{"id":"2cb40312-7f30-4257-bd98-452f2983e17d","arxiv_id":"1908.08245","paper_version":6,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Consensus-plus-innovations decentralized estimation converges in mean square and almost surely under stochastic spatio-temporal persistence of excitation, even with correlated random graphs, observation matrices, and communication delays.","lead":"This paper proves convergence conditions for a decentralized estimation algorithm where sensors share estimates over randomly changing networks with random delays and random measurement matrices. It shows that mean square and almost sure convergence hold under a persistence of excitation condition, even when randomness is correlated across time and nodes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pathwise spatio-temporal PE condition is assumed for general environments; only the Markovian corollary provides a verifiable sufficient case, limiting the claimed generality.","rationale":"The reader's weakest-assumption analysis is accurate: the pathwise stochastic spatio-temporal PE condition is the load-bearing premise. I agree that this is the point where the paper's generality is limited. However, I do not find an internal inconsistency in the proofs. The delayed-system equivalence asserted in Section V is correct: substituting g(k-q)=e(k-q+1)-F(k-q)e(k-q) into (15) and using (16) reproduces (10) by induction. The Markovian Corollary IV.1 does provide a nontrivial, verifiable sufficient condition (stationary balanced spanning tree plus joint observability) under which (c.1) holds for all sample paths. The numerical slips in Section VI (f≈2.45e-4 vs 5e-4, and the 0.01 factor vs ≈0.00823) affect only the illustrative example, not the theorems. The pathwise PE condition remains a real limitation: it is not derived from A1–A2, and for non-Markovian stationary processes with occasional poorly excited blocks it can fail with positive probability. This supports the CONDITIONAL verdict rather than ACCEPT, but does not warrant REJECT because the central conditional claim and its Markovian specialization appear sound.","tokens_in":40501,"tokens_out":29683,"duration_ms":279343,"concrete_test":"Construct a stationary ergodic two-node network satisfying A1–A2 with conditionally balanced mean graph and H_i(k)=s_i(k)I, where s_i(k) are i.i.d. Bernoulli(0.5); the stationary model is spatio-temporally jointly observable. With h=2, there is positive probability that a block has all s_i(k)=0, so the conditional expectation in (b.1) has λ_min=0 on those sample paths. Compute the left side of (b.1) over all m; if it is not ≥ c(m) a.s., the pathwise PE condition is not implied by the environment, confirming that the general theorems exclude such stationary processes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorems IV.1, V.1 and Corollaries V.1–V.3 all hinge on the stochastic spatio-temporal persistence of excitation condition, e.g. (b.1): λ_min of Σ_{k=mh}^{(m+1)h-1} (b(k)E[L̂_G(k)|F(mh-1)]⊗I + a(k)E[H(k)^T H(k)|F(mh-1)]) ≥ c(m) a.s. with Σ c(m)=∞. This is a pathwise requirement on every history; the standing assumptions A1–A2 and C1 do not imply it. The paper supplies one verifiable sufficient case (Corollary IV.1) via uniform ergodicity of a Markov chain, but for general non-Markovian environments the condition is simply assumed, and it can fail for stationary ergodic processes that occasionally have poorly excited blocks (e.g., long intervals with H=0 and disconnected graphs) with positive probability. The numerical example does not close this gap: it plots Λ^2_m for m≤500, not for all m, and contains apparent arithmetic slips (f_{1,1,β_H,4,4}(0.01)≈2.45e-4, not 5e-4; the weighted delay factor is ≈0.00823, not 0.01). Thus the central claim's scope is narrower than the abstract suggests: the general theorems are conditional on a condition that is not derived and may be hard to certify.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies decentralized cooperative online estimation of a fixed unknown parameter by a network of nodes, where each node takes linear measurements with random, possibly correlated observation matrices, and communicates over random directed graphs with random bounded time-varying delays. The algorithm combines a consensus term using delayed neighbor estimates with a local innovation term. The authors derive conditions under which all node estimates converge to the true parameter in mean square and almost surely (delay-free case) or in mean square (delayed case). The key tool is a stochastic spatio-temporal persistence-of-excitation condition on block sums of conditional expectations of graph Laplacians and observation matrices. For Markovian switching graphs and observation matrices, this condition is shown to follow from balancedness with a spanning tree of the stationary graph plus spatio-temporal joint observability. For the delayed case, the error system is transformed into an augmented system involving an auxiliary filter, and convergence conditions are expressed through conditional expectations of delay matrices, graph matrices, and observation