Abstract
The mean absolute deviation about the mean is an alternative to the standard deviation for measuring dispersion in a sample or in a population. For stationary, ergodic time series with a finite first moment, an asymptotic expansion for the sample mean absolute deviation is proposed. The expansion yields the asymptotic distribution of the sample mean absolute deviation under a wide range of settings, allowing for serial dependence or an infinite second moment.
Introduction
In the study of statistical dispersion, the standard deviation has long been the default metric. However, its reliance on squared distances makes it exceptionally sensitive to outliers, which can distort the perceived volatility or risk in a dataset. The Mean Absolute Deviation (MAD) about the mean—defined as the average absolute distance of observations from the sample mean—offers a more robust alternative.
Despite its intuitive appeal and practical robustness, the mathematical properties of MAD are surprisingly complex. The "absolute value" function introduces a non-differentiable "kink" at the origin, which prevents the straightforward application of standard calculus-based tools like the Delta Method. Historically, researchers relied on sophisticated "empirical process theory" to derive the asymptotic behavior of MAD, often under restrictive assumptions such as independent and identically distributed (IID) data with finite variance. This paper by Johan Segers provides a new, "first principles" asymptotic expansion that extends the utility of MAD to more complex data environments, including dependent time series and heavy-tailed distributions.
Methodology: A First Principles Approach
The central contribution of this work is the derivation of a linear asymptotic expansion for the sample MAD. Rather than employing the "sledgehammer" of empirical process theory, the author develops Lemma 2.1, which provides a linear approximation of the difference between the sample MAD ($D_n$) and the population MAD ($d$).
The expansion is built upon several key technical pillars:
- Birkhoff Ergodic Theorem: This allows the framework to accommodate stationary, ergodic time series where observations are not independent but follow a consistent probabilistic structure over time.
- Case-by-Case Analysis: The author meticulously decomposes the absolute value function to account for the fluctuations of the sample mean ($\bar{X}_n$) around the true population mean ($\mu$).
- Handling Distributional "Atoms": A significant advancement in this paper is the explicit treatment of distributions that possess a point mass (an "atom") exactly at the mean ($P[X = \mu] > 0$). While previous literature often assumed continuity at the mean, Segers demonstrates that the presence of an atom fundamentally alters the asymptotic limit.
Key Insights: The "Atom" Effect and Unified Framework
One of the paper's most striking revelations concerns the "Atom" effect. If a distribution has a positive probability of an observation equaling the mean, the MAD does not necessarily converge to a standard Gaussian (normal) distribution in large samples. Instead, the limit becomes a non-linear function of a bivariate normal distribution. This finding is critical for practitioners working with discrete or mixed distributions where such "spikes" are common.
Furthermore, the paper proposes a unified "recipe" for determining the limit distribution across three distinct scenarios:
- Independent Data (IID): Providing a simpler path to classical results.
- Dependent Time Series: Specifically "strongly mixing" processes where the influence of the past on the future diminishes over time.
- Heavy-Tailed Data: Scenarios where the variance is infinite (e.g., Pareto distributions), provided the first moment (the mean) remains finite.
Results and Applications
The theoretical framework yields several high-impact propositions:
Proposition 3.1 (Time Series): For strongly mixing time series, the MAD converges to a Gaussian limit. However, the variance of this limit is not the standard IID variance; it is adjusted to incorporate the serial correlation (autocovariance) of the data. This allows for more accurate confidence intervals in econometric and financial modeling.
Proposition 3.3 (Infinite Variance): In the case of independent data with infinite variance—specifically those within the domain of attraction of a stable law with index $\alpha \in (1, 2)$—the MAD converges to a stable distribution rather than a normal one. The author provides the precise scale parameter for this limit, reinforcing MAD's reliability in high-risk, heavy-tailed environments where standard deviation fails to converge.
Limitations and Future Directions
While the paper significantly advances the theoretical understanding of MAD, it notes certain boundaries. In the heavy-tailed case (Proposition 3.3), the author assumes the distribution is continuous at the mean. If an atom exists at the mean in an infinite-variance scenario, the resulting limit law is even more complex and ceases to be a stable distribution.
Additionally, the focus remains strictly on the mean absolute deviation about the mean. While the techniques might be adaptable to the median absolute deviation about the median (another common "MAD"), that remains outside the scope of this specific work. Finally, the paper provides the theoretical limit laws but leaves the practical estimation of parameters (like mixing coefficients or stable indices) for finite samples as a challenge for applied researchers.
Conclusion
By returning to first principles, this research clarifies the asymptotic behavior of one of statistics' most robust measures of dispersion. The provided expansion offers a versatile toolkit for researchers dealing with the complexities of real-world data, from the serial dependence of financial markets to the heavy tails of insurance risks.
Summary updated