{"id":"5034268e-6b13-4bd8-9887-d3c86f1f7e35","arxiv_id":"2509.05167","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Splitting quantum optimal control into repeated short-horizon MPC problems, with terminal constraints or optimized setpoints, gives faster and more robust qubit state preparation in simulations.","lead":"Model predictive control is applied to quantum systems by splitting a long optimal-control problem into short receding-horizon subproblems. The paper reports faster qubit state preparation and improved robustness to model mismatch on simulated examples, with a conditional stability guarantee for one scheme.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem IV.1's Assumption IV.1.1 is false for the pure-state representation used in the numerical examples: global phase makes ||X−X_ref||^2>0 while 1−F=0, so the Lyapunov argument in (22)-(23) does not apply as stated.","rationale":"The paper's headline theoretical contribution is Theorem IV.1, which converts a quadratic cost-controllability bound into exponential infidelity decay. The proof is a standard MPC Lyapunov argument, and it is structurally sound IF Assumptions IV.1.1 and IV.1.5 hold in the same state space where the dynamics are written and where the numerics are run. My central concern is that the state space is not fixed consistently. In Section II A, pure-state evolution is written as X(t)=|ψ(t)>; in Appendix A, the implementation decomposes |ψ>=a+ib and optimizes over a,b. On this state space, Assumption IV.1.1 is demonstrably false because global phase changes the norm but not the fidelity (6). The theorem therefore cannot be applied to the numerical examples in Section VII, which are all pure-state state-vector problems. This is more severe than 'unverified': it is an internal inconsistency in the standing assumptions. The reader's chosen weakest assumption, IV.1.5, is certainly unverified, and IV.1.4 is violated by the |+> -> |-> experiment, but those are limitations/scope issues; they do not make the proof invalid for the stated pure-state setup. The phase issue does. A fix is available: formulate the theory for density matrices or projective states, where F=1 implies X=X_ref, and re-run the examples in that representation. Because the framework is modular and the numerical demonstrations are plausible, I do not think the paper should be rejected; it should be conditionally accepted with a required revision that either restricts the stability theorem to density-matrix/projective systems or develops a phase-invariant Lyapunov proof.","tokens_in":17165,"tokens_out":9405,"duration_ms":111804,"concrete_test":"Check Assumption IV.1.1 literally in the representation used by the paper: take X1=|0>, X2=e^{iπ/2}|0>, and compute ||X1−X2||^2 = |1−e^{iπ/2}|^2 = 2 while 1−F = 1−|⟨0|e^{iπ/2}|0⟩|^2 = 0. This violates (17) for any finite c_1. To settle whether the theorem is salvageable, re-run the TEC example from Section VII B with X(t) replaced by the rank-one projector ρ(t)=|ψ(t)⟩⟨ψ(t)|, using the same controls and a state-space norm for which (17) holds; if the exponential infidelity bound is recovered for ρ, the theorem can be restated in the density-matrix formalism, but the paper must make that restriction explicit.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing concern is not the unverified quadratic controllability bound (IV.1.5) but the standing state representation. Section II A defines X(t)=|ψ(t)> for pure states, and Section VII/Appendix A solve the optimization in complex vector form (a+ib). Assumption IV.1.1 asserts ||X1−X2||^2 ≤ c_1(1−F(X1,X2)). For pure-state vectors with F from (6), choose X2=e^{iθ}X1. Then F=1, so the RHS is 0, while ||X1−X2||^2=|1−e^{iθ}|^2>0; no finite c_1 can satisfy (17). The proof of Theorem IV.1 uses (17) to turn J*(X_t)≤c_u||X_t−X_ref||^2 into J*(X_t)≤c_u c_1(1−F(X_t,X_ref)) (Eq. 22), and then obtains the contraction (23); without (17) this step collapses. If X is instead a density matrix (so global phase is quotiented), (17) can hold with a suitable norm, but then the paper's dynamics (4), terminal constraint (16d), and the numerical implementation in Appendix A are all for state vectors and must be re-expressed for ρ. As written, Theorem IV.1 does not cover the pure-state examples that constitute the numerical evidence. The paper can be repaired by working projectively or density-matrix-valued, but the current text is internally inconsistent.