{"id":"238084a7-380a-4df2-8fd4-7b9619682f61","arxiv_id":"1908.03514","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Modeling choices such as time horizon, normalization, and control set size substantially change controllability metrics in structural brain networks, and the new spatial adjacency connectivity and energy landscape complexity measures are complementary extensions.","lead":"This paper tests how four common network control measurements of the brain change when researchers pick different time systems, time horizons, normalization strengths, or sets of controlled nodes. It also introduces a new complexity metric for the control energy landscape and a spatial-adjacency version of brain connectivity, grounded in diffusion imaging from ten adults.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Null-model validation of the new complexity metric is internally inconsistent; with n=10 a Wilcoxon W=65 cannot occur, so the strongest empirical claim needs re-analysis.","rationale":"In good faith, the paper does what it claims: a systematic sensitivity analysis of four control metrics under time system, horizon, normalization, and control set size, with two extensions. The sensitivity results are internally coherent and the paper explicitly flags the linear, time-invariant, noiseless model as a simplification (§2.1, §5). I therefore do not treat the linear-model caveat as the decisive objection; it is a well-known modeling boundary that the paper itself documents. The load-bearing weak point is instead the empirical validation of the new complexity metric C, which the reader correctly identifies as the strongest specific finding. That validation depends on a Wilcoxon comparison whose reported statistic and p-values are not reproducible from the stated sample sizes: with 10 subjects, a signed-rank W cannot exceed 55, and the number of null instantiations is inconsistent (1000 in §3.9 vs 100 in Fig. 8). Since the manuscript provides no code or per-subject values, this is not a checkable arithmetic detail; it is the main evidence that brain networks have a more homogeneous energy landscape than null models. The correlations used to claim complementarity are also low-power (n=10) and largely not significant. These facts do not prove the claim false, but they make the central empirical result conditional on re-analysis. The existing CONDITIONAL verdict is therefore unchanged, with the required condition made concrete: recompute the null-model test and confirm the reported statistics.","tokens_in":25207,"tokens_out":10406,"duration_ms":116703,"concrete_test":"Obtain from the authors the per-subject C values and the full set of null-model C values (or the connectomes and null-generation code). Recompute C for each subject and each null model, then run one Wilcoxon rank-sum (or signed-rank) test with the actual n1=10 and n2 equal to the true null count. Report the test statistic and exact p-value for each null model. If W=65 with p=8e-8 is not reproduced, the §4.8 validation fails and the complexity metric's main evidence is void; if corrected p-values remain <0.05 after a suitable correction, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's most striking empirical result is §4.8: brain energy-landscape complexity C is significantly lower than topological, spatial, and combined null models (W=65, 0, 2498; p=8e-8, 5e-8, 6e-3). This is the only direct validation of the proposed metric and is highlighted in the abstract and discussion. The statistical basis is not reproducible as stated. §3.9 says 1000 null instantiations per model; the Fig. 8 caption says 100. More importantly, with n=10 participants, a Wilcoxon signed-rank statistic W (sum of positive ranks) has maximum 55, so W=65 is impossible under that test; if W is instead a rank-sum or U statistic, the reported p-values do not follow from the stated design. No code or per-subject C values are provided, so the discrepancy cannot be resolved from the manuscript. The complementarity claim is also weak: correlations with existing metrics in §4.8 are based on n=10 and mostly not significant. Because the complexity metric's utility rests on this validation, the central claim about C is not yet settled. The linearity caveat, by contrast, is explicitly acknowledged in §2.1 and §5 and does not undermine the internal sensitivity analysis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a methodological primer for network control theory applied to structural brain networks. Using diffusion imaging data from ten healthy adults, the authors systematically vary four modeling choices—time system, time horizon, normalization, and control set size—and quantify how these choices affect average controllability, modal controllability, minimum control energy, and optimal control energy. They also propose two extensions: a spatial-adjacency structural connectivity measure S, and a new complexity metric C defined as the interquartile range of eigenvalues of the inverted controllability Gramian. The paper reports that modeling choices can induce qualitatively different control regimes, that S provides information complementary to tractography-based connectivity, and that brain networks have significantly lower energy-landscape complexity than several null models. It closes with concrete modeling recommendations.","tokens_in":25415,"tokens_out":5476,"duration_ms":59091,"significance":"If the sensitivity results hold, the paper fills a genuine need in an active field: it provides an accessible, systematic account of how seemingly arbitrary choices affect commonly used controllability metrics, and it gives actionable recommendations. The treatment of the linearity and noise assumptions is unusually transparent, and the null-model framework for the new complexity metric is a principled way