acceptodds
Under review as a conference paper at ICLR 2027

Asymptotic Anytime-Valid Inference for Federated Learning

Abstract

Federated models are often inspected or released repeatedly during training, making uncertainty quantification at a single fixed horizon insufficient. We develop asymptotic anytime-valid inference for smooth, locally identifiable federated -estimation with full-participation Local SGD and deterministic intermittent communication. For the Polyak-Ruppert average of synchronized iterates, we establish an almost-sure Gaussian approximation that accounts for both total local computation and the variance clock of communication-round averages. The approximation retains the exact finite-schedule normalization and controls the accumulated local drift and noise-freezing errors. Combining it with Gaussian time-uniform boundaries and a strongly consistent covariance estimator yields confidence sequences with asymptotic time-uniform coverage after any fixed, pre-specified multiplicative inflation. We also prove strong consistency of a federated sandwich estimator based on fresh samples. Experiments on generalized linear models and a public-health dataset examine coverage, precision, and sensitivity to communication schedules.

Primary area:
learning theory
Keywords:
Anytime-valid inference, Federated learning, Local SGD, Strong Gaussian approximation.

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_Xv4GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies asymptotic anytime-valid inference for Polyak–Ruppert averages of synchronized Local-SGD iterates under full participation and deterministic intermittent communication. Its main result is an almost-sure Gaussian approximation on the variance clock v_T, retaining the exact schedule factor nu_T. Gaussian boundaries and a strongly consistent fresh-sample sandwich estimator yield asymptotic confidence sequences with fixed positive inflation. Validation uses three regression models, schedule and heterogeneity studies, and a BRFSS logistic-regression application.

Critical strengths
  1. Theoretical contribution: Theorem 1 and Lemmas 3–5 explicitly control accumulated local drift and noise-freezing errors. This addresses the pathwise control needed for repeated monitoring beyond a fixed-horizon CLT.

  2. Correct schedule accounting: Keeping nu_T = t_T v_T / T^2 avoids replacing a finite-schedule quantity by its limit. Table B.5 demonstrates its practical importance: using the reference normalization raises Log-schedule failure from 2.6% to 57.8%.

  3. Substantial sensitivity analysis: The experiments include 500-replication main studies, oracle versus plug-in covariance, monitoring starts, heterogeneity, participation, and covariance cost. The BRFSS application reports 358,927 cleaned records and 50 clients, providing a substantial real-data illustration.

Critical weaknesses
  1. Coverage guarantee versus evaluated procedure: Corollary 2 requires a fixed delta > 0, but all reported regions use delta = 0 (pp. 7–8). Definition 1 also supplies no usable finite monitoring start. Table B.2 reports 11.0% failure [7.9%, 15.0%] at nominal 5% for K = 50, h = 0 under the fixed total budget. This does not contradict the asymptotic theorem, but shows that the experiments do not establish reliable calibration for the procedure users would run.

  2. Covariance estimator mismatch: Proposition 1 proves consistency with independent fresh samples, whereas Section 3.1 and Appendix B.6 reuse the first training observation of each round. The reported agreement with oracle covariance does not establish strong consistency of this implemented estimator, which is required by Corollary 2. A proof for the reuse construction or experiments with the proved construction would close this gap.

  3. Restricted participation guarantee: The theory requires full participation and deterministic schedules. Table B.3 shows that applying the full-participation calibration at q = 0.5 produces 51.2% pathwise failure; the q-adjusted alternative has only empirical support. Although acknowledged, this restriction materially limits applicability to federated systems with client availability variation.

  4. Limited evidence for scalable inference: Most simulations use d = 5, and the dimension study stops at d = 30. Table B.4 already reports a covariance-estimation cost ratio of 8.19 times training-only cost at d = 30. The suggested Hessian-vector and low-dimensional-functional alternatives are not evaluated, and communication cost is not quantified. Thus efficient deployment beyond small regression models remains unsubstantiated.

  5. Real-data validation does not assess coverage: Appendix B.2 compares intervals with a centralized estimator fitted on the same data, rather than a known population target. It also does not clearly specify whether local observations are sampled with replacement, without replacement, or reused, which matters for the independent-stream assumptions. The dataset is sizeable, but containing the centralized benchmark is not evidence of nominal population coverage.

Questions for authors

Clarifications that could change my assessment:

  1. Does the sample-reuse estimator in Section 3.1 satisfy the almost-sure consistency condition of Corollary 2? Please provide the corresponding argument or results for the fresh-sample construction.

  2. How do fixed positive inflation values affect coverage and width in the undercoverage settings of Table B.2? Can you give a defensible rule for choosing the monitoring start?

  3. In BRFSS, what is the precise sampling/reuse protocol and inferential target: the empirical-distribution optimizer or a population parameter? How is this protocol connected to the stated assumptions?

6 comments

Thank you for your time and valuable comments. We respond to each point below.

W2 / Q1. We will revise Assumption 5, Proposition 1 and its proof to establish strong consistency for the completed-round sample-reuse construction of Section 3.1; the revised statement and proof are given in our response to Reviewer EeQq (W1). Under Assumptions 1–5, the positive-definite fallback is measurable with respect to completed-round information, is positive definite at every round and converges almost surely to the target covariance. It therefore supplies the covariance-estimation condition of Corollary 2 under the other stated assumptions of that corollary.

W1. Corollary 2 is stated for a fixed δ>0\delta > 0, whereas the regions reported in Section 3 use δ=0\delta=0 to evaluate the first-order construction. On every sample path, a larger δ\delta gives a larger region and multiplies every half-width by exactly 1+δ1+\delta, so failure (the time-uniform type-I error) is nonincreasing in δ\delta. We have added experiments with positive inflation δ∈0.01,0.02,0.05,0.10\delta\in\\{0.01,0.02,0.05,0.10\\} to quantify this trade-off. Table R1 shows representative results from the main simulations with the plug-in covariance. Failure stays below the nominal level for every δ\delta, and its maximum across the main simulations falls from 4.8% at δ=0\delta=0 to 1.0% at δ=0.10\delta=0.10.

Table R1. Time-uniform type-I error (%) and width under positive inflation in the main simulations (plug-in covariance, K=10K=10, T=50,000T=50{,}000, d=5d=5, 500 replications, GM boundary, n0=20,000n_0=20{,}000). Each cell lists failure (%) / terminal projection half-width (units of 10−310^{-3}).

