Elamur Groupantsprobabilistic-inference
The groupant beside models/probabilistic-inference.groupant in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.
Facts
Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.
Components
| component | label | where |
|---|---|---|
| approximant | partial source model | forward |
| ast | not materialized | |
| composition | partial source model | composition |
| cost | partial source model | cost |
| descriptant | partial source model | epoch-memory |
| invariant | partial source model | invariant |
| tests | bounded executed | wave26-evidence |
Obernants
The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.
| id | type | label | where |
|---|---|---|---|
probabilistic-inference-behavior-complement | behavior_complement | partial relational specification | behavior-complement |
probabilistic-inference-reverse-reconstruction | reverse_reconstruction | partial relational specification | reverse-reconstruction |
probabilistic-inference-max_sum_separation | max_sum_separation | partial relational specification | max_sum_separation |
probabilistic-inference-sampling_mass | sampling_mass | partial relational specification | sampling_mass |
probabilistic-inference-support_convergence | support_convergence | partial relational specification | support_convergence |
probabilistic-inference-weight_normalization | weight_normalization | partial relational specification | weight_normalization |
Model text
model
probabilistic-inference
forward
Tagged weighted observations: alias-table sampling decisions, most-probable sequence decoding, forward/backward posterior mass, Sinkhorn matrix scaling and weighted covariance. Deterministic supplied uniforms and probability normalization are separate from external randomness. HMM w(s)=pi(s0) product_t emission_t(st) product_{t>0}transition(st-1,st); Z=sum_s w(s),V=max_s w(s),0<=V<=Z. Posterior path sums divide by Z when Z>0. Sinkhorn X=diag(u)Kdiag(v),X1=r,X^T1=c. Population covariance=sum_i w_i(x_i-mu)(x_i-mu)^T/sum_i w_i under leaf denominator rules.
invariant
Nonnegative mass, finite values, support, zero-total handling, sequence-state indexes and normalization/status conventions remain leaf-specific. Maximum-path score is not summed evidence; posterior marginal consistency is not joint-path optimality.
cost
Alias preparation is linear and lookup constant within leaf bounds. Viterbi/forward-backward cost sequence length times transition work; Sinkhorn cost iterations times matrix cells plus marginal checks; weighted covariance includes sample/feature accumulation. Costs of producing random draws or building upstream models are outside these kernels.
composition
Alias may sample a normalized posterior row only after row validity and index mapping are checked; a sampled marginal sequence is not a joint HMM sample. Viterbi and forward/backward can share initial/transition/emission data but compare max versus sum semantics, with consistent log/scaled conventions. Sinkhorn outputs can represent couplings only under its support/marginal convergence conditions. Weighted covariance consumes weights, not probabilities necessarily summing to one; normalization must match its population/sample convention.
limitations
Successful allocations and each leaf's finite domain/error precedence. These are tagged relations, not automatic production adoption. Floating residuals are not solution-error guarantees. No generic 1/2/5/10% composition-wide profile; finite evidence remains separate from analytical claims.
epoch-memory
Wave26, 2026-10-30. Tagged library tools; no automatic training adoption, source-volume quality claim or generic numerical profile.
checks
["cargo test -p mova-algorithms alias_method::tests","cargo test -p mova-algorithms viterbi::tests","cargo test -p mova-algorithms forward_backward::tests","cargo test -p mova-algorithms sinkhorn_scaling::tests","cargo test -p mova-algorithms weighted_covariance::tests"]
wave26-executed-checks
probabilistic-inference: wave26 checks
15 source-local test declarations across five modules. Every module has a compiled failing production mutation followed by exact restoration. The final whole library passed 877 checks twice debug and twice release. See integrated evidence for scopes and limitations. These tagged APIs are not an invented serial pipeline.
catalog
Typed family relations
behavior-complement
Type: behavior_complement.
Exclude V>Z for identical nonnegative path weights beyond documented rounding tolerance; posterior normalization at Z=0 contrary to status/error rules; invalid alias variates; exhausted Sinkhorn residuals treated as exact marginals; covariance with a substituted denominator.
reverse-reconstruction
Type: reverse_reconstruction.
Z admits all HMMs with that path-weight sum; Viterbi constrains maximizing path/score but leaves other weights ambiguous. Alias outcomes do not reconstruct random source. Sinkhorn u->a*u,v->v/a preserves X for a>0. Covariance retains moments; many weighted multisets share them.
max_sum_separation
Type: max_sum_separation.
Viterbi optimizes a path; forward/backward sums paths. Their scores generally differ even when the selected path is valid.
sampling_mass
Type: sampling_mass.
Alias thresholds and supplied variates preserve their discrete mass model only under specified endpoints; no random-number source quality is certified.
support_convergence
Type: support_convergence.
Zero support can make desired marginals unattainable. Sinkhorn residual/exhaustion statuses cannot be recast as a successful exact coupling.
weight_normalization
Type: weight_normalization.
Weighted covariance normalization and effective degrees of freedom must match the consumer; uniform rescaling and zero mass need explicit handling.
wave26-evidence
probabilistic-inference: wave26 checks
15 source-local test declarations across five modules. Every module has a compiled failing production mutation followed by exact restoration. The final whole library passed 877 checks twice debug and twice release. See integrated evidence for scopes and limitations. These tagged APIs are not an invented serial pipeline.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/models/probabilistic-inference.groupant/