{"id":"ac57210b-6724-4341-acc6-6132e259db3e","arxiv_id":"2505.11762","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A parameterized Wasserstein Hamiltonian flow, implemented through Neural ODEs, is proposed as a sampling-based solver for the time-dependent Schrödinger equation.","lead":"This paper rewrites the time-dependent Schrödinger equation as a Hamiltonian flow of push-forward maps, a structure borrowed from generative models, and solves that flow with neural-network parameters instead of a grid. The method is tested on a harmonic oscillator, Gross-Pitaevskii equations, and a nine-dimensional three-particle problem, with qualitative agreement in low dimensions and no quantitative error assessment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The finite-dimensional reduction in Section 5 is not shown to approximate the full Wasserstein Hamiltonian flow: no error bound connects (5.4) to (4.5), and Section 8 lists this as open, so the central 'approximate solution' claim rests on an unproven closure assumption.","rationale":"The reader identified the reduced-order parameterization as the weakest assumption, and my assessment agrees: this is the single most load-bearing concern. The paper's main contribution is not merely the WHF reformulation, which is a clean and largely standard derivation, but the claim that a Neural-ODE parameterization yields a practical, training-free solver. That claim depends on the unproven assumption that the finite-dimensional manifold Oθ captures the infinite-dimensional flow well enough for the solution of (5.4) to approximate the solution of (4.5). This is not a matter of disagreement with existing consensus; it is an internal gap between the exact infinite-dimensional theorem and the numerical algorithm. The paper explicitly acknowledges the missing error bound in Section 8, which supports classifying this as an open issue rather than a fabricated objection. The numerical experiments are qualitative, particularly the 9D example, which has no reference solution and no convergence check; energy conservation alone is insufficient because the reduced system is itself Hamiltonian and can preserve its own energy while diverging from the true WHF trajectory. I therefore do not see a basis for rejecting the paper outright, since the infinite-dimensional framework and the affine exact-solution check are genuine supporting evidence. However, the central algorithmic claim cannot be accepted as established until the reduction step is either theoretically justified or validated by a quantitative convergence test on a problem with a known exact solution. Since the reader already issued CONDITIONAL, my independent read does not change that verdict.","tokens_in":17259,"tokens_out":6677,"duration_ms":76085,"concrete_test":"Run the 1D GPE experiment of Section 7.2 with the exact soliton solution (7.10), using the same Neural-ODE architecture but two very different widths (e.g., 20 and 200 hidden units per layer) and two horizons (t=1 and t=2). At each terminal time, compute the empirical 1-Wasserstein distance between Tθ(t)♯λ (evaluated from the Algorithm 6.1 samples) and the exact density (7.10). If the error does not decrease when the network width increases, or if it grows superlinearly with the horizon, then the reduced dynamics (5.4) are not converging to the full WHF flow, and the 'approximate solution' claim is not supported; if the error does shrink with width and remains bounded over the longer horizon, the concern is mitigated for this representative case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest part of the paper is the infinite-dimensional reformulation. Theorem 4.1 derives the second-order system (4.5) from the action (4.4), and the harmonic-oscillator example in Section 7.1 confirms that an affine subfamily yields the exact density evolution in that special case. The load-bearing weakness is the step from this exact WHF to the Neural-ODE parameterization in Section 5. The paper asserts that solving (5.4) 'can provide an approximate solution' to (4.5), but this requires that the reduced dynamics on Oθ faithfully track the full flow. For a generic submanifold Oθ, the Euler-Lagrange equations in the induced metric coincide with the restriction of the full Euler-Lagrange equations only if Oθ is totally geodesic, or if the true trajectory happens to remain in Oθ. Neither condition is established or even discussed for the chosen Neural-ODE family. No a priori or a posteriori bound is given relating the projected density Tθ(t)♯λ to the true solution ρ(t); the Wasserstein error question is explicitly left open in Section 8. The numerical evidence does not close this gap: the 9D example has no ground truth, and the Hamiltonian preservation shown in Figure 8 is a structural property of the reduced system that can hold even when the trajectory is far from the true WHF solution. Thus the central algorithmic claim — that the parameterized ODEs yield a controlled approximation of the TDSE density — is currently unsupported, even though the underlying WHF reformulation appears mathematically sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reformulates the time-dependent Schrödinger equation (TDSE) (1.1) via the Madelung transform as a Wasserstein Hamiltonian flow, and then as a Hamiltonian system on the space of push-forward diffeomorphisms. Theorem 4.1 derives the second-order Euler-Lagrange equation (4.5) for the critical point of the action (4.4). The authors then restrict the maps to a Neural-ODE parameterized family, pull back the L2(λ) metric to parameter space, and obtain the finite-dimensional Hamiltonian system (5.4). A semi-implicit symplectic algorithm is proposed and tested on three examples: a quantum harmonic oscillator, the Gross-Pitaevskii equation in one, three, and six dimensions, and a 9D three-particle system. The paper claims that solving the parameterized system provides an approximate solution to the original infinite-dimensional flow, with no training of the neural network against data.","tokens_in":17618,"tokens_out":8101,"duration_ms":84261,"significance":"If the central claim were established, the method would offer a sampling-based, training-free alternative for high-dimensional TDSE, with potential practical value. The strongest parts of the paper are the clean derivation of Theorem 4.1 and the exact harmonic-oscillator example in Section 7.1, where an affine subfamily reproduces the true density evolution. The authors are also transparent in Section 8 that the Wasserstein error bound is open. However, the load-bearing claim that the reduced system (5.4) approximates the full flow (4.5) is not supported by any error estimate, and the numerical validation is largely qualitative. The significance is therefore conditional: the paper presents an interesting reformulation, but the algorithmic claim needs substantial additional support.","major_comments":[{"comment":"The assertion that solving the parameterized system (5.4) \"can provide an approximate solution Tθ(t) to the original flow (4.5)\" is the central algorithmic claim of the paper, but no error bound or convergence statement connects the reduced dynamics to the full Wasserstein Hamiltonian flow. The reduced variational problem (5.3) minimizes the action over a prescribed finite-dimensional family Oθ; the Euler-Lagrange equations on Oθ coincide with the restriction of the full equations only if Oθ is totally geodesic or if the exact trajectory remains in Oθ, and neither condition is shown for the Neural-ODE family. Section 8 explicitly leaves the theoretical error bound in the Wasserstein metric as an open problem. To support the claim, the paper needs at least an a posteriori error estimate or a substantially weakened statement; without this, the algorithm's output is not shown to be a controlled approximation of the TDSE density.","section":"Section 5, after Eq. (5.5)"},{"comment":"For the Gross-Pitaevskii equation, the exact solution (7.10) is available, so reporting histograms only is not sufficient to quantify accuracy. The paper should include quantitative comparisons, such as L1 or L2 errors of the density, empirical Wasserstein distances, or errors in the mean and variance, as functions of time, and should study the sensitivity of the results to the number of samples N, the step size h, and the Neural-ODE width and depth. For the 9D example in Section 7.3 there is no ground truth; the Hamiltonian preservation shown in Figure 8 is a structural property of the reduced system (5.5) and can hold even when the trajectory is far from the true solution, so it does not by itself validate the accuracy of the method.","section":"Section 7.2, Figs. 2-5; Section 7.3, Fig. 8"},{"comment":"Proposition 5.1 is imported from the authors' earlier work [43] with the phrase \"Mimicking the derivation of PWHF as detailed in [43],\" yet this proposition is the foundation of the numerical method. The paper should either give a self-contained derivation or state the precise hypotheses needed, in particular the rank and conditioning of G(θ), the meaning and well-posedness of the pseudo-inverse G†, and whether the momentum p is guaranteed to lie in the range of G. Without this, the use of the pseudo-inverse in (5.4) and in Algorithm 6.1 is not fully justified.","section":"Section 5, Proposition 