The expectation-maximization (EM) algorithm is an iterative method for finding maximum-likelihood (or maximum a posteriori) estimates of parameters θ in statistical models that depend on unobserved latent variables z. At each iteration t it alternates between: (1) an E-step — computing Q(θ | θ^(t)) = E_{z|x,θ^(t)}[log p(x, z | θ)], the expected complete-data log-likelihood given the observed data x and current parameters; and (2) an M-step — updating θ^(t+1) = argmax_θ Q(θ | θ^(t)), maximizing this expectation. Its parameters are the observed data x, the latent-variable structure z, the complete-data likelihood p(x, z | θ), and the convergence criterion (change in log-likelihood below threshold, or maximum iterations). It persists as a standard algorithm in mixture-model fitting (Gaussian mixture models), hidden Markov models (Baum-Welch), missing-data imputation, and topic models (LDA), taught in every statistical-learning curriculum and implemented in all major libraries, so that iterative procedures for latent-variable models converge to stable fixed points. [formal: expectatio-maximatio | substrate: mind | horizon: a life | explicit: yes | epoch: 0.01]
Accepted ontology entry
expectation-maximization
The expectation-maximization (EM) algorithm is an iterative method for finding maximum-likelihood (or maximum a posteriori) estimates of parameters θ in statistical models that depend on unobserved latent variables z. At each iteration t i…
Definition
Why it is in scope
A human-made iterative optimization algorithm designed to find maximum-likelihood estimates in statistical models with latent variables, built to persist as a standard procedure in statistics, machine learning, and computational linguistics.
Names and aliases
- expectation-maximizationen · CANONICAL
Relations from this entry
- likelihoodDEPENDS_ON →
EM's E-step computes Q(θ|θ^(t)) = E[log p(x,z|θ)] — the expected complete-data log-likelihood. The M-step maximizes this over θ. The entire iterative procedure is a likelihood maximization algorithm for models with latent variables. Remove the likelihood function and EM has no mathematical object to optimize; it stops operating entirely. Constitutive present-tense dependency (Law 8b/8c).
- maximum-likelihood-estimationSERVES →
Law 8d 'for whose sake?': EM is built to find maximum-likelihood (and MAP) parameter estimates in models with unobserved latent variables — that estimation task is its designed purpose, and its two steps (E-step posterior, M-step maximization of the expected complete-data log-likelihood) exist only to produce the MLE. Servant (the iterative algorithm) points at master (the estimation method it was built for). Service need not be exclusive (it also serves MAP estimation); the pinned sense is MLE as the primary designed target.
- cmsm0cr9y00331q137lslor2hDEPENDS_ON →
EM operates on a statistical model with latent variables: the E-step computes posterior over latent variables given observed data and current parameters under the model's complete-data likelihood, and the M-step maximizes expected complete-data log-likelihood with respect to model parameters. Remove the statistical model and EM has no data-generating family, no parameter space, no latent structure to optimize — the iterative procedure stops operating entirely.
Relations to this entry
No accepted relations in this direction.
Record identity
- Created
- Sep 3, 2026, 1:32 PM UTC
- Content hash
- b4a1787618f37d05e44d8b682aeda40b2bba28962e5d7bc31feffc948d9d99dc