{"id":"11812f39-7682-4052-b325-5e602e836601","arxiv_id":"2411.10243","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A data-driven SDP design yields decentralized stabilizing state-feedback gains for discrete-time large-scale interconnected systems without identifying system matrices.","lead":"Large-scale systems with unknown dynamics are stabilized using a data-driven method that constructs decentralized controllers directly from measured state, input, and interconnection signals, bypassing model identification. The approach solves a semi-definite program and is demonstrated on a five-mass spring chain.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4 lacks an explicit symmetry/positive-definiteness constraint on S_i = X_i Q_i, making the Lyapunov matrix and the congruence step in the proof not well-defined as stated.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that verdict, but the single most load-bearing issue is more specific than the reader's stated weakest assumption. The reader's weakest_assumption concerns measurability of g_{ij} and known W_{ij}; that is an important practical limitation and is honestly stated in Section 2.1, so it is not an internal flaw. The symmetry gap, by contrast, sits inside the proof of the main theorem: S_i = X_i Q_i is used as a Lyapunov matrix, yet no condition guarantees S_i is symmetric positive definite. The congruence step (23) to (29) and the subsequent Schur complement only produce the claimed Z_i blocks when S_i is symmetric, because H_{i,1}^T = (Q_i S_i^{-1})^T = S_i^{-T}Q_i^T, which equals S_i^{-1}Q_i^T only if S_i^T = S_i. Moreover, if S_i is nonsymmetric, -S_i in the (1,1) position makes the LMI ill-posed. A charitable reading of the star notation supplies the missing symmetry, and then -S_i ≺ 0 gives S_i ≻ 0, so the theorem is repairable; this is why the appropriate verdict is CONDITIONAL rather than REJECT. The concrete test, namely a symbolic re-derivation plus a 2D nonsymmetric feasibility search, would settle whether the omission is merely presentational or hides an infeasibility. Overall, the data-driven parameterization and the sector-bound Lyapunov argument are plausible, and the spring-mass example gives no evidence of a deeper failure. I therefore leave the reader's verdict unchanged.","tokens_in":9854,"tokens_out":21462,"duration_ms":199064,"concrete_test":"Re-derive the step from (23) to (29) symbolically without assuming S_i^T = S_i, and check whether the (1,3) block after congruence equals H_{i,1}^T X_{i,[1,T]}^T. Then, for a 2-state subsystem with randomly generated full-rank data, parameterize S_i as [a b; c d] and search for a feasible point of (23) with b ≠ c; if such a nonsymmetric S_i exists, the theorem as stated allows invalid Lyapunov matrices, confirming the need for the explicit symmetry/positive-definiteness constraint. If no such point exists, the omission is only presentational and the proof can be repaired by adding the constraint S_i = S_i^T ≻ 0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is internal to Theorem 4: the matrix S_i is defined as S_i = X_{i,[0,T-1]}Q_i with Q_i a free decision variable, but the theorem never states that S_i must be symmetric positive definite. The proof's Lyapunov function V_i = x_i^T S_i^{-1}x_i requires S_i^{-1} to be positive definite, and the key step 'pre- and post-multiplying (23) with diag{S_i^{-1},I,I,I}' yields the claimed block (1,3) equal to H_{i,1}^T X_{i,[1,T]}^T only when S_i^T = S_i. If S_i is not symmetric, the matrix in (23) is not a valid negative-definite LMI because its (1,1) block -S_i is nonsymmetric, so the expression is undefined. A charitable reading of the ⋆ convention implicitly enforces symmetry, and then -S_i ≺ 0 forces S_i ≻ 0, which would repair the statement. But the theorem as written does not say this, and a free search over Q_i will generally not produce symmetric S_i. This is not a fatal flaw in the approach, but it is a genuine gap in the central theorem: without an explicit S_i = S_i^T ≻ 0 constraint, or a symmetric slack variable in place of S_i, the proof does not go through and the formula K_i = U_i Q_i S_i^{-1} may be evaluated with an invalid Lyapunov matrix.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven decentralized state-feedback stabilization method for discrete-time large-scale systems with unknown subsystem dynamics and sector-bounded interconnections. Under a rank/persistency-of-excitation condition, Lemmas 1–2 and Theorem 3 show that the open-loop and closed-loop subsystem dynamics can be represented exactly using collected data matrices U_{i,[0,T-1]}, Φ_{i,[0,T-1]}, X_{i,[0,T-1]}, and X_{i,[1,T]}. The main result, Theorem 4, formulates an LMI in variables Q_i, with S_i = X_{i,[0,T-1]}Q_i and the constraint Φ_{i,[0,T-1]}Q_i = 0, and gives the decentralized gains K_i = U_{i,[0,T-1]}Q_i S_i^{-1}. Stability is proved via a sum of local Lyapunov functions and the sector bounds on interconnections. A simulation on a five-mass spring-chain illustrates the design.","tokens_in":10156,"tokens_out":14436,"duration_ms":131744,"significance":"If the central theorem is repaired, the paper makes a useful incremental