Robust Mean Estimation Without Moments for Symmetric Distributions

We study the problem of robustly estimating the mean or location parameter without moment assumptions. We show that for a large class of symmetric distributions, the same error as in the Gaussian setting can be achieved efficiently. The distributions we study include products of arbitrary symmetric one-dimensional distributions, such as product Cauchy distributions, as well as elliptical distributions. For product distributions and elliptical distributions with known scatter (covariance) matrix, we show that given an $\varepsilon$-corrupted sample, we can with probability at least $1-\delta$ estimate its location up to error $O(\varepsilon \sqrt{\log(1/\varepsilon)})$ using $\tfrac{d\log(d) + \log(1/\delta)}{\varepsilon^2 \log(1/\varepsilon)}$ samples. This result matches the best-known guarantees for the Gaussian distribution and known SQ lower bounds (up to the $\log(d)$ factor). For elliptical distributions with unknown scatter (covariance) matrix, we propose a sequence of efficient algorithms that approaches this optimal error. Specifically, for every $k \in \mathbb{N}$, we design an estimator using time and samples $\tilde{O}({d^k})$ achieving error $O(\varepsilon^{1-\frac{1}{2k}})$. This matches the error and running time guarantees when assuming certifiably bounded moments of order up to $k$. For unknown covariance, such error bounds of $o(\sqrt{\varepsilon})$ are not even known for (general) sub-Gaussian distributions. Our algorithms are based on a generalization of the well-known filtering technique. We show how this machinery can be combined with Huber-loss-based techniques to work with projections of the noise that behave more nicely than the initial noise. Moreover, we show how SoS proofs can be used to obtain algorithmic guarantees even for distributions without a first moment. We believe that this approach may find other applications in future works.

Paper

Similar papers

Peer review

Reviewer DMaa5/10 · confidence 3/52023-06-19

Summary

This paper studies outlier-robust location estimation for symmetric(-like) distributions. For "semi-product" distributions, the authors show that one can achieve $O(\epsilon \sqrt{\log 1/\epsilon})$ asymptotic error with a polynomial number of samples (and time), and $O(\epsilon)$ asymptotic error when given quasipolynomially many samples (and time poly in the sample set size). For elliptical distributions, the authors show that as long as the scatter matrix has $\Omega(\log d)$ effective rank (hiding $\epsilon$-dependence here), then in quasipolynomial time we can yield asymptotic error $O(\epsilon \sqrt{\log 1/\epsilon})$. Crucially, for the last result, the scatter matrix is not known to the algorithm.

Strengths

Prior works in (algorithmic) robust statistics get "stuck" at the $\sqrt{\epsilon}$ asymptotic error, when the covariance of the distribution is not known, even when the underlying distribution is guaranteed to have higher moments or might even be sub-Gaussian. This paper identifies semi-product distributions and elliptical distributions as special classes for which further progress can be made. The author(s) adapts the filtering framework, and proposes using variants of the Huber loss as score as opposed to the basic quadratic score in the filtering step. This allows the author(s) to beat the $\sqrt{\epsilon}$ error even when covariance/scatter matrices are unknown to the algorithm.

Weaknesses

I find some parts of the writing clarity can be improved. In particular, I'm still a bit confused about how the guarantees of this work compare with prior works, re: knowledge assumptions on the algorithms. Some of the technical claims are also a bit over-sold (unless I misread or am misunderstanding). See the "Questions" section for more details. Another weakness, for me, is the motivation: the practical relevance of semi-product and elliptical distributions seems to be rooted in mathematical finance. However, at least from the cited works, the relevance of these distributions appears to more or less be "because we can write down theorems". I hope the authors will consider adding a more self-contained discussion on why we should care about these distributions; I think it will make the paper more convincing.

