acceptodds
Under review as a conference paper at ICLR 2027

Finite-Shot Quantum Federated Learning: Trajectory-Dependent Update Discrepancies and Nonconvex Convergence under Proximal Dynamics

Abstract

We study quantum federated learning with variational quantum classifiers under finite-shot measurements, heterogeneous clients, multiple local updates, and partial participation. Finite-shot parameter-shift gradients are evaluated along recursively evolving proximal client trajectories, so their effect cannot be represented by simply adding an O(1/S) variance term to a standard proximal federated bound. We therefore analyze the trajectory-dependent discrepancy between the actual federated update and an ideal global-gradient step. At the circuit level, we establish unbiasedness of the parameter-shift estimator, coordinatewise variance O(1/S), and gradient mean-squared error . At the optimization level, we establish first-order stationarity guarantees for smooth nonconvex objectives with time-varying learning rates and proximal coefficients, without a Polyak–\Lojasiewicz condition. The analysis identifies two finite-shot effects: a direct stochastic-gradient residual and an indirect trajectory-mediated residual. It also shows that heterogeneity does not create an independent common-iterate error; instead, it acts through subsequent client drift. For uniform sampling without replacement, we derive an additional correction proportional to and terminal drift energy, which vanishes under full participation. Experiments confirm near-inverse-shot estimator scaling and illustrate the resulting drift and stationarity behavior.

Primary area:
optimization
Keywords:
quantum federated learning, finite-shot measurements, parameter-shift gradients, nonconvex optimization, proximal federated learning, client heterogeneity

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 4 reviews, above 27% of the others.
Reviewer_A9pDOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript considers the problem of quantum federated learning, but its theoretical contribution is not clearly established.

  • the convergence analysis in Section 3.2 does not explain which assumptions, proof arguments, or conclusions depend on properties specific to the quantum setting.
  • The manuscript also lacks an explicit comparison with relevant convergence results for classical federated learning and stochastic optimization. Thus, it is difficult to determine whether the analysis provides new theoretical insights or follows from established results under an adapted formulation. The relationship to prior work should be more explicit.
Critical strengths

see the summary.

Critical weaknesses

see the summary.

Questions for authors

see the summary.

4 comments
AuthorsOpenReview ↗

Response to Reviewer A9pD: Question 1

We thank the reviewer for raising important questions about the theoretical contribution, its relationship to classical stochastic federated optimization, and the role of quantum measurements. The revised manuscript makes explicit a three-level structure that was already present in the submitted analysis: circuit-level finite-shot estimation (Lemma 3.1), the complete-gradient oracle condition (Assumption 3.4), and error propagation through proximal federated trajectories (Lemmas 3.8--3.10, Theorem 3.11, Corollary 3.12, and Theorem 3.15). The revised Introduction, Table 1, and Sections 3.1--3.2 now state this structure and its relationship to classical work directly. All results and evidence cited below are contained in the revised manuscript.

R1. Which assumptions, proof steps, and conclusions are quantum-specific?

The manuscript first links the shot budget to circuit-derivative estimation error and then analyzes the federated recursion under a separately stated complete-gradient oracle condition.

Circuit-level finite-shot estimation. For bounded observables and independently measured parameter shifts, Lemma 3.1 establishes that each finite-shot circuit-expectation derivative is unbiased with coordinatewise variance O(1/S)O(1/S) and that the corresponding dqd_q-coordinate derivative vector has mean-squared error O(dq/S)O(d_q/S). This result connects the physical measurement budget to circuit-derivative estimation error. We do not claim that parameter-shift estimation itself is new; it supplies the finite-shot specialization used in the subsequent federated analysis.

Complete-gradient oracle condition. An unbiased circuit-expectation derivative does not necessarily yield an unbiased complete gradient for an arbitrary nonlinear loss. Assumption 3.4 therefore separately requires conditional unbiasedness and bounded variance for the complete stochastic local-gradient oracle. For example, a nonlinear finite-shot plug-in estimator can have the form g^q=J^z⊤h(z^,y),\widehat{\mathbf g}_q = \widehat{\mathbf J}_{\mathbf z}^{\top} \mathbf h(\widehat{\mathbf z},y), for which unbiasedness of the readout and Jacobian estimators does not generally imply E[g^q]=Jz⊤h(z,y).\mathbb E[\widehat{\mathbf g}_q] = \mathbf J_{\mathbf z}^{\top}\mathbf h(\mathbf z,y). Theorem 3.11 is therefore conditional on Assumption 3.4; Lemma 3.1 alone does not establish that condition for every finite-shot cross-entropy implementation. Section 3.1 and Appendix C.6.1 state this limitation and the corresponding biased-oracle model. The manuscript does not claim an unconditional stationarity theorem for arbitrary nonlinear gradient bias.

Proximal federated trajectory analysis. Once a valid complete-gradient oracle has been specified, the smoothness, descent, Young's-inequality, finite-population, and telescoping arguments are classical and are not claimed to be uniquely quantum. The contribution is the derivation and control of the error and drift terms produced when stochastic gradients are evaluated repeatedly along heterogeneous proximal client trajectories:

  • Lemma 3.8 derives the schedule-dependent proximal-memory factor At,τ=∑r=0τ−1(1−ηtμt)rA_{t,\tau} = \sum_{r=0}^{\tau-1}(1-\eta_t\mu_t)^r and the resulting client-drift envelope. This factor determines the weights with which earlier stochastic local gradients contribute to the current client displacement.
  • Lemma 3.9 establishes the exact pathwise global-update discrepancy Et=−ηt∑τ=0T−1[N(t,τ)+W(t,τ)],\mathcal E_t = -\eta_t\sum_{\tau=0}^{\mathcal T-1} \left[ \mathbf N^{(t,\tau)} + \mathbf W^{(t,\tau)} \right], where N(t,τ)\mathbf N^{(t,\tau)} is the conditionally centered stochastic-oracle fluctuation and W(t,τ)\mathbf W^{(t,\tau)} contains the trajectory-induced gradient deviation and proximal restoring term. These components are separated by their roles in the update, not by an independence assumption; their dependence is retained when bounding the complete discrepancy.
  • Lemma 3.10 bounds the stochastic-oracle, trajectory-deviation, proximal, and client-drift terms in a one-round descent inequality. Theorem 3.11 then establishes finite-horizon first-order stationarity for smooth, lower-bounded nonconvex objectives without a PL condition while retaining client-drift energy explicitly.
  • Corollary 3.12 closes the drift recursion under the heterogeneity condition. Because weighted client-gradient deviations cancel at the common broadcast iterate, heterogeneity enters through the client trajectories followed after broadcast, rather than through an independent common-iterate O(ζ2)O(\zeta^2) bias.
  • Theorem 3.15 conditions on the realized stochastic client endpoints and derives the partial-participation correction K−mm(K−1) D‾t,T.\frac{K-m}{m(K-1)}\,\overline D_{t,\mathcal T}. Thus, client-sampling error depends on both the number of participating clients and the terminal dispersion of their proximal trajectories. The correction vanishes exactly under full participation.
AuthorsOpenReview ↗

Response to Reviewer A9pD: Questions 1 and 2

Finite-shot uncertainty consequently enters through two channels: directly through current stochastic-oracle fluctuations and indirectly by changing subsequent client iterates and gradient evaluations. We do not claim that the optimization tools used to analyze these effects apply only to quantum stochasticity. QProxFL instantiates the oracle using finite-shot parameter-shift estimation and derives how the resulting uncertainty propagates through the analyzed proximal federated recursion.

The paper also does not claim to characterize every source of quantum-gradient error. Lemma 3.1 treats finite-shot parameter-shift estimation. The primary contribution is the derivation of the proximal-memory factor, exact update-discrepancy decomposition, heterogeneity-dependent drift closure, and endpoint-dispersion participation correction for the federated recursion.

