Unbiased Compression Saves Communication in Distributed Optimization: When and How Much?

Communication compression is a common technique in distributed optimization that can alleviate communication overhead by transmitting compressed gradients and model parameters. However, compression can introduce information distortion, which slows down convergence and incurs more communication rounds to achieve desired solutions. Given the trade-off between lower per-round communication costs and additional rounds of communication, it is unclear whether communication compression reduces the total communication cost. This paper explores the conditions under which unbiased compression, a widely used form of compression, can reduce the total communication cost, as well as the extent to which it can do so. To this end, we present the first theoretical formulation for characterizing the total communication cost in distributed optimization with communication compression. We demonstrate that unbiased compression alone does not necessarily save the total communication cost, but this outcome can be achieved if the compressors used by all workers are further assumed independent. We establish lower bounds on the communication rounds required by algorithms using independent unbiased compressors to minimize smooth convex functions and show that these lower bounds are tight by refining the analysis for ADIANA. Our results reveal that using independent unbiased compression can reduce the total communication cost by a factor of up to $Θ(\sqrt{\min\{n, κ\}})$ when all local smoothness constants are constrained by a common upper bound, where $n$ is the number of workers and $κ$ is the condition number of the functions being minimized. These theoretical findings are supported by experimental results.

Paper

References (69)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer yR987/10 · confidence 4/52023-06-16

Summary

The authors consider the distributed convex optimization problem in the centralized setting. In this setting, the authors provide the new lower bounds on the total communication cost under the assumption that nodes send unbiased independent compressed vectors to a master. They also provide improved analysis of the current state-of-the-art ADIANA method in the general convex and strongly convex regimes.

Strengths

The paper provides two important things to the distributed optimization community: 1. While the upper bounds obtained by the ADIANA and CANITA methods are well known, finally, the authors provide lower bounds in the described setting. 2. The improved analysis of ADIANA is very surprising. I checked the proof. I'm not 100% sure, but the improved analysis seems to be correct. I believe that it is an interesting contribution on its own. All in all, I think that this paper deserves to be published in NeurIPS.

Weaknesses

However, I would like to point out some very important weakness that I want the authors to fix. The authors compare the total communication cost of compressed methods with the accelerated (Nesterov's) method. They claim that it is possible to improve the complexity by $\min [n, \kappa].$ I agree with this fact but only under the assumption that **the local smoothness constants are equal to the global smoothness constant.** The author ignore the fact (or I missed it in the text, then sorry me for that) that the local smoothness constants can be $n$ times larger than the global one. The authors define the local $L$--smoothness constant as $$f_i(y) \leq f_i(x) + <\nabla f_i(x), y - x> + \frac{L}{2} ||x - y||^2 \quad \forall i$$ While the non-compressed methods define it as $$f(y) \leq f(x) + <\nabla f(x), y - x> + \frac{L}{2} ||x - y||^2.$$ The $L$--constant of $f_i$ can be $n$ times larger than $L$--constant of $f$! I want to ask the authors to clarify it clearly in the paper. Ideally, in the abstract and in the contributions.

Questions

-

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

3 good

Contribution

3 good

Limitations

-

Reviewer JbF93/10 · confidence 2/52023-06-28

Summary

The authors study the extent to which unbiased compression can reduce the total communication cost of optimizing smooth convex functions. Additionally, they refine the analysis of ADIANA ([25]) to show that the bounds are near-optimal.

Strengths

Quantifying how much unbiased compression helps is an interesting topic. While I believe that nearly anyone (that uses compression) works with unbiased compressors in the distributed case, it is interesting to learn by how much it helps compared with biased compressors.

Weaknesses

It is not very surprising to me what without independence, unbiasedness cannot help in general. Further, I do not recall seeing works that do not use independence to get the error-cancellation, *other than works that use correlated compression to drive the error lower than that of i.i.d. compressors* (e.g., Correlated quantization for distributed mean estimation and optimization, ICML 2022). Also, the writing is hard to follow and seems to be contradictory in places. For example, You write, "Compared to lower bounds when using unbiased compression without independence [13], our lower bounds demonstrate significant improvements when n and κ are large". However, the upper and lower bounds of [13] seem to match in your Table 1, while your lower bound seems *weaker* than that of [13]. In general, there cannot be two matching sets of upper and lower bounds. Some of the claims also seem inflated; I don't view $O(\omega \epsilon^{-O(1)})$ as being near optimal for a lower bound of $\Omega(\omega\log\epsilon^{-1})$ -- the dependency on $\epsilon$ is exponentially worse. The improvement in the analysis of ADIANA is only applicable to some parameter range as it shaves an additive factor. (As a side note, it took me a minute to understand that by **ADIANA (Ours)** you don't actually mean that it's your *algorithm*, but just a refined analysis.)

Questions

Can you please clarify the differences in the asymptotics with [13]? (Also, what is specified in the table seems to be different than in their paper, where they have an additional additive term.)

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

Seems fine

Authorsrebuttal2023-08-14

Need more clarifications?

Dear reviewer JbF9, We thank you for your valuable comments. We have made detailed responses to address your concerns, but we have not received your replies to our current clarifications yet. We thus kindly ask if our responses have addressed all your concerns. If not, we are more than happy to provide more clarifications. Best, The authors of paper 5699

Reviewer MRty7/10 · confidence 4/52023-07-06

Summary

This paper investigates the communication saving of unbiased compression on distributed optimization. The main contributions are: i) a communication lower bound of distributed optimization algorithms with (not necessarily independent) unbiased compression; ii) a communication lower bound of distributed optimization algorithms with (independent) unbiased compression; iii) refined analysis of ADIANA to show tightness of these bounds; iv) discussions on the importance of independence to the communication saving.

Strengths

i) Overall, this paper is well written and easy to follow. ii) The investigated problem is important to distributed optimization, and the results are convincing.

Weaknesses

i) Section 1 claims that quantization and sparsification are modeled as unbiased compressors in some papers. But they can be biased too. Please clarify. ii) The error cancellation explanation of independence to unbiased compression makes sense. But, if $x_i$ at different nodes are quite different, do we still need independence to achieve the communication saving? iii) Table I considers both convex and strongly convex cases. Are there any difficulties in analyzing the non-convex case? iv) Some bounds in Table 1 are summations of two terms, for example, $\ln(1/\epsilon)$ and $\epsilon^{-1/2}$, $\epsilon^{-1/3}$ and $\epsilon^{-1/2}$, etc. Is it possible to reduce to one term? v) Which compressors satisfy Assumption 2 (unbiased compressor) and Assumption 3 (independent compressor)? Please comment below the assumptions. vi) The communication saving is discussed via comparing the lower bounds. However, in some datasets some algorithms might perform much better than the lower bounds. In this case, the comparisons would yield different conclusions. Please comment on this. vii) I give a score of 6 at this stage, and would like to change it after the rebuttal period.

Questions

i) Section 1 claims that quantization and sparsification are modeled as unbiased compressors in some papers. But they can be biased too. Please clarify. ii) The error cancellation explanation of independence to unbiased compression makes sense. But, if $x_i$ at different nodes are quite different, do we still need independence to achieve the communication saving? iii) Table I considers both convex and strongly convex cases. Are there any difficulties in analyzing the non-convex case? iv) Some bounds in Table 1 are summations of two terms, for example, $\ln(1/\epsilon)$ and $\epsilon^{-1/2}$, $\epsilon^{-1/3}$ and $\epsilon^{-1/2}$, etc. Is it possible to reduce to one term? v) Which compressors satisfy Assumption 2 (unbiased compressor) and Assumption 3 (independent compressor)? Please comment below the assumptions. vi) The communication saving is discussed via comparing the lower bounds. However, in some datasets some algorithms might perform much better than the lower bounds. In this case, the comparisons would yield different conclusions. Please comment on this.

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

3 good

Contribution

3 good

Limitations

N/A.

Authorsrebuttal2023-08-14

Need more clarifications?

Dear reviewer MRty, We thank you for your valuable comments. We have made detailed responses to address your concerns, but we have not received your replies to our current clarifications yet. We thus kindly ask if our responses have addressed all your concerns. If not, we are more than happy to provide more clarifications. Best, The authors of paper 5699

Reviewer u1Ro7/10 · confidence 3/52023-07-18

Summary

The paper analyzes the communication cost of unbiased compression algorithms for distributed optimization and proves that unbiased compression alone cannot reduce communication cost since whatever is saved via compression is lost due to an increase in the number of communication rounds needed for convergence. The authors then show that if in addition to being unbiased if the compressors are also independent then the lower bound on the number of communication rounds is reduced by a factor of $\Theta(\sqrt(\min(n, \kappa)))$ where $n$ is the number of nodes and $\kappa$ is the condition number of the function being minimized. They then improve the convergence analysis of ADIANA, an existing algorithm for communication compresssion in distributed optimization, and show that the lower bounds can be matched upto a log factor. Simulations on distributed least squares and distributed logistic regression corroborate their analysis and validate their claims.

Strengths

1. The authors identify an important gap in the existing literature on communication compression where emphasis is primarily laid on the unbiasedness of compressors, without also considering the effect of independence. They systematically build and explain their analysis providing theoretical proofs for all their claims. The theoretical analysis coupled with the clear explanation can make this an important and impactful addition to the literature. 2. Considering the total communication cost (per-round cost x number of rounds) makes the analysis more realistic and complete than previous works which just show savings in either per-round cost or on the number of rounds. Moreover, by proposing a general lower bound for all distributed optimization algorithms with unbiased and independent communication compression this work clarifies the target communication cost for all such communication compression algorithms which can make them easier to evaluate and analyze theoretically.

Weaknesses

1. The writing is a bit heavy on notation and the analysis may be hard to follow for readers unfamiliar with the literature. Especially since their is not clear example or intuition provided in the main paper to indicate why independence may help. 2. It is a bit unclear why ADIANA is chosen for analysis. The transition from deriving a general lower bound to analyzing a specific algorithm feels a bit abrupt. If possible, it would be good to provide some intuition for the kind of properties that would make a communication compression algorithm likely to be close to the lower bound thereby justifying the choice of ADIANA for analysis. This is a minor point however, as I understand that such an intuition may not be available at this time so it is okay if that is the case.

Questions

1. To address the first point under weaknesses above, I would recommend moving the content of Appendix A which provides intuition on why independence helps to the main pape. Algorithm 1 can be moved to the Appendix to make space, since it appears to be just a reproduction of the ADIANA algorithm from [25]. 2. Can you also provide an example of a compression scheme where the compressors are unbiased but not mutually independent and provide some intution for its sub-optimality or show it through any of the experiments in Section 6? 3. Were any changes made to the ADIANA algorithm at all in your analysis or experiments or was the algorithm the same as that in [25] everywhere?

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 have identified the major limitation of the analysis being limited to convex functions only and have marked that as an issue to be addressed in future work which is fine with me. I have identified a couple of more minor limitations under Weaknesses above and look forward to the authors' responses to those.

Reviewer tEWM7/10 · confidence 1/52023-07-26

Summary

Several communication compression strategies have been proposed in the last few years to improve distributed optimization. The authors consider the tradeoffs between lowering the communication costs per round vs the number of communication rounds and understand how they affect total communication costs. They theoretically formulate this and prove that under some assumptions, unbiased compression might not save total communication. However, if the compressors used by all workers are further assumed independent, they prove that total communication cost can be decreased. They also prove lower bounds on certain classes of algorithms by refining the analysis for ADIANA.

Strengths

The results are provably improved. The presentation is clear and easy to follow.

Weaknesses

The lower bounds are proved only for certain classes of algorithms.

Questions

It would be better if some experiments from the appendix were moved to main body.

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

1: Your assessment is an educated guess. The submission is not in your area or the submission was difficult to understand. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

NA

Reviewer oD7w7/10 · confidence 2/52023-07-27

Summary

The paper explores the conditions of unbiased compression to reduce the communication cost in distributed optimization. Specifically, the paper first presents a theoretical formalization of the total communication cost (TCC) in distributed optimization. With this formulation, the paper proves that unbiased compressor alone cannot necessarily save TCC. Then the paper proves lower bounds on the convergence complexity and shows that independent unbiased compressor provably saves TCC. The paper also improves ADIANA and provides the upper bound of the TCC that can be reduced. Experiments are also provided to support the theoretical findings.

