acceptodds
Under review as a conference paper at ICLR 2027

Dimension-Free Decentralized Nonsmooth Nonconvex Stochastic Optimization

12 views

Abstract

We investigate decentralized nonsmooth nonconvex stochastic optimization over a network of nodes, with the goal of finding an -Goldstein stationary point. The best existing algorithm achieves sample complexity and communication complexity, where is the problem dimension and is the spectral gap of the communication matrix. However, the polynomial dependence on can be a major bottleneck in high-dimensional regimes. In this paper, we propose a novel algorithm that achieves sample complexity and communication complexity. The primary technique is an elegant decentralized online-to-nonconvex conversion that reduces the original problem to a decentralized online convex optimization (D-OCO) problem. A key property of our conversion is that its consensus requirements can be inherited directly from the consensus of the underlying D-OCO decisions. In particular, this property enables us to establish an explicit connection between the dimension dependence and the consensus error, which in turn shows that the polynomial dependence on can be removed with only logarithmic additional communication.

Primary area:
optimization
Keywords:
Decentralized Optimization, Nonconvex Stochastic Optimization, Online-to-nonconvex Conversion

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_JD7UOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper considers the decentralized nonconvex nonsmooth optimization problem. The existing SOTA method achieves a linear dependence on the dimension dd in the sample complexity. By leveraging the shared random interpolation weight across nodes, the paper achieves a sample complexity that is independent of the dimension dd.

Critical strengths
  1. The proposed method achieves a dimension-free sample complexity for solving the decentralized nonconvex nonsmooth problem.

  2. The authors provided a theoretical convergence analysis of the proposed method, and it is correct and sound.

  3. The overall paper writing is clear and easy to follow.

Critical weaknesses
  1. Although the authors claimed they used a shared interpolated weight induced by a shared seed, they only used the shared weight to compute the gradient on a single machine, while all other machines were idle. In this case, why not just running the exact O2N algorithm on a single machine, and distributes all the final weights to all the machine.

  2. In Algorithm 1, the algorithm computes the weight averaged over all machines and all iterations, which is quite impractical for large-scale distributed system.

  3. No numerical experiments are performed to compare the proposed method with existing methods, such as the DOC$^2$S method.

Questions for authors
  1. Would the DOC$^2$S achieves the same rate as your method if they also adopted the shared random seed methodology?

  2. Is it possible to obtain a per-machine average for the output instead of a network-average output?

  3. Is it possible to extend the current methodology to achieve a linear speedup?

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Many thanks for the constructive review! We hope the reviewer could reevaluate our paper, and are very happy to respond to any additional questions during the discussion period.


Q1: ... they only ... compute the gradient on a single machine, while all other machines were idle ... why not just running the exact O2N algorithm on a single machine, and distributes all the final weights ...

A1: First, although only one machine queries its stochastic gradient at each iteration, we want to clarify that the other machines are not idle: all nodes maintain their local optimization variables and participate in the D-OCO updates and neighbor-to-neighbor communication.

Second, running O2N on a single fixed machine using only its local oracle would optimize the local objective fif_i, rather than the global objective f=1n∑ifif=\frac1n\sum_i f_i. Broadcasting the final solution would not resolve this objective mismatch. Therefore, running O2N on a single machine and broadcasting its output cannot directly replace our decentralized procedure.


Q2: ... the algorithm computes the weight averaged over all machines and all iterations ... Is it possible to obtain a per-machine average for the output ...

A2: Thank you for pointing out this issue and providing the helpful suggestion. We have revisited our analysis and proved that each node can return its own local average instead of the network-averaged output, with the same stationarity guarantee.

Specifically, node ii computes wik=1T∑t=1Twt,ik\mathbf{w}_i^k=\frac{1}{T}\sum_{t=1}^T \mathbf{w}_{t,i}^k and returns wi∼Unif⁡wi1,…,wiK\mathbf{w}_i\sim\operatorname{Unif}\\{\mathbf{w}_i^1,\ldots,\mathbf{w}_i^K\\}. The key insight is that the local points w1,ik,…,wT,ik\mathbf{w}_{1,i}^k,\ldots,\mathbf{w}_{T,i}^k of each node ii lie within distance TD=δ/2TD=\delta/2 of their average wik\mathbf{w}_i^k, allowing us to construct a Goldstein stationarity certificate directly for the local output. Moreover, our existing consensus guarantee ensures that replacing wˉtk\bar{\mathbf{w}}_t^k with wt,ik\mathbf{w}_{t,i}^k as the evaluation point of the smoothed global gradient increases the gradient discrepancy bound by at most a factor of 2.

Consequently, only the last term in Theorem 1 increases by a factor of 2. The original parameter choices in Corollary 1 already accommodate this increase, so both the parameter settings and the resulting sample and communication complexities remain unchanged. We have revised Algorithm 1, Theorem 1, Corollary 1, and the corresponding proofs accordingly.


Q3: No numerical experiments are performed ...

A3: We have added experiments on two nonsmooth nonconvex tasks: capped-ℓ1\ell_1-regularized SVM training on a9a and rcv1, and two-layer ReLU network training on MNIST and Fashion-MNIST. Across all four datasets, D-ONC reduces the global objective faster than DOC$^2$S and ME-DOL as either the sample or communication budget increases, supporting its practical effectiveness. The results are presented in the new “Preliminary Experiments” section, with detailed task formulations and parameter settings in the appendix.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)


Q4: Would the DOC$^2$S achieves the same rate as your method if they also adopted the shared random seed methodology?

A4: Yes. Actually, from Reviewer eDwY’s comment, we became aware that the recent revision of Chen et al. (2026, arXiv:2601.19381v2) also uses shared random interpolation and achieves sample and communication complexities matching those of our work. However, this revision was released on Oct 5, 2026, after our submission deadline of September 25, 2026. We therefore respectfully suggest that the subsequent improvements in v2 should not detract from our independently developed contributions.

Moreover, we want to emphasize that there are important distinctions in our algorithm and analysis.

  1. Our analysis bypasses a key technical challenge identified by Sahinoglu & Shahrampour (2024) and also addressed by Chen et al. (2026). Specifically, their analysis of the global objective change f(xˉtk)−f(xˉt−1k)f(\bar{\mathbf x}_t^k)-f(\bar{\mathbf x}_{t-1}^k) introduces the following discrepancy -\frac1n\sum_{i=1}^n\int_0^1 \nabla f_i\\!\left(\mathbf x_{t-1,i}^k+s\Delta_{t,i}^k\right)\\,ds,$$ which they bound using the smoothness of the local functions. We instead start from local objective changes (our Eq. (11)) and bound the same global objective change using only function Lipschitzness and iterate consensus.
  2. We establish a black-box reduction to D-OCO by reducing all three decentralization challenges to update-direction consensus. All gossip operations are handled by the D-OCO subroutine, and our Theorem 1 holds for any valid upper bound CT,K\mathcal C_{T,K} on its expected decision-consensus error. In contrast, Chen et al. (2026) separately apply accelerated gossip to both update directions and optimization iterates. Their Goldstein certificate construction also requires a radius-dependent consensus threshold to ensure neighborhood inclusion, whereas our reduction accommodates general consensus bounds.

Q5: Is it possible to extend the current methodology to achieve a linear speedup?

A5: At present, we do not have a theoretical guarantee of linear speedup. Specifically, a natural starting point is to replace client sampling with full participation, which allows us to average gradient estimates across nodes and thereby reduce their stochastic variance in the analysis. Although full participation is accommodated by our framework, the current D-OCO regret bound depends on the second moment of the gradient estimates, rather than their variance alone. Therefore, this variance reduction does not directly yield linear speedup. This limitation also arises in the full-participation analysis of Sahinoglu & Shahrampour (2024). Accordingly, establishing linear speedup may require a sharper analysis of the D-OCO regret, and we regard this as an important direction for future work.

I thank the authors for their careful clarification, and they have addressed most of my concerns. I have raised my score by 1.

Dear Reviewer JD7U,

Thank you for your reassessment and positive update.

Best,

Authors

Reviewer_Z6JNOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a decentralized online-to-nonconvex conversion (D-ONC) for finding (δ,ε)(\delta,\varepsilon)-Goldstein stationary points. It uses a shared random interpolation weight stks_t^k across nodes, client sampling, and randomized smoothing in the analysis. A general reduction (Thm 1) bounds stationarity in terms of D-OCO regret and consensus error, with the dimension entering only through a d CT,K\sqrt{d}\,C_{T,K} term. Instantiating it with AD-FTRL (Wan et al., 2024) without the blocking mechanism drives the consensus error to ζD\zeta D at a logarithmic communication cost. The result is O(δ−1ε−3)O(\delta^{-1}\varepsilon^{-3}) samples and O~(γ−1/2δ−1ε−3)\tilde O(\gamma^{-1/2}\delta^{-1}\varepsilon^{-3}) communication rounds, removing the dε−1d\varepsilon^{-1} term that the authors attribute to DOC2S. There are no experiments.

Critical strengths

(1) To my knowledge, this is the first decentralized first-order result for nonsmooth nonconvex stochastic optimization without polynomial dependence on dd, and its sample complexity matches the centralized lower bound of Cutkosky et al. (2023).

(2) The reduction (Theorem 1) is general: it applies to any D-OCO algorithm with the required regret and consensus guarantees. It also explains why, according to the authors, prior conversions (ME-DOL, DOC2S) retained a d\sqrt d factor. With independently drawn interpolation weights, the disagreement of the interpolation points contains a term that does not shrink with the D-OCO consensus error, so the d\sqrt d smoothness factor cannot be cancelled by additional communication; a shared interpolation weight removes this term. The proofs are short and easy to verify.

Critical weaknesses

(1) The paper would be considerably stronger with lower bounds, and it relies almost entirely on existing tools. The paper provides no lower bound for communication complexity, and the conclusion itself leaves near-optimality open. A matching lower bound for decentralized nonsmooth nonconvex optimization, showing whether O~(γ−1/2δ−1ε−3)\tilde O(\gamma^{-1/2}\delta^{-1}\varepsilon^{-3}) communication rounds are necessary, would turn this work from an improvement of a prior upper bound into a complete characterization of the problem, and would substantially raise its significance. The upper bound itself is obtained by assembling existing components: the O2NC conversion of Cutkosky et al., randomized smoothing and the Goldstein-subdifferential inclusion of Kornowski & Shamir, client sampling and the comparison framework of DOC2S, and AD-FTRL from Wan et al. with the blocking mechanism removed. The extension of AD-FTRL from bounded to second-moment-bounded gradients (Appendix D) is routine. The key new ingredient is a common random interpolation weight; once this observation is made, the rest of the analysis follows standard arguments. The improvement over DOC2S is also limited in scope. The sample complexity improves from δ−1(ε−3+dε−1)\delta^{-1}(\varepsilon^{-3}+d\varepsilon^{-1}) (as stated for DOC2S in the paper) to δ−1ε−3\delta^{-1}\varepsilon^{-3}, which is strict only when d≳ε−2d\gtrsim\varepsilon^{-2}, and the communication complexity has the same dependence on γ\gamma and ε\varepsilon as DOC2S. Overall, the paper is clear and, as far as I checked, correct, and it closes a relevant gap; however, the contribution is a focused upper-bound improvement obtained without new techniques, and lower bounds would considerably strengthen it.

(2) No mechanism corrects the accumulated disagreement, and decentralization brings no speedup. The local iterates are never mixed: their consensus is only inherited from the D-OCO decisions. By Lemma 3, the disagreement is not contracted but accumulates linearly over the whole horizon, E∥xt,ik−xˉtk∥≤((k−1)T+t) CT,K\mathbb{E}\Vert x^k_{t,i}-\bar x^k_t\Vert\le((k-1)T+t)\,C_{T,K}. This is what forces the consensus error to be extremely small, ζ≲ε/(LKd)\zeta\lesssim\varepsilon/(LK\sqrt d), and it means that the method has no mechanism to correct any deviation from exact synchronized execution: desynchronized random streams for stks_t^k and itki_t^k, dropped messages or numerical errors accumulate without correction. In addition, because of client sampling (inherited from DOC2S), exactly one node computes a stochastic gradient per iteration, so there is no speedup in the number of nodes.

Questions for authors

(a) Is there a lower bound on the communication complexity for this problem class? Is O~(γ−1/2δ−1ε−3)\tilde O(\gamma^{-1/2}\delta^{-1}\varepsilon^{-3}) communication necessary? Even a lower bound for a restricted class of algorithms would considerably strengthen the paper.

(b) Beyond the shared interpolation weight, which component of the analysis is technically new relative to DOC2S and Wan et al. (2024)? Since the improvement in sample complexity appears only when d≳ε−2d\gtrsim\varepsilon^{-2}, could you also discuss this regime explicitly?

