Beyond CARMA: Building Complex Processes from Lévy Noise

by

Author: Denis Avetisyan

This review introduces supCARMA processes-superpositions of continuous-time autoregressive moving average models driven by Lévy processes-and details their theoretical properties.

The correlation functions of supCAR(2)(2)-III processes demonstrate a transition in behavior based on shape parameters; when <span class="katex-eq" data-katex-display="false">\alpha + 3 \in (3,4]</span>, these processes exhibit non-oscillatory long-range dependence, but as the shape parameter increases beyond this range, oscillations emerge, indicating a shift in the underlying dynamic characteristics.
The correlation functions of supCAR(2)(2)-III processes demonstrate a transition in behavior based on shape parameters; when \alpha + 3 \in (3,4], these processes exhibit non-oscillatory long-range dependence, but as the shape parameter increases beyond this range, oscillations emerge, indicating a shift in the underlying dynamic characteristics.

The paper investigates existence conditions, stationarity, and correlation structures of supCARMA processes based on their underlying CAR(2)(2) matrix and infinitely divisible driving noise.

While existing time series models often struggle to capture complex dependencies and non-monotonic correlations, this paper introduces a flexible class of processes-supCARMA-defined as superpositions of Lévy-driven CARMA processes, extending the well-established supOU framework. Specifically, we analyze supCAR$(2)$ processes, revealing that their characteristics are fundamentally determined by the eigenstructure of the underlying CAR$(2)$ matrix, leading to distinct types with associated existence conditions and explicit correlation functions. These resulting processes can exhibit both long-range dependence and non-monotonicity, offering a powerful tool for modeling oscillatory or strongly dependent time series data-but how might these supCARMA processes be adapted for high-dimensional data analysis and real-time forecasting?

Unveiling Continuous Dynamics: The SupCARMA Process

Numerous processes in nature and various scientific fields unfold continuously, rather than in distinct steps; consider the gradual diffusion of a substance, the fluctuating price of a stock, or even the subtle shifts in physiological states. Traditional modeling techniques often rely on discrete-time approximations, dividing continuous phenomena into fixed intervals to facilitate computation. However, this simplification can introduce significant inaccuracies, particularly when capturing rapid changes or subtle dependencies. These discrete models struggle to represent the inherent smoothness and interconnectedness of continuous dynamics, potentially overlooking crucial information and leading to flawed predictions. Consequently, a need exists for methodologies capable of directly addressing the continuous nature of these phenomena, allowing for a more faithful and precise representation of their behavior and underlying mechanisms.

The SupCARMA process distinguishes itself by representing dynamic systems not through difference equations, but through continuous-time integrals. This integral representation allows for a remarkably flexible approach to modeling, accommodating a wider range of temporal dependencies than traditional discrete-time methods. Instead of approximating change as occurring at specific points in time, the process describes evolution as the continuous accumulation of infinitesimal changes driven by random measures. This means that the model can capture phenomena where events occur at any moment, and their influence persists continuously – a feature crucial for accurately simulating processes in fields like finance, physics, and biology. The framework’s adaptability stems from its ability to incorporate diverse “driving” random measures, effectively tailoring the model’s behavior to match the nuances of the system under investigation and offering a powerful alternative for analyzing continuous-time phenomena.

The SupCARMA process distinguishes itself through its utilization of infinitely divisible random measures as the engine driving its dynamic behavior. These measures, unlike their discrete counterparts, allow for variations of any size, enabling the model to capture subtle, continuous fluctuations inherent in many real-world systems. This characteristic is not merely a mathematical detail; it unlocks a remarkably diverse family of possible models. By carefully selecting the properties of this driving random measure – its intensity, covariance structure, and higher-order moments – researchers can tailor the SupCARMA process to precisely mimic the temporal dependencies observed in a wide range of phenomena, from financial markets to neuronal activity. The flexibility arises because these measures offer a fundamentally richer set of possibilities for generating stochastic variation than traditional approaches, facilitating a more nuanced and accurate representation of continuous-time dynamics.

The SupCARMA process gains its mathematical foundation from the concept of a Lévy Basis, a cornerstone of modern stochastic analysis. A Lévy Basis extends the properties of Brownian motion to encompass jumps and more general stochastic behaviors, providing a flexible framework for modeling unpredictable changes over time. Specifically, the SupCARMA process is constructed as the integral of a Lévy Basis with respect to a predictable measure, allowing it to capture a wide range of continuous-time dynamics. This connection isn’t merely technical; it means the well-established theory surrounding Lévy processes – including techniques for analyzing their properties and simulating their behavior – directly applies to understanding and utilizing the SupCARMA process. Consequently, this linkage facilitates the development of robust statistical inference and forecasting methods for systems modeled using this integral representation, making it a powerful tool for analyzing phenomena exhibiting complex, continuous-time evolution.

