Method of Stationary Phase

A startling number of the integrals physics actually needs cannot be evaluated in closed form yet must be known as some large parameter runs off to infinity: the amplitude of a wave a thousand wavelengths from an aperture, a Bessel function at large argument, the semiclassical propagator as $\hbar\to0$ , the long-time tail of a spreading wave packet. Each has the shape $I(\lambda)=\int_a^b g(t)\,e^{i\lambda f(t)}\,\mathrm{d}t$ with $\lambda\to\infty$ , and each resists the obvious expansion in $1/\lambda$ , because the small parameter is trapped inside a phase, where it produces not decay but ever-faster oscillation. The method of stationary phase is built for exactly this shape, and returns not just a number but a picture of why. The picture is interference. Where the phase $f$ is monotone the phasor $e^{i\lambda f(t)}$ winds faster and faster as $\lambda$ grows, so adjacent slices point in nearly opposite directions and annihilate. Cancellation fails only at the stationary points where $f'(t_0)=0$ : there the phase is momentarily frozen, neighbouring phasors align, and a residue survives. That residue scales as $\lambda^{-1/2}$ and carries a universal $\pm\pi/4$ phase — the fingerprint of a Gaussian rotated ninety degrees into the imaginary direction. This one mechanism reproduces geometrical optics from wave optics, the group velocity from a Fourier packet, the Bohr–Sommerfeld rule from a path integral, and the caustics of catastrophe theory from coalescing critical points. Stationary phase is less a trick than the doorway to all of semiclassical physics.

Key takeaways

Stationary phase estimates oscillatory integrals $I(\lambda)=\int g(t)\,e^{i\lambda f(t)}\,\mathrm{d}t$ as $\lambda\to\infty$ , where rapid winding of the phasor makes almost the whole integrand self-cancel.
The survivors are the stationary points $f'(t_0)=0$ ; everywhere else, integration by parts shows the contribution is only $O(\lambda^{-1})$ and comes from the endpoints.
A non-degenerate stationary point gives $I\sim g(t_0)\sqrt{\tfrac{2\pi}{\lambda|f''(t_0)|}}\,e^{i\lambda f(t_0)+i\frac{\pi}{4}\operatorname{sgn}f''(t_0)}$ — amplitude $\lambda^{-1/2}$ , phase $\pm\pi/4$ .
The $\pm\pi/4$ is the phase of $\sqrt{i}$ ; physically it is the Gouy/Maslov phase slip a wave picks up at a focus or turning point.
Degeneracy $f''=0$ slows the decay to $\lambda^{-1/p}$ and swaps the Gaussian for a higher canonical integral; order three is the Airy function of a caustic or WKB turning point.
It is the real-axis limit of steepest descent and the analytic engine of WKB, diffraction, dispersive waves, and special-function asymptotics.
Key idea: rapid oscillation kills the integral everywhere except at stationary phase points $f'(t_0)=0$ , where phasors add coherently to leave a residue of size $\lambda^{-1/2}$ carrying a universal $\pm\pi/4$ Fresnel phase.

The core insight

Rapid oscillation kills the integral everywhere except at stationary phase points $f'(t_0)=0$ , where the coherent addition of phasors leaves a residue of size $\lambda^{-1/2}$ carrying a universal $\pm\pi/4$ Fresnel phase.

Where this lives today

Stationary phase is a working instrument across modern physics. In radar and seismology, matched-filter and chirp-compression algorithms rest on the stationary-phase approximation of a linear-FM waveform — the chirp of the first figure is literally the transmitted pulse, its stationary point where the matched filter concentrates the return. In gravitational-wave astronomy the templates LIGO and Virgo cross-correlate against data are built in the frequency domain by the stationary-phase approximation to the inspiral waveform: the SPA phase $\Psi(f)$ of a compact-binary chirp is a textbook stationary-phase integral, accurate to a fraction of a radian and far cheaper than transforming hundreds of thousands of waveforms numerically. In optics the passage from a smooth caustic to its Airy and Pearcey fringes — rainbow supernumeraries, the bright cusp at the bottom of a teacup — is the degenerate regime of Derivation 2. And every semiclassical calculation, up to the Gutzwiller trace formula for chaotic billiards, is a sum over classical paths $e^{iS/\hbar}$ carrying Maslov phases. The frontier is uniform asymptotics: stitching the ordinary, Airy, and higher catastrophe integrals (Pearcey, swallowtail) into one expression valid as stationary points coalesce.

