Lecture Notes in Mathematics Edited by A. Dold and B. Eckmann 749 V. Girault P.-A. Raviart Finite Element Approximation of the Navier-Stokes Equations & Springer-Verlag Berlin Heidelberg New York 1979Authors, Vivette Girault Pierre-Arnaud Raviart Analyse Numérique Tour 55-65, 5éme étage Université Pierre et Marie Curie 4, Place Jussieu F-75280 Paris Cedex 05 AMS Subject Classifications (1970): 35 Q10, 65-06, 65MXX, 65 M15, 65N30, 76-XX ISBN 3-540-09557-8 Springer-Verlag Berlin Heidelberg New York ISBN 0-387-09557-8 Springer-Verlag New York Heidelberg Berlin Library of Congress Catalogingin Publication Data Girault, Vivette, 1943- Finite element approximation of the Navier-Stokes equations. (Lecture notes in mathematics; 749) Bibliography: p. Includes index. 1. Viscous flow. 2, Navier-Stokes equations Numerical solutions. 3. Finite element method. I. Raviar P. A, 1989 joint author. Il. Title. Il Series: Lecture notes in mathematics (Berlin), 749, QA3,L28 no. 749 [A929] 510°8s (519.4) 79-21917 ISBN 0-387-09557-8 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically those of transiation, reprinting, re-use of illustrations, broadcasting, reproduction by photocopying machine ot similar means, and storage in data banks. Under § 54 of the German Copyright Law where copies are made for other than private use, a fee is payable to the publisher, the amount of the fee to be determined by agreement with the publisher. © by Springer-Verlag Berlin Heidelberg 1979 Printed in Germany Printing and binding: Beitz Offsetdruck, Hemsbach/Bergstr. 2141/3140-543210INTRODUCTION ‘The contents of this publication have been taught at the University Pierre & Marie Curie as a graduate course in mmerical analysis during the academic year 1977-78. In the last few years, many engineers and mathematicians have concentrated their efforts on the numerical solution of the Navier-Stokes equations by finite element methods. The purpose of this series of lectures is to provide a fairly comprehensive treatment of the most recent mathematical developments in that field. t is not intended to give an exhaustive treatment of all finite element methods available for solving the Navier-Stokes equations. But instead, it places a great emphasis on the finite element methods of mixed type which play a fundamental part nowadays in numerical hydrodynamics. Consequently, these lecture notes can also be viewed as an introduction to the mixed finite element theory. We have tried as much as possible to make this text self-contained. In this respect, we have recalled a number of theoretical results on the pure mathematical aspect of the Navier-Stokes problem and we have frequently referred to the recent book by R. Temam [44]. The reader will find in chis reference further mathematical material. Besides R. Temam, the authors are gratefully indebted to M. Crouzeix for many helpful discussions and for providing original proofs of a number of theorems.CONTENTS CHAPTER I, MATHEMATICAL FOUNDATION OF THE STOKES PROBLEM $1, GENERALITTES ON SOME ELLTPPIC BOUNDARY VALUE PROBLEMS... 1.1, Basic concepts on Sobolev spaces ++. 1.2. Abstract elliptic theoxy . Gon1908 5 1.3. Example | : Dirichlet's harmonic problem . 6 1.4, Example 2 : Neumann's harmonic problem sssssseee 8 1.5. Example 3 : Dirichlet's biharmonic problem s.....++ n § 2. SOME FUNCTION SPACES ..ssseee beeen ween eens 12 Qi. The space H(div 5 Q) seseeseeeeee fore 2 13 2.2. The space H(curl 3 2) sesesseee seneearecenes 19 § 3. A DECOMPOSITION OF VECTOR FIELDS . 22 3.1. Existence of the stream function of a divergence-free vector field .. epep0500 00400008 22 3.2, A decomposition of [1?(9)]™ ..eseeee 29 3.3, A decomposition of [HE(A)]™ ss. Sogedeeqsee05cGc0 + 32 § 4. ANALYSIS OF AN ABSTRACT VARIATIONAL PROBLEM ......++ sence 39 4.1, Statement and solution of the problem sesesesesseeeesseeeeeeees 39 4.2, A saddlepoint approach eseseseeeeeeeees 43 4.3, Numerical solution by regularization ssseesesseseesesseeeeeens 45 4.4. Numerical solution by duality «+++ ar § 5. THEORY OF THE STOKES PROBLEM ...+.e+0e0+ 5.1. The " velocity-pressure " formulation . + 50 5.2. The " stream function " formulation « 55w CHAPTER II. NUMERICAL SOLUTION OF THE STOKES PROBLEM, A CLASSICAL METHOD .. 58 $1, AN ABSTRACT APPROXIMATION RESULT . § 2. A FIRST METHOD FOR SOLVING THE STOKES PROBLEM 2.1, The general approximation .++-+++++ 2.2. Example 1: a first-order approximation 2.3. Example 2 : a second-order approximation . 2.4, Numerical solution by regularization .. CHAPTER III. A MIXED FINITE ELEMENT METHOD FOR SOLVING THE STOKES PROBLEM . 86 § 1, MIXED APPROXIMATION OF AN ABSTRACT PROBLEM . 4A mixed variational problem «+. ce 86 1.2, Abstract mixed approximtion ... § 2, APPLICATION 10 THE HOMOGENEOUS STOKES PROBLEM ...... oo 2.1, A mixed formulation of Stokes equations seessseeeeeeeeeeeereres 93 2.2. A mixed method for Stokes problem --+esseseeeeeeeesaeees 98 2,3. Application to finite elements of degree % . g00c0G 101 CHAPTER IV, THE STATIONARY NAVIER-STOKES EQUATIONS + 108 § 1. A CIASS OF NON-LINEAR PROBLEMS 0+. ++ see0ee0 + 108 § 2. APPLICATION TO THE NAVIER-STOKES EQUATIONS . 