Stationary Phase and the Classical Limit

The path integral hands you a scandal: to compute how a particle gets from here to there, you are told to add up a contribution from every conceivable trajectory — the straight one, the loop-the-loop, the one that darts to Andromeda and back — each carrying a phase $e^{iS[x]/\hbar}$ of exactly unit modulus. Nothing is bigger than anything else. And yet a thrown ball follows one sharp parabola, not a democratic blur. How does the single classical history emerge from this infinity of equals? The answer is interference, made quantitative by one theorem: the method of stationary phase. Because the action of a macroscopic process is enormous in units of $\hbar$ — a gram-scale bead has $S/\hbar\sim 10^{27}$ — the phase $S/\hbar$ whirls through billions of turns under the tiniest deformation of the path. Neighbouring trajectories therefore cancel in pairs, and the sum survives only where the phase momentarily stops changing: where $\delta S = 0$ . That stationarity condition is the Euler–Lagrange equation. The classical path is not selected by a law added on top of quantum mechanics; it is the saddle point of the quantum sum. You arrive holding The Feynman Path Integral Formulation (the construction of $\int\mathcal{D}x\,e^{iS/\hbar}$ ) and its exact Gaussian evaluations in Free-Particle and Harmonic-Oscillator Path Integrals. Here we prove the general stationary-phase asymptotics against the exact Fresnel integral, watch $\delta S=0$ deliver Newton's law, show that the surviving classical phase $S_{\mathrm{cl}}(x,t)$ obeys the Hamilton–Jacobi equation, and read off the width $\sim\sqrt{\hbar}$ of the surviving tube of paths. This is the analytic engine of the classical limit — and, as The WKB Approximation and Bohr-Sommerfeld and Maslov Quantization show, of the entire semiclassical toolkit.

Key takeaways

The propagator is a sum over paths weighted by $e^{iS[x]/\hbar}$ ; as $\hbar\to0$ the phase oscillates so violently that all paths cancel except where $S$ is stationary, $\delta S=0$ .
Stationarity of the action is exactly the Euler–Lagrange equation, so the classical trajectory emerges as the saddle point of the quantum path integral — no extra postulate is needed.
The stationary-phase theorem $\int g\,e^{iS/\hbar}\to g(x_*)\sqrt{2\pi i\hbar/S''(x_*)}\,e^{iS(x_*)/\hbar}$ is verified against the exact Fresnel integral $\int e^{iax^2}\mathrm dx=\sqrt{\pi/a}\,e^{i\pi/4}$ — the $\pi/4$ is a real, computable number.
The surviving classical phase $S_{\mathrm{cl}}(x,t)$ satisfies the Hamilton–Jacobi equation $\partial_tS+H(x,\partial_xS)=0$ ; the free case $S_{\mathrm{cl}}=m(\Delta x)^2/2T$ solves it identically.
The tube of contributing paths has width $\sim\sqrt{\hbar}$ , and the leading quantum correction to the saddle result is $O(\hbar)$ — a controlled expansion, not a switch.
The same machinery is WKB, the Van Vleck prefactor, Maslov phases, and the Gutzwiller trace formula of quantum chaos.
Key idea: classical mechanics is the stationary-phase saddle point of the quantum path integral — as $\hbar\to0$ only the trajectory with $\delta S=0$ survives, and its phase obeys Hamilton–Jacobi.

The core insight

As $\hbar\to0$ the oscillatory weight $e^{iS/\hbar}$ makes every path cancel except where the action is stationary; that stationary path is the classical trajectory, its phase satisfies the Hamilton–Jacobi equation, and Gaussian fluctuations about it give the semiclassical propagator.

Where this lives today

Stationary phase is the beating heart of semiclassical physics, which is anything but a historical relic. The Gutzwiller trace formula — the density of quantum states written as a sum over the classical periodic orbits of a chaotic system, each weighted by a stability (Van Vleck) amplitude and a Maslov phase — is the foundation of quantum chaos and underlies the statistics of energy levels in quantum dots, microwave billiards, and the Riemann zeros. In molecular and chemical physics, semiclassical initial-value representations (Herman–Kluk propagators) evaluate the path integral by exactly this saddle-point machinery to simulate reaction dynamics too big for a full grid. And every time an attosecond-physics calculation invokes the "three-step model" of high-harmonic generation, it is computing a stationary-phase saddle of an electron's action in a laser field. The generalization to non-separable tori and chaotic orbits — EBK quantization and the Gutzwiller trace formula — is the natural next step this article points toward but does not prove.

The physical picture

Give every path a little clock hand of fixed length one, pointing at the angle $S[x]/\hbar$ . To find the amplitude you lay all these arrows tip-to-tail and read off the resultant. For a randomly chosen region of path space the angles are scattered uniformly around the circle, and the arrows curl up into a tight, useless spiral that sums to almost nothing. The exception is the neighbourhood of a path where $S$ barely changes as you wiggle the trajectory — there the arrows all point nearly the same way and add coherently. That special neighbourhood is the classical path, and "stationary phase" is just the statement that the arrows only line up where the clock hands stop spinning.

Figure

The clock-hand picture: only near the stationary path do the phasors align

Macroscopic actions are astronomically large in units of $\hbar$ . A one-gram bead moving one centimetre in one second has $S\approx\tfrac12 mv^2T\approx 5\times10^{-8}\,\mathrm{J\,s}$ , so $S/\hbar\approx 5\times10^{26}$ . A path deformation far too small to see changes the phase by billions of radians, which is why a thrown ball has a razor-sharp trajectory with no measurable quantum spread. For an electron in an atom, by contrast, $S/\hbar\sim 1$ : the phase barely turns, no single path dominates, and the full quantum blur survives. The whole distinction between "classical" and "quantum" is the single dimensionless number $S/\hbar$ . The width of the tube of paths that actually contribute scales like $\sqrt{\hbar}$ . Because the action is stationary at $x_{\mathrm{cl}}$ , it varies quadratically under a fluctuation $\eta$ : $S\approx S_{\mathrm{cl}}+\tfrac12 S''\eta^2$ . The phase stays coherent — within about one radian — for $\eta\lesssim\sqrt{\hbar/S''}$ . Quantum fuzziness is nothing but the residual width of this surviving tube, and it closes like $\sqrt{\hbar}$ as we approach the classical world.

Figure

As ℏ shrinks, the surviving tube of paths collapses onto the classical trajectory

The core equations

The starting point is the amplitude The Feynman Path Integral Formulation constructs by time-slicing: the propagator as a coherent sum over paths, each weighted by the exponential of its classical action. We recall it here as the object we are about to evaluate asymptotically — the construction and its properties are owned there; here we take the $\hbar\to0$ limit.

K(x_f,T;x_i,0)=\int_{x(0)=x_i}^{x(T)=x_f}\mathcal{D}x(t)\;\exp\!\Big(\frac{i}{\hbar}\,S[x]\Big),\qquad S[x]=\int_0^T L\,\mathrm{d}t \qquad (1)

where: $K$ — the transition amplitude (propagator); $\mathcal{D}x$ — the functional measure over all paths with the fixed endpoints; $S[x]$ — the classical action; $L(x,\dot x,t)$ — the Lagrangian. Reading it physically: every path contributes with the same magnitude $|e^{iS/\hbar}|=1$ and differs only in phase, so the entire content of the classical limit is in how those phases interfere. The engine that controls that interference is the stationary-phase theorem for oscillatory integrals. For an ordinary one-dimensional integral it reads as follows, and this is the statement we own and verify.

\int g(x)\,e^{iS(x)/\hbar}\,\mathrm{d}x\;\xrightarrow{\;\hbar\to0\;}\;\sum_{x_*}g(x_*)\sqrt{\frac{2\pi\hbar}{|S''(x_*)|}}\;e^{\,iS(x_*)/\hbar\,+\,i\,\mathrm{sgn}\!\big(S''(x_*)\big)\pi/4} \qquad (2)

where: $x_*$ — a stationary point, $S'(x_*)=0$ ; $S''$ — the phase curvature there; $\mathrm{sgn}(S'')=\pm1$ fixes the celebrated $\pm\pi/4$ phase. Reading it physically: the integral forgets everything except the immediate neighbourhood of each stationary point, where the amplitude $g$ is sampled and the curvature sets both the width $\sqrt{\hbar/|S''|}$ and the Fresnel phase. That $\pi/4$ is not a fudge — it is the exact phase of the Fresnel integral, computed below. Applied to the path integral, the "stationary point" is a whole trajectory, and $S'(x_*)=0$ becomes the vanishing of the functional derivative — the classical equation of motion.

