Harmonic approximation and improvement of flatness in a singular perturbation problem 4 1 0 Kelei Wang ∗ 2 Wuhan Institute of Physics and Mathematics, n a Chinese Academy of Sciences, Wuhan 430071, China J 5 [email protected] 1 ] P A Abstract . h t We study the De Giorgi type conjecture, that is, one dimensional symmetry prob- a m lem for entire solutions of an two components elliptic system in Rn, for all n ≥ 2. We prove that, if a solution (u,v) has a linear growth at infinity, then it is one [ dimensional, that is, depending only on one variable. The main ingredient is an 1 improvement of flatness estimate, which is achieved by the harmonic approximation v 7 technique adapted in the singularly perturbed situation. 1 5 Keywords: elliptic systems, phase separation, one dimensional symmetry, harmonic approxima- 3 . tion. 1 0 AMS Subject Classification (2010): 35B06, 35B08, 35B25, 35J91. 4 1 : v 1 Introduction i X r a In this paper, we continue our study in [18] on the De Giorgi type conjecture, i.e. one dimensional symmetry problem for solutions of the following two component elliptic system in Rn: ∆u = uv2, ∆v = vu2, u,v > 0 in Rn. (1.1) Weremovetheenergyminimizingconditionin[18]andprovetheonedimensionalsymmetry only under the linear growth condition. More precisely we prove ∗ The author is supported by NSFC No. 11301522. I would like to thank Prof. Fang Hua Lin for simulating discussions which clarify the geometric meanings of some arguments in this paper. 1 Theorem 1.1. If (u,v) is a solution of the problem (1.1), and there exists a constant C > 0 such that for any x ∈ Rn, u(x)+v(x) ≤ C(1+|x|), (1.2) then after a suitable rotation in Rn, u(x) ≡ u(x ), v(x) ≡ v(x ). n n The linear growth condition is sharp, as shown by the examples constructed in [4], where u − v is asymptotic to a homogeneous harmonic polynomial of degree d ≥ 2. For more discussions on (1.1), we refer to [3, 4, 6, 14]. Through suitable rescalings, the problem (1.1) is closely related to the following singu- larly perturbed problem (see Theorem 2.1 below) ∆u = κu v2, κ κ κ (1.3) ∆v = κv u2, ( κ κ κ which is used to describe the “phase separation” phenomena. When κ → +∞, the conver- gence ofsolutions (u ,v ) of (1.3)andtheir singular limit were studied by Caffarelli andLin κ κ [6], Noris-Tavares-Terracini-Verzini [14] and Tavares-Terracini [16] (see also Dancer-Zhang and the author [9]). The main ingredient of our proof is an improvement of flatness estimate for the singular perturbation problem (1.3), which is achieved by the blow up (harmonic approximation) technique. This type of arguments, first introduced by De Giorgi in his work on the regularity of minimal hypersurfaces [10], are by now classical in the elliptic regularity theory. It plays an important role in the establishment of many ε -regularity theorems, such as in the theory of stationary varifolds (cf. Allard [1], see also [13, Section 6.5] for an account), harmonic maps (cf. L. Simon [17]) and nonlinear elliptic systems (the indirect method, see for example Chen-Wu [7, Chapter 12]), just to name a few