4.4 Lebesgue Measure
Lebesgue outer measure, Carathéodory's measurable sets, σ-algebras and Borel sets, null sets, the Cantor set and fat Cantor sets, Vitali's non-measurable set, measurable and simple functions, and the theorems of Egorov and Lusin.
To integrate in Lebesgue's way — slicing the range instead of the domain, as in the teaser of the chapter introduction — we must be able to say how large the set $\{x: f(x)\ge y\}$ is, and for a wildly behaved $f$ that set need not be a finite union of intervals. It may be the rationals, a Cantor-like dust, or a dense open set with tiny total length. So we want a measure $m$ that assigns a length to every reasonable subset of $\mathbb{R}$ and behaves like length: (i) $m(I)=\ell(I)$, the length of an interval $I$; (ii) $m(E+t)=m(E)$ (translation invariance); (iii) $m\big(\bigcup_kE_k\big)=\sum_km(E_k)$ for pairwise disjoint $E_1,E_2,\dots$ (countable additivity). In §4.1 we already met the sets that such a measure must call null. This section builds the measure itself: first an outer measure defined for every set, then the class of measurable sets on which it behaves perfectly, and finally the measurable functions that Lebesgue's integral (§4.5) will accept.
There is a price. Vitali showed in 1905 that no function satisfying (i)–(iii) can be defined on all subsets of $\mathbb{R}$, so some sets must remain unmeasured. In practice this never bites — every set you can write down explicitly is measurable — but it explains why the theory is organised as it is. CSIR NET examines this material through concrete computations (the measure of a set built from intervals, rationals and Cantor-type constructions) and, above all, through Part C statements about null, dense, nowhere dense, Borel and measurable sets, whose wrong options are the counterexamples collected below.
Lebesgue Outer Measure
Throughout, $\ell(I)$ denotes the length of an interval $I$ (so $\ell((a,b))=\ell([a,b])=b-a$, and $\ell(I)=\infty$ for an unbounded interval); in §4.1 we wrote $|J|$ for the same thing. A cover of a set $A$ is a finite or countable family of open intervals whose union contains $A$ (a finite family can be padded with empty intervals, so we may always index covers by $k\in\mathbb N$).
The definition makes sense for every subset of $\mathbb{R}$, and a set has measure zero in the sense of Definition 4.5 exactly when $m^*(A)=0$: both say that $A$ can be covered by intervals of arbitrarily small total length.
- (Monotonicity.) $m^*(\varnothing)=0$, and $A\subseteq B$ implies $m^*(A)\le m^*(B)$.
- (Countable subadditivity.) For any sets $A_1,A_2,\dots$: $\ m^*\big(\bigcup_kA_k\big)\le\sum_km^*(A_k)$.
- (Length of intervals.) $m^*(I)=\ell(I)$ for every interval $I$.
- (Invariance.) $m^*(A+t)=m^*(A)$ and $m^*(cA)=|c|\,m^*(A)$ for all $t,c\in\mathbb{R}$.
- (Null sets.) Every countable set has outer measure $0$; and if $m^*(N)=0$ then $m^*(A\cup N)=m^*(A)$ for every $A$.
- (Outer regularity.) $m^*(A)=\inf\{m^*(U):U\text{ open},\ U\supseteq A\}$.
Proof. (1) The empty set is covered by empty intervals; every cover of $B$ covers $A$, so the infimum for $A$ is over a larger family.
(2) We may assume every $m^*(A_k)$ is finite. Let $\varepsilon\gt0$ and choose for each $k$ a cover $(I_{k,j})_j$ of $A_k$ with $\sum_j\ell(I_{k,j})\lt m^*(A_k)+\varepsilon/2^k$. All the $I_{k,j}$ together form a countable cover of $\bigcup_kA_k$, so $m^*\big(\bigcup A_k\big)\le\sum_k\big(m^*(A_k)+\varepsilon/2^k\big)$, which is $\sum_km^*(A_k)+\varepsilon$. Let $\varepsilon\to0$.
(3) First let $I=[a,b]$. The single interval $(a-\varepsilon,b+\varepsilon)$ covers it, so $m^*([a,b])\le b-a+2\varepsilon$ for every $\varepsilon$, i.e. $m^*([a,b])\le b-a$. For the reverse inequality take any cover $(I_k)$ of $[a,b]$. By the Heine–Borel theorem (§3.1) finitely many of the $I_k$ already cover $[a,b]$, and it suffices to show that their lengths add up to more than $b-a$. Pick one of them, $J_1=(a_1,b_1)$, containing $a$. If $b_1\le b$, then $b_1\in[a,b]$ but $b_1\notin J_1$, so another interval $J_2=(a_2,b_2)$ of the finite family contains $b_1$: $a_2\lt b_1\lt b_2$. Continue in this way; since the family is finite and each step uses a new interval, we stop at some $J_r$ with $b_r\gt b$. Then, using $a_{i+1}\lt b_i$, $$\sum_{i=1}^r\ell(J_i)=b_r-a_1+\sum_{i=1}^{r-1}(b_i-a_{i+1})\gt b_r-a_1\gt b-a.$$ Hence $m^*([a,b])=b-a$. A bounded interval $I$ with end points $a\lt b$ satisfies $[a+\varepsilon,b-\varepsilon]\subseteq I\subseteq[a,b]$, so by monotonicity $b-a-2\varepsilon\le m^*(I)\le b-a$. An unbounded interval contains closed intervals of every length, so its outer measure is $\infty$.
(4) $(I_k)$ covers $A$ if and only if $(I_k+t)$ covers $A+t$, and the lengths are the same; for $c\ne0$, $(cI_k)$ covers $cA$ with lengths multiplied by $|c|$ ($c=0$ is trivial).
(5) Cover the $k$-th point of a countable set by an interval of length $\varepsilon/2^k$. If $m^*(N)=0$, then $m^*(A)\le m^*(A\cup N)\le m^*(A)+m^*(N)=m^*(A)$ by (1) and (2).
