{"id":"74a1ed65-0d8c-454a-8b68-ea81607ddde3","arxiv_id":"1909.02192","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For closed-loop data, the REDAR algorithm (VARX fit plus balanced reduction) has one-step-ahead prediction error bounded by the optimal error plus terms that decay with model order p and with sample size T as O(1/√T).","lead":"This paper proves finite-sample error bounds for a closed-loop system identification method that first fits an autoregressive model and then reduces its order. The result gives engineers a data-dependent guarantee on prediction accuracy when data are collected under feedback control.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 8/10 use λmin(Γ) as a lower bound on λmin(Q); the asserted inequality is backwards and false for a simple AR(1), so the finite-data proof is not established as written.","rationale":"The reader's nominated weakest assumption is Gaussian noise. That is an explicit assumption of Section II-B, so it is not a gap in the paper's own claim; it would only matter for extending the result to non-Gaussian noise. The load-bearing issue is internal: the finite-data analysis requires a positive lower bound on Q=E[d_t d_t^T], and the proof chooses λmin(Γ) but asserts the wrong inequality. The scalar AR(1) example shows λmin(Q) can be smaller, so Lemma 8's bound is not valid as written. This does not necessarily falsify Theorem 1, because Q is fixed for a fixed system and λmin(Q)>0; replacing ξ by λmin(Q) would likely restore the same O(1/√T) rate with constants depending on that covariance. For that reason I would keep the conditional verdict, but the authors should correct Lemma 8 and Lemma 10. I also note the printed Theorem 1's finite-order term appears to miss the square on Lρ^{p+1}/(1−ρ) and on ||z||_P, which Section III-B shows should be squared; this looks like a typo rather than a conceptual problem. No code is supplied, so the simulation cannot be independently checked, but the reader's conditional verdict already accounts for that.","tokens_in":14401,"tokens_out":23123,"duration_ms":238400,"concrete_test":"Run the scalar closed-loop example with a=0.9, b=0.1, p=2, e_t,v_t iid N(0,1). Compute Q=E[d_t d_t^T] numerically (d_t=[z_{t-2};z_{t-1}]) and verify λmin(Q)≈0.526<1=λmin(Γ). Then with ΔQ=0 and α/T=10^{-6}, compare the true ||(Q+α I/T)^{-1}||≈1/0.526≈1.90 with Lemma 8's claimed upper bound 1/(1−10^{-6})≈1.000000001. If the claimed bound is smaller than the true norm, Lemma 8 is false for this system, confirming the proof gap. A repair should re-derive Lemma 10 with λmin(Q) in place of λmin(Γ) and check that the O(1/√T) term remains of the form k/√T.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-D (before Lemma 7) defines ξ=λmin(Γ) and states ξ≥λmin(Q). Lemma 8 then claims ||(Q+ΔQ+α I/T)^{-1}||≤1/(ξ−||ΔQ||−α/T). This requires Q⪰ξI, i.e. λmin(Q)≥ξ; the proof uses the opposite direction. The asserted direction is false in general. For y_t=a y_{t-1}+b u_{t-1}+e_t, u_t=v_t, with e_t,v_t iid N(0,I), p=2, Q=E[d_t d_t^T] has a y-y principal block (1/(1−a^2))[[1,a],[a,1]], whose minimum eigenvalue is 1/(1+|a|); with a=0.9 this is 0.526<1=λmin(Γ). By Cauchy interlacing, λmin(Q)≤0.526. Thus Lemma 8's inverse-norm bound can be smaller than the true norm, so the derivation of Lemma 10's k/√T term and Theorem 1's stated dependence on λmin(Γ) is unsupported. The O(1/√T) rate can likely be repaired by replacing ξ with λmin(Q)>0, a different system-dependent constant, but the written proof does not do this.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes REDAR, a two-step algorithm that fits a VARX model to closed-loop data and then applies balanced model reduction to obtain a state-space predictor. The main result (Theorem 1) is a non-asymptotic upper bound on the expected squared prediction error, decomposed into the optimal one-step prediction error, a model-order bias term, a model-reduction term, and a finite-data term of order O(1/√T). The proof proceeds through an error decomposition (Lemma 1), a bound on the Kalman filter tail (Lemma 2), concentration inequalities for Gaussian quadratic forms (Lemmas 4-6), and a delta-epsilon argument for the finite-data error (Lemmas 7-10). A second result (Theorem 2) gives a high-probability H∞ bound between the estimated model and the finite-horizon Kalman filter. The paper is self-contained and presents explicit rates, including exponential decay in the autoregressive order p and a 1/√T finite-data term.","tokens_in":14684,"tokens_out":11347,"duration_ms":98931,"significance":"If the bounds are valid, the paper fills a real gap: non-asymptotic guarantees for closed-loop identification are scarce. The algorithm is simple and practically motivated, and the analysis is self-contained, with no fitted constants or circular reasoning. The proof structure is clear, with standard concentration machinery and a clean decomposition of the prediction error into noise, model order, model reduction, and finite-data terms. The paper also provides a high-probability H∞ model-error bound that can be useful for robust control. The main limitation is that the finite-sample bounds require Gaussian noise; for non-Gaussian noise the exponential concentration inequalities are not established.","major_comments":[{"comment":"The proof uses ξ = λmin(Γ) as a lower bound for λmin(Q) when bounding ||Q^{-1}|| by 1/ξ, but the stated inequality ξ ≥ λmin(Q) gives the opposite direction. Since ξ ≥ λmin(Q), we have 1/ξ ≤ ||Q^{-1}||, so the claimed upper bounds in Lemmas 7 and 8 are not valid. A concrete counterexample is the AR(1) system y_t = a y_{t-1} + b u_{t-1} + e_t with u_t = v_t and e_t, v_t iid N(0,1), p=2. The y-y block of Q is (1/(1−a^2))[[1,a],[a,1]] with λmin = 1/(1+|a|), which for a=0.9 equals 0.526, while Γ=I_2 gives ξ=1. Thus Lemma 8's bound 1/(ξ−c2δ−α/T) can be smaller than the true norm ||(Q+ΔQ+α/T I)^{-1}||. Consequently the finite-data error bound in Lemma 10 and the k/√T term in Theorem 1 are not established as written. This is repairable by replacing ξ with λmin(Q) throughout the proof, but the constants in Theorem 1 and Lemma 10 would then depend on λmin(Q) rather than λmin(Γ), so the theorem statement must be revised.","section":"Theorem 1, model-order term"},{"comment":"The statement of Theorem 1 is inconsistent with the proof in Section III-B. Lemma 2 gives ||H_Tail||∞ ≤ Lρ^{p+1}/(1−ρ), and the bound on E[||y*_{t|−∞:t} − y_head_t||^2] is ||H_Tail||^2∞ ||z||^2_P. The theorem, however, is printed with a term Lρ^{p+1}/(1−ρ) ||z||_P, missing the squares on both the H∞ bound and the signal power. Please correct the theorem statement to match the proof.","section":"Theorem 1, model-order term"}],"minor_comments":[{"comment":"The eigenvalues of SR are indexed by i=1,...,T+p, but SR is a (T+p)n_z × (T+p)n_z matrix, so there are (T+p)n_z eigenvalues. The subsequent bound uses 2T for the number of eigenvalues; it should be 2Tn_z. This introduces a missing factor of n_z in the concentration constants, which does not change the exponential rate but should be corrected for dimensional consistency.","section":"Section III-C, Lemma 4"},{"comment":"There is a typo in the statement: the quadratic form should be (z_{t−k})_i (z_{t−ℓ})_j, not (z_{t−k})_i (z_{t−ℓ})_i, since the proof uses e_i e_j^T. Please fix the index.","section":"Section III-C, Lemma 5"},{"comment":"The Gaussian noise assumption is stated in Section II-B, but the discussion would benefit from explicitly noting that the finite-sample concentration bounds are not established for non-Gaussian noise, as this limits their applicability to general industrial data.","section":"Section II-B and Remark 1"},{"comment":"The probability bound in Theorem 2 uses δ with factor 4||J||^2∞, while Lemma 6 has a factor of 8||J||^2∞ in the linear term after the union bound. Checking the constants would improve the presentation; as printed, the δ expressions appear slightly inconsistent between the lemma and the theorem.","section":"Theorem 2 and Lemma 6"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is the first non-asymptotic guarantee for the REDAR closed-loop identification scheme, and the main result is plausible. The paper does well at laying out the decomposition, the tail bounds, and the high-probability H∞ version in Theorem 2. It is honest that the algorithm itself is not new, and it does not oversell the bound's tightness. The simulation is fine but no code is shipped, so that part is not independently checkable.\n\nThe real problem is in the finite-data error section. The stress-test note is right: ξ is defined as λmin(Γ), and the paper claims ξ ≥ λmin(Q). That is backwards. Q is the covariance of the regressor vector, and its minimum eigenvalue can be strictly smaller than λmin(Γ), even in a simple AR(1) with white excitation. Lemma 8's inverse-norm bound requires Q ⪰ ξI, so the proof as written does not support the k/√T term in Lemma 10. This is not a fatal flaw in the result—you can replace ξ with λmin(Q) and likely get the same rate with different constants—but the current text does not do that, and it is a load-bearing step for Theorem 1. There is also a minor typo in Lemma 10's d1 bound (ξ² appears where ξ⁴ is needed), which is easy to fix.\n\nWhat is genuinely new: the finite-sample analysis of an algorithm that had only asymptotic guarantees, and the careful handling of feedback correlation in the concentration bounds. The proof borrows standard machinery from Goldenshluger–Zeevi and Sarkar et al., but the adaptation to closed-loop data is nontrivial and the citation patterns are appropriate. The paper does not fit parameters to make the bound hold; the constants are explicit functions of the system and algorithm inputs.\n\nWho is this for: anyone working on finite-sample system identification, especially in closed-loop settings. The contribution is a solid subfield result, not a breakthrough. It deserves a serious referee, and the revision should be major: repair or restate Lemma 8/10 using λmin(Q), fix the k1 typo, and ideally release the simulation script. If those changes come through, Theorem 1 stands.","headline":"First finite-sample bound for closed-loop VARX + balanced reduction, but the proof has a real eigenvalue-direction gap that needs fixing before the rate is established.","tokens_in":15206,"tokens_out":2196,"would_cite":true,"duration_ms":24143,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B30","93E24","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves a finite-sample upper bound on the prediction error of a closed-loop identification algorithm that fits a long VARX model and then applies Hankel model reduction.","keywords":["closed-loop system identification","non-asymptotic bounds","VARX models","balanced model reduction","Kalman filter","generalization error","finite-sample analysis","H∞ norm"],"falsifier":"Run REDAR on a stable closed-loop LTI system with identical noise covariances but non-Gaussian heavy-tailed noises (e.g., Student-t or Laplace), and check on fresh data whether the mean squared prediction error decays like $O(1/\\sqrt{T})$ and stays below the claimed bound; a violation falsifies the central claim. Equivalent check: compare the empirical tail probabilities of the entries of $\\Delta Q$ and $\\Delta N$ under such noises with the exponential bound of Lemma 4.","tokens_in":14195,"feed_emoji":"📉","tokens_out":14113,"duration_ms":125764,"temperature":0.7,"pith_summary":"Closed-loop data breaks the standard bias-free guarantee of open-loop identification, because feedback correlates past noise with future inputs. This paper analyzes REDAR, an algorithm that first fits a high-order vector autoregressive model with exogenous inputs (VARX) to the closed-loop data and then reduces the model order by balanced model reduction with a specified H∞ error tolerance φ. The main theorem gives an explicit upper bound on the expected squared prediction error of the resulting one-step predictor, valid for any finite sample size T. If the bound is correct, an engineer can choose the VARX order p, the reduction tolerance φ, and the data length T to meet a target accuracy, and can compare closed-loop identification algorithms on equal terms. This matters because most existing non-asymptotic guarantees in system identification assume open-loop data, whereas practical data are often collected under feedback.","feed_headline":"Finite-data error bound proven for closed-loop system identification","feed_subtitle":"The theorem lets engineers trade data length, model order, and reduction error against prediction accuracy.","key_machinery":"The load-bearing object is the REDAR estimator, $\\hat y_t = H_R(q)z_t$, obtained from two steps: a ridge-regularized least-squares fit of $y_t$ on $p$ lags of the input/output vector (the VARX step), and balanced model reduction of the fitted transfer matrix with a guaranteed H∞ error at most $\\varphi$. The proof decomposes the prediction error into four parts — the optimal noise floor $\\|e\\|_P^2$, the tail of the steady-state Kalman filter operator $H_\\star$, the model-reduction error, and the finite-data least-squares error — and bounds each separately. The finite-data term is controlled by Gaussian concentration on the closed-loop regressor vector, namely a quadratic-form tail bound from the Gaussian moment generating function and a union bound over the entries of the empirical covariance and cross-covariance matrices.","core_discovery":"The paper's central claim is Theorem 1. Suppose the steady-state Kalman filter $H_\\star(z)$ is bounded by $L$ on or outside a circle of radius $\\rho<1$. Then for every $T\\ge T_0$ the REDAR predictor $\\hat y_t$ satisfies $\\mathbb{E}[\\|y_t-\\hat y_t\\|^2] \\le \\|e\\|_P^2 + \\frac{L\\rho^{p+1}}{1-\\rho}\\|z\\|_P^2 + 2\\varphi\\|z\\|_P^2 + \\frac{2k}{\\sqrt{T}}\\|z\\|_P^2$, where $\\|e\\|_P^2$ is the optimal one-step prediction error of the true model, $z$ stacks the measured input and output, and $k$ and $T_0$ depend on the dimensions, the regularization, the closed-loop H∞ norm, and the minimum noise eigenvalue. The first extra term is the bias from truncating the optimal filter to $p$ past lags, and it decays exponentially in $p$; the second term is the balanced-reduction tolerance; the third is the finite-data estimation error, decaying as $O(1/\\sqrt{T})$. The same analysis also gives a high-probability bound on the H∞ distance between the identified predictor and the finite-horizon Kalman filter.","pith_inferences":["The O(1/√T) rate matches known open-loop non-asymptotic identification rates, suggesting that feedback by itself may not worsen the sample complexity; instead, the closed-loop effect appears in constants such as the closed-loop H∞ norm rather than in the rate.","The proof structure suggests the same four-term decomposition would work for other reduction schemes, such as Hankel-norm approximation or truncation under a different error metric, whenever an analogous operator error bound is available.","A natural testable extension is to replace the Gaussian concentration argument with sub-Gaussian or martingale concentration; this would reveal whether the Gaussian assumption is a proof artifact or a real barrier.","Because the paper's own simulation shows the bound is loose, the practical constants could likely be improved by tightening the intermediate union-bound and norm-estimation steps without changing the algorithm."],"forward_implications":["For any chosen accuracy, an engineer can solve for a sufficient data length T, together with p and φ, from the closed-form bound, giving a sample-complexity certificate for closed-loop identification.","Because the model-order bias term decays exponentially in p, increasing the VARX order is a cheap way to shrink that term, and the 1/√T finite-data term dominates for large T.","The balanced-reduction tolerance φ enters linearly, so there is a quantitative trade-off between model simplification and prediction accuracy on unseen closed-loop data.","The high-probability H∞ bound on the error to the finite-horizon Kalman filter provides a quantified model uncertainty that could support robust control synthesis from finite closed-loop data."],"supporting_citations":[{"why":"Supplies the innovation-form model, the identifiability assumptions, and the notion of prediction error used throughout.","marker":"[1]"},{"why":"Provides the lemma, modified as Lemma 2, that bounds the decaying coefficients of the Kalman filter operator and yields the exponential-in-p term.","marker":"[4]"},{"why":"Motivates fitting high-order VARX models to closed-loop data to remove the feedback bias that spoils direct open-loop methods.","marker":"[16]"},{"why":"Establishes the asymptotic relation between the VARX-plus-balanced-reduction scheme and CCA, the algorithm whose finite-sample behavior this paper characterizes.","marker":"[21]"},{"why":"Guarantees that balanced reduction can deliver a model with H∞ error at most φ, giving the model-reduction term in the bound.","marker":"[22]"},{"why":"Supplies the Gaussian tail bound used to integrate the high-probability concentration events into the expectation bound.","marker":"[23]"}],"fun_headline_variants":["Finite-data error bounds for closed-loop identification","Non-asymptotic guarantees for closed-loop system ID","Closed-loop ID: data length vs prediction accuracy","Proven: finite-data error trade-off in closed-loop ID"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument's finite-sample probability bounds assume the process noise and the excitation noise are exactly Gaussian, so the regressor vector is Gaussian; if the noises are not Gaussian, the exponential concentration and the O(1/√T) term are not established by this paper.","fun_headline_variants_meta":{"raw":{"variants":["Finite-data error bounds for closed-loop identification","Non-asymptotic guarantees for closed-loop system ID","Closed-loop ID: data length vs prediction accuracy","Proven: finite-data error trade-off in closed-loop ID"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1581,"prompt_tokens":988,"completion_tokens":593,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":530}},"tokens_in":604,"tokens_out":593,"duration_ms":6462,"temperature":1.0,"reasoning_tokens":530,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:59:21.331007+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run REDAR on a stable closed-loop LTI system with identical noise covariances but non-Gaussian heavy-tailed noises (e.g., Student-t or Laplace), and check on fresh data whether the mean squared prediction error decays like $O(1/\\sqrt{T})$ and stays below the claimed bound; a violation falsifies the central claim. Equivalent check: compare the empirical tail probabilities of the entries of $\\Delta Q$ and $\\Delta N$ under such noises with the exponential bound of Lemma 4.","supporting_citations":[{"cited_title":"Ljung, System identification: theory for the user , 2nd ed","cited_arxiv_id":null,"evidence_quote":"Supplies the innovation-form model, the identifiability assumptions, and the notion of prediction error used throughout."},{"cited_title":"Nonasymptotic bounds for autore- gressive time series modeling,","cited_arxiv_id":null,"evidence_quote":"Provides the lemma, modified as Lemma 2, that bounds the decaying coefficients of the Kalman filter operator and yields the exponential-in-p term."},{"cited_title":"Subspace identification from closed loop data,","cited_arxiv_id":null,"evidence_quote":"Motivates fitting high-order VARX models to closed-loop data to remove the feedback bias that spoils direct open-loop methods."},{"cited_title":"The relation of the cca subspace method to a balanced reduction of an autoregressive model,","cited_arxiv_id":null,"evidence_quote":"Establishes the asymptotic relation between the VARX-plus-balanced-reduction scheme and CCA, the algorithm whose finite-sample behavior this paper characterizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Guarantees that balanced reduction can deliver a model with H∞ error at most φ, giving the model-reduction term in the bound."}],"review_version":1}