The Hamilton–Jacobi equation is a partial differential equation for a scalar function whose spatial derivatives enter through a prescribed function called the Hamiltonian. Its standard time-dependent form is
where denotes position, denotes time, and is the spatial gradient of . In classical mechanics, can represent an action function, and its gradient gives momentum. In optimization and control, it can represent an optimal cost; in wave propagation, it can represent a phase or an arrival time.
The equation is important because it converts problems about families of trajectories into a single equation for a scalar field. Its classical theory connects differential equations, mechanics, and the calculus of variations. Its modern theory addresses the singularities that typically arise even from smooth initial data, using generalized solutions selected by comparison principles, variational formulas, and small-diffusion limits. These developments have made Hamilton–Jacobi equations central to optimal control, front propagation, numerical computation, and the geometry of dynamical systems.
Mathematical form and historical origins
For , a Hamiltonian is a function , with representing momentum or a gradient variable. The initial-value problem is commonly written
It is first order because only first derivatives of appear, and it is generally nonlinear because need not depend linearly on . For instance, the Hamiltonian of a particle of mass moving in a potential is
which gives
Here the action has units of energy multiplied by time, its spatial gradient has units of momentum, and every term in the equation has units of energy.
More generally, a first-order equation may be written . Stationary Hamilton–Jacobi equations form an important subclass, including , where is a constant, and equations with explicit dependence on , such as . An initial condition determines an evolving field, whereas a stationary boundary-value problem may specify arrival times, costs, or values on part of a boundary. Boundary conditions are not interchangeable with those for second-order elliptic equations: characteristics can enter or leave the domain, and prescribing arbitrary data everywhere may be incompatible with a classical solution.
William Rowan Hamilton developed his characteristic-function approach in papers on dynamics published in 1834 and 1835, building on his earlier work in geometrical optics. Carl Gustav Jacob Jacobi subsequently developed the integration theory, with an important memoir appearing in 1837. Their work showed that solving one first-order equation for an action function could encode the solutions of an entire mechanical system. The resulting theory extended the analytical mechanics associated with Leonhard Euler and Joseph-Louis Lagrange.
During the twentieth century, the equation acquired interpretations far beyond particle mechanics. Richard Bellman’s dynamic programming linked it to optimal decision-making, while control theory produced Hamilton–Jacobi–Bellman equations. The development of viscosity solutions by Michael Crandall and Pierre-Louis Lions in the early 1980s, including their foundational 1983 paper, supplied a framework for solutions that are continuous but not differentiable. This changed the central question from finding smooth formulas alone to identifying a unique, stable solution after smoothness has broken down.
Action principles and characteristic trajectories
The mechanical origin of the equation is a variational problem. Let be a Lagrangian, depending on position, velocity, and time. For a path , its action is
Stationary paths satisfy the Euler–Lagrange equations
When the velocity-to-momentum map is invertible, the Hamiltonian is obtained by the Legendre transform:
For a strictly convex Lagrangian with suitable growth, this is equivalently the supremum of over velocities.
Suppose that is the action along a smoothly varying family of extremal paths from a fixed starting event to . Endpoint variation gives
and hence the Hamilton–Jacobi equation. This statement is local: a single smooth branch can cease to exist when several extremal paths reach the same endpoint. In minimum-action problems, taking the least action among competing paths gives a single-valued function, but that function need not remain differentiable.
Conversely, a smooth solution generates trajectories through the method of characteristics. For , they satisfy
with along a characteristic. The first two equations are Hamilton’s canonical equations. The third follows from the chain rule:
When the Legendre transform applies, , so the change in along a characteristic is the accumulated Lagrangian action.
For a free particle, , momentum is constant. A characteristic starting at has
This formula produces a smooth solution only while the map from to remains locally invertible. Its derivative is
If this matrix becomes singular, the projection of the characteristic family into position space can fold. Distinct trajectories may then arrive at the same position with different momenta. The characteristic system itself can remain smooth while the scalar solution loses smoothness; this distinction is fundamental to the theory.
Complete integrals, canonical transformations, and separation
In classical integration theory, a complete integral is a family of solutions depending on essential parameters , with a nondegeneracy condition such as
An additive constant is usually treated separately. The parameters supply enough independent information to recover the trajectories of a mechanical system with configuration coordinates. This is stronger than finding one particular solution of the partial differential equation.
