4.5 The Lebesgue Integral and Convergence Theorems
Lebesgue's integral for simple, non-negative and integrable functions; the monotone convergence theorem, Fatou's lemma and the dominated convergence theorem; when limits and integrals may be exchanged and when they may not; Riemann versus Lebesgue; and the space $L^1$.
A large part of analysis consists of exchanging two limiting operations: a limit with an integral, a series with an integral, a derivative with an integral. With the Riemann integral this is a delicate business. A pointwise limit of integrable functions need not be integrable at all (the indicators of finite sets of rationals increase to the Dirichlet function), and even uniform convergence, the usual Riemann-world sufficient condition, is far too strong for practice and useless on unbounded intervals. Lebesgue's integral, built on the measure of §4.4, comes with three theorems that settle almost every such question in one line: the monotone convergence theorem (increasing non-negative sequences), Fatou's lemma (an inequality that always holds for non-negative functions), and the dominated convergence theorem (anything bounded by a fixed integrable function).
These theorems are the working tools of probability (an expectation is a Lebesgue integral), Fourier analysis, partial differential equations and mathematical physics, and they are the most examined part of this chapter. CSIR NET Part B asks for limits such as $\lim_n\int_0^n(1-\frac xn)^n\,dx$ or for Lebesgue integrals of functions defined differently on rationals, Cantor sets and their complements; Part C offers sequences $f_n$ and asks for which of them $\lim\int f_n=\int\lim f_n$, or which statements about $L^1$ functions must hold. The wrong options are always the four classical ways in which "mass" can escape — collected in Example 4 and in the ⚠️ box.
The Integral of Non-negative Functions
Throughout, $E$ is a measurable subset of $\mathbb{R}$ and "measurable" means Lebesgue measurable (§4.4). We use the conventions $0\cdot\infty=0$ and $a+\infty=\infty$ for $a\in[0,\infty]$.
If $\varphi=\sum_{i=1}^ra_i\mathbf 1_{E_i}$ is simple on $E$ with $a_i\ge0$ and $E_1,\dots,E_r$ disjoint measurable sets covering $E$, its integral is $$\int_E\varphi=\sum_{i=1}^ra_i\,m(E_i)\in[0,\infty],$$ and for measurable $A\subseteq E$, $\int_A\varphi=\int_E\varphi\,\mathbf 1_A=\sum_ia_i\,m(A\cap E_i)$.
The value does not depend on the representation: if also $\varphi=\sum_jb_j\mathbf 1_{F_j}$ with disjoint $F_j$, then on each non-empty $E_i\cap F_j$ we have $a_i=b_j$, and both sums equal $\sum_{i,j}a_i\,m(E_i\cap F_j)$ by additivity of $m$. For example $\int_{[0,1]}\mathbf 1_{\mathbb Q}=1\cdot m(\mathbb Q\cap[0,1])+0\cdot m([0,1]\setminus\mathbb Q)=0$.
Proof. Write $\varphi=\sum a_i\mathbf 1_{E_i}$, $\psi=\sum b_j\mathbf 1_{F_j}$ and use the common refinement $\{E_i\cap F_j\}$: on $E_i\cap F_j$, $a\varphi+b\psi$ takes the value $aa_i+bb_j$, so $\int(a\varphi+b\psi)=\sum_{i,j}(aa_i+bb_j)m(E_i\cap F_j)$, which is $a\int\varphi+b\int\psi$. If $\varphi\le\psi$, then $a_i\le b_j$ whenever $E_i\cap F_j\ne\varnothing$, which gives (2). For (3), $\int_A\varphi=\sum_ia_i\,m(A\cap E_i)=\sum_ia_i\sum_km(A_k\cap E_i)$, and this equals $\sum_k\int_{A_k}\varphi$ by countable additivity of $m$ and rearrangement of non-negative terms. ∎
Compare with the Riemann integral: there, $f$ was approximated from below by step functions, constant on intervals; here by simple functions, constant on arbitrary measurable sets. That single change is what lets $\mathbf 1_{\mathbb Q}$ have an integral.
- If $f\le g$ then $\int_Ef\le\int_Eg$; if $A\subseteq B\subseteq E$ are measurable then $\int_Af\le\int_Bf$.
- $\int_Ecf=c\int_Ef$ for $c\in[0,\infty)$.
- If $m(E)=0$ then $\int_Ef=0$.
- (Markov's inequality.) $m(\{f\ge t\})\le\dfrac1t\displaystyle\int_Ef$ for every $t\gt0$.
- $\int_Ef=0$ if and only if $f=0$ a.e. on $E$.
- If $\int_Ef\lt\infty$ then $f\lt\infty$ a.e. on $E$.
