Why a specific formulation of the Heston pricing integral eliminates a numerical pitfall that causes wrong option prices in production systems
Fourier-based pricing is the workhorse for stochastic volatility models. For the Heston model, the standard approach is to express the option price as an integral of the model's characteristic function over the real line, then evaluate that integral numerically. The procedure is fast and exact in principle. In practice, however, a specific numerical issue — a discontinuity in the complex logarithm arising from its multi-valued nature — has caused materially wrong option prices in live pricing systems, particularly for long maturities and high volatility-of-volatility regimes. This paper by Dr. Ingo Fahrner, published in the International Journal of Theoretical and Applied Finance (2007), provides a clean theoretical proof that the problem does not arise when the integral is written in the correct form.
The Heston setup. The Heston model specifies joint dynamics for the asset price and its instantaneous variance: dSt = rSt dt + √vt St dWt, dvt = κ(θ − vt) dt + ξ√vt dZt, dWt dZt = ρ dt. Option prices are computed by Fourier inversion: the price is written as an integral over real u of the characteristic function φ(u, T) = 𝔼[eiu ln ST]. The characteristic function has a closed form that involves a complex square root and, critically, a complex logarithm of an expression that depends on u.
The branch-cut problem. The complex logarithm is multi-valued. Every implementation must choose a branch, and the standard choice restricts ln to its principal value, which is discontinuous along the negative real axis. In the original Heston form of the characteristic function, as u increases from 0 to ∞ along the real integration axis, the argument of the logarithm can rotate in the complex plane and cross the negative real axis. Each crossing introduces a jump of 2πi in the integrand. That jump propagates directly into the computed option price as a discrete error. For parameter combinations with high vol-of-vol ξ, strong correlation |ρ|, or long maturities T, such crossings occur within the standard integration range, producing prices that are wrong by a non-trivial and parameter-dependent amount.
The Lewis-Lipton form. The Lewis-Lipton pricing formula writes the same integral in a mathematically equivalent but analytically different factoring of the characteristic function. The key structural difference is the position of the branch cut relative to the integration path. In the Lewis-Lipton form, the expression inside the complex logarithm, as a function of real u, traces a path in the complex plane that stays strictly away from the negative real axis. The argument never crosses the branch cut because the real part of the logarithm's argument remains bounded away from zero throughout the integration range.
The proof. Fahrner demonstrates this with a direct and elementary argument. For real u, the logarithm's argument in the Lewis-Lipton form can be written as a ratio of quantities whose real part satisfies an explicit lower bound that depends only on the model parameters κ, ξ, ρ, and the time horizon T. The bound shows that the real part remains positive for all real u, which means the argument is always in the right half of the complex plane: the principal branch of ln is continuous there, and no tracking of rotations or winding numbers is required. The result holds for all admissible parameter values, all strikes, and all maturities.
The practical bottom line. The paper resolves a numerical stability problem that is not obvious from the model equations but has caused pricing errors in production systems. The fix requires no algorithmic change to the integration routine and no branch-tracking logic: it follows entirely from choosing the right formulation of the characteristic function before integrating. For any team implementing or validating Heston pricing via Fourier inversion, the prescription is direct: use the Lewis-Lipton form, apply the standard principal branch of the complex logarithm, and integrate along the real axis in the usual way. The result is numerically correct and stable across all parameter regimes, without additional complexity or computational overhead.
