Hamilton's Characteristic Function

Before Hamilton ever wrote down a Hamiltonian, he had already compressed the whole of geometrical optics into one function. His characteristic function assigned to any two points the optical path length of the ray connecting them; from its partial derivatives, every lens, mirror and caustic in an instrument could be read off without tracing a single ray. When he turned to mechanics in 1834, he brought the object with him — and in the mechanics of conservative systems it survives, essentially unchanged, as $W(q,\alpha)$ : Hamilton's characteristic function, the time-free half of Hamilton's principal function. The split is disarmingly simple. When the Hamiltonian carries no explicit time, energy is conserved and the principal function separates as $S = W(q,\alpha) - Et$ : a static landscape $W$ over configuration space, plus a uniform drift. All the geometry of the motion at energy $E$ — where the system can go, with what momentum, along which paths — lives in $W$ . Its level surfaces are wavefronts; the momentum $\mathbf{p} = \nabla W$ pierces them at right angles; the trajectories are the rays. This is the optical-mechanical analogy in its sharpest classical form, and a century later it is precisely the picture Schrödinger promoted into wave mechanics. But $W$ is more than a geometric shadow of $S$ . It is a generating function in its own right, producing a canonical transformation with a different — and for bound systems, more useful — target: instead of freezing everything, it freezes the momenta and lets a single clock-like coordinate run. This article develops $W$ as an object (the abbreviated action), as a generator (the $K = E$ transformation), and as the bridge to Maupertuis' principle and action-angle theory.

Key takeaways

For a conservative system ( $\partial H/\partial t = 0$ ) the principal function splits as $S(q,\alpha,t) = W(q,\alpha) - Et$ ; the time-free piece $W$ is Hamilton's characteristic function.
$W$ obeys the Time-Independent Hamilton-Jacobi Equation $H(q,\partial W/\partial q) = E$ and encodes all the geometry of motion at fixed energy.
Used as a type-2 generator, $W$ produces a canonical transformation with new Hamiltonian $K = E = \alpha_1$ : all new momenta are constants, all new coordinates are frozen except $\beta_1 = \partial W/\partial\alpha_1$ , which advances as $t + \mathrm{const}$ — time itself emerges as the coordinate conjugate to energy.
Along the orbit $W = \int \mathbf{p}\cdot\mathrm{d}\mathbf{q}$ — the abbreviated action of Maupertuis; over one cycle of a bound coordinate it accumulates the action variable $J = \oint p\,\mathrm{d}q$ .
Geometrically, surfaces of constant $W$ are wavefronts with $\mathbf{p} = \nabla W$ normal to them; trajectories are the orthogonal rays, and $|\nabla W|^2 = 2m(E - V)$ is the classical eikonal equation.
Key idea: for a conservative system, factor out time via $S = W - Et$ ; the remaining function $W(q,\alpha)$ obeys $H(q,\partial W/\partial q) = E$ and generates a canonical map whose new Hamiltonian is $E$ itself — so all but one new coordinate are frozen and the motion is solved by quadrature.

Where this lives today

The eikonal equation (3) never left the workbench — it just changed employers. Seismologists solve $|\nabla W| = s(\mathbf{r})$ (slowness playing the role of $\sqrt{2m(E-V)}$ ) across billions of grid cells to compute first-arrival traveltimes through the Earth; the fast-marching and fast-sweeping algorithms that do it are, quite literally, numerical constructions of Hamilton's wavefronts, and the same solvers route autonomous robots around obstacles by treating "time-to-reach" as a characteristic function over the map. In 2026 the fashionable accelerant is the physics-informed neural eikonal solver — networks trained to satisfy $|\nabla W|^2 = 2m(E-V)$ pointwise — which has cut seismic traveltime tomography and medical ultrasound aberration correction from cluster-hours to laptop-minutes. The conceptual descendants run deeper still. Every WKB calculation in quantum mechanics opens with $\psi \sim e^{iW/\hbar}$ — the characteristic function as the phase of a wavefunction — and the semiclassical machinery of chemistry (instanton rates, tunneling splittings) is applied $W$ -craft. Even Feynman's path integral honors the split (1): for a conservative system, the propagator's phase at stationary points is exactly $W - Et$ . The static landscape Hamilton drew for light in 1828 turned out to be the phase of matter.

