The Action Functional

Here is a strange idea that turned out to run the universe. Take a mechanical system — a thrown ball, a pendulum, a planet — and consider not its state at an instant but an entire candidate history: a whole path $q(t)$ from a fixed starting configuration at time $t_1$ to a fixed final configuration at time $t_2$ . Now grade that history with a single number: integrate the Lagrangian $L = T - V$ along it. The grade is called the action, $S[q] = \int_{t_1}^{t_2} L\,\mathrm{d}t$ , and the square brackets announce something genuinely new in your mathematical life: $S$ is a functional — a machine that eats a whole function and returns one real number. Different histories between the same two endpoints get different grades. Why grade histories at all? Because nature, it turns out, is a grader. Among all conceivable paths between the two endpoints, the one actually followed is the path whose grade is stationary — unchanged, to first order, by any small deformation. That statement, Hamilton's principle, and the machinery that converts it into differential equations, belong to the next two articles. This one is about the object itself: what a functional is, how to actually compute the action of a path (we will do it numerically, several times, and watch stationarity emerge from raw arithmetic), why $S$ carries the units of Planck's constant, and the two exact structural properties — endpoint gauge freedom and the endpoint derivatives that turn $S$ into Hamilton's principal function — that make the action the most consequential single number in physics, from optimal rocket trajectories to the phase of every quantum amplitude.

Key takeaways

The action $S[q] = \int_{t_1}^{t_2} L(q,\dot q,t)\,\mathrm{d}t$ is a functional: it assigns one real number to every entire path $q(t)$ joining fixed endpoints at fixed times.
Its dimensions are energy × time ( $\mathrm{J\,s}$ ) — the dimensions of Planck's constant $\hbar$ and of angular momentum; $S/\hbar$ is the dimensionless phase of quantum mechanics, and $S \gg \hbar$ is what "classical" means.
Computing $S$ for explicit competing paths is honest arithmetic: for a free particle the straight path scores $S = \tfrac12 mv^2 T$ and every detour scores more, with the difference quadratic (never linear) in the deformation — stationarity visible in numbers.
Adding a total time derivative $\mathrm{d}F(q,t)/\mathrm{d}t$ to $L$ shifts $S$ only by the fixed endpoint values $F_2 - F_1$ : the gauge freedom of the Lagrangian, exact at the level of the action.
Evaluated along the true path and viewed as a function of its endpoint, the action becomes Hamilton's principal function with $\partial S/\partial q = p$ and $\partial S/\partial t = -h$ — the seed of Hamilton–Jacobi theory.
Key idea: the action $S[q]=\int_{t_1}^{t_2}L\,\mathrm{d}t$ converts a dynamical problem into a variational one: the physical path is the stationary point of a single scalar functional, and demanding $\delta S=0$ generates the equations of motion in any coordinate system.

Where this lives today

The most computationally expensive numbers humanity produces are values of an action. Lattice QCD — the working theory of quarks and gluons — computes particle masses by importance-sampling millions of field configurations with probability weight $e^{-S}$ , where $S$ is the (Euclidean) action functional of the gluon field on a spacetime grid; the 2026-era exascale campaigns at Frontier and Jupiter spend hundreds of millions of GPU-hours essentially evaluating equation (1)'s field-theory descendant, over and over, for configurations nobody will ever "see." The functional this article defined for one puck is, scaled up, the central object of computational particle physics. Closer to the ground, every rocket that lands itself is solving a fixed-endpoint path-grading problem in real time. Guidance algorithms of the direct-collocation family discretize a candidate trajectory into a few hundred knots, attach a scalar cost functional — fuel burned, a close cousin of the action, sometimes literally augmented with $\int L\,\mathrm{d}t$ terms — and drive its discrete gradient to zero: $\delta S = 0$ as a convex optimization, run onboard in milliseconds during descent. The mathematical structure is byte-for-byte the one on this page: fixed endpoints, a functional over whole histories, and the winning history found by demanding first-order insensitivity to deformation.

The physical picture

An ordinary function is a machine with a number slot: feed in $x = 3$ , get back $f(3)$ . A functional is a machine with a curve slot: feed in an entire path — all of $q(t)$ from $t_1$ to $t_2$ , infinitely many values at once — and get back one number. The action is the canonical example. Its recipe is concrete: ride along the candidate path, and at each instant read off the kinetic energy and the potential energy; accumulate $T - V$ , second by second; the total at arrival is $S$ . A path that lingers where the potential is low earns a low score; a path that rushes (large $T$ ) pays for it; the action is the running ledger of that competition, settled once per whole history.

