{"id":"bc484eb5-4fb5-498c-a3b8-d3cc182112e3","arxiv_id":"1908.03588","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A passivity-based fault detection and isolation framework for nonlinear diffusively-coupled multi-agent networks, with graph-theoretic guarantees on the number of simultaneously isolable link faults.","lead":"This paper designs monitoring protocols that detect when communication links fail in multi-agent control networks and identify which links are broken, even when the agent dynamics are nonlinear. The method adds random test signals to the link controllers and uses passivity theory to guarantee that different missing-link patterns produce different steady-state outputs, so up to k-2 simultaneous failures can be isolated if the network graph is k-connected.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 8 is conditional on a perfect convergence-assertion oracle; the proposed oracle constructions require assumptions not present in the theorem, so the stated guarantee overstates what is proved.","rationale":"The graph-theoretic machinery in Theorems 4 and 5 is convincing, and the random edge-indication vector construction is a genuine contribution. However, the headline result Theorem 8 is not established for general MEIP agents: every detection and isolation guarantee passes through Assumption 4, a perfect finite-time convergence-assertion oracle. The only concrete candidate algorithms for this oracle require additional structural assumptions (Assumptions 2 and 3, full state measurement) and regularity conditions (compact sublevel sets, C^1 power-law behavior of Ω_i) that are not part of Theorem 8's hypotheses. The paper itself acknowledges the oracle assumption in Section VI, but the theorem statements and the abstract do not carry that caveat. This is a fixable but substantive gap: the central claim should be restated with the oracle as an explicit condition, or with the additional hypotheses needed to realize it. Because the underlying construction is promising and the identified gaps are of a completeness/assumption-matching nature rather than a fundamental flaw, the CONDITIONAL verdict is appropriate; no change to the reader's verdict is needed.","tokens_in":28226,"tokens_out":16475,"duration_ms":169030,"concrete_test":"Check whether Algorithm 7 can be initialized and run for an MEIP system satisfying Assumption 1 but not Assumption 3, with bounded output h (e.g., h(x)=tanh(x), q(x)=1, f increasing so the agent is MEIP) and a static edge controller g. Specifically, compute S(0) for a large initial condition and see whether the inverse Ω of ω is defined there; if S(0) exceeds the domain of Ω, Algorithm 7 has no value for M and Theorem 11's proof breaks at its first step. This directly tests whether the convergence-assertion oracle is realizable under the theorem's stated assumptions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V's Theorems 6–8 all invoke Assumption 4, the existence of a convergence-assertion algorithm A that never gives false positives or false negatives. The two algorithms proposed in Section VI do not realize this assumption under the hypotheses of Theorem 8. The high-rate-sampling method (Theorem 9, via Propositions 4–5) requires the maxima in Proposition 4 to be finite, which demands compactness of sublevel sets of S that is not stated in Assumption 1; the bound M can be infinite. The convergence-profile method (Algorithm 7) assumes Assumption 3 (control-affine agents and static edge controllers) and requires full knowledge of x(t); neither is assumed in Theorem 8. Moreover, Theorem 10 constructs Ω as the inverse of ω(θ)=sup_{x:(h(x)−h(x̂))^2≤θ}S(x); as the paper itself notes, ω(θ) can be infinite when h is bounded, so Ω is not defined for large S(0). Algorithm 7's step 6 then has no well-defined M. Proposition 6 further requires the Ω_i to be C^1 with power-law behavior, while Theorem 10 only yields a strictly ascending (possibly non-C^1) inverse. Thus the guarantee of Theorem 8—that Algorithms 5 and 6 solve Problem 2 with probability 1 under Assumption 1 alone—does not follow from the supplied proofs; it holds only conditionally on an oracle whose concrete instantiations need extra structural and regularity assumptions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops a passivity-based framework for detecting and isolating communication-link faults in diffusively coupled networks of maximal equilibrium-independent passive (MEIP) systems. It introduces \"edge-indication vectors,\" constant biases added to the edge-controller outputs, and proves (Theorems 4 and 5) that a generically chosen bias makes distinct underlying graphs have distinct steady-state outputs; in the 2-connected case the bias can be chosen in the kernel of the nominal incidence operator so that the faultless equilibrium is preserved. Assuming a perfect \"convergence assertion\" oracle (Assumption 4), the paper then gives detection, adversarial-planning, and fault-isolation algorithms, with a k-connected graph isolating up to k-2 link faults (Theorems 6-8). Section VI proposes two concrete oracle implementations: a high-rate sampling inequality check (Theorem 9) and a convergence-profile method based on storage functions and Omega functions (Theorem 11). A vehicle-velocity coordination case study illustrates the protocol.","tokens_in":28431,"tokens_out":14518,"duration_ms":149668,"significance":"The paper's main conceptual contribution is the edge-indication vector construction. Theorem 5 is a genuine and elegant extension of the authors' earlier network-identification work: the measure-zero intersection argument together with the two-coloring lemma gives a graph-theoretic condition under which the nominal steady state is preserved while faulty configurations are asymptotically separated. The k-connected isolation guarantee for up to k-2 faults is appealing and, if fully established, would be a substantial advance over LTI-only FDI methods. The paper is also honest in separating the asymptotic differentiation part from the online assertion part. However, the advertised guarantees in Theorems 6-8 are conditional on an oracle that, as shown below, is not instantiated under the same assumptions; the revision should either strengthen the hypotheses or weaken the claims accordingly.","major_comments":[{"comment":"Theorem 8 is stated under Assumption 1 alone ('Then, with probability 1, Algorithms 5 and 6 ... solve Problem 2'), but its proof invokes Assumption 4 through the instances of the convergence assertion protocol in Algorithm 6. The concrete protocols in Section VI do not realize Assumption 4 under Assumption 1: Algorithm 7 requires Assumption 3 and full knowledge of x(t), while Theorem 9's sampling interval is only defined when the bound M in Proposition 4 is finite. The stated guarantee therefore overstates what is proved; the theorem should be stated as conditional on Assumption 4, or the additional hypotheses needed by the chosen instantiation should be included in the theorem statement.","section":"Section V-C, Theorem 8"},{"comment":"Finiteness of the constant M is not guaranteed by the stated hypotheses. The maxima M_dot{x}, M_delta{y}, and M_delta{mu} are taken over the sublevel set B = {S <= S(x(t_k), eta(t_k))}; Assumption 1 does not imply that B is compact or that these maxima are finite, since the storage functions need not be coercive. The authors themselves note in the proof of Theorem 10 that omega(theta) can be infinite when h is bounded while S is not. If any of these maxima is infinite, the choice Delta t = delta/M in Theorem 9 and the term (M/2) Delta t^2 in equation (12) are undefined, so the high-rate sampling protocol does not implement Assumption 4 under the hypotheses stated.","section":"Section VI-A, Propositions 4-5 and Theorem 9"},{"comment":"The constructed function Omega is only defined as the inverse of omega on the set {theta : omega(theta) < infinity}; when h is bounded and S is unbounded, omega(theta) = infinity for large theta, so Omega is not a function [0,infinity) -> [0,infinity). Moreover, Theorem 10 yields only a strictly monotone inverse, while Proposition 6 and Corollary 2 require each Omega_i to be C^1 and to have a positive power-law limit at 0. Consequently the inequality dot{S} <= -C Omega_star(S) is not established from Theorem 10, and Algorithm 7's step 6 ('M = min_{x: S(x) >= delta} Omega(S(x))') has no well-defined value for states where S lies outside the domain of Omega.","section":"Section VI-B, Theorem 10 and Proposition 6"},{"comment":"The proof of Corollary 1 contains a sign error. From |dG/dt| <= M, the displayed inequality G(t) <= G(t_{k+1}) + M|t - t_{k+1}| becomes a lower bound on S(t_{k+1}) - S(t_k) after multiplying by -1 and integrating; it does not give the claimed upper bound. To obtain the stated inequality (12) one must instead use the lower Lipschitz bound G(t) >= G(t_k) - M(t - t_k) and integrate. As printed, the proof does not establish equation (12).","section":"Section VI-A, Corollary 1"}],"minor_comments":[{"comment":"The loop bound 'i_l < m' should be 'i_l <= m'; Algorithm 2 uses the correct bound.","section":"Algorithm 5, line 3"},{"comment":"The instruction 'Run steps 1-4 of Algorithm 1' should specify that the synthesis is performed for the graph Graphs(j), not for the original graph G; otherwise IP(j) is only defined for G and cannot serve as the stable-phase protocol for a faulty graph in Theorem 8.","section":"Algorithm 5, line 11"},{"comment":"The compact set written as [0,D]^n \\ {x : ||x|| > r} contains the origin, where F is undefined; the intended set should remove a neighborhood of the origin, e.g., {x : ||x|| >= r}.","section":"Proposition 6 proof"},{"comment":"There is a typographical error in the definition of M_delta{mu}: '|| psi((eta, E_G^T h(x)) - mu ||' is missing a closing parenthesis and should read '|| psi(eta, E_G^T h(x)) - mu ||'.","section":"Proposition 4"},{"comment":"The complexity statement first gives O(n^{cr}) for a universal constant c and then derives O(n^{2r}); the two bounds should be reconciled so that the exponent is stated exactly.","section":"Theorem 7 proof"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is self-aware about the oracle dependence in places (abstract, Assumption 4, Remark 4), and Theorems 4 and 5 appear sound, so the gaps are addressable by adding hypotheses to the oracle instantiations and by restating Theorems 6-8 as conditional on those hypotheses. I would not reject: the edge-indication vector framework is a real contribution. However, the current paper does not prove the headline guarantee as stated, because the convergence-assertion protocols in Section VI require extra regularity, coercivity, and measurement assumptions that are absent from Theorem 8."