acceptodds
Desk rejected at ICLR 2027

Learning Scalar Quadratic Contrasts for High-Dimensional Change Point Localization

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

A high-dimensional distribution can change through many weak interactions while every coordinate mean and variance remains fixed. We study how to aggregate this signal into one quadratic contrast for change point localization. A normalized matrix representation identifies the Frobenius jump with the change in the degree-two density projection. For bounded densities with low-rank jumps in their degree-two coefficient matrices, a sufficiently strong pilot signal allows the matrix scan to learn a quadratic contrast retaining at least half of the Frobenius jump. With enough data to estimate its two means, independent evaluation observations then yield a small-jump localization error proportional to , up to confidence factors, for a fixed density bound. The pilot pays for learning the aggregate contrast, while its retained signal governs scalar precision. Explicit matrix bounds extend detection and isolation to three temporal dependence models. Under geometric absolute regularity, deterministic spacing also yields a complete refinement guarantee in original time units. Experiments connect retained signal to localization accuracy, explain the effect of working rank, and quantify the sampling and resolution costs of spacing.

Primary area:
learning theory
Keywords:
Change-point detection, High-dimensional statistics, Nonparametric methods, Low-rank matrix estimation, Density projection

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_ZpnhOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how to locate changes in the distribution of a sequence of high-dimensional observations, including changes in dependence between coordinates that leave their individual means and variances unchanged. The method uses part of the data to learn a quadratic function whose expectation differs across a change, then applies that function to separate observations and locates the resulting scalar mean change. Assuming bounded densities and a low-rank matrix of changes in first and second moments, the authors prove detection and localization guarantees, with extensions to temporally dependent observations. Their experiments examine how accuracy depends on dimension, signal strength, the chosen rank, and temporal dependence.

Critical strengths

The method combines changes in many pairwise interactions into a single statistic obtained through a quadratic form, allowing it to detect changes that are difficult to see in individual coordinates. The theory gives explicit conditions under which the estimated statistic preserves enough of the signal for accurate localization.

Critical weaknesses
  1. The paper is extremely poorly-written and almost impossible to read. A clear, self-contained explanation of when the method works and what it improves over existing methods is missing. This is my primary concern, since it makes the contribution difficult to assess even after careful reading.
  2. The advantage over existing methods is not convincingly demonstrated. The experiments do not establish whether combining more directions provides a substantial benefit once the additional data needed to estimate them is taken into account.
Questions for authors
  1. Can you give a self-contained statement of the assumptions and sample size requirements for your main result, and explain precisely what it improves over the closest existing method? In particular, can you improve the writing throughout the paper?
  2. Do you have theoretical or experimental evidence that combining several changing directions substantially improves accuracy over using the strongest direction alone, after accounting for the additional data needed to estimate those directions?
5 comments
AuthorsOpenReview ↗

Weakness 1. When combining directions helps locate a change

Thank you for pointing out that the paper needs an example showing when the extra directions help. Consider two groups of coordinates whose individual means and variances stay fixed. What changes is how the groups move together. In our example, adding the coordinates within each group and multiplying the two sums gives a scalar whose mean changes. If the coefficient matrices are normalized to unit Frobenius norm, the best single direction captures at most 70.7% of the available quadratic moment signal, measured by the change in the scalar mean. Combining two directions with known weights captures all of it.

The question in practice is whether we can learn those weights. For independent observations, we use the odd observations to learn a weighted sum of coordinates, their squares and their pairwise products. We apply that function to the even observations and locate the change in its mean. The guarantee applies when the change in these moments has low rank and each segment is long enough both to learn the weights and to estimate the two scalar means. With high probability the learned function then retains at least half of each jump. Once it is learned, the location error bound has no further factor for the number of coordinates (Theorem 3.3).

