Sequential Change Point Detection via Denoising Score Matching

Liyan Xie; Shixiang Zhu; Wenbin Zhou; Zhigang Peng

REVIEW 4 major objections 5 minor 35 references

Sequential Change Point Detection via Denoising Score Matching

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proposes DSM-CUSUM, a change-point detector that uses denoising score matching to learn unknown distributions and proves worst-case detection delay bounds set by the Fisher divergence between pre- and post-change distributions.

desk verdict The WADD theorem has a factor-2 bug and the online proof has a measurability gap, but the DSM-CUSUM idea and the earthquake experiment are worth referee time. read the letter →

arxiv 2501.12667 v1 pith:K2GP6CTT submitted 2025-01-22 stat.ML cs.LG

Wenbin Zhou , Liyan Xie , Zhigang Peng , Shixiang Zhu This is my paper · ORCID

classification stat.MLcs.LG

keywords sequentialchange-pointdetectiondenoisingscorematchingCUSUMHyvärinenFisherdivergenceworst-caseaveragedelaynoiseinjectionscore-based

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a sequential change-point detection method, DSM-CUSUM, for streams whose pre- and post-change distributions are unknown and possibly high-dimensional. Instead of fitting densities, it trains two score functions, the gradients of the log-densities, with denoising score matching, injecting Gaussian noise to improve estimation in sparse regions, and builds a CUSUM statistic from the difference of their Hyvärinen scores. The paper proves that, provided the trained scores approximate the true perturbed scores within a small error, the worst-case average detection delay is bounded by a constant divided by the Fisher divergence between the two distributions, plus an error term. It argues that the injected noise creates a tradeoff in which the decrease in score-estimation error outweighs the bias from perturbing the distribution, so a moderate noise level lowers the overall error and speeds detection. The method is demonstrated on two synthetic problems and on geophysical signals preceding an earthquake, where it flags a precursor around mid-June 2014.

What carries the argument

The load-bearing machinery is the pair (denoising score matching, Hyvärinen-score CUSUM). Denoising score matching trains $\hat{s}_0$ and $\hat{s}_1$ by adding Gaussian noise and minimizing $\sum_{x_j \in D_i} \sum_{k=1}^K \|s(x_j + \epsilon_{jk}; \theta) + \epsilon_{jk}/\sigma^2\|_2^2$, which targets the true perturbed score $\nabla \log p_{i,\sigma}(x)$. The CUSUM statistic accumulates $\Delta(x) = H(x; \hat{s}_0) - H(x; \hat{s}_1)$, whose expectation is tied to the Fisher divergence by the identity in Lemma 2; Lemma 1 quantifies how estimation error $\epsilon_{\mathrm{est}}(\sigma)$ and perturbation error $\epsilon_{\mathrm{pert}}(\sigma)$ translate into a combined error $\epsilon(\sigma)$. Wald's identity (offline) and windowed random-walk overshoot estimates (online) turn that divergence-approximation result into worst-case detection delay bounds.

What would settle it

On a dataset with known ground-truth densities (for example, the paper's 2D Gaussian mixture), compute the squared error between the learned score and the true score of the noise-perturbed distribution on held-out samples for several noise levels and training efforts; if the error does not respect the assumed $\epsilon_{\mathrm{est}}(\sigma)$ with the stated high probability, or if the measured worst-case delay violates Theorem 1's bound, the assumed error model fails.

Watch

Extended reading notes

Core claim

At the paper's core is the claim that change-point detection can proceed without any density model: the score functions $\nabla \log p_0$ and $\nabla \log p_1$ are learned by denoising score matching, and the detection statistic is the accumulated difference of their Hyvärinen scores. Lemma 2 shows that in the ideal case the expected increment equals the Fisher divergence $DF(p_1\parallel p_0)$, and Lemma 1 bounds the error introduced by estimated and noise-perturbed scores as $\epsilon(\sigma)$. Theorem 1 then gives $WADD \le \tau/[DF(p_1\parallel p_0) - \epsilon(\sigma)] + \mathbb{E}_{x\sim p_1}[\Delta(x)^2]/[DF(p_1\parallel p_0) - \epsilon(\sigma)]^2$ for offline score estimates, while Theorem 2 gives a windowed analogue for online estimation. The paper also argues that injecting Gaussian noise improves score estimation in poorly sampled regions faster than it biases the perturbed distribution, so a moderate $\sigma$ reduces the overall error $\epsilon(\sigma)$ and improves detection delay. Real-data results on geophysical signals identify a precursor around mid-June 2014, consistent with domain experts.

