betaFeedback

WorksMathematics

The Square Maximizes the First Nonzero Steklov Eigenvalue Among Convex Quadrilaterals

Abstract

We show that among all convex planar polygons of prescribed perimeter with at most four sides, the square uniquely maximizes the first positive Steklov eigenvalue. If k∈(0,π/2)k \in (0, \pi/2) denotes the unique root of tan⁡ktanh⁡k=1\tan k \tanh k = 1, then Lσ1(Ω)≤8ktanh⁡kL\sigma_1(\Omega) \le 8k\tanh k, with equality if and only if Ω\Omega is a square. By the quadrilateral isoperimetric inequality, the square also uniquely maximizes the area-normalized eigenvalue ∣Ω∣ σ1(Ω)\sqrt{|\Omega|}\,\sigma_1(\Omega) in the same class. Our proof splits the four-dimensional moduli space of convex quadrilaterals into a neighborhood of the square and its complement. Near the square, a weighted combination of the two first Steklov eigenfunctions of the square produces a negative definite Rayleigh defect; its sign on an explicit parameter box is established via an exact degree-2222 Taylor expansion in ball arithmetic together with an analytic Cauchy remainder estimate. Far from the square, a boundary-variance inequality along a diagonal reduces the remaining shapes to a compact region, which we cover by 940940 closed subboxes using affine and harmonic quadratic trial functions checked in outward-rounded integer interval arithmetic.

Introduction and main results

Let Ω⊂R2\Omega\subset\mathbb{R}^2 be a bounded convex polygon of positive area ∣Ω∣|\Omega| and perimeter L=∣∂Ω∣L = |\partial\Omega|. The Steklov eigenvalue problem on Ω\Omega asks for nonzero functions u∈H1(Ω)u \in H^1(\Omega) and real numbers σ\sigma satisfying

Δu=0in Ω,∂νu=σuon ∂Ω\Delta u = 0 \quad\text{in } \Omega,\qquad\partial_\nu u = \sigma u \quad\text{on } \partial\Omega

in the weak sense, where ∂ν\partial_\nu is the outward normal derivative along ∂Ω\partial\Omega. Because Ω\Omega is a bounded Lipschitz domain, the trace operator H1(Ω)→L2(∂Ω)H^1(\Omega) \to L^2(\partial\Omega) is compact, and the Steklov spectrum forms an unbounded sequence

0=σ0(Ω)<σ1(Ω)≤σ2(Ω)≤⋯→+∞0 = \sigma_0(\Omega) < \sigma_1(\Omega) \le\sigma_2(\Omega) \le\cdots\to+\infty

counted with multiplicity; we refer to the surveys [7, 5] for background on the spectral geometry of this problem. Writing dμ=ds/Ld\mu= ds/L for normalized arclength measure on ∂Ω\partial\Omega and Ef=∫Ω∣∇f∣2E_f = \int_\Omega|\nabla f|^2 for the Dirichlet energy of f∈H1(Ω)f \in H^1(\Omega), the scale-invariant functional Σ(Ω):=Lσ1(Ω)\Sigma(\Omega) := L\sigma_1(\Omega) admits the variational characterization

Σ(Ω):=Lσ1(Ω)=inf⁡f∈H1(Ω)Var⁡μ(f)>0EfVar⁡μ(f),Var⁡μ(f)=∫∂Ωf2 dμ−(∫∂Ωf dμ)2(1)\Sigma(\Omega) := L\sigma_1(\Omega) = \inf_{\substack{f\in H^1(\Omega)\\ \operatorname{Var}_\mu(f)>0}} \frac{E_f}{\operatorname{Var}_\mu(f)},\qquad\operatorname{Var}_\mu(f) = \int_{\partial\Omega} f^2\,d\mu- \left(\int_{\partial\Omega} f\,d\mu\right)^2 \tag*{(1)}

and is invariant under Euclidean similarities.

By Weinstock’s classical inequality [9], every simply connected planar Lipschitz domain satisfies Σ(Ω)≤2π\Sigma(\Omega) \le2\pi, with equality only for disks. In higher dimensions, the perimeter-normalized bound extends to convex domains [3], whereas the volume-normalized inequality is due to Brock [2] and holds for all bounded Lipschitz domains; in dimension two, the area-normalized bound follows directly from Weinstock’s theorem and the isoperimetric inequality.

2020 Mathematics Subject Classification. Primary 35P15; Secondary 49R05, 52A10, 52A40, 65G20.

Key words and phrases. Steklov eigenvalue problem, convex quadrilaterals, Weinstock inequality, polygonal shape optimization, interval arithmetic. A natural polygonal counterpart, analogous to the Pólya–Szegő conjecture for the first Dirichlet eigenvalue and Laugesen’s problem for the first positive Neumann eigenvalue, asks whether the regular NN-gon maximizes Σ(Ω)\Sigma(\Omega) among convex polygons of given perimeter with at most NN sides. Unlike the Dirichlet problem for triangles and quadrilaterals, where Steiner symmetrization settles the sharp bounds, symmetrization does not preserve the boundary L2L^2-norm and perimeter in a way compatible with (1). The polygonal Steklov problem was raised by Oudet in the 2018 AIM Steklov problem list [1] and formulated explicitly for convex polygons with at most NN sides by Cheng, Gui, Hu, Li, and Yao [4 Conjecture 1.1], who also established that Σ\Sigma increases monotonically along the sequence of regular NN-gons [4 Theorem 1.1].

Using separation of variables and careful estimates for the resulting transcendental equations, Girouard, Lagacé, Polterovich, and Savo [6] proved that the square maximizes σ1\sigma_1 among rectangles of fixed perimeter or fixed area. For N=3N=3, a preprint by Davies gives a computer-assisted proof that the equilateral triangle uniquely maximizes Σ\Sigma among all triangles. Here we prove the case N=4N=4. For each triangular competitor TT, Lemma 3 gives the independent bound Σ(T)≤33\Sigma(T) \le3\sqrt{3}, strictly below the value for the square.

Theorem 1. Let Ω⊂R2\Omega\subset\mathbb{R}^2 be a convex polygon of positive area with at most four sides, and let k∈(0,π/2)k \in(0,\pi/2) be the unique root of tan⁡ktanh⁡k=1\tan k \tanh k = 1. Then

Σ(Ω)≤8ktanh⁡k=5.50602…,\Sigma(\Omega) \le8k\tanh k = 5.50602\ldots,

with equality if and only if Ω\Omega is a square.