The connection is expressed through a canonical transformation generated by :
For a time-dependent generating function of this type, the transformed Hamiltonian is . If satisfies the Hamilton–Jacobi equation, then , so the new variables and are constant along trajectories. Recovering and then becomes a problem of solving these implicit relations. The partial differential equation has effectively replaced integration of the original coupled ordinary differential equations.
For an autonomous Hamiltonian, the standard ansatz is
which reduces the equation to
Here is Hamilton’s characteristic function and is the energy. If suitable coordinates permit
the stationary equation separates into ordinary differential equations, with separation constants encoding conserved quantities. Separation is a special structural property of a system and its coordinates, not a general feature of Hamilton–Jacobi equations.
For a one-dimensional harmonic oscillator,
the stationary equation gives
Integrating yields on branches between turning points. The relation recovers the oscillatory trajectory. In a central-force problem, polar or spherical coordinates separate angular momentum from radial motion. For planar motion, the radial equation is
where is the conserved angular momentum.
In integrable systems, separated solutions lead to action–angle variables. Actions are defined by integrals over independent cycles,
When the Hamiltonian depends only on the actions, the conjugate angles advance at frequencies . Such coordinates describe regular quasiperiodic motion, but turning points, global topology, and nonintegrability can prevent one smooth generating function from covering an entire phase space.
Geometrical optics and the semiclassical limit
In geometrical optics, a wave is described locally by rapidly varying phase and more slowly varying amplitude. Substitution of a high-frequency ansatz into a wave equation separates the leading phase equation from lower-order transport equations for amplitude. The leading equation is an eikonal equation, a form of Hamilton–Jacobi equation. It describes wavefront geometry without resolving every oscillation of the wave.
For an isotropic medium with positive propagation speed , the first-arrival time from a source satisfies
away from the source and appropriate singular sets. Equivalently, , where is slowness. If the speed is constant and the source is the origin, . In heterogeneous media, first-arrival times can be characterized by minimizing
over admissible paths. This is the travel-time form of Fermat’s principle, associated with Pierre de Fermat. Refraction results from changes in the local travel-time cost.
A related evolutionary formulation represents a moving interface as the zero set of a function . If the interface moves with normal speed , measured in the direction , then
This is the basic equation of the level-set method. Corners and changes of topology can be represented without separately parameterizing each component of the interface. The sign of the speed term depends on the chosen orientation and on whether positive denotes the inside or outside.
Hamilton–Jacobi theory also emerges from quantum mechanics. For the Schrödinger equation
write locally , with real and . Where this decomposition is smooth and , the real part gives
If the amplitude varies on scales for which the final term is negligible, the leading phase satisfies the classical Hamilton–Jacobi equation. The imaginary part supplies an amplitude transport equation.
This connection underlies the WKB approximation, but it is not uniformly valid everywhere. At turning points or caustics, the simplest amplitude approximation can become singular, and several phase branches may contribute to the wave. A minimum-action viscosity solution selects a single scalar value and generally does not preserve all interference information. Wave propagation and viscosity theory therefore share a Hamiltonian structure without supplying interchangeable descriptions of the underlying physical field.
Loss of smoothness and viscosity solutions
Smooth initial data do not generally produce globally smooth Hamilton–Jacobi solutions. For the one-dimensional equation
differentiation gives, wherever is smooth,
This is the inviscid Burgers equation. Regions with larger move faster, so characteristics can collide. The gradient may develop a jump while remains continuous. After that event, the equation cannot be interpreted pointwise everywhere.
Requiring the equation to hold almost everywhere is often too weak to select a unique solution. Unlike a conservation law, a general Hamilton–Jacobi equation does not have a canonical distributional formulation, because nonlinear functions of distributional derivatives need not be defined. The useful generalized notion is instead based on smooth test functions that touch the candidate solution from above or below.
For
an upper semicontinuous function is a viscosity subsolution if every smooth for which has a local maximum at an interior point satisfies
there. A lower semicontinuous viscosity supersolution satisfies the reverse inequality at local minima of . A continuous function that is both is a viscosity solution. These tests impose directional information at corners without assigning a nonexistent classical gradient. Every classical solution is a viscosity solution, and the viscosity equation agrees with the classical equation wherever an appropriate solution is differentiable.
