{"id":"5d24ba61-fa74-4667-a6de-d6de2f8a607f","arxiv_id":"2601.09075","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Padé-rational fits to Volterra memory kernels can predict non-Markovian qubit trajectories across initial states even when the underlying kernels are not uniquely recoverable.","lead":"This paper introduces a Padé-based regression method to learn Volterra memory kernels of non-Markovian qubit dynamics from trajectory data. On three synthetic qubit-bath models, the method reproduces state trajectories across initial states, but the recovered kernels do not match the true ones.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3)'s convolution ansatz is violated by the exact kernels of Test Problems 2 and 3, so the learned B is not the physical memory kernel.","rationale":"The paper's central claim is that it learns the memory kernel. For this to be true, the hypothesis class must contain the true kernel. Equation (3) posits B(t−τ), but the exact generators in Secs. IV and V have kernels with explicit absolute-time dependence. In Test Problem 3, the authors themselves display B(t,τ) in Eq. (20), and in Test Problem 2 the e^{−2iε₀t} term in Eq. (15b) cannot be captured by any convolution kernel on the chosen state vector. This is not merely numerical ill-conditioning; it is a structural mismatch. The nonlocal Crank-Nicolson scheme in Appendix A is derived for B(t−τ) only, so—absent an undocumented modification—the solver is inconsistent with the stated data-generation model in Test Problem 3. Consequently, the learned functions are at best effective kernels that reproduce trajectories on [0,T]; they are not the physical memory kernels. The paper honestly reports poor kernel recovery in Figs. 3–4 and Sec. VI, but the conclusion overstates by claiming kernel identification. This directly undermines the title and abstract. The suggested re-fit with the correct B(t,τ) would settle whether the misspecification is the cause or whether even the correct model fails due to identifiability; either way, the current claim is unsupported. The reader's weakest assumption identifies the same issue, so I agree and recommend no change to the REJECT verdict.","tokens_in":15326,"tokens_out":7745,"duration_ms":76462,"concrete_test":"Re-run Test Problem 3's optimization using the exact B(t,τ) from Eq. (20) in the Crank-Nicolson solver, evaluating the history integral with B(t_n, t_k) rather than B(t_n−t_k), while keeping C_αβ(t−τ) as Padé functions. Compare the learned C_αβ with the prescribed C_αβ. If the non-convolution model recovers the kernels (kernel error comparable to Test Problem 1) while the lag-only convolution model does not, the misspecification is confirmed as the cause of kernel non-recovery.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the framework identifies B in the Volterra equation. But Eq. (3) assumes a time-translation-invariant kernel B(t−τ), while the exact dynamics used to generate data are not of that form. In Test Problem 2, Eq. (15b) contains an explicit factor e^{−2iε₀t} multiplying ρ10(s) inside the integral. This cannot be written as B(t−s)ρ10(s) for any B, since the t-dependence does not factor through the lag. In Test Problem 3, the data-generating kernel is explicitly written as B(t,τ) in Eqs. (19)–(20), with matrices M_αβ(t) multiplying scalar C_αβ(t−τ), so the kernel depends on t and τ separately, not only on t−τ. The Appendix A solver and the Padé-in-lag parameterization are derived for B(t−τ); applying them to B(t,τ) is a misspecified regression. A model that cannot represent the target object can fit trajectories well while producing an arbitrary effective kernel, which is exactly what the paper admits: 'the learned correlation functions are not close by any metric to the numerically evaluated correlation functions' (Sec. IV, Fig. 3). Thus the stated objective—identifying A and B from data—is not met for Test Problems 2 and 3, where no exact B(t−τ) exists.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven framework for extracting the linear operator A and the memory kernel B in the Volterra integro-differential equation dx/dt = Ax + ∫_0^t B(t−τ)x(τ)dτ from qubit trajectory data. Each component of B is parameterized as a Padé rational function; the parameters are optimized with a BFGS quasi-Newton method against a sum of trajectory misfits plus an H^1 regularization. The method is applied to three synthetic models: pure dephasing, an energy-exchange model, and a noncommuting σx/σz coupling model. The paper includes a well-posedness proof for the regularized optimization problem and a Crank-Nicolson discretization of the Volterra equation.","tokens_in":15724,"tokens_out":9855,"duration_ms":82838,"significance":"If the claims were fully supported, the paper would provide a simple, interpretable baseline for memory-kernel identification in open quantum systems. The pure-dephasing test (Sec. III) is a clean