acceptodds
Under review as a conference paper at ICLR 2027

Learning Brain Hypergraphs from Pairwise-Irreducible Interactions

Abstract

Brain hypergraph and other higher-order models aim to capture information beyond pairwise functional connectivity. Current methods, however, assign higher-order status by construction, from group size, statistical order, or architecture, so neither multiregional coactivation nor improved accuracy shows that a model uses information that pairwise statistics cannot explain. We propose HyPI (Hypergraph Learning from Pairwise-Irreducible Interactions), which replaces structure-defined with evidence-defined higher-order modeling. For each group of three or four regions, HyPI measures how far the observed joint activity departs from a pairwise maximum-entropy reference that matches all univariate and pairwise activation statistics. Departures are calibrated against temporal surrogates and screened for split-half reproducibility, yielding a hypergraph with a shared topology and subject-specific interaction vectors, which together form a subject's interaction profile. These vectors gate a residual branch over a frozen pairwise predictor, so a group contributes nothing when its calibrated interaction is zero. To certify higher-order use, we compare held-out risk against the pairwise predictor and against a capacity-matched model trained on reassigned profiles, and we replace the profiles at a fixed checkpoint to test whether predictions depend on the correct subject-specific profile. Synthetic experiments show recovery of planted interactions and identification of label-relevant ones. On ABIDE, REST-meta-MDD, ADNI, and ADHD-200, HyPI attains the highest mean accuracy across eight cohort–parcellation settings, and the controls show that models with similar accuracy differ markedly in certified higher-order use.

Primary area:
applications to neuroscience & cognitive science
Keywords:
higher-order, hypergraph, fmri

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NoaZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes HyPI, a method for learning higher-order interactions from resting-state fMRI data. It aims to identify brain interactions that cannot be explained by pairwise relationships. The method compares the activity of groups of brain regions with a pairwise maximum-entropy model. It then uses statistical tests and reliability checks to select hyperedges. These interactions are added to a frozen pairwise prediction model through a residual branch.

The authors also propose three criteria to evaluate higher-order interactions: statistical reliability, additional predictive value, and whether the model uses the correct subject-specific interactions. They evaluate the method on synthetic data and four clinical fMRI datasets. The reported results show improved classification accuracy and stronger evidence of higher-order interaction use compared with the tested methods.

Critical strengths
  • The research question is meaningful. The paper aims to distinguish whether higher-order interactions contain information beyond pairwise relationships, whether they improve prediction, and whether the model actually uses this information.

  • The evaluation framework is valuable. The authors go beyond classification accuracy by introducing capacity-matched controls and counterfactual interventions to examine whether the model depends on subject-specific higher-order interactions.

Critical weaknesses
  • The paper is difficult to follow. Throughout the manuscript, the authors introduce many concepts, technical terms, and mathematical notations without sufficiently explaining their roles in the overall framework and how they support the main scientific claims. The transitions between different parts are often unclear, making it easy to lose track of the main argument. The lengthy appendices add many technical details but do not sufficiently help readers understand the overall logic.

  • Insufficient theoretical and experimental support for the proposed certification protocol. The paper proposes P1–P3 as criteria for certifying higher-order information use. However, I do not think the theoretical analysis and experiments provide sufficient support for this protocol. The theoretical results mainly establish properties of individual components, but do not clearly justify why the three criteria together can establish genuine higher-order information use. For example, Appendix C.3 acknowledges that the temporal-null test does not directly test pairwise irreducibility. The experiments evaluate HyPI using the proposed criteria, but do not adequately validate the criteria themselves under known ground-truth conditions.

  • Incomplete discussion of related work. The paper overlooks two relevant studies: Zhan et al. (NeurIPS 2025) on higher-order brain network modeling, and Liu et al. (Nature Methods, 2025) on the effects of different pairwise interaction measures. Although Santoro et al. (Nature Communications, 2026) is cited, its systematic comparison of higher-order methods is not discussed sufficiently. A clearer discussion of these studies would help establish the contribution of HyPI and justify its choice of pairwise representations.

Questions for authors
  1. Could the authors explain more clearly why the main components of HyPI are needed and why they are combined in this way? In particular, how do the design choices in Section 3.4 contribute to the paper's main objectives?

  2. What theoretical arguments support the use of P1–P3 together as a certification protocol? How does P1 establish pairwise irreducibility when Appendix C.3 acknowledges that the temporal-null test does not directly test this hypothesis?

  3. How are the synthetic datasets generated, including the planted interactions and labels? Do these experiments provide evidence that the proposed certification protocol can correctly distinguish genuine higher-order information use from cases where it is absent?

  4. The full certification comparison with confidence intervals is reported for ABIDE (AAL116), while related analyses are presented for other settings. Could the authors clarify whether the joint certification results are consistent across the remaining datasets and parcellations?

  5. Liu et al. (Nature Methods, 2025) systematically compared 239 pairwise interaction statistics and showed that their choice can substantially affect downstream results. Given these findings, how robust are HyPI's reported higher-order predictive advantages and certification results to the choice of pairwise representation?

9 comments
AuthorsOpenReview ↗

Response to NoaZ · Comment 1/5

We thank the Reviewer for the careful review and for finding the research question meaningful and the evaluation framework valuable.

We have uploaded a revised PDF with changes in blue. Below, AHOA_{\mathrm{HO}} and DHOD_{\mathrm{HO}} are in 10−310^{-3} nats, and a method is certified when the 95% lower bounds of both are above zero.

Q1 & W1. Why is each component needed, and how do the components fit together?

Response.

Each component serves one of three requirements, and the revision now states this mapping throughout.

(a) What changed for readability.

We revised the paper so that its logic is easier to follow:

  • We changed Figure 1 to show two groups with the same pairwise links but different joint activity, which a structure-defined hyperedge treats alike and HyPI separates.
  • We added to Section 3.1 a mapping of every component to the requirement (P1–P3) it serves.
  • We added to Section 3.2 the three uses of "pairwise": the reference that defines P1, the surrogates that calibrate it, and the predictor input for P2 and P3.
  • We added the Bayes decomposition behind the residual branch at the start of Section 3.4, and the target null to Section 3.5.
  • We added an appendix guide (Table 3, page 16) that maps each appendix section to the requirement it supports.

(b) One argument.

A claim that a model uses higher-order information requires three things:

  • (P1) A group's joint activity departs reliably from what its pairwise relations predict.
  • (P2) The correct interaction profile lowers held-out risk beyond the pairwise predictor and beyond a model of identical capacity.
  • (P3) The trained model depends on the subject's own profile.

Each component of HyPI serves one of these (see Table R-N1 below).

