bayesmith.exact.proposals

Source-derived signatures and contracts. See the reference index for coverage and conventions.

Module contract

Small dense Gaussian proposals with optional model-coordinate MH correction.

These are explicitly budgeted proposal builders. They do not replace the
matrix-free GCR path. With MH enabled, approximation affects efficiency, while the original
full graph density and both normalized proposal directions define the target.
Explicitly disabling correction instead applies unadjusted approximate updates,
while retaining support and numerical-validity rejection.

DenseGaussianProposal MHInfo ProposalDiagnostics mh_transition full_support ProposalBlock CompiledProposals compile_proposals

DenseGaussianProposal

Defined in bayesmith.exact.proposals, line 32.

class DenseGaussianProposal(NamedTuple)
mean: jax.Array
precision_cholesky: jax.Array
valid: Any = True

A normalized law with precision ``L @ L.T`` in flat model coordinates.

DenseGaussianProposal.sample

sample(self, key)

DenseGaussianProposal.log_prob

log_prob(self, value)

DenseGaussianProposal.is_valid

is_valid(self)

MHInfo

Defined in bayesmith.exact.proposals, line 62.

class MHInfo(NamedTuple)
attempted: jax.Array
accepted: jax.Array
numerical_failure: jax.Array
support_rejected: jax.Array
log_accept_ratio: jax.Array

No additional source docstring. Read the signature and owning module contract above.

ProposalDiagnostics

Defined in bayesmith.exact.proposals, line 70.

class ProposalDiagnostics(NamedTuple)
attempted: jax.Array
accepted: jax.Array
numerical_failure: jax.Array
support_rejected: jax.Array

One counter per block; ``accepted`` counts applied proposals in either mode.

mh_transition

Defined in bayesmith.exact.proposals, line 79.

mh_transition(key, current, build, log_target, support=None, *, mh_correction=True)

One proposal attempt, with original-target MH enabled by default.

``build`` and ``log_target`` accept flat vectors. A support rejection is
distinct from a numerical failure. No retry or clipping is performed.
With ``mh_correction=False`` no reverse law or MH ratio is evaluated:
a supported draw with finite target density is applied, and the absent
log acceptance ratio is represented by NaN. This is an unadjusted kernel,
not a claim to sample the original target or a known approximate joint.

full_support

Defined in bayesmith.exact.proposals, line 149.

full_support(graph, state)

Check every latent against its distribution in the candidate environment.

ProposalBlock

Defined in bayesmith.exact.proposals, line 204.

class ProposalBlock
graph: Any
policy: Any
names: tuple[str, ...]
unpack: Any
terms: tuple[_Term, ...]

Static flattening metadata and a traceable conditional proposal builder.

ProposalBlock.pack

pack(self, state)

ProposalBlock.state_at

state_at(self, state, flat)

ProposalBlock.build

build(self, state)

ProposalBlock.step

step(self, key, state)

CompiledProposals

Defined in bayesmith.exact.proposals, line 344.

class CompiledProposals
names: tuple[str, ...]
remainder: tuple[str, ...]
blocks: tuple[ProposalBlock, ...]
initial_values: dict[str, Any]

No additional source docstring. Read the signature and owning module contract above.

CompiledProposals.mh_corrected

mh_corrected(self)

Whether every scheduled proposal preserves the original target by MH.

CompiledProposals.sweep

sweep(self, key, full_state)

compile_proposals

Defined in bayesmith.exact.proposals, line 475.

compile_proposals(graph, policies, initial_values)

Validate explicit blocks and dense budgets before tracing transitions.

Block priors may be Normal or Uniform; other factors remain in the target.
Observations used for a proposal must have the existing Gaussian precision
interface. Log proposals additionally accept canonical LogNormal data.
Complex/discrete blocks are left to the existing inference routes.