Load-bearing premise

The entire delay bound is conditional on assuming the trained score functions stay within a small, fixed error of the true noise-perturbed scores, an error the paper does not derive from the training objective or network size.

Editorial extensions

If this is right

If the score errors stay below the Fisher divergence, the detector has finite worst-case delay that grows with the threshold and shrinks as the two distributions become more distinguishable.
The offline version has a tighter bound than the online one, but the online bound's extra terms vanish as the threshold increases, so the two become comparable under strict false-alarm control.
The noise scale $\sigma$ becomes a tuning lever: picking it to minimize the combined estimation-plus-perturbation error speeds detection, and the paper's evidence indicates a small positive noise level beats no noise.
Because the method needs only scores, it can be applied where densities are unavailable or normalizing constants are intractable, which covers many high-dimensional monitoring settings.
Thresholds can be set from pre-change data alone to meet an average-run-length constraint, so the detector does not need samples from the post-change distribution in the offline training phase.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, a constructive next step is to derive $\epsilon_{\mathrm{est}}(\sigma)$ and $\epsilon_{\mathrm{pert}}(\sigma)$ from the denoising score matching objective and network capacity, which would turn the conditional bound into a recipe for choosing $\sigma$.
Beyond the paper, the online proof assumes the increment at time $t$ is predictable from past observations, while Algorithm 1 updates the score model with the current observation before computing that increment; closing this gap is needed for the online WADD theorem to cover the implemented procedure.
A testable extension the paper leaves implicit is an adaptive noise scale that responds to local sample density, since score-estimation error concentrates in low-density regions while perturbation bias is also local, potentially yielding faster detection in heterogeneous streams.
If the bound is tight, the statistic itself estimates the Fisher divergence online, which could serve as a continuous change-severity measure in monitoring applications rather than only a trigger for an alarm.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know before you spend time on this: the central WADD bound has a factor-of-2 bug in Lemma 2, and the online proof has a measurability gap. Both are fixable, but the theorems as stated don't hold.

What is actually new: combining denoising score matching with the Hyvärinen-score CUSUM statistic is a legitimate extension of existing score-based change detection. Prior work either assumed known unnormalized densities or used likelihood-ratio networks; here scores are estimated via DSM, and the noise-injection tradeoff is a reasonable angle. The paper is honest about its limitations—the conclusion explicitly says the theory doesn't capture the noise-error relationship—and the earthquake precursor application is a nice, nontrivial testbed.

Where the soft spots are: Lemma 2 claims E_p[H(q)-H(p)] = DF(p||q), but integration by parts gives (1/2)DF. I checked this in a 1D Gaussian example; the factor is definitely 1/2. Since Lemma 1 uses that identity for the drift, the WADD bounds in Theorems 1 and 2 inherit the error. The fix is mechanical—multiply the statistic by 2 or redefine H—but the claim as written is wrong. The online proof also assumes ∆(x_t) is F_{t-1}-measurable, while the algorithm updates the post-change model on x_t before computing ∆(x_t). That is a real gap between proof and algorithm. Separately, Assumptions 1 and 2 are strong and not derived; the bounds are conditional on unverified constants. That's a limitation the authors acknowledge, not a fatal flaw. The experiments would benefit from error bars and a code release, but the comparisons against exact SCUSUM and the SM ablation are useful.

Who this is for: researchers in quickest change detection who want a nonparametric, score-based alternative to likelihood CUSUM. The empirical idea is worth exploring despite the theoretical issues.

Recommendation: send it to a serious referee. The core idea deserves scrutiny and the theorems are repairable, but the current version needs major revision before its theoretical claims can be trusted.

Referee Report

4 major / 5 minor

Summary. The paper proposes DSM-CUSUM, a score-based CUSUM procedure for sequential change-point detection. The pre- and post-change distributions are not modeled parametrically; instead, their score functions are estimated via denoising score matching with Gaussian noise injection. Two estimation modes are considered: offline, using reference data from both regimes, and online, where the post-change score is updated from a sliding window. The theoretical section states worst-case average detection delay bounds in terms of the Fisher divergence between the two distributions and of estimation/perturbation errors. The method is evaluated on two synthetic datasets and a real-world seismic precursor-detection task. The paper is clearly written and the algorithmic framework is interesting, but the main theoretical results contain a load-bearing factor-of-two error and a measurability gap between the online algorithm and the proof.

