Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Fine-tuning a diffusion model under general costs is equivalent to solving a forward-backward SDE system.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 09:42 UTC pith:D52T4KUO

load-bearing objection Good framing, but the main theorem is false as stated: quadratic rewards break the bounded-gradient claim and the BSDE existence proof. the 4 major comments →

arxiv 2607.22660 v1 pith:D52T4KUO submitted 2026-06-28 math.OC math.PR

Generalized Fine-Tuning of Diffusion Models via Stochastic Control and FBSDEs

Zirui Wang , Lu Wang This is my paper
classification math.OC math.PR MSC 49L2560H1093E20
keywords diffusion modelsfine-tuningstochastic controlforward-backward SDEHamilton-Jacobi-Bellmanoptimal controlrobust controlreward collapse
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Fine-tuning a pretrained diffusion model is usually formulated as entropy-regularized control with a quadratic penalty on the control. This paper attempts to broaden that to an arbitrary running cost ℓ(t,x,u), yielding a relative generalized path cost that includes KL divergence and Wasserstein-like transport costs as special cases. It claims that the resulting fully nonlinear HJB equation has a unique classical solution, and that the value function is the Y-component of a coupled forward-backward SDE whose generator is the minimized Hamiltonian inf_u{ℓ+u·z/σ̃}. If true, this gives a unified solution recipe — compute (Y,Z) with a deep BSDE solver and read off the optimal feedback control — and extends to robust fine-tuning under model ambiguity by adding a δ||Z|| penalty.

Core claim

The paper's central claim is that the generalized fine-tuning problem — minimize E[∫ℓ(t,X,u)dt + r(X_T)] over controls u driving the reverse diffusion — is well-posed and solvable by FBSDE. Theorem 3.1 asserts that under H1-H7 the HJB equation has a unique classical solution v∈C^{1,2} with globally bounded gradient and Hessian, and a unique globally Lipschitz optimal feedback u*(t,x)=argmin_u{ℓ(t,x,u)+u·∇v(t,x)}. Theorem 3.3 identifies v with the Y-component of the FBSDE dX_t=eb dt+σ̃ dB_t, dY_t=-inf_u{ℓ+u·Z/σ̃}dt+Z dB_t, Y_T=r(X_T), and the optimal control as the argmin in the generator. The framework unifies entropy-regularized fine-tuning (quadratic ℓ), sparsity-inducing costs λ|u|₁, time

What carries the argument

The central object is the running cost ℓ and the induced relative generalized path cost Cℓ(P‖Q)=inf_{u:L(X^u)=P} E^{Q^u}[∫ℓ dt]. The argument turns on the fully nonlinear Hamiltonian H(t,x,z)=inf_u{ℓ(t,x,u)+u·z/σ̃(t)}. The usual decoupling-field argument supplies the value function, and the BSDE's Z-process — Z_t=σ̃(t)∇v(t,X_t) — is the key bridge: it serves simultaneously as the martingale representation of the backward process and the variable through which the optimal control is expressed. In the ambiguity extension, the support function δ||Z|| is added to the generator.

Load-bearing premise

The proof of BSDE existence assumes that the reward's exponential moments exp(γ|r(X_T)|) are finite for every γ, and it tries to derive this from r∈C^{2+α} with bounded second derivatives; that derivation is invalid because such a reward can grow quadratically (e.g., ||x||²), for which the required exponential moments are not finite.

