Fast Risk Certification of Candidate Trajectories under Uncertain Time-Varying Constraints

Srimanta Santra, Oleksii Molodchyk, Matti Noack, Timm FaulwasserarXiv 2026 · DOI · arXiv

TL;DR

This paper introduces a computationally efficient method to verify if a pre-planned trajectory is safe under uncertain, time-varying conditions. It achieves this by building a "surrogate" model offline and then using it for rapid, conservative risk assessments online, significantly outperforming direct simulation for complex systems.

What Problem Does This Paper Address?

In safety-critical applications like autonomous driving or chemical process control, it's not enough for a planned trajectory to be safe under ideal conditions. It must also be demonstrably safe when faced with uncertainties, such as unpredictable environmental factors, sensor noise, or variations in system parameters. Directly checking this safety for every possible uncertain scenario is computationally prohibitive, especially for complex, time-consuming models. This paper tackles the challenge of providing a fast, rigorous, and reliable safety certificate for a fixed candidate trajectory without needing to re-plan or re-simulate the entire system repeatedly.

Methods

The core of the paper's approach is an offline-online scheme:

  • Offline Stage:
    • Surrogate Model Construction: The paper proposes two methods to create a "surrogate" model that approximates the constraint violation behavior of the system under uncertainty.
  • Polynomial Chaos Expansion (PCE): Used when the uncertainty distribution is known and the system's behavior is smooth. PCE builds a structured, low-dimensional representation of the uncertainty.
  • Kernel Regression: A more flexible, data-driven approach used when the uncertainty distribution is unknown or the system's behavior is complex. It learns the relationship from sampled data.
  • Uniform Surrogate Error Bound: An estimate of how accurate the surrogate model is across the entire range of uncertainties is computed.
  • Online Stage:
    • Conservative VaR Calculation: The surrogate model is used to quickly estimate the worst-case Value-at-Risk (VaR) margin. VaR is a standard risk measure that quantifies the potential for losses (constraint violations) at a given confidence level.
  • Finite-Sample Guarantees: The paper derives theoretical guarantees for the conservatism of the VaR estimate. These guarantees are "distribution-free," meaning they don't assume a specific probability distribution for the constraint violations, relying instead on statistical inequalities (like the Dvoretzky-Kiefer-Wolfowitz inequality) and a union bound over all constraints and time points.
  • Certification Decision: If the computed conservative VaR margin is below a safety threshold (e.g., zero, indicating no violation probability above the risk level), the trajectory is certified as safe.

Datasets

The paper does not introduce new datasets. Instead, it uses a simulated dataset derived from a crystallization population balance model (PBM). This model is chosen because it represents a complex, computationally expensive system where direct Monte Carlo simulations for risk assessment are infeasible for real-time applications. The PBM's uncertainty stems from variations in growth and nucleation kinetics.

Key Insights and Ideas

  • Separation of Concerns: The key insight is to separate the computationally intensive task of understanding the system's uncertain behavior (offline surrogate construction) from the rapid decision-making process (online VaR evaluation).
  • Surrogates as Accelerators: Surrogate models (PCE and kernel regression) are powerful tools for approximating complex functions, enabling drastically faster risk assessments compared to direct simulations.
  • Rigorous, Distribution-Free Guarantees: The paper provides theoretical backing for the safety certificate, ensuring it's not just a heuristic but a statistically sound guarantee, even without knowing the exact probability distribution of violations.
  • Worst-Case VaR for Comprehensive Safety: Aggregating pointwise VaR into a worst-case VaR margin provides a single, interpretable metric for overall trajectory safety across all constraints and time steps.

Key Results

  • Significant Speedup: The case study demonstrated a dramatic reduction in evaluation time. Direct Monte Carlo simulation of the PBM took approximately 7164.6 ms, while the PCE and kernel surrogates achieved the same risk assessment in 0.034 ms and 0.189 ms, respectively.
  • Conservative Risk Tracking: The surrogate-based risk estimates closely tracked the results from direct Monte Carlo simulations, indicating that the surrogates capture the essential risk dynamics.
  • Practical Certification: The framework successfully provided a conservative upper bound on the risk, allowing for a clear "certify or reject" decision for the candidate trajectory.
  • Theoretical Guarantees: The paper provides finite-sample upper bounds for the VaR margin and shows how surrogate approximation errors translate to the certified VaR, ensuring the conservatism of the final decision.

Limitations and Future Work

  • Sample Size Requirement: The finite-sample guarantees depend on the number of samples used for training and evaluation. For very low risk levels (e.g., δ=0.10) and high confidence (e.g., β=0.05), a large number of samples might be required to satisfy the theoretical conditions for a formal certificate, as seen in the case study where the sample size was insufficient for a strict guarantee.
  • Continuous-Time Interpolation: The current framework operates on a discrete grid. Future work could explore continuous-time interpolation guarantees.
  • Less Conservative Bounds: The paper acknowledges that the derived bounds can be conservative. Developing less conservative confidence bounds and operational constructions of surrogate error bounds is a promising direction.
  • Joint Chance Guarantees: The current certificate is pointwise. Extending it to provide joint chance guarantees over the entire horizon would be valuable.

Summary updated

Abstract

This paper studies the certification of a fixed candidate trajectory on a finite certification grid under parametric uncertainty. For each constraint-time pair, we define a scalar measure of constraint violation and aggregate the resulting pointwise chance constraints into a worst-case Value-at-Risk (VaR) margin. The goal is not to generate a new trajectory, but to assess online whether a trajectory produced by a planner or predictive controller is sufficiently safe on the certification grid. Direct evaluation requires repeated uncertainty propagation and is often too expensive for computationally demanding models. We therefore adopt an offline-online scheme: offline, a surrogate of the constraint violation map along the candidate trajectory is constructed using polynomial chaos expansion (PCE) when the uncertainty law is known, or kernel regression when only sampled input-output data are available; online, the surrogate is sampled to evaluate conservative VaR bounds at low computational cost. On the theoretical side, we derive a finite-sample upper bound for the grid-based VaR margin using empirical quantiles, the Dvoretzky-Kiefer-Wolfowitz (DKW) inequality, and a union bound over all constraint-time pairs, without assuming a parametric family for the underlying violation distribution. We also show how a uniform surrogate error bound transfers to the certified VaR margin. The approach is illustrated on a crystallization population balance model, where the surrogate-based risk estimates track direct Monte Carlo results while substantially reducing online evaluation time.
BibTeX
@misc{santra2026fast,
  author = {Srimanta Santra and Oleksii Molodchyk and Matti Noack and Timm Faulwasser},
  title = {{Fast Risk Certification of Candidate Trajectories under Uncertain Time-Varying Constraints}},
  year = {2026},
  eprint = {2606.31566},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.31566},
}
Figure