110 2.1, Some results of functional analysis ..... no 2.2, Solutions of the Navier-Stokes problem « 3 § 8. A FIRS? METHOD FOR APPROXINATING THE NAVIER-STOKES EQUATIONS 121 3.1, The uniqueness case « 121 3.2. The non~uniqueness case «.. 2) § 4. A MIXED METHOD FOR APPROXIMATING THE NAVI +135 4.1, mixed formulation .seeeeeeeees 135 4.2, bn abstract mixed approximation ppeccoooon 138 4.3, Applications «+++ aepcpoceoocco0 43vu CHAPTER V. THE TIME-DEPENDENT NAVIER-STOKES EQUATIONS .. § 1. THE CONTINUOUS PROBLEM ... t Some vector-valued function spaces «. 1.2, Formulation of the Navier-Stokes problem « 1.3. Existence and uniqueness of the solution +++++++++ § 2. NUMERICAL SOLUTION BY SEMI-DISCRETIZATION : A ONE-STEP METHOD ... § 3. SEMI-DISCRETIZATION WITH A MULTISTEP METHOD ..... 3.1. Generalities about muleistep methods s..++ 3.2. Multistep methods for solving the Navier-Stokes problem «+.+-+ 3.3. Convergence of a family of two-step methods « BIBLIOGRAPHICAL NOTES «+++ REFERENCES .... INDEX «0.45 148 148 48. 156 161 170 179 179 182 185 192 195 200CHAPTER I MATHEMATICAL FOUNDATION OF THE STOKES — PROBLEM § 1 - GENERALITIES ON SOME ELLIPTIC BOUNDARY VALUE PROBLEMS In this paragraph we study briefly the Dirichlet's and Neumann's problems for the harmonic and biharmonic operators. 1.1. Basic concepts on Sobolev spaces Our purpose here is to recall the main notions and results, concerning the classical Sobolev spaces, which we shall use later on. Most results are stated without proof. The reader will find more details in the references listed at the end of this text. To simplify the discussion, we shall work from now on with real-valued func- tions, but of course every result stated here will carry on to complex-valued functions. Let 9 denote an open subset of R" with boundary [. We define D(0) to be the linear space of functions infinitely differentiable and with compact support on 9. Then, we set DH = fy see De) , or equivalently, if © denotes any open subset of &° such that HCO, DO) = tg 3 ee DOD. Now, let <2() denote the dual space of -D(2), often called the space of distri butions on ®. We denote by <.,.> the duality between D'(n) and vee Din; . tty se if ue CNG then atu = A 5 % = ax, sae ox, ‘a For mEN and pER with 1. When p= 2, W™(n) is usually denoted by Ha), and if there is no ambi- guity, we drop the subscript p= 2 when refering to its norm and seminorm. H"(n) ig a Hilbert space for the scalar product : a3) (av) - ) { u(x) av (ade « n,2 iC lols m Jn particular, we write the scalar product of 12(n) with no subseript at all. As Din) ¢ H™(0) , we define (4) way = FH" , ise. UO(A) is the closure of (9) for the norm Il. Il, a * We denote by W(o) the dual space of HT(9) normed by + (sy tet g = su, [stl p8 venta) vi, v40 9The following lemma characterizes the functionals of H(A) + e141. A distribution f£ belongs to (9) if and only if there exist functions in W2(a), for al ty kl is anorm on uw!/2(ry, and w!/2(r) is a Hilbert 12,7 - 2 space for this norm. tet i !/2(r) be the corresponding dual space of u!/*(r), normed by a" yw a9) sup , 12,0 TT ena 1/240 w #0 where again <.,. > denotes the duality between w'/¢r) and u!/?(r) . We remark that <.,.> {8 an extension of the scalar product of 12(r) in the sense that when yp" € 12(r) , we can identify with (oul) do tr Let Us (vy, ses) ¥,) de the unit outward normal to T which exists almost everywhere on [ thanks to the hypothesis of Lipschitz continuity. If v is a function in ¥?(Q), we define its normal derivative by (1-10) & 1 Xe can be proved chat the mapping v= 26 cua) 5 wl/2¢r)). Woreover, we can characterize H(g) as follows : ‘THEOREM 1.4. F(a) = 2(9) 5 na 2s 0} HC) = tv EHC) § yv=O and ZOO. When T is sufficiently smooch, che range space of y, can also be extended as follows. For mEN,m>1, we define H™'/?(r) as the image of u™%0) by the transformation y , equipped with the norm : vf intel ve Hmay me yet m1/2,0 ~yr3/2 Then, it can be checked that 22 € (2) for u in 0), and the following result holds : ‘THEOREM 1.5. The mapping wry x 2, defined on ¥°(9) is onto We close this section with two useful applications of the Green's formula. LE, 1.2. 1) Let u and ven(n) . then, for 1 denote the duality between V' and V. Let (u,v) + a(u,v) be a real bilinear formon VV, 2% an element of V" and consider the following problem : Find u€V_ such that (Py a(uyv) =< ave VvEV.The following theorem is due to Lax and Milgram [35] + ‘THEOREM 1.6. We assume that a is continuous and elliptic on V, i.e. there exist two constants M and a> 0 such that \ Ten , (1.14) a(u,v) ~- a(usv) belongs to wl(n) . Therefore, thanks to the Lax-Milgram theorem, problem (D') has one and only one solution u in l(a) . Tt remains only to prove that u may be characterized as the unique solution of problem (0)+ Taking ve D(a) in (1.20) gives : ayy) #-s vve Da). Hence u satisfies (ai) wu, € Bay, @) 2 (ain = bust in y- Conversely, every solution of (D,) is a solution of (D') by the density of Deny in wha) . Bue uno emle itt yun. therefore problems (D,) and (D) ate the same. As far as the regularity of u is concerned, we know, from the Lax-Milgram's theorem that the mapping & ++ u~u, is an isomorphism fron ¥'(9) onto ul(a) . Therefore, dunu, < : clearly, < tfty g We: Hence Huts g % %y MEO eu yg Vu, €u'(a) such that y,u, = @ + From definition (1.8) this implies that lu < oc, Wen. 10 nat tel) Thus, we have proved the following proposition +PROPOSITION 1.1. Problem (D) has one and only one solution u in u'(o) and a2y Duty g S$ CMEN; 9 + MEM pd » ies u depends continuously upon the data of (D)- 1°) Let mEN,m>1. When Pf is sufficiently smooth, it can be shown that if f£eW™? (0) and geu™'/?