Figure

A functional eats a whole path and returns one number — three histories, three scores

Notice what is fixed and what is free. Both endpoints are nailed down — where the system starts and ends, and when. Everything in between is up for grabs: the functional happily grades wild histories that no force could ever produce. That is the point. The action is defined on the space of all kinematically conceivable paths precisely so that the true dynamical path can be characterized by how its score compares with its neighbors'. It is the same move an economist makes by defining profit over all conceivable strategies in order to characterize the optimal one — except that mechanics' optimizer is exactly solvable. The numbers on the scoreboard above are real (they are computed in Derivation A and Worked Example 1): for a free particle crossing 2 m in 2 s, the straight path scores $1.000\ \mathrm{J\,s}$ , a gentle sinusoidal detour scores $1.077$ , a bigger one $1.308$ . Two features of those numbers preview the whole subject. The true path scores lowest here — and more importantly, deforming away from it by an amount $\varepsilon$ raises the score by something proportional to $\varepsilon^2$ , never $\varepsilon$ . A ledger whose first-order sensitivity to deformation vanishes is the fingerprint of a stationary point. Nature's paths sit at exactly such points; making that principle precise is the next article's job. One more feature hides in the units. Energy times time is the dimension of $\hbar = 1.055\times10^{-34}\ \mathrm{J\,s}$ , and the comparison $S/\hbar$ is the deepest dimensionless number in physics: it is the phase, in radians, that quantum mechanics attaches to a history. Your 1-J·s free particle carries a phase of $\sim 10^{34}$ radians; an electron traversing an atom carries a phase of order one. The first is why baseballs follow a single crisp path, the second is why electrons do not — and both statements are readings of the same functional this article defines.

From paths to numbers

Everything begins with the definition, so let it carry its full weight. For a system with generalized coordinates $q = (q_1,\dots,q_n)$ and Lagrangian $L(q,\dot q,t)$ , the action of a path $q(t)$ joining fixed endpoints $q(t_1) = q^{(1)}$ and $q(t_2) = q^{(2)}$ is

S[q] = \int_{t_1}^{t_2} L\big(q(t),\dot q(t),t\big)\,\mathrm{d}t \qquad (1)

where: $S[q]$ — a single real number, the functional's output for the whole input path $q(t)$ ; $L$ — typically $T - V$ , evaluated instant by instant along the candidate path, whether or not that path obeys any equation of motion; $t_1, t_2$ — fixed initial and final times. The square brackets are load-bearing notation: $S[q]$ depends on the entire function $q(\cdot)$ , not on any single value of it. Dimensions: energy × time, $\mathrm{J\,s}$ — the same as $\hbar$ and as angular momentum, a coincidence with consequences we return to below. Computing (1) for an explicit path is nothing more than freshman calculus, and doing it a few times is the fastest way to make the functional concept concrete. The recipe: substitute the candidate $q(t)$ and its derivative $\dot q(t)$ into $L$ , obtaining an ordinary function of $t$ ; integrate. To compare many candidates systematically, embed them in a one-parameter family

q_\varepsilon(t) = q(t) + \varepsilon\,\eta(t), \qquad \eta(t_1) = \eta(t_2) = 0 \qquad (2)

where: $\varepsilon$ — a dial that interpolates between the reference path ( $\varepsilon = 0$ ) and its deformations; $\eta(t)$ — the deformation profile, forced to vanish at the endpoints so every member of the family respects the boundary conditions. The action of the family, $S(\varepsilon)$ , is then an ordinary function of one variable — a profile through path-space along one direction — and its shape near $\varepsilon = 0$ diagnoses the reference path: a nonzero slope $S'(0)$ means the reference can be beaten to first order; $S'(0) = 0$ in every direction $\eta$ is what "stationary" means. Derivation A computes such a profile exactly.

Figure

The action along one deformation direction — an ordinary parabola whose flat bottom is the true path

Which statement about S answers which question

you want	use	notes
the score of one explicit path	eq. (1), direct integration	works for any path, physical or not
to compare a family of paths	$S(\varepsilon)$ via eq. (2)	stationarity ⇔ no linear term in $\varepsilon$
how $S$ responds to a Lagrangian gauge shift	eq. (3)	endpoint terms only — same physics
the momentum/energy content of the true path's score	eqs. (4), (5)	Hamilton's principal function
the quantum meaning of the score	phase $= S/\hbar$	classical limit when $S \gg \hbar$

The first exact structural property of the action concerns the non-uniqueness of the Lagrangian. Two Lagrangians differing by a total time derivative describe identical physics, and at the level of the action the mechanism is one line of calculus:

