I took the final of my first analysis course and I saw the following problem:

Let $f: [0,1] \rightarrow \Bbb{R}$ be a continuously differentiable function. Show that $$ \lim_{n\rightarrow \infty} \bigg[ \sum_{k = 1}^n f\bigg(\frac{k}{n}\bigg) -n\int_0^1 f(x)\, dx \bigg] = \frac{f(1)-f(0)}{2}$$


I was totally clueless. I tried putting $f(x) = x$ and see if that helps me to come up with a proof. Sitting over there for an hour or so, I suspect the formula is somehow related to the trapezoidal rule. But I still left that question blank. Now I continue doing this exam problem:

For each $n \in \Bbb{N}^+$, let $$T_n = \frac{f(0)+f(\frac{1}{n})}{2n} + \frac{f(\frac{1}{n})+f(\frac{2}{n})}{2n}+\cdots +\frac{f(\frac{n-1}{n})+f(1)}{2n} = \frac{f(0)-f(1)}{2n} +\frac{1}{n}\sum_{k=1}^n f\bigg(\frac{k}{n}\bigg)$$

Then we are acutally proving that

$$n\bigg( \underbrace{T_n -\int_0^1f(x)\,dx }_{E_n} \bigg) \rightarrow 0$$

I have looked at some documents about the error term in trapezoidal rule, we can rewrite $E_n$ using integration by parts. Let's write $a_k = \frac{k-1}{n}$ and $b_k = \frac{k}{n}$ and $c_k = \frac{a_k + b_k}{2}$

$$E_n = \sum_{k =1 }^n \bigg(\frac{f(\frac{k-1}{n})+f(\frac{k}{n})}{2}-\int_{\frac{k-1}{n}}^{\frac{k}{n}} f(x)\,dx \bigg)=\sum_{k=1}^n \bigg[ \int_{a_k}^{b_k} f'(x)(x-c_k)\,dx \bigg]$$

But we could not apply integration by parts again since we just assumed $f \in C^1[0,1]$ only, not $C^2$...I can just do the following now. For $f'$ is continuous on $[0,1]$, $|f|$ is bounded above by some number $M$. Then

$$ \begin{align}|E_n| & \leq \sum_{k=1}^n \bigg| \int_{a_k}^{b_k} f'(x)(x-c_k)\,dx \bigg|\leq \sum_{k=1}^n \bigg(\int_{a_k}^{b_k} |f'(x)||x-c_k|\,dx\bigg) \\ &\leq \sum_{k=1}^n \bigg( M\int_{a_k}^{b_k}|x-c_k|\,dx \bigg)= n \cdot \frac{M}{4n^2} =\frac{M}{4n} \end{align} $$

From this, we can only conclude that $|nE_n| = n|E_n| \leq M/4$ for each $n$, which is not what we want.

I believe that proposition is true since it holds for some functions in $C^1$ but not $C^2$, such as $x^{3/2}$ and $|x|^{2.1}\sin(1/|x|)$. So that the $C^1$ assumption should be sufficient.

Perhaps we shall use the continuity of $f'$ more fully in order to prove the proposition. For example, in the $\epsilon$-$\delta$ definition, I can take $N$ so big such that $1/N < \delta$, but don't know how to prove $|NE_N| <\epsilon'$ for this big $N$.

  • $\begingroup$ check EULER MACLAURIN in wikipedia $\endgroup$ – Jose Garcia May 31 '17 at 10:57
  • $\begingroup$ I know that is a powerful connection between sum and integrals just now. But I saw the remainder term is again a integral in the form $\int f'(x)P_1(x)\,dx$, where $P_1$ is some linear polynomial similar to what I have above. I think I have to approximate for that, but how? $\endgroup$ – Li Chun Min May 31 '17 at 11:10

Yes, it's the trapezoidal rule in action. To get the desired convergence and not only the boundedness, note that

$$\int_{a_k}^{b_k} f'(c_k)(x-c_k)\,dx = 0.$$

Hence you can estimate

$$\Biggl\lvert \int_{a_k}^{b_k} f'(x)(x-c_k)\,dx\Biggr\rvert \leqslant \int_{a_k}^{b_k} \lvert f'(x) - f'(c_k)\rvert\cdot \lvert x - c_k\rvert\,dx \leqslant \omega_{f'}\bigl(\tfrac{1}{2n}\bigr) \int_{a_k}^{b_k} \lvert x-c_k\rvert\,dx,$$

where $\omega_{f'}$ is the modulus of continuity of $f'$,

$$\omega_{f'}(\delta) = \max \:\{\lvert f'(x) - f'(y)\rvert : \lvert x-y\rvert \leqslant \delta\}.$$

Since $f'$ is uniformly continuous, we have $\lim\limits_{\delta \to 0} \omega_{f'}(\delta) = 0$, as needed, leading to

$$\Biggl\lvert \sum_{k = 1}^n f\biggl(\frac{k}{n}\biggr) - n\int_0^1 f(x)\,dx - \frac{f(1) - f(0)}{2}\Biggr\rvert \leqslant K\cdot \omega_{f'}\biggl(\frac{1}{2n}\biggr).$$

  • $\begingroup$ Thanks! that makes sense! Again I know the direction didn't know how to complete the details (sigh $\endgroup$ – Li Chun Min May 31 '17 at 12:01

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.