acceptodds
Under review as a conference paper at ICLR 2027

Local Riemannian Stochastic Gradient Descent: Learning and Uncertainty Quantification

Abstract

We study statistical inference with infrequent communication for the full-participation Riemannian Federated Averaging Gradient Streams (RFedAGS) recursion with growing communication intervals, where client drift and geometric error must be controlled at the scale of all local observations. A moment-sensitive synchronized-state bound makes these errors negligible and yields a central limit theorem (CLT) for a computable fixed-anchor average in a regular local basin. The covariance separates the objective Hessian, client noise and communication schedule. With a fixed burn-in, admissible slowly varying intervals preserve first-order precision while communication frequency vanishes. Fresh pooled scores and exact or sample-based Hessians consistently estimate this covariance, giving feasible Wald regions. Gaussian and sphere experiments at matched total oracle budgets show that a prespecified half-tail attains near-nominal coverage with substantially lower communication cost than a fixed-interval baseline, at the cost of a larger uncertainty scale. A four-site principal-subspace application illustrates the construction.

Primary area:
optimization
Keywords:
Central Limit Theorem, Federated learning, Riemannian SGD, Uncertainty quantification

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_6egQOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies statistical inference for RFedAGS with growing communication intervals. It derives a refined mean-square bound, establishes a CLT for a computable fixed-anchor average, and constructs consistent covariance estimates and Wald regions. Experiments quantify the trade-off between communication cost, coverage, and precision.

Critical strengths
  1. The refined state-error bound retains within-round variance reduction and permits substantially growing communication intervals. Controlling geometric and client-drift errors at the total-observation scale is a meaningful technical contribution.
  2. The paper provides feasible inference and counts its additional oracle, communication, and storage costs. The experiments report both calibrated cases and severe undercoverage, making the practical limitations clear.
Critical weaknesses
  1. The theory assumes that synchronized points, all local iterates, and reported averages remain in a sufficiently small neighborhood after a fixed burn-in. It does not establish this containment for the general unmodified recursion, limiting the applicability of the inference guarantees.
  2. The main communication comparison changes the interval schedule, step parameters, and discarded fraction together. Although tail ablations are provided, the advantage over a comparably tuned fixed-interval baseline remains unclear, especially at matched uncertainty size.
Questions for authors
  1. For the curved examples, can the required neighborhood containment be established for the unmodified recursion, rather than assumed?
  2. How does the growing-interval method compare with fixed intervals when both use the same tail fraction and are tuned to achieve comparable coverage and uncertainty size?
3 comments

Thank you for your time and valuable comments.

W1. The curved inference result requires the trajectory to remain in one fixed contraction neighborhood. The explicit radius and step restrictions are given below. This supplies a common domain for inverse retractions, transports and local expansions. The sphere and rotation data have bounded supports; their measured trajectories provide finite-run diagnostics, while the manifold inference theorem assumes the stated all-future containment. We extend the Gaussian quadratic result to finite-moment noise: for the common-Hessian quadratic model Gk(x,ξ)=H(x−ak)+ξG_k(x,\xi)=H(x-a_k)+\xi, H≻0H\succ0, with fixed dimension and client count. The client noises have fixed centered laws, are independent across calls and clients, and have finite (2+δ)(2+\delta) moments for some 0<δ≤20 < \delta\le2; let S=∑kpk2Cov⁡(ξk)≻0S=\sum_kp_k^2\operatorname{Cov}(\xi^k)\succ0. For the original unstopped recursion from any deterministic start, take Em=max⁡1,⌊cEmβ⌋E_m=\max\\{1,\lfloor c_Em^\beta\rfloor\\} and γm=cγ(m+m0)−α\gamma_m=c_\gamma(m+m_0)^{-\alpha} with 0≤β<10\le\beta < 1 and (1+β)/2<α<1(1+\beta)/2 < \alpha < 1. Then E∣xm−x⋆∣2=O(ηm)\mathbb E\\|x_m-x_\star\\|^2=O(\eta_m). Averaging rounds after bT=max⁡b0,⌊ρT⌋b_T=\max\\{b_0,\lfloor\rho T\rfloor\\}, for fixed b0b_0 and 0≤ρ<10\le\rho < 1, gives

