HANDBOOK NUMERICAL ANALYSIS P. G. CIARLET and J. L. LIONS « Editors atti Finite Difference Methods (Part 1) Solution of Equations in R” (va a8) NORTH-HOLLANDPLAIVINE SmNCE B ‘sare Harpers 25 ed Mow 200, 1000 Ak Aureserdam. The Netherbnek (0 Nb eevee Sademce FL, All ies eserves ‘This work f proucted unter coppight hy Meir Scie, ana the alicswingy ienen amd cumiaioes apply toa we ‘pa plete single haps oy he ke perso wend Wy matin copyright Peter pujrwat 3 Ye saga! oe a nr pctocopying cag tug Us spec oops ins Og bt serio ot parodvnl pepe fesse al JU fete of deewrase delves, Spor ans ae arabe fo edacexal Stotakon tt nach soaks pooenagees ov nw pf elect castinn wie FRoertsicen may he seep cee fet Ehevier's Scene & Tic Right gore i shard, UK: poe Hi 10s ARR nH) 195 NTN mnt permnomtechceronmy Yowrny ab cone: yuu rest ie athe lieve Seine Mowefsier OMpsAn levi uve, Py suk CHMORET OqYRY an fa Uae Pemntsten. (le LA nes my sear persian aad ri pyc hava ie Cope Chsanane Cn, tne 222 Rasemunad (Orie, Dens, 01903, USA. phone ¢ 11 (27H) 7ROSMA Fas+1 (FT) SMITH, atin the Uk terug he Copa [cerning Aspncy Rape barance Scrvic: (CLAUS). 0 Toahoahan Ciaat Real Leaks WIPOLP. UK pce: tl) 207 (601 S55, fan feHy N07 61 S90 Cer cownaris ny faa reprngrapi rie agency fr pays. Deerative Wark “Tables of consents fe epeonaea ta tera cacubaon, Pet permaicion of Hacer Scere hoger scr resale cor uiresennt of suck nara ‘Vermnciner ff Padvicern expered bs all abe deriva wotks sachalies conmplations and tarmatnons ‘Weornke Siac Doms Fete ofthe Pub rgd nor wr eke any materi consid inthis ek, icing ay chaper ‘pata coger [except accatimed show. mo part of thts seorknus hereon sore serevad yskoa F Waasnrted ay aT “ep reute, chewoak, nechonical pheucepying. recanlaag or ofereine without iar rticn peice al the Babbler ‘abe perenstac reyuests a1 Elsev ies cenwe de Texts Rights Depart nothe pnee, x amd emul abo pote shoes Retice ls resposseity bs ame hy the Paitbalcr Koray amary andor damage be prcons cr prope a wir of ect lati, repigcie: orvthcreisc. or roetay wae we operaicn any methods, progr neater ca contain fa eseral Fetein Hecan of raped thane othe relia enor, ie prtouls ane cericnbotn cago a srg sesapes tend ean ins eaion 1080) Sexes inmpaeiane 12 are enpressiet 3103 Library of ConmgesscCandoging in Publication Tata asccck of menencal analysiipescra eters, PAG, Che, 1 Lon. roan Anke isbeprapalieferem.rs Cretan: +f Fate sles anthods tp 1): Suet of ‘options BBN. 166.7 WiMuenecal anaten. (Cikt, Mapper W Lins, pte Los, QASIMIST (060 Teka ea or Wns; tL ta 5 The pape sc spin ots ht mopeds of ANSLEY 20-48-1003 Ponce of Ppt. Vrnkalun The SetitandeIntroduction G.I. Marchuk The finite difference method is a universal and efficient numerical method for solving differential equations. Its intensive development, which began at the end of 1940s and the beginning of 1950s, was stimulated by the need to cope with a number of complex problems of science and technology. Powerful computers provided an impetus of paramount importance for the development and application of the finite difference method which in itself is sufficiently simple in utilization and can be conveniently realized using computers of different architecture. A large number of complicated multidimensional problems in electrodynamics, elasticity theory, fluid mechanics, gas dynamics, theory of particle and radiation transfer, atmosphere and ocean dynamics, and plasma physics were solved employing the finite difference techniques. Numerous spectacular results have been obtained in the theory of finite difference methods during the last four decades. Tn ordinary differential equations, the stability of the main classical finite difference methods was investigated and the relevant accuracy estimates were constructed, a large number of new versions of these methods were constructed, and efficient algorithms were suggested for their realization in a wide field of applications-oriented problems. The needs of electronics, kinetics, and catalysis stimulated the development of a broad class of methods for solving stiff systems of equations. Problems in control theory, biology, and medicine were important for the progress in finite difference methods of solving delay ordinary differential equations. In partial differential equations, the achievements of the finite diflerence method operators of mathematical physics were constructed, including those with conser- vation properties, that is, those obeying the discrete counterparts of the laws of conservation. An elegant theory of approximation, stability, and convergence of the finite difference method was constructed. HANDBOOK OF NUMERICAL ANALYSIS, VOL. I Finite Difference Methods (Part 1}—Solution of Equations in H" (Part 1) Edited by P.G. Ciarlet and J.L, Lions © 1990, Elsevier Science Publishers B.V. (North-Holland) 34 Introduction ‘The efforts of the specialists in differential equations and numerical mathematics yielded a convenient and efficient apparatus of the finite difference method, including spectral analysis, discrete maximum principle, and energy method. Considerable progress was achieved in the methods on a sequence of grids, including extrapolation methods. A flood of new results in the theory of the finite element method greatly stimulated new achievements in the theory of the finite difference method. An important stage in the progress of finite difference methods was the development of the alternating direction implicit method, the fractional steps cihud, and ite spiiiting mcihod. The reaiizaiion of these meiiods consisis in solving a large number of one-dimensional problems. A considerable number of versions of this class of methods have been suggested, having high approximation accuracy and absolute stability. Numerous multidimensional problems in physics, mechanics, and geophysical hydrodynamics have been solved using these methods. The theory of the finite difference method is far from having been completed, especially in the field of nonlinear partial differential equations. Life never ceases to offer new complex problems, and the method of finite differences remains a powerful approach to solving them. It would be impossible in the Handbook of Numerical Analysis to cover even all the basic achievements of the theory. Nevertheless, I do not doubt that its publication will help making the finite difference method interesting to new people and will attract new researchers to solving its problems.Finite Difference a P Vidar Thomée Department of Mathematics Chalmers University of Technology S-41296 Géteborg, Sweden HANDBOOK OF NUMERICAL ANALYSIS, VOL. I Finite Difference Methods (Part 1) Solution of Equations in R* (Part 1) Edited by P.G. Ciarlet and J.L. Lions © 1990, Elsevier Science Publishers B.V. (North-Holland)Contents PREFACE Crarrer L Introduction 1, The pure initial value problem 2. The mixed initial boundary value problem Cuaprer II. The Pure Initial Value Problem 3, Finite difference schemes Ly theory for finite difference schemes with constant voelficients Particular single-step schemes for constant coefficients problems Some multistep difference schemes with constant coefficients John’s approach to maximum-norm estimates Stability of difference schemes for general parabolic equations Convergence estimates (Cuavrex Il, The Mixed Initial Boundary Value Problem 10, The energy method 11. Monotonicity and maximum principle type arguments 12. Swabiiity anaiysis by special mctivuds 13. Various additional topics REFERENCES List oF SYMBOLS Suwecr INDEX i ul ai 3 2 a 32 76 90 7 109 10 134 167 183, 191 195Preface This article is devoted to the numerical solution of linear partial differential equations of parabolic type by means of finite difference methods. The emphasis in our presentation will be on concepts and basic principles of the analysis, and weshall therefore often restrict our considerations to model problems in as much as the choice of the parabolic equation, the regularity of its coefficients, the underlying geometry and the boundary conditions are concerned. In the beginning of the article proofs are provided for the principal results, but as the situation under investigation Tequires more technical machinery, the analysis will become more sketchy. The reader will then be referred to the literature quoted for fuller accounts of both results and theoretical foundations. The article is divided into three chapters. The first of these is of an introductory nature and presents some of the basic problems and concepts of the theory for the model heat equation in one space dimension, This chapter i is: subdivided i into two 2, devoted to th: value problem, respectively. This division is then the a is for the plan t the rest of the article, where Chapters II and III treat these two classes of problems in greater depth and generality. Whereas most of the theory in Chapter II depends on Fouricr analysis, that of Chapter III relies heavily on energy and monotonicity type arguments. The finite difference method for partial differential equations has a relatively short history. After the fundamental theoretical paper by COURANT, FRIEDRICHS and LEWY [1928] on the solution of the problems of mathematical physics by means of finite differences, the subject lay dormant till the period of, and immediately following, the Second World War, when considerable theoretical progress was made, and large scale practical applications became possible with the aid of computers. In this context a major role was piayed by the work of von Neumann, partly reported in O'Brian, HyMAN and Kaptan [1951]. For parabolic equations a highlight of the early theory was the celebrated paper by Jou [1952], which had a great influence on the subsequent research. The field then had its golden age during the 1950s and 1960s, and major contributions were given by Douglas, Kreiss, Lees, Samarskii, Widlund and others. At the end of this period the theory for the pure initial value problem had become asonably well developed and complete, and this was essentially also true for mixed initial boundary value problems in one space dimension. For multidimensional problems in general domains the situation was less satisfactory, partly because the finite difference method employs the values of the solution at the points of a uniform 910 V. Thomée mesh, which does not necessarily fit the domain. This led the development into a different direction based on variational formulations of the boundary value problems, and using piecewise polynomial approximating functions on more flexible partitions of the domain. This approach, the finite element method, was better suited for complex geometries and many numerical analysts (including the author of this article) abandoned the classical finite difference method to work with finitc elements. The papers on parabolic equations using finite differences after 1970 are few, particularly in the West. It should be said, however, that finite elements and finite differences have many poinis in common, aud it may be more appropriaic iv ihiuk of ihe new development as a continuation of the established theory rather than a break away from it. In the Russian literature, for instance, finite element methods are often referred to as variational difference schemes, and variational thinking was, in fact, used already in the paper by Courant, Friedrichs and Lewy quoted above. The finite element theory owes much of its present level of development and sophistication to. the foundation provided by the finite difference theory. However, in the present Handbook, the two subjects are separated, and we shall only very briefly touch upon their interrelation below. In our presentation here we shall not discuss techniques for solving the algebraic linear systems of equations that result from the discretization of the initial boundary value problems, but refer to other articles of this Handbook concerning such matters. Neither shall we treat the related area of alternating direction implicit methods, or fractional step methods, which are designed to reduce the amount of computation needed in multidimensional, particularly rectangular, domains, and to which a special article of this volume is devoted. Several textbooks exist