{"id":"7e9dd99e-e977-4262-ab02-8f0c793199e4","arxiv_id":"1909.01288","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"For mutualistic and gene regulatory networks, nonlinear tipping-point control importance favors high-degree nodes, while linear controllability importance favors low-degree nodes, indicating a systematic mismatch.","lead":"This paper compares which nodes matter most for controlling real ecological and genetic networks, under two very different assumptions about the dynamics. It finds that the two methods rank nodes in opposite order, suggesting that popular linear control theory may not apply to these nonlinear living systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Linear importance measure relies on equal-weight adjacency whose degeneracies create the many driver sets; a Jacobian-based test could overturn the opposite-trend claim.","rationale":"The reader's weakest assumption concerns the representativeness of the 1000 random minimum controller sets and the correctness of the driver-set enumeration. That is a valid reproducibility concern: the sampling algorithm is unspecified and the row-dependence criterion is only sketched, so a biased sample could produce a spurious negative correlation. However, the more load-bearing assumption is earlier in the chain: the entire linear importance measure depends on the existence of many equivalent minimum driver sets, which in turn depends on eigenvalue degeneracies of the adjacency matrix. With equal unit link weights, bipartite mutualistic networks have large zero-eigenvalue multiplicities, creating the degeneracy that makes R_L a nontrivial distribution. If the network is represented with the natural linearization of the nonlinear dynamics (the Jacobian at the equilibrium), the weights are heterogeneous, the degeneracy is generically absent, and the driver-set ensemble—and thus R_L—changes qualitatively. The observed negative correlation between R_L and R_NL is therefore not established as a property of linear controllability per se; it may be an artifact of the equal-weight topology-only model. This concern does not invalidate the paper's broader cautionary message about topology-only linear controllability, and the paper deserves credit for the empirical breadth across 43 networks and for highlighting the non-uniqueness of minimum driver sets. But the headline claim of 'irrelevance' needs the linear model to be representative of actual interactions. The proposed Jacobian-based test would settle whether the negative correlation is robust. If it persists, the paper's conclusion is strengthened; if it disappears, the claim must be narrowed or rejected. Because this test has not been performed, the appropriate verdict remains conditional, matching the reader's verdict; hence UNCHANGED.","tokens_in":17526,"tokens_out":13702,"duration_ms":143219,"concrete_test":"Compute the Jacobian J of the nonlinear mutualistic dynamics used for Eq. (1) at the stable high-abundance equilibrium for network A with the same parameters as in Fig. 2. Use J as the adjacency matrix A in Eqs. (3)-(7), determine N_D and sample valid minimum driver sets via the PBH condition in Eq. (4), and recompute the linear importance R_L for the pollinator nodes. If the Pearson correlation between R_L and R_NL is no longer negative, or if N_D = 1 and R_L becomes nearly uniform, then the opposite-trend result is an artifact of equal unit link weights rather than a robust property of linear controllability. A simpler variant: replace unit weights with independent lognormal weights and repeat the same analysis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The linear importance ranking R_L^i in Eq. (7) is computed from minimum driver sets of the exact controllability of the unweighted adjacency matrix, i.e., all links are assigned unit weight. For an undirected network, Eq. (6) sets N_D equal to the maximum algebraic multiplicity of the adjacency eigenvalues. In a bipartite mutualistic network with unit link weights, the zero eigenvalue has a large multiplicity (at least |N_A - N_P|), which is exactly what produces the vast number of equivalent minimum controller sets and the degree-dependent driver frequencies that drive the negative correlation with nonlinear importance. If, instead, the adjacency matrix is taken from the Jacobian of the very nonlinear model used for the tipping-point control—evaluated at the stable high-abundance equilibrium—the off-diagonal entries are not equal, the eigenvalue degeneracy is generically broken, and N_D collapses to 1 for a connected network. Then the driver-set-frequency importance measure in Eq. (7) either becomes trivial (all single-node sets are equivalent) or yields a fundamentally different ranking. The paper's central empirical finding, the opposite trends in linear versus nonlinear importance, may therefore be an artifact of the equal-weight modeling assumption rather than a robust property of linear controllability applied to the actual interaction strengths. The paper never justifies this equal-weight choice or shows that the negative correlation survives generic weights or the natural linearization of