An inverse spectral problem for the wave equation with a varying Robin parameter at the boundary

1.1. Problem setting and main results.

• The eigenvalues at zero Robin parameter and their derivatives uniquely determine the wave speed, with the derivatives encoding boundary behavior of eigenfunctions.The numerical problem approximates these derivatives from two finite spectra.Source: §1.1 · p. 2
• The forward spectral map is Lipschitz continuous, while the inverse map is Hölder continuous on a smooth subclass with a buffer near the endpoint.The inverse result holds for exponents below κ/2; the subclass requires c to equal 1 outside the specified interval.Source: §1.1 · p. 3
• The numerical inverse problem uses two finite spectra, approximating eigenvalue derivatives by finite differences to construct a stable, computationally feasible inverse.The derivative approximation uses a sufficiently small fixed Robin parameter.Source: §1.1 · pp. 3–4

1.2. Outline of the paper and further results.

• The reconstruction identifies the Neumann-to-Dirichlet map from spectral data, then uses that map in boundary control to recover domains of influence and the wave speed.The numerical implementation uses truncated spectra and finite-difference estimates of eigenvalue derivatives.Source: §1.1 · p. 4, §1.2 · pp. 5–6
• Combining L2 and H1 penalties yields higher-regularity controls whose convergence to domain-of-influence indicators is uniform over the allowed radii.One pair of regularization parameters can therefore be fixed across the grid of radii used to compute travel-time volumes.Source: §1.2 · p. 6
• The experiments test smooth and piecewise constant profiles, verify that computed sources approximate domain-of-influence indicators, and examine sensitivity to spectral noise.The outline reports accurate recovery and errors concentrated near endpoints for smooth profiles and at jumps for piecewise constant profiles.Source: §1.2 · p. 6

2.1. Solutions to the impedance spectral problem.

• The point spectrum of P consists of positive reciprocals of A’s eigenvalues, and each weak solution space equals the corresponding eigenspace of A.For λ in σp(P), the weak solutions are exactly Ker(A − λ^-1 I).Source: §2.1 · p. 8

2.2. Analyticity of perturbations.

• The resolvent and each eigenvalue’s spectral projector depend analytically on the Robin parameter ω throughout [0, 1).The perturbation is represented by Pω = P − ωγ, enabling analytic dependence of the resolvent and spectral projectors.Source: §2.2 · pp. 11–12
• Every eigenvalue is simple, and each eigenspace has a unique unit-norm weak solution selected by requiring its value at 0 to be positive.The eigenfunction is normalized in L2_c; the sign convention v(0) > 0 selects one solution.Source: §2.2 · p. 12
• For each n, the eigenvalue λn(ω) and normalized eigenfunction ϕn,ω are analytic for 0 ≤ ω < 1.Source: §2.2 · p. 13

2.3. Generic behaviour of eigenvalues.

• At ω = 0, the derivative of each impedance eigenvalue curve recovers the squared boundary value of the corresponding unperturbed eigenfunction.The unperturbed problem is the Neumann-boundary spectral problem.Source: §2.3 · p. 14

3.1. Energy estimates and finite speed of propagation.

• For L2 Neumann data, the weak solution is unique and depends continuously on the input.Source: §3.1 · pp. 15–16
• For Neumann data in the compatible H1 subspace, the wave equation has a unique strong solution, and its solution operator is bounded into H2.The compatibility condition is imposed at t=0.Source: §3.1 · p. 16
• At time T, a Neumann input supported in [T−r,T] produces a wave whose support lies within the domain of influence Ω(r).Source: §3.1 · p. 16

3.2. Representation of the Neumann-to-Dirichlet map.

• The Neumann-to-Dirichlet map is bounded on L2(0,T) and admits a representation derived from spectral data.The representation is established through the eigenvalue-based formula.Source: §3.2 · pp. 17–18
• Energy estimates bound the Galerkin approximations in H1, supporting passage to the weak solution used in the map representation.Source: §3.2 · pp. 17–18

4.1. Preliminaries.

• Blagovestchenskii identities compute inner products of interior wave fields using boundary-time operators acting on prescribed boundary inputs.Source: §4.1 · pp. 18–19
• The interval reconstruction requires r ≤ τ(1), the exact condition ensuring the endpoint has not influenced the solution by observation time T.Source: §4.1 · p. 19

4.2. Reconstruction of wave speed function.

• The regularized controls depend continuously on r and converge uniformly as α and β tend to zero.Source: §4.2 · p. 20
• The reconstruction builds the Neumann-to-Dirichlet map from spectral data, forms the Blagovestchenskii operator, and uses it to determine the travel-time volume.Source: §4.2 · pp. 22–23
• The travel-time volume determines the wave speed through c(x) = 1/V′(τ(x)).Source: §4.2 · p. 23
• Equal spectral data imply equal wave speeds on the full interval, so the forward spectral map is injective.Source: §4.2 · p. 23

5.1. Forward stability.

• Eigenvalue differences satisfy |λc1_n − λc2_n| ≲ n^2∥c1 − c2∥C2(Ω) for every n.Source: §5.1 · p. 24
• Robin-parameter eigenvalue-derivative differences satisfy |(λc1_n)′(0) − (λc2_n)′(0)| ≲ ∥c1 − c2∥C2(Ω), uniformly in n.Source: §5.1 · p. 26
• The forward spectral map is Lipschitz when eigenvalue differences are weighted by n^-2 and Robin-parameter derivative differences are unweighted; weight s=2 is minimal.Source: §5.1 · p. 29

