Acta N umeric a ActaN umeric a 1993 Managing Editor A.I serles DAMTP, University of Cambridge, Silver Street Cambridge CB3 9EW, England Editorial Board C. de Booi; University of Wisconsin, Madison, USA F. Brezzi, lnstituto di Analisi Numerica del CNR, Italy J.C. Butchei; University of Auckland, New Zealand RG. Ciarlet, Universite'Paris VI, France G.H. Golub, Stanford University, USA H.B. Kellei; California Institute of Technology, USA H.O . Kreiss, University of California, Los Angeles, USA K.W. Morton, University of Oxford, England M.J.D. Powell, University of Cambridge, England R. Temam, Universite Paris Sud, France eta Nu meric a 1993 CAMBRIDGE UNIVERSITY PRES S Published by the Press Syndicate oft he University of Cambridge The Pitt Building, Trumpington Street, Cambridge CB2 1RP 40 West 20th Street, New York, NY 100114 211 , USA 10S tamford Road, Oakleigh, Melbourne 3166,A ustralia © Cambridge University Press 1993 First published 1993 Printed in Canada Library of Congress cataloging in publication data available A catalogue record for this book is available from the British Library ISBN 05 214 4356 3 hardback Contents Continuation and path following 1 Eugene L. Allgower and Kurt Georg Multivariate piecewise polynomials 65 C. de Boor Parallel numerical linear algebra Il l James W. Demmel, Michael T. Heath and Henk A. van der Vorst Linear stability analysis in the numerical solution of initial value problems 199 J.L.M. van Dorsselaer, J.F.B.M. Kraaijevanger and M.N. Spijker Finite element solution of the Navier-Stokes equations 239 Michel Fortin Old and new convergence proofs for multigrid methods 285 Harry Yserentant Ada Numerica (1993), pp.1 6 4 Continuation and path following Eugene L. Allgower* an d Kur t Georg * Department of Mathematics Colorado State University Ft. Collins, CO 80523, USA E-mail: [email protected] The main ideas of path following by predictorcorrecto r and piecewiselinea r methods, and their application int hed irection ofh omotopy methods and non linear eigenvalue problems are reviewed. Further new applications t o areas such as polynomial systems of equations, linear eigenvalue problems, interior methods forl inear programming, parametric programming and complex bi furcation ar es urveyed. Complexity issues an da vailable software ar ea lso discussed. CONTENTS 1 Introduction 1 2 Th e basics of predictorcorrecto r pat h following 3 3 Aspects of implementations 7 4 Applications 15 5 Piecewiselinea r methods 3 4 6 Complexity 4 1 7 Available software 4 4 References 4 7 1. Introductio n Continuation, embedding or homotopy methods have long served as useful theoretical tools i nm odern mathematics . Their use can be traced backa t least t o suchv enerated worksa st hoseo fP oincare (18811 886) , Klein (1882 1883) and Bernstein (1910). Leray and Schauder (1934) refined th e tool and presented i t as a g lobal result i nt opology, viz. th e homotopy invariance of degree. The use ofd eformations t os olve nonlinear systems of equations may b et raced back a t least t o Lahaye (1934). Th e classical embedding * Partially supported by the National Science Foundation via grant no. DMS9 104058 . 2 E. ALLGOWER AND K. GEORG methods weret h e first deformation methods t o be numerically implemented and may be regarded as a forerunner of th e predictorcorrecto r methods for path following which wew ill discuss here. Becauseo ft heir versatility and robustness, numericalc ontinuation or pat h following methods haven owb eenf indinge ver wider usei n scientific applica tions. Our aim here ist o present someo ft h e recent advances in this subject regarding new adaptations , applications, and analysis ofe fficiency and com plexity. To make th e discussion relatively selfc ontained , we review some of th e background of numerical continuation methods. Introductions into aspects of th e subject may be found in th e books by Garcia and Zangwill (1981), Gould and Tolle (1983), Keller (1987), Rheinboldt (1986), Seydel (1988) and Todd (1976a). Th e philosophy and notatio n of th e present arti cle will be tha t of our book Allgower and Georg (1990), which also contains an extensive bibliography u p t o 1990. The viewpoint which will be adopted here is tha t numerical continuation methods ar e techniques for numerically approximating a solution curve c which is implicitly defined by an underdetermined system of equations. In the literatur e of numerical analysis, th e term s numerical continuation and path following are used interchangeably. There ar e various objectives for which th e numerical approximation of c canb eu seda nd , depending upon th eo bjective, th ea pproximating technique is adapte d accordingly. In fact, continuation is a unifying concept, under which various numerical methods may be subsumed which may otherwise have very littl e in common. For example, simplicial fixed point methods for solving problems in mathematica l economics, th e generation of bifurca tion diagrams of nonlinear eigenvalue problems involving partia l differential equations, and th e recently developed interior point methods for solvingl in ear programming problems seem t o be quite unrelated. Nevertheless, there is some benefit in considering the m as special cases of pat h following. We personally are struck by th e remarkable fact tha t a technique which wasi ni tially developed for solving difficult nonlinear problems nowt urn