acceptodds
Desk rejected at ICLR 2027

Spectral Compression for Cubic Change Detection

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

Abstract

Changes in third-order interactions can preserve means and covariances. A change spread across many weak cubic coefficients may nevertheless concentrate in a few singular directions. We study when these directions can be learned and used for inference. For bounded marginal densities, unfolding SVD separates a spectral tail from estimation noise weighted by rank. A geometric inequality converts this error into a retained scalar jump, and a common spectral-energy condition supplies both an initial window and an informative direction. Under geometric absolute regularity, we derive the cubic matrix covariance scale and obtain a uniform concentration bound with an improved leading term. A valid dense dependent model exhibits a strict separation between the resulting sufficient conditions and those from residue-class coupling, including coefficient alternatives. For independent observations, validation and anchor conditions yield a calibrated test and a leading inverse-squared-jump localization bound in the small-signal regime. Related guarantees under dependence follow through analytic scanning and a separate joint-law comparison after thinning. Numerical illustrations examine the spectral construction and selected signal.

Primary area:
learning theory
Keywords:
Change-point detection, High-dimensional statistics, Nonparametric methods, Low-rank tensors, Higher-order statistics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_31CXOpenReview ↗
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
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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

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

The paper studies change-point detection for cubic interactions that can leave means and covariances unchanged. It links unfolding-SVD approximation error to the scalar jump retained by a learned projection, with a common spectral-energy condition supporting both initial window selection and direction estimation. For independent observations, a three-fold procedure provides permutation calibrated testing, and a localization bound whose leading small-signal term is inverse-squared in the retained jump. The paper derives covariance-based concentration bounds and constructs a dense alternative separating the resulting sufficient conditions from residue-class coupling bounds. Calibrated refinement under dependence is handled separately through temporal thinning. Simulation results support the proposed methods.

Critical strengths
  1. A common spectral condition connects pilot localization, direction learning, and subsequent inference.
  2. The dense cubic construction is statistically valid and demonstrates a meaningful consequence of the dependence analysis.
Critical weaknesses
  1. The novelty of the central spectral approximation result is insufficiently differentiated from prior work.

The paper does not adequately compare its "spectral tail plus noise" inequality with directly related gap-free SVD bounds, making the distinctive contribution of Theorem 3.1 difficult to assess. The common pilot fitting condition and dependence analysis are meaningful additions, but their significance relative to existing results is not established clearly enough.

  1. The experiments do not directly evaluate the paper’s most distinctive advantage.

The reported power experiments primarily use sparse signals and a coefficient based pilot. The dense construction is evaluated through population geometry, while the spectral pilot and temporal-dependence guarantees are developed analytically rather than tested in corresponding experiments.

Questions for authors

Except the weaknesses,

  1. How would the density and mixing bounds L,B,cL,B,c be specified in applications where they are not known a priori?

  2. For the dense construction, at what finite sample sizes and dimensions does the full explicit concentration threshold, including its envelope correction, certify detection?

  3. In the reported experiments, to what extent are missed detections attributable to pilot placement, direction estimation, and the validation threshold?

  4. Cubic change detection is a subset of distributional change detection, and both methods can preserve means and covariances. Could the authors clarify the distinctions or provide a detailed comparison between them?

8 comments
AuthorsOpenReview ↗

Weakness 1: From spectral approximation to a retained change signal

Theorem 3.1 uses the established gap free tail and noise tradeoff to control a quantity specific to change detection: the scalar population jump along a direction learned from a noisy, data selected window.

Write A=aD(1)A=aD_{(1)}, Z=E(1)Z=E_{(1)}, and let τr(A)\tau_r(A) be the Frobenius norm of the singular value tail of AA. With P^r\widehat P_r projecting onto the leading left singular vectors of A+ZA+Z, Drineas and Ipsen (2019, Corollary 2.4) give ∣(I−P^r)A∣F≤τr(A)+2∣Z∣F.\\|(I-\widehat P_r)A\\|_F\le\tau_r(A)+2\\|Z\\|_F. Kumar et al. (arXiv:2509.17382, Theorem 1) bound the actual noisy truncation by ∣(A+Z)r−A∣F≤(2+2)τr(A)+r,∣Z∣op.\\|(A+Z)_r-A\\|_F\le(2+\sqrt 2)\\{\tau_r(A)+\sqrt r\\,\\|Z\\|_{\rm op}\\}. These results establish the gap free matrix approximation step. Our fit folds and symmetrizes that matrix estimate. If its error is bounded by er=aτr(D)+22rδ<aκe_r=a\tau_r(D)+2\sqrt{2r}\delta<a\kappa, then its normalized direction retains at least κ1−er/(aκ)2.\kappa\sqrt{1-\\{e_r/(a\kappa)\\}^2}. Here DD is the cubic jump, κ=∣D∣F\kappa=\\|D\\|_F, aa is the local contrast amplitude, τr(D)\tau_r(D) is the unfolding tail, and δ\delta bounds the unfolding noise in operator norm. This statement controls the population jump even when the fitted direction varies across tied singular vectors.

The distinct role of the spectrum comes next. The common condition requires the clipped energy ∑jmin⁡σj(D(1))2,8ℓ2\sum_j\min\\{\sigma_j(D_{(1)})^2,8\ell^2\\} to be small relative to the jump energy, with a cutoff that also accounts for segment balance and window length. At normalized noise level ℓ\ell, this condition supplies both enough leading singular signal to place the pilot and a rank in the library with a large retained jump. A uniform noise event makes that comparison valid at the chosen window. Under independence, the validation permutations apply to the whole candidate maximum, and the retained jump sets the scale of the subsequent localization error. Under dependence, Lemma 5.1 preserves the q/pq/p matrix variance scale for weighted cubic sums, with p=d+1p=d+1 and q=(d+33)q=\binom{d+3}{3}, and the dependence bound supplies the noise level to the same pilot and fitting argument. Calibrated refinement under dependence uses the separate thinning comparison. The contribution is this connection between observable cubic contrasts, recoverable signal and inference, with the assumptions for each stage stated separately.

AuthorsOpenReview ↗

Weakness 2: Finite sample evidence for the dense spectral mechanism

The finite sample experiments now test whether a spectral direction learned from observations retains a dense cubic shift. In the paper's two frequency construction at d=16d=16, all coordinates participate and 816 of the 969 polynomial features carry cubic signal. With 192,000 independent observations per sequence, the two fitted families share the same data, raw unfolding operator pilot, folds and validation permutations. SVD retains 83.2% of the true jump on average and detects 40/40 sequences. Basis retains 4.6% and detects 9/40. The retained fraction is the signed population jump along the selected and oriented unit direction divided by the full jump norm.

A separate bounded factor experiment reaches the same question at d=200d=200. All coordinates contribute to a cubic jump with four dominant unfolding directions, and the mean and covariance remain constant across the change. At 12,000 independent observations per sequence and jump norm 2, SVD selects rank 4 in 50 of 60 cyclic orderings from 20 independent sequences. Under the shared raw spectral pilot and permutation protocol, SVD and Basis retain 30.3% and 0.14% and detect 19/20 and 1/20, respectively. The SVD library here uses dyadic ranks through 128 and the full rank 201. The observed rank choices and the retained fractions show how the population structure remains useful for fitting.

These experiments evaluate the dense signal and the spectral pilot under independent sampling. For temporal dependence, we evaluated the full analytic threshold for the refresh model, including the envelope correction. This gives concrete sufficient sample sizes in dimensions 16, 32 and 64, detailed in Question 2. These dependent results are detection certificates from the bound. The reported empirical rejection rates use independent observations.

AuthorsOpenReview ↗

Question 1: Specifying density and mixing envelopes in applications

The analytic dependent threshold uses upper envelopes for the sampling law. In an application we would set LL from a justified bound for the marginal densities relative to the uniform reference measure on the feature domain. For example, a specified bounded density model can give this bound directly. The two frequency construction gives L=3/2L=3/2 from the amplitude of the perturbation. Restricting the support alone does not give such an upper bound.

