acceptodds
Under review as a conference paper at ICLR 2027

Decentralized Machine Unlearning under Client Disconnections

Abstract

Decentralized machine unlearning seeks to remove the influence of a client without centralized coordination. A distinct challenge arises when a *retained client* becomes unavailable during unlearning, as its learned contribution must remain part of the desired model. We propose a collaborative decentralized unlearning framework that addresses this setting by reconstructing the counterfactual learning dynamics. For active retained clients, we adapt trajectory-based Taylor–L-BFGS reconstruction to reduce direct gradient evaluations. For a disconnected retained client, its former neighbors maintain distributed *ghost replicas* that reconstruct both its missing local updates and its participation in decentralized mixing using finite-step consensus over the active network. We analyze the resulting algorithm from an optimization-dynamics perspective, showing that its deviation from counterfactual retraining decomposes into gradient-reconstruction and ghost-reconstruction errors, with the latter decreasing with the gossip budget and network mixing quality. Experiments show that the proposed method removes the contribution targeted for unlearning while preserving that of an unavailable client, closely approaching oracle retraining.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Decentralized learning, Machine unlearning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_cnw1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies decentralized client unlearning when a different, retained client disconnects during unlearning. It combines trajectory-based Taylor–L-BFGS reconstruction with distributed ghost replicas to approximate the unavailable client’s updates and mixing operations. The analysis bounds deviation from counterfactual retraining, while experiments evaluate utility, backdoor removal, and consensus.

Critical strengths
  1. The paper identifies a meaningful distinction between forgetting a client and losing access to a retained client. Retraining only over active clients would remove information that should be preserved.
  2. Ghost replicas reconstruct both local updates and mixing participation. The no-ghost ablation shows a clear benefit when the disconnected client holds unique-label informatio
Critical weaknesses
  1. The tracking theorem assumes bounded gradient-reconstruction errors without establishing their size for the proposed frozen L-BFGS mechanism. The gradient-magnitude test does not ensure approximation accuracy, and increasing gossip cannot correct this error. Thus, the theorem does not establish when the method closely approximates retraining.
  2. The forgetting evaluation relies mainly on backdoor ASR. Low ASR demonstrates removal of the injected behavior, but does not establish removal of the forgotten client’s broader influence. Accuracy and consensus also provide limited evidence of closeness to the oracle model.
Questions for authors
  1. Can you quantify the ghost-gradient reconstruction error after disconnection, either theoretically or through measurements against an accessible oracle?
  2. Does the method match oracle retraining in prediction discrepancy and membership-inference results on the forgotten client’s clean data, beyond backdoor ASR?
1 comment

Thank you for the constructive comments. We address both concerns below, supported by additional experiments included in the revised appendix E.2.

W1/Q1 – Ghost-gradient reconstruction accuracy. Our theoretical analysis was designed to characterize how gradient-reconstruction and ghost-consensus errors affect the deviation from counterfactual retraining. Accordingly, Theorem 1 provides a conditional tracking bound, without assuming a specific accuracy for the L-BFGS approximation. We agree that the gradient-magnitude safeguard does not guarantee reconstruction accuracy. To further substantiate the effectiveness of the proposed mechanism, we measured the cosine similarity between reconstructed ghost gradients and exact gradients evaluated at the same iterates, obtaining an average of 0.714 (range: 0.577–0.806). We also measured the final relative model-weight error with respect to oracle retraining:

MethodRelative model-weight error (%)
Ours19.88
No Ghost35.01

Thus, ghost reconstruction reduces the deviation from oracle retraining by 43.2% compared with the No-Ghost ablation. These results provide empirical support for the reconstruction mechanism, while the theoretical guarantee remains conditional on the approximation errors.