(c) Since the iterates are never mixed and their disagreement accumulates linearly (Lemma 3), how robust is the method to deviations from exact synchronized execution (e.g., desynchronized random streams or dropped messages)? Could an explicit correction step, such as occasional gossip on the iterates, be added without losing the dimension-free guarantee?

3 comments

Many thanks for the constructive review!


Q1: Lower bound for communication complexity.

A1: Thank you for the helpful suggestion. However, we currently do not have a matching lower bound for communication complexity. Establishing one would require a hard instance that jointly captures the difficulty of finding a Goldstein stationary point for a nonsmooth nonconvex objective and the information-propagation constraints of a decentralized network. It is highly non-trivial to combine these two sources of difficulty, because of the interaction between communication, local computation, and stochastic oracle queries. Thus, we leave such a lower bound as an important open question.


Q2: Technical novelty beyond the shared interpolation weight.

A2: Thank you for recognizing the technical contribution of our shared random interpolation. Beyond this ingredient, our approach differs from DOC$^2$S (Chen et al., 2026) in both analysis and algorithm design.

  1. Our analysis bypasses a key technical challenge identified by Sahinoglu & Shahrampour (2024) and also addressed by Chen et al. (2026). Their analysis of the global objective change f(xˉtk)−f(xˉt−1k)f(\bar{\mathbf x}_t^k)-f(\bar{\mathbf x}_{t-1}^k) introduces the following discrepancy -\frac1n\sum_{i=1}^n\int_0^1 \nabla f_i\\!\left(\mathbf x_{t-1,i}^k+s\Delta_{t,i}^k\right)\\,ds,$$ which they bound using the smoothness of the local functions. We instead start from local objective changes (our Eq. (11)) and bound the same global objective change using only function Lipschitzness and iterate consensus.
  2. We establish a black-box reduction to D-OCO by reducing all three decentralization challenges to update-direction consensus. All gossip operations are handled by the D-OCO subroutine, and our Theorem 1 holds for any valid upper bound CT,K\mathcal C_{T,K} on its expected decision-consensus error. In contrast, Chen et al. (2026) separately apply accelerated gossip to both update directions and optimization iterates. Their Goldstein certificate construction also requires a radius-dependent consensus threshold to ensure neighborhood inclusion, whereas our reduction accommodates general consensus bounds.

Moreover, regarding Wan et al. (2024), we actually do not claim a new D-OCO technique. Our contribution is the reduction itself, which connects D-OCO with decentralized nonsmooth nonconvex optimization through regret and consensus guarantees. Its value therefore extends beyond a particular algorithm: advances in D-OCO can be incorporated by substituting the corresponding bounds into our theorem.


Q3: Explicit discussion on the improvement regime.

A3: Thank you for the helpful suggestion. We have added an explicit discussion in the remark following Corollary 1.


Q4: No correct mechanism for the accumulated disagreement.

A4: First, we want to clarify that not mixing the local iterates is a deliberate design choice that enables a clean black-box reduction to D-OCO. As you noted, although iterate disagreement appears to accumulate, it can be controlled through sufficiently accurate consensus of the D-OCO update directions.

Second, although we agree that deviations from exact synchronized execution are a practical concern, they fall outside the scope of this work. Extending our guarantees to these settings would require further algorithmic development and analysis. Thus, we cannot currently provide a definitive robustness guarantee.

Third, an explicit correction step, such as occasional gossip on the iterates, can be incorporated into our framework. Actually, this would bring our algorithm closer to those of Sahinoglu & Shahrampour (2024) and Chen et al. (2026), which have already applied additional (accelerated) gossip to the iterates. From their analyses, such a correction step is not the source of dimension dependence. Our analysis can likewise accommodate this correction while preserving the dimension-free guarantee.


Q5: No speedup in the number of nodes.

A5: We agree that our current guarantees do not establish a speedup in the number of nodes. However, this limitation is not solely due to client sampling. Actually, our framework also accommodates full participation, allowing us to exploit the variance reduction from averaging gradient estimates across nodes in the analysis. Nevertheless, the current D-OCO regret bound depends on the second moment of the gradient estimates, rather than their variance alone, so this variance reduction does not directly yield speedup. This limitation also arises in the full-participation analysis of Sahinoglu & Shahrampour (2024). Accordingly, establishing speedup may require a sharper analysis of the D-OCO regret, and we regard this as an important direction for future work.