L \to L + \frac{\mathrm{d}F(q,t)}{\mathrm{d}t} \quad\Longrightarrow\quad S \to S + F\big(q^{(2)},t_2\big) - F\big(q^{(1)},t_1\big) \qquad (3)

where: $F(q,t)$ — any smooth function of coordinates and time. The added integral collapses, by the fundamental theorem of calculus, to endpoint values — and the endpoints are fixed, the same for every candidate path. So the entire scoreboard shifts by one common constant: every path's score changes, but every difference of scores, and hence the identity of the stationary path, is untouched. This is the cleanest way to see that the pointwise value of $S$ (like that of $L$ ) is unphysical; only comparisons carry content. In quantum mechanics the same shift multiplies every amplitude by a common phase $e^{i(F_2 - F_1)/\hbar}$ — equally unobservable. The second structural property looks past the competition to the winner. Let the equations of motion be solved, and evaluate $S$ along the true path as a function of where and when that path ends: $S(q, t)$ , with the final endpoint now a variable. This object — the action as a function of its endpoint, called Hamilton's principal function — has derivatives that turn out to be old friends:

\frac{\partial S}{\partial q_k} = p_k \qquad (4)

where: $p_k = \partial L/\partial \dot q_k$ — the generalized momentum with which the true path arrives at the endpoint. Move the target a little and the minimal ledger changes by momentum times displacement.

\frac{\partial S}{\partial t} = -\,h \qquad (5)

where: $h = \sum_k p_k\dot q_k - L$ — the energy function (Jacobi integral) of the arriving path. Grant the journey a little more time and the ledger changes by minus the energy. Derivation C proves both from the boundary term that Hamilton's principle normally throws away, and together they seed the Hamilton–Jacobi reformulation of mechanics, where (4)–(5) are promoted into a partial differential equation for $S(q,t)$ itself. At this node, take them as the demonstration that the action is not a bookkeeping curiosity: evaluated on the dynamics, it contains the momenta and the energy as its gradients.

Derivation AEvaluating the Action on a Family of Free Particle Paths

The goal is to see stationarity arise from explicit integration, with no variational machinery at all. We take the simplest possible system — a free particle of mass $m$ , for which $L = \tfrac12 m\dot q^2$ — send it from $q = 0$ at $t = 0$ to $q = d$ at $t = T$ , and compute the action of the entire sinusoidal-detour family $q_\varepsilon(t) = vt + \varepsilon\sin(\pi t/T)$ , where $v = d/T$ is the average speed every candidate must share. The strategy is direct: substitute, expand the square, integrate term by term. The thing to watch is the cross term, linear in $\varepsilon$ — its integral will vanish identically, and that vanishing is exactly the stationarity of the straight path, caught in the act.

Step 1. Differentiate the family: $\dot q_\varepsilon(t) = v + \varepsilon\,\dfrac{\pi}{T}\cos\!\big(\tfrac{\pi t}{T}\big)$ . Every member starts at $0$ and ends at $d$ , since $\sin(0) = \sin(\pi) = 0$ — the deformation respects the fixed endpoints, as equation (2) demands.

Step 2. Substitute into $S(\varepsilon) = \int_0^T \tfrac12 m\,\dot q_\varepsilon^2\,\mathrm{d}t$ and expand the square:

S(\varepsilon) = \frac{m}{2}\int_0^T \left[ v^2 + 2v\varepsilon\,\frac{\pi}{T}\cos\!\Big(\frac{\pi t}{T}\Big) + \varepsilon^2\frac{\pi^2}{T^2}\cos^2\!\Big(\frac{\pi t}{T}\Big) \right] \mathrm{d}t

Step 3. Integrate each term. The first gives $\tfrac12 mv^2 T$ — the straight path's score. The cross term integrates to $2v\varepsilon\,\frac{\pi}{T}\cdot\big[\frac{T}{\pi}\sin(\pi t/T)\big]_0^T = 0$ . The quadratic term uses $\int_0^T \cos^2(\pi t/T)\,\mathrm{d}t = T/2$ :

S(\varepsilon) = \frac{1}{2}mv^2\,T + \frac{m\pi^2}{4T}\,\varepsilon^2

Why the vanishing cross term is the whole story: the term linear in $\varepsilon$ is precisely the first-order sensitivity of the ledger to this deformation direction. Its disappearance says the straight path cannot be beaten — or even changed — to first order: it is a stationary point of $S$ along this direction, and the surviving $\varepsilon^2$ term says it is in fact a local minimum.

Step 4. Put in numbers: $m = 1\ \mathrm{kg}$ , $d = 2\ \mathrm{m}$ , $T = 2\ \mathrm{s}$ , so $v = 1\ \mathrm{m/s}$ and $S(0) = 1.000\ \mathrm{J\,s}$ . The detour penalty is $\frac{m\pi^2}{4T}\varepsilon^2 = 1.234\,\varepsilon^2$ : for $\varepsilon = 0.25\ \mathrm{m}$ , $S = 1.077\ \mathrm{J\,s}$ ; for $\varepsilon = 0.5\ \mathrm{m}$ , $S = 1.308\ \mathrm{J\,s}$ — the scoreboard of the hero figure, now derived rather than asserted. Verified: numerical quadrature of $S(\varepsilon)$ on a 20 000-point grid reproduces 1.0000, 1.0771, and 1.3084 J·s for $\varepsilon = 0, 0.25, 0.5$ , matching the closed form to four decimals; the quadratic-only dependence on $\varepsilon$ confirms Landau–Lifshitz §2's general claim for the free particle.