W2/Q2 – Unlearning effectiveness and closeness to oracle retraining. Our original evaluation used backdoor removal to assess forgetting, since the injected behavior provides a measurable signature of the forgotten client's influence. We agree that this alone does not fully characterize unlearning effectiveness. We therefore extended our evaluation to include membership inference attacks based on LiRA (Carlini et al. 2022) on the forgotten client's clean data (see Appendix E.2.3):

MethodLiRA AUC
Original model0.6560
Ours0.4847
No Ghost0.4874
Retraining w. Departure0.5014
Oracle retraining0.4965

Our method reduces the attack AUC to approximately 0.5, closely matching oracle retraining. Furthermore, to complement the original accuracy and consensus results, we evaluated prediction similarity to the oracle:

MethodFull test: Agreement (%)Full test: TVDExclusive label: Agreement (%)Exclusive label: TVD
Ours99.270.007595.290.0452
No Ghost92.970.070613.260.8275
Retraining w. Departure93.120.067814.400.8176

These results demonstrate close agreement with oracle predictions, particularly on the disconnected client's exclusive-label data, where preserving its contribution is essential. Together with the original backdoor-removal experiments, they provide stronger empirical evidence of effective forgetting while preserving the contribution of the disconnected retained client.

References:

Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramèr. Membership inference attacks from first principles. In IEEE Symposium on Security and Privacy (SP), pp. 1897–1914, 2022.

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

This work studies decentralized client-level unlearning when a different retained client becomes unavailable during the unlearning procedure. It adapts Taylor–L-BFGS trajectory reconstruction for active clients and introduces distributed ghost replicas and gossip to approximate the missing retained client’s gradients and mixing updates. Theoretical tracking bounds and experiments on MNIST/CIFAR-10 compare the method with decentralized unlearning baselines and retraining references.

Critical strengths
  1. The paper investigates a relatively new decentralized unlearning scenario. It explicitly distinguishes between a client requesting unlearning and a retained client that becomes disconnected due to device failures or network unavailability. These two types of clients have different requirements regarding knowledge removal and preservation, making the problem worth investigating.
  2. The proposed algorithm has certain technical novelty. The authors not only reconstruct the gradients of disconnected clients but also use distributed consensus to recover the model aggregation operations that these clients would have performed. This provides a more complete reconstruction mechanism than simply approximating missing gradients.
  3. The experimental evaluation is relatively comprehensive. The authors compare their method with existing decentralized unlearning approaches and investigate different data and network settings. The implementation is also publicly available, which supports reproducibility.
Critical weaknesses
  1. Restrictive theoretical assumptions. The theoretical analysis relies on relatively strong assumptions, including strong convexity and bounded gradient-reconstruction errors. Even without establishing new theoretical results, the authors should discuss how the conclusions might change under weaker conditions, such as general convexity or the Polyak–Łojasiewicz (PL) condition.
  2. Insufficient theoretical guarantees for unlearning effectiveness. Theorem 1 only establishes an upper bound on the deviation between the reconstructed model trajectory and counterfactual retraining. While this is useful for analyzing reconstruction accuracy, it does not establish the distributional indistinguishability commonly required in certified unlearning. Since the objective of machine unlearning is not merely to maintain model performance but also to remove the influence of requested data, the current theoretical guarantee is insufficient to fully support the unlearning claim.
  3. Insufficient motivation for the proposed setting. The paper introduces a new decentralized unlearning scenario in which a disconnected retained client's contribution must be preserved. However, it does not sufficiently explain why existing decentralized unlearning methods cannot accommodate this setting. In particular, the paper should better distinguish its problem formulation from existing certified decentralized unlearning methods that already consider dynamic network topologies. The necessity and practical significance of the proposed setting require further clarification.
  4. Presentation quality could be improved. Figure 1 appears blurry and is difficult to read. Since the figure is relatively simple, the authors are encouraged to replace it with a higher-quality vector graphic.
Questions for authors
  1. How can the authors further validate the effectiveness of unlearning? My main concern is that the theoretical analysis only provides a trajectory-reconstruction error bound rather than a certified-unlearning guarantee. In machine unlearning, effectiveness can be supported by theoretical indistinguishability guarantees or evaluated empirically using privacy attacks and other forgetting metrics. For example, PDUDT includes membership inference attack (MIA) experiments despite providing theoretical guarantees. However, the current paper does not report MIA results or consider more recent evaluation approaches, such as Unlearning Evaluation through Subset Statistical Independence (ICLR 2026). Although the backdoor attack experiments provide useful evidence, they mainly evaluate the removal of a specific injected behavior and do not fully characterize the remaining influence of deleted data. Could the authors provide MIA-based evaluations or other suitable empirical forgetting metrics to further substantiate the unlearning effectiveness?
  2. Why cannot existing decentralized unlearning methods address the proposed disconnection scenario? For example, PDUDT considers dynamic network topologies, including changes in communication graphs across rounds. Could such a framework be adapted to handle client disconnections? More importantly, what fundamental challenges arise when preserving the contribution of an unavailable retained client that cannot be addressed by existing dynamic-topology methods? While the proposed reconstruction mechanism is reasonable, the motivation and distinction from previous work should be more clearly established.
  3. How sensitive are the theoretical results to the assumptions? In particular, can the authors discuss whether the counterfactual tracking bound and the ghost-reconstruction analysis can be extended to general convex objectives or objectives satisfying the PL condition? If these extensions are not straightforward, which parts of the analysis would no longer hold?
3 comments

