Asymptotically Optimal Block Quantization
Zador, Gersho, Bucklew–Wise, and task-aware high-rate theory
- Transfer to `OPEN-D-HIGH-RATE`
these papers supply the expected -type machinery, density/shape separation, and a task-aware precedent. What does not transfer: none treats the nonlinear logdet of retained conditional-score Fisher information or the profiled criterion, so the ScoreQuant high-rate claim remains open.
Zador and Gersho establish the classical high-rate scaling and geometric templates for fixed additive distortion; Bucklew & Wise give a rigorous multidimensional -power distortion limit. Gupta & Hero show that an inference objective—binary-detection loss—can admit high-rate vector- quantization calculus and an optimal point density.
- Bucklew–Wise DOI: https://doi.org/10.1109/TIT.1982.1056486
- Gupta–Hero DOI: https://doi.org/10.1109/TIT.2003.814482