Significance. If the proof issues are repaired, this is a useful contribution: it extends score-based quickest detection (Wu et al., 2023) to data-driven settings in which the score functions are learned from data, and it formulates a clean error decomposition into estimation error and perturbation error. The paper is also unusually honest in its limitations section, which states that the precise noise-scale-to-error relationship is not captured by the assumptions. The real-data experiment on earthquake precursors is a valuable empirical addition. However, the central theoretical claim is currently not correct as written: Lemma 2 is off by a factor of two, and the proof of the online bound assumes a measurability property that the algorithm does not possess. These issues are fixable, but they affect every stated WADD bound, so the present version cannot be accepted without substantial revision.

major comments (4)

[Appendix A, Lemma 2] Lemma 2 is off by a factor of two. With H(x;s) = div s(x) + (1/2)||s(x)||^2, integration by parts gives E_{x∼p}[H(x;q) − H(x;p)] = (1/2)E_{x∼p}||∇log p(x) − ∇log q(x)||^2 = (1/2)DF(p||q), not DF(p||q). The proof appeals to an identity attributed to [31], but the displayed identity `DF(p||q) = E_p[(1/2)||∇log p||^2 + H(x;q)]` is itself missing a factor of 1/2 on the left. A concrete check: for p=N(0,1) and q=N(μ,1), the left side equals μ^2/2 while DF(p||q)=μ^2. Because Lemma 1 uses this identity for the term E_1[H(x;p0)−H(x;p1)], the error bound in Lemma 1 and the denominators in Theorems 1 and 2 are off by a factor of two. The fix is mechanical — for example, redefine the statistic as 2[H(x;s0)−H(x;s1)] — but all theorem statements need to be corrected; the bounds are incorrect even in the oracle case ϵ(σ)=0.
[Theorem 2 proof, Eq. (12)] The proof of Theorem 2 relies on the inequality E_1[∆(x_t)|F_{t−1}] ≥ DF(p1||p0) − ϵ(σ), which requires ∆(x_t) to be F_{t−1}-measurable. But in the online algorithm, Eq. (7) updates the post-change score model using the current observation x_t before the statistic ∆(x_t) is computed, as Algorithm 1 makes explicit. Hence ∆(x_t) depends on x_t through the updated model parameters and is not F_{t−1}-measurable. This is a gap between the algorithm and the proof. The proof can be repaired by updating the score model using only observations up to time t−1 before computing ∆(x_t), or by defining a different filtration, but as written the assertion in Eq. (12) does not hold for the stated procedure.
[Section IV, Remark and Fig. 1] The abstract and Section I claim that the paper demonstrates through theoretical analysis that denoising score matching enhances detection power by controlling the injected noise scale. The theorems, however, only bound WADD in terms of a generic error ϵ(σ); they do not establish that ϵ(σ) is minimized by a finite positive σ. The claim that ϵ_est decreases faster than ϵ_pert increases is supported only by the single simulation in Fig. 1, not by a theorem or a general derivation. This should be stated as an empirical observation or a conjecture, or supported by a proof, to match the strength of the paper's claims.
[Section IV, Assumption 1] Assumption 1 postulates the existence of error bounds ϵ_est(σ) and ϵ_pert(σ) but does not derive them from the DSM objective or from the capacity of the score network. The paper's own conclusion acknowledges this limitation. This is acceptable for a conditional result, but it means the central theoretical contribution is an implication of an assumed error model rather than a guarantee of the proposed DSM procedure. Since the constant ϵ(σ) is what ultimately determines the WADD bound, the theorem would be substantially stronger with a concrete derivation or a worked example showing that the assumed error bounds are achievable for the DSM estimator.

minor comments (5)

