\documentclass[reqno]{amsart} \usepackage{hyperref} \usepackage{graphicx} \AtBeginDocument{{\noindent\small \emph{Electronic Journal of Differential Equations}, Vol. 2011 (2011), No. 132, pp. 1--30.\newline ISSN: 1072-6691. URL: http://ejde.math.txstate.edu or http://ejde.math.unt.edu \newline ftp ejde.math.txstate.edu (login: ftp)} \thanks{\copyright 2011 Texas State University - San Marcos.} \vspace{9mm}} \begin{document} \title[\hfilneg EJDE-2011/132\hfil Smallest eigenvalues of the $p$-Laplace operator] {Numerical investigation of the smallest eigenvalues of the $p$-Laplace operator on planar domains} \author[J. Hor\'ak\hfil EJDE-2011/132\hfilneg] {Ji\v{r}\'{\i} Hor\'ak} \address{Ji\v{r}\'{\i} Hor\'ak \newline Universit\"at zu K\"oln, Mathematisches Institut, 50923 K\"oln, Germany} \email{jhorak@math.uni-koeln.de} \urladdr{www.mi.uni-koeln.de/~jhorak/plaplace} \thanks{Submitted September 22, 2011. Published October 13, 2011.} \subjclass[2000]{35J92, 49M30, 49R05} \keywords{$p$-Laplace operator; eigenvalue; mountain pass algorithm; symmetry} \begin{abstract} The eigenvalue problem for the $p$-Laplace operator with $p>1$ on planar domains with zero Dirichlet boundary condition is considered. The Constrained Descent Method and the Constrained Mountain Pass Algorithm are used in the Sobolev space setting to numerically investigate the dependence of the two smallest eigenvalues on $p$. Computations are conducted for values of $p$ between 1.1 and 10. Symmetry properties of the second eigenfunction are also examined numerically. While for the disk an odd symmetry about the nodal line dividing the disk in halves is maintained for all the considered values of $p$, for rectangles and triangles symmetry changes as $p$ varies. Based on the numerical evidence the change of symmetry in this case occurs at a certain value $p_0$ which depends on the domain. \end{abstract} \maketitle \numberwithin{equation}{section} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{definition}[theorem]{Definition} \newtheorem{remark}[theorem]{Remark} \newcommand{\dualpairing}[2]{\langle #1,#2\rangle} \section{Introduction} For a bounded domain $\Omega\subset\mathbb{R}^N$, $N\in\mathbb{N}$ and a parameter $p\in(1,\infty)$ consider the nonlinear eigenvalue problem \begin{equation} \label{eq:evproblem} \begin{gathered} -\Delta_p u = \lambda |u|^{p-2} u \quad\text{in }\Omega, \\ u = 0 \quad\text{on }\partial\Omega \end{gathered} \end{equation} to be solved for a real function $u:\Omega\to\mathbb{R}$ and a parameter $\lambda\in\mathbb{R}$. The operator $\Delta_p u := \operatorname{div} \left(|\nabla u|^{p-2}\nabla u\right)$ is called the $p$-Laplace operator. If for a certain $\lambda$ a nontrivial weak solution $u\in W^{1,p}_0(\Omega)$ of~\eqref{eq:evproblem} exists, we call $\lambda$ and $u$ Dirichlet eigenvalue and eigenfunction of the $p$-Laplace operator, respectively. Problem \eqref{eq:evproblem} is homogeneous but in general not additive in $u$. From \cite{Anane,Lindqvist1,Lindqvist1ad,BelloniKawohl1} and others it is well known that there exists a smallest eigenvalue $\lambda_1$ and that it is positive, isolated and simple (i.e., the corresponding eigenfunction $u_1$ is unique up to multiplication by a constant). Moreover, for any eigenfunction $u$ it holds: $u$ corresponds to $\lambda_1$ if and only if it does not change its sign on $\Omega$. In \cite{GarciaPeral} the authors constructed a nondecreasing sequence of eigenvalues accumulating at infinity using a variational approach. Since between $\lambda_1$ and the next member of this sequence there are no other eigenvalues, as it was shown in \cite{AnaneTsouli1}, we call this second smallest eigenvalue $\lambda_2$ and a corresponding eigenfunction $u_2$. In general, however, it is not known yet whether this sequence contains all the eigenvalues. Nodal domains of variational eigenfunctions were studied in \cite{DrabekRobinson}. The regularity results of \cite{DiBenedetto} imply that any eigenfunction (perhaps redefined on a class of measure zero) is of class $C^{1,\alpha}(\Omega)$ for some $\alpha>0$. An early attempt at computing several eigenpairs of the $p$-Laplace operator on a planar domain ($N=2$) numerically is due to Brown and Reichel \cite{BrownReichel}. Under the assumption of radial symmetry they used a shooting method for the resulting ordinary differential equation. The first genuinely two-dimensional approach was taken by Yao and Zhou \cite{YaoZhou1} using their local minimax method based on a variational formulation. For a square $\Omega=\{(x_1,x_2) : x_1,x_2\in(0,2)\}$ and $p\in\{1.75, 2.5, 3.0\}$ the authors computed approximations to seven eigenvalues and corresponding eigenfunctions. They observed that the found eigenfunction $u_2$ has an odd symmetry about $x_1=1$ for $p<2$ and about $x_1=x_2$ for $p>2$. The goal of the current work is to apply the numerical variational methods of \cite{Ho1} to compute approximations of the two smallest eigenvalues and to visualize the corresponding eigenfunctions on a planar domain. In particular the focus is \begin{itemize} \item to extend the Constrained Mountain Pass Algorithm from the Hilbert space setting (as described in~\cite{ChMcK1} and \cite{Ho1}) to the Banach space $W^{1,p}_0(\Omega)$; to verify that this algorithm is suitable even for computations with $p$ ``far'' from 2; \item to observe the behavior of the eigenpairs for a large range of $p$ and compare it with the known theoretical results about the asymptotics for $p\to 1$ and $p\to\infty$; \item to observe changes in symmetry of $u_2$ on various domains. \end{itemize} In Section~\ref{sec:background} we review known results about the variational properties of $\lambda_1$ and $\lambda_2$ and their asymptotic behavior. The variational numerical methods applied to compute the eigenpairs are summarized in Sec.~\ref{sec:num_method}. The choice of a descent direction in the Banach space $W^{1,p}_0(\Omega)$ is discussed in detail here, too. In Sec.~\ref{sec:numerical_results} we present the numerical results for several planar domains. We pay a particular attention to the dependence of the eigenvalues on $p$ and changes of symmetry of the second eigenfunction. Several issues concerning the application of the numerical methods (like mesh refinement, choice of parameters, etc.) are addressed in Sec.~\ref{sec:numerics_remarks}. Finally, Sec.~\ref{sec:conclusion} summarizes our numerical observations and the Appendix provides proofs of claims used in Sec.~\ref{sec:num_method}. \section{Background material}\label{sec:background} In the Introduction we mentioned the existence of the first two eigenvalues $\lambda_1$ and $\lambda_2$. Now we review some known results about their variational characterization based on the above references and their asymptotic behavior for $p$ close to 1 and $p$ large. \subsection{Variational characterization of $\lambda_1$ and $\lambda_2$} Define two continuously Fr\'echet differentiable functionals $I,J\in C^1\big(W^{1,p}_0(\Omega),\mathbb{R}\big)$: \begin{equation} \label{eq:IJ} I(u) := \int_\Omega |\nabla u|^p\ dx, \quad J(u) := \int_\Omega |u|^p\ dx. \end{equation} Their Fr\'echet derivatives $I'(u),J'(u)$ are members of the dual space of $W^{1,p}_0(\Omega)$ which we denote by $W^{-1,q}(\Omega)$, where $\frac{1}{p}+\frac{1}{q}=1$, and are given by \begin{equation} \label{eq:IJprime} \dualpairing{I'(u)}{\phi} = p\int_\Omega |\nabla u|^{p-2}\nabla u\nabla\phi\ dx ,\quad \dualpairing{J'(u)}{\phi} = p\int_\Omega |u|^{p-2} u\phi\ dx . \end{equation} Two observation can be made: 1.~After testing \eqref{eq:evproblem} with $\phi\in W^{1,p}_0(\Omega)$ and integrating by parts it becomes clear that \eqref{eq:evproblem} is the Euler-Lagrange equation $I'(u)-\lambda J'(u)=0$ (up to the factor $p$) which all critical points of $I$ with respect to the constraint \begin{equation} \label{eq:S} S:=\big\{u\in W^{1,p}_0(\Omega) : J(u)=1\big\} \end{equation} must satisfy for some value of the Lagrange multiplier $\lambda$. 2.