Universal Online Learning with Gradient Variations: A Multi-layer Online Ensemble Approach

In this paper, we propose an online convex optimization approach with two different levels of adaptivity. On a higher level, our approach is agnostic to the unknown types and curvatures of the online functions, while at a lower level, it can exploit the unknown niceness of the environments and attain problem-dependent guarantees. Specifically, we obtain $\mathcal{O}(\log V_T)$, $\mathcal{O}(d \log V_T)$ and $\hat{\mathcal{O}}(\sqrt{V_T})$ regret bounds for strongly convex, exp-concave and convex loss functions, respectively, where $d$ is the dimension, $V_T$ denotes problem-dependent gradient variations and the $\hat{\mathcal{O}}(\cdot)$-notation omits $\log V_T$ factors. Our result not only safeguards the worst-case guarantees but also directly implies the small-loss bounds in analysis. Moreover, when applied to adversarial/stochastic convex optimization and game theory problems, our result enhances the existing universal guarantees. Our approach is based on a multi-layer online ensemble framework incorporating novel ingredients, including a carefully designed optimism for unifying diverse function types and cascaded corrections for algorithmic stability. Notably, despite its multi-layer structure, our algorithm necessitates only one gradient query per round, making it favorable when the gradient evaluation is time-consuming. This is facilitated by a novel regret decomposition equipped with carefully designed surrogate losses.

Paper

References (29)

Scroll for more · 17 remaining

Similar papers

Peer review

Reviewer rs9j8/10 · confidence 4/52023-07-04

Summary

In this work, the authors aim to develop an algorithm for online convex optimization that can adapt to the specific curvature of the loss function without knowing it in advance. The goal is to develop an algorithm that can adapt to strongly convex, exp-concave and convex loss functions dynamically, while also deriving some problem-dependent guarantees. They also ensure that their method is computationally efficient as it only requires one gradient query per round. To do so, they present a three layer algorithm that estimates both the type of loss functions and its parameters. Crucially, this kind of multi-layer approach generally comes at the cost of logarithmic factors. These three algorithms follow pre existing structures (MSMWC (Chen et al. 2021) for the meta algorithms and an optimistic Online Mirror Descent for the base algorithm) but the authors rely on several tricks and refinements of the analysis to exploit negative terms in the regret analysis and obtain competitive regret bounds. Finally, they also provide a novel decomposition of the regret which allows to reduce the number of gradients queries from $O(\log^2 T)$ to $O(1)$ at each round. This result is of independent interest, as they show that it can be used with different meta algorithms, such as Adapt-ML-Prod (van Erven and Koolen 2016) to recover worst-case guarantees (whereas the main results of the paper are stated in terms of problem dependent terms).

Strengths

This paper provides significant contributions to the field of OCO. Achieving optimal results simultaneously across several environments has been an important topic in learning theory in the past few years, and, by providing universal guarantees, this paper fits in this line of work. The presented algorithm builds upon existing results and provides non-trivial improvements to their analysis. These technical contributions are interesting and likely to be build upon in the future. It is worth noting that this paper is particularly well written. The body of the paper focuses on providing detailed explanations for the choice of methods, and the authors take a particular care in higlighting the connections and differences with the rest of the litterature. Discussionss such as in Section 3.2.1 which highlight the different approaches that could have been considered and were discarded are not that common and should be encouraged. The proofs, deferred to the appendix, are clear and detailed. Appendix A provides several applications for the results, showing that the method can be used to bridge the gap between stochastic and adversarial online convex optimization and for two-player zero-sum games. In both cases, the proposed algorithm obtains regret bounds that are at least matching up to log-factors and sometimes improving upon the state of the art while considering a broader setting.

Weaknesses

While many comparisons with other results in the litterature have been provided, it would be interesting to get comparisons with lower bounds, in particular for the small loss bounds that depend in F and V. As the results achieved by the current method are matching these of Zhang et al. (2022), though with a better runtime complexity, it could also be interesting to get experimental results to compare them.