Combining Theorem 1 with the quadrilateral isoperimetric inequality gives the corresponding bound at fixed area.

Corollary 2. Among all convex planar polygons of positive area with at most four sides,

σ1(Ω)∣Ω∣≤2ktanh⁡k,\sigma_1(\Omega)\sqrt{|\Omega|} \le2k\tanh k,

with equality if and only if Ω\Omega is a square.

Proof. Let ℓ1,ℓ2,ℓ3,ℓ4\ell_1,\ell_2,\ell_3,\ell_4 be the successive side lengths of Ω\Omega, inserting a fourth collinear vertex if Ω\Omega is a triangle. Splitting Ω\Omega along each diagonal in turn yields 2∣Ω∣≤ℓ1ℓ2+ℓ3ℓ42|\Omega| \le\ell_1\ell_2+\ell_3\ell_4 and 2∣Ω∣≤ℓ2ℓ3+ℓ4ℓ12|\Omega| \le\ell_2\ell_3+\ell_4\ell_1. Adding these inequalities gives

4∣Ω∣≤(ℓ1+ℓ3)(ℓ2+ℓ4)≤(ℓ1+ℓ2+ℓ3+ℓ4)24=L24,4|\Omega| \le(\ell_1+\ell_3)(\ell_2+\ell_4) \le\frac{(\ell_1+\ell_2+\ell_3+\ell_4)^2}{4} = \frac{L^2}{4},

so ∣Ω∣/L≤1/4\sqrt{|\Omega|}/L \le1/4, with equality precisely for squares. Multiplying Theorem 1 by ∣Ω∣/L≤1/4\sqrt{|\Omega|}/L \le1/4 and checking equality on a square completes the proof. □

Two features distinguish the general quadrilateral problem from the rectangular and triangular cases. First, the space of similarity classes of convex quadrilaterals is four-dimensional: beyond the two affine shape parameters governing parallelograms, two nonaffine parameters control trapezoidal and kite-like distortions. Second, the first positive Steklov eigenvalue of the square has multiplicity two, so Σ\Sigma is not smooth at the square and a single trial function cannot detect all perturbation directions simultaneously.

Our argument proceeds in three stages. In § 2, we parametrize convex quadrilaterals up to similarity as bilinear images of the reference square S=[−1,1]2S=[-1,1]^2 in four coordinates (a,c,p,q)(a,c,p,q), and we prove an elementary diagonal variance bound (Lemma 3) that excludes all triangles together with all quadrilaterals satisfying a≥0.21a\ge0.21 or c≥0.045c\ge0.045. In § 3, we form a weighted combination of the two first Steklov eigenfunctions of SS whose Rayleigh defect T(a,c,p,q)T(a,c,p,q) is invariant under the symmetries of the square and has a negative definite Hessian at the origin; an exact degree-22 Taylor expansion in ball arithmetic combined with an analytic Cauchy tail bound shows that T<0T<0 throughout the central box ∣a∣≤0.21|a|\le0.21, ∣c∣≤0.045|c|\le0.045, ∣p∣≤0.20|p| \le0.20, ∣q∣≤0.28|q| \le0.28. In § 4, we treat the remaining compact region by partitioning it into 940 closed subboxes on which explicit affine or harmonic quadratic trial functions yield Σ(Ω)<11/2<8ktanh⁡k\Sigma(\Omega) < 11/2 < 8k \tanh k, completing the proof of Theorem 1. § 5 gathers several concluding remarks on quantitative stability, possible extensions, and computational details.

Shape coordinates and a diagonal estimate

Let S=[−1,1]2S = [-1,1]^2. Bilinear interpolation of the four positively ordered vertices of a convex quadrilateral Ω\Omega represents Ω\Omega as the image of SS under

Z(x,y)=(X(x,y),Y(x,y))=ex+fy+hxy,(2)\mathbf{Z}(x,y) = (X(x,y),Y(x,y)) = \mathbf{e}x + \mathbf{f}y + \mathbf{h}xy, \tag*{(2)}

where, after applying a suitable Euclidean similarity,

e=(1+a,c),f=(c,1−a),h=(p,q).\mathbf{e} = (1+a,c), \qquad\mathbf{f} = (c,1-a), \qquad\mathbf{h} = (p,q).

To reach this normalization, translate the arithmetic mean of the four vertices to the origin. The columns e,f\mathbf{e},\mathbf{f} of the linear part of Z\mathbf{Z} are then two consecutive semidiagonals of the Varignon midpoint parallelogram (Figure 1, left), and positive vertex ordering ensures det⁡(e,f)>0\det(\mathbf{e},\mathbf{f}) > 0. Applying the orthogonal factor from the polar decomposition of (e f)(\mathbf{e}\ \mathbf{f}) rotates the linear part into a symmetric positive definite matrix, and a dilation rescales its trace to 2.

Bilinear chart and boundary arc decomposition

Figure 1. Left: the bilinear chart Z:[−1,1]2→Ω\mathbf{Z}: [-1,1]^2 \to\Omega and the inscribed midpoint parallelogram (dashed) with semidiagonals e,f\mathbf{e},\mathbf{f}. Right: the boundary arc decomposition along a normalized diagonal in Lemma 3.

Setting

Δ=1−a2−c2>0,jx=(1+a)q−cp,jy=(1−a)p−cq,\Delta= 1-a^2-c^2 > 0, \qquad j_x = (1+a)q-cp,\qquad j_y = (1-a)p-cq,
J(x,y)=det⁡DZ(x,y)=Δ+jxx+jyy,∣Ω∣=4Δ,(3)J(x,y) = \det D\mathbf{Z}(x,y) = \Delta+ j_xx+j_yy,\qquad|\Omega|=4\Delta, \tag*{(3)}

we see that for a quadrilateral with four strictly convex corners, consecutive edges make positive turns, so JJ is positive at all four corners of SS and hence throughout SS by affinity. The map Z\mathbf{Z} has topological degree 1 at every interior point of Ω\Omega and degree 0 outside Ω\Omega; since every interior preimage has positive Jacobian, Z\mathbf{Z} maps (−1,1)2(-1,1)^2 diffeomorphically onto the interior of Ω\Omega. In particular, nonnegativity of JJ at the corners of SS reads

∣jx∣+∣jy∣≤Δ.|j_x|+|j_y| \le\Delta.

Since h=(jye+jxf)/Δ\mathbf{h}=(j_y\mathbf{e}+j_x\mathbf{f})/\Delta, this convexity condition says precisely that h∈conv⁡(±e,±f)\mathbf{h}\in\operatorname{conv}(\pm\mathbf{e},\pm\mathbf{f}). The remaining discrete symmetries of the square act on (a,c,p,q)(a,c,p,q) as follows: exchanging the two coordinate axes sends (a,c,p,q)↦(−a,c,q,p)(a,c,p,q)\mapsto(-a,c,q,p); reflecting across one axis sends (a,c,p,q)↦(a,−c,p,−q)(a,c,p,q)\mapsto(a,-c,p,-q); and inversion through the origin sends (a,c,p,q)↦(a,c,−p,−q)(a,c,p,q)\mapsto(a,c,-p,-q). Composing these symmetries allows us to assume without loss of generality that a≥0a\ge0, c≥0c\ge0, and p≥0p\ge0. Whenever 0≤a≤0.210\le a\le0.21 and 0≤c≤0.0450\le c\le0.045, the inclusion h∈conv⁡(±e,±f)\mathbf{h}\in\operatorname{conv}(\pm\mathbf{e},\pm\mathbf{f}) bounds the components of h\mathbf{h} by

0≤p≤max⁡(1+a,c)≤1.21,∣q∣≤max⁡(c,1−a)≤1.0 \le p \le\max(1+a,c) \le1.21,\qquad|q| \le\max(c,1-a) \le1.

Lemma 3. Let DD be the length of either diagonal of a convex quadrilateral Ω\Omega. Then

Σ(Ω)≤12∣Ω∣D2.\Sigma(\Omega) \le\frac{12|\Omega|}{D^2}.

The same inequality holds when Ω\Omega is a triangle and DD is the length of any of its sides.

Proof. Align the chosen diagonal with the horizontal axis and rescale so that its endpoints lie at (−1,0)(-1,0) and $(1,0)) (Figure 1, right). Each of the two polygonal arcs of ∂Ω\partial\Omega joining these endpoints consists of two line segments meeting at an intermediate vertex (x0,y0)(x_0,y_0), whose lengths L1,L2L_1,L_2 satisfy L12−L22=(x0+1)2−(x0−1)2=4x0L_1^2-L_2^2=(x_0+1)^2-(x_0-1)^2=4x_0. Under normalized arclength on this arc, set δ=(L1−L2)/(L1+L2)\delta=(L_1-L_2)/(L_1+L_2), so the two segments carry weights (1+δ)/2(1+\delta)/2 and (1−δ)/2(1-\delta)/2. Conditional on each segment, the horizontal coordinate is uniform on [−1,x0][-1,x_0] or [x0,1][x_0,1], with conditional variances (x0+1)2/12(x_0+1)^2/12 and (x0−1)2/12(x_0-1)^2/12 and conditional means (x0−1)/2(x_0-1)/2 and (x0+1)/2(x_0+1)/2 differing by $1$. By the law of total variance, the variance of the horizontal coordinate along this arc equals

1+δ2(x0+1)212+1−δ2(x0−1)212+1−δ24=13+(x0−δ)(x0+3δ)12.\frac{1+\delta}{2}\frac{(x_0+1)^2}{12}+\frac{1-\delta}{2}\frac{(x_0-1)^2}{12}+\frac{1-\delta^2}{4}=\frac{1}{3}+\frac{(x_0-\delta)(x_0+3\delta)}{12}.

Because L1+L2≥2L_1+L_2\ge2 by the triangle inequality, the identity δ=4x0/(L1+L2)2\delta=4x_0/(L_1+L_2)^2 implies that δ\delta has the same sign as x0x_0 and ∣δ∣≤∣x0∣|\delta|\le|x_0|. Thus (x0−δ)(x0+3δ)≥0(x_0-\delta)(x_0+3\delta)\ge0, so the variance on each arc is at least 1/31/3. Since the variance of a mixture of two probability distributions is at least the average of their component variances, the boundary variance over all of ∂Ω\partial\Omega is also at least 1/31/3. Restoring the length scale gives Var⁡μ(f)≥D2/12\operatorname{Var}_{\mu}(f)\ge D^2/12 for the unit-slope linear projection ff, whose Dirichlet energy is Ef=∣Ω∣E_f=|\Omega|. Inserting ff into (1) establishes the claim. For a triangle, subdividing the chosen side at its midpoint yields two arcs of the same two-segment type. □

For a triangle Ω\Omega, choosing DD to be its longest side forces the third vertex to lie in the intersection of the two disks of radius DD centered at the endpoints of that side, so its altitude is at most 3D/2\sqrt{3}D/2 and ∣Ω∣≤3D2/4|\Omega|\le\sqrt{3}D^2/4. Lemma 3 then gives

Σ(Ω)≤33<112.(4)\Sigma(\Omega)\le3\sqrt{3}<\frac{11}{2}. \tag*{(4)}

For a convex quadrilateral in the chart (2) with c≥0c\ge0, the longer diagonal has length D=2∣e+f∣D=2|e+f|, where ∣e+f∣2=2(1+a2+c2+2c)|e+f|^2=2(1+a^2+c^2+2c). Applying Lemma 3 yields

Σ(Ω)≤6Δ1+a2+c2+2c<112(5)\Sigma(\Omega)\le\frac{6\Delta}{1+a^2+c^2+2c}<\frac{11}{2} \tag*{(5)}

whenever 23(a2+c2)+22c>123(a^2+c^2)+22c>1. In particular, (5) holds whenever a≥0.21a\ge0.21 or c≥0.045c\ge0.045.

Local analysis near the square

On the reference square S=[−1,1]2S=[-1,1]^2, let k∈(0,π/2)k\in(0,\pi/2) be the unique solution of tan⁡ktanh⁡k=1\tan k\tanh k=1, and put s=ktanh⁡ks=k\tanh k. By [6 Theorem 1.6 and Proposition 4.2], the first positive Steklov eigenvalue of SS is σ1(S)=s\sigma_1(S)=s, with multiplicity 22, and the corresponding eigenspace is spanned by the harmonic functions

u(X,Y)=sin⁡(kX)cosh⁡(kY),v(X,Y)=cosh⁡(kX)sin⁡(kY).u(X,Y)=\sin(kX)\cosh(kY),\qquad v(X,Y)=\cosh(kX)\sin(kY).

In the coordinates (2), write LL for the perimeter of Ω\Omega, put l=L/8l=L/8, and define for f∈H1(Ω)f\in H^1(\Omega)

Ef=∫Ω∣∇f∣2,Bf=∫∂Ωf2,mf=∫∂Ωf,Hf=l2Ef−slBf+s8mf2.E_f=\int_{\Omega}|\nabla f|^2,\qquad B_f=\int_{\partial\Omega}f^2,\qquad m_f=\int_{\partial\Omega}f,\qquad H_f=l^2E_f-slB_f+\frac{s}{8}m_f^2.

Whenever Hf<0H_f<0, testing the centered function f−mf/Lf-m_f/L in (1) yields Lσ1(Ω)<8sL\sigma_1(\Omega)<8s. To probe both modes of the double eigenspace at once, we form the weighted Rayleigh defect

T(a,c,p,q)=(1+a)Hu+(1−a)Hv,T(a,c,p,q)=(1+a)H_u+(1-a)H_v,

whose weights 1±a1 \pm a remain positive throughout the region considered below. The three discrete symmetries of the square act on (a,c,p,q)(a,c,p,q) by

(a,c,p,q)↦(a,−c,p,−q),(a,c,−p,−q),(−a,c,q,p).(a,c,p,q) \mapsto(a,-c,p,-q), \qquad(a,c,-p,-q), \qquad(-a,c,q,p).

leaving TT invariant. At the origin (0,0,0,0)(0,0,0,0) we have T(0)=0T(0)=0, since uu and vv are zero-mean Steklov eigenfunctions on SS and l=1l=1. Inversion eliminates every monomial of odd total degree in (p,q)(p,q) (in particular p,q,ap,aq,cp,cqp,q,ap,aq,cp,cq in degrees 1 and 2); reflection across one axis eliminates c,ac,pqc,ac,pq; and coordinate exchange removes the linear term in aa while equating the coefficients of p2p^2 and q2q^2. Consequently, the quadratic Taylor polynomial of TT at the origin takes the diagonal form γaa2+γcc2+γh(p2+q2)\gamma_a a^2+\gamma_c c^2+\gamma_h(p^2+q^2). Evaluating these second-order coefficients via the exact formulas of § 3.1 gives

−5.874<γa<−5.873,−28.171<γc<−28.170,−1.863<γh<−1.862.-5.874 < \gamma_a < -5.873, \qquad-28.171 < \gamma_c < -28.170, \qquad-1.863 < \gamma_h < -1.862.

Thus the square is a strict local maximizer of Σ\Sigma across all four shape coordinates, including both nonaffine directions (p,q)(p,q). We now turn this local negativity into an explicit neighborhood of the origin.

Exact Taylor coefficients

For ε=±1\varepsilon=\pm1, the constant boundary speeds along the horizontal edges y=εy=\varepsilon and vertical edges x=εx=\varepsilon of SS are

w0,ε=(1+a+εp)2+(c+εq)2,wε,0=(1−a+εq)2+(c+εp)2.w_{0,\varepsilon}=\sqrt{(1+a+\varepsilon p)^2+(c+\varepsilon q)^2}, \qquad w_{\varepsilon,0}=\sqrt{(1-a+\varepsilon q)^2+(c+\varepsilon p)^2}.

and the normalized perimeter is 4l=∑ε=±1(w0,ε+wε,0)4l=\sum_{\varepsilon=\pm1}(w_{0,\varepsilon}+w_{\varepsilon,0}). The area Jacobian (3) expands as

J(x,y)=1−a2−c2+(q+aq−cp)x+(p−ap−cq)y.J(x,y)=1-a^2-c^2+(q+aq-cp)x+(p-ap-cq)y.

For u(X,Y)=sin⁡(kX)cosh⁡(kY)u(X,Y)=\sin(kX)\cosh(kY), trigonometric and hyperbolic identities give

∣∇u∣2=12k2[cos⁡(2kX)+cosh⁡(2kY)],u2=14[1−cos⁡(2kX)][1+cosh⁡(2kY)].\lvert\nabla u\rvert^2=\frac{1}{2}k^2\left[\cos(2kX)+\cosh(2kY)\right], \qquad u^2=\frac{1}{4}\left[1-\cos(2kX)\right]\left[1+\cosh(2kY)\right].

while the corresponding formulas for vv follow by replacing (a,c,p,q)(a,c,p,q) with (−a,c,q,p)(-a,c,q,p). Every area or boundary integral entering EuE_u, BuB_u, mum_u is a linear combination of integrals of exponentials exp⁡(k(iαX+βY))\exp(k(i\alpha X+\beta Y)) with α,β∈{0,±1,±2}\alpha,\beta\in\{0,\pm1,\pm2\}, where i=−1i=\sqrt{-1}. Writing

iαX+βY=(iαx+βy)+a(iαx−βy)+c(iαy+βx)+p(iαxy)+q(βxy),i\alpha X+\beta Y=(i\alpha x+\beta y)+a(i\alpha x-\beta y)+c(i\alpha y+\beta x)+p(i\alpha xy)+q(\beta xy),

the coefficient of aν1cν2pν3qν4a^{\nu_1}c^{\nu_2}p^{\nu_3}q^{\nu_4} in the Taylor series of exp⁡(k(iαX+βY))\exp(k(i\alpha X+\beta Y)) is

ek(iαx+βy)kν1+ν2+ν3+ν4ν1! ν2! ν3! ν4!(iαx−βy)ν1(iαy+βx)ν2(iαxy)ν3(βxy)ν4.e^{k(i\alpha x+\beta y)}\frac{k^{\nu_1+\nu_2+\nu_3+\nu_4}}{\nu_1!\,\nu_2!\,\nu_3!\,\nu_4!}(i\alpha x-\beta y)^{\nu_1}(i\alpha y+\beta x)^{\nu_2}(i\alpha xy)^{\nu_3}(\beta xy)^{\nu_4}.

Each resulting one-dimensional integral of xneλxx^n e^{\lambda x} reduces either to an endpoint evaluation (along an edge) or to the elementary integral In(λ)=∫−11xneλx dxI_n(\lambda)=\int_{-1}^{1}x^n e^{\lambda x}\,dx, which obeys the recurrence

I0(λ)=eλ−e−λλ,In(λ)=eλ−(−1)ne−λ−nIn−1(λ)λ(λ≠0),I_0(\lambda)=\frac{e^\lambda-e^{-\lambda}}{\lambda}, \qquad I_n(\lambda)=\frac{e^\lambda-(-1)^n e^{-\lambda}-nI_{n-1}(\lambda)}{\lambda}\qquad(\lambda\ne0),

with In(0)=2/(n+1)I_n(0)=2/(n+1) for even nn and In(0)=0I_n(0)=0 for odd $n. Combining these exact integrals with the finite binomial expansions of the four edge speeds w0,ε,wε,0w_{0,\varepsilon},w_{\varepsilon,0} through degree N0N_0 determines every Taylor coefficient of TT algebraically, without numerical quadrature. Isolating kk by bisection on tan⁡ktanh⁡k=1\tan k\tanh k=1 and evaluating the recurrences and series products in outward-rounded Arb ball arithmetic [8] yields rigorous enclosures for all coefficients.

Cauchy remainder and the local box

Write z=(a,c,p,q)z=(a,c,p,q), let R1,R2,R3,R4>0R_1,R_2,R_3,R_4>0 be prescribed box half-widths (also denoted Ra,Rc,Rp,RqR_a,R_c,R_p,R_q), and define the positive definite quadratic form

Q(z)=5.8a2+28c2+1.8p2+1.8q2=∑i=14κizi2.Q(z)=5.8a^2+28c^2+1.8p^2+1.8q^2=\sum_{i=1}^{4}\kappa_i z_i^2.

From the enclosures on γa,γc,γh\gamma_a,\gamma_c,\gamma_h above, the degree-2 part of T(z)T(z) is bounded above by −Q(z)-Q(z). For any monomial zνz^\nu of total degree n≥3n\geq3, if every exponent νi\nu_i is even and its Taylor coefficient is nonpositive, the term is nonpositive throughout the box and may be dropped; otherwise, factoring out either zi2z_i^2 (when νi≥2\nu_i\geq2) or ∣zizj∣|z_i z_j| (when νi,νj≥1\nu_i,\nu_j\geq1) and estimating the remaining factors by ∣zr∣≤Rr|z_r|\leq R_r gives

∣zν∣≤Q(z)Rνmin⁡({1κiRi2:νi≥2}∪{12κiκjRiRj:i<j, νiνj>0}).|z^\nu|\leq Q(z)R^\nu\min\left(\left\{\frac{1}{\kappa_iR_i^2}:\nu_i\geq2\right\}\cup\left\{\frac{1}{2\sqrt{\kappa_i\kappa_j}R_iR_j}:i<j,\ \nu_i\nu_j>0\right\}\right).

To control the Taylor tail of degree ≥N0+1\geq N_0+1, write z=tζz=t\zeta with ∣ζi∣≤Ri|\zeta_i|\leq R_i and ∣ζi∣=Ri|\zeta_i|=R_i for at least one index ii, so that 0≤t≤10\leq t\leq1 and Q(z)≥t2min⁡iκiRi2Q(z)\geq t^2\min_i\kappa_iR_i^2. Set Rh=max⁡(Rp,Rq)R_h=\max(R_p,R_q) and choose a radius ρ>1\rho>1 satisfying

ρ2[(Ra+Rp)2+(Rc+Rq)2]<1,ρ2[(Ra+Rq)2+(Rc+Rp)2]<1.\rho^2\left[(R_a+R_p)^2+(R_c+R_q)^2\right]<1,\qquad\rho^2\left[(R_a+R_q)^2+(R_c+R_p)^2\right]<1.

Each squared edge speed factors as (1+t(ξ+iη))(1+t(ξ−iη))(1+t(\xi+i\eta))(1+t(\xi-i\eta)) with real ξ,η\xi,\eta obeying ρ2(ξ2+η2)<1\rho^2(\xi^2+\eta^2)<1, so the principal square roots (positive at t=0t=0) extend holomorphically to the closed disk ∣t∣≤ρ|t|\leq\rho. Every other term in T(tζ)T(t\zeta) is entire in tt. On the circle ∣t∣=ρ|t|=\rho, define

θ0=ρ(Ra+Rc+Rh),U=cosh⁡(θ0)cosh⁡(1+θ0),\theta_0=\rho(R_a+R_c+R_h),\qquad U=\cosh(\theta_0)\cosh(1+\theta_0),
G=12[cosh⁡(2θ0)+cosh⁡(2+2θ0)],w∗2=[1+ρ(Ra+Rh)]2+ρ2(Rc+Rh)2,G=\frac{1}{2}\left[\cosh(2\theta_0)+\cosh(2+2\theta_0)\right],\qquad w_*^2=\left[1+\rho(R_a+R_h)\right]^2+\rho^2(R_c+R_h)^2,

and

J∗=1+ρ(Rp+Rq)+ρ2[Ra2+Rc2+(Ra+Rc)(Rp+Rq)],J_*=1+\rho(R_p+R_q)+\rho^2\left[R_a^2+R_c^2+(R_a+R_c)(R_p+R_q)\right],
M=2(1+ρRa)w∗2[4J∗G+16sU2].M=2(1+\rho R_a)w_*^2[4J_*G+16sU^2].

For every (x,y)∈S(x,y)\in S and ∣t∣=ρ|t|=\rho, the complexified coordinates satisfy ∣Im⁡X∣,∣Im⁡Y∣≤θ0|\operatorname{Im}X|,|\operatorname{Im}Y|\leq\theta_0 and ∣Re⁡X∣,∣Re⁡Y∣≤1+θ0|\operatorname{Re}X|,|\operatorname{Re}Y|\leq1+\theta_0. Since 0<k<10<k<1, the bounds ∣sin⁡(kX)∣≤cosh⁡(θ0)|\sin(kX)|\leq\cosh(\theta_0) and ∣cosh⁡(kY)∣≤cosh⁡(1+θ0)|\cosh(kY)|\leq\cosh(1+\theta_0) give ∣u∣,∣v∣≤U|u|,|v|\leq U, and the analytically continued energy densities have moduli at most GG. Likewise, ∣J∣≤J∗|J|\leq J_* on SS and each continued edge speed has modulus at most w∗w_*. Therefore,

∣Ef∣≤4J∗G,∣Bf∣≤8w∗U2,∣mf∣≤8w∗U,∣l∣≤w∗.|E_f|\leq4J_*G,\qquad|B_f|\leq8w_*U^2,\qquad|m_f|\leq8w_*U,\qquad|l|\leq w_*.

whence ∣T(tζ)∣≤M|T(t\zeta)|\leq M on ∣t∣=ρ|t|=\rho. By Cauchy’s estimate, the Taylor tail of degree ≥N0+1\geq N_0+1 divided by Q(z)Q(z) is bounded uniformly by

Mρ−(N0+1)(1−1/ρ)min⁡iκiRi2.\frac{M\rho^{-(N_0+1)}}{(1-1/\rho)\min_i\kappa_iR_i^2}.

Whenever the sum of the relative bounds from degrees 3 through N0N_0 and this Cauchy remainder is less than 1, it follows that T(z)<0T(z)<0 for every nonzero zz in the box.

We choose the parameters

(Ra,Rc,Rp,Rq)=(0.21,0.045,0.20,0.28),N0=22,ρ=1.82.(R_a,R_c,R_p,R_q)=(0.21,0.045,0.20,0.28),\qquad N_0=22,\qquad\rho=1.82.

Evaluating the Taylor coefficients and remainder estimate in Arb ball arithmetic shows that terms of degrees 3 through 22 contribute at most 0.5553Q0.5553Q, that M<5,240M<5,240, and that the Cauchy tail contributes at most 0.2140 QQ. Their sum lies below 0.7693 QQ, yielding

T(a,c,p,q)≤−0.23Q(a,c,p,q)<0(6)T(a,c,p,q) \le-0.23Q(a,c,p,q) < 0 \tag*{(6)}

at every nonzero point of the central box. The same enclosure for kk verifies that 8s>11/28s > 11/2.

Exterior region: affine and harmonic quadratic trial functions

Retaining s=ktanh⁡ks=k\tanh k and combining the symmetry reduction of § 2 with (5), it remains to consider convex quadrilaterals in the parameter rectangle

0≤a≤0.21,0≤c≤0.045,0≤p≤1.21,−1≤q≤10 \le a \le0.21,\qquad0 \le c \le0.045,\qquad0 \le p \le1.21,\qquad-1 \le q \le1

lying outside the central box p≤0.20p \le0.20, ∣q∣≤0.28|q| \le0.28. On this complement, it suffices to exhibit a trial function f∈H1(Ω)f \in H^1(\Omega) satisfying

11Var⁡μ(f)−2Ef>0,(7)11\operatorname{Var}_{\mu}(f)-2E_f > 0, \tag*{(7)}

for then (1) implies Σ(Ω)<11/2<8s\Sigma(\Omega) < 11/2 < 8s.

Under the bilinear map (2), the four edges of Ω\Omega are parametrized by Mi+Viτ\mathbf{M}_i+\mathbf{V}_i\tau for −1≤τ≤1-1 \le\tau\le1 (i=1,2,3,4i=1,2,3,4), where the pairs (Mi,Vi)(\mathbf{M}_i,\mathbf{V}_i) are (up to orientation)

(f,e+h),(−f,e−h),(e,f+h),(−e,f−h).(\mathbf{f},\mathbf{e}+\mathbf{h}),\qquad(-\mathbf{f},\mathbf{e}-\mathbf{h}),\qquad(\mathbf{e},\mathbf{f}+\mathbf{h}),\qquad(-\mathbf{e},\mathbf{f}-\mathbf{h}).

Set wi=∣Vi∣w_i=|\mathbf{V}_i| and W=∑i=14wiW=\sum_{i=1}^{4}w_i, so that L=2WL=2W.

Affine trial functions

For a constant vector d∈R2\mathbf{d}\in\mathbb{R}^2, the linear function f(Z)=d⋅Zf(\mathbf{Z})=\mathbf{d}\cdot\mathbf{Z} has Dirichlet energy Ef=4Δ∣d∣2E_f=4\Delta|\mathbf{d}|^2.

Given a scalar shift b0∈Rb_0\in\mathbb{R}, put Ai=d⋅Mi−b0A_i=\mathbf{d}\cdot\mathbf{M}_i-b_0 and Bi=d⋅ViB_i=\mathbf{d}\cdot\mathbf{V}_i. Since 12∫−11τ2 dτ=1/3\frac{1}{2}\int_{-1}^{1}\tau^2\,\mathrm{d}\tau=1/3, condition (7) is equivalent to

F1:=∑i=14wi(33Ai2+11Bi2−24Δ∣d∣2),WF1>33(∑i=14wiAi)2,(8)F_1:=\sum_{i=1}^{4}w_i(33A_i^2+11B_i^2-24\Delta|\mathbf{d}|^2),\qquad WF_1>33\left(\sum_{i=1}^{4}w_iA_i\right)^2, \tag*{(8)}

because the difference between the left- and right-hand sides equals 3W2(11Var⁡μ(f)−2Ef)3W^2(11\operatorname{Var}_{\mu}(f)-2E_f).

Harmonic quadratic trial functions

Let d∈R2\mathbf{d}\in\mathbb{R}^2 and let K=(r1r2r2−r1)K=\begin{pmatrix}r_1&r_2\\r_2&-r_1\end{pmatrix} be a traceless symmetric 2×22\times2 matrix, and consider the harmonic quadratic polynomial

f(Z)=d⋅Z+ZTKZ.f(\mathbf{Z})=\mathbf{d}\cdot\mathbf{Z}+\mathbf{Z}^{T}K\mathbf{Z}.

Its gradient is ∇f(Z)=d+2KZ\nabla f(\mathbf{Z})=\mathbf{d}+2K\mathbf{Z}, and K2=(r12+r22)IK^2=(r_1^2+r_2^2)I. Integrating (3) over the symmetric square SS yields the three physical area moments

∫Ω1=4Δ,∫ΩZ=43(jxe+jyf),\int_{\Omega}1=4\Delta,\qquad\int_{\Omega}\mathbf{Z}=\frac{4}{3}(j_x\mathbf{e}+j_y\mathbf{f}),
9∫Ω∣Z∣2=4Δ(3∣e∣2+3∣f∣2+∣h∣2)+8(jxf⋅h+jye⋅h),9\int_{\Omega}|\mathbf{Z}|^2=4\Delta(3|\mathbf{e}|^2+3|\mathbf{f}|^2+|\mathbf{h}|^2)+8(j_x\mathbf{f}\cdot\mathbf{h}+j_y\mathbf{e}\cdot\mathbf{h}),

from which the Dirichlet energy follows explicitly as

Ef=4Δ∣d∣2+4(Kd)⋅∫ΩZ+4(r12+r22)∫Ω∣Z∣2.E_f=4\Delta|\mathbf{d}|^2+4(K\mathbf{d})\cdot\int_{\Omega}\mathbf{Z}+4(r_1^2+r_2^2)\int_{\Omega}|\mathbf{Z}|^2.

Along edge ii, write f(Mi+Viτ)−b0=Ai+Biτ+Ciτ2f(\mathbf{M}_i+\mathbf{V}_i\tau)-b_0=A_i+B_i\tau+C_i\tau^2, where

Ai=d⋅Mi+MiTKMi−b0,Bi=d⋅Vi+2MiTKVi,Ci=ViTKVi.A_i=\mathbf{d}\cdot\mathbf{M}_i+\mathbf{M}_i^{T}K\mathbf{M}_i-b_0,\qquad B_i=\mathbf{d}\cdot\mathbf{V}_i+2\mathbf{M}_i^{T}K\mathbf{V}_i,\qquad C_i=\mathbf{V}_i^{T}K\mathbf{V}_i.

Put ϕi=3Ai+Ci\phi_i = 3A_i + C_i and ψi=5ϕi2+15Bi2+4Ci2\psi_i = 5\phi_i^2 + 15B_i^2 + 4C_i^2. Direct integration over τ∈[−1,1]\tau\in[-1,1] shows that the edge averages of f−b0f-b_0 and (f−b0)2(f-b_0)^2 equal ϕi/3\phi_i/3 and ψi/45\psi_i/45, respectively. Clearing denominators in (7) yields the equivalent polynomial inequality

F2:=∑i=14wi(11ψi−90Ef),WF2>55(∑i=14wiϕi)2(9)F_2 := \sum_{i=1}^{4} w_i(11\psi_i - 90E_f), \qquad WF_2 > 55\left(\sum_{i=1}^{4}w_i\phi_i\right)^2 \tag*{(9)}

whose left-hand side minus its right-hand side equals 45W2(11Var⁡μ(f)−2Ef)45W^2(11\operatorname{Var}_{\mu}(f)-2E_f).

Subdivision of the exterior region

We decompose the exterior of the central box inside [0,0.21]×[0,0.045]×[0,1.21]×[−1,1][0,0.21] \times[0,0.045] \times[0,1.21] \times[-1,1] into six closed root boxes sharing the same (a,c)(a,c) intervals (Table 1): the first three cover the complement of [0,0.40]×[−0.60,0.60][0,0.40] \times[-0.60,0.60] in the (p,q)(p,q) rectangle, and the remaining three cover the portion of [0,0.40]×[−0.60,0.60][0,0.40] \times[-0.60,0.60] outside [0,0.20]×[−0.28,0.28][0,0.20] \times[-0.28,0.28]. Because every box in the partition is closed, all shared boundary faces are accounted for.

A subbox is eliminated if outward-rounded interval enclosures establish either ∣jx∣+∣jy∣>Δ|j_x|+|j_y|>\Delta throughout the subbox (so that it contains no admissible convex quadrilateral), or one of the spectral inequalities (5), (8), or (9). All interval endpoints are integer multiples of 2−322^{-32}, with rational root endpoints rounded outward; interval products use integer floor and ceiling, and square roots use integer square roots rounded outward. On each subbox, a midpoint covariance computation or generalized eigenproblem in floating point suggests candidate affine or harmonic quadratic coefficients, which are rounded to integers (with b0b_0 rounded to 2−32Z2^{-32}\mathbb{Z}). Once these explicit trial parameters are fixed, the verification of F1>0F_1>0 and (8), or F2>0F_2>0 and (9), is carried out entirely in integer interval arithmetic and does not rely on the accuracy of the floating-point step.

When none of the four criteria succeeds on a subbox, we bisect the coordinate of largest weighted width using weights (2,2,1,1)(2,2,1,1) on (a,c,p,q)(a,c,p,q), keeping the midpoint in both closed child subboxes. Across the six root boxes, 934 bisections suffice, yielding a closed cover of 940 subboxes (Table 1) on every one of which an elimination criterion holds.

Completion of the proof of Theorem 1. Every triangle satisfies (4), and 11/2<8s11/2<8s. For a convex quadrilateral Ω\Omega, pass to the normalized chart (2) with a,c,p≥0a,c,p\geq0. If a≥0.21a\geq0.21 or c≥0.045c\geq0.045, inequality (5) yields Σ(Ω)<11/2<8s\Sigma(\Omega)<11/2<8s. Otherwise, convexity restricts (p,q)(p,q) to [0,1.21]×[−1,1][0,1.21]\times[-1,1]. On the central box p≤0.20p\leq0.20, ∣q∣≤0.28|q|\leq0.28, inequality (6) gives T<0T<0 and hence Σ(Ω)<8s\Sigma(\Omega)<8s whenever (a,c,p,q)≠(0,0,0,0)(a,c,p,q)\ne(0,0,0,0). Every remaining admissible tuple (a,c,p,q)(a,c,p,q) lies in at least one of the 940 closed subboxes of Table 1, where again Σ(Ω)<11/2<8s\Sigma(\Omega)<11/2<8s. At (a,c,p,q)=(0,0,0,0)(a,c,p,q)=(0,0,0,0), the domain is the reference square S=[−1,1]2S = [-1,1]^2 with Σ(S)=8s\Sigma(S) = 8s. Undoing the similarity normalization shows that equality holds precisely for squares.

ppqqAffineDiagonalInfeasibleQuadraticTotal
[0.40,1.21][0.40,1.21][−1,1][-1,1]6508073
[0,0.40][0,0.40][0.60,1][0.60,1]6090069
[0,0.40][0,0.40][−1,−0.60][-1,-0.60]1873610224
[0.20,0.40][0.20,0.40][−0.60,0.60][-0.60,0.60]3700180217
[0,0.20][0,0.20][0.28,0.60][0.28,0.60]1910107127
[0,0.20][0,0.20][−0.60,−0.28][-0.60,-0.28]1930208230
Total387499495940

Table 1. Subboxes in the exterior partition, classified by elimination criterion. All six root boxes share a∈[0,0.21]a \in[0,0.21] and c∈[0,0.045]c \in[0,0.045].

Concluding remarks

We close with three observations concerning stability, possible extensions, and the computational implementation.

Quantitative stability. Although Theorem 1 is stated as a sharp inequality with equality characterization, the proof yields a quantitative stability estimate near the square. Indeed, for any f∈H1(Ω)f \in H^1(\Omega) we have

Hf=ℓ2(Ef−8sVar⁡μ(f))≥ℓ2Var⁡μ(f)(Σ(Ω)−8s),H_f = \ell^2\left(E_f - 8s\operatorname{Var}_{\mu}(f)\right) \ge\ell^2\operatorname{Var}_{\mu}(f)\left(\Sigma(\Omega)-8s\right),

so (6) implies that

8s−Σ(Ω)≥0.23Q(a,c,p,q)ℓ2[(1+a)Var⁡μ(u)+(1−a)Var⁡μ(v)]8s-\Sigma(\Omega) \ge\frac{0.23Q(a,c,p,q)}{\ell^2\left[(1+a)\operatorname{Var}_{\mu}(u)+(1-a)\operatorname{Var}_{\mu}(v)\right]}

throughout the central box. Because the denominator is continuous and bounded above on the compact central box, while 8s−Σ(Ω)>8s−11/2>08s-\Sigma(\Omega)>8s-11/2>0 outside it and a2+c2+p2+q2a^2+c^2+p^2+q^2 is uniformly bounded across all convex quadrilaterals (since a2+c2<1a^2+c^2<1 and h∈conv⁡(±e,±f)\mathbf{h}\in\operatorname{conv}(\pm\mathbf{e},\pm\mathbf{f})), the Steklov deficit 8ktanh⁡k−Σ(Ω)8k\tanh k-\Sigma(\Omega) controls a2+c2+p2+q2a^2+c^2+p^2+q^2 in the normalized chart (2) up to an explicit positive constant.

Nonconvex quadrilaterals and higher polygons. Convexity enters our proof in two places: the bilinear diffeomorphism (2) and the interior-diagonal decomposition of Lemma 3. It is natural to ask whether Theorem 1 remains valid for nonconvex simple quadrilaterals. Note that reflecting the two edges incident to a reflex vertex across the line joining its neighbors strictly increases the area while preserving the perimeter, so the quadrilateral isoperimetric bound ∣Ω∣/L<1/4\sqrt{|\Omega|}/L<1/4 still holds for all nonconvex simple quadrilaterals; consequently, any extension of Theorem 1 to nonconvex quadrilaterals would automatically carry Corollary 2 along with it. For convex NN-gons with N≥5N\ge5, two new obstacles arise: the dimension 2N−42N-4 of the moduli space grows with NN, and explicit closed-form Steklov eigenfunctions are no longer available for regular NN-gons when N≥5N\ge5, so a local analysis near the regular NN-gon would require rigorous approximations to the first Steklov eigenfunctions together with a lower bound on the spectral gap σ3−σ1\sigma_3-\sigma_1.

Computational details. The finite calculations of § 3 and § 4 are performed by compact, self-contained Python programs. For the local box, the recurrences of § 3.1 and the Cauchy tail estimate of § 3.2 are evaluated using Arb ball arithmetic [8] at 200-bit working precision, isolating kk to an interval of width at most 2−1502^{-150}. For the exterior region, the subdivision of § 4.3 uses exact integer interval arithmetic on 2−32Z2^{-32}\mathbb{Z}. All analytic steps connecting these finite inequalities to the Steklov spectrum are contained in § 2–§ 4. Soochow University, Suzhou, China

Nanjing University of Information Science and Technology, Nanjing, China

Xiamen University, Xiamen, China

References

  1. [1]American Institute of Mathematics, Steklov eigenproblems, 2018 workshop problem list, Section 4, Problem 4.6 (É. Oudet).
  2. [2]F. Brock, An isoperimetric inequality for eigenvalues of the Steklov problem, Z. Angew. Math. Mech. 81 (2001), no. 1, 69–71.doi.org/10.1002/1521-4001(200101)81:1<69::aid-zamm69>3.0.co;2-#
  3. [3]D. Bucur, V. Ferone, C. Nitsch, and C. Trombetti, Weinstock inequality in higher dimensions, J. Differential Geom. 118 (2021), no. 1, 1–21.arxiv.org/abs/1710.04587
  4. [4]Z. Cheng, C. Gui, Y. Hu, Q. Li, and R. Yao, Global strict monotonicity and asymptotics of the first Steklov eigenvalue for regular polygons, preprint (2026), arXiv:2603.25116v2.arxiv.org/abs/2603.25116
  5. [5]B. Colbois, A. Girouard, C. Gordon, and D. Sher, Some recent developments on the Steklov eigenvalue problem, Rev. Mat. Complut. 37 (2024), no. 1, 1–161.arxiv.org/abs/2212.12528
  6. [6]A. Girouard, J. Lagacé, I. Polterovich, and A. Savo, The Steklov spectrum of cuboids, Mathematika 65 (2019), no. 2, 272–310. arXiv:1711.03075.DOI
  7. [7]A. Girouard and I. Polterovich, Spectral geometry of the Steklov problem, J. Spectr. Theory 7 (2017), no. 2, 321–359.
  8. [8]F. Johansson, Arb: efficient arbitrary-precision midpoint-radius interval arithmetic, IEEE Trans. Comput. 66 (2017), no. 8, 1281–1292. doi:10.1109/TC.2017.2690633.DOI
  9. [9]R. Weinstock, Inequalities for a classical eigenvalue problem, J. Rational Mech. Anal. 3 (1954), 745–753.

Paper details

vlnc:W_7kh5sxch
Version
v1 · latest
Published
Fresh — published within the last 30 days
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