Part IV, Chapter 20 Galerkin approximation In the previous chapter, we have presented conditions that guarantee the wellposedness of the following model problem: Find u ∈ V such that (20.1) a(u, w) = ℓ(w), ∀w ∈ W, where V and W are Banach spaces, a is a bounded bilinear (or sesquilinear) form on V ×W , and ℓ is a bounded (anti)linear form on W . In this chapter, we study the approximation of (20.1) by the Galerkin method. The starting point is to replace the infinite-dimensional spaces V and W in (20.1) by finite-dimensional spaces Vh and Wh , and possibly the forms a and ℓ by some approximate forms ah and ℓh defined on Vh ×Wh and Wh , respectively. The subscript h refers to the characteristic size of the mesh that is used to build the spaces Vh and Wh ; examples can be found in Chapters 13 and 15. We give discrete versions of the Lax–Milgram Lemma and the BNB Theorem, which are respectively, sufficient conditions, and necessary and sufficient conditions for well-posedness of the approximate problem. Finally, we derive a bound on the approximation error (u − uh ) in a simple setting; the thorough error analysis is postponed to Chapter 21. 20.1 Setting The discrete problem takes the following form: Find uh ∈ Vh such that ah (uh , wh ) = ℓh (wh ), ∀wh ∈ Wh , (20.2) where ah is a bounded bilinear (or sesquilinear) form on Vh ×Wh and ℓh is a bounded (anti)linear form on Wh . Since the spaces Vh and Wh are finitedimensional, (20.2) is called a discrete problem. The space Vh is called the discrete solution space (or trial space), and Wh the discrete test space. 262 Chapter 20. Galerkin approximation Definition 20.1 (Standard Galerkin, Petrov–Galerkin). The Galerkin approximation is said to be standard when Wh = Vh in the discrete problem (20.2). Otherwise, (20.2) is called a Petrov–Galerkin approximation or a nonstandard Galerkin approximation. Definition 20.2 (Conformity). The approximation is said to be conforming if Vh ⊂ V and Wh ⊂ W . There are circumstances when considering nonconforming approximations is useful. Two important examples are discontinuous Galerkin methods where discrete functions are discontinuous across the mesh interfaces (see Chapters 30 and 45) and boundary penalty methods where boundary conditions are enforced weakly (see Chapters 29 and 44). Nonconforming approximations make it often necessary to work with discrete bilinear (and sometimes linear) formsRthat differ from their continuous counterparts. For instance, the bilinear form D ∇v·∇w dx does not make sense if its arguments are discontinuous. Another important example leading to a modification of the bilinear and linear forms at the discrete level is the use of quadratures (see Chapter 26). 20.2 Discrete well-posedness Our goal in this section is to study the well-posedness of the discrete problem (20.2). We equip the discrete spaces Vh and Wh with norms denoted k·kVh and k·kWh . These norms can differ from those of V and W ; one reason can be that the approximation is nonconforming. 