~If $(\lambda,u)$ is an eigenpair and we test \eqref{eq:evproblem} with $u$, we obtain the Rayleigh quotient \begin{equation} \label{eq:rayleigh} \lambda=\frac{\int_\Omega|\nabla u|^p\ dx}{\int_\Omega|u|^p\ dx}. \end{equation} Since both its numerator and denominator are homogeneous of the same degree in $u$, finding the smallest eigenvalue $\lambda$ is the same as minimizing $I$ on $S$: \begin{equation} \label{eq:lambda1} \lambda_1=\min_{u\in S} I(u) . \end{equation} A variational minimax characterization of the second eigenvalue $\lambda_2$ based on the Krasnoselskii genus was given in \cite{GarciaPeral}. Alternatively, since for $u_1\in S$ both $u_1$ and $-u_1$ are local minimizers of $I$ on $S$, a mountain pass characterization of $\lambda_2$ is also possible \cite{CuestadeFigueiredoGossez}: \begin{equation} \label{eq:lambda2} \lambda_2=\inf_{\gamma\in\Gamma}\max_{u\in\gamma([0,1])}I(u), \end{equation} where $\Gamma=\{\gamma\in C([0,1],S) : \gamma(0)=u_1,\ \gamma(1)=-u_1\}$ is the family of all paths in $S$ connecting the two local minimizers. Hence for the numerical computations we have the setting required by the Constrained Mountain Pass Algorithm of \cite{Ho1}. \subsection{Asymptotic behavior of $\lambda_1$ and $\lambda_2$ as $p\to 1$} To make the dependence of an eigenvalue on the domain $\Omega$ and the parameter $p$ explicit in our notation we will write $\lambda(\Omega;p)$ if necessary (and similarly for eigenfunctions). The main result of \cite{KawohlFridman} implies that for $\Omega$ with a Lipschitz boundary \begin{equation} \label{eq:lambda1to1} \lim_{p\to 1}\lambda_1(\Omega;p) = h_1(\Omega), \quad\text{where } h_1(\Omega):=\min_{D\subset\Omega}\frac{\operatorname{Per}(D)}{|D|} \end{equation} is called Cheeger constant, $\operatorname{Per}(D)$ denotes the perimeter of $D$ measured with respect to $\mathbb{R}^N$ and $|D|$ its $N$-dimensional Lebesgue measure. A minimizing set in the definition of $h_1(\Omega)$ is called a Cheeger set of $\Omega$. Furthermore, any convex planar domain $\Omega$ possesses a unique Cheeger set $\mathcal{C}_\Omega$ and \begin{equation} \label{eq:u1to1} \lim_{p\to 1}u_1(\Omega;p)=\chi_{\mathcal{C}_\Omega} \quad\text{in $L^1$ along a subsequence.} \end{equation} Here the eigenfunctions $u_1(\Omega;p)$ have been normalized to 1 in the $L^\infty$-norm, $\chi_{\mathcal{C}_\Omega}$ is the indicator function of $\mathcal{C}_\Omega$. A detailed description of how to find the Cheeger set $\mathcal{C}_\Omega$ for a convex planar domain $\Omega$ is given in \cite{LachandRobertKawohl}. Its main property is \begin{equation} \label{eq:Cheegerset} \mathcal{C}_\Omega=\cup \{B\subset\Omega : B\text{ is a ball of radius } \textstyle\frac{1}{h_1(\Omega)}\}. \end{equation} In \cite{Parini} it was shown that for $\Omega$ with a Lipschitz boundary it holds: \begin{equation} \label{eq:lambda2to1} \lim_{p\to 1}\lambda_2(\Omega;p) = h_2(\Omega), \end{equation} where \begin{equation} \label{eq:h2} h_2(\Omega):=\min\{\mu\in\mathbb{R}:\exists D_1,D_2\subset\Omega, D_1\cap D_2=\emptyset\text{ and } \max_{i=1,2}\frac{\operatorname{Per}(D_i)}{|D_i|}\leq\mu\} \end{equation} is called the second Cheeger constant and the convention $\operatorname{Per}(D)/|D|=\infty$ is used if $|D|=0$. Any two sets $D_1, D_2$ for which the minimum in the definition of $h_2(\Omega)$ is achieved are called coupled Cheeger sets of $\Omega$. For a result about the $L^1$-convergence of the second eigenfunctions we refer to \cite[Thm.~5.11]{Parini}. \subsection{Asymptotic behavior of $\lambda_1$ and $\lambda_2$ as $p\to\infty$} For a bounded domain $\Omega$ of $\mathbb{R}^N$ a limit problem of \eqref{eq:evproblem} as $p\to\infty$ is studied in \cite{JuutinenLindqvistManfredi,JuutinenLindqvist} for an unknown function $u$ and an unknown real parameter $\Lambda$ (see \cite[Definition 2.1]{JuutinenLindqvist}). The smallest $\Lambda$ for which this limit problem admits a nontrivial viscosity solution is called the first $\infty$-eigenvalue and denoted $\Lambda_1$. For $\Lambda_1$ there exists a positive viscosity solution and it holds: \begin{gather} \label{eq:lambda1toinfty} \lim_{p\to\infty}\big(\lambda_1(\Omega;p)\big)^{1/p} = \Lambda_1(\Omega),\\ \Lambda_1(\Omega) =\frac{1}{r_1}, \quad\text{where } r_1:=\sup\{r>0: \exists \text{ an open ball } B\subset\Omega \text{ of radius } r\}, \label{eq:Lambda1} \\ \Lambda_1(\Omega) =\min\big\{\frac{\|\nabla u\|_{L^\infty(\Omega)}} {\|u\|_{L^\infty(\Omega)}}: u\in W^{1,\infty}_0(\Omega)\setminus\{0\}\big\}.\label{eq:Lambda1var} \end{gather} The characterization \eqref{eq:Lambda1var} is an analogy of \eqref{eq:rayleigh} and \eqref{eq:lambda1}. Furthermore, for any sequence $\{u_1(\Omega;p_i)\}_{i=1}^\infty$ with $p_i\to\infty$ and $\|u_1(\Omega;p_i)\|_{L^{p_i}(\Omega)}=1$ there exists a subsequence converging uniformly to a viscosity solution of the limit problem for $\Lambda_1(\Omega)$. The smallest $\Lambda$ for which the limit problem admits a viscosity solution with at least two nodal domains is called the second $\infty$-eigenvalue and denoted $\Lambda_2$. From the definition it follows that $\Lambda_1\leq\Lambda_2$. If this inequality is strict, then zero is the only solution of the limit problem for $\Lambda\in(\Lambda_1,\Lambda_2)$. It holds: \begin{gather} \label{eq:lambda2toinfty} \lim_{p\to\infty}\big(\lambda_2(\Omega;p)\big)^{1/p} = \Lambda_2(\Omega), \\ \Lambda_2(\Omega) =\frac{1}{r_2}, \label{eq:Lambda2} \end{gather} where $r_2:=\sup\{r>0:\exists \text{ disjoint open balls } B_1, B_2\subset\Omega \text{ of radius } r\}$, \begin{equation} \Lambda_2(\Omega) =\inf_{\gamma\in\Gamma}\max_{u\in\gamma([0,1])} \|\nabla u\|_{L^\infty(\Omega)}, \label{eq:Lambda2var} \end{equation} where $\Gamma$ is defined as in \eqref{eq:lambda2}, $u_1$ is any first $\infty$-eigenfunction and $S:=\{u\in W^{1,\infty}_0(\Omega) : \|u\|_{L^\infty(\Omega)}=1\}$. Furthermore, for any sequence $\{u_2(\Omega;p_i)\}_{i=1}^\infty$ with $p_i\to\infty$ and $\|u_2(\Omega;p_i)\|_{L^{p_i}(\Omega)}=1$ there exists a subsequence converging uniformly to a viscosity solution of the limit problem for $\Lambda_2(\Omega)$ which has at least two nodal domains. \section{Numerical methods}\label{sec:num_method} An overview of the numerical methods used to compute approximations of the first and the second Dirichlet eigenpair of the $p$-Laplace operator is given in Fig.~\ref{fig:flowchart}. In this section we will describe these methods. Our goal is to find $u_1$ as a minimizer of $I$ on $S$ according to \eqref{eq:lambda1} and $u_2$ as a mountain pass point of $I$ on $S$ according to \eqref{eq:lambda2}. We first discretize the planar domain $\Omega$ using a mesh of triangles and apply the finite element method to approximate $W^{1,p}_0(\Omega)$ by a finite dimensional subspace. Then we fix $p\in(1,\infty)$ and use a variant of the \emph{Constrained Steepest Descent Method} (CDM) to find the first eigenpair, and the \emph{Constrained Mountain Pass Algorithm} (CMPA) to find the second eigenpair. We implement both methods based on \cite{Ho1}. There are, however, several important issues arising from the fact that we work in a Banach space and not a Hilbert space as in \cite{Ho1}. How to deal with these issues will also be explained in this section. For the computation of the descent direction the Augmented Lagrangian Method of \cite{GlowinskiMarroco} is applied. \begin{figure}[h] \centering \setlength{\unitlength}{1mm} \begin{picture}(108,56)(0,14) \put(-1,10){\includegraphics{figs/flowchart.pdf}} \put(2.5,29.5){\parbox{23mm}{\centering\small Triangulation of~$\Omega$, FEM}} \put(39.5,29.5){\parbox{23mm}{\centering\small CDM (iterative)}} \put(76.5,29.5){\parbox{23mm}{\centering\small CMPA (iterative)}} \put(57.6,49){\parbox{24mm}{\centering\small Aug.\ Lagrangian for $(-\Delta_p)^{-1}$ (iterative)}} \put(34,14.5){\parbox{23mm}{\centering\small $e_0$}} \put(45.5,13.5){\parbox{23mm}{\centering\small $(\lambda_1,u_1)$}} \put(68,31.5){\small $u_1$} \put(71,14.5){\parbox{23mm}{\centering\small $e_\mathrm{M}$}} \put(85,13.5){\parbox{23mm}{\centering\small $(\lambda_2,u_2),\dots$}} \put(28, 64.5){\parbox{83mm}{\centering\small loop over a range of values of $p\in(1,\infty)$}} \end{picture} \caption{Flowchart of the numerical computations.