A global stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.
Ada: We're looking at the paper: "A global stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition".
Emmy: Here's the gist of the paper. A global stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition is obtained by introducing a new probability measure to handle the lack of BMO martingale properties under unbounded terminal conditions Problem Formulation and Context The paper studies a stochastic optimal control problem for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition where the control domain is not necessarily convex The system involves an SDE for the state process X(t) and a BSDE for Y(t) The problem (O) seeks a control u∗(·) that minimizes the cost functional J(u(·)) = Y(0) Existing literature has established results for bounded terminal values, but this paper addresses the case with unbounded terminal conditions and a non-convex control domain Methodological Innovations The authors introduce a new probability measure Q∗ to analyze the well-posedness of linear BSDEs arising from the adjoint and variational equations This measure is defined by dQ∗dP = E
Z · 0 q∗(s)⊤dW(s): Under this measure, Girsanov’s transformation eliminates terms involving fz from the generators of the linear BSDEs, enabling well-posedness and estimates A new approach is provided for estimating (Yˆ ε,Zˆε) under this measure Key Analytical Results The paper establishes several key analytical results under the new measure Q∗ A stochastic maximum principle is obtained for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition.
Ada: The Hamiltonian function H is defined in terms of P1, Q1, and P2 The theorem states that ∆H(t, u) ⩾ 0 for all u ∈ U a.e., a.s.
Emmy: The estimate for (Yˆ ε,Zˆε) is shown to be O(ε2) This is achieved by combining estimates derived from different probability measures and Taylor expansions The estimate for (Yˆ ε,Zˆε) is shown to be O(ε2) This result is derived from combining the estimates for Yˆ ε, Xˆ ε, and Zˆε under the measure Q∗ Adjoint Equations and Variational Analysis The paper presents the first-order and second-order adjoint equations (5.1) and (5.2) derived using Itô's formula and Taylor expansions The well-posedness of these adjoint equations is established under the new measure Q∗ The variational equation (5.4) for Yˆ(t) is given in terms of the adjoint variables and other terms The estimate for (Yˆ ε,Zˆε) is derived by analyzing the terms in the variational equation Final Estimate and Maximum Principle The final result shows that Yˆε 2(t) = Hε 2(T) − Z T t n˜f2εy(s)Yˆ ε 2(s) + J1(s) o ds + Z T t Zˆε 1(s)dWq∗(s), where J1 is defined in terms of various components The estimate for Yˆε 2(t) is bounded by a term involving Kε2 This leads to the conclusion that E
sup t∈[0,T: Yˆ ε 2(t)2] = O(ε2) The general stochastic maximum principle is stated in Theorem 8.1 which asserts that ∆H(t, u) ⩾ 0 for all u ∈ U a.e., a.s.
Ada: References
1: P Barrieu and N El Karoui, Monotone stability of quadratic semimartingales with applications to unbounded general quadratic BSDEs, Ann Probab., 41 (2013), pp.
Paper summary: Ada: So we're looking at this paper by Hu and Li about a global stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition. What they're tackling is a problem where the control domain isn't necessarily convex, which is a big deal in control theory.
Emmy: Exactly, it’s about finding an optimal strategy when you can’t assume the setup is nicely shaped. The paper claims they get a global stochastic maximum principle under these conditions using a new probability measure.
Kolya: Before we get into the technical bits, I want to ask what this means for us practically. If we can find this principle, what does that actually help us figure out when designing these systems?
Ada: It means they've addressed a gap in the literature where previous methods didn't apply because of those unbounded terminal conditions. They introduce a new measure specifically to handle that lack of BMO martingale properties under these conditions.
Emmy: That new measure is defined by dQ star dP equals EZ dot zero q star s top dW s. It lets them analyze the well-posedness of the linear backward stochastic differential equations that pop up in both the adjoint and variational equations.
Kolya: Okay, so they're replacing a standard tool because it doesn't work when things get unbounded. What’s the main result they use this measure to establish?
Ada: The main thing is that under this new measure Q star, Girsanov’s transformation eliminates those pesky terms involving f z from the generators of the linear backward stochastic differential equations. That makes them well-posed and allows them to get estimates.
Emmy: And they use this to prove a new approach for estimating both Y hat epsilon and Z hat epsilon under this measure, which is crucial for their subsequent analysis of the adjoint equations.
Kolya: So we’re hearing that the core difficulty was proving that the BSDEs were well-behaved enough to start with. What are the actual estimates they manage to get once they have that new measure in place?
Ada: They establish a stochastic maximum principle where, for example, the Hamiltonian function H has a property called delta H t u greater than or equal to zero for all control u almost surely. That’s the theorem of Theorem eight point one <ref:2609.21451#pg2>.
