1 Definition and basic properties

An atomic measure is a measure that assigns positive mass to at least one point in its space. In many common settings, it is used to describe distributions or data structures that place weight on isolated outcomes rather than spreading continuously across an interval or region. Atomic measures are especially useful in measure theory and probability, where they provide a rigorous language for point masses and discrete components.

1.1 Measure-theoretic definition

Let \((X,\Sigma)\) be a measurable space and \(\mu\) a measure on it. A point \(x \in X\) is called an atom of \(\mu\) if \(\mu(\{x\}) > 0\). More generally, some texts define an atom as a measurable set \(A\) with positive measure such that every measurable subset \(B \subseteq A\) has either \(\mu(B)=0\) or \(\mu(B)=\mu(A)\). For measures on standard spaces, these viewpoints align closely with point masses.

1.2 Atoms and point masses

A point mass is the amount of measure concentrated at a single point. If \(\mu(\{x\}) = a > 0\), then \(x\) carries atomic weight \(a\). Several atoms may occur in one measure, and their masses may differ. In probability theory, these masses represent the probability assigned to exact outcomes.

1.3 Support of an atomic measure

The support of an atomic measure is the collection of points where mass is concentrated, often understood as the set of atoms or the closure of that set in a topological setting. For an atomic measure, the support is made up of isolated contributions rather than a continuous density. This gives such measures a distinctly discrete character.

1.3.1 Countable versus finite support

An atomic measure may have finitely many atoms or countably many atoms. Finite support is common in simple categorical models, while countable support appears in sequences of weighted outcomes or distributions on the integers. In many practical situations, countable support is still manageable because the total mass can be expressed as a convergent sum.

1.3.2 Purely atomic measures

A purely atomic measure is one that is entirely concentrated on atoms. Equivalently, the measure can be written as a sum of point masses, possibly over a finite or countable index set. Such measures have no continuous part, and every measurable set of positive measure contains at least one atom of positive weight.

1.4 Examples of atomic measures

Common examples include a Dirac measure at a point, a finite sum of Dirac measures, and the counting measure on a countable set. Empirical distributions formed from observed data are also atomic, since they place equal or weighted mass on sample points. Categorical probability distributions provide another familiar instance.

2 Decomposition of measures

Many measures can be decomposed into parts with different structural features. Atomic measures represent the discrete component in such decompositions, often alongside non-atomic or continuous components. This perspective is central in understanding complex measures built from mixtures of simple pieces.

2.1 Atomic and non-atomic components

The atomic part of a measure consists of all point masses. The non-atomic part assigns no positive mass to singletons. When both are present, the measure combines discrete spikes with a spread-out background. This separation is useful in analysis and probability, where discrete and continuous behavior are often studied together.

2.2 Lebesgue decomposition context

In the setting of Lebesgue decomposition, a measure may be decomposed relative to another reference measure into absolutely continuous, singular continuous, and atomic components. Atomicity concerns whether some of the mass is concentrated at points. Although the exact decomposition depends on the ambient space and reference measure, the atomic contribution is often easy to identify once point masses are present.

2.3 Discrete-continuous mixtures

Mixed measures arise when a distribution has both a continuous density and discrete spikes. Such mixtures are common in applied modeling, for example when a variable can take a special fixed value and otherwise vary continuously. The atomic terms capture the fixed outcomes, while the continuous term describes the remaining variability.

2.4 Singular measures and atomicity

Not every singular measure is atomic. A singular measure may concentrate on a thin set without assigning positive mass to points, as in the case of certain fractal-like distributions. Atomic measures are a special, highly concentrated subclass in which mass accumulates directly at isolated points.

3 Atomic measures in probability theory

In probability theory, atomic measures correspond to discrete probability distributions. They are foundational for describing random variables with countable ranges, exact outcomes, and empirical or mixed models. Their structure makes them easy to interpret and often straightforward to compute with.

3.1 Discrete probability distributions

A discrete probability distribution is a probability measure whose mass is concentrated on a finite or countable set of outcomes. Each outcome has a probability equal to the measure of its singleton set. Examples include the Bernoulli, binomial, geometric, and Poisson distributions, though the last has countably many atoms rather than finitely many.

3.2 Probability mass functions

For a discrete random variable, the probability mass function assigns to each value the probability of that value occurring. This function is the numerical representation of an atomic probability measure. The total probability is the sum of all point masses, which must equal one.

3.3 Empirical measures

Empirical measures are formed from observed data by placing mass on observed sample points. They provide a simple statistical approximation to an unknown distribution. Because they assign positive mass only to observed outcomes, they are purely atomic.

3.3.1 Sample-based distributions

In the simplest form, each observation receives equal weight \(1/n\), where \(n\) is the sample size. The resulting measure summarizes the sample without imposing a parametric shape. It is widely used in nonparametric statistics and resampling methods.

3.3.2 Weighted empirical measures

Weighted empirical measures allow different observations to contribute different amounts of mass. This is useful when observations have survey weights, importance weights, or varying reliability. The measure remains atomic, but the atom sizes reflect the chosen weighting scheme.

3.4 Mixture models with atomic parts

Mixture models may combine discrete and continuous components within a single distribution. An atomic component can represent special outcomes, structural zeros, or repeated fixed values. Such models are common when observed data show both exact concentrations and broader variation.

