{"id":"938ee962-00bd-42f7-acea-4de3194d2dbb","arxiv_id":"2507.01313","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"This paper repackages the deep FBSDE algorithm as a Neural Hamiltonian Operator, proves a universal approximation result, and offers convergence theorems that assume, without evidence, the very properties needed.","lead":"This paper defines a 'Neural Hamiltonian Operator' around the existing deep FBSDE method for high-dimensional stochastic control, proving that such operators can approximate optimal dynamics. Its training-convergence guarantees depend on unverified assumptions, and the numerical tests lack comparisons to prior methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero terminal loss is achievable by the decoupling field of any admissible feedback control, so minimizing J does not select the optimal Hamiltonian operator; the central recovery claim is unsupported.","rationale":"The reader's weakest_assumption correctly identifies the identifiability condition in Theorem 4.3 as the load-bearing premise. My stress-test sharpens this: the assumption is not merely unverified but contradicted by the structure of the objective. Section 3.3's loss (5) is a terminal shooting condition for an FBSDE with a candidate feedback control. Standard FBSDE theory implies that every sufficiently regular admissible control admits a decoupling field matching the terminal boundary; optimality (Hamiltonian maximization) is an independent condition that never appears in J. Therefore the zero sublevel set of J generically contains non-optimal operators, and Theorems 4.3 and 4.6 only prove convergence to some zero-loss operator whose control need not be optimal. This does not invalidate Theorem 4.1, which approximates a given target operator, nor the numerical experiments as heuristics, but it invalidates the paper's stated central claim. A fix would require adding an explicit Hamiltonian-maximization residual or adversarial max term to the loss and re-deriving the guarantees. As written, the central recovery claim is unsupported, so the verdict should move from CONDITIONAL to REJECT.","tokens_in":21859,"tokens_out":7873,"duration_ms":107905,"concrete_test":"Construct the 1D LQ example dS_t=(a S_t+α_t)dt+dZ_t with running cost 1/2 α_t^2 and terminal cost 1/2 q_T S_T^2. Compute the optimal Riccati feedback α^*. Choose a suboptimal linear feedback α(t,s)=k s with k≠k^*. For this chosen α, solve the scalar semilinear PDE for its decoupling field v^α(t,s)=θ_t s+φ_t with terminal condition v^α(T,s)=q_T s. Substituting Φ=v^α into (3)-(5) gives J(α,v^α)=0 by construction while α≠α^*, directly falsifying Assumption 1. A numerical cross-check is to take the paper's Problem 1, freeze the control network at a non-optimal constant, and train only Φ_ξ: if the terminal loss converges to approximately zero while the control remains far from ∇V, the identifiability assumption is empirically violated.","verdict_should_be":"REJECT","load_bearing_attack":"The load-bearing problem is not merely that Assumption 1 of Theorem 4.3 is unproved; the assumption is generically false. The objective (5) enforces only the PMP terminal boundary condition, with no term enforcing the Hamiltonian-maximization condition. For a fixed admissible Markov control α, the candidate system (3)-(4) is a Markovian BSDE with generator ∇_sH(t,s,α(t,s),Φ,q_Ψ) and terminal condition ∇G. Under the smoothness assumptions of Section 2, this BSDE has a decoupling field v^α satisfying v^α(T,s)=∇G(s), with q=∇_s v^α σ. Taking Φ_ξ=v^α makes the terminal loss identically zero, J(α,v^α)=0, even though α need not maximize H and L_Ψ≠L^*. Thus Assumption 1's injectivity between terminal conditions and operator coefficients fails in ordinary smooth problems, including linear-quadratic ones. Remark 4.4 calls such operators 'spurious' and assumes them away, but they form a generic zero set of J. Consequently, minimizing the empirical risk cannot, by itself, identify the optimal feedback control, and the central claim that the NHO 'recovers the optimal feedback control and value-gradient dynamics' does not follow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates high-dimensional stochastic control through a Neural Hamiltonian Operator (NHO), a parameterized generator for the FBSDE system associated with Pontryagin's maximum principle. The parameters Ψ=(ω,ξ) define a candidate control 𝛼𝜔 and a decoupling-field ansatz Φ𝜉; the coupled system (3)-(4) is simulated forward, and training minimizes the terminal mismatch loss (5). Section 4 proves a universal approximation theorem for the NHO coefficients (Theorem 4.1) and two SGD convergence theorems (Theorems 4.3 and 4.6) under smoothness, Polyak-Łojasiewicz, and identifiability assumptions. Section 5 adapts the framework to ergodic