(r), then ue HA) and (1.22) Det © OEM ag Hela tyr 2°) Wnen Tis only Lipschitz continuous, the same result is still valid for m= 2, provided @ is convex. ™ 1.4, Example 2+ Neumann's hamnic problem Here, we assume in addition that is connected and we deal with the non- homogeneous Neumann's problem + Find u such that : (3) tut i au ad 124) = or, where £ 12(9) and g€W'/2(r) satisfy the relation + (1.25) I faxt+,* o r Since problem (N) only involves the derivatives of u, it is clear that its solution is never unique. We turn the difficulty by seeking u in the quotient space u'(n)/R equipped with the quotient norm (1-26) 1 inf Ivy - Mate * et, 10 The theorem below states an important property of this space ; its proof can be found in Netas [39] .THEOREM 1.7 The space won /e is a Hilbert space for the quotient norm (1.26). Moreover, on this space the seminorm ¢ ++ | is a norm, equivalent to (1.26). 1a With this space, we can put problem (N) in the abstract setting of problem (®). et ve nlye, a(i,¥) = (grad u,grad v) yuei,vet and (1.27) eo tv exec eve vved. a z Note that the right-hand side of (1-27) is independent of the particular v€ > thanks to the compatibility condition (1.25). Furthermore, #€ V' because, owing to (1-8), we have : +g typ) inf Ivy, 12 ee a if, fv de + < aw >,| < EL, ier Thus (1.28) Wty, S REM, gt lelyy p Obviously, a(a,#) is continuous on Vx V , and by virtue of Theorem 1.7, vei? 9 #1 (a) /R Hence, by the Lax-Milgram's theorem, the following problem Find & in w'(a)/R satisfying a i 1.29) ase yeen(ae , has a unique solution & © H'(a)/R . Let us interpret problem (N'). When v is restricted to D(a), (1.29) yields + (1430) caus £ in L(t) yued . Next, by taking the scalar product of (1-30) with v and comparing with (1.29), we find : aap (Fad u, Frad v) = (-,v) + < gv > vveHl(a) . :10 Therefore, problem (N') is equivalent to find u in HI() satisfying (1.30) and (1.31). It remains to interpret (1-31) as a boundary condition. At the present stage this cannot be done without assuming that u € H°(q) . Then Green's formila (1.13) yields + 2) ao = : veg! [By eeecee ven, ive. bu _ As u is supposed to belong to H7(n), this implies in particular that ge nl/?(r), In that case, problens (N) and (N') are equivalent. Of course this is not entirely satisfactory in that the existence of a solution of problem (N) is subjected to the regularity of the solution of (N‘). Further on, with more powerful tools, we shall be able to eliminate this regularity hypothesis. Now, let us examine the regularity of u. According to the Lax-Milgram' s theorem, (1.28) and the equivalence Theorem 1.7, we obtain : u < + lel Celso tle ed -1/2,P 1,9 We have thus proved the following result. PROPOSITION Let the solution & of problem (N') belong to E°(9)/R. Then & is the only solution of problem (N) and each u € 4 is continuous with respect to the data, ire. 1.32) luli < CUED, p+ ely « Remark 1.2. As in the previous example, if T is very smooth and if £© H™°(n) and Haeea(t) BE with m> 2, then ic can be shown that & © WY 0)/R and 1.33) lulaa SCUEM 4 gt MRM a/9,7) for every veh.u 1.5. Example 3 + Dirichlet's biharmenic problem Consider the non-homogeneous fourth order problem : For £ given in #7 (a), g, piven in w/?(r) and gy in w/2¢r), (find u such thae : (1.34) Musf ino, Ze weg, on T (B) ‘ and au (1.36) Bre or. \ The function space natoratly attached to this probten 1s H2(0) and the a(u,v) = (Au,Av) « This form is elliptic on 12(a) because the mapping v m= HOY I, 9 ie 4 notm a on H2(R) equivatent to the norm L. lly 9+ Indeed, for v in D(A), we can 2.0 easily show by integrating by parts and interchanging derivatives that 2 (1.37) Wav Woo By density, the same result holds for the functions of H2(0). The equivalence follows from Theorem 1.1. According to Theorem 1.5, if T is smooth enough, there exists a function u, in W(n) such that 38) a a coe ‘Thus we turn to the following problem : in Find in H°(A) such that 1 2 (BD) G39) uu, © Ht) 2 \ c.40) a(uruy) =< fv > -alusv) Vv Ee HEC) « By the Lax-Milgran's theorem, problem (B') has exactly one solution w in H(A) Owing to (1.38) and (1.39), u satisfies the boundary conditions yout 2 on Cee et 2 aye .12 Besides that, by restricting the test functions of (1-40) to D(0), we Find ust in WC) Therefore, u is a solution of (B). Conversely, as in the case of the harmonic operator, we can show that problem (B) has at most one solution in