A c t a Numerica O Q ) Volume 7 0 o D 01 • ooo o ool • oio O oi I O oooo • oool • Acta Numeric a 199 8 Managing editor A. Iserles DAMPT, University of Cambridge, Silver Street Cambridge CBS 9EW, England Editorial Board C. de Boor, University of Wisconsin, Madison, USA F. Brezzi, Instituto di Analisi Numerica del CNR, Italy J. C. Butcher, University of Auckland, New Zealand P.G . Ciarlet, Universite Paris VI, France G.H. Golub, Stanford University, USA H.B . Keller, 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 umerica 1998 Volume 7 CAMBRIDGE UNIVERSITY PRES S Published by the Press Syndicate of the University of Cambridge The Pitt Building, Trumpington Street, Cambridge CB2 1RP 40 West 20th Street, New York, NY 100114 211 ,U SA 10 Stamford Road, Oakleigh, Melbourne 3166, Australia © Cambridge University Press 1998 First published 1998 Printed in Great Britain at the University Press, Cambridge Library of Congress cataloguing in publication data available A catalogue record for this book is available from the British Library ISBN 05 216 4316 3 hardback ISSN 09624 92 9 Contents Monte Carlo an d quasiMont e Carl o method s 1 Russel E. Caflisch Nonlinear approximatio n 51 Ronald A. DeVore Relative perturbatio n result s for matri x eigenvalue s and singula r value s 151 Use C. F. Ipsen Stability for timedependen t differentia l equation s 203 Heinz-Otto Kreiss and Jens Lorenz Direct searc h algorithm s for optimizatio n calculation s 287 M. J. D. Powell Choice of norm s for dat a fitting an d function approximatio n 337 G. A. Watson Ada Numerica (1998), pp .1 4 9 © Cambridge University Press, 1998 Monte Carl o an d quasiM ont e Carl o methods Russel E . Caflisch* Mathematics Department, UCLA, Los Angeles, CA 90095-1555, USA E-mail: [email protected] Monte Carlo iso neo ft he most versatile and widely used numerical methods. Its convergence rate, O(N~1^2), isi ndependent ofd imension, which shows Monte Carlo t ob ev ery robust bu ta lso slow. This article presents a ni ntro duction t o Monte Carlo methods fori ntegration problems, including conver gence theory, sampling methods and variance reduction techniques. Acceler ated convergence forM onte Carlo quadrature isa ttained using quasir ando m (also called lowd iscrepancy ) sequences, which area d eterministic alternative to random orp seudor ando m sequences. Th ep oints i na quasir ando ms e quence are correlated to provide greater uniformity. The resulting quadrature method, called quasiM ont e Carlo, hasa convergence rate ofa pproximately O((log N^N'1). For quasiM ont e Carlo, both theoretical error estimates and practical limitations arep resented. Although th ee mphasis i nt his articlei s on integration, Monte Carlo simulation ofr arefied gasd ynamics isa lso dis cussed. In the limit of small mean free path (that is, the fluid dynamic limit), Monte Carlol osesi ts effectiveness because the collisional distance ism uch less than the fluid dynamic length scale. Computational examples ar ep resented throughout the text t o illustrate the theory. A number ofo pen problemsa re described. Research supported in part by the Army Research Office under grant number DAAH04 951 0 155 . R. E. CAFLISCH CONTENTS 1 Introductio n 2 2 Monte Carlo integratio n 4 3 Generation and sampling method s 8 4 Variance reduction 13 5 Quasirando m numbers 23 6 QuasiM ont e Carlo techniques 33 7 Monte Carlo method s forr arefied gas dynamics 42 References 46 1. Introductio n Monte Carlo provides a d irect method forp erforming simulation and integ ration. Because i t is simple an dd irect , Monte Carlo is easy t o use. I ti s also robust , since it sa ccuracy depends o no nly th ec rudest measure oft h e complexity of th e problem. For example, Monte Carlo integration converges at a rat e O{N~1/2) tha t is independent oft h e dimension oft h e integral. For thi s reason, Monte Carlo ist h e only viable method fora wide range of highd imensiona l problems, ranging from atomic physics t o finance. The price fori t sr obustness ist ha t Monte Carlo ca nb ee xtremely slow. The order O(N~1^2) convergence rat e is decelerating, since a n additional factor of4 i ncrease in computationa l effort only provides an additional factor of 2i mprovement i na ccuracy. Th e result oft hi s combination ofe ase of use, wider ange ofa pplicability and slowc onvergence ist ha t ane normous amount of computer time iss pent o nM onte Carlo computations . This represents a great opportunit y forr esearchers i nc omputational sci ence. Even modest improvements i n th e Monte Carlo method ca nh ave substantial impact o nt h e efficiency an d range of applicability for Monte Carlo methods . Indeed, much oft h e effort i n th e development of Monte Carlo ha s been i n construction ofv ariance reduction methods which speed up th ec omputation . A description ofs ome oft h e most common variance reduction method s isg iven i nS ection4 . Variance reduction method s accelerate th ec onvergence rat e byr educing the constan t i n front oft h e O(N~1/2) forM onte Carlo methods using ran dom o rp seudorando m sequences. A na lternativ e approach t o acceleration ist oc hange th e choice ofs equence. QuasiM