{"id":"b2290075-4c18-4dea-a593-bbc775663a1c","arxiv_id":"1908.01421","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"The steady-state disagreement variance of a networked linear control system equals a sum, over the nonzero Laplacian eigenvalues, of a rational function determined by the node dynamics and feedback gain.","lead":"Networked control systems with identical agents reach consensus; this paper expresses the steady-state variance of their disagreements as a sum of rational functions of the communication graph's Laplacian eigenvalues. The result turns a large high-dimensional performance calculation into scalar evaluations and yields design rules and scaling laws for feedback gains, observers, and composite networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (8) misstates the measurement-noise feedback term, so the printed closed-loop model does not imply the decoupling (10); the central formula survives only after an unstated correction.","rationale":"The reader's weakest_assumption field emphasizes the Kronecker/product/symmetry restrictions; those are scope conditions of the stated class of systems and not a hidden flaw in the derivation. The operative correctness issue is the measurement-noise term in Eq. (8), which the reader's rationale also flags. I checked the rest of the chain: the trace argument in Appendix A, the Cramer's-rule rationality proof, the observer separation in Appendix B, and the composite-network decomposition in Theorem 9 are internally consistent once (8) is corrected. Theorem 5's reliance on [12] is disclosed and is a citation rather than a correctness defect. Hence the right disposition is the reader's CONDITIONAL verdict, not a rejection: the central spectral-summation result is sound after a mechanical but load-bearing correction to the displayed closed-loop model.","tokens_in":27875,"tokens_out":23512,"duration_ms":257599,"concrete_test":"Re-derive the closed-loop dynamics by substituting u=-(L\\otimes K)y and y=(I_N\\otimes H)x+\\sigma\\eta into (2); verify that \\dot x=(I_N\\otimes A-L\\otimes BKH)x+(I_N\\otimes E)\\xi-\\sigma(L\\otimes BK)\\eta, and that the change of variables (9) yields exactly (10). Then recompute the sigma-dependent expressions (53), (55), and (56) from this corrected model; if any of them uses the printed -(L\\otimes\\sigma I_{m_3})\\eta rather than -\\sigma(L\\otimes BK)\\eta, the displayed numerical formulas need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"From (2) and (3), with y=(I_N\\otimes H)x+\\sigma\\eta, the feedback is u=-(L\\otimes K)(Hx+\\sigma\\eta), so the noise enters the closed loop as -\\sigma(L\\otimes BK)\\eta. The displayed Eq. (8) instead has -(L\\otimes\\sigma I_{m_3})\\eta: this is dimensionally inconsistent (an Nm_3-dimensional vector cannot be added to the Nn-dimensional state equation unless n=m_3 and BK is dropped) and it cannot produce the -\\lambda_i BK\\sigma\\gamma_i term in Proposition 1. All of Theorem 1, the Lyapunov equation (13), and the sigma-dependent examples depend on the BK-multiplied noise term. Thus the central claim is supported by the decoupled systems (10), but Proposition 1 as stated does not follow from the printed closed-loop model; a reader following (8) literally cannot reproduce (10). The inconsistency is mechanical, since the intended model is clear from (10), but it sits at the base of the paper's main result, and the paper also uses m_3 interchangeably with the output dimension q without defining H.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a network of N identical linear time-invariant agents coupled by relative output feedback over a weighted undirected graph, with external disturbances and measurement noise. Its main claim (Theorem 1) is that the steady-state variance of the deviation from the network average equals a sum, over the nonzero Laplacian eigenvalues, of a low-dimensional rational function phi(lambda,K) obtained from a decoupled Lyapunov equation. The paper then develops connectivity-threshold design conditions for state-feedback and observer-based output-feedback gains, derives performance lower bounds and scaling laws for path/cycle graphs, extends the framework to composite networks, and illustrates the results with many closed-form examples, including double- and triple-integrator agents, platoons, harmonic oscillators, and an aircraft formation.","tokens_in":28050,"tokens_out":26288,"duration_ms":260193,"significance":"If the central formula is correct after the necessary correction, this is a useful unification and generalization of earlier H2/coherency results for first- and second-order consensus networks to arbitrary identical linear nodal dynamics with output feedback and measurement noise. The main strength of the paper