Fourier Series and Transforms

Motivation: why this matters in quant finance

The Fourier transform is the single most practically important "advanced" mathematical tool for working quants beyond basic calculus and probability. Its role in finance:

Characteristic function pricing. Many stochastic volatility and jump models (Heston, variance-gamma, CGMY) do not have closed-form densities, but they do have closed-form characteristic functions. The option price is recovered by an inverse Fourier transform. This is the standard calibration method for stochastic vol models.
FFT-based option pricing (Carr-Madan). The Fast Fourier Transform (FFT) computes option prices across an entire grid of strikes simultaneously, turning what would be thousands of separate numerical integrations into a single $O(N \log N)$ computation. This makes real-time calibration of vol surfaces feasible.
Spectral analysis of time series. Fourier analysis decomposes a signal into frequency components, detecting periodicity, seasonality, and cycles in financial data.

Fourier series

Definition

A periodic function $f(x)$ with period $2L$ can be represented as:

f(x) = \frac{a_0}{2} + \sum_{n=1}^{\infty}\left[a_n\cos\frac{n\pi x}{L} + b_n\sin\frac{n\pi x}{L}\right]

where the Fourier coefficients are:

a_n = \frac{1}{L}\int_{-L}^{L} f(x)\cos\frac{n\pi x}{L}\,dx, \qquad b_n = \frac{1}{L}\int_{-L}^{L} f(x)\sin\frac{n\pi x}{L}\,dx

Equivalently, in complex exponential form:

f(x) = \sum_{n=-\infty}^{\infty} c_n\,e^{in\pi x/L}, \qquad c_n = \frac{1}{2L}\int_{-L}^{L} f(x)\,e^{-in\pi x/L}\,dx

The Fourier series decomposes a function into a sum of sinusoidal components at different frequencies. Unlike Taylor series (which approximate locally around a point), Fourier series approximate globally over an entire period.

Convergence

f

is piecewise smooth, the Fourier series converges to

f(x)

at every point where

f

is continuous, and to

\frac{1}{2}[f(x^-) + f(x^+)]

at jump discontinuities.

Parseval's theorem:

\frac{1}{2L}\int_{-L}^{L} |f(x)|^2\,dx = \sum_{n=-\infty}^{\infty} |c_n|^2

. The "energy" of the signal equals the sum of energies in each frequency component.

The Fourier transform

Definition

For a non-periodic function

f: \mathbb{R} \to \mathbb{R}

(or

\mathbb{C}

), the Fourier transform is:

\hat{f}(\xi) = \mathcal{F}\{f\}(\xi) = \int_{-\infty}^{\infty} f(x)\,e^{-2\pi i \xi x}\,dx

and the inverse Fourier transform recovers

f

f(x) = \int_{-\infty}^{\infty} \hat{f}(\xi)\,e^{2\pi i \xi x}\,d\xi

(Convention varies by field; quant finance typically uses the convention $\hat{f}(\xi) = \int f(x)e^{i\xi x}\,dx$ with $1/(2\pi)$ in the inverse. The substance is identical.)

The Fourier transform extends Fourier series from periodic to non-periodic functions, replacing the discrete sum $\sum c_n e^{in\pi x/L}$ with a continuous integral $\int \hat{f}(\xi)e^{2\pi i\xi x}\,d\xi$ .

Key properties

Property	Time domain	Frequency domain
Linearity	$af + bg$	$a\hat{f} + b\hat{g}$
Shifting	$f(x - a)$	$e^{-2\pi i a\xi}\hat{f}(\xi)$
Scaling	$f(ax)$	$\frac{1}{
Derivative	$f'(x)$	$2\pi i\xi\,\hat{f}(\xi)$
Convolution	$(f * g)(x)$	$\hat{f}(\xi)\cdot\hat{g}(\xi)$

The convolution theorem is crucial: convolution in the time/space domain becomes multiplication in the frequency domain. Since the density of a sum of independent random variables is the convolution of their densities, the Fourier transform of the sum's density is the product of the individual Fourier transforms — this is why characteristic functions multiply for independent sums.