Strengths

Originality: The paper focuses on how to save TCC, which is a more practical concern in distribution optimization. Though the paper relies on the findings (independence on unbiased compressors) of previous works, it provides new findings (on the upper bound of the TTC that can be reduced) and applicable modifications of existing methods to obtain an match the bound. Overall the paper seems novel to me. Quality: The claims in the paper are well supported by theoretical analysis. Clarity: The paper is well organized and easy to read. The authors give good introduction and clearly list the questions (Q1, Q2) they need to answer. They also provide detailed intuition on why independent unbiased compressor saves communications.

Weaknesses

Overall the paper looks good to me. One concern is that the current dataset is relatively small. It would be better if the author can provide experiments on a larger real dataset to better prove the effectiveness of the theoretical findings.

Questions

N/A

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

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

N/A

Reviewer gfiB5/10 · confidence 4/52023-07-29

Summary

The paper investigated the overall communication costs for federated learning algorithms. A lower bound on the per-round communication cost is presented, then an analysis of an existing algorithm, ADIANA, is provided to obtain an improved upper bound.

Strengths

The general ideas for the settings and the results are clearly presented. Besides, the presented bounds are clean compared to related works in this field.

Weaknesses

There is an absence of several crucial details, or perhaps typos, which makes it difficult to concretely understand the results. For example, one central concept required in this work is the unbiased compressor, which based on the current manuscript, applies to all real input values. However, if we require it to output a discrete variable with a bounded number of bins, such a compressor may not exist as it is not clear how the unbiased condition could hold for all $x\in\mathbb{R}$.

Questions

There are possible places of improvement in terms of clarity. For example: 1. Generally, it would be better if any concept is defined before it is formally used. For example, in the current manuscript, the definition of “linear spanning” is lost in the texts. 2. Similarly, the assumption of fixed communication load per round could be introduced before the definition of $T_{\epsilon}$ in Sec. 2.4, otherwise, the meaning of minimization in its definition is not clear. 3. For strongly convex results, it might be cleaner to replace $L/\mu$ with $\kappa$. 4. The notation of $U^{int}_{\omega}$ (or $U_\omega$) can be removed (or moved to appendices) if not used.

Rating

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

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

2 fair

Contribution

2 fair

Limitations

The limitation is appropriately discussed in Section 7.

Authorsrebuttal2023-08-14

Need more clarifications?

Dear reviewer gfiB, We thank you for your valuable comments. We have made detailed responses to address your concerns, but we have not received your replies to our current clarifications yet. We thus kindly ask if our responses have addressed all your concerns. If not, we are more than happy to provide more clarifications. Best, The authors of paper 5699

Reviewer gfiB2023-08-20

Consistency between unbiased compressor and fixed communication

We would like to thank the author for the response. My concern about the unbiased compressor remains, as the communication in this work is measured in bits (as far as I understand, as stated on line 208 or in Figure 1), but the well-known results in [25,26] deliver continuous variables, which require infinitely many bits to make it unbiased. I suppose the authors may be intended to apply an additional quantization step which would introduce some bias, but this is unclear from the manuscripts. Especially when the authors assume fixed and deterministic load per round (instead of variable length), the only class of compression function C that enables decodability has to map the input to a fixed finite set, and any input larger than the maximum in that set can not be mapped to a distribution with unbiased expectation. So for the results to be meaningful, a modification for at least some of the assumptions in the current manuscript is needed. Here are the possibilities. 1. The end-to-end compression function is not unbiased. 2. The communication load can be non-deterministic. 3. The compressor only applies to a subset of the real inputs. 4. There is a random seed shared between all machines and the randomness of C is taken over the shared value. 5. There is an allowed exceptional event with a small probability that the assumptions on the compressors can be violated. It would also be helpful if a concrete example of the unbiased estimator is presented that requires finitely many bits for communication.