What would settle it

Take r(x)=||x||², which satisfies the stated smoothness assumption H4, and take a simple diffusion with Gaussian marginal (e.g., linear drift and constant diffusion). Then E exp(γ||X_T||²) is infinite for large γ, so the terminal condition fails the exponential-moment condition required by the quadratic-growth BSDE theorem invoked in the proof of Theorem 3.2. Running the proposed BSDE-FT algorithm on this reward and observing non-convergence or non-uniqueness of the deep solution would settle whether the FBSDE representation holds under the stated assumptions.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Entropy-regularized fine-tuning with ℓ=(α/2)||u/σ̃||² is recovered as the special case where the HJB equation is semilinear and the optimal control is σ̃²∇v/α; the FBSDE framework then contains the existing approach.
  • A single deep BSDE solver can handle non-quadratic costs such as ℓ=λ|u|₁ and ℓ=β(t)||u||², where no analytic argmin is available, as long as the generator remains tractable.
  • The optimal control is nonlinear feedback through Z_t, so the BSDE gradient component is the sufficient statistic for fine-tuning; users do not need to solve a separate PDE.
  • Robust fine-tuning under uncertainty about the score model is formulated as a minimax BSDE with generator +δ||Z||, and the worst-case perturbation is aligned with ∇v; this yields robust controls without changing the numerical scheme.
  • Well-posedness of the generalized fine-tuning objective follows from the claimed uniqueness of the classical HJB solution, guaranteeing that the reward-collapse trade-off has a well-defined optimum.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to benchmark the LQI example ℓ=(α/2σ̃²)||u||²+θ1·u+θ2 against a direct HJB solver in low dimension; agreement of Y0 would confirm the FBSDE characterization beyond the formal proof.
  • The framework implicitly suggests that any divergence expressible as an optimal control path cost — not just KL and p-Wasserstein — can be fine-tuned by the same machinery; a concrete next step is to write f-divergence penalties in this form.
  • The exponential-integrability gap in the existence proof leaves the deep BSDE algorithm without full theoretical support for rewards that grow quadratically; adding a bounded-reward assumption or explicit growth bound is a minimal amendment.
  • The robust formulation implies an adversarial interpretation: fine-tuning against the worst-case drift perturbation δ·sign(∇v) could be used to quantify how reward optimization is degraded by score-estimation error, which a practitioner could measure empirically.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a generalized stochastic-control framework for fine-tuning diffusion models. A general running cost ℓ(t,x,u) is introduced, giving a "relative generalized path cost" C_ℓ(P∥Q) that is claimed to include KL divergence and Wasserstein distance as special cases. The value function of the resulting Generalized Fine-Tuning (GFT) problem is asserted to be the unique classical C^{1,2} solution of a fully nonlinear HJB equation (Theorem 3.1), to admit a BSDE representation of the form (3.18)–(3.19) (Theorem 3.3), and to satisfy the existence/uniqueness estimates of Theorem 3.2. A deep BSDE algorithm and a robust-ambiguity extension are also presented.

Significance. The high-level idea — encode general divergence penalties through a running cost and solve the HJB equation via FBSDEs — is attractive and, if rigorously established, would provide a unified view of entropy-regularized and transport-based fine-tuning. The paper also makes a concrete algorithmic proposal (BSDE-FT). However, the central mathematical claims are not proven as stated. In fact, Theorem 3.1(i) fails on a simple admissible one-dimensional example with quadratic reward, and the proof of Theorem 3.2 uses a linear-growth/integrability assertion that is false under H4. These are load-bearing errors: the FBSDE representation and the proposed solution method rest on them. No numerical experiments or applications are included, so the algorithmic contribution is not demonstrated.

