Characteristic Function Pricing

Pricing an option normally means integrating the payoff against the probability density of the stock price at expiry. For Black-Scholes that density is a clean lognormal curve, easy to write down. For a model like Heston's stochastic volatility, or one with random jumps layered in, nobody has ever found a closed-form expression for that density — it may not even exist in a usable form. What these models do have, almost always, is a closed-form characteristic function: the Fourier transform of that same unknown density. It turns out that's enough.

The recipe you can write down, even when the dish is a mystery

A chord on a piano isn't described note by note when a sound engineer analyses it — it's described by its frequency spectrum, which fully captures the sound without ever drawing the waveform itself. Two completely different-looking waveforms can share the same spectrum, and knowing the spectrum is exactly as good as knowing the waveform for most purposes. A characteristic function is a distribution's frequency spectrum: it contains everything the density does, in a different, often more tractable, form.

$$\phi(u) = \mathbb{E}\left[e^{iuX}\right]$$

In plain English: $X$ is the log stock price at expiry, $u$ is a dummy frequency variable, and $\phi(u)$ is the expected value of a rotating complex number whose angle is tied to $X$. Crucially, for models like Heston, an explicit formula for $\phi(u)$ exists even though no explicit formula for the density of $X$ does — the randomness in volatility that makes the density intractable turns out not to block the characteristic function at all.

Worked example 1: sanity check on a case you can verify. For an ordinary normal distribution with mean $\mu$ and variance $\sigma^2$, the characteristic function is known in closed form: $\phi(u)=e^{i\mu u - \frac12\sigma^2 u^2}$. Take $\mu=0$, $\sigma=1$, $u=1$: $\phi(1) = e^{i\cdot 0 - 0.5} = e^{-0.5} \approx 0.6065$ (with an angle of zero here since $\mu=0$, so the imaginary part is zero and the answer is a plain real number). This matches independently computable moments of the standard normal — the characteristic function isn't a black box, it's just a different, equally checkable, encoding of the same distribution.

Drag the mean and standard deviation here and picture each resulting curve as also being fully described by its own $\phi(u)=e^{i\mu u - \frac12\sigma^2 u^2}$ — every shape you can draw with the sliders has an exact Fourier counterpart, computed the same way.

Worked example 2: why this matters for pricing. The price of a European call can be written entirely in terms of $\phi$, via the Gil-Pelaez inversion formula, roughly $C = S - \frac{K}{\pi}\int_0^\infty \text{Re}\left[\frac{e^{-iu\ln K}\phi(u-i)}{iu\,\phi(-i)}\right] du$ (the exact form varies by convention). The important thing isn't memorising the integral — it's that every ingredient on the right-hand side is $\phi$, evaluated at different points, and nothing requires ever knowing the density $f(x)$ itself. For Heston, $\phi(u)$ has a known closed form with about five parameters (initial variance, long-run variance, mean-reversion speed, vol-of-vol, correlation); plug those in, run the integral numerically, and you get an option price without ever simulating a single path.

What this means in practice

Every option pricing model built to capture volatility skew or jumps efficiently — Heston, Variance Gamma, Merton jump-diffusion, most affine models — is calibrated and priced through its characteristic function rather than its density. Carr-Madan FFT Pricing takes this one step further, batching the inversion integral for every strike at once via the FFT instead of running it strike by strike.