Decoding Dynamic Characteristics: Eigenvalues and State-Space Representation

The eigenvalues of the matrix A in a SupCARMA (Superparameterized CARMA) process directly determine the process’s dynamic characteristics, including its stability and the rate at which correlations decay. Specifically, if all eigenvalues of A have magnitudes strictly less than one, the process is guaranteed to be stationary and mean-reverting. The location of the eigenvalues in the complex plane dictates the oscillatory behavior of the process; complex conjugate pairs induce oscillations, while real negative eigenvalues govern the rate of convergence to the mean. Furthermore, the spectral density of the process, which describes the distribution of variance across different frequencies, is fundamentally shaped by the eigenvalues of A, with eigenvalues closer to the unit circle corresponding to stronger low-frequency components. \lambda_i represents each eigenvalue of A, and its magnitude determines the contribution of that mode to the overall process behavior.

A state-space representation, also known as a space-state model, provides a framework for modeling the SupCARMA process as a system of first-order difference equations. This involves defining a state vector, \mathbf{x}_t , that encapsulates the process’s relevant past information, and relating its current value to both past states and current and past noise terms. Specifically, the process is expressed through two equations: a state equation that describes the evolution of \mathbf{x}_t and an observation equation that links the state to the observed time series. This formulation allows for the application of powerful tools from control theory and time-series analysis, enabling a detailed investigation of the process’s internal dynamics, including its stability, response to initial conditions, and prediction properties. The state-space form also facilitates efficient computation of quantities such as the autocovariance function and spectral density, which are crucial for characterizing the process’s statistical properties.

A state-space representation of the SupCARMA process enables detailed analysis of its stability through examination of the system matrix. Specifically, the eigenvalues of this matrix directly determine the process’s stability; eigenvalues with negative real parts indicate stability, while those with positive real parts signify instability. Furthermore, the state-space formulation allows for the explicit calculation of the process’s long-term behavior, including its mean and autocovariance functions, by analyzing the system’s dynamics as time approaches infinity. This is achieved through the matrix exponential of the system matrix, e^{At}, which governs the evolution of the state vector over time and, consequently, the process’s long-term characteristics.

The correlation function of the SupCARMA process is fundamentally connected to its state-space representation through the observation equation. Specifically, the autocovariance at any given lag can be computed directly from the state-space parameters – the state transition matrix and the observation matrix. This linkage allows for the explicit calculation of the process’s dependence structure; given the state-space model, the correlation between any two points in the time series can be determined. Furthermore, analyzing the correlation function derived from the state-space representation reveals information about the process’s memory and the range of its dependence, effectively characterizing how past values influence future values within the time series.

Dissecting the SupCAR(2)(2): Unveiling Distinct Cases

The SupCAR(2)(2) process is a specialized case within the broader SupCARMA framework, defined by a specific autoregressive structure denoted as CAR(2)(2). This structure indicates that the process utilizes two lagged values of the dependent variable and two lagged values of the driving process in its current value calculation. Formally, a CAR(2)(2) process can be expressed as y_t = \phi_1 y_{t-1} + \phi_2 y_{t-2} + \psi_1 w_{t-1} + \psi_2 w_{t-2} + w_t , where y_t is the time series variable, w_t is a white noise error term, and φ and ψ represent the autoregressive and moving average coefficients, respectively. The SupCAR(2)(2) process applies this CAR(2)(2) structure within the SupCARMA methodology to model and analyze time series data with specific characteristics.

The SupCAR(2)(2) process, a specific implementation of the SupCARMA methodology, exhibits differing behaviors determined by the eigenvalues of the matrix A. These variations result in three distinct process forms: SupCAR(2)(2)-I, SupCAR(2)(2)-II, and SupCAR(2)(2)-III. The classification is directly linked to the characteristics of these eigenvalues; specifically, SupCAR(2)(2)-I is observed when A possesses real and equal eigenvalues. Conversely, SupCAR(2)(2)-II arises with distinct eigenvalues, while SupCAR(2)(2)-III is characterized by complex conjugate eigenvalues. This eigenvalue-based categorization is crucial for understanding and modeling the dynamic properties of the SupCAR(2)(2) process.

The SupCAR(2)(2)-I case is characterized by the matrix A possessing real and equal eigenvalues. This eigenvalue structure results in a predictable, non-oscillatory behavior of the process. Specifically, the state vector converges to a stable equilibrium point without exhibiting the damped oscillations seen in cases involving distinct or complex conjugate eigenvalues. The convergence rate is determined by the specific value of the repeated eigenvalue and the elements of the system matrix; a larger eigenvalue magnitude generally implies faster convergence, while the system matrix details influence the direction of convergence and the stability of the equilibrium.

The SupCAR(2)(2)-II and SupCAR(2)(2)-III processes diverge in behavior based on the eigenvalues of the associated matrix A. SupCAR(2)(2)-II arises when matrix A possesses distinct real eigenvalues, leading to differing rates of convergence or oscillation in the process. Conversely, SupCAR(2)(2)-III emerges when matrix A’s eigenvalues are complex conjugates. This results in oscillatory behavior with an exponentially decaying or growing amplitude, characterized by a frequency determined by the imaginary component of the eigenvalues and a damping rate determined by the real component. The specific characteristics of each case – convergence rate, oscillation frequency, and amplitude – are directly governed by the values of these eigenvalues within the 2 \times 2 matrix A.

Anchoring Stability: The Significance of Stationarity

The SupCARMA process, a valuable tool in modeling complex phenomena, demands adherence to specific stationarity conditions to guarantee its reliability and predictive power. These conditions aren’t merely technicalities; they fundamentally dictate whether the process will yield meaningful, stable results or devolve into unpredictable behavior. Specifically, these conditions center on the Lévy measure, a component that defines the probabilistic characteristics of the process’s jumps. A crucial requirement for existence and stationarity is that the integral of log|x| multiplied by the Lévy measure μ(dx), over all values of x greater than one, remains finite. Failing to meet this criterion introduces instability, rendering the SupCARMA process untrustworthy for accurate forecasting or robust analysis, and highlighting the importance of careful parameter selection and validation.

The very foundation of a stable SupCARMA process rests upon specific properties of its driving Lévy measure, a mathematical function describing the probabilistic characteristics of jumps in the process. Specifically, for the process to exist and maintain stationarity-meaning its statistical properties don’t change over time-the integral of |x|\log|x| multiplied by the Lévy measure μ(dx), evaluated over all values of x greater than one, must be finite. This condition essentially demands that extreme jumps, those with very large magnitudes, occur infrequently enough to prevent the process from becoming unbounded or exhibiting erratic behavior. Without this constraint, the statistical properties of the SupCARMA process become unpredictable, undermining its usefulness in modeling real-world phenomena where stable and consistent behavior is essential.

The variance of the SupCAR(2)(2) process, categorized as Type III, is not a simple, directly observable quantity but rather a calculated value deeply intertwined with the underlying stochastic components. Specifically, Var(X(0)) = (1/4)Var(L(1)) ∫_{0}^{∞}∫_{π/2}^{π} 1/(r³|cosψ|) π(dr,dψ), where Var(L(1)) represents the variance of the Lévy process driving the system, and the integral reflects the complex interplay between radial distance r, angle ψ, and the Lévy measure π(dr,dψ). This formulation indicates that the process’s instantaneous variance is proportionally related to the Lévy process’s variance, scaled by a factor determined by the integration over a specific quadrant in polar coordinates, highlighting the process’s inherent anisotropy and dependence on the characteristics of the random fluctuations driving it.

The statistical relationship between points in time within a SupCAR(2)(2) process, known as its correlation function, is defined by a complex integral expression: r(τ) = Var(L(1))/(2Var(X(0))) ∫(0,∞)×(π/2,π) e^(rτcosψ)/(r³sin²ψ) sin(rτsinψ − ψ) π(dr,dψ). This formula reveals that the degree of dependency between observations at different time lags τ is not simply determined by a few core parameters. Instead, it hinges on the interplay of the variance of the driving Lévy process Var(L(1)), the process’s overall variance Var(X(0)), and a double integral evaluated across a range of radial distances r and angles ψ. The exponential and trigonometric components within the integral demonstrate that even subtle changes in the process’s parameters can lead to significant alterations in the correlation structure, highlighting the nuanced and potentially non-intuitive behavior of this stochastic model.

The exploration of supCARMA processes, as detailed in this study, hinges on a rigorous understanding of how superimposed Lévy-driven components interact to define the process’s overall behavior. This careful construction mirrors a fundamental principle of pattern recognition: that complex systems are built from simpler, underlying structures. As Erwin Schrödinger noted, “One can never obtain more than one’s share of the truth.” The supCARMA model exemplifies this by demonstrating how a process’s stationarity and long-range dependence-key characteristics analyzed within-are intrinsically linked to the properties of its constituent Lévy basis and the carefully chosen CAR(2)(2) matrix. A thorough examination of these boundaries is crucial to avoid misinterpreting spurious correlations, highlighting the need for meticulous data analysis.

Where Do We Go From Here?

The introduction of supCARMA processes, while offering a structured approach to modeling long-range dependence via superposition, predictably opens more questions than it closes. The existence conditions, tied as they are to the spectral properties of the underlying CAR(2)(2) matrix, suggest a natural extension: exploring the richness of process behavior when these matrices flirt with instability. A rigorous classification of supCARMA processes based on these spectral characteristics-identifying those that exhibit genuinely novel correlation structures-remains a compelling, if technically challenging, endeavor.

Currently, the analysis heavily relies on the stationarity assumption. The behavior of supCARMA processes when subjected to non-stationary Lévy bases, or when the underlying CAR(2)(2) parameters themselves evolve, presents a significant, and perhaps more realistic, avenue for investigation. The interplay between the Lévy process driving the superposition and the inherent dynamics of the CARMA components deserves particular attention; does one dominate, or do they engage in a complex, mutually influencing dance?

Ultimately, the true test will lie in application. While the theoretical framework provides a powerful tool for generating and analyzing complex time series, the value of supCARMA processes hinges on their ability to meaningfully represent real-world phenomena. Finding empirical data where this superposition structure offers a demonstrable advantage over existing models-that is the puzzle that will truly define the legacy of this work.

Original article: https://arxiv.org/pdf/2601.15796.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-01-26 05:30