### Current browse context:

stat.ML

### Change to browse by:

### References & Citations

# Statistics > Machine Learning

# Title: Distribution Compression in Near-linear Time

(Submitted on 15 Nov 2021 (v1), last revised 18 Oct 2022 (this version, v6))

Abstract: In distribution compression, one aims to accurately summarize a probability distribution $\mathbb{P}$ using a small number of representative points. Near-optimal thinning procedures achieve this goal by sampling $n$ points from a Markov chain and identifying $\sqrt{n}$ points with $\widetilde{\mathcal{O}}(1/\sqrt{n})$ discrepancy to $\mathbb{P}$. Unfortunately, these algorithms suffer from quadratic or super-quadratic runtime in the sample size $n$. To address this deficiency, we introduce Compress++, a simple meta-procedure for speeding up any thinning algorithm while suffering at most a factor of $4$ in error. When combined with the quadratic-time kernel halving and kernel thinning algorithms of Dwivedi and Mackey (2021), Compress++ delivers $\sqrt{n}$ points with $\mathcal{O}(\sqrt{\log n/n})$ integration error and better-than-Monte-Carlo maximum mean discrepancy in $\mathcal{O}(n \log^3 n)$ time and $\mathcal{O}( \sqrt{n} \log^2 n )$ space. Moreover, Compress++ enjoys the same near-linear runtime given any quadratic-time input and reduces the runtime of super-quadratic algorithms by a square-root factor. In our benchmarks with high-dimensional Monte Carlo samples and Markov chains targeting challenging differential equation posteriors, Compress++ matches or nearly matches the accuracy of its input algorithm in orders of magnitude less time.

## Submission history

From: Lester Mackey [view email]**[v1]**Mon, 15 Nov 2021 17:42:57 GMT (730kb,D)

**[v2]**Wed, 17 Nov 2021 01:49:21 GMT (734kb,D)

**[v3]**Thu, 24 Mar 2022 22:46:34 GMT (792kb,D)

**[v4]**Tue, 14 Jun 2022 12:36:23 GMT (789kb,D)

**[v5]**Tue, 13 Sep 2022 17:57:45 GMT (789kb,D)

**[v6]**Tue, 18 Oct 2022 01:29:37 GMT (789kb,D)

Link back to: arXiv, form interface, contact.