WorksMathematics
The equilateral triangle maximises the first Steklov eigenvalue among triangles
Abstract
Let be the first nonzero Steklov eigenvalue of a triangle and its perimeter. Among triangles the equilateral triangle uniquely maximises . If the shortest and longest sides of have lengths , then . Since is also maximal at , the equilateral triangle uniquely maximises as well. This settles the case of the Steklov analogue of the Pólya–Szegő conjecture for -gons, for both normalisations. The extremal value satisfies . For side length one, . Under displacement of the apex over a fixed unit base, the first two perimeter-normalised eigenvalues split linearly with equal and opposite slopes independent of direction. The proof is computer-assisted. Near we transplant the first eigenspace of affinely; away from the trial function suffices. We verify the finite matrix and interval inequalities in ball arithmetic.
AI transparency statement. The main result, its computer-assisted proof, the partial Lean formalisation and the manuscript were produced by artificial intelligence. The human author initiated the broader research and directed the AI systems, but did not write the text and has not checked the arguments. Details are given at the end of the paper.
Introduction
For a bounded Lipschitz domain , the Steklov problem in , on has eigenvalues . They satisfy
The scale-invariant quantity is
Weinstock’s inequality [16] states that for simply connected planar domains, with equality only for the disk; see the survey [10]. For polygons with a prescribed number of sides, the natural conjecture is the analogue of Pólya–Szegő’s conjecture for the Dirichlet Laplacian:
Problem (Oudet; AIM problem list Steklov eigenproblems, Problem 4.6 [1]). Do regular -gons maximise among -gons of given area, or of given perimeter?
Dominguez, Nigam and Shahriari [7] gave numerical evidence for triangles using finite elements and Bayesian optimisation. Nigam also posed the perimeter question as Open Problem 8 at the 2020 miniconference on sharp eigenvalue estimates [13]. Among rectangles of given perimeter or area, the square maximises : this is the planar case of the cuboid inequality of Girouard, Lagacé, Polterovich and Savo [9 Thm. 1.6]. Cheng, Gui, Hu, Li and Yao [4] state the conjecture for convex polygons with at most sides at fixed perimeter (their Conjecture 1.1), and leave it open for every . They prove that of the perimeter-normalised regular -gon increases strictly with , and that its first eigenspace is two-dimensional. Dryden, Gordon, Moreno, Rowlett and Villegas-Blas [8] prove for triangles, where is the smallest angle. This shows that degenerate triangles are far from optimal, but it does not compare triangles with .
For the Neumann problem, Laugesen and Siudeja [11] proved that the equilateral triangle maximises the first nonzero Neumann eigenvalue times among triangles, and hence times the area. Near the equilateral triangle their trial functions are the eigenfunctions of the equilateral triangle, transplanted by the linear map onto the given triangle. This affine transplantation is the precedent for our local argument. The Steklov problem adds two difficulties: the first eigenvalue of is double, so the maximum is a cone and the comparison degenerates at ; and a rigorous lower bound for is needed. We are not aware of a proof of the Steklov case for any .
Theorem 1.1. Let be an equilateral triangle. For every triangle ,
with equality if and only if is equilateral. More precisely, normalise a longest side to and write the apex as with and . If , then
where and are the extreme side lengths of . Moreover
Girouard and Polterovich computed the first sixty normalised eigenvalues of the equilateral triangle numerically [10 Example 3.2.1]. For side length one we obtain (Proposition 3.4). The double eigenvalue agrees with [4 Thm. 1.2] for .
Corollary 1.2. For every triangle ,
with equality if and only if is equilateral.
Proof. . By Heron’s formula and the AM–GM inequality, , with equality only for equilateral triangles. Now apply Theorem 1.1.
Method
Upper bounds for come directly from (1); a lower bound is needed only for the single domain . We parametrise triangles by the apex over the base (Lemma 2.3), and let be the distance of from the equilateral apex.
Near (). The trial space is the exact first eigenspace of , transplanted by the affine map . The comparison involves only a few moments of the eigenspace and two exact identities. These identities let us divide out the factor (Lemma 4.2), which turns the degenerate comparison at into a sign condition. We check the sign condition on boxes in ball arithmetic. The moments are enclosed using a harmonic-polynomial approximation, a residual bound for the Dirichlet-to-Neumann map, and spectral isolation of a cubic eigenfunction by a Crouzeix–Raviart bound for .
