{"id":"82aca238-88e8-4531-9064-4b8b320bcd92","arxiv_id":"1908.00976","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper extends local module identification in dynamic networks to correlated process noise by embedding the target module in a MIMO prediction-error setup with graph-based signal selection.","lead":"This paper gives a method for identifying one connection inside a network of interconnected dynamic systems when the noise on different nodes is correlated. It selects which signals to measure and model together so the target connection is estimated consistently and with minimum variance.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Consistency claim hinges on an uncharacterized data-informativity condition involving an unmeasured innovation, so the algorithms alone do not guarantee consistent estimates.","rationale":"The reader's weakest_assumption correctly flags the need for exact Boolean correlation structure, and that is a real limitation. However, I judge the more load-bearing gap to be the data-informativity condition in Theorem 2(b), because it is an uncharacterized condition in the paper's own consistency theorem, it involves the unmeasured innovation ξQ, and the paper explicitly defers its specification to future work (Section X). Even if the correlation structure is known perfectly, the algorithms only enforce the invariance conditions of Theorem 1 and the delay conditions of Theorem 2(c); they do not enforce or characterize informativity. This means the central guarantee of consistent and maximum-likelihood estimation is not actionable in any concrete network unless the user independently establishes Φκ > 0. The reader's rationale already mentions this as a secondary issue, so my concern is complementary rather than contradictory. I maintain the CONDITIONAL verdict because the theoretical framework is coherent and the gap is clearly acknowledged; a conditional acceptance with a requirement to characterize informativity (or provide simulations) is appropriate.","tokens_in":27712,"tokens_out":18699,"duration_ms":185570,"concrete_test":"Apply Algorithm 1 to the Example 4 network with no external excitation (r = 0) and choose transfer-function parameters so that the spectral density Φκ is singular at some frequency (e.g., by introducing a delay-free path that makes one predictor input a static linear combination of another). Run a Monte Carlo prediction-error identification of G12 and check whether the criterion has multiple minima and the estimates are biased. If they are, the missing informativity characterization is material; if not, the gap is purely technical. A complementary analytical test is to derive path-based sufficient conditions for Φκ > 0 for the full-input algorithm and verify them on this network.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is the uncharacterized data-informativity condition in Theorem 2(b), which is central to the consistency claim. Condition (b) requires Φκ(ω) > 0 for κ(t) = [wD(t)', ξQ(t)', wo(t)]', where ξQ is an unmeasured innovation of the transformed noise. The paper explicitly states in Section X that 'A specification of path-based conditions for data-informativity is beyond the scope of this paper.' Consequently, the three selection algorithms in Sections VII–IX guarantee only module invariance (Theorem 1) and, through Proposition 2, the delay conditions; they do not ensure that the experimental setup is informative. Thus the central claim that the target module 'can be consistently estimated' holds only when an unverified and partly unverifiable spectral condition is satisfied. This is not merely a matter of misspecified noise correlations: even with the Boolean correlation structure known exactly, a user cannot check from measurable data whether the selected setup yields uniqueness in the prediction-error criterion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers the identification of a single module G_ji in a dynamic network with known topology, when process noises at different nodes are correlated. The authors propose a local direct method in which the target module is embedded in a MIMO prediction-error setup with suitably chosen predictor inputs w_D and predicted outputs w_Y, allowing correlated noise to be modeled by a full noise spectral density. The main theoretical results are: (i) a module invariance theorem (Theorem 1) giving conditions under which the target module appearing in the transformed identification model equals the original G_ji; (ii) a consistency theorem (Theorem 2) for the MIMO prediction-error estimate, requiring model-set membership, a data-informativity condition on Φ_κ, and delay conditions on paths/loops; and (iii) a maximum-likelihood result (Theorem 3) under Gaussian innovations. Three algorithms are presented for selecting the node signals in the full-input, minimum-input, and user-selection cases, each illustrated on examples. Proofs are collected in an appendix.","tokens_in":27917,"tokens_out":7832,"duration_ms":79215,"significance":"If the stated conditions are satisfied, the paper offers a systematic way to obtain consistent, asymptotically efficient estimates of a local module without identifying the entire network, extending earlier MISO results to the correlated-noise case. The explicit treatment of direct and indirect confounding variables, the block-structured transformation into a representation without confounding variables, and the three signal-selection algorithms are useful and nontrivial contributions to the dynamic-network identification literature. A notable strength is that the paper is transparent about the limitations: Section X explicitly states that the algorithms do not by themselves guarantee consistency and that path-based