1 Rates of convergence for approximations of viscosity solutions and homogenization Panagiotis E. Souganidis The University of Texas at Austin Nonlinear approximation techniques using L1 Texas A&M University May 2008 2 F(D2 u, Du, u, x) = 0 in U 3 F(D2 u, Du, u, x) = 0 in U numerical approximations monotone, stable, consistent approximations Fh (D2h uh , Dh uh , uh , x) = 0 in Uh uh → u and mesh size h, numerical nonlinearity Fh , solution uh , domain Uh = U ∩ ZNh uh − u = O(σ(h)) 4 F(D2 u, Du, u, x) = 0 in U numerical approximations monotone, stable, consistent approximations Fh (D2h uh , Dh uh , uh , x) = 0 in Uh uh → u and uh − u = O(σ(h)) mesh size h, numerical nonlinearity Fh , solution uh , domain Uh = U ∩ ZNh homogenization x F(D2 uε , Duε , uε , , ω) = 0 in U ε uε → u0 and F0 (D2 u0 , Du0 , u0 ) = 0 in U uε − u0 = O(σ(ε)) 5 NUMERICAL APPROXIMATIONS 6 NUMERICAL APPROXIMATIONS monotone schemes convergence uh −→ u rate of convergence ku − uh k = O(hα ) first-order (Hamilton-Jacobi) α = 1/2 Crandall-Lions, Souganidis α = 1/27 α = 1/5 α = 1/2 α small Krylov second-order convex, deg. elliptic uniformly elliptic h→0 Crandall-Lions, Souganidis Barles-Souganidis Barles-Jakobsen Krylov Caffarelli-Souganidis 7 NUMERICAL APPROXIMATIONS monotone schemes convergence uh −→ u rate of convergence ku − uh k = O(hα ) first-order (Hamilton-Jacobi) α = 1/2 Crandall-Lions, Souganidis α = 1/27 α = 1/5 α = 1/2 α small Krylov second-order convex, deg. elliptic uniformly elliptic h→0 Crandall-Lions, Souganidis Barles-Souganidis Barles-Jakobsen Krylov Caffarelli-Souganidis non-monotone schemes few convergence results for non monotone schemes TVD filtered and higher-order schemes Osher-Tadmor Lions-Souganidis 8 STOCHASTIC HOMOGENIZATION F(D2 uε , Duε , εx , ω) = 0 F uniformly elliptic stationary ergodic* * stationarity: f (y, ω) is stationary if µ({ω : f (y, ω) > α}) independent of y ergodicity: all translation invariant quantities are constant a.s. in ω 9 STOCHASTIC HOMOGENIZATION F(D2 uε , Duε , εx , ω) = 0 convergence F uniformly elliptic stationary ergodic* uε (·, ω) → u0 a.s. & F0 (D2 u0 , Du0 ) = 0 linear Papanicolaou-Varadhan, Kozlov nonlinear Caffarell-Souganidis-Wang * stationarity: f (y, ω) is stationary if µ({ω : f (y, ω) > α}) independent of y ergodicity: all translation invariant quantities are constant a.s. in ω 10 STOCHASTIC HOMOGENIZATION F(D2 uε , Duε , εx , ω) = 0 convergence F uniformly elliptic stationary ergodic* uε (·, ω) → u0 a.s. F0 (D2 u0 , Du0 ) = 0 & linear Papanicolaou-Varadhan, Kozlov nonlinear Caffarell-Souganidis-Wang rates of convergence strongly mixing media with algebraic rate linear kuε − u0 k = O(εγ ) a.s. nonlinear Yurinskii −c| ln ε|1/2 kuε − u0 k = O(e ) off −c| ln ε|1/2 a set with probability O(e ) * stationarity: f (y, ω) is stationary if µ({ω : f (y, ω) > α}) independent of y ergodicity: all translation invariant quantities are constant a.s. in ω Caffarelli-Souganidis 11 CONVERGENCE OF MONOTONE APPROXIMATIONS (F) F(D2 u, Du, u, x) = 0 12 CONVERGENCE OF MONOTONE APPROXIMATIONS (F) F(D2 u, Du, u, x) = 0 F degenerate elliptic (X ≦ Y =⇒ F(X, p, r, x) ≧ F(Y, p, r, x)) 13 CONVERGENCE OF MONOTONE APPROXIMATIONS (F) F(D2 u, Du, u, x) = 0 F degenerate elliptic approximation scheme (X ≦ Y =⇒ F(X, p, r, x) ≧ F(Y, p, r, x)) S([uh ]x , uh (x), x, h) = 0 14 CONVERGENCE OF MONOTONE APPROXIMATIONS (F) F(D2 u, Du, u, x) = 0 F degenerate elliptic approximation scheme monotone (X ≦ Y =⇒ F(X, p, r, x) ≧ F(Y, p, r, x)) S([uh ]x , uh (x), x, h) = 0 u≥v =⇒ S([u]x , s, x, h) ≦ S([v]x , s, x, h) 15 CONVERGENCE OF MONOTONE APPROXIMATIONS (F) F(D2 u, Du, u, x) = 0 F degenerate elliptic approximation scheme monotone stable (X ≦ Y =⇒ F(X, p, r, x) ≧ F(Y, p, r, x)) S([uh ]x , uh (x), x, h) = 0 u≥v =⇒ kuh k ≦ C S([u]x , s, x, h) ≦ S([v]x , s, x, h) independent of h 16 CONVERGENCE OF MONOTONE APPROXIMATIONS (F) F(D2 u, Du, u, x) = 0 F degenerate elliptic approximation scheme monotone stable consistent (X ≦ Y =⇒ F(X, p, r, x) ≧ F(Y, p, r, x)) S([uh ]x , uh (x), x, h) = 0 u≥v =⇒ S([u]x , s, x, h) ≦ S([v]x , s, x, h) kuh k ≦ C independent of h S([φ + ξ]x , φ(y) + ξ, y, h) −→ F(D2 φ(x), Dφ(x), φ(x), x) h→0 y→x ξ→0 (φ smooth) 17 CONVERGENCE OF MONOTONE APPROXIMATIONS (F) F(D2 u, Du, u, x) = 