The physical picture

Drop a pebble's worth of thought onto a landscape: over every point $q$ of configuration space, at a chosen energy $E$ , there sits a value of $W$ . The landscape is static — time has been peeled off and set aside. Yet it contains the complete kinematics of every orbit at that energy, stored geometrically: the contours of $W$ are surfaces the motion always crosses at right angles, and the steepness of $W$ at any point is the momentum the particle must have there, $|\nabla W| = \sqrt{2m(E-V)}$ . Where the potential is deep the contours crowd together (fast particle, steep $W$ ); near a turning point they spread apart and the gradient dies (slow particle, flat $W$ ).

Figure

The static landscape of W — wavefronts of constant W with trajectories as their orthogonal rays

This is optics wearing mechanics' clothes. In the eikonal limit of wave optics, light rays are the orthogonal trajectories of the wavefronts of constant phase; the phase obeys an equation formally identical to $|\nabla W|^2 = 2m(E-V)$ , with refractive index playing the role of $\sqrt{E - V}$ . Hamilton discovered the mechanics version because he owned the optics version. And the analogy carries a famous sting, worth meeting early: if you restore time via $S = W - Et$ , the surfaces of constant $S$ are moving wavefronts sweeping across the static $W$ -landscape — and they move slowest where the particle moves fastest (their speed is $E/|\nabla W| = E/p$ ). Nothing is wrong: the front speed is a phase velocity, not the particle's velocity. Resolving that tension honestly is what led Schrödinger from this exact picture to wave mechanics. The second face of $W$ is dynamical rather than geometric. As a generating function, $W$ performs a softer freeze than the principal function. In Generating Function to a Trivial Hamiltonian, the time-dependent generator $S$ drives the new Hamiltonian to zero and stops everything. The time-independent $W$ cannot do that — with $\partial W/\partial t = 0$ the transformation law gives $K = H = E$ , not zero. But that residue is itself a triumph of economy: a Hamiltonian equal to one of the new momenta is the next best thing to zero. All momenta stay constant, all coordinates but one stay frozen, and the single survivor — the coordinate conjugate to the energy — advances at exactly unit rate. The system's own clock has been promoted to a canonical coordinate.

Figure

Two freezes compared — the principal function stops everything, the characteristic function leaves one clock running

Third face: $W$ is a quantity you can accumulate along the orbit. Its rate of growth is $\sum_i p_i\dot q_i$ , so $W = \int \mathbf{p}\cdot\mathrm{d}\mathbf{q}$ — the abbreviated action of Maupertuis, the eighteenth-century action that predates Hamilton's. For a bound oscillation, one full cycle adds to $W$ the enclosed phase-space area $\oint p\,\mathrm{d}q$ — the action variable $J$ that Action Integrals over Invariant Tori builds into the coordinates of choice for integrable systems, and that the old quantum theory set equal to $nh$ . Every payoff of the Hamilton-Jacobi chapter routes through this one function.

The two functions and the split

Start from the relation that defines $W$ . When $\partial H/\partial t = 0$ , insert the ansatz $S = W(q,\alpha) - Et$ into the Hamilton-Jacobi Equation $H(q,\partial S/\partial q,t) + \partial S/\partial t = 0$ . Since $\partial S/\partial q_i = \partial W/\partial q_i$ and $\partial S/\partial t = -E$ , the time drops out wholesale:

S(q,\alpha,t) = W(q,\alpha) - Et, \qquad H\!\left(q,\frac{\partial W}{\partial q}\right) = E \qquad (1)

where: $W(q,\alpha)$ — Hamilton's characteristic function, carrying $n$ constants $\alpha_i$ of which the energy is one (conventionally $\alpha_1 = E$ , though any invertible relabeling works). The right-hand member is the Time-Independent Hamilton-Jacobi Equation, a PDE over configuration space alone — its structure, separation and quadratures are that article's business. Here the point is the decomposition: for conservative systems, the principal function is a static landscape minus a uniform energy drift, so all the problem-solving effort concentrates on the time-free $W$ . The momenta come from the landscape's gradient, exactly as they came from $S$ (the $-Et$ term carries no $q$ -dependence):

p_i = \frac{\partial W}{\partial q_i} \qquad (2)

