What the Spectrum Can See of a Drift: WKB Quantization for Direction-Dependent Wave Operators

Many physical media carry waves faster in one direction than the opposite one. Sound travels faster downwind, a fire front spreads faster with the wind, and a robot climbs a slope at a different cost than it descends. Geometers model this with a Finsler metric, a notion of length in which the cost of a small step depends on both position and direction. The simplest non-reversible case is a Randers metric, $F(x,y)=\sqrt{\alpha_x(y,y)}+\beta_x(y)$. Here $\alpha$ is an ordinary Riemannian metric and $\beta$ is a one-form, a "drift" that adds cost for motion against it and subtracts cost for motion along it.

Existing work on such media is either kinematic or heuristic. It recovers travel times and metrics from geodesic lengths, or it builds operators on meshes without analyzing their spectra. Nobody has asked the basic semiclassical question: when you look at the resonant frequencies of a cavity filled with a directional medium, which parts of the drift do those frequencies actually encode? We propose to answer it with WKB analysis. WKB is the classical high-frequency approximation that writes a wave as a slowly varying amplitude times a rapidly oscillating phase. We then use the answer to build an identifiability-aware method for recovering drifts from spectra. The method covers exactly the non-reversible regime where the known Finsler inversion theorems do not apply.

Core Claim

Take a cavity with a slow background flow, which is the "acoustic waves in a moving fluid" setting that the broken-scattering paper of de Hoop et al. explicitly excludes because its metric is non-reversible. WKB quantization (Bohr–Sommerfeld/EBK) of the convected wave operator says that a closed ray $\gamma$ contributes resonances at actions $\oint_\gamma F$. The drift enters these actions only through its circulation $\oint_\gamma \beta$, with opposite sign for the two traversal directions.

This leads to three falsifiable predictions at leading order in the small parameter $h = 1/\omega$.

First, the exact part of the drift is spectrally invisible at leading order. Exact means $\beta = df$, as in a potential flow. It shifts every travel time by a boundary-telescoping term and leaves all closed-orbit actions unchanged.

Second, degenerate pairs of modes rotating in opposite directions split by an amount set by the circulation along the corresponding periodic orbits. Writing $T_\gamma$ for the orbit's travel time, $$ \delta\omega_\gamma \approx \frac{2\pi}{T_\gamma}\cdot\frac{2\,\oint_\gamma \beta}{\oint_\gamma \sqrt{\alpha(\dot\gamma,\dot\gamma)}}\,\omega . $$ This form comes from differencing the Bohr–Sommerfeld conditions for the two traversal directions.

Third, a curl-free but topologically nontrivial drift still produces splitting. An example is a potential vortex in an annulus, which has $d\beta = 0$ but nonzero circulation. This is an acoustic analogue of the Aharonov–Bohm effect.

Direction-dependent curvature enters at the next order. In Finsler geometry the volume form is not determined by the length function alone. Its rate of distortion along geodesics is the S-curvature, and the mean Berwald curvature studied in Li's Schur-type lemma is its fiberwise Hessian. Our derivation target is that the $O(h)$ correction to the quantization condition is the orbit integral of this distortion term, together with the Maslov index. This would give the first point at which a directional curvature, rather than the metric itself, becomes spectrally measurable.

The resulting inverse statement is the main product. Splittings determine the ray transform of the non-exact part of $\beta$ over periodic orbits. The exact part must come from boundary travel times. The practical method follows directly: parametrize the drift by a Hodge-type split into an exact part and a co-exact-plus-harmonic part, fit the second from splittings, and fit the first from first-arrival times.

Research Program

Analysis. We derive the WKB hierarchy for the time-harmonic convected wave equation in two-dimensional cavities with low-Mach flow. The leading-order phase equation is the Randers/Zermelo eikonal. The transport equation for the amplitude is written along Randers geodesics, which are reparametrized magnetic geodesics of $\alpha$ with field $d\beta$.

Caustics are the main technical obstacle, because standard WKB amplitudes blow up where rays focus. We will not use standard shape operators there. Instead we transport amplitudes using the closed Jacobi-field ODE system of de Hoop, Ilmavirta and Lassas. That paper reconstructs the fundamental tensor along a single geodesic and stays stable through conjugate points; we reuse it as a numerical device for computing Maslov indices and monodromy. They work with ray data along isolated geodesics, whereas we need standing-wave quantization over families of closed orbits. The paper therefore supplies a method for us rather than competing with us.

For integrable cavities (disk and annulus with rotationally symmetric flows), EBK quantization is explicit. For a chaotic stadium with flow, we test the trace-formula version: the splitting of the length spectrum by $\pm\oint\beta$.

Numerics. We compute ground-truth spectra with a high-order finite-element discretization of the quadratic eigenvalue problem in $\omega$, up to several hundred modes. We use three flow classes chosen to separate the predictions. Potential flow ($\beta$ exact) should give no leading-order splitting. Solid-body rotation ($d\beta\neq 0$) should give splitting that tracks circulation. A potential vortex in an annulus ($d\beta=0$, nonzero circulation) should give purely topological splitting.

Inversion. We recover drifts from (i) a measured list of split frequencies and (ii) optionally a handful of boundary first-arrival times. The Hodge-split parametrization is fit by gradient descent through a differentiable WKB forward model. The exact component is fit only through travel times, using the implicit-differentiation eikonal machinery of the differentiable Randers–Finsler wildfire solver.

Application check. The Finsler–Laplace–Beltrami operator of Weber et al. builds a Randers-based operator on meshes via a small-asymmetry "distance trick" and uses its eigenfunctions for shape matching, with no spectral analysis. If our gauge-invariance prediction transfers to their operator, part of the drift field it learns or chooses is spectrally inert. Removing that part should then leave its correspondence accuracy unchanged.

Relation to Prior Work

The kinematic inversion results motivate the gap, since they explicitly exclude flows. Boundary-distance determination of Finsler manifolds requires fiberwise analyticity. Broken-scattering determination requires reversibility, convex foliations and dimension at least three. Our setting is two-dimensional, non-reversible and spectral, so it falls outside both.

Lassas, Uhlmann and Wang supply the methodological precedent: high-frequency expansions combined with microlocal analysis to recover geometry. They work with Lorentzian (quadratic, symmetric) light cones and time-domain sources. We work with non-quadratic, asymmetric cones and stationary spectra.

On the WKB side, Tomasini, Sclocchi and Wyart show both the power of WKB for predicting spectra (one-dimensional, isotropic) and its failure regime: below a critical ridge, sensitivity to boundary data breaks the prediction. That result constrains our design. We therefore measure where our prediction fails as a function of mode index and Mach number rather than assuming uniform validity.

The scarring construction on hyperbolic manifolds is confined to constant negative curvature and exact normal forms. We do not attempt quasimode scarring. Our claim is the narrower and testable one about eigenvalue splitting.

Andreux et al.'s anisotropic Laplace–Beltrami operator is symmetric by construction. It therefore cannot produce the directional splitting that is our signal, and it serves as a null model.

Baselines

The naive reference is the drift-free Riemannian prediction: Weyl/EBK eigenvalues of $\alpha$ alone. It predicts zero splitting and serves as the floor for the forward-model experiments. Any method must explain the splittings this reference misses.

The strongest forward competitor is first-order Rayleigh–Schrödinger perturbation theory in $\beta$ around the drift-free modes. It is exact to first order in Mach number and needs no ray geometry. We run it on the same cavities and mode ranges, using the same finite-element unperturbed modes. We count WKB as beating it if WKB's splitting error is lower at Mach numbers above roughly 0.1 or at mode indices above roughly 100. In those regimes, near-degenerate mixing should break first-order perturbation theory. If it does not, WKB's advantage reduces to interpretability.

The strongest inverse competitor is black-box differentiable spectral inversion. Here we backpropagate through the finite-element eigensolver to fit a fully parametrized $\beta$ to the same splitting data, with the same optimizer, iteration budget and regularization sweep. For a travel-time reference we run eikonal-only inversion with the solver of the wildfire paper. That paper reports drift-field error below 3% and metric error around 6% on its synthetic inverse problem. Those figures come from dense fire-perimeter data on flat grids, so they calibrate the eikonal approach rather than providing a like-for-like number. No reported result was found on Lacuna for drift recovery from spectral data. We will report all inversions as relative $L^2$ error separately on the exact and non-exact components of $\beta$. Beating the black-box inversion means lower non-exact-component error at equal compute, or equal error with an order of magnitude fewer forward solves.

The key ablation removes the Hodge split, so that the WKB forward model fits a single unstructured $\beta$ to the spectra alone. Our prediction is that this version shows a large, unresolvable error confined to the exact component. Seeing that error would confirm that the gauge null space is real, not a modeling artifact.

For the shape benchmark, the baseline is FLBO's reported accuracy at $r=0$ on FAUST Remeshed: 62.67%, against 61.95% for ACSCNN (Weber et al.). On SCAPE Remeshed the corresponding figures are 46.94% against 46.13%.

Why not just…?

Why not just solve the eigenproblem numerically and fit the drift by brute force? This is the real threat for pure prediction accuracy, and we run it as the main inverse baseline. Its weakness is identifiability. The black-box fit will converge to some $\beta$, but nothing in it reveals that the exact part is unconstrained, so it will silently report confident errors there. The ablation and the potential-flow cavity test exactly this. If the black-box inverter, given the same spectra, recovers the exact component accurately, our central gauge-invariance claim is wrong and the proposal fails.

Why not just use perturbation theory in the drift? For weak flow and low modes, perturbation theory should match WKB. We expect it to win on accuracy there and concede that regime. Perturbation theory also hides the structure we care about: it gives splittings as mode-coupling integrals, not as orbit circulations. It also needs the unperturbed eigenfunctions, which become expensive and nearly degenerate at high frequency. Sweeping Mach number and mode index settles where each method holds.

Why not just invert travel times with an eikonal solver? First arrivals do see the exact part of the drift, which spectra cannot. This is why our method uses them for that part. Travel times require timed sources and receivers on the boundary, though, and in a region with trapped or closed rays they miss the circulation on orbits that never reach the boundary. Splittings measure that circulation directly. In the annulus potential-vortex case, travel-time inversion with the drift supported away from boundary-reaching rays should fail, while splittings should succeed.

Why not just use a symmetric anisotropic operator? Andreux et al.'s operator is symmetric under direction reversal. It therefore produces no splitting between opposite traversal directions, which is the entire signal here. We include it only as a null check that measured splittings are not discretization artifacts.

Proposed Experiments

The first decisive check is the three-flow test in the disk and annulus. The prediction fails if potential flow produces leading-order splitting that scales like the Mach number, or if the potential vortex produces none. We report relative eigenvalue error and splitting error against finite-element ground truth as functions of mode index, from tens up to several hundred, and of Mach number from 0.01 to 0.3.

The second check is the $O(h)$ correction. We compare residual eigenvalue error after leading-order EBK with the predicted orbit integral of the volume-distortion term, using Busemann–Hausdorff and Holmes–Thompson volume conventions. Each convention gives a different S-curvature and so a different prediction. If the residual does not track the S-curvature term, the curvature claim is falsified even if the leading-order claim holds.

The third check is inversion in a stadium cavity with random smooth flows. We report component-wise drift error for spectra alone, travel times alone, and both combined, against the baselines above. We also locate where accuracy collapses with mode count, in analogy with Tomasini et al.'s critical threshold.

The fourth check is on FAUST Remeshed and SCAPE Remeshed. We retrain FLBO-based matching after projecting its drift field onto its co-exact-plus-harmonic part and onto its exact part separately. If accuracy with the exact part alone matches the symmetric-operator baseline, and accuracy with the co-exact part matches the full 62.67%, the gauge structure carries over to discrete shape operators. If not, we learn that the small-asymmetry FLBO is not a consistent discretization of a Randers wave operator, which is itself worth reporting.