(6) "$\le$" is monotonicity. If $m^*(A)\lt\infty$ and $\varepsilon\gt0$, take a cover with $\sum\ell(I_k)\lt m^*(A)+\varepsilon$; then $U=\bigcup I_k$ is open, contains $A$, and $m^*(U)\le\sum\ell(I_k)\lt m^*(A)+\varepsilon$ by (2) and (3). ∎
Part (3) is less innocent than it looks: it is the statement that $[a,b]$ cannot be covered by countably many intervals of smaller total length, and its proof uses the completeness of $\mathbb{R}$ (through Heine–Borel). Over $\mathbb Q$ it fails — $\mathbb Q\cap[0,1]$ can be covered by intervals of total length $\varepsilon$. What outer measure does not have is additivity: we shall see (Theorem 4.40) disjoint sets $A,B$ with $m^*(A\cup B)\lt m^*(A)+m^*(B)$.
Measurable Sets, σ-Algebras and Borel Sets
Carathéodory's idea is to keep only those sets $E$ that split every set additively. Such a set is "well-behaved enough" that cutting along it never loses or creates outer measure.
A set $E\subseteq\mathbb{R}$ is Lebesgue measurable if $$m^*(A)=m^*(A\cap E)+m^*(A\cap E^c)\qquad\text{for every }A\subseteq\mathbb{R}.$$ The family of measurable sets is denoted $\mathcal M$, and for $E\in\mathcal M$ we write $m(E)=m^*(E)$, the Lebesgue measure of $E$.
Since $A=(A\cap E)\cup(A\cap E^c)$, subadditivity always gives "$\le$". So $E$ is measurable as soon as $m^*(A)\ge m^*(A\cap E)+m^*(A\cap E^c)$ for every test set $A$ with $m^*(A)\lt\infty$.
A family $\Sigma$ of subsets of $\mathbb{R}$ is a σ-algebra if $\varnothing\in\Sigma$, $E\in\Sigma\Rightarrow E^c\in\Sigma$, and $E_1,E_2,\dots\in\Sigma\Rightarrow\bigcup_kE_k\in\Sigma$. (Then $\Sigma$ is also closed under countable intersections and differences.) The intersection of all σ-algebras containing the open sets is the Borel σ-algebra $\mathcal B(\mathbb{R})$; its members are the Borel sets.
A $G_\delta$ set is a countable intersection of open sets; an $F_\sigma$ set is a countable union of closed sets. Both are Borel. Examples: every closed set is $G_\delta$ ($F=\bigcap_n\{x:\operatorname{dist}(x,F)\lt\frac1n\}$); $\mathbb{Q}$ is $F_\sigma$ (a countable union of points); the irrationals form a $G_\delta$ set ($\bigcap_{q\in\mathbb Q}(\mathbb R\setminus\{q\})$).
- $\mathcal M$ is a σ-algebra.
- $\mathcal M$ contains every set of outer measure $0$ and every interval; hence every Borel set is measurable: $\mathcal B(\mathbb{R})\subseteq\mathcal M$.
- $m$ is countably additive on $\mathcal M$: if $E_1,E_2,\dots\in\mathcal M$ are pairwise disjoint, then $m\big(\bigcup_kE_k\big)=\sum_km(E_k)$.
Proof. Null sets and complements. If $m^*(E)=0$ then for any $A$, $m^*(A\cap E)\le m^*(E)=0$ and $m^*(A\cap E^c)\le m^*(A)$, so the criterion holds. The criterion is symmetric in $E$ and $E^c$, so $E\in\mathcal M\Rightarrow E^c\in\mathcal M$; and $\varnothing\in\mathcal M$.
Finite unions. Let $E_1,E_2\in\mathcal M$ and $A\subseteq\mathbb{R}$. Apply the criterion for $E_1$ to $A$, and then for $E_2$ to $A\cap E_1^c$: $$\begin{aligned}m^*(A)&=m^*(A\cap E_1)+m^*(A\cap E_1^c\cap E_2)\\&\quad+m^*(A\cap E_1^c\cap E_2^c).\end{aligned}$$ The first two sets have union $A\cap(E_1\cup E_2)$, and the third is $A\cap(E_1\cup E_2)^c$; by subadditivity, $m^*(A)\ge m^*(A\cap(E_1\cup E_2))$ $+\,m^*(A\cap(E_1\cup E_2)^c)$. So $E_1\cup E_2\in\mathcal M$, and by induction and complements $\mathcal M$ is closed under finite unions, intersections and differences.
Finite additivity on slices. If $E_1,\dots,E_n\in\mathcal M$ are disjoint and $F_n=E_1\cup\dots\cup E_n$, then for every $A$ $$m^*(A\cap F_n)=\sum_{k=1}^nm^*(A\cap E_k).$$ Indeed, testing $E_n$ on the set $A\cap F_n$ gives $m^*(A\cap F_n)=m^*(A\cap E_n)+m^*(A\cap F_{n-1})$; induct.
Countable unions. Let $E_1,E_2,\dots\in\mathcal M$ be disjoint, $E=\bigcup_kE_k$. Since $F_n\in\mathcal M$ and $F_n^c\supseteq E^c$, $$\begin{aligned}m^*(A)&=m^*(A\cap F_n)+m^*(A\cap F_n^c)\\&\ge\sum_{k=1}^nm^*(A\cap E_k)+m^*(A\cap E^c).\end{aligned}$$ Let $n\to\infty$ and use subadditivity: $m^*(A)\ge\sum_{k=1}^\infty m^*(A\cap E_k)+m^*(A\cap E^c)$, which is at least $m^*(A\cap E)+m^*(A\cap E^c)$. So $E\in\mathcal M$; and with $A=E$ all these inequalities become equalities, giving $m(E)=\sum_km(E_k)$, which is (3). A general sequence in $\mathcal M$ is first made disjoint, $E_k'=E_k\setminus(E_1\cup\dots\cup E_{k-1})$, without changing the union. This proves (1).
