📊 statistics

On Functional Deviations About Functionals

This paper develops a general framework for measuring dispersion by applying statistical functionals to absolute deviations from a reference value, establishes key inequalities comparing these measures, and illustrates their behavior across various distributions through simulation.

Original authors: STHITADHI DAS

Published 2026-10-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: STHITADHI DAS

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of data, knowing the average value of a group is often not enough. If a statistician tells you that the average income in a town is fifty thousand dollars, you still do not know if everyone earns close to that amount, or if a few billionaires are pulling the number up while most people struggle to make ends meet. To understand the true nature of a group, you must also measure how spread out the values are. This spread is called dispersion. For over a century, scientists have used a few standard tools to measure this spread, such as the average distance of all numbers from the center, or the square-root of the average of the squared distances. These tools work well for simple, balanced groups of numbers, but they can become misleading when the data is lopsided, has extreme outliers, or comes from a complex mixture of different sources. The question of how to best describe this spread, and which mathematical tools to use when the data is messy, remains a vital part of statistical science.

A researcher at Brainware University in India has recently proposed a new way to think about these measurements. Instead of treating the calculation of spread as a single, fixed recipe, the researcher suggests viewing it as a two-step process where the choices made at each step matter independently. The first step is choosing a reference point, or a "center," from which to measure the distance of every number in the group. This center could be the average, the middle value, or even the most frequent value. The second step is deciding how to summarize those distances. You could take the average of all the distances, find the middle distance, or look at the distance that covers the top ninety percent of the group. By separating these two choices, the researcher created a flexible framework that can generate many different ways to measure spread, including familiar ones and some that have not been formally compared before.

The study began by establishing a set of mathematical rules that govern how these different measurements relate to one another. The researcher proved that if you want to minimize the average distance of all numbers from a center, the middle value of the group is always a better choice than the average. In other words, the average distance of numbers from the middle value will never be larger than the average distance of numbers from the average itself. This holds true for any group of numbers, provided the numbers are not infinite. The researcher also showed that this average distance is always smaller than the standard measure of spread used in most textbooks, which relies on squaring the distances. These rules provide a solid foundation, showing that for certain types of data, one method is mathematically guaranteed to be more efficient than another.

To see how these ideas play out in the real world, the researcher ran a massive computer simulation. They generated three hundred thousand random numbers for six different types of distributions, representing everything from perfectly balanced groups to highly skewed ones with long tails of extreme values. They then calculated the various measures of spread for each group. The results confirmed the mathematical rules: for the balanced groups, like the normal bell curve or a uniform spread, the average distance from the middle value and the average distance from the average were almost identical. However, for the skewed groups, such as the exponential distribution or the log-normal distribution, the difference became significant. In these cases, the average distance calculated from the middle value was noticeably smaller than the one calculated from the average, sometimes by more than ten percent. This confirmed that when data is lopsided, the choice of center changes the story you tell about the spread.

The simulation also explored what happens when you look at the extremes of the spread rather than just the average. By calculating specific points in the distribution of distances, such as the point below which ninety percent of the distances fall, the researcher found that these "quantile" measures revealed details that the average missed. For groups with heavy tails or extreme outliers, the upper end of the spread was much larger than the average suggested. This means that relying on a single number to describe the spread can hide the fact that a small portion of the data is behaving very differently from the rest. The study showed that using a combination of these different measures gives a much richer and more accurate picture of the data than any single tool could provide.

One of the more complex parts of the study involved the "mode," which is the most frequent value in a group. Unlike the average or the middle value, the mode is not defined by a simple calculation but by finding the peak of a curve that represents the data. The researcher found that this method is much more sensitive to how the data is smoothed out during the calculation. Small changes in the method used to estimate the peak could lead to different results, making the mode-based measure of spread less stable than the others. This suggests that while the mode can be useful for identifying the most common value, using it as a center for measuring spread requires careful handling and cannot be treated with the same mathematical certainty as the average or the middle value.

The work concludes that there is no single "best" way to measure how spread out data is. The best choice depends entirely on the shape of the data and what question you are trying to answer. If the data is balanced, the traditional methods work well. If the data is skewed or has extreme values, using the middle value as a center and looking at different points in the distribution of distances provides a clearer view. The researcher's framework does not replace the old tools but organizes them into a coherent system, showing exactly when and why one tool should be preferred over another. By separating the choice of the center from the method of summarizing the distances, the study offers a more precise language for describing the variability that exists in everything from financial markets to biological measurements.

Technical Summary: On Functional Deviations About Functionals

Problem Statement
The paper addresses the need for a unified framework to describe statistical dispersion. Classical measures of dispersion (e.g., variance, standard deviation, mean absolute deviation, median absolute deviation) are often treated as distinct entities. However, they share a common construction: applying a statistical functional to the absolute deviations of a random variable from a reference location. The author argues that the choice of the reference location (e.g., mean, median, quantile) and the choice of the functional used to summarize the resulting deviations (e.g., expectation, median, quantile) are often conflated. This conflation obscures the distinct roles these choices play, particularly in non-symmetric, heavy-tailed, or multimodal distributions where different location functionals yield different reference points. Furthermore, while inequalities exist between specific measures (e.g., mean absolute deviation about the median vs. the mean), a general theoretical structure connecting these measures across different functionals is lacking.

Methodology
The paper develops a general mathematical framework and validates it through theoretical analysis and Monte Carlo simulation.

