Simplified Gentlest Ascent Dynamics for Saddle Points in Non-gradient Systems

TL;DR

This paper introduces a simplified gentlest ascent dynamics method for locating saddle points in non-gradient systems, reducing computational cost and complexity while maintaining convergence properties, with applications demonstrated on various models.

Contribution

A new simplified GAD method requiring only one direction variable for non-gradient systems, improving efficiency and avoiding transpose operations.

Findings

Reduces computational cost by half for direction variables.

Maintains convergence properties of the original GAD.

Successfully applied to multiple non-gradient models, including PDEs.

Abstract

The gentlest ascent dynamics (GAD) (Nonlinearity, vol. 24, no. 6, p1831, 2011) is a continuous time dynamics coupling both the position and the direction variables to efficiently locate the saddle point with a given index. These saddle points play important roles in the activated process of the randomly perturbed dynamical systems. For index-1 saddle points in non-gradient systems, the GAD requires two direction variables to approximate the eigenvectors of the Jacobian matrix and its transpose, respectively, while in the gradient systems, these two directions collapse to be the single min mode of the Hessian matrix. In this note, we present a simplified GAD which only needs one direction variable even for non-gradient systems. This new method not only reduces computational cost for directions by half, but also can avoid inconvenient operations on the transpose of Jacobian matrix. We prove the same convergence property for the simplified GAD as for the original GAD. The motivation of our simplified GAD is its formal analogy to the Hamiltonian dynamics governing the exit dynamics when the system is perturbed by small noise. Several non-gradient examples are presented to demonstrate our method, including the two dimensional models and the Allen-Cahn equation in the presence of shear flow.