I thank the authors for the detailed response and the revised version.

The added remark after Corollary 1 clearly states the high-dimensional regime in which the bounds improve over DOC2S, which addresses my question. The clarification of the technical differences from DOC2S, in particular bounding the global objective change using only Lipschitzness and iterate consensus rather than smoothness of the local functions, is helpful and makes the contribution clearer. I also appreciate the honest answers on the lower bound for communication complexity, on the absence of speedup in the number of nodes, and on the robustness to deviations from synchronized execution; I agree that these are natural directions for future work rather than shortcomings of this paper.

Overall, the paper makes a clear and self-contained contribution: it closes a known gap by giving the first dimension-free guarantee with optimal sample complexity, through a clean and general reduction that can directly benefit from future advances in D-OCO. The main reason I do not rate it higher is that the result is obtained mostly by combining existing techniques, so its technical depth is moderate. Taking both aspects into account, I believe that a rating of 3 is the most appropriate for this work, and I support its acceptance.

Dear Reviewer Z6JN,

Thank you for your prompt response and support for our paper.

Best,

Authors

Reviewer_eDwYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies decentralized nonsmooth nonconvex stochastic optimization, aiming to find an (δ,ϵ)(\delta,\epsilon)-Goldstein stationary point. The authors propose a decentralized online-to-nonconvex conversion (D-ONC) that reduces the original problem to decentralized online convex optimization (D-OCO). The key idea is to use a shared random interpolation weight across nodes, allowing the consensus error of the interpolation points to be controlled by that of the D-OCO decisions. By combining this conversion with accelerated decentralized FTRL and client sampling, the proposed algorithm achieves O(δ−1ϵ−3)O(\delta^{-1}\epsilon^{-3}) sample complexity and O~(γ−1/2δ−1ϵ−3)\widetilde O(\gamma^{-1/2}\delta^{-1}\epsilon^{-3}) communication complexity, eliminating the polynomial dependence on the problem dimension.

Critical strengths
  1. The paper establishes sample and communication complexity bounds without polynomial dependence on the problem dimension. The general reduction in Theorem 1 explicitly relates the dimension-dependent term to the consensus error of the underlying D-OCO algorithm, providing useful insight into how dimension dependence can be controlled through communication.

  2. The use of a shared random interpolation weight is simple but effective. It allows the consensus errors of the local iterates and interpolation points to be controlled through the underlying D-OCO decisions. The theoretical development is generally clear and well organized.

Critical weaknesses
  1. The paper is entirely theoretical and does not provide numerical experiments to validate the proposed method. Although the theoretical complexity improvements are interesting, experiments comparing the proposed algorithm with existing methods, particularly under different problem dimensions and network connectivity conditions, would help demonstrate the practical benefits of the dimension-free guarantees.

  2. The decentralized implementation of the final output is unclear. Algorithm 1 returns a global average of the local interpolation points, which is not directly available to individual nodes in a decentralized network. The paper does not explicitly explain how this output is computed using only neighbor-to-neighbor communication. Additional consensus steps may be required, and their effect on both the stationarity guarantee and communication complexity should be addressed.

Questions for authors

The recent revision of Chen et al. (2026, arXiv:2601.19381v2) also reports dimension-free sample complexity and nearly dimension-free communication complexity, and employs shared random interpolation across nodes. Could the authors clarify the main differences between the two approaches, particularly in terms of algorithm design, key proof techniques, and theoretical guarantees? Specifically, what are the main technical distinctions between the proposed D-OCO conversion and the analysis in Chen et al.?

4 comments

Many thanks for the constructive review!


Q1: ... does not provide numerical experiments ... would help demonstrate the practical benefits ...

A1: Thank you for the helpful suggestion. We have added experiments on two nonsmooth nonconvex tasks: capped-ℓ1\ell_1-regularized SVM training on a9a and rcv1, and two-layer ReLU network training on MNIST and Fashion-MNIST. Across all four datasets, D-ONC reduces the global objective faster than DOC$^2$S and ME-DOL as either the sample or communication budget increases, supporting its practical effectiveness. The results are presented in the new “Preliminary Experiments” section, with detailed task formulations and parameter settings in the appendix.