Intervals. It is enough to show $(a,\infty)\in\mathcal M$: then $(-\infty,b]$ (a complement), $(a,b]=(a,\infty)\cap(-\infty,b]$ and $(a,b)=\bigcup_n(a,b-\frac1n]$ are measurable; every open set is a countable union of open intervals (with rational end points), so every open set is in $\mathcal M$, and since $\mathcal M$ is a σ-algebra containing the open sets, it contains $\mathcal B(\mathbb R)$. Let $m^*(A)\lt\infty$, $\varepsilon\gt0$, and take a cover $(I_k)$ of $A$ with $\sum\ell(I_k)\lt m^*(A)+\varepsilon$. The sets $I_k'=I_k\cap(a,\infty)$ and $I_k''=I_k\cap(-\infty,a]$ are intervals (possibly empty) with $\ell(I_k')+\ell(I_k'')=\ell(I_k)$; they cover $A\cap(a,\infty)$ and $A\cap(-\infty,a]$ respectively. By subadditivity and Theorem 4.34(3), $$\begin{aligned}m^*(A\cap(a,\infty))+m^*(A\cap(-\infty,a])&\le\sum_k\ell(I_k')+\sum_k\ell(I_k'')\\&=\sum_k\ell(I_k)\lt m^*(A)+\varepsilon.\end{aligned}$$ Let $\varepsilon\to0$. ∎
- (From below.) If $E_1\subseteq E_2\subseteq\cdots$, then $m\big(\bigcup_nE_n\big)=\lim_{n\to\infty}m(E_n)$.
- (From above.) If $E_1\supseteq E_2\supseteq\cdots$ and $m(E_1)\lt\infty$, then $m\big(\bigcap_nE_n\big)=\lim_{n\to\infty}m(E_n)$.
Proof. (1) The "rings" $D_1=E_1$, $D_k=E_k\setminus E_{k-1}$ are disjoint and measurable, $E_n=D_1\cup\dots\cup D_n$ and $\bigcup_nE_n=\bigcup_kD_k$. By countable additivity, $m\big(\bigcup E_n\big)=\sum_km(D_k)$, and this is $\lim_n\sum_{k\le n}m(D_k)=\lim_nm(E_n)$. (2) The sets $E_1\setminus E_n$ increase to $E_1\setminus\bigcap_nE_n$, so by (1), $m(E_1)-m(E_n)\to m(E_1)-m\big(\bigcap E_n\big)$; the subtractions are legitimate because all measures involved are at most $m(E_1)\lt\infty$. ∎
The finiteness hypothesis in (2) is essential: $E_n=[n,\infty)$ decrease to $\varnothing$, yet $m(E_n)=\infty$ for all $n$.
- $E$ is measurable;
- for every $\varepsilon\gt0$ there is an open $U\supseteq E$ with $m^*(U\setminus E)\lt\varepsilon$;
- for every $\varepsilon\gt0$ there is a closed $F\subseteq E$ with $m^*(E\setminus F)\lt\varepsilon$;
- $E=G\setminus N$ with $G$ a $G_\delta$ set and $m^*(N)=0$;
- $E=H\cup N$ with $H$ an $F_\sigma$ set and $m^*(N)=0$.
Proof. (1)⇒(2). If $m(E)\lt\infty$, Theorem 4.34(6) gives an open $U\supseteq E$ with $m(U)\lt m(E)+\varepsilon$, and since $E$ is measurable, $m(U\setminus E)=m(U)-m(E)\lt\varepsilon$. In general apply this to $E_n=E\cap[-n,n]$ with $\varepsilon/2^n$, obtaining $U_n$; then $U=\bigcup U_n$ is open, contains $E$, and $U\setminus E\subseteq\bigcup_n(U_n\setminus E_n)$ has outer measure $\lt\varepsilon$.
(2)⇒(4). Choose open $U_n\supseteq E$ with $m^*(U_n\setminus E)\lt\frac1n$, and let $G=\bigcap_nU_n$, a $G_\delta$ set containing $E$. Then $N=G\setminus E\subseteq U_n\setminus E$ for every $n$, so $m^*(N)=0$, and $E=G\setminus N$.
(4)⇒(1) and (5)⇒(1). Borel sets and null sets are measurable (Theorem 4.35), and $\mathcal M$ is closed under differences and unions.
(1)⇒(3)⇒(5). Apply (2) to the measurable set $E^c$: there is an open $U\supseteq E^c$ with $m(U\setminus E^c)\lt\varepsilon$. Then $F=U^c$ is closed, $F\subseteq E$, and $E\setminus F=E\cap U=U\setminus E^c$. Taking closed $F_n\subseteq E$ with $m^*(E\setminus F_n)\lt\frac1n$, $H=\bigcup F_n$ is $F_\sigma$ and $N=E\setminus H$ is null. Finally, the sets $F\cap[-n,n]$ are compact and increase to $F$, so continuity from below gives the inner approximation by compact sets. ∎
So a measurable set is a Borel set "up to a null set" — but not every measurable set is Borel. One can show that $\mathcal B(\mathbb R)$ has cardinality $\mathfrak c$. On the other hand, every subset of the Cantor set $C$ (below) is null, hence measurable, and $C$ has cardinality $\mathfrak c$, so $\mathcal M$ has at least $2^{\mathfrak c}\gt\mathfrak c$ members (Cantor's theorem, §1.1). Hence there are Lebesgue measurable sets that are not Borel sets.
Null Sets, the Cantor Set and Fat Cantor Sets
Countable sets are null, but null sets can be much larger. The classical example is built by repeatedly deleting middle thirds.
- $C$ is compact and non-empty, and $m(C)=0$.
- $C$ is exactly the set of $x\in[0,1]$ having a ternary expansion $x=\sum_{k\ge1}a_k3^{-k}$ with every $a_k\in\{0,2\}$.
- $C$ is uncountable; in fact $|C|=\mathfrak c$.
- $C$ contains no interval of positive length; it is nowhere dense.
- $C$ is perfect: every point of $C$ is a limit point of $C$.
Proof. (1) Each $C_n$ is closed and bounded, so $C$ is compact; the end points $0,1,\frac13,\frac23,\dots$ of the intervals of the $C_n$ are never removed, so $C\ne\varnothing$. As $C\subseteq C_n$, $m(C)\le m(C_n)=2^n3^{-n}=(\frac23)^n\to0$.
(2) By induction on $n$, the intervals of $C_n$ are $[\,0.a_1\dots a_n,\ 0.a_1\dots a_n+3^{-n}]$ (in base $3$) with $a_1,\dots,a_n\in\{0,2\}$; so $x\in C_n$ if and only if $x$ has a ternary expansion whose first $n$ digits lie in $\{0,2\}$ (removing the open middle third of such an interval removes exactly the points whose $(n+1)$-st digit is forced to be $1$). A number has at most two ternary expansions. If $x\in C$, then for every $n$ one of them has its first $n$ digits in $\{0,2\}$, so one of the two does so for infinitely many $n$, hence for all $n$. The converse is clear.