Go deeper · Derivation B · 1 minThe Action Under a Gauge Shift of the Lagrangian

Derivation BThe Action Under a Gauge Shift of the Lagrangian

We prove equation (3): adding a total time derivative to the Lagrangian shifts the action of every path by one and the same endpoint constant, leaving all comparisons between paths — and hence the physics — untouched. The strategy is the fundamental theorem of calculus plus one observation about what "fixed endpoints" buys. Watch how little is assumed about the path: the argument holds for every candidate in the competition, not just the true one, which is what makes the gauge freedom exact rather than approximate.

Step 1. Let $L' = L + \dfrac{\mathrm{d}F(q,t)}{\mathrm{d}t}$ with $F$ any smooth function of coordinates and time, and compute the new action of an arbitrary candidate path $q(t)$ :

S'[q] = S[q] + \int_{t_1}^{t_2} \frac{\mathrm{d}}{\mathrm{d}t}F\big(q(t),t\big)\,\mathrm{d}t

Why the integrand is a genuine total derivative along the path: evaluated on a specific $q(t)$ , $F(q(t),t)$ is an ordinary function of one variable $t$ , and $\mathrm{d}F/\mathrm{d}t = \sum_k \frac{\partial F}{\partial q_k}\dot q_k + \frac{\partial F}{\partial t}$ is its ordinary derivative.

Step 2. Apply the fundamental theorem of calculus:

S'[q] = S[q] + F\big(q(t_2),t_2\big) - F\big(q(t_1),t_1\big) = S[q] + F\big(q^{(2)},t_2\big) - F\big(q^{(1)},t_1\big)

Step 3. Observe that the added constant is the same for every candidate: all paths in the competition share the endpoint values $q^{(1)}, q^{(2)}$ and the times $t_1, t_2$ by construction. The whole scoreboard translates rigidly; differences of scores, the location of the stationary point, and the second-order structure around it are all invariant. Why this matters beyond tidiness: it licenses every convenient rewriting of a Lagrangian — dropping total derivatives mid-calculation, the freedom to choose electromagnetic potentials — and it is the classical shadow of an unobservable global phase in quantum mechanics. *Verified: for the free particle with $F = \tfrac12 m q^2/T$ (an arbitrary test choice), explicit integration of $\mathrm{d}F/\mathrm{d}t$ along both the straight path and the $\varepsilon = 0.5$ detour of Derivation A adds the identical $F(d,T) - F(0,0) = md^2/2T = 1.0\ \mathrm{J\,s}$ to both scores, preserving their difference of 0.308 J·s. Consistent with Landau–Lifshitz §2 and Goldstein §1.4.*

Go deeper · Derivation C · 2 minThe Endpoint Derivatives of the Action Along the True Path

Derivation CThe Endpoint Derivatives of the Action Along the True Path

The goal is the pair (4)–(5): when the action is evaluated along the solved motion and regarded as a function $S(q,t)$ of the final endpoint, its gradient with respect to the endpoint position is the arriving momentum, and its time derivative is minus the energy function. The strategy is to vary the endpoint and keep, for once, the boundary term that fixed-endpoint arguments discard — that term is not garbage, it is the answer. The step to watch is the time derivative: moving the arrival time is not the same as moving the arrival point, and detangling the two is where the minus sign and the energy function appear.

Step 1. Vary the final position at fixed times. Comparing true paths that arrive at neighboring endpoints $q^{(2)}$ and $q^{(2)} + \delta q$ , the change in action is the usual first variation, which splits into a bulk integral and a boundary term. Along a true path the bulk integral vanishes — the path satisfies its equations of motion — leaving only

\delta S = \left[\sum_k \frac{\partial L}{\partial \dot q_k}\,\delta q_k\right]_{t_2} = \sum_k p_k\,\delta q_k \quad\Longrightarrow\quad \frac{\partial S}{\partial q_k} = p_k

Why only the upper boundary contributes: the initial endpoint stays nailed down, $\delta q(t_1) = 0$ ; and why the bulk dies: this derivation, unlike Derivation A, compares only true paths — stationarity is being used, not proved.

Step 2. Vary the arrival time. Along the true path, the total rate of change of the accumulated integral with its upper limit is just the integrand there: $\dfrac{\mathrm{d}S}{\mathrm{d}t_2} = L(t_2)$ .

