torchsky.map.likelihood¶

Bases: _GuardedEvidence

Fixed-model sampling variance, mean transfer and solve diagnostics.

For ridge strength d, mean_response is H (H+d Id)^-1 g. It equals g for an accurately solved unregularized estimable query. Otherwise its difference from g describes possible bias and is not a confidence bound. This result supplies no conformal or misspecification coverage guarantee. Public numeric reads validate the factor evidence and query direction; rebuild after changing either source.

_evidence_guard: Callable[[], None] | None = field(
    default=None, repr=False, compare=False
)

_guarded_fields = frozenset(
    {
        "sampling_variance",
        "mean_response",
        "solve_residual_norm",
        "damping",
    }
)

damping: float

mean_response: Tensor

nuisance_treatments: tuple[NuisanceTreatment, ...]

sampling_variance: Tensor

solve_residual_norm: Tensor

One fixed Gaussian quadratic, with centered native data and lineage.

measurement_ids identify disjoint native sample groups, including the constituent groups of a derived product. A mosaic and its inputs therefore share identifiers. Callers own the truthful lineage and sampling partition. eliminated_nuisance_ids identify unknowns profiled or integrated in this factor; common unknowns must be eliminated jointly, not separately twice.

Supply data centered by the declared calibrated additive mean. Precision callbacks own observed-support and real/complex noise conventions. This is a quadratic likelihood kernel. For a conditional or calibrated Gaussian density, supply its scalar log_normalization (e.g. half of n log(2 pi) + logdet(C) in real coordinates). Otherwise absolute density evaluation fails closed. Profiled nuisance objectives are not normalized predictive densities. Rebuild after changing response, calibration, noise, or data; tensors are not copied.

Data and normalization versions are tracked automatically. Supply an EvidenceSnapshot in dependencies for callback matrices, module parameters, noise, calibration and support, plus a revision getter for mutable non-tensor state or replaced tensor bindings. Callback closures are not inspected. Inference tensors require an explicit revision; callers own revision changes for writes that bypass Torch mutation versions.

_operator_binding: tuple[int, ...] = field(
    init=False, repr=False, compare=False
)

_snapshot: EvidenceSnapshot = field(
    init=False, repr=False, compare=False
)

calibration_id: str

data: Tensor

dependencies: EvidenceSnapshot | None = None

eliminated_nuisance_ids: tuple[str, ...] = ()

log_normalization: Tensor | None = None

measurement_ids: tuple[str, ...]

nuisance_treatment: NuisanceTreatment = 'conditional'

operator: BoundObservationOperator

provenance_id: str

__post_init__() -> None

_binding() -> tuple[int, ...]

precision(value: Tensor) -> Tensor

residual(state: Tensor) -> Tensor

validate_evidence() -> None

Independent fixed factors over a common scene, without a field solve.

Operators may emit native pixels, vectors, timelines or complex arrays. Nonlinear/Poisson likelihoods cannot use this quadratic compression. The system requires independent factors under their declared nuisance treatment; unrecorded correlation cannot be detected from metadata alone.

_operators: BoundObservationOperatorSet = field(
    init=False, repr=False
)

exposure_chunk_size: int = 1

factors: tuple[GaussianLikelihoodFactor, ...]

measurement_ids: tuple[str, ...]

__post_init__() -> None

functional_uncertainty(
    functional: Tensor,
    *,
    iterations: int = 64,
    damping: float = 0.0,
    rtol: float = 1e-05,
    atol: float = 0.0,
    preconditioner: Callable[[Tensor], Tensor]
    | Tensor
    | None = None,
    dependencies: EvidenceSnapshot | None = None,
) -> FunctionalSamplingUncertainty

information(state: Tensor) -> Tensor

negative_log_likelihood(state: Tensor) -> Tensor

normal_equations() -> GaussianNormalEquations

quadratic(state: Tensor) -> Tensor

rhs() -> Tensor

solve(
    *,
    iterations: int = 64,
    damping: float = 0.0,
    preconditioner: Callable[[Tensor], Tensor]
    | Tensor
    | None = None,
) -> LinearSolveResult

validate_evidence() -> None

Bases: _GuardedEvidence

Full quadratic evidence x^T H x - 2 x^T b + c for fixed callbacks.

The full information action is retained. No diagonal approximation or reconstructed image is presented as sufficient evidence. c excludes determinant terms; optional log_normalization must be supplied by the caller from the declared covariance and observed support.

Statistics returned by a likelihood system validate its declared sources on public reads and information calls. A previously retrieved tensor is an ordinary tensor; its later use cannot be intercepted.

_evidence_guard: Callable[[], None] | None = field(
    default=None, repr=False, compare=False
)

_guarded_fields = frozenset(
    {
        "rhs",
        "data_quadratic",
        "information",
        "log_normalization",
    }
)

data_quadratic: Tensor

information: Callable[[Tensor], Tensor]

log_normalization: Tensor | None = None

measurement_ids: tuple[str, ...]

rhs: Tensor

__post_init__() -> None

negative_log_likelihood(state: Tensor) -> Tensor

quadratic(state: Tensor) -> Tensor