{"id":"3f42950b-aca2-459c-8f1e-7ad0cff76320","arxiv_id":"2602.10936","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A state-space trajectory predictor makes indirect data-driven predictive control a special case of linear model predictive control, unlocking classical MPC theory for it.","lead":"The paper groups many indirect data-driven predictive controllers under one framework and shows that one predictor makes them exactly equivalent to standard linear model predictive control. This could let data-driven methods borrow classical MPC guarantees, although the paper only proves the equivalence conditionally.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TPC–MPC equivalence is algebraically sound, but the 'inherits MPC guarantees' claim needs formal transfer conditions for estimated models; the paper itself defers these certificates.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the TPC-to-MPC reduction is exact for the identified model, but the inheritance of stability and recursive-feasibility guarantees for the real plant requires conditions on model fidelity or uncertainty that the paper does not provide. Theorem 3 and Corollary 4 are internal algebraic equivalences and appear correct; the problem is with the external claim that these guarantees 'therefore' apply. This is an overclaim rather than an internal inconsistency. The numerical evidence, while suggestive, is a single second-order example without error bars and without code, so it does not close the gap. The Limitations section's admission that certificates 'could be formalized' supports the reader's conditional assessment. No stronger objection is warranted: the unifying framework and the data-efficiency of the state-space predictor are real contributions, and the algebraic equivalence is a solid basis for future robust/stochastic extensions. The verdict should remain CONDITIONAL, as the reader recommended, pending a precise statement of the conditions under which the MPC guarantees transfer to the true closed loop and more robust numerical evidence.","tokens_in":17461,"tokens_out":12747,"duration_ms":120161,"concrete_test":"Re-run the §V double-integrator experiment for d=30, 50, 100 closed-loop training samples, with 1000 Monte Carlo runs. For each run, estimate A, B, C, D via Algorithm 2 and simulate the resulting TPC/MPC control law on the true stochastic plant for 400 steps. Record the empirical frequency of instability (e.g., |y| exceeding a large threshold) and of recursive infeasibility. Then, for each estimated model, check whether the standard MPC terminal-cost/terminal-constraint or sufficiently-long-horizon conditions [18] hold. If the estimated-model conditions hold in a run where the true closed loop is unstable, the 'inherits guarantees' claim fails in the sense that matters for control; if no such run appears, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The algebraic core is sound: Theorem 3 and Corollary 4 show that TPC with the state-space predictor is exactly an MPC problem with the A, B, C, D matrices obtained from the one-step ARX fit in Algorithm 2. This is a genuine contribution. The load-bearing gap is the subsequent claim that TPC 'inherits the theoretical guarantees that linear MPC enjoys, such as conditions for stability and recursive feasibility' (end of §IV-A). The equivalence is only with respect to the estimated model. Algorithm 2 fits C and D by least squares from finitely many noisy trajectories; the §V plant has both process noise w and measurement noise v, so the estimated finite-order ARX model is at best an approximation of the true stochastic plant. Classical MPC stability/feasibility theorems [18,19] assume the prediction model is exact, or that a known uncertainty set is used. No theorem in the paper states conditions under which stability/feasibility of the estimated-model MPC transfers to the true closed loop. The paper's own Limitations section says 'Certificates of stability and recursive feasibility could be formalized by reducing TPC to MPC,' which concedes that these certificates are not actually proven. Thus the inheritance claim is an overstatement until the required model-accuracy/uncertainty conditions are stated and proved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines trajectory predictive control (TPC) as a class of indirect data-driven predictive control methods whose output predictor is linear in past inputs/outputs and future inputs, and shows that several DDPC variants (SPC, γ-DDPC, causal-γ-DDPC, transient predictive control, closed-loop SPC) are instances of TPC with specific predictors. The main theoretical contribution is a new 'state-space predictor' constructed from a one-step ARX fit, which the authors claim makes TPC exactly equivalent to linear MPC with a state-space model whose state is the recent input/output history. Corollary 4 states that the TPC and MPC optimizations have identical optimal solutions. The