bayesmith.cumulants

CumulantContractions

Defined in bayesmith.cumulants.operators, line 18.

class CumulantContractions(Protocol)

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

CumulantContractions.contract

contract(self, orders: tuple[int, ...]) -> jax.Array

Contract the product of these cumulants with H_(sum orders).

CumulantOperator

Defined in bayesmith.cumulants.operators, line 24.

class CumulantOperator(Protocol)
event_shape: tuple[int, ...]
orders: tuple[int, ...]

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

CumulantOperator.batch_shape

batch_shape(self) -> tuple[int, ...]

CumulantOperator.prepare

prepare(self, value: jax.Array, reference: Any) -> CumulantContractions

Prepare contractions against this same Gaussian reference.

DenseCumulants

Defined in bayesmith.cumulants.operators, line 243.

DenseCumulants(cumulants: Mapping[int, Any], *, event_shape)

Small-field tensors, each shaped batch + (N,)*r in flattened C order.

Computes dense Hermite derivatives up to degree 3*(order-2). Memory can
grow as N**degree; use LowRankCumulants or a custom contraction operator
for maps/cubes with many elements. Tensor symmetrization is implicit in
the contraction with a symmetric Hermite tensor.

DenseCumulants.batch_shape

batch_shape(self)

DenseCumulants.prepare

prepare(self, value, reference)

FieldEdgeworth

Defined in bayesmith.cumulants.field, line 39.

FieldEdgeworth(base: dist.Distribution, cumulants: CumulantOperator | None=None, *, order: int=2, validate_args=None)

Manually truncated joint non-Gaussian likelihood for one array event.

``base`` is a real Gaussian NumPyro distribution whose event_shape is the
map/cube shape. ``cumulants`` supplies contractions in those coordinates.
``order=N`` retains Edgeworth weight <= N-2, including cross-products.
This class knows nothing about grids, spectra or how contractions execute.

As with the scalar EdgeworthExpansion, negative corrections yield NaN and
are never clipped or renormalized. This is an approximation, not a globally
certified density. Non-Gaussian simulation/predictive sampling and automatic
order selection are intentionally not supplied. Order 2 samples ``base``.

FieldEdgeworth.support

support(self)

FieldEdgeworth.correction

correction(self, value)

Signed Gaussian multiplier, one per field event.

FieldEdgeworth.log_prob

log_prob(self, value)

FieldEdgeworth.sample

sample(self, key, sample_shape=())

pytree_data_fields = ('base', 'cumulants')
pytree_aux_fields = ('order',)

LowRankCumulants

Defined in bayesmith.cumulants.operators, line 144.

LowRankCumulants(directions, cumulants: Mapping[int, Any], *, event_shape=None)

Symmetric CP cumulants in the array's own coordinates.

``directions`` has shape ``batch + (rank,) + event_shape``; by default its
first axis is rank and all remaining axes are one field. ``cumulants[r]``
has shape ``batch + (rank,)`` (a scalar broadcasts). Coefficients are raw
cumulants, with no factorial or variance scaling. Missing orders are not
inferred as zero. The same directions may carry different weights at each
order; changing their normalization requires changing those weights.

LowRankCumulants.batch_shape

batch_shape(self)

LowRankCumulants.prepare

prepare(self, value, reference)

PeriodicNormal

Defined in bayesmith.cumulants.fourier, line 22.

PeriodicNormal(loc, power, *, event_ndim=None, validate_args=None)

Real Gaussian field with a periodic covariance diagonalized by FFT.

``power`` contains covariance eigenvalues on the FULL FFT grid (orthonormal
convention), not an rFFT half-grid or a continuum power spectrum. It must
be real, positive, finite and exactly even under simultaneous mode reversal.
``loc`` and power broadcast; the last ``event_ndim`` axes form one field.
The default makes all axes of loc one event. No global settings change.

PeriodicNormal.support

support(self)

PeriodicNormal.log_prob

log_prob(self, value)

PeriodicNormal.sample

sample(self, key, sample_shape=())

PeriodicNormal.mean

mean(self)

PeriodicNormal.variance

variance(self)

pytree_data_fields = ('loc', 'power')
pytree_aux_fields = ('field_ndim',)

low_rank_from_fourier

Defined in bayesmith.cumulants.fourier, line 108.

low_rank_from_fourier(coefficients, cumulants, *, event_shape)

Build real-coordinate CP cumulants from full-grid Fourier directions.

The trailing axes have ``event_shape``; the preceding axis is CP rank.
Inverts an orthonormal full FFT, keeping the coefficients at each cumulant
order unchanged. This transforms every tensor index through its direction
factors. It is NOT a change of the observed random variable or a general
arbitrary-polyspectrum factorization algorithm.

Reality is checked with a 64-epsilon, amplitude-relative numerical allowance
for FFT roundoff (a practical allowance, not a certified FFT error bound).
Only that roundoff imaginary residue is discarded; complex fields and
unpaired complex modes are rejected, including under JIT.