acceptodds
Under review as a conference paper at ICLR 2027

ProbFedZO: Probabilistic Step Size Adaptation for Federated Zeroth-Order Optimization

Abstract

Zeroth-order optimization allows clients with forward-only access to participate in federated learning while uploading only a few scalars. However, it does not determine how far to move along the estimated direction. Fixed schedules cannot adapt to the current round, whereas line search tests each candidate through additional client communication. We propose two algorithms that select the step size on the server without further client contact. Both collect one shared batch of probe evaluations and fit a polynomial surrogate of the global objective. Under full participation, SurrFedZO applies an Armijo test to the surrogate and takes the largest passing step. Under partial participation, ProbFedZO corrects the test for the uncertainty caused by client sampling and certifies a step only when its predicted decrease exceeds that uncertainty. We bound the error of the surrogate by the probe radius and show that accepted steps decrease the global objective, deterministically for SurrFedZO and with a prescribed probability for ProbFedZO. On federated logistic regression and LoRA fine-tuning, both algorithms achieve competitive or improved risk at matched communication budgets, including settings where fixed and diminishing schedules stall.

Primary area:
optimization
Keywords:
Federated Learning, Optimization, Probabilistic Numerics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_S7ucOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies step-size adaptation in federated zeroth-order optimization. The server broadcasts shared probe points, and each client uploads only scalar changes of its local objective evaluated at these probes. The server fits a polynomial surrogate model from these scalar responses and selects both the update direction and the step size on the server side using an Armijo condition applied to the surrogate. Under full participation, SurrFedZO directly applies the Armijo test to the surrogate; under partial participation, ProbFedZO further incorporates a correction for client sampling uncertainty, certifying a step only when the predicted decrease exceeds the estimated uncertainty. The paper also provides bounds on surrogate model error, probabilistic descent guarantees under sampling uncertainty, and experiments on logistic regression and LoRA fine-tuning.

Critical strengths

1.Reusing the probe responses already collected to construct a surrogate and selecting the step size on the server is a sensible communication-efficient idea.

2.ProbFedZO propagates the client-sampling covariance to the surrogate gradient and Armijo residual, and certifies candidate steps using an uncertainty-corrected condition. This design is more suitable for the partial-participation setting than plain line search or fixed step sizes.

3.Experiments include two-dimensional nonconvex functions, federated logistic regression, and LoRA fine-tuning of OPT-125M and Qwen3-0.6B-base, together with several ablations, indicating substantial experimental effort.

Critical weaknesses

1.Key symbols such as Vt,ϵH,λV_t, \epsilon_H, \lambda, the kernel degree pp, and the candidate step-size grid are not adequately explained in the main text. The reason why the direction dtd_t lies in RrR^r rather than RdR^d and its relation to the subspace VtV_t need to be clarified. Too much content is deferred to Appendix Table 1, hurting the self-containedness of the main text.

2.Theorem 1 directly relies on an “Assumption 5” that is not stated in the main text. The smoothness assumptions in Lemma 1, the poisedness condition on the probes, etc., are also not fully presented. Readers cannot easily judge the applicability of the theorems without consulting the appendix.

3.The statement of Proposition 1 in the main text is phrased using the estimated covariance, but the appendix proof actually uses the population covariance CtC_t for the uncertainty scale. The algorithm uses C^t\hat{C}_t in place of CtC_t; whether the uniform probability bound remains valid after this substitution is not fully proved in the paper. In addition, the theory requires κt>dteff\kappa_t > \sqrt{d^{eff}_t} to obtain a positive probability guarantee, whereas the experiments use κ=ρ/(1−ρ)\kappa = \sqrt{\rho/(1-\rho)} under the “uninflated plug-in rule.” The theoretical status of this choice needs to be clearly stated.

Questions for authors

1.Some minor questions: (1) why dt∈Rr\mathrm{d}_t \in \mathbb{R}^r not Rd\mathbb{R}^d?; (2) why i∈[m]i\in [m] not [N][N] in equstion 3?;

2.How is VtV_t generated or updated? Is the direction dtd_t in subspace coordinates, with the actual update being VtdtV_t d_t? Please clarify at the first occurrence in the main text.

