Global Stochastic Maximum Principle

Updated 27 January 2026

Global Stochastic Maximum Principle is a framework providing necessary conditions for optimality in diverse stochastic control systems with nonconvex control domains and control-dependent diffusion.
It employs spike variation analysis and variational expansions alongside first and second-order adjoint equations to rigorously relate perturbations in control to changes in the Hamiltonian.
The principle extends to complex models—including mean-field, jump, delay, and infinite-dimensional systems—and offers robust insights for advanced stochastic control challenges.

The global stochastic maximum principle (SMP) provides necessary conditions for optimality in a wide range of stochastic control systems, including finite and infinite dimensional SDEs, FBSDEs, mean-field models, systems with jumps, delays, regime switching, partial information, and recursive utility frameworks. In contrast to local or "Pontryagin-type" principles, the global SMP accommodates nonconvex control domains and fully nonclassical stochasticity—where controls may enter both drift and diffusion, affect quadratic or mean-field couplings, and connect to backward (cost) equations. The principle is formulated via spike variation analysis, variational expansions, and adjoint BSDEs or SPDEs, resulting in a variational inequality involving the Hamiltonian and its second-order terms.

1. Formal Statement of the Principle

Given a stochastic control system—represented either by SDEs, FBSDEs, SPDEs, or their mean-field/jump/delay variants—the global SMP asserts that if $u^*(\cdot)$ is an optimal control, then for almost every $t$ and almost surely, the following condition must hold for all admissible alternatives $v \in U$ : $\mathcal{H}(t, X^*,u^*,p,q,P) \geq \mathcal{H}(t, X^*,v,p,q,P)$ Here, $\mathcal{H}$ is the (possibly augmented) Hamiltonian involving the state $X^*$ , adjoint processes $p$ , $q$ (first-order), and $P$ (second-order); the precise formulation depends on model specifics, including the involvement of jumps, mean-field dependencies, delay, partial information, etc. The second-order term is essential for nonconvex domains or when the control affects the diffusion.

For example, in the context of fully coupled FBSDEs,

$\mathcal{H}(t,x,y,z,u,p,q,P) = g(t,x,y,z,u) + p\,b(t,x,y,z,u) + q\,\sigma(t,x,y,z,u) + \frac{1}{2}\!\left[\sigma(t,x,y,z,u)-\sigma(t,x^*,y^*,z^*,u^*)\right]^{{\top}}P \left[\sigma(t,x,y,z,u)-\sigma(t,x^*,y^*,z^*,u^*)\right]$

and the SMP requires maximizing this function over $u$ pointwise in time (Hu et al., 2018, 1803.02109).

2. Adjoint Process Hierarchy

The necessary optimality conditions are formulated using adjoint equations, which encode sensitivity of the cost functional to perturbations of the control. Typically:

First-order adjoint BSDEs/SPDEs: These equations track linear sensitivity and are solved backward from the terminal condition involving the running cost or target payoff. In the FBSDE case or mean-field models with measure derivatives, these may be quadratic, matrix-valued, or involve "anticipated" terms in the presence of delay or regime switching (Buckdahn et al., 2024, Hao et al., 2022, Meng et al., 2019, Li, 20 Jan 2026).
Second-order adjoint equations: When control enters the diffusion or the domain is nonconvex, second-order sensitivity becomes essential. In infinite dimensions these can be operator-valued or conditionally expected backward stochastic integral equations (Liu et al., 2021, Du et al., 2012, Du et al., 2012). In finite dimensions, second-order BSDEs may be matrix-valued or possess jump terms and mean-field couplings (Hu et al., 2020, Hao et al., 2018).
Boundary conditions: The terminal/boundary data for adjoints (e.g., $P_T$ ) is generally tied to the second derivatives of the terminal cost or utility, and may combine standard and measure-derivative terms in mean-field models.

3. Spike Variation Technique and Taylor Expansion

The foundational proof mechanism is the spike/needle variation: perturb the optimal control on a small, measurable set $E_\varepsilon$ and expand both the state equation and the cost functional to first and second order in $\varepsilon$ . The procedure involves:

Linearizing the state equation to derive first-order terms (typically $O(\sqrt{\varepsilon})$ ),
Quadratic expansion for second-order corrections ( $O(\varepsilon)$ ),
Application of Itô's formula and adjoint duality to identify the correspondence between cost differences and variations in the Hamiltonian,
In models with delay, anticipated terms and cross-adjoint equations arise (Meng et al., 2019, Li, 20 Jan 2026).

