acceptodds
Under review as a conference paper at ICLR 2027

Geometry-Aware Distributional Change-Point Inference

Abstract

A distributional change in independent manifold-valued observations can preserve location summaries and remain hard to detect even after a representation makes it visible. We study offline testing and localization through the contrast retained by geometric features and the variation assigned to that contrast. Our scan combines multiscale diffusion features with geodesic distance critics and controls finite-sample type-I error by jointly permuting all channels under the iid null. For fixed channels with stable standardizers, we derive local feature-drift limits and consistent localization whenever a retained contrast is nonzero. The leading trace-standardized signal is the fraction of pooled feature variance explained by the change. This characterization motivates projecting distance features off a finite diffusion span: when the diffusion contrast vanishes, projection preserves the distance contrast and yields an explicit variance-reduction factor. Circle, sphere and torus experiments demonstrate sensitivity to matched-moment changes and distinguish detection from localization. A frequency-band comparison recovers the main residual gain at lower cost, showing that variance allocation, as well as signal retention, determines the value of a geometric representation.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Change point, CUSUM, Manifold-valued data

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_StzUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies how to detect and locate one change in the distribution of independent observations. For each possible position of the change, the method compares averages before and after that position. To do that, the paper uses functions derived from diffusion and distances to selected anchored points. The largest result across all positions and all groups of measurements is described using common random permutations. The main idea is that including information about a change is not enough. A method can still have low quality if most of the variance comes from quantities unrelated to that change. The authors then propose predicting the distances from the diffusion measurements and then using the remaining part of those distances and under a precise condition, this preserves the difference between the two distributions while reducing variance.

Critical strengths

Strengths. The statistical construction is good. The permutation procedure accounts for the search over both possible change positions and different measurements. The paper also explains when quantities fitted using all observations can be reused after permutation and the experiments compare methods on the same generated samples and report uncertainty from repeated simulation.

The interpretation of signal strength is interesting. Equation (9) expresses the leading signal as the fraction of total variance explained by the difference between the two parts of the sequence. This explains well why a kernel that can distinguish every pair of distributions may nevertheless perform poorly with a limited number of observations.

The circle example makes this explanation concrete. Removing variation unrelated to the change increases the leading signal by a significant factor and this is a factor for the theoretical signal.

Finally, the authors are honest and report cases where their method performs worse. Adding distances reduces power in one torus experiment. Wasserstein distance performs better in a circle concentration experiment, and Energy detects some changes more often. This reporting helps readers understand the limits of their theory

Critical weaknesses

My main concerns are the following.

  1. The practical benefit of the regression step remains unclear to me as the explicit guarantee applies when the retained diffusion measurements have the same expectations before and after the change. Under this condition, the regression preserves the difference in mean distances while reducing variance. This is a sufficient condition for the stated gain, but to me not a necessary condition for every possible improvement. The main example places the change at frequency 3 while retaining only frequencies 1 and 2 in the diffusion measurements. It demonstrates the proposed mechanism correct but, the main experiments use a much richer Nyström approximation, with a higher rank. The simpler Band method is also competitive as it separates frequencies into groups and divides each group by its own variance. At frequency 3, it reaches the same power as the regression method, (0.996). These results support the importance of treating variation at different frequencies separately. They provide less evidence that regressing distances on diffusion measurements is the preferred solution.

  2. Adding distances provides little improvement in the main experiments. Indeed, in Table 5, the method using distances does not show a clear advantage over the method using diffusion alone. It seems that the combined procedure usually performs similarly to diffusion alone. Thus, stronger evidence is needed to support the practical value of combining diffusion measurements with distances. Evaluating the distances after regression in these same experiments would be particularly informative.

  3. Some pertinent direct comparisons are missing. I would use a Gaussian kernel on the usual vector representation of the observations, with several kernel widths. This would help determine whether the improvement comes from using the intrinsic geometry or from comparing several scales. Methods that rescale the calculated coordinates using their full covariance matrix are also relevant.

  4. The applications are not yet demonstrated. All experiments use synthetic data on the circle, sphere, or torus. The motivating examples in the introduction include wind observations and sequences of covariance matrices, where dependence between successive observations can be important but the current theoretical guarantees assume independent observations. The synthetic experiments respect this assumption, but the connection to the motivating applications therefore remains narrow. An application to real observations would help.

Important : the presentation could also be more focused because the main ideas are obscured by very dense terminology and several additional results

Questions for authors
  1. Which theoretical results do you point as the main new contributions beyond permutation, cumulative sum, and regression ? 2.Does replacing the original distances by the remaining distances after regression improve the main experiments on the circle, sphere, and torus ? 3.How does the method compare with a Gaussian kernel using several widths, and with rescaling based on the full covariance matrix? 4.When should a practitioner choose the regression method rather than the simpler Band method?
  2. What are the total computation times for the main experiments?
7 comments

Thank you for your time and valuable comments.

W1. The regression step addresses a specific statistical obstacle: a distance feature can contain a change omitted by a finite diffusion representation, while most of its variance repeats directions already represented by diffusion. Adding that feature makes the change visible, but does not necessarily make its signal large enough to detect. For feature mean contrast δ=PΨ−QΨ\delta=P\Psi-Q\Psi and change fraction θ\theta, the leading trace-standardized signal is

θ(1−θ)∣δ∣2θEP∣Ψ−PΨ∣2+(1−θ)EQ∣Ψ−QΨ∣2+θ(1−θ)∣δ∣2.\frac{\theta(1-\theta)\\|\delta\\|^2}{\theta E_P\\|\Psi-P\Psi\\|^2+(1-\theta)E_Q\\|\Psi-Q\Psi\\|^2+\theta(1-\theta)\\|\delta\\|^2}.

This is the fraction of pooled feature variance explained by the change.

The paper contains a direct ablation of this obstacle. In the third-harmonic experiment, all methods use the same samples and permutations, and their feature rules are fixed across the tested frequencies:

MethodRejection frequencyUnconditional location MAE
Diffusion modes 1–2, H20.0500.237
H2 + raw distances0.0540.232
H2 + residual distances0.9960.029
Diffusion modes 1–6, H60.1700.092
Frequency-band diffusion, Band0.9960.028

The residual versus raw power gain is 0.942. The explicit circle calculation explains its magnitude: projection removes approximately 98.6% of the distance variance, preserves the changing third harmonic contrast, and increases the population signal slope from approximately 0.00037 to 0.0258, a factor of 69.1. H6 already contains the changing harmonic and has the same number of scalar coordinates as the residual construction, yet its power is 0.170. Thus neither raw distance augmentation nor increasing the representation size reproduces the improvement. The regression is a substantive variance allocation step in this construction.

Zero diffusion contrast is a sufficient condition for the stated gain; the general tradeoff also allows nonzero diffusion contrast. The general pooled-regression calculation gives

\qquad v_\perp=\operatorname{tr}\\!\left(\Sigma_{FF}-\Sigma_{FH}\Sigma_{HH}^{-1}\Sigma_{HF}\right).$$ Writing $v_F=\operatorname{tr}(\Sigma_{FF})$, the residual channel has a larger population signal slope precisely when $$\frac{\\|\delta_F-B_H\delta_H\\|^2}{\\|\delta_F\\|^2} > \frac{v_\perp}{v_F},$$ provided $\delta_F\ne0$ and both variances are positive. In words, the proportion of squared contrast retained must exceed the proportion of variance retained. This criterion allows $\delta_H\ne0$ and makes explicit the tradeoff that the zero-contrast example isolates. Because the original diffusion channels are retained, the invertible transformation $(H,F)\mapsto(H,F-B_H H)$ also preserves the existence of a nonzero joint mean contrast. This preserves retained information and consistency under the stated assumptions; finite-sample power also depends on variance allocation and joint calibration. A concrete extension illustrates this point. For balanced circle distributions $$p_\pm(x)=\frac{1\pm\\{0.10\cos x+0.35\cos(3x)\\}}{2\pi},$$ with the same H2 span and 16 equally spaced distance landmarks, the diffusion contrast is nonzero. Applying the residual contrast and covariance formulas above gives raw and residual distance slopes of 0.002836 and 0.025759, respectively, a **9.08-fold increase**. Projection removes some distance contrast, but removes a much larger proportion of its variance. In the mixed-harmonic experiment, the raw and residual combinations have powers **0.186 and 0.986**, respectively. Band also has power 0.986. The full-covariance comparison appears in W3 below. This extends the evidence beyond the sufficient zero diffusion contrast condition and shows how regression implements effective variance allocation. Band's strong performance supports the same signal and variance explanation. It offers an economical solution when useful spectral groups are available: at the third harmonic it achieves the same power in 5.9 ms versus 7.2 ms for residualization, and at the fourth harmonic it performs better (0.344 versus 0.072). Pooled regression offers a complementary implementation that uses a selected feature span and a distance dictionary without requiring a prescribed frequency partition. Its fitted values are invariant to an invertible change of basis within the diffusion span. We recommend Band when an interpretable spectral partition is available, and residual distances when complementary distance contrasts coexist with substantial variation predictable from the retained span. The theory and experiments support this conditional recommendation. (revised Section 3.2, Section 4.2 and Proposition D.1). Continued in the next part

W2. The role of distances is to supply complementary contrasts outside a computed finite diffusion span. Exact heat kernel parametrizes the infinite representation; it does not guarantee that a finite approximation retains every contrast or weights it effectively. The third harmonic example demonstrates this distinction directly: H2 has zero population contrast, while the distance dictionary has a nonzero contrast that becomes effective after regression. Consequently, the residual distance construction detects a change that the retained diffusion channel alone cannot detect consistently through its mean contrast.

The additional ordinary directional mixture study provides another, more modest example of distance complementarity. With the fixed H2 representation, adding raw distances increases power from 0.436 to 0.468 in the clean weak mixture setting and from 0.362 to 0.396 with common measurement noise; the paired gains are 0.032 and 0.034. Residualization gives 0.418 and 0.350 in those same cells, underscoring that the value of distances and the benefit of projecting them are distinct questions.

