur-
faces. We show how the first level of data structures and polynomial solving
implementations are composed to build such algorithms. Such higher level
operations on geometric objects are embedded in the geometric modeler
AXEL^, as a dedicated plugin. We describe the design and techniques we
have developed to handle a hierarchy of generic and specialized implemen-
tations, based on a view mechanism. This approach is extended to build
plugins, which provide the equivalent functions in an interactive environ-
ment. In particular, we show how template mechanisms can be exploited
to transform static strongly typed code into dynamic polymorphic and
generic functions, assembled into a plugin that can be loaded easily in an
external tool.
1. Solvers of polynomial equations. A critical operation, which
we will have to perform in geometric computations on curves and surfaces,
is to solve polynomial equations. In such a computation, we start with
input polynomial equations (possibly with some incertitude on the coef-
ficient) and we want to compute an approximation of the (real) roots of
these equations or boxes containing these roots. Such operation should
be performed very efficiently and with guarantee, since they will be used
intensively in geometric computation.
In sections 1.1, 1.2, 1.3, we describe subdivision solvers which are
based on certified exclusion criteria. In other words, starting from an ini-
tial bounded domain, we remove subdomains which are guaranteed not to
contain a real solution of the polynomial equations. A parameter e > 0
is controlling the size of the boxes that are kept. For univariate polyno-
mials, existence and uniqueness criteria are applied to produce certified
isolation intervals which contain a single root. Such criteria also exist in
the multivariate case, but are not yet available in our current implemen-
tation. The interest of these subdivision methods, compared to homotopy
solvers [34], [15] or algebraic solvers [13], [33] is that only local informa-
tion related to the initial domain are used and it avoids the representation
or approximation of all the complex roots of the system. The methods

•^http: //axel. inria.fr/

 BERNARD MOURRAIN ET AL. 83

are particularly efficient for systems where the number of real roots is
much smaller that the number of complex roots or where the complex
roots are far from the domain of interest. However multiple roots usually
reduces their performance if their isolation is required, in addition to their
approximation.
1.1. Univariate subdivision solvers. Let us consider first an exact
polynomial / = J2i=o^i^^ ^ Q[^]- Our objective is to isolate the real
roots of / , i.e. to compute intervals with rational endpoints that contain
one and only one root of / , as well as the multiplicity of every real root.
The algorithms take these exact input polynomials and output certified
isolation intervals of their real roots. Some parts of the computation are
performed with exact integer or rational arithmetic (using the library GMP),
but some other parts might be implemented using floating point arithmetic.
It uses adapted rounding modes, to be able to certify the result. Here is
the general scheme of the subdivision solver that we consider, augmented
appropriately so that it also outputs the multiplicities. It uses an external
function V{f, I), which bounds the number of roots of / in the interval / .
ALGORITHM 1.1. R E A L R O O T ISOLATION

I N P U T : A polynomial / e Z[x], such that deg(/) = d and £ ( / ) = r.

O U T P U T : A list of intervals with rational endpoints, which contain one and only one
real root of / and the multiplicity of every real root.
1. Compute the square-free part of / , i.e. fred
2. Compute an interval /o = {—B, B) with rational endpoints that contains
all the real roots. Initialize a queue Q with /o.
3. While Q is not empty do
a) Pop an interval / from Q and compute v := V{f,I).
b)\f v = 0, discard / .
c) If i; = 1, output I.
d) \f V > 2, split / into II and IR and push them to Q.
4. Determine the multiplicities of the real roots, using the square-free factor-
ization of / .

Two families of solvers have been developed. One using Sturm theo-
rem, where V{f, I) returns the exact number (counted without multiplici-
ties) of the real roots of / in / . The second one based on Descartes' rule and
Bernstein representation, where V{f,I) bounds the number of real roots of
/ i n 7 (counted with multiplicities). As analyzed in [10], the bit complexity
of both approaches is in OB(d^r^), if / € Z[a;], deg(/) = dis the degree
of / and C{f) =T the maximal bitsize of its coefficients. Notice that with
the same complexity bound, we can also compute the multiplicities of the
real roots. However in practice, the behavior is not exactly the same, as
we will illustrate it.
84 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

1.1.1. Sturm subdivision solver. We recall here the main ingredi-

ents related to Sturm(-Habicht) sequence computations.
Let / = Yll=ofkx'',9 = 121=09kx'' E Z[x] where deg(/) =p>q =
deg{g) and C{f) = C{g) = r. We denote by rem(/, p) and quo {f,g) the
remainder and the quotient, respectively, of the Euclidean division of / by
g, in Q{x].
DEFINITION 1.2. [35, 3] The signed polynomial remainder sequence
of f and g, denoted by S P R S {f,g), is the polynomial sequence

Ro = f,Ri=g,R2 = - rem {f,g),... ,Rk = - rem {Rk-2, Rk-i),

with k the minimum index such that Tem{Rk-i,Rk) = 0. The quotient
sequence of f and g is the polynomial sequence {Qi}o<i<k-i, where Qi =
quo{Ri,Ri+i) and the quotient boot is {Qo,Qi,•••,Qk-i,Rk)-
Another construction yields the Sturm-Habicht sequence of / and
g, i.e. S t H a ( / , p ) , which achieves better bounds on the bit size of the
coefficients.
Let Mj be the matrix which has as rows the coefficient vectors of
the polynomials /x'~^~^, fx''~'^~^, ... ,fx,f, g,gx,... ,gxP~'^^^,gx'^~^''^
with respect to the monomial basis a;P+'~^~-', arP+'"~^"^, ...,x,l. The
dimension of Mj is {p + q- 2j) x (p + q — j)- For I = 0,... ,p + q — 1 — j let
Mj be the square matrix of dimension {p + q — 2j) x ip + q — 2j) obtained
by taking the first p + q — 1 — 2j columns and the 1 + {p + q — 2j) column
oiMj.
DEFINITION 1.3. The Sturm-Habicht sequence of f and g, is the se-
quence StH&{f,g) = {Hp =^ Hp{f,g), ...,Ho = Ho{f,g)), where Hp = / ,
Hp_i=g, Hj = (_l)(p+9-j)(p+9-^-i)/2 E t o d e t (Mj)x'.
For two polynomials of degree p and q and of bit size bounded by r ,
such sequences and their evaluation at a rational point a, where a S Q U
{±00} and £ (a) =CTcan be done respectively with complexity OB {p^qT)
and OB(gmax{pT, gcr}). For more details, see [35, 3, 18, 19].
The structure SturmSeq encodes these Sturm sequences in SYNAPS.
Several constructions are implemented, specified by a class in the construc-
tor, Euclidean, primitive and subresultant polynomial remainder sequences.
Let us present an example of code for constructing the Sturm-Habicht se-
quence s of two polynomials p, q £ Z[a;]. The implementation corresponds
to a variant of the inductive construction described in [3].
UPolDse<ZZ> p("3*x-5+23*x-3-x-2+234"), q("10*x-4+200*x-2-13243");
SturmSeq<ZZ> s(p,q,HABICHT());

The result is a sequence of polynomials, with coefficients in the initial

ring ZZ:
[3*x-5+23*x-3-x-2+234, 10*x-4+200*x-2-I3243,
3700*x-3+100*x-2-397290*x-23400, -174378300*x-2-4685100*x+1813200700,
796993090279590*x+51961697166600, -37867420670503735668763]
Such a sequence can be used to count roots in an interval. Let W(^f^g) (a)
denote the number of modified sign changes of the evaluation of StHa(/, g)
at a.
 BERNARD MOURRAIN ET AL. 85

THEOREM 1.1. [3, 36] Let f,g E Z[a;], where f is square-free and f
is the derivative of f and its leading coefficient fd > 0. If a < h are both
non-roots of f andj ranges over the roots of f in (a, b), then M^(/,3)(a) —
Wu,g)ib) = E ^ sign (/'(7)5(7))- If 5 = / ' then StHa(/, / ' ) is a Sturm
sequence and Th. 1.1 counts the number of distinct real roots of / i n (a, b).
For the Sturm solver V{f,[a,b]) will denote y ( / , [a,b]) = M/(y//)(a) -
Wif,nib).
1.1.2. Bernstein subdivision solver. In this section, we recall the
background of Bernstein polynomial representation and how it is used in
the subdivision solver. Given an arbitrary univariate polynomial function
/(x) € K, we can convert it to a representation of degree d in Bernstein
basis, which is defined by:

fix) ^ Yl ^^^'(^)' ^"d ^'(^) = ( i ) ^'(1 - ^)'~' (1-1)

where bi is usually referred as controlling coefficients. Such conversion is

done through a basis conversion [11]. The above formula can be generalized
to an arbitrary interval [a, b] by a variable substitution x' = (b — a)x + a.
We denote by 5^(x; a, b) = {fj{x — ay{b ~ a;)'*"'(6 — a)~'^ the correspond-
ing Bernstein basis on [a, 6]. There are several useful properties regarding
Bernstein basis given as follows:
• Convex-Hull Properties: Since ^iB'^{x;a,b) = 1 and Vx € {a,b],
B^{x;a,b) > 0 where i = 0, ...,d, the graph of / ( x ) = 0, which is
given by (x, /(x)), should always lie within the convex-hull defined
by the control coefficients (3,^1) [11].
• Subdivision (de Casteljau): Given to £ [0,1], /(x) can be repre-
sented piece-wisely by:
d d
fix) = Y,bfBiix;a,c) = Ybf~'^Biix-c,b), where (1.2)
i=0 1=0

6f) = (1 - t o ) 6 f - ' ) -f to6l+7'^ and c = (1 - to)a + tob. (1.3)

Another interesting property of this representation related to Descartes

rule of signs is that there is a simple and yet efficient test for the existence
of real roots in a given interval. It is based on the number of sign variation
y(bfe) of the sequence hk — [bi,..., bk] that we define recursively as follows:

f l,iibkbk+i <0
Vihk+i) = Vihk) + l ^^^^^ (1.4)

With this definition, we have:

86 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

PROPOSITION 1.1. Given a polynomial f{x) = YHhBf(x;a,h), the

number N of real roots of f on ]a,h[ is less than or equal to V{b), where
b = (fej)i=i,...,n o,nd N = V{b)mod 2. With this proposition,
• if V^(b) = 0, the number of real roots of / in [a, b] is 0;
• if V{h) = 1, the number of real roots of / in [a, 6] is 1.
This function V yields another variant of subdivision algorithm 1.1. In
order to analyze its behavior, a partial inverse of Descartes' rule and lower
bounds on the distance between roots of a polynomial have been used. It is
proved that the complexity of isolating the roots of a polynomial of degree
rf, with integer coefficients of bit size < T is bounded by O (^^T^) up to
some poly-logarithmic factors. See [8, 10] for more details.
Notice that this localization algorithm extends naturally to B-splines,
which are piecewise polynomial functions [11].
The approach can also be extended to polynomials with interval co-
efficients, by counting 1 sign variation for a sign sub-sequence -|-, ?, — or
— , ? , + ; 2 sign variations for a sign sub-sequence -I-, ?, -|- or —,?,—; 1 sign
variation for a sign sub-sequence ?, ?, where ? is the sign of an interval
containing 0. Again in this case, if a family f of polynomials is represented
by the sequence of intervals b = [bp,..., b^] in the Bernstein basis of the
interval [a, 6]
• if y ( b ) = 1, all the polynomials of the family f have one root in
[a,6],_
• if F(b) = 0, all the polynomials of the family f have no roots in
[a, 6].
This subdivision algorithm, using interval arithmetic, yields either intervals
of size smaller than e, which might contain the roots of / = 0 in [a, 6]
or isolating intervals for all the pol3niomials of the family defined by the
interval coefficients.
A variant of such approach in the monomial basis is called Uspensky's
method, e.g. [8, 29] and references therein. Another variant, using Cauchy's
lower bound on the positive roots of the polynomial, isolates the real roots
by computing their continued fraction expansion, c.f. [9] and references
therein. The expected complexity of this variant is the same as the worst
case bound of the subdivision solvers, i.e 0 B ( C / ^ T ^ ) .

1.1.3. Experimentation. In this section, we describe the experimen-

tal behavior of some of the implemented subdivision solvers on specific data
sets. The solvers take as input, a polynomial f with integer, rational or
interval coefficients and output intervals with rational endpoints. All use
the same initial interval, given by Cauchy bound.
The following graphs illustrate the behavior of various univariate sol-
vers, on random (D\) and Mignotte (D5) polynomials of maximum coeffi-
cient bit size 50 bits:
 BERNARD MOURRAIN ET AL. 87

The different solvers are: IslSturmQQ based on the construction of the

Sturm-Habicht sequence and subdivisions, using rational numbers or large
integers; IslBzIntegerZZ implementing the Bernstein subdivision solver,
using extended integer coefficients; IslBzBdgSturmQQ combining two sol-
vers (in a first part, the polynomial is converted to the Bernstein represen-
tation on the initial interval, using rational arithmetic and its coefficients
are rounded to double intervals. The Bernstein solver is applied on the
polynomial with interval coefficients. If the size of the domain is too small,
the Sturm solver is launched), CORE [16] and SlvAberthQQ corresponding
to MPSOLVE, a numerical solver based on Aberth's method [4] and imple-
mented by G. Fiorentino and D. Bini.
The average time over 100 runs is in seconds. The experiments were
performed on a Pentium (2.6 GHz), using g-f-f- 3.4.4 (Suse 10). The ex-
tended arithmetic is based on the hbrary GMP. For polynomials with few,
distinct and well separated real roots (Di), we observe that the Bernstein
subdivision solver perform well. When there are roots that are very close
(Ps), the computation time of the Sturm-Habicht sequence is negligible.
A combined solver based on numerical solvers such as MPSOLVE and sub-
division techniques using for instance the Bernstein representation seems
to be the most efficient approach. For more details, the reader may refer
to [10}.

1.2. Algebraic numbers. Algebraic numbers are of particular im-

portance in geometric problems such as arrangement or topology compu-
tation. In geometric modehng the treatment of algebraic curves or surfaces
leads impUcitly or explicitly to the manipulation of algebraic numbers. A
package of the hbrary is devoted to such problems. It is dealing with real
algebraic numbers, i.e. those real numbers that satisfy a polynomial equa-
tion with integer coefficients, form a real closed field denoted by Raig = Q-
Prom all integer polynomials that have an algebraic number a as root,
the primitive one (the gcd of the coefficients is 1) with the minimum de-
gree is called minimal. The minimal polynomial is unique (up to a sign),
88 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

primitive and irreducible, e.g. [36]. For the computation with real alge-
braic numbers, we use Sturm-Habicht sequences, hence it suffices to deal
with algebraic numbers, as roots of a square-free polynomial and not as
roots of their minimal one. In order to represent a real algebraic number
we choose the isolating interval representation, i.e. by a square-free poly-
nomial and an isolating interval, that is an interval with rational endpoints,
which contains only one root of the polynomial:

r = ( / ( X ) , [a, b]), where / € Z[X]and a, 6 e Q.

In the geometric applications (topology of algebraic curves and surfaces,

arrangement computation,...) that we are targeting, this representation is
enough. This is the reason why, we have not considered towers of algebraic
extensions (algebraic numbers which defining polynomials are also algebraic
numbers).
In order to achieve high performance for problems involving small de-
gree polynomials (typically geometric problems for conies), the treatment
of polynomials and algebraic numbers of degree up to 4, is preprocessed.
We use precomputed discrimination systems (Sturm-Habicht sequences) in
order to determine the square-free polynomial and to compute the isolating
interval as function in the coefficients of the polynomial (and to compare
algebraic numbers).
For poljTiomials of higher degree, we use a Sturm-like solver in order
to isolate the roots of the polynomial, but we can use any other solver
that can return isolating intervals. Evidently, a real algebraic number is
constructed by solving (in our case by isolating) the real roots of an integer
univariate polynomial.
Let us demonstrate the capabilities of the library by an example:
1: #include <syiiaps/usolve/Algebraic. h>

2: using namespace::SYNAPS;
3: using naiespace::std;

4: typedef ZZ NT;
5: typedef Alg6braic<NT> SOLVER;

6: int mainO {
7: SOLVER::Poly f("x-9-29*x-8+349*x-7-2309*x-6+9307*x-5-23771*x-4
+38511*x-3-38151*x-2+20952*x-4860");
8: Seq<SOLVHl::R0_t> sol= solve(f, SOLVERO);
9: for (unsigned i=0; i< sol.sizeO; ++i)
10: cout « "(" « i « ") " « sol[i] « endl;
11: return 0; }

First the user declares the number type of the coefficients of the poly-
nomials that he wants to deal with. In our case we use ZZ, which corre-
spond to GMP integers. In the sequel, he declares the solve algorithm, which
means that he chooses an algorithm in order to isolate the real roots of an
integer polynomial. There are many solvers in SYNAPS and each of them
 BERNARD MOURRAIN ET AL. 89

has a similar structure. In our example, we choose the Sturm subdivision

solver. For the various univariate solvers available in SYNAPS, the reader
may refer to the previous section or to [10]. Inside the main routine, the
user constructs a polynomial and solves it using the solve function. This
function constructs a sequence of real algebraic numbers that are printed
subsequently.
The implementation of the algebraic numbers is in the namespace
ALGEBRAIC. Since algebraic numbers need a lot of information concerning
the ring and the field number type, the number type of the approximation
etc, we gathered all this information into a s t r u c t called ALG_STURM<RT>,
which takes only the ring number type RT as parameter and from this
class we can determine all the other types. The class which implements
the real algebraic numbers is root_of <RT>. It provides construction func-
tions (such as solve, RootOf), comparison functions, sign function and
extensions to bivariate problems, considered as univariate over a univari-
ate pol)niomial ring. In order to compare two algebraic numbers, filtering
techniques improving the numerical approximation combined with explicit
methods based on Sturm's theorem are used. For the complexity of these
operations, the reader may refer to [10] and references therein.
Moreover, projection-based algorithms exists for constructing pairs of
real algebraic numbers that are real solutions of bivariate polynomials sys-
tems, as well as functions for the computing the sign of a bivariate integer
polynomial, evaluated over two real algebraic numbers.

1.3. Multivariate Bernstein subdivision solver. We consider

here the problem of computing the solutions of a polynomial system

fl{xi,...,Xn) = 0,

fs{xi,...,Xn) =0,

in a box B := [ai, 6i] x • • • x [a„, 6„] C R". The method for approximating
the real roots of this system, that we describe now uses the representation
of multivariate polynomials in Bernstein basis, analysis of sign variations
and univariate solvers (Section 1.1.2). The output is a set of small-enough
boxes, which contain these roots. The boxes which are removed are guar-
anteed to contain no root, but the existence and uniqueness of a root in
each output box is not provided, neither the multiplicity. The computation
is done with floating point arithmetic, using controlled rounding modes.
The subdivision solver [24] that we describe now, can be seen as an
improvement of the Interval Projected Polyhedron algorithm in [31].
In the following, we use the Bernstein basis representation of a multi-
variate polynomial / of the domain I := [ai,bi] x • • • x [a„, 6„] C M":
90 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

dl dn
f{xi,...,Xn) = X ] ••• X ] ^ii'-'in -B^i(a;i;oi,6i)---B^^a;(a;„;a„,6„).
41=0 in=0

DEFINITION 1.4. For any f e M[x] and j = l,...,n, let

r^Q{0<tk<dk,kfy} '

Mj{f;xj) = Y^ max bi,,,„,i^x B'J{xj;aj,bj).

THEOREM 1.2 (Projection Lemma). For any u = ( u i , . . . , u „ ) € / ,

and any j = 1,... ,n, we have

mif;Uj)<f{u)<M{f;uj).

As a direct consequence, we obtain the following corollary:

COROLLARY 1.1. Foranyrootu = ( u i , . . . , u „ ) of the equation f {^) =
0 in the domain I, we have fi. <Uj <'pj where

• ^. (resp. JIj) is either a root ofmj{f;Xj) = 0 or Mj{f;Xj) = 0 in

[aj,bj] or aj (resp. bj) ifmj{f;Xj) = 0 (resp. Mj(f;xj) =0) has
no root on \aj,bj],
• rnj{f;u) < 0 < Mj{f;u) on ^£.,71,].

The general scheme of the solver implementation consists in

1. applying a preconditioning step to the equations;
2. reducing the domain;
3. if the reduction ratio is too small, we split the domain
until the size of the domain is smaller than a given epsilon.
The following important ingredients of the algorithm parametrize its
implementation:
Preconditioner. It is a transformation of the initial system into a sys-
tem, which has a better numerical behavior. Solving the system f = 0 is
equivalent to solving the system M f = 0, where M is an s x s invertible ma-
trix. As such a transformation may increase the degree of some equations,
with respect to some variables, it has a cost, which might not be negligible
in some cases. Moreover, if for each polynomial of the system not all the
variables are involved, that is if the system is sparse with respect to the
variables, such a preconditioner may transform it into a system which is not
sparse anymore. In this case, we would prefer a partial preconditioner on
 BERNARD MOURRAIN ET AL. 91

a subset of the equations sharing a subset of variables. We consider Global

transformations, which minimize the distance between the equations, con-
sidered as vectors in an afiine space of polynomials of a given degree and
Local straightening (for s = n), which transform locally the system f into
a system J~^f, where J = (9xi/j(uo)i<i,j<s is the Jacobian matrix of f at
a point u of the domain I, where it is invertible.
It can be proved that the reduction based on the polynomial bounds
m and M behaves like Newton iteration near a simple root, that is we have
a quadratic convergence, with this transformation.

Reduction strategy. It is the technique used to reduce the initial do-

main, for searching the roots of the system. It can be based on Convex
hull properties as in [31] or on Root localization, which is a direct improve-
ment of the convex hull reduction and consists in computing the first (resp.
last) root of the polynomial mj{fk;Uj), (resp. Mj{fk;Uj)), in the interval
[aj,bj]. The current implementation of this reduction steps allows us to
consider the convex hull reduction, as one iteration step of this reduction
process. The guarantee that the computed intervals contain the roots of / ,
is obtained by controlling the rounding mode of the operations during the
de Casteljau computation.

Subdivision strategy. It is the technique used to subdivide the domain,

in order to simplify the forthcoming steps, for searching the roots of the
system. Some simple rules that can be used to subdivide a domain and
reduce its size. The approach, that we are using in our implementation is
the parameter domain bisection: The domain B is then split in half in a
direction j for which \bj — aj\ is maximal. But instead of choosing the size
of the interval as a criterion for the direction in which we split, we may
choose other criterion depending also on the value the functions rui, Mj or
fj (for instance where M, — rrij is maximal). A bound for the complexity
of this method is detailed in [24]. It involves metric quantities related to
the system f = 0, such as the Lipschitz constant of f in B, the entropy of
its near-zero level sets, a bound d on the degree of the equations in each
variable and the dimension n.

Examples. Here are some comparisons of the different strategies, de-

scribing the number of iterations in the main loop, the number of subdivi-
sion of a domain, the number of boxes produced by the method, the time
it takes. We compare the method sbd, a pure subdivision approach, rd a
method doing first reduction and based on a univariate root-solver using
the Descarte's rule, sbds a subdivision approach using the global precon-
ditioner, rds a reduction approach using the global preconditioner, rdl, a
reduction approach using the Jacobian preconditioner. The first example
is a bivariate system, with equations of degree 12) in each variable, the
second example is of bi-degree (8,8).
92 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

(a) (b)

Example a:
method iterations subdivisions results time (ms)
sbd 4826 4826 220 217
rd 2071 1437 128 114
sbds 3286 3286 152 180
rds 1113 748 88 117
rdl 389 116 78 44

Example b:
method iterations subdivisions results time (ms)
sbd 84887 84887 28896 3820
rd 82873 51100 20336 4553
sbds 6076 6076 364 333
rds 1486 920 144 163
rdl 1055 305 60 120

For more details on this solver, see [28], [24].

1.4, Resultant-based methods. A projection operator is an oper-
ator which associates to an overdetermined polynomial system in several
variables a polynomial depending only on the coefficients of this system,
which vanishes when the system has a solution. This projection operation
is a basic ingredient of many methods in Effective Algebraic Geometry. It
has important applications in CAGD (Computer Aided Geometric Design),
such as for the problem of implicitization of parametric surfaces, or for sur-
face parametrization inversion, intersection, and detection of singularities
of a parametrized surface. The library implements a set of resultants
Such approach based on resultant constructions yields a preprocessing
step in which we generate a dedicated code for the problem we want to
handle. The effective resolution, which then requires the instantiation of
 BERNARD MOURRAIN ET AL. 93

the parameters of the problems and the numerical solving, is thus highly
accelerated. Such solvers, exploiting adequately tools from numerical linear
algebra, are numerically robust and compute efficiently approximations of
all the roots, even if they have multiplicities. They can be used directly in
geometric applications (see eg. [20]) or embedded in an algorithm which
validates afterward the approximation. Such validation step is not yet
provided in our current implementation of these resultant based methods.
The library SYNAPS contains several types of resultant constructions,
such that the projective resultant, the toric resultant (based on an imple-
mentation by I. Emiris), and the Bezoutian resultant. Using these resultant
matrix formulations, solving a polynomial problem can be reduced to solv-
ing the generalized eigenvector problem T*{x)v = 0, where T{x) is a matrix
of size N X N with polynomial coefficients, or equivalently a polynomial
with N X N matrix coefficients.
li d = maxij{deg{Tij{x))}, we obtain T{x) =:= Tdx"^ + Td-ix'^'^ +
\-To, where Ti are nx n matrices. The problem is transformed into a
generalized eigenvalue problem {A — XB)v = 0 where

/O
(I 0 0 \
0 \
B = 0 • • •
0 ••• 0 / / 0
\n n ••• n_j 0 -nj
Vo ...
where A, B are N x d constant matrices. And we have the following inter-
esting property :

/ V \

T\x)v = 0<^{A-xB) = 0.