\nu_\rho=\frac{1-\rho^{1-\beta}}{(1-\rho)^2(1-\beta^2)}.$$ Independent fresh pooled scores consistently estimate $S$; exact Hessians and this retained-round variance factor therefore give asymptotically calibrated Wald regions. The proof uses the exact linear round transition, finite-moment Lyapunov control and the fresh-score law of large numbers. It covers ideal $t_5$ and symmetric Pareto shape-2.8 noise, for example with $\delta=0.5$. The submitted Pareto simulation uses a clamped sampler and therefore has all moments; the finite-moment extension above applies to the ideal, unclamped shape-2.8 law as well. The curved result retains its explicit local stability hypothesis. The finite-moment result is Proposition K.1; the Pareto sampler is described in Section 4.2. **W2.** We compare fixed $E_m=8$ with growing $E_m=\max\\{1,\lfloor\sqrt m\rfloor\\}$ in the Gaussian quadratic model, holding $\alpha=0.8$, the gain and the discarded fraction equal. The following independent paired experiment uses 1000 replications per configuration, with per-client budgets counting both optimization and fresh covariance-score observations. Both the following table and the subsequent fraction-0.1 comparisons use these new streams. At discarded fraction 0.5 and gain 0.8: | Per-client budget | Fixed coverage | Growing coverage | Fixed size | Growing size | Fixed MB | Growing MB | |---|---:|---:|---:|---:|---:|---:| | 16384 | 0.977 | 0.969 | 0.004474 | 0.003880 | 3.084 | 1.350 | | 65536 | 0.965 | 0.957 | 0.002238 | 0.001931 | 12.336 | 3.407 | With these identical tuning choices, the growing schedule gives smaller regions and lower communication at both budgets. The submitted fixed baseline instead uses fraction 0.1 and exponent 0.65, which explains the different size comparison there. At discarded fraction 0.1 and gain 1.6, fixed/growing sizes are 0.003415/0.003343 and 0.001707/0.001655, with coverage 0.959/0.923 and 0.944/0.933. With fraction 0.1 and gain 0.8, coverage is 0.948/0.723 and 0.939/0.907. For the growing gain-0.8 schedule at budget 16384, this rerun gives 0.723/0.969 at fractions 0.1/0.5; the submitted batch gives 0.710/0.962. The communication advantage depends on the schedules and tuning, as the validation-selected comparison below shows. These comparisons appear in Section 4.2, Table 2 and Appendix B.5.

Q1. For the sphere and rotation recursions at their experimental initializations, all-future containment in the required contraction basin remains an assumption. We make the radius restriction explicit below and use that same neighborhood for contraction and inference.

A conservative sufficient radius can be obtained from the constants already appearing in the proof. Let

\qquad b_{m+1}=\gamma_m r_{m+1},$$ and let $C_1 > 0$ be a uniform constant in the existing bound $$\mathbb E\\!\left[\\|b_{m+1}\\|\mid\mathcal F_m\right] \le C_1\gamma_m\bigl(\gamma_m+\\|\Delta_m\\|^2\bigr).$$ Fix this constant on a reference neighborhood where the geometric and oracle bounds hold. If $r_{\mathrm{geom}}$ is an admissible chart radius, a sufficient requirement on the neighborhood used in (G3) is $$\sup_{x\in\mathcal U} \\|R_{x^\star}^{-1}(x)\\| \le r \le \min\left\\{ r_{\mathrm{geom}},\frac{\mu}{8C_1} \right\\}.$$ The same fixed reference neighborhood also supplies constants $C_2,C_{\mathrm{noise}}$ such that, writing $\mathbb E_m=\mathbb E[\\,\cdot\mid\mathcal F_m]$ and $\kappa=\min(\delta,1)$, $$\mathbb E_m\\|b_{m+1}\\|^2\le C_2\\{\gamma_m^2\\|\Delta_m\\|^4+\gamma_m^{2+\kappa}\\},\qquad \mathbb E_m\\|\epsilon_{m+1}\\|^2\le C_{\mathrm{noise}}/E_m,\quad \mathbb E_m\epsilon_{m+1}=0.$$ With $M=\\|H_\star\\|_{\mathrm{op}}$, impose after the fixed burn-in $$\gamma_m\le\min\left\\{1,M^{-1},\frac{\mu}{2(M^2+3C_2r^2)}\right\\}.$$ For $D_m=(I-\gamma_mH_\star)\Delta_m$, the state–remainder cross term is bounded by $$2|\mathbb E_m\langle D_m,b_{m+1}\rangle| \le(\mu/4+2C_1r)\gamma_m\\|\Delta_m\\|^2+(4C_1^2/\mu)\gamma_m^3.$$ The noise–remainder cross term is at most $\gamma_m^2\mathbb E_m\\|\epsilon_{m+1}\\|^2/2+2\mathbb E_m\\|b_{m+1}\\|^2$. Together with the radius and step restrictions, this gives $$\mathbb E_m\\|\Delta_{m+1}\\|^2 \le(1-\mu\gamma_m)\\|\Delta_m\\|^2 +\tfrac32C_{\mathrm{noise}}\gamma_m\eta_m +(3C_2+4C_1^2/\mu)\gamma_m^{2+\kappa}.$$ Thus the original state rate follows with explicit constants. G3 requires containment under the original oracle law on this same predetermined neighborhood; the decreasing step schedule supplies the eventual step restriction. The experimental configurations are: - In the main sphere experiment, the tangent Gaussian observations are **truncated to radius $0.90$** before being mapped to the sphere. The initialization is $\operatorname{Exp}_{x^\star}(0.42,0.18)$, at distance approximately $0.457$ from the target. - In the rotation experiment, the observation rotation vectors have norm at most $0.5$. The initialization is $(0.3,0.2,-0.2)$, with norm approximately $0.412$. - The Euclidean experiments use unbounded Gaussian and Student noise and a clamped Pareto sampler. The bounded supports in the curved experiments allow local moment and smoothness conditions to be checked directly. In particular, preservation of a convex ball by an individual exponential SGD step does not automatically imply preservation under the RFedAGS server update, which aggregates transported gradient streams. The reported rotation diagnostic—all realized local rotation angles below $0.414$ in the 600 finite runs—is evidence about those simulated trajectories. The normalized conservative certificate obtained by taking $C_1\ge1$ has radius at most $1/8$ in these intrinsic-mean models, since $H_\star\preceq I$. For every deterministic burn-in, their sampling laws admit positive-probability finite sequences that remain outside this particular basin. On the sphere, observations can stay near the initial point. On rotations, $v_*=0.13(1,1,-1)/\sqrt3$ lies inside every client support because $\max_k\\|v_*-\mu_k\\|^2\le 0.13^2+0.2^2+0.052/\sqrt3 < 0.3^2$; the geodesic from the actual initialization to this point stays at target distance at least 0.13. Repeated observations at a common point $y$ give the exact gradient-stream update $x_{m+1}=\mathrm{Exp}_{x_m}[-(1-η_m)^{E_m}\mathrm{Log}_{x_m}v]$ The local iterates, synchronized points and fixed-anchor averages remain on the same segment; continuity extends each finite construction to a positive-probability event. Consequently this normalized small-basin certificate leaves the curved experiments outside its verified applicability. A sharper constant or a larger-basin/random-entry argument is a separate analytical route. The finite-trajectory diagnostics describe the observed stability, and the obstruction concerns this sufficient certificate alone. The radius and step restrictions appear in Assumption C.1(G3) and Lemma G.1; Appendix G gives the scope of the normalized certificate.

Q2. We compare fixed and growing intervals with a common tail fraction ρ=0.5\rho=0.5 and independent validation-loss tuning. Both families receive equal search budgets; tuning uses observed validation loss. The selected Gaussian fixed/growing settings are (E,α,cγ)=(16,0.65,0.8)(E,\alpha,c_\gamma)=(16,0.65,0.8) and (Em,α,cγ)=(max⁡1,⌊.5m⌋,0.9,1.6)(E_m,\alpha,c_\gamma)=(\max\\{1,\lfloor.5\sqrt m\rfloor\\},0.9,1.6). The sphere settings are (8,0.8,0.8)(8,0.8,0.8) and (max⁡1,⌊.5m⌋,0.8,0.8)(\max\\{1,\lfloor.5\sqrt m\rfloor\\},0.8,0.8). A separate independent pilot selects budgets to match uncertainty size. Equal-budget and matched-size comparisons use separate 500-replication confirmation batches. Size is tr⁡(Σ^T/tT)1/2\\{\operatorname{tr}(\widehat\Sigma_T/t_T)\\}^{1/2}; communication includes the dense optimization and inference messages.

Model / comparisonBudget, fixed/growingFixed coverageGrowing coverageFixed sizeGrowing sizeFixed MBGrowing MB
Gaussian, equal budget16384/163840.9500.9700.00441130.00390211.58722.1440
Gaussian, matched size16384/122880.9440.9420.00441870.00452141.58721.7696
Sphere, equal budget16384/163840.9820.9380.0009904110.0008646110.8634880.600320
Sphere, matched size16384/122880.9680.8900.0009903580.0009979530.8634880.495488

The selected Gaussian schedule Em=max⁡1,⌊.5m⌋E_m=\max\\{1,\lfloor.5\sqrt m\rfloor\\} communicates more often than fixed E=16E=16 at these budgets. At matched size it has similar coverage and more communication. On the sphere, the growing schedule reduces communication by 42.6% while coverage falls from 0.968 to 0.890. On the sphere, validation-loss tuning matches size but gives lower coverage for the growing schedule. These results quantify the calibration–communication tradeoff after equal-tail, data-based tuning.

A separate model-specific calibration exploration selects settings using the known simulation target, followed by independent confirmation. At matched sphere uncertainty size, growing coverage is 0.952 versus 0.982 for fixed intervals, with 63.9% lower dense communication and actual per-client sample counts 12269 versus 16380. Its MSE is 8.9319×10−78.9319\times10^{-7} versus 7.4036×10−77.4036\times10^{-7}. This is evidence about attainable configurations in that simulation model; the validation-loss comparison above is the data-based tuning procedure. The validation-selected results are in Table 5 and Appendix B.5.

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

Reviewer_oqy4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies statistical inference for Riemannian Local SGD, providing theoretical guarantees and a practical uncertainty quantification procedure.

I will also consider the comments of other reviewers before making my final assessment.


After rebuttal.

My main concerns are about the assumptions and the novelty of the work compared with Li et al. (2022). Although I am not an expert in statistical inference, I find the authors' responses reasonable.

Overall, I view this work as an incremental yet solid contribution building on local Riemannian SGD (Huang et al., 2026) and statistical inference (Li et al., 2022). Therefore, I have decided to give a borderline rating, leaning toward acceptance.

Critical strengths

The paper investigates statistical inference for Riemannian Local SGD, extending existing results to manifold-valued optimization.

Critical weaknesses
  1. Presentation and clarity. The paper needs further polishing, particularly the Introduction.

    • Motivation and background: Since the paper focuses on statistical inference, the Introduction should better explain its importance and motivation in distributed machine learning. The first three paragraphs should present the development of this research area with a clearer logical flow, rather than primarily listing related work.
    • Contributions: The fourth and fifth paragraphs should explain the main contributions more intuitively, preferably through explicit comparisons with existing work. As a reader without direct expertise in this area, I found it difficult to identify the key contributions from the technical descriptions (the mathematical expressions) alone.
    • Formatting: Figure 1 is placed too far from its corresponding discussion.
  2. Strong assumptions. Some assumptions appear restrictive:

    • Assumption C.1 (G3): All within-round local iterates and the reported geometric averages are required to remain in U\mathcal U. Can this condition be relaxed or established under more practical assumptions?