where: for a single particle in a potential this reads $\mathbf{p} = \nabla W$ , and squaring it against $H = \mathbf{p}^2/2m + V = E$ gives the classical eikonal equation

|\nabla W|^2 = 2m\big(E - V(\mathbf{r})\big) \qquad (3)

where: the local steepness of $W$ is fixed pointwise by energy conservation — steep where the particle is fast, flat at turning points where $E = V$ . Equation (3) is a first-order PDE whose solutions are built from the front-and-ray geometry of the hero figure; it is formally identical to the eikonal equation of geometrical optics with effective refractive index $n \propto \sqrt{E-V}$ , which is the precise content of the optical-mechanical analogy. Along an actual orbit, $W$ accumulates the abbreviated action. Differentiating $W(q(t),\alpha)$ along the motion (the $\alpha_i$ are constants) and using (2):

\frac{\mathrm{d}W}{\mathrm{d}t} = \sum_i \frac{\partial W}{\partial q_i}\,\dot q_i = \sum_i p_i \dot q_i \quad\Longrightarrow\quad W = \int \mathbf{p}\cdot\mathrm{d}\mathbf{q} \qquad (4)

where: the integral runs along the trajectory at energy $E$ . Compare the principal function's $\mathrm{d}S/\mathrm{d}t = L$ : the two rates differ by exactly $\mathrm{d}W/\mathrm{d}t - \mathrm{d}S/\mathrm{d}t = \sum p\dot q - L = H = E$ , which is the split (1) read as rates. $S$ accumulates the Lagrangian action; $W$ accumulates the Maupertuis action; they disagree by the steady metronome $Et$ .

Which function answers which question

you want	use	notes
a conservative system solved with least machinery	$W$ : eq. (1) right member, then eq. (6)	the default for every textbook problem
a driven / time-dependent Hamiltonian	$S$ only	no conserved $E$ to split off; see Generating Function to a Trivial Hamiltonian
the orbit's geometry (path, not timetable)	$\beta_j = \partial W/\partial\alpha_j$ , $j\geq 2$	time-free relations among the $q_i$ — the orbit equation directly
the timetable along the path	$\beta_1 = \partial W/\partial E = t + \mathrm{const}$	time-of-flight from an energy derivative; worked example 2
phase-space area / quantization / frequencies	$J = \oint p\,\mathrm{d}q$ from eq. (4)	gateway to Action-Angle Variables
wavefront/eikonal geometry	eq. (3)	$W$ = classical phase; WKB writes $\psi \sim e^{iW/\hbar}$

Derivation AThe canonical transformation generated by the characteristic function

We are deriving the dynamical content of $W$ : that used as a time-independent type-2 generator with the constants $\alpha_i$ as new momenta, it maps a conservative system to the Hamiltonian $K = \alpha_1 = E$ , so that every new variable is frozen except the single coordinate conjugate to the energy, which advances at unit rate. The strategy is the same three-move sequence used for the principal function — write the type-2 relations, transform the Hamiltonian, run Hamilton's equations in the new chart — with one crucial difference to watch: $\partial W/\partial t = 0$ , so no cancellation of $H$ is available and the transformed Hamiltonian survives as the energy itself. The payoff of that survival is that the surviving coordinate integrates to $t + \mathrm{const}$ , handing over the system's timetable by differentiation.

Step 1. Write the type-2 relations with generator $F_2 = W(q,\alpha)$ , taking the new momenta to be the constants: $P_i = \alpha_i$ with $\alpha_1 = E$ . Then

p_i = \frac{\partial W}{\partial q_i}, \qquad \beta_i = \frac{\partial W}{\partial \alpha_i}

Why this interface: the $\alpha_i$ enter $W$ as parameters of the complete integral of the time-independent HJ equation; making them the new momenta is what turns the parameter family into a canonical chart.

Step 2. Transform the Hamiltonian. The type-2 law is $K = H + \partial F_2/\partial t$ , and $W$ carries no explicit time:

K = H\!\left(q,\frac{\partial W}{\partial q}\right) = E = \alpha_1

Why the last equality is the whole trick: the time-independent HJ equation forces the combination $H(q,\partial W/\partial q)$ to equal the constant $E$ — not as an on-shell accident but identically in $q$ . Expressed in the new variables, $K$ is therefore the pure momentum function $K(\alpha) = \alpha_1$ : every coordinate is cyclic.