[Section II, Eq. (1)] In Eq. (1), the index on the left-hand side is t but the index in the description on the right is k; this should be made consistent (t = 1, ..., ν−1 and t = ν, ν+1, ...).
[Section IV, Assumption 1] The notation for norms in Assumption 1 is ambiguous: the displayed statements write `∥bs_i(x) − ∇log p_{i,σ}(x)∥_2 ≤ ϵ_est(σ)` while the proof in Appendix A uses `E_1∥bs_i(x) − ∇log p_{i,σ}(x)∥_2 ≤ ϵ_est(σ)`. Please clarify whether the object being bounded is the norm or the squared norm, and use a uniform notation throughout.
[Section IV, Theorem 1 and Theorem 2] The notation `p_∞` appears in the asymptotic remark after Theorem 1 and in the proof of Theorem 2 (e.g., `DF(p1||p_∞)+ϵ(σ)`); this appears to be a typo for `p_0` and should be corrected.
[Appendix A, Lemma 3] The intermediate inequality in the proof of Lemma 3, `(1/2)E[AB] ≤ (1/2)E[A]E[B] + (1/2)Cov(A,B)`, is not valid when the covariance is negative. The final bound can be obtained directly from E[AB] ≤ E[A]E[B] + sqrt(Var(A)Var(B)) combined with the factor 1/2, so the proof should be revised to skip the invalid intermediate step.
[Section V-A, Fig. 2] The claim that DSM-CUSUM significantly outperforms the baselines is based on a visual comparison of WADD-vs-ARL curves, but no error bars, standard deviations, or repeated-run variability are reported; adding such information would make the comparison more convincing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the WADD bounds are conditional error-propagation consequences of Assumptions 1-2, and the noise-injection benefit is simulation-based rather than a fitted prediction.

full rationale

The derivation chain is not circular. Theorem 1 and Theorem 2 follow from Lemma 1 (a derived bound on |E1[Δ] − DF| in terms of assumed ϵest, ϵpert, ϵdiv), Wald's identity, and the CUSUM stopping-time inequality WADD ≤ E1[T'] imported from [17]; none of these steps assumes the theorem it is used to prove. The claim that denoising score matching reduces total error is explicitly empirical: Section IV's Remark states "A key empirical observation is that ϵest typically decreases at a much faster rate than ϵpert increases," with Figure 1 supplying the simulation, and Section VI concedes "the preliminary nature of the theoretical assumptions, which do not explicitly capture the precise relationship between noise scale and error values," so the noise-benefit claim is not a fitted parameter renamed as a prediction. The one load-bearing self-citation is [17] (sharing author L. Xie) for the inequality WADD ≤ E1[T'] and an overshoot bound; this is an externally published result about window-limited CUSUM, not a restatement of the DSM-CUSUM bound, so under the review rules it is independent support rather than circularity. For completeness, non-circular correctness concerns remain: Lemma 2's factor-of-2 identity appears inconsistent with Definition 1 (the expectation equals (1/2)DF, not DF), and Theorem 2's proof assumes Δ(xt) is Ft−1-measurable while Algorithm 1 updates the score using xt in Eq. (7); these affect validity but do not make the argument circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central theoretical claims rest on two error-model assumptions that are asserted rather than derived, plus a prior inequality from the authors' own work. No new physical or modeling entities are introduced.

free parameters (4)

noise scale σ = 1.0 (used in experiments)
Central to the claimed noise tradeoff; hand-chosen, not derived or data-fitted. The theory only bounds the error given σ, it does not select it.
window size w = 10 (real data), not specified for synthetic
Controls online score update; chosen ad hoc.
learning rate η = not specified
Online gradient step in Eq. (7); value not reported, affecting reproducibility.
threshold τ = varies with target ARL, chosen by simulation (Appendix B-A)
Standard CPD tuning parameter; the paper uses an exponential approximation to set it.

assumptions (4)

ad hoc to paper Assumption 1: There exist constants δ, ϵ_est(σ), ϵ_pert(σ) such that with probability ≥1−δ the trained score models satisfy the stated approximation bounds.
This is the key enabling premise for all theorems; it is assumed rather than derived from the DSM training objective.
ad hoc to paper Assumption 2: Variance of the errors is bounded by C1·ϵ_est^2 and C2·ϵ_pert^2, and divergences approximate well, with bounded moments.
Same as above; supplies the moment control used in Lemma 1.
standard math Inequality WADD ≤ E1[T'] from Theorem 1 of [17] (Xie, Moustakides, Xie, 2023), with the overshoot bound from [34] (Lorden).
Used without proof as a prior result; note [17] shares an author with this paper, so this is a self-citation but a published peer-reviewed result.
domain assumption Gaussian kernel for noise injection K(x'|x) = N(x'; x, σ^2 I).
Simplifies the DSM objective and the perturbed-score analysis; other kernels are not explored.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Sequential Change Point Detection via Denoising Score Matching." pith.science (2026). https://pith.science/paper/K2GP6CTT

