acceptodds
Under review as a conference paper at ICLR 2027

Beyond Gaussian Assumptions: Distribution-Aware Channel Capacity for Effective Connectivity

10 views

Abstract

Effective-connectivity estimation from brain signals often relies on Gaussian residual modeling, which enables tractable estimation but can discard informative distributional structure and distort inferred directed interactions when empirical residuals are non-Gaussian. We show across multiple modalities, species, and experimental conditions that both brain signals and fitted channel residuals frequently deviate from Gaussianity. We therefore introduce a distribution-aware, information-theoretic measure of effective connectivity based on channel capacity under general residual distributions. To estimate the resulting capacity from empirical, potentially non-Gaussian residuals, we develop a dual-flow min–max estimator based on normalizing flows, in which a generator searches over admissible input distributions under a power constraint while an observer estimates output entropy. We provide a theoretical characterization of the estimator, showing that the observer objective recovers differential entropy up to a KL approximation term, that the formulation reduces to classical Gaussian capacity as a special case, and that residual entropy can alter achievable information rates beyond variance; game-gap and error analyses further characterize optimization and approximation sources. In brain-like simulations with known directed connectivity, Dual-flow achieves the highest AUROC and AUPRC across ten conditions spanning diverse network topologies, hidden drivers, feedback, and heterogeneous hemodynamics, compared with Gaussian capacity, Granger causality, VAR-LiNGAM, and GIMME. Applied to multimodal brain signals, the method reveals time- and condition-resolved directed interactions consistent with known neurobiological circuitry. Together, these results establish a principled distribution-aware framework for effective-connectivity estimation beyond Gaussian residual modeling.

Primary area:
applications to neuroscience & cognitive science
Keywords:
effective connectivity, channel capacity, information theory, non-Gaussian modeling, normalizing flows, min–max optimization, game theory, neuroimaging

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_bAmROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

this work proposed a distribution aware effective connectivity measure based on channel capacity and a dual normalizing flow estimator that models empirical residuals while optimizing the input distribution.

Critical strengths

the comparison to the matched gaussian provides an ablation of whether non-gaussian residual structure matters. The simulations cover hidden drivers, feedback, heterogeneous hemodynamics, and multiple network sizes.

Critical weaknesses

Capacity remains on a locally linear FIR channel with additive noise, so the resulting score doesn't explain the direct neural causation under omitted variables or model misspecification. One needs to restrict the assumptions in settings where nonlinear or temporally structured brain dynamics are important.

Questions for authors
  • How would the conclusions change with multivariate conditioning that explicitly accounts for common drivers?
  • Do authors think this paper is the right for ICLR? or ICLR reviewers have the capacity to judge the correctness in terms of the neuroscience?
3 comments
AuthorsOpenReview ↗

Response to Weakness / Question 1: Channel assumptions and common-driver comparisons

Linear-channel assumption

Capacity remains on a locally linear FIR channel with additive noise.

Response. Thank you. Our FIR model describes relationships at the measured BOLD scale. Nozari et al., Nature Biomedical Engineering (2024) reported that their best-performing linear autoregressive models matched or exceeded the tested nonlinear models in cross-validated one-step-ahead prediction for both resting-state fMRI and resting-state intracranial EEG. Their analysis shows how spatial and temporal averaging can reduce or mask microscopic nonlinearities, providing empirical motivation for linear modeling of macroscopic brain recordings. Such recordings are widely used to study effective connectivity between brain regions (Smith et al., 2011). We have added this empirical motivation in Appendix A.1. GC, VAR-LiNGAM and GIMME also use linear models. Our simulations test recovery after a nonlinear Balloon–Windkessel transformation from neural activity to BOLD. Linearity and Gaussian residuals are separate assumptions; Dual-flow retains the linear channel while modeling residual structure beyond variance. Strongly nonlinear interactions remain a limitation (Appendix C.1).

Temporal structure

One needs to restrict the assumptions in settings where nonlinear or temporally structured brain dynamics are important.

Response. Thank you. As discussed above, linear modeling is motivated by evidence from resting-state fMRI and intracranial EEG. For the continuous LFP–BOLD recordings, we refit the FIR channel and estimate its residual distribution in 50-s sliding windows with 25-s steps, allowing estimated connectivity to vary across the scan (Section 3.4; Appendix D.3). Our simulation benchmark also tests directed-edge recovery with temporally correlated AR(1) measurement noise (ρ = 0.4; Appendix E.1). Strong temporal dependence in the fitted residuals remains a limitation of the current estimator (Appendix C.1).

Causal interpretation

The resulting score doesn't explain the direct neural causation under omitted variables or model misspecification.

Response. Thank you for raising this concern. As already explained in Appendix A.1, effective connectivity (EC) denotes a model-based directed interaction, conditional on the fitted model and observed variables. Shared latent inputs or model misspecification can affect our estimate; it does not by itself establish direct anatomical connectivity or neural causation (Valdés-Sosa et al., 2011). Following established practice in EC analysis, we evaluate recovery of known directed connectivity using the benchmark simulation framework (Smith et al., 2011).

Explicit multivariate conditioning and common drivers

How would the conclusions change with multivariate conditioning that explicitly accounts for common drivers?

Response. Thank you. Our simulations already test robustness to common and hidden drivers. CD-e/CD-u contain observed common drivers; Hidden Diamond omits the driver from analysis; Comp12 combines two hidden drivers, feedback, indirect paths and heterogeneous HRFs. Explicit conditioning could change fitted channels, residuals and edge rankings by accounting for other observed regions; its quantitative effect on Dual-flow remains unmeasured.

We compare against multivariable linear network models: VAR-LiNGAM combines vector autoregression with contemporaneous relations; GIMME estimates group- and individual-level network paths. Both model multiple observed regions jointly. Dual-flow achieves higher mean AUROC and AUPRC than both comparators in these conditions:

ConditionDual-flowVAR-LiNGAMGIMME
Equal common driver.888/.857.725/.739.498/.333
Unequal common driver.858/.830.713/.726.750/.667
Hidden Diamond.910/.844.787/.675.502/.260
Combined confounds (Comp12).879/.724.692/.428.566/.313

Entries are mean AUROC/AUPRC across 50 realizations from Table 5. Expanded lag/pruning and GIMME cutoff searches retain this ranking (Appendix G.1); tuned GIMME improves Comp12 to .599/.356, below Dual-flow's .879/.724.

The simulator defines present and absent edges before estimation. Table 5 reports sensitivity (recovered fraction of true edges), precision (true fraction of detected edges), and FPR (falsely detected fraction of absent edges), with thresholds calibrated on the other 49 realizations. Dual-flow FPRs in these four conditions are .070, .040, .022 and .066. Existing tests therefore quantify both recovery and false positives under these confounds. Explicitly conditioned Dual-flow remains a model extension.

AuthorsOpenReview ↗

Response to Question 2: ICLR relevance, neuroscience/neuroimaging motivation, and applications

Relevance to ICLR

Do authors think this paper is the right for ICLR?

Response. Thank you. Our contribution is a novel learning algorithm for estimating effective connectivity, with neuroscience and neuroimaging providing the motivation and application. The machine-learning contribution is a min–max framework that learns a power-constrained input distribution while fitting flow-based entropy estimators to generated channel outputs and empirical residual samples. The entropy/KL identity, GCap recovery and known-target experiments validate the estimator; simulations assess connectivity recovery against specified ground truth.

Accepted ICLR papers address fMRI effective connectivity, functional-connectivity analysis and functional-network learning:

Broader neural-recording precedents include:

Assessing the neuroscience interpretation

Do ICLR reviewers have the capacity to judge the correctness in terms of the neuroscience?

Response. We provide supporting neuroscience literature, including Manto et al. (2012) for cerebellar motor control and Thompson et al. (2013) for LFP–BOLD correspondence. Our real-data studies test practical relevance; their interpretation rests on this literature. The contribution is methodological rather than a new neuroscience discovery.

In neuroscience, functional connectivity describes statistical dependence; effective connectivity estimates directed influence under a model (Smith et al., 2011). These estimates alone do not establish neural causation or synaptic connections (Friston, 2011; Valdés-Sosa et al., 2011). As discussed in Appendix A.1, our capacity measure is conditional on the fitted model and observed variables. Our HCP analysis concerns directed BOLD interactions, interpreted using the well-established physiology literature cited in the paper.

Reviewer_bAmROpenReview ↗

score increased from 2 to 3

I thank authors for the replies

Reviewer_Csp8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper challenges the common assumption that "residual or noise is approximately Gaussian" in effective connection analysis. In this paper, each directed ROI pair is modeled as a local finite impulse response channel, and the fitting residual is regarded as an empirical sample of unknown continuous noise distribution, and the channel capacity is used as an effective connection metric for distribution perception. In order to estimate the capacity, this paper proposes Dual-flow: the generator searches the input distribution under power constraints, the observer estimates the output entropy with normalized flow, and estimates the noise entropy through residual resampling; This framework can restore Gaussian capacity as a special case. The verification includes ten kinds of brain-like simulations, four predefined ROI analysis of tongue-motion fMRI, and BOLD-BLP cross-modal time-resolved correspondence analysis.

Critical strengths
  1. Cross-modal Gaussian diagnosis has practical methodological value. This paper not only assumes the existence of non-Gaussian, but shows the non-Gaussian and time-varying of signal and fitting residual across species and modes, and reports the correlation between signal-level non-Gaussian statistics and the difference between Gaussian capacity and Dual-flow capacity.
  2. The method has theoretical anchor points, and the coverage of simulation verification is wide. In this paper, the observer's negative logarithmic likelihood is related to the target entropy and approximate error, which shows that Gaussian capacity can be recovered as a special case, and the upper and lower bounds of entropy power are used to demonstrate that the matched variance non-Gaussian residual can still correspond to different achievable information rates.
Critical weaknesses
  1. Real data verification is not enough to support the core proposition that "distributed aware effective connection is superior to existing methods".
  2. The core assumption that "residuals can be regarded as independent samples from stable continuous distribution and uniformly resampled" has tension with short-window and time-related residuals, and lacks critical pressure test.
Questions for authors
  1. Please clarify the rule of edge selection in tongue-motion real data: What is the specific threshold of coefficient of variation below the 30th percentile? Which sides were selected by GCap and Dual-flow respectively?
  2. Please report the finite impulse response order of each ROI pair of tongue-motion, the sample number of residual pool, the convergence and variance of capacity estimation, and the game gap or observer approximate error diagnosis.
  3. The simulation contains time-dependent measurement noise, but Dual-flow resamples the residual independently and uniformly. Have Dual-flow been evaluated under time-dependent residuals?
  4. In the simulation, GIMME is close to random, and GC and VAR-LiNGAM are obviously weak. Please indicate whether a fair superparameter and model order search has been conducted for each baseline, especially considering the recommended range of ROI number and short sequence limitation of GIMME.
4 comments
AuthorsOpenReview ↗

Response to Weakness / Question 1: Contribution and tongue-motion findings

W1 — Contribution and real-data findings

Real data verification is not enough to support the core proposition that "distributed aware effective connection is superior to existing methods".

Response. Thank you for raising this concern. We apologize for any ambiguity in our wording that may have suggested such a proposition. Our contribution is distribution-aware channel capacity as a model-based EC measure and a dual-flow estimator using empirical residual samples and input-distribution optimization, incorporating residual structure beyond Gaussian second-order statistics. The submitted Contributions section describes “superior directed-edge recovery on simulated brain-like signals with ground-truth EC,” rather than general superiority.

Our primary validation of recovery accuracy uses simulated networks with known directed edges, where Dual-flow achieved higher mean AUROC and AUPRC than the evaluated baselines. HCP, which lacks connectivity ground truth, provides a short-window stress test of task sensitivity and physiological plausibility. We have clarified these distinct findings in the Abstract, Contributions, real-data interpretation, and Conclusion.

Q1.A — Selection rule and numerical CV cutoff

Please clarify the rule of edge selection in tongue-motion real data: What is the specific threshold of coefficient of variation below the 30th percentile?

Response. Thank you. Appendix D.2 requires both a task-related increase in a right-tailed paired test across 44 participants (FDR q<0.1) and CV below the 30th percentile of all 12 connections. Table 3 reports each edge's pFDR and CV. The 30th-percentile CV cutoffs are: 3.6685 (Dual-flow), 3.7187 (GCap), 6.3932 (Granger causality), and 3.9541 (VAR-LiNGAM).

Q1.B — Selected directed connections

Which sides were selected by GCap and Dual-flow respectively?

Response. Figure 5 and Appendix Figure 8 show the selected edges; Table 3 provides their statistics:

  • Dual-flow (three edges detected): VI.R→BA3b,OP4.L; BA3b,OP4.R→BA3b,OP4.L; BA3b,OP4.L→BA3b,OP4.R.
  • GCap (four edges detected): VI.R→VI.L; VI.R→BA3b,OP4.R; VI.R→BA3b,OP4.L; VI.L→BA3b,OP4.L.

As described in Section 4.3, paragraph 2, all three edges detected by Dual-flow and the VI.R→BA3b,OP4.L edge detected by GCap correspond to the biologically supported candidate pathways. These connections are bolded above.

AuthorsOpenReview ↗

Response to Weakness / Question 2: Short-window validation and estimation diagnostics

W2 — Short observed residual pools

The core assumption that "residuals can be regarded as independent samples from stable continuous distribution and uniformly resampled" has tension with short-window and time-related residuals, and lacks critical pressure test.

Response. Thank you. We refitted the complete pipeline on independent Chain records (31/62/124 input timepoints; Nmax=15), yielding T=17/48/110 residuals for Dual-flow/GCap. Mean ± SD:

Observed TInput time pointsMethodAUROCAUPRCPrecisionSensitivityFPR
1731Dual-flow0.580±0.1110.422±0.1250.352±0.2760.204±0.1590.121±0.075
GCap0.579±0.1190.406±0.1250.355±0.3330.144±0.1280.081±0.049
VAR-LiNGAM0.497±0.1510.348±0.1300.227±0.3070.108±0.1290.124±0.083
GC0.554±0.1350.401±0.1190.396±0.3900.140±0.1290.075±0.063
GIMMEN/EN/EN/EN/EN/E
4862Dual-flow0.725±0.1200.548±0.1420.535±0.3530.236±0.1700.080±0.071
GCap0.664±0.1090.454±0.1200.384±0.3370.172±0.1510.093±0.074
VAR-LiNGAM0.523±0.1730.375±0.1480.280±0.3310.124±0.1330.111±0.081
GC0.561±0.1320.425±0.1520.449±0.4420.156±0.1530.067±0.059
GIMMEN/EN/EN/EN/EN/E
110124Dual-flow0.855±0.0740.747±0.1050.820±0.2140.356±0.1630.041±0.060
GCap0.717±0.0910.488±0.0910.424±0.2520.220±0.1580.099±0.054
VAR-LiNGAM0.607±0.1720.473±0.1870.480±0.4250.172±0.1760.084±0.088
GC0.544±0.1200.384±0.1150.289±0.3610.116±0.1400.095±0.065
GIMME0.521±0.0760.312±0.0900.320±0.4380.088±0.1220.037±0.052

Dual-flow has the highest mean AUROC/AUPRC at each tested T; both improve with T. Recovery at T=17 remains limited. N/E: GIMME excluded below its recommended 100 observations. Temporal dependence: Q3.

We extend the Gaussian-capacity validation in Figure 6A to shorter observed T, comparing estimated capacity with analytic Gaussian capacity (input length T+Nmax−1):

Capacity errorT=16T=32T=64T=128
Mean±SD−0.3755±0.2160−0.1851±0.1539−0.0611±0.1052−0.0248±0.0643
[Min, Max][−0.8389, 0.2535][−0.4674, 0.1182][−0.2577, 0.1244][−0.1738, 0.1328]

Mean error approaches zero and SD decreases with T. Figure 6A-ii reports Gaussian FIR capacity-estimation bias over 60 runs against analytic capacity.

Q2.A — FIR order and residual-pool size

Please report the finite impulse response order of each ROI pair of tongue-motion and the sample number of residual pool.

Response. Each 12-second trial (TR=0.72 s) has 17 valid residuals. With Nmax=15, we supplied 17+15−1=31 points, including 14 history points; all orders use y(15:31). MATLAB lsqnonneg fits nonnegative FIR coefficients, and AICc selects the order per pair/trial. We corrected “16 points” to “17 points” in the manuscript.

Entries show mean ± SD [min–max].

Selected orders in task data

from\toVI.RVI.LBA3b,OP4.RBA3b,OP4.L
VI.R--1.5167±1.3180 [0-6]1.9833±1.7544 [0-8]2.0167±1.7287 [0-7]
VI.L1.4222±1.4375 [0-9]--2.2167±1.8102 [0-10]2.1611±1.8616 [0-8]
BA3b,OP4.R1.1944±1.0416 [0-5]1.2833±1.0318 [0-5]--1.9833±1.2837 [0-6]
BA3b,OP4.L1.1833±1.0700 [0-5]1.1000±0.8975 [0-5]1.6556±0.9591 [0-5]--

Selected orders in rest data

from\toVI.RVI.LBA3b,OP4.RBA3b,OP4.L
VI.R--0.7944±1.0394 [0-7]1.2667±1.7097 [0-10]1.2056±1.4862 [0-6]
VI.L0.8722±1.0936 [0-6]--1.1833±1.4396 [0-9]1.2389±1.3593 [0-7]
BA3b,OP4.R0.7778±1.0547 [0-6]0.8667±1.0954 [0-6]--1.4000±1.1704 [0-6]
BA3b,OP4.L0.6778±0.8565 [0-4]0.8833±1.0099 [0-8]1.5611±1.3502 [0-9]--

Q2.B — Convergence, estimation variability, and observer-approximation error

Please report the convergence and variance of capacity estimation, as well as the game gap or observer approximate error diagnosis.

Response.

1. Convergence. We use exponential moving averages EMA(CC) and EMA(CC−EMA(CC)), both with factor 0.9, and stop when the latter’s absolute value is <10^−6. The HCP example below meets this threshold at iteration 1527 (−4.54e−7).

2. Estimation bias and variability. Figure 6A-ii reports capacity-estimation bias and variability over 60 runs using simulated Gaussian FIR channels with known analytic capacities. For HCP motor-task data, absolute bias cannot be assessed because the true capacity is unknown.

3. Observer-approximation error. Figure 6A-i reports observer entropy-estimation error over 60 runs using distributions with known entropy. HCP example:

iterCCEMA(CC)e^=\hat{e}= CC-EMA(CC)EMA(e^\hat{e})
1000.747024.742724.004300-.004720
1100.765983.753976.012007.005790
1200.767941.773500-.005559.000207
1300.790945.795314-.004368-.001143
1400.82334.810266.013074-.000041
1500.823948.819271.004677-0.000789
1527.822059.837805-.01575-4.54e-7
AuthorsOpenReview ↗

Response to Question 3 / Question 4: Temporal dependence and baseline comparisons

Q3 — Evaluation with time-dependent noise and residuals

The simulation contains time-dependent measurement noise, but Dual-flow resamples the residual independently and uniformly. Have Dual-flow been evaluated under time-dependent residuals?

Response. Yes. Our original simulations evaluated the complete pipeline—including FIR fitting and independent residual resampling—with temporally correlated AR(1) measurement noise (ρ=0.4). Dual-flow achieved the highest mean AUROC and AUPRC for directed-edge recovery among the methods evaluated in each of the ten conditions (Table 5). We also examined the fitted HCP residuals, whose lag-1 autocorrelation was 0.1857±0.2814 (mean±SD). Independent resampling does not preserve residual time order; the simulation results provide empirical evidence of recovery performance with this approximation under the tested correlated-noise conditions.

Q4.A — Original baseline results and settings

In the simulation, GIMME is close to random, and GC and VAR-LiNGAM are obviously weak.

Response. Thank you. For Granger causality, BIC selected target and source orders up to our original maximum lag of 4. BIC balances fit against model complexity. The Chandler implementation uses BIC, and Smith et al. (2011) also evaluated BIC-selected pairwise Granger causality. Maximum lag 4 was our study-specific cap.

For VAR-LiNGAM, we used lags=1, criterion="bic", and prune=False. Lag 1 and BIC match the documented defaults; disabling pruning retained the unpruned connection coefficients. Pruning is an additional sparse coefficient-estimation step, so we explicitly compare both choices below. For GIMME, ar=TRUE, subgroup=FALSE, and groupcutoff=.75 match the documented standard settings, with group-level path search followed by individual-level estimation.

Q4.B — Fair parameter searches and updated comparisons

Please indicate whether a fair superparameter and model order search has been conducted for each baseline.

Response. We tested maximum lags {1,2,3,4,10,20} for corrected Granger causality and VAR-LiNGAM, retaining BIC order selection; VAR-LiNGAM was tested with and without pruning. The grid includes Smith et al.'s pairwise BIC search limits {1,2,3,10,20} and our original cap 4. These are search limits, rather than fixed fitted orders. For 300-point, 28-ROI Macq28, VAR-LiNGAM orders were restricted to the admissible range 1–9 when larger limits were requested.

For GIMME, we varied groupcutoff over {.50,.75,1.00}, keeping the other settings fixed. All individual fits across the three cutoff settings converged. Each evaluated condition used 50 realizations.

Tuning improves GIMME on Chain (HRF): reducing its cutoff from .75 to .50 raises AUROC/AUPRC from .605/.409 to .900/.850. Dual-flow achieves .939/.891, with paired gains of .039 [95% CI .023,.054] and .041 [.015,.065]. After the expanded searches, Dual-flow retains the highest mean AUROC and AUPRC across all ten tested conditions; GIMME was evaluated in nine conditions.

Revised Table 5 and Figure 4 report corrected GC at each condition’s highest-mean-AUROC setting, with all five metrics taken from that setting. Appendix G, Tables 6–9, reports the full lag, pruning and cutoff searches; Table 10 reports paired comparisons against GC, GCap and GIMME. Selection and evaluation use the same 50 realizations, and the confidence intervals condition on the selected settings and saved fits. These comparisons describe performance within the tested parameter grids.

Q4.C — GIMME applicability

Please consider the recommended range of ROI number and short sequence limitation of GIMME.

Response. GIMME's nine evaluated conditions contain 3–10 ROIs and 300 observations per realization, within the developer's reported 3–20-variable range and the FAQ's recommendation of at least 100 observations. Macq28 and the short HCP segments were excluded.

Thank you again for your thoughtful review. As the discussion period draws to a close, we would appreciate it if you could take our additional results and analyses into account in your final assessment. If any questions remain, we would be happy to clarify them.

Reviewer_1W1JOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper fits a directional FIR channel to each ROI pair and estimates its input-power-constrained capacity using empirical residual samples. A normalizing-flow generator optimizes the input distribution, while output and residual observer flows estimate an entropy difference. The theory recovers Gaussian-channel capacity and explains how residual entropy can affect achievable rates beyond variance. Controlled checks test entropy/capacity estimation, ten brain-network conditions test directed-edge ranking, and applications examine HCP tongue-task fMRI and rat BOLD-LFP correspondence. Dual-flow reports the highest mean AUROC and AUPRC in all ten network conditions.

Critical strengths
  1. A clear distribution-aware comparison within a fixed channel model

The comparison with Gaussian capacity within the same fitted channel is a good way to ask what residual distribution shape contributes. The entropy/KL identity makes the observer's approximation error explicit, and the Gaussian special case gives an analytic reference. The paper also distinguishes the population formulation from its finite-sample approximation and states the limits of pairwise causal interpretation (pp. 15, 20–21). I find this framing useful because it identifies a quantity that can be checked under known channels, rather than relying entirely on anatomical plausibility.

  1. Substantial controlled validation

The known-entropy and analytic Gaussian FIR checks are meaningful validation steps. The ten network conditions, each with 50 realizations, probe hidden drivers, feedback, heterogeneous HRFs and increased network size. Dual-flow has the best mean AUROC and AUPRC in those reported comparisons. The participant-level task tests and scan-clustered cross-modal bootstrap also use appropriate units of inference. The code now makes the generator/observer training concrete, including independent residual resampling, a unit power budget and EMA stopping. That detail helps separate a concern about practical estimation from the merit of the population objective.

Critical weaknesses
  1. The finite-residual estimator has not been validated in the regime of the real application

The HCP task/rest analysis fits each pairwise channel to only 16 time points (p. 26). trainDualNet.m then samples individual residual values independently with randi and reuses this pool at every iteration (lines 29–34). estimateCapacity.m generates length-1,024 sequences in batches of 256, but those sequences still contain draws from the same small observed residual pool. Varying generated length L therefore tests optimization sensitivity, not the information available to estimate a continuous residual entropy. The paper's known-distribution checks and network rankings do not establish capacity bias or null-edge behavior at this observed length.

The release adds a configuration issue that matters especially for short segments. nnls_vec.m fixes Y = y(Nmax:end) before selecting an order, so every candidate uses T-Nmax+1 observations rather than the paper's order-dependent residual length. The demo uses Nmax = 15; applying that setting to 16 points would leave two observations. I need the actual application lag limits and guards for AICc and rank before interpreting the stress test. The fit also uses nonnegative least squares, whose constraint should be specified when interpreting directional channels. Finally, independent scalar resampling discards residual temporal dependence. The observed BOLD is filtered and some network conditions contain AR(1) measurement noise, so residual-whiteness and input-dependence diagnostics are relevant to the channel assumptions.

  1. The released GC statistic cannot support the comparison without correction

granger_cause.m specifies the previously unclear output: it returns an F statistic. Its restricted model uses x(x_lag+1:T), whereas the unrestricted model uses x(x_lag+y_lag+1:T) (lines 72–89 and 108–129). Their residual sums of squares are then subtracted in the F numerator (line 138). A nested-model F comparison requires the same response observations in both fits. The denominator also uses T-(x_lag+y_lag+1), while the unrestricted design has T-x_lag-y_lag rows and 1+x_lag+y_lag columns. Its residual degrees of freedom are consequently T-2(x_lag+y_lag)-1 when the design is full rank. The implementation has no corresponding rank or short-series guard. The returned quantity consequently lacks the standard nested-model F-test interpretation.

I need to know whether this function produced the approximately 10^30 GC values in Figure 8 and the reported GC comparisons. Those comparisons should be repeated with a common observation set, admissible lags and correct degrees of freedom. The release also calculates Gaussian capacity using log2 while the neural entropy difference uses natural logarithms; absolute capacity comparisons need an explicit unit conversion. That conversion would not change within-method edge rankings. The absent experiment driver makes it difficult to trace these choices to the paper. Corrected baselines and a direct paired comparison are needed before significance-threshold crossings can support a superiority claim.

Questions for authors
  1. What exact Nmax, selected orders and valid residual counts were used for the 16-point HCP segments? Please validate the complete fitted-channel pipeline as observed T increases, including capacity bias/variance and null-edge behavior. This needs independent observed residual pools, rather than only longer resampled training sequences.

  2. Was the bundled granger_cause.m used for Figure 8 or the network benchmark? If so, how do the comparisons change after fitting both models on identical observations, enforcing full rank and positive residual degrees of freedom, and correcting the F denominator? Please report paired participant/scan-level differences against the corrected GC and matched Gaussian-capacity baselines.

  3. Please provide the experiment driver that produced the reported results and explain how it handles nonnegative FIR fitting, order-dependent residual lengths, capacity units and stopping settings. What do residual autocorrelation/source-dependence diagnostics show, and does the result persist with an appropriate dependence-preserving or prewhitened residual model?

4 comments
AuthorsOpenReview ↗

Response to Weakness / Question 1: FIR fitting and short-window validation

Q1.A — Fitting window and selected orders

What exact Nmax, selected orders and valid residual counts were used for the 16-point HCP segments?

Response. We corrected “16” to 17 points. Each 12-s trial (TR=.72 s; Nmax=15) uses 31 points, including 14 history points. Aligning x(t−14:t) with y(t), t=15:31, gives all orders 17 residuals. lsqnonneg fits nonnegative FIR coefficients; AICc selects each pair/trial’s order.

Entries show mean ± SD [min–max].

Task

from\toVI.RVI.LBA3b,OP4.RBA3b,OP4.L
VI.R--1.5167±1.3180 [0-6]1.9833±1.7544 [0-8]2.0167±1.7287 [0-7]
VI.L1.4222±1.4375 [0-9]--2.2167±1.8102 [0-10]2.1611±1.8616 [0-8]
BA3b,OP4.R1.1944±1.0416 [0-5]1.2833±1.0318 [0-5]--1.9833±1.2837 [0-6]
BA3b,OP4.L1.1833±1.0700 [0-5]1.1000±.8975 [0-5]1.6556±.9591 [0-5]--

Rest

from\toVI.RVI.LBA3b,OP4.RBA3b,OP4.L
VI.R--.7944±1.0394 [0-7]1.2667±1.7097 [0-10]1.2056±1.4862 [0-6]
VI.L.8722±1.0936 [0-6]--1.1833±1.4396 [0-9]1.2389±1.3593 [0-7]
BA3b,OP4.R.7778±1.0547 [0-6].8667±1.0954 [0-6]--1.4000±1.1704 [0-6]
BA3b,OP4.L.6778±.8565 [0-4].8833±1.0099 [0-8]1.5611±1.3502 [0-9]--

W1.A — Fixed versus order-dependent residual counts

nnls_vec.m fixes Y = y(Nmax:end) before selecting an order, so every candidate uses T-Nmax+1 observations rather than the paper's order-dependent residual length.

Response. See Q1.A for the common response window.

W1.B — AICc and rank checks

I need the actual application lag limits and guards for AICc and rank before interpreting the stress test.

Response. Thank you.

  1. Lag limits. Nh=0–15; 97% of selected orders were ≤5 (Q1.A).
  2. AICc for short windows. AICc corrects finite-sample bias; BIC approximates model evidence asymptotically (Burnham and Anderson, 2004). In the regression example of Hurvich and Tsai (1989, Table 1), AICc selected the correct model in 96/100 runs with 10 observations. We verified parameter counts and positive AICc denominators at n=17.
  3. Rank checks. With Nh lagged predictors, the row rank and column rank of X are equal and at most Nh. We center the signals instead of introducing an intercept. Since T=17 and Nmax=15, rank(X) ≤ Nh ≤ 15 < T. All evaluated fits had positive residual sums of squares.

Q1.B / W1 — Short-window performance

Please validate the complete fitted-channel pipeline as observed T increases, including null-edge behavior. This needs independent observed residual pools, rather than only longer resampled training sequences.

Response. We refitted each method on independent 31/62/124-point Chain records (TR=1 s). Nmax=15 gives Dual-flow/GCap T=17/48/110 observed residuals.

Observed TInput time pointsMethodAUROCAUPRCPrecisionSensitivityFPR
1731Dual-flow.580±.111.422±.125.352±.276.204±.159.121±.075
GCap.579±.119.406±.125.355±.333.144±.128.081±.049
VAR-LiNGAM.497±.151.348±.130.227±.307.108±.129.124±.083
GC.554±.135.401±.119.396±.390.140±.129.075±.063
GIMMEN/EN/EN/EN/EN/E
4862Dual-flow.725±.120.548±.142.535±.353.236±.170.080±.071
GCap.664±.109.454±.120.384±.337.172±.151.093±.074
VAR-LiNGAM.523±.173.375±.148.280±.331.124±.133.111±.081
GC.561±.132.425±.152.449±.442.156±.153.067±.059
GIMMEN/EN/EN/EN/EN/E
110124Dual-flow.855±.074.747±.105.820±.214.356±.163.041±.060
GCap.717±.091.488±.091.424±.252.220±.158.099±.054
VAR-LiNGAM.607±.172.473±.187.480±.425.172±.176.084±.088
GC.544±.120.384±.115.289±.361.116±.140.095±.065
GIMME.521±.076.312±.090.320±.438.088±.122.037±.052

N/E: GIMME excluded below its recommended ≥100 observations. FPR=FP/(FP+TN): fraction of null edges falsely detected as connections.

Dual-flow has the highest mean AUROC/AUPRC at each tested T; both improve with T. HCP complements this ground-truth validation with a short-window stress test of task sensitivity across 44 participants.

Q1.C / W1 — Capacity bias and variance

Please validate the complete fitted-channel pipeline as observed T increases, including capacity bias/variance.

Response. We extend the Gaussian-capacity validation in Figure 6A to shorter observed lengths T={16,32,64,128}, comparing estimated capacity with analytic Gaussian capacity (input length T+Nmax−1). Error=estimated capacity - analytic Gaussian-channel capacity.

ErrorT=16T=32T=64T=128
Mean±SD−.3755±.2160−.1851±.1539−.0611±.1052−.0248±.0643
[Min, Max][−.8389, .2535][−.4674, .1182][−.2577, .1244][−.1738, .1328]

The mean error (bias) approaches zero, and the SD decreases as T increases.

AuthorsOpenReview ↗

Response to Weakness / Question 2: Granger causality correction and updated results

Q2.A / W2 — GC implementation and correction

Was the bundled granger_cause.m used for Figure 8 or the network benchmark? If so, how do the comparisons change after fitting both models on identical observations, enforcing full rank and positive residual degrees of freedom, and correcting the F denominator?

Response. Thank you for identifying these issues. We used the bundled granger_cause.m, based on Chandler’s implementation from MATLAB Central for all ten simulations and HCP Figure 8. Published uses include Smith et al. (2011) and Ji et al. (2016).

We confirmed both defects and apologize for overlooking them. At target/source orders p,q, restricted and unrestricted responses began at p+1 and p+q+1, giving residual sums of squares (RSS) on different observations. The denominator used T−(1+p+q), not the actual unrestricted residual degrees of freedom.

Correction: all BIC candidates use M+1:T, where M is the maximum lag. After selecting p,q, both nested fits use max(p,q)+1:T, sharing N=T−max(p,q) observations. With k_U=1+p+q coefficients including the intercept, ν=N−k_U and

F=[(RSS_R−RSS_U)/q]/(RSS_U/ν).

The critical value uses the same ν. The code checks T−M>1+2M, requires full rank and positive residual degrees of freedom, and excludes invalid candidates. It rejects RSS_U≤0 and materially negative RSS_R−RSS_U; only roundoff-sized negative F values are reset to zero.

Q2.B / W2 — Corrected simulation comparisons

Those comparisons should be repeated with a common observation set, admissible lags and correct degrees of freedom.

Response. To isolate the correction, we retained BIC and maximum lag 4 and reran all ten conditions, with 50 realizations each:

ConditionAUROC: original → correctedAUPRC: original → corrected
CD-e.580 → .570.572 → .557
CD-u.610 → .608.576 → .586
Diamond.607 → .612.453 → .456
Diamond (H).589 → .619.494 → .516
Chain.603 → .609.428 → .445
Chain (HRF).752 → .757.631 → .654
Feedback.514 → .514.495 → .498
Mod10.630 → .636.266 → .276
Comp12.736 → .743.483 → .490
Macq28.679 → .679.145 → .144

Mean GC changes are at most .030 AUROC and .023 AUPRC. Dual-flow remains highest in both across all ten conditions. Its paired gains over GC at maximum lag 4 are .136–.376 AUROC and .234–.539 AUPRC; all pointwise 95% bootstrap intervals are positive.

Table 5/Figure 4 use each condition’s best-mean-AUROC GC setting from maximum lags {1,2,3,4,10,20}, reporting all five metrics from that setting.

Q2.C / W2 — Corrected HCP GC results

I need to know whether this function produced the approximately 10^30 GC values in Figure 8 and the reported GC comparisons.

Response. Yes. In 444/1,056 fits, T=17 and p=q=4 gave nine observations for nine coefficients: zero residual degrees of freedom. Dividing near-zero RSS_U by an incorrect denominator of 8 produced GC scores near 10^30. Corrected fits at these orders use 13 observations and ν=4. Across corrected fits, ν=4–13 and F=6.772×10^(−6)–35.1928. None of the 12 pairs shows a significant task–rest increase (FDR-adjusted p=.7144–.9281; required q<.1).

Q2.D — Direct paired comparisons

Please report paired participant/scan-level differences against the corrected GC and matched Gaussian-capacity baselines.

Response. EC is a model-based readout; raw GC F statistics and capacity estimates are not directly comparable.

1. Simulation comparisons. Following established practice in EC analysis, we validate EC estimates through recovery of known directed edges in simulations (Smith et al., 2011). Appendix G.2 compares AUROC/AUPRC between methods paired on the same simulated realization.

2. HCP task sensitivity. HCP provides a short-window stress test of task sensitivity with 17 fitted observations per trial. Table 3/Figure 8 report edges showing significant task–rest increases across 44 participants and meeting each method's CV criterion. In response to your request, we added Table 11 (Appendix G.3) to directly compare participants' task–rest changes between Dual-flow and GCap in common capacity units. The paired 95% CIs include zero for all 12 possible directed connections among the four ROIs.

3. BOLD–LFP correspondence. Within each method, we correlate BOLD-derived and LFP-band-power-derived EC across matched windows to assess the neural–BOLD correspondence supported by the cited physiology literature (Appendix D.3). At your request, Appendix G.4 reports mean ± SD across 18 matched windows per scan, side by side in each method’s units. As noted above, raw values in different units are not directly comparable.

AuthorsOpenReview ↗

Response to Question 3: Experiment driver, residual diagnostics and dependence-preserving comparison

Q3.A — Driver, units and stopping

Please provide the experiment driver that produced the reported results and explain how it handles nonnegative FIR fitting, order-dependent residual lengths, capacity units and stopping settings.

Response. We included the matlab scripts to produce the results and figures in the supplemetary materials for your review. We used matlab function lsqnonneg in nnls_vec.m to fit nonnegative FIR coefficients. As we explained in Block 1, Q1.A, the residual lengths are fixed at 17 and independent of the selected order.

We have converted all units to nats in our analysis: GCap in bits × ln(2) gives values in nats. The comparaison and the rankings were all correct. To avoid any confusion, we updated the GCap calculation in nnls_vec.m using natural log.

trainDualNet.m tracks a capacity EMA (factor .9) and a second EMA (.9) of the latest estimate minus the capacity EMA. The training stops when the second EMA’s absolute value is < 10^−6. This setting helps to ensure that the final result is relatively stable against the resampling of noise samples and the oscillation due to alternative gradients.

Q3.B — Residual diagnostics

What do residual autocorrelation/source-dependence diagnostics show?

Response. HCP tongue diagnostics (mean ± SD):

Lag012345678910111213141516
Residual autocorrelation1±0.1857±.2814−.0164±.2555−.0366±.2144−.0356±.2083−.0742±.2007−.1063±.1907−.0899±.1876−.0741±.1819−.0693±.1737−.0491±.1647−.0376±.1516−.0349±.1382−.0243±.1244−.0177±.1120−.0133±.0906-0.0064±0.0659
Source–residual correlation−.0086±.0431−.0140±.0582−.0098±.0434−.0097±.0381−.0033±.0183−.0046±.0309−.0212±.0752−.0200±.0529.0000±.0000−.0000±.0000———————

Lag-0 autocorrelation is 1 by definition; — means unreported beyond selected Nh≤10.

Absolute mean source–residual correlations are ≤.0212. Residual autocorrelation is .1857±.2814 at lag 1; absolute means at lags 2–16 are ≤.1063.

Q3.C — Dependence-preserving comparison

Does the result persist with an appropriate dependence-preserving or prewhitened residual model?

Response. We appreciate this suggestion. Our current estimator independently resamples fitted residuals. Under temporally correlated AR(1) measurement noise (ρ=0.4), Dual-flow achieved the highest mean AUROC and AUPRC for directed-edge recovery among the methods evaluated in each of the ten conditions (Table 5). Evaluating whether these results persist with dependence-preserving or prewhitened residual models remains future work.

Thank you again for your thoughtful review. As the discussion period draws to a close, we would appreciate it if you could take our additional results and analyses into account in your final assessment. If any questions remain, we would be happy to clarify them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.