---
title: Path-Dependent Hamilton–Jacobi Equations
url: https://www.emergentmind.com/topics/path-dependent-hamilton-jacobi-equations
type: topic
---

# Path-Dependent Hamilton–Jacobi Equations

Path-dependent Hamilton–Jacobi equations are functional partial differential equations in which the unknown depends on an entire past trajectory rather than only on a current state. In the deterministic first-order setting, a canonical form is
\[
\partial_t \varphi(t, x(\cdot)) + H\bigl(t, x(\cdot), \nabla \varphi(t, x(\cdot))\bigr)=0,
\qquad
\varphi(T,x(\cdot))=\sigma(x(\cdot)),
\]
posed on spaces of continuous paths and interpreted through coinvariant or co-invariant derivatives; in stochastic control, second-order path-dependent Hamilton–Jacobi–Bellman and Hamilton–Jacobi–Isaacs equations include horizontal and vertical derivatives in the sense of functional Itô calculus, or stochastic derivatives when the value function is a random field [2204.09275] [2603.16168] [2004.02095] [2107.05959] [2006.13043] [2307.08882]. The subject lies at the intersection of optimal control, differential games, delay systems, stochastic analysis, and infinite-dimensional PDE theory, and its central analytic themes are dynamic programming, generalized solution concepts, comparison principles, and stability.

## 1. Canonical equations and path-space state variables

A basic deterministic Cauchy problem studied in several papers is the first-order path-dependent Hamilton–Jacobi equation on \(C([-h,T],\mathbb{R}^n)\),
\[
\partial_t \varphi(t, x(\cdot)) + H\bigl(t, x(\cdot), \nabla \varphi(t, x(\cdot))\bigr) = 0,
\qquad
\varphi(T, x(\cdot)) = \sigma(x(\cdot)),
\]
where \(\varphi\) is non-anticipative, meaning that \(\varphi(t,x(\cdot))\) depends only on \(x(\cdot \wedge t)\) [2204.09275] [2603.16168] [2010.06190]. This formulation is natural for time-delay systems and other hereditary dynamics because the history segment is part of the effective state.

In controlled stochastic settings, the canonical equation becomes second-order. One representative form is the path-dependent HJB equation
\[
\partial^H_t u(t,x) + \sup_{a\in A} \left\{ \langle b(t,x,a), \partial^V_x u(t,x)\rangle + \frac{1}{2}\operatorname{Tr}\big[(\sigma\sigma^\top)(t,x,a)\partial^V_{xx}u(t,x)\big] + f(t,x,a) \right\} = 0,
\]
associated with controlled SDEs whose coefficients are non-anticipative in the path argument [2107.05959]. A related second-order path-dependent HJB equation arising from path-dependent SDEs and BSDE cost functionals is
\[
\partial_t V(\gamma_t) + H\left(\gamma_t, V(\gamma_t), \partial_x V(\gamma_t), \partial_{xx} V(\gamma_t) \right)=0,
\qquad
V(\gamma_T)=\phi(\gamma_T),
\]
with Hamiltonian
\[
H(\gamma_t, r, p, \iota)=\sup_{u\in U}\left\{\langle p,b(\gamma_t,u)\rangle+\tfrac12\operatorname{tr}\big[\iota \sigma(\gamma_t,u)\sigma(\gamma_t,u)^\top\big]+q(\gamma_t,r,\sigma^\top(\gamma_t,u)p,u)\right\}
\]
[2005.05309].

Further variants replace deterministic PDE structure by stochastic path-dependent evolution. In one formulation, the value function is a random field satisfying
\[
-\mathfrak{d}_t u(t,x_t)-\mathcal{H}(t,x_t,\nabla u(t,x_t))=0,
\qquad
u(T,x_T)=G(x_T),
\]
where the Hamiltonian contains the drift operator on a Gelfand triple [2307.08882]. Another formulation introduces a stochastic path-dependent HJB equation
\[
-\mathcal{O}_t u(t,x_t)-H(t,x_t,\nabla u,\nabla^2 u,\mathcal{D}_\omega \nabla u)=0,
\qquad
u(T,x)=G(x),
\]
for controlled SDEs with random path-dependent coefficients [2006.13043].

A common misconception is that path-dependence can always be removed by augmenting the state with finitely many variables. The supplied literature repeatedly formulates the value function on full path spaces \(C([0,T];\mathbb{R}^d)\), \(D([0,T],\mathbb{R}^d)\), \(C([-h,T],\mathbb{R}^n)\), or \(C([0,T],H)\), and explicitly describes the resulting equations as infinite-dimensional and path-dependent [2107.05959] [2408.02145] [1706.07297].

## 2. Derivative structures on path space

Two derivative frameworks dominate the literature. The first is Dupire’s functional Itô calculus, which uses horizontal and vertical derivatives. In this setting, the horizontal derivative measures pathwise time increments, and the first and second vertical derivatives measure sensitivity to instantaneous perturbations of the current endpoint or future continuation. This framework underlies first-order and second-order path-dependent HJB equations in finite dimensions and Hilbert spaces, as well as functional Itô formulas used in verification arguments [2107.05959] [2004.02095] [2005.05309] [2009.05367].

The second is the coinvariant or co-invariant derivative framework on spaces of continuous histories. A functional \(\varphi\) is ci-differentiable at \((t,x(\cdot))\) if there exist \(\partial_t \varphi(t,x(\cdot))\) and \(\nabla \varphi(t,x(\cdot))\) such that, for admissible extensions \(y(\cdot)\),
\[
\varphi(\tau,y(\cdot))-\varphi(t,x(\cdot))
=
\partial_t \varphi(t,x(\cdot))(\tau-t)
+
\langle \nabla \varphi(t,x(\cdot)), y(\tau)-x(t)\rangle
+o(\tau-t),
\]
with the remainder controlled relative to the path increment [2204.09275] [2603.16168] [2412.17388]. This derivative is tailored to non-anticipative functionals for delay equations and differential games.

In infinite-dimensional Hilbert settings, the derivative structure must also account for unbounded or nonlinear drift operators. For path-dependent Hamilton–Jacobi equations of the form
\[
\partial_t u(t,x)-\langle A(t,x(t)),\partial_x u(t,x)\rangle+F(t,x,\partial_x u(t,x))=0,
\]
the operator term is treated through path derivatives and trajectory-based test functions adapted to nonlinear monotone and coercive operators on a Gelfand triple \(V\subseteq H\subseteq V^*\) [1706.07297]. A later formulation decomposes test functionals into a smooth part and a singular part
\[
\tilde{\varphi}^z(t,x)=\int_{t_0}^t \langle A(s,x(s)),z\rangle\,ds,
\]
which absorbs the singular infinite-dimensional behavior of \(A\) [2509.16015].

The stochastic path-dependent setting introduces additional derivatives with respect to randomness. In the SPHJ literature, \(\mathfrak{d}_t\) is a generalized horizontal derivative and \(\mathfrak{d}_\omega\) is a noise derivative, while another line uses \(\mathcal{O}_t\) and \(\mathcal{D}_\omega\) for the drift and random-field components of the stochastic path-dependent HJB operator [2307.08882] [2006.13043].

## 3. Generalized solution concepts

Viscosity and minimax solutions are the principal generalized solution notions. In the Crandall–Lions path-dependent viscosity framework for second-order HJB equations, a subsolution or supersolution is tested against smooth functionals \(\varphi\) for which \(u-\varphi\) attains a global maximum or minimum on a one-sided set, and the HJB inequality is imposed using horizontal and vertical derivatives [2107.05959]. For first-order path-dependent HJB equations in Dupire form, a notion of viscosity solution is defined via test function classes on compact subsets of path space, and the value functional is characterized as the unique viscosity solution in the corresponding class [2004.02095].