3.How is κt\kappa_t chosen in the experiments? The theory requires κt>dteff\kappa_t > \sqrt{d^{eff}_t}, but the experiments use κ=ρ/(1−ρ)\kappa = \sqrt{\rho/(1-\rho)}. Is this choice heuristic or formally calibrated?

4.How is the candidate grid Ht\mathcal{H}_t chosen?

5.How is the “matched communication budget” defined? How much communication is consumed by the fixed step size, diminishing step size, and distributed line search baselines? Are local evaluation counts and uploaded scalar counts unified into the same budget?

2 comments
AuthorsOpenReview ↗

Reply to Reviewer S7uc

We sincerely thank the reviewer for the close reading of both the main text and the proofs. Your comments on Proposition 1 in particular helped us state our guarantees more accurately.

W1, Q2. Notation and self-containedness. Thank you, we agree that too much was left to the appendix. The items you listed are now defined in the main text at first use, and your reading of V_t and d_t is correct:

ItemNow statedWhere
V_td×r Gaussian matrix, orthonormalized, drawn each round from a shared seed, never transmittedSec. 3.1
d_t in R^rsubspace coordinates; the update is w_{t+1} = w_t + V_t(η_t d_t)Sec. 3.1
Kernel degreep = 3 in all experimentsSec. 3.3
Step-size grid80 geometric values from η_max,t = min{η_0, Δ/‖d_t‖} down to 10^-3 · η_max,t; Δ = 0.5, η_0 = 1Sec. 3.4, App. E.11

W2. Assumptions behind Lemma 1 and Theorem 1. Thank you for pointing this out. Sec. 5 now opens with an "Assumptions" paragraph: (A1) exact responses, C^p smoothness with a Lipschitz p-th derivative, and a poised probe set; (A2) uniform client sampling; (A3) a lower bound on F and L-smoothness. Lemma 1 and Theorem 1 now cite (A1) directly.

W3, Q3. Estimated versus population covariance, and the choice of κ. We are grateful for this observation, which is correct: the statement of Prop. 1 was imprecise. It is now stated explicitly for the population covariance, and a new Remark 1 separates what is proved from what is heuristic:

Covariance in the testκStatusWhere
Population C_tκ > sqrt(d_eff)proved, probability ≥ 1 − d_eff/κ²Prop. 1, Sec. 5.2
Inflated plug-insameproved under bounded responsesProp. 5, App. C.10
Uninflated plug-in (experiments)sqrt(ρ/(1−ρ))heuristic, measured empiricallyRemark 1, App. E.6

So the experiments use a heuristic calibration (Cantelli's, for a single candidate fixed before sampling), and we now say so in the main text and list its certification as open. Because it is a heuristic, we measure its reliability directly: the fraction of certified steps that decrease the true objective rises from 0.77–0.89 at ρ = 0.1 to 0.86–0.98 at ρ = 0.9 across three datasets (Table 3).

Q1. Notation in Eq. (3). Thank you for catching this. Eq. (3) is written with S and m so that it serves both algorithms; under full participation S = {1, …, N} and m = N. This is now stated right after the equation.

Q4. Candidate grid. The grid is given in the table above. Every candidate is tested on the surrogate, so the grid costs only server arithmetic and no client evaluations.

Q5. Matched communication budget. Thank you for asking, as this was not defined clearly. The budget is a cap on the total bits uploaded by clients, at 32 bits per scalar; downlink is not counted because V_t and the probes come from a shared seed. Every method runs until it reaches the same total. Per round, constant and diminishing steps and our rules upload mK scalars, while the line search uploads mK + mJ (J ≤ 8) and so completes fewer rounds. Each uploaded scalar corresponds to one local evaluation (Sec. 6, App. E.11, Table 9).

We thank the reviewer again for the detailed and helpful feedback, and we would be glad to clarify any remaining point.

Thanks for the authors' detailed response. It addresses my concern. I will increase my score.

Reviewer_6atdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Adaptive step size methods in federated zeroth-order optimization are studied in this manuscript, where clients provide function evaluations instead of gradients. Two algorithms are proposed: SurfFedZO for full client participation and ProbFedZO for partial client participation. In SurfFedZO, the authors construct a polynomial surrogate of the global objective from shared probe evaluations and use an Armijo line search to choose the update step size without additional communication rounds. In ProbFedZO, sampling uncertainty is considered. Theoretical results are provided to present approximation errors and probabilistic descent guarantees. Several simulation results are shown to demonstrate the benefits of the proposed methods.