control with a Hamiltonian-variance objective and Lyapunov regularization, and Section 6 reports d=50 numerical experiments on three problems.","tokens_in":22053,"tokens_out":11064,"duration_ms":131558,"significance":"The universal approximation theorem is a solid and useful result: it reduces the problem to C1 approximation of the optimal control, the value gradient, and its Jacobian, and it is independent of the training procedure. The operator-theoretic framing of deep FBSDE solvers is a helpful organizing perspective, and the numerical experiments demonstrate that the terminal-mismatch objective can be minimized in high dimensions. However, the central claim that the NHO recovers the optimal feedback control and value-gradient dynamics is not established. The loss (5) enforces only the terminal condition of the PMP; it does not enforce the Hamiltonian-maximization condition. The identifiability assumption that is needed to bridge this gap is not proved and is generically false in the idealized function class used by the approximation theorem. The convergence theorems are therefore conditional on a condition that fails in the intended setting, and the paper's main claim is unsupported.","major_comments":[{"comment":"The terminal mismatch loss (5) contains no term that enforces the PMP Hamiltonian-maximization condition. For any fixed admissible Markov feedback control α, the Markovian BSDE with generator ∇_sH(t,s,α(t,s),v,∇_s v σ) and terminal condition v(T,s)=∇G(s) has a decoupling field v^α under the smoothness assumptions of Section 2. Taking Φξ=v^α and 𝛼𝜔=α in (3)-(4) yields p̃_T=v^α(T,S_T)=∇G(S_T), so J(Ψ)=0 even though L_Ψ is the operator associated with the non-optimal control α. This is not a measure-zero 'spurious' phenomenon: it holds for every admissible α, including the zero control in a linear-quadratic problem. Thus Assumption 1, which the paper calls 'subtle yet crucial' in Remark 4.4, is false in the continuum limit of the network class used by Theorem 4.1, and the observability remark does not supply a replacement condition. The paper needs either a proof of identifiability under explicit conditions or a different objective that includes the maximization condition.","section":"§3.3, Eq. (5); §4.1, Assumption 1"},{"comment":"Both convergence theorems conclude that the expected loss tends to the minimum value, lim_k E[J(Ψ_k)]=0 or lim_k E[J_λ(Ψ_k)]=J_λ(Ψ*). These are statements about the training objective, not about the learned operator. The abstract and Section 1 interpret the minimizer as recovering the optimal feedback control and value-gradient dynamics; that interpretation requires the injectivity asserted in Assumption 1. Since the construction in the previous comment shows that J=0 is attained by non-optimal operators, the convergence results do not substantiate the recovery claim. The theorems should be stated as conditional results about loss minimization under the stated assumptions, and the paper should not claim recovery of the optimal control unless an identifiability result is proved.","section":"§4.1, Theorem 4.3; §4.2, Theorem 4.6"}],"minor_comments":[{"comment":"The statement approximates 𝛼𝜔 to 𝛼∗ but does not specify how the network output is constrained to the compact set K. A projection onto K or a suitable output-layer design should be included so that μ, σ, and f are evaluated on their domain; otherwise the coefficients bΨ and ΣΨ are not even defined for arbitrary network outputs.","section":"Theorem 4.1"},{"comment":"The initial-state distribution for the expectation in J is not specified. Remark 4.4's identifiability argument explicitly relies on a rich distribution over initial states; the definition of J should state the distribution (e.g., a sampling measure ν on R^d).","section":"§3.3, Eq. (5)"},{"comment":"The smoothing parameter ε used to approximate the non-smooth cost |α|^(3/2) is not reported; include the numerical value for reproducibility.","section":"Remark 6.1"},{"comment":"The legend entries 'Reg(0.1·||p||²)' and 'Reg(0.01·||p||²)' do not match the regularizer in Eq. (6), which is an integrated penalty on ||∇_s Φ_ξ||²_F; align the notation or explain the abbreviation.","section":"Figures 2, 4, and 6"},{"comment":"The conclusion states the method is validated 'in up to 100 dimensions,' but all reported experiments are in d=50; correct the discrepancy.","section":"Section 7"}],"recommendation":"reject","confidential_remarks":"The central issue is not presentation: the terminal-mismatch objective is not identifiable, and the identifiability assumption needed for the recovery claim is generically false in the intended setting. A revised manuscript would need to change the learning objective or substantially narrow the claims. I would be open to reconsidering if the authors add a Hamiltonian-maximization term or a rigorous identifiability condition, but in its current form the paper's main theorem cannot support the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does two things well: it gives a crisp operator-theoretic framing of the deep FBSDE method, and it proves a C^1 universal approximation theorem for the parameterized family. The extension to continuous-martingale drivers is largely cosmetic but harmless. What it does not do is back its central claim—that training on the terminal-mismatch loss J(Ψ)=E||p_T − ∇G(S_T)||^2 recovers the optimal feedback control.