acceptodds
Under review as a conference paper at ICLR 2027

Controlling False Discovery in Arbitrarily Structured Hypothesis Spaces via Reproducing Kernels.

Abstract

Large-scale hypothesis testing is central to modern science, where controlling the False Discovery Rate (FDR) has become the standard approach to managing false positives across many simultaneous tests. Hypotheses rarely exist in isolation; they often exhibit structure through proximity, connectivity, or hierarchy. This structure represents both a challenge and an opportunity: while classical methods treat these dependencies as obstacles requiring conservative correction, leveraging them can substantially increase discovery power. Here, we reframe structured FDR control as a regularized learning problem. By optimizing within a suitable Reproducing Kernel Hilbert Space (RKHS), we introduce a framework that unifies continuous domains, graphs, and hierarchies under a single algorithm through kernel choice alone. This formulation enables smooth solutions in place of the piecewise-constant fits of prior methods, principled likelihood-based hyperparameter selection rather than heuristic tuning, and inference at unobserved locations which in turn supports sample-efficient experimental design. Building on this estimator, we provide two decision rules which we prove to control the FDR. We validate our method on two sources: spatial locations derived from high-dimensional real-world datasets, and a differential gene expression task utilizing protein-protein interaction graphs.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
FDR, Multiple hypotheses, RKHS.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4qy5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies FDR control over structured hypothesis spaces. It estimates this probability using a regularized mixture likelihood and proposes two decision rules: one combines spatial local-fdr ranking with marginal local-fdr accounting, while the other uses mirror statistics. The paper provides conditional FDR guarantees for the first rule, an mFDR bound with a slack term for the second, and an exact FDR guarantee for a masked variant.

Critical strengths

The paper addresses an interesting and practically relevant problem: FDR control in structured hypothesis spaces. Using relationships among hypotheses to improve discovery power is a worthwhile direction, and the proposed kernel formulation offers a flexible way to incorporate different types of structure.

Critical weaknesses
  1. The validity guarantee for Rule 1 does not cover the implemented procedure. Theorem C.1 requires the PC condition, but the paper does not establish that the proposed RKHS-based ordering satisfies it. The theorem also uses the true marginal parameters (αˉ,f0,f1)(\bar\alpha,f_0,f_1), whereas the implemented procedure uses their estimates, without analyzing the effect of this substitution. These gaps leave the statistical validity of the proposed Rule 1 unresolved.

  2. The practical meaning of Rule 2's control guarantee is unclear. The bound mFDR(R)≤q+O(1/λreg)+O(1/N)\mathrm{mFDR}(R)\le q+O(1/\lambda_{\mathrm{reg}})+O(1/N) contains a term that does not vanish with NN when λreg\lambda_{\mathrm{reg}} is fixed. The paper does not report the selected values or establish that the resulting bound is informative in the experimental settings. Meanwhile, the masked variant has an exact FDR guarantee, but its performance is not separately reported. Consequently, the experiments do not show whether the version with an exact guarantee retains the claimed power advantage.

Questions for authors
  1. Under what conditions does the proposed RKHS-based ordering satisfy PC? Does the control guarantee remain valid when the true marginal parameters are replaced by the estimates used in the experiments?

  2. What search range and selected values of λreg\lambda_{\mathrm{reg}} were used in each experiment? Can the resulting error bound be quantified sufficiently to determine whether it is nontrivial? Does the analysis cover the cross-validation selection used in the implementation?

  3. Why do several BH results exceed the nominal FDR level under the stated independent, uniform-null simulation setup? The conditions of BH should be weaker than your methods, right? Are the results of the simulation correct?

  4. Was the masked Rule 2 evaluated? If so, please report its FDR and power under the same settings. If not, can a direct comparison be provided to establish whether the claimed power advantage persists for the version with an FDR guarantee?

