Lebesgue on Riemann Integration

When is a bounded, real-valued function f:[a,b]\rightarrow\mathbb{R} Riemann integrable? I ask this question for a few reasons. The first is because I assigned a couple exercises on the my students’ take-home final exam introducing ‘measure zero’ and the types of discontinuities of functions. The second is that I was reading the statement of Weyl equidistribution theorem, where the condition is Riemann integrable, as opposed to Lebesgue integrable. In fact, the theorem is false if you use the latter. The third and last is that I have known the answer for some time, but I have never worked through the proof in full detail. Well, I plan to change that with today’s post.

Theorem 1. (Lebesgue) A bounded function f: [a,b]\rightarrow\mathbb{R} is Riemann integrable if and only if the set of discontinuities of f is of measure zero.

It turns out that a bounded function is Riemann integrable if and only if it is continuous and its discontinuities form a set of measure zero. The French mathematician Henri Lebesgue proved this result over a century ago. It is important to not confuse this ‘Lebesgue’s theorem’ with other results about the Lebesgue integral, a more powerful integral based on measure theory. However, the limitations of the Riemann integral shown by Lebesgue motivate the study of his integral.

Indeed, we would think that a function zero almost everywhere (a.e.) would be Riemann integrable with integral zero. This is not the case. Consider the Dirichlet function

\displaystyle f(x):=\begin{cases}1&{x\in\mathbb{Q}}\\ 0&{x\in\mathbb{R}\setminus\mathbb{Q}}\end{cases}

which is zero a.e. However, f is not Riemann integrable on any compact interval [a,b]. Given any partition P=\left\{a=x_{0}<x_{1}<\cdots<x_{n}=b\right\}, the upper Riemann sum

\displaystyle\overline{S}(f;P):=\sum_{j=1}^{n}\underbrace{\sup_{t\in[x_{j-1},x_{j}]}f(t)}_{M_{j}}\underbrace{(x_{j}-x_{j-1})}_{\Delta x_{j}}

is 1 while the lower Riemann sum

\displaystyle\underline{S}(f;P):=\sum_{j=1}^{n}\underbrace{\inf_{t\in[x_{j-1},x_{j}]}f(t)}_{m_{j}}\underbrace{(x_{j}-x_{j-1})}_{\Delta x_{j}}

is 0 by the density of the rationals in the reals.

Enough with the motivation, let’s get down to the details. To prove Lebesgue’s theorem, we will need to define the oscillation of a bounded function f at a point x, which quantifies how much f varies in an arbitrarily small interval about x. Precisely, the oscillation of f:[a,b]\rightarrow\mathbb{C} at the point x\in[a,b] is 

\displaystyle\text{osc}(f;x):=\lim_{h\rightarrow0^{+}}\sup_{x',x''\in(x-h,x+h)}\left|f(x')-f(x'')\right|

Note that the supremum is well-defined since f is bounded. The oscillation of f at a point x is a useful notion, since it compactly encodes the continuity of a function.

Lemma 2. f:[a,b]\rightarrow\mathbb{C} is continuous at a point x\in[a,b] if and only if \text{osc}(f;x)=0.

\displaystyle\forall x'\in(x-\delta,x+\delta) \ \left|f(x)-f(x')\right|<\dfrac{\varepsilon}{2},

which implies by the triangle inequality that

\displaystyle\forall x',x''\in(x-\delta,x+\delta) \ \left|f(x')-f(x'')\right|<\varepsilon\Longrightarrow\sup_{x',x''\in(x-\delta,x+\delta)}\left|f(x')-f(x'')\right|\leq\varepsilon

Letting \delta\rightarrow0, we see that \text{osc}(f;p)\leq\varepsilon. Since \varepsilon was arbitrary, we conclude that \text{osc}(f;p)=0.

The other direction of the proof essentially follows from retracing our steps above, and I leave it as an exercise to the reader. \Box

Lemma 3. Let f: [a,b]\rightarrow\mathbb{C} be a bounded function. Then for any s>0,

\displaystyle A_{s}:=\left\{x\in[a,b]:\text{osc}(f;x)\geq s\right\}

is compact.