which treat finite difference methods for parabolic problems, and we refer. in particular. to Ricutmyer and Morton [1967] and Samarsktt and Guin [1973] for thorough accounts of the field, but also (in chronological order of publication) to Co.tatz [1955], Forsyte and Wasow [1960], Ruanenx and Fiuiprow [1960], Fox [1962], Gopunov and RyABENKI [1964], Sauvev [1964], Smit [1965], Bapuska, Pracer and Virdsex [1966], Mircwewe [1969], and Samarsku [1971]. In addition we would like to mention the 2 1 £1960], We have includ. ist of references a large number ofo original papers, not alll of which are quoted in our text. For treatises on the theory of parabolic differential equations, covering existence, uniqueness and regularity results such as needed here, we refer to FRIEDMAN [1964] and LapyZENSKAsA, SOLONNIKOV and Uraticeva [1968]. I would like to take this opportunity to thank Chalmers University of Technology for granting me a reduction of my tcaching load while writing this article, to Ann-Britt Karlsson and Yumi Karlsson for typing the manuscript, and to Mohammad Asadzadeh for proofreading the entire work.Carter I Introduction In this first intraductory chapter our purpose is to use the simplest possible model problems to present some basic concepts which are important for the understanding of the formulation and analysis of finite difference methods for parabolic partial differential equations. The chapter is subdivided into two sections corresponding to the two basic problems discussed in the rest of this article, namely the pure initial value problem and the mixed initial boundary value problem. In the first section we thus consider the pure initial value problem for the heat equation in one space dimension. We begin with the simplest example of an explicit one-step, or two-level, finite difference approximation, discuss its stability with respect to the maximum norm and relate its formal accuracy to its rate of convergence to the exact solution. We also present an example of the construction of a more accurate explicit scheme. We then introduce the application of Fourier iccimigues in ite auaiysis uf bil siability, mow with respect io the ZL. accuracy, and convergence. We finally touch upon the possibility of using more than two time levels in our approximations. Section 2 is devoted to the mixed initial boundary value problem for the same basic parabolic equation, with Dirichlet type boundary conditions at the endpoints of a finite interval in the space variable. Here we discuss the possibility and advantage of using implicit methods, requiring the solution of a linear system of equations at each time level. Stability and error analysis is carried out for the simplest such methods, the backward Euler method and the more accurate Crank-Nicolson method. Again both maximum-norm estimates based on positivity properties of the difference scheme and [,-norm estimates derived by Fourier analysis are treated. A brief mention is made of the possibility of extending some initial boundary value problems to periodic pure initial value problems. The material in this chapter is standard and we refer to the basic textbooks quoted in our preface for further details and references. 1. The pure initial value problem In this first section of our introduction to the solution of parabolic problems by means of finite difference methods we shall discuss several such methods for the pure initial value problem for the simple homogeneous heat equation in one space dimension.2 V. Thomée CHAPTER 1 We thus wish to find the solution of the pure initial value problem Pu s=aa fk R, 120, a oe OF XE ty u(x,0)=0(x) for xe R, where R denotes the real axis and v is a given smooth bounded function. It is well known that this problem admits a unique solution, many properties of which may be deduced, for instance, from the representation 1 = | cM yy dys . u(x, t) Jami fe (9) dy =(E(v) (x). Here we think of the right-hand side as defining the solution operator E(¢) of (1.1).In particular, we may note that this solution operator is bounded, or, more precisely, sup| E(o(x)] =sup|ulx,t)| 0. (12) For the numerical solution of the problem (1.1) by finite differences we introduce a grid of mesh points (x, t)=(jh,mk) where h and k are mesh parameters which are small and thought of: ‘as tending to zero, and where j and nare integers, n>0. We then look for an appre te of h will be denoted by U3, by solving a problem in which the derivatives in (1.1) awe been replaced by finite difference quotients. Define thus for functions defined on the grid the forward and backward difference quotients 8,Uj=h (UF. — UD), 6, Uj=h (U5 —U5_.), and similarly, for instance, OU} =k" (UF*!— U9). The simplest finite difference equation corresponding to (1.1) is then for —w4 the method is unstable. To see this, we may choose v,=(— 1)/e where ¢ is a small positive number so that ||v|| =e. Then = [A= 19-1 41-24) (— 1) + A - 1) Je =(1—44)(— 1%, or, more generally, Uj=(L—4ay(— Ie, whence |U"]=(44-1'e00 as n00. We thus find that in this case, even though the initial data are very small, the discrete svluiivu icuds iv infinity » for insianes, ai i=nk= i, when k=1/n>0. This may be interpreted to mean that very small perturbations of the initial data (for instance by roundoff errors) may cause such big changes in the discrete solution at later times that it becomes useless. We now restrict the considerations to the stable case, 4<+4, and we shall show that, provided the 1 data and thus the exact solution of (1.1) are smooth enough, the discrete solution converges to the exact solution as the mesh parameters tend to zero. In order to demonstrate this, we need to notice that the exact solution satisfies the difference equation with a small error, which tends to zero with h and k. In fact, setting uj=u(jh, nk) we have by Taylor series expansion =0,u}—0,5;uj = Ok + h?) = O(N), if A<4, where the constants involved in the order relations will depend on upper bounds for in wom as n> vy, iia14 V. Thomée (CHAPTER 1 @7u/dt? and @*u/dx*. The expression t4 is referred to as the truncation or local discretization error. We have now the following result. Here and below we denote by C™" the set of functions of (x, t) with bounded derivatives of orders at most m and n with respect to x and t, respectively. Tueorem 1.1. Assume that we C*? and k/h? = 1<4. Then there is a C= Clu, T) such that |U"=u"| 1. Take ; UZAY, acl" Po BE oo, ? and hence, by repeated application, [|U"| =1EE oI" 00 as no, which proves the theorem. (J The condition (1.8) is a special case of what is referred to as von Neumann’s condition, and which was first proposed in O’BRiAN, Hyman and Kaptan [1951]. ‘We shall now see that in a slightly different setting this is also sutticient for stability. The symbol is particularly suited for investigating stability in the framework of Fourier analysis. It is then most convenient to use the /, -norm to measure the mesh functions. Let thus V={V,}*,,, be a mesh function in the space variable and set wo 12 Wb.=(t > Ww) : The set of mesh functions thus normed will be denoted /,,, below. Let us also define, for such a mesh function, its discrete Fourier transform HO=hy Ve, 7 where we asume that the sum is absolutely convergent. Recall the Parseval relation Wika x5 [moras « 1 = | [Pokey ag. (1) “ahSECTION 1 Introduction 7 We may now define stability with respect to the norm ||-{|,,,, oF in I, ,, to mean, analogously to above, WERV lan0, which is equivalent to (1.8) (and thus with C=1 in (1.10). ‘The convergence analysis may also be carried out in ||*||2,, We introduce as before the truncation error Mako yet lg yey Ty Ey and say that the difference scheme is accurate of order yp (assuming 4=k/h?= constant) if, formally, t*=O(h") as h-»0. Under the appropriate assumptions concerning to the smoothness and asymptotic behavior for large |x| of the exact solution this assumption may be translated into an estimate of the form Wella pS CU", (Lt) and we may show the following error estimate. Tueorem 1.4. Assume that E, is accurate of order p and stable in |, ,. Then under the appropriate regularity assumptions on the exact solution of (1.1) we have18 V. Thomée CHAPTER | Proor. As above we have, for 27=Ut—uf, ant aa=-k YEE ey iS so that by (1.10) and (1.11), W2"lla.nSk Y WER illaask YS Uhlan 0 So which proves the theorem. [ Consider again the fourth-order scheme (1.5). Recall that the stability with respect to the maximum norm was deduced above for $ —1 for A<%. It follows that \E(@|< {for all real if A<3. Since E(n)=1+84(4—3)> 1 if 4>} the von Neumann condition is not satisfied then. Thus, in particular, the present method is stable in /, , for 4<3 and unstable in both 1, , and maximum norm if 2>3. We note that if 1 <4, we have strict inequality for #2nn, or [E(Q|<1 for 0<|é|0. Such a two-step or three-level scheme would formally be accurate of second order in both x and t. Although, as we shall see in Section 3 below, the particular scheme (1.14) tums out to be unstable for any combination of h and k, other multistep schemes are useful in applications. 2. The mixed initial boundary value problem In physical situations our above pure initial value model problem (1.1) is generally not adequate, and instead it is required to solve the parabolic equation on a finite interval with boundary values given at the endpoints of the interval for positive time. We shall therefore have reason to consider the following model problem: ou Ou ou ou > aa for xe [0,1], 120, u(0, )=u(1,=0, for t>0, (2.1) u(x, 0)=0(x), for xe [0, 1). For the approximate solution we may again cover the domain with a grid of mesh points, this time by dividing [0, 1] into subintervals of equal length h= UM, where M isa positive integer, and consider the mesh points (jh, nk) with j= and n=O, i,.... ing as above Uj deuvic ihe apprvaimaic suiuiivn ai (ji, nh), the 4, i b22 V. Thomée CHAPTER 1 natural explicit finite difference scheme is now 0,U}=0,0,U}, for for n>0, (2.2) M, .5M-I, 20, for j=0,... or, for Uj, j=0,...,.M, given, Up AUG. + Ufys)A(L=2A)UF, j=, vo This scheme is referred to as the forward Euler scheme. This time we are looking for a sequence of vectors U"=(U3,...,U%y)" with U%=U%,=0 which we first norm by (U"ll...=max |U5]. j When 1=k/h? <4 we conclude, as before, that VOT SU | yas or, defining the local solution operator Ey, in the obvious way, NEtePll oS Ulloa 220, so that the scheme is maximum-norm stable for 1<3. Here, in order to see that this condilion is also necessary for stability, we modify our above example so as to incorporate the boundary conditions and set M. —1 sin nh, j= Then, by a simple calculation, Ut=(1-22-2A.008 mhy"V;, j=0,... If 4>4 we have for h sufficiently small cos nhl >y>1 and hence, if nk=1, say, WO" on 2ZYUPll or7 00 a8 hk. In the presence of stability we may define the local discretization or truncation error tf and show convergence in cxactly the same way as before, and obtain the following theorem: TueoreM 2.1. Assume that ue C4? and that U" is the solution of the forward Euler scheme (2.2) with 1<4. Then there is a constant C=C(u, T) such that 1, since then for some interior mesh point of [0, 1] the equation uses mesh points outside this interval. In such a case the equation has to be modified near the endpoints, which significantly complicates the analysis. The stability requirement k < 4h? used in the forward Euler method above is quite Testrictive in practice and it would be desirable to be able to use A and k, for instance, of the same order of magnitude. For this purpose one may define an approximate solution, instead as above, implicitly by the following equation, which is referred to as the backward Euler scheme and which was proposed first in Laasowen [1949]: wo M-1, 120, 6,Uj**=0,6,UF", j= US*!