Illumination-Aware Regularization for Direction-Dependent Travel-Time Inversion

In many physical media a wave's speed depends on the direction it travels as well as on where it is. Seismic waves move faster along rock layers than across them. Fire runs downwind faster than upwind. Mathematicians describe such media with Finsler geometry, a generalization of ordinary curved-space geometry in which the "length" of a small step depends on the step's direction. They have proved that arrival-time measurements at a boundary determine such a medium, but only along the directions that the fastest rays actually travel. Practitioners who try to invert real data hit a different wall: their results swing wildly with the choice of regularization weight, the knob that trades data fit against smoothness, and they end up grid-searching it for each new dataset.

This note argues that these are the same problem seen from two sides. The theory says each location and direction pair is either constrained by the data or not, and the two cases are sharply separated. A single scalar regularization weight is asked to serve both cases at once, and it cannot. We propose to measure, per location and per direction, how strongly the data constrain the medium. That measurement comes almost for free from the gradient of a differentiable travel-time solver. We then use it to set the regularization separately for well-observed and unobserved directions. The prediction is concrete: one global weight then transfers across datasets and sensor layouts. In the same experiments, the never-implemented recovery theorems get their first empirical test.

Core Claim

The direction-dependent recovery theorems already say where tuning sensitivity must come from. Determination of a compact Finsler manifold from its boundary distance map proves that boundary travel times fix the Finsler function on a set swept by distance-minimizing rays (Theorem 1.3). It proves recovery everywhere only if the function is real-analytic in the direction variable (Theorem 1.5). The broken-scattering and sphere-data papers (de Hoop et al., broken scattering; de Hoop et al., reconstruction along a geodesic) stress the same point: knowing the speed in one direction at a point says almost nothing about the other directions there. Our claim has two parts. In a discretized inversion, almost all of the dependence of the recovered medium on the regularization weight sits in the poorly illuminated (location, direction) pairs. Weighting the regularizer by a per-direction illumination map removes most of that dependence. All three papers above are pure theory with no numerical results. We use them as the motivation and as the test target, not as a method.

Research Program

The forward model follows Wildfire Simulation with Differentiable Randers-Finsler Eikonal Solvers. The eikonal equation is the equation for first-arrival times. That paper solves it with a fast sweeping method, a grid solver that updates arrival times in alternating sweep orders. It uses a Randers metric $F(x,v)=\sqrt{v^\top G(x)v}+b(x)^\top v$, where $G$ encodes direction-dependent speed and $b$ a one-way drift such as wind. It obtains exact gradients by implicit differentiation of the converged solution, which it reports as about 755x faster than finite differences. That paper never studies how the loss balance or regularization scale affects results. Filling that gap is our starting point.

We add a parameterization of the speed profile at each grid cell by a low-order Fourier series in angle, with $F(x,\theta)$ for 2D angle $\theta$. A truncated Fourier series is real-analytic in direction, so it mimics the hypothesis of Theorem 1.5. This lets us test whether that hypothesis does practical work. The parameterization contains Randers and elliptic metrics as special cases, and it can also represent non-elliptic qP-like profiles.

The illumination map is the diagonal of the Gauss-Newton matrix, which approximates the curvature of the data-fit loss:

$$H(x,\theta)=\sum_{k}\Big(\frac{\partial T_k}{\partial F(x,\theta)}\Big)^2.$$

Here $T_k$ is the predicted arrival time for source-receiver pair $k$. Each term is obtained by one adjoint solve through the implicit-differentiation machinery. $H$ is nonzero only where rays pass, and only in the directions they travel. The objective is

$$\mathcal{L}=\sum_k \frac{(T_k-T_k^{\rm obs})^2}{\sigma^2}+\lambda_0\sum_{x,\theta}\frac{\bar H}{H(x,\theta)+\bar H}\,\|\nabla_{x,\theta}F\|^2 .$$

In this objective, $\sigma$ is the noise level, $\bar H$ is the median of $H$ over illuminated cells, and $\lambda_0$ is the single remaining weight.

The logic is simple. Where $H=0$, only the smoothness term acts. Minimizing a smoothness penalty alone gives a solution that does not depend on its scale, so dark directions are filled by smooth extension from lit ones whatever $\lambda_0$ is. Where $H$ is large, the penalty is down-weighted in proportion, so the data decide. Rays bend as $F$ changes, so we recompute $H$ every few outer iterations.

The program then asks three things. First, does the recovery error, stratified by illumination, reproduce the theorems' prediction: accurate where minimizing rays sample, arbitrary elsewhere? Second, does the analytic angular basis shrink the error in dark directions when the true medium is analytic in direction, but not when it is not? Third, does $\lambda_0$ transfer across media and sensor layouts?

Baselines

The naive reference point is the starting model with no inversion, plus an isotropic inversion that ignores direction dependence. Together they show how much signal direction-dependent modeling recovers at all.

The strongest existing method is the scalar-weight Randers inversion of the wildfire Randers-Finsler paper. On Sim2Real-Fire it reports synthetic recovery of about 6% error on isotropic metrics and under 3% on drift fields. It reports perimeter-prediction test correlations of 0.824 within scene and 0.766 across scenes. We give it a uniform smoothness weight, tuned two ways: by oracle selection on held-out ground truth, which is an upper bound, and by Bayesian optimization. The Bayesian optimization gets the same total number of forward-adjoint solves that our method spends on recomputing $H$ plus its own run. Same grids, data, noise, and solver are used throughout. Beating it means matching its oracle-tuned error with a fixed $\lambda_0$ that was never retuned. It also means degrading much less as $\lambda_0$ moves over two orders of magnitude. Our pre-set criterion is that worst-case error over that range is within 1.2x of the best, and we report the baseline's ratio alongside.

A classical automatic rule, Morozov's discrepancy principle (choose the scalar weight so the misfit equals the noise level), is a third baseline. Two ablations isolate the key ingredient. One collapses $H$ over angle into a location-only illumination map, which tests whether the direction resolution matters. The other randomly permutes $H$ across cells, which tests whether the specific map matters rather than just any spatially varying weight.

Why not just…?

Why not just tune the scalar weight with Bayesian optimization or grid search? In 2D this is affordable, so it is a real competitor. It is the reason the oracle-tuned baseline exists. The issue is transfer, not cost. Efficient Continuous Control with Double Actors and Regularized Critics found that regularization weights tuned on one task fail on another. In 3D seismic settings, one inversion took about three days in Full Waveform Inversion Guided by Travel Time Tomography, so repeated tuning is out of reach. The decisive test is to tune on one sensor layout and evaluate on another. If tuned scalar weights transfer as well as our fixed $\lambda_0$, the proposal fails.

Why not just use the discrepancy principle? It sets one weight automatically from the noise level, which removes the knob without new machinery. But it still applies one weight to lit and dark directions alike. Our analysis predicts it will over-smooth well-sampled directions in order to tame unsampled ones. The direction-collapsed ablation and the illumination-stratified error plots show directly whether this happens.

Why not just restrict to elliptic (Riemannian) anisotropy, with fewer parameters and less to tune? This is a real threat on media that really are elliptic. There we expect it to tie, and we will report that. The sphere-data reconstruction paper argues that a Riemannian restriction loses qP-wave features. An elliptic metric also cannot represent the one-way drift that the wildfire paper models. We include non-elliptic and drift-dominated test media precisely to measure the cost of this mis-specification.

Why not just let a neural encoder regularize implicitly, as the wildfire paper does? An encoder learns a prior over media, but it does not know which directions this particular dataset failed to observe. We run the encoder pipeline with and without our weighting. If the encoder alone removes sensitivity to its loss weights, the illumination map adds nothing in learned settings, and we would say so.

Proposed Experiments

On synthetic media, we take velocity structures from OpenFWI, from the easy FlatVel family through the hard CurveFault family. OpenFWI itself is isotropic acoustic data, with no anisotropic travel-time result reported on Lacuna, so we impose controlled direction dependence on these structures. We use three types: elliptic, qP-like non-elliptic, and Randers drift. We compute boundary travel times under two sensor layouts. One is dense. The other is sparse and mirrors the 5-shot, 70-receiver geometry of OpenFWI 2.0. We sweep noise levels.

The metrics are relative error in $F$, stratified by illumination quantile; mean absolute error and structural similarity of the recovered speed maps; and the worst-to-best error ratio across a $\lambda_0$ sweep.

The first decisive check tests the theory. Error should collapse in high-illumination strata for every method and every $\lambda_0$. Variation across $\lambda_0$ should concentrate in low-illumination strata for the scalar baselines and nearly vanish for ours.

The second check is transfer: fix $\lambda_0$ on FlatVel with dense sensors, then evaluate unchanged on CurveFault with sparse sensors.

The third check is on Sim2Real-Fire, where fire spreads mostly downwind, so upwind directions are almost never observed. This makes it a natural test of dark-direction behavior. We compare drift recovery and perimeter correlation against the reported 0.824 and 0.766. Here the bar is not a higher correlation. It is matching it without per-scene retuning while reporting calibrated uncertainty in upwind directions.

The project fails if the scalar baselines' sensitivity turns out not to concentrate in poorly illuminated directions. That outcome would itself be a useful negative result about how far the theorems carry into practice.