2

Let $p, q \in \mathbb{N}$ (with $0 \in \mathbb{N}$), and consider the integral $$I_{p,q} := \int_0^{\pi} \int_0^{\pi} \frac{\sin(x) \sin(px) \cos(qy)}{\sin(x)^2 + \sin(y)^2} dx~dy.$$

For $p$ even or $q$ odd, a simple argument of anti-symmetry about $\frac{\pi}{2}$ shows that $I_{p,q} = 0$. For $p$ odd and $q$ even, $I_{p,q} \neq 0$.

However, numerical experiments suggest that for any $p_0, q_0 \in \mathbb{N}$, $$\lim\limits_{q\to\infty} I_{p_0,q} = 0 \qquad \lim\limits_{p\to\infty} I_{p,q_0} = 0$$ The plot below shows a numerical approximation (computed with MATLAB's integrate2) to $I_{p,q}$ as a function of $p$ and $q$.

<span class=$I_{p,q}$ as a function of $p$ and $q$" />

How can I convince myself that this is true?

Edit: I asked this question originally as "Is there a closed form for $I_{p,q}$?" As it has been a few weeks and so far that has proven too ambitious, I have relaxed it to the version above.

  • This comes from a reformulation of my previous (not yet answered, at the moment of writing this) question https://math.stackexchange.com/questions/3814366/this-finite-sum-involving-roots-of-unity-is-bounded-but-why – DashDotDashDot Sep 08 '20 at 21:28
  • 1
    Perhaps one can expand $\sin(px)$ and $\cos(qy)$ using Chebyshev polynomials but I doubt it would help much. Personally I would just resort to numerical methods on this one. It certainly doesn't help that the integrand contains two singularities. – K.defaoite Sep 08 '20 at 21:52
  • @DashDotDashDot For high values of $p,q$ the integrand becomes highly oscillatory and the numerical integration is unreliable. I wouldn't trust the numerical result without some theoretical corroboration. – PierreCarre Oct 21 '20 at 10:57
  • The expression "$\sin(x)^2$" means $(\sin x)^2$ or $\sin (x^2)$? – Alex Ravsky Oct 21 '20 at 15:32
  • @PierreCarre Absolutely! For high values of $p$ and $q$, the numerical results start becoming very noisy as one would expect -- hence the plot stopping at 50, and this question :) – DashDotDashDot Oct 21 '20 at 16:17
  • 1
    @AlexRavsky The former. – DashDotDashDot Oct 21 '20 at 16:18
  • 1
    I guess the required result can follow from the next fact, which is easy to show. Let $a<b$ be real numbers and $f:[a,b]\to\Bbb R$ be a continuous function. Then $\lim_{n\to\infty}\int_a^b f(t)\sin nt dt=\lim_{n\to\infty}\int_a^b f(t)\cos nt dt 0$. Moreover, I guess if $a=0$, $b=\pi$ and $f(a)=f(b)$ then the bounds for the magnitude of the members of the sequences ${\int_a^b f(t)\sin nt dt}$ and ${\int_a^b f(t)\cos nt dt}$ are similar to those for magnitude of Fourier coefficients of $f$. – Alex Ravsky Oct 21 '20 at 16:35

1 Answers1

5

Proposition. For any natural number $p$ we have $I_{p,\infty}=0$.

Proof. As $I_{p,q}$ is even we have $$I_{p,q}=\frac14\int_{-\pi}^\pi\int_{-\pi}^\pi\frac{\sin x\sin px\cos qy}{\sin^2x+\sin^2y}\,dx\,dy\tag1.$$ Define $a:=\sin^2y$. Let $w=e^{ix}$ and $C=\{w:|w|=1\}$ so that \begin{align}\int_{-\pi}^\pi\frac{\sin x\sin px}{\sin^2x+\sin^2y}\,dx&=\oint_C\frac{\frac{w-1/w}{2i}\cdot\frac{w^p-1/w^p}{2i}}{\frac{(w-1/w)^2}{-4}+a}\frac{dw}{iw}=-i\oint_C\frac{(w^2-1)(w^{2p}-1)}{w^p(w^4-(4a+2)w^2+1)}\,dw.\end{align} Inside $C$ there is a pole at $w_0=0$ of order $p$ and simple poles at $w_{1,2}=\pm\sqrt{2a+1-2\sqrt{a^2+a}}$. Letting $f(w)$ be the integrand and $t=w_{1,2}^2$, we obtain \begin{align}\int_{-\pi}^\pi\frac{\sin x\sin px}{\sin^2x+\sin^2y}\,dx&=2\pi(\operatorname{Res}(f,0)+2\operatorname{Res}(f,\sqrt t))\end{align} where $\operatorname{Res}(f,\sqrt t)=(t-1)(t^p-1)t^{-p/2}(a^2+a)^{-1/2}/4$. To evaluate the residue at $0$, note that \begin{align}(p-1)!\operatorname{Res}(f,0)=\lim_{w\to0}\frac{d^{p-1}}{dw^{p-1}}\frac{(w^2-1)(w^{2p}-1)}{w^4-(4a+2)w^2+1}=\lim_{w\to0}\frac{d^{p-1}}{dw^{p-1}}\frac{-w^2+1}{w^4-(4a+2)w^2+1}\end{align} after clearing terms of orders higher than $p-1$. Define $h(w):=(w^4-2(2a+1)w^2+1)^{-1}$ and invoking Leibniz's rule yields $$(p-1)!\operatorname{Res}(f,0)=h^{(p-1)}(0)-2\binom{p-1}2h^{(p-3)}(0)\implies\operatorname{Res}(f,0)=\frac{h^{(p-1)}(0)}{(p-1)!}-\frac{h^{(p-3)}(0)}{(p-3)!}$$ Substituting this into $(1)$ gives \begin{align}I_{p,q}&=2\pi\Re\int_0^{\pi/2}\left(\operatorname{Res}(f,0)+2\operatorname{Res}(f,\sqrt t)\right)e^{iqy}\,dy\end{align} and it follows from the Riemann-Lebesgue lemma that $I_{p,\infty}=0$ provided that the sum of the residues is $L^1$-integrable. It is evident that $|\operatorname{Res}(f,0)|<\infty$ as it is a polynomial in $2\sin^2y+1$, since for all derivatives, the denominator at zero is $1$. Finally, as $t(a)$ is decreasing and non-negative, $$\operatorname{Res}(f,\sqrt t)<\frac{(t-1)(t^p-1)}{\sqrt{a^2+a}}t(a=1)^{-p/2}=2t(a=1)^{-p/2}\left(\frac a{a^2+a}-1\right)(t^p-1)<\infty$$ so the criterion is met.