20.2.1 Discrete Lax–Milgram Lemma 20.3 (Discrete Lax–Milgram). Let Vh be a finite-dimensional space over R. Set Wh = Vh in problem (20.2). Let ah be a bounded bilinear form on Vh ×Vh and let ℓh be a bounded linear form on Vh . Assume that the bilinear form ah is coercive on Vh , i.e., there is a real number αh > 0 and ξ = ±1 such that ξah (vh , vh ) ≥ αh kvh k2Vh , ∀vh ∈ Vh . (20.3) Then, the problem (20.2) is well-posed and the a priori estimate kuh kVh ≤ 1 ′ αh kℓh kVh holds. In the complex case, the same conclusions hold if there is a real number αh > 0 and a complex number ξ with |ξ| = 1 such that ℜ (ξah (vh , vh )) ≥ αh kvh k2Vh , ∀vh ∈ Vh . (20.4) Proof. We give the proof in the real case; the complex case is treated similarly. Let Ah : Vh → Vh′ be the linear operator such that hAh vh , wh iVh′ ,Vh = Part IV. Galerkin Approximation 263 ah (vh , wh ) for all vh , wh ∈ Vh . Well-posedness of (20.2) is equivalent to Ah being an isomorphism, which owing to the finite-dimensionality of Vh , is equivalent to Ah being injective, i.e., ker(Ah ) = {0}, since dim(Vh ) = dim(Vh′ ) < ∞. Let vh ∈ ker(Ah ), then 0 = ξhAh vh , vh iVh′ ,Vh = ξah (vh , vh ) ≥ αh kvh k2Vh , which proves that vh = 0. Hence, ker(Ah ) = {0}. Remark 20.4 (Variational formulation). As in the continuous setting, see Proposition 19.8, if ah is coercive (with ξ = 1) and symmetric in the real case, uh solves (20.2) with Wh = Vh if and only if uh minimizes the functional Eh (vh ) = 21 ah (vh , vh ) − ℓh (vh ) over Vh . In the conforming case where Vh ⊂ V , ah = a, and ℓh = ℓ, we observe that Eh (uh ) = E(uh ) ≥ E(u). ⊓ ⊔ 20.2.2 Discrete BNB Theorem 20.5 (Discrete BNB). Let Vh and Wh be finite-dimensional spaces over R. Let ah be a bilinear form on Vh ×Wh and let ℓh be a linear form on Wh . Then, problem (20.2) is well-posed if and only if inf sup vh ∈Vh wh ∈Wh ah (vh , wh ) =: αh > 0, kvh kVh kwh kWh dim(Vh ) = dim(Wh ). (20.5a) (20.5b) Moreover, the a priori estimate kuh kVh ≤ α1h kℓh kWh′ holds. Condition (20.5a) is equivalent to the following inf-sup condition: ∃αh > 0, αh kvh kVh ≤ sup wh ∈Wh ah (vh , wh ) , kwh kWh ∀vh ∈ Vh . (20.6) In the complex case, ah (vh , wh ) must be replaced by ℜ(ah (vh , wh )) in the numerator in (20.5a). Proof. We give the proof in the real case; the complex case is treated similarly. Let Ah : Vh → Wh′ be the linear operator such that hAh vh , wh iWh′ ,Wh = ah (vh , wh ), ∀(vh , wh ) ∈ Vh × Wh . (20.7) The well-posedness of (20.2) is equivalent to Ah being an isomorphism, which owing to the finite-dimensional setting and the Rank Theorem, is equivalent to the following conditions: ker(Ah ) = {0} (i.e., Ah is injective), dim(Vh ) = dim(Wh′ ). (20.8a) (20.8b) Since dim(Wh ) = dim(Wh′ ), (20.5b) is equivalent to (20.8b). Let us prove that (20.8a) is equivalent to the inf-sup condition (20.5a). Assume first that (20.5a) holds and let vh ∈ Vh be such that Ah vh = 0. Then, 264 Chapter 20. Galerkin approximation αh kvh kVh ≤ sup wh ∈Wh hAh vh , wh iWh′ ,Wh = 0, kwh kWh whence vh = 0. Conversely, assume ker(Ah ) = {0} and let us prove the inf-sup condition (20.5a). An equivalent statement of (20.5a) is that there is n0 ∈ N∗ such that for all Vh ∈ Vh with kvh kVh = 1, supwh ∈Wh 1 n0 . hAh vh ,wh iW ′ ,W h h kwh kWh > Reasoning by contradiction, consider a sequence (vhn )n∈N in Vh with hAh vhn ,wh iW ′ ,W h h ≤ n1 . Since Vh is finite-dimensional, kvhn kVh = 1 and supwh ∈Wh kwh kWh its unit sphere is compact. Therefore, up to a subsequence, vhn converges to a certain vh in Vh . The limit vh satisfies kvh kVh = 1 and ah (vh , wh ) = 0 for all wh ∈ Wh . The latter property implies vh ∈ ker(Ah ) which contradicts kvh kVh = 1 since ker(Ah ) = {0}. Finally, the a priori estimate is proved as in the continuous case. ⊓ ⊔ Remark 20.6 (Verifying (20.5)). Since (20.5b) is, in general, simple to verify, the key property for discrete well-posedness is (20.5a). ⊓ ⊔ Remark 20.7 (Discrete vs. continuous BNB Theorems). The inf-sup condition (20.5a) is nothing but the first condition (bnb1) of the BNB Theorem 19.10. The second condition (bnb2) amounts to ∀wh ∈ Wh , [ ah (vh , wh ) = 0, ∀vh ∈ Vh ] =⇒ [ wh = 0 ]. (20.9) Let us introduce the adjoint operator A∗h : Wh → Vh′ (note that Wh is reflexive, being finite-dimensional) such that hA∗h wh , vh iVh′ ,Vh = ah (vh , wh ), ∀(vh , wh ) ∈ Vh × Wh . (20.10) Then, (20.9) amounts to A∗h being injective, and this statement is equivalent to (20.5b) provided (20.5a) holds; see Exercise 20.1. ⊓ ⊔ Remark 20.8 (Well-posedness of discrete adjoint problem). Consider the adjoint operator A∗h : Wh → Vh′ defined in (20.10). Then, Ah is an isomorphism if and only if A∗h is an isomorphism; see Exercise 20.2. In other words, the well-posedness of the discrete problem (20.2) is equivalent to the well-posedness of the following adjoint problem: Find yh ∈ Wh such that ah (vh , yh ) = gh (vh ), ∀vh ∈ Vh , for any gh ∈ Vh′ . Moreover, owing to Corollary A.61 (and since Vh is reflexive, being finite-dimensional), Ah and A∗h satisfy the inf-sup condition (20.5a) with the same constant αh , i.e., inf sup vh ∈Vh wh ∈Wh hAh vh , wh iWh′ ,Wh hAh vh , wh iWh′ ,Wh = inf sup . wh ∈Wh vh ∈Vh kvh kVh kwh kWh kvh kVh kwh kWh (20.11) (Note that the numerator of the right-hand side equals hA∗h wh , vh iVh′ ,Vh .) ⊓ ⊔ Part IV. Galerkin Approximation 265 20.3 Fortin’s criterion(♦) Consider a conforming approximation, Vh ⊂ V and Wh ⊂ W , and assume that the discrete problem (20.2) is formulated using the same bilinear form as in the continuous problem, i.e., ah = a. Then, if a is coercive on V with coercivity parameter α, a also coercive on Vh with coercivity parameter αh = α, and the discrete problem (20.2) with Wh = Vh is automatically well-posed. In the more general case of the BNB Theorem 19.10, the situation is different since neither condition (bnb1) nor condition (bnb2) imply its discrete counterpart. h ,w) For instance, since Vh ⊂ V , (bnb1) implies that supw∈W a(v kwkW ≥ αkvh kV for all vh ∈ Vh , but it is not clear that the bound still holds if the argument in the supremum is taken in the subspace Wh ⊂ W . This means that the inf-sup condition (20.5a) needs to be proved. A powerful tool to achieve this is due to Fortin [235]; see also Boffi et al. [64, Prop. 5.4.3]. Here, we consider the Banach space setting and we include a converse statement. Lemma 20.9 (Fortin’s