    • Assumption C.3 (21): The stochastic gradient is assumed to be uniformly bounded E[∥∣Gk(x,ξ)∥∣2+δ∣G]≤CG\mathbb{E}[\|| G_k (x,\xi)\||^{2+\delta}\mid \mathcal{G}] \leq C_G. Could this be replaced by a weaker condition on the centered gradient noise, such as E[∥∣Gk(x,ξ)−∇fk(x)∥∣2+δ∣G]≤CG\mathbb E[\||G_k(x,\xi)-\nabla f_k(x)\||^{2+\delta}\mid\mathcal G]\le C_G?

    • Some key assumptions should also be summarized in the main body.

  3. Novelty. The work appears to extend the statistical inference framework for Euclidean Local SGD (Li et al., 2022) to Riemannian Local SGD (Huang et al., 2026). The authors should more explicitly highlight the technical differences and novel contributions relative to these works.

Questions for authors

None beyond the issues raised above.

8 comments

Thank you for your time and valuable comments.

W1. We replace the opening with: “Distributed principal component analysis needs a sampling law and feasible uncertainty scale, not only optimization error.” Less frequent communication changes both the stochastic averaging scale and the geometric remainder terms. Relative to Li et al. (2022), the sharper state bound permits power schedules excluded by their displayed schedule condition, while controlling transport and retraction errors. Relative to Huang et al. (2026), the target is a sampling law and feasible uncertainty quantification for the full-participation RFedAGS recursion.

We move the transport illustration beside its first discussion. It compares client-to-server-to-target transport with direct client-to-target transport: both preserve lengths, but their compositions can differ on a curved manifold. Controlling this difference connects the geometric analysis to the covariance estimate and uncertainty regions. The opening and literature comparison are in the Introduction; Figure 1 is beside the transport discussion on page 2.

W2. Assumption (G3) specifies the local stability regime in which we study statistical fluctuations around a nondegenerate local minimizer x⋆x^\star. The neighborhood U\mathcal U provides a common domain for the inverse retractions, vector transports, and local expansions used by the inference procedure. This requirement concerns the region visited by the algorithm.

Similar localization assumptions are used in Riemannian stochastic optimization. Bonnabel (2013, Theorems 1–2) assumes that the iterates remain in a compact set. Tripuraneni, Flammarion, Bach, and Jordan (2018, Assumption 2) study averaged Riemannian SGD in a compact, connected, totally retractive region. Huang et al. (2026, Assumption 3.2) require both the outer and inner RFedAGS iterates to remain in a common totally retractive region contained in a compact connected set. Thus, localization of the within-round iterates has a direct precedent in the optimization framework underlying our paper. Such localization is also compatible with noncompact manifolds. For example, Zhang and Sra (2016) analyze projected stochastic methods on bounded geodesically convex regions of Hadamard manifolds. The relevant distinction is between the global geometry of the parameter space and the local region in which a stable optimization trajectory is analyzed.

Under our existing assumptions, the proposed condition is equivalent to the moment requirement in (21), up to its numerical constant.

Let p=2+δp=2+\delta and write

=G_k(x,\xi)-\nabla f_k(x).$$ The geometric and objective-regularity assumptions specify a compact neighborhood $\mathcal U$ on which every client gradient is continuous. Since the number of clients is fixed, $$B := \max_{1\le k\le K} \sup_{x\in\mathcal U} \\|\nabla f_k(x)\\| < \infty.$$ Here all norms are the Riemannian norms in the corresponding tangent spaces. For any $x\in\mathcal U$, the inequality $$\\|u+v\\|^p \le 2^{p-1}\bigl(\\|u\\|^p+\\|v\\|^p\bigr)$$ gives $$\mathbb E\\!\left[ \\|G_k(x,\xi)\\|^p\mid\mathcal G \right] \le 2^{p-1} \left\\{ \mathbb E\\!\left[ \\|\varepsilon_k(x,\xi)\\|^p\mid\mathcal G \right] +B^p \right\\},$$ almost surely. Consequently, the reviewer’s proposed uniform bound $C_\varepsilon$ implies (21) with $$C_G=2^{p-1}(C_\varepsilon+B^p).$$ Conversely, $$\mathbb E\\!\left[ \\|\varepsilon_k(x,\xi)\\|^p\mid\mathcal G \right] \le 2^{p-1}(C_G+B^p).$$ Thus, the two uniform conditional moment requirements describe the same class of admissible oracles under C.1–C.2. The argument requires bounded mean gradients on the localized parameter region and permits unbounded observations and stochastic gradients. In particular, the individual client gradients at $x^\star$ need not vanish; the argument allows the client heterogeneity considered in our paper. The assumptions cover geometry, path containment, local identification, oracle moments, covariance regularity and schedules. Section 3 summarizes these assumptions, and Appendix C.1 gives the equivalent local moment formulations. In the curved models, the inference guarantee is conditional on all-future containment in the same fixed contraction basin; bounded observations and finite-run trajectory diagnostics alone establish neither this condition nor its probability-one form. We extend the Gaussian quadratic result to finite-moment noise: for the common-Hessian quadratic model $G_k(x,\xi)=H(x-a_k)+\xi$, $H\succ0$, with fixed dimension and client count. The client noises have fixed centered laws, are independent across calls and clients, and have finite $(2+\delta)$ moments for some $0 < \delta\le2$; let $S=\sum_kp_k^2\operatorname{Cov}(\xi^k)\succ0$. For the original unstopped recursion from any deterministic start, take $E_m=\max\\{1,\lfloor c_Em^\beta\rfloor\\}$ and $\gamma_m=c_\gamma(m+m_0)^{-\alpha}$ with $0\le\beta < 1$ and $(1+\beta)/2 < \alpha < 1$. Then $\mathbb E\\|x_m-x_\star\\|^2=O(\eta_m)$. Averaging rounds after $b_T=\max\\{b_0,\lfloor\rho T\rfloor\\}$, for fixed $b_0$ and $0\le\rho < 1$, gives $$\sqrt{t_T}(\widehat x_T-x_\star)\Rightarrow N(0,\nu_\rho H^{-1}SH^{-1}),\qquad \nu_\rho=\frac{1-\rho^{1-\beta}}{(1-\rho)^2(1-\beta^2)}.$$ Independent fresh pooled scores consistently estimate $S$; exact Hessians and this retained-round variance factor therefore give asymptotically calibrated Wald regions. The proof uses the exact linear round transition, finite-moment Lyapunov control and the fresh-score law of large numbers. It covers ideal $t_5$ and symmetric Pareto shape-2.8 noise, for example with $\delta=0.5$. The submitted Pareto simulation uses a clamped sampler and therefore has all moments; the finite-moment extension above applies to the ideal, unclamped shape-2.8 law as well. The curved result retains its explicit local stability hypothesis. The finite-moment result is Proposition K.1; the Pareto sampler is described in Section 4.2.