Step 3. Expand that total derivative in terms of the two ways the endpoint data change as $t_2$ advances: the clock advances and the arrival point slides along the trajectory at rate $\dot q(t_2)$ :

\frac{\mathrm{d}S}{\mathrm{d}t_2} = \frac{\partial S}{\partial t_2} + \sum_k \frac{\partial S}{\partial q_k}\,\dot q_k = \frac{\partial S}{\partial t_2} + \sum_k p_k\dot q_k

Step 4. Set the two expressions equal and solve:

\frac{\partial S}{\partial t} = L - \sum_k p_k \dot q_k = -\,h

Why the combination $\sum p_k\dot q_k - L$ deserves its own name: it is the energy function (Jacobi integral) of the arriving motion, equal to $T + V$ for time-independent constraints — see the dedicated article. The action loses one unit of $h$ per unit of extra travel time granted. *Verified: for the free particle, $S(q,t) = mq^2/2t$ evaluated along true paths from the origin; then $\partial S/\partial q = mq/t = mv = p$ and $\partial S/\partial t = -mq^2/2t^2 = -\tfrac12 mv^2 = -E$ — both identities exact. Matches Goldstein §10.1 and Landau–Lifshitz §43.*

Limiting cases

Case	Result	Meaning
free particle, straight path	$S = \tfrac12 mv^2\,T$	pure kinetic ledger; every detour adds a positive $O(\varepsilon^2)$ penalty — a genuine minimum
short time interval	stationary path is a true minimum	before any conjugate point the $\varepsilon^2$ coefficient is positive in every direction; "least action" is honest here
gauge-shifted Lagrangian	$S' = S + F\vert_{1}^{2}$	rigid translation of the scoreboard; physics (differences) invariant
endpoint as variable, on-shell	$\partial S/\partial q = p$ , $\partial S/\partial t = -h$	the action becomes Hamilton's principal function — gateway to Hamilton–Jacobi theory
$S \gg \hbar$	classical mechanics	quantum phases $e^{iS/\hbar}$ of neighboring paths cancel except near the stationary path
$S \sim \hbar$	quantum regime	no single path dominates; the electron in an atom accumulates $S$ of order $\hbar$ per orbit

Common trapThe action is just another function of time that grows as the motion unfolds

It is not, and conflating the two objects that share the letter $S$ derails more students at this node than any algebra. The action functional $S[q]$ of equation (1) has no time argument at all: its input is a complete path with both endpoints already specified, and its output is one number — asking for " $S$ at time $t$ " or " $\mathrm{d}S/\mathrm{d}t$ " of a functional is a category error, like asking for the temperature of Tuesday's average. The running integral $\int_{t_1}^{t} L\,\mathrm{d}t'$ is a function of its upper limit, but it is a different object, and only becomes canonical when evaluated along the true path — at which point it earns the name Hamilton's principal function $S(q,t)$ , whose derivatives are the momenta and (minus) the energy, per equations (4)–(5). Keep the two straight with one question: "am I grading a whole candidate history (functional, brackets), or am I tracking the accumulated score of the actual motion (function, parentheses)?" The first is the arena of the variational principle; the second is the seed of Hamilton–Jacobi theory.

Worked examples

1 · Grading three histories of a free crossing (confidence builder)

A puck of mass $m = 1\ \mathrm{kg}$ slides frictionlessly from $x = 0$ to $x = 2\ \mathrm{m}$ , taking exactly $T = 2\ \mathrm{s}$ . Grade three candidate histories: (a) the uniform glide; (b) "procrastinate": rest for 1 s, then cross at $2\ \mathrm{m/s}$ ; (c) the sinusoidal detour $x(t) = t + 0.5\sin(\pi t/2)$ .

Step 1. Path (a): $\dot x = 1\ \mathrm{m/s}$ throughout, $V = 0$ , so $S_a = \int_0^2 \tfrac12(1)(1)^2\,\mathrm{d}t = \boxed{1.000\ \mathrm{J\,s}}$ .

Step 2. Path (b): the first second contributes nothing ( $\dot x = 0$ ); the second second has $\dot x = 2\ \mathrm{m/s}$ , contributing $\tfrac12(1)(4)(1)$ . Total $S_b = \boxed{2.000\ \mathrm{J\,s}}$ — procrastination doubles the bill, because the kinetic ledger is quadratic in speed: crossing the same distance in half the time costs $2^2 \times \tfrac12 = 2$ times the uniform score.

Step 3. Path (c): this is the $\varepsilon = 0.5$ member of Derivation A's family, so $S_c = 1.000 + \frac{(1)\pi^2}{4(2)}(0.5)^2 = 1.000 + 1.2337 \times 0.25 = \boxed{1.308\ \mathrm{J\,s}}$ .

