Stochastic Optimal Guaranteed Cost Control

• Stochastic optimal guaranteed cost control is a framework that provides explicit performance certificates under uncertainty, including worst-case, risk-based, and target constraints.
• It integrates diverse methodologies such as reflected SDEs with recursive BSDE cost functionals, stochastic target formulations, and convex relaxations to certify control bounds.
• Practical insights include applications in finance, quantum systems, and reinforcement learning, with computational strategies like SOS relaxations and Malliavin-based gradient methods.

Stochastic optimal guaranteed cost control is a family of stochastic control formulations in which a controller minimizes a performance criterion while certifying a bound, target, or risk level under uncertainty. In the literature, this certification takes several mathematically distinct forms: a worst-case bound under model uncertainty; a minimal achievable expected or average cost under stochastic dynamics; exact satisfaction of stochastic target or chance-type constraints; or certified upper and lower bounds on the optimal value function. Representative formulations include reflected stochastic differential systems with recursive backward-cost functionals, finite-horizon stochastic target problems, dynamic time-consistent risk-constrained Markov decision processes, robust control under $G$-Brownian motion, linearly solvable stochastic optimal control with semidefinite relaxations, stochastic PDE-constrained optimization with mean–variance objectives, and guaranteed-cost LQG control for uncertain linear quantum stochastic systems (Li et al., 2012, Chow et al., 2015, 0807.4619).

1. Conceptual scope and performance criteria

The core object is a stochastic control system, either continuous-time or discrete-time, together with a cost functional and a class of admissible controls. In continuous time, one standard form is

$$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$

with controlled SDE dynamics, while in discrete time one encounters both finite-horizon total-cost formulations and infinite-horizon average-cost criteria (Amidzadeh, 15 Mar 2026, Lai et al., 2020). In the average-cost setting for linear systems with multiplicative and additive Gaussian noise, the optimal long-run performance is

$$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$

and the optimal gain minimizes $\lambda$ over admissible linear policies (Lai et al., 2020).

What counts as “guaranteed cost” depends on the uncertainty model. In robust formulations under $G$-expectation, the cost already contains a worst-case operator,

$\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$

so the value function is the minimal worst-case cost over a family of volatility scenarios (Hu et al., 2016). In stochastic target problems, the guarantee is not a worst-case expectation but almost-sure satisfaction of a terminal target such as

$$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$

with the controller minimizing effort among all controls that satisfy this hard stochastic constraint (Dolinsky et al., 2019). In value-function approximation and global optimization, the guarantee is again different: one constructs certified lower and upper bounds on the optimal stochastic cost, either through SOS/SDP approximations of a desirability PDE or through convex/concave relaxations of the expected terminal cost (Horowitz et al., 2014, Shao et al., 2017).

This diversity suggests that stochastic optimal guaranteed cost control is not a single criterion but a collection of certification paradigms tied to how uncertainty, admissibility, and performance are modeled. A common thread is that the controller is not assessed only by nominal optimality; it is assessed by a mathematically explicit performance certificate.

2. Recursive costs, reflected dynamics, and nonlinear HJB characterizations

A particularly rich formulation arises for stochastic systems reflected in a domain, where the state is constrained to remain in a bounded convex set $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$ with reflection in the inward normal direction $\nabla\phi(x)$ on $\partial D$. The controlled reflected SDE is

$$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$0

with $$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$1 increasing only when the state lies on the boundary, thereby enforcing hard state constraints (Li et al., 2012).

The cost is recursive and is defined by a generalized BSDE,

$$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$2

so the cost process depends nonlinearly on its own continuation value. The control criterion is $$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$3, and the value function is the essential supremum over admissible controls. The paper shows that this quantity is deterministic, continuous in time, locally Lipschitz in space, and satisfies a generalized dynamic programming principle through a backward stochastic semigroup (Li et al., 2012).

