betaFeedback

WorksMathematics

Diffusion-Score-Based Joint Topology Identification and Generative Particle Filtering

Abstract

Inferring the interaction graph of a networked system while tracking its latent state is central to infrastructure, environmental, and brain networks. Joint identification of interaction topology and latent network states is challenging when observations are intermittent and transition noise is unknown and non-Gaussian. We develop TopoScore-FAF, a diffusion-score-based framework for joint topology identification and generative particle filtering. The graph governs a leaky-diffusion conditional mean, while a frozen diffusion emulator represents the centered transition residual, avoiding the structural ambiguity of unrestricted generative models. First, we establish a score–topology duality that expresses edge-wise transition likelihood gradients as projections of the residual diffusion score. This enables topology learning without evaluating transition densities or differentiating through the generative sampler. Second, we develop diffusion-score-based topology identification with guarantees of identifiability under intermittent observations and invariance of the population minimizer across diffusion noise levels. Sufficient smoothing yields global strong convexity for bounded-plus-Gaussian residuals. As smoothing vanishes, the asymptotic covariance of the state-pair estimator decreases toward the Cramér–Rao bound, with a strict advantage over Gaussian quasi-likelihood under an edge-relevant non-Gaussianity condition. Third, we construct a topology-conditioned generative particle filter that combines guided residual sampling with score-driven graph updates. We establish mean-square closed-loop error bounds under a step-size-independent small-gain condition. The resulting framework uses the graph to structure state propagation and the estimated states to refine the graph. Synthetic and real-data experiments demonstrate its benefits under non-Gaussian forcing.

Paper details

vlnc:W_mstakvxg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…