Questions

Could you discuss if and how you think these results could be further improved?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

4 excellent

Contribution

4 excellent

Limitations

NA

Reviewer osvq9/10 · confidence 3/52023-07-09

Summary

This paper put together and designed new techniques to reach a high level of adaptivity in online convex optimization. It is a timely contribution that is useful to put together various algorithms optimal for different problem setups and to better adapt to the characteristic of each setup.

Strengths

This is an excellent paper in presenting the results and the qualitative explanation of the techniques or proofs. The adaptivity contribution of the paper is also very satisfying in terms of general knowledge of machine learning.

Weaknesses

I do not see any weakness in this paper. However, I have not checked the proofs in the appendix.

Questions

No question.

Rating

9: Very Strong Accept: Technically flawless paper with groundbreaking impact on at least one area of AI/ML and excellent impact on multiple areas of AI/ML, with flawless evaluation, resources, and 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

4 excellent

Limitations

None.

Reviewer wYJC7/10 · confidence 4/52023-07-09

Summary

This paper proposes algorithms that have data-dependent regrets and adaptive to the loss function classes. The algorithm improves on Zhang et al., 2022 in the sense that 1) it also achieves a gradient variation bound for convex Lipschitz losses and 2) it requires only one gradient evaluation at each round. The first is achieved by the introduction of optimism. The second is achieved by appropriately designed surrogate functions for the base learners. --- The authors have addressed my questions. I keep the original rating.

Strengths

1. The algorithm is new and improves on Zhang et al., 2022 (see the summary). 2. The paper clearly addresses the ideas behind algorithm design and analysis.

Weaknesses

1. **Comparison of complexity.** Please also compare the number of the instances of the base learners required for each algorithm in Table 1. This information helps evaluate the computational complexity. 2. **Algorithm 1.** It is said that the base learners are set "as specified in Section 2," but Section 2 (preliminaries) does not give any algorithm. 1. **Presentation.** - Section 4 is presented in a hasty way. The reader may not be able to understand the ideas without reading Appendix C. - Table 1: The Lipschitzness and smoothness assumptions are missing. The assumptions may not hold in, e.g., online portfolio selection. - Ln. 35: That "the learner requires to know the function information (type and curvature) in advance to select suitable algorithms" only holds for textbook OCO algorithms and not general OCO. - The three aspects in Ln. 57--60 can be skipped as that will be addressed in detail in Ln. 79--89. - Ln. 58: The term "small-loss bound" has not been explained before. It is better to explain it as the term is not common sense to general machine learning researchers. - Ln. 125: That $N \approx 1 + \log T + \log T$ looks weird. Perhaps there is some typo. 2. **Typos.** - Ln. 57: aspect*s* - Ln. 219: *for* our purpose

Questions

Please address the first two weaknesses and the possible typo in Ln. 125.

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

This is a theory paper. All assumptions are explicitly written. The potential issue regarding the computational complexity has been raised in the weakness block above.

Reviewer ZhYs6/10 · confidence 3/52023-07-18

Summary

A new algorithm for online learning is provided, together with improved regret bounds compared to the state-of-the-art: The new regret guarantees are summarised in table 1 on page 2: in the case of a strongly convex function, the regret guarantee is of $O(min(\log(V_T),\log (F_T)))$ where $V_T$ and $F_T$ are the gradient variation and cumulative loss of the best comparator respectively (cf. equation on lines 53 to 54). In the case of exp-concave functions, the rate is the same, whereas in the case of a convex function, the rate is $\widetilde{O}(\min(\sqrt{V_t},\sqrt{F_T})$ instead. In terms of the regret bound, only the convex case is an improvement over the existing results ([3], $O(\sqrt{F_T})$). An additional claimed advantage is that the gradient query is reduced to $O(1)$ instead of $O(\log(T))$: this is valid for an improved algorithm (separate from the main algorithm presented in the paper) with additional surrogate losses, which is only presented in the appendix on page 18 (Algorithm 2) The main algorithm itself is a combination of several existing techniques: at each iteration, a set of $KN$ experts $(k,i)\in [K]\times [N]$ each consisting of different gradient descent methods is maintained. Each expert computes the gradient update as per Optimistic OMD [3]. Then, for each fixed $k$, the experts $(k,i')$ for $i'\leq N$ are aggregated exactly as in MsMwC (cf. [1] and equation (2.3)), then the resulting experts are aggregated again via another use of MsMwC, with different parameters. The proof is a very long list of re-expressions of the regret via various existing results and additional calculations, drawing a lot of inspiration from [1], which itself improves on techniques in [2]. =================Post-Rebuttal============= After discussing with the authors, I am convinced that the presentation issues are theoretically fixable, and I still think the amount of work is adequate. Thus, I will keep my score and hope the authors fix all the minor errors as promised and try hard to make the paper more accessible to a broader audience. ============== **References** [1] Sijia Chen, Wei-Wei Tu, Peng Zhao and Lijun Zhang. Optimistic Online Mirror Descent for Bridging Stochastic and Adversarial Online Convex Optimization. ICML 2023. [2] Pierre Gaillard, Gilles Stoltz and Tim Van Erven. A Second-order Bound with Excess Losses. COLT 2014. [3] Lijun Zhang, Guanghui Wang, Jinfeng Yi and Tianbao Yang. A Simple yet Universal Strategy for Online Convex Optimization. ICML 2022.

Strengths

This is a highly technical and long paper and it is clear the authors know what they are doing. The parts of the proofs I managed to check are generally correct with only minor typos (though they are not reader friendly).

Weaknesses

In my opinion, the main paper is poorly written. I understand that the authors are trying to explain the intuition for why they construct their algorithm in this way, but there isn't enough context or revision of existing methods to make the paper readable. The paper reads like a soliloquy of a researcher trying to solve the problem at hand: for instance, in lines 140 to 148, the authors are saying "if we used Adapt-ML [2], bad things ill happen, what can be done about this?" then later in Section 3, their solution is proposed, which consists in using [3] instead, with a purportedly novel choice of optimism. See also how the authors are using expressions like "Another try is to" on line 206. This is an issue not just in terms of grammar but also because it occurs in the context of a vague exposition that follows the stream of consciousness of the author rather than trying to present a cogent story. Here are examples of more concrete **issues with the presentation**: **1** There is a lot of talk in the main paper of "how to cancel out terms" which comes a long time before the part where the authors introduce their main algorithm. When they finally decide to do so on page 7, it is only the basic version of the algorithm (not the one with the $O(1)$ gradient calls) and it still vaguely refers to "the base learners described in Section 2", which I can only imagine refers to what is introduced in line 168 (section 3.1), but is still presented as new in line 177 "to unify all kinds of functions, we propose...." **2** The concept of "optimism" isn't properly introduced either in the main or in the supplementary; the only definition is on line 72 in the intro: "a hyperparameter encoding historical information", which I am given to understand means slightly different things in different algorithms. A few more words of explanation here would be nice. Definitely, the base learners should be defined properly in the main text. In the current form, not even the reference to [3] is especially clear outside of Table 1. The same goes for the section names, terms such as "exogeneous negativity" are not very informative. In footnote 2, the authors even speak of "implementing $m_t$", which is definitely very informal and should be explained more clearly. **3** Note also that the Bregman Divergence is note defined in lines 486 to 487 or 211-212 (page 6), despite the fact that it would make the description more complete and that it is indeed defined in the reference [1] when the same equation is introduced. **4** In the proof of Lemma 2, the first equation is hard to understand without background in the field. the only explanation is "by standard analysis of optimistic OMD". It would be nice to say that this refers specifically to Theorem 1 of [4], which is explained better later in the appendix when the argument arises again (line 767 page 30). Since it is natural for the reader to read the proof of Lemma 2 first, the more detailed explanation should be present there as well. **5** Similarly, the proof of Lemma 1 on page 14 uses equation E.1 from page 29, which is completely unnecessary as E.1. is derived from first principles. **6** The proof of Lemma 2 is done for "an arbitrary comparator $u\in\Delta$, but the concept of comparator isn't really defined there, though it becomes clear from context after some effort. This also seems to run counter to the fact that in the main paper the authors write elsewhere in the paper that they "focus on the proof for fixed learning rate", which seems not to be the case in the case of the proof of Lemma 2. **7** What does the notation $\|.\|_{U_t}$ mean on line 761? Is it explained somewhere? I understand that I am not an expert in the field and my difficulties understanding the bigger picture is mostly due to this, but I do feel like in the case of well written papers, I am often able to make better progress, in less time, towards understanding more complex papers than this one in equally distant areas. However, I am not sure, it may be that the prerequisites are intrinsically larger. **References** [1] Sijia Chen, Wei-Wei Tu, Peng Zhao and Lijun Zhang. Optimistic Online Mirror Descent for Bridging Stochastic and Adversarial Online Convex Optimization. ICML 2023. [2] Pierre Gaillard, Gilles Stoltz and Tim Van Erven. A Second-order Bound with Excess Losses. COLT 2014. [3] Lijun Zhang, Guanghui Wang, Jinfeng Yi and Tianbao Yang. A Simple yet Universal Strategy for Online Convex Optimization. ICML 2022. [4] Zhao, Peng ; Zhang, Yu-Jie ; Zhang, Lijun ; Zhou, Zhi-Hua. Adaptivity and Non-stationarity: Problem-dependent Dynamic Regret for Online Convex Optimization. ArXiv. May 2023.

