A function often appears before the measure does. We may write down a temperature field, a random payoff, a candidate density, or a solution of a differential equation before asking whether it can be integrated. The first obstruction is not size, continuity, or differentiability. It is observability: if we ask whether the value of the function lies in a set we can observe in the codomain, does that question correspond to a set we can observe in the domain? This is the reason measurability is phrased through inverse images. A function $f: E \to G$ transports questions backward: an event $A \subset G$ becomes the event $f^{-1}(A) \subset E$. The function is measurable exactly when every observable question about values has an observable answer on the source. [example: A Function That Cannot Be Integrated] Let $V \subset [0,1]$ be non-Lebesgue-measurable, and define $f:[0,1]\to\mathbb{R}$ by $f=\mathbb{1}_V$. Thus, for each $x\in[0,1]$, \begin{align*} f(x)=1 \text{ if } x\in V, \qquad f(x)=0 \text{ if } x\notin V. \end{align*} We compute the threshold set at level $1/2$. If $x\in V$, then $f(x)=1>1/2$, so $x\in\{y\in[0,1]:f(y)>1/2\}$. If $x\notin V$, then $f(x)=0\le 1/2$, so $x\notin\{y\in[0,1]:f(y)>1/2\}$. Therefore the two inclusions give \begin{align*} \{x\in[0,1]:f(x)>1/2\}=V. \end{align*} When $[0,1]$ is equipped with the Lebesgue measurable sets, this level set is not measurable because $V$ was chosen to be non-Lebesgue-measurable. Hence $f$ is not a permissible Lebesgue integrand: the basic value-question “is $f$ larger than $1/2$?” pulls back to a set outside the measurable structure on the domain. [/example] The example shows that measurability is not a cosmetic condition added after a function is already understood. It decides whether the function belongs to the measure-theoretic universe at all. From this point of view, [measurable functions](/page/Measurable%20Functions) are the morphisms of measure theory: they preserve the structure that matters. ## Definition ### Measurable Structures Before defining measurable functions, we need to name the measurable structure that lives on a set. The relevant structure is not a measure yet, but a collection of subsets stable under the set operations that countable limiting arguments require. [definition: Sigma-Algebra] Let $E$ be a set. A sigma-algebra on $E$ is a collection $\mathcal{E} \subset \mathcal{P}(E)$ such that $E \in \mathcal{E}$, such that $E \setminus A \in \mathcal{E}$ whenever $A \in \mathcal{E}$, and such that $\bigcup_{n=1}^{\infty} A_n \in \mathcal{E}$ whenever $A_n \in \mathcal{E}$ for every $n \in \mathbb{N}$. [/definition] A sigma-algebra by itself is only a list of admissible subsets; it does not name the ambient set as part of the object. Since measurability depends on both the points being mapped and the subsets allowed to be observed, the basic domain for measure theory must keep these two pieces together. This becomes important whenever the same set carries different sigma-algebras, because the identity map may be measurable for one choice of observable sets and not for another. To make later definitions unambiguous, we need a single object that records both the underlying set and the sigma-algebra of observable subsets on it. [definition: Measurable Space] A measurable space is a pair $(E, \mathcal{E})$ where $E$ is a set and $\mathcal{E}$ is a sigma-algebra on $E$. [/definition] ### The Inverse-Image Condition With measurable spaces defined, the central problem is how a function should respect their observable sets. A function $f:E\to G$ transports questions backward: a set $A\subset G$ of observable values becomes the inverse image $f^{-1}(A)\subset E$. The function is admissible precisely when every observable value-question in the target pulls back to an observable source-event. [definition: Measurable Function] Let $(E, \mathcal{E})$ and $(G, \mathcal{G})$ be measurable spaces. A function $f: E \to G$ is $\mathcal{E}/\mathcal{G}$-measurable if $f^{-1}(A) \in \mathcal{E}$ for every $A \in \mathcal{G}$. [/definition] When the sigma-algebras are understood, we simply say that $f$ is measurable. The notation $\mathcal{E}/\mathcal{G}$ is useful when several sigma-algebras live on the same underlying sets, since the same pointwise function may be measurable for one choice and not for another. Thus the definition of measurable function