State-Estimation of Partially-Observed Markov Chains: Decomposition, Convergence, and Component Identification
Abstract
A partially-observed Markov chain (s,Y) consists of an N-state Markov chain S, along with a process Y of noisy observations of the transitions of S. A metric on stochastic N-vectors and a generalized ergodic coefficient on the transition probability matrices of (S,Y) are defined, resulting in a notion (similar to weak ergodicity) of deteriorating dependence on initial value in a process of distributions of the state (of S) conditioned on past observations. If (S, Y) is stationary, then S may be decomposed into M < or = N components, where M=1 neither implies nor is implied by ergodicity of S, such that conditional state distributions within each component geometrically approach an initial-value-independent process in the manner described above, and one or more equivalent components eventually dominate the others. A method for drift-free finite-memory approximation (or realization) of this process is also introduced.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 15, 1977
- Accession Number
- ADA050344
Entities
People
- Loren K. Platzman
Organizations
- Massachusetts Institute of Technology