Step 3. Run Hamilton's equations in the new chart:

\dot\alpha_i = -\frac{\partial K}{\partial\beta_i} = 0, \qquad \dot\beta_i = \frac{\partial K}{\partial\alpha_i} = \delta_{i1}

Why the Kronecker delta: $K = \alpha_1$ depends on nothing but $\alpha_1$ , so the only nonvanishing rate in the entire chart is $\dot\beta_1 = 1$ .

Step 4. Integrate and invert:

\beta_1 = t + \beta_1^0 = \frac{\partial W}{\partial E}, \qquad \beta_j = \frac{\partial W}{\partial \alpha_j} = \mathrm{const}\ (j \geq 2)

The $n-1$ frozen relations $\beta_j = \partial W/\partial\alpha_j$ contain no time: they are the orbit equations, fixing the path's geometry. The single running relation $\partial W/\partial E = t + \beta_1^0$ is the timetable, locating the system on that path at each instant. Solving the set for $q_i(t)$ completes the solution. As a check: setting $n = 1$ reduces the entire scheme to the elementary quadrature of 1D energy conservation, as it must.

The structural moral of Step 4 deserves its own sentence: the characteristic function separates geometry from schedule. The frozen $\beta_j$ answer "what curve does the system trace?" — a question that never mentions time — while $\beta_1$ answers "when is it where?". Central-force theory exploits this daily: the orbit equation $\phi(r)$ comes from $\partial W/\partial\ell$ , the time along the orbit from $\partial W/\partial E$ , and the two calculations never interfere. (The systematic exploitation, coordinate system by coordinate system, is Separation of Variables in Hamilton-Jacobi Theory.)

Go deeper · Derivation B · 1 minMaupertuis' principle falls out of the characteristic function

Derivation BMaupertuis' principle falls out of the characteristic function

The goal is the principle of least abbreviated action: among all paths joining two configuration-space points at the same energy $E$ , the true path makes $W = \int\mathbf{p}\cdot\mathrm{d}\mathbf{q}$ stationary. The strategy is to relate $\delta W$ to the variation of the full action $S$ , whose stationarity is Hamilton's principle, and track carefully what is held fixed — this is the classic place where the two action principles get conflated, so the bookkeeping of endpoint-time variations is the step to watch. Hamilton's principle fixes the traversal time and lets energy vary between comparison paths; Maupertuis fixes the energy and lets the traversal time vary. The characteristic function is exactly the object built to absorb that exchange.

Step 1. Relate the two actions along any comparison path with conserved $H = E$ : from the split (1) read as accumulations,

W = S + Et_{12}

where $t_{12}$ is the traversal time of the path. Why: $\mathrm{d}W/\mathrm{d}t - \mathrm{d}S/\mathrm{d}t = H = E$ pointwise along the path, integrated over the traversal.

Step 2. Vary within the fixed-energy family: $\delta W = \delta S + E\,\delta t_{12}$ , since $E$ is common to all comparison paths. Why $\delta t_{12} \neq 0$ : paths of equal energy but different shape take different times — the variation cannot hold both $E$ and $t_{12}$ fixed, and does not try to.

Step 3. Evaluate $\delta S$ for variations with fixed spatial endpoints but movable end time. The general endpoint variation of the action is $\delta S = [\,\mathbf{p}\cdot\delta\mathbf{q} - H\,\delta t\,]_1^2 + (\text{bulk term})$ ; the bulk term vanishes on the true path by Hamilton's principle, and the spatial endpoint displacements are zero, leaving

\delta S = -E\,\delta t_{12}

Why the boundary term survives only in time: the comparison paths all start and end at the same configuration points, but arrive at different moments — the movable-end-time term $-H\delta t$ is the entire variation.

Step 4. Combine Steps 2 and 3:

\delta W = -E\,\delta t_{12} + E\,\delta t_{12} = 0

Among all paths of energy $E$ between fixed endpoints, the true trajectory makes the abbreviated action stationary — Maupertuis' principle, in Jacobi's clean formulation. As a check: for a free particle the result reduces to $\delta\int p\,\mathrm{d}s = p\,\delta(\mathrm{length}) = 0$ — straight lines, as it must.

