Nodal Sets of Steklov Eigenfuntions by Katarı́na Bellová A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy Department of Mathematics New York University May, 2012 Professor Fanghua Lin Abstract We study the nodal set of the Steklov eigenfunctions on the boundary of a smooth bounded domain in Rn – the eigenfunctions of the Dirichlet-to-Neumann map Λ. For a bounded Lipschitz domain Ω ⊂ Rn , this map associates to each function u defined on the boundary ∂Ω, the normal derivative of the harmonic function on Ω with boundary data u. Under the assumption that the domain Ω is C 2 , we prove a doubling property for the eigenfunction u. The main goal of this Thesis is to estimate the Hausdorff Hn−2 -measure of the nodal set of u|∂Ω in terms of the eigenvalue λ as λ grows to infinity, provided Ω is fixed. In case that the domain Ω is analytic, we prove a polynomial bound (Cλ6 ). My methods, which build on the work of Lin, Garofalo and Han [Garofalo and Lin, CPAM 40 (1987), no. 3; Lin, CPAM 42 (1989), no. 6; Han and Lin, JPDE 7 (1994), no. 2], can be used also for solutions to more general elliptic equations and/or boundary conditions. iii Contents Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii Introduction 0.1 Problem setting and main results . . . . . . . . . . . . . . . . . . . . . . . 0.2 Other related problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 0.3 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2 3 4 1 Steklov Eigenfunctions and the Doubling Condition 6 1.1 Global Integral quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 1.2 The Frequency Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.3 Doubling Condition for Steklov Eigenfunctions . . . . . . . . . . . . . . . 18 2 Size of the Nodal Sets of Steklov Eigenfunctions 39 2.1 Nodal Sets of Steklov Eigenfunctions on Analytic Domains . . . . . . . . 39 2.2 Nodal Sets of Steklov Eigenfunctions on Smooth Domains . . . . . . . . . 44 Bibliography 45 iv Introduction Many classical results in linear elliptic partial differential equations are motivated by complex analysis: the maximum principle (for modulus of holomorphic functions), the unique continuation properties, the Cauchy integral representation formula, and the interior gradient estimates of holomorphic functions were all generalized to the solutions of linear second order elliptic PDEs with suitably smooth coefficients. A lot is known about the nodal (zero) sets of holomorphic functions in complex plane, and it is an important general research topic to find analogues for the nodal and critical point sets of solutions to PDEs. In some cases, properties of nodal sets of solutions are themselves the primary concern: in a study of moving defects in nematic liquid crystals by Lin [28], the singular set of optical axes (i.e. defects) of liquid crystals in motion can be described precisely by the nodal set of solutions to certain parabolic equations. In other cases, nodal sets provide important information in the study of other properties of solutions. In this Thesis, we are studying the nodal set of the Steklov eigenfunctions on the boundary of a smooth bounded domain in Rn – the eigenfunctions of the Dirichlet-toNeumann map Λ. For a bounded Lipschitz domain Ω ⊂ Rn , this map associates to each function u defined on the boundary ∂Ω, the normal derivative of the harmonic function on Ω with boundary data u. More generally, one can consider an n-dimensional smooth Riemannian manifold (M, g) instead of Ω, and replace the Laplacian by the LaplaceBeltrami operator ∆g . Our methods, which build on the work of Lin, Garofalo, and Han ([18, 29, 23]), can be used also for solutions to more general elliptic equations and/or boundary conditions. The Steklov eigenfunctions were introduced by Steklov [34] in 1902 for bounded domains in the plane. He was motivated by physics – the functions represent the steady state temperature on Ω such that the flux on the boundary is proportional to the temperature. The problem can also be interpreted as vibration of a free membrane with the mass uniformly distributed on the boundary. Note that the eigenfunction’s nodal set represents the stationary points on the boundary. The Steklov problem is also important in conformal geometry, Sobolev trace inequalities, and inverse problems. The Thesis is organized as follows. The remainder of the introduction summarizes the main results, describes related problems and fixes some basic notation. In Chapter 1 we prove a doubling condition for Steklov eigenfunctions on a C 2 -domain Ω, which will serve as a cornerstone for the nodal set estimate in both analytic and smooth setting. In Chapter 2 we prove an explicit estimate for the nodal set of Steklov eigenfunctions in the case that Ω has analytic boundary. We also discuss the case when Ω is smooth, but 1 not analytic, and outline a possible direction of research in that setting. 0.1 Problem setting and main results Let Ω ⊂ Rn be a Lipschitz domain. The Dirichlet-to-Neumann operator Λ : H 1/2 (∂Ω) → H −1/2 (∂Ω) is defined as follows. For f ∈ H 1/2 (∂Ω), we solve the Laplace equation ∆u = 0 in Ω, u=f on ∂Ω. This gives a solution u ∈ H 1 (Ω), and we set (Λf ) to be the trace of ∂u ∂ν on ∂Ω, where ν is the exterior unit normal. We obtain a bounded self-adjoint operator from H 1/2 (∂Ω) to H −1/2 (∂Ω). It has a discrete spectrum {λj }∞ j=0 , 0 = λ0 < λ1 ≤ λ2 ≤ λ3 ≤ ..., limj→∞ λj = ∞. The eigenfunctions of Λ (called Steklov eigenfunctions) corresponding to eigenvalue λ can be identified with the trace on ∂Ω of their harmonic extensions to Ω, which satisfy ∆u = 0 in Ω, (1) ∂u = λu on ∂Ω. ∂ν The main goal of this Thesis is to estimate the size (the Hausdorff Hn−2 -measure) of the nodal set of u|∂Ω in terms of λ as λ grows to infinity, provided Ω is fixed. In Chapter 2.1 we prove the following bound in case that Ω has analytic boundary: Theorem 1. Let Ω ⊂ Rn be an analytic domain. Then there exists a constant C depending only on Ω and n such that for any λ > 0 and u which is a (classical) solution to (1) there holds Hn−2 ({x ∈ ∂Ω : u(x) = 0}) ≤ Cλ6 . (2) The scaling λ6 in (2) is not optimal. Actually, our proof gives approximately λ5.6 , but we believe that the optimal scaling is λ. However, even this polynomial bound is valuable in a problem like this – the main difficulty in the estimate is to avoid an exponential bound eCλ . As in [29] and [23], we use a doubling condition, i.e. a control of the L2 -norm of u on a ball B2r (x) by the L2 -norm of u on a smaller ball Br (x), as the crucial tool to estimate the nodal set. We prove a doubling condition in the following form in Chapter 1: Theorem 2. Let Ω ⊂ Rn be a C 2 domain. Then there exist constants r0 , C > 0 depending only on Ω and n such that for any r ≤ r0 /λ, x ∈ ∂Ω, and u which is a (classical) solution to (1), there holds Z Z 2 Cλ5 u2 . u ≤2 B(x,r)∩∂Ω B(x,2r)∩∂Ω To put our work into context, note that our problem is similar in nature to the classical question of estimating the size of nodal sets of eigenfunctions of the Laplace operator in a compact manifold. The following conjecture was proposed by Yau in [39]: 2 Conjecture 0.1.1. Suppose (M n , g) is a smooth n-dimensional connected and compact Riemannian manifold without boundary. Consider an eigenfunction u corresponding to the eigenvalue λ, i.e. , ∆g u + λu = 0 on M. Then there holds √ √ c1 λ ≤ Hn−1 ({x ∈ M ; u(x) = 0}) ≤ c2 λ, where c1 and c2 are positive constants depending only on (M, g). This conjecture was proved in case that (M n , g) is analytic by Donnelly and Fefferman in [10]. It is still open whether Conjecture 0.1.1 holds if (M n , g) is only smooth. The known results for the smooth case are far from optimal, the upper bound remaining exponential (see [25]). Another similar problem has been studied for the Neumann eigenfunctions on a piecewise analytic plane domain Ω ⊂ R2 in [36]. This paper is concerned about the asymptotics of the number of nodal points of the eigenfunctions on the boundary ∂Ω, as the eigenvalue λ increases to infinity. It proves that this number is bounded above by CΩ λ. 0.2 Other related problems In this section we briefly present some other problems related to the Dirichlet-to-Neumann map and its eigenvalues/eigenfunctions. They are classical problems which illustrate the importance of this map. The classical Yamabe Problem consists in showing that every Riemannian compact manifold, without boundary, admits a conformally related metric with constant scalar curvature. It was formulated by Yamabe [38] in 1960 and proved by Aubin ([4], 1976) and Schoen ([33], 1984). A related problem for a manifold (M, g) with boundary asks to find a conformally related metric with zero scalar curvature on M and constant mean curvature on the boundary. It was solved in most cases by Escobar ([12]) and by Marques ([31]) in the remaining ones, and can be thought of as a generalization of the Riemann mapping theorem to higher dimensions. This problem can be reformulated as finding a solution to ∆g u = 0 in M, n du n − 2 on ∂M, + hg u = λu n−2 dν 2 u>0 in M, where hg is the mean curvature on the boundary. That corresponds to solving Λg (u) + f (u) = 0 n (3) n−2 . Equations of this type often come on ∂M with a nonlinearity f (u) = n−2 2 hg u − λu up when studying the geometry of a manifold. Much like in [7] for a different problem, the understanding of Steklov eigenfunctions is helpful in the study of such equations. 3 Another well-known example is looking for the optimal constant in the Sobolev trace inequality. The first non-zero eigenvalue for the Steklov problem, λ1 , has a variational characterization: R |∇f |2 λ1 = R min RM . 2 ∂M f =0 ∂M f From this characterization follows this Sobolev trace inequality: for all functions f ∈ H 1 (M ), we have Z Z 1 2 |f − f | ≤ |∇f |2 , (4) λ 1 ∂M M where f is the mean value of f when restricted to the boundary. The last inequality is fundamental in the study of existence and regularity of solutions of some boundary value problems. Hence it is important to know the dependence of λ1 on the geometry of M . It has been studied e.g. in [37, 32, 13, 16]. Estimates for the higher eigenvalues and their distribution have also been studied, see e.g. [5, 26, 9]. Inequality (4) implies the embedding H 1 (M ) ֒→ L2 (∂M ). This is not the optimal embedding. For simplicity, consider the half-space Rn+ . For any function f on Rn+ , sufficiently smooth and decaying fast enough at infinity, there holds ||f ||L2(n−1)/(n−2) (∂Rn ) ≤ C(n)||∇f ||L2 (Rn+ ) , + (5) where C(n) is a known constant. The functions f for which equality holds in (5) satisfy the Euler-Lagrange equation ∆f = 0 ∂f − Q(Rn+ )f n/(n−2) = 0 ∂ν on Rn+ , on ∂Rn+ , where Q(Rn+ ) = 1/C(n)2 . This is again an equation of type (3). See [11] for details, and [27] for a generalization for manifolds. There are many more problems related to the Steklov eigenfunctions. The understanding of the problem could have applications in inverse conductivity problems ([8, 35]), cloaking ([2]), and modeling of sloshing of a perfect fluid in a tank ([15, 6]). 0.3 Notation Throughout this Thesis, Br = B(0, r) will denote the open ball in Rn with center 0 and radius r, and B(x, r) the open ball in Rn with center x and radius r. Similarly we use k B n−1 (x, r) for ball in Rn−1 and B C (z, r) for ball in Ck (k = n, n − 1). We denote the coordinates of a vector b ∈ Rn by bi , i = 1, 2, . . . , n. We abbreviate 2w ∂w the partial derivatives as ∂x = wxi , ∂x∂i ∂x = wxi ,xj . i j We use the letter C as a constant (usually depending only on n and Ω), which can change from line to line. For a given Ω, there are only finitely many eigenvalues λ of the Dirichlet-to-Neumann map which are less than 1. Hence, when we are proving upper bounds in the form of 4 Cλk with specific k and C depending on Ω, without loss of generality we can assume that the eigenvalue λ is larger or equal to 1, and we often do so without pointing it out. 5 Chapter 1 Steklov Eigenfunctions and the Doubling Condition 1.1 Global Integral quantities In this section, we will derive some properties of the gradient of Steklov eigenfunctions. Although we will not use them in the proofs of the main results, they illustrate the special structure of the Steklov eigenfunctions. Let u be a solution of (1) on a Lipschitz domain Ω. By normalization, assume Z u2 = 1. (1.1) ∂Ω Proposition 1.1.1. Z Ω |∇u|2 = λ. (1.2) Proof. By integration by parts and using (1) and (1.1), we have Z Z Z Z Z ∂u ∂u 2 u |∇u| = u u∆u = u2 = λ. − =λ ∂Ω ∂ν Ω ∂Ω ∂ν Ω ∂Ω Next, we are interested in the size of |∇u|2 integrated over ∂Ω. For the gradient in the normal direction we have 2 Z ∂u = λ2 u 2 = λ2 . (1.3) ∂Ω ∂ν ∂Ω We will show that the gradient in the tangential directions is of comparable size. Z Theorem 3. 2 ∂u ≈ λ2 , (1.4) ∂Ω ∂τ where the sign ≈ means that there exist constants C1 , C2 depending on Ω and n such R 2 ≤ C 2 λ2 . that C1 λ2 ≤ ∂Ω ∂u ∂τ Z 6 Proof. To simplify the notation, we will use the summation convention – summing over the indices i and j (from 1 to n) wherever they come up. By multiplying the equation ∆u = 0 by x · ∇u and integrating by parts, we get Z Z Z Z ∂u 2 0= x i u xi u xj xj = x · ∇u − |∇u| − x i u xi xj u xj . (1.5) ∂ν Ω ∂Ω Ω Ω Further integration by parts yields Z Z Z Z 2 2 x i u xi xj u xj , x · ν|∇u| − n |∇u| − x i u xi xj u xj = Ω ΩZ ∂Ω ZΩ Z 1 n x i u xi xj u xj = x · ν|∇u|2 − |∇u|2 . 2 ∂Ω 2 Ω Ω Hence from (1.5) we get Z Z 2 (n − 2) |∇u| = Ω 2 ∂Ω x · ν|∇u| − 2 Z ∂Ω 2 Z 2 ∂u ∂u x · ν − = x · ν − 2 ∂τ ∂ν ∂Ω ∂Ω Z ∂u ∂ν n−1 XZ x · ∇u k=1 ∂Ω x · τk ∂u ∂u , ∂τk ∂ν (1.6) 2 where τ1 , . . . , τn−1 are the mutually perpendicular tangential vectors on ∂Ω, and | ∂u ∂τ | = Pn−1 ∂u 2 ′ k=1 | ∂τk | . Note that (1.6) is valid also if we replace Ω by any subdomain Ω ⊂ Ω. Let us first assume that Ω is star-shaped. Then there exists x0 ∈ Ω such that for each x ∈ ∂Ω the segment x0 x lies in Ω, and (x − x0 ) · ν(x) ≥ δ for some δ > 0 independent of x. Without loss of generality, we can assume x0 = 0 (otherwise we can derive (1.6) with x − x0 in place of x). Since Ω is bounded, we also have |x| ≤ K for x ∈ ∂Ω and some constant K. Hence from (1.6), using (1.2) and (1.3) and the Cauchy-Schwartz inequality, we get 2 Z Z 2 ∂u ∂u 1 ≤ x · ν ∂τ δ ∂Ω ∂τ ∂Ω ! 2 Z Z n−1 XZ ∂u ∂u 1 ∂u x · τk = (n − 2) |∇u|2 + x · ν + 2 δ ∂ν ∂τk ∂ν Ω ∂Ω k=1 ∂Ω Z 2 2 ∂u K 1 n−2 + 2(n − 1)K λ2 . λ + λ2 + ≤ δ δ 2 ∂Ω ∂τ δ2 Therefore, Z ∂Ω 2 ∂u ≤ C 2 λ2 , ∂τ where C2 is a constant depending only on Ω and n. 7 On the other hand, analogously 2 Z 2 Z ∂u ∂u ≥ 1 x · ν ∂τ K ∂Ω ∂Ω ∂τ ! 2 Z Z n−1 XZ ∂u 1 ∂u ∂u 2 x · ν + 2 = (n − 2) |∇u| + x · τk K ∂ν ∂τk ∂ν ∂Ω Ω ∂Ω k=1 Z ∂u 2 δ 2(n − 1)K n−2 − δ λ2 , λ + λ2 − ≥ K K δ 2K ∂Ω ∂τ whence Z 2 ∂u ≥ C 1 λ2 , ∂Ω ∂τ where C1 > 0 is a constant depending only on Ω and n. We have thus proved (1.4) for star-shaped Ω. In the general case, we cover the boundary ∂Ω by p pieces ∂Ω = ∪pα=1 Dα such that each Dα = ∂Ω ∩ A′α , where A′α is a rectangular box and Dα is connected. Moreover, we require that Aα = A′α ∩ Ω is star-shaped (this can be assured because Ω is Lipschitz), and 2n−1 [ Sα,i ∪ Dα , ∂Aα = i=1 where each Sα,i lies in a hyperplane Hα,i and if we shift every Hα,i parallelly by ǫ > 0 in the direction away from Aα , the body formed by these shifted hyperplanes and ∂Ω containing Dα stays in Ω. Such covering of ∂Ω exists, with ǫ > 0 depending on Ω. β Now take one Dα and denote by Hα,i the hyperplane parallel to Hα,i in the distance of β away from A′α , 0 ≤ β ≤ ǫ. By Fubini’s theorem and (1.2), Z Z ǫZ 2 |∇u|2 = λ, |∇u| dS dβ ≤ 0 β Hα,i ∩Ω Ω so there exists β ∈ [0, ǫ] such that Z β Hα,i ∩Ω |∇u|2 dS ≤ Cλ, where C depends on Ω (actually, C = 1/ǫ). Hence, by possibly shifting the walls Sα,i of Dα , we can assume Z |∇u|2 ≤ Cλ. (1.7) Sα,i Now apply (1.6) to Aα instead of Ω (see the remark under (1.6)): (n − 2) Z Aα |∇u|2 = Z 2 Z 2 n−1 XZ ∂u ∂u ∂u ∂u x · ν − x · τk . (1.8) x · ν − 2 ∂τ ∂ν ∂τ k ∂ν ∂Aα ∂Aα ∂Aα k=1 8 For the left-hand side we have Z Z 2 (n − 2) |∇u| ≤ (n − 2) |∇u|2 = (n − 2)λ. Aα Ω If we split the integrals on S the right-hand side of (1.8) into integrals over Dα and S2n−1 2n−1 i=1 Sα,i , all integrals over i=1 Sα,i are at most of order λ by (1.7). Since Z 2 Z 2 ∂u ∂u ≤ = λ2 ∂ν Dα ∂Ω ∂ν and for some α Z 2 Z 2 ∂u ∂u 1 = 1 λ2 , ≥ ∂ν p p ∂Ω ∂ν Dα the same argument as in the star-shaped case gives us the existence of constants C1 , C2 > 0 depending on Ω such that Z 2 ∂u ≤ C 2 λ2 for all α, Dα ∂τ Z 2 ∂u ≥ C 1 λ2 for some α. ∂τ Dα The theorem follows. 1.2 The Frequency Function In this section we will review the theory developed in [17] and [18] about the frequency functions for both harmonic functions and solutions to general elliptic equations. We will heavily rely on these results in the following section (Doubling Condition for Steklov Eigenfunctions). 1.2.1 Frequency function and doubling condition for harmonic functions We use the frequency function as the main tool to derive the doubling condition, as e.g. in [17], [18], [29], [20]. Since we will use the theory also for harmonic functions and the exposition is easier in this special case, let us present it first in this setting. Definition. For a harmonic function u on ball B1 and r < 1, the frequency N (r) is defined as rD(r) , (1.9) N (r) = H(r) where Z D(r) = |∇u|2 dx, Z Br H(r) = u2 dσ, ∂Br 9 where σ is the surface measure. For a harmonic function on a ball B(a, r), N (a, r), D(a, r) and H(a, r) are defined analogously. The frequency is a way how to measure the growth of a harmonic function. If u is a homogeneous harmonic polynomial, its frequency is exactly its degree. See [20] or [22] for more examples. Let us list some important properties. We refer to the survey paper [20] for proofs, although they are known much longer and sketches of the proofs can be found e.g. in [17] or [29]. The following monotonicity property of the frequency function is attributed to F. J. Almgren, Jr., [1]. Proposition 1.2.1. Let u be a harmonic function in B1 . Then N (r) is a nondecreasing function of r ∈ (0, 1). Proof. See [20], Theorem 1.2. Corollary 1.2.2. The limit limr→0 N (r) exists and equals to the vanishing order of u at 0. Proof. See [20], Corollary 1.3. Proposition 1.2.3. Let u be a harmonic function in B1 . For any r ∈ (0, 1), there holds d H(r) N (r) log n−1 = 2 . (1.10) dr r r Integrating (1.10), we obtain that for any 0 < r1 < r2 < 1, there holds Z r2 H(r2 ) H(r1 ) N (r) . = n−1 exp 2 r r2n−1 r1 r1 (1.11) Using the monotonicity of N (Proposition 1.2.1), it follows that H(r2 ) ≤ r2n−1 r2 r1 2N (r2 ) H(r1 ) . r1n−1 (1.12) Proof. See [20], Corollary 1.4. Corollary 1.2.4. Let u be a harmonic function in B1 . Then the function Z r 7→ − u2 (1.13) ∂Br is increasing with respect to r, r ∈ (0, 1), and Z Z 2 − u ≤− Br u2 . (1.14) ∂Br Proof. It follows directly from (1.10) that the function (1.13) is increasing. By integration, we get (1.14). 10 The following result, if we take η = 1/2, is called the doubling condition. It is a counterpart of Corollary 1.2.4. Corollary 1.2.5. Let u be a harmonic function in B1 . For any R, η ∈ (0, 1), there holds Z Z 2 −2N (R) − u ≤η − u2 , (1.15) ∂BR ∂B Z Z ηR − u2 ≤ η −2N (R) − u2 . BR (1.16) BηR Proof. Taking r1 = ηR, r2 = R in (1.12), we obtain H(R) H(ηR) ≤ η −2N (R) , n−1 R (ηR)n−1 which is (1.15). Integrating (1.15) from 0 to R, and using the monotonicity of N (Proposition 1.2.1), we obtain (1.16): Z Z Z RZ Z R −2N (r) 1 2 2 u2 dσ∂Bηr dr η u = u dσ∂Br dr ≤ n−1 η ∂Bηr BR 0 ∂Br 0 Z ηR Z Z 1 1 u2 dσ∂Bρ dρ = η −2N (R) n u2 . ≤ η −2N (R) n η 0 η ∂Bρ BηR Next, we show that not only having a bound on the frequency implies a doubling condition (Corollary 1.2.5), but also knowing a doubling condition to be true implies a bound on the frequency. Lemma 1.2.6. Let u be a harmonic function in Br , r > 0. Let 0 < α < γ < 1, and assume Z Z 2 − u ≥ κ− u2 (1.17) Bαr for some κ > 0. Then N (αr) ≤ Br − log (κ(1 − γ n )) . 2 log (γ/α) In particular, for any β < α, there holds Z Z log(α/β) 2 n log(γ/α) − u ≥ (κ(1 − γ )) − Bβr (1.18) u2 . Bαr Proof. Using Corollary 1.2.4 and the assumption (1.17), we obtain Z rZ Z Z Z 1 u2 dσdρ − u2 ≥ − u2 ≥ κ− u2 ≥ κ n ω r n γr ∂Bρ ∂Bαr Bαr Br Z r n−1 Z Z ρ 1 2 n dρ u = κ(1 − γ )− u2 . ≥κ ωn rn γr γr ∂Bγr ∂Bγr 11 (1.19) Using (1.11) and the monotonicity of N , we further obtain Z γr H(γr)/(γr)n−1 N (ρ) n − log(κ(1 − γ )) ≥ log ≥ 2N (αr) log(γ/α), =2 n−1 H(αr)/(αr) ρ αr which is (1.18). The inequality (1.19) follows from this bound on N (αr) and the doubling condition (1.16): 2N (αr) Z − log(κ(1−γ n )) Z log(γ/α) β β − u2 u2 ≥ − u2 ≥ α α Bαr Bβr Bαr Z log(α/β) = (κ(1 − γ n )) log(γ/α) − u2 . Z − Bαr Next, we recall a result estimating the frequency in a given point by the frequency in a different point. This is an important property whose adaptation we will use to obtain global estimates. Proposition 1.2.7. Let u be a harmonic function in B1 . For any R ∈ (0, 1), there exists a constant N0 = N0 (R) ≪ 1 such that the following holds. If N (0, 1) ≤ N0 , then u does not vanish in BR . If N (0, 1) ≥ N0 , then there holds 1 N p, (1 − R) ≤ CN (0, 1), for any p ∈ BR , 2 where C is a positive constant depending only on n and R. In particular, the vanishing order of u at any point in BR never exceeds c(n, R)N (0, 1). Proof. See [20], Theorem 1.6. 