{"id":"4d98a52d-626d-44ba-92a3-2dd22dadd09a","arxiv_id":"1908.06250","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Two distributed protocols built from a cofactor-weighted mirror graph achieve signed-average consensus on strongly connected signed digraphs if and only if the digraph is structurally balanced, with a nonlinear fixed-time variant.","lead":"This paper designs new rules for groups of connected agents with both cooperative and antagonistic interactions, ensuring they settle on a signed average value even when the interaction graph is directed and uneven. The result extends known consensus theory to a broader class of signed networks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's proof uses a false equality 1^T W L = 0 for the signed Laplacian; the balanced-case limit is not established as written.","rationale":"The most load-bearing concern is the false linear algebra assertion in the proof of Theorem 2, because it sits inside the proof of the central if-and-only-if claim for protocol (12). The reader's weakest_assumption was the digon sign-symmetry assumption, which is a real scope limitation and is also noted in the rationale. However, the false assertion is more directly threatening: it makes the balanced-case proof invalid even within the paper's stated assumptions. The counterexample above is within the stated assumptions (strongly connected, digon sign-symmetric, balanced), so the proof's intermediate step is objectively false. The theorem's conclusion may still be salvageable because the left eigenvector D1 can be verified directly, but the manuscript as written does not contain that verification. The reader's verdict of CONDITIONAL remains appropriate: this is a proof gap that can be repaired, not a demonstrated counterexample to the theorem itself. The digon sign-symmetry scope issue is secondary; it narrows the advertised generality but is an explicit assumption in Section III-A. I therefore recommend no change to the reader's conditional verdict.","tokens_in":18309,"tokens_out":29820,"duration_ms":248716,"concrete_test":"Take the 3-node graph G with all weights equal to 1: edges (1,2),(2,1) positive; (2,3),(3,2) negative; (3,1),(1,3) negative. Compute unsigned Laplacian L̄ and its cofactor matrix W (here W = 3I), and signed Laplacian L_s = diag(2,2,2) - A with A = [[0,1,-1],[1,0,-1],[-1,-1,0]]. Verify 1^T W L_s = [6,6,12] ≠ 0, while for D = diag(1,1,-1), (D1)^T W L_s = 0 and W L_s D1 = 0. Then re-derive the limit of ẋ = -WL_s x using the left eigenvector D1 and confirm x(∞) = (1^T D x0/3) D1. This settles whether the false intermediate is essential: if the terminal formula holds without it, Theorem 2's result stands with a corrected proof; if not, the theorem's balanced case fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 2 asserts '[det(L11),...,det(Lnn)]L = 0^T_n' and 'It directly produces 1^T_n W L = 0^T_n', where L must be the signed Laplacian appearing in ẋ = -WLx. The cofactor left-null vector is a property of the induced unsigned Laplacian L̄, not of the signed Laplacian L_s. For any signed Laplacian with negative off-diagonal entries, 1^T W L_s = w^T L_s is generically nonzero. Concrete instance: the digon sign-symmetric, strongly connected, structurally balanced 3-node digraph with edges (1,2),(2,1) weight +1, (2,3),(3,2) weight -1, (3,1),(1,3) weight -1. Here L̄ = [[2,-1,-1],[-1,2,-1],[-1,-1,2]], W = 3I, and L_s = [[2,-1,1],[-1,2,1],[1,1,2]]. Then 1^T W L_s = [6,6,12] ≠ 0. The claimed equality is therefore false. The subsequent step that ν_l = D1 is a left eigenvector of WL is derived from this false equality; although (D1)^T W L_s = 0 can be proved directly for D = diag(1,1,-1), the published proof does not establish it. Consequently the balanced direction of Theorem 2, including the signed-average value formula, lacks a valid proof as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies distributed averaging over signed directed networks. It introduces two graph-theoretic constructions—the induced unsigned digraph and the mirror (undirected) signed graph—and uses cofactors of the induced unsigned Laplacian to define an \"improved Laplacian potential\" that is a quadratic form in the signed Laplacian even when the digraph is not weight balanced. Two linear nearest-neighbor protocols are proposed, one based on the mirror Laplacian and one on the cofactor-weighted signed Laplacian, and the paper claims that signed-average consensus (respectively, state stability) is achieved if and only if the strongly connected signed digraph is structurally balanced (respectively, unbalanced). A nonlinear protocol is then claimed to achieve fixed-time bipartite consensus or state stability with a settling-time bound independent of initial conditions. The simulations illustrate the claims. All results are explicitly conditional on a digon sign-symmetry assumption stated in Section III-A.","tokens_in":18611,"tokens_out":16902,"duration_ms":145848,"significance":"If