Spectral properties of the Dirichlet-to-Neumann map for the Helmholtz equation

Abstract: The study of the Dirichlet-to-Neumann map and the associated Steklov problem for the Laplace equation has been a central topic in spectral geometry over the past decade. In this survey, we consider a more general framework in which the Laplace equation is replaced by the Helmholtz equation. We examine how the properties of the Dirichlet-to-Neumann eigenvalues and eigenfunctions depend on the parameter in the Helmholtz equation and describe new phenomena arising when this parameter is nonzero, as opposed to the Laplace case. In particular, we present various eigenvalue inequalities, analyse spectral asymptotics in different regimes, and investigate nodal domains and other features of eigenfunctions. We also discuss applications where the Helmholtz parameter plays an essential role, as well as challenges encountered in the numerical computation of the Dirichlet-to-Neumann spectrum.

• The paper develops an explicit matrix formula expressing the Helmholtz DtN map at any parameter through a reference DtN spectrum and Dirichlet eigenfunction data, enabling analytic perturbation analysis.
• The paper proves that DtN eigenvalues decrease strictly between Dirichlet eigenvalues and that their spectra form real-analytic branches, resolving an open question on Robin eigenvalue limits.
• The paper unifies asymptotic, nodal, positivity, geometric, and numerical results while identifying open problems involving rough boundaries, sharp inequalities, boundary nodal counts, and polygonal domains.

This survey by Grebenkov, Levitin, and Polterovich (2604.11526) develops the spectral theory of the Dirichlet-to-Neumann (DtN) map $\mathcal{D}_\Lambda$ associated with the Helmholtz equation $-\Delta U = \Lambda U$ on a bounded Lipschitz domain $\Omega \subset \mathbb{R}^d$, generalising the classical Steklov problem ($\Lambda=0$) to arbitrary real values of the Helmholtz parameter. The paper is a survey with new results: an explicit matrix representation of $\mathcal{D}_\Lambda$ in terms of a fixed DtN spectrum and Dirichlet spectral data, the decomposition of the union of all DtN spectra into real-analytic curves, a positive answer to an open problem of Bucur–Freitas–Kennedy on Robin eigenvalue branches, and several conjectures supported by numerical evidence.

Setting and basic spectral theory

For $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$, the operator $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$ maps a boundary datum $u$ to the normal derivative of its unique $\Lambda$-harmonic extension $U = \mathcal{E}_\Lambda u$. On bounded Lipschitz domains it is self-adjoint, semi-bounded below, and has compact resolvent; for smooth boundaries it is an elliptic pseudodifferential operator of order one behaving as $-\Delta U = \Lambda U$0 up to order zero terms. Its spectrum $-\Delta U = \Lambda U$1 coincides with that of the Steklov-type problem

$-\Delta U = \Lambda U$2

The paper emphasises the Robin–Dirichlet-to-Neumann duality: $-\Delta U = \Lambda U$3 if and only if $-\Delta U = \Lambda U$4, with matching multiplicities and eigenfunctions related by the $-\Delta U = \Lambda U$5-harmonic extension. This duality is the workhorse throughout: it transfers monotonicity, asymptotics, nodal count, and positivity results from the Robin Laplacian to the DtN map.

A subtle point is the variational characterisation. For $-\Delta U = \Lambda U$6, the minimax principle can be written over the full space $-\Delta U = \Lambda U$7; for $-\Delta U = \Lambda U$8 this fails — the authors reproduce Friedlander's example showing the Rayleigh quotient over $-\Delta U = \Lambda U$9 diverges to $\Omega \subset \mathbb{R}^d$0, so the minimisation must be restricted to the space $\Omega \subset \mathbb{R}^d$1 of $\Omega \subset \mathbb{R}^d$2-harmonic functions. This restriction has direct consequences: the simple test function $\Omega \subset \mathbb{R}^d$3 yields $\Omega \subset \mathbb{R}^d$4 for $\Omega \subset \mathbb{R}^d$5.