\frac{\delta S}{\delta x(t)}=0\;\Longrightarrow\;\frac{\mathrm{d}}{\mathrm{d}t}\frac{\partial L}{\partial\dot x}-\frac{\partial L}{\partial x}=0 \qquad (3)

where: $\delta S/\delta x$ — the variational derivative of the action; the right-hand side is the Euler–Lagrange equation. Reading it physically: the path that survives the quantum sum is precisely the one obeying Newton's law in Lagrangian dress. Hamilton's principle of stationary action is not an independent axiom — it is the saddle-point condition of $\int\mathcal{D}x\,e^{iS/\hbar}$ . Expanding the action about that classical path leaves the classical phase out front and a Gaussian integral over fluctuations, which produces the semiclassical propagator. The prefactor — owned in Free-Particle and Harmonic-Oscillator Path Integrals — is the Van Vleck determinant.

K_{\mathrm{sc}}=\Big(\frac{1}{2\pi i\hbar}\Big)^{\!n/2}\Big|\det\!\Big(-\frac{\partial^2 S_{\mathrm{cl}}}{\partial x_i\,\partial x_f}\Big)\Big|^{1/2}\exp\!\Big(\frac{i}{\hbar}S_{\mathrm{cl}}-\frac{i\pi\nu}{2}\Big) \qquad (4)

where: $n$ — spatial dimension; $\partial^2 S_{\mathrm{cl}}/\partial x_i\partial x_f$ — the Van Vleck matrix of mixed second derivatives of the classical action; $\nu$ — the Maslov index counting caustics. Reading it physically: the amplitude at the detector is large where classical trajectories focus (the Van Vleck determinant is big) and each focusing event stamps a $-\pi/2$ Maslov phase, a bookkeeping owned by Bohr-Sommerfeld and Maslov Quantization. Finally, the classical phase $S_{\mathrm{cl}}(x,t)$ is not an arbitrary function: substituting the semiclassical ansatz into the Schrödinger equation shows it satisfies the Hamilton–Jacobi equation at leading order, with a transport equation at next order.

\psi=A(x,t)\,e^{iW(x,t)/\hbar}:\qquad \underbrace{\partial_t W+H\!\Big(x,\partial_x W\Big)=0}_{\mathcal O(\hbar^0)\ \text{Hamilton–Jacobi}},\qquad \underbrace{\partial_t(A^2)+\partial_x\!\big(A^2\,\partial_x W/m\big)=0}_{\mathcal O(\hbar^1)\ \text{transport}} \qquad (5)

where: $W$ — Hamilton's principal function (the leading phase, identical to $S_{\mathrm{cl}}$ ); $A$ — a slowly varying amplitude; $H=p^2/2m+V$ . Reading it physically: the wave's phase evolves by classical mechanics (Hamilton–Jacobi) and its probability $A^2$ is simply carried along the classical trajectories (continuity). This is the precise sense in which quantum mechanics is wave optics whose geometric-optics limit is Newtonian dynamics.

Derivation AThe stationary-phase theorem and its Fresnel phase

Our goal is to establish equation (2) — that an oscillatory integral collapses, as $\hbar\to0$ , onto its stationary points with the Gaussian width $\sqrt{2\pi\hbar/|S''|}$ and the signature phase $e^{i\,\mathrm{sgn}(S'')\pi/4}$ . The strategy is to expand the phase to second order about a stationary point, do the resulting Fresnel integral in closed form, and pin the $\pi/4$ by comparison with the exact Gaussian-with-imaginary-coefficient integral. The one thing to watch is that the coefficient of $x^2$ in the phase is imaginary, so the "Gaussian" is a Fresnel integral whose convergence is by oscillation, not decay — we handle it by an $i\epsilon$ -style contour rotation.

Step 1. Near a stationary point $x_*$ where $S'(x_*)=0$ , Taylor-expand the phase, keeping the amplitude at leading order: $S(x)\approx S(x_*)+\tfrac12 S''(x_*)(x-x_*)^2$ and $g(x)\approx g(x_*)$ . The linear term is absent by construction — that is what "stationary" buys us. Why: the first-order term is exactly $S'(x_*)(x-x_*)$ , and it vanishes because $S'(x_*)=0$ ; away from a stationary point this term would survive and make the integrand oscillate to death.

Step 2. The remaining integral is the Fresnel integral $\int_{-\infty}^{\infty}e^{iau^2}\mathrm du$ with $a=S''(x_*)/(2\hbar)$ . Its exact value is $\sqrt{\pi/|a|}\,e^{i\,\mathrm{sgn}(a)\pi/4}$ , obtained by rotating the contour by $\pm\pi/4$ so the integrand becomes a genuine decaying Gaussian. Why: writing $\int e^{iau^2}\mathrm du=\sqrt{\pi/(-ia)}$ and taking the branch with a positive real part gives $\sqrt{\pi/|a|}$ times $e^{i\pi/4}$ for $a>0$ and $e^{-i\pi/4}$ for $a<0$ ; this is the origin of the $\pm\pi/4$ .

Step 3. Restore $a=S''/2\hbar$ : $\sqrt{\pi/|a|}=\sqrt{2\pi\hbar/|S''|}$ . Multiplying by the frozen amplitude $g(x_*)$ and the phase $e^{iS(x_*)/\hbar}$ reproduces one term of (2); summing over all stationary points gives the full formula. Why: each stationary point is an independent saddle, and for well-separated saddles their Gaussian neighbourhoods do not overlap, so the contributions simply add — with interference encoded in the different phases $e^{iS(x_*)/\hbar}$ . *Verified: the verifier evaluates $\int e^{iax^2}\mathrm dx$ symbolically for $a=0.7,3,12$ and matches $\sqrt{\pi/a}\,e^{i\pi/4}$ to $10^{-9}$ ; the $a<0$ case reproduces $e^{-i\pi/4}$ ; and $e^{i\pi/4}=(1+i)/\sqrt2$ is confirmed. Against a Gaussian-amplitude model the relative error of the leading term shrinks as $0.025\to0.005\to0.001$ for $\hbar=0.05\to0.01\to0.002$ , i.e. linearly in $\hbar$ . Standard: Bleistein & Handelsman, "Asymptotic Expansions of Integrals," Ch. 6.*

Go deeper · Derivation B · 2 minStationarity is Hamilton's principle, and the classical phase solves Hamilton–Jacobi

Derivation BStationarity is Hamilton's principle, and the classical phase solves Hamilton–Jacobi

The goal is twofold: first, that the surviving path obeys the Euler–Lagrange equation (so "classical" means "stationary phase"); second, that the classical action $S_{\mathrm{cl}}(x,t)$ , viewed as a function of its endpoint, satisfies the Hamilton–Jacobi equation. The route is to compute the change of phase between a path and its neighbour, note when it survives the $\hbar\to0$ limit, and then feed the semiclassical wavefunction into the Schrödinger equation and collect powers of $\hbar$ . The subtlety is keeping track of which object is a functional (the action along a path) and which is an ordinary function (the action at a fixed endpoint).

Step 1. The phase difference between a path $x$ and a neighbour $x+\delta x$ is $\Delta(S/\hbar)=\tfrac1\hbar\int\frac{\delta S}{\delta x(t)}\,\delta x(t)\,\mathrm dt$ . If $\delta S/\delta x\neq0$ this diverges like $1/\hbar$ , so neighbouring phasors point randomly and cancel; only if $\delta S/\delta x=0$ does the phase change at second order, $\mathcal O(\delta x^2/\hbar)$ , and the arrows stay aligned over a tube of width $\sim\sqrt\hbar$ . Why: as $\hbar\to0$ , any first-order phase difference sweeps through many multiples of $2\pi$ under an arbitrarily small $\delta x$ ; the only stationary configuration is $\delta S/\delta x=0$ , which is the Euler–Lagrange equation.

Step 2. For the free particle, solving Euler–Lagrange ( $\ddot x=0$ ) with endpoints $x(0)=x_i,\,x(T)=x_f$ gives the straight line $x_{\mathrm{cl}}(t)=x_i+(x_f-x_i)t/T$ , and integrating $L=\tfrac12 m\dot x^2$ along it yields $S_{\mathrm{cl}}=m(x_f-x_i)^2/2T$ . This is the classical action as a function of the endpoint $(x_f,T)$ . Why: constant velocity makes the integrand constant, so the time integral is trivial; the result depends only on the endpoint separation and the elapsed time.

Step 3. Now treat $S(x,t)\equiv m(x-x_i)^2/2t$ as an ordinary function of the endpoint and test the Hamilton–Jacobi equation $\partial_t S+(\partial_x S)^2/2m=0$ . Compute $\partial_x S=m(x-x_i)/t$ (which is exactly the classical momentum $p$ ) and $\partial_t S=-m(x-x_i)^2/2t^2$ ; their combination vanishes identically. Why: $\partial_t S+(\partial_x S)^2/2m=-\tfrac{m(x-x_i)^2}{2t^2}+\tfrac{1}{2m}\cdot\tfrac{m^2(x-x_i)^2}{t^2}=0$ — the free classical action is a solution of Hamilton–Jacobi, and $\partial_x S=p$ recovers the momentum–action relation.