paper further claims that TPC therefore inherits the stability and recursive-feasibility guarantees of linear MPC. Numerical experiments on a double-integrator example compare the predictors and show that the state-space predictor is competitive, particularly with small training datasets.","tokens_in":17769,"tokens_out":15227,"duration_ms":126802,"significance":"If the central equivalence is correct, the paper provides a clean bridge between indirect DDPC and conventional linear MPC, which would be a valuable conceptual contribution: it would allow the application of a mature body of MPC theory to a broad family of data-driven controllers. The state-space predictor also appears to be genuinely simpler and more data-efficient than other unstructured predictors, and the unification of earlier DDPC variants is useful. The paper's algebraic derivations and the equivalence theorem are the strongest parts; the numerical experiments are reproducible in principle. However, the claimed inheritance of MPC guarantees is overstated for the real closed loop unless model-error conditions are stated and proved.","major_comments":[{"comment":"The sentence 'TPC with the state-space predictor therefore inherits the theoretical guarantees that linear MPC enjoys, such as conditions for stability and recursive feasibility' is not supported by the results. The equivalence in Theorem 3 and Corollary 4 holds for the identified model, whose matrices A,B,C,D are least-squares estimates from finite noisy data (Algorithm 2). Classical MPC stability and feasibility theorems [18,19] assume the prediction model is exact or that a known uncertainty set is used. The paper's own Limitations section states that 'Certificates of stability and recursive feasibility could be formalized by reducing TPC to MPC', which concedes that these certificates are not actually proved. The claim should be qualified to the design model, or the authors should prove transfer conditions (e.g., robust MPC with an uncertainty set, or a stochastic framework) under wh","section":"§IV-A, after Corollary 4; §VI"},{"comment":"There is a notation conflict that obscures the central theorem. Algorithm 2 first defines A and B from (18), then overwrites them via 'A = A - KC, B = B - KD'. Theorem 3 says it uses A,B 'according to Algorithm 2', but the MPC problem (24) explicitly uses 'A, B, C, and D defined by (18)', i.e., the original matrices. The proof's final sentence also refers to 'A,B,K defined as in (18)', contradicting the theorem statement. This makes it impossible for a reader to verify Corollary 4 without reconstructing the intended algebra. Use distinct symbols (e.g., Ā = A - KC, B̄ = B - KD) for the intermediate quantities and state explicitly whether (23) and (24) use A,B from (18) or the overwritten Ā,B̄.","section":"Theorem 3, Algorithm 2, Eq. (24)"},{"comment":"The dimension statement before (16) says C ∈ R^{h ny × mnz} and D ∈ R^{h ny × h nu}, but y(t) is ny-dimensional and in Lemma 2 C = P1 ∈ R^{ny × mnz}, D = F11 ∈ R^{ny × nu}. This appears to be a typo, but it is confusing in a paper whose core is a state-space equivalence. Correct the dimensions.","section":"Eq. (16) and Algorithm 2"}],"minor_comments":[{"comment":"The phrase 'For mA' appears to be a typo; should read 'For A, B, C, D, K, P, F, and Φy constructed according to Algorithm 2'.","section":"Theorem 3 statement"},{"comment":"In the top-left panel, the d=30 bars for predictors other than the state-space predictor extend off the vertical axis; the text mentions values up to 0.38. It would help to show these values or use a broken axis, since the comparison at small d is a key empirical claim.","section":"§V-A, Fig. 2"},{"comment":"The statement that the AIC 'usually suggests a memory of one or two time steps and rarely more than three' is a qualitative summary; consider reporting the selected m values or a small table for reproducibility.","section":"§V-A, AIC discussion"},{"comment":"The first paragraph repeats the unqualified inheritance claim from §IV-A; even if the earlier text is revised, ensure the summary and conclusion are consistent with the qualified statement.","section":"§VI"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central algebraic contribution appears sound and is likely to be of interest to the DDPC community. The main issue is that the abstract, introduction, and §IV-A overstate the consequences of the equivalence by claiming that TPC 'inherits' MPC stability and recursive-feasibility guarantees without addressing model error from finite noisy data. The Limitations section itself concedes that certificates are not formalized. I recommend major revision to correct the claims and to clarify the notation in Theorem 3 and Algorithm 2. The numerical evidence is narrow (single double-integrator, unconstrained), but that alone would not block publication if the theoretical claims are accurately stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a real result: it defines TPC as a unifying class of indirect DDPC, introduces a state-space predictor whose state is the recent input/output history, and proves that TPC with that predictor is exactly linear MPC on the identified model. I checked Lemma 2, Theorem 3, and Corollary 4; the algebra is careful and correct. The equivalence is not just a formality—it means every existing indirect DDPC method in the paper's taxonomy, plus the new predictor, can be viewed through one lens. The parameter-count comparison and the data-efficiency argument for the state-space predictor are concrete and useful. The paper also deserves credit for showing that 'relax-and-regularize' TPC can exhibit optimism bias, a useful caution.