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a modular model predictive control framework for quantum optimal control (MPQC). The basic idea is to replace a long-horizon quantum optimal control problem with a sequence of shorter-horizon problems, applying only the first M steps of each solution. Three schemes are presented: a basic scheme (Section III), a terminal-equality-constrained (TEC) scheme for which an exponential stability theorem is stated (Section IV), and a setpoint-optimization scheme (Section V). The paper also discusses how standard QOC algorithms such as GRAPE, Krotov's method, and CRAB can be used within the framework, and it reports single-qubit numerical experiments illustrating efficiency gains in open-loop use and robustness gains in closed-loop use.","tokens_in":17574,"tokens_out":4790,"duration_ms":58416,"significance":"If the stability theorem were fully valid for the systems treated in the numerical sections, the paper would be a useful bridge between classical MPC theory and quantum optimal control, and the modularity claim is attractive: existing QOC solvers could be reused in a receding-horizon scheme with formal guarantees. The paper also makes a fair effort to situate itself in the existing literature on MPC for quantum systems. However, the central theoretical result is currently stated for a state representation that is inconsistent with the pure-state formulation used in the examples, and one of the key numerical demonstrations violates a stated assumption of the theorem. The framework is promising, but the gap between the theorem and the evidence needs to be closed before the claims in the abstract can be accepted.","major_comments":[{"comment":"Assumption IV.1.1 is false for the pure-state representation used throughout the paper. In Section II A, X(t) is taken to be |ψ(t)> for pure states, and the fidelity is defined in Eq. (6) as |<ψ_ref|ψ>|^2. For any X1 and X2 = e^{iθ} X1 with θ not a multiple of 2π, we have F(X1,X2)=1 but ||X1-X2||^2 = |1-e^{iθ}|^2 ||X1||^2 > 0, so no finite c1 can satisfy (17). This inequality is used in Eq. (22) to turn the quadratic controllability bound (18) into a bound on J*(X_t) in terms of infidelity; without (17), the Lyapunov argument collapses. The statement after Assumption IV.1.1 that (17) holds 'for pure states with the diamond norm' does not fix this, because the diamond norm is not the norm induced by the state-vector representation and the numerical implementation in Appendix A works with complex vectors a+ib. A consistent repair would require reformulating the theory for density matrices","section":"Section IV, Assumption IV.1.1 and proof of Theorem IV.1"},{"comment":"The terminal-equality-constrained experiment in Section VII B transfers |+> to |-> for the Hamiltonian H(t)=ωσ_z + u(t)σ_x. The paper itself notes that this target is not an eigenstate of H(u_ref) for any constant u_ref, i.e., condition (9) of Assumption IV.1.4 is violated. Therefore Theorem IV.1 does not cover this experiment, and it cannot serve as numerical validation of the exponential stability claim. The same issue affects the corresponding rows of Table III. A valid numerical test of the theorem would need to use a target satisfying Assumption IV.1.4, or the theorem would need to be generalized to targets that are not eigenstates of the controlled Hamiltonian.","section":"Section VII B, TEC experiment"},{"comment":"The quadratic controllability-cost bound J*(X) ≤ c_u ||X-X_ref||^2 is a nontrivial assumption and is not verified for any of the systems in the paper. It is not merely a technical convenience: the constant c_u enters the contraction factor γ in Eq. (23) and hence the exponential rate in Theorem IV.1. For the terminal-constrained problem (16), feasibility already requires exact steering to X_ref (up to phase), and the cost depends on the stage cost weights and the numerical optimizer used. The citation to classical MPC literature is not by itself evidence that this bound holds for bilinear quantum dynamics. The authors should either prove the bound for a class of systems, verify it numerically for the reported examples, or explicitly state that the exponential-rate conclusion is conditional on an unverified controllability-with-cost assumption.","section":"Assumption IV.1.5"},{"comment":"The proof of Theorem IV.1 is a sketch. The key inequality J*(X_{t+1}) - J*(X_t) ≤ -ℓ(X_t,u_t) is asserted with the comment 'Using classical MPC arguments [16]' and no derivation. For a rigorous theorem in a quantum setting, the reader needs to see how recursive feasibility follows from Assumption IV.1.4, especially given that the terminal constraint (16d) only enforces F=1 and hence allows a global-phase difference between X_L and X_ref. It should be shown explicitly that the tail of the optimal input, augmented by the holding input u_ref, is feasible at the next sampling time and yields the stated cost decrease. This is standard in MPC but should be written out for the fidelity-based stage cost and the phase ambiguity, since those are the nonstandard parts.","section":"Proof of Theorem IV.1"}],"minor_comments":[{"comment":"The stage cost in the setpoint-optimization problem appears to contain a typo: the expressions α(1-F(X_t,X_s(t))) and ||u_t-u_s(t)||^2_R do not depend on the summation index k and seem intended to be α(1-F(X̄_k(t),X_s(t))) and ||ū_k(t)-u_s(t)||^2_R. As printed, the sum is L times the same term and does not