Dispersion Function¶

The convex location-indexed functional \(D_X(u)=\mathbb{E}|X-u|\), whose slopes recover an integrable real distribution.

Version: v1 · 2026-08-30 · History
Domain-specific #: 1692
Origin domain: probability theory
Subdomain: stochastic orders

Core Idea¶

For an integrable real random variable \(X\), the dispersion function is

\[ D_X(u)=\mathbb{E}|X-u|,\qquad u\in\mathbb{R}. \]

It turns one distribution into an entire location-indexed curve: at every proposed reference point \(u\), the curve reports expected absolute distance from that point. Muñoz-Pérez and Sánchez-Gómez introduced this named functional as a characterization of a distribution and as a basis for comparing dispersion.^[1] The crucial identity is stronger than “another variability measure.” Convexity and the one-sided slopes preserve the whole cumulative distribution:

\[ D'_{X,+}(u)=2F_X(u)-1,\qquad D'_{X,-}(u)=2F_X(u-)-1. \]

Thus the right derivative recovers \(F_X(u)=(D'_{X,+}(u)+1)/2\), including atoms; ordinary differentiability should not be asserted at a jump. The recognition invariant is the complete absolute-loss profile over every location, not a single mean absolute deviation evaluated at the mean or median.

This functional has an autonomous role because it connects robust loss, distribution recovery, convex geometry, medians, moments, and stochastic comparison through one curve. It is domain-specific rather than prime: its name and validity conditions remain probability-theoretic, even though convex absolute loss appears in many applications.

Structural Signature¶

Recognition roles: an integrable real random variable — a candidate reference location \(u\) — absolute displacement \(|X-u|\) — probability-weighted expectation — a finite convex function on \(\mathbb R\) — one-sided slopes encoding the CDF — tail asymptotes encoding the mean.

Integrable law. \(\mathbb E|X|<\infty\) ensures \(D_X(u)\) is finite for every finite \(u\).
Moving reference point. The argument \(u\) ranges over the whole real line. Fixing one \(u\) produces only one absolute moment.
Absolute-loss expectation. The functional uses power one, preserving robustness and convexity.
Convex profile. Every function \(u\mapsto|x-u|\) is convex; expectation preserves convexity.
Distribution-bearing slopes. Right and left derivatives distinguish cumulative mass through and before \(u\), so their gap is \(2P(X=u)\).
Location summary. Minimizers are medians, not necessarily the mean.
Asymptotic calibration. For integrable \(X\), \(D_X(u)-u\to-\mathbb EX\) as \(u\to+\infty\), while \(D_X(u)+u\to\mathbb EX\) as \(u\to-\infty\).

Recognition test: verify the curve for all \(u\), compute one-sided slopes, and check that they yield a nondecreasing right-continuous CDF with limits zero and one. A scalar dispersion statistic, a mean–variance relation, or an unrelated function called “dispersion” does not qualify.

What It Is Not¶

It is not the variance function \(V(\mu)\), which maps a model's mean to conditional variance. Here \(u\) is a freely varying reference location and \(D_X\) encodes one complete law.
It is not variance. Variance is the scalar \(\mathbb E(X-\mathbb EX)^2\); the dispersion function uses absolute loss and retains a full curve.
It is not mean absolute deviation alone. Evaluating \(D_X\) at one center discards the slope information that recovers the CDF.
It is not prime:dispersion in this catalog. That prime is deterministic propagation-rate separation of a launched bundle, a different identity sharing a word.
It is not the plasma dispersion function, a special complex function in kinetic plasma theory.
It is not a generic loss function. Absolute loss is an ingredient, but the named object is its expectation as a function of every candidate location under a probability law.

Scope of Application¶

The primary scope is univariate probability theory and mathematical statistics for laws with finite first moment. It provides a representation of distributions, a graphical absolute-distance profile, a route to medians, and an object for studying spread comparisons. The original paper explicitly establishes distribution characterization and an induced dispersive ordering.^[1]

In stochastic-order theory, the curve belongs among integrated-distribution transforms used to compare laws beyond their means. Shaked and Shanthikumar document stochastic orders as tools for uncertainty comparison and for structural analysis of stochastic systems.^[2] The dispersion function's exact ordering convention must be stated rather than replaced by a vague rule that “larger curve means more variable,” because translations and location differences can also move curves.