Table R-N1. Role of each component of HyPI. Ablation values are accuracy (%) / AHOA_{\mathrm{HO}} / DHOD_{\mathrm{HO}} (point estimates, 10−310^{-3} nats) on ABIDE (AAL116). The full model gives 72.84 / 5.48 / 6.47, and higher is better (↑) for all three.

ComponentServesWhy it is neededWithout it (acc. ↑ / AHOA_{\mathrm{HO}} ↑ / DHOD_{\mathrm{HO}} ↑)
Pairwise max-entropy referenceP1Defines what pairwise statistics already explain71.92 / 1.83 / 3.05
Temporal-null calibrationP1Calibrates the score against second-order-matched surrogates72.58 / 4.26 / 5.12
Split-half reliabilityP1Keeps only reproducible individual differences72.15 / 3.45 / 3.78
Frozen backbone with logit residualP2Any gain is attributable to the interaction branch72.41 / 2.21 / 3.42
Interaction gatingP3Messages come from the profile, not from which groups are present72.65 / 3.82 / 2.14
Hypergraph message passingP2, P3Profiles are read through the regions the groups share72.36 / 3.61 / 4.86
  • Removing any component lowers both AHOA_{\mathrm{HO}} and DHOD_{\mathrm{HO}}, so each is needed.
  • The pairwise reference and the frozen backbone matter most for AHOA_{\mathrm{HO}} (1.83 and 2.21), and gating matters most for DHOD_{\mathrm{HO}} (2.14).

(c) Design of the residual branch.

By Bayes' rule,

log⁡p(y∣S(2),FHO)=log⁡p(y∣S(2))+log⁡p(FHO∣y,S(2))p(FHO∣S(2)),\log p\big(y \mid S^{(2)}, F^{\mathrm{HO}}\big) = \log p\big(y \mid S^{(2)}\big) + \log \frac{p\big(F^{\mathrm{HO}} \mid y, S^{(2)}\big)}{p\big(F^{\mathrm{HO}} \mid S^{(2)}\big)},

so optimal use of higher-order information is an additive log-space correction to the best pairwise predictor, and its expected reduction in log-loss is the conditional mutual information I(Y;FHO∣S(2))I(Y; F^{\mathrm{HO}} \mid S^{(2)}). HyPI follows this identity:

  • Frozen pairwise predictor and logit residual. Only the second term is learned, so any change in risk comes from the interaction branch (P2).
  • Zero preservation. The interaction projections, the readout, and the classifier are bias-free, so a group with a zero interaction vector sends no message, and with all vectors at zero the model reduces exactly to the pairwise predictor. Replacing profiles in the Counterfactual thus changes only the interaction pathway (P3). The context networks have biases but only gate a message that the interaction vector produces.
  • Interaction gating. Pairwise context modulates how a profile is expressed at each region but cannot produce a message alone. Without gating, DHOD_{\mathrm{HO}} drops the most (Table R-N1), so gating is what makes predictions depend on the profile.
AuthorsOpenReview ↗

Response to NoaZ · Comment 2/5

Q2 & W2. What justifies using P1–P3 together, and how does P1 establish pairwise irreducibility?

Response.

P1–P3 together test one null, and each rules out an explanation that the other two cannot. P1 defines irreducibility by the score and uses the temporal null only to calibrate it.

(a) Target null and its connection to conditional independence.

We added to Section 3.5 the null that P1–P3 test together:

  • Higher-order use means that the profile carries label information that pairwise statistics do not, so the null is H0:Y⊥ ⁣ ⁣ ⁣⊥FHO∣S(2)H_0: Y \perp\!\!\!\perp F^{\mathrm{HO}} \mid S^{(2)}.
  • Conditional randomization would test it by resampling profiles from p(FHO∣S(2))p(F^{\mathrm{HO}} \mid S^{(2)}). This conditional is unknown, so we give each subject the profile of a pairwise-matched donor.
  • The capacity-matched Null is then a permute-and-relearn control, and the Counterfactual is a matched form of model reliance.
  • For a Bayes-optimal predictor and exact replacement, min⁡{RR,RN}−RF\min\{R_R, R_N\}-R_F equals I(Y;FHO∣S(2))I(Y; F^{\mathrm{HO}} \mid S^{(2)}).

The protocol is thus related to conditional randomization but is not an exact conditional randomization test, so certification is evidence relative to the chosen profiles, surrogates, learner, and matching.

(b) Why the three criteria are needed together.

Each rules out an explanation that the other two cannot:

  • P1 rules out profiles explained by pairwise statistics or estimation artifacts. Without it, a model could pass P2 and P3 with an auxiliary input computed from connectivity.
  • P2 rules out gains from added capacity or from re-optimizing the pairwise model. Without it, a model can depend on an input (P3) that adds no label information, as MvHo-IB does in most settings.
  • P3 rules out a model that ignores whose profile it receives. P2 compares models trained on true and reassigned profiles, so a gain can arise during training although the final predictions do not change when the profile is replaced. P3 tests the fixed checkpoint directly.

(c) The decision rule.

A trained predictor can also gain from pairwise information that its backbone underuses, as the synthetic negative control of Q3 shows. We thus added RSR_S, the risk of the same branch trained on surrogate-scan profiles, which keep each scan's spectra and connectivity but remove higher-order structure, so that AHO=min⁡{RR,RN,RS}−RFA_{\mathrm{HO}}=\min\{R_R, R_N, R_S\}-R_F (Section 3.5).

  • Certification requires the 95% lower bounds of AHOA_{\mathrm{HO}} and DHOD_{\mathrm{HO}} to exceed zero, which makes it an intersection–union test over four comparisons that share the bootstrap resamples and need no adjustment.
  • Under the idealization of (a), AHO=I(Y;FHO∣S(2))−I(Y;Fsur∣S(2))A_{\mathrm{HO}}=I(Y; F^{\mathrm{HO}} \mid S^{(2)})-I(Y; F^{\mathrm{sur}} \mid S^{(2)}) for the surrogate-scan profile FsurF^{\mathrm{sur}}, so a positive value still implies I(Y;FHO∣S(2))>0I(Y; F^{\mathrm{HO}} \mid S^{(2)})>0.
  • With the fitted models fixed, HyPI is certified under this rule in all eight settings, and the rule lowers its bounds by at most 0.03.

(d) How P1 establishes pairwise irreducibility.

