How Well Can You Hear a Shape? Fisher Forecasts for Inverse Spectral Problems

Kac's old question, "can one hear the shape of a drum?", asks whether the list of vibration frequencies of a space determines its geometry. Its modern descendants are all yes/no answers. The discrete Laplacian fixes a discrete surface metric. Scattering data fix conical singularities on hyperbolic surfaces. In practice, though, nobody ever has the whole spectrum without error. A physicist has a handful of Kaluza-Klein masses: particle masses that equal the vibration frequencies of the hidden, curled-up extra dimensions. A shape analyst has the first fifty eigenvalues of a graph Laplacian built from a noisy point cloud. The question that matters for them is quantitative. Given $K$ eigenvalues measured with known noise, how precisely do they pin down the shape parameters? Which eigenvalues carry the information? Where does recovery become impossible?

We propose to answer this with the standard tool for that kind of question in observational cosmology: Fisher matrix forecasting. This method predicts, before any data are collected, the smallest achievable error bars on parameters, known as the Cramér-Rao bound, together with the parameter combinations that are degenerate. Applying it to spectra is not a plug-in exercise, for two reasons. First, eigenvalues cross and become degenerate as the geometry deforms. Second, eigenvalues estimated from point clouds are biased by the choice of kernel bandwidth. The contribution here is a spectral Fisher forecast that stays well defined through crossings and accounts for bandwidth bias. We validate it against brute-force Monte Carlo inversion on two testbeds: deformed spheres sampled as point clouds, and the one-parameter family of quintic Calabi-Yau manifolds.

Core Claim

Let $t = (t^1,\dots,t^p)$ parameterize a family of geometries. These could be the axis lengths of an ellipsoid or the complex-structure modulus of the quintic. Let $s(t) \in \mathbb{R}^M$ be a vector of spectral summaries, and let $C$ be their noise covariance. The forecast is

$$F_{ab} = \sum_{k,l} \frac{\partial s_k}{\partial t^a}\,(C^{-1})_{kl}\,\frac{\partial s_l}{\partial t^b}, \qquad \delta t^a = \sum_b (F^{-1})^{ab} \sum_{k,l} \frac{\partial s_k}{\partial t^b} (C^{-1})_{kl}\, \beta_l .$$

Here $F$ is the Fisher matrix. Its inverse lower-bounds the covariance of any unbiased estimate of $t$. The vector $\beta$ is the systematic bias in the measured summaries, for example the oversmoothing bias of a graph Laplacian at bandwidth $h$. So $\delta t$ is the forecast shift in the recovered parameters caused by that bias. This second formula is the standard cosmology device for propagating systematics.

The derivatives come from first-order perturbation theory, also called the Hellmann-Feynman rule. For an isolated eigenpair $(\lambda_k, \psi_k)$, $\partial \lambda_k / \partial t^a = \langle \psi_k, (\partial_a \Delta)\, \psi_k \rangle$, so no finite differences across the family are needed.

The claim has three parts. First, the naive recipe breaks near level crossings and degeneracies. That recipe treats sorted eigenvalues as independent Gaussian measurements, which is the obvious transplant of Measuring Cosmological Parameters with Galaxy Surveys, where noise power spectra are projected into a Fisher matrix to forecast survey errors and degeneracies. Near a crossing, the sorted eigenvalues are not differentiable and their estimation noise is strongly correlated.

Second, the fix is to replace each near-degenerate cluster of eigenvalues by its symmetric functions, such as cluster trace and cluster spread. These are smooth in $t$. Their derivatives come from degenerate perturbation theory: the eigenvalues of the small matrix $\langle \psi_i, \partial_a \Delta\, \psi_j\rangle$ restricted to the cluster. The covariance $C$ should be the full cross-mode covariance, estimated by resampling.

Third, a forecast built this way should match the actual spread of a maximum-likelihood inversion. It should also predict in advance where identifiability is lost, meaning where $F$ becomes nearly singular.

Why This Gap Is Real

Several papers motivate the question by showing spectra determine geometry in principle, and they constrain it by showing how hard spectra are to compute. Discrete Laplace-Beltrami Operator Determines Discrete Riemannian Metric proves that the cotangent Laplacian uniquely encodes a mesh metric. Conic singularities, generalized scattering matrix, and inverse scattering on asymptotically hyperbolic surfaces proves that scattering data uniquely determine cone points and cusps. Both give uniqueness with full, exact data. Neither says how many eigenvalues suffice or how noise propagates, and that is exactly what our forecast supplies.

