{"id":"ef50f090-a050-4718-b91c-95bcc7c6447d","arxiv_id":"2509.14047","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Data-driven decentralized control of interconnected discrete-time LTI systems is achieved by synthesizing local dissipative controllers and certifying global stability with LMIs from local data and noise bounds.","lead":"The paper gives a data-driven method to design local state-feedback controllers that stabilize an interconnected system using only local data, without knowing the system model or the coupling. It turns local dissipativity and global stability conditions into linear matrix inequalities and validates the approach on simulated DC microgrids.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Inertia condition (17) is not enforced by the LMI relaxation; Algorithm 1's stability certificate relies on a post-hoc check that can fail.","rationale":"The reader's weakest_assumption field points to the noise bounds in Assumptions 1 and 2. That is a legitimate premise, but it is an explicit, standard modeling assumption: every data-driven robust control result is conditional on the uncertainty description. The more load-bearing concern, in my view, is the internal gap between the formal statement of Algorithm 1 (which requires the nonconvex inertia condition (17)) and its implementable LMI relaxation (which omits it and checks afterwards). This directly affects whether the algorithm as practically used actually delivers the advertised stability guarantee. The reader's rationale does mention this as a 'principal weakness', but it is not the field identified as the weakest assumption, hence 'partial' agreement. The paper's theoretical results (Theorem 1 and Theorem 2) appear mathematically sound under their stated assumptions; the S-lemma and dual QMI steps are standard, and the numerical experiments are encouraging. The weakness is in the algorithm's implementability and the guarantee attached to its output. This does not overturn the paper's contribution, but it justifies a conditional verdict rather than unconditional acceptance. I therefore recommend keeping the reader's CONDITIONAL verdict unchanged.","tokens_in":18814,"tokens_out":21031,"duration_ms":215084,"concrete_test":"Run the SDP relaxation (20),(27b),(33) on the 100 random microgrid instances from Section VI, and record the inertia of Q_i = [H_i G_i^T; G_i F_i] for every returned solution. If any instance yields In(Q_i) != (p_i,0,p_i), then the implemented Algorithm 1 violated its own formal constraint (17). For such an instance, test whether the closed-loop system with K_i = L_i P_i^{-1} is actually stable: if it is unstable, the concern lands; if it is stable, then the dual LMI may still be sufficient in practice, which would weaken the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Algorithm 1 computes local gains K_i that stabilize the nominal global system (14). The proof of Theorem 1 uses Lemma 1, which requires the supply-rate matrix S_i = [H_i G_i^T; G_i F_i]^{-1} to have inertia (p_i,0,p_i). This is exactly condition (17). However, Algorithm 1's step 3 formulates the search as LMIs (20),(27b),(33) and only mentions (17) as an additional constraint; the text admits that step 3 is an LMI only when (17) is omitted. The numerical implementation solves the LMIs without (17) and checks it afterwards. If the SDP returns (F_i,G_i,H_i) violating (17), then the equivalence in Lemma 1 between the primal dissipativity LMI (4) and the dual form (23) no longer holds. Consequently, the certified dissipativity from (20) is not guaranteed to imply actual dissipativity of the closed-loop subsystem, and the chain leading to global stability via Proposition 4/Theorem 2 is broken. The paper provides no convex characterization of (17), so the algorithm as stated is not a pure LMI method; its success depends on a non-robust post-hoc verification. This is an internal gap, distinct from the standard assumption on noise bounds (Assumptions 1 and 2), which are explicit and can be checked against the application at hand.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven, decentralized state-feedback synthesis method for interconnected discrete-time LTI systems. The authors first derive a local LMI condition (Theorem 1) that, for a given data set satisfying a QMI noise bound, yields a gain K_i rendering the nominal closed-loop subsystem dissipative with respect to a supply rate parameterized by (F_i,G_i,H_i). They then use a dual QMI argument (Theorem 2) to translate an interconnection stability condition from [15] into a decentralized data-driven LMI using measurements of v_i and \\tilde{y}_i. By treating (F_i,G_i,H_i) as decision variables, the two LMIs are combined into Algorithm 1; a specialized diffusive-coupling variant (Algorithm 2) replaces the interconnection LMI with a condition based on the maximum weighted degree consistent with the data. Numerical experiments on a 50-DGU microgrid, including comparisons with a centralized structured data-driven controller, are reported.","tokens_in":19034,"tokens_out":14588,"duration_ms":137556,"significance":"If the claims hold, the paper is a clean contribution to direct data-driven control of networked systems: it is fully decentralized, avoids system identification, handles process/measurement/interconnection noise via QMIs, and provides a unified S-lemma/dissipativity framework. The benchmarks for data length and computation time are meaningful. However, the load-bearing gap in the treatment of the inertia condition (17) tempers the significance until it is resolved.","major_comments":[{"comment":"The inertia condition (17) is essential for the equivalence in Lemma 1 between the primal dissipativity LMI (4) and the dual form (22) used in the proof of Theorem 1. However, Algorithm 1 lists (17) as a constraint while the immediately following text admits that step 3 is an LMI 'when we omit the inertia condition (17)'. Section VI confirms that the implementation solves the LMIs without (17) and checks it a posteriori. Since (17) is not convex, the relaxed LMI can return (F_i,G_i,H_i) for which (22) is not equivalent to dissipativity; the subsequent stability certificate via Proposition 4/Theorem 2 then has no basis. The paper provides no convex characterization of (17), nor does it report how often the post-hoc check succeeds in the experiments. As stated, Algorithm 1 is not an LMI algorithm, and the reported 'feasible instances' in Table II may overstate the number of systems for whi","section":"Algorithm 1 (Section IV-C), Theorem 1 (Section IV-A), and Section VI"},{"comment":"Theorem 1's 'if and only if' uses the converse of the matrix S-lemma (Proposition 1), which requires J_i to have at least one positive eigenvalue. This assumption is stated in the theorem but is not part of the constraints in Algorithm 1 and is not checked in the numerical implementation. The sufficiency direction holds without the converse, so the algorithm's stability guarantee is not destroyed, but the theorem overclaims an equivalence that the algorithm does not actually verify. The authors should either weaken the statement to a sufficient condition or add the positive-eigenvalue check to the algorithm and report its success rate.","section":"Theorem 1 (Section IV-A) and Algorithm 1"}],"minor_comments":[{"comment":"The abstract claims that 'both conditions take the form of linear matrix inequalities', but Algorithm 1 includes the non-convex inertia condition (17). The sentence 'Step 3 is an LMI condition when we omit the inertia condition (17)' should be reconciled with the abstract and the title's framing.","section":"Abstract and Section IV-C"},{"comment":"The computation of d_max_i via (39) assumes that Theta_i is invertible. The paragraph on the invertibility assumption explains that a small perturbation of Psi_i restores it, but the numerical section does not state whether Theta_i was actually invertible in all generated instances, nor whether the perturbation was needed.","section":"Section V-B and Algorithm 2"},{"comment":"The authors report 'feasible instances' for Algorithms 1 and 2 in Table II, but they also state that the inertia condition is checked after solving the LMIs. To avoid misleading claims, the table should report how many of those feasible instances also passed the inertia check, or clarify that the reported numbers refer only to LMI feasibility.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the paper is generally well-written, but the inertia condition gap is a serious technical issue that must be addressed before publication. The authors should either prove that solutions of the relaxed LMIs automatically satisfy (17) under the stated assumptions, or modify the algorithm to enforce a convex sufficient condition and update the numerical results accordingly. The 'if and only if' statements in Theorems 1 and 2 also need to be aligned with the assumptions actually used in the implementation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — here's my take on arXiv:2509.14047.\n\nThe paper does something genuinely useful: it gives a data-driven, fully decentralized way to stabilize an interconnected LTI system when both local dynamics and the interconnection matrix are unknown. The key trick is to treat the supply-rate matrices (F_i, G_i, H_i) as decision variables, so the local dissipativity LMI and the decentralized stability LMI can be solved jointly. That is a real extension of the van Waarde et al. S-lemma machinery and of Martinelli et al.'s discrete-time dissipativity interconnection condition. The diffusive-coupling variant, with a data-driven upper bound on the weighted degree, is neat.\n\nThe math is generally sound. Theorems 1 and 2 are standard S-lemma/dual-QMI applications, and the proofs check out. The microgrid experiments are careful: 100 random systems, varying noise/data/topology, and a comparison against a centralized structured controller. The computational savings are real, though the baseline comparison is predictable because it requires global data.\n\nThe soft spot is the inertia condition (17). Lemma 1 requires the supply-rate matrix to have inertia (p_i,0,p_i). In Algorithm 1, that condition is listed as a constraint, but the LMI relaxation omits it; the text acknowledges this, and the implementation solves without it and checks afterwards. So the guarantee is conditional: if the SDP returns matrices violating (17), the equivalence between the primal dissipativity LMI and the dual form breaks, and the chain to global stability is not rigorous. This isn't fatal — in the experiments it held every time — but it is an internal gap that should be fixed, either with a convex relaxation that enforces the inertia or by explicitly stating it as a verification step and analyzing failure rates.\n\nThe other assumption — the noise QMI bounds Phi_i and Psi_i — is standard for this literature and not a hidden flaw. It is user-specified, but that's explicit.\n\nWorth refereeing. I would accept it with a request to address the inertia-condition handling. The paper is honest, the derivations are reproducible, and the code is available. A serious referee can push on the inertia issue and on the conservatism of the decentralized condition.\n\nIf I worked in data-driven networked control, I'd cite it.","headline":"Solid data-driven decentralized control pipeline with a real but fixable gap in how Algorithm 1 handles the inertia condition.","tokens_in":19617,"tokens_out":1804,"would_cite":true,"duration_ms":20792,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93A14","93B52","93C55","93D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that local noisy data alone, plus neighbor output measurements, are enough to synthesize decentralized stabilizing controllers for an unknown interconnected system.","keywords":["decentralized control","data-driven control","dissipativity theory","interconnected systems","linear matrix inequalities","state-feedback control","discrete-time LTI systems","microgrids"],"falsifier":"Generate random systems whose noise stays inside the assumed quadratic bounds, solve the LMIs (20), (27b), (33), and simulate the noiseless closed loop with the true system and true interconnection. If any feasible controller yields a diverging trajectory, the claimed sufficiency chain is false; because the paper proves sufficiency, a single such counterexample settles the question.","tokens_in":18622,"feed_emoji":"⚡","tokens_out":7167,"duration_ms":70786,"temperature":0.7,"pith_summary":"The paper tries to show that stabilizing controllers for an interconnected system can be designed from local data alone, without any model of the subsystems or of the interconnections. It gives a two-step pipeline: first, a data-driven LMI renders each closed-loop subsystem dissipative with respect to a chosen supply rate; second, a data-driven decentralized LMI checks that these supply rates are compatible with the unknown symmetric interconnection. Because the supply-rate matrices appear linearly in both LMIs, they can be treated as decision variables, so local dissipativity design and global stability certification are solved jointly. The resulting local state-feedback gains stabilize the nominal global closed-loop system for every system and every interconnection consistent with the measured data. A special case exploiting diffusive coupling computes a data-consistent upper bound on the coupling strength and uses it in a simpler stability condition.","feed_headline":"Local data alone stabilize an unknown interconnected network","feed_subtitle":"A per-subsystem optimization certifies stability for the whole unknown grid, from noisy local measurements alone.","key_machinery":"The machinery is the matrix S-lemma together with the dual quadratic matrix inequality. The data-consistent set of systems Sigma_i is described by a QMI built from local data and the noise bound Phi_i; the S-lemma turns 'dissipative for all systems in Sigma_i' into the single LMI (20) with multiplier alpha_i. Similarly, the interconnection data, through the dual QMI, turn 'stability inequality holds for all consistent interconnection matrices' into LMI (33) with multiplier tau_i. The supply-rate matrices (F_i, G_i, H_i), which enter linearly, act as the coupling variables, while the storage function P_i and the gain-related matrix L_i are the other decision variables. For diffusive coupling,","core_discovery":"The central claim is that, under quadratic bounds on process noise, measurement noise, and interconnection noise, the feasibility of the data-driven LMIs (20), (27b), and (33), with the supply-rate matrices (F_i, G_i, H_i) as decision variables, certifies that each closed-loop subsystem is dissipative with a supply rate that, together with the dual QMI of the interconnection data, guarantees asymptotic stability of the nominal global closed-loop system. Concretely, the computed gain K_i = L_i P_i^{-1} and the storage function V_i(x_i) = x_i^T P_i^{-1} x_i witness dissipativity, and the stability inequality holds for every interconnection matrix consistent with the neighbor data. This turns r","pith_inferences":["If the same pipeline extends to directed interconnections with a suitable local dissipativity condition, the data-driven decomposition could apply to non-symmetric networks; the paper only treats symmetric coupling M = M^T.","Treating F_i, G_i, H_i as free variables suggests a natural way to tune performance: add objective terms on these matrices to shape the supply rate, for instance toward passivity or L2 gain, without changing the algorithm's structure. The paper leaves this open.","The feasibility drop of Algorithm 2 as noise or coupling density increases hints that the diffusive-coupling shortcut is most useful in lightly loaded, low-noise regimes; the authors conjecture this but do not quantify the conservatism.","Because the local conditions are independent LMIs, the approach could run online in a plug-and-play fashion when a new subsystem joins, provided its data and neighbor set are available; this is not tested in the paper."],"forward_implications":["Controllers can be synthesized entirely from local data and neighbor output measurements; no global model, no identification step, and no centralized optimization are needed.","The closed-loop guarantee is robust: it holds for all systems in the data-consistent set Sigma_i and all interconnections consistent with the data, not just the observed trajectory.","Since the supply-rate matrices are decision variables, the design automatically finds a dissipativity certificate compatible with the interconnection, making the local and global conditions jointly feasible rather than iterated.","Because the nominal system is asymptotically stable, the actual noisy system is input-to-state stable with respect to process, measurement, and interconnection noise, so bounded noise leads to bounded state deviations.","For networks coupled by diffusion, the quadratic program (39) yields the tightest data-consistent upper bound on the weighted degree, which is then plugged into the stability condition to design the local gains."],"fun_headline_variants":["Local data alone certify decentralized stability for unknown grids","Data-driven dissipativity stabilizes interconnected systems locally","Stabilize a network without a model: local data suffices","Per-subsystem data yields global stability in interconnected systems","From local measurements to global stability: a data-driven route"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The guarantee rests on knowing quadratic bounds for the unmeasured noise and interconnection noise, plus sufficiently rich data; if those bounds mischaracterize the real noise, the data-consistent sets are wrong and the certified controller may not stabilize the actual system.","fun_headline_variants_meta":{"raw":{"variants":["Local data alone certify decentralized stability for unknown grids","Data-driven dissipativity stabilizes interconnected systems locally","Stabilize a network without a model: local data suffices","Per-subsystem data yields global stability in interconnected systems","From local measurements to global stability: a data-driven route"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000681,"raw_usage":{"total_tokens":2883,"prompt_tokens":650,"completion_tokens":2233,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":394,"completion_tokens_details":{"reasoning_tokens":2154}},"tokens_in":394,"tokens_out":2233,"duration_ms":17593,"temperature":1.0,"reasoning_tokens":2154,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:26:58.643343+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate random systems whose noise stays inside the assumed quadratic bounds, solve the LMIs (20), (27b), (33), and simulate the noiseless closed loop with the true system and true interconnection. If any feasible controller yields a diverging trajectory, the claimed sufficiency chain is false; because the paper proves sufficiency, a single such counterexample settles the question.","supporting_citations":[],"review_version":1}