Lanczos Algorithms for Large Symmetric Eigenvalue Computations. by Jane K.

Cullum; Ralph
A. Willoughby
Review by: Joe H. Chow
SIAM Review, Vol. 29, No. 2 (Jun., 1987), pp. 307-309
Published by: Society for Industrial and Applied Mathematics
Stable URL: http://www.jstor.org/stable/2031672 .
Accessed: 14/06/2014 05:35

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact [email protected].

Society for Industrial and Applied Mathematics is collaborating with JSTOR to digitize, preserve and extend
access to SIAM Review.

http://www.jstor.org

This content downloaded from 194.29.185.109 on Sat, 14 Jun 2014 05:35:38 AM

All use subject to JSTOR Terms and Conditions
SIAM REVIEW ? 1987 SocietyforIndustrialand Applied Mathematics
Vol. 28, No. 2, June 1987 007

BOOK REVIEWS
EDITED BY GEORGE HANDELMAN

are invitedto send booksfor reviewto Professor

Publishers of
GeorgeHandelman,Department
Sciences,Rensselaer
Mathematical Troy,NewYork12180-3590.
Institute,
Polytechnic

for Large Symmetric

Lanezos Algorithms By JANEK.
EigenvalueComputations.
Birkhauser,Boston, 1985. Vol. I: Theory.
CULLUMANDRALPHA. WILLOUGHBY.
xiv + 273 pp. $24.95. ISBN 0-8176-3058-9. Progress in Scientific Computing,
Vol. 3. Vol. II: Programs. viii+ 500 pp. $49.95. ISBN 0-8176-3294-8. Progress
in ScientificComputing, Vol. 4. (Completeset:ISBN 0-8176-3292-6.)
This is a researchmonographon a specificLanczos techniqueforcomputing
eigenvalues oflarge,sparsesymmetric matrices, whichcannotbe practically analyzed
byEISPACK-typeprocedures. It documents theresultsofover15 yearsofresearch
and computation experience bytheauthorsand is presented in 2 volumes.VolumeI
is on thetheoryand computation procedures oftheCullum-Willoughby's Lanczos
algorithms. VolumeII liststheFORTRAN code oftherelatedcomputerprograms.
Thisreviewis mostlyon VolumeI.
This advancedmonograph is intendedforengineers, scientists and mathemati-
cianswhoare interested in eigenelement computational procedures forlargematrix
problems. Largeeigenelement computation is nota routinetasklikeusingEISPACK.
Understanding thetechnique,in additionto theproblemit is supposedto solve,is
crucialto itssuccessful use.Fromthispointofview,thismonograph is recommended
readingfortheCullum-Willoughby's Lanczosalgorithm. Although thismonograph
is largelyself-contained, familiarity withmatrixcomputation techniques,such as
thosein Stewart'sbook Introduction to MatrixComputations, is essentialin fully
appreciating thematerial.
The purposeofthemonograph can perhapsbe bestexplainedwithsomeback-
groundon thedevelopment of theLanczos procedure.Lanczosintroduced (in the
early1950s)his techniqueof computing eigenvalues and eigenvectors of real sym-
metricmatrices byfirstperforming a Lanczosrecursion to transform thesystem to a
tridiagonal matrix.However,implementation ofthebasicLanczosrecursion yielded
erraticresults,a factattributed to theloss of orthogonality of the Lanczos vectors
whichperform the transformation. The loss of orthogonality was assumedto be
causedbyroundoff and cancellation errorsand wasdeemedunavoidable.As a result
interest in thetechniquefaded.However,Paigeshowedin his 1971 thesisthatthe
loss of orthogonality was actuallydue to a combination of the effects of roundoff
errorstogetherwiththe convergence of one or more of the eigenvaluesof the
tridiagonal Lanczosmatrixas itssizeis increased. Thisresultpointedoutpossibilities
ofhowthebasic Lanczosrecursion can be modified suchthataccurateeigenvalues
and eigenvectors can be computed.Sincethen,manyLanczos methodshavebeen
proposedforthecomputation ofeigenfunctions oflarge,sparsematrices, becausethe
methoddoes notproducefill-ins, and thegeneration ofLanczosmatrixis economic.
Two basicapproachesare used.One approachmaintains that"globalorthogonality
is crucial"whiletheotherapproachmaintains that"localorthogonality is sufficient."
The approachin thismonographbelongsto the latterapproachwhichdoes not
perform anyreorthogonalization.
Volume I discussesthe Cullum-Willoughby algorithmsof computingthe
eigenelements of real symmetric, Hermitianand nondefective complexsymmetric
307