s out t o be extremely useful for treatin g various problems which are essentially linear: e.g. linear eigenvalue problems, and linear programming and complementar ity problems. The remainder of th e article is organized as follows. Section 2 contains the basic ideas of predictorcorrecto r pat h following methods. In Section 3 somet echnical aspects of implementing predictorcorrecto r methods are ad dressed, e.g. th e numerical linear algebra involved and steplength strategies. Section 4 deals with various applications of pat h following methods. We begin with a brief discussion of homotopy methods forf ixedp oint problems and global Newton methods . Then wea ddress th e problem off indingm ulti ple solutions. In particular , wed iscuss recent homotopy methods for finding all solutions of polynomial systems of equations. Next wes urvey some pat h CONTINUATION AND PATH FOLLOWING 3 following aspectso fn onlinear eigenvaluep roblems, an d address th e question of handling bifurcations. Finally, thre e new developments i np at h following are discussed: (1) Th e solution ofl inear eigenvalue problems via special ho motopy approaches; (2)t h e handling ofp arametri c programming problems by following certain branches ofc ritical points via active se ts trategies; an d (3) th e pat h following aspects involved i n th e interior point methods for solving linear and quadrati c programming problems. Section 5 presents a n introduction t o th e principles of piecewise linear methods. These methods view pat h following i na d ifferent light: instead of approximately following a s mooth solution curve, the y exactly follow a n ap proximate curve (i.e. a p olygonal path) . Some instances where these meth ods are useful are discussed, e.g. linear complementarity problems o rh omo topy methods where predictorcorrecto r methods ar en o t implementable, because ofl ack ofs moothness. We also briefly address th er elated topico f approximating implicitly defined surfaces. The issueo f th e computational complexity of pat h following is considered in Section 6. This issue isr elated t ot h e NewtonK antorovic h theory andi s currently ofc onsiderable interest i nt h e context ofi nterior point methods . We conclude b yl isting some available software related t o pat h following and indicate how th e reader might access these codes. No attemp t t oc om pare or evaluate th e various codes iso ffered. I na ny case, our opinion is tha t path following codes always need t ob ec onsiderably adapte d t ot h e special purposes forw hich they ar ed esigned. Th e pat h following literatur e offers various toolsf ora ccomplishing such tasks . Although ther e are some general purpose codes, probably none will slay every dragon. The extensive bibliography contains only cited items. Space considera tions prohibited th e addressing ofs ome importan t topics, and consequently some significant recent contributions t o th ef ielda r en o t contained i n th e bibliography. 2. The basics of predictor-corrector path following The simplest (an d most frequently occurring) case ofa n underdetermined system ofn onlinear equations contains jus t one degree of freedom: H(u) = 0 where H : RN+1 » RN isa smooth map . (2.1 ) When we sayt ha t a map iss mooth we shall mean tha t i t ha sa s many continuous derivatives a st h e context oft h e discussion requires. For conve nience,t h e reader may assume C°°. I no rder t oa pply th e Implicit Function Theorem, we need th e following standar d Definition 2.1 We call u a regular point of H ift h e Jacobian H'(u)h a s maximal rank. We call y a regular value of H if u i s a regular point of 4 E. ALLGOWER AND K. GEORG H whenever H(u) = y. If a point or value is not regular, then it is called singular. Let UQ e RN+1 be a regular point of H such tha t H(UQ) = 0. I t follows from th e Implicit Function Theorem tha t th e solution set H~ 1(0) can be locally parametrized abou t uo with respect t o some coordinate. By a re parametrization (according t o arclength) , weo btain a smooth curve c : J — J^AT+I £ open interval J containing zero such tha t for all s € J: or som e c(0) = u (2.2) 0 H'(c(s))c(s) = 0, (2.3) ||c(a)|| = 1, (2.4) det Thesec onditions uniquely determine th et angent c(s). Herea nd int h e fol lowing, ( . )* denotes th e Hermitian transpos e and || . || th e Euclidean norm. Condition (2.4) normalizes th e parametrizatio n t o arclength. This is only for theoretical convenience, and it is not an intrinsic restriction. Condition (2.5) chooses one of th e two possible orientations. The preceding discussion motivates th e following Definition 2.2 Let A be an (JV,iV|l)m atri x with maximal rank. For the purpose of our exposition, th e unique vector t(A) 6 R^ +1 satisfying th e conditions At = 0, (2.6) ll'll = 1, (27 ) det (£ } >0 , (2.8) will be called th e tangent vector induced by A. Making use of thi s definition, solution curve c(s) is characterized as th e solution of th e initial value problem u = t(H'(u)), ti(0) = u (2.9) o which in thi s context is occasionally attribute d t o Davidenko (1953), see also Branin (1972). Note tha t th e domain {u € R^ +1 : u is a regular point } is open. This differential equation is not used in efficient pat h following algorithms, bu t it serves as a useful device in analysing th e path . Two examples are: Lemma 2. 3 Let (a, b)b e th e maximal interval of existence for (2.9). If a