Theoretical Framework:
- Definitions: The author defines a "functional deviation" $D_{S|T}(X) = S(|X - T(X)|)$ , where $T$ is a reference functional determining the location and $S$ is a summary functional applied to the distribution of absolute deviations. A "powered" version $D^{(r)}_{S|T}$ is also defined.
- Properties: The paper establishes fundamental properties of these deviations, including non-negativity, translation invariance, and positive scale equivariance, provided the underlying functionals satisfy specific conditions.
- Inequalities: Rigorous inequalities are derived under explicit moment and integrability assumptions. Key results include:
  - The expected absolute deviation about a median is less than or equal to the expected absolute deviation about the mean ( $E|X-m| \le E|X-\mu|$ ).
  - The mean absolute deviation about the mean is bounded by the standard deviation ( $E|X-\mu| \le \sigma$ ).
  - Bounds relating the median and quantiles of absolute deviations to their expectations (e.g., $Q_Y(p) \le E[Y]/(1-p)$ for non-negative $Y$ ).
- Mode Limitations: The paper explicitly notes that universal inequalities analogous to those for the mean and median cannot be asserted for the mode due to its lack of a minimization property for absolute loss and potential non-uniqueness.
Simulation Study:
- Design: A Monte Carlo simulation was conducted using $N=300,000$ observations for six distinct distributions: Standard Normal, Uniform, Exponential, Student's $t$ (3 degrees of freedom), Lognormal, and a two-component Gaussian Mixture.
- Objectives: The study aimed to empirically verify the theoretical inequalities and examine how distributional characteristics (symmetry, skewness, tail weight, multimodality) influence the numerical relationships between different functional deviations.
- Extensions: A repeated-sampling experiment ( $B=2000$ replications) was performed across varying sample sizes ( $n \in \{50, \dots, 1000\}$ ) to assess finite-sample stability, bias, and variability. A separate sensitivity analysis was conducted for mode-based deviations using kernel density estimation with varying bandwidths.

Key Contributions

Unified Formulation: The paper provides a common mathematical representation for a wide class of dispersion measures by separating the reference location functional ( $T$ ) from the deviation-summary functional ( $S$ ). This allows for the systematic comparison of mixed-functional deviations (e.g., mean absolute deviation about the median).
Rigorous Inequalities: The author establishes distribution-free inequalities for selected members of the functional deviation family, clarifying the conditions (e.g., existence of moments) under which they hold.
Empirical Characterization: The study provides detailed numerical evidence of how the choice of reference and summary functionals affects dispersion measures across diverse distributional shapes. It highlights that while symmetric distributions often yield coincident measures, skewed and heavy-tailed distributions exhibit significant divergence.
Mode Sensitivity Analysis: The paper explicitly investigates the instability of mode-based deviations, demonstrating their sensitivity to bandwidth selection in density estimation, and argues for their treatment as distinct from mean/median-based measures.

Results

Inequality Verification: The simulation results confirmed the theoretical inequalities. Specifically, the mean absolute deviation about the median was consistently less than or equal to the mean absolute deviation about the mean across all distributions. The mean absolute deviation about the mean was consistently less than the standard deviation.
Distributional Impact:
- Symmetric Distributions: For Normal, Uniform, and Student's $t(3)$ , the mean and median coincided, resulting in nearly identical mean-centered and median-centered absolute deviations.
- Skewed Distributions: For Exponential and Lognormal distributions, the median-centered mean absolute deviation was significantly smaller than the mean-centered version (reductions of ~5.8% and ~10.8%, respectively), reflecting the minimization property of the median.
- Heavy Tails: The Student's $t(3)$ and Lognormal distributions showed that standard deviations were substantially larger than mean absolute deviations, illustrating the greater sensitivity of quadratic measures to extreme values.
- Quantiles: Quantile-based deviations revealed that upper-tail behavior varies significantly across distributions, providing information not captured by single-point summaries like the mean or median.
Finite-Sample Behavior: The repeated sampling experiment showed that while the ratio of median-centered to mean-centered deviations approaches 1 for symmetric distributions as sample size increases, the reduction persists for skewed distributions. Variability in estimators was found to be higher for heavy-tailed and skewed distributions.
Mode Instability: The estimated mode of absolute deviations was shown to be highly sensitive to the bandwidth used in kernel density estimation, with the estimated global mode and the number of local maxima varying significantly with smoothing parameters.

Significance and Claims
The paper claims to provide a structured basis for determining which comparisons between dispersion measures are mathematically valid and which depend on the underlying distribution. It emphasizes that the proposed framework does not impose a universal ranking on all functional deviations but rather clarifies the roles of reference and summary functionals.

The author modestly positions the work as a unification of existing concepts and a source of rigorous inequalities, rather than a discovery of new universal laws. The significance lies in:

Conceptual Clarity: Separating the "where" (reference) from the "how" (summary) of measuring dispersion.
Theoretical Rigor: Providing proofs for inequalities that are often assumed or observed only numerically.
Practical Guidance: Demonstrating through simulation that the choice of functional matters significantly in asymmetric or heavy-tailed settings, and cautioning against the uncritical use of mode-based measures without addressing estimation sensitivity.

The paper concludes that while theoretical inequalities hold under specific assumptions, numerical agreement in finite samples constitutes empirical evidence rather than proof, and that future work should focus on extending these frameworks to multivariate data, functional data, and robustness under contamination.

More like this