The physical picture

Think of the integral as a walk in the complex plane. Each slice $g(t)\,\mathrm{d}t$ is a tiny step pointing at angle $\lambda f(t)$ ; as $t$ sweeps the interval the running sum traces a curve. Where $f$ is steep the direction spins so fast the curve coils into a tight closed loop with almost no net displacement — destructive interference made visible. Only where the spinning stalls, at $f'(t_0)=0$ , do consecutive steps align and the sum lurch forward. The whole integral is essentially that one lurch, and its size is set by how wide the stall is. The first figure shows the integrand itself: a chirp that races at the edges and freezes in the middle.

Figure

The chirp cos(λt²/2) with λ = 6 has a stationary point at t₀ = 0 where f′ = 0. Away from it the phase races and neighbouring slices cancel; inside the coherent window of half-width √(2/λ) ≈ 0.58 the phase is frozen and the slices add, leaving a net residue of order λ^−1/2.

How wide is the stall? The phase departs from stationarity like $\tfrac12\lambda f''(t_0)(t-t_0)^2$ , and coherence is lost once that reaches order one, at $|t-t_0|\sim(\lambda|f''|)^{-1/2}$ . So the window shrinks as $\lambda^{-1/2}$ , the integrand is $O(1)$ across it, and the product is $O(\lambda^{-1/2})$ — the answer up to a constant, which the Fresnel integral then fixes. The second figure puts the method on trial against a function known to every decimal — the Bessel function $J_0$ .

Figure

Stationary phase against the exact Bessel function J₀(x): the one-line estimate √(2/πx) cos(x − π/4) in amber locks onto the true J₀ in blue once x ≳ 3, and at x = 10 it predicts −0.247 versus the exact −0.246, an error near 0.3 percent that keeps shrinking as x grows.

The core equations

The object of study is one family of integrals; everything below is a definition, a condition locating its dominant points, or a formula for the residue they leave.

I(\lambda)=\int_a^b g(t)\,e^{i\lambda f(t)}\,\mathrm{d}t,\qquad \lambda\to+\infty \qquad (1)

where: $\lambda$ — large real positive parameter (a frequency, a distance in wavelengths, $1/\hbar$ , a time); $f(t)$ — real-valued phase function, smooth on $[a,b]$ ; $g(t)$ — slowly varying complex amplitude or weight; $[a,b]$ — real interval, possibly infinite. A fast oscillation $e^{i\lambda f}$ modulated by a gentle envelope $g$ ; the whole art is deciding which parts of $[a,b]$ actually contribute. Since the integrand's magnitude is fixed at $|g|$ , the only thing that can make $I$ small is cancellation, which fails only where the phase stops advancing — the condition locating every interior contribution.

f'(t_0)=0 \qquad (2)

where: $t_0$ — a stationary (critical) point of the phase. At $t_0$ the wave is momentarily at rest in phase, so its neighbourhood radiates in step; these are the only interior points that matter at leading order. The residue comes from reducing the local integral to one master integral — the Fresnel integral, the analytic continuation of the Gaussian to a purely imaginary exponent. It is worth memorizing.

\int_{-\infty}^{\infty} e^{i\alpha u^2}\,\mathrm{d}u=\sqrt{\dfrac{\pi}{|\alpha|}}\;e^{\,i\frac{\pi}{4}\operatorname{sgn}\alpha} \qquad (3)

where: $\alpha$ — real nonzero coefficient, here $\alpha=\tfrac12\lambda f''(t_0)$ ; $\operatorname{sgn}\alpha=\pm1$ . Reading: the magnitude $\sqrt{\pi/|\alpha|}$ is the ordinary Gaussian width, and the extra factor $e^{\pm i\pi/4}$ — the phase of $\sqrt{\pm i}$ — is the entire origin of the celebrated quarter-turn. Combining (2) and (3) with a Taylor expansion of the phase gives the workhorse of the whole subject.