Step 4. Substituting $\psi=Ae^{iW/\hbar}$ into $i\hbar\partial_t\psi=-\tfrac{\hbar^2}{2m}\partial_x^2\psi+V\psi$ and collecting powers of $\hbar$ gives, at $\mathcal O(\hbar^0)$ , the Hamilton–Jacobi equation $\partial_t W+(\partial_x W)^2/2m+V=0$ , and at $\mathcal O(\hbar^1)$ the transport equation $\partial_t(A^2)+\partial_x(A^2\partial_x W/m)=0$ . Why: the $\hbar^0$ term is purely the phase equation (no derivatives of $A$ ), identifying $W=S_{\mathrm{cl}}$ ; the $\hbar^1$ term is the imaginary part, a continuity equation carrying probability $A^2$ along the classical flow $v=\partial_x W/m$ . *Verified: the verifier confirms the free $S_{\mathrm{cl}}$ gives HJ residual exactly $0$ and $\partial_x S=m(x_f-x_i)/T$ ; it inserts $Ae^{iW/\hbar}$ into the Schrödinger equation symbolically and finds the $\mathcal O(\hbar^0)$ coefficient is $-1$ times the HJ combination $\partial_t W+(\partial_x W)^2/2m+V$ , with a nonzero transport term at $\mathcal O(\hbar^1)$ . Standard: Sakurai & Napolitano §2.6.*

Go deeper · Derivation C · 2 minThe harmonic oscillator: Van Vleck prefactor, the free limit, and caustics

Derivation CThe harmonic oscillator: Van Vleck prefactor, the free limit, and caustics

The goal is to exhibit the semiclassical propagator (4) in a nontrivial yet exactly Gaussian case, where stationary phase is not merely asymptotic but exact. We compute the oscillator classical action, read off its Van Vleck prefactor, check that it reduces to the free propagator as $\omega\to0$ , and locate the caustics at $\omega T=n\pi$ where the prefactor diverges. The point to watch is that for a quadratic action the fluctuation expansion terminates, so the saddle result is the full answer — this is why the free particle and oscillator are the two exact laboratories for the classical limit.

Step 1. Solving $\ddot x+\omega^2 x=0$ with the given endpoints and integrating $L=\tfrac12 m\dot x^2-\tfrac12 m\omega^2 x^2$ gives $S_{\mathrm{cl}}=\dfrac{m\omega}{2\sin\omega T}\big[(x_i^2+x_f^2)\cos\omega T-2x_ix_f\big]$ . Why: the classical solution is a sinusoid matched to both endpoints; the action integral of a quadratic Lagrangian along it produces the compact $\cos\omega T/\sin\omega T$ structure.

Step 2. The Van Vleck prefactor is $-\partial^2 S_{\mathrm{cl}}/\partial x_i\partial x_f = m\omega/\sin\omega T$ . This is the semiclassical amplitude, and being an exact Gaussian problem it is the complete prefactor, not just the leading term. Why: differentiating $S_{\mathrm{cl}}$ once in $x_i$ and once in $x_f$ picks out the $-2x_ix_f$ cross term, leaving $m\omega/\sin\omega T$ ; because the action is quadratic there are no higher fluctuation corrections.

Step 3. As $\omega\to0$ , $\sin\omega T\to\omega T$ and $\cos\omega T\to1$ , so $S_{\mathrm{cl}}\to m(x_f-x_i)^2/2T$ and the prefactor $\to m/T$ — exactly the free propagator. The oscillator smoothly contains the free particle. Why: the $\omega\to0$ limit is a consistency check that any oscillator formula must pass, and it does so term by term.

Step 4. The prefactor $\propto1/\sqrt{\sin\omega T}$ diverges whenever $\sin\omega T=0$ , i.e. $\omega T=n\pi$ . These are the caustics — instants at which every trajectory launched from $x_i$ refocuses. Each crossing advances the Maslov index by one, contributing the $-i\pi\nu/2$ phase of (4). Why: at $\omega T=n\pi$ all classical paths from a fixed start return to a common point regardless of initial momentum, so the density of classical arrivals is infinite and the semiclassical amplitude blows up; the finite exact kernel is recovered by the Maslov phase and, near the caustic, by an Airy-type uniform approximation. *Verified: the verifier confirms $-\partial^2 S_{\mathrm{cl}}/\partial x_i\partial x_f=m\omega/\sin\omega T$ symbolically, that $\lim_{\omega\to0}S_{\mathrm{cl}}=m(x_f-x_i)^2/2T$ exactly, and that $\sin(n\pi)=0$ for $n=1,2,3$ (the caustic condition). Standard: Feynman & Hibbs §3-5; Schulman Ch. 6.*

Limiting cases

case	result	meaning
$\hbar\to0$	$K\sim A\,e^{iS_{\mathrm{cl}}/\hbar}$	only the stationary (classical) path survives; Newtonian/Hamiltonian dynamics recovered from the propagator.
quadratic action	stationary phase is exact	free particle and harmonic oscillator: no $\hbar$ corrections, because fluctuations are exactly Gaussian.
several stationary paths	$K\sim\sum_n A_n\,e^{iS_n/\hbar-i\nu_n\pi/2}$	classical paths interfere; produces double-slit fringes, rainbow scattering, and the Gutzwiller periodic-orbit sum.
caustic / conjugate point	Van Vleck $\to\infty$	neighbouring classical paths focus; the naive Gaussian prefactor diverges and Airy-uniform matching plus a Maslov phase is required.

Common trap“The classical path is the only path” — stationary phase does not say that

Stationary phase is an asymptotic statement, not a claim that non-classical paths disappear at finite $\hbar$ . They still contribute; their net effect is merely suppressed by powers of $\hbar$ , and the leading correction to the saddle result is $O(\hbar)$ . Two further traps: "stationary" means extremal or saddle, not necessarily a minimum of the action (light can take a path of stationary, even maximal, optical length); and at a caustic the Gaussian prefactor $|\det\partial^2 S_{\mathrm{cl}}|^{1/2}$ diverges, so the semiclassical formula must be repaired with a uniform (Airy) approximation and the correct Maslov count — you cannot just plug in.

Worked examples

Example 1 — The Fresnel integral and its $\pi/4$ phase (the engine of stationary phase). We verify the single most-used fact of the method: the value and phase of the Fresnel integral $\int_{-\infty}^{\infty}e^{iax^2}\mathrm dx$ , which is the $\hbar\to0$ skeleton of every path-integral saddle.

Step 1. Rotate the contour. Write $e^{iax^2}=e^{-(-ia)x^2}$ ; for $a>0$ the coefficient $-ia$ has a positive real part after a $-\pi/4$ contour rotation, turning the oscillation into a convergent Gaussian.

Step 2. Apply the Gaussian formula $\int e^{-bx^2}\mathrm dx=\sqrt{\pi/b}$ with $b=-ia$ : the result is $\sqrt{\pi/(-ia)}=\sqrt{\pi/|a|}\,e^{i\pi/4}$ , since $(-i)^{-1/2}=e^{i\pi/4}$ .

Step 3. Check numerically at $a=3$ : $\sqrt{\pi/3}=1.02333$ , and $e^{i\pi/4}=(1+i)/\sqrt2=0.70711(1+i)$ , so the integral equals $0.72360(1+i)$ . The exact symbolic integral agrees to $10^{-9}$ .

Step 4. For $a<0$ the rotation is $+\pi/4$ instead, giving the conjugate phase $e^{-i\pi/4}$ . The sign of the phase is $\mathrm{sgn}(a)$ — this is exactly the $\mathrm{sgn}(S'')$ in equation (2).

Result: $\int_{-\infty}^{\infty}e^{iax^2}\mathrm dx=\sqrt{\pi/|a|}\,e^{i\,\mathrm{sgn}(a)\pi/4}$ , and at $a=3$ this is $\mathbf{0.72360\,(1+i)}$ . The $\pi/4$ is the seed of every Maslov phase.

Example 2 — The free classical action solves Hamilton–Jacobi (and $\partial_x S=p$ ). We show the free-particle $S_{\mathrm{cl}}$ , treated as a function of its endpoint, is an exact solution of the Hamilton–Jacobi equation, and that its space derivative is the momentum.

Step 1. Take $S(x,t)=\dfrac{m(x-x_i)^2}{2t}$ with $x_i$ fixed (the endpoint action from Derivation 2).

Step 2. Differentiate: $\partial_x S=\dfrac{m(x-x_i)}{t}$ and $\partial_t S=-\dfrac{m(x-x_i)^2}{2t^2}$ .

Step 3. Form the Hamilton–Jacobi combination for $H=p^2/2m$ : $\partial_t S+\dfrac{(\partial_x S)^2}{2m}=-\dfrac{m(x-x_i)^2}{2t^2}+\dfrac{1}{2m}\cdot\dfrac{m^2(x-x_i)^2}{t^2}=-\dfrac{m(x-x_i)^2}{2t^2}+\dfrac{m(x-x_i)^2}{2t^2}$ .