ont e Carlo methods use quasi random (also known a s lowd iscrepancy ) sequences instead of random o r pseudorandom . Unlike pseudorando m sequences, quasir ando m sequences do no ta ttemp t t oi mitat e th e behaviour ofr andom sequences. Instead ,t h e elements ofa q uasirando m sequence are correlated t om ake them more uni form tha n rando m sequences. For thi s reason, Monte Carloi ntegration using MONTE CARLO AND QUASIMONT E CARLO 3 quasir ando m points converges more rapidly, at a rat e O(N~1 (log N)k), for somec onstant k. Quasir ando m sequences are described in Sections 5a nd 6. In spite of their importance in applications, Monte Carlo methods re ceive relatively littl e attentio n from numerical analysts and applied math ematicians. Instead, it is number theorist s and statistician s who design th e pseudor andom , quasir ando m and other types of sequence tha t are used in Monte Carlo, while th e innovations in Monte Carlo techniques are de veloped mainly by practitioners , including physicists, systems engineers and statisticians. The reasons for th e near neglect of Monte Carlo in numerical analysis and applied mathematics are related t o it s robustness. First , Monte Carlo meth ods require less sophisticated mathematics tha n other numerical methods . Finite difference and finite element methods, for example, require careful mathematical analysis because of possible stability problems, bu t stability is not an issue for Monte Carlo. Instead , Monte Carlo nearly always gives an answer tha t is qualitatively correct, bu t acceleration (error reduction) is always needed. Second, Monte Carlo methods are often phrased in non mathematical terms . In rarefied gas dynamics, for example, Monte Carlo allows for direct simulation of th e dynamics of th e gas of particles, as de scribed in Section 7. Finally, it iso ften difficult t o obtain definitive resultso n Monte Carlo, because of th e random noise. Thu s computationa l improve ments often come more from experience tha n from a particula r insightful calculation. Thisa rticle isi ntended t op rovide ani ntroduction t o MonteC arlo methods for numerical analysts and applied mathematicians . In spite of th e reasons cited above, there are ample opportunitie s for thi s community t o make sig nificant contributions t o Monte Carlo. Firs t of all, any improvements can have a big impact, because of th e prevalence of Monte Carlo computations . Second, the methodology of numerical analysis and applied mathematics , including well controlled computational experiments on canonical problems, is needed for Monte Carlo. Finally, ther e are some outstandin g problems on which a numerical analysis or applied mathematic s viewpoint is clearly needed; for example: design ofM onteC arlos imulation for transpor t problems int h e diffusion limit (Section 7) formulation of a quasiM ont e Carlo method for th e Metropolis al gorithm (Section 6) explanation of why quasiM ont e Carlo behaves like standar d Monte Carlo when th e dimension is large and th e number of simulation is of moderate size (Section 6). Someo lder, bu t still very good, general references on Monte Carlo are Kalos and Whitlock (1986) and Hammersley and Handscomb (1965). 4 R. E. CAFLISCH The focus oft hi s article iso n Monte Carlof or integration problems. Integ ration problems ar es imply stated , bu t they ca nb ee xtremely challenging. In addition, integration problems contain most of th e difficulties tha t ar e found i n more general Monte Carlo computations, such a s simulationa n d optimization. The next section formulates th eM onte Carlo method fori ntegration and describes it s convergence. Section 3 describes random number generators and sampling methods . Variance reduction methods ar ed iscussed i nS ec tion 4a n dq uasiM ont e Carlo methods i n Section 5. Effective use ofq uasi Monte Carlor equiress omem odification ofs tandar d Monte Carlot echniques, as described i n Section 6. Monte Carlo methods for rarefied ga sd ynamics are described i n Section 7, with emphasis on th e loss of effectiveness for Monte Carlo i nt h e fluid dynamic limit. 2. Mont e Carl o integratio n The integral ofa L ebesgue integrable function f(x) can b ee xpressed as th e average or expectation of th e function / evaluated a t a random location. Consider a ni ntegral ont h eo ned imensiona l unit interval nn= f1 f(x)dx = f. (2.i ) Jo Let a; bea r andom variable tha t isu niformly distribute d ont h e unit interval. Then I[f] = E[f(x)]. (2.2 ) For a ni ntegral ont h eu nit cube Id = [0, l]di n dd imensions, )dx, (2.3) in which a;i sa uniformly distribute d vector i nt h eu nit cube. The Monte Carlo quadratur e formula isb ased ont h ep robabilistic inter pretation ofa n integral. Consider as equence {x } sampled from th e uniform n distribution. Then a ne mpirical approximation t ot h ee xpectationi s M/] = ^ £/(*n) . (2.4) n=l According t ot h eS trong Law ofL arge Numbers (Feller 1971), thi s approx imation i sc onvergent with probability one; tha ti s , (2.5) In addition, i ti su nbiased, which means tha t th e average of Iff[ /] ise xactly /[/] for any N; tha ti s , E[I [f]] = I[f], (2.6) N