} \label{fig:flowchart} \end{figure} \subsection{Finite element method}\label{sec:fem} A finite element approximation of the $p$-Laplacian was studied in \cite{BarrettLiu}. We adopt this approach for our computations. The planar domain $\Omega$ is approximated by a polygonal domain $\Omega^h$ which is partitioned into a finite number of triangles of diameter at most $h$. Let $\{a_i\}_{i=1}^k$ be the set of those triangle vertices which lie in the interior of $\Omega^h$. Functions $\{\phi_i\}_{i=1}^k$ forming a basis of the $k$-dimensional subspace $V^h_0$ of $W^{1,p}_0(\Omega^h)$ are chosen linear on each triangle with $\phi_i(a_j)=\delta_{ij}$, where $\delta_{ij}$ is the Kronecker delta, and zero on $\partial\Omega^h$. The space $V^h_0$ is our finite element approximation of the Sobolev space $W^{1,p}_0(\Omega)$. In \cite{BarrettLiu} a detailed description of this method was given for the boundary value problem \begin{equation} \label{eq:basicproblem} \begin{gathered} -\Delta_p u = f \quad\text{in }\Omega, \\ u = 0 \quad\text{on }\partial\Omega \end{gathered} \end{equation} with the right-hand side $f\in L^2(\Omega)$. Since it is a straightforward task to adapt it to our problem \eqref{eq:evproblem} with $\lambda |u|^{p-2}u$ on the right-hand side, we will not show the details here. We will however mention one additional technical detail involved. The evaluation of the functionals given in \eqref{eq:IJ} and \eqref{eq:IJprime} for functions from $V^h_0$ amounts to adding up the contributions of the individual triangles that make up $\Omega^h$. For example, for $I(u)$ with $u\in V^h_0$ one merely needs to integrate a constant on every triangle. The situation is different for $J(u)$. Let $T\subset\Omega^h$ be a triangle with area $|T|$ and vertices $A$, $B$, and $C$ and let $u$ be a linear function on $T$ with values $u_A$, $u_B$, and $u_C$ at these vertices, respectively. If these values are mutually different, then the following formula holds: \begin{equation} \label{eq:Jontriangle} \begin{split} \int_T |u|^p\,dx &= \frac{2\,|T|}{(p+1)(p+2)(u_C-u_A)}\\ &\quad\times \Big(\frac{|u_C|^{p+2}-|u_B|^{p+2}}{u_C-u_B} - \frac{|u_A|^{p+2}-|u_B|^{p+2}}{u_A-u_B} \Big). \end{split} \end{equation} By inspecting this formula we see that great care must be taken when implementing it to avoid numerical cancellations. This is crucial for the success of our method. A similar situation occurs when evaluating $\dualpairing{J'(u)}{\phi}$ (or the right-hand side of the equation in \eqref{eq:evproblem} in the weak formulation). \subsection{Direction of descent}\label{sec:descent} An important ingredient of the variational numerical methods CDM and CMPA is finding a descent direction of the functional $I$ on the constraint set $S$. How this is accomplished in the Hilbert space setting was shown in~\cite{Ho1}: Let $\nabla I(u)$ be the Riesz representation of $I'(u)$ (i.e., the gradient) and $P_u$ the orthogonal projection on the tangent space of $S$ at $u\in S$. Then \begin{equation}\label{eq:orthogonalprojection} w_u=-P_u\nabla I(u),\quad u\in S \end{equation} gives the steepest descent direction of $I$ at $u$ with respect to $S$. Because of the lack of orthogonality in the Banach space $W^{1,p}_0(\Omega)$ we need to take a different approach. Let \begin{equation} \label{eq:tgspace} T_uS:=\big\{v\in W^{1,p}_0(\Omega):\dualpairing{J'(u)}{v}=0\big\},\quad \|v\|:=\Big(\int_\Omega|\nabla v|^p\,dx\Big)^{1/p} \end{equation} denote the tangent space of $S$ at $u\in S$ and the norm of $v\in W^{1,p}_0(\Omega)$, respectively. The problem of finding the steepest descent direction of $I$ with respect to $S$ can be written as follows: for a given $u\in S$ which is not a critical point of $I$ with respect to $S$ \begin{equation} \label{eq:steepestdescentBanach} \text{minimize } \dualpairing{I'(u)}{w} \quad \text{subject to}\quad w\in\{v\in T_uS : \|v\|=1\}. \end{equation} The existence of a unique solution can be proved in a straightforward way using a minimizing sequence (in a reflexive Banach space, in general). This solution must satisfy the Euler-Lagrange equation \begin{equation} \label{eq:elg_steepest_descent} -\Delta_p w = \beta\left( -\Delta_p u - \alpha |u|^{p-2}u \right), \end{equation} where $\alpha,\beta\in\mathbb{R}$ are unknown. The coefficient $\beta$ comes from the requirement $\|w\|=1$. After testing \eqref{eq:elg_steepest_descent} by $w$ it can be seen that $\beta<0$ since the minimum in \eqref{eq:steepestdescentBanach} is negative. For $p\neq 2$ finding the right $\alpha$ is not an easy problem. We will try to find a different convenient descent direction instead, not necessarily the steepest one. A simple calculation shows that under no constraints the steepest descent direction of $I$ at $u\in W^{1,p}_0(\Omega)$ is given by $-u$. For $u\in S$ we consider the point $w_u\in T_u S$ closest to $-u$, i.e., the unique solution of the minimization problem \begin{equation} \label{eq:descentBanach} \text{minimize } \|w+u\| \quad \text{subject to}\quad w\in T_uS. \end{equation} The minimizer must satisfy the Euler-Lagrange equation \begin{equation} \label{eq:elg_descent} -\Delta_p(w+u)=\alpha |u|^{p-2}u \end{equation} for some $\alpha\in\mathbb{R}$. Unlike \eqref{eq:elg_steepest_descent}, this equation can be solved easily for $w$: \begin{equation} \label{eq:descent_direction} w_u=-u+\frac{1}{\int_\Omega |u|^{p-2}u \,v_u\,dx}\ v_u,\quad \text{where }v_u:=(-\Delta_p)^{-1}\left(|u|^{p-2}u\right). \end{equation} The operator $(-\Delta_p)^{-1}$ is discussed later in Sec.~\ref{sec:inverse}. Lemma \ref{lemma2} in the Appendix shows that $w_u$ is, indeed, a descent direction of $I$ with respect to $S$. This descent direction is used in our implementation of the variational numerical methods CDM and CMPA. \begin{remark} \label{rmk1} \rm 1.~Observe that if $-\Delta_p$ were linear, equations \eqref{eq:elg_steepest_descent} and \eqref{eq:elg_descent} would coincide (after setting $\beta=-1$). Hence in case $p=2$ they yield the same descent direction (which is the one given by \eqref{eq:orthogonalprojection}). 2.~With $\beta=-1$ in \eqref{eq:elg_steepest_descent}, both equations \eqref{eq:elg_steepest_descent} and \eqref{eq:elg_descent} yield a zero solution if and only if $u$ is a critical point of $I$ with respect to $S$. \end{remark} \subsection{Inverse of the $p$-Laplace operator}\label{sec:inverse} A classical result (see, e.g., \cite[Theorem 1.3]{Struwe}) says that for any $f\in W^{-1,q}(\Omega)$, the dual of $W^{1,p}_0(\Omega)$ with $\frac{1}{p}+\frac{1}{q}=1$, the problem \begin{equation} \label{eq:constant_rhs} \begin{gathered} -\Delta_p u = f \quad\text{in }\Omega, \\ u = 0 \quad\text{on }\partial\Omega \end{gathered} \end{equation} has a unique weak solution in $W^{1,p}_0(\Omega)$. This means that the operator \begin{equation} \label{eq:operator} -\Delta_p: W^{1,p}_0(\Omega)\to W^{-1,q}(\Omega) \quad\text{given by}\quad \dualpairing{-\Delta_p u}{v}=\int_\Omega|\nabla u|^{p-2}\nabla u\nabla v\,dx \end{equation} is invertible. We denote its inverse by $(-\Delta_p)^{-1}$. In order to use the descent direction given by \eqref{eq:descent_direction} we need to compute $v_u$ first, i.e., we need to solve problem \eqref{eq:constant_rhs} numerically. For that we apply the Augmented Lagrangian Method of \cite{GlowinskiMarroco}. Here we give a brief description of this method. Let $V^h_0$ be again the subspace of continuous functions of $W^{1,p}_0(\Omega^h)$ which are linear on every triangle of a triangulation of $\Omega^h$, $D^h$ the space of functions with values in $\mathbb{R}^2$ defined on $\Omega^h$ which are constant on each triangle, and $r>0$ a parameter. For the Augmented Lagrangian \begin{equation} \label{eq:augmented_lagrangian} \mathcal{L}_r(v,t,\mu):=\frac{1}{p}\int_\Omega |t|^p\,dx - \dualpairing{f}{v}+ \frac{r}{2}\int_\Omega|\nabla v-t|^2\,dx + \int_\Omega\mu\cdot(\nabla v-t)\,dx, \end{equation} where $v\in V^h_0$ and $t,\mu\in D^h$, a saddle point $(u,s,\eta)$ is searched for such that \begin{equation} \label{eq:al_saddle_point} \mathcal{L}_r(u,s,\mu)\leq\mathcal{L}_r(u,s,\eta)\leq\mathcal{L}_r(v,t,\eta) \quad \forall (v,t,\mu)\in V^h_0\times D^h\times D^h. \end{equation} A sequence $\big(u^{(n)},s^{(n)},\eta^{(n)}\big)$ approximating $(u,s,\eta)$ is constructed as follows: choose $\big(s^{(0)},\eta^{(1)}\big)\in D^h\times D^h$ and for $n\in\mathbb{N}$ solve \begin{gather} %% The first two lines belong together and should share one label %% (similar as in \ref{eq:constant_rhs}) \begin{gathered} -r\Delta u^{(n)} = f+\eta^{(n)}\cdot\nabla - r\,s^{(n-1)}\cdot\nabla \quad\text{in }\Omega, \\ u^{(n)} =0 \quad\text{on }\partial\Omega, \end{gathered}\label{al_step1}\\ \big|s^{(n)}\big|^{p-2}s^{(n)}+r\,s^{(n)} = r\nabla u^{(n)}+\eta^{(n)}, \label{al_step2}\\ \eta^{(n+1)}=\eta^{(n)}+r\big(\nabla u^{(n)}-s^{(n)}\big).