{"id":"f09e1ad4-7c24-40fd-9531-d5229f96ea86","arxiv_id":"2411.11542","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A single semidefinite program computes a structured static state-feedback gain with a guaranteed H2 bound for all linear systems consistent with noisy measured data.","lead":"This paper develops a convex optimization method that designs a structured state-feedback controller, one whose gain matrix respects a fixed sparsity or coordination pattern, directly from noisy input-state data. The controller comes with a certified worst-case H2 performance bound that holds for every plant consistent with the observed data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (16) defines Ψ as G^T Φ G with incompatible dimensions, so the central object Σ_D and Theorem 3.2 are not well-posed as written; the intended formula appears to be G Φ G^T.","rationale":"I read the paper in good faith and the intended construction is recognizable: combine an extended H2 LMI, the Ferrante-Dabbene-Ravazzi structured-control condition, and the matrix S-procedure to certify worst-case H2 performance over the data-consistent plant set. The numerical examples and DOI-provided code are real evidence, and the broad approach is standard. The reader's chosen weakest point (missing Slater conditions in Lemma 3.1) is less damaging than thought because the implication actually used in Theorem 3.2 is the S-procedure's sufficient direction, which does not require strict feasibility; the converse would only matter for exactness/optimality guarantees beyond what the theorem states. Instead, the load-bearing defect is in Eq. (16): the displayed product G^T Φ G is dimension-inconsistent given Φ ∈ S^{n+T} and G of size (2n+m)×(n+T). Since Σ_D, Lemma 3.1, and the main theorem all use this Ψ, the theorem is not well-posed as written. This is an internal inconsistency, not a disagreement with consensus, and it is fixable; hence the appropriate verdict remains CONDITIONAL rather than REJECT, but the typo must be corrected and the proof rechecked. I would keep the reader's verdict but for a different, more concrete reason, hence partial agreement.","tokens_in":11905,"tokens_out":29315,"duration_ms":269306,"concrete_test":"Independently derive Ψ from the noise bound: substitute W_- = X_+ - A X_- - B U_- into [I; W_-^T]^T Φ [I; W_-^T] ⪰ 0, and verify the resulting matrix is G Φ G^T rather than G^T Φ G. Then correct Eq. (16), rebuild Σ_D, and re-solve (18) on the two numerical examples; check that the reported H2 bounds in Tables I-IV are reproduced and that the LMI dimensions are consistent (Ψ ∈ S^{2n+m}). If the corrected SDP yields different values or becomes infeasible, the numerical support for Theorem 3.2 is materially affected.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The theorem's proof rests entirely on the set Σ_D of data-consistent plants. In Eq. (16), let G = [[I, X_+],[0,-X_-],[0,-U_-]], which has size (2n+m) × (n+T). The paper defines Ψ := G^T Φ G. But Φ ∈ S^{n+T}, so G^T is (n+T)×(2n+m) and G^T Φ has inner dimensions 2n+m and n+T; the product is undefined (unless 2n+m = n+T). The same mismatch makes G^T Φ G ill-typed, so Σ_D in (17) is not defined, and the SDP (18) that subtracts αΨ cannot be interpreted. Re-deriving the consistency set by substituting W_- = X_+ - A X_- - B U_- into [I; W_-^T]^T Φ [I; W_-^T] ⪰ 0 gives Ψ = G Φ G^T (here G^T z = [I; W_-^T] with z = [I; A^T; B^T]); the transpose is on the wrong factor in the printed equation. The Slater/strict-feasibility issue in Lemma 3.1 is secondary: the direction actually used in Theorem 3.2 (feasibility of the multiplier LMI implies performance for all plants in Σ_D) is the unconditional S-procedure sufficiency, so even if the stated 'if and only if' needs regularity, the forward implication survives. The dimension typo, however, blocks any formal check of the central claim and should be corrected before the theorem can be evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a convex LMI-based method for data-driven H2-suboptimal control with structured static state-feedback gains. Given input-state data collected under a bounded process noise satisfying a quadratic matrix inequality (QMI), the authors characterize the set of all plants consistent with the data via a matrix ellipsoid. They then impose the subspace constraint on the controller gain using a convex sufficient condition from [7] and apply a matrix S-lemma to convert the worst-case H2 performance condition into an LMI. The main result, Theorem 3.2, states that a feasible solution of the resulting SDP yields a structured gain that is gamma*-suboptimal for every plant in the data-consistent set, and the SDP size is independent of the number of data samples. Two numerical examples, including sparse and sharing control, illustrate the method.","tokens_in":12183,"tokens_out":10630,"duration_ms":93727,"significance":"If the main theorem is correct, the paper offers a computationally appealing convex formulation for a class of structured H2 control problems that are generally NP-hard, with guaranteed worst-case performance over all data-consistent plants. The dimension of the SDP being independent of the data length T is a genuine practical strength, and the authors provide publicly available MATLAB code for reproducibility. The contribution is an incremental combination of existing