A model of the dependence can supply B,cB,c. For example, a Markov transition with a verified common refresh component of probability at least ϵ∈(0,1)\epsilon\in(0,1) at each step, including across a change, permits coupling at the first shared refresh. This gives βX(k)≤(1−ϵ)k\beta_X(k)\le(1-\epsilon)^k, so one can take B=1B=1 and c=−log⁡(1−ϵ)c=-\log(1-\epsilon). The refresh model in the paper gives the same bound with 1−ϵ1-\epsilon equal to its retention probability. The envelope must cover the entire time series, including the transition across the change.

Without justified envelopes, a threshold computed over a range of plausible values of L,B,cL,B,c is a sensitivity analysis. A point estimate of density or a decay rate inferred from sample autocorrelations does not certify the stated finite sample level. The independent validation permutation test requires neither LL nor mixing constants for its null size guarantee under independent, identically distributed observations. For dependent data, the separate thinning route can use any justified bound on βX(k)\beta_X(k) at its chosen spacing, with the comparison error subtracted from the testing budget. These are two concrete ways to use the method, depending on what the application can justify about sampling.

Reviewer_i7hmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies distributional changes visible through polynomial features of degree at most three, including changes that preserve means and covariances. It represents these features as a symmetric tensor and relates spectral approximation error to the signal retained by a learned projection. For independent observations, a three-fold procedure combines window and direction estimation, permutation testing, and scalar localization under explicit signal conditions. Under geometric absolute regularity, a cubic covariance bound yields an asymptotic detection guarantee for a dense construction where specified residue-class coupling bounds are inconclusive. The evidence consists of proofs, population calculations, and simulations with independent observations.

Critical strengths
  1. The paper connects approximation quality to a calibrated statistical procedure. Theorems 3.1 and 4.1 relate spectral tails to retained signal while accounting for an unknown fitting window and direction selection. The separation of fitting, validation, and anchored localization is careful. In particular, permuting the full library maximum gives finite-sample size control under an iid null without the bounded-density assumptions used in the signal analysis.

  2. Lemma 5.1 preserves the cubic unfolding's matrix-variance scale under temporal dependence. The dense construction provides valid bounded densities with unchanged lower-order moments, weak individual cubic coefficients, and an asymptotically rank-one unfolding. Proposition 5.3 uses this construction to separate the stated sufficient conditions for all sufficiently large dimensions. This is a useful structural and asymptotic result; a finite-sample performance advantage has not been established.

Critical weaknesses
  1. The dependence comparison leaves the added benefit of the BMY concentration route unclear. Lemma 5.1 also supports a blocking argument: control the variances of block sums, compare the odd and even block sequences with independent blocks, and apply matrix Bernstein. This removes the extra square-root block-length factor in the leading variance term, while retaining a block-envelope term and a coupling-failure probability. Proposition 5.3 remains valid for its stated residue-class comparator, and this alternative still uses the paper's covariance lemma. However, the comparison does not isolate the benefit of BMY over blocking with that same lemma.

    Section 5.1 identifies a real distinction: the BMY route avoids an additional sample-law comparison probability. The consequence of that distinction should be assessed at the same total failure probability. Finite-sample interpretation also needs care. In Appendix H.4's example, with L=3/2 and mixing bounded by 2−k2^{-k}, the leading thresholds at an equal tail parameter cross only above a coupling gap of approximately 767,000, or 2700 with the process-specific variance. These comparisons concern conservative upper bounds and omit envelope corrections; they do not show that either detector fails in practice. They do leave the usable range of the analytic improvement uncertain.

  2. The simulations do not establish a finite-sample spectral advantage in the dense regime motivating the method. The cubic-family comparisons use the coefficient pilot on independent data. The dense construction receives population calculations, while the additional rank-one simulation has only three active coordinates (Appendix I.3). In Table 1 at d=16, unfolding SVD detects 20%, 7.5%, and 2.5% of quadratic–linear, triple, and mixed changes, compared with 100%, 77.5%, and 82.5% for sparse basis. These sparse alternatives favor coefficient selection, so the results do not refute the dense-signal motivation. They leave unanswered whether the favorable spectral geometry survives estimation when many individually weak coefficients carry the change. A targeted comparison on a valid, suitably scaled dense construction would address this gap.

Questions for authors
  1. Is there a concrete finite-sample setting in the refresh model where the full covariance threshold is informative? Please compare it with the full residue-class threshold at the same overall failure probability, distinguishing the general and process-specific variance bounds and retaining the envelope terms.

  2. Section 5.1 emphasizes that BMY avoids an additional coupling-failure probability. Compared with odd/even block coupling using Lemma 5.1, what benefit does this provide after both methods allocate the same total failure probability? Are there relevant regimes where it changes the detection condition or the contrasts that can be covered?

  3. Are existing results available for the dense construction using the spectral pilot, with coefficient-based alternatives given comparable data and calibration? A suitably rescaled model at d=16 would retain the existing experiment's 969-feature dimension, making it a focused extension. Retained signal and rejection rates would be particularly informative.

6 comments
AuthorsOpenReview ↗

Weakness 1: What the covariance lemma and BMY each contribute

Lemma 5.1 supplies the leading variance gain for both BMY and the odd and even block construction you propose. Applying the lemma to the weighted sums within each block keeps the variance on each parity at the scale of its squared weights. Rectangular matrix Bernstein then gives a leading term of 2Vxb2V\sqrt{x_b} and a bounded summand term of 43B3b∣w∣∞xb\frac43 B_3 b\\|w\\|_\infty x_b, where ∣w∣2=1\\|w\\|_2=1, V2V^2 is the variance bound from Lemma 5.1 and B3=63qB_3=6\sqrt{3q} is the observation envelope. For a family of MM contrasts, xb=log⁡2M(p+p2)/(δ−εb)x_b=\log\\{2M(p+p^2)/(\delta-\varepsilon_b)\\}, where p=d+1p=d+1, q=(d+33)q=\binom{d+3}{3}, and εb=(⌈n/b⌉−2)+βX(b+1)\varepsilon_b=(\lceil n/b\rceil-2)_+\beta_X(b+1) bounds the total comparison cost of the two parity vectors. This cost is shared across the family.

At fixed geometric mixing and fixed δ\delta, taking bb of order log⁡n\log n gives the same leading variance order as BMY. Proposition 5.3 concerns the stated residue class bounds. The strengthened comparison shows that the variance gain also carries over to blocking when it uses Lemma 5.1.

BMY can allocate the full failure budget to the deviation bound. This changes the sufficient detection condition when the budget shrinks rapidly. In the refresh model with fixed dimension and jump size, n=4hn=4h, and a change at 2h2h, take all full interval cuts in [h,3h][h,3h] and their radius hh local contrasts. For δn=exp⁡(−na)\delta_n=\exp(-n^a) with 1/2<a<11/2<a<1, their maximum weight is at most 2/n2/\sqrt n. The complete BMY bound is o(n)o(\sqrt n). The block Bernstein threshold above requires b=Ω(na)b=\Omega(n^a) to meet the same budget, giving an envelope term of order at least n2a−1/2n^{2a-1/2}. The population peak grows as n\sqrt n, so the displayed BMY bound certifies detection in this regime.

For the fixed budget comparison, we evaluated the full thresholds with δ=0.05\delta=0.05, including the comparison costs and bounded summand terms. In the d=16d=16 refresh example with L=3/2L=3/2 and retention probability 1/21/2, the general BMY bound certifies detection at n=4×1014n=4\times10^{14}. This evaluates the analytic sufficient condition directly at a finite sample size.

AuthorsOpenReview ↗

Weakness 2: Learning a low rank direction for a dense cubic change

In a bounded factor experiment with four dominant unfolding directions and all 200 coordinates involved in the cubic change, the SVD family selected rank 4 in 50 of 60 cyclic orderings. These orderings come from 20 independent sequences of 12,000 observations each. The cubic jump has norm 2, with the mean and covariance held constant across the change.

