betaFeedback

WorksMathematics

On the Flucher–Rumpf isoperimetric question for the interior Bernoulli free boundary problem: a conformal reformulation, quantitative bounds and the second variation at the disc

Abstract

Let c(Ω)c(Ω) be the Bernoulli constant of a bounded domain Ω⊂RNΩ\subset\R^N, that is, the infimum of the parameters λλ for which the interior Bernoulli free boundary problem Δu=0Δu=0 in Ω∖KΩ\setminus K, u=0u=0 on ∂Ω\partialΩ, u=1u=1 and ∣∇u∣=λ|\nabla u|=λ on ∂K\partial K admits a classical solution. Flucher and Rumpf asked whether c(Ω)≥c(Ω∗)c(Ω)\ge c(Ω^*), where Ω∗Ω^* is the ball of the same volume. The variational analogue was settled by Daners and Kawohl, but the variational constant differs from cc already for the ball, and the original question remains open. We prove the following. (i) Through a conformal parametrisation of Ω∖KΩ\setminus K by an annulus, every classical solution in a simply connected planar domain with connected, filled free set satisfies λ≥μ/Rλ\geμ/R, where πR2=∣Ω∣πR^2=|Ω| and μ≈2.485268μ\approx 2.485268; hence c(Ω)≥0.9142 c(Ω∗)c(Ω)\ge 0.9142\,c(Ω^*) for every bounded planar convex domain of class C1C^1. (ii) In this class, the Flucher–Rumpf inequality is equivalent to an area inequality for analytic functions in an annulus, which also provides an explicit generator of exact solutions. (iii) A Rellich–Pohozaev identity yields, in every dimension N≥2N\ge2 and for every strictly star-shaped domain, c(Ω)≥max⁡{θ(Ω),κN} c(Ω∗)c(Ω)\ge\max\{\sqrt{θ(Ω)},κ_N\}\,c(Ω^*), where θ(Ω)∈(0,1]θ(Ω)\in(0,1] is an explicit geometric quantity equal to 11 only for balls and κNκ_N is an explicit constant with κ3≈0.650κ_3\approx0.650 and κN→1κ_N\to1 as N→∞N\to\infty.

Introduction

The interior Bernoulli problem

Let Ω⊂RN\Omega\subset\mathbb{R}^N, N≥2N \ge2, be a bounded domain and λ>0\lambda> 0. The interior Bernoulli problem consists in finding a compact set K⊂ΩK \subset\Omega and a function uu such that

Δu=0 in Ω∖K,u=0 on ∂Ω,u=1 and ∣∇u∣=λ on ∂K.(1)\Delta u = 0 \text{ in } \Omega\setminus K,\qquad u = 0 \text{ on } \partial\Omega,\qquad u = 1 \text{ and } |\nabla u| = \lambda\text{ on } \partial K. \tag*{(1)}

A classical solution is a pair (K,u)(K,u) such that ∂K\partial K is of class C1,αC^{1,\alpha}, uu is continuous on Ω‾∖K∘\overline{\Omega} \setminus K^\circ, u∈C1(N∖K∘)u \in C^1(\mathcal{N} \setminus K^\circ) for a neighbourhood N\mathcal{N} of ∂K\partial K, and (1.1) holds pointwise, the gradient on ∂K\partial K being the limit from Ω∖K\Omega\setminus K. Problems of this type arise in ideal fluid flow, electrochemical machining and shape optimisation; see Flucher and Rumpf [11] and the survey [21].

For K⊂ΩK \subset\Omega compact we denote by

cap⁡(K;Ω)=inf⁡{∫Ω∣∇v∣2 dx:v∈W01,2(Ω), v≥1 a.e. on K}\operatorname{cap}(K;\Omega) = \inf\left\{ \int_{\Omega} |\nabla v|^2\,dx : v \in W_0^{1,2}(\Omega),\ v \ge1\ \text{a.e. on } K \right\}

the condenser capacity of KK relative to Ω\Omega, and by uKu_K the capacitary potential, i.e. the minimiser. If (K,u)(K,u) is a classical solution, then u=uKu=u_K on Ω∖K\Omega\setminus K and, by Green’s formula,

cap⁡(K;Ω)=∫∂K∣∇u∣ dσ=λPer⁡(K).\operatorname{cap}(K;\Omega)=\int_{\partial K}|\nabla u|\,d\sigma=\lambda\operatorname{Per}(K).

For convex Ω\Omega there is a threshold: by the work of Beurling [3], Acker [1], Henrot and Shahgholian [13, 14], and Cardaliaguet and Tahraoui [7], there exists c(Ω)>0c(\Omega)>0 such that (1) has no solution for λ<c(Ω)\lambda<c(\Omega) and admits a classical solution with KK convex for every λ≥c(Ω)\lambda\ge c(\Omega); moreover, the solution at λ=c(Ω)\lambda=c(\Omega) is unique [7]. In general we set

c(Ω)=inf⁡{λ>0:(1.1) admits a classical solution},(2)c(\Omega)=\inf\{\lambda>0:(1.1)\text{ admits a classical solution}\}, \tag*{(2)}

the Bernoulli constant of Ω\Omega (with c(Ω)=+∞c(\Omega)=+\infty if no classical solution exists). Since uu is invariant and ∣∇u∣|\nabla u| scales like t−1t^{-1} under x↦txx\mapsto tx, one has c(tΩ)=c(Ω)/tc(t\Omega)=c(\Omega)/t, and cc is invariant under rigid motions.

The question of Flucher and Rumpf

In [11 p. 202] Flucher and Rumpf asked whether

c(Ω)≥c(Ω∗),(3)c(\Omega)\ge c(\Omega^*), \tag*{(3)}

Ω∗\Omega^* being the ball with ∣Ω∗∣=∣Ω∣|\Omega^*|=|\Omega|. For the ball BR⊂RNB_R\subset\mathbb{R}^N the radial solutions are explicit. With K=BsK=B_s one has u(r)=log⁡(R/r)/log⁡(R/s)u(r)=\log(R/r)/\log(R/s) if N=2N=2 and u(r)=(r2−N−R2−N)/(s2−N−R2−N)u(r)=(r^{2-N}-R^{2-N})/(s^{2-N}-R^{2-N}) if N≥3N\ge3, so that on ∂Bs\partial B_s

λrad(s)=1slog⁡(R/s)(N=2),λrad(s)=N−2s(1−(s/R)N−2)(N≥3).(4)\lambda_{\mathrm{rad}}(s)=\frac{1}{s\log(R/s)}\quad(N=2),\qquad\lambda_{\mathrm{rad}}(s)=\frac{N-2}{s\left(1-(s/R)^{N-2}\right)}\quad(N\ge3). \tag*{(4)}

Minimising in s∈(0,R)s\in(0,R) gives

c(BR)=eR(N=2, s=R/e),c(BR)=(N−1)N−1N−2R(N≥3, s=R(N−1)−1N−2),(5)c(B_R)=\frac{e}{R}\quad(N=2,\ s=R/e),\qquad c(B_R)=\frac{(N-1)^{\frac{N-1}{N-2}}}{R}\quad(N\ge3,\ s=R(N-1)^{-\frac{1}{N-2}}), \tag*{(5)}

in agreement with [4] (case p=2p=2). For λ>c(BR)\lambda>c(B_R) two radial solutions coexist (the elliptic and the hyperbolic one) and they merge at λ=c(BR)\lambda=c(B_R): the Bernoulli constant is a turning point (fold) of the solution branch.

Several partial answers are known. Ly and Seck [15] proved, for the pp-Laplacian and star-shaped domains, the existence of classical solutions for all parameters above an explicit threshold, which yields an explicit upper bound for the Bernoulli constant. Cardaliaguet and Tahraoui [7] gave an upper bound in terms of the harmonic radius and the lower bound c(Ω)≥c(BR‾)c(\Omega)\ge c(B_{\overline{R}}), where BR‾B_{\overline{R}} is the smallest ball containing Ω\Omega; this lower bound degenerates for elongated domains. Bianchini and Salani [4] established a Brunn–Minkowski inequality for cc in the convex class, from which an Urysohn-type inequality follows: among convex sets of given mean width, the ball minimises cc. As they observe, this does not imply (3). Daners and Kawohl [9] considered the minimal parameter Λ(Ω)\Lambda(\Omega) for which the Alt–Caffarelli functional

Jλ(v)=∫Ω∣∇v∣2 dx+λ2∣{v<1}∣,v∈W01,2(Ω),\mathcal{J}_{\lambda}(v)=\int_{\Omega}|\nabla v|^2\,dx+\lambda^2|\{v<1\}|,\qquad v\in W_0^{1,2}(\Omega),

has a nontrivial minimiser, and proved by Schwarz symmetrisation that Λ(Ω)≥Λ(Ω∗)\Lambda(\Omega)\ge\Lambda(\Omega^*), with equality essentially only for balls. However, Λ\Lambda and cc are different: for the disc Λ(BR)=2e/R≈3.30/R\Lambda(B_R)=2\sqrt{e}/R\approx3.30/R, whereas c(BR)=e/R≈2.72/Rc(B_R)=e/R\approx2.72/R; for c(BR)≤λ<Λ(BR)c(B_R)\le\lambda<\Lambda(B_R) solutions exist, but none of them has lower energy than v≡0v\equiv0. Inequality (3) itself remains open, even for convex planar domains; see also [4, 7].

Main results

In the plane we use conformal mapping. Throughout Sections 2–3, Ω⊂C\Omega\subset\mathbb{C} is a bounded simply connected domain, and we only consider solutions whose free set KK is connected and filled (C∖K\mathbb{C} \setminus K connected); we call them admissible solutions. For convex Ω\Omega of class C1C^1, the solution at λ=c(Ω)\lambda= c(\Omega) is admissible by [14].

Theorem A (quantitative isoperimetric bound in the plane). Let Ω⊂C\Omega\subset\mathbb{C} be bounded and simply connected with ∣Ω∣=πR2|\Omega| = \pi R^2, and let (K,u)(K,u) be an admissible classical solution of (1). Then

λ≥μR,μ=min⁡0<ρ<11−ρ2ρlog⁡(1/ρ)=1ρ∗1−ρ∗2≈2.485268,\lambda\ge\frac{\mu}{R}, \qquad\mu= \min_{0<\rho<1} \frac{\sqrt{1-\rho^2}}{\rho\log(1/\rho)} = \frac{1}{\rho_*\sqrt{1-\rho_*^2}} \approx2.485268,

where ρ∗≈0.450764\rho_* \approx0.450764 is the unique root in (0,1)(0,1) of log⁡(1/ρ)=1−ρ2\log(1/\rho) = 1-\rho^2. In particular, for every bounded convex domain Ω⊂R2\Omega\subset\mathbb{R}^2 of class C1C^1,

c(Ω)≥μec(Ω∗)≈0.914279c(Ω∗).c(\Omega) \ge\frac{\mu}{e}c(\Omega^*) \approx0.914279c(\Omega^*).

To state the converse we introduce, for 0<ρ<10 < \rho< 1, the annulus Aρ={ρ<∣w∣<1}A_\rho= \{\rho< |w| < 1\} and the following class.