tools: the QMI/set-membership framework of [11], [12], the structured-control relaxation of [7], and the extended H2 LMI of [34]. That said, the central claim is clear and the intended derivation is plausible, aside from the load-bearing issues detailed below. The paper is a reasonable fit for the journal, provided the major concerns are addressed.","major_comments":[{"comment":"The definition of Ψ in Eq. (16) is dimensionally invalid as printed. The matrix G has size (2n+m) × (n+T) while Φ has size (n+T) × (n+T); consequently the product G^T Φ G is undefined unless 2n+m = n+T, which is not assumed. Since Ψ is used to define the consistency set Σ_D in (17) and appears in the main SDP (18a), the central claim of Theorem 3.2 is not well-posed as written. The correct expression is Ψ = G Φ G^T, which follows by substituting W_- = X_+ - A X_- - B U_- into (14) and comparing with the quadratic form in (17). Please correct this and confirm that all subsequent blocks, especially the embedding in (18a), are dimensionally consistent.","section":"Section III-A, Eq. (16)"},{"comment":"Lemma 3.1 is stated as an if-and-only-if but does not restate the strict feasibility or regularity conditions required by the cited matrix S-lemma [11, Cor. 4.13]. The forward (sufficiency) direction, which is the only direction used in the proof of Theorem 3.2, is the standard unconditional S-procedure argument and is not endangered. However, the claimed equivalence, and the statement that (18a) is 'equivalent' to the robust performance condition, are not justified without those conditions. Please either state the required regularity conditions or reformulate Lemma 3.1 as the sufficiency result that the theorem actually needs.","section":"Section III-B, Lemma 3.1"},{"comment":"The last paragraph of Section III-B claims that 'the only non-conservative step leading up to Theorem 3.2 is the convex relaxation based on Lemma 2.3' and that this is 'also the only source of conservatism.' This is contradicted by the paper's own discussion in Section IV.A and Table II, where the authors note that increasing the data length T can lead to more conservative bounds due to the approximation of the per-sample noise bound by a single ellipsoid in (14). The ellipsoidal overapproximation of the noise set is a distinct source of conservatism. Please revise this claim so that it accurately lists all sources of conservatism, and ensure that Lemma 3.1's missing regularity conditions, if not addressed, are also acknowledged.","section":"Section III-B, end of Section III-B"}],"minor_comments":[{"comment":"The row labels '3 (P = R diag.)' and '3 (R diag.)' repeat the number 3; the first of these should presumably be '2 (P = R diag.)' to match the design numbering used elsewhere.","section":"Table IV"},{"comment":"The notation '13×2' in the sparsity pattern appears to be a typographical artifact; it should likely be '1_{3\\times 2}' or another explicit block notation.","section":"Section IV.B, Eq. (21a)"},{"comment":"After fixing Eq. (16), please make the block structure of the matrix in (18a) explicit, in particular the dimensions of the Ψ embedding and the size of the full constraint, so that the 'scales independently of T' claim is directly verifiable.","section":"Section III-C and (18a)"},{"comment":"The proof is a single sentence appealing to a Schur complement and Lemma 3.1. Given the dimension issue in Eq. (16), a few more steps showing how the 4x4 block in (18a) arises from the Schur complement of the lemma would greatly improve readability and verifiability.","section":"Proof of Theorem 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental but useful combination of established tools: the QMI/set-membership data-driven framework, the structured-control relaxation of Ferrante et al., and the extended H2 LMI of de Oliveira et al. The main theorem appears sound once the transpose error in Eq. (16) is fixed and the S-lemma regularity conditions are stated properly. The 'only source of conservatism' claim is an overstatement and should be corrected. The novelty is modest but acceptable for the journal; the availability of reproducible code is a positive aspect. I recommend major revision rather than rejection because the issues are correctable and the core idea is defensible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, workmanlike paper that combines three known ingredients—the QMI data-consistency framework, the matrix S-lemma, and a convex sufficient condition for structured feedback—into one SDP for data-driven structured H2 control. The problem is genuinely new in this form, and the central theorem is plausible. The manuscript has one blocking typo and one overclaim, both fixable.\n\nWhat's actually new: Theorem 3.2's SDP (18) is the first convex formulation I know of that gives a hard worst-case H2 guarantee for structured static state feedback using only measured data and a noise bound, with problem size independent of the number of samples. The combination is straightforward—the paper doesn't invent a new framework—but it is a useful application and the structured-control condition from [7] is a good fit for the QMI data model. The examples are small but sensible; Design 4 (R in Υ(S)) clearly beats the diagonal-R alternatives, and the code is available via DOI.\n\nSoft spots, in order of severity:\n\n1. Eq (16) is ill-typed as printed. The block matrix G is (2n+m)×(n+T) and Φ is (n+T)×(n+T), so GᵀΦG doesn't parse unless 2n+m = n+T. The intended definition is clearly Ψ = GΦGᵀ, which is (2n+m)×(2n+m) and makes Σ_D in (17) well-formed. This looks like a simple transpose typo, but it sits in the central object and blocks any formal verification of the theorem. Fix it.\n\n2. The paper claims the only source of conservatism is the convex relaxation from Lemma 2.3. That's contradicted by their own Remark 3 and Table II discussion, where overapproximating per-sample noise bounds by a single ellipsoid is identified as another source. The claim should be retracted or substantially qualified.\n\n3. Lemma 3.1 states an if-and-only-if without restating the regularity conditions from [11, Cor. 4.13]. The direction actually used in Theorem 3.2 is the unconditional S-procedure sufficiency, so the main theorem doesn't depend on those conditions—but the lemma as written is too strong.\n\nMinor: the examples are demonstrations rather than a systematic benchmark, and the numerical code depends on Mosek and YALMIP with no commit hash. That's fine for a first paper on this topic, but the reproducibility claim is a bit weaker than the DOI suggests.\n\nWho this is for: researchers working on data-driven control or structured/decentralized control. They'll find the SDP useful and the paper readable, despite the typos. It deserves a serious referee—the idea is sound and the community will want to see it. My recommendation: accept after the typo is fixed and the conservatism claim is corrected.","headline":"Solid combination of known tools into a new SDP for data-driven structured H2 control; a dimension typo and an overclaim about conservatism need fixing before the details are trustworthy.","tokens_in":12756,"tokens_out":4632,"would_cite":true,"duration_ms":37620,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B52","93B51","93C55","90C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that one convex linear-matrix-inequality program, built only from measured input-state data and a known noise bound, certifies a worst-case H2 performance bound for a structured state-feedback controller on every plant…","keywords":["data-driven control","structured control","H2 suboptimal control","linear matrix inequality","set-membership","matrix S-lemma","quadratic matrix inequality","state feedback"],"falsifier":"Construct a data set for which Sigma_D is nonempty but has empty interior (for example, exact noiseless data with fewer samples than states), solve SDP (18), and compare the certified gamma* against an exhaustive grid search for the worst-case H2 norm over all A,B satisfying the QMI; a certified bound below the grid-search worst case would show the hidden regularity condition is needed for the theorem to hold.","tokens_in":11659,"feed_emoji":"🎛️","tokens_out":4997,"duration_ms":48614,"temperature":0.7,"pith_summary":"This paper claims that H2-suboptimal regulation with a structured state-feedback gain can be certified directly from noisy input-state data, without identifying the matrices A and B. It combines a known convex relaxation for structured control with a matrix S-lemma that turns the requirement \"performance holds for every plant consistent with the data\" into one linear matrix inequality. The result is a single convex program whose size depends on the state and input dimensions but not on the number of collected samples. A feasible solution produces a structured gain K = $LR^{{-1}}$ that meets the reported H2 bound on the true plant and on every plant in the data-consistent set. The significance is that networked or communication-limited controllers can be designed with guaranteed worst-case performance from data alone.","feed_headline":"Data alone certifies structured H2 control bounds","feed_subtitle":"A single convex program turns noisy input-state data into a state-feedback gain with guaranteed worst-case H2 performance.","key_machinery":"The paper's central objects are the QMI-ellipsoid set Sigma_D defined in (17) from the measured matrices X-, U-, X+ and the noise bound Phi; the matrix S-lemma (Lemma 3.1) that converts the uniform performance condition over Sigma_D into an LMI with multipliers $\\alpha$, $\\beta$; and the subspace sets Upsilon(S) from Lemma 2.3, which give a convex sufficient condition R in Upsilon(S), L in S implies $LR^{{-1}}$ in S. These are assembled into the SDP (18), where R acts as a slack Lyapunov factor and L = KR. The SDP's dimensions scale with n and m but not with the trajectory length T.","core_discovery":"Theorem 3.2 is the central claim: given input-state data collected under a known quadratic noise bound and a subspace S of allowed controller gains, if the semidefinite program (18) is feasible with optimal value gamma*, then K = $LR^{{-1}}$ belongs to S and is a gamma*-suboptimal H2 controller for every system (A,B) in Sigma_D, the set of plants compatible with the data. Because the true plant is one of these compatible plants, the controller is guaranteed to regulate the actual system as well. The authors identify the convex condition of Lemma 2.3, used to enforce the subspace constraint, as the only source of conservatism in the design procedure.","pith_inferences":["The same QMI-plus-S-lemma template likely extends to H-infinity or mixed-sensitivity objectives, as the paper's future-work section suggests, but each extension needs a new QMI representation of the performance condition.","A quantitative trade-off map between sparsity, data length, noise level, and achievable H2 bound could be extracted from the SDP's feasible region or dual variables; the paper only provides example-based observations.","If Slater regularity fails, a small robustification of the data ellipsoid (slightly shrinking Phi or adding regularization) could restore the exactness of the S-procedure step, at the cost of a slightly conservative bound; this is a testable fix.","The non-monotonic behavior in T seen in Tables II and IV stems from approximating per-sample noise by a single ellipsoid, so alternative noise models such as polytopic bounds might yield monotone guarantees."],"forward_implications":["A single convex LMI certifies a worst-case H2 bound on every plant consistent with the data, so no separate system identification step is needed.","The SDP's size scales with the state and input dimensions but not with the number of collected samples, so long trajectories do not increase the optimization cost.","Structured constraints are encoded through Upsilon(S), so sparsity patterns, decentralized information constraints, or sharing constraints can be enforced while retaining convexity.","The resulting controller K is guaranteed to meet the reported bound on the true underlying system even though A and B are unknown."],"supporting_citations":[{"why":"Supplies the convex sufficient condition R in Upsilon(S), L in S implies LR^{-1} in S that underlies the structured-control relaxation in Lemma 2.3.","marker":"[7]"},{"why":"Provides the QMI set-membership framework and the matrix S-lemma corollary (Corollary 4.13) on which Lemma 3.1's equivalence is based.","marker":"[11]"},{"why":"Shows how noisy data can directly yield robust feedback controllers via a matrix S-lemma, the template this paper extends to structured H2 control.","marker":"[12]"},{"why":"Supplies the extended H2 LMI with the slack variable R on which Lemma 2.2 and the R/L parametrization in the SDP are built.","marker":"[34]"},{"why":"Is the classical S-procedure result that underlies the matrix-form S-lemma used in the data-driven certification.","marker":"[13]"}],"fun_headline_variants":["One SDP proves structured H2 control from data","Data-driven H2 regulation with worst-case guarantees","Guaranteed H2 bounds via data-driven LMIs","Structured H2 control from noisy input-state data","Convex certifier for data-driven structured control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the S-procedure step being exact: the data set must contain at least one strictly data-consistent plant in the interior of the ellipsoid, and if that Slater-type condition fails, the LMI certificate may not reflect the true worst case.","fun_headline_variants_meta":{"raw":{"variants":["One SDP proves structured H2 control from data","Data-driven H2 regulation with worst-case guarantees","Guaranteed H2 bounds via data-driven LMIs","Structured H2 control from noisy input-state data","Convex certifier for data-driven structured control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000277,"raw_usage":{"total_tokens":1578,"prompt_tokens":798,"completion_tokens":780,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":414,"completion_tokens_details":{"reasoning_tokens":705}},"tokens_in":414,"tokens_out":780,"duration_ms":8194,"temperature":1.0,"reasoning_tokens":705,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:24:12.613089+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a data set for which Sigma_D is nonempty but has empty interior (for example, exact noiseless data with fewer samples than states), solve SDP (18), and compare the certified gamma* against an exhaustive grid search for the worst-case H2 norm over all A,B satisfying the QMI; a certified bound below the grid-search worst case would show the hidden regularity condition is needed for the theorem to hold.","supporting_citations":[{"cited_title":"On the design of structured stabilizers for LTI systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the convex sufficient condition R in Upsilon(S), L in S implies LR^{-1} in S that underlies the structured-control relaxation in Lemma 2.3."},{"cited_title":"Quadratic matrix inequalities with applications to data- based control,","cited_arxiv_id":null,"evidence_quote":"Provides the QMI set-membership framework and the matrix S-lemma corollary (Corollary 4.13) on which Lemma 3.1's equivalence is based."},{"cited_title":"From no isy data to feedback controllers: Nonconservative design via a matrix S- lemma,","cited_arxiv_id":null,"evidence_quote":"Shows how noisy data can directly yield robust feedback controllers via a matrix S-lemma, the template this paper extends to structured H2 control."},{"cited_title":"Exte nded H2 and H∞ norm characterizations and controller parametrizations f or discrete-time systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the extended H2 LMI with the slack variable R on which Lemma 2.2 and the R/L parametrization in the SDP are built."},{"cited_title":"S-Procedure in Nonlinear Control Th eory,","cited_arxiv_id":null,"evidence_quote":"Is the classical S-procedure result that underlies the matrix-form S-lemma used in the data-driven certification."}],"review_version":1}