The corresponding HJB equation is a fully nonlinear parabolic PDE with nonlinear Neumann boundary condition: $$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$4 with Hamiltonian

$$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$5

The reflection term in the state dynamics becomes the Neumann-type boundary condition, and the boundary generator $$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$6 appears directly in the boundary cost term. The value function is the unique viscosity solution of this HJB system (Li et al., 2012).

In guaranteed-cost language, this framework provides a precise mathematical object for the optimal performance level under state constraints and recursive, BSDE-based costs. The paper itself formulates the value as a maximization; under a minimizing sign convention, the same structure yields a minimal guaranteed cost. The recursive dependence on $$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$7 and $$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$8 also allows nonlinear valuation of future costs, including risk-sensitive or dynamically risk-adjusted structures.

3. Stochastic targets, dynamic risk constraints, and randomized policies

A different branch of the subject treats guarantees as explicit feasibility requirements. In the stochastic target formulation, the controlled state is

$$J(u,x_0)=\mathbb{E}\Bigg[\int_0^T L(x_t^u,u_t,t)\,dt+h(x_T^u)\Bigg],$$9

and admissible controls must satisfy

$$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$0

The main power-cost problem is

$$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$1

For this problem, the value function reduces, after scaling, to a one-dimensional function $$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$2 that is the minimal positive solution of the semi-linear ODE

$$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$3

with boundary conditions $$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$4, $$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$5. The associated BSDE has singular terminal condition $$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$6, and the optimal control is given explicitly in feedback form through the conditional probability process $$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$7 (Dolinsky et al., 2019). For exponential running costs, by contrast, the optimal controller becomes deterministic and constant: $$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$8 and the value is

$$\lambda=\lim_{N\to\infty}\frac{1}{N}\mathbb{E}\Big[\sum_{t=0}^{N-1}c(x_t,u_t)\Big],$$9

which the paper describes as a “trivial” optimal control (Dolinsky et al., 2019).

Time-consistent risk-constrained control introduces a different guarantee. In a finite-horizon finite-state MDP with stage-wise performance cost $\lambda$0 and risk cost $\lambda$1, the objective is to minimize

$\lambda$2

subject to a dynamic, time-consistent risk constraint

$\lambda$3

The dynamic risk measure is built recursively from coherent one-step conditional risk mappings, and the resulting dynamic program works on an augmented state $\lambda$4, where $\lambda$5 is a risk budget carried forward through the Bellman operator (Chow et al., 2015). This formulation replaces a single expected-cost constraint by a multistage guarantee on tail risk, with explicit time consistency.

Randomized control enters when the feasible performance–constraint set is nonconvex. For a finite-horizon stochastic optimal control problem with $\lambda$6 stochastic constraints, initial randomization among at most $\lambda$7 deterministic policies—“$\lambda$8-randomization”—is sufficient to attain the optimal mixed-strategy performance. The mixed strategy convexifies the achievable cost–constraint set, and the cost reduction relative to the optimal pure strategy is exactly the duality gap: $\lambda$9 The paper gives necessary and sufficient optimality conditions for randomized solutions and a dual-optimization-based construction; for $G$0, the dual problem can be solved by root finding (Ono et al., 2016). In this line of work, the guaranteed-cost interpretation is that the randomized controller achieves the minimal expected cost compatible with the stochastic constraints, whereas deterministic policies can be conservative when the underlying problem is nonconvex.

4. Robustness, worst-case formulations, and recurrence guarantees

Robust stochastic guaranteed cost control is explicit in the $G$1-Brownian framework. The controlled state satisfies a $G$2-SDE,

$G$3

and the cost is defined by an infinite-horizon $G$4-BSDE. Because

$G$5

the control problem is intrinsically an $G$6 problem and can be seen as a robust optimal control problem. The value function $G$7 is deterministic, continuous, and the unique viscosity solution of the elliptic HJBI equation

$G$8

where the operator $G$9 encodes the supremum over volatility matrices (Hu et al., 2016). In this setting, the guaranteed cost is the minimal worst-case infinite-horizon cost under volatility uncertainty.