Definition 1.1. Let A\mathcal{A} be the set of triples (ρ,m,φ)(\rho,m,\varphi) with ρ∈(0,1)\rho\in(0,1), m>0m > 0 and φ\varphi analytic in AρA_\rho, continuous on Aρ∪{∣w∣=ρ}A_\rho\cup\{|w|=\rho\}, with Re⁡φ=0\operatorname{Re}\varphi= 0 on {∣w∣=ρ}\{|w|=\rho\}, such that meφm e^\varphi has a primitive FF on AρA_\rho (equivalently, Res⁡w=0eφ=0\operatorname{Res}_{w=0} e^\varphi= 0 in the Laurent sense) which is bounded and extends to an injective continuous map on {ρ≤∣w∣<1}\{\rho\le|w| < 1\}. We write KFK_F for the compact set bounded by the Jordan curve F({∣w∣=ρ})F(\{|w|=\rho\}) and ΩF=F(Aρ)∪KF\Omega_F = F(A_\rho) \cup K_F.

Proposition B (equivalent extremal problem). The Flucher–Rumpf inequality λ≥e/R\lambda\ge e/R holds for all admissible solutions in all bounded simply connected planar domains if and only if, for every (ρ,m,φ)∈A(\rho,m,\varphi) \in\mathcal{A},

∣ΩF∣≥πm2(eρlog⁡(1/ρ))2.(6)|\Omega_F| \ge\pi m^2\left(e\rho\log(1/\rho)\right)^2. \tag*{(6)}

Moreover, every element of A\mathcal{A} produces an explicit admissible solution in ΩF\Omega_F.

Proposition B is a reformulation rather than a new inequality: all the difficulty of (3) is transferred to the univalence constraint in Definition 1.1, which is global and non-convex. Its interest lies in the converse part, which generates exact solutions on which (3) can be tested (§ 6).

Proposition C (second variation at the disc; formal). Let k≥1k \ge1 and Ωε={r<(1+εcos⁡kθ)/1+ε2/2}\Omega_\varepsilon= \{r < (1+\varepsilon\cos k\theta)/\sqrt{1+\varepsilon^2/2}\}, so that ∣Ωε∣=π|\Omega_\varepsilon| = \pi. At the level of a formal expansion (see Remark 3.1),

c(Ωε)=e+Akε2+O(ε4),Ak=e(k2−1)[(k−1)−2(k2−1)xk−(k+1)xk2]2[(k−1)+(k+1)xk]2,c(\Omega_\varepsilon) = e + A_k\varepsilon^2 + O(\varepsilon^4), \qquad A_k = \frac{e(k^2-1)[(k-1)-2(k^2-1)x_k-(k+1)x_k^2]}{2[(k-1)+(k+1)x_k]^2},

where xk=e−2kx_k=e^{-2k}. One has A1=0A_1=0 (translations), Ak>0A_k>0 for all k≥2k\ge2, and Ak∼e(k+1)/2A_k\sim e(k+1)/2 as k→∞k\to\infty.

The expansion is even in ε\varepsilon, because the rotation θ↦θ+π/k\theta\mapsto\theta+\pi/k maps Ωε\Omega_\varepsilon onto Ω−ε\Omega_{-\varepsilon}. Proposition C is derived by two independent computations (the conformal parametrisation and a domain expansion in the physical plane), which agree identically and match the numerical values to four digits.

Our last main result is not restricted to the plane. For a bounded domain Ω⊂RN\Omega\subset\mathbb{R}^N of class C1,αC^{1,\alpha} which is strictly star-shaped with respect to the origin, i.e. x⋅ν(x)>0x\cdot\nu(x)>0 on ∂Ω\partial\Omega, set

I(Ω)=∫∂Ωdσx⋅ν,θ(Ω)=I(Ω∗)I(Ω),I(\Omega)=\int_{\partial\Omega}\frac{d\sigma}{x\cdot\nu}, \qquad\theta(\Omega)=\frac{I(\Omega^*)}{I(\Omega)},

where Ω∗\Omega^* is the ball centred at the origin with ∣Ω∗∣=∣Ω∣|\Omega^*| = |\Omega|. We show in § 4 that θ(Ω)∈(0,1]\theta(\Omega) \in(0,1], with θ(Ω)=1\theta(\Omega)=1 if and only if Ω\Omega is a ball centred at the origin. Define also κ2=0\kappa_2=0 and, for N≥3N\geq3,

κN=N−2NN(N/2)1N−2(N−1)N−1N−2.(7)\kappa_N=\sqrt{\frac{N-2}{N}}\frac{N(N/2)^{\frac{1}{N-2}}}{(N-1)^{\frac{N-1}{N-2}}}. \tag*{(7)}

Theorem D (Rellich–Pohozaev bounds in every dimension). Let N≥2N\geq2, let Ω⊂RN\Omega\subset\mathbb{R}^N be a bounded domain of class C1,αC^{1,\alpha}, strictly star-shaped with respect to the origin, and let (K,u)(K,u) be a classical solution of (1.1) with parameter λ\lambda (no assumption is made on the topology of KK). Then

λ2N∣K∣=∫∂Ω∣∇u∣2(x⋅ν) dσ+(N−2)cap⁡(K;Ω),(8)\lambda^2N|K|=\int_{\partial\Omega}|\nabla u|^2(x\cdot\nu)\,\mathrm{d}\sigma+(N-2)\operatorname{cap}(K;\Omega), \tag*{(8)}

and

λ≥max⁡{θ(Ω),κN} c(Ω∗).\lambda\geq\max\{\sqrt{\theta(\Omega)},\kappa_N\}\,c(\Omega^*).

Consequently c(Ω)≥max⁡{θ(Ω),κN} c(Ω∗)c(\Omega)\geq\max\{\sqrt{\theta(\Omega)},\kappa_N\}\,c(\Omega^*).

NN3451020100
κN\kappa_N0.64950.76980.82780.92340.96360.9930

Table 1. The constant κN\kappa_N of Theorem D; κN→1\kappa_N\to1 as N→∞N\to\infty.

For N≥3N\geq3, Theorem D gives a uniform bound of the type of Theorem A for all strictly star-shaped domains, convex or not. For N=2N=2 the coefficient N−2N-2 in (1.8) vanishes; this is why a uniform constant in the plane requires a different tool, here the conformal parametrisation. In the plane, the two results are complementary.

Corollary E. Let Ω⊂R2\Omega\subset\mathbb{R}^2 be a bounded convex domain of class C1,αC^{1,\alpha} and x0∈Ωx_0\in\Omega. Then

c(Ω)≥max⁡{μe,θx0(Ω)}c(Ω∗),c(\Omega)\geq\max\left\{\frac{\mu}{e},\sqrt{\theta_{x_0}(\Omega)}\right\}c(\Omega^*),

where θx0\theta_{x_0} is computed with respect to x0x_0. For an ellipse of aspect ratio qq, centred at x0x_0, θ=2/(q+q−1)\theta=2/(q+q^{-1}), and the second bound is the better one for q<1.8529q<1.8529.

Finally, § 5 explains why the natural symmetrisation route through polarization fails at the level of subsolutions, while it does work at the level of energies.

Organisation. § 2 contains the conformal dictionary and the proofs of Theorem A and Proposition B. § 3 is devoted to Proposition C. § 4 proves Theorem D and Corollary E. § 5 deals with polarization. § 6 reports the numerical experiments, and § 7 lists open problems.

Organisation

The conformal dictionary and the proofs of Theorem A and Proposition B

From solutions to analytic functions

2.1. From solutions to analytic functions. Let (K,u)(K,u) be an admissible classical solution in a bounded simply connected domain Ω⊂C\Omega\subset\mathbb{C}. Since KK is connected and filled and ∂K\partial K is a C1,αC^{1,\alpha} curve, KK has nonempty interior and ∂K\partial K is a single C1,αC^{1,\alpha} Jordan curve. As ∂Ω\partial\Omega is a nondegenerate continuum, Ω∖K\Omega\setminus K is a doubly connected domain with two nondegenerate boundary components. Hence there exist a unique ρ∈(0,1)\rho\in(0,1) and a conformal map

F:Aρ={ρ<∣w∣<1}⟶Ω∖KF:A_\rho=\{\rho<|w|<1\}\longrightarrow\Omega\setminus K

such that F(w)→∂ΩF(w) \to\partial\Omega as ∣w∣→1|w| \to1 and F(w)→∂KF(w) \to\partial K as ∣w∣→ρ|w| \to\rho.

Lemma 2.1 (boundary regularity at the inner circle). FF extends to an injective map of class C1,αC^{1,\alpha} on {ρ≤∣w∣<1}\{\rho\le|w| < 1\}, and F′F' does not vanish on {∣w∣=ρ}\{|w| = \rho\}.

Proof. The Kellogg–Warschawski theorem [19 Thm. 3.6] concerns simply connected domains, so we reduce to that case. Fix z0∈K∘z_0 \in K^\circ and let T(z)=1/(z−z0)T(z) = 1/(z-z_0). Then W=T(C^∖K)W = T(\widehat{\mathbb{C}} \setminus K) is a bounded simply connected domain containing 0=T(∞)0 = T(\infty), whose boundary T(∂K)T(\partial K) is a C1,αC^{1,\alpha} Jordan curve. Let g:D→Wg : \mathbb{D} \to W be a Riemann map. By the Kellogg–Warschawski theorem, g∈C1,α(D‾)g \in C^{1,\alpha}(\overline{\mathbb{D}}) and g′≠0g' \ne0 on D‾\overline{\mathbb{D}}. Moreover T(Ω∖K)=W∖ET(\Omega\setminus K) = W \setminus E, where E=T(C^∖Ω)E = T(\widehat{\mathbb{C}} \setminus\Omega) is a compact connected subset of WW.

The map h=g−1∘T∘Fh = g^{-1} \circ T \circ F is a conformal map of AρA_\rho onto the doubly connected domain D∖g−1(E)\mathbb{D} \setminus g^{-1}(E), and ∣h(w)∣→1|h(w)| \to1 as ∣w∣→ρ|w| \to\rho. By the Schwarz reflection principle, hh extends analytically to a neighbourhood of {∣w∣=ρ}\{|w| = \rho\} by h(w)=1/h(ρ2/w‾)‾h(w) = 1/\overline{h(\rho^2/\overline{w})}. This extension is injective near the circle: the two half-collars are mapped into D\mathbb{D} and into C^∖D‾\widehat{\mathbb{C}} \setminus\overline{\mathbb{D}} respectively, and the boundary map is a homeomorphism of circles. Hence h′≠0h' \ne0 on {∣w∣=ρ}\{|w| = \rho\}.

