3
$\begingroup$

The Riemann sum over an interval $[a,b]$ is usually defined as $$\lim\limits_{N\to\infty}\sum\limits_{k=0}^Nf\left(a+k\cdot\frac{b-a}{N}\right)\frac{b-a}{N}$$

Thus if we encounter a sum of the form $$\lim\limits_{N\to\infty}\sum\limits_{k=0}^Nf\left(k\cdot\frac{1}{N}\right)\frac{1}{N}$$ we can conclude that it is equal to an integral over the interval $[0,1]$. $$\lim\limits_{N\to\infty}\sum\limits_{k=0}^Nf\left(k\cdot\frac{1}{N}\right)\frac{1}{N}=\int_0^1f(x)dx\tag{1}\label{1}$$

What can we conclude about the following sum

$$\lim\limits_{N\to\infty}\lim\limits_{M\to\infty}\sum\limits_{k=0}^M f\left(k\cdot\frac{1}{N}\right)\frac{1}{N}\tag{2}\label{2}$$

To clarify, this is an infinite sum \eqref{2}, that differs from the Riemann sum \eqref{1}, in the upper limit of the sum. In the Riemann sum \eqref{1}, there is a relation between $M$ and $N$, namely $N=M$, while there is no such relation specified in \eqref{2}. If we can equate it to an integral, how are we to determine the limits of integration?

The equation \eqref{2} is to be taken, that the $M\to\infty$, we thus have an infinite sum (suppose it is convergent). Than we form a sequence of infinite sums, where $N$ increases for each element of the sequence. That is $$S_N=\lim\limits_{M\to\infty}\sum\limits_{k=0}^M f\left(k\cdot\frac{1}{N}\right)\frac{1}{N}$$ What does this sequence tend to?

Is it true that (or when is it true) $$\lim\limits_{N\to\infty}S_N=\int_0^\infty f(x)dx$$

Also the general term in \eqref{2} is $C_k=f\left(k\cdot\frac{1}{N}\right)$. How does it behave in the limit, namely $$\lim\limits_{N\to \infty}\lim\limits_{M\to \infty}f\left(M\cdot\frac{1}{N}\right)$$

$\endgroup$
  • $\begingroup$ Your sum $(2)$ is not well defined for $M>N.$ $\endgroup$ – zhw. Feb 4 '17 at 20:36
  • $\begingroup$ All derivations (I encountered) of the inverse fourier transform as a limiting fourier series use a sum of this form. I have similar objections as you, but seeing how so many texts use it, I though that perhaps I don't understand something. Looking at my question here I recreate the argument from those texts, where this sum shows up. I asked a separate question to seperate it from the fourier environment. $\endgroup$ – LeastSquaresWonderer Feb 4 '17 at 20:44
  • 1
    $\begingroup$ If $f(x)$ is defined on $[0,\infty)$ then $\sum_{k=0}^\infty f(k/N) \frac{1}{N}$ is an approximation for $\int_0^\infty f(x){\rm d}x$ with step-size $\frac{1}{N}$. $\endgroup$ – Winther Feb 4 '17 at 20:44
  • $\begingroup$ Could you please derive this for me? It would also resolve my unanswered question here. I have been thinking about this for few days, and I would really be grateful if you can derive this result for me. $\endgroup$ – LeastSquaresWonderer Feb 4 '17 at 20:46
  • $\begingroup$ So when we take the limit as N--> infinity, the approximation becomes the integral? I would really be grateful for a derivation and justification of the statement. $\endgroup$ – LeastSquaresWonderer Feb 4 '17 at 20:49
2
$\begingroup$

Suppose $f$ is Riemann integrable on $[0,b]$ for every $b > 0$ and the improper integral over $[0, \infty)$ is convergent.

We first consider the case where $f$ is nonnegative and non-increasing, as suggested by @Winther, where we have

$$\frac{f((k+1)/N)}{N} \leqslant \int_{k/N}^{(k+1)/N} f(x) \, dx \leqslant \frac{f(k/N)}{N}. $$

This implies

$$\int_0^{(M+1)/N} f(x) \, dx \leqslant \frac{1}{N} \sum_{k=0}^{M} f(k/N) \leqslant \frac{f(0)}{N} + \int_0^{M/N} f(x) \, dx. $$

The sequence of partial sums is increasing and bounded , hence convergent as $M \to \infty$, with

$$\int_0^{\infty} f(x) \, dx \leqslant\frac{1}{N} \sum_{k=0}^{\infty} f(k/N) \leqslant \frac{f(0)}{N} + \int_0^{\infty} f(x) \, dx. $$

Therefore,

$$\lim_{N \to \infty} \lim_{ M \to \infty}\frac{1}{N} \sum_{k=0}^{M} f(k/N) = \int_0^\infty f(x) \, dx.$$

Can this still hold if $f$ is not monotonic?

For example, consider $f(x) = \sin x /x$, where

$$\int_0^\infty \frac{\sin x}{x} \, dx = \frac{\pi}{2}.$$

Examining the corresponding series (WLOG starting with $k=1$) we find

$$\frac{1}{N}\sum_{k = 1}^{\infty} \frac{\sin (k/N)}{k/N} = \sum_{k = 1}^{\infty} \frac{\sin (k/N)}{k} \\ = \frac{\pi}{2}-\frac{1}{2N} \\ \longrightarrow_{N \to \infty} \frac{\pi}{2}.$$

I have not yet found a counterexample for a non-monotone function. As the integral test can be generalized to $C^1$ functions of bounded variation, I suspect this may characterize a wider class of functions for which this result holds.

This could be shown by considering

$$\left|\int_{k/N}^{(k+1)/N} f(x) \, dx - \frac{f(k/N)}{N} \right| \leqslant \int_{k/N}^{(k+1)/N} |f(x) - f(k/N)| \, dx $$

and then summing over $k$, applying the mean value theorem when $f$ is differentiable, and using $\int_0^\infty |f'(x)|\, dx < \infty$ to show that the sum converges to the integral as $N \to \infty.$

$\endgroup$
  • $\begingroup$ Very nice answer, thank you. If I may ask, does the order you specify the limits in matter? First M-->\infty, than N-->\infty ? The question relates to k/N. Becase k-->M, and k/N-->\infty (because the count goes to infinity, hence the upper limit of integration),this implies that k "overgrows" N, hence M does too. I hope my wording is understandable. $\endgroup$ – LeastSquaresWonderer Feb 5 '17 at 9:13
  • $\begingroup$ You're welcome. The order does matter. Here we take the limit as $M \to \infty$ first followed by the limit as $N \to \infty$. If we switch, then with $M$ fixed we have $\frac{1}{N} \sum_{k=0}^M f(k/N) \to 0$ as $N \to \infty$ if for example $f$ is continuous and $f(0)$ is finite. $\endgroup$ – RRL Feb 5 '17 at 9:26
  • $\begingroup$ This was my hypothesis, and the original reason for thinking about this question. Thank you for helping me understand. $\endgroup$ – LeastSquaresWonderer Feb 5 '17 at 9:31

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.