Questions

**1** Can you describe the base learner from [3] concisely? Perhaps rewrite the main paper in a more careful way? **2** I am a bit confused about what you mean by the $O$ notation sometimes: which constants are considered to be "$O(1)$ exactly? In page 31, it seems like $\gamma$ is treated as a variable (not $O(1)$), but it is less clear in the case of $D$ and $G$: at the end of the proof of Lemma 10, **the $O(1)$ notation hides a term of $4DG^2$**. If you can absorb this into the $O(1)$ notation, why do you still need to write $\gamma D^2$? I can sort of see that the reason is that there is a $\gamma$ in there, but providing a bit of context might help. **3** Do you really mean an equal sign at the second equality (first line) describing the Adaptivity term in page 30? it seems like if $1/\gamma<\frac{D}{\sqrt{1+\bar{V_T}}}$, **the equality wouldn't hold**. More worrying still, the inequality would be in the wrong direction. **4** In lemma 3, you make the assumption that $(\ell_{t,k -m_{t,k}})^2 =\langle l_{t,k}-m_{t,k},p_{t,k}\rangle^2$. Could you provide more context there? **5** Could you summarize the difference between your work and [1] more concisely? Why was the base learner [3] not used directly in [1]? What would happen if we used the correct base learner but did not use the hierarchical approach from your method? **Minor mathematical errors:** **1** I think the term $-8\sum_{t=2}^T \|p_t-p_{t-1}\|$ should be $-4\sum_{t=2}^T \|p_t-p_{t-1}\|$ in Lemma 2. Indeed, the definition of "term C" excludes the factor of $1/2$ at the bottom of page 14 (as evidenced not just by the curly bracket but also by the use of Pinsker's inequality in line 498 as there would need to be a factor of 16 instead of 32 if the 1/2 were included): this factor of $1/2$ needs to be reintroduced in line 513 on page 16. It would also be nice to quote Pinsker's inequality and perhaps add a citation at line 499 (page 15). **2** In the calculations on page 15, lines 497-498, at the third equality on the first line, the $b_i$ should be at the numerator instead of the denominator. **3** In the second line of the series of equations for the optimality gap in the proof of Lemma 10 on page 30, I think it should be $\leq$ instead of $=$ because there is a term of $-\|x^*-\hat{x}_{T+1}\|^2\frac{1}{2\eta_T}$ which is dropped out. **4** In the statement of Lemma 5, it should be $\|p-q\|_1^2$ instead of $\|p-p\|_1^2$ **5** In the proof of Lemma 1 on page 14, at the second line of equalities after line 482, I think it should be $\leq $ and not $=$ because of a missing $\langle g_T-g_{T+1},x_T-x_{T,i*}\rangle$ term which is added to make the sum look simpler. **6** In the definitions of the iterations of $p_t$ and $q_t$ from [1] in line 486 page 14 and also in line 211 page 6, the last term should be $D(p,q_{t-1})$ instead of $D(p,q_{t})$, consistent with equation 6 in [1]. **NB**: I have used q instead of \hat{p} due to issues with markdown. **Typos** line 7: "the O notation" (missing "the") line 146: constants=> constant line 161, same thing line 205 "it is still open that whether..>" ==> "it is still an open problem to determine whether" line 217 "same of" ==> "same as" line 252 "same of" ==> "same as" line 488: "follows the similar logic of" ==> "follows a similar logic as" line 506 "which equals to" ==> "which is equivalent to" (this somewhat changes the meaning and makes the line more understandable. Also, one might want to consider turning some of these inline equations into {align} blocks). line 739 "it the optimal step size" ==> "if the optimal step size"