Critical strengths

The paper proposes two adaptive step size selection methods, i.e., SurfFedZO and ProbFedZO, for federated zeroth-order optimization under full and partial client participation, respectively. The proposed algorithms are supported by theoretical analysis of descent guarantees and probabilistic step size certification, and simulation results demonstrating their effectiveness on several learning tasks.

Critical weaknesses
  1. This paper lacks a convergence analysis to a stationary point for nonconvex optimization. Furthermore, the computational and communication complexities of the proposed algorithms are not given, particularly regarding the total number of function evaluations and communication rounds needed to obtain a target accuracy. Since adaptive step size selection and communication efficiency are the main motivations of this work, these theoretical results are important. The authors should provide convergence rate, computational, and communication complexities analysis for both SurfFedZO and ProbFedZO. These results should be compared with existing federated zeroth-order optimization algorithms, e.g., FedZO, DeComFL, and HiSo.
  2. In the current version of this paper, the simulation results are mainly about comparing the proposed algorithms with fixed and diminishing step size schedules and a distributed line search baseline. However, there are existing federated zeroth-order optimization methods, including FedZO, DeComFL, HiSo, and ZO-AdaFL. The authors need to provide additional simulation results of these baselines to demonstrate whether the method proposed by the authors has sufficient performance improvement compared to existing methods, and reduces the costs of computation and communication.
Questions for authors
  1. Please provide analysis of the convergence rate, computation, and communication complexities for SurfFedZO and ProbFedZO.
  2. Please provide a comparison with existing federated zeroth-order optimization algorithms, e.g., FedZO, DeComFL, and HiSo, in terms of the convergence rate, computation and communication complexities.
  3. Better to place the theoretical assumptions in the main content.
  4. Can the authors add experiments including existing federated zeroth-order optimization algorithms, e.g., FedZO, DeComFL, HiSo, and ZO-AdaFL, in terms of convergence performance, communication and computational costs?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 6atd

We sincerely thank the reviewer for the thoughtful comments. They led us to add a convergence analysis, a comparison of rates and costs, and new experiments, and we believe the paper is stronger for it.

W1, Q1. Convergence to a stationary point and complexity. Thank you for raising this. We agree that these results matter for a paper motivated by step-size selection and communication. They are now in Sec. 5.3 (Prop. 2), with proofs in App. F.

SurrFedZO (full participation)ProbFedZO (partial participation)
Rate for min_t E‖∇F(w_t)‖²O(d/r · 1/T)O(d/(rπ) · 1/√T), conditional on certification probability π
Non-decaying termsmodel error O(Δ^2p), removed by a shrinking radiusmodel error, sampling variance O((N−m)/(Nm)), false certification O(d_eff/κ²)
Rounds to ε-stationarityO(d/(rε))–
Evaluations / scalarsO(dN(K+1)/(rε)) / O(dNK/(rε)) in totalm(K+1) / mK per round

We would like to be transparent about one limitation: the rate for ProbFedZO is conditional on a lower bound on the certification probability. We explain in App. F.4 why this cannot be removed without further assumptions, and we list an unconditional rate as an open problem.

Q2. Comparison with existing methods in rate and complexity. We added App. G (Tables 13 and 14). A summary:

MethodRate in TEvaluations per roundScalars per round
FedZOO(sqrt(d/(mHT·b1·b2)))mH·b1(b2+1)md
ZO-AdaFLO(1/T) to a neighborhood, full participation2mKmd
DeComFLO(sqrt(d/(mPKT)))mK(P+1)mKP
HiSoO(1/√T)mτ(P+1)mτP
SurrFedZOO(d/r · 1/T)N(K+1)NK
ProbFedZOO(d/(rπ) · 1/√T)m(K+1)mK

We respectfully note that these rates are proved under different assumptions (stochastic versus exact local evaluations), so we do not claim a faster rate. What our method changes is the per-round cost, which equals that of a fixed step.

Q3. Assumptions in the main text. Thank you, we agree. Sec. 5 now opens with an "Assumptions" paragraph (A1 to A3), and Lemma 1, Theorem 1 and Prop. 1 cite them directly.