says exactly that a map preserves observable value-events by pulling them back to observable source-events. ### Borel and Probabilistic Targets The most common codomain in analysis is $\mathbb{R}$, not with every subset declared measurable, but with the sigma-algebra generated by open sets. The next definition is needed because it turns a [topological space](/page/Topological%20Space) into a measurable space without discarding the open-set information used in continuity. [definition: Borel Sigma-Algebra] Let $(X, \tau)$ be a topological space. The Borel sigma-algebra on $X$, denoted $\mathcal{B}(X)$, is the smallest sigma-algebra on $X$ containing every [open set](/page/Open%20Set) in $\tau$. [/definition] Topology often supplies the observable sets in the codomain: open sets are the primitive data, and the Borel sigma-algebra records all measurable events forced by them. Thus a function into a topological space can be tested for measurability without choosing a separate sigma-algebra by hand. The resulting notion is especially useful because it includes continuous maps but also many discontinuous maps whose level events are still observable. This motivates a named class of maps whose source has measurable structure and whose target is observed through its Borel sets. [definition: Borel Measurable Function] Let $(E, \mathcal{E})$ be a measurable space and let $(G, \tau)$ be a topological space. A function $f: E \to G$ is Borel measurable if it is $\mathcal{E}/\mathcal{B}(G)$-measurable. [/definition] For real-valued functions on a [measure space](/page/Measure%20Space), the codomain sigma-algebra is almost always $\mathcal{B}(\mathbb{R})$. Naming this convention prevents ambiguity later: the measure $\mu$ lives on the source, while the measurable value-events are Borel subsets of the real line. This terminology also separates two roles that are often written together: $\mu$ determines which source sets are negligible or integrable, while the Borel sigma-algebra determines which real value-events can be pulled back. Because integration will always use the measure on the source space, it is useful to attach the phrase $\mu$-measurable to this real-valued situation explicitly. The next definition fixes the convention used throughout integration theory: a real-valued function is tested against Borel subsets of $\mathbb{R}$, but the resulting inverse images must belong to the sigma-algebra on which $\mu$ is defined. This is the precise condition that lets later expressions such as level sets, truncations, and integrals be interpreted inside the given measure space. [definition: $\mu$-Measurable Real-Valued Function] Let $(X, \mathcal{M}, \mu)$ be a measure space. A function $f: X \to \mathbb{R}$ is $\mu$-measurable if $f$ is $\mathcal{M}/\mathcal{B}(\mathbb{R})$-measurable. [/definition] In probability, one wants to assign probabilities to value-events such as $\{\omega:X(\omega)\in A\}$. This is possible only when every measurable event $A$ in the value space pulls back to an event in the sample space, so the usual probabilistic term is just measurability in this setting. Without this condition, expressions such as $\mathbb{P}(X\in A)$ would not necessarily be defined, even though $X$ is a pointwise function on outcomes. The definition below fixes that terminology by treating a [random variable](/page/Random%20Variable) as an observable map from outcomes to values, rather than as an arbitrary function on the sample space. [definition: Random Variable] Let $(\Omega, \mathcal{F}, \mathbb{P})$ be a probability space and let $(E, \mathcal{E})$ be a measurable space. A random variable with values in $(E, \mathcal{E})$ is a measurable function $X: (\Omega, \mathcal{F}) \to (E, \mathcal{E})$. [/definition] Thus random variables, Borel maps, and ordinary measurable functions are not separate ideas. They are the same inverse-image condition used in different mathematical languages. ## Testing Measurability ### Generators The definition asks us to check every measurable subset of the codomain. That is usually impossible directly, because a Borel sigma-algebra contains far more sets than the open intervals used to generate it. The central technique is to test measurability on a generating family and then let closure under sigma-algebra operations do the rest. A generated sigma-algebra packages this technique. The