The minimax approach is formulated differently. In the coinvariant framework, a minimax solution is defined by weak invariance properties of the epigraph and subgraph, or equivalently by inequalities involving characteristic-type path evolutions. One representative definition requires, for every \((t,x(\cdot))\), \(\tau\), and \(s\in\mathbb{R}^n\),
\[
\inf_{z(\cdot)\in Lip_{c_H}(t,x(\cdot))}
\left[
\varphi_s(\tau,z_\tau(\cdot))
+
\int_t^\tau H(\xi,z_\xi(\cdot),s)\,d\xi
\right]
\le \varphi_s(t,x(\cdot)),
\]
with the corresponding reversed inequality for the lower solution [2204.09275]. A related 2026 formulation defines upper and lower solutions through integral inequalities of the type
\[
\varphi(\tau,y(\cdot))-\int_t^\tau \left(\langle s,\dot y(\xi)\rangle-H(\xi,y(\cdot),s)\right)d\xi \le \varphi(t,x(\cdot)),
\]
again with reversed inequality for the opposite semisolution [2603.16168].

Several papers prove equivalence results between these notions, but only under specific frameworks. For path-dependent Hamilton–Jacobi equations with coinvariant derivatives over the space of continuous functions, minimax and viscosity solutions are proved equivalent, yielding comparison and uniqueness as corollaries [2204.09275]. In infinite-dimensional Hilbert spaces with nonlinear monotone and coercive operators, a new notion of viscosity solution is shown equivalent to the minimax notion, and this equivalence implies existence and uniqueness for viscosity solutions as well [2509.16015].

Other generalized notions also appear. For path-dependent Hamilton–Jacobi equations with super-quadratic growth in the gradient, Dini subsolutions, Dini supersolutions, maximal Dini subsolutions, and lower semicontinuous minimax solutions are used instead of a standard viscosity framework, particularly in connection with vanishing viscosity limits and lower semicontinuous terminal costs [2102.00038]. In state-and-control path-dependent stochastic zero-sum games, viscosity solutions are defined on compact subsets of a \(\kappa\)-Hölder space, with test functionals satisfying a predictable dependence condition on control histories [1911.00315].

## 4. Dynamic programming, comparison, uniqueness, and stability

The dynamic programming principle is the structural bridge from control and game formulations to path-dependent Hamilton–Jacobi equations. For path-dependent stochastic control under general conditions, one rigorous form is
\[
v(t,x)=\sup_{\alpha}\mathbb{E}\left[\int_t^s f(r,X^{t,x,\alpha},\alpha_r)\,dr+v(s,X^{t,x,\alpha})\right],
\qquad t\le s\le T,
\]
and this principle yields the associated path-dependent HJB equation [2107.05959]. A 2012 paper on stochastic functional differential systems already stated, at the abstract level, that within the framework of functional Itô calculus it builds the dynamic programming principle and the related path-dependent HJB equation, and proves that the value function is the viscosity solution of that equation [1207.1194].

Comparison principles are the core uniqueness mechanism, but the proofs vary sharply across frameworks. For Crandall–Lions viscosity solutions of path-dependent HJB equations, uniqueness is proved by showing that the value function is bigger than any viscosity subsolution and smaller than any viscosity supersolution; the argument uses a smooth variational principle, gauge functions, cylindrical approximations by finite-dimensional PDEs, and regularity results for parabolic equations, rather than a direct doubling-of-variables proof [2107.05959]. In the first-order Dupire framework, uniqueness is obtained by adapting doubling-of-variables arguments through auxiliary functionals equivalent to the square of the sup norm [2004.02095]. In the revised minimax theory for ci-equations, a Lyapunov–Krasovskii functional equivalent in some sense to the square of the uniform norm of the path variable provides the key comparison tool [2010.06190].