\ x^-'v )
We apply it for implicit curve intersection problems in [5]. Given two
polynomials p, g € Q[a;, y], we compute their resultant matrix, with respect
to y. This yields a matrix T(x), from which we deduce the coordinates
of the intersection points by solving the generalized eigenvector problem
T{xf V = 0. The case of multiple roots in the resultant and of intersection
points with the same x-coordinates is analyzed in details, so that we can
recover their multiplicities.
Examples. Here are two examples which illustrate the behavior of the
implementation. The eigenvalue computation is using the routine dgegv
from the FORTRAN library LAPACK. The connection with this external
library is performed transparently, through the mechanisms of views, as
detailed in Section 3.
94 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

p = X* — 2x^2/ + y^ v^ -y^
q = y - Ix^
(xl, yl) = (1.6e-09,0) at multiplicity 4
(x2, y2) = (-0.5,0.5) of multiplicity 2
(x3, y3) = (0.5.0.5) of multiplicity 2
Execution time = 0.005s; Eps = 10'-6.

= ar'' + Zx^y^ + Zx^y^ + y" - 42o,2

p,--,.6. a;^y
q = y - x^ + x^
(xl, yl) (-3e-16,2e-17) of multiplicity 8
(x2, y2) (-0.60,-0.76) of multiplicity 1
(x3, y3) (-0.60,0.76) of multiplicity 1
(x4, y4) (0.72, -0.37) of multiplicity 1
(x5. y5) (0.72,0.37) of multiplicity 1
Execution time = 0.011s; Eps = 10"-3.

Such tools have also been used in the following CAD modeling problem: the
extraction of a set of geometric primitives properly describing a given 3D
point cloud obtained by scanning a real scene. If the extraction of planes
is considered as a well-solved problem, the extraction of circular cylinders,
these geometric primitives are basically used to represent "pipes" in an
industrial environment, is not easy and has been recently addressed. We
describe an application of resultant based method to this problem which has
been experimented in collaboration with Th. Chaperon from the MENSI
company. It proceeds as follows: First, we devise a polynomial dedicated
solver, which given 5 points randomly selected in our 3D point cloud, com-
putes the cyhnders passing through them (recall that 5 is the minimum
number of points defining generically a finite number of cylinders, actually
6 in the complex numbers). Then we apply it for almost all (or randomly
chosen) sets of 5 points in the whole point could, and extract the "clusters
of directions" as a primitive cylinder. This requires the intensive resolution
of thousands of small polynomial systems. Classical resultant or residual
resultant constructions are used in this case, to compute quickly their roots,
even in the presence of multiple roots.

1.5. Generalized normal form. The solver that we describe now

computes all the complex roots of a zero-dimensional polynomial systems
in C". It proceeds in two steps:
• Computation of the generalized normal form modulo the ideal
(/l: •••,/«)•
• Computation of the roots from eigencomputation.
Hereafter, the polynomials are in the ring K [ x i , . . . , x„] with coefficients in
the field K (eg. Q, R, C) and variables x j , . . . , X n .
A classical approach for normal form computation is through Grobner
basis [6]. Unfortunately, their behavior on approximate data is not satis-
factory [23]. In SYNAPS, a generaUzed normal form method is implemented
which allows us to treat poljmomial systems with approximate coefficients.
 BERNARD MOURRAIN ET AL. 95

more safely. Such a normal form computation is available with exact arith-
metic (rational numbers or modular numbers) and with extended floating
point arithmetic. The numerical stability of the generalized normal form
computation is improved by the allowing column pivoting on the coefficient
matrices of the involved polynomials, which is not possible for Grobner ba-
sis computation. In the case of floating point arithmetic, no certification is
provided yet in the current implementation.
The method constructs a set of monomials B and a rewriting family
F on the monomial set B, which have the following property: V/, / ' € F,
• / has exactly one monomial that we denoted by 7 ( / ) (also called
the leading monomial of / ) in dB,
. supp(/) C B+, supp(/ - 7(/)) C B,
• i f 7 ( / ) = 7 ( / ' ) t h e n / = /',
where B'^ = BUxiBU •• -UXnB, dB = B'^\B. Moreover, the monomial
set B will be connected to 1: for every monomial m in B, there exists
a finite sequence of variables (xj^ )je[i,;] such that 1 £ B, Ilj=i..,i'Xij €
B, W e [1, /] and n^eli.i] ^ij = '^•
The normal form construction is based on the following criteria:
THEOREM 1.3. [22] Let B be a monomial set connected to 1. Let
Rp : 5 + -^ B be a projection from J5+ to B, with kernel F and let I — (F)
be the ideal generated by F. Let Mi : B —* B be defined by b >-* RF{xib).
Then, the two properties are equivalent:
1. For all I <i,j <n, Mj o Mj = Mj o Mi.
2. K [ x i , . . . , x „ ] = ( B ) © J .
The algorithm consists in constructing degree by degree the set B and
in checking at each level, whether the partial operator of multiplication
commutes. For more details, see [33, 25, 26, 23].
Prom this normal form N, we deduce the roots of the system as follows.
We use the properties of the operators of multiplication by elements of
A = R/{fi,..,fs)), as follows (see [2], [21], [32]):

ALGORITHM 1.5. SOLVING IN THE CASE OF SIMPLE ROOTS

Let a& A such that a{C,i) # o ( 0 ) ^^ i¥^ J (which is generically the case).
1. Compute the Ma the multiplication matrix by a in the basis x^ =
( l , x i , . . . , x „ , . . . ) of A
2. Compute the eigenvectors A = (Ai, Kx^,..., Ax„,...) of M^.
3. For each eigenvector A with A i ^ 0, compute and return the point C =
V Ai '••• ' Ai ) •

The case of multiple roots is treated by simultaneous triangulation of sev-

eral multiphcation matrices. The main ingredients which are involved here
are sparse linear algebra (implemented from the direct sparse linear solver
superLU), and eigenvalue and eigenvector computation (based on LAPACK
96 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

routines). These different types of tools are combined together, in order

to obtain an efficient zero-dimensional polynomial system solver. We are
currently working on techniques adapted from thoses of [30, 27] to certify a
posteriori the numerical approximation of the roots obtained by such eigen-
values/eigenvectors computation. These techniques will also allow certified
root refinement. Eventually a purely symbolic representation of the roots
will be added.
2. Geometric computing with curves and surfaces. A special
context where algebraic operations on curves and surfaces are critical is
non-Unear computational geometry. Shapes are modeled by semi-algebraic
sets, where the local primitives are implicit or parametric curves or surfaces.
This is typically the case in CAGD, where NURBS or Bspline functions [11]
are used standard primitives. The degree of the piecewise polynomial mod-
els is usually 3 (or 5) in each variable. In this section, we describe some
functions from the module shape of the hbrary SYNAPS. We focus on tools
for computing the topology of impHcit curves and surfaces, which are fun-
damental in arrangement problems. All these methods rely on polynomial
solvers as fundamental ingredients. They guaranty their results, provided
the polynomial solvers are able to perform certified root isolation.

2.1. Sweeping of planar curves. Let C be a planar implicit curve

defined by a polynomial equation f{x,y) = 0, where / e Q[x,yj. Here is
the scheme of an algorithm, which outputs a graph of points in the plane,
isotopic to the curve C, and with the following properties:

ALGORITHM 2.1. TOPOLOGY OF AN IMPLICIT CURVE

I N P U T : an algebraic curve C given by a square-free equation f{x,y) = 0 with

/GQ[X,2/].
• Choose a direction (say the x-axis direction), and consider a virtual line
orthogonal to this direction, sweeping the plane;
• Detect at which critical position, the topology of the intersection of the line
and the curve changes (the roots of the resultant iesy{f,dyf){x) = 0.
• -Compute the corresponding intersection points / ( a , y) = 0 for a a root of
resy(/,aj,/)(a5) = 0.
• Compute the intersection points for sample lines in between these critical
positions.
• Check the generic position (at most one critical point per sweeping line)
and connect the computed points by segments, in order to get a graph
isotopic to C.

In this algorithm, we compute the Sturm-Habicht sequence of the two poly-

nomials / , dyf with respect to the variable y. The last term of this sequence
is their resultant r{x) G Q{x]. We solve this equation. Let ai < •• < as
 BERNARD MOURRAIN ET AL. 97

be the real roots of r(x) = 0. For each a^, we compute the correspond-
ing y such that f{ai,y) = 0. At this point, we also check that the curve
is in generic position, that is, there is at most one multiple solution of
f{ai, y) = 0. This construction involve the manipulation of algebraic num-
bers and the use of univariate solvers (see Sections 1.1 and 1.2).
Once these points are computed, we compute intermediate (rational)
points Ho < Hi < • • • < Us, such that fii-i < ai < Hi and the corresponding
y such that f{Hi,y) — 0. Here again we use a univariate solver, knowing
that the roots are simple (using a Bernstein subdivision solver).
All these points are collected and sorted by lexicographic order such
that X > y. The subset of points, corresponding to two consecutive x-
coordinates are connected, according to their y-coordinates. See [17] for
more details.
Examples.

Here is how the function and the external viewer AXEL in SYNAPS are called:
MPol<(3Q> p("5*x"4*y-2*x"6-4*x"2*y"2+y"3+y"5-2*y"4*x"2-y"4+2*x"2*y"3");
topology::point_graph<double> g;
topology(g, p,TopSueep2d<double,SIvBzBdg<double> >()) ;
axel::ostream os("tmp.axl") ; os«g; os.viewO;

It computes a graph of points g (of type topology: :point^raph<C>)

which is isotopic to the curve C, defined by the polynomial p. The coeffi-
cients of the points in the computed graph axe of type C. The polynomial p
is converted to a polynomial with rational coefiicients, if needed. The class
TopSweep2d<C, SLV> depends on two parameters:
» C, which is the type of coefiicients of the points in the result,
• SLV, which is the type of univariate solver to be used. The default
value is SlvbzBdg<double>.
2.2. Subdivision method for planar curves. In this section, we
consider a curve C in K^, defined by the equation f{x,y) = 0 with / G
Q[x, y] and a domain B = [a, b] x[c,d\ CM.^.
In order to compute the topology of this curve in a box, we use the
notion of regularity:
DEFINITION 2.2. We say that the curve C is y-regular (resp. x-regular)
in B, if C has no tangent parallel to the y-direction (resp. x-direction)
in B.
98 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

Notice that if C is x-regular (or y-regular) it is smooth in B since it

cannot have singular points in B. A curve is regular in B, if it is a;-regular
or j/-regular in B.
We use the property that if C is x-regular in B, then its topology can
be deduced from its intersection with the boundary dB.
PROPOSITION 2.1. [17] If C is regular in B, its topology in B is
uniquely determined by its intersection C n dB with the boundary of B.
Here is a simple test for the regularity of a curve in a domain, which
extends in some way the criterion in [12]:
PROPOSITION 2.2. [17] If the coefficients of dyf{x,y) ^ 0 (resp.
dxf{x, y) ^ 0^ in the Bernstein basis of the domain B = [a, b\ x [c, d] C R"^
have the same sign € {—1,1}, then the curve C is regular on B.
In this case, we have the following connection algorithm, which output
a set of segments isotopic to the curve in the domain.
• Compute the points of C fl dB, repeating a point if its multiphcity
is even.
• Sort them by lexicographic order so that x > y: C := {pi,P2, • • •}
• Connect them by pair [pi, P2] > [P3> P4] > • • • of consecutive points in £.
This yields the following subdivision algorithm for a planar implicit curve:

ALGORITHM 2.3. TOPOLOGY OF A PLANAR IMPLICIT CURVE

I N P U T : A box Bo C K^, a curve C defined by the squarefree polynomial equation

f{x,y) =: 0 with / e Q[x,y] and e > 0.
• Compute the x and y critical points of f{x,y) = 0.
• C = {Bo}
• While £ is not empty,
— Choose and remove a domain B of C;
— Test if there is a unique critical point, which is not singular in B;
— If it is the case, compute the topology of C in B,
— else if | B | > e, subdivide the box into subdomains and add them to
C.
— otherwise connect the center of the box to the points of C n dB.
OUTPUT: a set of points and a set of (smooth) arcs connecting these points.

PROPOSITION 2.3. For e > 0 small enough, the algorithm 2.3 com-
putes a graph of points which is isotopic to the curve COB.
The subdivision level (controlled by e) can be improved significantly
by analyzing the number of real branches at a singular point of the curve
C, using topological degree computation. The solver used to compute these
singular points should be able to isolate them from other extremal points
of / (real roots of dxf = dyf = 0) to guaranty the topology of the curve.
 BERNARD MOURRAIN ET AL. 99

Example c:

This curve is the discriminant curve of a bivariate system with few mono-
mials used in [7] to give a counter-example to Kushnirenko's conjecture.
It is of degree 47 in x and y, and the maximal bit size of its coefficient
is of order 300. It takes less that 10 s. (on an Intel M 2.00GHz i686)
to compute the topological graph, by rounding up and down the Bernstein
coefficients of the polynomial to the nearest double machine precision num-
bers, applying the subdivision techniques on the enveloping pol5Tiomials.
We observe a very tiny domain, which at first sight looks like a cusp point,
but which contains in reality, 3 cusps points and 3 crossing points. The
central region near these cusp points is the region where counter-examples
have been found.
Example d: This ciu-ve is the projection onto the {x, y) plane
of the curve of points with tangent parallel to the
-z-direction for a surface of degree 4. It is defined by
the equation of degree 12, and has 4 real cusps and
2 real crossing points. The size of the coefficients
is small and the topological graph is computed in
less than 1 s. It defines 4 connected regions in the
plane, one of these being very small and difficult to
' see on the picture.
2.3. Subdivision approach for space curves. In this section, we
consider a curve C of E^. We suppose that I{C) = (/i, /a, • • •, /*;) and for
two polynomials/(x,y,z),g(x,y,z) £ 7(C), we define t = v{f)/\S7{9)- We
are interested in the topology of C in a box B = [ao, ho] x [a\, 6i] x [a2,62]-
Similar to the 2D case, we can represent / , g and each component of t in the
Bernstein basis for the domain B. As we will see, the sign changes of the
resulting Bernstein coefficients will make it possible to test the regularity
of the curve with minimal effort.
Here is a criteria of regularity of space curves which allows us to deduce
the topology of C in the domain:
PROPOSITION 2.4. [17] Let C he a 3D spatial curve defined by f = 0
and g = 0. If
100 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

• tx{x) ^ 0 on B, and
• dyh ^ 0 on z-faces, and dzh ^ 0 and it has the same sign on both
y-faces of B, for h = f or h = g,

then the topology of C is uniquely determined from the points C D dB.

A similar criterion applies by symmetry, exchanging the roles of the x,
y, z coordinates. If one of these criteria applies with tj(x) ^ 0 on S (for
i = x,y, z), we will say that C is i-regular on B.
Prom a practical point of view, the test ti{x) 7^ 0 or di{h) / 0 for
i = x,y or z, h = f OT g can be replaced by the stronger condition that
their coefficients in the Bernstein basis of B have a constant sign, which is
straightforward to check. Similarly, such a property on the faces of B is
also direct to check, since the coefficients of a polynomial on a facet form
a subset of the coefficients of this polynomial in the box.
In addition to these tests, we also test whether both surfaces penetrate
the cell, since a point on the curve must lie on both surfaces. This test
could be done by looking at the sign changes of the Bernstein coefficients
of the surfaces with respect to that cell. If no sign change occurs, we can
rule out the possibility that the cell contains any portion of the curve C,
and thus terminate the subdivision early. In this case, we will also say that
the cell is regular.
This regularity criterion is sufficient for us to uniquely construct the
topological graph g oiC within B. Without loss of generaUty, we suppose
that the curve C is x-regular in B. Hence, there is no singularity of C in
B. Furthermore, this also guarantees that there is no 'turning-back' of the
curve tangent along x-direction, so the mapping of C onto the x axis is in-
jective. Intuitively, the mapped curve should be a series of non-overlapping
line segments, whose ends correspond to the intersections between the curve
C and the cell, and such mapping is injective.
This property leads us to a unique way to connect those intersection
points, once they are computed in order to obtain a graph representing the
topology of C, similar to the 2D method.
In order to apply this algorithm, we need to compute the points of
C n B, that is to solve a bivariate system of each facet of B. This is
performed by applying the algorithm described in Section 1.3.
The special treatment of points of C on an edge of B or where C
is tangent to a face requires the computation of tangency information at
these points. This is performed by evaluating the derivatives of the defining
equations of C at these points.
Collecting these properties, we have the following subdivision algo-
rithm, which subdivides the domain B until some size e, if the curve is not
regular in B. It relies on a bivariate solver, for computing the intersection
of the curve with the faces of the box.
 BERNARD MOURRAIN ET AL. 101

ALGORITHM 2.4. TOPOLOGY OF A SPACE CURVE

INPUT: a curve C defined by equations /i = 0, /2 = Q,... ,fk = 0 and a domain

B = [ao,6o] X [ai,bi] x [02,62] C K^ and e > 0.
• For 1 <i < j < k,
— compute the Bernstein coefficients of the x,y,z coordinates of V / i A
Vfj in B
— check that they are of the same sign for one of the coordinates
(say x ) ;
— check the x-regularity condition on the facets of B.
• If such a pair (i,j) satisfying the previous regularity condition exists,
— Compute the points of C n dB and connect them.
• else if | B | > e, subdivide B and proceed recursively on each subdomain.
o
• otherwise find a point p in B, compute the point COdB and connect them
to p.
O U T P U T : a set of points p and a set of arcs connecting them.

As in the 2D case, we have the following "convergence" property:

PROPOSITION 2.5. For e > 0 small enough, the graph of points and
arcs computed by the algorithm has the same topology as C (iB.
Example. This curve is defined by fix,y,z) = 0,
dzf{x,y,z) •• 0 where
f{x,y,z)^4{T^x^-y^)(T^y'' • Z^){T^ Z^ - l2)

-il + 2T)(x^-{-y^ + z^-iy,

r = i ( l + V5),
of degree 6 is defining the surface S called Earth's
Sextic. This curve, called the polar curve of S in
the z direction, is of degree 30 = 6 x 5. We compute
the topology by approximating the coefficients of /
and dzf by floating point numbers.
2.4. Subdivision method for surfaces. In this section, we consider
a surface S defined by the equation f(x,y,z) = 0, with / G Q[x,y,z].
We assume that / is squarefree, that is / has no irreducible factors of
multipUcity > 2. For more details, see [1].
Unlike in the 2 dimensional case, the topology of the singular locus
and the way to smooth locus is attached to it can be really complicated.
Topologically we can characterize the topological situation as follows:
• Near a 2-dimensional stratum the topology is the same as a hyper-
plane.
• Near a 1-dimensional stratum the topology is the same as a cylinder
on a singular planar curve.
102 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

• Near a 0-dimensional stratum the topology is the same as a cone

with base the surface intersect a small ball containing the 0-dimen-
sional stratum.
Moreover, we know only one of these three situations can and will happen
locally. So we just have to design a solution for each one of the above
three cases.
For efhciency reasons the criteria we have designed work for situations
more general than the limit three cases. The 2-dimensional strata criterion
can succeed even with several hyperplanes in the box not only just one. For
the 1-dimensional strata we can triangulate even when some patches of the
2-dimensional strata lie in the box even though they are disconnected from
the singular locus (in the box). In the case of the 0-dimensional strata
we need to have the exact topology in the box. Of course the criteria
eventually succeed if the box is small enough. We now describe each one
of these criteria and the matching connection algorithm.
Let B = [a, b] x[c,d]x [e, f] C R^ and a surface S C B. The boundary
of B is denoted hereafter by dB. The x-faces (resp. y, -z-facet) of B are
the planar domains of the boundary dB of B, orthogonal to the direction
X (resp. y, z).
DEFINITION 2.5. The surface S is z-regular (resp. y, z-regular) in the
domain B if,
• S has no tangent line parallel to the z-direction (reps, x, y-direc-
tion),
• Sr\F is regular, for F a z-facet (resp. x, y-facet) of B.
We will say that <S is regular in B if it is regular in B for the direction
X, y or z. Here again, if a point p S <S is singular, then any line through
this point is tangent to >S at p. Thus a surface in B which is regular is
also smooth.
PROPOSITION 2.6. [1] IfS is regular in B, then its topology is uniquely
determined by its intersection with the edges of B.
As in the 2D case, simple tests of regularity can be derived from the
representation of / in the Bernstein basis.
PROPOSITION 2.7. Let {u,v,w) be any permutation of{x,y,z). Sup-
pose thai the coefficients of duf in the Bernstein basis of B have the same
sign S {—1,1} and that the coefficients ofdvf or dy,f on the u-facets of B
are also of the same sign € {—1,1}. Then C is regular in B.
This criterion implies that in the valid cells, the derivative of / in
one direction is of constant sign and on the two faces transverse to this
direction, another derivative is of constant sign. This may be difficult to
obtain, when a point of the surface where two derivatives vanish is on (or
near) the boundary of the cell. A situation where duf 7^ 0 but where both
derivative dyf, d^f are not of constant sign on a n-facet F oi B, can be
handled by applying recursively the 2D algorithm of the facet F.
For handling the singularities of an algebraic surface f{x, y, z) — 0,
we exploit the properties of the polar variety '^ziS) defined by / ( x , y, z) =
 BERNARD MOURRAIN ET AL. 103

0,dzf{x,y,z) = 0. For that purpose, we apply the implicit space curve

algorithm of Section 2.3.
ALGORITHM 2.6. TOPOLOGY OF A SURFACE

I N P U T : a surface S defined by a squarefree equation f{x,y,z) = 0, a domain

-Bo = [ao,bo] X [ai,6i] x [ 0 2 , M C R^ and e > 0.
• Compute generators g\,...,gk of the ideal I(<lz{S)) = {f,dzf) :
Jif,d.f).
• Compute the Bernstein coefficients of the / and gi in the Bernstein basis
of B := Bo-
• If 5 is regular in B, compute its topological structure.
• Else if the polar variety €z{S) is regular and connected in B, compute the
topological structure of <S n dB by Algorithm 2.3 on each facet of B
• Else if |JB| > e, subdivide the box B and proceed recursively on each sub-
domain.
o
• Otherwise find a point p in B, compute the topological structure of . S n ^ B
by Algorithm 2.3 and its link over p.
OUTPUT: a set of points, arcs and patches and adjacency re-
lations describing the topology of 5 n Bo arcs connecting them.

As in the previous case, for e > 0 small enough, the output of this
algorithm is topologically equivalent to <$ PI B.
Example.
Here is the Barth sextic surface whose polar variety
has been computed in section 2.3. This surface of
degree 6 has the maximum number of isolated sin-
gularities for this degree, that is 65. These singular
points are also singular points of its polar variety.

3. The design of the library. As illustrated in previous sections,

various internal representations (eg. dense Bernstein basis or sparse repre-
sentations for polynomials) for the abstract data^types (eg. vectors, poly-
nomials) are required, together with algorithm specializations on these rep-
resentations. The library SYNAPS makes it possible to define parametrized
but efficient data structures for fundamental algebraic objects such as vec-
tors, matrices, monomials and polsoiomials . . . which can easily be used in
the construction of more elaborated algorithms.
We pay special attention to genericity in designing structures for which
effectiveness can be maintained. Thanks to the parametrization of the code
using templates and to the control of their instantiations using traits and
template expressions [14], they offer generic programming without losing
effectiveness. We need to combine generic implementations, which allow to
104 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

reuse code on different types of data representation, with specialized imple-

mentations tuned to specific critical operations on some of these data struc-
tures. This is typically the case if we want to use external or third-party
libraries, such as LAPACK (Fortran library for numerical Unear algebra),
GMP (C Mbrary for extended arithmetics), or MPSOLVE (C univariate solver
implementation, using extended multiprecision arithmetic). For instance
LAPACK routines should coexist with generic ones on matrices. In order
to optimize the implementation, while avoiding rewriting several times the
same code, we need to consider hierarchical classes of data-structures and
a way to implement specializations for each level in this hierarchy. In this
section, we describe the design of the library, which allows such a combina-
tion. Since, it is aimed to be used in dedicated applications, we should also
take care of techniques to embed the library functionalities into an external
apphcation. In this section, we also describe the method, that we adopt
to build dynamic modules, combining generic and special code through a
transparent plugin mechanism. This approach is illustrated by the con-
nection with the geometric modeler AXEL, which uses SYNAPS library for
geometric computation on algebraic curves and surfaces.

3.1. The view hierarchy. A mathematical object can have different

logic representations. Moreover, the fact that a data structure represents
a particular mathematical object is a meaningful information in generic
programming in order to dispatch algorithms. Template functions can be
defined for a formal type parameter T, assuming a semantically and syn-
tactically well defined interface. This interface is called a concept. When
an instantiation on T of a generic algorithm makes sense, we say that T is
a representation of the concept assumed by the algorithm.
In order to dispatch a function call to the right generic version, people
have to write some tjfpe declarations associated to T, specifying what kind
of concept it implements. This can be done using traits, which are types
wearing explicit information about types. In our development, we choose
another solution, based on a view mechanism.
The idea behind a view is to reify a concept in order to dispatch algo-
rithms by concepts using function overloading: when a data type T imple-
ments a concept C, we say that it can be seen as a C. A view is then defined
as a generic type parametrized by a formal type T that is assumed to imple-
ment the concept associated to the view. As an example, Polynomial<T>
is a view associated to the Poljmomial concept, reifying the information
that T implements the required interface for the Polynomial concept. Let's
assume we write a generic function f expecting a type implementing the
Polynomial concept, written in the namespace POLYNOMIAL:
namespace POLYNOMIAL {
// generic implwoentation of f assvuning the Polynomial concept
template<typename Polynomial> void f(const Polynomial & p)
{ ... ; g(p) ; ... }
} ;
 BERNARD MOURRAIN E T AL. 105

and the following function:

template<class PolynoniialType> struct Polynomial {} ;

template<class PolynomialType> void f(const Polynomial<PolynomialType> fe p)

{
using namespace POLYNOMIAL ; f ((const PolynranieLlTypefc)p) ;
}
Then, by specifying that a data type T inherits from Polynomial<T>,
the preceding code ensures that the generic function POLYNOMIAL: :f will
be used when there is no version of f written for T. A polynomial having
dense representation can allow a better implementation of f, in this case,
we can say that it implements the DensePolynomial concept which is a
sub-concept of Polynomial.
To reify the idea of sub-concept, we define the DensePolynoinial<T>
view as inheriting from Polynomial<T>. This way, a type T inheriting from
DensePolynomial<T>, will have an implementation of f corresponding to
the first one available, seeing the type successively as a T, a DensePolyno-
mial<T> and a Pol3momial<T>. This mechanism also allows to specialize
a function for a given representation (eg. representation in the Bernstein
basis b e z i e r : :repld<C>), by providing the function
namespace b e z i e r { template<class C> void f ( c o n s t bezier::repld<C> ftp);}

3.2. Interfacing with an interactive environment. The preced-

ing view mechanism is in fact a way to associate functions to objects con-
sidering a hierarchy of concepts. It is actually closely linked to another
kind of design by virtual classes, which we used to make the library collab-
orate with the modeler AXEL. In this framework, the external tool defines
a virtual hierarchy of objects and each interface defines a set of member
functions inheriting one from another, and, corresponds to a concept, each
inheritance relationship being seen as a sub-concept declaration. We de-
scribe here the procedure used to link the static view hierarchy with the
dynamic virtual hierarchy.
The interfaces I are implemented as classes with virtual functions.
In order to automatically construct a class which implements the set of
functions for the interface I and the representation R, we define a wrapper
class W<I,R>-
To choose, at compilation time (see section 3.1), the most specialized
function for a given data type, we define a view class V<I ,R>, where R is a
representation implementing the concept associated to the interface I.
Suppose that we have defined a function a, which associates to a type
T, its interface class I (implemented by the traits class i n t e r f aceof <T>),
and, that we also have defined /?(T) operating on types, which computes
the base class of T (accessible as T: :base_t). Then for a given represen-
tation tj^e R, the following inheritance relationship of classes defined by
induction as:
R -> W<a(R),R>;W<I,R> -^ V<I,R>;V<I,R> -^ W</?(I) ,R>;V<0,R> -» a(R)
106 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

yields the following inheritance chain:

R -* W<Io,R> -^ V<Io,R> -^ W<Ii,R> - » . . . - * / , ,
where /Q is the interface of R and 7 i , . . . , 7*, the upper level of classes in
the hierarchy. This allows us to define the specialized implementations of
this hierarchy level by level and to automatically choose most optimized
functions.
Let's assume that we have defined two length generic functions for
the ParametricCurve and BSplineCurve concepts, that is, for the types
V<IParametricCurve,R> and V<IBSplinGCurve,R>:
template<class C, c l a s s R> C lengthCconst V<IParametricCurve, R>
k c, const C k eps)
{
using namespace PARAMETRICCURVE ; r e t u r n l e n g t h C c . r e p O , eps) ;
}

teiaplate<class C, c l a s s R> C lengthCconst V<IBSplineCurve, K>

k c, const C k eps)
{
using namespace RATIONALCURVE ; return lengthCc.repO, eps) ;
}
We define now the implementation of the interface, using the wrapper
class associated to the IParametricCurve:
teaplate<class R> struct W<IParaaetricCurve,R> : V<IParametricCurve,R>
{
double lengthCdouble eps) const { return lengthCrepC), eps) ; }
} ;
In order to insert a specific representation of rational curves MPolDse-
RationalCurve<C> in the hierarchy, we use the following construction
which specifies its interface and its implementation:
template<claiss C> s t r u c t iBterfaceof< MPolDseRationalCurve<C> >
{
typedef IRationeilCurve T ;
} ;

teHq>late<class C> s t r u c t HPolDseRationalCurve

: W< IRationalCurve,MPol0seRationalChirve<C> >
{
typedef IRationalSurface base_t ;

} ;
This technique allows us to easily embed the Ubrary into an external
interactive environment such as the geometric modeler AXEL, as we illus-
trate it now. This yields a plugin ShapePlugin, which will be compiled
separately and loaded dynamicEoly into the modeler. To build this plugin,
we first furnish a factory fi-om a list of tjqjes:
typedef type::gentlist<
MPolDseRationalCurve<double>,
MPolDseRationalSurface<double>,

>::T TypeList ;
 BERNARD MOURRAIN E T AL. 107

void ShapePlugin::init(void)
{
factory = new WShapGFactory<TypeList> ;
}
This factory allows to maintain a map from the interfaces of external tools
(as in AXEL) to SYNAPS interfaces, which is enriched each time an object
is created:
void ShapePlugin::create(QRationalCurve * c)
{
IShape * shape = f a c t o r y - > i n t e r f a c e ( " I R a t i o n a l C u r v e " ) ;
IRationalCurve * r e ;
IPsurametricCurve * pc ;
r e = dynamic_cast<IRationsLLCurve *> (shape) ;
pc = dynamic_cast<IParametricCurve *>(shape) ;
rc->setEquations(c->pw(),c->px(),c->py(),"t") ;
pc->setRange(c->siBin() ,c->smax()) ;
m a p . i n s e r t ( c , shape) ;
}
The application forwards function calls over its objects to calls of functions
in the plugin, following the same pattern. This approach allows to make
code designed for different purposes, occasionally in different languages,
coexist.
4. Conclusion. SYNAPS is an open source C++ library for sjnnbolic
and numeric computations, that provides algebraic algorithms and data
structures for manipulating poljniomials, for solving polsoiomial equations
and computing with real algebraic numbers. The wide range of the alge-
braic operations that are implemented in the hbrary, as well as the design
of it, based on the view mechanism, allows us to tune the algebraic op-
erations in order to tackle difficult problems in non-linear computational
geometry, such as the computation of the topology of real plane and space
curves and surfaces. Last but not least, the design of the library permits
the development of algebraic plugins that can be used as algebraic primi-
tives to external software packages specialized to geometric modeling, such
as AXEL.
SYNAPS is a continuous effort to combine symbolic and numeric tech-
niques in algebraic computing under the generic programming paradigm.
The library contains more than 200,000 hnes of codes. In a short term,
we plan to structure it into autonomous subpackages and to extend its
capability to build dedicated plugins for external interactive tools, while
improving the implementations for polynomials.
Acknowledgments. B. Mourrain, J.P. Pavone, and J. Wintz are par-
tially supported by the European projects ACS (Algorithms for Complex
Shapes, 1ST FET Open 006413), AIM@Shape (1ST Network of ExceUence
506766) and the ANR project GECKO (Geometry and Complexity).
108 SYNAPS: A LIBRARY FOR SYMBOLIC-NUMERIC COMPUTATION

REFERENCES

[1] L. ALBERTI, G . COMTE, AND B . MOURRAIN. Meshing implicit algebraic surfaces:

the smooth case. In L.L. Schumaker M. Ma«hlen, K. Morken (editors), Mathe-
matical Methods for Curves and Surfaces: Tromso'04, pages 11-26. Nashboro,
2005.
[2] W . AuziNGER AND H. J. STETTER. An elimination algorithm for the computation
of all zeros of a system of multivariate polynomial equations. In Proc. Intern.
Conf. on Numerical Math., Vol. 8 6 of Int. Series of Nuraerical Math, pages
12-30. Birkhauser Verlag, 1988.
[3] S. BASU, R . POLLACK, AND M . - F . ROY. Algorithms in Real Algebraic Geometry.
Springer-Verlag, Berlin, 2003. ISBN 3-540-00973-6.
[4] D. BINI. Numerical computation of polynomial zeros by means of aberth's method.
Numerical Algorithms, 13, 1996.
[5] L. BuSE, H. KHALIL, AND B . MOURRAIN. Resultant-based method for plane curves
intersection problems. In Proc. of the conf. Computer Algebra in Scientifc
Computing, Vol. 3 7 1 8 of LNCS, pages 75-92. Springer, 2005.
[6] D. Cox, J. LITTLE, AND D . O'SHBA. Ideals, Varieties, and Algorithms, 2nd edition.
Undergraduate Texts in Mathematics. Springer, New York, 1997.
[7] A. DICKENSTEIN, M.J. R O J A S , K . RUSEKZ, AND J. SHIHX. Extremal real algebraic
geometry and A-discriminants. Preprint, 2007.
[8] A. ElGENWiLLiG, V. SHARMA, AND C. K. Y A P . Almost tight recursion tree bounds
for t h e descartes method. In ISSAC '06: Proceedings of the 2006 Interna-
tional Symposium on Symbolic and Algebraic Computation, pages 71-78. ACM
Press. New York, NY, USA, 2006.
[9] I.Z. EMIMS AND E . P . TsiGARlDAS. Univariate polynomial real root isolation: Con-
tinued fractions revisited. In 14th Annual European Symposium on Algo-
rithms, Vol. 4 1 6 8 of LNCS, pages 817-828. Springer, 2006.
[10] I.Z. EMIRIS, B . MOURRAIN, AND E . P . TSIGARIDAS. Real algebraic numbers:
Complexity analysis eind experimentations. In P. Hertling, C M . Hoff-
mann, W. Luther, N. Revol (editors). Reliable Implementation of Real
Number Algorithms: Theory and Practice, LNCS, Springer-Verlag, 2007
(to appear).
[11] G. FARIN. Curves and surfaces for computer aided geometric design: A practical
guide. Comp. Science and Sci. Computing. Acad. Press, 1990.
[12] M.S. FLOATER. On zero curves of bivariate polynomials. Journal Advances in
Computational Mathematics, 5(1):399-415, 1996.
[13] J . C . FAUGERE AND F . ROUILLIER, FGB/RS Grobner basis computation and real
root isolation, h t t p : / / f g b r s . l i p 6 . f r / s a l s a / S o f t w a r e /
[14] E R I C H GAMMA, RICHARD H E L M , R A L P H JOHNSON, AND J O H N VLISSIDES. Design
.patterns: Abstraction and reuse of object-oriented design. Lecture Notes in
Computer Science, 707:406-431, 1993.
[15] TAKAYUKI G U N J I , SUNYOUNG K I M , MASAKAZU KOJIMA, AKIKO TAKEDA, KATSUKI
FUJISAWA, AND TOMOHIKO MIZUTANI, PHOM - A Polyhedral Homotopy Con-
tinuation Method for Polynomial Systems, Research Report B-386, Dept. of
Mathematical and Computing Sciences, Tokyo Institute of Technology, Me-
guro, Tokyo 152-8552, Japan, December 2002,
[16] V. KARAMCHETI, C . L I , I. PECHTCHANSKI, AND C . Y A P . A CORE library for robust
numeric and geometric computation. In 15th ACM Sym,p. on Computational
Geometry, 1999.
[17] CHEN LIANG, B . MOURRAIN, AND J . P . PAVONE. Subdivision methods for 2d and
3d implicit curvra. In Computational Methods for Algebraic Spline Surfaces.
Springer-Verlag, 2006 (to appear).
[18] T . LICKTEIG AND M . - F . ROY. Sylvester-habicht sequences and fast cauchy index
computations. J. of Symbolic Computation, 31:315-341, 2001.
 BERNARD MOURRAIN E T AL. 109

[19] H. LoMBARDi, M.-F. ROY, AND M . SAFEY EL DIN. New structure theorems for
subresultants. J. of Symbolic Computation, 29:663-690, 2000. Special Issue
Symbolic Computation in Algebra, Analysis, and Geometry.
[20] V.J. MiLENKOVIC AND E. SACKS. An approximate arrangement algorithm for
semi-algebraic curves. Proceedings of the 22th Annual Symposium on Com-
putational Geometry, ACM, pages 237-245, June 2006.
[21] B. MOURRAIN. Computing isolated polynomial roots by matrix methods. J. of
Symbolic Computation, Special Issue on Symbolic-Numeric Algebra for Poly-
nomials, 26(6):715-738, Dec. 1998.
[22] B. MOURRAIN. A new criterion for normal form algorithms. In M. Fossorier,
H. Imai, Shu Lin, and A. Poli, editors, Proc. AAECC, Vol. 1 7 1 9 of LNCS,
pages 430-443. Springer, Berlin, 1999.
[23] B. MOURRAIN. Pythagore's dilemma, Symbolic-Numeric Computation and the
Border Basis Method, pages 223-243. Mathematics and Visualisation.
Birkhauser, 2006.
[24] B. MoURRAIN AND J . - P . PAVONE. Subdivision methods for solving polynomial
equations. Technical Report 5658, INRIA Sophia-Antipolis, 2005.
[25] B. MOURRAIN AND P H . TREBUCHET. Solving projective complete intersection
faster. In C. Traverso, editor, Proc. Intern. Symp. on Symbolic and Alge-
braic Computation, pages 231-238. New-York, ACM Press., 2000.
[26] B. MouRRAiN AND P H . TREBUCHET. Generalised normal forms and polynomial
system solving. In M. Kauers, editor, Proc. Intern. Symp. on Symbolic and
Algebraic Computation, pages 253-260. New-York, ACM Press., 2005.
[27] TAKESHI OGITA, SHIN'ICHI OISHI. Fast Inclusion of Interval Matrix Multiplication.
Reliable Computing 11 (3): 191-205 (2005)
[28] J . P . PAVONE. Auto-intersection de surfaces pamatries reelles. P h D thesis, Uni-
versite de Nice Sophia-Antipolis, 2004.
[29] F. ROUILLIER AND P . ZIMMERMANN. Efficient isolation of a polynomial real roots.
Journal of Computaticmal and Applied Mathematics, 162(l):33-50, 2003.
[30] S. R U M P , Computational error bounds for multiple or nearly multiple eigenvalues.
Linear Algebra and its Applications, 3 2 4 (2001), pp. 209-226.
[31] E.C. SHERBROOKE AND N . M . PATRIKALAKIS. Computation of the solutions
of nonlinear polynomial systems. Comput. Aided Geom. Design, 10(5):
379-405, 1993.
[32] H.J. STETTER. Numerical polynomial algebra. Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, PA, 2004.
[33] P H . TREBUCHET. Vers une rdsolution stable et rapide des Equations algibrigues.
P h D thesis, Universite Pierre et Marie Curie, 2002.
[34] J. VERSCHBLDE. Algorithm 795: PHCpack: A general-purpose solver for polyno-
mial systems by homotopy continuation. ACM IVansactions on Mathematical
Software 25(2):251-276, 1999
[35] J. VON ZUR GATHEN AND J. GERHARD. Modem computer algebra. Cambridge
- University Press, New York, 1999.
[36] C.K. Y A P . Fundamental Problems in Algorithmic Algebra. Oxford University
Press, 2000. ftp:/Preliminary/cs.nyu.edu/pub/local/yap/algebra^bk.html.
 TROPICAL IMPLICITIZATION A N D
M I X E D FIBER POLYTOPES
BERND STURMFELS* AND JOSEPHINE YUt

A b s t r a c t . The software Trim offers implementations of tropical implicitization and

tropical elimination, as developed by Tevelev and the authors. Given a polynomial map
with generic coefficients, Trim computes the tropical variety of the image. When the
image is a hypersurface, the output is the Newton polytope of the defining polynomial.
Trim can thus be used to compute mixed fiber polytopes, including secondary polytopes.

K e y w o r d s . Elimination theory, fiber polytope, implicitization, mixed volume,

Newton polytope, tropical algebraic geometry, secondary polytope.

A M S ( M O S ) s u b j e c t classifications. 14Q10, 52B20, 52B55, 65D18.

1. I n t r o d u c t i o n . Implicitization is the problem of transforming a

given parametric representation of an algebraic variety into its implicit
representation as the zero set of polynomials. Most algorithms for elimi-
nation and implicitization are based on multivariate resultants or Grobner
bases, but current implementations of these methods are often too slow.
When the variety is a h3q)ersurface, the coefficients of the implicit equation
can also be computed by way of numerical linear algebra [3, 7], provided
the Newton polytope of that implicit equation can be predicted a priori.
The problem of predicting the Newton polytope was recently solved
independently by three sets of authors, namely, by Emiris, Konaxis and
Palios [8], Esterov and Khovanskii [13], and in our joint papers with Tevelev
[18, 19]. A main conclusion of these papers can be summarized as follows:
The Newton polytope of the implicit equation is a mixed fiber polytope.
The first objective of the present article is to explain this conclusion
and to present the software package Trim for computing such mixed fiber
polytopes. The name of our program stands for Tropical Implicitization,
and it underlines our view that the prediction of Newton polytopes is best
understood within the larger context of tropical algebraic geometry. The
general theory of tropical eUmination developed in [18] unifies earlier results
on discriminants [4] and on generic pol5momial maps whose images can have
any codimension [19]. The second objective of this article is to explain the
main results of tropical elimination theory and their implementation in
Trim. Numerous hands-on examples will illustrate the use of the software.
At various places we give precise pointers to [8] and [13], so as to highlight
similarities and differences among the different approaches to the subject.
Our presentation is organized as follows. In Section 2 we start out with
a quick guide to Trim by showing some simple computations. In Section

* University of California, Berkeley, CA 94720 ([email protected]).

tMassachusetts Institute of Technology, Cambridge, MA 02139 (jyu©math.mit.edu).
Ill
112 BERND STURMFELS AND JOSEPHINE YU

3 we explain mixed fiber polytopes. That exposition is self-contained and

may be of independent interest to combinatorialists. In Section 4 we discuss
the computation of mixed fiber polytopes in the context of elimination
theory, and in Section 5 we show how the tropical implicitization problem
is solved in Trim. Theorem 5.1 expresses the Newton polytope of the
impUcit equation as a mixed fiber polytope. In Section 6 we present results
in tropical geometry on which the development of Trim is based, and we
explain various details concerning our algorithms and their implementation.
2. How t o use Trim. The first step is to download Trim from the
following website which contains information for installation in Linux:
http://math.mit.edu/~jyu/TrIm
Trim is a collection of C + + programs which are glued together and in-
tegrated with the external software polynaie [9] using p e r l scripts. The
language p e r l was chosen for ease of interfacing between various programs.
The fundamental problem in tropical implicitization is to compute the
Newton polytope of a hypersurface which is parametrized by Laurent poly-
nomials with sufficiently generic coeflacients. As an example we consider
the following three Laurent polynomials in two unknowns x and y with
sufiiciently generic coefficients ai,a2,as,/3i,/32,/93,71,72,73:
1
u = a i • ^5-T- + a2- X + aa-xy
X'^y'
V = /3i • x'^ + 132 • y +,3 13:
. _
X
..2 1 1
w = 71-2/ + 72 • 1- 73 • - •
xy y
We seek the unique (up to scaling) irreducible polynomial F{u, v, w) which
vanishes on the image of the corresponding morphism (C*)^ -^ C^. Using
our software Trim, the Newton polytope of the polynomial F{u, v, w) can
be computed as follows. We first create a file input with the contents

[x,y]
[x~(-2)*y~(-2) + X + x*y,
x~2 + y + x~(-l).
y-2 + x-(-i)*y-(-l) + y - ( - l ) ]

Here the coefficients are suppressed: they are tacitly assumed to be generic.
We next run a p e r l script using the command . / T r l m . p r l input. The
output produced by this program call is quite long. It includes the fines

VERTICES
19 20
10 92
1 090
 TROPICAL IMPLICITIZATION 113

1 0 0 9
1 6 6 0
1 2 2 8
1 0 6 6
1 2 8 2
1 6 0 6
1 0 0 0
1 9 0 0
1 2 0 9
1 8 2 2

Ignoring the initial 1, this list consists of 13 lattice points in E^, and these
are precisely the vertices of the Newton polytope of F{u, v, w). The above
output format is compatible with the polyhedral software Polymake [9].
We find that the Newton polytope has 10 facets, 21 edges, and 13 vertices.
Further down in the output. Trim prints a list of all lattice points in the
Newton polytope, and it ends by telling us the number of lattice points:

N.LATTICE.POINTS
383

Each of the 383 lattice points {i,j, k) represents a monomial u'^v^w'' which
might occur with non-zero coefficient in the expansion oiF{u, v, w). Hence,
to recover the coefficients of F{u, v, w) we must solve a linear system of 382
equations with 383 unknowns. Interestingly, in this example, 39 of the 383
monomials always have coefficient zero in F{u, v, w). Even when a i , . . . , 73
are completely generic, the number of monomials in F{u, v, w) is only 344.
The command . /Trim. p r l implements a certain algorithm, to be de-
scribed in the next sections, whose input consists of n lattice polytopes
in R"~^ and whose output consists of one lattice polytope in R". In our
example, with n = 3, the input consists of three triangles and the output
consisted of a three-dimensional polytope. These are depicted in Figure 1.
The program also works in higher dimensions but the running time
quickly increases. For instance, consider the hypersurface in C* represented
by the following four Laurent polynomials in x,y,z, written in Trim format:

[x,y,z]
[x*y + z + 1,
x*z + y + 1,
y*2 + X + 1,
x*3 + y*5 + z*7]

It takes Trim a few moments to inform us that the Newton polytope of

this hypersurface has 40 vertices and contains precisely 5026 lattice points.
The /-vector of this four-dimensional polytope equals (40, 111, 103,32).
114 BERND STURMFELS AND JOSEPHINE YU

Input Output
FIG. 1. Tropical implicitization constructs the three-dimensional Newton polytope
of a parametrized surface from the three Newton polygons of the given parametrization.

REMARK 2.1. The examples above may serve as illustrations for the
results in the papers [8] ajid [13]. Emiris, Konaxis and Palios [8] place the
emphasis on computational complexity, they present a precise formula for
plane parametric curves, and they allow for the map to given by rational
functions. Esterov and Khovanskii develop a general theory of polyhedral
ehmination, which parallels the tropical approach in [18], and which in-
cludes implicitization as a very special case. A formula for the leading
coefficients of the implicit equation is given in [8, §4]. This formula is cur-
rently not implemented in Trim but it could be added in a future version.

What distinguishes Trim from the approaches in [8] and [13] is the
command TrCI which computes the tropical variety of a generic complete
intersection. The relevant mathematics will be reviewed in Section 6. This
command is one of the ingredients in the implementation of tropical implic-
itization. The input again consists of m Laurent polynomials in n variables
whose coefficients are tacitly assumed to be generic, or, equivalently, of m
lattice polytopes in n-space. Here it is assumed that m < n. If equality
holds then the program simply computes the mixed volume of the given
polytopes. As an example, delete the last hne from the previous input file:

k,y,z3
[x*y + z + 1,
x*z + y + 1 ,
y*z + X + 1]
 TROPICAL IMPLICITIZATION 115

The command . /TrCI. p r l input computes the mixed volume of the three
given lattice polytopes in R^. Here the given polytopes are triangles. The
last line in the output shows that their mixed volume equals five.
Now repeat the experiment with the input file input as follows:

[x*y + z + 1, x*z + y + 1]

The output is a one-dimensional tropical variety given by five rays in R^:

DIM
1

RAYS
0 -1 -1
0 0 1
1 0 0
0 1 0
- 1 1 1

MAXIMAL.CONES
0
1
2
3
4

MULTIPLICITIES
2
1
1
1
1

Note that the first ray, here indexed by 0, has multiplicity two. This scaling
ensures that the sum of the five RAYS equals the zero vector (0,0,0).
For a more interesting example, let us tropicalize the complete inter-
section of two generic hypersurfaces in C^. We prepare input as follows:

[a,b,c,d,e]
[a*b + b*c + c»d + d*e + a*e + 1,
a*b*c + b*c*d + c*d*e + d*e*a + e*a*b]

When applied to these two polynomials in five unknowns, the command

. / T r C I . p r l input produces a three-dimensional fan in R^. This fan has
26 rays and it has 60 maximal cones. Each maximal cone is the cone over
a triangle or a quadrangle, and it has multipUcity one. The rays are
116 BERND STURMFELS AND JOSEPHINE YU

RAYS
- 1 1 0 0 1
- 1 1 1 - 1 3
0 1 0 0 1
- 1 3 - 1 1 1

The rays are labeled 0 , 1 , . . . , 25, in the order in which they were printed.
The maximal (three-dimensional) cones appear output in the format

MAXIMAL.CONES
0 12 3
0 1 7 10
0 1 12
0 3 4 7
0 3 12
0 7 12

It is instructive to compute the tropical intersection of two generic

hypersurfaces with the same support. For example, consider the input file

[x.y.z]
[ l + x + y + z + x*y + x*z + y*z + x*y*z,
l + x + y + z + x*y + x*z + y*z + x*y*z]

As before, the reader should imagine that the coefficients are generic ra-
tional numbers instead of one's. The tropical complete intersection de-
termined by these two equations consists of the six rays normal to the six
facets of the given three-dimensional cube. The same output would be pro-
duced by Jensen's software GFan [2, 12], which computes arbitrary tropical
varieties, provided we input the two equations with generic coefficients.

