Module lloyd 

Lloyd-Max optimal scalar quantizer tables for N(0,1), precomputed offline.

Centroids and decision boundaries are computed via Lloyd’s algorithm (2000 iterations, convergence tolerance 1e-12) on the standard normal distribution, then embedded as constants. No runtime training pass.

4-bit: 16 centroids / 15 boundaries. 2-bit: 4 centroids / 3 boundaries.

Constants

BOUNDARIES_4BIT

4-bit decision boundaries (index i separates centroid i-1 and i).

BOUNDARIES_5BIT

5-bit decision boundaries.

BOUNDARIES_6BIT

6-bit decision boundaries.

CENTROIDS_4BIT

4-bit Lloyd-Max centroids for N(0,1).

CENTROIDS_5BIT

5-bit Lloyd-Max centroids for N(0,1).

CENTROIDS_6BIT

6-bit Lloyd-Max centroids for N(0,1).

Functions

boundaries

Decision boundaries for a bit width.

centroids

Codebook centroids for a bit width (alias of the per-width constants).

dequantize

Dequantize a code at the given bit width (direct centroid index).

dequantize_4bit

Dequantize a 4-bit index back to its centroid value.

levels

Number of codebook levels for a supported bit width (4 -> 16, 5 -> 32, 6 -> 64).

quantize

Quantize one N(0,1) value at the given bit width via binary search.

quantize_4bit

Quantize one N(0,1)-distributed value to a 4-bit index via binary search over the decision boundaries.