Escape over a saddle by coloured noise: theory and numerics

Technical Summary: Escape over a Saddle by Coloured Noise: Theory and Numerics

Problem Statement
Stochastic dynamical systems are essential for modeling transitions induced by random disturbances, particularly escapes from attracting equilibria across the stable manifold of a saddle point. While the small-noise limit is theoretically well-understood via the Freidlin-Wentzell large deviation principle (LDP), existing computational methods face significant limitations. Current approaches typically struggle with:

1. Degenerate noise: Situations where noise does not act in all phase space directions (e.g., noise entering only through momentum equations in mechanical systems).
2. Coloured noise: Realistic forcing with finite correlation times, often modeled via linear filters on white noise, rather than idealized white noise.
3. Unbounded time-horizons: The physical reality that escape events occur over infinite time, whereas numerical methods often rely on finite intervals, leading to inefficiencies and time-translation degeneracies.

Existing methods for computing barrier heights on infinite time-horizons often rely on the invertibility of the noise covariance matrix, rendering them inapplicable to degenerate noise cases. Furthermore, direct sampling of rare events is computationally prohibitive, and standard discretization of infinite-time trajectories wastes resources on regions where dynamics are nearly linear.

Methodology
The authors propose a unified framework combining three distinct methodological innovations to address these challenges:

1. Hamiltonian Optimal Control (HOC) for Degenerate and Coloured Noise:

   • To handle degenerate noise (where the covariance matrix $C = \sigma\sigma^T$ is singular), the paper adopts the Hamiltonian Optimal Control method. Instead of minimizing the action functional involving $C^{-1}$ (which is ill-defined for degenerate noise), the problem is reformulated as a deterministic variational task in an extended phase space.
   • By introducing conjugate variables (adjoints) for each state component, the method derives Hamilton's equations of motion. This approach avoids the inversion of the noise covariance matrix entirely, naturally accommodating degeneracy without requiring explicit pre-conditioning or strict hypoellipticity assumptions.
   • Coloured noise is modeled by augmenting the state space with filter variables governed by linear stochastic differential equations (Ornstein-Uhlenbeck processes). This transforms the coloured noise problem into a standard SDE framework with a higher-dimensional state vector, where the noise acts only on the filter variables, automatically resulting in a degenerate noise structure for the original system.

2. The Method of Division (MOD) for Infinite Time-Horizons:

   • To address the computational infeasibility of simulating infinite time-intervals and the numerical instability caused by time-translation degeneracy, the authors introduce the Method of Division.
   • The transition trajectory is decomposed into three distinct phases:
     • Initial Escape: The system escapes the stable equilibrium (sink) over $(-\infty, 0]$. This phase is treated analytically using linearized dynamics around the sink. The action is computed via the controllability Gramian, solving a Lyapunov equation.
     • Transitional Phase: The core non-linear transition occurs over a finite interval $[0, T]$. This is the only part solved numerically using the HOC method.
     • Final Approach: The system approaches the saddle over $[T, \infty)$. This phase is also treated analytically by linearizing dynamics around the saddle. The unstable direction is handled explicitly, while stable directions flow naturally toward the saddle.
   • Breaking Time-Translation Degeneracy: To fix the numerical ambiguity where optimal trajectories can slide freely along the time axis, the method imposes a terminal constraint. The trajectory is required to cross a predefined transverse ellipsoid around the saddle at time $T$. This constraint breaks the symmetry, ensuring a well-conditioned numerical problem.

3. Numerical Implementation:

   • The overall optimization problem minimizes the sum of the actions from the three phases with respect to the transition path and the optimal endpoints ($\phi_0$ near the sink and $\phi_T$ on the ellipsoid).
   • The Augmented Lagrangian method is employed to solve the constrained optimization problem. This approach convexifies the objective function and updates Lagrange multipliers and penalty parameters iteratively to satisfy the terminal ellipsoid constraint without requiring infinite penalty parameters, thus avoiding ill-conditioning.
   • The algorithm utilizes non-linear conjugate gradient methods for efficient convergence.

Key Contributions

• Extension to Degenerate Noise: The paper provides a robust computational framework for calculating transition rates and most probable paths in systems with degenerate noise, a scenario previously inaccessible to methods relying on covariance matrix inversion.
• Handling Coloured Noise: By integrating linear filters into the state space and utilizing HOC, the method effectively handles physically realistic coloured noise without resorting to white noise approximations.
• Infinite Time-Horizon Solution: The Method of Division offers a novel computational strategy for rare events over infinite time. It isolates the computationally expensive non-linear transition region while analytically resolving the asymptotic behavior near equilibria, significantly enhancing computational efficiency and convergence.
• Removal of Time-Translation Degeneracy: The introduction of a transverse terminal constraint provides a rigorous mechanism to stabilize numerical solutions for infinite-time trajectories.

Results and Applications
The effectiveness of the proposed framework is demonstrated through two specific case studies:

1. Inverted Double-Well Potential: A simplified model used to illustrate the theoretical capabilities of the method in handling degenerate and coloured noise.
2. Ship Capsize Model: A simplified roll–heave model representing a critical application in marine engineering. The method successfully computes the most likely escape paths and the exponential scaling laws of transition rates for a ship under stochastic wave forcing, capturing the dominant exponential contribution to the capsize rate.

Significance
The paper positions this work as a bridge between stochastic analysis and non-autonomous dynamical systems theory, specifically addressing the challenge of ship capsize. It claims to offer a physically realistic and numerically robust approach for safety and reliability engineering, where estimating the rate of very unlikely transitions (such as capsizing) is crucial. By enabling the computation of barrier heights and most probable paths for degenerate and coloured noise over infinite time-horizons, the method provides a rigorous basis for long-term safety assessments that were previously out of reach for existing computational techniques. The work serves as a stochastic companion to previous deterministic approaches in non-autonomous dynamics, integrating these fields to better understand state transitions in complex physical systems.