0 F degenerate elliptic approximation scheme monotone stable consistent Theorem: (X ≦ Y =⇒ F(X, p, r, x) ≧ F(Y, p, r, x)) S([uh ]x , uh (x), x, h) = 0 u≥v S([u]x , s, x, h) ≦ S([v]x , s, x, h) =⇒ kuh k ≦ C independent of h S([φ + ξ]x , φ(y) + ξ, y, h) −→ F(D2 φ(x), Dφ(x), φ(x), x) h→0 y→x ξ→0 uh −→ u h→0 u solution of (F) (φ smooth) 18 Proof 19 Proof S([uh ]x , uh (x), x, h) = 0 20 Proof S([uh ]x , uh (x), x, h) = 0 u∗ (x) = lim uh (y) y→x stability =⇒ h→0 u∗ (x) = lim uh (y) y→0 h→0 exist 21 Proof S([uh ]x , uh (x), x, h) = 0 u∗ (x) = lim uh (y) y→x stability h→0 =⇒ exist u∗ (x) = lim uh (y) y→0 h→0 u∗ subsolution monotonicity =⇒ consistency of u∗ supersolution F(D2 u, Du, u, x) = 0 22 Proof S([uh ]x , uh (x), x, h) = 0 u∗ (x) = lim uh (y) y→x stability h→0 =⇒ exist u∗ (x) = lim uh (y) y→0 h→0 u∗ subsolution monotonicity =⇒ consistency comparison for definition of of F(D2 u, Du, u, x) = 0 u∗ supersolution (F) =⇒ u∗ ≦ u∗ u∗ , u∗ =⇒ u∗ ≦ u∗ ) u∗ = u∗ = u solution of (F) and =⇒ uh −→ u h→0 23 ∗ u subsolution iff ( for all smooth φ and all max x of u∗ − φ F(D2 φ(x), Dφ(x), u∗ (x), x) ≦ 0 24 ∗ u subsolution iff ( for all smooth φ and all max x of u∗ − φ F(D2 φ(x), Dφ(x), u∗ (x), x) ≦ 0 fix φ smooth x0 max of u∗ − φ and u∗ (x0 ) = φ(x0 ) 25 ∗ u subsolution iff ( for all smooth φ and all max x of u∗ − φ F(D2 φ(x), Dφ(x), u∗ (x), x) ≦ 0 fix φ smooth x0 max of u∗ − φ and “uh → u∗ ” =⇒ u∗ (x0 ) = φ(x0 ) xh max of uh − φ =⇒ uh ≦ φ − ξh xh → x0 and ξh = uh (xh ) − φ(xh ) → 0 as h → 0 26 ∗ u subsolution iff ( for all smooth φ and all max x of u∗ − φ F(D2 φ(x), Dφ(x), u∗ (x), x) ≦ 0 fix φ smooth x0 max of u∗ − φ and “uh → u∗ ” =⇒ u∗ (x0 ) = φ(x0 ) xh max of uh − φ =⇒ uh ≦ φ − ξh xh → x0 and ξh = uh (xh ) − φ(xh ) → 0 as h → 0 uh ≦ φ + ξh w w monotonicity S([φ + ξh ]x , φ(xh ) + ξh , xh , h) ≦ 0 = S([uh ]x , uh (xh ), xh , h) 27 ∗ u subsolution iff ( for all smooth φ and all max x of u∗ − φ F(D2 φ(x), Dφ(x), u∗ (x), x) ≦ 0 fix φ smooth x0 max of u∗ − φ and “uh → u∗ ” =⇒ u∗ (x0 ) = φ(x0 ) xh max of uh − φ =⇒ uh ≦ φ − ξh xh → x0 and ξh = uh (xh ) − φ(xh ) → 0 as h → 0 uh ≦ φ + ξh w w monotonicity S([φ + ξh ]x , φ(xh ) + ξh , xh , h) ≦ 0 = S([uh ]x , uh (xh ), xh , h) w w consistency xh → 0 ξh → 0 F(D2 φ(x0 ), Dφ(x0 ), u∗ (x0 ), x0 ) ≦ 0 28 Examples • Hamilton-Jacobi equation ut + H(ux ) = 0 29 Examples • Hamilton-Jacobi equation ut + H(ux ) = 0 r − u(x, t − h) h » “ – ” θ u(x + λh) − 2u(x) + u(x − λh) u(x + λh) − u(x − λh) − −h H 2λh λ λh S([u]x , r, x, t, h) = Lax-Friedrichs “ n n n n ff n ” Uj+1 − Uj−1 θ Uj+1 − 2Uj + Uj−1 − Ujn+1 = Ujn − ∆t H 2∆x λ ∆x (CFL) 2θ − λkH ′ k ≧ 0 =⇒ monotonicity (λ = ∆t ) ∆x 30 • Isaacs-Bellman equation ut + F(D2 u, Du, x) = 0 31 • Isaacs-Bellman equation ut + F(D2 u, Du, x) = 0 F(X, p, x) = max min[−tr (aα,β (x)X) − bα,β (x) · p] α β stochastic differential games 32 • Isaacs-Bellman equation ut + F(D2 u, Du, x) = 0 F(X, p, x) = max min[−tr (aα,β (x)X) − bα,β (x) · p] α β 8 u(x + ei h) − u(x) > > < h h,α,β u= uxi ≈ Di > u(x − hei ) − u(x) > : h stochastic differential games if bα,β (x) ≧ 0 i if bα,β (x) < 0 i 33 • Isaacs-Bellman equation ut + F(D2 u, Du, x) = 0 F(X, p, x) = max min[−tr (aα,β (x)X) − bα,β (x) · p] α β 8 u(x + ei h) − u(x) > > < h h,α,β u= uxi ≈ Di > u(x − hei ) − u(x) > : h uxi xj stochastic differential games if bα,β (x) ≧ 0 i if bα,β (x) < 0 i 8 u(x + hei ) − 2u(x) + u(x − hei ) > if i = j > > > h2 > > > 8 > > > <> > > h,αβ > ≈ Dij u(x) = > if aα,β ≧0 ij >< > > if i 6= j > > > > α,β > > > > if a < 0 ij > > >> : : 34 S([u]h , r, x, t, h) = r − u(x, t − h) h h i h,α,β −h max min − aα,β u(x, t − h) − bα,β (x)Dh,α,β u(x, t − h) ij (x) · Dij i i α β 35 S([u]h , r, x, t, h) = r − u(x, t − h) h h i h,α,β −h max min − aα,β u(x, t − h) − bα,β (x)Dh,α,β u(x, t − h) ij (x) · Dij i i α β aα,β diagonally dominant aα,β ii (x) − X j6=i w w S monotone |aα,β ij (x)| ≧ 0 for all i, α, β, x 36 RATES OF CONVERGENCE – Hamilton-Jacobi 37 RATES OF CONVERGENCE – Hamilton-Jacobi F(Du, u, x) = 0 S([uh ]x , uh (x), x, h) = 0 38 RATES OF CONVERGENCE – Hamilton-Jacobi F(Du, u, x) = 0 S([uh ]x , uh (x), x, h) = 0 monotonicity u≦v and m≧0 =⇒ S([u + m]x , r + m, x, h) ≧ λm + S([v]x , r, x, h) 39 RATES OF CONVERGENCE – Hamilton-Jacobi F(Du, u, x) = 0 S([uh ]x , uh (x), x, h) = 