\n\nThe stress-test note is right. The loss only enforces the PMP terminal boundary. For any admissible Markov control α, the corresponding FBSDE has a decoupling field v^α, and setting Φ = v^α makes J identically zero. So the zero set of J is huge; it contains an infinite-dimensional family of spurious operators. Assumption 1 in Theorem 4.3 asserts that zero loss forces the operator coefficients to equal the true optimal ones, but that is generically false, even in smooth linear-quadratic problems. The paper's Remark 4.4 calls such operators 'spurious' and assumes them away, but they are not a measure-zero nuisance—they are a generic phenomenon. Without identifiability, the SGD convergence theorems, rigorous as they are, are theorems about converging to some point in the zero set. The universal approximation theorem does not rescue this.\n\nCredit where due: the paper is honest about the gap. Remark 2.2 flags the classical-solution limitation, Conjecture 5.5 lays out the viscosity roadblock, and Remark 4.4 explicitly calls identifiability 'subtle yet crucial' and leaves it open. The writing is clear and the appendices are standard but complete. The numerical section is thin: no baselines, no code, no error bars, and the conclusion says 'up to 100 dimensions' while the experiments use d=50. That is a minor point next to the identifiability problem.\n\nBottom line: this is a framework paper with a real structural flaw in its main argument. It deserves a serious referee because the problem is important and the formalization is a useful step. But the author needs to either add the Hamiltonian-maximization condition to the loss or prove some restricted identifiability result. As it stands, the abstract's claim that the optimal NHO 'can be found' by this training procedure is unsupported.","headline":"Clean operator-theoretic framing of deep FBSDE, but the central recovery claim fails: the terminal-mismatch loss has a huge zero set that does not single out the optimal Hamiltonian operator.","tokens_in":22611,"tokens_out":3501,"would_cite":false,"duration_ms":39445,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E20","60H30","68T07","49L20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that high-dimensional stochastic control can be solved as a statistical inference problem: training the Neural Hamiltonian Operator—neural networks for the feedback control and the value-function gradient—on the terminal…","keywords":["Neural Hamiltonian Operator","Pontryagin's Maximum Principle","Deep FBSDE","High-Dimensional Stochastic Control","Universal Approximation","Operator Learning","Stochastic control","Backward stochastic differential equations"],"falsifier":"On a scalar problem with an even terminal payoff, so that $G(s)=G(-s)$, train two NHOs from different initializations until $J(\\Psi)$ reaches machine zero; if the two parameter sets imply measurably different controls or state paths from the same initial state, the identifiability assumption behind Theorem 4.3 is false. A complementary check is to run the same training on a problem whose known value function is only a viscosity solution, not classically differentiable, and compare the implied control with the known optimal one.","tokens_in":21569,"feed_emoji":"🎛️","tokens_out":12028,"duration_ms":188690,"temperature":0.7,"pith_summary":"Stochastic control problems in high dimensions are usually attacked either by solving the Hamilton-Jacobi-Bellman equation, which suffers from the curse of dimensionality, or by converting the problem into a coupled forward-backward stochastic differential equation via Pontryagin's Maximum Principle. This paper claims that the FBSDE route can be viewed as a statistical inference problem: one defines a Neural Hamiltonian Operator (NHO), a differential operator whose coefficients are neural networks for the feedback control and for the spatial gradient of the value function, and trains those networks by minimizing the squared mismatch between the simulated adjoint process at terminal time and the gradient of the terminal payoff. The paper proves that the family of NHOs is dense in the space of true Hamiltonian operators, so approximation error can be made