This leads to a pointwise maximum inequality at almost every time, potentially involving indicator or shift terms if delay or partial information is present.

4. Hamiltonian Structure and Quadratic Correction

The Hamiltonian $\mathcal{H}$ encompasses the first-order contributions involving running cost, drift, and diffusion terms weighted by adjoint processes, and, crucially, second-order terms such as: $\frac{1}{2} \left(g_t(v)-g_t(u^*)\right)^{\top} P_t \left(g_t(v)-g_t(u^*)\right)$ for systems with nonconvex control domains and control-dependent diffusion (Du et al., 2012, Du et al., 2012). In mean-field, jump-diffusion, or FBSDE models, the Hamiltonian may also rely on measure derivatives, expectation terms, or conditional expectations to encode globality (Buckdahn et al., 2024, Hao et al., 2018, Hao et al., 2022). The quadratic term ensures the global nature of the principle and cannot be neglected except in specific convex setups.

5. Model Extensions: Mean-Field, Jumps, Delay, Infinite Dimensions, Partial Observation

The SMP has been established in broad settings:

Mean-field type: Coefficients depend on the law of the state (McKean–Vlasov), leading to measure-derivative adjoints, conditional expectation Hamiltonians, and law-dependent boundary conditions (Buckdahn et al., 2024, Hao et al., 2018).
Jumps/Teugel's Martingales: Lévy processes and multidimensional jumps require special stochastic integrals, jump-adjoint equations, and the inclusion of jump increments and cross-terms in the maximum condition (Lin, 2012, Hao et al., 2018, Zheng et al., 2021).
Delay systems: State or control delay induces anticipated backward equations and extra cross-term adjoints, as well as compensating shift terms in the maximum principle for proper accounting of future values (Meng et al., 2019, Li, 20 Jan 2026).
Infinite-dimensional systems: Control of SPDEs or abstract evolution equations leads to operator-valued adjoints and requires specialized analysis for existence in weak/generalized solution spaces (Du et al., 2012, Liu et al., 2021, Al-Hussein, 2012, Du et al., 2012).
Partial information/Observation: Observations affected by noise or jumps require filtering methods (Girsanov transform, separation principle) and the maximum principle applies to conditional expectations in the enlarged filtration (Zheng et al., 2022, Zheng et al., 2021).

6. Linear-Quadratic and Stackelberg Systems

In linear-quadratic (LQ) setups, adjoints reduce to matrix Riccati equations and SMP becomes a stationarity condition for the classical quadratic Hamiltonian. The explicit feedback control is given in terms of adjoint solutions (Hu et al., 2020, 1803.02109, Bensoussan et al., 2012). Stackelberg differential games require nested applications of the principle: the follower solves an SMP for their optimal response given the leader's control, and the leader then solves a FBSDE-driven SMP for their own optimal policy (Bensoussan et al., 2012).

7. Sufficiency, Verification, and Novel Features

The global SMP is a necessary condition for optimality; when the Hamiltonian is convex in state and control variables and the terminal cost is convex, it also becomes sufficient (Buckdahn et al., 2024, Hao et al., 2022, Hu et al., 2020). Recent advances include:

BMO-martingale estimates for multi-dimensional quadratic BSDEs allowing global results without convexity, time-horizon or norm-expansion restrictions (Hu et al., 2020).
Treatment of SMP in mean-field FBSDEs with jumps, including law-derivative estimates via Lions–Cardaliaguet theory (Hao et al., 2018).
Operators valued in infinite-dimensional spaces using the notion of generalized solution for the second-order adjoint in SPDEs (Du et al., 2012, Liu et al., 2021).
Stochastic maximum principles for partially observed systems with fractional Brownian motion—eliminating fractional noise via deterministic transformation and then applying classical SMP (Zheng et al., 2022).
Nonlinear filtering and feedback BAEs in partially observed progressive control problems with Poisson-jump correlations (Zheng et al., 2021).