W2, Q4. Experiments against existing methods. Thank you for this request. We agree that a new optimizer should be measured against the strongest existing methods. All baselines in the paper run DeComFL's protocol, which we now state explicitly, and we added HiSo on all six language-model tasks at the same budget, using the authors' released code (App. E.14, Table 12). Final training loss, three seeds:

TaskProbFedZO (one setting)HiSo, best rule per taskHiSo, η = 10^-4HiSo, η = 10^-3HiSo, diminishing
CB0.7880.8380.8470.8420.838
AG News1.3421.2681.4221.2681.389
TREC1.6531.6451.8021.6871.645
RTE0.4430.4430.4431.0910.731
QNLI0.5340.5340.5340.8090.719
SST-20.2380.2690.2690.3670.708

Against HiSo's best rule on each task, ProbFedZO is lower on CB and SST-2, equal on RTE and QNLI, within one standard deviation on TREC, and higher on AG News. With one setting per method, ProbFedZO is lower than or equal to every HiSo rule on at least five of the six tasks.

We would like to stress one point, because it is central to this comparison. A best step size exists for each task and can be very good, but it has to be found. In federated learning the server holds no data, so it has no validation set on which to compare candidate step sizes. Each candidate costs a complete federated run, and selecting the winner requires the global objective, which the server cannot evaluate under partial participation. The "best rule per task" column was obtained in exactly this way in simulation, at a cost the budget does not count. ProbFedZO selects its step inside each round from responses it already receives, so its reported budget is its whole cost.

FedZO and ZO-AdaFL upload full model updates, so a single round exceeds our entire bit budget. We therefore compare with them by rate and per-round cost in App. G.

We are grateful for these suggestions and would be happy to provide any further clarification.

Dear Reviewer 6atd,

Thank you again for your constructive feedback, which helped us substantially improve our manuscript.

We have addressed your main concerns by adding a nonconvex stationary-point convergence analysis, computational and communication complexity comparisons, and additional experiments with HiSo.

As the discussion period is approaching its end, we would greatly appreciate it if you could review our response and let us know whether these additions adequately address your concerns, or if any issues remain unresolved.

Thank you very much for your time and consideration.

Thank you for the detailed responses. I appreciate the additional theoretical analysis and experiments, which have addressed several of my original concerns. However, two theoretical points remain.

1. Complexity of SurrFedZO. In Corollary 6, the probe radius Δ\Delta needs to be sufficiently small as ϵ\epsilon decrease to control the model-error floor. However, Eq. (99) shows that ηmin⁡\eta_{\min} also depends on Δ\Delta. Since the iteration complexity contains a factor of 1/ηmin⁡1/\eta_{\min}, could the authors clarify how this dependence is accounted for in the claimed O(d/(rϵ))O(d/(r\epsilon)) communication complexity? It seems that the dependence of ηmin⁡\eta_{\min} on ϵ\epsilon may affect the stated rate.

2. Convergence of ProbFedZO. I appreciate the authors' transparency regarding the limitations of Proposition 7. However, its convergence guarantee requires a uniform positive certification probability, which Remark 13 acknowledges may not hold near stationarity. Furthermore, the guarantee does not directly cover the uninflated plug-in covariance used in the experiments. Therefore, the convergence analysis still does not fully establish a stationarity guarantee for the implemented ProbFedZO algorithm. I would appreciate clarification on how the main theoretical claims should be interpreted in light of these limitations.

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

The paper studies step-size selection in federated zeroth-order optimization. Fixed and diminishing schedules may adapt poorly to local geometry, while conventional distributed line search requires additional client queries for each candidate step. The proposed approach reuses function-value differences from a shared batch of probes to fit a polynomial surrogate anchored at the origin. Its gradient and a modified Hessian, with eigenvalue magnitudes bounded below, determine the search direction.

Under full participation, SurrFedZO applies an Armijo test to the surrogate. Under partial participation, ProbFedZO estimates sampling uncertainty using the sample covariance of client responses and adds uncertainty margins to the directional slope and Armijo residual. Candidate selection takes place entirely on the server. If certification fails, the algorithm uses a gradient fallback with a diminishing step size.

