Let $f:[a,b]\to \mathbb{R}$ be a monotone function (say strictly increasing).

Then, do for every $\epsilon>0$ exist two step functions $h,g$ so that $g\le f\le h$ and $0\le h-g\le \epsilon$?

Does there exist some closed form of these functions like the one below?

I encountered the problem when trying to prove the Riemann Integrability of monotone functions via the traditional definition of the Riemann Integral (not the one with Darboux sums). The book I am reading proves that a continuous function is integrable via the process below:

Let $\mathcal{P}=\left\{ a=x_0<...<x_n=b \right\}$ be a partition of $[a,b]$. Define $m_i=\min_{x\in [x_{i-1},x_i]}f(x)$ and $M_i=\max_{x\in [x_{i-1},x_i]}f(x)$. By the Extreme Value theorem $M_i,m_i$ are well defined. We approximate $f$ with step functions: \begin{gather}g=m_1\chi_{[x_0,x_1]}+\sum_{i=2}^nm_i\chi_{(x_{i-1},x_i]}\\ h=M_1\chi_{[x_0,x_1]}+\sum_{i=2}^nM_i\chi_{(x_{i-1},x_i]}\end{gather} It is easily seen that they satisfy the "step function approximation" and as $0\le \int_a^bh-g\le \epsilon$ (uniform continuity) by a previous theorem $f$ is integrable.

The book then goes on to generalise by discussing regulated functions. I would like however to see a self contained proof similar to the previous one if $f$ is monotone. This question is reduced to the two questions asked in the beggining of the post


Let $\varepsilon>0$ and $N_\varepsilon$ the smallest $n \in \mathbb{N}$ such that $$ \frac{1}{n}(b-a)(f(b)-f(a)) \le \varepsilon $$ For $n \ge \max\{2,N_\varepsilon\}$ consider the following partition of $[a,b]$: $$ \mathcal{P}=\{a=x_0<x_1<\ldots<x_n=b\}, \ x_i=a+i\frac{b-a}{n}\ 0 \le i \le n. $$ Set $$ A_i=\begin{cases} [x_0,x_1] & \text{ for } i=0\\ (x_i,x_{i+1}]& \text{ for } 1 \le i \le n-1 \end{cases} $$ Since $f$ is strictly increasing for each $i \in \{0,\ldots,n-1\}$ we have $$ f(x_i) \le f(x) \le f(x_{i+1}) \quad \forall\ x_i \le x \le x_{i+1}. $$ Setting $$ h=\sum_{i=0}^{n-1}f(x_{i+1})\chi_{A_i},\ g=\sum_{i=0}^{n-1}f(x_i)\chi_{A_i} $$ we have $$ g(x)\le f(x) \le h(x) \quad \forall\ x \in [a,b], $$ and $$ h(x)-g(x)=\sum_{i=0}^{n-1}\Big(f(x_{i+1})-f(x_i)\Big)\chi_{A_i}(x)>0 \quad \forall\ x \in [a,b]. $$ In addition \begin{eqnarray} \int_a^b(h-g)&=&\sum_{i=0}^{n-1}(f(x_{i+1})-f(x_i))\int_a^b\chi_{A_i}=\sum_{i=0}^{n-1}(f(x_{i+1})-f(x_i))(x_{i+1}-x_i)\\ &=&\frac{b-a}{n}\sum_{i=0}^{n-1}(f(x_{i+1})-f(x_i))=\frac{1}{n}(b-a)(f(b)-f(a))\\ &\le&\frac{1}{N_\varepsilon}(b-a)(f(b)-f(a))\le \varepsilon. \end{eqnarray}

  • 1
    $\begingroup$ You didn't have to include so many details but thank you sincerely $\endgroup$ – Optional Jan 22 '13 at 11:19

Let $P$ be a finite partition of $[f(a),f(b)]$ into subintervals of length at most $\epsilon$. For $A\in P$, write $m_A<M_A$ for the endpoints of $A$ and define $$g=\sum_{A\in P}m_A\chi_{f^{-1}(A)}\qquad\mbox{and}\qquad h=\sum_{A\in P} M_A\chi_{f^{-1}(A)}.$$

  • $\begingroup$ Thanks for the answer. Can you suggest such a partition $P$ (that is preferably simple)? $\endgroup$ – Optional Jan 21 '13 at 19:19
  • $\begingroup$ Sure. Pick $n\ge1$ large enough that $\delta:=(f(b)-f(a))/n\le \epsilon$, and take $P=\{A_1,A_2,\dots,A_n\}$ where $A_1=[f(a),f(a)+\delta)$, $A_2=[f(a)+\delta,f(a)+2\delta)$, ... , $A_{n-1}=[f(a)+(n-2)\delta,f(a)+(n-1)\delta)$, $A_n=[f(a)+(n-1)\delta,f(a)+n\delta]$. $\endgroup$ – user108903 Jan 21 '13 at 19:23
  • $\begingroup$ Yeah, thats what I had in mind. The whole process of partitioning the range instead of the domain is a little complicated. Can we partition the domain instead? $\endgroup$ – Optional Jan 21 '13 at 19:28
  • $\begingroup$ Well you are partitioning the domain (even into intervals), but you're using a partition of the range to find the partition of the domain. $\endgroup$ – user108903 Jan 21 '13 at 19:29
  • $\begingroup$ But how do we know $f^{-1}(A)$ is an interval and so the two functions are integrable? $\endgroup$ – Optional Jan 21 '13 at 19:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.