data-informativity conditions are beyond the scope of the paper. The appendix proofs are detailed, and the paper correctly distinguishes between the invariance guarantee and the conditional consistency result.","major_comments":[{"comment":"The data-informativity condition in Theorem 2(b), Φ_κ(ω) > 0 with κ = [w_D^T, ξ_Q^T, w_o^T]^T, involves the unmeasured innovation ξ_Q, and Section X explicitly states that 'a specification of path-based conditions for data-informativity is beyond the scope of this paper.' Consequently, the three algorithms of Sections VII–IX guarantee only module invariance (Theorem 1) and, through Proposition 2, the delay conditions; they do not ensure that the selected setup is informative. The paper's central claim that the target module can be consistently estimated with maximum-likelihood properties therefore holds only under an additional, unverified spectral condition that cannot be checked from measured data alone. The authors should either provide constructive, checkable conditions for data-informativity (for example, conditions on the external excitation r) or, at minimum, reframe the main contribution as invariance plus conditional consistency and add a detailed discussion of how a user might verify or satisfy this condition in practice.","section":"Section X, Theorem 2(b)"},{"comment":"The maximum-likelihood result is not derived in the paper; the proof is deferred by stating 'Can be shown by following a similar reasoning as in Theorem 1 of [8].' Since the present identification setup differs from [8] in important respects—signals in the set Q act simultaneously as inputs and outputs, and the noise model has the block-diagonal structure (43)—the applicability of the argument in [8] is not immediate. The authors should include a self-contained proof or a precise reduction of the present setup to that of [8, Theorem 1]. Without this, the ML claim is not independently verifiable and the reader cannot judge whether the particular parametrization and the presence of the block-diagonal noise model affect the result.","section":"Theorem 3"}],"minor_comments":[{"comment":"The assumption that the Boolean correlation structure of Φ_v is known a priori is strong and is used explicitly by all three algorithms. The paper should add a brief discussion of how the results degrade if this structure is misspecified, even though data-driven estimation is deferred to future work.","section":"Section II"},{"comment":"The notation 'w_YYF' is unclear and likely a typesetting artifact; it should be written as w_Y ∪ w_F or similar. The same issue appears in the reformulated condition in Proposition 2.","section":"Theorem 2(c), Proposition 2(c)"},{"comment":"There is a typo: 'maximum likehood' should be 'maximum likelihood.'","section":"Example 4"},{"comment":"The statement that the minimum-input algorithm 'adds the smallest number of additional signals to be measured' is plausible but not proven. A formal argument or a counterexample discussion would strengthen the claim.","section":"Section VIII"},{"comment":"The text 'conditon 2b' should be 'condition 2b.'","section":"Proof of Theorem 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper gives a constructive way to build MIMO prediction-error setups for identifying a single module when process noises are correlated across nodes. That is a real extension: earlier local direct methods assumed uncorrelated noise, and indirect or two-stage methods bought consistency at the price of giving up minimum-variance/ML properties. The three node-selection algorithms (full input, minimum input, user selection) are clearly specified, and the module-invariance theorem is proved in the appendix under explicit path-based conditions. I also credit the authors for being honest about what their algorithms do and do not guarantee. Section X states plainly that the algorithms do not guarantee consistency—data-informativity and delay conditions still need to be satisfied. That is a refreshingly direct statement of a real limitation.\n\nThe soft spots, in proportion:\n\n1. The data-informativity condition in Theorem 2(b) is uncharacterized. The regression vector includes an unmeasured innovation term, and the paper says a path-based informativity condition is beyond scope. So in practice a user cannot verify from measured signals whether the chosen setup is informative. This is a genuine gap, though not a fatal one—prediction-error identification routinely relies on similarly abstract informativity assumptions. But given the paper's constructive tone, the gap between the algorithms and a guaranteed consistent estimate is larger than one might initially think.\n\n2. The maximum-likelihood claim in Theorem 3 is deferred to 'a similar reasoning as in Theorem 1 of [8]'. That is probably fine for a specialized audience, but a referee should ask for a sketch of why the existing proof transfers to this setup, especially because the predictor here includes simultaneous input/output signals in Q.\n\n3. The Boolean noise correlation structure is assumed known. The paper mentions future work on estimating it, but as presented, the method requires prior knowledge that may be hard to get.\n\n4. Examples are illustrative; there is no Monte Carlo validation. Minor, given the theoretical nature.