This content downloaded from 194.29.185.109 on Sat, 14 Jun 2014 05:35:38 AM

All use subject to JSTOR Terms and Conditions
308 BOOK REVIEWS

matrices, and singularvaluesofrealrectangular matrices. The materialis organized

intoan Introduction and 8 chapters (Chapters0 to 7).
The Introduction is a summary ofthemonograph, including whereto findthe
relatedcode in VolumeII. Chapter0 introduces thenotationand definitions usedin
themonograph. Chapter1 is a reviewof realsymmetric matrices-their properties
and theiroccurrence in physicalsystemmodels.
Chapter2 reviews thebasicLanczosrecursion procedure andtheproperties when
performed withexactmathematics. Then it discussesthepracticalsituationof how
themethodbecomesunreliablewhenperformed withfiniteprecisionarithmetic. It
drawson Paige'sresults on attributing thelossoforthogonality oftheLanczosvectors
to roundoff errorsand convergences of certaineigenvalues.The emphasisis on
understanding and interpreting Paige'sresults;proofsof theorems are skipped.To
helpthereadergeta clearerunderstanding of thetradeoffs involved,it presentsa
dissectionoftheessentialchoices,suchas thetypeofrecursions and theamountof
reorthogonalization, inbuilding a practicalLanczosprocedure. Thechapter concludes
witha survey ofrecentmethods, listedaccording totheamountofreorthogonalization
required.A previewoftheCullum-Willoughby procedure is presented here.
Chapter3 preparesthereaderfortheCullum-Willoughby procedure presented
in Chapter4 witha reviewon the Sturmsequencingtechniqueand the inverse
iterationmethodforcomputing eigenvaluesand eigenvectors oftridiagonal matrices.
As mentioned earlier,theCullum-Willoughby's Lanczosalgorithm requiresno re-
orthogonalization. This is based on the"LanczosPhenomenon"thateverydistinct
eigenvalue oftherealsymmetric matrix A willappearas an eigenvalue ofa sufficiently
largeLanczosmatrix.Chapter4 providesthejustification oftheprocedure. It proves
the equivalence,underexact arithmetic, betweenthe Lanczos iterations and the
conjugate gradient approachforsolvingsystems oflinearequations.In finite precision
arithmetic the authorsshow that undercertainmild assumptions, the Lanczos
algorithm yieldsconjugategradients approximately. Thisobservation is usedto infer
thatthe eigenvaluesof A will be near-zeros of the characteristic polynomialof a
sufficiently largeLanczos matrix.Since largeLanczos matricescontaineigenvalues
thatdo notbelongto A, theauthorsdevisean identification testforseparating the
goodeigenvalues from the The is
spuriouseigenvalues. procedure illustrated detail in
withseverallargematrixexamples.
Chapters5 and 6 aretheextension oftheCullum-Willoughby methodto other
matrices. Chapter5 discussesthecomputation ofsingular valuesofa realrectangular
matrixA. It includesa helpfulreviewofsingularvaluesand theirapplications. The
techniquehereis to createa symmetric matrixwithzerodiagonalblocks,andA and
AT as its upperand loweroff-diagonal blocks,respectively. In this setup,even
dimensionLanczos matricesare required.The procedureis illustrated withseveral
examples.Chapter6 discussestheeigenelement computation fornondefective com-
plexsymmetric matrices. A two-sided Lanczosrecursion is used,resulting in complex
symmetric tridiagonal matrices. A QL algorithm forcomputing eigenvalues of such
matrices is includedforcompleteness.
Chapter7 presents an iterative blockLanczosprocedure witha conjugate gradient
procedureforselecting the updatedapproximation of the eigenvector matrix.The
approachis different fromthatpresented in Chapter4. The procedureis included
becauseitcan computemultiple eigenvaluesofa realsymmetric matrix, whereasthe
singlevectorLanczosprocedurecannot.Practicalimplementation oftheprocedure
is discussed.A faster hybrid blockschemeis also presented.
VolumeII is a collection ofFORTRAN codeoftheprocedures foundin Volume
I. Each programis precededby a briefsummaryof the algorithm. A glimpseof