matrices.","tokens_in":40844,"tokens_out":4780,"duration_ms":48822,"significance":"If the theorems are correct, the paper makes a useful contribution: it relaxes the usual i.i.d. or Markovian independence assumptions on graphs and observation matrices, allows unbalanced and time-varying mean graphs, and treats nonuniform random delays through a novel delay-matrix formalism. The proof strategy via binomial expansion of random matrix products is sophisticated and appears to be largely self-contained. The Markovian corollary provides a genuinely verifiable sufficient condition and clarifies that neither local observability nor instantaneous global observability is necessary. However, the advertised generality is tempered by the fact that the central stochastic spatio-temporal PE condition is assumed, not certified, for general non-Markovian environments, and the delayed-case analysis rests on an equivalence that is asserted but not proved.","major_comments":[{"comment":"The equivalence between the delayed error system (10) and the augmented system (15)-(16) is asserted with 'It can be verified' but no proof is given. This equivalence is load-bearing: the proof of Theorem V.1 and all delayed-case corollaries rely on r(k) being equal to e(k). I request a complete derivation or a precise reference that establishes this equivalence, including the construction of F(k) and C_q(k) for all k and the treatment of the initial conditions r(k)=e(k) for -d<=k<=-1.","section":"Section V, Eqs. (15)-(16)"},{"comment":"The stochastic spatio-temporal persistence of excitation condition is a pathwise requirement on every history and is not implied by Assumptions A1-A2 and Conditions C1. The paper supplies a verifiable sufficient case only in the Markovian setting (Corollary IV.1). For general non-Markovian environments, the main theorems are conditional on a premise that is not derived and may fail for stationary ergodic processes that occasionally have long blocks with H(k)=0 and disconnected graphs. The authors should either provide verifiable sufficient conditions for (b.1)/(c.1) in the non-Markovian case or explicitly state that the general theorems apply only when this condition can be certified, and adjust the abstract and introduction accordingly.","section":"Section IV, conditions (b.1) and (c.1)"},{"comment":"The numerical claim lambda_min = 0.5821 in the two-node example is incorrect. For a12=1, a21=0.3, H1=0, H2=1, the matrix Lhat + H^T H equals [[1, -0.65], [-0.65, 1.3]], whose minimum eigenvalue is approximately 0.4829, not 0.5821. The illustrative claim that the condition holds with c(m)=0.5821/(m+1) should be corrected and re-verified.","section":"Remark 5"},{"comment":"The numerical verification of Corollary V.1 contains arithmetic errors. With C1=1, beta_a=1, beta_H=4.07, N=4, d=4 and psi_2=0.01, the function f_{C1,beta_a,beta_H,N,d}(0.01) evaluates to about 2.45e-4, not 5e-4. Similarly, the weighted delay factor sum 0.5 * sum_{q=0}^4 p_q * ((1+psi_2)^q - 1)/(2 - (1+psi_2)^q) is approximately 0.00823, not 0.01. These values are used to set b(0) and to verify inequality (21); the example should be recomputed with correct values and the plots re-examined.","section":"Section VI"}],"minor_comments":[{"comment":"There are several typographical errors: 'applicaitons', 'Conditon', 'SCDA system' (should be SCADA), and 'sigma-filed' for 'sigma-field'. A careful proofread is recommended.","section":"Throughout"},{"comment":"The caption of Figure 5 uses 'Lambda^2_m - 0.01(b(2m-1)+b(2m))' while the text and the y-axis label use 'Lambda^2_m - 0.01 * sum_{k=2m}^{2m+1} b(k)'. These should be made consistent.","section":"Figure 5 caption"},{"comment":"The notation p_{ji,q}(k) is used for the probability that the delay equals q, but the delay variable is written as lambda_{ji}(k). Clarifying the indexing (link from j to i) in a sentence would help the reader avoid confusion with the q-th delay matrix.","section":"Section II.B, Eq. (3)"},{"comment":"The definition of the set Gamma_1 requires E[A_G(k)|F(k-1)] to be nonnegative and its associated random graph balanced almost surely. It would be helpful to state explicitly that this is a condition on the conditional mean graph, not on the sample graph, in the main text preceding Theorem IV.2.","section":"Section IV, Theorem IV.2"}],"recommendation":"major_revision","confidential_remarks":"The paper relies on two self-cited references ([36] for a lemma and [41] for a convergence tool). This is not fatal, but the authors should ensure the borrowed lemmas are fully stated and that the citations are appropriate. The main missing ingredient is the proof of the delayed-system equivalence (15)-(16); without it the delayed-case results are not fully established. The numerical errors in Remark 5 and Section VI should be corrected, but they are not signs of a fundamentally flawed central proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a real theoretical advance: it proves mean-square and almost-sure convergence for consensus-plus-innovations estimation under a block-wise stochastic spatio-temporal persistence-of-excitation condition, without requiring observation matrices and graphs to be i.i.d., Markovian, or mutually independent, and it introduces delay matrices to handle random time-varying delays. Second, the main theorems are conditional on that PE condition, which is assumed for general environments and only verified for the Markovian case; and the delayed-case equivalence is asserted without proof. I recommend peer review, not desk reject, but with requested revisions.