Step 4. Read the scoreboard: $S_a < S_c < S_b$ , and no smooth deformation of (a) that respects the endpoints can lower its score even to first order (the vanishing cross term of Derivation A). For the free particle, the true path is not merely stationary but the outright minimum — the straightest possible use of the allotted time.

2 · The projectile's negative action, and beating the wrong paths (exam level)

A ball of mass $m = 0.2\ \mathrm{kg}$ is thrown straight up from the ground and returns to the ground $T = 2\ \mathrm{s}$ later ( $g = 9.81\ \mathrm{m/s^2}$ ). (i) Find the true path and its action. (ii) Grade the "lazy" history $y \equiv 0$ (never leave the ground) and the "triangle" history (rise at constant speed to the same peak height, descend symmetrically). (iii) Interpret, and count the quantum phase.

Step 1. The true path must launch with $v_0 = gT/2 = 9.81\ \mathrm{m/s}$ to return at $t = T$ : $y(t) = v_0 t - \tfrac12 g t^2$ , peaking at $y_{\max} = v_0^2/2g = 4.905\ \mathrm{m}$ .

Step 2. Its kinetic ledger: $\int_0^T \tfrac12 m(v_0 - gt)^2\,\mathrm{d}t = \tfrac12 m\big(v_0^2 T - v_0 g T^2 + \tfrac13 g^2 T^3\big) = 6.416\ \mathrm{J\,s}$ . Its potential ledger: $\int_0^T mg\,y\,\mathrm{d}t = mg\big(\tfrac12 v_0 T^2 - \tfrac16 g T^3\big) = 12.831\ \mathrm{J\,s}$ . Note the exact ratio $\int V\,\mathrm{d}t = 2\int T\,\mathrm{d}t$ — a virial-flavored identity special to the linear potential, and a strong arithmetic cross-check. The action: $S_{\mathrm{true}} = 6.416 - 12.831 = \boxed{-6.416\ \mathrm{J\,s}}$ .

Step 3. The lazy history $y \equiv 0$ : both $T$ and $V$ vanish identically, so $S_{\mathrm{lazy}} = \boxed{0}$ . The triangle history: constant climb speed $y_{\max}/(T/2) = 4.905\ \mathrm{m/s}$ , so $\int T\,\mathrm{d}t = \tfrac12(0.2)(4.905^2)(2) = 4.812\ \mathrm{J\,s}$ ; the average height is $y_{\max}/2$ , so $\int V\,\mathrm{d}t = mg\,\bar y\,T = (0.2)(9.81)(2.4525)(2) = 9.624\ \mathrm{J\,s}$ ; $S_{\mathrm{tri}} = 4.812 - 9.624 = \boxed{-4.812\ \mathrm{J\,s}}$ .

Step 4. Interpret: $S_{\mathrm{true}} = -6.416 < S_{\mathrm{tri}} = -4.812 < S_{\mathrm{lazy}} = 0$ . The parabola beats both rivals, and the sign of its score is meaningless — actions can be negative at will (shift $V$ by a constant and every score shifts), only comparisons count. The lazy path loses because it forfeits the huge $-V$ earnings available at altitude; the triangle loses because its abrupt speed profile overpays kinetic cost for the altitude it buys; the parabola is the perfectly negotiated compromise. For gravity's linear potential the second variation is positive along every deformation — no conjugate points exist — so the true path is a genuine minimum, not just a saddle.

Step 5. Quantum phase: $|S_{\mathrm{true}}|/\hbar = 6.416/(1.055\times10^{-34}) = 6.1\times10^{34}\ \mathrm{radians}$ . Neighboring histories a hair's breadth away differ by billions of full turns of phase and annihilate by interference; only the stationary path's neighborhood survives. That single number is why the ball has a trajectory at all.

Origins · 1 minPierre Louis Maupertuis and how this was discovered

Pierre Louis Maupertuis

1698–1759 · Paris Academy → Lapland expedition

Maupertuis led the 1736 expedition to Lapland that measured a degree of the meridian and proved Newton right about the flattened Earth — "the man who flattened the Earth and the Cassinis," Voltaire quipped, back when they were friends. In 1744 he announced something grander: a principle of least action, the claim that nature accomplishes every motion with the minimum expenditure of a quantity he called action. His version (essentially $\int mv\,\mathrm{d}s$ , the abbreviated action) was half theology — proof of a wise Creator's economy — and his mathematics was shaky; Euler, the same year, made it rigorous, and generously credited Maupertuis anyway. The priority battle that followed was the scientific scandal of the age: the mathematician König produced a letter attributing the idea to Leibniz, Maupertuis had him condemned by the Berlin Academy he presided over, and Voltaire — now an enemy — destroyed him in the satirical Histoire du docteur Akakia. Maupertuis never recovered. The physics did: purged of teleology by Lagrange and Hamilton, his "least action" became the organizing principle of classical and quantum mechanics alike.

