When a parameter grows, integrals get simple
Up to now your contour work has chased exact answers: the residue theorem hands you a closed form, and that is that. But an enormous class of integrals important in physics and number theory has no closed form at all — and yet they still want to be understood. The trick is to stop asking for the exact value and instead ask how the integral behaves as some parameter, call it the large number N, runs off to infinity. That is asymptotics: not the answer, but the answer's shape for large N, often a single leading term plus a controlled error that shrinks as N grows.
Here is the one picture that runs through this whole guide. Take an integral of the form integral of e^(N phi(z)) dz, where N is large and phi is some holomorphic function. Where phi has a large real part, e^(N phi) is astronomically big; where the real part is even slightly smaller, e^(N phi) is exponentially smaller still — and the gap explodes as N grows. So for huge N the integrand is essentially zero everywhere except in a tiny neighbourhood of the point where Re(phi) is largest. The whole integral is decided by what happens at one point. Asymptotics is the art of finding that point and reading off its local contribution.
Laplace's method: the real-line warm-up
Start on the real line, where there is no contour to bend yet, just to get the mechanism in your hands. Suppose phi is real, smooth, and has a single interior maximum at the point t_0. Near t_0 the function looks like a downward parabola, phi(t) is approximately phi(t_0) - (1/2)|phi''(t_0)|(t - t_0)^2, because the first derivative vanishes at a maximum. So e^(N phi(t)) is approximately e^(N phi(t_0)) times a Gaussian bump in t, narrower and narrower as N grows. This is Laplace's method, and it reduces a hard integral to a Gaussian integral you already know how to do.
Doing that Gaussian integral gives the headline formula: integral of e^(N phi(t)) dt is approximately e^(N phi(t_0)) times the square root of (2 pi / (N |phi''(t_0)|)), for large N. Read the anatomy slowly, because it repeats everywhere. The exponential e^(N phi(t_0)) is the value at the peak — the dominant scale. The square root of 1/N is the width of the bump — narrower contributions for larger N. And phi''(t_0), the curvature of the peak, controls how sharp the bump is: a sharper peak gives a smaller integral. Three ingredients, one peak.
One refinement makes Laplace's method industrial. Often the integrand is e^(N phi(t)) times an extra slowly varying factor f(t), or the maximum sits at an endpoint rather than an interior peak. Watson's lemma handles exactly the endpoint case: when the action piles up at t = 0 and the integrand expands in a power series there, you may integrate that series term by term against the exponential and recover a full asymptotic expansion in powers of 1/N, not just the leading term. It is the workhorse for turning a single dominant point into a whole series of corrections.
Steepest descent: bending the contour to find the peak
Now the genuinely complex story. In the complex plane phi(z) is holomorphic, and its real part Re(phi) — the thing controlling the size of e^(N phi) — is a harmonic function. Harmonic functions have no interior maxima or minima; the only flat spots are saddle points, places where phi'(z_0) = 0. A saddle looks like a mountain pass: uphill in two opposing directions, downhill in the perpendicular two. There is no peak to sit on. So the Laplace picture seems to break — until you remember you are allowed to move the contour.
This is where Cauchy's theory earns its keep. Because e^(N phi) is holomorphic, deforming the contour across a region with no singularities does not change the integral one bit — that is just the Cauchy integral theorem in action. So you are free to slide your path of integration until it crosses the saddle in the most convenient way. The most convenient way is the path along which Re(phi) drops as fast as possible on both sides of the saddle — the path of steepest descent. On that path the integrand decays like a clean Gaussian away from z_0, and Laplace's method applies verbatim. This is the method of steepest descent.
There is a gorgeous bonus hidden in the geometry. The level curves where Re(phi) is constant and the curves where Im(phi) is constant cross at right angles — that orthogonality is exactly the Cauchy-Riemann equations showing their face. The steepest-descent direction, where Re(phi) falls fastest, is the direction along which Im(phi) stays constant. So on the magic path the imaginary part of phi is frozen, the integrand has no oscillation at all, and the wild cancellation that makes oscillatory integrals so painful simply evaporates. You have traded a wobbling integrand for a placid Gaussian. That trade is the whole point, and the reason it is called the saddle-point method.