Rating

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

1 poor

Contribution

2 fair

Limitations

The improvement compared to [3] is somewhat disappointing compared to the much larger complexity of the algorithm: it seems that eh main improvement is to use a single algorithm for all types of functions and to reach similar rates, with a better rate in the case of convex functions. This is interesting theoretically but the algorithm seems too unwieldy to be used in either practice or future work. I cannot fully vouch for the correctness as I often had to read the supplementary line by line without fully grasping the bigger picture. It is also hard to vouch for the originality compared to [3] as I am not familiar enough with it. Nevertheless, this seems like solid work, assuming the presentation was improved.

Reviewer x4hL6/10 · confidence 3/52023-07-25

Summary

This paper proposes an online convex optimization algorithm that is adaptive in two ways to the environment. Specifically, it can achieve logarithmic regrets for problems with good properties such as strong convexity and exp-concavity, as well as $\sqrt{T}$ regret in the worst-case. Further, it achieves improved regret bounds for problems with small variation of the function gradients. In addition, it works with only one gradient call per round.

Strengths

- The paper is well structured and easy to follow. - Existing studies are adequately reviewed and the paper is informative for the reader. - The paper clearly explains the structure of the proposed method, the idea of the proof, and its implication in great detail.

Weaknesses

Compared to the study by Zhang et al. [2022], the magnitude of the contribution of this paper appears somewhat limited. Although the design of the algorithm is quite sophisticated, the main component is a combination of already known techniques. It would be better to have a more detailed description of the challenges in algorithm design and analysis, as well as the new techniques proposed in this paper.

Questions

Can you address the concerns described in Weaknesses above?

Rating

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

2 fair

Limitations

I have no concerns about the limitations and potential negative societal impact are adequately addressed.

Reviewer FwDP6/10 · confidence 4/52023-07-26

Summary

Addressing to the drawback of Zhang et al. [2022], which does not enjoy gradient-variation bounds for convex functions, using an optimistic version of ADAPT-ML-PROD with a second-order bound of [Wei et al., 2016],three-layer online ensemble structure with two-layer meta learner running MSMWC [Chen et al., 2021] with different parameter configurations, and each MSMWC-MID is further connected with N base learners to explore the unknown function information, is developed.$\mathcal{O}(\ln V_T)$, $\mathcal{O}(d \ln V_T)$ and $\hat{\mathcal{O}}(\sqrt{V_T})$ regret bounds for strongly convex, exp-concave and convex loss functions is obtained. two different levels of adaptivity with problem-dependent gradient variations, e.g., adaptive to unknown function information and benign Environments are fulfilled.

Strengths

In Zhang et al. [2022], the convex case has not been addressed. a two-layer framework In Zhang et al. [2022] is generalized to three-layer online ensemble structure. a single algorithm with gradient-variation bounds for all kinds of functions is designed.

