spectral_entropy

probly.quantification.spectral_entropy = <flextype.singledispatch.Flexdispatch object>

Compute spectral uncertainty as Von Neumann entropy of an RBF kernel [WGD+26].

Parameters:
  • embeddings – Embeddings with one protected trailing embedding axis. Samples carry their own sample axes; for a sample of embedding samples, the groups and the samples within them are flattened into one set.

  • sample_dim – Batch axis or axes containing samples. Multiple axes are flattened.

  • gamma – RBF bandwidth parameter.

  • normalized – Whether embeddings are L2-normalized.

  • eps – Numerical cutoff for trace and eigenvalues.

Returns:

Spectral entropy over the remaining batch shape.