contribution by extending the Persis–Tesi data-driven framework to decentralized control of interconnected large-scale systems. The data-driven representation in Lemma 2 and Theorem 3 is sound under the stated rank condition, and the SDP formulation is direct: the controller gains are parameterized by data matrices without an explicit model identification step. The authors openly acknowledge the main limitations (no measurement noise, no delays or uncertainties). The numerical example is simple, but the theory is the primary contribution. The paper is not circular: Q_i are decision variables of an LMI, not parameters fitted to reproduce a desired stability outcome.","major_comments":[{"comment":"The variable S_i = X_{i,[0,T-1]} Q_i is not constrained to be symmetric or positive definite. For the block matrix in (23) to be a valid negative-definite LMI, the diagonal blocks -S_i must be symmetric, and negative definiteness would then imply S_i ≻ 0 and hence invertibility. However, the theorem only says 'there exist matrices Q_i'; a free Q_i will generally produce a nonsymmetric S_i, making the expression in (23) undefined as an LMI. This also affects the proof: the Lyapunov function V_i = x_i^T S_i^{-1} x_i in (26) and the congruence transformation with diag{S_i^{-1}, I, I, I} before (29) are justified only if S_i is symmetric positive definite. This is a load-bearing gap, because (23) is the only synthesis condition and (25) is undefined if S_i is not invertible. The authors should add the explicit constraints S_i = S_i^T ≻ 0 (e.g., declare S_i symmetric and impose the linear equality S_i = X_{i,[0,T-1]} Q_i), or introduce a symmetric slack variable in place of S_i.","section":"Theorem 4, Eqs. (23)–(26)"}],"minor_comments":[{"comment":"The definition of G_i repeats G_{i1}; it should read G_i = [G_{i1} G_{i2} ··· G_{iM}].","section":"Section 2.1, Eq. (3)"},{"comment":"The term (I - Y†Y)w is dimensionally inconsistent; the general least-squares solution should be Ξ* = X Y† + Z(I - Y†Y) for a matrix Z of size n_i × T. The conclusion is unaffected because the projector vanishes under (11).","section":"Appendix A, Eq. (A.2)"},{"comment":"The dimensions of the decision variables should be stated, e.g., Q_i ∈ R^{T × n_i}, and the matrix W_i should be defined in the theorem statement rather than only in the proof.","section":"Theorem 4"},{"comment":"The estimator \\bar w_{ij} = max ||g_{ij}(x_j)||/||x_j|| is undefined at x_j = 0; it should be stated as a supremum over x_j ≠ 0.","section":"Remark 1"},{"comment":"The instructions say to form U_{i,[0,T-1]} again after it was formed in Step (2); rephrase to avoid duplication.","section":"Section 4, Step (3)"},{"comment":"Exact measurability of all interconnection inputs g_{ij}(x_j) is a strong requirement for large-scale systems; the paper would benefit from a discussion of how these signals are obtained in practice, particularly in the spring-mass example.","section":"Assumption 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable incremental extension of the data-driven control literature to decentralized systems. The main theorem is not stated rigorously because of the missing symmetry/positive-definiteness condition on S_i; the fix is straightforward, so I recommend major revision rather than rejection. The simulation is only a proof of concept and does not compare with model-based baselines; this is acceptable for the paper's scope but should be acknowledged. No concerns about plagiarism or citation practices."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know: this paper takes the well-known data-driven control formulas of Persis and Tesi and extends them to decentralized stabilization of a class of discrete-time large-scale systems with sector-bounded interconnections. The core idea—treating the interconnection signals as measurable 'inputs' alongside the control inputs and using a rank condition to get a data-driven representation—is sound and genuinely new in this combination. The paper is clearly organized and the derivations are mostly straightforward.\n\nThe main theorem, Theorem 4, has a real gap in its statement. The LMI (23) uses -S_i as the (1,1) block, with S_i = X_i Q_i, and the proof defines a Lyapunov function x_i^T S_i^{-1} x_i, which requires S_i to be symmetric positive definite. But the theorem only says 'there exist matrices Q_i'; it doesn't explicitly require S_i to be symmetric or PD. A charitable reading of the ⋆ notation might implicitly enforce symmetry, and then -S_i ≺ 0 would give S_i ≻ 0. But that should be stated. As written, a feasible search over unconstrained Q_i will not generally produce symmetric S_i, and the congruence step and Lyapunov argument are not well-defined. This is fixable by adding S_i = S_i^T ≻ 0 (or a symmetric slack variable) to the conditions, but it needs to be fixed before the theorem is correct.\n\nThe other soft spot is practical rather than mathematical. The method requires exact measurement of every interconnection input g_ij(x_j) and known sector-bound matrices W_ij. That's a strong assumption for large-scale systems; interconnections are often not directly measurable. The paper mentions it, but the limitation deserves more emphasis. The simulation on a mass-spring chain is a feasibility demo: no baseline, no noise, no code or data. It shows the method works on a simple example, but little more.