penalize the predicted trajectory.","section":"Equation (26a)"},{"comment":"The sentence 'we allow S=0' is unclear notation; S is a set defined in (27), so it should be something like 'we allow S = {0}' or 'we allow the constraint (26e) to be dropped.' Please rephrase.","section":"Section V, after Eq. (26e)"},{"comment":"There is a typo: 'expresssed' should be 'expressed.' Also, the real-valued reformulation (A2) is stated for H(u) with a real part H_r and imaginary part H_i; for the Schrödinger equation including the factor -i, the signs should be checked carefully. The current text says the representation is 'mathematically equivalent,' but it would be helpful to state the precise correspondence to Eq. (2).","section":"Appendix A"},{"comment":"The claim that (17) holds 'for pure states with the diamond norm' is misleading. The diamond norm is defined for quantum operations and is not a natural norm on pure state vectors. For density matrices, a trace-norm or Hilbert-Schmidt norm bound can be appropriate, but the paper should be consistent about whether X denotes a state vector or a density matrix.","section":"After Assumption IV.1.1"},{"comment":"The statement that setpoint optimization 'can fail to stabilize the target setpoint' when condition (9) is violated is not quantified or illustrated. Since Section VII C otherwise reports perfect fidelity for all targets, this caveat deserves either a supporting experiment or a reference to a known counterexample.","section":"Section VII C, last paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible first step toward bringing MPC machinery into quantum optimal control, and the modularity idea is genuinely appealing. My main concern is that the central stability theorem is stated for a representation that is inconsistent with the pure-state implementation used in the numerics, and one of the headline TEC experiments violates an assumption of the theorem. These are fixable in principle by reformulating the theory for density matrices or projective states and by choosing numerical examples that satisfy the assumptions, but the revision will require more than cosmetic changes. I would not recommend rejection at this stage, because the conceptual framework is sound and the errors are localized in the statement and verification of the assumptions. However, the authors need to either prove or verify Assumption IV.1.5 and provide a self-contained proof of the key cost-decrease inequality."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe MPQC framework is a reasonable piece of engineering: it wraps existing QOC solvers in an MPC loop, develops a terminal-equality-constrained variant with a claimed exponential stability guarantee, and adds a setpoint-optimization variant. The tutorial parts and the modular discussion of GRAPE/Krotov/CRAB are useful. The authors are also honest about limitations—they explicitly say the |+> to |-> example violates the eigenstate condition (9).\n\nThe problem is Theorem IV.1. Assumption IV.1.1 asserts ||X1-X2||^2 ≤ c1(1-F(X1,X2)). For the pure-state representation X=|ψ> used in Section II and in all numerics, this is false: take X2=e^{iθ}X1. Fidelity is 1, RHS is 0, LHS is positive. No finite c1 exists. The proof uses (17) to get the Lyapunov bound (22), so the exponential stability result collapses for the stated setting. The authors even call the phase “unimportant” in the text, but the norm they use does not quotient it out. This is not a minor typo; it is the load-bearing step of the theorem. The same issue likely underlies Assumption IV.1.5, which is asserted but never verified: a quadratic upper bound on optimal cost in the vector norm cannot hold when the norm distinguishes phase-equivalent states.\n\nThe fix is straightforward in principle: formulate the theory for density matrices or for a phase-invariant quotient norm. The exponential-stability argument itself is standard MPC and probably goes through once the geometry is right. But as written, the theorem does not cover the pure-state examples that make up the numerical evidence.\n\nThe numerical section is honest but thin: single-qubit tests, no error bars, no code, and Eq. (26a) has an indexing problem (the terms inside the sum don't depend on k, as printed). The robustness comparison to GRAPE is suggestive but not conclusive.\n\nBottom line: this is a useful framework and a clearly written paper, but the central theoretical guarantee is not valid in the current formalism. I'd send it to review—the flaw is repairable and the modular contribution is worth refereeing—but any reviewer should be asked to verify the state-space representation before the theorem is accepted.\n\nBest.","headline":"The modular MPC framework is useful, but Theorem IV.1 rests on an assumption that is false for the pure-state representation used in all the numerics.","tokens_in":18048,"tokens_out":3522,"would_cite":false,"duration_ms":37528,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that repeated short-horizon optimal control, with a terminal equality constraint, makes a quantum target exponentially stable and cuts computational cost.","keywords":["quantum optimal control","model predictive control","receding-horizon control","closed-loop quantum control","stability guarantees","terminal equality constraint","setpoint optimization","state transfer"],"falsifier":"For the single-qubit Hamiltonian in Section VII B, compute the optimal terminal-constrained cost J*(X) for initial states approaching the target and plot J*/||X-X_ref||^2; if that ratio grows without bound as X approaches the target, Assumption IV.1.5 fails and the exponential bound of Theorem IV.1 cannot hold for that system.","tokens_in":17062,"feed_emoji":"⚛️","tokens_out":5149,"duration_ms":45539,"temperature":0.7,"pith_summary":"The paper tries to establish that model predictive control—solve a short-horizon quantum optimal control problem, apply only the first few input steps, remeasure, and repeat—can be applied to quantum systems with both practical and theoretical benefits. It claims this receding-horizon approach improves efficiency in open-loop use and robustness in closed-loop use, while remaining compatible with existing quantum control solvers. Its central theoretical result is that, when each short problem is required to end exactly at the target state, the scheme drives the fidelity to one exponentially fast. A setpoint-optimization variant relaxes that hard constraint and, in the numerical examples, achieves perfect fidelity with much shorter horizons and far lower runtime.","feed_headline":"Quantum control gets a receding-horizon recipe with proof","feed_subtitle":"Model predictive quantum control cuts runtime, adds feedback robustness, and guarantees exponential arrival at the target.","key_machinery":"The carrying mechanism is the receding-horizon loop: at each step, solve a shorter QOC problem of horizon L, apply only the first M steps, update the state by simulation (open loop) or measurement (closed loop), and repeat. For the stability guarantee, the machinery is the optimal value function J*(X) of the terminal-constrained problem, used as a Lyapunov function. The contraction factor γ = 1 − α/(c_u c_1) comes from combining the quadratic upper bound J*(X) ≤ c_u ||X−X_ref||^2 with the norm-fidelity inequality ||X1−X2||^2 ≤ c_1(1−F(X1,X2)) and the positive-definite stage cost. The setpoint-optimization variant replaces the hard terminal constraint by an artificial steady-state setpoint, p","core_discovery":"The paper's central claim is that model predictive control—solve a short-horizon optimal control problem, apply only the first few input steps, reinitialize from the current state, and repeat—can be applied to quantum systems and delivers both practical and theoretical benefits. Its main theorem states that when each short problem is required to end exactly at the target state (a terminal equality constraint), the receding-horizon scheme makes the target exponentially stable: 1−F(X_t,X_ref) ≤ C γ^t (1−F(X_0,X_ref)). The proof uses the optimal cost of the short problem as a Lyapunov function; the key bound is that this cost is quadratically bounded by the distance to the target. The paper als","pith_inferences":["If the efficiency gains persist at larger system sizes, MPQC could serve as a warm-start or approximation engine for pulse compilation, converting one hard long-horizon nonconvex problem into several easier short-horizon ones.","The Lyapunov argument suggests that under small measurement noise the scheme should inherit a form of input-to-state stability, but the paper does not prove this; a testable extension is to bound final infidelity by a function of the noise level.","For targets that are not eigenstates of the controlled Hamiltonian, the setpoint scheme can fail, as the paper notes; a rotating-frame or time-dependent setpoint variant might extend guarantees to that case.","Closed-loop MPQC currently assumes full state tomography, which is costly; the paper's suggestion of shadow tomography points to a concrete way to make the feedback loop practical, and one could test MPQC with very few measurement samples."],"forward_implications":["A long-horizon QOC problem can be decomposed into short subproblems with only minor performance loss; numerical results show runtime drops of roughly an order of magnitude while final fidelity stays at 1.","With feedback, MPQC becomes robust: for a single-qubit model with unknown drift error up to ±1, closed-loop terminal-constrained MPQC keeps final fidelity high where open-loop QOC degrades.","The exponential-stability theorem gives a quantitative convergence rate for state preparation whenever the terminal-constrained subproblem is feasible and the quadratic cost bound holds.","Setpoint optimization reduces the minimal prediction horizon from L=15-20 (terminal equality) to L=2 in the