This is also why raw distances and residual distances should be distinguished in the original study. The primary multiscale Nyström Combined method adds raw distances to a much richer Nyström representation. The rich-representation results show that the raw distance channel usually contributes little in those designs. It also reveals a genuine search cost: for the strongest torus dependence change, Heat has power 0.734 and Combined 0.668 because the additional channel contributes no new rejecting samples while increasing the joint threshold.

The following comparison uses the submitted implementation with 500 new paired replications per condition at n=500n=500 and 199 common permutations. It replaces raw distances by residuals against a basis of the retained diffusion span. The submitted Heat/Combined powers for these three conditions are 0.068/0.068, 0.238/0.226 and 0.092/0.088; the independent rerun below evaluates all additional methods on common samples.

Original-design conditionHeatRaw CombinedResidual CombinedExact Gaussian5Rank64 white
Circle shape, middle0.0660.0660.0660.0680.064
Sphere shape, middle0.2660.2560.2660.2840.692
Torus dependence, middle0.0620.0600.0620.0580.362

Exact Gaussian5 uses full ambient Gaussian Gram matrices exp⁡−D2/(8γmD)\exp\\{-D^2/(8\gamma m_D)\\} at γ∈2−6,2−4,2−2,1,4\gamma\in\\{2^{-6},2^{-4},2^{-2},1,4\\}, where mDm_D is the pooled median squared ambient distance. Rank64 white retains the leading 64 pooled principal coordinates of the concatenated heat/distance features and rescales them to unit covariance. Every method calibrates its complete split search with the same permutations. Rank64 white has higher power than Combined in the sphere and torus rows (0.692 versus 0.256 and 0.362 versus 0.060), showing the value of covariance weighting in these designs. The folded-embedding comparison in W3 examines a separate role for intrinsic geometry.

We ran a fine-scale change experiment using the original multiscale Nyström implementation. The two segment densities on the known unit circle are p±(θ)=[1±acos⁡(kθ)]/(2π)p_\pm(\theta)=[1\pm a\cos(k\theta)]/(2\pi), with 1000 observations and a change at observation 500. Every method uses the same observations and 199 common permutations, trim 0.10, and nominal level 0.05. The five diffusion-time multipliers, rank cap 128 and 64 distance landmarks are unchanged; residualization uses a numerical basis of the retained diffusion span. The representation receives neither the changing frequency nor its amplitude. The Gaussian comparison uses the standard unit-circle embedding (cos⁡θ,sin⁡θ)(\cos\theta,\sin\theta) and nine bandwidths. The table reports 500 independent repetitions per setting.

Here the covariance control uses the unnormalized joint heat/distance coordinates with λ=tr⁡(Σ^)/(qn)\lambda=\operatorname{tr}(\widehat\Sigma)/(q\sqrt n), where qq is their total number.

SettingHeatCombinedResidual CombinedUnnormalized covariance ridgeGaussian (9 scales)
Uniform null0.0660.0660.0660.0480.066
k=19k=19, a=0.35a=0.350.0540.0540.7320.0480.054
k=19k=19, a=0.70a=0.700.0560.0561.0000.0820.054
k=23k=23, a=0.70a=0.700.0420.0421.0000.0600.042

At k=19k=19, a=0.70a=0.70, residualization raises power from 0.056 to 1.000, and at k=23k=23, a=0.70a=0.70, from 0.042 to 1.000. This provides a multiscale Nyström example in which distance residuals retain a fine-scale contrast while regression removes variation shared with diffusion. Together with the coarse-scale results above, it identifies when residualization is useful. (revised Section 4.2, Table 4 and Appendix B.7, Table 12).

Continued in the next part

W3. We compared the original multiscale Heat and Combined procedures with exact ambient Gaussian kernels and a full-covariance scan. For a known unit-speed circle embedding

ιε,k(x)=(1−ε2kcos⁡kx,1−ε2ksin⁡kx,εcos⁡x,εsin⁡x),0<ε<1,\iota_{\varepsilon,k}(x)=\left(\frac{\sqrt{1-\varepsilon^2}}{k}\cos kx,\frac{\sqrt{1-\varepsilon^2}}{k}\sin kx,\varepsilon\cos x,\varepsilon\sin x\right),\qquad 0 < \varepsilon < 1,

the derivative has norm one and the last coordinate pair identifies the angle. The round case is k=1k=1, ε=0.30\varepsilon=0.30, which has the same chord distances as the standard unit circle. Intrinsic arc distances and the heat kernel are unchanged. At k=3k=3, locations separated intrinsically by 2π/32\pi/3 are only 3ε\sqrt3\varepsilon apart in the ambient space. This isolates the value of the supplied intrinsic metric.