Near the circle, T∘F=g∘hT \circ F = g \circ h stays away from 00, so F=z0+1/(g∘h)F = z_0 + 1/(g \circ h) is of class C1,αC^{1,\alpha} up to {∣w∣=ρ}\{|w| = \rho\}. Finally, F′=−(g′∘h)h′/(g∘h)2≠0F' = -(g' \circ h)h'/(g \circ h)^2 \ne0 there. □

The function w↦log⁡∣w∣/log⁡ρw \mapsto\log|w|/\log\rho is harmonic in AρA_\rho, equal to 00 on ∣w∣=1|w| = 1 and to 11 on ∣w∣=ρ|w| = \rho. The function u∘Fu \circ F is bounded and harmonic in AρA_\rho, with boundary limits 00 as ∣w∣→1|w| \to1 (since uu is continuous up to ∂Ω\partial\Omega) and 11 as ∣w∣→ρ|w| \to\rho. By uniqueness for the Dirichlet problem,

u∘F(w)=log⁡∣w∣log⁡ρ,w∈Aρ.(9)u \circ F(w) = \frac{\log|w|}{\log\rho}, \qquad w \in A_\rho. \tag*{(9)}

Set L=log⁡(1/ρ)L = \log(1/\rho). By conformal invariance, ∣∇u(F(w))∣∣F′(w)∣=∣∇(u∘F)(w)∣=1/(∣w∣L)|\nabla u(F(w))| |F'(w)| = |\nabla(u \circ F)(w)| = 1/(|w|L). Letting ∣w∣→ρ|w| \to\rho and using Lemma 2.1, we obtain on ∂K\partial K

∣∇u(F(w))∣=1ρL∣F′(w)∣,∣w∣=ρ.(10)|\nabla u(F(w))| = \frac{1}{\rho L |F'(w)|}, \qquad|w| = \rho. \tag*{(10)}

The Bernoulli condition ∣∇u∣=λ|\nabla u| = \lambda is therefore equivalent to

∣F′(w)∣=m on ∣w∣=ρ,λ=1mρL.(11)|F'(w)| = m \text{ on } |w| = \rho, \qquad\lambda= \frac{1}{m\rho L}. \tag*{(11)}

Lemma 2.2. Let h=log⁡∣F′∣h = \log|F'|, which is harmonic in AρA_\rho. Then for every r∈(ρ,1)r \in(\rho,1)

12π∫02πh(reiθ) dθ=log⁡m.\frac{1}{2\pi}\int_{0}^{2\pi} h(re^{i\theta})\,d\theta= \log m.

Proof. For a harmonic function in an annulus, the circular mean M(r)M(r) satisfies (rM′(r))′=0(rM'(r))' = 0, so M(r)=a+βlog⁡rM(r) = a + \beta\log r with

β=rM′(r)=12π∫02πr ∂rh(reiθ) dθ=12π∮∣w∣=rdarg⁡F′(w),\beta= rM'(r) = \frac{1}{2\pi}\int_{0}^{2\pi} r\,\partial_r h(re^{i\theta})\,d\theta= \frac{1}{2\pi}\oint_{|w|=r} d\arg F'(w),

by the Cauchy–Riemann equations (r∂rlog⁡∣f∣=∂θarg⁡f)(r\partial_r \log|f| = \partial_\theta\arg f). Thus β\beta is the winding number of F′F' along ∣w∣=r|w| = r.

The curve γr=F({∣w∣=r})\gamma_r = F(\{|w| = r\}) is an analytic Jordan curve which separates KK from ∂Ω\partial\Omega. Since FF preserves orientation and maps the bounded region {ρ<∣w∣<r}\{\rho< |w| < r\} onto the region between ∂K\partial K and γr\gamma_r, the curve γr\gamma_r is positively oriented. Its tangent vector is ∂θF(reiθ)=ireiθF′(reiθ)\partial_\theta F(re^{i\theta}) = ire^{i\theta}F'(re^{i\theta}). By the Umlaufsatz, the argument of the tangent increases by 2π2\pi along γr\gamma_r. The argument of ireiθire^{i\theta} also increases by 2π2\pi, hence the winding number of F′F' is 00.

Thus MM is constant on (ρ,1)(\rho,1). By Lemma 2.1, hh is uniformly continuous on {ρ≤∣w∣≤r1}\{\rho\le|w| \le r_1\} for every r1<1r_1<1, so M(r)→12π∫02πh(ρeiθ) dθ=log⁡mM(r) \to\frac{1}{2\pi}\int_0^{2\pi}h(\rho e^{i\theta})\,d\theta=\log m as r→ρr\to\rho.

Lemma 2.2 shows that log⁡F′\log F' has a single-valued branch in AρA_\rho. Writing F′=meφF'=me^\varphi, the function φ\varphi is analytic in AρA_\rho, continuous up to ∣w∣=ρ|w|=\rho, and Re⁡φ=0\operatorname{Re}\varphi=0 there. In Laurent series this means

φ(w)=iγ+∑n≥1(cnwn−cn‾ρ2nw−n),γ∈R,(12)\varphi(w)=i\gamma+\sum_{n\ge1}\left(c_nw^n-\overline{c_n}\rho^{2n}w^{-n}\right),\qquad\gamma\in\mathbb{R}, \tag*{(12)}

since on ∣w∣=ρ|w|=\rho one has w−n=w‾ nρ−2nw^{-n}=\overline{w}^{\,n}\rho^{-2n} and each bracket equals 2iIm⁡(cnwn)2i\operatorname{Im}(c_nw^n). Single-valuedness of the primitive FF is the condition Res⁡w=0eφ=0\operatorname{Res}_{w=0}e^\varphi=0. Since FF is bounded and, by Lemma 2.1, injective on {ρ≤∣w∣<1}\{\rho\le|w|<1\}, we have (ρ,m,φ)∈A(\rho,m,\varphi)\in\mathcal{A}, K=KFK=K_F and Ω=ΩF\Omega=\Omega_F.

From analytic functions to solutions

Conversely, let (ρ,m,φ)∈A(\rho,m,\varphi)\in\mathcal{A} with primitive FF. Since Re⁡φ=0\operatorname{Re}\varphi=0 on ∣w∣=ρ|w|=\rho, the Schwarz reflection φ(w)=−φ(ρ2/w‾)\varphi(w)=-\varphi(\rho^2/\overline{w}) extends φ\varphi analytically across {∣w∣=ρ}\{|w|=\rho\}. Hence FF is analytic across that circle with ∣F′∣=m≠0|F'|=m\ne0 there, and Γρ=F({∣w∣=ρ})\Gamma_\rho=F(\{|w|=\rho\}) is an analytic Jordan curve.

Let KFK_F be the closed region it bounds and ΩF=F(Aρ)∪KF\Omega_F=F(A_\rho)\cup K_F. Then ΩF\Omega_F is open, bounded (as FF is bounded) and simply connected: its complement is the unbounded component of C∖F(Aρ)\mathbb{C}\setminus F(A_\rho). Define u=log⁡∣F−1∣/log⁡ρu=\log|F^{-1}|/\log\rho in F(Aρ)F(A_\rho) and u=1u=1 on KFK_F. Then uu is harmonic in ΩF∖KF\Omega_F\setminus K_F and of class C1C^1 up to Γρ\Gamma_\rho, since F−1F^{-1} is analytic up to Γρ\Gamma_\rho. As z→∂ΩFz\to\partial\Omega_F, F−1(z)F^{-1}(z) leaves every compact subset of {ρ≤∣w∣<1}\{\rho\le|w|<1\}, so ∣F−1(z)∣→1|F^{-1}(z)|\to1 and u(z)→0u(z)\to0. By (10)–(11), (KF,u)(K_F,u) is an admissible classical solution of (1.1) in ΩF\Omega_F with λ=1/(mρL)\lambda=1/(m\rho L). Moreover

∣ΩF∣=∣KF∣+∬Aρ∣F′(w)∣2 dA(w).(13)|\Omega_F|=|K_F|+\iint_{A_\rho}|F'(w)|^2\,dA(w). \tag*{(13)}

This gives an explicit way to generate exact solutions, which we use in Section 6.

Proof of Theorem A

By Jensen’s inequality and Lemma 2.2, for every r∈(ρ,1)r\in(\rho,1),

12π∫02π∣F′(reiθ)∣2 dθ=12π∫02πe2h(reiθ) dθ≥exp⁡(1π∫02πh(reiθ) dθ)=m2.\frac{1}{2\pi}\int_0^{2\pi}|F'(re^{i\theta})|^2\,d\theta =\frac{1}{2\pi}\int_0^{2\pi}e^{2h(re^{i\theta})}\,d\theta \ge\exp\left(\frac{1}{\pi}\int_0^{2\pi}h(re^{i\theta})\,d\theta\right)=m^2.

Integrating in rr and using (13) with ∣K∣≥0|K|\ge0,

πR2=∣Ω∣≥∣Ω∖K∣=∫ρ1∫02π∣F′(reiθ)∣2r dθ dr≥πm2(1−ρ2).\pi R^2=|\Omega|\ge|\Omega\setminus K|=\int_\rho^1\int_0^{2\pi}|F'(re^{i\theta})|^2r\,d\theta\,dr\ge\pi m^2(1-\rho^2).

Hence m≤R/1−ρ2m\le R/\sqrt{1-\rho^2} and, by (11),

λ=1mρL≥1−ρ2Rρlog⁡(1/ρ)≥μR.\lambda=\frac{1}{m\rho L}\ge\frac{\sqrt{1-\rho^2}}{R\rho\log(1/\rho)}\ge\frac{\mu}{R}.

It remains to compute μ\mu. Let f(ρ)=1−ρ2/(ρlog⁡(1/ρ))f(\rho)=\sqrt{1-\rho^2}/(\rho\log(1/\rho)), which tends to +∞+\infty as ρ→0\rho\to0 and as ρ→1\rho\to1. A direct computation gives

f′f(ρ)=−ρ1−ρ2−1ρ+1ρlog⁡(1/ρ),\frac{f'}{f}(\rho)=-\frac{\rho}{1-\rho^2}-\frac{1}{\rho}+\frac{1}{\rho\log(1/\rho)},

which vanishes if and only if log⁡(1/ρ)=1−ρ2\log(1/\rho)=1-\rho^2. The function ψ(ρ)=log⁡(1/ρ)−1+ρ2\psi(\rho)=\log(1/\rho)-1+\rho^2 satisfies ψ′(ρ)=2ρ−1/ρ\psi'(\rho)=2\rho-1/\rho. It is therefore decreasing on (0,2−1/2)(0,2^{-1/2}) and increasing on (2−1/2,1)(2^{-1/2},1); it tends to +∞+\infty at 00, is negative at 2−1/22^{-1/2} and vanishes at ρ=1\rho=1. Hence it has a unique root ρ∗∈(0,2−1/2)\rho_*\in(0,2^{-1/2}), and at this root f(ρ∗)=1/(ρ∗1−ρ∗2)f(\rho_*)=1/(\rho_*\sqrt{1-\rho_*^2}). Numerically, ρ∗=0.4507636…\rho_*=0.4507636\ldots and μ=2.4852675…\mu=2.4852675\ldots.