1.2.2 Frequency and doubling condition for solutions of general elliptic equations We will need a generalization of the frequency (1.9) for more general elliptic equations. Let us recall the theory developed in [17] and [18]. In the unit ball B1 ⊂ Rn , consider the equation Lw = div(A(x)∇w) + b(x) · ∇w + c(x)w = 0, (1.20) where A(x) = (aij (x))ni,j=1 is a real symmetric matrix-valued function on B1 , b(x) is a vector valued and c(x) is a scalar function on B1 . We assume (i) there exists α ∈ (0, 1) such that, for every x ∈ B1 and ξ ∈ Rn , α|ξ|2 ≤ n X i,j=1 12 aij (x)ξi ξj ; (1.21) (ii) there exists Γ > 0 such that, for every x, y ∈ B1 , |aij (x) − aij (y)| ≤ Γ|x − y|, i, j = 1, . . . , n; (iii) there exists K > 0 such that X X ||aij ||L∞ (B1 ) + ||bj ||L∞ (B1 ) + ||c||L∞ (B1 ) ≤ K. i,j (1.22) (1.23) j As in [17] and [18], we introduce a Riemannian metric associated with the principal part of the operator L in (1.20). Set ḡij (x) = aij (x)(det A(x))1/(n−2) , where (aij (x)) = A−1 (x). Let (ḡ ij )(x) = (ḡij (x))−1 , r(x)2 = ḡij (0)xi xj , η(x) = ḡ kl (x) ∂r ∂r (x) (x). ∂xk ∂xl Finally introduce a new metric tensor gij (x)dxi ⊗ dxj by defining gij = η(x)ḡij (x). Then gij are Lipschitz functions with a Lipschitz constant depending on Γ in (1.22). Let M be the Riemannian manifold (B1 , gij ). As proved in Section 3 of [3], in the intrinsic geodesic polar coordinates with pole at zero, the metric tensor gij dxi ⊗ dxj becomes dr ⊗ dr + r2 bij (r, θ)dθi ⊗ dθj , where ∂ bij (0, 0) = δij , bij (r, θ) ≤ Λ, i, j = 1, 2, . . . , n − 1, (1.24) ∂r and Λ depends on n, α, Γ. Hence the geodesic ball of radius r and center at x = 0 coincides with the Euclidean ball Br with center at zero. Notice that if A(x) = I (identity matrix), then g(x) = I and b(r, θ) = I in the point (r, θ) corresponding to x. In general, from (1.24) we obtain that there exists r0 > 0 depending on n, α, Γ such that 1 I ≤ (bij (r, θ)) ≤ 2I 2 i.e. for 0 ≤ r ≤ r0 . (1.25) Denote by ∇M and divM , respectively, the intrinsic gradient and divergence on M , ∇M w = n X g ij i,j=1 ∂w ∂ , ∂xi ∂xj n 1 X ∂ √ divM X = √ ( gXi ), g ∂xi i,j=1 13 where (g ij (x)) is the inverse matrix of G = (gij (x)) and g(x) = | det(gij (x))|. Then (1.20) can be rewritten as divM (µ(x)∇M w) + bM (x) · ∇M w + cM (x)w = 0. (1.26) Here µ(x) = η 1−n/2 (x) is a Lipschitz function. Written in polar coordinates, µ(x) = µ(r, θ) satisfies ∂ µ(0, 0) = 1, µ(r, θ) ≤ Λ, ∂r C1 ≤ µ(x) ≤ C2 , (1.27) where C1 , C2 are positive constants depending only on n, α, Γ. Finally √ bM = G(b/ g), √ cM = c/ g. √ Since ( g)−1 is a bounded Lipschitz factor whose bounds depend only on n, α, and Γ, the coefficients bM and cM satisfy the same bounds (1.23) as b, c, up to a multiplicative constant depending on n, α, Γ. 1,2 Notation 1.2.8. For a solution w ∈ Wloc (B1 ) of (1.20) in B1 and 0 < r < 1, let Z H(r) = µw2 dV∂Br Z ∂Br D(r) = µ|∇M w|2 dVM , Br Z I(r) = (µ|∇M w|2 + ubM · ∇M w + cM w2 ) dVM , Br N (r) = rI(r) , H(r) the last quantity being defined only if H(r) > 0. Analogously we define these quantities not only for solutions w on the unit ball B1 = B(0, 1), but also for solutions on any ball B(x0 , r0 ) with center x0 and radius r0 . Then H(r), D(r), I(r) and N (r) also depend on w and x0 , and if the function and the center are not clear from the context, we will denote them Hw (x0 , r), Dw (x0 , r), Iw (x0 , r), Nw (x0 , r) (we will skip either x0 or w if just one is not clear from the context). Note that unlike in the harmonic case, it is not clear anymore where is N (r) defined and whether N (r) > 0: we have D(r) > 0, H(r) ≥ 0, but the sign of I(r) is not obvious. However, it can be shown that at least on some interval, N (r) is defined and N (r)/r ≥ −C for some constant C. 1,2 Lemma 1.2.9. Let w ∈ Wloc (B1 ) be a nonzero solution to (1.20) in B1 . Then there exist constants r0 , C > 0, depending on n, α, Γ, K, such that for any r ∈ (0, r0 ), D(r) ≤ 2I(r) + CH(r), 1 D(r) ≥ I(r) − CH(r), 2 H(r) > 0. 14 (1.28) Proof. The lemma can be found in [22] as Lemma 3.2.3 and Corollary 3.2.5. In [18], the last inequality is Lemma 2.2. and the previous two easily follow from the estimates done in the proof of this lemma. 1,2 Corollary 1.2.10. Let w ∈ Wloc (B1 ) be a nonzero solution to (1.20) in B1 . Then there exist constants r0 , C > 0, depending on n, α, Γ, K, such that for any r ∈ (0, r0 ), N (r) is defined and N (r) ≥ −C. r Proof. Take r0 from Lemma 1.2.9. Then for r ∈ (0, r0 ), H(r) is positive, so N (r) is defined and by (1.28), we obtain N (r) I(r) D(r) = ≥ − C ≥ −C. r H(r) 2H(r) Corresponding to the monotonicity of N in the harmonic case, there holds the following bound in the general elliptic setting. 1,2 Theorem 4. Let w ∈ Wloc (B1 ) be a nonzero solution to (1.20) in B1 . Then there exist constants r0 , c1 , c2 > 0, depending on n, α, Γ, K, such that N (R1 ) ≤ c1 + c2 N (R2 ) for any 0 < R1 < R2 ≤ r0 . (1.29) Proof. This theorem follows from Theorem 2.1 in [18], using our bounds (1.23) instead of their more general assumptions on b and c, and checking the L∞ bound on N in the proof. It can also be found in a more similar form in [22] as Theorem 3.2.1. The only difference in our formulation is that we state the estimate for any R2 ≤ r0 , not just R2 = r0 . However, going through the proofs of the theorems in [18] or [22], one can check that the estimate is true with our formulation without any changes to the proofs. Another way how to verify inequality (1.29) for R2 < r0 if we know that it holds for R2 = r0 is to consider a solution wr0 /R2 = w( Rr20x ) of a scaled equation (which satisfies the assumptions (1.21), (1.22) and (1.23) with the same constants as the original equation), f2 = r0 , and then use the scaling of N explained in the next use (1.29) for wr0 /R2 and R section to deduce (1.29) for w and the original R2 < r0 . Next, we will recall a few results for solutions of general elliptic equations corresponding to Proposition 1.2.3, Corollary 1.2.4 and Corollary 1.2.5. Instead of Proposition 1.2.3, we have the following. 1,2 Proposition 1.2.11. Let w ∈ Wloc (B1 ) be a nonzero solution to (1.20) in B1 . Then H(r) N (r) d log n−1 = O(1) + 2 , (1.30) dr r r 15 where O(1) denotes a function bounded by a constant depending on n, α, Γ and the L∞ bound on the leading order coefficients A. Integrating (1.30), we obtain that for any 0 < R1 < R2 < 1, there holds Z R2 H(R2 ) H(R1 ) N (r) = n−1 exp O(1)(R2 − R1 ) + 2 . (1.31) r R2n−1 R1 R1 Using Theorem 4 (the bound for N ), it follows that if R2 ≤ r0 (where r0 comes from Theorem 4 and depends on n, α, Γ, K), then c1 +c2 N (R2 ) H(R1 ) H(R2 ) (R2 −R1 ) R2 , (1.32) n−1 ≤ C R1 R2 R1n−1 where C, c1 , c2 depend on n, α, Γ, K. Proof. Equation (1.30) can be found in [18] as equation (2.16). The fact that O(1) does not depend on the coefficients b and c can be verified by going back to the proof of this equation. The O(1) is the same as in [17] for equations with no lower order terms. The following corollary corresponds to Corollary 1.2.4. 1,2 Corollary 1.2.12. Let w ∈ Wloc (B1 ) be a nonzero solution to (1.20) in B1 . Then there exist constants r0 , C > 0 depending only on n, α, Γ, K such that Z Z 2 − w ≤ C− w2 for any 0 < s < r < r0 (1.33) ∂Bs and ∂Br Z Z 2 − w ≤ C− Br w2 ∂Br for r ∈ (0, r0 ). (1.34) Proof. Using Corollary 1.2.10, i.e. N (r)/r ≥ −C, and (1.30), we obtain H(r) N (r) d ≥ −C log n−1 = O(1) + 2 dr r r for some constants r0 , C and r < r0 . Hence the Rfunction eCr H(r)/rn−1 is increasing. Next, the function µ in the definition of H(r) = ∂Br µw2 dV∂Br is bounded both from below and above by positive constants ((1.27)), and by (1.25) we can adjust r0 so that the measure dV∂Br , r < r0 is comparable with Hn−1 . Then we easily obtain the estimate (1.33) for the average integrals. By integration, we get (1.34). Next, we state the doubling condition. 1,2 Theorem 5. Let w ∈ Wloc (B1 ) be a nonzero solution to (1.20) in B1 . Then there exist constants r0 , C, c1 , c2 > 0, depending on n, α, Γ, K, such that for any 0 < R1 < R2 < r0 , c1 +c2 N (R2 ) Z Z R2 2 w2 , (1.35) − w ≤C − R 1 ∂BR1 ∂BR2 c1 +c2 N (R2 ) Z Z R2 − w2 ≤ C − w2 . (1.36) R 1 BR BR 2 1 16 Proof. This can be found in a slightly different form in [18] as Theorem 1.2, or inR [22] as Theorem 3.2.7. It also follows from (1.32) using that H(r) is comparable with ∂Br w2 for r small enough. Finally, let us show the equivalent of Lemma 1.2.6: knowing a doubling condition to be true implies a bound on the frequency. 1,2 Lemma 1.2.13. Let w ∈ Wloc (B1 ) be a nonzero solution to (1.20) in B1 . Assume Z Z 2 − w ≥ κ− w2 (1.37) Bζr Br for some κ, ζ ∈ (0, 1) and r ∈ (0, r0 ], where r0 depends on n, α, Γ, K and is chosen so that the previous properties in this subsection hold. Then there exists a constant Cζ > 0 depending on n, α, Γ, K and ζ such that N (ζr) ≤ Cζ (1 − log κ). (1.38) In particular, for any β < ζ, there exist constants C1 , C2 depending on n, α, Γ, K and ζ, β such that Z Z 1 C2 κ − w2 . (1.39) − w2 ≥ C1 Bζr Bβr Proof. We copy the proof of Lemma 1.2.6. Choose γ such that ζ < γ < 1, e.g. γ = 1+ζ 2 . Using Corollary 1.2.12 and the assumption (1.37), we obtain Z Z Z Z rZ 1 1 1 1 2 2 2 − w ≥ − w ≥ κ− w ≥ κ w2 dσdρ n C C C ω r n γr ∂Bρ Bζr Br ∂Bζr Z Z r n−1 Z 1 ρ 1 1 n 2 ≥ κ dρ w = κ(1 − γ )− w2 . C ωn rn γr γr C ∂Bγr ∂Bγr R 2 Using (1.31), Theorem 4 and the comparability of H(r) and ∂Br w , we further obtain H(γr)/(γr)n−1 n C − log(κ(1 − γ )) ≥ log H(ζr)/(ζr)n−1 Z γr N (ρ) = O(1)(γ − ζ)r + 2 ρ ζr 1 ≥ −C + N (ζr) log(γ/ζ), C and (1.38) follows. The inequality (1.39) follows from this bound on N (ζr) and the doubling condition (1.36): f f Z Z Z 1 β C1 −C2 log κ 1 β c1 +c2 N (ζr) 2 2 − w2 − w ≥ − w ≥ C ζ C ζ B Bζr Bβr Z ζr 1 C2 κ − w2 . = C1 Bζr 17 1.3 Doubling Condition for Steklov Eigenfunctions In this section, Ω ⊂ Rn will denote a C 2 domain, and u will be a Steklov eigenfunction corresponding to eigenvalue λ, harmonically extended to Ω, i.e satisfying (1): ∆u = 0 ∂u = λu ∂ν in Ω, on ∂Ω. The doubling condition, i.e. controlling the L2 -norm of u on a ball B2r (x) by the L2 -norm of u on a smaller ball Br (x), is an important property when trying to estimate the nodal set. The main result of this section is the following theorem. Theorem 2. There exist constants r0 , C > 0 depending only on Ω and n such that for any r ≤ r0 /λ and x ∈ ∂Ω, Z Z 2 Cλ5 u2 . (1.40) u ≤2 B(x,r)∩∂Ω B(x,2r)∩∂Ω Remark 1.3.1. The exponent λ5 in (1.40) is not optimal. The proof below essentially gives approximately λ4.6 , but we believe that the optimal scaling is λ. However, even this polynomial bound is good in a problem like ours – by a more direct application of the results in [18], one gets only exponential bound (i.e. eCλ instead of our λ5 ). 