Repeating the process above for $q$, we have $$\int_{-\pi}^\pi\frac{\cos qy}{\sin^2x+\sin^2y}\,dy=2i\oint_D\frac{z^{2q}+1}{z^{q-1}(z^4-(4b+2)z^2+1)}\,dz$$ where $D$ is the unit disk defined by $z=e^{iy}$ and $b:=\sin^2x$. Notice the denominator of the integrand $g(z)$ is essentially the same as that of $f(w)$ so the computations will be very similar. Hence $$I_{p,q}=-2\pi\Im\int_0^\pi\sin x\cdot(\operatorname{Res}(g,0)+2\operatorname{Res}(g,z^*))e^{ipx}\,dx$$ where $z^*=\sqrt{2b+1-2\sqrt{b^2+b}}$ and the same conclusion should follow.

  • Thank you for your answer, TheSimpliFire. I've been going through it and there are a couple of points that aren't clear to me -- apologies as I'm surely missing something. First: why is $\operatorname{Res}(f,\sqrt t)=(t-1)(t^p-1)t^{-(p+3)/2}/4$? If $t$ is as in your answer, and $s$ is such that $z^2 - (4a+2)z + 1 = (z-t)(z-s)$, wouldn't one have $\operatorname{Res}(f,\sqrt t)=(t-1)(t^p-1)t^{-(p+1)/2}(t-s)^{-1}/2$? I don't think that $t = -s$ holds. Second, why is $\int_{-\pi}^{\pi} \operatorname{Res}(f,0) cos(qy)dy = 0$? – DashDotDashDot Oct 29 '20 at 21:55
  • @DashDotDashDot $w^4-(4a+2)w^2+1=(w^2-t)(w^2-s)=(w-\sqrt t)(w+\sqrt t)(w^2-s)$ where $s=2a+1+2\sqrt{a^2+a}$ but this is outside the contour as $s>1$. – ə̷̶̸͇̘̜́̍͗̂̄︣͟ Oct 30 '20 at 09:20
  • Thank you, yes, this is clear :) And then, $ \lim_{w \to \sqrt t} (w-\sqrt t) / ( (w - \sqrt t)(w + \sqrt t)(w^2 -s) ) = ( 2\sqrt t (t - s) )^{-1}$. Whereas, if I understand correctly from your expression of $\operatorname{Res}(f,\sqrt t)$, you arrive to a value of $(4 t \sqrt t)^{-1}$ for that limit. Is this so or have I misunderstood? – DashDotDashDot Oct 30 '20 at 09:54
  • Yes, so we obtain $(t-1)(t^p-1)t^{-p/2}(4t\sqrt t)^{-1}=(t-1)(t^p-1)t^{-(p+3)/2}/4$. I will correct the result on $\operatorname{Res}(f,0)$ as it is not zero, but we can still use Riemann-Lebesgue. – ə̷̶̸͇̘̜́̍͗̂̄︣͟ Oct 30 '20 at 09:56
  • Thank you again! But what is not clear to me is why $2 \sqrt t (t-s) = 4t \sqrt t$ as I think you use? I think $t-s = 4\sqrt{a^2 + a}$. And thank you for the clarification about the residual at 0; I see how R-L still applies but I was confused by that step :) – DashDotDashDot Oct 30 '20 at 10:01
  • You are correct! Thanks for spotting that, will amend. – ə̷̶̸͇̘̜́̍͗̂̄︣͟ Oct 30 '20 at 10:06
  • Thank you so much! I will see if I can find a way around it myself too. And fwiw, I find your argument pure brilliance otherwise :) I tried using the residue theorem before, but the residue at 0 defeated me -- your proof is beautiful! – DashDotDashDot Oct 30 '20 at 10:38
  • See revised @DashDotDashDot – ə̷̶̸͇̘̜́̍͗̂̄︣͟ Oct 30 '20 at 11:26
  • @DashDotDashDot Revisiting this again after someone pointed this out to me: I'm sorry I made a mistake regarding the Chebyshev polynomials -- they cannot be applied in this case since $|2a+1|>1$ so the generating function is undefined. It turns out there is a much simpler argument for boundedness, which is that the denominator is independent of $a$ for all derivatives of $h$ which makes each of them polynomials. – ə̷̶̸͇̘̜́̍͗̂̄︣͟ Oct 02 '22 at 11:16