Thank you for the constructive comments. We address both concerns below, supported by additional experiments included in the revised appendix E.2

W1/Q3 – Theoretical assumptions. Our theoretical contribution focuses on analyzing the learning dynamics, characterizing how closely the unlearning trajectory tracks counterfactual retraining and how reconstruction errors affect this correspondence. Such analyses commonly impose regularity assumptions to derive informative convergence bounds, as in decentralized optimization (Yuan et al., 2016) and inexact gradient methods (d'Aspremont, 2008). Notably, the closely related FedRecover framework (Cao et al., 2023) adopts similar assumptions to analyze reconstruction accuracy while demonstrating empirical effectiveness on nonconvex neural networks. In our case, these assumptions enable geometric contraction and isolate the effects of gradient reconstruction and ghost consensus. Under general convexity, geometric contraction would no longer hold, while under the PL condition, objective-value convergence may still be established but trajectory tracking would require additional arguments. Although the formal guarantees are restricted to the assumed setting, our experiments on nonconvex neural networks demonstrate the practical relevance of the resulting insights beyond it.

W2/Q1 – Unlearning effectiveness and certification. Our theoretical contribution focuses on tracking counterfactual retraining, rather than providing certified unlearning through distributional indistinguishability. Accordingly, the original evaluation measured forgetting through backdoor removal, where the injected behavior provides a measurable signature of the forgotten client's influence. We agree that additional forgetting metrics would strengthen the evaluation. We therefore conducted membership inference experiments using LiRA (Carlini et al. 2022) on the forgotten client's clean data (see Appendix E.2.3):

MethodLiRA AUC
Original model0.6560
Ours0.4847
No Ghost0.4874
Retraining w. Departure0.5014
Oracle retraining0.4965

Our method reduces the attack AUC to approximately 0.5, closely matching oracle retraining. To further assess reconstruction fidelity, we also measured prediction agreement and mean total variation distance (TVD) relative to oracle retraining:

MethodFull test: Agreement (%)Full test: TVDExclusive label: Agreement (%)Exclusive label: TVD
Ours99.270.007595.290.0452
No Ghost92.970.070613.260.8275
Retraining w. Departure93.120.067814.400.8176

These results provide complementary evidence of effective forgetting and close agreement with oracle retraining, particularly in preserving the disconnected client's exclusive-label information. While they do not establish certified unlearning, they substantially strengthen the empirical validation of our approach.

W3/Q2 – Motivation and distinction from existing methods. We agree that existing methods, including PDUDT, can accommodate dynamic communication topologies. However, topology changes and the unavailability of a retained client's data and computations are distinct challenges. Existing dynamic-topology methods adapt decentralized unlearning to changes in connectivity, but do not reconstruct the missing optimization updates and mixing operations of a retained client whose contribution must remain in the target model. Simply excluding this client would instead approximate retraining without both the forgotten and disconnected clients, which is not the desired counterfactual. Our ghost mechanism explicitly addresses this distinction, reconstructing both the missing gradients and the client's mixing participation. The No-Ghost ablation and the additional prediction-discrepancy results above demonstrate the significance of this mechanism.

W4 – Figure quality. We agree and will replace Figure 1 with a higher-quality vector graphic.

References:

Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramèr. Membership inference attacks from first principles. In IEEE Symposium on Security and Privacy (SP), pp. 1897–1914, 2022.

Kun Yuan, Qing Ling, and Wotao Yin. On the convergence of decentralized gradient descent. SIAM Journal on Optimization, 26(3):1835–1854, 2016.

Alexandre d’Aspremont. Smooth optimization with approximate gradient. SIAM Journal on Optimization, 19(3):1171–1183, 2008.