Proof. Since [a,b] is compact, it suffices to show that A_{s} is a closed subset of [a,b], or A_{s}^{c} is open in [a,b]. Let x_{0}\in A_{s}^{c}, so that \text{osc}(f;x_{0}) =: t <s. Set \varepsilon:=\frac{s-t}{2}. By definition of oscillation, there exists a constant \delta=\delta(\epsilon)>0 such that 0<h\leq\delta implies that

\displaystyle\left|t-\sup_{x',x''\in (x_{0}-h,x_{0}+h)}\left|f(x')-f(x'')\right|\right|<\varepsilon

By the triangle inequality,

\displaystyle\sup_{x',x''\in (x_{0}-\delta,x_{0}+\delta)}\left|f(x')-f(x'')\right|+t<\frac{s-t}{2}+t=\frac{s+t}{2}<s,

So, for any y_{0}\in (x_{0}-\delta,x_{0}+\delta), we can take h>0 sufficiently small so that (y_{0}-h,y_{0}+h)\subset(x_{0}-\delta,x_{0}+\delta) and therefore

\displaystyle\text{osc}(f;y_{0})=\lim_{h\rightarrow0^{+}}\sup_{x',x''\in (y_{0}-h,y_{0}+h)}\left|f(x')-f(x'')\right|\leq\frac{s+t}{2}<s

Equivalently, (x_{0}-\delta,x_{0}+\delta)\subset A_{s}^{c}. \Box

Let M and m be the supremum and infimum of f, respectively, on [a,b]. Denote the set of discontinuities of f by A, and note that 

\displaystyle A:=\left\{x\in[a,b]:\text{osc}(f;x)>0\right\},

which can be represented as the countable union

\displaystyle A=\bigcup_{k=1}^{\infty}A_{k},\indent A_{k}:=\left\{x\in[a,b]:\text{osc}(f;x)\geq\frac{1}{k}\right\}

First, suppose that f is Riemann integrable. Then, for any \varepsilon>0 given, there exists a partition

\displaystyle P:=\left\{a=x_{0}<x_{1}<\cdots<x_{n}=b\right\}

such that

\displaystyle\overline{S}(f;{P})-\underline{S}(f;{P})=\sum_{j=1}^{n}\left[M_{j}-m_{j}\right]\underbrace{(x_{j-1}-x_{j})}_{\Delta x_{j}}<\dfrac{\varepsilon}{k}

If x\in(x_{j-1},x_{j})\cap A_{k}, for some j\in\left\{1,\cdots,n\right\} and k\in\mathbb{N}, then

\displaystyle(M_{j}-m_{j})\Delta x_{j}\geq\dfrac{1}{k}\Delta x_{j}

Hence, if I=\left\{1\leq j\leq n: A_{k}\cap(x_{j-1},x_{j})\neq\emptyset\right\}, then

\displaystyle\dfrac{\varepsilon}{k}>\sum_{j=1}^{n}[M_{j}-m_{j}]\Delta x_{j}\geq\sum_{j\in I}[M_{j}-m_{j}]\Delta x_{j}\geq\dfrac{1}{k}\sum_{j\in I}\Delta x_{j}

Multiplying both sides of the inequality by k, we see that \sum_{j\in I}\Delta x_{j}<\varepsilon. The intervals (x_{j-1},x_{j}), for j\in I, form an open cover of A_{k}\setminus\left\{x_{0},\cdots,x_{n}\right\}, which is equal in measure to A_{k}. Hence,

\displaystyle\left|A_{k}\right|\leq\sum_{j\in I}\Delta x_{j}<\varepsilon,

which implies that \left|A_{k}\right|=0. Since the countable union of null sets is again a null set, we conclude that \left|A\right|=0.

Now suppose that \left|A\right|=0. Since A_{s}\subset A, we see that \left|A_{s}\right|=0, for all s>0. Hence, for \varepsilon>0 be given, there exists a countable collection of open intervals \left\{I_{k}\right\}_{k=1}^{\infty} such that

\displaystyle A_{\frac{\varepsilon}{2(b-a)}}\subset\bigcup_{k=1}^{\infty}I_{k},\indent\sum_{k=1}^{\infty}\left|I_{k}\right|<\dfrac{\varepsilon}{2(M-m)}

The preceding lemma implies that A_{\varepsilon} is compact, hence there exists a finite subcover \left\{I_{1},\cdots,I_{N}\right\} of A_{\varepsilon}. If x\in[a,b]\setminus\bigcup_{j=1}^{N}I_{j}, then \text{osc}(f;x)<\varepsilon. So for such an x, there exists \delta_{x}>0 such that for all x',x''\in (x-\delta_{x},x+\delta_{x}), \left|f(x')-f(x'')\right|<\frac{\varepsilon}{2(b-a)}. The collection of intervals

\displaystyle\left\{(x-\delta_{x},x+\delta_{x}): x\in [a,b]\setminus\bigcup_{j=1}^{N}I_{j}\right\}

forms an open cover of [a,b]\setminus\bigcup_{j=1}^{N}I_{j}, which is compact. Hence, there exists a finite subcover

