{"id":"3fa22b3b-45b1-462f-8079-f2816b65f2fb","arxiv_id":"2509.19887","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A phase factor in norm-preserving quantum state diffusion can be chosen to minimize local variance, yielding DO-QSD, which is locally at least as good as any jump unraveling.","lead":"Stochastic unraveling turns a Lindblad master equation into random wave-function trajectories, with a trade-off between memory and statistical noise. This paper derives a tunable phase in the noise that makes the variance of an observable grow as slowly as possible, and proves this choice beats jump-process unravelings locally in time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's pointwise optimal phase need not be continuous; for a qubit with L=σ_y, O=σ_x it is discontinuous at zeros of R, so DO-QSD may fall outside the admissible family (10).","rationale":"The paper's main analytical contribution is the explicit pointwise minimization in Theorem 4, and the reader correctly identified that the state-dependent phase is chosen without checking continuity or measurability near zeros of R. My independent analysis confirms that this is not merely a technical annoyance: for a concrete qubit example, R has isolated zeros with nonzero winding, forcing e^{iθ⋆} to be discontinuous, so the DO-QSD coefficients are not in the continuous family (10) that Theorem 2 characterizes. This makes the theorem, as stated, imprecise: the infimum may be attained only in a generalized sense, not by an admissible continuous unraveling. The numerical method itself may still be useful, and the local dominance over jump processes likely survives as an infimum statement, so I do not think the paper should be rejected. The reader's CONDITIONAL verdict is appropriate; my concern reinforces it rather than changing it. I also note the absence of released code/data, but that is secondary to the mathematical regularity gap.","tokens_in":26458,"tokens_out":9893,"duration_ms":120809,"concrete_test":"Take a qubit with L=σ_y, O=σ_x. Parametrize states by Bloch vector r=(√(1−ε²), ε cos φ, ε sin φ) and compute R(ψ)=−r_x r_y + i r_z. Evaluate the DO-QSD phase e^{iθ⋆(ψ)}=i R*/|R|. Show that as ε→0, this quantity depends on φ (it approaches a direction-dependent unit complex number), so e^{iθ⋆} has no continuous extension at r=(1,0,0). Then check the proof of Theorem 4: because it uses continuity of the coefficients in (10), the theorem as stated excludes DO-QSD. This single qubit computation settles whether the regularity gap is real; if the authors instead intend measurable coefficients, the theorem must be restated and the SDE well-posedness re-argued.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DO-QSD is an explicit global minimizer in the continuous family (10) is load-bearing and is not established. In the proof of Theorem 4 (Appendix A.2), θ⋆ is chosen pointwise as e^{iθ⋆}=i e^{-iP(ψ)}, where P is the phase of R(ψ)=⟨Oψ,Lψ⟩−⟨L⟩_ψ⟨ψ,Oψ⟩. This is only defined where R≠0, and the implied unit field e^{-iP}=R*/|R| need not admit a continuous extension across zeros of R. The family (10) and Theorem 2 explicitly require continuous η, θ, γ; hence the constructed 'minimizer' may not be admissible. This is not a pedantic edge case: for a qubit with L=σ_y and O=σ_x, R(ψ)=−r_x r_y + i r_z in Bloch coordinates, which has simple zeros at r=(±1,0,0) and (0,±1,0). Near r=(1,0,0), R≈−r_y+i r_z, so e^{iθ⋆}≈±(r_z−i r_y)/sqrt(r_y²+r_z²), which has no continuous limit at the zero. Indeed, no continuous unit complex field can satisfy Re{e^{iθ}R}=0 in a neighborhood of such a zero, so the loss L2 has infimum 0 but is not attained in the continuous class (10). The optimal-value formula (15) and the diffusion≤jump comparison may survive as infimum statements, but the advertised 'global minimizer' requires either a relaxed admissible class (e.g., measurable bounded coefficients) or an approximation argument, neither of which is supplied.