Near-optimal learning with average Hölder smoothness

We generalize the notion of average Lipschitz smoothness proposed by Ashlagi et al. (COLT 2021) by extending it to Hölder smoothness. This measure of the "effective smoothness" of a function is sensitive to the underlying distribution and can be dramatically smaller than its classic "worst-case" Hölder constant. We consider both the realizable and the agnostic (noisy) regression settings, proving upper and lower risk bounds in terms of the average Hölder smoothness; these rates improve upon both previously known rates even in the special case of average Lipschitz smoothness. Moreover, our lower bound is tight in the realizable setting up to log factors, thus we establish the minimax rate. From an algorithmic perspective, since our notion of average smoothness is defined with respect to the unknown underlying distribution, the learner does not have an explicit representation of the function class, hence is unable to execute ERM. Nevertheless, we provide distinct learning algorithms that achieve both (nearly) optimal learning rates. Our results hold in any totally bounded metric space, and are stated in terms of its intrinsic geometry. Overall, our results show that the classic worst-case notion of Hölder smoothness can be essentially replaced by its average, yielding considerably sharper guarantees.

Paper

Similar papers

Peer review

Reviewer uWsQ7/10 · confidence 3/52023-06-13

Summary

The authors proposed a notion named average Hölder smoothness, which is a generalization of average Lipschitz smoothness. They gave sample complexity bounds of learning the averge Hölder smooth function class in both the realizable and agnostic setting with corresponding algorithms. They also gave lower bounds which matches the upper bound up to logarithmic factors.

Strengths

1. The writing is clear. The paper is easy to follow and the authors explain their contributions well. The technical part in the main article, although concise, highlights the major part of the proof. 2. The contributions are valuable. The sample complexity of learning an average smooth Hölder function class is previous unknown even in the special case of average Lipschitz smoothness and so is the lower bound. The authors also showed the class of average smooth Hölder functions are strictly richer than previous notions like average smooth Lipschitz functions. 3. The algorithm is novel and has great potential. Indeed the classical ERM fails on learning the average smooth Hölder function class in the distribution free setting so it is conceivable that the algorithm could be used to solve some real problems arising from practice that couldn't be solved before. 4. The results are quite complete. The upper bound is tight up to only logarithmic factors and the bounds is analogous to classical PAC learning sample complexity bounds.

Weaknesses

1. I'm not quite convinced that there's no efficient algorithm for the agnostic case unless more evidence is provided. The reduction from agnostic setting to realizable setting is quite generic. But for this learning problem, it feels quite possible that we can find an efficient algorithm. 2. There are some typos. In line 182 it should be "There exits a positive number $N$" instead of a "function".

Questions

1. Is it possible to derive tighter bounds for subclasses of average Hölder smooth function classes? In the paper only learning from the entire class is considered but it could be much easier to learn from a subclass in some cases, right? Is it possible to derive a bound dependent on the specific function class? If so, will the average Hölder smooth parameter still show up somewhere in the bound? 2. I think it's better to give some concrete examples in applications, like giving a setting where ERM fails and we can only use your algorithm. Learning finite sums with a distribution concentrating on positions where the smooth parameter is small may be a good example. 3. I think maybe you should compare your bound more with the classical PAC learning bound to help reader better understand. 4. Are there any other works giving sample complexity bounds on distribution free learning for generic or specific function classes? Maybe adding these to the related work section.

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors addressed the limitation of their theoretical work properly.

Reviewer VoQT7/10 · confidence 3/52023-07-07

Summary

This paper introduces the notion of "average Hölder smoothness", generalizing the notion of Holder smoothness. Authors motivate the introduction of this class, explaining that the Holder constant is a supremum over many slopes that can eventually not be representative of the general behavior of this slope. After a long recall on needed notions, authors listed their theorems. The first notable one (Thm 3.1) controls the backeting entropy of the newly defined class of functions with its covering number. This theorem is crucial as 1. the bracketing entropy is used to encode the number of "packs" (brackets) one needs to enclose all the average holder smooths functions, and then naturally bounds the generalization error (See Thm 3.2), 2. the covering number is more intuitive and known in classical case. Then, Thm 3.4 follows, providing the generalization error gap on this class of function depending on the number of sampled data. The next 2 sections discuss learning algorithms to minimize the empirical function keeping the generalization gap low. Note that an approximation of holder constants is used since its exact value cannot be computed. Finally, lower bounds are computed in the section 6. Overall, this paper seems self contained and complete, generalizing results obtained on average Lipschitz continuous function class.

Strengths

The paper is self-contained, well written, notions are explained clearly and the reasoning is explained step by step to obtain a generalization bound on a newly proposed class of functions that is much larger than the classical ones and describes functions better.

Weaknesses

Minor: - l.147: What about N=1? What is the log basis?
I checked in [Ashlagi et al., 2021, Lemma 22]: the log basis is e and there is a small mistake at the very end of their proof. The actual bound should be
$E[Z] \leq (1 + log(N)) W[Z]$. Finally, the provided bound only holds for $N\geq 3$.

Questions

- What motivates these definitions of « semi-local » holder constants? Would we be able to say something with $\Lambda_f^{\beta}(x) = inf_{\varepsilon>0} sup_{y\in B(x, \varepsilon) -\lbrace x \rbrace} \frac{|f(x)-f(y)|}{\rho(x, y)^{\beta}}$ ? - If I understand correctly, Thm 3.1 and 3.4 are claimed as contribution, but not Proposition 3.2. Where can we find it? - Thm 3.4: Why did we lose the $log(1/\varepsilon)$ term from Th 3.1?

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

4 excellent

Contribution

3 good

Limitations

- Cor 3.5: Number of needed data for good generalization is exponential in d.

Reviewer LpFP6/10 · confidence 2/52023-07-25

Summary

This paper generalizes the notion of average Lipschitz smoothness proposed by Ashlagi et al. by extending it to Hölder smoothness. Eventually, the class of the average holder smooth function is much larger than the class of the worst-case holder smooth function. Further, the paper provides upper and lower risk bounds in terms of the average Hölder smoothness for realizable regression and noisy regression.

Strengths

Though I am not an expert in this area, the results provided in the paper are clearly better than the previous result for the average Lipschitz case. The bounds provided in the paper for bracketing number is distribution-free. Using this distribution free bound on the bracketing number, the authors prove sample complexity for obtaining uniform convergence of average-Hölder functions in the realizable case. the learner does not have an explicit representation of the function class, running an ERM is not possible in this case. Hence, the authors construct a non-parametric regression algorithm with nearly an optimal rate. The authors also provide results for agnostic settings using the reduction from agnostic to realizable settings proposed in the previous work of Hopkins et al. [2022]. A matching lower bound is also provided in the paper.

Weaknesses

Overall, I find the paper quite well written. However, it would be great if more details about the algorithm could be provided in the paper. Again, this is coming from non an expert in this area. To obtain the result in theorem 4.1, what is the modification needed in the algorithm as running ERM is not possible? It would be great if that could be explained in detail. The same goes for the agnostic result. Also, the worst-case rate in remark 4.2 is exponential in dimension, can the authors please comment that under what conditions we might get rid of exponential dimension independence? For example, approximation results in sobolev space get rid of exponential dependence if the function is differentiable O(d) times.

Questions

See weaknesses.

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

3 good

Limitations

yes.

Reviewer LpFP2023-08-14

Thanks for the response. I keep my original score.

Reviewer VoQT2023-08-19

Thank you

I thank the authors for their detailed answers to my (minor) concerns.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC