kernel_entropy.entropy

Kernel Language Entropy calculation.

Transforms similarity matrix W into Von Neumann Entropy through: W -> Laplacian -> Heat Kernel -> Density Matrix -> VNE

Functions

kle_from_similarity(W[, t])	Compute Kernel Language Entropy from similarity matrix.
kle_to_certainty(entropy, n_samples)	Map KLE (nats) to a [0, 1] certainty.

kernel_entropy.entropy.kle_from_similarity(W: Tensor, t: float = 1.0) → float

Compute Kernel Language Entropy from similarity matrix.

Parameters:

• W – N×N symmetric similarity matrix on CUDA (from NLI scoring)

• t – Heat kernel lengthscale (default: 1.0)

Returns:

Von Neumann Entropy (float)

kernel_entropy.entropy.kle_to_certainty(entropy: float, n_samples: int) → float | None

Map KLE (nats) to a [0, 1] certainty.

Returns None when n_samples < 2 (insufficient samples to define entropy). Otherwise returns 1 - entropy / log(n_samples), clamped to [0, 1] to absorb numerical noise near the log(N) ceiling. See the response aggregation doc for the log(N) upper bound.