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript studies stochastic unravelings of Lindblad master equations. For one Lindblad operator and one noise term, it gives a parametric characterization of diffusion and jump unravelings (Theorems 2 and 6), and then derives dynamically optimal unravelings by minimizing the local growth rate of E|⟨ψ,Oψ⟩|² (Theorems 4 and 8). The principal advertised results are an explicit formula for the optimal diffusion coefficient/phase and a proof that optimized diffusion is locally no worse than any jump process. Numerical experiments on a two-level system and a cavity-QED model compare DO-QSD with rQSD, cQSD, and a machine-learned unraveling.","tokens_in":26858,"tokens_out":12494,"duration_ms":92021,"significance":"The paper contains a genuinely useful observation: a state-dependent phase freedom in the unraveling family, previously neglected, can be chosen in closed form to reduce the local second-moment growth of a target observable. The Itô calculus derivation is clean, the DO-QSD formula is explicit and free of fitted parameters, and the diffusion-vs-jump dominance inequality is a strong, falsifiable statement. If the regularity gap identified below is repaired, the paper would be a solid contribution to trajectory methods for Lindblad simulation. The numerical evidence is suggestive but limited to two small models, and no code or data are provided.","major_comments":[{"comment":"The pointwise optimizer e^{iθ*}=i e^{-iP(ψ)} is defined only where R(ψ)=⟨Oψ,Lψ⟩-⟨L⟩_ψ⟨ψ,Oψ⟩≠0, and no continuous selection across zeros of R is established. This is not a pedantic edge case: for L=σ_y, O=σ_x, R=-r_x r_y+i r_z in Bloch coordinates, and near r=(1,0,0) the required phase behaves like (r_z-i r_y)/√(r_y²+r_z²), which has no continuous limit at the zero; no continuous unit field can achieve Re{e^{iθ}R}=0 in a neighborhood. Since the admissible family (10) and Theorem 2 require continuous η,θ,γ, the claimed 'global minimizer' may fall outside the class in which the optimization is posed. The theorem should be restated as an infimum over continuous admissible coefficients, or the class should be enlarged to measurable bounded coefficients with a proof of SDE well-posedness. The comparison inequality in Theorem 8(i) survives as an infimum comparison, but Theorem 4's attainment cl","section":"Section 2.3 / Theorem 4 / Appendix A.2, Eq. (A.7)"}],"minor_comments":[{"comment":"The text says the atom 'starts in the maximally entangled state' when the initial state is |+⟩⟨+|⊗|0⟩⟨0|. A single-qubit superposition is not maximally entangled; the wording should be corrected (e.g., 'maximally coherent state' or 'equal superposition').","section":"Section 4.2"},{"comment":"Lemma 10 is used to produce the explicit minimizer in Theorem 8(ii), but its proof is omitted as 'elementary'. Given that the theorem is an advertised explicit result, a short derivation of (B.8)-(B.9) should be included.","section":"Appendix B.3, Lemma 10"},{"comment":"The parameter Λ is introduced as a uniform upper bound on the jump rate, but the paper does not discuss how Λ should be chosen in practice or how sensitive the DO-QJP formulas are to it. Since DO-QJP is not implemented numerically, a brief comment on the role of Λ would improve interpretability.","section":"Section 3.3, Eq. (24)"},{"comment":"No code or data availability statement is provided. For a numerical methods paper, releasing the simulation scripts would significantly aid reproducibility.","section":"Appendix D / Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the central idea is sound, but the regularity gap in Theorem 4 is a genuine correctness issue in a load-bearing claim. It is fixable by restating the result as an infimum over the continuous admissible class or by extending the admissible class to measurable bounded coefficients. I would not reject the paper on this basis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper gives a clean characterization of norm-preserving stochastic unravelings for a single Lindblad operator, and the state-dependent phase factor e^{iθ} in the diffusion term is a genuine addition to Gisin–Percival. Second, the claimed explicit global minimizer DO-QSD (Theorem 4) is not actually admissible in the continuous class the paper defines, because the optimal phase is discontinuous at zeros of R(ψ). That is a load-bearing gap, not a technicality.\n\nWhere the paper earns its keep: Lemma 9's Itô calculus is clean, the reduction of the variance derivative to a single loss term is the right way to view the problem, and the dominance inequality dV_diffusion ≤ dV_jump in Theorem 8 follows from a nonnegative extra term. The DO-QSD formulas are simple and cheap, and the numerics, while small, show genuine variance reductions. The authors are honest about local-in-time and problem-dependent scope, and they do not oversell the machine-learning comparison.