The decisive uniqueness mechanism is a comparison principle: under suitable assumptions, a subsolution lying below a supersolution on the relevant initial or boundary set remains below it in the domain. The assumptions depend on the problem and can involve continuity, controlled spatial dependence, coercivity, boundedness, or growth conditions. Comparison is not automatic for every Hamiltonian and every boundary prescription. When it holds, it gives uniqueness and order stability, and often yields a contraction estimate such as
for suitable evolution problems.
The name “viscosity solution” reflects the selection obtained from small diffusion:
Under appropriate compactness and comparison hypotheses, solutions converge as to the viscosity solution. The Laplacian smooths the approximation, but the limit need not be differentiable. More generally, viscosity solutions are stable under suitable locally uniform limits of equations and solutions. This stability makes the definition effective for existence proofs, approximation schemes, and limits involving rapidly varying coefficients.
Variational formulas and conservation laws
For a Hamiltonian depending only on momentum, explicit variational representations are available under convexity and growth assumptions. Suppose is convex and superlinear, and define
The Hopf–Lax formula is
With appropriate hypotheses on the initial function, this gives the viscosity solution of . The formula minimizes the initial value plus the action required to travel from to in time .
For , the Legendre transform is , so
If the minimizer is unique and the necessary differentiability holds, then
Multiple minimizers indicate competing optimal histories and commonly correspond to nondifferentiability of . The formula makes the selection mechanism visible: rather than retaining every characteristic branch, the solution takes the smallest admissible cost.
For example, in one dimension let . The quadratic Hopf–Lax formula gives
For , the minimizing starting point is ; for , it is . At , both and minimize. The solution has a persistent downward corner. Its derivative is to the left and to the right, forming a stationary entropy shock for Burgers’ equation. Thus the loss of differentiability records a genuine collision of selected characteristics rather than a defect in the solution concept.
For one space dimension and sufficiently regular , differentiation formally connects Hamilton–Jacobi equations with scalar conservation laws:
Under suitable assumptions, the derivative of the viscosity solution corresponds to the entropy solution of the conservation law. The two admissibility theories are closely related, but their unknowns differ: remains continuous while its derivative can jump. In several dimensions, the gradient satisfies a constrained vector system, so the relation is not simply the full theory of scalar conservation laws in higher dimensions.
For position- and time-dependent Hamiltonians, straight-line costs are replaced by a path minimization:
The associated evolution operators are often called the Lax–Oleinik operators. Their composition property expresses the fact that an optimal path can be divided at an intermediate time without changing the optimal total cost. This property is also the bridge from variational mechanics to dynamic programming.
Optimal control, stochastic dynamics, and differential games
In deterministic optimal control, a state follows
where is an admissible control. With running cost , terminal cost , and terminal time , define the value function
The dynamic programming principle states that optimization over the full interval can be split into optimization over a short first interval followed by the optimal remaining cost. A smooth expansion over that short interval yields the Hamilton–Jacobi–Bellman equation.
For a terminal-value problem, a standard formulation is
At smooth points, this is equivalent to
For generalized solutions, the time orientation and subsolution conventions must be kept consistent. This is a backward problem: the terminal cost determines values at earlier times. Under suitable assumptions, the value function is characterized as its viscosity solution, even when an optimal feedback changes discontinuously or has corners.
As a basic example, consider motion with bounded speed, , , and unit running cost until a target is reached. The stationary minimum travel time satisfies
hence , with zero value on the target. The same eikonal equation therefore describes both first-arrival waves and time-optimal motion. More elaborate control constraints produce anisotropic Hamiltonians, for which the cost of moving in a direction depends on that direction.
If the state follows a stochastic differential equation,
the corresponding equation includes a second-order term:
Here is a Wiener process. This is a possibly degenerate second-order Hamilton–Jacobi–Bellman equation. The diffusion term arises from the generator of the stochastic dynamics, not merely from numerical smoothing.
In differential games, two players influence the dynamics or cost, producing minimax Hamiltonians. Exchanging the order of minimization and maximization need not leave the Hamiltonian unchanged. Equality of the two pointwise optimization orders is the Isaacs condition, which is important in establishing a common game value. These equations occur in pursuit–evasion, robust control, and reachability analysis. Practical use is constrained by the curse of dimensionality: a grid with nodes in each of state coordinates contains nodes.
Numerical approximation and front computation
Numerical methods must capture the selected generalized solution, not merely approximate derivatives accurately where the solution is smooth. Centered spatial differences can generate oscillations near corners or select inappropriate limits. Upwind methods instead use the direction of characteristic propagation. For a one-dimensional equation, a typical explicit finite-difference scheme is
where and are backward and forward differences, and is a consistent numerical Hamiltonian.
For , a local Lax–Friedrichs numerical Hamiltonian takes the form
where bounds over the relevant gradient range. The final term introduces numerical dissipation. With a suitable relation between and the grid spacing, the update is monotone: increasing input grid values cannot decrease the corresponding updated value. This order property is distinct from high-order accuracy and is central to robust approximation of viscosity solutions.
The Barles–Souganidis convergence framework, published in 1991, shows that monotone, stable, consistent schemes converge to the viscosity solution when the limiting equation has an appropriate comparison principle and the other required hypotheses hold. Consistency connects the discrete equation to smooth test functions; stability bounds the approximations; monotonicity supplies the order structure needed at nonsmooth points. Higher-order ENO and WENO reconstructions improve resolution in smooth regions, but their convergence cannot be inferred merely from formal order. Filtered schemes combine a reliable monotone method with a more accurate approximation where the latter is admissible.
For isotropic eikonal equations with positive speed, the fast marching method exploits the causal ordering of increasing arrival times. Grid values are accepted in approximately increasing order using local upwind equations. A binary-heap implementation commonly has complexity for grid nodes. Fast sweeping methods instead make alternating directional passes through the grid; their efficiency depends on the characteristic geometry and the number of sweeps required. Standard isotropic fast marching is not directly suitable for every anisotropic Hamiltonian, because its local acceptance rule relies on a particular causality structure.
Semi-Lagrangian schemes approximate the control interpretation directly, tracing short trajectories and combining interpolation with local optimization. They can permit larger time steps than some explicit finite-difference schemes, although monotonicity, interpolation error, and minimization accuracy remain important. High-dimensional alternatives include sparse representations, low-rank methods, and learned approximations. Their computational advantages do not by themselves provide the comparison-based guarantees available for classical monotone grid schemes; verification of approximation quality remains a separate issue.
Geometric structure, long-time behavior, and homogenization
On a configuration manifold , momentum belongs naturally to the cotangent bundle . A smooth function defines the graph of the covector field . In symplectic geometry, this graph is a Lagrangian submanifold: it has half the dimension of the cotangent bundle, and the symplectic form restricts to zero on it. Hamiltonian evolution moves such a graph through phase space. A Hamilton–Jacobi solution describes the evolution only as long as the resulting submanifold can still be represented as a suitable graph over configuration space.
This viewpoint explains why classical solutions can fail without any singularity in the underlying Hamiltonian flow. The evolving Lagrangian submanifold may remain smooth while its projection onto folds. These projection singularities form caustics. A single-valued viscosity solution selects an admissible scalar branch, whereas geometrical and semiclassical descriptions can retain a more elaborate multibranch structure. Moreover, not every globally defined Lagrangian graph is the differential of a single-valued function; global exactness is an additional topological condition.
Long-time analysis often leads to the stationary additive eigenvalue problem
If this equation has an appropriate solution, then solves the autonomous evolution equation. Under additional assumptions, general solutions may approach this form, but neither convergence nor uniqueness of is automatic. On compact manifolds, convexity and superlinearity in momentum support the weak KAM theory developed by Albert Fathi in the 1990s. It relates stationary viscosity solutions, minimizing trajectories, and invariant sets from Aubry–Mather theory. In that setting, additive constants in remain irrelevant, while the admissible value of is a central dynamical quantity.
In homogenization, a representative rapidly oscillating problem is
When the spatial dependence is periodic and suitable coercivity and regularity conditions hold, the solutions can converge to an effective equation
The effective Hamiltonian is characterized through a cell problem: for each fixed momentum , one seeks a periodic corrector and a constant satisfying
Correctors are generally interpreted in the viscosity sense. The effective Hamiltonian is determined by how paths exploit the microscopic medium, rather than by a simple pointwise average of the original Hamiltonian.
These problems connect the local equation to global transport and minimizing dynamics. Coercivity, often expressed as growth of as , helps control gradients and obtain compactness. Convexity supports variational representations and effective-Hamiltonian constructions, while nonconvex Hamiltonians can require different arguments and may behave differently in homogenization or long-time limits. The existence of stationary solutions, the structure of their singular sets, and the asymptotic behavior of evolution therefore depend on the specific Hamiltonian, domain, and boundary conditions.