Hamilton-Jacobi Equation

Definition

The Hamilton-Jacobi equation is a single partial differential equation for one scalar function — the action S(q, t):

∂S/∂t + H(q, ∂S/∂q, t) = 0

The trick is in the second argument of the Hamiltonian. Wherever a canonical momentum p would normally appear, you substitute the gradient of S:

p = ∂S/∂q

That single substitution turns the Hamiltonian — a function of position and momentum — into a function of position and the slope of S. The result is one PDE whose solution, called Hamilton's principal function, contains the entire dynamics of the system. There are no separate equations of motion to integrate: once you know S, you read the trajectories off by differentiation.

S is not an abstract bookkeeping device. It is the classical action:

S(q, t) = ∫ L dt    (along the actual path from a fixed start to (q, t))

where L = T − V is the Lagrangian. The action you minimize in the principle of least action is the very same S that solves the Hamilton-Jacobi equation.

How it works — wavefronts and rays

The deepest way to read the equation is geometric. Fix a value of S and ask: where in space does the action equal that value at time t? The answer is a surface — a wavefront of constant action. As t advances, that surface moves through configuration space like the crest of a wave.

Now consider the momentum. Because p = ∂S/∂q is the gradient of S, momentum always points in the direction of steepest increase of S — that is, perpendicular to the surfaces of constant action. A particle's velocity is parallel to its momentum, so:

• Surfaces of constant S are the wavefronts.
• Trajectories are the rays, always orthogonal to those wavefronts.

This is precisely the relationship between wavefronts and light rays in geometric optics. Hamilton spotted the analogy in the 1830s while studying optics, then realized the same mathematics governs mechanics. A mechanical system propagates like a wave whose rays are the orbits — the foundation of what physicists call the optical-mechanical analogy. The same picture underlies Fermat's principle and Huygens' construction.

The time-independent form

If the Hamiltonian does not depend explicitly on time (energy is conserved), you can separate the time variable. Write:

S(q, t) = W(q) − E·t

Here W(q) is Hamilton's characteristic function and E is the conserved energy. Substituting back, the −E·t term supplies ∂S/∂t = −E, and the equation collapses to the time-independent Hamilton-Jacobi equation:

H(q, ∂W/∂q) = E

For a single particle of mass m in a potential V(q), with H = p²/2m + V, this reads:

(1/2m)(∂W/∂q)² + V(q) = E

Rearranged, ∂W/∂q = ±√(2m[E − V]), so the momentum at every point is fixed by energy conservation — and W is just the accumulated ∫ p dq. This equation is identical in form to the eikonal equation of geometric optics, (∇W)² = n², with the local "index of refraction" playing the role of √(2m[E−V]). Mechanics literally is ray optics in a medium whose refractive index is set by the potential.

A worked example with numbers — free particle in 1D

Take the simplest possible case: a free particle of mass m = 2 kg, no potential (V = 0), energy E = 4 J. The time-independent HJ equation is:

(1/2m)(dW/dx)² = E
(dW/dx)² = 2mE = 2·2·4 = 16  (kg²·m²/s²)
dW/dx = √16 = 4 kg·m/s

So p = ∂W/∂x = 4 kg·m/s — constant, as expected for a free particle. Integrating, W(x) = 4x, and the full principal function is:

S(x, t) = W(x) − E·t = 4x − 4t

Now extract the trajectory. The Jacobi prescription says set ∂S/∂E = constant = β:

∂S/∂E = (∂W/∂E) − t = (x·√(m/2E)) − t = β
With m=2, E=4:  √(m/2E) = √(2/8) = 0.5
0.5·x − t = β   ⟹   x(t) = 2t + 2β

That is a straight line moving at constant velocity v = 2 m/s. Check it against elementary mechanics: KE = ½mv² = ½·2·2² = 4 J = E. ✓ And p = mv = 2·2 = 4 kg·m/s = ∂W/∂x. ✓ The Hamilton-Jacobi machinery reproduces the obvious answer — but the same recipe (write S, separate, differentiate with respect to constants) works unchanged for the Kepler problem, a charged particle in a field, or a symmetric top, where the elementary route is hopeless.

The bridge to quantum mechanics — ψ ~ e^(iS/ℏ)

Here is why physicists revere this equation. Take the time-dependent Schrödinger equation and substitute a wavefunction of the form:

ψ(q, t) = A(q, t) · e^(iS(q,t)/ℏ)

Expand in powers of ℏ. The leading term (order ℏ⁰) is exactly the Hamilton-Jacobi equation:

∂S/∂t + (1/2m)(∂S/∂q)² + V = 0

The next term (order ℏ¹) is a continuity equation that propagates the amplitude A along the rays. This is the WKB approximation. The physical message is striking: the quantum phase is the classical action divided by ℏ. Because S/ℏ is enormous for macroscopic motion, the phase oscillates so fast that all paths cancel by destructive interference — except where S is stationary. Stationary action is exactly the principle of least action, so the surviving paths are the classical trajectories. As ℏ → 0, quantum mechanics collapses onto Hamilton-Jacobi.

How big is S/ℏ? For a 0.5 kg ball thrown for 1 second with action of order S ≈ E·t ≈ 1 J·s, the phase S/ℏ ≈ 1 / (1.05×10⁻³⁴) ≈ 10³⁴ radians. That astronomically large number is why the everyday world looks classical. For an electron in an atom, S ~ ℏ, the phase is order 1, and interference cannot be ignored — you must use the full Schrödinger equation. Schrödinger himself reverse-engineered his equation from exactly this analogy: he asked what wave equation would have the Hamilton-Jacobi equation as its short-wavelength limit.