For each candidate interval, Wang, Yu and Rinaldo (2021) learn a projection along which the variance changes. Our gain comes from combining several changing components in one scalar. In our 64 dimensional experiment with one known change and unknown location, both ranks of our estimator learned from the same 2,048 observations and evaluated on the same other 2,048. Using two directions reduced the median absolute location error from 24.5 to 4.5 time points over 40 trials (Table 1).

AuthorsOpenReview ↗

Weakness 2. Localization accuracy at the same training and evaluation budgets

Thank you for asking whether the gain survives the cost of estimating the second direction. Table 1 lets us make that comparison within each trial. Both estimators see exactly the same sequence. Each scans the same candidate locations using the odd observations and learns its own weights. One uses the strongest estimated direction and the other uses two. On the even observations, both use the first and last quarters to estimate the two means and the same scalar least squares criterion to locate the change. Thus any difficulty estimating the extra direction already affects the reported error.

Both estimators are told that there is one change and search the central half of the sequence. The true location is drawn uniformly among the integers from 0.4n0.4n to 0.6n0.6n, with nn the total sample size. The dense rank two jump has Frobenius norm 0.9. Each entry in the last two columns is the median absolute location error in time points over 40 trials.

Dimension   Learning   Evaluation   One direction   Two directions
       16      2,048        2,048             7              2.5
       32      2,048        2,048             9.5            4.5
       64      2,048        2,048            24.5            4.5
      128      2,048        2,048            41              9.5
      200      2,048        2,048           618.5          472
      200      4,096        4,096            12              6
      200      8,192        8,192             9.5            3.5

At dimension 200, with 4,096 observations for learning and another 4,096 for evaluation, including the second direction halved the median error from 12 to 6. On those same 40 sequences, two directions gave a closer estimate in 24 trials, tied in 7, and gave a larger error in 9. This comparison isolates the choice of rank within a common procedure.

AuthorsOpenReview ↗

Question 1. Assumptions, sample size and comparison with covariance projection

Theorem 3.3 treats independent observations in [−1,1]d[-1,1]^d with K≥1K\geq1 changes. Within each segment the distribution is constant, with density at most LL relative to uniform probability on the cube. Coordinates within an observation may be dependent.

We measure change kk by DkD_k, the difference between the mean feature matrices after and before the change. The symmetric feature matrix has H00=1H_{00}=1, H0j=3/2 XjH_{0j}=\sqrt{3/2}\,X_j, Hjj=5(3Xj2−1)/2H_{jj}=\sqrt{5}(3X_j^2-1)/2, and Hij=3XiXj/2H_{ij}=3X_iX_j/\sqrt{2} for 1≤i<j≤d1\leq i<j\leq d. Assume each DkD_k has rank at most rr. Its Frobenius norm κk\kappa_k is the jump size, and κ=min⁡kκk>0\kappa=\min_k\kappa_k>0. A coefficient matrix VV of unit Frobenius norm defines the scalar ⟨H(X),V⟩F\langle H(X),V\rangle_F, whose mean changes by ⟨Dk,V⟩F\langle D_k,V\rangle_F.

Split the data into T=⌊n/2⌋T=\lfloor n/2\rfloor pairs. The odd observations learn the directions and the even observations locate the changes. Write ΔU\Delta_U for the number of odd observations in the shortest segment, including the end segments. Choose a minimum scan window mm that is a multiple of four with m≤ΔU/8m\leq\Delta_U/8. Two requirements determine how much data is sufficient.

  1. To learn the directions, require ΔUκ2>6272rR2\Delta_U\kappa^2>6272rR^2 and use threshold 6r R6\sqrt r\,R. One conservative pilot noise bound is R=2L(1+d/2)ℓ+4d3/m ℓR=\sqrt{2L(1+d/2)\ell}+4d\sqrt{3/m}\,\ell, where δ∈(0,1)\delta\in(0,1) is the failure probability, ℓ=log⁡(4(d+1)N/δ)\ell=\log(4(d+1)N/\delta) and NN counts the candidate window and split pairs, including expanded windows. This count is at most 6T(1+⌊log⁡2(T/m)⌋)6T(1+\lfloor\log_2(T/m)\rfloor).
  2. To estimate the scalar means on each side, it suffices that mm exceed a universal constant times (L/κ2+1/κ)log⁡(16K/δ)(L/\kappa^2+1/\kappa)\log(16K/\delta).

