PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources
The paper presents PrismQuant, a method for rate-distortion-optimal vector quantization tailored for Gaussian-mixture sources. It addresses the challenges posed by multimodal sources and introduces a new approach to bit allocation across heterogeneous branches. Experimental results indicate that PrismQuant performs competitively against existing methods while maintaining a smaller model size.
- ▪PrismQuant is designed for Gaussian-mixture sources and aims to optimize rate-distortion performance.
- ▪The method utilizes a global reverse-waterfilling level for efficient bit allocation across different components.
- ▪Experiments show that PrismQuant closely approaches theoretical rate-distortion bounds and outperforms transformer-based codecs.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Information Theory arXiv:2605.15507 (cs) [Submitted on 15 May 2026] Title:PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources Authors:Bumsu Park, Chanho Park, Youngmok Park, Namyoon Lee View a PDF of the paper titled PrismQuant: Rate-Distortion-Optimal Vector Quantization for Gaussian-Mixture Sources, by Bumsu Park and 3 other authors View PDF HTML (experimental) Abstract:For a Gaussian source under mean-squared error (MSE), classical transform coding is rate--distortion (RD) optimal: the Karhunen--Loeve transform (KLT) diagonalizes the covariance, reverse waterfilling allocates the bits, and scalar quantization closes the loop.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
