acceptodds
Under review as a conference paper at ICLR 2027

Online Inference with Momentum and Nonsmooth Scores

Abstract

Momentum changes the transient behavior of stochastic gradient algorithms. We ask when these changes preserve statistical inference after Polyak–Ruppert averaging. We develop a framework for fixed stable score filters that includes SGD, fixed preconditioning, scalar and coordinate momentum, and higher-order filters. Under explicit local and trajectory conditions, the filter gain cancels from the first-order influence function. Averaged estimators therefore share the usual sandwich covariance for the same estimating equation and retained observations. The analysis permits nondifferentiable or discontinuous sample scores by imposing differentiability on their population mean and local quadratic-mean continuity. We construct one-pass sandwich intervals, including an online Jacobian estimator for quantile regression that leaves the nonsmooth update unchanged. An integrated process limit also justifies path-based self-normalization without the additional maximal boundary condition used here to establish uniform path convergence. Simulations show lower error for momentum in one least-squares setting with , and online quantile coverage near the nominal level across the reported regular settings.

Primary area:
optimization
Keywords:
Nonsmooth stochastic approximation, Momentum, Online learning, Self normalization

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 4 reviews, above 85% of the others.
Reviewer_PyThOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates statistical inference for Polyak–Ruppert averages of stochastic recursions based on estimating equations. These recursions incorporate fixed, stable, finite-dimensional score filters—such as scalar or coordinate-wise momentum terms, fixed preconditioners, and higher-order filters. Furthermore, the work develops a Wald-type online inference approach that employs an estimator for the Jacobian matrix based on residual kernels—specifically within the context of quantile regression.

Critical strengths
  1. The transformation ut=et−ηtDvtu_t=e_t-\eta_tDv_t leads to the stated exact recursion, and $ (KA)^{-1}K=A^{-1} $ correctly accounts for "gain cancellation" without requiring KK and AA to commute. The proof of Theorem 3.4 incorporates the relevant boundary terms, terms accounting for variations in step size and scores, as well as the nonlinear residual terms.

  2. The requirement that the "population score"—as opposed to the scores of individual samples—must be differentiable allows for the inclusion of quantile regression in a natural way. In this context, the computable evaluation of the kernel-Jacobian matrix is ​​significant.

  3. The paper correctly treats the endpoint as an independent coordinate, given that evaluation at s=1s=1 is not continuous in L2L^2 space. The mapping onto the integrated self-normalizer is continuous within the proposed topology; furthermore, the proof addresses the positive-definite property of the limiting matrix, which is based on a Brownian bridge.

  4. The paper distinguishes between covariance consistency and asymptotic normality, and provides a fully specified example of a projected quantile regression. Moreover, the paper critically examines the limitations of its simplified reference models, the use of a fixed numerical ridge regularization, and the illustration of the approach using real-world data.

Critical weaknesses
  1. Assumption 3.3 presupposes long-term localization, a mean-square convergence rate with ϱ>max⁡(0,2α−1)\varrho > \max(0, 2\alpha-1), a uniformly bounded second moment of the filter state, and a negligible accumulated nonlinear residual term. Demonstrating these properties is a fundamental aspect of the stochastic approximation problem. While Theorems C.5 and C.6 provide useful sufficient conditions for the scalar momentum approach, no such proof exists for coordinate-wise filters or two-pole filters.

  2. The diagnostic procedure based on equal-sized blocks employs a growth rule for the block size—a rule that the article itself acknowledges may not dominate the correlation length of the trajectory. Similarly, the approach using smoothed quantiles utilizes an uncorrected interval and does not represent a calibrated implementation of the cited methods. Consequently, the resulting significant undercoverage (falling well below the nominal coverage probability) offers only limited evidence of an advantage over the aforementioned methods. Furthermore, Table 2 compares Wald-type intervals—based on the average of extreme values ​​(“tail-average”)—with self-normalized intervals based on the average over the entire range (“full-average”); in this comparison, both the point estimator and the normalization method are varied. Although this is addressed in the text, it precludes a clear attribution of the observed differences in coverage probability. Providing data on interval lengths and bias, as well as comparisons based on identical averaging windows, would be more informative than focusing exclusively on coverage probability and RMSE. 3. The greatest improvement is observed in a scenario selected via preliminary tests (pilot experiments) based on the least-squares method; Appendix A.1 notes that the selection criterion also included a requirement for bias reduction. While independent replications confirm the result reported for this specific configuration, they do not demonstrate robustness against slightly altered settings or separately optimized SGD-based benchmarks (baselines).

  3. The simulations employ a fixed ridge parameter, even though consistency would require the condition λn→0\lambda_n \to 0. In the general case, a fixed λ>0\lambda > 0 leads to the following limiting approximation of the inverse: $ R_\lambda = (A^\top A + \lambda I)^{-1} A^\top, $ which results in a covariance matrix RλSRλ⊤R_\lambda S R_\lambda^\top that generally differs from A−1SA−⊤A^{-1} S A^{-\top}.

  4. Minor points:

  • Specify the matrices (B,C,P)(B, C, P) used in the experiments involving coordinate and two-pole filters; the specific coefficients for the coordinate filters are not provided in the manuscript. – Reformulate the requirement for a positive ridge parameter that approaches zero in Theorem 3.11 so that the statement regarding the consistency of the inverse is self-contained.

  • Include a concise table summarizing the assumptions, distinguishing between the requirements imposed on the estimator (central limit theorem, covariance consistency, limit for integrated paths, and limit for uniform paths). - Consolidate redundant definitions and move the verified example of a non-smooth case closer to the main results. – Correct the incomplete phrasing “supports both inference and self-normalized inference” in Section 5, line 477.

Questions for authors
  1. Can you state the fundamental stability and convergence rate conditions for the coordinate and two-pole filters, and explain which non-projected simulation configurations are covered by a complete verification of Assumption 3.3?

  2. Can you provide a specific model or a set of trajectories for which Theorem 3.7 holds, but the additional maximal condition (20) is not satisfied—or for which even stronger assumptions are required?

  3. Does the advantage gained by the least-squares approach at d=100d=100 persist when varying step sizes, momentum coefficients, dimensions, and time horizons—even if the SGD method is granted its own dedicated optimization budget?

  4. How sensitive are the coverage results to a "vanishing ridge scheme" and to the kernel bandwidth? Can you quantify the numerical objective error in the experiments involving misspecified logistic models, and provide the exact filter matrices required to reproduce the comparisons?

8 comments

Thank you for your time and valuable comments.

Note: Because of the 5000-character limit, we divided our response into six comments.

W1. The added Proposition R1 fills this gap for compactly projected fixed filters. It establishes the state second-moment bound, the mean-square rate, localization, and accumulated-bias control under explicit conditions on KgKg and KξK\xi. It also proves that projection eventually becomes inactive, so the limiting inference uses the original estimating equation. The statement and proof for this projected class are given under Q1 in Parts 4/6–5/6.

Table R6 compares the assumptions with established averaging and online-inference results. In particular, high-level trajectory confinement and mean-square bounds also appear explicitly in An–Huo's dynamic-preconditioning analysis. Our claim is that the conditions have clear mathematical roles and are now verified for the specified projected class; the other unprojected nonlinear examples remain conditional.

The following comparisons explain why local regularity, trajectory control, and covariance regularity are separated. Each comparison identifies the relevant assumptions and the verification required for its own setting.

Table R6. Verified assumption comparisons with related averaging and online-inference results.

ReferenceRelevant condition and comparison
Polyak and Juditsky (1992), Assumptions 3.1–3.3Lyapunov stability, local mean-field expansion with a higher-order remainder, and conditional mean-square control of the state-dependent noise perturbation. These are classical precedents for the roles of our local and trajectory conditions.
Fort (2015), Theorem 3.2 and AVER2Averaging is formulated conditional on convergence to a target, with an accumulated perturbation negligible on the square-root scale. This supports the conditional architecture, while leaving filter-specific verification necessary.
Chen et al. (2020), Assumptions 3.1–3.2Strong population convexity, a globally Lipschitz population gradient, differentiable sample losses, and covariance/fourth-moment noise conditions. These are sufficient regularity conditions for their inference analysis. Our local mean-score and finite-moment formulation has a different scope and retains explicit trajectory obligations.
An and Huo (2026b), Assumptions 4–5Local mean expansion with trajectory confinement, and the rate E∥xt−x∗∥2≤Ct−α\mathbb E\lVert x_t-x^{\ast}\rVert^2\le C t^{-\alpha}, followed by primitive sufficient conditions. This is a particularly close precedent for separating a high-level inferential theorem from stability verification.
Lee et al. (2025), Assumption 3.1Their quantile uniform-path theorem uses density regularity, local curvature, and E[(∥X∥6+1)exp⁡(∥X∥2)]<∞\mathbb E[(\lVert X\rVert^6+1)\exp(\lVert X\rVert^2)] < \infty. Our integrated-path route uses finite-moment conditions instead, conditional on the stated trajectory properties. Primitive stability for an unprojected quantile recursion requires its own verification.

W2. A tail average means the last 7070\\% of iterations, rather than extreme-valued observations. In the original quantile comparison, kernel and oracle intervals share the same tail-average estimator, but the self-normalized interval uses the full average.