The pre/post laws are Pδ=(1/2+δ)vM(0,4)+(1/2−δ)vM(2π/3,4)P_\delta=(1/2+\delta)\mathrm{vM}(0,4)+(1/2-\delta)\mathrm{vM}(2\pi/3,4) and Qδ=P−δQ_\delta=P_{-\delta}. For n=500n=500 and change fraction 0.5, C and H are the original Combined and Heat scans; G9 is the exact nine-width Gaussian, WG9 applies it after pooled ambient whitening, and UG9 applies it after recovering the known circle coordinates. Block-normalized covariance ridge first normalizes each heat/distance block to unit pooled trace and jointly calibrates ridge factors 0.1,1,100.1,1,10.

Signal / embeddingCHG9WG9UG9Block-normalized covariance ridge
Null, round0.0440.0460.0460.0480.0460.042
Null, folded ε=0.10\varepsilon=0.100.0440.0460.0440.0460.0460.042
δ=0.08\delta=0.08, round0.7880.7920.7940.7000.7940.480
δ=0.08\delta=0.08, folded ε=0.10\varepsilon=0.100.7880.7920.0680.5760.7940.480
δ=0.12\delta=0.12, folded ε=0.10\varepsilon=0.100.9940.9940.0960.9660.9940.924

The geometric representation raises power from 0.068 to 0.788 relative to the nine-width ambient Gaussian in the folded δ=0.08\delta=0.08 case, and from 0.576 to 0.788 relative to its ambient-whitened version. The geometry-assisted Gaussian reaches 0.794, showing that supplying the same geometric preprocessing can also make that baseline competitive. Heat and Combined perform similarly here; their shared gain comes from the intrinsic representation.

The Gaussian kernels are exp⁡−D2/(8γmD)\exp\\{-D^2/(8\gamma m_D)\\}, where mDm_D is the pooled median squared distance and γ∈2−14,2−12,…,22\gamma\in\\{2^{-14},2^{-12},\ldots,2^2\\}. Each method calibrates the entire declared scale-and-split search using the same 199 permutations. The ridge-regularized full-covariance scan is our implementation of the reviewer's suggested covariance rescaling: joint coordinates are transformed by an estimated covariance with ridge regularization for numerical stability. It is distinct from ambient-coordinate whitening. On the rich features, the unnormalized ridge rule λ=tr⁡(Σ^)/(qn)\lambda=\operatorname{tr}(\widehat\Sigma)/(q\sqrt n) has null 0.050 and powers 0.108/0.460/0.908 at δ=0.04/0.08/0.12\delta=0.04/0.08/0.12; the stronger block-normalized three-ridge control has null 0.042 and powers 0.118/0.480/0.924.

For the restricted Fourier comparisons, the Fourier covariance-ridge benchmark uses H2 at time 1/161/16 and 16 distance landmarks, with λ=tr⁡(Σ^)/(20n)\lambda=\operatorname{tr}(\widehat\Sigma)/(20\sqrt n) and trace normalization after transformation. In the mixed-harmonic comparison, residual distances have null 0.042, power 0.986 and unconditional MAE 0.0306; the Fourier covariance-ridge scan has null 0.052, power 0.946 and MAE 0.0389. For the weak ordinary mixture, the same covariance benchmark has power 0.504 clean, 0.408 with wrapped noise, and 0.400 under 5% uniform contamination at change fraction 0.30. The covariance benchmark is stronger than raw augmentation in the clean and noisy weak-mixture cells (0.504 versus 0.468 and 0.408 versus 0.396); raw augmentation is stronger under contamination (0.428 versus 0.400).

Continued in the next part

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

The paper proposes a technique for an offline scan to identify distributional changes. If I understand correctly, the case where distributional data samples are supported on a manifold, geometric representations are particularly important and can have major consequences on whether distributional changes can be detected. To that end, the paper provides theory regarding the effects of geometric features and statistical tests so that distributional changes on data with rich geometric structure can be reliably and accurately detected.

Critical strengths

The paper addresses a real mathematical problem arising from actual tasks in wind-farm forecasting. This adds credibility to its utility in applications. The paper also does an excellent job of mathematically formalizing the problem at hand and presenting the proposed method with strong technical rigor.

Critical weaknesses
  1. Despite the paper's introduction of wind-farm forecasting in the first sentence, this application appears to never be addressed again. This makes it unclear whether the results presented in this work are actually relevant to the motivating problem.
  2. The writing is frequently hard to follow. While the work appears highly technically rigorous, Sections 2 and 3 in particular would benefit from better narrative framing connecting the definitions and theoretical results.
Questions for authors
  1. Experiments appear to be on simpler tasks, and their relevance to the motivating wind-farm forecasting task is unclear. Are these simpler tasks representative in some sense of the motivating problem at the beginning of the paper, and if not, what are the challenges inhibiting this to be directly applied to a more realistic dataset?
  2. Are there any limitations to this work that could be included in the conclusions?
3 comments

Thank you for your time and valuable comments.

W1. The intended connection is to the regime-detection stage of a forecasting pipeline: detecting a change in the distribution of directional observations and estimating the boundary of the current regime. Assessing forecasting gains from that boundary requires a downstream application evaluation.