(3) For $y\in[0,1)$ take its binary expansion $y=\sum b_k2^{-k}$ not ending in an infinite run of $1$s, and map it to $\sum_k(2b_k)3^{-k}\in C$. Distinct digit sequences in $\{0,2\}$ give distinct numbers (two ternary expansions of the same number differ by a block $\dots c\,000\dots$ against $\dots(c-1)\,222\dots$, and one of these contains the digit $1$), so this is an injection $[0,1)\to C$. Hence $\mathfrak c\le|C|\le|\mathbb{R}|=\mathfrak c$ (Schröder–Bernstein, §1.1).
(4) An interval contained in $C$ is a connected subset of $C_n$, hence lies inside one of its intervals, so its length is at most $3^{-n}$ for every $n$. A closed set with empty interior is nowhere dense.
(5) If $x\in C$ and $n\ge1$, $x$ lies in an interval $J$ of $C_n$ of length $3^{-n}$; both end points of $J$ belong to $C$ and at least one differs from $x$. So points of $C\setminus\{x\}$ come within $3^{-n}$ of $x$. ∎
Two further facts are worth knowing. First, $C+C=\{x+y:x,y\in C\}=[0,2]$ (Exercise II.5): the sum of two null sets can be an interval. Second, the Cantor function $c:[0,1]\to[0,1]$ sends $\sum a_k3^{-k}\in C$ (digits $a_k\in\{0,2\}$) to $\sum\frac{a_k}2\,2^{-k}$ and is constant on each removed interval. It is increasing, continuous and onto $[0,1]$, and $c'=0$ on the open set $[0,1]\setminus C$ of full measure — yet $c(1)-c(0)=1$. It is monotone, so of bounded variation (§4.3), but it is not the integral of its derivative; this "devil's staircase" will reappear in §4.5.
Proof. The construction works. Let $L_n$ be the common length of the $2^n$ intervals of $K_n$, so $L_0=1$ and $L_n=\frac12(L_{n-1}-\alpha3^{-n})$. By induction $L_n\ge3^{-n}$: if $L_{n-1}\ge3^{-(n-1)}$ then $L_{n-1}-\alpha3^{-n}\ge3\cdot3^{-n}-3^{-n}=2\cdot3^{-n}$, so the middle interval fits and $L_n\ge3^{-n}$. Also $L_n\le\frac12L_{n-1}$, so $L_n\le2^{-n}\to0$.
Measure. The removed open intervals are pairwise disjoint, with total length $\sum_{n\ge1}2^{n-1}\alpha3^{-n}=\frac\alpha2\sum_{n\ge1}(\tfrac23)^n=\frac\alpha2\cdot2=\alpha$. Their union $U$ is open and $K_\alpha=[0,1]\setminus U$, so $m(K_\alpha)=1-\alpha$.
Topology. $K_\alpha$ is an intersection of closed sets, hence compact. An interval inside $K_\alpha$ lies in one interval of each $K_n$, so has length $\le L_n\to0$: $K_\alpha$ has empty interior. End points of intervals of $K_n$ are never removed, which gives perfectness exactly as for $C$.
Riemann integrability. At $x\in K_\alpha$, $\mathbf 1_{K_\alpha}(x)=1$, but every neighbourhood of $x$ meets $U$ (as $K_\alpha$ has empty interior), where the function is $0$: so $\mathbf 1_{K_\alpha}$ is discontinuous at every point of $K_\alpha$ (and continuous on the open set $U$). Its set of discontinuities is $K_\alpha$, of measure $1-\alpha\gt0$, and Lebesgue's criterion (Theorem 4.5) says it is not Riemann integrable. ∎
In Lebesgue's theory, $\int_0^1\mathbf 1_{K_\alpha}$ will simply be $m(K_\alpha)=1-\alpha$ (§4.5). Interactive 1 lets you watch both constructions; note that the number of intervals, $2^n$, is the same for every $\alpha$ — only their lengths differ.
A Non-Measurable Set
Proof. Call $x,y\in[0,1]$ equivalent if $x-y\in\mathbb Q$; this is an equivalence relation (§1.1), and each class is countable (of the form $(x+\mathbb Q)\cap[0,1]$), so there are uncountably many classes. By the axiom of choice there is a set $V\subseteq[0,1]$ containing exactly one point of each class. Let $r_1,r_2,\dots$ enumerate $\mathbb Q\cap[-1,1]$ and put $V_k=V+r_k$.
The $V_k$ are pairwise disjoint. If $x\in V_j\cap V_k$, then $x=v+r_j=v'+r_k$ with $v,v'\in V$, so $v-v'=r_k-r_j\in\mathbb Q$: $v$ and $v'$ are equivalent, hence equal (one point per class), and then $r_j=r_k$, $j=k$.
Their union $S$ satisfies $[0,1]\subseteq S\subseteq[-1,2]$. The second inclusion is clear. If $x\in[0,1]$, let $v\in V$ be the representative of its class; then $x-v\in\mathbb Q\cap[-1,1]$, say $x-v=r_k$, so $x\in V_k$.
Contradiction. If $V$ were measurable, each $V_k$ would be measurable with $m(V_k)=m(V)$ (translation invariance), and countable additivity would give $m(S)=\sum_km(V)$, which is $0$ if $m(V)=0$ and $\infty$ if $m(V)\gt0$. But $1\le m(S)\le3$. So $V\notin\mathcal M$. The same argument, using only countable additivity, translation invariance and $\mu([0,1])=1$, rules out any such $\mu$ on all subsets. Finally, since $V$ fails Carathéodory's criterion, some test set $A$ has $m^*(A)\lt m^*(A\cap V)+m^*(A\cap V^c)$; the sets $A\cap V$ and $A\setminus V$ are disjoint with union $A$. ∎
Some remarks for orientation. (1) The argument adapts to show that every set of positive outer measure contains a non-measurable subset. (2) Choice cannot be avoided: Solovay (1970) built a model of set theory without full choice in which every subset of $\mathbb R$ is measurable. (3) A non-measurable set can never be written down explicitly; in examinations, "measurable" is the default for any set built from intervals by countable operations, and for any subset of a null set. (4) The Banach–Tarski paradox shows the analogous problem in $\mathbb{R}^3$ is even more dramatic for finitely additive "volumes" invariant under rotations.
Measurable Functions, Simple Functions, Egorov and Lusin
We allow functions with values in the extended reals $[-\infty,\infty]$ (useful for limits and suprema), with the obvious order.