This content downloaded from 194.29.185.109 on Sat, 14 Jun 2014 05:35:38 AM

All use subject to JSTOR Terms and Conditions
 BOOK REVIEWS 309

FORTRAN code indicatesthatthe programs are welldocumented. The programs

havebeen checkedby thePFORT compiler.Nonportable instructionsare listedat
theheaderofeach program. As forreliability,thisreviewer can onlyrelatehis own
experience.Severalyearsago,thisreviewer obtainedan earlierversionofthecodefor
realsymmetric matrices fromtheauthorsand usedit on a computer differentfrom
thatusedbytheauthors.The program hasruneversincewithout a change.
The monograph canbe readat severallevelsofinterest. Forthoseinterested only
in background oftheLanczosalgorithm, a readingofChapter2 is adequate.Those
interestedin usingtheCullum-Willoughby algorithm can go on to Chapter4, and if
necessary,toChapters 5, 6 and 7. Mostoftheproofscanbe skippedwithout impeding
on theunderstanding oftheprocedures. Theyareintendedforseriousresearchers.
The materialpresented in thismonograph is quitecomplex.However,itscom-
prehensionis made much easier by the authors'carefulorganization and clear
presentation. Explanations and interpretations are givenafterall lemmasand theo-
rems.Computation algorithms arealwayslistedoutseparately. Examplesto illustrate
the proceduresare explainedin detail.The largematrixproblems,besidesbeing
are quite fascinating
illustrative, to read, as each of thempresenteda different
challenge.
In summary, thismonograph is wellwritten and worthcarefulstudying bythose
interestedin largeeigenelement computation problems.
JOE H. CHOW
Institute
Polytechnic
Rensselaer

NumericalOptimization Techniques.By YURIJG. EVTUSHENKO. Optimization Soft-

ware,Inc., New York, 1985. xiv + 561 pp. $68.00. ISBN 0-911575-07-3. A
volumein theTranslations Seriesin Mathematics and Engineering.
Ifyouhaveonlyone nuitericaloptimization bookon yourshelf, thisshouldnot
be it.Too manymodemand important topicsareeithertreatedmuchtoo briefly or
ignoredentirely. For example,thereis virtually nothing on themostelegantoptimi-
zationtopicofourtime:positivedefinite secantupdatesforunconstrained minimi-
zation.On theotherhand,no optimization bookshelf is complete withoutthisvolume.
Severalstandard topicsarepresented froma verynonstandard (andinteresting)point
ofview.Furthermore, a numberofnonstandard (butimportant) topicsarediscussed
at length.
The firstchapterprovidesa comprehensive discussionofoptimality conditions.
Besidestheusuallocalcharacterizations forsmoothfunctions, conditions fornondif-
ferentiableproblems, minimax, and optimalcontrolproblemsarealso considered in
somedetail.
Western readerswillprobably be enlightened byChapter2. Stability theoryand
themethodof Lyapunovfunctions are subjectsnot usuallyfoundin optimization
texts.However,suchtoolslead to interesting convergence analysis.Chapter2 is rich
withsuchresults.Also includedin thischapterare convergence resultsforiterative
processes(for singleand multi-valued mappings),local and global Netwon-like
methodsand analysisfornonlinearequations,and methodsforfinding a minimax.
The latteris anothertopicfrequently overlooked in optimization books.
The discussionon globalNewton-like methodsfornonlinear equationsis some-
whatweak.Thereis no mentionofthepopularLevenberg-Marquardt method(orits
derivatives)and the mostimportant optimization problemof all-nonlinearleast
squares-goes unnoticed.Moreover,thisis the onlychapterwhereunconstrained
minimization is considered and thereis onlya briefmentionhere.

This content downloaded from 194.29.185.109 on Sat, 14 Jun 2014 05:35:38 AM

All use subject to JSTOR Terms and Conditions