\label{al_step3} \end{gather} For given $s^{(n-1)}$ and $\eta^{(n)}$ the boundary value problem \eqref{al_step1} can be solved for $u^{(n)}$. For this, one just needs some standard algorithm for finding the inverse of the Laplace operator with Dirichlet boundary conditions. The equation in \eqref{al_step1} is understood in the weak sense: for example, the term $\eta^{(n)}\cdot\nabla$ is evaluated as $\int_\Omega\eta^{(n)}\cdot\nabla\phi\,dx$ for a test function $\phi\in V^h_0$. Next, equation \eqref{al_step2} is used to find $s^{(n)}$. The $\mathbb{R}^2$-norm of $s^{(n)}$ must satisfy \begin{equation} \label{eq:al_step2a} \big|s^{(n)}\big|^{p-1}+r\,\big|s^{(n)}\big|=\big|r\,\nabla u^{(n)}+\eta^{(n)}\big|, \end{equation} which on each triangle is just a scalar nonlinear equation with one unknown. For each triangle it can be solved, e.g., by Newton's method. After $|s^{(n)}|$ has been obtained, $s^{(n)}$ can be computed immediately from \eqref{al_step2}. At the end $\eta$ is updated according to \eqref{al_step3} and a new iteration step can be started. The convergence of this method was studied in \cite{GlowinskiMarroco}. We use the norm of $\nabla u^{(n)}-s^{(n)}$ to measure the convergence. \subsection{Constrained Descent Method}\label{sec:CDM} The Constrained Descent Method (CDM) is applied to find the first eigenpair of the $p$-Laplace operator: $u_1$ is found as the minimizer of $I$ with respect to $S$, $\lambda_1=I(u_1)$. As mentioned above, it differs from the Constrained Steepest Descent Methods of \cite{Ho1} in the way the descent direction is chosen. The method solves numerically the initial value problem \begin{equation} \label{eq:cdm_ivp} \frac{d}{d t}u(t)=w_{u(t)},\quad u(0)=e_0\in S, \end{equation} where $w_u$ is given by \eqref{eq:descent_direction} for $u\in S$. Proposition~\ref{proposition1} in the Appendix states that this problem has a unique solution $u(t)\in S$ for $t\in(0,\infty)$ and that $u(t)$ gets arbitrarily close to a critical point of $I$ with respect to $S$ as $t\to\infty$. After choosing the starting point $e_0\in S$ and setting $u^{(0)}:=e_0$ the initial value problem is solved by repeating the following two steps: First (Euler's step), given $u^{(n-1)}$ find $\bar u^{(n)}=u^{(n-1)}+\Delta t^{(n)}\,w_{u^{(n-1)}}$ with some small value $\Delta t^{(n)}>0$. Second (scaling), define $u^{(n)}=c\bar u^{(n)}$, where the coefficient $c\in\mathbb{R}$ is chosen such that $u^{(n)}\in S$. In case $I\big(u^{(n)}\big) > I\big(u^{(n-1)}\big)$, halve the step $\Delta t^{(n)}$ and compute $\bar u^{(n)}$ and $u^{(n)}$ again. If this halving has to be repeated and $\Delta t^{(n)}$ becomes very small (smaller than a prescribed threshold value), stop the algorithm. The norm of the descent direction $\|w_{u^{(n-1)}}\|$ is used to measure convergence of $u^{(n)}$ to an eigenfunction $u$. When computing $w_u$ according to \eqref{eq:descent_direction} the integral $\nu:=\int_\Omega |u|^{p-2}u \,v_u\,dx$ has to be evaluated. If $\|w_{u^{(n-1)}}\|$ is small, then $\big(1/\nu^{(n-1)}\big)^{p-1}$ approximates the eigenvalue. We note that at every step of CDM the Augmented Lagrangian Method of Sec.~\ref{sec:inverse} has to be applied to compute the descent direction $w_{u^{(n-1)}}$. \subsection{Constrained Mountain Pass Algorithm}\label{sec:CMPA} Suppose that an approximation of the first eigenvalue $\lambda_1$ and eigenfunction $u_1$ of the $p$-Laplace operator have been computed. Constrained Mountain Pass Algorithm (CMPA) is applied to find the second eigenpair: $u_2$ is found as a mountain pass point of $I$ on $S$ lying ``between'' the two local minimizers $u_1$ and $-u_1$, $\lambda_2=I(u_2)$. Again, it differs from CMPA described in detail in \cite{Ho1} in the choice of the descent direction. We give a short summary here based on the original description of the Mountain Pass Algorithm by Choi and McKenna \cite{ChMcK1}: Take a discretized path $\{z_j\}_{j=0}^{P}\subset S$ connecting $z_0:=u_1$ with $z_P:=-u_1$. After finding the path point $z_m=:z^\mathrm{max}$ at which $I$ is maximal along the path, move this point a small distance in the tangent space to~$S$ at $z^\mathrm{max}$ in the descent direction $w_{z^\mathrm{max}}$ and then scale it (as in CDM) to come back to $S$. Thus the path has been deformed on~$S$ and the maximum of $I$ lowered. Repeat this deforming of the path until the maximum along the path cannot be lowered anymore: a mountain pass point of $I$ with respect to~$S$ has been reached. To construct the initial path connecting $u_1$ and $-u_1$ in $S$ we choose an intermediate point $e_\mathrm{M}\in S\setminus\{\pm u_1\}$, set $k:=[P/2]$ and define: \begin{gather*} \bar z_j:=u_1+\frac{j}{k}(e_\mathrm{M}-u_1) \quad\text{for }j\in\{0,\dots,k\}, \\ \bar z_j:=e_\mathrm{M}+\frac{j-k}{P-k}(-u_1-e_\mathrm{M}) \quad\text{for }j\in\{k,\dots,P\}, \\ z_j:=c_j\bar z_j\in S\text{ (scaling to $S$ as in Sec.~\ref{sec:CDM})} \quad\text{for }j\in\{0,\dots,P\}. \end{gather*} Connecting $u_1$ and $-u_1$ by a line segment without the intermediate point $e_\mathrm{M}$ would not work. Such a line segment passes through $0$ and cannot be scaled to get to $S$. Finally, as in CDM, $\|w_{z^\mathrm{max}}\|$ is used to measure convergence to an eigenfunction $u$. The corresponding eigenvalue $\lambda$ is computed as in CDM, too. At every step of CMPA the Augmented Lagrangian Method of Sec.~\ref{sec:inverse} has to be applied to compute the descent direction $w_{z^\mathrm{max}}$. \section{Numerical results}\label{sec:numerical_results} In this section numerical results will be given for the following planar domains: the unit disk, the square with side length 2, the rectangle with sides 2 and 7/4, the isosceles triangle with base and height 1, the isosceles triangle with base 1 and height 3/4, and the equilateral triangle with side 1. Unless explicitly stated otherwise the computed eigenfunctions will be plotted as a surface over the domain with heights given by the function values and as a contour plot of these values (like, e.g., in Fig.~\ref{fig:disk_u2}). In order to better compare the shapes, the eigenfunctions in these figures have been scaled to have the same maximum value. We do not explicitly differentiate between two eigenfunctions $u$ and $\tilde u$ if $\tilde u(x)=cu(Tx)$, where $c\in\mathbb{R}$ is a scaling coefficient and $T:\Omega\to\Omega$ is some symmetry transformation of $\Omega$ (e.g., for a square a rotation by $\pi/2$ about the center of the square). Animated graphs of the computed eigenfunctions $u_1$ and $u_2$ for the considered domains are available as an electronic data supplement to this paper from author's home page and using pointers on the journal's web page. \subsection{Unit Disk} Let \[ \Omega:=\left\{(x_1,x_2)\in\mathbb{R}^2:x_1^2+x_2^2<1\right\}. \] Before presenting the numerical results we make a remark about the radially symmetric case. It is known that the first eigenfunction for the disk is radially symmetric. One important question about the second eigenfunction for the disk has been whether it is radially symmetric, too. In \cite{Parini,BenDrabGirg} the authors proved that for $p$ close to 1 the answer is no. The eigenvalue problem \eqref{eq:evproblem} under the assumption of radial symmetry $u=u(r)$, $r\in(0,1)$ becomes \begin{equation} \label{eq:evproblem_radsym} \begin{gathered} -\left(r|u'|^{p-2}u'\right)^\prime = \lambda r |u|^{p-2}u, \\ u'(0)=0,\quad u(1)=0. \end{gathered} \end{equation} This and a related problem are treated, for example, in \cite{BrownReichel} and \cite{BenDrabGirg}, where numerical approaches play an important role. For the current numerical investigation we adapt our \emph{genuine 2D method} described in Sec.~\ref{sec:num_method} in the following ways: \begin{itemize} \item all integrals are one-dimensional, \item the weight $r$ is introduced, \item the natural boundary condition is implemented at $r=0$ (the zero boundary condition stays at $r=1$). \end{itemize} Since these modifications are rather elementary, we will not describe them in more detail. We will refer to this method as \emph{radial 2D method}. For the computations carried out by the genuine 2D method the domain $\Omega$ was approximated by a polygon and discretized using 68,608 triangles. For the computations carried out by the radial 2D method the interval $(0,1)$ was divided into 1,000 subintervals of the same length. \begin{table}[h] \centering\vspace{1ex} \begin{tabular}[t]{c|c|c|c|c|cc|c|c|c|} \cline{2-5}\cline{8-10} (a) & $p$ & $\lambda_1$ & $\lambda_2$ & \multicolumn{1}{c|}{$\lambda_2^\mathrm{rad}$} & \quad\quad & (b) & $p$ & $\lambda_1$ & $\lambda_2^\mathrm{rad}$ \\\cline{2-5}\cline{8-10} \multicolumn{10}{c}{}\\\cline{2-5}\cline{8-10} & 1.1 & 2.5690 & 4.2008 & 5.6809 &&& 1.1 & 2.5688 & 5.6762 \\\cline{2-5}\cline{8-10} & 1.2 & 2.9654 & 5.0707 & 7.2277 &&& 1.2 & 2.9653 & 7.2251 \\\cline{2-5}\cline{8-10} & 1.3 & 3.3263 & 5.9604 & 8.9302 &&& 1.3 & 3.3260 & 8.9279 \\\cline{2-5}\cline{8-10} & 1.4 & 3.6740 & 6.9072 & 10.861 &&& 1.4 & 3.6739 & 10.858 \\\cline{2-5}\cline{8-10} & 1.5 & 4.0179 & 7.9310 & \multicolumn{1}{c}{} &&& 1.5 & 