proof-of-principle: the convolution assumption is exact there, and the learned Padé kernel appears to reproduce both the trajectory and the correlation function. The existence theorem (App. B), while not novel, is a useful guarantee for the regularized finite-dimensional problem. The paper also deserves credit for explicitly acknowledging the ill-posedness of kernel recovery. However, the main claim of 'identifying A and B directly from data' is established only for the first test problem. For the other two problems the data-generating dynamics are not in the assumed convolution class, so the learned B is an effective object rather than the physical memory kernel. This confines the actual contribution to trajectory forecasting with effective convolution models.","major_comments":[{"comment":"The central hypothesis (3) requires B(t−τ) to depend only on the lag. The exact dynamics used for Test Problems 2 and 3 violate this. In Eq. (15b), the second term contains e^{−2iε0t} e^{i(ε0−ω)(t−s)}; for ε0=1 this equals e^{−i(ε0+ω)t} e^{−i(ε0−ω)s}, which depends on t and s separately, not on t−s alone. Thus no matrix kernel B(t−τ) can reproduce Eq. (15). The statement in Sec. IV that 'the dynamics can be written in the form of eq. (3)' is therefore incorrect. For Test Problem 3, Eq. (20) explicitly defines B(t,τ) with t-dependent matrices B_αβ(t). The optimization over O in Eq. (5) is searching the wrong class for these problems.","section":"Sec. II Eq. (3); Sec. IV Eq. (15b); Sec. V Eq. (20)"},{"comment":"The paper's own results show that the learned B is not the physical kernel for Problems 2 and 3. The caption of Fig. 3 states 'the learned correlation functions are not close by any metric to the numerically evaluated correlation functions embedded in eqs (15).' Sec. V states 'Pointwise kernel recovery is, as in problem 2, significantly more challenging.' Since the true kernel is outside the model class, a good trajectory fit is not evidence of kernel identification; it is an effective-fit phenomenon. The abstract's 'identifying non-Markovian kernels' and Sec. II's 'central objective ... identify A and B directly from data' are therefore overstated and should be rephrased, or the experiments must be restricted to convolution-compatible models.","section":"Sec. IV Fig. 3; Sec. V; Abstract; Sec. II"},{"comment":"The numerical scheme is derived for Eq. (A1) with B(t−τ), and the update (A4)–(A5) evaluates B at entries such as B(t_n−t_k). For the two-argument kernel B(t,τ) in Eq. (19), the scheme as written is not defined. The manuscript does not describe how the nonlocal Crank–Nicolson update is modified for Test Problems 2 and 3. This is a reproducibility gap in the numerical sections and also indicates that the implemented solver does not solve the stated state equation for the nonconvolution part.","section":"Appendix A, Eqs. (A1)–(A5)"},{"comment":"The existence and regularity theorem is proved only for the convolution state equation (B1). It does not cover the nonconvolution kernels B(t,τ) of Test Problem 3 or the second term of Eq. (15b) in Test Problem 2. Thus the claimed 'well-posedness of the learning problem' does not extend to two of the three numerical test problems. The theorem should either be extended or the theoretical claims qualified.","section":"Appendix B, Eq. (B1)"}],"minor_comments":[{"comment":"Spelling is inconsistent: both 'Crank-Nicholson' and 'Crank-Nicolson' appear, and 'Padé' accents are inconsistent.","section":"Throughout"},{"comment":"The text says the sub-Ohmic, Ohmic, and super-Ohmic regimes are all studied, but the figure caption lists only p=1/2 and p=2. The Ohmic case is missing from the displayed results.","section":"Sec. III, Fig. 1"},{"comment":"The [q/r] notation is nonstandard because the denominator degree appears to be q+r+1; the convention should be defined explicitly.","section":"Eq. (4)"},{"comment":"The sentence 'therefore, we do not investigate how our trained model generalizes given the uniqueness of integral curves from the dynamics' is unclear and should be rewritten.","section":"Sec. III"}],"recommendation":"reject","confidential_remarks":"The paper is not circular and does not appear to hide prior work; it is also honest about its limitations. My rejection is based solely on the model-class mismatch between the convolution ansatz and the actual test problems. The paper could be reconsidered as a study of effective convolution models for trajectory prediction, but as written the abstract and conclusions overstate the kernel-identification achievement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main thing to know: this is a decent data-driven baseline for predicting non-Markovian qubit trajectories, but it does not do what the title and abstract claim. The kernel recovery fails in two of the three test problems, and the paper itself admits this. If reframed as trajectory prediction with effective kernels, it's a plausible contribution; as a kernel identification method, it overreaches.\n\nWhat's actually new: the specific package — vectorized qubit Volterra dynamics, Padé rational kernels, Tikhonov regularization — is a straightforward extension of existing memory-kernel learning in GLEs and transfer-tensor methods, but the first test problem (pure dephasing) is clean and shows accurate kernel recovery in that special case. The observation that accurate trajectory fits coexist with kernel non-identifiability is worth stating, and the authors are honest about it in the conclusions. The well-posedness proof is standard but fine.\n\nThe soft spots are serious. Equation (3) assumes a time-translation-invariant kernel B(t−τ). Test Problem 2 has an explicit e^{−2iε0 t} factor in the integrand (Eq. 15b), and Test Problem 3 is written directly as B(t,τ) with t-dependent matrices (Eq. 20). The model cannot represent these kernels, so the learned B is an artifact of misspecification — not the physical memory kernel. The paper's own Figure 3 caption says the learned correlation functions \"are not close by any metric\" to the true ones. That is the headline result, not the identification claim. Also, the arXiv abstract promises a fourth test problem (quantum Rabi model) and a \"parametrization-invariant sensitivity analysis\" that do not appear in the body. Maybe those are in a longer version, but as submitted, it's a mismatch.\n\nNone of this is fatal to the trajectory-prediction story, which does generalize out-of-sample on physically admissible initial states. But the authors need to revise the framing substantially: title, abstract, and claims should be about learning effective kernels for prediction, not identifying the true Nakajima–Zwanzig memory kernel.\n\nI'd send this to peer review. A good referee can push for the reframing and check the noise/regularization behavior more carefully. The reader's REJECT is a bit harsh if the authors are willing to revise; the underlying numerical work seems honest and reproducible in principle. For a reading group, it's a maybe — useful as a cautionary example of identifiability versus prediction in data-driven quantum dynamics.","headline":"Trajectory prediction works, but the paper's central claim of identifying memory kernels is not supported: the model assumes a convolution form that the exact kernels of two test problems violate, and the abstract promises more than the body delivers.","tokens_in":16146,"tokens_out":2299,"would_cite":false,"duration_ms":22937,"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":"Padé rational fits to qubit trajectory data learn effective Volterra memory kernels for non-Markovian dynamics, and the regularized learning problem is provably well-posed.","keywords":["non-Markovian dynamics","memory kernel","Volterra integro-differential equation","Padé approximants","open quantum systems","Nakajima-Zwanzig","kernel identification","qubit decoherence"],"falsifier":"A decisive test: use the trained kernel from Test Problem 2 to predict the state at a time well beyond the training horizon (say t=10) and compare to exact integration of Eqs. (15); a divergence would confirm that the learned B is an effective finite-window kernel rather than the true memory kernel. A more local falsifier is to compare the learned kernel's prediction for d²ρ/dt² at t=0 (which is governed by B(0)) against the exact short-time expansion.","tokens_in":15204,"feed_emoji":"⚛️","tokens_out":9271,"duration_ms":72127,"temperature":0.7,"pith_summary":"The paper develops a data-driven method to infer the non-Markovian equation of motion of a qubit from time-series data. Starting from the Nakajima-Zwanzig form, the reduced density matrix is vectorized and the dynamics are cast as a Volterra integro-differential equation with an operator-valued memory kernel. Each kernel entry is modeled as a Padé rational function, and the free parameters are fit by minimizing a regularized trajectory-misfit functional. The authors prove the optimization is well-posed and demonstrate on three synthetic test problems that the learned models reproduce and generalize the state dynamics. Crucially, they find that accurate trajectory fits do not imply pointwise recovery of the true kernel, so what is learned is an effective kernel that supports prediction, not necessarily the microscopic one.","feed_headline":"Padé fits learn effective qubit memory kernels from data","feed_subtitle":"Minimal rational functions match open-qubit trajectories even where the exact memory kernel cannot be recovered pointwise.","key_machinery":"The central object is the vectorized Volterra equation dx/dt = A x(t) + ∫_0^t B(t−τ)x(τ)dτ, where x ∈ C^4 is the vectorized qubit density matrix, A is the instantaneous generator, and B is the operator-valued memory kernel. Each scalar entry B_ij is modeled as a [q/r] Padé approximant in the lag variable—a fixed-order rational function with coefficients