The quantum case gives a more classical guaranteed-cost statement. For uncertain linear quantum stochastic systems, the performance index is

$\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$0

The uncertainty matrix $\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$1 satisfies $\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$2. By establishing an exact correspondence with an auxiliary classical uncertain system, the problem reduces to a classical minimax LQG design. Under the Riccati conditions in the paper, the resulting dynamic output-feedback controller guarantees

$\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$3

for all admissible uncertainties (0807.4619). This is a direct robust guaranteed-cost result, with an explicit worst-case upper bound on expected quadratic cost.

A further performance guarantee appears in the recent discrete-time infinite-horizon discounted literature. For nonlinear stochastic systems

$\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$4

with discounted value function

$\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$5

the paper introduces stochastic cost-controllability and detectability conditions involving a state measure $\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$6 and proves uniform semi-global practical recurrence for closed-loop systems controlled by optimal or near-optimal inputs. Under additional continuity assumptions, this property is robust with respect to small strictly causal perturbations (Moldenhauer et al., 29 Apr 2025). The guarantee here is not a min–max bound of the LQG type; it is a high-probability recurrence and boundedness certificate for closed-loop trajectories induced by discounted optimal control.

5. Certified computation, approximation, and learning

A substantial part of the literature studies how guaranteed-cost properties can be computed or certified numerically. One influential route begins from the linearly solvable stochastic optimal control transformation

$\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$7

under the matching condition

$\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$8

This converts the nonlinear HJB into a linear PDE for the desirability function $\mathbb{E}^G[X]=\sup_{P\in\mathcal P}E^P[X],$9. Sum-of-squares relaxations of the PDE yield semidefinite programs whose solutions are guaranteed lower and upper bounds on $$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$0, hence guaranteed upper and lower bounds on the value function $$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$1 through the logarithmic transform. The resulting hierarchy has monotonically decreasing residual bound $$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$2, and the approximate feedback law is obtained from

$$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$3

The main guarantee is a certified interval containing the optimal stochastic cost-to-go (Horowitz et al., 2014).

A complementary certification framework treats nonlinear stochastic optimal control with expected terminal cost

$$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$4

Using dynamic convex/concave relaxations of the state and cost integrand on $$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$5, combined with a partition $$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$6 of the uncertainty space and Jensen’s inequality, the paper constructs finitely computable convex and concave relaxations

$$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$7

with no sample-based approximation error. These give rigorous lower and upper bounds on the optimal objective value and are intended for use in spatial branch-and-bound algorithms (Shao et al., 2017).

For stochastic PDE-constrained control, Rosseel and Wells formulate optimization directly in stochastic function spaces and exploit the stochastic dimension to penalize statistical moments of the response. Two representative cost functionals are

$$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$8

and

$$X_T^{x,u}\ge \mathbb{I}_{\{W_T>c\}} \quad \text{a.s.},$$9

The optimal control can be decomposed as $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$0, where $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$1 is designed and $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$2 is a known zero-mean stochastic component. One-shot stochastic finite element methods then solve the coupled state–adjoint system. A key computational observation is that stochastic collocation loses its usual decoupling advantage when deterministic controls or moment terms are present, because the collocation points become coupled (Rosseel et al., 2011).

Learning-based computation appears in discrete-time average-cost control. For linear systems with multiplicative and additive Gaussian noise, the value function under a fixed linear gain $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$3 is quadratic,

$D=\{x\in\mathbb{R}^d:\phi(x)>0\}$4

and the optimal gain satisfies a stochastic algebraic Riccati equation. The paper defines a quadratic Q-function with kernel matrix $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$5, proposes a model-free policy-evaluation/policy-improvement scheme using recursive least squares, and proves convergence of the learned kernel matrix and control gain to the optimal ones (Lai et al., 2020). In that setting, the learned controller asymptotically achieves the minimal average cost among admissible linear policies.