4.0177 & 13.073 \\\cline{2-4}\cline{8-10} & 1.6 & 4.3623 & 9.0465 & \multicolumn{1}{c}{} &&& 1.6 & 4.3621 & 15.626 \\\cline{2-4}\cline{8-10} & 1.7 & 4.7097 & 10.266 & \multicolumn{1}{c}{} &&& 1.7 & 4.7095 & 18.574 \\\cline{2-4}\cline{8-10} & 1.8 & 5.0618 & 11.604 & \multicolumn{1}{c}{} &&& 1.8 & 5.0616 & 21.982 \\\cline{2-4}\cline{8-10} & 1.9 & 5.4194 & 13.072 & \multicolumn{1}{c}{} &&& 1.9 & 5.4192 & 25.921 \\\cline{2-4}\cline{8-10} & 2.0 & 5.7834 & 14.683 & \multicolumn{1}{c}{} &&& 2.0 & 5.7831 & 30.471 \\\cline{2-4}\cline{8-10} & 2.1 & 6.1542 & 16.452 & \multicolumn{1}{c}{} &&& 2.1 & 6.1539 & 35.725 \\\cline{2-4}\cline{8-10} & 2.2 & 6.5320 & 18.395 & \multicolumn{1}{c}{} &&& 2.2 & 6.5317 & 41.788 \\\cline{2-4}\cline{8-10} & 2.3 & 6.9173 & 20.527 & \multicolumn{1}{c}{} &&& 2.3 & 6.9169 & 48.780 \\\cline{2-4}\cline{8-10} & 2.4 & 7.3102 & 22.866 & \multicolumn{1}{c}{} &&& 2.4 & 7.3097 & 56.836 \\\cline{2-4}\cline{8-10} & 2.5 & 7.7107 & 25.432 & \multicolumn{1}{c}{} &&& 2.5 & 7.7102 & 66.112 \\\cline{2-4}\cline{8-10} & 3.0 & 9.8323 & 42.460 & \multicolumn{1}{c}{} &&& 3.0 & 9.8314 & 137.93 \\\cline{2-4}\cline{8-10} & 4.0 & 14.683 & 110.71 & \multicolumn{1}{c}{} &&& 4.0 & 14.681 & 559.02 \\\cline{2-4}\cline{8-10} & 5.0 & 20.351 & 273.00 & \multicolumn{1}{c}{} &&& 5.0 & 20.347 & 2,132.7 \\\cline{2-4}\cline{8-10} & 6.0 & 26.832 & 649.47 & \multicolumn{1}{c}{} &&& 6.0 & 26.823 & 7,822.6 \\\cline{2-4}\cline{8-10} & 8.0 & 42.210 & 3,430.1 & \multicolumn{1}{c}{} &&& 8.0 & 42.182 & 97,462 \\\cline{2-4}\cline{8-10} & 10.0 & 60.784 & 17,071 & \multicolumn{1}{c}{} &&& 10.0 & 60.715 & $1.1359\cdot 10^6$ \\\cline{2-4}\cline{8-10} \end{tabular}\vspace{2ex} \caption{Eigenvalues for the disk computed numerically by: (a) the genuine 2D method, (b) the radial 2D method.} \label{tab:disk_lambda} \end{table} Figure \ref{fig:disk_u2} shows the eigenfunction $u_2$ computed by the genuine 2D method for several values of $p$. The computed eigenvalues $\lambda_1$ and $\lambda_2$ for these and other values of $p$ are listed in Table~\ref{tab:disk_lambda}(a). Figure~\ref{fig:disk_em}(a) shows the shape of the intermediate point $e_\mathrm{M}$ on the initial path connecting $u_1$ and $-u_1$ we used for CMPA to find $u_2$ for all the listed values of $p$. The function $u_2$ found this way seems to posses an odd symmetry with respect to its nodal line. The slope of this nodal line in the coordinate system $(x_1,x_2)$ depends on the computation. For the depiction in Fig.~\ref{fig:disk_u2} we rotated $\Omega$ in each case to make the slope appear the same. CMPA needed between 120 and 600 iterations to converge. \newcommand{\dsutwographs}[2]{\begin{picture}(25,55) \put(0,0){\makebox(25,4){$p=#1$}} \put(.5,28){\includegraphics[width=24mm, clip=true, viewport= 175 75 450 395]{figs/D_disk/#2.jpg}} \put(1,6){\includegraphics[width=23mm]{figs/D_disk/#2.pdf}} \end{picture} } \begin{figure}[t] \centering \setlength{\unitlength}{1mm} \begin{picture}(127,55) \put(0,0){\dsutwographs{1.1}{out_u2_p_1-1b}} \put(25.5,0){\dsutwographs{1.4}{out_u2_p_1-4}} \put(51,0){\dsutwographs{2.0}{out_u2_p_2-0}} \put(76.5,0){\dsutwographs{2.5}{out_u2_p_2-5}} \put(102,0){\dsutwographs{8.0}{out_u2_p_8-0}} \end{picture} \caption{The numerically computed second eigenfunction $u_2$ for the disk.} \label{fig:disk_u2} \end{figure} \begin{figure}[t] \centering \setlength{\unitlength}{1mm} \begin{picture}(127,55) \put(0,0){\dsutwographs{1.1}{out_u2rad_p_1-1}} \put(0,52){(a)} \put(27,29){\includegraphics[width=54mm]{figs/D_disk/out_u2rad_1d_p_1-1.pdf}} \put(77,29){\includegraphics[width=54mm]{figs/D_disk/out_u2rad_1d_p_1-4.pdf}} \put(27,0){\includegraphics[width=54mm]{figs/D_disk/out_u2rad_1d_p_2-5.pdf}} \put(77,0){\includegraphics[width=54mm]{figs/D_disk/out_u2rad_1d_p_8-0.pdf}} \put(65,50){$p=1.1$} \put(115,50){$p=1.4$} \put(65,19){$p=2.5$} \put(115,19){$p=8.0$} \put(27,52){(b)} \end{picture} \caption{The numerically computed radially symmetric second eigenfunction $u_2^\mathrm{rad}$: (a) using the genuine 2D method; (b) using the radial 2D method (the profile for the radial coordinate $r\in(0,1)$ is shown, scaled such that $\|u_2^\mathrm{rad}\|_p=1$).} \label{fig:disk_u2rad} \end{figure} Figure~\ref{fig:disk_em}(b) shows an alternative shape of $e_\mathrm{M}$. With such an initial path CMPA converged for $p=1.1$ and $p=1.2$ to a radially symmetric function we call $u_2^\mathrm{rad}$ (but for higher values of $p$ to the oddly symmetric function $u_2$). Figure~\ref{fig:disk_u2rad}(a) shows $u_2^\mathrm{rad}$ for $p=1.1$. Figure~\ref{fig:disk_em}(c) shows yet another choice of $e_\mathrm{M}$ (radially symmetric). With this intermediate point of the initial path and for $p=1.3$ and $p=1.4$ (but not larger) CMPA seems to converge to a radially symmetric function first but after many iterations the path slips down and the algorithm converges eventually to the oddly symmetric $u_2$. The graph in Fig.~\ref{fig:disk_em}(d) shows how the maximum value of the Dirichlet functional $I$ along the path develops during the run of the algorithm (for $p=1.3$). The horizontal axis shows the number of iterations. The flat part between iterations 70 and 260 indicates that the path is staying close to a critical point. When now the norm of the descent direction $w_{z^\mathrm{max}}^{}$ given in \eqref{eq:descent_direction} computed at the ``highest'' point $z^\mathrm{max}$ of the path gets small enough, we stop the algorithm and save this highest point. Since it displays a radial symmetry, we call it $u_2^\mathrm{rad}$ again. \newcommand{\dsemgraphs}[2]{% \begin{picture}(25,48) \put(.5,22){\includegraphics[width=24mm, clip=true, viewport= 175 75 450 395]{figs/D_disk/#2.jpg}} \put(1,0){\includegraphics[width=23mm]{figs/D_disk/#2.pdf}} \put(0,45){#1} \end{picture} } \begin{figure}[t] \centering \setlength{\unitlength}{1mm} \begin{picture}(127,48) \put(0,0){\dsemgraphs{(a)}{em_mountain_p_1-8}} \put(25.5,0){\dsemgraphs{(b)}{em_mountain_rad_p_1-1}} \put(51,0){\dsemgraphs{(c)}{em_mountain_rad_p_1-3}} \put(79,3){\includegraphics[width=52.5mm]{figs/D_disk/Ialongpath.pdf}} \put(113,.5){\small iterations} \put(80,40){$I(z^\mathrm{max})$} \put(77.3,7.8){\scriptsize 5.9604} \put(77.3,15.7){\scriptsize 8.9302} \put(97,18.5){$u_2^\mathrm{rad}$} \put(123,10){$u_2$} \put(110,30){$p=1.3$} \put(76.5,45){(d)} \end{picture} \caption{(a)--(c) Intermediate point $e_\mathrm{M}$ of the initial path used in CMPA. (d) Maximum value of the Dirichlet functional $I$ along the path during the run of CMPA with $e_\mathrm{M}$ shown in (c) for $p=1.3$.} \label{fig:disk_em} \end{figure} The eigenvalues $\lambda_2^\mathrm{rad}$ corresponding to the found $u_2^\mathrm{rad}$ are also listed in Table~\ref{tab:disk_lambda}(a). Figure~\ref{fig:disk_u2rad}(b) shows profiles of the eigenfunction $u_2^\mathrm{rad}$ computed by the radial 2D method for several values of $p$. The eigenvalues $\lambda_1$ and $\lambda_2^\mathrm{rad}$ computed by this method for these and other values of $p$ are listed in Table~\ref{tab:disk_lambda}(b). The convergence of CMPA does not seem to be sensitive to the choice of $e_\mathrm{M}$ in this case. By comparing the values of $\lambda_1$ and $\lambda_2^\mathrm{rad}$ in Table~\ref{tab:disk_lambda}(a) with those in Table~\ref{tab:disk_lambda}(b) which were computed by the two different numerical methods we observe that their first three digits coincide in almost all the cases. Also, the profiles of $u_1$, $u_2^\mathrm{rad}$ are very close for both methods, respectively (cf.~Fig~\ref{fig:disk_u2rad}(a) and the top left graph in (b) for $u_2^\mathrm{rad}$ and $p=1.1$). We conclude that these are numerical approximations of the same eigenvalue-eigenfunction pairs. The behavior of CMPA suggests that although $u_2^\mathrm{rad}$ is a constrained mountain pass point of $I$ among radially symmetric functions, it is not a constrained mountain pass point with no assumption on the symmetry (cf.~Fig.~\ref{fig:disk_em}(d)). The cases of $p=1.1$ and $p=1.2$ when CMPA with $e_\mathrm{M}$ from Fig.