Questions

More important: - Knowledge assumption on algorithm in Theorems 1.4 -- 1.6: one of the selling points from the introduction is that the algorithm doesn't need to know the covariance or scatter matrix. However, it was unclear to me whether this means the algorithm knows absolutely nothing about the scatter, or if it only knows very little (such as an upper bound). The theorems can be a lot more explicit in saying exactly what the algorithm needs to know (or explicitly say that it needs to know absolutely nothing about the scatter matrix). For example, does the algorithm in Theorem 1.4 need to know the parameter $\rho$, which is essentially an upper bound to the covariance, at least in the Gaussian case? This distinction seems pretty important to me. - Result statement of Theorems 1.4 -- 1.5: the statements are of the form "if $n$ is sufficiently large, then we get this almost-sub-Gaussian error rate plus the claimed asymptotic error". However, it seems that the required $n$ lower bound is large enough that the almost-sub-Gaussian error is already dominated by the asymptotic error. Yet the prose following the theorems seem to sell how the theorem gets close to the sub-Gaussian rate (for Theorem 1.4, and for constant $\delta$ and large $C$ for Theorem 1.5). Am I misunderstanding something? - Theorem 1.6: Line 170 claims that Theorem 1.6 is only slightly sub-optimal in the no-corruption part of the error. But that error is multiplicative in terms of the effective rank and $\log d/\delta$, which is much worse than the sub-Gaussian rate of *additive* error. This error is equal to the naive analysis of entry-wise median-of-means. In fact, entry-wise median-of-means does not actually need the $d$ in $\log d/\delta$ from the naive union bound, and can get error $\sqrt{\mathrm{Tr}(\Sigma)\log\frac{1}{\delta}/n}$, so the error in Theorem 1.6 is even worse than entry-wise median-of-means. Am I missing the point here? - It is unclear to me that even the sub-Gaussian rate is the correct error rate for the non-robust part of these theorems. For 1-d symmetric distributions, and without corruptions, it was shown by Stone (1975, Maximum Likelihood Estimators of a Location Parameter) that asymptotically one can achieve Fisher information rates, which can be much better than the variance-based sub-Gaussian rate. - The appendix was hard to follow. There are three appendices with filtering algorithms, but Appendix I was titled only "Filtering Algorithm". It also seemed a bit weird that Appendices H and I appeared last, even though they prove Theorem 1.4, which was introduced earlier than Theorems 1.5 and 1.6. In terms of the actual content, there doesn't seem to be much discussion on how the techniques in the appendices relate to each other. I could see that Appendix E generalizes Appendix D, but how does Appendix I (and relatedly, Appendix H) relate to Appendix E (and C)? Less important: - Notation is slightly confusing, with the bold and capital letters. There is no distinction between random scalars and random vectors (e.g. in defining elliptical distributions). - Page 3, it was claimed that the first 2 properties of semi-product distributions allow for accurate non-robust (i.i.d.) location estimation by using entry-wise median. This seems true only when $\rho$ is small? - Footnote 1, does $\alpha$ need to be known by the algorithm? - Footnote 2, there is an $O(\log 1/e)$ which I assume is a typo? Is "$e$" supposed to be "$\epsilon$"? - CTBJ22 reference: they show how to do mean estimation robustly when only $1+\alpha$ moments exist, but (1) they do not handle the full adversarial corruption model because of sample-splitting issues and (2) they require an upper bound to the $1+\alpha$ moment in every direction.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

N/A

Reviewer QF3Z6/10 · confidence 4/52023-07-03

Summary

This paper considers sample complexity of (robust) mean estimation possibly without moments. Two typical examples are product distributions and elliptical distributions. The main technical contribution of this paper is to adapt the filtering techniques to the setting with less restrict moment assumptions.

Strengths

The results of the paper are interesting and new; while the idea may not be. The basic argument relies on filtering, which was previously proposed by [DKK+17] and [DKK+19]. The authors of the paper make a good observation on how the idea of filtering (and coupling) can be used with less restrictive moment assumptions (but it finally replaced by some concentration). The paper is well written, and I enjoyed every minute reading it.

Weaknesses

As mentioned, the main concern is that the authors replace the moment conditions with "concentration", i.e. $P(|\eta| \le \rho) > \frac{1}{100}$ and $P(R \le \sqrt{2d}) \ge \frac{1}{100}$. Moreover, from technique viewpoint, the main components, filtering and coupling (identifiability), are not novel. The time complexity is not explicit (though polynomial). Also I think the paper is lack of (synthetic or real) experiments.

