We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

math

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Information Theory

Title: A Class of Nonbinary Symmetric Information Bottleneck Problems

Abstract: We study two dual settings of information processing. Let $ \mathsf{Y} \rightarrow \mathsf{X} \rightarrow \mathsf{W} $ be a Markov chain with fixed joint probability mass function $ \mathsf{P}_{\mathsf{X}\mathsf{Y}} $ and a mutual information constraint on the pair $ (\mathsf{W},\mathsf{X}) $. For the first problem, known as Information Bottleneck, we aim to maximize the mutual information between the random variables $ \mathsf{Y} $ and $ \mathsf{W} $, while for the second problem, termed as Privacy Funnel, our goal is to minimize it. In particular, we analyze the scenario for which $ \mathsf{X} $ is the input, and $ \mathsf{Y} $ is the output of modulo-additive noise channel. We provide analytical characterization of the optimal information rates and the achieving distributions.
Comments: 7 pages, 4 figures, Submitted to the 2022 International Zurich Seminar on Information and Communication
Subjects: Information Theory (cs.IT)
Cite as: arXiv:2110.00985 [cs.IT]
  (or arXiv:2110.00985v1 [cs.IT] for this version)

Submission history

From: Michael Dikshtein [view email]
[v1] Sun, 3 Oct 2021 11:28:23 GMT (35kb)

Link back to: arXiv, form interface, contact.