criterion). Let V and W be two real Banach spaces and let a ∈ B(V, W ). Assume that a satisfies the inf-sup condition (bnb1) with constant α > 0. Let Vh ⊂ V and let Wh ⊂ W be equipped with the norms of V and W , respectively. Consider the following two statements: (i) There exists a mapping Πh : W → Wh and a real number γΠ > 0 such that a(vh , Πh w − w) = 0, for all (vh , w) ∈ Vh × W , and γΠ kΠh wkW ≤ kwkW for all w ∈ W . (ii) a satisfies the inf-sup condition inf sup vh ∈Vh wh ∈Wh a(vh , wh ) ≥ β > 0, kvh kV kwh kW (20.12) In the complex case, a(vh , wh ) is replaced by ℜ(a(vh , wh )) in (20.12). Then, (i) ⇒ (ii) with β = γΠ α and (ii) ⇒ (i) with γΠ = β kakB(V,W ) . Proof. We give the proof in the real case; the complex case is treated similarly. (1) Assuming (i), we have sup wh ∈Wh a(vh , wh ) a(vh , Πh w) a(vh , w) ≥ sup = sup kwh kW w∈W kΠh wkW w∈W kΠh wkW ≥ γΠ sup w∈W a(vh , w) ≥ γΠ α kvh kV , kwkW since a satisfies (bnb1) and Vh ⊂ V . This proves (20.12) with β = γΠ α. (2) Conversely, assume that a satisfies (20.12). Let Ah : Vh → Wh′ be the operator defined in (20.7), i.e., hAh vh , wh iWh′ ,Wh = a(vh , wh ). We now apply the converse of Lemma A.57 with Wh in lieu of V , Vh′ in lieu of W , and A∗h in lieu of A; note that we identify Vh and Vh′′ , Wh and Wh′′ , since Vh and Wh are finite-dimensional. According to Lemma A.57, (20.12) implies 266 Chapter 20. Galerkin approximation that for all θh ∈ Vh′ , there exists Φh (θh ) ∈ Wh such that A∗h Φh (θh ) = θh and βkΦh (θh )kW ≤ kθh kVh′ . Let w ∈ W and define θh (w) ∈ Vh′ by setting hθh (w), vh iVh′ ,Vh := a(vh , w) for all vh ∈ Vh . Defining Πh (w) = Φh (θh (w)) yields a(vh , Πh (w)) = hAh vh , Φh (θh (w))iWh′ ,Wh = hA∗h Φh (θh (w)), vh iVh′ ,Vh = hθh (w), vh iVh′ ,Vh = a(vh , w), which proves that a(vh , Πh (w) − w) = 0 for all w ∈ W . Moreover βkΠh (w)kW = βkΦh (θh (w))kW ≤ kθh (w)kVh′ ≤ kakB(V,W ) kwkW , which proves that β kakB(V,W ) kΠh (w)kW ≤ kwkW . ⊓ ⊔ Remark 20.10 (Equivalence, linearity). Only the forward implication in Fortin’s criterion is usually found in the literature. Note that there is a gap in the stability constant γΠ between the direct and converse statements, kakB(V,W ) since the ratio of the two is equal to (which is independent of the α discrete setting). Moreover, in the converse statement, the mapping Πh can be nonlinear (since the mapping Φh can be nonlinear). The existence of a linear mapping is granted in a Hilbertian setting, see Remark A.58. ⊓ ⊔ 20.4 Basic error estimate In this section, we bound the approximation error (u − uh ). We consider the relatively simple setting of a conforming approximation (Vh ⊂ V and Wh ⊂ W ) with ah = a and ℓh = ℓ. We assume that the inf-sup condition (20.5a) holds, as well as dim(Vh ) = dim(Wh ), so that the discrete problem (20.2) is well-posed. Simple sufficient conditions for (20.5a) to hold are Wh = Vh and a coercive on Vh (for which a sufficient condition is that a be coercive on V ). Lemma 20.11 (Galerkin’s orthogonality). Let u solve (20.1) and let uh solve (20.2) with Vh ⊂ V , Wh ⊂ W , ah = a, and ℓh = ℓ. Then, the following holds: a(u − uh , wh ) = 