Seminar: The Jacob Ziv Communication and Information Theory seminar

ECE Women Community

A Gaussian Quantizer Is Not Too Bad

Date: November,19,2026 Start Time: 14:30 - 15:30
Location: 1061, Meyer Building
Add to:
Lecturer: Dr. Anatoly Khina
How well can one compressor perform when only the source mean and variance are known? Universality matters in dynamical settings such as predictive coding and quantized-feedback control, where past quantization errors make the distribution of later signals hard to track.

I will present sharp rate bounds for Gaussian-designed scalar entropy-coded dithered quantization (ECDQ), a variable-length compression scheme. Under quadratic distortion, the maximal rate across all sources with given mean and variance is determined and shown to approach that of a Gaussian source as the distortion tends to zero. The maximal rate is further determined for the classes of log-concave sources and symmetric log-concave sources. Symmetric log-concave coordinate densities arise when TurboQuant randomly rotates normalized vectors for LLM key–value cache compression.

In linear predictive coding of Gaussian sources, the residuals entering the ECDQ are linear combinations of independent uniform and Gaussian variables. The proposed scheme achieves the same distortion at no higher average rate than if these residuals were Gaussian with the same mean and variance. Time permitting, I will present parallel results for fixed-length compression using Gaussian-designed Lloyd–Max (k-means) quantizers for selected scenarios.

Anatoly Khina is a faculty member in the School of Electrical and Computer Engineering at Tel Aviv University, where he earned his B.Sc., M.Sc., and Ph.D. degrees in 2006, 2010, and 2016, respectively. During his studies, he held engineering roles spanning algorithms, data science, software, and hardware. From 2015 to 2018, he was a postdoctoral scholar in the Department of Electrical Engineering at the California Institute of Technology (Caltech). In spring 2018, he was a research fellow at the Simons Institute for the Theory of Computing at the University of California, Berkeley. His research interests include information theory, control theory, signal processing, machine learning, and statistics.

 

All Seminars