We added to Section 3.2 that the score defines irreducibility and the temporal null calibrates it:

  • Js,gHOJ^{\mathrm{HO}}_{s,g} is the KL divergence from a group's empirical joint-state distribution to its projection onto the pairwise log-linear family. By the Pythagorean relation, it is the divergence that no pairwise model can remove, and it is zero if and only if the distribution lies in the pairwise family.
  • Finite scans, autocorrelation, and binarization produce positive divergences even for a stationary Gaussian process, so the temporal null calibrates the score against them.
  • Non-Gaussian marginals, nonstationarity, or motion could also exceed the null, so we tested them directly (Appendix F.2). Profiles on spectrum-preserving surrogates, with or without each ROI's amplitude distribution, are never certified. Under the submitted rule, certification holds under drift-preserving and time-varying nulls, FD censoring, motion-balanced classes, and same-site donors (lower bounds of AHOA_{\mathrm{HO}} 0.23 to 1.36).
AuthorsOpenReview ↗

Response to NoaZ · Comment 3/5

Q3 & W2. Does the protocol separate higher-order use from its absence under known ground truth?

Response.

Yes. HyPI profiles are certified in at most 1% of null cohorts and in 80% of positive-control cohorts. We changed the synthetic setup of the submission (50 subjects, 50 time points) to longer, fMRI-like scans, and added this validation to Section 4.2 and Appendix G.

(a) Generator.

  • Each cohort has 240 subjects, 30 ROIs, and 200 time points of autoregressive background activity (lag-one coefficient 0.5).
  • We plant four triplets and two quadruplets. In each planted group, the binary state of one member agrees with the parity of the other members with a subject-specific probability drawn from U(0.5,0.9)U(0.5, 0.9), so every planted group has zero pairwise correlation.
  • Hyperedges are fixed to the planted groups to isolate the protocol from discovery.ixed to the planted groups to isolate the protocol from discovery.

(b) Scenarios.

In the positive control, the label shifts the interaction strength of three planted groups. In four null scenarios, the planted interactions are label-free:

  • (i) The label shifts the correlation of an ROI pair outside every planted group.
  • (ii) The label shifts the correlation of two non-target members inside three planted groups.
  • (iii) The label sets this correlation to ±0.30\pm 0.30 with a random sign for one class and to 0 for the other, which a linear predictor on connectivity cannot read.
  • (iv) The label is drawn from a logistic model on the scan connectivity of these member pairs.

Scenarios (i)–(iii) act on population parameters, so H0H_0 holds for the population pairwise structure, but interaction estimates could still carry information about the latent coupling beyond the scan connectivity. In (iv), H0H_0 holds exactly for the model input. We run 200 cohorts per null scenario and 100 for the positive control.

(c) Results.

  • HyPI profiles are certified in at most 1% of the cohorts of every null scenario, consistent with the 2.5% level of each one-sided bound, and in 80% of the positive-control cohorts (see Table R-N2 below).
  • The table also shows a limitation of the submitted rule. Profiles built from the connectivity of the same groups, for which H0H_0 holds exactly, pass that rule in 56.0% to 76.5% of the cohorts with label information inside the groups, because the frozen predictor on all 435 connectivity entries underuses the few informative ones.
  • We thus added RSR_S (Q2(c)). These connectivity profiles equal their surrogate-scan profiles, so RS=RFR_S=R_F and the revised rule excludes them by construction, which is not calibration evidence. The calibration evidence is the rate for HyPI profiles, which keep 80% in the positive control because their surrogate-scan branch gains at most 1.5% of RR−RFR_R-R_F over the pairwise predictor.
  • In the full pipeline with discovery, 10 of 12 cohorts are certified when the planted interactions carry the label, and HyPI is 19.4% more accurate than the pairwise predictor. When the same interactions are present but only connectivity carries the label, all six cohorts retain the interactions and none is certified, so discovery alone does not produce certification.

Table R-N2. Percentage of synthetic cohorts certified (95% Clopper–Pearson intervals) under the two-way rule of the submission, for surrogate-scan profiles, and under the three-way rule. Higher-order structure carries no label information in all but the last row, and H0H_0 holds exactly for the model input in (iv). In the two-way and three-way columns, lower is better (↓) for the null rows and higher is better (↑) for planted interactions. The surrogate-scan column should be low in every row.

Label carried byProfileBetterTwo-waySurrogate-scanThree-way
Connectivity of other regionsHyPI↓0.5 [0.0, 2.8]0.0 [0.0, 1.8]0.0 [0.0, 1.8]
Group connectivity↓0.0 [0.0, 1.8]0.0 [0.0, 1.8]0.0 [0.0, 1.8]
Coupling inside groupsHyPI↓1.0 [0.1, 3.6]0.0 [0.0, 1.8]1.0 [0.1, 3.6]
Group connectivity↓69.0 [62.1, 75.3]67.5 [60.5, 73.9]0.0 [0.0, 1.8]
Sign-symmetric couplingHyPI↓0.0 [0.0, 1.8]0.5 [0.0, 2.8]0.0 [0.0, 1.8]
Group connectivity↓76.5 [70.0, 82.2]77.0 [70.5, 82.6]0.0 [0.0, 1.8]
Model input (iv)HyPI↓1.0 [0.1, 3.6]0.5 [0.0, 2.8]1.0 [0.1, 3.6]
Group connectivity↓56.0 [48.8, 63.0]54.5 [47.3, 61.5]0.0 [0.0, 1.8]
Planted interactionsHyPI↑80.0 [70.8, 87.3]0.0 [0.0, 3.6]80.0 [70.8, 87.3]
Reviewer_mwVDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HyPI, a framework for identifying and exploiting higher-order interactions in resting-state fMRI. Rather than defining higher-order relationships solely through hypergraph construction, the method compares observed joint binary activation distributions with pairwise maximum-entropy references. Candidate interactions are further subjected to temporal-surrogate calibration and split-half reliability screening. The resulting subject-specific interaction representations are incorporated through a zero-preserving residual branch over a frozen pairwise Brain Network Transformer. The paper also introduces an evaluation protocol that distinguishes controlled predictive advantage from functional dependence through capacity-matched null models and fixed-checkpoint counterfactual interventions. Experiments are conducted on four clinical fMRI datasets under two brain parcellations. HyPI obtains the highest reported mean classification accuracy across all eight settings. Synthetic experiments and additional interventions are used to evaluate interaction recovery and model dependence on higher-order representations.

Critical strengths
  1. A meaningful distinction between higher-order structure and its actual use by a predictive model. The paper addresses an important methodological gap: constructing hyperedges or achieving higher classification accuracy does not, by itself, establish that a model exploits information beyond pairwise representations. The separation between controlled predictive advantage and functional dependence is well motivated. The capacity-matched null and fixed-checkpoint counterfactual evaluation provide useful controls against alternative explanations based on model capacity or auxiliary-input presence.
  2. A carefully structured modeling and evaluation pipeline. The proposal-confirmation procedure, reliability screening, frozen backbone, and zero-preserving residual design provide a relatively transparent decomposition of the method. The experiments span four clinical cohorts and two parcellations, and the appendix documents many implementation and data-partition details. These aspects strengthen the technical credibility of the work.
Critical weaknesses
  1. The inferential evidence does not fully isolate higher-order neural interactions from alternative temporal and nuisance effects. The distinction between the pairwise maximum-entropy reference and the temporal-surrogate null requires more careful interpretation. Appendix C.3 explicitly states that the surrogate-based hypothesis test concerns invariance under shared-phase rotations rather than the hypothesis that joint binary states are fully explained by pairwise statistics. It also acknowledges that non-Gaussian marginals, nonstationarity, and transient artifacts can cause rejection. Consequently, although the KL divergence measures departure from the specified binary pairwise reference, the additional significance and reliability criteria do not necessarily establish that the retained interactions reflect robust, biologically meaningful higher-order mechanisms. Split-half agreement within the same scan may also retain stable nuisance effects. This concern is consequential because reliable pairwise-irreducible interaction discovery is central to the claimed contribution. I would like to see stronger controls using pairwise-only synthetic processes with realistic nuisance characteristics, sensitivity analyses for motion and binarization, and empirical false-discovery assessments for overlapping candidate groups under the employed multiple-testing procedure.
  2. The evidence for robust predictive superiority and generalizable certification remains limited. Although HyPI obtains the highest mean accuracy in all eight settings, its advantage over the strongest competing method is approximately 0.77–2.02 percentage points. Table 1 reports cross-fold standard deviations but does not establish whether these paired performance differences are statistically reliable. Furthermore, the full numerical joint-certification comparison in Appendix E.4 is presented for ABIDE with AAL116 parcellation. The reported bootstrap intervals are conditional on fitted models and fixed evaluation conditions, and do not incorporate retraining or rematching variability. The experiments also rely on subject-wise rather than leave-one-site-out cross-validation, despite the multisite nature of relevant cohorts. Appendix G reports noticeable demographic imbalance in ADHD-200, which motivates additional control analyses. Finally, the classification accuracy of the exact four-channel pairwise backbone used by HyPI should be explicitly reported across all settings to clarify the contribution of the higher-order residual branch. These issues do not invalidate the proposed architecture, but they limit the strength of the empirical conclusions.
Questions for authors
  1. Can the authors provide null simulations or controlled experiments that distinguish genuine higher-order effects from non-Gaussian temporal structure, nonstationarity, and motion-related artifacts? How robust are the retained interactions to the binarization threshold and scan length? Given the overlap among candidate ROI groups, what evidence supports false-discovery control under the dependence structure of the tests?
  2. Can the authors report the performance of the exact four-channel BNT-only backbone, the factual HyPI model, and the strongest competing models under matched input conditions, together with paired confidence intervals or statistical comparisons across the eight evaluation settings?
  3. How stable are the higher-order-use certification results across datasets, random training runs, and independently generated matched reassignment mappings? Can the authors provide additional joint-certification results beyond ABIDE-AAL116?
  4. Have the authors evaluated the potential effects of acquisition site, demographic imbalance, and head motion on classification and selected hyperedges? Site-held-out evaluation or appropriate nuisance-controlled analyses would materially strengthen the conclusions.
6 comments
AuthorsOpenReview ↗

Response to mwVD · Comment 1/4

We thank the Reviewer for the detailed review and for recognizing the distinction between higher-order structure and its use by a model.

We have uploaded a revised PDF with changes in blue. Below, AHOA_{\mathrm{HO}} and DHOD_{\mathrm{HO}} are in 10−310^{-3} nats, and a method is certified when the 95% lower bounds of both are above zero.

Q1 & W1. Can higher-order interactions be separated from temporal and nuisance effects?

Response.

Yes. We now test nuisance directly, at discovery and at certification.

(a) What defines irreducibility.

We added to Section 3.2 that the two parts of P1 play different roles:

  • The score Js,gHOJ^{\mathrm{HO}}_{s,g} is the divergence that no pairwise model can remove (Pythagorean relation), so it defines pairwise irreducibility.
  • The temporal null only calibrates this score against finite scans, autocorrelation, and binarization, so nuisance that exceeds the null must be tested directly, as in (b) to (d).

(b) Nuisance at discovery.

We added 100 pairwise-only null cohorts for each of four conditions (clean, motion-like spikes, slow variance drifts, state switching), and a robust discovery that adds FD censoring and a time-varying null, which recolors each group's surrogate with its sliding-window covariance (Appendices F.2 and G).

Table R-M1. Percentage of 100 null cohorts per condition in which discovery admits a group (95% Clopper–Pearson intervals). Lower is better (↓).

Null cohortsDefault null ↓Robust discovery ↓
Clean1 [0.0, 5.4]1 [0.0, 5.4]
Motion-like spikes31 [22.1, 41.0]7 [2.9, 13.9]
Slow variance drifts18 [11.0, 26.9]1 [0.0, 5.4]
State switching47 [36.9, 57.2]4 [1.1, 9.9]
  • Under the default null, nuisance causes admissions in 18% to 47% of cohorts, because subjects with stronger nuisance keep their ranks across scan halves.
  • Robust discovery lowers this to 1% to 7% (Table R-M1), so most nuisance no longer passes discovery.
  • The 7% for spikes comes from spikes below the censoring threshold, and masking the true spike frames lowers it to 2%.
  • Without nuisance, the time-varying null still admits 5.5 planted groups per cohort, against 5.8 by default, so it costs little power.
  • On real data, HyPI stays certified under robust discovery in all eight settings (Q3).

(c) Nuisance at certification.

We added profiles recomputed on surrogates of each subject's own scan and recertified the retrained branch (Appendix F.2). Surrogate-scan profiles keep spectra and connectivity, and amplitude-matched profiles also keep each ROI's amplitude distribution.

  • Neither is certified in the four tested settings, and their point estimates of AHOA_{\mathrm{HO}} lie within 0.1 of zero, so spectra and non-Gaussian marginals do not produce the certified gain.
  • True profiles beat the branch trained on surrogate-scan profiles in all eight settings, a comparison the revised rule now requires (Q3).
  • Under the submitted rule, certification holds with ROI spectra in the pairwise input (lower bounds of AHOA_{\mathrm{HO}} 0.41 to 1.31), under a drift-preserving null (0.47 to 1.28), and under the time-varying null (0.33 to 1.19), so slow drifts do not explain it.

(d) Binarization, scan length, preprocessing, and multiple testing.

We added these analyses on ABIDE (AAL116) (Appendices F.3 and F.4).

  • Thresholds at the median and at the mean ±\pm 0.5 SD give retained sets with a Jaccard overlap of 0.71 to 0.88 with the default set and keep certification, so the result does not hinge on the threshold.
  • Truncating scans to 150 volumes keeps certification (0.81 / 1.46), and 91% of the groups retained from truncated scans are also retained from full scans, so short scans make discovery more conservative, not different.
  • Global signal regression keeps certification (1.11 / 1.93).
  • Overlapping candidates may violate the positive dependence that BH needs, so we also ran Benjamini–Yekutieli, valid under any dependence (Section 3.3). It retains fewer groups (30.4 triplets and 15.9 quadruplets, against 38.2 and 21.6) and keeps certification (1.02 / 1.79).
  • Neither procedure controls the false-discovery rate of the set that the reliability screen keeps, so we measured it on synthetic data (Appendix G). In 12 cohorts with both decoy blocks, all 69 admitted groups are planted.

(e) Scope.

HyPI certifies a departure from a pairwise reference that a trained model uses, not a neural mechanism. We added this to Section 4.6, which already called the anatomy descriptive.

AuthorsOpenReview ↗

Response to mwVD · Comment 2/4

Q2 & W2. Are the accuracy gains statistically reliable, and how accurate is the four-channel backbone?

Response.

The gain over the four-channel backbone is reliable at the subject level. Moreover, our method is certified in all 8 settings, and the closest competitor in none.

(a) Matched inputs.

In the submission, the backbone of HyPI received three dynamic channels that the pairwise baselines did not, so part of the gain could come from a richer input. We changed Table 1 so that all pairwise baselines receive the same four-channel input, which makes BNT the exact frozen predictor of HyPI, and higher-order baselines that accept pairwise connectivity also receive the dynamic channels (see Table R-M2 below).

Table R-M2. Accuracy (%) under matched inputs, averaged over five outer folds and three seeds. BNT (four-channel) is the frozen predictor of HyPI. Higher is better (↑).

Method (accuracy ↑)ABIDE AAL116ABIDE CC200REST-meta-MDD AAL116REST-meta-MDD CC200ADNI AAL116ADNI CC200ADHD-200 AAL116ADHD-200 CC200
BNT (static, as submitted)67.6868.3267.1567.8865.1265.7865.4565.15
BNT (four-channel)70.4171.0669.8370.7168.3768.7168.8868.62
MvHo-IB (+ dynamic channels)71.7672.8471.5472.2169.5470.0669.8169.62
HyperDiag (+ dynamic channels)71.6272.3771.4771.9569.3870.1469.6169.23
HA-STA (+ dynamic channels)72.5273.0672.2172.6770.4770.8769.5070.22
HyPI72.8473.3172.4772.9570.5971.1671.4771.28
  • The dynamic channels raise BNT by 2.7% to 3.5%, so the input mismatch was real.
  • HyPI still adds 2.2% to 2.7% over its frozen backbone, so the residual branch contributes beyond the richer input.
  • HyPI is the most accurate method in all eight settings.

(b) Statistical tests.

In Appendix, we have added paired tests at two levels, with Holm adjustment over the eight settings (Appendix F.1: per-setting accuracies in Table 15, fold-level pp-values in Table 16, paired intervals in Table 17, AUC and balanced accuracy in Table 18).

  • Subject level. With a paired sign-flip test on pooled out-of-fold predictions, seeds averaged, the gain over the four-channel BNT is significant in every setting (adjusted p≤p\le 0.017), and its 95% paired interval excludes zero in all eight settings (lower bounds 0.41% to 1.71%).
  • Subject level, competitors. The differences from MvHo-IB (0.47% to 1.66%) and HA-STA (0.12% to 1.97%) are not significant (unadjusted p≥p\ge 0.13), and their 95% paired intervals include zero in all eight settings.
  • Fold level. With the Nadeau–Bengio test on five paired fold accuracies, the gain over BNT has unadjusted pp from 0.04 to 0.22 and is not significant after adjustment, since five folds give little power.

The subject-level test conditions on the trained models, so we claim the gain over BNT as significant at the subject level only.

(c) Why certification is needed.

Accuracy cannot separate HyPI from HA-STA and MvHo-IB, so the revision rests the claim of higher-order use on certification.

  • HyPI is certified in all eight settings, and HA-STA and MvHo-IB in none (Q3).
  • The paired lower bound of the difference in DHOD_{\mathrm{HO}} between HyPI and HA-STA is positive in all eight settings (1.96 to 2.66), so HyPI depends on its profiles more than HA-STA does.
  • Under the submitted rule, the paired lower bound of the difference in AHOA_{\mathrm{HO}} is positive in seven settings against HA-STA and in six against MvHo-IB.

The abstract, introduction, and conclusion now report that HyPI is certified in all eight settings.

AuthorsOpenReview ↗

Response to mwVD · Comment 3/4

Q3 & W2. Is certification stable across datasets, training runs, and matchings?

Response.

Yes, in all eight settings with fitted models fixed, and on ABIDE and REST-meta-MDD under the strictest analysis.

(a) Datasets and parcellations.

Table R-M3 (below) reports all six certifiable methods in all eight settings under a revised rule:

A synthetic negative control showed that the submitted rule, AHO=min⁡{RR,RN}−RFA_{\mathrm{HO}}=\min\{R_R, R_N\}-R_F, can credit a branch that reads pairwise information more efficiently than its frozen predictor. We thus added RSR_S, the risk of the same branch trained on surrogate-scan profiles, which keep each scan's spectra and connectivity but remove higher-order structure, so that AHO=min⁡{RR,RN,RS}−RFA_{\mathrm{HO}}=\min\{R_R, R_N, R_S\}-R_F.

Table R-M3. 95% lower bounds of AHOA_{\mathrm{HO}} / DHOD_{\mathrm{HO}} (10−310^{-3} nats) in all eight settings under the revised rule. Higher is better (↑), and certified cells (both bounds above zero) are in bold. All baselines are retrained with matched inputs. †^\daggerCertified under the submitted rule (0.21 / 1.18).

Method (AHOA_{\mathrm{HO}} / DHOD_{\mathrm{HO}} ↑)ABIDE AAL116ABIDE CC200REST-meta-MDD AAL116REST-meta-MDD CC200ADNI AAL116ADNI CC200ADHD-200 AAL116ADHD-200 CC200Certified ↑
HyPI1.24 / 2.041.52 / 1.831.08 / 2.261.32 / 1.710.71 / 1.580.90 / 1.390.57 / 1.470.80 / 1.128/8
MvHo-IB−1.39-1.39 / 1.36−1.02-1.02 / 1.62−1.42-1.42 / 0.87−0.92-0.92 / 1.31−0.85-0.85 / 0.95−1.61-1.61 / 1.18$^\dagger$−1.50-1.50 / 0.42−1.20-1.20 / −0.21-0.210/8
HA-STA−0.97-0.97 / −2.44-2.44−0.11-0.11 / −1.94-1.94−0.95-0.95 / −2.12-2.12−0.59-0.59 / −1.68-1.68−1.06-1.06 / −2.47-2.47−0.78-0.78 / −1.83-1.83−1.29-1.29 / −2.61-2.61−0.95-0.95 / −2.05-2.050/8
HyperDiag−2.62-2.62 / −3.88-3.88−1.92-1.92 / −3.27-3.27−2.84-2.84 / −3.58-3.58−2.01-2.01 / −3.02-3.02−2.66-2.66 / −4.12-4.12−2.26-2.26 / −3.49-3.49−3.11-3.11 / −3.95-3.95−2.46-2.46 / −3.66-3.660/8
HGST−3.66-3.66 / −3.47-3.47−3.11-3.11 / −2.71-2.71−3.91-3.91 / −3.36-3.36−2.73-2.73 / −2.94-2.94−3.67-3.67 / −3.88-3.88−3.26-3.26 / −3.07-3.07−4.10-4.10 / −3.73-3.73−3.42-3.42 / −2.86-2.860/8
HOI-Brain−3.14-3.14 / −3.89-3.89−2.57-2.57 / −3.35-3.35−3.32-3.32 / −3.91-3.91−2.70-2.70 / −3.12-3.12−3.19-3.19 / −4.24-4.24−2.80-2.80 / −3.57-3.57−3.59-3.59 / −4.06-4.06−2.88-2.88 / −3.69-3.690/8
  • HyPI is certified in every setting, and every comparison behind the rule has an interval above zero.
  • The five baselines, retrained with matched inputs, are certified in none, so certified higher-order use is specific to HyPI.
  • For HyPI, the revised rule lowers the bounds by at most 0.03, since the surrogate-scan branch gains at most 0.02×10−30.02\times10^{-3} nats over the pairwise predictor.
  • The one certification of MvHo-IB under the submitted rule, on ADNI (CC200), does not hold under the revised rule (lower bound of AHOA_{\mathrm{HO}} −1.61-1.61) as desired (its surrogate-scan profiles, which lack higher-order structure, also passed there).
  • Fed through our residual branch over the same frozen BNT, MvHo-IB and HA-STA are certified in one and no settings under the revised rule, so the separation comes mainly from the profiles.

These bounds keep the fitted models fixed.

(b) Retraining and rematching.

We added 60 runs per setting, combining 20 random near-neighbor matchings with three seeds that each retrain the backbone and every branch (Appendix E.7), and tested three progressively stricter analyses:

  1. Retraining and rematching. AHOA_{\mathrm{HO}} is positive in 98.3% to 100% of the runs, and the pooled lower bounds stay positive in all eight settings (0.05 to 0.80 under the submitted rule), so certification does not hinge on one run or matching.
  2. Robust discovery. With fixed models, robust discovery of Q1(b) keeps all eight settings certified (0.22 to 1.16), so certification survives the stricter discovery.
  3. Both together. HyPI stays certified on ABIDE and REST-meta-MDD (0.20 to 0.47) and is inconclusive on ADNI and ADHD-200 (−0.38-0.38 to −0.13-0.13). These are the two smallest cohorts (825 and 766 subjects), so the extra variance widens their intervals past zero, while DHOD_{\mathrm{HO}} stays positive (0.16 to 0.52).

The abstract, introduction, Section 4.1, and conclusion now report both results.

(c) Quality of the matched replacement.

We added matching diagnostics (Appendix E.7).

  • Matched donors are about a quarter closer to the receiver than random donors (distance ratio 0.71 to 0.78), so donors resemble the receiver in pairwise input.
  • A classifier separates true from reassigned (pairwise input, profile) pairs with an AUC of only 0.52 to 0.56, against 0.57 to 0.62 for random pairing, so little mismatch remains.
Reviewer_S6pYOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HyPI for learning higher-order brain interactions from resting-state fMRI. It measures deviations from a pairwise maximum-entropy reference, then applies temporal surrogate calibration and split-half reliability screening to select hyperedges. Subject-specific interaction profiles are incorporated through a residual branch over a frozen pairwise predictor. The paper also introduces capacity-matched and counterfactual controls to assess the predictive contribution and actual use of these profiles. Experiments on synthetic data and four fMRI datasets show improvements over the compared methods, supported by ablation and sensitivity analyses.

Critical strengths

S1. Evidence-based hypergraph construction.

Rather than defining higher-order interactions solely by group size or hypergraph structure, HyPI evaluates deviations from a pairwise maximum-entropy reference. The temporal calibration and split-half screening further account for estimation variability, making the selected interactions more interpretable.

S2. Evaluation of higher-order information use.

The frozen pairwise backbone and capacity-matched control help isolate the predictive contribution of interaction profiles. The same-checkpoint counterfactual test further examines whether predictions depend on the correct subject-specific profiles. These controls provide evidence beyond classification accuracy alone.

Critical weaknesses

W1. Limited coverage of the certification analysis.

The detailed certification results are reported mainly for ABIDE (AAL116), although classification performance is evaluated across eight dataset–parcellation settings. Since certification is a central contribution, results from additional settings would help establish whether the observed predictive advantage and profile dependence are consistent across datasets.

W2. Unclear robustness to acquisition-related variability.

The evaluation uses subject-wise cross-validation rather than site-independent splits. Although temporal surrogate calibration and reliability screening are included, their effectiveness against site differences, head motion, and scan-length variation is not fully established. This leaves some uncertainty about the stability of the discovered interactions under different acquisition conditions.

Questions for authors

Q1. Does HyPI satisfy the joint certification criterion on datasets and parcellations beyond ABIDE (AAL116)? Could the authors report the corresponding results?

Q2. How sensitive are the discovered hyperedges to acquisition-site differences, head motion, and scan length? Are there additional analyses supporting their robustness?

2 comments
AuthorsOpenReview ↗

Response to S6pY · Comment 1/2

We thank the Reviewer for the careful reading and for recognizing the evidence-based hyperedge admission and the certification of higher-order use.

We have uploaded a revised PDF with changes in blue. Below, AHOA_{\mathrm{HO}} and DHOD_{\mathrm{HO}} are in 10−310^{-3} nats, and a method is certified when the 95% lower bounds of both are above zero.

Q1 & W1. Is joint certification consistent beyond ABIDE (AAL116)?

Response.

Yes. The table below uses a revised rule:

A synthetic negative control showed that the submitted rule, AHO=min⁡{RR,RN}−RFA_{\mathrm{HO}}=\min\{R_R, R_N\}-R_F, can credit a branch that reads pairwise information more efficiently than its frozen predictor. We thus added RSR_S, the risk of the same branch trained on surrogate-scan profiles, which keep each scan's spectra and connectivity but remove higher-order structure, so that AHO=min⁡{RR,RN,RS}−RFA_{\mathrm{HO}}=\min\{R_R, R_N, R_S\}-R_F.

Table R-S1. 95% lower bounds of AHOA_{\mathrm{HO}} / DHOD_{\mathrm{HO}} (10−310^{-3} nats) in all eight settings under the revised rule. Certified cells are in bold. All baselines are retrained with matched inputs. †^\daggerCertified under the submitted rule (0.21 / 1.18). Higher is better (↑).

Method (AHOA_{\mathrm{HO}} / DHOD_{\mathrm{HO}} ↑)ABIDE AAL116ABIDE CC200REST-meta-MDD AAL116REST-meta-MDD CC200ADNI AAL116ADNI CC200ADHD-200 AAL116ADHD-200 CC200Certified ↑
HyPI1.24 / 2.041.52 / 1.831.08 / 2.261.32 / 1.710.71 / 1.580.90 / 1.390.57 / 1.470.80 / 1.128/8
MvHo-IB−1.39-1.39 / 1.36−1.02-1.02 / 1.62−1.42-1.42 / 0.87−0.92-0.92 / 1.31−0.85-0.85 / 0.95−1.61-1.61 / 1.18$^\dagger$−1.50-1.50 / 0.42−1.20-1.20 / −0.21-0.210/8
HA-STA−0.97-0.97 / −2.44-2.44−0.11-0.11 / −1.94-1.94−0.95-0.95 / −2.12-2.12−0.59-0.59 / −1.68-1.68−1.06-1.06 / −2.47-2.47−0.78-0.78 / −1.83-1.83−1.29-1.29 / −2.61-2.61−0.95-0.95 / −2.05-2.050/8
HyperDiag−2.62-2.62 / −3.88-3.88−1.92-1.92 / −3.27-3.27−2.84-2.84 / −3.58-3.58−2.01-2.01 / −3.02-3.02−2.66-2.66 / −4.12-4.12−2.26-2.26 / −3.49-3.49−3.11-3.11 / −3.95-3.95−2.46-2.46 / −3.66-3.660/8
HGST−3.66-3.66 / −3.47-3.47−3.11-3.11 / −2.71-2.71−3.91-3.91 / −3.36-3.36−2.73-2.73 / −2.94-2.94−3.67-3.67 / −3.88-3.88−3.26-3.26 / −3.07-3.07−4.10-4.10 / −3.73-3.73−3.42-3.42 / −2.86-2.860/8
HOI-Brain−3.14-3.14 / −3.89-3.89−2.57-2.57 / −3.35-3.35−3.32-3.32 / −3.91-3.91−2.70-2.70 / −3.12-3.12−3.19-3.19 / −4.24-4.24−2.80-2.80 / −3.57-3.57−3.59-3.59 / −4.06-4.06−2.88-2.88 / −3.69-3.690/8

Table R-S1 reports all six certifiable methods in all eight settings:

  • HyPI is certified in every setting, with lower bounds of AHOA_{\mathrm{HO}} from 0.57 to 1.52 and of DHOD_{\mathrm{HO}} from 1.12 to 2.26.
  • The five baselines, retrained with the dynamic channels that the backbone of HyPI receives, are certified in none (as expected).
  • For HyPI, the revised rule lowers the bounds by at most 0.03, since the surrogate-scan branch gains at most 0.02×10−30.02\times10^{-3} nats over the pairwise predictor.
  • The one certification of MvHo-IB under the submitted rule, on ADNI (CC200), does not hold under the revised rule (lower bound of AHOA_{\mathrm{HO}} −1.61-1.61) as desired (its surrogate-scan profiles, which lack higher-order structure, also passed there).

These bounds keep the fitted models fixed.

We also tested stability under three progressively stricter analyses (Appendices E.7 and F.2):

  1. Retraining and rematching. Three seeds, each retraining the backbone and every branch, are combined with 20 random near-neighbor matchings for the Null, giving 60 runs per setting. AHOA_{\mathrm{HO}} is positive in 98.3% to 100% of the runs, and the lower bounds pooled over the 60 runs stay positive in all eight settings (0.05 to 0.80 under the submitted rule, which the revised rule lowers by at most 0.03).
  2. Robust discovery. Discovery adds FD censoring and a time-varying null, which reproduces the slowly varying covariance of each candidate group. With fixed models, HyPI remains certified in all eight settings (lower bounds of AHOA_{\mathrm{HO}} 0.22 to 1.16).
  3. Both together. With robust discovery and the 60 runs pooled, HyPI remains certified on ABIDE and REST-meta-MDD (lower bounds of AHOA_{\mathrm{HO}} 0.20 to 0.47) and is inconclusive on ADNI and ADHD-200 (−0.38-0.38 to −0.13-0.13). These are the two smallest cohorts (825 and 766 subjects), so the extra variance widens their intervals past zero, while DHOD_{\mathrm{HO}} stays positive (0.16 to 0.52).

Section 4.1 now covers all eight settings.

AuthorsOpenReview ↗

Response to S6pY · Comment 2/2

Q2 & W2. How robust are the hyperedges and certification to site, head motion, and scan length?

Response.

Certification holds for each factor. The controls below were run under the submitted rule as sensitivity analyses of that rule, and Appendices F.2, F.3, and F.5 give the full tables.

(a) Site.

  • Site explains only 1.9% to 3.1% of the variance of the mean interaction norm.
  • Restricting reassignment to same-site donors keeps certification in the four tested settings (lower bounds of AHOA_{\mathrm{HO}} 0.35 to 1.24).
  • We also evaluated on sites unseen during discovery and training, with leave-one-site-out on ABIDE (AAL116, 17 folds) and ADHD-200 (AAL116, 7 folds) and five site-disjoint folds over the 24 sites of REST-meta-MDD (AAL116):
  • All methods lose accuracy. HyPI remains the most accurate (68.94%, 66.95%, and 69.18%, against 67.41%, 65.40%, and 68.02% for HA-STA), and it is the only tested method certified in all three cohorts (lower bounds 0.42 / 1.08, 0.21 / 0.74, and 0.49 / 1.21).
  • The retained hyperedge sets overlap across site folds with a mean Jaccard index of 0.61 to 0.70, close to 0.57 to 0.66 under subject-wise folds.

(b) Head motion.

  • The mean interaction norm correlates only weakly with mean framewise displacement (FD), with Spearman ρ\rho from 0.06 to 0.13.
  • Certification holds with motion-balanced classes (lower bounds of AHOA_{\mathrm{HO}} 0.23 to 1.12).
  • Certification also holds when we censor every frame with FD >0.5>0.5 mm, one frame before it, and two after it, and rerun the full pipeline (0.31 to 1.36).
  • On ABIDE (AAL116), censoring retains 34.6 triplets and 19.2 quadruplets, against 38.2 and 21.6 by default, with a Jaccard overlap of 0.74. On ADHD-200 (AAL116), the cohort with the most motion, the overlap is 0.66.

(c) Scan length.

  • Truncating every ABIDE scan to its first 150 volumes shortens 74% of the scans.
  • On AAL116, fewer groups pass the reliability screen (26.3 triplets and 12.8 quadruplets), and 91% of them per fold are also retained with full scans. Accuracy is 72.18%, against 69.86% for the frozen BNT on the same scans, and certification holds (0.81 / 1.46).
  • Scan length is set by the site protocol and is shared by both classes within a site.
  • A classifier on scan length alone reaches a balanced accuracy of 51.6%, and adding scan length to the frozen predictor and the matching keeps certification (1.17 / 1.95).
  • On full scans, the fraction of zero interaction vectors is 35.2%, 31.0%, and 28.0% in the shortest, middle, and longest thirds of ABIDE scans.
  • On synthetic cohorts with 150 time points and strong autocorrelation, the reliability screen retains 0.7 groups on average, so short scans make HyPI more conservative.

(d) Stability across subjects.

Across the five outer folds, which use partly different discovery subjects, the retained sets overlap with a mean Jaccard index of 0.49 to 0.66, and no fold yields an empty hypergraph.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

General Response

We thank the AC / Reviewers for valuable comments.

All reviewers found the question of what brain hypergraphs learn beyond pairwise connectivity meaningful and the certification framework valuable but have also further questions.

The table below summarizes the key issues, how we addressed them, and if they are resolved. The revised PDF marks changes in blue. A method is certified when the 95% lower bounds of AHOA_{\mathrm{HO}} and DHOD_{\mathrm{HO}} are both above zero.

Key issue (raised by)How we addressed itStatus
Certification beyond ABIDE (AAL116) (S6pY, NoaZ)Rev. thought that certification was shown only for ABIDE (AAL116). We added certification of all six methods in all eight settings (Appendix E.6). A synthetic control exposed a false pass of the submitted rule, so the rule now also compares against surrogate-scan profiles. HyPI is certified in all eight settings, the five baselines in none.Resolved
Fair baselines / statistical reliability (mwVD)Rev. thought that part of the gain came from a richer input, as our frozen backbone had 3 dynamic channels that the pairwise baselines lacked. We removed this mismatch: pairwise baselines now get the same four-channel input, and higher-order baselines its dynamic channels. Every method sees the same information now. In paired tests (Appendix F.1), HyPI gains 2.2% to 2.7% over its backbone (significant at the subject level), but its margins over HA-STA / MvHo-IB are not significant, so accuracy alone cannot show higher-order use.Resolved. Certification separates these models
Temporal and nuisance effects (mwVD)Rev. thought that discovery could admit head motion, slow drifts, or state switching. We added a robust discovery with FD censoring and a time-varying null, and tested it on null cohorts with spikes, drifts, and state switching (Appendix G). False admissions fall from 18%–47% to 1%–7%.Resolved. The 7% for spikes comes from spikes below the censoring threshold
Robustness to site, motion, scan length, demographics (S6pY, mwVD)Revs. thought that these factors could drive the results under subject-wise folds, so we added a control for each: leave-one-site-out evaluation on three cohorts, same-site donors, motion-balanced classes, FD censoring, truncation to 150 volumes, and sex / age control on ADHD-200 (Appendices F.2, F.3, F.6). Certification holds.Resolved
Stability over retraining / rematching (mwVD)Rev. felt that the intervals ignore retraining / rematching. We fixed it by pooling 60 runs per setting from 3 seeds and 20 matchings (Appendix E.7). Pooled bounds stay positive in all eight settings, and with robust discovery also applied, on ABIDE and REST-meta-MDD.Addressed: HyPI stays certified on ABIDE / REST-meta-MDD, ADNI / ADHD-200, the two smallest cohorts, are inconclusive (as expected). Abstract, Section 4.1, conclusion updated.
Logic / presentation (NoaZ)Rev. thought that the roles of many concepts were unclear. We added a component-to-requirement map (Section 3.1), the Bayes decomposition behind the residual branch (Section 3.4), and an appendix guide, and changed Figure 1 to a concrete example.Addressed
Justification of P1–P3 (NoaZ)Rev. thought that P1–P3 lack a joint justification and that the temporal null does not test pairwise irreducibility. We added the null that P1–P3 test together, Y⊥FHO∣S(2)Y \perp F^{\mathrm{HO}} \mid S^{(2)} (Section 3.5), and stated that the score defines irreducibility while the temporal null only calibrates it (Section 3.2). Each criterion rules out one explanation the others cannot.Resolved
Validation under known ground truth (NoaZ)Rev. thought that P1–P3 were not validated under known ground truth. We changed the synthetic setup of Section 4.2 to 240 subjects and 200 time points and added four null scenarios and a positive control (Appendix G). HyPI is certified in at most 1% of null cohorts and 80% of positive controls, so the protocol rarely certifies absent use and detects it when present.Resolved
Related work and pairwise statistic (NoaZ)Rev. thought that Zhan et al. (NeurIPS 2025) and Liu et al. (Nat. Methods 2025) were missing, Santoro et al. (Nat. Commun. 2026) was underdiscussed, and results could depend on the pairwise statistic. We added the first two to Section 2, changed the discussion of the third, and added GCM and three further pairwise statistics in the input (Appendix F.7). Certification holds for every input.Resolved
Biological interpretation (mwVD)Rev. thought that the criteria do not establish biologically meaningful mechanisms. We changed Section 4.6 to state that the anatomy makes no claim about neural mechanism. Certification shows only a departure from a pairwise reference that a trained model uses.Resolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.