acceptodds
Under review as a conference paper at ICLR 2027

Diffusion-Scale Martingale MMD Tests on Compact Riemannian Manifolds

Abstract

Two-sample testing on a compact Riemannian manifold requires a kernel that respects geometry and remains sensitive to differences at unknown spatial scales. Diffusion kernels provide an intrinsic family, but smoothing attenuates both signal and noise, and neighboring scales can give redundant tests. We study this tradeoff within martingale maximum mean discrepancy testing. An exact spectral criterion shows that additional diffusion improves local asymptotic efficiency precisely when the average null-noise frequency exceeds the average signal frequency. For a fixed grid, thresholding the covariance pseudoinverse accommodates singularity and yields an asymptotic chi-square null limit with consistently estimated positive rank. Simulations on the circle, torus, and sphere support the predicted frequency dependence and conservative-to-near-nominal size. A matched torus comparison shows how the resampling-free calculation can improve detection by processing more available observations in similar time. These results connect intrinsic scale choice to both statistical power and computational cost.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Kernel method, Martingale, MMD test, Manifold data

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_3anfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies two-sample testing on compact Riemannian manifolds using diffusion kernels within the martingale MMD framework. It derives a local-efficiency identity comparing diffusion-weighted signal and null-noise frequencies, and extends multi kernel inference to singular covariance through a thresholded pseudo inverse and estimated rank. It provides fixed-grid null and local-alternative limits, an exact finite-feature implementation, and experiments on the circle, torus, sphere, and archived pigeon bearings. The empirical analysis separates equal-sample power from the benefit of processing more observations under a similar computation budget.

Critical strengths
  1. The spectral efficiency calculation gives a clear interpretation of smoothing: additional diffusion helps when it removes null variation faster than signal. The treatment includes covariance between modes under nonuniform sampling, rather than relying only on an orthogonal uniform-null example.
  2. The singular-covariance extension is carefully formulated. The null theorem separates covariance-estimation error from the rank threshold, and the local theorem states the additional covariance-convergence and threshold conditions. These conditions matter when some contrast directions vanish or are nearly redundant.
  3. The experiments are unusually transparent about tradeoffs. They include duplicate and nearby scales, finite-support nulls, truncation controls, permutation calibration of the same rank statistic, and both equal-n and similar-time FUSE comparisons. The paper explicitly reports cases where FUSE or cross-witness testing is more powerful.
  4. The application is accompanied by relevant checks. The analysis documents independent-of-bearing subsampling, multiplicity correction, conservative small-sample calibration, and sensitivity to ordering/subsampling. It does not interpret the archive comparison causally.
Critical weaknesses
  1. The methodological advance is narrower than a new general testing framework. Chatterjee and Ramdas (2026, Section 7.1, Theorem 9) already provide multi kernel mMMD with a chi-square reference using covariance inversion. Applying mMMD to a diffusion-kernel family and computing it through finite features are natural specializations, as the paper itself notes in Proposition D.1. The main additions are therefore the frequency-based interpretation and the treatment of singular directions. Both are useful, but the efficiency identity is conditional on an unknown alternative direction and does not provide an adaptive scale-selection or optimality guarantee. The current fixed-grid, known-spectrum setting also limits the breadth of the result.
  2. The rank-recovery theorem and practical spectral truncation operate in substantially different regimes. Table 4 reports median retained rank two for both the exact-duplicate and nearby-time circle grids. Under a uniform circle law with the stated 100-mode truncation, these grids have population ranks five and six, respectively. Using the paper’s normalization, the population covariance for times a and b is 8∑k=150exp[−(a+b)k2]8∑_{k=1}^{50}exp[−(a+b)k²]; at n=500n=500, the default threshold is approximately 0.633, and only two population eigenvalues exceed it in either grid. Thus these experiments demonstrate aggressive finite-sample truncation, not recovery of the population rank. The paper acknowledges this distinction, but the practical connection between its asymptotic theorem and the calibrated test remains incomplete. The lost signal along discarded directions, sensitivity near the threshold, and the recommendation to use different exponents in local experiments deserve a more direct explanation.
  3. The computation–power advantage is established against a limited set of resource choices. The larger-sample comparison fixes FUSE at 199 permutations; it does not include a similarly efficient finite-feature cross-witness or other inexpensive Gaussian-calibrated baseline under the same budget. Such a comparison is pertinent because Table 5 already shows competitive cross-witness power in some settings. At equal n=5000, FUSE has higher high-frequency power, 0.767 versus 0.613; the similar-time advantage arises when RA processes 5000 pairs and FUSE processes 2500. This is a valid and clearly disclosed comparison, but it establishes an advantage for those particular procedures and budgets. A small extension of the existing comparison would clarify whether rank aggregation itself offers a compelling computational choice. Anirban Chatterjee and Aaditya Ramdas. A Martingale Kernel Two-Sample Test. ALT 2026,PMLR 313,pp. 1–44。​
Questions for authors
  1. Can you report the population/empirical eigenvalue spectra relative to the threshold for the rank-stress examples, and relate discarded directions to power? What single prespecified cutoff would you recommend when it is unknown whether an alternative is local?
  2. Could the existing larger-sample benchmark include a finite-feature cross-witness baseline on the same predetermined scales, with a valid prespecified aggregation rule, and a modest FUSE permutation-budget sensitivity check? This would directly test the claimed computational use case without introducing a new application.
  3. How should a practitioner interpret the application result given that the Holm-adjusted rejection fractions over order/subsample reruns are 0.610 and 0.548 for the two olfactory contrasts? In particular, what guidance can be given for a single prespecified analysis without selecting a favorable ordering?
4 comments

