A functional limit theorem for partial sums of dependent random variables with infinite variance

Bojan Basrak, Danijel Krizmanić, Johan SegersarXiv 2010 · arXiv

TL;DR

This paper establishes a functional limit theorem for the partial sums of stationary, heavy-tailed (infinite variance) random variables that exhibit dependence. The authors prove that under specific "mixing" and "regular variation" conditions, these sums converge to a stable Lévy process, provided we use Skorohod’s $M_1$ topology to account for the "clustering" of extreme values.

What Problem Does This Paper Address?

In classical probability, the Central Limit Theorem tells us that sums of independent variables with finite variance converge to a Gaussian distribution. When variance is infinite (heavy tails), sums instead converge to non-Gaussian "stable" distributions.

The "functional" version of this—looking at the entire path of the sum over time—usually results in a stable Lévy process. However, most existing theory assumes the variables are independent and identically distributed (i.i.d.). In the real world (especially in finance or teletraffic), data is often dependent.

Dependence creates a major mathematical headache: clustering. Large values tend to happen in groups. In the limit, a cluster of several large jumps might "collapse" into a single jump. Standard mathematical frameworks (like Skorohod’s $J_1$ topology) are too rigid to handle this because they require jumps to stay separated. This paper seeks to bridge that gap by providing a rigorous proof for dependent sequences using a more flexible framework.

Methods

The authors employ several sophisticated tools from modern probability:

  • Regular Variation: They assume the sequence is "jointly regularly varying," a mathematical way of saying the "tails" of the distribution follow a power law, which is characteristic of infinite variance.
  • Tail Processes: They use the concept of a "tail process" to describe the behavior of the sequence specifically when it takes on extreme values. This allows them to characterize how clusters form.
  • Point Process Convergence: The core of the proof involves showing that a "time-space point process" (tracking when and how large the values are) converges to a Poisson cluster process.
  • Skorohod’s $M_1$ Topology: This is the "secret sauce." Unlike the standard $J_1$ topology, the $M_1$ topology considers the "completed graph" of a function (including vertical segments at jumps). This allows multiple jumps occurring very close together in a cluster to merge into a single jump in the limit without breaking the convergence.

Datasets

Not discussed in the paper. As a theoretical probability paper, it focuses on mathematical proofs and provides illustrative examples using well-known stochastic models rather than empirical datasets.

Key Insights and Ideas

  • The $M_1$ Necessity: The most significant insight is that the $J_1$ topology is fundamentally inappropriate for dependent heavy-tailed data. Because dependent extremes cluster, the partial sum process exhibits "rapid successions of jumps" that the $J_1$ metric cannot reconcile with a single-jump limit.
  • Lévy Triple Modification: The authors show that clustering doesn't just change the speed of convergence; it changes the "Lévy triple" (the parameters defining the limit process). The limit for a dependent sequence can look very different from the i.i.d. case with the same marginal distribution.
  • Asymptotic Independence: The theorem relies on the idea that while values are dependent locally (forming clusters), the clusters themselves must be "far enough apart" to be treated as asymptotically independent.

Key Results

  • Theorem 3.4 (The Main Result): The authors prove that for a stationary sequence satisfying their conditions, the normalized partial sum process $V_n(t)$ converges weakly to an $\alpha$-stable Lévy process in the space of càdlàg functions endowed with the $M_1$ topology.
  • Explicit Parameters: They provide formulas to calculate the parameters of the limit process based on the "tail process" of the original sequence.
  • Model Validation: The paper demonstrates that these conditions are satisfied by several popular time-series models:
  • Moving Average (MA) processes with heavy tails.
  • Squared GARCH(1,1) processes, which are widely used to model volatility in financial markets.
  • Stochastic Volatility models.

Limitations and Future Work

  • The Sign Constraint: A key assumption in the main theorem is that all extreme values within a single cluster must have the same sign (all positive or all negative). If a cluster contains both large positive and large negative values that cancel each other out, the $M_1$ topology may still fail.
  • Topology Alternatives: The authors briefly mention that if the sign constraint is violated, one might need to move to an even weaker topology, such as the $L_1$ topology.
  • Checking Conditions: While the paper provides sufficient conditions (like $\rho$-mixing), these can be notoriously difficult to verify for complex, real-world processes. Future work could involve simplifying these conditions for broader classes of models.

Summary updated

Abstract

Under an appropriate regular variation condition, the affinely normalized partial sums of a sequence of independent and identically distributed random variables converges weakly to a non-Gaussian stable random variable. A functional version of this is known to be true as well, the limit process being a stable Levy process. The main result in the paper is that for a stationary, regularly varying sequence for which clusters of high-threshold excesses can be broken down into asymptotically independent blocks, the properly centered partial sum process still converges to a stable Levy process. Due to clustering, the Levy triple of the limit process can be different from the one in the independent case. The convergence takes place in the space of cadlag functions endowed with Skorohod's $M_1$ topology, the more usual $J_1$ topology being inappropriate as the partial sum processes may exhibit rapid successions of jumps within temporal clusters of large values, collapsing in the limit to a single jump. The result rests on a new limit theorem for point processes which is of independent interest. The theory is applied to moving average processes, squared GARCH(1,1) processes and stochastic volatility models.
BibTeX
@misc{basrak2010functional,
  author = {Bojan Basrak and Danijel Krizmanić and Johan Segers},
  title = {{A functional limit theorem for partial sums of dependent random variables with infinite variance}},
  year = {2010},
  eprint = {1001.1345},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1001.1345},
}

Loading related content...

Figure