Together these conditions guarantee, with probability at least 1−δ1-\delta, that the procedure finds exactly KK changes and retains at least κk/2\kappa_k/2 of each jump. The location error at change kk is at most a universal constant times 1+(L/κk2+1/κk)log⁡(8K/δ)1+(L/\kappa_k^2+1/\kappa_k)\log(8K/\delta) time points.

Wang, Yu and Rinaldo’s WBSIP also learns projections on separate data. For centered sub-Gaussian observations it measures a covariance change of any rank by its largest absolute eigenvalue. Our method targets low rank changes in the normalized moment matrix defined above. Its best single direction captures the largest absolute eigenvalue of DkD_k, while the best rank rr contrast captures the square root of the sum of its squared eigenvalues. Both methods give inverse squared dependence on their respective jump sizes for small jumps. The improvement here is that several components, even with opposite signs, can increase the scalar jump used for localization. The first sample requirement accounts for learning them. Table 1 compares the two ranks within our procedure. Wang, Yu and Rinaldo (2021)

For the revision, we propose opening with the two group example, defining the two samples before the algorithm, stating the learning and localization requirements together before the proofs, and leading the experiments with the comparison at equal sample sizes.

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

The paper addresses the problem of high dimensional change point localization. In particular, they propose a solution to aggregate dense interaction changes into one scalar quadratic contrast: under the condition of a sufficiently strong pilot signal, they guarantee the retention of at least half of the Frobenius jump.

The experimental evaluation is performed on synthetic datasets

Critical strengths
  1. The problem of high dimensional change point localization is very realistic and of practical interest.

  2. The proposed low-rank contrast aggregates many interaction changes of both signs into a single scalar contrast, which captures changes that are missed by entrywise or single-direction projection methods.

Critical weaknesses
  1. some critical parts of the methodology are not clearly described and the notation happens to be used before being introduced. Additionally, it is not clear how a sufficiently strong pilot is selected and if this assumption is realistic.

  2. the experimental setting appears limited: the tests are almost totally performed on a specifically tailored synthetic dataset. The only test on a real-world dataset (UCI HAR) is presented very shortly at the end of the experimental section and, from the description in the Appendix, appears semi-synthetic. In addition, no figures on the experimental findings (e.g. localization of changes) are reported.

  3. the Related Work section is very short and omits other important parts of the change point literature.

Questions for authors
  1. could you add pseudocode and better state the pilot and anchor condition from the beginning?

  2. the almost totality of the tests are performed on a synthetic dataset characterized by an exact rank-two jump. How does the proposed method behave on general low-rank/full rank and different types of changes?

  3. can you evaluate the proposed method on real sequences with natural change points? The HAR experiments construct synthetically the change points with resampling.

  4. the method does not win at d=200. Why does this happen? What happens when you increase n at different dimensions?

  5. why the proposed method is not compared to any existing works in the literature, such as Wang et al. (2021)? Could you add this (or other state-of-art solutions) in the experimental evaluation?

  6. what is the computational cost of the proposed method? Can you provide an analysis with respect to baseline and competitors?

9 comments
AuthorsOpenReview ↗

Weakness 1. How the pilot is chosen and what makes it informative

The split into pilot and evaluation observations is fixed before looking at the data. For independent observations, the odd observations form the pilot and the even observations are reserved for localization. A pilot window is a candidate time interval on the odd observations. A contrast is a weighted sum of first and second moments learned on that interval. The two anchors are blocks of evaluation observations on either side of an isolated change. They estimate the two scalar means.

