A Generalized Graph Signal Processing Framework for Multiple Hypothesis Testing over Networks

Abdelhak M. Zoubir; Feng Ji; Martin G\"olz; Wee Peng Tay; Xingchao Jian

REVIEW 3 major objections 5 minor 21 references

A Generalized Graph Signal Processing Framework for Multiple Hypothesis Testing over Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Multiple hypothesis testing over a graph–time domain can be driven by a generalized graph signal that varies continuously, and the paper proposes an empirical-Bayes strategy that it claims controls false discovery rate asymptotically…

desk verdict The GGSP parameterization of continuously varying p-value mixtures is a genuine step beyond homogeneous or finite-state MHT, but the FDR-control theorem is delegated to a companion paper, and the simulations are too thin to support the 'best power' claim. read the letter →

arxiv 2506.03496 v1 pith:PZJJF2PE submitted 2025-06-04 eess.SP cs.ITmath.IT

Xingchao Jian , Martin Gölz , Feng Ji , Wee Peng Tay , Abdelhak M. Zoubir This is my paper

classification eess.SPcs.ITmath.IT MSC 62F0362F1262C1094A12

keywords multiplehypothesistestingfalsediscoveryrategeneralizedgraphsignalprocessingempiricalBayeslocalsensornetworksseismicdetectiongraph-timesignals

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

The paper tries to establish that multiple hypothesis testing over a joint graph–time domain can be made more powerful by letting both the probability that a hypothesis is null and the distribution of p-values under the alternative vary continuously across the domain. It models this inhomogeneity as a generalized graph signal — a function on the product of a graph's vertices and a measure space — and estimates that signal by maximum likelihood under a bandlimited assumption. Substituting the estimate into the oracle local false discovery rate rule, the paper claims asymptotic false discovery rate control and reports the best detection power among compared methods on a seismic sensor network task. The payoff would be a principled way to borrow strength from nearby nodes and nearby time points without forcing all hypotheses to share one p-value distribution.

What carries the argument

The central object is the bandlimited generalized graph signal $\gamma(v,t)=\sum_{k_1=1}^{K_1}\sum_{k_2=1}^{K_2}\xi_{k_1,k_2}\varphi_{k_1}(v)\psi_{k_2}(t)$, defined on the Cartesian product of the graph vertex set and a measure space via the graph Fourier basis and an orthonormal basis of $L^2(T)$. The load-bearing identity is the mixture model $f_{\mathrm{mix}}(p\mid\gamma(v,t))=\pi_0\circ\gamma(v,t)\,f_0(p;(v,t))+(1-\pi_0\circ\gamma(v,t))\,f_1(p;\gamma(v,t))$, from which the lfdr is computed as $\pi_0 f_0/f_{\mathrm{mix}}$. The machinery works by estimating the coefficient matrix $\Xi$ with the MLE, obtaining a uniformly consistent estimate of $\gamma$ and hence of the lfdr surface, and then thresholding the estimated lfdr at a level $\eta$ chosen to keep the estimated marginal FDR below the nominal level $\alpha$.

What would settle it

Generate synthetic network-time data with known ground truth where the null prior $\pi_0(v,t)$ is constant while the alternative p-value density varies independently, with a signal that is not exactly bandlimited; apply MHT-GGSP with the paper's one-parameter family and BIC-selected $(K_1,K_2)$. If the empirical FDR over many replicates exceeds the nominal level $\alpha$ beyond sampling error, or the estimated lfdr is far from the true lfdr, the central claim fails for this misspecified family.