\n\nWhat is actually new: the relaxation of independence assumptions; the delay-matrix formulation; the use of binomial expansion of random matrix products to turn convergence analysis into analysis of expectations of matrix products; and Corollary IV.1, which gives a nice sufficient condition — balanced stationary graph with spanning tree plus spatio-temporal joint observability, without requiring local or instantaneous global observability. The appendix proof is long but the structure is coherent: block conditional expectations, product decay from a non-summable c(m), and martingale arguments. The self-citations to [36] and [41] are used as lemmas; that is not a problem.\n\nSoft spots, in order of seriousness. (1) The stochastic spatio-temporal PE condition is not derived for general environments. It is a pathwise lower bound on each length-h block of conditional expectations. For a stationary ergodic process with occasionally poorly excited blocks, it can fail. The theorems honestly state it as a condition, so this is not circular, but the claimed generality — random graphs and observation matrices without special statistical properties — is narrower than it appears: only the Markovian corollary gives a verifiable sufficient case. The paper should say this explicitly. (2) The transformation from the delayed error system (10) to the augmented system (15)-(16) is asserted as 'it can be verified' but no proof is given. All delayed-case results rest on that equivalence; either prove it or give a precise citation. (3) Numerical checks: I believe the λ_min in Remark 5 is about 0.483, not 0.5821, and the values f = 0.0005 and the 0.01 delay factor in Section VI are not consistent with their own formulas (I get roughly 0.000245 and 0.00407). If my arithmetic is right, the figures need to be redone. Minor, but it undercuts the claim that the conditions are verified.\n\nWho this is for: people working on distributed estimation, consensus algorithms, or adaptive networked control who need convergence guarantees under correlated random topologies and delays. They get a serious, if conditional, convergence theory.\n\nMy recommendation: send to peer review. The core mathematical structure is sound and the result is a genuine extension. Request fixes for the numerical slips, a proof or reference for the equivalence, and a clearer statement that the general PE condition is assumed, with the Markovian case as the main verifiable path.","headline":"A genuine theoretical extension of consensus-plus-innovations convergence under a block-wise stochastic persistence-of-excitation condition, with random delays included; the main theorems are conditional and the delayed-case equivalence and numerical verification need tightening, but this deserves peer review.","tokens_in":41277,"tokens_out":5018,"would_cite":true,"duration_ms":51639,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E10","93E15","93A14","60G42"],"pacs":[],"model":"deepseek-v4-flash","headline":"A network of estimators can converge on a shared unknown parameter even when observation matrices, communication graphs, and communication delays are all random, time-varying, and statistically dependent, provided a stochastic…","keywords":["decentralized online estimation","cooperative estimation","random communication graph","random observation matrix","random time delay","persistence of excitation","consensus plus innovations","mean square convergence"],"falsifier":"Run the delay-free algorithm on a Markovian switching chain whose stationary graph is balanced with a spanning tree and whose stationary Gram matrix satisfies (14), set gains like $a(k)=b(k)=1/(k+1)^{0.6}$, and compute the averaged relative error after a long horizon; Corollary IV.1 predicts the error approaches zero, so a nonzero limiting error would refute the central claim. Equivalently, construct a process satisfying all assumptions except (b.1) by making nodes isolated and setting all observation matrices to zero: the PE condition fails and the estimates remain frozen, confirming that the condition is indispensable.","tokens_in":1711,"feed_emoji":"📡","tokens_out":2005,"duration_ms":75954,"temperature":0.7,"pith_summary":"This paper proves that a group of networked estimators can cooperatively track an unknown parameter even when each node's measurements, the communication links, and the delivery delays are all randomly time-varying and can even be statistically dependent. The central result is a condition called stochastic spatio-temporal persistence of excitation: over each fixed-length block, the pooled conditional expectation of graph Laplacians and observation matrices must have a minimum eigenvalue bounded below by a non-summable sequence. When that condition holds, properly chosen vanishing gains drive all nodes' estimates to the true parameter in mean square and, in the delay-free case, almost surely. This matters because existing analyses typically required i.i.d. graphs, balanced mean graphs, or independence between sensing and communication. The paper's condition allows unbalanced, stepwise-disconnected graphs and correlated random environments, and it handles random nonuniform delays by introducing delay matrices and reducing the delayed system to an equivalent delay-free one.","feed_headline":"Network estimators converge despite random links, sensors, delays","feed_subtitle":"A block-wise persistence-of-excitation condition lets consensus-plus-innovations updates reach the truth.","key_machinery":"The load-bearing object is the stochastic spatio-temporal persistence-of-excitation condition: for a fixed block length $h$, the minimum eigenvalue of the sum of conditional expectations of $b(k)$ times the symmetrized graph Laplacian and $a(k)$ times the Gram observation matrix over each block is required to be at least $c(m)$, with $\\sum_m c(m)=\\infty$. This condition controls the binomial expansion of the random matrix product $\\Phi_P(k,0)$, forcing the expected error covariance to decay through products of factors $(1-c(m)+\\text{small remainder})$. For the delayed case, the companion machinery is the delay matrix $I(k,q)$ and the equivalent delay-free system $r(k+1)=F(k)r(k)+g(k)$, whose invertibility is guaranteed by a small initial consensus gain (Lemma V.1); the same blockwise persistence condition then controls the products $\\Phi_F(k,0)$.","core_discovery":"The paper's central claim is that consensus-plus-innovations estimation over a random network converges whenever the random environment is persistently exciting in a joint spatio-temporal sense. Concretely, for a block length $h$, the key condition requires that the minimum eigenvalue of $\\sum_{k=mh}^{(m+1)h-1}(b(k)E[\\hat{L}_G(k)|\\mathcal{F}(mh-1)]\\otimes I_n + a(k)E[H^T(k)H(k)|\\mathcal{F}(mh-1)])$ be at least $c(m)$ almost surely, with $\\sum_m c(m)=\\infty$. Under this condition, Theorem IV.1 gives mean-square and almost-sure convergence for the delay-free algorithm, and Theorem IV.2 gives the same under a constant lower bound when the digraphs are conditionally balanced. For Markovian switching environments, Corollary IV.1 derives the condition from a stationary balanced graph with a spanning tree plus spatio-temporal joint observability, so neither local observability nor instantaneous global observability is needed. With delays, the paper models the random lags with delay matrices $I(k,q)$ and, through binomial expansion of random matrix products, converts convergence of the delayed error system into convergence of expectations of products of transformed matrices $F(k)$; Corollary V.3 states that under the same persistence-of-excitation condition, any bounded delay can be tolerated by choosing gains sufficiently small when the digraphs are conditionally balanced.","pith_inferences":["The condition could be monitored online by computing running blockwise minimum eigenvalues; when it fails, a network could adaptively increase block length or switch to more informative measurements to restore convergence.","The dependence of the admissible delay on graph intensity suggests a protocol-level design rule: if delay distributions are heavy, links must be used more frequently or observation matrices must be richer.","A natural extension is to replace the almost-sure blockwise lower bound with a high-probability version; such a relaxation might give practical stopping rules while preserving convergence in probability rather than almost surely.","The blockwise condition could also be used to compare sensor scheduling policies, since any policy that keeps the blockwise minimum eigenvalue non-summable is sufficient for convergence."],"forward_implications":["Observation matrices, communication graphs, and delays need not be mutually independent or spatio-temporally independent for convergence; all three may be correlated.","Balanced mean graphs are not necessary: the persistence condition can hold even when sample paths and mean graphs are unbalanced.","For Markovian switching environments, convergence follows when the stationary graph is balanced with a spanning tree and the measurement model is spatio-temporally jointly observable, so neither local observability nor instantaneous global observability is required.","With conditionally balanced digraphs, any bounded random delay can be accommodated by choosing sufficiently small algorithm gains, and the maximum tolerable delay bound is expressed in terms of graph weights, delay probabilities, and gains.","In the limit of completely isolated nodes, the condition degenerates to independent stochastic persistence-of-excitation conditions for centralized estimation."],"supporting_citations":[{"why":"Supplies the baseline consensus-plus-innovations algorithm whose error dynamics this paper extends to random graphs, observation matrices, and delays.","marker":"[23]"},{"why":"Introduces the stochastic persistence-of-excitation condition for centralized estimation that the paper generalizes to the decentralized spatio-temporal setting.","marker":"[38]"},{"why":"Treats decentralized estimation with random graphs and observation matrices under i.i.d. assumptions, which the paper relaxes.","marker":"[22]"},{"why":"Provides the Markovian switching graph analysis that Corollary IV.1 extends to spatio-temporally jointly observable models.","marker":"[24]"},{"why":"Supplies the transformation of delayed consensus systems into equivalent delay-free systems, adapted here to estimation with random delays.","marker":"[32]"},{"why":"Supports the binomial-expansion and state-augmentation technique used to handle time-varying delays in the error dynamics.","marker":"[33]"},{"why":"Provides the almost-supermartingale convergence theorem used to establish almost-sure convergence of the estimates.","marker":"[42]"}],"fun_headline_variants":["Random links, sensors, and delays don't break convergence if excitation persists","Consensus-plus-innovations converges under joint spatio-temporal excitation","Without local observability, a joint condition still guarantees network learning","Bounded delays are tolerable in distributed estimation under one condition","Even random graphs and measurements yield convergence with proper gains"],"cache_read_input_tokens":43520,"weakest_assumption_plain":"The load-bearing premise is the stochastic spatio-temporal persistence-of-excitation condition itself: over every block the pooled conditional expectations of graph Laplacians and observation matrices must have a minimum eigenvalue bounded below by a non-summable sequence, on every sample path, and the paper's other assumptions do not guarantee this.","fun_headline_variants_meta":{"raw":{"variants":["Random links, sensors, and delays don't break convergence if excitation persists","Consensus-plus-innovations converges under joint spatio-temporal excitation","Without local observability, a joint condition still guarantees network learning","Bounded delays are tolerable in distributed estimation under one condition","Even random graphs and measurements yield convergence with proper gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3513,"prompt_tokens":1085,"completion_tokens":2428,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":2342}},"tokens_in":701,"tokens_out":2428,"duration_ms":19663,"temperature":1.0,"reasoning_tokens":2342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:44:47.471199+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the delay-free algorithm on a Markovian switching chain whose stationary graph is balanced with a spanning tree and whose stationary Gram matrix satisfies (14), set gains like $a(k)=b(k)=1/(k+1)^{0.6}$, and compute the averaged relative error after a long horizon; Corollary IV.1 predicts the error approaches zero, so a nonzero limiting error would refute the central claim. Equivalently, construct a process satisfying all assumptions except (b.1) by making nodes isolated and setting all observation matrices to zero: the PE condition fails and the estimates remain frozen, confirming that the condition is indispensable.","supporting_citations":[{"cited_title":"Consensus+innovations distributed inference over networks: Cooperation and sensing in networked systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline consensus-plus-innovations algorithm whose error dynamics this paper extends to random graphs, observation matrices, and delays."},{"cited_title":"Estimating time-varying parameters by the Kalman-filter based algorithm,","cited_arxiv_id":null,"evidence_quote":"Introduces the stochastic persistence-of-excitation condition for centralized estimation that the paper generalizes to the decentralized spatio-temporal setting."},{"cited_title":"Distributed parameter estimation in sensor networks: Nonlinear observation models and imperfect communication,","cited_arxiv_id":null,"evidence_quote":"Treats decentralized estimation with random graphs and observation matrices under i.i.d. assumptions, which the paper relaxes."},{"cited_title":"Distributed parameter estimation over unreliable networks with markovian switching topologies,","cited_arxiv_id":null,"evidence_quote":"Provides the Markovian switching graph analysis that Corollary IV.1 extends to spatio-temporally jointly observable models."},{"cited_title":"Distributed consensus for multiagent systems with communication delays and limited data rate,","cited_arxiv_id":null,"evidence_quote":"Supplies the transformation of delayed consensus systems into equivalent delay-free systems, adapted here to estimation with random delays."},{"cited_title":"Distributed consensus for multi-agent systems with delays and noises in transmission channels,","cited_arxiv_id":null,"evidence_quote":"Supports the binomial-expansion and state-augmentation technique used to handle time-varying delays in the error dynamics."},{"cited_title":"A convergence theorem for nonnegative almost supermartingales and some applications,","cited_arxiv_id":null,"evidence_quote":"Provides the almost-supermartingale convergence theorem used to establish almost-sure convergence of the estimates."}],"review_version":1}