\n\nThe module-invariance proof is the load-bearing part and it looks solid. The consistency theorem is standard conditional on informativity. This is a well-written, honest paper that deserves a serious referee. The authors should be asked to tighten the informativity discussion, perhaps by pointing to the cited path-based results or stating what would need to be checked in practice. If you are considering a special issue on dynamic networks, send it out.","headline":"A principled, honest extension of local module identification to correlated noise, with the main gap—data-informativity—left explicitly open rather than solved.","tokens_in":28394,"tokens_out":2132,"would_cite":true,"duration_ms":25036,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For dynamic networks with correlated node noise, a local module can be consistently estimated with maximum-likelihood properties by embedding it in a MIMO prediction-error setup.","keywords":["dynamic networks","local module identification","correlated noise","system identification","predictor input selection","predicted output selection","confounding variables","maximum likelihood"],"falsifier":"Simulate the paper's two-node example with dynamically correlated $v_1$ and $v_2$: a SISO direct-method estimate of $G_{21}$ should be biased, while the MIMO setup with both nodes as outputs should become unbiased and approach the asymptotic variance bound as record length grows. A single network satisfying all conditions of Theorems 1 and 2 in which the MIMO direct estimate stays biased would refute the central claim; likewise, a misspecified Boolean correlation pattern that nevertheless yields unbiased estimates would show the knowability assumption is not necessary.","tokens_in":27552,"feed_emoji":"🕸️","tokens_out":7660,"duration_ms":67310,"temperature":0.7,"pith_summary":"The paper extends local module identification in dynamic networks from the setting of uncorrelated process noises to the setting where noises on different nodes are correlated. It claims that, when the network topology and the Boolean correlation structure of the disturbances are known, a target module can be consistently estimated with asymptotic maximum-likelihood properties by embedding it in a MIMO prediction-error setup rather than the usual MISO setup. The core of the method is the selection of measured node signals as predictor inputs and predicted outputs so that confounding variables, unmeasured disturbances that affect both sides of the estimation problem, are neutralized by the noise model. Three algorithms are given for this selection, and the paper proves target-module invariance, consistency, and maximum-likelihood properties under the relevant conditions. The main caveat is that the entire construction assumes exact knowledge of which noise correlations are nonzero.","feed_headline":"Correlated noise no longer blocks local module estimates","feed_subtitle":"Embedding the target module in a MIMO prediction-error setup restores consistency and minimum-variance identification","key_machinery":"The load-bearing object is the transformed network representation (8), built by first removing unmeasured node signals $w_Z$ (immersion) and then reshaping the noise model so that no confounding variables remain for the estimation problem $w_U \\to w_Y$. A confounding variable is an unmeasured noise component that has paths to both an input and an output of the estimation problem. The node sets are decomposed as $Y = Q \\cup \\{o\\}$, $D = Q \\cup U$, and $U = A \\cup B$, and the main conditions are Condition 1 (parallel path and loop condition), which keeps the target module invariant under immersion, and Condition 2 (no confounding variables for $w_A \\to w_Y$ and $w_A \\to w_B$), which keeps it invariant under the noise reshaping. Theorem 1 combines these into the module invariance result; Theorem 2 adds data-informativity, expressed as $\\Phi_\\kappa(\\omega) > 0$, and delay-in-path conditions for consistency; Theorem 3 gives the maximum-likelihood estimate. The three selection algorithms in Sections VII–IX operationalize these conditions.","core_discovery":"On the paper's own terms, the discovery is that a single module in a dynamic network with known topology can be identified locally, with consistency and asymptotic maximum-likelihood (minimum-variance) properties, even when process noises are correlated across nodes, provided the identification problem is reformulated as a MIMO prediction-error estimation. The target module $G_{ji}$ is embedded in an estimation problem with predictor inputs $w_D$ and predicted outputs $w_Y$, where some signals act as both input and output; the correlation in the noise is absorbed by a full multivariate noise model. The paper proves that if the selected node sets satisfy the parallel path and loop condition, the no-confounding conditions of Theorem 1, data-informativity, and the delay conditions of Theorem 2, then the target module appearing in the transformed equations is exactly the original module and is estimated consistently, with the maximum-likelihood formula of Theorem 3. Three algorithms—full input, minimum input, and user selection—construct such setups from the network topology and the Boolean noise correlation structure.","pith_inferences":["Editorial inference: The paper's conditions are purely graphical once the Boolean correlation structure is known, so the signal-selection step should be automatable at scale; the paper mentions this possibility but does not implement it.","Editorial inference: In practice the correlation structure would often be estimated from data, and the method's guarantees would then inherit the estimation error in that structure; the invariance and consistency results would need to be re-derived under estimated rather than known correlation patterns.","Editorial inference: Because the three algorithms produce different experimental setups, the variance of the final module estimate will generally differ across them; a user could in principle select a setup by comparing the asymptotic variance bounds, a comparison the paper leaves for future work."],"forward_implications":["The target module can be estimated consistently with maximum-likelihood properties even when node noises are correlated, provided the selected node signals satisfy the invariance, informativity, and delay conditions.","Correlated noise no longer forces a choice between consistency (indirect or two-stage methods) and minimum variance: the MIMO direct method delivers both.","Each of the three selection algorithms (full input, minimum input, user selection) yields a valid identification setup under the stated conditions, so users can trade measurement cost against setup complexity.","Only local measurements are needed; the method avoids collecting node signals far from the target module and avoids identifying unnecessary modules."],"supporting_citations":[{"why":"Baseline MISO direct method for local module identification under uncorrelated noise; the setup this paper generalizes.","marker":"[15]"},{"why":"Source of the parallel path and loop condition and the immersion of unmeasured nodes used in Condition 1 and Theorem 1.","marker":"[23]"},{"why":"Two-node correlated-noise example and the reasoning that predicting additional outputs removes bias; starting point of the MIMO approach.","marker":"[25]"},{"why":"Handling of confounding variables in dynamic networks, which the paper extends to correlated noise and MIMO setups.","marker":"[26]"},{"why":"Prediction-error identification for full dynamic networks with noise models; basis for the consistency and maximum-likelihood arguments in Theorems 2 and 3.","marker":"[8]"},{"why":"Classical prediction-error identification framework and regularity conditions underlying the direct method.","marker":"[21]"},{"why":"Preliminary full-input-case version of the MIMO embedding that the paper generalizes.","marker":"[27]"},{"why":"Spectral factorization used to construct the monic, stable, minimum-phase noise representation in Proposition 1.","marker":"[40]"}],"fun_headline_variants":["Correlated noise tamed for local module ID","Local modules identified despite node noise correlation","MIMO reformulation enables local module estimation","Consistent local module ID with correlated process noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the user knows in advance exactly which pairs of node noises are correlated (the zero/nonzero pattern of the noise spectrum); if that pattern is missing or wrong, the selected node sets can violate Condition 2 and the invariance, consistency, and maximum-likelihood guarantees no longer follow.","fun_headline_variants_meta":{"raw":{"variants":["Correlated noise tamed for local module ID","Local modules identified despite node noise correlation","MIMO reformulation enables local module estimation","Consistent local module ID with correlated process noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1551,"prompt_tokens":943,"completion_tokens":608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":552}},"tokens_in":559,"tokens_out":608,"duration_ms":6131,"temperature":1.0,"reasoning_tokens":552,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:31:18.585382+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the paper's two-node example with dynamically correlated $v_1$ and $v_2$: a SISO direct-method estimate of $G_{21}$ should be biased, while the MIMO setup with both nodes as outputs should become unbiased and approach the asymptotic variance bound as record length grows. A single network satisfying all conditions of Theorems 1 and 2 in which the MIMO direct estimate stays biased would refute the central claim; likewise, a misspecified Boolean correlation pattern that nevertheless yields unbiased estimates would show the knowability assumption is not necessary.","supporting_citations":[{"cited_title":"Identification of dynamic models in complex networks with prediction error methods - basic methods for consistent module estimates,","cited_arxiv_id":null,"evidence_quote":"Baseline MISO direct method for local module identification under uncorrelated noise; the setup this paper generalizes."},{"cited_title":"Identification of dynamic models in complex networks with prediction error methods: Predictor input selection,","cited_arxiv_id":null,"evidence_quote":"Source of the parallel path and loop condition and the immersion of unmeasured nodes used in Condition 1 and Theorem 1."},{"cited_title":"From closed-loop identification to dynamic networks: generalization of the direct method,","cited_arxiv_id":null,"evidence_quote":"Two-node correlated-noise example and the reasoning that predicting additional outputs removes bias; starting point of the MIMO approach."},{"cited_title":"Conditions for handling confounding variables in dynamic networks,","cited_arxiv_id":null,"evidence_quote":"Handling of confounding variables in dynamic networks, which the paper extends to correlated noise and MIMO setups."},{"cited_title":"Prediction error identification of linear dynamic networks with rank-reduced noise,","cited_arxiv_id":null,"evidence_quote":"Prediction-error identification for full dynamic networks with noise models; basis for the consistency and maximum-likelihood arguments in Theorems 2 and 3."},{"cited_title":"Ljung, System Identification: Theory for the User","cited_arxiv_id":null,"evidence_quote":"Classical prediction-error identification framework and regularity conditions underlying the direct method."},{"cited_title":"Local module identification in dynamic networks with correlated noise: the full input case,","cited_arxiv_id":null,"evidence_quote":"Preliminary full-input-case version of the MIMO embedding that the paper generalizes."},{"cited_title":"On the factorization of rational matrices,","cited_arxiv_id":null,"evidence_quote":"Spectral factorization used to construct the monic, stable, minimum-phase noise representation in Proposition 1."}],"review_version":1}