New neural method detects hidden structures in high-dimensional systems

A new paper on arXiv (2605.24136) by Taj Jones-McCormick tackles the challenge of detecting metastable basins—dynamically distinct regions where a system lingers before rare transitions—in high-dimensional stochastic systems. Traditional approaches rely on spatial discretization or spectral analysis of transition operators, which break down in high dimensions or when basin geometry is nonlinear. The author proves a key insight: if two initial states belong to the same basin, a Bayes-optimal classifier distinguishing their trajectory distributions achieves risk near 1/2 (random guessing); if they are in different basins, risk approaches zero. This transforms basin detection into a two-sample discrimination problem.

Building on this principle, the author develops a neural algorithm that receives candidate basin representatives and iteratively merges them by estimating classification risk with a neural network approximating the Bayes classifier. Evaluated on synthetic systems where low-dimensional dynamics are embedded in high-dimensional noisy ambient spaces, the method accurately recovers basin structure where standard spectral and clustering methods fail. This highlights trajectory discrimination as an effective tool for dynamical analysis in high dimensions.