Introduction
The Cox-Ingersoll-Ross (CIR) process is a cornerstone of mathematical finance, most notably serving as the stochastic volatility component in the Heston model. Defined by the stochastic differential equation (SDE):
$$dX_t = (\delta - \gamma X_t) dt + \beta \sqrt{X_t} dW_t$$
where $\delta, \gamma, \beta > 0$, the CIR process is prized for its ability to model mean-reverting, non-negative quantities like interest rates and volatility. However, the "square-root" diffusion coefficient $\beta \sqrt{X_t}$ presents a significant challenge for numerical analysis. Because the square root function is not Lipschitz continuous at zero—its derivative tends toward infinity—standard convergence proofs for numerical schemes like the Euler-Maruyama method do not apply.
The Challenge of the Accessible Boundary
The behavior of the CIR process near zero is governed by the relationship between its parameters, often summarized by the Feller condition. If $2\delta \ge \beta^2$, the process never reaches the boundary point zero (the boundary is "inaccessible"). In this regime, positive strong convergence rates for numerical approximations have been well-established.
However, empirical calibrations of the Heston model frequently yield parameters where $2\delta < \beta^2$. In this case, the process can and does hit zero (the boundary is "accessible"). Until recently, establishing a positive strong convergence rate for numerical approximations in this accessible regime remained a significant open problem in the field. This paper addresses this gap, focusing on the drift-implicit square-root Euler approximation.
Methodology: Drift-Implicit Schemes and Bessel Processes
The authors analyze the drift-implicit square-root Euler scheme proposed by Alfonsi (2005). This specific scheme is designed to preserve the non-negativity of the process, a physical requirement for financial variables. The scheme's stability is particularly valuable when the process approaches the non-Lipschitz boundary at zero.
The theoretical breakthrough in this paper stems from a deep investigation into the temporal regularity of Bessel processes. Since a CIR process can be viewed as a squared Bessel process (through a specific transformation), the smoothness of the Bessel process over time is directly linked to the convergence of CIR approximations.
Key Theoretical Steps:
- Temporal Regularity: The authors prove that Bessel processes are temporally $1/2$-Hölder continuous in $L^p$ for all $p \in (0, \infty)$. This confirms that the process behaves as smoothly as a standard Brownian motion in an $L^p$ sense, even when interacting with the zero boundary.
- Inverse Moment Estimates: To handle the singularity at zero, the authors developed new estimates for the exponential inverse moments of Bessel-type processes.
- Error Analysis: By combining these regularity results with the properties of the drift-implicit scheme, the authors translate the temporal smoothness of the process into a formal convergence rate.
Key Results
The paper provides two primary theorems that advance our understanding of CIR approximations:
1. Temporal Regularity (Theorem 1.2)
For any $p \in (0, \infty)$, the Bessel process $Z$ satisfies: $$\sup_{s \neq t} \frac{\|Z_t - Z_s\|_{L^p}}{|t-s|^{1/2}} < \infty$$ This result is foundational, as it establishes the necessary regularity to bound the error introduced by discrete-time approximations.
2. Strong Convergence (Theorem 1.1)
In the parameter regime where $2\delta/\beta^2 > 1/2$, the authors prove a positive strong $L^p$-convergence rate for the drift-implicit square-root Euler approximation. Specifically, the error is bounded by $C \cdot h^\alpha$, where $h$ is the step size and $\alpha$ is a positive exponent determined by the process parameters. This covers the "upper half" of the accessible boundary regime ($1/2 < 2\delta/\beta^2 < 1$).
Implications and Future Work
This research provides the first theoretical guarantee for the accuracy of CIR approximations in a regime where the process hits zero. For practitioners using the Heston model, this justifies the use of implicit schemes for pricing derivatives even when the Feller condition is violated.
However, several questions remain:
- The Lower Regime: For the highly volatile regime where $2\delta/\beta^2 \le 1/2$, a positive strong convergence rate is still unproven.
- Rate Sharpness: Numerical simulations often suggest faster convergence than what can currently be proven theoretically. Future research may close this gap between observed performance and theoretical bounds.
- Path-Dependency: Extending these results to path-dependent options (such as Asian or Lookback options) would further bridge the gap between numerical analysis and practical financial engineering.
Summary updated