“Sufficiently strong” describes the signal available in a segment. The algorithm scans a fixed grid of windows at several lengths and selects the shortest eligible windows whose scores exceed the threshold. In the unknown count experiment in Section 4.4, that threshold was calibrated on 80 separate null sequences before testing. Under the stated rank bound, for a segment long enough relative to the jump size and matrix noise, the learned contrast retains at least half of the population jump. Theorem 3.3 also requires enough observations in the two anchors to estimate the means.

We can check what this means empirically when the true direction is known. At d=200d=200 and the same jump size, the learned direction retained 73% of the jump on average at n=8192n=8192 and 87% at n=16384n=16384 (Section 4.1, 40 trials each). These are observed signal fractions. The theorem gives a sufficient condition for retention, and the scan itself uses observable scores.

AuthorsOpenReview ↗

Weakness 2. Localization errors in the activity and synthetic experiments

The activity experiment is a controlled transition between two distributions of measured smartphone features. We select the 128 coordinates using training data, then draw the two segments from separate test pools for walking and walking downstairs. A random sign applied to each whole observation removes its population mean while preserving every pairwise product. The switch between pools gives a known reference location against which to measure error.

Here is the distribution of absolute localization errors for the rank ten contrast in all 40 activity trials, in units of observations.

Absolute error     0     1     2 to 3     4 to 7     8 or more
Number of trials   8    15          8          8             1

The median error was 1 observation. The median was also 1 for the full quadratic scan, 2 for the kernel feature scan and 291.5 for the mean scan. The activity comparison therefore shows how much localization information survives the removal of mean differences in these measured features.

For the synthetic experiments, Figure 2 (p. 7) plots location error and signal retention as dimension and sample size change. Figure 3 (p. 8) varies the jump size, Figure 4 (p. 9) shows the effect of temporal dependence, and Figure 5 (p. 31) compares the rank choices. Table 2 gives the activity error and its interquartile range.

AuthorsOpenReview ↗

Weakness 3. Connections to segmentation and distributional change detection

Thank you for pointing out the need for a broader comparison. Global segmentation and distributional testing provide two useful points of reference. Killick, Fearnhead and Eckley (2012) use pruning to minimize a penalized sum of segment costs in PELT. This addresses the search over segmentations once a cost has been specified. Our search uses seeded intervals and shortest significant windows. The statistic on those windows estimates which interactions changed. PELT paper

Matteson and James (2014) use energy distances to detect general multivariate distributional changes. Our target is the part of a change visible in first and second moments. For that target, the Frobenius norm of the moment jump quantifies the signal, and the learned scalar contrast lets us relate the retained signal to location error. Energy distance paper

Among projection methods, Wang and Samworth (2018) target sparse mean changes, while Wang, Yu and Rinaldo (2021) use independent projections for covariance changes. Our dense interaction model keeps all coordinate means and variances fixed and combines several changing spectral components. This places our contribution at the intersection of projection learning and multiscale localization.

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

This paper studies change point localization in high-dimensional data, particularly when the dependence structure changes while individual coordinate means and variances remain unchanged. The authors propose learning a low-rank quadratic contrast through a matrix CUSUM procedure, which aggregates high-dimensional changes into a scalar mean shift for more accurate localization. The paper provides theoretical guarantees showing that, under certain low-rank and signal-strength assumptions, the learned contrast retains at least half of the original change signal and achieves a localization error scaling with the inverse squared jump strength. The analysis also covers several temporal dependence settings, with experiments on synthetic data and activity-derived distributions supporting the theoretical findings.

Critical strengths
  1. The paper addresses an interesting and challenging problem where changes occur in the interactions between variables rather than their individual means or variances. This setting is important because standard mean-based change detection methods may fail to capture such changes.
  2. The idea of learning a low-rank quadratic contrast to turn a high-dimensional change into a scalar mean shift is interesting. It allows the method to aggregate weak signals across many interactions while simplifying the final localization problem.
  3. The paper goes beyond independent observations and studies three temporal dependence settings, including geometric absolute regularity, Markov dependence, and physical dependence. This makes the theoretical analysis more comprehensive and relevant to time-series applications.
Critical weaknesses
  1. The main theoretical results rely on several fairly strong assumptions, including bounded densities, low-rank change matrices, and sufficiently strong signals. These conditions may not always hold in real high-dimensional data. Although the paper discusses approximately low-rank changes, the full end-to-end guarantees still require exact low rank. This limits how broadly the theoretical results can be applied. 2.Most experiments rely on the same synthetic distribution family with a rank-two change structure that closely matches the method's assumptions. Although the paper includes an experiment using UCI activity recognition data, the change points are artificially constructed through resampling and sign flipping. More diverse datasets with naturally occurring changes would provide stronger evidence of the method's practical effectiveness.
  2. The experimental comparisons mainly focus on full Frobenius, max-entry, mean, and kernel-feature scans. While these baselines help explain the behavior of different statistics, the evaluation lacks direct comparisons with stronger high-dimensional covariance change point methods or other structured projection approaches. In particular, some closely related methods discussed in the related work are not included in the experiments, making it harder to assess the improvement over existing approaches.
Questions for authors
  1. Most of the experiments use rank-two changes, and the main theoretical guarantees assume exact low rank. Could you show what happens with approximately low-rank or higher-rank changes?
  2. In practice, we usually don't know the true rank of the change matrix. How much does performance drop when the chosen rank is too small or too large? Is there a practical way to select it from the data?
7 comments
AuthorsOpenReview ↗

Weakness 1. When the localization guarantee applies

Theorem 3.3 allows any prespecified bound rr on the rank of the change between adjacent moment matrices. Each segment matrix can have full rank. For example, in Proposition G.1 every interaction between two groups of coordinates changes, yet their common pattern gives a jump of rank two. The joint density in that example is bounded by 1+∣θ∣1+|\theta| relative to product uniform measure, independently of the number of coordinates.

Two sample requirements enter the localization result. The pilot needs enough observations between changes to learn a direction that retains at least half the jump. The two outer blocks of independent evaluation observations must also be long enough to estimate the scalar means on either side. The density bound LL controls the variance of the learned contrast on those separate observations. Under independence, for fixed LL and small jumps, the location error scales with the inverse squared jump size up to a logarithmic factor in the number of changes and the confidence level. Section 4.1 and Table 1 illustrate the learning step at 200 dimensions. For the same dense jump, increasing the sample size from 8192 to 16384 raised the average fraction of signal retained from 73% to 87%, and median absolute location error fell from 6 to 3.5 observations (40 single change trials per setting).

For an approximately low rank jump, Proposition D.4 quantifies the effect of the discarded spectrum on the scan score. The guarantee for recovering all change points and refining their locations uses the stated rank bound.

AuthorsOpenReview ↗

Weakness 2. What the activity recordings contribute to the evaluation

The UCI experiment asks whether a quadratic contrast can locate a transition between two measured activity distributions after their coordinate means have been made identical. The interaction structure comes from smartphone measurements of walking and walking downstairs. We selected 128 features using the training split and sampled the two segments from separate activity pools in the test split. Applying a random sign to each entire observation gives both activities mean zero and preserves every product XiXjX_iX_j. The switch between pools supplies a reference time against which location error can be measured.

Over 40 sequences of length 1600, the median absolute location error was 1 observation for the rank ten contrast, 2 for the kernel feature scan and 291.5 for the mean scan (Section 4.4). The full quadratic scan also had median error 1. This comparison shows that quadratic information in the recorded activities sufficed to locate the controlled transition. The records supply the features and their interactions, while the resampling specifies the transition time. Localization of naturally timed changes in uninterrupted recordings remains a separate empirical question.

