From a peak to a spin
In the previous guide you met Laplace's method: an integral of e^{M phi(t)} with a large M is dominated by the single tallest peak of phi, because the exponential weight is overwhelmingly concentrated there, like a spotlight on one spot of a dark stage. You found the maximum, fitted a parabola, did the Gaussian integral, and read off the answer. This guide handles the integral's restless cousin: replace the real growing exponential by an imaginary, oscillating one. The weight no longer piles up at a peak — it spins, and a fresh idea is needed to tame it.
The integral we now care about looks like the integral of g(t) e^{i k phi(t)} dt with a large real k. Here e^{i k phi} is a unit complex number that winds around the circle: its size is always 1, but its angle k phi(t) races as t moves. Nothing is large or small in magnitude anywhere — so where could the answer possibly come from? The honest, beautiful answer is cancellation. Picture adding millions of little arrows, each pointing in a direction that twirls faster and faster. Neighbouring arrows point nearly opposite ways and annihilate in pairs; the sum collapses to almost nothing.
Almost nothing — but not quite. There is one kind of place where the cancellation fails: a point where the phase momentarily stops changing, so neighbouring arrows briefly point the same way and reinforce instead of cancelling. Those are the stationary points, where phi'(t0) = 0 — the very derivative you set to zero to find a flat spot back in Volume I. Everything that survives the great cancellation lives in their tiny neighbourhoods. That single observation is the method of stationary phase.
Stationary phase: where the waves agree
Let us make the estimate concrete. Near a stationary point t0 the phase is locally flat, so its Taylor expansion starts at second order: phi(t) is approximately phi(t0) + (1/2) phi''(t0) (t - t0)^2. The linear term is gone precisely because phi'(t0) = 0. Slot this into the integrand and the surviving piece becomes g(t0) e^{i k phi(t0)} times the integral of e^{i (k/2) phi''(t0) (t - t0)^2} dt — a Gaussian, but with an imaginary exponent. That imaginary Gaussian is the whole trick, and it does converge, giving a clean closed form.
integral of g(t) e^{i k phi(t)} dt (large k)
only phi'(t0) = 0 survives ->
~ g(t0) * sqrt( 2*pi / (k * |phi''(t0)|) )
* e^{ i k phi(t0) }
* e^{ i * sigma * pi/4 }, sigma = sign of phi''(t0)
size of the surviving piece ~ 1 / sqrt(k) (slow decay)Compare this with Laplace's method and one difference jumps out. Laplace's peak gave a contribution that shrinks like 1/sqrt(M) but multiplied by a possibly huge or tiny e^{M phi(t0)} — exponential weighting. Stationary phase has no exponential weighting at all; magnitude stays 1 everywhere, so the answer decays only as 1/sqrt(k). Oscillation cancels slowly and grudgingly, whereas an exponential peak localises with brutal efficiency. There is also a new fingerprint: that extra phase factor e^{i sigma pi/4}, with sigma the sign of phi''(t0). It comes straight from rotating the Gaussian into the imaginary axis, and it is the honest cost of the phase being imaginary rather than real.
Why optics and waves live here
This is not an abstract trick — it is how nature draws light. A wave field at a faraway screen is a sum (an integral) over every path the wave could take, each carrying a phase k phi proportional to its travel time. For most paths the phase spins wildly as you nudge the path, so they cancel. The paths that survive are exactly the ones where the travel time is stationary: phi'(t0) = 0. That is Fermat's principle — light travels along paths of stationary time — falling out of stationary phase as a theorem, not an axiom. The bright spots of a diffraction pattern, the focus of a lens, a rainbow's caustic: all are stationary-phase survivors.
The same machinery gives the group velocity of a wave packet and how it spreads as it travels, the far-field radiation pattern of an antenna, and the connection formulae of the WKB method you will meet next. Whenever you hear that a system 'picks out' a stationary path, time, or action — Fermat's principle, the principle of least action, the semiclassical limit — there is a stationary-phase or saddle-point integral quietly doing the picking.
Into the complex plane: steepest descent
Stationary phase works, but the imaginary Gaussian and that pi/4 twist feel like patches. The method of steepest descent gives the cleaner, deeper view by unifying the oscillating case and the peaked case into one picture in the complex plane. Take an integral of e^{M f(z)} dz where f is an analytic function of a complex variable z and M is large. Whether the exponent looks growing or oscillating on the real line, in the complex plane it is just one analytic landscape — and we are free to walk a different route across it.
Write f(z) = u(z) + i v(z). The integrand's magnitude is |e^{M f}| = e^{M u}, set entirely by the real part u; the imaginary part v is the pure phase that spins. On the real axis v might be racing — pure oscillation, the hard case. But moving the path into the complex plane is a legitimate, exact move, not a cheat: because f is analytic, Cauchy's theorem says the integral is unchanged when we deform the contour through any region where f stays analytic, just as it underwrites the residue theorem you use to crack real integrals. So we go hunting for a better path — one tailor-made to kill the oscillation.
The key landmark is a saddle point z0, a place where f'(z0) = 0. Here analyticity forces something striking: an analytic function's real part u can never have an ordinary hilltop or valley-bottom in the plane (that is the maximum principle for a harmonic function, which u always is). The only critical landform u can build is a saddle — a mountain pass that rises toward two peaks one way and falls into two valleys the perpendicular way. Picture standing on that pass. There is exactly one direction in which the ground drops away as steeply as possible: the path of steepest descent.
Here is the magic. Along that steepest-descent path through the saddle, the imaginary part v stays exactly constant — the phase freezes, the spinning stops dead. (This is no coincidence: for an analytic function the curves of constant phase and the curves of steepest height-change are one and the same.) On that path the integrand is a real, sharply peaked bump in e^{M u}, and we are back in Laplace's-method territory — fit a parabola, do a real Gaussian, done. The wild oscillation has been traded, exactly, for a tame peak just by choosing the route.
The recipe, and where it bites
- Find the saddle points: solve f'(z0) = 0 in the complex plane. There may be several; you will keep only the ones your contour actually needs.
- Deform the original contour, using Cauchy's theorem, onto the steepest-descent path that runs through the chosen saddle — the curve along which the imaginary part of f stays constant.
- Expand f to second order at the saddle, f(z) approx f(z0) + (1/2) f''(z0) (z - z0)^2, turning the local integrand into a real Gaussian bump.
- Do the Gaussian integral along the descent direction. The leading term is e^{M f(z0)} times sqrt(2 pi / (M |f''(z0)|)), carrying a phase factor fixed by the orientation of the descent path.
Notice the leading answer has the very same shape as Laplace's method — e^{M f(z0)} sqrt(2 pi / (M |f''(z0)|)) — just evaluated at a complex saddle, with a phase from the path's tilt. And just as Watson's lemma upgrades Laplace's leading estimate to a full asymptotic series, expanding f past second order at the saddle hands you the higher corrections here too. This one method delivers the large-argument behaviour of an astonishing roster of special functions: Bessel functions, Airy functions, the gamma function off the real axis, and most integral transforms.