Proof. (1) Every simple $\varphi\le f$ also satisfies $\varphi\le g$; and $\varphi\mathbf 1_A\le\varphi\mathbf 1_B$. (2) For $c\gt0$, $\varphi\mapsto c\varphi$ is a bijection between simple functions below $f$ and below $cf$; $c=0$ is clear. (3) Every simple $\varphi$ has $\int_E\varphi=\sum a_im(E_i)=0$. (4) The simple function $t\,\mathbf 1_{\{f\ge t\}}$ lies below $f$, so $t\,m(\{f\ge t\})\le\int f$. (5) If $f=0$ outside a null set $N$, every simple $\varphi\le f$ vanishes outside $N$, so $\int\varphi=\int_N\varphi=0$. Conversely, if $\int f=0$ then by (4) each $\{f\ge\frac1k\}$ is null, and so is their union $\{f\gt0\}$. (6) $\{f=\infty\}\subseteq\{f\ge t\}$, so $m(\{f=\infty\})\le\frac1t\int f\to0$ as $t\to\infty$. ∎
Proof. Changing $f_n$ and $f$ on a null set $N$ changes none of the integrals (for simple $\varphi$, $\int_E\varphi=\int_{E\setminus N}\varphi$ by Theorem 4.45(3); take suprema), so assume convergence everywhere. $f$ is measurable (Theorem 4.41). Since $f_n\le f_{n+1}\le f$, the integrals $\int f_n$ increase to a limit $L\le\int f$.
For the reverse inequality, fix a simple $\varphi$ with $0\le\varphi\le f$ and a number $c\in(0,1)$, and let $A_n=\{x\in E:f_n(x)\ge c\,\varphi(x)\}$. These sets are measurable and increase with $n$. They exhaust $E$: if $\varphi(x)=0$ then $x\in A_1$; if $\varphi(x)\gt0$ then $f(x)\ge\varphi(x)\gt c\varphi(x)$, so $f_n(x)\ge c\varphi(x)$ for large $n$. Hence $$\int_Ef_n\ge\int_{A_n}f_n\ge c\int_{A_n}\varphi\ \longrightarrow\ c\int_E\varphi,$$ where the limit uses the countable additivity of $A\mapsto\int_A\varphi$ (Theorem 4.45(3)) and continuity from below (Theorem 4.36(1)). So $L\ge c\int_E\varphi$. Letting $c\to1$ and taking the supremum over $\varphi$ gives $L\ge\int_Ef$. ∎
The factor $c\lt1$ is the whole trick: with $c=1$ the sets $A_n=\{f_n\ge\varphi\}$ might never exhaust $E$ (if $f_n\lt f=\varphi$ everywhere).
- $\int_E(f+g)=\int_Ef+\int_Eg$.
- (Term-by-term integration.) $\displaystyle\int_E\sum_{k=1}^\infty g_k=\sum_{k=1}^\infty\int_Eg_k$.
- $\int_Ef=\lim_n\int_E\varphi_n$ for the simple functions $\varphi_n=\min(n,2^{-n}\lfloor2^nf\rfloor)$ of Theorem 4.42.
- $\nu(A)=\int_Af$ defines a countably additive function on the measurable subsets of $E$, and $\nu(A)=0$ whenever $m(A)=0$.
Proof. (3) is Theorem 4.47 applied to $\varphi_n\uparrow f$. (1) If $\varphi_n\uparrow f$ and $\psi_n\uparrow g$ are simple, then $\varphi_n+\psi_n\uparrow f+g$, and by MCT and Theorem 4.45(1), $\int(f+g)=\lim\int(\varphi_n+\psi_n)$, which is $\lim\big(\int\varphi_n+\int\psi_n\big)=\int f+\int g$. (2) By (1) and induction, $\int\sum_{k\le n}g_k=\sum_{k\le n}\int g_k$; the partial sums increase to $\sum_kg_k$, so MCT finishes. (4) Apply (2) to $g_k=f\mathbf 1_{A_k}$; the last claim is Theorem 4.46(3). ∎
Part (3) is Lebesgue's "slice the range" integral, and with Tonelli's theorem (Theorem 4.57) it becomes the layer-cake formula $\int_Ef=\int_0^\infty m(\{f\gt t\})\,dt$ — the formula behind the Lebesgue sums in the teaser of the chapter introduction.
Fatou's Lemma and Integrable Functions
Proof. Let $g_n=\inf_{k\ge n}f_k$. Then $g_n$ is measurable, $0\le g_1\le g_2\le\cdots$, and $g_n\uparrow\liminf_nf_n$. By MCT, $\int\liminf f_n=\lim_n\int g_n$. But $g_n\le f_k$ for every $k\ge n$, so $\int g_n\le\inf_{k\ge n}\int f_k$, and letting $n\to\infty$, $\lim\int g_n\le\liminf\int f_k$. ∎
The inequality can be strict. With $f_n=\mathbf 1_{[0,1]}$ for odd $n$ and $f_n=\mathbf 1_{[1,2]}$ for even $n$, $\liminf f_n=\mathbf 1_{\{1\}}$ has integral $0$ while every $\int f_n=1$. With $f_n=n\mathbf 1_{(0,1/n)}$, $\int\liminf f_n=0\lt1$. The hypothesis $f_n\ge0$ matters too: $f_n=-\mathbf 1_{[n,n+1]}$ has $\liminf f_n=0$ but $\int f_n=-1$.
Lebesgue integrability is absolute: $f$ is integrable if and only if $|f|$ is. There is no Lebesgue analogue of a conditionally convergent improper integral; $\frac{\sin x}{x}$, whose improper integral over $(0,\infty)$ is $\frac\pi2$ (§4.2), is not Lebesgue integrable on $(0,\infty)$ (Example 2).
- (Linearity.) $af+bg\in\mathcal L^1(E)$ and $\int(af+bg)=a\int f+b\int g$ (where $f,g$ are finite, i.e. a.e.).
- (Monotonicity.) $f\le g$ a.e. implies $\int f\le\int g$; and $\big|\int f\big|\le\int|f|$.
- If $f=g$ a.e. then $\int f=\int g$; and $f$ is finite a.e.
- (Absolute continuity.) For every $\varepsilon\gt0$ there is $\delta\gt0$ such that $\int_A|f|\lt\varepsilon$ for every measurable $A\subseteq E$ with $m(A)\lt\delta$.
Proof. (3) $f$ is finite a.e. by Theorem 4.46(6) applied to $|f|$. If $f=g$ outside a null set $N$, then $\int f^\pm=\int_{E\setminus N}f^\pm+\int_Nf^\pm=\int_{E\setminus N}g^\pm$ (Theorem 4.48(4)), and the same for $g$.
(1) By (3) we may assume $f,g$ real-valued. $|af+bg|\le|a||f|+|b||g|$, so $af+bg$ is integrable. For $h=f+g$: from $h^+-h^-=f^+-f^-+g^+-g^-$ we get $h^++f^-+g^-=h^-+f^++g^+$, a sum of non-negative functions on each side. Integrate using Theorem 4.48(1); all terms are finite, so we can rearrange: $\int h^+-\int h^-=\int f^+-\int f^-+\int g^+-\int g^-$. For $c\ge0$, $(cf)^\pm=cf^\pm$; for $c\lt0$, $(cf)^\pm=|c|f^\mp$; in both cases $\int cf=c\int f$.
(2) $g-f\ge0$ a.e., so $0\le\int(g-f)=\int g-\int f$ by (1). Apply this to $-|f|\le f\le|f|$.
(4) The functions $\min(|f|,n)$ increase to $|f|$, so by MCT there is $n$ with $\int\big(|f|-\min(|f|,n)\big)\lt\frac\varepsilon2$. Put $\delta=\frac\varepsilon{2n}$. If $m(A)\lt\delta$, then $$\int_A|f|\le\int\big(|f|-\min(|f|,n)\big)+\int_A\min(|f|,n),$$ which is less than $\frac\varepsilon2+n\,m(A)\lt\varepsilon$. ∎
The Dominated Convergence Theorem
Proof. Discarding a null set, assume $f_n\to f$ and $|f_n|\le g$ everywhere, with $g$ finite. Then $|f|\le g$, so $f$ is measurable (Theorem 4.41) and integrable. Apply Fatou's lemma to the non-negative functions $h_n=2g-|f_n-f|$ (non-negative because $|f_n-f|\le|f_n|+|f|\le2g$), which converge to $2g$: $$\int2g\le\liminf_n\int\big(2g-|f_n-f|\big)=\int2g-\limsup_n\int|f_n-f|.$$ Since $\int2g$ is finite we may cancel it: $\limsup_n\int|f_n-f|\le0$. Finally $\big|\int f_n-\int f\big|\le\int|f_n-f|\to0$. ∎
- (Bounded convergence theorem.) If $m(E)\lt\infty$, $f_n\to f$ a.e. on $E$ and $|f_n|\le M$ for all $n$, then $\int_Ef_n\to\int_Ef$.
- (Series.) If $g_k$ are measurable and $\sum_k\int_E|g_k|\lt\infty$, then $\sum_kg_k(x)$ converges absolutely for a.e. $x$, its sum is integrable, and $\int_E\sum_kg_k=\sum_k\int_Eg_k$.
- (Continuity in a parameter.) Let $f(x,t)$ be defined for $x\in E$, $t\in(a,b)$, measurable in $x$ for each $t$, with $|f(x,t)|\le g(x)$ for all $t$, where $g\in\mathcal L^1(E)$. If $t\mapsto f(x,t)$ is continuous at $t_0$ for a.e. $x$, then $F(t)=\int_Ef(x,t)\,dx$ is continuous at $t_0$.
- (Differentiation under the integral sign.) If, in addition, $f(\cdot,t)\in\mathcal L^1(E)$, $\frac{\partial f}{\partial t}(x,t)$ exists for all $t\in(a,b)$ and all $x\in E$, and $\big|\frac{\partial f}{\partial t}(x,t)\big|\le h(x)$ for all $x,t$ with $h\in\mathcal L^1(E)$, then $F$ is differentiable and $F'(t)=\int_E\frac{\partial f}{\partial t}(x,t)\,dx$.