5.1"}],"minor_comments":[{"comment":"Equation (4.12) has a sign error on the right-hand side: as printed it gives d/dt(λTdot) = + ∇(...), which would imply Tddot = +∇(...) rather than the stated (4.5). The sign inconsistency between (4.12) and (4.5) should be corrected.","section":"Section 4, proof of Theorem 4.1, Eq. (4.12)"},{"comment":"The algorithm computes ξ as the solution of G(θl)ξ = pl via MINRES, but Eq. (5.4a) defines θdot = G†p. When G is singular, the two are not equivalent; the relation between the pseudo-inverse formulation and the linear solve used in the algorithm should be clarified, as should the role of the fixed-point parameter γ and the stopping criterion for the inner iteration.","section":"Algorithm 6.1, lines 6-8"},{"comment":"There are numerous typos, including \"physcis\" in the introduction and \"historgram\" in the captions of Figures 2, 4, and 5; the manuscript would benefit from a careful proofreading pass.","section":"Throughout"},{"comment":"The notation for the Fisher information is not consistent: FQ is used in (3.6)-(3.7) and in Appendix A, but the functional is also referred to as FQ in the text. The authors should define the notation once and use it consistently throughout.","section":"Sections 3.2 and 3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is heavily built on the authors' prior works [10] and [43], especially the PWHF framework; the genuinely new elements here are the inclusion of the Fisher-information term via Neural ODE and the application to the TDSE. The central closure gap (no error bound between the reduced parameterized dynamics and the full Wasserstein Hamiltonian flow) is inherited from the previous framework, and the numerical evidence is not yet strong enough to compensate. The editor may wish to ask the authors to state clearly which parts are new relative to [43] and to provide either an error estimate or a significantly more thorough numerical validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I think this is a genuinely useful paper, with a big caveat. The infinite-dimensional reformulation is clean, and the harmonic oscillator check is real. But the step from exact WHF to the neural-ODE reduction is an assertion, not a theorem, and the numerics don't close the gap.\n\nWhat's actually new: the paper packages known pieces — WHF for TDSE via Madelung, parameterized WHF from their own [43], Neural ODE — into a training-free sampling algorithm for the TDSE. The Fisher information estimator using Neural ODE is a reasonable and useful trick. Theorem 4.1 is derived cleanly; the push-forward variation argument is standard but correct. The affine harmonic oscillator example is a nice sanity check because it is one of the few settings where the reduced family is exactly invariant.\n\nSoft spots, in order. Section 5 rests entirely on [43]: Proposition 5.1 is imported without derivation, and the sentence \"Solving (5.4) can provide an approximate solution\" is doing a lot of work. For a generic submanifold Oθ there is no reason the projected Euler-Lagrange dynamics track the full flow. The paper itself lists the Wasserstein error bound as open in Section 8, so this is not a hidden flaw — it is acknowledged. The numerics are qualitative: histograms, no error norms, no baselines, and the 9D example has no ground truth. Hamiltonian preservation (Figure 8) is a structural property of the reduced system and would hold even far from the true trajectory. The self-citation load is heavy but not improper; [43] really does contain the parameterized machinery.\n\nI don't think there is a fatal error in the core derivation. The paper is honest about what is open. But the central algorithmic claim — controlled approximation of the TDSE density — is currently supported only by a special case and by plausibility.\n\nWho should read it: people working on particle/sampling methods for PDEs and on WHF. It would be a reasonable paper to cite for the WHF-TDSE reformulation, not yet for the numerical method's reliability. A serious referee could push for an error estimate (even a posteriori) or, failing that, a counterexample-style study of when the reduced flow leaves the true density. The paper deserves peer review, not desk rejection.","headline":"Clean WHF reformulation of the TDSE with a useful Neural-ODE Fisher trick, but the parameterized reduction's approximation claim is unproven and the numerics stay qualitative.","tokens_in":18147,"tokens_out":1904,"would_cite":true,"duration_ms":19596,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q41","49Q22","65P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the time-dependent Schrödinger equation can be reformulated as a Hamiltonian flow on the space of probability push-forward maps, and that parameterizing these maps with Neural ODEs produces a training-free…","keywords":["Schrödinger equation","Wasserstein Hamiltonian flow","Madelung transform","push-forward map","Neural ODE","generative model","optimal transport","high-dimensional numerical PDE"],"falsifier":"Run the proposed algorithm on a one- or two-dimensional linear Schrödinger equation whose exact density is sharply peaked and oscillates rapidly, compare the Wasserstein distance between the sample-based density $\\rho_\\theta(t)$ and the true $|\\psi(t,\\cdot)|^2$ over a long time interval, and monitor whether the distance remains small as the network width and sample count are increased; if the distance grows without bound, the parameterization does not capture the true evolution.","tokens_in":17059,"feed_emoji":"⚛️","tokens_out":12945,"duration_ms":102229,"temperature":0.7,"pith_summary":"The paper tries to establish a new numerical route to the time-dependent Schrödinger equation that avoids spatial grids. Using the Madelung transform, the quantum density and phase are governed by a Wasserstein Hamiltonian flow on the space of probability densities. The paper then restricts the dynamics to a family of push-forward maps realized by Neural ODEs, pulls back the Wasserstein metric onto the parameter space, and derives an ODE system for the network parameters. Solving this finite-dimensional ODE system yields approximate sample trajectories for the quantum density. The value of the claim is that it offers a potential way to handle high-dimensional systems where classical discretization methods fail.","feed_headline":"Neural ODE flows carry the Schrödinger density","feed_subtitle":"No grid, no training: the Schrödinger equation becomes a Hamiltonian ODE in network parameters.","key_machinery":"The load-bearing machinery is the Madelung transform (3.4), which maps the complex wavefunction to a density–phase pair, and the identification of the TDSE with the Wasserstein Hamiltonian flow whose Hamiltonian (3.6) contains the kinetic energy, the external potential functional, and Fisher information. The paper lifts this flow to the space of diffeomorphisms via the push-forward map $T$, proving the Euler–Lagrange equation (4.5). For computation, it restricts $T$ to Neural-ODE push-forward maps and forms the pullback metric $G(\\theta) = \\int \\partial_\\theta T_\\theta(z)^\\top \\partial_\\theta T_\\theta(z) \\lambda(z)\\, dz$. This metric converts the infinite-dimensional flow into the finite-dimensional Hamiltonian system (5.4) for the parameters $\\theta$ and momenta $p$. Neural ODEs are chosen because they are invertible and give convenient access to the logarithmic density and hence the Fisher information term.","core_discovery":"The central discovery is that the TDSE, via the Madelung transform and the identification of the quantum potential as the first variation of Fisher information, is exactly a Wasserstein Hamiltonian flow with Hamiltonian (3.6). Theorem 4.1 states that the critical point of the associated action satisfies the operator equation $\\ddot{T}(z) = -\\nabla_X \\frac{\\delta}{\\delta\\rho} \\mathcal{F}(T_\\sharp\\lambda)\\circ T(z)$, and that the push-forward density $T_\\sharp\\lambda$ solves the WHF. When $T$ is parameterized by Neural ODEs, the metric $G(\\theta)$ in parameter space turns this into the Hamiltonian ODE system (5.4). The paper presents numerical experiments on a harmonic oscillator, a Gross-Pitaevskii equation, and a three-particle interactive system in 9 dimensions, showing sample histograms that match analytic solutions and a preserved Hamiltonian. The paper stops short of proving that the parameterized system approximates the infinite-dimensional flow with a controlled error; it lists this as an open question.","pith_inferences":["If a rigorous Wasserstein error bound were established between the parameterized flow and the true TDSE density, the method would become a practical contender for dimensions around 4–10, where current grid methods struggle.","The same pullback-of-the-Wasserstein-metric construction could be applied to other Hamiltonian PDEs written as Wasserstein Hamiltonian flows, such as the Vlasov–Poisson system or mean-field game equations.","A natural test is to compare the algorithm against known solutions for wavefunctions with rapid phase oscillations, where the Fisher information term dominates and approximation errors would be most visible."],"forward_implications":["The method requires no spatial grid or basis; it only needs samples from a reference distribution and an ODE solver, so it scales