Reader question: If an option model gives us a characteristic function rather than a closed-form option price, how can one algorithm convert that frequency-domain object into prices for hundreds or thousands of strikes at once?
The Carr–Madan method answers by reframing European option pricing as a Fourier-inversion problem and then exploiting the Fast Fourier Transform. The key move is not simply “apply an FFT.” A raw call-price function is not integrable enough for the transform Carr and Madan need. The algorithm first damps the call price, builds a transform that can be written directly from the risk-neutral characteristic function, discretises the frequency integral, and then uses an FFT to recover an entire log-strike grid in one operation.
This article owns the characteristic-function-to-strike-grid pricing problem: a risk-neutral characteristic function + discounting + numerical grid choices → European option prices across many strikes, together with numerical diagnostics for damping, truncation, aliasing, quadrature, interpolation and convergence.
It does not own the stochastic model that generates the characteristic function, the estimation of Heston parameters, early exercise, path-dependent contracts, or the construction of the volatility surface. Those are separate jobs. The purpose here is computational mathematics: how a Fourier representation becomes a reliable pricing engine.
This is public mathematical education, not personalised financial advice or a trading recommendation.
1. The characteristic function is often easier than the density
Let XT = ln ST. Under a risk-neutral model, suppose we know the characteristic function:
φ(u) = E[eiuXT].
For Black–Scholes this is immediate. For models such as Heston, Variance Gamma and other affine or Lévy specifications, the characteristic function can remain available in closed form even when the terminal density or option price does not have a comparably simple expression.
A characteristic function is therefore a compact description of the distribution in the frequency domain. The Carr–Madan algorithm asks: can we recover option prices from it efficiently?
2. Why the raw call-price transform is troublesome
Write log strike as:
k = ln K.
Let C(k) be the time-0 price of a European call with log strike k. For very negative k, the strike approaches zero and the call value approaches the asset value rather than decaying to zero. That prevents the raw call-price function from having the integrability properties needed for the direct Fourier transform used by the algorithm.
Carr and Madan therefore introduce a damping factor:
cα(k) = eαkC(k),
with α > 0 chosen so that the damped call-price function is sufficiently integrable.
This is the first important lesson: the damping parameter is part of the mathematical validity of the transform, not merely a cosmetic numerical tuning knob.
3. The transformed call price
Under the standard Carr–Madan convention, the Fourier transform of the damped call price can be expressed through the characteristic function of ln ST:
ψ(v) = e−rT φ(v − i(α+1)) / [α² + α − v² + i(2α+1)v].
The call price is then recovered from:
C(k) = e−αk/π ∫0∞ Re[e−ivkψ(v)]dv.
The formula separates the problem into two parts:
- the economic model, contained in φ;
- the numerical inversion machinery, contained in the damping, grid and Fourier transform.
This modularity is why one FFT engine can be reused across many characteristic-function models.
4. The moment condition behind α
Evaluating φ(v − i(α+1)) effectively asks the model for a positive moment of ST. A valid damping parameter must lie inside a strip where that complex-valued characteristic function exists and remains numerically stable.
So the algorithm cannot choose α independently of the model. A value that works for Black–Scholes may fail for a heavy-tailed model if the required moment does not exist.
Falsifier: if prices change materially across nearby valid damping values, or if the characteristic function becomes singular or explosive at the shifted argument, the chosen damping regime is not trustworthy.
5. Discretise the frequency axis
Choose N frequency points:
vj = jη, j = 0,1,…,N−1,
where η is the frequency spacing.
The infinite integral is truncated at approximately:
vmax ≈ (N−1)η.
That immediately creates a trade-off.
- Small η samples the transform finely but, for fixed N, lowers the maximum sampled frequency.
- Large η extends the frequency range but coarsens the sampling.
- Larger N can improve both reach and resolution, but increases cost and memory.
The FFT itself is fast. Choosing a grid that represents the integral well is the harder problem.
6. The strike-grid spacing is not independent
The discrete Fourier transform couples the frequency grid to the log-strike grid. A common relation is:
λ = 2π/(Nη),
where λ is log-strike spacing.
If the log-strike grid is written:
km = −b + mλ,
then a common symmetric choice is:
b = Nλ/2.
This is a subtle but essential constraint: reducing frequency spacing η increases log-strike spacing λ unless N is also increased. A user cannot demand arbitrary strike resolution, frequency resolution and finite array length simultaneously.
7. One FFT gives many strikes
After quadrature weights and a phase-shift factor are applied, the discretised integral takes the form of a discrete Fourier transform. The FFT evaluates that transform in approximately:
O(N log N)
operations rather than the O(N²) work of evaluating every strike-frequency pair directly.
The efficiency matters most when the use case needs many strikes for the same maturity and model parameters. In calibration, the model may be repriced repeatedly while an optimiser changes parameters. One transform can populate an entire strike strip at once.
For a single isolated strike, direct numerical integration can be competitive or simpler. The algorithm’s comparative advantage comes from batch structure.
8. Quadrature still matters inside an FFT
The FFT accelerates a sum, but the sum is still approximating an integral. The frequency integral therefore needs quadrature weights.
Common implementations use trapezoidal or Simpson-type weights. Simpson weighting can reduce quadrature error for smooth integrands, but it does not cure poor truncation, invalid damping, aliasing or characteristic-function instability.
A useful diagnostic is to run the same grid with two reasonable quadrature rules. If the option prices differ materially, the frequency grid is not yet resolving the integral well enough.
9. Aliasing is a periodicity problem
A discrete Fourier transform behaves as though the sampled function repeats periodically. If the effective strike domain is too narrow or the transform has not decayed enough at the frequency boundary, information can wrap around the finite grid and contaminate prices.
This is aliasing.
It can appear as oscillatory prices, negative values for deeply out-of-the-money calls, or strange behaviour near the ends of the strike grid.
Diagnostic: enlarge N, alter η, or shift the strike domain. A true economic price should not move materially simply because the periodic numerical box was widened.
10. Truncation error is different from aliasing
The theoretical integral extends to infinity. The numerical integral stops at vmax. If ψ(v) has not decayed enough by that point, the omitted tail creates truncation error.
A practical convergence check is to increase vmax while holding the strike region of interest well resolved. If prices continue to move, the original truncation was inadequate.
Aliasing and truncation can interact, but they are conceptually different:
- truncation: the missing high-frequency tail;
- aliasing: periodic wrap-around from discrete sampling.
11. Interpolation is part of the price
The FFT returns values only at its own log-strike nodes. Market strikes rarely line up exactly with those nodes.
An implementation therefore interpolates from the FFT grid to the requested strike. Linear interpolation is simple and often adequate on a fine grid, while spline-based interpolation can be smoother but may introduce overshoot.
This creates another verification requirement:
price-grid error and interpolation error must be separated.
One way is to test at strikes that lie exactly on the FFT grid and then compare off-grid interpolation against a high-accuracy direct-integral benchmark.
12. A minimal algorithm
- Choose maturity, rates and model parameters.
- Construct the risk-neutral characteristic function φ(u).
- Select a valid damping parameter α.
- Choose N and frequency spacing η.
- Compute the linked log-strike spacing λ and strike range.
- Evaluate ψ(vj) on the frequency grid.
- Apply the phase shift and quadrature weights.
- Run the FFT.
- Take the real part, undo the damping and map log strikes back to strikes.
- Interpolate to requested strikes.
- Check arbitrage bounds and benchmark prices.
- Repeat on a refined grid before trusting production output.
13. Inputs and outputs
Inputs can include:
- spot or forward price;
- discount factors or rate curve;
- dividend or carry assumptions;
- maturity;
- model parameters;
- risk-neutral characteristic-function implementation;
- damping parameter α;
- FFT length N;
- frequency step η;
- quadrature rule;
- strike-range centring convention;
- interpolation method.
Outputs can include:
- log-strike grid;
- strike grid;
- call-price grid;
- put prices derived consistently if required;
- interpolated target-strike prices;
- grid-convergence diagnostics;
- arbitrage-bound flags;
- timing and numerical-error comparisons.
14. Benchmark first with Black–Scholes
Black–Scholes provides an ideal unit test because both its characteristic function and closed-form European option price are known.
Run the FFT engine using the Black–Scholes characteristic function and compare the output against the analytic formula over:
- deep in-the-money strikes;
- near-the-money strikes;
- deep out-of-the-money strikes;
- short maturities;
- long maturities;
- low and high volatility regimes.
One correct at-the-money price is weak evidence. A full error surface across strikes and maturities is much stronger.
15. Heston is a second benchmark
Heston’s stochastic-volatility model has a closed-form characteristic function and well-known semi-analytic pricing representations. That makes it a useful cross-check after Black–Scholes.
If the FFT and an independently coded Heston integral agree across a grid, confidence in both the characteristic-function shift and the FFT discretisation rises.
See Heston calibration algorithms for the model-calibration problem that sits above the pricer.
16. Evidence polarity
Evidence for confidence includes agreement with analytic Black–Scholes values, agreement with an independent direct Fourier integral, stable prices under moderate changes in α, N and η, small interpolation error, nonnegative call prices, monotonicity in strike, convexity in strike where appropriate, and smooth convergence when the grid is refined.
Evidence against confidence includes negative option values, oscillations near the grid edges, prices that move sharply with damping, large differences between Simpson and trapezoidal quadrature, material sensitivity to the strike-domain shift, failure of put–call parity, or a transform that becomes unstable at the shifted characteristic-function argument.
17. Counterexample: α is valid numerically but invalid mathematically
A programmer may discover that α = 4 produces finite machine numbers and therefore assume it is safe. But if the model does not possess the corresponding positive moment, the shifted characteristic function is outside its valid transform strip.
The code has produced a number, not a justified price.
Falsifier: verify the model’s moment/exponential-integrability condition and test nearby damping values inside the valid strip.
18. Counterexample: coarse strike spacing hides local smile structure
A large frequency step can produce a very coarse log-strike grid. Interpolation may then smooth over local curvature that matters for calibration.
Falsifier: double N or reduce λ while preserving adequate frequency reach. If fitted model parameters change materially, the original calibration was partly fitting interpolation error.
19. Counterexample: a larger N does not fix every error
Increasing N while leaving η and damping poorly chosen may extend the grid without resolving the numerical problem that matters.
Grid design is multidimensional. Diagnostics should vary N, η, α and strike-centering separately rather than treating “more FFT points” as a universal cure.
20. Counterexample: exact FFT, wrong characteristic function
The FFT can invert the supplied transform perfectly while the characteristic function itself is coded with a wrong sign, wrong drift, wrong branch convention or inconsistent dividend treatment.
Falsifier: evaluate the characteristic function at known identities such as φ(0)=1, verify martingale conditions where applicable, and benchmark the same function through an independent pricing route.
21. Counterexample: European machinery applied to American exercise
The basic Carr–Madan transform prices European payoffs. Early exercise changes the valuation problem because the holder can stop before maturity.
If the contract is genuinely American, a direct European FFT price is solving the wrong problem.
See finite-difference option-pricing algorithms and American-option binomial trees for methods that naturally represent early exercise.
22. Alternatives
Direct Fourier integration prices one strike at a time and is useful as an independent benchmark.
COS methods expand the density or payoff in a Fourier-cosine basis and can converge rapidly for many models.
Fractional FFT relaxes part of the rigid spacing relation between the input and output grids.
Finite-difference PDEs are attractive for low-dimensional local or stochastic-volatility models with boundaries and early exercise.
Monte Carlo handles high dimension and path dependence but gives noisy estimates unless variance is carefully controlled.
The algorithm should follow the contract structure and required output, not simply the existence of a fast transform.
23. Connections to the Bukit Timah Tutor knowledge estate
Black–Scholes option-pricing algorithms provide the clean analytic benchmark.
Heston calibration algorithms show how a characteristic-function pricer sits inside a nonlinear parameter-fitting loop.
Monte Carlo pricing algorithms provide a contrasting numerical route based on path simulation rather than transform inversion.
The lane hub is Finance & Banking Algorithms | Applied Mathematics in Real Financial Systems.
24. What would falsify confidence?
Confidence should be withdrawn if analytic benchmarks are not recovered; if prices fail arbitrage bounds; if results move materially under modest valid changes to damping or grid parameters; if high-frequency truncation remains material; if grid-edge aliasing contaminates the strike region of interest; if interpolation dominates the error; if characteristic-function identities fail; or if the contract requires features the European transform does not model.
25. Verification and update triggers
Preserve the characteristic-function version, damping parameter, FFT length, frequency step, strike-grid mapping, quadrature rule, interpolation method and benchmark error surface for every release. Revalidate after changes to the characteristic function, discounting convention, carry model, FFT library, complex-number implementation or interpolation code.
Trigger review when calibration starts selecting extreme parameter values, grid-edge prices become negative, convergence deteriorates, new products require wider strike or maturity ranges, or a model’s moment conditions change the admissible damping strip.
26. Primary and high-quality references
- Peter Carr and Dilip B. Madan, Option Valuation Using the Fast Fourier Transform, Journal of Computational Finance, 1999.
- Steven L. Heston, A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options, Review of Financial Studies, 1993.
- Fang Fang and Cornelis W. Oosterlee, A Novel Pricing Method for European Options Based on Fourier-Cosine Series Expansions, SIAM Journal on Scientific Computing.
- Federal Reserve Board and OCC, SR 11-7 — Guidance on Model Risk Management, for independent validation and ongoing monitoring principles.
Educational boundary: FFT accuracy is a numerical property. It does not prove that the risk-neutral model generating the characteristic function is empirically adequate.