The characteristic function

Definition

The characteristic function of a random variable

X

is:

\varphi_X(t) = \mathbb{E}[e^{itX}] = \int_{-\infty}^{\infty} e^{itx}\,f_X(x)\,dx

This is the Fourier transform of the density

f_X

(up to sign convention). Unlike the MGF, the characteristic function always exists because

|e^{itx}| = 1

for real

t

and

x

, so the integral converges for any distribution.

Key properties

$\varphi_X(0) = 1$ always.
$\varphi_X(-t) = \overline{\varphi_X(t)}$ (conjugate symmetry).
Moments: $\mathbb{E}[X^n] = \frac{1}{i^n}\varphi_X^{(n)}(0)$ (when they exist).
Independence: if $X, Y$ are independent, $\varphi_{X+Y}(t) = \varphi_X(t) \cdot \varphi_Y(t)$ .

Key characteristic functions

Distribution	$\varphi_X(t)$
$\mathcal{N}(\mu, \sigma^2)$	$e^{i\mu t - \sigma^2 t^2/2}$
Poisson( $\lambda$ )	$e^{\lambda(e^{it} - 1)}$
Exponential( $\lambda$ )	$\lambda/(\lambda - it)$
Heston (log-price)	Closed form in terms of $t$ , vol-of-vol, correlation, mean reversion

The Heston characteristic function is the key to the Heston model's practical utility: although the density has no closed form, the characteristic function does, enabling Fourier-based pricing.

Fourier pricing methods

The Gil-Pelaez inversion

The CDF of $X$ can be recovered from the characteristic function:

F_X(x) = \frac{1}{2} - \frac{1}{\pi}\int_0^{\infty} \text{Re}\left[\frac{e^{-itx}\varphi_X(t)}{it}\right]dt

This expresses the CDF (and hence tail probabilities) as a single one-dimensional integral, even when the density has no closed form.

The Carr-Madan formula

For a European call with strike $K = e^k$ (log-strike $k$ ), the price is:

C(k) = \frac{e^{-\alpha k}}{\pi}\int_0^{\infty} e^{-ivk}\,\psi(v)\,dv

where $\psi(v)$ is a modified characteristic function incorporating the risk-free rate, dampening parameter $\alpha$ , and the characteristic function of the log-price.

The key insight: this integral has the form of an inverse Fourier transform, so it can be evaluated using the Fast Fourier Transform (FFT), which computes

N

values simultaneously in

O(N\log N)

operations. Instead of computing one option price per numerical integration, you get prices for an entire grid of

N

strikes in a single FFT call.

Practical impact: Calibrating a stochastic vol model (e.g., Heston) to an implied vol surface with 50–100 strikes and 5–10 maturities requires evaluating hundreds of option prices. With Carr-Madan + FFT, this takes milliseconds instead of seconds — fast enough for real-time calibration on a trading desk.

The COS method

An alternative Fourier method that expands the density in a cosine series:

f(x) \approx \sum_{k=0}^{N-1} A_k \cos\left(k\pi\frac{x - a}{b - a}\right)

where the coefficients $A_k$ are computed from the characteristic function. The COS method is often faster and more accurate than Carr-Madan for individual option prices and is particularly efficient for early-exercise (American/Bermudan) options via backward induction in the Fourier domain.

The Discrete Fourier Transform (DFT) and FFT

DFT

For a sequence $\{x_0, x_1, \ldots, x_{N-1}\}$ :

X_k = \sum_{n=0}^{N-1} x_n\,e^{-2\pi ikn/N}, \quad k = 0, 1, \ldots, N-1

This is a discrete approximation to the continuous Fourier transform. The inverse DFT recovers $\{x_n\}$ from $\{X_k\}$ .

FFT

The Fast Fourier Transform (Cooley-Tukey algorithm) computes the DFT in

O(N\log N)

operations instead of

O(N^2)

. For

N = 4096

(a typical grid size in Carr-Madan), this is a speedup of roughly 300×.

In quant finance, the FFT is used in Carr-Madan pricing, COS method pricing, spectral analysis of return time series, and fast convolution for loss distribution aggregation in credit risk.

Examples and applications

Example 1: Heston model calibration

The Heston model has parameters $\{v_0, \kappa, \theta, \xi, \rho\}$ (initial variance, mean reversion speed, long-run variance, vol-of-vol, correlation). Calibration minimises:

\sum_{i} \left(C_{\text{model}}(K_i, T_i; \text{params}) - C_{\text{market}}(K_i, T_i)\right)^2

Each $C_{\text{model}}$ is computed via Carr-Madan (FFT) or COS method using the Heston characteristic function. The optimiser calls this pricing routine hundreds of times during calibration. Without Fourier methods, this would be computationally intractable.

Example 2: recovering the density from the characteristic function

Given the characteristic function $\varphi(t)$ of a variance-gamma process, the density is:

f(x) = \frac{1}{2\pi}\int_{-\infty}^{\infty} e^{-itx}\varphi(t)\,dt

This improper integral is computed numerically (truncating at

|t| = T_{\max}

and discretising). The result gives the full density, from which you can compute any option price, risk measure, or probability.

Example 3: why characteristic functions multiply

If $X$ and $Y$ are independent with log-prices $x = \ln S_1$ , $y = \ln S_2$ , then:

\varphi_{X+Y}(t) = \mathbb{E}[e^{it(X+Y)}] = \mathbb{E}[e^{itX}]\mathbb{E}[e^{itY}] = \varphi_X(t)\varphi_Y(t)

In the Fourier domain, adding independent random variables is just multiplying their characteristic functions. This is the reason Fourier methods are natural for models built from sums of independent components (compound Poisson, Lévy processes, variance-gamma).

Common confusions and pitfalls

"The characteristic function is just the MGF with $it$ instead of $t$ ." Structurally yes (

\varphi(t) = M(it)

), but the crucial difference is that

|e^{itx}| = 1

while

|e^{tx}|

can be infinite. The CF always exists; the MGF may not. For heavy-tailed models (Student's

t

, stable distributions), you must use the CF.

"Fourier methods are only for European options." The COS method extends to early-exercise options via backward induction in the Fourier domain. Barrier options, Bermudans, and some path-dependent options can also be priced using Fourier techniques with appropriate modifications.

"The FFT gives exact prices." The FFT discretises and truncates the Fourier integral, introducing truncation error and aliasing. The dampening parameter

\alpha

in Carr-Madan must be chosen carefully — too small and the integrand decays slowly; too large and numerical precision suffers.

Where this goes next

Fourier methods connect to:

Power Series: The characteristic function is the complex analogue of the MGF, which is a power series. When the MGF exists, $\varphi(t) = M(it)$ .
Taylor Series: Small- $t$ expansion of the CF recovers moments; the COS method is a cosine-series (Fourier) expansion of the density.
Numerical Integration: Fourier pricing reduces to evaluating a one-dimensional integral, computed by quadrature or FFT.
Change of Variables: The Fourier transform is a change of representation from the "price domain" to the "frequency domain."
Normal Distribution: The Gaussian CF $e^{i\mu t - \sigma^2 t^2/2}$ is the simplest non-trivial example and the building block for many models.
Asymptotic Expansions: When the Fourier integral cannot be computed exactly, saddle-point and stationary-phase approximations provide asymptotic results.

References

Stewart, J. (2008). Single Variable Calculus: Early Transcendentals (6th ed.). Thomson Brooks/Cole. The uploaded single-variable resource does not substantially cover Fourier analysis; this note should be grounded in a dedicated Fourier/analysis source when one is added.