Watch

Extended reading notes

Core claim

The paper extends the two-groups empirical-Bayes model to the joint domain $V\times T$, where each sample point $(v,t)$ carries its own hypothesis, and defines a random generalized graph signal $\gamma(v,t)$ that parameterizes both the null prior probability $\pi_0\circ\gamma(v,t)$ and the alternative p-value density $f_1(\cdot;\gamma(v,t))$. Under identifiability assumptions (Assumption 1), the local false discovery rate $\mathrm{lfdr}(p;\gamma(v,t))$ is identified from the mixture density. The paper proves that the optimal thresholding rule rejects exactly those hypotheses whose lfdr is below a common level $\eta$ (Theorem 1, a modification of a result proved in the companion paper), and proposes the strategy MHT-GGSP: estimate the Fourier coefficients $\Xi$ of $\gamma$ by maximum likelihood under the bandlimited assumption (Assumption 2), plug the estimate into the oracle rule, and select bandwidths $K_1,K_2$ by BIC. Theorem 2 gives uniform consistency of the MLE estimator under mild conditions, so the estimated lfdr converges to the true lfdr, and the paper asserts asymptotic FDR control. In the seismic experiment against BH, lfdr-sMoM, FDR-smoothing, SABHA, and AdaPT, MHT-GGSP is reported to achieve the best empirical power while controlling FDR.

Load-bearing premise

Everything rests on the assumption that a single scalar $\gamma$ at each point correctly captures both the probability that the null hypothesis is true and the shape of the alternative p-value distribution, and that $\gamma$ is exactly bandlimited with finite bandwidths $K_1$ and $K_2$; if the true data-generating process has prior and alternative density varying independently, or has spectral content beyond the assumed bandwidths, the estimated lfdr does not converge to the true lfdr and asymptotic FDR control is not guaranteed.

Editorial extensions

If this is right

If the central claim is right, any detection problem over a sensor network with a time axis can be addressed with a continuously varying null prior and alternative density, rather than a small number of discrete mixture classes.
The method yields an explicit recipe: choose the graph shift operator and temporal basis, run MLE, select bandwidths by BIC, and threshold the estimated lfdr; the FDR is then asymptotically controlled at any nominal level $\alpha$.
The oracle optimality result implies that within this model family no other thresholding rule can achieve higher marginal power at the same marginal FDR, so the comparison baselines are operating below the achievable frontier.
The consistency result implies that with enough samples the estimated signal converges uniformly to the true signal, so the rejection set converges to the oracle rejection set in the limit.
The framework is not limited to seismic data: the same machinery transfers to any graph whose vertices are paired with a continuous covariate space such as space, frequency, or feature coordinates.

Reading between the lines

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

Inference: The paper's one-parameter family $f_{\mathrm{mix}}(p\mid\gamma)=\mathrm{sigmoid}(\gamma)\,p^{\mathrm{sigmoid}(\gamma)-1}$ ties the null prior to the alternative shape; if real data have these two ingredients varying independently, the method would be misspecified, and a two-parameter family within the same GGSP machinery would be the natural robustness upgrade.
Inference: A finite-sample FDR guarantee is a plausible extension the author did not develop; because the lfdr estimates come from an MLE on a compact parameter space, a non-asymptotic bound would likely require a Lipschitz or bracketing condition on the family $f_{\mathrm{mix}}$.
Inference: The bandlimitedness assumption could be tested empirically by comparing the BIC-fitted $\gamma$ against a flexible nonparametric estimate of the lfdr surface; large residuals would indicate that the low-dimensional subspace assumption is not adequate for the data at hand.

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

You should know two things about arXiv:2506.03496. The genuinely new idea is to model the two-groups mixture densities as continuously varying over a graph-time domain using bandlimited generalized graph signal coefficients. That is a real departure from the homogeneous models in AdaPT-style work and from methods with a finite number of states. The other thing is that the paper's central claim—asymptotic FDR control for the resulting MHT-GGSP strategy—is not proved here. It is delegated to a companion paper [20, Theorem 3], whose conditions are summarized as "mild" without being stated.

What the paper does well: the empirical-Bayes setup is coherent. The identification assumptions (Assumption 1) are standard in the lfdr literature and let them recover pi0 and f1 from the mixture. Theorem 1's oracle solution as an lfdr level set is a reasonable generalization of AdaPT, and the BIC-based selection of K1,K2 is sensible. The seismic detection task is a natural application for this kind of joint graph-time structure.

The soft spots are real but not all equally severe. The biggest one is the missing FDR theorem. Theorem 1 and Theorem 2 are also proved only by pointers to [20]; for a paper whose main selling point is a strategy "with asymptotic FDR control", this is a load-bearing gap. The step from pointwise MLE consistency of gamma to uniform control of the estimated r_M(eta) and the resulting lfdr thresholds is nontrivial and is not sketched in the text. If the companion paper is not public or if its assumptions differ from Assumptions 1-2, the advertised guarantee is unsupported.

Second, the parametric family fmix(p|gamma)=sigmoid(gamma) p^{sigmoid(gamma)-1} forces a single scalar to control both the null prior and the alternative density shape. That is restrictive, and the paper gives no misspecification analysis. If the model is wrong, Theorem 2's consistency is consistency for the wrong parameters; it does not imply the true lfdr is recovered, so FDR control is not inherited.

Third, the numerical section uses one dataset, one noise level, and 20 repetitions with no error bars. The statement that MHT-GGSP "observes the best power" is stronger than the evidence supports, especially since the FDR inflation of the competitors may be specific to this setup.

None of this is fatal on its own. The framework is coherent and the modeling idea is worth taking seriously. But the version I read is conditional: it needs the FDR theorem stated and proved (or a public companion with the full proof), error bars in the simulations, and at least some discussion of robustness to misspecification before the performance claims can be trusted.

The paper deserves a serious referee; I would send it out, but I would expect major revision. It is most useful for readers working on structured MHT in signal processing or spatial-temporal statistics. I would not cite it yet in its current form.

Referee Report

3 major / 5 minor

Summary. The paper proposes an empirical-Bayes multiple hypothesis testing (MHT) framework for hypotheses located on a joint graph-measure ('time') domain. Each hypothesis has a p-value, and the null prior probability and the alternative p-value density are assumed to vary smoothly over the joint domain through a latent random field gamma. The latent field is modeled as a bandlimited generalized graph signal with Fourier parameter matrix Xi. The authors derive an oracle decision rule that thresholds the local false discovery rate (lfdr), estimate Xi by maximum likelihood, and claim that plugging the estimated lfdr into the oracle rule yields asymptotic FDR control. The method is demonstrated on a seismic sensor-network detection task, where it is reported to achieve the best power among compared methods while controlling FDR.

Significance. If the central claim is correct, the paper offers a principled way to borrow strength across both graph structure and continuous side information in MHT, going beyond homogeneous or finite-state models. The model formulation is a genuine extension of the two-groups model, and the oracle derivation, though deferred, follows the known lfdr-thresholding logic. The authors should be credited for identifying a relevant application domain and for a concrete estimation procedure. However, the paper's theoretical core — the asymptotic FDR control theorem — is not stated or proved in the manuscript; it is only referenced to the authors' companion paper. This makes the advertised guarantee unverifiable without access to that work.

major comments (3)

[Section IV, final paragraph before Section V] The sentence 'It can be proved that, under mild conditions, this strategy admits asymptotic FDR control [20, Theorem 3]' is the paper's main advertised guarantee, yet the theorem is not stated and the 'mild conditions' are never enumerated. The reader cannot verify whether Assumptions 1 and 2 are sufficient, whether additional compactness or smoothness conditions are needed, or whether the plug-in step (replacing lfdr with its estimate) is covered. This is a load-bearing gap: the FDR control claim currently rests entirely on an external, unpublished theorem. Please state the theorem precisely and either prove it or give a self-contained argument that the conditions of [20, Theorem 3] are satisfied.
[Section IV, Theorem 2 and Eq. (20)] The uniform convergence statement sup_{(v,t) in J} |gamma_hat(v,t) - gamma(v,t)| -> 0 in probability is essential for the subsequent plug-in of the estimated lfdr into the quantities r_M(eta), d1,M(eta), d0,M(eta) (Eqs. (13)-(15)). The proof is deferred to [20, Appendix C]. Moreover, the paper does not explain how the sup-norm convergence of gamma_hat translates into uniform convergence of the estimated lfdr over the threshold eta, nor how the difference between the empirical ratio r_M(eta) and mFDR is controlled. These steps are not immediate consequences of pointwise MLE consistency and need to be supplied.
[Section V, first paragraph; Section III, Eq. (4)] The numerical implementation fixes the mixing density to fmix(p|gamma) = sigmoid(gamma) p^{sigmoid(gamma)-1}, so a single scalar gamma simultaneously determines the null prior and the alternative density shape. The theoretical results, however, are stated for a general parametric family f1(.;zeta). If this one-parameter family is misspecified, the consistency of the MLE (Theorem 2) does not imply that the estimated lfdr converges to the true lfdr, and the claimed asymptotic FDR control from [20, Theorem 3] could fail. The paper should either prove that the chosen family is sufficiently flexible or discuss the misspecification robustness of the procedure.

minor comments (5)

[Section II] The notation uses 'S' both for the sample set and, in the Notations paragraph, for matrices; this is confusing when reading equations such as (8). Please disambiguate, e.g., use a different symbol for matrices.
[Section IV, Theorem 1] The proof of Theorem 1 is deferred to [20, Appendix B]; since Theorem 1 is a minor generalization of existing results, a short proof sketch or an explicit statement of the modified arguments would make the paper more self-contained.
[Section V, Fig. 2] The figure caption states 'Each point is obtained by 20 repetitions,' but no error bars or standard deviations are shown. Given the paper's claim that MHT-GGSP 'observes the best power,' reporting uncertainty across repetitions would strengthen the empirical comparison.
[Section V, noise setup] The synthetic experiment uses a single noise multiplier (900 times the background noise). A sensitivity analysis over a range of signal-to-noise ratios would help establish the practical robustness of the proposed method.
[Throughout] There are several minor typos and inconsistent spacing (e.g., 'J' vs 'J ', 'BIC =K1K2 lnM−2l ∗ K1,K2' with a missing '=' sign and a stray 'l' exponent format). A careful proofreading pass is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity by construction; the asymptotic FDR guarantee is delegated to the companion paper [20], which is an omitted-proof and self-citation issue rather than a circular reduction.

full rationale

No circular derivation is exhibited. The statistical model in equations (3)-(4) defines the mixture density and the generalized graph signal gamma, while the identifiability identities (5)-(6) are algebraic consequences of Assumption 1 and do not presuppose the oracle or the FDR result. The oracle threshold (12)-(16) is a standard lfdr level-set construction, and Theorem 1 is presented as a known optimality result referenced to [18, Theorem 3] and [20, Appendix B], not as an input-output tautology. The MLE (19) and its consistency in Theorem 2 are standard under the compactness and identifiability assumptions in Assumption 2, with the proof deferred to [20, Appendix C]. The final claim, 'It can be proved that, under mild conditions, this strategy admits asymptotic FDR control [20, Theorem 3]', is the single place where the central guarantee is not proven in the manuscript; this is an omitted proof and a self-citation to the authors' companion paper, but it is not circular by construction because the companion theorem is not shown to be equivalent to the model definition or to a fitted parameter. The simulation compares methods on a seismic benchmark with synthetic noise and reports empirical FDR and power; no quantity that is fitted, such as gamma or the lfdr, is renamed as a predicted rejection, and the threshold is chosen to control an estimated FDR rather than to match the reported power. Thus the derivation chain does not reduce to its own inputs; the score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the known-null density, identifiability conditions, exact bandlimitedness, correct parametric family, i.i.d. sampling, and the companion paper's FDR theorem. The latent field gamma is a model construct with no independent evidence channel; the Fourier coefficients and BIC-selected bandwidths are fitted from data.

free parameters (4)

Fourier coefficients of the latent signal gamma = data-dependent, estimated by MLE in Eq. (19)
The latent signal gamma, and hence the lfdr and the detection thresholds, is fully determined by these MLE-fitted coefficients.
Bandwidth hyperparameters K1, K2 = chosen by BIC, dataset-dependent
Assumption 2(i) treats the truncation as exact; BIC selects K, but the approximation error is not included in the FDR theorem.
Alternative density parameterization a = sigmoid(gamma) = varies over the domain
The numerical model fixes fmix(p|gamma)=a p^{a-1}; this is a modeling choice that constrains the shape of the mixture.
Synthetic noise multiplier = 900
The numerical experiment adds AWGN with energy 900 times the background; the reported power comparison is conditional on this simulation setting.

assumptions (6)

domain assumption Null p-value density f0 is known, and under a simple null f0=Unif(0,1).
Used throughout the statistical model and in the seismic z-test, where the null z-statistic is taken to be standard normal.
domain assumption Assumption 1(i)-(iv): f0 non-decreasing, f1 non-increasing, f1(1)=0, and continuity; this ensures identifiability via Eqs. (5)-(6).
Core identifiability for lfdr; if violated, gamma and pi0 are not recoverable from the marginal p-value density.
domain assumption Assumption 2(i): gamma is exactly bandlimited in the product graph-Fourier/time-Fourier basis with finite K1, K2 and compact coefficient set.
MLE consistency in Theorem 2 and uniform convergence in Eq. (20) rely on this truncation being true rather than approximate.
domain assumption Assumption 2(iii)-(iv): fmix is identifiable in the Fourier coefficients, and the expected sup-norm log-likelihood is finite.
These are technical conditions needed for consistency of the maximum likelihood estimator.
domain assumption Sample points (v_m, t_m) are i.i.d. from a positive measure rho and are independent of gamma.
The likelihood in Eq. (19) treats all M p-values as independent draws; temporal or spatial dependence is excluded.
ad hoc to paper The companion paper's Theorem 3 gives asymptotic FDR control under mild conditions.
The theorem is not stated or proved in this preprint; the entire FDR control guarantee is delegated to [20].

invented entities (1)

Latent random scalar field gamma on the joint graph-time domain
purpose: Parameterizes the inhomogeneous null prior probability and the alternative p-value density over the joint domain.
gamma is not observed directly; it is inferred from the same p-values used to evaluate detection. No external measurement validates gamma.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Generalized Graph Signal Processing Framework for Multiple Hypothesis Testing over Networks." pith.science (2026). https://pith.science/paper/PZJJF2PE

@misc{pith2026250603496,
  author       = {Pith},
  title        = {Pith review of: A Generalized Graph Signal Processing Framework for Multiple Hypothesis Testing over Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PZJJF2PE}},
  note         = {Machine review of arXiv:2506.03496}
}

abstract

We consider the multiple hypothesis testing (MHT) problem over the joint domain formed by a graph and a measure space. On each sample point of this joint domain, we assign a hypothesis test and a corresponding $p$-value. The goal is to make decisions for all hypotheses simultaneously, using all available $p$-values. In practice, this problem resembles the detection problem over a sensor network during a period of time. To solve this problem, we extend the traditional two-groups model such that the prior probability of the null hypothesis and the alternative distribution of $p$-values can be inhomogeneous over the joint domain. We model the inhomogeneity via a generalized graph signal. This more flexible statistical model yields a more powerful detection strategy by leveraging the information from the joint domain.

Figures

Figures reproduced from arXiv: 2506.03496 by the authors.

