uniform_fitting#

Uniform-layer fitting: the packed SamplingKind classes and the least-squares Problem.

The uniform twins of t3toolbox.backend.fitting’s kinds – subclasses that override the five layer-specific operations and, for the probe kinds, where omega’s axes sit in the PACKED output. Each carries the fixed rank it was built at (shape + the plain-UT3 masks) as FIELDS, so a rebuilt kind of the same rank is the same jax cache key (t3toolbox.backend.geometry follows the same rule; the reasoning is docs/contributor/parameters_not_closures.md).

uniform_least_squares_problem() packs the loop-invariant sample + data ONCE and returns the shared backend Problem, so the optimizers run fully packed – no per-matvec pack/unpack. The geometry half lives in t3toolbox.backend.geometry.

Protocol note (review H2-9, ruled keep + document): the kinds’ forward / transpose keep the ragged Kind protocol’s (v, sample, frame_data, sweep) signature even though the packed sweep already carries the sample and the frame – the extra arguments are deliberately unread. Protocol uniformity is the point: generic code (the shared LocalModel) calls every kind the same way, ragged or uniform.

Classes#

UniformApplyKind

The uniform twin of APPLY.

UniformEntriesKind

The uniform twin of ENTRIES.

UniformProbeKind

The uniform twin of ProbeKind.

UniformApplyDerivativesKind

The uniform twin of ApplyDerivativesKind.

UniformEntriesDerivativesKind

The uniform twin of EntriesDerivativesKind.

UniformProbeDerivativesKind

The uniform twin of ProbeDerivativesKind.

Functions#

uniform_sampling_kind(name, x0_data[, weight])

Build the uniform plain sampling kind by name, at x0's fixed rank. Only the vector-valued

uniform_derivatives_kind(name, x0_data, order[, ...])

Build the uniform derivative sampling kind by name, at x0's fixed rank.

pack_sample(name, sample, N)

Pack the loop-invariant mode-vectors of sample once (mirror-tolerant: packed input is kept).

pack_data(name, data, N)

Pack the observed data (or a residual of the same shape) once: probe kinds -> a packed

uniform_minimal(x0[, sharing])

Reduce x0 to its structurally-minimal ranks -- the SAME tensor, with any unrealizable nominal

uniform_least_squares_problem(geometry, kind_name, x0, ...)

Assemble a fully-packed uniform least-squares Problem.