examples, cutting runtime to about one second while preserving unit fidelity.","Because the framework is modular, any existing QOC solver can be plugged into the basic scheme; only the guaranteed schemes require solvers that handle terminal constraints."],"supporting_citations":[{"why":"Supplies the classical MPC feasibility and contraction arguments that the proof of Theorem IV.1 adapts to fidelity costs.","marker":"[16]"},{"why":"Justifies the norm-fidelity inequality (17) for pure states, unitaries, and mixed states.","marker":"[30]"},{"why":"Provides the controllability background invoked for the quadratic cost bound in Assumption IV.1.5.","marker":"[31]"},{"why":"Gradient-ascent pulse engineering: the standard QOC solver that MPQC is benchmarked against and that can implement the basic scheme.","marker":"[11]"},{"why":"Provides the numerical optimization platform used for the implementations of the MPQC subproblems.","marker":"[37]"},{"why":"A monotonic-convergence QOC technique discussed as compatible with the basic scheme but not with terminal-constrained schemes.","marker":"[12]"},{"why":"MPC-for-tracking literature that the setpoint-optimization variant builds on.","marker":"[32-34]"},{"why":"Shows terminal-cost-only stability without terminal constraints, used to argue that basic MPQC can still be made rigorous.","marker":"[48]"}],"fun_headline_variants":["Receding-horizon control gives quantum targets exponential stability","Quantum MPC: modular recipe for faster, sturdier control","Model predictive control offers efficient quantum robustness with proof","Short-horizon quantum control with guarantee: receding-horizon works","Quantum control's new trick: solve small, apply first, repeat"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The proof assumes the cheapest way to exactly reach the target from a nearby state costs at most a constant times the squared distance to it; the experiments never verify this for the systems they simulate.","fun_headline_variants_meta":{"raw":{"variants":["Receding-horizon control gives quantum targets exponential stability","Quantum MPC: modular recipe for faster, sturdier control","Model predictive control offers efficient quantum robustness with proof","Short-horizon quantum control with guarantee: receding-horizon works","Quantum control's new trick: solve small, apply first, repeat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1355,"prompt_tokens":666,"completion_tokens":689,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":410,"completion_tokens_details":{"reasoning_tokens":603}},"tokens_in":410,"tokens_out":689,"duration_ms":6821,"temperature":1.0,"reasoning_tokens":603,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:32:53.767171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the single-qubit Hamiltonian in Section VII B, compute the optimal terminal-constrained cost J*(X) for initial states approaching the target and plot J*/||X-X_ref||^2; if that ratio grows without bound as X approaches the target, Assumption IV.1.5 fails and the exponential bound of Theorem IV.1 cannot hold for that system.","supporting_citations":[{"cited_title":"Introduction to the Pontryagin maximum principle for quantum optimal control,","cited_arxiv_id":null,"evidence_quote":"Supplies the classical MPC feasibility and contraction arguments that the proof of Theorem IV.1 adapts to fidelity costs."},{"cited_title":"Model predictive control for robust quantum state preparation,","cited_arxiv_id":null,"evidence_quote":"Justifies the norm-fidelity inequality (17) for pure states, unitaries, and mixed states."},{"cited_title":"Comparing, optimizing, and benchmarking quantum- control algorithms in a unifying programming frame- work,","cited_arxiv_id":null,"evidence_quote":"Provides the controllability background invoked for the quadratic cost bound in Assumption IV.1.5."},{"cited_title":"Quantum estimation, con- trol and learning: opportunities and challenges,","cited_arxiv_id":null,"evidence_quote":"Gradient-ascent pulse engineering: the standard QOC solver that MPQC is benchmarked against and that can implement the basic scheme."},{"cited_title":"Dis- tance measures to compare real and ideal quantum pro- cesses,","cited_arxiv_id":null,"evidence_quote":"Provides the numerical optimization platform used for the implementations of the MPQC subproblems."},{"cited_title":"Quantum optimal control in quantum technologies. strategic report on current status, visions and goals for research in Eu- rope,","cited_arxiv_id":null,"evidence_quote":"A monotonic-convergence QOC technique discussed as compatible with the basic scheme but not with terminal-constrained schemes."},{"cited_title":"Efficient nu- merical methods for nonlinear MPC and moving horizon estimation,","cited_arxiv_id":null,"evidence_quote":"Shows terminal-cost-only stability without terminal constraints, used to argue that basic MPQC can still be made rigorous."}],"review_version":1}