arbitrarily small, and proves convergence of stochastic gradient descent under identifiability, smoothness, and Polyak-Lojasiewicz-type assumptions. If these claims hold, deep FBSDE solvers acquire a rigorous footing: they are doing nonparametric inference of an unknown generator of optimal dynamics from simulated trajectories, and the remaining barriers are optimization and stability rather than approximation.","feed_headline":"Terminal mismatch alone recovers optimal controls in high dimensions","feed_subtitle":"A new operator view gives deep FBSDE solvers a statistical-inference footing for high-dimensional stochastic control.","key_machinery":"The central object is the Neural Hamiltonian Operator $L_\\Psi$, the infinitesimal generator of the forward SDE for the extended state $X_t=(S_t,\\tilde p_t)$. On a test function $g$, it acts as $(L_\\Psi g)(t,x)=\\nabla_x g(x)^\\top b_\\Psi(t,s)+\\frac{1}{2}\\mathrm{Tr}(D_\\Psi(t,s)\\nabla_x^2 g(x))$, where $b_\\Psi$ stacks the drift of the state and the candidate adjoint, and $D_\\Psi=\\Sigma_\\Psi C_t \\Sigma_\\Psi^\\top$ is built from the diffusion blocks $\\Sigma_\\Psi=(\\sigma(\\cdot,\\alpha_\\omega)^\\top, q_\\Psi^\\top)^\\top$ with $q_\\Psi=(\\nabla_s\\Phi_\\xi)\\sigma(\\cdot,\\alpha_\\omega)$. This operator carries the whole argument because its coefficients are exactly the feedback control $\\alpha_\\omega$ and the decoupling-field ansatz $\\Phi_\\xi\\approx\\nabla_s V$, so minimizing the terminal mismatch over $L_\\Psi$ is equivalent to searching over candidate optimal dynamics.","core_discovery":"The paper's central claim is that the optimal feedback control of a stochastic control problem coincides with the dynamics of the unique Neural Hamiltonian Operator that satisfies the terminal condition of the PMP. Concretely, for network parameters $\\Psi=(\\omega,\\xi)$, the candidate adjoint is $\\tilde p_t$, initialized as $\\Phi_\\xi(t_0,s_0)$ and evolved forward through the SDE whose generator is $L_\\Psi$, while the terminal condition requires $\\tilde p_T=\\nabla G(S_T)$. Training minimizes $J(\\Psi)=\\mathbb{E}[\\|\\tilde p_T-\\nabla G(S_T)\\|^2]$, and the paper asserts that a minimizer of this loss gives an operator whose coefficients match those of the true optimal Hamiltonian system. Theorem 4.1 states that the NHO family is dense in the space of true Hamiltonian operators in coefficient sup-norm for any continuous square-integrable martingale driver, and Theorems 4.3 and 4.6 state that SGD reaches the optimum in expectation under identifiability, $L$-smoothness, and global or local Polyak-Lojasiewicz conditions. The extensions to ergodic control add a loss based on the variance of the Hamiltonian along paths, whose zero minimizer is shown to imply a constant Hamiltonian almost surely under the invariant measure.","pith_inferences":["Beyond the paper, a trained NHO could be reused as an initialization for a different horizon or terminal cost, because the generator is local in time; the paper does not test this transfer.","Beyond the paper, disagreement among multiple near-zero-loss initializations would be a practical red flag for the identifiability assumption, suggesting the terminal condition alone is too weak and the initial-state distribution should be widened.","Beyond the paper, the classical-solution assumption implies the method is most fragile in viscosity regimes; smoothing a non-differentiable terminal cost and watching whether the learned control tracks the unique optimal one would test that boundary.","Beyond the paper, the Lyapunov regularizer in the ergodic loss can serve as an online stability certificate: estimating the drift of $\\|S_t\\|^2$ from simulated paths and checking that it stays negative would certify ergodicity of a learned policy."],"forward_implications":["Under Theorem 4.1, any sufficiently smooth optimal Hamiltonian system can be matched to arbitrary precision by an NHO, so in the smooth regime the method carries no irreducible approximation error.","When the identifiability, smoothness, and Polyak-Lojasiewicz assumptions hold, SGD on the terminal mismatch converges in expectation, making the deep FBSDE training procedure a provable statistical estimator rather than a heuristic.","Because the approximation and operator definitions only require a continuous square-integrable martingale driver, the same training scheme applies to models whose noise is not Brownian, including drivers with stochastic covariance.","In the ergodic infinite-horizon setting, a zero of the Hamiltonian-variance loss forces the learned Hamiltonian to be constant almost surely under the invariant measure, giving a necessary-condition check for learned policies.","The reported $d=50$ experiments on a known analytic control, a double-well potential, and portfolio