Wulff-Sphere Directional Sampling for Rare Failures

Most tools for estimating tiny failure probabilities of autonomous systems quietly assume the random disturbances are Gaussian, so that "how unlikely is this input" is measured by a round Euclidean distance. Many real disturbances do not look like that. Examples include actuator commands clipped to a polytope, sensor errors with Laplace-shaped tails, and gusts that are stronger along one axis in a non-elliptical way. For these, the natural measure of how extreme an input is comes from a general convex "norm" $F$ whose unit ball can be any symmetric convex body, such as a cube, a diamond, or a crystal-like polytope. We propose to handle this case exactly rather than approximately. When the disturbance density depends on the input only through $F$, every sample splits into two independent parts: a scalar radius, which has a known one-dimensional law, and a direction, drawn from a fixed measure on the boundary of the unit $F$-ball. The failure probability then equals an average, over directions, of a one-dimensional tail probability evaluated at the distance along that ray where failure begins.

This matters for two reasons. First, the estimator needs no smoothness of the failure boundary, only root-finding along rays, so it fits black-box simulators with hard software thresholds. Second, its variance has a clean geometric meaning, and that geometry is exactly what anisotropic capacity analysis studies. The estimator has zero variance when the failure boundary is a scaled copy of the unit $F$-ball (the "Wulff shape" of the geometric-analysis literature). Its difficulty near the most likely failure direction is controlled by how the boundary's curvature, measured in the $F$ geometry, departs from the curvature of that shape. We would build the estimator, prove its asymptotic behavior, and test whether it beats Gaussianize-then-sample pipelines on non-Gaussian versions of standard failure benchmarks.

Core Claim

Write the disturbance density as $c\,h(F(x))$ on $\mathbb{R}^n$, where $F$ is a convex, positively homogeneous gauge and $h$ is decreasing (Laplace-type $h(r)=e^{-r}$, generalized Gaussian $h(r)=e^{-r^\alpha}$, and others). Then $X = R\,\Theta$. Here $R$ has density proportional to $r^{n-1}h(r)$, and $\Theta$ is independent of $R$ and follows the cone measure on the unit $F$-sphere $\{F=1\}$; the cone measure is what you get by sampling uniformly inside the unit ball and pushing each point radially out to its boundary. With $\bar G$ the tail function of $R$ and $\rho(\theta)=\inf\{r>0: r\theta \in E\}$ the distance along direction $\theta$ at which the failure set $E$ begins,

$$ p = \mathbb{P}(X\in E) = \mathbb{E}_{\Theta}\big[\bar G(\rho(\Theta))\big] $$

holds exactly whenever each ray meets $E$ in a single half-line; a star-shaped complement of $E$ guarantees this. For general $E$, the tail term is replaced by a sum over the intervals where the ray lies in $E$. For the Gaussian case, where $F$ is the Euclidean norm and $R$ is chi-distributed, this reduces to the classical directional simulation of structural reliability. The new content is threefold. We carry the construction to arbitrary gauges. We pair it with importance sampling over directions, centered on anisotropic dominating directions $\theta_* \propto \arg\min_{x\in E}F(x)$. And we give a curvature-based account of its variance and asymptotics.

The analytic hypothesis is as follows. Let $r_*=\rho(\theta_*)$. Near $\theta_*$, $\rho$ grows quadratically with a Hessian set by the gap between the $F$-relative shape operator of $\partial E$ and that of the Wulff sphere of radius $r_*$ touching $E$ at the dominating point. Laplace's method on the $F$-sphere should then give $p \approx \bar G(r_*)\cdot C(\theta_*, r_*)\cdot \det(S)^{-1/2}$. In this formula $S$ is that curvature gap, scaled by $r_*$, and $C$ collects the cone-measure density at $\theta_*$ and dimension-dependent constants. We treat this as a conjecture to prove and check numerically. It has two consequences. It gives a closed-form anisotropic analogue of second-order reliability corrections. It also says that the width of the optimal angular proposal is dictated by the same anisotropic second fundamental form that drives the Wulff-rigid inequalities in capacity theory.

Relation to Prior Work

Two papers by Li and Xiong supply the geometric language and the conditions this proposal borrows. Sharp bounds for the anisotropic p-capacity of Euclidean compact sets proves sharp capacity bounds via inverse anisotropic mean curvature flow, with equality exactly for Wulff shapes. Sharp upper bounds for the capacity in the hyperbolic and Euclidean spaces does the analogous work in hyperbolic and Euclidean space for star-shaped, mean-convex sets. Both parametrize star-shaped boundaries as radial graphs over a sphere, which is precisely our $\rho(\theta)$, and both identify Wulff shapes as the extremal case. Our zero-variance condition is the probabilistic mirror of their rigidity statements.

We deliberately do not claim that anisotropic $p$-capacity is itself a failure probability. For $p=2$ and quadratic $F$, capacity does govern hitting probabilities of a linear diffusion, but that case is removable by whitening. For non-quadratic $F$, the anisotropic Laplacian is nonlinear and has no equally clean probabilistic reading. So capacity theory enters as the source of geometry (radial parametrization, $F$-curvature, Wulff rigidity), not as a plug-in bound. Coşkunuzer's embedding result for hyperbolic least-area planes (Least area planes in hyperbolic 3-space are properly embedded) is not used.

The closest rare-event work is Rare-event Simulation for Neural Network and Random Forest Predictors. It finds Euclidean dominating points on piecewise-linear failure regions by mixed-integer programming (MIP) and centers Gaussian mixtures there, obtaining asymptotic optimality under Gaussian inputs. We reuse its MIP machinery, since polyhedral gauges such as $\ell_1$, $\ell_\infty$, or actuator polytopes keep the program linear. We differ in the metric, and in replacing shifted Gaussian proposals by exact radial integration plus angular sampling. Shifted Gaussians are mis-specified under Laplace-tailed inputs, because their weights $P/Q$ grow without bound in the tails.

Evaluating Trajectory Collision Probability through Adaptive Importance Sampling for Safe Motion Planning uses Mahalanobis-distance collision modes, which is the ellipsoidal special case of our geometry. The affine-invariant Langevin samplers (Affine invariant interacting Langevin dynamics in Markov chain importance sampling for rare event estimation, Affine Invariant Langevin Dynamics for rare-event sampling) adapt to linear stretching through ensemble covariances. They therefore cannot represent non-elliptical level sets, and they require gradient-like information we avoid. Generative samplers, namely Rare Event Probability Learning by Normalizing Flows (NOFIS) and Diffusion-Based Failure Sampling for Evaluating Safety-Critical Autonomous Systems (DiFS), learn failure distributions flexibly but pay in simulator calls and training time. They are the strongest general-purpose competitors, not structural relatives. Finally, the documented mode collapse of the cross-entropy method in Failure Probability Estimation for Black-Box Autonomous Systems using State-Dependent Importance Sampling Proposals (SPAIS) motivates using multiple dominating directions.

Research Program

The first component is the estimator. Directions are drawn either from the cone measure (sample uniformly in the unit $F$-ball by hit-and-run, then normalize) or from an angular importance proposal. For each direction, a coarse scan followed by bisection along the ray locates every entry into $E$, and the exact radial tail mass is summed. The angular proposal is a mixture of pushed-forward distributions: sample a point from a local law around each dominating point $x_k$, then project it radially onto $\{F=1\}$. Its density on the sphere is a one-dimensional integral along the ray, computed numerically, so importance weights remain exact.

Dominating points come from a MIP with polyhedral $F$ when the system is a ReLU network or random forest, as in the MIP paper. For black-box simulators they come from multi-start minimization of $F(x)$ subject to failure, using the same ray search. Mixture weights are set by the curvature-corrected asymptotic formula rather than equally, and then adapted.

The second component is analysis. We aim to prove unbiasedness when every ray entry is found. We aim to prove bounded relative error as $p\to 0$ for angular proposals whose width matches the curvature gap $S$, compared with only logarithmic efficiency for untuned widths. We will also characterize zero variance as occurring exactly when $\partial E$ is a centered Wulff sphere. Finally, we will derive the anisotropic Laplace formula above, including the degenerate case where $\partial E$ has a flat face aligned with a face of a polyhedral $F$. In that case $S$ is singular and the asymptotics change order, which matters for polyhedral gauges.

The third component is benchmarks with non-Gaussian disturbances. Here the ground truth is known or cheaply computable to high precision, and the disturbance model is plausible.

Baselines

The naive reference is plain Monte Carlo under the true non-Gaussian law. Its coefficient of variation is about $1/\sqrt{Np}$, so it serves only on problems with $p \ge 10^{-4}$ and as a check of the ground truth. On the SPAIS pendulum it has a reported absolute error of 0.81 ± 0.65, against 0.06 ± 0.03 for SPAIS (Failure Probability Estimation for Black-Box Autonomous Systems using State-Dependent Importance Sampling Proposals); that comparison is under the original disturbances.

The strongest structural competitor is a Gaussianize-then-sample pipeline. Each input is mapped to a standard Gaussian by its marginal quantile transform, or by a Rosenblatt transform when the components are dependent, and the Euclidean dominating-point mixture of Rare-event Simulation for Neural Network and Random Forest Predictors or classical Gaussian directional simulation is run in the transformed space. This is the method a practitioner would reach for first, and for independent-coordinate laws it is exactly valid. The paper's reported gain on MAGIC and synthetic piecewise-linear models is that the ratio of relative errors between uniform importance sampling and its dominating-point sampler rose from 2 to 5–6 as rarity increased, under Gaussian inputs.

The strongest general-purpose competitors are NOFIS, which reports error 0.61 on the 62-dimensional ResNet18 reliability task using 18,000 samples (Rare Event Probability Learning by Normalizing Flows), and DiFS, which reports failure discovery rate above 0.66 on its benchmarks (Diffusion-Based Failure Sampling for Evaluating Safety-Critical Autonomous Systems). We also include the cross-entropy method with heavy-tailed (Laplace or Student-t) mixture proposals, and subset simulation. For every non-Gaussian variant we construct, there is no reported result found on Lacuna, so all numbers there will come from our own runs.

Fairness rests on counting simulator calls, not samples. Our method spends a scan plus a bisection per direction, typically tens of calls. Every method gets the same total call budget, the same hyperparameter-search budget measured in calls, and the same wall-clock limit. Results are averaged over at least 50 independent repetitions.

Beating the competition means a lower relative root-mean-square error at equal calls. It also means a relative error that stays flat, rather than growing, as $p$ is pushed from $10^{-4}$ to $10^{-10}$ by scaling the failure threshold. Three ablations isolate the ingredients. The first replaces the $F$-geometry with Euclidean directional simulation in the original coordinates. The second uses uniform instead of dominating-direction angular sampling. The third uses equal instead of curvature-based mixture weights.

Why not just…?

Why not just transform the inputs to Gaussians and use existing methods? This is the real threat, and we treat it as the main baseline rather than dismiss it. For independent Laplace or generalized-Gaussian coordinates, the marginal transform is exact. After that, Gaussian dominating-point sampling is fully legitimate, although the transform bends flat failure faces into curved ones. We expect it to remain competitive there, and if it matches us on independent-coordinate laws we will say so.

The expected advantage appears when $F$ is not a product norm. A rotated $\ell_1$ ball, an actuator polytope coupling axes, or a crystalline gauge has no closed-form Rosenblatt transform. Numerical transforms add error precisely in the tails, and the transformed failure set can become non-star-shaped and multimodal. The experiments sweep from product to strongly non-product gauges, so they will show where, if anywhere, the crossover lies.

Why not just put a Gaussian mixture at the dominating points, as the MIP paper does? Under a Laplace-tailed nominal law, a shifted Gaussian proposal has lighter tails than the target. The weight $P/Q$ is then unbounded and the estimator's variance can be infinite, so the method is not merely less efficient. A Euclidean dominating point is also generally not the most likely failure point under $F$. We will report the empirical weight tails and the pointwise error of the Euclidean method on problems where the $F$-dominating point and the Euclidean one differ.

Why not just train NOFIS or DiFS on the non-Gaussian law? They can represent the right distribution in principle. NOFIS, however, reported needing 18,000 calls on a 62-dimensional task, and DiFS reported training in hours against minutes for simpler proposals. Neither returns the analytic curvature prediction we test. If they match our error at equal simulator calls, the structural approach loses its practical case and keeps only its theoretical content. The equal-call comparison is designed to settle exactly that.

Proposed Experiments

The first tier uses synthetic problems with exact answers. Under Laplace, $\ell_\infty$-symmetric, and random polyhedral gauges in dimensions 2 to 40, we use half-spaces, complements of Wulff spheres, and the Leaf, Cube, and Rosenbrock limit states from the NOFIS suite. Ground truth comes from one-dimensional characteristic-function inversion for half-spaces, and from very long runs otherwise. The rarity threshold is swept so that $p$ ranges from $10^{-4}$ to $10^{-10}$. Two checks are decisive here. The estimator must have numerically zero variance on centered Wulff spheres. And the curvature-based asymptotic formula must match the true $p$ to within a few percent at $p=10^{-8}$ on smooth curved boundaries. A systematic miss would falsify the curvature claim.

The second tier uses piecewise-linear predictors. These are ReLU networks and random forests trained on UCI MAGIC Gamma Telescope, with misclassification probability under Laplace or polytope perturbations of the 10 features, and the ResNet18 62-dimensional reliability task from NOFIS under the same non-Gaussian perturbations. Dominating points come from a linear MIP. We will also rerun the Gaussian versions to confirm that our method, reduced to directional simulation, is not worse than the reported NOFIS figure. That rerun checks our implementation; it is not a claim of superiority.

The third tier uses sequential black-box systems. These are the SPAIS and DiFS pendulum and the F-16 ground collision avoidance model, with torque or observation noise replaced by saturated, polytope-shaped, or Laplace-tailed sequences. Rays here scale the whole disturbance sequence. The metrics are relative error at equal calls, and coverage of both falling directions of the pendulum, which is where the cross-entropy method collapsed in SPAIS. The decisive test of the whole proposal is simple. On non-product gauges, at equal simulator calls, the $F$-directional estimator must hold bounded relative error across four decades of rarity, while the Gaussianize-then-sample baseline and the Euclidean ablation must degrade. If they do not degrade, the anisotropic geometry is not earning its keep.