Exponential localization of Steklov eigenfunctions on warped product manifolds: the flea on the elephant phenomenon (Q6081634)

Let \((M,g)\) be a smooth compact connected Riemannian manifold of dimension \(n \geq 2\) with boundary \(\partial M\). The authors consider Steklov spectrum of the Dirichlet to Neumann operator \(\Lambda_g(\omega)\) defined by \[ \Lambda_g(\omega) \psi = (\partial_\nu u)_{|\partial M}, \] where for each \(\psi \in H^{\frac{1}{2}}(\partial M)\) the function \(u \in H^1(M)\) is the unique solution of the Dirichlet problem \[ \left\{ \begin{array}{cl} -\Delta_g u = \omega \,u, & \text{on } M, \\ u = \psi, & \text{on } \partial M, \end{array} \right.\tag{1} \] while \(\omega\) is a fixed frequency, which is not in the Dirichlet spectrum of the positive Laplace-Beltrami operator \(-\Delta_g\) on \((M,g)\). The spectrum of the operator \(\Lambda_g(\omega)\) is discrete and forms an increasing sequence \[ \lambda_0 < \lambda_1 \leq \dots \leq \lambda_m \to +\infty. \] By \((\phi_m)_{m \geq 0}\) we denote the corresponding normalized eigenfunctions in \(L^2(\partial M)\) and by \((\varphi_m)_{m \geq 0}\) the corresponding solutions of the above Dirichlet problem with \(\psi = \phi_m\). The authors analyse the Steklov eigenfunctions on warped product Riemannian manifolds, the boundaries of which have two distinct connected components, namely, they consider a cylinder \(M = [0,1] \times K\) with \(K\) a \((n-1)\)-dimensional smooth closed manifold equipped with a Riemannian metric having the form \[ g = f(x) [dx^2 + g_K], \] where \(f\) is a smooth positive function on \([0,1]\), and \(g_K\) is any smooth Riemannian metric on \(K\). The second-order elliptic selfadjoint operator Laplace-Beltrami operator \(-\Delta_K\) on \(K\) has a discrete spectrum \((\mu_m)_{m\geq0}\) \[ 0 = \mu_0 < \mu_1 \leq \mu_2 \leq \dots \leq \mu_m \longrightarrow +\infty \] with a sequence \((Y_m)_{m \geq 0}\) of associated normalized eigenfunctions. The transversal Laplace-Beltrami operator \(-\Delta_K\) commutes with the Laplace-Beltrami operator \(-\Delta_g\) on the whole manifold \((M,g)\) as well as with the self-adjoint DN operator \(\Lambda_g(\omega)\) on \(L^2(\partial M)\). This is why, to each eigenvalue \(\mu_m, \ m\geq0\), one can associate two distinct Steklov eigenvalues \(\lambda_m^\pm\), (with \(\lambda_m^+ > \lambda_m^-\)), and the the set of pairs \((\mu_m,\lambda_m^\pm)\) forms the joint spectrum of \((-\Delta_K,\Lambda_g(\omega))\). The main results concern the asymptotic behavior of the Steklov eigenvalues \(\lambda_m^\pm\) and of their gap \(d_m:=|\lambda_m^+-\lambda_m^-|\), as well as the localization of the Steklov eigenfunctions \(\varphi_m^\pm\). The authors distinguish four cases and in each of them the main result is as follows. Case I [(\(n=2\), \(\omega\ne0\)) or (\(n\geq 3\))]. Let \(f(x) = f (1-x)\), \(x\in[0,1]\). Under the assumption of symmetry on \(f\) , we have as \(m\to\infty\), \[ \lambda_m^\pm=\frac{\sqrt{\mu_m}}{\sqrt{f(0)}}+O(1) \] and \[ \lambda_m^+-\lambda_m^-=\frac{2}{\sqrt{f(0)}}\sqrt{\mu_n}e^{-\sqrt{\mu_m}} \left(1+O\left(\frac{1}{\sqrt{\mu_m}}\right)\right). \] Moreover, there exist two constants \(C > 0\) and \(m_0 > 0\) such that for all \(m \geq m_0\) and for all \((x,\theta)\in [0, 1] \times K\) \[ |\varphi_m^\pm(x,\theta)|\leq C\mu_m^{\frac{n-2}{4}} \big(e^{-\sqrt{\mu_m}x} + e^{-\sqrt{\mu_m}(1-x)}\big). \] Case II.A. [(\(n=2\), \(\omega\ne0\)) or (\(n\geq 3\))]. Assume here that the Taylor series of \(f(x)\) and \(f (1-x)\) differ at the order \(k\) at \(x = 0\), i.e. \(k\geq 0\) is the smallest integer such that \(f^{(k)}(0)\ne(-1)^k f^{(k)}(1)\). Let \begin{align*} &a_0=b_0=\frac{1}{\sqrt{f(0)}}-\frac{1}{\sqrt{f(1)}}, \\ &a_k=-\frac{\omega}{2^{k+1}\sqrt{f(0)}} \big(f^{(k)}(0)-(-1)^k f^{(k)}(1)\big),\qquad k\geq 1, \\ &b_k=\frac{n-2}{2^{k+1}f^{\frac{3}{2}}(0)} \big(f^{(k)}(0)-(-1)^k f^{(k)}(1)\big),\qquad k\geq 1. \end{align*} Under the above assumptions, we have as \(m\to \infty\), \[ \lambda_m^+ = \max \left( \frac{1}{\sqrt{f(1)}},\frac{1}{ {\sqrt{f(0)}}}\right) \sqrt{\mu_m} + O (1) , \quad \lambda_m^- = \min \left( \frac{1}{\sqrt{f(1)}},\frac{1}{ {\sqrt{f(0)}}}\right) \sqrt{\mu_m} + O (1). \] The distance between the two eigenvalues \(\lambda_m^\pm\) satisfies in dimension \(n=2\) with \(\omega \not=0\), \begin{align*} d_m =& |a_0| \ \sqrt{\mu_m} + O \left( 1 \right) \mbox{ if } \ k=0, \\ d_m =& |a_k| \ \mu_m^{-\frac{1+k}{2}} + O \left( \mu_m^{-\frac{k+2}{2}}\right) \mbox{ if } \ k \geq 1 , \end{align*} whereas in dimension \(n \geq 3\), \[ d_m = |b_k| \ \mu_m^{\frac{1-k}{2}} + O \left( \mu_m^{-\frac{k}{2}}\right) \mbox{ if } k \geq 0. \] Moreover, there exist constants \(C_k>0\) and \(m_0 > 0\) such that for all \(m \geq m_0\) and for all \((x,\theta) \in [0,1] \times K\), we have: \begin{itemize} \item[1.] For \(n=2\) with \(\omega \not =0\), if \(a_k <0\), \begin{align*} & |\varphi_m^+(x,\theta)| \leq C_0 \left( e^{-\sqrt{\mu_m}(1+x)} + e^{-\sqrt{\mu_m}(1-x)} \right), \mbox{ if } \ k =0\, .\\ & |\varphi_m^-(x,\theta)| \leq C_0 \left( e^{-\sqrt{\mu_m}x} + e^{-\sqrt{\mu_m}(2-x)} \right), \mbox{ if } \ k =0 \,.\\ & |\varphi_m^+(x,\theta)| \leq C_k \left( \mu_m^{\frac{k + 2}{2}} e^{-\sqrt{\mu_m}(1+x)} + e^{-\sqrt{\mu_m}(1-x)} \right), \mbox{ if } \ k \geq 1\,. \\ & |\varphi_m^-(x,\theta)| \leq C_k \left( e^{-\sqrt{\mu_m}x} + \mu_m^{\frac{k + 2}{2}} e^{-\sqrt{\mu_m}(2-x)} \right), \mbox{ if } \ k \geq 1\,. \end{align*} and, the same estimates hold with \(+\) and \(-\) inverted if \(a_k>0\)\,. \item[2.] For \(n \geq 3\), if \(b_k <0\), \begin{align*} & |\varphi_m^+(x,\theta)| \leq C_k \left( \mu_m^{\frac{n+2k - 2}{4}} e^{-\sqrt{\mu_m}(1+x)} + \mu_m^{\frac{n - 2}{4}} e^{-\sqrt{\mu_m}(1-x)} \right), \\ & |\varphi_m^-(x,\theta)| \leq C_k \left( \mu_m^{\frac{n - 2}{4}} e^{-\sqrt{\mu_m}x} + \mu_m^{\frac{n +2k - 2}{4}} e^{-\sqrt{\mu_m}(2-x)} \right), \end{align*} and, the same estimates hold with \(+\) and \(-\) inverted if \(b_k>0\)\,. \end{itemize} {Case II.B}. Assume