Fisher transforms

FisherTransform dataclass

FisherTransform(
    matrix: ndarray,
    eigenvectors: ndarray,
    eigenvalues: ndarray,
    retained_eigenvalues: ndarray,
    rank_rtol: float,
    threshold: float,
    whiten: bool,
)

Projection onto informative Fisher directions, optionally whitened.

Scores are never mean-centered. matrix maps raw score vectors from shape [..., P] to optimization coordinates [..., R].

Attributes:

• matrix (ndarray) –

  Projection or whitening matrix with shape [P, R].

• eigenvectors, eigenvalues –

  Complete eigendecomposition of the symmetrized input Fisher matrix.

• retained_eigenvalues (ndarray) –

  Eigenvalues above the numerical rank threshold.

• rank_rtol, threshold –

  Relative and absolute rank thresholds.

• whiten (bool) –

  Whether retained directions are scaled by inverse square root eigenvalues.

input_dim property

input_dim: int

Return the raw score dimension P.

rank property

rank: int

Return the retained informative rank R.

dropped_directions property

dropped_directions: int

Return the number of projected-out score directions.

apply

apply(
    scores: ArrayLike,
    *,
    execution: ExecutionConfig | None = None,
) -> ndarray

Map raw scores into the fitted informative coordinate system.

to_dict

to_dict() -> dict[str, JsonValue]

Return a JSON-compatible representation.