Multi-Rate Distributed Unscented Kalman Filtering Under Collective Detectability

What the math gives to ML

The paper provides a transferable information-form mechanism for decentralized estimation when no individual node is observable but the network is collectively detectable over a time window. Each node diffuses information matrices and information vectors, rather than averaging means or covariances; this preserves a uniformly positive network information margin and avoids requiring a contraction-mixing condition for the fused mean. The most promising neural-network transfer is asynchronous multi-agent or multi-view latent-state inference: let each encoder maintain a Gaussian latent posterior, perform local likelihood updates at its own rate, and diffuse natural parameters over the communication graph. The paper also supplies a measurable design signature: bounded quotient covariance should appear once the windowed collective information Gramian exceeds a positive threshold, while arithmetic covariance or mean fusion can become unstable when individual agents miss latent modes.