The action principle and variational calculus

Intuition [Beginner]

Newton's laws say: here are the forces, now compute the acceleration, then integrate twice to get the motion. The action principle says something radically different: of all the paths a particle could take between two points at two times, the one it actually takes is the path that makes a certain number — the action — as small as possible (more precisely, stationary).

What is this number? For each candidate path, you compute the Lagrangian at every point along the path. The Lagrangian is kinetic energy minus potential energy: $L=T-V$. The action is the accumulated sum of $L$ over the entire path — for a continuous path, this is $S=$ the area under the $L$-versus-time curve from $t_1$ to $t_2$, which equals the area under the $T$ curve minus the area under the $V$ curve.

Nature picks the path that makes $S$ stationary — meaning that small wiggles of the path do not change $S$ to first order. For a ball thrown in a parabolic arc, the actual parabola is the one path between the throw and the catch for which $S$ is stationary. Every other nearby curve (a slightly higher arc, a slightly lower one) gives a larger action.

Why should you care? Three reasons.

First, the action principle is equivalent to Newton's laws for standard mechanical systems. Every Newtonian $F=ma$ problem can be reformulated as "find the path that makes $S$ stationary," and you get the same answer. But the action principle works in situations where Newton's laws are awkward: systems with constraints (a bead on a wire, a pendulum with a rigid rod), systems with many particles, systems in curvilinear coordinates.

Second, the action principle is coordinate-independent. Newton's $F=ma$ requires you to choose coordinates and decompose forces into components. The action $S$ is a single number — it does not care what coordinates you compute it in. This makes it the natural starting point for general relativity and quantum field theory.

Third, the action principle is the bridge to quantum mechanics. In Feynman's path-integral formulation, every path contributes, weighted by $e^{iS/\hslash }$. In the classical limit ($\hslash \to 0$), the paths near the stationary-action path reinforce each other and all others cancel — recovering the action principle. Classical mechanics is quantum mechanics in the limit where the action is much larger than Planck's constant.

Visual [Beginner]

Figure: A particle must travel from point A (left) at time $t_1$ to point B (right) at time $t_2$. Several candidate paths are drawn: the actual physical path (solid curve, a parabola for projectile motion) and two nearby alternative paths (dashed curves, one slightly higher, one slightly lower). Below each path is a number: the action $S$ is smallest for the physical path. The physical path is labelled "stationary action." A small arrow shows a perturbation of the physical path; the text reads "a small wiggle changes $S$ by zero to first order."

Worked example [Beginner]

A free particle (no forces, so $V=0$) moves in one dimension from $x(0)=0$ to $x(2)=10$ metres in 2 seconds. Its Lagrangian is $L=T=\frac{1}{2}m{\dot{x}}^2$ (kinetic energy only, no potential). What path $x(t)$ makes the action stationary?

The action is $S=$ the accumulated kinetic energy over 2 seconds. For a straight-line path $x(t)=5t$ (constant velocity $v=5$ m/s), the kinetic energy is constant at $\frac{1}{2}m(25)$, so $S=\frac{1}{2}m\times 25\times 2=25m$ joule-seconds.

Now try a detoured path: $x(t)=5t+ϵ\sin (\pi t)$, which also connects the endpoints (the sine term is zero at $t=0$ and $t=2$). Its velocity is $\dot{x}=5+ϵ\pi \cos (\pi t)$, so ${\dot{x}}^2=25+10ϵ\pi \cos (\pi t)+{ϵ}^2{\pi }^2{\cos }^2(\pi t)$. The action is:

$S=\frac{1}{2}m\times 25\times 2+\frac{1}{2}m\times 10ϵ\pi \times 0+\frac{1}{2}m{ϵ}^2{\pi }^2\times 1=25m+0+\frac{1}{2}m{ϵ}^2{\pi }^2$.

The first-order term vanishes (the cosine term averages to zero over the full interval). The action differs from the straight-line value only at second order in $ϵ$. The straight-line path has stationary action — the first variation $\delta S$ is zero.

What this tells us: a free particle, subject to no forces, travels in a straight line at constant speed — Newton's first law recovered from the action principle. The action is stationary (not necessarily minimal) at the physical path. For this particular example it happens to be a minimum; for other systems it can be a saddle point.

Check your understanding [Beginner]

Formal definition [Intermediate+]

Let $Q$ be the configuration space of a mechanical system (a smooth manifold of dimension $n$, the number of degrees of freedom). A path is a smooth curve $\gamma :[t_1,t_2]\to Q$ with fixed endpoints $\gamma (t_1)=q_1$ and $\gamma (t_2)=q_2$.

The Lagrangian is a smooth function $L:TQ\times \mathbb{R}\to \mathbb{R}$ on the tangent bundle (position-velocity space) and time. For a standard mechanical system, $L(q,\dot{q},t)=T(q,\dot{q})-V(q,t)$ where $T$ is the kinetic energy and $V$ the potential energy.

The action functional $S:\mathcal{P}\to \mathbb{R}$ is defined on the space $\mathcal{P}$ of smooth paths with fixed endpoints by

$$S[\gamma ]:={\int }_{t_1}^{t_2}L(\gamma (t),\dot{\gamma }(t),t)dt.$$

Hamilton's principle of stationary action. The physical trajectory of the system is the path $\gamma$ for which the action is stationary under variations that vanish at the endpoints:

$$\delta S=0\text{for all }\delta \gamma \text{ with }\delta \gamma (t_1)=\delta \gamma (t_2)=0.$$

To compute $\delta S$, consider a one-parameter family of paths ${\gamma }_{ϵ}(t)=\gamma (t)+ϵ\eta (t)$ where $\eta (t_1)=\eta (t_2)=0$ and $ϵ$ is small. The first variation is

$$\delta S:=\frac{d}{dϵ}{|}_{ϵ=0}S[{\gamma }_{ϵ}]={\int }_{t_1}^{t_2}\left(\frac{\partial L}{\partial q^i}{\eta }^i+\frac{\partial L}{\partial {\dot{q}}^i}{\dot{\eta }}^i\right)dt.$$

Integrating the second term by parts (the boundary term vanishes because $\eta (t_1)=\eta (t_2)=0$):

$$\delta S={\int }_{t_1}^{t_2}\left(\frac{\partial L}{\partial q^i}-\frac{d}{dt}\frac{\partial L}{\partial {\dot{q}}^i}\right){\eta }^{i}dt.$$

Since ${\eta }^i(t)$ is arbitrary (subject to the endpoint conditions), $\delta S=0$ for all $\eta$ if and only if

$$\boxed{\frac{\partial L}{\partial q^i}-\frac{d}{dt}\frac{\partial L}{\partial {\dot{q}}^i}=0,i=1,\dots ,n.}$$

These are the Euler-Lagrange equations, derived fully and applied in 09.02.02 pending. For this unit, the key point is the principle itself: the Euler-Lagrange equations are the condition for the action to be stationary.

The action principle and Newton's second law. For a single particle in Cartesian coordinates with $L=\frac{1}{2}m|\dot{\mathbf{r}}{|}^2-V(\mathbf{r})$, the Euler-Lagrange equations give $m\ddot{\mathbf{r}}=-\nabla V=\mathbf{F}$. So Newton's second law is recovered as a special case of the action principle for the standard Lagrangian $L=T-V$.

Counterexamples to common slips

• The action is not necessarily minimised. The principle is "stationary action," not "least action." For most simple mechanical systems over short times the stationary path is indeed a minimum, but for the harmonic oscillator with period $T$, paths of duration greater than $T/2$ have the stationary path as a saddle point. The correct terminology is Hamilton's principle or the principle of stationary action.

• The Lagrangian is not unique. If $L$ and $L^{\prime }=L+\frac{d}{dt}F(q,t)$ differ by a total time derivative, they yield the same Euler-Lagrange equations and the same physical trajectories. This gauge freedom in the Lagrangian is important for electromagnetic potentials and gauge field theories.

• The action principle requires fixed endpoints. The variation $\delta \gamma$ must vanish at $t_1$ and $t_2$. Free-endpoint variations give additional boundary conditions (natural boundary conditions) but change the variational problem.

Key theorem with proof [Intermediate+]

Theorem (Hamilton's principle is equivalent to Newton's second law for standard mechanical systems). Let a particle of mass $m$ move in a potential $V(\mathbf{r})$ with no constraints. The action $S[\gamma ]={\int }_{t_1}^{t_2}(\frac{1}{2}m|\dot{\mathbf{r}}{|}^2-V(\mathbf{r}))dt$ is stationary at $\gamma$ if and only if $\gamma$ satisfies $m\ddot{\mathbf{r}}=-\nabla V$ (Newton's second law with conservative force).

Proof. The Lagrangian is $L=\frac{1}{2}m({\dot{x}}^2+{\dot{y}}^2+{\dot{z}}^2)-V(x,y,z)$. The Euler-Lagrange equation for the $x$-component is:

$$\frac{\partial L}{\partial x}-\frac{d}{dt}\frac{\partial L}{\partial \dot{x}}=-\frac{\partial V}{\partial x}-\frac{d}{dt}(m\dot{x})=-\frac{\partial V}{\partial x}-m\ddot{x}=0.$$

Rearranging: $m\ddot{x}=-\partial V/\partial x$. The $y$ and $z$ components give identical equations. Together:

$$m\ddot{\mathbf{r}}=-\nabla V=\mathbf{F},$$

which is Newton's second law for a conservative force. The converse follows by running the derivation in reverse: if $m\ddot{\mathbf{r}}=-\nabla V$, then the Euler-Lagrange equations hold, and therefore $\delta S=0$ for all variations. ∎

Corollary. For any system with a standard Lagrangian $L=T(\dot{q})-V(q)$, Hamilton's principle reproduces Newton's second law in generalised coordinates. The Euler-Lagrange equations are the coordinate-independent generalisation of $F=ma$.

The power of this equivalence is that the action principle works in any coordinate system without recomputing force components. For a pendulum, writing $L$ in the angular coordinate $\theta$ and applying the action principle gives the equation of motion directly, without resolving forces along and perpendicular to the string.

Worked example: the brachistochrone

Find the curve $y(x)$ between $(0,0)$ and $(x_0,y_0)$ (with $y$ positive downward) along which a bead slides frictionlessly under gravity in the least time.

The transit time is $T=\int ds/v$ where $ds=\sqrt{1+y^{\prime 2}}dx$ and $v=\sqrt{2gy}$ (energy conservation). So the functional to minimise is:

$$T[y]={\int }_0^{x_0}\sqrt{\frac{1+y^{\prime 2}}{2gy}}dx.$$

The integrand $f(y,y^{\prime })=\sqrt{(1+y^{\prime 2})/(2gy)}$ has no explicit $x$-dependence, so the Beltrami identity gives $f-y^{\prime }(\partial f/\partial y^{\prime })=C$. Computing:

$$\frac{\partial f}{\partial y^{\prime }}=\frac{y^{\prime }}{\sqrt{2gy(1+y^{\prime 2})}},f-y^{\prime }\frac{\partial f}{\partial y^{\prime }}=\frac{1}{\sqrt{2gy(1+y^{\prime 2})}}=\frac{1}{\sqrt{2g}}\cdot \frac{1}{\sqrt{y(1+y^{\prime 2})}}.$$

Setting this equal to a constant $1/\sqrt{2ga}$: $y(1+y^{\prime 2})=2a$. Parametrising $y^{\prime }=\cot (\theta /2)$ gives the cycloid $x=a(\theta -\sin \theta )$, $y=a(1-\cos \theta )$. The brachistochrone is a cycloid — the same curve generated by a point on a rolling wheel. This was the problem that launched the calculus of variations (Johann Bernoulli, 1696).

