A first attempt at measuring sets is seductively simple: assign a size to every subset of the line. Intervals should have their usual lengths, translating a set should not change its size, and a countable disjoint union should have size equal to the sum of the sizes. These requirements match the way length behaves in calculus, probability, and geometry. The surprise is that the three demands cannot coexist on the whole power set of $\mathbb R$.

The obstruction is not a technical inconvenience. It says that the domain of a measure is part of the mathematical structure. Before asking how large a set is, we must decide which sets are measurable. A sigma-algebra is the device that makes this decision stable under complements, countable unions, and countable intersections. It is the language in which limiting events, level sets of functions, Borel sets, product spaces, and conditional information are all expressed.

[example: The Vitali Obstruction] Let $\sim$ be the [equivalence relation](/page/Equivalence%20Relation) on $[0,1]$ defined by $x\sim y$ exactly when $x-y\in\mathbb Q$. By the [axiom of choice](/page/Axiom%20of%20Choice), choose a set $V\subset[0,1]$ containing exactly one point from each equivalence class. Suppose, toward a contradiction, that a countably additive translation-invariant measure $m$ is defined on every subset of $\mathbb R$ and agrees with interval length. For $q\in\mathbb Q\cap[-1,1]$, write \begin{align*} V+q=\{v+q:v\in V\}. \end{align*} These translates are pairwise disjoint. Indeed, if $z\in (V+q_1)\cap(V+q_2)$, then for some $v_1,v_2\in V$, \begin{align*} z=v_1+q_1=v_2+q_2. \end{align*} Hence \begin{align*} v_1-v_2=q_2-q_1\in\mathbb Q, \end{align*} so $v_1$ and $v_2$ lie in the same equivalence class. Since $V$ contains exactly one representative from each class, $v_1=v_2$, and then \begin{align*} v_1+q_1=v_1+q_2 \end{align*} forces $q_1=q_2$. The rational translates cover $[0,1]$ and stay inside $[-1,2]$. If $x\in[0,1]$, let $v\in V$ be the chosen representative of the equivalence class of $x$. Then $x-v\in\mathbb Q$, and because $x,v\in[0,1]$, \begin{align*} -1\le x-v\le 1. \end{align*} Thus, with $q=x-v$, we have $q\in\mathbb Q\cap[-1,1]$ and $x=v+q\in V+q$. Conversely, if $y\in V+q$ with $q\in\mathbb Q\cap[-1,1]$, then $y=v+q$ for some $v\in[0,1]$, so \begin{align*} -1\le v+q\le 2. \end{align*} Therefore \begin{align*} [0,1]\subset \bigcup_{q\in\mathbb Q\cap[-1,1]}(V+q)\subset[-1,2]. \end{align*} Since $\mathbb Q\cap[-1,1]$ is countable, enumerate it as $(q_n)_{n\in\mathbb N}$. Countable additivity applies to the pairwise disjoint family $(V+q_n)_{n\in\mathbb N}$, giving \begin{align*} m\left(\bigcup_{q\in\mathbb Q\cap[-1,1]}(V+q)\right) &=m\left(\bigcup_{n=1}^{\infty}(V+q_n)\right)\\ &=\sum_{n=1}^{\infty}m(V+q_n). \end{align*} By [translation invariance](/theorems/4911), \begin{align*} m(V+q_n)=m(V) \end{align*} for every $n$, so \begin{align*} m\left(\bigcup_{q\in\mathbb Q\cap[-1,1]}(V+q)\right) =\sum_{n=1}^{\infty}m(V). \end{align*} Monotonicity follows from countable additivity: if $A\subset B$, then $B=A\cup(B\setminus A)$ is a disjoint union, so \begin{align*} m(B)=m(A)+m(B\setminus A)\ge m(A). \end{align*} Using the inclusions above and agreement with interval length, \begin{align*} 1=m([0,1]) \le m\left(\bigcup_{q\in\mathbb Q\cap[-1,1]}(V+q)\right) \le m([-1,2])=3. \end{align*} Now $V\subset[0,1]$, so monotonicity gives \begin{align*} 0\le m(V)\le m([0,1])=1. \end{align*} If $m(V)=0$, then \begin{align*} \sum_{n=1}^{\infty}m(V)=\sum_{n=1}^{\infty}0=0, \end{align*} contradicting the lower bound $1$. If $m(V)>0$, then the partial sums satisfy \begin{align*} \sum_{n=1}^{N}m(V)=N\,m(V), \end{align*} which exceed $3$ for sufficiently large $N$, so \begin{align*} \sum_{n=1}^{\infty}m(V)=\infty, \end{align*} contradicting the upper bound $3$. Hence no measure on all subsets of $\mathbb R$ can simultaneously be countably additive, translation-invariant, and agree with interval length. [/example]

The example shows why measure theory begins by restricting attention to a well-behaved family of sets. The family must still be large enough for analysis: if a limiting process gives a countable union of measurable sets, the result should remain measurable; if an event is measurable, its complement should also be measurable. The definition packages precisely those closure requirements.

## Definition

The basic object is not a measure, but the domain on which a measure could live. The empty set and the whole space must be available, complements must be allowed because negation is a basic operation on events, and countable unions must be allowed because limits in analysis produce countably many approximations.