SVD and Basis shared the data and raw spectral pilot, and calibrated both tests at nominal level 0.05 with 199 permutations. The SVD grid used dyadic ranks through 128 and the full rank 201. Basis used support sizes 1, 4 and 16. The average share of the true jump along the selected and oriented directions was 30.3% for SVD and 0.14% for Basis. The corresponding tests detected the change in 19/20 and 1/20 sequences.

We also compared SVD with the full cubic direction in a separate 200 dimensional case with one dominant unfolding direction, using the same sequence length and jump norm. SVD selected rank 1 in all 60 cyclic orderings. With the common spectral pilot, both methods detected the change in all 20 sequences. SVD retained 64.6% of the jump on average and had a median absolute localization error of 9 observations. The full fit retained 8.7% and had a median error of 120 observations.

AuthorsOpenReview ↗

Question 1: A finite refresh calculation at a common failure probability

Yes. In a valid refresh model at n=4×1014n=4\times10^{14}, the complete general and refresh specific BMY thresholds both satisfy the detection condition at total failure probability 0.050.05.

Set d=16d=16, p=17p=17, q=969q=969, L=3/2L=3/2, and retention probability ϱ=1/2\varrho=1/2. The two frequency densities f±=1±2(1+c16)−1g16f_\pm=1\pm\\{2(1+c_{16})\\}^{-1}g_{16} lie in [1/2,3/2][1/2,3/2], agree through degree two, and give a cubic jump norm of 0.108357250.10835725. Take n=4hn=4h with the change at 2h2h, and include every full interval cut in [h,3h][h,3h] and its radius hh local contrast. Thus M=4h+2M=4h+2 and ∣w∣∞≤h−1/2\\|w\\|_\infty\le h^{-1/2}. With c=log⁡2c=\log 2, B3=63qB_3=6\sqrt{3q}, and x=log⁡2M(p+p2)/0.05x=\log\\{2M(p+p^2)/0.05\\}, the closed uniform BMY upper bound is

+\frac{2B_3}{\sqrt h}\\{4\sqrt{x/c}+\gamma(c,n)x\\},\qquad \gamma(c,n)=\frac{\log n}{\log 2}\max\left\\{2,\frac{32\log n}{c\log 2}\right\\}.$$ We evaluate it with each variance input $$V_{\rm gen}^2=\frac qp\left\\{L+16K_4^2\sqrt{L}\frac{\sqrt{\varrho}}{1-\sqrt{\varrho}}\right\\}, \qquad V_{\rm ref}^2=\frac qp L\frac{1+\varrho}{1-\varrho}, \quad K_4^2=27.$$ For the stated residue class threshold, use $k=\lceil\log(2n/0.05)/c\rceil$ and assign $\zeta=0.05-(n-k)2^{-k}$ to the tail bound. Its complete threshold is $$U_{\rm res}=\sqrt{2L(q/p)k x_c}+\frac{2B_3 kx_c}{3\sqrt h}, \qquad x_c=\log\\{2Mk(p+p^2+2q)/\zeta\\}.$$ The original residue threshold includes the coefficient scans in its dimension factor. Its tail and law comparison allocations sum to 0.05, the same budget as the BMY bounds. ```text n BMY general BMY refresh Residue Population peak 4 x 10^14 75446.01 4759.95 676.84 1074524 ``` The population peak is $\sqrt h\\,\sigma_1(D)$, where $D$ is the cubic tensor jump and $\sigma_1(D)$ is the largest singular value of its unfolding. For any displayed threshold $U$, a peak above $2U$ yields size at most 0.05 and power at least 0.95. At this finite setting that inequality holds with both BMY variance inputs. With the refresh variance, the detection condition also holds at $n=4\times10^{12}$. The population peak is approximately 107,452, above twice the full BMY threshold of 6,223.95.

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

Related papers

Discussion (0)

Sign in to comment.