The circle, sphere, and torus represent an azimuth, a three-dimensional direction, and a pair of angular measurements, respectively. The primary experiments cover ordinary location and concentration changes, changes in multimodality with the mean preserved, and changes in dependence between angular coordinates with their marginal laws preserved. These are features that a directional regime detector may need to distinguish. For example, in the strongest matched-mean sphere mixture, Combined rejects in 99.0% of repetitions while the extrinsic mean scan rejects in 4.0%, despite the unchanged population mean vector. This demonstrates sensitivity to distributional information beyond the mean.

A concrete idealized application is a sequence of independent directional observations from successive controlled runs or independently sampled episodes, with one distributional switch. The geometry is supplied by the measurement itself. Under this sampling model, the permutation argument provides finite-sample permutation validity; the feature-drift and permutation-consistency results give localization and detection consistency when a fixed representation retains a nonzero contrast and the stated conditions hold. The fitted-circle representation result additionally addresses pooled fitted features on the circle under its specific conditions.

To supplement the mechanism examples, we also examine an ordinary circle mixture change: P=vM(0,4)P=\mathrm{vM}(0,4) before the switch and Q=0.7,vM(0,4)+0.3,vM(π/2,8)Q=0.7\\,\mathrm{vM}(0,4)+0.3\\,\mathrm{vM}(\pi/2,8) afterward, at change fraction 0.30. This represents the appearance of an additional preferred direction without matching moments or selecting a changing harmonic. The observation variants add wrapped Gaussian noise with SD 0.25 radians or independently replace 5% of observations by uniform circle draws. H2 retains diffusion modes 1–2 and H6 modes 1–6 at times 1/161/16 and 1/41/4; Band separates modes 1–2 and 3–6 into separately standardized channels at both times. Raw and residual distances use 16 equally spaced circle landmarks. At n=500n=500, H2, raw- and residual-distance augmentation, H6 and Band each reach power 1.000 in these weight-0.30 settings. With mixture weight 0.10, the raw-distance powers are 0.468/0.396/0.428 for clean/noisy/contaminated observations, versus 0.436/0.362/0.402 for H2.

The obstacle to applying the current theorem directly to a chronological wind series is temporal dependence, together with the need to specify relevant covariates, seasonal variation, and the target regime change. Validity for dependent observations requires a resampling method and analysis appropriate to the complete multichannel search. The present experiments provide controlled evidence about directional distribution changes under the specified sampling model. (revised Section 4.1 and Appendix B.8, Table 13).

W2. We simplify the procedural explanation to the following:

We construct diffusion and landmark-distance features, normalize their pooled variation, and calibrate the joint maximum over channels and splits by common permutations. The analysis connects retained contrast to the triangular CUSUM drift, localization and the fraction of pooled variance explained by the change.

This introduces the statistical role of each construction before the harmonic calculation and technical details. (revised Section 2).

Q1. The regime-detection example in W1 can be tested directly with independent directional runs. For the ordinary circle mixture with weight 0.30, n=500n=500 and change fraction 0.30, H2, raw-distance augmentation, residual-distance augmentation, H6 and Band each have power 1.000 under clean observations, wrapped Gaussian noise and 5% uniform contamination. With mixture weight 0.10, raw-distance augmentation has powers 0.468/0.396/0.428, compared with 0.436/0.362/0.402 for H2. These results quantify detection of a new directional component under the stated observation mechanisms. Forecasting gains require a downstream evaluation using the detected regime boundary. (revised Appendix B.8, Table 13).

Continued in the next part

Q2. The applicable regime is explicit: The primary guarantees concern independent observations and one change in the interior of the trimmed search interval. Intrinsic diffusion requires a specified geometric structure, while a finite representation detects only contrasts that survive its approximation. Regression can remove useful signal as well as variance when the diffusion contrast is nonzero, and adding channels can increase the joint threshold. Consequently, the full combination is not uniformly preferable to simpler alternatives such as Band.

The fitted-representation consistency theorem applies to the circle under its stated regularity conditions. Unknown-manifold recovery and fitted-feature local-alternative rates require further analysis. Localization MAE and tolerance probabilities quantify point-estimation accuracy; interval inference additionally requires uncertainty calibration. Further work is needed for dependent observations, uncertain geometry, and multiple regime changes. (revised Conclusion, Section 5).

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 detailed responses and clarification! I maintain my original positive rating of the paper.

Reviewer_XvFkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies offline detection and localization of a single distributional change point in independent manifold-valued observations. It proposes a multichannel CUSUM-type scan that combines multiscale diffusion/heat-kernel features with geodesic distance critics, standardizes channels using pooled variance, and jointly calibrates the maximum over candidate splits and channels using permutations. For fixed features, the theory connects detection and localization to the retained feature contrast and to the fraction of pooled feature variance explained by the change. The paper further introduces a residual distance channel obtained by projecting landmark-distance features onto the orthogonal complement of the retained diffusion span, and proves that this projection can preserve distance contrasts not represented by diffusion features while reducing irrelevant shared variance. Experiments on the circle, sphere, and torus evaluate type-I error, detection power, and localization accuracy under location, concentration, shape, and dependence changes, with comparisons to extrinsic mean scans, exact Wasserstein-1, and energy-distance baselines.