the nonlinear dynamics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares nodal importance under two control paradigms for a set of empirical mutualistic networks, the S. cerevisiae gene regulatory network, and the C. elegans connectome. Nonlinear importance is defined by the recovery point of the system after a tipping-point collapse when a single species or gene is maintained at a fixed abundance or activity level (Eqs. 1 and 2). Linear importance is defined as the frequency with which a node appears in randomly sampled minimum driver sets of the unweighted adjacency matrix under the exact controllability framework (Eq. 7). The authors report that the two importance rankings exhibit opposite degree trends, with nonlinear importance favoring high-degree nodes and linear importance favoring low-degree nodes, and that Pearson correlations are negative and cosine distances are large for most of the 43 mutualistic networks and for the two yeast subnetworks. They also argue that a previous linear-controllability analysis of the C. elegans connectome cannot predict neuron function because the driver-set importance is approximately uniform and muscle cells appear more important than neurons.","tokens_in":17771,"tokens_out":6153,"duration_ms":64328,"significance":"If the central empirical finding is robust, the paper would make an important contribution by providing a quantitative, network-based argument against applying linear controllability rankings to nonlinear biological networks, and by drawing explicit attention to the non-uniqueness of minimum driver sets. The comparison metric is concrete and falsifiable: nonlinear importance is tied to a specific tipping-point recovery protocol, and linear importance is measured by driver-set frequency. The breadth of empirical networks and the inclusion of a gene regulatory network are strengths, as is the explicit use of Pearson correlation and cosine distance to summarize the disagreement between the two rankings. The paper is less strong on methodology: the equal-weight adjacency assumption, the asymmetry between the node sets used for the two importance measures, and the unspecified driver-set sampling algorithm leave the main claim not fully established.","major_comments":[{"comment":"The central comparison is not yet established because the linear importance measure R_L in Eq. (7) is computed from the unweighted adjacency matrix of each network. For the bipartite mutualistic networks, the large zero-eigenvalue multiplicity of the binary matrix is precisely what creates the vast number of equivalent minimum controller sets that produces the degree-dependent driver frequencies in Figs. 2(e-h) and 3(e). If the adjacency matrix were replaced by the Jacobian of the nonlinear mutualistic dynamics evaluated at the stable high-abundance state, the eigenvalue degeneracy would generically be broken, and for a connected network Eq. (6) would give N_D=1, making the driver-set-frequency measure R_L uniform or nearly so. The reported negative correlation between R_NL and R_L could therefore be an artifact of the equal-weight structural assumption rather than a property of linear controllability of the actual nonlinear system. The authors should either justify the unweighted choice as the intended linear model or repeat the analysis using Jacobian-based and generically weighted adjacency matrices and show that the opposite-trend finding survives.","section":"Methods, 'Linear control importance ranking', Eqs. (3)-(7)"},{"comment":"Nonlinear importance in Eq. (1) is defined only for pollinator species, while linear importance in Eq. (7) is computed for all nodes, including plants. The text reports correlations for 'the five mutualistic networks' and then for 43 networks in Fig. 5 without stating whether the Pearson correlation and cosine distance are computed only over the intersecting node set (pollinators) or over all nodes. If the latter, the comparison mixes nodes for which R_NL is defined with nodes for which it is not, and this asymmetry alone could produce the negative correlations. The authors should specify the node sets used for every correlation and, if necessary, recompute Fig. 5 on the common pollinator subset.","section":"Results, 'Nonlinear and linear control importance', Fig. 2 and Fig. 5"},{"comment":"The driver-set sampling underlying Eq. (7) is not specified. The paper states that 1000 random minimum controller sets are used and illustrates the row-dependence criterion with a 10-node example, but it does not provide the algorithm for enumerating or sampling the minimum controller sets, nor a test that 1000 samples are representative when the total number of sets is of order 10^12 for network E. Without this information, R_L cannot be reproduced, and it is unclear whether the reported frequencies are unbiased estimates of the true driver-set probabilities. The authors should specify the sampling algorithm and provide convergence diagnostics.","section":"Methods, Eq. (7) and Fig. 6"}],"minor_comments":[{"comment":"The phrase 'suggesting strongly irrelevance of linear controllability' is ungrammatical; it should be 'suggesting the strong irrelevance of linear controllability' or 'suggesting that linear controllability is irrelevant'.","section":"Abstract"},{"comment":"The text refers to 'Figures 7(b) and 7(c)' as two realizations of the minimum controller set and to Fig. 7(d) as the importance ranking, whereas the caption labels the importance ranking as (b) and the realizations as (c,d). The labeling should be made consistent.","section":"Appendix A, Fig. 7"},{"comment":"The word 'Where' at the start of the sentence defining the gene regulatory importance should be lowercase 'where'.","section":"Methods, Eq. (2)"},{"comment":"The sentence 'the number of ways to choose the latter is 54' appears without a superscript and the relation between linearly dependent rows and admissible driver nodes is described only cryptically; please clarify the counting and the criterion for choosing driver rows.","section":"Fig. 6 caption"},{"comment":"Table II contains typographical errors such as 'Greenladn' and 'Aores Island', and the species names in Table I are formatted inconsistently; these should be corrected.","section":"Appendix B and Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the network control community, but the equal-weight adjacency issue is a serious correctness risk: the main opposite-trend result may be an artifact of using the binary adjacency matrix rather than the Jacobian of the nonlinear dynamics. If the authors can show the result persists under Jacobian-based or generically weighted adjacency matrices, the paper would be publishable. The title and abstract claim 'irrelevance' more strongly than the presented evidence supports, and the manuscript would benefit from a more measured framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The paper does something concrete: it compares, for 43 empirical mutualistic networks and the yeast transcriptional network, a nonlinear control importance (ability to recover from a tipping point by maintaining one pollinator's abundance) with a linear control importance (frequency in minimum driver sets under exact controllability of the unweighted adjacency matrix). The headline result is that the two rankings move in opposite directions with degree — nonlinear favors hubs, linear favors small-degree nodes. That is a genuine empirical pattern, and the C. elegans reanalysis is a useful corrective to the overclaiming in Yan et al. The paper earns credit for taking driver-set non-uniqueness seriously; most of the linear controllability literature ignores it. The main soft spot is exactly the one in the stress-test note. The linear importance is computed from the unweighted, equal-weight adjacency matrix, and in bipartite mutualistic networks the large multiplicity of the zero eigenvalue is what creates the enormous degeneracy of minimum driver sets. Use the Jacobian of the actual nonlinear model at the stable high-abundance equilibrium and the eigenvalues are generically nondegenerate; the minimum driver count falls, and the frequency measure in Eq. (7) either becomes trivial or produces a different ranking. The paper never justifies the equal-weight choice, nor does it test whether the opposite trend survives generic weights or the natural linearization. That said, much of the applied linear-controllability literature really does use structural/unweighted graphs, so the paper is a fair critique of that literature even if the broader \"irrelevance\" claim is stronger than what is tested. Two smaller issues: nonlinear importance is computed only for pollinators while linear importance covers all nodes, and the 1000 sampled driver sets are described without an algorithm or error bars. The negative correlation across 43 networks makes the sampling concern minor, but the pollinator/whole-network asymmetry could bias the comparison. Overall: this is a serious paper with a real empirical finding, aimed at the right target. The weak point is the unweighted adjacency choice, and a Jacobian-based check would either strengthen or narrow the conclusion. I would send it to an expert referee and ask for that check before acceptance; as is, it deserves peer review, not desk rejection.","headline":"Empirical opposite-trend result that deserves peer review, but its linear-control measure rests on unweighted adjacency and may not survive weighting by the actual Jacobian.","tokens_in":18279,"tokens_out":2304,"would_cite":true,"duration_ms":25546,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B05","05C82"],"pacs":[],"model":"deepseek-v4-flash","headline":"For nonlinear biological networks, linear controllability produces opposite, and therefore useless, node rankings.","keywords":["linear controllability","nonlinear dynamical networks","nodal importance","tipping point control","minimum driver set","mutualistic networks","gene regulatory networks","C. elegans connectome"],"falsifier":"Take a small network for which every minimum driver set can be enumerated exhaustively, compute the exact linear importance for every node, and compare it with the nonlinear recovery-based importance; if the exact correlation turns out positive rather than negative, the reported opposite trend would be an artifact of sampling a limited number of driver sets.","tokens_in":17308,"feed_emoji":"🦋","tokens_out":3225,"duration_ms":36936,"temperature":0.7,"pith_summary":"The paper asks whether linear controllability theory, which assumes nodes obey linear time-invariant dynamics, can say anything useful about real networks whose dynamics are nonlinear. It answers by comparing how individual nodes rank in importance for the two kinds of control. For nonlinear control, importance is measured by how well a single managed node can pull a system back from a tipping-point collapse. For linear control, importance is the probability that a node appears in a minimum driver set. Across many empirical mutualistic networks and the yeast gene regulatory network, the two rankings move in opposite directions: nonlinear control favors high-degree nodes, while linear control favors low-degree nodes. The paper concludes that linear controllability generates information that is not useful for controlling tipping-point dynamics in these nonlinear systems.","feed_headline":"Linear controllability ranks biological networks backwards","feed_subtitle":"For mutualistic and gene regulatory networks, nonlinear control favors hubs; linear control theory favors small-degree nodes.","key_machinery":"The key object is the linear control importance measure of Eq. (7): for a network with many equivalent minimum driver sets, the importance of node i is the fraction of those sets in which it appears. The minimum driver sets themselves are built from the PBH rank condition of exact controllability, which says that the control rank deficiency equals the maximum geometric multiplicity of an eigenvalue of the adjacency matrix; the linearly dependent rows of the matrix give the candidate driver nodes. This is compared against the nonlinear control importance of Eq. (1), defined by the normalized recovery threshold when a single node is externally maintained. The comparison produces the paper's Pearson-correlation and cosine-distance evidence for opposite ranking trends.","core_discovery":"The central claim is that nodal importance rankings from linear and nonlinear control are characteristically opposite for empirical biological networks. In nonlinear control of tipping points, the paper defines importance by the recovery point: a node is more important if holding its abundance or activity at a fixed level lets the whole system recover at a harsher parameter value. For linear control, importance is defined as the frequency with which a node appears among equivalent minimum driver sets, computed with the exact controllability theory. In most of the 43 mutualistic networks tested and in two subnetworks of the S. cerevisiae gene regulatory network, nonlinear importance correlates positively with degree while linear importance correlates negatively with degree. The paper also finds that, for the C. elegans connectome treated as a linear network, the minimum driver set is not unique and nodal importance is nearly uniform, with muscle cells appearing about twice as often as motor neurons, a result it argues is biologically meaningless and contradicts the claim that linear control principles predict neuron function.","pith_inferences":["The paper's opposite-trend result likely extends beyond tipping-point recovery: because high-degree nodes dominate nonlinear resilience in many dynamical models, any linear controllability measure that systematically avoids hubs will tend to mismatch nonlinear control objectives in general.","A natural testable extension is to compute linear importance with the structural controllability maximum-matching method instead of exact controllability; the paper's argument suggests the same low-degree bias should appear, since unmatched nodes in maximum matchings tend to be low-degree.","One could also test whether the negative correlation persists for weighted networks, since exact controllability's eigenvalue degeneracy, and hence the set of minimum drivers, depends on the weight values rather than only on topology.","The implicit deeper claim, that nonlinear nodal importance is generically heterogeneous while linear importance is often nearly uniform, could be probed with other nonlinear control goals such as synchronization or resilience maintenance to see whether heterogeneity is a common feature."],"forward_implications":["If the ranking is accepted, linear controllability should not be used to select driver nodes for controlling tipping points in mutualistic or gene regulatory networks, because the nodes it favors are often exactly the ones that cannot trigger recovery in the nonlinear system.","For networks with many equivalent minimum driver sets, any single minimum driver set is nearly arbitrary, so linear controllability cannot serve as a reliable centrality or node-importance ranking.","The C. elegans analysis implies that linear structural controllability cannot single out special neurons for function prediction when the linear importance distribution is approximatively uniform across all neurons.","The negative correlation between nonlinear and linear importance suggests that using network structure alone, as linear controllability does, will typically miss the nodes that matter for nonlinear resilience and recovery.","For the gene regulatory network, genes with zero nonlinear control importance can have very high linear importance, so a linear-theory-based intervention could be completely ineffective or even harmful."],"supporting_citations":[{"why":"Supplies the exact controllability theory used to determine minimum controller sets and the candidate driver nodes from the PBH rank condition.","marker":"[9]"},{"why":"The C. elegans connectome claim under examination, which the paper argues overstates what linear controllability can predict.","marker":"[31]"},{"why":"Documents that many equivalent minimum controller sets exist, justifying the paper's probability-based linear importance measure.","marker":"[60]"},{"why":"Introduced the maximum-matching linear structural controllability framework that the paper argues is irrelevant to nonlinear control.","marker":"[4]"},{"why":"Provides the S. cerevisiae transcriptional regulatory network data used for the gene regulatory network analysis.","marker":"[51]"},{"why":"Supplies the empirical Tenerife mutualistic network used as the main illustrative example.","marker":"[61]"},{"why":"Supplies an empirical mutualistic network used in the correlation and cosine-distance comparisons.","marker":"[68]"},{"why":"Supplies an empirical mutualistic network used in the correlation and cosine-distance comparisons.","marker":"[69]"}],"fun_headline_variants":["Linear control ranks hubs as least important for mutualistic networks","Nonlinear control favors hubs; linear control favors small nodes","Opposite importance: nonlinear networks need hubs, linear says small","Why linear controllability fails for biological networks","C. elegans connectome: linear control rankings disputed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The linear importance ranking rests on the assumption that the 1000 randomly sampled minimum driver sets fairly represent the full ensemble of equivalent sets, so that a node's appearance frequency is its true linear-control importance.","fun_headline_variants_meta":{"raw":{"variants":["Linear control ranks hubs as least important for mutualistic networks","Nonlinear control favors hubs; linear control favors small nodes","Opposite importance: nonlinear networks need hubs, linear says small","Why linear controllability fails for biological networks","C. elegans connectome: linear control rankings disputed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001255,"raw_usage":{"total_tokens":5117,"prompt_tokens":892,"completion_tokens":4225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":4147}},"tokens_in":508,"tokens_out":4225,"duration_ms":29395,"temperature":1.0,"reasoning_tokens":4147,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:23:10.410537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small network for which every minimum driver set can be enumerated exhaustively, compute the exact linear importance for every node, and compare it with the nonlinear recovery-based importance; if the exact correlation turns out positive rather than negative, the reported opposite trend would be an artifact of sampling a limited number of driver sets.","supporting_citations":[{"cited_title":"Controlling edge dynamics in complex networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the exact controllability theory used to determine minimum controller sets and the candidate driver nodes from the PBH rank condition."},{"cited_title":"Nijmeijer and A","cited_arxiv_id":null,"evidence_quote":"The C. elegans connectome claim under examination, which the paper argues overstates what linear controllability can predict."},{"cited_title":"Tipping points in ecological networks,","cited_arxiv_id":null,"evidence_quote":"Documents that many equivalent minimum controller sets exist, justifying the paper's probability-based linear importance measure."},{"cited_title":"Controllability of multi-agent systems from a graph-theoretic perspective,","cited_arxiv_id":null,"evidence_quote":"Introduced the maximum-matching linear structural controllability framework that the paper argues is irrelevant to nonlinear control."},{"cited_title":"Alon, An Introduction to Systems Biology: Design Principles of Biological Circuits (CRC press, 2006)","cited_arxiv_id":null,"evidence_quote":"Provides the S. cerevisiae transcriptional regulatory network data used for the gene regulatory network analysis."},{"cited_title":"Correlations in the degeneracy of structurally control lable topologies for networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the empirical Tenerife mutualistic network used as the main illustrative example."},{"cited_title":"Resilience and stability of ecological systems,","cited_arxiv_id":null,"evidence_quote":"Supplies an empirical mutualistic network used in the correlation and cosine-distance comparisons."},{"cited_title":"The structure of a plant-pollinator food web,","cited_arxiv_id":null,"evidence_quote":"Supplies an empirical mutualistic network used in the correlation and cosine-distance comparisons."}],"review_version":1}