- $f$ is measurable iff $\{f\ge a\}\in\mathcal M$ for all $a$, iff $\{f\lt a\}\in\mathcal M$ for all $a$, iff $\{f\le a\}\in\mathcal M$ for all $a$. If $f$ is measurable and real-valued, $f^{-1}(B)\in\mathcal M$ for every Borel set $B$.
- Continuous functions on $E$, monotone functions on intervals, and indicators $\mathbf 1_A$ of measurable sets $A$ are measurable. ($\mathbf 1_A$ is measurable iff $A\in\mathcal M$.)
- If $f,g$ are measurable and real-valued and $c\in\mathbb{R}$, then $f+g$, $cf$, $fg$, $\max(f,g)$, $\min(f,g)$, $|f|$, $f^+=\max(f,0)$, $f^-=\max(-f,0)$ are measurable, and so is $\varphi\circ f$ for every continuous $\varphi:\mathbb{R}\to\mathbb{R}$.
- If $f_1,f_2,\dots$ are measurable, so are $\sup_nf_n$, $\inf_nf_n$, $\limsup_nf_n$, $\liminf_nf_n$; in particular a pointwise limit of measurable functions is measurable.
- If $f$ is measurable and $g=f$ a.e., then $g$ is measurable.
Proof. (1) $\{f\ge a\}=\bigcap_n\{f\gt a-\frac1n\}$ and $\{f\gt a\}=\bigcup_n\{f\ge a+\frac1n\}$; the other two sets are complements. For real-valued $f$, the family $\{B\subseteq\mathbb R: f^{-1}(B)\in\mathcal M\}$ is a σ-algebra (preimages commute with complements and unions) containing every $(a,\infty)$ and hence every $(a,b)=(a,\infty)\setminus[b,\infty)$ and every open set; so it contains $\mathcal B(\mathbb R)$.
(2) If $f$ is continuous on $E$, $\{f\gt a\}=E\cap U$ for some open $U$. If $f$ is increasing on an interval $J$, $\{f\gt a\}$ is an interval (if it contains $x$, it contains every $y\in J$ with $y\gt x$); similarly for decreasing $f$. $\{\mathbf 1_A\gt a\}$ is $\varnothing$, $A$ or $\mathbb{R}$.
(3) $\{f+g\gt a\}=\bigcup_{q\in\mathbb Q}\big(\{f\gt q\}\cap\{g\gt a-q\}\big)$: if $f(x)+g(x)\gt a$, choose a rational $q$ with $f(x)\gt q\gt a-g(x)$. For $\varphi$ continuous, $\{\varphi\circ f\gt a\}=f^{-1}(\varphi^{-1}(a,\infty))$, the preimage of an open set; this covers $cf$, $|f|$, $f^\pm$ and $f^2$. Then $fg=\frac14\big((f+g)^2-(f-g)^2\big)$, $\max(f,g)=\frac12(f+g+|f-g|)$, $\min(f,g)=\frac12(f+g-|f-g|)$.
(4) $\{\sup_nf_n\gt a\}=\bigcup_n\{f_n\gt a\}$; $\inf_nf_n=-\sup_n(-f_n)$; $\limsup_nf_n=\inf_n\sup_{k\ge n}f_k$ and $\liminf_nf_n=\sup_n\inf_{k\ge n}f_k$. Where the limit exists it equals $\limsup$.
(5) Let $N\supseteq\{f\ne g\}$ be null. Then $\{g\gt a\}=\big(\{f\gt a\}\setminus N\big)\cup\big(\{g\gt a\}\cap N\big)$; the first set is measurable and the second is a subset of a null set, hence null, hence measurable (Theorem 4.35(2)). ∎
Part (5) uses the completeness of Lebesgue measure (subsets of null sets are measurable). This is why statements such as "$f=g$ a.e." are harmless in Lebesgue theory, and why we may ignore the values of a function on any null set.
Proof. $\varphi_n$ takes the values $k/2^n$, $0\le k\le n2^n$, on the measurable sets $\{k/2^n\le f\lt(k+1)/2^n\}$ (for $k\lt n2^n$) and $\{f\ge n\}$ (for $k=n2^n$), so it is simple. Since $2\lfloor t\rfloor\le\lfloor2t\rfloor$, we have $\frac{\lfloor2^nt\rfloor}{2^n}\le\frac{\lfloor2^{n+1}t\rfloor}{2^{n+1}}\le t$, and $\min(n,\cdot)\le\min(n+1,\cdot)$; so $\varphi_n\le\varphi_{n+1}\le f$. If $f(x)\lt\infty$, then for $n\gt f(x)$ we get $0\le f(x)-\varphi_n(x)\lt2^{-n}$; if $f(x)=\infty$, $\varphi_n(x)=n\to\infty$. If $f\le M$ on a set $A$, the bound $0\le f-\varphi_n\lt2^{-n}$ holds on all of $A$ as soon as $n\gt M$: uniform convergence. ∎
The $\varphi_n$ are built by slicing the range of $f$ into levels of height $2^{-n}$ — exactly Lebesgue's idea. They are the bridge to the integral: §4.5 defines $\int\varphi_n=\sum_k\frac k{2^n}\,m\big(\{\varphi_n=\frac k{2^n}\}\big)$ and $\int f=\lim\int\varphi_n$.
Pointwise convergence of measurable functions is much weaker than uniform convergence — but on a set of finite measure it is uniform if we are allowed to discard a set of arbitrarily small measure. This is Littlewood's "third principle"; the first two say that a measurable set is nearly a finite union of intervals (Theorem 4.37 and Exercise I.9) and a measurable function is nearly continuous (Lusin's theorem below).
Proof. Remove from $E$ the null set where $f_n(x)\not\to f(x)$; this changes no measure, so assume $f_n\to f$ everywhere on $E$. For $k,n\in\mathbb N$ let $$E_n^k=\bigcup_{j\ge n}\{x\in E:|f_j(x)-f(x)|\ge\tfrac1k\}.$$ These sets are measurable and decrease in $n$, and $\bigcap_nE_n^k=\varnothing$: a point in every $E_n^k$ would have $|f_j(x)-f(x)|\ge\frac1k$ for infinitely many $j$, contradicting convergence. Since $m(E)\lt\infty$, continuity from above (Theorem 4.36(2)) gives $m(E_n^k)\to0$ as $n\to\infty$; choose $n_k$ with $m(E_{n_k}^k)\lt\varepsilon/2^{k+1}$. Let $A=\bigcup_kE_{n_k}^k$, so $m(A)\lt\varepsilon/2$, and put $F_0=E\setminus A$. For $x\in F_0$ and every $k$, $|f_j(x)-f(x)|\lt\frac1k$ for all $j\ge n_k$: that is uniform convergence on $F_0$. Finally, by Theorem 4.37(3) choose a closed $F\subseteq F_0$ with $m(F_0\setminus F)\lt\varepsilon/2$; uniform convergence persists on the smaller set and $m(E\setminus F)\lt\varepsilon$. ∎
Proof (careful sketch). Simple functions. If $\varphi=\sum_{i=1}^ra_i\mathbf 1_{E_i}$ in canonical form, choose closed $F_i\subseteq E_i$ with $m(E_i\setminus F_i)\lt\varepsilon/r$ (Theorem 4.37). Then $F=F_1\cup\dots\cup F_r$ is closed, $m([a,b]\setminus F)\lt\varepsilon$, and $\varphi$ is constant on each $F_i$. The $F_i$ are disjoint compact sets, so they are at positive distance from one another; hence each point of $F_i$ has a neighbourhood meeting $F$ only in $F_i$, and $\varphi|_F$ is continuous.
General $f$. By Theorem 4.42 there are simple $\varphi_n\to f$ pointwise. Choose closed $F_n$ with $m([a,b]\setminus F_n)\lt\varepsilon/2^{n+1}$ on which $\varphi_n$ is continuous, and by Egorov a closed $F_0$ with $m([a,b]\setminus F_0)\lt\varepsilon/2$ on which $\varphi_n\to f$ uniformly. Then $F=\bigcap_{n\ge0}F_n$ is closed, $m([a,b]\setminus F)\lt\varepsilon$, and $f|_F$ is a uniform limit of the continuous functions $\varphi_n|_F$, hence continuous (the $\frac\varepsilon3$ argument of §5.1).
Extension. $[a,b]\setminus F$ is relatively open, a countable union of disjoint intervals; on each bounded gap $(c,d)$ with $c,d\in F$ define $g$ linearly between $f(c)$ and $f(d)$ (and constant on a gap at an end of $[a,b]$). One checks that $g$ is continuous on $[a,b]$ — this is the one-dimensional case of the Tietze extension theorem. ∎
Read Lusin's theorem carefully: it says that the restriction $f|_F$ is continuous, not that $f$ is continuous at the points of $F$. For $f=\mathbf 1_{\mathbb Q}$, take $F$ a closed set of irrationals with $m([0,1]\setminus F)\lt\varepsilon$ (Example 2(c)); $f|_F\equiv0$ is continuous, while $f$ itself is continuous nowhere.
- Null does not mean countable. The Cantor set is null with cardinality $\mathfrak c$; so is the set of numbers in $[0,1]$ with a decimal expansion avoiding the digit $7$. (Countable does imply null.)
- Empty interior does not mean null. The irrationals in $[0,1]$ have measure $1$; a fat Cantor set is closed, nowhere dense and perfect, with measure $1-\alpha$.
- Dense does not mean large. $\mathbb Q$ is dense and null; a dense open subset of $[0,1]$ can have measure less than any given $\varepsilon\gt0$. But a set of full measure in $[0,1]$ is dense.
- Outer measure is only subadditive. On non-measurable sets it fails to be additive (Theorem 4.40). Additivity is guaranteed for measurable sets, and for sets at positive distance.
- Continuity from above needs a finite measure. $[n,\infty)\downarrow\varnothing$ with $m([n,\infty))=\infty$. Continuity from below needs nothing.
- Measurable is not the same as Borel. There are $2^{\mathfrak c}$ measurable sets (all subsets of $C$) but only $\mathfrak c$ Borel sets. Every measurable set differs from a Borel ($G_\delta$ or $F_\sigma$) set by a null set.
- Continuous images of measurable sets need not be measurable. $\psi(x)=x+c(x)$ ($c$ the Cantor function) is a homeomorphism of $[0,1]$ onto $[0,2]$ that maps $C$ onto a set of measure $1$; a non-measurable subset $W$ of $\psi(C)$ has preimage $\psi^{-1}(W)\subseteq C$, which is null and so measurable. Hence also: $\mathbf 1_{\psi^{-1}(W)}\circ\psi^{-1}$ is a measurable function composed with a continuous one, and it is not measurable.
- Positive measure does not give an interval — only the Steinhaus property that $E-E$ contains an interval around $0$.
- $|f|$ measurable, or all level sets measurable, does not make $f$ measurable (Example 4(c), (d)). But $f$ measurable makes $|f|$, $f^\pm$, $\varphi\circ f$ ($\varphi$ continuous) measurable.
- Egorov needs $m(E)\lt\infty$, and the bad set cannot be null. $\mathbf 1_{[n,n+1]}\to0$ on $\mathbb{R}$ is uniform off no set of finite measure; $x^n$ is not uniform on $[0,1)$, a set of full measure.
- Lusin is about the restriction. $f|_F$ continuous does not mean $f$ is continuous at the points of $F$: $\mathbf 1_{\mathbb Q}$ restricted to a closed set of irrationals is $0$.
Exercises
- 1 Using only the definition, show $m^*\big(\{\frac1n:n\in\mathbb N\}\big)=0$.
- 2 Find $m\big((0,3)\setminus\mathbb Q\big)$ and $m\big([0,1]\cup(\mathbb{Q}\cap[2,5])\cup(\sqrt2+\mathbb Z)\big)$.
- 3 Find $m\big(\bigcup_{n\ge1}[n,n+\frac1{n^2}]\big)$.
- 4 Show that if $E\in\mathcal M$ and $m(E\,\triangle\,F)=0$, then $F\in\mathcal M$ and $m(F)=m(E)$.
- 5 Let $E\subseteq[0,1]$ be measurable with $m(E)=1$. Prove that $E$ is dense in $[0,1]$. Is the converse true?
- 6 Prove that $m(E_1\cup E_2)+m(E_1\cap E_2)=m(E_1)+m(E_2)$ for $E_1,E_2\in\mathcal M$.