to validate a derived quantity. The regional-level correlations are mostly strong and based on a large number of regions, which supports the main sensitivity claims. The paper is less convincing on the individual level, where n = 10 yields wide confidence intervals and many non-significant correlations. The validation of the complexity metric, which is highlighted in the abstract and discussion, currently rests on internally inconsistent statistics and weak complementarity evidence, so the central claims about C are not yet fully supported.","major_comments":[{"comment":"The headline null-model validation of the complexity metric C is internally inconsistent. The text reports Wilcoxon W = 65, 0, 2498 with p = 8×10−8, 5×10−8, and 6×10−3 for the topological, spatial, and combined null models. With n = 10 participants, a Wilcoxon signed-rank statistic cannot exceed 55, so W = 65 and W = 2498 are impossible under that test; moreover, a signed-rank W = 0 would have exact two-sided p ≈ 0.002, not 5×10−8. If a two-sample rank-sum test against the null instantiations is intended instead, the test is not described, and the sample sizes are ambiguous because §3.9 states 1000 random instantiations per model while the Fig. 8 caption states 100. No per-subject C values or code are provided to resolve the discrepancy. Since the claimed lower complexity of brain energy landscapes is the only direct validation of C and is emphasized in the abstract and discussion, this issue must be fixed by re-analyzing the data and reporting the exact test, sample sizes, and per-subject values before the central claims about C can be assessed.","section":"§4.8, Eq. (9); §3.9; Fig. 8"},{"comment":"The complementarity of C with existing controllability metrics is asserted but rests almost entirely on n = 10 correlations. The reported Pearson correlations are r = -0.15 (p = 0.68) with average controllability, r = -0.67 (p = 0.04) with modal controllability, and r = -0.40 (p = 0.26) with minimum control energy. Two of these are not statistically significant, and the one nominally significant result would not survive correction for multiple comparisons. The statement that the complexity metric is 'complementary' to existing controllability metrics therefore goes beyond the evidence reported. Please provide confidence intervals and clearly framed effect-size statements, or temper the complementarity claim to reflect the preliminary nature of this evidence.","section":"§4.8"}],"minor_comments":[{"comment":"The number of null-model instantiations is inconsistent: §3.9 says 1000 per model, while the Fig. 8 caption says 100. Please reconcile these numbers and state the exact test used.","section":"Fig. 8 caption vs. §3.9"},{"comment":"Reference [10] contains malformed BibTeX text in the author field and should be corrected.","section":"References"},{"comment":"The reported mean control energies (e.g., meanmin = 43.8 and meanopt = 56.0) are given without units or a clear statement of the normalization of the state vectors; please define these units explicitly.","section":"§4.5"},{"comment":"The paper would benefit from a data and code availability statement, especially for the new complexity metric whose validation currently cannot be reproduced from the manuscript alone.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central sensitivity analysis is defensible and likely useful, but the validation of the new complexity metric needs to be redone or carefully re-reported. The Wilcoxon statistics as written cannot be reproduced from the stated design, and the complementarity correlations are too weak to support the current wording. I do not see this as a reject if the authors can supply the exact tests, correct the null-model counts, and tone down or properly support the complementarity claim. I would also encourage the editor to request that the per-subject C values be made available as supplementary material."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the controllability guide. It's a genuinely useful paper for a growing field, and the sensitivity analysis is exactly the sort of thing that should exist. But the one new headline result — the complexity metric validating against null models — has a statistics problem that needs to be sorted before anyone trusts it.\n\nWhat the paper does well: it runs a systematic, one-at-a-time sweep of four modeling choices (time system, time horizon, normalization, control set size) against four common metrics, and reports rank-based consistencies with mostly clear figures. The didactic intro is solid, and the authors are upfront about the linear, time-invariant, noise-free model's limits. The two extensions — a spatial adjacency matrix S and the energy-landscape complexity C — are genuinely new, and the idea of measuring the spread of inverse Gramian eigenvalues is reasonable. Give credit for the null models, too, in principle.\n\nThe soft spot is serious. Section 4.8 reports W = 65, 0, and 2498 against three null models, with p-values 8e-8, 5e-8, 6e-3. With n = 10 participants, a Wilcoxon signed-rank statistic caps at 55, so W = 65 can't come from that test. If W is a rank-sum against 1000 null instantiations, then W = 0 should yield a p-value much smaller than 5e-8. The text says 1000 nulls in §3.9; the Figure 8 caption says 100. I can't resolve this from the manuscript, and no per-subject values or code are given. Since this null-model comparison is the only validation of C, the central claim about the metric is not yet supported as stated. The rest of the paper, the sensitivity analysis, doesn't inherit this problem; those claims stand on the correlations and are internally consistent.\n\nThe individual-level correlations are also thin (n=10), and many are non-significant — the authors say so, but a reader should weigh the recommendations accordingly. That said, the recommendations themselves are reasonable and clearly derived.\n\nWho should read this: anyone applying network control theory to structural connectomes. It's a practical field guide, not a breakthrough, and it should get a serious peer review — the statistical issues are correctable, not fatal. I'd require a corrected analysis of §4.8 and, ideally, public code and data before endorsing it as best practice.","headline":"Useful sensitivity guide with a broken complexity-metric validation; fix the null-model statistics and it's publishable.","tokens_in":25982,"tokens_out":3177,"would_cite":false,"duration_ms":32761,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Brain controllability metrics shift with modeling choices, and a new metric shows brain energy landscapes are more homogeneous than rewired networks.","keywords":["network control theory","structural brain networks","controllability metrics","energy landscape complexity","spatial adjacency connectivity","diffusion imaging","modeling choices","modal controllability"],"falsifier":"Recompute average controllability, minimum control energy, and the complexity metric on the same ten connectomes after replacing the linear noiseless model with a stochastic or nonlinear neural mass model, or after replacing $A$ with a functional or effective connectivity matrix. If the short-horizon and fast-stabilization control regimes and the brain's lower complexity relative to null models vanish or reverse, then the reported properties belong to the linear tractography model rather than to brain architecture.","tokens_in":24990,"feed_emoji":"🧠","tokens_out":7538,"duration_ms":66575,"temperature":0.7,"pith_summary":"This paper argues that the numbers produced by network control theory on structural brain networks are not fixed properties of the brain but depend on modeling choices: time system, time horizon, normalization, and control set size. It tests four standard metrics (average controllability, modal controllability, minimum control energy, and optimal control energy) on diffusion-imaging connectomes from ten adults and shows that short time horizons and fast stabilization create distinct control regimes. It also proposes two extensions: a spatial adjacency connectivity matrix based on face-touching voxels, and a complexity metric for the energy landscape. The paper's strongest new finding is that brain networks have a significantly more homogeneous energy landscape than topological, spatial, or combined null models, implying that the brain requires similar control energy for diverse state transitions.","feed_headline":"Brain control energy is more uniform than rewiring predicts","feed_subtitle":"Modeling choices shift brain controllability metrics; two new measures add complementary information.","key_machinery":"The machinery is the linear controlled system $\\dot{x}(t) = Ax(t) + B_\\kappa u_\\kappa(t)$, with $A$ the weighted structural adjacency matrix (233 nodes, zero diagonal) and $B_\\kappa$ selecting control nodes. From this come the controllability Gramian $W_{\\kappa,T} = \\int_0^T e^{At} B_\\kappa B_\\kappa^\\top e^{A^\\top t} dt$ and the four metrics: average controllability, modal controllability $\\phi_i = \\sum_{j=1}^{N} (1 - e^{\\lambda_j(A)}) v_{ij}^2$, minimum control energy, and optimal control energy. The paper's new complexity metric $C$ measures the spread of the eigenvalue distribution of $W_{\\kappa,T}^{-1}$, capturing the heterogeneity of minimum control energy across all possible state transitions. The spatial adjacency matrix counts face-touching voxels between parcels and is averaged with $A$ to form a combined connectivity matrix.","core_discovery":"The central claim is that controllability statistics of structural brain networks should be interpreted relative to the modeling choices that produced them. Under the standard linear time-invariant model $\\dot{x} = Ax + B_\\kappa u_\\kappa$, the paper shows that discrete versus continuous time, the time horizon $T$, the normalization parameter $c$, and the size and composition of the control set $B_\\kappa$ each shift metric values enough to change regional rankings and even introduce qualitatively different control regimes. The authors therefore recommend verifying results in the alternative time system, checking multiple time horizons, scaling $c$ to the largest eigenvalue of $A$, and controlling sufficiently many regions. The paper further defines spatial adjacency connectivity as the count of face-touching voxels between parcels and energy landscape complexity as the interquartile range of the eigenvalues of the inverse controllability Gramian. It shows that the spatial adjacency measure is complementary to tractography-based connectivity, and that the brain's energy landscape complexity is lower than all three null models tested.","pith_inferences":["If the complexity metric proves reproducible, it could be tested as a state-independent marker in disorders where neural dynamics become rigid or stereotyped; this is not tested in the paper.","The trade-off between controlling fast and slow modes at the 50% threshold suggests a testable regional specialization: subcortical regions favor fast-mode control and frontoparietal regions favor slow-mode control in larger samples.","Because the energy landscape metric is built on the inverse Gramian with full-brain control, applying it to partial control sets will require regularization or a different estimator; the paper notes the inverse Gramian is ill-conditioned for small control sets.","One could test whether the alternative control regimes induced by short horizons and fast stabilization correspond to stimulation protocols of different durations in intracranial recording data, linking the modeling regimes to observable neural responses."],"forward_implications":["Studies comparing controllability results across datasets must report time system, time horizon, normalization parameter, and control set size, since each shifts the metrics enough to change regional rankings.","Control energy drops exponentially as control set size grows, so partial control sets produce energy estimates that are reliable only above roughly 28–30% of nodes and can follow qualitatively different state trajectories than full-brain control.","Short time horizons and fast stabilization form a distinct control regime, so results obtained at one time scale should not be extrapolated to another without rechecking the alternative.","The lower complexity of the brain's energy landscape relative to all three null models indicates that diverse state transitions require more similar control energy in brain networks than in rewired networks.","Spatial adjacency connectivity is complementary to tractography-based connectivity, suggesting that combining both measures may capture controllability information that either measure alone misses."],"supporting_citations":[{"why":"Supplies the linear-systems foundations, including the controllability Gramian that underlies all four metrics.","marker":"[9]"},{"why":"Defines the controllability metrics and the exponential energy-versus-control-set relation used throughout.","marker":"[11]"},{"why":"Shows how graph architecture scales control energy and motivates the new energy-complexity measure.","marker":"[30]"},{"why":"Establishes average and modal controllability on structural brain networks and is the baseline for the paper's consistency checks.","marker":"[31]"},{"why":"Provides the optimal control energy framework and the default parameters (time horizon, rho) used in the simulations.","marker":"[35]"},{"why":"Shows the role of network topology in optimally controlling the human connectome, informing the control-set-size analyses.","marker":"[36]"},{"why":"Supports the claim that brain networks are controllable from any single region via structural controllability of symmetric networks.","marker":"[60]"},{"why":"Defines the seven cognitive systems used to build the initial and target brain states for control energy simulations.","marker":"[72]"},{"why":"Provides the topological null model preserving degree and strength, against which the energy landscape complexity is compared.","marker":"[75]"},{"why":"Supplies the spatial and combined null models that preserve distance-weight relationships, used for the same comparison.","marker":"[76]"}],"fun_headline_variants":["Modeling choices skew brain controllability metrics","Brain control metrics shift with model assumptions","New measures for brain network energy landscapes","How to avoid pitfalls in brain controllability studies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the linear, time-invariant, noise-free equation $\\dot{x} = Ax + B_\\kappa u_\\kappa$, with $A$ taken from deterministic tractography streamline counts and diagonal zeroed, adequately captures the neural dynamics that controllability metrics describe; the paper itself states that linearity, time invariance, and noise-freedom are approximations.","fun_headline_variants_meta":{"raw":{"variants":["Modeling choices skew brain controllability metrics","Brain control metrics shift with model assumptions","New measures for brain network energy landscapes","How to avoid pitfalls in brain controllability studies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000482,"raw_usage":{"total_tokens":2403,"prompt_tokens":990,"completion_tokens":1413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":1359}},"tokens_in":606,"tokens_out":1413,"duration_ms":12167,"temperature":1.0,"reasoning_tokens":1359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:10:35.654233+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute average controllability, minimum control energy, and the complexity metric on the same ten connectomes after replacing the linear noiseless model with a stochastic or nonlinear neural mass model, or after replacing $A$ with a functional or effective connectivity matrix. If the short-horizon and fast-stabilization control regimes and the brain's lower complexity relative to null models vanish or reverse, then the reported properties belong to the linear tractography model rather than to brain architecture.","supporting_citations":[{"cited_title":"Optimal trajectories of brain state transitions","cited_arxiv_id":null,"evidence_quote":"Provides the optimal control energy framework and the default parameters (time horizon, rho) used in the simulations."},{"cited_title":"Optimally controlling the human connectome: the role of network topology","cited_arxiv_id":null,"evidence_quote":"Shows the role of network topology in optimally controlling the human connectome, informing the control-set-size analyses."},{"cited_title":"Structural controllability of symmetric networks","cited_arxiv_id":null,"evidence_quote":"Supports the claim that brain networks are controllable from any single region via structural controllability of symmetric networks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the seven cognitive systems used to build the initial and target brain states for control energy simulations."},{"cited_title":"Complex network measures of brain connectivity: uses and interpreta- tions","cited_arxiv_id":null,"evidence_quote":"Provides the topological null model preserving degree and strength, against which the energy landscape complexity is compared."},{"cited_title":"The contribution of geometry to the human connectome","cited_arxiv_id":null,"evidence_quote":"Supplies the spatial and combined null models that preserve distance-weight relationships, used for the same comparison."}],"review_version":1}