=Uy'=0, 020, (23) U9=V,=0(jh), O=1,...,M. For U" given this defines U"** by means of the linear system of equations (42a aust erat) Us =U '=0, In matrix notation it may be written as Bo" =0", where O"*! and U” are thought of as vectors with (M — 1) components and B is the diagonally dominant, symmetric, tridiagonal matrix 1428-4 0 0 0 eB 1422-2 0 + 0 | BH o A i42A -A o he L 0 -2 veal Clearly this system may easily be solved for U**!. Introducing the finite- dimensional space [? of (M + 1)-vectors {V;}é with Vy = V,, =Oand the operator By, on I? defined by By Vi =(1+24V; — AV 5-1 +Vj41)=V,—k0,6,V;,, f=l,...,M—1, the above system may also be written Routt pe Bol! us4 V. Thomée Cuaprer 1 or U"*! = By! U"=E,,U". We shall show now that this method is stable in maximum-norm without any Testrictions on k and h. In fact, with 4 arbitrary, we have WU" Lo aSIU" oy 20. (24) For, with suitable jg, Vor w= 15S 1, which is the interesting case if we want to be able to take h and k of the same order, one obtains instead (L$ AO SAU uF QA— DIU" Ico. which does not show maximum-norm stability, since 2A —1> 1. For A<1 we have immediately as before a O(k* +h?) =O(h?) convergence estimate. We shall return later to the question of maximum-norm stability and convergence for A> 1. We turn now instead to an analysis in a real I,-type space. We introduce thus for vectors V =(Vo,..., Vy)" the inner product ah M V.W)=hY VW, imo and the corresponding norm M 42 IV lana, men(r x v3) i=0 We denote by /3,, the space /? equipped with this inner product and norm and note that this space is spanned by the (M—1) vectors g,, p=1,...,.M—1, with components Opj=/2sin npjh, j=0,...,M, and that these form an orthonormal basis with respect to the above inner product, or 1, ifp=4, (op.0.4=e.= ft itpea. We also observe that the g, are eigenfunctions of the finite difference operators —6,6, in the sense that =0,8,0,)=Q/h2\(I—cosmphg,s, j= 1,....M—1 We shall now use these notions to discuss the stability of the three difference:Section 2 Introduction 27 schemes considered above. Let V be given initial data in /2,,. Thon wot V=Y cep, where ¢,=(V. ope The forward Euler method associates with these initial data 1 U}=V,+k,4Vj;= Y ¢,(1—24(1—cos mph)y,, j=1,...,M—1, pat Uj=Un=0, or, more generally, Met Us= LY cok (aph)"g,;, J=0,.-- (2.6) pat where E(@)=1-24+2A.c08 €. By Parseval’s relation we have thus <1 12 WWha=( x etepn*) 0. The discrete problem (2.8) thus reduces to 2M equations at each time level. Similarly we may apply the backward Euler equations (1+ 2AU5** —AUFZ3 + UFZ{)=Uj, j=0, +1, +2,..., 29) and look for a solution which is periodic with period 2M in j, thus reducing (2.9) to a system of 2M equations for Uj*', j= —M,...,M—1, say, with the matrix P1424 A 0 Ow -jA 7 Ah 1424-4 Oo - 0 0 —A 1424 -A 0 -A La 142A This matrix is clearly invertible since it is diagonally dominant. Similar consider- ations hold for the Crank-Nicolson method. Inthe case that V is odd and satisfies Vy —0 the same holds for U"and the systems reduce to our old finite-dimensional systems of order (M — 1). Certain other problems may as well be restated in terms of periodicity require- du Ou a Bx? in [0,1] for t>0, 10,9 =S441,)=0 for t>0, — u(x,0)=o(x) in [0, 1]. Inthis case we may extend v first to [0, 2] by requiring u(x) =o(2— x) and then extend this function to a 4-periodic odd function on R. Looking for a solution with these properties will then lead to a solution of our original problem. Again, the finite difference methods suggested by the above may be applied for the periodic pure initial value problem and yields an approximate solution of our boundary value problem. Let us therefore, more generally, consider the periodic initial value problem, which we normalize to have period 2n, oa _dw a ax” u(x,0)=Ax), ox+2n)=(x) for xeR, and let us also consider a finite difference approximation defined by Yb,U"**(x—ph)=¥a,U"(e— ph) for n>0, > > xeR, 1>0, 2:10) T(x) = of,30 V. Thomée Cuaprer 1 where h—2n/M for some integer M and where the sums are finite. For the analysis of this situation it is again convenient to use Fourier analysis and to work in the space L$ of 2n-periodic functions with norm r 12 Hola. -( [ i}*ax) For such functions we have the Fourier series representation if img. on J o(xje"'* dx, v(x)=) Ge", where 6; i and we recall Parseval’s relation Woll3,«=20 Y |6,. Jere Setting now urd =¥, Oe, 7 we find, by (2.10), xb Uyt een a, U5eu="™, or, with a{E)= x a,e~'Pe tnd similarly for b(), Smimoy tet) af jh) O Fe". i i Assuming b(£) 40 for all real € we thus conclude O5*1=Eh)O5, where E(E)=b(2)"‘a(é), and hence U(x) =P EL ihy'o,e*. i By Parseval’s relation we find for the operator E, defined by U**! =E,U" that 1 NER, «= (25 ieuAr®,?) 7 1. The approximation by finite differences may be explicit or implicit. In the latter case the algebraic system of equations to be solved at a specific time level generally results in a finite algorithm only when the problem is periodic in space. We shall allow the differential equations and the approximating difference schemes to have variable but smooth coefficients. Our main concerns are the stability of the finite difference scheme with respect to various norms and the rate of convergence of the approximate, or discrete, solution to the caact, or continuous, solution of In Section 3 we introduce the basic concepts of the theory, such as explicit and implicit schemes, one-step and multistep schemes, truncation errors, orders of accuracy, and various notions of stability and convergence. We also give a brief account of the Lax-Richtmyer theory of the relation between stability and convergence. Section 4 is devoted to the Fourier method of analysis of one-step constant coefficient methods in the space L,(R*). In this case, using Parseval’s relation, the stability may be translated to the boundedness of a family of powers of trigonometric rational functions, which are matrix-valued in the systems case. The well-known von Neumann necessary condition for stability then relates to the eigenvalues of these matrix-valued functions. A somewhat stronger condition, introduced in a special case in Joun [1952], is shown to be sufficient for stability and to imply regularity properties of the difference methods similar to those of the continuous problem. The Fourier method is also applied to derive estimates for the rate of convergence for parabolic difference schemes. In Section 5 several examples of specific standard finite difference methods are investigated, in light of the theory, with respect to stability and accuracy. In Section 6 multistep schemes with constant coefficients are brought into the framework of one-step methods and particular examples are analyzed. Section 7 develops some of the material from the fundamental paper by John quoted above, in which Fourier methods, somewhat more refined than those of the ic given problem. aa32 V. Thomée (Cuarrer IL earlier sections, are used to derive stability bounds in the maximum-norm for single- step explicit schemes in the case of a scalar second-order equation (which may have variable coefficients) in one space dimension. In Section 8 we describe generali- zations of these results, principally by Widlund, to explicit and implicit, one- and multistep schemes for parabolic systems of equations in an arbitrary number of space dimensions. In the final Section 9 we discuss the rate of convergence of finite difference schemes for parabolic equations and systems. Here more attention than before is paid to the relation between the regularity of the data of the continuous problem and the convergence properties of the approximating scheme, Tu order iv desuribe this the Besov spaces Bs are introduced into the analysis. Direct results are presented which show that a specific degree of regularity together with a given order of accuracy and parabolicity of the method implies a certain order of convergence. Often this convergence is of lower order for nonregular data even when the method has high accuracy, but we also indicate how a preliminary smoothing of data can recover the convergence rate caused by low regularity. We finally present examples of inverse results which make it possible to draw conclusions about the data from the observed rate of convergence. 3. Finite difference schemes In this section we shall, in a more systematic and general way than in Chapter I, introduce the basic concepts relating to finite difference schemes for the numerical solution of the pure initial value problem ou By Pes t, Dut+f = ¥ Pyx)Dut f(x) in R’xR,, Gl) alee u(x, 0)=v(x) on R’, where «=(z,,...,0,) is a multi-index, |a|=a, +---+a, and D*v=(8/0x, F*---(0/x,)*0. Although our purpose is the solution of parabolic problems we shall not specify the type of the linear differential operator P(x, t, D) at present, but simply assume, to begin with, that the initial value problem (3.1) has a sufficiently regular solution for our purposes. We shall return in Sections 4 and 8 below to more precise descriptions of parabolic equations. We begin our discussion here by considering the case of the homogeneous equation so that f =0, and we first consider an explicit finite difference scheme U0) =Y. aglx, nk, AUC — Bh) *Section 3 Pure initial value problem 3B where U" stands for the approximate solution at time nk, where kis the time step and where / is the mesh size in space, B=(,,...,f,) is a multi-index with integer components, and the summation is finite. The space varible x may range either over R¢ or over the mesh points fh. We shall also allow the scheme to be implicit, ic. defined by the equation By,U"*=AE,U", n=0,1,. where Bj,, is an operator of the same form as 4f.,, but where we also assume that #,, has a bounded inverse (in the space of functions under consideration) so that the iatter equation may de soived for U*** and written USB UE The solution of this discrete problem may then, in both cases, be written U"= Exim 1 Exim 2 Fant 0U%, or, if we set U° =v, and introduce the product tn 1 Ean 27° Ems (3.2) as In many cases we shall assume that k and h are tied together by a relation such as k/h=A=constant, most often with q=M, the order of the operator P(x, t, D). We shall then omit h in the notation and write, for instance, E,_,, instead of E, ,,,. When the coefficients are independent of t, we may write simply E, and the solution in such a case is simply U"= Efe. For the nonhomogeneous equation we will use difference approximations of the form Bi,U"'=AL,U"+kME Rf, n=0,1,..., 3.3) where Mf,,is an operator which could be of the same type as AZ, and BE.,, or,¢.g. an integral average such as nek Mis SQ=k* [ L(x, 0) dt. nk We find then, with Mj,,=(Bi,)"'Mfa, USE URES and the solution of the discrete problem takes the form naa UNSERRU°+k Y ERit! MES G4) i=0 i kand hare reiaicd as in the simpie vase thai the cocflicienis are independent4 ¥. Thomée Cuaprer IL above, that 4j,,f-f/ and U°~» this reduces to not - Ut=Epotk Y Bp ip, J=0 The difference equation is said to be consistent with the differential equation if, formally, for smooth solutions of the differential equation, the difference cquation is satisfied with an error which is o(k) as k,h0. More precisely, introducing the truncation error, Tren = KO" (Bi yl! — AE it!) — Miah the scheme is said to be accurate of order in x and vin if, uniformly for 00, so that the requirement for Mj is RSQ + HS FV C)+OM*) as h>0. One choice satisfying these conditions is, as is easily seen, MES(QQ=A SO) +I t HFSS) + TE I"—A). We shall now introduce the concept of stability of a difference scheme. For this Purpose we assume that the functions under consideration belong to a normed linear space 1’, with norm ||. The difference scheme introduced above is said to be stable with respect to if, with C=C,, JEEZo|| 0. Thus, if G/and V° are close to F/ and U®, respectively, then V" is close to U" in the sense of the above inequality. As a special case one may prove the following convergence result: THEOREM 3.2. Assume that the difference scheme is accurate of order win x and v in t, and that E, 4,q 1s stable with respect to 4”. Then, if U° =v, and if the exact solution uis sufficiently smooth, we have U"—u"ll < Clu. Th" +k’) for 00 given, we may choose é such that val 0 as k0. te(0,T] Such an operator is stable if, with C=C,, \Efvll 0 asn+co, nk. The proof of the sufficiency of stability for convergence is essentially the same as our proof of Theorem 3.2 above. The proof of its necessity is based on one of the fundamental theorems of functional analysis, the Banach—-Steinhaus theorem (or the principle of uniform boundedness). been gencraiized iv cover a variety of situationsSECTION 3 Pure initial value problem 39 such as time-dependent, nonhomogeneous or even nonlinear equations, see e.g. Srerrer [1959], THOMPSON [1964], ANsoRGE and Hass [1970]. We shall now briefly discuss the case of a multistep finite difference scheme. Such a scheme is based on a difference equation of the form Buy Ut = Abas Ut + Ab am UO EME af where m>2. This formula may only be uscd for n>m—1 and thus requires, in addition to the natural initial condition U°=v, that U',..., U"~! be prescribed. A common way to analyze such a scheme is to reduce it to a two-level or one-step scheme of the form discussed above. For this pur puse oue ints duces ihe compouid vector-valued unknown function On=(us, U1 yam ty and the matrices of operators AL At ce A am a1 ABn,2 ih, Bi, 0 0 T 0 0 7 oo! 0 Agn= 0 f 0 and Bey=| | . : : : 0 I a a 10 With this notation the scheme can be written as By, 0"! = AL, U"+kF" forn>m—1, O=y, where F*=(ME,,f,0,...,0)". Again we may introduce the discrete local solution operator of the homogeneous equation, Ex nn= (Bh) Abn and define its stability and consistency in the obvious way. In particular. if A” is a normed linear space in which we consider our functions, then 0" is sought in the product space 4” x --- x W” with m factors, and the stability is measured with respect to the corresponding norm, In the discussion of the accuracy of the scheme special attention has to be given the choice of the starting values U',...,U"7'. Similarly as before, with @ =(u", u"~',...,u"""*")" for the exact solution we now have, for n>m—1, n-1 Otek (Oma) YEE tea G9) emt where the truncation error only stems from the first components of the compound solutions. It follows that, in the presence of stability, if the truncation error is Gect_get fh" +4"), say, then the ylubal e111 is of the same order provided O77" —a40 V. Thomée Cuarter I matches this. Note that, since the initial error is a local error, a lower order approximation may be used in the initial steps, because an extra factor k is available in this term, which is not needed to compensate for the summation in (3.9), see the example below. As an illustration, consider for the solution of the model homogeneous one-dimensional heat equation the three-level equation mle) — end rH m eee) (x) _ Ux +h) — Soro en Ay (.10) or, if A=k/h?, U"* 1) =(22U"(x +h) —40U"(x) + 22U"(x~ A) + U9, In system form this may be expressed in terms of 0" =(U", U"~1)" =(U", V")" as U"™* (x) =22. U(x-+ A) —42U"(x) + 22x —h) + V0), vT*(x)= Ux). or Or (x)= (By O"Y(a) [2a 01, 449 1712, [oa o1_ “Lo of? wr] of? “+9 of” (x—h) for n> 1. By the symmetry around (x, nk) the exact solution satisfies (3.10) with an error of Oth? + k?), which translates into we ” tue [ fe [F | soar) as hk. If we assume A constant the order term reduces to O(h?) as k tends to 0. With U° =U! =p the initial accuracy is also O(k)= O(h?) for this case, but if k and h are independent, a more natural choice is U(x) =0(x) +k0,9,0(x) ux, k) | OUP | kh?) as Ayke +0. @G.11) n of stability is now interpreted to mean the boundedness of 0" in U?,U°Y', and this would imply convergence of order O(h?) or O(h? +k?), respectively, in the two cases, for sufficiently smooth However, it tuns out that the present scheme is unstable for any che example, if the scheme is considered at the mesh points jh and we set U,=U(jh) and A =1, ,, We may choose the initial values Oo}SECTION 3 Pure initial value problem 4 where o is a fixed 2-vector, and then find ~, J-sa iyet of af a. For any A the 2x 2 matrix entering here has two distinct real eigenvalues, one of which is less than — 1. If we choose for a the corresponding eigenvector, it is clear that 0" is unbounded as n grows and thus the scheme is unstable. As we shall see in a later chapter this scheme may be stabilized, for any constant A, ‘ing LI" by the average 4(I/"*1 4. 17*~1) co that the scheme hecomes UN *()—UN A(x) Ux +h) U(x) — U1) + UA) sr 7 (3.12) which was proposed in Du Fort and FRrankeL [1953]. As another example we could consider the implicit three-level method derived by selecting a second-order accurate backward difference approximation to Ou/Ot at t=(n+ Ik. This requirement defines the scheme uniquely as (QU"**()—2U%(a) +4" "()/k=0,0,U"* (0), (3.13) and, as we shall see in Section 6 below, it is unconditionally stable. The truncation error here is O(h? +k?) and. in order to match this order we may set U° =v and choose U? by (3.11). Note that thus the first-order accurate forward Euler method is accurate enough to match the second-order method (3.13). Note also that no stahility restriction needs to he imposed for this first step hecause the formula is only used once. We shall end this section by making an observation concerning the accuracy of the Du Fort-Frankel scheme (3.12). Let u be a sufficiently smooth function. With 8,5, as before and correspondingly for 3,5, and with 6,, the symmetric difference quotient, defined by Gyula, 1) = (ul, + k)— u(x, t—W))/(2K)= 4G, +8,)ulx, 1), we have for the truncation error of (3.12) uP 4() u(x) 2k u(x th)—ul* xy— u(x) +(x) Ese =6,u(x, nk) —0,5,.(x, nk) + (k/h)?8,5,u(x, nk) fa ? RO = S ]os ne a osm) +0(2)+ 01H?) +OLk/H"] Consistency with the heat equation therefore req Than) = es that k/h tends to zero, which iga2 V. Thomée (Cuarrer I the case e.g. if k/h? — 1 —constant. However, if instead k/h=1 —constant, we obtain Teheal)= [= oc pou SF omen") as h0, which shows that the scheme is then consistent, not with the heat equation, but with the second-order hyperbolic equation 207 ue ou Ou eet O Ox? We shall return in Section 6 to discuss the stability of the Du Fort~Frankel scheme. 4, L, theory for finite difference schemes with constant coefficients In this section we shall use the Fourier transform systematically to express and analyze the notions of consistency, stability and convergence for constant coefficient single-step finite difference methods applied to the pure initial value problem for a homogeneous parabolic equation, or system of equations, in d space dimensions As we have seen earlier such material is also relevant to the study of initial boundary value problems when the boundary conditions may be interpreted as periodicity 4 rst, we define the Fourier transform over R¢ by a a= fovea ae, x= PF xy, ca Re ; where, as always below, we asume that » is small enough for large |x| that the definitions and subsequent calculations are justified. We recall Fourier’s inversion formula, =n 4—9)=08)~| HE)e** dE, Pa and Parseval’s relation, lo] =Qn)- 4? oll, where, as for the rest of this section, ||-|| denotes the norm in L,(R*). We introduce as above, for «=(1,..., #4) amulti-index and ja|=, +-+--+a,, the mixed derivative of order |a, Dt =(8/0x,)"---(6/Ox,)"*, and recall that ill E06), where *= E41. value problem ‘alue problem (4.1)SECTION 4 Pure initial value problem 4B % _ p= Y, P,D'u for xeR!, 120, or Ja} 0, feR’, 4 \i2 where |é| = (5 3) : c We recall that this is the same as saying that — P(D) is strongly elliptic. Sometimes, we shall allow (4.2) to be an N x N system. In this case the coefficients P, and the polynomial P(é) are N x N matrices and the condition (4.4) is replaced by APGD)<—clEIM+C, — e>0, EER’, (4.5) where A(A)=max Re ,, 2 (44) if {2,}}., are the cigenvalues of A. Clearly (4.4) is satisfied with M=2 for the second-order operator in (4.3) since then ‘ ‘ Re PGE)=— YL patjtr— ¥ (mp,é,+Re po, kad JaL <-elEP +CEl+<—delEP+C,, EER Equations for which (4.4) or (4.5) are satisfied are referred to below as parabolic in ihe scuse of Peixovenu [1937].44 ¥, Thomée (Cuapter IL Letting 0,1) denote the Fourier transform with respect to x of the solution at time t, we obtain by virtue of (4.1) the ordinary differential equation 4 HG )~ PEED) for 20, (6,0) = 6(2), so that, formally, a(g, )=exp(ePg)a(2), from which u(x, t) may be found by means of the inverse Fourier transform. Using this together with Parseval’s relation we find at once for the solution E(t)v of (4.2), IE@ol< suplexptrPGis)| lol, (4.6) and hence, by (4.4), [E@ol| 0, cee so that, in particular, the initial value problem (4.2) is correctly posed in L. This estimate is valid also in the case of a system if we interpret the modulus in (4.6) as the matrix norm subordinate to the vector norm used. In fact, for an N x N matrix A we have with the above notation (ct. ¢.g. GELFAND and SCHILOW [1964) wea lexp Al0, k, s S$ kek, , K. In the case of a system one may use the fact that for a N x N matrix A with spectral radius p= p(A) onc has (cf. Tr L066) radius p= p(A) onc has (of. Toombs [1966}) [A"| N. (This may be thought of as a discrete analogue of (4.7).) Applied to E,(h7 '), using that consistency implies nlE,(h 12) 0, and by (4.20), [E)10. (4.22) Writing E,()" —exp(nk PGs) not = x EXE! (EO) —expkPRe))expUkPCig)), (4.23)50 V. Thomée (Cuarrer I we hence have [ECE exp (nkP(ig))] < CnkhM(L + [EMC < CHAI +E"), AlS;\0, we have, with C=C, 7. [Uw || 0 we have Ke] EM He REM < Cam] | ik OM, Hence, |E,()" —exp(nkP(id))] 0, (5.1) special cases of the so-called @-method defined by 0,U"x) = 00,0, U"* (x) +(1 —0)0,6,U"(x), with @=0 for the forward Euler, @=1 for the backward Euler and @=} for the Crank-Nicolson method. Written in the basic form (4.10) we have (1—0k0,5,)U"* =(1+(1—9)k3,5,)U", and we find for the symbol of the corresponding operator EySection 5 Pure initial value problem 53 where 1=k/h is considered to be fixed. Assuming 0<0<1 we have EQ®<1 for eR, and the stability requirement reduces to 1-401 -O)A min Ee aa or 62 (-2< oo Hence the @-method is unconditionally L,-stable, ic. stable for all 4, if @>4, whereas for 0<4 stability holds if and only if ; 1 4ST Hy" ‘We note that if strict inequality holds in (5.2), then the method is parabolic, since in this case JE@|<1 for 0<|é|4 this is always satisfied, and for 9 <4 the stability condition reads max |E(é)| = max, é I ‘SH y 4 D()=2 ya —cosé;) & =lkP? sy DY EF+ OU) as £0, a and hence, by a simple calculation, ‘ E()=1—-Algl? +734 YY EF +007 1E|* + O(1EI"), it and, since now P(2)=léI?, exp(AP(ic))=1—AJEl? + 4271144 O(1EI9). We conclude that the difference scheme is always second-order accurate and never of higher order if d>1. Returning to the one-dimensional equation (5.1) we shall show now that for any given positive integer v there cxists a unique explicit method for (5.1) of accuracy y= 2v using the (2v+ 1) points x + jh, j= —v,..., v. This method will be shown to be parabolic for A sufficiently small. We shall prove this by constructing an even trigonometric polynomial E(Q)=EE, A= ¥, a(Acos je =o56 V. Thomée Charter II such that EG, Ay=e"*’ +0(@"*?) as E50, (5.3) and such that, for small A, [EG Al<1 for 0<|é|0, r= and, by taking the (2/)th power and replacing z by 4é, t= Y Bi,24(2sin? 36)! itt Bi2Ai—coséy’ with p, 2,>0. TMs i Hence if 10 for 0 0.0). fest with the Q,(£) bounded, This implies (5.4) for small 2 and thus completes the‘SECTION 5 Pure initial value problem 57 For v=2 we easily find Ty,4(2) =2(1 —cos €) + 4(1 —cos )? = $—§c0s E+ 4 00s 2¢, T,4(€)=4(1 —cos ¢)? = 6—8 cos ¢ +2 cos 26, and hence EE, A)= 1-444 32? + (§4—41? Joos & + (A? —42)oos 26, finite difference schemes for general parabolic equations, even in the case of variable coefficients, See also Srranc [1963]. One other way of defining finite difference methods for the model equation (5.1) with prescribed accuracy and stability properties is to proceed as follows. As the purpose is to find a trigonometric rational function E(Z) with the property EQ =e +0(6*7) as 0, (7) it is natural to determine first a trigonometric polynomial A(¢) such that A= Dae” = — 2 4 O(E**2) as E50, and then choose a rational function r(z) such that rz)=e" +O(z"'*) as 740. Then, if 2q>p, we have that EQ =HAAG)=e4® + O(4*Y) Se FOUND 4 O(E2+2) = oA 4 O(EH*), which is the desired property (5.7). Now, if —A4, r=2/(1— 20) if 0<4.and we recognize our above conditions for accuracy and stability for this method. This procedure for constructing a finite difference operator may be thought of, in the following way, as a discretization in the space variable, followed by a separate discretization in time: Replace first the differential operator 0? /@x? in (5.1) by a finite difference operator A,V(x)=h" Da, Vex~ ph), ? and consider thus the initial value problem a =A,U for t>0. Introducing the