Monotonicity, counting functions, and analytic branches

Three structural facts organise the dependence on $\Omega \subset \mathbb{R}^d$6. First, each eigenvalue $\Omega \subset \mathbb{R}^d$7 is strictly decreasing between consecutive Dirichlet eigenvalues; at a Dirichlet eigenvalue of multiplicity $\Omega \subset \mathbb{R}^d$8, exactly $\Omega \subset \mathbb{R}^d$9 eigenvalues tend to $\Lambda=0$0 from the left. Second, zero belongs to $\Lambda=0$1 precisely when $\Lambda=0$2 is a Neumann eigenvalue, giving the counting identity $\Lambda=0$3 and hence Friedlander's inequality $\Lambda=0$4 for $\Lambda=0$5. Third, the principal eigenvalue is positive for $\Lambda=0$6 and strictly negative for $\Lambda=0$7 — the latter proved via the plane-wave test function $\Lambda=0$8 with $\Lambda=0$9.

The central new result is an explicit matrix identity. Given the eigenbasis $\mathcal{D}_\Lambda$0 of $\mathcal{D}_\Lambda$1 and the Dirichlet spectral data, the matrix of $\mathcal{D}_\Lambda$2 satisfies

$\mathcal{D}_\Lambda$3

where $\mathcal{D}_\Lambda$4 collects inner products of $\mathcal{D}_\Lambda$5 against normal derivatives of Dirichlet eigenfunctions and $\mathcal{D}_\Lambda$6 is diagonal with entries $\mathcal{D}_\Lambda$7. Since $\mathcal{D}_\Lambda$8 is thereby a Kato type (A) analytic family on every interval avoiding the Dirichlet spectrum, the union of spectra decomposes into real-analytic curves defined either on semi-infinite intervals $\mathcal{D}_\Lambda$9 or finite intervals between Dirichlet eigenvalues. As a corollary via the duality, every analytic branch of Robin eigenvalues tends either to $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$0 or to a Dirichlet eigenvalue as $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$1, resolving Open Problem 4.11 of Bucur–Freitas–Kennedy. The authors also derive first and second derivatives of simple eigenvalue branches; the second-derivative formula appears not to have been stated previously in this generality, and implies concavity of $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$2 on $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$3.

Domain monotonicity fails in general (a dumbbell counterexample even for $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$4), but a monotonicity under exclusions holds for $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$5: removing a compactly contained obstacle cannot increase any eigenvalue. For $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$6 this too fails, since $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$7 as $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$8 while the punctured domain retains a bounded principal eigenvalue.

Explicit examples and inequalities

Eigenvalues are computable by separation of variables for intervals, disks, balls, annuli, and cuboids. For the unit disk, $$\Lambda \notin \operatorname{Spec}(-\Delta_\Omega^{Dir})$$9 for $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$0 and $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$1 for $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$2; an appendix proves via order-derivatives of Bessel functions that these curves do not cross for $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$3, so the Steklov ordering persists. Notably, disk eigenfunctions are independent of $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$4 — an exceptional consequence of rotational symmetry; for cuboids, branches can intersect even at negative $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$5.

On the isoperimetric side, the survey recalls Weinstock's inequality $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$6 (sharp only among simply connected planar domains), the Hersch–Payne–Schiffer bound $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$7, and the sharp constant $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$8 for arbitrary planar domains. For $$\mathcal{D}_\Lambda : H^{1/2}(\partial\Omega) \to H^{-1/2}(\partial\Omega)$$9, available bounds are coarser: $u$0 with universal $u$1, and Menezes–Lima's two-eigenvalue inequality recovering Weinstock in the limit. Extensions to higher dimensions are explicitly noted as not yet worked out. A uniform estimate of Girouard–Karpukhin–Levitin–Polterovich gives $u$2 for $u$3 on smooth boundaries, and the elementary exponential test function yields $u$4 for all Lipschitz domains. Based on numerics, the authors conjecture the branch-wise refinement $u$5 along fixed analytic branches for $u$6; this is verified for balls but open in general.

Asymptotic regimes

Near $u$7, perturbation theory gives $u$8, with the second-order term expressed through Neumann eigenfunctions; e.g., $u$9.

As $\Lambda$0 from below, the first $\Lambda$1 eigenvalues blow up like $\Lambda$2, where $\Lambda$3 are eigenvalues of the Gram matrix of normal derivatives of the corresponding Dirichlet eigenfunctions.