I(\lambda)\sim g(t_0)\,\sqrt{\dfrac{2\pi}{\lambda\,|f''(t_0)|}}\;e^{\,i\lambda f(t_0)+i\frac{\pi}{4}\operatorname{sgn} f''(t_0)} \qquad (4)

where: $f''(t_0)$ — curvature of the phase at the stationary point, assumed nonzero (the non-degenerate case). Reading: amplitude $\propto\lambda^{-1/2}$ , a rapidly rotating carrier phase $e^{i\lambda f(t_0)}$ , and the sign of the curvature deciding whether the Fresnel twist is $+\pi/4$ or $-\pi/4$ . When several stationary points lie in the interval, each contributes a copy of (4), added as complex numbers whose relative phases produce observable fringes.

I(\lambda)\sim \sum_{j}\,g(t_j)\,\sqrt{\dfrac{2\pi}{\lambda\,|f''(t_j)|}}\;e^{\,i\lambda f(t_j)+i\frac{\pi}{4}\operatorname{sgn} f''(t_j)} \qquad (5)

where: $t_j$ — all interior non-degenerate stationary points. The sum is coherent: two points beat with period set by $\lambda[f(t_1)-f(t_2)]$ — the interference of a two-slit or a rainbow. If the phase has no stationary point at all, the leading behaviour retreats to the endpoints, and one integration by parts exhibits it exactly, at the subdominant order $\lambda^{-1}$ .

I(\lambda)\sim \dfrac{1}{i\lambda}\left[\dfrac{g(b)}{f'(b)}\,e^{i\lambda f(b)}-\dfrac{g(a)}{f'(a)}\,e^{i\lambda f(a)}\right]+O(\lambda^{-2}) \qquad (6)

where: $f'\neq0$ throughout $[a,b]$ . Reading: the endpoints behave like diffracting edges; because $\lambda^{-1}$ is smaller than $\lambda^{-1/2}$ , an interior stationary point always dominates the edges when one is present. In several dimensions the same machinery gives a determinant of the Hessian in the amplitude and a signature $\sigma=n_+-n_-$ in the phase, $I\sim g(\mathbf{x}_0)(2\pi/\lambda)^{d/2}e^{i\lambda f(\mathbf{x}_0)+i\frac{\pi}{4}\sigma}/\sqrt{|\det\operatorname{Hess}f|}$ — each dimension a factor $\lambda^{-1/2}$ , each negative Hessian direction a further $-\pi/2$ of phase, the seed of the half-integer Maslov index (see the limiting-cases table).

Derivation ALeading order from a quadratic phase and the Fresnel reduction

The goal is to turn "the stationary point dominates" into equation (4), tracking both the $\lambda^{-1/2}$ amplitude and the $\pm\pi/4$ phase. Assume one interior stationary point $t_0$ with $f'(t_0)=0$ , $f''(t_0)\neq0$ , $g(t_0)\neq0$ . The route: localize to a shrinking neighbourhood, replace the phase by its quadratic Taylor polynomial, freeze the amplitude, extend the limits to infinity at exponentially small cost, and evaluate the Fresnel integral by rotating a Gaussian into the complex plane — each step dropping errors one power of $\lambda^{-1/2}$ below.

Step 1. Split the interval into a small window $[t_0-\delta,t_0+\delta]$ around the stationary point and the two remaining stationary-free pieces, and discard the pieces. Why: on the stationary-free pieces $f'\neq0$ , so equation (6) shows their contribution is $O(\lambda^{-1})$ , negligible against the $O(\lambda^{-1/2})$ we are about to find; only the window matters.

Step 2. Inside the window write $f(t)=f(t_0)+\tfrac12 f''(t_0)(t-t_0)^2+O((t-t_0)^3)$ ; the linear term is absent by stationarity. Why: $f(t_0)$ factors out as the carrier $e^{i\lambda f(t_0)}$ , the quadratic term is what the Fresnel integral wants, and the cubic remainder feeds only subleading orders, since over the window of width $(\lambda|f''|)^{-1/2}$ one has $\lambda(t-t_0)^3\sim\lambda^{-1/2}\to0$ .

Step 3. Replace $g(t)$ by the constant $g(t_0)$ and substitute $u=t-t_0$ , giving $I\approx g(t_0)\,e^{i\lambda f(t_0)}\int_{-\delta}^{\delta}e^{\,i\frac{\lambda f''(t_0)}{2}u^2}\,\mathrm{d}u$ , then extend the limits to $\pm\infty$ . Why: $g$ varies on an $O(1)$ scale while the action lives on the $\lambda^{-1/2}$ scale, so it is effectively constant; the tails beyond $\delta$ oscillate with no stationary point, so extending the limits costs only $O(\lambda^{-1})$ .