R2. How does this differ from existing classical and quantum optimization results?

The contribution is the joint derivation of proximal memory, stochastic-oracle and trajectory discrepancies, client-drift closure, and endpoint-dependent participation error.

Client drift, nonconvex proximal federated optimization, and partial-participation variance are established classical topics. SCAFFOLD establishes control-variate drift correction; Yuan--Li provide proximal-point-based nonconvex FedProx analysis and a stochastic extension; and FedVARP provides server-side participation-variance reduction. We do not claim novelty for these ingredients individually.

However, replacing a generic stochastic-gradient variance constant by O(dq/S)O(d_q/S) in an existing bound does not by itself produce:

  • the schedule-dependent geometric-memory factor At,τA_{t,\tau};
  • the exact N+W\mathbf N+\mathbf W global-update discrepancy;
  • the drift-explicit stationarity theorem followed by a post-broadcast heterogeneity closure; or
  • the finite-population correction coupled to terminal proximal endpoint dispersion.

These results require first unrolling the stochastic proximal client recursion, then decomposing the resulting server update, and finally bounding the stochastic-oracle, trajectory-deviation, proximal, and client-drift terms before applying descent and telescoping arguments. None of the cited classical results or compared QFL methods jointly derive these four results under finite-shot parameter-shift uncertainty.

The revised manuscript also makes the comparison with quantum work explicit. Sweke et al. establish finite-measurement stochastic-gradient results for hybrid quantum--classical optimization, while Chung et al. study parameter-shift-based quantum federated gradient aggregation. Q-ANCHOR analyzes nonconvex finite-shot quantum FedAvg with ZNE/control-variate hardware-bias correction; Q-ShiftDP studies parameter-shift sensitivity, finite-shot error, and differential privacy; and DUQFL-Prox develops an adaptive deep-unfolded SPSA-based proximal QFL method with empirical drift stabilization.

We do not claim to be the first work on finite-shot QFL or proximal QFL. These methods use different gradient estimators, hardware-bias or privacy mechanisms, optimization objectives, and convergence guarantees. Our distinct result is the analysis of the time-varying stochastic Prox-SGD recursion summarized above. The revised Introduction and Table 1 compare the optimization setting, uncertainty model, local-update dynamics, participation model, and the error and drift terms retained in the convergence guarantees. Our comparisons with these methods are theorem-level; we do not claim empirical superiority based on unmatched implementations.

Estimator and trajectory diagnostics.

The revised manuscript supplements a parameter-shift variance measurement at a fixed circuit checkpoint with a matched experiment that tracks finite-shot and analytic-expectation trajectories. Section 4.1 evaluates the parameter-shift estimator at a fixed circuit checkpoint and obtains a fitted variance exponent ρ^=−0.9817\widehat{\rho}=-0.9817, supporting the near-S−1S^{-1} scaling in Lemma 3.1. This diagnostic concerns circuit-expectation derivatives and does not verify complete-gradient unbiasedness.

Section 4.2 and Appendix D.5 examine the distinct trajectory-level question using a four-qubit, one-layer VQC, K=5K=5 heterogeneous clients, T=4\mathcal T=4 local proximal-SGD steps, full participation, eight communication rounds, eight paired seeds, and shot budgets S∈{32,128,512}S\in\{32,128,512\}. Within each seed, the finite-shot trajectories and their analytic-expectation reference share the initialization, preprocessed features, client partition, scheduled mini-batches, learning-rate sequence, and proximal schedule.

AuthorsOpenReview ↗

Response to Reviewer A9pD: Questions 2

Increasing SS from 3232 to 512512 reduces the mean final squared trajectory discrepancy from 0.1304910.130491 to 0.0788500.078850, a 39.639.6% reduction, with a reduction in every paired seed. An exact update-map counterfactual separates the current-round discrepancy from the discrepancy inherited through earlier parameter trajectories while retaining their cross term. The mean squared history component exceeds the current-round component at the final round for all three shot budgets. Because the final-round learning rate is small, the current-round discrepancy is comparatively small in these runs; we therefore do not infer that history-dependent discrepancies generally dominate current-round discrepancies. We also do not claim componentwise monotonicity in SS.

The manuscript explicitly states that these empirical components are not term-by-term realizations of the analytical components in Lemma 3.9. The complete protocol, decomposition, limitations, and seed-level values appear in Figure 2 and Tables 13--15.

The manuscript also distinguishes the theory-aligned proximal-SGD diagnostics from the practical Adam/AdamW classification experiments. Neither experiment is presented as direct verification of the complete-gradient oracle condition or Theorem 3.11. The experimental stationarity quantity is defined as a finite-shot squared-gradient diagnostic evaluated using S=128S=128 shots on a fixed 32-example subset, rather than the full-objective stationarity quantity in the theorem.

The theory-aligned experiments compare Q-FedAvg, constant Q-FedProx, and annealed Q-FedProx using matched initializations, client partitions, mini-batch schedules, shot budgets, and learning-rate sequences. Appendix D.8 also reports centralized and local-only VQC results while stating explicitly that their circuit-evaluation budgets are not matched to those of the federated methods. These results are not presented as evidence of quantum advantage, intrinsic federated superiority, or uniform superiority of either proximal schedule.

To improve accessibility, the revised Introduction presents the measurement-constrained QFL problem before the detailed notation, explains the PL condition at first use, separates circuit-derivative estimation from the complete-gradient oracle condition, and states explicitly that finite-shot uncertainty acts directly through current stochastic-gradient errors and indirectly by changing subsequent client iterates and gradient evaluations. The revision also standardizes the terminology for the MNIST and six-class MedMNIST experiments.

AI-use disclosure.

Following the Program Chairs' clarification regarding AI-use reporting, the revised manuscript includes the requested disclosure:

Generative AI tools were used for language editing, proofreading, manuscript organization, presentation refinement, and feedback on the clarity and positioning of the work. The authors developed the research problem, theoretical models and claims, proofs, algorithms, experimental methodology and implementation, and data analyses. The authors independently verified all mathematical arguments, numerical results, citations, comparisons, and AI-assisted text, and take full responsibility for the final content.

This disclosure distinguishes language and presentation assistance from the scientific contributions and confirms the authors' responsibility for all mathematical, experimental, comparative, and bibliographic content.

AuthorsOpenReview ↗

Response to Reviewer A9pD: Follow-up

As the discussion period concludes tomorrow, we wanted to follow up on our response and the revised manuscript. We clarified which components are quantum-specific, which proof tools are classical, and what is newly derived for the time-varying stochastic proximal recursion. The revision also states the conditional scope of Theorem 3.11 explicitly, provides detailed comparisons with the closest classical and QFL results, reports matched finite-shot/analytic trajectory experiments, and includes the requested AI-use disclosure.

We would greatly appreciate any feedback on whether these revisions address your concerns and remain happy to clarify any outstanding point before the discussion period ends.

Reviewer_UHq3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents a convergence analysis of finite-shot quantum federated learning with variational quantum classifiers, heterogeneous clients, proximal local updates, and partial participation. It derives nonconvex first-order stationarity bounds without assuming the Polyak–Łojasiewicz condition, distinguishing direct finite-shot gradient uncertainty from its indirect effects through evolving client trajectories. The analysis further characterizes the influence of client heterogeneity, proximal regularization, and partial participation on convergence. Experiments validate inverse-shot gradient-variance scaling and investigate proximal dynamics, complemented by federated classification experiments on MNIST and medical-image datasets.

Critical strengths

Comprehensive theoretical analysis

The paper develops a unified convergence framework accounting for finite-shot measurement uncertainty, heterogeneous client trajectories, proximal regularization, and partial participation. Establishing nonconvex first-order stationarity without a PL assumption is a useful extension of existing QFL convergence analyses, particularly because the bounds explicitly distinguish direct stochastic-gradient errors from trajectory-mediated effects.