Step 4. The two terms cancel identically. Moreover $\partial_x S=m(x-x_i)/t=mv=p$ : the space-gradient of the classical action is the classical momentum, as Hamilton–Jacobi theory demands.

Result: $\partial_t S+(\partial_x S)^2/2m=\mathbf{0}$ exactly, and $\partial_x S=\mathbf{p}$ . The classical phase of the path integral obeys the classical Hamilton–Jacobi equation — the cleanest statement that the saddle point of $\int\mathcal Dx\,e^{iS/\hbar}$ is classical mechanics.

Origins · 1 minhamilton-jacobi equation and how this was discovered

hamilton-jacobi equation

Hamilton 1834–35, Jacobi 1837 · The optical-mechanical analogy

William Rowan Hamilton, working on geometrical optics in the 1830s, discovered that mechanics and optics share a single mathematical skeleton: a "principal function" whose gradient gives momentum and whose level surfaces are wavefronts. Carl Gustav Jacob Jacobi completed the theory in 1837 into the equation now bearing both names, $\partial_t S+H(x,\partial_x S)=0$ . A century before Schrödinger, Hamilton had written the eikonal equation of a wave theory of matter — he simply had no reason to guess the wavelength was $h/p$ . The classical limit of quantum mechanics is, quite literally, the recovery of Hamilton's 1834 vision.

History

1834

— Hamilton introduces the characteristic and principal functions, unifying optics and mechanics

1837

— Jacobi casts the theory as the Hamilton–Jacobi partial differential equation

1847

— Stokes formulates the method of stationary phase for oscillatory integrals (later refined by Kelvin, 1887)

1928

— Van Vleck writes down the correspondence-principle (semiclassical) determinant prefactor

1948

— Feynman's path integral makes stationary phase the origin of classical mechanics

Going further

References. Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §2.6 (path integrals, the classical limit) · Feynman & Hibbs, Quantum Mechanics and Path Integrals, Ch. 2–4 · Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 8 & 16 · Schulman, Techniques and Applications of Path Integration, Ch. 6, 12–13 · Bleistein & Handelsman, Asymptotic Expansions of Integrals, Ch. 6 · Gutzwiller, Chaos in Classical and Quantum Mechanics, Ch. 12.

Provenance. Upgraded from the AllQFT base article (dry: no history, no figures, no worked numbers). Every propagator and limit is machine-verified in `verify_aq-qm-stationary-phase-classical-limit.py`: (1) the Fresnel integral $\int e^{iax^2}\mathrm dx=\sqrt{\pi/|a|}\,e^{i\,\mathrm{sgn}(a)\pi/4}$ at $a=0.7,3,12$ to $10^{-9}$ ; (2) the free $S_{\mathrm{cl}}$ satisfies Hamilton–Jacobi with residual exactly $0$ and $\partial_x S=p$ ; (3) the oscillator Van Vleck prefactor $m\omega/\sin\omega T$ , its $\omega\to0$ reduction to the free action, and caustics at $\omega T=n\pi$ ; (4) the $\mathcal O(\hbar^0)$ Hamilton–Jacobi and $\mathcal O(\hbar^1)$ transport equations from the Schrödinger $\hbar$ -expansion; (5) the linear-in- $\hbar$ shrinkage of the leading correction; (6) the macroscopic estimate $S/\hbar\approx5\times10^{26}$ versus $S/\hbar\sim1$ for an electron in an atom. The construction of the path integral is owned by The Feynman Path Integral Formulation; the exact Gaussian propagators and Van Vleck prefactor by Free-Particle and Harmonic-Oscillator Path Integrals; the eikonal/ $p^{-1/2}$ WKB amplitude by The WKB Approximation; the Maslov index and caustic bookkeeping by Bohr-Sommerfeld and Maslov Quantization — recalled here by name, never re-derived. The operator-language companion statement (that a localized packet's centroid follows Newton's law) is Ehrenfest's Theorem.

Builds on: Lagrangian and Hamiltonian mechanics · The Feynman Path Integral Formulation · Gaussian and saddle-point integrals · The WKB Approximation. Leads to: Hamilton–Jacobi theory · Instantons and Quantum Tunneling · the Gutzwiller trace formula and EBK torus quantization (semiclassical quantum chaos).

If every path has |e^(iS/ℏ)| = 1, how can one path "dominate"?

No path is larger in magnitude — domination is about the phase coherence of a neighbourhood, not the size of a single path. Around the classical trajectory the action is stationary, so a whole tube of paths of width $\sim\sqrt\hbar$ shares nearly the same phase and adds constructively. Everywhere else the phase sweeps through many multiples of $2\pi$ under tiny deformations and the contributions cancel in pairs. So "the classical path dominates" is shorthand for "the classical path is the one surrounded by constructively interfering neighbours," a statement about the second variation $\delta^2 S$ , not about any path's modulus.

