Home / Curated / Model Families / Boltzmann chain

Boltzmann chain

Model Families AIMech:BoltzmannChain DRAFT

A family of tied-weight undirected Boltzmann sequence models with hidden temporal variables, observed variables and chain-derived connectivity chosen to permit exact elimination.

Identity Notes & Concept Scope The Saul–Jordan construction includes a strictly linear categorical chain and specified coupled-parallel or looped variants. It denotes an energy parameterization and temporal graph family, not a sampled Monte Carlo chain, an RBM without hidden-hidden connections, or a generic directed factorial HMM. The linear positive-potential case overlaps HMM distributions under the terminal conditions established by MacKay; different parameter names alone do not establish additional distributional expressiveness. More complex source architectures are not automatically covered by that linear equivalence. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 379 preceding native IDs/names/aliases for the four researched candidate names. Both full-checkout fixed-string searches included ignored and hidden files, with 7 and 7 matching lines. All 6 unique matched contents were read: pending leads and preceding source/review context, not existing records of these three new identities. Expanded source IDs, plurals, Gibbs-machine and mixture-transition terminology were included. All 63 semantic label/synonym hits, 17 exact fully annotated AIO CLASS nodes with outgoing edges, their generated YAMLs and six native neighbors were read or byte-verified after prior complete reading. Markov mixtures of experts substantially overlaps the existing IOHMM family, so no fourth ID or unconditional alias is created. Absence is bounded to the pinned vocabulary and inspected neighborhood.

Claims & Grounded Evidence

2 Claims

Mechanism #method

In the linear categorical case, for fixed length L and inverse temperature beta, assign joint probability proportional to exp(beta[Pi_(s_1)+sum_(t=1)^(L-1) A_(s_t,s_(t+1))+sum_(t=1)^L B_(s_t,y_t)]), normalized over all hidden paths and observation sequences by Z_L. A and B are tied across time and need not themselves be log-normalized transition or emission probabilities. Observed likelihood is the ratio of a clamped and a free partition function. A log-likelihood derivative with respect to a tied interaction is beta times its expected count under the clamped distribution minus its expected count under the free distribution. For the selected tractable graphs, exact variable elimination computes these quantities: pruning sums out leaves, decimation sums over an intermediate state, and parallel interactions combine additively in the energy. The paper uses particular coupled chains at different time resolutions and selected loops that preserve an eliminable structure; tractability depends on intermediate state-space size, not merely the presence of temporal edges.

supports Source: Boltzmann Chains and Hidden Markov Models · Locator: Sections 2.1–2.3, Equations (1)–(5), printed pp.436–438 / PDF pp.2–4; Section 3, Figure 2, pp.438–439 / PDF pp.4–5; Sections 4.1–4.2 and Figures 3–5, pp.439–442 / PDF pp.5–8

Scope #linear-hmm-boundary

For the strictly linear finite-energy model, a positive eigenvector of the transfer matrix allows interaction weights to be transformed into normalized HMM parameters. A remaining factor depends only on the final hidden state. Conditioning both models on the same fixed terminal state removes that factor from normalized relative probabilities; more complex architectures are outside this equivalence result.

supports Source: Equivalence of Linear Boltzmann Chains and Hidden Markov Models · Locator: Abstract and Equations (1)–(4), printed/PDF pp.1–2; positive-matrix proof and terminal pseudo-prior discussion, p.2; architectural scope discussion, p.3

⚠️ Curated Limitations & Negative Results

The model is globally normalized at each sequence length. Separately normalizing each energy table is generally not an equivalent model transformation.
MacKay’s positive-matrix gauge transformation produces normalized HMM parameters plus a final-state factor. Equivalence requires the same constrained final state and normalization over that constrained sample space, or explicit handling of the terminal factor. It is not evidence that every unrestricted-terminal HMM and linear energy parameterization agree.
The inspected equivalence proof assumes strictly positive transfer entries from finite energies. Reducible or zero-support cases need separate conditions. It does not cover every coupled or looped graph.
Exact elimination can become exponential in graph width or state size. The source’s chosen chain/loop motifs do not establish efficient exact inference for arbitrary Boltzmann networks.
First-order hidden dynamics can retain long-lag statistical correlations. The source’s wording about absent long-term dependence is not interpreted as zero correlation at all nonadjacent times; extra direct long-range links are a different structural property.
The synthetic learning curves do not certify universal gains, real-data effectiveness or scalability. No author implementation, model fitting or benchmark was executed.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:BoltzmannMachineNetwork` Boltzmann Machine Network	broad_match	The source explicitly defines the temporal model as a tied-weight Boltzmann machine with categorical units and undirected pair interactions. The AIO target describes the broader symmetrically connected stochastic network/Markov-random-field family; the selected temporal graphs are a narrower construction.
`AIO:MarkovChain` Markov Chain	related_match	The strictly linear hidden-chain structure has a qualified relation to HMMs and their latent Markov transitions. Coupled and looped energy graphs, plus terminal normalization conditions, preclude treating the whole family or its observations as the pinned first-order process.

Evidence Sources & Primary Literature

2 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`method-source`	Boltzmann Chains and Hidden Markov Models Lawrence K. Saul, Michael I. Jordan	primary paper full_text	ID: https://papers.neurips.cc/paper_files/paper/1994/hash/4e0cb6fb5fb446d1c92ede2ed8780188-Abstract.html sha256:583d728c7adf…
`linear-equivalence`	Equivalence of Linear Boltzmann Chains and Hidden Markov Models David J. C. MacKay	author technical report full_text	ID: https://www.inference.org.uk/mackay/jordan.pdf sha256:30728362614b…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T20:01:01Z

Updated: 2026-10-11T20:01:50Z

Source File: data/curated/BoltzmannChain.yaml