Iterative conditional estimation
Algorithms AIMech:IterativeConditionalEstimation DRAFT
Aliases: ICE
An incomplete-data parameter-estimation procedure that repeatedly replaces a chosen complete-data estimator by its conditional expectation given observed data under the current parameter estimate.
Identity Notes & Concept Scope
ICE specifies expectation of an estimator, not maximization of an expected complete-data log likelihood as in EM. The estimator may be maximum likelihood, least squares or another supported complete-data estimator. Shared applications to hidden Markov models do not make ICE a model family, an ICM label optimizer or a synonym for every stochastic EM method. Its behavior depends on the selected estimator and parameterization. The short acronym is scoped to this estimation procedure. No normalized identity match was found across all 442 pinned AIO labels and all scoped synonyms and 394 preceding native IDs, names and aliases for these three concepts. Full-checkout fixed-string searches included ignored and hidden files in worktree and main, yielding 64 and 64 matched lines; all 44 unique contents were read. Hits represent neighboring records, citations and discovery leads rather than records of these identities. All 97 semantic label/synonym hits were read. Twenty-one fully annotated exact CLASS nodes and outgoing edges were verified identical to the preceding inspection, and three additional nodes were read. Twenty-eight contextual YAMLs were byte-verified after prior full reading and four freshly read, totaling 24 generated and eight native contextual records. Short acronyms were included in the complete normalized identity comparison. The absence assessment is bounded to the pin, native collection and inspected semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
Let observed data be y and hidden data h, and choose a complete-data estimator T(h,y) of a parameter theta. From a valid current theta_k, set theta_(k+1)=E_(theta_k)[T(H,y)|Y=y], when this expectation exists and remains admissible. A finite hidden space permits a posterior-weighted sum; otherwise average T over appropriate posterior samples as a Monte Carlo approximation. This average is taken after evaluating the estimator on each completed datum. It generally differs from applying a nonlinear estimator to expected sufficient statistics. With a complete-data maximum-likelihood estimator, ICE averages an argmax, whereas EM takes an argmax after averaging the log likelihood. For a regular full identifiable exponential family with a valid finite interior complete-data MLE and an invertible mean-parameter map, the expectation coordinate psi=E_theta[s(H,Y)] has complete-data estimator s(H,Y), so ICE and EM coincide in that coordinate. Nonlinear reparameterization need not preserve equality of their updates.
⚠️ Curated Limitations & Negative Results
- The complete-data estimator must be defined on posterior-supported completed data and have the required moments. Empty mixture components, boundary or infinite MLEs, constrained parameter spaces and nonidentifiability need explicit conventions; a conditional mean can leave a nonconvex parameter space.
- Conditional expectation is the least-squares approximation to an integrable squared-error target under a fixed law. That fact does not prove decreasing true parameter error, observed-likelihood monotonicity or global convergence across ICE iterations with changing laws.
- The Delmas equivalence uses the mean/expectation parameter psi, although the source calls it canonical; its natural parameter phi is distinct. The claim is not automatic for arbitrary coordinates, curved/constrained exponential families or boundary estimates. A nonlinear transformation does not generally commute with conditional expectation.
- For Gaussian mixtures, expectation of a complete-data class mean or variance containing a random class count differs in general from a ratio of expected sufficient statistics. The usual Gaussian-parameter ICE update therefore cannot simply be replaced by an EM update without checking the conditions.
- Monte Carlo approximation requires valid posterior sampling and appropriate convergence conditions. Correlated Gibbs samples, finite sample counts and apparently steady parameters do not by themselves certify exact expectations or convergence. A one-sample version can resemble stochastic EM for a specific complete-data estimator; it is not every stochastic-approximation EM algorithm.
- The inspected examples in Delmas contain scaling/covariance notation issues, and its introductory suggestion that mixed discrete-continuous data make likelihood irrelevant is not adopted. The definition and qualified coordinate comparison do not certify all printed example formulas. Neither inspected paper provides a universal empirical performance guarantee; no fitting, benchmark or author-code run was performed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:MachineLearningTask
Machine Learning Task
|
related_match | ICE is a general parameter-learning procedure for incomplete-data statistical models. MachineLearningTask supplies a broad learning context but does not identify conditional-estimator iteration or distinguish it from EM. A related mapping records that conceptual connection without classifying the algorithm as a task or assuming a supervised/unsupervised subtype. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Wojciech Pieczynski
|
primary paper
full_text
|
sha256:3fd1eb9cfae0…
|
em-comparison
|
Jean Pierre Delmas
|
primary paper
full_text
|
sha256:065593e02a66…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T21:20:23Z
Updated: 2026-10-11T21:20:23Z
Source File:
data/curated/IterativeConditionalEstimation.yaml