A stochastic model in which a discrete-time latent Markov chain {S_t} evolves with specified transition probabilities and each hidden state emits observable outputs according to a state-dependent probability distribution; inference and learning aim to estimate the posterior distribution over hidden states given observations and/or to fit transition and emission parameters from data.

Principle

Conditional independence in time: given the current hidden state S_t, the observation at time t is independent of past and future observations and states except through S_t; the hidden states themselves satisfy the Markov property (P(S_t | S_{t-1},...) = P(S_t | S_{t-1})).

Demonstration

Illustrative scenario → A sequence of acoustic feature vectors is modeled as emissions from an unobserved sequence of phoneme states; recognition proceeds by computing P(S_{1:T} | observations) via the forward–backward (sum–product) algorithm or Viterbi decoding for the most probable state path; parameter learning can use the Baum–Welch (expectation–maximization) procedure to adjust transition and emission parameters from labeled or unlabeled sequences.

Misapplication

Treating the observed sequence as Markov and directly using transition counts on observations to infer latent structure; this confuses observable transitions with latent-state transitions and can produce biased state estimates when emissions are noisy or many-to-one.

Consequence

When correctly specified, an HMM provides a principled way to infer hidden state trajectories, segment sequences, and estimate generative parameters; when mis-specified (wrong state space, inappropriate emission family, or violated Markov assumption) inference and learned parameters can be systematically biased and predictive performance degraded.

Reversal

The usual HMM assumptions fail when latent dynamics are non-Markovian, when state durations have non-geometric distributions (requiring Hidden semi-Markov models), or when latent variables are continuous (favoring Kalman/linear dynamical models) — in these cases, alternative model classes or augmented state representations are required.

Boundary

Clearly within: discrete-time finite-state latent chain with emissions per time step. Boundary case: models with explicit duration modeling (HSMM) or state-dependent autoregressive emissions. Clearly outside: models with continuous latent variables governed by linear Gaussian dynamics (Kalman filters) or models that do not posit a latent state generating observations.

Semantic Tension

Expressiveness (richer state spaces or emission models) versus tractability (exact inference requires low treewidth or specialized algorithms); adding flexibility often raises computational or identifiability challenges.

Synthesis

An HMM separates qualitative temporal structure (a Markovian latent process) from quantitative observation models (emissions); this decomposition enables tractable inference algorithms when the Markov and emission assumptions are appropriate, but those same assumptions determine both the model's applicability and its limitations.