Stochastic processes with jumps, such as Lévy processes, are widely used to model discontinuous phenomena in finance, physics, and engineering. A Lévy process is completely characterized by its drift vector, its diffusion matrix, and its Lévy measure—a function that dictates the frequency and size of its jumps.
While the forward problem—computing the distribution or characteristic function from a known Lévy measure—is straightforward via the classical Lévy-Khintchine formula, the inverse problem is notoriously difficult. In practice, researchers often only observe a process at discrete, low-frequency intervals, yet they need to recover the underlying, continuous-time jump behavior.
The core insight of this research direction is to operate entirely within the transform domain (spectral domain). By estimating characteristic functions or moment generating functions from data and applying specialized, regularized inversion formulas, it is possible to map frequency-domain representations directly back to latent jump measures.
Bypassing Densities via the Cumulant Function
In classical probability, the Gil-Pelaez or Lévy inversion formulas recover a probability density function (PDF) from a characteristic function. However, recovering a Lévy measure is fundamentally different because the measure can have a singularity at zero, representing an infinite number of infinitesimally small jumps.
To bridge this gap, Evgeny Burnaev developed a direct analytical pathway in Формула обращения для безгранично делимых распределений and Inversion formula for infinitely divisible distributions. Instead of inverting the characteristic function directly, the author introduces a modified version of the cumulant function (the logarithm of the characteristic function), denoted as $\hat{\psi}(z)$.
By manipulating this function and its derivatives, the singularity at zero is isolated and regularized. This transforms the problem into a modified Fourier transform of a specific spectral function. This mathematical bridge allows standard inversion techniques to extract the precise jump intensity directly from frequency-domain data, bypassing the need for indirect or approximate time-domain calculations.
Stabilizing the Inverse Problem against Noise
While analytical formulas work perfectly with theoretical characteristic functions, statistical inference introduces empirical noise. Empirical characteristic functions estimated from finite, low-frequency data fluctuate wildly at high frequencies. When passed through an inverse Fourier transform, these high-frequency fluctuations blow up, making direct inversion highly unstable and ill-posed.
To resolve this, Denis Belomestny introduced regularized spectral estimation frameworks. In Statistical inference for time-changed Lévy processes via composite characteristic function estimation, the author leverages composite characteristic functions to establish a stable relationship between observed time-changed processes and latent Lévy processes. Similarly, in Spectral estimation of the Lévy density in partially observed affine models, the author transforms the estimation problem by exploiting the log-affine structure of conditional characteristic functions, facilitating stable estimation even when direct observation of the process is limited.
By applying a regularized inverse Fourier transform to the estimated characteristic exponent, Tatiana Orlova et al. demonstrate in Statistical inference for moving‐average Lévy‐driven processes: Fourier‐based approach that these spectral estimators can achieve optimal convergence rates. The regularization acts as a low-pass filter, dampening high-frequency empirical noise while preserving the structural information of the underlying Lévy density.
Generalizing to Other Integral Transforms
The intuition of mapping transform-domain properties back to spatial distributions extends beyond standard Fourier-based characteristic functions. Depending on how the stochastic process is framed, other integral transforms offer more natural coordinate systems to decouple the underlying dynamics:
- Conditional Characteristic Functions (CCFs): When dealing with Markov processes where transition densities lack closed-form expressions, Liang Peng et al. in Parameter estimation and model testing for Markov processes via conditional characteristic functions utilize CCFs within an empirical likelihood framework. This allows for robust parameter estimation of complex dynamics, such as infinite-activity jump processes, directly from the Fourier transform of the transition probability.
- Mellin Transforms: For the statistical Skorohod embedding problem—where the goal is to recover the distribution of a random time $T$ from a time-scaled Brownian motion—John Schoenmakers in Statistical Skorohod embedding problem: Optimality and asymptotic normality constructs a density estimator via an empirical Mellin transform. A regularized inverse Mellin transform, incorporating a kernel function and a bandwidth parameter, successfully stabilizes the underlying deconvolution.
- Laplace Transforms: In signal processing contexts, Sung Sik Nam et al. in An MGF-Based Unified Framework to Determine the Joint Statistics of Partial Sums of Ordered i.n.d. Random Variables derive joint Moment Generating Functions (MGFs) for ranked inputs. Applying an inverse Laplace transform to these MGFs provides a systematic pathway to recover the real-valued joint probability density functions.
Across all these facets, the core mechanism remains identical: map a complex, convolution-heavy spatial problem into a transform domain where the latent components decouple algebraically, and then apply a regularized inversion operator to safely navigate high-frequency noise and singularities.
Key Foundational Work
Inversion formula for infinitely divisible distributions (2006) — contributed 9 concept elements to this direction.
Формула обращения для безгранично делимых распределений (2006) — contributed 5 concept elements to this direction.
Spectral estimation of the Lévy density in partially observed affine models (2011) — contributed 2 concept elements to this direction.