Variants and related equations

Form	Equation	Used for
Time-dependent HJ	∂S/∂t + H(q, ∂S/∂q, t) = 0	General dynamics, time-varying forces
Time-independent HJ	H(q, ∂W/∂q) = E	Conservative systems; separation of variables
Eikonal equation	(∇W)² = n²(q)	Geometric optics — the optical twin of HJ
Action-angle form	Jᵢ = ∮ pᵢ dqᵢ, ωᵢ = ∂H/∂Jᵢ	Integrable systems, orbital frequencies, old quantum theory
WKB / semiclassical	ψ ~ A·e^(iS/ℏ)	Tunneling rates, quantization conditions, classical limit
Hamilton-Jacobi-Bellman	∂V/∂t + min_u[c + (∂V/∂x)·f] = 0	Optimal control, reinforcement learning, finance

Separation of variables and integrability

The reason Hamilton-Jacobi is so powerful in practice is separation of variables. For a conservative system you first peel off time with S = W − Et. Then you guess that the characteristic function splits into a sum over coordinates:

W(q₁, q₂, …, qₙ) = W₁(q₁) + W₂(q₂) + … + Wₙ(qₙ)

When this works, each coordinate decouples and brings its own separation constant, which is a conserved quantity. A system that separates completely is called integrable: it has as many conserved quantities as degrees of freedom. The separation constants are the new momenta of a canonical transformation that makes the dynamics trivial — every coordinate either stays constant or advances linearly in time.

The Kepler problem separates in spherical coordinates, giving energy, total angular momentum, and the z-component of angular momentum as the three constants. The action variables Jᵢ = ∮ pᵢ dqᵢ built from these were exactly what Bohr and Sommerfeld quantized (Jᵢ = nᵢh) in the old quantum theory — the direct ancestor of the modern quantum numbers.

Common pitfalls and misconceptions

• "S is a potential energy." No. S is the action, with units of energy × time (joule-seconds), the same units as angular momentum and as ℏ. Its gradient is momentum and its time derivative is −energy; S itself is neither.
• "You need to solve Hamilton's equations first." The opposite — a complete solution of the single HJ PDE replaces the entire set of 2n ordinary differential equations. You get the orbit by differentiating S with respect to the constants, not by integrating forces.
• "The wavefronts move at the particle's speed." They don't. As in optics, the wavefronts (surfaces of constant S) and the rays (trajectories) move at different speeds; the wavefront "phase velocity" can even exceed the particle speed. Only the directions are locked: rays are perpendicular to wavefronts.
• "It only works for separable systems." The equation is always valid; it is only easy when the system separates. Chaotic, non-integrable systems still obey HJ, but a smooth global S may not exist — it develops caustics and multivaluedness, which is exactly where the semiclassical WKB approximation breaks down and needs Maslov corrections.
• "ψ ~ e^(iS/ℏ) is exact." It is the leading semiclassical term. The amplitude A and higher ℏ corrections matter near turning points and caustics, where the simple WKB form diverges.
• "HJ is just a curiosity." Its optimal-control descendant, the Hamilton-Jacobi-Bellman equation, is the backbone of modern control theory and reinforcement learning. The same PDE-with-optimal-characteristics idea runs spacecraft trajectories and trains robots.

Where it shows up

• Celestial mechanics and orbits. Separating the HJ equation gives action-angle variables and the slow precession frequencies of planetary orbits — the natural language of perturbation theory.
• Semiclassical quantum mechanics. Tunneling rates, Bohr-Sommerfeld quantization, and the WKB connection formulas all flow from ψ ~ e^(iS/ℏ).
• General relativity. Geodesics of test particles are found by separating the HJ equation in metrics like Schwarzschild and Kerr — how light bending and orbital precession around black holes are computed.
• Geometric optics. The eikonal equation that designs lenses and traces rays is the HJ equation in disguise (see Fermat's principle).
• Optimal control and AI. The Hamilton-Jacobi-Bellman equation underlies trajectory optimization, autonomous-vehicle planning, and value functions in reinforcement learning.
• Plasma and accelerator physics. Canonical perturbation theory built on HJ describes particle drifts and adiabatic invariants in magnetic confinement.

Derivation analysis — why one PDE beats 2n ODEs

The Hamilton-Jacobi equation arises from a canonical transformation. We look for a transformation from old variables (q, p) to new variables (Q, P) so cleverly chosen that the new Hamiltonian K is identically zero. If K = 0, then by Hamilton's equations Q̇ = ∂K/∂P = 0 and Ṗ = −∂K/∂Q = 0 — every new coordinate and momentum is constant in time. The generating function of that transformation, of "type 2" denoted S(q, P, t), satisfies p = ∂S/∂q and K = H + ∂S/∂t. Demanding K = 0 gives precisely:

H(q, ∂S/∂q, t) + ∂S/∂t = 0

The payoff is a change in the kind of work required. Hamilton's equations are 2n coupled nonlinear ODEs you must integrate step by step. A complete solution of the HJ PDE — one carrying n independent constants α₁…αₙ — lets you find the orbit purely by algebra: set βᵢ = ∂S/∂αᵢ (n new constants) and solve those equations for q(t). No time integration of the trajectory is needed. For an integrable system this is a spectacular shortcut. For a non-integrable system the complete solution doesn't exist globally, and that failure is itself diagnostic: it signals chaos, the breakdown of smooth invariant tori, and the limits of the semiclassical picture.

The cost is conceptual difficulty (you must guess a separation ansatz) traded for computational elegance. When the trade pays off — Kepler, the symmetric top, a charge in crossed fields, geodesics in Kerr spacetime — Hamilton-Jacobi is the most efficient formulation of mechanics ever written, and the only one that turns smoothly into quantum mechanics as you let ℏ stop being negligible.