Table R2 therefore holds the full-average point estimator fixed and reports both interval lengths and intercept bias. It shows that neither normalization uniformly resolves the upper-tail undercoverage. The tail-Wald results on the same paths give the additional window diagnostic. Separately, Table R3 retains the oracle-tuned least-squares comparison, while Table R1 gives the new nonsmooth comparison with data-only tuning. These comparisons isolate the estimator and averaging-window choices. The equal-block and smoothed procedures remain diagnostic; a calibrated external comparison on a common recursion and window remains open.

Continued in Part 2/6: W2 continued: Tables R1–R3; W3, Table R4, and the W4 introduction.

Part 2/6 of our response to Reviewer PyTh (continued): W2 continued: Tables R1–R3; W3, Table R4, and the W4 introduction.

Table R1. Quantile regression with data-only tuning and fresh paired evaluation. Coverage is for the nominal 9595\\% first-slope Wald interval; RMSE is (500d)−1∑r=1500∣θˉr−θτ∣21/2\\{(500d)^{-1}\sum_{r=1}^{500}\\|\bar\theta_r-\theta_\tau\\|^2\\}^{1/2} and length is the mean interval length. Dimension is d=5d=5 except in the last row. All methods use the nonsmooth score, the same inference construction and the same selection budget.

SGDMomentum
SettingCoverageRMSELengthCoverageRMSELength
Median t3t_3, het.0.9560.004780.025880.9500.004870.02588
Upper exponential, het.0.9540.018890.092650.9740.018060.09252
Upper exponential, hom.0.9520.019800.083420.9600.019770.08348
Lower exponential, hom.0.9700.002270.009410.9540.002280.00930
Median t3t_3, het., d=100d=1000.9580.004630.027930.9680.004610.02784

Table R2. Additional upper-quantile diagnostics for normalized momentum, τ=0.9\tau=0.9. Wald and self-normalized (SN) full-window intervals share the same point estimator. These are fresh evaluations, distinct from the original quantile evaluations. The tail average uses the last 7070\\% of iterations, Nn=28000N_n=28000; the Jacobian uses the full history.

DiagnosticAsymmetric, heteroscedastict3t_3
Full-average Wald coverage0.8420.768
Full-average SN coverage0.7900.848
Wald minus SN coverage0.052−0.080-0.080
Full-average Wald mean length0.049090.03387
Full-average SN mean length0.054210.05029
Full-average intercept bias0.010050.00999
Tail-average Wald coverage0.9360.964
Tail-average intercept bias0.007210.00405

Table R3. Paired comparisons in the high-dimensional least-squares setting. Differences use momentum minus SGD. RMSE differences use the unrounded estimates and can differ slightly from subtraction of displayed RMSEs. Both tuning rules used simulation truth.

QuantityOriginal common tuningSeparate equal-budget tuning
Coverage difference0.0540.076
RMSE difference−0.01998-0.01998−0.001309-0.001309
Mean length: momentum0.065590.06644
Mean length: SGD0.067110.06522

W3. For the general inverse in equation (5), a fixed λ>0\lambda > 0 leads to

\qquad R_\lambda S R_\lambda^\mathsf T,$$ which generally differs from $A^{-1}SA^{-\mathsf T}$. For the direct quantile inverse, a fixed ridge similarly gives $(A_\tau+\lambda I)^{-1}$ rather than $A_\tau^{-1}$. The fixed $10^{-6}$ ridge in the submitted experiments is a numerical choice; the theoretical guarantee uses a positive vanishing ridge. We state the ridge condition explicitly: “Let $\lambda_n > 0$ be a deterministic sequence with $\lambda_n\to0$.” The additional paired check in Table R4 compares the fixed ridge with two sequences that satisfy this condition. Their empirical coverages agree in this design and their interval lengths are close. This addresses sensitivity at the reported horizon without changing the theorem's requirement. **Table R4.** Ridge sensitivity for the upper-tail asymmetric heteroscedastic quantile design (intercept, matching Table 2), using paired recursions and direct ridge of the online quantile Jacobian. | Ridge | Coverage | Mean interval length | | --- | --- | --- | | Fixed $10^{-6}$ | 0.948 | 0.05865 | | $\lambda_n=n^{-1}$ | 0.948 | 0.05864 | | $\lambda_n=n^{-3/4}$ | 0.948 | 0.05847 | **W4.** The filter realizations in Part 3/6 give explicit normalized coordinate and two pole realizations and their gains. The coordinate coefficients reported in the numerical response are $\beta_j=0.5+0.45(j-1)/(d-1)$ for the simulated $d > 1$, and the cascade poles are $(a,b)=(0.5,\beta)$. We specify the stage order directly in the equations in Part 3/6: the second stage uses the current first-stage state. **Ridge condition.** We state explicitly: “Let $\lambda_n > 0$ be a deterministic sequence with $\lambda_n\to0$.” *Continued in Part 3/6: W4 continued: assumption map, Table R5, filter realizations, and Q1 introduction.*

Part 3/6 of our response to Reviewer PyTh (continued): W4 continued: assumption map, Table R5, filter realizations, and Q1 introduction.

Assumption map. Table R5 distinguishes the requirements for the estimator limit, covariance consistency, integrated-path inference, and a uniform-path upgrade. Wald coverage combines covariance consistency with the estimator CLT.

Table R5. Conditions for the four guarantees, under the standing i.i.d. and predictable-iterate setup. The theorem and assumption identifiers retain their original numbering.

GuaranteeRequirements and result
ALR and estimator CLTAssumptions 2.1 and 3.1–3.3, and retained-window condition (2); Theorem 3.4.
Covariance consistencyA root of the population score, nonsingular AA, finite target-score second moment, and Assumption 3.5, including score L2L^2 convergence, Jacobian consistency, and a positive vanishing ridge; deterministic windows with Nn→∞N_n\to\infty. Theorem 3.6. Quantile Jacobian consistency uses the additional conditions in Theorem 3.11. Wald coverage also requires the estimator CLT.
Integrated path and endpointAssumptions 2.1 and 3.1–3.3; Theorem 3.7. For the self-normalized pivots, the relevant contrast covariance is nondegenerate; Theorem 3.8.
Uniform-path upgradeTheorem 3.7's conditions plus maximal condition (20); Proposition C.8.

Organization and wording. We make the concrete nonsmooth example explicit and separate its conditions from the inferential conclusions. The concise explanation reads:

“For the bounded-regressor quantile model with shifted logistic errors, compact convex projection gives positive radial drift, a bounded score second moment, and a quadratic mean remainder. Proposition R1 therefore applies to fixed stable normalized coordinate-momentum and two pole filters with K=IdK=I_d and α∈(1/2,1)\alpha\in(1/2,1). The bounded, Lipschitz density also verifies the feasible kernel-Jacobian conditions. This supplies a concrete projected nonsmooth model supporting both Wald and self-normalized inference.”

The full model and verification are given under Q1 in Parts 4/6–5/6. We complete the concluding statement as: “The framework supports both Wald and self-normalized inference.” The condition map above distinguishes the estimator CLT, covariance consistency, integrated-path limit, and uniform-path upgrade.

Coordinate momentum. Let B=diag⁡(β1,…,βd)B=\operatorname{diag}(\beta_1,\ldots,\beta_d), C=Id−BC=I_d-B, and P=IdP=I_d, with 0≤βj<10\le\beta_j < 1. Then

vt=Bvt−1+(Id−B)Gt,K=P(Id−B)−1C=Id.v_t=Bv_{t-1}+(I_d-B)G_t,\qquad K=P(I_d-B)^{-1}C=I_d.

The numerical coefficients reported in the response are βj=0.5+0.45(j−1)/(d−1)\beta_j=0.5+0.45(j-1)/(d-1) for d>1d > 1.

two pole cascade with contemporaneous second-stage input. An explicit normalized realization is

m_t^{(2)}=b m_{t-1}^{(2)}+(1-b)m_t^{(1)},$$ with update direction $m_t^{(2)}$. For $v_t=((m_t^{(1)})^\mathsf T,(m_t^{(2)})^\mathsf T)^\mathsf T$, this gives $$B=\begin{pmatrix}aI_d&0\\\a(1-b)I_d&bI_d\end{pmatrix},\qquad C=\begin{pmatrix}(1-a)I_d\\\\(1-a)(1-b)I_d\end{pmatrix},\qquad P=\begin{pmatrix}0&I_d\end{pmatrix}.$$ For $a,b\in[0,1)$, $\rho(B)=\max\\{a,b\\} < 1$ and $K=I_d$. The reported poles are $a=0.5$ and $b=\beta$. These equations specify the realization used for the stability verification. If the second stage instead uses $m_{t-1}^{(1)}$, the lower-left block of $B$ is $(1-b)I_d$ and the lower block of $C$ is zero; the gain is still $I_d$, but the finite sample trajectory differs. Thus the stage order must accompany the poles when documenting the numerical implementation. **Q1.** For either normalized filter, $K=I_d$. Under compact convex projection, Proposition R1 requires positive radial drift of $g$ on that set, a linear bound on its magnitude, a quadratic second-moment bound on the centered score, and the stated local mean remainder. The conclusion is $\mathbb E\lVert e_t\rVert^2=O(t^{-\alpha})$ and a uniformly bounded filter-state second moment; $\alpha(1+\nu) > 1$ gives the negligible accumulated bias. The argument requires both stability of $B$ and the stated drift condition. Among the submitted unprojected configurations, the scalar theorem completely verifies SGD and normalized scalar momentum in the submitted least-squares designs. Primitive verification remains open for the submitted unprojected logistic, quantile, coordinate and two pole configurations. The least-squares moment check and the projected nonsmooth example are given in Part 5/6. *Continued in Part 4/6: Q1 continued: Proposition R1 and the first proof steps.*
Reviewer_G8KbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper explores when momentum-type modifications to SGD preserve valid statistical inference after Polyak--Ruppert averaging, including when sample scores are nonsmooth or discontinuous (e.g., quantile regression). It studies a general class of fixed linear score filters.