examples. In singular perturbation problems, Savin’s proof of the De Giorgi conjecture for Allen- Cahn equation [15] also uses an improvement of flatness estimate and the harmonic ap- proximation type argument. However, there the quantity to be improved is different from the classical energy quantity. Indeed, the method developed in [15] is mainly on the vis- cosity (or Krylov-Safonov) side and corresponds to the Harnack inequality approach to the regularity of minimal hypersurfaces as developed in Caffarelli-Cordoba [5]. In this paper we will explore some aspects of harmonic approximation arguments in the singular perturbation problem (1.3), from the variational side. Thus in our estimate we still use an energy type quantity, which is similar to the excess used in Allard’s regularity theory. In this sense, our method may be viewed as a direct generalization of the classical harmonic approximation technique in this singular perturbation problem. 2 However, in order to get a harmonic function in the blow up limit, we use the stationary condition arising from the equation, but not the equation (1.3) itself. Let us first recall the stationary condition. Given a κ > 0 fixed, any solution of (1.3), (u ,v ) is smooth. Let Y κ κ be a smooth vector field with compact support, then by considering domain variations in the form ut(x) := u (x+tY(x)), vt(x) := v (x+tY(x)), for |t| small, κ κ κ κ we have d |∇ut(x)|2 +|∇vt(x)|2 +κut(x)2vt(x)2 dx = 0. dt κ κ κ κ t=0 Z (cid:12) (cid:0) (cid:1) Through some integration by parts we obtain the stationary cond(cid:12)ition for (u ,v ), (cid:12) κ κ |∇u |2 +|∇v |2 +κu2v2 divY −2DY(∇u ,∇u )−2DY(∇v ,∇v ) = 0. (1.4) κ κ κ κ κ κ κ κ Z (cid:0) (cid:1) Here div is the divergence operator, and for a function u, n ∂Yα ∂u ∂u DY(∇u,∇u) = . ∂x ∂x ∂x β α β α,β=1 X For the problem (1.3), we have better control and convergence on the energy level, while the equation itself is badly behaved. This is why we choose to blow up the stationary condition to get a harmonic function in the limit. Here we would like to mention that the stationary condition appears more naturally in some other singular perturbation problems, such as the Allen-Cahn model (cf. Hutchinson-Tonegawa [12]) and the Ginzburg-Landau model (cf. Bethuel-Brezis-Orlandi [2]). In these problems, the stationary condition is directly linked to the limit problem, i.e. the stationary condition for varifolds (in the sense of Allard [1]). In the remaining part of this paper, a solution (u,v) of the problem (1.1) will be fixed. We use the notation a = O(b ), if there exists a constant C such that, as κ → +∞, κ κ |a | ≤ Cb , κ κ and we say a = o(b ) if κ κ a κ lim = 0. κ→+∞ bκ We use C to denote various universal constants, which are independent of the base point x ∈ Rn and the radius R. (In some cases it depends on the solution itself.) It may be different from line to line. Hs is used to denote the s−dimensional Hausdorff measure. 