0, ∀wh ∈ Wh . (20.13) Proof. For all wh ∈ Wh , we observe that a(u, wh ) = ℓ(wh ) since Wh ⊂ W and ℓ(wh ) = a(uh , wh ) since ah = a and ℓh = ℓ. ⊓ ⊔ The starting point of the error analysis is to make sure that (20.2) is consistent with the original problem (20.1). Loosely speaking, consistency is checked by inserting the exact solution into the discrete problem and verifying that the result is small. In the present situation, we infer that a(u, wh ) = l(wh ) for all wh ∈ Wh (since ah = a, ℓh = ℓ, and Wh ⊂ W ), i.e., the approximate Part IV. Galerkin Approximation 267 problem is exactly consistent with the continuous one. This exact consistency property is just a reformulation of Galerkin’s orthogonality. Let us now bound the approximation error. Since the approximation setting is conforming, we equip the subspaces Vh ⊂ V and Wh ⊂ W with the norms k·kV and k·kW , respectively. Recall that kakB(V,W ) := . sup(v,w)∈V ×W kvka(v,w) V kwkW Lemma 20.12 (Basic error estimate). Assume that problems (20.1) and (20.2) are well-posed with solutions u and uh . Assume that the approximation setting is conforming with ah = a and ℓh = ℓ. Let αh > 0 be the constant in the inf-sup condition (20.5a). Then, the following holds: kakB(V,W ) inf ku − vh kV . (20.14) ku − uh kV ≤ 1 + αh vh ∈Vh Consequently, if approximability holds in the sense that lim inf kv − vh kV = 0, h→0 vh ∈Vh ∀v ∈ V, (20.15) then limh→0 ku − uh kV = 0. Proof. Let vh ∈ Vh . Using stability (i.e., the inf-sup condition (20.5a)), consistency (i.e., Galerkin’s orthogonality), and the boundedness of a, we infer that αh kuh − vh kV ≤ sup wh ∈Wh = sup wh ∈Wh a(uh − vh , wh ) kwh kW a(u − vh , wh ) ≤ kakB(V,W ) ku − vh kV . kwh kW ⊓ ⊔ We conclude using the triangle inequality. Lemma 20.12 is derived in Babuška and Aziz [33]. Note that (20.14) implies that if the exact solution turns out to be in Vh , then uh = u. Whenever V and W are Hilbert spaces, the error estimate can be sharpened as follows: kakB(V,W ) (20.16) inf ku − vh kV , ku − uh kV ≤ αh vh ∈Vh see Xu and Zikatanov [495] and Exercise 20.7. Moreover, if the bilinear form a is coercive (with parameter α), W = V , and Wh = Vh , the error estimate is known as Céa’s Lemma [136] and takes the form (see Exercise 20.6) kakB(V,V ) inf ku − vh kV . (20.17) ku − uh kV ≤ α vh ∈Vh 268 Chapter 20. Galerkin approximation 20.5 Stability loss(♦) The purpose of this section is to illustrate on a simple one-dimensional transport (advection) equation the loss of stability when applying a naive Galerkin approximation technique to a first-order PDE. We consider the onedimensional transport equation u′ = f in D, u(0) = 0, (20.18) with D := (0, 1) and given f ∈ L2 (D). Following §18.2.3), we consider the L2 based weak formulation with solution space V := {v ∈ H 1 (D) | v(0) = 0} and R1 test space W := L2 (D) (we drop the superscripts). Setting a(v, w) = 0 v ′ w dt for all (v, w) ∈ V ×W , the weak problem consists of finding u ∈ V such that R1 a(u, w) = 0 f w dt for all w ∈ W ; this problem is well-posed, see Exercise 19.8. Consider H 1 -conforming P1 (Lagrange) finite elements. Let N be a