The main theory contributions are: Common asymptotic linearity (Thm.3.4); feasible Wald inference (Thms. 3.6, 3.11); and self-normalized (random-scaling) inference (Thms.3.7--3.8). The simulation studies consider 15 configurations with 500 replications each. The tuning (α=0.6\alpha = 0.6, 30% burn-in, step size and horizon multipliers) was chosen in independent pilot runs whose selection criterion included coverage. The real data example is on the UCI combined-cycle power-plant data, the method is used for descriptive quantile inference across five supplied row orders.

Critical strengths

Strength 1 : The article gives a clean, general mechanism for why filtering does not change the asymptotic covariance. The exact change of variables ut=et−ηtDvtu_t = e_t - \eta_t D v_t (Prop. C.1) is purely algebraic, and it shows directly that a single gain KK multiplies both the local drift KAeKAe and the noise Kξ⋆K\xi^\star. The gain therefore cancels from the influence function, leaving −A−1ξt⋆-A^{-1}\xi_t^\star. This is a structural explanation rather than a case-by-case asymptotic argument, so one proof covers SGD, fixed preconditioning, scalar and coordinate momentum, and higher-order filters at once.

Strength 2: The article gives feasible one-pass inference for nonsmooth quantile regression that holds up empirically. The online kernel Jacobian (Thm. 3.11) confines smoothing to covariance estimation, so the parameter recursion keeps the original indicator score. Because the kernel is evaluated at predictable iterates, the bias bound is $C|\theta_{t-1} - \theta_\tau|^{\nu_A} + C h_t^{\nu_A}, $ and no rate condition linking parameter error to bandwidth is needed.

Strength 3: In the simulations, the resulting Wald intervals cover at 0.940--0.972 across seven quantile settings, including heavy-tailed, asymmetric, and heteroscedastic designs.

Critical weaknesses

A main result (Thm. 3.4) depends on Assumption 3.3. That assumption imposes eventual localization, a mean-square rate E∥et∥2≤Let−ϱ\mathbb{E}\|e_t\|^2 \le L_e t^{-\varrho} with ϱ>2α−1\varrho > 2\alpha - 1, a uniform filter-state bound sup⁡tE∥vt∥2<∞\sup_t \mathbb{E}\|v_t\|^2 < \infty, and negligibility of the accumulated nonlinear remainder (11). Given these conditions, the remaining argument is largely the algebraic change of variables (Prop. C.1) followed by standard Abel-summation and martingale steps. In other words, the stability analysis, which is where momentum and higher-order filters actually differ from SGD, is assumed rather than proved. The paper verifies these conditions only for scalar}momentum, and only in two cases. The incremental contribution over Tang et al. (2024) is modest. It amounts to (i) a gain-cancellation identity that is elementary once written down and (ii) an extension to nonsmooth scores whose trajectory requirements are left largely unverified.

The experiments do little to show that filtering matters for inference. In 14 of 15 configurations, gain-matched filters produce RMSE and coverage that are identical or nearly identical to SGD at the reported precision. The single setting with a clear benefit is least squares with d=100d = 100 and β=0.99\beta = 0.99. Even there, the batch estimator remains better, and batch quantile regression often has lower RMSE than every recursive method.

Questions for authors

Can Assumption 3.3 be verified for coordinate momentum, the two-pole cascade, or general stable (B,C,P)(B, C, P)? It would be enough to do this under projection, for example through a Lyapunov argument that relies only on ρ(B)<1\rho(B) < 1 together with a drift condition on KgK g. If no such result is available, please state precisely which filters the paper's results cover unconditionally.

The logistic and quantile simulations run unprojected recursions, which lie outside Thms.C.5--C.6. Do the conclusions in Tables 1, 2, and 5 change under the projected recursions that the theory covers? Alternatively, is there a projection-free stability result for quantile regression, perhaps using the boundedness of the quantile score?

SGD and momentum share η0=0.16\eta_0 = 0.16, a value chosen by the pilot. Does SGD with its own tuned η0\eta_0 (or α\alpha) close the gap in RMSE (0.037 vs. 0.017) and coverage (0.910 vs. 0.964)? If it does, the only empirical evidence that filtering helps would reduce to a difference in stepsize tuning.

Beyond the extension to nonsmooth scores, what is technically new relative to Tang et al. (2024)? Please also clarify the relationship to An & Huo (2026a,b), which study the Polyak--Ruppert CLT under momentum and dynamic preconditioning. In particular, does any result in the present paper follow as a special case of those works?

9 comments

Thank you for your time and valuable comments.

Note: Because of the 5000-character limit, we divided our response into six comments.

W1. The main inferential theorem and the primitive stability results have different roles. Theorem 3.4 identifies conditions under which filtering preserves inference for a given score and averaging window. The submitted Theorems C.5–C.6 verify those conditions for scalar momentum. Applying the general inferential theorem requires verification of its trajectory conditions for the chosen filter.

We address the missing projected verification directly through Proposition R1. The proof must handle a stable matrix BB that need not contract in Euclidean norm and a projection correction that enters the transformed recursion. A Lyapunov norm controls the first issue; the bound

∥ut∥2≤∥at∥2+ηt2∥Dvt∥2\lVert u_t\rVert^2\le \lVert a_t\rVert^2+\eta_t^2\lVert Dv_t\rVert^2

controls the second, where ata_t is the transformed update before the projection correction. Together they give the state bound, the O(ηt)O(\eta_t) rate, and eventual inactivity of projection for general fixed filters under the stated drift conditions.

The contribution beyond an endpoint equivalence also includes the joint L2L^2-path and endpoint limit and the feasible matrix Jacobian for the unchanged quantile recursion. The technical obstacles and the precise relationship to prior results are detailed below. Question 4 in Part 5/6 addresses Tang and An–Huo separately.

Our contribution is to establish a usable inference framework across finite memory filters and discontinuous estimating equations, while retaining the original score update. The substantive work lies in the following three steps.

  1. Controlling the filtered recursion without sample score differentiability. The quantile indicator is discontinuous, so the analysis uses the population mean expansion and score-continuity bounds. In addition, the filter introduces a state whose weighted boundary and stepsize variation terms must vanish on the inferential scale. Theorem 3.4 combines the exact state transformation with predictable localization, local quadratic mean continuity of the centered score, a mean square trajectory rate, and control of accumulated nonlinear bias. These arguments yield the influence representation for the original average, including a retained averaging window and matrix-valued filters. The sample loss smoothness in Assumption A1 of Tang et al. (2024) excludes the quantile indicator score; their scalar momentum equivalence is an important starting point for this extension.

  2. Making quantile inference feasible without smoothing the parameter update. The indicator score requires a separate estimator of Aτ=Ef(XTθτ∣X)XXTA_\tau=\mathbb E\\{f(X^\mathsf T\theta_\tau\mid X)XX^\mathsf T\\}. Theorem 3.11 controls a covariate weighted matrix kernel estimator evaluated along predictable, filtered iterates. Integrating over the fresh response gives a conditional bias bound of the form

C∥θt−1−θτ∥νA+ChtνA,C\lVert \theta_{t-1}-\theta_\tau\rVert^{\nu_A}+C h_t^{\nu_A},

which avoids a requirement that the parameter error be o(ht)o(h_t). The proof also controls the matrix martingale fluctuations under the weighted pp-moment condition, for some p>1p > 1. Recursive scalar density estimators already appear in Amiri and Thiam (2014) and Wei et al. (2026); the inference method is our contribution here.

  1. Obtaining the joint limit required by the self-normalizer. The path-based denominator requires the endpoint jointly with the L2([0,1])L^2([0,1]) path limit. Theorems 3.7–3.8 establish both jointly and control the integrated filter remainders. This extends the established random-scaling construction of Lee et al. (2025) to the present filters without imposing the additional maximal boundary condition used in our uniform-path result.

W2. Agreement in most submitted settings is consistent with the common influence function. Batch estimators have lower displayed RMSE in most settings, with some ties after rounding. They use all observations, whereas the recursive Wald estimates retain the last 7070\\% of iterates; the comparison therefore changes both the estimator and the information retained by its average.

The additional smooth-score evidence is the oracle-tuned least-squares comparison retained in Table R3. It preserves a smaller advantage for momentum in that setting. Table R1 separately reports the nonsmooth experiment with data-only tuning. For quantile regression, the supported contribution is feasible online inference with the nonsmooth update: scalar-momentum kernel coverage is 0.9460.946–0.9700.970 in the submitted designs, close to its same-estimator oracle comparison.

Continued in Part 2/6: W2 continued: Tables R1 and R3; Q1 and Proposition R1.

Part 2/6 of our response to Reviewer G8Kb (continued): W2 continued: Tables R1 and R3; Q1 and Proposition R1.