[definition: Sigma-Algebra] Let $X$ be a set. A sigma-algebra on $X$ is a collection $\mathcal F \subset \mathcal P(X)$ such that: 1. $X \in \mathcal F$. 2. If $A \in \mathcal F$, then $X \setminus A \in \mathcal F$. 3. If $(A_n)_{n \in \mathbb N}$ is a sequence of sets in $\mathcal F$, then $\bigcup_{n=1}^{\infty} A_n \in \mathcal F$. [/definition]

This definition is the only structure needed before measures enter. It gives us a universe of admissible sets and guarantees that the usual limiting operations on events stay inside that universe.

## Basic Measurable Structures

A collection of subsets cannot be interpreted without knowing the universe in which complements are taken. This creates the need to name the underlying set and the sigma-algebra as a single object, so that measurability always has a specified ambient space.

[definition: Measurable Space] A measurable space is a pair $(X, \mathcal F)$ where $X$ is a set and $\mathcal F$ is a sigma-algebra on $X$. [/definition]

The members of $\mathcal F$ are called measurable sets. At this stage no sizes have been assigned; we have only specified which questions about membership are allowed. The next step is to put numerical size on those allowed questions. This cannot be done by a function on arbitrary subsets unless the domain is already closed under the countable unions appearing in additivity, so the natural object is a size function whose domain is the sigma-algebra of a measurable space.

[definition: Measure on a Measurable Space] Let $(X,\mathcal F)$ be a measurable space. A measure on $(X,\mathcal F)$ is a function \begin{align*} \mu:\mathcal F &\to [0,\infty] \end{align*} such that $\mu(\varnothing)=0$ and, whenever $(A_n)_{n\in\mathbb N}$ is a pairwise disjoint sequence in $\mathcal F$, \begin{align*} \mu\left(\bigcup_{n=1}^{\infty}A_n\right)=\sum_{n=1}^{\infty}\mu(A_n). \end{align*} [/definition]

This definition is not meant to replace the separate page on measures. Its role here is to show why the sigma-algebra axioms are exactly the closure assumptions needed before countable additivity can even be stated. Once the domain has been separated from the size function, it becomes meaningful to compare different domains of measurable information.

The smallest possible domain arises when the intended model has no observable distinction between individual points: the only reliable question is whether the outcome lies somewhere in $X$ or nowhere at all. Naming this degenerate case gives a lower bound against which richer measurable structures can be measured.

[definition: Indiscrete Sigma-Algebra] Let $X$ be a set. The indiscrete sigma-algebra on $X$ is \begin{align*} \{\varnothing, X\}. \end{align*} [/definition]

The indiscrete sigma-algebra forgets every proper subset of $X$. A function out of such a space has very few measurable level sets, so this sigma-algebra is too small for most analytic purposes. At the other end, finite and countable sample spaces often come with complete observational access: after a die roll, for instance, asking whether the outcome belongs to any specified subset is a legitimate event. The next definition records the measurable structure forced by allowing every membership question.

[definition: Discrete Sigma-Algebra] Let $X$ be a set. The discrete sigma-algebra on $X$ is the power set $\mathcal P(X)$. [/definition]

On a [countable set](/page/Countable%20Set), the discrete sigma-algebra is often the natural choice: every subset can be counted, summed over, or assigned probability mass. On uncountable spaces, the same choice can conflict with geometric requirements such as translation invariance.

[example: Coin Flips and the Power Set] Let $X=\{H,T\}$ be the sample space for one coin flip. Its power set is obtained by listing all subsets of the two-point set: \begin{align*} \mathcal P(X)=\{\varnothing,\{H\},\{T\},\{H,T\}\} =\{\varnothing,\{H\},\{T\},X\}. \end{align*} Thus the discrete sigma-algebra on $X$ records every membership question about the outcome: $\{H\}$ records that heads occurred, $\{T\}$ records that tails occurred, $X$ records that some outcome occurred, and $\varnothing$ records the impossible event. Now consider two flips with sample space \begin{align*} X^2=\{(H,H),(H,T),(T,H),(T,T)\}. \end{align*} If the observer records only the first flip, then the two distinguishable nontrivial events are \begin{align*} \{H\}\times X&=\{(H,H),(H,T)\},\\ \{T\}\times X&=\{(T,H),(T,T)\}. \end{align*} These two sets are complements in $X^2$, since \begin{align*} X^2\setminus(\{H\}\times X)=\{T\}\times X, \qquad X^2\setminus(\{T\}\times X)=\{H\}\times X. \end{align*} Their union is the whole space and their intersection is empty: \begin{align*} (\{H\}\times X)\cup(\{T\}\times X)&=X^2,\\ (\{H\}\times X)\cap(\{T\}\times X)&=\varnothing. \end{align*} So the sigma-algebra generated by first-flip information is \begin{align*} \{\varnothing,\{H\}\times X,\{T\}\times X,X^2\}. \end{align*} It contains the event that the first flip was heads and the event that the first flip was tails, but it does not contain $\{(H,T)\}$ because the only nonempty proper sets in the displayed sigma-algebra are $\{(H,H),(H,T)\}$ and $\{(T,H),(T,T)\}$. Thus the underlying outcome space still has four points, while this sigma-algebra retains only the information needed to distinguish the first coordinate. [/example]