Thank you for your time and valuable comments.

W1. We identify the contributions as the diffusion-frequency efficiency identity and singular covariance inference. Chatterjee and Ramdas provide the martingale and full rank multikernel construction; our row-average convention is asymptotically equivalent to theirs. Our efficiency identity compares signal and null attenuation, including nonuniform mode covariance, and the null and local-alternative limits handle covariance estimation, thresholding and positive-rank recovery. Even two characteristic kernels kk and 2k2k have collinear contrasts. Rank adaptation handles such redundant covariance directions; diffusion times are specified independently of the testing observations. For an unknown direction we aggregate the predetermined grid. An independent pilot may choose among a fixed finite menu while preserving the test-sample level; the efficiency identity describes attenuation for a specified alternative, with optimal scale selection remaining a separate problem. A fixed bounded numerical kernel can supply the represented contrasts when an analytic spectrum is unavailable, subject to its own covariance and rank conditions.

W2. We distinguish population rank from the directions retained at a finite sample size. For the normalized-volume circle basis with 50 sine/cosine pairs, direct expansion gives

Γab=8∑k=150e−(a+b)k2.\Gamma_{ab}=8\sum_{k=1}^{50}e^{-(a+b)k^2}.

The duplicate grid is (0.025,0.05,0.10,0.25,0.25,1)(0.025,0.05,0.10,0.25,0.25,1); the nearby grid replaces the second 0.25 by 0.251. An additional 500-replication circle experiment uses common observations across the two grids and both cutoffs, independently of the submitted rank-stress experiment. The null is uniform and the alternative has density 1+0.3cos⁡(7θ)1+0.3\cos(7\theta) relative to normalized volume. At n=500n=500:

EigenvalueDuplicate populationEmpirical medianNearby populationEmpirical median
165.256067.40565.234967.385
25.297005.19865.297535.2003
30.493520.456310.494230.45682
40.064610.0553850.064500.055289
50.007120.00600610.007130.0060189
6002.7622,10−82.7622\\,10^{-8}2.3419,10−82.3419\\,10^{-8}

Population ranks are five and six. The historical cutoff 0.63335 and prospective cutoff 0.79584 both retain two population directions. The third empirical eigenvalue exceeds the historical cutoff in 0.114 of the null runs for each grid. For a fixed Gaussian signal vector bb and covariance eigenpairs (λj,uj)(\lambda_j,u_j), discarding positive directions loses noncentrality

∑0<λj≤κ(ujTb)2λj.\sum_{0 < \lambda_j\le\kappa}\frac{(u_j^Tb)^2}{\lambda_j}.

