Headline finding
TurboQuant is a data-oblivious, online vector quantization framework that achieves near-optimal distortion rates, within a small constant factor (approximately 2.7) of the information-theoretic lower bound, across all bit-widths and dimensions for both mean-squared error (MSE) and inner product distortion.
The theoretical upper and lower bounds are rigorously proved in Theorems 1–3, and the empirical distortion plots in Figure 3 align with these bounds.
“Our data-oblivious algorithms, suitable for online applications, achieve near-optimal distortion rates (within a small constant factor) across all bit-widths and dimensions.” · Show in paper