HJBI Framework: Stochastic Differential Games

• HJBI framework is a comprehensive mathematical structure for modeling zero-sum stochastic differential games using nonlinear PDEs and BSPDEs.
• It integrates dynamic programming principles and viscosity solutions to bridge probabilistic representations with analytic formulations.
• Advanced numerical methods like discontinuous Galerkin and C⁰-interior penalty FEM ensure stability, convergence, and practical applicability.

The Hamilton-Jacobi-Bellman-Isaacs (HJBI) framework provides the analytic and probabilistic foundation for modeling zero-sum stochastic differential games involving two players, typically framed as a minimax optimization over a value function governed by a fully nonlinear (often second-order) PDE or, in stochastic settings, a backward stochastic partial differential equation (BSPDE). The HJBI equation generalizes the Hamilton-Jacobi-Bellman equation of stochastic control (single-player) and the deterministic Hamilton-Jacobi-Isaacs equation of pursuit-evasion or robust control, encoding the dynamics, cost structure, and strategic interaction through min-max operators and, in general, nonlocal and non-smooth elements.

1. HJBI Equation: Analytic and Probabilistic Form

The classical analytic form on a domain $\Omega\subset\mathbb{R}^d$ specifies the Isaacs/HJBI PDE as

$$F(D^2u(x), u(x), x) := \inf_{\alpha \in A} \sup_{\beta \in B} \Big\{-\operatorname{Tr}[a^{\alpha\beta}(x) D^2 u(x)] - b^{\alpha\beta}(x)\cdot Du(x) - c^{\alpha\beta}(x) u(x) - f^{\alpha\beta}(x)\Big\} = 0,$$

where $A, B$ are compact metric control (or game) parameter sets and $$(a^{\alpha\beta}, b^{\alpha\beta}, c^{\alpha\beta}, f^{\alpha\beta})$$ are the data of the game (Kawecki et al., 2020). This form encompasses both fully nonlinear elliptic and parabolic regimes and generalizes to quasi-variational inequalities (QVIs) to accommodate impulse or state constraints (Zhang, 2019, Lee et al., 2021).

In the stochastic setting, denoting $(X_t, Y_t, Z_t)$ as state, cost, and control processes,

• State dynamics: $$dX_t = b(t, X_t, u_t, v_t)dt + \sigma(t, X_t, u_t, v_t)dW_t$$,
• Recursive cost (BSDE): $dY_t = -f(t, X_t, Y_t, Z_t, u_t, v_t)dt + Z_t dW_t$, $Y_T = g(X_T)$.

Value functions are defined either as lower/upper values via Elliott–Kalton non-anticipative strategies or as viscosity solutions to the associated PDE/BSPDE (Wang et al., 2024, Qiu et al., 2020).

In the presence of jumps or nonlocal terms, the HJBI equation incorporates additional integral or impulse operators, as in (Luo et al., 2023, Yoshioka et al., 2021).

2. Dynamic Programming and Strategic Constructions

The HJBI framework rests on a dynamic programming principle (DPP), linking probabilistic and analytic representations. Informally, for value function $V$ and admissible strategies: $V(t,x) = \essinf_{\beta} \esssup_u \mathbb{E}\left[ V(t+\delta, X_{t+\delta}^{t,x;u,\beta(u)})\right],$ which, upon localization and small time increments, yields the infinitesimal generator appearing in the PDE form (Zhang, 2019, Wang et al., 2024, Qiu et al., 2020).

For problems with impulse or state-constraint mechanisms, the DPP integrates both the evolution under continuous controls and discrete jumps, producing quasi-variational inequalities with one or more obstacles (Zhang, 2019, Lee et al., 2021).

The Isaacs condition (i.e., $$F(D^2u(x), u(x), x) := \inf_{\alpha \in A} \sup_{\beta \in B} \Big\{-\operatorname{Tr}[a^{\alpha\beta}(x) D^2 u(x)] - b^{\alpha\beta}(x)\cdot Du(x) - c^{\alpha\beta}(x) u(x) - f^{\alpha\beta}(x)\Big\} = 0,$$0) is crucial for coincident upper/lower value functions and the existence of a single value/PDE (Qiu et al., 2020, Zhang, 2019).

3. Viscosity Solutions, Comparison, and Regularity

HJBI equations—especially in non-smooth or stochastic settings—are typically analyzed in the viscosity solution framework. Sub- and supersolution notions are localized via test functions/semimartingale expansions (in the stochastic setting), with the min-max structure of HJBI encoded at the level of the viscosity criterion (Qiu et al., 2020, Wang et al., 2024, Luo et al., 2023).

Comparison principles—often leveraging monotonicity or special structure in the data—yield uniqueness. Under monotonic BSDE generators, strong regularity (e.g., Lipschitz continuity in $$F(D^2u(x), u(x), x) := \inf_{\alpha \in A} \sup_{\beta \in B} \Big\{-\operatorname{Tr}[a^{\alpha\beta}(x) D^2 u(x)] - b^{\alpha\beta}(x)\cdot Du(x) - c^{\alpha\beta}(x) u(x) - f^{\alpha\beta}(x)\Big\} = 0,$$1, improved time regularity) can be obtained, even when nonlinearities are non-Lipschitz in the primary cost variable (Wang et al., 2024). For systems with jumps or coupled equations (as in multi-mode or robust control), viscosity theory is developed for systems of integral–PDEs (Luo et al., 2023, Yoshioka et al., 2021).

4. Numerical Approximation: FEM, DG, and Adaptivity

For elliptic and parabolic HJBI equations with sufficiently regular (Cordes) coefficients, discontinuous Galerkin (DG) and $$F(D^2u(x), u(x), x) := \inf_{\alpha \in A} \sup_{\beta \in B} \Big\{-\operatorname{Tr}[a^{\alpha\beta}(x) D^2 u(x)] - b^{\alpha\beta}(x)\cdot Du(x) - c^{\alpha\beta}(x) u(x) - f^{\alpha\beta}(x)\Big\} = 0,$$2-interior penalty (C$$F(D^2u(x), u(x), x) := \inf_{\alpha \in A} \sup_{\beta \in B} \Big\{-\operatorname{Tr}[a^{\alpha\beta}(x) D^2 u(x)] - b^{\alpha\beta}(x)\cdot Du(x) - c^{\alpha\beta}(x) u(x) - f^{\alpha\beta}(x)\Big\} = 0,$$3-IP) finite element methods are developed, with strong monotonicity and Lipschitz properties in mesh-dependent norms underpinning well-posedness and convergence (Kawecki et al., 2020, Kawecki et al., 2021, Kawecki et al., 2020). Error analysis relies on a posteriori estimators built from local residuals and jump terms, and a priori quasi-optimality follows from abstract monotonicity/consistency conditions (see Table).

Method	Key property	Applicability
DG/C$$F(D^2u(x), u(x), x) := \inf_{\alpha \in A} \sup_{\beta \in B} \Big\{-\operatorname{Tr}[a^{\alpha\beta}(x) D^2 u(x)] - b^{\alpha\beta}(x)\cdot Du(x) - c^{\alpha\beta}(x) u(x) - f^{\alpha\beta}(x)\Big\} = 0,$$4-IP FEM	Strong monotonicity, a posteriori estimation	HJBI with Cordes coefficients
Monotone FD	Discrete comparison, Barles–Souganidis convergence	Nonlocal or PIDE HJBI (Yoshioka et al., 2021)

For periodic and homogenized HJBI settings, DG and C$$F(D^2u(x), u(x), x) := \inf_{\alpha \in A} \sup_{\beta \in B} \Big\{-\operatorname{Tr}[a^{\alpha\beta}(x) D^2 u(x)] - b^{\alpha\beta}(x)\cdot Du(x) - c^{\alpha\beta}(x) u(x) - f^{\alpha\beta}(x)\Big\} = 0,$$5-IP methods enable effective computation of cell problems and ergodic effective Hamiltonians (Kawecki et al., 2021). Limit-space arguments allow convergence analysis for adaptively refined, nonconforming meshes (Kawecki et al., 2020).

5. Stochastic Games, Coupled Systems & Nonlocal HJBI

Recent work extends the HJBI formalism to:

• Systems of coupled HJBI equations with integral–differential operators (to model switching, robust, or multi-mode games) (Luo et al., 2023),
• HJBI equations on discrete structures such as finite graphs, where comparison principles, monotonicity, and min–max Laplacian representations parallel the PDE setting (Forcillo et al., 10 Nov 2025),
• Stochastic control with impulse, rational inattention, or robustification via relative entropy, leading to nonlocal PIDE-HJBI equations with viscosity solutions and monotone finite-difference discretizations (Yoshioka et al., 2021).

Such generalizations address practical control and game problems under nonstandard uncertainty, high-frequency interventions, or information constraints.

6. Representative Applications and Examples

Applications of the HJBI framework span domains:

• Financial engineering, e.g., in recursive utility maximization or portfolio optimization with Epstein–Zin (non-Lipschitz) utility (Wang et al., 2024),
• State-constrained safety/reachability in engineering systems using QVIs to encode barrier objectives (Lee et al., 2021),
• Multi-mode robust control, as in coupled HJBI systems (Luo et al., 2023),
• Environmental management under stochastic and information-limited conditions (Yoshioka et al., 2021),
• Bellman–Isaacs representation on networks and Markov chains (Forcillo et al., 10 Nov 2025).

Model problem structures (see Table) demonstrate the diverse PDE and game-theoretic forms encompassed.

Context	PDE type	Special features
Impulse/Obstacles	QVI with double obstacles (Zhang, 2019)	Nonlocal obstacle operators
Stochastic zero-sum	Fully nonlinear SPDE (Qiu et al., 2020)	BSPDE, random Hamiltonians
State-constraints	HJBI with obstacles (Lee et al., 2021)	Free-boundary, multi-regime
Coupled systems	Integral–PDE systems (Luo et al., 2023)	BSDE with multiple Poisson jumps
Nonlocal/fractional	PIDE HJBI (Yoshioka et al., 2021)	Robust/rational inattention

7. Key Theoretical Insights and Open Directions

• The Cordes condition is fundamental for well-posedness of strong solutions and the development of stable numerical methods in the fully nonlinear regime (Kawecki et al., 2020).
• The viscosity solution framework ensures existence/uniqueness in the presence of non-smooth, nonlocal, or even discontinuous data (Wang et al., 2024, Luo et al., 2023).
• Stochastic HJBI equations induce random-field value functions, requiring stochastic analogues of DPP and viscosity theory (Qiu et al., 2020).
• The Isaacs condition remains pivotal throughout: when it fails, upper and lower value functions may diverge, and a value for the game may not exist (Zhang, 2019, Qiu et al., 2020).

Extensions toward degenerate diffusions, rough coefficients, or path-dependent/HJB–PPDE settings remain active areas, as do robust and learning-based formulations applying HJBI in reinforcement learning and PDE-constrained optimization contexts.