As $\Lambda$4, one always has $\Lambda$5, and for $\Lambda$6 boundaries $\Lambda$7 for every fixed $\Lambda$8. For $\Lambda$9 boundaries, inversion of Pankrashkin–Popoff's Robin asymptotics refines this to $U = \mathcal{E}_\Lambda u$0, where $U = \mathcal{E}_\Lambda u$1 is the maximal mean curvature. For polygons the coefficient drops below one: for a curvilinear polygon with smallest angle $U = \mathcal{E}_\Lambda u$2, $U = \mathcal{E}_\Lambda u$3. The survey reformulates a conjecture of Chaigneau–Grebenkov identifying all coefficients $U = \mathcal{E}_\Lambda u$4 with the multiset built from $U = \mathcal{E}_\Lambda u$5 over corners; its validity reduces to whether the count of discrete Robin eigenvalues below the essential spectrum of a sector equals the known lower bound $U = \mathcal{E}_\Lambda u$6 — known to hold when all angles are at least $U = \mathcal{E}_\Lambda u$7. A caveat is recorded: Dietze and Pankrashkin exhibited Lipschitz domains where even $U = \mathcal{E}_\Lambda u$8 lacks a linear-in-$U = \mathcal{E}_\Lambda u$9 asymptotic, so regularity genuinely matters here.

The Weyl law $-\Delta U = \Lambda U$00 holds for smooth boundaries and extends to $-\Delta U = \Lambda U$01 ($-\Delta U = \Lambda U$02 planar) boundaries for $-\Delta U = \Lambda U$03; for merely Lipschitz boundaries it is established only for $-\Delta U = \Lambda U$04 (Karpukhin–Lagacé–Polterovich, Rozenblum), and extending it to $-\Delta U = \Lambda U$05 is posed as an open question. In dimension two, simply connected domains enjoy superpolynomial closeness of their eigenvalues to those of the equal-perimeter disk, extended by Lagacé–St-Amant to $-\Delta U = \Lambda U$06 with a complete expansion whose first correction is linear in $-\Delta U = \Lambda U$07.

Eigenfunctions: nodal structure, positivity, localisation

Bulk eigenfunctions $-\Delta U = \Lambda U$08 are real-analytic inside $-\Delta U = \Lambda U$09 and continuous up to a Lipschitz boundary ($-\Delta U = \Lambda U$10 up to a $-\Delta U = \Lambda U$11 boundary). Courant's theorem extends with a shift: $-\Delta U = \Lambda U$12 has at most $-\Delta U = \Lambda U$13 nodal domains, where $-\Delta U = \Lambda U$14 counts Dirichlet eigenvalues below $-\Delta U = \Lambda U$15 — obtained via the Robin duality. For boundary eigenfunctions $-\Delta U = \Lambda U$16 the situation is sharply different: since $-\Delta U = \Lambda U$17 is nonlocal, no Courant-type bound holds in general; Enciso–Pistoia–Provenzano constructed Riemannian manifolds of dimension $-\Delta U = \Lambda U$18 where finitely many low boundary eigenfunctions have arbitrarily many nodal domains. Whether an asymptotic $-\Delta U = \Lambda U$19 bound holds for $-\Delta U = \Lambda U$20 remains open. In two dimensions, a Hopf–Oleinik argument shows that for $-\Delta U = \Lambda U$21 boundaries the zeros of $-\Delta U = \Lambda U$22 are exactly endpoints of nodal lines of $-\Delta U = \Lambda U$23, which would transfer the Courant bound; the argument breaks down for merely Lipschitz boundaries, and the authors flag finding an alternative approach as an interesting question.

Strict positivity of the ground state $-\Delta U = \Lambda U$24 on $-\Delta U = \Lambda U$25 holds for $-\Delta U = \Lambda U$26, implying simplicity of $-\Delta U = \Lambda U$27; beyond $-\Delta U = \Lambda U$28, Daners' examples show neither semigroup positivity nor ground-state positivity admits a clean criterion. High eigenfunctions localise near the boundary: polynomial decay $-\Delta U = \Lambda U$29 away from $-\Delta U = \Lambda U$30 for smooth boundaries and $-\Delta U = \Lambda U$31, upgraded to exponential decay $-\Delta U = \Lambda U$32 for real-analytic boundaries. Helffer and Kachmar asked whether $-\Delta U = \Lambda U$33 suffices for the exponential rate; this remains unresolved.