Emmy: And they also get an estimate for Y hat epsilon and Z hat epsilon that is O epsilon squared by combining results from different probability measures and using Taylor expansions.
Kolya: O epsilon squared sounds good, but what does that actually tell us about the stability or error of the solution in the control problem? Is it a tight bound?
Paper summary: Ada: It implies that once you get those estimates for Y hat epsilon and Z hat epsilon under this new measure, you can see that they are bounded by a term involving K epsilon squared. This ultimately leads to the conclusion that the expected supremum of Y hat epsilon squared over time is O epsilon squared.
Emmy: That means the error in the solution scales quadratically with some small parameter, which is a pretty strong statement about convergence. It shows that their approach works when things are unbounded, provided you use this new measure Q star.
Kolya: So if we translate that for someone who isn't deep in control theory, what’s the big picture takeaway here? What does this paper actually change for the field of stochastic control?
Ada: It fills a significant gap because previous results relied on the BMO martingale property, which just doesn't hold when you have unbounded terminal conditions. This paper provides a new way to proceed without that assumption.
Emmy: The implication is that we can now apply maximum principle techniques to forward-backward systems even when the control domain isn't smooth or convex, as long as we introduce this specialized probability measure.
Kolya: So, what are the limitations they admit? What’s the caveat on this new setup?
Ada: They note that they have to rely on certain assumptions for their work to hold under their current framework. For instance, in example one four of their analysis, if the stochastic process u is unbounded and defined by something like W t, then they cannot guarantee that a specific term involving e epsilon X hat T squared remains integrable.
Emmy: So even with this new measure and the O epsilon squared estimate for Y hat epsilon squared, they still hit a wall when the control itself becomes unbounded in a way that violates integrability requirements.
Kolya: That's fair. So, if we look at this paper by Hu and Li about forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition, what does it really add to the literature on optimal control?
Ada: It adds a global stochastic maximum principle for this specific class of systems where the control domain isn't necessarily convex. That’s a major addition because we’ve seen work on related problems, like those with bounded terminal values, but not this exact setup.
Emmy: The paper shows that by introducing the right mathematical tools—specifically that new probability measure Q star—we can overcome the hurdle posed by unbounded terminal conditions and non-convex constraints in these forward-backward systems.
Kolya: It’s about showing how to find a global optimum even when the standard machinery breaks down due to those boundary conditions. The main thing is that they've provided a path forward for solving this class of control problems.
Conclusion: Ada: So, we’re wrapping up our look at this paper about forward-backward stochastic control systems. The authors are tackling those tricky unbounded terminal conditions and non-convex control domains with a global maximum principle.
Emmy: It’s a dense piece of math, but the core idea is that they found a way to get an optimal strategy even when the setup isn't nice or bounded. They introduced this new probability measure Q star to handle the technical problems that pop up with those unbounded conditions.
Kolya: What does this really mean for us? It means we can finally apply these powerful maximum principle tools to control problems that used to be off-limits because of the growth of the terminal condition.
Ada: Exactly, it fills a real gap in the literature where previous results just didn't apply anymore because of those unbounded conditions. The authors show that with this new approach, we can establish a global maximum principle for these complex systems.
Emmy: From an algebraic standpoint, they managed to build the whole argument around this new measure Q star so that the linear backward stochastic differential equations become well-posed and solvable in a controlled way.
Kolya: And the results they get, like that big O of n squared estimate for Y hat epsilon squared, shows how stable these solutions are when you get those estimates under this measure. It gives us a concrete idea of the error scaling.
Ada: So, to sum it up, this paper provides a robust method for finding an optimal control in forward-backward systems with quadratic generators and unbounded terminal conditions, even when the control domain isn't convex.
Emmy: It’s about proving that structure—the global maximum principle—still holds true under these challenging mathematical circumstances.
Kolya: This opens the door to solving a whole class of problems we previously couldn't touch with existing literature. What we look at next is exactly how they prove those key estimates, which is where the real heavy lifting happens.
Ying Hu, Feng Li, Jiaqiang Wen
University of Rennes · Southern University of Science and Technology
math.OC
Submitted: 2026-09-18
Updated: 2026-11-08
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 80/100
The gist: A global stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition is obtained by introducing a new probability…
Key concepts
- Forward-Backward Stochastic Control Systems
- These are mathematical models involving two coupled processes: an SDE (Stochastic Differential Equation) for the state process X(t) and a BSDE (Backward Stochastic Differential Equation) for the cost functional Y(t). The goal is to find an optimal control strategy u* that minimizes the total cost J(u).
- Quadratic Convex Generator
- This describes the structure of the system's dynamics, where the generator related to Y(t) has a quadratic term. 'Convex' implies certain desirable mathematical properties for analysis, which is crucial for establishing the existence and uniqueness of solutions in this control problem.
- New Probability Measure (Q*)
- This is a specially constructed measure used to analyze the well-posedness of linear BSDEs that arise from the adjoint and variational equations. It is introduced because standard methods fail when terminal conditions are unbounded, allowing the authors to simplify the analysis by eliminating problematic terms.
- Stochastic Maximum Principle
- This is a fundamental result in optimal control theory. It provides necessary conditions for finding an optimal control u*. The paper establishes a global version of this principle, meaning it holds for all possible controls within the allowed set, under the specific system conditions studied.
Terminology
Summary
A global stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition is obtained by introducing a new probability measure to handle the lack of BMO martingale properties under unbounded terminal conditions
Problem Formulation and Context
The paper studies a stochastic optimal control problem for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition where the control domain is not necessarily convex The system involves an SDE for the state process X(t) and a BSDE for Y(t) The problem (O) seeks a control u∗(·) that minimizes the cost functional J(u(·)) = Y(0) Existing literature has established results for bounded terminal values, but this paper addresses the case with unbounded terminal conditions and a non-convex control domain
Methodological Innovations
The authors introduce a new probability measure Q∗ to analyze the well-posedness of linear BSDEs arising from the adjoint and variational equations This measure is defined by dQ∗dP = E[Z · 0 q∗(s)⊤dW(s)] Under this measure, Girsanov’s transformation eliminates terms involving fz from the generators of the linear BSDEs, enabling well-posedness and estimates A new approach is provided for estimating (Yˆ ε,Zˆε) under this measure
Key Analytical Results
The paper establishes several key analytical results under the new measure Q∗
** A stochastic maximum principle is obtained for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition. The Hamiltonian function H is defined in terms of P1, Q1, and P2 The theorem states that ∆H(t, u) ⩾ 0 for all u ∈ U a. p. 2e., a.s. **
** The estimate for (Yˆ ε,Zˆε) is shown to be O(ε2) This is achieved by combining estimates derived from different probability measures and Taylor expansions **
** The estimate for (Yˆ ε,Zˆε) is shown to be O(ε2) This result is derived from combining the estimates for Yˆ ε, Xˆ ε, and Zˆε under the measure Q∗ **
Adjoint Equations and Variational Analysis
The paper presents the first-order and second-order adjoint equations (5.1) and (5.2) derived using Itô's formula and Taylor expansions The well-posedness of these adjoint equations is established under the new measure Q∗ The variational equation (5.4) for Yˆ(t) is given in terms of the adjoint variables and other terms The estimate for (Yˆ ε,Zˆε) is derived by analyzing the terms in the variational equation
Final Estimate and Maximum Principle
The final result shows that Yˆε 2(t) = Hε 2(T) − Z T t n˜f2εy(s)Yˆ ε 2(s) + J1(s) o ds + Z T t Zˆε 1(s)dWq∗(s), where J1 is defined in terms of various components The estimate for Yˆε 2(t) is bounded by a term involving Kε2 This leads to the conclusion that E[sup t∈[0,T] Yˆ ε 2(t)2] = O(ε2) The general stochastic maximum principle is stated in Theorem 8.1 which asserts that ∆H(t, u) ⩾ 0 for all u ∈ U a.e., a.s.