next definition is needed because it lets us replace a large sigma-algebra by a smaller collection of basic sets that force it. [definition: Generated Sigma-Algebra] Let $G$ be a set and let $\mathcal{A} \subset \mathcal{P}(G)$. The sigma-algebra generated by $\mathcal{A}$, denoted $\sigma(\mathcal{A})$, is the intersection of all sigma-algebras on $G$ that contain $\mathcal{A}$. [/definition] The direct definition of measurability can demand checking an entire generated sigma-algebra, even though that sigma-algebra may be far too large to describe explicitly. Since inverse images preserve complements and countable unions, it should be enough to verify the inverse-image condition on the smaller family that generates the target sigma-algebra. The point that needs justification is that no new obstruction appears when complements and countable unions are formed after taking inverse images. The useful test is therefore to ask whether control on a generating collection forces control on every set generated from it. [quotetheorem:525] For real-valued functions, even checking all open sets is more work than necessary. The order structure of $\mathbb{R}$ suggests a sharper test: threshold events should determine the Borel behavior of the function because intervals and rays generate the usual Borel sigma-algebra. This matters in practice because sublevel sets are often visible from inequalities, monotonicity, or formulas, while arbitrary Borel preimages are not. The criterion below replaces arbitrary Borel preimages by the much simpler family of sublevel sets. [quotetheorem:4878] The version with sets of the form $\{f \le t\}$ is especially common in undergraduate measure theory. Equivalent versions use $\{f<t\}$, $\{f>t\}$, or $\{f \ge t\}$ for every real $t$; the equivalence follows from countable unions, countable intersections, and complements. [example: Checking Measurability of a Step Function] Let $(X,\mathcal{M})$ be a measurable space, let $A,B\in\mathcal{M}$, and define $f:X\to\mathbb{R}$ by \begin{align*} f(x)=2\mathbb{1}_A(x)-\mathbb{1}_B(x). \end{align*} The four membership patterns partition $X$. Since $\mathcal{M}$ is closed under complements and finite intersections, the sets $A\cap B$, $A\cap B^c$, $A^c\cap B$, and $A^c\cap B^c$ all belong to $\mathcal{M}$. On each atom, the value of $f$ is fixed. If $x\in A\cap B$, then \begin{align*} f(x)=2\cdot 1-1=1. \end{align*} If $x\in A\cap B^c$, then \begin{align*} f(x)=2\cdot 1-0=2. \end{align*} If $x\in A^c\cap B$, then \begin{align*} f(x)=2\cdot 0-1=-1. \end{align*} If $x\in A^c\cap B^c$, then \begin{align*} f(x)=2\cdot 0-0=0. \end{align*} We now compute the threshold sets $\{x\in X:f(x)>t\}$. If $t<-1$, all four possible values $-1,0,1,2$ are greater than $t$, so \begin{align*} \{x\in X:f(x)>t\}=X. \end{align*} If $-1\le t<0$, the values $0,1,2$ are greater than $t$ and the value $-1$ is not, so \begin{align*} \{x\in X:f(x)>t\}=(A\cap B)\cup(A\cap B^c)\cup(A^c\cap B^c). \end{align*} If $0\le t<1$, the values $1,2$ are greater than $t$ and the values $-1,0$ are not, so \begin{align*} \{x\in X:f(x)>t\}=(A\cap B)\cup(A\cap B^c)=A. \end{align*} If $1\le t<2$, only the value $2$ is greater than $t$, so \begin{align*} \{x\in X:f(x)>t\}=A\cap B^c. \end{align*} If $t\ge 2$, no possible value of $f$ is greater than $t$, so \begin{align*} \{x\in X:f(x)>t\}=\varnothing. \end{align*} Each displayed threshold set is a finite union of sets in $\mathcal{M}$, hence belongs to $\mathcal{M}$. Therefore the step function $f=2\mathbb{1}_A-\mathbb{1}_B$ is measurable: its value-level questions reduce to finitely many measurable membership questions about $A$ and $B$. [/example] ### Continuity as a Source of Measurability Continuity controls inverse images of open sets, while Borel measurability asks about inverse images of every Borel set. The obstruction is that the Borel sigma-algebra contains many non-open sets; the generator principle bridges this gap by extending the open-set check to all Borel sets. [quotetheorem:8506] This result explains why measurable functions are far more flexible than continuous functions. Continuity is enough for measurability, but it is not necessary. Indicators, step functions, monotone functions, and pointwise limits of continuous functions are often measurable even when discontinuities are unavoidable. ## Building Measurable Functions ### Composition Once measurability has been verified for a few basic functions, the next question is whether common constructions preserve it. Composition is the most structural operation, because it says that measurable observations can be fed into further measurable observations. [quotetheorem:1019] This is why applying a Borel function to a measurable real-valued function keeps measurability. Absolute values, powers, exponentials, logarithms on their domains, maxima, minima, and truncations all fit this pattern. [example: Absolute Value and Truncation] Let $(X,\mathcal{M})$ be a measurable space and let $f:X\to\mathbb{R}$ be measurable. Define $\phi:\mathbb{R}\to\mathbb{R}$ by $\phi(y)=|y|$. Since $\phi$ is continuous, it is Borel measurable by *[Continuous Maps Are Borel Measurable](/theorems/8506)*. Therefore, by *[Composition of Measurable Functions](/theorems/1019)*, the composition $\phi\circ f$ is measurable, and for every $x\in X$, \begin{align*} (\phi\circ f)(x)=\phi(f(x))=|f(x)|. \end{align*} Fix $n\in\mathbb{N}$ and define $\psi_n:\mathbb{R}\to\mathbb{R}$ by $\psi_n(y)=\min\{|y|,n\}$. The identity \begin{align*} \min\{|y|,n\}=\frac{|y|+n-\bigl||y|-n\bigr|}{2} \end{align*} shows that $\psi_n$ is continuous, since it is obtained from continuous real functions using addition, subtraction, scalar multiplication, and absolute value. Hence $\psi_n$ is Borel measurable by *Continuous Maps Are Borel Measurable*. Applying *Composition of Measurable Functions* again, $\psi_n\circ f$ is measurable, and for every $x\in X$, \begin{align*} (\psi_n\circ f)(x)=\psi_n(f(x))=\min\{|f(x)|,n\}. \end{align*} The truncation is bounded because for every $x\in X$, \begin{align*} 0\le \min\{|f(x)|,n\}\le n. \end{align*} Since $f(x)\in\mathbb{R}$, the number $|f(x)|$ is finite, so once $n\ge |f(x)|$ one has $\min\{|f(x)|,n\}=|f(x)|$. Thus the bounded measurable functions $x\mapsto\min\{|f(x)|,n\}$ increase pointwise to $x\mapsto |f(x)|$, giving bounded measurable approximations to the possibly unbounded function $|f|$. [/example] ### Products and Algebra Algebraic operations on two measurable functions first form the pair of their values. To ask whether this paired value is measurable, the product set must carry a sigma-algebra that recognizes measurable information from each coordinate. [definition: Product Sigma-Algebra] Let $(E, \mathcal{E})$ and $(G, \mathcal{G})$ be measurable spaces. The product sigma-algebra on $E \times G$, denoted $\mathcal{E} \otimes \mathcal{G}$, is the sigma-algebra generated by all measurable rectangles $A \times B$ with $A \in \mathcal{E}$ and $B \in \mathcal{G}$. [/definition] After putting a sigma-algebra on the product, the key question is whether two separate measurable observations can be bundled without losing measurability. This must be checked because the pair map has values in a product space, not in either original codomain alone. [quotetheorem:8507] Once measurable functions can be paired, ordinary real algebra becomes available only if the operation applied to the pair is itself measurable. Addition, multiplication, maxima, and related operations are controlled by their continuity on $\mathbb{R}^2$, which is what turns pointwise algebra into measurable algebra. The obstruction to rule out is that a formula built pointwise from measurable functions might introduce new value-events that were not measurable in the original functions. The result below shows that continuous real operations do not create this problem for the standard pointwise operations. [quotetheorem:8508] Signed functions create a difficulty for integration because positive and negative contributions can cancel while each side may need separate control. Splitting a real-valued function into its upward and downward components gives two nonnegative functions that retain the original function's size information. [definition: Positive and Negative Parts] Let $f: X \to \mathbb{R}$ be a function. The positive part and negative part of $f$ are the functions $f^+, f^-: X \to [0,\infty)$ defined by $f^+(x)=\max\{f(x),0\}$ and $f^-(x)=\max\{-f(x),0\}$. [/definition] After this definition, the identities $f=f^+-f^-$ and $|f|=f^+ + f^-$ explain why these parts are useful. If $f$ is measurable, then $f^+$ and $f^-$ are measurable because they are obtained by composing $f$ with continuous functions. ## Real-Valued and Extended-Valued