In robust statistical reasoning, \(D_X(u)\) is population absolute loss for a constant predictor \(u\). Its minimizers are medians, and empirical analogues replace expectation with a sample average. That connection is literal, but the population functional should not be merged with a particular sample estimator or optimization algorithm.

Clarity¶

The name separates three information levels. A mean gives one location number. Mean absolute deviation gives one spread number relative to a selected center. The dispersion function records expected absolute deviation from every center, thereby retaining the entire CDF. This explains why two laws can share mean and mean absolute deviation yet have different dispersion functions.

The one-sided derivative convention adds critical clarity at atoms. For a continuous distribution, the ordinary derivative exists and \(F_X=(D'_X+1)/2\). At an atom, the convex curve has a kink; the derivative jump is evidence rather than a defect. Using only “the derivative” would erase precisely the discrete mass. A good analysis reports the relevant side.

Manages Complexity¶

A probability law may require a density, a mass function, or a mixture of both. The dispersion function represents all three through one finite convex curve. Its geometry compresses multiple tasks: slopes give probabilities, kinks give atoms, minima give medians, and tail intercepts give the mean. This is useful when convex analysis is more tractable than manipulating a distribution measure directly.

The compression keeps integrability, location, slope, and asymptotic behavior explicit. It discards no distributional information in principle, but it does transform that information: local probability becomes slope and point mass becomes a corner. Numerical use still requires estimating an expectation across many \(u\), so the conceptual compression is not a claim of cost-free computation.

Abstract Reasoning¶

Characterization: equal dispersion functions imply equal one-sided derivatives and hence equal CDFs. Location: any minimizer of a convex \(D_X\) satisfies a zero-in-subgradient condition, which is exactly the median condition \(F_X(u-)\le1/2\le F_X(u)\). Atom detection: \(D'_{+}(u)-D'_{-}(u)=2P(X=u)\). Translation: for \(Y=X+a\), \(D_Y(u)=D_X(u-a)\). Scaling: for real \(b\), \(D_{bX}(u)=|b|D_X(u/b)\) when \(b\ne0\). These relations can be checked without assuming a density.

When \(X\) has finite variance and mean \(\mu\), convexity gives \(D_X(u)\ge|\mu-u|\); integrating the gap across \(u\) yields the variance under the standard identity reported in the dispersion-function literature. This surprising relation is secondary to the first-moment definition and requires the additional second-moment condition; it must not be used for heavy-tailed laws lacking finite variance.

Knowledge Transfer¶

Within probability and statistics, the full object transfers from distribution characterization to robust location, empirical absolute-loss curves, and stochastic comparison. The equations survive unchanged. A sample version \(n^{-1}\sum_i|X_i-u|\) approximates the population curve and makes the geometry visible, although inferential guarantees need their own assumptions.

Outside those fields, absolute-loss minimization and convexity transfer through existing abstractions such as prime:expected_value, but “dispersion function” does not. The same phrase also denotes unrelated objects in waves and plasma physics, demonstrating that lexical travel is not structural recurrence.

Examples¶

Two-point law¶

Let \(P(X=-1)=P(X=1)=1/2\). Then

\[ D_X(u)=\tfrac12|-1-u|+\tfrac12|1-u| =\begin{cases}-u,&u<-1,\\1,&-1\le u\le1,\\u,&u>1.\end{cases} \]

The flat minimum shows that every \(u\in[-1,1]\) is a median. Slopes jump by one at each atom, matching \(2P(X=\pm1)=1\). The right slopes \((-1,0,1)\) recover CDF values \((0,1/2,1)\). This maps every signature role and shows why ordinary differentiability fails at atoms.

Uniform law¶

For \(X\sim\mathrm{Uniform}(0,1)\), direct integration gives \(D_X(u)=1/2-u\) for \(u\le0\), \(D_X(u)=u^2-u+1/2\) for \(0\le u\le1\), and \(D_X(u)=u-1/2\) for \(u\ge1\). On the support, \(D'_X(u)=2u-1\), so \((D'_X(u)+1)/2=u=F_X(u)\). The minimum at \(u=1/2\) identifies the median, while the tail intercepts identify the mean \(1/2\).