3. Mixed fiber polytopes. We now describe the construction of

mixed fiber polytopes. These generalize ordinary fiber polytopes [1], and
hence they generaUze secondary polytopes [10]. The existence of mixed
fiber pol5rtopes was predicted by McDonald [14] and Michiels and Cools [16]
in the context of polynomial systems solving. They were first constructed
by McMuUen [15], and later independently by Esterov and Khovanskii [13].
The presentation in this section is written entirely in the language of
combinatorial geometry, and it should be of independent interest to some
of the readers of Ziegler's text book [20]. There are no polynomials or
varieties in this section, neither classical nor tropical. The connection to
efimination and tropical geometry will be explained in subsequent sections.
Consider a linear map n : MP -^ R9 and a p-dimensional poljd;ope
P CW whose image Q = 7r(P) is a g-dimensional polytope in R^. If a; is
any point in the interior of Q then its fiber 7r~^(a:) fl P is a polytope of
dimension p — q. The fiber polytope is defined as the Minkowski integral
 TROPICAL IMPLICITIZATION 117

S„(P) = [ {Tr-\x) n P) dx. (3.1)

JQ
It was shown in [1] that this integral defines a polytope of dinaension p — q.
The fiber polytope E^ (P) lies in an affine subspace of W which is a par-
allel translate of kernel(7r). Billera and Sturmfels [1] used the notation
S(P, Q) for the fiber polytope, and they showed that its faces are in bijec-
tion with the coherent polyhedral subdivisions of Q which are induced from
the boundary of P. We here prefer the notation S,r(P) over the notation
S(P, Q), so as to highlight the dependence on TC for fixed P and varying TT.
EXAMPLE 3.1. Let p = 3 and take P to be the standard 3-cube

P = conv{(000),(001),(010),(011),(100),(101),(110),(lll)}.

We also set q' = 1 and we fix the linear map

n : E ^ - > R \ iu,v,w) i-^u + 2v + 3w.

Then Q = 7r(P) is the line segment [0,6]. For 0 < a; < 6, each fiber
7r~^(a;) n P is either a triangle, a quadrangle or a pentagon. Since the
fibers have a fixed normal fan over each open segment {i,i + l), we find
5 -i+l 5 -.
ST(P) = J2 {7r-\^)nP)dx = -£{^-^1+ )np).
i=0 "'* i=0 ^

Hence the fiber polygon is really just the Minkowski sum of two triangles,
two quadrangles and two pentagon, and this turns out to be a hexagon:

E , ( P ) = conv{(l, 10,5), (1,4,9), (5,2,9), (11,2,7), (11,8,3), (7,10,3)}.

In the next section we shall demonstrate how Trim can be used to compute
fiber polytopes. The output produced will be the planar hexagon which is
gotten from the coordinates above by applying the Hnear map (w, v, w) —
i »
{w — 3,v + w — 9). Hence Trim produces the following coordinatization:

E„(P) = conv{(2,6), (6,4), (6,2), (4,0), (0,2), (0,4)}. (3.2)

It is no big news to polytope aficionados that the fiber polygon of the 3-

cube is a hexagon. Indeed, by [20, Example 9.8], the fiber polytope obtained
by projecting the p-dimensional cube onto a Une is the permutohedron of
dimension p—l. For p = 3 the vertices of the hexagon 5]^(P) correspond
to the six monotone edge paths on the 3-cube from (000) to (111). D
As a special case of the construction of fiber polytopes we get the
secondary polytopes. Suppose that P is a polytope with n vertices in W
and let A denote the standard (n — l)-simplex in M". There exists a linear
map /9 : M" -* RP such that p{A) = P, and this Unear map is unique if
we prescribe a bijection from the vertices of A onto the vertices of P . The
118 BERND STURMFELS AND JOSEPHINE YU

polytope Sp(A) is called the secondary polytope of P; see [20, Definition

9.9]. Secondary polytopes were first introduced in an algebraic context by
Gel'fand, Kapranov and Zelevinsky [10]. For example, if we take P to be
the 3-dimensional cube as above, then the simplex A is 7-dimensional, and
the secondary polytope Sp(A) is a 4-dimensional polytope with 74 vertices.
These vertices are in bijection with the 74 triangulations of the 3-cube.
A detailed introduction to triangulations and a range of methods for
computing secondary poljrtopes can be found in the forthcoming book [5].
We note that the computation of fiber polytopes can in principle be reduced
to the computation of secondary pols^opes, by means of the formula
E,(p(A)) = p(E,op(A)). (3.3)
Here TT o p is the composition of the following two linear maps of polytopes:
A -^ P ^ Q.
The formula (3.3) appears in [1, Lemma 2.3] and in [20, Exercise 9.6].
The algorithm of Emiris et al. [8, §4] for computing Newton polytopes of
specialized resultants is based on a variant of (3.3). Neither our software
Trim nor the Esterov-Khovanskii construction [13] uses the formula (3.3).
We now come to the main point of this section, namely, the construc-
tion of mixed fiber polytopes. This is primarily due to McMuUen [15], but
was rediscovered in the context of elimination theory by Khovanskii and
Esterov [13, §3]. We fix a linear map TT : E^ —+ R ' as above, but we
now consider a collection of c polytopes P i , . . . , P c in E^. We consider
the Minkowski sum P\ = AiPi + • • • + XcPc where A = ( A i , . . . , Ac) is a
parameter vector of unspecified positive real numbers. We shall assume
that Px is of full dimension p, but we do allow its summands Pi to be
lower-dimensional. The image of Px under the map ir is the g-dimensional
pol5d;ope

7r(PA) = A i - 7 r ( P i ) + --- + Ae-7r(Pe).

The following result concerns the fiber pol5d;ope from P \ onto 7r(Px).
THEOREM 3.2 ([15, 13]). The fiber polytope E^(P\) depends poly-
nomially on the parameter vector A. This polynomial is homogeneous of
degree q + 1- Moreover, there exist unique polytopes Mi^i^...i^ such that

S . ( A i P i + ----fAePc) = Y. Ai'A^--AJ=-Mi,i,...i,. (3.4)

n-l—i-<c=q+l
To appreciate this theorem, it helps to begin with the case c = 1. That
corresponds to scaling the polytopes P and Q above by the same factor A.
This results in the Minkowski integral (3.1) being scaled by the factor A^"*"^.
More generally, the coefficients of the pure powers A]+^ in the expansion
(3.4) are precisely the fiber polytopes of the individual Pj, that is,

-^0,...,0,g+l,0,...,0 = 5]^(Pj)-
 TROPICAL IMPLICITIZATION 119

On the other extreme, we may consider ii = J2 = • • • = *c = 1, which is the

term of interest for ehmination theory. Of course, if all ij 's are equal to one
then the number c of polytopes Pj is one more than the dimension q of the
image of TT. We now assume that this holds, i.e., we assume that c = q + 1.
We define the mixed fiber polytope to be the coefficient of the monomial
A1A2 • • • Ac in the formula (3.4). The mixed fiber pol5^ope is denoted

S^(Pi,P2,...,Pc) := M11...1. (3.5)

The smallest non-trivial case arises when p = 3, c = 2 and g = 1, where we

are projecting two pol5d;opes Pi and P2 in M^. Their mixed fiber poljrtope
with respect to a hnear form TT : M^ —» R^ is the coefficient of A1A2 in

E,(AiPi+A2P2) = A ? - E , ( P i ) + AiA2-S,(Pi,P2) + A i - S , ( P 2 ) .

The following is [18, Example 4.10]. It wiU be revisited in Example 4.2.

EXAMPLE 3.3. Consider the following two tetrahedra in three-space:

Pi = conv{0,3ei,3e2,363} and P2 = conv{0, —2ei, —2e2, —263}.

Their Minkowski sum Pi 4- P2 has 12 vertices, 24 edges and 14 facets. If

we take TT : R^ —» R^ to be the hnear form (w, v,w) i~^ u — 2v + w then the
fiber polytope Iln{Pi + P2) = .^^20 + Mn + M02 is a polygon with ten
vertices. Its summands M20 = ^^(Pi) and M02 = S T ( P 2 ) are quadrangles,
while the mixed fiber polytope M n = E^(Pi,P2) is a hexagon. D
We remark that fiber polj^opes are special instances of mixed fiber
polytopes. Suppose that Pi = P2 = • • • = Pc are all equal to the same
fixed poljd;ope P in MP. Then the fiber polytope S,r(PA) in (3.4) equals

E(AIPI + • • • + AePe) = (Ai-f--- + Ae)^-S,(P).

Hence the fiber pols^tope S^(P) is the mixed fiber polytope i;,r(-P, • • • ^ P )
scaled by a factor of 1/c!. Similarly, any of the coefficients in the expansion
(3.4) can be expressed as mixed fiber polj^opes. Up to scaling, we have

Mi^i^-.-ic — ^ 7 r ( f l ; - • •,Pl,P2: • • • 1-P2; • • • , Pc, • •• ,Pc)-

i\ times 12 times ic times

In the next section we shall explain how mixed fiber polytopes, and hence
also fiber polj^opes and secondary polytopes, can be computed using Trim.
4. Elimination. Let fi,f2,---,fc £ C[xf^,xf^,...,x^^] be Lau-
rent polynomials whose Newton polytopes are P i , P 2 , . . . , P c C R^, and
suppose that the coefficients of the fi are generic. This means that

Ji\X) = / ^ Ci^a • ^ 1 ^ 2 • • • •''p )

aePiOZ"
120 BERND STURMFELS AND JOSEPHINE YU

where the coeflBcients a^a are assumed to be sufficiently generic non-zero

complex numbers. The corresponding variety
X = {u€ {C*r •• hiu) = f2{u) = ••• = fciu) = 0}
is a complete intersection of codimension c in the algebraic torus (C*)^.
We set r = p — c 4-1 and we fix an integer matrix A = (a,j) of format
rxp where the rows of A are assumed to be linearly independent. We also
let Tt -.W —* W~^ be any linear map whose kernel equals the row space of
A. The matrix A induces the following monomial map:

a:(CT->(C*r'=+S (xi,...,Xp)H.(nxf%...,nx;-). (4.1)

Let Y be the closure in (C*)''"'^+^ of the image a{X). Then y is a hyper-

surface, and we are interested in its Newton polytope. By this we mean
the Newton polytope of the irreducible equation of that hypersurface.
THEOREM 4.1 (Khovanskii and Esterov [13]). The Newton polytope
of Y is affinely isomorphic to the mixed fiber polytope Sx(Pi, •. •, Pc)-
A proof of this result using tropical geometry is given in [18]. The
computation of the hypersurface Y from the defining equations / i , . . . , /c
of X is a key problem of elimination theory. Theorem 4.1 offers a tropical
solution to this problem. It predicts the Newton polj^ope of Y. This
information is useful for symbolic-numeric software. Knowing the Newton
polytopes reduces computing the equation of Y to linear algebra.
The numerical mathematics of this Unear algebra problem is interesting
and challenging, as seen in [3] and confirmed by the experiments reported
in [19, §5.2). We hope that our software Trim will eventually be integrated
with software exact linear algebra or numerical linear algebra (e.g. LAPack).
Such a combination would have the potential of becoming a useful tool for
practitioners of non-hnear computational geometry.
In what follows, we demonstrate how Trim computes the Newton poly-
tope of Y and hence the mixed fiber polytope EB^(PI, . . . , Pc). The input
consists of the poljrtiopes Pi,...,Pc and the matrix A. The map TT is tacitly
understood as the map from W onto the cokernel of the transpose of A.
EXAMPLE 4.2. Let p = 3,c = 2 and consider [18, Example 1.3]. Here
the variety X is the curve in (C*)^ defined by the two Laurent polynomials
/ i = aixl + a2xl + asxl + 04 and /2 = fSix^'^ + ^2X2'^ + Psx^'^ + 13A-
We seek to compute the Newton polygon of the image curve Y in (C*)^
where A = ( . j . The curve is written on a file input as follows:

[xl,x2,x3]
[xl-3 + x2-3 + x3-3 + 1, xl~(-2)+x2-(-2)+x3-(-2)+l]
 TROPICAL IMPLICITIZATION 121

We also prepare a second input file A.matrix as follows:

LINEAR.MAP
111
0 12

We now execute the following two commands in Trim:

./TrCI.prl input > fan

./project.prl fan A.matrix

The output we obtain is the Newton polygon of the curve Y:

VERTICES
1 36 0
1 0 36
1 30 12
1 18 12
1 6 24
1 18 24

This hexagon coincides with the hexagon in [18, Examples 1.3 and 4.10].
It is isomorphic to the mixed fiber polytope S,r(Pi, P2) in Example 3.3. D
We may use Trim to compute arbitrary fiber polytopes. For example,
to carry out the computation of Example 3.1, we prepare input as

[x.y.z]
[ l + x + y + z + x*y + x*z + y*z + x*y*z,
l + x + y + z + x*y + x*z + y*z + x*y*z]

and A.matrix as

LIMEAR_MAP
1 1 - 1
2 - 1 0

The two commands above now produce the hexagon in (3.2). Our next
example shows how to compute secondary polytopes using Trim.
EXAMPLE 4.3. Following [20, Example 9.11], we consider the hexagon
with vertices (0,0), (1,1), (2,4), (3,9), (4,16), (5,25). This hexagon is rep-
resented in Trim by the following file A.matrix. The rows of this matrix
span the linear relations among the five non-zero vertices of the hexagon:

LIMEAR.MAP
3 - 3 1 0 0
8 - 6 0 1 0
15 -10 0 0 1
122 BERND STURMFELS AND JOSEPHINE YU

On the file input we take three copies of the standard 5-simplex:

[a.b.c.d.e]
[a+b+c+d+e+1, a+b+c+d+e+1, a+b+c+d+e+1]

Running our two commands, we obtain a 3-dimensional polj^ope with 14

vertices, 21 edges and 9 facets. That polytope is the associahedron [20]. D
We close this section with another application of tropical elimination.
EXAMPLE 4.4. For two subvarieties Xi and X2 of (C*)" we define their
coordinate-wise product Xi * X2 to be the closure of the set of all points
{uivi,..., UnV„) where ( w i , . . . , u„) G Xi and {vi,...,Vn) £ X2. The
expected dimension of Xi * X2 is the sum of the dimensions of Xi and X2,
so we can expect Xi-kX2 to be a hypersurface when dim(Xi) +dim(X2) =
n — 1. Assuming that Xi and X2 are generic complete intersections then
the Newton polytope of that hypersurface can be computed using Trim as
follows. Let p = 2n and define X as the direct product Xi x X2- Then
Xi • X2 is the image of X under the monomial map

a: ( C * ) ^ " - ^ ( C * ) " , {ui,...,Un,Vi,...,Vn) >-* {uiVi,...,UnVn).

Here is an example where Xi and X2 are curves in three-dimensional space

(n = 3). The two input curves are specified on the file input as follows:

[ul,u2,u3, vl,v2,v3]
[ul + u2 + u3 + 1,
ul*u2 + ul*u3 + u2*u3 + u i + u2 + u 3 ,
vl*v2 + vi*v3 + v2*v3 + v i + v2 + v3 + 1,
vl*v2*v3 + vl*v2 + vl*v3 + v2*v3 + v i + v2 + v3]

T h e multiplication m a p a : (C*)^ x (C*)^ —» (C*)^ is specified on A . m a t r i x :

LINEAR_MAP
10 0 10 0
0 10 0 10
0 0 10 0 1

T h e image of X i x X2 under the m a p a is t h e surface X i * X2. We find

t h a t the Newton polytope of this surface has ten vertices and seven facets:

VERTICES
18 4 0
10 8 4
10 8 0
10 0 8
14 8 0
14 0 8
10 4 8
 TROPICAL IMPLICITIZATION 123

18 0 0
10 0 0
18 0 4

FACETS
128 -16 0 0
0 0 0 32
128 0 -16 0
0 32 0 0
0 0 64 0
128 0 0 -16
192 -16 -16 -16

5. Implicitization. Implicitization is a special case of elimination.

Suppose we are given n Laurent poljmomials gi,...,gn in C[ii , . . • , t „ l i ]
which have Newton poljd;opes Qi,..., Qn C R"""^ and whose coefficients
are generic complex numbers. These data defines the morphism

g : ( e ) " - i ^ ( C * r , i ^ (51 ( f ) , . . . , 3 „ _ i W ) . (5.1)

Under mild hypotheses, the closure of the image of 5 is a hypersurface

Y in (C*)". Our problem is to compute the Newton poljdiope of this
hypersurface. A first example of how this is done in Trim was shown in the
beginning of Section 2, and more examples will be featured in this section.
The problem of implicitization is reduced to the ehmination compu-
tation in the previous section as follows. We introduce n new variables
y i , . . . , 2/n and we consider the following n auxiliary Laurent polynomials:

fi{x) = 9iit) - yi, hix) = g-iit) -yi, •••, fn{x) = 9n{t) - yn- (5.2)

Here we set p = 2n — 1 and {x\,..., Xp) = {ti,..., tn-i,yi,. • •, Vn) so as to

match the earlier notation. The subvariety of (C*)P = (C*)""^ x (C*)"
defined by / i , . . . , / „ is a generic complete intersection of codimension n,
namely, it is the graph of the map g. The image of g is obtained by pro-
jecting the variety {/i = • • • = / „ = 0} onto the last n coordinates. This
projection is the monomial map a specified by the n x ^ m a t r i x .4 = (0 / )
where 0 is the n x (n—1) matrix of zeroes and / is the nxn identity matrix.
This shows that we can solve the implicitization problem by doing the
same calculation as in the previous section. Since that calculation is a main
application of Trim, we have hard-wired it in the command . / T r i m . p r l .
Here is an example that illustrates the advantage of using tropical implic-
itization in analyzing parametric surfaces of high degree in three-space.
EXAMPLE 5.1. Consider the parametric surface specified by the input

[x.y]
[x~7*y"2 + x*y + x~2*y-7 + 1,
124 BERND STURMFELS AND JOSEPHINE YU

x*8*y~8 + x"3*y"4 + x*4*y"3 + 1,

x"6*y + x*y"6 + x"3*y~2 + x"2*y"3 + x + y]

Using the technique shown in Section 2, we learn in a few seconds to learn

that the irreducible equation of this surface has degree 90. The command
. / T r l m . p r l input reveals that its Newton polytope has six vertices

1 80 0 0
1 0 45 0
1 0 0 80
1 0 10 80
1 0 0 0
1 28 0 54

This polytope also has six facets, namely four triangles and two quadran-
gles. The expected number of monomials in the implicit equation equals

N_UTTICE_POINTS
62778

At this point the user can make an informed choice as to whether she wishes
to attempt solving for the coefficients using numerical linear algebra. D
Returning to our polyhedral discussion in Section 3, we next give a
conceptual formula for the Newton polytope of the implicit equation as
a mixed fiber polytope. The given input is a Ust of n lattice polytopes
Qii Q2, • • •, <3n in R"""^. Taking the direct product of R"~^ with the space
E " with standard basis {ei, 6 2 , . . . , e„}, we consider the auxiUary polytopes

Qi X {ei} , Q2 X {62} , . . . , Q„ X {e„} C R""^ x R".

These are the Newton polytopes of the equations / i , /2, . . . , / „ in (5.2). We
now define 7r to be the projection onto the first n — 1 coordinates

7 r : R " - l X R " - * R " - 1 , {m, . . . ,Un-l,VuV2, . . . ,Vn) ^ (Ul,...,u„_l).

The following result is an immediate corollary to Theorem 4.1. We propose
that it be named the Fundamental Theorem of Tropical Implicitization.
THEOREM 5.2. The Newton polytope of an irreducible hypersurface in
(C*)" which is parametrically represented by generic Laurent polynomials
with given Newton polytopes Qi,. ••, Qn equals the mixed fiber polytope

Sx(Qi X {ei}, Q2 X {62},..., Q„ X {en}) (5.3)

This theorem is the geometric characterization of the Newton poly-
tope of the implicit equation, and it summarizes the essence of the recent
progress obtained by Emiris, Konaxis and Palios [8], Esterov and Khovan-
skii [13], and Sturmfels, Tevelev and Yu [18, 19]. Our implementation of in
 TROPICAL IMPLICITIZATION 125

Trim computes the mixed fiber poljrtope (5.3) for any given Qi,Q2, • • •, Qn,
and it suggests that the Fundamental Theorem of Tropical Implicitization
will be a tool of considerable practical value for computational algebra.
EXAMPLE 5.3. We consider a threefold in C^ which is parametrically
represented by four trivariate polynomials. On the file input we write

[x,y,z]
[x + y + z + 1,
x~2*z + y"2*x + z"2*y + 1,
x"2*y + y*2*z + z"2*x + 1,
x*y + x*z + y*z + x + y + z]

The Newton polytope of this threefold has the f-vector (8,16,14,6), and it
contains precisely 619 lattice points. The eight vertices among them are

VERTICES
1 15 0 0 0
1 0 6 0 0
1 0 0 0 9
1 0 0 6 0
1 0 0 0 0
1 12 0 0 3
1 9 3 0 0
1 9 0 3 0

This four-dimensional polytope is the mixed fiber polytope (5.3) for the
tetrahedra Qi,Q2,Qs and the octahedron Q4 specified in the file input. D
In (5.1) we assumed that the gi{t) are Laurent polynomials but this
hypothesis can be relaxed to other settings discussed in [8, 13]. In partic-
ular, the Fundamental Theorem of Tropical Imphcitization extends to the
case when the gi{t) are rational functions. Here is how this works in Trim.
EXAMPLE 5.4. Let a i , . . . , as and / 3 i , . . . , ^5 be general complex num-
bers and consider the plane curve which has the rational parametrization

_ ait^ + aat + ^3 , _ Pit* + (32t^ + ^3 ,^ ..

"" - 04*2 + a s ^"^ ^ - /34i2 + /35 • ^^-^^

This curve appears in [8, Example 4.7]. The input for Trim is as follows:

[ t , X. y ]
[ t"3 + t + 1 + x*t"2 + X ,
t " 4 + t " 3 + 1 + y*t"2 + y ]

The equation of the plane curve is gotten by eliminating the unknown

t from the two equations (5.4). Tropical elimination using Trim predicts
that the Newton polygon of that plane curve is the following pentagon:
126 BERND STURMFELS AND JOSEPHINE YU

POINTS
14 2
10 3
12 3
10 0
14 0

This prediction is the correct Newton polygon for generic coefficients aj

and Pj. In particular, generically the pairs of coefficients (a4,a5) and
(A,/?5) in the denominators are distinct. If we assume, however, that
the denominators are the same, then the correct Newton polytope is a
quadrangle inside this pentagon, as computed in [8, Example 5.6]. D

6. Tropical varieties. We now explain the mathematics on which

Trim is based. The key idea is to embed the study of Newton polytopes
into the context of tropical geometry [2, 4, 18, 19]. Let / be any ideal in
the Laurent polynomial ring C[xf ^ , . . . ,x^\ Then its tropical variety is

T(/) = {w €iW : \nw{I) does not contain a monomial}.

Here inu,(/) is the ideal of C[xf ^,... ,Xp ^] which is generated by the w-
initial forms of all elements in / . The set T ( / ) can be given the structure of
a polyhedral fan, for instance, by restricting the Grobner fan of any homog-
enization of / . A point w in T{I) is called regular if it lies in the interior
of a maximal cone in some fan structure on T ( / ) . Every regular point w
naturally comes with a multiplicity m^, which is a positive integer. We
can define the sum of multiplicities of all minimal associate primes
of the initial ideal in^(/). The multipHcities m^ on T(7) are independent
of the fan structure and they satisfy the balancing condition [18, Def. 3.3].
If 7 is a principal ideal, generated by one Laurent polynomial f{x),
then T{I) is the union of all codimension one cones in the normal fan of
the Newton polytope P of f(x). A point w € T(7) is regular if and only
if w supports an edge of P, and m^ is the lattice length of that edge. It
is important to note that the polytope P can be reconstructed uniquely,
up to translation, from the tropical hypersurface T(7) together with its
multiplicities m^. The following Trim example shows how to go back and
forth between the Newton polytope P and its tropical hypersurface T(7).
EXAMPLE 6.1. We write the following polynomial onto the file poly:

[ X, y, z ]
[ X + y + z + x~2*y*2 + x"2*z*2 + y"2*z*2 ]

The command . / T r C I . p r l poly > fan writes the tropical surface defined
by this polynomial onto a file fan. That output file starts out like this:

AMBIENT_DIM
 TROPICAL IMPLICITIZATION 127

DIM
2

The tropical surface consists of 12 two-dimensional cones on 8 rays in R^.

Three of the 12 cones have multiplicity two, while the others have multi-
plicity one. Combinatorially, this surface is the edge graph of the 3-cube.
The Newton polytope P is an octahedron, and it can be recovered from the
data on fan after we place the 3 x 3 identity matrix in the file A.matrix:

LINEAR.MAP
10 0
0 10
0 0 1

Our familiar command

./project.prl fan A.matrix

now reproduces the Newton octahedron P in the familiar format:

VERTICES
12 2 0
10 2 2
10 10
12 0 2
110 0
10 0 1

Note that the edge lengths of P are the multiplicities on the tropical
surface. Q
The implementation of the command p r o j e c t . p r l is based on the
formula given in [19, Theorem 5.2]. See [4, §2] for a more general version.
This rssult translates into an algorithm for Trim which can be described as
follows. Given a generic vector u; G R*^ such that faceu,(P) is a vertex v, the
«*'* coordinate vi is the number of intersections, counted with multiplicities,
of the ray w + R>oei with the tropical variety. Here, the multiplicity of the
intersection with a cone T is the multiplicity of T times the absolute value
of the i*'' coordinate of the primitive normal vector to the cone F.
Intuitively, what we are doing in our software is the following. We
wish to determine the coordinates of the extreme vertex v = faceiu(P) of
a polytope P in a given direction w. Our polytope is placed so that it
lies in the positive orthant and touches all the coordinate hyperplanes. To
compute the i*'' coordinate of the vertex v, we can walk from v toward the
i*'' hyperplane along the edges of P, while keeping track of the edge lengths
128 BERND STURMFELS AND JOSEPHINE YU

in the z*'' direction. A systematic way to carry out the walk is to follow the
edges whose inner normal cone intersect the ray to + R>oei. Recall that the
multiplicity of a codimension one normal cone is the lattice length of the
corresponding edge. Using this subroutine for computing extreme vertices,
the whole polytope is now constructed using the method of Huggins [11].
To compute the tropical variety T(I) for an arbitrary ideal I one can
use the Grobner-based software GFan due to Jensen [2, 12]. Our polyhedral
software Trim performs the same computation faster when the generators
of / are Laurent polynomials fi,f2,---,fc that are generic relative to their
Newton polytopes Pi, P2, • • •, Pc- It implements the following combinato-
rial formula for the tropical variety T{I) of the complete intersection I.
THEOREM 6.2. The tropical variety T{I) is supported on a subfan of
the normal fan of the Minkowski sum Yli=i Pi- -^ point w is in T{I) if and
only if the polytope tacGwi^j^ J Pj) has dimension > \J\forJC {l,...,c}.
The multiplicity of T{I) at a regular point w is the mixed volume

ruw = mixedvolume(faceTO(Pi),faceTO(^2), • • • >fa.ceTO(-Pc)), (6.1)

where we normalize volume respect to the affine lattice parallel to Yljej Pj-
For a proof of this theorem see [18, §4]. We already saw some examples
in the second half of Section 2. Here is one more such illustration:
EXAMPLE 6.3. We consider the generic complete intersection of codi-
mension three in six-dimensional space (C*)^ given by the polynomials

[a,b,c,d,e,f]
[a*b*c*d*e*f + a + b + c + d + e + f + l ,
a*b + b*c + c*d + d*e + e*f + f * a ,
a + b + c + d + e + f + 1]

The application of . /TrCI. p r l to this input internally constructs the cor-

responding six-dimensional polytope Pi + P2 + Pa- It lists all facets and
their normal vectors, it generates all three-dimensional cones in the nor-
mal fan, and it picks a representative vector w in the relative interior of
each such cone. The three polytopes i&ce^{Pi), face^(P2) and i&cewiPa)
are translated to lie in the same three-dimensional space, and their mixed
volume m.^ is computed. If m^ is positive then Trim outputs the rays of
that cone and the mixed volume m.^,. The output reveals that this tropical
threefold in R® consists of 117 three-dimensional cones on 22 rays. D
In our implementation of Theorem 6.2 in Trim, the Minkowski sum
is computed using the iB4e library [11], enumerating the d-dimensional
cones is done by Polymake [9], the mixed volumes are computed using the
mixed volume l i b r a r y [6], and integer linear algebra for lattice indices
is done using NTL [17]. What oversees all the computations is the p e r l
script TrCI . p r l . The output format is consistent with that of the current
 TROPICAL IMPLICITIZATION 129

version of Gf an [12] and is intended to interface with the polyhedral software

Polymake [9] when it supports polyhedral complexes and fans in the future.
Elimination theory is concerned with computing the image of an alge-
braic variety whose ideal / we know under a morphism a : (C*)P -> (C*)'".
We write /? for the corresponding homomorphism from the Laurent poly-
nomial ring in r unknowns to the Laurent polynomial ring in p unknowns.
Then J = (3"^(I) is the ideal of the image variety, and, ideally, we would
hke to find generators for J. That problem is too hard, and what we do
instead is to apply tropical elimination theory as follows. We assume that
a is a monomial map, specified by an r x p integer matrix A = (ay) as
in (4.1). Rather than computing the variety of J from the variety of / , we
instead compute the tropical variety T{J) from the tropical variety T{I).
This is done by the following theorem which characterizes the multiplicities.
THEOREM 6.4. The tropical variety T{J) equals the image of T{I)
under the linear map A. If the monomial map a induces a generically
finite morphism of degree S from the variety of I onto the variety of J then
the multiplicity ofT{J) at a regular point w is computed by the formula

m-u, = -r-^rn^- index (L„ n F : A(L„ n Z^)). (6.2)

The sum is over all points v in T{I) with Av ^ w. We assume that the
number of these points is finite, they are all regular in T{I), and L„ is the
linear span of a neighborhood of v in T ( / ) , and similarly for w G T{J).
The formula (6.2) can be regarded as a push-forward formula in in-
tersection theory on toric varieties, and it constitutes the workhorse inside
the Trim command p r o j e c t . p r l . When the tropical variety T{J) has
codimension one, then that tropical hypersurface determines the Newton
poljrtope of the generator of J. The transformation from tropical hyper-
surface to mixed fiber pol5rtope was behind all our earlier examples. That
transformation was shown explicitly for an octahedron in Example 6.1.
In all our examples so far, the ideal / was tacitly assumed to be a
generic complete intersection, and Theorem 6.2 was used to determine the
tropical variety T{I) and the multiphcities m„. In other words, the com-
mand TrCI furnished the ingredients for the formula (6.2). In particular,
then I is the ideal of the graph of a morphism, as in Section 5, then Theo-
rem 6.4 specializes to the formula for tropical implicitization given in [19].
It is important to note, however, that Theorem 6.4 applies to any ideal
/ whose tropical variety happens to be known, even if I is not a generic
complete intersection. For instance, T{I) might be the output of a GFan
computation, or it might be one of the special tropical varieties which have
already been described in the literature. Any tropical variety with known
multiphcities can serve as the input to the Trim command project .prl.
Acknowledgments. We are grateful to Peter Huggins for his help
at many stages of the Trim project. His iB4e software became a crucial
130 BERND STURMFELS AND JOSEPHINE YU

ingredient in our implementation. Bernd Sturmfels was partially supported

by the National Science Foundation (DMS-0456960), and Josephine Yu was
supported by a UC Berkeley Graduate Opportunity Fellowship and by the
Institute for Mathematics and its Applications (IMA) in Minneapolis.

REFERENCES

[1] LOUIS BILLERA AND BERND STURMFELS, Fiber polytopes, Annals of Mathematics
135 (1992), 527-549.
[2] TRISTRAM BOGART, ANDERS JENSEN, DAVID SPEYER, BERND STURMFELS, AND
REKHA THOMAS, Computing tropical varieties, Journal of Symbolic Compu-
tation 4 2 (2007), 54-73.
[3] ROBERT CORLESS, M A R K GIBSBRECHT, ILIAS KOTSIREAS, AND STEVEN W A T T , NU-
m,erical im,plicitization of parametric hypersurfaces with linear algebra, in:
Artificial Intelligence and Symbolic Computation, Springer Lecture Notes in
Computer Science, 1 9 3 0 (2000), 174-183.
[4] ALICIA DICKENSTEIN, EVA MARIA FBICHTNER, AND BERND STURMFELS, Tropical
discriminants, J. Amer. Math. Soc. 2 0 (2007), 1111-1133.
[5] JESUS D E LOERA, J O R G RAMBAU, AND FRANCISCO SANTOS, Triangulations: Ap-
plications, Structures and Algorithms, Algorithms and Computation in Math-
ematics, Springer Verlag, Heidelberg, t o appear.
[6] IOANNIS EMIRIS AND JOHN CANNY, Efficient incremental algorithms for the sparse
resultant and the mixed volume, J. Symbolic Computation 20 (1995), 117-150.
[7] IOANNIS EMIRIS AND ILIAS KOTSIREAS, Implicitization exploiting sparseness, in D.
Dutta, M. Smid, R. Janardan (eds.). Geometric And Algorithmic Aspects Of
Computer-8iided Design And Majiufacturing, pp. 281-298, DIMACS Series in
Discrete Mathematics and Theoretical Computer Science 6 7 , American Math-
ematical Society, Providence RI, 2005.
[8] IOANNIS EMIRIS, CHRISTOS KONAXIS, AND LEONIDAS PALIOS, Computing the New-
ton polytope of specialized resultants. Presented at the conference MEGA 2007
(Effective Methods in Algebraic Geometry), Strobl, Austria, June 2007.
[9] EWGENIJ GAWRILOW AND MICHAEL JOSWIG, Polymake: A framework for analyzing
convex polytopes. In Polytopes—Combinatorics and Computation (Oberwol-
fach, 1997), DMV Seminar, Vol. 2 9 , pages 43-73. Birkhauser, Basel, 2000.
[10] ISRAEL GEL'FAND, MIKHAEL KAPRANOV, AND ANDREI ZELEVINSKY, Discriminants,
Resultants, and Multidimensional Determinants; Birkhauser, Boston, 1994.
[11] P E T E R HUGGINS, iB4e: A software framework for parametrizing specialized LP
problems, A Iglesias, N Takayama (Eds.), Mathematical Software - ICMS
2006, Second International Congress on Mathematical Software, Castro Ur-
diales, Spain, September 1-3, 2006, Proceedings. Lecture Notes in Computer
Science 4151 Springer 2006 (pp. 245-247).
[12] ANDERS N . JENSEN, Gfan, a software system for Grobner fans. Available at
h t t p : //home. imf. au. dk/aj ensen/sof tweire/gf an/gf an. html.
[13] AsKOLD KHOVANSKII AND ALEXANDER ESTEROV, Elimination theory and Newton
polyhedra. Preprint: math.AG/0611107.
[14] JOHN MCDONALD, Fractional power series solutions for systems of equations, Dis-
crete and Computational Geometry 2 7 (2002), 501-529.
[15] P E T E R MCMULLEN, Mixed fibre polytopes. Discrete and Computational Geometry
32 (2004), 521-532.
[16] T O M MICHIELS AND RONALD COOLS, Decomposing the secondary Cayley polytope.
Discrete and Computational Geometry 2 3 (2000), 367-380.
[17] VICTOR SHOUP, N T L : A Library for doing Number Theory, available at
http://www.shoup.net/index.html.
 TROPICAL IMPLICITIZATION 131

[18] BERND STURMFBLS AND JENIA TEVELEV, Elimination theory for tropical varieties,
Preprint: math.AG/0704347.
[19] BERND STURMFELS, JENIA TEVELEV, AND JOSEPHINE Y U , The Newton polytope of
the implicit equation, Moscow Mathematical Journal 7 (2007), 327-346.
[20] GiJNTER ZiEGLER, Lectures on Polytopes, Graduate Texts in Mathematics 152,
Springer-Verlag, New York, 1995.
 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES:
C O M P U T I N G ALL EQUILIBRIA W I T H G A M B I T
A N D PHCPACK
T H E O D O R E L. TUROCY*

A b s t r a c t . This paper describes a new implementation of an algorithm to find all

isolated Nash equilibria in a finite strategic game. The implementation uses the game
theory software package Gambit to generate systems of polynomial equations which are
necessary conditions for a Nash equilibrium, and polyhedral homotopy continuation via
the package PHCpack to compute solutions to the systems. Numerical experiments
to characterize the performance of the implementation are reported. In addition, the
current and future roles of support enumeration methods in the context of methods for
computing Nash equilibria are discussed.

K e y words. Game theory, Nash equilibrium, numerical solution of polynomial

equations.

A M S ( M O S ) subject classifications. 91A05, 91A06, 91A12, 12-04.

1. Introduction. Game theory is a cross-disciplinary field whose aim

is to model and understand how rational decision-makers make choices.
Game theory focuses on situations where the outcome depends on the
choices of two or more such decision-makers. Some of game theory's most
significant success has been in the social sciences, most notably economics.
In recognition of this, the 1994 Nobel Memorial Prize in Economics was
awarded to John Nash, John Harsanyi, and Pleinhard Selten, three pioneers
of the field. The scope of the influence of game theory is not limited to
social science; for example, MAYNARD SMITH [8] applied game theoretic
concepts to the study of evolutionary biology. Computer scientists have
applied game theory to inform the design of automated agents.
A central concept of game theory is the Nash equilibrium. Nash equi-
librium can be thought of as a condition of mutual consistency among the
beliefs and actions of a collection of agents, each of whom seeks to maximize
his own objective function. (A formal definition is given in Section 2.) Nash
equilibria thus have a normative attraction, in that they represent points
which-are stable with respect to unilateral changes in action by any one
player. At a Nash equilibrium, no individual player can change his action
in such a way as to improve his outcome in the game.
For games in which there are a finite number of players, each of whom
can choose one action from a finite set of possible actions, the existence of
an equilibrium was established by NASH [12]. Nash's existence argument
allows players to choose from the set of probability distributions over their
actions. For many games, this is essential; the game "rock, paper, scissors"
(RPS) is a familiar example in which randomization is necessary for the

'Department of Economics, Texas A&M University, College Station, TX 77843

(turocy®econmail.tamu.edu).
133
134 THEODORE L. TUROCY

existence of equilibrium. While the mutual consistency property of Nash

equilibrium suggests a fixed-point formulation, the Nash equilibrium prob-
lem can be stated in several equivalent ways; see the survey of M C K E L V E Y
AND MCLENNAN [9] for more details. Each formulation suggests a different
computational approach for locating equilibria.
The mathematical versatility and beauty of the Nash equilibrium for-
mulation aside, practitioners in the various fields in which the concept
is applied are primarily interested in getting the set of equilibria of their
games of interest, and in estimating what games are feasible to solve with a
given computational budget. The software package Gambit ( M C K E L V E Y ,
M C L E N N A N , AND T U R O C Y [11]) aims to provide such users a set of rou-
tines for computing the equilibria of an arbitrary finite game. Currently,
Gambit provides a number of methods which can compute one or more
equihbria in games with more than two players. However, clear guidance
from experience as to which method or methods to use remains absent.
This paper documents one step in the direction of "black-boxing" the pro-
cess of computing equilibria in those games.
One class of approaches which has received recent attention is based
on the solution to systems of poljmomial equations. The calculation is
organized by enumerating supports, where a support is a set of strategies
which are assigned positive probability. The Nash equihbrium conditions
imply a set of polynomial equations, the solutions of which contain the set
of Nash equihbria on that support.
This approach is not new; in fact, this is the way most humans go
about trying to manually compute equihbria. DiCKHAUT AND KAPLAN [4]
first documented a program to compute equilibria using support enumer-
ation. A method which attempts to compute all isolated Nash equilibria,
using essentially the support enumeration described in this paper, has been
present in Gambit since the mid-1990s.
This paper reports computational results on a new implementation,
using Gambit as the frontend for defining games and generating the sys-
tems of polynomial equations to be solved. Polyhedral homotopy continu-
ation (HUBER AND STURMFELS [6]) is used as the backend to generate the
solutions of these systems, using PHCpack (VERSCHELDE [16]).
The program bridges a gap between two previous papers. HERINGS
AND P E E T E R S [5] develop essentially the same algorithm, decomposing the
game by means of "admissible" supports and showing how to solve the
resulting systems of polynomial equations using a homotopy method. This
paper augments their work by giving and proving the correctness of an
explicit algorithm for finding all admissible supports, and characterizing
the scaling of the algorithm as the size of the game increases. The choice
of PHCpack as the backend is inspired in part by the results of DATTA
[3], who gives a performance report on solving for all totally mixed Nash
equilibria of a game using PHCpack. Those results show that the numerical
 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES 135

approach in PHCpack outperforms programs to solve polynomial equations

via computer algebra using Grobner bases.
The paper proceeds as follows. Section 2 outlines the Nash equilibrium
problem. It describes the support enumeration process to compute all sup-
ports which might harbor Nash equilibria, and argues the correctness of the
process. Section 3 works through the computational process for a sample
game. Section 4 evaluates the performance and scalability of the program.
Section 5 discusses games in which the set of Nash equilibria has compo-
nents with positive dimension, and the imphcations for implementations
of this method as a "black box" solver. Finally, Section 6 concludes with
a discussion of future directions and extensions, and the role of support
enumeration methods within the toolkit of computation in game theory.

2. The algorithm.

2.1. Games in strategic form. A game in strategic form is played

by a collection of N players, indexed i = 1,2,..., AT. Each player i in-
dependently and simultaneously chooses from a set of strategies Si =
{ 1 , 2 , . . . , Jj}. These choices determine an outcome {ji,J2, • • • ,JN}- Play-
ers rank outcomes according to a payoff function Ui{ji,J2,---,JN), with
larger values being preferred to smaller ones.
For example, consider the game played by two drivers approaching
each other on a country road. Each driver can choose to steer his vehicle
to the right (strategy number 1) or to the left (strategy number 2). The
drivers prefer not having a collision to having one, but do not care whether
they pass each other on the right or the left. This ranking of outcomes
is captured by the payoff function iti(l,l) = MJ(2, 2) = 1 and itj(l,2) =
Ui{2,1) = 0 for both drivers i = 1,2.
In some types of games, players may find it in their best interest to
be unpredictable. In RPS, for example, a player who chooses rock with
certainty will lose a majority of games, while a player who randomizes his
play equally across the three choices can guarantee he will win at least one-
third of games and lose no more than one-third of games. A randomized
strategy for player i will be denoted TTJ, with iTij being the probability
that player i chooses his jth "pure" strategy. The set of all randomized
strategies for player i is S j . The payoS^ function is extended to account for
mixed strategies by taking the expected value

"i(7ri,7r2,...,7rAr)

= X ] 5 Z • • • 5 Z "»(jl'J2, • • • ,JN)T^lnT^2J2 • • • T^NJM-

Thus the payoff function is polynomial, or more specifically multilinear, in

the probabilities TTJJ .
136 THEODORE L. TUROCY

2.2. Nash equilibrium. One objective of game theory is to predict

how players will make their choices. The central concept of classical game
theory is the Nash equilibrium. In words, a Nash equilibrium is a situation
in which each player makes a choice which maximizes his expected payoif,
given the choices all other players make. Returning to the example of the
two drivers, it is a Nash equilibrium for both drivers to steer to the right.
If driver 1 expects driver 2 to steer to the right, then driver 1 wants also
to steer to the right. The same is true for driver 2, if he expects driver 1
to steer to the right. This is not the unique Nash equilibrium, however; it
is also a Nash equilibrium for both drivers to steer to the left.
Nash equiUbria may also involve active randomization by one or more
players. Consider RPS, and assume both players get a payoff of 1 if they
win, -1 if they lose, and 0 for a tie. Suppose player 1 anticipates player 2
will choose paper; then, player 1 prefers to play scissors. But, if player 1
is going to play scissors, player 2 does not want to play paper; rather, he
would want to play rock. Continuing this chain of reasoning leads to a cycle.
RPS does not have any Nash equiUbria in which players do not randomize.
NASH [12] estabUshed that an equihbrium point in randomized strategies
always exists for games with a finite number of players and strategies.
Formally, a randomized strategy profile TT = (Trj)^^ is a Nash equilib-
rium if and only if it satisfies

Wj(7r) >'Ui(7ri,7r2,...,/9i,...,7rjv)

for all Pi £ Sj for alH = 1,2,..., TV. That is to say, for each player i there is
no other randomized strategy pi which gives strictly higher expected value
t h a n TTj.
Because the expected payoff to player i is linear in the probabilities
TTij assigned to the player's own strategies, a profile can only be a Nash
equilibrium if it assigns positive probability only to pure strategies in the
set argmaXjg^. Ui{iTi,..., j,...,TTAT). Therefore, all strategies which are
played with positive probability under TTJ must give the same expected
payoff: tij(s,7r_j) = •Uj(t,7r_i) for any two strategies s,t G Si such that
TTjg > 0 and TTjt > 0. This means the Nash equilibrium conditions can be
written^

Mj(s,7r_i) = Ut(t, 7r_j)Vs,t € 5i such that TTJS > OandTTjt > 0;

Ui(s,7r_i) > Mj(t,7r_j)Vs G 5, such that TTJS > 0,Vt G Sf,
ns > OVs G Si-,
Ji
E^'^^1- (2.2)

^The abusive subscript notation —i indicates the components of an object belonging

to all players other than player i.
 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES 137

In RPS, it is a Nash equilibrium for each player to play each of the

three choices rock, paper, and scissors with probability g. Each player then
will win one-third of the games and lose one-third of the games, with the
remaining one-third of the games resulting in draws, for an expected payoff
of zero. Furthermore, conditional on playing rock, each player will win one-
third, lose one-third, and draw one-third; the same is true for paper and
scissors. The equality conditions in (2.2) are therefore satisfied.
It turns out that this is the unique Nash equilibrium for RPS. To verify
this, it is necessary to check all possible "supports," or sets of strategies
played with positive probability. For example, one would need to verify
that there are no solutions satisfying (2.2) when player 1 plays rock and
scissors with positive probability, but never plays paper, while player 2
plays paper and scissors with positive probability, but never plays rock.
For any given game, the number of possible supports is
N

since each strategy Sj G Si can independently be in or out of the support

(giving 2"^'), except the support where all strategies are out is not valid
(minus one). The combinatoric burden can be lessened somewhat by iden-
tifying supports for which it can be easily determined that solutions to
(2.2) are impossible.
2.3. E x a m i n i n g possible equilibrium s u p p o r t s . A method for
efiiciently enumerating "admissible" supports (following the terminology
of HERINGS AND P E E T E R S [5}) has been part of the Gambit package
(MCKELVEY ET AL [11]) since the mid-1990s. This was independently
described in HERINGS AND P E E T E R S [5]. This section presents a formal
correctness argument.
In what follows, a capital letter, for example T, refers to a subset of the
strategies in a game. A subscripted capital refers to the strategies in that
set belonging to a player; thus, Ti denotes the strategies in T belonging to
player i. Calligraphic capitals refer to collections of sets of strategies, or
collections of supports. A support is defined a subset of the strategies in a
game that satisfies the additional requirement that each player has at least
one strategy in the set. That is, a support T is a subset of 5 = U^iSi
such that, for each player i,TiCSi and Ti^%.
A strategy Si G Si strictly dominates another strategy r^ G 5, if
Ui{si,t) > Ui{ri,t) for all i G S-i. A strictly dominated strategy can-
not be played with positive probability in a Nash equihbrium, since the
conditions in (2.2) cannot be satisfied. This concept can be generalized in
the context of determining the set of admissible supports. A mixed strategy
TTi strictly dominates a pure strategy Si against a support T if

TTj y-T Si = Ui{-Ki,t-i) > Ui{Si,t^i)\/t-i G XjjLjT-i.

138 THEODORE L. TUROCY

The operator ^T is monotonic in T in that, if TTJ ^T Si, then TTJ >-V Si for
any support V C T.
The undominated strategies Si in a support T are then defined as

U{T) = {s e T :^ TTi e Ei such that TTi ^T Si}.

The conditions (2.2) cannot be satisfied if T contains any strategies which

are dominated against T, i.e., if U{T) c T. Let U*{T) be the support
resulting from iteratively applying the operator U to T. Since the game
has a finite number of strategies, U*{T) can be computed in a finite number
of steps.
The set of admissible supports is

Af = {T:U{T) = T}.

The task is to efficiently enumerate the set M. For any two sets X,Y C S
with X n F = 0, define

V{X,Y) = {T G AA : X G Tforallx G X,andy ^Tiorally G Y}. (2.3)

In words, •P(X, Y) is the set of admissible supports in which all the strate-
gies in X are present, and all the strategies in Y are not present. Observe
that P ( 0 , 0 ) = A f . I i X u Y = S, then

V(YV\-i {^} ^ ^ ^ W = X a n d X i ^ 0 foralH,

f^(^,r) | 0 if{/(X)7^XorXi=0 forsomei.

li X UY C 5, then V{X, Y) can be decomposed into the union of two

disjoint sets

V{X, Y) = P{X U {s}, y ) U V{X, Y U {s}) (2.4)

forany s G 5 \ ( X u y ) .
The following recursive procedure based on (2.4) efficiently uses dom-
inance information to compute Af. For any partition of S into X, Y, and
Z = S\(XUY),
• Compute U*{XUZ).
— If some player i has no strategies in U*{X U Z), P{X, Y) = 0.
- If some strategy s G X but s ^ U*{X U Z), P{X, Y) = 0.
• Let Y' = Yu{s e Z : s ^ U*{X U Z)}, Z' = Z\{s & Z : s ^
U*{X\JZ)}.
• IiZ' = 0, then V{X,Y) = {X}.
• If Z ' 9^ 0, select any strategy z £ Z'. Compute V{X, Y) recur-
sively by

ViX, Y) = V{X U {z}, Y') U V{X, Y' U {z}).

 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES 139

Thus the following proposition holds.

PROPOSITION 2.1. Every support containing a Nash equilibrium is in
the set V{$,$). Furthermore, the support enumeration process visits each
of these supports exactly one time.
Proof. •p(0,0) = jV follows from the definition in (2.3), observing that
the universal quantifiers hold trivially because X and Y are empty. The
process visits any support at most once because the sets in the decompo-
sition in (2.4) are disjoint. D
3. A worked e x a m p l e . To illustrate the computational process, con-
sider the three-player game in Table 1. In this game, each player has two
strategies, labeled t and b for Player 1, / and r for Player 2, and u and
d for Player 3. Each row gives the respective payoffs to the three players
for each of the possible strategy combinations. The set of Nash equihbria
of this game consists of nine isolated points, which has been shown to be
the maximal number a game of this dimension can have ( M C K E L V E Y AND
MCLENNAN [10]).