Proof. (1) Dominate by the constant $M$, which is integrable because $m(E)\lt\infty$. (2) By Theorem 4.48(2), $G=\sum_k|g_k|$ has $\int G=\sum\int|g_k|\lt\infty$, so $G\lt\infty$ a.e. (Theorem 4.46(6)); where $G(x)\lt\infty$ the series $\sum g_k(x)$ converges absolutely. Its partial sums are dominated by $G$, and DCT gives the formula. (3) By the sequential criterion (§3.1) it suffices to show $F(t_n)\to F(t_0)$ for every sequence $t_n\to t_0$; and $f(\cdot,t_n)\to f(\cdot,t_0)$ a.e. with domination by $g$. (4) Fix $t$ and $s_n\to0$ with $s_n\ne0$. By the mean value theorem (§3.3), the difference quotients $q_n(x)=\frac{f(x,t+s_n)-f(x,t)}{s_n}$ equal $\frac{\partial f}{\partial t}(x,\xi_n)$ for some $\xi_n$ between $t$ and $t+s_n$, so $|q_n|\le h$; and $q_n(x)\to\frac{\partial f}{\partial t}(x,t)$. DCT gives $\frac{F(t+s_n)-F(t)}{s_n}=\int q_n\to\int\frac{\partial f}{\partial t}(x,t)\,dx$. ∎
- Find the pointwise limit $f(x)=\lim f_n(x)$ (a.e. is enough). Expressions such as $(1\pm\frac xn)^n\to e^{\pm x}$, $n\sin\frac xn\to x$, $x^{1/n}\to1$ $(x\gt0)$, $x^n\to0$ $(0\le x\lt1)$ recur constantly.
- Non-negative and increasing? Use MCT — no integrability needed, and the answer may be $\infty$.
- Look for a dominating function $g$ independent of $n$, with $\int g\lt\infty$. Standard bounds: $(1-\frac xn)^n\le e^{-x}$ on $[0,n]$; $(1+\frac xn)^n\ge1+x$, and for $n\ge2$ $(1+\frac xn)^{n}\ge(1+\frac x2)^2$; $|\sin u|\le|u|$; $\frac{nx}{1+n^2x^2}\le\frac12$; $x^{1/n}\ge x$ on $[0,1]$. The best candidate is $\sup_n|f_n|$ — DCT applies (with $g=\sup_{n}|f_n|$) exactly when this is integrable, possibly after discarding finitely many $n$.
- Bounded functions on a set of finite measure? Bounded convergence theorem.
- No domination? Then compute $\int f_n$ directly (often by substitution); expect the answer to differ from $\int f$, and identify the escaping mass: a spike at a point, a bump sliding to infinity, or a function spreading thinly over an ever larger set (Example 4). Fatou's inequality $\int f\le\liminf\int f_n$ still holds when $f_n\ge0$.
- (Riesz–Fischer.) $L^1(E)$ is complete: every Cauchy sequence in $\|\cdot\|_1$ converges in $L^1$.
- Simple integrable functions, step functions, and continuous functions vanishing outside a bounded interval are dense in $L^1(\mathbb{R})$.
Proof (sketch). (1) Given a Cauchy sequence, choose $n_1\lt n_2\lt\cdots$ with $\|f_{n_{k+1}}-f_{n_k}\|_1\lt2^{-k}$. By Theorem 4.52(2), $\sum_k|f_{n_{k+1}}-f_{n_k}|$ is integrable and finite a.e., so the telescoping series $f_{n_1}+\sum_k(f_{n_{k+1}}-f_{n_k})$ converges a.e. to some $f$, dominated by $|f_{n_1}|+\sum_k|f_{n_{k+1}}-f_{n_k}|\in\mathcal L^1$. DCT gives $f_{n_k}\to f$ in $L^1$, and a Cauchy sequence with a convergent subsequence converges. (2) For $f\ge0$, the simple $\varphi_n\uparrow f$ of Theorem 4.42 satisfy $\|f-\varphi_n\|_1\to0$ by DCT (dominated by $f$); split a general $f$ into $f^\pm$. A simple integrable function is a combination of indicators $\mathbf 1_A$ with $m(A)\lt\infty$, and by Exercise I.9 of §4.4 each $A$ is close in measure to a finite union of intervals, i.e. $\mathbf 1_A$ is close in $L^1$ to a step function; finally a step function is approximated in $L^1$ by continuous piecewise-linear functions (round off each jump over a short interval). ∎
Proof. Choose $n_k$ increasing with $\|f_{n_k}-f\|_1\lt2^{-k}$; then $\sum_k\int|f_{n_k}-f|\lt\infty$, so by Theorem 4.52(2) $\sum_k|f_{n_k}(x)-f(x)|\lt\infty$ for a.e. $x$, and the terms tend to $0$. The typewriter sequence of Example 4(d) converges in $L^1$ but nowhere pointwise; the spike $n\mathbf 1_{(0,1/n)}$ converges everywhere but not in $L^1$ ($\|f_n-0\|_1=1$). ∎
Riemann versus Lebesgue
Proof. $f$ is bounded, $|f|\le M$. By Riemann's criterion (§4.1) choose partitions $P_1\subseteq P_2\subseteq\cdots$, each refining the previous one, with $U(f,P_n)-L(f,P_n)\to0$. For $P_n$ with subintervals $I_k$ and $m_k=\inf_{I_k}f$, $M_k=\sup_{I_k}f$, define step functions $\ell_n=\sum_km_k\mathbf 1_{I_k}$ and $u_n=\sum_kM_k\mathbf 1_{I_k}$ (using half-open subintervals, say). Off the countable set $Q$ of all partition points, $\ell_n\le f\le u_n$; the $\ell_n$ increase and the $u_n$ decrease (refinement), and their Lebesgue integrals are exactly the Darboux sums: $\int\ell_n=L(f,P_n)$, $\int u_n=U(f,P_n)$.
