Tempered Stable Lévy Noise

• Tempered stable Lévy noise is a class of infinitely divisible pure-jump processes derived by tempering α-stable laws with an exponential cutoff to ensure finite moments.
• These processes interpolate between heavy-tailed α-stable noise and Gaussian models, effectively capturing large but bounded jumps in various systems.
• Applications in finance and physics benefit from using these models for accurate volatility calibration and realistic simulation of anomalous transport phenomena.

Tempered stable Lévy noise refers to a class of infinitely divisible pure-jump stochastic processes whose Lévy measure is a deformation of the α-stable law via an exponential (or more general) cutoff. This construction retains power-law tails at intermediate scales while ensuring that all moments are finite due to the tempering effect. Tempered stable Lévy processes interpolate between heavy-tailed, infinite activity α-stable noise and exponentially light-tailed (or even Gaussian) processes, and are foundational in modeling a wide range of systems with rare, large jumps whose amplitude is physically or statistically bounded. The classical Normal-Tempered-Stable (NTS), Variance Gamma (VG), and Normal-Inverse-Gaussian (NIG) subclasses provide key examples. Applications are prominent in physics (turbulence, anomalous transport), finance (asset dynamics, volatility modeling), and stochastic dynamics out of equilibrium.

1. Definition and Mathematical Structure

Tempered stable Lévy processes generalize the α-stable family by introducing a tempering function into the Lévy measure, typically of exponential type. The one-dimensional Lévy density for an α-stable law ($0 < \alpha < 2$) is

$$\nu(x) = C_+ x^{-1-\alpha} 1_{x>0} + C_- |x|^{-1-\alpha} 1_{x<0},$$

which yields infinite variance for $\alpha < 2$ and infinite mean for $\alpha < 1$. Tempering modifies this to

$$\nu(x) = C_+ x^{-1-\alpha} e^{-\lambda_+ x} 1_{x>0} + C_- |x|^{-1-\alpha} e^{-\lambda_- |x|} 1_{x<0},$$

where $\lambda_+, \lambda_- > 0$ are tempering parameters. For $\alpha = 0$, this recovers the Variance Gamma (VG) process; $\alpha = 1/2$ yields the NIG process—all as special cases of NTS models (Azzone et al., 2019).

The associated characteristic (Lévy–Khintchine) exponent for the NTS case is

$$\psi(u) = i u \varphi + \int_{\mathbb{R}} \left(e^{i u x} - 1 - i u x 1_{|x|<1}\right) \nu(x) dx,$$

with explicit forms available for tempered subordinators and time-changed Brownian constructions. Tempered stable laws always remain infinitely divisible and possess transition densities expressible using special functions (e.g., modified Bessel functions for NTS (Azzone et al., 2019)).

2. Stochastic Process Construction

A tempered stable process $X_t$ may be constructed as a time-change of Brownian motion: $X_t = \mu S_t + \sigma W_{S_t},$ where $S_t$ is an $\alpha$-stable subordinator with exponentially tempered jumps: $$\mathbb{E}[e^{-q S_t}] = \exp \left\{ - t \phi_S(q) \right\}, \qquad \phi_S(q) = \frac{1-\alpha}{\alpha k} \left[(\lambda + q)^{\alpha} - \lambda^{\alpha} \right],$$ with $k, \lambda > 0$ and $W$ standard Brownian motion. This construction recovers the NTS and VG cases as limiting forms for $\alpha = 1/2$ and $\alpha = 0$, respectively (Azzone et al., 2019).

The process is termed additive if the parameters $(\mu, \sigma, k, \lambda)$ are allowed to be time-dependent and deterministic, yielding independent but non-stationary increments (Azzone et al., 2019). This enables the exact fit of options-implied volatility surfaces across maturities and assists in modeling time-inhomogeneous jump activities.

3. Moments, Scaling Laws, and Asymptotics

Finite moments are a key property distinguishing tempered stable laws from their α-stable progenitors. For increments $f_t$, expanding the cumulant generating function yields: $$\kappa_1(t) = \varphi_t t - m_t t, \qquad \kappa_2(t) = [\sigma_t^2 + k_t m_t^2] t,$$ where $m_t = (\frac{1}{2}+\eta_t)\sigma_t^2$ (Azzone et al., 2019). For power-law parameterizations $k_t = \bar{k} t^{\beta}$, $\eta_t = \bar{\eta} t^{\delta}$, one finds

$$v_t = \bar{\sigma}^2 + \bar{k} \bar{\sigma}^4 t^{\beta} \left( \frac{1}{2} + \bar{\eta} t^{\delta} \right)^2$$

and the variance of increments follows accordingly. Small-time increments retain the α-stable scaling, with jumps dominating, while large-time asymptotics approach Gaussianity via the central limit theorem, with the diffusion term eventually dominating (Azzone et al., 2019). This bridges heavy-tail effects at fine scales and regular behavior at macro scales.

4. Calibration and Empirical Performance

Calibrating additive NTS or VG models to equity implied volatility surfaces, one finds that all maturities $T$ yield parameters $k_T, \eta_T, \sigma_T$ obeying scaling laws: $$\hat{k}_\theta \approx \bar{k} \theta^{\beta}, \qquad \hat{\eta}_\theta \approx \bar{\eta} \theta^{\delta}, \qquad \theta = T \sigma_T^2,$$ with empirical fits showing $\beta \approx 1$, $\delta \approx -1/2$, and strictly positive scale constants. Across all maturities, this results in substantially better calibration (mean squared error improved by two orders of magnitude) than stationary Lévy or self-similar alternatives. The full volatility smile and skew can be reproduced both at short (few days) and long (years) maturities (Azzone et al., 2019).

5. Applications and Implications

Tempered stable Lévy noise models are widely used in financial mathematics, especially for modeling asset price returns and the stochastic volatility surfaces observed in equity derivatives, matching both the empirically observed heavy tails and the regularization at extreme values. In addition, these processes model anomalous transport and turbulence in physics, where bounded energy or finite domain sizes require a mechanism to temper otherwise unphysical jump amplitudes with realistic cutoffs (Azzone et al., 2019). The construction as additive processes with independent increments and scalable parameters allows for consistent fitting and forecasting in both empirical data and simulation-driven theoretical studies.

6. Extensions and Related Classes

Tempered stable families admit further generalization, such as the use of geometric or Mittag-Leffler tempering functions, which allow for tunable intermediate regimes between pure power-law and exponential decay (Torricelli, 2023). Properties such as absolute continuity with respect to the parent stable law, spectral density characterization, and explicit cumulant formulas (expressed via special functions) are established, ensuring analytic tractability and simulation feasibility. The correspondence between tempered subordinators and the time-changed Brownian framework underpins both practical implementation and theoretical analysis.

7. Summary Table: Core Properties of Tempered Stable Lévy Noise

Feature	α-stable	Tempered Stable	Gaussian (limit)
Decay of tails	$|x|^{-1-\alpha}$	$|x|^{-1-\alpha} e^{-\lambda|x|}$	$e^{-x^2/(2\sigma^2)}$
Moments	Some diverge	All finite	All finite
Infinite activity?	Yes	Yes	No
Limiting behavior	Heavy-tailed	Interpolates, Gaussian for large λ	Gaussian
Calibration to equity	Poor for large T	Excellent, robust	Poor for short T

• (Azzone et al., 2019): Additive normal tempered stable processes for equity derivatives and power law scaling
• (Torricelli, 2023): Tempered geometric stable distributions and processes