Limiting cases

Case	Result	Meaning
free particle (1D)	$W = \sqrt{2mE}\,x$	constant-gradient landscape: uniform momentum, evenly spaced fronts; $\partial W/\partial E = t$ gives $x = \sqrt{2E/m}\,t$
harmonic oscillator	$W = \int\!\sqrt{2m(E - \tfrac12 m\omega^2x^2)}\,\mathrm{d}x$	fronts crowd at the center, flatten at turning points; one cycle accumulates $\oint p\,\mathrm{d}x = 2\pi E/\omega$
turning point ( $E = V$ )	$\nabla W = 0$	the landscape goes flat; $W$ 's two branches (outgoing/incoming momentum) meet — the seed of WKB connection formulas
separable system	$W = \sum_i W_i(q_i;\alpha)$	the landscape factorizes into one-dimensional profiles, each a quadrature — see Separation of Variables in Hamilton-Jacobi Theory
explicitly time-dependent $H$	no $W$ exists	nothing conserved to split off; retreat to the full principal function $S$

Common trapThe characteristic function does not make the new Hamiltonian vanish

It is tempting to file $W$ as "the principal function for conservative systems" and assume it, too, drives $K$ to zero. It does not, and the difference is not pedantry — it changes the equations of motion in the new chart. With $S$ the new Hamiltonian vanishes and all $2n$ new variables are frozen. With $W$ the new Hamiltonian is $K = \alpha_1 = E$ : the momenta and $n-1$ coordinates are frozen, but the coordinate conjugate to the energy runs as $\beta_1 = t + \mathrm{const}$ . Forgetting the running coordinate produces "solutions" with no time dependence at all; forgetting which coordinate runs scrambles the timetable. A second trap: the two functions differ by $Et$ , so mixing their derivative relations — e.g., using $\partial S/\partial E$ where $\partial W/\partial E$ is meant — silently shifts every time by $t$ itself. Keep the ledger straight: $S$ freezes everything and satisfies $\partial S/\partial t = -H$ ; $W$ freezes all but the clock and satisfies $\partial W/\partial t = 0$ .

Worked examples

1 · The harmonic oscillator's landscape, clock and area (confidence builder)

A mass $m = 2\ \mathrm{kg}$ on a spring with $\omega = 5\ \mathrm{rad/s}$ carries energy $E = 100\ \mathrm{J}$ . Using only the characteristic function: find the amplitude, the trajectory, the momentum at the center, and the abbreviated action accumulated over one full cycle.

Step 1. The landscape. From the eikonal relation (3) in one dimension, $p(x) = \partial W/\partial x = \sqrt{2m(E - \tfrac12 m\omega^2 x^2)}$ , so $W(x,E) = \int_0^x \sqrt{2m(E - \tfrac12 m\omega^2 x'^2)}\,\mathrm{d}x'$ . The turning points sit where the gradient dies: $x_\pm = \pm\sqrt{2E/m\omega^2} = \pm\sqrt{200/50} = \boxed{\pm 2\ \mathrm{m}}$ .

Step 2. The clock. The timetable relation is $t + \beta^0 = \dfrac{\partial W}{\partial E} = \displaystyle\int_0^x \frac{m\,\mathrm{d}x'}{\sqrt{2m(E - \tfrac12 m\omega^2 x'^2)}} = \frac{1}{\omega}\arcsin\!\left(x\sqrt{\frac{m\omega^2}{2E}}\right)$ .

Step 3. Invert for the trajectory (choose $\beta^0 = 0$ ): $x(t) = \sqrt{\dfrac{2E}{m\omega^2}}\,\sin\omega t = \boxed{2\sin(5t)\ \mathrm{m}}$ — amplitude and phase delivered by differentiation and inversion, no equation of motion ever integrated.

Step 4. Momentum at the center: $p(0) = \sqrt{2mE} = \sqrt{2\times 2\times 100} = \boxed{20\ \mathrm{kg\,m/s}}$ , i.e. $v = 10\ \mathrm{m/s}$ — which checks against $v_{\max} = \omega A = 5\times 2 = 10\ \mathrm{m/s}$ ✓.