@misc{pith2026250112667,
  author       = {Pith},
  title        = {Pith review of: Sequential Change Point Detection via Denoising Score Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2GP6CTT}},
  note         = {Machine review of arXiv:2501.12667}
}

read the original abstract

Sequential change-point detection plays a critical role in numerous real-world applications, where timely identification of distributional shifts can greatly mitigate adverse outcomes. Classical methods commonly rely on parametric density assumptions of pre- and post-change distributions, limiting their effectiveness for high-dimensional, complex data streams. This paper proposes a score-based CUSUM change-point detection, in which the score functions of the data distribution are estimated by injecting noise and applying denoising score matching. We consider both offline and online versions of score estimation. Through theoretical analysis, we demonstrate that denoising score matching can enhance detection power by effectively controlling the injected noise scale. Finally, we validate the practical efficacy of our method through numerical experiments on two synthetic datasets and a real-world earthquake precursor detection task, demonstrating its effectiveness in challenging scenarios.

Figures

Figures reproduced from arXiv: 2501.12667 by the authors.

**Figure 3.** Geophysical signal datasets: comparison of our method (red star) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 5.** Visualization of the two datasets. The left column depicts [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 28 canonical work pages

[31]

Estimation of non-normalized statistical models by score matching

A. Hyvärinen and P. Dayan, “Estimation of non-normalized statistical models by score matching.” Journal of Machine Learning Research , vol. 6, no. 4, 2005

work page 2005
[1]

Quickest change detection,

V . V . Veeravalli and T. Banerjee, “Quickest change detection,”Academic Press Library in Signal Processing: Array and Statistical Signal Processing, vol. 3, pp. 209–256, 2013

work page 2013
[2]

Siegmund, Sequential Analysis: Tests and Confidence Intervals

D. Siegmund, Sequential Analysis: Tests and Confidence Intervals . Springer, 1985

work page 1985
[3]

Tartakovsky, I

A. Tartakovsky, I. Nikiforov, and M. Basseville, Sequential Analysis: Hypothesis Testing and Changepoint Detection . ser. Monographs on Statistics and Applied Probability 136. Boca Raton, London, New York: Chapman & Hall/CRC Press, Taylor & Francis Group, 2015

work page 2015
[4]

Change-point detection in the conditional correlation structure of multivariate volatility models,

M. Barassi, L. Horvath, and Y . Zhao, “Change-point detection in the conditional correlation structure of multivariate volatility models,”Journal of Business & Economic Statistics , vol. 38, no. 2, pp. 340–349, 2020

work page 2020
[5]

Sequential adversarial anomaly detection for one-class event data,

S. Zhu, H. S. Yuchi, M. Zhang, and Y . Xie, “Sequential adversarial anomaly detection for one-class event data,” INFORMS Journal on Data Science, vol. 2, no. 1, pp. 45–59, 2023

work page 2023
[6]

Sequential changepoint detection in quality control and dynamical systems,

T. L. Lai, “Sequential changepoint detection in quality control and dynamical systems,” Journal of the Royal Statistical Society: Series B (Methodological), vol. 57, no. 4, pp. 613–644, 1995

work page 1995
[7]

Early detection of COVID-19 hotspots using spatio-temporal data,

S. Zhu, A. Bukharin, L. Xie, K. Yamin, S. Yang, P. Keskinocak, and Y . Xie, “Early detection of COVID-19 hotspots using spatio-temporal data,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 2, pp. 250–260, 2022

work page 2022

Show all 35 references

[8]

Conditional generative modeling for high-dimensional marked temporal point processes,

Z. Dong, Z. Fan, and S. Zhu, “Conditional generative modeling for high-dimensional marked temporal point processes,” arXiv preprint arXiv:2305.12569, 2023

2023 arXiv
[9]