Xiaoyu Cao, Jinyuan Jia, Zaixi Zhang, and Neil Zhenqiang Gong. FedRecover: Recovering from poisoning attacks in federated learning using historical information. In IEEE Symposium on Security and Privacy (SP), pp. 1366–1383, 2023.

Thank you for the detailed response and additional experiments, which help clarify the proposed setting and strengthen the empirical evaluation. However, my main concern regarding the theoretical contribution remains. While the adopted assumptions are common in related work, the analysis primarily establishes trajectory deviation bounds under restrictive conditions, and its theoretical novelty beyond existing optimization analyses remains unclear. Therefore, I will maintain my original score.

We thank the reviewer for reconsidering our submission and for acknowledging the additional empirical evidence. We would like to clarify an important aspect of our theoretical contribution that may not have been sufficiently emphasized in our previous response.

The main mathematical tool that we rely upon in our theoretical analysis, namely, contraction arguments, are indeed established techniques. We do not claim these as novel mathematical tools. However, we would like to emphasize that our theoretical analysis indeed brings novelty that goes beyond existing optimization analyses. Specifically, our theoretical contribution lies in characterizing the specific reconstruction errors arising when a retained client becomes unavailable, and establishing how these errors affect the ability to track counterfactual retraining.

In particular, our analysis goes beyond a generic trajectory-deviation bound in two respects. First, Lemma 1 and Corollary 2 (Appendix C.2) explicitly characterize the ghost-reconstruction error in terms of the gossip budget, network spectral properties, and disagreement among distributed ghost replicas. Under the stated conditions, the error decreases geometrically with the number of gossip iterations. Second, Theorem 1 establishes how this ghost-reconstruction error and the gradient-approximation error separately affect counterfactual tracking, showing that they can be controlled through different mechanisms. This analysis is specific to our setting, where the unavailable client's contribution must be preserved despite the absence of its data and computations, and is not addressed by existing dynamic-topology unlearning methods (e.g., PDUDT).

Furthermore, motivated by the reviewers' comments, we have extended the formal analysis in Appendix C.5 to multiple forgetting requests and retained-client disconnections, deriving an aggregate ghost-reconstruction error bound (Lemma 3) and a generalized tracking guarantee (Corollary 4).

We acknowledge that our guarantees remain restricted to the stated regularity assumptions. Nevertheless, we believe that identifying and theoretically characterizing these new error mechanisms, together with the proposed reconstruction algorithm and its empirical validation, constitute a meaningful technical contribution beyond existing decentralized optimization analyses. We hope this clarification helps the reviewer reconsider the theoretical contribution and overall assessment.

Reviewer_gnyaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper combines (i) machine unlearning and (ii) client disconnection in a unified framework for decentralized learning. These two phenomena both involve removing a node from the decentralized network, however, for (i) we want to unlearn the node's information from the network, whereas for (ii) we need to retain the information of the node in the network. Towards this, the paper (i) employs an adapted trajectory-based Taylor–L-BFGS for unlearning, and (ii) uses ghost replicas of the disconnected node to conduct multiple consensus iterations to retain its information in the network.

Critical strengths
  1. The consideration of machine unlearning alongside node disconnection is an interesting approach, both from a practical point of view and also the mathematical framework needed to analyze it.

  2. The paper is written well, which makes it easy to understand the motivation and each piece of the methodology.

Critical weaknesses
  1. The assumptions used in the paper are relatively strict, i.e., μ\mu-strong convexity and uniformly bounded gradient reconstruction errors. The authors also do not discuss whether making these assumptions are necessarily (are they common machine unlearning literature) and whether others have relaxed the assumptions to non-convex models and non-uniformly bounded gradients.

  2. The work is limited to only unlearning a single node NN, and the disconnection event of a single client rr. In decentralized graphs, the network conditions are often not so predictable, and thus it would be more compelling if the authors considering an arbitrary subset of disconnection and unlearning of an arbitrary number of nodes.

Questions for authors
  1. The counterfactual terminology is used in the paper to refer to the new model that would have been learned had the training did not see the data of a specific client. Is this prevalent in the machine unlearning literature? This nomenclature has a specific meaning in causal inference, and a connection has not been made it the paper.

  2. The two consensus operations required in parallel to recover the state of the disconnected client, i.e., matrix AA and vector bb, resemble the push-sum algorithm as a consensus mechanism over directed graphs. However, no connections are made in the paper, and it is not justified why this could not be done via a single consensus operation in the paper.

