betaFeedback

WorksMathematics

The Escobar Constant for Two-dimensional Disks

Abstract

Let (M,g)(M,g) be a smooth compact Riemannian surface homeomorphic to a closed disk, with boundary length LL. We establish sharp upper bounds for its Escobar constant under upper bounds on the Gaussian curvature. If Kg≤0K_g\leq 0, then I(M,g)≤2/πI(M,g)\leq 2/π, with equality if and only if (M,g)(M,g) is isometric to a Euclidean disk. If Kg≤κK_g\leqκ for some κ>0κ>0, L<2π/κL<2π/\sqrtκ, and Area⁡g(M)<2π/κ\operatorname{Area}_g(M)<2π/κ, then I(M,g)≤4κLarcsin⁡(κL2π)I(M,g)\leq\frac{4}{\sqrtκL}\arcsin(\frac{\sqrtκL}{2π}), with equality if and only if (M,g)(M,g) is isometric to the geodesic disk of the same perimeter contained in an open hemisphere of the sphere of constant curvature κκ. We also discuss the higher order Escobar constant for planar domains. In particular, we construct counterexamples to Hassannezhad and Siffert's conjecture.

Introduction

Let (Mn,g)(M^n,g), (n≥2)(n \ge2) be a compact connected Riemannian manifold with nonempty smooth boundary. Motivated by the search for geometric lower bounds for the first nonzero Steklov eigenvalue, Escobar [4] introduced the isoperimetric constant

I(M):=inf⁡Ω∣∂IΩ∣min⁡{∣∂EΩ∣,∣∂M∖∂EΩ∣}.I(M) := \inf_{\Omega} \frac{|\partial_I\Omega|}{\min\{|\partial_E\Omega|,|\partial M\setminus\partial_E\Omega|\}}.

where ∂IΩ=∂Ω∩M∘\partial_I\Omega= \partial\Omega\cap M^\circ, ∂EΩ=∂Ω∩∂M\partial_E\Omega= \partial\Omega\cap\partial M, and ∣⋅∣|\cdot| denotes (n−1)(n-1)-dimensional Riemannian measure. An admissible set Ω⊂M\Omega\subset M is a relatively open subset with finitely many connected components and piecewise smooth boundary. The above infimum is taken over all admissible sets Ω⊂M\Omega\subset M such that both terms in the denominator are positive. This constant provides a boundary analogue of the classical Cheeger constant. In particular, Escobar proved a Cheeger-type inequality yielding a lower bound for the first Steklov eigenvalue of MM in terms of I(M)I(M) and the first Robin eigenvalue of MM.

In this note, we are concerned with the Escobar constant for two-dimensional Riemannian disks with Gauss curvature bounded from above. For a Riemannian disk MM, this constant admits a particularly simple representation (see Lemma 2.1)

I(M)=inf⁡x,y∈∂Mx≠ydM(x,y)d∂M(x,y).I(M)=\inf_{\substack{x,y\in\partial M\\x\ne y}}\frac{d_M(x,y)}{d_{\partial M}(x,y)}.

where dMd_M is the intrinsic distance on MM and d∂M(x,y)d_{\partial M}(x,y) is the length of the shorter boundary arc joining xx and yy. Thus, in this setting, the Escobar constant is determined entirely by the comparison between distances measured within the disk and along its boundary. In the Euclidean setting, Escobar [5 Theorem 2] proved the sharp inequality

I(W)≤I(D)=2πI(W) \le I(\mathbb{D}) = \frac{2}{\pi}

for bounded planar domains WW with rectifiable boundary, with equality if and only if WW is a round disk. For convex planar domains, the same sharp comparison also follows from the distortion inequality for closed curves; see Kusner–Sullivan [8 Proposition 2.1] for a proof and its attribution to Gromov.

The corresponding Euclidean maximization problem has also been resolved in higher dimensions. Cianchi, Ferone, Nitsch and Trombetti [3 Theorem 1.1] proved that balls minimize the optimal median-normalized boundary trace constant in BVBV among admissible domains in Rn\mathbb{R}^n, for every n≥2n \ge2. For a bounded connected smooth domain W⊂RnW \subset\mathbb{R}^n, this constant, denoted by Cmed(W)C_{\mathrm{med}}(W), is the smallest constant such that

inf⁡c∈R∫∂W∣Tu−c∣ dHn−1≤Cmed(W)∣Du∣(W),u∈BV(W),\inf_{c\in\mathbb{R}} \int_{\partial W} |Tu-c|\,d\mathcal{H}^{n-1} \le C_{\mathrm{med}}(W)|Du|(W), \qquad u \in BV(W),

where TuTu denotes the boundary trace and ∣Du∣(W)|Du|(W) the total variation of uu. Its geometric characterization gives Cmed(W)=I(W)−1C_{\mathrm{med}}(W)=I(W)^{-1}. Consequently, their theorem yields the sharp inequality

I(W)≤I(Bn)=2ωn−1nωn.I(W) \le I(B^n) = \frac{2\omega_{n-1}}{n\omega_n}.

Within the class of smooth domains, equality holds if and only if WW is a Euclidean ball. Here BnB^n is the unit ball in Rn\mathbb{R}^n and ωj\omega_j denotes the volume of the unit ball in Rj\mathbb{R}^j.

We establish two sharp comparison theorems of Escobar constants of two-dimensional disks with Gauss curvature bounded from above.

Theorem 1.1. Let (M,g)(M,g) be a smooth compact Riemannian surface homeomorphic to a closed disk. If Kg≤0K_g \le0, then

I(M)≤2π.I(M) \le\frac{2}{\pi}.

Equality holds if and only if (M,g)(M,g) is isometric to a Euclidean disk.

A corresponding sharp comparison holds when the Gaussian curvature has a positive upper bound, under suitable restrictions on the boundary length and the area.

Theorem 1.2. Let (M,g)(M,g) be a smooth compact Riemannian surface homeomorphic to a closed disk, and set L:=ℓg(∂M)L := \ell_g(\partial M). Suppose that, for some κ>0\kappa> 0,

Kg≤κ,L<2πκ,Area⁡g(M)<2πκ.K_g \le\kappa,\qquad L < \frac{2\pi}{\sqrt{\kappa}},\qquad\operatorname{Area}_g(M) < \frac{2\pi}{\kappa}.

Then

I(M)≤4κLarcsin⁡(κL2π).I(M) \le\frac{4}{\sqrt{\kappa L}}\arcsin\left(\frac{\sqrt{\kappa}L}{2\pi}\right).

The equality holds if and only if (M,g)(M,g) is isometric to the closed geodesic disk of radius

rκ,L:=1κarcsin⁡(κL2π)<π2κr_{\kappa,L} := \frac{1}{\sqrt{\kappa}}\arcsin\left(\frac{\sqrt{\kappa}L}{2\pi}\right) < \frac{\pi}{2\sqrt{\kappa}}

in the round sphere of constant Gaussian curvature κ\kappa.

It is convenient to introduce a boundary partition isoperimetric profile. Let L=ℓg(∂M)L = \ell_g(\partial M). For an admissible set Ω⊂M\Omega\subset M, write ΣΩ=∂IΩ,ΓΩ=∂EΩ\Sigma_\Omega= \partial_I\Omega,\Gamma_\Omega= \partial_E\Omega for the interior and exterior parts of its boundary, respectively. For 0<s<L0 < s < L, define

PM(s)=inf⁡{ℓg(ΣΩ):Ω⊂M is admissible, ℓg(ΓΩ)=s}.\mathcal{P}_M(s) = \inf\{\ell_g(\Sigma_\Omega) : \Omega\subset M\text{ is admissible},\ \ell_g(\Gamma_\Omega)=s\}.

Taking complements gives PM(s)=PM(L−s)\mathcal{P}_M(s) = \mathcal{P}_M(L-s), and it follows from the definition that

I(M,g)=inf⁡0<s≤L/2PM(s)s.I(M,g) = \inf_{0<s\leq L/2}\frac{\mathcal{P}_M(s)}{s}.

Theorem 1.3. Let (M,g)(M,g) be a smooth compact Riemannian surface homeomorphic to a closed disk with Kg≤0K_g \leq0. Then

PM(s)≤Lπsin⁡(πsL),0<s<L,\mathcal{P}_M(s) \leq\frac{L}{\pi}\sin\left(\frac{\pi s}{L}\right),\qquad0 < s < L,

where L=ℓg(∂M)L = \ell_g(\partial M). Moreover, the equality at s=L/2s = L/2 holds if and only if (M,g)(M,g) is isometric to the Euclidean disk of perimeter LL.

The boundary partition profiles of Euclidean and hyperbolic disks can be computed explicitly. Let D0,LD_{0,L} denote the Euclidean disk of perimeter LL, and let Dκ,LD_{\kappa,L} denote the geodesic disk of perimeter LL in the hyperbolic plane of constant curvature κ=−a2<0\kappa= -a^2 < 0. A direct computation gives

PD0,L(s)=Lπsin⁡(πsL),0<s<L,(1)\mathcal{P}_{D_{0,L}}(s) = \frac{L}{\pi}\sin\left(\frac{\pi s}{L}\right),\qquad0 < s < L, \tag*{(1)}

and

PDκ,L(s)=2aarsinh⁡(aL2πsin⁡(πsL)),0<s<L.\mathcal{P}_{D_{\kappa,L}}(s) = \frac{2}{a}\operatorname{arsinh}\left(\frac{aL}{2\pi}\sin\left(\frac{\pi s}{L}\right)\right),\qquad0 < s < L.

Since P(s)/s\mathcal{P}(s)/s is nonincreasing on (0,L/2](0,L/2], we obtain

I(D0,L)=2π,I(Dκ,L)=4aLarsinh⁡(aL2π)<2π.I(D_{0,L})=\frac{2}{\pi},\qquad I(D_{\kappa,L})=\frac{4}{aL}\operatorname{arsinh}\left(\frac{aL}{2\pi}\right)<\frac{2}{\pi}.

For every fixed κ<0\kappa<0, however,

lim⁡L→0I(Dκ,L)=2π.\lim_{L\to0}I(D_{\kappa,L})=\frac{2}{\pi}.

Consequently, even under the stronger curvature assumption Kg≤κ<0K_g\leq\kappa<0, no upper bound strictly smaller than 2/π2/\pi can hold independently of the perimeter. Thus the constant 2/π2/\pi remains optimal in this class, with optimality witnessed by shrinking hyperbolic disks.

We also consider higher order Escobar constants, introduced by Hassannezhad and Miclo [6]. For an integer k≥2k\geq2, these are defined by

Ik(M)=inf⁡(Ω1,…,Ωk)∈Ak(M)max⁡1≤j≤kℓg(∂IΩj)ℓg(∂EΩj).I_k(M)=\inf_{(\Omega_1,\ldots,\Omega_k)\in\mathcal{A}_k(M)}\max_{1\leq j\leq k}\frac{\ell_g(\partial_I\Omega_j)}{\ell_g(\partial_E\Omega_j)}.

where Ak(M)\mathcal{A}_k(M) consists of kk-tuples of pairwise disjoint nonempty admissible sets, which are not required to cover MM. The case k=2k=2 recovers the Escobar constant I(M)I(M). Hassannezhad and Siffert [7] conjectured that the Euclidean disk maximizes IkI_k among bounded planar domains with rectifiable boundary for every k≥3k\geq3.

We disprove this conjecture by constructing counterexamples among smooth strictly convex planar domains. More precisely, let D\mathbb{D} be the Euclidean unit disk and set Cε=diag⁡(eε,e−ε)DC_\varepsilon=\operatorname{diag}(e^\varepsilon,e^{-\varepsilon})\mathbb{D}. For every fixed integer k≥3k\geq3, we show that

Ik(Cε)>Ik(D)=kπsin⁡(πk)I_k(C_\varepsilon)>I_k(\mathbb{D})=\frac{k}{\pi}\sin\left(\frac{\pi}{k}\right)

for all sufficiently small nonzero ε\varepsilon; see Proposition 5.2. Thus arbitrarily small elliptic deformations of the disk increase each fixed higher order Escobar constant, in contrast to the maximizing property of the disk for I2I_2.

The paper is organized as follows. Section 2 gives a detailed proof of the distance representation of the Escobar constant. Section 3 recalls the necessary background on CAT(κ\kappa) spaces and Reshetnyak’s majorization theorem. Section 4 proves the main comparison theorems. Finally, Section 5 discusses higher-order Escobar constants, gives a short proof of the formula of Hassannezhad and Siffert for Euclidean disks, and constructs counterexamples to their conjecture.

Distance representation of the Escobar constant

Throughout this section, (M,g)(M,g) is a smooth compact Riemannian disk, and L:=ℓg(∂M)L:=\ell_g(\partial M). We denote by dMd_M the intrinsic length distance on MM and by d∂M(x,y)d_{\partial M}(x,y) the length of the shorter boundary arc joining x,y∈∂Mx,y\in\partial M.

The following observation reduces the Escobar constant of a disk to a comparison of interior and boundary distances. It was already stated by Escobar for planar domains [5 p. 102, Eq. (3)]. For completeness, we give a detailed proof in the setting of smooth Riemannian disks, allowing admissible sets and their boundary traces to be disconnected. We also record the corresponding estimate for the boundary partition profile, which will be used in the comparison arguments.

Lemma 2.1. For every smooth compact Riemannian disk (M,g)(M,g),

I(M)=inf⁡x,y∈∂Mx≠ydM(x,y)d∂M(x,y).(2)I(M)=\inf_{\substack{x,y\in\partial M\\x\ne y}}\frac{d_M(x,y)}{d_{\partial M}(x,y)}. \tag*{(2)}

Moreover, if a boundary arc of length s∈(0,L)s\in(0,L) has endpoints x,yx,y, then

PM(s)≤dM(x,y).(3)\mathcal{P}_M(s)\le d_M(x,y). \tag*{(3)}

Proof. Fix distinct points x,y∈∂Mx,y\in\partial M. For every ε>0\varepsilon>0, there exists a properly embedded piecewise smooth arc α\alpha joining xx to yy, with α∖{x,y}⊂M∘\alpha\setminus\{x,y\}\subset M^\circ, such that

ℓg(α)≤dM(x,y)+ε.\ell_g(\alpha)\le d_M(x,y)+\varepsilon.

Indeed, start with a curve of length less than dM(x,y)+ε/2d_M(x,y)+\varepsilon/2, push its boundary portions slightly into a smooth collar, and then remove loops and smooth corners. The endpoints remain fixed, and the additional length can be made smaller than ε/2\varepsilon/2.

The arc α\alpha separates MM into two components. The component adjacent to the prescribed boundary arc has boundary trace of length ss. Letting ε↓0\varepsilon\downarrow0 proves (2.2). Using the component adjacent to a shorter boundary arc also gives

I(M)≤dM(x,y)d∂M(x,y).I(M)\le\frac{d_M(x,y)}{d_{\partial M}(x,y)}.

For the reverse inequality, set

cM:=inf⁡x,y∈∂Mx≠ydM(x,y)d∂M(x,y).c_M:=\inf_{\substack{x,y\in\partial M\\x\ne y}}\frac{d_M(x,y)}{d_{\partial M}(x,y)}.

First consider an admissible set Ω\Omega for which ΣΩ‾\overline{\Sigma_\Omega} is a finite disjoint union of piecewise smooth embedded arcs meeting ∂M\partial M transversely and simple closed curves in M∘M^\circ. Let α1,…,αN\alpha_1,\ldots,\alpha_N be the arc components, with endpoint pairs (xj,yj)(x_j,y_j). Choose a shorter boundary arc AjA_j joining xjx_j to yjy_j, and put

aj:=ℓg(Aj)=d∂M(xj,yj).a_j:=\ell_g(A_j)=d_{\partial M}(x_j,y_j).

The characteristic functions of ΓΩ\Gamma_\Omega and of the symmetric difference A1△⋯△ANA_1\triangle\cdots\triangle A_N have the same jumps modulo two around the boundary circle. Consequently, this symmetric difference agrees, up to endpoints, with either ΓΩ\Gamma_{\Omega} or its complement. Since the length of a symmetric difference is at most the sum of the lengths, we obtain

min⁡{ℓg(ΓΩ),L−ℓg(ΓΩ)}≤∑j=1Naj.\min\{\ell_g(\Gamma_{\Omega}),L-\ell_g(\Gamma_{\Omega})\}\leq\sum_{j=1}^{N}a_j.

Closed components of ΣΩ\Sigma_{\Omega} only add to its length. Therefore,

ℓg(ΣΩ)≥∑j=1Nℓg(αj)≥∑j=1NdM(xj,yj)≥cM∑j=1Naj≥cMmin⁡{ℓg(ΓΩ),L−ℓg(ΓΩ)}.\begin{aligned} \ell_g(\Sigma_{\Omega})\geq\sum_{j=1}^{N}\ell_g(\alpha_j)\geq\sum_{j=1}^{N}d_M(x_j,y_j)\geq c_M\sum_{j=1}^{N}a_j \\ &\geq c_M\min\{\ell_g(\Gamma_{\Omega}),L-\ell_g(\Gamma_{\Omega})\}. \end{aligned}

For a general piecewise smooth admissible set, local smoothing and transverse perturbations in a boundary collar give regular sets Ωm\Omega_m such that

ℓg(ΣΩm)⟶ℓg(ΣΩ),1ΓΩm⟶1ΓΩin L1(∂M).\ell_g(\Sigma_{\Omega_m})\longrightarrow\ell_g(\Sigma_{\Omega}),\qquad\mathbf{1}_{\Gamma_{\Omega_m}}\longrightarrow\mathbf{1}_{\Gamma_{\Omega}}\quad\text{in }L^1(\partial M).

The preceding inequality therefore passes to the limit. Taking the infimum over Ω\Omega proves (2.1). □

CAT(κ\kappa) spaces and majorization

In this section, we collect the metric-geometric facts used in the comparison proofs, beginning with a brief review of CAT(κ\kappa) spaces.

For κ∈R\kappa\in\mathbb{R}, let Mκ2\mathbb{M}^2_{\kappa} denote the complete simply connected surface of constant curvature κ\kappa, and set

Dκ:={π/κ,κ>0,+∞,κ≤0.D_{\kappa}:=\begin{cases} \pi/\sqrt{\kappa}, & \kappa>0,\\ +\infty, & \kappa\leq0. \end{cases}

Thus M02=R2\mathbb{M}^2_0=\mathbb{R}^2, while for κ>0\kappa>0,

Mκ2=SR2,R=1κ.\mathbb{M}^2_{\kappa}=S^2_R,\qquad R=\frac{1}{\sqrt{\kappa}}.

A geodesic metric space XX is called CAT(κ\kappa) if every geodesic triangle of perimeter less than 2Dκ2D_{\kappa} satisfies the following comparison: the distance between any two points on its sides is at most the distance between the corresponding points on the comparison triangle in Mκ2\mathbb{M}^2_{\kappa}. The corresponding points are determined by their distances along the sides. For κ≤0\kappa\leq0, the perimeter restriction is vacuous. A space is locally CAT(κ\kappa) if every point has a neighborhood satisfying this comparison.

Under the hypotheses of Theorem 1.1 and Theorem 1.2, the Riemannian disk MM is a CAT(κ\kappa) space.

Lemma 3.1. Let (M,g)(M,g) be a smooth compact Riemannian disk.

  1. If Kg≤0K_g \le0, then (M,dM)(M,d_M) is CAT(0).

  1. If κ>0\kappa> 0, Kg≤κK_g \le\kappa, ℓ(∂M)<2π/κ\ell(\partial M) < 2\pi/\sqrt{\kappa}, and Area⁡g(M)<2π/κ\operatorname{Area}_g(M) < 2\pi/\kappa, then (M,dM)(M,d_M) is CAT(κ\kappa).

Proof. For the first statement, equip MM with its intrinsic length metric. The characterization theorem of Alexander–Berg–Bishop [1] implies that MM is locally CAT(0), including at boundary points. In dimension two, the additional conditions on sectional curvatures of the boundary are vacuous, since the boundary is one-dimensional. By the Cartan–Hadamard theorem [1 Corollary 1], MM is CAT(0).

For the second statement, Bowditch [2 Lemma 4.1.3] proved that if a disk MM is locally CAT(1) and ℓ(∂M)<2π\ell(\partial M) < 2\pi, then MM is globally CAT(1) if and only if Area⁡(M)<2π\operatorname{Area}(M) < 2\pi. The general case follows by rescaling. □

A key tool we use in the proof is the following form of Reshetnyak’s majorization theorem [11]; see also [10 Section 3.2 and Section 17].

Theorem 3.2. Let XX be a complete CAT(κ\kappa) space, and let γ:R/LZ→X\gamma: \mathbb{R}/L\mathbb{Z} \to X be an arclength parametrization of a rectifiable Jordan curve. Assume that L<2DκL < 2D_\kappa. Then there exist a compact geodesically convex domain C⊂Mκ2C \subset\mathbb{M}^2_\kappa, an arclength parametrization c:R/LZ→∂Cc : \mathbb{R}/L\mathbb{Z} \to\partial C, and a 1-Lipschitz map F:C→XF : C \to X such that F∘c=γF \circ c = \gamma. In particular, F∣∂CF|_{\partial C} is an arclength-preserving homeomorphism onto the image of γ\gamma, and ℓ(∂C)=L\ell(\partial C) = L. When κ>0\kappa> 0, the domain CC lies in an open hemisphere of Mκ2\mathbb{M}^2_\kappa.

Here the metric on CC is the restriction of the model distance, which agrees with its intrinsic length metric by convexity.

Lemma 3.3. Let (M,g)(M,g) be a smooth compact Riemannian disk, and set L=ℓg(∂M)L = \ell_g(\partial M). Assume either

κ=0,Kg≤0,\kappa= 0,\qquad K_g \le0,

or

κ>0,Kg≤κ,Area⁡g(M)<2πκ,L<2πκ.\kappa> 0,\qquad K_g \le\kappa,\qquad\operatorname{Area}_g(M) < \frac{2\pi}{\kappa},\qquad L < \frac{2\pi}{\sqrt{\kappa}}.

Then there exist a compact geodesically convex domain C⊂Mκ2C \subset\mathbb{M}^2_\kappa and a surjective 1-Lipschitz map F:C→MF : C \to M whose boundary restriction is an arclength-preserving homeomorphism

F∣∂C:∂C→∂M.F|_{\partial C} : \partial C \to\partial M.

In particular, ℓ(∂C)=L\ell(\partial C) = L.

Proof. By Lemma 3.1, (M,dM)(M,d_M) is CAT(κ\kappa). Apply Theorem 3.2 to an arclength parametrization of ∂M\partial M. The length condition is automatic when κ=0\kappa= 0.

It remains to prove that FF is surjective. Every boundary point lies in its image. If an interior point q∈Mq \in M were omitted, then the boundary homeomorphism would extend to a continuous map F:C→M∖{q}F : C \to M \setminus\{q\}. This would make the boundary circle null-homotopic in a punctured disk, which is impossible. Therefore FF is surjective. □

Proofs of the comparison theorems

We first establish two auxiliary facts used in both curvature regimes: a mean-chord inequality in Euclidean space and a rigidity statement for maps preserving the lengths of diameters. The following inequality can be found in [9].

Lemma 4.1. Let m≥2m \ge2, and let c:R/LZ⟶Rmc : \mathbb{R}/L\mathbb{Z} \longrightarrow\mathbb{R}^{m} be a closed Lipschitz curve parametrized by arclength. For every 0<s≤L/20 < s \le L/2,

1L∫0L∣c(t+s)−c(t)∣2 dt≤L2π2sin⁡2(πsL).(4)\frac{1}{L}\int_{0}^{L} \lvert c(t+s)-c(t)\rvert^{2}\,\mathrm{d}t \le\frac{L^{2}}{\pi^{2}}\sin^{2}\left(\frac{\pi s}{L}\right). \tag*{(4)}

Equality at s=L/2s = L/2 holds if and only if cc parametrizes a round circle in an affine two-dimensional plane once at unit speed.

Proof. Rescale the curve and its parameter so that L=2πL = 2\pi. Write its vector-valued Fourier expansion as

c(t)=∑n∈Zc^neint,c^n∈Cm.c(t)=\sum_{n\in\mathbb{Z}}\widehat{c}_{n}e^{\mathrm{i}nt},\qquad\widehat{c}_{n}\in\mathbb{C}^{m}.

Since ∣c′(t)∣=1\lvert c'(t)\rvert= 1 almost everywhere, Parseval’s identity gives

∑n∈Zn2∣c^n∣2=12π∫02π∣c′(t)∣2 dt=1.\sum_{n\in\mathbb{Z}}n^{2}\lvert\widehat{c}_{n}\rvert^{2}=\frac{1}{2\pi}\int_{0}^{2\pi}\lvert c'(t)\rvert^{2}\,\mathrm{d}t=1.

For 0<a≤π0 < a \le\pi, we therefore obtain

12π∫02π∣c(t+a)−c(t)∣2 dt=4∑n∈Zsin⁡2(na2)∣c^n∣2\frac{1}{2\pi}\int_{0}^{2\pi}\lvert c(t+a)-c(t)\rvert^{2}\,\mathrm{d}t=4\sum_{n\in\mathbb{Z}}\sin^{2}\left(\frac{na}{2}\right)\lvert\widehat{c}_{n}\rvert^{2}
≤4sin⁡2(a2)∑n∈Zn2∣c^n∣2=4sin⁡2(a2).\begin{aligned} \le4\sin^{2}\left(\frac{a}{2}\right)\sum_{n\in\mathbb{Z}}n^{2}\lvert\widehat{c}_{n}\rvert^{2} \\ &=4\sin^{2}\left(\frac{a}{2}\right). \end{aligned}

Here we used

∣sin⁡(nx)∣≤∣n∣sin⁡x,n∈Z,0≤x≤π/2.\lvert\sin(nx)\rvert\le\lvert n\rvert\sin x,\qquad n\in\mathbb{Z},\quad0\le x\le\pi/2.

Rescaling proves (4.1).

Suppose equality holds at a=πa=\pi. For every ∣n∣≥2\lvert n\rvert\ge2,

sin⁡2(nπ/2)<n2,\sin^{2}(n\pi/2)<n^{2},

so c^n=0\widehat{c}_{n}=0 for all ∣n∣≥2\lvert n\rvert\ge2. Since cc is real-valued,

c(t)=c0+Acos⁡t+Bsin⁡tc(t)=c_{0}+A\cos t+B\sin t

for vectors A,B∈RmA,B\in\mathbb{R}^{m}. The unit-speed condition implies

∣A∣=∣B∣=1,⟨A,B⟩=0.\lvert A\rvert=\lvert B\rvert=1,\qquad\langle A,B\rangle=0.

Thus cc parametrizes a unit circle in an affine two-dimensional plane. The converse follows by direct calculation. □

Lemma 4.2. Let κ≥0\kappa\ge0, let XX be a CAT(κ)(\kappa) space, and let B=B‾Mκ2(o,r)B = \overline{B}_{\mathbb{M}^2_\kappa}(o,r), where r>0r > 0 and, when κ>0\kappa> 0, r<Dκ/2r < D_\kappa/2. For x∈Bx \in B, denote by x∗x^* the point on the opposite radial segment with

dMκ2(o,x∗)=dMκ2(o,x).d_{\mathbb{M}^2_\kappa}(o,x^*) = d_{\mathbb{M}^2_\kappa}(o,x).

Let F:B→XF:B \to X be 11-Lipschitz. If

dX(F(u),F(u∗))=2rfor every u∈∂B,(5)d_X(F(u),F(u^*)) = 2r \qquad\text{for every } u \in\partial B, \tag*{(5)}

then FF is an isometric embedding.

Proof. Write δ\delta for the distance in Mκ2\mathbb{M}^2_\kappa. The 11-Lipschitz property and (4.2) imply that FF preserves all distances along each diameter: the distance between its endpoints equals the length of the diameter, so no subsegment can be shortened. Consequently,

dX(F(o),F(x))=δ(o,x),dX(F(x),F(x∗))=2δ(o,x),(6)d_X(F(o),F(x)) = \delta(o,x), \qquad d_X(F(x),F(x^*)) = 2\delta(o,x), \tag*{(6)}

and F(o)F(o) is the midpoint of the geodesic joining F(x)F(x) to F(x∗)F(x^*).

First suppose κ=0\kappa= 0, and identify oo with the origin. For x,y∈Bx,y \in B with x≠0x \ne0, the CAT(0)(0) midpoint inequality gives

∣y∣2=dX(F(o),F(y))2≤12dX(F(x),F(y))2+12dX(F(−x),F(y))2−14dX(F(x),F(−x))2≤12∣x−y∣2+12∣x+y∣2−∣x∣2=∣y∣2.\begin{aligned} |y|^2 &= d_X(F(o),F(y))^2 \\ &\le\frac{1}{2}d_X(F(x),F(y))^2 + \frac{1}{2}d_X(F(-x),F(y))^2 - \frac{1}{4}d_X(F(x),F(-x))^2 \\ &\le\frac{1}{2}|x-y|^2 + \frac{1}{2}|x+y|^2 - |x|^2 \\ &= |y|^2. \end{aligned}

Every inequality is an equality. In particular, dX(F(x),F(y))=∣x−y∣d_X(F(x),F(y)) = |x-y|. The case x=0x = 0 follows from (4.3).

Now suppose κ>0\kappa> 0, and set R=1/κR = 1/\sqrt{\kappa}. Fix x,y∈Bx,y \in B with x≠ox \ne o, and write

ρ:=δ(o,x),τ:=δ(o,y),a:=δ(x,y),b:=δ(x∗,y).\rho:= \delta(o,x), \qquad\tau:= \delta(o,y), \qquad a := \delta(x,y), \qquad b := \delta(x^*,y).

The spherical midpoint identity gives

cos⁡(a/R)+cos⁡(b/R)=2cos⁡(ρ/R)cos⁡(τ/R)>0.(7)\cos(a/R) + \cos(b/R) = 2\cos(\rho/R)\cos(\tau/R) > 0. \tag*{(7)}

Since a,b<πRa,b < \pi R, this implies a+b<πRa+b < \pi R. Also, 2ρ<πR2\rho< \pi R. Thus the triangle with vertices x,x∗,yx,x^*,y has perimeter

a+b+2ρ<2πR.a+b+2\rho< 2\pi R.

The corresponding triangle in XX has no larger perimeter, so the CAT(κ)(\kappa) midpoint comparison applies. Using (4.3), (4.4), and cos⁡(ρ/R)>0\cos(\rho/R)>0, we obtain

cos⁡(τ/R)=cos⁡(a/R)+cos⁡(b/R)2cos⁡(ρ/R)≤cos⁡(dX(F(x),F(y))/R)+cos⁡(dX(F(x∗),F(y))/R)2cos⁡(ρ/R)≤cos⁡(dX(F(o),F(y))/R)=cos⁡(τ/R).\cos(\tau/R)=\frac{\cos(a/R)+\cos(b/R)}{2\cos(\rho/R)} \leq\frac{\cos(d_X(F(x),F(y))/R)+\cos(d_X(F(x^*),F(y))/R)}{2\cos(\rho/R)} \leq\cos(d_X(F(o),F(y))/R) =\cos(\tau/R).

The first inequality follows from the 1-Lipschitz property and the monotonicity of cosine on [0,π][0,\pi]. The second is the CAT(κ\kappa) midpoint inequality, obtained by applying the spherical midpoint identity to the comparison triangle.

All inequalities are therefore equalities. Each of the two terms in the numerator can only have increased in the first inequality, so equality forces

dX(F(x),F(y))=a=δ(x,y).d_X(F(x),F(y))=a=\delta(x,y).

Again, the case x=ox=o follows from (4.3). Thus FF preserves all distances. □

Proof of Theorems 1.1 and 1.3. Assume Kg≤0K_g\leq0. Let C⊂R2C\subset\mathbb{R}^2 and F:C→MF:C\to M be supplied by Lemma 3.3. Choose an arclength parametrization

c:R/LZ⟶∂C,c:\mathbb{R}/L\mathbb{Z}\longrightarrow\partial C,

and set γ:=F∘c\gamma:=F\circ c. Then γ\gamma is an arclength parametrization of ∂M\partial M, and

dM(γ(t+s),γ(t))≤∣c(t+s)−c(t)∣.(8)d_M(\gamma(t+s),\gamma(t))\leq|c(t+s)-c(t)|. \tag*{(8)}

Fix 0<s≤L/20<s\leq L/2. Lemma 4.1 implies that some t0∈R/LZt_0\in\mathbb{R}/L\mathbb{Z} satisfies

∣c(t0+s)−c(t0)∣≤Lπsin⁡(πsL).|c(t_0+s)-c(t_0)|\leq\frac{L}{\pi}\sin\left(\frac{\pi s}{L}\right).

The boundary arc γ([t0,t0+s])\gamma([t_0,t_0+s]) has length ss. By Lemma 2.1 and (4.5),

PM(s)≤dM(γ(t0+s),γ(t0))≤∣c(t0+s)−c(t0)∣≤Lπsin⁡(πsL).\mathcal{P}_M(s)\leq d_M(\gamma(t_0+s),\gamma(t_0)) \leq|c(t_0+s)-c(t_0)| \leq\frac{L}{\pi}\sin\left(\frac{\pi s}{L}\right).

Symmetry gives the same estimate for L/2<s<LL/2<s<L. Taking s=L/2s=L/2 yields

I(M,g)≤2LPM(L/2)≤2π.(9)I(M,g)\leq\frac{2}{L}\mathcal{P}_M(L/2)\leq\frac{2}{\pi}. \tag*{(9)}

Suppose PM(L/2)=Lπ\mathcal{P}_M(L/2)=\frac{L}{\pi}. For every tt, Lemma 2.1 gives

Lπ≤dM(γ(t+L/2),γ(t))≤∣c(t+L/2)−c(t)∣.\frac{L}{\pi}\le d_M(\gamma(t+L/2),\gamma(t))\le\lvert c(t+L/2)-c(t)\rvert.

On the other hand,

1L∫0L∣c(t+L/2)−c(t)∣2 dt≤L2π2.\frac{1}{L}\int_0^L\lvert c(t+L/2)-c(t)\rvert^2\,\mathrm{d}t\le\frac{L^2}{\pi^2}.

by Lemma 4.1. Hence equality holds in the mean-chord inequality. Its equality statement shows that CC is a Euclidean disk of radius r0=L2πr_0=\frac{L}{2\pi}. The preceding pointwise inequalities now give

dM(F(u),F(u∗))=2r0for every u∈∂C.d_M(F(u),F(u^*))=2r_0\qquad\text{for every }u\in\partial C.

Lemma 4.2 shows that FF is an isometric embedding. It is surjective by Lemma 3.3, so (M,g)(M,g) is isometric to that Euclidean disk.

If I(M,g)=2/πI(M,g)=2/\pi, then (9) forces PM(L/2)=Lπ\mathcal{P}_M(L/2)=\frac{L}{\pi}. The preceding argument therefore applies. Conversely, (1) shows that a Euclidean disk attains equality in both comparisons.

Proof of Theorem 1.2. Assume that κ>0\kappa>0 and

Kg≤κ,L<2πκ,Area⁡g(M)<2πκ.K_g\le\kappa,\qquad L<\frac{2\pi}{\sqrt{\kappa}},\qquad\operatorname{Area}_g(M)<\frac{2\pi}{\kappa}.

Set

R:=1κ,q:=L2πR∈(0,1),r:=Rarcsin⁡q∈(0,πR/2).R:=\frac{1}{\sqrt{\kappa}},\qquad q:=\frac{L}{2\pi R}\in(0,1),\qquad r:=R\arcsin q\in(0,\pi R/2).

Let C⊂SR2C\subset S_R^2 and F:C→MF:C\to M be supplied by Lemma 3.3. Choose an arclength parametrization

c:R/LZ⟶∂C⊂SR2⊂R3,c:\mathbb{R}/L\mathbb{Z}\longrightarrow\partial C\subset S_R^2\subset\mathbb{R}^3,

and set γ:=F∘c\gamma:=F\circ c.

The comparison estimate. Viewed as a curve in R3\mathbb{R}^3, cc is still parametrized by arclength. For every 0<s≤L/20<s\le L/2, Lemma 4.1 gives some t0t_0 such that

∣c(t0+s)−c(t0)∣≤Lπsin⁡(πsL).\lvert c(t_0+s)-c(t_0)\rvert\le\frac{L}{\pi}\sin\left(\frac{\pi s}{L}\right).

For u,v∈SR2u,v\in S_R^2, the spherical distance is related to the Euclidean chord length by

dSR2(u,v)=2Rarcsin⁡(∣u−v∣2R).d_{S_R^2}(u,v)=2R\arcsin\left(\frac{\lvert u-v\rvert}{2R}\right).

Using Lemma 2.1 and the 1-Lipschitz property of FF, we obtain

PM(s)≤dM(γ(t0+s),γ(t0))≤dSR2(c(t0+s),c(t0))≤2Rarcsin⁡(qsin⁡(πsL)).\begin{aligned} \mathcal{P}_M(s) &\le d_M(\gamma(t_0+s),\gamma(t_0))\\ &\le d_{S_R^2}(c(t_0+s),c(t_0))\\ &\le2R\arcsin\left(q\sin\left(\frac{\pi s}{L}\right)\right). \end{aligned}

By symmetry, this estimate holds for every 0<s<L0<s<L. Thus

PM(s)≤2Rarcsin⁡(qsin⁡(πsL)),0<s<L.\mathcal{P}_M(s)\le2R\arcsin\left(q\sin\left(\frac{\pi s}{L}\right)\right),\qquad0<s<L.

In particular,

I(M,g)≤2LPM(L/2)≤4rL=4κLarcsin⁡(κL2π).(10)I(M,g)\le\frac{2}{L}\mathcal{P}_M(L/2)\le\frac{4r}{L}=\frac{4}{\sqrt{\kappa}L}\arcsin\left(\frac{\sqrt{\kappa}L}{2\pi}\right). \tag*{(10)}

Suppose equality holds in the Escobar comparison. Then (10) implies PM(L/2)=2r\mathcal{P}_M(L/2)=2r. For every tt, Lemma 2.1 gives

2r≤dM(γ(t+L/2),γ(t))≤dSR2(c(t+L/2),c(t)).2r\le d_M(\gamma(t+L/2),\gamma(t))\le d_{S_R^2}(c(t+L/2),c(t)).

Since spherical distance lies in [0,πR][0,\pi R], the chord-length formula implies

∣c(t+L/2)−c(t)∣≥2Rsin⁡(r/R)=2Rq=Lπ.\left|c(t+L/2)-c(t)\right|\ge2R\sin(r/R)=2Rq=\frac{L}{\pi}.

Meanwhile, Lemma 4.1 gives

1L∫0L∣c(t+L/2)−c(t)∣2 dt≤L2π2.\frac{1}{L}\int_0^L\left|c(t+L/2)-c(t)\right|^2\,dt\le\frac{L^2}{\pi^2}.

Thus equality holds in the mean-chord inequality, and cc parametrizes a Euclidean round circle in an affine plane in R3\mathbb{R}^3.

This circle lies on SR2S_R^2 and has Euclidean radius L/(2π)=Rq<RL/(2\pi)=Rq<R. Since CC lies in an open hemisphere, it is the smaller spherical disk bounded by this circle. Consequently, C=B‾SR2(o,r)C=\overline{B}_{S_R^2}(o,r) for some o∈SR2o\in S_R^2.

Points whose boundary parameters differ by L/2L/2 are now the opposite endpoints of a diameter of this spherical disk. Their spherical distance is 2r2r, so the preceding inequalities give

dM(F(u),F(u∗))=2rfor every u∈∂C.d_M(F(u),F(u^*))=2r\qquad\text{for every }u\in\partial C.

Lemma 4.2 shows that FF is an isometric embedding. Surjectivity then proves that (M,g)(M,g) is isometric to the spherical disk of radius rr.

Conversely, let

B=B‾SR2(o,r),0<r<πR/2.B=\overline{B}_{S_R^2}(o,r),\qquad0<r<\pi R/2.

Its perimeter is L=2πRsin⁡(r/R)L = 2\pi R\sin(r/R), so q=sin⁡(r/R)q = \sin(r/R). Because BB is geodesically convex, two boundary points separated by a shorter boundary arc of length s∈(0,L/2]s \in(0,L/2] have intrinsic distance

H(s):=2Rarcsin⁡(qsin⁡(πsL)).H(s) := 2R\arcsin\left(q\sin\left(\frac{\pi s}{L}\right)\right).

To determine the minimum of H(s)/sH(s)/s, put

f(x):=arcsin⁡(qsin⁡x),0≤x≤π/2.f(x) := \arcsin(q\sin x), \qquad0 \le x \le\pi/2.

Then f(0)=0f(0) = 0 and

f′′(x)=−q(1−q2)sin⁡x(1−q2sin⁡2x)3/2≤0.f''(x) = -\frac{q(1-q^2)\sin x}{(1-q^2\sin^2 x)^{3/2}} \le0.

Hence f(x)/xf(x)/x is nonincreasing for x>0x > 0, and therefore H(s)/sH(s)/s is nonincreasing on (0,L/2](0,L/2]. Lemma 2.1 now gives

I(B)=inf⁡0<s≤L/2H(s)s=H(L/2)L/2=4rL.I(B) = \inf_{0<s\le L/2}\frac{H(s)}{s} = \frac{H(L/2)}{L/2} = \frac{4r}{L}.

Thus the spherical disk attains equality. □

Higher order Escobar constants

We conclude by considering higher order Escobar constants. These constants were introduced by Hassannezhad–Miclo [6] in connection with lower bounds for higher Steklov eigenvalues, and were subsequently studied for planar domains by Hassannezhad and Siffert [7].

Let (M,g)(M,g) be a compact Riemannian surface with boundary. For a nonempty admissible set Ω⊂M\Omega\subset M, retain the notation ΣΩ,ΓΩ\Sigma_\Omega,\Gamma_\Omega and define

ηM(Ω):=ℓg(ΣΩ)ℓg(ΓΩ).\eta_M(\Omega) := \frac{\ell_g(\Sigma_\Omega)}{\ell_g(\Gamma_\Omega)}.

A quotient with vanishing denominator is interpreted as +∞+\infty.

For an integer k≥1k \ge1, let Ak(M)\mathcal{A}_k(M) be the family of kk-tuples (Ω1,…,Ωk)(\Omega_1,\ldots,\Omega_k) of nonempty mutually disjoint admissible sets. The sets are not required to cover MM, and their closures may meet. Each Ωj\Omega_j may have several connected components; thus kk counts the number of sets in the tuple, rather than the total number of their connected components. Both boundary lengths in ηM(Ωj)\eta_M(\Omega_j) include all components of Ωj\Omega_j.

The kk-th Escobar constant is

Ik(M):=inf⁡(Ω1,…,Ωk)∈Ak(M)max⁡1≤j≤kηM(Ωj).(11)I_k(M) := \inf_{(\Omega_1,\ldots,\Omega_k)\in\mathcal{A}_k(M)}\max_{1\le j\le k}\eta_M(\Omega_j). \tag*{(11)}

In particular, I1(M)=0I_1(M) = 0 and I2(M)=I(M)I_2(M) = I(M). Indeed, an admissible set and its complementary region give I2(M)≤I(M)I_2(M) \le I(M). Conversely, in any admissible pair, at least one boundary trace has length at most ℓg(∂M)/2\ell_g(\partial M)/2, and the corresponding ratio is at least I(M)I(M).

Let D\mathbb{D} denote the Euclidean unit disk. Hassannezhad and Siffert [7 Conjecture 1] conjectured that

Ik(M)≤Ik(D)(12)I_k(M) \le I_k(\mathbb{D}) \tag*{(12)}

for every bounded planar domain MM with rectifiable boundary and every integer k≥3k \ge3.

In this section, we use the boundary partition profile to give a short proof of the computation of Ik(D)I_k(\mathbb{D}) by Hassannezhad–Siffert [7 Theorem 1.3]. Then we show that (12) fails even for smooth strictly convex planar domains.

Proposition 5.1 (Hassannezhad–Siffert). For every Euclidean disk D\mathbb{D} and every integer k≥2k \ge2,

Ik(D)=kπsin⁡(πk)≔ck.I_k(\mathbb{D}) = \frac{k}{\pi}\sin\left(\frac{\pi}{k}\right) \coloneqq c_k.

Proof. By scale invariance, it suffices to consider the unit disk D\mathbb{D}. Its boundary partition profile is

PD(s)=2sin⁡(s/2),0<s<2π.\mathcal{P}_{\mathbb{D}}(s) = 2\sin(s/2), \qquad0 < s < 2\pi.

Let (Ω1,…,Ωk)∈Ak(D)(\Omega_1,\ldots,\Omega_k) \in\mathcal{A}_k(\mathbb{D}). If one of the exterior boundary lengths vanishes, the required lower bound is immediate. Otherwise, disjointness gives

∑j=1kℓ(ΓΩj)≤2π.\sum_{j=1}^{k} \ell(\Gamma_{\Omega_j}) \le2\pi.

Hence there exists an index j0j_0 such that 0<s≔ℓ(ΓΩj0)≤2πk0 < s \coloneqq\ell(\Gamma_{\Omega_{j_0}}) \le\frac{2\pi}{k}. Since s↦2sin⁡(s/2)/ss \mapsto2\sin(s/2)/s is decreasing on (0,π](0,\pi],

max⁡1≤j≤kℓ(ΣΩj)ℓ(ΓΩj)≥ℓ(ΣΩj0)s≥PD(s)s=2sin⁡(s/2)s≥kπsin⁡(πk).\max_{1\le j\le k}\frac{\ell(\Sigma_{\Omega_j})}{\ell(\Gamma_{\Omega_j})} \ge\frac{\ell(\Sigma_{\Omega_{j_0}})}{s} \ge\frac{\mathcal{P}_{\mathbb{D}}(s)}{s} = \frac{2\sin(s/2)}{s} \ge\frac{k}{\pi}\sin\left(\frac{\pi}{k}\right).

Taking the infimum proves the lower bound.

For the upper bound, divide ∂D\partial\mathbb{D} into kk equal arcs and take the circular segments bounded by these arcs and their chords, with common boundaries omitted. These sets are pairwise disjoint, and each satisfies

ℓ(ΣΩj)ℓ(ΓΩj)=2sin⁡(π/k)2π/k=kπsin⁡(πk).\frac{\ell(\Sigma_{\Omega_j})}{\ell(\Gamma_{\Omega_j})} = \frac{2\sin(\pi/k)}{2\pi/k} = \frac{k}{\pi}\sin\left(\frac{\pi}{k}\right).

This proves the matching upper bound. □

Next, we construct counterexamples to (12). For ε∈R\varepsilon\in\mathbb{R}, set

Aε=(eε00e−ε),Cε=AεD.A_\varepsilon= \begin{pmatrix} e^\varepsilon& 0 \\ 0 & e^{-\varepsilon} \end{pmatrix}, \qquad C_\varepsilon= A_\varepsilon\mathbb{D}.

Thus CεC_\varepsilon is an ellipse of area π\pi.

Proposition 5.2. For every integer k≥3k \ge3, there exists εk>0\varepsilon_k > 0 such that

Ik(Cε)>Ik(D)for 0<∣ε∣<εk.I_k(C_\varepsilon) > I_k(\mathbb{D}) \qquad\text{for } 0 < |\varepsilon| < \varepsilon_k.

To prove the proposition, we first reduce the problem to regions bounded by one boundary arc and its chord. We call such a region a cap; after straightening its interior boundary, it is a type-1 region in the terminology of [7].

Lemma 5.3. Fix k≥3k \ge3. There exist δ>0\delta> 0 and ε0>0\varepsilon_0 > 0 such that, whenever ∣ε∣<ε0|\varepsilon| < \varepsilon_0, any admissible kk-tuple in CεC_\varepsilon satisfying

max⁡jηCε(Ωj)<ck+δ\max_j \eta_{C_\varepsilon}(\Omega_j) < c_k + \delta

can be replaced by a kk-tuple of caps without increasing its maximal ratio, up to an arbitrarily small error.

Proof. Choose a number ρ\rho such that ck<ρ<ck+1c_k < \rho< c_{k+1}, and choose δ,ε0>0\delta, \varepsilon_0 > 0 sufficiently small that e2ε0(ck+δ)<ρe^{2\varepsilon_0}(c_k+\delta) < \rho. Since the singular values of AεA_\varepsilon are e∣ε∣e^{|\varepsilon|} and e−∣ε∣e^{-|\varepsilon|}, every admissible set Ω⊂Cε\Omega\subset C_\varepsilon satisfies

ηD(Aε−1Ω)≤e2∣ε∣ηCε(Ω).(13)\eta_{\mathbb{D}}(A_\varepsilon^{-1}\Omega) \le e^{2|\varepsilon|}\eta_{C_\varepsilon}(\Omega). \tag*{(13)}

We first perform the usual planar reductions. From each set, retain a connected component whose ratio does not exceed that of the whole set. Fill interior holes and replace interior boundary arcs by their chords. These operations do not increase the ratios. Convexity and the noninterlacing of the boundary traces allow the resulting regions to remain disjoint. Coincident endpoints or interfaces can be separated by arbitrarily small perturbations. This is the chord-straightening reduction used in [7, proof of Theorem 1.3].

Absorbing leaf

Figure 1. Absorbing leaf

The resulting noncrossing chords partition the ellipse into regions. The graph having these regions as vertices and the chords as edges is a tree. If a leaf region does not belong to the chosen kk-tuple, absorb it into its adjacent chosen region (see the figure above). This removes an interior chord and adds an exterior boundary arc, so the corresponding ratio decreases. Repeating this procedure, we may assume that every leaf region belongs to the tuple.

Pull the resulting configuration back to D\mathbb{D}. By (13), all its ratios are strictly smaller than ρ\rho. Let a∈(0,π/2)a \in(0,\pi/2) be determined by

sin⁡aa=ρ.\frac{\sin a}{a}=\rho.

The monotonicity of sin⁡x/x\sin x/x gives

πk+1<a<πk.\frac{\pi}{k+1}<a<\frac{\pi}{k}.

A leaf region in the disk is a circular segment. If its boundary arc has length bb, its ratio is 2sin⁡(b/2)b\frac{2\sin(b/2)}{b}. Consequently, every leaf region has boundary arc length greater than 2a2a.

Removing any edge of a finite tree leaves a leaf of the original tree on each side. Therefore, each side of every chord contains the boundary trace of a chosen leaf region. Both boundary arcs cut off by that chord have length greater than 2a2a. Its Euclidean length is thus at least 2sin⁡a2\sin a.

For the jj-th chosen region in the disk, let djd_j be the number of its interior chords, and let bjb_j be the length of its exterior boundary. Then

ρbj≥ℓ(ΣΩj)≥2djsin⁡a.\rho b_j \geq\ell(\Sigma_{\Omega_j}) \geq2d_j\sin a.

Since the exterior traces are disjoint, ∑jbj≤2π\sum_j b_j \leq2\pi, and hence

∑j=1kdj≤πρsin⁡a=πa<k+1.\sum_{j=1}^{k}d_j\leq\frac{\pi\rho}{\sin a}=\frac{\pi}{a}<k+1.

Each djd_j is a positive integer. It follows that dj=1d_j=1 for every jj, so every chosen region is a cap (see the figure below). The same is true after applying AεA_\varepsilon.

An ellipse partitioned into regions, with a grey diagonal region

Figure 2. The grey region cannot occur in such a configuration.

Proof of Proposition 5.2. Fix k≥3k \ge3 and put

h0=πk,f(h)=sinh⁡hh.h_0 = \frac{\pi}{k}, \qquad f(h) = \frac{\sinh h}{h}.

Thus ck=f(h0)c_k=f(h_0).

Parametrize the boundary of the ellipse by

γε(t)=(eεcos⁡t,e−εsin⁡t),t∈R/2πZ,\gamma_\varepsilon(t) = \left(e^\varepsilon\cos t,e^{-\varepsilon}\sin t\right), \qquad t \in\mathbb{R}/2\pi\mathbb{Z},

and write

vε(t)=∣γε′(t)∣=e2εsin⁡2t+e−2εcos⁡2t.v_\varepsilon(t) = \lvert\gamma_\varepsilon'(t)\rvert= \sqrt{e^{2\varepsilon}\sin^2t+e^{-2\varepsilon}\cos^2t}.

A cap corresponding to the parameter interval [m−h,m+h][m-h,m+h] has ratio

Rε(m,h)=2sinh⁡h vε(m)∫m−hm+hvε(t) dt.(14)R_\varepsilon(m,h)=\frac{2\sinh h\,v_\varepsilon(m)}{\int_{m-h}^{m+h}v_\varepsilon(t)\,\mathrm{d}t}. \tag*{(14)}

Here 2h2h is a parameter interval length, rather than the arclength of the corresponding elliptic arc.

We first explain the reduction to equal-ratio cap configurations. Applying AεA_\varepsilon to an optimal disk configuration gives

Ik(Cε)≤e2∣ε∣ck.I_k(C_\varepsilon)\le e^{2\lvert\varepsilon\rvert}c_k.

Lemma 5.3 therefore applies to minimizing sequences when ∣ε∣\lvert\varepsilon\rvert is sufficiently small. For such a cap configuration, let 2hj2h_j be the lengths of its boundary parameter intervals. Pulling back to the disk and using (13) gives

f(hj)≤ck+Ok(∣ε∣).f(h_j)\le c_k+O_k(\lvert\varepsilon\rvert).

Together with ∑jhj≤π\sum_j h_j\le\pi and the strict monotonicity of ff, this implies

hj=h0+Ok(∣ε∣)h_j=h_0+O_k(\lvert\varepsilon\rvert)

for every jj, with an additional error tending to zero along the minimizing sequence.

At ε=0\varepsilon=0, extending either endpoint of a cap strictly decreases its ratio near h=h0h=h_0, since f′(h0)<0f'(h_0)<0. This remains true for small ∣ε∣\lvert\varepsilon\rvert. We may therefore extend the boundary intervals to fill the gaps without increasing the maximal ratio. The configuration is then described by

t0<t1<⋯<tk=t0+2π,hj=tj+1−tj2,mj=tj+1+tj2,t_0<t_1<\cdots<t_k=t_0+2\pi,\qquad h_j=\frac{t_{j+1}-t_j}{2},\qquad m_j=\frac{t_{j+1}+t_j}{2},

where j=0,…,k−1j=0,\ldots,k-1. The resulting minimization takes place in a compact neighborhood, modulo rotation, of the regular partitions. At a minimizing partition all kk ratios are equal. Indeed, at ε=0\varepsilon=0 their linearizations with respect to the endpoints are

f′(h0)2(uj+1−uj).\frac{f'(h_0)}{2}(u_{j+1}-u_j).

Any proper subset of these rows is linearly independent, and this remains true for sufficiently small ∣ε∣|\varepsilon|. If only a proper subset of the ratios attained the maximum, an endpoint variation could strictly decrease all those ratios while keeping the others below the previous maximum, a contradiction.

We now parametrize the equal-ratio configurations. For θ∈R/2πZ\theta\in\mathbb{R}/2\pi\mathbb{Z}, consider

{Rε(mj,hj)=q,j=0,…,k−1,∑j=0k−1(tj−θ−2jh0)=0.(15)\begin{cases} R_\varepsilon(m_j,h_j)=q, & j=0,\ldots,k-1,\\ \sum_{j=0}^{k-1}(t_j-\theta-2jh_0)=0. \tag*{(15)} \end{cases}

with tk=t0+2πt_k=t_0+2\pi. At ε=0\varepsilon=0, the solution is

tj0=θ+2jh0,q=ck.t_j^0=\theta+2jh_0,\qquad q=c_k.

The linearization in (t0,…,tk−1,q)(t_0,\ldots,t_{k-1},q) is invertible: summing the first kk linearized equations determines the variation of qq, the cyclic differences then determine the endpoint variations up to a common constant, and the last equation fixes that constant.

The implicit function theorem thus gives smooth functions tj(θ,ε)t_j(\theta,\varepsilon) and q(θ,ε)q(\theta,\varepsilon), uniformly in θ\theta. By the preceding reduction and uniqueness near the regular partitions,

Ik(Cε)=min⁡θq(θ,ε).(16)I_k(C_\varepsilon)=\min_\theta q(\theta,\varepsilon). \tag*{(16)}

We compute the second-order expansion of qq. Uniformly in tt,

vε(t)=1−εcos⁡2t+ε2(34−14cos⁡4t)+O(∣ε∣3).v_\varepsilon(t)=1-\varepsilon\cos2t+\varepsilon^2\left(\frac{3}{4}-\frac{1}{4}\cos4t\right)+O(|\varepsilon|^3).

Set

A(h)=sin⁡2h2h,B(h)=sin⁡4h4h.A(h)=\frac{\sin2h}{2h},\qquad B(h)=\frac{\sin4h}{4h}.

Substitution into (14) gives

Rε(m,h)=f(h)[1+ε(A(h)−1)cos⁡2m+ε2(A(h)(A(h)−1)cos⁡22m+B(h)−14cos⁡4m)]+O(∣ε∣3).(17)R_\varepsilon(m,h)=f(h)\left[1+\varepsilon(A(h)-1)\cos2m+\varepsilon^2\left(A(h)(A(h)-1)\cos^2 2m+\frac{B(h)-1}{4}\cos4m\right)\right]+O(|\varepsilon|^3). \tag*{(17)}

Write a dot for differentiation in ε\varepsilon at ε=0\varepsilon=0, and set mj0=θ+(2j+1)h0m_j^0=\theta+(2j+1)h_0. Differentiating (5.5) yields

f′(h0)h˙j+f(h0)(A(h0)−1)cos⁡2mj0=q˙.f'(h_0)\dot{h}_j+f(h_0)(A(h_0)-1)\cos2m_j^0=\dot{q}.

Since

∑j=0k−1h˙j=0,∑j=0k−1cos⁡2mj0=0,\sum_{j=0}^{k-1}\dot{h}_j=0,\qquad\sum_{j=0}^{k-1}\cos2m_j^0=0,

we obtain q˙=0\dot{q}=0. Define

β(h)=−f(h)f′(h)(A(h)−1).\beta(h)=-\frac{f(h)}{f'(h)}(A(h)-1).

Solving the cyclic difference equations, with the normalization in (5.5), gives

h˙j=β(h0)cos⁡2mj0,m˙j=β(h0)cot⁡(2h0)sin⁡2mj0.(18)\dot{h}_j=\beta(h_0)\cos2m_j^0,\qquad\dot{m}_j=\beta(h_0)\cot(2h_0)\sin2m_j^0. \tag*{(18)}

Suppose first that k=3k=3 or k≥5k\geq5. Then

1k∑j=0k−1cos⁡22mj0=1k∑j=0k−1sin⁡22mj0=12,∑j=0k−1cos⁡4mj0=0.\frac{1}{k}\sum_{j=0}^{k-1}\cos^2 2m_j^0=\frac{1}{k}\sum_{j=0}^{k-1}\sin^2 2m_j^0=\frac{1}{2},\qquad\sum_{j=0}^{k-1}\cos4m_j^0=0.

Expanding (5.5) to second order, using (5.7) and (5.8), and averaging over jj, we find

q(θ,ε)=ck(1+a(h0)ε2)+Ok(∣ε∣3),(19)q(\theta,\varepsilon)=c_k\left(1+a(h_0)\varepsilon^2\right)+O_k(|\varepsilon|^3), \tag*{(19)}

where the remainder is uniform in θ\theta and

a(h)=(h−sinh⁡hcosh⁡h)(2tanh⁡h+sinh⁡hcosh⁡h−3h)4(1−hcoth⁡h)2.a(h)=\frac{(h-\sinh h\cosh h)(2\tanh h+\sinh h\cosh h-3h)}{4(1-h\coth h)^2}.

Both factors in the numerator are positive for 0<h<π/20<h<\pi/2. Indeed, they vanish at h=0h=0, and

ddh(h−sinh⁡hcosh⁡h)=2sin⁡2h>0,\frac{d}{dh}(h-\sinh h\cosh h)=2\sin^2 h>0,

while

ddh(2tanh⁡h+sinh⁡hcosh⁡h−3h)=2sin⁡4hcos⁡2h>0.\frac{d}{dh}(2\tanh h+\sinh h\cosh h-3h)=\frac{2\sin^4 h}{\cos^2 h}>0.

Consequently, a(h0)>0a(h_0)>0.

For k=4k=4, the fourth Fourier mode survives the averaging. Writing

a4=a(π4)=(π−2)(10−3π)4(4−π)2,a_4=a\left(\frac{\pi}{4}\right)=\frac{(\pi-2)(10-3\pi)}{4(4-\pi)^2},

the same expansion gives

q(θ,ε)=c4[1+(a4+(14−a4)cos⁡4θ)ε2]+O(∣ε∣3).q(\theta,\varepsilon)=c_4\left[1+\left(a_4+\left(\frac{1}{4}-a_4\right)\cos4\theta\right)\varepsilon^2\right]+O(|\varepsilon|^3).

Here

14−a4=(π−3)2(4−π)2>0,\frac{1}{4}-a_4=\frac{(\pi-3)^2}{(4-\pi)^2}>0,

and therefore the quadratic coefficient is bounded below by

2a4−14=−7π2+40π−564(4−π)2>0.2a_4-\frac{1}{4}=\frac{-7\pi^2+40\pi-56}{4(4-\pi)^2}>0.

Thus, for every fixed k≥3k \ge3, there is a constant μk>0\mu_k>0 such that, uniformly in θ\theta,

q(θ,ε)≥ck(1+μkε2)−Ok(∣ε∣3).q(\theta,\varepsilon)\ge c_k(1+\mu_k\varepsilon^2)-O_k(|\varepsilon|^3).

Taking the minimum in (5.6) proves Ik(Cε)>ck=Ik(D)I_k(C_\varepsilon)>c_k=I_k(\mathbb{D}) for all sufficiently small nonzero ε\varepsilon. □\square

Declaration of AI use We use ChatGPT 6 Astra for proof exploration and editing. We take responsibility for verification and exposition.

School of Mathematics, Nanjing University, Nanjing 210093, Jiangsu, China

Email address: [email protected]

School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai 200240, China

Email address: [email protected]

School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai 200240, China

Email address: [email protected]

References

  1. [1]S. B. Alexander, J. D. Berg, and R. L. Bishop, Geometric curvature bounds in Riemannian manifolds with boundary, Trans. Amer. Math. Soc. 339 (1993), no. 2, 703–716.DOI
  2. [2]B. H. Bowditch, Notes on locally CAT(1) spaces, in Geometric Group Theory, Ohio State Univ. Math. Res. Inst. Publ. 3, de Gruyter, Berlin, 1995, 1–48.
  3. [3]A. Cianchi, V. Ferone, C. Nitsch, and C. Trombetti, Balls minimize trace constants in BV, J. Reine Angew. Math. 725 (2017), 41–61.
  4. [4]J. F. Escobar, The geometry of the first non-zero Stekloff eigenvalue, J. Funct. Anal. 150 (1997), no. 2, 544–556.DOI
  5. [5]J. F. Escobar, An isoperimetric inequality and the first Steklov eigenvalue, J. Funct. Anal. 165 (1999), no. 1, 101–116.DOI
  6. [6]A. Hassannezhad and L. Miclo, Higher order Cheeger inequalities for Steklov eigenvalues, Ann. Sci. Éc. Norm. Supér. (4) 53 (2020), no. 1, 43–88.arxiv.org/abs/1705.08643
  7. [7]A. Hassannezhad and A. Siffert, Escobar constants of planar domains, Ann. Global Anal. Geom. 61 (2022), 215–233.arxiv.org/abs/1905.07634
  8. [8]R. B. Kusner and J. M. Sullivan, On distortion and thickness of knots, in Topology and Geometry in Polymer Science (Minneapolis, MN, 1996), IMA Vol. Math. Appl., vol. 103, Springer, New York, 1998, pp. 67–78.arxiv.org/abs/dg-ga/9702001
  9. [9]G. Lukó, On the mean length of the chords of a closed curve, Israel J. Math. 4 (1966), 23–32.DOI
  10. [10]A. Lytchak and S. Wenger, Isoperimetric characterization of upper curvature bounds, Acta Math. 221 (2018), no. 1, 159–202.arxiv.org/abs/1611.05261
  11. [11]Yu. G. Reshetnyak, Inextensible mappings in a space of curvature no greater than K, Siberian Math. J. 9 (1968), no. 4, 683–689.DOI

Paper details

vlnc:W_atvys69x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents