Replicable Uniformity Testing

Uniformity testing is arguably one of the most fundamental distribution testing problems. Given sample access to an unknown distribution $\mathbf{p}$ on $[n]$, one must decide if $\mathbf{p}$ is uniform or $\varepsilon$-far from uniform (in total variation distance). A long line of work established that uniformity testing has sample complexity $\Theta(\sqrt{n}\varepsilon^{-2})$. However, when the input distribution is neither uniform nor far from uniform, known algorithms may have highly non-replicable behavior. Consequently, if these algorithms are applied in scientific studies, they may lead to contradictory results that erode public trust in science. In this work, we revisit uniformity testing under the framework of algorithmic replicability [STOC '22], requiring the algorithm to be replicable under arbitrary distributions. While replicability typically incurs a $\rho^{-2}$ factor overhead in sample complexity, we obtain a replicable uniformity tester using only $\tilde{O}(\sqrt{n} \varepsilon^{-2} \rho^{-1})$ samples. To our knowledge, this is the first replicable learning algorithm with (nearly) linear dependence on $\rho$. Lastly, we consider a class of ``symmetric"algorithms [FOCS '00] whose outputs are invariant under relabeling of the domain $[n]$, which includes all existing uniformity testers (including ours). For this natural class of algorithms, we prove a nearly matching sample complexity lower bound for replicable uniformity testing.

Paper

Similar papers

Peer review

Reviewer g6cx7/10 · confidence 4/52024-06-22

Summary

In this paper the authors study the problem of replicable uniformity testing. The non-replicable version of the problem can be stated as follows: given some $\varepsilon > 0$ and sample access to some distribution $p$ on $[n]$ what is the minimum number of samples $m$ to distinguish whether $p$ is the uniform distribution or $\varepsilon$ far from it in TV distance? A long line of work has shown that the sample complexity for this task is $\Theta(\sqrt{n}/\varepsilon^2)$. The authors ask for algorithms that have the additional replicability requirement based on the definition of Impagliazzo et al. (2022), meaning that when the algorithm is executed twice on two i.i.d. sets of samples from the same distribution it will make the same decision with probability at least $1-\rho$. The authors design an algorithm that requires $\tilde{\Theta}(\sqrt{n}/(\varepsilon^2\rho) + 1/(\varepsilon^2\rho^2))$ many samples from the distribution $p$. This bound is achieved by considering a (non-replicable) uniformity tester that is based on some $\ell_1$ statistic, since testers based on $\ell_2$ statistics would incur a $tilde{\Theta}(\sqrt{n}/(\varepsilon^2\rho^2))$ sample complexity. Moreover, they provide an (almost) matching lower bound for a natural class of "symmetric" algorithms.

Strengths

I find this work pretty interesting. I think uniformity testing is a natural application domain for the definition of replicability, since the decisions of the algorithm are binary so the discrete metric on the output space (which is what the definition is asking for) is a natural one for this problem. Moreover, the $\sqrt{n}/(\varepsilon^2 \rho)$ sample complexity is interesting and has not appeared in the replicability line of work before. From a technical point of view, the upper bound is achieved by a standard modification of some non-replicable algorithm using a random thresholding trick. However the correctness is based on a concentration argument that requires subtle technical work, dependent on different regimes of the parameters that come into play. Similarly, the lower bound follows the high-level template that has been introduced by prior work. Nevertheless, the technical details are not straightforward. I think another aspect of the results that the authors could consider highlighting is that, in this setting, replicability does not require any blow-up in the sample complexity with respect to the ambient dimension. This is in contrast with other tasks such as replicable mean estimation. The straightforward generalization of the approach of [Impagliazzo, Lei, Pitassi, Sorrell '22] to the $\ell_\infty$ estimation of the means of $n$ coins requires a blow-up of $n^2/\rho^2$ samples. There is a (computationally inefficient) approach by [Karbasi, Velegkas, Yang, Zhou '23] that shaves off a factor of $n$, and was recently shown to be optimal [Hopkins, Impagliazzo, Kane, Liu, Ye '24]. Similar results hold for the $\ell_2$ estimation. Overall, I believe that the paper studies a useful problem, the results are interesting, and the technical contribution is above the bar for NeurIPS.

Weaknesses

I think that the discussion prior to the statement of the main result could be improved by mentioning that there is a $1/(\varepsilon^2\rho^2)$ additive term in the sample complexity (maybe even in the discussion in the abstract). I am bit confused about how the $\tilde{O}(1/(\varepsilon^2\rho^2))$ shows up in the upper bound. The proof sketch that is presented in the main body only considers the $\tilde{O}(\sqrt{n}/(varepsilon^2\rho))$ term. I understand that this is required even for $n=2$, I just don't see how it is used in the general case. Could you please elaborate on it?