I(N) = integral over C of e^( N phi(z) ) dz , N large 1. find the saddle: phi'(z_0) = 0 2. deform C through z_0 along steepest descent (Im phi constant) 3. expand: phi(z) ~ phi(z_0) + (1/2) phi''(z_0) (z - z_0)^2 4. do the Gaussian: I(N) ~ e^( N phi(z_0) ) * sqrt( 2 pi / ( N * (-phi''(z_0)) ) )
Stationary phase: when the integrand only oscillates
A close cousin appears constantly in wave physics and signal theory, where integrals look like integral of g(t) e^(i N psi(t)) dt with N large and everything in sight real. Now there is no decay anywhere — the integrand has constant modulus |g(t)| and merely spins around the unit circle faster and faster as N grows. Naively you might expect such a fast-spinning integral to be large, but the opposite is true: rapid oscillation means neighbouring contributions point in opposite directions and cancel almost perfectly. The integral is tiny — except where the spinning momentarily stops.
The spinning stops where the phase is momentarily flat, that is, where psi'(t_0) = 0 — a stationary point of the phase. Near such a point the phase barely changes, neighbouring contributions line up instead of cancelling, and that small window dominates the whole integral. This is the method of stationary phase, and structurally it is the twin of steepest descent: in both, one special point where the derivative of the exponent vanishes swallows the integral. The difference is only whether the exponent's effect is decay (real phi) or oscillation (imaginary phi).
Carrying out the local Gaussian — now a Fresnel integral, the oscillatory sibling of the Gaussian — gives the leading term: the integral is approximately g(t_0) e^(i N psi(t_0)) times the square root of (2 pi / (N |psi''(t_0)|)), with an extra phase factor e^(i sigma pi/4) where sigma is the sign of psi''(t_0). Notice the 1/N falloff again, but now there is no exponential smallness, only an algebraic decay — oscillatory integrals die off far more gently than steepest-descent ones, which is why diffraction and wave fronts decay so slowly.
A real prize: Stirling's formula for n!
Let us cash all of this in on something famous. The factorial n! is captured by the Gamma function through Gamma(n + 1) = n!, and Gamma has the Euler integral Gamma(n + 1) = integral from 0 to infinity of t^n e^(-t) dt. Substitute t = N s to pull out the large parameter N = n, and the integrand becomes e^(N(log s - s)) times a constant — exactly the e^(N phi) shape, with phi(s) = log s - s. We are one saddle away from an answer.
- Find the saddle: phi(s) = log s - s gives phi'(s) = 1/s - 1, which vanishes at s_0 = 1.
- Read the peak value: phi(1) = log 1 - 1 = -1, and the curvature phi''(s) = -1/s^2 gives phi''(1) = -1.
- Apply the Laplace formula: integral ~ e^(N phi(1)) sqrt(2 pi / (N |phi''(1)|)) = e^(-N) sqrt(2 pi / N).
- Restore the constants from t = N s (a factor N^(N+1)) to assemble Gamma(n + 1).
Putting the pieces together with N = n delivers Stirling's formula: n! is approximately the square root of (2 pi n) times (n/e)^n, for large n. Stop and admire what just happened. A purely real, deeply combinatorial quantity — the number of ways to arrange n objects — fell straight out of finding where the derivative of log s - s vanishes in the complex plane and doing one Gaussian integral. The e in (n/e)^n is the peak value e^(-1); the square root of (2 pi n) is the width of the bump. Every symbol traces back to the saddle. And Watson's lemma, applied to the same integral, hands you the whole correction series 1 + 1/(12n) + 1/(288 n^2) - ... for free.
Where it works, where it bites
Be honest about the assumptions, because each one has teeth. First, deforming the contour to the saddle is legal only across regions free of singularities — if a pole or branch cut sits between your starting path and the steepest-descent path, sliding past it changes the integral, and you must add the residues you sweep across or route around the cut. The deformation is not free; it is paid for by Cauchy's theorem, and that theorem demands an obstacle-free passage.
Second, the leading term is honest only when the saddle is simple, meaning phi''(z_0) is nonzero. When phi''(z_0) vanishes too, two saddles have merged, the Gaussian approximation collapses, and the local model is no longer a parabola but a cubic — the integral is governed by an Airy function rather than a square root. That is exactly the regime of Airy-function asymptotics, the mathematics of a caustic, the bright edge of a rainbow where two light rays coalesce. Coalescing saddles are not an edge case; they are the gateway to the next layer of the subject.
One last guardrail. The added arcs you discard when closing a contour really must vanish — that is not decoration, it is the same discipline as Jordan's lemma from your residue work, and skipping the check is how wrong answers sneak in. With those caveats respected, you now hold a remarkably general engine: any integral wearing an e^(N phi) coat, real or oscillatory, surrenders its large-N behaviour to a saddle and a Gaussian. The next guide turns the same contour-bending instinct on the Bromwich contour, inverting Laplace transforms by the very deformations you just learned to trust.