Hamilton-Jacobi-Isaacs Equation

Updated 7 January 2026

Hamilton-Jacobi-Isaacs equation is a central PDE in zero-sum differential games characterized by a minimax structure and nonanticipative control strategies.
It generalizes the Hamilton-Jacobi-Bellman equation by incorporating dual objectives, making it key for robust control and stochastic as well as path-dependent systems.
Numerical and analytical methods, including viscosity solution theory and policy iteration, are used to ensure well-posedness and convergence under the Isaacs condition.

The Hamilton-Jacobi-Isaacs (HJI) equation is the central partial differential equation (PDE) of zero-sum differential games, with deep connections to stochastic control, robust control, and path-dependent systems. Formally, the HJI equation generalizes the Hamilton-Jacobi-Bellman (HJB) equation by encoding a minimax (or sup-inf) structure in the Hamiltonian, reflecting the antagonistic objectives of two players optimizing over nonanticipative control strategies. The HJI equation arises in a broad array of applications, from the robust stabilization of nonlinear PDEs, to singularly perturbed slow-fast systems, stochastic games with delays or path/fractional-order dependencies, and graph-based or nonlocal control models.

1. Mathematical Formulation and General Structure

Consider a two-player zero-sum stochastic differential game on a filtered probability space, with controlled state dynamics

$dX_s = f(X_s, u_s, v_s) ds + \sigma(X_s, u_s, v_s) dB_s, \quad X_t = x$

where $u_s \in U$ and $v_s \in V$ are admissible controls for Players I and II respectively, taking values in compact sets. Associated with nonanticipative control strategies, the value function for the game is classically defined by

$V(t, x) = \sup_{\beta \in \mathcal{B}} \inf_{u \in \mathcal{U}} J(t, x; u, \beta[u]),$

with payoff functional $J$ defined via terminal and running costs.

The corresponding (second-order) Hamilton-Jacobi-Isaacs equation is

$\begin{aligned} - \partial_t V(t,x) + H(t,x, \nabla_x V, D_{xx}V, V, \dots) &= 0, \quad t < T\ V(T, x) &= h(x). \end{aligned}$

The HJI Hamiltonian $H$ encodes the min-max (or max-min) over the control sets: $H(t,x,p,A,y,z) = \sup_{u \in U} \inf_{v \in V} \Big\{ \frac{1}{2} \mathrm{Tr}[\sigma \sigma^\top(t,x,u,v)A] + b(t,x,u,v)^\top p + f(t,x,y,z+\sigma^\top p, u,v) \Big\}.$ Under the Isaacs condition, the two orders (inf-sup vs sup-inf) coincide (Qiu et al., 2020), ensuring well-posedness of the game.

Path-dependent variants allow the data and controls to depend on entire histories; the associated PHJI are formulated using the functional Itô calculus and accommodate infinite-dimensional state spaces and generalized derivatives (Moon, 2019, Gomoyunov, 2021).

2. Viscosity Solution Theory and Well-Posedness

In general, HJI equations are fully nonlinear, often degenerate-elliptic PIDEs or SPDEs. Classical solutions rarely exist; the correct solution framework is the viscosity solution, which adapts to cases of non-smooth data and value functions.

Crandall–Lions viscosity theory extends to state and path-dependent HJI equations via appropriate test functionals, including on Hölder-compact path spaces and with functional derivatives (Dupire/vertical/horizontal) (Moon, 2019).
Stochastic HJI and HJBI equations involve random coefficients and require a stochastic viscosity-solution framework, with sublinear expectations tied to the associated BSDEs (Qiu et al., 2020).
For fractional-order and delay systems, coinvariant and Caputo-type derivatives are used. Here, viscosity solutions are defined using test functionals and sub- and superdifferentials adapted to the memory structure of the system (Gomoyunov, 2021, Plaksin, 2020).
Existence follows from the dynamic programming principle (DPP) and regularity of coefficients; uniqueness typically holds under comparison principles and the Isaacs condition; for certain nonlinear and non-smooth cases, uniqueness may be open or require additional monotonicity (Moon, 2019, Qiu et al., 2020, Wang et al., 2024).

3. Methods of Solution: Numerical and Analytical Approaches

A variety of computational techniques have been developed for high- and infinite-dimensional HJI equations:

Policy iteration and Galerkin methods: High-dimensional HJI PDEs arising from robust $\mathcal{H}_\infty$ control of PDEs are tackled by spectral and polynomial tensor approaches, with policy iteration providing convergence even for non-quadratic value functions (Kalise et al., 2019).
Mesh-free and PINN-based methods: Physics-informed neural networks integrated with policy iteration permit the solution of nonconvex, high-dimensional stochastic HJI equations. The method iteratively approximates the value function via supervised learning and alternating policy improvement, yielding equi-Lipschitz iterates and proven convergence to the viscosity solution under suitable conditions (Yang et al., 21 Jul 2025).
Graph-based discrete Isaacs equations: Discrete analogues of the HJI equation posed on finite graphs admit fully discrete viscosity theory. Existence and uniqueness are guaranteed through global comparison and monotonicity properties of the (min-max) operators. Such schemes converge to their continuum analogues as graph mesh size vanishes (Forcillo et al., 10 Nov 2025).
Eulerian and Cauchy-type variational schemes: For reachability and optimal control, variational (successive sweep) algorithms based on local Taylor expansions along nominal trajectories allow polynomial-time overapproximation of reachable sets, circumventing the curse of dimensionality for sufficiently smooth HJI data (Molu et al., 2022).
Discontinuous Galerkin and $C^0$ -interior penalty finite element: Fully nonlinear strong-solution theory (under Cordes-type structural conditions) admits quasi-optimal a priori and a posteriori error bounds. The DG/ $C^0$ -IP schemes extend to periodic, homogenization, and effective Hamiltonian computation settings (Kawecki et al., 2020, Kawecki et al., 2021).

4. Structural Generalizations: Path-Dependence, Fractional Dynamics, Delays

Modern developments extend the HJI framework far beyond classical Markovian control:

State and Control Path-Dependence: Value functionals depend on entire state and control histories; the PDE is replaced with a path-dependent HJI, involving functional Itô derivatives. Viscosity solutions are defined in function spaces with Hölder regularity, using backward semigroups and functional Taylor expansions. The Isaacs equivalence and DPP hold under nonanticipative strategies (Moon, 2019).
Fractional and Delay Systems: Caputo and coinvariant derivatives encode memory effects in the system, requiring new viscosity frameworks. In these cases, the corresponding path-dependent HJBI equations are uniquely solved by value functionals in the class of nonanticipative and locally Lipschitz functionals satisfying adapted viscosity inequalities (Gomoyunov, 2021, Plaksin, 2020).
Stochastic, Non-Lipschitzian, and Infinite-Dimensional Extensions: Games with cost functionals given by BSDEs with monotone but possibly non-Lipschitz generators admit a full viscosity solution and regularity theory for the HJBI equation (Wang et al., 2024).

5. Applications: Robust and Stochastic Control, Reachability, Homogenization

HJI equations serve as the core tool in robust control, stochastic games, mean-field interactions, and nonlocal problems:

Robust $\mathcal{H}_\infty$ Feedback for PDEs: Numerical solution of high-dimensional HJI PDEs enables robust feedback synthesis for nonlinear PDEs, achieving strong disturbance rejection and model-uncertainty mitigation—a key step beyond Riccati-based linearizations (Kalise et al., 2019).
Discounted and Infinite-Horizon Control: The discounted HJI PDE, particularly relevant for $\mathcal{H}_\infty$ control over infinite horizons, connects to the theory of contact Hamiltonian systems and the stable manifold approach for controller synthesis (Chen et al., 2024).
Mean Field Games and Singular Perturbations: In slow-fast systems, singular perturbations of HJI equations admit sharp quantitative convergence rates ( $O(\varepsilon)$ ) between fast/slow game values and reduced homogenized equations. This has impact in multi-agent mean-field game analysis and acceleration models (Cannarsa et al., 2024).
Nonlocal and Integro-Differential Models: The Isaacs structure naturally accommodates jump processes, nonlocal diffusion (integro-differential HJI), and rational inattention or robust filtering in environmental and economic modelling (Yoshioka et al., 2021).
Graph and Discrete Control: Discrete HJI operators on graphs model Markov chains, discrete stochastic control and pursuit-evasion games on networks, converging under graph refinement to continuum PDEs (Forcillo et al., 10 Nov 2025).

6. Connections to Dynamic Programming and Isaacs Condition

The HJI equation is the infinitesimal representation of the dynamic programming principle (DPP) for the value functional in zero-sum games. The DPP, via backward semigroups and stochastic calculus, provides the link between optimal game strategies and viscosity solutions to the HJI (or path/HJBI) equation (Qiu et al., 2020, Wang et al., 2024).

The Isaacs condition, i.e., the coincidence of min-sup and sup-min Hamiltonians, is essential for the well-posedness of the unique value and the governing HJI PDE. When the Isaacs condition fails, the upper and lower value functions may not coincide, and the HJI PDE no longer describes a well-defined game value (Moon, 2019). For many stochastic, nonlocal, and delay settings, Isaacs-type conditions and monotonicity remain central to existence and uniqueness theory.

In conclusion, the Hamilton-Jacobi-Isaacs equation is the foundational PDE of dynamic game theory, synthesizing minimax optimization, dynamic programming, and nonlinear PDE analysis in one unifying framework. Advances in solution theory, including path/fractional/delay dependence, robust and nonconvex control, high-dimensional computation, and discrete/graph settings, have established HJI analysis as the technical backbone for modern differential games, stochastic control, and robust feedback design (Moon, 2019, Gomoyunov, 2021, Qiu et al., 2020, Yang et al., 21 Jul 2025, Wang et al., 2024, Kalise et al., 2019, Forcillo et al., 10 Nov 2025, Yoshioka et al., 2021).