If Ω\Omega is convex of class C1C^{1}, then by [14] problem (1.1) has a classical solution with KK convex at λ=c(Ω)\lambda=c(\Omega). This solution is admissible, so c(Ω)≥μ/R=(μ/e)c(Ω∗)c(\Omega)\geq\mu/R=(\mu/e)c(\Omega^{*}) by (5). □\square

Remark 2.3. Two inequalities are lossy in this argument: Jensen’s inequality, which is an equality only if ∣F′∣|F'| is constant on each circle, and the bound ∣K∣≥0|K|\geq0. For the disc (F(w)=mw, ρ=1/e)(F(w)=mw,\ \rho=1/e) one has ∣K∣=πm2ρ2|K|=\pi m^{2}\rho^{2}. Any lower bound of the form ∣K∣≥πm2ρ2Θ|K|\geq\pi m^{2}\rho^{2}\Theta, with Θ>0\Theta>0 depending on how far 12π∫∣F′∣2\frac{1}{2\pi}\int|F'|^{2} is from m2m^{2}, would improve the constant μ/e\mu/e.

Proof of Proposition B

Assume that (1.6) holds on A\mathcal{A}, and let (K,u)(K,u) be an admissible solution in Ω\Omega with ∣Ω∣=πR2|\Omega|=\pi R^{2}. The dictionary above gives (ρ,m,φ)∈A(\rho,m,\varphi)\in\mathcal{A} with Ω=ΩF\Omega=\Omega_{F} and λ=1/(mρL)\lambda=1/(m\rho L). Then (1.6) reads πR2≥π(e/λ)2\pi R^{2}\geq\pi(e/\lambda)^{2}, i.e. λ≥e/R\lambda\geq e/R.

Conversely, assume the Flucher–Rumpf inequality for admissible solutions. Each (ρ,m,φ)∈A(\rho,m,\varphi)\in\mathcal{A} produces, by § 2.2, the admissible solution (KF,u)(K_{F},u) in ΩF\Omega_{F} with λ=1/(mρL)\lambda=1/(m\rho L). Hence λ≥e/R\lambda\geq e/R with πR2=∣ΩF∣\pi R^{2}=|\Omega_{F}|, that is, ∣ΩF∣≥πe2/λ2=πm2(eρL)2|\Omega_{F}|\geq\pi e^{2}/\lambda^{2}=\pi m^{2}(e\rho L)^{2}. □\square

Remark 2.4. The disc corresponds to φ≡iγ\varphi\equiv i\gamma and ρ=1/e\rho=1/e, where (1.6) is an equality. Inequality (1.6) with the factor (eρlog⁡(1/ρ))2(e\rho\log(1/\rho))^{2} replaced by 11 is false for ρ\rho close to 11 (see (15) below, where Qk(ρ)<0Q_{k}(\rho)<0 for ρ\rho near 11). Thus the factor (eρlog⁡(1/ρ))2≤1(e\rho\log(1/\rho))^{2}\leq1, which vanishes as ρ→1\rho\to1, is essential.

Second variation at the disc

Reduction at the turning point

Let Ωε\Omega_{\varepsilon} be as in Proposition C. Along the branch of solutions in Ωε\Omega_{\varepsilon}, parametrised by a smooth parameter ss with s=s0s=s_{0} at the turning point of the disc, we write

λ(s,ε)=λ0(s)+ελ1(s)+ε2λ2(s)+O(ε3).\lambda(s,\varepsilon)=\lambda_{0}(s)+\varepsilon\lambda_{1}(s)+\varepsilon^{2}\lambda_{2}(s)+O(\varepsilon^{3}).

The rotation θ↦θ+π/k\theta\mapsto\theta+\pi/k maps Ωε\Omega_{\varepsilon} onto Ω−ε\Omega_{-\varepsilon}, so λ1≡0\lambda_{1}\equiv0 and the expansion is even in ε\varepsilon. Since λ0′(s0)=0\lambda_{0}'(s_{0})=0 and λ0′′(s0)>0\lambda_{0}''(s_{0})>0, the minimum in ss moves by O(ε2)O(\varepsilon^{2}), which changes the minimal value only by O(ε4)O(\varepsilon^{4}). Hence

c(Ωε)=min⁡sλ(s,ε)=e+ε2λ2(s0)+O(ε4).(14)c(\Omega_{\varepsilon})=\min_{s}\lambda(s,\varepsilon)=e+\varepsilon^{2}\lambda_{2}(s_{0})+O(\varepsilon^{4}). \tag*{(14)}

For the same reason λ2(s0)\lambda_{2}(s_{0}) does not depend on the choice of parametrisation of the branch. This is the simplest instance of a Lyapunov–Schmidt reduction at a fold.

First derivation: conformal parametrisation

We use the conformal modulus ρ\rho as the branch parameter (ρ0=1/e)(\rho_{0}=1/e). Take k≥2k\geq2, a>0a>0 and

φε(w)=εa(wk−ρ2kw−k),\varphi_{\varepsilon}(w)=\varepsilon a\left(w^{k}-\rho^{2k}w^{-k}\right),

which is of the form (2.4). All exponents in eφε=∑jgjwje^{\varphi_{\varepsilon}}=\sum_{j}g_{j}w^{j} are multiples of kk, so g−1=0g_{-1}=0 and

F(w)=m∑jgjj+1wj+1F(w)=m\sum_{j}\frac{g_{j}}{j+1}w^{j+1}

is single-valued and analytic in C∖{0}\mathbb{C}\setminus\{0\}. For ε\varepsilon small it is a C1C^{1}-small perturbation of mwmw on a neighbourhood of Aρ‾\overline{A_{\rho}}, hence univalent there. Thus (ρ,m,φε)∈A(\rho,m,\varphi_{\varepsilon})\in\mathcal{A} and, by § 2.2, we obtain an exact solution with λ=1/(mρL)\lambda=1/(m\rho L).

Area. If F(eiθ)=∑naneinθF(e^{i\theta})=\sum_{n}a_{n}e^{in\theta}, the area enclosed by the Jordan curve F({∣w∣=1})F(\{|w|=1\}) is ∣ΩF∣=π∑nn∣an∣2|\Omega_{F}|=\pi\sum_{n}n|a_{n}|^{2}. Here aj+1=mgj/(j+1)a_{j+1}=mg_{j}/(j+1). From eφ=1+φ+φ2/2+O(ε3)e^{\varphi}=1+\varphi+\varphi^{2}/2+O(\varepsilon^{3}) we get

g0=1−ε2a2ρ2k+O(ε4),gk=εa+O(ε3),g−k=−εaρ2k+O(ε3),g±2k=O(ε2),g_{0}=1-\varepsilon^{2}a^{2}\rho^{2k}+O(\varepsilon^{4}),\qquad g_{k}=\varepsilon a+O(\varepsilon^{3}),\qquad g_{-k}=-\varepsilon a\rho^{2k}+O(\varepsilon^{3}),\qquad g_{\pm2k}=O(\varepsilon^{2}),

and therefore

∣ΩF∣πm2=∑j≠−1∣gj∣2j+1=1+ε2a2Qk(ρ)+O(ε4),Qk(ρ)=1k+1−2ρ2k−ρ4kk−1.(15)\frac{|\Omega_F|}{\pi m^2}=\sum_{j\ne-1}\frac{|g_j|^2}{j+1}=1+\varepsilon^2a^2Q_k(\rho)+O(\varepsilon^4),\qquad Q_k(\rho)=\frac{1}{k+1}-2\rho^{2k}-\frac{\rho^{4k}}{k-1}. \tag*{(15)}

Shape of the outer boundary. On the outer circle,

F(eiθ)=meiθ(1+εa[eikθk+1+ρ2ke−ikθk−1])+O(ε2).F(e^{i\theta})=me^{i\theta}\left(1+\varepsilon a\left[\frac{e^{ik\theta}}{k+1}+\frac{\rho^{2k}e^{-ik\theta}}{k-1}\right]\right)+O(\varepsilon^2).

Up to an O(ε)O(\varepsilon) tangential reparametrisation, ∂ΩF\partial\Omega_F is therefore the curve r=m(1+δcos⁡kψ+O(ε2))r=m(1+\delta\cos k\psi+O(\varepsilon^2)), with

δ=εaSk(ρ),Sk(ρ)=1k+1+ρ2kk−1.\delta=\varepsilon aS_k(\rho),\qquad S_k(\rho)=\frac{1}{k+1}+\frac{\rho^{2k}}{k-1}.

Matching. The O(ε2)O(\varepsilon^2) corrections to φ\varphi needed to match a prescribed outer boundary exactly involve only the modes 00 (with imaginary coefficient, i.e. a rotation) and ±2k\pm2k. They do not affect (3.2) at order ε2\varepsilon^2. A mode 2k2k of amplitude O(δ2)O(\delta^2) in the boundary changes cc only at order δ4\delta^4: by rotational invariance, the cubic terms of the expansion of cc couple it only to the square of the kk-th mode. Normalising ∣ΩF∣=π|\Omega_F|=\pi and using (2.3), the solution of conformal modulus ρ\rho in the domain with boundary amplitude δ\delta satisfies

λ(ρ,δ)2=1+δ2Qk(ρ)/Sk(ρ)2(ρlog⁡(1/ρ))2+O(δ4).\lambda(\rho,\delta)^2=\frac{1+\delta^2Q_k(\rho)/S_k(\rho)^2}{(\rho\log(1/\rho))^2}+O(\delta^4).

By (3.1), c=e(1+δ2Qk(1/e)/(2Sk(1/e)2))+O(δ4)c=e(1+\delta^2Q_k(1/e)/(2S_k(1/e)^2))+O(\delta^4), that is,

Ak=eQk(1/e)2Sk(1/e)2.(16)A_k=\frac{eQ_k(1/e)}{2S_k(1/e)^2}. \tag*{(16)}

With xk=e−2kx_k=e^{-2k} one has Qk(1/e)=[(k−1)−2(k2−1)xk−(k+1)xk2]/(k2−1)Q_k(1/e)=[(k-1)-2(k^2-1)x_k-(k+1)x_k^2]/(k^2-1) and Sk(1/e)=[(k−1)+(k+1)xk]/(k2−1)S_k(1/e)=[(k-1)+(k+1)x_k]/(k^2-1), which gives the formula of Proposition C.

Second derivation: domain expansion

We check (3.3) by a direct expansion in the physical plane. Let Ω~ε={r<1+εcos⁡kθ}\widetilde{\Omega}_\varepsilon=\{r<1+\varepsilon\cos k\theta\}, whose area is π(1+ε2/2)\pi(1+\varepsilon^2/2), and write c=cos⁡kθc=\cos k\theta. Let

∂K={r=s+εbc+ε2g2(θ)},∫02πg2 dθ=0,\partial K=\{r=s+\varepsilon b c+\varepsilon^2g_2(\theta)\},\qquad\int_0^{2\pi}g_2\,d\theta=0,

so that ss is the mean radius, and set ℓ=log⁡s<0\ell=\log s<0. We look for

u=u0+εu1+ε2u2,u0=log⁡rℓ,u1=(αrk+βr−k)c,u=u_0+\varepsilon u_1+\varepsilon^2u_2,\qquad u_0=\frac{\log r}{\ell},\qquad u_1=(\alpha r^k+\beta r^{-k})c,

where u2u_2 has a radial part c0+d0log⁡rc_0+d_0\log r plus modes that average to zero on circles. Below, ⟨⋅⟩\langle\cdot\rangle denotes the mean in θ\theta; note ⟨c2⟩=12\langle c^2\rangle=\frac12.

First order. The conditions u=0u=0 on the outer boundary, u=1u=1 on ∂K\partial K and ∣∇u∣|\nabla u| constant on ∂K\partial K give

α+β=−1ℓ,αsk+βs−k=−bsℓ,k(αsk−βs−k)=bsℓ.\alpha+\beta=-\frac{1}{\ell},\qquad\alpha s^k+\beta s^{-k}=-\frac{b}{s\ell},\qquad k(\alpha s^k-\beta s^{-k})=\frac{b}{s\ell}.

With B=b/(sℓ)B=b/(s\ell) and D(s)=(k−1)s−k+(k+1)skD(s)=(k-1)s^{-k}+(k+1)s^k, the solution is

αsk=−(k−1)B2k,βs−k=−(k+1)B2k,B=2kℓD(s),b=2ksD(s).\alpha s^k=-\frac{(k-1)B}{2k},\qquad\beta s^{-k}=-\frac{(k+1)B}{2k},\qquad B=\frac{2k}{\ell D(s)},\qquad b=\frac{2ks}{D(s)}.

In particular $\lambda_1=0.

Second order, outer boundary. Expanding u(1+εc,θ)=0u(1+\varepsilon c,\theta)=0 to order ε2\varepsilon^2 and averaging gives

−14ℓ+k(α−β)2+c0=0,hencec0=14ℓ+B4[(k−1)s−k−(k+1)sk].-\frac{1}{4\ell}+\frac{k(\alpha-\beta)}{2}+c_0=0, \qquad\text{hence}\qquad c_0=\frac{1}{4\ell}+\frac{B}{4}\left[(k-1)s^{-k}-(k+1)s^k\right].

Second order, inner boundary. With η=εbc+ε2g2\eta=\varepsilon bc+\varepsilon^2g_2 one has u0(s+η)=1+(η/s−η2/(2s2))/ℓ+O(ε3)u_0(s+\eta)=1+(\eta/s-\eta^2/(2s^2))/\ell+O(\varepsilon^3). The averaged ε2\varepsilon^2-part of u=1u=1 on ∂K\partial K reads

−b24s2ℓ+bk2s(αsk−βs−k)+c0+d0ℓ=0.-\frac{b^2}{4s^2\ell}+\frac{bk}{2s}\left(\alpha s^k-\beta s^{-k}\right)+c_0+d_0\ell=0.

Since b2/(4s2ℓ)=B2ℓ/4b^2/(4s^2\ell)=B^2\ell/4 and bk(αsk−βs−k)/(2s)=B2ℓ/2bk(\alpha s^k-\beta s^{-k})/(2s)=B^2\ell/2, this gives d0=−c0/ℓ−B2/4d_0=-c_0/\ell-B^2/4.

Second order, gradient condition. On ∂K\partial K we use ∣∇u∣2=(∂ru)2+r−2(∂θu)2|\nabla u|^2=(\partial_r u)^2+r^{-2}(\partial_\theta u)^2. The first-order part of ∂ru\partial_r u vanishes by the first-order equations, so

∂ru=1sℓ+ε2X+O(ε3),X=−g2s2ℓ+b2c2s3ℓ+bc ∂r2u1(s)+∂ru2(s),\partial_r u=\frac{1}{s\ell}+\varepsilon^2X+O(\varepsilon^3), \qquad X=-\frac{g_2}{s^2\ell}+\frac{b^2c^2}{s^3\ell}+bc\,\partial_r^2u_1(s)+\partial_ru_2(s),

and r−2(∂θu)2=ε2k2B2sin⁡2(kθ)/s2+O(ε3)r^{-2}(\partial_\theta u)^2=\varepsilon^2k^2B^2\sin^2(k\theta)/s^2+O(\varepsilon^3), because αsk+βs−k=−B\alpha s^k+\beta s^{-k}=-B. Writing λ=λ0+ε2λ2\lambda=\lambda_0+\varepsilon^2\lambda_2 with λ0=−1/(sℓ)>0\lambda_0=-1/(s\ell)>0 and averaging λ2=∣∇u∣2\lambda^2=|\nabla u|^2, we get

λ2=−⟨X⟩+k2B24s2λ0=−⟨X⟩−k2B2ℓ4s.\lambda_2=-\langle X\rangle+\frac{k^2B^2}{4s^2\lambda_0} =-\langle X\rangle-\frac{k^2B^2\ell}{4s}.

Using ∂r2u1(s)=−B(k2+1)s−2c\partial_r^2u_1(s)=-B(k^2+1)s^{-2}c, one finds ⟨X⟩=−k2B2ℓ/(2s)+d0/s\langle X\rangle=-k^2B^2\ell/(2s)+d_0/s, and therefore

λ2(s)=1s[k2B2ℓ4+14ℓ2+B4ℓ((k−1)s−k−(k+1)sk)+B24].(17)\lambda_2(s)=\frac{1}{s}\left[\frac{k^2B^2\ell}{4}+\frac{1}{4\ell^2}+\frac{B}{4\ell}\left((k-1)s^{-k}-(k+1)s^k\right)+\frac{B^2}{4}\right]. \tag*{(17)}

At s=1/es=1/e (ℓ=−1\ell=-1, B=−2k/DB=-2k/D, D=(k−1)ek+(k+1)e−kD=(k-1)e^k+(k+1)e^{-k}) this becomes

λ2(1/e)=e[14−k2+k(k−1)ekD−k2(k2−1)D2].\lambda_2(1/e)=e\left[\frac{1}{4}-\frac{k}{2}+\frac{k(k-1)e^k}{D}-\frac{k^2(k^2-1)}{D^2}\right].

Since c(tΩ)=c(Ω)/tc(t\Omega)=c(\Omega)/t and Ωε=Ω~ε/1+ε2/2\Omega_\varepsilon=\widetilde{\Omega}_\varepsilon/\sqrt{1+\varepsilon^2/2}, the area normalisation contributes +e/4+e/4. Hence

Ak=e[1−k2+kΦk−k2Φk(1−Φk)],Φk=(k−1)e2k(k−1)e2k+k+1.A_k=e\left[\frac{1-k}{2}+k\Phi_k-k^2\Phi_k(1-\Phi_k)\right], \qquad \Phi_k=\frac{(k-1)e^{2k}}{(k-1)e^{2k}+k+1}.

Put xk=e−2kx_k=e^{-2k} and Dk′=(k−1)+(k+1)xkD'_k=(k-1)+(k+1)x_k, so that Φk=(k−1)/Dk′\Phi_k=(k-1)/D'_k. One checks

2Dk′2[1−k2+kΦk−k2Φk(1−Φk)]=(k−1)2(k+1)−2(k2−1)2xk−(k−1)(k+1)2xk2.2D_k^{\prime2}\left[\frac{1-k}{2}+k\Phi_k-k^2\Phi_k(1-\Phi_k)\right] =(k-1)^2(k+1)-2(k^2-1)^2x_k-(k-1)(k+1)^2x_k^2.

The right-hand side is exactly (k2−1)[(k−1)−2(k2−1)xk−(k+1)xk2](k^2-1)[(k-1)-2(k^2-1)x_k-(k+1)x_k^2], so both derivations agree identically.

Positivity

For k=1k=1 the factor k2−1k^2-1 vanishes, so A1=0A_1=0, as it must be, since translations leave cc invariant. For k≥2k\ge2, (k+1)xk=(k+1)e−2k≤3e−4<0.055(k+1)x_k=(k+1)e^{-2k}\le3e^{-4}<0.055, because k↦(k+1)e−2kk\mapsto(k+1)e^{-2k} is decreasing. Hence

(k−1)−2(k2−1)xk−(k+1)xk2≥(k−1)(1−0.11)−0.055xk>0,(k-1)-2(k^2-1)x_k-(k+1)x_k^2\ge(k-1)(1-0.11)-0.055x_k>0,

and Ak>0A_k>0. As k→∞k\to\infty, Ak=e(k+1)/2+O(k3e−2k)A_k=e(k+1)/2+O(k^3e^{-2k}).

By rotational invariance, the second variation is invariant under rotations; on each two-dimensional space spanned by cos⁡kθ\cos k\theta and sin⁡kθ\sin k\theta the rotation group acts irreducibly, so the second variation is a multiple of the identity there. Thus it is the diagonal quadratic form

c(Ω)=e+ε2∑k≥2Ak(ak2+bk2)+O(ε3)for ∂Ω: r∝1+ε∑k(akcos⁡kθ+bksin⁡kθ),c(\Omega)=e+\varepsilon^2\sum_{k\ge2}A_k(a_k^2+b_k^2)+O(\varepsilon^3) \qquad\text{for }\partial\Omega:\ r\propto1+\varepsilon\sum_k(a_k\cos k\theta+b_k\sin k\theta),

which is positive definite modulo translations and, since Ak≍kA_k \asymp k, coercive in the H1/2H^{1/2} norm of the boundary perturbation.

Remark 3.1. To turn Proposition C into a rigorous local minimality statement, one needs: (a) the implicit function theorem for the branch of solutions near the fold, in Hölder spaces, to justify the expansions; (b) a stability argument of Dambrine–Pierre type [8] to pass from H1/2H^{1/2}-coercivity of the second variation to strict local minimality in a C2,αC^{2,\alpha} neighbourhood (the usual two-norm discrepancy). We do not carry this out here. The numerical values in Table 4 confirm the formula independently.

kkε=0.03\varepsilon=0.03ε=0.06\varepsilon=0.06extrapolatedAkA_k (Prop. C)
23.258603.262083.25743.257430
35.278155.283965.27625.276246
46.761656.751196.76516.765339
58.133888.089918.14858.149293

Table 4. Second variation coefficients.

A Rellich–Pohozaev identity and the proof of Theorem D

Throughout this section Ω⊂RN\Omega\subset\mathbb{R}^{N}, N≥2N \ge2, is a bounded domain of class C1,αC^{1,\alpha}, strictly star-shaped with respect to the origin, and (K,u)(K,u) is a classical solution of (1.1). Set D=Ω∖KD=\Omega\setminus K. By the Schauder estimates, u∈C1,αu \in C^{1,\alpha} up to ∂Ω\partial\Omega, and u∈C1u \in C^1 up to ∂K\partial K by assumption, so u∈C2(D)∩C1(D‾)u \in C^2(D) \cap C^1(\overline{D}). We denote by ν\nu the outer unit normal to Ω\Omega on ∂Ω\partial\Omega and to KK on ∂K\partial K.

Two integral identities

Lemma 4.1.

∫∂Ω∣∇u∣ dσ=∫∂K∣∇u∣ dσ=cap⁡(K;Ω)=λPer⁡(K).\int_{\partial\Omega} |\nabla u|\,d\sigma= \int_{\partial K} |\nabla u|\,d\sigma= \operatorname{cap}(K;\Omega) = \lambda\operatorname{Per}(K).

Proof. By the strong maximum principle 0<u<10 < u < 1 in DD, so ∣∇u∣=−∂νu|\nabla u|=-\partial_{\nu}u on ∂Ω\partial\Omega and ∣∇u∣=∂νu|\nabla u|=\partial_{\nu}u on ∂K\partial K (one-sided derivatives from DD). Since uu is harmonic, ∫∂D∂νDu dσ=0\int_{\partial D}\partial_{\nu_D}u\,d\sigma=0, which gives the first equality. Green’s formula ∫D∣∇u∣2=∫∂Du∂νDu dσ=∫∂K∣∇u∣ dσ\int_D|\nabla u|^2=\int_{\partial D}u\partial_{\nu_D}u\,d\sigma=\int_{\partial K}|\nabla u|\,d\sigma and the minimality of uK=uu_K=u give the second; the third follows from ∣∇u∣=λ|\nabla u|=\lambda on ∂K\partial K. □

Lemma 4.2 (Rellich–Pohozaev identity). Identity (1.8) holds.

Proof. Consider the vector field V=(x⋅∇u)∇u−12∣∇u∣2xV=(x\cdot\nabla u)\nabla u-\frac{1}{2}|\nabla u|^2x [20, 17]. For harmonic uu,

div⁡((x⋅∇u)∇u)=∣∇u∣2+x⋅D2u∇u,div⁡(12∣∇u∣2x)=N2∣∇u∣2+x⋅D2u∇u,\begin{aligned} \operatorname{div}\bigl((x\cdot\nabla u)\nabla u\bigr) &= |\nabla u|^2+x\cdot D^2u\nabla u,\\ \operatorname{div}\left(\frac{1}{2}|\nabla u|^2x\right) &= \frac{N}{2}|\nabla u|^2+x\cdot D^2u\nabla u, \end{aligned}

so div⁡V=2−N2∣∇u∣2\operatorname{div}V=\frac{2-N}{2}|\nabla u|^2. Since V∈C(D‾)∩C1(D)V\in C(\overline{D})\cap C^1(D), div⁡V∈L1(D)\operatorname{div}V\in L^1(D) and ∂D\partial D is of class C1C^1, the divergence theorem applies (e.g. by exhausting DD by smooth subdomains).

On each component of ∂D\partial D, uu is constant, so ∇u=(∂νDu)νD\nabla u=(\partial_{\nu_D}u)\nu_D and V⋅νD=12(x⋅νD)∣∇u∣2V\cdot\nu_D=\frac{1}{2}(x\cdot\nu_D)|\nabla u|^2. On ∂K\partial K, νD=−ν\nu_D=-\nu and ∣∇u∣=λ|\nabla u|=\lambda, while ∫∂Kx⋅ν dσ=N∣K∣\int_{\partial K}x\cdot\nu\,d\sigma=N|K| by the divergence theorem in KK. Using Lemma 4.1,

2−N2cap⁡(K;Ω)=12∫∂Ω∣∇u∣2(x⋅ν) dσ−λ22N∣K∣,\frac{2-N}{2}\operatorname{cap}(K;\Omega)=\frac{1}{2}\int_{\partial\Omega}|\nabla u|^2(x\cdot\nu)\,d\sigma-\frac{\lambda^2}{2}N|K|,

which is (8). □

The geometric factor

Lemma 4.3. I(Ω)≥∣∂Ω∣2N∣Ω∣≥I(Ω∗)=∣∂Ω∗∣RI(\Omega)\ge\frac{|\partial\Omega|^2}{N|\Omega|}\ge I(\Omega^*)=\frac{|\partial\Omega^*|}{R}, where ∣Ω∗∣=∣BR∣|\Omega^*|=|B_R|. Hence θ(Ω)∈(0,1]\theta(\Omega)\in(0,1], with equality if and only if Ω\Omega is a ball centred at the origin.

Proof. By Cauchy–Schwarz, ∣∂Ω∣2≤I(Ω)∫∂Ωx⋅ν dσ=I(Ω)N∣Ω∣|\partial\Omega|^2\le I(\Omega)\int_{\partial\Omega}x\cdot\nu\,d\sigma=I(\Omega)N|\Omega|. The isoperimetric inequality gives ∣∂Ω∣2N∣Ω∣≥∣∂BR∣2N∣BR∣=∣∂BR∣R=I(BR)\frac{|\partial\Omega|^2}{N|\Omega|}\ge\frac{|\partial B_R|^2}{N|B_R|}=\frac{|\partial B_R|}{R}=I(B_R). Equality in the isoperimetric inequality forces Ω\Omega to be a ball; equality in Cauchy–Schwarz forces x⋅νx\cdot\nu to be constant on ∂Ω\partial\Omega, so the ball is centred at the origin. □

Proof of Theorem D

Step 1: a lower bound by the capacity. By Lemma 4.1 and Cauchy–Schwarz,

cap⁡(K;Ω)2=(∫∂Ω∣∇u∣ dσ)2≤∫∂Ω∣∇u∣2(x⋅ν) dσ⋅I(Ω).\operatorname{cap}(K;\Omega)^2=\left(\int_{\partial\Omega}|\nabla u|\,d\sigma\right)^2\leq\int_{\partial\Omega}|\nabla u|^2(x\cdot\nu)\,d\sigma\cdot I(\Omega).

Inserting this in (8) gives

λ2N∣K∣≥ΦΩ(cap⁡(K;Ω)),ΦΩ(t)=t2I(Ω)+(N−2)t,(18)\lambda^2N|K|\geq\Phi_\Omega(\operatorname{cap}(K;\Omega)),\qquad\Phi_\Omega(t)=\frac{t^2}{I(\Omega)}+(N-2)t, \tag*{(18)}

and ΦΩ\Phi_\Omega is increasing on [0,∞)[0,\infty).

Step 2: symmetrisation. Let RR and s<Rs<R be defined by ∣BR∣=∣Ω∣|B_R|=|\Omega| and ∣Bs∣=∣K∣|B_s|=|K|, and put C(s)=cap⁡(Bs‾;BR)C(s)=\operatorname{cap}(\overline{B_s};B_R), that is,

C(s)=2πlog⁡(R/s)(N=2),C(s)=(N−2)∣SN−1∣s2−N−R2−N(N≥3).C(s)=\frac{2\pi}{\log(R/s)}\quad(N=2),\qquad C(s)=\frac{(N-2)|S^{N-1}|}{s^{2-N}-R^{2-N}}\quad(N\geq3).

Extend uu by 11 on KK and by 00 outside Ω\Omega, and let u∗u^* be its Schwarz symmetrisation. By the Pólya–Szegő inequality [18, 6], u∗∈W01(BR)u^*\in W_0^1(B_R) and ∫∣∇u∗∣2≤∫∣∇u∣2\int|\nabla u^*|^2\leq\int|\nabla u|^2. Since 0<u<10<u<1 in DD, the set {u∗=1}\{u^*=1\} is the closed ball of measure ∣{u=1}∣=∣K∣|\{u=1\}|=|K|, namely Bs‾\overline{B_s}. Thus u∗u^* is admissible for cap⁡(Bs‾;BR)\operatorname{cap}(\overline{B_s};B_R), and

cap⁡(K;Ω)≥C(s).\operatorname{cap}(K;\Omega)\geq C(s).

By (18) and the monotonicity of ΦΩ\Phi_\Omega,

λ2N∣Bs∣≥C(s)2I(Ω)+(N−2)C(s).(19)\lambda^2N|B_s|\geq\frac{C(s)^2}{I(\Omega)}+(N-2)C(s). \tag*{(19)}

Step 3: the bound θ\sqrt{\theta}. For the concentric radial solution K=Bs‾K=\overline{B_s} in Ω=BR\Omega=B_R, the gradient ∣∇u∣=C(s)/∣∂BR∣|\nabla u|=C(s)/|\partial B_R| is constant on ∂BR\partial B_R and x⋅ν=Rx\cdot\nu=R. Hence both Cauchy–Schwarz steps above are equalities, and (8) reads

λrad(s)2N∣Bs∣=C(s)2I(BR)+(N−2)C(s).(20)\lambda_{\mathrm{rad}}(s)^2N|B_s|=\frac{C(s)^2}{I(B_R)}+(N-2)C(s). \tag*{(20)}

Since I(Ω)=I(BR)/θI(\Omega)=I(B_R)/\theta with θ=θ(Ω)≤1\theta=\theta(\Omega)\leq1 (Lemma 4.3), (19) gives

λ2N∣Bs∣≥θ[C(s)2I(BR)+(N−2)C(s)]≥θ[C(s)2I(BR)+(N−2)C(s)]=θλrad(s)2N∣Bs∣.\lambda^2N|B_s|\geq\theta\left[\frac{C(s)^2}{I(B_R)}+(N-2)C(s)\right]\geq\theta\left[\frac{C(s)^2}{I(B_R)}+(N-2)C(s)\right]=\theta\lambda_{\mathrm{rad}}(s)^2N|B_s|.

Hence λ≥θλrad(s)≥θc(BR)\lambda\geq\sqrt{\theta}\lambda_{\mathrm{rad}}(s)\geq\sqrt{\theta}c(B_R), by (5).

Step 4: the bound κN\kappa_N (N≥3N\geq3). Dropping the first term in (19),

λ2≥N−2NC(s)∣Bs∣=N−2NN(N−2)R2t2(1−tN−2),t=sR.\lambda^2\geq\frac{N-2}{N}\frac{C(s)}{|B_s|}=\frac{N-2}{N}\frac{N(N-2)}{R^2t^2(1-t^{N-2})},\qquad t=\frac{s}{R}.

The function t↦t2−tNt\mapsto t^2-t^N is maximal on (0,1)(0,1) at tN−2=2/Nt^{N-2}=2/N, where it equals t2(N−2)/Nt^2(N-2)/N. Hence

λ2≥N−2NΛN(R)2,ΛN(R)=N(N/2)1N−2R.\lambda^2\geq\frac{N-2}{N}\Lambda_N(R)^2,\qquad\Lambda_N(R)=\frac{N(N/2)^{\frac{1}{N-2}}}{R}.

Dividing by c(BR)=(N−1)N−1N−2/Rc(B_R)=(N-1)^{\frac{N-1}{N-2}}/R gives λ≥κNc(BR)\lambda\geq\kappa_Nc(B_R) with κN\kappa_N as in (7). Since c(Ω∗)=c(BR)c(\Omega^*)=c(B_R), Theorem D follows by taking the infimum over all classical solutions. □\square

Remark 4.4. ΛN(R)\Lambda_N(R) is the Daners–Kawohl constant of BRB_R (for N=2N = 2, Λ2(R)=2e/R\Lambda_2(R) = 2\sqrt{e}/R). Comparing Jλ(v)J_\lambda(v) with Jλ(0)J_\lambda(0) shows that Λ(Ω)2=inf⁡Kcap⁡(K;Ω)/∣K∣\Lambda(\Omega)^2 = \inf_K \operatorname{cap}(K;\Omega)/|K|. Dropping only the boundary term in (1.8), Step 4 yields more precisely

c(Ω)≥N−2N Λ(Ω)(N≥3).c(\Omega) \ge\sqrt{\frac{N-2}{N}}\,\Lambda(\Omega) \qquad(N \ge3).

so that, for star-shaped domains, the two constants cc and Λ\Lambda are comparable. Combined with Λ(Ω)≥Λ(Ω∗)\Lambda(\Omega) \ge\Lambda(\Omega^*) [9], this gives the bound κN\kappa_N again.

Remark 4.5. As N→∞N \to\infty, R c(BR)=N−1+log⁡N+o(1)R\,c(B_R) = N - 1 + \log N + o(1) and RΛN(R)=N+log⁡(N/2)+o(1)R\Lambda_N(R) = N + \log(N/2) + o(1), so κN→1\kappa_N \to1. For Ω\Omega close to a ball, I(Ω)/I(Ω∗)−1I(\Omega)/I(\Omega^*) - 1 is quadratic in the deformation (it is minimal at the ball), so the bound θ\sqrt{\theta} deviates from 1 only at second order. This is the same order as in Proposition C, but with the opposite sign: it does not prove local minimality.

Proof of Corollary E

A bounded convex domain of class C1,αC^{1,\alpha} is strictly star-shaped with respect to any interior point x0x_0, since (x−x0)⋅ν≥dist⁡(x0,∂Ω)>0(x-x_0)\cdot\nu\ge\operatorname{dist}(x_0,\partial\Omega)>0. By translation invariance of cc, Theorem D applies with the origin at x0x_0, and Theorem A gives the other bound. For the ellipse x=(acos⁡t,bsin⁡t)x=(a\cos t,b\sin t) centred at x0=0x_0=0, one has dσ=a2sin⁡2t+b2cos⁡2t dt\mathrm{d}\sigma=\sqrt{a^2\sin^2t+b^2\cos^2t}\,\mathrm{d}t and x⋅ν=ab/a2sin⁡2t+b2cos⁡2tx\cdot\nu=ab/\sqrt{a^2\sin^2t+b^2\cos^2t}, so

I=π(a2+b2)ab,θ=2aba2+b2=2q+q−1,q=ab.I=\frac{\pi(a^2+b^2)}{ab}, \qquad\theta=\frac{2ab}{a^2+b^2}=\frac{2}{q+q^{-1}}, \qquad q=\frac{a}{b}.

The inequality θ>μ/e\sqrt{\theta}>\mu/e is equivalent to q+q−1<2e2/μ2≈2.39261q+q^{-1}<2e^2/\mu^2\approx2.39261, i.e. q<1.8529q<1.8529. □

Polarization: an obstruction and what survives

For a closed half-space HH with reflection σ\sigma, the polarization of a set EE is

EH=((E∪σE)∩H)∪((E∩σE)∩Hc),E^H=((E\cup\sigma E)\cap H)\cup((E\cap\sigma E)\cap H^c),

and the polarization of a function vv is max⁡(v,v∘σ)\max(v,v\circ\sigma) on HH and min⁡(v,v∘σ)\min(v,v\circ\sigma) on HcH^c; see [5]. A natural strategy towards (1.3) is to show that c(ΩH)≤c(Ω)c(\Omega^H)\le c(\Omega) and to conclude by approximating Ω∗\Omega^* by iterated polarizations. If cc admits the Beurling-type characterisation

c(Ω)=inf⁡K⊂Ωsup⁡∂K∣∇uK∣,c(\Omega)=\inf_{K\subset\Omega}\sup_{\partial K}|\nabla u_K|,

where uKu_K is the capacitary potential of KK in Ω\Omega, it would suffice to transport subsolutions, i.e. to deduce sup⁡∂KH∣∇uKH∣≤λ\sup_{\partial K^H}|\nabla u_{K^H}|\le\lambda in ΩH\Omega^H from sup⁡∂K∣∇uK∣≤λ\sup_{\partial K}|\nabla u_K|\le\lambda. For convex domains, such a characterisation is closely related to the subsolution methods of Beurling [3], Acker [1] and Henrot–Shahgholian [14]; our numerical tests (§ 6.5) support it. The next lemma shows that the polarized free set itself is never a good competitor. Note that it does not disprove c(ΩH)≤c(Ω)c(\Omega^H)\le c(\Omega): other free sets in ΩH\Omega^H may still do better.

Lemma 5.1 (corner obstruction). Assume that ∂K\partial K and ∂(σK)\partial(\sigma K) are C1,αC^{1,\alpha} curves that cross transversally at a point P∈int⁡(Hc)∩ΩHP\in\operatorname{int}(H^c)\cap\Omega^H. Then sup⁡∂KH∣∇uKH∣=+∞\sup_{\partial K^H}|\nabla u_{K^H}|=+\infty.

Proof. Near PP, KHK^H coincides with K∩σKK\cap\sigma K, which has at PP a convex corner of interior angle ϑ∈(0,π)\vartheta\in(0,\pi). Hence ΩH∖KH\Omega^H\setminus K^H has at PP an opening angle ω=2π−ϑ>π\omega=2\pi-\vartheta>\pi. The function v=1−uKHv=1-u_{K^H} is positive and harmonic in ΩH∖KH\Omega^H\setminus K^H and vanishes on both arcs of ∂KH\partial K^H through PP. By the corner asymptotics [12 Ch. 4–5], after straightening the two arcs by a conformal map, v=arπ/ωsin⁡(πϕ/ω)+o(rπ/ω)v=a r^{\pi/\omega}\sin(\pi\phi/\omega)+o(r^{\pi/\omega}), with a>0a>0 by the Hopf lemma at corners. Hence ∣∇v∣∼a(π/ω)rπ/ω−1|\nabla v|\sim a(\pi/\omega)r^{\pi/\omega-1} along ∂KH\partial K^H, which is unbounded since π/ω<1\pi/\omega<1. □

On the side HH, the union K∪σKK \cup\sigma K creates reentrant corners of KHK^{H}, at which the gradient vanishes. The obstruction is therefore entirely located in HcH^{c}, where the polarized function is a minimum of harmonic functions and is only superharmonic. The critical free sets of nonsymmetric convex domains are smooth and convex, and for most HH they cross their reflections transversally. What survives is the energy inequality.

Proposition 5.2. For K⊂ΩK \subset\Omega set Eλ(K;Ω)=cap⁡(K;Ω)−λ2∣K∣E_{\lambda}(K;\Omega)=\operatorname{cap}(K;\Omega)-\lambda^{2}|K|. Then, for every λ\lambda,

Eλ(KH;ΩH)≤Eλ(K;Ω).E_{\lambda}(K^{H};\Omega^{H}) \leq E_{\lambda}(K;\Omega).

Proof. Extend uKu_{K} by 00 outside Ω\Omega; it is continuous on RN\mathbb{R}^{N} and {uK>0}=Ω\{u_{K}>0\}=\Omega, {uK≥1}=K\{u_{K}\geq1\}=K. Its polarization uKHu_{K}^{H} is continuous, vanishes outside {uK>0}H=ΩH\{u_{K}>0\}^{H}=\Omega^{H}, and therefore belongs to W01,2(ΩH)W_{0}^{1,2}(\Omega^{H}). It satisfies uKH≥1u_{K}^{H}\geq1 on {uK≥1}H=KH\{u_{K}\geq1\}^{H}=K^{H}, and has the same Dirichlet integral as uKu_{K} [5]. Hence cap⁡(KH;ΩH)≤cap⁡(K;Ω)\operatorname{cap}(K^{H};\Omega^{H})\leq\operatorname{cap}(K;\Omega), while ∣KH∣=∣K∣|K^{H}|=|K|. □

By the Hadamard formula δcap⁡(K;Ω)=∫∂K∣∇uK∣2Vn dσ\delta\operatorname{cap}(K;\Omega)=\int_{\partial K}|\nabla u_{K}|^{2}V_{n}\,d\sigma, solutions of (1) are exactly the critical points of EλE_{\lambda}. For the disc, the elliptic solution is a local minimum and the hyperbolic one a saddle, and they merge into an inflection point at λ=c\lambda=c. Proposition 5.2 implies that polarization decreases the Daners–Kawohl constant Λ\Lambda (the threshold for inf⁡Eλ<0\inf E_{\lambda}<0). It also decreases the capacitary profile CΩ(v)=inf⁡{cap⁡(K;Ω):∣K∣=v}C_{\Omega}(v)=\inf\{\operatorname{cap}(K;\Omega):|K|=v\}. But cc is the threshold of existence of a critical point, not an energy level, and it is not monotone under a decrease of EλE_{\lambda}.

In the convex class, Cardaliaguet and Tahraoui [7] studied the profile

FΩ(σ)=inf⁡{cap⁡(D;Ω):D⊂⊂Ω convex, ∣D∣=σ}.F_{\Omega}(\sigma)=\inf\{\operatorname{cap}(D;\Omega):D\subset\subset\Omega\text{ convex},\ |D|=\sigma\}.

They showed that, at points where FΩF_{\Omega} is differentiable, convex minimisers are solutions of level FΩ′(σ)\sqrt{F_{\Omega}'(\sigma)}. This suggests the following question.

Question 5.3. Find a characterisation of c(Ω)c(\Omega) in terms of EλE_{\lambda} (or of FΩF_{\Omega}) which is monotone under pointwise decrease of the energy. For instance: is

c(Ω)=inf⁡{λ:Eλ(⋅;Ω) has a strict local minimiser},c(\Omega)=\inf\{\lambda:E_{\lambda}(\cdot;\Omega)\text{ has a strict local minimiser}\},

and does a strict local minimiser for Ω\Omega produce one for ΩH\Omega^{H}?

Numerical experiments

Method

We solve (1) in the plane by the method of fundamental solutions. The potential is represented as u(x)=a0+∑jajlog⁡∣x−zj∣u(x)=a_{0}+\sum_{j}a_{j}\log|x-z_{j}|, with source points zjz_{j} placed outside Ω\Omega and inside KK. The Dirichlet conditions u=0u=0 on ∂Ω\partial\Omega and u=1u=1 on ∂K\partial K are imposed in the least-squares sense at 240–320 collocation points. The free boundary is parametrised in polar form, ρ(θ)=s+∑k≤M(akcos⁡kθ+bksin⁡kθ)\rho(\theta)=s+\sum_{k\leq M}(a_{k}\cos k\theta+b_{k}\sin k\theta), M=12M=12–20. For fixed mean radius ss, the equation ∣∇u∣=λ|\nabla u|=\lambda on ∂K\partial K is solved for (ak,bk,λ)(a_{k},b_{k},\lambda) by a Levenberg–Marquardt method. The Bernoulli constant is then the minimum of λ(s)\lambda(s) along the branch. Every domain is normalised to area π\pi, so that c(Ω∗)=ec(\Omega^{*})=e.

Validation. For the unit disc the method returns c=2.718281828459045c=2.718281828459045 at s=0.36787943s=0.36787943, i.e. ee and 1/e1/e to machine precision. Capacities of concentric annuli are reproduced to 10−1410^{-14}. Residuals of the free boundary condition are below 10−510^{-5} in all reported runs, and usually below 10−1010^{-10}.

Ellipses and Steiner symmetrisation

Table 2 shows that cc increases with the aspect ratio of ellipses. Values are stable to 10−910^{-9} between M=12M=12 and M=20M=20. For aspect ratio 3 the polar parametrisation of KK is no longer adequate, and we do not report it. The last column gives the lower bound of Corollary E.

qqc(Ω)c(\Omega)c(Ω)/ec(\Omega)/emax⁡{μ/e,θ}\max\{\mu/e,\sqrt{\theta}\}
12.71828211
1.22.7452821.0099330.991769
1.252.7586751.0148600.987730
1.52.8504621.0486260.960769
23.0954361.1387470.914279

Table 2. Ellipses of area π\pi and aspect ratio qq.

To test the Steiner route we consider Ωf={(x,y):∣y−f(x)∣<0.81−x2}\Omega_f=\{(x,y):|y-f(x)|<0.8\sqrt{1-x^2}\}, whose Steiner symmetral with respect to yy is the ellipse with semi-axes 1 and 0.8. Table 3 shows that symmetrisation decreases cc in all cases, including a non-convex one.

Domainf(x)f(x)c/ec/e
Steiner symmetral (ellipse)01.014860
Mild banana0.25x20.25x^21.024310
S-shape0.4x30.4x^31.032890
Strong banana (non-convex)0.5x20.5x^21.049881

Table 3. Steiner symmetrisation; values normalised to area π\pi.

Second variation

For Ωε\Omega_\varepsilon as in Proposition C, we computed (c(Ωε)−e)/ε2(c(\Omega_\varepsilon)-e)/\varepsilon^2 at ε=0.03\varepsilon=0.03 and ε=0.06\varepsilon=0.06 and Richardson-extrapolated in ε2\varepsilon^2, which is legitimate since the expansion is even in ε\varepsilon. The agreement with the closed formula is about 10−410^{-4} (Table 4).

A capacitary Cheeger inequality fails

Since cap⁡(K;Ω)=∫∂K∣∇uK∣ dσ\operatorname{cap}(K;\Omega)=\int_{\partial K}|\nabla u_K|\,d\sigma, one has sup⁡∂K∣∇uK∣≥cap⁡(K;Ω)/Per⁡(K)\sup_{\partial K}|\nabla u_K|\ge\operatorname{cap}(K;\Omega)/\operatorname{Per}(K), with equality for concentric discs. One might hope to prove (1.3) via inf⁡Kcap⁡(K;Ω)/Per⁡(K)≥c(Ω∗)\inf_K\operatorname{cap}(K;\Omega)/\operatorname{Per}(K)\ge c(\Omega^*). This is false. Oscillating boundaries make Per⁡(K)\operatorname{Per}(K) arbitrarily large while the capacity stays bounded, so the infimum is 0 for every Ω\Omega. In the unit disc, already ρ(θ)=s(1+0.2cos⁡4θ)\rho(\theta)=s(1+0.2\cos4\theta) gives cap⁡/Per⁡=2.5169<e\operatorname{cap}/\operatorname{Per}=2.5169<e, while sup⁡∣∇uK∣=3.9453>e\sup|\nabla u_K|=3.9453>e. The supremum cannot be replaced by a mean.

Testing the Beurling characterisation

In the unit disc, 129 random compact sets KK (five Fourier modes, mean radius in [0.2,0.45][0.2,0.45]) all satisfy sup⁡∂K∣∇uK∣≥3.65>e\sup_{\partial K}|\nabla u_K| \ge3.65 > e. Random perturbations of the critical free set KcK_c, with 40 samples for each amplitude, always increase the supremum. The smallest increments are 0.12, 0.64 and 1.43 for amplitudes 0.002, 0.01 and 0.03 in the disc, and 0.18, 0.64 and 1.67 for the ellipse q=2q=2.

Testing Proposition B on exact solutions

The converse part of the dictionary turns every element of A\mathcal{A} into an exact solution. We sampled ρ∈[0.15,0.8]\rho\in[0.15,0.8] and 2 to 5 modes cnc_n in (2.4), and adjusted c1c_1 to enforce Res⁡eφ=0\operatorname{Res}e^\varphi=0. We computed FF and the areas spectrally on each circle (with 1024 Fourier modes), and checked univalence (simple, positively oriented, nested boundary curves). Among 313 univalent random samples, the minimum of λR/e\lambda R/e is 1.000032. We also ran 24 Nelder–Mead minimisations of λR/e\lambda R/e under the univalence constraint, starting from random configurations. All stopped above 1, the lowest value being 1.0012 near the disc (ρ=0.366\rho=0.366). No violation of (1.6) was found.

The Rellich–Pohozaev identity

For the critical solution in the ellipse q=1.5q=1.5, the two sides of (1.8) (with N=2N=2, i.e. 2λ2∣K∣=∫∂Ω∣∇u∣2x⋅ν dσ2\lambda^{2}|K|=\int_{\partial\Omega}|\nabla u|^{2}x\cdot\nu\,d\sigma) evaluate to 6.72954776011166 and 6.72954776011170. The bound of Theorem D, θ=0.960769\sqrt{\theta}=0.960769, is compared with the computed values of c/ec/e in Table 2.

Concluding remarks and open problems

  1. Sharp constant in Theorem A. The constant 0.914 comes from neglecting ∣K∣|K| and from Jensen’s inequality. A quantitative stability estimate, saying that circle means of ∣F∣2|F|^2 close to m2m^2 force FF to be close to a homothety and hence ∣K∣≈πm2ρ2|K| \approx\pi m^2\rho^2, would improve it. Proving (1.6) would settle (1.3) for convex planar domains.

  1. Higher dimensions. Improve the constant κN\kappa_N of Theorem D, in particular for N=3N=3, where κ3≈0.650\kappa_3 \approx0.650. The boundary term ∫∂Ω∣∇u∣2x⋅ν dσ\int_{\partial\Omega}|\nabla u|^2x\cdot\nu\,d\sigma, which is only estimated through I(Ω)I(\Omega), carries the missing information.

  1. Local minimality. Make Proposition C rigorous along the lines of Remark 3.1. The domain expansion of Section 3 extends to N≥3N\ge3 with spherical harmonics, u1=(αrℓ+βr2−N−ℓ)Yℓu_1=(\alpha r^\ell+\beta r^{2-N-\ell})Y_\ell; it remains to extend it to the pp-Laplacian.

  1. Polarization. Answer the Question of Section 5.

  1. Riemannian setting. Bernoulli problems on manifolds have been studied by Djité and Seck [10] (exterior problem, pp-Laplacian, compact manifolds) and, in the one-phase framework, by Peng, Zhang and Zhu [16]. The argument of [9] only uses the characterisation of Λ\Lambda as a capacity/volume quotient and the Pólya–Szegő inequality, which hold for geodesic symmetrisation on SN\mathbb{S}^{N} and HN\mathbb{H}^{N}. We therefore expect Λ(Ω)≥Λ(B)\Lambda(\Omega) \ge\Lambda(B) for geodesic balls BB of the same volume in these space forms. The classical constant cc is open there as well.

Acknowledgements

Acknowledgements. The authors acknowledge the use of Claude (Anthropic), an AI assistant, for symbolic and numerical verifications, for implementing the numerical experiments and for help in drafting and revising the manuscript. All results, proofs and computations were checked by the authors, who take full responsibility for the content of this paper.

Université Cheikh Anta Diop de Dakar, Sénégal Email address: [email protected]

Université Cheikh Anta Diop de Dakar, Sénégal Email address: [email protected]

References

  1. [1]A. Acker, Uniqueness and monotonicity of solutions for the interior Bernoulli free boundary problem in the convex n-dimensional case, Nonlinear Anal. 13 (1989), 1409–1425.DOI
  2. [3]A. Beurling, On free-boundary problems for the Laplace equation, Seminars on Analytic Functions, vol. 1, Institute for Advanced Study, Princeton, 1957, 248–263.
  3. [4]C. Bianchini and P. Salani, Concavity properties for elliptic free boundary problems, arXiv:0810.4842.arxiv.org/abs/0810.4842
  4. [5]F. Brock and A. Yu. Solynin, An approach to symmetrization via polarization, Trans. Amer. Math. Soc. 352 (2000), 1759–1796.DOI
  5. [6]J. E. Brothers and W. P. Ziemer, Minimal rearrangements of Sobolev functions, J. Reine Angew. Math. 384 (1988), 153–179.DOI
  6. [7]P. Cardaliaguet and R. Tahraoui, Some uniqueness results for Bernoulli interior free-boundary problems in convex domains, Electron. J. Differential Equations 2002, No. 102, 1–16.
  7. [8]M. Dambrine and M. Pierre, About stability of equilibrium shapes, M2AN Math. Model. Numer. Anal. 34 (2000), 811–834.DOI
  8. [9]D. Danners and B. Kawohl, An isoperimetric inequality related to a Bernoulli problem, Calc. Var. Partial Differential Equations 39 (2010), 547–555.
  9. [10]A. S. Djité and D. Seck, Shape optimization and Bernoulli free boundary problems for the Riemannian p-Laplacian, arXiv:2609.00571.arxiv.org/abs/2609.00571
  10. [11]M. Flucher and M. Rumpf, Bernoulli’s free-boundary problem, qualitative theory and numerical approximation, J. Reine Angew. Math. 486 (1997), 165–204.DOI
  11. [12]P. Grisvard, Elliptic Problems in Nonsmooth Domains, Pitman, Boston, 1985.
  12. [13]A. Henrot and H. Shahgholian, Existence of classical solutions to a free boundary problem for the p-Laplace operator. I. The exterior convex case, J. Reine Angew. Math. 521 (2000), 85–97.DOI
  13. [14]A. Henrot and H. Shahgholian, Existence of classical solutions to a free boundary problem for the p-Laplace operator. II. The interior convex case, Indiana Univ. Math. J. 49 (2000), 311–323.DOI
  14. [15]I. Ly and D. Seck, Isoperimetric inequality for an interior free boundary problem with p-Laplacian operator, Electron. J. Differential Equations 2004, No. 109, 1–12.
  15. [16]Yu. Peng, H.-C. Zhang and X.-P. Zhu, One-phase free boundary problems on Riemannian complexes, Tr. Mat. Inst. Steklova 330 (2025).
  16. [17]S. I. Pohozaev, On the eigenfunctions of the equation Δu + λf(u) = 0, Soviet Math. Dokl. 6 (1965), 1408–1411.
  17. [18]G. Pólya and G. Szegő, Isoperimetric Inequalities in Mathematical Physics, Princeton University Press, 1951.
  18. [19]Ch. Pommerenke, Boundary Behaviour of Conformal Maps, Springer, Berlin, 1992.DOI
  19. [20]F. Rellich, Darstellung der Eigenwerte von Δu + λu = 0 durch ein Randintegral, Math. Z. 46 (1940), 635–636.DOI
  20. [21]D. Seck, On an isoperimetric inequality and various methods for the Bernoulli’s free boundary problems, São Paulo J. Math. Sci. 10 (2016), 36–59.DOI

Paper details

vlnc:W_sptdpd96
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude

View statement
Looking for related papers…
Contents