Structural Tensions¶

T1: Scalar simplicity versus full-distribution retention. Each evaluation is merely an expected absolute deviation, yet the complete continuum of evaluations preserves the entire law. Diagnostic: Is the analysis using one value \(D_X(u_0)\), or the whole curve and its slopes?

T2: Smooth characterization versus informative kinks. Continuous laws often yield a differentiable profile, while atoms create nondifferentiable corners. Smoothing those corners can erase mass. Diagnostic: Are left and right derivatives equal, and if not, does their gap match an atom probability?

T3: Spread ordering versus location contamination. Curves change under translation as well as dispersion. A pointwise comparison without alignment can mistake a shifted law for a more dispersed one. Diagnostic: Has the intended stochastic order and any required centering or quantile condition been stated explicitly?

T4: Autonomous functional versus Expected Value plus Absolute Loss. Those ingredients generate the formula, but do not alone foreground the entire location-indexed curve, CDF recovery, kink diagnostics, and stochastic-order role. Diagnostic: Can the proposed composite recover a law from one-sided slopes without reintroducing the named profile? If not, an autonomous residual remains.

Structural–Framed Character¶

This is a mathematically structural but domain-specific abstraction. Its evaluation is not institutional or normative, and the defining equations do not depend on human practice. Yet its vocabulary and recognition conditions—random variable, CDF, integrability, one-sided derivative, stochastic order—are native to probability theory.

The term is particularly unsafe across domains because physics uses “dispersion function” for unrelated wave and plasma objects. Import versus recognition is therefore decided by the absolute-loss expectation and CDF-recovery identities, not the name.

Structural Core vs. Domain Accent¶

The structural core is a family of expected losses indexed by candidate locations, with convex geometry translating hidden distributional information into slopes. The domain accent supplies probability measures, integrability, CDFs, atoms, medians, and stochastic orders. Without that apparatus, one has a generic convex objective.

The node does not clear the prime bar: it does not recur literally in three unrelated substrates, and its name has homonyms rather than transfers. It clears the domain-specific bar because it has a stable formula, recognition test, diagnostics, inference rules, and literature-defined autonomy.

The dispersion function uses prime:expected_value: for each \(u\), it averages the random quantity \(|X-u|\) against the probability measure. The proposed DAG relation is compositional because a function-valued profile is not a subtype of the scalar expectation operation.

domain_specific:probability_distribution is the source object and is recoverable from the profile, but the functional is not a subtype of a distribution. prime:dispersion is explicitly declined as a parent because its propagation-separation identity is unrelated. domain_specific:variance_function is a neighbor with a different argument and output.

Relationships to Other Abstractions¶

Current abstraction Dispersion Function Domain-specific

Parents (1) — more general patterns this builds on

Dispersion Function presupposes Expected Value Prime

The dispersion function uses prime:expected_value: for each \(u\), it averages the random quantity \(|X-u|\) against the probability measure.

Hierarchy paths (3) — routes to 2 parentless roots

Dispersion Function → Expected Value → Aggregation → Micro Macro Linkage

Show alternative paths (2)

Neighborhood in Abstraction Space¶

Dispersion Function sits in a sparse region of the domain-specific corpus (71^st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Variance function: conditional variance as a function of mean; test the argument \(\mu\) versus free reference point \(u\).
Mean absolute deviation: one evaluation at a chosen center; test whether the whole curve is retained.
Expected absolute loss: the generic objective; the dispersion function is its complete population profile for one law.
Plasma dispersion function: a complex special function in kinetic plasma theory.
Wave dispersion relation: connects frequency and wave number, not location to expected distance.
Dispersive stochastic order: an order relation between laws; the function may help define or study an order but is not itself the order.

References¶

[1] J. Muñoz-Pérez and A. Sánchez-Gómez, “A Characterization of the Distribution Function: The Dispersion Function,” Statistics & Probability Letters 10, no. 3 (1990): 235–239, https://doi.org/10.1016/0167-7152(90)90080-Q. registry ↩a ↩b

[2] Moshe Shaked and J. George Shanthikumar, Stochastic Orders (Springer, 2007), https://doi.org/10.1007/978-0-387-34675-5. registry ↩