major comments (4)
  1. [§3.2, Theorem 3.1(i) / Eq. (3.7)] Theorem 3.1(i) claims that, under H1–H7, the HJB equation (3.7) has a unique classical solution with ∥∇_x v∥_∞ ≤ L_1. This is false under the stated assumptions. Take d=1, b(t,x)=0, σ̃(t)=1, ℓ(t,x,u)=u²/2, and r(x)=x². Assumptions H1–H7 hold (H4 holds because r has bounded second derivative; H2 holds with p=2). The HJB equation is v_t + ½v_xx − ½(v_x)² = 0, v(T,x)=x². Its explicit classical solution is v(t,x)=x²/(1+2(T−t)) + ½ log(1+2(T−t)), whose gradient 2x/(1+2(T−t)) is unbounded in x. Thus the asserted a priori bound is not merely unproven; it is wrong for admissible data. This also breaks the claimed regularity needed for the FBSDE representation and for the feedback formula (3.8).
  2. [§3.3, Step 2 / Eq. (3.16)] The proof of Theorem 3.2 asserts that H4 implies |r(x)| ≤ L_0^r∥x∥ + |r(0)| and hence E[exp(γ|r(X_T)|)] < ∞ for every γ>0. This is incorrect: H4 only requires r∈C^{2+α} with bounded second derivatives, which permits quadratic growth, e.g. r(x)=∥x∥². For such r and a Gaussian X_T, E[exp(γ|r(X_T)|)] = ∞ for γ ≥ 1/(2T), so (3.16) is false. The cited Briand–Hu existence theorem for quadratic BSDEs is therefore not applicable under the stated hypotheses. The theorem needs an explicit sublinear-growth or bounded-reward condition, or a different integrability argument.
  3. [§3.3, Step 1 and Assumption H2] Theorem 3.2 is stated for H1–H7, but H2 only assumes coercivity with p≥1. In Step 1 the proof silently upgrades to p≥2, writing "By Assumption (H2), ... for p≥2." Without p≥2 the Legendre–Fenchel conjugate of ℓ can grow faster than quadratically in z, and the quadratic-growth BSDE theory invoked in the proof does not apply. This is another mismatch between the assumptions and the proof. At minimum the theorem must restrict H2 to p≥2; but even with that restriction, Major Comment 2 shows the result still fails for r(x)=∥x∥².
  4. [Proof of Theorem 3.1, Steps 1–3] The fixed-point proof is incomplete in several places. Step 1 asserts the existence of a decoupling field v^u∈C^{1,2} with uniform bounds ∥∇_x v^u∥≤C_1(R), ∥∇²_xx v^u∥≤C_2(R) for every u in the weighted Lipschitz ball B_R. No assumptions are given that guarantee these bounds; indeed, the example in Major Comment 1 shows that such uniform bounded-gradient bounds fail for a valid problem. Step 2 then defines R_0 through an implicit inequality involving C_2(R), but no proof is given that this inequality has a solution. Lemma 3's gradient stability estimate is asserted by passing from difference quotients to a uniform sup over x without controlling the constant independently of the initial point; as written, the contraction argument does not rigorously close. The proof therefore does not establish the theorem's central well-posedness claim.
minor comments (5)
  1. [Definition 3.1, Eq. (3.1)] The SDE (3.1) is typeset incorrectly: "+eσ(t), dB_t" should read "+eσ(t) dB_t".
  2. [Assumption H5, Eq. (3.3)] The first inequality is written with ∥x_1−x_2∥² but is called an L-Lipschitz condition; it should be ∥x_1−x_2∥ (or the terminology should be adjusted).
  3. [Proposition 3.2, Wasserstein case] The Wasserstein case assumes σ̃(t)≡0, which is incompatible with the uniform ellipticity in H7. The equivalence to the p-Wasserstein distance is only sketched and requires separate assumptions; as stated it is outside the paper's framework.
  4. [§3.4.1, Eq. (3.22)–(3.24)] The algebra for the linear-quadratic-interaction cost appears to be incorrect: θ_2(t,x) should not be inside the squared norm in f, and the optimal control should be u*(t)=−(σ̃²/α)θ_1(t,X_t)−(σ̃/α)Z_t, with the sign of the θ_1 term checked. The current Eq. (3.22)–(3.24) is not consistent with a direct minimization of (3.21) plus u·z/σ̃.
  5. [Theorem 5.2, Proof] The proof assumes "Since σ̃(t)>0 is a scalar", but H7 allows a matrix σ̃(t). For matrix σ̃, the support function sup_{∥η∥≤δ} η·(σ̃∇v) equals δ∥σ̃^T∇v∥, not δ∥σ̃∇v∥. The robust HJB equation and the candidate η* need to be corrected for the matrix case.

Circularity Check

0 steps flagged

No significant circularity: the FBSDE representation is a standard Itô-formula equivalence and the derivation relies on external, independent theorems.

full rationale

The paper's derivation chain is: define the generalized path cost C_l(P||Q) and the GFT control problem; derive the HJB equation from the dynamic programming principle (Proposition 3.3); prove regularity and well-posedness via a fixed-point argument citing external FBSDE results (Ma-Yong-Zhang, Delarue); prove BSDE existence using Briand-Hu; and finally obtain the FBSDE representation by applying Itô's formula to v(t,X_t) and setting Z_t = sigma(t) grad v(t,X_t). None of these steps substitutes a fitted parameter for a prediction, renames an input as an output, or imports an unverified uniqueness theorem from the authors' own prior work. The generalized path cost is introduced as a definition, and the KL and Wasserstein identifications are checked against external benchmarks rather than being assumed. The cited results (Pardoux-Peng, Briand-Hu, Ma-Yong-Zhang, Delarue, Kobylanski) are standard external theorems, not self-citations. The skeptical note about H4 allowing quadratic rewards is a real mathematical-assumption gap: H4 does not imply the linear growth |r(x)| <= L_0||x||+|r(0)| used in Theorem 3.2, and the claimed exponential integrability can fail for r(x)=||x||^2. However, this is a correctness/rigor defect in the stated assumptions, not circularity: the theorem's conclusion is not an input of its proof, even if the proof as written is invalid for admissible data. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The paper introduces no fitted constants; α and δ are arbitrary hyperparameters, not data-fitted. The main debts are unproven or silently strengthened assumptions: the false linear-growth inference from H4, the silent p≥2 upgrade, and the reliance on external theorems whose hypotheses are not verified.