Authorsrebuttal2023-08-21

Author response (Part 1/2)

We really appreciate the reviewer's deep thought and sharp observation on the concept of unbiased compressors. Below are our responses. **1. No finite-bit unbiased compressor over the entire real space.** We agree with the reviewer. With finite bits, a compressor cannot estimate an arbitrarily large value in an unbiased manner and thus cannot facilitate the compression for all real numbers/vectors. **2. It is a common issue in literature** The reviewer's sharp insights indeed expose a common issue existing in a large body of literature. For most literature on unbiased quantization (such as Q-SGD [3] and natural compression [14]), while the quantization schemes therein are defined over the entire real space, they only apply these schemes to values represented by float32 or float64. Apparently, quantizing arbitrarily large (or small) values in the entire real space will result in infinite bits, which contradicts the purpose of saving communication. Thus it makes more sense to quantize values originally represented with float32 or float64 to values with much fewer bits. Our work follows this convention. While we define the unbiased compressor over the entire real space, we apply it to input values represented with float32 or float64. Since the input values have already been represented with finite bits, the output values of our defined unbiased compressor are also with finite bits, which is consistent with our setting that "each communication round has fixed and deterministic load". This convention is also followed by other literature such as [R1], which was accepted by NeurIPS2022, to compare the amount of saved communicated bits in comparison to other baselines. [R1] Wang, B., Safaryan, M. and Richtárik, P., 2022. Theoretically better and numerically faster distributed optimization with smoothness-aware quantization techniques. Advances in Neural Information Processing Systems, 35, pp.9841-9852.

Authorsrebuttal2023-08-21

Author response (Part 2/2)

**3. A refined definition for unbiased compressor** While we are following the convention in literature, we are glad to provide a more precise definition for unbiased compressors stuided in the manuscript to clarify the reviewer's confusion. In fact, an unbiased compressor with finite bits exists if and only if the following conditions hold: - **The valid input value must be bounded.** It's clear that using finite bits cannot provide an unbiased estimate for arbitrarily large (or small) values. - **The valid non-zero input must be bounded away from zero.** Since the compression uses finite bits, there exists the smallest distance that bounds all possible input values away from 0. Given the above conditions, we will modify the classical definition of unbiased compressor as follows: **Assumption 2’** We assume unbiased compressors $\\{C_i\\}_{i=1}^n$ satisfy \begin{equation} \mathbb{E}[C_i(x)]=x,\quad\mathbb{E}[\\|C_i(x)-x\\|^2]\le\omega\\|x\\|^2, \end{equation} for a constant $\omega\ge0$ and any input $x\in\mathcal{X}\subset\mathbb{R}^d$, where $\mathcal{X}$ satisfies the following condition: **There exists positive constants $0<\epsilon_m<M$ so that it holds for all $ x=(x_1,\cdots,x_d)^\top\in\mathcal{X}$ that $|x_i|\in\\{0\\}\cup[\epsilon_m,M]$, $i=1,\cdots,d$.** We believe it is a proper definition because - **It is well justified in practice.