0 monotonicity u≦v stability and m≧0 =⇒ kuh k ≦ C S([u + m]x , r + m, x, h) ≧ λm + S([v]x , r, x, h) independent of h 40 RATES OF CONVERGENCE – Hamilton-Jacobi F(Du, u, x) = 0 S([uh ]x , uh (x), x, h) = 0 monotonicity u≦v stability consistency and m≧0 =⇒ kuh k ≦ C S([u + m]x , r + m, x, h) ≧ λm + S([v]x , r, x, h) independent of h |S([ψ]x , ψ(x), x, h) − F(Dψ(x), ψ(x), x)| ≦ C(1 + |D2 ψ|)h (all ψ smooth) 41 RATES OF CONVERGENCE – Hamilton-Jacobi F(Du, u, x) = 0 S([uh ]x , uh (x), x, h) = 0 monotonicity u≦v stability consistency Theorem: and m≧0 =⇒ kuh k ≦ C S([u + m]x , r + m, x, h) ≧ λm + S([v]x , r, x, h) independent of h |S([ψ]x , ψ(x), x, h) − F(Dψ(x), ψ(x), x)| ≦ C(1 + |D2 ψ|)h (all ψ smooth) u Lipschitz solution of F = 0 |u − uh | ≦ Kh1/2 F convex −Kh≦u−uh ≦Kh1/2 =⇒ (K = K(F, kDuk)) Capuzzo-Dolcetta-Ishii 42 proof of uh ≦ u + Kh1/2 43 inf - and sup - convolution regularizations 44 inf - and sup - convolution regularizations u bounded, continuous ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] and uε (x) = inf[u(y) + (2ε)−1 |x − y|2 ] 45 inf - and sup - convolution regularizations u bounded, continuous ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] and • ūε , uε Lipschitz continuous and uε (x) = inf[u(y) + (2ε)−1 |x − y|2 ] ūε ↓ u, uε ↑ u 46 inf - and sup - convolution regularizations u bounded, continuous ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] and • ūε , uε Lipschitz continuous and uε (x) = inf[u(y) + (2ε)−1 |x − y|2 ] ūε ↓ u, uε ↑ u ūε (x) = u(y(x)) − (2ε) −1 |x − y(x)| 2 ūε (x + hχ) − 2ūε (x) + ūε (x − hχ) • ūε uε semi-convex semi-concave (D ūε ≧ − εI ) (D2 uε ≦ εI ) ≧ u(y(x)) − (2ε) 2 (−2ε) −2u(y(x)) + (2ε) (−2ε) −1 2 −1 −1 −1 |x − y(x) + hχ| 2 2 |x − y(x)| + u(y(x)) |x − y(x) − hχ|2 = 2 2 −1 2 [|x−y(x)+hχ| −2|x−y(x)| +|x−y(x)−hχ| ] = −2 h 47 inf - and sup - convolution regularizations u bounded, continuous ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] and • ūε , uε Lipschitz continuous and uε (x) = inf[u(y) + (2ε)−1 |x − y|2 ] ūε ↓ u, uε ↑ u ūε (x) = u(y(x)) − (2ε) −1 |x − y(x)| 2 ūε (x + hχ) − 2ūε (x) + ūε (x − hχ) • ūε uε semi-convex semi-concave (D ūε ≧ − εI ) (D2 uε ≦ εI ) (−2ε) • ūε , uε ≧ u(y(x)) − (2ε) 2 twice differentiable a.e. −2u(y(x)) + (2ε) (−2ε) −1 2 −1 −1 −1 |x − y(x) + hχ| 2 2 |x − y(x)| + u(y(x)) |x − y(x) − hχ|2 = 2 2 −1 2 [|x−y(x)+hχ| −2|x−y(x)| +|x−y(x)−hχ| ] = −2 h 48 inf - and sup - convolution regularizations u bounded, continuous ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] and • ūε , uε Lipschitz continuous and uε (x) = inf[u(y) + (2ε)−1 |x − y|2 ] ūε ↓ u, uε ↑ u ūε (x) = u(y(x)) − (2ε) −1 |x − y(x)| 2 ūε (x + hχ) − 2ūε (x) + ūε (x − hχ) • ūε uε semi-convex semi-concave (D ūε ≧ − εI ) (D2 uε ≦ εI ) (−2ε) • ūε , uε ≧ u(y(x)) − (2ε) 2 −2u(y(x)) + (2ε) (−2ε) −1 −1 −1 −1 |x − y(x) + hχ| 2 2 |x − y(x)| + u(y(x)) |x − y(x) − hχ|2 = 2 2 2 −1 2 [|x−y(x)+hχ| −2|x−y(x)| +|x−y(x)−hχ| ] = −2 twice differentiable a.e. • u Lipschitz continuous ⇒ kūε − uk ≦ kDukε, kuε − uk ≦ kDukε h 49 inf - and sup - convolution regularizations u bounded, continuous ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] and • ūε , uε Lipschitz continuous and uε (x) = inf[u(y) + (2ε)−1 |x − y|2 ] ūε ↓ u, uε ↑ u ūε (x) = u(y(x)) − (2ε) −1 |x − y(x)| 2 ūε (x + hχ) − 2ūε (x) + ūε (x − hχ) • ūε uε semi-convex semi-concave (D ūε ≧ − εI ) (D2 uε ≦ εI ) (−2ε) • ūε , uε ≧ u(y(x)) − (2ε) 2 −2u(y(x)) + (2ε) (−2ε) −1 −1 −1 −1 |x − y(x) + hχ| 2 2 |x − y(x)| + u(y(x)) |x − y(x) − hχ|2 = 2 2 2 −1 2 [|x−y(x)+hχ| −2|x−y(x)| +|x−y(x)−hχ| ] = −2 twice differentiable a.e. • u Lipschitz continuous ⇒ kūε − uk ≦ kDukε, kuε − uk ≦ kDukε • F(D2 u, Du, u, x) ≦ 0 ⇒ “F(D2 ūε , Dūε , ūε , x) ≦ ε” • F(D2 u, Du, u, x) ≧ 0 ⇒ “F(D2 uε , Duε , uε , x) ≧ −ε” h 50 u(x) = −|x| uε (x) = sup[−|y| − (2ε)−1 |x − y|2 ] 8 1 > <− |x|2 for |x| ≦ ε ε 2ε u (x) = > :−|x| + ε for |x| ≧ ε 3 uε (x) = inf[−|y| + (2ε)−1 |x − y|2 ] uε (x) = −|x| − 3ε 2 51 parallel surface regularization 52 parallel surface regularization d(x, y) = dist((x, y), graph u) (x ∈ RN , y ∈ R) d(x, u(x)) = 0 graph ūε = {(x, y) ∈ RN+1 : d(x, y) = ε and y ≧ u(x)} fi d(x, ū (x)) = ε ε graph uε = {(x, y) ∈ RN+1 : d(x, y) = ε and y ≦ u(x)} fi d(x, u (x)) = ε ε • ūε ūε ≧ u uε ≦ u “F(D2 ūε , Dūε , ūε , x) ≦ ε” semi-concave, ūε ↓ u , semi-convex, uε ↑ u , “F(D2 uε , Duε , uε , x) ≧ −ε” Lip. continuous, • uε 53 proof of uh ≦ u + Kh1/2 54 proof of • u Lipschitz =⇒ uh ≦ u + Kh1/2 |u − uε | ≦ kDukε 55 proof of • u Lipschitz • compare =⇒ uh and uε uh ≦ u + Kh1/2 |u − uε | ≦ kDukε 56 proof of • u Lipschitz • compare • use uε =⇒ uh uh ≦ u + Kh1/2 |u − uε | ≦ kDukε and uε as a test function m = max(uh − uε ) =⇒ uh ≦ uε + m 57 proof of • u Lipschitz • compare • use • consistency uε =⇒ uh uh ≦ u + Kh1/2 |u − uε | ≦ kDukε and uε as a test function m = max(uh − uε ) =⇒ uh ≦ uε + m “ 1” F(Duε , uε , x) ≧ 0 =⇒ S([uε ], uε (x), x, h) ≧ −C(1 + |D2 uε |)h ≧ −C 1 + h ε 58 proof of • u Lipschitz • compare • use • consistency uε =⇒ uh uh ≦ u + Kh1/2 |u − uε | ≦ kDukε and uε as a test function m = max(uh − uε ) =⇒ uh ≦ uε + m “ 1” F(Duε , uε , x) ≧ 0 =⇒ S([uε ], uε (x), x, h) ≧ −C(1 + |D2 uε |)h ≧ −C 1 + h ε • monotonicity S([uε ]x , uε (x), x, h) ≦ S([uh − m]x , uh (x) − m, x, h) ≦ −λm + S([uh ]x , uh (x), x, h) ” “ =⇒ λm ≦ C 1 + 1ε h 59 proof of • u Lipschitz • compare • use • consistency uε =⇒ uh uh ≦ u + Kh1/2 |u − uε | ≦ kDukε and uε as a test function m = max(uh − uε ) =⇒ uh ≦ uε + m “ 1” F(Duε , uε , x) ≧ 0 =⇒ S([uε ], uε (x), x, h) ≧ −C(1 + |D2 uε |)h ≧ −C 1 + h ε • monotonicity S([uε ]x , uε (x), x, h) ≦ S([uh − m]x , uh (x) − m, x, h) ≦ −λm + S([uh ]x , uh (x), x, h) ” “ =⇒ λm ≦ C 1 + 1ε h ” “ • total error kDukε + C 1 + 1ε h ≈ Kh1/2 60 RATES OF CONVERGENCE – uniformly elliptic equations 61 RATES OF CONVERGENCE – uniformly elliptic equations F(D2 u, Du, x) = 0 S([uh ]x , uh (x), x, h) = 0 consistency |S([φ]x , φ(x), x, h) − F(D2 φ(x), Dφ(x), x)| ≦ K(1 + |D3 φ|)h 62 RATES OF CONVERGENCE – uniformly elliptic equations F(D2 u, Du, x) = 0 S([uh ]x , uh (x), x, h) = 0 consistency |S([φ]x , φ(x), x, h) − F(D2 φ(x), Dφ(x), x)| ≦ K(1 + |D3 φ|)h problem no regularization of viscosity solutions controlling “third-derivatives” and “preserving” equation 63 RATES OF CONVERGENCE – uniformly elliptic equations F(D2 u, Du, x) = 0 S([uh ]x , uh (x), x, h) = 0 consistency |S([φ]x , φ(x), x, h) − F(D2 φ(x), Dφ(x), x)| ≦ K(1 + |D3 φ|)h problem F convex no regularization of viscosity solutions controlling “third-derivatives” and “preserving” equation stochastic control representation, special schemes pde-switching systems Krylov Barles-Jakobsen 64 RATES OF CONVERGENCE – uniformly elliptic equations F(D2 u, Du, x) = 0 S([uh ]x , uh (x), x, h) = 0 consistency |S([φ]x , φ(x), x, h) − F(D2 φ(x), Dφ(x), x)| ≦ K(1 + |D3 φ|)h problem F convex no regularization of viscosity solutions controlling “third-derivatives” and “preserving” equation stochastic control representation, special schemes pde-switching systems F uniformly elliptic new regularity, δ-solutions Krylov Barles-Jakobsen Caffarelli-Souganidis 65 GENERAL STRATEGY (∗) 8 <F(D2 u) = f :u = g on in U F uniformly elliptic ∂U 66 GENERAL STRATEGY (∗) 8 <F(D2 u) = f :u = g • δ-viscosity solutions on in U F uniformly elliptic ∂U 67 GENERAL STRATEGY (∗) 8 <F(D2 u) = f :u = g • δ-viscosity solutions • new regularity result on in U F uniformly elliptic ∂U 68 GENERAL STRATEGY (∗) 8 <F(D2 u) = f :u = g on in U F uniformly elliptic ∂U • δ-viscosity solutions • new regularity result • Theorem A: u ∈ C0,1 (Ū) solves (∗), u± δ-sub- (super-) solution of (∗) and ku± − uk = O(δ η ) on ∂U, then there exists a universal θ > 0 st ku − u± k = O(δ θ ) in U. 69 GENERAL STRATEGY (∗) 8 <F(D2 u) = f :u = g on in U F uniformly elliptic ∂U • δ-viscosity solutions • new regularity result • Theorem A: u ∈ C0,1 (Ū) solves (∗), u± δ-sub- (super-) solution of (∗) and ku± − uk = O(δ η ) on ∂U, then there exists a universal θ > 0 • Theorem B: st ku − u± k = O(δ θ ) in U. numerical approximations are δ-solutions for δ = δ(h). 70 GENERAL STRATEGY (∗) 8 <F(D2 u) = f :u = g on in U F uniformly elliptic ∂U • δ-viscosity solutions • new regularity result • Theorem A: u ∈ C0,1 (Ū) solves (∗), u± δ-sub- (super-) solution of (∗) and ku± − uk = O(δ η ) on ∂U, then there exists a universal θ > 0 • Theorem B: st ku − u± k = O(δ θ ) in U. numerical approximations are δ-solutions for δ = δ(h). • Theorem C: oscillatory solutions are δ-solutions for δ = δ(ε) off a set of ω’s with probability less than δ. 71 δ-viscosity solution of F(D2 u) = f in U 72 δ-viscosity solution of F(D2 u) = f in U u viscosity subsolution iff ( for all x ∈ U and all quadratics P touching u from above at x, F(D2 P) ≦ f (x) P touches u from above at x u(y) ≦ u(x) + P(y − x) + o(|y − x|2 ) ≦ u(x) + (P + εI)(x − y) in B(x, δ(ε)) 73 δ-viscosity solution of F(D2 u) = f in U u viscosity subsolution iff ( for all x ∈ U and all quadratics P touching u from above at x, F(D2 P) ≦ f (x) P touches u from above at x u(y) ≦ u(x) + P(y − x) + o(|y − x|2 ) ≦ u(x) + (P + εI)(x − y) in B(x, δ(ε)) u δ-viscosity subsolution iff 8 > <for all B(x, δ) ⊂ U and all quadratics P such that u ≦ P in B(x, δ) , u(x) = P(x) and D2 P = O(δ −α ) for some α > 0 > : F(D2 P) ≦ f (x) 74 δ-viscosity solution of F(D2 u) = f in U u viscosity subsolution iff ( for all x ∈ U and all quadratics P touching u from above at x, F(D2 P) ≦ f (x) P touches u from above at x u(y) ≦ u(x) + P(y − x) + o(|y − x|2 ) ≦ u(x) + (P + εI)(x − y) in B(x, δ(ε)) u δ-viscosity subsolution iff 8 > <for all B(x, δ) ⊂ U and all quadratics P such that u ≦ P in B(x, δ) , u(x) = P(x) and D2 P = O(δ −α ) for some α > 0 > : F(D2 P) ≦ f (x) subsolutions are always δ-subsolutions δ-subsolutions are not always subsolutions 75 Lemma: Any monotone, consistent approximation uh of F(D2 u) = f is an h-solution of F(D2 w) = f ± Kh. 8 < all B(x, δ) ⊂ U and all quadratics P such that u δ-viscosity subsolution iff u ≦ P in B(x, δ), u(x) = P(x) and |D2 P| = O(δ −α ) (α > 0) : F(D2 P) ≦ f (x) 76 Lemma: Any monotone, consistent approximation uh of F(D2 u) = f is an h-solution of F(D2 w) = f ± Kh. 8 < all B(x, δ) ⊂ U and all quadratics P such that u δ-viscosity subsolution iff u ≦ P in B(x, δ), u(x) = P(x) and |D2 P| = O(δ −α ) (α > 0) : F(D2 P) ≦ f (x) Proof: uh ≦ Q monotonicity =⇒ S([Q]x , Q(x), x, h) ≦ S([uh ]x , uh (x), x, h) = 0 consistency =⇒ S([Q]x , Q(x), x, h) ≧ F(D2 Q) − f − Kh Theorem: kuh − uk = O(hα ) in B(x, δ), uh (x) = Q(x) α ∈ (0, 1) 77 HOMOGENIZATION x F(D2 uε , , ω) = 0 in U ε uε −→ u0 a.s. ε→0 and F0 (D2 u0 ) = 0 in U F uniformly elliptic, stationary ergodic 78 HOMOGENIZATION x F(D2 uε , , ω) = 0 in U ε uε −→ u0 a.s. ε→0 and F uniformly elliptic, stationary ergodic F0 (D2 u0 ) = 0 in U For each Q ∈ SN , F0 (Q) is the unique constant st • 8 <F(D2 uε , εx , ω) = F0 (Q) in B1 , then kuε (·, ω) − QkC(B̄1 ) → 0 a.s. if : uε = Q on ∂B1 if 8 ! <F(D2 uε , x, ω) = F0 (Q) in B1/ε , then kε2 uε (·, ω) − QkC(B̄1/ε ) → 0 a.s. : uε = Q on ∂B1/ε uε (x) = ε2 uε ( εx ) Caffarelli-Souganidis-Wang 79 • Lemma: strongly mixing media with algebraic rate ∃ Aε ⊂ Ω st P(Aε ) ≦ Ce−c| ln ε| 1/2 and kuε (·, ω) − QkC(B̄1 ) ≦ C(1 + kQk)e−c| ln ε| 1/2 in Acε 80 • Lemma: strongly mixing media with algebraic rate ∃ Aε ⊂ Ω st P(Aε ) ≦ Ce−c| ln ε| 1/2 and kuε (·, ω) − QkC(B̄1 ) ≦ C(1 + kQk)e−c| ln ε| • Lemma: If uε F(D2 uε , εx , ω) = 0 in U, is e−c| ln ε| 1/2 – solution off a set 1/2 in Acε then Aε ∈ Ω st P(Aε ) ≦ Ce−c| ln ε| 1/2 81 back to sup- and inf-convolutions some key properties of the regularizations of Lipschitz sub- and super-solutions 82 back to sup- and inf-convolutions some key properties of the regularizations of Lipschitz sub- and super-solutions F(D2 u, Du, u, x) = 0 graph ūε = {(x, y) ∈ RN+1 : d(x, y) = ε and y ≧ u(x)} ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] y(x) “maximizer” for x ūε (x) = u(y(x)) − (2ε) −1 F uniformly elliptic y(x) “maximizer” for x 2 |x − y(x)| d((x, ūε (x)), (y(x), u(y(x))) = ε 83 ∃ C > 0 (depending ONLY on ellipticity constants and N) st • |x1 − x2 | ≦ C|y(x1 ) − y(x2 )| (Jacobian of y 7→ y−1 (x) is bdd) • if a quadratic P touches ūε from above at x, then u is touched at y(x) from above by a quadratic Pε and 2 D ūε (x) ≧ D u(y(x)) + Cε2 |D2 u(y(x))|2 • ∃ t0 , σ st for t ≧ t0 2 ∃ Aεt st |Aεt | ≦ t−σ and uε has a second order expansion from above with error of size t in Aε,c t ūε has a second order expansion from below with error of size t in Aε,c t 84 > > u 85 > > u x > > D2 uε (x) ≧ − εc u 86 > > u x > > > y(x) u D2 uε (x) ≧ − εc 87 > > u x > > > y(x) u > D2 uε (x) ≧ − C ε 2 |D u(y(x))| ≦ C ε 88 > > u |A*t ,c | <= C|Atc | A*t x > > > y(x) u > At D2 uε (x) ≧ − C ε 2 |D u(y(x))| ≦ C ε 89 Sketch of proof 90 Sketch of proof • ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] = u(y(x)) − (2ε)−1 |x − y(x)|2 91 Sketch of proof • ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] = u(y(x)) − (2ε)−1 |x − y(x)|2 P • ūε touched from above by a quadratic P at x ⇒ x y(x) u−ε u P(z) − P(x) ≧ ūε (z) − ūε (x) ≧ u(z) − (2ε)−1 |z − y|2 − (u(y(x)) − (2ε)−1 |x − y(x)|2 ) ⇒ u(y) ≦ u(y(x)) − (2ε)−1 [|z − y|2 − |x − y(x)|2 ] ⇒ u is touched from above at y(x) by a quadratic of opening ε−1 92 Sketch of proof • ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] = u(y(x)) − (2ε)−1 |x − y(x)|2 P • ūε touched from above by a quadratic P at x ⇒ x y(x) u−ε u P(z) − P(x) ≧ ūε (z) − ūε (x) ≧ u(z) − (2ε)−1 |z − y|2 − (u(y(x)) − (2ε)−1 |x − y(x)|2 ) ⇒ u(y) ≦ u(y(x)) − (2ε)−1 [|z − y|2 − |x − y(x)|2 ] ⇒ u is touched from above at y(x) by a quadratic of opening ε−1 P • u solves u uniformly elliptic equation x y(x) u−ε u Harnack inequality =⇒ u is also touched from below at y(x) by a quadratic of opening c/ε 93 Sketch of proof • ūε (x) = sup[u(y) − (2ε)−1 |x − y|2 ] = u(y(x)) − (2ε)−1 |x − y(x)|2 P • ūε touched from above by a quadratic P at x ⇒ x y(x) u−ε u P(z) − P(x) ≧ ūε (z) − ūε (x) ≧ u(z) − (2ε)−1 |z − y|2 − (u(y(x)) − (2ε)−1 |x − y(x)|2 ) ⇒ u(y) ≦ u(y(x)) − (2ε)−1 [|z − y|2 − |x − y(x)|2 ] ⇒ u is touched from above at y(x) by a quadratic of opening ε−1 P • u solves u uniformly elliptic equation x y(x) u−ε u Harnack inequality =⇒ u is also touched from below at y(x) by a quadratic of opening c/ε u is differentiable at y(x) and has C1 -contact from above and below with convex and concave envelops of paraboloids with opening c/ε • x = y(x) − εDu(y(x)) =⇒ |Du(y(x1 )) − Du(y(x2 ))| ≦ cε−1 |y(x1 ) − y(x2 )| =⇒ |x1 − x2 | ≦ (1 + c)|y(x1 ) − y(x2 )| 94 (NEW) REGULARITY RESULT F(D2 u) = f in U 95 (NEW) REGULARITY RESULT F(D2 u) = f in U • F uniformly elliptic, u, f Lip, U = B1 =⇒ ∃ t0 , σ depending on ellipticity and N st for t ≧ t0 ∃ At ⊂ B1 st |(B1 \ At ) ∩ B1/2 | ≦ t−σ , and for all x0 ∈ At ∩ B1/2 ∃ quadratic Qt,x0 such that F(D2 Qt,x0 ) = f (x0 ), |D2 Qt,x0 | ≦ t, and u(x) = u(x0 ) + Qt,x0 (x − x0 ) + O(t|x − x0 |3 ) in B1 Caffarelli 96 (NEW) REGULARITY RESULT F(D2 u) = f in U • F uniformly elliptic, u, f Lip, U = B1 =⇒ ∃ t0 , σ depending on ellipticity and N st for t ≧ t0 ∃ At ⊂ B1 st |(B1 \ At ) ∩ B1/2 | ≦ t−σ , and for all x0 ∈ At ∩ B1/2 ∃ quadratic Qt,x0 such that F(D2 Qt,x0 ) = f (x0 ), |D2 Qt,x0 | ≦ t, and u(x) = u(x0 ) + Qt,x0 (x − x0 ) + O(t|x − x0 |3 ) in B1 Caffarelli •• F uniformly elliptic, u, f Lip, U = B1 u± ε sup, inf-convolution ∃ t0 , σ depending on ellipticity and N st for t ≧ t0 ∃ Aεt ⊂ B1 st |(B1 \ Aεt ) ∩ B1/2 | ≦ t−σ and for all x0 ∈ Aεt ∩ B1/2 ∃ quadratic Qεt,x0 ∈ SN such that F(D2 Qεt,x0 ) ≈ f (x0 ), |D2 Qεt,x0 | ≦ t and ε 3 u± ε (x) ≈ uε (x0 ) + Qt,x0 (x − x0 ) + O(t|x − x0 | ) in B1 97 Proof of regularity result F(D2 u) = f in B1 98 Proof of regularity result F(D2 u) = f in B1 tr Dx FD2 uxi = fxi in B1 ∃ “universal” t0 , σ st, for all t ≥ t0 , v = uxi is touched from above and below i,t in At ∩ B1/2 by quadratics Pi,t x0 and P̄x0 with opening t and C −σ |At | ≦ C(kDuk∞ + kDx f kN )t =⇒ i,t uxi differentiable in At ∩ B1/2 , Duxi (x0 ) = DPi,t x0 (x0 ) = DPx0 (x0 ) and |Du(x) − Du(x0 ) − D2 u(x0 )(x − x0 )| ≦ Ct|x − x0 |2 99 Sketch of proof of u solution, v δ-supersolution of F(D2 w) = f in U u ≦ v + O(δ γ ) on ∂U ff =⇒ u ≦ v + cδ θ • several approximations/regularizations u −→ u subsolution F(D2 w) = f − δ β 2 v −→ v δ-supersolution F(D v) = f D2 u ≧ −δ −2ζ I D2 v ≦ δ 2ζ I (θ > 0) 100 Sketch of proof of u solution, v δ-supersolution of F(D2 w) = f in U u ≦ v + O(δ γ ) on ∂U ff =⇒ u ≦ v + cδ θ • several approximations/regularizations u −→ u subsolution F(D2 w) = f − δ β 2 v −→ v δ-supersolution F(D v) = f • Γw concave envelope D2 u ≧ −δ −2ζ I D2 v ≦ δ 2ζ I w=u−v of |D2 Γw | ≦ cδ 2ζ on contact set ABP-estimate =⇒ sup w ≦ cδ −2ζ |{Γw = w}|1/N (θ > 0) 101 U • if δ θ ≦ sup w , Uδ (∗∗) {Γw =w} then δ (θ+2ζ)N ≦ c|{Γw = w}| 102 U • if δ θ ≦ sup w , Uδ (∗∗) then {Γw =w} δ (θ+2ζ)N ≦ c|{Γw = w}| • covering argument and (∗∗) =⇒ ∃ B(xi , δ γ ) st |B(xi , 21 δ γ ) ∩ {Γw = w}| ≧ cδ (θ+2ζ+γ)N B(xi, 2-1δγ) 103 U • if δ θ ≦ sup w , Uδ (∗∗) then {Γw =w} δ (θ+2ζ)N ≦ c|{Γw = w}| • covering argument and (∗∗) =⇒ ∃ B(xi , δ γ ) st |B(xi , 21 δ γ ) ∩ {Γw = w}| ≧ cδ (θ+2ζ+γ)N 1 • apply regularity result to B(xi , δ γ ) with t = δ − σ (θ+2ζ+γ)N ∃ x0 ∈ contact set ∩ B(xi , 2−1 δ γ ) and 2 |D Qt | ≦ t, 2 F(D Qt ) ≦ f − δ β quadratic Qt and u(x) = u(x0 ) + Qt (x − x0 ) + O(t|x − x0 |3 ) x0 At such that γ B(xi, 2-1δ ) 104 U • if δ θ ≦ sup w , Uδ (∗∗) then {Γw =w} δ (θ+2ζ)N ≦ c|{Γw = w}| • covering argument and (∗∗) =⇒ ∃ B(xi , δ γ ) st |B(xi , 21 δ γ ) ∩ {Γw = w}| ≧ cδ (θ+2ζ+γ)N 1 • apply regularity result to B(xi , δ γ ) with t = δ − σ (θ+2ζ+γ)N ∃ x0 ∈ contact set ∩ B(xi , 2−1 δ γ ) and 2 |D Qt | ≦ t, 2 F(D Qt ) ≦ f − δ β quadratic Qt and u(x) = u(x0 ) + Qt (x − x0 ) + O(t|x − x0 |3 ) • v δ-supersolution u≦ v+ℓ =⇒ ℓ linear F(D2 Qt − tδ) ≧ f (x0 ) x0 At such that γ B(xi, 2-1δ ) 105 U • if δ θ ≦ sup w , Uδ (∗∗) then {Γw =w} δ (θ+2ζ)N ≦ c|{Γw = w}| • covering argument and (∗∗) =⇒ ∃ B(xi , δ γ ) st |B(xi , 21 δ γ ) ∩ {Γw = w}| ≧ cδ (θ+2ζ+γ)N 1 • apply regularity result to B(xi , δ γ ) with t = δ − σ (θ+2ζ+γ)N ∃ x0 ∈ contact set ∩ B(xi , 2−1 δ γ ) and 2 |D Qt | ≦ t, 2 F(D Qt ) ≦ f − δ β quadratic Qt and u(x) = u(x0 ) + Qt (x − x0 ) + O(t|x − x0 |3 ) • v δ-supersolution u≦ v+ℓ =⇒ ℓ linear F(D2 Qt − tδ) ≧ f (x0 ) 1 • uniform ellipticity =⇒ δ β ≦ cδ 1− σ (θ+2ζ+γ)N a contradiction if σ − (θ + 2ζ + γ) > σβ x0 At such that γ B(xi, 2-1δ ) 106 HOMOGENIZATION 8 <F(D2 uε , εx , ω) = 0 in U : uε = g on ∂U F uniformly elliptic, stationary, ergodic 107 HOMOGENIZATION 8 <F(D2 uε , εx , ω) = 0 in U : uε = g on ∂U ∃ F0 uniformly elliptic st uε → u0 in C(Ū) ε→0 and a.s. F uniformly elliptic, stationary, ergodic 8 <F0 (D2 u0 ) = 0 in U : u0 = g on ∂U Caffarelli-Souganidis-Wang 108 HOMOGENIZATION 8 <F(D2 uε , εx , ω) = 0 in U : uε = g on ∂U ∃ F0 uniformly elliptic st uε → u0 in C(Ū) ε→0 rate of convergence and a.s. F uniformly elliptic, stationary, ergodic 8 <F0 (D2 u0 ) = 0 in U : u0 = g on ∂U kuε − u0 k = O(σ(ε)) Caffarelli-Souganidis-Wang 109 HOMOGENIZATION 8 <F(D2 uε , εx , ω) = 0 in U : uε = g on ∂U ∃ F0 uniformly elliptic st uε → u0 in C(Ū) and a.s. ε→0 rate of convergence F uniformly elliptic, stationary, ergodic 8 <F0 (D2 u0 ) = 0 in U Caffarelli-Souganidis-Wang : u0 = g on ∂U kuε − u0 k = O(σ(ε)) strongly mixing with algebraic rate* F linear σ(ε) = εγ Yurinskii * F strongly mixing with rate φ E[F(x, ·)F(y, ·) − EF(x, ·)EF(y, ·)] ≦ φ(|x − y|)(E(F(x, ·))2 E(F(y, ·))2 )1/2 110 HOMOGENIZATION 8 <F(D2 uε , εx , ω) = 0 in U : uε = g on ∂U ∃ F0 uniformly elliptic st uε → u0 in C(Ū) and a.s. ε→0 rate of convergence F uniformly elliptic, stationary, ergodic 8 <F0 (D2 u0 ) = 0 in U Caffarelli-Souganidis-Wang : u0 = g on ∂U kuε − u0 k = O(σ(ε)) strongly mixing with algebraic rate* F linear F nonlinear σ(ε) = εγ σ(ε) = e−c| ln ε| Yurinskii 1/2 Caffarelli-Souganidis * F strongly mixing with rate φ E[F(x, ·)F(y, ·) − EF(x, ·)EF(y, ·)] ≦ φ(|x − y|)(E(F(x, ·))2 E(F(y, ·))2 )1/2 111 • rate for quadratic data =⇒ rate for general data F(D2 uε , εx , ω) = 0 in U =⇒ uε is δ-subsolution of F0 (D2 w) = −δ α δ = Ce−| ln ε| 1/2 112 • rate for quadratic data =⇒ rate for general data F(D2 uε , εx , ω) = 0 in U =⇒ uε is δ-subsolution of F0 (D2 w) = −δ α δ = Ce−| ln ε| 1/2 • P quadratic st |D2 P| ≦ δ −σ , uε ≦ P in B(x0 , δ), uε (x0 ) = P(x0 ) P • assume uε F0 (D2 P) < −δ α x0 Bδ(x0) 113 • =⇒ rate for general data rate for quadratic data F(D2 uε , εx , ω) = 0 in U =⇒ uε is δ-subsolution of F0 (D2 w) = −δ α δ = Ce−| ln ε| 1/2 • P quadratic st |D2 P| ≦ δ −σ , uε ≦ P in B(x0 , δ), uε (x0 ) = P(x0 ) • P F0 (D2 P) < −δ α • assume “lower” P to Pδ (x) = P(x) − ηδ α (δ 2 − |x − x0 |2 ) uε (x0 ) = P(x0 ) ⇒ Pδ Pδ (x0 ) − uε (x0 ) = ηδ 2+α uniform ellipticity of F0 ⇒ F0 (D2 Pδ ) < 0 x0 Bδ(x0) uε 114 • =⇒ rate for general data rate for quadratic data F(D2 uε , εx , ω) = 0 in U =⇒ uε is δ-subsolution of F0 (D2 w) = −δ α δ = Ce−| ln ε| 1/2 • P quadratic st |D2 P| ≦ δ −σ , uε ≦ P in B(x0 , δ), uε (x0 ) = P(x0 ) • α uεδ 2 Pδ 2 “lower” P to Pδ (x) = P(x) − ηδ (δ − |x − x0 | ) uε (x0 ) = P(x0 ) ⇒ ⇒ F0 (D2 Pδ ) < 0 ( F(D2 uδε , εx , ω) = F0(D2 Pδ ) < 0 in B(x0 , δ) uδε = Pδ on ∂B(x0 , δ) uε Pδ (x0 ) − uε (x0 ) = ηδ 2+α uniform ellipticity of F0 • P F0 (D2 P) < −δ α • assume • ⇒ • x0 Bδ(x0) uδε ≧ uε kuδε −Pδ k ≦ Cδ 2 (1+δ −σ )e−c| ln(ε/δ)| 1/2 in Aεδ 115 • =⇒ rate for general data rate for quadratic data F(D2 uε , εx , ω) = 0 in U =⇒ uε is δ-subsolution of F0 (D2 w) = −δ α δ = Ce−| ln ε| 1/2 • P quadratic st |D2 P| ≦ δ −σ , uε ≦ P in B(x0 , δ), uε (x0 ) = P(x0 ) • α uεδ 2 Pδ 2 “lower” P to Pδ (x) = P(x) − ηδ (δ − |x − x0 | ) uε (x0 ) = P(x0 ) ⇒ uε Pδ (x0 ) − uε (x0 ) = ηδ 2+α uniform ellipticity of F0 • P F0 (D2 P) < −δ α • assume ⇒ F0 (D2 Pδ ) < 0 ( F(D2 uδε , εx , ω) = F0(D2 Pδ ) < 0 in B(x0 , δ) uδε = Pδ on ∂B(x0 , δ) • ⇒ • x0 Bδ(x0) uδε ≧ uε kuδε −Pδ k ≦ Cδ 2 (1+δ −σ )e−c| ln(ε/δ)| • 0 ≦ uδε (x0 ) − uε (x0 ) = uδε (x0 ) − P(x0 ) + P(x0 ) − uε (x0 ) ≦ Cδ 2 (1 + δ −σ )e−c| ln(ε/δ)| 1/2 − ηδ α+2 < 0 for δ = Ce−c| ln ε| 1/2 1/2 in Aεδ 116 • rate for quadratic data 8 <F(D2 uε , εx , ω) = F0 (D2 P) in Q1 : uε = P on ∂Q1 uε −→ P a.s. ε→0 8 <F0 (D2 u0 ) = F0 (D2 P) in Q1 :u = P on ∂Q 0 1 Q1 unit cube 117 • rate for quadratic data 8 <F(D2 uε , εx , ω) = F0 (D2 P) in Q1 : uε = P on ∂Q1 • estimate uε −→ P a.s. ε→0 uε − P 8 <F0 (D2 u0 ) = F0 (D2 P) in Q1 :u = P on ∂Q 0 1 Q1 unit cube 118 • rate for quadratic data 8 <F(D2 uε , εx , ω) = F0 (D2 P) in Q1 : uε = P on ∂Q1 • estimate or (better) uε −→ P a.s. ε→0 uε − P 8 <F0 (D2 u0 ) = F0 (D2 P) in Q1 :u = P on ∂Q 0 1 Q1 unit cube v± ε −P x (F(D2 P, εx , ω)−ℓ)+ in Q1 F(D2 v+ ε , ε , ω) = ℓ+1{v+ ε =P} obstacle problem with obstacle P from below x F(D2 v− (F(D2 P, εx , ω)−ℓ)− in Q1 ε , ε , ω) = ℓ−1{v− ε =P} obstacle problem with obstacle P from above − v+ ε = vε = P on ∂Q1 (ℓ = F0 (D2 P)) 119 • rate for quadratic data 8 <F(D2 uε , εx , ω) = F0 (D2 P) in Q1 : uε = P on ∂Q1 • estimate or (better) uε −→ P a.s. ε→0 uε − P 8 <F0 (D2 u0 ) = F0 (D2 P) in Q1 :u = P on ∂Q 0 1 Q1 unit cube v± ε −P x (F(D2 P, εx , ω)−ℓ)+ in Q1 F(D2 v+ ε , ε , ω) = ℓ+1{v+ ε =P} obstacle problem with obstacle P from below x F(D2 v− (F(D2 P, εx , ω)−ℓ)− in Q1 ε , ε , ω) = ℓ−1{v− ε =P} obstacle problem with obstacle P from above − v+ ε = vε = P on ∂Q1 • (ℓ = F0 (D2 P)) N sup(v± ε − P) is controlled by L -norm of rhs (total mass) 120 REVIEW OF KEY FACTS ABOUT UNIFORMLY ELLIPTIC PDE 121 REVIEW OF KEY FACTS ABOUT UNIFORMLY ELLIPTIC PDE • linearization F(D2 u, x) = 0 F(D2 v, x) = 0 =⇒ w=u−v −aij (x)wxi xj = 0 aij bdd meas 122 REVIEW OF KEY FACTS ABOUT UNIFORMLY ELLIPTIC PDE • linearization F(D2 u, x) = 0 F(D2 v, x) = 0 =⇒ w=u−v −aij (x)wxi xj = 0 aij bdd meas • Alexandrov-Bakelman-Pucci (ABP)-estimate −aij wxi xj = f in B1 =⇒ sup w+ ≦ sup w+ + Ckf+ kLN B1 ∂B1 (C universal) 123 REVIEW OF KEY FACTS ABOUT UNIFORMLY ELLIPTIC PDE • linearization F(D2 u, x) = 0 F(D2 v, x) = 0 =⇒ w=u−v −aij (x)wxi xj = 0 aij bdd meas • Alexandrov-Bakelman-Pucci (ABP)-estimate −aij wxi xj = f in B1 =⇒ sup w+ ≦ sup w+ + Ckf+ kLN B1 (C universal) ∂B1 • Fabes-Stroock (FS)-estimate ( −aij wxi xj = f in B1 0<f <1 =⇒ w|B1/2 ≧ Ckf kLM w = 0 on ∂B1 (C, M universal, M large) • obstacle problem u smallest st • u = Q on ∂B1 8 2 <F(D2 u, x) ≧ 0 in B1=⇒ • F(D u, x) = 0 in {u > Q} • 0 ≦ (u−Q)(y) ≦ C|x − y|2 (u(x) = Q(x)) :u ≧ Q in B1 • F(D2 u, x) = 1{u=Q} F(Q, x)+ 124 • rate for quadratic data P – wlog F0 (D2 P) = 0 v± ε solution of obstacle problem from above (below) obstacle P uε solution 125 • rate for quadratic data P – wlog F0 (D2 P) = 0 v± ε solution of obstacle problem from above (below) obstacle P uε solution • − enough to find rate for kv+ ε − Pk and kvε − Pk ± kv± ε − Pk ≦ Chε (ω) R y 2 N 1/N Λ± h± ε contact set ε (ω) = [ Q ∩Λ± [F(D P, ε , ω)± ] dy] 1 ε 126 • rate for quadratic data P – wlog F0 (D2 P) = 0 v± ε solution of obstacle problem from above (below) obstacle P uε solution • − enough to find rate for kv+ ε − Pk and kvε − Pk • ± kv± ε − Pk ≦ Chε (ω) R y 2 N 1/N Λ± h± ε contact set ε (ω) = [ Q ∩Λ± [F(D P, ε , ω)± ] dy] 1 ε estimate − 2 2 Hk = E(h+ k ) E(hk ) (ε = 3−k ) 127 KEY STEP Assume h± k (ω) ≥ θ > 0 FS k ± N + − N − ⇒ v+ F(D2 v± k , 3 x, ω) = (hk ) ⇒ −aij (vk − vk ) ≧ 2θ k − vk |Q monotonicity of obstacle problem ⇒ 1/2 ≧ 2θMN ± ± ± Λ± k+ℓ ⊂ Λ ⇒ if vk+ℓ = P ⇒ vk = P k - scale each cell of k + ℓ-scale has diameter 3−k−ℓ − MN v+ − P + P − v− = v+ k − vk ≧ 2θ |k {z } | {z k} ≧0 ≧0 either must have if k + l scale MN v+ k −P ≧ θ Λ+ k+ℓ ∩ B1/2 6= ∅ 6⇒ or MN P − v− k ≧ θ 3−2(k+ℓ) C ≧ θMN impossible for ℓ large! • need to go to k + ℓ instead of k + 1 leads to slow rate