- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
this work studies distributional changes that keep means and covariances unchanged but are visible through cubic features and developed spectral compression guarantees that connect tensor approximation error. A major concern is if this paper is the right fit for ICLR and if ICLR audience would be interested in it or able to judge it.
Critical strengths
this work follows the typical PDE methodology that works on spectral approximation error. The dependence analysis and dense cubic construction provide a concrete regime where exploiting spectral structure gives stronger sufficient conditions.
Critical weaknesses
- Strong assumptions are required for the framework, which makes practical usage uncertain.
- no empirical evidence except synthetic ones
Questions for authors
- Could you include a real high-dim sequence or a broader benchmark against the current kernel or learned change-point methods under comparable assumptions?
- Do authors think if ICLR audience would be interested in this work or able to judge this work?
4 comments
Weakness 2: Learning a change direction across all 200 coordinates
The additional empirical evidence comes from controlled simulations. Knowing the true cubic jump lets us measure how much of it a learned direction preserves. In the 200 dimensional scenario with four dominant unfolding directions, all 200 coordinates participate in the change while the mean and covariance stay fixed. We generated 20 independent sequences of 12,000 independent observations each. The SVD and untruncated cubic fits share the fitting window chosen by the raw unfolding operator norm, the folds and the validation permutations.
SVD selects rank four in 50 of the 60 fold rotations. Its fitted direction retains 30.3% of the population cubic jump on average, compared with 6.5% for the untruncated fit. Here retention is the jump in the mean of the fitted and oriented scalar feature divided by the norm of the full cubic jump. SVD detects 19 of the 20 changes, and the untruncated fit detects 16. On the same 16 sequences detected by both, the median absolute localization errors are 21 and 150 observations, respectively. With the window and inference procedure shared, this comparison shows how spectral fitting can recover a useful direction when the shift spans all coordinates.
The experiment also evaluates symmetric HOSVD (Tucker), which projects all three modes onto a common leading singular subspace. Using the same fitting windows, this candidate retains 69.7% of the population jump on average in this scenario.
Question 1:A matched comparison with Gaussian kernel features
The high dimensional comparison we can report uses simulated sequences and a fixed Gaussian random Fourier feature (RFF) projection baseline. It extends the manuscript's comparison to 100 and 200 variables. Each sequence consists of 12,000 independent observations. The table reports the compressible cases from this study, with one or four dominant cubic unfolding directions. The cubic jump has the same norm across these settings, and each change preserves the mean and covariance.
The comparator uses 256 random Fourier features with squared bandwidth fixed at one third of the input dimension. It places its window using the norm of the feature contrast and uses the normalized local contrast to project validation observations to a scalar. Spectral SVD uses the raw unfolding operator norm to place its window. Both methods use the same sequences, interlaced folds, window length, 199 validation permutations per rotation and nominal 5% level after Bonferroni correction.
Each entry below counts rejections out of 20 independent sequences. The matched no change sequences retain the distribution before the change throughout.
Change detected False alarms
Dimension Spectrum SVD RFF proj. SVD RFF proj.
100 1 dir. 20/20 1/20 1/20 2/20
100 4 dir. 20/20 1/20 0/20 0/20
200 1 dir. 20/20 0/20 0/20 2/20
200 4 dir. 19/20 1/20 0/20 2/20
Spectral SVD detects more changes than this RFF projection baseline in all four compressible settings under the same nominal calibration.
Question 2: Learning a representation for changes in interactions
Yes. In a sensor system, the expected product of three readings can change while the individual means and pairwise covariances stay fixed. With many sensors, the changing interactions may be unknown and dispersed across many products. We learn a scalar feature that combines these interactions from unlabelled observations and test it on held out data.
The representation has a concrete learning objective. Its mean should change enough between the two segments to support calibrated detection and accurate localization. The paper quantifies the tradeoff between the population signal discarded by compression and the estimation noise in the fitted contrast. Readers can assess the resulting representation by the retained population jump, false alarms after direction selection, and the error in locating the change. These quantities make the role of representation learning explicit in the downstream inference task
Weakness 1: Nonparametric detection and localization with spectral compression
Our procedure learns a scalar change score from high dimensional vector observations without specifying parametric distributions for the two segments. We form cumulative sum (CUSUM) contrasts of Legendre features of total degree at most three and represent them as third order tensors. This makes cubic interactions available to the detector even when the means and covariance stay fixed. Truncating the tensor unfolding gives a fitted direction for the scalar score. In simulations with 12,000 observations in 100 and 200 dimensions, the implementation scans every admissible split within each pilot fold without storing a feature matrix for all observations. At 200 dimensions the representation includes 1,373,700 nonconstant features.
The theory connects window selection, direction learning, detection, and localization. An exact second moment calculation under the reference measure gives an unfolding variance bound for the correlated polynomial features. Matrix Bernstein controls the resulting noise uniformly over candidate windows. Theorem 3.1 converts the discarded singular spectrum and the rank dependent fitting noise into a lower bound on the retained jump. In Theorem 4.1 a common spectral condition supports both placement of the normalized pilot window and fitting an informative direction. With the stated validation and anchor conditions, the retained jump controls detection and the leading localization error bound scales as its inverse square in the small signal regime at fixed confidence and bounded density.
Power and localization require adequate segment lengths, a density bound relative to the uniform measure on the input cube, and a change visible in the polynomial features with enough signal relative to estimation noise. The spectral condition concerns that jump and allows approximate low rank and tied singular values. The coordinates within each observation may be dependent. When observations are independent and identically distributed under the null, the permutation test controls false alarms in finite samples without a density bound or knowledge of the true rank. Each permutation repeats the maximum over the fitted directions on independent validation data (Theorem D.1). For temporally dependent data, Section 5 derives the unfolding noise bound under geometric mixing, retaining the independent dimension scaling in the leading variance term and accounting for a separate logarithmic correction. Its separate calibration result uses a predetermined grid and deducts a justified mixing error from the test level.