** The inputs of compressors under machinery computations; therefore, their maximum value and non-zero lower bound are constrained by the computation precision of machines. For instance, when performing computations using float32, the resulting values are bounded above by approximately 3.4e38 and away from zero by at least about 1.18e-38 if non-zero. - **It is consistent with unbiased compressors proposed in previous works.** Natural compression [14] is a good unbiased compression scheme, following which we can provide a concrete example of unbiased compressors satisfying Assumption $2^\prime$: Letting $\mathcal{C}(0)=0$ and \begin{equation} \mathcal{C}(t)=\left\\{ \begin{array}{2} sign(t)\cdot2^{\lfloor\log_2|t|\rfloor}, &\text{with } p(t),\\\\ sign(t)\cdot2^{\lceil\log_2|t|\rceil}, &\text{with } 1-p(t), \end{array} \right. \end{equation} where $t\ne0$ and probability $p(t)=\frac{2^{\lceil\log_2|t|\rceil}-|t|}{2^{\lfloor\log_2|t|\rfloor}}$, it's clear that $d(2+\lceil\log_2(\lceil\log_2(M)\rceil-\lfloor\log_2(\epsilon_m)\rfloor+1)\rceil)$ bits are sufficient for constructing such a compressor that satisfies Assumption $2^\prime$ with $\omega=1/8$. - **It does not affect the theoretical results of our work.** -- **Upper bound**. Since the convergence criterion $\epsilon$ is typically much larger than the machine precision $\epsilon_m$ (e.g.,1.18e-38), the distortion incurred by finite bits has little effect on the ideal computations in the real space. In this case, all inputs of the compressors, which are calculated by machinery computations, naturally fall into the valid input set $\mathcal{X}$ we consider. As a result, machine precision $\epsilon_m$ does not affect our established upper bound for ADIANA. In fact, it is convection in optimization literature to ignore the effect of machine precision on the convergence rate and complexity. -- **Lower bound**. In our paper, we established the lower bound for $\inf_{A\in\mathcal{A}}\sup_{C_i \in \mathcal{U}\_\omega} T(A, \\{C_i\\})$ where $\mathcal{A}$ is the class of algorithms defined in Definition 1, $\mathcal{U}\_\omega$ is the class of unbiased compressor defined in Assumption 2, and $T(A, \\{C_i\\})$ is the convergence complexity achieved with algorithm $A$ and compressors $C_i$'s. When we consider the new class of unbiased compressor $\mathcal{U}\_\omega^\prime$ specified by Assumption $2^\prime$, it is easy to see that $\mathcal{U}\_\omega \subseteq \mathcal{U}\_\omega^\prime$. Let $T_L$ be the lower bound of $\inf_{A\in\mathcal{A}}\sup_{C_i \in \mathcal{U}\_\omega} T(A, \\{C_i\\})$, it naturally holds that \begin{equation} T_L \le\inf_{A\in\mathcal{A}}\sup_{C_i \in \mathcal{U}\_\omega} T(A, \\{C_i\\}) \le \inf_{A\in\mathcal{A}}\sup_{C_i \in \mathcal{U}'\_\omega} T(A, \\{C_i\\}) \end{equation} In other words, our established lower bound also holds for the new class of unbiased compressors defined by Assumption $2^\prime$. Since the upper bound and lower bound are not affected by the newly introduced compressor assumption up to machine precision $\epsilon_m$, they still nearly match each other. In other words, the modified family of unbiased compressors does not affect our theoretical results. **Summary** We really appreciate the sharp observation and useful suggestions proposed by the reviewer. It is actually a common issue in literature, and it can be practically addressed by considering finite-bit inputs only. We hope this can resolve the reviewer's concerns, and we are looking forward to the follow-up discussions.

