Switching Observers for Linear Systems: Beyond Individual Observability

What the math gives to ML

The paper provides a constructive switched-observer mechanism for systems whose individual measurement channels may be non-detectable, while their combined outputs are observable. Its transferable asset is an offline set of channel-specific correction gains together with a dwell-time-dependent stability certificate for the switched estimation-error dynamics. A neural implementation can use this structure for an intermittent or multimodal latent-state model: each available modality applies its own observer correction, while stability is certified over complete switching cycles rather than requiring every modality to stabilize the latent state independently. The key falsifiable prediction is that the latent estimation error decays when the spectral radius of the cycle transition is below one, even when individual channel transitions have spectral radius above one.