University of Connecticut Mathematics A gradient bound and a Liouville theorem for nonlinear Poisson equations Mingfeng Zhao 1 Email: mingfeng.zhao@uconn.edu May 22, 2011 1 Department of Mathematics, University of Connecticut, 196 Auditorium Road, Unit 3009, Storrs, CT 06269-3009 Contents 1. Main ideas 1 2. A useful lemma 2 3. A gradient estimates for Poisson’s equation 3 4. Important auxiliary functions 6 5. Best Gradient Estimate 10 6. Liouville Type Theorem 20 References 21 In this note, we will give a detailed explanation of Modica’s paper[3]. 1. Main ideas (1) ∆x u(x) = F 0 (x) is translation invariant for the statement of the Theorem; (2) Since u is bounded in Rn , then inf n |Du(x)| = 0; x∈R (3) Use the standard gradient estimate for the Poisson equation, we know the function P (x) = |∇u(x)|2 − 2F (u(x)) is bounded in Rn ; (4) Construct a special function η,ρ0 satisfying some special properties; (5) Let v(x) = η,ρ0 (x)P (x) for |x| ≥ ρ0 , we have v(x) ≤ max {, max P (x)}, |x|=ρ0 for all |x| ≥ ρ0 . (6) Consider |∇u(x)|2 for the extremum point of v(x), both bolds for the previous step 5; (7) Take & 0, we obtain P (x) ≤ max 0, max P (x) , |x|=ρ0 for all |x| ≥ ρ0 . And take ρ0 & 0, we get P (x) ≤ max {0, P (0)} ≤ |Du(0)|2 < δ. (8) By the standard Taylor’s expansion when considering the minimum point of u, and use Gronwall’s inequality to show the u(x) ≡ constant in Rn . –1– 2. A useful lemma Theorem 2.1. For any u ∈ C 1 (Rn ) ∩ L∞ (Rn ), then inf x∈Rn |Du(x)| = 0. Proof. If the above statement is not true, then there exists some u ∈ C 1 (Rn ) inf x∈Rn Let v(x) = T L∞ (Rn ) such that |Du(x)| = δ0 > 0. u(x) , then v ∈ C 1 (Rn ) ∩ L∞ (Rn ), and δ0 for all x ∈ Rn . |Dv(x)| ≥ 1, Consider the ODE: (1) ẋ(t) = Dv(x(t), x(0) = 0. t>0 Since Dv ∈ C(Rn ), then by the ODE’s Theory, we know the ODE (1) has a solution x(t). Now consider the function: f (t) = v(x(t)), t ≥ 0. Since v ∈ C 1 (Rn ), and x(t) ∈ C 1 ((0, ∞)), then we have f 0 (t) = Dv(x(t)) · ẋ(t) = Dv(x(t) · Dv(x(t)) = |Dv(x(t))|2 ≥ 1, for all t > 0. Since f 0 is continuous, then either f 0 (t) ≥ 1 for all t > 0; or f 0 (t) ≤ −1 for all t > 0. Without loss of generality, we assume f 0 (t) ≥ 1 for all t > 0. Then by integrating both sides, we have f (t) − f (0) ≥ t. So v(x(t)) ≥ f (0) − t = v(x(0)) − t = v(0) − t for all t > 0. When take t → ∞, we know v(x(t)) → ∞, as t → ∞. In particular, we know v(x) is unbounded in Rn . Since v(x) = –2– u(x) δ0 , then u(x) is unbounded in Rn ., contradiction. Remark 2.1. Let u ∈ C 1,α (Rn ) for some 0 < α ≤ 1, then infn |Du(x)| = 0. In fact, let M = sup u(x), then x∈R x∈Rn n k k n there exists some sequence {xk }∞ k=1 ⊂ R such that u(x ) → M , as k → ∞. Let vk (x) = u(x + x ) for all x ∈ R , then vk ∈ C 1,α (Rn ), and kvk kC 1,α (Rn ) = kukC 1,α (Rn ) for all k ≥ 1. By the Ascoli-Arzela’s Theorem, there exists some 1 v ∈ Cloc (Rn ), vk (x) ≤ v(0) for all x ∈ Rn and all k ≥ 1, and some subsequence of {vk }∞ k=1 , without loss of generality, 1 n n we assume {vk }∞ k=1 itself such that vk → v in Cloc (R ), as k → ∞. In particular, we get v(x) ≤ v(0) for all x ∈ R , 1 which implies that Dv(0) = 0. Since vk → v in Cloc (Rn ), as k → ∞, then Dvk (0) = Du(xk ) → 0 = Dv(0), as k → ∞. Therefore, we can conclude that infn |Du(x)| = 0. x∈R 3. A gradient estimates for Poisson’s equation Theorem 3.1 (Theorem 3.9, in Page 41, in Gilbarg and Trudinger [1]). For any u ∈ C 2 (Rn ), f ∈ C(Rn ) and ∆x u(x) = f (x), for all x ∈ Rn . Then for any x0 ∈ Rn , any d > 0, we have |Du(x0 )| ≤ n d sup |u(x)| + sup |f (x)|. d x∈∂Qd (x0 ) 2 x∈Qd (x0 ) Where Qd (x0 ) denote the cube with center x0 , and length 2d, that is n Qd (x0 ) = x ∈ R : max |xi − (x0 )i | < d . 1≤i≤n Proof. Claim I: For any u ∈ C 2 (Rn ), f ∈ C(Rn ) and ∆x u(x) = f (x), for all x ∈ Rn . Then for d > 0, we have |Du(0)| ≤ n d sup |u(x)| + sup |f (x)|. d x∈∂Qd (0) 2 x∈Qd (0) In fact, for any d > 0, 1 ≤ k ≤ n, we consider the k-th upper half cub n Qk+ (0) = x ∈ R : max |x | < d, 0 < x < d . i k d 1≤i≤n Then for any x ∈ Qk+ d (0), we know |xi | < d for all i = 1, · · · , k − 1, k + 1, · · · , n, and 0 < xk < d. Then −d < −xk < 0, so (x1 , · · · , −xk , · · · , xn ) ∈ Qd (0). And n ∂Qk+ d (0) = {x ∈ R : |xi | = k for some 1 ≤ i ≤ n, i 6= k, 0 ≤ xk ≤ d} [ {x ∈ Rn : |xi | ≤ d, xk = 0.} Consider the function v(x) = 1 [u(x1 , · · · , xk , · · · , xn ) − u(x1 , · · · , −xk , · · · , xn )], 2 for all x ∈ Qk+ d (0). Let M = supx∈∂Qd (0) |u(x)|, and N = supx∈Qd (0) |f (x)|. Then we know • v(x1 , · · · , xk−1 , 0, xk+1 , · · · , xn ) = 0 for all (x1 , · · · , xk−1 , 0, xk+1 , · · · , xn ) ∈ ∂Qk+ d (0); –3– • For all x ∈ ∂Qk+ d (0), we have |v(x)| = ≤ ≤ = 1 [u(x1 , · · · , xk , · · · , xn ) − u(x1 , · · · , −xk , · · · , xn )] 2 1 [|u(x1 , · · · , xk , · · · , xn )| + |u(x1 , · · · , −xk , · · · , xn )|] 2 1 [2M ] 2 M So we have |v(x)| ≤ M. sup x∈∂Qk+ d (0) • For any x ∈ Qk+ d (0), we have ∆x v(x) 1 [∆x u(x1 , · · · , xk , · · · , xn ) − ∆x u(x1 , · · · , −xk , · · · , xn )] 2 1 [f (x1 , · · · , xk , · · · , xn ) − f (x1 , · · · , −xk , · · · , xn )]. 2 = = So for any x ∈ Qk+ d (0), we get |∆x v(x)| ≤ ≤ = Consider another function M w(x) = 2 d 1 [2N ] 2 1 [|f (x1 , · · · , xk , · · · , xn )| + |f (x1 , · · · , −xk , · · · , xn )|] 2 N. X |xi |2 + xk (nd − (n − 1)xk ) + 1≤i≤n,i6=k N xk (d − xk ) 2 for all x ∈ Qk+ d (0). So we have • For all (x1 , · · · , xk−1 , 0, xk+1 , · · · , xn ) ∈ ∂Qk+ d (0), we have w(x1 , · · · , xk−1 , 0, xk+1 , · · · , xn ) = M d2 X |xi |2 1≤i≤n,i6=k ≥ 0. • For all x such that |xj | = d for some 1 ≤ j ≤ n, j 6= k, and 0 ≤ xk ≤ d, then we know P 1≤i≤n,i6=k |xi |2 ≥ |xj |2 = d2 . Since 0 ≤ xk ≤ d, then we know xk (nd − (n − 1)xk ) ≥ 0, and xk (d − xk ) ≥ 0. Hence we have w(x) ≥ M 2 ·d d2 = M. –4– • For any 1 ≤ i ≤ n, i 6= k, x ∈ Qk+ d (0), we have 2M ∂ 2 w(x) = 2 2 ∂xi d For any x ∈ Qk+ d (0), we have ∂ 2 w(x) ∂x2k = M N · (−2(n − 1)) − ·2 d2 2 = − 2(n − 1)M − N. d2 Therefore, for any x ∈ Qk+ d (0), we have ∆x w(x) = (n − 1) · 2(n − 1)M 2M − −N d2 d2 = −N. By the properties of v(x) and w(x), we know ∆x (w ± v)(x) ≤ 0, in Qk+ (0) d w(x) ± v(x) ≥ 0, on ∂Qk+ d (0). Then by the maximum principle, we know for all x ∈ Qk+ d (0), we have w(x) ± v(x) ≥ 0. That is, |v(x)| ≤ w(x) for all x ∈ Qk+ d (0). Now take x1 = · · · = xk−1 = xk+1 = · · · = xn = 0, and xk & 0, we have ∂u(0) ∂xk = = ≤ = = = lim xk &0 u(0, · · · , xk , · · · , 0) − u(0, · · · , −xk , · · · , 0) 2xk lim v(0, · · · , xk , · · · , 0) xk lim w(0, · · · , xk , · · · , 0) xk lim M N [xk (nd − (n − 1)xk )] + xk (d − xk ) d2 xk 2xk lim M N (nd − (n − 1)xk ) + (d − xk ) 2 d 2 xk &0 xk &0 xk &0 xk &0 d n M + N. d 2 By the definition of M and N , we get ∂u(0) n d ≤ sup |u(x)| + sup |f (x)|. ∂xk d x∈∂Qd (0 2 x∈Qd (0 Since 1 ≤ k ≤ n is arbitrary, then we get |Du(0)| ≤ n d sup |u(x)| + sup |f (x)|. d x∈∂Qd (0) 2 x∈Qd (0) –5– Claim II: For any u ∈ C 2 (Rn ), f ∈ C(Rn ) and ∆x u(x) = f (x), for all x ∈ Rn . Then for any x0 ∈ Rn , any d > 0, we have |Du(x0 )| ≤ n d sup |u(x)| + sup |f (x)|. d x∈∂Qd (x0 ) 2 x∈Qd (x0 ) In fact, for any x0 , we consider v(x) = u(x0 + x), for all x ∈ Rn . Then ∆x v(x) = f (x + x0 ). And by the result of the Claim I, we get for and d > 0, we have |Dv(0)| ≤ n d sup |v(x)| + sup |f (x + x0 )| d x∈∂Qd (0) 2 x∈Qd (0) = d n sup |u(x)| + sup |f (x)| d x∈∂Qd (x0 ) 2 x∈Qd (x0 ) Since Dv(0) = Du(x0 ), so we get |Du(x0 )| ≤ n d sup |u(x)| + sup |f (x)|. d x∈∂Qd (x0 ) 2 x∈Qd (x0 ) Remark 3.1. The key point of the proof is to find these two auxiliary functions v(x) and w(x). 4. Important auxiliary functions Theorem 4.1. Let M > 0, N > 0 be any positive numbers and fix them, and integer n ≥ 1. Then for any α > 0, β > 0, there exists some C 2 functions: ηα,β : [β, ∞) −→ R. with the following good properties: i. For any α > 0, β > 0, we have ηα,β (β) = 1. ii. For any α > 0, β > 0, then for any t ≥ β, we have ηα,β (t) > 0, and 0 ηα,β (t) < 0. In particular, we know for all t ≥ β, we have ηα,β (t) ≤ 1. iii For any α > 0, β > 0, we have lim ηα,β (t) = 0. t%∞ iv For any α > 0, β > 0, then for all t ≥ β, we have lim ηα,β (t) = 1. α&0 –6– v For any α > 0, β > 0, then for all t ≥ β, we have h i2 !2 0 0 2 η (t) (n − 1)ηα,β (t) α,β α ηα,β (t) M 0 00 · − η (t) − ηα,β (t) − = . 0 (t) ηα,β ηα,β (t) α α,β t N Proof. For any α > 0, we know Z 0 1 α e− N s ds = s2 ∞ Z αt t2 e− N 1 ∞ Z 1 dt t2 Let t = 1 s αt e− N dt = 1 ∞ N − αt e N α = − = N −α e N α 1 Since α > 0, N > 0 < ∞. So for any α > 0, and for all 0 ≤ t ≤ 1, we can define Z 1 −α e Ns gα (t) = ds s2 t Z 1t αr 1 r2 e− N 2 dr = r 1 Z 1t αr = e− N dr Let r = 1 s 1 1 = = > N αt t − e− N α 1 α N −α · e N − e− N t α 0. So we have α gα0 (t) = − e− N t < 0, t2 for all 0 ≤ t ≤ 1. α −N That means that gα is strictly decreasing, so gα has inverse function gα−1 which is also strictly decreasing on 0, N , αe and (gα−1 )0 (t) < 0, for all 0 ≤< For any α > 0, β > 0, then for any ∞ ≥ t ≥ β, we define Z hα,β (t) = t β Then we know h0α,β (t) = Mt e− α tn−1 Ms e− α ds. sn−1 > 0 for all t ≥ β, and hα,β (∞) = ∞ Ms e− α ds sn−1 β –7– Z N −α e N. α = = = < Ms ∞ e− α ds β n−1 β Z Ms 1 e− α ds β n−1 β α Ms 1 · − · e− α β n−1 M Mβ α e− α n−1 Mβ Z ≤ ∞ β ∞. So 0 < hα,β (∞) < ∞. And for any t ≥ β, we have ≤ 0 = < gα (0) · hα,β (t) hα,β (∞) α N −N αe hα,β (∞) · hα,β (t) N −α e N. α So for all t ≥ β, we can define βα,β (t) = gα−1 gα (0) · hα,β (t) . hα,β (∞) Claim: ηα,β (t) satisfy our conditions. i. For any α > 0, β > 0, we have ηα,β (β) = gα−1 gα (0) · hα,β (β) hα,β (∞) = gα−1 (0) = 1. ii. For any α > 0, β > 0, then for any t ≥ β, since the image of gα−1 is (0, 1], then we have gα (0) βα,β (t) = gα−1 · hα,β (t) > 0. hα,β (∞) And also we have 0 ηα,β (t) 0 gα (0) · hα,β (β) hα,β (∞) 0 gα (0) gα−1 · hα,β (β) · hα,β (∞) 0 gα (0) gα−1 · hα,β (β) · hα,β (∞) = = = gα−1 gα (0) · h0 (β) hα,β (∞) α,β < 0. So we know ηα,β (t) is strictly decreasing, so for all t ≥ β, we have ηα,β (t) ≤ 1. –8– Mt gα (0) e− α · n−1 hα,β (∞) t iii. For any α > 0, β > 0, then lim ηα,β (t) = ηα,β (∞) t%∞ = gα−1 gα (0) · hα,β (∞) hα,β (∞) = gα−1 (gα (0)) = 0. iv. For any α > 0, β > 0, then for any t ≥ β, since Z lim hα,β (t) α&0 = α&0 t Z = β Ms e− α ds sn−1 Ms e− α ds By the Monotone Convergence Theorem α&0 sn−1 lim β = t lim 0. Because gα−1 (0) = 1 for all α > 0, and by the continuity of gα−1 , we know lim ηα,β (t) = 1. α&0 v. For any α > 0, β > 0, then for any t ≥ β, since βα,β (t) = gα−1 gα (0) · hα,β (t) . hα,β (∞) Then gα (ηα,η (t)) = gα (0) · hα,β (t). hα,β (∞) That is, for any t ≥ β, we have Z α 1 ηα,β (t) gα (0) e− N s ds = · s2 hα,β (∞) Z t β Ms e− α ds. sn−1 Differentiating on both sides, we can get − α Mt e N ηα,β (t) gα (0) e− α 0 − · η (t) = · . α,β [ηα,β (t)]2 hα,β (∞) tn−1 0 Since ηα,β (t) < 0, and ηα,β (t) > 0, then we can take the logarithm on both sides, then we have 0 ln(−ηα,β (t)) − α N ηα,β (t) − 2 ln ηα,β (t) = ln gα (0) hα,β (∞) − Mt − (n − 1) ln t. α Differentiating on both sides, we have 0 0 00 −ηα,β (t) αηα,β (t) 2ηα,β (t) M n−1 + − =− − . 0 −ηα,β (t) N [ηα,β (t)]2 [ηα,β (t)]2 α t –9– Both sides are multiplied by α = N ηα,β (t) 0 (t) ηα,β !2 [ηα,β (t)]2 0 (t) , ηα,β , then for all t ≥ β, we get h i2 0 0 2 η (t) (n − 1)ηα,β (t) α,β M 0 00 · − η (t) − ηα,β (t) − . ηα,β (t) α α,β t 5. Best Gradient Estimate Theorem 5.1. Let F ∈ C 2 (Rn ), and F (t) ≥ 0 for all t ∈ R. For any u ∈ C 3 (Rn ) T the PDEs: ∆x u(x) = F 0 (u(x)), for all x ∈ Rn . Then |Du(x)|2 ≤ 2F (u(x)), for all x ∈ Rn . Proof. a. If u(x) ≡ C in Rn , then Du(x) = 0. But F (u(x)) ≥ 0, then we have 0 = |Du(x)|2 ≤ 2F (u(x)), b. u(x) 6≡ Constant in Rn . Since u ∈ C 3 (Rn ) T for all x ∈ Rn . L∞ (Rn ), by Theorem 2.1, we know inf x∈Rn |Du(x)| = 0. So for any δ > 0, there exists some x0 ∈ Rn such that |Du(x0 )|2 < δ. Now let v(x) = u(x + x0 ), so ∆x v(x) = F 0 (v(x)) in Rn , and we have |Dv(0)|2 < δ. Define the P -function of the PDE ∆v = F 0 (v) as: P (x) = |Dv(x)|2 − 2F (v(x)), Since u ∈ C 3 (Rn ) T for all x ∈ Rn . L∞ (Rn ), so there exists some A > 0 such that kukL∞ (Rn ) ≤ A. So for all x ∈ Rn , we have |F 0 (u(x))| ≤ kF 0 kL∞ ([−A,A]) = B. – 10 – L∞ (Rn ) which is a solution of So by Theorem 3.1, we just take d = 1, we can get for all x ∈ Rn , we have |Du(x)| ≤ n |u(x)| + sup y∈∂Q1 (x) ≤ nA + 1 sup |F 0 (y)| 2 y∈Q1 (x) B . 2 So for all x ∈ Rn , we have |P (x)| ≤ nA + B +B 2 = C. Since u ∈ C 3 (Rn ), and F ∈ C 2 (R), then we know P (x) ∈ C 2 (Rn ), and for all 1 ≤ i ≤ n, all x ∈ Rn , we have ∂P (x) ∂xi = ∂ 2 P (x) ∂xi ∂xi = = n X ∂v(x) ∂ 2 v(x) ∂v(x) − 2F 0 (v(x)) · ∂x ∂x ∂x ∂xi j i j j=1 " # 2 2 n X ∂ 2 v(x) ∂v(x) ∂ 3 v(x) ∂v(x) ∂ 2 v(x) 00 2 + · − 2F (v(x)) − 2F 0 (v(x)) ∂xi ∂xj ∂xj ∂xi ∂xi ∂xj ∂xi ∂xi ∂xi j=1 2 2 · 2 n 2 X ∂ v(x) j=1 ∂xi ∂xj −2F 00 (v(x)) · +2 n X ∂v(x) j=1 ∂v(x) ∂xi 2 ∂xj · ∂ 3 v(x) ∂xi ∂xi ∂xj − 2F 0 (v(x)) · ∂ 2 v(x) . ∂xi ∂xi Hence, we have ∆x P (x) = 2 2 n X n 2 X ∂ v(x) i=1 j=1 00 ∂xi ∂xj −2F (v(x)) · +2 n X n X ∂v(x) 2 n X ∂v(x) i=1 ∂xj i=1 j=1 ∂xi · ∂ 3 v(x) ∂xi ∂xi ∂xj − 2F 0 (v(x)) · n X ∂ 2 v(x) ∂xi ∂xi i=1 2 n 2 n n X X ∂ v(x) ∂v(x) X ∂ 3 v(x) = 2 +2 · − 2F 00 (v(x))|∇v(x)|2 ∂x ∂x ∂x ∂x ∂x ∂x i j j i i j i,j=1 j=1 i=1 −2F 0 (v(x)) · ∆x v(x) 2 n n 2 X X ∂v(x) ∂∆x v(x) ∂ v(x) +2 · − 2F 00 (v(x))|∇v(x)|2 − 2F 0 (v(x)) · F 0 (v(x)) = 2 ∂x ∂x ∂x ∂x i j j j j=1 i,j=1 = 2 2 n 2 n X X ∂ v(x) ∂v(x) ∂F 0 (v(x)) +2 · − 2F 00 (v(x))|∇v(x)|2 − 2[F 0 (v(x))]2 ∂x ∂x ∂x ∂x i j j j i,j=1 j=1 = 2 2 n 2 n X X ∂ v(x) ∂v(x) ∂v(x) +2 · F 00 (v(x)) · − 2F 00 (v(x))|∇v(x)|2 − 2[F 0 (v(x))]2 ∂x ∂x ∂x ∂x i j j j i,j=1 j=1 2 n 2 X ∂ v(x) = 2 + 2F 00 (v(x))|∇v(x)|2 − 2F 00 (v(x))|∇v(x)|2 − 2[F 0 (v(x))]2 ∂x ∂x i j i,j=1 2 n 2 X ∂ v(x) = 2 − 2[F 0 (v(x))]2 . ∂x ∂x i j i,j=1 – 11 – On the other hand, we know for any 1 ≤ i ≤ n, we have ∂P (x) ∂v(x) + 2F 0 (v(x)) · ∂xi ∂xi = 2 n X ∂v(x) j=1 ∂xj · ∂ 2 v(x) ∂xi ∂xj So we get ∂P (x) ∂v(x) + 2F 0 (v(x)) · ∂xi ∂xi ≤ 2 n X ∂ 2 v(x) ∂v(x) · ∂xj ∂xi ∂xj i=1 " ≤ 2 2 n X ∂v(x) i=1 = ∂xj = · 2 n 2 X ∂ v(x) i=1 # 12 ∂xi ∂xj By Cauchy-Schwarz’s inequality #1 " n X ∂ 2 v(x) 2 2 2|∇v(x)| ∂xi ∂xj i=1 = # 12 " 2|∇v(x)| √ ∆x P (x) + 2[F 0 (v(x))]2 2 21 1 2|∇v(x)|[∆x P (x) + 2[F 0 (v(x))]2 ] 2 Hence, we get 2|∇v(x)|2 [∆x P (x) + 2[F 0 (v(x))]2 ] = 2|∇v(x)|2 ∆x P (x) + 4|∇v(x)|2 · [F 0 (v(x))]2 n 2 X ∂v(x) ∂P (x) + 2F 0 (v(x)) · ∂xi ∂xi i=1 n n X ∂P (x) 2 X ∂P (x) ∂v(x) 0 = + 2F (v(x)) · · ∂x ∂xi ∂xi i i=1 i=1 2 n X ∂v(x) +4[F 0 (v(x))]2 · ∂xi i=1 ≥ = |∇P (x)|2 + 2F 0 (v(x))(∇v(x) · ∇P (x)) + 4[F 0 (v(x))]2 · |∇v(x)|2 Therefore, we have (2) |∇v(x)|2 ∆x P (x) ≥ 1 |∇P (x)|2 + 2F 0 (v(x))(∇v(x) · ∇P (x)). 2 Let M = supx∈Rn 2|F 0 (v(x))||∇v(x)|, and N = supx∈Rn 2|∇v(x)|2 . So we know 0 ≤ M, N < ∞. If N = 0, so for all x ∈ Rn , we have Dv(x) = 0. This contradicts with v(x) 6≡ constant in Rn . So we have N > 0. If M = 0, so for all x ∈ Rn , we have |F 0 (v(x))||∇v(x)| ≡ 0. That is, we have ∇F (v(x)) ≡ 0. So F (v(x)) ≡ Constant for all x ∈ Rn . So ∆x v(x) = 0, that is v(x) is a bounded harmonic function, so v(x) is constant, contradiction. Hence, we have M > 0. – 12 – Hence, by Theorem 4.1, for M, N > 0 and fix M and N , then for any α > 0, β > 0, there exists some C 2 functions: ηα,β : [β, ∞) −→ R. with the following good properties: i. For any α > 0, β > 0, we have ηα,β (β) = 1. ii. For any α > 0, β > 0, then for any t ≥ β, we have 0 and ηα,β (t) < 0. ηα,β (t) > 0, In particular, we know for all t ≥ β, we have ηα,β (t) ≤ 1. iii For any α > 0, β > 0, we have lim ηα,β (t) = 0. t%∞ iv For any α > 0, β > 0, then for all t ≥ β, we have lim ηα,β (t) = 1. α&0 v For any α > 0, β > 0, then for all t ≥ β, we have h i2 !2 0 0 (t) 2 η (n − 1)ηα,β (t) α,β ηα,β (t) M 0 α 00 · − η (t) − ηα,β (t) − = . 0 (t) ηα,β ηα,β (t) α α,β t N Now consider wα,β (x) = ηα,β (|x|)P (x), for all x ∈ Rn . Claim: For all aα > 0, β, then for all |x| ≥ β, we have wα,β (x) ≤ max α, max wα,β (x) = max α, max P (x) . |x|=β |x|=β When |x| = β, we know wα,β (x) = ηα,β (|x|)P (x) = ηα,β (β)P (x) = P (x) Since ηα,β (β) = 1. Hence we have max α, max wα,β (x) |x|=β = max α, max P (x) . |x|=β If sup|x|≥β wα,β (x) ≤ 0, since max α, max P (x) ≥ α > 0. |x|=β – 13 – So obviously, for all |x| ≥ β, we get wα,β (x) ≤ 0 ≤ max α, max P (x) . |x|=β If Tα,β = sup|x|≥β wα,β (x) > 0, then there exists some x1 ∈ Rn such that wα,β (x1 ) > 0 By the property iii of ηα,β (t), we know lim ηα,β (|x|) = 0. |x|%∞ By the definition of wα,β (x), and kP kL∞ (Rn ) ≤ C, then we get lim |x|%∞ wα,β (x) = 0. Since wα,β (x1 ) > 0, so the maximum of wα,β (x) is attained in {x ∈ Rn : |x| ≥ β}. For any x ∈ {x ∈ Rn : |x| = β} such that wα,β (x) = T , then we know wα,β (x) ≤ Tα,β ≤ max α, max wα,β (x) = max α, max P (x) . |x|=β |x|=β On the other hand, for any x∗ ∈ {x ∈ Rn : |x| > β} such that wα,β (x∗ ) = Tα,β , then we know x∗ is an interior maximum point of wα,β (x) in {x ∈ Rn : |x| ≥ β}. So have ∇wα,β (x∗ ) = 0, and D2 wα,β (x∗ ) ≤ 0. And also, by the definition of wα,β (x) = ηα,β (|x|)P (x), then we have 0 = ∇wα,β (x∗ ) = P (x∗ )∇ηα,β (|x∗ |) + ηα,β (|x∗ |)∇P (x∗ ). Since ηα,β (|x∗ |) > 0, so we get ∇P (x∗ ) = −P (x∗ ) · ∇ηα,β (|x∗ |) . ηα,β (|x∗ |) Since wα,β (x) = ηα,β (|x|)P (x), then we have ∆x wα,β (x) = P (x)∆x ηα,β (|x|) + 2∇P (x) · ∇ηα,β (|x|) + ηα,β (|x|)∆x P (x). So we can get |∇v(x∗ )|2 ∆x wα,β (x∗ ) = |∇v(x∗ )|2 [P (x∗ )∆x ηα,β (|x∗ |) + 2∇P (x∗ ) · ∇ηα,β (|x∗ |) + ηα,β (|x∗ |)∆x P (x∗ )] = |∇v(x∗ )|2 P (x)∆x ηα,β (|x∗ |) + 2|∇v(x∗ )|2 (∇P (x∗ ) · ∇ηα,β (|x∗ |)) +ηα,β (|x∗ |)|∇v(x∗ )|2 ∆x P (x∗ ) ≥ |∇v(x∗ )|2 P (x∗ )∆x ηα,β (|x∗ |) + 2|∇v(x∗ )|2 (∇P (x∗ ) · ∇ηα,β (|x∗ |)) – 14 – 1 |∇P (x∗ )|2 + 2F 0 (v(x∗ ))(∇v(x∗ ) · ∇P (x∗ )) By (2) 2 ∇ηα,β (|x∗ |) ∗ ∗ ∗ 2 ∗ ∗ 2 ∗ · ∇ηα,β (|x |) = P (x )|∇v(x )| ∆x ηα,β (|x |) + 2|∇v(x )| −P (x ) · ηα,β (|x∗ |) " 2 ∇ηα,β (|x∗ |) 1 ∗ + 2F 0 (v(x∗ ))ηα,β (|x∗ |)· −P (x∗ ) · +ηα,β (|x |) 2 ηα,β (|x∗ |) ∇ηα,β (|x∗ |) ∇v(x∗ ) · −P (x∗ ) · ηα,β (|x∗ |) +ηα,β (|x∗ |) = P (x∗ )|∇v(x∗ )|2 ∆x ηα,β (|x∗ |) − 2P (x∗ ) · |∇v(x∗ )|2 |∇ηα,β (|x∗ |)|2 ηα,β (|x∗ |) [P (x∗ )]2 |∇ηα,β (|x∗ |)|2 − 2P (x∗ )F 0 (v(x∗ ))(∇v(x∗ ) · ∇ηα,β (x∗ )) 2ηα,β (|x∗ |) |∇v(x∗ )|2 |∇ηα,β (|x∗ |)|2 0 ∗ ∗ ∗ − 2F P (x∗ ) |∇v(x∗ )|2 ∆x ηα,β (|x∗ |) − 2 · (v(x ))(∇v(x ) · ∇η (x )) α,β ηα,β (|x∗ |) + = + [P (x∗ )]2 |∇ηα,β (|x∗ |)|2 2ηα,β (|x∗ |) Since wα,β (x∗ ) = ηα,β (|x∗ |)P (x∗ ) = T > 0. By the Property I of ηα,β (t), we know ηα,β (|x∗ |) > 0. So we get P (x∗ ) > 0. By D2 wα,β (x∗ ) ≤ 0, then ∆wα,β (x∗ ) ≤ 0. Hence, we can get 2|∇v(x∗ )|2 |∇ηα,β (|x∗ |)|2 P (x∗ )|∇ηα,β (|x∗ |)|2 ≤ + 2F 0 (v(x∗ ))(∇v(x∗ ) · ∇ηα,β (x∗ )) − |∇v(x∗ )|2 ∆x ηα,β (|x∗ |). ∗ 2ηα,β (|x |) ηα,β (|x∗ |) If |∇v(x∗ )|2 ≤ α, then we have Tα,β = wα,β (x∗ ) = ηα,β (|x∗ |)P (x∗ ) ≤ P (x∗ ) Since 0 < ηα,β (|x∗ |) ≤ 1, and P (x∗ ) > 0 = |∇v(x∗ )|2 − F (v(x∗ )) ≤ |∇v(x∗ )|2 Since F (v(x∗ )) ≥ 0 ≤ α. If |∇v(x∗ )|2 > α, then we have Tα,β = wα,β (x∗ ) = ηα,β (|x∗ |)P (x∗ ) – 15 – = ≤ ≤ ≤ 2[ηα,β (|x∗ )]2 P (x∗ )|∇ηα,β (|x∗ |)|2 · |∇ηα,β (|x∗ )|2 2ηα,β (|x∗ |) 2[ηα,β (|x∗ )]2 2|∇v(x∗ )|2 |∇ηα,β (|x∗ |)|2 0 ∗ ∗ ∗ ∗ 2 ∗ + 2F (v(x ))(∇v(x ) · ∇η (x )) − |∇v(x )| ∆ η (|x |) α,β x α,β |∇ηα,β (|x∗ )|2 ηα,β (|x∗ |) 2[ηα,β (|x∗ )]2 2|∇v(x∗ )|2 |∇ηα,β (|x∗ |)|2 0 ∗ ∗ ∗ ∗ 2 ∗ + 2|F (v(x ))||(∇v(x ) · ∇ηα,β (x ))| − |∇v(x )| ∆x ηα,β (|x |) |∇ηα,β (|x∗ )|2 ηα,β (|x∗ |) 2[ηα,β (|x∗ )]2 2|∇v(x∗ )|2 |∇ηα,β (|x∗ |)|2 0 ∗ ∗ ∗ ∗ 2 ∗ + 2|F (v(x ))||∇v(x )||∇η (x )| − |∇v(x )| ∆ η (|x |) α,β x α,β |∇ηα,β (|x∗ )|2 ηα,β (|x∗ |) By the Cauchy-Schwarz’s inequality |∇v(x∗ )|2 · 2[ηα,β (|x∗ )]2 2|∇ηα,β (|x∗ |)|2 1 0 ∗ ∗ ∗ ∗ = · 2|F (v(x ))||∇v(x )||∇ηα,β (x )| − ∆x ηα,β (|x |) + |∇ηα,β (|x∗ )|2 ηα,β (|x∗ |) |∇v(x∗ )|2 [ηα,β (|x∗ )]2 2|∇ηα,β (|x∗ |)|2 M ∗ ∗ ≤ N· + |∇η (|x |)| − ∆ η (|x |) α,β x α,β |∇ηα,β (|x∗ )|2 ηα,β (|x∗ |) α By the definition of M and N , and |∇v(x∗ )|2 > α 0 Since ηα,β is radial function, and ηα,β (|x∗ |) < 0, then we know 0 |∇ηα,β (|x∗ |)| = −ηα,β (|x∗ |) ∆x ηα,β (|x∗ |) 00 = ηα,β (|x∗ |) + n−1 0 η (|x∗ |). |x∗ | α,β So we get Tα,β = wα,β (x∗ ) [ηα,β (|x∗ )]2 2|∇ηα,β (|x∗ |)|2 M ∗ ∗ + |∇η (|x |)| − ∆ η (|x |) α,β x α,β |∇ηα,β (|x∗ )|2 ηα,β (|x∗ |) α [ηα,β (|x∗ )]2 M 0 n−1 0 2|∇ηα,β (|x∗ |)|2 ∗ 00 ∗ ∗ = N· − η (|x |) − ηα,β (|x |) − η (|x |) 0 (|x∗ |)]2 [−ηα,β ηα,β (|x∗ |) α α,β |x∗ | α,β α = N· By the Property v of ηα,β (t) N ≤ N· = α. Therefore, for all |x| ≥ β, we have wα,β (x) ≤ Tα,β = wα,β (x∗ ) ≤ α. In a summary, we have Therefore, for all |x| ≥ β, we have wα,β (x) = ηα,β (|x|)P (x) ≤ max α, max P (x) . |x|=β Take α & 0, by the Property iv of ηα,β (t), we know P (x) ≤ max {0, P (0)}, for all x ∈ Rn . Since |∇v(0)|2 < δ, then we have P (0) = |∇v(0)|2 − 2F (v(x)) – 16 – ≤ |∇v(0)|2 < δ. Since F (v(x)) ≥ 0 So we get P (x) = |∇v(x)|2 − 2F (v(x)) < δ, for all x ∈ Rn . Since v(x) = u(x + x0 ), then |∇u(x + x0 )|2 − 2F (u(x + x0 )) < δ, for all x ∈ Rn . By the arbitrary of x in Rn , we can get |∇u(x)|2 − 2F (u(x)) < δ, for all x ∈ Rn . By the arbitrary of δ > 0, we have |∇u(x)|2 ≤ 2F (u(x)), for all x ∈ Rn . Approach II of Theorem 5.1, Proposition 5.1 on Page 24 in Gui[2]. Define 1 H(x) = F (u(x)) − |∇u(x)|2 , 2 ∀x ∈ Rn . By Theorem 3.1, we know that there exists some C1 > 0 such that H(x) > −C1 , ∀x ∈ Rn . So we can define A = infn H(x). x∈R n Then there exists a sequence xk ∈ R such that lim H(xk ) = A. k→∞ For any k ≥ 1, let uk (x) = u(x + xk ) in Rn , by the standard elliptic estimates, we know that there exists some C2 > 0 such that kuk kC 2,α (Rn ) ≤ C2 . By the Ascoli-Arzela Theorem, there exists some v ∈ C 2 (Rn ) such that uk → v in C 2 (Rn ), as k → ∞, which implies that ∆uk (x) → ∆v(x), F (uk (x)) → F (v(x)) and |∇uk (x)| → |∇v(x)| for all x ∈ Rn , as k → ∞. Therefore, we know that v satisfies ∆v(x) = F 0 (v(x)) in Rn , and 1 A = F (v(0)) − |∇v(0)|2 = infn x∈R 2 – 17 – 1 F (v(x)) − |∇v(x)|2 . 2 Let 1 G(x) = F (v(x)) − |∇v(x)|2 , 2 ∀x ∈ Rn . Direct computations show that |∇v(x)|2 ∆G(x) = n X bi (x)Gxi (x) i=1 2 n n n X X X + uxj (x)uxi xj (x) − |∇u(x)|2 |uxi xj (x)|2 i=1 ≤ n X j=1 i,j=1 bi (x)Gxi (x), By Cauchy-Schwarz’s inequality i=1 That is, we get −|∇v(x)|2 ∆G(x) + n X bi (x)Gxi (x) ≥ 0, ∀x ∈ Rn . i=1 If ∇v(0) = 0, then A = F (v(0)) ≥ 0. If ∇v(0) 6= 0, since G(x) attains its minimum at x = 0, by the Strong Maximum Principle, we know that G(x) ≡ A for all x ∈ Rn . On the other hand, by Theorem 2.1, we know that A ≥ 0. In summary, we have G(x) ≥ A ≥ 0 in Rn , which implies that |∇u(x)|2 ≤ 1 F (u(x)), 2 ∀x ∈ Rn . Remark 5.1. In the proof of the above theorem, we find that if we assume F (t) ≥ 0 for [−M, M ], where kukL∞ (Rn ) ≤ M , then the above proof still works, so the theorem holds as well. Remark 5.2. The above gradient estimate is optimal. Look at the PDE: ∆x u(x) = sin u(x), in Rn . Letu(x) = 4 arctan ex1 , then we know ∂u(x) ∂x1 = 4ex1 1 + e2x1 ∂ 2 u(x) ∂x1 ∂x1 = 4[ex1 (1 + e2x1 ) − ex1 · 2e2x1 ] [1 + e2x1 ]2 = 4[ex1 + e3x1 − 2e3x1 ] [1 + e2x1 ]2 = 4[ex1 − e3x1 ] [1 + e2x1 ]2 = 4(t − t3 ) [1 + t2 ]2 Let t = ex1 > 0 So we have ∆x u(x) = ∂u(x) 4(t − t3 ) = . ∂x1 [1 + t2 ]2 – 18 – On the other hand, we know sin u(x) = sin(4 arctan ex1 ) = sin(4 arctan t) = 2 tan(2 arctan t) 1 + tan2 (2 arctan t) For tan(2 arctan t), we have tan(2 arctan t) = = 2 tan(arctan t) 1 − tan2 (arctan t) 2t . 1 − t2 Hence, we get sin u(x) = 2 tan(2 arctan t) 1 + tan2 (2 arctan t) = 2t 2 · 1−t 2 2 2t 1 + 1−t 2 = 4t[1 − t2 ] [1 − t2 ]2 + 4t2 = 4[t − t3 ] [1 + t2 ]2 = ∆x u(x). Then u(x) = 4 arctan ex1 is a bounded solution of this PDE, and 0 < u(x) < 2π for all x ∈ Rn . Take F (t) = 1 − cos t, and t = ex1 , we have 2F (u(x)) = 2 − 2 cos(4 arctan ex1 ) = 2 − 2 cos(4 arctan t) 4 sin2 (2 arctan t) 2 2 tan(arctan t) = 4 1 + tan2 (arctan t) 2 2t = 4 1 + t2 = = |∇u(x)|2 16t2 [1 + t2 ]2 4ex1 1 + e2x1 4t = 1 + t2 – 19 – = 2 = 16t2 [1 + t2 ]2 = 2F (u(x)). In a summary, we can find u(x) such that |∇u(x)|2 = 2F (u(x)), for all x ∈ Rn . 6. Liouville Type Theorem Theorem 6.1. Let F ∈ C 2 (Rn ), and F (t) ≥ 0 for all t ∈ R. For any u ∈ C 3 (Rn ) T L∞ (Rn ) which is a solution of the PDEs: ∆x u(x) = F 0 (u(x)), for all x ∈ Rn . Furthermore, we assume that there exists some x0 ∈ Rn such that F (u(x0 )) = 0. Then u(x) ≡ u(x0 ), for all x ∈ Rn . Proof. Let a = u(x0 ) ∈ R, and define Γa (u) = {x ∈ Rn : u(x) = a}. Then Γa (u) 6= φ. Since u ∈ C 3 (Rn ), then we know that Γa (u) is a closed subset of Rn . Now we show that Γa (u) is also an open subset of Rn . For any x∗ ∈ Γa (u), since F (t) ≥ 0 for all t ∈ R, and F (a) = 0, then t = a is a minimum point of F in R. So F 0 (a) = 0, and F 00 (a) ≥ 0. By the Taylor’s expansion of F , we know there exist some small δ > 0 such that whenever |t − a| ≤ δ, we have F (t) F 00 (ξt,a ) (t − a)2 2 = F (a) + F 0 (a)(t − a) + = F 00 (ξt,a ) (t − a)2 2 ≤ kF 00 kL∞ ([a−δ,a+δ]) (t − a)2 = k(t − a)2 . Since u ∈ C 3 (Rn ), then there exists some > 0 such that whenever |x − x∗ | < , we have |u(x) − u(x∗ )| < δ. For any ω ∈ Rn with |ω| = 1, we define φ(t) = u(x∗ + tω) − u(x∗ ) = u(x∗ + tω) − a, Then we have φ0 (t) = ∇u(x∗ + tω) · ω, so |φ0 (t)|2 = |∇u(x∗ + tω) · ω|2 – 20 – for all |t| < . ≤ |∇u(x∗ + tω)|2 ≤ 2F (u(x∗ + tω)) By Cauchy-Schwarz’s inequality, and |ω| = 1 By Theorem 5.1 ≤ 2k[u(x∗ + tω) − a]2 = 2k[φ(t)]2 . Then by the Fundamental Theorem of Calculus, and φ(0) = 0, we know for all |t| < , we have Z t φ0 (s)ds |φ(t)| = 0 t Z |φ0 (s)|ds ≤ 0 ≤ Z t√ 2k|φ(s)|ds 0 √ = Z t |φ(s)|ds. 2k 0 By the Bellman’s inequality or the Integral form of the Gronwall’s inequality (Theorem 2, in Page 2, in [4]), all |t| < , we have |φ(t)| ≤ 0. That is φ(t) = for all |t| < . So we get, for all |t| < , and |ω| = 1, we have u(x∗ + tω) = u(x∗ ) = a. In particular, we know B (x∗ ) ⊂ Γa (u). So Γa (u) is open. So Γa (u) = Rn . Therefore, we know u(x) ≡ u(x0 ), for all x ∈ Rn . References [1] David Gilbarg and Neil S. Trudinger. Elliptic partial differential equations of second order. Springer, 2001. [2] Changfeng Gui. Allen-Cahn equation and its generalizations. Notes, 2009. [3] Luciano Modica. A gradient bound and a liouville theorem for nonlinear poisson equations. Communications on Pure and Applied Mathematics, 38:679–684, 1985. [4] Mingfeng Zhao. The Gronwall’s inequality. Personal Notes on Mathematical Problems or Theorems, 2011. – 21 –