Weaknesses

1 the derived results substantially depend on Zhang et al. [2022], [Chen et al., 2021], Zhao et al. [2021] and [Wei et al., 2016]. 2 The results of the algorithm still belong to the class of MSMWC [Chen et al., 2021]. the decomposed regret and injecting cascaded correction terms to both the top and middle layer make the work replicate published findings without adding substantial knowledge, leading to lack of novelty.

Questions

I believe that exists other gradient-variation bounds, which is similar to [Wei et al., 2016], Can better results be obtained by means of it?

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

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

the authors adequately addressed the limitations.

Authorsrebuttal2023-08-10

Response to other questions

In this reply, we respond to presentation/typo issues and other questions, hoping to help you understand our work better. Since some LaTeX commends of `\hat{}`, `\boldsymbol{}` do not work well in OpenReview, we use other notations instead or simply omit them. --- In this part, we clarify the **presentation issues**. **Q1**. The presentation order. **A1**. We truly appreciate your advice. We will reorder the content in the revision to provide readers with an overview of the entire paper before delving into the specifics. **Q2**. The concept of "optimism"; The definition of base learners; Section names such as "exogenous negativity"; "implementing $m_t$". **A2**. We greatly appreciate your suggestions. We will provide more explanations about optimism when it first appears. Given that an extra page is allowed in the camera-ready version, we will include a more comprehensive formalization of base learners. And we will carefully revise the paper to remove not informative statements. **Q3**. The Bregman divergence is not defined. **A3**. Thanks. We will include the definition in the revised version. **Q4**. In the proof of Lemma 2, the first equation is hard to understand without a background in the field. **A4**. We appreciate your keen observation! We will cite Theorem 1 of Zhao et al. [2023] at this point. **Q5**. Referring to equation E.1 in the proof of Lemma 1 is unnecessary. **A5**. Thanks. We will remove it. **Q6**. The comparator and learning rate in Lemma 2. **A6**. Thanks for highlighting this. In Lemma 2, we provided a more general result with an arbitrary comparator and changing learning rates. This was done hoping that the negative stability term analysis would be comprehensive enough for readers interested solely in the MsMwC algorithm. We will state this clearly to prevent misunderstandings. **Q7**. What does the notation $\|\cdot\|_{U_t}$ mean on line 761? Is it explained somewhere? **A7**. Thanks for your sharp observation. It refers to the matrix norm: $\|\mathbf{x}\|_{U_t} = \sqrt{\mathbf{x}^\top U_t \mathbf{x}}$. We will provide its definition in the revised version. --- In this part, we answer your **questions**. **Q8**. The $O$-notation. **A8**. Thanks for raising this. The gradient norm $G$ and domain diameter $D$ are generally considered constants in constrained online learning. As a result, we incorporate them into the $O(1)$ term. Regarding $\gamma$, since it can take different values in various scenarios (e.g., please refer to Line 694 and Line 704), we do not include it in $O(1)$, even though $\gamma$ is always a constant. **Q9**. Do you really mean an equal sign at the second equality (first line) describing the Adaptivity term in page 30? **A9**. Thanks for the question. The correct derivation should use $\le$ due to $\eta_t = \min\{D/\sqrt{1+\bar{V}_t}, 1/\gamma\} \le D/\sqrt{1+\bar{V}_t}$. **Q10**. In lemma 3, you make the assumption that $(\ell_{t,k} - m_{t,k})^2 = \langle \ell_{t,k} - m_{t,k}, p_{t,k} \rangle$. Could you provide more context there? **A10**. We appreciate your insightful question. This is not an assumption but rather a *condition* that is inherently satisfied by our algorithm. For context, this condition serves to ensure a self-contained proof for our meta algorithm and has been validated, for instance, in Line 595. We appreciate this inquiry and will refine the manuscript to underscore that this is a condition, not an assumption. --- In this part, we answer your remarks on some minor **mathematical errors**. **Q11**. The constant before $\|p_t - p_{t-1}\|_1^2$ and the citation to Pinsker's inequality. **A11**. Thanks for the sharp observation! It is indeed a typo. We will rectify this typo by adjusting the constant to 4 and will cite Pinsker's inequality. **Q12**. In the calculations on page 15, lines 497-498, at the third equality on the first line, the $b_i$ should be at the numerator instead of the denominator. **A12**. Thanks. $b_i$ should be at the numerator, and we will revise it. **Q13**. In the second line of the series of equations for the optimality gap in the proof of Lemma 10 on page 30, I think it should be $\le$ instead of $=$. **A13**. Thanks. We will correct it to be $\le$. **Q14**. In the statement of Lemma 5, it should be $\|p - q\|_1^2$ instead of $\|p - p\|_1^2$. **A14**. Thanks for catching this error! It is a typo, and we will correct it. **Q15**. In the proof of Lemma 1 on page 14, at the second line of equalities after line 482, I think it should be $\le$ and not $=$. **A15**. Thanks. It is a typo, and we will revise it to be $\le$. **Q16**. In the definitions of the iterations of $p_t$ and $q_t$ from [1] in line 486 page 14 and also in line 211 page 6, the last term should be $D(p, q_{t-1})$, where $q$ denotes `\hat{p}`. **A16**. Thanks. We will revise it in the follow-up version. And thanks for finding some additional typos in both the main paper and the appendix. We will revise them carefully.