Fourier transform O(€, 1) of U(x, 1) with respect to x, we find for this function the ordinary differential equation £ 06.9=b-*AE)OE.) for 1>0, (613 or, by integration between ¢, =nk and t,,, =(n+ Dk, OC ty = exp A(hO(E, ty). (5.13) This may now be approximated, with r as above, by OP" *(E)=r(AA(hg)O%E) = Hhg)O"). The property (5.10) is referred to as ellipticity of the finite diflerence operator A,. The use of special rational functions r(z) to replace e” in (5.13) may be interpreted as the use of a Runge-Kutta type method for the integration of the ordinary differential equation (5.17) (of eg Henrict [1967])SECTION 5 Pure initial value problem 59 The discretization in space only, leading to the above system of ordinary differential equations with respect to time, with the values at the mesh points as the unknowns, is sometimes referred to in the literature as the method of lines (cf. e.g. FRANKLIN [1969], Lees [1961], Varca [1962]). Discretization in time only, or the method of RorHe [1931], can sometimes also be used as a preliminary step. These considerations may be generalized to a scalar parabolic equation of the form & _ RDW= > P,D*u, (5.14) a ini=a where, for simplicity, the elliptic operator P(D) only includes derivatives of the highest order M and has real coefficients. We may then first approximate P(D) by a finite difference operator Q,V(x)=h-™ S gy V(x — Bh), (6.15) # for which we introduce the symbol Q(¢) satisfying O8= Lage #8 = Pid + OEM") as E50, B where p is the order of accuracy. The finite difference operator (5.15) is said to be (cf. Tuomas [1964]) if QO<0 for dp, E() =e72© 4 OWE Mt) eH 4 OUE|MN, The discussion of the stability is analogous to above. In particular, if we can construct an elliptic finite difference operator consistent with P(D), then the 0-method with 0=0, 1 or } gives an explicit forward Euler, implicit backward Euler or a Crank-Nicolson method for (5.14), which is L,-stable or parabolic under the appropriate assumption. Recalling that — P(D) is elliptic if —PUd)=—(-I)"? Yo Piet >elEM [al=M for E40, c>0, it is natural to try to construct an elliptic finite difference operator consistent with P(D) simply by replacing the derivatives in P(D) by symmetric finite difference quotients, which corresponds to taking Qe) = Pisin 6)= PA) + O(|EIM*?) as 0, where sin € =(sin é,,...,sin &,). The polynomial O46) thus defined is nonpositive0 V. Thomée Cuaprer I E(2)=r(4Q(2) with r(z) as above. However, as Q(é)=0 if, for instance, 5, this operator is not elliptic in the above sense, and the corresponding finite difference scheme for (5.14) thus not parabolic. In order to make Q, elliptic one may modify the definition by setting : O(é)=Plising)— ¥ (1 ~cosz,, which does not change the consistency since the additional term is of order O(|é|?") as £0. The operator thus defined uses other than the closest possible neighbors at the mesh point x. For instance, for the second-order operator a Ou P(D)= — (D) > Pm the term p,,0?u/0x? is replaced by u(x + 2he,)—2u(x)-+ u(x — 2he,) , 4h? : ‘ic finite diflerence operator which does not have this Py Another possible elli disadvantage is ‘ R= Y Puds,Se jake with symbol ‘ OH)= ¥ pale*—1)-—-e*), 21 which is elliptic as 4 -Q= iL Pl —cos E; —isin E,(1 —cos é, +isin &,) ikea = Yall —cosé,)(1—cos é)+ ¥ pasin €,sin &, Ak dk ¥ (-cosé,). iva veh (cos é,)? + sin? é,)=2c = Another common choice is 9,V0)= x 2 1j0x, 5x, V(x) + & Pda, Ox,V00). where 6,,=40,,+5,,), or ~O9=25 Pull —cos &,)+ >,Pasin é sin &.,SECTION 5 Pure initial value problem 61 which is easily seen to imply that Q, is elliptic. For d=2 the latter choice corresponds to 2 OV(x)= ¥ py(V(x + he,) —2V(x)+ V(x—he,))/h? = + P12(V(x+ he, +he,)—V(x+he, —he,) —V(x—he, +he,)+ V(x—he, —he,)V/h?, and thus uses the nine points x, x-the,, x-the, the. discussion, we know that there exists a trigonometric polynomial of order v such that (cf. (5.5)) AE) = — Ty ay) =- z By,2,(1 —cos &)! it = —0 +0(67"*7) as £0, and such that by (5.6) the corresponding difference operator, —h0,8.¥. is elliptic. The above makes it natural to ask for rational functions r(z) which approximate e* near z=0 and which satisfy the appropriate boundedness conditions. The most commonly used functions of this type are the Padé approximants (cf. e.g. VARGA [1961, 1962]) which are defined by 1e)=r, (2) =e +0(@"*8*1) as 250, where d, , and n,, are polynomials of degree p and q, respectively. One may show that these polynomials are uniquely determined by _ + _et4a-. i "d= & praia il and p 4,42)= ¥ J=o (p+ ate For p,q 1 we recognize the rational functions ro 1, ry,9 and r, , corresponding to the forward and backward Euler method and Crank—Nicolson’s method. It is known that Irpd2<1 for z<0, ifp>q.@ V. Thomée Cuarrer IL For p0. We recall that the Crank-Nicolson scheme for this equation is defined bySECTION 5 Pure initial value problem 6 where A,V=0,,0,,V+0,,04,% and observe that the strictly two-dimensional discrete elliptic operator I — kA, is involved in solving for U"*!. The purpose of the alternating direction method is to reduce the computational labor by solving instead two one-dimensional equations. For this purpose we introduce an intermediate value U"*?? for the solution at t=(n+4)k by the equations uein_y 1k en BU, UM yi or (= 4k0,,8, JU" = (44k9,,8,.U" (149,83, )U" =(1+3k3,,3, JUN. The first step is thus implicit with respect to the x, variable and explicit with respect to x,, and in the second step the roles of the variables are reversed. Elimination of ur*"? gives, sinve ihe vasious operaiors commuie, U"* 1 =E,U" =(1—4kd, M+ 5kO,,8,,)( — 3k0,,5,,) (+ $k0,,8,,)U", and we find for the symbol of E,, with 4=k/h?, 1=A(1 ~cos é,), 1-A(1 cos é,) OT —cosE,) 1+A1—cosé,)" It is clear from this representation that E, is unconditionally stable in L,. By an obvious calculation we have E()=(e*' + OF) (e+ O()) =e +O(IE|*) as 60, so that the scheme is second-order accurate. (In fact, the scheme is easily seen to be second-order accurate in both space and time when h and k are allowed to vary independently of each other.) We shall derive this method in a slightly different way, which immediately generalizes to the heat equation in several space dimensions, ou 6 Ou 3p ANE x ae (5.16) ic fraciionai step method and uses the aux Tie mviiod is uow referred io as64 V. Thomée (Cuarrer IT values U"*, j untild_ pated Until, yntu~ tna ee 1,...,4, with the final value U"*1, defined by or Urtild (1 $0.5)" (U4 $k0,,5,) UNO 4 ‘ UN SB, U"= T] -H40, 5) 440, JU jt In this approach one thus approximates the operators in the sum in (5.16) separately, and obtains U"*! as a product of one-dimensional [?-stable operators acting on U". We have again for the symbol 1-A(1—cosE) EO= T] — ee = TT Ty 7G cosé) =e FP + O(LEI*) as 0, so that the operator E, is of second order. We refer to the special article by Marchuk on “fractional step methods” in this volume for a more thorough discussion of these questions. 6. Some multistep difference schemes with constant coefficients We shall naw consider the stability in I, of some specific multistep difference schemes for a simple constant coefficient scalar parabolic problem of the form a lai=ar (6.1) ux,0)=(x) for xeR’. Let thus h and k be the mesh sizes in space and time and let kh” = 4=constant. Recall that a multistep scheme is then defined by a relation of the form B,U"** =A, U"+ + Ag J"! for n>m—1, (6.2) where m>2, together with the initial conditions for j=0,...,m—1, where the »! are determined in some fashion from theSECTION 6 Pure initial value problem 65 that the 4,, and B, are finite difference operators with constant coefficients, Ayolx)= Layoex—Bh) for | a Byo(x)= z bgv(x— Bh), where the a, and by are constant and the sums finite. Introducing as in | Section 3 the vectors Ge =(u", u"4,...,U"-"*")F and the corresponding matrices A, and B, equation (6.2) takes the form B,0"*'=A4,0" for n>m—1, B e or O80" forn>m-1, (63) where Beta Beta Br‘ Atm I 0 aa 0 A=] 0 i 0 Loo io 4 We say that (6.2) is stable in L, if (6.3) is stable in the product space L? so that the powers of the operator E, are bounded on [3 Introducing the symbols or characteristic polynomials of the operators B, and Ay, WO= Lay and som, EER, af2)= Page, 1 6 the symbol, or amplification matrix, of (6.3) is bg) be) 1 0 0 EG=| 0 1 066 V. Thomée Cuarrer I the matrix £(), The von Neumann condition is pEE)<1 for EER’, where p denotes the spectral radius, and we say that (6.2) is parabolic (in the sense of John) if, with ¢>0, AE) 1, (6.5) where d,, 6, and 6, denote forward, backward and centered difference quotients, or, with A=k/h?, xeR, t>0. Ut 1G) = 2H U(x +h) —2U%x) + Ux—h)) + UP), (6.6) This method was mentioned briefly already in Section 1 and we showed in Section 3 that it is unstable in the maximum norm. To consider it in the present framework we note that here b)=a(=1, and a,(2)=4i(cosé—1) for EER, so that (6.4) becomes 42 —4d(cos E— I)u—1=0. For ¢#2nr this equation has two distinct real eigenvalues whose product is 1 and hence von Neumann’s condition is violated for any 4. This method is thus, as expected, unconditionally unstable in L,. We recall from Section 3 also the Du Fort-Frankel scheme which may be obtained by replacing 2U"(x) in (6.5) or (6.6) by U"**(x)+ U"~ (x), so that the method isSECTION 6 Pure initial value problem 67 Here the amplification matrix is 4a 1-24 Bey=|tr2A 8° TRA], 0 and its characteristic equation is 1-24 1424 we ~ ens Fe p— “1424 The eigenvalues of E(é) are therefore cos E+ 2A Toa y - I. In the former case, we have immediately 24 {1 142 142A and since the product of the roots of (6.7) equals Ilo P= |1-2al1+2<1, we have, with the proper ordering, lI0, ie. the Du Fort-Frankel scheme is parabolic. We shall now show that the method is actually stable in L,. We recall that for any (2x 2) matrix A =(a,,) there is a unitary matrix U which transforms A to a triangular68 V. Thomée Cuarrer II matrix, or m U*AU= a al where #1,» are the eigenvalues of A and I hy My My 2 Fp Mo Wy +4 2M M2 + Az 2tlr M2, so that, in particular, 2 imix Z iagi- jhe Applied to our matrix E(2) this shows |m| <> Fle sel+ | +1<4. in cH The L,-stability of the method is now equivalent to the uniform boundedness of the powers of the triangular matrix mista) (‘ m i n>, nt mee ba)=m Fabs and, since |j;| <1 for j=1,2, to the boundedness of m,(j,, #2). In the case (6.9) we have Im(y 42) l, EER. Yogether these estimates compiete the proof of the L,-stability of the Du Fort-Frankel scheme for any fixed 1<0o. More generally, let us consider the initial value problem (6.1) where P= P(D)is an elliptic operator of principal type, PUd)=(—1)¥?_ Yo P,g<0 for E40. laieae One way to define a multistep finite difference approximation is to proceed as in Section 5 and discretize first in space by replacing P by a finite difference operator and then apply a known multistep discretization mcthod in time to the resulting ordinary differential equation. Let thus Q, be finite difference operator of the form Qnvlx)=h-™ Y° Ogvlx— Bh), ?SECTION 6 Pure initial value problem 69 which we assume to be elliptic, so that OW) =LO,e% <0 for léjl0, (6.12) which we now want to scretize in time. fe equations which has been extensively investigated is the linear multistep methods, which for equation (6.12) takes the form ¥ aU" Fok Y BQ,U" IF forn>m—1, (6.13) izo j=o where m is the number of steps. Assuming that a >0, By >0, it is clear by (6.11) that (0) —BoQ,)"! exists on L,, and we may solve (6.13) for U"*? to obtain Ur t= F (ay —hfo 4) Maj —K#,0,)U" Far t tigation of the stability we consider the characteristic equation Y (4B 21)" 4 =0, which is thus the characteristic equation of the one-step system formulation of (6.13) as well as of the scalar problem. For the Fourier transform of the solution we have Orta z (2 —BoAQ OY” Mes BA QO 4, where A=kh™™. We shall consider some specific choices of linear multistep methods which are used for ordinary differential equations. Let us being with the Adams methods. They consist in writing (6.12) in the form fag Ulta + 1)= U(t,) + [ Q,U(s) ds, i and then replacing U(s) in the integrand by a Lagrange interpolation polynomial. Using thus the polynomial determined by the values at f,,...,f,41— one obtains the m-step Adams-Bashforth method U™ =U" YL by QUI, (6.14) fm where the coefficients b,,; may be found in e.g. Henrict [1962]. If instead we use the70 V. Thomée Charrer IL polynomial interpolating also at t,,, we obtain the m-step Adams-Moulton method (1 kbSoQ,)U"?