positive integer. Set h = N1 and xi = ih for all i ∈ {0: N }. Let us use the finite element space Vh = {vh ∈ C 0 (D) | ∀i ∈ {0: N −1}, vh|[xi ,xi+1 ] ∈ P1 | vh (0) = 0} both as solution space and as test space. Since Vh is a subspace of V and of W , the approximation is conforming, and we can consider using the bilinear form a in the discrete problem in a standard Galerkin setting. R 1The discrete problem consists of finding uh ∈ Vh such that a(uh , wh ) = 0 f wh dt for all wh ∈ Vh . (The resulting linear system is identical to that obtained with centered finite differences.) The well-posedness of the discrete problem falls into the framework of Theorem 20.5. Since the discrete solution and test spaces have the same dimension, well-posedness amounts to inf sup vh ∈Vh wh ∈Vh a(vh , wh ) =: αh > 0. kvh kH 1 (D) kwh kL2 (D) Unfortunately, it turns out that αh is not bounded away from zero from below uniformly with respect to h. Proposition 20.13 (Stability loss). There are c1 > 0 and c2 > 0, uniform with respect to h, such that c1 h ≤ αh ≤ c2 h. Proof. Assume that N is even, the odd case being treated similarly. Denote by {ϕ1 , . . . , ϕN } the nodal basis of Vh . Recall from §22.2 the operator Rϕ : RN → Vh that reconstructs a function in Vh from its coordinate vector. A direct computation using Simpson’s rule (see §??) shows that 61 hkY k2ℓ2 (RN ) ≤ kRϕ (Y )k2L2 (D) ≤ hkY k2ℓ2 (RN ) for all Y ∈ RN . These bounds imply that 1 h− 2 kGkℓ2 (RN ) ≤ sup wh ∈Vh √ 1 a(vh , wh ) ≤ 6h− 2 kGkℓ2 (RN ) , kwh kL2 (D) (20.19) for all vh ∈ Vh with G ∈ RN such that Gi = a(vh , ϕi ) for all i ∈ {1: N }. (2) The bound from above. The idea consists of constructing an oscillating Part IV. Galerkin Approximation 269 PN function vh ∈ Vh ; namely, set vh = i=1 Vi ϕi with V2i = 2ih if 1 ≤ i ≤ N2 and V2i+1 = 1 if 0 ≤ i ≤ N2 − 1. Set V0 = 0 and denote by [·] the integer part operator. Then, hkvh′ k2L2 (D) = N −1 X i=0 N 2 ≥ −1 X i=0 h Z xi+1 xi 2 (vh′ ) dt = N −1 X i=0 [N 4 ] (1 − 2ih)2 ≥ X i=0 (Vi+1 − Vi )2 (1 − 2ih)2 ≥ 41 ([ N4 ] + 1), 1 since 1 − 2ih ≥ 12 if i ≤ [ N4 ]. Using the inequality [ N4 ] + 1 > N4 = 4h , yields 1 ′ 2 kvh kL (D) ≥ 4h . Furthermore, Gi = 0 if i is even, and Gi = h otherwise. Hence, kGkℓ2 (RN ) ≤ h1/2 . Using the rightmost bound in (20.19) yields the √ bound from above in (20.13) with c2 = 4 6. (3) The bound from below. Let vh be arbitrary in Vh . Set V0 = 0 and VN +1 = VN . Since Gi = 12 (Vi+1 − Vi−1 ) for all i ∈ {1: N }, we infer that, Pi−1 for even indices, |V2i | ≤ 2 k=0 |G2k+1 |, while for odd indices, |V2i−1 | ≤ PN PN P N2 |G2k | ≤ 2 k=1 |Gk |. Therefore, kV kℓ∞ (RN ) ≤ 2 k=1 |Gk | ≤ |VN +1 | + 2 k=i 2h−1/2 kGkℓ2 (RN ) owing to the Cauchy–Schwarz inequality. Furthermore, a di√ rect computation shows that kvh′ kL2 (D) ≤ 2h−1 kV kℓ∞ (RN ) . Using the left√ most bound in (20.19) and the Poincaré inequality 2kvh′ kL2 (D) ≥ kvh kH 1 (D) yields the bound from below in (20.13) with c2 = 14 . ⊓ ⊔ The main conclusion we can draw from Proposition 20.13 is that a naive Galerkin approximation of first-order PDEs cannot produce optimal error estimates, even though it yields an invertible