Careful separation of theoretical assumptions and empirical evidence

The authors explicitly distinguish unbiased circuit-level parameter-shift estimation from the stronger assumptions required for complete loss gradients. They also separate theory-aligned proximal-SGD diagnostics from practical Adam/AdamW experiments and acknowledge statistically inconclusive results. This transparency strengthens the credibility and interpretation of the analysis.

Critical weaknesses

Insufficient motivation and difficult presentation

The paper provides little motivation for quantum federated learning or why variational quantum circuits should be used instead of classical models in the first place. This makes the significance of the investigated problem difficult to assess for a general ICLR audience. Moreover, the presentation is highly technical and often difficult to follow without substantial background in quantum computing and optimization. Mathematical expressions are introduced already in the introduction with little explanation of their variables, while important concepts such as the Polyak–Łojasiewicz condition are insufficiently introduced. The large number of individual contributions, dense mathematical exposition, and small figures further reduce accessibility. The experimental description also contains ambiguities, particularly the apparent confusion between the MedNIST and MedMNIST datasets.

Insufficient positioning against prior work and unclear quantum-specific contribution

The paper does not adequately distinguish its theoretical contributions from existing results in classical stochastic federated optimization. Client drift under heterogeneous objectives [1], nonconvex FedProx convergence without PL assumptions [2], and variance from partial client participation [3] have already been studied extensively. Finite-shot stochastic-gradient convergence in quantum optimization [4] and parameter-shift-based quantum federated learning [5] are also relevant. More fundamentally, the main convergence analysis relies on generic stochastic-gradient assumptions, with the quantum-specific contribution largely entering through the known inverse-shot variance scaling. It remains unclear which theoretical results are specific to quantum measurements rather than applicable to arbitrary stochastic gradient estimators. Furthermore, unbiasedness of the complete loss gradient is assumed but not established for the implemented nonlinear finite-shot classifiers, limiting the direct applicability of the convergence guarantees.

[1] Karimireddy, Sai Praneeth, et al. "SCAFFOLD: Stochastic Controlled Averaging for Federated Learning." Proceedings of the 37th International Conference on Machine Learning (ICML). PMLR, 2020.

[2] Yuan, Xiaotong, and Ping Li. "On Convergence of FedProx: Local Dissimilarity Invariant Bounds, Non-smoothness and Beyond." Advances in Neural Information Processing Systems (NeurIPS) 35 (2022).

[3] Jhunjhunwala, Divyansh, et al. "Fedvarp: Tackling the variance due to partial client participation in federated learning." Proceedings of the 38th Conference on Uncertainty in Artificial Intelligence (UAI). PMLR, 2022.

[4] Sweke, Ryan, et al. "Stochastic gradient descent for hybrid quantum-classical optimization." Quantum 4 (2020): 314.

[5] Chung, Jaehyun, et al. "Quantum Federated Gradient Aggregation Using the Parameter-Shift Rule." IEEE Internet of Things Journal (2026).

Questions for authors

1. Motivation and practical relevance (clarification)

What motivates the use of quantum models in federated learning rather than conventional classical models? Are there particular applications or properties of variational quantum circuits that make the proposed analysis especially relevant? A clearer motivation would help establish the significance of the work for a broader ML audience.

2. Novelty and quantum-specific contributions (clarification)

How do the theoretical results differ from existing analyses of classical stochastic federated learning, particularly regarding client drift [1], nonconvex FedProx convergence [2], and partial participation [3]? Since the convergence analysis relies on generic stochastic-gradient assumptions, which results fundamentally depend on quantum measurements rather than applying equally to arbitrary stochastic-gradient estimators? How does the analysis extend existing finite-shot quantum optimization results [4,5]?

3. Applicability of the convergence guarantees (clarification and additional evidence)

The main convergence theorem assumes conditionally unbiased stochastic gradients of the complete loss, while the paper acknowledges that unbiased parameter-shift estimates do not necessarily yield unbiased gradients for nonlinear losses such as cross-entropy. Does the gradient estimator used in the experiments satisfy this assumption? If not, could the authors clarify which theoretical guarantees apply to the implemented classifiers and whether the bias-aware extension can account for the resulting finite-shot gradient bias?

8 comments
AuthorsOpenReview ↗

Response to Reviewer UHq3: Question 1

We thank the reviewer for the careful reading and for recognizing both the paper's joint analysis of stationarity, client drift, and partial participation and its separation of theoretical oracle assumptions from empirical evidence. The revised manuscript makes the motivation, relationship to classical federated optimization, scope of the complete-gradient oracle condition, and limitations of the experimental evidence explicit. All arguments and results cited below are contained in the revised manuscript.

R1. Why study quantum models in federated learning?

We do not claim that VQCs are intrinsically preferable to classical models or that our experiments establish quantum advantage. The paper studies a measurement-constrained distributed optimization problem that arises when variational quantum or hybrid quantum--classical models are trained across clients holding decentralized data.

For a circuit with dqd_q trainable parameters, parameter-shift gradient estimation requires shifted circuit evaluations, and each circuit expectation on quantum hardware must be estimated from finitely many measurements. The shot budget SS is therefore both a physical execution resource and a source of stochastic uncertainty. In federated training, clients repeatedly evaluate these estimates at client-specific models over multiple local steps before aggregation. A finite-shot perturbation can consequently change not only the current gradient estimate but also the model at which every subsequent local gradient is evaluated.

The central question is therefore:

How does finite-shot measurement uncertainty propagate when heterogeneous clients perform multiple proximal updates before aggregation?

This question is meaningful independently of any claim of quantum speedup, predictive superiority, or computational advantage. The revised Introduction now states this motivation before the detailed notation, and Appendix D.9 explicitly calibrates the empirical claims and excludes a quantum-advantage interpretation.

The centralized and local-only VQC results in Appendix D.8 compare training architectures, but their circuit-evaluation budgets are not matched to those of the federated methods. They are not presented as evidence of quantum advantage or as compute-matched comparisons with classical representation-learning models.

AuthorsOpenReview ↗

Response to Reviewer UHq3: Question 2

R2. Which parts are quantum-specific, and how does the analysis differ from classical federated optimization?

The manuscript separates the quantum measurement interface from the classical optimization tools used to analyze the resulting distributed recursion.

At the circuit level, Lemma 3.1 provides the quantum measurement interface. For bounded observables and independent shifted measurements, it establishes unbiased finite-shot parameter-shift estimation of each circuit-expectation derivative, coordinatewise variance O(1/S)O(1/S), and dqd_q-coordinate mean-squared error O(dq/S)O(d_q/S). This result connects the physical measurement budget to an estimator-level stochastic-error model.

Once a valid complete-gradient oracle has been specified, the subsequent smoothness, descent, and finite-population tools are not claimed to be uniquely quantum. The contribution is the analysis of how the stated oracle uncertainty propagates through an explicitly unrolled, time-varying stochastic proximal federated recursion. Specifically:

  • Lemma 3.8 derives the schedule-dependent proximal-memory factor At,τ=∑r=0τ−1(1−ηtμt)rA_{t,\tau} = \sum_{r=0}^{\tau-1}(1-\eta_t\mu_t)^r and the corresponding client-drift envelope. This quantity makes explicit how the proximal schedule controls the persistence of earlier stochastic local gradients.
  • Lemma 3.9 establishes the exact global-update discrepancy Et=−ηt∑τ=0T−1[N(t,τ)+W(t,τ)],\mathcal E_t = -\eta_t\sum_{\tau=0}^{\mathcal T-1} \left[ \mathbf N^{(t,\tau)} + \mathbf W^{(t,\tau)} \right], separating the conditionally centered stochastic-oracle fluctuation from the trajectory-induced gradient deviation and proximal restoring term.
  • Theorem 3.11 establishes finite-horizon first-order stationarity for smooth, lower-bounded nonconvex objectives without a PL condition while retaining client-drift energy explicitly. Corollary 3.12 then closes the drift recursion under the heterogeneity condition.
  • Theorem 3.15 conditions on the realized stochastic client endpoints and couples partial participation to terminal endpoint dispersion through cm=K−mm(K−1).c_m=\frac{K-m}{m(K-1)}. The corresponding participation correction vanishes exactly under full participation.