u°(@) « From (1-40) and the equivalence of norms, we derive the bound Hut, gS CWE, gt tugly g) Yu, satisfying (1.38) , 2,9 ieee Waly gS Cy EM yy + MBI ayo 7 + HB yep) « These results are summed up in the proposition below : PROPOSITION 2.3. If T ig sufficiently smooth, problem (B) has exactly one solution u in won), bounded as follows : aa) Wolly, SCWe ML 248 * Ns, "sya,r 7 Wey “2,P Remark 1.3 When T is sufficiently differentiable it can be shom that, if £ © H™*(a), ey? 3/2 8, (7) and gy W727) for me ZEN, then UE HO) and Hata SS WE gg t M8, Nea yoyp tl Babee 3 /2,P § 2 - SOME FUNCTION sPacES Throughout this paragraph, we assume that 9 is an open subset of R” with a bounded and Lipschitz continuous boundary [ . We introduce here special Hilbert spaces that are particularly well suited to incompressible flows and other problems arising in mechanics. Several results are stated without proof ; these can be found in the book by Duvaut and Lions [22 ]+2.1, The space H(div 3 9) From now on, we shall often deal with vector-valued functions. We shat? distinguish vectors by means of arrows and extend naturally all the previous norms to vectors as follows : if ¥ = (v. + ¥,) then 3 Se Ue Woy =(2 0 Hap? . Poi divv= 7 Then, we introduce the following spaces + uaiv sa) = GE Gy)” s div Fev} , normed by . roe 2 Qa) WF tycaiy gy 7 WF AS g tHaiv Fs gH? and uj(aiv sa) = Da) MED Clearly, H(div ; 9) is a Wilbert space for the norm (2-1). [D@1" is dense in W(divs 9) « PROOF. Let U€ H(div ; 9). First, let us show that there exists a sequence of functions of H(div ; 9) with compact support, that tends to U. Let a denote any positive real number and let y, be a function of Die”) such that + 1 for |x| | SiS plresp.f,) denote the component of O-% dounded by F,(resp. f, and 96). Lee Vx {classes ¥ 5¥jo © H(O;)/8) and consider the following problem : Find #€©V satisfying r { _ arad iegrad ¢ dx = < 2,0 > yoev , 0.8 where @ ; meas@,) e 2.2) = 0. Finally, Le avel se, Wty 5a) lvl ge Yvev. since we can assume that ™ is continuous. Hence £€ V' . As a consequence, problem (N) has a unique solution w€V that satisfies: (sia re a Tas { nojaivd ao. o0 © era © Therefore the restriction to F of é + u isa function of (D(@))" that converges toward U for the normof (div; 2). = Remark 2.1. We shall see later on that, in this theorem, U is extended by matching its normal component with that of its extension. This process will be used several times. . Again, let ¥ denote the unit exterior normal along [ . The next theorem concerns the boundary values of functions of H(div ; 2). ‘THEOREM 2.2. The mapping yt ¥ + WU p Sefined on (D(@))" can_be extended by continuity to a linear and continuous mapping, still denoted b u(div 5.9) into w/2¢ry, » from Let v € DG) and ve (DG))". The following Green's formta holds : (2.4) Gerad 9) + (div ¥ye) = { ood do. r As D(H is dense in ul (a), (2.4) is still valid for g in u'(o) and in (D(H) . therefore 2.) if oF3 aol < UF acaiy y gy We Bye YR EHD» v TeDE" - 3 : Now, let u be any elenent of u!/(r), Then there exists an element y of7 u(9) such that y,e =u + Hence (2.5) implies that If y ¥ed ao] <0 Iu vuewoy,vte De). : ucaiv 50)" "ayer Thus VSM SWF lyase 5 0) + Hence, the neat mapping y,¥ r+ Gh), defined on (DG) ts continuous for the norm of H(div 5 9). Since ( My Mecuaiv 5 m3 HF ¢r) 7+PROOF. tet u’en/2(p) and consider the problem + ; Find g in H'(o) such that 2.9) -aetes0 in a, 2.10) Be on re Unlike the Neumann's problem of § 1.4, this problem has exactly one solution in ul(q). We denote it by p and set ¥= grad . Then VE H(div; 9) and yo #o" . Moreover le i wetyy op, v Ue Wcurt 3 9) , r vee na We now turn to the case % = 5. When ¥€ (D'(0))°, we define its curl by av, 9vy bv avy av ga ge (3 eee (2.15) arg (2 eee eee), & fag? Tay Fa’ Fa, HH) Wich this operator, we introduce the following space + u(edet 5 9) = WE C2 (a)? 5 cder VE CL? (my)73 normed by * 2 ee iia 2.16) ¥ > = uy + eden ¥ 7 (2.16) WS Wacaier gay 7 MENG g t Hebel Fm gd ‘THEOREM 2.4. The space (D())? is dense in Medel 3 9) This theorem is proved in Duvaut & Lions [22 ] -a x3 denote the vector product of J and ¥ . The mapping |r defined on (D(i))? can be extended by continuity to a Linear continuous mapping still written VY ++ ¥x ¥), from H(curl 3 2) onto Ir arte), PROOF. The proof is similar to that of Theorem 2.2. For all $ and % in (D(H)? , the foltowing Green's formta is valid : - Gein + Geter 3) = (Grd F ae. r By density, this equality also holds for alt @ in (u'(n))>. Therefore if eye * . * 2 y te clea} I, GeYS dol SUF Wa cter 5 al yg YE DM .¥ ee CHAD) Now, for each jin (ii!