2

In my research, I have reached a point where I need to find a nonnegative function which satisfies the following properties:

  1. It's an even function:

$$f(-\theta) = f(\theta)$$

  1. It's normalised as:

$$\int^\pi_0f(\theta)d\theta = \frac{\pi}{2}$$

  1. For any $\phi$ in the interval $[0,\pi]$,the following property is satisfied:

$$\int^\pi_0[f(\theta)+f(\theta - \phi)-2f(\theta)f(\theta - \phi)]d\theta = \pi\sin^2\frac{\phi}{2}$$

I would be very much in your debt if you help me find such a function.

Arthur
  • 83

2 Answers2

4

This falls into the problem of computing convolutional square roots, modulo the flaws already outlined by Cameron Williams.

A more sensible problem would be; compute a function $f$ on the circle (that is, a $2\pi$ periodic function) that satisfies $f(\theta)=f(-\theta)$ and such that $$\tag{1} f\ast f (\phi) = \pi \cos^2\frac \phi 2.$$ Here $$ f\ast f(\phi):= \int_{-\pi}^\pi f(\phi-\theta)f(\theta)\,d\theta.$$ The normalisation condition is already encoded in (1); indeed, evaluating (1) at $\phi=0$, $$ f\ast f(0)= 2\int_0^\pi f^2(\theta)\, d\theta = \pi.$$

Convolution square roots have been discussed before on Math.StackExchange. (Here there is an answer of Julián Aguirre (who, unfortunately, recently died). In the comments to that question, I have linked a related problem, answered by Terry Tao. All of these links, however, describe the problem on $\mathbb R$. I doubt they can be useful to the OP).

  • Thank you very much, I will look into it thoroughly. – Arthur Feb 11 '21 at 18:51
  • 1
    Yeah, well, there’s not much in what I wrote. You should write $f$ as a cosine series. The convolution will turn into a product of the Fourier coefficients. Then notice that $\cos^2\phi/2=\frac12 -\frac12 \cos\phi$ and equate the coefficients. I think you’ll get your answer rather quickly. – Giuseppe Negro Feb 11 '21 at 18:57
  • @GiuseppeNegro, I think it's supposed to be $(1+\cos \phi)/2$, not $-$. Otherwise, I agree with your advice, which is in fact a method employed by the other answer, even though they went wrong somethere – Yuriy S Feb 11 '21 at 19:06
  • @YuriyS: oh yeah, thanks, that’s a typo. – Giuseppe Negro Feb 11 '21 at 19:07
  • Also, the normalization condition in the OP is different, it doesn't contain a square – Yuriy S Feb 11 '21 at 19:10
  • 2
    @YuriyS: I noticed that, thanks. My point is that the original problem is overdetermined, I should have been clearer. The equation (1) already forces a normalisation, with the square, which may or may not be compatible with the other normalisation requested by the OP. Unless there is a typo in the OP and there’s just a missing square there. – Giuseppe Negro Feb 11 '21 at 19:38
3

There are two functions that satisfy the updated conditions.

Let us start by writing $f(\theta)$ as a Fourier series involving only cosines (which are even functions):

\begin{equation} f(\theta)=\sum_{n=0}^{+\infty}f_n\cos(n\theta) \end{equation}

the first condition is immediately satisfied. Let us impose the second condition:

\begin{equation} \int_{0}^{\pi}d\theta \,\,f(\theta)=\sum_{n=0}^{+\infty}f_n\int_{0}^{\pi}\cos(n\theta)=\sum_{n=0}^{+\infty}f_n\frac{\sin{(\pi n)}}{n} \stackrel{!}{=}\frac{\pi}{2} \end{equation}

this expression vanishes except when $n=0$, in that case \begin{equation} \frac{\sin{(\pi n)}}{n}=\pi \end{equation} or, in other words, $f_0=\frac{1}{2}$. Hence,

\begin{equation} f(\theta)=\frac{1}{2}+\sum_{n=1}^{+\infty}f_n\cos(n\theta) \end{equation}

Now we have to look at condition 3. We have
\begin{equation} \int_{0}^{\pi}d\theta\,\,\cos\left(n(\theta-\phi)\right)=\frac{\sin(n\phi )}{n}[1-(-1)^n], \end{equation}

\begin{equation} \begin{split} \int_0^{\pi}d\theta\,\,\left(\frac{1}{2}+\sum_{n= 1}^{+\infty}f_n\cos(n\theta)\right)\left(\frac{1}{2}+\sum_{m=1}^{+\infty}f_m\cos(m(\theta-\phi))\right) \\ \\ =\frac{\pi}{4}+\frac{1}{2}\sum_{n=1}^{+\infty}\frac{\sin(n\phi )}{n}\left[1-(-1)^n\right] +\sum_{n,m\neq0}f_nf_m\int_0^{\pi}d\theta\,\,\cos(n\theta)\cos(m(\theta-\phi)) \end{split} \end{equation}