Step 4. Evaluate the Fresnel integral with $\alpha=\tfrac12\lambda f''(t_0)$ by analytically continuing the Gaussian: $\int e^{-\beta u^2}\mathrm{d}u=\sqrt{\pi/\beta}$ holds for $\operatorname{Re}\beta>0$ , and letting $\beta=\epsilon-i\alpha\to-i\alpha$ gives $\int e^{i\alpha u^2}\mathrm{d}u=\sqrt{\pi/(-i\alpha)}=\sqrt{\pi/|\alpha|}\,e^{i\frac{\pi}{4}\operatorname{sgn}\alpha}$ . Why: $-i=e^{-i\pi/2}$ so $(-i)^{-1/2}=e^{i\pi/4}$ ; approaching the imaginary axis through positive real part keeps the integral convergent, licensing the rotation and pinning the branch.

Step 5. Substitute $\sqrt{\pi/|\alpha|}=\sqrt{2\pi/(\lambda|f''(t_0)|)}$ and $\operatorname{sgn}\alpha=\operatorname{sgn}f''(t_0)$ and multiply the factors together to obtain equation (4). Why: the width of the Gaussian delivered the $\lambda^{-1/2}$ , the rotation delivered the $\pm\pi/4$ , and the carrier $e^{i\lambda f(t_0)}$ came along untouched from Step 2 — the three signatures of stationary phase, assembled.

Go deeper · Derivation B · 1 minDegenerate stationary points and the birth of the Airy function

Derivation BDegenerate stationary points and the birth of the Airy function

The goal is to see what replaces equation (4) when the curvature itself vanishes, $f''(t_0)=0$ , so the quadratic reduction fails. Suppose the first non-vanishing derivative is of order $p$ : $f^{(k)}(t_0)=0$ for $1\le k<p$ while $f^{(p)}(t_0)\neq0$ . The upshot is a slower decay $\lambda^{-1/p}$ and a canonical integral $\int e^{\pm iu^p}\,\mathrm{d}u$ in place of the Fresnel one; the case $p=3$ matters most as the universal profile of a caustic and a WKB turning point — the Airy function.

Step 1. Expand the phase to its leading non-vanishing order, $f(t)\approx f(t_0)+\dfrac{f^{(p)}(t_0)}{p!}(t-t_0)^p$ . Why: with all lower derivatives gone, this single term governs the shape of the phase near $t_0$ , and its exponent $p$ will set both the power of $\lambda$ and the canonical integral.

Step 2. Rescale to absorb $\lambda$ by setting $u=\bigl(\lambda|f^{(p)}(t_0)|/p!\bigr)^{1/p}(t-t_0)$ , so the phase becomes $\pm u^p$ and the Jacobian pulls out a factor $\bigl(p!/(\lambda|f^{(p)}(t_0)|)\bigr)^{1/p}$ . Why: the exponent $1/p$ on $\lambda$ is now explicit — the wider the flat region a higher $p$ produces, the more slowly the integral decays.

Step 3. The reduced integral is $\int_{-\infty}^{\infty}e^{\pm iu^p}\,\mathrm{d}u$ , evaluated by rotating the contour onto the ray of steepest descent through angle $\pi/(2p)$ , giving $\int_0^{\infty}e^{iu^p}\mathrm{d}u=\tfrac1p\Gamma\!\bigl(\tfrac1p\bigr)e^{i\pi/(2p)}$ . Why: for $p=2$ this returns $\tfrac12\sqrt{\pi}\,e^{i\pi/4}$ , half the Fresnel integral, so the degenerate formula reduces to the ordinary one; for general $p$ the Gamma function and phase $\pi/(2p)$ replace $\sqrt{\pi}$ and $\pi/4$ .

Step 4. Read off the decay $\lambda^{-1/p}$ ; for $p=3$ recognize that a control parameter (distance to the turning point or caustic) enters linearly in the phase and the local integral becomes $\operatorname{Ai}(x)=\tfrac1\pi\int_0^\infty\cos(\tfrac{t^3}{3}+xt)\,\mathrm{d}t$ . Why: a cubic phase is a pair of stationary points about to collide; the Airy function interpolates between the two-point interference of equation (5) on one side and exponential decay on the other — the behaviour of a wavefield across a caustic.