Let $\ell=\lim\ell_n$ and $u=\lim u_n$; they are measurable, bounded by $M$, and $\ell\le f\le u$ off $Q$. By the bounded convergence theorem, $$\int_{[a,b]}(u-\ell)=\lim_n\big(U(f,P_n)-L(f,P_n)\big)=0.$$ Since $u-\ell\ge0$, Theorem 4.46(5) gives $u=\ell$ a.e.; hence $f=\ell$ a.e., so $f$ is measurable (Theorem 4.41(5), completeness of Lebesgue measure) and $\int_{[a,b]}f=\int\ell=\lim L(f,P_n)$, which is the Riemann integral. ∎
The converse fails badly: $\mathbf 1_{\mathbb Q}$ and $\mathbf 1_{K_\alpha}$ (a fat Cantor set) are Lebesgue integrable on $[0,1]$, with integrals $0$ and $1-\alpha$, but not Riemann integrable. The first is at least equal a.e. to a Riemann integrable function (the zero function); the second is not even that — any $g=\mathbf 1_{K_\alpha}$ a.e. is still discontinuous at each point of $K_\alpha$, since every neighbourhood of such a point meets both $K_\alpha$ and its complement in sets of positive measure.
Proof. The functions $|f|\mathbf 1_{[a,n]}$ increase to $|f|$; by MCT and Theorem 4.55, $\int_{[a,\infty)}|f|=\lim_n\int_a^n|f|$, which is finite exactly when the improper integral of $|f|$ converges (the partial integrals increase). If it does, then for any $b_k\to\infty$ the functions $f\mathbf 1_{[a,b_k]}$ converge to $f$ and are dominated by $|f|\in\mathcal L^1$, so DCT gives $\int_a^{b_k}f\to\int_{[a,\infty)}f$; thus the improper integral exists and has the same value. ∎
So absolutely convergent improper integrals are Lebesgue integrals, while conditionally convergent ones ($\int_0^\infty\frac{\sin x}x$, $\int_1^\infty\sin(x^2)\,dx$) are not; they are limits of Lebesgue integrals over $[0,b]$.
- (Tonelli.) If $f\ge0$, then $x\mapsto\int f(x,y)\,dy$ and $y\mapsto\int f(x,y)\,dx$ are measurable (defined a.e.) and $$\iint f\,dm_2=\int\Big(\int f(x,y)\,dy\Big)dx=\int\Big(\int f(x,y)\,dx\Big)dy,$$ all three possibly $\infty$.
- (Fubini.) If $\iint|f|\,dm_2\lt\infty$ — which, by Tonelli, may be checked on either iterated integral of $|f|$ — the same equalities hold, with finite values.
The integrability hypothesis in Fubini's theorem is essential. For $f(x,y)=\frac{x^2-y^2}{(x^2+y^2)^2}$ on $(0,1]^2$, $\frac{\partial}{\partial y}\frac{y}{x^2+y^2}=f(x,y)$, so $\int_0^1f(x,y)\,dy=\frac1{1+x^2}$ and $\int_0^1\!\int_0^1f\,dy\,dx=\frac\pi4$; since $f(y,x)=-f(x,y)$, the other order gives $-\frac\pi4$. By Tonelli, $\iint|f|=\infty$. (§4.2 used Tonelli, legitimately, for the non-negative integrand in $\Gamma(p)\Gamma(q)=\Gamma(p+q)B(p,q)$.)
- $\lim\int f_n\ne\int\lim f_n$ in general. Spike $n\mathbf 1_{(0,1/n)}$, sliding bump $\mathbf 1_{[n,n+1]}$, flattening $\frac1n\mathbf 1_{[0,n]}$, concentration $n^2xe^{-nx}$: all tend to $0$ pointwise with integrals tending to $1$.
- Uniform convergence is not enough on sets of infinite measure. $\frac1n\mathbf 1_{[0,n]}\to0$ uniformly on $\mathbb{R}$ with $\int=1$. On a set of finite measure, uniform convergence (indeed bounded convergence) suffices.
- Fatou is an inequality, in one direction, for non-negative functions. It can be strict; $\int\liminf f_n\ge\liminf\int f_n$ is false; and without $f_n\ge0$ even Fatou fails ($f_n=-\mathbf 1_{[n,n+1]}$).
- MCT is for increasing sequences. For decreasing ones it fails: $\mathbf 1_{[n,\infty)}\downarrow0$ with integrals $\infty$. It holds for decreasing sequences when $\int f_1\lt\infty$ (apply DCT, or MCT to $f_1-f_n$).
- Improper Riemann integrable $\not\Rightarrow$ Lebesgue integrable. $\frac{\sin x}{x}$ on $(0,\infty)$; $\sum\frac{(-1)^n}n\mathbf 1_{[n,n+1)}$. Only absolutely convergent improper integrals are Lebesgue integrals.
