Let $f:\left[a,b\right]\to\mathbb{R}$ be a function that is derivative so that $f'$ is continuous then $$ \lim_{n\to\infty}\int\limits _{a}^{b}\sin\left(nt\right)f\left(t\right)dt=0 $$

My attempt: I want to show that $\forall\epsilon>0 \ \exists \ n_0 \in \mathbb{N}$ so that $\forall n>n_0 \ \ \ \left|\int\limits _{a}^{b}\sin\left(nt\right)f\left(t\right)dt\right|<\epsilon$. Let there be $\epsilon>0$ because $f$ and $\sin$ are derivative and $f'$ is continuous we can use integration by parts so that $$ \left|\int\limits _{a}^{b}f'\left(t\right)\sin\left(nt\right)dt\right|=\left|\left[f\left(t\right)\sin\left(nt\right)\right]_{a}^{b}-n\int\limits _{a}^{b}f\left(t\right)\cos\left(nt\right)\right|\leq\left|\left[f\left(t\right)\sin\left(nt\right)\right]_{a}^{b}\right|+\left|n\int\limits _{a}^{b}f\left(t\right)\cos\left(nt\right)\right| $$

That is my main idea. any suggestions?

  • $\begingroup$ Integrate the original by parts. $\endgroup$ May 9 '13 at 22:44
  • $\begingroup$ For something that is nearly identical, check out the Riemann-Lebesgue lemma. Essentially what this is saying is that you can cover $[a,b]$ with intervals on which $f(t)\sin(nt)$ goes through an entire period, and hence integrates to zero. As $n$ gets larger these periods get smaller, and so the "left over" part that you're not covering is a smaller and smaller set. Its not hard to see what happens in the limit. $\endgroup$ May 9 '13 at 22:46
  • $\begingroup$ Deja-vu I asked it yesterday :P math.stackexchange.com/questions/386273/… $\endgroup$ May 10 '13 at 0:18

Good idea. You just did the wrong integration by parts. Instead $$ \int_a^b\sin(nt)f(t)dt=\frac{-\cos(nt)f(t)}{n}\Big|_a^b+\frac{1}{n}\int_a^b\cos(nt)f'(t)dt. $$ Now it is easy to see why the rhs tends to $0$.

Note: this follows from Riemann-Lebesgue lemma as soon as $f$ is Lebesgue integrable over the interval. But actually, a convenient proof of the latter goes by proving it for $C^1$ functions and then using their density in $L^1$. Riemann-Lebesgue holds over unbounded intervals also. Then one can consider compactly supported $C^1$ functions.

Fourier coefficients: if you integrate over an interval of length $2\pi$, you get a Fourier coefficient of $f$. So in general, when $f$ is integrable, the Fourier coefficients tend to $0$. But what you see here is that when $f$ is $C^1$, the coefficients are $o\big(\frac{1}{n}\big)$. And going on like this, the more regular the function, the faster the coefficients converge to $0$. If $f$ is $C^k$, the coefficients are $o\big(\frac{1}{n^k} \big)$.


For future references, this is a special case of the Riemann Lebesgue Lemma (well, this is the Riemann Lebesgue Lemma, really). It states:

T Let $f:[a,b]\to\Bbb R$ be Riemann integrable over $[a,b]$. Then $$\lim_{\lambda\to\infty}\int_a^b f(t)\sin\lambda t dt=0$$

P First, observe this is true for constant functions, since $$\int_a^b K\sin\lambda tdt=K\frac{-\cos(\lambda b)+\cos(\lambda a)}{\lambda}\to 0$$

It follows that it is also true for a step function, that is, a piecewise constant function, say $$s(x)=\sum_{i=1}^n s_i{\bf 1}_{X_i}$$ where the $X_i=[x_{i-1},x_i]$ are subintervals of $[a,b]$, since $$\int_a^b s(t) \sin\lambda tdt=\sum_{i=1}^n s_i\int_{x_{i-1}}^{x_i} \sin\lambda tdt$$

Let $\epsilon >0$ be given. Since $f$ is Riemann integrable there exists a step function $s\leq f$ defined on $[a,b]$ such that $$\int_a^b f-\int_a^b s<\frac{\epsilon}2$$

Since the lemma is true for step functions, there exists $M$ such that $\lambda>M$ implies $$\left| {\int_a^b {s\left( t \right)\sin \lambda tdt} } \right|<\frac{\epsilon}2$$

But $$\begin{align} \left| {\int_a^b f (t)\sin \lambda tdt} \right| &= \left| {\int_a^b {\left( {f(t) - s\left( t \right) + s\left( t \right)} \right)\sin \lambda tdt} } \right| \\ &\leq \left| {\int_a^b {\left( {f(t) - s\left( t \right)} \right)\sin \lambda tdt} } \right| + \left| {\int_a^b {s\left( t \right)\sin \lambda tdt} } \right| \\ &\leq \int_a^b {\left| {f(t) - s\left( t \right)} \right|dt} + \left| {\int_a^b {s\left( t \right)\sin \lambda tdt} } \right| \\ &= \int_a^b {\left( {f(t) - s\left( t \right)} \right)dt} + \left| {\int_a^b {s\left( t \right)\sin \lambda tdt} } \right| \\ &< \frac{\epsilon }{2} + \frac{\epsilon }{2} = \epsilon \end{align} $$

so the theorem is proven. $\blacktriangle$.

  • $\begingroup$ Note that $\lambda\to\infty$ is taken as a "limit in $\Bbb R$", not as a sequential limit, i.e. "$n\to\infty$". $\endgroup$
    – Pedro Tamaroff
    May 9 '13 at 23:20
  • 1
    $\begingroup$ Since $f$ is differentiable, this can be proven more easily than proving the Riemann-Lebesgue Lemma. Nice proof however. $\endgroup$
    – robjohn
    May 9 '13 at 23:27
  • $\begingroup$ @robjohn Thanks. Yes, the point was that this was true in more generality, and maybe the OP will bump into this again. $\endgroup$
    – Pedro Tamaroff
    May 9 '13 at 23:28
  • $\begingroup$ Note however that a classic proof goes by doing the $C^1$ case first, and then obtain the result by $L^1$ density of the latter. What you did is another classic argument, by density of step functions. $\endgroup$
    – Julien
    May 9 '13 at 23:30

You're integrating the wrong thing by parts. Use the technique on the actual expression you want the limit of, differentiating $f$ and integrating $\sin$ to get a $1/n$ which is small.

Then have a go at bounding the thing it multiplies by something finite. A very crude estimate will do!


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.