1.3.1 Reduction to Neumann boundary condition and reflection across boundary Notation 1.3.2. Since ∂Ω is compact and C 2 , it has bounded curvature and there exists δ > 0 such that the map (y, t) 7→ y + tν(y) is one-to-one from ∂Ω × (−δ, δ) onto δ-neighborhood of ∂Ω. For any ρ ≤ δ, denote {y + tν(y) | y ∈ ∂Ω, t ∈ (−ρ, 0)} = {x ∈ Rn | dist(x, ∂Ω) < ρ} ∩ Ω = Ωρ , {y + tν(y) | y ∈ ∂Ω, t ∈ (0, ρ)} = {x ∈ Rn | dist(x, ∂Ω) < ρ} \ Ω̄ = Ω′ρ . This means that for each x ∈ Ωδ , there exists a unique closest point on the boundary ∂Ω, and for each y ∈ ∂Ω there exists x ∈ Ωδ such that dist(x, y) = δ and B(x, δ) ⊂ Ω. Furthermore, since Ω\Ωδ is connected, for each two points y1 , y2 ∈ ∂Ω and the corresponding points x1 , x2 ∈ Ωδ such that dist(x1 , y1 ) = dist(x2 , y2 ) = δ and B(x1 , δ), B(x2 , δ) ⊂ Ω, there exists a curve Γx1 ,x2 in Ω \ Ωδ with endpoints x1 , x2 . If we look at Γ as a set of points in Ω, then {x ∈ Rn : dist(x, Γ) < δ} ⊂ Ω. We will need these properties later. Fix this δ (depending on Ω) for the rest of this section. We want to extend the function u defined on Ωδ ∪ ∂Ω to Ωδ ∪ ∂Ω ∪ Ω′δ = D, 18 so that the boundary ∂Ω becomes a hypersurface in D. In order to do so, define v(x) := u(x)eλd(x) for x ∈ Ωδ ∪ ∂Ω, where d(x) = dist(x, ∂Ω) is the distance function. Fix this notation throughout the rest of this section. Note that v(x) = 0 if and only if u(x) = 0 and v(x) = u(x) on ∂Ω. Since ∂Ω is C 2 , so is d(x) on Ωδ . For y ∈ ∂Ω and x = y + tν(y) ∈ Ωδ , we have ∇d(x) = ∇d(y) = −ν(y), ∆d(x) = − n−1 X i=1 (1.41) κi (y) , 1 − κi (y)d(x) where {κi } are the principal curvatures of ∂Ω, see the Appendix, pp. 381-383 in [19]. From (1) we get that v satisfies the equation div(A(x)∇v) + b(x) · ∇v + c(x)v = 0 ∂v =0 ∂ν where in Ωδ , on ∂Ω, A=I b = −2λ∇d in Ωδ . c = λ2 − λ∆d (1.42) (1.43) Write each x ∈ Ωδ as x = y + tν(y), y ∈ ∂Ω and t ∈ (0, −δ), and consider the reflection map Ψ : Ωδ → Ω′δ Ψ(x) = y − tν(y). Since ∂Ω is C 2 , Ψ is C 2 . We will also use the notation Ψ(x) = x′ . For x′ ∈ Ω′δ define v(x′ ) := v(Ψ−1 (x′ )) = v(x). ∂v Then v ∈ C 2 (Ω′δ ), and since ∂ν = 0 on ∂Ω, we also have that ∇v is Lipschitz in D, and hence v is twice weakly differentiable in D. By (1.42), v satisfies div(A∇v) + b · ∇v + cv = 0 in Ω′δ , (1.44) with A = (aij )ni,j=1 , aij (x′ ) = n X ∂Ψi k=1 bi (x′ ) = − ∂xk (x) ∂Ψj (x) ∂xk ∂ ij ′ a (x ) + ∆Ψi (x) + ∇Ψi (x) · b(x) ∂x′j c(x′ ) = c(x). 19 (1.45) Hence v satisfies div(A∇v) + b · ∇v + cv = 0 (1.46) a.e. in D, where by (1.43) and (1.45) we have the bounds ||A||L∞ (D) ≤ C, ||b||L∞ (D) ≤ Cλ, (1.47) 2 ||c||L∞ (D) ≤ Cλ , with the constant C depending only on the domain Ω. Since v is also twice weakly differentiable, it satisfies equation (1.46) in the strong sense (as in [19, Chapter 9]). We will see soon that v is also a weak solution of (1.46) (this is not yet clear since we do not know how A behaves across the hypersurface ∂Ω). Note that A(x′ ) = ∇Ψ(x)(∇Ψ)T (x) for x′ ∈ Ω′δ . (1.48) Proposition 1.3.3. A is uniformly Lipschitz in D, with the Lipschitz constant depending on the boundary ∂Ω. Proof. This is clear in Ωδ and Ω′δ ; what we are concerned about is the hypersurface ∂Ω. First let us show that ∇Ψ(x)(∇Ψ)T (x) = I for x ∈ ∂Ω, (1.49) where we take Ψ(x) = x for x ∈ ∂Ω (and consider the gradient of Ψ in points on ∂Ω only from the side of within Ω). Take any x0 ∈ ∂Ω. If the tangent plane to ∂Ω at x0 is parallel to the plane {xn = 0}, then it is easy to see that 1 0 ··· 0 0 0 1 · · · 0 0 .. .. ∇Ψ(x) = ... . . 0 0 · · · 1 0 0 0 · · · 0 −1 and indeed ∇Ψ(x0 )(∇Ψ)T (x0 ) = I. If the tangent plane to ∂Ω at x0 is general, there exists a rotation R which maps this tangent plane to a tangent plane parallel to {xn = 0}. Then by the change of coordinates y = Rx, the mapping Ψ̃(y) = RΨ(RT y) will be exactly like Ψ in the special case discussed, and we will have I = ∇Ψ̃(y0 )(∇Ψ̃)T (y0 ) = R∇Ψ(x0 )RT R(∇Ψ)T (x0 )RT = = R∇Ψ(x0 )(∇Ψ)T (x0 )RT , I = ∇Ψ(x0 )(∇Ψ)T (x0 ) (we are using the rotation property R−1 = RT a few times). 20 Hence (1.49) is true, and we can define A(x) = I = ∇Ψ(x)(∇Ψ)T (x) for x ∈ ∂Ω. Together with (1.43) we get that A is Lipschitz in Ωδ . Since ∂Ω is smooth, so is ∇Ψ and together with (1.45) we get that A is Lipschitz in Ω′δ . Hence, A is Lipschitz in the whole D: for x ∈ Ωδ and y ∈ Ω′δ , the segment from x to y intersects ∂Ω in some z, and then |A(x) − A(y)| ≤ |A(x) − A(z)| + |A(z) − A(y)| ≤ K(|x − z| + |z − y|) = K|x − y|. Remark 1.3.4. Note that if we start with a more general elliptic matrix A than the Laplacian in our domain Ωδ , then after reflexion we do not necessarily end up with a Lipschitz extension of A. In general, the reflected solution will satisfy (1.44) with A(x′ ) = ∇Ψ(x)A(x)(∇Ψ)T (x). For instance, if we consider the problem in R2 and take 1 0 0 1 A(x0 ) = , ∇Ψ(x0 ) = 0 2 1 0 for some x0 ∈ ∂Ω (having the tangent line in x0 parallel to the line x1 − x2 = 0 gives us this ∇Ψ), we get 2 0 T ∇Ψ(x0 )A(x0 )(∇Ψ) (x0 ) = 6= A(x0 ), 0 1 so A will be discontinuous at x0 . Proposition 1.3.5. Equation (1.46) is uniformly elliptic in D. This is clear for all x ∈ Ωδ , and for x ∈ Ω′δ we have by (1.48) ξ T A(x′ )ξ = ξ T ∇Ψ(x)(∇Ψ)T (x)ξ = |(∇Ψ)T (x)ξ|2 ≥ 1 2 |ξ| , η2 where η = sup |(∇Ψ)(x)−1 | = sup |∇Ψ−1 (x′ )| x∈Ωδ x′ ∈Ω′δ depends only on Ω. Proposition 1.3.6. The function v satisfies (1.46) in the weak sense, i.e. for all ϕ ∈ C0∞ (D) we have Z −A∇v · ∇ϕ + b∇vϕ + cvϕ = 0. D 21 Proof. Since v ∈ C 2 (Ωδ ) ∩ C 2 (Ω′δ ) ∩ C 1 (D) and it satisfies (1.46) in Ωδ and Ω′δ pointwise, we have Z (−A∇v · ∇ϕ + b∇vϕ + cvϕ) DZ Z (−A∇v · ∇ϕ + b∇vϕ + cvϕ) = (−A∇v · ∇ϕ + b∇vϕ + cvϕ) + = Z Ωδ (div(A∇v)ϕ + b∇vϕ + cvϕ) + Ωδ + = Z Z ∂Ω Z Ω′δ ∂Ωδ (div(A∇v)ϕ + b∇vϕ + cvϕ) + Ω′δ A∇vϕ · ν − Z ∂Ω Z A∇vϕ · νΩδ ∂Ωδ A∇vϕ · νΩ′δ A∇vϕ · ν = 0. We used that ϕ = 0 on ∂D ⊃ (∂Ωδ \ ∂Ω) ∪ (∂Ω′δ \ ∂Ω), the exterior normals for Ωδ and Ω′δ on ∂Ω satisfy νΩδ = ν = −νΩ′δ , and A is continuous across ∂Ω. Doubling condition for v: To prove Theorem 2, we will first derive an analogous doubling condition on the solid D for v, and then use the approach as in [29] (used for parabolic equations) to deduce Theorem 2: Theorem 7. There exist constants r0 , C > 0 depending only on Ω and n such that for any r ≤ r0 /2λ and x ∈ ∂Ω, Z Z 5 v2. (1.50) v 2 ≤ 2Cλ B(x,r) B(x,2r) The main difficulty we encounter is that for a general elliptic equation, an adaptation of the results in [18] gives us a bound for the frequency only on small balls of radius of order 1/λ. That globally translates into the suboptimal exponential bound mentioned in Remark 1.3.1. Therefore we move on from the more general equation (1.46), which is satisfied by v in the neighborhood of ∂Ω, back to the original Laplace equation inside of Ω, and then back to the neighborhood of the boundary with the equation (1.46) again. 1.3.2 Frequency for v Near the boundary ∂Ω, we will consider the equation (1.46). Hence we use the generalization of the frequency (1.9) for more general elliptic equations from Subsection 1.2.2. Let us apply the results from that subsection to our function v. Recall Theorem 4: for a solution w to a general elliptic equation on B1 , it gives us the bound for the frequency of the form N (R1 ) ≤ c1 + c2 N (R2 ), if 0 < R1 < R2 ≤ r0 . However, it does not give an explicit bound for r0 , c1 , c2 in terms of the L∞ bounds on the coefficients of the equation. If we use the bounds (1.47) instead of (1.23) and go 22 through the proof of the theorem, we obtain that (1.29) holds for r0 ∼ 1/λ, c1 , c2 ∼ 1. We can obtain these bounds easier from Theorem 4 by scaling. Scaling the equation: Recall that the function v satisfies the equation (1.46): div(A∇v) + b · ∇v + cv = 0 a.e. in D, with A elliptic and Lipschitz, the ellipticity and Lipschitz constants depending only on Ω. By (1.47) we have ||A||L∞ (D) ≤ C, ||b||L∞ (D) ≤ Cλ, ||c||L∞ (D) ≤ Cλ2 , with the constant C depending only on the domain Ω. Consider this equation in a ball B(x0 , r1 /λ) ⊂ D and define vx0 ,λ (x) := v(x0 + x/λ) for x ∈ B(0, r1 ). Then vx0 ,λ satisfies div(Ax0 ,λ ∇vx0 ,λ ) + bx0 ,λ · ∇vx0 ,λ + cx0 ,λ vx0 ,λ = 0 a.e. in B(0, r1 ), with (1.51) Ax0 ,λ (y) = A(x0 + y/λ), bx0 ,λ (y) = λ−1 b(x0 + y/λ), cx0 ,λ (y) = λ−2 c(x0 + y/λ), so Ax0 ,λ (y), bx0 ,λ (y) and cx0 ,λ (y) are bounded uniformly in L∞ by a constant depending only on Ω. The ellipticity constant of A does not change and the Lipschitz constant of A only improves (since λ ≥ 1), so we also have bounds on them depending only on Ω. A simple change of variables yields that for y ∈ B(0, r1 ) and r ≤ r1 − |y|, Hvx0 ,λ (y, r) = λn−1 Hv (x0 + y/λ, r/λ), Dvx0 ,λ (y, r) = λn−2 Dv (x0 + y/λ, r/λ), Ivx0 ,λ (y, r) = λn−2 Iv (x0 + y/λ, r/λ), Nvx0 ,λ (y, r) = Nv (x0 + y/λ, r/λ). Remark 1.3.7. Notice that for x ∈ Ω ∩ D, the principal part of equation (1.46) is the Laplacian, i.e. A = I. Hence, in points from Ω ∩ D, the metric g associated with A is trivial, and in these points we have Z Hv (r) = v2, Z ∂Br Dv (r) = |∇ v|2 , Br Z Iv (r) = (|∇ v|2 + ub · ∇ v + cv 2 ). Br 23 In particular, we can skip the argument about H(r) and they are equal in this case. R ∂Br v 2 being comparable – Next, we apply Theorem 4, Proposition 1.2.11, Corollary 1.2.12, Theorem 5 and Lemma 1.2.13 to the function vx0 ,λ , and using the scaling above we rewrite the results in terms of v. We immediately obtain the following results. Proposition 1.3.8. Let B(x0 , r1 /λ) ⊂ D. Then there exist constants c1 , c2 > 0, r0 ∈ (0, r1 ) depending only on r1 and Ω such that Nv (x0 , R1 ) ≤ c1 + c2 Nv (x0 , R2 ) for any 0 < R1 < R2 ≤ r0 /λ. (1.52) Proposition 1.3.9. Let B(x0 , r1 /λ) ⊂ D. Then there exist constants C, c1 , c2 > 0, r0 ∈ (0, r1 ) depending only on r1 and Ω such that for any 0 < R1 < R2 < r0 /λ, Z R2 c1 +c2 Nv (x0 ,R2 ) v ≤C v2, − R1 ∂B(x0 ,R2 ) ∂B(x0 ,R1 ) c1 +c2 Nv (x0 ,R2 ) Z Z R2 v2 ≤ C − v2. − R 1 B(x0 ,R2 ) B(x0 ,R1 ) Z − 2 Lemma 1.3.10. Let B(x0 , r1 /λ) ⊂ D. Assume Z Z 2 v ≥ κ− − v2 (1.53) (1.54) (1.55) B(x0 ,r) B(x0 ,ζr) for some κ, ζ ∈ (0, 1) and r ∈ (0, r0 /λ], where r0 depends on r1 and Ω and is chosen so that Propositions 1.3.8 and 1.3.9, and the analogies of Proposition 1.2.11 and Corollary 1.2.12 hold. Then there exists a constant Cζ > 0 depending on r1 , Ω and ζ such that Nv (x0 , ζr) ≤ Cζ (1 − log κ). (1.56) In particular, for any β < ζ, there exist constants C1 , C2 depending on r1 , Ω and ζ, β such that Z Z 1 C2 v2 ≥ v2. (1.57) − κ − C 1 B(x0 ,βr) B(x0 ,ζr) 1.3.3 Proof of Theorem 7 We will use an argument of the type as in Proposition 1.2.7 in a chain. For that, we will need to start in a point where the frequency is reasonably bounded/where we have a doubling condition with a reasonable constant. We will require even somewhat more: the integral of u2 over a small ball with center in this special starting point y∗ ∈ ∂Ω will control the global integral of u2 . Recall the Notation 1.3.2: δ > 0 is a constant depending on Ω such that each point in Ωδ = {x ∈ Rn | dist(x, ∂Ω) < δ} ∩ Ω has a single closest point on ∂Ω. 24 Lemma 1.3.11. For any ρ < δ, there exists a point y∗ ∈ ∂Ω such that for some constant C depending only on Ω, there holds Z Z 2 −2n+1 u2 . u ≤ Cρ B(y∗ ,ρ)∩Ω Ω In the proof we crucially use this interior estimate for harmonic functions: Proposition 1.3.12. Let w be a harmonic function in B1 . Then sup |w| ≤ C B1/2 Z w 2 B1 1/2 , where C is a constant depending only on n. Proof. This can be found e.g. in [24], Remark 1.19. Proof of Lemma 1.3.11. Choose a ρ/2-net of points y1 , y2 , . . . , ym ∈ ∂Ω, i.e. for each y ∈ ∂Ω there exists 1 ≤ i ≤ m such that |y − yi | ≤ ρ/2. We can always make m ≤ Cρ−(n−1) , where C depends only on Ω. Then the balls {B(yi , ρ)}m i=1 cover Ωρ/2 , and therefore there exists y∗ ∈ {y1 , y2 , . . . , ym } such that Z Z u2 . (1.58) u2 ≤ Cρ−(n−1) Ωρ/2 Now we just need to bound R Ωu B(y∗ ,ρ)∩Ω 2 in terms of R Ωρ/2 u2 . Let n ρo = Ω \ Ωρ . Ωcρ/4 = x ∈ Ω : dist(x, ∂Ω) ≥ 4 4 Then ∂Ωcρ/4 = {x ∈ Ω : dist(x, ∂Ω) = ρ/4} and for any x ∈ ∂Ωcρ/4 , by the scaling of the interior estimate in Proposition 1.3.12 we obtain Z Z sup u2 ≤ Cρ−n u2 ≤ Cρ−n u2 . B(x,ρ/4) B(x,ρ/8) Ωρ/2 Hence, by the maximum principle for the harmonic function u inside Ωcρ/4 , 2 2 sup u ≤ sup u ≤ Ωcρ/4 Z ∂Ωcρ/4 2 u = Ω Z sup u + Ω\Ωρ/2 ≤ Cρ−n sup u ≤ Cρ −n x∈∂Ωcρ/4 B(x,ρ/8) 2 Z 2 Z 2 Ωρ/2 2 u ≤ C sup u + Ωcρ/4 u2 . Ωρ/2 Combining this with (1.58), we obtain the desired estimate. 25 Z Z u2 , Ωρ/2 u2 Ωρ/2 Proof of Theorem 7. Fix r1 := δλ1 , where λ1 is the smallest positive Steklov eigenvalue for Ω. Then B(x0 , r1 /λ) ⊂ D for every x0 ∈ ∂Ω and every Steklov eigenvalue λ. Fix r0 < r1 so that Propositions 1.3.8 and 1.3.9 and Lemma 1.3.10 hold with this r0 . Note that r1 and r0 depend only on Ω. To prove the Theorem, it is enough to prove that there exists a constant C depending only on Ω and n such that for every y ∈ ∂Ω Z Z 2 Cλ5 v2. (1.59) v ≤2 B(y,r0 /2λ) B(y,r0 /λ) Indeed, this will prove (1.50) for r = r0 /2λ. Once we know it for this r, from Lemma 1.3.10 we obtain a bound for the frequency N (r0 /2λ) ≤ Cλ5 , and from Propositions 1.3.8 and 1.3.9 we obtain (1.50) for any r ≤ r0 /2λ. The rest of the proof is organized as follows. First, we prove a doubling condition for v in a special point y = y∗ ∈ ∂Ω, move to a close-by point inside Ω, and switch to u. As the next step, we show a doubling condition for u in a point near y∗ for a ball with fixed radius (independent of λ), and propagate the doubling condition estimate through Ω into a neighborhood of any other given point y0 ∈ ∂Ω. In step 3, we gradually pass from the ball with fixed radius to a small ball in ∼ 1/λ-neighborhood of ∂Ω, still for u and within Ω. In the last step, we switch back to v and deduce the doubling condition (1.59) in the point y0 ∈ ∂Ω. Step 1: We start in a point y = y∗ ∈ ∂Ω, which we obtain from Lemma 1.3.11 for ρ = r0 /2λ, i.e. there holds Z r −2n+1 Z 0 2 u ≤C u2 . 2λ Ω B(y∗ ,r0 /2λ)∩Ω In Ω, we have |v(x)| = |u(x)eλd(x) | ≥ |u(x)|. Hence Z Z Z 1 r0 2n−1 2 2 u2 . u ≥ v ≥ C 2λ Ω B(y∗ ,r0 /2λ)∩Ω B(y∗ ,r0 /2λ) (1.60) R R On the other hand, we show that the integral B(y,r0 /λ) v 2 is controlled by Ω u2 for any y ∈ ∂Ω. Fix y ∈ ∂Ω. Notice that r0 /λ < δ. Recall that the reflection map Ψ introduced in Subsection 1.3.1 maps any x ∈ Ωδ , written as x = z + tν(z), where z ∈ ∂Ω, into Ψ(x) = z − tν(z), and this is a 1-to-1 map from Ωδ onto Ω′δ = D \ Ω. Since B(y, r0 /λ) \ Ω ⊂ Ω′r0 /λ ⊂ Ω′δ , the map Ψ−1 maps B(y, r0 /λ) \ Ω onto a subset of Ωr0 /λ , and by change of variables, using the boundedness of the Jacobian of Ψ−1 (the bound depends only on Ω), we obtain Z Z 2 v ≤C v2. Ωr0 /λ B(y,r0 /λ)\Ω On Ωr0 /λ we have |v(x)| = |u(x)eλd(x) | ≤ |u(x)|er0 ≤ C|u(x)|, 26 (1.61) with C depending only on Ω. Hence, Z Z 2 v = B(y,r0 /λ) 2 v + B(y,r0 /λ)∩Ω ≤C Z 2 Ωr0 /λ u ≤C Z Z 2 B(y,r0 /λ)\Ω v ≤C Z v2 Ωr0 /λ (1.62) 2 u . Ω Now we just need to propagate the estimate (1.60) into any point y ∈ ∂Ω – combined with (1.62), we will obtain a doubling condition for v. Using (1.60) and (1.62) in y = y∗ , we know that Z Z 1 1 2 v2. v ≥ C λ2n−1 B(y∗ ,r0 /λ) B(y∗ ,r0 /2λ) Hence, from Lemma 1.3.10 used for r = rλ0 , ζ = 21 , κ = C1 λ−2n+1 , β = 41 , we get Z Z 1 1 2 v ≥ v2, (1.63) C1 λC2 B(y∗ ,r0 /2λ) B(y∗ ,r0 /4λ) where C1 , C2 depend only on Ω. Now fix y = y0 ∈ ∂Ω in which we want to prove (1.59). Recalling Notation 1.3.2, we know that there exist points z1 , z2 ∈ Ω such that dist(z1 , y∗ ) = dist(z2 , y0 ) = δ, B(z1 , δ), B(z2 , δ) ⊂ Ω, and there is a curve Γ in Ω with endpoints z1 , z2 such that if we look at Γ as a set of points in Ω, then {x ∈ Rn : dist(x, Γ) < δ} ⊂ Ω. We will propagate the doubling condition estimate along this curve. First choose x1 on the segment y∗ z1 such that dist(y∗ , x1 ) = r0 /4λ. Then B(x1 , r0 /4λ) lies inside Ω and touches ∂Ω in y∗ . We also have B(y∗ , r0 /4λ) ⊂ B(x1 , r0 /2λ) ⊂ B(x1 , 3r0 /4λ) ⊂ B(y∗ , r0 /λ). Hence, using (1.63) and (1.60), Z B(x1 ,r0 /2λ) 2 v ≥ Z v2 B(y ,r /4λ) ∗ 0 Z 1 1 v2 ≥ C1 λC2 B(y∗ ,r0 /2λ) Z 1 1 ≥ u2 . C 1 λC 2 Ω (1.64) On the other hand, from B(x1 , 3r0 /4λ) ⊂ B(y∗ , r0 /λ) and (1.62) (used for y = y∗ ), we obtain Z Z Z u2 , v2 ≤ C v2 ≤ B(y∗ ,r0 /λ) B(x1 ,3r0 /4λ) so together with (1.64) we have Z 1 1 v ≥ C 1 λC 2 B(x1 ,r0 /2λ) 2 27 Z Ω v2. B(x1 ,3r0 /4λ) Hence it follows from Lemma 1.3.10 used in the center x1 for r = 3r0 /4λ, ζ = 2/3, κ = C11 λC1 2 , β = 1/6, that Z B(x1 ,r0 /8λ) v2 ≥ 1 1 f1 λCf2 C Z B(x1 ,r0 /2λ) v2 ≥ 1 1 C 1 λC 2 Z u2 , Ω where in the last inequality we used (1.64). Notice that starting from the estimate (1.60) for a ball with center y∗ ∈ ∂Ω, we moved to an estimate for a ball B(x1 , r0 /8λ) inside Ω. Since B(x1 , r0 /8λ) ⊂ Ωr0 /λ , using (1.61) we can move from v towards the harmonic function u: Z Z Z 1 1 1 u2 ≥ v2 ≥ u2 . (1.65) C2 C C λ 1 B(x1 ,r0 /8λ) B(x1 ,r0 /8λ) Ω Step 2: Now we are ready to move towards balls with fixed radii which will not depend on λ. Take the point z1 for which B(z1 , δ) lies inside Ω and touches ∂Ω in y∗ . Since x1 lies on the segment y∗ z1 in distance r0 /4λ from y∗ , we have B(x1 , r0 /8λ) ⊂ B(z1 , δ − r0 /8λ) ⊂ B(z1 , δ) ⊂ Ω. Hence, by (1.65) we obtain Z Z 2 u ≥ 1 1 u ≥ C C 1λ 2 B(x1 ,r0 /8λ) B(z1 ,δ−r0 /8λ) 2 Z 1 1 u ≥ C C 1λ 2 Ω 2 Z u2 . (1.66) B(z1 ,δ) Now we can use Lemma 1.2.6 for the harmonic function u on the ball B(z1 , δ) and r0 r0 α = 1 − 8λδ , γ = 1 − 16λδ , κ = C11 λC1 2 , β = 1/2. For this choice, we have r0 r0 n 1 ≤1− γn = 1 − =1− , 16λδ 16λδ Cλ 1 1 − γn ≥ , Cλ r0 r0 1 − 16λδ r0 γ = > e 32λδ , r0 ≥ 1 + α 1 − 8λδ 16λδ 1 log(γ/α) > , Cλ log(α/β) ≤ log 2, and Lemma 1.2.6 together with (1.66) yield Z z1 2 B(z1 ,δ/2) u ≥ 1 1 C 1 λC 2 Cλ e Z Cλ log λ Z 1 u ≥ u2 . 2 B(z1 ,δ−r0 /8λ) Ω 2 (1.67) Now we propagate the estimate (1.67) along Γ into the point z2 . We choose points = z1 , z 2 , . . . , z k = z2 on Γ such that dist(z i , z i+1 ) ≤ δ/4 for i = 1, 2, . . . , k − 1. 28 Notice that k is bounded by a constant depending only on Ω. Using induction, assume Z Ci λ log λ Z 1 u ≥ u2 2 i B(z ,δ/2) Ω 2 (1.68) for some constant Ci depending only on Ω – for i = 1, this is (1.67). Then Z Ci λ log λ Z 1 u2 u ≥ 2 i i B(z ,δ) B(z ,δ/2) 2 and the inclusion B(z i+1 , δ/2) ⊃ B(z i , δ/4), Lemma 1.2.6 and (1.68) imply Z Z 2 u2 u ≥ B(z i ,δ/4) B(z i+1 ,δ/2) Cλ log λ Z Ci+1 λ log λ Z 1 1 2 ≥ u ≥ u2 2 2 i B(z ,δ/2) Ω (here C comes from Lemma 1.2.6 and depends on Ci ), which is (1.68) for i + 1 instead of i. After k − 1 steps, the constant Ck will be bounded by a constant depending only on Ω, and we obtain Cλ log λ Z Z 1 2 u2 . (1.69) u ≥ 2 Ω B(z2 ,δ/2) Step 3: Now we need to get from the ball B(z2 , δ/2) closer to the boundary, where we will be able to switch back to the function v. In each step, we multiply the distance from y0 by a factor of 5/6. Hence, we will need ∼ log λ steps to get to a neighborhood of ∂Ω of order 1/λ. Recall that B(z2 , δ) lies inside Ω and touches ∂Ω in y0 . Consider the sequence i 0 x , x1 , . . . , xl of points on the segment z2 y0 such that dist(xi , y0 ) = δ 65 (i.e. x0 = z2 ) r0 and l is chosen to be the smallest integer such that dist(xl , y0 ) ≤ 4λ , i.e. l= log λ + C, log 6/5 4δ−log r0 4δ−log r0 ≤ C < log log + 1. Denote ri := δ where log log 6/5 6/5 lies in Ω and touches ∂Ω in y0 . By induction, we will show that Z 5 i 6 , so that the ball B(xi , ri ) C2i λ log λ Z 1 2i −1 u ≥ u2 , τ 2 B(xi ,r i /6) Ω 2 (1.70) 25 n where τ = 24 − 1. For i = 0, this follows from (1.69) and Lemma 1.2.6 used for r = δ, α = 1/2, β = 1/6. 29 Now assume that (1.70) holds for some i ∈ {0, 1, . . . , l − 1}, and we will show it for i+1. Since dist(xi , xi+1 ) = ri /6, we have B(xi+1 , 2ri+1 /5) = B(xi+1 , ri /3) ⊃ B(xi , ri /6). On the other hand, we have B(xi+1 , ri+1 ) ⊂ Ω. Hence, (1.70) implies Z C2i λ log λ Z 1 2i −1 u ≥ u ≥ τ u2 2 i i i+1 i+1 B(x ,r /6) B(x ,2r /5) Ω C2i λ log λ Z 1 i τ 2 −1 u2 . ≥ 2 B(xi+1 ,r i+1 ) 2 Z 2 (1.71) Apply Lemma 1.2.6 in point xi+1 for r = ri+1 , α = 2/5, β = 1/6, γ = 24/25, κ = i 5 n 1 C2 λ log λ 2i −1 τ . This is chosen so that the exponent in (1.19) is one: 2 2 log(α/β) = 1, log(γ/α) and our computation becomes much easier. We obtain Z u2 B(xi+1 ,r i+1 /6) ! n Z n C2i λ log λ 24 5 1 (1/6)n i 2 −1 u2 τ 1− ≥ n (2/5) 2 2 25 B(xi+1 ,2r i+1 /5) C2i+1 λ log λ Z 1 2i+1 −1 ≥ τ u2 2 i+1 i+1 B(x ,r ) (the last inequality follows from (1.71)), which is (1.70) for i + 1. Using (1.70) for i = l, we obtain Z C2l λ log λ Z 1 2l −1 τ u2 u ≥ 2 l l Ω B(x ,r /6) 2 log 2 Cλ1+ log(6/5) log λ Z 1 u2 ≥ 2 Ω Cλ5 Z 1 u2 . ≥ 2 Ω (1.72) r0 Step 4: Recall that l was chosen so that rl = dist(xl , y0 ) ≤ 4λ . Hence, B(y0 , r0 /2λ)∩ l l Ω ⊃ B(x , r /6), and using that |v(x)| ≥ |u(x)| in Ωδ and (1.72), we obtain Cλ5 Z 1 u2 . u ≥ u ≥ v ≥ 2 Ω B(xl ,r l /6) B(y0 ,r0 /2λ)∩Ω B(y0 ,r0 /2λ) R Together with (1.62) used for y = y0 which bounds B(y0 ,r0 /λ) v 2 by a constant times R 2 Ω u , we obtain the desired doubling condition (1.59) in y = y0 , what we wanted. Z 2 Z 2 30 Z 2 Remark 1.3.13. In Step 3, instead of our explicit choice for decreasing the distance to the boundary by the multiplicative factor of 5/6, the radius of the smaller ball in control being 1/6 of the distance to the boundary, and the auxiliary constant γ = 24/25, we can do the calculation with general constants. While the computation becomes more log 2 complicated, we do not gain too much: instead of our exponent 1 + log(6/5) ≈ 4.8, a numerical calculation suggests that the optimal choice of the constants gives an exponent of roughly 4.6. 1.3.4 Doubling condition on ∂Ω In this section, we will use Theorem 7 to prove Theorem 2: Theorem 2. There exist constants r0 , C depending only on Ω and n such that for any r ≤ r0 /λ and x ∈ ∂Ω, Z Z 2 Cλ5 u2 . (1.40) u ≤2 B(x,r)∩∂Ω B(x,2r)∩∂Ω As a connection between the boundary integrals and integrals over solid balls, we will use a quantitative Cauchy uniqueness theorem which was proved in [29]. In [29] it is formulated for solutions of general elliptic equations under the same assumptions as used in Subsection 1.2.2: consider the equation n X (aij (x)wxi )xj + in B1 (0) ⊂ bi (x)wxi + c(x)w = 0 (1.73) i=1 i,j=1 Rn n X with coefficients satisfying the following assumptions: aij (x)ξi ξj ≥ α|ξ|2 , ∀ξ ∈ Rn , x ∈ B1 (0); X X (ii) |aij (x)| + |bi (x)| + |c(x)| ≤ K, ∀x ∈ B1 (0); (i) i,j (iii) X i,j i |aij (x) − aij (y)| ≤ γ|x − y|, (1.74) ∀x, y ∈ B1 (0); for some positive constants α, K, γ. Lemma 1.3.14 ([29], Lemma 4.3). Let w be a solution of (1.73) in B1+ ⊂ Rn with the coefficients of the equation satisfying (1.74) and ||w||L2 (B + ) ≤ 1. Suppose that 1 ||w||H 1 (Γ) + ||(∂w/∂xn )||L2 (Γ) ≤ ǫ ≪ 1, where Γ = {(x′ , 0) ∈ Rn : |x′ | < 3/4}. Then ||w||L2 (B + 1/2 constants C, β which depend only on n, α, K, γ. ) ≤ Cǫβ for some positive Notice that Lemma 1.3.14 relates the H 1 -norm on a hypersurface to the L2 -norm in the solid ball. For a doubling condition, we need the L2 -norm also on the hypersurface. The following lemma gives us the missing connection between the H 1 -norm on 31 the hypersurface and the L2 -norm on the hypersurface; the additional H 2 -norm on the solid will be bounded using interior elliptic estimates. Technically it is a combination of a J. L. Lions-type lemma estimating the H 1 (Rn−1 )-norm in terms of the H 3/2 (Rn−1 )-norm and the L2 (Rn−1 )-norm, and the standard trace theorem. Lemma 1.3.15. Let w ∈ H 2 (Rn ) and consider the trace of w onto {x ∈ Rn : xn = 0} = Rn−1 , which we denote by w. Then there exists a constant C depending only on n such that for any η > 0, C ||∇w||L2 (Rn−1 ) ≤ η||w||H 2 (Rn ) + 2 ||w||L2 (Rn−1 ) . (1.75) η Proof. First we will show that for any w ∈ H 3/2 (Rn−1 ), η > 0, there holds ||∇w||L2 (Rn−1 ) ≤ η||w||Ḣ 3/2 (Rn−1 ) + 1 ||w||L2 (Rn−1 ) . η2 (1.76) Using the Fourier transform ŵ(ξ) = 1 (2π)(n−1)/2 Z w(x)e−ix·ξ dx, Rn−1 Plancherel identity and Young’s inequality |ξ|2 ≤ 23 η 2 |ξ|3 + 31 η14 , we have Z 2 |iξ ŵ|2 dξ ||∇w||L2 (Rn−1 ) = n−1 R Z Z 2 2 1 1 3 2 ≤ η |ŵ|2 dξ |ξ| |ŵ| dξ + 3 3 η 4 Rn−1 Rn−1 1 1 2 ||w||2L2 (Rn−1 ) , = η 2 ||w||2Ḣ 3/2 (Rn−1 ) + 3 3 η4 and (1.76) follows. Using the trace theorem H 3/2 (Rn−1 ) ⊂ H 2 (Rn ) (continuously), for any w ∈ H 2 (Rn ) we obtain from (1.76) ||∇w||L2 (Rn−1 ) ≤ Cη||w||H 2 (Rn ) + 1 ||w||L2 (Rn−1 ) . η2 Using η/C instead of η, we obtain (1.75). Proof of Theorem 2. Since u = v on ∂Ω, we can rewrite (1.40) (which we want to prove for r small enough) as Z Z 2 Cλ5 v2. (1.77) v ≤2 B(x,r)∩∂Ω B(x,2r)∩∂Ω From Theorem 7, we know that there exist constants r0 , C depending only on Ω and n such that for any r ≤ r0 /2λ and x ∈ ∂Ω, Z Z 2 Cλ5 v2. (1.78) v ≤2 B(x,r) B(x,2r) 32 Fix this r0 and also fix the point x = x0 ∈ ∂Ω in which we want to prove (1.77). In order to have coefficients of the underlying equation bounded independently of λ as in (1.74), we use the scaling from Section 1.3.2: vx0 ,λ (x) = v(x0 + x/λ), x ∈ B(0, r0 ). Then vx0 ,λ satisfies (1.51) with the coefficients Ax0 ,λ (y), bx0 ,λ (y) and cx0 ,λ (y) bounded uniformly in L∞ by a constant depending only on Ω, i.e. they satisfy (1.74) with α, K, γ independent of λ (the Lipschitz constant γ only improves with the scaling, ellipticity α stays the same). Proving Theorem 2 is equivalent to proving, for any r < r1 (where r1 is to be determined, depending on Ω), Z Z 2 Cλ5 vx0 ,λ ≤ 2 vx20 ,λ , (1.79) B(0,2r)∩∂Ωx0 ,λ B(0,r)∩∂Ωx0 ,λ where Ωx0 ,λ = {x : x0 + x/λ ∈ Ω}. From (1.78), we know that for any r ≤ r0 /2, Z Z 2 Cλ5 vx20 ,λ . (1.80) vx0 ,λ ≤ 2 B(0,r) B(0,2r) Since we want to use the lemmas above, we need to flatten out the hypersurface ∂Ω. Since Ω is a C 2 -domain, we can assume that in B(x0 , r0 /λ) it is a graph of a C 2 -function whose C 2 -norm is bounded by a constant M independent of x0 and λ – otherwise we just diminish r0 . After scaling by λ, this bound only improves. Hence, we can assume that B(0, r0 ) ∩ ∂Ωx0 ,λ = {x = (x′ , xn ) ∈ B(0, r0 ) : xn = Φ(x′ )}, where Φ ∈ C 2 (B n−1 (0, r0 )), Φ(0) = 0, ∇Φ(0) = 0, ||Φ||C 2 ≤ M , and hence (by possibly diminishing r0 ) ||Φ||C 1 (B n−1 (0,r0 )) ≤ ǫ, (1.81) where ǫ ≤ 0.1 will be chosen later. Now define the injective map F : B n−1 (0, r0 )×R → Rn by F (x′ , xn ) := (x′ , xn + Φ(x′ )). Using (1.81) one can easily compute that F (B(0, r)) ⊂ B(0, (1 + ǫ)r) for any r ≤ r0 . (1.82) The inverse F −1 (x′ , xn ) = (x′ , xn − Φ(x′ )) is well-defined on B(0, r0 ) and similarly, F (B(0, r)) ⊃ B(0, r/(1 + ǫ)) for any r ≤ r0 . (1.83) Hence we have for any r < r0 /(1 + ǫ) that F (B n−1 (0, r/(1 + ǫ)) × {0}) ⊂ B(0, r) ∩ ∂Ωx0 ,λ ⊂ F (B n−1 (0, (1 + ǫ)r) × {0}). (1.84) Therefore, to prove (1.79), it is enough to show Z Z 5 vx20 ,λ ≤ 2Cλ F (B n−1 (0,2r)×{0}) 33 F (B n−1 (0,r)×{0}) vx20 ,λ (1.85) for all r ≤ r1 for some r1 ≤ r0 /2(1 + ǫ) depending only on Ω (we can use this doubling 2 2 2 ≥ 1.12 2 > condition twice to account for the loss of the factor (1+ǫ)2 in radii: (1+ǫ) 2 2). Now define w(x) := vx0 ,λ (F (x)) for x ∈ B(0, r0 /(1 + ǫ)). We apply area formula to both sides of (1.85) under the map F |B n−1 (0,2r)×{0} . The (generalized) Jacobian of this map will be q q p T J = det((In−1 , ∇Φ)(In−1 , ∇Φ) ) = det(In−1 + (∇Φ)(∇Φ)T ) = 1 + |∇Φ|2 , so using (1.81), we get 1 ≤ J ≤ 1 + ǫ ≤ 1.1. Hence, to prove (1.85), it is enough to prove Z Z 2 Cλ5 w2 (1.86) w ≤2 B n−1 (0,r)×{0} B n−1 (0,2r)×{0} for all r ≤ r1 for some r1 ≤ r0 /2(1 + ǫ) depending only on Ω. Using the same arguments ((1.82), (1.83), using the doubling condition twice and the Jacobian of F being 1), it follows from (1.80) that Z Z B(0,2r) w2 ≤ 2Cλ 5 w2 (1.87) B(0,r) for any r ≤ r0 /2(1 + ǫ). Using that vx0 ,λ solves the equation (1.51), we obtain that w = vx0 ,λ ◦ F solves div(Ã∇w) + b̃ · ∇w + c̃w = 0, (1.88) where Ã(x) = Ax0 ,λ (F (x)) + P (F (x)), b̃(x) = bx0 ,λ (F (x)) + Q(F (x)), c̃(x) = cx0 ,λ (F (x)). The perturbation P is symmetric and for x = (x′ , xn ), P ij (x) = 0 P in (x) = − P nn n−1 X for i, j = 1, . . . , n − 1, aij (x)Φxj (x′ ) for i = 1, . . . , n − 1, j=1 (x) = −2 n−1 X in ′ a (x)Φxi (x ) + n−1 X aij (x)Φxi (x′ )Φxj (x′ ). i,j=1 i=1 For the perturbation Q we have Qi (x) = n−1 X ′ aij xn (x)Φxj (x ) j=1 Qn (x) = − n−1 X i=1 bi (x)Φxi (x′ ) + for i = 1, . . . , n − 1, n−1 X i=1 ′ ain xn (x)Φxi (x ) − 34 n−1 X i,j=1 ′ ′ aij xn (x)Φxi (x )Φxj (x ). Hence, using the bound (1.81) on the derivatives of Φ, we have ||P (x)||L∞ ≤ CKǫ, ||Q(x)||L∞ ≤ C(K + γ)ǫ, and the coefficients of equation (1.88) will satisfy L∞ -bounds depending only on Ω, the Lipschitz constant of à as well, and if we choose ǫ small enough, we also get the ellipticity of à depending only on Ω (or even an absolute bound like α̃ ≥ 1/2). Now we are ready to prove (1.86) for r small enough. Fix r ≤ r0 /2(1 + ǫ), so that (1.87) holds. We use scaling again so that we move to balls of fixed radii: denote w̃(x) := c̃w(rx), where the constant c̃ is chosen so that Z x ∈ B(0, 2), (1.89) w̃2 = 1. (1.90) B(0,2) Using the doubling condition (1.87) twice, we obtain Z 5 w̃2 ≥ 2−Cλ . (1.91) B(0,1/2) Note that w̃ satisfies an equation of type (1.73) with the coefficients satisfying (1.74) with α, K,γ depending only on Ω: compared to the equation (1.88) satisfied by w, after the scaling, as we have seen before, the L∞ bounds of the lower-order coefficients only improve, while the leading order coefficients keep the L∞ -bounds and ellipticity and improve the Lipschitz constant. Using Lemma 1.3.14, we will show that ||w̃||H 1 (B n−1 (0,3/4)) + ||(∂ w̃/∂xn )||L2 (B n−1 (0,3/4)) ≥ 2−Cλ 5 (1.92) with a different constant C, which still depends only on Ω. Take the constants C = C1 , β from Lemma 1.3.14 used for w̃ which only depend on n and Ω. Then if we have ||w̃||H 1 (B n−1 (0,3/4)) + ||(∂ w̃/∂xn )||L2 (B n−1 (0,3/4)) = ǫ̃ ≪ 1, then Lemma 1.3.14 and (1.90) imply ||w̃||L2 (B(0,1/2)) ≤ C1 ǫ̃β . Since we have the lower bound (1.91), it follows that −1/β −Cλ5 /2β ǫ̃ ≥ C1 2 , so we get (1.92) with the constant (C + 2 log2 C1 )/2β. Now we will show that ||∇w̃||L2 (B n−1 (0,3/4)) ≥ 1 ||(∂ w̃/∂xn )||L2 (B n−1 (0,3/4)) , C 35 (1.93) and hence from (1.92) we will get a lower bound for ||w̃||H 1 (B n−1 (0,3/4)) . After that, we will show that ||w̃||L2 (B n−1 (0,7/8)) controls ||w̃||H 1 (B n−1 (0,3/4)) , and hence we will get a lower bound for it, which will be a significant part of the doubling condition on the hypersurface. Recall that w̃(x) = c̃w(rx) = c̃vx0 ,λ (F (rx)), F (B n−1 (0, 2r) × {0}) ⊂ ∂Ωx0 ,λ , and the normal derivative of vx0 ,λ on ∂Ωx0 ,λ is zero. After using the area formula as above, we obtain (∇T denotes the tangential gradient): ||∇w̃||L2 (B n−1 (0,3/4)) = c̃r− ≥ c̃r n−3 2 − n−3 2 = c̃r− n−3 2 ||(∂ w̃/∂xn )||L2 (B n−1 (0,3/4)) = c̃r− n−3 2 ≤ c̃r − n−3 2 ||∇w||L2 (B n−1 (0,3r/4)) C −1 ||∇T vx0 ,λ ||L2 (F (B n−1 (0,3r/4))) C −1 ||∇vx0 ,λ ||L2 (F (B n−1 (0,3r/4))) , ||(∂w/∂xn )||L2 (B n−1 (0,3r/4)) C||∇vx0 ,λ ||L2 (F (B n−1 (0,3r/4))) , and (1.93) follows. Together with (1.92), we hence obtain (with a different constant than in (1.92)) 5 ||w̃||H 1 (B n−1 (0,3/4)) ≥ 2−Cλ . (1.94) Now we will show that ||w̃||H 1 (B n−1 (0,3/4)) ≥ ǫ̃ ⇒ ||w̃||L2 (B n−1 (0,7/8)) ≥ ǫ̃3 /C. (1.95) We will use Lemma 1.3.15. Since it involves functions defined in the whole Rn , introduce a cut-off function ϕ ∈ C ∞ (Rn ) such that ϕ=1 on B(0, 3/4), ϕ=0 on Rn \ B(0, 7/8), ||ϕ||C 2 (Rn ) ≤ C. Then using Lemma 1.3.15 for the function w̃′ = w̃ · ϕ (extended by 0 beyond Rn \ B(0, 7r/8)), we get that for any η > 0, ||∇w̃||L2 (B n−1 (0,3/4)) ≤ ||∇w̃′ ||L2 (Rn−1 ) C ||w̃′ ||L2 (Rn−1 ) η2 C ≤ Cη||w̃||H 2 (B(0,7/8)) + 2 ||w̃||L2 (B n−1 (0,7/8)) η C ≤ Cη||w̃||L2 (B(0,1)) + 2 ||w̃||L2 (B n−1 (0,7/8)) η C ≤ Cη + 2 ||w̃||L2 (B n−1 (0,7/8)) , η ≤ η||w̃′ ||H 2 (Rn ) + 36 (1.96) where in the last two lines we used the interior elliptic estimate for strong solutions ||w̃||H 2 (B(0,7/8)) ≤ C||w̃||L2 (B(0,1)) (see [19, Theorem 9.11]) and (1.90). Choosing η = ǫ̃/2C in (1.96), where ||w̃||H 1 (B n−1 (0,3/4)) ≥ ǫ̃, we obtain ǫ̃ − ||w̃||L2 (B n−1 (0,3/4)) ≤ ||∇w̃||L2 (B n−1 (0,3/4)) C ≤ ǫ̃/2 + 2 ||w̃||L2 (B n−1 (0,7/8)) , ǫ̃ and (1.95) follows. Together with (1.94), we obtain (with a different constant) 5 ||w̃||L2 (B n−1 (0,7/8)) ≥ 2−Cλ . (1.97) On the other hand, the trace theorem, the interior elliptic estimate [19, Theorem 9.11], and (1.90) imply ||w̃||L2 (B n−1 (0,7/4)) ≤ C||w̃||H 1 (B n (0,7/4)) ≤ C̃||w̃||L2 (B(0,2)) = C̃, so 5 ||w̃||L2 (B n−1 (0,7/8)) ≥ 2−Cλ ||w̃||L2 (B n−1 (0,7/4)) , which translates into 5 ||w||L2 (B n−1 (0,7r/8)) ≥ 2−Cλ ||w||L2 (B n−1 (0,7r/4)) . Since r ≤ r0 /2(1+ǫ) was arbitrary, we proved (1.86) for any r ≤ r1 , r1 = 7r0 /(16·1.1). In the process of the above proof, we showed that the average integral of u2 over a small ball on the boundary controls the average integral of u2 over a small solid ball with the same center. Since we will use this result in the next chapter (independently of the doubling condition), let us formulate it precisely. Corollary 1.3.16. There exist C, r0 > 0 depending only on Ω such that for all r ≤ r0 and x0 ∈ ∂Ω, there holds Z Z 2 −Cλ5 λ u ≥2 u2 . (1.98) r r B (x0 , 2r )∩Ω B ( x0 , λ )∩∂Ω λ Proof. This follows from (1.97). Looking back at the definition of c̃ by (1.89) and (1.90), we have Z Z 1 1 2 w (rx)dx = vx20 ,λ (F (x)) = n c̃2 r B(0,2) B(0,2r) Z n Z λ λn 2 u2 . ≥ n v ≥ n r B x0 , 2r r B x0 , 2r ∩Ω (1+ǫ)λ (1+ǫ)λ 37 We used a change of variables by the map F , whose Jacobian is 1 and for which (1.82) holds. On the other hand, from (1.97), we obtain Z Z 1 −Cλ5 1 1 2 w̃ = n−1 w2 2 ≤ 2 c̃2 c̃ B n−1 (0,7/8) r B n−1 (0,7/8r) Z 1 vx20 ,λ (F (x)) = n−1 r n−1 B (0,7/8r) Z Z n−1 λ λn−1 2 ≤ n−1 v = n−1 u2 . 7(1+ǫ) 7(1+ǫ) r r B x0 , B x0 , r ∩∂Ω r ∩∂Ω 8 8 Here we used a change of variables by the map F |B n−1 (0,r)×{0} , whose Jacobian is larger or equal to 1 and for which (1.84) holds. Combining the above two estimates and using ǫ such that (1 + ǫ)2 < 8/7 (this gives us r0 > 0 depending only on Ω such that the estimates hold for r ≤ r0 ), we obtain (1.98). 38 Chapter 2 Size of the Nodal Sets of Steklov Eigenfunctions 2.1 Nodal Sets of Steklov Eigenfunctions on Analytic Domains In this section we prove Theorem 1: Theorem 1. Let Ω ⊂ Rn be an analytic domain. Then there exists a constant C depending only on Ω and n such that for any λ > 0 and u which is a (classical) solution to (1) there holds Hn−2 ({x ∈ ∂Ω : u(x) = 0}) ≤ Cλ6 . (2.1) The proof relies on the doubling condition on the boundary (Theorem 2) and the approach used in [29] for analytic solutions of elliptic equations on a solid domain. For the rest of this section, assume the assumptions of Theorem 1: Ω ⊂ Rn is an analytic domain and u is a solution to (1) with eigenvalue λ > 0. 2.1.1 Analyticity of the eigenfunctions First we will see that u is analytic up to the boundary ∂Ω, and get an estimate on its derivatives which we will need in a complexification argument. Proposition 2.1.1. The eigenfunction u is real-analytic on a neighborhood of Ω. More precisely, there exists r0 > 0 depending only on Ω such that u can be harmonically extended onto an rλ0 -neighborhood of Ω. Moreover, there exists a constant C such that for every x0 ∈ ∂Ω and r ≤ r0 , α! |Dα u(x0 )| ≤ C (r/2neλ)|α| 39 Z − B(x0 ,r/λ) u2 !1/2 . (2.2) Proof. By the theorem on elliptic iterates of Lions and Magenes ([30, Theorem VIII.1.2]), u is analytic on a neighborhood of Ω. To determine the size of this neighborhood, for a fixed x0 ∈ ∂Ω, in a similar way as in Section 1.3.1, denote ux0 ,λ (x) := u(x0 + x/λ), x ∈ Ωx0 ,λ , where Ωx0 ,λ = {x : x0 + x/λ ∈ Ω}. By (1) we have ∆ux0 ,λ = 0 ∂u =u ∂ν in Ωx0 ,λ , on ∂Ωx0 ,λ . It follows from sections VIII.1 and VIII.2 in [30] that ux0 ,λ can be analytically extended onto B(0, r0 ), where r0 depends only on Ω, not on x0 or λ (the domain Ωx0 ,λ changes, but only becomes flatter, i.e. better with increasing λ). This means that u can be analytically extended onto B(x0 , r0 /λ). Then ∆u is also analytic on B(x0 , r0 /λ) and is equal to zero on the open set B(x0 , r0 /λ) ∩ Ω, hence it is zero in B(x0 , r0 /λ) and u is harmonic in B(x0 , r0 /λ). Hence, from Proposition 1.13 and Remark 1.19 in [24], we obtain |Dα u(x0 )| ≤ C α! (r/2neλ)|α| α! ≤C (r/2neλ)|α| max B(x0 ,r/2λ) |u| Z − u2 B(x0 ,r/λ) !1/2 . Corollary 2.1.2. Let r0 be as in Proposition 2.1.1, x0 ∈ ∂Ω. Then u can be extended n into the complex ball B C (x0 , r0 /6nλ) such that for any r ≤ r0 , sup n z∈B C (x0 ,r/6nλ) Z |u(z)| ≤ C − B(x0 ,r/λ) u 2 !1/2 . (2.3) Proof. Write the real Taylor polynomial of u in x0 . By the estimate (2.2) on its con efficients |Dα u(x0 )|/α!, it converges for all complex z ∈ B C (x0 , r0 /2neλ), and for n z ∈ B C (x0 , r/6nλ) we also obtain the estimate (2.3). 2.1.2 Proof of Theorem 1 Flattening the boundary and complexification: As in the proof of Theorem 2 in Section 1.3.4, we cover ∂Ω by finitely many pieces Γi ⊂ ∂Ω, i = 1, 2, . . . , k, such that k depends only on Ω and each Γi is a graph of a (this time) analytic function (in some coordinate system). We require some additional properties of this cover. For each Γi = Γ, after choosing the right coordinate system, we want Γ = {x = (x′ , xn ) ∈ Rn : x′ ∈ Γ̃, xn = Φ(x′ )}, 40 where Γ̃ ⊂ Rn−1 is a compact set, Φ is analytic on the larger set 2r0 ′ n−1 Γ̃ = x ∈ R : dist(x, Γ̃) < , λ1 where r0 > 0 is a constant depending on Ω, {(x′ , Φ(x′ )) : x′ ∈ Γ̃′ } ⊂ ∂Ω, ∇Φ(x0 ) = 0 for some point x0 ∈ Γ̃, and Φ can be extended to a complex analytic function on −r0 r0 n−1 ′ , Γ̃ × λ1 λ1 with ∇z Φ(x0 ) = 0, ′ |∇z Φ(z)| ≤ 0.1 for z ∈ Γ̃ × −r0 r0 , λ1 λ1 n−1 . (2.4) This can be achieved by considering a neighborhood of each x0 ∈ ∂Ω which can be parametrized this way and then choosing a finite cover of ∂Ω by these neighborhoods. Denote r0 ′ n−1 Γ̃m = x ∈ R : dist(x, Γ̃) < λ1 an “intermediate” set (Γ̃ ⊂ Γ̃′m ⊂ Γ̃′ ). In a similar way as in the proof of Theorem 2, denote w(z ′ ) := u(z ′ , Φ(z ′ )). By the definition of Φ and Corollary 2.1.2, this is well-defined if −r0 r0 n−1 r0 ′ ′ z ∈ Γ̃ × , , and dist((z ′ , Φ(z ′ )), ∂Ω) ≤ λ1 λ1 6nλ where r0 is the minimum of r0 ’s coming up in Corollary 2.1.2 and in the discussion above about properties of Φ. If z ′ = x′ + iy ′ , x′ ∈ Γ̃′ , |y ′ | ≤ λr01 , then (x′ , Φ(x′ )) ∈ ∂Ω, and we have dist((z ′ , Φ(z ′ )), ∂Ω) ≤ |(x′ + iy ′ , Φ(x′ + iy ′ )) − (x′ , Φ(x′ ))| ≤ |y ′ | + |y ′ | sup |∇z Φ(x′ + ity ′ )| ≤ 1.1|y ′ | t∈[0,1] by (2.4). Hence, w is well-defined and analytic for z ′ = x′ + iy ′ , x′ ∈ Γ̃′ , |y ′ | ≤ r0 , 6.6nλ and by Corollary 2.1.2, for all r ≤ r0 , x′ ∈ Γ̃′m we also have sup r ) (x′ , 6.6nλ n−1 BC Z |w| ≤ C − B((x′ ,Φ(x′ )),r/λ) 41 u2 !1/2 . (2.5) Instead of the bound on the right hand side which involves an integral over Rn -balls, using Corollary 1.3.16 we can bound the left hand side by an integral over balls on ∂Ω. Denote x0 = (x′ , Φ(x′ )). Although Corollary 1.3.16 gives a bound only on an integral over B(x0 , r/λ) ∩ Ω, in our case u is harmonic also on B(x0 , r/λ) ∩ Ωc and satisfies the c almost identical boundary condition ∂u ∂ν = −λu on ∂Ω (the normal has opposite sign). Since all arguments in the proof of Corollary 1.3.16 are local and the sign of λ never mattered, applying it to B(x0 , r/λ) ∩ Ωc gives us Z Z 2 −Cλ5 λ u ≥2 u2 r r B (x0 , r )∩Ωc B (x0 , 2λ )∩∂Ω λ and together with Corollary 1.3.16 for regular Ω and (2.5), we obtain !1/2 n−1 Z λ 2 u r B (x0 , r )∩∂Ω Cλ5 BC sup |w| ≤ 2 r (x′ , 3.3nλ ) n−1 λ for all x′ ∈ Γ̃′m , x0 = (x′ , Φ(x′ )), and r ≤ r0 , where r0 depends only on Ω (it is different than above). By a change of variables as in the proof of Theorem 2, using (2.4) for bounding the Jacobian, this translates into Cλ5 sup |w| ≤ 2 r (x′ , 3.3nλ ) B n−1 (x′ , 1.1r λ ) n−1 BC (since x′ ∈ Γ̃′m , we have B n−1 x′ , 1.1r λ Hence we proved Z − w2 !1/2 ⊂ Γ̃′ and the right hand side is well-defined). Cλ5 |w| ≤ 2 sup r ) (x′ , 4nλ Z − r B n−1 (x′ , λ ) n−1 BC w2 !1/2 (2.6) for all x′ ∈ Γ̃′m and r ≤ r0 , where r0 > 0 depends only on Ω (different from r0 above). Also note that the doubling condition, Theorem 2, translates into the following for w: Z Z 2 Cλ5 w ≤2 w2 (2.7) r B n−1 (x′ , λ ) r B n−1 (x′ , 2λ ) for all x′ ∈ Γ̃′m and r ≤ r0 , where r0 > 0 depends only on Ω (different from r0 above). We actually proved this in the process of proving Theorem 2 as (1.86) (the w|B n−1 (0,r0 )×{0} in there was defined exactly as our w shifted to 0 and scaled by λ). By the change of variable as in the proof of Theorem 2, using (2.4), to prove (2.1), it is enough to prove Hn−2 ({x′ ∈ Γ̃ : w(x′ ) = 0}) ≤ Cλ6 . (2.8) To estimate the nodal set, we use an estimate on the number of zero points for analytic functions. 42 Lemma 2.1.3. Suppose f : B1 ⊂ C → C is analytic with |f (0)| = 1 sup |f | ≤ 2N , and B1 for some positive constant N . Then for any r ∈ (0, 1) there holds #{z ∈ Br : f (z) = 0} ≤ cN, where c is a positive constant depending only on r. For r = 1/2, we have #{z ∈ B1/2 : f (z) = 0} ≤ N. Proof. A similar version of this lemma was first proved in [10]. This version can be found in [22] as Lemma 2.3.2. Proof of Theorem 1. It remains to prove (2.8). Take r0 such that both (2.6) and (2.7) r0 hold for all r ≤ r0 and x′ ∈ Γ̃′m . Choose a cover of Γ̃ by balls of radii R = (16n+1)λ . n−1 We need Cλ balls in the cover, where C depends only on Ω. Take one of them and denote it B(p, R), i.e. the center is p ∈ Γ̃. Choose xp ∈ B(p, R) ⊂ Γ̃′m such that Z |w(xp )| ≥ − B(p,R) |w| 2 !1/2 . (2.9) Then B(p, R) ⊂ B(xp , 2R), so the Cλn−1 balls B(xp , 2R) cover Γ̃. Next, B(xp , 16nR) ⊂ B(p, (16n + 1)R), and since 4R < r0 /4nλ, by (2.6) we have Cλ5 sup BC n−1 (xp ,4R) |w| ≤ 2 Z − w 2 B n−1 (xp ,16nR) Cλ5 ≤2 Z − !1/2 w B n−1 (p,(16n+1)R) 2 !1/2 . Now use the doubling condition (2.7) with center p, ⌈log2 (16n + 1)⌉-times, and obtain sup B Cn−1 (xp ,4R) Cλ5 |w| ≤ 2 Z − w B n−1 (p,R) 2 !1/2 5 ≤ 2Cλ |w(xp )| (2.10) (we used (2.9) in the last inequality). For each direction ω ∈ Rn−1 , |ω| = 1, consider the function fω (z) = w(xp + 4Rzω) of one complex variable z defined on B C (0, 1). From (2.10) and Lemma 2.1.3, we obtain #{x ∈ B n−1 (xp , 2R) : x − xp ||ω, w(x) = 0} ≤ #{z ∈ B1/2 : fω (z) = 0} = N (ω) ≤ Cλ5 . 43 By the integral geometry estimate [14, 3.2.22], we obtain Z n−2 n−1 n−2 H {x ∈ B (xp , 2R) : w(x) = 0} ≤ c(n)R Sn−2 N (ω)dω ≤ C 1 λn−2 λ5 . Since there were Cλn−1 balls B n−1 (xp , 2R) which cover Γ̃, summing up these estimates gives (2.8). 2.2 Nodal Sets of Steklov Eigenfunctions on Smooth Domains In this section we briefly discuss the possible directions of research to obtain an estimate for the size of the nodal set of Steklov eigenfunctions in the case that Ω is not analytic, using the doubling condition which we proved also in this setting. Following an argument first employed in [25], the following bound was proved for solutions of general elliptic equations with non-analytic coefficients in [22] as Corollary 5.3.8: Theorem 10. Suppose that w is a nonzero solution of (1.20) in B1 , with the coefficients of the equation satisfying (1.21), (1.22) and (1.23). Assume that for each B(p, r) ⊂ B1 , there holds R r B(p,r) |Dw|2 R ≤ N. (2.11) 2 ∂B(p,r) w Then Hn−1 ({x ∈ B1/2 : w(x) = 0}) ≤ (c1 N )c2 N , where c1 and c2 are positive constants depending only on n, α, Γ and K. The proof of this theorem splits the nodal set of w into a good part, where the gradient of w is large, and a bad part, where the gradient is small. The good part can be well approximated by the nodal set of a harmonic function and estimated that way. For the bad part, the fact that the singular set where both w and Dw vanish has a lower dimension, is used. Unfortunately, that part does not translate into our case of Steklov eigenfunctions: since we consider them on the hypersurface ∂Ω, the singular set can have the same dimension as the nodal set. The Hn−2 -measure of the singular set can be estimated, but the known results in [21] require the coefficients of the equation to be smooth and use their C M -norms, where M depends (among others) on N from (2.11), which in our case translates to requiring C M -smoothness of the domain Ω with M depending on λ. The estimate on the size of the singular set is also not explicit. Another way how to estimate the size of the nodal set of Steklov eigenfunctions on non-analytic Ω is to follow the approach used in [23] to estimate the measure of slices of nodal sets. However, this approach also uses the C M -norms of coefficients, M depending on N much like above, and the estimate is also not explicit. Using either approach, one can hope to prove that for C ∞ -smooth Ω, the Hn−2 measure of the nodal set of Steklov eigenfunctions is finite for any λ. 44 Bibliography [1] Frederick J. Almgren, Jr., Dirichlet’s problem for multiple valued functions and the regularity of mass minimizing integral currents, Minimal submanifolds and geodesics (Proc. Japan-United States Sem., Tokyo, 1977), North-Holland, Amsterdam, 1979, pp. 1–6. MR 574247 (82g:49038) [2] Habib Ammari, Hyeonbae Kang, Hyundae Lee, and Mikyoung Lim, Enhancement of near cloaking using generalized polarization tensors vanishing structures. part I: The conductivity problem, Submitted to Communications in Mathematical Physics, 2011. [3] Nachman Aronszajn, Andrzej Krzywicki, and Jacek Szarski, A unique continuation theorem for exterior differential forms on Riemannian manifolds, Ark. Mat. 4 (1962), 417–453 (1962). MR 0140031 (25 #3455) [4] Thierry Aubin, Équations différentielles non linéaires et problème de Yamabe concernant la courbure scalaire, J. Math. Pures Appl. (9) 55 (1976), no. 3, 269–296. MR 0431287 (55 #4288) [5] Catherine Bandle, Über des Stekloffsche Eigenwertproblem: Isoperimetrische Ungleichungen für symmetrische Gebiete, Z. Angew. Math. Phys. 19 (1968), 627–637. MR 0235459 (38 #3768) [6] Rodrigo Bañuelos, Tadeusz Kulczycki, Iosif Polterovich, and Bartlomiej Siudeja, Eigenvalue inequalities for mixed Steklov problems, Operator theory and its applications, Amer. Math. Soc. Transl. Ser. 2, vol. 231, Amer. Math. Soc., Providence, RI, 2010, pp. 19–34. MR 2758960 [7] Haı̈m Brézis and Louis Nirenberg, Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents, Comm. Pure Appl. Math. 36 (1983), no. 4, 437–477. MR 709644 (84h:35059) [8] Alberto-P. Calderón, On an inverse boundary value problem, Seminar on Numerical Analysis and its Applications to Continuum Physics (Rio de Janeiro, 1980), Soc. Brasil. Mat., Rio de Janeiro, 1980, pp. 65–73. MR 590275 (81k:35160) [9] Bodo Dittmar, Sums of reciprocal Stekloff eigenvalues, Math. Nachr. 268 (2004), 44–49. MR 2054531 (2005d:35188) 45 [10] Harold Donnelly and Charles Fefferman, Nodal sets of eigenfunctions on Riemannian manifolds, Invent. Math. 93 (1988), no. 1, 161–183. MR 943927 (89m:58207) [11] José F. Escobar, Sharp constant in a Sobolev trace inequality, Indiana Univ. Math. J. 37 (1988), no. 3, 687–698. MR 962929 (90a:46071) [12] , Conformal deformation of a Riemannian metric to a scalar flat metric with constant mean curvature on the boundary, Ann. of Math. (2) 136 (1992), no. 1, 1–50. MR 1173925 (93e:53046) [13] , The geometry of the first non-zero Stekloff eigenvalue, J. Funct. Anal. 150 (1997), no. 2, 544–556. MR 1479552 (98g:58180) [14] Herbert Federer, Geometric measure theory, Die Grundlehren der mathematischen Wissenschaften, Band 153, Springer-Verlag New York Inc., New York, 1969. MR 0257325 (41 #1976) [15] David W. Fox and James R. Kuttler, Sloshing frequencies, Z. Angew. Math. Phys. 34 (1983), no. 5, 668–696. MR 723140 (84k:76027) [16] Ailana Fraser and Richard Schoen, The first Steklov eigenvalue, conformal geometry, and minimal surfaces, Adv. Math. 226 (2011), no. 5, 4011–4030. MR 2770439 [17] Nicola Garofalo and Fang-Hua Lin, Monotonicity properties of variational integrals, Ap weights and unique continuation, Indiana Univ. Math. J. 35 (1986), no. 2, 245– 268. MR 833393 (88b:35059) [18] , Unique continuation for elliptic operators: a geometric-variational approach, Comm. Pure Appl. Math. 40 (1987), no. 3, 347–366. MR 882069 (88j:35046) [19] David Gilbarg and Neil S. Trudinger, Elliptic partial differential equations of second order, Springer-Verlag, Berlin, 1977, Grundlehren der Mathematischen Wissenschaften, Vol. 224. MR 0473443 (57 #13109) [20] Qing Han, Nodal sets of harmonic functions, Pure Appl. Math. Q. 3 (2007), no. 3, part 2, 647–688. MR 2351641 (2008j:31001) [21] Qing Han, Robert Hardt, and Fanghua Lin, Geometric measure of singular sets of elliptic equations, Comm. Pure Appl. Math. 51 (1998), no. 11-12, 1425–1443. MR 1639155 (99h:35032a) [22] Qing Han and Fang-Hua Lin, Nodal sets of solutions of elliptic differential equations, in preparation. [23] , On the geometric measure of nodal sets of solutions, J. Partial Differential Equations 7 (1994), no. 2, 111–131. MR 1280175 (95c:35079) [24] Qing Han and Fanghua Lin, Elliptic partial differential equations, Courant Lecture Notes in Mathematics, vol. 1, New York University Courant Institute of Mathematical Sciences, New York, 1997. MR 1669352 (2001d:35035) 46 [25] Robert Hardt and Leon Simon, Nodal sets for solutions of elliptic equations, J. Differential Geom. 30 (1989), no. 2, 505–522. MR 1010169 (90m:58031) [26] Joseph Hersch, Lawrence E. Payne, and Menahem M. Schiffer, Some inequalities for Stekloff eigenvalues, Arch. Rational Mech. Anal. 57 (1975), 99–114. MR 0387837 (52 #8676) [27] Yanyan Li and Meijun Zhu, Sharp Sobolev trace inequalities on Riemannian manifolds with boundaries, Comm. Pure Appl. Math. 50 (1997), no. 5, 449–487. MR 1443055 (98c:46065) [28] Fang-Hua Lin, Nonlinear theory of defects in nematic liquid crystals; phase transition and flow phenomena, Comm. Pure Appl. Math. 42 (1989), no. 6, 789–814. MR 1003435 (90g:82076) [29] , Nodal sets of solutions of elliptic and parabolic equations, Comm. Pure Appl. Math. 44 (1991), no. 3, 287–308. MR 1090434 (92b:58224) [30] Jacques-Louis Lions and Enrico Magenes, Non-homogeneous boundary value problems and applications. Vol. III, Springer-Verlag, New York, 1973, Translated from the French by P. Kenneth, Die Grundlehren der mathematischen Wissenschaften, Band 183. MR 0350179 (50 #2672) [31] Fernando C. Marques, Conformal deformations to scalar-flat metrics with constant mean curvature on the boundary, Comm. Anal. Geom. 15 (2007), no. 2, 381–405. MR 2344328 (2008i:53046) [32] Lawrence E. Payne, Some isoperimetric inequalities for harmonic functions, SIAM J. Math. Anal. 1 (1970), 354–359. MR 0437782 (55 #10705a) [33] Richard Schoen, Conformal deformation of a Riemannian metric to constant scalar curvature, J. Differential Geom. 20 (1984), no. 2, 479–495. MR 788292 (86i:58137) [34] Wladimir Stekloff, Sur les problèmes fondamentaux de la physique mathématique, Ann. Sci. École Norm. Sup. (3) 19 (1902), 191–259. MR 1509012 [35] John Sylvester and Gunther Uhlmann, A uniqueness theorem for an inverse boundary value problem in electrical prospection, Comm. Pure Appl. Math. 39 (1986), no. 1, 91–112. MR 820341 (87j:35377) [36] John A. Toth and Steve Zelditch, Counting nodal lines which touch the boundary of an analytic domain, J. Differential Geom. 81 (2009), no. 3, 649–686. MR 2487604 (2010i:58030) [37] Robert Weinstock, Inequalities for a classical eigenvalue problem, J. Rational Mech. Anal. 3 (1954), 745–753. MR 0064989 (16,368c) [38] Hidehiko Yamabe, On a deformation of Riemannian structures on compact manifolds, Osaka Math. J. 12 (1960), 21–37. MR 0125546 (23 #A2847) 47 [39] Shing Tung Yau, Problem section, Seminar on Differential Geometry, Ann. of Math. Stud., vol. 102, Princeton Univ. Press, Princeton, N.J., 1982, pp. 669–706. MR 645762 (83e:53029) 48