Model/scheduleδ=0.01\delta=0.01δ=0.05\delta=0.05δ=0.10\delta=0.10
Linear/C1C_12.8/15.4521.6/16.0641.0/16.829
Linear/C5C_51.0/7.4180.4/7.7120.0/8.079
Linear/Log0.0/5.5810.0/5.8020.0/6.078
Logistic/C1C_14.2/15.3152.6/15.9210.8/16.679
Logistic/C5C_51.2/7.3510.4/7.6430.2/8.006
Logistic/Log0.4/5.5310.0/5.7500.0/6.023
Poisson/C1C_12.0/5.6431.0/5.8670.8/6.146
Poisson/C5C_51.0/2.7110.8/2.8180.2/2.953
Poisson/Log0.2/2.0400.2/2.1210.0/2.222

We also evaluated positive inflation in the settings of Table B.2 on 500 new replications. Table R2 includes the settings with failure above 5% and the corresponding longer run. Failure decreases with δ\delta. At δ=0.05\delta=0.05, failure is below the nominal level in every evaluated setting except (K,h,T)=(50,0,4,000)(K,h,T)=(50,0,4{,}000), where the excess comes from the early monitored rounds and decreases with a later start (see also our response to Reviewer EeQq, W3).

Table R2. Each numeric cell lists oracle/plug-in values. Representative positive-inflation results in the Linear C5C_5 setting (d=5d=5, 500 replications, GM boundary, zero initialization). Failure (%) denotes at least one noncoverage after the 10% monitoring start. Terminal half-widths are in units of 10−310^{-3} and are multiplied by exactly 1+δ1+\delta at δ>0\delta > 0.

(K,h,T)(K,h,T)δ=0\delta=0δ=0.05\delta=0.05δ=0.10\delta=0.10Half-width, δ=0\delta=0
(50,0,4000)(50,0,4000)9.6/8.06.4/5.24.0/2.44.633/4.660
(50,1,4000)(50,1,4000)6.6/5.63.8/3.22.8/2.06.871/6.895
(50,0,20000)(50,0,20000)3.8/3.42.0/2.01.8/1.82.072/2.075

Corollary 2 holds for any fixed δ>0\delta > 0; in the proof, δ\delta absorbs coupling, Brownian interpolation and covariance-estimation errors that vanish almost surely. In finite samples, δ\delta multiplies every half-width by exactly 1+δ1+\delta, so a larger δ\delta is more conservative. We therefore recommend a small δ\delta, chosen by the extra width that is acceptable and fixed before training together with the monitoring start. We have clarified this choice.

Definition 1 is a limit as T0→∞T_0\to\infty. In practice, T0T_0 is set by the application: it is the first round at which the global model is inspected or released (Waudby-Smith et al., 2024; Bibaut et al., 2022). With the estimator and boundary fixed, a later T0T_0 shortens the monitoring window and preserves the width at every round, so failure is nonincreasing. When no inspection round is specified, we use T0=⌈0.10H⌉+1T_0=\lceil0.10H\rceil+1, where HH is the planned number of rounds. We describe this practical choice in Section 2.3.

We also report failure in the same main simulations when monitoring starts at 5%, 10% or 20% of the rounds, at δ=0\delta=0. Failure is at most 4.8% for every start, and moving the start from 5% to 20% of the rounds changes it by at most 0.6 percentage points. At δ=0.05\delta=0.05, the maxima are 2.8%, 2.8% and 2.6%, respectively. This is consistent with an initialization bias that is small relative to the radius once monitoring starts, so that the earliest monitored rounds contribute few noncoverages.

In the 4,0004{,}000-round setting, the noncoverages occur in the early monitored rounds, where the initialization bias of yˉT\bar y_T is still large relative to the radius: the bias decays as T−1T^{-1}, whereas the radius decays as T−1/2T^{-1/2} (see also our response to Reviewer C5pb, W1). We also evaluate averaging after a warm-up to reduce the effect of initialization. The variance clock and plug-in covariance estimator use the same rounds as the average. We provide an asymptotic guarantee when the number of warm-up rounds bb remains fixed as T0→∞T_0\to\infty. The 5% warm-up and 10% monitoring start are practical choices examined in Table R3.

At δ=0\delta=0, failure after warm-up averaging is at most 3.8% in the main simulations and 4.6% in the settings of Table B.2; with the oracle covariance the maxima are 3.6% and 4.2%. Relative to the full average, the terminal half-width changes by between −10.3-10.3\\% under Log and +2.2+2.2\\% under C1C_1 in the main simulations, where LIL-UB failure is at most 0.6% for every reported δ\delta.

Table R3. Representative results for averaging after the warm-up (plug-in covariance, GM boundary, d=5d=5, 500 replications). For the main simulations, K=10K=10, T=50,000T=50{,}000, n0=20,000n_0=20{,}000 and b=2,500b=2{,}500. The (K,h,T)(K,h,T) rows are Linear C5C_5 settings and use b=200b=200 for T=4,000T=4{,}000 and b=1,000b=1{,}000 for T=20,000T=20{,}000, with n0n_0 the geometric mean of vT−vbv_T-v_b at T0T_0 and at TT. The average, variance clock and plug-in covariance all start after the bb warm-up rounds. The columns headed by δ\delta report failure (%), which denotes at least one noncoverage after the 10% monitoring start. Terminal half-widths are in units of 10−310^{-3} and are multiplied by exactly 1+δ1+\delta at δ>0\delta > 0.

Model/schedule or (K,h,T)(K,h,T)δ=0\delta=0δ=0.05\delta=0.05δ=0.10\delta=0.10Half-width, δ=0\delta=0
Linear/C1C_13.81.60.815.639
Linear/C5C_50.60.20.07.004
Linear/Log0.00.00.04.956
Logistic/C1C_13.21.00.415.500
Poisson/C1C_13.21.40.85.713
(50,0,4000)(50,0,4000)4.22.21.04.587
(50,1,4000)(50,1,4000)2.41.41.06.803
(50,0,20000)(50,0,20000)2.20.60.42.051
(10,1,20000)(10,1,20000)3.42.00.46.801
(5,0,20000)(5,0,20000)4.62.01.06.487

Q2. Table R2 above reports positive inflation in the undercoverage settings of Table B.2: for (K,h,T)=(50,0,4,000)(K,h,T)=(50,0,4{,}000), increasing δ\delta from 0 to 0.05 and 0.10 reduces oracle failure from 9.6% to 6.4% and 4.0%, and plug-in failure from 8.0% to 5.2% and 2.4%, while multiplying each half-width by exactly 1+δ1+\delta. We recommend fixing δ\delta by the acceptable extra width before training and setting the monitoring start to the application's first inspection or release round. When that round is unspecified, we use T0=⌈0.10H⌉+1T_0=\lceil0.10H\rceil+1, with HH the planned number of rounds, as described in W1.

W3. Our theoretical guarantee covers full participation with deterministic communication schedules (see the abstract and Section 4). Extending it to partial participation requires accounting for client-selection variance and new strong-approximation bounds for local drift and noise freezing under sampling. Appendices A.6 and B.5 examine an empirical adjustment in this setting.

With a fraction qq of clients participating, fewer independent client contributions are averaged, increasing the data-noise covariance from A1A_1 to Aq=A1/qA_q=A_1/q in the uniform-sampling setting of Appendix B.5. Client heterogeneity introduces an additional source of variation: the client population gradients cancel at the global optimum under full participation, while a randomly selected subset has a fluctuating aggregate gradient. This produces the selection component BqB_q, giving the leading terminal covariance T−2(vTAq+TBq)T^{-2}(v_TA_q+TB_q). The omitted data-noise inflation and selection variation explain the 51.2% pathwise failure of full-participation calibration at q=0.5q=0.5 in Table B.3. At q=0.5q=0.5, the participation-adjusted GM calibration using A1/qA_1/q gives 4.6% failure; the adjusted failure rates at q=0.8q=0.8 and q=0.6q=0.6 are 1.8% and 2.8%, respectively. These experiments provide empirical evidence for participation-aware calibration in the settings examined. Establishing theory under varying client availability and relaxing deterministic communication remain directions for future work.

We will clarify this empirical scope in Appendix B.5. We refer to the partial-participation covariance T−2(vTAq+TBq)T^{-2}(v_TA_q+TB_q), computed from the known AqA_q and BqB_q, as the PP oracle. The qq-adjusted GM calibration uses only AqA_q: AqA_q accumulates on the vTv_T clock and BqB_q on the TT clock (Appendix A.6), so their combination generally has a time-varying matrix on the scalar GM clock. We use the PP oracle at the terminal round.

W4. We have run d=50d=50 and d=100d=100 and will add these results to the appendix. Table R4 reports the Logistic model under C5C_5, and Table R5 shows representative Linear results.

Table R4. Error and volume cells list oracle/plug-in values. Time-uniform type-I error (%) and average confidence-set volume in the Logistic C5C_5 setting (K=10K=10, T=50,000T=50{,}000, n0=20,000n_0=20{,}000, 500 replications). Error is measured after the 10% monitoring start, with nominal level 5%; the volume root at TT is in units of 10−310^{-3}. The average starts after a warm-up of 2,500 rounds, c0=0.5c_0=0.5 and δ=0.05\delta=0.05.

dd/boundaryErrorAvg CS Volume$^{1/d}$
50/Fixed67.8/69.26.94/6.94
50/LIL-UB0.0/0.016.24/16.24
50/GM0.0/0.08.14/8.15
100/Fixed85.4/87.26.79/6.80
100/LIL-UB0.2/0.016.64/16.65
100/GM0.0/0.07.76/7.77

LIL-UB and GM stay below the nominal level, with GM giving smaller regions, whereas the fixed-time boundary fails far more often. Oracle and plug-in covariance give similar results, and the Linear model behaves similarly.

Table R5. Error and volume cells list oracle/plug-in values. Representative Linear results at d=100d=100, with the same design as Table R4. Time-uniform type-I error is in percent; the volume root at TT is in units of 10−310^{-3}.

Schedule/boundaryErrorAvg CS Volume$^{1/d}$
C1C_1/Fixed44.6/37.463.89/64.52
C1C_1/LIL-UB0.0/0.0157.80/159.35
C1C_1/GM0.6/0.473.34/74.06
C5C_5/Fixed43.4/47.628.50/28.54
C5C_5/LIL-UB0.0/0.069.85/69.94
C5C_5/GM0.0/0.032.57/32.61
Log/Fixed68.4/76.216.56/16.56
Log/LIL-UB0.0/0.040.33/40.34
Log/GM0.0/0.022.36/22.37

The margin below the nominal level comes from the boundaries: under C5C_5 with δ=0\delta=0, simulating the Gaussian limit of Theorem 1 on the same clock and window gives GM failure of 0.46% at d=50d=50 and 0.41% at d=100d=100, against 0.8% and 0.4% with the plug-in. For GM, n0=20000n_0=20000 exceeds the terminal clock under C5C_5 and Log, so the boundary stays above its minimum, and the inflation δ>0\delta > 0 required by Corollary 2 adds to the margin; for LIL-UB, the margin comes from the finite-LIL constants. The norm of a Gaussian vector concentrates as dd grows, so a small increase in the radius lowers failure sharply, and at d=100d=100 the GM radius, inflation included, is only 14% above the fixed-time radius.

For the reported 8.19×8.19\times cost ratio, the cost drops substantially if the covariance is refactorized periodically. Corollary 2 requires an HT\mathcal H_T-measurable, strongly consistent estimator; we will add the following remark to the paper. Let g(T)g(T) be the largest point not exceeding TT of a deterministic increasing grid. Then Σ^g(T)\widehat\Sigma_{g(T)} is HT\mathcal H_T-measurable and converges to Σ\Sigma almost surely, so Corollary 2 applies with Σ~T\widetilde\Sigma_T the positive-definite fallback of Σ^g(T)\widehat\Sigma_{g(T)}, with the statistic evaluated at every round. Table R6 reports new timings for the Logistic model under C5C_5; with a refresh every 100 rounds the cost ratio is 1.69×1.69\times at d=100d=100.

Table R6. Training-plus-covariance/training-only time, both including data generation (same definition as Table B.4); columns give the refactorization interval. Logistic C5C_5, 10 trajectories, 2,000 rounds, median of 3 paired runs, on a different host from Table B.4.

ddEvery roundEvery 100 rounds
304.611.29
508.961.36
10023.471.69

We now evaluate Hessian-vector products and low-dimensional functionals. For a contrast c⊤x⋆c^\top x^\star, the one-dimensional interval needs only the scalar variance c⊤Σc=u⊤Suc^\top\Sigma c=u^\top Su with u=G−1cu=G^{-1}c, and uu can be solved online from Hessian-vector products, giving a matrix-free implementation. The server broadcasts umu_m, and each client returns its Hessian-vector product and its score projected on the running average uˉm\bar u_m. The server updates um+1=um−(m+1)−0.6(H^mum−c)u_{m+1}=u_m-(m+1)^{-0.6}(\widehat H_mu_m-c), averages umu_m from 1% of the rounds, and estimates the variance by the average of (uˉm⊤gˉm)2(\bar u_m^\top\bar g_m)^2 from 2% of the rounds. Each round, every client returns one Hessian-vector product and one score projection. Table R7 reports the type-I error of the resulting interval for a zero coefficient (e6e_6), a nonzero coefficient (e5e_5) and the direction of x⋆x^\star.

Reviewer_C5pbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper develops asymptotic anytime-valid inference for smooth, locally identifiable federated MM-estimation.It considers full-participation Local SGD with deterministic intermittent communication. The authors establish an almost-sure Gaussian approximation for the Polyak--Ruppert average of synchronized iterates. The approximation retains the exact finite-schedule normalization and controls accumulated local drift and noise-freezing errors. Gaussian time-uniform boundaries and a strongly consistent covariance estimator yield asymptotic confidence sequences after fixed, pre-specified multiplicative inflation. Experiments on generalized linear models and a public-health dataset examine coverage, precision, and sensitivity to communication schedules.

Critical strengths
  • The paper addresses uncertainty quantification under repeated monitoring of federated training. This is relevant when model inspection or stopping depends on the observed training path.

  • The Gaussian approximation accounts for total local computation and the variance clock. It retains the exact finite-schedule normalization and controls accumulated local drift and noise-freezing errors.

  • The paper provides explicit Gaussian time-uniform boundaries. It proves strong consistency of a federated sandwich estimator based on fresh samples.

  • The experiments examine coverage and precision across communication schedules, heterogeneity levels, and dimensions. They also assess covariance-estimation accuracy and computational cost.

Critical weaknesses
  • Table B.2 reports 11.0% pathwise failure with oracle covariance, above the nominal 5% level. The monitoring-start analysis in Appendix~B.3 does not cover this setting. The effects of later monitoring and multiplicative inflation in this setting remain unclear.

  • Proposition~1 establishes strong consistency for the fresh-sample sandwich estimator. However, the main experiments reuse each round's first training observation. The consistency argument for this implemented estimator is not explicitly provided.

  • Appendix~B.2 does not specify how observations are sampled within each BRFSS client. It is also unclear whether the confidence sequences target an empirical minimizer or a population parameter. These details matter for relating the application to the independent-stream assumptions.

Questions for authors

Questions

  • Why are synchronized iterates equally weighted when the communication intervals EmE_m vary? How does this choice affect statistical efficiency compared with weighting by EmE_m?

  • How is the mean-square stability condition in Assumption 4 justified for the Poisson regression model in Section 3.1?

  • The experiments reuse each round's first training observation at the round-start parameter. How can the proof of Proposition~1 be adapted to establish strong consistency for this estimator?

  • Table B.2 reports 11.0% pathwise failure with oracle covariance at the nominal 5% level. How do the monitoring start and inflation affect coverage and interval width in this setting?

Suggestions

  • Explain the choice of equal weighting under varying communication intervals. Discuss its implications for statistical efficiency.

  • Clarify whether the Poisson regression model satisfies Assumption 4. Provide supporting arguments or state any additional conditions needed.

  • Provide a strong consistency argument for the estimator that reuses each round's first training observation.

  • Investigate whether the initialization bias contributes to the 11.0% pathwise failure in Table B.2. Examine sensitivity to the monitoring start and multiplicative inflation. Report both pathwise failure and interval width.

8 comments

Thank you for your time and valuable comments. We respond to each point below.

W2 / Q3 / S3. We will revise Assumption 5, Proposition 1 and its proof to establish strong consistency for the completed-round sample-reuse construction of Section 3.1. For round mm, define

g_m=\sum_{k=1}^Kp_k\nabla f_k(\bar x_{t_m};\xi_{t_m}^k),$$ and use the completed-round averages $\widehat G_T=T^{-1}\sum_{m < T}H_m$ and $\widehat S_T=T^{-1}\sum_{m < T}g_mg_m^\top$. Under Assumptions 1–5, the revised proposition establishes $\widehat G_T\to G$, $\widehat S_T\to S$, and $\widehat\Sigma_T\to\Sigma=G^{-1}SG^{-\top}$ almost surely, where $\widehat\Sigma_T=\widehat G_T^{-1}\widehat S_T\widehat G_T^{-\top}$ when $\widehat G_T$ is nonsingular and is zero otherwise. The round-start iterate is measurable before the round's observations are drawn. Conditional on that information, the observations are independent across clients. Localizing near $x^\star$ gives square-integrable martingale differences, and martingale convergence and Kronecker's lemma give the strong law for the completed-round averages. The detailed revised statement and proof are included in our response to Reviewer EeQq (W1). **W1 / Q4 / S4.** The submitted Table B.2 reports $11.0\\%$ oracle pathwise failure in this setting. Corollary 2 is a limit result as the number of rounds grows. At the finite budget of $4{,}000$ rounds, the noncoverages occur in the early monitored rounds, where the initialization bias of $\bar y_T$ is still large relative to the radius: the bias decays as $T^{-1}$, whereas the radius decays as $T^{-1/2}$. At this finite budget the initialization bias does contribute, and its effect vanishes as the number of rounds grows. Using the known simulation target $x^\star$ as a diagnostic initializer, we rerun the same replications from $x^\star$ instead of $0$; at $T=4{,}000$ this removes the bias and brings failure at $\delta=0$ to about the nominal level. With monitoring from $10\\%$ of the rounds, so that the start moves later as $T$ grows, the gap in failure between the two initializations shrinks with $T$ and is negligible at $T=20{,}000$ with a warm-up of $1{,}000$ rounds (Table R2), consistent with a ratio of bias to radius that decays as $t^{-1/2}$. Averaging after the warm-up keeps $x_0=0$ and brings failure below $5\\%$ in this setting: with the plug-in covariance and GM boundary, failure is $4.2\\%$ at $\delta=0$ and $2.2\\%$ at $\delta=0.05$, and the terminal half-width at $\delta=0$ is $4.587\times10^{-3}$. The average, variance clock and covariance estimator exclude the same 200 warm-up rounds, and monitoring starts after $10\\%$ of the rounds. On every sample path, a later start or a larger $\delta$ weakly reduces failure; a later start leaves the width at each round unchanged, and $\delta$ multiplies every half-width by exactly $1+\delta$. Table R1 reports representative values. The two choices can be combined: at $\delta=0.05$ with monitoring from $20\\%$ of the rounds, failure is $5.0\\%$ with the oracle covariance and $4.4\\%$ with the plug-in covariance, at a half-width $5\\%$ wider than at $\delta=0$. We also extend the monitoring-start comparison to the main simulations and will add it to the manuscript. Table R1. Positive inflation and monitoring start in the Linear $C_5$ setting ($K=50$, $h=0$, $T=4{,}000$, $d=5$, 500 new replications, GM boundary, zero initialization). Failure (%) denotes at least one noncoverage after the indicated monitoring start. Terminal half-widths are in units of $10^{-3}$. Each pair is oracle/plug-in. |$\delta$|Failure:10% start|Failure:20% start|Half-width| |---|---|---|---| |0|9.6/8.0|7.0/6.0|4.633/4.660| |0.05|6.4/5.2|5.0/4.4|4.864/4.893| |0.10|4.0/2.4|3.2/2.2|5.096/5.126| Table R2. Initialization in the Linear $C_5$ setting ($K=50$, $h=0$, $d=5$, 500 paired replications, GM boundary, $\delta=0$, $T_0=\lceil0.1T\rceil+1$). Entries are pathwise failure (%), paired as oracle/plug-in. Initialization at the known simulation target $x^\star$ is a diagnostic. The rows with 200 warm-up rounds and $T > 4{,}000$ extend the same runs and use $n_0=\sqrt{v_{401}v_{4000}}$; the other rows use $n_0=\sqrt{v_{T_0}v_T}$. |$T$|Warm-up|$x_0=0$|$x_0=x^\star$| |---|---|---|---| |4,000|200|9.6/8.0|5.0/5.2| |8,000|200|7.2/6.2|4.0/3.6| |20,000|200|4.8/4.0|3.4/3.0| |20,000|1,000|3.8/3.4|3.6/3.4|

Q1 / S1. We use the equal-round Polyak–Ruppert average of synchronized iterates, yˉT=1T∑m=0T−1xˉtm+1,\bar y_T=\frac1T\sum_{m=0}^{T-1}\bar x_{t_{m+1}}, following the estimator studied by Li et al. (2022). Its statistical efficiency is quantified by the schedule factor νT\nu_T. Weighting the synchronized iterates by EmE_m defines a different estimator, yˉTE=1tT∑m=0T−1Emxˉtm+1,\bar y_T^E=\frac1{t_T}\sum_{m=0}^{T-1}E_m\bar x_{t_{m+1}}, with tT=∑m<TEm.t_T=\sum_{m < T}E_m. The distinction between weighting round noises and weighting iterates is important because each synchronized iterate depends on all preceding rounds.

Recall that the independent frozen round noises ε~m\widetilde\varepsilon_m have covariance S/EmS/E_m. The decomposition underlying Theorem 1 gives

\bar y_T-x^\star =-\frac1T\sum_{m < T}G^{-1}\widetilde\varepsilon_m +o_{\mathrm{a.s.}}\\!\left(\frac{\sqrt{v_T}}{T}\right), \qquad v_T=\sum_{m < T}E_m^{-1}.$$ Consequently, its first-order covariance is $$\tag{3} \frac{v_T}{T^2}\Sigma=\frac{\nu_T}{t_T}\Sigma, \qquad \nu_T= \frac{(\sum_{m < T}E_m)(\sum_{m < T}E_m^{-1})}{T^2} \ge1,$$ where the inequality follows from Cauchy–Schwarz. In contrast, the $E_m$-weighted frozen-noise average $$Z_T^E=-\frac1{t_T}\sum_{m < T}E_mG^{-1}\widetilde\varepsilon_m$$ has covariance exactly $\Sigma/t_T$. Thus the frozen-noise comparison favors weighting by $E_m$. To transfer this improvement to the actual iterates for time-uniform inference, one would need to establish $$\tag{4} \bar y_T^E-x^\star=Z_T^E+o_{a.s.}(t_T^{-1/2}).$$ The alternating schedule below gives a different limiting covariance under our assumptions. Using the notation of Appendix A.2, write $$s_m=\bar x_{t_m}-x^\star, \qquad s_{m+1}=B_ms_m-\gamma_m(r_m+\delta_m+\varepsilon_m), \qquad B_m=I-\gamma_mG.$$ For the equal-round average, the coefficient propagating the noise from round $j$ is $$A_j^{T-1} =\gamma_j\sum_{m=j}^{T-1}\prod_{i=j+1}^{m}B_i.$$ For the $E_m$-weighted average, the exact decomposition is instead $$\bar y_T^E-x^\star =I_T^E-\frac1{t_T}\sum_{j < T}A_{T,j}^E (r_j+\delta_j+\varepsilon_j),$$ where $$\tag{5} A_{T,j}^E =\gamma_j\sum_{m=j}^{T-1}E_m\prod_{i=j+1}^{m}B_i, \qquad I_T^E =\frac1{t_T}\sum_{m < T}E_m \left(\prod_{i=0}^{m}B_i\right)s_0.$$ The noise from round $j$ affects every later iterate, and therefore receives the weights of those later rounds. The exact difference is $$\tag{6} A_{T,j}^E-E_jG^{-1} =E_j(A_j^{T-1}-G^{-1}) +\gamma_j\sum_{m=j}^{T-1} (E_m-E_j)\prod_{i=j+1}^{m}B_i.$$ The second term is an additional schedule-variation term. Fix any $K\ge1$, take $p_k=1/K$, and let $$f_k(x;\xi)=\tfrac12x^2-\xi x, \qquad \xi_t^k\stackrel{\mathrm{i.i.d.}}{\sim}N(0,K).$$ Then $f(x)=x^2/2$, $x^\star=0$, and $G=S=\Sigma=1$. The noise variance $K$ normalizes the aggregated gradient-noise variance to one. Initialize all clients at zero and set $$\tag{7} \gamma_m=\tfrac12(m+1)^{-3/5}, \qquad \eta_m=\gamma_m/E_m, \qquad E_{2j}=1,\quad E_{2j+1}=9.$$ Each client performs the actual local updates $$x_{t+1}^k=(1-\eta_m)x_t^k+\eta_m\xi_t^k, \qquad t_m\le t < t_{m+1},$$ followed by synchronization. We can show that $$\bar y_T^E-\bar y_T=o_{\mathrm{a.s.}}(T^{-1/2}), \quad \sqrt{t_T}\\,\bar y_T^E\\ \Rightarrow\\ N(0,25/9),\quad \sqrt{t_T}Z_T^E\sim N(0,1),\quad \sqrt{t_T}(\bar y_T^E-Z_T^E) \Rightarrow N(0,16/9).$$ For this alternating schedule, weighting synchronized iterates by $E_m$ retains the factor $\nu_T$ in the limiting covariance. We will add this example and its proof, given below, to the appendix. To prove this, write $X_m=\bar x_{t_m}$ and $\zeta_{m,r}=K^{-1}\sum_k\xi_{t_m+r}^k$. The $\zeta_{m,r}$ are independent standard normal variables. Solving the local linear recursion and aggregating gives the exact round-level recursion $$\tag{8} X_{m+1}=a_mX_m+U_m, \qquad a_m=(1-\eta_m)^{E_m}, \qquad U_m=\eta_m\sum_{r=0}^{E_m-1} (1-\eta_m)^{E_m-1-r}\zeta_{m,r}.$$ Here $U_m$ is independent of the information available at the round start, with mean zero, and $$1-a_m\le\gamma_m, \qquad a_m^2\le e^{-2\gamma_m}\le1-\gamma_m, \qquad \mathbb E U_m^2 \le\eta_m^2E_m =\gamma_m^2/E_m \le\gamma_m^2.$$ Thus $$\mathbb E(X_{m+1}^2\mid\mathcal F_{t_m}) \le(1-\gamma_m)X_m^2+\gamma_m^2.$$ Since $\sum_m\gamma_m=\infty$ and $\sum_m\gamma_m^2 < \infty$, the Robbins–Siegmund theorem gives $X_m\to0$ almost surely. Moreover, $\gamma_m-\gamma_{m+1}=o(\gamma_m^2)$, so induction in the second-moment recursion yields $\mathbb E X_m^2\le C\gamma_m$. Indeed, eventually $\gamma_m-\gamma_{m+1}\le\gamma_m^2/2$ and, for $C\ge2$ chosen to cover the initial indices, $$(1-\gamma_m)C\gamma_m+\gamma_m^2 \le C\gamma_m-\tfrac C2\gamma_m^2 \le C\gamma_{m+1}.$$

This verifies Assumption 4. Assumptions 1–3 hold with λ=1\lambda=1, p=4p=4, a=3/5a=3/5, and b=0b=0; the communication intervals are uniformly bounded.

Next, (8) gives

\le(1-a_m)\mathbb E|X_m| +(\mathbb E U_m^2)^{1/2} \le C\gamma_m.$$ Consequently, $$\mathbb E\sum_{m=0}^{\infty} \frac{|X_{m+1}-X_m|}{\sqrt{m+1}} \le C\sum_{m=0}^{\infty}(m+1)^{-11/10} < \infty.$$ Tonelli's theorem and Kronecker's lemma imply $$\tag{9} \sum_{m=0}^{T-1}|X_{m+1}-X_m| =o_{\mathrm{a.s.}}(\sqrt T).$$ For even $T$, $t_T=5T$, and the difference of the two averages is particularly transparent: $$\bar y_T^E-\bar y_T =\frac4{5T} \sum_{j=0}^{T/2-1}(X_{2j+2}-X_{2j+1}).$$ For a bound valid for every $T$, define $$C_{m,T} =\sum_{r=0}^{m}\left(E_r-\frac{t_T}{T}\right).$$ The alternating schedule gives $\max_{m < T}|C_{m,T}|\le4$ and $C_{T-1,T}=0$. Summation by parts therefore yields $$\begin{aligned} \bar y_T^E-\bar y_T &=\frac1{t_T}\sum_{m < T} \left(E_m-\frac{t_T}{T}\right)X_{m+1}\\\\ &=\frac1{t_T}\sum_{m=0}^{T-2} C_{m,T}(X_{m+1}-X_{m+2}), \end{aligned}$$ and hence, by (9) and $t_T\asymp T$, $$|\bar y_T^E-\bar y_T| \le\frac4{t_T} \sum_{m=0}^{T-2}|X_{m+2}-X_{m+1}| =o_{\mathrm{a.s.}}(T^{-1/2}).$$ Finally, $$\frac{t_T}{T}\longrightarrow5, \qquad \frac{v_T}{T}\longrightarrow\frac59, \qquad \nu_T\longrightarrow\frac{25}{9}.$$ Corollary 1 gives $\sqrt{t_T}\\,\bar y_T\Rightarrow N(0,25/9)$. The almost-sure equivalence and Slutsky's theorem give the same limit for $\sqrt{t_T}\\,\bar y_T^E$. Writing $$\bar\zeta_m =E_m^{-1}\sum_{r=0}^{E_m-1}\zeta_{m,r}, \qquad \widetilde\varepsilon_m=-\bar\zeta_m,$$ we obtain $$Z_T^E=\frac1{t_T}\sum_{m < T}E_m\bar\zeta_m, \qquad \sqrt{t_T}Z_T^E\sim N(0,1).$$ Thus, $$\sqrt{t_T}(\bar y_T^E-Z_T^E) =\sqrt{t_T}\sum_{m < T} \left(\frac1T-\frac{E_m}{t_T}\right)\bar\zeta_m +o_{\mathrm{a.s.}}(1).$$ The displayed Gaussian sum has variance $$t_T\sum_{m < T}\frac1{E_m} \left(\frac1T-\frac{E_m}{t_T}\right)^2 =\nu_T-1\longrightarrow\frac{16}{9}.$$ Therefore, $\sqrt{t_T}(\bar y_T^E-Z_T^E) \Rightarrow N(0,16/9)$. An exact recursion for the joint second moments of the synchronized iterate and the two running sums gives $t_T\operatorname{Var}(\bar y_T)=2.894$, $2.819$ and $2.801$, and $t_T\operatorname{Var}(\bar y_T^E)=2.879$, $2.816$ and $2.799$, at $T=5{,}000$, $50{,}000$ and $200{,}000$. Both approach the limit $25/9\approx2.778$ and stay far from $t_T\operatorname{Var}(Z_T^E)=1$. We will add these values to Appendix B.7. Table R3 compares the two averages on the same sample paths in the Linear setting ($T=50{,}000$, 500 replications). Besides $C_1$, $C_5$, Log and AltBlock, it includes a schedule Alt that alternates $E_m=1$ and $E_m=9$ every round after the warm-up, so that its $\nu_T$ is close to that of AltBlock. The normalized MSE, which equals $1$ for an unbiased estimator with covariance $\Sigma/t_T$, is close to $\nu_T$ for the equal-round average under every schedule. Weighting by $E_m$ reduces it by $11\\%$ under $C_5$, where $E_m$ changes only once, at the end of the warm-up, and by $36\\%$ under Log and AltBlock, where $E_m$ is constant over long blocks of rounds. Under Alt, the two averages have nearly the same normalized MSE (ratio $0.996$). Table R3. Equal-round and $E_m$-weighted averages in the Linear setting ($K=10$, $d=5$, $T=50{,}000$, 500 paired replications). Both averages include all rounds and use the same sample paths. MSE pairs are equal-round/$E_m$-weighted. Normalized MSE is $t_T\\,\mathbb E\\{(\hat x-x^\star)^\top\Sigma^{-1}(\hat x-x^\star)\\}/d$, computed with the true $\Sigma$ of each replication and averaged over replications; it includes the finite-sample bias. Ratio is the normalized MSE of the $E_m$-weighted average divided by that of the equal-round average. Alt alternates $E_m=1$ and $E_m=9$ every round after the $2{,}500$-round warm-up. |Schedule|$\nu_T$|MSE:equal/weighted|Ratio| |---|---|---|---| |$C_1$|1.000|1.018/1.018|1.000| |$C_5$|1.152|1.127/0.998|0.886| |Log|1.622|1.641/1.044|0.636| |AltBlock|2.766|2.726/1.733|0.636| |Alt|2.773|2.724/2.714|0.996|
Reviewer_EeQqOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper constructs confidence regions that can be monitored repeatedly during Local-SGD training. For Polyak-Ruppert averages under full participation, it proves an asymptotic Gaussian approximation that accounts for varying communication intervals. The construction combines time-uniform Gaussian bounds with covariance estimation. Experiments study coverage and interval width in simulations and a public-health application.

Critical strengths

• The paper addresses the useful question of how to make valid inference when training is monitored repeatedly rather than stopped at a fixed round. The analysis explicitly accounts for drift from local updates and errors from freezing the noise process.

• Keeping the exact normalization for the communication schedule makes the treatment of varying intervals more informative. The experiments compare oracle and estimated covariance and vary client count and heterogeneity.

Critical weaknesses

• The covariance estimator is proved consistent using fresh samples independent of training, but the experiments reuse training data. Does consistency still hold with this reuse? The current proof does not cover the estimator used in the experiments.

• The coverage theorem requires a positive inflation chosen in advance, whereas every experiment uses zero inflation. Please show the coverage and interval widths with positive inflation and explain how its value should be chosen in practice.

• In Table B.2, with 50 clients and 4,000 rounds, 11% of runs miss the true parameter at least once during monitoring, exceeding the nominal 5%. This occurs even with the true covariance. Under what training budgets does coverage approach the stated level?

Questions for authors

Please see weaknesses.

5 comments

Thank you for your time and valuable comments.

W1. We will revise Proposition 1 and its proof to cover the completed-round sample-reuse construction described in Section 3.1. We state this construction and its strong-consistency argument explicitly below.

For round mm, define

\quad g_m=\sum_{k=1}^K p_k\nabla f_k(\bar x_{t_m};\xi_{t_m}^k).$$ Let $$\widehat G_T=\frac1T\sum_{m=0}^{T-1}H_m, \qquad \widehat S_T=\frac1T\sum_{m=0}^{T-1}g_mg_m^\top,$$ and $$\widehat\Sigma_T= \begin{cases} \widehat G_T^{-1}\widehat S_T\widehat G_T^{-\top}, &\widehat G_T\text{ is nonsingular},\\\\ 0,&\widehat G_T\text{ is singular}. \end{cases}$$ Thus all $T$ completed rounds, including warm-up, enter the estimator. Assumption 5 (Covariance estimation). All gradients and sample Hessians in the construction are defined and finite at the evaluated parameters. There is a neighborhood $\mathcal N$ of $x^\star$ such that, for every client $k$, the sample Hessian exists on $\mathcal N$ and differentiation can be interchanged with expectation: $$\mathbb E\\{\nabla^2 f_k(x;\xi^k)\\}=\nabla^2 f_k(x), \qquad x\in\mathcal N, \qquad G=\sum_{k=1}^Kp_k\nabla^2 f_k(x^\star).$$ Moreover, $\mathbb E\\|\nabla^2f_k(x^\star;\xi^k)\\|_F^2 < \infty,$ and, for some finite $L_H$ and all $x,y\in\mathcal N$, $$\left\\{\mathbb E \\|\nabla^2f_k(x;\xi^k)-\nabla^2f_k(y;\xi^k)\\|_F^2 \right\\}^{1/2} \le L_H\\|x-y\\|.$$ In the same neighborhood, $\sup_{x\in\mathcal N} \mathbb E\\|\nabla f_k(x;\xi^k)\\|^4 < \infty.$ The Frobenius norm is used for matrix moment bounds; in fixed dimension it is equivalent to the operator norm used in the manuscript. Proposition 1 (Strong consistency). Suppose Assumptions 1-5 hold. Then, $$\widehat G_T\longrightarrow G, \qquad \widehat S_T\longrightarrow S, \qquad \widehat\Sigma_T\longrightarrow\Sigma \qquad\text{almost surely},$$ where $$S=\sum_{k=1}^Kp_k^2S_k, \qquad S_k=\operatorname{Cov}\\{\nabla f_k(x^\star;\xi^k)\\}, \qquad \Sigma=G^{-1}SG^{-\top}.$$ Almost surely, there exists a finite $T_0$ such that, for all $T\ge T_0$, $$\lambda_{\min}\\!\left( \frac{\widehat G_T+\widehat G_T^\top}{2} \right) \ge\frac12\lambda_{\min}(G) > 0, \qquad \lambda_{\min}(\widehat S_T) \ge\frac12\lambda_{\min}(S) > 0, \qquad \widehat\Sigma_T\succ0.$$ Proof of Proposition 1. Let $Q_m=g_mg_m^\top$. Choose $r > 0$ within the common neighborhoods in Assumptions 2 and 5, and set $J_m=\mathbf 1\\{\\|\bar x_{t_m}-x^\star\\|\le r\\}$. Both $\bar x_{t_m}$ and $J_m$ are $\mathcal F_{t_m}$-measurable, and $J_m=1$ eventually almost surely by Assumption 4. Conditional on $\mathcal F_{t_m}$, the observations $\xi_{t_m}^k$ are independent with laws $P_k$. Assumption 5 and Jensen's inequality therefore give $$\sup_m\mathbb E\bigl[J_m\\|H_m\\|_F^2+J_m\\|g_m\\|^4\bigr] < \infty.$$ For $Z_m=H_m$ or $Q_m$, define \$ D_m^Z=J_mZ_m-\\mathbb E[J_mZ_m\\mid\\mathcal F_{t_m}]. \$ These are $\mathcal F_{t_m+1}$-measurable, have conditional mean zero given $\mathcal F_{t_m}$, and satisfy $$\sum_{m=0}^\infty \frac{\mathbb E\\|D_m^Z\\|_F^2}{(m+1)^2} < \infty.$$ Since $\mathcal F_{t_m+1}\subseteq\mathcal F_{t_{m+1}}$, martingale convergence on the round filtration and Kronecker's lemma yield, entrywise, $$\frac1T\sum_{m=0}^{T-1}D_m^Z\longrightarrow0 \qquad\text{almost surely}.$$ For independent $\xi^k\sim P_k$, define locally $$g(x;\xi)=\sum_{k=1}^Kp_k\nabla f_k(x;\xi^k),\qquad h(x)=\sum_{k=1}^Kp_k\nabla^2f_k(x),\qquad V(x)=\mathbb E[g(x;\xi)g(x;\xi)^\top].$$ Extend $h$ and $V$ by zero outside this neighborhood. Then $$\mathbb E[J_mH_m\mid\mathcal F_{t_m}]=J_mh(\bar x_{t_m}), \qquad \mathbb E[J_mQ_m\mid\mathcal F_{t_m}]=J_mV(\bar x_{t_m}).$$ Independence across clients and unbiasedness give, locally, $$V(x)=\sum_{k=1}^Kp_k^2 \operatorname{Cov}\\{\nabla f_k(x;\xi^k)\\} +\nabla f(x)\nabla f(x)^\top.$$ Thus $h(x^\star)=G$ and $V(x^\star)=S$, since $\nabla f(x^\star)=0$. Locally, Assumption 5 gives $\\|h(x)-G\\|_F\le C\\|x-x^\star\\|$, while Assumption 2 and Cauchy–Schwarz imply $$\\|V(x)-S\\|_F \le \left\\{\mathbb E\\|g(x;\xi)-g(x^\star;\xi)\\|^2\right\\}^{1/2} \left\\{\mathbb E\bigl(\\|g(x;\xi)\\|+\\|g(x^\star;\xi)\\|\bigr)^2 \right\\}^{1/2}\le C\\|x-x^\star\\|.$$

By xˉtm→x⋆\bar x_{t_m}\to x^\star and Cesàro averaging, the averages of the two conditional means converge almost surely to GG and SS. Combining this with the martingale limits proves convergence of the localized averages. Since Jm=1J_m=1 eventually and all evaluations are finite, removing JmJ_m changes each average by Oa.s.(T−1)O_{\mathrm{a.s.}}(T^{-1}). Hence G^T→G\widehat G_T\to G and S^T→S\widehat S_T\to S almost surely.

