This is the authors' abstract. We don't add key points for this paper.
Nonlinear state estimation under indistinguishability consideration or significant measurement noise remains a challenging problem. While the Kazantzis-Kravaris-Luenberger (KKL) observer offers a robust framework by immersing nonlinear dynamics into a stable linear filter, reconstructing the physical state from the latent space is often ill-posed. In singular regions, the transformation is non-injective, and standard neural regression yields physically inconsistent mean estimates. In noisy scenarios, exact inversion amplifies measurement errors. n this paper, we propose KKL-CFM, a novel architecture coupling linear KKL dynamics with Conditional Flow Matching (CFM). Instead of learning a deterministic inverse, we approximate the conditional distribution of the state. This generative approach allows for (i) capturing complex multimodal distributions when observability is lost, and (ii) empirical resilience to noise in stochastic environments. Furthermore, we introduce a Clustering-and-Tracking scheme to track the system's distinct physical branches over time.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Peralez et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: