Quantitative Finance Fourier-Cosine Series Spectral Convergence

Fang & Oosterlee (2008) Fourier-Cosine (COS) Option Pricing Lab

Price European options with exponential / geometric convergence $\mathcal{O}(e^{-c N})$ using Fourier-cosine series expansions. Explore cumulant-based truncation intervals $[a, b]$, closed-form payoff coefficients $\chi_k$ & $\psi_k$, and exact analytical Greeks.

Expansion: $V = e^{-rT} \sum_{k=0}^{N-1}{'} \operatorname{Re}[\phi(\frac{k\pi}{b-a}) e^{-i k\pi \frac{a}{b-a}}] V_k$ Truncation: $[c_1 - L\sqrt{c_2 + \sqrt{c_4}}, c_1 + L\sqrt{c_2 + \sqrt{c_4}}]$ Convergence: Exponential ($\sim 128$ terms)
Institutional Highlights
  • • Gold standard Fourier option pricer across quantitative trading desks
  • • Outperforms Carr-Madan FFT with geometric error decay
  • • Free from rigid Nyquist reciprocal strike grid spacing constraints
  • • Closed-form analytical Delta and Gamma with zero bumping noise
ATM Call Price
$0.00
COS Inversion at $K \approx S_0$
ATM Put Price
$0.00
Analytical Payoff $V_k^{\text{put}}$
Analytical Delta ($\Delta$)
0.000
Closed-Form Derivative
Analytical Gamma ($\Gamma$)
0.0000
Via Cosine Density Expansion
Max Absolute Error
< 1e-12
vs Analytical BSM Benchmark
Data Export
Full Strikes & Greeks CSV
Calibrated Market Presets:
Loading baseline preset...

Model & Numerical Controls

1.00

Fourier-Cosine Series Parameters

10
Domain is $[c_1 - L \cdot \text{scale}, c_1 + L \cdot \text{scale}]$. $L = 10$ ensures truncation errors $< 10^{-14}$.
COS Interval & Cumulants
Domain Lower Bound $a$: --
Domain Upper Bound $b$: --
Truncation Factor $L$: --
Expansion Terms $N$: --
First Cumulant $c_1$: --
Second Cumulant $c_2$: --

COS Spectral Valuation Spectrum & Greeks

Key Algorithmic Advantage: Switch to the Spectral Convergence tab to witness the steep geometric decline in error $\mathcal{O}(e^{-c N})$. Unlike the Carr-Madan FFT which requires thousands of points, the COS method reaches machine precision with just $N = 128$ cosine terms.

Expansion Terms $N$ vs. Truncation Factor $L$

Maximum Absolute Error (MAE) across strikes against the analytical Black-Scholes benchmark ($S_0 = 100, T = 1$).

Factor $L$ \ Terms $N$ 32 64 128 256 512
Notice how $L \ge 10$ and $N \ge 128$ consistently deliver machine-precision error ($< 10^{-12}$).

Maturity $T$ vs. Volatility $\sigma$

At-the-Money European Call Option Value ($S_0 = 100, K = 100$) across term structure and volatility surfaces.

Vol $\sigma$ \ Expiry $T$ 0.25y 0.50y 1.00y 2.00y 3.00y
Illustrates option price monotonicity across increasing time to maturity and volatility parameters.

Mathematical Derivations: Fang-Oosterlee (2008) COS Method

1. Fourier-Cosine Series Expansion of Risk-Neutral Density

Let $y \equiv \ln(S_T / S_0)$ be the log-asset return. For a truncated interval $[a, b] \subset \mathbb{R}$, any smooth probability density function $f(y)$ can be expanded in terms of a Fourier-cosine series:

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

where $\sum'$ indicates that the prime summation has its first term ($k=0$) multiplied by $1/2$. The series coefficients $A_k$ are determined directly from the characteristic function $\phi(u) = \mathbb{E}[e^{i u y}]$:

$$A_k = \frac{2}{b - a} \operatorname{Re}\left[ \phi\left( \frac{k\pi}{b - a} \right) \exp\left( -i \frac{k\pi a}{b - a} \right) \right]$$

No numerical integration or fast Fourier transform is needed to compute $A_k$; they are directly sampled from the analytical characteristic function!

2. Analytical Closed-Form Payoff Coefficients $V_k$

The European call option price is $V(S_0, K, T) = e^{-rT} \int_a^b (S_0 e^y - K)^+ f(y) \, dy$. Substituting the cosine expansion of $f(y)$ yields:

$$V(S_0, K, T) \approx e^{-rT} \frac{b - a}{2} {\sum_{k=0}^{N-1}}' A_k V_k$$

With log-strike cutoff $x = \ln(K / S_0)$, the call payoff coefficients $V_k^{\text{call}}$ integrate analytically:

$$V_k^{\text{call}} = \frac{2}{b - a} \left[ S_0 \chi_k(x, b) - K \psi_k(x, b) \right]$$

where $\chi_k(c, d) = \int_c^d e^y \cos(\omega_k(y-a)) dy$ and $\psi_k(c, d) = \int_c^d \cos(\omega_k(y-a)) dy$ have exact closed-form trigonometric solutions.

3. Cumulant-Based Integration Domain $[a, b]$

To prevent both truncation error (cutting off density tails) and aliasing error (too wide domain requiring excessive $N$), the bounds $[a, b]$ are calibrated to the log-asset distribution cumulants $c_1, c_2, c_4$:

$$a = c_1 - L \sqrt{c_2 + \sqrt{c_4}}, \quad b = c_1 + L \sqrt{c_2 + \sqrt{c_4}}$$

By choosing $L \in [8, 12]$ (default $L=10$), the tail probability outside $[a, b]$ is bounded by:

$$\mathbb{P}(y \notin [a, b]) \le \frac{1}{L^4} \sim 10^{-4} \implies \text{Option Truncation Error } < 10^{-14}$$

4. Exact Analytical Greeks via Fourier Differentiation

Unlike finite-difference bump-and-reprice methods that suffer from round-off error, the COS method yields analytical Greeks directly:

  • Delta ($\Delta$): $$\Delta = e^{-rT} {\sum_{k=0}^{N-1}}' A_k \cdot \chi_k(x, b)$$
  • Gamma ($\Gamma$): Evaluating the cosine density expansion directly at $x = \ln(K / S_0)$: $$\Gamma = e^{-rT} \frac{f(x)}{S_0 K} = e^{-rT} \frac{1}{S_0 K} \frac{2}{b - a} {\sum_{k=0}^{N-1}}' A_k \cos\left( k\pi \frac{x - a}{b - a} \right)$$

Because $f(x)$ is computed from the same precalculated coefficients $A_k$, Gamma is evaluated essentially for free with machine precision!

Interactive Mastery Assessment: Fang-Oosterlee COS Method

1. What is the fundamental convergence rate of the Fang-Oosterlee COS method for smooth asset densities?

2. Why does the COS method avoid numerical quadrature when evaluating option payoffs?

3. How are the integration bounds $[a, b]$ derived in the COS framework?

4. What is a primary operational advantage of the COS method over Carr-Madan FFT?

Frequently Asked Questions

The COS method, introduced by Fang and Oosterlee (2008), reconstructs the risk-neutral probability density function from its characteristic function via a Fourier-cosine series expansion on a truncated interval $[a, b]$. Because the cosine coefficients of European call and put payoffs have exact analytical closed-form integrals, options can be priced with exponential spectral convergence using only 64 to 128 terms.

For underlying asset distributions with infinitely smooth ($C^\infty$) probability density functions, Fourier-cosine series coefficients decay exponentially fast as $\mathcal{O}(e^{-c N})$. Consequently, the numerical truncation error drops geometrically, allowing the algorithm to achieve machine precision ($10^{-12}$) with a tiny fraction of the expansion terms required by discrete Fast Fourier Transform methods.

Fang and Oosterlee formulate the integration domain $[a, b]$ using the cumulants of the log-asset price: $a = c_1 - L \sqrt{c_2 + \sqrt{c_4}}$ and $b = c_1 + L \sqrt{c_2 + \sqrt{c_4}}$, where $c_1, c_2,$ and $c_4$ represent the mean, variance, and fourth cumulant. Setting $L$ between 8 and 12 guarantees that truncation probability errors remain below $10^{-12}$ while avoiding unnecessary domain expansion.

Because the option price is represented as a linear combination of analytical cosine payoff integrals, Greeks such as Delta and Gamma are obtained in closed form by directly differentiating the payoff coefficients or evaluating the cosine expansion of the risk-neutral density at the log-strike. This eliminates numerical bumping noise, step-size tuning, and re-pricing overhead.

While Carr-Madan FFT links integration step $\eta$ and strike resolution $\lambda$ via the rigid reciprocal constraint $\lambda \eta = \frac{2\pi}{N}$, the COS method prices arbitrary individual or multiple strikes directly. Furthermore, the COS method avoids arbitrary damping parameters and achieves exponential convergence rather than algebraic $\mathcal{O}(N^{-2})$ or $\mathcal{O}(N^{-4})$ Simpson error rates.

Related Quantitative Finance & Fourier Valuation Labs

Carr-Madan FFT Option Pricing Lab

Invert characteristic functions across full strike surfaces simultaneously in $\mathcal{O}(N \log N)$ operations.

Carr-Madan Option Replication Lab

Synthesize generalized nonlinear payoffs and variance swaps through static option continuum portfolios.

Heston Stochastic Volatility Lab

Model mean-reverting CIR variance processes, leverage correlation skews, and Feller boundary condition dynamics.

Longstaff-Schwartz LSM Lab

Value American options via backward induction, dynamic early exercise frontiers, and OLS regression.