- Lebesgue integrable $\not\Rightarrow$ Riemann integrable, even on $[a,b]$ and even for bounded functions: $\mathbf 1_{\mathbb Q}$, $\mathbf 1_{K_\alpha}$; and $\mathbf 1_{K_\alpha}$ is not a.e. equal to any Riemann integrable function.
- $L^1$ convergence and a.e. convergence are independent. Typewriter: $L^1$ but nowhere; spike: everywhere but not $L^1$. $L^1$ convergence does give an a.e. convergent subsequence.
- $f\in L^1(\mathbb{R})$ does not mean $f\to0$ at $\infty$, nor $f$ bounded — not even for continuous $f$ (spikes of height $n$ and area $n^{-2}$).
- $\int f=0$ gives $f=0$ a.e. only for $f\ge0$ (or when $\int_Af=0$ for every measurable $A$, Exercise II.7).
- Fubini needs $\iint|f|\lt\infty$. $\frac{x^2-y^2}{(x^2+y^2)^2}$ on $(0,1]^2$ has iterated integrals $\frac\pi4$ and $-\frac\pi4$. Tonelli (no integrability needed) is for $f\ge0$.
- The fundamental theorem of calculus can fail for Lebesgue integrals. The Cantor function $c$ has $c'=0$ a.e., so $\int_0^1c'=0\ne1=c(1)-c(0)$; $F(x)-F(a)=\int_a^xF'$ needs $F$ absolutely continuous, a strengthening of bounded variation (§4.3).
Exercises
- 1 On $[0,1]$ let $f=1$ on the Cantor set $C$, $f=2$ on $(\mathbb Q\cap[0,1])\setminus C$, and $f=3$ elsewhere. Compute $\int_{[0,1]}f$.
- 2 Show that $\int_{(0,1]}\lfloor1/x\rfloor\,dx=\infty$, although $\lfloor1/x\rfloor$ is finite at every point, and that $x\lfloor 1/x\rfloor$ is integrable on $(0,1]$.
- 3 Prove $\displaystyle\int_0^1\frac{-\log(1-x)}{x}\,dx=\frac{\pi^2}6$.
- 4 Prove $\displaystyle\int_0^\infty\frac{x^2}{e^x-1}\,dx=2\zeta(3)$, where $\zeta(3)=\sum_{k\ge1}k^{-3}$.
- 5 Let $f\ge0$ be measurable on $[0,1]$ with $\int f=2$. Show $m(\{f\ge10\})\le\frac15$, and find an $f$ for which equality holds.
- 6 Show that MCT fails for $f_n=\mathbf 1_{[n,\infty)}$ (a decreasing sequence), and prove: if $f_1\ge f_2\ge\cdots\ge0$ and $\int f_1\lt\infty$, then $\int f_n\to\int\lim f_n$.
- 7 Let $(q_n)$ enumerate $\mathbb Q$ and $f(x)=\sum_{n\ge1}2^{-n}|x-q_n|^{-1/2}\mathbf 1_{\{|x-q_n|\lt1\}}$. Show that $\int_{\mathbb{R}}f=4$, hence $f\lt\infty$ a.e., although $f$ is unbounded on every interval.
- 8 Let $f\ge0$ be integrable on $\mathbb{R}$. Show that $F(x)=\int_{(-\infty,x]}f$ is continuous and increasing, with $\lim_{x\to\infty}F(x)=\int_{\mathbb{R}}f$.
- 1 For $f_n=\mathbf 1_{[0,1]}$ ($n$ odd) and $f_n=\mathbf 1_{[1,2]}$ ($n$ even), compute $\int\liminf f_n$, $\liminf\int f_n$, $\int\limsup f_n$ and $\limsup\int f_n$.
- 2 Show that the conclusion of Fatou's lemma fails for $f_n=-\mathbf 1_{[n,n+1]}$ on $\mathbb{R}$. Which hypothesis is violated?
- 3 Decide: $\frac{\sin x}x\in\mathcal L^1(0,\infty)$? $\ \frac{\sin x}{x^2}\in\mathcal L^1(1,\infty)$? $\ \frac{\sin x}{x^{3/2}}\in\mathcal L^1(0,\infty)$?
- 4 For which real $a$ is $x^ae^{-x}\in\mathcal L^1(0,\infty)$? For which $a$ is $x^a\in\mathcal L^1(0,1)$, and $x^a\in\mathcal L^1(1,\infty)$?
- 5 Let $f\in\mathcal L^1(\mathbb{R})$. Prove that $\int_{\{|x|\gt n\}}|f|\to0$ and $n\,m(\{|f|\gt n\})\to0$.
- 6 Give a continuous $f:\mathbb{R}\to(0,\infty)$ in $\mathcal L^1(\mathbb{R})$ with $\limsup_{x\to\infty}f(x)=\infty$.
- 7 Let $f\in\mathcal L^1(E)$ with $\int_Af=0$ for every measurable $A\subseteq E$. Prove $f=0$ a.e.