Limiting cases

The decay rate of $I(\lambda)$ is fixed by the local geometry of the phase at its critical point and by the dimension; the table lays out the hierarchy from generic to degenerate and multidimensional.

Case	Result	Meaning
Non-degenerate interior stationary point	$I\sim g(t_0)\sqrt{\dfrac{2\pi}{\lambda\lvert f''(t_0)\rvert}}\,e^{i\lambda f(t_0)+i\frac{\pi}{4}\operatorname{sgn}f''(t_0)}$	the generic case: amplitude $\propto\lambda^{-1/2}$ , universal $\pm\pi/4$ Fresnel phase
No stationary point (endpoint dominated)	$I\sim\dfrac{g(b)e^{i\lambda f(b)}}{i\lambda f'(b)}-\dfrac{g(a)e^{i\lambda f(a)}}{i\lambda f'(a)}$	integration by parts; decay $\propto\lambda^{-1}$ , one power faster, from diffracting edges
Degenerate point of order $p$	$I\sim g(t_0)\,\dfrac{\Gamma(1/p)}{p}\left(\dfrac{p!}{\lambda\lvert f^{(p)}(t_0)\rvert}\right)^{1/p}e^{i\lambda f(t_0)\pm i\frac{\pi}{2p}}$	slower decay $\propto\lambda^{-1/p}$ ; $p=2$ recovers the generic case, $p=3$ gives Airy/caustic behaviour
$d$ -dimensional non-degenerate critical point	$I\sim g(\mathbf{x}_0)\left(\dfrac{2\pi}{\lambda}\right)^{d/2}\dfrac{e^{i\lambda f(\mathbf{x}_0)+i\frac{\pi}{4}\sigma}}{\sqrt{\lvert\det\operatorname{Hess}f\rvert}}$	decay $\propto\lambda^{-d/2}$ ; signature $\sigma=n_+-n_-$ is the seed of the Maslov index
Stationary point at an endpoint	$I\sim\tfrac12 g(t_0)\sqrt{\dfrac{2\pi}{\lambda\lvert f''(t_0)\rvert}}\,e^{i\lambda f(t_0)+i\frac{\pi}{4}\operatorname{sgn}f''(t_0)}$	only half the Fresnel window lies in range, so exactly half the interior residue survives

Common trapStationary phase and Laplace's method are the same formula, so they give the same answer

They are not, and conflating them is the most common error. Both give magnitude $\sqrt{2\pi/(\lambda|f''|)}$ , but Laplace's method (real exponent $\int g\,e^{-\lambda\phi}\,\mathrm{d}t$ ) has a genuine peak and no phase, while stationary phase (imaginary exponent, $|e^{i\lambda f}|=1$ everywhere) is a cancellation carrying the extra $e^{\pm i\pi/4}$ from rotating the Gaussian, with algebraic rather than exponentially small corrections. Second trap: never drop the endpoints when there is no interior stationary point — the leading behaviour is then the $O(\lambda^{-1})$ of equation (6), not $\lambda^{-1/2}$ .

Worked examples

1 · The Bessel function from one line of stationary phase (fully worked)

Use the integral representation $J_0(x)=\tfrac1\pi\int_0^\pi\cos(x\sin\theta)\,\mathrm{d}\theta$ to find the large- $x$ behaviour of $J_0$ , and test it numerically at $x=10$ .

Step 1. Write the cosine as the real part of an exponential: $J_0(x)=\tfrac1\pi\operatorname{Re}\int_0^\pi e^{ix\sin\theta}\,\mathrm{d}\theta$ , so the phase is $f(\theta)=\sin\theta$ with $g=1$ and $\lambda=x$ . The problem is now exactly of the form (1).

Step 2. Locate the stationary point: $f'(\theta)=\cos\theta=0$ inside $(0,\pi)$ at $\theta_0=\pi/2$ . There $f(\theta_0)=\sin(\pi/2)=1$ and $f''(\theta_0)=-\sin(\pi/2)=-1$ , so the point is non-degenerate with negative curvature.

Step 3. Apply equation (4) with $g(\theta_0)=1$ , $|f''|=1$ , $\operatorname{sgn}f''=-1$ : the interior integral is $\sqrt{2\pi/x}\,e^{i x-i\pi/4}$ .

Step 4. Take the real part and divide by $\pi$ : $J_0(x)\sim\tfrac1\pi\sqrt{2\pi/x}\cos\!\bigl(x-\tfrac\pi4\bigr)=\sqrt{\dfrac{2}{\pi x}}\cos\!\bigl(x-\tfrac\pi4\bigr)$ . This is the textbook asymptotic, and the tell-tale $-\pi/4$ is nothing but the Fresnel phase of the stationary point.

Step 5. Test at $x=10$ : $\sqrt{2/(10\pi)}=0.25231$ and $\cos(10-\pi/4)=\cos(9.2146)=-0.97800$ , so the estimate is $\mathbf{-0.2468}$ against the exact $J_0(10)=\mathbf{-0.24594}$ — an error of $0.0008$ , about 0.3 percent, and falling like $1/x$ . Stationary phase turns an intractable oscillatory integral into $\sqrt{2/\pi x}\cos(x-\pi/4)$ , accurate to sub-percent by $x=10$ .

2 · Two stationary points, interference, and the Airy connection

Estimate $I(\lambda)=\int_{-\infty}^{\infty}e^{i\lambda(t^3/3-t)}\,\mathrm{d}t$ for large $\lambda$ and interpret the result.

Step 1. The phase $f(t)=t^3/3-t$ has $f'(t)=t^2-1$ , so there are two stationary points, $t=+1$ and $t=-1$ — the case of equation (5).

Step 2. Evaluate the local data. At $t=+1$ : $f=-\tfrac23$ , $f''=2t=+2$ . At $t=-1$ : $f=+\tfrac23$ , $f''=-2$ . Both are non-degenerate with $|f''|=2$ , but opposite curvature signs, so one carries $+\pi/4$ and the other $-\pi/4$ .

Step 3. Add the two copies of (4): $I\sim\sqrt{\tfrac{2\pi}{2\lambda}}\bigl[e^{-i2\lambda/3+i\pi/4}+e^{+i2\lambda/3-i\pi/4}\bigr]=2\sqrt{\dfrac{\pi}{\lambda}}\,\cos\!\Bigl(\dfrac{2\lambda}{3}-\dfrac{\pi}{4}\Bigr)$ . The two points beat against each other, producing a real oscillation whose frequency is set by their phase difference $\tfrac43\lambda$ .

Step 4. Recognize the exact object: the integral is $2\pi\,\lambda^{-1/3}\operatorname{Ai}(-\lambda^{2/3})$ , and the Airy asymptotic $\operatorname{Ai}(-z)\sim\pi^{-1/2}z^{-1/4}\cos(\tfrac23 z^{3/2}-\tfrac\pi4)$ reproduces the stationary-phase sum term for term — the two colliding points are the Airy oscillation, the connection that governs a WKB turning point.

Step 5. Check the numbers at $\lambda=10$ : the stationary-phase sum gives $2\sqrt{\pi/10}\cos(20/3-\pi/4)=\mathbf{1.0317}$ , while the exact $2\pi\cdot10^{-1/3}\operatorname{Ai}(-10^{2/3})=\mathbf{1.0263}$ . Agreement to 0.5 percent from two interfering stationary points. Two stationary points of opposite curvature interfere into $2\sqrt{\pi/\lambda}\cos(2\lambda/3-\pi/4)$ , the oscillatory face of the Airy function.

3 · When the stationary point sits on the boundary

Find the leading behaviour of $I(\lambda)=\int_0^1 e^{i\lambda t^2}\,\mathrm{d}t$ as $\lambda\to\infty$ , and test it at $\lambda=50$ .

Step 1. The phase $f(t)=t^2$ has $f'(t)=2t$ , which vanishes only at $t=0$ — but that is the endpoint of the interval, not an interior point.

Step 2. A stationary point at the boundary contributes only the half of the Fresnel window that lies inside $[0,1]$ , so it gives exactly half of equation (4). With $f''=2$ , $\operatorname{sgn}f''=+1$ , $g=1$ : the boundary-stationary term is $\tfrac12\sqrt{2\pi/(2\lambda)}\,e^{i\pi/4}=\tfrac12\sqrt{\pi/\lambda}\,e^{i\pi/4}$ .

Step 3. The other endpoint $t=1$ has $f'(1)=2\neq0$ , so it contributes an ordinary edge term from equation (6): $\dfrac{g(1)e^{i\lambda}}{i\lambda f'(1)}=\dfrac{e^{i\lambda}}{2i\lambda}$ , of order $\lambda^{-1}$ .