- 7 (Borel–Cantelli.) Let $E_n\in\mathcal M$ with $\sum_nm(E_n)\lt\infty$. Prove that the set of points lying in infinitely many $E_n$ is null.
- 8 Let $E\subseteq[0,1]$ be measurable. Show that $t\mapsto m(E\cap[0,t])$ is continuous (indeed $1$-Lipschitz), and deduce that for each $c\in[0,m(E)]$, $E$ has a measurable subset of measure $c$.
- 9 Prove that if $m(E)\lt\infty$ and $\varepsilon\gt0$, there is a finite union $W$ of open intervals with $m(E\,\triangle\,W)\lt\varepsilon$. Prove also that $m^*(A\cup B)=m^*(A)+m^*(B)$ whenever $\operatorname{dist}(A,B)\gt0$.
- 1 Show that $\frac34\in C$ and $\frac1{10}\in C$, and that $\frac12\notin C$.
- 2 Compute the total length of the open intervals removed in constructing $C$, and the sum of the lengths removed at stages $1$–$4$.
- 3 Let $T$ be the set of $x\in[0,1]$ having a base-$5$ expansion using only the digits $0,2,4$. Show that $T$ is compact, uncountable and null.
- 4 Choose $\alpha$ so that the fat Cantor set $K_\alpha$ of Definition 4.19 has measure $0.9$, and find the length $L_2$ of each interval of $K_2$ for that $\alpha$.
- 5 Prove that $C+C=[0,2]$. (Hint: $\frac12C$ consists of the numbers with ternary digits in $\{0,1\}$.)
- 6 Show that $\mathbf 1_C$ is Riemann integrable on $[0,1]$ while $\mathbf 1_{K_{1/2}}$ is not.
- 7 Compute $c(\frac14)$, $c(\frac79)$ and $c(\frac12)$ for the Cantor function $c$.
- 8 Is there a closed set $F\subseteq[0,1]$ with empty interior and $m(F)=0.999$? Is there one with $m(F)=1$?
- 1 Show that $f$ is measurable if and only if $\{f\gt q\}\in\mathcal M$ for every rational $q$.
- 2 Let $f$ be measurable on $E$. Show that $\{x\in E:f(x)=c\}$ is measurable for every $c$, and that $f^2$ and $\sqrt{|f|}$ are measurable.
- 3 Let $f:\mathbb{R}\to\mathbb{R}$ be continuous a.e. (its points of discontinuity form a null set). Is $f$ measurable?
- 4 Let $f_n$ be measurable on $E$. Show that $\{x:\lim_nf_n(x)\text{ exists in }\mathbb{R}\}$ is measurable.
- 5 Let $f(x)=x$ on $[0,1]$. Write the simple function $\varphi_3$ of Theorem 4.42 in canonical form, and compute $\sum_k\frac k8\,m(\{\varphi_3=\frac k8\})$.
- 6 Give an example of measurable $f_n$ on $[0,1]$ with $f_n\to0$ a.e. but $f_n(x)\to0$ for no $x$ in a dense set.
- 7 Prove that if $f:\mathbb{R}\to\mathbb{R}$ is measurable and $f(x+y)=f(x)+f(y)$ for all $x,y$, then $f(x)=cx$. (You may use Lusin's theorem and Steinhaus's theorem; a sketch suffices.)
- 1 For $f_n(x)=nxe^{1-nx}$ on $[0,1]$, find the pointwise limit and $\sup_{[0,1]}f_n$. For $\varepsilon=0.1$, find a set $F=[\varepsilon,1]$ as in Egorov's theorem and the least $N$ with $\sup_F f_n\lt0.05$ for all $n\ge N$.
- 2 Show that Egorov's theorem fails on $[0,\infty)$ for $f_n(x)=\frac xn$ (which converges to $0$ pointwise).
- 3 Find a closed $F\subseteq[0,1]$ with $m([0,1]\setminus F)\lt\varepsilon$ such that $\mathbf 1_{\mathbb Q}|_F$ is continuous. Is there one with $m([0,1]\setminus F)=0$?
- 4 Construct an open set $U\subseteq\mathbb{R}$ that is dense in $\mathbb{R}$ and has $m(U)\lt1$.
- 5 Let $E\in\mathcal M$ and suppose $m(E\cap I)\le\frac12\ell(I)$ for every bounded interval $I$. Prove that $m(E)=0$.
- 6 Show that there is no countably additive, translation-invariant $\mu$ defined on all subsets of $\mathbb R$ with $0\lt\mu([0,1])\lt\infty$.
- 7 Prove that every set $A\subseteq\mathbb{R}$ (measurable or not) is contained in a $G_\delta$ set $G$ with $m(G)=m^*(A)$.
E1. CSIR Part B Let $A=\big([0,3]\setminus\mathbb{Q}\big)\cup(5+C)\cup\bigcup_{n\ge1}\big(10+n,\ 10+n+3^{-n}\big)$, where $C$ is the Cantor set. Then $m(A)$ equals (A) $3$; (B) $\frac72$; (C) $4$; (D) $\frac92$.
E2. CSIR Part B GATE MA From $[0,1]$ remove, at stage $n=1,2,3,\dots$, an open interval of length $5^{-n}$ from the middle of each of the $2^{n-1}$ closed intervals that remain. The measure of the set that is left is (A) $0$; (B) $\frac13$; (C) $\frac12$; (D) $\frac23$.
E3. CSIR Part B Let $E$ be the set of $x\in[0,1]$ having a decimal expansion that uses neither the digit $3$ nor the digit $8$. Then (A) $E$ is countable and $m(E)=0$; (B) $E$ is uncountable and $m(E)=0$; (C) $m(E)=\frac45$; (D) $E$ is not Lebesgue measurable.
E4. CSIR Part C Which of the following statements are true? (One or more options may be correct.) (A) Every uncountable subset of $\mathbb{R}$ has positive outer measure. (B) If $A\subseteq[0,1]$ and $m^*(A)=1$, then $A$ is dense in $[0,1]$. (C) Every closed subset of $[0,1]$ with empty interior has measure $0$. (D) Every Lebesgue measurable set of positive measure contains a subset that is not Lebesgue measurable.