Machine learning and earthquake forecasting—next steps,

G. C. Beroza, M. Segou, and S. Mostafa Mousavi, “Machine learning and earthquake forecasting—next steps,” Nature communications, vol. 12, no. 1, p. 4761, 2021

2021
[10]

New opportunities to study earthquake precursors,

M. E. Pritchard, R. M. Allen, T. W. Becker, M. D. Behn, E. E. Brodsky, R. Bürgmann, C. Ebinger, J. T. Freymueller, M. Gerstenberger, B. Haines et al., “New opportunities to study earthquake precursors,” pp. 2444–2447, 2020

2020
[11]

Looking for earthquake precursors from space: A critical review,

P. Picozza, L. Conti, and A. Sotgiu, “Looking for earthquake precursors from space: A critical review,” Frontiers in Earth Science , vol. 9, p. 676775, 2021

2021
[12]

A critical review of ground based observations of earthquake precursors,

L. Conti, P. Picozza, and A. Sotgiu, “A critical review of ground based observations of earthquake precursors,” Frontiers in Earth Science, vol. 9, p. 676766, 2021

2021
[13]

Physical mechanisms of earthquake nucleation and foreshocks: cascade triggering, aseismic slip, or fluid flows?

Z. Peng and X. Lei, “Physical mechanisms of earthquake nucleation and foreshocks: cascade triggering, aseismic slip, or fluid flows?” Earthquake Research Advances, p. 100349, 2024

2024
[14]

Machine learning in earthquake seismology,

S. M. Mousavi and G. C. Beroza, “Machine learning in earthquake seismology,” Annual Review of Earth and Planetary Sciences , vol. 51, no. 1, pp. 105–129, 2023

2023
[15]

A unified data-adaptive framework for high dimensional change point detection,

B. Liu, C. Zhou, X. Zhang, and Y . Liu, “A unified data-adaptive framework for high dimensional change point detection,” Journal of the Royal Statistical Society Series B: Statistical Methodology , vol. 82, no. 4, pp. 933–963, 2020

2020
[16]

Information bounds and quick detection of parameter changes in stochastic systems,

T. L. Lai, “Information bounds and quick detection of parameter changes in stochastic systems,” IEEE Transactions on Information theory , vol. 44, no. 7, pp. 2917–2929, 1998

1998
[17]

Window-limited cusum for sequential change detection,

L. Xie, G. V . Moustakides, and Y . Xie, “Window-limited cusum for sequential change detection,” IEEE Transactions on Information Theory , vol. 69, no. 9, pp. 5990–6005, 2023

2023
[18]

M-statistic for kernel change-point detection,

S. Li, Y . Xie, H. Dai, and L. Song, “M-statistic for kernel change-point detection,” Advances in Neural Information Processing Systems , vol. 28, 2015

2015
[19]

Scan b-statistic for kernel change-point detection,

——, “Scan b-statistic for kernel change-point detection,” Sequential Analysis, vol. 38, no. 4, pp. 503–544, 2019

2019
[20]

Sliced score matching: A scalable approach to density and score estimation,

Y . Song, S. Garg, J. Shi, and S. Ermon, “Sliced score matching: A scalable approach to density and score estimation,” in Uncertainty in Artificial Intelligence, 2020, pp. 574–584

2020
[21]

A connection between score matching and denoising autoencoders,

P. Vincent, “A connection between score matching and denoising autoencoders,” Neural computation, vol. 23, no. 7, pp. 1661–1674, 2011

2011
[22]

Generative modeling by estimating gradients of the data distribution,

Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” Advances in neural information processing systems , vol. 32, 2019

2019
[23]

H. V . Poor and O. Hadjiliadis,Quickest Detection. Cambridge University Press, 2008

2008
[24]

Sequential (quickest) change detection: Classical results and new directions,

L. Xie, S. Zou, Y . Xie, and V . V . Veeravalli, “Sequential (quickest) change detection: Classical results and new directions,” IEEE Journal on Selected Areas in Information Theory , vol. 2, no. 2, pp. 494–514, 2021

2021
[25]

Continuous inspection schemes,

E. S. Page, “Continuous inspection schemes,” Biometrika, vol. 41, no. 1/2, pp. 100–115, 1954

1954
[26]

Quickest change detection with leave- one-out density estimation,

Y . Liang and V . V . Veeravalli, “Quickest change detection with leave- one-out density estimation,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