Functions Many natural limits of measurable functions are not finite everywhere. A sequence of increasing nonnegative functions may diverge to $\infty$, and a potential may take the value $-\infty$ on a singular set. The next definition is needed so that these limiting objects have a codomain. [definition: Extended Real Line] The extended real line is the ordered set $[-\infty,\infty]=\mathbb{R} \cup \{-\infty,\infty\}$ with the usual order extended by $-\infty \le t \le \infty$ for every $t \in \mathbb{R}$. [/definition] When infinite values are allowed, ordinary inverse images of Borel subsets can obscure the simple order tests used in integration. Real threshold sets still make sense and detect how the function sits below every finite level, while handling $\infty$ and $-\infty$ through the order rather than as ordinary [real numbers](/page/Real%20Numbers). [definition: Extended-Valued Measurable Function] Let $(X, \mathcal{M})$ be a measurable space. A function $f: X \to [-\infty,\infty]$ is measurable if $\{x \in X : f(x) \le t\} \in \mathcal{M}$ for every $t \in \mathbb{R}$. [/definition] This definition agrees with the usual Borel-measurable definition for the natural Borel structure on $[-\infty,\infty]$. However, the threshold condition only mentions finite real levels, so it is not immediate that the exceptional sets where the function equals $\infty$ or $-\infty$ are measurable. That gap matters whenever infinite values are used to encode constraints or singular behavior: one must still be able to isolate the set where the constraint is infinite or where the singularity occurs. The next result supplies this missing measurability of the infinite-value sets from the finite-threshold definition. [quotetheorem:8509] The extended-valued setting is not merely a technical enlargement. It is the natural language for monotone convergence, essential suprema, lower semicontinuous functions, and variational problems where infinite values encode constraints. [example: Encoding a Constraint by an Infinite Value] Let $X=\mathbb{R}$ with $\mathcal{B}(\mathbb{R})$, and define $I:\mathbb{R}\to[-\infty,\infty]$ by setting $I(x)=x^2$ for $x\ge 0$ and $I(x)=\infty$ for $x<0$. We compute the sublevel set $\{x\in\mathbb{R}:I(x)\le t\}$ for an arbitrary real number $t$. If $t<0$ and $x<0$, then $I(x)=\infty$, so $I(x)\le t$ is false. If $t<0$ and $x\ge 0$, then \begin{align*} I(x)=x^2\ge 0>t, \end{align*} so $I(x)\le t$ is again false. Hence \begin{align*} \{x\in\mathbb{R}:I(x)\le t\}=\varnothing \qquad \text{for } t<0. \end{align*} Now suppose $t\ge 0$. If $x<0$, then $I(x)=\infty$, so $I(x)\le t$ is false. If $x\ge 0$, then $I(x)=x^2$, and for nonnegative $x$ and $t$, \begin{align*} x^2\le t \quad \text{if and only if} \quad 0\le x\le \sqrt{t}. \end{align*} Therefore \begin{align*} \{x\in\mathbb{R}:I(x)\le t\}=[0,\sqrt{t}] \qquad \text{for } t\ge 0. \end{align*} The set $\varnothing$ and every closed interval $[0,\sqrt{t}]$ belong to $\mathcal{B}(\mathbb{R})$, so $\{x\in\mathbb{R}:I(x)\le t\}$ is Borel for every $t\in\mathbb{R}$. Thus $I$ is an extended-valued measurable function, and the value $\infty$ records that negative $x$ are forbidden rather than merely costly. [/example] ## Measurability and Integration The [Lebesgue integral](/page/Lebesgue%20Integral) is not defined by summing values point by point. It is built by measuring the sizes of sets on which a function takes certain values. The next definition is needed because indicator functions are the first functions whose integrals are determined directly by a measure. [definition: Indicator Function] Let $X$ be a set and let $A \subset X$. The indicator function of $A$ is the function $\mathbb{1}_A: X \to \mathbb{R}$ that takes the value $1$ on $A$ and the value $0$ on $X \setminus A$. [/definition] The indicator function translates set membership into a numerical function. The possible obstruction is that a numerical function may be measurable only when the set whose membership it records is itself measurable. Thus indicators are the bridge between measurable sets and the first measurable functions. [quotetheorem:8510] Indicators are too rigid to approximate general functions by themselves: each one only records membership in a single set. Integration needs a larger finite class that can distinguish several measurable regions, assign a different height