Table R1. Quantile regression with data-only tuning and fresh paired evaluation. Coverage is for the nominal 9595\\% first-slope Wald interval; RMSE is (500d)−1∑r=1500∣θˉr−θτ∣21/2\\{(500d)^{-1}\sum_{r=1}^{500}\\|\bar\theta_r-\theta_\tau\\|^2\\}^{1/2} and length is the mean interval length. Dimension is d=5d=5 except in the last row. All methods use the nonsmooth score, the same inference construction and the same selection budget.

SGDMomentum
SettingCoverageRMSELengthCoverageRMSELength
Median t3t_3, het.0.9560.004780.025880.9500.004870.02588
Upper exponential, het.0.9540.018890.092650.9740.018060.09252
Upper exponential, hom.0.9520.019800.083420.9600.019770.08348
Lower exponential, hom.0.9700.002270.009410.9540.002280.00930
Median t3t_3, het., d=100d=1000.9580.004630.027930.9680.004610.02784

Table R3. Paired comparisons in the high-dimensional least-squares setting. Differences use momentum minus SGD. RMSE differences use the unrounded estimates and can differ slightly from subtraction of displayed RMSEs. Both tuning rules used simulation truth.

QuantityOriginal common tuningSeparate equal-budget tuning
Coverage difference0.0540.076
RMSE difference−0.01998-0.01998−0.001309-0.001309
Mean length: momentum0.065590.06644
Mean length: SGD0.067110.06522

Q1. Yes, with a drift condition on the effective mean score KgKg, not from ρ(B)<1\rho(B) < 1 alone. Proposition R1 states sufficient conditions and provides a complete proof. On a known compact convex set C\mathcal C containing the target in its interior, it assumes

\qquad \lVert Kg(\theta_\star+e)\rVert\le L\lVert e\rVert,$$ and a quadratic second-moment bound on $K\xi$. The matrix $W=\sum_{j\ge0}(B^j)^\mathsf T B^j$ supplies the Lyapunov norm. For a remainder $O(\lVert e\rVert^{1+\nu})$ and $\alpha(1+\nu) > 1$, the result verifies Assumption 3.3 with $\varrho=\alpha$ and transfers the CLT and integrated path limit to the projected recursion. The normalized coordinate and cascade constructions in the filter realizations in Part 4/6 have $K=I_d$. Thus the projected quantile model of Example C.18 satisfies these conditions for both. The result concerns each fixed filter; its constants are not uniform as the filter approaches instability. **Proposition R1 (Projected fixed filters).** Use the paper's measurable-score and i.i.d. observation setup. Let $\mathcal C\subset U$ be a known compact convex set with $\theta_\star\in\operatorname{int}(\mathcal C)$. Consider $$v_t=Bv_{t-1}+CG(\theta_{t-1},Z_t),\qquad \theta_t=\Pi_{\mathcal C}(\theta_{t-1}-\eta_tPv_t),\qquad \eta_t=\eta_0t^{-\alpha},\quad \eta_0 > 0.$$ The matrices are fixed, $\rho(B) < 1$, and $K=P(I-B)^{-1}C$ is nonsingular. Suppose $\theta_0\in\mathcal C$, $\mathbb E\lVert v_0\rVert^2 < \infty$, and the initialization is independent of the observations. Write $\xi(\theta,Z)=G(\theta,Z)-g(\theta)$. Assume that, for every $e$ with $\theta_\star+e\in\mathcal C$, $$\begin{aligned} \langle e,Kg(\theta_\star+e)\rangle&\ge\mu\lVert e\rVert^2, &\lVert Kg(\theta_\star+e)\rVert&\le L\lVert e\rVert,\\\\ \mathbb E\lVert K\xi(\theta_\star+e,Z)\rVert^2&\le\sigma^2+L_\xi^2\lVert e\rVert^2, \end{aligned}$$ where $\mu > 0$ and all constants are finite. Suppose also that $g(\theta_\star+e)=Ae+r(e)$ on $U$, with nonsingular $A$, and $\lVert r(e)\rVert\le L_r\lVert e\rVert^{1+\nu}$ on $\mathcal C$, where $\nu\in(0,1]$. If $0 < \alpha < 1$ and $\alpha(1+\nu) > 1$, then $$\sup_t\mathbb E\lVert v_t\rVert^2 < \infty,\qquad \mathbb E\lVert \theta_t-\theta_\star\rVert^2=O(\eta_t),\qquad \theta_t\longrightarrow\theta_\star\quad\text{almost surely}.$$ Projection is eventually inactive almost surely, and Assumption 3.3 holds with $\varrho=\alpha$. If Assumption 3.2 also holds, the joint integrated-path and endpoint limit holds; for retained windows satisfying condition (2), the asymptotic linear representation and CLT hold as well. *Continued in Part 3/6: Q1 continued: proof through almost-sure convergence.*

Part 3/6 of our response to Reviewer G8Kb (continued): Q1 continued: proof through almost-sure convergence.

Proof. Filter-state bound. Compactness, the growth and noise bounds, and nonsingularity of KK imply sup⁡θ∈CE∥G(θ,Z)∥2<∞\sup_{\theta\in\mathcal C}\mathbb E\lVert G(\theta,Z)\rVert^2 < \infty. Define

W=∑j=0∞(Bj)TBj.W=\sum_{j=0}^{\infty}(B^j)^\mathsf T B^j.

Then W≻0W\succ0 and BTWB=W−IB^\mathsf T WB=W-I. Thus BB contracts in the WW-norm, even when it does not contract in Euclidean norm. Young's inequality and the conditional score bound give, for some fixed q<1q < 1,

\le q\lVert v_{t-1}\rVert_W^2+C_0.$$ Iteration and equivalence of finite-dimensional norms give the uniform second-moment bound. Throughout this proof, $C_j$ denote finite constants depending on the fixed filter and the stated bounds. **Projection and the transformed recursion.** Let $e_t=\theta_t-\theta_\star$, $D=PB(I-B)^{-1}$, $u_t=e_t-\eta_tDv_t$, and $u_0=e_0-\eta_1Dv_0$. Set $$x_t=\theta_{t-1}-\eta_tPv_t,\qquad q_t=\Pi_{\mathcal C}(x_t)-x_t,\qquad \delta_t=(\eta_{t-1}-\eta_t)Dv_{t-1},$$ with $\eta_0=\eta_1$ in the initial-index convention. The exact identity is $$u_t=a_t+q_t,\qquad a_t=u_{t-1}-\eta_tKG(\theta_{t-1},Z_t)+\delta_t.$$ The projection variational inequality gives $\langle q_t,e_t\rangle\le0$. Consequently, $$\begin{aligned} \lVert u_t\rVert^2 &=\lVert a_t\rVert^2+2\langle u_t,q_t\rangle-\lVert q_t\rVert^2\\\\ &\le\lVert a_t\rVert^2-2\eta_t\langle Dv_t,q_t\rangle-\lVert q_t\rVert^2\\\\ &\le\lVert a_t\rVert^2+\eta_t^2\lVert Dv_t\rVert^2. \end{aligned}$$ This estimate applies to general fixed stable filters, including matrices $B$ without an orthogonal diagonalization. **Mean-square rate.** Since $e_{t-1}=u_{t-1}+\eta_{t-1}Dv_{t-1}$, the drift and growth bounds imply $$\langle u_{t-1},Kg(\theta_{t-1})\rangle \ge\frac\mu4\lVert u_{t-1}\rVert^2 -C_1\eta_{t-1}^2\lVert v_{t-1}\rVert^2.$$ Also $\lVert \delta_t\rVert\le C_2\eta_tt^{-1}\lVert v_{t-1}\rVert$. Expanding $\lVert a_t\rVert^2$ conditionally, using centered fresh noise, and absorbing the cross terms into a small multiple of $\eta_t\lVert u_{t-1}\rVert^2$ yields, for all sufficiently large $t$, $$\mathbb E(\lVert a_t\rVert^2\mid\mathcal F_{t-1}) \le(1-c_0\eta_t)\lVert u_{t-1}\rVert^2 +C_3\eta_t^2+C_4\eta_t^3\lVert v_{t-1}\rVert^2.$$ Here the stepsize-variation cost satisfies \$\\lVert \\delta\_t\\rVert^2/\\eta\_t \\le C\_5\\eta\_t t^{-2}\\lVert v\_{t-1}\\rVert^2 \\le C\_6\\eta\_t^3\\lVert v\_{t-1}\\rVert^2\$, since $\alpha < 1$. The conditional state bound and the projection inequality therefore give $$\mathbb E(\lVert u_t\rVert^2\mid\mathcal F_{t-1}) \le(1-c\eta_t)\lVert u_{t-1}\rVert^2 +C_7\eta_t^2(1+\lVert v_{t-1}\rVert^2).\tag{1}$$ Taking expectations gives $U_t\le(1-c\eta_t)U_{t-1}+C_8\eta_t^2$, where $U_t=\mathbb E\lVert u_t\rVert^2$. Since $\eta_{t-1}/\eta_t=1+O(t^{-1})$ and $t^{-1}=o(\eta_t)$, the deterministic comparison argument gives $U_t=O(\eta_t)$. Then $e_t=u_t+\eta_tDv_t$ gives $\mathbb E\lVert e_t\rVert^2=O(\eta_t)$. **Almost-sure convergence and eventual inactivity.** The condition $\alpha(1+\nu) > 1$, with $\nu\le1$, implies $\alpha > 1/2$. The additive terms in (1) have summable expectations. The nonnegative almost-supermartingale argument gives convergence of $\lVert u_t\rVert^2$ and $\sum_t\eta_t\lVert u_{t-1}\rVert^2 < \infty$ almost surely. Since $\sum_t\eta_t=\infty$, the limit is zero. Moreover, $\mathbb E\sum_t\eta_t^2\lVert v_t\rVert^2 < \infty$, so $\eta_tv_t\to0$ almost surely. It follows that $e_t\to0$ almost surely. An open ball about $\theta_\star$ lies in $\mathcal C$. Eventually both $\theta_{t-1}$ and $x_t$ lie in this ball because $\eta_tPv_t\to0$. Thus $q_t=0$ eventually almost surely. *Continued in Part 4/6: Q1 proof completion, nonsmooth model, and filter realizations; Q2.*
Reviewer_E5CmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies inference from Polyak-Ruppert averages of stochastic recursions with fixed finite-dimensional score filters, including momentum. Under local population-score regularity and explicit trajectory conditions, it establishes a common asymptotic linear representation whose influence function does not depend on the filter. It develops feasible sandwich inference, including a predictable kernel Jacobian estimator for quantile regression that leaves the indicator-score update unchanged. A joint limit for the partial-sum path in L2 and its endpoint also justifies integrated self-normalization. Scalar-momentum stability results, regression specializations, simulations, and a descriptive power-plant quantile example support and delimit the framework.

Critical strengths
  • Theorems 3.4, 3.7, 3.8 give a coherent inference argument for fixed filters and nonsmooth scores. The proof controls the transformed-state boundary terms and establishes the endpoint jointly with the L2 path, providing the joint convergence needed for path normalization.

  • Theorem 3.11 gives a useful feasible quantile-Jacobian construction without smoothing the parameter update. Under the stated density-envelope and weighted-moment assumptions, conditional response integration removes the need to make the parameter error negligible relative to the bandwidth.

Critical weaknesses
  • Applicability remains conditional for much of the filter class in the general formulation. Assumption 3.3 requires localization, a mean-square rate, a bounded filter-state second moment, and negligible accumulated nonlinear bias; stability of B does not establish these. The primitive verification results concern scalar momentum, with projected constructions for logistic/quantile regression, whereas the reported simulations are unprojected and also include other filters. The paper acknowledges this distinction, but it limits how broadly the theoretical guarantee can currently be applied and makes my acceptance recommendation weak.

  • The empirical evidence gives limited guidance on choosing the proposed inference over calibrated alternatives. The equal-block diagnostic in Table 2 and smoothed prototype in Table 5 are explicitly not calibrated implementations of the cited methods, while the momentum improvement is concentrated in one pilot-selected least-squares configuration. Appendix A.1 omits the pilot grid and numerical selection thresholds, and the paired comparisons lack Monte Carlo uncertainty.

Questions for authors
  • Is there a sufficient condition argument beyond Theorems C.5–C.6 for any reported unprojected quantile or non-scalar-filter configuration?
  • What were the pilot grid, objective weights, and numerical stability/coverage/RMSE/bias-reduction thresholds? Please also give paired uncertainty for the highlighted estimation and coverage comparisons. Are results available for a calibrated online covariance baseline using the same recursion and retained averaging window?
9 comments

Thank you for your time and valuable comments.

Note: Because of the 5000-character limit, we divided our response into six comments.

W1. Proposition R1 directly extends the primitive verification to projected general fixed filters. It establishes a uniform filter-state second moment, the mean-square rate O(ηt)O(\eta_t), almost-sure convergence, and eventual inactivity of projection. With the stated Hölder remainder condition, these conclusions verify Assumption 3.3. Its application to Example C.18 gives a concrete nonsmooth example for coordinate and two pole filters, rather than only a conditional theorem for those filters.

The distinction between local regularity and trajectory control has precise precedents. Polyak and Juditsky (1992; Assumptions 3.1–3.3) combine Lyapunov stability, local mean-field expansion, and mean-square noise continuity. More recently, An and Huo (2026b; Assumptions 4–5) explicitly impose trajectory confinement and an iterate mean-square rate, then provide sufficient conditions. Table R6 gives the comparisons. These precedents explain the assumption structure. Our filter-specific verification is as follows: the projected class is verified by Proposition R1, unprojected scalar least squares by Theorem C.5, and the other reported unprojected trajectories remain conditional.

The following comparisons explain why local regularity, trajectory control, and covariance regularity are separated. Each comparison identifies the relevant assumptions and the verification required for its own setting.

Table R6. Verified assumption comparisons with related averaging and online-inference results.

ReferenceRelevant condition and comparison
Polyak and Juditsky (1992), Assumptions 3.1–3.3Lyapunov stability, local mean-field expansion with a higher-order remainder, and conditional mean-square control of the state-dependent noise perturbation. These are classical precedents for the roles of our local and trajectory conditions.
Fort (2015), Theorem 3.2 and AVER2Averaging is formulated conditional on convergence to a target, with an accumulated perturbation negligible on the square-root scale. This supports the conditional architecture, while leaving filter-specific verification necessary.
Chen et al. (2020), Assumptions 3.1–3.2Strong population convexity, a globally Lipschitz population gradient, differentiable sample losses, and covariance/fourth-moment noise conditions. These are sufficient regularity conditions for their inference analysis. Our local mean-score and finite-moment formulation has a different scope and retains explicit trajectory obligations.
An and Huo (2026b), Assumptions 4–5Local mean expansion with trajectory confinement, and the rate E∥xt−x∗∥2≤Ct−α\mathbb E\lVert x_t-x^{\ast}\rVert^2\le C t^{-\alpha}, followed by primitive sufficient conditions. This is a particularly close precedent for separating a high-level inferential theorem from stability verification.
Lee et al. (2025), Assumption 3.1Their quantile uniform-path theorem uses density regularity, local curvature, and E[(∥X∥6+1)exp⁡(∥X∥2)]<∞\mathbb E[(\lVert X\rVert^6+1)\exp(\lVert X\rVert^2)] < \infty. Our integrated-path route uses finite-moment conditions instead, conditional on the stated trajectory properties. Primitive stability for an unprojected quantile recursion requires its own verification.

W2. We distinguish the supported findings: feasible quantile covariance estimation tracks the oracle on the same estimator and window, and a finite sample momentum advantage remains in one high-dimensional stress setting after oracle-based separate tuning. Table R3 retains this smooth-score diagnostic and its interval lengths; Table R1 instead reports the new nonsmooth comparison with data-only selection. The empirical conclusions concern these matched-estimator and separately tuned comparisons.

Question 2 (Parts 4/6–6/6) distinguishes the original truth-based selection from the new validation-loss rule and gives the additional paired comparisons.

Q1. The added result goes beyond the scalar theorems for the projected nonscalar versions: Proposition R1 covers arbitrary fixed stable BB, with the full-set drift and noise conditions stated there. The reported unprojected quantile and nonscalar runs retain their conditional interpretation. For the original unprojected experiments, the complete primitive verification currently applies to SGD and normalized scalar momentum in the submitted least-squares designs, as verified in Part 4/6. We explicitly distinguish the projected guarantee from the conditional interpretation of the unprojected experiments.

Continued in Part 2/6: Q1 continued: Proposition R1 and the first proof steps.

Part 2/6 of our response to Reviewer E5Cm (continued): Q1 continued: Proposition R1 and the first proof steps.

Proposition R1 (Projected fixed filters). Use the paper's measurable-score and i.i.d. observation setup. Let C⊂U\mathcal C\subset U be a known compact convex set with θ⋆∈int⁡(C)\theta_\star\in\operatorname{int}(\mathcal C). Consider

\theta_t=\Pi_{\mathcal C}(\theta_{t-1}-\eta_tPv_t),\qquad \eta_t=\eta_0t^{-\alpha},\quad \eta_0 > 0.$$ The matrices are fixed, $\rho(B) < 1$, and $K=P(I-B)^{-1}C$ is nonsingular. Suppose $\theta_0\in\mathcal C$, $\mathbb E\lVert v_0\rVert^2 < \infty$, and the initialization is independent of the observations. Write $\xi(\theta,Z)=G(\theta,Z)-g(\theta)$. Assume that, for every $e$ with $\theta_\star+e\in\mathcal C$, $$\begin{aligned} \langle e,Kg(\theta_\star+e)\rangle&\ge\mu\lVert e\rVert^2, &\lVert Kg(\theta_\star+e)\rVert&\le L\lVert e\rVert,\\\\ \mathbb E\lVert K\xi(\theta_\star+e,Z)\rVert^2&\le\sigma^2+L_\xi^2\lVert e\rVert^2, \end{aligned}$$ where $\mu > 0$ and all constants are finite. Suppose also that $g(\theta_\star+e)=Ae+r(e)$ on $U$, with nonsingular $A$, and $\lVert r(e)\rVert\le L_r\lVert e\rVert^{1+\nu}$ on $\mathcal C$, where $\nu\in(0,1]$. If $0 < \alpha < 1$ and $\alpha(1+\nu) > 1$, then $$\sup_t\mathbb E\lVert v_t\rVert^2 < \infty,\qquad \mathbb E\lVert \theta_t-\theta_\star\rVert^2=O(\eta_t),\qquad \theta_t\longrightarrow\theta_\star\quad\text{almost surely}.$$ Projection is eventually inactive almost surely, and Assumption 3.3 holds with $\varrho=\alpha$. If Assumption 3.2 also holds, the joint integrated-path and endpoint limit holds; for retained windows satisfying condition (2), the asymptotic linear representation and CLT hold as well. **Proof.** **Filter-state bound.** Compactness, the growth and noise bounds, and nonsingularity of $K$ imply $\sup_{\theta\in\mathcal C}\mathbb E\lVert G(\theta,Z)\rVert^2 < \infty$. Define $$W=\sum_{j=0}^{\infty}(B^j)^\mathsf T B^j.$$ Then $W\succ0$ and $B^\mathsf T WB=W-I$. Thus $B$ contracts in the $W$-norm, even when it does not contract in Euclidean norm. Young's inequality and the conditional score bound give, for some fixed $q < 1$, $$\mathbb E(\lVert v_t\rVert_W^2\mid\mathcal F_{t-1}) \le q\lVert v_{t-1}\rVert_W^2+C_0.$$ Iteration and equivalence of finite-dimensional norms give the uniform second-moment bound. Below, $C_j$ denote finite constants depending on the fixed filter and the stated bounds. **Projection and the transformed recursion.** Let $e_t=\theta_t-\theta_\star$, $D=PB(I-B)^{-1}$, $u_t=e_t-\eta_tDv_t$, and $u_0=e_0-\eta_1Dv_0$. Set $$x_t=\theta_{t-1}-\eta_tPv_t,\qquad q_t=\Pi_{\mathcal C}(x_t)-x_t,\qquad \delta_t=(\eta_{t-1}-\eta_t)Dv_{t-1},$$ with $\eta_0=\eta_1$ in the initial-index convention. The exact identity is $$u_t=a_t+q_t,\qquad a_t=u_{t-1}-\eta_tKG(\theta_{t-1},Z_t)+\delta_t.$$ The projection variational inequality gives $\langle q_t,e_t\rangle\le0$. Consequently, $$\begin{aligned} \lVert u_t\rVert^2 &=\lVert a_t\rVert^2+2\langle u_t,q_t\rangle-\lVert q_t\rVert^2\\\\ &\le\lVert a_t\rVert^2-2\eta_t\langle Dv_t,q_t\rangle-\lVert q_t\rVert^2\\\\ &\le\lVert a_t\rVert^2+\eta_t^2\lVert Dv_t\rVert^2. \end{aligned}$$ This estimate applies to general fixed stable filters, including matrices $B$ without an orthogonal diagonalization. *Continued in Part 3/6: Q1 continued: remaining proof steps and the nonsmooth model.*

Part 3/6 of our response to Reviewer E5Cm (continued): Q1 continued: remaining proof steps and the nonsmooth model.

Mean-square rate. Since et−1=ut−1+ηt−1Dvt−1e_{t-1}=u_{t-1}+\eta_{t-1}Dv_{t-1}, the drift and growth bounds imply

\ge\frac\mu4\lVert u_{t-1}\rVert^2 -C_1\eta_{t-1}^2\lVert v_{t-1}\rVert^2.$$ Also $\lVert \delta_t\rVert\le C_2\eta_tt^{-1}\lVert v_{t-1}\rVert$. Expanding $\lVert a_t\rVert^2$ conditionally, using centered fresh noise, and absorbing the cross terms into a small multiple of $\eta_t\lVert u_{t-1}\rVert^2$ yields, for all sufficiently large $t$, $$\mathbb E(\lVert a_t\rVert^2\mid\mathcal F_{t-1}) \le(1-c_0\eta_t)\lVert u_{t-1}\rVert^2 +C_3\eta_t^2+C_4\eta_t^3\lVert v_{t-1}\rVert^2.$$ Here the stepsize-variation cost satisfies \$\\lVert \\delta\_t\\rVert^2/\\eta\_t \\le C\_5\\eta\_t t^{-2}\\lVert v\_{t-1}\\rVert^2 \\le C\_6\\eta\_t^3\\lVert v\_{t-1}\\rVert^2\$, since $\alpha < 1$. The conditional state bound and the projection inequality therefore give $$\mathbb E(\lVert u_t\rVert^2\mid\mathcal F_{t-1}) \le(1-c\eta_t)\lVert u_{t-1}\rVert^2 +C_7\eta_t^2(1+\lVert v_{t-1}\rVert^2).\tag{1}$$ Taking expectations gives $U_t\le(1-c\eta_t)U_{t-1}+C_8\eta_t^2$, where $U_t=\mathbb E\lVert u_t\rVert^2$. Since $\eta_{t-1}/\eta_t=1+O(t^{-1})$ and $t^{-1}=o(\eta_t)$, the deterministic comparison argument gives $U_t=O(\eta_t)$. Then $e_t=u_t+\eta_tDv_t$ gives $\mathbb E\lVert e_t\rVert^2=O(\eta_t)$. **Almost-sure convergence and eventual inactivity.** The condition $\alpha(1+\nu) > 1$, with $\nu\le1$, implies $\alpha > 1/2$. The additive terms in (1) have summable expectations. The nonnegative almost-supermartingale argument gives convergence of $\lVert u_t\rVert^2$ and $\sum_t\eta_t\lVert u_{t-1}\rVert^2 < \infty$ almost surely. Since $\sum_t\eta_t=\infty$, the limit is zero. Moreover, $\mathbb E\sum_t\eta_t^2\lVert v_t\rVert^2 < \infty$, so $\eta_tv_t\to0$ almost surely. It follows that $e_t\to0$ almost surely. An open ball about $\theta_\star$ lies in $\mathcal C$. Eventually both $\theta_{t-1}$ and $x_t$ lie in this ball because $\eta_tPv_t\to0$. Thus $q_t=0$ eventually almost surely. **Accumulated remainder and inference.** Lyapunov's inequality gives $$\frac1{\sqrt n}\sum_{t=1}^n\mathbb E\lVert r(e_{t-1})\rVert \le\frac{C_9}{\sqrt n} \left(1+\sum_{t=2}^n t^{-\alpha(1+\nu)/2}\right)=o(1).$$ Markov's inequality verifies the accumulated-remainder condition. Since $\alpha > \max\\{0,2\alpha-1\\}$, all trajectory requirements hold with $\varrho=\alpha$. The projection adds a term proportional to $q_t/\eta_t$ in the estimating-equation decomposition. If $T < \infty$ is its last active time, then $$\frac1{\sqrt n}\max_{k\le n} \lVert \sum_{t=1}^k\frac{q_t}{\eta_t}\rVert \le\frac1{\sqrt n}\sum_{t=1}^T\frac{\lVert q_t\rVert}{\eta_t} \longrightarrow0\quad\text{almost surely}.$$ For a retained window the analogous bound uses $\sqrt{N_n}$, which diverges under (2). The remaining averaging and integrated-path arguments apply to the actual projected trajectory, as in submitted Proposition C.26. ∎ **Concrete nonsmooth verification.** In Example C.18, $X=(1,U_1,\ldots,U_{d-1})^\mathsf T/\sqrt d$ with independent Rademacher coordinates, $\lVert X\rVert=1$, and $\mathbb E XX^\mathsf T=I_d/d$. The quantile target satisfies $\lVert \theta_\tau\rVert\le1$ and $Y=X^\mathsf T\theta_\tau+\varepsilon$, where $\varepsilon=T-\log\\{\tau/(1-\tau)\\}$ and $T$ is standard logistic. On $\mathcal C=\\{\theta:\lVert \theta\rVert\le2\\}$, $$e^\mathsf T g_\tau(\theta_\tau+e)\ge(m_\tau/d)\lVert e\rVert^2, \qquad \lVert g_\tau(\theta_\tau+e)\rVert\le\tfrac14\lVert e\rVert, \qquad m_\tau=\min_{|s|\le3}f_\varepsilon(s) > 0.$$ The score is bounded and the mean remainder is quadratic. Therefore Proposition R1 applies to every fixed stable filter with $K=I_d$, including the coordinate-momentum and two pole constructions, for $\alpha\in(1/2,1)$. The bounded, Lipschitz density also verifies the kernel-Jacobian conditions with, for example, $h_t=h_0t^{-1/3}$, $\lambda_n=n^{-1/4}$, and the Epanechnikov kernel. This gives a complete projected nonsmooth model with both Wald and self-normalized inference for the nonscalar filters. *Continued in Part 4/6: Q1 filter realizations; Q2 and Table R3.*
Reviewer_uPH2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies online inference for Polyak–Ruppert averages of stochastic approximation algorithms with fixed stable score filters, which includes scalar momentum, coordinate momentum and higher-order filters. Under local score regularity and trajectory conditions, the authors establish an asymptotic linear representation and show that the leading covariance doesn't depend on the filter. The paper also proposes an online kernel estimator of the Jacobian for quantile regression and establishes an L2L^2 process limit for self-normalized inference. Simulations and a real-data application using the power-plant dataset evaluate estimation error, confidence-interval coverage, and sensitivity to data ordering.

Critical strengths
  1. The paper gives a unified analysis of fixed stable score filters with nonsmooth sample scores, including quantile regression with momentum. The authors separate the local score conditions from the trajectory conditions required for inference. They also establish sufficient conditions for scalar momentum and provide a verified projected quantile example in C.18. These results give the framework concrete applications and show that the analysis does not rely entirely on assumed trajectory properties.

  2. The joint L2L^2 process and endpoint limit directly justifies the proposed self-normalizer without the additional maximal boundary condition required for uniform path convergence. The weak moment assumptions also cover settings with heavy-tailed covariates. For quantile regression, the paper develops an online kernel estimator of the Jacobian while retaining the original nonsmooth score update. Also, its consistency doesn't require ∥θt−θ⋆∥=o(ht)\Vert \theta_t-\theta^\star \Vert = o(h_t), and the resulting confidence intervals have coverage close to those based on the oracle Jacobian in the simulations.

Critical weaknesses
  1. My main concern is the size of the theoretical advance relative to existing online inference results. Tang et al. (2024) already establish inference for averaged scalar momentum in smooth settings, while Lee et al. (2025) establish random scaling inference for averaged quantile SGD. The present paper extends these results to general fixed stable score filters and nonsmooth sample scores. However, for the endpoint CLT, the filter remainder becomes a telescoping boundary term, and Fort (2015) covers an important subclass under stronger smoothness and stability conditions. The online kernel Jacobian estimator also has close precedents in Wei et al. (2026) and Amiri and Thiam (2014), including recursive kernel estimation and the conditional-integration argument. The weaker moment conditions and general-filter formulation are useful, but the general result assumes the trajectory conditions in Assumption 3.3, which are not verified for the coordinate or higher-order filters.

  2. The experiments mainly establish feasibility rather than a clear practical advantage. The pilot selects the tuning parameters using coverage and RMSE computed relative to the true parameter, leaving practical tuning unresolved despite the independent evaluation. The equal-block and smoothed comparators are explicitly diagnostic, so their poor coverage does not establish an advantage over calibrated existing procedures. Interval lengths are also missing. Self-normalized coverage falls to 0.778 and 0.834 in two quantile settings, which does not contradict the asymptotic result but raises concerns about its calibration at the reported sample sizes.

References:

  • Fort (2015). Central limit theorems for stochastic approximation with controlled Markov chain dynamics. ESAIM: Probability and Statistics, 19, 60–80.
  • Wei, Li, Chen, and Wu (2026). Central Limit Theorems for Stochastic Gradient Descent Quantile Estimators. arXiv:2503.02178, June 2026 revision.
  • Amiri and Thiam (2014). A smoothing stochastic algorithm for quantile estimation. Statistics & Probability Letters, 93, 116–125.
Questions for authors
  1. Could the authors clarify which parts of Theorem 3.4 go beyond general averaging results such as Fort (2015) given that the filter remainder telescopes? Which additional settings covered by the theorem are most important for applications?

  2. Beyond the projected scalar momentum example, which nonsmooth models and filters have their trajectory conditions verified from primitive assumptions? In particular, are any of the unprojected quantile configurations in the experiments fully covered by the theory?

  3. Do the reported gains persist when the tuning rule doesn't use the true parameter and the method is compared with a calibrated existing online inference procedure? Reporting both coverage and interval length would help assess its practical performance. Could the authors also explain the substantial undercoverage of the self-normalized intervals in the upper-quantile settings?

9 comments

Thank you for your time and valuable comments.

Note: Because of the 5000-character limit, we divided our response into six comments.

W1. Our contribution is to establish a usable inference framework across finite memory filters and discontinuous estimating equations, while retaining the original score update. The substantive work lies in the following three steps.

  1. Controlling the filtered recursion without sample score differentiability. The quantile indicator is discontinuous, so the analysis uses the population mean expansion and score-continuity bounds. In addition, the filter introduces a state whose weighted boundary and stepsize variation terms must vanish on the inferential scale. Theorem 3.4 combines the exact state transformation with predictable localization, local quadratic mean continuity of the centered score, a mean square trajectory rate, and control of accumulated nonlinear bias. These arguments yield the influence representation for the original average, including a retained averaging window and matrix-valued filters. The sample loss smoothness in Assumption A1 of Tang et al. (2024) excludes the quantile indicator score; their scalar momentum equivalence is an important starting point for this extension.

  2. Making quantile inference feasible without smoothing the parameter update. The indicator score requires a separate estimator of Aτ=Ef(XTθτ∣X)XXTA_\tau=\mathbb E\\{f(X^\mathsf T\theta_\tau\mid X)XX^\mathsf T\\}. Theorem 3.11 controls a covariate weighted matrix kernel estimator evaluated along predictable, filtered iterates. Integrating over the fresh response gives a conditional bias bound of the form

C∥θt−1−θτ∥νA+ChtνA,C\lVert \theta_{t-1}-\theta_\tau\rVert^{\nu_A}+C h_t^{\nu_A},

which avoids a requirement that the parameter error be o(ht)o(h_t). The proof also controls the matrix martingale fluctuations under the weighted pp-moment condition, for some p>1p > 1. Recursive scalar density estimators already appear in Amiri and Thiam (2014) and Wei et al. (2026); the inference method is our contribution here.

  1. Obtaining the joint limit required by the self-normalizer. The path-based denominator requires the endpoint jointly with the L2([0,1])L^2([0,1]) path limit. Theorems 3.7–3.8 establish both jointly and control the integrated filter remainders. This extends Lee et al.'s (2025) random-scaling construction to the present filters without imposing the additional maximal boundary condition used in our uniform path result.

We also address the stability objection with Proposition R1 in Part 1/6 and its proof in Parts 2/6–3/6. Under explicit drift and noise conditions on a compact convex set, it verifies the trajectory assumptions for arbitrary fixed stable filters, including coordinate momentum and two pole cascades. This supplies a primitive verification beyond the submitted scalar results. Parameter convergence uses both filter stability and the stated drift condition.

Proposition R1 (Projected fixed filters). Use the paper's measurable score and i.i.d. observation setup. Let C⊂U\mathcal C\subset U be a known compact convex set with θ⋆∈int⁡(C)\theta_\star\in\operatorname{int}(\mathcal C). Consider

\theta_t=\Pi_{\mathcal C}(\theta_{t-1}-\eta_tPv_t),\qquad \eta_t=\eta_0t^{-\alpha},\quad \eta_0 > 0.$$ The matrices are fixed, $\rho(B) < 1$, and $K=P(I-B)^{-1}C$ is nonsingular. Suppose $\theta_0\in\mathcal C$, $\mathbb E\lVert v_0\rVert^2 < \infty$, and the initialization is independent of the observations. Write $\xi(\theta,Z)=G(\theta,Z)-g(\theta)$. Assume that, for every $e$ with $\theta_\star+e\in\mathcal C$, $$\begin{aligned} \langle e,Kg(\theta_\star+e)\rangle&\ge\mu\lVert e\rVert^2, &\lVert Kg(\theta_\star+e)\rVert&\le L\lVert e\rVert,\\\\ \mathbb E\lVert K\xi(\theta_\star+e,Z)\rVert^2&\le\sigma^2+L_\xi^2\lVert e\rVert^2, \end{aligned}$$ where $\mu > 0$ and all constants are finite. Suppose also that $g(\theta_\star+e)=Ae+r(e)$ on $U$, with nonsingular $A$, and $\lVert r(e)\rVert\le L_r\lVert e\rVert^{1+\nu}$ on $\mathcal C$, where $\nu\in(0,1]$. If $0 < \alpha < 1$ and $\alpha(1+\nu) > 1$, then $$\sup_t\mathbb E\lVert v_t\rVert^2 < \infty,\qquad \mathbb E\lVert \theta_t-\theta_\star\rVert^2=O(\eta_t),\qquad \theta_t\longrightarrow\theta_\star\quad\text{almost surely}.$$ Projection is eventually inactive almost surely, and Assumption 3.3 holds with $\varrho=\alpha$. If Assumption 3.2 also holds, the joint integrated-path and endpoint limit holds; for retained windows satisfying condition (2), the asymptotic linear representation and CLT hold as well. *Continued in Part 2/6: W1 continued: filter-state, projection, and mean-square proofs.*

Part 2/6 of our response to Reviewer uPH2 (continued): W1 continued: filter-state, projection, and mean-square proofs.

Proof. Filter-state bound. Compactness, the growth and noise bounds, and nonsingularity of KK imply sup⁡θ∈CE∥G(θ,Z)∥2<∞\sup_{\theta\in\mathcal C}\mathbb E\lVert G(\theta,Z)\rVert^2 < \infty. Define

W=∑j=0∞(Bj)TBj.W=\sum_{j=0}^{\infty}(B^j)^\mathsf T B^j.

Then W≻0W\succ0 and BTWB=W−IB^\mathsf T WB=W-I. Thus BB contracts in the WW-norm, even when it does not contract in Euclidean norm. Young's inequality and the conditional score bound give, for some fixed q<1q < 1,

\le q\lVert v_{t-1}\rVert_W^2+C_0.$$ Iteration and equivalence of finite-dimensional norms give the uniform second-moment bound. Below, $C_j$ denote finite constants depending on the fixed filter and the stated bounds. **Projection and the transformed recursion.** Let $e_t=\theta_t-\theta_\star$, $D=PB(I-B)^{-1}$, $u_t=e_t-\eta_tDv_t$, and $u_0=e_0-\eta_1Dv_0$. Set $$x_t=\theta_{t-1}-\eta_tPv_t,\qquad q_t=\Pi_{\mathcal C}(x_t)-x_t,\qquad \delta_t=(\eta_{t-1}-\eta_t)Dv_{t-1},$$ with $\eta_0=\eta_1$ in the initial-index convention. The exact identity is $$u_t=a_t+q_t,\qquad a_t=u_{t-1}-\eta_tKG(\theta_{t-1},Z_t)+\delta_t.$$ The projection variational inequality gives $\langle q_t,e_t\rangle\le0$. Consequently, $$\begin{aligned} \lVert u_t\rVert^2 &=\lVert a_t\rVert^2+2\langle u_t,q_t\rangle-\lVert q_t\rVert^2\\\\ &\le\lVert a_t\rVert^2-2\eta_t\langle Dv_t,q_t\rangle-\lVert q_t\rVert^2\\\\ &\le\lVert a_t\rVert^2+\eta_t^2\lVert Dv_t\rVert^2. \end{aligned}$$ This estimate applies to general fixed stable filters, including matrices $B$ without an orthogonal diagonalization. **Mean-square rate.** Since $e_{t-1}=u_{t-1}+\eta_{t-1}Dv_{t-1}$, the drift and growth bounds imply $$\langle u_{t-1},Kg(\theta_{t-1})\rangle \ge\frac\mu4\lVert u_{t-1}\rVert^2 -C_1\eta_{t-1}^2\lVert v_{t-1}\rVert^2.$$ Also $\lVert \delta_t\rVert\le C_2\eta_tt^{-1}\lVert v_{t-1}\rVert$. Expanding $\lVert a_t\rVert^2$ conditionally, using centered fresh noise, and absorbing the cross terms into a small multiple of $\eta_t\lVert u_{t-1}\rVert^2$ yields, for all sufficiently large $t$, $$\mathbb E(\lVert a_t\rVert^2\mid\mathcal F_{t-1}) \le(1-c_0\eta_t)\lVert u_{t-1}\rVert^2 +C_3\eta_t^2+C_4\eta_t^3\lVert v_{t-1}\rVert^2.$$ Here the stepsize-variation cost satisfies \$\\lVert \\delta\_t\\rVert^2/\\eta\_t \\le C\_5\\eta\_t t^{-2}\\lVert v\_{t-1}\\rVert^2 \\le C\_6\\eta\_t^3\\lVert v\_{t-1}\\rVert^2\$, since $\alpha < 1$. The conditional state bound and the projection inequality therefore give $$\mathbb E(\lVert u_t\rVert^2\mid\mathcal F_{t-1}) \le(1-c\eta_t)\lVert u_{t-1}\rVert^2 +C_7\eta_t^2(1+\lVert v_{t-1}\rVert^2).\tag{1}$$ Taking expectations gives $U_t\le(1-c\eta_t)U_{t-1}+C_8\eta_t^2$, where $U_t=\mathbb E\lVert u_t\rVert^2$. Since $\eta_{t-1}/\eta_t=1+O(t^{-1})$ and $t^{-1}=o(\eta_t)$, the deterministic comparison argument gives $U_t=O(\eta_t)$. Then $e_t=u_t+\eta_tDv_t$ gives $\mathbb E\lVert e_t\rVert^2=O(\eta_t)$. *Continued in Part 3/6: W1 proof completion and the W2 experimental setup.*

Part 3/6 of our response to Reviewer uPH2 (continued): W1 proof completion and the W2 experimental setup.

Almost-sure convergence and eventual inactivity. The condition α(1+ν)>1\alpha(1+\nu) > 1, with ν≤1\nu\le1, implies α>1/2\alpha > 1/2. The additive terms in (1), given in Part 2/6, have summable expectations. The nonnegative almost-supermartingale argument gives convergence of ∥ut∥2\lVert u_t\rVert^2 and ∑tηt∥ut−1∥2<∞\sum_t\eta_t\lVert u_{t-1}\rVert^2 < \infty almost surely. Since ∑tηt=∞\sum_t\eta_t=\infty, the limit is zero. Moreover, E∑tηt2∥vt∥2<∞\mathbb E\sum_t\eta_t^2\lVert v_t\rVert^2 < \infty, so ηtvt→0\eta_tv_t\to0 almost surely. It follows that et→0e_t\to0 almost surely. An open ball about θ⋆\theta_\star lies in C\mathcal C. Eventually both θt−1\theta_{t-1} and xtx_t lie in this ball because ηtPvt→0\eta_tPv_t\to0. Thus qt=0q_t=0 eventually almost surely.

Accumulated remainder and inference. Lyapunov's inequality gives

\le\frac{C_9}{\sqrt n} \left(1+\sum_{t=2}^n t^{-\alpha(1+\nu)/2}\right)=o(1).$$ Markov's inequality verifies the accumulated-remainder condition. Since $\alpha > \max\\{0,2\alpha-1\\}$, all trajectory requirements hold with $\varrho=\alpha$. The projection adds a term proportional to $q_t/\eta_t$ in the estimating-equation decomposition. If $T < \infty$ is its last active time, then $$\frac1{\sqrt n}\max_{k\le n} \lVert \sum_{t=1}^k\frac{q_t}{\eta_t}\rVert \le\frac1{\sqrt n}\sum_{t=1}^T\frac{\lVert q_t\rVert}{\eta_t} \longrightarrow0\quad\text{almost surely}.$$ For a retained window the analogous bound uses $\sqrt{N_n}$, which diverges under (2). The remaining averaging and integrated path arguments apply to the actual projected trajectory, as in submitted Proposition C.26. ∎ **W2.** The strongest submitted evidence for the proposed quantile covariance estimator is the *same-estimator, same-window* comparison: kernel and oracle coverage differ by at most $0.012$ in the submitted designs. This supports the feasibility of the Jacobian estimator on the same estimator and averaging window. Table R1 (Part 4/6) compares SGD and normalized momentum using the original nonsmooth quantile update $X_t\\{1(Y_t\le X_t^\mathsf T\theta_{t-1})-\tau\\}$ and selection based only on observed data. The settings use $n=40000$ and an intercept. The first four coefficients are $(0.25,0.55,-0.45,0.30)$ and the rest are zero, with $Y=X^\mathsf T\theta_\tau+\sigma(X)\varepsilon_\tau$. Regressors have AR(1) correlation $0.5$ across coordinates, with independent observations over time: independent standardized $t_5$ innovations are used in the median $t_3$ and lower-exponential settings, and Gaussian innovations in the upper-exponential settings. Errors are $t_3/\sqrt3$ or $U+\log(1-\tau)$ with $U\sim\operatorname{Exp}(1)$, so that the error has zero $\tau$-quantile. The heteroscedastic scale is $0.5+0.5|X_2|$; the homoscedastic scale is one. The median and lower/upper levels are $0.5$ and $0.1/0.9$, respectively. The comparison includes the submitted quantile designs, a homoscedastic upper-tail counterpart and a $d=100$ median counterpart. Both methods start at zero, retain the last $70\\%$ of iterates, and use $\eta_t=\eta_0t^{-\alpha}$. Momentum is normalized with the submitted quantile value $\beta=0.9$. Each method searches $\alpha\in\\{0.55,0.60,0.65\\}$ and $\eta_0\in\\{0.04,0.08,0.16,0.32,0.64\\}$. Twenty independent training/validation pairs rank all 15 candidates by mean validation check loss $\rho_\tau(u)=u\\{\tau-1(u < 0)\\}$ evaluated at the fitted tail average; 50 new pairs select among each method's best three. Each training and validation set has $40000$ observations, shared across candidates and methods within a pair. Ties favor smaller $\eta_0$, then smaller $\alpha$. The selection uses neither the true coefficient nor oracle covariance or coverage. Evaluation uses fresh paired data after selection is frozen. Pilot grid search and validation precede the final single pass estimation and inference run. The Wald intervals use the feasible online Epanechnikov Jacobian with $h_t=t^{-0.3}$ and the observed score covariance. The Jacobian uses all observations, the score covariance uses the retained window, and the standard error is $\sqrt{\widehat\Sigma_{22}/N_n}$ with $N_n=28000$, where $\widehat\Sigma$ is the estimated sandwich matrix. The direct ridge is $\lambda_n=0.04/n$, which vanishes and equals $10^{-6}$ at this horizon. Smoothing is confined to the Jacobian. The reported contrast is the first slope, coordinate 2 after the intercept, rather than the intercept used in the submitted quantile tables. *Continued in Part 4/6: W2 continued: Table R1 and discussion; Q1.*

Comments on the paper

1 comment

Thank you for your time and valuable comments. We have supplied a primitive stability result and proof for projected fixed filters, including normalized coordinate momentum and two-pole filters, and clarified the distinct conditions for estimator limits, feasible covariance estimation, and self-normalized inference. We have added a nonsmooth quantile comparison with separate tuning based only on observed training and validation data, using the same search budget for SGD and momentum. We report coverage together with estimation error and interval length, retain the smooth-score comparison separately, and include common-window and ridge/bandwidth checks. These results and explanations are incorporated into the paper and marked in blue. We hope the detailed responses below address your concerns. Thank you again for your valuable feedback; we welcome further questions and discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.