PR #6 Empirical Evaluation · PREREG 61c552b + Ext fb54260 · + larger_t.py Sweep

PR #6 Codebooks vs. Scalar Rounding & E8 at Equal Bits per Weight

Status: Complete (Weight Error Only) Predictions: 6/8 Plan + 3/3 Ext Held

Executive Summary: What Held, What Failed, and What the Sweep Found

Testing the three weight-only codebook families proposed in PR #6 against equal-bit scalar mixes (S) and E8 mixes across Gaussian (G), heavy-tailed Student-t4 (T), and 32 real speech-model matrices (Real) yields a sharp split: as proposed, every candidate loses to both scalar and E8 on every family because pyramidal Voronoi cells carry a 0.68–1.69 dB normalized second-moment penalty over cubes. However, an unplanned sweep (larger_t.py) letting the axis offset t move to 3.2–4.0 uncovered an outlier-escape regime that beats scalar rounding in 38/40 synthetic cells (up to −1.72 dB on Student-t) and beats E8 in 8/40 low-bit cells.

Proposed F_n / CaB_n
+0.01 to +3.54 dB
Worse on every family
Free-t Sweep (t ≈ 3.4)
38/40 Beat Scalar
Up to −1.72 dB on T
Vortex V16 / V64
−9.34 / −14.56 dB
Ties scalar; behind E8
L880 (0.70 bits/wt)
−2.42 dB (−2.79w)
Floor is −4.21 dB (closed)
Six Headline Findings & Geometric Diagnosis

Why Packing Fraction (ρ) Does Not Govern Quantization Error (G)

PR #6 cited sphere-packing fractions (ρ = 0.2233 for F3 and φ = 0.746865 for CaB6) as evidence of quantization quality. Sphere packing governs channel noise resistance, whereas fine-cell source coding distortion obeys D ≈ G · V2/n, where G is the dimensionless normalized second moment of the Voronoi cell. The cube achieves G = 1/12 = 0.083333 and E8 achieves 0.071682; pyramidal Voronoi cells in Fn and CaBn have higher second moments, losing 0.68 to 1.69 dB before data distribution effects even enter.

Normalized Second Moment G: Closed-Form vs. Measured (bounds.out §7)

Precision Check: All 14 measured G values sit within 1.7 standard errors of the new closed-form expressions (0.1 and 0.9 SE on 40M-point checks for n = 8, 16). Even placing t at the pyramid centroid rather than the incentre still leaves F8 1.48 dB worse than a cube.

The 6 Empirical Conclusions in Short

1. Proposed F_n & CaB_n lose everywhere: Behind scalar mix by 0.4–1.3 dB (Gaussian) and 0.0–0.8 dB (Student-t), and behind E8 by 0.7–3.5 dB across every rate tested.
2. The cell shape is the culprit: Pyramidal Voronoi cells lose 1.69 dB (F8), 1.21 dB (F16), 1.04 dB (CaB7), and 0.68 dB (CaB15) relative to a cube.
3. Zero-search decoding holds for any 0 < t < 1: Rounding to 2m and picking the largest coordinate remainder |xk − 2mk| is exact for all t ∈ (0, 1)—the incentre plays no special role in it.
4. Free t ∈ [3.2, 4.0] creates an outlier-escape code: Five to seven times the incentre (0.586), large t abandons pyramid cells and lets one coordinate per block jump 1.7 cells past the grid edge, beating scalar in 38/40 synthetic cells.
5. Vortex codebooks match scalar, trail E8: V16 (−9.34 dB) and V64 (−14.56 dB) match optimal scalar quantisers (−9.30, −14.62 dB) but lag E8 (−10.12, −15.55 dB).
6. L880 loses to random points; 2-bit target is closed: L880 trails 880 random Gaussian points by 0.8 dB (0.4 dB whitened); at 0.70 bits/wt, Shannon's floor is −4.21 dB, ruling out matching 2-bit codes (~−9.3 dB) on weight error.

Proof Summary: Why Pyramidal Voronoi Decoding Holds for Any t ∈ (0, 1)

Step 1 · Cell Reduction

Translate x into [-1, 1]ⁿ with x1 ≥ |xj| for all j. Any point of Fn has the form p = 2m + s·t·ek with 0 < t < 1.

Step 2 · Coordinate Inequality

For l ≠ k, (xl − 2ml)² ≥ xl² since |xl| ≤ 1. On coordinate k, the nearest of 2m ± t is sign(xk)·t because midpoints sit at 0 and ±1.

Step 3 · Exact Verification

|x − p|² ≥ |x|² − 2t|xk| + t² ≥ |x − t e1|². Checked against exhaustive search on 200,000 points for n = 2..5 and 75 full codebooks (40 main + 35 sweep): max difference = 0.