that there exists \(a \in [0,\tfrac{1}{2})\) such that \(f(x) = f(1-x)\) for all \(x \in [0,a]\). If \(a=0\), then assume only that \(f(x)\) and \(f(1-x)\) have the same Taylor series at \(x=0\). Assume that there exists \(0<\delta <\tfrac{1}{2} - a\) such that for all \(x \in (a,a+\delta]\), we have either \(q_f(x) - q_f(1-x) > 0\), or \(q_f(x) - q_f(1-x) < 0\). Under the above assumptions, if there exists \(\delta>0\) such that \(q_f(x) - q_f(1-x) > 0\) for all \(x \in (a,a+\delta]\), we have as \(m \to \infty\), \[ \lambda_m^- = \frac{\sqrt{\mu_m}}{\sqrt{f(0)}} + O (1), \quad \lambda_m^+ = \frac{\sqrt{\mu_m}}{\sqrt{f(1)}} + O (1) . \] If \(a>0\), for all \(\varepsilon>0\) small enough, there exists three constants \(c_\varepsilon>0\), \(C_\varepsilon >0\), and \(m_\varepsilon > 0\) such that for all \(m \geq m_\varepsilon\) \[ c_\varepsilon \ e^{-2(a+\varepsilon) \sqrt{\mu_m}}\leq \lambda_m^+ -\lambda_m^- \leq C_\varepsilon \ e^{-2 (a-\varepsilon) \sqrt{\mu_m} }, \] If \(a=0\), for all \(\varepsilon>0\) small enough, there exists two constants \(c_\varepsilon>0\) and \(m_{\varepsilon} > 0\) such that for all \(m \geq m_{\varepsilon}\), \[ \lambda_m^+ -\lambda_m^- \geq c_\varepsilon \ e^{-2\varepsilon \sqrt{\mu_m}}. \] \noindent Moreover, for all \(m \geq m_\varepsilon\) and for all \((x,\theta) \in [0,1] \times K\), \begin{align*} & |\varphi_m^-(x,\theta)| \leq C_\varepsilon \left( e^{-\sqrt{\mu_m}(1 - 2(a+\varepsilon)+x)} + \sqrt{\mu_m}^{\frac{n-2}{2}} e^{-\sqrt{\mu_m}(1-x)} \right), \\ & |\varphi_m^+(x,\theta)| \leq C_\varepsilon \left( \sqrt{\mu_m}^{\frac{n-2}{2}} e^{-\sqrt{\mu_m}x} + e^{-\sqrt{\mu_m}(2 - 2(a+\varepsilon)-x)} \right). \end{align*} If for all \(x \in (a,a+\delta]\), we have \(q_f(x) - q_f(1-x) < 0\), the same estimates hold with \(+\) and \(-\) inverted. {Case II.C}. Assume that \(f(x)\) and \(f(1-x)\) have the same Taylor series at \(x=0\). Under the above assumptions, for all \(m \geq 0\), \[ \lambda_m^+ -\lambda_m^- \geq \frac{2 \ \mu_m}{(f(0))^{1/2}} \ \frac {1}{\sqrt{\mu_m} \sinh (\sqrt{\mu_m}) + e^{\sqrt{\mu_m} + \|q_f\|}}, \] where \(\|q_f\|\) is the \(L^2\)-norm of the potential \(q_f\). \\ Moreover, there exists a subsequence \((m_k)_{k\geq 0}\) such that we have either \[ \lambda_{m_k}^- = \frac{\sqrt{\mu_{m_k}}}{\sqrt{f(0)}} + O (1), \qquad \lambda_{m_k}^+ = \frac{\sqrt{\mu_{m_k}}}{\sqrt{f(1)}} + O (1) \mbox{ as } m_k \to \infty. \] and moreover, there exists a constant \(C > 0\) such that for all \(\varepsilon > 0\), there exists \(m_\varepsilon > 0\) such that for all \(m_k \geq m_\varepsilon\) and for all \((x,\theta) \in [0,1] \times K\) \begin{align*} & |\varphi_{m_k}^-(x,\theta)| \leq C \sqrt{\mu_{m_k}}^{\frac{n-2}{2}} \left( \varepsilon e^{-\sqrt{\mu_{m_k}}x} + e^{-\sqrt{\mu_{m_k}}(1-x)} \right), \\ & |\varphi_{m_k}^+(x,\theta)| \leq C \sqrt{\mu_{m_k}}^{\frac{n-2}{2}} \left( e^{-\sqrt{\mu_{m_k}}x} + \varepsilon e^{-\sqrt{\mu_{m_k}}(1 - x} \right), \end{align*} or the same estimates hold with \(+\) and \(-\) inverted.