These results expose direct stochastic-oracle and trajectory-mediated finite-shot contributions. They also show that, under full participation, weighted client-gradient deviations cancel at the common broadcast model. Heterogeneity therefore enters through the distinct trajectories followed after clients leave the broadcast iterate, rather than through an independent common-iterate O(ζ2)O(\zeta^2) bias.

We agree that stochastic gradients, client drift, proximal updates, and participation variance are established classical topics. SCAFFOLD controls drift through control variates; Yuan--Li analyze nonconvex FedProx through proximal-point methods and a stochastic extension; and FedVARP reduces partial-participation variance using stored server updates.

These works establish control-variate drift correction, proximal-point-based FedProx analysis, and server-side participation variance reduction, respectively. We do not claim novelty for these ingredients individually. However, none of these classical results or the compared QFL methods jointly derive the schedule-dependent proximal-memory envelope, the N+W\mathbf N+\mathbf W update-discrepancy decomposition, the post-broadcast heterogeneity closure, and the partial-participation correction coupled to terminal endpoint dispersion under finite-shot parameter-shift uncertainty. Merely replacing a generic variance constant with O(dq/S)O(d_q/S) in an existing federated convergence bound does not produce these four results.

Similarly, Sweke et al. establish finite-measurement stochastic-gradient results for hybrid quantum--classical optimization, while Chung et al. study parameter-shift-based quantum federated gradient aggregation. These works motivate the circuit-level interface but do not derive the proximal-trajectory, heterogeneity, and endpoint-dispersion results above. The revised Introduction and Table 1 make these classical and quantum comparisons explicit and also distinguish QProxFL from Q-ANCHOR, Q-ShiftDP, and DUQFL-Prox. We do not claim that finite-shot QFL or proximal QFL is new in isolation.

AuthorsOpenReview ↗

Response to Reviewer UHq3: Question 3

R3. Do the guarantees apply to the nonlinear classifiers used experimentally?

Not unconditionally. Unbiased parameter-shift estimates of circuit-expectation derivatives do not automatically imply an unbiased complete gradient for a nonlinear loss such as cross-entropy. Lemma 3.1 concerns individual circuit-expectation derivatives, whereas Assumption 3.4 separately requires conditional unbiasedness and bounded variance for the complete stochastic local-gradient oracle.

For a nonlinear finite-shot plug-in estimator, g^q=J^z⊤h(z^,y),\widehat{\mathbf g}_q = \widehat{\mathbf J}_{\mathbf z}^{\top} \mathbf h(\widehat{\mathbf z},y), unbiasedness of z^\widehat{\mathbf z} and J^z\widehat{\mathbf J}_{\mathbf z} does not generally imply E[g^q]=Jz⊤h(z,y),\mathbb E[\widehat{\mathbf g}_q] = \mathbf J_{\mathbf z}^{\top}\mathbf h(\mathbf z,y), because of nonlinear composition and statistical interaction. Accordingly, \textbf{we do not claim that the finite-shot nonlinear loss gradients used in either the controlled proximal-SGD experiment or the practical Adam/AdamW classifiers satisfy Assumption 3.4 exactly}. Theorem 3.11 is a conditional stationarity theorem for complete-gradient estimators satisfying that assumption, not an unconditional guarantee for every nonlinear finite-shot classifier.

Section 4.1 provides a separate estimator-level diagnostic: the measured parameter-shift variance follows a near-S−1S^{-1} law with fitted exponent ρ^=−0.9817\widehat{\rho}=-0.9817. This result supports the circuit-expectation derivative scaling in Lemma 3.1, but it does not establish conditional unbiasedness of the complete nonlinear loss gradient.

If the complete-gradient oracle instead has conditional bias, E[gk(t,τ)∣Ft,τ]=∇Fk(θk(t,τ))+bk(t,τ),\mathbb E[ \mathbf g_k^{(t,\tau)} \mid \mathcal F_{t,\tau}] = \nabla F_k(\boldsymbol{\theta}_k^{(t,\tau)}) + \mathbf b_k^{(t,\tau)}, the unbiased-oracle stationarity theorem cannot be applied unchanged. Section 3.1 and Appendix C.6.1 state this biased-oracle model explicitly. Finite-shot nonlinear plug-in bias requires analysis at the complete-gradient level and need not obey the same O(1/S)O(1/S) variance law as an individual parameter-shift derivative. Systematic hardware bias is different again: unlike sampling variance, it does not generally vanish as S→∞S\to\infty. The manuscript therefore does not absorb either effect into Lemma 3.1 or claim an unconditional replacement theorem for arbitrary biased nonlinear gradients.

Estimator and trajectory diagnostics.

The revised manuscript supplements the parameter-shift variance measurement at a fixed circuit checkpoint with a matched experiment that tracks finite-shot and analytic-expectation trajectories. The experiment uses a four-qubit, one-layer VQC, K=5K=5 heterogeneous clients, T=4\mathcal T=4 local proximal-SGD steps per round, full participation, eight communication rounds, eight paired seeds, and shot budgets S∈{32,128,512}S\in\{32,128,512\}. Within each seed, the finite-shot trajectories and their analytic-expectation reference share the initialization, preprocessed features, client partition, scheduled mini-batches, learning-rate sequence, and proximal schedule.

Increasing SS from 3232 to 512512 reduces the mean final squared trajectory discrepancy from 0.1304910.130491 to 0.0788500.078850, a 39.639.6% reduction, with a reduction in every paired seed. An exact update-map counterfactual separates the current-round discrepancy from the discrepancy inherited through earlier parameter trajectories while retaining their cross term. The mean squared history component is larger at the final round for all three tested shot budgets. Because the final-round learning rate is small, we do not interpret this ordering as a universal history-amplification factor. We also explicitly state that these empirical components are not term-by-term realizations of the analytical components in Lemma 3.9. The complete protocol, decomposition, limitations, and seed-level results appear in Section 4.2, Appendix D.5, Figure 2, and Tables 13--15.

Appendix D.4 reports all eight seeds for the theory-aligned proximal diagnostic and the sensitivity to seed 6. Across all seeds, the mean paired terminal-diagnostic changes are −26.28-26.28% for constant proximal regularization and −14.70-14.70% for annealing. Excluding seed 6 changes these means to −16.29-16.29% and −3.24-3.24%, respectively. The median terminal diagnostic values--1.1611.161 for Q-FedAvg, 1.1171.117 for constant Q-FedProx, and 1.2041.204 for annealed Q-FedProx--show much smaller separation than the mean paired percentages, confirming substantial seed-level variability.

For seed 6 specifically, the diagnostic decreases by 96.2396.23% and 94.9594.95% under the constant and annealed schedules, respectively, while test accuracy decreases by 7.007.00 and 6.506.50 percentage points. This demonstrates why a smaller terminal optimization diagnostic should not be interpreted as improved prediction or generalization. We therefore do not claim uniform or predictive superiority of either proximal schedule.

Reviewer_6Sn6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates finite-shot quantum federated learning with heterogeneous clients, multiple local updates, proximal regularization, and partial participation. The authors analyze how quantum measurement uncertainty propagates through client optimization trajectories and derive nonconvex first-order stationarity guarantees. The analysis separates direct stochastic-gradient errors from trajectory-mediated effects and characterizes the influence of client heterogeneity and partial participation.

Critical strengths
  • The paper provides a systematic theoretical analysis of finite-shot uncertainty, heterogeneous client trajectories, and proximal regularization in quantum federated learning.

  • The convergence analysis covers smooth nonconvex objectives without requiring the Polyak–Lojasiewicz condition and explicitly accounts for partial client participation.

Critical weaknesses

The theoretical novelty is limited, with substantial overlap with existing quantum federated learning literature. Q-ANCHOR [1] has already established nonconvex stationarity guarantees without the PL condition under finite-shot quantum gradient uncertainty, heterogeneous clients, multiple local updates, and partial participation. Its convergence analysis explicitly incorporates quantum measurement variance into client-drift bounds, indicating that the propagation of stochastic quantum errors through local optimization trajectories has already been considered. Furthermore, Q-ShiftDP [2] studies finite-shot parameter-shift gradient estimation and establishes gradient estimation error bounds, while DUQFL-Prox [3] has already investigated proximal regularization for controlling client drift in quantum federated learning. Although the present paper provides a proximal-specific convergence analysis, its main ingredients and analytical mechanisms substantially overlap with these works. The paper does not adequately distinguish its theoretical contributions from these closely related results or demonstrate a fundamentally stronger convergence guarantee.

The quantum-specific significance and applicability of the theoretical guarantees remain unclear. The convergence analysis relies on Assumption 3.4, which directly assumes conditional unbiasedness and a bounded variance for the complete stochastic loss-gradient oracle. However, Lemma 3.1 only establishes unbiasedness and variance bounds for circuit-expectation derivatives, which do not automatically extend to nonlinear loss gradients. Consequently, a crucial connection between finite-shot quantum measurements and the optimization-level assumptions is not established for the practical learning objectives. Moreover, the proposed global-update discrepancy decomposition follows standard stochastic federated optimization arguments, while the partial-participation correction uses a classical finite-population sampling identity. It remains unclear what genuinely quantum-specific theoretical insight is obtained beyond incorporating an inverse-shot variance bound into an existing proximal federated optimization framework.

The experimental evidence does not convincingly establish the significance of the proposed trajectory-dependent analysis. The inverse-shot variance experiment primarily confirms the well-known (O(1/S)) behavior of parameter-shift estimators rather than validating the claimed distinction between direct and trajectory-mediated measurement errors. The theory-aligned proximal experiments use only eight paired seeds and a small four-qubit VQC, with the reported confidence intervals for cumulative and peak drift improvements including zero. Although constant proximal regularization improves the stationarity surrogate, this does not directly verify the proposed decomposition of finite-shot effects. Furthermore, the practical classification experiments use Adam/AdamW rather than the Prox-SGD dynamics covered by the convergence theorems, and comparisons against closely related quantum federated approaches such as Q-ANCHOR and DUQFL-Prox are missing. These limitations make it difficult to assess whether the proposed theoretical characterization provides meaningful practical advantages over existing methods.

[1] Hoang M. Ngo, Quan Nguyen, Wanli Xing, and My T. Thai. Q-ANCHOR: Federated Quantum Learning with ZNE-guided Correction. arXiv preprint arXiv:2605.30075, 2026. https://arxiv.org/abs/2605.30075

[2] Hoang M. Ngo, Nhat Hoang-Xuan, Quan Minh Nguyen, Nguyen Hoang Khoi Do, Incheol Shin, and My T. Thai. Q-ShiftDP: A Differentially Private Parameter-Shift Rule for Quantum Machine Learning. Proceedings of the 29th International Conference on Artificial Intelligence and Statistics (AISTATS), PMLR 300:4447–4455, 2026. https://proceedings.mlr.press/v300/ngo26b.html

[3] Shanika Iroshi Nanayakkara and Shiva Raj Pokhrel. A Drift Stable Quantum Federated Learning for Intelligent Services. arXiv preprint arXiv:2607.21647, 2026. https://arxiv.org/abs/2607.21647

Questions for authors
  1. What specific convergence guarantee or theoretical insight cannot be obtained from the existing analyses of Q-ANCHOR and classical stochastic FedProx? Can the authors provide a direct theorem-level comparison?

2How is the conditional unbiasedness of the complete loss-gradient oracle justified for the nonlinear quantum classification objectives used in the experiments?

3.Can the authors experimentally isolate and quantify the direct and trajectory-mediated contributions of finite-shot uncertainty, and compare against more closely related quantum federated learning methods?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 6Sn6: Question 1

We thank the reviewer for recognizing the paper's systematic nonconvex analysis and for requesting a more explicit comparison with Q-ANCHOR, Q-ShiftDP, DUQFL-Prox, and classical federated optimization. The revised Introduction and Table 1 now provide theorem-level comparisons in terms of update rules, oracle assumptions, treatment of client drift and measurement error, heterogeneity assumptions, participation models, and convergence guarantees. All arguments and evidence cited below are contained in the revised manuscript.

We do not claim that finite-shot variance, nonconvex QFL convergence, client drift, proximal regularization, or stochastic-error propagation is individually new. Nor do we claim that classical stochastic-optimization tools are inapplicable to quantum-generated errors. Once conditional unbiasedness and bounded variance are established for the complete-gradient oracle, classical smoothness, descent, and finite-population arguments can be applied. Our contribution is the analysis of an explicitly unrolled, time-varying stochastic proximal recursion. This analysis derives a proximal-memory factor, an exact decomposition of stochastic-oracle and trajectory-induced update discrepancies, a recursive client-drift closure, and a partial-participation correction coupled to terminal client-endpoint dispersion. These terms do not follow merely by replacing a generic variance constant with O(dq/S)O(d_q/S) in a bound that evaluates client gradients at the common broadcast model.

R1. What guarantee or insight is not obtained from Q-ANCHOR or classical stochastic FedProx?

Q-ANCHOR and QProxFL analyze different update structures and derive different propagated quantities. Q-ANCHOR already establishes nonconvex finite-shot QFL convergence with multiple local updates, heterogeneous clients, partial participation, and ZNE/control-variate hardware-bias correction. We therefore do not claim to be the first work on nonconvex finite-shot QFL. QProxFL instead analyzes the time-varying stochastic proximal recursion dk(t,τ+1)=(1−ηtμt)dk(t,τ)−ηtgk(t,τ),\mathbf d_k^{(t,\tau+1)} = (1-\eta_t\mu_t)\mathbf d_k^{(t,\tau)} -\eta_t\mathbf g_k^{(t,\tau)}, which yields the following four coupled results.