Step 5. The area. Over one cycle the abbreviated action accumulates the phase-space ellipse's area, semi-axes $\sqrt{2mE} = 20$ and $\sqrt{2E/m\omega^2} = 2$ : $\oint p\,\mathrm{d}x = \pi\times 20\times 2 = 125.7\ \mathrm{J\,s}$ . Cross-check with the closed form $2\pi E/\omega = 2\pi\times 100/5 = \boxed{125.7\ \mathrm{J\,s}}$ ✓ — the action variable $J$ that Action-Angle Variables will make a coordinate.

2 · Time-of-flight from a derivative, and the eikonal check (exam level)

Same oscillator ( $m = 2\ \mathrm{kg}$ , $\omega = 5\ \mathrm{rad/s}$ , $E = 100\ \mathrm{J}$ ). Without solving for the trajectory, find (a) the time to travel from the center $x = 0$ to $x = 1\ \mathrm{m}$ , and (b) the momentum at $x = 1\ \mathrm{m}$ , checking it against the eikonal equation. (c) Show the same time-of-flight trick on a free particle and interpret.

Step 1. (a) The timetable relation of the $W$ -transformation says: time-of-flight between two configurations is the difference of an energy derivative — $t_{0\to 1} = \dfrac{\partial W}{\partial E}\Big|_{x=1} - \dfrac{\partial W}{\partial E}\Big|_{x=0}$ .

Step 2. Evaluate with the Step 2 integral of example 1: $t_{0\to 1} = \dfrac{1}{\omega}\arcsin\!\left(x\sqrt{\dfrac{m\omega^2}{2E}}\right)\Big|_{x=1} = \dfrac{1}{5}\arcsin(0.5) = \dfrac{\pi/6}{5} = \boxed{0.105\ \mathrm{s}}$ . Sanity check from the known motion: $1 = 2\sin 5t$ gives $5t = \pi/6$ , $t = 0.105\ \mathrm{s}$ ✓ — but note the derivative route never used $x(t)$ .

Step 3. (b) Momentum from the landscape's steepness at $x = 1$ : $p = \sqrt{2m(E - \tfrac12 m\omega^2 x^2)} = \sqrt{4\,(100 - 25)} = \sqrt{300} = \boxed{17.3\ \mathrm{kg\,m/s}}$ .

Step 4. Check via the trajectory: $v(t) = 10\cos 5t$ at $5t = \pi/6$ gives $v = 10\times(\sqrt{3}/2) = 8.66\ \mathrm{m/s}$ , so $p = 2\times 8.66 = 17.3\ \mathrm{kg\,m/s}$ ✓. The eikonal equation (3) reads $|\nabla W|^2 = 2m(E-V)$ : $300 = 4\times(100 - 25)$ ✓ exactly.

Step 5. (c) Free particle: $W = \sqrt{2mE}\,x$ , so $t = \partial W/\partial E = \sqrt{m/2E}\;x = x/v$ . The energy-derivative of the abbreviated action is literally distance over speed — the trick is general: differentiating the static landscape with respect to energy yields the schedule, because $E$ and $t$ are a conjugate pair. That conjugacy — energy paired with time exactly as momentum pairs with position — is the classical shadow of a relation that returns, transformed, in quantum mechanics.

Origins · 1 minSir William Rowan Hamilton and how this was discovered

Sir William Rowan Hamilton

1805–1865 · Trinity College Dublin and Dunsink Observatory

Hamilton built the characteristic function twice. The first time was in optics: his Theory of Systems of Rays (begun at age 21) showed that an optical instrument of any complexity is completely summarized by one function — the optical path length between endpoints — from whose derivatives all ray directions follow. He prized above all the method's generality: aberrations, caustics and focal properties tumbled out of partial differentiation. The 1832 sensation of conical refraction — a genuinely new optical phenomenon predicted from the characteristic function before being observed — made him famous at 27. The second construction, in the 1834-35 dynamics papers, transplanted the function to mechanics, where the role of optical path length passed to the action. Hamilton's own formulation was clumsy — his function obeyed two PDEs simultaneously and seemed to require knowing the motion in advance — and it was Jacobi who pruned it to the single equation used today. The deep dividend arrived in 1926, when Schrödinger, reading the optical-mechanical analogy in reverse, asked what wave theory has mechanics as its eikonal limit — and wrote the equation of wave mechanics.

History

1744

— Maupertuis announces the principle of least action for

\int p\,\mathrm{d}s

; Euler makes it precise for a single particle — the abbreviated action is born a century before

W

1828

— Hamilton's Theory of Systems of Rays compresses geometrical optics into one characteristic function; supplements through 1832 predict conical refraction, confirmed within months.