Critical strengths

1.The method reuses one shared batch of client probe evaluations to construct a curvature-aware update direction and perform surrogate line search on the server. Testing additional candidate step sizes therefore requires no further client evaluations or communication, directly addressing a key cost of line search in federated zeroth-order optimization.

2.The paper extends the method to partial client participation by accounting for variation across client responses. Its analysis separates surrogate approximation error from client-sampling error and addresses the adaptive selection of directions and step sizes. Experiments on nonconvex objectives, federated logistic regression, and LoRA fine-tuning examine the method across several settings under matched communication budgets.

Critical weaknesses

1.There is no direct comparison with ordinary surrogate line search using the same direction and probe budget, so the contribution of the uncertainty correction remains unclear. The tuning of the fixed and diminishing step-size baselines is also unclear, making the reported gains difficult to assess.

2.The proposed line search avoids additional client communication for candidate steps, but each participating client still performs K+1 loss evaluations per round. Results plotted against communication bits alone do not show whether the method reduces total client computation or wall-clock training time.

Questions for authors

The theoretical surrogate error bound requires a sufficiently large, well-conditioned probe set, while each additional probe increases client-side loss evaluations and communication. Could the authors report a sensitivity analysis varying K, including values near the minimum required for the polynomial surrogate, and show its effect on surrogate accuracy, the fraction of accepted steps that decrease the true objective, final risk, and total computation and communication cost?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Fuiw

We sincerely thank the reviewer for the careful reading and the constructive suggestions. Each of your points led to a new experiment in the revision.

W1. Contribution of the uncertainty correction, and baseline tuning. Thank you for pointing this out. We agree that the original comparison did not separate the step-size rule from the direction. We added an ablation that keeps the protocol, probes, surrogate and direction fixed and changes only the step-size rule (App. E.12, Fig. 12, Table 10). Final risk after 5×10^5 bits, OPT-125M, m = 10 of N = 100, α = 0.01, three seeds:

Step-size ruleAG NewsCBTREC
Diminishing schedule1.4040.8601.721
Distributed line search1.3400.8201.705
ProbFedZO1.3030.7951.640

ProbFedZO is not the fastest rule early in training, because it declines to certify a step when client disagreement is larger than the predicted decrease. It reaches the lowest final risk on all three tasks. To show the effect of the correction's size, we sweep the multiplier κ from 0.33 to 3.0: this removes 55% to 87% of the falsely certified steps while keeping 81% to 84% of the correct ones (App. E.6, Table 3).

On baseline tuning, we apologize that this was unclear. The settings are now stated in Sec. 6 and App. E.11: constant steps over {10^-3, 10^-2, 10^-1} for logistic regression and {10^-7, …, 10^-1} for language models, with every tested value plotted; the diminishing schedule η_0/sqrt(t+1) with η_0 = 1; and a line search over {1, …, 10^-7} with at most 8 trials per round.

W2. Client computation and wall-clock time. We agree that bits alone do not show client computation. We added a per-round cost table (Table 9, App. E.11), with m clients, K probes and J line-search trials:

RuleLocal evaluationsUploaded scalarsExtra round-trips
Constant or diminishing stepm(K+1)mK0
Distributed line searchm(K+1) + mJmK + mJJ
SurrFedZO / ProbFedZOm(K+1)mK0

Our rules cost a client exactly what a fixed step costs. The only additional cost is on the server, measured at 2.1 ms per round, against 0.64 s for a single client forward pass on Qwen3-0.6B. Wall-clock time is therefore dominated by the client evaluations counted above.

Q1. Sensitivity to the probe count K. Thank you for this suggestion, which we found very useful. We ran the sweep on a9a with the cubic kernel, r = 2 and three seeds (App. E.13, Fig. 13, Table 11). The poised minimum is K = 9 and our default is K = 10.

K57910121520
Accepted steps that decrease the true objective0.760.880.970.981.001.001.00
Evaluations per round6m8m10m11m13m16m21m
Total bits, 100 rounds, N = 100 (×10^6)1.602.242.883.203.844.806.40

Surrogate accuracy (Hessian SNR) rises from about 12 dB at K = 5 to 30 dB at K = 9 and saturates at 38 to 39 dB for K ≥ 12. The risk curves for all K ≥ 9 agree within one standard error, while K = 5 is clearly worse. Reliability therefore turns on at the threshold predicted by Lemma 1, and probes beyond K of about 10 to 12 add cost without adding progress.

We hope these additions address your concerns, and we would be happy to clarify anything further.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Author responses: SurrFedZO / ProbFedZO

We sincerely thank all three reviewers and the area chair for the time and care they put into our paper. The comments were constructive and have made the paper clearer and more complete. Changes in the revised PDF are colored by reviewer (green: Fuiw, blue: 6atd, orange: S7uc), and each change carries a tag such as [6atd W1] that matches the items in our replies.

Since several comments relate to it, we would first like to clarify what the method does.

What the method is. Existing federated zeroth-order methods estimate a direction by finite differences and then need a step size that is set in advance. SurrFedZO and ProbFedZO take a different route: the server fits a polynomial surrogate to the clients' scalar responses and takes both the update direction and the step size from that model. This is a new server-side optimizer. What it keeps from existing work is the communication protocol (DeComFL's), at the per-round client cost of a fixed step.

Fixed or scheduled stepDistributed line searchOurs
Adapts the step per roundnoyesyes
Needs per-task tuningyesnono
Extra client cost per roundnoneone evaluation and one scalar per client per trialnone
Accounts for client samplingnonoyes (ProbFedZO)

Main additions in the revision.

AdditionRequested byWhere
Convergence rate and complexity6atdSec. 5.3, App. F
Rates and costs vs. FedZO, ZO-AdaFL, DeComFL, HiSo6atdApp. G
Experiments against HiSo (six tasks)6atdSec. 6.2, App. E.14
Step-rule ablation, baseline settings, cost accountingFuiw, S7ucSec. 6, App. E.11, E.12
Probe-count sweepFuiwApp. E.13
Notation and assumptions in main text; Prop. 1 correctedS7uc, 6atdSec. 3.1, Sec. 5

We have also tried to be open about the limits of our results, and we point them out in the individual replies. We are grateful for the chance to improve the paper and would be glad to answer any further questions.

Dear Reviewers and Area Chair,

We sincerely thank you for your careful reviews, constructive feedback, and thoughtful discussion.

Your suggestions have helped us substantially improve the theoretical analysis, experimental evaluation, and clarity of our manuscript.

We greatly appreciate your time and consideration.

Best regards, The Authors

AuthorsOpenReview ↗

**Summary of the revision and the discussion**

We thank all reviewers and the area chair for their time. For convenience, we summarize what changed in the revised paper and where to find it.

Reviewer Fuiw

  • Baselines and tuning (W1): all baselines run the same protocol, and all tested step sizes are reported (Section 6, Appendix E.11). Appendix E.12 compares the step-size rules on the same direction and shows the selected step sizes.
  • Cost (W2): per-round client cost and measured server time (Appendix E.11, Table 9).
  • Sensitivity to the probe count (Q1): Appendix E.13.

Reviewer S7uc

  • Notation, subspace and grid (W1, Q2, Q4): Sections 3.1 and 3.4, Appendix E.11.
  • Assumptions in the main text (W2): the assumptions paragraph of Section 5.
  • Population versus estimated covariance, and the choice of κ\kappa (W3, Q3): Proposition 1 and Remark 1 now state what is proved and what is heuristic.

Reviewer 6atd

  • Convergence rate and complexity (W1, Q1, Q2): Section 5.3, Appendices F and G. Following the reviewer's follow-up comment, we corrected the complexity of SurrFedZO in Corollary 6: the number of rounds is O((d/(rε))2p/(2p−1))O\big((d/(r\varepsilon))^{2p/(2p-1)}\big), that is, exponent 6/5 for the cubic surrogate.
  • Comparison with HiSo and other methods (W2, Q4): Appendix E.14, Table 12.

What we claim and what we do not. SurrFedZO converges to a stationary point. For ProbFedZO we prove that each certified step is safe with a stated probability, and we give a rate that is conditional on the certification probability. We do not claim a convergence guarantee for the implemented ProbFedZO; its performance under partial participation is supported by the experiments. The abstract and the contribution list now state this.

We are grateful for the reviewers' comments, which made the paper more accurate.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.