Entropy of row-wise top-v truncated weights for every truncation level

Computes, in closed form, the mean normalized row entropy that calc_fused_weight_entropy(postprocess_fused_weight(W, top_v = v)) would return for every v from 1 to ncol(W), plus the no-truncation value. Each row is sorted once; the entropy of its top-v renormalized weights is log(S_v) - T_v / S_v with S_v the cumulative weight and T_v the cumulative w * log(w), so the whole curve costs one sort per row.

Usage

fused_entropy_curve(W, keep_ties = TRUE, eps = 1e-12)

Arguments

W

Square numeric weight matrix (rows are samples).

keep_ties

Logical; whether truncation keeps ties at the cutoff, as truncate_top_v_rows() does by default.

eps

Weights at or below this value are treated as zero.

Value

A list with v (integer vector 1:ncol(W)), entropy (mean normalized row entropy at each v), and entropy_inf (no truncation).