\n\nThe soft spot is the inheritance claim. After Corollary 4 the paper says TPC 'inherits the theoretical guarantees that linear MPC enjoys, such as conditions for stability and recursive feasibility.' That is an overstatement. The equivalence is to MPC with A,B,C,D estimated from finitely many noisy trajectories. Classical linear MPC stability/recursive-feasibility theorems assume the model is exact or that a known uncertainty set is used. No theorem in the paper states conditions under which guarantees transfer from the estimated model to the true plant. The Limitations section itself says 'Certificates of stability and recursive feasibility could be formalized,' which concedes they are not proven. That is not a fatal flaw—the algebraic contribution stands on its own—but the manuscript should either soften the inheritance language or state the model-accuracy/uncertainty conditions required.\n\nThe numerical evidence is modest: one double-integrator example from the prior literature, Monte Carlo over 1,000 runs, but no error bars or variance information in the figures, and no code or data shipped. For a paper whose main practical claim is data efficiency, a second system and some uncertainty quantification would strengthen the case. That said, the experiments do show the state-space predictor's advantage at small sample sizes, which is consistent with the parameter-count argument.\n\nThe citation pattern looks fair and current; the relevant SPC, γ-DDPC, causal-γ-DDPC, transient, and fixed-length papers are all there, and the prior work on bias is engaged honestly. The authors do not overclaim novelty; they acknowledge the ARX/state-space construction is classical.\n\nWho this is for: researchers working on data-driven predictive control, especially those comparing direct and indirect methods. It gives the indirect camp a clean theoretical handle and a practical reason to prefer the state-space predictor for small datasets. It deserves a serious referee—send it out. The main request to the authors should be to either prove or clearly scope the guarantee-inheritance claim, and to consider releasing code/data.","headline":"A genuinely useful unification: TPC with the state-space predictor is exactly linear MPC on the identified model; the inheritance claim for MPC guarantees is the one real overreach.","tokens_in":18253,"tokens_out":2058,"would_cite":true,"duration_ms":19505,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Trajectory predictive control with the state-space predictor is exactly linear MPC, so data-driven controllers inherit MPC's stability and feasibility theory.","keywords":["trajectory predictive control","data-driven predictive control","state-space predictor","model predictive control","system identification","output feedback","recursive feasibility","subspace predictive control"],"falsifier":"Simulate TPC with the state-space predictor on a constrained LTI plant using training data gathered under a fixed controller with weak excitation; if the closed loop violates constraints or diverges even though the equivalent MPC (24) is recursively feasible for the identified model, the transfer of guarantees to the real plant is falsified.","tokens_in":17339,"feed_emoji":"⚙️","tokens_out":6655,"duration_ms":59812,"temperature":0.7,"pith_summary":"This paper defines trajectory predictive control (TPC) as a family of indirect data-driven predictive controllers—encompassing subspace predictive control, γ-DDPC, transient predictive control, and related methods—in which future outputs are a linear function of recent inputs/outputs and planned inputs. It introduces the state-space predictor, which represents the same prediction using an LTI state-space model whose state is the recent input/output history. The paper proves that TPC with this predictor is a special case of linear MPC: the two optimization problems have identical optimal solutions. If correct, the entire established theory of linear MPC—stability conditions, recursive feasibility, and robust, stochastic, and scenario extensions—applies to this data-driven controller. Numerical experiments show the state-space predictor is the most data-efficient and performs best on small training