naturally with dimension.","The numerical scheme is training-free: the parameters are advanced by a symplectic Euler update and linear solves, not by minimizing a loss.","The reformulation combines density evolution and particle trajectories in a single self-contained system, in contrast to guiding-wave formulations which require a separately supplied guiding field.","If the approach generalizes, it offers a sampling-based alternative for high-dimensional quantum dynamics beyond the dimensions reachable by spectral or finite-difference methods."],"supporting_citations":[{"why":"Supplies the Wasserstein Hamiltonian flow formulation used to reinterpret the Madelung system as a Hamiltonian flow on densities.","marker":"[10]"},{"why":"Provides the Neural ODE architecture, invertibility, and instantaneous change-of-variables formula used to compute the logarithm density and Fisher information.","marker":"[9]"},{"why":"Provides the parameterized Wasserstein Hamiltonian flow framework and the symplectic Euler scheme used for the parameter system (5.4).","marker":"[43]"},{"why":"Gives the optimal transport view of the Schrödinger equation that the paper draws on for the Hamiltonian (3.6).","marker":"[40]"},{"why":"Gives the quantum potential identity used to express the Fisher information variation.","marker":"[7]"},{"why":"Identifies the single-valued-phase assumption that removes the potential inequivalence between the Schrödinger equation and the Madelung system.","marker":"[41]"},{"why":"Supplies the mathematical treatment of the Madelung transform as a symplectic transformation.","marker":"[33]"}],"fun_headline_variants":["Wasserstein flow turns Schrödinger into a neural ODE","No grid: Schrödinger solved as Hamiltonian flow in parameters","Quantum density flows via Wasserstein Hamiltonian neural nets","Schrödinger's equation becomes a learned push-forward flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The finite-dimensional family of Neural-ODE push-forward maps is assumed to be expressive enough that the solution of the projected Hamiltonian system (5.4) stays close to the true infinite-dimensional flow; the paper offers no error bound, and its own conclusion lists 'can the error be bounded theoretically in the Wasserstein metric?' as an open question.","fun_headline_variants_meta":{"raw":{"variants":["Wasserstein flow turns Schrödinger into a neural ODE","No grid: Schrödinger solved as Hamiltonian flow in parameters","Quantum density flows via Wasserstein Hamiltonian neural nets","Schrödinger's equation becomes a learned push-forward flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000287,"raw_usage":{"total_tokens":1679,"prompt_tokens":929,"completion_tokens":750,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":684}},"tokens_in":545,"tokens_out":750,"duration_ms":6841,"temperature":1.0,"reasoning_tokens":684,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:48:35.130928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed algorithm on a one- or two-dimensional linear Schrödinger equation whose exact density is sharply peaked and oscillates rapidly, compare the Wasserstein distance between the sample-based density $\\rho_\\theta(t)$ and the true $|\\psi(t,\\cdot)|^2$ over a long time interval, and monitor whether the distance remains small as the network width and sample count are increased; if the distance grows without bound, the parameterization does not capture the true evolution.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Wasserstein Hamiltonian flow formulation used to reinterpret the Madelung system as a Hamiltonian flow on densities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Neural ODE architecture, invertibility, and instantaneous change-of-variables formula used to compute the logarithm density and Fisher information."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the parameterized Wasserstein Hamiltonian flow framework and the symplectic Euler scheme used for the parameter system (5.4)."},{"cited_title":"von Renesse , An optimal transport view of schr¨ odinger’s equation , Canadian Mathematical Bulletin, 55 (2012), p","cited_arxiv_id":null,"evidence_quote":"Gives the optimal transport view of the Schrödinger equation that the paper draws on for the Hamiltonian (3.6)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Identifies the single-valued-phase assumption that removes the potential inequivalence between the Schrödinger equation and the Madelung system."},{"cited_title":"Reddiger and B","cited_arxiv_id":null,"evidence_quote":"Supplies the mathematical treatment of the Madelung transform as a symplectic transformation."}],"review_version":1}