Why is the phase exactly π/4 and not some other angle?

The $\pi/4$ is the phase of the Fresnel integral $\int e^{iax^2}\mathrm dx=\sqrt{\pi/|a|}\,e^{i\,\mathrm{sgn}(a)\pi/4}$ , which you get by rotating the integration contour by $\pm45^\circ$ to turn the oscillating integrand into a decaying Gaussian. Algebraically it is $(-i)^{-1/2}=e^{i\pi/4}$ . Physically it is the phase lag a wave accumulates in passing through a focus in one transverse dimension; in $n$ dimensions each independent Gaussian direction contributes its own $\pm\pi/4$ , and the accumulated multiples of $\pi/2$ are precisely the Maslov index. So the $\pi/4$ is not arbitrary — it is the atom out of which every caustic phase is built.

Is the semiclassical propagator ever exact, or always an approximation?

It is exact for any system whose action is at most quadratic in the coordinates: the free particle, the harmonic oscillator, a particle in a uniform force, and a driven oscillator. For these the expansion of the action about the classical path terminates at second order, so the fluctuation integral is a genuine (not approximate) Gaussian and there are no $\hbar$ corrections — the Van Vleck formula equals the full propagator. For anything with cubic or higher terms in the fluctuation (a general potential) the saddle-point result is only the leading term of an asymptotic series in $\hbar$ , with the first correction of order $\hbar$ relative.

How is stationary phase related to WKB?

They are the same physics in different clothing. WKB writes $\psi=A\,e^{iW/\hbar}$ and expands the Schrödinger equation in powers of $\hbar$ ; stationary phase evaluates the path integral $\int\mathcal Dx\,e^{iS/\hbar}$ by the saddle point. Both isolate the classical action $W=S_{\mathrm{cl}}$ at leading order (Hamilton–Jacobi), both produce the amplitude $A\propto p^{-1/2}$ (owned in The WKB Approximation as the transport-equation amplitude, and here as the 1D Van Vleck factor), and both carry the same Maslov phases at turning points/caustics. The path-integral route is more general because it survives in situations with no separable eikonal; the direct-Schrödinger WKB route is the cleaner bookkeeping in one dimension.

What exactly diverges at a caustic, and does the true propagator blow up too?

The semiclassical prefactor diverges, not the true quantum propagator. The Van Vleck determinant $\det(\partial^2 S_{\mathrm{cl}}/\partial x_i\partial x_f)$ measures the focusing of classical trajectories; at a caustic a whole pencil of trajectories refocuses to a point, so the classical density of arrivals is infinite and the amplitude $\propto|\det|^{1/2}$ blows up. The exact kernel stays finite because diffraction smooths the focus — near the caustic the correct description is an Airy function, not a Gaussian. Practically, you cross the caustic by switching to a uniform (Airy) approximation and adding a $-\pi/2$ Maslov phase; the divergence is an artifact of the leading Gaussian approximation, and the oscillator kernel is finite at $\omega T=n\pi$ once the phase is tracked correctly.

Does stationary phase select a minimum of the action?

No — it selects a stationary point, which may be a minimum, a maximum, or a saddle. Hamilton's principle is often loosely called "least action," but the correct statement is "stationary action," $\delta S=0$ . For short times the classical path is indeed a local minimum, but past the first conjugate point (caustic) it becomes a saddle of the action functional, and the sign changes are exactly what the Maslov index counts. The interference does not care whether $S$ is smallest — only that it is stationary, so that neighbouring phases reinforce. This is the direct analog of Fermat's principle in optics, where light can take a path of stationary (sometimes maximal) optical length.

How does this connect to Ehrenfest's theorem and the classicality of wave packets?

Ehrenfest's Theorem is the operator-language twin: it states $m\,\mathrm d^2\langle x\rangle/\mathrm dt^2=-\langle V'(x)\rangle$ . Stationary phase explains why a localized packet's centroid tracks the classical trajectory — the packet is a superposition of de Broglie waves whose phases reinforce along the path where $\delta S=0$ . The two agree when the packet is narrow enough that $\langle V'(x)\rangle\approx V'(\langle x\rangle)$ , the same condition $\lambda_{\mathrm{dB}}\ll L$ that validates the $\hbar$ -expansion. Both fail near caustics and for delocalized states — exactly where quantum mechanics reasserts itself.