linear system (c1 6= 0). In practice, this problem manifests itself through the presence of spurious wiggles in the approximate solution. Exercises Exercise 20.1 (Condition (bnb2)). Prove that (20.9) is equivalent to (20.5b) provided (20.5a) holds. (Hint: use that dim(Wh ) = rank(Ah ) + dim(ker(A∗h )), with A∗h defined in (20.10), together with the Rank Theorem.) Exercise 20.2 (Bijectivity of A∗h ). Prove that Ah is an isomorphism if and only if A∗h is an isomorphism. (Hint: use dim(Vh ) = rank(A∗h ) + dim(ker(Ah )) and dim(Wh ) = rank(Ah ) + dim(ker(A∗h )).) Exercise 20.3 (Minimal residual, Petrov–Galerkin). Let V, W be two Hilbert spaces, let A ∈ L(V ; W ′ ) be an isomorphism, and let ℓ ∈ W ′ . Consider 270 Chapter 20. Galerkin approximation a finite-dimensional subspace Vh ⊂ V . We want to approximate the unique solution u ∈ V satisfying Au = ℓ by using a conforming Petrov–Galerkin approx−1 imation with Wh = JW AVh ⊂ W where JW : W → W ′ is the Riesz–Fréchet isomorphism. The discrete bilinear form is ah (vh , wh ) = hAvh , wh iW ′ ,W , and the discrete linear form is ℓh (wh ) = ℓ(wh ). (i) Prove that the discrete problem (20.2) is well-posed. (ii) Show that its unique solution minimizes the residual functional R(v) = kAv − ℓkW ′ over Vh . Exercise 20.4 (Variant of Fortin’s criterion). Write a variant of Fortin’s criterion assuming V, W reflexive and using this time an operator Πh : V → Vh such that a(Πh v − v, wh ) = 0 for all (v, wh ) ∈ V × Wh and γΠ kΠh vkV ≤ kvkV for all v ∈ V for some γΠ > 0. (Hint: use (20.11) and Corollary A.61.) Exercise 20.5 (Fortin’s criterion). Let V and W be two real Banach spaces and let a ∈ B(V, W ). Let Vh ⊂ V and let Wh ⊂ W be equipped with the norms of V and W , respectively. Assume that there are two operators (not necessarily linear) Π1,h , Π2,h : W −→ Wh and two uniform constants c1 , c2 > 0 such that kΠ1,h wkW ≤ c1 kwkW , kΠ2,h (I − Π1,h )wkW ≤ c2 kwkW and a(vh , Π2,h w) = a(vh , Π2,h w) for all vh ∈ Vh , w ∈ W . Prove that Πh = Π1,h + Π2,h (I − Π1,h ) satisfies Fortin’s criterion. Exercise 20.6 (Céa’s Lemma). Assume that a is coercive on V with parameter α. Take W = V and Wh = Vh . Assume Vh ⊂ V and ah = a. Prove (20.17) 1 kakB(V,V ) 2 inf vh ∈Vh ku − vh kV if a is symmetric. and ku − uh kV ≤ α Exercise 20.7 (Xu–Zikatanov Lemma). Let V, W be two Hilbert spaces, let Vh ⊂ V and Wh ⊂ W with dim(Vh ) = dim(Wh ), and let a be a bounded bilinear form on V ×W satisfying the stability property (??). Let Ph : V → Vh map u ∈ V to the unique solution of a(uh − u, wh ) = 0 for all wh ∈ Wh . kak B(V,W ) (i) Prove that Ph ∈ L(V ; V ) with kPh kL(V ;V ) ≤ . αh (ii) Prove that Ph ◦Ph = Ph and that (I − Ph )(vh ) = 0 for all vh ∈ Vh . (iii) Prove (20.16). (Hint: see the proof of Theorem 4.14.) Exercise 20.8 (One-dimensional transport). Consider the bilinear form a and the discrete space Vh from §20.5. Set Wh = {wh ∈ L∞ (D) | ∀i ∈ {0: N −1}, wh|[xi ,xi+1 ] ∈ P0 }. Verify that dim(Vh ) = dim(Wh ) and prove that, a(vh ,wh ) uniformly with respect to h, inf vh ∈Vh supwh ∈Wh kvh k 1,1 kwh kL∞ (D) > 0. W (Hint: see the proof of Proposition 19.16.) (D)