ξ. A nonlocal Crank-Nicolson update solves the forward dynamics, and the learning objective is a Tikhonov-regularized L^2 trajectory misfit. The paper proves existence of minimizers via coercivity of the H^1 penalty, Sobolev compactness, and weak lower semicontinuity.","core_discovery":"The central claim is that a low-parameter rational representation—fixed-order Padé approximants for each entry of the matrix-valued memory kernel—suffices to capture nontrivial non-Markovian features such as oscillatory memory, algebraic tails, and phase-sensitive coherence transfer. Learning is posed as a constrained optimization over an admissible operator space, regularized by an H^1 penalty, and the Volterra state equation is solved by a nonlocal Crank-Nicolson scheme. The paper proves existence of minimizers and sequential continuity of the loss along bounded operator sequences. Across three increasingly complex synthetic testbeds, the learned models accurately reproduce and generalize","pith_inferences":["A consequence the authors leave implicit: because the learned kernel is effective rather than pointwise-correct, any downstream quantity that depends on the kernel's values—multi-time correlations, fidelities under pulse sequences, or entanglement with the environment—must be validated separately; trajectory fidelity alone does not certify them.","The observed non-identifiability suggests a general limit for Nakajima-Zwanzig kernel fitting from state trajectories alone: without additional structural assumptions (complete positivity, known commutator structure, longer observation windows), the true memory kernel cannot be disentangled from the effective one.","A natural experimental extension would be to feed noisy process-tomography data from a real qubit into this pipeline; the noise-sensitivity analysis in Test Problem 1 indicates that a small H^1 regularization can suppress spurious kernel oscillations, which would be essential for experimental data."],"forward_implications":["For pure dephasing, the Padé model reproduces the correlation function accurately in sub-Ohmic, Ohmic, and super-Ohmic regimes, so the method can be used as a fast emulator (about 0.01 s per evaluation) for spectral-density sweeps.","Learned Padé-kernel models can predict non-Markovian qubit trajectories on moderate time scales from a modest number of training trajectories (30 for the multi-channel problems).","Because state trajectories are insensitive to the unrecoverable parts of the kernel, trajectory-level prediction is robust to the severe ill-conditioning of kernel recovery.","The approach avoids symbolic libraries of special functions and deep-network architectures, needing only fixed-order rational functions plus a standard quasi-Newton optimizer.","The well-posedness theorem guarantees that the regularized learning objective always has a minimizer, so numerical fitting is on solid ground even when the kernel is not unique."],"fun_headline_variants":["Padé learns non-Markovian qubit memory from data","Data-driven rational fits capture qubit memory kernels","Padé approximants recover qubit memory from trajectories","Padé fits recover qubit memory despite ill-conditioning","Low-order Padé fits learn qubit memory kernels"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the true reduced dynamics are exactly representable as a time-translation-invariant Volterra convolution, dx/dt = A x(t) + ∫_0^t B(t−τ)x(τ)dτ with B depending only on the lag; but the exact kernels of the paper's own Test Problems 2 and 3 contain explicit absolute-time factors (e.g., e^{-2iε0t} in Eq. (15b) and the t-dependent matrices B_αβ(t) in Eq. (20)), so the convolution ansatz is already violated in the testbeds.","fun_headline_variants_meta":{"raw":{"variants":["Padé learns non-Markovian qubit memory from data","Data-driven rational fits capture qubit memory kernels","Padé approximants recover qubit memory from trajectories","Padé fits recover qubit memory despite ill-conditioning","Low-order Padé fits learn qubit memory kernels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001547,"raw_usage":{"total_tokens":6062,"prompt_tokens":821,"completion_tokens":5241,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":5162}},"tokens_in":565,"tokens_out":5241,"duration_ms":33129,"temperature":1.0,"reasoning_tokens":5162,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:44:30.740073+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: use the trained kernel from Test Problem 2 to predict the state at a time well beyond the training horizon (say t=10) and compare to exact integration of Eqs. (15); a divergence would confirm that the learned B is an effective finite-window kernel rather than the true memory kernel. A more local falsifier is to compare the learned kernel's prediction for d²ρ/dt² at t=0 (which is governed by B(0)) against the exact short-time expansion.","supporting_citations":[],"review_version":1}