The last integral is \begin{equation} \int_0^{\pi}d\theta\,\,\cos(n\theta)\cos(m(\theta-\phi))=\frac{m}{m^2-n^2}[\sin(m\phi )-(-1)^n\sin(m(\phi-\pi))] \end{equation}


Let us study the limit $n\to m$ \begin{equation} \lim_{m\to n}\frac{m}{m^2-n^2}[\sin(m\phi )-(-1)^n\sin(m(\phi-\pi))]=\frac{\pi}{2}\cos\left(n\phi\right) \end{equation}


Hence, imposing condition 3 means that

\begin{equation} \frac{\pi}{2}+\frac{\pi}{2}+\sum_{n=1}^{\infty}f_n \frac{\sin(\phi n)}{n}[1-(-1)^n]-2\bigg\{\frac{\pi}{4}+\frac{1}{2}\sum_{n=1}^{+\infty}f_n\frac{\sin(n\phi)}{n}[1-(-1)^n]\\ +\frac{\pi}{2}\sum_{n=1}^{+\infty}f^2_n\cos(n\phi)+\sum_{n\neq m\neq 0}f_nf_m \frac{m}{m^2-n^2}\sin(m\phi )[1-(-1)^{n+m}]\bigg\}\stackrel{!}{=} \pi\sin^2\left(\frac{\phi}{2}\right) \end{equation}


Now the right hand side can be written as

\begin{equation} \pi\sin^2\left(\frac{\phi}{2}\right)=\frac{\pi}{2}(1-\cos\phi) \end{equation}

Therefore we have obtained the equation

\begin{equation} \begin{split} \pi\sum_{n=1}^{+\infty}f^2_n\cos(n\phi)+2\sum_{n\neq m\neq 0}f_nf_m \frac{m}{m^2-n^2}\sin(m\phi )[1-(-1)^{n+m}]=\frac{\pi}{2}\cos(\phi) \end{split} \end{equation}


We can satisfy this equation if $f^2_1\pi=\frac{\pi}{2}$ and if $f_{k}=0$ whenever $k>1$. This means $f_1=\pm \frac{1}{\sqrt{2}}$.

We conclude that the functions that you are looking for are

\begin{equation} f(\theta)=\frac{1}{2}\pm\frac{1}{\sqrt{2}}\cos\theta \end{equation}

If you want $f$ to be non negative, then just choose the plus sign.

  • (+1) for your systematic approach. – euler_med Feb 10 '21 at 12:20
  • 3
    I doubt that $f(\theta)$ is allowed to depend on $\phi$. – Noiralef Feb 10 '21 at 12:26
  • I interpreted $\phi$ as a constant, fixed term, because $f$ depends only on $\theta$ and $f(\theta-\phi)$ would not have sense if $\phi$ is a variable. – Ruben Campos Delgado Feb 10 '21 at 12:31
  • 2
    The question clearly says "for any $\phi$", so $\phi$ is an independent variable. – Javier Feb 10 '21 at 12:37
  • 1
    Thank you all for the engagement. I appreciate the effort very much. My problem though, is that $f$ cannot depend on $\phi$. And regarding your question about the paper, yes, it will lead to the publication of a paper. And I would be very glad to give you the acknowledgement you deserve if you helped me find the function :) – Arthur Feb 10 '21 at 13:49
  • @Arthur I think I have found your function. – Ruben Campos Delgado Feb 11 '21 at 19:45
  • @RubenCamposDelgado Unfortunately, this function doesn't work. I Think the reason is because you thought that the integral:

    $\int^\pi_0\cos(m(\theta-\phi))d\theta=0$

    but it's not :/

    – Arthur Feb 12 '21 at 00:02
  • The third condition doesn't work for your new function. We have $$\int_0^\pi \cos(\theta-\phi) d\theta=2\sin \phi$$ which is a term not present on the right hand side – Yuriy S Feb 12 '21 at 07:29
  • Yes, sorry, I see. Thank you for pointing this out. I try again – Ruben Campos Delgado Feb 12 '21 at 07:59
  • I tried again. I think such a function does not exist. Equating the Fourier coefficients leads to a contradiction, it is not possible to find Fourier coefficients such that the equation is satisfied. – Ruben Campos Delgado Feb 12 '21 at 13:35
  • 1
    I suspect that you are right. I have reached the same conclusion. In fact, if we extended the integral limits to $[-\pi,\pi]$, then as Giuseppe Negro posted in the other answer, the problem would be easily solved, and your previous solution would have been correct. You get my upvote for all your work – Yuriy S Feb 12 '21 at 14:00
  • Thank you for your effort, in fact a Greek physicist of mine has sent me some different calculations that might prove such a function doesn't exist. However, as a last hope, I can save my model if I find a function that has somewhat different properties, I will edit my question here, maybe we can get somewhere! – Arthur Feb 13 '21 at 00:50
  • @Arthur With the new condition the function does exist. I have modified my answer – Ruben Campos Delgado Feb 13 '21 at 10:07