=U"+k Y, bS;Q,U"I4, Jt where the coefficients b%,; are also easy to determine. By their construction these methods are of order O(k") and O(k"* +), respectively, in time, to which the error of the discretization in space, O(h'), say, has to be added. In view of the relation kh A=constant the total order is thus O(h™!*"™-®) and Myint + 1YMoy OF ‘b For m=2 we have the Adams-Bashforth method Unt =(1+3k0,)U"—$k0,U"!, and the Adams-Moulton method (1—#skQ,)U"** =(1 + 3kQ,)U" — FskQ,U"*, (6.15) with the characteristic equations (14 QE) + Q(E)=0, (1 F40(E)p? — (1+ H40(6))u + 406) =0, (6.16) respect: By direct computation it follows that von Neumann’s condition is satisfied for the Adams-Bashforth method if and only if -AQ(EV<1 for EER’ (6.17) For the standard approximation to the one-dimensional heat equation we have Q(f)=2(cos ¢ — 1) and the condition (6.17) holds if and only if 4 <4. This requirement is thus more severe than for our previous methods and the method is therefore not competitive. Turning now to the Adams-Moulton formula (6.15) we easily see that this method, although implicit, cannot be unconditionally stable. In fact, this would demand that both roots of (6.16) are in the unit disk for all values of p=4Q(2)<0. However as p— — oo the equation becomes fu? +3u—ie=0, which has the root t= —1(4+.,/21) outside the unit disk. The method is thus unstable for large 4 and hence also inferior to our previous methods. We now turn to the method of backward differencing to construct a finite difference operator from (6.12). It consists in replacing the time derivative in (6.12) by the derivative of the interpolating polynomial, based on the values at t,,1s tye ---fns1—m eValuated at f,, ,, to obtain an implicit scheme. This yields a method of the form kau} z BU" 7] =0,U"*!, iatSECTION 6 Pure initial value problem n or (Gq hOQU = SF By 2A (6.18) a It would also be possible to construct in this way an explicit scheme by evaluating at ty, Le. eu ait hous & pau (6.19) or, at some earlier point t,_,_»j=2,...,m. Since the order of accuracy in the replacement of the time derivative is O{k") we find that the total error, as in (6.14), is O(h?+k"), or, with 4=kh~™ constant, Ofhminta.sem, For m=1 these schemes reduce to the backward and forward Euler methods discussed above, but since we are interested now in multistep methods we consider only m>2. We briefly discuss the stability of these schemes and begin with (6.18). The characteristic equation is (Cm — AQ)" — YL Byte F=0. (6.20) For m=2 and m=3 we have, in particular, for (6.18), G—kQ,)U"*! =2U"—4$u""*, (6.21) fd — kQ,U"* 1 =3U"— SU" + 4U"-2, In order to decide whether the roots of the characteristic polynomials correspond- ing to these methods are in the unit disk we transform it to the left half-plane by setting z=(1+)/(1—n) and apply the Hurwitz criterion: In order that the equation Lo 7)1/=0 with y.>0 has all roots in the left half-plane if suffices that D,=1,>0, Mr 3 Ps Yaka Yo Yo Ya “Vana OM ¥ Yx-3|>0 for k=2, 0 Ye where ;=0 for j>m. We now apply this criterion to show that the method (6.21) always satisfies von Neumann’s condition. In fact, setting p=AQ(€) we obtain for the transformed equation (4—p)y? +(2—2p)y— p=0. (6.22)nD V. Thomée ‘Chaprer I Here for p<0 we have D,=2—2p>0, For m=3 we have EP pn? +(6— 3p)? +(2— 3p — p=0, and, for p<0, D,=2-3p>0, 2-3p P-p|_. D, =p 6-3p) = 8p? — Fp +12, and 2-39 B-p 0 Ds=| -p 6-39 0 |=@P—p)D,>0. | 0 2-3p %-p| Both methods thus satisfy von Neumann’s condition for any 4 and are parabolic. For m=2 and 3 the explicit schemes (6.19) are SU" =kQ,U"+4U""! for n>1, and fu" 1 =(—444Q,)U"4+ U1 —4U"-? for n>2. For the first the characteristic equation is 1? —2pp—1=0 with the roots w= p+./p?+1, one of which is always outside the unit disk if p <0. Hence this method is unstable for any choice of 4; it contains (6.5) as a special case. For the second method the characteristic equation is 12 +3(3—p)e? —3n+4=0, and with p=0 the roots are z= 1 and p=4(—5-+./33) so that this method is also unconditionally unstable. We shall now descrihe two higher-order three-level methods fram Nona as andSECTION 6 Pure initial value problem B Gunw [1963] for the model heat equation in d space dimensions (with d<4), wn xeR, 120, (6.23) a we D5 with initial data u(x, 0)=0(x). These methods will have the special feature that they only use the immediate neighbors at a given mesh point, which 1s, of course, particularly desirable when applying them to a mixed initial boundary value problem. However, as we shall see, the stability results we shall present do not have exactly the same character as the ones discussed above. Let us first note that for u smooth, with ux the) — u(x) — 2,00) = nk), we have aw Row oy =2.Beal Ty gag + OCH) =40, Ou" tutta” » Be OU +e) as h,k0. Setting, as in Section 5, A, 2 2.8.9 we have then nett ST at tut) — BP 35 er towteey (6.24) Thus, if we can approximate the sum appearing as a coefficient of h? on the right to second order in h, then we will have produced a O(h*+k?) finite difference approximation for (6.23). Note that, for the exact solution of (6.23), «atu ou Aus Yt? Yea Au, =, er OX; it Ox1Oxz4 V. Thomée Chapter I and hence, using the fact that A?7u=u,,, wet —2utpur! — -~o -2 F 0,5,0,5,u"+ OW? +K2) as h,k0, (6.25) iy 8s, and also, since A7u=Au, © Oty (ytttigent ean (Ce) \ on / =2 F0,,5,,0,,0,u+ OW? +k +h4/k) as hk. (6.26) ici The relation (6.24) together with (6.25) and (6.26) thus suggest the following two difference analogues of (6.23), namely CC = “=4A(U"! 4Ut UY) Ak qr =2Y 0,8202, uw") (62 and, with A=K/0?, ~ yeu =4A,((1—1/(84)U"* + U" H+(L+ABANUT) + 3h 2B 2s8.020,0" (6.28) The crucial matter is then the stability of these schemes. In this regard we have the following two results from Douctas and Gunn [1963] (where they were expressed for a cubic domain and in discrete /, -norm), TueoreM 6.1. The difference scheme (6.27) is conditionally stable in L,(R‘) ford <4 in the sense that if U° =0 and 1 =k/h? is hounded away from Q and 0 (or hounded away from 0 if d<3) then, with ||-||=lI+ In yeas JU 1. (629) THEOREM 6.2. The difference scheme (6.28)is unconditionally stable in Ly(R*) for d<3 in the sense that (6.29) holds if U° =0. We note that these results do not show stability in the sense of boundedness for all U® and U', but are restricted to data with U°=0. ication of the Fourier transform, equations (6.27) and (6.28) may both be written in the formSECTION 6 Pure initial value problem 15 and the corresponding characteristic equation is HOw? — a, (Ee a,(2)=0, with roots j4,= 44,(¢), j= 1, 2. The solution then satisfies (assuming distinct roots; the case of coinciding roots has to be treated separately) G)=c Dish" + co(Oya(hey', nB2, with initial conditions c4(€) +e2(€)= 0%) =0, (ue (he) +e2(2)u(he)=0 (8), which implies HCH — walhey" Hy (h2)— (he) The stability stated then requires the boundedness of the ratio on the right as hé © R*. One may show that for the method (6.27) this ratio is bounded by H644447' 46) if d<4 and for (6.28) it is bounded by 4 if d<3. Although this stability concept is more restrictive than earlier, it suffices nevertheless in order to derive convergence estimates if we choose U® =u°, as then cror29=U9 49-0. ¥ Ulis oe= (6). chosen so that Ulu" || 1 from W" and W*~! namely hy defining intermediate16 V. Thomée Cuaprer II valucs Weth}, 7. aeayunnts (5 8-4, )W"— A.W = ine (4 B)W' Wht BW=0, j= and then setting weet per ke, ay set 5, 44/4, 10,5, j-23, and for (6.28), By=— FL 1/BAR.Bz, j= 12,3. In both cases (for d=3), it is shown in Douctas and Gunn [1963, 1964] that [Wow lla nS CO +k), in the former case for 4 bounded away from 0, in the latter for 1%. 7. John’s approach to maximum-norm estimates In this section our main purpose is to describe some results and techniques developed in the important paper of Joun [1952] concerning the maximum-norm stability of explicit finite difference schemes for general second-order parabolic equations in one-space variable. We shall also include some material concerning related work based on Fourier analysis, such as discussions of unconditional maximum-norm stability of certain implicit methods and of the use of Fourier multipliers in stability analysis. t the general Ou So pene os ate ot +po(x. ut f(x,t) in Rx [0,7], ma) where p,(x, )>4>0, under the initial condition u{x,0)=0(x) for xe R, (7.2) and an explicit finite difference approximation of the form U"™()=Y ai(x, t, NU"(x— Ih) +Kf (x, ) for n>0, ‘ (73) U%x)=0x), xER,SECTION 7 Pure initial value problem n where (x, #)—(jh, nk) varies over the mesh points in & x [0, T]. The summation is over a finite set of ts, |l| vi which proves the result in the case considered. The case of variable coefficients may be considered as a perturbation of the constant coefficient case. We sketch the proof of the stability in the case that the coefficients depend on x but not on t or h, so that Ep)=E a,(x)o(x— Th). Letting U"= Efo we shall want to estimate U"(xo), for xq arbitrary, by the maximum norm of v. We then fix the coefficients of E, at xq to obtain the representation E,olx) =, Gyo(x— th) + ¥ by(x)o(x— 1h), T T82 V. Thomée Carrer I UP WO =P gUN(x — th) +E b,C)U"x— Ih) T 7 =E,U(x)+ 00), where the latter equality defines E, and f*, and hence nxt UMxo)=Etvlxo)+ LER! Xo)- (7.10) m=o Setting Ef o(x) =, 4,0 —ph), 7 we may write ERS") = Spf 1 (Xo — Ph) ? =r Sno 2. bylXq— phy U"™*—™(X9— (p+ Dh) ? = Ymil%o)U"- "(Xo ih), (7.11) where ImlXo)=Z Gn, 5-121 —jt+ ih). T Setting now B(xo, D=Y bio —jh+ the, T we have x He | Bevo, Blo, De az Ymsl Using the consistency relations one may show, by arguments similar to the ones used to estimate the a,, above, for 4 fixed, nk < T, that with C independent of xo, _f1 iol Ino 0. (7.12) This is relevant for the analysis of the convergence of difference quotients of the approximate solution to derivatives of the exact solution of (7.1) and was used in Joun [1952] in the study of smoothness properties of the exact solution. The solution U" of the difference scheme defined by (7.3) may also be written in the form UE) =HY. dy HOCH) + bY Dyers PMC TA) ; wo or m1 AY dn.0,500j-1+h x Laue jin and the coefficients g,,,,;,(#) may be thought of as determining a discrete fundamental solution. For the particular case that the coefficients of (7.3) are independent of x, t and h, we are in the situation discussed in detail in the proof of Theorem 7.2 above, with Sram, j= Nyy, pte The maximum-norm stability may be expressed in terms of the fundamental solution as sup h Yam iKI0 and yl, is an even natural number, such that EE, + = EC) expla, —ByL(1+O(1)) as E10. (714) The proof of this result may be carried out by the above technique of John. It follows, of course, in particular, that von Neumann’s condition is necessary for stability. For an operator E, which is consistent with the heat equation condition (7.14) is satisfied for €,=0, with «,=0, 8,=/ and y,=2. If there are no further points ¢, in |é|0, a constant C=C(A) such norm Eto < Chol. Since it would be desirable, in practice, to take h and k of the same order, and thus A of order 1/h, it is of interest to ask whether the same constant can be chosen for all 4, such as in the case of L,-stability. This problem was considered in SerpYUKOVA [1964, 1967] who showed | Efol|<23|o], nzl, 4>0 (for earlicr related work, cf. also Juncosa and Younc [1957] and Wasow [1958)). Somewhat more general results were obtained in HAKBERG [1970] and NORDMARK [1974]. We shall now briefly describe Hakberg’s result. Consider the initiai value probiem for ihe parabolic equation ou 1 Mu a ae (7.15) M=2m, xeR, t>0, and consider a consistent finite difference operator of the form E,v=R(kA,)v, (7.16) where R is a real rational function and Ayx)=h-™ Yo dplx—jh), Liles deR.86 V. Thomée ‘Cnaprter If The operator thus defined is consistent with (7.15) if R(y)=1—y+oly) as y0, (1.17) and A(B)= —EM+0(EM) as E+0, (7.18) where A(Z) is the symbol of Ay. The symbol of E, is then E()=R(AA(E), where A=kh-™, and we say that