/(ry)? there exists a 3 in ('(a))? such that 3 Hence < UW lceder 5 ay FU y jor + Therefore, se * . =)3 (2.17) W xvii © Mladen gay Y YE (DO). This permits us to extend the mapping ¥ ++ ¥x¥), by continuity to a mapping r from H(ert 39) into (W'/? (ry)? | satisfying (2.17) for % in H(cdrl ; 9). Tn addition, Green's formuta becomes + (2.18) -@,cdrt ¥) + Gycdrt 8) =< Vx Saye op ¥ Ye mcurt 3 9), 9 € clay). Let us introduce the following subspace of W(cliet $9) H(curl 5 9) Hy (curt 3 9) = (D(a))3 . ‘THEOREM 2.6. uted 5) = Pender Dy Ha Ved Remark 2.3. Tg VE H,(cérl ; 9) , its tangential components vanish on I . .22 § 3 - A DBOOMPOSITION OF VECTOR FIELDS In this paragraph, ve shall prove that every vector of (17(s))" is che sun of a divergence-free vector and a gradient vector. This will lead to an interesting decomposition of (17(9))" and (ui(a))” as a direct sum of orthogonal spaces. Me shall make the following assumptions on 9:9 is bounded, eventually multiply-connected, and its boundary T is Lipschite continuous. Like in § 2, ve shall denote by T, the exterior boundary of and by T, 51 0 for Ot, the set C-H is not connected and again we call that component which is bounded by T, , for 1 == gad y ax yee De’) , that is [ > caiviw > =-[ Safco ar- J [ datdey ax. a iso Ja, As ¥ and @ are both divergence-free, Green's formula (2.7) and (3.4) yield . P (3.5) < div We > = Gy Yere op H oo <1 Mere on and therefore inside 9 « In the sum, the normal to T, is directed outside lence each etm of this sum cancels 4 term of < y,uryyp Pp « Therefore a 2. oo 7s ed 7 Therefore a divs -0. ©) Let us introduce the Fourier transform of v a ~2iw (x, 6) . . oe (3.6) yi) 7 ic e vj dx, j = 1,2, where (x,€) = x8) Eo + Note that each 3, (4) is a holonorphic function of the complex variables §, and z__since the support of 3, is compact. : Ja tems of Foutier transforms, the condition div ¥= 0 becones (3.7) 5, + 55%, =0 and vs cirly holds if and only if (3-8) ¥ = inte » U, = - 2ing, Now, if we take 3 = ¥, then, thanks to (3.7) both equalities of (3-8) are valid, ive, G= gh = . Therefore, the inverse transform of 9 2 . 2 Lee? is the required stream function of ¥, provided ¢€17(R,) . As ¥;€ LORD ,25 it suffices to show that @ is bounded in a neighborhood of the origin. According to (3.7), we have FO 0. Hence, using the holomorphy of the function %, , we obtain . av. qi - 1 (eo) + ofl. KE) = FE, GO #0 I » so that 3 =i + 5 PEE) = Bey BE, Ee # OUIEQD Clearly, this inplies that @ is bounded in a neighborhood of zero. # Remark 3:1 If is connected, then clearly the stream function y of V is umique up to an additive constant. = * Suppose again that @ is connected (otherwise, we deal with each of its components separately) and consider the space He (7 H(div 5 9) 3 div ¥= 0}. Then every stream function y of ¥ satisfies O, thatis + we alr, opp, * Aconstant o, 5 for O 0- Then, Isl, q and Wl are two equivalent norms on the space (ve H(A) vip = 0}Then, the stream function of $ that vanishes on T, is also the only solution of the problem : Find g in @ such that (3.9) (etrt gcirt = Gcirl yy VX Ee Let us interpret (3.9). If we restrict x to D(2) , we get : ap = curv in Dm. Next, by applying formally (2.8) to y , we obtain (ort vyctet = - ed +< BZ oyxop xe ee Likewise, (2.14) yields formally Gyokiet O = (curt Fy) - < Myx op + HER. Finally, by comparing these two equalities with (3.9) , we get t 0 veer. These results are summarized in the following corollary. OPOUIARY 3.1. Let © be Like in Lema 3.1. For TEM, the relation : 7a establishes a one-tovone correspondence betueen che spaces Hand . Furthermore, the stream function » can be characterized as the solution of problem (3.2). This problem has the following interpretation - de =curl¥ in DQ), fe iep where the constants or O, = 0. For O is divergence-free and has a compact support. Again, let %, be the Fourier transform of v, 5.) = { ew ee 7 i Then %, is holomorphic in R2 since the support of %, is compact. The conditior i é div ¥ = 0 becomes + (3.12) 8,70 such that curl gs, Now, we must find a function @ in [ ise. such that t9, = Bim(EPy - EP) » G13) = 2in(E, — E)8q) » 9, = 2im(E,8, - £9)) « The third equation of (3:13) is @ consequence of the first two and (3.12). Therefore, in order to fix @ , we add to (3.13) the condition 3 G14) I g@,c0- The unique sotution of (3.12), (3-13) and (3.14) is # 1 5 " BETET™ (Ea%2 7 b%s) > 1 Gad %o " Bndede (61%3 7 89) and $4 = ste (69, - 8,5)? se zine ceo) gt? 3 where Ici’ = J g2 . By inspection, these equations imply that oo ist 3,1 , provided that 2 is bounded at the origin. First, we observe that (3.12) impties that @,(0) = 0. Then, as is holomorphic, it follows that 9,(e) = (0) + of?) in a neighborhood of 0. Hence # is bounded as € tends to zero. By restricting to 9 the inverse transform 3 of $ , we thus find a function ¢ in (u'(9))3 such that cirt $= and moreover, dive =0, which is slightly stronger than the statenent of the theorem. # Remarks3.2. 1) The divergence-free stream function 3 of v is unique, but this does not hold for ¥ , since the extension of ¥ is not unique. 2) From the identity in (D"(9))? (3.15) Curl (cUrt ¢) = - a9 + grad dive , we see that 9 satisfies29 (3.16) w= carry. The choice of the boundary conditions that must be added to (3.16) in order to characterize @ can be found in Bernardi [ 8 ]. 3) In the above proof, we have chosen a divergence-free stream function, but of course this is not the only possibility. For instance, we might have replaced (3.14) by Guay oO. e Then (3-12), (3413) and (3.14!) have the only solution . 