Representations, extensions, and applications

Layer-potential theory yields $-\Delta U = \Lambda U$34, independent of the choice of fundamental solution — important because for $-\Delta U = \Lambda U$35 the fundamental solution is complex-valued (Hankel function satisfying the Sommerfeld condition) yet the map is real. Green's function representations express Robin Green's functions through bulk eigenfunctions with explicit $-\Delta U = \Lambda U$36-dependence, $-\Delta U = \Lambda U$37, and identify the boundary trace of the Neumann Green's function with the kernel of $-\Delta U = \Lambda U$38.

Extensions covered include mixed Steklov–Dirichlet–Neumann problems (with the sloshing problem as the canonical example, and domain monotonicity restored in mixed settings à la Bañuelos–Kulczycki–Polterovich–Siudeja), exterior problems (well-posed for $-\Delta U = \Lambda U$39 with exponentially decaying solutions; delicate at $-\Delta U = \Lambda U$40 where several constructions coincide up to rank-one perturbations, with the Dirichlet truncation giving the canonical operator; non-self-adjoint for $-\Delta U = \Lambda U$41), and complex $-\Delta U = \Lambda U$42, where $-\Delta U = \Lambda U$43 is $-\Delta U = \Lambda U$44-sectorial with meromorphic dependence on $-\Delta U = \Lambda U$45 whose poles are exactly the Dirichlet eigenvalues, and the Robin duality survives verbatim.

Numerically, the survey reviews FEM, boundary element methods based on the boundary integral equation, and the method of fundamental solutions, noting a common difficulty: accuracy deteriorates and negative eigenvalues become hard to resolve near Dirichlet eigenvalues. Two applications receive detailed treatment: domain decomposition, where partial DtN maps on artificial interfaces reduce waveguide problems to bounded subdomains, and diffusion-controlled reactions, where the encounter-based framework uses the heat kernel $-\Delta U = \Lambda U$46 of the DtN semigroup on the boundary local time; averaging over an exponentially distributed threshold recovers the classical Robin flux density, while other threshold distributions model encounter-dependent surface reactivity beyond Robin boundary conditions.

Limitations and open questions

Several limitations are conceded explicitly. The variational principle over $-\Delta U = \Lambda U$47 is unavailable for $-\Delta U = \Lambda U$48, restricting many techniques to the subcritical regime. Isoperimetric inequalities for $-\Delta U = \Lambda U$49 are crude, dimensionally restricted to planar domains, and their higher-dimensional analogues are undeveloped. The conjectures on branch-wise $-\Delta U = \Lambda U$50 bounds and polygonal coefficients $-\Delta U = \Lambda U$51 rest on numerics and on the unproven equality case for discrete eigenvalues of Robin sectors. Weyl asymptotics with remainder control, pointwise Weyl laws, and $-\Delta U = \Lambda U$52 eigenfunction bounds are unverified for non-smooth boundaries when $-\Delta U = \Lambda U$53. No Courant-type nodal bound exists for boundary eigenfunctions in dimensions $-\Delta U = \Lambda U$54, and the Euclidean analogue of the Enciso–Pistoia–Provenzano construction is unknown. Exponential boundary localisation is proven only for real-analytic boundaries. Finally, almost nothing is known about DtN-isospectrality: no pair of non-isometric Euclidean domains is known to be $-\Delta U = \Lambda U$55-isospectral for nonzero $-\Delta U = \Lambda U$56, nor Robin-isospectral for nonzero Robin parameter.

Conclusion

The survey consolidates the spectral theory of the Helmholtz Dirichlet-to-Neumann map around three pillars: the Robin–DtN duality, the analytic-curve decomposition of the parameter-dependent spectrum, and the explicit matrix formula relating $-\Delta U = \Lambda U$57 to $-\Delta U = \Lambda U$58 plus Dirichlet data. These tools yield quantitative asymptotics in the regimes $-\Delta U = \Lambda U$59, $-\Delta U = \Lambda U$60, and $-\Delta U = \Lambda U$61, extend Courant and positivity theory to the Helmholtz setting, and connect the subject to sloshing, scattering, and diffusion-mediated surface chemistry. The residual gaps — rough-boundary asymptotics, sharp isoperimetric inequalities for $-\Delta U = \Lambda U$62, nodal counts for boundary eigenfunctions, and the sector eigenvalue count underlying the polygonal conjecture — delineate the precise agenda the paper leaves to subsequent work.