On the Properties of Moments of Matrix Exponential Distributions and Matrix Exponential Processes

Levente Bodrog, András Horváth, Miklós TelekDROPS (Schloss Dagstuhl – Leibniz Center for Informatics) 2008 · DOI

TL;DR

This paper introduces a novel method to compute moments of matrix exponential (ME) distributions and processes without needing their underlying representations. This is achieved by leveraging relationships between moments, enabling the generation of higher-order moments from lower-order ones, which is crucial for analyzing complex stochastic systems.

What Problem Does This Paper Address?

Many real-world systems, particularly in areas like telecommunications and queueing theory, can be modeled using stochastic processes. Matrix exponential (ME) distributions and processes offer a powerful, generalized framework for these models, extending beyond simpler phase-type (PH) distributions and Markovian arrival processes (MAPs). However, working with ME models is challenging because their underlying representations (matrices and initial vectors) are often complex and difficult to determine directly. This paper addresses the problem of characterizing these ME distributions and processes by focusing on their moments. Specifically, it aims to develop methods to compute any finite-dimensional moments, including joint moments of processes, using only a set of known lower-order moments, thereby bypassing the need to explicitly compute the ME representation. This is vital for tasks like moment matching and fitting ME models to empirical data.

Methods

The core of the paper's methodology lies in exploiting the relationships between different moments of ME distributions and processes.

  • Hankel Matrix Properties: The authors introduce and utilize the properties of doubly infinite Hankel matrices formed by the moments. They demonstrate that the rank of these matrices is directly related to the order of the ME distribution or process. This property allows them to derive determinantal equations that govern the relationships between moments.
  • Moment Generation Algorithm: Based on these determinantal equations, an algorithm is presented to compute any finite-dimensional moments. This algorithm works by iteratively using lower-order moments to derive higher-order ones, without requiring the explicit (a, A) representation for distributions or (A0, A1) for processes.
  • Reduced Moments: The paper defines and uses "reduced moments" (e.g., $r_i$ for distributions and $E(X_0^{k_1} \dots X_m^{k_m})$ for processes) as the fundamental building blocks.
  • MEP Joint Moment Relations: For matrix exponential processes (MEPs), the paper extends these concepts to joint moments of inter-arrival times. It establishes relationships between these joint moments using matrix structures ($M_1$ and $M_2$) and their determinants, similar to the Hankel matrix approach for distributions.
  • Basic Moment Set: A "basic moment set" is defined, which consists of a minimal set of moments (reduced moments and certain joint moments) that uniquely determine all other moments of a given MEP.

Datasets

Not discussed in the paper. The paper focuses on theoretical derivations and algorithmic development, illustrated with examples rather than empirical datasets.

Key Insights and Ideas

  • Moment-Centric Approach: The most significant insight is that the behavior of ME distributions and processes can be fully characterized by their moments, and these moments are intricately related. This allows for a powerful alternative to working directly with the underlying matrix representations.
  • Determinantal Equations as Constraints: The determinantal equations derived from Hankel matrices (for distributions) and related matrix structures (for processes) are not just theoretical curiosities; they are the operational tools that enable the computation of unknown moments from known ones.
  • Uniqueness and Redundancy: The paper highlights that ME representations are often non-unique. However, the moment relationships are fundamental and can uniquely define the distribution or process, provided the representation is non-redundant.
  • MEP Complexity: The paper demonstrates that for MEPs, a specific set of moments (the "basic moment set") is sufficient to determine all other moments, offering a structured way to understand their complexity.
  • Identical Moments, Different Processes: A surprising finding is that different MEPs can share the same marginal distribution and lag-correlations, yet have distinct joint densities, as illustrated in Section 3.5. This underscores the importance of considering higher-order joint moments for a complete characterization.

Key Results

  • Algorithm for Moment Computation: The paper presents an explicit algorithm to compute any finite-dimensional moments of ME distributions and joint moments of MEPs based on a set of low-order moments.
  • Rank Properties: The rank of specific Hankel-like matrices is shown to be equal to the order of the ME distribution or process, providing a fundamental structural property.
  • Feasibility Check: The derived moment relationships can be used to check if a given set of moments is feasible for an ME distribution or MEP.
  • Examples: The paper provides concrete examples of generating moments for ME distributions and checking the feasibility of MEPs, demonstrating the practical application of the developed theory.

Limitations and Future Work

  • Computational Complexity: While the algorithms bypass the need for explicit representations, computing high-order moments can still become computationally intensive.
  • Numerical Stability: The paper implicitly relies on determinant calculations, which can be sensitive to numerical precision for very high-order moments or ill-conditioned matrices.
  • Future Work: The authors explicitly state their intention to apply this compact moment representation of MEPs to "moment matching and fitting." This suggests a future direction focused on using these theoretical results for practical model construction and parameter estimation from data. They also mention that the presented results are directly applicable to PH distributions and MAPs, implying potential for further exploration in these specific, widely used sub-classes.

Summary updated

Abstract

In this paper we provide properties of moments of matrix exponential distributions and joint moments of matrix exponential processes. Based on the provided properties, an algorithm is presented to compute any finite dimensional moments of these processes based on a set of required (low order) moments. This algorithm does not require the computation of any representation of the given process. We present some related examples to demonstrate the potential use of the properties of moments.
BibTeX
@inproceedings{bodrog2008properties,
  author = {Levente Bodrog and András Horváth and Miklós Telek},
  title = {{On the Properties of Moments of Matrix Exponential Distributions and Matrix Exponential Processes}},
  booktitle = {DROPS (Schloss Dagstuhl – Leibniz Center for Informatics)},
  year = {2008},
  doi = {10.4230/dagsemproc.07461.12},
}
Figure