Questions

Please see the question in the weaknesses section. Could you also please elaborate on the dependence of the sample complexity with respect to some error parameter? I'm referring to your discussion in footnote 1.

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

Addressed.

Reviewer sPzn8/10 · confidence 4/52024-07-12

Summary

This work studies uniformity testing in the context of replicable algorithms. Knon uniformity testing algorithms are non-replicable in the following sense: a) If the unknown distribution equals the uniform, then they output 1 whp, b) if they are epsilon-away from uniform, output 0 whp c) when the distance is between 0 and epsilon, then they output 1 with arbitrary (could be 1/2) probability. Thus when (c) happens, two different runs of the algorithm may give two different answers, thus are not replicable. The goal of this work is to design algorithms so that even when (c) happens, the algorithms are \rho-replicable (jn the sense of ILPS22). The obvious \rho-replicable algorithms will have a blow-up 1/\rho^2 in the sample complexity (compared to non-replicable algorithms). The main contribution of this work is to show that this can be achieved via only a 1/\rho-blowup in sample complexity. This work also shows that 1/\rho-bloup is necessary if the algorithms have certain property (permutation-invariant).

Strengths

The cost of replicability (in terms of sample complexity) is not well understood. ILPS shows that a blow-up of 1/\rho^2 is needed when the sample space size in 2. I belive this is the only known result. It is natural to expect that a similar lowerbound holds for high-dimensional distributions. However, rather surprisingly, this work shows that this is not necessarily the case. Further elucidating the point hat the tradeoff between sample complexity and replicability needs much further investigation by the community. I really like the contribution of the work.

Weaknesses

I am hoping to uncover a high-level explanation as why the dependency on \rho is inverse-linear when $n$ is high. The additive term has 1/\rho^2 term which is consistent with the lowerbound of ILPS. To me, this suggests that there is an intuitive explanation of the sample complexity, however, I am not able to uncover such an explanation. The overview that the authors provide is more technical than intuitive.

Questions

1. Related to weakness. Can you the high-level intuition?

Rating

8

Confidence

4

Soundness

4

Presentation

3

Contribution

4

Limitations

None

Reviewer MK1R4/10 · confidence 5/52024-07-13

Summary

The concept of "reproducible" learning was introduced in STOC 22 paper and the concept is a very relevant to modern day research. They also showed how algorithms based on statistical estimations can be easily converted to reproducible algorithm with a little overhead. In this paper the authors have tried to produce reproducible testing algorithm for uniformity testing. Uniformity is indeed a very important problem is a lot of applications.

Strengths

NA

Weaknesses

There are many algorithms for uniformity testing. Many of them are either already reproducible or since they are based on statistical estimations they can be easily converted into reproducible algorithms. Neither has the literature of the uniformity testing done thoroughly, nor has it been discussed why this algorithm is new. The same, or similar, algorithm has already been used in the literature.

Questions

NA

Rating

4

Confidence

5

Soundness

1

Presentation

1

Contribution

1

Limitations

NA

Reviewer H8kM5/10 · confidence 3/52024-07-15

Summary

This paper studies uniformity testing under the replicability constraint. Given samples from an unknown distribution P, we need to decide whether P is uniform or eps far from uniform with high probability. Additionally, the algorithm needs to report the same answer in two different random input samples with high probability. The main contribution is an algorithm that takes 1/rho factor more samples than the non-replicable counterpart for which tight sample complexity is well-known. When the sample space size is 2, a 1/rho^2-factor blowup is known to be necessary from prior work. Thus, the new algorithm shaves off a rho factor. Matching sample complexity lower bound is shown only under symmetric assumptions where the algorithm behaves identically under a renaming of the sample space.

Strengths

The main strength of the paper is the improved upper bound. At this point several different approaches for the non-replicable uniformity testing are well-known. The authors manage to show that the empirical TV approach in fact gives the best bounds under replicability constraints.

Weaknesses

The main weakness I believe is the lack of an unconditional tight sample-complexity lower bound.

Questions

None.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

None.

Reviewer sPzn2024-08-09

Thank you for the response.

Reviewer g6cx2024-08-09

Thank you for your response, after reading the rest of the reviews and your rebuttal I remain positive about the paper.

Reviewer MK1R2024-08-13

Thanks for the detailed response.

Reviewer H8kM2024-08-14

Read the rebuttal

Thanks for the reply from the authors. I have not found enough evidence to change my assessment of the paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC