Hamilton's Principle of Stationary Action

Modern fundamental physics is written action-first. General relativity is the stationarity of the Einstein–Hilbert action; the Standard Model is defined by its Lagrangian density and $\delta S=0$; string theory's starting point is the area action of a worldsheet. When lattice-QCD collaborations burn hundreds of millions of core-hours on exascale machines in 2026, what the machines actually sample is $e^{-S}$ over field histories — Feynman's sum over paths in imaginary time, with Hamilton's stationary configurations dominating the integral. The full story of how $\delta S=0$ becomes quantum amplitude lives in the action-principle-to-QFT article. The principle also earns its keep as an algorithm. Discretize the action itself — not the differential equation — and demand stationarity of the sum: the result is a variational integrator, a time-stepping scheme that inherits the structural blessings of the principle (exact conservation of the symplectic form and momentum maps, no secular energy drift). This is why spacecraft trajectory designers and molecular-dynamics packages favor symplectic/variational schemes for million-orbit and billion-step simulations: a Runge–Kutta orbit spirals unphysically after enough steps, while a variational integrator's energy error stays bounded for the entire run. The same discrete-mechanics machinery now drives robotics trajectory optimization, where "find the motion connecting these two configurations" is solved directly as a discrete stationary-action problem — Hamilton's boundary-value question, asked by a motion planner sixty times a second.

Pin down two events: the system is at configuration $q_1$ at time $t_1$, and at configuration $q_2$ at time $t_2$. Between those pins runs an infinity of conceivable histories — fast ones, lazy ones, looping ones. The action functional grades each of them with one number, $S=\int(T-V)\,\mathrm{d}t$. Hamilton's principle says: nature's path is the one sitting at a flat spot of that grading. Wiggle the true path slightly anywhere in the interior and the score changes only at second order in the wiggle — exactly the way a function sits flat at the bottom of a valley or on a mountain pass.

Why should the flat spot be the physical path, and why is the score $T-V$ rather than, say, $T+V$? Watch a ball tossed straight up. Kinetic energy wants the trip over quickly (less $\int T$); potential energy rewards loitering up high (more $\int V$, hence less $S=\int T\,\mathrm{d}t-\int V\,\mathrm{d}t$). Those two pressures pull the schedule in opposite directions, and the true parabolic flight is the perfect compromise — quick enough that kinetic cost stays modest, high enough and long enough at the top that the potential credit is collected. The stationary point of $S$ is exactly the schedule at which trading a little speed for a little altitude no longer pays either way. The minus sign is the whole mechanism: with $T+V$ both terms would push the same way and nothing would balance.

One more picture belongs up front, because it corrects the single most repeated error about this subject. The principle is stationary action, not least action. For a harmonic oscillator, launch a family of paths from one point with slightly different velocities: they fan out, and — this is the special property of the oscillator — all reconverge to a point after exactly half a period, no matter the amplitude. That refocusing point is called the kinetic focus (or conjugate point). For trips shorter than the focus the true path genuinely minimizes $S$; for trips that pass through it, there exist wiggles that lower the action, and the true path is only a saddle — still perfectly stationary, no longer a minimum.

Finally, the apparent teleology. How can a ball "know" at launch that it must arrive at $q_2$ at $t_2$? It doesn't have to. Stationarity of a global integral against local wiggles is a condition enforceable instant by instant — it compresses into a differential equation at each interior time, which is exactly the Euler–Lagrange machinery. The global statement and the local law carry identical information; the principle just packages all the local conditions into one scalar sentence. And beneath that equivalence sits Feynman's quantum answer: every path is actually taken, each contributing a phase $e^{iS/\hbar}$; away from the stationary path the phases spin wildly and cancel, while near $\delta S=0$ they reinforce. Classical purposefulness is quantum bookkeeping seen from far away.

Everything begins with the score being graded. For a system described by generalized coordinates $q=(q_1,\dots,q_n)$ with Lagrangian $L(q,\dot q,t)$ — built as $T-V$ in whatever coordinates make the constraints invisible — each admissible history between the pinned endpoints is assigned its action. This is the functional studied in detail in the action-functional article; here it becomes the protagonist of a law of nature:

where: $S$ — the action, one real number per path (units J·s); $t_1,t_2$ — fixed initial and final times; $q(t)$ — any twice-differentiable path satisfying $q(t_1)=q_1$, $q(t_2)=q_2$. Note what is being held and what is free: both endpoint configurations and both times are data; the entire interior of the path is variable. Hamilton's principle is then the statement that the physical motion is a stationary point of this functional. Writing $\delta$ for the first-order change under an infinitesimal path deformation:

where: $\delta q_i(t)$ — an arbitrary smooth virtual displacement of coordinate $i$, free in the interior, clamped to zero at both ends. The clamping is not a technicality: it is what makes the boundary term in the variation vanish, and it encodes the physical question being asked — which path connects these two given configurations in the given time? To make $\delta$ concrete, embed the candidate path in a one-parameter family. This device — comparison paths labelled by a single dial $\alpha$ — is what turns functional calculus into ordinary calculus:

where: $\alpha$ — the smallness dial ($\alpha=0$ is the candidate path); $\eta_i(t)$ — an arbitrary fixed deformation shape. Stationarity means $\mathrm{d}S/\mathrm{d}\alpha|_{\alpha=0}=0$ for every admissible shape $\eta_i$ — a far stronger demand than flatness in one direction, and the reason a single scalar condition can determine $n$ functions of time. What does the demand deliver? Differentiating under the integral and integrating the velocity term by parts (the boundary term dies by the endpoint clamping) leaves

where: the bracket multiplying each independent $\delta q_i$ must vanish separately, by the fundamental lemma of the calculus of variations (proved carefully in the calculus-of-variations article — the machinery is borrowed, not rebuilt here). The result is the local content of the global principle:

where: one second-order ODE per degree of freedom — the Euler–Lagrange equations, whose anatomy, solution craft, and worked classics belong to their own article. The point to absorb here is the equivalence itself: the teleological-looking global statement (2) and the causal local law (5) are the same fact. Nothing "knows the future"; a path that is flat against all wiggles is flat against localized wiggles at each instant, and that instant-by-instant flatness is the differential equation. Two structural properties of the principle do belong to this article. First, the Lagrangian entering (1) is not unique. Add to $L$ the total time derivative of any function of coordinates and time, and the action shifts by a pure boundary constant:

where: $F(q,t)$ — any differentiable function. Because endpoints are pinned, the added term is the same constant for every competitor path — it moves the whole grading up or down without tilting it, so the flat spot (and hence the physics) is untouched. This gauge freedom is why the electromagnetic Lagrangian can change under gauge transformations while the Lorentz force stays put. Second, the principle as stated requires monogenic forces — every force derivable from a potential $V(q,t)$ or a generalized velocity-dependent potential $U(q,\dot q,t)$, as for the Lorentz force. Dissipation breaks the pattern, but does not kill the method: the variational statement extends to

where: $Q_i^{(\mathrm{nc})}$ — the generalized nonpotential forces (friction, drag). Equation (7) is no longer the stationarity of any single functional — there is no "action for friction" — but it remains a valid and useful variational bookkeeping, and it reduces to (2) the moment the $Q^{(\mathrm{nc})}$ vanish.

Which statement answers which question

The name "principle of least action" now needs its overdue correction. Stationarity guarantees flatness, not depth — and for mechanical Lagrangians the flat spot can genuinely stop being a minimum. The clean way to see it is to compute the second variation and watch its sign.

One more boundary needs patrolling, because two historical action principles coexist and answer different questions. Hamilton's principle pins the two times and lets varied paths have any energy. The older principle of Maupertuis pins the energy and lets the transit time float. Confusing them is a classic exam disaster.

1 · Grading three schedules for a tossed ball (confidence builder)

A ball of mass $m=0.20\ \mathrm{kg}$ is thrown straight up and caught at the same height $1.00\ \mathrm{s}$ later (take $g=10\ \mathrm{m/s^2}$ for clean numbers, so the true flight is $x(t)=5t-5t^2$, peaking at $1.25\ \mathrm{m}$). Compute the action for (a) the true parabola, (b) the "lazy parabola" family $x_c(t)=c\,t(1-t)$, and (c) a constant-speed triangle to the same peak height. Confirm the true path is stationary — and here, minimal.

Step 1. Set up the family. For $x_c(t)=c\,t(1-t)$: velocity $\dot x_c = c(1-2t)$, so $T(t)=\tfrac12(0.2)\,c^2(1-2t)^2$ and $V(t)=(0.2)(10)\,c\,t(1-t)=2c\,t(1-t)$.

Step 2. Integrate the kinetic term: $\int_0^1(1-2t)^2\mathrm{d}t = \tfrac13$, so $\int T\,\mathrm{d}t = 0.1\,c^2\cdot\tfrac13 = c^2/30$. Integrate the potential term: $\int_0^1 t(1-t)\,\mathrm{d}t=\tfrac16$, so $\int V\,\mathrm{d}t = 2c\cdot\tfrac16 = c/3$.

Step 3. The action of the family is an ordinary parabola in the dial $c$:

where: $c$ — the launch-speed dial (the true path has $c=5$, since $x=5t-5t^2$). Setting $\mathrm{d}S/\mathrm{d}c = 2c/30 - 1/3 = 0$ gives $c=5$ exactly: the stationary member of the family is the true flight, no equations of motion consulted.

Step 4. Put in numbers. $S(5) = 25/30 - 5/3 = \boxed{-0.833\ \mathrm{J\cdot s}}$. The equal-distance neighbours: $S(4) = 16/30-4/3 = -0.800\ \mathrm{J\cdot s}$ and $S(6)=36/30-2 = -0.800\ \mathrm{J\cdot s}$ — symmetric and both larger, the signature of a flat minimum.

Step 5. Grade the outsider. The triangle rises at constant $2.5\ \mathrm{m/s}$ to $1.25\ \mathrm{m}$ and returns: $\int T\,\mathrm{d}t = \tfrac12(0.2)(2.5)^2\cdot 1 = 0.625\ \mathrm{J\cdot s}$; mean height $0.625\ \mathrm{m}$ gives $\int V\,\mathrm{d}t = (0.2)(10)(0.625) = 1.25\ \mathrm{J\cdot s}$; so $S_{\triangle} = \boxed{-0.625\ \mathrm{J\cdot s}}$ — worse than every parabola tried. Same endpoints, same duration, same peak height: the schedule is what the action grades.

2 · Locating the kinetic focus of an oscillator numerically (exam level)

A glider of mass $m=0.50\ \mathrm{kg}$ on a spring has $\omega = 2.0\ \mathrm{rad/s}$ (period $3.14\ \mathrm{s}$). For a trip of duration (a) $T=1.0\ \mathrm{s}$ and (b) $T=2.0\ \mathrm{s}$, evaluate the second variation $\delta^2 S$ for the half-wave wiggle of amplitude $a=1.0\ \mathrm{cm}$, and interpret.

Step 1. The tool is the boxed result of the kinetic-focus derivation: $\delta^2 S = \dfrac{ma^2}{2}\Big(\dfrac{\pi^2}{T}-\omega^2 T\Big)$, exact for the oscillator. The prefactor: $\tfrac12 m a^2 = \tfrac12(0.50)(0.010)^2 = 2.5\times10^{-5}\ \mathrm{kg\,m^2}$.

Step 2. Case (a), $T=1.0\ \mathrm{s}$: bracket $= \pi^2/1.0 - (2.0)^2(1.0) = 9.870 - 4.000 = 5.870\ \mathrm{s^{-1}}$. So $\delta^2 S = 2.5\times10^{-5}\times 5.870 = \boxed{+1.47\times10^{-4}\ \mathrm{J\cdot s}}$. Positive: the wiggle raises the action — the true path is a minimum here.

Step 3. Case (b), $T=2.0\ \mathrm{s}$: bracket $= \pi^2/2.0 - 4.0\times2.0 = 4.935 - 8.000 = -3.065\ \mathrm{s^{-1}}$. So $\delta^2 S = 2.5\times10^{-5}\times(-3.065) = \boxed{-7.66\times10^{-5}\ \mathrm{J\cdot s}}$. Negative: this gentle half-wave lowers the action below the true path's — a saddle, past the focus.

Step 4. Locate the crossover: $\pi^2/T = \omega^2 T$ gives $T = \pi/\omega = \pi/2.0 = \boxed{1.57\ \mathrm{s}}$ — exactly half the $3.14\ \mathrm{s}$ period, where the fan of oscillator paths refocuses. A $1.0\ \mathrm{s}$ trip sits safely inside "least action"; a $2.0\ \mathrm{s}$ trip has already left it. The principle $\delta S=0$ holds identically in both cases — only the word least comes and goes.

References. Goldstein, Poole & Safko, Classical Mechanics, 3rd ed., §2.1–2.3 · Landau & Lifshitz, Mechanics, 3rd ed., §2 · Morin, Introduction to Classical Mechanics, Ch. 6 · Taylor, Classical Mechanics, Ch. 7 · Lanczos, The Variational Principles of Mechanics, Ch. V

Provenance. Equations follow Goldstein §2.1–2.3, Landau-Lifshitz §2 and §44, and Morin Ch. 6. A verifier should: recompute the integration-by-parts yielding the Euler–Lagrange equations; verify the second-variation result $\delta^2 S=\tfrac{ma^2}{2}(\pi^2/T-\omega^2 T)$ and its sign change at $T=\pi/\omega$ (cf. Gray & Taylor, AJP 75, 434, 2007); confirm the Lorentz-force Lagrangian $L=\tfrac12 m v^2 - q\phi + q\,\mathbf v\cdot\mathbf A$ reproduces $m\ddot{\mathbf r}=q(\mathbf E+\mathbf v\times\mathbf B)$; and check the Maupertuis relation $W=S+E(t_2-t_1)$. No empirical constants are quoted. Worked-example numbers on this page ($S(5)=-0.833\ \mathrm{J\cdot s}$, $S(4)=S(6)=-0.800\ \mathrm{J\cdot s}$, $S_\triangle=-0.625\ \mathrm{J\cdot s}$; $\delta^2S=+1.47\times10^{-4}$ and $-7.66\times10^{-5}\ \mathrm{J\cdot s}$, focus at $1.57\ \mathrm{s}$) are illustrative arithmetic recomputed from the boxed formulas.

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