Alternating Expectation Conditional Maximization
Algorithms AIMech:AlternatingExpectationConditionalMaximization DRAFT
Aliases: AECM, AECM Algorithm
A likelihood-optimization framework that alternates cycles of expectation and constrained maximization, permitting the admissible data augmentation and constraint functions to change between cycles while collectively covering the parameter directions.
Identity Notes & Concept Scope
AECM generalizes fixed-augmentation ECM and partitioned SAGE through cycle-specific augmentations and general constraints. One cycle contains an E-step and one or more CM steps sharing that expectation; an iteration completes the stated space-filling coverage. Working parameters index augmentation choices without changing the observed model. This is distinct from PX-EM expansion/reduction and from Gibbs sampling, even when particular examples have matching updates. All 442 pinned AIO labels/scoped synonyms and 466 preceding native IDs, names and aliases were compared. The sole normalized search-term collision, bare SAGE, identifies the existing sparse additive text model; the new optimization method uses qualified aliases. Ignored-inclusive fixed-string searches of both full checkouts returned 436 and 459 line instances. Context windows for all 143 distinct normalized matched texts were inspected; duplicate ledger/review mentions, unrelated environment strings and the stochastic-EM lead do not establish these four method identities. All 225 semantic label/synonym rows and 17 annotated AIO nodes with outgoing edges were inspected; sixteen nodes were byte-verified unchanged after prior reading and UncertaintyBias was newly read. Thirty-one contextual YAMLs comprise 17 AIO and 14 native records, with 27 verified prior readings and four full new readings; the existing sparse text SAGE was also reread. Absence is bounded to this pin, maintained corpus and inspected neighborhood.
Claims & Grounded Evidence
4 Claims
Mechanism
#cycle-specific-augmentation
At the current parameter choose an augmentation that yields the same observed-data model and compute its expected augmented log likelihood conditional on the observed data. Perform the ordered CM steps on that fixed expectation, each on a constraint surface containing the most recent intermediate parameter. Before changing augmentation or starting the next cycle, refresh the expectation at the current parameter. Constraints may be more general than disjoint coordinate blocks.
Mechanism
#cycle-ascent-and-coverage
The fixed expected objective increases within a cycle, so observed likelihood is nondecreasing between completed cycles and consequently completed iterations. This need not hold between every pair of CM substeps within one frozen-expectation cycle. Parameter-space coverage is additional to ascent: the source groups consecutive cycles into iterations when their constraint-gradient spaces have trivial intersection. The stationary-limit result further requires its stated regularity, interior, CM uniqueness or tie-handling, and iteration-mapping conditions.
Scope
#working-parameter-t-example
For a multivariate t model, the source indexes augmentations by q(a)=|Sigma|^(-a)q. With fixed degrees of freedom nu and dimension p, a=1/(nu+p) gives the weighted location and scatter normalized by the sum of latent weights rather than sample count. For unknown nu, location/scatter and nu updates form different cycles, with the design parameter recomputed from the current fixed nu. This family-specific information comparison does not assert optimality over all augmentations or algorithms.
Limitation
#rate-theorem-limits
The general matrix-rate formula assumes convergence to an interior point and a fixed cycle composition with the specified differentiability conditions. It does not imply AECM is always faster than EM. Improving a local asymptotic rate or an individual block step does not guarantee fewer elapsed seconds, and the theorem does not establish convergence at nonnegative parameter boundaries.
⚠️ Curated Limitations & Negative Results
- Monotone bounded likelihood values alone do not establish parameter convergence or global optimality. Incomplete coverage, inexact unconstrained numerical steps or stale expectations after changing augmentation require separate analysis.
- The main paper’s interior convergence results do not automatically cover its Poisson image example at zero intensities; the discussion explicitly refers to separate SAGE analysis. The Poisson pseudobackground example overlaps SAGE3 and is not assigned a duplicate concept ID.
- Working-parameter selection may simplify or complicate the E/CM subproblems. A deterministic optimization result is not a posterior-sampling mixing guarantee.
- All 58 source PDF pages were inspected, including discussions and author corrections. Historical simulations, CPU comparisons, image reconstructions and author software were not reproduced; no universal speed or clinical-performance claim is made.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:MachineLearningTask
Machine Learning Task
|
related_match | This deterministic likelihood-fitting algorithm learns model parameters from data. The inspected AIO target denotes a learning process/task, so this is algorithm-to-task relatedness without an is-a or equivalence assertion. Statistical latent-data augmentation does not identify the pinned transformation-based DataAugmentation term. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Xiao-Li Meng, David van Dyk
|
primary paper
full_text
|
ID: 10.1111/1467-9868.00082
sha256:ddd0a69b5e5d…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-12T03:40:12Z
Updated: 2026-10-12T03:45:56Z