For frequency one, three and seven, the fractions of noncentrality lost are 0.178/0.656/0.797 on the duplicate grid and 0.177/0.723/0.825 on the nearby grid. These quantify the Gaussian projection mechanism. The actual frequency seven rejection rates at n=500n=500 are 0.118 with the historical cutoff and 0.104 with the prospective cutoff; their null rates are 0.026 and 0.024. The two grids have the same displayed rejection rates in this paired batch; the submission's independently seeded rank-stress batch gave 0.034 and 0.040 at exponent 1/41/4. Thus finite sample truncation trades weak-direction retention for stability, while the fixed-law theorem gives eventual rank recovery as threshold and estimation error vanish, for a fixed positive spectrum.

Continued in the next part

W3. We compared RA, MMD-FUSE and a finite-feature cross-witness grid on exactly the same observations, with n=5000n=5000 observations per group for every method. The first distribution is uniform on the torus; the second has density 1+Acos⁡(2πmx1)1+A\cos(2\pi m x_1). All methods use the same 400 nonconstant normalized-volume features and six predetermined diffusion times. Frequencies 9 and 11 extend the original frequency-seven comparison within the retained feature basis. The displayed results use an independent new batch of 500 replications per setting after an exploratory frequency comparison, separate from the preceding 300-repetition frequency-seven study. RA uses the recommended cutoff (log⁡n)−1/8(\log n)^{-1/8}, with the historical exponent 1/41/4 shown separately. Cross-witness fits each scale on one half, studentizes on the independent half and aggregates by pgrid=min⁡(1,6min⁡ℓpℓ)p_{\mathrm{grid}}=\min(1,6\min_\ell p_\ell); FUSE uses 19, 49 or 199 nested permutations. Fixed bounded features, independent halves and Bonferroni give asymptotic level for the declared cross-witness aggregation.

MethodNull rejectionPower, (m,A)=(7,0.3)(m,A)=(7,0.3)Power, (9,0.3)(9,0.3)Power, (11,0.3)(11,0.3)Power, (9,0.6)(9,0.6)Power, (11,0.6)(11,0.6)
RA, cutoff exponent 1/81/80.0440.5900.2240.0700.9980.578
RA, cutoff exponent 1/41/40.0540.5980.3200.0980.9980.786
FUSE, 19 permutations0.0400.6520.2040.0580.9880.396
FUSE, 49 permutations0.0400.6720.1840.0520.9880.410
FUSE, 199 permutations0.0560.7500.2200.0780.9980.490
Cross-witness grid0.0160.3240.0840.0420.8860.176

At frequency 11 and amplitude 0.6, the recommended RA test has power 0.578 versus 0.396/0.410/0.490 for FUSE and 0.176 for cross-witness, gains of 18.2/16.8/8.8 and 40.2 percentage points, respectively. At the original frequency seven and amplitude 0.3, FUSE has higher power (0.652/0.672/0.750 versus 0.590); at frequency nine and amplitude 0.3, RA and 199-permutation FUSE are close (0.224 versus 0.220). At frequency 11 and amplitude 0.3 both methods have low power (0.070 for RA and 0.078 for 199-permutation FUSE); at frequency nine and amplitude 0.6 both reach 0.998. The advantage is clearest at frequency 11 and amplitude 0.6, alongside the frequency-seven setting favoring FUSE. The historical exponent gives power 0.786 at frequency 11 and amplitude 0.6, with null rejection 0.054; the recommended exponent remains the primary comparison.