  3. Is there a way to remove the connectivity requirement of the undirected graph after node rr is disconnected? Since it's model is copied to all its neighbors, it seems its model information could be retained in both of the disconnected graphs. Would there be a way to also accomplish machine unlearning of node NN in both of the disjoint graphs?

4 comments

Thank you for the constructive comments. We address both concerns below, supported by additional experiments included in the revised appendix E.2

W1 – Theoretical assumptions. We would like to clarify that our theoretical contribution is an analysis of the learning dynamics, characterizing how closely the unlearning trajectory tracks counterfactual retraining and how different reconstruction errors affect this correspondence. Such analyses commonly impose regularity assumptions to obtain informative convergence bounds, as in decentralized optimization (Yuan et al., 2016) and inexact gradient methods (d'Aspremont, 2008). In fact, the closely related FedRecover framework (Cao et al., 2023) adopts such assumptions to analyze tracking of retraining, while demonstrating empirical effectiveness on nonconvex neural networks. Similarly, our assumptions enable us to isolate the effects of gradient reconstruction and ghost consensus, yielding insights into their impact on unlearning dynamics. While the theoretical guarantees are restricted to the assumed setting, our experiments with nonconvex neural networks demonstrate that these insights remain practically relevant beyond it. We clarify the scope and limitations of these assumptions in Appendix F of the revised manuscript.

W2 – Multiple unlearning requests and client disconnections. The original formulation considers one forgotten and one disconnected client to simplify the exposition and theoretical analysis. However, the proposed framework is not inherently restricted to a single event. To demonstrate this, we conducted an additional experiment on MNIST with two simultaneous unlearning requests and three sequential retained-client disconnections. The forgotten clients contain distinct backdoors, while the departing clients hold unique-label information that must be preserved. The results of the study (which is detailed in Appendix E.2.2 of the revised manuscript) are summarized below:

MethodAccuracy (%)ASR 1 (%)ASR 2 (%)
Ours87.140.760.88
No Ghost69.470.930.86
Retraining w. Departure71.880.580.78
Oracle retraining88.280.140.70

Our method remains within 1.14 percentage points of oracle retraining, while successfully removing both backdoors and substantially outperforming the alternatives in accuracy. These results demonstrate that the proposed reconstruction mechanism can accommodate multiple forgetting requests and sequential disconnections, although the present theoretical guarantee does not cover arbitrary subsets.

References

Kun Yuan, Qing Ling, and Wotao Yin. On the convergence of decentralized gradient descent. SIAM Journal on Optimization, 26(3):1835–1854, 2016.

Alexandre d’Aspremont. Smooth optimization with approximate gradient. SIAM Journal on Optimization, 19(3):1171–1183, 2008.

Xiaoyu Cao, Jinyuan Jia, Zaixi Zhang, and Neil Zhenqiang Gong. FedRecover: Recovering from poisoning attacks in federated learning using historical information. In IEEE Symposium on Security and Privacy (SP), pp. 1366–1383, 2023.

Q1 – Counterfactual terminology. We use counterfactual retraining to denote the hypothetical training trajectory that would have been obtained had the forgotten client never participated. This usage is consistent with existing machine unlearning literature, which refers to retraining without the deleted data as a counterfactual reference (Murata et al., 2026), including the notion of a counterfactual unlearning trajectory (Benarroch et al., 2026). Our use of the term does not invoke a causal inference framework, but rather identifies the reference trajectory that the unlearning procedure seeks to approximate.

Q2 – Consensus mechanism. The reviewer correctly observes that our construction follows the ratio-consensus (push-sum) principle (Kempe et al., 2003). The two parallel iterations aggregate the distributed contributions to the disconnected client's mixing update and provide their normalization, respectively. Their ratio recovers the desired weighted sum without requiring global knowledge of the number of active clients. In our undirected setting, a single average-consensus process followed by appropriate scaling would also suffice if this number were known. We will clarify this connection and the rationale for our formulation in Section 4.2.