3 2 The improvement of flatness First, to explain why our main Theorem 1.1 is related to a singular perturbation problem, let us recall the following result, which is essentially [18, Lemma 5.2]. Theorem 2.1. For any ε > 0, there exists an R such that if R ≥ R and x ∈ {u = v}, 0 0 0 by defining 1 1 u (x) := u(x +Rx), v (x) := v(x +Rx), (2.1) κ 0 κ 0 R R there exists a constant c independent of x ∈ {u = v} and R, and a vector e satisfying 0 0 |e| ≥ c , such that 0 |∇u −∇v −e|2 ≤ ε2. κ κ ZB1(0) Note that (u ,v ) satisfies (1.3) with κ = R4. For a proof of this theorem see [18, κ κ Lemma 5.2]. The only new point is that, if sup |u −v −e·x|2 ≤ ε2, κ κ B2(0) thenu −v isalsoclosetoe·xinH1(B (0))topology. Thiscanbeprovedbyacontradiction κ κ 1 argument, using the H1 strong convergence for solutions of (1.3) (cf. [14, Theorem 1.2]). Note that we can replace the global uniform H¨older estimate used in [14] by the interior uniform H¨older estimate [18, Theorem 2.6]. Throughout this paper, (u ,v ) always denotes a solution defined as in (2.1). The κ κ followingimprovement ofdecayestimatewillbethemainingredientinourproofofTheorem 1.1. Theorem 2.2. There exist four universal constants θ ∈ (0,1/2), ε small and K ,C(n) 0 0 large such that, if (u ,v ) is a solution of (1.3) in B (0), satisfying κ κ 1 |∇u −∇v −e|2 = ε2 ≤ ε2, (2.2) κ κ 0 ZB1(0) where e is a vector satisfying |e| ≥ c /2, and κ1/4ε2 ≥ K , then there exists another vector 0 0 e˜, with |e˜−e| ≤ C(n)ε, such that 1 θ−n |∇u −∇v −e˜|2 ≤ ε2. κ κ 2 ZBθ(0) 4 The proof will be given later. Note that this theorem is not a local result. It depends on the global Lipschitz estimate established in [18], which is stated for solutions of (1.1) defined on the entire space Rn. This decay estimate can be used to prove Theorem 2.3. There exists a constant C > 0 such that, for any x ∈ {u = v} and R > 1, there exists a vector e , with x,R |e | ≥ c /2, x,R 0 such that |∇u−∇v −e |2 ≤ CRn−1. x,R ZBR(x) Proof. Fix an R > 1 and x ∈ {u = v}, which we assume to be the origin 0. For each 0 i > 0, denote R := Rθ−i. i Let E and the vector e be defined by i i E := minR1−n |∇u−∇v −e|2 = R1−n |∇u−∇v−e |2. i e∈Rn i i i ZBRi(0) ZBRi(0) Note that for any fixed e, R1−n |∇u−∇v −e|2 ≤ R1−n |∇u−∇v −e|2 i i ZBRi(0) ZBθ−1Ri(0) ≤ θ1−n θ−1R 1−n |∇u−∇v −e|2. i (cid:0) (cid:1) ZBθ−1Ri(0) Hence we always have E ≤ θ1−nE . (2.3) i i+1 Furthermore, since (see [18, Theorem 5.1]) sup(|∇u|+|∇v|) < +∞, Rn there exists a constant C, which is independent of i, such that E ≤ Cθ−i. (2.4) i By Theorem 2.1, for any sequence i → +∞, there exists a subsequence (still denoted by i) such that u (x) := R−1u(R x) → (e·x)+, v (x) := R−1v(R x) → (e·x)−. i i i i i i 5 Here e is a vector in Rn satisfying |e| ≥ c , and the convergence is in C (Rn) and also in 0 loc H1 (Rn). Note that (u ,v ) satisfies (1.3) with κ = R4. loc i i i i Indeed, by Theorem 2.1, if R ≥ R , where R is a constant depending only on ε , there i 0 0 0 exists a vector e¯ with |e¯| ≥ c such that i i 0 |∇u−∇v−e¯|2 ≤ ε2Rn. i 0 i ZBRi(0) By definition, if we replace e¯ by e , we can get the same estimate. Thus by Theorem 2.2, i i if we also have E ≥ K , or equivalently, i+1 0 |∇u −∇v −e |2 ≥ K R−1 = K κ−14, i+1 i+1 i+1 0 i+1 0 i+1 ZB1(0) then there