TABLE 1
A three-player game with nine isolated Nash equilibria.

1 2 3 Wl, '"2, Uz
t / u 9, 8, 12
t / d 0,0,0
t r u 0,0,0
t r d 3,4,6
b I u 0,0,0
b I d 3,4,6
b r u 9,8,2
b r d 0,0,0

At initialization, set X = F = 0, so Z = {t, b, I, r, u, d}. By inspection,

no strategies are dominated against XUZ. Therefore, the algorithm selects
a strategy from Z, say t, and recursively computes P{{t}, 0) and V{$, {t}).
For X = {t} and Y — $, X LJ Z = S; as noted before, no strategies
are dominated. Therefore, select a strategy from Z, for example b, and
recursively compute P({t, b}, 0) and VHt}, {b}).
Take the case X = {t}, Y = {b}. No strategies are dominated against
XuZ = {t, I, r, u, d}. Select I from Z, and recursively compute P({f, /}, {b})
Bnd Vi{t},{b,l}).
Take the case X = {t}, Y = {b,l}. Here, b and / are guaranteed
not to be used by players 1 and 2, respectively. Therefore, only the two
rows in Table 1 that are relevant are those where player 1 chooses t and
player 2 chooses r. Against the support {t,r,u,d}, d dominates u for
player 3, so U{{t,r,u,d}) = {t,r,d}. Therefore, X = {t}, Y' = {b,l,u},
140 THEODORE L. TUROCY

and Z' = {r, d}. At this point, r and d must be assigned to X, since
otherwise the resulting support will not be vahd. Therefore, {t, r, d} is an
admissible support.
Proceeding in this way, there are 11 admissible supports, out of
3 X 3 X 3 = 27 possible.
M ={{t, b, I, r, u, d}, {t, b,l,r,u}, {t, b, I, r, d}
{t, b, I, u, d}, {t, b, r, u, d}, {t, l,r, u, d},
{t,I,u}, {t,r,d}, {b,l,r,u,d}, {b,I,d}, {b,r,u}}.

For each support, consider whether conditions (2.2) can be satisfied. An

immediate consequence of admissibihty is that any support in A/" with ex-
actly one strategy for each player is automatically an equihbrium.
Now consider the full support {t, b,l,r, u, d}. Let TTI be the probabihty
player 1 chooses t, TT; the probabihty player 2 chooses /, and 7r„ the proba-
bility player 3 chooses u. The equal-payoff conditions in (2.2) for the three
players are

97r;7r„ + 3(1 - 7ri){l - 7r„) = 37rz(l - 7r„) + 9(1 - 7ri)7r„

87rt7r„ + 4(1 - 7rt)(l - 7r„) = 47rt(l - 7r„) + 8(1 - 7rt)7r„
127rt7ri + 2(1 - 7rt)(l - n) = 67rt(l - TT;) + 6(1 - Trt)Tri.

This system of equations has two solutions: (7rt,7r;,7r„) = ( | , | , 5) and

(7rt,7ri,7r„) = (5, | , | ) . Since all strategies are used with positive probabil-
ity, all the inequality conditions in (2.2) hold automatically, and so both
these solutions are Nash equihbria.
Next, consider the admissible support {t,l,r,u,d}. The equal-payoff
conditions for players 2 and 3 simphfy to

4(1 - TTu) = Situ

2(1 - TTi) = e-rci.
This system has solution (7ri,7r„) = ( | , | ) , making (7r(,7r;,7ru) = (O, \, 5)
a candidate Nash equihbrium. To verify that this is an equihbrium, the
condition tii(6;7r_i) > ui(t;7r_i) must hold; it does, since
3
ui(6;7r_i) = 37ri(l -7r„)-|-9(l -7r()7ru = 4 -

ui(i;7r_i) = 97r;7r„4-3(l - n){l - ^u) = 2 - .

Not all solutions of the polynomial equations correspond to Nash equi-

libria. Consider the admissible support {b, l,r,u,d}. The equal-payoff con-
ditions are

STTU = 4(1 - 7r„)

127ri = 6(1 - TT/),
 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES 141

which has solution (TTIJ-KU) = (5, 5), making (7rt,7r;,7r„) = (O, | , | ) a can-
didate Nash equilibrium. However, the condition ui(f;7r_i) > ui(6;7r_i)
fails to hold, since ui{t; 7r_i) = 2 | but ui{b; 7r_i) = 2 | . Thus, this solution
does not correspond to a Nash equilibrium.
In total, eight of the admissible supports have at least one Nash equi-
librium, with the full support having two, bringing the total number of
Nash equilibria to nine:
(TTt, TT,, 7r„) - {(1/2,2/5,1/4), (2/5,1/2,1/3),
(1/2,1/2,1), (1/3,1,1/4),
(1,1,1), (1,0,0), (0,1/4,1/3), (0,1,0), (0,0,1)}.
In addition, on three supports, {t,b,l,r,d}, {t,b,r,u,d}, and {t,l,r,u,d},
there is a solution to the equal-payoff conditions that does not correspond
to a Nash equilibrium.
4. Performance and scalability. The support enumeration method
has been implemented in Gambit. The program uses the Gambit library
for the support enumeration, and hands off the systems of equations gen-
erated to PHCpack for solution. The numerical experiments in this section
characterize the scalabiUty of the method on games of various sizes, as a
rough guide to the practical hmitations of the algorithm in the current
implementation.
4.1. Games with randomly-drawn payoffs. The first set of re-
sults evaluates the computational time for games of a given dimension
with randomly-selected payoffs. Each payoff is drawn independently from
the uniform distribution on [0,1].

TABLE 2
Some summary statistics on the average performance of the program on games
with randomly-drawn payoffs.

Runtime (sec) #Ga-

Dimension Supports NE non-NE Enum PHCpack mes
2^ = 2x2x2 2.98 1.92 1.63 0.01 0.02 100
-2^ 12.75 3.86 22.15 0.27 0.28 100
2^ 54.98 7.34 195.51 0.48 4.14 100
2« 224.48 15.52 1559.54 3.19 82.48 100
2^ 866.11 34.09 12538.22 21.57 1881.51 100
2« 3123.45 78.45 106185.27 164.70 51362.03 11
3x3x3 55.45 4.68 111.93 0.27 1.49 100
4x4x4 909.80 11.61 3469.58 6.40 108.71 100
5x5x5 12616.64 28.24 94374.07 136.99 6652.47 87
3x3x3x3 805.34 18.45 6101.96 8.90 318.53 100

Table 2 presents summary statistics on the average performance of the

program. The dimension column gives the number of strategies for each
142 THEODORE L. TUROCY

player. The second column gives the average number of admissible supports
considered in the games sampled. The third and fourth column together
report the average number of solutions found by PHCpack, divided into
those which correspond to Nash equilibria and those which do not. The
fifth and sixth columns give the average runtime, in seconds, on a Pentium
IV workstation with a single 2.8GHz processor. The column "Enum" gives
the time spent performing the enumeration, and "PHCpack" the time spent
solving the systems of equations.
The striking feature of Table 2 is that the time per invocation of PHC-
pack, and the number of solutions found that do not correspond to Nash
equilibria, increase very rapidly in the size of the game. The average num-
ber of supports visited ranges from roughly one-third to one-half of the
maximum possible for eax;h size game.

4.2. Games with more structure: a coordination game ex-

ample. It is possible, however, that these results are unduly pessimistic.
Games which are of interest to practitioners are not drawn randomly from
the set of possible games; rather, they often have a particular structure.
This point is made, for example, by NUDELMAN ET AL [14], whose GAMUT
database of games collects examples from several such classes. For some
classes of games, the support enumeration process may be able to exploit
this structure by eliminating many supports from consideration. Based on
the timing results in the previous section, reducing the number of polyno-
mial systems to be considered should permit significantly larger games to
be solved feasibly.
As an example, consider the family of "minimum effort" games. Each
player selects a level of effort firom an interval of integers { 1 , 2 , . . . , J } .
A player selecting a level of effort e incurs a cost of effort c(e), where
the function c(-) is increasing in effort. Each player receives a benefit
6(min(ei, 6 2 , . . . , BAT)), which depends only on the lowest level of effort cho-
sen. Each player's payoff is given by

Ui{ei, 6 2 , . . . , BN) = 6(min(ei, 6 2 , . . . , ejv)) - c{ei).

It is an equilibrium in pure strategies for all players to choose the same

effort level e, for each possible effort level e. In addition, there are many
equilibria in which randomization is present.
For example, consider the case of c(e) = 2e and 6(e) = 3e, with Wj = 0.
For N = 3 and J = 4, there are 336 admissible supports, and total runtime
is 33.3 seconds. Similarly, for A'' = 3 and J = 5, there are 2677 admissible
supports, and the program runs in 1590 seconds, and when N = i and
J = 4, there are 3203 admissible supports, and runtime is 9138 seconds.
These compare favorably with the "average" games from Table 2. In this
example, the symmetry of the game is not exploited, which would further
significantly decrease the computational load.
 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES 143

5. Games with non-isolated equilibria. In all the examples in

Section 4, the set of Nash equilibria consists of a collection of isolated points
in the space of randomized strategies. This phenomenon is generically true,
in that it occurs with probability one if payoff values are drawn randomly.
However, games which are of interest in applications often do not fall into
this category. In these games, the set of equilibria may have components
with positive dimension.
NAU ET AL [13] provide an example of such a game, with payoffs as
given in Table 3.

TABLE 3
A three-player game with a positive-dimension component of Nash equilibria span-
ning five admissible supports.

1 2 3 Ui,U2,U3
t / u 0,0,2
t / d 1,1,0
t r u 0, 3,0
t r d 0,0,0
b I u 3,0,0
b I d 0,0,0
b r u 0,0,0
b r d 0,0,3

This game has three Nash equilibria in pure strategies: (7rt,7r;,7r„) =

(1,0,1), (0,1,1), (0,0,0). The first two are extreme points of one connected
component of equilibria U1UU2UU3, where

Ui = < (7rt,7ri,7ru) = ( l , 0 , a ) f o r a £
i'
U2 = < (TTt, TTi, ^„) = (0,1, a ) for a e
i^]}
% = |(7rt,7ri,7r„)= L _ ,^ " , a , ^ j fora € [ 0 , 1 ] | .

The set of Nash equilibria is shown in Figure 1. The connected component

contains equilibria on a total of five admissible supports.
This example illustrates the need for a full "black-box" solver to iden-
tify supports on which a positive-dimension set of solutions exists, and to
identify components which span multiple admissible supports. This initial
implementation does neither, and is able to identify only the two pure-
strategy endpoints of the connected component, as well as the two "corner"
points.
To understand the types of degeneracy which may arise, consider the
polynomial systems generated by some of the admissible supports. On the
144 T H E O D O R E L. TUROCY

F I G . 1. The set of Nash equilibria {nt,TTi,nn) for the game in Table 3.

support {i,fe, r, d}, the sole equal-payoif constraint reduces to 0 = 0; this

occurs on a total of six admissible supports. The implementation flags these
as "singular," meaning the user needs to examine these supports manually
for possible equilibria. The corner points are identified, but only as edge
cases on larger supports. For example, the Nash equilibrium (0,1, | ) is
identified as a solution on the support {t, b, I, u, d}.
While it is fairly straightforward to flag some supports as singular,
checking for possible positive-dimension solution sets remains imperfect.
The positive-dimension component in this game passes through the interior
of the cube in Figure 1, which corresponds to the full support {t, b, I, r, u, d}.
On this support, the equal-payoff conditions are
Tri{l — 7r«) = STT/TTU
7rt(l — 7r„) = BirtTTu
2-Kt-ni - 3(1 - 7rt)(l • 7 r „ ) .
Since TTt, TT; ^ 0, the first two equations are redundant, and the system has
solutions along the curve U^. The current implementation does not identify
this redundancy. It does identify two solution points which are not Nash
equilibria, in which itt or TT; fall outside [0,1].
6. Discussion. Support enumeration methods will hkely play a sig-
nificant role in the computation of Nash equifibria in finite games. This
method is the only one currently known that is guaranteed to find all iso-
lated Nash equilibria in finite time. It will therefore be an essential tool
for applications requiring identification of all equihbria. BAJARI ET AL [1]
use this method in econometric estimation of games and equilibria based
on empirical data.
 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES 145

In experimental economics, researchers design the parameters of games

to be played by paid subjects in controlled laboratory settings. The prop-
erties of the set of Nash equilibria are one criterion often used in selecting
those parameters. For example, CHEN ET AL [2] applied the program to a
cost-sharing game to establish the nonexistence of Nash equilibria in which
any player randomizes. With this knowledge, it is possible to test whether
Nash equilibrium is a useful organizing principle for the observed data.

The support enumeration method can be specialized for particular

tasks. For example, one might want to know whether a strategy is played
with positive probability in any equilibrium. To do so, one only needs to
consider the admissible supports containing that strategy. In other cases,
one might be interested in finding out whether the game has a unique
equilibrium. An effective method to accomplish this should minimize the
time it takes to find two equilibria, if two are present. Some heuristics
in this direction are investigated by P O R T E R ET AL [15], who find that
support enumeration again stacks up well against other methods for finding
a sample equilibrium.

Despair at the combinatoric complexity of the support enumeration

process should also be tempered by the observation that many games are
naturally represented in extensive, or game tree, form. Using the sequence
form representation of KOLLER, MEGIDDO, AND VON STENGEL [7], the
strategy space for each player may be represented more efficiently. Con-
verting an extensive game to strategic form for use with the support enu-
meration method described here may result in a strategic form which is
exponential in the size of the original tree. The sequence form represen-
tation, on the other hand, is hnear in the size of the tree. A support
enumeration method similar to the one described here can be used on the
sequence form, and the Nash equilibrium conditions for each support result
in a system of polynomial equations and inequalities, just as in the strate-
gic form. An implementation of the support enumeration process has been
available in Gambit since the mid-1990s; interfacing this with the PHCpack
solver is a future objective.

A second avenue for future development is in solving the systems

of equations generated by the Nash equilibrium conditions on a support.
These systems have some regular structures. The equations generated by
the indifference conditions for player i involve only the probabilities of
strategies of players other than player i. In addition, the current imple-
mentation ignores the nonnegativity constraints on the variables in solving
the systems. The equations generated by one support are similar to those
generated by other supports which differ in the addition or removal of one
strategy. In addition, HERINGS AND P E E T E R S [5] list several other possible
optimizations for special-case supports. None of these possible optimiza-
tions are currently exploited by this implementation.
146 T H E O D O R E L. TUROCY

7. Acknowledgements. This paper is based on a talk given at the

Institute for Mathematics and its Applications at the University of Min-
nesota, during the Software for Algebraic Geometry workshop in Octo-
ber 2006. Comments and suggestions from the participants there were
greatly appreciated. Suggestions from three anonymous referees did much
to sharpen the exposition.
The programs described here, including the scripts used to build
the randomly-drawn and minimum-effort games, are all available un-
der the GNU General PubUc License from the Gambit website at
http://gambit.sourceforge.net.

REFERENCES

[I] PATRICK B A J A R I , H A N H O N G , AND STEPHEN RYAN. Identification and estimation

of discrete games of complete information. Working paper, University of Min-
nesota, 2007.
[2] YAN C H E N , LAURA RAZZOLINI, AND T H E O D O R E L . TUROCY. Congestion Allocation
for Distributed Networks: An Experimental Study. Economic Theory, 33:121-
143, 2007.
[3] R.S. DATTA. Using Computer Algebra To Compute Nash Equilibria. In Proceed-
ings of the 2003 International Symposium on Symbolic and Algebraic Compu-
tation (ISSAC 2003), pages 74-79, 2003.
[4] JOHN DICKHAUT AND T O D D KAPLAN. A Program for Finding Nash Equilibria. T h e
Mathematica Journal, l(4):87-93, 1992.
[5] P . J . J . HERINGS AND R . J . A . P . PEETERS. A globally convergent algorithm to com-
pute all Nash equilibria for n-Person games. Annals of Operations Research,
137:349-368, 2005.
[6] BiRK HUBER AND BERND STURMFELS. A polyhedral method for solving sparse
polynomial systems. Mathematics of Computation, 64:1541-1555, 1995.
[7] DAPHNE ROLLER, NIMROD MEGIDDO, AND BERNHARD VON STENGEL. Efficient com-
putation of equilibria for extensive two-person games. Games and Economic
Behavior, 14:247-259, 1996.
[8] JOHN MAYNARD SMITH. Evolution and the Theory of Games. Cambridge UP,
Cambridge, 1982.
[9] RICHARD D . M C K E L V E Y AND ANDREW M . M C L E N N A N . Computation of Equilibria
in Finite Games. In Hans Amman, David A. Kendrick, and John Rust, editors,
T h e Handbook of Computational Economics, Volume I, pages 87-142. Elsevier,
1996.
[10] RICHARD D . M C K E L V E Y AND ANDREW M . M C L E N N A N . The maximal number of
regular totally mixed Nash equilibria. Journal of Economic Theory, 72:411-
425, 1997.
[II] RICHARD D . M C K E L V E Y , ANDREW M . M C L E N N A N , AND THEODORE L . T U R O C Y .
Gambit: Software Tools for Game Theory, http://gambit.sourceforge.net.
[12] JOHN F . NASH. Equilibrium points in n-person games. Proceedings of the National
Academy of Sciences, 36:48-49, 1950.
[13] ROBERT N A U , SABRINA GOMEZ CANOVAS, AND P I E R R E HANSEN. On the Geometry
of Nash Equilibria and Correlated Equilibria. International Journal of Game
Theory, 32:443-453, 2004.
[14] E. NUDELMAN, J. WORTMAN, Y . SHOHAM, AND K . LEYTON-BROWN. Run the
GAMUT: A comprehensive approach to evaluating game-theoretic algorithms.
In Proceedings of t h e Third International Joint Conference on Autonomous
Agents & Multi Agent Systems, 2004.
 TOWARDS A BLACK-BOX SOLVER FOR FINITE GAMES 147

[15] RYAN W . P O R T E R , E U G E N E NUDELMAN, AND YOAV SHOHAM. Simple search

methods for finding a Nash equilibrium. Gaines and Economic Behavior,
forthcoming.
[16] JAN VERSCHELDE. Algorithm, 795: PHCpack: a general-purpose solver for polyno-
mial systems by homotopy continuation. ACM Transactions on Mathematical
Software, 25(2), 1999.
 A P A T O O L S : A SOFTWARE TOOLBOX FOR
A P P R O X I M A T E POLYNOMIAL A L G E B R A
ZHONGGANG ZENG*

A b s t r a c t . Approximate polynomial algebra becomes an emerging area of study

in recent years with a broad spectrum of applications. In this paper, we present a
software toolbox ApaTools for approximate polynomial algebra. This package includes
Maple and Matlab functions implementing advanced numerical algorithms for practical
applications, as well as basic utility routines that can be used as building blocks for
developing other numerical and symbolic methods in computational algebra.

K e y words, software, polynomial, GCD, factorization, elimination, Jordan Canon-

ical Form, multiplicity, rank.

A M S ( M O S ) subject classifications. 11R09,12D05,65F15,65F20,65F22,65H10.

1. Introduction. Approximate polynomial algebra emerges as a

growing area of study in recent years. With rich theories and abun-
dant symbohc algorithms already developed in commutative algebra and
algebraic geometry, a solid foundation is in place for building numerical
and hybrid computing methods. In fact, many robust numerical and
numeric-symbolic algorithms have been developed for solving polynomial
systems [1,11, 17, 25, 28], univariate factorization [32, 33], univariate GCD
[5, 9, 13, 15, 23, 30], multivariate GCD [10, 34, 35], computing the dual
bases and multiplicity structures of polynomial ideals [2, 7, 20], multivariate
factorization [4, 10, 34], and polynomial elimination [3, 8, 31], etc. Those
algorithms have a broad spectrum of applications in scientific computing
such as robotics [19, 24, 25, 26], control [6, 14], image processing [21, 22],
computational biology and chemistry [8], and so on.
In this paper, we present a software toolbox ApaTools for approximate
polynomial algebra. This toolbox includes Matlab and Maple implemen-
tations of basic algorithms, utility procedures and test suites. The Matlab
version of ApaTools is also named Apalab. Those functions can be used
either directly in applications or as building blocks for more advanced com-
puting methods.
One of the main difficulties for numerical computation in polynomial
algebra is the ill-posedness that frequently occurs when a problem has a
discontinuous solution with respect to data. Those ill-posed problems are
not directly suitable for floating point arithmetic since the solutions are
infinitely sensitive to rounding errors. It has been shown in recent de-
velopment that such difficulty can be overcome by seeking an approximate
solution as formulated based on the three principles of backward nearness,

* Department of Mathematics, Northeastern Illinois University, Chicago, IL 60625,

USA. Research supported in part by NSF under Grant DMS-0412003 and DMS-0715127
(zzengtoeiu.edu).
149
150 ZHONGGANG ZENG

maximum co-dimension and minimal distance [30, 33, 36] which we shall
elaborate in §2. With such formulations, algorithms can be constructed
using a two-staged strategy: identifying the solution structure first followed
by solving a least squares problem for an approximate solution. ApaTools
includes generic routines for matrix building and Gauss-Newton iteration
that are the main engines for both stages of computation.
Algorithms implemented in ApaTools are designed to regularize the
problem for removing ill-posedness rather than extending machine pre-
cision for accurate computation. The Maple version of ApaTools can
nonetheless take advantage of variable machine precision via setting the
worksheet "Digits" conveniently. The Matlab version (Apalab) uses the
hardware precision throughout, and is generally much faster than the Maple
counterparts.
ApaTools is at its start as an on-going project. At this stage, the
main objective is to provide researchers in polynomial algebra with generic
and versatile tools that simplify and accelerate algorithm development, ex-
perimentation, and implementation. Particularly on the Maple platform,
enabling convenient and faster coding take higher priority over fast execu-
tion. Moreover, we emphasize on achieving the highest possible accuracy
and robustness in algorithm design and implementation. Nevertheless, the
functions in ApaTools such as approximate GCD appear to be "the most
efficient and rehable" [27, page 223].
ApaTools is maintained at the author's website and fi-eely accessible
to academic researchers and educators for the foreseeable future.

2. Exact and approximate polynomial algebra. Conventional

symbolic computation assumes both data and arithmetic to be exact. In
practical applications, however, problem data are likely to be approximate.
As a result, exact solutions of those inexact problems may not serve the
practical purposes. Alternatively, an approximate solution is sought to
approximate the underlying exact solution. The comparison can be seen
in following examples.
EXAMPLE 1. Exact and approximate G C D . The following poly-
nomial pair has an exact GCD in gci{f,g) =x^ — 2xy + 3z:

f{x,y,z) = IX^^-^o^-^z-fx^y+f^^yz+fzx-^z^,
g{x,y,z) = §0^4+ a | x 2 j ; - i x 3 j ; - § x y 2 + 1 1 ^ 2 ^ + | Z j ; ^ . ^^•'>

When polynomials are represented with finite precision floating point num-
bers in coefficients

f{x, y, z) = 1.57142857x^ - 1.35294H8x^z - 3.14285714x^2/ + 2.70588235x^2

+4.7142857120; - 4.05882353z2,
g{x,y,z)= 1.06896552X''+ 1.26086957x^y-2.13793103x^2/-2.52173913xj/^
+3.20689655a;2z + 3.78260870^2:,
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 151

FIG. 1. Matlab results for the polynomial. (2.3).

the exact GCD degrades to 3cd{f,g) = 1. The approximate GCD within

e, as calculated by our ApaTools function mvGCD for 2 x 10"^ < e < 10"^,

^a^(/, g) = .99999999993;^ - 1.999999994xj/ + 2.999999994z

is an accurate approximation to the underlying exact GCD. D

EXAMPLE 2. Exact and approximate univariate factorization.
Univariate factorization in complex field C is equivalent to root-finding.
Consider univariate polynomial

p{x) = x^°° — 400x199 + 79500x'9» + . . . + 2.04126914035338 • 10*" x^°°

— 3.55467815448396 • 10*® x^^ + . . . + 1.261349023419937 • 10^^ x^ (2.3)
— 1.977831229290266 • 10^^ x + 1.541167191654753 • lO'"'
« (x - lf°{x - 2f°{x ~ 3)*°(x - 4)20 (2.4)

with coefficients in hardware precision (16-digits). Using the coefficient

vector of p as input, the standard Matlab root-finder outputs a cluster of
200 roots as shown in Figure 1.
In contrast, our Matlab function uvFactor in Apalab calculates an
approximate factorization:
» [F.res.cond] =• uvFactor(p);
THE COSDITIOH NUMBER: 2.57705
THE BACKWARD ERROR: 7.62e-016
THE ESTIMATED FORWARD ROOT ERROR: 3.93e-015

FACTORS

( X- 4.000000000000008 )-20
( X- 2.999999999999994 )-40
( X- 2.000000000000002 )-60
( X - 1.000000000000000 )-80
152 ZHONGGANG ZENG

codimension

n(2,2)
n(4); n(i,i,2) -n(i,1,1,1)
n(i,3)

F I G . 2. Stratification of manifolds of degree 4 polynomials, -with "—•" denoting

"in the closure of".