Step 4. Add them: $I\sim\tfrac12\sqrt{\pi/\lambda}\,e^{i\pi/4}-\dfrac{i}{2\lambda}e^{i\lambda}$ . The first term ( $\lambda^{-1/2}$ ) dominates; the second ( $\lambda^{-1}$ ) is the leading correction.

Step 5. Test at $\lambda=50$ : the formula gives $0.08600+0.07897\,i$ (so $|I|=0.11676$ ), while the exact integral, in terms of Fresnel integrals, is $0.08590+0.07900\,i$ ( $|I|=0.11671$ ). The two agree to 0.04 percent — and dropping the endpoint term would spoil the third decimal. A boundary stationary point gives half the interior residue; the far endpoint rides along at order $\lambda^{-1}$ .

Physics of the portrait

The method has a memorable birthday: it was invented not for quantum mechanics or optics but to answer a question anyone can pose from a bridge — why does the wake behind a moving ship always open at the same angle?

Origins · 1 minLord Kelvin and how this was discovered

Lord Kelvin

1824–1907, born William Thomson · the principle of stationary phase for dispersive water waves

William Thomson — Lord Kelvin from 1892 — was the most versatile physicist of the Victorian age, and stationary phase is one of his quietly influential inventions. In an 1887 paper on the waves from a single impulse in deep water, he faced an integral over all wavelengths, each travelling at its own dispersive speed, and reasoned that at a distant point only wavelengths whose phase is stationary in wavenumber survive the mutual cancellation. From that he deduced that a ship's wake is confined within a wedge of half-angle $\arcsin(1/3)\approx19.47^\circ$ , independent of speed — the Kelvin wedge you can still check from any bridge. He had been anticipated by George Stokes, who a generation earlier wrestled the asymptotics of the Airy integral and discovered the Stokes phenomenon. Kelvin's principle was later made rigorous by van der Corput's lemmas, generalized to the complex plane as Debye's steepest descent, and equipped with the Maslov index counting the $\pi/2$ jumps a wave accumulates at caustics. What began as an explanation of boat wakes is now the spine of semiclassical quantum mechanics.

History

1838

— Airy introduces his rainbow integral to describe the light intensity near a caustic

1850

— Stokes computes the Airy integral's asymptotics and uncovers the Stokes phenomenon

1887

— Kelvin states the principle of stationary phase and derives the ship-wave wedge

1909

— Debye recasts the idea as the method of steepest descent for large-order Bessel functions

1965

— Maslov introduces the index counting the π/2 phase jumps accumulated at caustics

Going further

Texts. Bender and Orszag, Advanced Mathematical Methods for Scientists and Engineers, ch. 6 (the standard physicist's account) · Stone and Goldbart, Mathematics for Physics · Arfken, Weber and Harris, Mathematical Methods for Physicists, 7th ed. · Hassani, Mathematical Physics · Wong, Asymptotic Approximations of Integrals (the rigorous reference) · Erdélyi, Asymptotic Expansions. Every formula here is standard. Key numerical benchmarks: the Fresnel master integral (3) follows by analytic continuation of the Gaussian; $J_0(x)\sim\sqrt{2/(\pi x)}\cos(x-\pi/4)$ agrees with the exact value to 0.3 percent at $x=10$ (DLMF 10.17); the two-point stationary-phase sum equals the Airy asymptotic $2\pi\lambda^{-1/3}\operatorname{Ai}(-\lambda^{2/3})$ to 0.5 percent at $\lambda=10$ ; and $\int_0^1 e^{i\lambda t^2}\mathrm{d}t$ agrees with the exact Fresnel-integral value to 0.04 percent at $\lambda=50$ .

Builds on: Asymptotic Series and Poincaré Expansions · Order Symbols and Gauge Functions · Laplace's Method for Integrals · Watson's Lemma. Leads to: Method of Steepest Descent and Saddle Points · WKB Approximation · Connection Formulae and Turning Points · Bohr–Sommerfeld Quantization · Stokes Phenomenon and Anti-Stokes Lines.

Ask the tutor: Give me any oscillatory integral you are stuck on — a Bessel or Airy asymptotic, a diffraction amplitude, a WKB turning point — and I will walk you through locating its stationary points and reading off the $\lambda^{-1/2}$ amplitude and $\pm\pi/4$ phase.

Why is the integral small when the integrand never gets small — its magnitude is 1 everywhere?

The integrand $g\,e^{i\lambda f}$ has magnitude $|g|$ , which shrinks nowhere, so the smallness of $I$ cannot come from its size — it comes entirely from cancellation. As $\lambda$ grows the phasor spins faster, and over any monotone stretch of $f$ the positive and negative excursions almost perfectly annihilate. This fails only where the phase stops advancing, at $f'=0$ , so the survivor is a delicate residue of an almost-complete cancellation. That delicacy is also why the corrections are only algebraically small, $O(\lambda^{-1})$ , rather than exponentially small as in Laplace's method.

What exactly is the difference between stationary phase and Laplace's method?

They are twins with a crucial sign. Laplace's method treats a real decaying exponent $\int g\,e^{-\lambda\phi}\,\mathrm{d}t$ , with a genuine peak at the minimum of $\phi$ and exponential suppression elsewhere; the answer $g(t_0)\sqrt{2\pi/(\lambda\phi'')}\,e^{-\lambda\phi(t_0)}$ is real and phaseless. Stationary phase treats an imaginary exponent $\int g\,e^{i\lambda f}\,\mathrm{d}t$ , with no peak and a cancellation instead — the same magnitude $\sqrt{2\pi/(\lambda|f''|)}$ but times the Fresnel phase $e^{\pm i\pi/4}$ . Formally they are the two real-axis faces of one complex object, the method of steepest descent; the $\pi/4$ is the angle between them.

Why is the phase shift exactly π/4, and what does it mean physically?

Algebraically the $\pi/4$ is the phase of $\sqrt{i}$ : reducing the local integral to $\int e^{i\alpha u^2}\mathrm{d}u=\sqrt{\pi/|\alpha|}\,e^{i\frac\pi4\operatorname{sgn}\alpha}$ rotates a Gaussian ninety degrees into the imaginary direction, and the square root of that rotation is $e^{\pm i\pi/4}$ . Physically it is the Gouy phase — the quarter-wavelength slip a wave acquires through a focus — and in the semiclassical setting the Maslov phase picked up at each turning point. Accumulated over an orbit these quarter-turns give the $+\tfrac12$ in $\oint p\,\mathrm{d}q=2\pi\hbar(n+\tfrac12)$ — the reason the oscillator ground state sits at $\tfrac12\hbar\omega$ rather than zero.

When do the endpoints of the interval matter?

Whenever no interior stationary point is present, and as a subleading correction even when one is. If $f'\neq0$ throughout $[a,b]$ the whole leading behaviour, of order $\lambda^{-1}$ , comes from the two endpoints via one integration by parts — physically, diffracting edges. If an interior stationary point exists it contributes at $\lambda^{-1/2}$ and dominates, but the endpoints remain at $\lambda^{-1}$ and must be kept for the first correction. The subtle case (worked example 3) is a stationary point exactly at an endpoint: only half the Fresnel window is in range, giving half the interior residue plus the far endpoint's edge term.

What happens when two stationary points collide?

They merge into a single degenerate point, and the ordinary formula (4), with $|f''|$ in its denominator, blows up — signalling its own breakdown. The cure is Derivation 2: near the merger the phase is locally cubic and the right description is the Airy function, decaying like $\lambda^{-1/3}$ . On one side of the collision the two points are real and interfere (the oscillations of worked example 2); on the other they are complex and the field decays; the Airy function stitches the regimes together. This is exactly a caustic in optics (a rainbow's bright edge and supernumerary fringes) and a WKB turning point — governed by the same Airy mathematics.

How is stationary phase related to the method of steepest descent?

Stationary phase is the real-axis restriction of steepest descent. In steepest descent one deforms a contour $\int g\,e^{\lambda h(z)}\,\mathrm{d}z$ through a saddle of $h$ along the direction where $\operatorname{Re}h$ falls fastest, turning the oscillatory integral into a Laplace-type peak. Stationary phase is the case $h=if$ purely imaginary on the real axis: the saddle is a real stationary point $f'(t_0)=0$ , and instead of deforming one rotates the local Gaussian by ninety degrees — which is what manufactures the $\pi/4$ . Both extract the contribution of critical points of the exponent; steepest descent is the general tool, stationary phase its gentler real-axis cousin.