free parameters (2)
  • α
    Positive hyperparameter in the linear-quadratic and product-quadratic running costs (3.21), (3.25). It is symbolic and arbitrary; no value is fitted or calibrated.
  • δ
    Worst-case ambiguity radius in the robust formulation (5.5). The paper states it 'could be fixed and selected arbitrarily or adaptively', so it is a free model parameter with no principled assignment.
axioms (5)
  • domain assumption ℓ(t,x,·) is strongly convex, lower semicontinuous, and coercive with ℓ≥c1||u||^p−c2 (H1-H2).
    This is the main structural assumption on the running cost; it guarantees minimizers exist and the generalized path cost Cℓ is well-defined.
  • ad hoc to paper H4 (r∈C^{2+α} with bounded second derivatives) implies |r(x)|≤L||x||+C and E exp(γ|r(X_T)|)<∞ for all γ.
    This implication is used in Theorem 3.2, Step 2, but is false: quadratic rewards satisfy H4 and violate the exponential-integrability claim. The theorem needs a separate boundedness or linear-growth assumption.
  • ad hoc to paper Coercivity exponent p≥2 in H2.
    H2 is stated with p≥1, but Theorem 3.2's proof silently assumes p≥2 so that the conjugate h grows at most quadratically in z. For 1≤p<2 the cited Briand-Hu quadratic BSDE theory does not apply.
  • domain assumption σ̃ is uniformly elliptic on the whole interval [0,T] (H7.2).
    Uniform ellipticity is used for FBSDE/PDE regularity and for the sub-Gaussian estimates in Theorem 3.2. Many diffusion schedules use noise that vanishes at endpoints, so the assumption may fail in practice.
  • standard math External quadratic-BSDE theorems (Briand-Hu 2006), the four-step scheme (Ma-Yong-Zhang), and the Delarue regularity results apply to the constructed FBSDE.
    The paper invokes these results without checking all of their hypotheses; several hypothesis checks are precisely where the stated assumptions fail (p≥2, terminal integrability).

pith-pipeline@v1.3.0-alltime-deepseek · 16036 in / 22800 out tokens · 202682 ms · 2026-08-02T09:42:02.446617+00:00 · methodology

0 comments
read the original abstract

We propose a generalized fine-tuning framework for diffusion models from the perspective of stochastic control. Beyond entropy-regularized formulations, we introduce a general running cost that induces a relative generalized path cost, encompassing both Kullback-Leibler divergence and optimal transport metrics as special cases. This leads to a fully nonlinear Hamilton-Jacobi-Bellman equation. We characterize the value function through a forward-backward stochastic differential equation system and establish existence and uniqueness under standard regularity conditions. The optimal control admits a nonlinear feedback form driven by the backward stochastic differential equation gradient component. Finally, we show that generalized fine-tuning naturally introduces an additional gradient-dependent penalty, providing a unified framework for diffusion fine-tuning under general distributional constraints.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 8 linked inside Pith

  1. [1]

    Deep un- supervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep un- supervised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. pmlr, 2015

    2015

  2. [2]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Er- mon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  3. [3]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  4. [4]

    Im- agen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Grit- senko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Im- agen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022. 21

    Pith/arXiv arXiv 2022

  5. [5]

    Equiv- ariant diffusion for molecule generation in 3d

    Emiel Hoogeboom, Vıctor Garcia Satorras, Cl´ ement Vignac, and Max Welling. Equiv- ariant diffusion for molecule generation in 3d. InInternational conference on machine learning, pages 8867–8887. PMLR, 2022

    2022

  6. [6]

    Geod- iff: A geometric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923, 2022

    Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geod- iff: A geometric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923, 2022

    Pith/arXiv arXiv 2022

  7. [7]

    Score-based diffusion models for accelerated mri

    Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical image analysis, 80:102479, 2022

    2022

  8. [8]