to each one, and still remain controllable by the algebra of measurable sets. This finite class is meant to preserve the set-based nature of the Lebesgue integral. If a function takes only finitely many values, then its behavior can be understood by measuring the regions on which each value occurs, rather than by listing individual points. Requiring those regions to be measurable keeps the construction compatible with the measure and makes the eventual integral formula unambiguous. To build functions with several possible values while keeping every level tied to measurable sets, we combine finitely many indicators with real coefficients. The point of the definition is not just finite-valuedness, but finite-valuedness whose pieces can be measured. Writing the function as a finite sum of weighted indicators records exactly which measurable regions contribute which numerical heights, so the later integral can be reduced to a finite calculation involving measures of sets. [definition: Simple Function] Let $(X, \mathcal{M})$ be a measurable space. A function $s: X \to \mathbb{R}$ is a [simple function](/page/Simple%20Function) if there exist $n \in \mathbb{N}$, real numbers $a_1,\dots,a_n \in \mathbb{R}$, and measurable sets $A_1,\dots,A_n \in \mathcal{M}$ such that $s=\sum_{i=1}^{n} a_i\mathbb{1}_{A_i}$. [/definition] A simple function is measurable because it is built from measurable indicators using finite algebraic operations. The remaining problem is how to assign its integral in a way that depends on the sizes of the regions where its values occur, rather than on individual points. For nonnegative simple functions, this can be done by summing value times measure over measurable pieces. [definition: Integral of a Nonnegative Simple Function] Let $(X, \mathcal{M}, \mu)$ be a measure space, and let $\mathcal{S}_+(X,\mathcal{M})$ denote the set of nonnegative simple measurable functions $s: X \to [0,\infty)$. The integral of nonnegative simple functions is the map \begin{align*} \int_X \cdot \, d\mu: \mathcal{S}_+(X,\mathcal{M}) &\to [0,\infty]. \end{align*} If $s=\sum_{i=1}^{n} a_i\mathbb{1}_{A_i}$, where $a_i \ge 0$ and the sets $A_i \in \mathcal{M}$ are pairwise disjoint, then \begin{align*} \int_X s \, d\mu=\sum_{i=1}^{n} a_i\mu(A_i). \end{align*} [/definition] Different disjoint representations of the same simple function give the same value for the sum, after refining both representations by their common measurable atoms. This well-definedness is the point that lets the formula depend on $s$ itself rather than on a particular way of writing $s$. This definition reveals why measurability is not optional. The expression $\mu(A_i)$ is meaningful only when each $A_i$ belongs to $\mathcal{M}$. A non-measurable indicator would ask the measure to evaluate a set outside its domain. [example: Why the Non-Measurable Indicator Blocks the Integral] Let $\mathcal{L}^1$ denote [Lebesgue measure](/page/Lebesgue%20Measure) on $\mathbb{R}$, whose domain is the Lebesgue sigma-algebra. Let $V \subset [0,1]$ be non-Lebesgue-measurable and define $f=\mathbb{1}_V$. For every $x\in\mathbb{R}$, \begin{align*} f(x)=1 \text{ if } x\in V, \qquad f(x)=0 \text{ if } x\notin V. \end{align*} The value set on which $f$ equals $1$ is exactly $V$. Indeed, if $x\in V$, then $f(x)=1$, so $x\in\{y\in\mathbb{R}:f(y)=1\}$. If $x\notin V$, then $f(x)=0$, so $x\notin\{y\in\mathbb{R}:f(y)=1\}$. Hence \begin{align*} \{x\in\mathbb{R}:f(x)=1\}=V. \end{align*} The simple-function integral formula applies to representations of the form \begin{align*} s=\sum_{i=1}^{n} a_i\mathbb{1}_{A_i} \end{align*} with measurable sets $A_i$ in the domain of the measure. For $f=\mathbb{1}_V$, the corresponding one-term representation would be \begin{align*} f=1\cdot \mathbb{1}_V. \end{align*} Substituting this representation into the simple-function formula would require \begin{align*} \int_{\mathbb{R}} f\,d\mathcal{L}^1=1\cdot \mathcal{L}^1(V)=\mathcal{L}^1(V). \end{align*} But $\mathcal{L}^1(V)$ is not defined, because $V$ is not in the Lebesgue sigma-algebra. The obstruction is therefore not the size or oscillation of $f$; it is that the set on which $f$ takes the value $1$ lies outside the measurable structure on which Lebesgue measure is defined. [/example] ## Approximation by Simple Functions The defining strength of measurable functions is that they can be recovered from countably many simple measurable observations. This is the measure-theoretic counterpart of approximating continuous functions by polynomials or smooth functions in other parts of analysis. The approximation is especially clean for nonnegative functions: divide the vertical axis into dyadic intervals, record which interval contains the value, and let the mesh size shrink while the cutoff height grows. The next theorem is needed because it is the bridge from simple-function integration to the integral of a general nonnegative measurable function. [quotetheorem:1020] This theorem is the reason the Lebesgue integral can be defined for all nonnegative measurable functions after it is defined for simple functions. The integral is obtained by taking suprema over simple functions below $f$, or equivalently by limits along such increasing approximations. [example: Dyadic Approximation on a Finite Interval] Let $f:[0,1]\to[0,\infty)$ be given by $f(x)=x$, with $[0,1]$ equipped with $\mathcal{B}([0,1])$. For $n\in\mathbb{N}$ and $k\in\{0,\dots,2^n-1\}$, set \begin{align*} I_{n,k}=[k2^{-n},(k+1)2^{-n})\cap[0,1]. \end{align*} Define $s_n:[0,1]\to\mathbb{R}$ by $s_n(x)=k2^{-n}$ when $x\in I_{n,k}$, and set $s_n(1)=1$. Equivalently, \begin{align*} s_n=\sum_{k=0}^{2^n-1} k2^{-n}\mathbb{1}_{I_{n,k}}+\mathbb{1}_{\{1\}}. \end{align*} Each $I_{n,k}$ is Borel in $[0,1]$: it is the intersection of $[0,1]$ with a half-open interval $[k2^{-n},(k+1)2^{-n})$, and \begin{align*} [k2^{-n},(k+1)2^{-n})=[k2^{-n},\infty)\cap(-\infty,(k+1)2^{-n}). \end{align*} The singleton $\{1\}$ is closed in $[0,1]$, hence Borel. Therefore $s_n$ is a simple measurable function because it is a finite linear combination of indicators of Borel sets. Now fix $x\in[0,1)$. There is a unique $k\in\{0,\dots,2^n-1\}$ such that \begin{align*} k2^{-n}\le x<(k+1)2^{-n}. \end{align*} For this $k$, the definition gives \begin{align*} s_n(x)=k2^{-n}. \end{align*} The left inequality gives \begin{align*} 0\le s_n(x)\le x. \end{align*} The right inequality gives \begin{align*} x<(k+1)2^{-n}=k2^{-n}+2^{-n}=s_n(x)+2^{-n}. \end{align*} Subtracting $s_n(x)$ from both sides yields \begin{align*} 0\le x-s_n(x)<2^{-n}. \end{align*} Since $2^{-n}\to0$, the squeeze estimate implies $s_n(x)\to x$ for every $x\in[0,1)$. At the endpoint, the definition gives $s_n(1)=1=f(1)$ for every $n$. Thus the dyadic simple functions $s_n$ converge pointwise to $f(x)=x$ on all of $[0,1]$, and they do so from below with an error at most one dyadic mesh length away from the endpoint. [/example] The approximation theorem also clarifies the role of countability. A sigma-algebra is not required to be closed under arbitrary unions, but dyadic approximation only asks for countably many measurable pieces and countably many limiting operations. ## Convergence and Limits ### Pointwise Stability Analysis is built from limiting processes. If measurable functions were not stable under pointwise limits, they would be too fragile for integration, probability, PDE, or functional analysis. The first limit operation to isolate is passage to a pointwise limit. Here $\overline{\mathbb{R}}$ denotes the extended real line $[-\infty,\infty]$, so the theorem covers limits that may take infinite endpoint values as well as ordinary real-valued limits. [quotetheorem:1024] The obstruction is that a pointwise limit may be highly discontinuous even when every approximating function is well behaved, so measurability cannot be inherited from continuity alone. This theorem is the measure-theoretic reason that pointwise limiting arguments can be used without leaving the class of permissible integrands. [example: A Discontinuous Limit of Measurable Functions] For $n\in\mathbb{N}$, define $f_n:[0,1]\to\mathbb{R}$ by $f_n(x)=x^n$. Each $f_n$ is continuous on $[0,1]$, since it is the restriction of the polynomial function $x\mapsto x^n$ on $\mathbb{R}$, and hence each $f_n$ is Borel measurable by *Continuous Maps Are Borel Measurable*. We compute the pointwise limit. If $x=0$, then for every $n\in\mathbb{N}$, \begin{align*} f_n(0)=0^n=0. \end{align*} If $0<x<1$, let $a=1/x$, so $a>1$ and $x=1/a$. Given $\varepsilon>0$, choose $N\in\mathbb{N}$ such that $a^N>1/\varepsilon$. Then for every $n\ge N$, \begin{align*} 0\le x^n=a^{-n}\le a^{-N}<\varepsilon. \end{align*} Thus $x^n\to0$ for every $0<x<1$. At the endpoint $x=1$, for every $n\in\mathbb{N}$, \begin{align*} f_n(1)=1^n=1. \end{align*} Therefore the pointwise limit $f:[0,1]\to\mathbb{R}$ is \begin{align*} f(x)=0 \text{ for } 0\le x<1, \qquad f(1)=1. \end{align*} This is exactly the indicator function of the singleton $\{1\}$, because $\mathbb{1}_{\{1\}}(x)=0$ when $x\ne1$ and $\mathbb{1}_{\{1\}}(1)=1$. Since $f$ is the pointwise limit of the measurable functions $f_n$, it is measurable by *[Closure Under Pointwise Limits](/theorems/1024)*. The limit is not continuous at $1$: if $x_m=1-1/m$, then $x_m\to1$, but \begin{align*} f(x_m)=0 \text{ for every } m\ge2, \end{align*} while $f(1)=1$. Thus measurability survives the pointwise limit even though continuity is lost at the endpoint. [/example] ### Null-Set Modifications A common warning is that changing a measurable function on a non-measurable set may destroy measurability. Almost-everywhere arguments therefore require a sharper principle: changing values on a set of measure zero should not alter measurability, but this is only reliable when every subset of a null set is still measurable. This is the role of completeness in a measure space, and it is exactly the condition needed for null-set modifications to preserve measurable functions. [quotetheorem:8511] Completeness is essential here. Without it, a subset of a measurable null set may fail to belong to $\mathcal{M}$, and changing a function on that subset can destroy measurability. The theorem is often used silently in $L^p$ spaces, the spaces of $p$-integrable functions modulo equality almost everywhere. The measurable representative still matters: the equivalence class belongs to $L^p$ only because it contains measurable functions. ## Beyond and Connected Topics Measurable functions are the entrance to [Lebesgue integration](/page/Integral). The integral begins with indicators, passes through simple functions, and reaches nonnegative measurable functions by monotone approximation. From there, signed and complex-valued integrals are built by decomposing functions into positive and negative or real and imaginary parts. In probability, a random variable is a measurable function. This viewpoint makes distributions into pushforward measures: if $X: (\Omega, \mathcal{F}, \mathbb{P}) \to (E, \mathcal{E})$ is a random variable, its law is the measure $\mathbb{P} \circ X^{-1}$ on $(E, \mathcal{E})$. Expectations are then integrals of measurable functions with respect to probability measures. In functional analysis, measurable functions become elements of the function space $L^p(E, \mathcal{E}, \mu)$ for $1 \le p \le \infty$, consisting of measurable functions with finite $p$-integrability norm on the measure space $(E, \mathcal{E}, \mu)$ after quotienting by almost-everywhere equality. This is where the subject connects to [Banach spaces](/page/Banach%20Space), [Hilbert spaces](/page/Hilbert%20Space), [weak convergence](/page/Weak%20Convergence), and duality. In geometric measure theory, measurability is the background condition behind BV functions, sets of finite perimeter, and integration against [Hausdorff measure](/page/Hausdorff%20Measure). The point is not just to measure domains, but to measure level sets, traces, approximate limits, and distributional derivatives. Measurability also interacts with topology through Borel maps, with descriptive set theory through the structure of Borel and analytic sets, and with PDE through weak solutions represented by measurable or Sobolev functions. Each direction keeps the same basic idea: a function is admissible only when its value-level questions define measurable sets. ## References Androma, [Integral](/page/Integral). Androma, [Banach Space](/page/Banach%20Space). Androma, [Hilbert Space](/page/Hilbert%20Space). Androma, [Cambridge II Analysis of Functions](/page/Cambridge%20II%20Analysis%20of%20Functions). Androma, [Geometric Measure Theory III: BV Functions and Sets of Finite Perimeter](/page/Geometric%20Measure%20Theory%20III%3A%20BV%20Functions%20and%20Sets%20of%20Finite%20Perimeter). Androma, [Cambridge II Linear Analysis](/page/Cambridge%20II%20Linear%20Analysis). Androma, [Cambridge III Functional Analysis](/page/Cambridge%20III%20Functional%20Analysis). Folland, *Real Analysis: Modern Techniques and Their Applications* (1999). Royden and Fitzpatrick, *Real Analysis* (2010). Bogachev, *Measure Theory* (2007).