2 . 1 . Seg eo g. a,-0. 17 ie, 2°” Bin, 3 The corresponding stream function ¢ of ¥ has the form (¢,¥s0)+ . 3.2. A decomposition of [1.7¢0)1" ‘THEOREM 3.3 For each function % of [1°(a)] , there exists a function q € u!(n) anda function p €H'(a) if n= 2 (respectively, 9 € (H(M)? ie n= 3), such that ( corte if n=2 + ee ding ie ones, (3.17) yy@- ala <0. Moreover, the functions cirl y(resp.cuirl ¢) and grad q are orthogonal in (17(9))". PROOF. Let ¥€(17(m]” and consider the following problen + fing q in u'(a) satisfying G18) (eFad q sad uw) = Gysrad yy) Vu enim. This Neumann's problem has a unique solution 4 in H'(9)/R. Any q€ 4 verifies dq div ¥ in wl) Hence ¥- gtddq is a divergence-free vector of H(div 3 9). Then, by applying30 Green's formula (2.7) to (3-18), we get O = (Hgrdd q.arad u) = < Greraa q)- Tu >) vuenta . — 7 This implies that G-grad q)-%),= 0 in w'/?(r), Therefore, we can apply to V- grad q the conclusion of Theorem 3.1 (or 3.2) and we find the required function y(or ¢). Note that this particular y (or ¢) satisfies (3.19) yy(ctrl #) = 0. To show the orthogonality of grad r and curl # , we use Green's formula and (3.19) (grad x, cart 3) = - (xdiv (cdet B)) + < y (cael Beyer >p so vrent. # Remarks3.3. 1) The decomposition of Theorem 3.3 determines a unique q , up to an additive constant. The function q is characterized as the solution of (3.18) that is , if ¥€ H(div ; 9) , then dq=div¥ in a and 28a FS on ee ae tf J is onty in [1°(A)]" , the last equality is format. 2) Wuen n= 2, the above decomposition and the condition yj, = 0 determine y uniquely. We can characterize y as being the only function of @ that satisfies (carl esctrl 9) = @ ~ grad aycirl vxed Hence, formally is the solution of is constant for 1, = 0 , problem (3.23) has a unique solution r b in w(a)/R. Moreover, since $0 € u'/2cr) then p € HP (A)/R , provided that T ie sufftetently smooth (cf. remark 1.2 n°2). Then, according to part 1, there exists a function u, in allay)? such that div, = 0 and 4, ~B- y,(erdd p) on Te Therefore u= ay + grad p is the required function. Remark 3.4. Jet us assume that I is infinitely differentiable and that belongs co G2 (ry), for maa, with f $3 do = 0. Since Theorem 1.5 and r remark 1.2 are both valid for any m , we can apply the above reasoning to show that there exists u in (HT(n))" such chat divieo and ySe%- 0 * Now, we define the following spaces : va Fe ain); dive =o, wea) = {© 1 (9) § (qo) = OF and we denote by Vi the orthogonal complenent of V in (1i!(9))” for the scalar product (grad U, grad ¥). IEMA 3.2. The divergence operator is an isomorphism fron Vi onto 12(n) « Proof. tet Fe calcay)” . By Green's forma :34 { aiv Fax = [Bi erco. Thus div € £((u}(a))" 5 12(a)). Let us show that div is @ one-to-one mapping fron tonto 12a). Since ¥ = Ker(ain), fe suffices to show ehat div naps 2 7 ‘ Hy 3 (a}qa)® onto 12(9). For this, let q be a function of 13(9) i we seek in aay) such that div ¥=q. As { is bounded, there exists some function @ in wa) such that seq in ne We set ¥, = grad a © (u'(n))" . Then div 3, = 08-4 5 moreover, by Green's formula [ BS ao-[ ave, ane{ aeceo. r a a atso y,%, € at!/(1)%, Therefore, we can apply Theorem 3.5: there exists W, in (HI(a))" such that div ®, = 0 and yQ6, = 7,¥, + Then $= %,- iG, is the required function since He (HL(A))" and div¥=q- Finally, it follows from the open mapping Theorem (cf. Yosida [46 ]) that the inverse of div is continuous from 12(n) onto Vi , therefore div is an isomorphism. = Renark 3. The usefulness of 12(9) arises not only from this lemma, but also from the fact that ¥ q © 12(a): (3.24) Waly, inf lqte I, gt hal . wn” ER 20 1? (a) /8 where 4 is the class of 17(n)/R containing q . Asa consequence, it is often very handy to work with 12(a) instead of L7(m/R. # THEOREM 3.6, Tet % betong to (H'(m))” and satisfy (3.25) -s0 vvev -35 Then, there exists exactly one function ¢ in 17(0) such that (26 «Bt [ pavdaxs- ve aay”. a PROOF. Consider the following problem Find & in Vi satisfying iv > veev. (3.27) (div t,div ¥) As div is an isomorphism from V¥ onto 12(n), it follows that 2 fe ic oe wi. > lz Vaiv 542, > all Hence, by the Lax-Milgram's Theorem, (3-27) has a unique solution % in v . Then, hypothesis (3.25) implies that W also satisfies (iv Laiv =< D> vee akan. We set v= dive 12(n) and we find (3.26). It remains to prove that y is unique in 12(A). But clearly, if v © L2(@) and (esdiv ) = 0 ¥ FE CHCA), then = 0 since div maps (Hi(A))™ onto ey. the next theorem states another apptteation of Lema 3.2, ‘THEOREM 3.7 There exists a constant > 0 such that (3.28) sup EHS iyi ve em. Fe alin? aya PROOF. fo pow eee oe Hew such that ve div% and |¥|). + (fda Larde D = - Gav) vee ala). Therefore, ~< bo > e< gagd> vee ala. Hence % satisfies Ge cil(a)" and - sum sfidq in o, z “los t= cay! grea Conversely, it is clear that (a) 'gragq¢v' for every a 2 a. Hinally, the proof of the density of Yin Vv is similar to that of the density of Ui in we. ®37 Pomark Since V is a closed subspace of hay)? > we have ala =vev - Hence, Theorem 3.8 