~\ref{fig:disk_em}(b) converged to a radially symmetric function and the path did not slip off to asymmetric functions with lower values of $I$ seem to contradict this. However, we assume that this was caused by the ``flat'' shape of the landscape of $I$ close to $u_2^\mathrm{rad}$ for $p$ close to 1 and by numerical inaccuracies. \begin{figure}[t] \centering \setlength{\unitlength}{1mm} \begin{picture}(127,43)\small \put(-3,-1){\includegraphics[width=67mm]{figs/D_disk/lambda_plot1a.pdf}} \put(60.5,-1){\includegraphics[width=67mm]{figs/D_disk/lambda_plot1bdotted.pdf}} \put(3,40){$\lambda$} \put(60,3.5){$p$} \put(48.5,25){$\lambda_1$} \put(19,31){$\lambda_2$} \put(7.5,35.5){$\lambda_2^\mathrm{rad}$} \put(66.5,40){$\lambda$} \put(123.5,3.5){$p$} \put(117,13){$\lambda_1$} \put(110,23){$\lambda_2$} \put(102,33){$\lambda_2^\mathrm{rad}$} \end{picture} \caption{Dependence of the numerically computed eigenvalues for the disk on $p$. The three cross symbols in the graph on the right mark the values $h_1(\Omega)=2$, $h_2(\Omega)\approx 3.1543$, and $h_2^\mathrm{rad}(\Omega)=4$.} \label{fig:disk_lambda_plot1} \end{figure} \begin{figure}[t] \centering \setlength{\unitlength}{1mm} \begin{picture}(127,45)(-17,1)\small \put(0,0){\includegraphics[width=90mm]{figs/D_disk/lambda_plot2.pdf}} \put(6.5,42.5){$\lambda^{1/p}$} \put(84.5,4.5){$p$} \put(45,15.7){$(\lambda_1)^{1/p}$} \put(45,24.5){$(\lambda_2)^{1/p}$} \put(45,35){$(\lambda_2^\mathrm{rad})^{1/p}$} \put(-15,13){\makebox(25,4)[r]{$h_1(\Omega)=2$}} \put(-15,21.3){\makebox(25,4)[r]{$h_2(\Omega)\approx 3.1543$}} \put(-15,27){\makebox(25,4)[r]{$h_2^\mathrm{rad}(\Omega)=4$}} \put(83.5,8){$\Lambda_1=1$} \put(83.5,14.5){$\Lambda_2=2$} \put(83.5,21.5){$\Lambda_2^\mathrm{rad}=3$} \end{picture} \caption{Dependence of the numerically computed eigenvalues for the disk raised to $1/p$ on $p$.} \label{fig:disk_lambda_plot2} \end{figure} The dependence of $\lambda_1$, $\lambda_2$, and $\lambda_2^\mathrm{rad}$ on $p$ is presented in Figs.~\ref{fig:disk_lambda_plot1} and \ref{fig:disk_lambda_plot2}. First of all we observe that for all the values of $p$ considered the inequality $\lambda_2<\lambda_2^\mathrm{rad}$ holds. Hence this is a numerical evidence that the second eigenfunction for the disk is not radially symmetric not only for small $p$ but for a large range of $p$. \begin{samepage} Second, we can observe the following asymptotic behavior: \begin{center}\renewcommand{\arraystretch}{1.3}\vspace{1ex} \begin{tabular}{|c|c|c|c|} \cline{2-4} \multicolumn{1}{c|}{} & $\lambda_1$ & $\lambda_2$ & $\lambda_2^\mathrm{rad}$ \\\hline $\lim_{p\to1+}\lambda$ & 2 & 3.1543 & 4 \\\hline $\lim_{p\to\infty}\lambda^{1/p}$ & 1 & 2 & 3 \\\hline \end{tabular}\renewcommand{\arraystretch}{1}\vspace{1ex} \end{center} \end{samepage} Theoretical results for $\lambda_1$ and $\lambda_2$ were summarized in Sec.~\ref{sec:background}. The values $h_1(\Omega)$, $\Lambda_1(\Omega)$, and $\Lambda_2(\Omega)$ for the disk are easy to compute. In \cite{Parini} it was proved that $h_2(\Omega)$ for the disk equals the first Cheeger constant for the half-disk which is approximately 3.1543. We could observe that $u_1$ converges to 1 for $p\to 1$ as explained in \cite{KawohlFridman} and to the distance function to the boundary for $p\to\infty$ as explained in \cite{JuutinenLindqvistManfredi}. In Fig.~\ref{fig:disk_u2} we observe that for $p\to1$ the function $u_2$ is getting close to the indicator function of the Cheeger set for the half-disk on each nodal domain. In \cite{BenDrabGirg} a numerical evidence is given leading to the conjecture for the asymptotic behavior of $\lambda_2^\mathrm{rad}$ given in the above table. Our numerical results (at least for $p\to1$) support this conjecture. To motivate these values and the profiles of $u_2^\mathrm{rad}$ in Fig.~\ref{fig:disk_u2rad} we make the following two remarks: \begin{remark} \label{rmk3.1} \rm 1.~For $r\in(0,1)$ let $D(r)$ be the disk of radius $r$ centered at the origin, $A(r)=\Omega\setminus\overline{D(r)}$ an annulus. It is easy to show that for $r=1/2$ both $D$ and $A$ have the same Cheeger constant $h_2^\mathrm{rad}(\Omega):=h_1(D(1/2))=h_1(A(1/2))=4$ (see, e.g., \cite{KrejcirikPratelli} for a result about the Cheeger constant of an annulus). The function $u_2^\mathrm{rad}$ with its profile shown in Fig.~\ref{fig:disk_u2rad} seems to get close to the indicator function of $D(1/2)$ and $A(1/2)$ on each nodal domain for $p\to1$. 2.~Under the assumption of radial symmetry two largest disjoint disks of the same radius inscribed in $\Omega$ have radius 1/3. Hence we define $\Lambda_2^\mathrm{rad}=\frac{1}{1/3}=3$. The function $u_2^\mathrm{rad}$ with its profile shown in Fig.~\ref{fig:disk_u2rad} seems to get close on each nodal domain to a multiple of the function giving the distance to the boundary on $D(1/3)$ and $A(1/3)$ for large $p$. \end{remark} \subsection{Square} Let \[ \Omega:=\left\{(x_1,x_2)\in\mathbb{R}^2:x_1,x_2\in(0,2)\right\}. \] This domain was discretized using 83,968 triangles. Figure \ref{fig:square_u2} shows the eigenfunction $u_2$ for several values of $p$. Table \ref{tab:square_lambda} lists the computed values of $\lambda_1$ and $\lambda_2$. \begin{table}[h] \centering\vspace{1ex} \begin{tabular}[t]{|c|c|c|c|c|c|c|c|c|} \cline{1-4}\cline{6-9} $p$ & $\lambda_1$ & $\lambda_2$ & $\lambda_{\mathcal{S}_2}$ & \quad\quad & $p$ & $\lambda_1$ & $\lambda_2$ & $\lambda_{\mathcal{S}_1}$ \\\cline{1-4}\cline{6-9} \multicolumn{9}{c}{}\\\cline{1-4}\cline{6-9} 1.1 & 2.3649 & 3.7586 & 3.8702 & & 2.0 & 4.9349 & 12.338 & 12.338 \\\cline{1-4}\cline{6-9} 1.2 & 2.6934 & 4.5012 & 4.6179 & & 2.1 & 5.2139 & 13.684 & 13.744 \\\cline{1-4}\cline{6-9} 1.3 & 2.9986 & 5.2500 & 5.3715 & & 2.2 & 5.4952 & 15.144 & 15.282 \\\cline{1-4}\cline{6-9} 1.4 & 3.2834 & 6.0385 & 6.1621 & & 2.3 & 5.7791 & 16.725 & 16.961 \\\cline{1-4}\cline{6-9} 1.5 & 3.5611 & 6.8835 & 7.0053 & & 2.4 & 6.0658 & 18.438 & 18.797 \\\cline{1-4}\cline{6-9} 1.6 & 3.8356 & 7.7971 & 7.9118 & & 2.5 & 6.3552 & 20.293 & 20.802 \\\cline{1-4}\cline{6-9} 1.7 & 4.1092 & 8.7897 & 8.8903 & & 3.0 & 7.8452 & 32.107 & 33.956 \\\cline{1-4}\cline{6-9} 1.8 & 4.3830 & 9.8708 & 9.9490 & & 4.0 & 11.038 & 74.757 & 85.447 \\\cline{1-4}\cline{6-9} 1.9 & 4.6581 & 11.050 & 11.095 & & 5.0 & 14.497 & 163.59 & 205.08 \\\cline{1-4}\cline{6-9} 2.0 & 4.9349 & 12.338 & 12.338 & & 6.0 & 18.194 & 343.77 & 477.60 \\\cline{1-4}\cline{6-9} \multicolumn{4}{c}{} & & 8.0 & 26.221 & 1,402.1 & 2,443.4 \\\cline{6-9} \multicolumn{4}{c}{} & & 10.0 & 34.990 & 5,339.0 & 11,888 \\\cline{6-9} \end{tabular}\vspace{2ex} \caption{Eigenvalues for the square.} \label{tab:square_lambda} \end{table} Various choices of the intermediate path point $e_\mathrm{M}$ were used to compute $u_2$. Only in case $p=2$ different choices of $e_\mathrm{M}$ caused CMPA to converge to different functions $u_2$. Since for the square the eigenspace corresponding to the second eigenvalue of the Laplace operator is two-dimensional, CMPA converges to some member of this eigenspace depending on the shape of the initial path. Figure \ref{fig:square_u2} shows one such eigenfunction. However, even for $p=2$ this has no influence on the computed value of $\lambda_2$. \newcommand{\squtwographs}[2]{% \begin{picture}(25,52) \put(0,0){\makebox(25,4){$p=#1$}} \put(0,30){\includegraphics[width=25mm, clip=true, viewport= 150 90 473 390]{figs/D_square2/#2.jpg}} \put(1,6){\includegraphics[width=23mm]{figs/D_square2/#2.pdf}} \end{picture} } \begin{figure}[p] \centering \setlength{\unitlength}{1mm} \begin{picture}(150,52) \put(0,0){\squtwographs{1.1}{out_u2_p_1-1b}} \put(25.5,0){\squtwographs{1.4}{out_u2_p_1-4}} \put(51,0){\squtwographs{2.0}{out_u2_p_2-0d}} \put(76.5,0){\squtwographs{2.5}{out_u2_p_2-5a}} \put(102,0){\squtwographs{8.0}{out_u2_p_8-0a}} \end{picture} \caption{The numerically computed second eigenfunction $u_2$ for the square.} \label{fig:square_u2} \end{figure} \begin{figure}[p] \centering \setlength{\unitlength}{1mm} \begin{picture}(127,45)\small \put(-3,-1){\includegraphics[width=67mm]{figs/D_square2/lambda_plot1a.pdf}} \put(60.5,-1){\includegraphics[width=67mm]{figs/D_square2/lambda_plot1bdotted.pdf}} \put(3,40.5){$\lambda$} \put(60,3.5){$p$} \put(51,29){$\lambda_1$} \put(19,33.5){$\lambda_2$} \put(7,18.5){$\lambda_{\mathcal{S}_2}$} \put(8,17.5){\vector(0,-1){7}} \put(10,33.5){$\lambda_{\mathcal{S}_1}$} \put(12,28.5){\line(0,1){4}} \put(12,28.5){\vector(1,0){3}} \put(66.5,40.5){$\lambda$} \put(123.5,3.5){$p$} \put(117,12.5){$\lambda_1$} \put(107,27){$\lambda_2$} \put(100.5,33){$\lambda_{\mathcal{S}_2}$} \end{picture} \caption{Dependence of the numerically computed eigenvalues for the square on $p$. The three cross symbols in the graph on the right mark the values of $h_1$ for $\Omega$, $\Omega^\mathrm{half}_1$, and $\Omega^\mathrm{half}_2$.