E, is uniformly maximum-norm stable if, with C independent of A, |Efvll 1, O02, + (7.21) holds. Also, M,(+) is a submultiplicative norm, so that M,{ab)0. It is easy to see that the second quantity is independent of # so that the conditionSECTION 7 Pure initial value problem 89 reduces to MfECY)0. (7.22) It follows from (7.20) and (7.21) that stability in maximum norm implies stability in L, for other p. We remark that it follows from our above discussion that the condition for stability in C(R) or I,,,, is the same as in W,,. In order to use these notions to prove stability it is thus needed to have access to some method to bound an expression such as the left-hand side in (7.22). One may then first show that in order to estimate a 2n-periodic multiplier a it suffices to estimate qa where is a function in C§(i*) which is identicaiiy i for j¢;j<7, JA 1,...,d, or M,(a 4d, i(2v) Ma)0, which is the desired result. The technique just described may be applied to give a simple proof of the sufficiency of the conditions of Theorem 7.3 and are also useful to show convergence estimates, cf. Section 9 below.90 V. Thomée CHaprer I 8 Stability of difference schemes for general parabolic equations In this section we shall consider the stability of finite difference schemes for general parabolic equations and systems. We first discuss equations or systems which are parabolic in the sense of Petrovskii and difference schemes which are parabolic in a sense generalizing the one introduced by John and described in Section 7 above. We also show analogues of some known smoothing properties of parabolic problems and touch upon the possibility of using such properties as definitions of parabolicity of difference schemes. thus first consider cyuations of the form Pla, Dz YL Px,Dy, (8.1) tain xeR’, 1>0, where, with a=(a,,...,a,) Dtu=(2/0x,)" ++ @/Oxy"u, lala, +o bay The equation is considered, as usual, under the initial condition u(x,0)=0(x), xeR*, Such an equation is said to be parabolic if P(x, t, D) is elliptic, or more precisely, if the real part of its characteristic polynomial, Pynig= Yo Pix, megs, ie is negative-definite, and uniformly parabolic in a domain in Ré x R,, if, for (x,¢) in this domain, Re P(x,t,it)<—clé|" for EER? with c>0. (8.2) It is clear that in this case M has to be an even positive number. We shail allow below also that (8.1) is a system of N equations in N unknowns is i ding u be au N-vevion u=(4,,.-.,uy)” and ihe P(N x NY matrices, which we always assume to be sufficiently smooth for our purposes. In the ion (8.2) by a corresponding condition for the eigenvalues. Setting, as in Section 4, for any (N x N) matrix with eigenvalues {4,}7, A(A)=max Re 1, i we thus say that (8.1) is parabolic in Petrovskii’s sense if AUP(X, 4,12) < CEM, GER, 620,SECTION 8 Pure initial value problem o1 solution I'(x,t,y,s) defined for x, ye R4, 0 0. (8.3) The fundamentai soiution aiso has the property [D2 DST(x, t,x +z,5)] SOL 5) 4+ 0 exp(—efziMt—s)- 9), with the power of t—s on the right independent of «. (In the constant coefficient case I depends only on z= y—x and t—s and thus I'(x,t,x+z,s) is independent of x.) It follows, in particular, from the above representation and (8.3) that the initial value problem in (8.1) is well posed in L,=L,(R‘) and W"=W2(R°), the Sobolev space defined by the norm Iole= ¥ ID%I,,- (alm Further, the solution has the regularity property corresponding to our previous result in L, for the constant coefficient case (cf. (4.9)): With u(x, t)= E(t)v we have, for 10, (8.5) where A,,,=A,(nk) and B,,,=B,(nk) are difference operators of the form Axo) = Y aglx,t, h)olx~ Bh), é By(ttx)= Yb, h)o(x— Bh), 7 with ay, b, smooth functions of x, ¢ and h and where 4=k/h™ = constant. We assume that B,(c) is invertible (in the appropriate L,-space) so that (8.5) may be written UN tS EU" with Ey,= Bint An or92 V. Thomée Cuapter I The scheme is stable in L., if UL, 0. (8.7) The following important result as then proved by WipLUND [1965, 1966] (cf. also Aronson [1963b]). Tueoren 8.2. If the difference scheme is parabolic in the sense of John, then it is stable in L, for 10 (6 arbitrary in the explicit case) 1254p x91 6, (n—m+ lk} The first of these are analogous to the corresponding estimates (8.3) for the continuous problem. The second estimate, which is not needed for explicit schemes as 6 is then unrestricted, shows that the contributions for large j are exponentially small. The above results may also be generalized to certain multistep schemes. Foliowing WIDLUND [i966] we thus consider schemes of the form BRU" 12 ABU AE QUE, (89) where the operators A%,;, j=1,-..,m, are of the form Aj,;=A,, Auk), where Ay Atwe) =a + Yay g(x, t,h)v(x — Bh), ? and similarly B(t)0(x)= Ay, o(0)0(). =1+ Yo glxyt, h)o(x— Bh), P with the sums extending only over finite sets of multi-indices 8. We note that the94 V. Thomée CHarrer EH natural generalizations to variable coefficients of the multistep schemes discussed in Section 6 may be written in this form. As earlier, in addition to the initial condition U® =0, itis necessary here to supply methods for the determination of U?, ..., U"~}, and we shall assume that this is done in some appropriate way, in particular so that these values are suitably bounded by the initial data v. In the same way as previously discussed such a scheme may be represented as a two-level scheme for an associated system in the variable 0" =(U",U"~},..., U™-"™* 1)", namely Br = AO", or (BYR ce (BD Alon On 8, 0"= Similarly to above we may introduce the principal symbol of the operator Eng = Ex(nk), Beste *A lx.) ~| Ex, t,)= Bix, t,x, = ' | where Alx O=al+ YL ajglxt, Oe", jf B and correspondingly for B(x, t, &). For a consistent scheme we have, in particular, E(x,1,0)=Al, where a % ... , 10... 0 A-i 0 1 0 0 0 i 0 It follows that the houndedness of A" n=O, 1, h we shall refer to as stahilitySECTION 8 Pure initial value problem 95 ofthe matrix 4, is a necessary condition for stability of the scheme (8.9), or of E,. AS is well known, A is a stable matrix if and only if all eigenvalues of A are in the closed unit disk and all eigenvalues on the unit circle are simple. Recall also that these eigenvalues are the roots of the equation mm Bop We say that the multistep scheme (8.9) is parabolic (in the sense of John) if B(x, t, 8) <1 —cl EM for iGiSm fei, d, uniformly in x and t. The following is now the main result of WipLUND [1966] about the stability of the multistep parabolic scheme. THEOREM 8.4. Assume that the multistep scheme (8.9) is consistent with (8.1) and that Aisa stable matrix. Then the scheme is stable in L, and VEU", 0, we have [eU"-D*ut-,t),, 70 ask+0, nk=t, For another concept of a parabolic difference operator, of single-step or multistep type, sce Ionkin and Moktn [1974] and Moin [1975]. This concept is associated with the use of a discrete Fourier transform in space and a discrete Laplace transform in time, and results in a priori estimates in discrete analogues of the Sobolev spaces W™. 9. Convergence estimates In the preceding sections we have shown a number of convergence results (cf. e.g. Theorem 3.2) to the effect that a stable finite difference scheme which is accurate of order 1 d the exact col emooth ig convergent to that came order pr.98 V. Thomée Cuarrer IT enough. We also know by the Lax equivalence theorem (Theorem 3.3) that convergence follows from stability and consistency without any regularity hypo- theses on the data other than that they belong to the space of functions under consideration. Our purpose in this section is to give more precise information about the relation between the rate of convergence and the smoothness of the exact solution. Consider thus an initial value problem Ou 5, = Pest, Du = Po PAx,oDu, xe’, 1>0, (9.1) lice u(x, 0)=0(x), xe RS, where the equation is parabolic (in the sense of Petrovskii), and a consistent finite difference scheme of the form U" SE, U"= BrP Ang", n>0, A=k/h™, (9.2) where Ayq=A,(nk) and By, = B,(nk) are finite difference operators of the form uscd in (8.5). We shall also assume that the scheme is accurate of order y. In Sections 5 and 6 we have become acquainted with several examples of difference schemes of various orders of accuracy and we aiso quoted a resuit in Kreis [i959a] where it is shown that schemes of arbitrarily high order of accuracy exist in great generality. In order to express our results we need to introduce certain Banach spaces, known as the Besov spaces B;*. Recall from Section 7 that W,= W, (IR) denotes L,(IR") for 10, we introduce the modulus of continuity in W, of order j,j= 1,2, by wp s(t, ¥)= sup I(T, —¥ollw,. Ilse where T,v(x)=ux + y). Let now s>0 and 10, we have for 1s5, or 3;=5;, dy wt then v=0. This shows thus that if the rate of convergence in W, is O(h*), where s0, . Amr XER 120 (0.7) and a corresponding constant coefficient scheme defined by E,ox)=Sa,0(x—jh), k/h?=A=constant. (9.8) J TuoreM 9.5, Consider the initial value problem for the parabolic equation (9.7) and a corresponding scheme (9.8) of order 1: which is parabolic in the sense of John. Let 10 and assume that ve L,, is such that \U"— a", 0. Then there exists a v € BS, such that lim sup h7*|U"—u"l|,,_ >0. k=1 ko In spite of the above, it is, however, possible to attain a O(h") convergence rate in the maximum-norm under weaker assumptions than v € B,(s0, oy Jx7X), X20, XalX) = {r 320 (9.10) Clearly many functions of one variable occurring in applications are linear combinations of functions of this form. As is easily seen, we have y, € Bs if and only if s 0, b! and b‘ coincide with multipliers on AL, for |é| <25 and |¢| > 6, respectively. Since the multipliers on FL, are simply the functions in L.,, the above conditions are seen to be satisfied for p=2 if GO=1+ O(lE") as 0, and, for any muiti-index 6 #0, G)=O(E—2fml") as €>2r, uniformly in f. Special examples of smoothing operators of orders 1 and 2, respectively, in the case d=1 are a2 MY vx=h7* [ v(x—y)dy (9.15) ana and kh MP (x) =h-* [t-te (9.16) oh and for general y, a smoothing operator of order can easily be constructed in the form M¥o(x)=ho! {ve ~ty)o(x—y) dy, where ys function which is 4, eA) for odd and (— jr 1 ys —!)for} ecewise a polynomial of degree (y~1) and which wen, Tn fact, MY,104 V. Thomée Cuarrer I J=1,2, corresponds to 60=( a), j=l and, generally, p,(sin $2) Ger * where p, is the polynomial of lowest degree such that p,(sin 2) =&" +O") as £0. For p=2, the operator M,, corresponding to afl K<6 so= fh I> 6, with 0 <6 1 (9.17) Proor. In order to shed some light on the mechanisms that are involved we shall present a proof for the one-dimensional heat equation in the case p=2. Recalling that then ag, =e" HE), FUnEP TLE) = FuSecTION 9 Pure initial value problem 105 we see at once by Parseval’s relation that the proof of (9.17) reduces to showing |E(hE'G(hE)—e- "| we have, by (9.14), EEG) < CEE)" -Isin 32)", and since the right-hand side is periodic it suffices to estimate it for |¢| <7. We have therefore, using the parabolicity of E,, [BCE GO| < Clee < Cn“? for |e|>n. Since le™*G(O1n, we conclude that \I|6, respectively, so that the parameter p in (9.14) is replaced by v. Our previous smoothing operators are then of orders (, 2) and it is easy to see, for instance, that the (9.15) ha. 4) Th ke d {0.48} ha: (2,8) Thi106 V. Thomée (CHAPTER IL TueoreM 9.10. Assume that (9.1) and (9.2) are parabolic in the sense of Petrovskii and John, respectively, and that (9.2) is accurate of order . Let d1—s. Then for U" the discrete solution with U°=Myp we have |U"=w'l, 0, u(0,)=u(t,t)=0 for t>0, (10.1) u(x, 0) =0(x) for 00. For mesh functions V=(Vo,..., Vy)" and W=(Wo,..., Wy)" we introduce the discrete inner product M WV, W)=h ¥ VW; ime and the corresponding norm ivi= (a vi) (104) 7 We shall employ our previous notation for forward and backward difference quotient operators 0, 8,, 6, and 6,. For mesh functions Vand W as above, which satisly Vj=V,,=0 and correspondingly for W, we may define, eg., Moi @.V,W)=h DY 0.V;-W;, ia where we note that the summation ends at j= M —1. With the obvious corres- ponding expression involving 6, we have by summation by parts 0,V, W)= -(V,3,W). (10.5) We now consider the backward Euler method for (10.1) defined by 8,U5=0,a_ 20,0); forj=l,...,M—1l, n>1, (10.6) Uj=Uy=0 = forn>1, U9 =V,=0(jh) for j= where we have set (a 1,2)! =a 2 = a(jh—$h, nk). Explicitly the difference equation may be written US UP" ahs se(UFe 1 — UV A~ a5 1ya(U5- UF )/h k h . It is easy to see by the arguments employed in Section 2 that this finite difference scheme is stable in the maximum norm, but we shall now use energy arguments to demonstrate its stability with respect to the discrete [,-norm introduced in (10.4). We thus multiply (10.6) by U7 and sum over j=1,...,M—1 and obtain, noting that U3=U%=0, that (a 23,07), UY), 0.7) fa i329, um umSECTION 10 Mixed initial boundary value problem 113 By the summation by parts formula (10.5) we have (,(2_ 4,25, UY", U")= (a 1/25, U", 8, U"). Further, we note that 8,05 Uj= ko '(Uj}— Up") UF = tk Up? (UP HUF UF =43[6,(U})? +6, U5)7]. (10.8) Consequently (10.7) yields, 40,|| U"||? +3 |6,U"||? + (a 1/25, U", 8,U")=0 for n>0, wed from which we conclude at once, for instance, SU"? <0, or jutsu", whence, by summation, Ur <|U =I", which shows the [,-stability for any choice of h and k Let us now consider instead the forward Euler method, 0,U5=0,(a_1,0,U)} forj=1,...,M—1, n>0, 3=U"=0 for n>0, US=V,=0(h) for j=0,...,M. The analysis is similar but (10.8) is replaced by 8,U3-UF=H18,(US? —k@O,UIP 1, and hence the energy identity (10.9) by 48, U" |? + (a 4,20, U", 6,0) =4k a," ||? (10.10) Since obviously 10, ¥ || <2h- "17, (10.11) we have 11,0" || = O(a 120,U)"| <2h~* Ja" 23,0", and hence from (10.10) and using (10.2), with A=k/h?, POU + HAE 2175.0" 2eWK Man 2a ym, < UK iat aU,4 V. Thomée CHaprer IL Therefore, if 1 is chosen small enough so that 2K <1, (10.12) where K is the constant in (10.2), we may conclude that aAUrieso, or ju" ].