Reviewer gfiB2023-08-22

This updated formulation looks good to me, and I'm willing to increase my score from 3 to 5, assuming that the rest of the paper is to be revised to make sure everything is consistent.

Reviewer yR982023-08-10

Respond

Thank you for the rebuttal! I didn't mean that you would start reproving everything for the "global smoothness" case. I only kindly asked you to add comments in the paper that the comparison between your results and the classical AGD method are only valid under the assumption that "that all local smoothness constants are constrained by a common upper bound." From the rebuttal, it is clear that we are on the same page! The authors addressed all my questions, and I will keep my score. BTW, I've recently found the paper \[1\]. They consider a slightly different setup, bidirectional communication. However, in their Contribution C, they say that they improve CANITA's $1 / \varepsilon^{1/3}$ rate to $\ln 1/ \varepsilon.$ In your paper, you have the gap between the upper bound and the lower bound ($1 / \varepsilon^{1/3}$ vs $\ln 1/ \varepsilon.$). It seems that this paper closes that gap, so your lower bound is tight. \[1\]: https://arxiv.org/pdf/2305.12379.pdf

Authorsrebuttal2023-08-11

We really appreciate your follow-up feedback. We'll definitely clarify the smoothness issue in the revision as we stated in our rebuttal. We are also happy to discuss the recent paper of 2Direction with the reviewer. As mentioned by their authors, the rate of 2Direction in the low accuracy regimes improves the $\Theta\left(\frac{1}{\epsilon^{1/3}}\right)$ term to $\Theta\left(\log\frac{1}{\epsilon}\right)$. However, when simplfiying to our setting where $r=0$, $L=L_{\max}$, $\alpha=1$, their dominant term in the high precision case (namely, $\epsilon$ is sufficiently small) $\left(1+\frac{\omega^{1/2}}{n^{1/6}}+\frac{\omega^{3/4}}{n^{1/4}}+\frac{\omega}{\sqrt{n}}\right)\frac{\sqrt{L\Delta}}{\sqrt{\epsilon}}$ is worse than $\left(1+\frac{\omega}{\sqrt{n}}\right)\epsilon^{-1/2}$, the major term in our lower bound. Therefore, despite of the inspiring work, it remains open to find an algorithm that, in the generally-convex case, can tightly match both the $\ln(1/\epsilon)$ term and the $1/\sqrt{\epsilon}$ term simultaneously. Finally, we thank the reviewer for pointing out this recent work (which came out after our submission) and we will also comment on this work in our later revisions.

Reviewer yR982023-08-11

Respond

I'm in no way trying to down the value of your contribution. Everything that you wrote is almost true. Indeed, they do not improve the high precision term of the original CANITA paper. BTW, using Young's inequality, note that $\frac{\omega^{1/2}}{n^{1/6}} \leq \frac{1}{3} 1^{3} + \frac{2}{3} \frac{\omega^{3/4}}{n^{1/4}}.$ so their complexity is $\Theta(1 + \frac{\omega^{3/4}}{n^{1/4}} + \frac{\omega}{\sqrt{n}})\frac{\sqrt{L \Delta}}{\sqrt{\varepsilon}}$ when $L = L_{\max}$ in the high precision case. Good luck with other rebuttals!

Reviewer u1Ro2023-08-15

Re

Thank you for the detailed response. The example and intuition provided for unbiased but dependent compressors makes sense to me and I think adding these points to the paper would definitely help readers understand why incorporating independence can lead to tighter bounds and lower communication cost in practice. I do not have any other questions or concerns and as I have already recommended acceptance, I will keep my score.

Authorsrebuttal2023-08-19

Thanks very much for the follow-up comments

We are very happy that your concerns have been addressed. Following your suggestions, we will add these examples and intuitions to the main paper in the revision. Thanks very much for your valuable feedback and comments.

Reviewer MRty2023-08-19

I have read the response and increased my score to 7.

Authorsrebuttal2023-08-19

Thanks very much for raising the score

We are very happy that your concerns have been addressed. Thanks very much for your valuable feedback and comments.

Authorsrebuttal2023-08-20

Can we have your response to our rebuttals?

Dear Reviewer JbF9, The reviewer-author discussion period will end **tomorrow**. Could you please let us know whether our rebuttal has resolved your concerns? If not, could you please point them out so that we can try to address them as best as we can? Thanks very much for your time and efforts to review our work.

Authorsrebuttal2023-08-21

A rebuttal summary before the reviewer-author discussion deadline

Dear Reviewers, We sincerely appreciate the time and effort you have devoted to reviewing our paper. Your insightful feedback has significantly improved the quality of our work. We are very delighted that reviewers find that our work "**novel**" and "**provide new findings**" on the lower bound, and establish results that "**match with**" the lower bound, that our claims are "**convincing**" and "**well-supported**" by theoretical analysis, and that our results are "**important and impactful addition**" to the community. We have taken all questions and concerns seriously, providing comprehensive rebuttals to each reviewer. We are glad that the issues raised by reviewers oD7w, tEWM, u1Ro, MRty, and yR98 have been sufficiently addressed, and we will integrate their feedback into the revision. We only had the chance to communicate with reviewer gfiB in the last day before the deadline. Our definition of unbiased compressor follows standard convention in literature, i.e., define the compressor over the entire real space but apply it to values represented with float32 or float64. However, to fully resolve reviewer gfiB's concern, we do provide an improved definition of unbiased compressor which is practical, applicable to many compressors established in literature, and does not affect our theoretical results. We hope this can resolve his or her concerns on the concept of unbiased compressor. While we attempted follow-up with reviewer JbF9, we unfortunately have not received any response. As detailed in our global response point 2, we have provided ample existing work exploiting correlated unbiased compressors. The other concerns are also thoroughly addressed in our response. We sincerely thank all reviewers for their thoughtful comments and questions. Their feedback has been invaluable for strengthening our work. Best regards, Submission 5699 Authors

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC