- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper studies sequential inference for degenerate degree-2 U-statistics, which generally have Gaussian chaos limits that depend on the spectrum of the underlying kernel operator and are therefore difficult to calibrate sequentially. To address this problem, the authors replace the ordinary U-statistic by a row-normalized version that admits a martingale representation on a harmonic variance clock and yields a strong Gaussian approximation at the cost of only logarithmic variance inflation. This gives spectrum-free, asymptotically time-uniform confidence sequences and sequential tests, and extends to fixed-dimensional vector-valued kernels. The benefits of the proposed method are illustrated with toy experiments on sequential kernel two-sample and independence testing, and compared against spectral calibration of the Gaussian chaos limit.
Critical strengths
- Technical elegance. The basic idea of the paper, using row normalization to turn a degenerate U-statistic into a martingale and thereby replacing Gaussian chaos calibration by ordinary Gaussian calibration at only logarithmic increase in variance, is elegant and a strong technical contribution. The benefits of the significantly simpler calibration compared to SAGE, which the paper views as the main baseline, are clear and convincing.
- Rigor and generality. Despite its mathematical rigor, the paper is clearly written. The fact that the argument applies to general symmetric degree-2 U-statistics is nice and the assumptions (i.i.d., finite fourth moments, fixed kernel) do not take away too much from the theory.
Critical weaknesses
- Unclear finite-sample convergence. The guarantees of the proposed procedure are asymptotic. That, by itself, is fair and appropriately signposted. However, the harmonic clock of the Gaussian approximation makes the speed of convergence a particularly salient issue for this paper. Even though the slow clock does not necessarily imply slow convergence of the test, the absence of either a discussion of this issue or an empirical study across increasing monitoring starts currently leaves an important practical question unresolved.
- Experimental scope and significance. The experiments are limited to synthetic toy settings and compare only against the spectrally-calibrated SAGE method of Cai et al. (2026). While the experiments provide illustrative evidence that the proposed martingalized calibration results in higher power than more cumbersome and computationally expensive spectral calibration, it does not establish a broad advantage over competing sequential inference procedures at the application level. This is consistent with how the contributions are framed, but limits the demonstrated significance of the contributions. Similarly, the related work section does a good job at motivating the proposed method against SAGE specifically and surveying alternative sequential testing procedures, but then, rather unsatisfyingly, leaves unclear why one should prefer the proposed method over existing methods beyond SAGE. The paper could do a better job at making clear the significance of its contributions, both in the related work section and through its experiments.
Overall, while the proposed method is elegant, rigorous, and well-presented, I hesitantly lean towards rejection because the paper positions itself in a somewhat narrow niche and currently does not convincingly establish a clear advantage against competing sequential inference procedures at the application level. I therefore suspect that the broader impact of the proposed method may be limited, though I acknowledge that this perception may stem partly from my own background being outside sequential testing specifically. I therefore encourage the authors to use the rebuttal to explicitly highlight the significance and practical value of their contributions.
Questions for authors
- Does your proof yield any quantitative insight into the speed of convergence, even through rough rates under stronger assumptions or in special cases? Is there reason to expect the resulting tests to be systematically conservative or anti-conservative for small monitoring starts ?
- Can you clarify when a practitioner should choose your proposed method over, say, the sequential two-sample test of Shekhar and Ramdas (2024) or the sequential independence test of Podkopaev et al. (2023), and why?
9 comments
Theorem 1 (Scalar GM). Under (A1), take for two-sided inference or for upper-tail testing, and let with the corresponding calibration function defined above. For , put
For , define
If , then
\mathbb{P}\\{\text{an error at some }n\ge m\\}\le\min\left\\{1,\\ p_m^{\rm s}(\delta)+F\\!\left(\sqrt{c^2-\Delta_m^{\rm sGM}(\delta)}\right)\right\\}.
Here the error event is for upper-tail testing, or for two-sided noncoverage, where .
Proof sketch. Use the common scalar preparation above. Write . The centered-scale identity in the proof of manuscript Theorem 1, , gives
On , the bounds defining and show that this implies
For upper-tail rejection, gives the same inclusion without absolute values. The manuscript’s Brownian calibration results (Lemmas 3 and 4) bound this crossing probability by . Adding proves the claim.
Theorem 2 (Scalar LIL). Under (A1), fix , and . Let for upper-tail testing and for two-sided inference, and write and .
For , put
\begin{aligned}\Delta_m^{\rm sLIL}(\delta)={}&s\ell_\eta\log\frac{1+\delta}{1-\delta}\\\\&+\frac{\delta}{4(1-\delta)}\left\\{\log\frac{J\zeta(s)}{\alpha}+s\log\frac{\max\\{1,\log\eta\\}}{\log\eta}+\frac{s\ell_\eta}{e}\right\\}\\\\&+\frac{c_\eta}{m-1}\left\\{\log\frac{J\zeta(s)}{\alpha}+s\log\frac{\max\\{1,\log\eta\\}}{\log\eta}+\frac{2s\ell_\eta}{e}\right\\}^2.\end{aligned}
Then the original scalar LIL procedure satisfies
\mathbb{P}\\{\text{an error at some }n\ge m\\}\le\min\left\\{1,\\ p_m^{\rm s}(\delta)+\alpha e^{\Delta_m^{\rm sLIL}(\delta)}\right\\}.
Proof sketch. Use the common scalar preparation above. Write the squared LIL boundary as , where
For and ,
With , the same centered-scale identity makes the squared error condition
Set . On , the preceding bounds and imply
For , rejection also requires . The manuscript’s Brownian calibration lemmas (Lemmas 3 and 4) bound this one- or two-sided crossing probability by when . Adding proves the claim. And if , the bound truncated at one is immediate.
Theorem 3 (Vector GM and LIL). Under (A2), fix and . For VEC-GM, define
For VEC-LIL, retain the original parameters , , and , and define
\begin{aligned}\Delta_m^{\rm VLIL}(\delta)={}&q\log\left(1+\frac{\log\\{(1+\delta)/(1-\delta)\\}}{\log a}\right)+\frac{d+1}{2}\log\frac{1+\delta}{1-\delta}\\\\&+\frac{\delta}{3(1-\delta)}\left\\{\log\frac{\zeta(q)\beta}{\alpha(1-\beta^{-1})}+d\log\frac{3\beta}{\varepsilon}+\frac{q}{e\log a}\right\\}\\\\&+\frac{2a}{(1-\varepsilon)^2(m-1)}\left\\{\log\frac{\zeta(q)\beta}{\alpha(1-\beta^{-1})}+d\log\frac{3\beta}{\varepsilon}+\frac{2q}{e\log a}\right\\}^2.\end{aligned}
For either procedure, use the corresponding . Then
The original centered covariance estimator is used, with whenever that estimator is singular.
Proof sketch. As in the scalar case, multiplication by preserves noncoverage and reduces the covariance to , with its original dependence retained in . For the normalized rows , let and define
where is the predictable quadratic variation. Every scalar projection is square-integrable, so Howard et al. (2020, Lemma 3(f)) gives the nonnegative supermartingales
\exp\left\\{\boldsymbol\lambda^{\mathsf T}\boldsymbol M_n-\tfrac12\boldsymbol\lambda^{\mathsf T}\boldsymbol P_n\boldsymbol\lambda\right\\},\qquad\boldsymbol\lambda\in\mathbb{R}^d,
starting at one. This verifies the conditions for the vector boundary results.
Define the common event
\mathcal E_m^{\rm v}=\left\\{\sup_{n\ge m}\\|\langle\boldsymbol M\rangle_n/T_n-I_d\\|_F\le\delta/2,\\ \sup_{n\ge m}\\|\boldsymbol Q_n-\langle\boldsymbol M\rangle_n\\|_F\le\delta T_m/2\right\\},
where is the identity matrix and is the Frobenius norm. A quantitative refinement of the covariance-clock argument in manuscript Lemma 5 gives . On , simultaneously for ,
where is the operator norm. Apply the Gaussian-mixture bound of Howard et al. (2021, Section 3.2) for GM and the self-normalized bound of Whitehouse et al. (2026, Theorem 4.1) for LIL, as in manuscript Lemma 7. Use Gaussian regularization and a LIL variance floor of , with LIL epoch weights , .
As in the scalar proofs, the remaining step transfers these boundaries to the original centered estimator. In the normalized coordinates, manuscript Lemma 5 gives
On , apply Sherman–Morrison and the displayed covariance bounds. Using and the same logarithm estimates bounds the covariance perturbation and centering losses by the corresponding . Hence noncoverage on implies crossing the cited boundary at level , which bounds its probability when below one. Adding and truncating at one proves the result.
For each of the four procedures, a known upper bound on the corresponding moment or gives a sufficient large start to choose such that the applicable explicit error bound is at most , for a prescribed tolerance . We will include the three theorems and their complete proofs in the appendix to the revision.
We are grateful to the reviewer for their time and valuable questions and suggestions.
Q1. We now strengthen the original asymptotic error-control results with explicit infinite-horizon error bounds for both scalar and vector procedures, under the same canonical-kernel and fourth-moment assumptions.
For the centered kernels h0=h−θ and h0=h−θ, let τ2=Eh02 and Ω=E(h0h0T). Write Tn=∑j=1n−1j−1, K=E∣h0(X1,X2)∣4/τ4 and Kd=E∣Ω−1/2h0(X1,X2)∣24. For a deterministic proof tolerance 0<δ≤1/2, define
pms(δ)=δ2Tm1408K+δ2Tm280K,pmv(δ)=δ2Tm1408Kd+δ2Tm212Kd.
Table 1 summarizes the bounds and each deterministic correction Δm is defined in the corresponding theorem below.
Table 1. Bounds on the probability of any error at n≥m. Each right-hand side is truncated at one.
For scalar GM, c=F−1(α) and the displayed bound requires ΔmsGM(δ)<c2. Here F(c)=21−Φ(c)+cϕ(c) for two-sided inference and F(c)=1−Φ(c)+cϕ(c)+ϕ(c)2/Φ(c) for upper-tail inference, where Φ and ϕ are the standard normal distribution and density. For every row, Δm(δ)=O(δ+(m−1)−1) with dimension, nominal level and boundary parameters fixed. Balancing this correction with K\*/(δ2Tm) by taking δ≍(K\*/Tm)1/3 gives the common rate
Pan error at some n≥m≤α+O((K\*/Tm)1/3),K\*=KorKd.
For each fixed admissible law this is O((logm)−1/3) excess error. These quantitative bounds directly imply the original scalar and vector asymptotic time-uniform coverage guarantees as m→∞.
We state the three theorems below using the notation of the manuscript. Due to space constraints, each theorem is followed by a proof sketch. Assumption (A1) requires E∣h0∣4<∞ and τ2>0; (A2) requires E∣h0∣24<∞ and Ω≻0. For scalar inference, the centered rows are Yi=(i−1)−1∑j<ih0(Xi,Xj), Mn=∑i=2nYi, and τn is the original centered row-scale estimate. The sets Cm,nell(α) are the original confidence ellipsoids for θ=Eh(X1,X2).
Common scalar proof ingredients. Since τ>0, replacing h0 by h0/τ scales both Mn and τn by 1/τ, preserving all crossing events. K=E∣h0∣4/τ4 refers to the original kernel, and the procedure still estimates the unknown scale. Write [M]n=∑i=2nYi2. The manuscript’s original Skorokhod embedding (Lemma 1) gives Mn=B(Rn), where B is standard Brownian motion and Rn are increasing stopping times with R1=0. Its moment condition permits Itô’s formula and isometry, which yield
supn≥2ERn−[M]n2=32∑i=2∞EYi4≤2K.
Combining this with the explicit clock estimates in the proof of manuscript Theorem 1 and the maximal inequality gives P((Ems)c)≤pms(δ), where Ems=⋂n≥m∣Rn/Tn−1∣≤δ,∣Rn−[M]n∣≤δTm/4. Both scalar proofs use this event and the existing Brownian boundary calibrations (manuscript Lemmas 3 and 4), comparing boundaries at Rn directly.