Median complete times at the same n=5000n=5000 are 0.1717 seconds for RA (including both cutoff calculations), 0.2611/0.2750/0.3491 for FUSE and 0.1145 for cross-witness. These streaming-implementation timings on uniform-null streams include feature evaluation, normalization, calibration and the final p-value, on one BLAS thread after warmup, excluding data generation. They compare methods within this benchmark; the submission reports its original implementation timings. RA uses 34.3%/37.6%/50.8% less time than FUSE with 19/49/199 permutations, respectively. Cross-witness is faster, with lower power in the displayed alternatives. Reducing FUSE's permutation budget leaves its pooled feature Gram normalization cost in place. The displayed power comparison fixes the sample size for all methods, and runtime is measured separately.

Q1. For a new analysis we specify κn=(log⁡n)−1/8\kappa_n=(\log n)^{-1/8} with the normalized-volume kernels. It satisfies both κnlog⁡n→∞\kappa_n\sqrt{\log n}\to\infty and κn(log⁡n)1/4→∞\kappa_n(\log n)^{1/4}\to\infty. The local result additionally assumes ∣Γn−Γ∣op=o(κn)\\|\Gamma_n-\Gamma\\|_{\mathrm{op}}=o(\kappa_n). Multiplying every kernel by cc multiplies covariance eigenvalues by c2c^2, so the cutoff is multiplied by c2c^2 to preserve the retained directions. Normalization, threshold and retained rank are therefore specified together. The submission already uses exponent 1/81/8 in the local spectral experiments and 1/41/4 in the non-local experiments. The latter satisfies the null condition but lies on the boundary of the stronger local condition; 1/81/8 satisfies both.

At n=500n=500, the population spectra and empirical medians in W2 place the third direction close to the cutoff. Changing the cutoff from 0.63335 to 0.79584 changes null rejection 0.026 to 0.024 and frequency-seven power 0.118 to 0.104 in the additional paired circle experiment (500 replications, amplitude 0.3). The projection formula in W2 quantifies why a small covariance eigenvalue may still carry standardized signal. For small-sample calibration under exchangeability, permutations recompute covariance, retained rank and the entire statistic.

Continued in the next part

Q2. The matched comparison in W3 includes the requested cross-witness baseline with a fixed six-scale Bonferroni aggregate and FUSE budgets 19, 49 and 199. Every method receives the same n=5000n=5000 observations per group. At frequency 11 and amplitude 0.6, recommended-cutoff RA reaches power 0.578 versus 0.176 for cross-witness and 0.396/0.410/0.490 for FUSE; at frequency seven and amplitude 0.3, FUSE reaches 0.652/0.672/0.750 versus 0.590 for RA. Median complete times at the same n=5000n=5000 are 0.1717 seconds for RA (including both cutoff calculations), 0.2611/0.2750/0.3491 for FUSE and 0.1145 for cross-witness. These streaming-implementation timings on uniform-null streams include feature evaluation, normalization, calibration and the final p-value, on one BLAS thread after warmup, excluding data generation. They compare methods within this benchmark; the submission reports its original implementation timings. RA uses 34.3%/37.6%/50.8% less time than FUSE with 19/49/199 permutations, respectively. Cross-witness is faster, with lower power in the displayed alternatives. Reducing FUSE's permutation budget leaves its pooled feature Gram normalization cost in place. This comparison holds the observations and sample size fixed across methods.

Q3. The 0.610 and 0.548 fractions measure sensitivity to ordering and subsampling of the same birds. The primary analysis uses bearing-blind seeded record selection and a fixed order. The 0.610/0.548 rejection frequencies show that the decision depends on the ordering and subsample. For a single analysis, prespecify the bearing-blind record-selection seed, pairing/order, normalization, cutoff, reference distribution and Holm family. The primary asymptotic Holm p-values are 0.0181 and 0.0989 for the two olfactory contrasts. For the same RA statistic, 9999 pooled permutations recomputing covariance and rank give Holm p-values 0.0039 and 0.0222. The latter calculation changes the calibration of the same observed statistic. These results present both the primary large-sample analysis and its calibration sensitivity.

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

Reviewer_3anfOpenReview ↗

I will maintain my current score

Thank you for the detailed responses and additional experiments. These largely address my main concerns, and I will maintain my current score.

Reviewer_dMFYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

For two-sample test on a compact Riemannian manifold, one can use diffusion kernels within martingale maximum mean discrepancy testing. The paper studies how do the different choices, notably the diffusion time, influences the resulting properties of the test. The theoretical results are supported by numerical experiments on the circle, torus and sphere along with an application for homing-pigeon bearings.

Critical strengths

The paper addresses a meaningful problem in MMD test on Riemannian manifolds. The empirical results verify the effectiveness of the proposed theoretical results.

Critical weaknesses
  1. The paper is mathematically heavy, and the presentation can be difficult to follow for readers without substantial background on MMD testing. While the geometric settings are described relatively well, the details and motivations of the considered martingale MMD and the diffusion kernels are less covered in the current manuscript. A more self-contained introduction to these topics would benefit the readers, and also make it clearer what were already known and what are the paper's novel contributions.

  2. The paper, notably the experiments, focus on manifolds with analytical Laplace-Beltrami eigenbases. It might be useful to clarify the theoretical and empirical implications when such eigenbases are not analytically available.

Questions for authors
  1. Theorem 3.3 discusses how diffusion time influence the efficiency. Can it be used to construct a practical method to choose the diffusion times for improved efficiencies?

  2. How would the proposed method behave for manifolds with intractable Laplace-Beltrami eigenbases? Notably, would the properties of the resulting statistical test change?

2 comments

Thank you for your time and valuable comments.

W1. We connect the construction explicitly to its statistical purpose. MMD compares mean features, ∣EPϕ(X)−EQϕ(Y)∣2\\|E_P\phi(X)-E_Q\phi(Y)\\|^2. For a fresh pair Zi=(Xi,Yi)Z_i=(X_i,Y_i), the increment

Di=1i−1∑j<ik(Xi,Xj)+k(Yi,Yj)−k(Xi,Yj)−k(Yi,Xj)D_i=\frac1{i-1}\sum_{j < i}\\{k(X_i,X_j)+k(Y_i,Y_j)-k(X_i,Y_j)-k(Y_i,X_j)\\}

evaluates a past data witness on that pair. Under the null the fresh observations have the same law, giving conditional mean zero; accumulated increments measure discrepancy and quadratic variation measures noise. Diffusion assigns mode jj weight e−aλje^{-a\lambda_j}, attenuating fine scale signal and noise together. The efficiency identity compares these rates. We credit the martingale/full rank construction to Chatterjee and Ramdas, and identify our additions as the diffusion-frequency interpretation and singular covariance inference under the stated threshold conditions.

W2. A fixed bounded positive semidefinite numerical kernel, or a feature map fitted on independent data, can replace an analytic eigenbasis. Conditional on that map, inference uses its contrast covariance and requires the stated covariance consistency and positive rank conditions. A finite map detects the differences represented by its features. Recovering the exact heat-kernel efficiency interpretation additionally requires control of spectral approximation and truncation errors.

Direct kernel evaluation costs O(Ln2+nL2)O(Ln^2+nL^2) arithmetic, compared with O(nLJ+nL2)O(nLJ+nL^2) once a JJ-feature representation is available; constructing a numerical spectrum adds its own cost. The present numerical studies use analytic eigenfunctions; a numerical-eigenbasis implementation has not been evaluated. These distinctions identify what changes computationally and statistically when an eigenbasis is intractable: a numerical PSD representation supplies a test for its represented contrasts, with calibration under its own covariance and rank assumptions.

Q1. For an unknown direction, our operational prescription jointly aggregates the predetermined grid

aλ1∈0.025,0.05,0.10,0.25,0.50,1.a\lambda_1\in\\{0.025,0.05,0.10,0.25,0.50,1\\}.

Theorem 3.3 gives the efficiency criterion for comparing scales under a specified signal and null spectrum. Pilot estimates of these quantities can rank members of a fixed menu; the theorem itself supplies a spectral comparison rather than an optimal data-adaptive selection rule. When independent pilot information is available, it may select among a fixed finite menu. If member jj has test-sample null rejection probability pj,n→αp_{j,n}\to\alpha and the pilot selects JnJ_n independently, then

\le\max_j|p_{j,n}-\alpha|\to0.$$ For conservative members the same argument gives the corresponding upper bound. This supplies asymptotic level protection for independent-pilot selection. The cutoff uses the testing-sample size; power depends on the selected scales and the alternative's retained frequencies. **Q2.** The bounded numerical-kernel interpretation in W2 supplies the practical alternative to an analytic eigenbasis. Holding the map fixed, or learning it independently, lets the testing calculation condition on that representation. The method then estimates the represented contrast covariance and retains its sufficiently large eigen-directions. The multiscale aggregation over the predetermined grid uses this same declared representation and cutoff. To transfer the exact heat-kernel efficiency formula, the numerical kernel must also approximate the heat kernel at the rates required by the corresponding asymptotic result. Thank you again for your valuable feedback. We hope our responses address your concerns and welcome any further questions or discussion.

I thank the authors for their response. I would like to maintain my original positive rating.

Reviewer_r4vFOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies how to detect differences between two groups of observations on a known geometric space, such as a circle or sphere. It uses diffusion kernels to compare the groups at different smoothing scales. Building on an existing MMD test, it explains when smoothing helps detection and how to combine scales that contain overlapping information. It computes p-values using a large-sample approximation, avoiding repeated random shuffling of group labels.

Experiments use simulated data on the circle, torus, and sphere, plus recorded pigeon flight directions. MMD-FUSE often detects differences better at the same sample size, but the proposed method is faster and can sometimes achieve better detection by processing more observations within a similar time budget.

Critical strengths
  1. The theory explains when smoothing helps and how to combine scales. Theorem 3.3 relates detection ability to how quickly smoothing removes signal and noise. Theorems 3.4–3.5 justify combining scales even when some provide redundant information. These results clarify the method's behavior and extend the existing martingale MMD framework.

  2. The experiments report both advantages and limitations. They check false-positive rates, detection ability, and runtime under several settings. MMD-FUSE often performs better with the same number of observations, while the proposed method can sometimes perform better within a similar time budget by processing more data.

Critical weaknesses
  1. The practical advantage needs broader comparisons. The main runtime comparison uses MMD-FUSE with 199 permutations. It is unclear whether the advantage persists with fewer permutations. Also, the theory explains which smoothing scales help, but does not directly tell a practitioner which scale to choose when the difference between groups is unknown.

  2. Results are sensitive to the threshold, kernel scaling, and, in the pigeon example, subsampling and observation order. Clearer guidance is needed for reliable use with small samples.

Questions for authors
  1. Does the runtime advantage persist with fewer MMD-FUSE permutations?

  2. How should users choose the threshold for discarding redundant directions, and how sensitive are false-positive rates and detection ability to that choice?

  3. For the pigeon data, do conclusions change if p-values are computed using permutations instead of the large-sample approximation, keeping the statistic, observations, and order fixed?

4 comments

Thank you for your time and valuable comments.

W1. We compared RA, MMD-FUSE and a finite-feature cross-witness grid on exactly the same observations, with n=5000n=5000 observations per group for every method. The first distribution is uniform on the torus; the second has density 1+Acos⁡(2πmx1)1+A\cos(2\pi m x_1). All methods use the same 400 nonconstant normalized-volume features and six predetermined diffusion times. Frequencies 9 and 11 extend the original frequency-seven comparison within the retained feature basis. The displayed results use an independent new batch of 500 replications per setting after an exploratory frequency comparison, separate from the preceding 300-repetition frequency-seven study. RA uses the recommended cutoff (log⁡n)−1/8(\log n)^{-1/8}, with the historical exponent 1/41/4 shown separately. Cross-witness fits each scale on one half, studentizes on the independent half and aggregates by pgrid=min⁡(1,6min⁡ℓpℓ)p_{\mathrm{grid}}=\min(1,6\min_\ell p_\ell); FUSE uses 19, 49 or 199 nested permutations. Fixed bounded features, independent halves and Bonferroni give asymptotic level for the declared cross-witness aggregation.

MethodNull rejectionPower, (m,A)=(7,0.3)(m,A)=(7,0.3)Power, (9,0.3)(9,0.3)Power, (11,0.3)(11,0.3)Power, (9,0.6)(9,0.6)Power, (11,0.6)(11,0.6)
RA, cutoff exponent 1/81/80.0440.5900.2240.0700.9980.578
RA, cutoff exponent 1/41/40.0540.5980.3200.0980.9980.786
FUSE, 19 permutations0.0400.6520.2040.0580.9880.396
FUSE, 49 permutations0.0400.6720.1840.0520.9880.410
FUSE, 199 permutations0.0560.7500.2200.0780.9980.490
Cross-witness grid0.0160.3240.0840.0420.8860.176

At frequency 11 and amplitude 0.6, the recommended RA test has power 0.578 versus 0.396/0.410/0.490 for FUSE and 0.176 for cross-witness, gains of 18.2/16.8/8.8 and 40.2 percentage points, respectively. At the original frequency seven and amplitude 0.3, FUSE has higher power (0.652/0.672/0.750 versus 0.590); at frequency nine and amplitude 0.3, RA and 199-permutation FUSE are close (0.224 versus 0.220). At frequency 11 and amplitude 0.3 both methods have low power (0.070 for RA and 0.078 for 199-permutation FUSE); at frequency nine and amplitude 0.6 both reach 0.998. The advantage is clearest at frequency 11 and amplitude 0.6, alongside the frequency-seven setting favoring FUSE. The historical exponent gives power 0.786 at frequency 11 and amplitude 0.6, with null rejection 0.054; the recommended exponent remains the primary comparison.

Median complete times at the same n=5000n=5000 are 0.1717 seconds for RA (including both cutoff calculations), 0.2611/0.2750/0.3491 for FUSE and 0.1145 for cross-witness. These streaming-implementation timings on uniform-null streams include feature evaluation, normalization, calibration and the final p-value, on one BLAS thread after warmup, excluding data generation. They compare methods within this benchmark; the submission reports its original implementation timings. RA uses 34.3%/37.6%/50.8% less time than FUSE with 19/49/199 permutations, respectively. Cross-witness is faster, with lower power in the displayed alternatives. Reducing FUSE's permutation budget leaves its pooled feature Gram normalization cost in place.

Continued in next part

W2. Our theoretical results allow a range of thresholds rather than requiring a particular exponent. Writing κn=(log⁡n)−β\kappa_n=(\log n)^{-\beta}, the null asymptotic result requires κnlog⁡n→∞\kappa_n\sqrt{\log n}\to\infty, which holds for 0<β<1/20 < \beta < 1/2. The stronger rate condition for the local-alternative result, κn(log⁡n)1/4→∞\kappa_n(\log n)^{1/4}\to\infty, holds for 0<β<1/40 < \beta < 1/4, subject to the additional covariance convergence assumption ∣Γn−Γ∣op=o(κn)\\|\Gamma_n-\Gamma\\|_{\mathrm{op}}=o(\kappa_n).

The submitted local spectral experiments use exponent 1/81/8, while the non-local cells use 1/41/4. In an additional 500-replication circle experiment, independent of the submission batch, we compare exponent β=1/4\beta=1/4 with β=1/8\beta=1/8 in the circle experiments containing duplicate and nearly duplicate diffusion times. The grids and cutoffs use common observations, with a uniform null and alternative density 1+0.3cos⁡(7θ)1+0.3\cos(7\theta). At n=500n=500, changing the exponent from 1/41/4 to 1/81/8 increases the threshold from 0.633350.63335 to 0.795840.79584. For both grids, the null rejection rate changes from 0.0260.026 to 0.0240.024, while power against the frequency-seven alternative changes from 0.1180.118 to 0.1040.104. The population covariance ranks are five and six for the duplicate and nearby grids, respectively, whereas both thresholds retain only two population directions. The third population eigenvalues are approximately 0.4940.494, below both thresholds. Consequently, the two choices induce the same population-level truncation in these examples. The discarded directions account for approximately 79.779.7\\% and 82.582.5\\% of the Gaussian noncentrality for the frequency-seven alternative on the duplicate and nearby grids, respectively. These calculations make the potential loss of signal from finite-sample truncation explicit. Kernel scaling and thresholding are specified together: multiplying the kernels by cc requires multiplying the covariance cutoff by c2c^2. For the pigeon analysis, fix the bearing-blind subsampling seed and pairing/order before testing; the 0.610/0.548 rejection frequencies across alternative orders and subsamples quantify sensitivity to those choices. At small sample sizes, pooled permutations recalibrate the entire statistic, including covariance and retained rank, under exchangeability.

Q1. At equal sample size, RA takes 0.1717 seconds versus 0.2611/0.2750/0.3491 for FUSE with 19/49/199 permutations, using 34.3%/37.6%/50.8% less time, respectively. The RA time includes both cutoff calculations. The displayed power comparison gives every method the same sample size, and runtime is measured separately. At frequency 11 and amplitude 0.6, RA reaches power 0.578 versus 0.396/0.410/0.490 for FUSE with 19/49/199 permutations using the same observations for every method. FUSE's pooled feature Gram normalization remains when the permutation count is reduced. Timings on uniform-null streams cover raw observations through the final p-value, including feature evaluation, normalization and calibration, with one BLAS thread after warmup; data generation is excluded.

For an unknown direction, our operational prescription jointly aggregates the predetermined grid

aλ1∈0.025,0.05,0.10,0.25,0.50,1.a\lambda_1\in\\{0.025,0.05,0.10,0.25,0.50,1\\}.

The efficiency identity explains how smoothing attenuates signal and noise at different frequencies. When independent pilot information is available, it may select among a fixed finite menu. If member jj has test-sample null rejection probability pj,n→αp_{j,n}\to\alpha and the pilot selects JnJ_n independently, then

\le\max_j|p_{j,n}-\alpha|\to0.$$ For conservative members the same argument gives the corresponding upper bound. This supplies asymptotic level protection for independent-pilot selection. The cutoff uses the testing-sample size; power depends on the selected scales and the alternative's retained frequencies. **Q2.** Every $\beta\in(0,1/4)$ satisfies both cutoff-rate conditions, together with the covariance convergence assumption stated in W2. We recommend $\beta=1/8$ as a simple prespecified choice; $\beta=1/6$ is also admissible. The submission uses $1/8$ for its local spectral experiments and $1/4$ for its non-local cells. Exponent $1/4$ satisfies the null requirement and lies on the boundary of the stronger local requirement. Continued in next part

Q3. We kept the selected observations, primary order, observed statistic and cutoff fixed and recomputed increments, covariance, rank and the full statistic under pooled permutations. With 9999 permutations, p=(1+\\#\\{Q_b\ge Q_{\mathrm{obs}}\\})/10000; Holm correction is applied separately to each three-test family.

ContrastAsymptotic Holm pRA-statistic permutation pRA-statistic permutation Holm p
Control–trigeminal0.22340.23190.2319
Control–olfactory0.01810.00130.0039
Trigeminal–olfactory0.09890.01110.0222

The first two decisions stay unchanged; trigeminal–olfactory rejects at the 0.05 Holm level under permutation calibration. This directly isolates the effect of calibration for the same observed analysis. The finite-sample permutation argument assumes pooled exchangeability; release-site or batch dependence requires a corresponding sampling model.

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

Reviewer_r4vFOpenReview ↗

Answer to rebuttal

Thank you for the detailed responses and the additional experiments, which answer my questions on the runtime comparison, and the choice of threshold. I encourage you to include these new results in the final version, and I will maintain my current score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.