AuthorsOpenReview ↗

Weakness 3. What the comparison with structured projections establishes

Thank you for asking how the method compares with established covariance approaches. The closest covariance method in Section 1.1 is WBSIP (Wang, Yu and Rinaldo), which learns a projection from independent data and measures the covariance jump by its operator norm. Our contrast uses the Frobenius norm of the jump DD in normalized first and second moments. On the same matrix, a single spectral direction captures at most max⁡j∣λj∣\max_j|\lambda_j|, while the ideal quadratic contrast combines the signed directions to capture (∑jλj2)1/2(\sum_j\lambda_j^2)^{1/2}. Here the λj\lambda_j are the eigenvalues of DD. In the dense model of Proposition G.1, two eigenvalues of equal magnitude and opposite sign give an aggregate jump 2\sqrt{2} times that of the strongest single direction. The sparse projection of Wang and Samworth addresses a different target, a change in the mean vector, which stays fixed in this model.

Table 1 tests the effect of aggregating directions within a common localization procedure. At d=128d=128, n=4096n=4096 and jump size 0.90.9, median absolute errors over 40 single change trials were 41 observations for rank one, 9.5 for rank two and 140.5 for the full direction. The three estimators share the pilot observations, evaluation observations and scalar objective. These numbers quantify the effect of choosing the projection within our procedure. The comparison with WBSIP itself is analytical.

Reviewer_UD95OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies high-dimensional change point detection and localization, particularly when changes occur in the relationships among variables rather than their individual means or variances. The method combines the first- and second-order moments into a matrix and assumes that the difference between the matrices before and after each change point is low rank. Based on this assumption, the authors construct a matrix CUSUM statistic that uses only the largest rr eigenvalues in absolute value. A seeded segmentation procedure is then used to detect and isolate multiple change points. For more accurate localization, the method uses a quadratic score from the matrix CUSUM and applies a least-squares change point estimator to the resulting scalar sequence using separate observations.

The paper establishes high-probability guarantees for recovering the correct number of change points and obtaining localization errors of order O(κ−2)O(\kappa^{-2}) under suitable signal and spacing conditions. The theoretical analysis also covers several types of temporal dependence. Numerical experiments on synthetic and activity-derived data compare the proposed method with other change point detection methods and examine how the rank, signal strength, sample size, and temporal dependence affect localization accuracy.

Critical strengths
  1. By assuming that the change in the matrix of first- and second-order moments is low rank, the proposed statistic combines information from multiple changing directions while reducing noise from other directions. This can be useful when individual changes are weak but their combined signal is substantial.

  2. The paper establishes finite-sample guarantees for detecting and isolating an unknown number of change points, followed by refined localization with an error bound of order O(κ−2)O(\kappa^{-2}) under suitable conditions. The experiments also illustrate the effects of rank selection, sample size, and signal strength.

Critical weaknesses
  1. Restrictive assumptions. The theory requires bounded observations, uniformly bounded densities, and exactly low-rank moment changes. Moreover, the dimension-free localization rate depends on a dimension-dependent condition for learning the projection direction, limiting its applicability to general high-dimensional settings.

  2. Unnecessarily complicated presentation and overloaded notation. The core idea is simply to combine first- and second-order moments into a matrix and exploit the low rank of its change. However, concepts such as "orthogonal projection onto polynomials of total degree at most two" and "density projections" unnecessarily complicate the presentation and obscure the actual method. Also, Section 3 introduces numerous auxiliary quantities, such as bd=6db_d=6d, vd=L(1+d/2)v_d=L(1+d/2), and ℓ=log⁡(2p∣G∣/δ)\ell=\log(2p|\mathcal G|/\delta), which are further combined into λiid\lambda_{\mathrm{iid}} and other conditions. These nested definitions make the main theoretical results difficult to follow.

Questions for authors
  1. Is the proposed method essentially a weighted CUSUM of first- and second-order moments followed by low-rank truncation? What is the essential role of the degree-two density projection formulation, and what theoretical advantage does the proposed statistic offer over existing covariance change point methods?

  2. Can the authors clarify which common high-dimensional models satisfy the uniformly bounded density and low-rank change assumptions? In particular, how would the theoretical guarantees change if the density bound LL grows with the dimension?

