For Z = Σ wᵢxᵢ with independent iid pairs and centroid-condition quantizers, NMSE_Z ≈ NMSE_w + NMSE_x + O(1/N⁴). The NMSE does not grow with dimension d, meaning quantization error in large matrix multiplications is controlled.
A matrix multiplication in a neural network is built from dot products, each summing thousands of weight × activation products. A natural worry: does quantization error accumulate as the dot product dimension grows? This corollary shows it does not — the NMSE (normalized mean squared error, the fraction of signal power lost to quantization) of a dot product is independent of its length.
This is a direct consequence of the product error decomposition, which shows how quantization error in individual products combines. It means QF8 (or any format using log-uniform quantization) can be used in arbitrarily large matrix multiplications without degradation.
For a dot product Z = Σᵢ₌₁ᵈ wᵢxᵢ where:
Then:
The NMSE does not grow with dimension d.
Under independence across dimensions, the expected squared error of the dot product is:
and the expected squared signal is:
The dimension d and the scale factor E[w²x²] cancel in the NMSE ratio, leaving only the per-element quantization quality.
Absolute error does accumulate: the MSE of a d-dimensional dot product grows linearly with d. But NMSE — the relative measure — does not, because both the signal and the noise scale equally with d.
This is analogous to the law of large numbers: individual product errors are random and partially cancel in the sum. The NMSE ratio captures this cancellation automatically.
Practical implication: A QF8 matrix multiplication with d = 4096 has the same NMSE as one with d = 32. The +6.6 dB SQNR advantage over FP8 holds at any scale, which is confirmed by benchmarks showing consistent SQNR across matrix sizes from 16×32×16 to 128×256×128.