Q2: ... decentralized implementation of the final output is unclear ...

A2: Thank you for pointing out this issue. We agree that the availability of the final output at individual nodes should be explicitly addressed. To this end, we have revisited our analysis and established a stronger result: each node can directly return a locally computable output with the same stationarity guarantee, without any additional communication.

Specifically, node ii computes wik=1T∑t=1Twt,ik\mathbf{w}_i^k=\frac{1}{T}\sum_{t=1}^T \mathbf{w}_{t,i}^k and returns wi∼Unif⁡wi1,…,wiK\mathbf{w}_i\sim\operatorname{Unif}\\{\mathbf{w}_i^1,\ldots,\mathbf{w}_i^K\\}. The key insight is that the local points w1,ik,…,wT,ik\mathbf{w}_{1,i}^k,\ldots,\mathbf{w}_{T,i}^k of each node ii lie within distance TD=δ/2TD=\delta/2 of their average wik\mathbf{w}_i^k, allowing us to construct a Goldstein stationarity certificate directly for the local output. Moreover, our existing consensus guarantee ensures that replacing wˉtk\bar{\mathbf{w}}_t^k with wt,ik\mathbf{w}_{t,i}^k as the evaluation point of the smoothed global gradient increases the gradient discrepancy bound by at most a factor of 2.

Consequently, only the last term in Theorem 1 increases by a factor of 2. The original parameter choices in Corollary 1 already accommodate this increase, so both the parameter settings and the resulting sample and communication complexities remain unchanged. We have revised Algorithm 1, Theorem 1, Corollary 1, and the corresponding proofs accordingly.


Q3: The recent revision of Chen et al. (2026, arXiv:2601.19381v2) also ... employs shared random interpolation across nodes. Could the authors clarify the main differences between the two approaches ... what are the main technical distinctions ...

A3: Thank you for pointing out Chen et al. (2026, arXiv:2601.19381v2). We agree that this version also uses shared random interpolation and achieves sample and communication complexities matching those of our work. However, this revision was released on Oct 5, 2026, after our submission deadline of September 25, 2026. We therefore respectfully suggest that the subsequent improvements in v2 should not detract from our independently developed contributions.

Moreover, we want to emphasize that there are important distinctions in our algorithm and analysis.

  1. Our analysis bypasses a key technical challenge identified by Sahinoglu & Shahrampour (2024) and also addressed by Chen et al. (2026). Specifically, their analysis of the global objective change f(xˉtk)−f(xˉt−1k)f(\bar{\mathbf x}_t^k)-f(\bar{\mathbf x}_{t-1}^k) introduces the following discrepancy -\frac1n\sum_{i=1}^n\int_0^1 \nabla f_i\\!\left(\mathbf x_{t-1,i}^k+s\Delta_{t,i}^k\right)\\,ds,$$ which they bound using the smoothness of the local functions. We instead start from local objective changes (our Eq. (11)) and bound the same global objective change using only function Lipschitzness and iterate consensus.
  2. We establish a black-box reduction to D-OCO by reducing all three decentralization challenges to update-direction consensus. All gossip operations are handled by the D-OCO subroutine, and our Theorem 1 holds for any valid upper bound CT,K\mathcal C_{T,K} on its expected decision-consensus error. In contrast, Chen et al. (2026) separately apply accelerated gossip to both update directions and optimization iterates. Their Goldstein certificate construction also requires a radius-dependent consensus threshold to ensure neighborhood inclusion, whereas our reduction accommodates general consensus bounds.
AuthorsOpenReview ↗

To Reviewer eDwY

Dear Reviewer eDwY,

Thank you again for your time and thoughtful review of our paper.

We would be grateful if you could take another look at our responses and the revised manuscript, and let us know whether we have adequately addressed your main concerns. We would be happy to clarify any remaining questions.

Best,

Authors

Thank you for your rebuttal My main concerns have been addressed and I maintain my current score.

Dear Reviewer eDwY,

Thank you for your response and for confirming that your main concerns have been addressed. We appreciate your time and consideration.

Best,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

To All Reviewers

We sincerely thank all reviewers for their constructive comments and questions, which have helped us improve the paper. We have responded to each reviewer’s comments point by point in the individual responses, and submitted a revised manuscript incorporating the corresponding changes. All changes are marked in blue in the revised version.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.