Away from (). The single trial function gives , well below . A one-dimensional check on 16 intervals suffices.
Preliminaries
The Dirichlet-to-Neumann map
Let be a bounded Lipschitz domain. The Dirichlet-to-Neumann map is the self-adjoint operator on associated with the closed form on , where is the harmonic extension. It has compact resolvent and eigenvalues , with an orthonormal basis of eigenvectors of ; is constant. If is a harmonic polynomial with trace , Green’s formula gives and . For and write .
Lemma 2.1 (Inclusion). Let have dimension , and suppose for all . Then has at least eigenvalues, counted with multiplicity, in .
Proof. Suppose not. Then some is orthogonal to the spectral subspace of , and
a contradiction.
Lemma 2.2 (Eigenvector enclosure; cf. [6]). Let be the spectral projection of onto . Suppose every eigenvalue of outside is either or at least , and put . Let be orthogonal to the constants, and . Then
Proof. for every , and the component along vanishes. The remaining with have , so , and
because is decreasing for when (its derivative is ).
Parametrisation
For with , write
and set
Since is invariant under similarities, it suffices to treat the triangles , .
Lemma 2.3. Every triangle is similar to for some , and is equilateral if and only if .
Proof. Rescale so that a longest side has length 1. Move it to with the third vertex in the upper half-plane, and reflect in if necessary so that . The side from to the apex is not longer than the base, so ; this forces . The other side then satisfies . The equilateral case is clear.
We write , for polar coordinates about the equilateral apex; is the Euclidean distance in the -plane.
Affine transplantation
Let , and . Write . For an edge of a polygon with unit tangent , put . Changing variables gives
because the arc length on is times that on . For a finite-dimensional space of trial functions, define the forms and on . If is positive definite on , then (1) gives .
The equilateral triangle
Let , with side 1 and . Let be its centroid and , so that the circumradius is 1 in and the apex lies on the positive real -axis. The dihedral group acts by and .
An exactly equivariant trial pair
Given real coefficients (, , fourteen in all), put , and . Rotation by multiplies by , and reflection conjugates it. Hence transforms under exactly as the coordinates do, whatever the are. In particular ; each is the real part of a holomorphic polynomial (since and ), and by Schur’s lemma all their Gram matrices are scalar. The come from a floating-point Ritz computation and are then fixed as exact binary numbers.
Let be the Rayleigh quotient of , rounded to binary64. All boundary integrals of products of and are integrals of polynomials along segments, and we compute them exactly in ball arithmetic. This gives the relative residual
By Lemma 2.1, has at least two eigenvalues in .
A Crouzeix–Raviart lower bound for
Subdivide uniformly into equilateral cells of side . Let be the Crouzeix–Raviart space [5]: piecewise linear functions that are continuous at the midpoints of interior edges. Let be the interpolation by edge means, . Write and , and let be the finite eigenvalues of the pencil on . The min–max principle holds over subspaces on which is positive definite.
Lemma 3.1 (Trace constant). Let be an equilateral triangle of side , an edge of , and with . Then
Proof. Let . Since has mean zero on , for every constant ; in particular . Let be the vertex opposite . The field has normal component on the two edges through , and times on . Its divergence is . The divergence theorem therefore gives
By the Payne–Weinberger inequality [15, 3] on the convex set of diameter , . Finally .
Lemma 3.2 (Projection bound; cf. [14, 17]). Let and . Then for each below the number of boundary edges,
Proof. (i) For every , . Indeed is constant on each cell , and because is linear and has the same edge means as .
(ii) Write . It has mean zero on every edge of every cell. Each boundary edge lies in one cell, and for a cell has at most two boundary edges (exactly two at the corners). Lemma 3.1 then gives .
(iii) Let be the span of the eigenfunctions of , and take with . Put . By (i), , and by (ii), . If the claim is trivial, since . Otherwise is injective on , and
The maximum is attained at . The min–max principle on the -dimensional space gives , which rearranges to the claim.
Proposition 3.3. .
Proof. Take , which gives 900 degrees of freedom and . Let and be the matrices of and in the edge basis of . On an equilateral cell the element stiffness matrix is
@ On a boundary edge, the basis function of that edge is and the other two are the linear function rising from to along the edge, so the element boundary-mass entries are , and . Let be a fixed matrix with binary64 entries (computed once as approximate eigenvectors of ). In ball arithmetic we check two things:
the Gershgorin discs of exclude , so is nonsingular;
the Gershgorin discs of consist of discs in and discs in .
Gershgorin’s counting theorem and Sylvester’s law of inertia give exactly four negative eigenvalues of . If , the pencil would be negative definite at this shift on the span of five discrete eigenfunctions, including the constant. Thus , and Lemma 3.2 gives .
Proposition 3.4. For the side-one equilateral triangle,
In particular the third positive eigenvalue is simple and
Proof. Use Cartesian coordinates centred at the centroid and put . The harmonic polynomial satisfies on . Rotation by preserves , so the same boundary condition holds on every edge. Thus is an eigenvalue. The inclusion interval lies strictly between and and contains at least two eigenvalues counted with multiplicity. Together with and these account, with multiplicity, for all eigenvalues below the lower bound in Proposition 3.3. Hence contains precisely , and is simple. Take in Lemma 2.2. Let be the spectral projection onto . It commutes with the -action, so is again an equivariant pair. Since , Lemma 2.2 gives ; since with equal norms, are linearly independent. They therefore span the two-dimensional range of , which carries the irreducible representation of . By Schur’s lemma acts on this range as a scalar.
Local maximality
Let be the eigenspace of , with basis (). For , , , we have with
Number the edges , and . Then , while
are the side lengths of , and .
The moments of are the following matrices and vectors, indexed by : the energy matrices , the edge Gram matrices , and the edge-mean vectors . They satisfy two exact identities:
The first holds because eigenfunctions of with are orthogonal to constants. The second is the weak eigen-equation on .
Lemma 4.1 (Variational comparison). Let a trial space on a polygon have energy matrix and positive definite boundary-variance matrix . For a reference value put . Then
In particular, if has a positive direction, the deficit exceeds .
Proof. Apply (1) to each trial function and subtract the resulting bound from .
Apply the lemma with , and on . At the eigen-equation gives . The following factorisation isolates the leading term.
Proof. First, , so .
Next, , which gives .
For the boundary terms, by (3). Also , using .
Hence , and . Subtracting these two equations and using (3) gives .
Lemma 4.3 (Moment enclosures). Let and . Then
Consequently each moment of differs from the corresponding moment of by at most the following amounts:
for the entry of ;
for the entry of ;
for the entry of .
Proof. Apply Lemma 2.2 to , with ; its hypotheses were verified in the proof of Proposition 3.4. The moment bounds follow from Cauchy–Schwarz.
Here , so and . We compute the moments of exactly by Green’s formula and for holomorphic . This turns every area integral into exact polynomial line integrals.
Proposition 4.4. Let be the binary64 number nearest . For and every direction, .
Proof. We cover by 907 closed boxes, where is the binary64 number nearest 6.3 and exceeds . The boxes lie in the rectangle, have disjoint interiors, and their exact rational areas sum to its area. They cover the rectangle since a nonempty relatively open complement would have positive area. We evaluate and on every box using Lemma 4.3. For the checks give
Thus has a positive direction. Since , Lemma 4.1 gives a deficit greater than . The other two sides have total length greater than the unit base, so .
Theorem 4.5 (Splitting of the first pair). For a fixed unit base let the apex move from by . As , uniformly in ,
where the exact constant satisfies .
Proof. Work on with the Dirichlet inner product. On this fixed space let and be the pulled-back forms of (2), now acting on all of . Their representing operators are real analytic in the Cartesian components of : the energy coefficients are rational in , and the edge lengths and perimeter are positive near . The operator is uniformly positive there. The boundary variance is unchanged by adding constants and the trace is compact into , so is compact. Positive eigenvalues of are the reciprocals of the normalised positive Steklov eigenvalues, by the weak equation on the quotient (choose the representative of zero weighted boundary mean). The inverse square root is analytic near by its power series.
At the top eigenvalue of is isolated and double by Proposition 3.4. The resolvent on a small enclosing circle has a convergent Neumann series for small . The Riesz projection defined by this contour is real analytic and has rank two by continuity. Project a basis at 0 and apply to obtain an analytic frame of the corresponding pencil subspace. Multiplication by the inverse square root of its positive -Gram matrix makes the frame -orthonormal. In this frame the matrix of is symmetric and real analytic, with eigenvalues . Write
where is linear and the remainder is in matrix norm, uniformly in direction.
For a unit vector the matrix represents on the unperturbed eigenspace in a -orthonormal basis. Indeed, differentiating the moving frame introduces two further terms, each zero by the weak eigen-equation against every vector of . In the original basis the same derivative is , since on that fixed space. Consequently
To see why the slope is independent of , identify the apex with . Cyclic relabelling of the vertices followed by a similarity replaces by . It fixes and . Thus the two normalised eigenvalues agree as an unordered pair at and . Both the trace and the determinant of are therefore invariant under . Comparing the terms of degree one in the trace and degree two in the determinant shows that and are invariant under rotation by . The only invariant linear form is zero, and every invariant real quadratic form is a multiple of (its symmetric matrix commutes with that rotation). Hence has eigenvalues , where
Ball evaluation using the enclosed moments of the projected eigenfunctions gives the stated positive enclosure for . The min–max principle bounds the change in each ordered eigenvalue by the norm of the matrix remainder, proving both expansions.
The far region and the proof of the theorem
For put
Lemma 5.1. (i) . (ii) For , with and ,
where
and .
Proof. (i) For we have , and means , where the right side is positive because . Since , this is equivalent to .
(ii) Take in (1). The vertical section of at abscissa has length for and for , so the substitution on gives
The antiderivative and the identity give the closed form of . The denominator of (1) is . On the side from to arc length is with , and similarly on the third side. So for every constant ,
Hence with .
Proposition 5.2. For every with , .
Proof. By Lemma 5.1(i) these are the points with , . The bound of Lemma 5.1(ii) is nondecreasing in , and . On an interval , the first entry of is increasing in and the second is decreasing, so
Moreover and . Hence on that interval
We evaluate the right side in ball arithmetic on the 16 intervals , , which cover . Each upper endpoint is below .
Proof of Theorem 1.1 Use the normalisation of Lemma 2.3. The deficit vanishes at and exceeds for by Proposition 4.4. For , Proposition 5.2 gives a deficit at least . Throughout we have and , so . Finally in these coordinates, and the reverse triangle inequality gives . This proves both deficit bounds and the equality case. The extremal value is enclosed in Proposition 3.4.
Corollary 5.3 (Fixed-profile thin limit). Fix , and set on . Define
Then and as .
Proof. In Henrot and Michetti’s thin-domain lemma [12 Lemma 3.2] take the upper profile , the lower profile zero, and thickness . Their lower-growth hypothesis holds since . Their first one-dimensional eigenvalue is , and their asymptotic formula multiplies it by . Now use .
Implementation
All certified steps use Arb [2] (python-flint 0.9.0), at 64–200 bits.
Certificate data. Floating-point arithmetic is used only to produce candidate data, which are then fixed and treated as exact binary numbers: the coefficients and the congruence matrix . These are committed to the repository, and the certified steps read only them. The shift is computed from the in ball arithmetic and rounded to binary64. Hence the certified results do not depend on the linear-algebra library or the processor; regenerating the data is a separate, optional step. Every number of this paper that a computation checks is kept in a single claims file. The checking script compares each computed enclosure with these claims in exact rational arithmetic, and the paper’s macros are generated from the same file.
Boxes. Box endpoints are exact binary numbers. A parameter box enters as the Arb union of its two exact endpoints, which contains the whole interval. Square roots of nonnegative quantities whose balls straddle 0 are clamped to . In the code, is the binary double nearest 0.6, which is , slightly below 0.6. The local and the far computations use this same value, so the two regions meet exactly. The tiling of the local boxes is checked with exact float comparisons and an exact rational area sum.
Consistency checks (not part of the proof).
The Green-formula moments agree with brute-force quadrature on random harmonic polynomials to .
The transplant bound of § 5, and the full Rayleigh quotient of , agree with direct quadrature on to about .
At 20 sample triangles both bounds lie above the floating-point Ritz value of , as upper bounds must.
The Crouzeix–Raviart discrete values are for and for , increasing towards .
Code names. Some identifiers differ from the notation here. In the Python code, is sig, the edge means are mm, are d1, d2, the are ce, and is D. In Lean, the are m0x, ..., m2y, the are c0, c1, c2, and are NA, DA.
Lean. A Lean 4/Mathlib development formalises the elementary steps:
the triangle isoperimetric inequality in Heron form, with its equality case, and the deduction of Corollary 1.2;
the algebra of Lemma 4.2;
the integral defining in Lemma 5.1, and the two edge integrals of used in the consistency checks.
It uses no sorry, and its only axioms are propext, Classical.choice and Quot.sound. It does not formalise the spectral theory, the trace and Payne–Weinberger inequalities, or the ball-arithmetic computations.
Availability. The code, the logs, the box certificates and the Lean files are in the accompanying repository2. The single command scripts/verify.py reruns all of them and writes a PASS/FAIL summary.
Numerical observations (not rigorous)
Except for the certified upper bounds in Table 1, the statements in this section are numerical. They use the harmonic-polynomial Ritz method, with trial space and all integrals computed on the boundary.
| polygon | certified upper bound for | estimate of |
| 60–60–60 | 3.8724615957729 | 3.8724615957728 |
| 50–50–80 | 2.8777309857377 | 2.87773098574 |
| 70–70–40 | 2.5730765728918 | 2.573076572892 |
| 40–60–80 | 2.5350358150935 | 2.53503581509 |
| 45–45–90 | 2.4960723700476 | 2.4960723700475 |
| 30–60–90 | 1.9013817577256 | 1.9013817577255 |
| 30–30–120 | 1.5603440800120 | 1.5603441 |
| square | 5.5060219386902 | 5.5060219386901 |
| regular pentagon | 5.9774725550439 | 5.9774726 |
| regular hexagon | 6.1355701455936 | 6.135570 |
Table 1. Values of . The values are heuristic error estimates. Rounding the central estimate to the displayed precision can place it slightly above the certified upper bound.
Observation 7.1 (Global picture). On a grid over (double precision, ) the largest value of is attained at the equilateral apex. Table 1 lists selected values, sorted by , together with the regular -gons for . They are computed as follows:
the Ritz matrices are exact polynomial line integrals, evaluated in ball arithmetic;
the eigenproblem is solved in 50- and 80-digit arithmetic with the full basis of directions, for ;
the estimate is the value for ;
its error is estimated (not rigorously) as twice the largest of the last change in , a geometric and a power-law extrapolation of the remaining changes, and the change between the two precisions; the estimate is printed only to the digits this error justifies;
the certified upper bound is the Rayleigh quotient of the fixed trial function, evaluated in ball arithmetic on the exact polygon.
Convergence is super-geometric for , the square, and the right-angled triangles. It is slowest for the polygons with an angle of or (the 30–30–120 triangle, the pentagon and the hexagon), whose corners produce the strongest singularities.
The regular -gon values increase with , consistent with [4].
Observation 7.2 (Thin triangles). For the limiting endpoint profile in the quotient of Corollary 5.3, the Euler–Lagrange equation gives . Numerically , and . The double-precision Ritz values () of at agree with these to relative. In particular for thin isosceles triangles, compared with the rigorous constant 45.6 of [8].
Remarks
Lemma 4.1 applies to any polygonal trial space for which the energy and boundary-variance forms can be enclosed. At a multiple reference eigenvalue the spectral gap controls the projection error, and the identity makes verification of a linear deficit possible up to .
For the affine group no longer acts transitively on similarity classes of -gons, so one needs a richer family of transplantations, for example piecewise affine or Schwarz–Christoffel maps. Parallelograms form a two-parameter affine family of the square. General quadrilaterals have four parameters modulo similarities, so a first-order comparison of a double eigenvalue would also have to address deformations in the kernel of the linear map into symmetric matrices.
Declaration of AI-generated research and authorship
This manuscript is the product of an AI-driven research process. During a broader investigation of open problems from the AIM problem lists, initiated by the human author, an AI system identified this problem as a candidate, found the proof, and wrote the code for the certified computations. The author then directed AI systems to package the result, formalise parts of it in Lean and prepare this manuscript. The AI systems developed the results and their presentation, including the mathematical exposition, the computations, the formalisation, and the organisation and writing of the text; the human author wrote no part of the manuscript. The system was Claude Opus 5.5 (Anthropic), used through Claude Code. OpenAI Codex subsequently revised the proof and computational certificates for stability, spectral isolation and splitting, and incorporated the thin-domain corollary.
The proof is computer-assisted and is reproduced by scripts/verify.py; the Lean development covers only the elementary steps listed in § 7.
The human author has not read this manuscript or checked its arguments. At the time of writing, the mathematics, the code and the formalisation have been checked only by AI systems and by the computer verification described above; no independent human expert has reviewed them. The author takes responsibility for the accuracy of this disclosure.
References
- [1]AIM Problem Lists, Steklov eigenproblems, Section 4 (Shape optimisation for Steklov), Problem 4.6 (É. Oudet), http://aimpl.org/stekloveigen/4.
- [2]F. Johansson, Arb: efficient arbitrary-precision midpoint-radius interval arithmetic, IEEE Trans. Comput. 66 (2017), 1281–1292. doi:10.1109/TC.2017.2690633.DOI
- [3]M. Bebendorf, A note on the Poincaré inequality for convex domains, Z. Anal. Anwend. 22 (2003), 751–756. doi:10.4171/ZAA/1170.DOI
- [4]Z. Cheng, C. Gui, Y. Hu, Q. Li, R. Yao, Global strict monotonicity and asymptotics of the first Steklov eigenvalue for regular polygons, arXiv:2603.25116v2 (2026). doi:10.48550/arXiv.2603.25116.DOI
- [5]M. Crouzeix, P.-A. Raviart, Conforming and nonconforming finite element methods for solving the stationary Stokes equations I, RAIRO Anal. Numér. 7 (1973), 33–75. doi:10.1051/m2an/197307R300331.DOI
- [6]C. Davis, W. M. Kahan, The rotation of eigenvectors by a perturbation. III, SIAM J. Numer. Anal. 7 (1970), 1–46. doi:10.1137/0707001.
- [7]S. Dominguez, N. Nigam, B. Shahriari, A combined finite element and Bayesian optimization framework for shape optimization in spectral geometry, Comput. Math. Appl. 74 (2017), no. 11, 2874–2896. doi:10.1016/j.camwa.2017.08.044.DOI
- [8]E. B. Dryden, C. Gordon, J. Moreno, J. Rowlett, C. Villegas-Blas, The Steklov spectrum of convex polygonal domains I: spectral finiteness, J. Geom. Anal. 35 (2025), no. 3, Paper No. 91. doi:10.1007/s12220-025-01922-8.DOI
- [9]A. Girouard, J. Lagacé, I. Polterovich, A. Savo, The Steklov spectrum of cuboids, Mathematika 65 (2019), no. 2, 272–310. doi:10.1112/S0025579318000414.DOI
- [10]A. Girouard, I. Polterovich, Spectral geometry of the Steklov problem, J. Spectr. Theory 7 (2017), 321–359. doi:10.4171/JST/164.DOI
- [11]R. S. Laugesen, B. A. Siudeja, Maximizing Neumann fundamental tones of triangles, J. Math. Phys. 50 (2009), 112903. doi:10.1063/1.3246834.DOI
- [12]A. Henrot and M. Michetti, A comparison between Neumann and Steklov eigenvalues, J. Spectr. Theory 12 (2022), 1405–1442, doi:10.4171/JST/429.DOI
- [13]N. Nigam, Open Problem 8 in Open problems from the miniconference on sharp eigenvalue estimates for partial differential operators (2020), problem list.
- [14]X. Liu, A framework of verified eigenvalue bounds for self-adjoint differential operators, Appl. Math. Comput. 267 (2015), 341–355. doi:10.1016/j.amc.2015.03.048.DOI
- [15]L. E. Payne, H. F. Weinberger, An optimal Poincaré inequality for convex domains, Arch. Rational Mech. Anal. 5 (1960), 286–292. doi:10.1007/BF00252910.DOI
- [16]R. Weinstock, Inequalities for a classical eigenvalue problem, J. Rational Mech. Anal. 3 (1954), 745–753. doi:10.1512/iumj.1954.3.53036.DOI
- [17]C. You, H. Xie, X. Liu, Guaranteed eigenvalue bounds for the Steklov eigenvalue problem, SIAM J. Numer. Anal. 57 (2019), 1395–1410. doi:10.1137/18M1189592.DOI