is the explicit separation between the graph spectrum and the low-dimensional gain-dependent rational function, which enables symbolic evaluation, design thresholds, and scaling laws. The paper also provides concrete falsifiable predictions, such as Theta(N^2) and Theta(N^4) scalings for path graphs, and a clean LMI construction for minimum-connectivity gains. The appendices contain detailed derivations, and the numerical examples are broadly consistent with the formulas. The significance is conditional, however, because the printed derivation of the main theorem contains an error in the treatment of the measurement-noise term.","major_comments":[{"comment":"The displayed closed-loop equation is not the consequence of the model in Section III. From (2) and (3), y=(I_N tensor H)x + sigma*eta, so u = -(L tensor K)(I_N tensor H)x - sigma(L tensor K)eta, and hence the state equation contains -sigma(L tensor BK)eta, not -(L tensor sigma I_{m3})eta. The printed term is dimensionally inconsistent unless n=m3=q and it omits the factor BK; consequently, a reader who follows (8) literally cannot derive the decoupled system (10), the Lyapunov equation (13), or the sigma-dependent examples. Equation (6) similarly drops the measurement-noise term without comment. The intended model is clear from (10), so this is a mechanical error, but it sits at the base of Theorem 1 and must be corrected, with the dimension relation m3=q stated explicitly and the derivations in Appendix A re-verified against the corrected noise term.","section":"Section IV, Eq. (8), and Proposition 1"}],"minor_comments":[{"comment":"The centering matrix should act on the performance-output dimension m2, not m1; as printed, z(t) has dimension N*m2, so the first factor should be M_N tensor I_{m2}.","section":"Section III, Eq. (5)"},{"comment":"The dimensions q and m3 are never related; because eta_i is added to Hx_i and because the observer formulas use F*gamma_i, the paper should state explicitly that m3=q.","section":"Section III, after Eq. (2)"},{"comment":"The limit in (37) is taken with respect to sigma, but sigma is the physical measurement-noise magnitude fixed by the problem; the dual of Theorem 7 requires an independent auxiliary parameter epsilon, as in (35)-(36), and (38) should be written with epsilon^{-2}. As printed, the statement conflates the design parameter with the noise level.","section":"Section V.D, Theorem 8, Eq. (37)"},{"comment":"For C=e1 and sigma=0, Table I gives phi=1/(2 k1 k2 lambda^2), whereas (53) states 1/(k1 k2 lambda^2); if the performance output is the full state, the expression should additionally contain a 1/(2 k2 lambda) term. The displayed formula should be reconciled with Table I.","section":"Example 7, Eq. (53)"},{"comment":"Compared with Eq. (42), both rational terms appear to be missing a factor 1/lambda^2 in the denominator; as printed, the sigma=0 term is (9 lambda^4 + 11 lambda^3 + 9 lambda^2 + 4 lambda + 1)/(2(3 lambda^2 + 1)), which does not blow up as lambda tends to zero and does not match (42).","section":"Example 7, Eq. (55)"},{"comment":"Choosing K* = 2*lambda_tilde(K)*K may not stabilize the system because the infimum in (28) need not be attained; the argument should choose some c > lambda_tilde(K) and take K* = c K.","section":"Appendix E, proof of Theorem 5"},{"comment":"The derivative argument shows that s is nondecreasing rather than strictly increasing unless phi' is strictly increasing; the equality-characterization sentence should be adjusted accordingly.","section":"Appendix K, proof of Theorem 11"},{"comment":"The paper should state explicitly that the exogenous signals are white-noise processes with unit spectral density; the phrase 'Gaussian with unit variance' alone does not imply the H2-norm variance interpretation used in (7).","section":"General modeling assumptions"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is salvageable and likely acceptable after revision. The main obstacle is the measurement-noise typo in Eq. (8), which is mechanical rather than fundamental because the intended model is unambiguous from Eq. (10). I recommend asking the authors to correct that equation, state m3=q, and clean up the several example-level typos listed in the minor comments. The relationship to prior work [12] and the conference version [14] is acknowledged, and I do not see a novelty or citation concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is solid and genuinely useful. The paper characterizes the steady-state variance of a consensus network with identical linear nodes as a rational function of the Laplacian eigenvalues, φ(λ, K), computed from a low-dimensional Lyapunov equation. That unifies and extends earlier first- and second-order results, and the extensions to observer-based output feedback with a separation principle and to composite networks add real value. The scaling laws and bounds in Sections VI and VIII are derived from the formulas, not fitted, and the numerical examples line up with the closed-form expressions. The authors also disclose in Remark 2 that part of Theorem 1 is hidden in [12], which is honest and appropriate.\n\nThe stress-test note is right about equation (8). As printed, the closed-loop noise term is −(L ⊗ σ I_{m3})η, which is dimensionally inconsistent and cannot produce the −λ_i BK σ γ_i term in Proposition 1. The intended model is clear from (10) and the appendix, so it is a mechanical error rather than a conceptual one, but it sits at the base of the main theorem. A reader following (8) literally cannot reproduce (10). That has to be fixed. There are also minor notational slips: the paper uses m_3 and q interchangeably without defining q = m_3, H appears without dimensions, and Theorem 3 writes ψ(λ, K) where the dependence is on F. These are easy cleanups.\n\nThe deeper assumptions — identical subsystems, undirected graph, factored gains K_{ij} = k_{ij} K — are stated clearly and are what make the diagonalization work. That is a scope condition, not a hidden flaw. Theorem 5 leans on [12] for part of the sufficiency proof, but the authors say so explicitly in Remark 4; a self-contained proof would be better but the reliance is disclosed.\n\nMy overall read: the central decoupling argument and its consequences hold up. The paper is honest, the derivations in the appendices are careful, and the results are a step forward for the subfield. It deserves peer review. I would recommend acceptance after the authors correct equation (8), define the missing dimensions, and fix the notational slips. The typo is embarrassing but not fatal; the ideas are worth the referee time.\n\nWho is this for? Anyone working on H2 performance of multi-agent systems, coherence of platoons, or spectral design of networked controllers. It is not a revolutionary paper, but it is a solid and reusable framework.","headline":"Worth a careful read: the spectral decoupling is real, but equation (8) is misprinted in a way that blocks reproduction of the main result until it is fixed.","tokens_in":28613,"tokens_out":2648,"would_cite":true,"duration_ms":28917,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93A14","93B52","05C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"One rational function predicts a network's noise response","keywords":["networked control systems","H2 performance","Laplacian spectrum","consensus","minimum connectivity threshold","scaling laws","multi-agent systems","decentralized observer"],"falsifier":"Take a four-node network of double integrators with $A=\\begin{bmatrix}0&1\\\\0&0\\end{bmatrix}$, $B=E=[0,1]^T$, $C=[1,0]$, $H=I_2$, unit disturbance on one channel, and an undirected weighted path with arbitrary weights; compute the steady-state variance $\\rho$ both by solving the full $8$-dimensional Lyapunov equation for (8) and by evaluating $\\sum_{i=2}^{4}\\varphi(\\lambda_i,K)$ from (13). If the two numbers disagree beyond numerical tolerance, the spectral formula is false. As an even more direct check, for a fixed $K$ evaluate $\\varphi(\\lambda,K)$ symbolically for a $\\lambda$ that is not a Laplacian eigenvalue and verify that it is rational with denominator $\\det((A-\\lambda BKH)\\otimes I_n+I_n\\otimes(A-\\lambda BKH))$.","tokens_in":27625,"feed_emoji":"🌐","tokens_out":8058,"duration_ms":79756,"temperature":0.7,"pith_summary":"This paper proves that for a network of identical linear subsystems coupled by relative output feedback over an undirected weighted graph, the steady-state variance of the deviations from consensus is exactly $\\rho(L,K)=\\sum_{i=2}^{N}\\varphi(\\lambda_i,K)$, where $\\lambda_i$ are the nonzero Laplacian eigenvalues and $\\varphi$ is a rational function of the eigenvalue and the feedback gain, computed from a low-dimensional Lyapunov equation. The result separates the graph topology from the nodal dynamics: once the Laplacian spectrum is known and $\\varphi$ is evaluated symbolically, the performance of the whole network is obtained without solving the full $Nn$-dimensional Lyapunov equation. The same machinery yields minimum-connectivity threshold designs for feedback and observer gains, fundamental performance limits for non-minimum-phase nodes, graph-theoretic bounds, scaling laws over path and cycle graphs, and a two-level formula for composite networks. A sympathetic reading is that this provides a spectral framework for $H_2$ performance analysis that covers first-order, second-order, and higher-order nodal dynamics in one derivation.","feed_headline":"One rational function predicts a network's noise response","feed_subtitle":"For identical nodes with relative feedback, the whole network's variance reduces to a sum over graph eigenvalues.","key_machinery":"The load-bearing mechanism is simultaneous diagonalization by the graph Laplacian eigenvectors. Because the subsystems are identical and the gain matrices factor as $K_{ij}=k_{ij}K$, the closed-loop matrix $I_N\\otimes A-L\\otimes BKH$ can be block-diagonalized by the orthogonal change of variables $r=(U^T\\otimes I_n)x$ into $N$ decoupled $n$-dimensional systems, one per Laplacian eigenvalue. The paper's performance function $\\varphi(\\lambda,K)$ is the $H_2$ squared norm of that decoupled subsystem with the given $\\lambda$, evaluated by solving the $n$-dimensional Lyapunov equation (13); rationality follows by vectorizing the Lyapunov equation and using Cramer's rule. A second named object is the minimum connectivity threshold $\\tilde\\lambda(K)=\\inf\\{\\lambda>0:(A-cBKH)\\text{ is Hurwitz for }c>\\lambda\\}$, which characterizes gains whose stabilizing property is monotone in graph connectivity and enables the LMI-based designs.","core_discovery":"The central discovery is a spectral summation formula. After applying the relative feedback law $u=-(L\\otimes KH)x$, the closed-loop network dynamics decouple under the coordinate change $r=(U^T\\otimes I_n)x$ into $N$ independent subsystems $\\Sigma_i:\\ \\dot r_i=(A-\\lambda_i BKH)r_i+E\\chi_i-\\lambda_i BK\\sigma\\gamma_i$. The paper shows that whenever $\\Sigma_2,\\dots,\\Sigma_N$ are asymptotically stable, the steady-state variance of the deviation from the mean output is $$\\rho(L,K)=\\sum_{i=2}^{N}\\varphi(\\lambda_i,K),\\qquad \\varphi(\\$\\lambda$,K)=\\operatorname{Tr}\\big(CP(\\$\\lambda$,K)C^T\\big),$$ where $P(\\lambda,K)$ is the unique positive-definite solution of the algebraic Lyapunov equation $$(A-\\$\\lambda$ BKH)P+P(A-\\$\\lambda$ BKH)^T+EE^T+\\$lambda^{2}$\\$sigma^{2}$ BK(BK)^T=0.$$ Because $\\varphi$ is rational in $\\lambda$ (Cramer's rule applied to the vectorized Lyapunov equation), the entire dependence on the communication graph enters only through the nonzero Laplacian eigenvalues. This converts a high-dimensional network performance computation into evaluating one low-dimensional function at each nonzero eigenvalue, and it reveals thresholds, bounds, and scaling laws that would be hard to see from the full matrices.","pith_inferences":["Inference: The two-level composite formula (66) could be iterated: building subnetworks of subnetworks would add one spectral sum per level, so a hierarchical network with $d$ levels would have performance expressed as a $d$-fold sum over the Laplacian spectra of the graphs at each level.","Inference: Because $\\varphi$ is rational and low-dimensional, its values can be computed once on a grid of $\\lambda$ or fit by regression, as the paper suggests, and then reused to evaluate thousands of candidate topologies; this makes spectral-formula-based topology search a plausible algorithmic route that the paper does not itself develop.","Inference: The paper's bounds rely on convexity of $\\varphi$, yet its aircraft example shows $\\varphi_1$ need not be convex; a natural extension is to seek similar lower bounds for non-convex $\\varphi$ using monotonicity or majorization constraints on Laplacian spectra, or to characterize exactly which nodal dynamics yield convex $\\varphi$.","Inference: A testable conjecture is that when the identical-subsystem or factored-gain assumptions are mildly violated, the formula should still hold approximately with an error controlled by the deviation from undirected symmetry or by the spread of the nodal matrices; the paper gives no such perturbation bound, so this would be new territory."],"forward_implications":["The performance of any large network in this class is computable from the Laplacian spectrum and one low-dimensional function; no high-dimensional Lyapunov solve is needed for each new graph.","Stabilizability of $(A,B)$ is equivalent to the existence of a feedback gain with finite $\\tilde\\lambda(K)$, and detectability plays the dual role for the observer design; both admit explicit LMI constructions.","For convex performance functions, purely graph-theoretic information (number of edges, maximum degree, total weight) yields lower bounds on performance and a sparsity--performance tradeoff; equality holds only for complete or star graphs in Theorem 11 and for complete equal-weight graphs in Theorem 12.","Over path and cycle graphs, the performance scales as $\\Theta(N\\Gamma_N)$ with $\\Gamma_N$ an explicit integral of $\\varphi$, yielding concrete laws such as $\\Theta(N^2)$ for single integrators and $\\Theta(N^4)$ for double integrators with $C=[1,0]$.","Non-minimum-phase nodal dynamics impose an unavoidable positive floor $\\operatorname{Tr}(E^TP_0E)$ on the achievable performance, independent of graph size or feedback gain.","The same spectral-sum decomposition also applies to the steady-state variance of the control effort $\\rho_u$, giving rational input functions $\\varphi_u(\\lambda,K)$ that inherit the design and scaling analysis."],"supporting_citations":[{"why":"Supplies the decoupled-subsystem analysis for state feedback that Theorem 1 extends to output feedback with measurement noise and explicit spectral expressions.","marker":"[12]"},{"why":"Supplies the Lyapunov-equation evaluation of the $H_2$ norm used to compute $\\varphi$ from the decoupled subsystems.","marker":"[26]"},{"why":"Provides the second-order consensus coherency baseline and scaling laws that the paper generalizes.","marker":"[6]"},{"why":"Supplies graph-theoretic $H_2$ bounds and scaling results that Theorem 11 and Corollary 2 extend.","marker":"[11]"},{"why":"Supplies the equivalence between stabilizability and LMI feasibility used in Theorem 5.","marker":"[16]"},{"why":"Supplies the convex-function eigenvalue inequality behind Theorem 11 and its equality cases.","marker":"[29]"},{"why":"Supplies the parametric-integral approximation idea behind Theorem 13 for path and cycle graphs.","marker":"[23]"}],"fun_headline_variants":["Network variance: rational sum over Laplacian eigenvalues","Steady-state variance: a sum of rational functions of eigenvalues","Control performance via graph spectrum: a unifying formula","Eigenvalue-based performance bounds for networked linear control","Closed-form network performance from Laplacian eigenvalues"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every subsystem is identical and every feedback gain factors as a scalar weight times a common matrix $K$ over a fixed undirected weighted graph; only then do the Laplacian eigenvectors simultaneously diagonalize the closed-loop dynamics, and if the graph were directed, the nodes heterogeneous, or the gains not in this factored form, the summation formula (11) would no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["Network variance: rational sum over Laplacian eigenvalues","Steady-state variance: a sum of rational functions of eigenvalues","Control performance via graph spectrum: a unifying formula","Eigenvalue-based performance bounds for networked linear control","Closed-form network performance from Laplacian eigenvalues"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1448,"prompt_tokens":941,"completion_tokens":507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":431}},"tokens_in":557,"tokens_out":507,"duration_ms":5378,"temperature":1.0,"reasoning_tokens":431,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:13:25.043148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a four-node network of double integrators with $A=\\begin{bmatrix}0&1\\\\0&0\\end{bmatrix}$, $B=E=[0,1]^T$, $C=[1,0]$, $H=I_2$, unit disturbance on one channel, and an undirected weighted path with arbitrary weights; compute the steady-state variance $\\rho$ both by solving the full $8$-dimensional Lyapunov equation for (8) and by evaluating $\\sum_{i=2}^{4}\\varphi(\\lambda_i,K)$ from (13). If the two numbers disagree beyond numerical tolerance, the spectral formula is false. As an even more direct check, for a fixed $K$ evaluate $\\varphi(\\lambda,K)$ symbolically for a $\\lambda$ that is not a Laplacian eigenvalue and verify that it is rational with denominator $\\det((A-\\lambda BKH)\\otimes I_n+I_n\\otimes(A-\\lambda BKH))$.","supporting_citations":[{"cited_title":"State- space solutions to standard H2 and H∞ control problems,","cited_arxiv_id":null,"evidence_quote":"Supplies the Lyapunov-equation evaluation of the $H_2$ norm used to compute $\\varphi$ from the decoupled subsystems."},{"cited_title":"Coherence in large-scale networks: Dimension-dependent limitations of local feed- back,","cited_arxiv_id":null,"evidence_quote":"Provides the second-order consensus coherency baseline and scaling laws that the paper generalizes."},{"cited_title":"Fundamental limits and tradeoffs on distur- bance propagation in linear dynamical networks,","cited_arxiv_id":null,"evidence_quote":"Supplies graph-theoretic $H_2$ bounds and scaling results that Theorem 11 and Corollary 2 extend."},{"cited_title":"On sum of powers of the laplacian eigenvalues of graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the convex-function eigenvalue inequality behind Theorem 11 and its equality cases."},{"cited_title":"Estrada index of cycles and paths,","cited_arxiv_id":null,"evidence_quote":"Supplies the parametric-integral approximation idea behind Theorem 13 for path and cycle graphs."}],"review_version":1}