History

1744

— Maupertuis announces the principle of least action to the Paris Academy, as much theology as mechanics; Euler's Methodus inveniendi makes the variational mathematics rigorous the same year.

1760

— Lagrange's

\delta

-calculus turns path comparison into an algorithm, freeing the principle from geometric case-by-case constructions.

1834

— Hamilton defines the action as

\int L\,\mathrm{d}t

over histories with fixed endpoints and times — the modern functional of this article — and shows all of dynamics follows from its stationarity.

1926

— Schrödinger, guided by the optical-mechanical analogy and Hamilton's principal function, writes the wave equation whose phase is

S/\hbar

1948

— Feynman promotes the action to the generator of quantum mechanics itself: every path contributes the amplitude

e^{iS/\hbar}

, and the classical world is the stationary-phase limit.

Going further

References. Morin, Introduction to Classical Mechanics, Ch. 6 · Goldstein, Poole & Safko, Classical Mechanics 3rd ed., Ch. 2 · Landau & Lifshitz, Mechanics (Vol. 1), §2 · Taylor, Classical Mechanics, Ch. 6–7 · Arnold, Mathematical Methods of Classical Mechanics, Ch. 3

Provenance. Definitions and derivations follow Goldstein (Classical Mechanics, Ch. 2), Landau & Lifshitz (Mechanics §2), Morin (Ch. 6), and Taylor (Ch. 6–7). A verifier should: (1) re-derive the Euler–Lagrange equations from $\delta S=0$ via integration by parts and the fundamental lemma; (2) confirm $L=\tfrac12 m\dot x^2-V$ yields $m\ddot x=-V'$ ; (3) check the polar-coordinate transform of the free-particle Lagrangian; (4) verify the dimension of $S$ as energy×time, matching that of $\hbar$ . All equations are standard textbook results; no numerical constants beyond dimensional statements are asserted. Worked-example numbers on this page (free particle: $S = 1.000 / 1.077 / 1.308 / 2.000\ \mathrm{J\,s}$ across the four graded paths; projectile: $S_{\mathrm{true}} = -6.416\ \mathrm{J\,s}$ with $\int V\,\mathrm{d}t = 2\int T\,\mathrm{d}t$ , $S_{\mathrm{tri}} = -4.812\ \mathrm{J\,s}$ , phase $6.1\times10^{34}\ \mathrm{rad}$ ) are illustrative arithmetic recomputed from the boxed formulas.

Builds on: The Lagrangian Function · Generalized Coordinates and Configuration Space · Virtual Displacements and the Principle of Virtual Work · d'Alembert's Principle Leads to: Hamilton's Principle of Stationary Action · The Euler-Lagrange Equations · Symmetries and Continuous Transformations · The Energy Function and Jacobi Integral

Is the action physically real, or just clever bookkeeping? Nothing seems to "have" 1 J·s.

Its value is bookkeeping; its comparisons are physics. The gauge shift (3) moves every path's score by a common constant, so no experiment can measure "the" action of a motion — in that respect $S$ is like potential energy, where only differences matter, pushed one level up. What is physical: differences of action between paths (they decide the stationary point), derivatives of the on-shell action (they are the momenta and energy, eqs. 4–5), and — decisively — the action modulo $2\pi\hbar$ , because $S/\hbar$ is the quantum phase of a history. Interferometers genuinely read out action differences: a neutron sent along two arms of different gravitational potential arrives with the measured phase difference $\Delta S/\hbar$ (the COW experiment). So the honest statement is: $S$ is exactly as real as a phase can be.

Why does action have the same units as angular momentum? Is that coincidence meaningful?

Both are instances of the pairing momentum × conjugate coordinate. Action accumulates $L\,\mathrm{d}t$ , and the endpoint derivatives (4)–(5) reveal its content as $p\,\mathrm{d}q - h\,\mathrm{d}t$ — momentum times displacement. Angular momentum is momentum times displacement around a loop, i.e. exactly the $\oint p\,\mathrm{d}q$ shape. That is why $\hbar$ can serve simultaneously as the quantum of action and the quantum of angular momentum: Bohr's old quantization $\oint p\,\mathrm{d}q = nh$ was a statement about an action integral, and spin-½ is an angular momentum of $\hbar/2$ . In the Hamiltonian formulation this pairing becomes the symplectic structure, and "actions" $J = \frac{1}{2\pi}\oint p\,\mathrm{d}q$ become half of the action-angle coordinate pairs — the same word, the same units, the same object viewed on closed loops.

Concretely, how is differentiating a functional different from differentiating a function?