sets.","feed_headline":"Data-driven controller proves identical to classic MPC","feed_subtitle":"With the new state-space predictor, the two controllers share the same optimal solutions and the same guarantees.","key_machinery":"The state-space predictor: a trajectory predictor built from a one-step ARX model, lifted into an LTI state-space model (16) with the recent input/output history zp(t) as the state. The construction (Lemma 2, Algorithm 2) defines A, B, C, D, K so that iterative application of the one-step predictor exactly matches the state-space model; this bijection is what makes TPC an instance of MPC and carries the whole argument.","core_discovery":"The central claim is that TPC with the state-space predictor is not an approximation of MPC but an exact special case. Theorem 3 establishes an equivalence between the trajectory predictor (future outputs as linear functions of recent history and planned inputs) and an innovations-form LTI state-space model whose state is the recent input/output history. Corollary 4 states that the TPC optimization (4) and the equivalent MPC optimization (24) have identical optimal solutions. The paper therefore argues that all theoretical guarantees of linear MPC—conditions for stability, recursive feasibility, and extensions to robust, stochastic, and scenario formulations—apply to this data-driven control","pith_inferences":["An implication left implicit is that any existing linear MPC implementation can be repurposed as a TPC controller by fitting the one-step ARX model, so data-driven control can immediately use mature MPC software, tuning rules, and warm-starting.","The parameter-count advantage suggests the state-space predictor should be the default choice for adaptive or online learning settings where training data accumulate slowly; a testable prediction is that its edge over other predictors grows with system dimension and horizon length.","Because the equivalence is algebraic, the hard part of applying the inherited guarantees is the quality of the one-step ARX fit; the paper's own limitations section notes that formal stability and feasibility certificates are not actually derived, so a follow-up could quantify how model error degrades those guarantees.","The same state-space structure could be extended to nonlinear or time-varying plants by replacing the linear ARX fit with a nonlinear predictor while keeping the MPC-equivalent optimization form."],"forward_implications":["TPC with the state-space predictor inherits linear MPC's stability conditions and recursive feasibility guarantees.","The same controller can be extended to robust, stochastic, or scenario MPC using established methods.","No state estimator is needed: the state is the perfectly observed recent input/output history, giving output-feedback control without estimation.","The state-space predictor requires the fewest parameters and fewest training examples of the predictors compared, making it the most data-efficient option.","Since TPC is equivalent to MPC, other linear control design tools—H2/H∞ synthesis, system-level synthesis—become available to data-driven controllers."],"fun_headline_variants":["Data-driven control equals classic MPC, exactly","State-space predictor makes data-driven control inherit MPC theory","New predictor shows data-driven MPC is exact linear MPC","Data-driven predictive control matches MPC guarantees","State-space predictor unifies data-driven and linear MPC"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the one-step ARX model fitted from finite, noisy training data is an accurate description of the true plant, since the equivalence holds for the identified model and stability/feasibility guarantees only transfer to the real closed loop if that model is good.","fun_headline_variants_meta":{"raw":{"variants":["Data-driven control equals classic MPC, exactly","State-space predictor makes data-driven control inherit MPC theory","New predictor shows data-driven MPC is exact linear MPC","Data-driven predictive control matches MPC guarantees","State-space predictor unifies data-driven and linear MPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000372,"raw_usage":{"total_tokens":1744,"prompt_tokens":577,"completion_tokens":1167,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":321,"completion_tokens_details":{"reasoning_tokens":1096}},"tokens_in":321,"tokens_out":1167,"duration_ms":7944,"temperature":1.0,"reasoning_tokens":1096,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T00:55:42.737107+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate TPC with the state-space predictor on a constrained LTI plant using training data gathered under a fixed controller with weak excitation; if the closed loop violates constraints or diverges even though the equivalent MPC (24) is recursively feasible for the identified model, the transfer of guarantees to the real plant is falsified.","supporting_citations":[],"review_version":1}