Questions

I have several questions and comments. (1) In most "useful" applications, some moment assumptions may be needed. For instance, as the authors mentioned, covariance structure may be helpful. Also I may suggest to use "location" parameter instead of "mean". One case is the product Cauchy but it does not have a mean. (2) p.1, line 16: It may be better to say SQ refers to statistical query (if I don't miss it). (3) p.3, Def 1.2: If I understand correctly, the variables $(|\eta_j|)_{j = 1}^d$ can be dependent. If so, this is worth commenting since it is (apparently may be slightly) stronger than the usual sense of product measure. (4) Theorems 1.4-1.6: In all theorems, it is stated that the time complexity is "polynomial", i.e. $n^{\mathcal{O}(1)}$. However, it is not clear whether it is $n$ or $n^{100}$ (this makes difference in high dimensions). I understand that exact or a bound may be difficult. But I hope the authors may clarify this.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer 371F7/10 · confidence 3/52023-07-03

Summary

This paper studies the robust mean estimation problem without any moment assumptions. Instead, they consider a class of symmetric distributions, that is, semi-product and elliptical distributions. They develop a method based on Huber loss and the classic filtering technique that can achieve the same error rate as if the underlining distributions are sub-Gaussian. Their sample complexities are nearly optimal (with additional log factors).

Strengths

- The idea to use Huber loss and develop a similar result for the classic filtering method is very smart. I think this idea can be generalized to other settings and can be of independent interest. - The paper is well-structured. The techniques part clearly provides the motivation and is very readable.

Weaknesses

- The title is a little bit overclaimed. This paper studies the robust mean estimation problem without moments, but with the constraint that the distribution should be symmetric. I think it would be better to reflect this constraint in the title. - The notation style is not consistent. Based on my understanding, the authors use bold font for random variables. However, in some cases, like line 95, some symbols are not in bold font. - A quick question: I can understand that Huber loss has many fantastic properties. Compared with l2-loss, it is more robust against outliers; compared with l1-loss, it is differentiable everywhere. However, l1-loss is only not differentiable at 0. Hence, my question is can we replace the Huber loss with l1-loss? If not, can you briefly mention what is the difficulty?

Questions

Please see the weakness part.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Please see the weakness part.

Reviewer 6FtY7/10 · confidence 4/52023-07-11

Summary

This work studies robust mean estimation for symmetric distributions. There may possibly no moments (not even the first moment, hence the term location instead of mean), so previous efficient algorithms that rely on strong distributional assumptions may not work on the heavy-tailed setting. The main result is that similar statistical guarantees as the Gaussian setting can be obtained for $\rho$-semi-product distributions, where a semi-product distribution is a set of distributions that covers slightly more than that of product distributions. For instance, semi-product distributions include elliptical distributions, which is not a product distribution. For the case of known covariance matrix (or scatter matrix for elliptical), the guarantee obtained nearly matches best-known guarantees for Gaussians. For the unknown case, they obtain an error guarantee of $O(\epsilon^{1-\frac{1}{2k}})$ using $\tilde{O}(d^k)$ samples. The key approach is to generalize the filtering idea for robust mean estimation with a Huber loss instead of a quadratic loss in order to be able to handle distributions without second moments. Once the Huber loss is properly incorporated into the filtering technique, the necessary certificate for filtering can be obtained algorithmically without using sum-of-squares proofs for symmetric distributions.

Strengths

- This work studies a setting which (partially*) generalizes previous works to handle heavy-tailed distributions in which first or second moments may not exist. Here, they incorporate Huber loss into the common filtering technique that has been used in much of the recent robust learning literature. This in itself is novel. Furthermore, by exploiting symmetry, the filtering technique is made more algorithmically feasible without relying SoS approaches. - Through studying heavy-tailed symmetric distributions, such as elliptical distributions, in the case of unknown covariance, the paper obtains error bounds of $o(\sqrt{\epsilon})$ that were not known for general subgaussian distributions. - The main result obtains strong guarantees that nearly match that of Gaussians. *refer to Weakness 1

Weaknesses

1. Only a minor weakness: though the abstact motivates the work by stating that previous efficient estimation algorithms assume strong distributional conditions, the symmetry distributional assumption seems also possibly strong. While it is able to incorporate heavy-tail distributions, it also limits itself in generality (as ever-slightly altered Gaussians are not symmetric but have strong concentrations to exploit). 2. While I enjoyed the content of the paper, he main body seems abruptly cut off without a conclusion or final algorithmic overview. The presesntation would be greatly improved with a final algorithmic description along with a retrospective conclusion.

Questions

- For the main theorems, there is an upper bound on $\epsilon$. What is the breakdown point here? If C < 1/2, what is the reason the proposed approach breaks down? - In Theorem 1.6, the intuitive definition of $k$ is unclear. Some exposition on $k$ would be helpful.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

No limtation addressed.

Reviewer DMaa2023-08-10

I thank the authors for their response. Before I consider changing my score, I have the following further remarks and questions/requests. - Knowledge assumptions: thank you for the clarification. This is important to add to the main paper and to discuss very clearly and explicitly. - Sub-Gaussian error and sample complexity: I'm ok with the error being written explicitly in terms of all the parameters, but I think the "close to sub-Gaussian error" claim does need to be dropped and changed to a weaker claim instead about almost-optimal sample complexity. - Theorem 1.6: I agree that's the algorithmic challenge. I'm just saying that I disagree with the claim that the result is almost optimal, and I think the claim needs to be dropped. - Fisher information rate: I believe this warrants a discussion in the paper (even if it's in the appendix for space reasons). [GLP23] shows that it's possible to achieve smoothed Fisher information rates for the non-robust (and 1-d) setting, which asymptotically converges to the true Fisher information rate. I'm just pointing out that we can perhaps aim for smoothed Fisher information rate for the non-robust part while retaining a good robust-error, and so the optimality claims in the paper needs to be toned down, given the symmetry assumption. - Motivating the distributional assumptions: I understand that space was limited in the rebuttals, but can the author respond with what they plan to say about the concrete motivations?

Authorsrebuttal2023-08-15

Dear Reviewer, Thank you for your additional comments. We will add a detailed discussion about what is known to the algorithm to the paper. We will rephrase the statements regarding the "non-robust error-rate" and its optimality, in particular, (a) changing them to statements about sample complexity and (b) mentioning that it is an exciting open question to go beyond the "Gaussian sample complexity/error rate", and hence, also beyond the guarantees our results achieve, and to aim for a sample complexity/error rate depending on the (smoothed) Fisher Information (similar to [GLP23]). We will also include a discussion about Fisher information similar to the one we had above. Regarding motivations for our distributional assumptions: We would like to recall that the motivation for our paper is mostly theoretical. In particular, our work is guided by the question in lines 66-67: "Do there exist classes of heavy-tailed distributions for which we can (efficiently) achieve the same robust error as for the Gaussian distribution?" A weaker version of this question is: "For what classes of distributions can we (efficiently) achieve robust error $o(\sqrt{\varepsilon})$?" We believe characterizing these distributions is an important foundational question in algorithmic robust statistics. We believe a natural approach to this question (and the first) is considering classes of distributions satisfying the following criteria - they generalize the Gaussian distribution - in the non-robust setting, they behave as nicely as the Gaussian distribution - they are general, i.e., contain many distributions for which the question is still open - they are theoretically well-studied One such class that has been studied before were sub-gaussian distributions. However, additional assumptions (identity-covariance or "certifiably" sub-gaussian) are needed to go beyond $o(\sqrt{\varepsilon})$ in this case. As stated in our submission, our work identifies two other classes satisfying the criteria above, for which error $o(\sqrt{\varepsilon})$ is achievable. 1. Semi-product distributions: This generalizes the standard Gaussian distribution. And we show how to match the robust error of the Gaussian setting with nearly optimal sample complexity. This was previously known only for sub-Gaussian distributions with covariance identity. 1. Elliptical distributions: This generalises the "unknown covariance" Gaussian distribution in which it is only known how to achieve error $o(\sqrt{\varepsilon})$ for the Gaussian distribution, due to the algebraic structure of its moments, and "certifiably" sub-Gaussian distributions. We find it very appealing that there exist classes of distributions that is well-studied in the statistical literature, for which this is possible. Kind regards, The Authors

Reviewer QF3Z2023-08-11

Many thanks for the detailed explanations. The score remains unchanged.

Reviewer 6FtY2023-08-14

Reply to Authors

My questions have been answered. Thank you for the response.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC