{"id":"c29ba04a-705d-42de-8d73-ef3e5ea2b058","arxiv_id":"2603.07634","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PDGC decomposes spectral multivariate Granger causality into unique, redundant, and synergistic components and reveals distinct high-order causal patterns under tilt in syncope patients versus controls.","lead":"The paper introduces Partial Decomposition of Granger Causality (PDGC), which splits multivariate spectral Granger causality into unique, redundant, and synergistic causal parts using partial information decomposition and state-space models. It then uses those parts to show that people prone to neurally mediated syncope respond differently to postural stress than healthy controls in cardiovascular and cerebrovascular control.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The min-redundancy definition (Eq. 13) is the load-bearing untested choice that can re-label unique/redundant/synergistic atoms even when full spectral GC is correct.","rationale":"The Reader correctly isolates the min-redundancy definition (Eq. 13) together with the linear SS/VAR backbone as the weakest assumption. That choice is load-bearing: every unique/redundant/synergistic atom and every band-limited physiological contrast is obtained from it via Möbius inversion and spectral integration. The abstract’s claim of benchmark-simulation validation is absent from the body, so there is no independent check that the atoms recover known causal mechanisms. The n=13 per group and missing code/data further limit strength of the clinical claims, but they are secondary to the definitional dependence. Because the mathematical construction remains coherent and the method is still useful once the missing validation and a sensitivity check are supplied, the Reader’s CONDITIONAL verdict is the right one; no stronger rejection is warranted. The concrete test above directly probes whether the reported group differences survive a change of redundancy function—the single check that would settle the concern.","tokens_in":16419,"tokens_out":647,"duration_ms":6040,"concrete_test":"Re-compute the coarse-grained LF and whole-band unique/redundant/synergistic GC values of Figs. 3–4 after replacing the min operator in Eq. 13 with an alternative spectral redundancy (e.g., the average of the subset spectral GCs, or a frequency-domain analogue of I_MMI). If any of the statistically significant REST-vs-HUT contrasts (unique SAP\to HP in nonSYNC; redundant/synergistic {MAP,HP}\to MCBV in SYNC) reverse or lose significance, the physiological claims are definition-dependent and the central claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on the spectral PID of multivariate GC (Eqs. 12–17) obtained by defining the redundant GC of an atom as the pointwise minimum of the spectral GCs of its source subsets (Eq. 13) and recovering atoms by Möbius inversion. This is a modeling choice, not a theorem: different redundancy functions (e.g., I_min vs. other PID candidates) can re-partition the same full GC into different unique/redundant/synergistic values. The manuscript never shows that the physiological conclusions (tilt-induced unique SAP\to HP rise in controls; LF redundant/synergistic MAP/HP\to MCBV rise in SYNC) are invariant under a different redundancy function, nor does it supply the benchmark simulations promised in the abstract that would demonstrate the atoms recover known ground-truth mechanisms. Because all subsequent coarse-graining (Fig. 1, Eq. 16) and band-limited physiological interpretations inherit this definition, a mis-specified min can produce the reported group differences even when the underlying linear spectral GC is correctly estimated. Linear VAR/SS assumptions compound the risk for short physiological series, but the min choice is the more fundamental soft spot.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces Partial Decomposition of Granger Causality (PDGC), which embeds multivariate spectral Granger causality (GC) computed from state-space models into the partial information decomposition (PID) framework. Multivariate GC from a set of drivers X to a target Y is dissected into unique, redundant, and synergistic atoms by defining a spectral redundant GC as the pointwise minimum of the spectral GCs of the atom’s source subsets (Eq. 13), recovering atoms by Möbius inversion (Eqs. 14–15), and coarse-graining into unique/redundant/synergistic components (Eqs. 16–17). Whole-band and band-limited (LF/HF) integrals recover time-domain measures. The method is applied to cardiovascular (RESP, SAP → HP) and cerebrovascular (RESP, MAP, HP → MCBV) networks in syncope patients versus controls at rest and during head-up tilt, reporting blunted unique/redundant SAP\to HP responses in patients and elevated LF redundant/synergistic MAP/HP\to MCBV effects that are interpreted as markers of autonomic and autoregulatory dysfunction.","tokens_in":16705,"tokens_out":1386,"duration_ms":11584,"significance":"If the construction is sound, PDGC supplies a directed, frequency-resolved high-order causality tool that is more interpretable than pairwise or conditional GC alone and is computationally convenient via closed SS reduced models. The physiological application is of genuine interest for network physiology: it refines known baroreflex and cerebral-autoregulation findings by attributing tilt responses to unique versus higher-order components in specific bands. Strengths include the SS formulation that avoids infinite-order VAR truncation, the spectral-integration property that links time- and frequency-domain PIDs, and the use of surrogate testing plus non-parametric group comparisons. The abstract’s claim of benchmark validation, if present and rigorous, would further strengthen the contribution for data-driven network science.","major_comments":[{"comment":"Abstract and §I promise “validation on benchmark simulations” showing that unique/redundant/synergistic GC “reflect the underlying causal mechanisms and are computationally reliable.” No such simulations appear in the manuscript (Methods, Results, or appendices). Without ground-truth recovery experiments (e.g., known redundant/synergistic VAR or SS networks), the central claim that the atoms correctly dissect causal mechanisms remains untested.","section":"Abstract / §I"},{"comment":"Eq. (13) defines spectral redundant GC as the pointwise minimum of the spectral GCs of the atom’s source subsets. This is a modeling choice (I_min-style), not a theorem. Different redundancy functions can re-partition the same full GC into different unique/redundant/synergistic values. The manuscript never demonstrates that the reported physiological conclusions (tilt-induced unique SAP\to HP rise in controls; LF redundant/synergistic MAP/HP\to MCBV rise in SYNC) are invariant under an alternative redundancy definition, nor does it compare against other PID candidates. Because all coarse-graining (Fig. 1, Eq. 16) and band-limited interpretations inherit this definition, sensitivity analysis is load-bearing.","section":"§IIB, Eq. (13)"},{"comment":"All spectral GC atoms are obtained from linear Gaussian VAR/SS models of short (N=250) physiological series. The manuscript does not assess robustness to mild nonlinearity, non-Gaussian innovations, or model-order misspecification, all of which are common in cardiovascular/cerebrovascular data. If the linear assumption fails, the unique/redundant/synergistic split can be distorted even when full multivariate GC is nonzero. At minimum, a brief residual diagnostic or comparison with a nonlinear GC baseline on the same data would be needed to support the physiological claims in §IIID.","section":"§IIC / §IIIB"}],"minor_comments":[{"comment":"Figure numbering is inconsistent: the CV results are introduced as “Fig. 4” but the caption and subsequent text refer to Fig. 3; CB results are also labeled Fig. 4. Correct the labels and cross-references.","section":"§IIIC"},{"comment":"The abstract states that validation “demonstrates … computational reliability,” yet no numerical stability, bias, or variance results are reported. Either add a short computational-reliability subsection or soften the abstract claim.","section":"Abstract"},{"comment":"Notation for the restricted process Z and the Cholesky-adjusted transfer function H̃ versus H is dense; a short table of symbols or an expanded sentence after Eq. (20) would help readers implement the method.","section":"§IIA–IIC"},{"comment":"Surrogate significance is assessed at the single-subject level with IAAFT (100 surrogates, 95th percentile). Clarify whether multiple-comparison correction across atoms/bands was considered, or state that none was applied.","section":"§IIIB"},{"comment":"Several self-citations to the authors’ PIRD and HOI papers are appropriate for the building blocks, but a brief explicit comparison of PDGC versus SURD and PIRD (beyond the one-sentence claim in §IV) would better position the novelty.","section":"§IV"}],"recommendation":"major_revision","confidential_remarks":"The core construction is a natural and useful marriage of spectral GC and PID; the main obstacle to acceptance is the missing benchmark validation promised in the abstract and the lack of sensitivity checks on the min-redundancy choice. If the authors can supply those two items (even in a short appendix) the paper would be a solid contribution for a methods-oriented journal. Heavy self-citation is noticeable but not disqualifying given the authors’ prior work on the same machinery."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean methods paper that does something useful: it takes multivariate spectral Granger causality, embeds it in a PID lattice via a pointwise min redundancy, recovers unique/redundant/synergistic atoms by Möbius inversion, and keeps the whole thing consistent under whole-band and band-limited integration. The state-space route (Barnett–Seth style) is the right engineering choice; it avoids the infinite-order reduced-VAR mess and makes the spectral GC well-defined. That construction is coherent, and the coarse-graining rules they adopt are already in the literature they cite.\n\nWhat is actually new is the explicit spectral PID of GC itself (their PDGC), not the individual ingredients. The authors’ own prior PIRD and HOI work supplies the scaffolding; this paper specializes it to directed spectral GC and applies it to a real CV/CB tilt dataset. The physiology is readable: controls show the expected tilt-driven rise in unique SAP→HP (and some redundancy), while SYNC patients show blunted CV response and elevated LF redundant/synergistic MAP/HP→MCBV. That pattern is more informative than full multivariate GC alone, and it matches known baroreflex/CA stories without forcing them.\n\nSoft spots, in proportion. The abstract promises benchmark simulations that “demonstrate the measures reflect underlying causal mechanisms”; they are not in the manuscript body. That is a real gap for a methods claim. The min-redundancy definition (Eq. 13) is a modeling choice, not a uniqueness theorem; different PID candidates can re-partition the same full GC. They do not show invariance of the group differences under an alternative redundancy function. Linear VAR/SS on short physiological series (n=13 per group) is standard for this literature but still limits how hard one can push clinical language. No code or data release. None of these sink the central construction; they just keep the paper conditional rather than finished.\n\nMath and citation pattern look solid: Geweke spectral GC, Cholesky handling of instantaneous effects, SS reduced models, and PID lattice are used correctly. Self-citation is heavy but points to building blocks, not circular proof of the syncope result.\n\nThis is for people who already do spectral GC or high-order information dynamics on oscillatory physiological networks. It deserves a serious referee. I would engage with it, cite the method when I need band-specific unique/redundant/synergistic GC, and ask for the missing simulations and a sensitivity check on the redundancy function.","headline":"Solid spectral PID of multivariate GC that organizes known pairwise/conditional intuitions and yields clean tilt contrasts; the min-redundancy choice is a modeling decision, not a theorem, and the promised simulations are missing from the body.","tokens_in":17365,"tokens_out":605,"would_cite":true,"duration_ms":5885,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Partial Decomposition of Granger Causality splits multivariate spectral causality into unique, redundant and synergistic atoms that distinguish syncope patients from healthy controls under postural stress.","keywords":["Granger causality","partial information decomposition","spectral analysis","high-order interactions","state-space models","network physiology","cardiovascular control","cerebrovascular autoregulation"],"falsifier":"On a controlled linear Gaussian network whose ground-truth unique, redundant and synergistic causal strengths are known, compute PDGC; if the recovered atoms deviate systematically from those known strengths, or if the same atoms change sign or disappear under mild non-Gaussian driving noise while full GC remains unchanged, the decomposition is falsified.","tokens_in":17277,"feed_emoji":"🫀","tokens_out":986,"duration_ms":9286,"temperature":0.7,"pith_summary":"The paper introduces Partial Decomposition of Granger Causality (PDGC), a method that takes the familiar multivariate Granger causality from several driver time series to a target and splits it into unique, redundant and synergistic pieces. It does so by embedding frequency-domain Granger measures inside the partial-information-decomposition lattice, defining redundancy at each frequency as the minimum of the drivers’ spectral causalities and recovering the atoms by Möbius inversion. Because the construction lives in the spectral domain, the same atoms can be integrated over any physiologically meaningful band or over the whole spectrum. On cardiovascular and cerebrovascular series the resulting unique, redundant and synergistic components show opposite responses to head-up tilt in healthy subjects versus patients prone to neurally-mediated syncope, revealing high-order control patterns that ordinary pairwise or full multivariate Granger causality leave opaque. The method therefore supplies a practical, frequency-resolved language for high-order directed interactions among oscillatory processes.","feed_headline":"Spectral Granger causality split into unique, redundant, synergistic atoms","feed_subtitle":"The split shows opposite postural-stress responses in syncope patients versus healthy controls","key_machinery":"The spectral redundant Granger causality f^∩_{X_α→Y}(ω) := min_j f_{X_αj→Y}(ω), together with its full-band integral and the subsequent Möbius inversion that yields the atomic GCs; the construction is realized by state-space models of the VAR process so that reduced models remain exact.","core_discovery":"Multivariate spectral Granger causality from a set of drivers to a target can be decomposed, via a pointwise min-redundancy function and Möbius inversion on the PID lattice, into non-negative unique, redundant and synergistic spectral atoms that integrate exactly to the classical time-domain Granger measures; these atoms, when evaluated on physiological series, expose distinctive high-order causal reorganizations under postural stress that separate syncope patients from matched controls.","pith_inferences":["The same spectral-min construction could be applied to other directed spectral measures (e.g., directed transfer function or partial directed coherence) to obtain analogous unique/redundant/synergistic decompositions.","If the linear-Gaussian assumption is relaxed by replacing the state-space GC with a model-free spectral transfer entropy, the resulting atoms would test whether the physiological findings survive nonlinear interactions.","The opposite tilt responses of CV versus CB high-order atoms suggest that a joint PDGC of the full five-variable network might reveal compensatory loops that are invisible when the two sub-networks are analysed separately."],"forward_implications":["Any existing multivariate spectral GC analysis can be re-run with PDGC to report unique, redundant and synergistic contributions band by band.","In network physiology the unique SAP→HP and the redundant/synergistic MAP/HP→MCBV components become candidate biomarkers of baroreflex and cerebral-autoregulation dysfunction under orthostatic stress.","Because the atoms integrate exactly across frequency, the same decomposition can be restricted to any a-priori band of interest without recomputing the underlying models.","The method supplies a directed, frequency-resolved counterpart to existing undirected high-order information measures, enabling comparison of pairwise versus multi-body causal routes in oscillatory networks."],"fun_headline_variants":["Spectral GC dissected into unique, redundant, synergistic atoms via PID","PDGC extracts high-order unique redundant synergistic causal atoms","Unique redundant synergistic GC atoms separate syncope patients from controls","Multivariate spectral GC split by partial info decomposition into three atoms","Postural stress flips high-order GC patterns in syncope versus healthy subjects"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Redundancy is defined as the pointwise minimum of the spectral Granger measures of the atom’s source subsets, and all measures are obtained from linear state-space models; if that minimum misrepresents shared causal information, or if important nonlinear causal structure is present, the unique/redundant/synergistic split can fail even when full multivariate GC is large.","fun_headline_variants_meta":{"raw":{"variants":["Spectral GC dissected into unique, redundant, synergistic atoms via PID","PDGC extracts high-order unique redundant synergistic causal atoms","Unique redundant synergistic GC atoms separate syncope patients from controls","Multivariate spectral GC split by partial info decomposition into three atoms","Postural stress flips high-order GC patterns in syncope versus healthy subjects"]},"model":"grok-4.5","effort":"low","cost_usd":0.005982,"raw_usage":{"total_tokens":1565,"prompt_tokens":804,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":59820000,"prompt_tokens_details":{"text_tokens":804,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":672,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":804,"tokens_out":89,"duration_ms":5780,"temperature":1.0,"reasoning_tokens":672,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T13:07:11.593780+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a controlled linear Gaussian network whose ground-truth unique, redundant and synergistic causal strengths are known, compute PDGC; if the recovered atoms deviate systematically from those known strengths, or if the same atoms change sign or disappear under mild non-Gaussian driving noise while full GC remains unchanged, the decomposition is falsified.","supporting_citations":[],"review_version":1}