} \label{fig:square_lambda_plot1} \end{figure} \begin{figure}[p] \centering \setlength{\unitlength}{1mm} \begin{picture}(127,48)(-17,1)\small \put(0,0){\includegraphics[width=90mm]{figs/D_square2/lambda_plot2.pdf}} \put(6.5,42.5){$\lambda^{1/p}$} \put(85.5,4){$p$} \put(45,10){$(\lambda_1)^{1/p}$} \put(45,23){$(\lambda_2)^{1/p}$} \put(48,35){$(\lambda_{\mathcal{S}_1})^{1/p}$} \put(39.5,36){\vector(0,-1){4.5}} \put(39.5,36){\line(1,0){8}} \put(22,41){$(\lambda_{\mathcal{S}_2})^{1/p}$} \put(13.5,42){\vector(0,-1){2}} \put(13.5,42){\line(1,0){8}} \put(-15,16){\makebox(25,4)[r]{$h_1(\Omega)\approx 1.89$}} \put(-15,27){\makebox(25,4)[r]{$h_1(\Omega^\mathrm{half}_1)\approx 2.85$}} \put(-15,31){\makebox(25,4)[r]{$h_1(\Omega^\mathrm{half}_2)\approx 2.96$}} \put(83.5,7){$\Lambda_1(\Omega)=1$} \put(83.5,14.5){$\Lambda_1(\Omega^\mathrm{half}_2)\approx 1.71$} \put(83.5,19){$\Lambda_1(\Omega^\mathrm{half}_1)=2$} \end{picture} \caption{Dependence of the numerically computed eigenvalues for the square raised to $1/p$ on $p$.} \label{fig:square_lambda_plot2} \end{figure} We say that a function has symmetry $\mathcal{S}_1$ (odd symmetry about $x_1=1$ and even symmetry about $x_2=1$) if it belongs to \[ \mathcal{S}_1 := \{u:\Omega\to\mathbb{R}: u(x_1,x_2)=-u(2-x_1,x_2), u(x_1,x_2)=u(x_1,2-x_2)\}, \] and symmetry $\mathcal{S}_2$ (odd symmetry about $x_1=x_2$ and even symmetry about $x_1=2-x_2$) if it belongs to \[ \mathcal{S}_2 := \{u:\Omega\to\mathbb{R}: u(x_1,x_2)=-u(x_2,x_1),u(x_1,x_2)=u(2-x_2,2-x_1)\}. \] As it was observed in \cite{YaoZhou1}, $u_2$ changes its symmetry at $p=2$ from $\mathcal{S}_1$ for $p<2$ to $\mathcal{S}_2$ for $p>2$. Let $\lambda_{\mathcal{S}_i}$ denote the smallest eigenvalue with an eigenfunction belonging to $\mathcal{S}_i$ where $i\in\{1,2\}$. The values of $\lambda_{\mathcal{S}_i}$ can be computed using CDM on $\Omega$ with additional boundary conditions $u(1,x_2)=0$ for $x_2\in(0,2)$ or $u(x,x)=0$ for $x\in(0,2)$, respectively, or as the first eigenvalue on the half-domain $\Omega^\mathrm{half}_i$, where \begin{gather*} \Omega^\mathrm{half}_1 := \{(x_1,x_2)\in\mathbb{R}^2: x_1\in(0,1), x_2\in(0,2)\}, \\ \Omega^\mathrm{half}_2 := \{(x_1,x_2)\in\mathbb{R}^2: x_1\in(0,2), x_2\in(0,x_1)\}. \end{gather*} \begin{table}[t] \begin{center}\renewcommand{\arraystretch}{1.3} \begin{tabular}{|c|c|c|c|c|c|c|c|} \cline{2-3}\cline{6-8} \multicolumn{1}{l|}{(a)} & $\lambda_{\mathcal{S}_1}$ & $\lambda_{\mathcal{S}_2}$ & \multicolumn{1}{c}{\rule{2mm}{0pt}} & \multicolumn{1}{l|}{(b)} & $\lambda_1$ & $\lambda_{\mathcal{S}_1}$ & $\lambda_{\mathcal{S}_2}$\\\cline{1-3}\cline{5-8} $p<2$ & $=\lambda_2$ & $>\lambda_2$ && $\lim_{p\to1+}\lambda$ & $1+\frac{1}{2}\sqrt{\pi}$ & $\frac{4-\pi}{3-\sqrt{1+2\pi}}$ & $1+\frac{1}{2}(\sqrt{2}+\sqrt{2\pi})$ \\\cline{1-3}\cline{5-8} $2
\lambda_2$ & $=\lambda_2$ && $\lim_{p\to\infty}\lambda^{1/p}$ &
1 & 2 & $1+\frac{1}{2}\sqrt{2}$ \\\cline{1-3}\cline{5-8}
\end{tabular}\renewcommand{\arraystretch}{1}
\end{center}\vspace{2ex}
\caption{The smallest eigenvalues $\lambda_{\mathcal{S}_1}$ and $\lambda_{\mathcal{S}_2}$
under symmetry assumptions for the square.
(a) Numerical comparison with $\lambda_2$.
(b) Asymptotic behavior: the first row shows values of $h_1$, the second row values of $\Lambda_1$
for $\Omega$, $\Omega^\mathrm{half}_1$, and $\Omega^\mathrm{half}_2$, respectively.}
\label{tab:square_observations}
\end{table}
Our numerical observations regarding these eigenvalues and $\lambda_2$
are summarized in Table \ref{tab:square_observations}(a) and the
computed values are listed in Table \ref{tab:square_lambda}. We stress
that $\lambda_2$ was computed with no a priori assumptions on
symmetry. The dependence of the eigenvalues $\lambda_1$, $\lambda_2$,
$\lambda_{\mathcal{S}_1}$ and $\lambda_{\mathcal{S}_2}$ on $p$ is
further plotted in Figures \ref{fig:square_lambda_plot1} and
\ref{fig:square_lambda_plot2}.
These figures and Table \ref{tab:square_observations}(b) also explain
the asymptotic behavior as $p\to 1$ and $p\to\infty$. While the table
shows the limit values as given by the theory, the graphs indicate
convergence to these values (at least for $p\to 1$; for $p\to\infty$
it seems a larger range of $p$ would be needed). The Cheeger constants
$h_1$ shown in the first row of Table \ref{tab:square_observations}(b)
have been computed according to \cite{KawohlFridman},
\cite{LachandRobertKawohl} by
\begin{equation}
\label{eq:square_cheeger}
h_1\big((0,a)\times(0,b)\big)=\frac{4-\pi}{a+b-\sqrt{(a-b)^2+\pi ab}} \quad\text{for }a,b>0.
\end{equation}
The evaluation of $\Lambda_1$ in the second row is straightforward.
\subsection{Rectangle}\label{rectangle}
Let
\[
\Omega:=\left\{(x_1,x_2)\in\mathbb{R}^2:x_1\in(0,2), x_2\in(0,1.75)\right\}.
\]
This domain was discretized using 77,312 triangles. The shape of the
first eigenfunction $u_1$ and the graph of the first eigenvalue
$\lambda_1(\Omega;p)$ are similar to those for the square. However,
the symmetry properties of the second eigenfunction $u_2$ are
different: According to our numerical observations, for $p\leq 3.6$
the eigenfunction $u_2$ preserves an odd symmetry about $x_1=1$ and an
even symmetry about $x_2=0.875$ (which we call $\mathcal{S}_1$ as in
the case of the square). For $p\geq 3.7$ this symmetry is lost and
$u_2$ maintains an odd symmetry with respect to $(1,0.875)$, the
center of $\Omega$. The contour lines of $u_2$ for several values of
$p$ are shown in Fig.~\ref{fig:rect_u2}.
\newcommand{\reutwographs}[2]{%
\begin{picture}(30,30)
\put(0,0){\makebox(30,4){$p=#1$}}
\put(1,6){\scalebox{-1}[1]{\includegraphics[width=28mm]{figs/D_rectangle_2_1.75/#2.pdf}}}
\end{picture}
}
\begin{figure}[h]
\centering
\setlength{\unitlength}{1mm}
\begin{picture}(127,30)
\put(.5,0){\reutwographs{3.6}{out_u2_p_3-6}}
\put(32.5,0){\reutwographs{3.7}{out_u2_p_3-7}}
\put(64.5,0){\reutwographs{3.8}{out_u2_p_3-8}}
\put(96.5,0){\reutwographs{8.0}{out_u2_p_8-0}}
\end{picture}
\caption{The numerically computed second eigenfunction $u_2$ for the rectangle.}
\label{fig:rect_u2}
\end{figure}
\begin{figure}[h]
\centering
\setlength{\unitlength}{1mm}
\begin{picture}(140,42)\small
\put(-3,0){\includegraphics[width=7.5cm]{figs/D_rectangle_2_1.75/lambda_plot.pdf}}
\put(1.5,39){$\lambda^{1/p}$}
\put(49,10){$(\lambda_1)^{1/p}$}
\put(49,23.5){$(\lambda_2)^{1/p}$}
\put(49,31.5){$(\lambda_{\mathcal{S}_1})^{1/p}$}
\put(68,3.5){$p$}
\put(76,0){\normalsize
\begin{tabular}[b]{|c|c|c|c|}
\multicolumn{4}{c}{$\Omega=(0,2)\times(0,1.75)$} \\[-5pt]
\multicolumn{4}{c}{}\\\hline
$p$ & $\lambda_1$ & $\lambda_2$ & $\lambda_{\mathcal{S}_1}$ \\ \hline
\multicolumn{4}{c}{}\\\hline
3.6 & 12.680 & 63.436 & 63.436 \\\hline
3.7 & 13.206 & 69.491 & 69.492 \\\hline
3.8 & 13.743 & 76.059 & 76.083 \\\hline
3.9 & 14.292 & 83.175 & 83.256 \\\hline
4.0 & 14.853 & 90.881 & 91.059 \\\hline
8.0 & 49.531 & 2,192.9 & 2,574.6 \\\hline
\end{tabular}
}
\end{picture}
\caption{Comparison of the numerically computed second eigenvalue
$\lambda_2$ and the smallest eigenvalue $\lambda_{\mathcal{S}_1}$
under the symmetry $\mathcal{S}_1$ for the rectangle
$\Omega$ ($\lambda_1$ is shown for reference).}
\label{fig:rect_lambda2}
\end{figure}
\begin{table}[h]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\multicolumn{4}{c}{(a)\quad Rectangle $(0,2)\times(0,1.9)$} &
\multicolumn{1}{c}{\rule{2mm}{0pt}}
& \multicolumn{4}{c}{(b)\quad Rectangle $(0,2)\times(0,1.6)$}\\[-5pt]
\multicolumn{8}{c}{}\\
\cline{1-4}\cline{6-9}
$p$ & $\lambda_1$ & $\lambda_2$ & $\lambda_{\mathcal{S}_1}$ &
& $p$ & $\lambda_1$ & $\lambda_2$ & $\lambda_{\mathcal{S}_1}$ \\
\cline{1-4}\cline{6-9}
\multicolumn{9}{c}{}\\
\cline{1-4}\cline{6-9}
2.44 & 6.5926 & 20.0177 & 20.0177 && 5.6 & 36.077 & 383.4648 & 383.4648
\\ \cline{1-4}\cline{6-9}
2.46 & 6.6579 & 20.4281 & 20.4281 && 5.8 & 38.898 & 453.2332 & 453.2333
\\ \cline{1-4}\cline{6-9}
2.48 & 6.7234 & 20.8451 & 20.8457 && 6.0 & 41.892 & 535.2007 & 535.2009
\\ \cline{1-4}\cline{6-9}
2.50 & 6.7891 & 21.2683 & 21.2708 && 6.2 & 45.068 & 631.4438 & 631.4478
\\ \cline{1-4}\cline{6-9}
2.60 & 7.1205 & 23.4816 & 23.5124 && 6.4 & 48.437 & 744.1846 & 744.4026
\\ \cline{1-4}\cline{6-9}
\end{tabular}\vspace{2ex}
\caption{Comparison of the numerically computed second eigenvalue
$\lambda_2$ and the smallest eigenvalue $\lambda_{\mathcal{S}_1}$
under the symmetry $\mathcal{S}_1$ for other rectangles.}
\label{tab:rect_more_lambda2}
\end{table}
For $p\geq 3.7$ the smallest eigenvalue $\lambda_{\mathcal{S}_1}$
corresponding to an eigenfunction with symmetry $\mathcal{S}_1$ is
larger than $\lambda_2$ (cf.\ Fig.~\ref{fig:rect_lambda2}). This
eigenpair can be computed by CDM on $\Omega$ with an additional
boundary condition $u(1,x_2)=0$ for $x_2\in(0,1.75)$ or as the first
eigenpair on the half-rectangle
\[
\Omega^\mathrm{half}:=\left\{(x_1,x_2)\in\mathbb{R}^2:x_1\in(0,1),
x_2\in(0,1.75)\right\}.