The approodmate factors calculated by uvFactor approximate the un-

derlying factors in (2.4) with 15 correct digits. In exact sense, however,
the perturbed polynomial p in (2.3) has only linear factors corresponding
to a cluster of simple roots. Figure 1 shows the root cluster computed by
Matlab function ROOTS, which is backward accurate in the sense that the
result consists of exact roots of a nearby pol5aiomial. D
Both examples show the effect of ill-posedness on inexact problems.
Since the solution structure can be completely altered by an infinitesimal
error in data, it may not be meaningful to compute the exact solution of a
perturbed ill-posed problem. Therefore, we need to formulate an approx-
imate solution for the given problem in order to remove the ill-posedness.
Using polynomial factorization as an example, we elaborate the formulation
process below.
Associating a general monic polynomial of degree 4 with a vector

p{x) = X'^ + PlX^ + P2X^ + PSX + Pi ~ [Pl>P2,P3,P4] e C^,

the set of all such polsmomials possessing the factorization of multiplicities

1 and 3

[x^ — (a + 3/3)X^ -I- 3(;9^ + a/3)X^ - (30/9^ + ll^)X + ap^

= {x-af{x-0f |a,/3eC,a7^/3}

denoted by 11(1,3) forms a manifold of codimension 2 with respect to

the metric topology induced by the norm ||p|| — \ / | p i p H h |P4P- On
the other hand, manifold n ( l , 3) is in the closure 11(1,1,2) of manifold
n ( l , l , 2 ) of codimension 1 since

hm [x - af{x - p + efix -13 ^ e'f = {x-afix-pf.

Likewise 11(1,1,2) C 11(1,1,1,1) = C^, and the five manifolds form a

stratification as shown in Figure 2.
EXAMPLES. When polynomial p e n ( l , 3 ) , say

p = x^ + Ax^ 16a;-16 = {x-2f{x + 2f

 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 153

is perturbed as p, say

P = x" + s.ggggx-'- i6x -16,

the original factoring structure (1,3) is lost since p € C — 11(1,1,1,1).
However, structure (1,3) can still be recovered from p since 11(1,3) is
the manifold of highest codimension passing through the £-neighborhood
of p for any e satisfying

0.0001 = lb - PII < £ < inf { | b - 9 | | I g e n ( 2 , 2 ) u n ( 4 ) } ,

say £ = 0.001. After identifying 11(1,3) this way, we define the ap-
proximate factorization of p within e as the exact factorization of
p = (x — aY{x — 13)^ that is the nearest polynomial to p in 11(1,3).
With this formulation, computing the approximate factorization of p is
a well-posed problem, and the roots a and $ oi p are approximation
to the intended roots a and j3 oi p with error bounded by a structure
preserving condition number [33]. We can verify this well-posedness using
ApaTools function uvFactor:
> F. r e s := uvFactor(x*4+3.9999*x"3-16»x-16,x,0.001):
> evalf(F);

.9999887784 (-2.000005124 + .9999999996 x ) (2.000004920 + 1.000000000 x)'*

with backward error | | p - p | | w 0.0000907 and the forward error 0.0000071

on the roots. The accuracy should be considered remarkable given the data
error of 0.0001. D
Generally, algebraic problems are often ill-posed because the set of
problems whose solutions possessing a distinct structure form a manifold of
positive codimension, and perturbations generically pushes a given prob-
lem away from the manifold. Our strategy starts with formulating the
approximate solution of an ill-posed algebraic problem following a "three
strikes" principle for removing the discontinuity:
Backward nearness: The approximate solution to the given (in-
exact) problem P is the exact solution of a nearby problem P
within £.
Maximum codimension: The nearby problem P is in the manifold
n possessing the highest codimension among all manifolds passing
through the £-neighborhood of P.
Minimum distance: Problem P is the nearest point in manifold
n to the given problem P
Here "the nearest point" may have a question mark in its uniqueness
at least theoretically. However, multiple occurrences of the identical min-
imum distance appear to be non-generic and we have not yet encountered
such cases in practical computation. At least, the nearest point is unique
when the backward nearness is sufficiently small [30].
154 ZHONGGANG ZENG

Following these principles, we can formulate the approximate polyno-

mial GCD [30], the approximate matrix Jordan Canonical Form [36], the
approximate factorization of polynomials [33], etc., as well-posed and of-
ten well-conditioned problems, so that it becomes feasible to calculate such
approximate solutions using floating point arithmetic. Based on those
formulation principles, computing the approximate solution can be carried
out in two optimization processes: maximizing the codimension of man-
ifolds followed by minimizing the distance to the manifold, leading to a
two-staged strategy for designing robust algorithms:
Stage I: Calculating the solution structure by finding the nearby
manifold 11 of highest codimension.
Stage II: Solving for the approximate solution by minimizing the
distance from given problem to manifold 11.
The main mechanism at Stage I is matrix rank-revealing, while Stage
II relies on solving nonlinear least squares problems.
We illustrate the two stage-approach using the univariate GCD com-
putation as an example. For given polynomials / and g of degrees
m > n > 0 respectively, the fc-th degree GCD manifold
- ^ r ' " = {(P'9) I deg{p) = m,deg(g) = n, deg(ja {p, q)) = k}
is a manifold of codimension k with respect to metric topology. When-
ever such a manifold passes through an e-neighborhood of (/, g), the fc-th
Sylvester matrix Skif,g) = [Cm-k{f),Cn-kig)] is rank-deficient by one in
approximate sense (c.f. convolution matrix in the next section). Conse-
quently, Stage I for finding the highest codimension GCD manifold nearby
is a successive rank-reveahng process on Sj{f,g) for j = n, n — 1, • • •
until the degree k is identified.
Once the GCD degree k is identified. Stage II becomes solving an
overdetermined quadratic system
<f>{u) --= 1
. , deg(u) = k, deg{t;) = m-k, deg{w) = n-k,
u-v == /
and a proper linear functional <>
/
u- w -= 9
for its least squares solution. The Gauss-Newton iteration (see §4) is
proven effective for this purpose.
The backward nearness threshold e to be provided by users generally
depends on the nature of the appHcation. The rule of thump is to set e
slightly larger than the magnitude of data error. More specifically, if one
knows the given (inexact) problem P is a small perturbation from the
underlying (exact) problem P that belongs to a manifold 11. Let S > 0
be the minimum distance firom P to manifold 11' 7^ 11 of higher or equal
codimension as 11. Then for ||P — P | | < £ < 5 , the three-strikes principle
ensures the correct manifold 11 to be identified, and the computed solution
of problem P will approximate the underlying solution of problem P .
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 155

Using the polynomial pair {f,g) in (2.2) as an example, its coefBcients

are 9-digits approximation to {f,g) in (2.1) with perturbation magnitude
1.1 X 10~^. Any other GCD manifold of equal or higher codimension is at
least 0.1 away from {f,g)- Thus any threshold e between 1.1x10"*
and 10~^ will ensure recovering the approximate GCD. In practice, s
should be set near the low end, say 10~^.
3. Matrix computation and matrix building tools. Finding the
structure of an approximate solution at Stage I almost always involve ma-
trix rank-revealing, whereas the minimization at Stage II can be accom-
plished using the Gauss-Newton iteration. For those considerations, the
base level tools in ApaTools include matrix builders, rank-revealing tools,
and the Gauss-Newton iteration routines.
Polynomials with a certain degree bound form a vector space V with
a monomial basis {pi,-- • ,Pn} arranged by a term order. The designated
term order can be considered a linear mapping
* : pk^4'4^---4' -^ efc € C"
that constitutes an isomorphism between V and C", where efc denotes
the fc-th canonical vector of C" (i.e. the fc-th column of the nxn identity
matrix). Let # i : Vi —-* C " and *2 : ^2 —> C" be isomorphisms
and £ : Vi —> P2 be a hnear transformation. Then there is a matrix
yl g £nxm |.jjg^|. makes the following diagram commute:

Pi — ^ V2
I*-
C" A

Matrix >1 = [ai, • • • , am] can then be generated column-by-column in the

loop as follows:
for j = 1,2,- • • ,m do
- p = ^iHej)
- q = C{p)
- aj = *2(g)
end do
Here e^ is the j - t h canonical vector in C " , namely the j-th
column of the m x m identity matrix. Implementation of software tools
for isomorphisms between polynomial vector spaces and C" constitutes
matrix building tools in ApaTools.
EXAMPLE 4. For instance, let V'' denote the vector space of
polynomials with total degree less than or equal to k. For a fixed poly-
nomial / , polynomial multiplication with / is a linear transformation
£ / : V" —^V"
Cfig) = f-g eV^ for all 5 6 P " (3.1)
156 ZHONGGANG ZENG

with m = deg(/) + n that corresponds to a convolution matrix Cn{f)- In

ApaTools, we include a convenient function LinearTransformMatrix as a
generic matrix builder for any linear transformation between vector spaces
of polynomials. The user needs to provide a subroutine for the linear
transformation as input for LinearTransf ormMatrix. Using the linear
transformation in (3.1) as an example, we write a simple Maple procedure
for Cf with / being an input parameter:
> PolynMultiply := proc( g::polynom, x::{name,list}, f:ipolynom);
return expand(f*g)
end proc:

The convolution matrix Ciif) for / ( x , y) = x + 2y + 3 can then be

generated by a simple call:
> T := LinearTransformMatrixCPolynMultxply,[x+2*y+3],[x,y],2,3);

3 0 0 0 0 0

1 3 0 0 0 0

2 0 3 0 0 0

0 1 0 3 0 0

0 2 1 0 3 0

0 0 2 0 0 3

D
In this example, the input PolynMultiply is the procediure name
for Cf, item [x+2*y+3] is the Ust of parameters for PoljmMultiply
besides [x,y] that stands for the list of indeterminates, and 2, 3 are
the degree bounds for the vector spaces V^ and 'P" respectively. The
graded lexicographical monomial order is used here as the default monomial
ordering.
With matrices being constructed, rank-revealing is frequently applied
in approximate polynomial algebra. As an example, a polynomial pair
(/i 9) G -^"^ X ^ " having a nontrivial GCD of degree k can be written
as f = uv and g = uw where u — gd (/, g) with cofactors v and w.
Consequently, polynomials v and ty satisfy fw — gv = 0, or equivalently
a homogeneous system of linear equations

w
Cn-k{f), Cm-kig) —V

where v and w are coefficient vectors of v and w respectively. As

a result, finding the GCD structure is equivalent to computing the rank of
the Sylvester subresultant matrix [C„-fc(/),Cm-fe(p)] for k decreasing
from min{m, n} until rank-deficiency occurs.
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 157

Another example is numerical elimination. For given polynomials /

and g in variables x and y, if there are polynomials p and q such
that variable x is eliminated from polynomial h= p- f + q • g, then

dxipf + qg) = [{dj) + f • da,]p + [(9x5) +9-d,]q = 0. (3.2)

Since mapping p —> [{dxf) + f-dx]p is a linear transformation, equation

(3.2) can be converted to a homogeneous system of linear equations in
matrix-vector form

M P 0
q
that becomes a rank-revealing and kernel-finding problem for the elimina-
tion matrix M.
In approximate polynomial algebra, we seek the approximate rank and
the approximate kernel of a matrix in contrast to seeking exact rank and
kernel in exact polynomial algebra. Following the same "three-strikes"
principle, the approximate rank and kernel of a matrix A € C"*"" are
the exact rank and kernel of a nearby matrix B € C""*". This matrix
B is the nearest matrix to A on the manifold Ilfc that possesses the
highest codimension among manifolds n i , n 2 , - - - passing through an e-
neighborhood of A, where Hj is the set of all mxn matrices with rank
j . Approximate rank/kernel can be efficiently computed using numerical
rank-revealing algorithms [16, 18] that are implemented as components of
ApaTools.
4. Nonlinear least squares and the Gauss-Newton iteration.
The minimization at Stage II is a nonlinear least squares problem that can
be solved using the Gauss-Newton iteration on an overdetermined system
of equations in the form of

f(z) = 0 for f : C " — > C " , z e C", m>n. (4.1)

We seek the least squares solution z satisfying

l|f(^)f = ii|P„{l|f(^)f} (4.2)

with a choice of || • || to be either weighted or straightforward Euclidean
norm. Let J(z) be the Jacobian of f(z). Then the minimum occurs at
z such that [33]

J(z)*f(z) = 0 (4.3)

where (•)* denotes the Hermitian transpose of a matrix or vector (•).

If f (z) is analj^ic, the Jacobian J(z) is injective so that the Moore-
Penrose inverse J(z)+ = (J(z)*J(z)) J(z)*, the minimum ||f(z)|| is
158 ZHONGGANG ZENG

small, and the initial iterate ZQ is close to z, then the Gauss-Newton

iteration

Zfe+i = Zfc - J(zfc)+f(zfc), A; = 0,1,- (4.4)

converges to z [33].
The Gauss-Newton iteration is extensively used in ApaTools. A typi-
cal case is computing the approximate GCD: For given polynomial pair p
and q with degrees m and n respectively, we seek a polynomial triplet
(u, V, w) such that deg(w) = k, deg{uv) = m, deg{uw) = n, and

\\{p,q) - {uv,uw) min {\\iP,Q) {uv,uw)\\^.

deg(u) = k
d e g ( « v ) = jn
deg(uiu) = n

Let p , q, u, V, and w be the coefficient vectors of p, q, u, v, and

w respectively, then the overdetermined system is constructed to have the
least squares solution to uv—p = 0 and uv — q = 0 along with a proper
scaling equation on u. This system can be written in matrix-vector form
as

r*u-l
f(u, v,w) = 0 for f(u, V, w) Ck{v)u-p (4.5)
Ck{w)u-q

where Cj{h) denotes the convolution matrix corresponding to the linear

transformation Ch '• 9 —* h-g GV^ on the vector space V^ (see §3) and
r is called a scaUng vector. The choices of r is random or parallel to
the initial estimate of u, which makes no noticeable difference in practical
computation either way. The Jacobian can be easily constructed as

7(u,v,w) Ckiv) Cm-kiu)

Ck{w) Cn-kiu)

The Gauss-Newton iteration (4.4) can thus be applied accordingly. The

auxilliary equation r*u—1 = 0 is crucial in making J(u, v , w ) injective.
It is essential to formulate the overdetermined system f(z) = 0
with enough equations such that the Jacobian J(z) is injective at the
least squares solution z and, as a result, its Moore-Penrose inverse
J(z)+ = (J(z)* J(z)) J(z)*, ensuring the Gauss-Newton iteration to be
well defined in a neighborhood of z. Moreover, the norm || J(z)"*" || serves
as a condition number of the approximate solution [30, 33, 36].
We provide a generic blackbox function GaussNewton for the general
purpose Gauss-Newton iteration. This function requires the user to pro-
vide a routine for computing function f(z) in (4.1), an optional routine
for computing the Jacobian J(z) and an initial iterate ZQ as its input.
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 159

Then GaussNewton carries out the Gauss-Newton iteration and outputs the
least squares solution z along with the residual ||f(z)||2- As a conve-
nient option, the Maple version of GaussNewton can forgo the requirement
of providing Jacobian routine by computing J(z) with Maple symbolic
manipulation.
EXAMPLE 5. Using the above example of GCD computation, the
generic Gauss-Newton function needs only a user-defined subroutine for
computing f (z) that can be as simple as
> GcdFunc := procC z, m, n, p, q, r)
local F, u, V, w;
u, V, w := z [ l . . r a ] , z[m+1..m+n], z [ m + n + l . . - 1 ] ;
F := <LinearAlgebraCDotProduct3(r,u)-l, C o n v o l u t i o n ( u , v ) - p ,
Convolutioii(u,w)-q>:
return F [ l . . - 1 , 1 ] ;
end p r o c :

Here, the Maple routine GcdFunc returns the vector value of f(z)
from input vector z consists of coefficients of u, i; and w, along with
parameters m, n, p , q, r representing the length of u, the length
of V, coefficient vectors p, q and r, respectively. The command
Convolution(u,v) and Convolution(u,w) produces coefficient vectors
of polynomial products u • v and u • v respectively by ApaTools function
Convolution. Then the Gauss-Newton iteration is carried out by a simple
call of GaussNewton:
> # get coef. vectors p, q
> p, q := PolynomialToolsiCoefficientVector]{x"5-x"3+5*x"2-6*x+10,x),
PolynomialTools[CoefficientVector](2*x"4+x"2-6,x);
> r := VectorC[.4,0,.23): # define scaling vector r
> zO := Vectored.99,0.01,1.01, # initial guess for u, v, «
4.99,-3.01,0,.99,
-3.01,0,1.99]):
> # Gauss-Newton iteration
> z,res := GaussNewton(GcdFunc,zO, [3,4,p,q,r] , [le-9,9,true]) :
Gauss-Kestoo step 0, rasidual - S.OOa-02
Gauss-Keutoa step 1, residual " 2.01e-04
Gauss-liewton step 2, residual - 3.16e-09
Gauss-Henton step 3, residual * 4.44a-t6

The approximate GCD and cofactors can then be retrieved from the
result jjf the iteration:
> CoefficientVector2UnivariatePolynoiiiial(z[l. .3] ,x) ;
CoefficientVector2tJnivariatePolynomial(zC4. .7] , x ) ;
CoefficientVector2UnivariatePolynomial(zC8..-1],x);

2.000000000000000 + 3.1779793 X 1 0 ~ 1 ^ x + 1.000000000000000 x^

4.999999999999999 - 3.000000000000000 I - 1.530197036 X 1 0 ~ ' ^ x^ + 0.9999999999999998 x ^

-2.999999999999999 + 2.54808329 X 1 0 ~ ^ ^ x + 2.000000000000000 x^

that are accurate approximation to the exact GCD u = x^ + 2, cofactors

V = x^ ~3x + 5 and w = 2x^ — 3. Notice that the Jacobian routine is not
necessarily an input item for GaussNewton. This type of generic routines
enables fast implementation of experimental algorithms. D
160 ZHONGGANG ZENG

Example 5 illustrates the most basic application of GaussNewton for

computing the least squares zero to a vector-valued function / ( z ) . In
approximate polynomial algebra, we often encounter a task of finding the
least squares zero of a function whose domain and range are lists of polyno-
mials. The function GaussNewton provides another convenient feature of
direct computation of the least squares zero in terms of a list of polynomials
as shown below.
EXAMPLE 6. Considering the same GCD problem and using the same
polynomials as in Example 5, computing gcd{f,g) is equivalent to finding
the least squares zero of the function

F{pi,P2,P3) = [ r * p i - l , pi-P2-f, Pi-Ps-g]- (4.6)

whose domain and range are both lists of polynomials. To prepare for
applying GaussNewton, we write a simple Maple procedure for the function
F in (4.6) in a straightforward translation:
> GcdPolynFunc ;= proc( p. x, f, g, r ) # procedure for the gcd function
return [ PolynomialDotProduct(r,p[13,x) - 1,
expand(p[l]*p[2])-f, expand(p[l]*pC3])-g ]
end proc:

We then set the input items and call GaussNewton in an intuitive

manner:
> f, g := x"5-x"3+5*x"2-6*x+10, 2*x~4 + x~2 - 6: # define input polynomial pair
> r := .4 + .2*x*2: # scaling polynomial
> u, V, w := 1.99 + 1.0i*x~2, # set initial guesses
4.99 - 3.01*x +0.99*x-3,
-3.01 + 1.99*x-2:
> s . r e s := GaussNewtonCGcdPolynFunc,[u,v,w3,[[x],f,g,r],Cle-i2,9,true3)

Gauss-Newton step 0, residual = 1,.34e-01

Gauss-Kewton step 1, residual = 2,,44e-04
Gauss-Newton step 2, residual = 2,.66e-09
Gauss-Newton step 3, residual = 3.,75e-15

> s [ i ] / l c o e f f ( s [ l ] ) ; s [ 2 ] / l c o e f f ( s [ 2 ] ) ; 2*s[33/lcoeff ( s [ 3 ] ) ; # show r e s u l t

2.000000000000000 + 1.000000000000000 x^
4.999999999999999 - 2.999999999999999 x + 0.9999999999999997 x^
-3.000000000000000 + 2.000000000000000 x^
This feature of GaussNewton enables direct manipulation of
polynomials and avoids the tedious task of translation between polyno-
mials and vectors. D
R e m a r k . Gauss-Newton iteration locally converges to a point satis-
fying (4.3) which is the necessary condition for a local minimum in (4.2).
A global minimum is not guaranteed in general. As a matter of fact, a
global minimum is difficult to verify, let alone to compute with any cer-
tainty. Even though a proposed method in [15] is claimed to have the
capability of finding the global minimum in polynomial time, that claim is
neither proved rigorously nor backed up with an implementation. On the
other hand, the backward accuracy of the computing result can be verified
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 161

by the actual backward nearness whose measurement is a by-product of

the Gauss-Newton iteration (i.e. the output r e s of GaussNewton). Conse-
quently, the lack of guarantee in finding global minimum does not appear
to be a serious concern in practical computation, or should not be until
such an application arises. Furthermore, it can be proved that the mini-
mum is indeed global if the given (inexact) problem is sufficiently close to
the underlying exact problem [30]. D
5. ApaTools overview. We give an itemized overview of our package
ApaTools on its functionality currently available.
1. Base level tools
• LinearTransf ormMatrix: A generic function for construct-
ing the matrix associated with a given linear transformation
between vector spaces of polynomials, as described in §3.
• GaussNewton: A generic function for carrying out the Gauss-
Newton iteration for a given overdetermined system of non-
hnear equations as described in §4.
2. M a t r i x c o m p u t a t i o n tools
• NulVector: The function for computing the smallest singu-
lar value and the corresponding singular vectors. In many
cases, it is desirable to determine if a matrix is approximately
rank-deficient without computing a full-blown singular value
decomposition. Function NulVector is designed for this pur-
pose. It also serves as a subroutine for ApproxiRank that
computes the approximate rank and approximate kernel.
• ApproxiRank: The function for computing the approximate
rank ranke ( A ) and approximate kernel /Ce ( J4 ) of a matrix
A within a given threshold 9. Here the approximate rank
is defined as

rankg (A) = min rank(B)

\\B-A\\2<e

and the approximate kernel /Cg ( A ) = /C ( B ) for

||B-A||2 = min ||C-A||2.

rank(C)=rank9( ^ )

Function ApproxiRank is eflScient when the nullity of A is

low. The standard singular value decomposition (SVD) may
be more suitable if the approximate rank of A is around one
half of the column dimension.
• ApproxiJCF: The function for computing the approximate
Jordan Canonical Form (JCF) of a given matrix A within a
threshold e. Computing the exact Jordan Canonical Form is
162 ZHONGGANG ZENG

an ill-posed problem that requires exact data with sjmibolic

computation. The approximate JCF is defined according to
the same "three strikes" principle and computed with a two-
staged algorithm elaborated in §2. As a result, ApproxiJCF
is capable of computing the Jordan canonical decomposition
accurately even if the matrix is perturbed [36].
3. Univariate polynomial tools:
• uvGCD: The function for computing the approximate greatest
common divisor, denoted by gcdg{f,g), of univariate poly-
nomial pair (/, g) within a given threshold 9 with definition
as follows [30].
Let V be the vector space of polynomial pairs (p, q) sat-
isfying deg(p) < deg(/) and deg(Q) < d.eg{g). Then
IIj = {(p,g) € V I deg{£cd{p,q)) = j } is a manifold of
codimension j in V associated with the metric topology in-
duced by the vector 2-norm for j = 0,1, • • •. Let 11^ be the
highest codimension manifold among Ho, Hi, • • • containing
a polynomial pair within distance 6 of {f,g)- Then the
approximate GCD gcd ^ (/, fl") is the exact GCD of pair {p, q),
where (p, q) is the nearest polynomial pair on the manifold
life to the given pair {f,g).
Function uvGCD accepts input polynomial pair {f,g) and
threshold 0. The user can choose to omit providing 6 value
in Matlab, or set it to be zero in Maple, to invoke the default
9 that is cube root of the square of the machine epsilon.
The output consists oi u = s'^dg{f,g), cofactors v and
w along with the residual ||(/,p) — {uv,uw)\\, where the
norm || • || is either the vector 2-norm chosen by users or
the default weighted 2-norm. The input polynomials are not
required to be monic and will be taken "as is" without scaling.
The output GCD and cofactors are generally not monic for
avoiding unbalanced scaling to the computing result.
There are several proposed algorithms in the literature for
computing the univariate approximate GCD, such as [5, 9, 13,
15, 23]. Our uvGCD is significantly more accurate and robust
than existing implementation due to the our employment of
the iterative refinement at Stage II.
• ExtendGCD: The function for computing a polynomial pair
(p, g) such that pf+ qg = gcdg{f,g) for given polynomial /
and g using the output of uvGCD. The extended GCD can be
applied to transforming matrices of polynomial entries, and
in particular, to computing the Smith Normal Form [29, 35].
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 163

• uvFactor: The function for computing an approximate fac-

torization

P0{X - Zir^ {X - Z2r' •••{X- ZkT"

for a given polynomial p(a;), as shown in Example 2. Exclud-

ing certain pathologically ill-conditioned cases, the factors of
polynomials with nontrivial multiplicities rrij > 1 can be
calculated accurately by uvFactor without extending the ma-
chine precision even if the coefficients of p{x) are perturbed
to certain extent [33).
4. Multivariate polynomial tools
• mvGCD: The function for computing the approximate GCD
of a given pair (/, 5) of multivariate polynomials. The
formulation of the multivariate approximate GCD is similar
to the univariate case. The computation requires applying
uvGCD repeatedly on the univariate polynomial pairs pro-
jected from (/, g) in determining the GCD structure before
calling GaussNewton for solving a least squares problem that
minimizes the distance from the given polynomial pair to a
GCD manifold. For details, see [34].
• Squaref reePactor: The function for computing an approx-
imate squarefree factorization of a given multivariate polyno-
mial p. Here, again, the notion of the approximate squarefree
factorization is formulated using the "three-strikes" principle.
Function Squaref r e e P a c t o r produces two types of squarefree
factorizations: a staircase squarefree factorization

P = {PlY{P2?---{Pkf

where pi, • • • ,Pk are coprime, or a flat-type squarefree fac-

torization

P = / i • /2 • • • /fc

where fj = PjPj+i •• -Pk for j = 1,- •• ,k, and ft divides

fj for all i>j. Each pi or fj is "squarefree", namely it
has no repeated nontrivial factors of its own.
• M u l t i p l i c i t y S t r u c t u r e : The function for computing the
multiplicity structure of a given polynomial system

/ i ( x i , - - - ,Xs) = 0
: (5.1)
ft{xi,--- ,a;s) = 0
164 ZHONGGANG ZENG

at a given zero x* = (xj, • • • ,x*).