5.2. Inverse stability.

• The Neumann-to-Dirichlet map is uniformly bounded from L2(0, 2T) to H1(0, 2T), with a bound depending only on Cmin, Cmax, and T.The proof extends the time interval to (0, 4T) and selects an energy-estimate time in [2T, 4T].Source: §5.2 · pp. 30–31
• The cited estimate is stated for any θ ∈(0, 1/2) and c1, c2 ∈C, with a constant depending on θ that is not bounded as θ →1.Source: §5.2 · p. 32
• The reconstruction strategy of requires a fixed control range below the unknown travel time that also covers the support of c−1, which is not possible for all of C.The subclass C′ ensures a common admissible range and that reconstruction over this range covers the wave speed throughout Ω.Source: §5.2 · pp. 34–35
• The inverse spectral map Φ−1 : D′ → C′ is Hölder continuous from ℓ∞^2 × ℓ∞^0 to C^2(Ω) for every exponent θ < κ/2.Source: §5.2 · p. 35
• The inverse estimate also yields C^{2,η}(Ω) stability with exponent θ(1−η) for η<1, but it does not imply stability in C^{2,1}(Ω) or C^3(Ω).The uniform C^3 bound does not ensure that third derivatives of uniformly close wave speeds are close.Source: §5.2 · pp. 35–36
• The inverse-map exponent κ=1/54 is inherited from the regularization strategy of; lower bounds and whether a strategy-independent Hölder exponent or subclass Lipschitz continuity holds remain open.Source: §5.2 · p. 36

6.1. Forward problem: finite element method.

• The forward problem uses a Ritz–Galerkin finite element approximation on a uniform spatial mesh with piecewise linear basis functions.The Dirichlet condition at x = 1 removes the final nodal basis function from the approximation space.Source: §6.1 · pp. 36–37
• The Robin impedance changes only the first diagonal entry of the stiffness matrix, while the wave speed enters through the weighted mass matrix.The local mass contributions are approximated using two-point Gauss–Legendre quadrature.Source: §6.1 · p. 38
• Finite element matrices yield generalized eigenvalue approximations for the unperturbed and perturbed problems, which provide the simulated spectral data.The mass matrix is symmetric positive-definite, and the stiffness matrix is symmetric, giving real discrete eigenvalues.Source: §6.1 · p. 39

6.2. Inverse problem: reconstruction of the wave speed via the boundary control method.

• The inverse procedure builds a Neumann-to-Dirichlet map from spectral data, forms a discrete Blagoveščenskiĭ operator, and uses regularized controls to approximate indicator functions.The computed control yields a volume profile that is differentiated and transformed to reconstruct the wave speed.Source: §6.2 · pp. 39–40
• The volume profile is differentiated through a regularized inversion of the cumulative integration operator, with total variation used for piecewise constant profiles.Spline or H1-Tikhonov differentiation is used for smooth profiles; total variation is reserved for profiles outside the admissible class because their derivatives are discontinuous.Source: §6.2 · p. 41
• The fully discrete scheme has no proved convergence, and its truncation, mesh, time-step, observation-margin, and control-regularization parameters are chosen empirically.The text notes that a convergence analysis would require coupling these parameters.Source: §6.2 · p. 41

7.1. Reconstruction of Neumann boundary sources.

• With α = 3e-5 and β = 3e-10, joint L2–H1 regularization gives an oscillation-free, sharper source reconstruction than the L2-only optimum.The recovered indicator has a steeper wavefront transition and closer agreement with the true characteristic function.Source: §7.1 · p. 45, Figure 7 · p. 45

7.2. Wave speed profiles.

• Below 0.004%, the relative L2 and L1 errors for c(x) = 1; for smooth profiles (b) and (c), both errors remain below 0.33%.For the smooth profiles, errors concentrate near the domain endpoints, where regularized differentiation of V has end effects.Source: §7.2 · p. 46
• For profile (d), linear extrapolation corrects right-end effects; piecewise-constant levels are accurate, with errors concentrated at jumps.In the multiple-jump case, TV and L∞ accuracy deteriorates; the authors leave a fuller explanation and comparison with TV-based methods for future work.Source: §7.2 · pp. 46–47

7.3. Sensitivity to noise in the spectral data.

• The noise threshold grows linearly with ω and is expected to fall like 1/L^2; below it, reconstruction matches noise-free accuracy, while above it results deteriorate or break down.The experiments report noise-free-level accuracy around ρ ≈ 0.5 and rapid deterioration around ρ ≈ 5.Source: §7.3 · p. 48

7.4. Conclusions.

• The numerical experiments reconstruct smooth and piecewise-constant wave speeds accurately, with stability under spectral noise when the noise is small relative to the Robin parameter.Errors concentrate near endpoints for smooth profiles and at jumps for piecewise-constant profiles.Source: §7.4 · pp. 48–49
• The reported errors reflect grid-search tuning against the exact wave speed, which is unavailable in practice, and interacting parameters cannot be tuned independently.The authors identify data-driven parameter choice and coupling parameters to discretization and noise as future work.Source: §7.4 · p. 49

• The reconstruction code is openly available at GitHub and archived at Zenodo.The reported results used version 1.0.Source: Data availability · p. 49
• All spectral data used in the reported results were simulated by the code; no experimental data were used.Source: Data availability · p. 49