the theorems are correct, the paper makes a valuable contribution by removing the weight-balance restriction for signed-average consensus on directed signed networks and by providing a Lyapunov-based convergence analysis that works for weight-unbalanced topologies. The fixed-time extension is a useful application of the improved potential. The paper builds transparently on established tools (the cofactor property from [37] and the fixed-time lemmas from [30]) rather than assuming its conclusions. However, the proof of Theorem 2 contains a false intermediate assertion, and the proof of Lemma 3 has a gap; these are repairable, but as written they compromise the validity of the main if-and-only-if claims.","major_comments":[{"comment":"The proof asserts '[det(L11),...,det(Lnn)]L = 0_n^T' for the signed Laplacian L appearing in (15), citing [37]. The cofactor vector is a left null vector of the induced unsigned Laplacian L̄ = Δ − |A|, not of the signed Laplacian L_s = Δ − A. For the digon sign-symmetric, strongly connected, structurally balanced digraph with adjacency A = [[0,1,-1],[1,0,-1],[-1,-1,0]], one computes L_s = [[2,-1,1],[-1,2,1],[1,1,2]], W = 3I_3, and 1_3^T W L_s = [6,6,12] ≠ 0. Thus the claimed equality is false, and the subsequent identification of ν_l = D_3 1_3 as a left eigenvector of WL_s is not established by the published proof. The conclusion is salvageable: from L_s = D L̄ D and w^T L̄ = 0 one obtains (D1)^T W L_s = (Dw)^T L̄ D = 0. This step must be written out, and the signed-average value formula then follows by normalizing the left eigenvector correctly. As it stands, the balanced direction of Theorem 2 lacks a valid proof.","section":"Theorem 2 proof, Section IV (system (15) and Eq. (3))"},{"comment":"In the converse direction of Lemma 3(1), the paper asserts that from N(L̂) = N(L) = span{D_n 1_n} one can derive D_n L D_n 1_n = L 1_n = 0_n, and that this implies D_n L D_n = L. Neither step is valid for a signed Laplacian: L 1_n is not generally zero (it equals the vector of signed row sums), and a single null vector does not determine the matrix. The structural-balance conclusion does follow from N(L) = span{D_n 1_n} via known results for signed Laplacians, but the argument as written does not establish it. Since Lemma 3 underpins Lemma 4 and therefore the 'only if' halves of Theorems 1 and 2, this gap needs to be repaired.","section":"Lemma 3, Appendix C"},{"comment":"The paper assumes throughout that the signed digraph is digon sign-symmetric (a_ij a_ji ≥ 0 for all i,j), yet the abstract promises 'general directed topologies' and the introduction claims 'arbitrary directed signed networks.' This assumption is essential: without it, the mirror signed graph in Lemma 2 need not preserve the sign pattern, and the equivalence of structural balance between G and Ĝ can fail. For example, a two-node graph with a_12 = 1 and a_21 = −1 has cofactors det(L̄_11) = det(L̄_22) = 1, giving â_12 = (1·1 + 1·(−1))/2 = 0, so the mirror graph is disconnected and Lemma 2(1) fails. The paper should either qualify its claims to 'strongly connected, digon sign-symmetric signed digraphs' or prominently state this restriction as a limitation of the proposed approach.","section":"Section III-A vs. Abstract and Introduction"}],"minor_comments":[{"comment":"The symbol L is used for both the signed Laplacian and the induced unsigned Laplacian; Definition 1 and Eq. (3) are ambiguous. The proof error in Theorem 2 is a direct consequence of this notational conflation. Please use L_s and L̄ consistently.","section":"Throughout"},{"comment":"The settling-time formula (19) appears to have a typesetting error: the term involving k_1 should be n^{(m-r)/(2r)}/k_1^{r/(m-r)}, as follows from Lemma 7 with α_1 = k_1 n^{(r-m)/(2r)} λ_2(L̂), rather than n^{(m-r)/(2r)} k_1^{r/(m-r)}. Please verify the displayed formula.","section":"Theorem 3, Eq. (19)"},{"comment":"The statement that Ĝ 'has the same sign pattern as G' is imprecise: Ĝ is undirected and may contain an edge where only one directed edge exists in G, and the relevant claim is about the sign of any nonzero â_ij. Consider rewording for clarity.","section":"Lemma 2"},{"comment":"Table I reports only the settling-time upper bounds; the eigenvalues λ_1(L̂) and λ_2(L̂) used in (19)-(20) should be reported so that the numerical values can be reproduced.","section":"Table I"},{"comment":"The derivation of (23) from (22) invokes [30, Lemma 4] but does not state it; please state the lemma or make the application explicit, since it is essential to the fixed-time bound.","section":"Proof of Theorem 3"},{"comment":"There are several typos and notational slips, e.g., 'Lapalcian' in Section III-A, and the repeated use of 'L = Δ − A' with different meanings in Definition 1 and the surrounding text. A careful notation and proofreading pass is needed.","section":"Editorial"}],"recommendation":"major_revision","confidential_remarks":"The core idea of the paper—cofactor-weighted Laplacian potentials for signed digraphs—is promising and likely publishable after a careful revision. The main proofs rely on external results (cofactor properties from [37] and fixed-time lemmas from [30]), so there is no circularity concern. The decisive issue is that the proof of Theorem 2 contains a false equality and the proof of Lemma 3 has an invalid implication; both are repairable, but the authors must also correct the notation that conflates the signed and unsigned Laplacians. I would ask the editor to treat the revision as requiring a full re-check of the appendices as well as the main theorems."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: this paper has a genuinely useful construction—the cofactor-weighted improved Laplacian potential—and a correct-looking main theorem for signed-average consensus on strongly connected directed signed graphs without weight balance. But the proof of Theorem 2 contains a false linear algebra step, and the 'general directed topologies' in the abstract oversells the assumptions.\n\nWhat's new: building Φ_e(x) = x^T(WL+L^TW)x with W = diag(det(\\bar{L}_{ii})) is a neat idea. It makes the Laplacian potential well-defined for weight-unbalanced digraphs, and Lemma 4's equivalence between Φ_e(x)=0 and the balanced/unbalanced subspace is correct. The mirror signed graph construction in Theorem 1 is also clean, and the fixed-time extension in Theorem 3 follows the standard comparison-principle template. Simulations support the claims.\n\nThe soft spots, in order of severity.\n\nFirst, Theorem 2's proof. The cofactor property from [37] applies to the induced unsigned Laplacian \\bar{L}: w^T \\bar{L}=0. The paper then asserts this gives 1^T W L=0 where L is the signed Laplacian. That is false. For a 3-node structurally balanced digraph with one positive edge and two negative edges, 1^T W L = [6,6,12] ≠ 0. The good news: the conclusion ν_l = D1 is still true, and it follows directly because L = D\\bar{L}D and DWD=W, so (D1)^T W L = 1^T W \\bar{L} D = 0. But the printed argument needs repair, and until then the signed-average value formula in Theorem 2 is unproven.\n\nSecond, the digon sign-symmetry assumption (a_ij a_ji ≥ 0) is not in the abstract. It's stated in Section III-A, but 'arbitrary directed signed networks' or 'general directed topologies' is misleading. This assumption is needed for Lemma 2 to preserve sign patterns.\n\nThird, the protocols are called nearest-neighbor, but each weight det(L_ii) or \\hat{a}_ij requires global knowledge of the graph topology. That's a standard caveat in this literature, but it should be explicit.\n\nThe citation pattern is fine; the self-citations are to published results that are actually used.\n\nWho should read this: anyone working on signed consensus or distributed averaging with antagonistic interactions. The improved potential will probably be cited even if the proof gap takes a round of revision. I'd send it to review; the fix is short, and the core construction is worth having in the literature.","headline":"Useful cofactor-weighted Laplacian construction for signed-average consensus on directed graphs, but Theorem 2's proof needs a short repair.","tokens_in":19124,"tokens_out":8721,"would_cite":true,"duration_ms":71754,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93A16","93D50","05C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that signed-average consensus on strongly connected directed signed networks is achievable exactly when the signed digraph is structurally balanced—and state stability exactly when it is unbalanced—without requiring weight…","keywords":["signed networks","signed-average consensus","distributed averaging","structural balance","directed signed digraphs","Laplacian potential","fixed-time consensus","multi-agent systems"],"falsifier":"Run protocol (12) on a strongly connected, digon sign-symmetric signed digraph that is structurally balanced but has unequal in- and out-degrees, starting from random initial states; the paper predicts the terminal state is exactly $\\frac{1}{n}(1_n^T D_n x_0) D_n 1_n$. If the computed terminal vector differs beyond numerical error, or if $\\hat L = (W L + L^T W)/2$ has a zero eigenvector not proportional to $D_n 1_n$, the core dichotomy fails. Repeating the test with a reciprocal edge pair whose weights have opposite signs, violating digon sign-symmetry, locates the boundary of the assumption.","tokens_in":1837,"feed_emoji":"⚖️","tokens_out":7006,"duration_ms":182116,"temperature":0.7,"pith_summary":"In a signed network, agents cooperate with some neighbors and compete with others, and the goal of distributed averaging is for every agent to end at the same magnitude, with signs prescribed by the network's camp structure. Earlier solutions required the digraph to be weight balanced, meaning each agent's total incoming influence had to equal its outgoing influence. This paper removes that restriction for strongly connected directed topologies by introducing an improved Laplacian potential, built by weighting each edge with a cofactor determinant of the induced unsigned Laplacian and symmetrizing the graph into an undirected mirror signed graph. Using either of two proposed nearest-neighbor protocols, the paper proves signed-average consensus if and only if the signed digraph is structurally balanced, and state stability, meaning all states converge to zero, if and only if it is structurally unbalanced. The same potential also yields fixed-time bipartite consensus or stability, with settling-time bounds independent of initial conditions.","feed_headline":"Signed-average consensus no longer needs weight balance","feed_subtitle":"Two local protocols achieve it exactly when the signed digraph is structurally balanced.","key_machinery":"The central object is the improved Laplacian potential function $\\Phi_e(x) = \\sum_{i,j} \\det(L_{ii}) |a_{ij}| (x_i - \\operatorname{sgn}(a_{ij}) x_j)^2 = x^T (W L + L^T W) x$, where $W$ is the diagonal matrix of cofactors $\\det(L_{ii})$ of the induced unsigned Laplacian. For a strongly connected digraph these cofactors are strictly positive and, as a vector, form a left eigenvector of $L$ for eigenvalue zero; this is what makes the mirror Laplacian $\\hat L = (W L + L^T W)/2$ symmetric with the same null-space structure as $L$. The potential does the work of a Lyapunov function: along $\\dot x = -\\hat L x$, the derivative of $V(x) = x^T x$ is exactly $-\\Phi_e(x)$, so LaSalle's invariance principle turns the null-space dichotomy of $\\hat L$ into the consensus-versus-stability dichotomy, with no weight-balance condition anywhere. The same potential supports the Lyapunov estimate for the fixed-time nonlinear protocol in Theorem 3.","core_discovery":"On the paper's own terms, the central discovery is that weight imbalance was never the real obstruction to signed-average consensus on directed signed networks; the sign pattern is. For a strongly connected signed digraph $G$, take absolute values of the edge weights to form the induced unsigned digraph, let $W = \\mathrm{diag}(\\det(L_{11}), \\dots, \\det(L_{nn}))$ be the diagonal matrix of principal-minor cofactors of its Laplacian, and define the mirror signed graph by the symmetrized Laplacian $\\hat L = (W L + L^T W)/2$. Because the vector of cofactors is a left eigenvector of $L$ for eigenvalue $0$, $\\hat L$ is a genuine symmetric Laplacian, and its null space is $\\mathrm{span}\\{D_n 1_n\\}$ exactly when $G$ is structurally balanced and $\\{0\\}$ otherwise. The systems $\\dot x = -\\hat L x$ (protocol (9)) and $\\dot x = -W L x$ (protocol (12)) therefore converge, respectively, to the signed average $(1/n)\\sum_j \\sigma_j x_{j0}$ applied with the gauge signs, or to zero. Theorems 1 and 2 state these as if-and-only-if results: structural balance is necessary as well as sufficient, and structural unbalance is necessary as well as sufficient for stability.","pith_inferences":["A natural next step, not pursued here, is relaxing digon sign-symmetry; the mirror-graph sign-preservation lemma is where the assumption enters, so mixed-sign reciprocal edge pairs are the first place to look for counterexamples.","The cofactor gains used in the protocols are principal minors of the global Laplacian, so although the update rules are local, the protocols are not fully distributed in information; estimating these gains online or replacing them with local surrogates would make the design practically distributed.","The symmetrization recipe—weighting a directed Laplacian by a positive diagonal matrix from its left null space before symmetrizing—does not depend on the averaging problem and could transfer other directed-network results, such as synchronization, formation control, and opinion dynamics, to weight-unbalanced digraphs.","A quantitative prediction is that the convergence rate of the mirror-Laplacian protocol is governed by the spectral gap of the mirror graph; changing edge weights while keeping sign patterns would test whether weight imbalance affects only the rate, not the outcome."],"forward_implications":["Signed-average consensus becomes attainable on every strongly connected directed signed network, weight-balanced or not, and the limiting value is the gauge-transformed average of the initial states.","The two linear protocols (9) and (12) give simple implementations: one is the gradient of the improved potential, the other applies cofactor gains to the classical neighbor rule.","The balanced/unbalanced dichotomy is sharp: structurally unbalanced digraphs drive all states to zero rather than to a polarized consensus, so the sign pattern alone decides the outcome.","With the nonlinear protocol (18), the same conclusions hold in fixed time, with explicit upper bounds on the settling time that do not depend on initial states.","Because unsigned networks are signed networks with all-positive edges, both protocols also solve average consensus on strongly connected unsigned digraphs without weight balance."],"supporting_citations":[{"why":"Defines bipartite and signed-average consensus on signed networks and gives the classical protocol that requires strong connectivity plus weight balance, the baseline this paper extends.","marker":"[17]"},{"why":"Supplies the M-matrix criteria for directed signed networks used in Lemma 5 to characterize eigenvalues of W L in the balanced and unbalanced cases.","marker":"[26]"},{"why":"Provides the key algebraic facts that det(L_ii) is positive and that the vector of these cofactors is a left eigenvector of L for eigenvalue zero, making the mirror Laplacian well defined.","marker":"[37]"},{"why":"Contributes the fixed-time comparison lemmas and the undirected-topology baseline that Theorem 3 extends to directed signed networks.","marker":"[30]"},{"why":"Provides LaSalle's invariance theorem, the comparison principle, and the asymptotic stability theorems used throughout the convergence proofs.","marker":"[38]"},{"why":"Gives the semi-cycle criterion for structural balance of marked digraphs, used to prove that the mirror signed graph preserves the sign pattern of G.","marker":"[39]"}],"fun_headline_variants":["Sign pattern, not weight balance, decides signed consensus","Structural balance is the true condition for signed-average consensus","Forget weight balance: structural balance gives signed consensus","Structural balance, not weight balance, enables signed consensus"],"cache_read_input_tokens":21248,"weakest_assumption_plain":"The load-bearing assumption is that the signed digraph is digon sign-symmetric—for every pair of agents, the product of the two reciprocal edge weights is nonnegative—so that no two agents influence each other with opposite signs; the mirror graph inherits the sign pattern only under this condition, and the if-and-only-if consensus and stability theorems depend on that inheritance.","fun_headline_variants_meta":{"raw":{"variants":["Sign pattern, not weight balance, decides signed consensus","Structural balance is the true condition for signed-average consensus","Forget weight balance: structural balance gives signed consensus","Structural balance, not weight balance, enables signed consensus"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00127,"raw_usage":{"total_tokens":5244,"prompt_tokens":1037,"completion_tokens":4207,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":4144}},"tokens_in":653,"tokens_out":4207,"duration_ms":31150,"temperature":1.0,"reasoning_tokens":4144,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:52:58.191510+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run protocol (12) on a strongly connected, digon sign-symmetric signed digraph that is structurally balanced but has unequal in- and out-degrees, starting from random initial states; the paper predicts the terminal state is exactly $\\frac{1}{n}(1_n^T D_n x_0) D_n 1_n$. If the computed terminal vector differs beyond numerical error, or if $\\hat L = (W L + L^T W)/2$ has a zero eigenvector not proportional to $D_n 1_n$, the core dichotomy fails. Repeating the test with a reciprocal edge pair whose weights have opposite signs, violating digon sign-symmetry, locates the boundary of the assumption.","supporting_citations":[{"cited_title":"Consensus problems on networks with antag onistic inter- actions,","cited_arxiv_id":null,"evidence_quote":"Defines bipartite and signed-average consensus on signed networks and gives the classical protocol that requires strong connectivity plus weight balance, the baseline this paper extends."},{"cited_title":"Convergence analysis of directed signed netw orks via an M matrix approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the M-matrix criteria for directed signed networks used in Lemma 5 to characterize eigenvalues of W L in the balanced and unbalanced cases."},{"cited_title":"Algebraic criteria for consensus prob lems of continuous-time network systems,","cited_arxiv_id":null,"evidence_quote":"Provides the key algebraic facts that det(L_ii) is positive and that the vector of these cofactors is a left eigenvector of L for eigenvalue zero, making the mirror Laplacian well defined."},{"cited_title":"Finite-time consensus for mu ltiagent sys- tems with cooperative and antagonistic interactions,","cited_arxiv_id":null,"evidence_quote":"Contributes the fixed-time comparison lemmas and the undirected-topology baseline that Theorem 3 extends to directed signed networks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides LaSalle's invariance theorem, the comparison principle, and the asymptotic stability theorems used throughout the convergence proofs."},{"cited_title":"Consistency in marked digr aphs,","cited_arxiv_id":null,"evidence_quote":"Gives the semi-cycle criterion for structural balance of marked digraphs, used to prove that the mirror signed graph preserves the sign pattern of G."}],"review_version":1}