A function of $n$ variables has $n$ partial derivatives; a functional has one "partial derivative" per point of the path — a continuum of them, collectively called the functional or variational derivative. The practical technology is the one-parameter family (2): pick a deformation direction $\eta(t)$ , reduce $S$ to the ordinary function $S(\varepsilon)$ , and differentiate at $\varepsilon = 0$ . Stationarity demands $S'(0) = 0$ for every admissible $\eta$ — infinitely many conditions, which is why it compresses into a differential equation holding at each instant rather than a finite set of equations. The systematic machinery ( $\delta$ -algebra, the fundamental lemma, the functional derivative $\delta S/\delta q(t)$ ) is built carefully in the calculus of variations article; this article's Derivation A is the machinery's content shown once by bare hands.

Does the particle somehow evaluate all paths before choosing? That sounds like teleology.

Classically, nothing is evaluated: the particle obeys a local differential equation instant by instant, and "the stationary path" is merely an equivalent global characterization of the same trajectory — two descriptions, one motion, no foresight required. But the discomfort has a genuine quantum resolution: in Feynman's formulation every path really does contribute an amplitude $e^{iS/\hbar}$ , and the sum over all of them is the propagator. No path is chosen; all are summed. When $S \gg \hbar$ , contributions from all but the stationary path's neighborhood cancel by destructive interference (Worked Example 2's $10^{34}$ radians), leaving what looks like a particle that selected the special path. The classical variational principle is thus the stationary-phase shadow of a democracy of histories — teleology dissolved into interference. The full story lives in the action principle in quantum theory.

My textbook also defines an "abbreviated action" with an integral of p dq. Is that the same S?

No — it is Maupertuis's older, closely related functional, and telling them apart matters. Hamilton's action (this article) is $S = \int L\,\mathrm{d}t$ over paths with fixed endpoints and fixed total time, and its stationarity gives the full equations of motion. The abbreviated action is $S_0 = \int \sum_k p_k\,\mathrm{d}q_k$ , compared over paths of fixed energy $E$ with the travel time left free; its stationarity (Maupertuis's principle) delivers the shape of the orbit but not the schedule along it. The two connect through $S = S_0 - E(t_2 - t_1)$ on shell, which is equation (5) integrated. Maupertuis's version is the one that morphs into Fermat's principle of optics (path length weighted by index of refraction) and underlies the optical-mechanical analogy Schrödinger exploited; Hamilton's version is the one quantum field theory generalizes. When a text says "least action" without qualification, check which functional and which constraint it means.

Can the action be negative, and does a more negative action mean anything energetically?

It can, freely — Worked Example 2's true path scores $-6.42\ \mathrm{J\,s}$ — and no, nothing energetic follows from the sign. Two separate freedoms make the sign conventional: shifting the potential by a constant $c$ shifts every score by $-c(t_2 - t_1)$ , and the gauge shift (3) adds arbitrary endpoint constants. A path of very negative action is one that lingers profitably where $V$ is high relative to the reference — the projectile earns $-V$ at altitude — not one that is "low energy": the energy is not even constant across the family of compared paths, and the true path's energy is encoded in the action's time derivative (eq. 5), not its value. Treat the action like a golf score with an arbitrary par: only relative standings decide the tournament.

Why must both endpoints be fixed in position and time? What if I know the initial position and velocity instead?

The variational competition needs a well-posed field of candidates, and fixed $(q_1, t_1) \to (q_2, t_2)$ is the choice that makes the boundary term of the first variation vanish, yielding clean local equations. It is a characterization, not a claim about how you launch balls: physically you usually know initial position and velocity, and the differential equations you extract are solved with those data as usual. The two viewpoints select the same trajectories — one slices the solution set by boundary values, the other by initial values. Occasionally the boundary-value slicing genuinely misbehaves: for some endpoint pairs several true paths exist (throw the ball high and fast or low and quick between the same two events), and at conjugate points neighboring true paths refocus, which is where minima degrade into saddles. Those subtleties — and what "stationary" means there — are exactly the subject of Hamilton's principle.

If only variations of S matter, why do lattice QCD computations care about the actual value of S?

Because quantum mechanics reads the action twice. In real time, amplitudes go as $e^{iS/\hbar}$ and only phase differences — hence action differences — are observable, consistent with the gauge story of equation (3). Lattice QCD works instead in imaginary time (Euclidean signature), where the weight becomes the real, positive $e^{-S_E}$ and the action functions as an energy-like cost in a statistical ensemble: configurations are sampled with probability proportional to that weight, so the relative values of $S_E$ across configurations control everything. It is still comparisons that matter — a constant shift of $S_E$ cancels between numerator and denominator of every expectation value — but now the comparisons span field configurations rather than particle paths. The functional concept is unchanged: a machine that eats an entire history (there, a field over spacetime) and returns one grading number.