Q3 – Connectivity requirement. Connectivity of the active graph is required for the global consensus guarantee, but is not strictly necessary for operating the proposed mechanism. Indeed, ghost replicas can preserve the disconnected client's contribution within separate network components. To examine this, we conducted an additional experiment (see Appendix E.2.2) in which a retained-client disconnection partitions the network into two disconnected components (termed A and B):

MethodAccuracy A (%)Accuracy B (%)
Ours87.0386.82
No Ghost78.4178.48
Retraining w. Departure79.3179.25
Oracle retraining87.4086.49

Despite the loss of connectivity, both components achieve accuracy close to oracle retraining, substantially outperforming the alternatives. This demonstrates empirical robustness beyond the connectivity assumption, although global consensus guarantees no longer apply. If the disconnected client subsequently requests unlearning, the target counterfactual must also exclude its contribution. Handling such sequential forgetting requests across disconnected components would require additional analysis and is beyond our current scope.

References

Hanna Benarroch, Jamal Atif, and Olivier Cappé. Counterfactual certification for sequential unlearning with long-run performance guarantees. arXiv preprint arXiv:2610.17993, 2026.

Naoki Murata, Yuhta Takida, Chieh-Hsin Lai, Toshimitsu Uesaka, Bac Nguyen, Stefano Ermon, and Yuki Mitsufuji. GUDA: Counterfactual group-wise training data attribution for diffusion models via unlearning. In Proceedings of the 43rd International Conference on Machine Learning, pp. 91363–91389. PMLR, 2026.

David Kempe, Alin Dobra, and Johannes Gehrke. Gossip-based computation of aggregate information. In IEEE Symposium on Foundations of Computer Science, pp. 482–491, 2003.

I would like to thank the authors for their rebuttal and the new experiments they ran for (i) multiple client disconnection and unlearning and (ii) disconnected graphs. However, both my comments were mostly concerned with generalizing them formally in the methodology and theoretically show the effects.

Adding those generalizations to the paper's framework will boost the paper's quality and insight, which might be hard to finish in this short rebuttal. For that reason, I would recommend a resubmission and would like to maintain my original score.

We thank the reviewer for the constructive follow-up and, in particular, for encouraging us to formally generalize our methodology and theoretical analysis. We agree that this extension strengthens the contribution, and, as our original derivation was naturally extendible, we are pleased to clarify that we have now incorporated it into the revised manuscript.

Specifically, Appendix C.5 formally extends the framework to multiple simultaneous unlearning requests and multiple retained-client disconnections occurring at different iterations. The extension is relatively natural: unlearning multiple clients amounts to redefining the retained-client set, while each disconnected retained client is represented by its own distributed ghost replicas, with the corresponding reconstruction errors aggregated across clients.

It is emphasized that the extension is not limited to the algorithmic formulation. Lemma 3 and Corollary 4 in Appendix C.5 establish an explicit bound on the aggregate ghost-reconstruction error and generalize Theorem 1, respectively. The resulting tracking bound retains the same structure, with the errors of multiple disconnected clients accounted for explicitly. We have also added a clarification following Algorithm 1 describing the generalized procedure.

The extension retains two network assumptions: disconnected retained clients are nonadjacent, and the active graph remains connected. The former prevents coupling between the ghost reconstructions of neighboring disconnected clients, while the latter ensures global information exchange and consensus. These assumptions are standard in spirit for decentralized optimization, where global convergence guarantees typically require connectivity or joint connectivity over time. There are existing distributed optimization analyses that accommodate graphs that are temporarily disconnected, provided sufficient connectivity is maintained over time (e.g., Nedić and Olshevsky, 2015; Nedić et al., 2017), still, permanent fragmentation into components that cannot communicate generally precludes convergence to a common global solution. Our additional fragmentation experiments demonstrate empirical behavior, but we do not claim a corresponding global guarantee.

We sincerely appreciate the reviewer's suggestion, which has led to a meaningful extension of both our method and theoretical results. We hope that the new formal analysis addresses the principal concern underlying the recommendation to resubmit, and that the reviewer will reconsider the assessment in light of these revisions.

References

Nedić, Angelia, and Alex Olshevsky. "Distributed optimization over time-varying directed graphs." IEEE Transactions on Automatic Control 60, no. 3 (2014): 601-615.

Nedic, Angelia, Alex Olshevsky, and Wei Shi. "Achieving geometric convergence for distributed optimization over time-varying graphs." SIAM Journal on Optimization 27, no. 4 (2017): 2597-2633.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.