On the physics side, Moduli-dependent KK towers and the swampland distance conjecture on the quintic Calabi-Yau manifold numerically computes scalar Laplacian eigenvalues along a one-parameter slice of moduli. It fits exponential mass decay, reporting $\alpha \approx 0.45$ overall and $m \approx 56.4\,e^{-(0.906 \pm 0.034)\, d}$ for the second-lightest mode. It also observes level crossings for which it has no predictive theory. We take its numerical pipeline as the method that generates spectra, and its crossings as the hardest test of our forecast. We do not attempt an analytic theory of which crossings occur. Instead we ask whether spectral information about the modulus survives them.

On the data side, Data driven estimation of Laplace-Beltrami operator shows that point-cloud Laplacians face a bandwidth bias-variance trade-off. It gives Lepski-adaptive pointwise rates of $n^{-1/(d+4)}$ on $S^2$ and leaves spectral convergence under adaptive bandwidth open. This constrains our design. Because no theorem gives the eigenvalue noise covariance, we estimate $C$ and $\beta$ empirically by resampling, and we treat bias explicitly rather than assuming unbiased measurements.

The closest Fisher-geometry work is Riemannian Laplace Approximation with the Fisher Metric. It uses the Fisher matrix as a metric for warping posterior samples along geodesics. Its "Laplace" is the integral approximation, not the operator, and it forecasts no measurement. Similarly, Personalized Federated Learning with Residual Fisher Information for Medical Image Segmentation combines planar Fourier amplitude swaps with diagonal Fisher masks. It does not touch Laplace-Beltrami spectra or Cramér-Rao bounds.

Research Program

We build a forecasting library with three layers.

The spectral layer computes eigenpairs and the operator derivatives $\partial_a \Delta$. It uses analytic metrics for ellipsoids, cotangent meshes for high-resolution "truth," and the Donaldson-algorithm Ricci-flat metrics plus Galerkin discretization from Eigenvalues and eigenfunctions of the scalar Laplace operator on Calabi-Yau manifolds for the quintic. Where large meshes need many eigenpairs, a multigrid solver in the spirit of Bootstrap Multigrid for the Laplace-Beltrami Eigenvalue Problem supplies them.

The summary layer groups eigenvalues into clusters wherever neighboring gaps fall below a noise-scaled threshold. Within each cluster it forms smooth symmetric functions, and it computes their derivatives by degenerate perturbation theory.

The noise layer estimates the full covariance $C(n,h)$ and bias $\beta(n,h)$ of the summaries. For point clouds it does this by bootstrap resampling and by comparison against mesh truth. For the Calabi-Yau setting it uses an explicit measurement model: fractional mass uncertainties applied to the first $K$ KK masses.

The library then answers design questions. How does the bound on each parameter shrink with $K$, with sample size $n$, and with bandwidth $h$? Which modes dominate each row of $F$? Along which directions in $t$ does $F$ approach singularity? The sphere is the ideal stress test. Its eigenvalues $l(l+1)$ have multiplicity $2l+1$, and deforming it into an ellipsoid splits every degeneracy. The naive sorted-eigenvalue forecast is therefore at its worst exactly at the most symmetric point.

Baselines

The do-nothing reference is the volume-and-Weyl forecast. By Weyl's law, the growth of eigenvalues reveals only total volume to leading order. So a forecast using only the counting function treats every volume-preserving deformation as unrecoverable. Any useful spectral forecast must beat this by bounding volume-preserving parameters, such as ellipsoid aspect ratios at fixed volume, which this reference cannot constrain at all.

The strongest existing forecasting method is the textbook Fisher recipe of Measuring Cosmological Parameters with Galaxy Surveys, transplanted to sorted eigenvalues with diagonal Gaussian noise and no bias term. It is the strongest competitor because it is what anyone would do first, and away from crossings it should be nearly correct. We run it on exactly the same eigenpairs, the same $K$, and the same noise draws as our forecast, with its diagonal variances taken from the same bootstrap. The only differences are clustering, off-diagonal covariance, and bias propagation.

The ground-truth comparator is Monte Carlo inversion. We draw many noisy spectra or point clouds, fit $t$ by maximum likelihood on the same summaries, and record the empirical covariance and mean shift. No reported result found on Lacuna for Fisher or Cramér-Rao forecasts of inverse spectral problems. The only quantitative anchors in the file are the pointwise rate $n^{-1/(d+4)}$, for which the file gives no $S^2$ constants, and the quintic mass fits quoted above, which our reproduction must recover before any forecast is trusted.

Beating the textbook recipe means two things. First, our forecast's predicted standard deviations stay within a fixed tolerance of the Monte Carlo spread, which we set at 20%, across the whole family including crossings, while the textbook forecast leaves that band near crossings. Second, our predicted bias shift $\delta t$ matches the Monte Carlo mean shift as bandwidth varies.

Three ablations isolate the key ingredients. We remove clustering while keeping full covariance and bias. We remove off-diagonal covariance while keeping clustering. We remove the bias term. Each ablation shows which ingredient carries the improvement and where in the family it matters.

Why not just…?

Why not just run the inversion and measure its error by simulation? For a one- or two-parameter family this is a real threat. Monte Carlo inversion is the gold standard, and we use it as ground truth. The forecast earns its place in two ways. Its cost is one eigen-solve and one bootstrap per parameter point, rather than thousands of inversions. More importantly, it decomposes information by mode and exposes near-singular directions, which simulation only reveals after the fact. We will report wall-clock cost for both methods. If the forecast is not substantially cheaper at matched accuracy for $p \ge 3$, its value reduces to interpretability, and we will say so.

Why not just use heat-trace invariants (volume, total curvature, boundary length) instead of individual eigenvalues? These are smooth and crossing-free, which makes them attractive. They are also the natural summaries for the integrated invariants that small-time heat expansions expose. But they are global integrals: they cannot separate deformations that preserve volume and total curvature. The experiment that settles this computes the forecast using only heat-trace coefficients and compares it with the forecast using clustered low eigenvalues on the ellipsoid family. If the heat-trace forecast is equally tight, the eigenvalue machinery is unnecessary for smooth deformations.

Why not just smooth through crossings by fitting curves to eigenvalue branches, as the quintic study fits exponentials? Curve fits along the family give derivatives. However, they assign the derivative to a sorted branch, and that branch changes identity at each crossing. Fits also say nothing about correlated estimation noise between the near-degenerate modes. Our ablation without clustering is exactly this approach, so its failure or success near the quintic crossings answers the question directly.

Proposed Experiments

The first experiment uses point clouds on deformed spheres. We sample $n$ points from ellipsoids with axes $(1, 1+t^1, 1+t^2)$, with $t$ ranging from $0$ to $0.3$. We build unnormalized graph Laplacians at fixed bandwidths and at Lepski-selected bandwidths following Data driven estimation of Laplace-Beltrami operator. We then compare the forecast and Monte Carlo covariances of $t$ for $K \in \{10, 30, 100\}$ and $n \in \{10^3, 10^4, 10^5\}$.

The decisive checks come in three steps. The first concerns calibration at the sphere: does the naive forecast blow up or collapse at $t = 0$, where degeneracy is maximal, while ours stays calibrated? The second concerns bias: does the forecast bias shift $\delta t$ track the empirical drift as $h$ grows? The third concerns sample size: does forecast precision scale with $n$ at the empirically observed rate? This last check gives indirect evidence on the open spectral-convergence question.

The second experiment uses the quintic Calabi-Yau family. We reproduce the scalar spectra along the one-parameter slice and require agreement with the reported $\alpha \approx 0.45$ and the second-mode fit before proceeding. We then forecast the bound on the modulus from the first $K$ masses at fixed fractional mass error, and locate parameter values where $F$ nearly vanishes. The decisive check is whether those predicted information holes coincide with the observed level crossings and symmetry-enhanced points. A further check is whether grouping modes by symmetry representation, as the clustering implies, restores information there. If information holes appear only in the naive forecast, crossings are a bookkeeping artifact. If they persist in ours, they are genuine limits on recovering geometry from KK masses, which is a physics statement the existing numerics cannot make.