**Figure 2.** FDR and detection power under different target FDR levels. Each [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 20 canonical work pages

[20]

A Graph Signal Processing Perspective of Network Multiple Hypothesis Testing with False Discovery Rate Control

X. Jian, M. Gölz, F. Ji, W. P. Tay, and A. M. Zoubir, “A graph signal processing perspective of network multiple hypothesis testing with false discovery rate control,”arXiv preprint arXiv:2408.03142, 2024

work page Pith review arXiv 2024
[1]

Controlling the false discovery rate: A practical and powerful approach to multiple testing,

Y . Benjamini and Y . Hochberg, “Controlling the false discovery rate: A practical and powerful approach to multiple testing,”Journal of the Royal Statistical Society: Series B (Methodological), vol. 57, no. 1, pp. 289–300, 1995

1995
[2]

A Direct Approach to False Discovery Rates,

J. D. Storey, “A Direct Approach to False Discovery Rates,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 64, no. 3, pp. 479–498, Aug. 2002

work page 2002
[3]

Multiple testing with the structure-adaptive benjamini–hochberg algorithm,

A. Li and R. F. Barber, “Multiple testing with the structure-adaptive benjamini–hochberg algorithm,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 81, no. 1, pp. 45–74, Feb. 2019

work page 2019
[4]

False discovery rate smoothing,

R. A. P. Wesley Tansey, Oluwasanmi Koyejo and J. G. Scott, “False discovery rate smoothing,”J. Amer. Statist. Assoc., vol. 113, no. 523, pp. 1156–1171, 2018

work page 2018
[5]

On large-scale multiple testing over networks: An asymptotic approach,

M. Pournaderi and Y . Xiang, “On large-scale multiple testing over networks: An asymptotic approach,”IEEE Trans. Signal Inf. Process. Netw., vol. 9, pp. 442–457, 2023

work page 2023
[6]

Multiple hypothesis testing framework for spatial signals,

M. Gölz, A. M. Zoubir, and V . Koivunen, “Multiple hypothesis testing framework for spatial signals,”IEEE Trans. Signal Inf. Process. Netw., vol. 8, pp. 771–787, 2022

work page 2022
[7]

Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,

A. Anis, A. Gadde, and A. Ortega, “Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,”IEEE Trans. Signal Process., vol. 64, no. 14, pp. 3775–3789, Mar. 2016

work page 2016

Show all 21 references

[8]

Sampling signals on graphs: From theory to applications,

Y . Tanaka, Y . C. Eldar, A. Ortega, and G. Cheung, “Sampling signals on graphs: From theory to applications,”IEEE Signal Process. Mag., vol. 37, no. 6, pp. 14–30, Nov. 2020

2020
[9]

Kernel-based reconstruction of graph signals,

D. Romero, M. Ma, and G. B. Giannakis, “Kernel-based reconstruction of graph signals,”IEEE Trans. Signal Process., vol. 65, no. 3, pp. 764–778, Oct. 2017

2017
[10]

Bayesian estimation of graph signals,

A. Kroizer, T. Routtenberg, and Y . C. Eldar, “Bayesian estimation of graph signals,”IEEE Trans. Signal Process., vol. 70, pp. 2207–2223, Mar. 2022

2022
[11]

Two channel filter banks on arbitrary graphs with positive semi definite variation operators,

E. Pavez, B. Girault, A. Ortega, and P. A. Chou, “Two channel filter banks on arbitrary graphs with positive semi definite variation operators,” IEEE Trans. Signal Process., vol. 71, pp. 917–932, Mar. 2023

2023
[12]

Graph filters for signal processing and machine learning on graphs,

E. Isufi, F. Gama, D. I. Shuman, and S. Segarra, “Graph filters for signal processing and machine learning on graphs,”IEEE Trans. Signal Process., pp. 1–32, 2024

2024
[13]

Generalized graph signal processing,

F. Ji and W. P. Tay, “Generalized graph signal processing,” inProc. IEEE Global Conf. on Signal and Information Processing, 2018, pp. 708–712

2018
[14]

A Hilbert space theory of generalized graph signal processing,

——, “A Hilbert space theory of generalized graph signal processing,” IEEE Trans. Signal Process., vol. 67, no. 24, pp. 6188–6203, Nov. 2019

2019
[15]