    Measurement-conditioned denoising diffusion proba- bilistic model for under-sampled medical image reconstruction

    Yutong Xie and Quanzheng Li. Measurement-conditioned denoising diffusion proba- bilistic model for under-sampled medical image reconstruction. InInternational Confer- ence on Medical Image Computing and Computer-Assisted Intervention, pages 655–664. Springer, 2022

    2022

  9. [9]

    Generative adversarial nets.Ad- vances in neural information processing systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Ad- vances in neural information processing systems, 27, 2014

    2014

  10. [10]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  11. [11]

    On fast sampling of diffusion probabilistic models.arXiv preprint arXiv:2106.00132, 2021

    Zhifeng Kong and Wei Ping. On fast sampling of diffusion probabilistic models.arXiv preprint arXiv:2106.00132, 2021

    Pith/arXiv arXiv 2021

  12. [12]

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation

    Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  13. [13]

    One-step diffusion with distribution matching distillation

    Tianwei Yin, Micha¨ el Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613–6623, 2024

    2024

  14. [14]

    Training diffusion models with reinforcement learning.arXiv preprint arXiv:2305.13301, 2023

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning.arXiv preprint arXiv:2305.13301, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Dpok: Reinforce- ment learning for fine-tuning text-to-image diffusion models.Advances in Neural Infor- mation Processing Systems, 36:79858–79885, 2023

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforce- ment learning for fine-tuning text-to-image diffusion models.Advances in Neural Infor- mation Processing Systems, 36:79858–79885, 2023

    2023

  16. [16]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500–22510, 2023

    2023

  17. [17]

    Directly fine-tuning diffusion models on differentiable rewards.arXiv preprint arXiv:2309.17400, 2023

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards.arXiv preprint arXiv:2309.17400, 2023. 22

    Pith/arXiv arXiv 2023

  18. [18]

    Fine- tuning of continuous-time diffusion models as entropy-regularized control.arXiv preprint arXiv:2402.15194, 2024

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Fine- tuning of continuous-time diffusion models as entropy-regularized control.arXiv preprint arXiv:2402.15194, 2024

    Pith/arXiv arXiv 2024

  19. [19]

    Fine-tuning of diffusion models via stochastic control: entropy regularization and beyond, 2025

    Wenpin Tang and Fuzhong Zhou. Fine-tuning of diffusion models via stochastic control: entropy regularization and beyond, 2025

    2025

  20. [20]

    Conjugate convex functions in optimal stochastic control.Journal of Mathematical Analysis and Applications, 44(2):384–404, 1973

    Jean-Michel Bismut. Conjugate convex functions in optimal stochastic control.Journal of Mathematical Analysis and Applications, 44(2):384–404, 1973

    1973

  21. [21]

    Adapted solution of a backward stochastic differential equation.Systems & Control Letters, 14(1):55–61, 1990

    Etienne Pardoux and Shige Peng. Adapted solution of a backward stochastic differential equation.Systems & Control Letters, 14(1):55–61, 1990

    1990

  22. [22]

    Stochastic hamilton–jacobi–bellman equations.SIAM Journal on Control and Optimization, 30(2):284–304, 1992

    Shige Peng. Stochastic hamilton–jacobi–bellman equations.SIAM Journal on Control and Optimization, 30(2):284–304, 1992

    1992

  23. [23]

    Backward stochastic differential equations and applications to optimal con- trol.Applied Mathematics and Optimization, 27(2):125–144, 1993

    Shige Peng. Backward stochastic differential equations and applications to optimal con- trol.Applied Mathematics and Optimization, 27(2):125–144, 1993

    1993

  24. [24]

    Solving high-dimensional partial differen- tial equations using deep learning.Proceedings of the National Academy of Sciences, 115(34):8505–8510, 2018

    Jiequn Han, Arnulf Jentzen, and Weinan E. Solving high-dimensional partial differen- tial equations using deep learning.Proceedings of the National Academy of Sciences, 115(34):8505–8510, 2018

    2018

  25. [25]

    Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    Brian DO Anderson. Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326, 1982

    1982

  26. [26]

    Backward stochastic differential equations and partial differen- tial equations with quadratic growth.The Annals of Probability, 28(2):558–602, 2000

    Magdalena Kobylanski. Backward stochastic differential equations and partial differen- tial equations with quadratic growth.The Annals of Probability, 28(2):558–602, 2000. 23

    2000