Exercises [Intermediate+]

Lean formalization [Intermediate+]

Mathlib has Frechet derivatives on Banach spaces (Analysis.Calculus.FrechetDeriv) and the smooth manifold structure needed for path spaces. It does not formalise the action functional as a map from a space of curves to the reals, nor Hamilton's principle as a first-variation condition, nor the derivation of the Euler-Lagrange equations by integration by parts. The pieces exist (integration, differentiation by parts, ODE theory) but are not assembled into a variational-calculus theorem. lean_status: none.

The variational principle on configuration manifolds [Master]

The intermediate-tier derivation uses local coordinates on the configuration manifold $Q$. The coordinate-free formulation lifts the action functional to the space of smooth paths in $Q$ and expresses the variational principle in terms of geometric objects on the tangent bundle $TQ$.

Let $\gamma :[t_1,t_2]\to Q$ be a smooth curve. Its lift to $TQ$ is the curve $\dot{\gamma }:[t_1,t_2]\to TQ$ given by $\dot{\gamma }(t)=(\gamma (t),\dot{\gamma }(t))$, where the second component is the velocity vector at time $t$. The Lagrangian $L:TQ\times \mathbb{R}\to \mathbb{R}$ pulls back along the lift to give the time-dependent function $L(\dot{\gamma }(t),t)$, and the action is the integral of this pullback.

A variation of $\gamma$ is a smooth map $\Gamma :(-ϵ,ϵ)\times [t_1,t_2]\to Q$ with $\Gamma (0,t)=\gamma (t)$ and $\Gamma (s,t_1)=\gamma (t_1)$, $\Gamma (s,t_2)=\gamma (t_2)$ for all $s$. The variation field $V(t)=\partial \Gamma /\partial s{|}_{s=0}$ is a vector field along $\gamma$ vanishing at the endpoints. The first variation is

$$\delta S=\frac{d}{ds}{|}_{s=0}{\int }_{t_1}^{t_2}L({\dot{\Gamma }}_s(t),t)dt={\int }_{t_1}^{t_2}\left(\frac{\partial L}{\partial q^i}-\frac{d}{dt}\frac{\partial L}{\partial {\dot{q}}^i}\right)V^{i}dt.$$

In local coordinates this recovers the Euler-Lagrange equations derived in the Intermediate tier. The coordinate-free content is that the equations hold in every chart, by the fundamental lemma of the calculus of variations applied to the component functions.

The deeper geometric structure is captured by the Poincare-Cartan 1-form on $TQ$:

$${\theta }_L:=\frac{\partial L}{\partial {\dot{q}}^i}d{q}^i-Ldt.$$

Its exterior derivative ${\omega }_L=d{\theta }_L$ is a 2-form on $TQ$. For a regular Lagrangian (one for which the fibre Hessian ${\partial }^{2}L/\partial {\dot{q}}^i\partial {\dot{q}}^j$ is non-degenerate), ${\omega }_L$ is a symplectic form. The Euler-Lagrange equations are equivalent to the condition that the lifted physical curve $\dot{\gamma }$ is an integral curve of the Lagrangian vector field $X_E$ defined by ${\omega }_L(X_E,\cdot )=dE$, where $E={\dot{q}}^i\partial L/\partial {\dot{q}}^i-L$ is the energy function.

This reformulation identifies the Euler-Lagrange dynamics with symplectic geometry on $TQ$: the action principle selects curves that are Hamiltonian flow lines of $X_E$ on the symplectic manifold $(TQ,{\omega }_L)$. The Legendre transform 09.04.01 pending maps this picture to the cotangent bundle $T^{\ast }Q$ with its canonical symplectic form ${\omega }_0=d{p}_i\wedge d{q}^i$, and the Hamiltonian vector field $X_H$ on $T^{\ast }Q$ is the Legendre-dual of $X_E$.

Proposition. If $L$ is a regular Lagrangian on $TQ$ and $\gamma$ is a solution of the Euler-Lagrange equations, then the lifted curve $\dot{\gamma }$ in $TQ$ is an integral curve of the Lagrangian vector field $X_E$ satisfying ${\omega }_L(X_E,\cdot )=dE$.

Proof. In local coordinates, write ${\omega }_L=d{p}_i\wedge d{q}^i-(\partial L/\partial q^i)d{q}^i\wedge dt$ where $p_i=\partial L/\partial {\dot{q}}^i$. Along the lifted curve $\dot{\gamma }$, the tangent vector has components $({\dot{q}}^i,{\ddot{q}}^i)$. The Euler-Lagrange equations give ${\dot{p}}_i=\partial L/\partial q^i$. Contracting $i_{{\dot{\gamma }}^{\prime }}{\omega }_L$ and using ${\dot{p}}_i=\partial L/\partial q^i$, the result $i_{{\dot{\gamma }}^{\prime }}{\omega }_L=dE({\dot{\gamma }}^{\prime })$ follows, confirming $\dot{\gamma }$ is an integral curve of $X_E$. $\square$

The second variation, Jacobi fields, and conjugate points [Master]

The first variation determines whether the action is stationary. The second variation ${\delta }^{2}S$ determines the character of the stationary point: minimum, maximum, or saddle.

Consider a variation $\Gamma (s,t)=\gamma (t)+sV(t)+\frac{1}{2}{s}^{2}W(t)+O(s^3)$ where $V$ is the variation field and $W$ is a second-order correction. The second variation of the action at a stationary path $\gamma$ is

$${\delta }^{2}S[V]={\int }_{t_1}^{t_2}\left(\frac{{\partial }^{2}L}{\partial {\dot{q}}^i\partial {\dot{q}}^j}{\dot{V}}^i{\dot{V}}^j+2\frac{{\partial }^{2}L}{\partial {\dot{q}}^i\partial q^j}{\dot{V}}^i{V}^j+\frac{{\partial }^{2}L}{\partial q^i\partial q^j}{V}^i{V}^j\right)dt.$$

For the geodesic Lagrangian $L=\frac{1}{2}{g}_{ij}{\dot{q}}^i{\dot{q}}^j$ on a Riemannian manifold $(Q,g)$, this simplifies to

$${\delta }^{2}S[V]={\int }_{t_1}^{t_2}\left(g_{ij}{\dot{V}}^i{\dot{V}}^j-R_{ikjl}{\dot{q}}^k{\dot{q}}^l{V}^i{V}^j\right)dt,$$

where $R_{ikjl}$ is the Riemann curvature tensor. This is the index form. A variation $V$ for which ${\delta }^{2}S[V]<0$ witnesses that the geodesic is not a minimum of the action.

A Jacobi field along $\gamma$ is a vector field $J(t)$ satisfying the Jacobi equation

$$\frac{D^2{J}^i}{d{t}^2}+R^i{}_{jkl}{\dot{\gamma }}^k{\dot{\gamma }}^l{J}^j=0,$$

where $D/dt$ is the covariant derivative along $\gamma$. Jacobi fields are variations through geodesics: if $\Gamma (s,t)$ is a family of geodesics with $\Gamma (0,t)=\gamma (t)$, then $J(t)=\partial \Gamma /\partial s{|}_{s=0}$ is a Jacobi field. Conversely, every Jacobi field arises this way.

Two points $q_1=\gamma (t_1)$ and $q_2=\gamma (t_2)$ on a geodesic are conjugate if there exists a non-zero Jacobi field $J$ along $\gamma$ with $J(t_1)=0$ and $J(t_2)=0$. Geometrically, conjugate points are where the exponential map ${\exp }_{q_1}$ fails to be a local diffeomorphism: $\gamma (t_2)$ is conjugate to $q_1$ along $\gamma$ if and only if $d({\exp }_{q_1}{)}_{t_2\dot{\gamma }(0)}$ has non-maximal rank.

Theorem (Morse index theorem for geodesics). The index of the quadratic form ${\delta }^{2}S$ on the space of variation fields vanishing at the endpoints — the number of independent negative-directions — equals the number of conjugate points along $\gamma$ in $(t_1,t_2)$, counted with multiplicity.

The Morse index theorem identifies the analytical property (index of the second variation) with the geometric property (count of conjugate points). A geodesic segment free of interior conjugate points locally minimises the energy functional; the appearance of the first conjugate point marks the transition from minimum to saddle.

Example. On the sphere $S^2$ with its round metric, any geodesic (great circle) starting at the north pole has its first conjugate point at the south pole, at distance $\pi$ along the geodesic. For paths shorter than $\pi$, the great-circle arc is the unique energy minimiser. For paths longer than $\pi$, it is a saddle point: one can shorten by "popping" the geodesic off the conjugate point. This is the geometric reason why airlines fly great-circle routes between nearby cities but not through antipodal points, where multiple shortest routes coexist.

Example. For the harmonic oscillator $L=\frac{1}{2}m{\dot{x}}^2-\frac{1}{2}k{x}^2$ with angular frequency $\omega =\sqrt{k/m}$, a path of duration greater than $T/2=\pi /\omega$ has conjugate points. The action changes from a minimum to a saddle at $t_2-t_1=T/2$, even though the path remains the unique physical trajectory. This confirms the "stationary action" terminology: the physical path need not minimise $S$.

The Hamilton-Jacobi equation and the optical analogy [Master]

The action $S$, regarded as a function of the upper endpoint $q_2$ and upper time limit $t_2$ (with the lower endpoint $q_1$ and lower time limit $t_1$ fixed), satisfies a first-order nonlinear PDE — the Hamilton-Jacobi equation. Write $S(q,t;q_1,t_1)$ for the action evaluated along the physical path from $(q_1,t_1)$ to $(q,t)$. By the chain rule and the Euler-Lagrange equations:

$$\frac{\partial S}{\partial t}+H\left(q,\frac{\partial S}{\partial q}\right)=0,$$

where $H(q,p)=p_i{\dot{q}}^i-L$ is the Hamiltonian and $p_i=\partial S/\partial q^i$ are the momenta. This is the Hamilton-Jacobi equation. It states that the time-dependent action function $S$ is a solution of a first-order PDE whose characteristic equations are Hamilton's equations of motion.

Derivation. Consider a solution path $\gamma$ from $(q_1,t_1)$ to $(q_2,t_2)$. The action along $\gamma$ is $S(q_2,t_2)={\int }_{t_1}^{t_2}Ldt$. Vary the upper endpoint $q_2\to q_2+\delta q$, $t_2\to t_2+\delta t$. The first variation gives (without the endpoint-vanishing condition):

$$\delta S=\frac{\partial L}{\partial {\dot{q}}^i}{|}_{t_2}\delta q^i+\left(L-\frac{\partial L}{\partial {\dot{q}}^i}{\dot{q}}^i\right){|}_{t_2}\delta t=p_i(t_2)\delta q^i-H(t_2)\delta t.$$

So $\partial S/\partial q^i=p_i$ and $\partial S/\partial t=-H$, giving the Hamilton-Jacobi equation immediately.

A complete integral of the Hamilton-Jacobi equation is a solution $S(q,t;\alpha )$ depending on $n$ constants of integration ${\alpha }_1,\dots ,{\alpha }_n$ (one for each degree of freedom) such that $\det ({\partial }^{2}S/\partial q^i\partial {\alpha }_j)\ne 0$. By the method of characteristics, a complete integral determines the full solution of the mechanical system: the trajectory is recovered from $\partial S/\partial {\alpha }_i={\beta }_i$ (constant), $p_i=\partial S/\partial q^i$, and the physical meaning of the constants ${\alpha }_i,{\beta }_i$ depends on the choice of coordinates and the symmetries of the system.

The optical analogy is exact. The eikonal equation of geometric optics is $(\nabla \varphi {)}^2=n^2(x)$, where $\varphi$ is the optical path length and $n(x)$ the refractive index. Light rays are the characteristics of the eikonal equation, just as mechanical trajectories are the characteristics of the Hamilton-Jacobi equation. Hamilton discovered this analogy in 1834 ^{[Hamilton 1834]}, and it directly inspired de Broglie's matter waves (1924) and Schrodinger's wave equation (1926), which was found by analogy with the Hamilton-Jacobi equation.

Example: the Kepler problem. For a particle of mass $m$ in a central potential $V=-k/r$, the Hamilton-Jacobi equation separates in spherical coordinates $(r,\theta ,\phi )$. The constants of integration are the energy $E$, the total angular momentum $L$, and the $z$-component of angular momentum $L_z$. The separation yields $S=S_r(r)+S_{\theta }(\theta )+S_{\phi }(\phi )-Et$, with each function determined by an ODE. The radial equation $S_r(r)=\int \sqrt{2m(E+k/r)-L^2/r^2}dr$ gives the turning points and the orbit shape, recovering the elliptic, parabolic, and hyperbolic orbits from the single variational principle.

Feynman's path integral and the classical limit [Master]

Feynman's 1948 formulation replaces the classical statement "the physical path has stationary action" with the quantum-mechanical statement "the probability amplitude for a particle to go from $q_1$ at $t_1$ to $q_2$ at $t_2$ is a sum over all paths, weighted by $e^{iS[\gamma ]/\hslash }$" ^{[Feynman 1948]}:

$$⟨q_2,t_2|q_1,t_1⟩={\int }_{\mathcal{P}}{e}^{iS[\gamma ]/\hslash }\mathcal{D}\gamma .$$

The integral is over all continuous paths $\gamma :[t_1,t_2]\to Q$ with $\gamma (t_1)=q_1$ and $\gamma (t_2)=q_2$, and $\mathcal{D}\gamma$ is a formal path-space measure. The rigorous construction of $\mathcal{D}\gamma$ requires either the Wiener measure (Euclidean time) or the Fresnel integral (real time); for this discussion the formal expression suffices.

In the classical limit $S\gg \hslash$, the phase $e^{iS/\hslash }$ oscillates rapidly for paths where $S$ is not stationary. These paths contribute nearly zero by destructive interference. Near the stationary-action path ${\gamma }_0$, expand $S[\gamma ]=S[{\gamma }_0]+\frac{1}{2}{\delta }^{2}S[{\gamma }_0](\gamma -{\gamma }_0)+\cdots$. The stationary-phase approximation gives

$$⟨q_2,t_2|q_1,t_1⟩\approx {\left(\frac{1}{2\pi i\hslash }\right)}^{n/2}\sqrt{\left|\det \frac{{\partial }^{2}S}{\partial q_2\partial q_1}\right|}{e}^{iS[{\gamma }_0]/\hslash -i\nu \pi /2},$$

where $\nu$ is the Morse index of ${\gamma }_0$ (the number of negative eigenvalues of ${\delta }^{2}S$, equal to the number of conjugate points along ${\gamma }_0$ by the Morse index theorem). The prefactor $\det ({\partial }^{2}S/\partial q_2\partial q_1)$ is the van Vleck determinant, named after van Vleck's 1928 work on correspondence principles ^{[VanVleck 1928]}.

The van Vleck determinant has a direct physical meaning. For the propagator $⟨q_2|q_1⟩$, the probability density $|⟨q_2|q_1⟩{|}^2$ is proportional to $|\det ({\partial }^{2}S/\partial q_2\partial q_1)|$. This determinant measures the density of classical trajectories in configuration space — the rate at which nearby initial conditions spread. The classical limit of the quantum propagator is the square root of this trajectory density, multiplied by the classical phase $e^{iS/\hslash }$.

The path integral provides a conceptual resolution of the "teleology" of the action principle. The classical action principle says "nature selects the stationary-action path." The path integral says "all paths contribute, and the stationary-action path dominates because of phase coherence." The action principle is not a teleological selection mechanism; it is a consequence of wave interference in the limit $\hslash \to 0$.

The connection to the Hamilton-Jacobi equation appears in the WKB approximation. Writing the wavefunction as $\psi (q,t)=A(q,t)e^{iS(q,t)/\hslash }$ and substituting into the Schrodinger equation, the leading order in $\hslash$ recovers the Hamilton-Jacobi equation for $S$, and the next order gives a transport equation for the amplitude $A$ related to the van Vleck determinant. The WKB hierarchy generalises to arbitrary order in $\hslash$ and is the semiclassical bridge between the path integral and the Hamilton-Jacobi theory 09.05.01 pending.

Field-theoretic extension of the action principle [Master]

The action principle extends from particle paths to field configurations. Replace the path $\gamma :[t_1,t_2]\to Q$ by a field $\varphi :\mathcal{M}\to {\mathbb{R}}^N$ on spacetime $\mathcal{M}$ (a Lorentzian manifold of dimension $d$). The Lagrangian density $\mathcal{L}(\varphi ,{\partial }_{\mu }\varphi ,x)$ is a function of the field and its first derivatives. The action functional is

$$S[\varphi ]={\int }_{\mathcal{M}}\mathcal{L}(\varphi ,{\partial }_{\mu }\varphi ,x)d^{d}x.$$

The field-theoretic Euler-Lagrange equation is derived by the same variational argument: vary $\varphi \to \varphi +ϵ\delta \varphi$ with $\delta \varphi$ vanishing on $\partial \mathcal{M}$, integrate the derivative term by parts, and apply the fundamental lemma:

$$\boxed{\frac{\partial \mathcal{L}}{\partial {\varphi }^a}-{\partial }_{\mu }\left(\frac{\partial \mathcal{L}}{\partial ({\partial }_{\mu }{\varphi }^a)}\right)=0.}$$

The index $a=1,\dots ,N$ labels the field components and $\mu =0,1,\dots ,d-1$ labels spacetime directions. Summation over repeated indices is implied.

Example: the Klein-Gordon field. For a real scalar field $\varphi$ on Minkowski spacetime with Lagrangian density $\mathcal{L}=\frac{1}{2}({\partial }_{\mu }\varphi )({\partial }^{\mu }\varphi )-\frac{1}{2}{m}^2{\varphi }^2$, the Euler-Lagrange equation gives $(\square +m^2)\varphi =0$, where $\square ={\partial }_{\mu }{\partial }^{\mu }$ is the d'Alembertian. This is the Klein-Gordon equation, the relativistic wave equation for a free scalar particle of mass $m$.

Example: Maxwell's equations from the electromagnetic action. The electromagnetic field is described by the potential 1-form $A_{\mu }$ with field strength $F_{\mu \nu }={\partial }_{\mu }{A}_{\nu }-{\partial }_{\nu }{A}_{\mu }$. The Lagrangian density is

$${\mathcal{L}}_{EM}=-\frac{1}{4}{F}_{\mu \nu }{F}^{\mu \nu }-J^{\mu }{A}_{\mu },$$

where $J^{\mu }$ is the current 4-vector. Varying with respect to $A_{\mu }$ gives ${\partial }_{\mu }{F}^{\mu \nu }=J^{\nu }$ — the inhomogeneous Maxwell equations. The homogeneous equations ${\partial }_{[\lambda }{F}_{\mu \nu ]}=0$ are satisfied automatically by the definition $F=dA$. This Lagrangian formulation of electromagnetism is the starting point for gauge field theory 10.09.01 pending and the Standard Model.

The stress-energy tensor arises from the field-theoretic action via Noether's theorem applied to spacetime translations:

$$T^{\mu \nu }=\frac{\partial \mathcal{L}}{\partial ({\partial }_{\mu }{\varphi }^a)}{\partial }^{\nu }{\varphi }^a-g^{\mu \nu }\mathcal{L},$$

where $g^{\mu \nu }$ is the inverse metric. Conservation ${\partial }_{\mu }{T}^{\mu \nu }=0$ follows from translation invariance of $\mathcal{L}$. This tensor is the source term in Einstein's field equations $G_{\mu \nu }=8\pi G{T}_{\mu \nu }$ 13.02.01, making the action principle the bridge from Lagrangian field theory to general relativity.

Synthesis. The action principle is the foundational reason that classical mechanics, optics, quantum mechanics, and field theory share a common variational architecture. The central insight is that the single functional $S[\gamma ]=\int Ldt$ simultaneously encodes the equations of motion (via $\delta S=0$), the symplectic structure (via the Poincare-Cartan form ${\theta }_L$), the Hamiltonian (via the energy function $E$), the quantum propagator (via the path integral $e^{iS/\hslash }$), and the conserved quantities (via Noether's theorem applied to symmetries of $L$). Putting these together, the Euler-Lagrange equations on $TQ$, the Hamilton-Jacobi PDE on $T^{\ast }Q$, the path integral on the space of curves, and the field-theoretic Euler-Lagrange equations on the space of field configurations are all manifestations of the same variational condition.

This is exactly the structure that identifies Lagrangian and Hamiltonian mechanics: the bridge is the Legendre transform $TQ\to T^{\ast }Q$, and the pattern generalises from finite-dimensional configuration spaces to infinite-dimensional field spaces, from classical mechanics to quantum field theory, and from particle paths to spacetime histories. The action principle appears again in 09.03.01 pending as the input to Noether's first theorem, builds toward 09.04.01 pending as the Legendre convert converts the variational problem into symplectic dynamics, and the pattern recurs in 12.10.01 pending where the classical action weights the Feynman path integral.

Full proof set [Master]

Proposition 1 (The Poincare-Cartan form encodes the Euler-Lagrange equations). Let $L$ be a regular Lagrangian on $TQ$ with Poincare-Cartan form ${\theta }_L=p_{i}d{q}^i-Ldt$, where $p_i=\partial L/\partial {\dot{q}}^i$. A curve $\gamma :[t_1,t_2]\to Q$ satisfies the Euler-Lagrange equations if and only if the exterior derivative $d{\theta }_L$ vanishes on the lifted curve $\dot{\gamma }$, in the sense that $i_{{\dot{\gamma }}^{\prime }}d{\theta }_L=0$ when restricted to variations vanishing at the endpoints.

Proof. In local coordinates, $d{\theta }_L=d{p}_i\wedge d{q}^i-dL\wedge dt$. Expand $dL=(\partial L/\partial q^i)d{q}^i+(\partial L/\partial {\dot{q}}^i)d{\dot{q}}^i+(\partial L/\partial t)dt$. Substituting:

$$d{\theta }_L=d{p}_i\wedge d{q}^i-\frac{\partial L}{\partial q^i}d{q}^i\wedge dt-p_{i}d{\dot{q}}^i\wedge dt.$$

Along the lifted curve $\dot{\gamma }$, the tangent vector is $X={\dot{q}}^i\partial /\partial q^i+{\ddot{q}}^i\partial /\partial {\dot{q}}^i+\partial /\partial t$. Contracting with $d{\theta }_L$:

$$i_{X}d{\theta }_L={\dot{q}}^{i}d{p}_i-{\dot{p}}_{i}d{q}^i-\frac{\partial L}{\partial q^i}{\dot{q}}^{i}dt+{\dot{q}}^i{p}_{i}dt+\text{boundary terms}.$$

Using $p_i=\partial L/\partial {\dot{q}}^i$ and simplifying, the condition $i_{X}d{\theta }_L=0$ restricted to variations gives ${\dot{p}}_i=\partial L/\partial q^i$, which are the Euler-Lagrange equations. $\square$

Proposition 2 (Conjugate points detect saddle behaviour). Let $\gamma :[t_1,t_2]\to Q$ be a geodesic on a Riemannian manifold. If $\gamma$ has no conjugate points in $(t_1,t_2)$, then $\gamma$ is a local minimum of the energy functional among curves with the same endpoints. If $\gamma$ has a conjugate point at some $t_ \in (t_1, t_2)$,thenthereexistsavariation$V$with$\delta^2 S[V] < 0$,and$\gamma$ is a saddle point.*

Proof. The second variation for the geodesic energy is the index form $I(V,V)={\int }_{t_1}^{t_2}(|DV/dt{|}^2-R(\dot{\gamma },V,\dot{\gamma },V))dt$. If there are no conjugate points in $(t_1,t_2)$, then the Jacobi fields vanishing at $t_1$ are all non-zero in the open interval, and the index form is positive definite on variation fields vanishing at both endpoints. This is the Jacobi field criterion for a minimum, established by comparing an arbitrary variation field $V$ with the Jacobi field $J$ sharing the same boundary data: integration by parts shows $I(V,V)=I(V-J,V-J)+I(J,J)$, and both terms are non-negative when no conjugate points are present.

If $t_{\ast }$ is conjugate to $t_1$, then there exists a Jacobi field $J$ with $J(t_1)=J(t_{\ast })=0$ and $J\not\equiv 0$. Extend $J$ by zero for $t>t_{\ast }$ and construct a variation supported in a small neighbourhood of $t_{\ast }$. The second variation of this piecewise-smooth variation is negative, because the curvature term dominates near the conjugate point where $J$ is small but $R(\dot{\gamma },J,\dot{\gamma },J)$ is non-vanishing. $\square$

Proposition 3 (Gauge invariance of the action). Two Lagrangians $L$ and $L^{\prime }=L+dF(q,t)/dt$ that differ by a total time derivative yield identical Euler-Lagrange equations and the same stationary paths.

Proof. The action difference is $\Delta S={\int }_{t_1}^{t_2}dF/dtdt=F(q(t_2),t_2)-F(q(t_1),t_1)$. This depends only on the fixed endpoints $(q(t_1),t_1)$ and $(q(t_2),t_2)$, not on the path between them. For any variation $\delta \gamma$ vanishing at the endpoints, $\delta (\Delta S)=0$. So $\delta S^{\prime }=\delta S+\delta (\Delta S)=\delta S$, and the stationary conditions are identical. $\square$

Connections [Master]

• Newton's laws 09.01.02 pending. The action principle recovers Newton's second law $F=ma$ in Cartesian coordinates with the standard Lagrangian $L=T-V$; the Euler-Lagrange equations are the coordinate-independent generalisation of this recovery.

• Conservation laws 09.01.03. Energy and momentum conservation are re-derived from the action principle via the Beltrami identity (energy conservation from time-translation invariance) and cyclic coordinates (momentum conservation from translation invariance in a coordinate), both special cases of Noether's theorem.

• Euler-Lagrange equations 09.02.02 pending. The differential-equation form of the stationary-action condition derived in this unit is developed in full computational detail there, including constrained systems and non-Cartesian coordinates.

• Noether's theorem 09.03.01 pending. Generalises the Beltrami identity and cyclic-coordinate observations: every continuous symmetry of $L$ produces a conserved quantity. The action principle is the input, and Noether's first theorem is the output.

• Legendre transform and Hamiltonian mechanics 09.04.01 pending. Converts the Lagrangian $L(q,\dot{q})$ to the Hamiltonian $H(q,p)$, mapping the action principle on $TQ$ to Hamilton's equations on $T^{\ast }Q$. The Poincare-Cartan form ${\theta }_L$ on $TQ$ maps to the canonical symplectic form ${\omega }_0$ on $T^{\ast }Q$.

• Hamilton-Jacobi theory 09.05.01 pending. The action function $S(q,t)$ satisfies the Hamilton-Jacobi PDE, whose characteristics are the physical trajectories. This is the PDE counterpart to the variational formulation.

• Classical field theory and electromagnetism 10.09.01 pending. Lifts the action principle from particle paths to field configurations, yielding Maxwell's equations from the EM Lagrangian ${\mathcal{L}}_{EM}=-\frac{1}{4}{F}_{\mu \nu }{F}^{\mu \nu }$.

• Feynman path integrals 12.10.01 pending. Replaces the classical stationary-action selection with a quantum sum over all paths weighted by $e^{iS/\hslash }$; the classical limit $\hslash \to 0$ recovers the action principle by stationary-phase dominance.

• General relativity 13.02.01. The Einstein-Hilbert action $S_{EH}=\int R\sqrt{-g}{d}^{4}x$ applies the variational principle to the spacetime metric itself; the field-theoretic Euler-Lagrange equation for the metric is Einstein's equation $G_{\mu \nu }=8\pi G{T}_{\mu \nu }$.

Historical & philosophical context [Master]

The action principle originated not in mechanics but in optics. Fermat (1662) proposed that light follows the path of least time between two points (Fermat's principle), explaining refraction ^{[Fermat 1662]}. Maupertuis (1744) conjectured a similar principle for mechanics — "nature acts by the shortest path" — but his formulation was vague and his priority dispute with Konig was acrimonious ^{[Maupertuis 1744]}.

Euler, responding to Maupertuis, formalised the variational calculus in Methodus inveniendi lineas curvas maximi minimive proprietate gaudentes (1744), deriving what are now called the Euler equations for variational problems ^{[Euler 1744]}. Lagrange, in Mecanique analytique (1788), reformulated all of mechanics as the consequence of a single variational principle — the principle of virtual work extended to dynamics via d'Alembert's principle — and derived the general equations of motion (the Euler-Lagrange equations in modern terminology) without drawing a single diagram ^{[Lagrange 1788]}.

Hamilton, in his two papers on dynamics (1834, 1835), unified the optical and mechanical variational principles by introducing the action $S$ as a function of both endpoints and time, deriving the Hamilton-Jacobi equation, and recognising the analogy with the eikonal equation of geometric optics ^{[Hamilton 1834]}. Jacobi (1843) developed the theory of complete integrals and the variational approach to canonical transformations. The optical-mechanical analogy was the direct inspiration for de Broglie's matter waves (1924) and Schrodinger's wave equation (1926), which was discovered by analogy with the Hamilton-Jacobi equation.

The philosophical content of the action principle is distinctive. Unlike Newton's laws, which are local (force at a point determines acceleration at that point), the action principle is global (the entire path is selected by a condition on the whole trajectory). This teleological character — nature "knows" the endpoint and "chooses" accordingly — troubled Maupertuis's contemporaries and remains a subject of philosophical analysis 20.07.01 pending. The resolution in quantum mechanics is that the action principle is an approximation: all paths contribute, and the stationary-action path dominates because of phase coherence, not because nature "chooses" it.

Morse (1934) extended the second-variation analysis to infinite-dimensional path spaces, founding Morse theory — the study of the relationship between the critical points of a function on a manifold and the topology of the manifold itself. The Morse index theorem for geodesics, stated in the second-variation section above, is the prototype for the general theory, which appears again in 05.01.01 (Morse theory on loop spaces).

Bibliography [Master]

@incollection{Fermat1662,
  author = {Fermat, Pierre de},
  title = {Synthesis ad refractiones},
  year = {1662},
  note = {Posthumous; collected in {\it Varia opera mathematica}, 1679}
}

@article{Maupertuis1744,
  author = {Maupertuis, Pierre-Louis},
  title = {Accord de diff\'erentes loix de la nature qui avaient jusqu'ici paru incompatibles},
  journal = {M\'emoires de l'Acad\'emie Royale des Sciences de Paris},
  year = {1744},
  pages = {417--426}
}

@book{Euler1744,
  author = {Euler, Leonhard},
  title = {Methodus inveniendi lineas curvas maximi minimive proprietate gaudentes},
  publisher = {Bousquet},
  address = {Lausanne},
  year = {1744}
}

@book{Lagrange1788,
  author = {Lagrange, Joseph-Louis},
  title = {M\'ecanique analytique},
  publisher = {Desaint},
  address = {Paris},
  year = {1788}
}

@article{Hamilton1834,
  author = {Hamilton, William Rowan},
  title = {On a General Method in Dynamics},
  journal = {Philosophical Transactions of the Royal Society},
  volume = {124},
  pages = {247--308},
  year = {1834}
}

@book{Arnold1989,
  author = {Arnold, Vladimir I.},
  title = {Mathematical Methods of Classical Mechanics},
  edition = {2nd},
  publisher = {Springer},
  series = {Graduate Texts in Mathematics},
  number = {60},
  year = {1989}
}

@book{LandauLifshitz1976,
  author = {Landau, Lev D. and Lifshitz, Evgeny M.},
  title = {Mechanics},
  edition = {3rd},
  publisher = {Pergamon},
  series = {Course of Theoretical Physics},
  volume = {1},
  year = {1976}
}

@book{FeynmanHibbs1965,
  author = {Feynman, Richard P. and Hibbs, Albert R.},
  title = {Quantum Mechanics and Path Integrals},
  publisher = {McGraw-Hill},
  year = {1965}
}

@book{Taylor2005,
  author = {Taylor, John R.},
  title = {Classical Mechanics},
  publisher = {University Science Books},
  year = {2005}
}

@book{SusskindHrabovsky2014,
  author = {Susskind, Leonard and Hrabovsky, George},
  title = {The Theoretical Minimum: Classical Mechanics},
  publisher = {Basic Books},
  year = {2014}
}

@book{Goldstein2002,
  author = {Goldstein, Herbert and Poole, Charles P. and Safko, John},
  title = {Classical Mechanics},
  edition = {3rd},
  publisher = {Pearson},
  year = {2002}
}

@book{TongDynamics,
  author = {Tong, David},
  title = {Classical Dynamics},
  series = {DAMTP Cambridge Lecture Notes},
  year = {2005},
  note = {\S2}
}

@article{VanVleck1928,
  author = {Van Vleck, John H.},
  title = {The Correspondence Principle in the Statistical Interpretation of Quantum Mechanics},
  journal = {Proceedings of the National Academy of Sciences},
  volume = {14},
  pages = {178--188},
  year = {1928}
}

@article{Feynman1948,
  author = {Feynman, Richard P.},
  title = {Space-Time Approach to Non-Relativistic Quantum Mechanics},
  journal = {Reviews of Modern Physics},
  volume = {20},
  pages = {367--387},
  year = {1948}
}

@book{Morse1934,
  author = {Morse, Marston},
  title = {The Calculus of Variations in the Large},
  publisher = {American Mathematical Society},
  series = {Colloquium Publications},
  volume = {18},
  year = {1934}
}

@book{Jacobi1843,
  author = {Jacobi, Carl G. J.},
  title = {Vorlesungen \"uber Dynamik},
  year = {1843},
  note = {Published posthumously, 1866}
}