\n\nThe soft spot: Theorem 4 claims a global minimizer in the family (10), where η, θ, γ are continuous. The pointwise optimum e^{iθ*} = i e^{-iP(ψ)} is the phase of R*/|R|. At a zero of R, that field has no continuous extension. For L=σ_y, O=σ_x, R = -r_x r_y + i r_z, with simple zeros on the Bloch equator; the phase has a vortex. In any neighborhood of such a zero, no continuous unit complex field satisfies Re{ e^{iθ} R } = 0 identically. So the infimum of the loss is zero but is not attained in the continuous class. The paper needs either to relax the admissible class to measurable bounded coefficients (then address SDE well-posedness for discontinuous coefficients) or to give an approximation argument. As written, the 'global minimizer' statement is too strong. Also, no code or data is released, and the two numerical examples are small; that is minor compared to the regularity gap.\n\nWho this is for: anyone simulating Lindblad equations with stochastic wave functions and caring about estimator variance. The characterization alone is worth a read. But the rigorous claim of optimality needs repair. I would send this to peer review, with the referee asked specifically to check Theorem 4's admissibility and the proposed fixes. The paper is not a desk reject; it's a revise.","headline":"A genuinely useful unraveling optimization, but the headline optimality theorem has a regularity gap at zeros of the phase.","tokens_in":27327,"tokens_out":4355,"would_cite":true,"duration_ms":37943,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81S22","81S25"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that among all norm-preserving diffusion-type stochastic unravelings of a Lindblad equation, a closed-form, state-dependent phase choice minimizes the local growth rate of the variance of any chosen observable, and the opt","keywords":["Lindblad equation","stochastic unraveling","quantum state diffusion","quantum jump process","variance reduction","open quantum systems","dynamically optimal","Monte Carlo simulation"],"falsifier":"Take a two-level system with L=√γ σ_- and O=σ_x, start from a state that crosses the zero set of R(ψ), and simulate the DO-QSD SDE with a fine time step. If the empirical local growth of the variance exceeds the predicted optimal value or the state norm drifts, the pointwise minimizer is not admissible.","tokens_in":26338,"feed_emoji":"⚛️","tokens_out":7821,"duration_ms":174815,"temperature":0.7,"pith_summary":"To simulate an open quantum system, one often represents the density matrix as an average over random pure-state trajectories. Different unravelings produce different statistical noise even though the ensemble average is identical, and picking the least noisy one matters for computing observables. This paper characterizes, for a single Lindblad operator, essentially every diffusion and jump unraveling that preserves the norm along each path. It then solves the greedy optimization problem of minimizing the instantaneous growth of E|⟨ψ,Oψ⟩|², obtaining an explicit optimal diffusion scheme (DO-QSD) whose phase factor e^{iθ}=ie^{-iP} cancels the variance-driving term. Because the same optimization can be performed for jump processes, the paper proves that optimized diffusion's local variance growth is never larger than that of any jump ansatz, and the closed-form scheme needs no training.","feed_headline":"Diffusion unraveling beats every jump scheme locally","feed_subtitle":"A closed-form phase twist makes quantum trajectory sampling no noisier than jump processes, with no training.","key_machinery":"The key object is the parametric family (10) of diffusion unravelings, with free complex coefficients η_k(ψ) and real phase functions θ_k, γ_k, and the norm-preservation constraint η_k=ih_k-e^{iθ_k}⟨L_k⟩_ψ. The optimization is driven by a loss function L1=E[Re{η⟨ψ,Oψ⟩+e^{iθ}⟨Oψ,Lψ⟩}²] that appears in the decomposition of (d/dt)E|⟨ψ,Oψ⟩|²; the scheme-independent terms depend only on L and O, so minimizing the loss is sufficient. The theorem pins down the phase e^{iθ⋆} so that the complex combination becomes purely imaginary, killing L1, and a comparison with the jump family shows that the extra term in the jump variance is nonnegative. The central identity behind the comparison is the formula","core_discovery":"The central discovery is that the earlier neglect of a state-dependent phase factor in the diffusion coefficient was not harmless: it is the phase that converts the stochastic unraveling from a fixed-noise ansatz into an adaptive, variance-minimizing one. For one Lindblad operator L and observable O, the paper proves that any norm-preserving diffusion can be written with drift and diffusion containing η(ψ)=ih(ψ)-e^{iθ(ψ)}⟨L⟩_ψ; the variance growth separates into a scheme-independent part plus the squared real part of a complex quantity. Minimizing pointwise yields η⋆=-e^{iθ⋆}⟨L⟩_ψ with e^{iθ⋆}=± i e^{-iP}, where P is the phase of ⟨Oψ,Lψ⟩-⟨L⟩_ψ⟨O⟩_ψ, making the loss identically zero at each s","pith_inferences":["If the regularity issue around R(ψ)=0 is resolved (for example, by a consistent phase convention), the pointwise local optimality could plausibly be lifted to finite-time variance bounds for a large class of Hamiltonians; this is not proved in the paper.","Because the optimal phase depends only on current expectations, the same greedy construction applies to time-dependent Lindbladians, which the authors note, and suggests adaptive 'online' unravelings for non-Markovian embeddings where the noise process can be tuned as the simulation evolves.","The paper skips the proof of Lemma 1 and does not prove continuity or measurability of the optimal phase; these are gaps rather than contradictions, but they mean the well-posedness of the DO-QSD SDE is established only away from the zero set of R. A concrete robustness test is to replace the singular phase at R=0 by its principal value and measure the finite-time variance penalty.","The comparison ignores discretization bias; for coarse time steps, jump processes may win in practice even though the continuous-time local variance is worse, so the 'diffusion-wins' conclusion is an ideal-solver statement."],"forward_implications":["For a single-observable simulation, DO-QSD gives the smallest possible local variance growth among all norm-preserving diffusion unravelings in family (10), so no neural-network training or parameter search is needed to reach that level.","Locally in time, DO-QSD's variance growth is no larger than that of any jump-process unraveling, including DO-QJP; thus diffusion is at least as sample-efficient in the small-time-step regime.","The optimal phase is invariant under adding a constant to the observable, so the estimator for tr(Oρ) inherits the same variance benefit for O+cI.","For multiple observables, an explicit multi-observable version (multi-DO-QSD) minimizes the total variance, though with a more constrained choice of phase.","The scheme is complementary to control-variate and multilevel Monte Carlo techniques, so it can be combined with other variance-reduction strategies."],"fun_headline_variants":["Phase twist yields provably optimal diffusion unraveling","Diffusion scheme beats jump processes in variance","Closed-form optimal unraveling for Lindblad equations","Variance-minimal unraveling derived for any observable","Noise-minimizing quantum trajectories without training"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The pointwise optimal phase e^{iθ⋆}=ie^{-iP} is chosen from the phase of R(ψ)=⟨Oψ,Lψ⟩-⟨L⟩_ψ⟨O⟩_ψ, and the paper assumes this selection is continuous enough for the stochastic differential equation to be well-posed; no regularity proof is given at states where R(ψ)=0.","fun_headline_variants_meta":{"raw":{"variants":["Phase twist yields provably optimal diffusion unraveling","Diffusion scheme beats jump processes in variance","Closed-form optimal unraveling for Lindblad equations","Variance-minimal unraveling derived for any observable","Noise-minimizing quantum trajectories without training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3014,"prompt_tokens":756,"completion_tokens":2258,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":2187}},"tokens_in":500,"tokens_out":2258,"duration_ms":631182,"temperature":1.0,"reasoning_tokens":2187,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T15:18:45.767704+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-level system with L=√γ σ_- and O=σ_x, start from a state that crosses the zero set of R(ψ), and simulate the DO-QSD SDE with a fine time step. If the empirical local growth of the variance exceeds the predicted optimal value or the state norm drifts, the pointwise minimizer is not admissible.","supporting_citations":[],"review_version":1}