More recent uniqueness work relaxes assumptions on the Hamiltonian. For path-dependent Hamilton–Jacobi equations with time-measurable Hamiltonians and \(u\)-dependence, minimax comparison and Perron-type existence yield existence and uniqueness under broad measurability and monotonicity hypotheses [2408.02145]. A 2026 uniqueness theory defines viscosity solutions via coinvariantly smooth test functionals and a dense family of compact subsets of the space of continuous functions, proving uniqueness either in the class of continuous viscosity solutions under sublinear growth of the local Lipschitz constant in the gradient variable, or in a locally Lipschitz subclass in the general case; the proofs use a standard doubling method combined with a novel penalty functional [2604.25305].

Stability is a recurrent well-posedness property. Stability theorems are stated for first-order and second-order path-dependent HJB equations in finite dimensions and Hilbert spaces, for minimax solutions under measurable-in-time Hamiltonians, and for infinite-dimensional viscosity solutions using half-relaxed limits [2004.02095] [2005.05309] [2009.05367] [2603.16168] [2509.16015]. A plausible implication is that stability has become a defining criterion for which generalized solution notions are analytically useful, especially when approximation by finite-dimensional, time-regularized, or mollified problems is essential.

## 5. Infinite-dimensional, stochastic, and vanishing-viscosity extensions

The theory extends beyond finite-dimensional deterministic path spaces. In infinite dimensions, one line studies fully nonlinear path-dependent PDEs with nonlinear, monotone, hemicontinuous, bounded, and coercive operators on Hilbert space, proving well-posedness for minimax solutions and introducing a hybrid minimax–viscosity methodology for comparison [1706.07297]. Another line develops second-order path-dependent HJB equations in Hilbert spaces for stochastic evolution equations, identifies the value functional as the unique viscosity solution, and proves consistency with classical solutions and a stability property [2009.05367].

Stochastic path-dependence introduces randomness into the value function itself. For controlled ordinary differential equations with random path-dependent coefficients, the value function becomes a random field on the path space and is characterized by a stochastic path-dependent Hamilton–Jacobi equation; a notion of viscosity solution is proposed and uniqueness is proved for that random-field equation [2005.01232]. For controlled stochastic differential equations with random path-dependent coefficients, the associated stochastic path-dependent HJB equation is studied together with notions of viscosity solution and classical solution, and the value function is proved to be the viscosity solution; a uniqueness result for viscosity solutions is given for certain superparabolic cases, while uniqueness of classical solutions is addressed more generally [2006.13043]. In infinite-dimensional stochastic control with measurable randomness and path-dependence, the value function is again a random field on path space and is characterized by a stochastic path-dependent Hamilton–Jacobi equation with a viscosity theory adapted to the Gelfand-triple setting [2307.08882].

A distinct extension is the non-Markovian vanishing-viscosity method. For path-dependent Hamilton–Jacobi equations with super-quadratic growth in the gradient, uniqueness of maximal subsolutions is obtained for viscous equations related to convex super-quadratic BSDEs, and well-posedness is established for the first-order Hamilton–Jacobi–Bellman equation associated to a Bolza problem of the calculus of variations with path-dependent terminal cost [2102.00038]. The value functions \(v_n\) for viscous problems converge uniformly on compacta to the first-order value function \(v_0\), which is characterized as the unique lower semicontinuous minimax solution or, in an alternative formulation, as the unique bounded maximal Dini subsolution [2102.00038].

These results show that “path-dependent Hamilton–Jacobi equations” is not a single analytic framework but a family of closely related frameworks indexed by the choice of path space, derivative notion, stochastic structure, and operator class.

## 6. Optimal control, differential games, and related models

The subject is driven by applications. In deterministic optimal control of functional differential equations with delays and discount factors, the value function satisfies a path-dependent Hamilton–Jacobi–Bellman equation with time-measurable data, and is the unique minimax solution under broad assumptions [2408.02145]. In stochastic optimal control of path-dependent SDEs or stochastic evolution equations, the value functional is identified as the unique viscosity solution of the associated second-order PHJB equation, including formulations with BSDE cost functionals and applications to backward stochastic HJB equations [2005.05309] [2009.05367].

Differential games supply the Isaacs counterpart. For zero-sum games with time-delay systems, the lower and upper value functionals are represented by minimax or viscosity solutions of path-dependent Hamilton–Jacobi equations with Hamiltonians
\[
H^-(\tau,y(\cdot),s)=\max_{v\in Q}\min_{u\in P}\big(\langle s,f(\tau,y(\cdot),u,v)\rangle-\chi(\tau,y(\cdot),u,v)\big),
\]
\[
H^+(\tau,y(\cdot),s)=\min_{u\in P}\max_{v\in Q}\big(\langle s,f(\tau,y(\cdot),u,v)\rangle-\chi(\tau,y(\cdot),u,v)\big),
\]
and the Isaacs condition yields existence of the game value [2603.16168]. In infinite-dimensional time-delay evolution equations, both Elliott–Kalton and Krasovskii–Subbotin strategy frameworks are treated, with the value characterized by the unique viscosity or minimax solution of the associated path-dependent Isaacs equation [2509.16015] [1706.07297].

The class of admissible dynamics is broader than delayed differential equations. For zero-sum games governed by nonlinear Volterra integral equations of Hammerstein type with weakly singular kernel,
\[
x(\tau)=y(\tau)+\int_0^\tau K(\tau,\xi)f\big(\xi,x(\xi),u(\xi),v(\xi)\big)\,d\xi,
\]
the lower and upper value functionals are viscosity solutions of a path-dependent Hamilton–Jacobi equation on a space of system positions, and under the Isaacs condition the common value is the unique viscosity solution [2404.10428]. In stochastic zero-sum games with state and control path-dependence, the lower and upper Hamilton–Jacobi–Isaacs equations are second-order state-and-control path-dependent nonlinear PDEs, and the lower and upper value functionals satisfy the dynamic programming principle and are viscosity solutions in a compact Hölder-space setting [1911.00315].

A plausible implication is that the path-dependent Hamilton–Jacobi formalism now functions as a unifying analytic language for delay systems, hereditary control, non-Markovian stochastic control, Volterra dynamics, and infinite-dimensional games.

## 7. Regularity, current directions, and unresolved issues

Regularity theory has become increasingly explicit. For minimax solutions of first-order path-dependent Hamilton–Jacobi equations with coinvariant derivatives, Lipschitz continuity in both time and functional variables has been proved under assumptions on the Hamiltonian and the boundary functional; the path-variable estimates are given both in the uniform norm and in a special norm combining endpoint and \(L^2\)-history terms [2412.17388]. This regularity is significant because some uniqueness frameworks require continuity or local Lipschitz conditions on candidate solutions [2604.25305].

Recent work also weakens structural assumptions on the Hamiltonian. Time-measurable dependence, \(u\)-dependence, and Carathéodory-type assumptions have been incorporated into minimax well-posedness and game-value results [2408.02145] [2603.16168]. In the infinite-dimensional Hilbert setting, existence on the whole path space and stability via half-relaxed limits have been obtained under more general assumptions than in the earlier literature [2509.16015].

At the same time, the literature does not present a single universal notion of generalized solution. Dupire-based viscosity solutions, coinvariant viscosity solutions, minimax solutions, Dini solutions, and stochastic viscosity notions coexist, with equivalence results proved only in certain settings [2204.09275] [2509.16015] [2102.00038]. A concrete unresolved point appears in stochastic zero-sum differential games with both state and control path-dependence: the supplied summary states that uniqueness of viscosity solutions is not fully established in that most general setting and is left as open future work, although uniqueness of classical solutions is proved in the state-path-dependent case [1911.00315].

The current research trajectory therefore combines two movements. One is expansion of admissible data—measurable-in-time Hamiltonians, broader operator classes, and more general path spaces. The other is consolidation—comparison, equivalence, stability, and regularity results that identify when distinct generalized formulations describe the same value functional.

Source: https://www.emergentmind.com/topics/path-dependent-hamilton-jacobi-equations