- 1 $\displaystyle\lim_{n\to\infty}\int_0^1\frac{1+nx^2}{(1+x^2)^n}\,dx$
- 2 $\displaystyle\lim_{n\to\infty}\int_0^\infty\frac{dx}{(1+x/n)^n\,x^{1/n}}$
- 3 $\displaystyle\lim_{n\to\infty}\int_0^n\Big(1-\frac xn\Big)^ne^{x/2}\,dx$
- 4 $\displaystyle\lim_{n\to\infty}\int_0^1\frac{n^{3/2}x}{1+n^2x^2}\,dx$
- 5 $\displaystyle\lim_{n\to\infty}\int_0^\infty ne^{-nx}\cos x\,dx$ — by DCT after substituting $u=nx$, and directly.
- 6 $\displaystyle\lim_{n\to\infty}\int_0^1g(x^n)\,dx$ for continuous $g:[0,1]\to\mathbb{R}$.
- 7 Prove $\displaystyle\int_0^1\frac{\log(1/x)}{1+x}\,dx=\frac{\pi^2}{12}$ (justify the term-by-term integration).
- 8 For $F(t)=\displaystyle\int_0^\infty e^{-x}\frac{\sin tx}{x}\,dx$, show $F'(t)=\frac1{1+t^2}$ and $F(t)=\tan^{-1}t$.
- 9 Give $f_n\ge0$ continuous on $[0,\infty)$ with $f_n\to0$ uniformly and $\int f_n=1$ for all $n$. Which hypothesis of the dominated convergence theorem fails?
- 1 Let $f(x)=x^2$ for rational $x$ and $f(x)=x^3$ for irrational $x$ in $[0,1]$. Compute $\int_{[0,1]}f$ and decide Riemann integrability.
- 2 Let $q_1,q_2,\dots$ enumerate $\mathbb{Q}\cap[0,1]$ and $f_n=\mathbf 1_{\{q_1,\dots,q_n\}}$. Show each $f_n$ is Riemann integrable with integral $0$, $f_n\uparrow\mathbf 1_{\mathbb Q\cap[0,1]}$, which is not Riemann integrable, and find its Lebesgue integral.
- 3 Compute $\displaystyle\int_{(0,1]}\frac{dx}{\lfloor1/x\rfloor^2}$ in closed form.
- 4 For $f(x,y)=\frac{x^2-y^2}{(x^2+y^2)^2}$ on $(0,1]^2$, compute both iterated integrals, and show directly that $\iint|f|=\infty$.
- 5 Let $f\in\mathcal L^1[0,1]$ with $\int_0^xf=0$ for every $x\in[0,1]$. Prove $f=0$ a.e.
- 6 (Riemann–Lebesgue lemma.) Prove that $\int_{\mathbb{R}}f(x)\sin(nx)\,dx\to0$ for every $f\in\mathcal L^1(\mathbb{R})$. (First take $f=\mathbf 1_{[a,b]}$, then use Theorem 4.53(2).)
- 7 Using Tonelli's theorem, prove the layer-cake formula $\int_Ef=\int_0^\infty m(\{f\gt t\})\,dt$ for measurable $f\ge0$, and use it to compute $\int_{(0,1]}x^{-1/2}dx$ by "slicing the range".
E1. CSIR Part B GATE MA The value of $\displaystyle\lim_{n\to\infty}\int_0^1\frac{n\,x^{n-1}}{1+x}\,dx$ is (A) $0$; (B) $\frac12$; (C) $\log2$; (D) $1$.
E2. CSIR Part B Let $f\ge0$ be Lebesgue measurable on $[0,1]$ with $\int_{[0,1]}f=\frac14$. The largest possible value of $m(\{x:f(x)\ge1\})$ is (A) $\frac1{16}$; (B) $\frac12$; (C) $\frac14$; (D) $1$.
E3. CSIR Part B $\displaystyle\lim_{n\to\infty}\int_0^n\Big(1+\frac xn\Big)^ne^{-2x}\,dx$ equals (A) $1$; (B) $\frac12$; (C) $2$; (D) $\infty$.
E4. CSIR Part C For which of the following sequences on $[0,1]$ is $\displaystyle\lim_{n\to\infty}\int_0^1f_n=\int_0^1\lim_{n\to\infty}f_n$? (One or more options may be correct.) (A) $f_n=n\mathbf 1_{(0,1/n)}$; (B) $f_n(x)=x^n$; (C) $f_n(x)=nx^n(1-x)$; (D) $f_n(x)=n^2xe^{-nx}$.
E5. CSIR Part C Let $f\in L^1(\mathbb{R})$. Which of the following must be true? (A) $f$ is bounded. (B) $\lim_{x\to\infty}f(x)=0$. (C) $\lim_{n\to\infty}\int_n^{n+1}|f(x)|\,dx=0$. (D) $m(\{x:|f(x)|\gt t\})\le\frac1t\int_{\mathbb{R}}|f|$ for every $t\gt0$.
E6. CSIR Part C Which of the following are true? (A) Every Riemann integrable function on $[a,b]$ is Lebesgue integrable, with the same integral. (B) If the improper Riemann integral $\int_0^\infty f(x)\,dx$ converges, then $f\in L^1(0,\infty)$. (C) The indicator function of a closed nowhere dense subset of $[0,1]$ of positive measure is Lebesgue integrable but not Riemann integrable. (D) If $f_n\to f$ in $L^1[0,1]$, then $f_n(x)\to f(x)$ for almost every $x$.
Suggested plan — 3 lectures × 60 min.
- Building the integral; MCT. Definitions 4.23–4.24 with the picture "simple functions are constant on measurable sets, not on intervals". Theorem 4.46 (do Markov carefully — it is examined), the MCT with its proof (spend time on why $c\lt1$ is needed), Theorem 4.48, Example 1. Project Interactive 1 and compare $x^{-1/2}$ with $\frac1x$.
- Fatou, integrable functions, DCT. Fatou with the three counterexamples in the text, Definition 4.25, Theorem 4.50 (prove absolute continuity), Example 2, then the DCT with its two-line proof via Fatou. The strategy box and Example 3 on the board; let students find the dominating function in (a)–(c) before you show it, and ask why (d) has none.
- Escaping mass, $L^1$, Riemann vs Lebesgue. Example 4 with Interactive 2 (all four modes), Theorems 4.52–4.54, Example 5. Theorem 4.55 with its proof, Theorem 4.56, Tonelli/Fubini with the counterexample (let a student compute both orders). Finish with the Exam corner E4 and E6 as a class vote.
Discussion questions. (1) MCT needs no integrability hypothesis and DCT needs a dominating function. Why is there no "DCT without domination", and what replaces it in probability (uniform integrability)? (2) In which sense is the Lebesgue integral "absolute", and what is lost compared with improper Riemann integrals? (3) Each of the four escaping-mass examples has $\sup_nf_n\notin L^1$. Is the converse true — does $\lim\int f_n\ne\int\lim f_n$ whenever $\sup_n f_n\notin L^1$? (Example 7(b) says no.)
Misconceptions to watch for. "Pointwise convergence lets me pass to the limit"; "uniform convergence is always enough"; using Fatou backwards or for sign-changing functions; applying MCT to decreasing sequences; believing $\frac{\sin x}x\in L^1(0,\infty)$ because $\int_0^\infty\frac{\sin x}x=\frac\pi2$; "$L^1$ functions tend to $0$ at infinity"; swapping iterated integrals without checking $\iint|f|\lt\infty$.
Suggested homework. I(2, 3, 5, 6, 7), II(1, 3, 4, 5), III(1, 2, 3, 6, 7, 9), IV(1, 2, 4, 5); Exam corner E1, E3, E4, E6.
Research question. A gambler plays "double or nothing" with a fair coin, starting with one unit and staking everything each round. Her expected fortune is $1$ after every round, yet she is ruined with probability $1$. How can $\mathbb E[X_n]=1$ for all $n$ while $X_n\to0$ almost surely — and what does this have to do with the dominated convergence theorem?
Why it matters. Expectations are Lebesgue integrals, and "$\lim\mathbb E[X_n]=\mathbb E[\lim X_n]$" is exactly the exchange of limit and integral. Its failure underlies the danger of doubling ("martingale") betting systems, the need for credit limits and margin requirements in finance, and the stopping-time conditions in the optional stopping theorem used to price American options and to analyse clinical trials that stop early.
What to do.
- Model the coin tosses on $([0,1],m)$: the $k$-th toss is the $k$-th binary digit of $\omega$. Show that the fortune after $n$ rounds is $X_n(\omega)=2^n\mathbf 1_{[0,2^{-n})}(\omega)$ — a dyadic version of the spike of Example 4(a) — so $\int X_n=1$ and $X_n\to0$ a.e.
- Compute $\sup_nX_n$ and show that $\int\sup_nX_n=\infty$, so no integrable function dominates the fortunes.
- Simulate (seeded generator) $N=10^5$ gamblers for $n=1,\dots,30$ rounds and plot the sample mean of $X_n$ against $n$. Explain why it collapses to $0$ for $n$ beyond about $\log_2N$, even though every $\mathbb E[X_n]=1$.
- Now give the gambler a credit limit: she stops when her fortune reaches $2^K$. Show the new fortunes are dominated by $2^K$, and use the bounded convergence theorem to compute $\lim_n\mathbb E[X_n]$.
Tools: Python or a spreadsheet; Interactive 2 (spike mode) for the picture.
Hand in: The computations of steps 1, 2 and 4, the plot of step 3, and a paragraph explaining which hypothesis of the dominated convergence theorem the unlimited game violates.
Going further. Read about uniform integrability and Vitali's convergence theorem: on a probability space, $X_n\to X$ a.s. and $(X_n)$ uniformly integrable imply $\mathbb E[X_n]\to\mathbb E[X]$, and uniform integrability is also necessary for $L^1$ convergence. Check that the doubling fortunes are not uniformly integrable.
🧪 Self-check quiz
Q1. The Lebesgue integral $\int_{[0,1]}\mathbf 1_{\mathbb Q}\,dm$ is
Q2. Fatou's lemma states that for measurable $f_n\ge0$,
Q3. For $f_n=n\mathbf 1_{(0,1/n)}$ on $[0,1]$, $\int f_n=1$ but $\int\lim f_n=0$. Which hypothesis of the dominated convergence theorem fails?
Q4. The function $\frac{\sin x}{x}$ on $(0,\infty)$
Q5. $\displaystyle\lim_{n\to\infty}\int_0^\infty\Big(1+\frac xn\Big)^{-n}dx$ equals