As G≻0G\succ0 and S≻0S\succ0, Weyl's inequality gives the stated eventual eigenvalue bounds. Continuity of inversion then yields Σ^T=G^T−1S^TG^T−⊤→G−1SG−⊤=Σ\widehat\Sigma_T =\widehat G_T^{-1}\widehat S_T\widehat G_T^{-\top} \to G^{-1}SG^{-\top}=\Sigma almost surely.

This argument establishes strong consistency for completed-round sample reuse directly. Further define

\widehat\Sigma_T^{\mathrm{sym}} =\frac{\widehat\Sigma_T+\widehat\Sigma_T^\top}{2}, \qquad \widetilde\Sigma_T= \begin{cases} \widehat\Sigma_T^{\mathrm{sym}}, &\lambda_{\min}(\widehat\Sigma_T^{\mathrm{sym}}) > 0,\\\\ I_d,&\text{otherwise}. \end{cases}$$ Then $\widetilde\Sigma_T$ is $\mathcal F_{t_T}$-measurable, positive definite for every $T\ge1$, and converges almost surely to $\Sigma$. It therefore satisfies the covariance-estimation requirements of Corollary 2. With $\mathcal H_T=\mathcal F_{t_T}$, it supplies the covariance condition in Corollary 2 under the other stated assumptions of that corollary. **W2.** Corollary 2 is stated for a fixed $\delta > 0$, whereas the regions reported in Section 3 use $\delta=0$ to evaluate the first-order construction. We have run experiments with positive inflation $\delta\in\\{0.01,0.02,0.05,0.10\\}$ to quantify the trade-off between coverage and width. On every sample path, a larger $\delta$ gives a larger region and multiplies every half-width by exactly $1+\delta$, so failure (the time-uniform type-I error) is nonincreasing in $\delta$. In the main simulations, failure stays below the nominal level for every $\delta$, and its maximum falls from $4.8\\%$ at $\delta=0$ to $1.0\\%$ at $\delta=0.10$. Representative results for the $C_1$ schedule are shown below. |Model|$\delta=0.01$|$\delta=0.05$|$\delta=0.10$| |---|---|---|---| |Linear|2.8/15.452|1.6/16.064|1.0/16.829| |Logistic|4.2/15.315|2.6/15.921|0.8/16.679| |Poisson|2.0/5.643|1.0/5.867|0.8/6.146| Plug-in covariance, $K=10$, $T=50{,}000$, $d=5$, 500 replications, GM boundary, $n_0=20{,}000$. Cells give failure (%)/terminal half-width ($10^{-3}$). Corollary 2 holds for any fixed $\delta > 0$; in the proof, $\delta$ absorbs coupling, Brownian interpolation and covariance-estimation errors that vanish almost surely. In finite samples, $\delta$ multiplies every half-width by exactly $1+\delta$, so a larger $\delta$ is more conservative. We therefore recommend a small $\delta$, chosen by the extra width that is acceptable and fixed before training together with the monitoring start. We will add these results and clarify the choice of $\delta$.

W3. Corollary 2 is a limit result as the number of rounds grows. The 1111\\% pathwise failure in Table B.2 with K=50K=50 and T=4,000T=4{,}000 occurs in the early monitored rounds, where the initialization bias of yˉT\bar y_T is still large relative to the radius: the bias decays as T−1T^{-1}, whereas the radius decays as T−1/2T^{-1/2}. This is why the excess appears with the true covariance as well. Failure counts a noncoverage at any round after the monitoring start.

Because the excess comes from the early monitored rounds, coverage depends on the budget through the monitoring start, so we vary both on the same runs using 500 new replications (Table R1). With the start fixed at round 401, failure stays near 1010\\% up to T=20,000T=20{,}000. With monitoring from 1010\\% of the rounds, so that the start moves later with the budget, oracle failure is 9.69.6\\%, 7.27.2\\% and 4.84.8\\% at T=4,000T=4{,}000, 8,0008{,}000 and 20,00020{,}000, and in this setting it is at or below 55\\% once monitoring starts at round 1,601. Coverage therefore approaches the stated level once monitoring starts after the initialization bias has decayed.

Table R1: Number of rounds and monitoring start in the Linear setting. Cells give oracle/plug-in values: failure (%) by monitoring start, then terminal half-width (10−310^{-3}).

TTRound 40110% of TT20% of TT40% of TTHalf-width
4,0009.6/8.09.6/8.07.0/6.04.6/3.84.633/4.660
8,0009.8/8.27.2/6.25.0/4.23.2/2.63.268/3.279
20,0009.8/8.24.8/4.02.4/2.60.6/0.62.162/2.166

K=50K=50, h=0h=0, d=5d=5, 500 replications, GM boundary, δ=0\delta=0, zero initialization, warm-up of 200 rounds, n0=v401v4000n_0=\sqrt{v_{401}v_{4000}}.

Averaging after the warm-up brings failure to 4.24.2\\% already at T=4,000T=4{,}000 with the plug-in covariance, using the same post-warm-up rounds for the average, variance clock and covariance estimator. Here the schedule is C5C_5, K=50K=50, h=0h=0, d=5d=5, b=200b=200 warm-up rounds, monitoring starts at 1010\\% of the rounds, δ=0\delta=0, the GM boundary is used, and the terminal half-width is 4.587×10−34.587\times10^{-3} (500 replications). The GM design point is the geometric mean of vT0−vbv_{T_0}-v_b and vT−vbv_T-v_b, where T0T_0 is the monitoring start. In the main simulations, failure also stays below 55\\% for every monitoring start examined.

Thank you again for your valuable feedback. We hope our responses have addressed your concerns and welcome any further questions or discussion.

Comments on the paper

1 comment

Thank you for your time and valuable comments.

We clarified the completed-round covariance estimator and its strong-consistency conditions, the role of fixed positive inflation and the monitoring start, and the relationship between communication schedules, the variance clock, and iterate weighting. We also clarified the interpretation of the real-data results and justified the applicability of our theory to projected Poisson regression. The numerical discussion now brings representative inflation and matrix-free contrast results into the main text, alongside supporting comparisons of initialization, client heterogeneity, higher dimensions, partial participation, and communication and covariance costs.

These changes will be incorporated into the revised manuscript, with the additions and modifications highlighted in blue. We hope our responses address your concerns and welcome further questions or discussion. Thank you again for your valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.