2023
[27]

Sequential change-point detection based on direct density-ratio estimation,

Y . Kawahara and M. Sugiyama, “Sequential change-point detection based on direct density-ratio estimation,” Statistical Analysis and Data Mining: The ASA Data Science Journal , vol. 5, no. 2, pp. 114–127, 2012

2012
[28]

Score-based quickest change detection for unnormalized models,

S. Wu, E. Diao, T. Banerjee, J. Ding, and V . Tarokh, “Score-based quickest change detection for unnormalized models,” in Proceedings of The 26th International Conference on Artificial Intelligence and Statistics , vol. 206. PMLR, 2023, pp. 10 546–10 565

2023
[29]

Training neural networks for likelihood/density ratio estimation,

G. V . Moustakides and K. Basioti, “Training neural networks for likelihood/density ratio estimation,” 2019. [Online]. Available: https://arxiv.org/abs/1911.00405

2019 arXiv
[30]

Training neural networks for sequential change-point detection,

J. Lee, Y . Xie, and X. Cheng, “Training neural networks for sequential change-point detection,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

2023
[32]

Score-based generative modeling through stochastic differential equations,

Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456 , 2020

2011 arXiv
[33]

Generative modeling with denoising auto-encoders and langevin sampling,

A. Block, Y . Mroueh, and A. Rakhlin, “Generative modeling with denoising auto-encoders and langevin sampling,” arXiv preprint arXiv:2002.00107, 2020

2002 arXiv
[34]

On excess over the boundary,

G. Lorden, “On excess over the boundary,” The Annals of Mathematical Statistics, vol. 41, no. 2, pp. 520–527, 1970

1970
[35]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in Neural Information Processing Systems , vol. 33, pp. 6840– 6851, 2020. APPENDIX A DETAILED PROOFS FOR SECTION IV We begin with introducing two useful technical lemmas (Lemma 2 and Lemma 3). ...

2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [31]

Estimation of non-normalized statistical models by score matching

A. Hyvärinen and P. Dayan, “Estimation of non-normalized statistical models by score matching.” Journal of Machine Learning Research , vol. 6, no. 4, 2005

work page 2005

[2] [1]

Quickest change detection,

V . V . Veeravalli and T. Banerjee, “Quickest change detection,”Academic Press Library in Signal Processing: Array and Statistical Signal Processing, vol. 3, pp. 209–256, 2013

work page 2013

[3] [2]

Siegmund, Sequential Analysis: Tests and Confidence Intervals

D. Siegmund, Sequential Analysis: Tests and Confidence Intervals . Springer, 1985

work page 1985

[4] [3]

Tartakovsky, I

A. Tartakovsky, I. Nikiforov, and M. Basseville, Sequential Analysis: Hypothesis Testing and Changepoint Detection . ser. Monographs on Statistics and Applied Probability 136. Boca Raton, London, New York: Chapman & Hall/CRC Press, Taylor & Francis Group, 2015

work page 2015

[5] [4]

Change-point detection in the conditional correlation structure of multivariate volatility models,

M. Barassi, L. Horvath, and Y . Zhao, “Change-point detection in the conditional correlation structure of multivariate volatility models,”Journal of Business & Economic Statistics , vol. 38, no. 2, pp. 340–349, 2020

work page 2020

[6] [5]

Sequential adversarial anomaly detection for one-class event data,

S. Zhu, H. S. Yuchi, M. Zhang, and Y . Xie, “Sequential adversarial anomaly detection for one-class event data,” INFORMS Journal on Data Science, vol. 2, no. 1, pp. 45–59, 2023

work page 2023

[7] [6]

Sequential changepoint detection in quality control and dynamical systems,

T. L. Lai, “Sequential changepoint detection in quality control and dynamical systems,” Journal of the Royal Statistical Society: Series B (Methodological), vol. 57, no. 4, pp. 613–644, 1995

work page 1995

[8] [7]

Early detection of COVID-19 hotspots using spatio-temporal data,

S. Zhu, A. Bukharin, L. Xie, K. Yamin, S. Yang, P. Keskinocak, and Y . Xie, “Early detection of COVID-19 hotspots using spatio-temporal data,” IEEE Journal of Selected Topics in Signal Processing , vol. 16, no. 2, pp. 250–260, 2022

work page 2022