Critical strengths
  1. Clear variance-allocation principle for geometric feature scans. The paper gives a useful characterization of when a representation detects a change: the relevant fixed-change signal depends not only on whether the feature contrast is nonzero, but also on how much pooled variance is irrelevant to that contrast. This helps explain why a rich geometric representation can still have poor finite-sample sensitivity, and it gives an intuitive rationale for residualizing distance features against the retained diffusion span. This is consequential because it moves the discussion beyond characteristic kernels or exact manifold embeddings toward practical statistical weighting of retained information.

  2. Joint permutation calibration is appropriate for the proposed scan. Since the statistic maximizes over candidate split points and multiple channels, the calibration is important. The paper correctly uses common random permutations and calibrates the entire maximum statistic under the iid null. This is consequential because many multichannel scan procedures can become anti-conservative if the search is ignored; here the proposed inference is at least finite-sample valid under the stated iid null and declared algorithm.

Critical weaknesses
  1. The problem setting feels overly idealized and the practical relevance is not convincingly demonstrated. The method assumes independent manifold-valued observations from a known or well-behaved closed Riemannian manifold, with a single distributional change at an unknown fixed fraction θ, and experiments are limited to circle, sphere, and torus designs with handcrafted matched-moment or harmonic alternatives. In many realistic settings, observations may be noisy, dependent, heterogeneous in dimension, embedded in higher-dimensional spaces, affected by drift, outliers, or nuisance variation, and the relevant manifold may not be known. The motivation involving wind-farm regimes is promising, but the paper does not evaluate the method on real or semi-real data, nor does it show that the diffusion/distance/residual machinery outperforms simpler nonparametric methods when the change is not deliberately matched to the theory. This matters because the central question is not just whether the proposed test can detect carefully constructed geometric changes, but whether such a setting is common enough or important enough to justify this level of geometric structure.

  2. The experiments are too simple to establish the paper’s main practical claim. The synthetic experiments are useful for checking theory, but they mostly test cases where the change is exactly aligned with a harmonic, landmark distance contrast, matched moment, or known manifold structure. The paper itself shows that frequency-band scans can recover most of the benefit of the more elaborate residualization, which weakens the practical appeal of the full diffusion–distance construction. There is also no robustness study to noise, unknown manifolds, nuisance drift, uneven segment lengths, near-boundary changes, multiple changes, or real data. This is consequential because the proposed method makes many assumptions and choices—diffusion times, landmarks, pooled fitting, channel weights, residual projection—and the paper does not convincingly show that those choices remain beneficial outside the clean theoretical regime.

Questions for authors
  1. Can you clarify the definition of τ_n in Eq. (2)? Please define τ_n as the integer change index, e.g., τ_n ∈ {1,...,n−1}, so that θ_n = τ_n/n and θ_n → θ.

  2. What is the intended practical scope of the problem setting? Could the authors provide evidence that such single-change, independent, manifold-valued sequences with a known geometry occur often enough in applications to motivate this method? A real-data or semi-real-data experiment would substantially strengthen the paper.

  3. How sensitive is the proposed scan to mis-specified or unknown geometry? What happens if the data lie near but not exactly on a low-dimensional manifold, contain noise, have outliers, or have drift that is not the target change?

  4. Why should a practitioner prefer the full diffusion–distance–residual procedure over simpler alternatives such as fixed frequency-band diffusion scans, distance-profile scans, kernel CUSUM, or Wasserstein-based scans? The current experiments suggest that simpler banding can recover much of the benefit at lower cost.

  5. How should diffusion times, landmarks, channel weights, and the residual projection be chosen in applications where the changing harmonic or geometric direction is unknown? Is there guidance or robustness results for these choices?

  6. The localization evaluation reports unconditional MAE, including non-rejections. Could the authors also report conditional localization accuracy given rejection, or confidence intervals around estimated change points? This would better separate detection quality from localization quality.

7 comments

Thank you for your time and valuable comments.

W1. Combined is our principal proposed scan: it jointly uses multiple diffusion/heat-kernel measurements and geodesic distance measurements. Residual Combined retains the diffusion measurements and replaces the distance measurements by their regression residuals against the retained diffusion span. The main study uses a multiscale Nyström implementation. The supplementary H2 + raw distances (H2+F) and H2 + residual distances (H2+F⊥F^\perp) are restricted Fourier proxies that isolate these two mechanisms; they are not identical to the full Nyström procedures. H6 and Band are the simpler spectral comparisons used below.

The target is offline detection and localization of one distributional change in independent observations with a common representation. We formulate the feature scan on a Polish metric space; a closed connected Riemannian manifold is used for the intrinsic heat-kernel construction. The landmark distance features require only a compact metric space. These are the geometric requirements of the respective constructions.