\n\nOverall, the paper is a reasonable incremental contribution. The data-driven representation and the LMI formulation are plausible, and the sector-bounded interconnection treatment is standard but clean. The missing symmetry constraint in Theorem 4 is a genuine but local flaw; I don't think it sinks the approach. I'd send it to peer review, with the expectation that the authors tighten Theorem 4 and discuss the measurability assumption more carefully. I probably wouldn't cite it in my own work this year, but it's worth a reading-group slot.","headline":"A useful but rough extension of Persis–Tesi to interconnected systems; Theorem 4 needs an explicit symmetry/PD constraint on the Lyapunov matrix before it is correct as stated.","tokens_in":10680,"tokens_out":3276,"would_cite":false,"duration_ms":30793,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93A14","93C55","93D20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Decentralized stabilizing controllers can be computed directly from measured state, control, and interconnection data, without identifying the system's model.","keywords":["data-driven control","decentralized control","large-scale systems","stabilization","linear matrix inequalities","discrete-time systems","persistency of excitation","interconnected systems"],"falsifier":"Take a two-subsystem instance of the spring-mass chain, deliberately set $W_{ij}$ to a value smaller than the true Lipschitz constant of $g_{ij}$, and record data while satisfying the rank condition. If the resulting LMI (23) is feasible but the closed-loop system under the computed $K_i$ fails to converge, that would show the stability certificate is invalid when the sector-bound knowledge is wrong. Alternatively, corrupt the measured interconnection inputs with small noise (e.g., $\\|\\eta(k)\\| \\approx 0.01\\|\\phi_i(k)\\|$) and re-run the design; instability under the resulting gains would demonstrate the method's sensitivity to exact measurability.","tokens_in":9622,"feed_emoji":"⚙️","tokens_out":5540,"duration_ms":49464,"temperature":0.7,"pith_summary":"This paper establishes that for a class of discrete-time large-scale systems composed of nonlinear interconnected subsystems, decentralized state-feedback controllers can be designed directly from measured data without first identifying the subsystem matrices. The key is an exact data-driven representation of the closed-loop system, valid when the recorded control and interconnection input sequences are persistently exciting. Under a sector-bound condition on the interconnection nonlinearities, the stabilizing gains are obtained by solving a semi-definite program (an LMI) per subsystem, with an auxiliary constraint that zeros out the interconnection-related data. The result is demonstrated on a chain of five spring-mass systems whose velocities converge to a target.","feed_headline":"Data alone yield decentralized stabilizing controllers for large-scale systems","feed_subtitle":"A semidefinite program over measured state, input, and interconnection data replaces system identification.","key_machinery":"The load-bearing mechanism is Theorem 3's exact data-driven representation of the closed loop, which converts the unknown matrices $\\{A_i, B_i, G_i\\}$ into the measured data block $X_{i,[1,T]}$ and two parameter matrices $H_{i,1}$, $H_{i,2}$. The matrix $H_{i,1}$ carries the controller gain through $K_i = U_{i,[0,T-1]} H_{i,1}$, while the constraint $\\Phi_{i,[0,T-1]} Q_i = 0$ ensures that the row block of data corresponding to interconnection inputs is annihilated, so the effective feedback only uses local state. Stability is established through the Lyapunov matrix $S_i^{-1}$ and the LMI (23), which is feasible if there exist $Q_i$ making the data-driven closed-loop matrix contractive with respect to the sector bounds $W_{ij}$.","core_discovery":"The central discovery is that the closed-loop behavior of each controlled subsystem can be written entirely in terms of data matrices: $x_i(k+1) = X_{i,[1,T]}(H_{i,1} x_i(k) + H_{i,2} \\phi_i(k))$, where $H_{i,1}$ and $H_{i,2}$ are solutions of linear equations built from collected state, input, and interconnection input matrices. Because these solutions exist whenever the input data is persistently exciting of order $n_i+1$, the feedback gain becomes a parameter of the data representation rather than a function of an unknown model. Stability is enforced by a Lyapunov function $V = \\sum_i x_i^\\top S_i^{-1} x_i$ with $S_i = X_{i,[0,T-1]} Q_i$ and an LMI equivalent to $\\sum_i (\\Delta V_i + \\hat{\\Theta}_i) < 0$, where the $\\hat{\\Theta}_i$ terms account for the interconnection bounds. Theorem 4 therefore reduces decentralized stabilization to a data-only semidefinite program: find $Q_i$ satisfying (23) and $\\Phi_{i,[0,T-1]} Q_i = 0$, and set $K_i = U_{i,[0,T-1]} Q_i S_i^{-1}$, which asymptotically stabilizes the large-scale system.","pith_inferences":["The paper requires exact measurement of interconnection inputs and known sector-bound matrices $W_{ij}$; if those are only estimated with error, the stability certificate degrades, and one could add robustness margins to the LMI to account for noise, though that extension is not made here.","The same data representation could be reused for other objectives, such as LQR-like performance or disturbance rejection, by replacing the Lyapunov inequality in (23) with the corresponding performance inequality while keeping the identification-free structure.","The persistence-of-excitation condition on interconnection inputs means that data collection experiments must actively excite the coupling channels; in applications where interconnections cannot be injected externally, one might rely on naturally occurring excitation or use closed-loop data, which would be a natural extension of the approach.","Since the LMI is solved per subsystem, the method is naturally parallelizable and could scale to very large networks, but the conservativeness incurred by bounding all interconnections with known $W_{ij}$ suggests that exploiting the actual sparse interconnection topology could reduce conservatism."],"forward_implications":["If the LMI (23) is feasible and the data satisfy the rank condition, the gains $K_i = U_{i,[0,T-1]} Q_i S_i^{-1}$ guarantee asymptotic stability of the interconnected closed-loop system, without any model identification step.","The design is fully decentralized: each subsystem's gain is computed from its own local data and the interconnection input data, and the resulting controller uses only local state feedback.","The approach applies to any interconnection nonlinearity satisfying the sector inequality $\\|g_{ij}(r)-g_{ij}(s)\\| \\le \\|W_{ij}(r-s)\\|$, with the $W_{ij}$ matrices entering the LMI; no other model information is needed.","The validation on a five-mass spring chain shows the method in action, with the velocities of all masses converging to the target in about four seconds.","Because the design is a semidefinite program per subsystem, the computational cost scales with subsystem size rather than the full large-scale system."],"supporting_citations":[{"why":"Supplies the persistency-of-excitation lemma used in Lemma 1 to guarantee the rank condition on the collected data.","marker":"Persis and Tesi (2020)"},{"why":"Establishes the notion of persistent excitation and the fundamental lemma that this paper generalizes to interconnected systems with interconnection inputs.","marker":"Willems et al. (2005)"},{"why":"Provides the Moore–Penrose inverse solution to the least-squares problem used in Lemma 2 to prove exact data-driven representation.","marker":"Penrose (1956)"}],"fun_headline_variants":["Data-only decentralized control stabilizes large-scale systems","Skip modeling, use data for decentralized stabilization","No model needed: data yield stabilizing controllers","One-step data-driven design for large-scale control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that every interconnection input $g_{ij}(x_j)$ can be measured exactly and that the sector-bound matrices $W_{ij}$ are known; if the coupling signals between subsystems cannot be measured or bounded, the data matrices and the LMI cannot be formed, and the whole design collapses.","fun_headline_variants_meta":{"raw":{"variants":["Data-only decentralized control stabilizes large-scale systems","Skip modeling, use data for decentralized stabilization","No model needed: data yield stabilizing controllers","One-step data-driven design for large-scale control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1343,"prompt_tokens":918,"completion_tokens":425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":368}},"tokens_in":534,"tokens_out":425,"duration_ms":4759,"temperature":1.0,"reasoning_tokens":368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:49:10.315726+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-subsystem instance of the spring-mass chain, deliberately set $W_{ij}$ to a value smaller than the true Lipschitz constant of $g_{ij}$, and record data while satisfying the rank condition. If the resulting LMI (23) is feasible but the closed-loop system under the computed $K_i$ fails to converge, that would show the stability certificate is invalid when the sector-bound knowledge is wrong. Alternatively, corrupt the measured interconnection inputs with small noise (e.g., $\\|\\eta(k)\\| \\approx 0.01\\|\\phi_i(k)\\|$) and re-run the design; instability under the resulting gains would demonstrate the method's sensitivity to exact measurability.","supporting_citations":[{"cited_title":"and Tesi, P","cited_arxiv_id":null,"evidence_quote":"Supplies the persistency-of-excitation lemma used in Lemma 1 to guarantee the rank condition on the collected data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the notion of persistent excitation and the fundamental lemma that this paper generalizes to interconnected systems with interconnection inputs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Moore–Penrose inverse solution to the least-squares problem used in Lemma 2 to prove exact data-driven representation."}],"review_version":1}