W3. We compare our work with the proposed literature. Li et al. (2022) establish statistical inference for Euclidean Local SGD. Huang et al. (2026) introduce RFedAGS and analyze its optimization behavior, including heterogeneous clients and partial participation. Our paper studies statistical inference for the full-participation RFedAGS recursion when communication intervals grow.

The main additional difficulty is that gradients evaluated at different client iterates belong to different tangent spaces. The algorithm transports these gradients to the server, and the asymptotic analysis transports them further to the target. The resulting transport and retraction errors interact with client drift and increasing local work. Establishing a sampling limit requires controlling their accumulated errors at the scale tT\sqrt{t_T}, where tTt_T counts the local oracle observations.

Our main technical contributions address this difficulty in three steps:

  1. Control of geometric and local-update errors. The round-wise representation follows the actual client-to-server-to-target transport path, without assuming that transport is transitive. The moment-dependent state bound
=O\\!\left( \eta_m+\gamma_m^{\\,1+\min(\delta,1)} \right)$$ preserves the variance reduction from within-round averaging and controls the boundary terms needed for inference with growing communication intervals. 2. **A computable manifold-valued estimator.** The fixed-anchor average connects the theoretical average in the unknown target’s tangent space to an implementable estimator. Its averaging step uses $O(d)$ memory and does not require the additional squared-retraction-distance convexity condition used for the recursive alternative. 3. **Feasible uncertainty quantification in the estimated tangent space.** We establish consistency of the transported score and Hessian estimators in the tangent space of the reported estimate, yielding feasible Wald regions. This requires accounting for the changing evaluation points, transport, and the estimated reporting point. Li et al. (2022, §4.1) pool client scores before forming outer products in Euclidean Local SGD. Our construction transports these messages to a common tangent space and controls the resulting geometric errors in the covariance consistency analysis. This comparison appears in the Introduction on page 2. **References** - Bonnabel, S. (2013). [Stochastic Gradient Descent on Riemannian Manifolds](https://arxiv.org/abs/1111.5280). *IEEE Transactions on Automatic Control*, **58**(9), 2217–2229. - Huang, Z., Huang, W., Jawanpuria, P., and Mishra, B. (2026). [Riemannian Federated Learning via Averaging Gradient Streams](https://proceedings.iclr.cc/paper_files/paper/2026/file/2295135a965127dd488e28aca753457b-Paper-Conference.pdf). *ICLR*. - Li, X., Liang, J., Chang, X., and Zhang, Z. (2022). [Statistical Estimation and Online Inference via Local SGD](https://proceedings.mlr.press/v178/li22b.html). *COLT*, PMLR **178**, 1613–1661. - Tripuraneni, N., Flammarion, N., Bach, F., and Jordan, M. I. (2018). [Averaging Stochastic Gradient Descent on Riemannian Manifolds](https://proceedings.mlr.press/v75/tripuraneni18a.html). *COLT*, PMLR **75**, 650–687. - Zhang, H., and Sra, S. (2016). [First-order Methods for Geodesically Convex Optimization](https://proceedings.mlr.press/v49/zhang16b.html). *COLT*, PMLR **49**, 1617–1638. Thank you again for your valuable feedback. We hope our responses address your concerns and welcome any further questions or discussion.
Reviewer_UmQiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies inference for the full-participation RFedAGS recursion (Riemannian Federated SGD, Huang et al., 2026). In each round, KK heterogeneous clients run EmE_m local Riemannian updates with a step size ηm\eta_m, starting from the synchronised point, and transport their gradients to the server's tangent space. The sever then takes a step of size γm=Emηm\gamma_m=E_m\eta_m.

The authors establish a mean-square bound for the synchronised iterate, E∥Δm∥2=O(ηm+γm1+min⁡(δ,1))\mathbb E\|\Delta_m\|^2 = O(\eta_m + \gamma_m^{1 + \min(\delta,1)}) under a (2+δ)(2+\delta) moment (Theorem 3.1). This allows taking γm≍m−α,Em≍mβ\gamma_m \asymp m^{-\alpha}, E_m \asymp m^\beta.

They also prove a central limit theorem at the optimization clock tT=∑m≤TEmt_T=\sum_{m\le T}E_m, with an asymptotic covariance νH⋆−1SH⋆−1\nu H_\star^{-1}SH_\star^{-1}, with pooled client noise S=∑kpk2SkS=\sum_kp_k^2S_k and an explicit factor ν=lim⁡(tT/T2)∑mEm−1\nu = \lim(t_T/T^2) \sum_m E_m^{-1} (Theorem 3.2).

This paper also shows that it is possible to consistently estimate the covariance using fresh pooled scores and transported Hessian messages, thus yielding Wald regions (Theorem 3.3, Corollary 3.4).

Gaussian, sphere and rotation simulations compare error, coverage, uncertainty size, and communication at matched optimization-plus-inference budgets. A four-site Grassmann principal-subspace example illustrates inference conditional on the UCI Heart Disease data.

Critical strengths

This paper presents a refined analysis of the interaction between local work and geometry.

When using the bound E∥Δm∥2=O(γm)E \|\Delta_m\|^2 = O(\gamma_m) from Li et al. (2022), the endpoint condition forces α+β<1\alpha + \beta < 1. Since α>(1+β)/2\alpha > (1+\beta)/2, this forces β<1/3\beta < 1/3.

Theorem 3.1 manages to get O(ηm+γm1+κ)O(\eta_m + \gamma_m^{1+\kappa}), removing the above restriction, and allowing (under a third moment) 0≤β<10\le\beta<1 and (1+β)/2<α<1(1+\beta)/2<\alpha<1 for Em≍mβE_m\asymp m^\beta and γm≍m−α\gamma_m\asymp m^{-\alpha}. This is a notable technical improvement.

For instance, the experiments use α=0.8,β=1/2>1/3\alpha = 0.8, \beta = 1/2 > 1/3, which lies outside of the region given by Li et al. (2022).

Critical weaknesses

The theoretical guarantees remain conditional on local stability.

Assumption C.1(G3) requires synchronized iterates, local iterates, and reported averages to remain in some neighborhood U\mathcal U after a fixed finite burn-in. Moreover, Lemma G.1 needs that neighborhood small enough to absorb nonlinear terms. This assumption is a property of the algorithm, and not of the problem. With unbounded noise (such as Gaussian, Student and Pareto used in Section 4.2), each step can exit U\mathcal U with a positive probability.

Proposition K.1 shows that for the quadratic Gaussian model, Assumption C.1 (G3) is not required ; for the other settings however, the theoretical results are established for a localised recursion.

Questions for authors
  1. Clarification of the localisation assumption. Is it possible to make the required basin size in Lemma G.1 explicit in Assumption C.1, and can G3 be verified for the actual sphere or rotation recursion and initialization used in the experiments?

  2. Comparison with Li et al. (2022). Could you state explicitly which schedules are acceptable for your result, but not for their Assumption 3.4(iv) (in the Euclidean special case)? Can the random-scaling approach of Li et al. (2022) be transported to the manifold setting, thus avoiding d2d^2 messages?

  3. Choice of tail fraction. Is there a data-driven rule for ρ\rho that avoids the severe undercoverage reported in Section 4.2 without pilot tuning?

6 comments

Thank you for your time and valuable comments.

W1. Assumption (G3) specifies the local stability regime in which we study statistical fluctuations around a nondegenerate local minimizer x⋆x^\star. The neighborhood U\mathcal U provides a common domain for the inverse retractions, vector transports, and local expansions used by the inference procedure. This requirement concerns the region visited by the algorithm.

Similar localization assumptions are used in Riemannian stochastic optimization. Bonnabel (2013, Theorems 1–2) assumes that the iterates remain in a compact set. Tripuraneni, Flammarion, Bach, and Jordan (2018, Assumption 2) study averaged Riemannian SGD in a compact, connected, totally retractive region. Huang et al. (2026, Assumption 3.2) require both the outer and inner RFedAGS iterates to remain in a common totally retractive region contained in a compact connected set. Thus, localization of the within-round iterates has a direct precedent in the optimization framework underlying our paper. Such localization is also compatible with noncompact manifolds. For example, Zhang and Sra (2016) analyze projected stochastic methods on bounded geodesically convex regions of Hadamard manifolds. The relevant distinction is between the global geometry of the parameter space and the local region in which a stable optimization trajectory is analyzed.

We extend the Gaussian quadratic result to finite-moment noise: for the common-Hessian quadratic model Gk(x,ξ)=H(x−ak)+ξG_k(x,\xi)=H(x-a_k)+\xi, H≻0H\succ0, with fixed dimension and client count. The client noises have fixed centered laws, are independent across calls and clients, and have finite (2+δ)(2+\delta) moments for some 0<δ≤20 < \delta\le2; let S=∑kpk2Cov⁡(ξk)≻0S=\sum_kp_k^2\operatorname{Cov}(\xi^k)\succ0. For the original unstopped recursion from any deterministic start, take Em=max⁡1,⌊cEmβ⌋E_m=\max\\{1,\lfloor c_Em^\beta\rfloor\\} and γm=cγ(m+m0)−α\gamma_m=c_\gamma(m+m_0)^{-\alpha} with 0≤β<10\le\beta < 1 and (1+β)/2<α<1(1+\beta)/2 < \alpha < 1. Then E∣xm−x⋆∣2=O(ηm)\mathbb E\\|x_m-x_\star\\|^2=O(\eta_m). Averaging rounds after bT=max⁡b0,⌊ρT⌋b_T=\max\\{b_0,\lfloor\rho T\rfloor\\}, for fixed b0b_0 and 0≤ρ<10\le\rho < 1, gives

\nu_\rho=\frac{1-\rho^{1-\beta}}{(1-\rho)^2(1-\beta^2)}.$$ Independent fresh pooled scores consistently estimate $S$; exact Hessians and this retained-round variance factor therefore give asymptotically calibrated Wald regions. The proof uses the exact linear round transition, finite-moment Lyapunov control and the fresh-score law of large numbers. It covers ideal $t_5$ and symmetric Pareto shape-2.8 noise, for example with $\delta=0.5$. The submitted Pareto simulation uses a clamped sampler and therefore has all moments; the finite-moment extension above applies to the ideal, unclamped shape-2.8 law as well. The curved result retains its explicit local stability hypothesis. The finite-moment result is Proposition K.1; the Pareto sampler is described in Section 4.2. Continued in the next part

Q1. For the sphere and rotation recursions at their experimental initializations, all-future containment in the required contraction basin remains an assumption. We make the radius restriction explicit below and use that same neighborhood for contraction and inference.

A conservative sufficient radius can be obtained from the constants already appearing in the proof. Let

\qquad b_{m+1}=\gamma_m r_{m+1},$$ and let $C_1 > 0$ be a uniform constant in the existing bound $$\mathbb E\\!\left[\\|b_{m+1}\\|\mid\mathcal F_m\right] \le C_1\gamma_m\bigl(\gamma_m+\\|\Delta_m\\|^2\bigr).$$ Fix this constant on a reference neighborhood where the geometric and oracle bounds hold. If $r_{\mathrm{geom}}$ is an admissible chart radius, a sufficient requirement on the neighborhood used in (G3) is $$\sup_{x\in\mathcal U} \\|R_{x^\star}^{-1}(x)\\| \le r \le \min\left\\{ r_{\mathrm{geom}},\frac{\mu}{8C_1} \right\\}.$$ The same fixed reference neighborhood also supplies constants $C_2,C_{\mathrm{noise}}$ such that, writing $\mathbb E_m=\mathbb E[\\,\cdot\mid\mathcal F_m]$ and $\kappa=\min(\delta,1)$, $$\mathbb E_m\\|b_{m+1}\\|^2\le C_2\\{\gamma_m^2\\|\Delta_m\\|^4+\gamma_m^{2+\kappa}\\},\qquad \mathbb E_m\\|\epsilon_{m+1}\\|^2\le C_{\mathrm{noise}}/E_m,\quad \mathbb E_m\epsilon_{m+1}=0.$$ With $M=\\|H_\star\\|_{\mathrm{op}}$, impose after the fixed burn-in $$\gamma_m\le\min\left\\{1,M^{-1},\frac{\mu}{2(M^2+3C_2r^2)}\right\\}.$$ For $D_m=(I-\gamma_mH_\star)\Delta_m$, the state–remainder cross term is bounded by $$2|\mathbb E_m\langle D_m,b_{m+1}\rangle| \le(\mu/4+2C_1r)\gamma_m\\|\Delta_m\\|^2+(4C_1^2/\mu)\gamma_m^3.$$ The noise–remainder cross term is at most $\gamma_m^2\mathbb E_m\\|\epsilon_{m+1}\\|^2/2+2\mathbb E_m\\|b_{m+1}\\|^2$. Together with the radius and step restrictions, this gives $$\mathbb E_m\\|\Delta_{m+1}\\|^2 \le(1-\mu\gamma_m)\\|\Delta_m\\|^2 +\tfrac32C_{\mathrm{noise}}\gamma_m\eta_m +(3C_2+4C_1^2/\mu)\gamma_m^{2+\kappa}.$$ Thus the original state rate follows with explicit constants. G3 requires containment under the original oracle law on this same predetermined neighborhood; the decreasing step schedule supplies the eventual step restriction. The experimental configurations are: - In the main sphere experiment, the tangent Gaussian observations are **truncated to radius $0.90$** before being mapped to the sphere. The initialization is $\operatorname{Exp}_{x^\star}(0.42,0.18)$, at distance approximately $0.457$ from the target. - In the rotation experiment, the observation rotation vectors have norm at most $0.5$. The initialization is $(0.3,0.2,-0.2)$, with norm approximately $0.412$. - The Euclidean experiments use unbounded Gaussian and Student noise and a clamped Pareto sampler. The bounded supports in the curved experiments allow local moment and smoothness conditions to be checked directly. In particular, preservation of a convex ball by an individual exponential SGD step does not automatically imply preservation under the RFedAGS server update, which aggregates transported gradient streams. The reported rotation diagnostic—all realized local rotation angles below $0.414$ in the 600 finite runs—is evidence about those simulated trajectories. The normalized conservative certificate obtained by taking $C_1\ge1$ has radius at most $1/8$ in these intrinsic-mean models, since $H_\star\preceq I$. For every deterministic burn-in, their sampling laws admit positive-probability finite sequences that remain outside this particular basin. On the sphere, observations can stay near the initial point. On rotations, $v_*=0.13(1,1,-1)/\sqrt3$ lies inside every client support because $\max_k\\|v_*-\mu_k\\|^2\le 0.13^2+0.2^2+0.052/\sqrt3 < 0.3^2$; the geodesic from the actual initialization to this point stays at target distance at least 0.13. Repeated observations at a common point $y$ give the exact gradient-stream update $$x_{m+1}=\operatorname{Exp}_{x_m}\\{[1-(1-\eta_m)^{E_m}]\operatorname{Log}_{x_m}y\\}.$$ The local iterates, synchronized points and fixed-anchor averages remain on the same segment; continuity extends each finite construction to a positive-probability event. Consequently this normalized small-basin certificate leaves the curved experiments outside its verified applicability. A sharper constant or a larger-basin/random-entry argument is a separate analytical route. The finite-trajectory diagnostics describe the observed stability, and the obstruction concerns this sufficient certificate alone. The radius and step restrictions appear in Assumption C.1(G3) and Lemma G.1; Appendix G gives the scope of the normalized certificate. Continued in the next part

Q2. The comparison can be stated explicitly in the Euclidean specialization. Consider

\gamma_m\asymp m^{-\alpha},\qquad \eta_m=\gamma_m/E_m,$$ with $0\le\beta < 1$ and $1/2 < \alpha < 1$. Writing $s_T=\sqrt{t_T}/T$, the two conditions displayed in Assumption 3.4(iv) of Li et al. (2022) are $$s_T\sum_{m\le T}\gamma_m\longrightarrow0, \qquad \frac{s_T}{\sqrt{\gamma_T}}\longrightarrow0.$$ Substituting the power schedules gives $$s_T\sum_{m\le T}\gamma_m \asymp T^{(1+\beta)/2-\alpha}, \qquad \frac{s_T}{\sqrt{\gamma_T}} \asymp T^{(\alpha+\beta-1)/2}.$$ Consequently, those displayed sufficient conditions require $$\frac{1+\beta}{2} < \alpha < 1-\beta,$$ which admits a choice of $\alpha$ only when $\beta < 1/3$. The table concerns the power schedules $E_m=\max\\{1,\lfloor c_E m^\beta\rfloor\\}$ and $\gamma_m=c_\gamma(m+m_0)^{-\alpha}$, with fixed offsets and the other geometric and oracle assumptions retained. The polynomial-schedule result gives the following comparison of sufficient hypotheses: | Schedule conditions | Admissible $\alpha$ for a given $\beta$ | Range of $\beta$ admitting such an $\alpha$ | |---|---|---| | Li et al., displayed Assumption 3.4(iv) | $(1+\beta)/2 < \alpha < 1-\beta$ | $0\le\beta < 1/3$ | | Our result, $0 < \delta < 1$ | $(1+\beta)/2 < \alpha < \min\\{1,(1-\beta)/(1-\delta)\\}$ | $0\le\beta < (1+\delta)/(3-\delta)$ | | Our result, $\delta\ge1$ | $(1+\beta)/2 < \alpha < 1$ | $0\le\beta < 1$ | The improvement follows from retaining the within-round variance reduction in $$\mathbb E\\|\Delta_m\\|^2 =O\\!\left(\eta_m+\gamma_m^{1+\min(\delta,1)}\right).$$ Random scaling might be a promising way to reduce the additional communication required for inference. Li et al.’s construction relies on a functional central limit theorem for the iterate path and uses the communication-dependent variance clock. A natural extension would represent the synchronized iterates in our fixed-anchor chart, $$v_m=R_a^{-1}(x_m)\in T_a\mathcal M,$$ and construct a path based normalizer in this common vector space. Its variance clock would depend on $$A_m=\sum_{j=1}^{m}E_j^{-1}.$$ This construction could use synchronized iterates already available at the server and therefore avoid additional fresh score and Hessian transmissions. Its validity requires a functional limit and uniform control of the geometric and averaging remainders over partial sums. The present terminal CLT establishes the endpoint distribution; the functional limit and uniform remainder bounds are the additional steps needed for this extension. Joint $d$-dimensional random scaling generally still requires a $d\times d$ normalization matrix at the server, even when it eliminates the $d^2$ Hessian messages from clients. Our current plug-in construction provides an explicit covariance estimator and Wald regions under the results established in the paper. **Q3.** Our current guarantee applies to a prespecified tail fraction. A pilot-free selector that prevents finite-sample undercoverage remains open. In the reported experiments, $\rho$ is fixed before the final replications. The ablations show that its effect can be substantial: with the square-root schedule and gain held fixed, changing $\rho$ from $0.1$ to $0.5$ increases coverage from $71.0\\%$ to $96.2\\%$ in the Gaussian experiment and from $81.2\\%$ to $95.4\\%$ on the sphere. These results quantify the gain from a half-tail in the specified configurations. For the polynomial schedule $E_m\asymp m^\beta$, $0\le\beta < 1$, a prespecified retained tail after $\lfloor\rho T\rfloor$ uses variance factor $\nu_\rho=\\{1-\rho^{1-\beta}\\}/\\{(1-\rho)^2(1-\beta^2)\\}$ in the $\sqrt{t_T}$ limit, subject to the retained-range endpoint and remainder conditions. The covariance estimate and Wald radius use this same factor. Finite-sample initialization bias and inference after selecting $\rho$ from the same trajectory require additional control. Comparing successive tail estimates and covariance estimates could provide useful stabilization diagnostics, but turning such diagnostics into an inferentially valid selection rule would require additional analysis. Continued in the next part

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.