E5. CSIR Part C Let $f_n,f$ be real-valued measurable functions with $f_n\to f$ pointwise on the set $E$. Which of the following are true? (A) If $E=[0,1]$, then for every $\varepsilon\gt0$ there is a measurable $F\subseteq E$ with $m(E\setminus F)\lt\varepsilon$ on which $f_n\to f$ uniformly. (B) If $E=[0,1]$, there is a measurable $F\subseteq E$ with $m(E\setminus F)=0$ on which $f_n\to f$ uniformly. (C) Statement (A) holds with $E=\mathbb{R}$. (D) If $E=[0,1]$, then $f$ is continuous at every point of a closed set $F$ with $m(E\setminus F)\lt\varepsilon$.
E6. CSIR Part C Which of the following are true? (A) The set of irrational numbers is a $G_\delta$ set. (B) Every Lebesgue measurable subset of $\mathbb{R}$ is a Borel set. (C) Every Lebesgue measurable set is the union of an $F_\sigma$ set and a null set. (D) The Cantor set is a Borel set.
Suggested plan — 3 lectures × 60 min.
- Outer measure. Motivate with the three wishes (length, translation invariance, countable additivity). Definition 4.15, Theorem 4.34 — prove (2) and (3) on the board (the chain-of-intervals argument is worth the time: ask why it fails over $\mathbb Q$). Example 1. Close with the question "is $m^*$ additive?" and leave it open.
- Measurable sets and the Cantor family. Carathéodory's criterion with the picture of a test set $A$ cut by $E$; Theorem 4.35 (prove the countable-union step in full, the interval step if time permits), continuity of measure, regularity (statement and the idea "open from outside, closed from inside"). Then the Cantor set and fat Cantor sets with Interactive 1 projected: move $\alpha$ from $1$ to $0.5$ and ask the class what changes and what does not. Example 2 and Example 3 as a quick-fire true/false round.
- Vitali, measurable functions, Egorov and Lusin. Vitali's set in 15 minutes (the three bullet steps). Definition 4.20, Theorem 4.41 (do the $f+g$ trick with rationals), Theorem 4.42 with Example 6. Egorov with Interactive 2 (compare the sliding bump with $x^n$), Lusin as a statement with the $\mathbf 1_{\mathbb Q}$ warning. Exam corner E4–E6 as a class discussion.
Discussion questions. (1) Why does the proof that $m^*([a,b])=b-a$ need Heine–Borel — what would go wrong with a cover by infinitely many intervals if we skipped it? (2) The Cantor set and a fat Cantor set are homeomorphic (both are compact, perfect and totally disconnected). So is "measure zero" a topological property? (3) Which of the three wishes would you give up to measure every set — and what do Vitali and Banach–Tarski say about each choice?
Misconceptions to watch for. "Null means countable"; "nowhere dense means null"; "dense means full measure"; confusing the measure of the Cantor set ($0$) with the total length removed ($1$); using continuity from above with infinite measure; believing every measurable set is Borel; reading Lusin's theorem as "$f$ is continuous at the points of $F$"; applying Egorov on $\mathbb{R}$.
Suggested homework. I(2, 3, 5, 7, 8), II(1, 3, 4, 5, 7), III(1, 3, 5), IV(1, 3, 5, 7); Exam corner E1, E3, E4, E6.
Research question. How large is the set of real numbers that can be approximated "unusually well" by fractions? For $\tau\gt0$ let $W_\tau$ be the set of $x\in[0,1]$ for which $\big|x-\frac pq\big|\lt q^{-\tau}$ holds for infinitely many pairs of integers $(p,q)$ with $q\ge1$. Is $W_\tau$ null, of full measure, or something in between?
Why it matters. Rational approximation decides whether small divisors $q\omega-p$ can make a perturbation series blow up. This is the heart of the KAM theory of stable planetary orbits, of resonances in the rings of Saturn and in particle accelerators, and of vibration problems in rotating machinery, where engineers avoid frequency ratios that are too close to fractions. Measure theory answers the question "how typical is a dangerous frequency?".
What to do.
- For fixed $q$, let $A_q=\bigcup_{p=0}^{q}\big(\frac pq-q^{-\tau},\frac pq+q^{-\tau}\big)\cap[0,1]$. Show $m(A_q)\le2(q+1)q^{-\tau}$, and use Borel–Cantelli (Exercise I.7) to prove that $W_\tau$ is null for every $\tau\gt2$.
- Read the statement of Dirichlet's approximation theorem and explain why $W_2$ contains every irrational number of $[0,1]$, so $m(W_2)=1$: the measure jumps from $1$ to $0$ at $\tau=2$.
- Experiment: draw $10^5$ random $x\in[0,1]$ (any seeded generator) and, for each, count the $q\le10^4$ with $\operatorname{dist}(qx,\mathbb Z)\lt q^{1-\tau}$. Compare the average counts for $\tau=2$ and $\tau=2.5$ with the predictions $\sum_{q\le Q}\min(1,2q^{1-\tau})$.
- Show that the Liouville number $\sum_{k\ge1}10^{-k!}$ lies in $W_\tau$ for every $\tau$, so $\bigcap_\tau W_\tau$ is a null set that is nevertheless dense and uncountable.
Tools: Python (NumPy) or a spreadsheet for step 3; Exercise I.7 and Interactive 1 for the covering picture.
Hand in: The proof for $\tau\gt2$, a table of average counts against $Q$ for both values of $\tau$, and a paragraph interpreting the jump at $\tau=2$.
Going further. Khintchine's theorem replaces $q^{-\tau}$ by $\psi(q)/q$ and decides the measure of the set by the convergence of $\sum\psi(q)$; the Jarník–Besicovitch theorem computes the Hausdorff dimension $\frac2\tau$ of the null sets $W_\tau$. Look up either, and explain how Hausdorff dimension distinguishes null sets such as $\mathbb{Q}$ (dimension $0$) and the Cantor set (dimension $\log2/\log3\approx0.631$).
🧪 Self-check quiz
Q1. The Lebesgue measure of $\big([0,1]\setminus\mathbb Q\big)\cup\big(\mathbb{Q}\cap[2,3]\big)$ is
Q2. The Cantor set $C$ is
Q3. A fat Cantor set $K\subseteq[0,1]$ with $m(K)=\frac12$
Q4. For $E_n=[n,\infty)$ we have $m(E_n)=\infty$ for all $n$ and $m\big(\bigcap_nE_n\big)=0$. This shows that
Q5. Which statement about a measurable $f:[0,1]\to\mathbb{R}$ is guaranteed?