exists another vector e˜ so that i+1 1 θ−n |∇u −∇v −e˜ |2 ≤ |∇u −∇v −e |2. i+1 i+1 i+1 i+1 i+1 i+1 2 ZBθ(0) ZB1(0) This can be rewritten as E ≤ R1−n |∇u−∇v −e˜ |2 i i i+1 ZBRi(0) = θ−nR |∇u −∇v −e˜ |2 i i+1 i+1 i+1 ZBθ(0) R ≤ i |∇u −∇v −e |2 (2.5) i+1 i+1 i+1 2 ZB1(0) 1 = R R−n |∇u−∇v −e |2 2 i i+1 i+1 ZBRi+1(0) θ = E . i+1 2 Now we claim that for all i ≥ min{logR0−logR,1}, |logθ0| E ≤ θ1−nK . (2.6) i 0 Assume by the contrary, there exists an i ≥ min{logR0−logR,1} such that E > θ1−nK . 0 |logθ0| i0 0 First by (2.3), E ≥ θn−1E > K . i0+1 i0 0 6 Thus the assumptions of Theorem 2.2 are satisfied and we have (2.5), which says 2 E ≥ E ≥ E > θ1−nK . i0+1 θ i0 i0 0 This can be iterated, and we get, for any j ≥ 0, j+1 2 2 E ≥ E ≥ E . i0+j+1 θ i0+j i0 θ (cid:18) (cid:19) However, since 2/θ > 1/θ, this contradicts (2.4) if j is large enough. Note that the constant θ1−nK in (2.6) is independent of the base point x ∈ {u = v} and the radius R. Thus we 0 0 get (2.6) for any R ≥ R and x ∈ {u = v}. Then by choosing a larger constant, this can 0 be extended to cover [1,R ] if we note the global Lipschitz bound of u and v. 0 We have shown the existence of e for any x ∈ {u = v} and R > 1. The lower bound x,R for |e | can be proved as in the proof of Theorem 2.1 by using the Alt-Caffarelli-Friedman x,R inequality (see [18, Theorem 4.3]). With this theorem in hand, we can use e ·(y−x) to replace the harmonic replacement x,R ϕ in [18, Section 7]. The following arguments to prove Theorem 1.1 are exactly the same R,x one in [18, Section 8 and 9]. The remaining part of this paper will be devoted to the proof of Theorem 2.2. 3 Some a priori estimates In this section, we present some a priori estimates for the solution (u,v). These estimates show that various quantities, when integrated on B (x), have a growth bound as Rn−1. R This is exactly what we expect for one dimensional solutions. Several estimates from [18] will be needed in this section. Lemma 3.1. There exist two positive constants C and M, such that for any R > CM and t ≥ M, Hn−1(B ∩{u = t}) ≤ CRn−1. R Proof. First, by [18, Lemma 5.2 and Lemma 5.4], there exists a constant c(M) > 0 such that, |∇u| ≥ c(M) on {u = t}. (3.1) Since u is smooth, by the implicit function theorem {u = t} is a smooth hypersurface. Now Hn−1(B ∩{u = t}) ≤ c(M)−1 |∇u|. (3.2) R ZBR∩{u=t} 7 Note that on {u = t} ∂u |∇u| = − , ∂ν where ν is the unit normal vector of {u = t} pointing to {u < t}. Then by the divergence theorem ∂u ∂u − = − ∆u ∂ν ∂r ZBR∩{u=t} Z∂BR∩{u>t} ZBR∩{u>t} ≤ |∇u|+ ∆u Z∂BR∩{u>t} ZBR ≤ 2 |∇u| ≤ CRn−1. Z∂BR Here we have used the global Lipschitz continuity of u, cf. [18, Theorem 5.1]. The same results also hold for v and u−v, which we do not repeat here. Next we give a measure estimate for the transition part {u ≤ T,v ≤ T}. Lemma 3.2. For any T > 1, there exists a constant C(T) > 0, such that for any R > 1 and x ∈ Rn, Hn(B (x)∩{u ≤ T,v ≤ T}) ≤ C(T)Rn−1. R Proof. First we have the Claim. For each T > 1, there exists a c(T) > 0 such that, if x ∈ {u ≤ T,v ≤ T}, then 0 u(x ) ≥ c(T), v(x ) ≥ c(T). (3.3) 0 0 By assuming this claim, we get Hn(B (x)∩{u ≤ T,v ≤ T}) ≤ c(T)−4 u2v2 ≤ C(T)Rn−1, R ZBR(x)∩{u≤T,v≤T} where in the last inequality we have used [18, Lemma 6.4]. To prove the claim, first we note that, there exists a constant C (T) such that 1 dist(x ,{u = v}) ≤ C (T). (3.4) 0 1 Indeed, if dist(x ,{u = v}) ≥ L (L large to be chosen), take y ∈ {u = v} to realize this 0 0 distance and define 1 1 u˜(x) = u(y +Lx), v˜(x) = v(y +Lx). 0 0 L L 8 Then by [18, Lemma 5.2], there exits a vector e and a universal constant C, with 1 ≤ |e| ≤ C, C such that |u˜(x)−(e·x)+|+|v˜(x)−(e·x)−| ≤ h(L), where h(L) is small if L large enough. Without loss of generality we can assume B (L−1(x −y )) ⊂ {u˜ > v˜}. By a geometric 1 0 0 consideration, we have 1 L−1(x −y )·e ≥ . 0 0 2C Consequently, 1 u˜(L−1(x −y )) ≥ L−1(x −y )·e−h(L) ≥ . 0 0 0 0 4C Thus u(x ) > T if L large, which is a contradiction. 0 After establishing (3.4), we can use the standard Harnack inequality and [18, Lemma 4.7] to deduce the claimed (3.3). Lemma 3.3. There exists a constant C > 0, such that for any R > 1 and x ∈ Rn, |∇u||∇v| ≤ CRn−1. ZBR(x) Proof. Fix a T > 0, which will be determined below. (It is independent of x and R.) We divide the estimate into three parts, {u ≤ T,v ≤ T}, {u > T} and {v > T}. Note that if T is large enough, by [18, Lemma 6.1], these three parts are disjoint. First in B (x) ∩ {u ≤ T,v ≤ T}, by the global Lipschitz continuity of u and v [18, R Theorem 5.1] and the previous lemma, we have |∇u||∇v| ≤ CHn(B (x)∩{u ≤ T,v ≤ T}) ≤ CRn−1. (3.5) R ZBR(x)∩{u≤T,v≤T} If T large, in {u > T}, |∇u| ≥ c(T) > 0 for a constant c(T) depending only on T (cf. the proof of Lemma 3.1). Furthermore, by the proof of [18, Lemma 6.3], there exists a constant C such that |∇v| ≤ Ce−Cu in {u > T}. Then by the co-area formula and Lemma 3.1, +∞ |∇u||∇v| = |∇v| dt ZBR(x)∩{u>T} ZT (cid:18)ZBR∩{u=t} (cid:19) 9 +∞ ≤ C e−ctHn−1(B ∩{u = t})dt R ZT ≤ CRn−1. The same estimate holds for {v > T}. Putting these together we can finish the proof. Lemma 3.4. There exists a constant C > 0, such that for any R > 1 and x ∈ Rn, uv3 +vu3 ≤ CRn−1. ZBR(x) Proof. We still choose a T > 0, which will be determined below, and divide the estimate into three parts, {u ≤ T,v ≤ T}, {u > T} and {v > T}. By the proof of [18, Lemma 6.1], we still have uv3 +vu3 ≤ C in Rn. Then in B (x)∩{u ≤ T,v ≤ T}, R uv3 +vu3 ≤ CHn(B (x)∩{u ≤ T,v ≤ T}) ≤ CRn−1. R ZBR(x)∩{u≤T,v≤T} Next, by the proof of [18, Lemma 6.1], there exists a constant C such that uv3 +vu3 ≤ Ce−Cu in {u > T}. Then by the co-area formula and the lower bound of |∇u| in {u > T} (i.e. (3.1)), +∞ uv3 +vu3 uv3 +vu3 = dt |∇u| ZBR(x)∩{u>T} ZM (cid:18)ZBR∩{u=t} (cid:19) +∞ ≤ C e−ctHn−1(B ∩{u = t})dt R ZM ≤ CRn−1. The same estimate holds for {v > T}. Putting these together we can finish the proof. 4 Blow up the stationary condition In this section and the next one, we prove Theorem 2.2. We argue by contradiction, so assume that as κ → +∞, there exists a sequence of solutions (u ,v ) satisfying the κ κ conditions but not the conclusions in that theorem, that is, |∇u −∇v −e|2 = ε2 → 0, (4.1) κ κ κ ZB1(0) 10