  1. Proximal geometric memory. Lemma 3.8 derives At,τ=∑r=0τ−1(1−ηtμt)r,Dt,τ≤ηt2G2At,τ2.A_{t,\tau} = \sum_{r=0}^{\tau-1}(1-\eta_t\mu_t)^r, \qquad D_{t,\tau} \leq \eta_t^2G^2A_{t,\tau}^2. Thus, the proximal schedule explicitly controls how previous stochastic local gradients persist along each client trajectory. In particular, At,τ=τA_{t,\tau}=\tau when μt=0\mu_t=0, whereas At,τ≤min⁡{τ,(ηtμt)−1}when μt>0.A_{t,\tau} \leq \min\{\tau,(\eta_t\mu_t)^{-1}\} \quad\text{when }\mu_t>0.
  2. Exact global-update discrepancy. Lemma 3.9 establishes the pathwise identity Et:=Δ(t)+ηtT∇F(θ(t))=−ηt∑τ=0T−1[N(t,τ)+W(t,τ)],\mathcal E_t := \boldsymbol{\Delta}^{(t)} +\eta_t\mathcal T\nabla F(\boldsymbol{\theta}^{(t)}) = -\eta_t\sum_{\tau=0}^{\mathcal T-1} \left[ \mathbf N^{(t,\tau)} + \mathbf W^{(t,\tau)} \right], where N(t,τ)\mathbf N^{(t,\tau)} is the conditionally centered stochastic-oracle fluctuation and W(t,τ)\mathbf W^{(t,\tau)} contains the trajectory-induced gradient deviation and proximal restoring term. This identity distinguishes the direct effect of finite-shot oracle fluctuations from their indirect effect through subsequent client iterates and gradients.
  3. Drift-explicit and heterogeneity-resolved stationarity. Theorem 3.11 establishes finite-horizon first-order stationarity for smooth, lower-bounded nonconvex objectives without a PL condition, while retaining the client-drift energy explicitly. This theorem does not require the heterogeneity condition in Assumption 3.6. Corollary 3.12 then closes the drift recursion under that condition. Because weighted client-gradient deviations cancel at the common broadcast iterate, heterogeneity enters through post-broadcast client trajectories rather than as an independent common-iterate O(ζ2)O(\zeta^2) bias.
  4. Participation coupled to endpoint dispersion. Under equal client weights and uniform sampling without replacement, Theorem 3.15 conditions first on the realized stochastic client endpoints and derives the correction cmD‾t,T,cm=K−mm(K−1).c_m\overline D_{t,\mathcal T}, \qquad c_m=\frac{K-m}{m(K-1)}. The participation penalty therefore depends jointly on the number of sampled clients and the terminal dispersion of their proximal trajectories, and it vanishes exactly under full participation (m=Km=K).

In summary, no closest result jointly provides \emph{schedule-dependent proximal memory, an exact stochastic-oracle/trajectory discrepancy, a post-broadcast heterogeneity closure, and a partial-participation residual coupled to terminal endpoint dispersion}. This joint characterization, rather than any one ingredient in isolation, is the paper's principal theoretical contribution.

AuthorsOpenReview ↗

Response to Reviewer 6Sn6: Questions 1 and 2

Corollary 3.12 further resolves the finite-shot component σ‾shot,t 2=∑kpkcps,kdqSk,t\overline{\sigma}_{\mathrm{shot},t}^{\,2} = \sum_k p_k\frac{c_{\mathrm{ps},k} d_q}{S_{k,t}} into a direct stochastic-oracle residual and a trajectory-mediated residual weighted by proximal memory and drift accumulation. These are distinct terms in the convergence bound; we do not assume that the underlying random effects are statistically independent.

The distinction from classical federated optimization is likewise specific. SCAFFOLD controls client drift using control variates; Yuan--Li analyze nonconvex FedProx through proximal-point methods and a stochastic extension; and FedVARP reduces partial-participation variance using stored server updates. These works provide essential foundations, but they do not jointly derive the proximal-memory envelope, the N+W\mathbf N+\mathbf W discrepancy, the post-broadcast heterogeneity closure, and the endpoint-dispersion participation residual under finite-shot parameter-shift uncertainty. Substituting O(dq/S)O(d_q/S) for a generic variance constant in an existing FedProx theorem does not by itself produce these coupled quantities.

Q-ShiftDP is directly relevant to parameter-shift sensitivity and finite-shot estimator error, but its target is differential privacy and privacy--utility guarantees rather than heterogeneous multi-step proximal federated stationarity. DUQFL-Prox develops an adaptive, deep-unfolded SPSA-based proximal QFL method and demonstrates empirical drift stabilization; it does not establish the general stochastic Prox-SGD stationarity theorem, exact update-discrepancy identity, heterogeneity closure, or endpoint-participation correction derived here. These distinctions are stated explicitly in the revised Introduction and Table 1.

R2. How is conditional unbiasedness justified for the complete nonlinear loss gradient?

It is imposed as an explicit oracle condition; it is not inferred from circuit-level parameter-shift unbiasedness. Lemma 3.1 establishes unbiasedness and O(1/S)O(1/S) variance for individual circuit-expectation derivatives. Assumption 3.4 separately requires conditional unbiasedness and bounded variance for the complete stochastic local-gradient oracle used in the convergence analysis.

For a nonlinear loss, a finite-shot plug-in estimator can have the form g^q=J^z⊤h(z^,y).\widehat{\mathbf g}_q = \widehat{\mathbf J}_{\mathbf z}^{\top} \mathbf h(\widehat{\mathbf z},y). Even when the readout and Jacobian estimators are individually unbiased, nonlinear composition and their statistical interaction can yield E[g^q]≠Jz⊤h(z,y).\mathbb E[\widehat{\mathbf g}_q] \neq \mathbf J_{\mathbf z}^{\top} \mathbf h(\mathbf z,y). Accordingly, we do not claim that the finite-shot nonlinear loss gradients used in either the controlled proximal-SGD experiment or the practical Adam/AdamW classifiers satisfy Assumption 3.4 exactly. Theorem 3.11 is a conditional optimization guarantee for complete-gradient estimators satisfying that assumption. The experiments are therefore not presented as direct numerical verification of the complete-gradient oracle condition or of Theorem 3.11.

Section 4.1 provides a separate estimator-level diagnostic: the measured parameter-shift variance follows a near-S−1S^{-1} law with fitted exponent ρ^\widehat{\rho} = −0.9817-0.9817. This supports the circuit-expectation derivative scaling in Lemma 3.1, but it does not establish conditional unbiasedness of the complete nonlinear loss gradient.

If the complete-gradient oracle instead has conditional bias, E[gk(t,τ)∣Ft,τ]=∇Fk(θk(t,τ))+bk(t,τ),\mathbb E[ \mathbf g_k^{(t,\tau)} \mid\mathcal F_{t,\tau}] = \nabla F_k(\boldsymbol{\theta}_k^{(t,\tau)}) + \mathbf b_k^{(t,\tau)}, then the unbiased-oracle stationarity theorem cannot be applied unchanged. Section 3.1 and Appendix C.6.1 make this limitation explicit, separate systematic bias from zero-mean finite-shot variation, and explain that increasing the shot budget suppresses sampling variance but does not generally remove nonlinear plug-in or hardware-induced bias. The manuscript does not claim an unconditional replacement theorem for arbitrary biased nonlinear gradients.

AuthorsOpenReview ↗

Response to Reviewer 6Sn6: Question 3

R3. Can the direct and trajectory-mediated effects be quantified experimentally, and are related QFL methods compared?

The revised manuscript reports a matched counterfactual experiment that separates current-round finite-shot effects from differences inherited through earlier trajectories. The experiment uses a four-qubit, one-layer VQC, K=5K=5 heterogeneous clients, T=4\mathcal T=4 local proximal-SGD steps per round, mini-batches of size four, full participation, eight communication rounds, eight paired seeds, and shot budgets S∈{32,128,512}S\in\{32,128,512\}. Within each seed, the finite-shot trajectories and their analytic-expectation reference share the initialization, preprocessed features, client partition, scheduled mini-batches, learning-rate sequence, and annealed proximal schedule.

Let Ψt(S)\Psi_t^{(S)} denote the complete one-round federated update map using SS shots per shifted-circuit evaluation, conditional on the fixed scheduled mini-batches, and let Ψt(∞)\Psi_t^{(\infty)} denote the corresponding analytic-expectation map. The analytic map removes finite-shot measurement sampling while preserving the same model, data, local-update rule, mini-batch schedule, and proximal dynamics; it is not an ideal global-gradient update. Adding and subtracting the analytic map applied to the finite-shot iterate gives δttotal=Ψt(S)(θt(S))−Ψt(∞)(θt(S))⏟δtdirect+Ψt(∞)(θt(S))−Ψt(∞)(θt(∞))⏟δthistory.\boldsymbol{\delta}_t^{\mathrm{total}} = \underbrace{ \Psi_t^{(S)}(\boldsymbol{\theta}_t^{(S)}) - \Psi_t^{(\infty)}(\boldsymbol{\theta}_t^{(S)}) }_{\boldsymbol{\delta}_t^{\mathrm{direct}}} + \underbrace{ \Psi_t^{(\infty)}(\boldsymbol{\theta}_t^{(S)}) - \Psi_t^{(\infty)}(\boldsymbol{\theta}_t^{(\infty)}) }_{\boldsymbol{\delta}_t^{\mathrm{history}}}. The direct term measures the current-round finite-shot/analytic difference from the same server parameter, including propagation through that round's local steps. The history term measures how parameter differences inherited from previous rounds alter the current analytic update. Their interaction is retained: ∥δttotal∥22=∥δtdirect∥22+∥δthistory∥22+2⟨δtdirect,δthistory⟩,\|\boldsymbol{\delta}_t^{\mathrm{total}}\|_2^2 = \|\boldsymbol{\delta}_t^{\mathrm{direct}}\|_2^2 + \|\boldsymbol{\delta}_t^{\mathrm{history}}\|_2^2 + 2\left\langle \boldsymbol{\delta}_t^{\mathrm{direct}}, \boldsymbol{\delta}_t^{\mathrm{history}} \right\rangle, so no orthogonality or independence assumption is imposed.

The final-round means over the eight paired seeds are

SSDirect2\mathrm{Direct}^2History2\mathrm{History}^2Cross\mathrm{Cross}Total2\mathrm{Total}^2
320.0001710.1285220.0017980.130491
1280.0000740.0864690.0022360.088778
5120.0000840.0783460.0004190.078850

These entries are components of squared parameter discrepancies, not gradient-estimator variances. Increasing SS from 3232 to 512512 reduces the mean final squared trajectory discrepancy by 39.639.6%, with a reduction in all eight paired seeds. The mean squared history component exceeds the direct component at the final round for every tested shot budget. Because the final-round learning rate is small, the current-round perturbation is comparatively small in these runs; we do not interpret this ordering as a universal history-amplification factor. We also do not claim componentwise monotonicity: the mean direct component is slightly larger at S=512S=512 than at S=128S=128, and individual trajectories need not be monotone in SS.

This counterfactual decomposition is complementary to, but not term-by-term identical to, Lemma 3.9. The complete protocol, decomposition, and seed-level values appear in Section 4.2, Appendix D.5.1, Figure 2, and Tables 13--14; the separate shot-switch pilot and its limitations are reported in Appendix D.5.2 and Table 15.

The theory-aligned experiments compare Q-FedAvg, constant Q-FedProx, and annealed Q-FedProx under matched dynamics. Appendix D.8 and Table 19 additionally provide structural comparisons with centralized and local-only VQCs while explicitly identifying their unmatched computational budgets. We have not conducted a head-to-head experiment against Q-ANCHOR, Q-ShiftDP, or DUQFL-Prox and therefore make no claim of empirical superiority over those methods. Their estimators, correction mechanisms, objectives, and experimental assumptions differ, so an unmatched numerical comparison would not isolate the contribution of the present analysis. Our comparison with these methods is therefore theorem-level, while our empirical claim is narrower: the controlled matched experiment directly quantifies the trajectory-propagation mechanism analyzed in QProxFL.

AuthorsOpenReview ↗

Response to Reviewer 6Sn6: Follow-up

As the discussion period concludes tomorrow, we wanted to follow up on our response and the revised manuscript. We provided theorem-level comparisons with Q-ANCHOR, Q-ShiftDP, DUQFL-Prox, and classical federated optimization; clarified the complete-gradient oracle condition; and added matched finite-shot/analytic experiments that quantify current-round and history-dependent trajectory discrepancies over eight paired seeds.

We would greatly appreciate any feedback on whether these revisions address your concerns regarding theoretical novelty, oracle validity, and empirical support, and we remain happy to clarify any outstanding point before the discussion period ends.

Reviewer_zJi6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper analyzes finite-shot quantum federated learning with heterogeneous clients and proximal local updates. It shows how measurement noise affects both gradient estimates and later client trajectories, and derives nonconvex convergence bounds covering client drift and partial participation.

Critical strengths

The paper addresses a clear and meaningful problem, and the theoretical analysis is sound and clearly structured. It connects circuit-level estimator properties to nonconvex convergence bounds, separating direct noise effects, trajectory-mediated effects, and partial-participation error. This is the main reason for my positive assessment.

Critical weaknesses
  1. The paper’s main point is that finite-shot noise affects both gradients and later client trajectories. The experiments mainly check gradient variance at a fixed checkpoint and compare training outcomes. They do not show how much each effect matters during training.

  2. The main diagnostic runs for only eight rounds with eight seeds. The drift changes are not statistically significant, and the stationarity gain is sensitive to one difficult seed.

Questions for authors

I would be happy to raise my score if the authors can address the following points clearly and provide supporting evidence:

  1. What exactly do you measure as “stationarity” in the experiments? Is it the full global gradient norm used in the theory, or an estimate from sampled data and measurements?

  2. Can you vary the shot budget during training and show the direct noise effect and the effect through later client updates separately? This is the main point of the paper, and I would like to see it tested directly.

  3. Can you share the results for all eight seeds and show how much the stationarity improvement changes without seed 6? I want to understand whether the benefit holds across runs or mainly comes from this one case.

These are requests to support the current claims, not suggestions for future work.

4 comments
AuthorsOpenReview ↗

Response to Reviewer zJi6: Question 1

We thank the reviewer for the constructive assessment and for identifying three points that required more explicit empirical documentation. The original submission established theoretically that finite-shot uncertainty affects proximal federated learning both directly through the stochastic-gradient oracle and indirectly by perturbing subsequent client trajectories. The revised manuscript now defines the experimental stationarity diagnostic precisely, reports a matched finite-shot/analytic trajectory study, and provides all eight seed-level results together with a sensitivity analysis excluding seed 6. All evidence cited below is contained in the revised manuscript.

R1. What is measured as ``stationarity''?

The experimental diagnostic is not the full-objective stationarity quantity controlled by Theorem 3.11. Theorem 3.11 bounds a learning-rate-weighted average of E[∥∇F(θ(t))∥22\mathbb{E} [ \|\nabla F(\boldsymbol{\theta}^{(t)})\|^2_2 for the complete federated objective. In contrast, for schedule a∈{FedAvg,Const,Anneal}a\in\{\mathrm{FedAvg},\mathrm{Const},\mathrm{Anneal}\} and seed ss, the experiment reports the terminal diagnostic

G^E,s(a):=∥g^E,128(θs,a(T))∥22,∣E∣=32.\widehat{\mathcal G}_{\mathcal E,s}^{(a)} := \left\| \widehat{\mathbf g}_{\mathcal E,128} \left(\boldsymbol{\theta}_{s,a}^{(T)}\right) \right\|_2^2, \quad |\mathcal E|=32.

Here, g^E,128\widehat{\mathbf g}_{\mathcal E,128} is the stochastic loss gradient with respect to the complete hybrid-model parameters, computed using S=128S=128-shot circuit evaluations on the same fixed 32-example subset E\mathcal E for every method and matched seed. Both the proximal-SGD training runs and this terminal diagnostic use S=128S=128 finite-shot circuit evaluations. The diagnostic therefore depends on the selected 3232-example subset and contains additional finite-shot variability from its evaluation.

We use this quantity only for matched finite-horizon comparisons among the three proximal schedules. We do not present it as an estimate of the full-objective quantity in Theorem 3.11, as direct numerical verification of that theorem, or as evidence that the complete-gradient oracle condition in Assumption 3.4 holds for the experimental classifier. This definition and its limitations are stated explicitly in Section 4.2 and Appendix D.4.3; the paired results appear in Table 8. The Adam/AdamW classification experiments remain complementary practical evaluations rather than tests of Theorem 3.11.

AuthorsOpenReview ↗

Response to Reviewer zJi6: Question 2

R2. Can the direct and trajectory-mediated effects be quantified separately?

Yes. The revised manuscript reports a matched counterfactual experiment that separates current-round finite-shot effects from differences inherited through earlier trajectories. The experiment uses a four-qubit, one-layer VQC, K=5K=5 heterogeneous clients, T=4\mathcal T=4 local proximal-SGD steps per round, mini-batches of size four, full participation, eight communication rounds, eight paired seeds, and shot budgets S∈{32,128,512}S\in\{32,128,512\}. Within each seed, every finite-shot run and its analytic-expectation reference share the initialization, preprocessed features, client partition, scheduled mini-batches, learning-rate sequence, and annealed proximal schedule.

Let Ψt(S)\Psi_t^{(S)} denote the complete one-round federated update map using SS shots per shifted-circuit evaluation, conditional on the fixed scheduled mini-batches, and let Ψt(∞)\Psi_t^{(\infty)} denote the corresponding analytic-expectation map. The analytic map removes finite-shot measurement sampling while preserving the same model, data, local-update rule, mini-batch schedule, and proximal dynamics; it is not an ideal global-gradient update. Adding and subtracting the analytic update applied to the finite-shot iterate gives the exact decomposition

θ(t+1,S)−θ(t+1,∞)=Ψt(S)(θ(t,S))−Ψt(∞)(θ(t,S))⏟δtdirect+Ψt(∞)(θ(t,S))−Ψt(∞)(θ(t,∞))⏟δthistory.\boldsymbol{\theta}^{(t+1,S)} - \boldsymbol{\theta}^{(t+1,\infty)} = \underbrace{\Psi_t^{(S)} \left(\boldsymbol{\theta}^{(t,S)}\right) - \Psi_t^{(\infty)} \left(\boldsymbol{\theta}^{(t,S)}\right)}_{\boldsymbol{\delta}_t^{\mathrm{direct}}} + \underbrace{\Psi_t^{(\infty)} \left(\boldsymbol{\theta}^{(t,S)}\right) - \Psi_t^{(\infty)} \left(\boldsymbol{\theta}^{(t,\infty)} \right)}_{\boldsymbol{\delta}_t^{\mathrm{history}}}.

The direct term measures the current-round difference between finite-shot and analytic updates from the same server parameter, including its propagation through the four local steps. The history term measures the response of the analytic update map to parameter differences inherited from previous rounds. We retain their interaction:

∥δttotal∥22=∥δtdirect∥22+∥δthistory∥22+2⟨δtdirect,δthistory⟩,\|\boldsymbol{\delta}_t^{\mathrm{total}}\|_2^2 = \|\boldsymbol{\delta}_t^{\mathrm{direct}}\|_2^2 + \|\boldsymbol{\delta}_t^{\mathrm{history}}\|_2^2 + 2\left\langle \boldsymbol{\delta}_t^{\mathrm{direct}}, \boldsymbol{\delta}_t^{\mathrm{history}} \right\rangle,

so no orthogonality or independence assumption is imposed.

The final-round means over the eight paired seeds are:

SSDirect2\mathrm{Direct}^2History2\mathrm{History}^2Cross\mathrm{Cross}Total2\mathrm{Total}^2
320.0001710.1285220.0017980.130491
1280.0000740.0864690.0022360.088778
5120.0000840.0783460.0004190.078850

These entries are components of squared parameter discrepancies, not gradient-estimator variances. Increasing SS from 3232 to 512512 reduces the mean final squared discrepancy by 39.639.6%, and Table 14 shows a reduction for every paired seed. The mean squared history component exceeds the direct component at the final round for all three shot budgets. Because the final-round learning rate is small, the current-round perturbation is comparatively small in these runs; we therefore do not interpret this ordering as a universal history-amplification factor. We also do not claim componentwise monotonicity: the mean direct component is slightly larger at S=512S=512 than at S=128S=128, and individual trajectories need not be monotone in SS.

Appendix D.5.2 additionally reports a 5050-round, single-seed shot-switch pilot. Following a 32→51232 \rightarrow 512 switch after round 25, the squared parameter gap relative to the constant-512512 reference decreases from 0.3424440.342444 at round 26 to 0.3079080.307908 at round 50 but remains nonzero while both runs use S=512S=512. The remaining gap is consistent with differences inherited from the first 25 rounds, but does not isolate that mechanism because differing post-switch measurement realizations may also contribute to the observed discrepancy. Parameter-space discrepancies also need not imply corresponding differences in objective value, stationarity, or predictive performance.

The complete protocol and results appear inSection 4.2, Appendices D.5.1--D.5.2, Figure 2, and Tables 13--15. This empirical counterfactual complements Lemma 3.9 but is not identified term-by-term with the stochastic-oracle and trajectory/proximal components in that analytical decomposition.

AuthorsOpenReview ↗

Response to Reviewer zJi6: Question 3

R3. How sensitive is the result to seed 6?

Seed 6 is influential, and the revised manuscript reports both the complete eight-seed comparison and the sensitivity result obtained after removing it. The paired percentage changes in the terminal squared-gradient diagnostic relative to Q-FedAvg are:

Seed01234567
Constant2.53-65.84-1.31-0.211.78-0.60-96.23-50.36
Annealed3.29-31.1061.390.44-7.10-3.36-94.95-46.21

Constant and annealed Q-FedProx reduce the diagnostic in 6/86/8 and 5/85/8 seeds, respectively. Across all eight seeds, their mean paired changes are −26.28-26.28% with a 9595% paired-bootstrap confidence interval of [−52.96-52.96%,−5.30-5.30%] and −14.70-14.70% with an interval of [−44.75-44.75%,14.2314.23%], respectively. Thus, the reported interval excludes zero only for the constant schedule.

Seed 6 contributes materially to the mean effect. Its terminal diagnostic values are approximately 117.98117.98 for Q-FedAvg, 4.454.45 for constant Q-FedProx, and 5.965.96 for annealed Q-FedProx, corresponding to reductions of 96.2396.23% and 94.9594.95%. Excluding seed 6 changes the mean paired reductions to 16.2916.29% and 3.243.24%, respectively. These seven-seed leave-one-out values are post hoc sensitivity diagnostics, not replacements for the primary eight-seed comparisons.

Removing seed 6 also changes the mean test-accuracy difference from −0.71-0.71 to +0.19+0.19 percentage points for constant Q-FedProx and from −0.85-0.85 to −0.05-0.05 percentage points for annealed Q-FedProx. This further confirms that seed 6 materially affects both the stationarity and predictive summaries. The stationarity effect is nevertheless not exclusive to seed 6: the constant schedule also reduces the diagnostic by 65.8465.84% for seed 1 and 50.3650.36% for seed 7.

A complementary robust summary gives median terminal diagnostic values of 1.1611.161 for Q-FedAvg, 1.1171.117 for constant Q-FedProx, and 1.2041.204 for annealed Q-FedProx. Their much smaller separation relative to the mean paired percentages confirms substantial seed-level variability. Moreover, a lower terminal diagnostic does not imply improved prediction: for seed 6, test accuracy decreases by 7.007.00 and 6.506.50 percentage points under the constant and annealed schedules, respectively.

We therefore make only the conclusion directly supported by these data: proximal regularization reduces the terminal finite-shot squared-gradient diagnostic for some trajectories in this controlled setting, including substantial reductions for seeds 1, 6, and 7 under the constant schedule. The results do not establish uniform superiority of either proximal schedule or improved classification performance. The aggregate results, all seed-level values, predictive comparisons, and sensitivity analysis appear in Appendices D.4.3--D.4.4 and Tables 8--12.

Thank you for your careful response.

Currently, I don't have any further question for you now. I believe the assessment with "Weak Acceptance" can reasonably reflect my opinion on this paper. So I will keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.