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

Download:

Current browse context:

math.ST

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

Mathematics > Statistics Theory

Title: Locally private non-asymptotic testing of discrete distributions is faster using interactive mechanisms

Abstract: We find separation rates for testing multinomial or more general discrete distributions under the constraint of local differential privacy. We construct efficient randomized algorithms and test procedures, in both the case where only non-interactive privacy mechanisms are allowed and also in the case where all sequentially interactive privacy mechanisms are allowed. The separation rates are faster in the latter case. We prove general information theoretical bounds that allow us to establish the optimality of our algorithms among all pairs of privacy mechanisms and test procedures, in most usual cases. Considered examples include testing uniform, polynomially and exponentially decreasing distributions.
Comments: 31 pages
Subjects: Statistics Theory (math.ST); Methodology (stat.ME); Machine Learning (stat.ML)
Cite as: arXiv:2005.12601 [math.ST]
  (or arXiv:2005.12601v1 [math.ST] for this version)

Submission history

From: Thomas Berrett [view email]
[v1] Tue, 26 May 2020 09:43:08 GMT (24kb)

Link back to: arXiv, form interface, contact.