References
[1] P Barrieu and N El Karoui, Monotone stability of quadratic semimartingales with applications to unbounded general quadratic BSDEs, Ann Probab., 41 (2013), pp. 1831–1863
[2] P Briand, B Delyon, Y Hu, E Pardoux, and L Stoica, L p solutions of backward stochastic differential equations, Stochastic Process. Appl., 108 (1) (2003), pp. 109–129
[3] P Briand and Y Hu, BSDE with quadratic growth and unbounded terminal value, Probab. Theory Relat. p. 1 p. 1 Fields, 136 (4) (2006), pp. 604–618
[4] P Briand and Y Hu, Quadratic BSDEs with convex generators and unbounded terminal conditions, Probab. Theory Relat. Fields, 141 (3) (2008), pp. 543–567
[5] R Buckdahn, J Li, Y Li, and Y Wang, A Global Stochastic Maximum Principle for MeanField Forward-Backward Stochastic Control Systems with Quadratic Generators, Ann Appl. Probab., 36 (1) (2026), pp. 275–318
[6] F Delarue and P Lavigne, Robust mean field control: stochastic maximum principle and variational mean field games, arXiv preprint arXiv: 2604.21641
[7] F Delbaen, Y Hu, and A Richou, On the uniqueness of solutions to quadratic BSDEs with convex generators and unbounded terminal conditions, Ann Inst Henri Poincar´e Probab Stat., 47 (2) (2011), pp. 559–574
[8] F Delbaen, Y Hu, and A Richou, On the uniqueness of solutions to quadratic BSDEs with convex generators and unbounded terminal conditions: the critical case, Discrete Contin. Dyn. Syst., 35 (11) (2015), pp. 5273–5283
[9] S Fan, Y Hu, and S Tang, On the uniqueness of solutions to quadratic BSDEs with non-convex generators and unbounded terminal conditions, C.R. Math Acad Sci Paris, 358 (2) (2020), pp. p. 0 227–235
[10] S Fan, Y Hu, and S Tang, Multi-dimensional backward stochastic differential equations of diagonally quadratic generators: the general result, J Differential Equations, 368 (2023), pp. 105– 140
[11] T Hao and Q Meng, A global maximum principle for optimal control of general mean-field forward-backward stochastic systems with jumps, ESAIM Contr Optim Calc Var., 26 (2020), 87
[12] M Hu, Stochastic global maximum principle for optimization with recursive utilities, Probab. Uncertain Quantit Risk, 2 (2017), pp. p. 1 1–20
[13] M Hu, S Ji, and R Xu, A global stochastic maximum principle for forward-backward stochastic control systems with quadratic generators, SIAM J Control Optim., 60 (3) (2022), 1791–1818
[14] M Hu, S Ji, R Xu, and X Xue, A BSDE approach to the asymmetric risk-sensitive optimization and its applications, ESAIM Control Optim. Calc Var., 32 (2026), 72
[15] M Hu, S Ji, and X Xue, A global stochastic maximum principle for fully coupled forwardbackward stochastic systems, SIAM J Control Optim., 56 (6) (2018), pp. 4309–4335
[16] M Hu and F Wang, Maximum principle for stochastic recursive optimal control problem under model uncertainty, SIAM J Control Optim., 58 (3) (2020), pp. 1341–1370
[17] Y Hu, P Imkeller, and M Muller, ¨ Utility maximization in incomplete markets, Ann Appl Probab., 15 (3) (2005), pp. 1691–1712
[18] Y Hu and S Tang, Multi-dimensional backward stochastic differential equations of diagonally quadratic generators, Stoch Process Appl., 126 (4) (2016), pp. 1066–1086
[19] Y Hu, S Tang, and F Wang, Quadratic G-BSDEs with convex generators and unbounded terminal conditions, Stoch Process Appl., 153, (2022), pp. 363–390
[20] J Huang and J Shi, Maximum principle for optimal control of fully coupled forward-backward stochastic differential delayed equations, ESAIM Contr Optim Calc Var., 18 (4) (2012), pp. 1073– 1096
[24] E Pardoux and S Peng, Adapted solution of a backward stochastic differential equation, Systems Control Lett., 14 (1) (1990), pp. 55–61
[25] S Peng, Open problems of backward stochastic differential equations, in: S Chen, et al. (Eds.), Control of Distributed Parameter and Stochastic Systems, Hangzhou, 1998, Kluwer Academic Publishers, Boston, 1999, pp. p. 0 265–273
[26] A Richou, Markovian quadratic and superquadratic BSDEs with an unbounded terminal condition, Stoch Process Appl., 122 (9) (2012), pp. 3173–3208
[27] J Shi and Z Wu, The maximum principle for fully coupled forward-backward stochastic control system, Acta Automat Sinica, 32 (2006), pp. 161–169
[28] E M Stein and R Shakarchi, Real analysis: measure theory, integration, and Hilbert spaces, Princeton University Press, 2009
[30] P Whittle, A risk-sensitive maximum principle, Systems Control Lett., 15 (3) (1990), pp. 183–192
[32] J Yong, Optimality variational principle for controlled forward-backward stochastic differential equations with mixed initial-terminal conditions, SIAM J Control Optim., 48 (2010), pp. 4119–4156
Page 3
The solvability of BSDEs with a Lipschitz continuous generator was first established by Pardoux and Peng [24] After that, the class of BSDEs, with generators having quadratic growth in the state variable Z, has attracted a lot of attention in recent years For the BSDE in (1.1) with a generator of linear growth, Yong [32] obtained a stochastic maximum principle for forward-backward stochastic control problems by treating Z as a control process, regarding the terminal condition as a constraint, and applying the Ekeland variational principle For the BSDE in (1. p. 11) with a generator of quadratic growth and bounded terminal condition, Hu, Ji, and Xu [13] obtained the stochastic maximum principle by using the theory of BMO martingales