  3. Can the authors express the main conditions of Theorem 3.3 more directly in terms of dimension, rank, sample size, change point spacing, and signal strength? This would help clarify when the claimed dimension-free localization rate is achievable.

  4. How should the working rank rr and detection threshold be selected in practice? Section 4.4 calibrates the threshold using independently generated null sequences, but such data may not be available in real applications. How sensitive are the results to these choices?

6 comments
AuthorsOpenReview ↗

Weakness1: Which part of localization depends on dimension?

Dimension enters when we learn which combination of moments changes. Once the learned contrast retains enough of the jump, the final localization step uses a scalar sequence whose noise is controlled by the density bound LL.

The two larger sample sizes in Table 1 illustrate this distinction at d=200d=200. We fixed the dense rank two jump at κ=0.9\kappa=0.9 and compared the learned direction with the true direction in 40 trials per setting, each with one change of unknown location. At n=8192n=8192 the learned direction retained 73% of the jump on average, and the median absolute location error was 6 observations. At n=16384n=16384 it retained 87%, and the median error was 3.5 observations. Using the true direction gave a median error of one observation at both sample sizes.

The assumptions enter at different steps. Bounded coordinates control the contribution of one observation to the scan. The joint density bound relative to product uniform measure controls the variance of every normalized quadratic contrast. The difference between adjacent moment matrices must have rank at most rr. Each segment matrix may have full rank. Theorem 3.3 gives the localization guarantee under independence and its signal and spacing conditions. For approximately low rank changes, Proposition D.4 bounds the truncation error in the scan score through the discarded spectral tail.

AuthorsOpenReview ↗

Weakness 2: The algorithm in terms of observable moments

Thank you for pointing out where the notation obscures the computation. For each candidate split, we compare the average first and second moments on its two sides. We arrange their differences in a matrix and score the split using the rr eigenvalues of largest magnitude. After isolating a change, we expand its window and estimate a quadratic contrast from the pilot observations in that window. Applying the contrast to separate evaluation observations gives one number per time point. We estimate the two means on the outer evaluation blocks, hold them fixed, and minimize the least squares loss over split locations in the central block.

All of these steps use observed moments. The density projection identifies exactly which part of a distributional change they measure and gives a common variance bound for the resulting scalar scores. The auxiliary quantities in Section 3 bound scan noise. The size of one matrix observation, its variance, and the number of candidate splits are the three inputs to that bound. Theorem 3.3 then asks whether a segment contains enough signal to learn the contrast and enough observations on either side to estimate its two scalar means.

AuthorsOpenReview ↗

Question 1: Which signal does the quadratic contrast collect?

Yes. The scan is a CUSUM of normalized first and second moments with spectral truncation. The degree two projection tells us what this computation measures. If DD is the jump in the moment matrix, Proposition A.1 identifies ∥D∥F\|D\|_F with the L2L^2 norm of the change in density visible to polynomials of degree at most two. The normalization also bounds the variance of every unit quadratic contrast by the density bound LL.

The distinction from the covariance method WBSIP of Wang, Yu and Rinaldo is the signal being used. WBSIP measures the covariance jump by its operator norm. On the same moment matrix DD, the strongest single spectral direction captures max⁡j∣λj∣\max_j|\lambda_j|, where the λj\lambda_j are the eigenvalues of DD. Our ideal contrast weights each component by its signed eigenvalue divided by ∥D∥F\|D\|_F. The resulting mean jump is ∥D∥F=(∑jλj2)1/2\|D\|_F=(\sum_j\lambda_j^2)^{1/2}. In the interaction model of Proposition G.1, the two nonzero eigenvalues have opposite signs and equal magnitude. Their contributions add, giving a scalar jump 2\sqrt{2} times as large as the strongest single component. Theorem 3.3 shows when the learned contrast retains at least half of the aggregate jump. This is the benefit of pooling several changing directions under the stated learning condition.

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

Related papers

Discussion (0)

Sign in to comment.