spectral_decomposition

probly.quantification.decomposition.spectral.torch.spectral_decomposition(embeddings: TorchEmbedding | Tensor, *, group_dim: int, sample_dim: int | tuple[int, ...], gamma: float = 1.0, normalized: bool = True, eps: float = 1e-12) → SpectralDecomposition

Decompose spectral uncertainty over grouped embedding samples.

Parameters:

• embeddings – Embeddings with a trailing protected embedding axis.

• group_dim – Batch axis containing uncertainty groups, such as clarification samples.

• sample_dim – Batch axis or axes containing samples within each group.

• gamma – RBF bandwidth parameter.

• normalized – Whether embeddings are L2-normalized.

• eps – Numerical cutoff for trace and eigenvalues.

Returns:

Additive spectral uncertainty decomposition.