4 Statistical applications

Atomic measures appear in many statistical settings where outcomes occur at isolated values or where models need explicit point masses. They are important in practical data analysis because many variables are partly discrete, partly continuous, or observed with ties and special values.

4.1 Modeling count data

Count data are naturally described by atomic measures, since they take values in a countable set such as the nonnegative integers. Models for event counts, occurrence frequencies, and tallies typically rely on discrete probability measures. The atomic structure reflects the integer-valued nature of the observations.

4.2 Bayesian priors with point masses

Bayesian modeling sometimes uses priors that place positive probability on exact parameter values. Such priors are atomic or mixed atomic-continuous measures. They are useful for encoding sparsity, selecting among competing hypotheses, or allowing special fixed values with nonzero prior weight.

4.3 Categorical and multinomial models

Categorical and multinomial models are built on atomic probability measures over a finite set of categories. Each category corresponds to a point mass, and the total probability is distributed among them. These models are standard tools for survey responses, classification, and discrete choice analysis.

4.4 Survival and event-time models

In survival and event-time analysis, atomic measures can model exact event times, especially when outcomes are recorded with limited time resolution or when certain times occur repeatedly. They are also useful for describing distributions with fixed failure times or administrative recording points.

4.4.1 Tied event times

Tied event times occur when multiple events are observed at the same time point. This can produce visible atoms in the estimated event-time distribution. Such ties may arise from rounding, measurement limits, or genuine clustering of occurrences.

4.4.2 Mass at zero or other fixed points

Some event-time variables have a positive probability of taking the value zero or another predetermined time. This feature is often modeled by a point mass at the fixed point, combined with a continuous distribution for positive times. The atomic component captures immediate events or structural concentrations.

5 Mathematical properties

Atomic measures have several useful formal properties that follow from the axioms of measure theory. These properties make them suitable for both theoretical work and practical calculations.

5.1 Total mass and normalization

A finite atomic measure has a finite total mass equal to the sum of the masses of its atoms. If the total mass is one, the measure is a probability measure. Normalization is often achieved by dividing by the total mass when the measure is nonzero and finite.

5.2 Sigma-additivity

Like all measures, atomic measures are countably additive over disjoint measurable sets. This means the measure of a countable union of disjoint sets equals the sum of their individual measures. For atomic measures, this property is often expressed as the sum of point masses over the atoms contained in the set.

5.3 Integrals with respect to atomic measures

Integration against an atomic measure reduces to a weighted sum over atoms. If \(\mu = \sum_i a_i \delta_{x_i}\), then integrating a function \(f\) gives \(\int f\,d\mu = \sum_i a_i f(x_i)\), provided the sum is well defined. This simplification is one reason atomic measures are convenient in computation.

5.4 Convergence of atomic measures

Sequences of atomic measures may converge in various senses, including weak convergence. Limits can retain, lose, or redistribute atomic mass depending on how the atoms move and how their weights change. Such convergence is important in approximation theory, estimation, and asymptotic analysis.

6 Estimation and inference

When atomic measures are used in statistical work, the main inferential tasks involve estimating atom locations, atom sizes, and the number of distinct support points. These tasks can be straightforward in simple settings but more challenging when atomic and continuous components are mixed.

6.1 Estimating point masses from data

Point masses are often estimated by relative frequencies or weighted frequencies. For observed discrete outcomes, the natural estimator of an atom’s mass is the proportion of observations at that point. When data are noisy or incomplete, estimation may require smoothing or latent-variable methods.

6.2 Maximum likelihood methods

Maximum likelihood estimation for atomic models typically reduces to fitting category probabilities or point-mass weights. In mixture settings, the likelihood may involve both discrete and continuous contributions. Optimization can then be more involved, especially when the locations of atoms are unknown.

6.3 Nonparametric estimation

Nonparametric methods frequently estimate an atomic distribution directly from the data without imposing a fixed parametric family. The empirical measure is the most basic example. More elaborate approaches may estimate sparse support, cluster repeated values, or separate discrete and continuous parts.

6.4 Testing for discrete support

Statistical tests may be used to determine whether a distribution contains point masses or whether observed repetitions are consistent with a continuous model. Such procedures can involve goodness-of-fit methods, tests for zero inflation, or diagnostics for excess ties. The goal is to identify whether atomic structure is present and how strong it is.

Atomic measures are closely linked to several standard objects in measure theory and probability. These related concepts clarify different ways in which mass can be distributed over a space.

7.1 Dirac measure

A Dirac measure places all mass at a single point. It is the simplest example of an atomic measure and serves as a building block for more complicated discrete distributions. It is often denoted by \(\delta_x\) for a point \(x\).

7.2 Counting measure

Counting measure assigns to a set the number of elements it contains, when that number is finite or countably infinite. On discrete spaces, it is a basic example of a purely atomic measure. Each singleton has mass one.

7.3 Discrete measure

A discrete measure is a measure supported on a countable set. It is often used interchangeably with atomic measure in probabilistic contexts, although some authors distinguish general atomic measures from those supported on a countable collection of points. In practice, both terms usually signal point-concentrated mass.

7.4 Non-atomic measure

A non-atomic measure assigns zero mass to every singleton. Such measures describe spread-out distributions without point masses. They provide the contrasting case to atomic measures and are central in the study of continuous phenomena.