Let dj denote a differential monomial

A vector of complex numbers a = [aj | j € N*] corresponds

to a differential functional a[x*] defined as

for any polynomial / in the ideal X = (/i,--- , / t ) . The

vector space

Dx.(I) = {a[x*] I a[x*](/) = 0 for all / e (/i, • • • , /*)}

is called the dual space of the ideal I at x*. The dimension

of the dual space I>x* (X) is the multiplicity of x* as a zero
to the system (5.1). The dual space itself constitutes the
multipUcity structure of the system (5.1) at zero x* [27].
The function M u l t i p l i c i t y S t r u c t u r e calculates the multi-
plicity, a basis for the dual space along with other invariants
such as breadth, depth and Hilbert function [7] even if the
system and zero are inexact.
For closely related algorithms, see [2, 20].
• PolynomialEliminate: The function for computing polyno-
mials p, q and h such that h = pf + qg belongs to a
specified elimination ideal generated by polynomials / and
g. In other words, PolynomialEliminate eliminates a spec-
ified variable Xj in h — pf + qg for given / and g in
variables xi,---,Xs, by solving the differential equation

^ ( p / + .5) = 0

for p and q via rank-reveahng tool ApproxiKernel on a

sequence of matrices generated by the matrix building tool
LinearTransf ormMatrix. Combined with mvGCD, this ehmi-
nation tool is particularly useful in solving polynomial sys-
tems whose solutions contain nonzero dimensional compo-
nents [31].
6. A brief comparison to existing software packages. For a sim-
ilar purpose, a software package SNAP (Symbohc-Numeric Algorithms for
Polynomials) [12] is included in recent releases of Maple. The scope
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 165

of ApaTools is much broader than current SNAP in functionahty. In

short, SNAP consists of only ten functions with three univariate GCD pro-
cedures (EpsilonGCD, QuasiGCD and QRGCD) with nearly identical objec-
tives, along with seven utility routines. In comparison, ApaTools' matrix
rank-revealing, univariate factorization, multivariate GCD, squarefree fac-
torization, multiplicity structure, polynomial elimination, etc. are original
developments. For the univariate GCD function that is included in both
ApaTools and SNAP, our uvGCD is more advanced in robustness and accu-
racy due to the iterative refinement at Stage II. A detailed comparison
on univariate GCD computation can be found in [30]. Our approach of
building on matrix building and least squares solving is also novel in
software design.
On Matlab platform, there appear to be no other packages that are
comparable to Apalab in functionality and comprehensiveness.
7. Future development. ApaTools is an on-going project. While
continuing to refine the existing functions, we plan to expand the package
by developing more algorithms and their implementations for various prob-
lems in approximate polynomial algebra. For instance, a project is under-
way for developing a numerical algorithm for computing the approximate
irreducible factorization of multivariate polynomials. The "three-strikes"
principle is again applied to formulate the notion of the approximate ir-
reducible factorization, and the algorithm follows the two-staged strategy.
Moreover, the algorithm and its implementation will be built on top of the
existing uvGCD, mvGCD, SquarefreeFactor, and ApproxiJCF.
Acknowledgements. This work is originated and partially completed
during the author's six week participation in the Thematic Year on Appli-
cation of Algebraic Geometry at the Institute for Mathematics and Its Ap-
pUcations (IMA) in 2006. The author is grateful to IMA for its refreshing
research environment.

REFERENCES

[1] D. BATES, J.D. HAUENSTERN, A.J. SOMMESE, AND C . W . WAMPLBR,

Bertini: Software for Nmerical Algebraic Geometry, http://www.nd.edu/
~sommese/bertini, 2006.
[2] D . J . BATES, C . P E T E R S O N , AND A.J. SOMMESE, A numerical-symbolic algorithm
for computing the multiplicity of a component of an algebraic set, J. of Com-
plexity, 2 2 (2006), pp. 475-489.
[3] L. BuSE, I. EMIRIS, AND B . MOURRAIN, Multires. http://www-sop.inria.fr/
galaad/mourrain/multires.htmL
[4] G. CHEZE AND A . GALLIGO, Four lessons on polynomial absolute factorization.
Solving Polynomial Equations: Foundations, Algorithms, and Applications, A.
Dickenstein and I.Z. Emiris, eds, Algorithms and Computation in Mathematics
14, Springer-Verlag (2005), pp. 339-392.
[5] R . M . CORLESS, S . M . WATT, AND L . ZHI, QR factoring to compute the GCD of uni-
variate approximate polynomials, IEEE Trans. SignaJ Processing, 5 2 (2003),
pp. 3394-3402.
166 ZHONGGANG ZENG

[6] B. DATTA AND K . DATTA, Toeplitz algorithms for controllability, GCD and Cauchy
index. Proceedings of t h e American Control Conference.
[7] B. DAYTON AND Z . ZENG, Computing the multiplicity structure in solving polyno-
mial systems. Proceedings of ISSAC '05, ACM Press, pp. 116-123, 2005.
[8] I.Z. EMIRIS, E . D . FRITZILAS, AND D . MANOCHA, Algebraic algorithms for deter-
mining structure in biological chemistry, Internatinal J. Quantum Chemistry,
106 (2006), pp. 190-210.
[9] I.Z. EMIRIS, A. GALLIGO, AND H . LOMBARDI, Certified approximate univariate
GCDs, J. Pure Appl. Algebra, 1 1 7 / 1 1 8 (1997), pp. 229-251.
[10] S. G A O , E . KALTOFEN, J. M A Y , Z . YANG, AND L . ZHI, Approximate factorization
of multivariate polynomials via differential equations. Proc. ISSAC '04, ACM
Press, pp. 167-174, 2004.
[11] T. G A O AND T . - Y . L I , MixedVol: A software package for mixed volume computa-
tion, ACM Trans. Math. Software, 3 1 (2005), pp. 555-560.
[12] C.-P. JEANNEROD AND G . LABAHN, The SNAP package for arithemetic with nu-
meric polynomials. In International Congress of Mathematical Software, World
Scientific, pp. 61-71, 2002.
[13] E. KALTOFEN, J. MAY, Z . YANG, AND L . ZHI, Structured low rank approximation of
Sylvester matrix, in Symbolic-Numeric Computation, Trends in Mathematics,
D. Wang and L. Zhi, editors, Birkhauser Verlag, Basel, Switzerlajid (2007),
pp. 69-83.
[14] N. KARCANIAS AND M . MITROULI, A matrix pencil based numerical method for the
computation of the GCD of polynomials, IEEE Trans, on Automatic Control,
39 (1994), pp. 977-981.
[15] N.K. KARMARKAR AND Y.N. LAKSHMAN, On approximate polynomial greatest com-
mon divisors, J. Symb. Comput., 26 (1998), pp. 653-666.
[16] T.-L. L E E , T . Y . L I , AND Z . ZENG, A rank-revealing method with updating, down-
dating and applications. Part 11. submitted, 2006.
[17] T.-Y. L I , Solving polynomial systems by the homotopy continuation method. Hand-
book of Numerical Analysis, XI, edited by P.G. Cizirlet, North-Holand, Ams-
terdam (2003), pp. 209-304.
[18] T.Y. L I AND Z . ZENG, A rank-revealing method mth updating, downdating and
applications, SIAM J. Matrix Anal. Appl., 26 (2005), pp. 918-946.
[19] J . P . MERLBT, Singular configuration of parallel manipulators and Grossman ge-
ometry, Int. J. Robitics Research, 8 (1989), pp. 45-59.
[20] B. MOURRAIN, Isolated points, duality and residues, J. of Pure and Applied Alge-
bra, 117 & 118 (1996), pp. 469-493. Special issue for the Proc. of the 4th
Int. Symp. on Effective Methods in Algebraic Geometry (MEGA).
[21] H.-T. P A I , A. BoviK, AND B . EVANS, Multi-channel blind image restoration,
TUBITAK Elektrik J. of Electrical Eng. and Comput. Sci., 5 (1997),
pp. 79-97.
[22] S. PILLAI AND B . LIANG, Blind image deconvolution using GCD approach, IEEE
TVans. Image Processing, 8 (1999), pp. 202-219.
[23] D. RUPPRECHT, An algorithm for computing certified approximate GCD of n uni-
variate polynomials, J. Pure and Appl. Alg., 1 3 9 (1999), pp. 255-284.
[24] A.J. SoMMBSE, J. VERSCHELDE, AND C . W . WAMPLER, Numerical factorization of
multivariate complex polynomials, Theoretical Computer Science, 3 1 5 (2003),
pp. 651-669.
[25] A.J. SoMMESE AND C.W. WAMPLER, The Numerical Solution of Systems of Poly-
nomials, World Scientific Pub., Hackensack, NJ, 2005.
[26] B . M . S T - O N G E AND C M . GOSSELIN, Singularity analysis and representation
of general Gough-Stewart platform, Int. J. Robitics Research, 19 (2000),
pp. 271-288.
[27] H.J. STETTBR, Numerical Polynomial Algebra, SIAM, 2004.
 APATOOLS FOR APPROXIMATE POLYNOMIAL ALGEBRA 167

[28] J. VERSCHELDE, Algorithm 795: PHCpack: A general-purpose solver for polyno-

mial systems by homotopy continuation, ACM Treins. Math. Software (1999),
pp. 251-276.
[29] J. VERSCHELDE AND Y . WANG, Computing dynamic output feedback laws, IEEE
Trans. Automatic Control (2004), pp. 1552-1571.
[30] Z. ZENG, The approximate GCD of inexact polynomials. Part I. Preprint, 2007.
[31] Z. ZENG, A polynomial elimination method for symbolic and numerical computa-
tion. Preprint, 2007.
[32] Z. ZENG, Algorithm 835: Multroot - A Matlab package for computing polynomial
roots and multiplicities, ACM Trans. Math. Software, 3 0 (2004), pp. 218-235.
[33] , Computing multiple roots of inexact polynomials, Math. Comp., 74 (2005),
pp. 869-903.
[34] Z. ZENG AND B . DAYTON, The approximate GCD of inexact polynomials. II: A
multivariate algorithm. Proceedings of ISSAC'04, ACM Press, pp. 320-327
(2006).
[35] , The approximate GCD of inexact polynomials. Part II. Preprint, 2007.
[36] Z. ZENG AND T . Y . L I , A numerical method for computing the Jordan Canonical
Form. Preprint, 2007.
 LIST OF W O R K S H O P PARTICIPANTS

• Douglas N. Arnold, Institute for Mathematics and its Applications,

University of Minnesota Twin Cities
• Donald G. Aronson, Institute for Mathematics and its Applica-
tions, University of Minnesota Twin Cities
• Daniel J. Bates, Institute for Mathematics and its Apphcations,
University of Minnesota Twin Cities
• Gian Mario Besana, Department of Computer Science-Telecom-
munications, DePaul University
• Anna M. Bigatti, Dipartimento di Matematica, Universita di Gen-
ova
• Rachelle Bouchat, Department of Mathematics, University of Ken-
tucky
• Massimo Caboara, Dipartimento di Matematica, Universita di
Genova
• lonut Ciocan-Fontanine, Institute for Mathematics and its Appli-
cations, University of Minnesota Twin Cities
• Barry H. Dayton, Department of Mathematics, Northeastern Illi-
nois University
• Wolfram Decker, Fachrichtung Mathematik, Universitat des Saar-
landes
• Alicia Dickenstein, Departamento de Matematica, Facultad de
Ciencias Exactas y Naturales, Universidad de Buenos Aires
• Kenneth R. Driessel, Department of Mathematics, Iowa State Uni-
versity
• Makan Fardad, Department of Electrical and Computer Engineer-
ing, University of Minnesota Twin Cities
• Xuhong Gao, Department of Mathematical Sciences, Clemson Uni-
versity
• Jiirgen Gerhard, Department of Research and Development,
Maplesoft
• Jason E. Gower, Institute for Mathematics and its Applications,
University of Minnesota Twin Cities,
• Daniel R. Grayson, Department of Mathematics, University of Illi-
nois at Urbana-Champaign
• Genhua Guan, Department of Mathematics, Clemson University
• Yun Guan, Department of Mathematics, Statistics and Computer
Science, University of Illinois at Chicago
• Tatsuyoshi Hamada, Department of Applied Mathematics,
Fukuoka University
• Marshall Hampton, Department of Mathematics and Statistics,
University of Minnesota
169
170 LIST OF WORKSHOP PARTICIPANTS

• Gloria Haro Ortega, Institute for Mathematics and its Applica-

tions, University of Minnesota Twin Cities
• Jonathan D. Hauenstein, Department of Mathematics, University
of Notre Dame
• Raymond Hemmecke, Otto-von-Guericke-Universitat Magdeburg
• Milena Hering, Institute for Mathematics and its Applications,
University of Minnesota Twin Cities
• Benjamin J. Howard, Institute for Mathematics and its Applica-
tions, University of Minnesota Twin Cities
• Evelyne Hubert, Project CAFE, Institut National de Recherche en
Informatique Automatique (INRIA)
• Farhad Jafari, Department of Mathematics, University of Wyo-
ming
• Abdul Salam Jarrah, Virginia Bioinformatics Institute, Virginia
Polytechnic Institute and State University
• Anders Nedergaard Jensen, Institut fiir Mathematik, Technische
Universitat Berlin
• Gabriela Jeronimo, Departamento de Matematica - FCEyN, Uni-
versity of Buenos Aires
• Masakazu Kojima, Department of Mathematical and Computing
Sciences, Tokyo Institute of Technology
• Song-Hwa Kwon, Institute for Mathematics and its Applications,
University of Minnesota Twin Cities
• Niels Lauritzen, Institut for Matematiske Fag, Aarhus University
• Gregoire Lecerf, Laboratoire de Mathematiques, Universit Ver-
sailles/Saint Quentin-en-Yvelines
• Anton Leykin, Institute for Mathematics and its Applications, Uni-
versity of Minnesota Twin Cities
• Hstau Y. Liao, Institute for Mathematics and its AppUcations,
University of Minnesota Twin Cities
• Gennady Lyubeznik, School of Mathematics, University of Min-
nesota Twin Cities
• Diane Maclagan, Department of Mathematics, Rutgers University
• .Peter Nicholas Malkin, CORE, Centre for Operations Research
and Econometrics, Universit Catholique de Louvain
• Hannah Markwig, Institute for Mathematics and its Applications,
University of Minnesota Twin Cities
• Thomas Markwig, Department of Mathematics, Universitat Kai-
serslautern
• Andrew A. McLennan, Department of Economics, University of
Minnesota Twin Cities
• Richard B. Moeckel, School of Mathematics, University of Min-
nesota Twin Cities
• Bernard Mourrain, Project GALAAD, Institut National de Recher-
che en Informatique Automatique (INRIA)
 LIST OF WORKSHOP PARTICIPANTS 171

Uwe Nagel, Department of Mathematics, University of Kentucky

Jiawang Nie, Institute of Mathematics and its Apphcations, Uni-
versity of Minnesota Twin Cities
Michael E. O'SuUivan, Department of Mathematics and Statistics,
San Diego State University
Rohit Pandita, Department of Aerospace Engineering and Mechan-
ics, University of Minnesota Twin Cities
Antonis Papachristodoulou, Engineering Science University of Ox-
ford
Pablo A. Parrilo, Laboratory for Information and Decision Sys-
tems, Massachusetts Institute of Technology
Chris Peterson, Department of Mathematics, Colorado State Uni-
versity
Sonja Petrovic, Department of Mathematics, University of Ken-
tucky
Kathy Wei Piret, Department of Mathematics, Statistics and Com-
puter Science, University of IlUnois
Sorin Popescu, Department of Mathematics, SUNY
Jacob Quant, University of Minnesota Twin Cities
Gregory J. Reid, Department of Applied Mathematics, University
of Western Ontario
Victor Reiner, School of Mathematics, University of Minnesota
Twin Cities
Joel Roberts, School of Mathematics, University of Minnesota
Twin Cities
Fabrice Rouillier, Projet SALSA, Institut National de Recherche
en Informatique Automatique (INRIA)
Bjarke Hammersholt Roune, Department of Mathematics, Aarhus
University
David Rusin, Department of Mathematical Sciences, Northern lUi-
nois University
Arnd Scheel, Institute for Mathematics and its Applications, Uni-
versity of Minnesota Twin Cities
Hans Schoenemann, Department of Mathematics Universitat
Kaiserslautern Eric Schost, LIX, Ecole Polytechnique
Mathias Schulze, Department of Mathematics, Oklahoma State
University
Robin Scott, Department of Applied Mathematics, University of
Western Ontario
Chehrzad Shakiban, Institute of Mathematics and its Applications,
University of Minnesota Twin Cities
Donald H. Singley, 3M
Andrew J. Sommese, Department of Mathematics, University of
Notre Dame
172 LIST OF WORKSHOP PARTICIPANTS

Steven Sperber, School of Mathematics, University of Minnesota

Twin Cities
Dumitru Stamate, School of Mathematics, University of Minnesota
Twin Cities
William Stein, Department of Mathematics, University of Wash-
ington
Brandilyn Stigler, Mathematical Biosciences Institute
Michael E. Stillman, Department of Mathematics, Cornell Univer-
sity
Erik Stokes, University of Kentucky
Nobuki Takayama, Department of Mathematics, Kobe University
Akiko Takeda, Department of Mathematical and Computing Sci-
ences, Tokyo Institute of Technology
AmeUa Taylor, Colorado College Enrique Augusto Tobis, Com-
puter Science and Mathematics, University of Buenos Aires
Carl Toews, Institute for Mathematics and its Applications, Uni-
versity of Minnesota Twin Cities
Ehas P. Tsigaridas, project GALAAD, INRIA Sophia AntipoUs
Theodore L. Turocy, Department of Economics, Texas A&M Uni-
versity
Balint Vanek, Department of Aerospace Engineering and Mechan-
ics, University of Minnesota Twin Cities
Jan Verschelde, Department of Mathematics, Statistics and Com-
puter Science, University of Ilhnois at Chicago
John Voight, Institute for Mathematics and its Applications, Uni-
versity of Minnesota Twin Cities
Charles W. Wampler II, General Motors Research and Develop-
ment
Stephen M. Watt, Department of Computer Science, University of
Western Ontario
Thomas Wolf, Department of Mathematics, Brock University
Wenyuan Wu, Department of Applied Mathematics, University of
Western Ontario
Zhonggang Zeng, Department of Mathematics, Northeastern Ilh-
nois University
Ailing Zhao, Department of Mathematics, Statistics and Computer
Science, University of Illinois
Ke Zhou, Electrical and Computer Engineering Department, Uni-
versity of Minnesota Twin Cities
Yan Zhuang, Department of Mathematics, Statistics and Com-
puter Science, University of Ilhnois
1997-1998 Emerging Applications of Djmamical Systems
1998-1999 Mathematics in Biology
1999-2000 Reactive Flows and Transport Phenomena
2000-2001 Mathematics in Multimedia
2001-2002 Mathematics in the Geosciences
2002-2003 Optimization
2003-2004 Probability and Statistics in Complex Systems: Genomics,
Networks, and Financial Engineering
2004-2005 Mathematics of Materials and Macromolecules: Multiple Scales,
Disorder, and Singularities
2005-2006 Imaging
2006-2007 Applications of Algebraic Geometry
2007-2008 Mathematics of Molecular and Cellular Biology
2008-2009 Mathematics and Chemistry
2009-2010 Complex Fluids and Complex Flows

IMA SUMMER PROGRAMS

1987 Robotics
1988 Signal Processing
1989 Robust Statistics and Diagnostics
1990 Radar and Sonar (June 18-29)
New Directions in Time Series Analysis (July 2-27)
1991 Semiconductors
1992 Enviroimaental Studies: Mathematical, Computational, and
Statistical Analysis
1993 Modehng, Mesh Generation, and Adaptive Numerical Methods
for Partial Differential Equations
1994 Molecular Biology
1995 Large Scale Optimizations with Applications to Inverse Problems,
Optimal Control and Design, and Molecular and Structural
Optimization
1996 Emerging Applications of Number Theory (July 15-26)
Theory of Random Sets (August 22-24)
1997 Statistics in the Health Sciences
1998 Coding and Cryptography (July 6-18)
Mathematical Modeling in Industry (July 22-31)
1999 Codes, Systems, and Graphical Models (August 2-13, 1999)
2000 Mathematical Modeling in Industry: A Workshop for Graduate
Students (July 19-28)
2001 Geometric Methods in Inverse Problems and PDE Control
(July 16-27)
2002 Special Functions in the Digital Age (July 22-August 2)
2003 Probability and Partial Differential Equations in Modern
Applied Mathematics (July 21-August 1)
2004 n-Categories: Foundations and Applications (June 7-18)
2005 Wireless Communications (June 22-July 1)
2006 Sjonmetries and Overdetermined Systems of Partial Differential
Equations (July 17-August 4)
2007 Classical and Quantum Approaches in Molecular Modeling
(July 23-August 3)
2008 Geometrical Singularities and Singular Geometries
(July 14-25)

I M A "HOT TOPICS" W O R K S H O P S
• Challenges and Opportunities in Genomics: Production, Storage,
Mining and Use, April 24-27, 1999
• Decision Making Under Uncertainty: Energy and Environmental
Models, July 20-24, 1999
• Analysis and Modeling of Optical Devices, September 9-10, 1999
• Decision Making under Uncertainty: Assessment of the Reliability
of Mathematical Models, September 16-17, 1999
• Scaling Phenomena in Communication Networks, October 22-24,
1999
• Text Mining, April 17-18, 2000
• Mathematical Challenges in Global Positioning Systems (GPS),
August 16-18, 2000
• Modeling and Analysis of Noise in Integrated Circuits and Systems,
August 29-30, 2000
• Mathematics of the Internet: E-Auction and Markets, December
3-5, 2000
• Analysis and Modeling of Industrial Jetting Processes, January
10-13, 2001
• Special Workshop: Mathematical Opportunities in Large-Scale Net-
work Dynamics, August 6-7, 2001
• Wireless Networks, August 8-10 2001
• Numerical Relativity, June 24-29, 2002
• Operational Modeling and Biodefense: Problems, Techniques, and
Opportunities, September 28, 2002
• Data-driven Control and Optimization, December 4-6, 2002
• Agent Based Modehng and Simulation, November 3-6, 2003
• Enhancing the Search of Mathematics, April 26-27, 2004
• Compatible Spatial Discretizations for Partial Differential Equa-
tions, May 11-15, 2004
• Adaptive Sensing and Multimode Data Inversion, June 27-30, 2004
• Mixed Integer Programming, July 25-29, 2005
• New Directions in Probability Theory, August 5-6, 2005
• Negative Index Materials, October 2-4, 2006
 • The Evolution of Mathematical Communication in the Age of Dig-
ital Libraries, December 8-9, 2006
• Math is Cool! and Who Wants to Be a Mathematician?, November
3,2006
• Special Workshop: Blackwell-Tapia Conference, November 3-4,
2006
• Stochastic Models for Intracellular Reaction Networks, May 11-13,
2008

S P R I N G E R L E C T U R E N O T E S FROM T H E IMA:

The Mathematics and Physics of Disordered Media

Editors: Barry Hughes and Barry Ninham
(Lecture Notes in Math., Volume 1035, 1983)
Orienting Polymers
Editor: J.L. Ericksen
(Lecture Notes in Math., Volume 1063, 1984)
New Perspectives in Thermodynamics
Editor: James Serrin
(Springer-Verlag, 1986)

Models of Economic Dynamics

Editor: Hugo Sonnenschein
(Lecture Notes in Econ., Volume 264, 1986)
 The IMA Volumes in Mathematics and its Applications

1 Volume 142: Compatible Spatial Discretizations

Editors: Douglas N. Arnold, Pavel B. Bochev, Richard B. Lehoucq,
Roy A. Nicolaides, and Mikhail Shashkov
2 Volume 143: Wireless Communications
Editors: Prathima Agrawal, Daniel Matthew Andrews, Philip J. Fleming,
George Yin, and Lisa Zhang
3 Volume 144: Symmetries and Overdetermined Systems of Partial
Differential Equations
Editors: Michael Eastwood and Willard Miller, Jr.
4 Volume 145: Topics in Stochastic Analysis and Nonparametric
Estimation
Editors: Pao-Liu Chow, Boris Mordukhovich, and George Yin
5 Volume 146: Algorithms in Algebraic Geometry
Editors: Alicia Dickenstein, Frank-Olaf Schreyer,
and Andrew Sommese
6 Volume 147: Symmetric Functionals on Random Matrices and
Random Matchings Problems
by Grzegorz A. Rempała and Jacek Wesołowski
7 Volume 148: Software for Algebraic Geometry
Editors: Michael E. Stillman, Nobuki Takayama, and Jan Verschelde

The full list of IMA books can be found at the Web site of Institute for
Mathematics and its Applications:
http://www.ima.umn.edu/springer/volumes.html