\displaystyle\left\{(\tilde{x}_{1}-\delta_{1},\tilde{x}_{1}+\delta_{1}),\cdots,(\tilde{x}_{k}-\delta_{k},\tilde{x}_{k}+\delta_{k})\right\}

This collection of intervals together with the intervals I_{1},\cdots,I_{N} form an open cover of [a,b]. We want to use this open cover to obtain a partition P of [a,b] such that \overline{S}(f;P)-\underline{S}(f;P)<\varepsilon. In particular, we want the partition a=x_{0}<x_{1}<\cdots<x_{n}=b to have the property that any interval [x_{j-1},x_{j}] is contained entirely in an element of the open cover, for which we will need Lebesgue’s number lemma. We postpone the proof of the lemma until the end in favor of completing the proof of Theorem 1.

Lemma 4. (Lebesgue) If (X,d) is a compact metric space and \mathcal{A} is an open cover of X, then there exists a \delta>0 such that, for every subset C\subset X with \text{diam } C<\delta, there exists some A\in\mathcal{A} such that C\subset A.

Divide P into two sets P_{1} and P_{2}, where

\displaystyle P_{1}:=\left\{x_{j}\in P:[x_{j-1},x_{j}]\subset I_{k}\text{ for some }k_{j}\right\},\indent P_{2}:=P\setminus C_{1}

For points in P_{1}, the oscillation can be ‘big’, but the length of the interval [x_{j-1}-x_{j}] is very small. For points in P_{2}, the oscillation is sufficiently ‘small’, and the total length of the intervals is at most b-a.

\begin{array}{lcl}\displaystyle\overline{S}(f;P)-\underline{S}(f;P)&=&\displaystyle\sum_{x_{j}\in C_{1}}[M_{j}-m_{j}]\Delta x_{j}+\sum_{x_{j}\in C_{2}}[M_{j}-m_{j}]\Delta x_{j}\\[.9 em]&\leq&\displaystyle(M-m)\sum_{x_{j}\in C_{1}}\left|I_{k_{j}}\right|+\dfrac{\varepsilon}{2(b-a)}\sum_{x_{j}\in C_{2}}\Delta x_{j}\\[.9 em]&\leq&\displaystyle(M-m)\cdot\dfrac{\varepsilon}{2(M-m)}+\dfrac{\varepsilon}{2(b-a)}\cdot(b-a)\\[.9 em]&=&\displaystyle\varepsilon\end{array}

Proof of Lebesgue’s Number Lemma. We may assume that X\notin\mathcal{A}, otherwise any \delta>0 works. Since X is compact, there exist finitely many elements A_{1},\cdots,A_{n} in \mathcal{A} such that X\subset\bigcup_{j=1}^{n}A_{j}. For each j=1,\cdots,n, define B_{j}:=X\setminus A_{j}. Define a real-valued function f on X by

\displaystyle f(x):=\dfrac{1}{n}\sum_{j=1}^{n}d(x,B_{j}),\indent\forall x\in X

I claim that f(x)>0 for all x\in X. Indeed, if x\in A_{j}, then we can take \varepsilon>0 sufficiently small so that the open ball B(x;\varepsilon)\subset A_{j}. Hence, d(x,B_{j})\geq\varepsilon and f(x)\geq\frac{\varepsilon}{n}.

f is continuous and X is compact, so by the extreme value theorem, f attains its minimum \delta>0 on X. Suppose we are given a set C\subset X such that \text{diam }C<\delta. For any point x_{0}\in C, the ball B(x_{0};\delta) contains the set C, by definition of diameter. If m denotes the index of a set satisfying d(x_{0},B_{m})=\max\left\{d(x_{0},B_{j}):1\leq j\leq n\right\}, then

\displaystyle\delta\leq f(x_{0})=\dfrac{1}{n}\sum_{j=1}^{n}d(x_{0},B_{j})\leq d(x_{0},B_{m}),

which implies that C\subset B(x_{0};\delta)\subset X\setminus B_{m}=A_{m}. \Box

Advertisements
This entry was posted in math.CA, math.GM, math.NT and tagged , , , , . Bookmark the permalink.

2 Responses to Lebesgue on Riemann Integration

  1. alice says:

    Dude.. I am not much into reading, but somehow I got to read lots of articles on your blog. Its amazing how interesting it is for me to visit you very often. –

  2. Where ever you go, I’ll stick to…Congrats!!!!

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s