liquidation show the terminal-loss training reaching small values while the implied control matches reference behavior."],"supporting_citations":[{"why":"Supplies the original deep BSDE shooting method that the NHO training objective is built on.","marker":"E, Han and Jentzen (2017)"},{"why":"Introduces the high-dimensional deep FBSDE solver whose computational scheme the NHO formalizes.","marker":"Han, Jentzen and E (2018)"},{"why":"Provides the PMP and FBSDE theory, including the identification of the adjoint process with the value-function gradient.","marker":"Yong and Zhou (1999)"},{"why":"Underlies Lemma A.1, the C^1 simultaneous approximation of a function and its Jacobian used in Theorem 4.1.","marker":"Hornik (1990)"},{"why":"Supplies the Polyak-Lojasiewicz condition that Theorems 4.3 and 4.6 use to prove SGD convergence.","marker":"Polyak (1963)"},{"why":"Provides the Foster-Lyapunov stochastic-stability criterion used in Theorem 5.3 for ergodicity.","marker":"Khasminskii (2011)"},{"why":"Supplies the Markov-chain stability theory also invoked in Theorem 5.3.","marker":"Meyn and Tweedie (2009)"},{"why":"Gives the monotone-scheme convergence framework that Conjecture 5.5 says a proof would need.","marker":"Barles and Souganidis (1991)"},{"why":"Defines viscosity solutions that the paper's Conjecture 5.5 aims to connect to.","marker":"Crandall, Ishii and Lions (1992)"}],"fun_headline_variants":["Neural Hamiltonian Operator learns optimal controls from terminal mismatch","Operator view gives deep FBSDE solvers statistical footing","Terminal condition alone defines optimal Neural Hamiltonians","Learn high-dim optimal controls via a single terminal loss","Neural Hamiltonian Operators: universal approximation for stochastic control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that zero terminal mismatch can only be produced by the true optimal dynamics: no other parameter setting can satisfy the end condition by coincidence, and the value function is assumed to be genuinely differentiable so that the adjoint process really is its gradient. If either part fails, a network can drive the loss to zero while generating dynamics that are not the optimal control.","fun_headline_variants_meta":{"raw":{"variants":["Neural Hamiltonian Operator learns optimal controls from terminal mismatch","Operator view gives deep FBSDE solvers statistical footing","Terminal condition alone defines optimal Neural Hamiltonians","Learn high-dim optimal controls via a single terminal loss","Neural Hamiltonian Operators: universal approximation for stochastic control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000585,"raw_usage":{"total_tokens":2773,"prompt_tokens":994,"completion_tokens":1779,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":1703}},"tokens_in":610,"tokens_out":1779,"duration_ms":14830,"temperature":1.0,"reasoning_tokens":1703,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:55:12.703886+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a scalar problem with an even terminal payoff, so that $G(s)=G(-s)$, train two NHOs from different initializations until $J(\\Psi)$ reaches machine zero; if the two parameter sets imply measurably different controls or state paths from the same initial state, the identifiability assumption behind Theorem 4.3 is false. A complementary check is to run the same training on a problem whose known value function is only a viscosity solution, not classically differentiable, and compare the implied control with the known optimal one.","supporting_citations":[{"cited_title":", Jentzen , Arnulf A","cited_arxiv_id":null,"evidence_quote":"Introduces the high-dimensional deep FBSDE solver whose computational scheme the NHO formalizes."},{"cited_title":"Zhou , Xun Yu X","cited_arxiv_id":null,"evidence_quote":"Provides the PMP and FBSDE theory, including the identification of the adjoint process with the value-function gradient."},{"cited_title":"( 1990 )","cited_arxiv_id":null,"evidence_quote":"Underlies Lemma A.1, the C^1 simultaneous approximation of a function and its Jacobian used in Theorem 4.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Polyak-Lojasiewicz condition that Theorems 4.3 and 4.6 use to prove SGD convergence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Foster-Lyapunov stochastic-stability criterion used in Theorem 5.3 for ergodicity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Markov-chain stability theory also invoked in Theorem 5.3."},{"cited_title":"Souganidis , Panagiotis E","cited_arxiv_id":null,"evidence_quote":"Gives the monotone-scheme convergence framework that Conjecture 5.5 says a proof would need."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines viscosity solutions that the paper's Conjecture 5.5 aims to connect to."}],"review_version":1}