Reviewer ZhYs2023-08-18

Follow-up questions

Thanks for the many clarifications and for agreeing to correct the errors. Please do not forget to include a clarification about which quantities you include in the O notation (Q8) and also an explanation of the difference with [Chen et al. 2023]. I still have a few remaining questions: 1I am still not sure about your answer to my point about the equality at the top of page 30 (Proof of Lemma 10). I see you agree with me that it should be an inequality rather than an equality, but still, that doesn't fully explain why you are using the formula $\eta_t=\min(\frac{D}{1+\bar{V}_t},\frac{1}{\gamma})$. Indeed, it seems this condition is taken from the assumptions in Lemma 8 instead. The assumption in Lemma 10 is instead $\eta_t=\frac{2}{\gamma+\lambda_t}$. How do you reconcile this? What is the correct assumption on $\eta_t$ (and how does it ensure that the equation just before line 771 still holds)? 2. In your answer to my Q10, you claim that the condition is "inherently satisfied". Can you explain from which equation in the definition this follows ? (both in this rebuttal and later, in the paper). 3. In A2, you claim that by applying your results to Chen et al, you solve a "Major open problem". This seems like an overstatement. Was this open problem suggested at any time before the Arxiv publication of Chen et al in february 2023?

Authorsrebuttal2023-08-19

Response to the follow-up questions

Thanks for the follow-up questions and helpful comments. We will revise the paper according to your suggestions. Below we answer your technical questions. --- **Q1**. The step size issue in Lemma 8 and Lemma 10. **A1**. We clarify that the setup of the step size $\eta_t$ is *not* an assumption but used in the base algorithms. Lemma 8 and Lemma 10 provide the regret guarantees of base learners for *different* kinds of functions (convex and strongly convex), respectively. Briefly, as explained in A1 in the first response about the base learners from [Zhang et al., 2022], we employ diverse base learners to hedge the uncertainty of the unknown function type and curvature coefficient. Therefore, it is necessary to establish the regret guarantee for each function type separately. To this end, we provide Lemma 8 (for convex functions), Lemma 9 (for exp-concave functions), and Lemma 10 (for strongly convex functions). Specifically, - Lemma 8 provides the regret guarantee of the base learner for *convex* functions, which runs OGD with the step size $\eta_t = \min(D/\sqrt{1 + \bar{V}_t},1/\gamma)$. - In contrast, Lemma 10 provides the regret guarantee of the base learner for *strongly convex* functions, which runs OGD but with a different step size $\eta_t = 2/(\gamma + \lambda t)$. - Additionally, Lemma 9 provides the regret guarantee of the base learner for *exp-concave* functions, which runs the optimistic OMD with a local-norm regularizer. Therefore, Lemma 8 and Lemma 10 do not conflict since they refer to different base learners designed for different function families. --- **Q2**. Condition validation of Lemma 3. **A2**. Lemma 3 is only used in the proof of Theorem 2. Concretely, we use it in Line 595 to analyze the meta regret. Before applying Lemma 3, we have validated that its conditions can be fully satisfied, as can be seen in the second line of Line 594 (with detailed validations in Line 591-594). We will revise the paper to make it more clearly. --- **Q3**. SEA-related questions. **A3**. We mention that applying our universal algorithm to the SEA model directly addresses the major open problem *originally* proposed by [Chen et al., 2023]. So the open problem didn’t appear before Feb 2023. In fact, the study of SEA