Asymptotic Expansions and the Saddle-Point Method

Some quantities you need, a deep out-of-the-money option's price, a rare-event probability in a tail-risk model, come from an integral with no closed form and no realistic hope of an exact answer. But you often don't need the exact answer; you need a very good approximation, especially in an extreme regime (very small time, very far strike, very large sample size). Asymptotic expansions supply exactly this: an approximating series, built specifically for a limiting regime, that can be far more accurate there than a "converges eventually" series ever would be, sometimes stunningly so, using just its first term or two. The saddle-point method is the standard technique for producing one from an integral by finding a special point where the integrand's behavior is easiest to capture.

The analogy: hiking through a mountain pass

Suppose the value of an integral is dominated by a huge exponential factor $e^{Ng(z)}$ for some function $g$ and a large parameter $N$, most of an integral's "mass" concentrates overwhelmingly near wherever $g$ is largest along the path, the way almost all the weight of a very peaked mountain range sits near its single highest ridge. If you route the integration path through a saddle point of $g$ in the complex plane, a mountain pass, high in one direction and low in the perpendicular direction, the path can be chosen so it descends steepest away from the pass in both directions, letting the whole integral be approximated by the simple, well-understood behavior right at that one point, ignoring everywhere else almost entirely.

The mechanics, one symbol at a time

For an integral of the form $\displaystyle I(N) = \int e^{N g(z)}\,dz$ with large $N$, find a saddle point $z_0$ where $g'(z_0)=0$ (a critical point of $g$, analogous to the zero-gradient points from ordinary calculus, but now in the complex plane). Near $z_0$, Taylor-expand: $g(z) \approx g(z_0) + \tfrac12 g''(z_0)(z-z_0)^2$. Deforming the contour to pass through $z_0$ along the direction of steepest descent turns the integral into a Gaussian-type integral, giving the leading-order asymptotic approximation

In plain English: the integral is approximately the peak height $e^{Ng(z_0)}$ times a correction factor that depends only on how sharply the function curves at that one peak (via $g''(z_0)$), everything else about the integrand, everywhere except right at the saddle, drops out of the leading approximation.

Worked example 1: recovering Stirling's approximation

The factorial has the integral representation $n! = \int_0^\infty t^n e^{-t}\,dt$. Writing $t = nz$ makes the exponent $n(\ln z - z)$, so $g(z) = \ln z - z$, with $g'(z) = 1/z - 1 = 0$ at $z_0=1$, and $g''(z_0) = -1/z_0^2 = -1$. The saddle-point formula gives $n! \sim n^n e^{-n}\sqrt{2\pi/n}$, Stirling's approximation. Checking $n=10$: exact $10! = 3{,}628{,}800$; Stirling gives $10^{10}e^{-10}\sqrt{2\pi/10} \approx 3{,}598{,}696$, an error of under 1%, remarkably accurate from a single saddle-point evaluation, with no summing of correction terms needed.

Worked example 2: a deep out-of-the-money tail probability

Suppose you need $P(X > 5)$ for a random variable with moment-generating function $M(\theta)=e^{\theta^2/2}$ (standard normal). The saddle-point approximation for a tail probability uses the point $\hat\theta$ solving $M'(\hat\theta)/M(\hat\theta) = 5$, i.e. $\hat\theta = 5$, giving the leading-order approximation

The exact standard normal tail probability $P(X>5)$ is about $2.87\times10^{-7}$, the saddle-point estimate is accurate to within about 3%, computed from one algebraic evaluation rather than numerically integrating a vanishingly small tail directly, which is often numerically unstable at this depth.

The steep, sharply peaked curves in this family are exactly the shape a saddle-point integrand takes near its dominant point, nearly all the area concentrates in a narrow region around the peak, which is precisely what the method exploits.

What this means in practice

Saddle-point and related asymptotic methods are the standard way to compute deep-tail option prices and rare-event probabilities that are numerically unstable or impossibly slow to get by brute-force integration or naive Monte Carlo, they're used directly to approximate characteristic-function option prices at extreme strikes and to estimate probabilities of large portfolio losses under credit risk models. The trade-off to remember: these expansions are usually divergent series that are only accurate for the regime they were built for (large $N$, deep tail); using more terms doesn't always help, and using the formula outside its intended regime can be badly wrong.