QD-PINN:A Physics-Informed Neural Network Approach to Bayesian Quickest Detection in High-Dimensional Diffusions

Hang Ye, Yi Liu, Haipeng XingarXiv 2026 · arXiv

This paper presents QD-PINN, a novel framework that leverages Physics-Informed Neural Networks (PINNs) to solve Bayesian quickest detection problems (QDPs) in high-dimensional settings. By reformulating the problem as a dual-network optimization task, the authors successfully estimate optimal stopping boundaries for diffusion processes in up to 1,000 dimensions, effectively overcoming the "curse of dimensionality" that plagues traditional numerical methods.

TL;DR

QD-PINN uses a dual-network physics-informed architecture to jointly learn the value function and the optimal stopping boundary for high-dimensional change-point detection. It provides a mathematically rigorous bridge between free-boundary PDEs and deep learning, demonstrating stable and accurate results in systems with up to 1,000 dimensions.

What Problem Does This Paper Address?

Sequential change-point detection (or "quickest detection") is the art of identifying a shift in a stochastic process's distribution as fast as possible while keeping false alarms to a minimum. This is critical in fields like high-frequency finance, cybersecurity, and industrial IoT monitoring.

Mathematically, these are optimal stopping problems characterized by Free-Boundary Problems (FBPs). The "free boundary" is the threshold that separates the "continuation region" (keep observing) from the "stopping region" (declare a change). Solving these in high dimensions is notoriously difficult because:

  1. Curse of Dimensionality: Traditional grid-based numerical schemes scale exponentially with the number of variables.
  2. Coupling: Spatial correlations (cross-derivatives in the PDE) prevent the problem from being broken down into simpler, independent parts.
  3. Unknown Geometry: Standard Monte Carlo methods struggle because the very boundary they need to find is what defines the stopping criteria.

Methods

The authors propose a mesh-free approach using two deep feedforward neural networks: one to approximate the value function $V(\pi, x)$ and another for the stopping boundary $S(x)$.

Key Technical Innovations:

  • Theoretical Equivalence: The authors prove that the free-boundary PDE is equivalent to an infinite-dimensional stochastic optimization problem. This provides the mathematical justification for using finite-dimensional neural networks to find the solution.
  • Dual-Network Architecture: Unlike standard PINNs that might only solve a PDE, this setup uses a "Boundary Network" to explicitly parameterize the stopping threshold, which is then fed into the "Value Network."
  • Fischer-Burmeister Loss: To handle the complementarity constraints (where one of two conditions must be zero), they use the Fischer-Burmeister function. This prevents the networks from converging to "degenerate" solutions where the model erroneously assigns the entire domain to just one region.
  • Kolmogorov Transformation: They employ this transformation to enforce "smooth-fit" conditions (ensuring the solution is differentiable at the boundary) directly in the original high-dimensional state space.
  • Adaptive Sampling: Because the boundary is the most critical part of the domain, they use a residual-based adaptive sampling strategy (Algorithm 1) to concentrate training points near the evolving estimated boundary.
  • Hutchinson Trace Estimation: To keep computation feasible at $n=1000$, they use this technique to estimate high-dimensional second-order operators (the Hessian) without explicitly constructing them.
Dual architecture neural network used to approximate the value function and stopping boundary.

Datasets

The method was evaluated using synthetic but mathematically rigorous benchmarks where analytical or high-precision numerical solutions are available for comparison:

  • High-Dimensional Brownian Motion: Detecting a change in drift.
  • Multivariate Ornstein-Uhlenbeck (OU) Processes: Detecting changes in mean reversion speed or long-term mean.
  • Correlation Structures: The experiments tested various spatial dependencies, including Tridiagonal and Toeplitz correlation matrices, to simulate real-world "coupled" data streams.

Key Insights and Ideas

  • Mesh-Free Scalability: The most significant "aha" moment is the demonstration that PINNs can handle 1,000 dimensions (Figure 9) for a problem type that usually caps out at 5-10 dimensions with grid-based methods.
  • Stability through Uncertainty: By using Deep Ensembles and Monte Carlo Dropout, the authors show that the learned boundaries are remarkably consistent across different random initializations.
  • The Importance of Sampling: Uniform sampling fails in high dimensions because the "volume" of the boundary region becomes vanishingly small relative to the whole space. The adaptive sampling strategy is what makes the convergence possible.

Key Results

  • High Accuracy: In Brownian motion tests, the PINN estimates for the stopping boundary were within $10^{-3}$ absolute error of the exact analytical solution ($S \approx 0.556$).
  • Robustness to Dimension: The training loss remained stable as the dimension $n$ increased from 1 to 500 (and eventually 1,000), showing that the method effectively bypasses the curse of dimensionality.
  • State-Dependent Boundaries: For OU processes, the model successfully captured U-shaped and other complex topological structures of the stopping boundary that vary based on the current state of the system.
Uncertainty quantification results for the Wiener process (top) and the OU process (bottom).

Limitations and Future Work

  • Synthetic Dynamics: The current work assumes the underlying SDE dynamics (drift and diffusion coefficients) are known. In many real-world scenarios, these must be estimated simultaneously.
  • Computational Cost: While it beats grid methods, training still takes 30–50 minutes on a high-end laptop GPU for $n=500$. Further optimization of the adaptive sampling could speed this up.
  • Future Directions: The authors suggest extending this to "non-Bayesian" formulations (like Lorden’s or Pollak’s criteria) and exploring more complex diffusion processes where the noise might be non-Gaussian.

Summary updated

Abstract

Quickest detection problems (QDPs) arise in a wide range of applications. However, the complementarity structure of the associated variational inequalities poses substantial challenges for traditional numerical methods, particularly in high-dimensional settings. We propose a PINN-based framework for estimating the optimal stopping boundary in QDPs with high-dimensional diffusion processes and spatial dependence, while overcoming the curse of dimensionality. Theoretically, we establish a novel equivalence between the free-boundary PDE and an infinite-dimensional stochastic optimization problem, rigorously justifying the use of finite-dimensional neural networks to approximate the true solution. Furthermore, we assess the robustness of the proposed framework by quantifying uncertainty in the learned stopping boundaries using deep ensembles and Monte Carlo dropout. The results show low variability across independent training runs and stochastic forward passes, indicating stable and consistent boundary estimates. We also compare our PINN results with reference numerical solutions where available and find close agreement.
BibTeX
@misc{ye2026qd,
  author = {Hang Ye and Yi Liu and Haipeng Xing},
  title = {{QD-PINN:A Physics-Informed Neural Network Approach to Bayesian Quickest Detection in High-Dimensional Diffusions}},
  year = {2026},
  eprint = {2610.37473},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.37473},
}
Figure