1834

— Hamilton's dynamics papers carry the characteristic function into mechanics, tying it to the accumulated action at fixed energy.

1837

— Jacobi trims Hamilton's two-equation scheme to a single PDE and shows one complete integral suffices — the working form of the theory.

1926

— Schrödinger inverts the optical-mechanical analogy: seeking the wave theory whose eikonal is

|\nabla W|^2 = 2m(E-V)

, he arrives at the wave equation of quantum mechanics.

Going further

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §10.3 · Landau & Lifshitz, Mechanics, 3rd ed., §47 · Morin, Introduction to Classical Mechanics, Ch. 15 (Hamilton-Jacobi) · Taylor, Classical Mechanics, Ch. 13 · Arnold, Mathematical Methods of Classical Mechanics, §46-47 Key results to keep in mind: substituting $S=W-Et$ into the HJ equation yields $H=E$ ; Hamilton's equations in the new chart give $K=\alpha_1$ and $\dot\beta_1=1$ ; the SHO action integral evaluates as $\oint p\,\mathrm{d}x=2\pi E/\omega$ (the phase-space ellipse area), with frequency $\nu=\omega/2\pi$ . The worked-example numbers throughout — amplitude $2\ \mathrm{m}$ , $p(0)=20\ \mathrm{kg\,m/s}$ , cycle action $125.7\ \mathrm{J\,s}$ , $t_{0\to 1}=0.105\ \mathrm{s}$ , $p(1\ \mathrm{m})=17.3\ \mathrm{kg\,m/s}$ — all follow directly from the boxed formulas.

Builds on: Hamilton-Jacobi Equation · Hamilton's Principal Function · Time-Independent Hamilton-Jacobi Equation · Generating Function to a Trivial Hamiltonian Leads to: Separation of Variables in Hamilton-Jacobi Theory · Action-Angle Variables · Action Integrals over Invariant Tori · Liouville Integrable Systems

Why would I ever accept K = E when the principal function gives me K = 0? Isn't zero strictly better?

Zero is a stronger freeze but costs a harder object: $S$ depends on $n+1$ variables and must be re-derived with its time dependence intact. For a conservative system that time dependence is pure boilerplate — always exactly $-Et$ — so carrying it explicitly buys nothing. Working with $W$ solves a PDE in one fewer variable, and the "loss" (one coordinate runs) is actually the deliverable: $\beta_1 = \partial W/\partial E = t + \mathrm{const}$ is the timetable you wanted anyway. For bound motion the $W$ -route is also the only sensible gateway to action-angle variables, where the goal is emphatically not to freeze the angles but to make them advance uniformly. Choose $S$ only when the Hamiltonian's time dependence is genuine.

Is it really legitimate that time becomes a coordinate? t is the evolution parameter, not a dynamical variable.

It is legitimate and exact, though the phrasing deserves care. Nothing in the canonical formalism forbids a new coordinate from happening to increase at unit rate and hence numerically track $t$ ; that is all $\beta_1 = t + \beta_1^0$ says. The deeper observation is that energy and time form a conjugate pair in precisely the sense position and momentum do: $\beta_1$ is conjugate to $\alpha_1 = E$ , and its Hamilton equation $\dot\beta_1 = \partial K/\partial E = 1$ is what "conjugate to energy" means dynamically. This conjugacy is structural, not accidental — it reappears as $\partial S/\partial t = -E$ for the principal function, in the extended-phase-space formalism where $(t,-E)$ is adjoined as a genuine canonical pair, and (transformed) in quantum mechanics.

For a bound orbit the particle returns to the same q with the opposite momentum. Doesn't p = ∇W then contradict itself?

Sharp observation — a single-valued $W$ cannot assign both $+p$ and $-p$ to one point. The resolution is that $W$ is multivalued: it has (in 1D) two branches, one for the outgoing motion with $W' = +\sqrt{2m(E-V)}$ and one for the return with the minus sign, joined at the turning points where the gradient vanishes. Following the motion continuously means walking up one sheet and back along the other; each full cycle adds the enclosed area $\oint p\,\mathrm{d}q = J$ to $W$ , so the function climbs by one action quantum per period rather than returning to its old value. This branch structure is not a defect — it is exactly what WKB connection formulas negotiate at turning points, and why $\oint p\,\mathrm{d}q$ (not $W$ itself) is the natural single-valued object for bound motion.

How do the wavefronts move if W is time-independent?

The $W$ -fronts do not move — they are a static scaffolding over configuration space. What moves are the fronts of constant principal function $S = W - Et$ : the condition $S = \mathrm{const}$ at increasing $t$ requires $W$ to increase at rate $E$ , so the moving front slides across the static $W$ -contours, occupying the contour $W = W_0 + Et$ at time $t$ . Its local speed is $E/|\nabla W| = E/p$ — the famous inversion: fastest fronts where the particle is slowest. Both pictures are useful: the static $W$ -landscape carries the geometry (rays orthogonal to fronts), while the sliding $S$ -front carries the phase evolution that the WKB wavefunction $\psi \sim e^{i(W-Et)/\hbar}$ inherits verbatim.

What exactly is the difference between Maupertuis' principle and Hamilton's principle?

Both are stationarity statements about the true path, but they compare different families of trial paths. Hamilton's principle varies $S = \int L\,\mathrm{d}t$ over paths with fixed endpoints and fixed traversal time; energy is allowed to differ between trial paths. Maupertuis' principle varies $W = \int\mathbf{p}\cdot\mathrm{d}\mathbf{q}$ over paths with fixed endpoints and fixed energy; the traversal time is allowed to differ. The derivation in this article shows they exchange exactly one boundary term, $E\,\delta t_{12}$ . Practically: Hamilton's principle is more general (works for time-dependent systems) and delivers equations of motion; Maupertuis' is geometric — it determines the shape of the orbit without its schedule, and in the Jacobi-metric form it turns trajectories into geodesics of a curved configuration-space metric.

Why is W called the abbreviated action — what got abbreviated?

The integrand. The full (Hamilton) action accumulates $L\,\mathrm{d}t = (\mathbf{p}\cdot\dot{\mathbf{q}} - H)\,\mathrm{d}t$ ; the abbreviated action keeps only the first piece, $\mathbf{p}\cdot\mathrm{d}\mathbf{q}$ , dropping the $-H\,\mathrm{d}t = -E\,\mathrm{d}t$ term that is pure bookkeeping when energy is conserved. Hence $W = S + Et$ : the "abbreviation" is the removal of the energy drift. Historically the order is reversed — Maupertuis' $\int p\,\mathrm{d}s$ (1744) predates Hamilton's $\int L\,\mathrm{d}t$ (1834) by ninety years — so it is the full action that is the latecomer, generalizing the abbreviated one to time-dependent situations where no fixed $E$ exists to strip out.

Does the Bohr-Sommerfeld rule quantize W itself?

It quantizes $W$ 's growth per cycle. The old quantum theory demands $\oint p_i\,\mathrm{d}q_i = n_i h$ for each separable degree of freedom — and that loop integral is exactly the amount the (multivalued) characteristic function climbs in one period of $q_i$ . So the rule says: the classical phase $W/\hbar$ must advance by an integer multiple of $2\pi$ per cycle, which is precisely the condition for the WKB wavefunction $\psi \sim e^{iW/\hbar}$ to be single-valued on the orbit. For the oscillator of worked example 1, $J = 2\pi E/\omega$ gives $E_n = n\hbar\omega$ — off from the true spectrum only by the half-integer zero-point shift that the turning-point branch structure (Maslov correction) supplies. The refined statement and its uses live in Action-Angle Variables.

Can I use W when the motion is unbounded, like scattering?

Yes — $W = \int\mathbf{p}\cdot\mathrm{d}\mathbf{q}$ and the whole generating-function machinery need only a conserved energy, not bound orbits. For scattering problems $W$ is single-valued (no cycles, no branch accumulation) and its derivatives give deflection functions and time delays: $\partial W/\partial\ell$ yields the scattering angle relation, $\partial W/\partial E$ the traversal time including the potential's delay. What specifically requires bound motion is the action variable $J = \oint p\,\mathrm{d}q$ — a closed loop needs a closed orbit (or a librating coordinate). So: $W$ always exists for conservative systems; $J$ and everything built on it (angles, frequencies, Bohr-Sommerfeld) belong to the bound chapter of the story.