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look. The new idea is genuinely good: choose a random bias w in ker(E_G P_G) so that the nominal network's steady state is unchanged, but every faulty subgraph gets pushed to a different steady-state output. That turns network FDI into a steady-state distinguishing problem, with clean graph-theoretic limits—detect any number of faults if the graph is 2-connected, isolate up to k-2 if it is k-connected. Theorems 4 and 5 are convincing under Assumption 1; the implicit-function and measure-zero arguments hold up, and the random-vector trick in the kernel is a real contribution, not just a repackaging of earlier network-identification work.\n\nThe soft spot is the distance between Theorem 8 and what the paper actually proves. Everything in Section V relies on Assumption 4—a convergence assertion oracle with zero false positives and zero false negatives. That is fine as a conditional framework, but the two oracle constructions in Section VI do not deliver Assumption 4 under Theorem 8's hypotheses. The high-rate sampling method (Theorem 9) needs the bounds M in Proposition 4 to be finite, which requires compact sublevel sets of the storage function; Assumption 1 doesn't give you that, and the bound can be infinite. The convergence-profile method (Algorithm 7) assumes Assumption 3 (static edge controllers and control-affine agents), which is also not in Theorem 8. And Theorem 10 itself flags the problem: ω(θ) can be infinite when h is bounded, so Ω is not always defined; Proposition 6 needs a C^1 power-law inverse, but Theorem 10 only gives a strictly ascending inverse. These are fixable—add the regularity assumptions to the theorem statements or weaken the claims to 'provided the oracle exists'—but as written, the headline guarantee overstates what is proved.\n\nThe literature handling is fair; the authors lean on their own passivity/network-optimization framework, but the new results do not reduce to it. The case study is illustrative; no code or repeated trials, so treat it as a sanity check, not evidence. The writing is clear and the proofs are readable.\n\nThis paper deserves a serious referee. The core idea is valuable, the gaps are localized, and the fix is to align the claims with the assumptions. I'd send it out, and I'd push the authors to tighten Section VI so the oracle assumption is either met or made explicit as a separate hypothesis. Whoever works on passivity-based FDI or network identification will want to cite the edge-indication construction.","headline":"Smart passivity-based FDI trick with real isolation guarantees; the headline theorem overclaims because its oracle assumption is not realized by the paper's own algorithms.","tokens_in":29017,"tokens_out":2438,"would_cite":true,"duration_ms":24783,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93A14","93C10","93D05","05C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a random bias to edge controllers makes every faulty subgraph converge to a distinct steady-state output, allowing online detection and isolation of up to k-2 link faults in k-connected nonlinear networks.","keywords":["fault detection and isolation","multi-agent systems","maximal equilibrium-independent passivity","network optimization","edge-indication vectors","graph connectivity","convergence assertion","nonlinear diffusively-coupled networks"],"falsifier":"Run the full protocol (Algorithms 5 and 6) on a 3-connected MEIP network with two links deleted a short time apart, using the convergence-profile assertion test with sample times whose gaps are bounded away from zero, as Proposition 7 permits. If the exploratory phase ever gets stuck with every instance having declared a fault, or if the output settles at a value different from $y^\\star$, then Theorem 8 is false. Alternatively, find a 2-connected MEIP network and a bias $w\\in\\ker E_G P_G$ for which two different faulty subgraphs have identical steady-state outputs; that would contradict Theorem 5.","tokens_in":27944,"feed_emoji":"🔗","tokens_out":9095,"duration_ms":87008,"temperature":0.7,"pith_summary":"This paper sets out to prove that communication-link failures in a nonlinear multi-agent network can be found and repaired online as long as the network is sufficiently connected and the agents and controllers belong to the broad passivity class called maximal equilibrium-independent passive (MEIP). The key move is to inject a constant bias into each edge controller: with probability one, a random bias makes every different faulty subgraph converge to a different steady-state output, and a bias chosen in a particular kernel leaves the faultless equilibrium untouched. Assuming a perfect \"convergence assertion\" test, a 2-connected graph allows detection of any number of link faults, and a k-connected graph allows isolation of up to k-2 faults while the closed loop still converges to the desired output $y^\\star$. The paper also constructs two such assertion tests from passivity storage functions, one based on high-rate sampling and one on convergence profiles, and demonstrates the full protocol on a 20-vehicle velocity-coordination example. The result matters because it gives graph-theoretic, model-independent guarantees for fault detection and isolation in nonlinear networks, not just linear ones.","feed_headline":"Random bias isolates up to k-2 link faults in nonlinear networks","feed_subtitle":"When links fail, steady-state outputs give away which ones—and the network still hits its target.","key_machinery":"The machinery is built from three pieces. First, the steady-state characterization of a diffusively-coupled network: by MEIP passivity and network optimization, the closed-loop steady-state output solves $k^{-1}(y)+E_H\\gamma(E_H^\\top y)+E_H P_H w=0$, where $k^{-1}$ is the inverse steady-state relation of the agents, $\\gamma$ that of the edge controllers, $E_H$ the incidence matrix, and $P_H$ the projection onto the edges of $H$. Second, edge-indication vectors, constant biases $w$ added to controller outputs, together with transversality arguments: because the Jacobian $\\nabla k^{-1}(y)+E_H\\nabla\\gamma(E_H^\\top y)$ is positive definite under Assumption 1, distinct subgraphs give manifolds of full dimension whose intersections have measure zero, so a random $w$ separates them almost surely. Third, the convergence assertion protocol, an oracle algorithm $A$ that decides whether the running network converges to a conjectured limit; two passivity-based implementations are given, one checking a sampled dissipation inequality and one using convergence profiles, functions $\\Omega$ with $\\Omega(S_i(x_i))\\le (y_i-\\bar y_i)^2$, to turn the storage-function decrease into the scalar inequality $\\dot S\\le -C\\Omega_\\star(S)$.","core_discovery":"On the paper's own terms, the central discovery is that network faults are asymptotically distinguishable by steady-state outputs: if every agent is MEIP and every controller is output-strictly MEIP, then for a constant bias vector $w$ added to the controller outputs, the steady-state output $y$ of the network with graph $H$ is the unique solution of $k^{-1}(y)+E_H\\gamma(E_H^\\top y)=-E_H P_H w$, so different subgraphs $H$ generally give different limits. A random $w$ is a $G$-edge-indication vector with probability 1 (Theorem 4), and when $G$ is 2-connected a random $w$ in $\\ker E_G P_G$ distinguishes the faultless graph $G$ from every faulty subgraph while leaving the nominal equilibrium $y^\\star$ unchanged (Theorem 5). With a convergence assertion algorithm $A$ that never lies (Assumption 4), this turns into online fault detection (Theorem 6), a winning planner strategy in an adversarial edge-removal game (Theorem 7), and, when $G$ is $k$-connected with $k\\ge 3$, detection and isolation of up to $r=k-2$ faults with guaranteed convergence to $y^\\star$ (Theorem 8). The paper further claims that $A$ can be realized by two passivity-based tests: high-rate sampling of a dissipation inequality (Theorem 9) and a convergence-profile test for control-affine agents (Theorem 11).","pith_inferences":["The paper's identification of faulty links as a side effect suggests the same steady-state separation could serve as a passive network-topology identification method driven by a single persistent bias, without the need to apply a sequence of distinct excitation signals.","If the perfect convergence-assertion oracle is replaced by a probabilistic or approximate test, the exact guarantees would likely become high-probability guarantees, and whether the connectivity thresholds survive under measurement noise is an open question.","Remark 2 indicates the template extends beyond edge failures to vertex failures and hybrid transceiver failure models, as long as the candidate faulty subgraphs can be enumerated and separated by steady-state outputs.","Since the separation proofs rely on transversality of full-dimensional manifolds in the bias space, the dimension of the bias space may impose a converse bound on how many distinct faulty graphs can be isolated, which could yield a lower-bound counterpart to Theorem 8."],"forward_implications":["A network whose interaction graph is 2-connected can be monitored for any number of link failures without knowing the failure times, as long as the convergence-assertion oracle is available.","If the graph is $k$-connected with $k\\ge 3$, up to $k-2$ links can fail, in any order, and the protocol still drives the outputs to $y^\\star$ while identifying the failed links as a side effect.","The number of isolable faults is a purely graph-theoretic quantity ($k-2$), so the same design works for any MEIP agents and controllers on that graph.","In the adversarial-game version, the planner wins with probability 1, with polynomial-time synthesis when $r=O(1)$ and a broadcast message of only $O(r\\log n)$ bits.","Two concrete assertion tests are available: high-rate sampling works for general MEIP networks, and the convergence-profile test works for control-affine agents with strictly monotone output maps."],"supporting_citations":[{"why":"Supplies the MEIP framework connecting steady-states of diffusively-coupled networks to network optimization, including the closed-loop convergence theorem used throughout.","marker":"[25]"},{"why":"Supplies the synthesis theorem and the steady-state condition $0\\in k^{-1}(y)+E_G\\gamma(E_G^\\top y)$ that the fault-monitoring scheme builds on.","marker":"[28]"},{"why":"Introduces the indication-vector idea from network identification that edge-indication vectors generalize to the fault-detection setting.","marker":"[14]"},{"why":"Provides the control-affine MEIP sufficient conditions and storage-function construction used in Theorem 1 and in the convergence-profile protocol.","marker":"[33]"},{"why":"Supplies Menger's theorem and the graph-connectivity notions underlying the 2-connected and k-connected guarantees.","marker":"[31]"},{"why":"Provides the maximal monotone relation and network-optimization duality formalism used to define MEIP and to justify the steady-state equations.","marker":"[26]"}],"fun_headline_variants":["Random bias isolates up to k-2 link faults in nonlinear nets","Steady-state outputs locate up to k-2 communication link faults","Detect link failures in nonlinear networks via output data","Up to k-2 link faults isolated by random bias in MEIP networks","Network fault detection via steady-state outputs and random bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on having a convergence-assertion algorithm that is a perfect oracle: it never says \"no\" when the network is converging to the guessed limit and eventually says \"no\" otherwise, and the paper's two concrete constructions only guarantee this under additional structural and regularity assumptions, so if no such oracle can be implemented, the detection and isolation guarantees degrade.","fun_headline_variants_meta":{"raw":{"variants":["Random bias isolates up to k-2 link faults in nonlinear nets","Steady-state outputs locate up to k-2 communication link faults","Detect link failures in nonlinear networks via output data","Up to k-2 link faults isolated by random bias in MEIP networks","Network fault detection via steady-state outputs and random bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000809,"raw_usage":{"total_tokens":3596,"prompt_tokens":1035,"completion_tokens":2561,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":2474}},"tokens_in":651,"tokens_out":2561,"duration_ms":21956,"temperature":1.0,"reasoning_tokens":2474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:08:37.815454+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full protocol (Algorithms 5 and 6) on a 3-connected MEIP network with two links deleted a short time apart, using the convergence-profile assertion test with sample times whose gaps are bounded away from zero, as Proposition 7 permits. If the exploratory phase ever gets stuck with every instance having declared a fault, or if the output settles at a value different from $y^\\star$, then Theorem 8 is false. Alternatively, find a 2-connected MEIP network and a bias $w\\in\\ker E_G P_G$ for which two different faulty subgraphs have identical steady-state outputs; that would contradict Theorem 5.","supporting_citations":[{"cited_title":"Duality and network theory in passivity-based cooperative control,","cited_arxiv_id":null,"evidence_quote":"Supplies the MEIP framework connecting steady-states of diffusively-coupled networks to network optimization, including the closed-loop convergence theorem used throughout."},{"cited_title":"Analysis and synthesis of MIMO multi-agent systems using network optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the synthesis theorem and the steady-state condition $0\\in k^{-1}(y)+E_G\\gamma(E_G^\\top y)$ that the fault-monitoring scheme builds on."},{"cited_title":"Network identification: A passivity and network optimization approach,","cited_arxiv_id":null,"evidence_quote":"Introduces the indication-vector idea from network identification that edge-indication vectors generalize to the fault-detection setting."},{"cited_title":"Model-Free Practical Cooperative Control for Diffusively Coupled Systems","cited_arxiv_id":"1906.05204","evidence_quote":"Provides the control-affine MEIP sufficient conditions and storage-function construction used in Theorem 1 and in the convergence-profile protocol."},{"cited_title":"Bondy and U","cited_arxiv_id":null,"evidence_quote":"Supplies Menger's theorem and the graph-connectivity notions underlying the 2-connected and k-connected guarantees."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the maximal monotone relation and network-optimization duality formalism used to define MEIP and to justify the steady-state equations."}],"review_version":1}