implies that every function of (HI(A))" can be written in the form (3.31) %= ca lerie q+ etd. The function q is uniquely determined in L2(0) by (3.31). # When n= 2, we can characterize entirely the stream function » in (3.31) that satisfies yj, = 0+ We introduce the following closed subspace of wn) + 0 O= (EW) 547, 505 xp fs constant, 1 Melo * llag > % Ielaye 19 > ovis g- Now, each function % of [H}(9]? has exactly one stream function » in @, and (3.31) implies that (3-32) curl $= - wy + curt(-a)"! grad g let x€@ and let cirl xy , GEV. From (3.32), we get + (-te,curl @) = (curl ¥,curl %) ~ (curl(-4) rad qycurl &) . But (curl (~a)"' gFda q,curt &) < (7a) lgrdé qyetrl curl & > : a ey and : ctrl curl G =-t0 + grad div-- ao, since %€V. Therefore (ourt(~a)"! grad qycurl &) = < (-a)"larad ay 8B > =< grad qi 0 since WV. Hence, we have :38 (3.33) (ip .8x) = - Ceurl ¥,4y) vxee. As tae Il, is a normon @ , problem (3.33) has a unique solution y in © and therefore this problem characterizes o « Let us interpret problem (3.33) in terms of a boundary value problem By restricting x to (8) , ve obtain : (3.34) By =~ aCeurl ) in wm). Then, by taking the scalar product of both sides of (3.34) with x in @, integrating by parts, and comparing with (3.33), we find (formally) ee. p vxee As x €@, this implies that, formally: { Le + curt Dao = 0 for 1 Si a Syloe 1¥do= 0 for 1. We introduce two continuous bilinear forms + AG) ERK RR, DG) EK EMR with norms Nall sup AY = cup b(vyy) ave x fully ivi vexyeN #0, v40 a v40u20 Uv ty Me ty Consider the following variational problem : 6 For £ given in X' and yx in M, find a pair (u,A) in XM such that + @ (41) a(uy) + bw, =e YVEX (4.2) diay) =< yur YREM. In order to study problem (Q), we require some extra notations. We associate with forms a and b two continuous, linear operators : AGL(X X') and BEL(XEM') defined by + (4.3) = atuy) Vu, vex, (4.4) < Byyy > = b(vyu) vvex , veM40 Let BY €£(M5 X') be the dual operator of B, ives (45) =< By > = div) YU EM , ¥vVEX . Te can be readily verified that (4.6) =llall WA lg cy 5 xy Ib. » UB Mey swt With these operators, we have an equivalent formulation of problem (Q) ¢ Find (uA) ©X«M satisfying + @) AutBA ea in x, Bu ex in Mw Next, we set V= Ker 8 in X and more generally, for each x €M' , we define the affine variety : VQ) = {vEXs Bye Xd. Equivalently, we can write that (4.2) Vd = EX DO) =< yur YEH , v= vO) - Moreover, the continuity of B implies that V is a closed subspace of X. Now, with (Q) we associate the following problem : Find u in V(x) such that >) (4.8) atv) =<2v> YveEV. Clearly, if (u,2) ©X*xM is a solution of (Q), then vu V(x) and u is a solution of (4.8), i.e. u is a solution of (P). The rest of this section is devoted to show the converse of this statement and the existence and uniqueness of the solution, under suitable assumptions. For this, we define the polar set V° of Voby + vr(eex' 5 0 such that4 4.9) co fon bw) ag vem vex Iviytvly (it) the operator B' is an isomorphism from M onto V° and. (4,10) IBtullys > ellully Wem 5 (sit) the operator B is an isomorphism from V! onto M' and a) Ipviyys > Biiviy vvex. 1) Let us show that (i) = (1i)+ By (4.5), statement (i) is equivalent to < Bu > sul aa Sell ' vex Wt 7 ve x that is, (4.9) is equivalent to (4.10). It remains to prove that 8’ is an isomorphism. Clearly, (4.10) implies that B! isa one-to-one operator from M onto its range A(B'). Moreover, it also implies that the inverse of B' is continuous. Hence 8' is an isomorphism from M onto @(B'). Thus, we are led to prove that aca") For this, we remark that &(B') is a closed subspace of x' , since B' is an isomorphism. Therefore, we can apply the closed range theorem of Banach (ef. Yosida [46 ] p-205) which says that &(B") = (Ker(B))° = VO. This proves part a°1. » (= Gin. First, we observe that V° can be identified isometrically with (v')'. Indeed, for vex tet vb denote the orthogonal projection of v on WV. Then, with each g€ (V')" we associate the element g of Xx’ defined by - cats vvex. Obviously @€V° and tt is easy to check that the correspondence g ++ &42 maps isometrically (V1)! onto V°. This permits to identity (W)' and Vv. As a consequence, statements (ii) and (ifi) are equivalent. = The condition (4.9) is usually called an" inf-sup " condition. ‘THEOREM 4.1. Let us nak “hypotheses : (5) There exists @ constant a> 0 such that (4.12) avy) > avi vveV. (i) The bilinear form > satisfies the inf-sup condition (4.9). Then problem (P) has a unique solution u in WOO unique A in M such that the pair (u,1) is the unique solution of problem (Q)- Moreover, the mapping (2.x) ++ (u,A) is an isomorphism from X' x M' onto xu PRO! From (4.9) and Lemma 4.1, ve see that there exists a unique element uv, in Vo such that 1 Buys x and Musty < flxly Therefore the following problem is equivalent to problem (P) Find wsu-u, in V_ satisfying e avy) = < fv > > a(uysv) vvev. Since a is V-elliptic, we can apply the Lax-Milgram Theorem to (P"). Thus problem (P) has a unique solution u in V(x) and Woy S CNL Ags + Mix ys) + where the constant ¢, depends only upon a, 8 and llall « Now, 2 - Au belongs to V° ; therefore, according to Lemma 4.1, there exists one and only one X in M such that BM 2 by 1 and Wr Wye SF Werdullyy < CMe ty, + Mx ty) «43 Hence (u,A) is the only solution of problem (Q)+ The mapping (24x) + (uA) is obviously an isomorphism from x' x mM! onto XxM. 0 ® 1) Under the hypotheses of Theorem 4.1, problems(P) and (Q) are equivalent. 