PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources
A recent preprint on arXiv (2605.15507) presents PrismQuant, a framework for vector quantization that optimally addresses rate-distortion for Gaussian-mixture sources. Traditional transform coding, which excels with unimodal Gaussian sources when using MSE, falls short for multimodal sources, as a single covariance fails to represent diverse local geometries. The authors demonstrate that the mixture structure only incurs a cost related to component labeling, while the genie-aided conditional RD function is determined by a unified global reverse-waterfilling level applicable to all components and eigenmodes. Building on this finding, PrismQuant achieves optimal quantization.
Key facts
- arXiv preprint 2605.15507
- Title: PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources
- Classical transform coding is RD-optimal for Gaussian sources under MSE
- KLT diagonalizes covariance, reverse waterfilling allocates bits
- Multimodal sources break the classical story
- Gaussian-mixture sources are revisited
- Mixture structure incurs only a component label cost
- Genie-aided conditional RD function uses a single global reverse-waterfilling level
Entities
Institutions
- arXiv