Kernel based reconstruction for generalized graph signal processing,

X. Jian, W. P. Tay, and Y . C. Eldar, “Kernel based reconstruction for generalized graph signal processing,”IEEE Trans. Signal Process., vol. 72, pp. 2308–2322, Apr. 2024

2024
[16]

Kernel ridge regression for generalized graph signal processing,

X. Jian and W. P. Tay, “Kernel ridge regression for generalized graph signal processing,” inProc. IEEE Int. Conf. Acoustics, Speech, and Signal Processing, Rhodes Island, Greece, 2023

2023
[17]

Efron,Large-Scale Inference: Empirical Bayes Methods for Estimation, Testing, and Prediction

B. Efron,Large-Scale Inference: Empirical Bayes Methods for Estimation, Testing, and Prediction. Cambridge: Cambridge University Press, 2010

2010
[18]

AdaPT: an interactive procedure for multiple testing with side information,

L. Lei and W. Fithian, “AdaPT: an interactive procedure for multiple testing with side information,”Journal of the Royal Statistical Society. Series B: Statistical Methodology, vol. 80, no. 4, pp. 649–679, 2018

2018
[19]

Optimal false discovery rate control for large scale multiple testing with auxiliary information,

H. Cao, J. Chen, and X. Zhang, “Optimal false discovery rate control for large scale multiple testing with auxiliary information,”Annals of Statistics, vol. 50, no. 2, pp. 807 – 857, 2022

2022
[21]

Seismic detectors: the state of the art,

J. Berger and R. L. Sax, “Seismic detectors: the state of the art,” SSR-R, Technical Report 80-4588, 1980

1980

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [20]

A Graph Signal Processing Perspective of Network Multiple Hypothesis Testing with False Discovery Rate Control

X. Jian, M. Gölz, F. Ji, W. P. Tay, and A. M. Zoubir, “A graph signal processing perspective of network multiple hypothesis testing with false discovery rate control,”arXiv preprint arXiv:2408.03142, 2024

work page Pith review arXiv 2024

[2] [1]

Controlling the false discovery rate: A practical and powerful approach to multiple testing,

Y . Benjamini and Y . Hochberg, “Controlling the false discovery rate: A practical and powerful approach to multiple testing,”Journal of the Royal Statistical Society: Series B (Methodological), vol. 57, no. 1, pp. 289–300, 1995

1995

[3] [2]

A Direct Approach to False Discovery Rates,

J. D. Storey, “A Direct Approach to False Discovery Rates,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 64, no. 3, pp. 479–498, Aug. 2002

work page 2002

[4] [3]

Multiple testing with the structure-adaptive benjamini–hochberg algorithm,

A. Li and R. F. Barber, “Multiple testing with the structure-adaptive benjamini–hochberg algorithm,”Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 81, no. 1, pp. 45–74, Feb. 2019

work page 2019

[5] [4]

False discovery rate smoothing,

R. A. P. Wesley Tansey, Oluwasanmi Koyejo and J. G. Scott, “False discovery rate smoothing,”J. Amer. Statist. Assoc., vol. 113, no. 523, pp. 1156–1171, 2018

work page 2018

[6] [5]

On large-scale multiple testing over networks: An asymptotic approach,

M. Pournaderi and Y . Xiang, “On large-scale multiple testing over networks: An asymptotic approach,”IEEE Trans. Signal Inf. Process. Netw., vol. 9, pp. 442–457, 2023

work page 2023

[7] [6]

Multiple hypothesis testing framework for spatial signals,

M. Gölz, A. M. Zoubir, and V . Koivunen, “Multiple hypothesis testing framework for spatial signals,”IEEE Trans. Signal Inf. Process. Netw., vol. 8, pp. 771–787, 2022

work page 2022

[8] [7]

Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,

A. Anis, A. Gadde, and A. Ortega, “Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,”IEEE Trans. Signal Process., vol. 64, no. 14, pp. 3775–3789, Mar. 2016

work page 2016