EN
Back to the archive

The encyclopedia · R&D & Science · Technical decision · 1977

Dempster-Laird-Rubin fit models with hidden data by alternating guess and maximize

Dempster, Laird and Rubin's 1977 EM algorithm fit maximum-likelihood models when data is missing by alternating between estimating hidden values and maximizing.

Harvard University

the move

Many models have variables that can't be observed — which cluster a point belongs to, the true value behind a censored measurement. Maximum likelihood is impossible when the data is incomplete.

The 1977 paper by Dempster, Laird and Rubin gave the field an algorithm: iteratively fill in the expected values of the hidden data (E-step), then maximize the parameters assuming those values are real (M-step). The likelihood is provably non-decreasing each round.

That guarantee is the point: EM needs only the current estimate, never the hidden truth, and is guaranteed to improve until it stops. It turned a hard, often unsolvable estimation problem into a loop anyone can run.

why it works

  • Filling in hidden values lets maximization proceed even with missing data.
  • The likelihood never falls, so the loop always makes progress.
  • It needs no gradient over the latent variables, only a tractable conditional expectation.
  • It handles clustering, mixture models, and imputation with one machinery.
the payoffAlternate: fill in hidden data, then maximize; repeatinspired

what transfers

When unobserved data blocks estimation, don't discard it. Repeatedly impute the hidden part from a current guess and re-estimate, since each round only improves fit.

what came after

EM became a standard tool in statistics and machine learning, behind Gaussian-mixture clustering, latent variable models, hidden Markov model training and recommendation systems. Its guarantee and simple loop make it the default for incomplete-data maximum likelihood.

references

spotted an error? The archive wants to know.

same kind of clever