Page 4
The forward-backward stochastic control problem was proposed by Peng [25] as an open problem For the BSDE in (1.1) with a generator of quadratic growth and unbounded terminal condition, to the best of our knowledge, only Hu, Ji, Xu, and Xue [14] studied the specific case and Delarue and Lavigne [6] studied the case of convex control domain This paper aims to obtain a global stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition, where the control domain is not necessarily convex
Page 5
Since h is unbounded, the existing literature does not guarantee that R · 0 Z(s)dW(s) has the BMO property, from which it follows that R · 0 fz(s)dW(s) is, in general, not a BMO martingale Consequently, in the case of unbounded terminal condition, the proofs of results analogous to Hu, Ji, and Xu [13] are no longer applicable Fortunately, from Delbaen, Hu, and Richou [7], we know that E R · 0 fz(s)dW(s) is in Llog L space Inspired by this, we introduce a new probability measure (see (3.5))
Page 6
The main contributions in this paper are as follows
** We obtain a stochastic maximum principle for forward-backward stochastic control systems with quadratic convex generator and unbounded terminal condition, which fills an important gap in the existing literature **
** We introduce a new measure to study the well-posedness of the linear BSDEs with random coefficients arising from the adjoint and variational equations **
** We provide a new approach for proving the estimate for (Yˆ ε,Zˆε) which plays a crucial role in the proofs of Proposition 6.1 and Lemma 7. p. 1 p. 11 **
Page 7
The main difficulty is to estimate (Yˆ ε,Zˆε) under the new probability measure (see Proposition 4.2) Once the estimate for (Yˆ ε,Zˆε) under the new probability measure has been established, the BSDEs arising in the subsequent analysis are, in fact, linear with respect to the new Brownian motion under the new measure
Page 8
When the generator in (1.1) takes the following special form: f(t, X(t), Y (t), Z(t), u(t) = γ2 Z(t)2 + g(t, X(t), u(t)), after applying a suitable transformation, one can observe that the problem (O) becomes risk-sensitive control problem In this setting, only the first-order variational equation X1 needs to be considered
Page 9
In contrast to Delarue and Lavigne [6], the drift term b and diffusion term σ in this paper are not necessarily linear, and the control domain is not necessarily convex The uniform boundedness of σ(t, u) ensures the exponential integrability needed to establish well-posedness and carry out subsequent second-order variational analysis
Page 10
The main contributions in this paper are as follows
** We provide a new approach for proving the estimate for (Yˆ ε,Zˆε) which plays a crucial role in the proofs of Proposition 6.1 and Lemma 7.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Risk-sensitive control problem analysis: The system can now analyze risk-sensitive control problems by treating them as special cases of forward-backward stochastic systems with quadratic generators, leveraging transformations to simplify the generator form to a risk-sensitive structure.
-
Global maximum principle for nonconvex domains: The AI system can obtain a global stochastic maximum principle for forward-backward stochastic control systems even when the control domain is not necessarily convex, which addresses a significant gap in existing literature.
-
New probability measure application: The system can employ a new probability measure defined by q star to study the well-posedness of linear BSDEs with random coefficients arising from adjoint and variational equations, allowing for the establishment of necessary estimates.
-
Variational equation estimation under Q star: The AI can perform rigorous estimations for the first and second-order variational equations X1, X2, X1 hat 1, X2 hat 2 under the new probability measure Q star to bound their norms in terms of epsilon.
-
Estimate for Y epsilon (0) minus Y star (0) minus Y hat (0): The system can provide an estimate for the difference between the optimal value function and its approximation, showing it is o(epsilon), which is crucial for proving convergence properties.
-
Stochastic maximum principle derivation: The AI can derive a general stochastic maximum principle by defining a Hamiltonian function H and showing that the difference in cost functional J(u epsilon) minus J(u star) is positive under the given assumptions.
Sources
Related papers
- An Accelerated Proximal Gradient Method for Bures-Wasserstein Space
- FOMTools: Development and AI-Assisted Code Generation for First-Order Methods
- Lions and Muons: Optimization via Stochastic Frank-Wolfe under Heavy-Tailed Noise
- Dynamic Service Recommendation with Congestion-Dependent Joining: Near-Optimal and Constant-Factor Approximation Algorithms
- Integrating Learning and Optimization for Open Transport Capacity Pooling under Disruption
- Zeroth-Order Gradient Estimators under Decision-Dependent Noise