The circle, sphere, and torus encode angles, directions, and paired angles. An application model within the assumptions is independent directional measurements from successive controlled runs, with one change in the generating regime. Dependence between coordinates within an observation is allowed, as in the torus experiment; temporal independence concerns different observations. A chronological wind series requires additional dependence-aware calibration, under its serial sampling structure.

The primary experiments include location, concentration, mixture-shape and dependence changes under standard directional distributions. For example, Combined detects the strongest matched-mean sphere mixture with power 0.990 while the extrinsic mean scan rejects with frequency 0.040. These designs establish sensitivity to relevant distributional features, while leaving the need for a real application clear.

We ran an ordinary directional-mixture experiment that does not match moments or prescribe a changing harmonic. Each observation is an independent direction on the known circle. Before the switch, P=vM(0,4)P=\mathrm{vM}(0,4); afterward,

Q=0.7,vM(0,4)+0.3,vM(π/2,8).Q=0.7\\,\mathrm{vM}(0,4)+0.3\\,\mathrm{vM}(\pi/2,8).

The switch occurs at θ=0.30\theta=0.30, so the segments have unequal lengths. We additionally examine common wrapped Gaussian measurement noise with standard deviation 0.25 radians, and a separate 5% uniform contamination condition applied independently with the same rule to both segments. The null in each case draws iid observations from PP through the same observation model. These are controlled illustrations of an independent run application. (revised Section 4.1 and Appendix B.8).

W2. The comparison uses n=500n=500, trim 0.10, 199 common permutations and nominal level 0.05. The methods reported here use the same observations and permutations, with each permutation repeating the complete channel-and-split maximum. The Fourier implementations use diffusion times 1/161/16 and 1/41/4 and 16 equally spaced distance landmarks. Their feature rules are fixed across scenarios.

A. Does adding distance information improve detection? We examine the ordinary-mixture change from W1 and its weaker version, Q=0.9,vM(0,4)+0.1,vM(π/2,8)Q=0.9\\,\mathrm{vM}(0,4)+0.1\\,\mathrm{vM}(\pi/2,8). The weaker follow-up was specified after the original weight-0.30 mixture reached the detection ceiling. The table reports clean, noisy and contaminated observations at change fraction 0.30. For the clean weight-0.10 mixture, near-boundary checks at fractions 0.15 and 0.85 give H2/raw/residual powers 0.240/0.240/0.228 and 0.308/0.354/0.300, respectively.

Weak-mixture results. Power for the weaker ordinary-mixture change. Noise is wrapped Gaussian with SD 0.25 radians; contamination is 5% uniform, applied independently by the same rule to both segments.

MethodClean, 0.30Noise, 0.30Contamination, 0.30
H20.4360.3620.402
H2 + raw distances0.4680.3960.428
H2 + residual distances0.4180.3500.378
H60.4300.3580.408
Band0.4040.2980.320

Adding raw distances to H2 increases observed power from 0.436 to 0.468 in the clean setting, from 0.362 to 0.396 under noise, and from 0.402 to 0.428 under contamination. The paired gains are 0.032, 0.034, and 0.026. Thus additional distance information provides modest gains in these ordinary directional changes, including the specified noise and contamination. Residualization does not improve these particular cells: its powers are 0.418, 0.350, and 0.378. Distance augmentation and residualization address different statistical questions.

continued in the next part

B. When is residualization beneficial? The balanced mixed-harmonic alternative

p±(x)=1±[0.10cos⁡x+0.35cos⁡(3x)]/(2π)p_\pm(x)=\\{1\pm[0.10\cos x+0.35\cos(3x)]\\}/(2\pi)

has a nonzero retained diffusion contrast and a uniform pooled null. It tests whether regression is useful when the change involves both a retained harmonic and an omitted harmonic.

Mixed-harmonic results. Mixed-harmonic change with nonzero retained diffusion contrast. MAE is measured as a fraction of the sequence length.

MethodNull rejectionPowerRejecting runsMAE, all runsMAE, rejecting runs
H20.0500.17286/5000.19730.1458
H2 + raw distances0.0600.18693/5000.18500.1390
H2 + residual distances0.0420.986493/5000.03060.0299
H60.0500.356178/5000.09100.0749
Band0.0420.986493/5000.03030.0292

The residual proxy raises power from 0.186 to 0.986 relative to raw augmentation, a difference of 0.800, while unconditional MAE falls from 0.1850 to 0.0306. The regression basis has rank four in every replication, and the median residual-to-raw distance variance ratio is 0.01437. Much of the distance variation is predictable from diffusion but contributes little useful change signal; regression removes that variation. The residual contrast and variance calculation quantifies the tradeoff between retained squared contrast and residual variance. For this design, the population distance-channel signal slopes are 0.002836 before and 0.025759 after regression, a 9.08-fold increase. The example therefore extends the variance-reduction mechanism beyond a zero diffusion contrast. Band also achieves power 0.986; its paired difference from the residual proxy is zero. This supports the variance-reduction mechanism and identifies Band as an effective alternative when suitable frequency groups are available.

