On matrix exponential approximations of the infimum of a spectrally negative Levy process

Florin Avram, András Horváth, Martijn PistoriusarXiv 2012 · DOI · arXiv

TL;DR

This paper tackles the challenge of approximating the infimum of spectrally negative Levy processes, crucial for problems like ruin probabilities in insurance. The authors introduce a new approximation for a specific model and generalize existing methods to fit an arbitrary number of moments, offering a more flexible approach to complex financial and queueing problems.

What Problem Does This Paper Address?

In fields like risk theory, queueing theory, and financial mathematics, understanding the behavior of stochastic processes is paramount. A key problem is approximating the "infimum" of a spectrally negative Levy process. This is directly related to calculating ruin probabilities (the chance an insurance company goes bankrupt), waiting times in queues, and pricing financial derivatives. Many existing approximation methods rely on fitting a small number of initial moments (like the mean and variance) of the process's distribution. However, for more accurate modeling, especially in "intermediate regimes" where simple asymptotic behaviors don't fully apply, fitting a higher number of moments is desirable. The challenge lies in developing approximations that are not only accurate but also "admissible" – meaning they correspond to valid probability distributions (e.g., non-negative densities and non-increasing survival functions). This paper addresses the scarcity of such high-order, admissible approximations and explores the difficulties in their construction.

Methods

The paper primarily focuses on Padé approximations and related rational function approximations of Laplace transforms. These methods are powerful because they can approximate complex functions using ratios of polynomials, and their coefficients can be directly related to the moments of the underlying distribution.

Key methodological contributions include:

  • New Approximation for Perturbed Cramér-Lundberg Model: The authors propose a novel approximation for a specific risk model that includes a Brownian perturbation.
  • Generalization of Johnson-Taaffe Approximations: They highlight and extend the work of Johnson and Taaffe, presenting a family of approximations that can fit an arbitrarily high number of moments. This is a significant generalization of earlier methods that were limited to fitting only a few moments.
  • Exploration of Admissibility: A significant portion of the paper is dedicated to discussing the crucial concept of "admissibility" for these approximations. This involves ensuring that the resulting approximations correspond to valid probability distributions, which becomes increasingly difficult with higher-order approximations. They explore conditions for admissibility, particularly in the context of phase-type representations and positive realization problems.
  • Two-Point Padé Approximations: The paper also delves into two-point Padé approximations, which incorporate information about the function's behavior at both the origin (related to moments) and at infinity (related to asymptotic behavior). These are noted as being particularly useful for perturbed models.

Datasets

Not discussed in the paper. The paper focuses on theoretical approximations and their properties, using examples with specific distributions (e.g., mixed exponential, Gamma) for numerical illustration rather than large datasets.

Key Insights and Ideas

  • High-Order Approximations are Challenging but Necessary: While fitting more moments generally leads to better approximations, ensuring the "admissibility" (i.e., validity as a probability distribution) of these high-order approximations is a major hurdle.
  • Johnson-Taaffe Approximations Offer Great Flexibility: The Johnson-Taaffe method provides a powerful framework for constructing approximations that can match an arbitrary number of moments, significantly extending the capabilities of earlier techniques.
  • Admissibility is Key: The paper repeatedly emphasizes that simply fitting moments isn't enough; the resulting approximations must represent valid probability distributions. This is a non-trivial constraint, especially for higher-order methods.
  • Two-Point Padé Approximations are Valuable for Perturbed Models: Incorporating behavior at infinity, in addition to moments, is crucial for accurately modeling processes with perturbations (like Brownian motion).
  • Phase-Type Representations Offer Automatic Admissibility: When an approximation can be represented as a phase-type distribution, admissibility is guaranteed, but finding such representations can be difficult.

Key Results

  • The paper introduces a new approximation for the perturbed Cramér-Lundberg model.
  • It highlights a family of Johnson-Taaffe approximations that can fit an arbitrarily high number of moments, greatly generalizing existing methods.
  • The paper provides a detailed discussion and examples of admissibility criteria for these approximations, particularly in the context of phase-type distributions.
  • Numerical examples (Figures 2-9, Tables 1-4) demonstrate the performance of various approximations (Renyi, De Vylder, Theorem 3, Theorem 5, etc.) for different claim distributions (mixed exponential, Gamma) and perturbation levels. These results show that newer approximations (like those from Theorem 5) often offer better accuracy, especially for certain ranges of parameters. For instance, Figure 3 shows that Theorem 5 generally has a lower relative error than De Vylder or Renyi.

Limitations and Future Work

  • Admissibility for High-Order Approximations Remains an Open Problem: While the paper discusses admissibility, ensuring it for arbitrary high-order approximations, especially those involving more than four terms or moments, is still an active research area.
  • Computational Complexity: Finding optimal parameters for high-order approximations can be computationally intensive.
  • Domain Identification for Method Preference: The paper poses the question of identifying specific domains where different approximation methods (e.g., Ramsay vs. Beekman-Bowers) are preferable, suggesting this as a direction for future work.
  • Fitting Laplace Transform at Infinity: The authors note that obtaining approximations that fit the Laplace transform at infinity, in addition to moments, would be highly beneficial.
  • Algorithmic Approach for Admissibility Testing: The paper points out the lack of a general algorithmic approach for testing admissibility of combinations of more than four negative exponentials.

Summary updated

Abstract

We recall four open problems concerning constructing high-order matrix-exponential approximations for the infimum of a spectrally negative Levy process (with applications to first-passage/ruin probabilities, the waiting time distribution in the M/G/1 queue, pricing of barrier options, etc). On the way, we provide a new approximation, for the perturbed Cramer-Lundberg model, and recall a remarkable family of (not minimal order) approximations of Johnson and Taaffe, which fit an arbitrarily high number of moments, greatly generalizing the currently used approximations of Renyi, De Vylder and Whitt-Ramsay. Obtaining such approximations which fit the Laplace transform at infinity as well would be quite useful.
BibTeX
@misc{avram2012matrix,
  author = {Florin Avram and András Horváth and Martijn Pistorius},
  title = {{On matrix exponential approximations of the infimum of a spectrally negative Levy process}},
  year = {2012},
  eprint = {1210.2611},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1210.2611},
}
Figure