2) If a is elliptic, chen the inf - sup condition (4.9) is necessary as well as sufficient for the mapping (us) + (24x) to be an isomorphism from Xx M onto x" x MW . Indeed, ve have already shown the sufficteney in the proof of Theorem 4.1 5 it remains to prove the necessity. Let x € Mand let (vj2) be the solution of (Q) with right-hand side (O.x)+ Then Busy and thus @(B) = N' , 60 that B is a continuous and one-to~ one mapping fron YF onto N'.. Therefore, 8 is an isomorphism from Vi onto M's Hence, by virtue of Lemma 4.1, the inf - sup condition is valid. = ® 4.2. A saddle-point approach Under adequate hypotheses, it is possible to formulate problem (Q) in terms of a saddle~point problem. In addition to the notations of the previous section, we introduce two quadratic functionals J: X +R and £:X*%M +R defined by + Gt ooo eats: (414) £(vyu) = J0v) * bOVW) = <> + £ is usually called the Lagrangian functional associated with problem (Q)+ Consider the following problem : { Find a saddle~point (u,x) in Xx M of the Lagrangian £ , (1) ) ive. find a pair (u,)) in XM such that (415) L(y) -< £(u,4) © LA) VvEX , VueM.44 ‘THEOREM 4.2. Under the hypotheses of Theorem 4.1 and if, moreover, the bilinear form a is symmetric _and semi positive definite on X + (4.16) av) > 0 VVEX, then problem (L) has @ unique solution (uy) in X wueEM. as y is any element of M this is equivalent co (4.2) btu s0 with 2 (4.19) cw > y telly vueM. let CEL(M; M') be defined by = cw) WA, WEN, Let © >0 be a parameter which will tend to zero. We conaider the problem + Find a pair(u,A°) © Xx M satisfying CQ) } (4.20) atu®,v) + bOWAAe) = < Bw > vvex V2 = eek + bMS =< eV eM. As C is non singular, equation (4.21) is equivalent to : (4.22) . thao : Hence we can eliminate 2° from (4,20) and derive another problem, which is obviously equivalent to (Q°) + ( Find u® © X such that Pe ce ¢ 1 rlaye so 1 : (4.23) aCuev) +2 < bv, tne >= vex.46 Remark 4.3, Clearly, when the bilinear forms are symmetric, solving problem (P") is equivalent to finding u° © X such that J(u) = inf 5o(v) where Jv) = vex The expression £ < By - x, @!(Bv-y) > is a penalty term corresponding to the constraint b(v,u) = 0. Thus problem (P") is a penalized version of problem (P). ® THEOREM 4 Under the hypotheses (4.9),(4.19) and if there exists a constant a > 0 such that : o\ : 2 (4.24) avy +< By lev> > allviiz vvex, then problems (Q) and (Q°) both have one and only one solution . Moreover, the following error bound holds for every sufficiently small ¢ (4.25) Nee ty tiara Sk eile 7 + ix x where the constant K depends upon a,8, all, lb ll and lle ll only PROOF. Hypothesis (4.24) implies that a is V-elliptic. Hence problem (Q) has a unique solution (uA) in KxM. Now, it follows from (4.19) and (4.24) that problem (P°) has exactly one solution u° in X- Therefore, if we define by (4.22) then (u",A°) is the only solution of (Q*). It remains to establish (4.25). From (4.20) and (4-1), (4.21) and (4.2), we get t a(ucu®,v) + b(v,AraS) = 0 vvex (4.26) bru", y) + © cA" u) = 0 vuem . The first equation, together with (4.9), yields +47 BPAY < sup PARI < pat null vey Ty x whence (4.27) eae ty < Flat wut. By taking v=u-u" and y= an in (4,26), we find + a(uruSyucu) = © e(Asdea8) = eeQeAEIed®) Se (AAA) owing to (4.19). Then (4.27) gives + © ut tall ut aceeh out) = ee a%) 2, ey 52 vvex, ~ CO Ag) + 0pb py pw) + Py HD YUEN,48 where the parameters p> 0 are arbitrary. Obviously, if a is elliptic and c is Melliptic, these equations determine uniquely and separately ui, and Ay We shall study a variant of this method which leads to a slightly more general algorithm obtained by a procedure known as the " augnented Lagrangian technique. The idea is that, because Bu,- x is supposed to tend to zero, the ellipticity of a can be strengthened by adding to it the term r < Bv,C /(Bv-y) > for some parameter r > 0. This yields the folloving problem. Find a sequence (ut in Xx M satisfying + 1 Be" Bune QD vvex, (4.28) (uv) +r < BvyC > + bia.) =< ave er < Bye ly > (429) = egy Agel) + eLbCo oe) = oT < Xen > vuew. Here again, the computation of u_,, is dissociated from that of ‘ntl ml * ‘Theorem 4.4. We assume that the hypotheses (4.9), (4-16), (4.19) and (4-24) hold « In addition ve assume that the bilinear form c(.,.) is symmetric and that there exists a constant a(r) > 0 such that ol 2 (4.30) a(v,v) +r < Bv,C (By > 3 a(r)IlBv Wage bd doa ‘Then the algorithm (Q,) uniquely determines a sequence (ujs4,) in XxM. Moreover, under the conditions + 4.31) < in sup p< 2 yan ‘ O< into, Soup by <2 vate) we_have Lim (lucul iy + Pagid = 0. PROOF. From (4.19), (4-24) and (4-16), we derive readily that a +2 ety, Cliv> Danity viz vvex.