C. What happens under more complicated conditions? At ordinary-mixture weight 0.30 and change fraction 0.30, all compared methods reach power 1.000 in the clean, noisy and contaminated conditions. The residual proxy has null rejection 0.034/0.036/0.054 and normalized MAE 0.0138/0.0139/0.0155, respectively; conditional and unconditional MAE coincide. The weaker-mixture and mixed-harmonic comparisons above distinguish the methods below this detection ceiling.

The weak-mixture and mixed-harmonic results distinguish distance augmentation from residualization under noise, contamination and multi-component changes. The near-boundary checks quantify the loss of power as a segment becomes short.

These restricted Fourier studies explain component behavior; the primary multiscale Nyström comparisons evaluate the richer proposed procedure. Together, they relate performance to retained contrast, variance allocation and joint calibration.

For noise and uncertain geometry, calibration and power must be distinguished. If a common observation mechanism is applied independently to iid observations, the observed null remains iid. The permutation argument therefore still applies to a pooled permutation-invariant feature construction, or to an order-dependent construction refitted within every permutation. Power depends on retention of the alternative contrast, which an observation mechanism can attenuate or erase.

There is also a useful stability implication of the existing results. The transport bound makes the fixed diffusion map Φt\Phi_t Lipschitz, with constant LtL_t, and the distance critics are 1-Lipschitz. For perturbed laws P~,Q~\widetilde P,\widetilde Q on the same known geometry,

\le L_t\\{W_1(\widetilde P,P)+W_1(\widetilde Q,Q)\\}.$$ Each distance contrast has the analogous scalar bound with constant 1. Thus a fixed contrast larger than its perturbation bound remains nonzero, and the fixed-feature consistency result continues to apply under its other assumptions. This explains a controlled stability regime for fixed features on the known geometry. Our new noise and contamination checks use the known circle and have that scope. Time-varying nuisance drift is a separate modeling issue. It changes the observed distribution and is therefore an alternative under the current homogeneity null. Distinguishing such drift from a target regime change requires an explicitly different target or a nuisance model, with rejection interpreted relative to that target. (revised Appendix B.8, Tables 13–14, and Proposition C.4). Continued in the next part

Q1. The change occurs at the integer observation index

\theta_n=\tau_n/n\longrightarrow\theta\in(0,1).$$ The localization theorem additionally requires $\theta\in(\eta,1-\eta)$. At a candidate split $r_k=k/n$, the exact bridge mean is $$\mathbb E B_{n,\psi}(r_k) =\sqrt n\\,b_{\theta_n}(r_k)(P_n\psi-Q_n\psi), \qquad b_a(r)=r\wedge a-ra,$$ by the exact bridge-mean identity. Finite-sample identities use $\theta_n$, and $\theta$ denotes its limit. The reported localization error is $|\widehat\tau_n-\tau_n|/n$. (revised Eq. (2) and Theorem 3.2). **Q2.** The intended scope is an offline single-change problem for independent observations with a common geometric representation. A direction recorded from an independently sampled operating episode naturally lies on the circle or sphere; paired angular measurements lie on the torus. The within-observation coordinates may be dependent, although observations must be independent across the sequence. As explained in W1, the primary experiments cover location, concentration, multimodality, and dependence changes beyond the harmonic construction. The new ordinary directional-mixture example in W2 supplies a further non-harmonic illustration, including unequal segment lengths, noise, and contamination. A chronological wind or covariance-matrix application requires specifying the observation units, preprocessing, temporal dependence and scientific change target, followed by dependence-aware calibration and a downstream forecasting evaluation. (revised Section 4.1, Appendix B.8 and Conclusion). **Q3.** The permutation argument guarantees finite-sample null calibration under iid observations when a learned representation or tuning rule is pooled and permutation invariant (or refitted under each permutation). This calibration guarantee is distinct from power: unknown or misspecified geometry can weaken or erase the retained feature contrast. The controlled noise and contamination results in W2 show how the fixed-Fourier procedure behaves on the **known circle** under wrapped Gaussian noise and stationary 5% uniform contamination. The diffusion-map Lipschitz bound and the one-Lipschitz distance critics provide the fixed-feature perturbation bound stated in W2. These perturbations keep observations on the known circle. Off-manifold observations and learned geometry require an observation model and additional approximation analysis. The original Combined scan at $n=500$ and change fractions 0.25/0.75 has power 0.036/0.052 for circle shape, 0.074/0.140 for sphere shape and 0.996/1.000 for torus shape. At the midpoint, increasing $n$ from 200 to 1000 changes sphere power from 0.050 to 0.962 and torus-shape power from 0.656 to 1.000; circle power remains 0.068/0.066. Our additional 0.15 and 0.85 circle experiments evaluate behavior nearer the trim boundaries. Multiple changes and drift unrelated to the scientific target remain outside the current guarantees. Under the paper’s homogeneity null, time-varying nuisance drift is itself a distributional change, so separating it from a target change would require a different null or an explicit nuisance model. (revised Appendix B.6 and Appendix B.8). Continued in the next part

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.