\]
Our conjecture is that for a rectangle $R=(0,a)\times(0,b)$ with
$02$ such that $u_2$ has two nodal domains which for
$p 0$ is a constant which does not depend on $w_n$ and
$u_n$. These inequalities and \eqref{eq:convergence_to_zero} imply
\begin{equation}
\label{eq:wn_to_zero}
w_n\to 0 \quad \text{strongly in }W^{1,p}_0(\Omega)\text{ as }n\to\infty.
\end{equation}
This and \eqref{eq:seq_wn} in turn imply that $\{u_n\}$ converges
strongly to $u$ and that
\begin{equation}
\label{eq:u_critical_point}
u=\frac{1}{\|v\|^p}(-\Delta_p)^{-1}\left(|u|^{p-2}u\right),
\end{equation}
which means that $u$ is a critical point of $I$ with respect to $S$.
\end{proof}
\begin{remark} \label{rmk56} \rm
To better understand the implications of the choice of the descent
direction we remark how the proof of the proposition would change if
we used the steepest descent direction instead of the descent
direction given by \eqref{eq:descent_direction1}. Up to normalization
the steepest descent direction $w$ defined by
\eqref{eq:steepestdescentBanach} can be written as the solution of
\[
-\Delta_p w=\Delta_p u+\alpha|u|^{p-2}u
\]
for a suitable $\alpha$. Testing this equation by $w$ and using
$w\in T_u S$ yields
$\|w\|^p=-\frac{1}{p}\dualpairing{I'(u)}{w}$. Hence equation
\eqref{eq:convergence_Iprimewn} would directly imply $w_n\to 0$. We
can write
\[
0\leftarrow \|w_n\|^{p-1}=\|-\Delta_p w_n\|_\ast = \left\|-\Delta_p u_n -\alpha_n|u_n|^{p-2}u_n\right\|_\ast=
\frac{1}{p}\left\|I'(u_n)-\alpha_n J'(u_n)\right\|_\ast,
\]
where $\|\cdot\|_\ast$ denotes the norm in the dual space
$W^{-1,q}(\Omega)$. If we define
$\|I'|_{S_u}\|:=\inf_{\alpha\in\mathbb{R}}\|I'(u)-\alpha J'(u)\|_\ast$ as in
\cite{Bon}, \cite{Ho1}, then we would obtain $\|I'|_{S_{u_n}}\|\to
0$. The Palais-Smale condition under constraints which was
formulated in \cite{Bon} states in its simplified form that if
$\{I'(u_n)\}$ is bounded and $\|I'|_{S_{u_n}}\|\to 0$ then $\{u_n\}$
possesses a convergent subsequence. In
\cite{CuestadeFigueiredoGossez} it was shown that this condition
holds in our setting. Hence the choice of the steepest descent
direction would yield a more ``classical'' proof of the proposition.
\end{remark}
\begin{thebibliography}{00}
\bibitem{Anane}
A.~Anane, \emph{Simplicit\'e et isolation de la premi\`ere valeur propre du
{$p$}-laplacien avec poids}, C. R. Acad. Sci. Paris S\'er. I Math.
\textbf{305} (1987), no.~16, 725--728.
\bibitem{AnaneTsouli1}
A.~Anane and N.~Tsouli, \emph{On the second eigenvalue of the
{$p$}-{L}aplacian}, Nonlinear partial differential equations ({F}\`es, 1994),
Pitman Res. Notes Math. Ser., vol. 343, Longman, Harlow, 1996, pp.~1--9.
\bibitem{BarrettLiu}
J.~W. Barrett and W.~B. Liu, \emph{Finite element approximation of the
{$p$}-{L}aplacian}, Math. Comp. \textbf{61} (1993), no.~204, 523--537.
\bibitem{BelloniKawohl1}
M.~Belloni and B.~Kawohl, \emph{A direct uniqueness proof for equations
involving the {$p$}-{L}aplace operator}, Manuscripta Math. \textbf{109}
(2002), no.~2, 229--231.
\bibitem{BenDrabGirg}
J.~Benedikt, P.~Dr\'abek, and P.~Girg, \emph{The second eigenfunction of the
{$p$}-{L}aplacian on the disk is not radial}, Preprint, 2010.
\bibitem{Bon}
A.~Bonnet, \emph{A deformation lemma on a {$C\sp 1$} manifold}, Manuscripta
Math. \textbf{81} (1993), no.~3-4, 339--359.
\bibitem{BrownReichel}
B.~M. Brown and W.~Reichel, \emph{Computing eigenvalues and {F}u\v c\'\i
k-spectrum of the radially symmetric {$p$}-{L}aplacian}, J. Comput. Appl.
Math. \textbf{148} (2002), no.~1, 183--211.
\bibitem{ChMcK1}
Y.~S. Choi and P.~J. McKenna, \emph{A mountain pass method for the numerical
solution of semilinear elliptic problems}, Nonlinear Anal. \textbf{20}
(1993), no.~4, 417--437.
\bibitem{CuestadeFigueiredoGossez}
M.~Cuesta, D.~de~Figueiredo, and J.-P. Gossez, \emph{The beginning of the
{F}u\v cik spectrum for the {$p$}-{L}aplacian}, J. Differential Equations
\textbf{159} (1999), no.~1, 212--238.
\bibitem{DiBenedetto}
E.~DiBenedetto, \emph{{$C^{1+\alpha }$} local regularity of weak solutions of
degenerate elliptic equations}, Nonlinear Anal. \textbf{7} (1983), no.~8,
827--850.
\bibitem{DrabekRobinson}
P.~Dr{\'a}bek and S.~B. Robinson, \emph{On the generalization of the {C}ourant
nodal domain theorem}, J. Differential Equations \textbf{181} (2002), no.~1,
58--71.
\bibitem{GarciaPeral}
J.~P. Garc{\'{\i}}a~Azorero and I.~Peral~Alonso, \emph{Existence and
nonuniqueness for the {$p$}-{L}aplacian: nonlinear eigenvalues}, Comm.
Partial Differential Equations \textbf{12} (1987), no.~12, 1389--1430.
\bibitem{GlowinskiMarroco}
R.~Glowinski and A.~Marroco, \emph{Sur l'approximation, par \'el\'ements finis
d'ordre un, et la r\'esolution, par p\'enalisation-dualit\'e, d'une classe de
probl\`emes de {D}irichlet non lin\'eaires}, RAIRO Analyse Num\'erique
\textbf{9} (1975), no.~R-2, 41--76.
\bibitem{Ho1}
J.~Hor{\'a}k, \emph{Constrained mountain pass algorithm for the numerical
solution of semilinear elliptic problems}, Numer. Math. \textbf{98} (2004),
no.~2, 251--276.
\bibitem{JuutinenLindqvist}
P.~Juutinen and P.~Lindqvist, \emph{On the higher eigenvalues for the
{$\infty$}-eigenvalue problem}, Calc. Var. Partial Differential Equations
\textbf{23} (2005), no.~2, 169--192.
\bibitem{JuutinenLindqvistManfredi}
P.~Juutinen, P.~Lindqvist, and J.~J. Manfredi, \emph{The {$\infty$}-eigenvalue
problem}, Arch. Ration. Mech. Anal. \textbf{148} (1999), no.~2, 89--105.
\bibitem{KawohlFridman}
B.~Kawohl and V.~Fridman, \emph{Isoperimetric estimates for the first
eigenvalue of the {$p$}-{L}aplace operator and the {C}heeger constant},
Comment. Math. Univ. Carolin. \textbf{44} (2003), no.~4, 659--667.
\bibitem{LachandRobertKawohl}
B.~Kawohl and T.~Lachand-Robert, \emph{Characterization of {C}heeger sets for
convex subsets of the plane}, Pacific J. Math. \textbf{225} (2006), no.~1,
103--118.
\bibitem{KrejcirikPratelli}
D.~Krej\v{c}i\v{r}\'{\i}k and A.~Pratelli, \emph{The {C}heeger constant of
curved strips}, arXiv:1011.3490v1, 2010.
\bibitem{Lindqvist1}
P.~Lindqvist, \emph{On the equation {${\rm div}\,(\vert \nabla u\vert
^{p-2}\nabla u)+\lambda\vert u\vert ^{p-2}u=0$}}, Proc. Amer. Math. Soc.
\textbf{109} (1990), no.~1, 157--164.
\bibitem{Lindqvist1ad}
P.~Lindqvist,\emph{Addendum: ``{O}n the equation {${\rm div}(\vert \nabla u\vert
^{p-2}\nabla u)+\lambda\vert u\vert ^{p-2}u=0$}''}, Proc. Amer. Math. Soc.
\textbf{116} (1992), no.~2, 583--584.
\bibitem{Parini}
E.~Parini, \emph{The second eigenvalue of the {$p$}-{L}aplacian as {$p$} goes
to 1}, Int. J. Differ. Equ. (2010), Art. ID 984671, 23.
\bibitem{Struwe}
M.~Struwe, \emph{Variational methods}, Springer-Verlag, Berlin, 1990.
\bibitem{YaoZhou1}
X.~Yao and J.~Zhou, \emph{Numerical methods for computing nonlinear eigenpairs.
{I}. {I}so-homogeneous cases}, SIAM J. Sci. Comput. \textbf{29} (2007),
no.~4, 1355--1374.
\end{thebibliography}
\end{document}