More recently, Malliavin-calculus-based stochastic maximum principle methods have been used to avoid adjoint BSDE simulation altogether. For continuous-time SOCPs with cost

$D=\{x\in\mathbb{R}^d:\phi(x)>0\}$6

the first variation is expressed in terms of Malliavin derivatives $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$7 and the stochastic flow $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$8, rather than a backward adjoint process. The resulting iterative algorithm updates piecewise-constant controls by projected gradient steps and proves convergence under Lipschitz, monotonicity, and consistency conditions: $D=\{x\in\mathbb{R}^d:\phi(x)>0\}$9 The guarantee here concerns algorithmic convergence to the optimal control, rather than a robust worst-case performance bound (Amidzadeh, 15 Mar 2026).

6. Applications, interpretations, and limitations

The application range is broad. In finance, stochastic target control provides a model of super-hedging with transaction costs in the Bachelier model, where the terminal requirement is to hold at least one share when a call option is in the money (Dolinsky et al., 2019). In path planning, finite-state MDP control, and entry–descent–and–landing, randomized mixed strategies lower expected cost under stochastic constraints by exploiting nonconvexity in the feasible cost–constraint set (Ono et al., 2016). In nonlinear continuous-time control, SOS relaxations and convex/concave expected-cost relaxations aim at globally valid certificates for the value function or optimal objective (Horowitz et al., 2014, Shao et al., 2017). In distributed-parameter systems, stochastic PDE control directly shapes both mean response and variance (Rosseel et al., 2011). In quantum optics, guaranteed-cost LQG control is implemented through classical measurement and feedback for uncertain optical cavities (0807.4619).

The literature also imposes substantial structural assumptions. Reflected recursive control requires a $\nabla\phi(x)$0 defining function for the domain, Lipschitz coefficients, and linear growth conditions (Li et al., 2012). The stochastic target ODE characterization leaves uniqueness of the positive solution as an open problem because the coefficient $\nabla\phi(x)$1 vanishes at $\nabla\phi(x)$2 (Dolinsky et al., 2019). The linearly solvable SOS framework relies on the restrictive noise-control matching condition $\nabla\phi(x)$3 and polynomial data on compact semialgebraic domains (Horowitz et al., 2014). The $\nabla\phi(x)$4-Brownian infinite-horizon theory requires dissipativity and monotonicity assumptions strong enough to make the $\nabla\phi(x)$5-BSDE well posed (Hu et al., 2016). Dynamic programming with time-consistent risk constraints is developed for finite-state, finite-action MDPs and requires optimization over future risk-budget functions (Chow et al., 2015). Stochastic PDE formulations inherit the high dimensionality of polynomial chaos and lose stochastic collocation non-intrusivity in the presence of deterministic controls or moment penalties (Rosseel et al., 2011). Reinforcement-learning and Malliavin-gradient methods provide computational convergence results, but their guarantees remain tied to the assumed model class and regularity (Lai et al., 2020, Amidzadeh, 15 Mar 2026).

A persistent misconception is that guaranteed cost control in stochastic systems must always mean a worst-case $\nabla\phi(x)$6-type bound. The literature shows a wider landscape. In some works the guarantee is a worst-case expectation over a set of models; in others it is a certified upper bound on the optimal value function; in others it is exact satisfaction of a stochastic target, a time-consistent risk budget, or asymptotic convergence of a learned average cost (Hu et al., 2016, Shao et al., 2017, Chow et al., 2015, Lai et al., 2020). A plausible implication is that the field is best organized by the type of certificate being enforced—worst-case, recursive, target-based, risk-based, or approximation-based—rather than by a single canonical performance index.

Taken together, these works establish stochastic optimal guaranteed cost control as a technically heterogeneous but conceptually coherent area: stochastic control under uncertainty with an explicit performance certificate. The certificate may be a viscosity-characterized value function, a risk budget, a target constraint, a Riccati-based bound, a convex relaxation envelope, a recurrent set reached with high probability, or a convergent computational approximation. The unifying problem is not merely to optimize, but to optimize with proof.