<||U"l, which is the [,-stability desired. The symbol of the difference operator is in this case E(x, t, €)= 1 —2da(x, 1)(1—cos €), (10.13) and we recognize in (10,12) the yon Neumann stability condition. In order to consider similarly the Crank—Nicolson method 8,Uj=0,AP BAU + UF"), J=b.-M—1, nB0, where a"*"/?=a(+, nk +4k), we set Up =UT+ UT), and obtain (0,U%, UN 12) 4 (ath) P5,U" 12, B,U8* 12)—0, (10.14) Here (@,U", UN 2) (UN? — Unk, U4 UT) =40,|U"?, and (10.14) thus immediately yields aU"? <0, which shows the unconditionai stability in this case. For the 0-method, BUF 2A 28,UF%), Ff (10.15) UR =UYt=0, n>0, Up =n), where a"**=a(-,(n+O)k)SecTION 10 Mixed initial boundary value problem 115 we have similarly 0,0", U"*9) + (a,8,U" 48,0" )=0, (10.16) Here @,U*, U"**)=40,|| UI? +0 —H)k||0,0" 7, (10.17) and hence, if 0>4, we obtain at once aU"? <0, witich shows unconditionai stability. For ¢<4 we may use the difference equation (10.15) and (10.11) to obtain (8,0 <2h-* a" 28.8, , so that, using also (10.2), 20,|]U"|? + | (ae f2)'75, U9)? <4 — AKIO, U" |? <(1~20)2AK ||(@U'f)"78,U" "7. The stabi 24(1—20)K <1. fuuk at the proofs abuve we for the backward Euler metho ty now follows under the condition ae Ur? +k Y [S,U7 |? mei For the forward Euler method we obtain similarly \UP+k Y |u|? 4 we find at once 8, l\(@_ 42)'78,U"||? <0, (10.22) from which (10.20) follows, in view of the bounds (10.2) for a. For 0<4 we have instead 0,0" |? +4, @_ 1j2)78,U"|? S<(3—O)Ah?K 0,6, U"||? <2(1 — 20) KA|}0,U"||?, and hence, since we now assume the mesh ratio condition (10.19), that (10.22) still holds, which again implies (10.20) Our abuve Lstabi equation nunhumugenevus Gu_ Of du fen (att Osxsl, 120, (10.23) with the same boundary and initial conditions as before. For instance, let us consider the @-method ,U}=0,(a92 BUF) + F, jal...,M—1, 30, (10.24) where F* could be chosen, for instance, as f(-, (n+ 0)k). Multiplication by U3*® and Summation over j gi ty iv (10.16), that (@,U", U"*) + (aef25,U" 79, 5,0") =(F", U8). For 024 we conclude, as above, for ¢>0 arbitrary, 39, UP +40, U |? 4 we conclude at once 8, N@— 12)"78,U" |? < FMP, and hence easily |],U"I?< C6, U9? + Ck y | FM? for nk0, with the boundary and initial conditions used above in (10.6). Employing the same approach as before we obtain now ©,U*, U") + (a 1)2)478,U" |? =(10,U"+ BBU", U") S, Hence, for small k, with different constants C, Wun Wu? se? U2, 2Ck and, by repeated application, JU" 0, 4k 18,U"? <2A(lal 5 ,28,U"| + Chil3,U" || + Ch"? S2AK(1 +8) (@% j2)"70,U" |? + CAC], UI? + 0" |?) 0, .. u(x, 0) = v(x), where a is a smooth function satisfying (10.2) and a and f are smooth bounded functions of t. In this ca: hh x-0 and and we set, for M a positive integer greater than 1, h=1/(M—1) and x;= —}h+jh,j=0,...,M. We then have xg= —JShand x,=1+4hand these points are thus outside the interval [0, 1] whercas the remaining points x,,..., xy are interior mesh points. We now pose the backward Euler type difference equations at the interior mesh points 6,U}=0,(a_1/26,U)} for j=1,...,.M—1, n21, (10.34) and supplement these equations with the boundary and initial conditions 6,Ut +4a"(UT + US) =3, UY, + 3B"(Uy + UY-1)=0 for n> 1, 8 (10.35) U=o(jh), j=1,...,M—1. The approximations of the boundary conditions have thus been chosen symmet- tically around the points x=0 and x=1 and are therefore second-order accurate. In analyzing the stability of this scheme by the energy method it is convenient to indicate the points included in the summation in the notation for the discrete inner products and set YMen=hX VW m where r=0, 1, 2 and s= M —1, M, depending on the case at hand, and similarly for12 V. Thomée Chaprer I] the discrete norms. Let us recall the summation by parts formula. OV. Wqg=h YA VW, = V8 Wrst. t Vers We VW, In particular, considering a mesh function U" satisfying the Neumann boundary conditions 0,(4- 25, U), U2 = (4% 28,0" BU 2,1) + 125. Ig U1 425, UF UF = (a2 23,UB.U"2,m 1) (10.36) We now apply this to a solution of (10.34), (10.35) with a=8=0. We multiply (10.34) by Ut and sum over the interior mesh points to obtain @.U".U" (O(a, .5,U)". U"), and hence, in view of the boundary conditions and (10.8), SG HUME any | SRO ray 4 (24 328.0% BU 2,41 for n> 1. We conclude as carlicr that 9, IU" lia) <0, or WU" lay SU" Nae ay which shows the stability in the case of Neumann boundary conditions. Note that the argument automatically shows also the existence of a solution of (10.34) and (10.35) since U"-! =0 implies U"=0. Consider now the general case of the boundary conditions (10.35) with «and £ not necessarily vanishing. In view of (10.35), the identity (10.36) may now be replaced by (4-128, 0% Ua. = = 1128, U", BU 2,01) t Oe 1/20, Ue Ue 1 47 28,0 Uy SAE 28, U5 BU 1 + UWB? +1U FP + [UI 12? + IUD) It is easy to prove that for ¢>0 arbitrarily small the a priori inequalities Ham tColiU ia vySECTION 10 Mixed initial boundary value problem 123 and [Uol?<2U,/2 +27, U, 7 <2|U,/ + 2hl15, Uae <2C+H)6,U [fray + CMU ll’ ae, hold, and similarly for U yy. Multiplication of (10.34) by U7 and summation over the interior mesh points therefore implies, in view of the general boundary conditions (10.35), that 38,1 U NG ae 1 FRI, UE ae 1) +40, U lan l, BUT + Ja(UT + U8)— GS, nel, (10.38) BUN +3B(Uirt+ Ub J= GI, n> 1, Up=ujh), j=1,...,M—1. Proceeding as in the above stability analysis we now obtain instead of (10.37) BO, |U" [cae 9 SKID, 02 ae) + 10¥U" lao 0, a 4 (AU,U)= Y (yd,,U,5,,U)27 Ye, U2. (10.41) agen at We shall now apply the @-method to the present situation and thus pose the to the solution # of (10.39) at r= nk, where crete problem to find an appro’126 V. Thomée Citaprer UI k is the time step, as a mesh function U" on Q,.T, for n>0, defined by ,U"=04,U"** +(1—9)A,U" in Q, n>0, u"=0 only, n>0, (10.42) U%(Bh)=v(Bh) for BheQ, where 0, and 9, as usual, denote a forward difference quotient and a parameter with 0<0< 1, respectively. Given U* equation (10.42) requires us to find U"*! vanishing on I’, such that (I-kOA,U"* =(F+ KL 8)A,)US and it follows immediately from the positivity of — A,, see (10.41), that this problem has a unique solution. We now turn to the stability. Multiplication of (10.42) by u"*8=9u"*4.(1—0U" gives ,U*, U"*9)=40,|| U"||? +O —Hk||G,U"P. (10.44) Hence, if 024 we obtain at once 8,|| U"ll? <0, (10.45) which shows the stability estimate |U"| <|U°| =|o] for n>, for any choice of k and h. For @<4 we may use the equation (10.42) to obtain a 8,0" =|A,U"*8. 0. We have, by the definition of Ly, in (10.48), Velika Vette VE lT,Vyt! on OPSECTION 10 Mixed initial boundary value problem 129 We multiply by V"*! and sum to obtain ‘ IVP aky 18,0" IP VV" 4 Mg, VO) SAV MF +4 VP ER LVI Vt), or Gyette 2 og, One pttod HR |e 2yk GV? 1 SVP +2 (Ly VY, Vet). (10.52) Let now Ij be the set of xeQ? with neighbors in Pf and Q} =22\F}. Set WIV = ne? S Vax +h! x Voo?, cr} x0} Ly V(x) on Q}, ” vow ALyV(x) on Th, Lo for x€szf. Then (LanV SV YS WE Vy VL (10.53) Now we have easily, for Ve 9,(22), ‘ IVI 0, u=0 on 6Q, t>0, (10.57) u(x,0)=(x) in 2, where, for simplicity, @ is a convex plane domain with smooth boundary. We let h be a small positive number and impose a triangulation of the plane by the 10.1). Note that the triangles near 62 have st curvilinear. We assume that the modifications near 0Q are done in such a way that the angles of the triangulations are bounded below, independently of h. The triangulations are denoted by 7, Let now S, denote the continuous functions on Q which are linear in each of the triangles of 7, and which vanish outside Q,, and let {P}4* be the interior vertices of J,. A function ye 5S, may then be written as Nn WOD= Y PIG), (10.58) i where {9,}¥4 are the basis functions in S, defined by p(P,)=1, 9(P,)=0 for I#j. For the purpose of defining an approximate solution in S, of the initial boundary132 Cuaprer TL Fic. 10.1, value problem (10.57) we first write this problem in a weak form: We multiply the heat equation by a smooth function ¢ which vanishes on 09, integrate over Q, and apply Green’s formula to the second term to obtain for all such g, with (v, w) now denoting the standard inner product in L,(Q), (u,,9)+(Vu,Vo)=(£@) for t>0. We may now pose the approximate problem to find u,(z), belonging to S, for each t, such that (ye DtVu,VN=(6%) for xeS,, 120, (10.59) together with the initial condition u,(0)= v4, where », is some approximation of v in S,, such as the interpolant Nn = Y Pio; it Since we have only discretized in the space variables, this is referred to as a semidiscrete problem In terms of a basis {g,}* for S, our semidiscrete problem may he stated: Find the coefficients U,(a) in & wod= ¥ Uli ist such that z Uidlenvd~ ¥ UHV; Vod="hos | (10.60) Fi =i Recail from (10.58) that the U;(z) are then the values of the approximate solution atSEcTION 10 Mixed initial boundary value problem 133 the mesh points P), In matrix formulation (10.60) may be expressed as BU()+AU()=7() for £20, (10.61) with U(0)=o, given, where B=(b,) is the mass matrix with clements by =(~ p91), (ay) the stiffness matrix with ay=(Vo;,V@,), 7 =(J)), the vector with entries A=(F4;) and U(p) the vector of the unknown coefficients U;(0), j=1,....Ny. We may now discretize in the time variable, for instance by means of the 6-method, to obtain yee k with U°=», given, where U"**=9U"*! +(1—6)U", 0<0<1. Let us now note that for P; =(/,h, [zh) an interior mesh point which together with its immediate neighbors belongs to the original square mesh, then the equation in (10.60) corresponding to that point may be written YL BUi-7- AU =h Loos i and the corresponding time discrete equation Y BUI; AUT =h 74 oy), (10.62) 7 where A,~=d,,0,,+0,,5,, is the standard discrete Laplacian, 2, the forward difference quotient and j ranges over (0,0),(+1,0),(0,41),(+ 1, #1) with Bo.o= 3 Bi1.0=Bo,+0=Bir.z 0 Note that even if the scheme is implicit as B is not diagonal, and has to be inverted. However, it is possible to modify the scheme slightly to replace B by a diagonal matrix. This modification is accomplished by evaluating the inner product in the first term in (10.59) and thus in (10.60) by means of numerical quadrature as follows: We set (o,»),= YL Qow), where for each triangle t€ J, (with vertices {P,,}}-1)s 3 Q()=sarea LY /P.,)- ms The method is then (unas Dit (Vin VO=(L VLES 120. Since by our definition (Gp )y=9 for j#1, the matrix B in (10.61) is now replaced by a diagonal matrix. The procedure is referred to as the lumped mass method since it may be interpreted as having been