Reviewer_JHfDOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a unified framework for structured false discovery rate control by fomulating the problem as regularized learning in a RKHS. By encoding spatial, graph, or hierarchical structures through kernel choices, the framework provides smooth estimates of prior probabilities, and supports likelihood-based hyperparameter selection. The auther develop two decision rules with theoretical guarantees. Experiments on synthetic and real-world datasets demonstrate improved statistical power over existing methods while maintaing error control.

Critical strengths

The paper formulates structured FDR control as an RKHS-regularized learning problem, accommodating continuous domains, graphs, and hierarchies through different kernel choices. Two decision rules are proposed with FDR-related guarantees that are robust to spatial model misspecification under certain conditions.

Critical weaknesses

The theoretical guarantees rely on independent observations, whereas spatial and network-structured hypotheses often exibit dependence.

Questions for authors
  1. Do the FDR guarantees extend to dependent p-values, such as spatial dependence?
  2. Does the soft penalty theoretically guarantee that the estimated mixing probabilities remain in [0,1]?
  3. How sensitive are the results to kernel misspecification or weak spatial structure?
  4. How should practitioners choose between Rule 1 and Rule 2?
Reviewer_9FSJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper concerns about false discovery rate (FDR) control in multiple testing. Unlike classical local false discovery rate (lfdr) approaches, the authors allow the prior null probability to depend on covariates, thereby incorporating spatial and other structural information into the lfdr framework. By leveraging reproducing kernel Hilbert spaces (RKHS), the authors further establish an algorithmic stability-based upper bound on the marginal FDR (mFDR). Further simulations and real-data experiments are conducted to support the proposed methodology and theoretical analysis.

Critical strengths

This paper combines lfdr with mirror statistics and provides a stability-based mFDR bound for a multiple testing procedure with covariate-dependent null priors. The theoretical development is accompanied by extensive empirical evaluations.

Critical weaknesses

Several closely related methods have already been developed in the literature. Moreover, the proposed framework requires independent and uniformly distributed null p-values, which is a relatively strong assumption for lfdr-based methods. The paper does not establish a clear theoretical advantage over existing approaches, such as covariate-adaptive BH procedures or other methods that incorporate spatial information while controlling the FDR.

Questions for authors
  • The FDR guarantee for Rule 1 assumes the true marginal lfdr, whereas the implemented procedure uses a plug-in estimator. The authors should clarify this distinction and establish theoretical guarantees for the actual plug-in procedure.

  • The assumption ∑i∈H0P(ηi≤w)≤ρwN\sum_{i\in\mathcal H_0}P(\eta_i\le w)\le \rho wNmay fail because the data-dependent cutoff t^\widehat t is selected from the observed scores. If the selected cutoff coincides with a null score with positive probability, then P(ηi=0)>0P(\eta_i=0)>0, contradicting the bound as w→0w\to0. The authors should justify this assumption or provide a suitable modification.

  • The stability analysis assumes fixed regularization parameters and kernels, whereas these are selected through cross-validation in practice. The authors should clarify whether the theoretical guarantees extend to the full data-adaptive procedure.

  • Under independent uniform null p-values, BH should control the FDR at the nominal level. Why does Figure 2(b) show substantial FDR inflation for BH, sometimes even exceeding that of the proposed lfdr methods?

  • The HIGGS experiment uses the event threshold mwwbb<0.8m_{wwbb}<0.8, but the choice of 0.8 is not justified. The authors should report results for several alternative thresholds mwwbb<cm_{wwbb}<cto assess the robustness of the proposed methods.

  • More simulation settings, including less informative spatial structures, would strengthen the evaluation. In particular, the authors should compare the same decision rule using a global null prior αˉ\bar\alpha, an RKHS-estimated prior α^(x)\widehat\alpha(x), and the oracle prior α(x)\alpha(x), alongside BH. Empirical mFDR should also be reported, as it directly corresponds to the theoretical guarantees.

  • The authors should investigate how informative or tight the theoretical mFDR upper bound is. For example, varying λreg\lambda_{\mathrm{reg}} and comparing empirical mFDR with the theoretical bound would help assess the practical significance of the stability analysis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.