was initiated very recently by [Sachs et al., NeurIPS 2022], whose manuscript was initially released in Feb 2022 as an arXiv version and later published at NeurIPS 2022. In the following, we provide a brief review of the development of the SEA model and clarify the importance of the aforementioned open problem. - The SEA (Stochastically Extended Adversarial) model, serving as a natural bridge between stochastic optimization and adversarial online optimization, was recently proposed by [Sachs et al., NeurIPS 2022] and continues to be an area of active research. - The most up-to-date findings are provided by [Chen et al., ICML 2023]. Although they introduced a unified OMD framework and achieved favorable guarantees for convex/exp-concave/strongly convex functions, respectively, their algorithms require separately designed parameter settings. Thus, a central open problem left by [Chen et al., 2023] is to design a (single) adaptive algorithm that can handle potentially different functions while maintaining optimal regret guarantees. - This adaptivity concern has garnered interest from the community. Notably, [Sachs et al., arXiv 2023], a work concurrent to ours (their manuscript was posted to arXiv online on Mar 6, slightly ahead of NeurIPS ddl), also investigates the adaptivity of the SEA model but only attains partial adaptability. Our results are *strictly better than theirs*, achieving the state-of-the-art. For an overview of comparisons, please kindly refer to Table 3 on Page 13. **References:** [1] Sachs et al., Between Stochastic and Adversarial Online Convex Optimization: Improved Regret Bounds via Smoothness, NeurIPS 2022 [2] Chen et al., Optimistic Online Mirror Descent for Bridging Stochastic and Adversarial Online Convex Optimization, ICML 2023 [3] Sachs et al., Accelerated Rates between Stochastic and Adversarial Online Convex Optimization, arXiv 2023 --- We greatly appreciate your thorough review and helpful feedback. We will work on refining the presentation to enhance clarity and avoid potential misunderstandings.

Reviewer wYJC2023-08-14

Thanks

Thanks for the response. I will keep the rating. **Q1.** Please include this information. Thanks. **Q2.** I raised this issue simply as a matter of presentation. The statement may confuse some readers, as Lines 125--129 are buried in a paragraph in the literature review. Please consider rewriting the related sentences to make things clearer. **Q3.** Please add a few words to explain the reason for the two $\log T$ terms.

Authorsrebuttal2023-08-14

Thanks for the suggestions!

**Response to Q1:** Thanks for the recommendation. We'll incorporate a comparison of base learner numbers from various methods in our revised version. **Response to Q2:** Thanks for the sincere advice. Given the extra page being permitted in the camera-ready version, we will rewrite this part to give a more detailed explanation of the base learners. **Response to Q3:** Thanks. We'll provide further clarifications in this section. Finally, thanks again for the thorough review and valuable suggestions!

Reviewer x4hL2023-08-16

Thank you very much for your very thorough response. I now have a better understanding of the importance of this research. I maintain my positive score.

Authorsrebuttal2023-08-20

Dear Reviewer FwDP, We sincerely appreciate your helpful comments. As the period for author-reviewer discussions is coming to an end, we kindly ask whether our response has properly addressed your concerns and potential misunderstandings. Please let us know if you have any more questions, and we are happy to provide further clarification. Thanks! Best Regards, Authors

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC