{"id":"bed31cf5-4fa7-43bb-b7be-c73f420a766d","arxiv_id":"2607.28123","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Projection-regularized indirect DPC reduces EIV prediction error versus SPC, adapts via covariance blending, and yields high-probability recursive feasibility and ISpS under martingale noise bounds.","lead":"PRPC is an indirect data-driven controller that regularizes the fundamental-lemma weight, collapses it to fixed-size covariances, and adapts online for noisy and time-varying plants. It matters because it aims to keep subspace-style speed while adding finite-sample safety under closed-loop noise.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection beyond the reader's L_\\Theta point; the EIV strict-reduction claim is only weakly supported by the derivative argument and modest numerics.","rationale":"The central architecture (regularized projection \to exact covariance collapse \to SPC recovery as \\lambda\to0, Prop. 1/Thm. 1) is algebraically sound and the SNM + tightened-constraint wrapper is standard robust-MPC reasoning once Assumption 1 and terminal ingredients are granted. The reader already flagged the load-bearing practical hole: L_\\Theta (and c_w) must be supplied a priori and are not identified from data, so the high-probability recursive feasibility/ISpS statements are conditional certificates rather than fully data-driven guarantees. My secondary observation on Prop. 4 does not overturn that assessment; it only notes that the EIV half of the strongest claim is modestly evidenced. No hidden inconsistency in the KKT collapse, the martingale bound derivation, or the recursive-feasibility induction was found. Hence the verdict remains CONDITIONAL at high confidence, with no adjustment required.","tokens_in":21191,"tokens_out":626,"duration_ms":13487,"concrete_test":"Re-run the Fig. 3/4 Monte-Carlo MSE-ratio experiment on the Boeing 747 LTI plant at the paper’s most favorable reported corner (M/T_h\\approx1.3, \\sigma_w/\\sigma_v=10, same horizons) over ≥50 independent noise realizations; report the fraction of trials in which min_\\lambda MSE(PRPC) < MSE(SPC) and the median reduction. If that fraction is appreciably below 1 or the median reduction stays single-digit, the “strictly reduces” language in the strongest claim should be softened to “can reduce.”","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (known uniform L_\\Theta and c_w in Assumption 1 feeding r_k, Theorems 2–4) is correctly identified and is the dominant practical gap for unknown LTV plants. Independently, the other half of the strongest claim—that PRPC strictly reduces predictor MSE vs SPC under process noise for some \\lambda*>0 (Prop. 4)—rests on a first-order derivative argument at \\lambda=0 (Appendix D) that is only qualitative: it requires the residual-variance term to dominate a bounded bias derivative without an explicit, checkable condition on (\\sigma_w, M, spectrum of \\Sigma_\\Phi0). The accompanying Monte-Carlo evidence shows only modest gains (~7% at M/T_h\\approx1.3 under mild noise; ~20% at \\sigma_w/\\sigma_v=10 and M/T_h\\approx2.6), so “strictly reduces” is true in a weak sense but not transformative. Neither issue is an internal contradiction; both are already reflected in the CONDITIONAL verdict.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Projection-Regularized Predictive Control (PRPC), an indirect data-driven predictive controller that retains a Tikhonov-regularized fundamental-lemma weight g subject to a hard future-input constraint and a soft past-data match. The KKT system is analytically collapsed onto fixed-dimension sample covariances (Proposition 1), recovering unregularized SPC as λ→0 (Theorem 1) while remaining well-posed under rank deficiency. A bias–variance analysis argues that under process noise (errors-in-variables) there exists λ*>0 strictly reducing predictor MSE relative to SPC (Propositions 3–4), and that PRPC stays defined when M<Th (Proposition 5). These properties support an adaptive sliding-window scheme via convex blending of offline and exponentially forgotten online covariances. Closed-loop prediction error is bounded uniformly in time by a self-normalized martingale radius plus a deterministic mismatch term (Theorem 2); embedding the radius into dynamically tightened constraints yields high-probability recursive feasibility and ISpS (Theorems 3–4). LTI (Boeing 747) and polytopic LTV simulations illustrate conditioning gains, modest MSE reductions, tracking, and empirical coverage of the bound.","tokens_in":21510,"tokens_out":1836,"duration_ms":44513,"significance":"If the claims hold, the work cleanly unifies computational advantages of subspace predictors with regularization-based robustness and supplies finite-sample, closed-loop-valid safety certificates for adaptive LTV data-driven MPC—addressing a recognized gap between DeePC-style methods and classical robust tube MPC under correlated online data. Strengths that deserve explicit credit: the covariance collapse (Proposition 1) is algebraically exact and numerically verified to machine precision against SPC (Fig. 1); the PRPC–SPC limit (Theorem 1) and rank-deficient well-posedness (Proposition 5) are sharp; the multi-output SNM argument and standard robust-MPC feasibility/ISpS scaffolding are carefully adapted to the blended-covariance predictor. The free design knobs (λ, γ, ρ_f) are standard and the circularity risk is low. The main practical caveat is that the safety radius inherits a known uniform mismatch bound L_Θ and noise proxy c_w, which limits turnkey use on fully unknown LTV plants; the EIV “strict reduction” claim is real but modest in the reported numerics.","major_comments":[{"comment":"Proposition 4 and Appendix D assert existence of λ*>0 with MSE(PRPC)<MSE(SPC) under EIV by showing ∂MSE/∂λ|λ=0<0. The argument is only qualitative: it requires the residual-variance term to dominate a bounded bias derivative “whenever σ_ζ² > θ_i² a_i(1−a_i)/s_i⁰ for the significant directions,” without an explicit, checkable condition in terms of (σ_w/σ_v, M, spectrum of Σ_Φ0). The Monte-Carlo evidence (Figs. 2–4) shows only modest gains (~7% at M/Th≈1.3 under mild process noise; ~20% at σ_w/σ_v=10 and M/Th≈2.6). Please either (i) supply a sufficient condition that can be verified from the offline covariances and noise proxies, or (ii) temper the abstract/introduction language (“strictly reduces,” “transformative advantages,” “up to a 7% reduction”) to match the derivative-level guarantee and the reported effect sizes.","section":"§4.2, Proposition 4, Appendix D; Abstract; §8"},{"comment":"Assumption 1 requires a known uniform operator-norm bound L_Θ on the entire parametric mismatch ΔΘ(k) for all k, together with a known conditional sub-Gaussian proxy c_w. Both enter the radius r_k (Theorem 2) and therefore the tightened sets Y_tight(k), the terminal set inclusion, recursive feasibility (Theorem 3), and the ISpS ultimate ball O(r_∞) (Theorem 4). For a truly unknown LTV plant this bound is not identified from data in the paper. The LTV experiments use a known polytopic hull, so L_Θ is available by construction; the manuscript should state clearly how L_Θ (and c_w) are obtained in practice, discuss sensitivity of the closed-loop certificates when L_Θ is over-/under-estimated, or provide a data-driven outer bound. Without this, the high-probability safety claim for adaptive LTV operation rests on an a-priori quantity that the method does not produce.","section":"§6.1 Assumption 1; Theorem 2; Theorems 3–4; §7.2"},{"comment":"Lemma 2 and the adaptive scheme rely on a strictly positive offline anchor weight γ∈(0,1) to guarantee Σ_act_pp(k) ⪰ γ Σ_off_pp ≻ 0 and hence uniform invertibility of W_p and S. The ablation (Fig. 9) shows that un-anchored recursion (γ=0) matches average-case cost under persistent excitation, so the anchor’s value is purely structural. Please quantify how large γ must be relative to the offline PE margin and the forgetting factor ρ_f so that the spectral floor remains useful (i.e., r_∞ does not become vacuous), and discuss the bias introduced into the predictor when the offline model is far from the current LTV vertex. This is load-bearing for the claim that the adaptive controller remains well-posed and non-conservative under loss of online excitation.","section":"§5, Lemma 2; §7.2 ablation; Lemma 3"}],"minor_comments":[{"comment":"Fig. 6 caption and text refer to “γ-DPC” and “GDPC” somewhat interchangeably; please align nomenclature with the cited GDPC reference and define every baseline acronym at first use in §7.1.","section":"§7.1, Fig. 6"},{"comment":"In (13) the factor M that “normalizes the update to the scale of the offline data” is not defined in the notation section; clarify whether it is the offline column count or a free gain.","section":"§5, Eq. (13)"},{"comment":"Theorem 2 writes β_k with √(1+λ ϱ(V(k)⁻¹)); a one-line justification that this factor absorbs the ridge initialization bias (beyond the sketch in Appendix E) would help readers reproduce the constant.","section":"Theorem 2; Appendix E"},{"comment":"Typos / style: “F undamental Lemma”, “V ariance”, “T uning”, “F easibility” (spurious spaces after capitals in several headings); “HV AC”; “minλ MSE” formatting in Fig. 3; “plim” vs. “p-lim” consistency in Appendix C.","section":"Throughout; §2–§4 headings"},{"comment":"The conclusion claims “reduces the dominant offline computational load by over 50% compared to unregularized SPC”; SPC’s classical cost is also O(T_h² M) or better with covariance form—please cite the exact baseline complexity being compared or soften the percentage claim.","section":"§8"},{"comment":"Table 1 / Table 2 matrices are hard to parse in text form; if the journal allows, provide them as supplementary files or increase column formatting for readability.","section":"§7.1–7.2"}],"recommendation":"major_revision","confidential_remarks":"The core algebra (covariance collapse, SPC limit, rank-deficient well-posedness) is solid and the SNM+robust-MPC scaffolding is competent. The two load-bearing soft spots are (i) the qualitative Prop. 4 / modest numerics versus strong abstract wording, and (ii) the a-priori L_Θ required for the LTV safety theorems. Both are fixable by tempered claims, an explicit sufficient condition or honest limitation paragraph, and a short discussion of how to set L_Θ/c_w; I would not reject on novelty or correctness grounds. Fit for a control journal that values data-driven MPC with certificates is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit here is the packaging, not a new paradigm. They keep the latent g from the fundamental lemma, Tikhonov-regularize the soft initial-condition match while hard-enforcing U_f g = u_N, then algebraically collapse everything onto fixed-size covariances (Prop. 1). That recovers SPC as λ→0 (Thm. 1, Fig. 1 to machine precision), stays well-posed when M < T_h, and gives a constant-cost online QP plus a cheap recursive update. The convex offline anchor (γ) is a simple, honest fix for loss of PE in closed loop. That part is clean and worth having on the shelf.\n\nThe EIV story is directionally right: under process noise the unregularized SPC map is inconsistent, and a first-order derivative argument plus Monte Carlo shows some λ*>0 can cut predictor MSE. Gains are real but modest—about 7% in scarce data under mild noise, up to ~20% when process noise dominates. Do not oversell “strictly reduces” as transformative; the Appendix D argument is qualitative and needs the residual-variance term to dominate without a sharp checkable condition. Numerics match that reading.\n\nStability is the usual robust-MPC wrapper with a vector SNM radius and dynamic tightening. Theorems 2–4 are carefully written, but they lean on a known uniform mismatch bound L_Θ, a known sub-Gaussian proxy c_w, and classical terminal ingredients. For a truly unknown LTV plant those are design inputs, not identified quantities. The paper is upfront enough about the setup; just do not treat the high-probability ISpS claim as free lunch.\n\nCitations sit in the right neighborhood (DeePC, SPC, GDPC, SMM, regularized and adaptive variants). No circularity games. No code shipped, which is a practical minus for reproduction.\n\nWho benefits: people already building indirect/adaptive DPC who want a fixed-dimension predictor that does not collapse in short windows, plus a finite-sample tube they can plug in if they are willing to supply L_Θ and terminals. Not required reading outside that lane.\n\nI would send it to peer review. Expect referees to press on L_Θ, the strength of Prop. 4, and code. Worth engaging if you work this stack; skip if you only want foundational behavioral results.","headline":"Clean incremental package: regularized-g collapsed to fixed covariances, modest EIV gains, and a standard SNM tube for adaptive SPC—solid systems work, not a foundational reset.","tokens_in":22200,"tokens_out":599,"would_cite":true,"duration_ms":22439,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A regularized projection of the fundamental-lemma weights cuts prediction error under process noise and keeps adaptive data-driven control recursively feasible with high probability.","keywords":["data-driven control","predictive control","regularization","adaptive control","linear time-varying systems","subspace methods","martingale concentration"],"falsifier":"On a process-noise plant with scarce data, check whether cross-validated PRPC yields lower predictor MSE than unregularized SPC; under slow LTV drift, check whether the mismatch-augmented radius achieves the claimed coverage while the naive noise-only radius does not. If the MSE ratio never falls below one, or coverage fails the target, the central claims fail.","tokens_in":22008,"feed_emoji":"⚙️","tokens_out":898,"duration_ms":36489,"temperature":0.7,"pith_summary":"Indirect data-driven predictive control is usually brittle when process noise corrupts the regressors or when sliding data windows lose rank. This paper keeps the latent weight vector from the fundamental lemma, applies a Tikhonov-regularized projection that matches past data softly and future inputs strictly, and shows the whole construction collapses analytically into fixed-dimension covariance matrices. The resulting Projection-Regularized Predictive Control (PRPC) strictly lowers mean-squared prediction error relative to unregularized subspace methods under errors-in-variables noise and remains well-defined when the sample size falls below the regressor dimension. The same covariances support a sliding-window adaptive controller for linear time-varying plants; closed-loop correlations are handled by a uniform-in-time finite-sample radius from vector-valued self-normalized martingales. Embedding that radius into dynamically tightened constraints yields robust recursive feasibility and input-to-state practical stability at a user-chosen confidence level, while the online problem stays a low-dimensional quadratic program independent of data length.","feed_headline":"Regularized projection cuts error under process noise","feed_subtitle":"Fixed-size covariances keep adaptive LTV control fast and high-probability safe","key_machinery":"Analytical covariance collapse of the regularized projection: the high-dimensional weight vector is eliminated offline into fixed-size matrices (Σpp, Σup, Σyp, …), so the online controller is a low-dimensional QP whose latency does not grow with data length; the same objects carry recursive blending and the self-normalized martingale radius that tightens the constraints.","core_discovery":"Under process noise and rank-deficient windows, the regularized fundamental-lemma projection, condensed into empirical cross-covariances, produces a multi-step predictor whose mean-squared error is strictly smaller than unregularized subspace predictive control for some positive regularization, and whose uniform statistical uncertainty radius, once used to tighten constraints, guarantees recursive feasibility and input-to-state practical stability with probability at least 1−δ.","pith_inferences":["The same covariance collapse could fix the online dimension of other regularized Hankel or instrumental-variable predictors.","If the drift bound were estimated from residual growth instead of supplied a priori, the guarantees would become fully data-driven for unknown time variation.","The bias–variance analysis implies the largest practical gains appear exactly in early commissioning and sudden plant-change regimes where data are scarce and process noise dominates."],"forward_implications":["Online controller latency is decoupled from the length of the collected trajectory.","Sliding-window LTV adaptation stays mathematically well-posed even when feedback kills persistent excitation.","Constraint tightening can be driven by a finite-sample, uniform-in-time statistical radius rather than purely worst-case tubes.","Under pure measurement noise and abundant data the method recovers ordinary subspace predictive control, so no extra conservatism is forced.","Rank-deficient windows no longer make the predictor undefined."],"fun_headline_variants":["Projection regularization cuts MSE under noise and rank deficiency","Fixed-dim covariances speed adaptive LTV predictive control","Martingale radius tightens constraints for high-prob ISpS","PRPC beats unregularized subspace predictors under process noise","Regularized fundamental-lemma projection lowers multi-step error"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Safety needs a known uniform bound on how far the true multi-step map can drift from the nominal predictor at every time, plus a known noise size; neither is learned from data in the paper.","fun_headline_variants_meta":{"raw":{"variants":["Projection regularization cuts MSE under noise and rank deficiency","Fixed-dim covariances speed adaptive LTV predictive control","Martingale radius tightens constraints for high-prob ISpS","PRPC beats unregularized subspace predictors under process noise","Regularized fundamental-lemma projection lowers multi-step error"]},"model":"grok-4.5","effort":"low","cost_usd":0.004739,"raw_usage":{"total_tokens":1343,"prompt_tokens":721,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":47388000,"prompt_tokens_details":{"text_tokens":721,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":558,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":721,"tokens_out":64,"duration_ms":8712,"temperature":1.0,"reasoning_tokens":558,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T17:09:31.051486+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a process-noise plant with scarce data, check whether cross-validated PRPC yields lower predictor MSE than unregularized SPC; under slow LTV drift, check whether the mismatch-augmented radius achieves the claimed coverage while the naive noise-only radius does not. If the MSE ratio never falls below one, or coverage fails the target, the central claims fail.","supporting_citations":[],"review_version":1}