2

How to determine the convergence of $\int_0^\infty e^{-x}\log(\cos^2x)dx$? Any hint is appreciated.

Sam
  • 367

2 Answers2

3

The function $\log(\cos^2)$ is nonpositive on $(0,+\infty)$ with period $\pi$ hence, splitting $(0,+\infty)$ into the intervals $n\pi+(0,\pi)$ for $n\geqslant0$, one sees that, finite or infinite, the integral is $$ I=\sum_{n\geqslant0}\int_0^\pi\mathrm e^{-x-n\pi}\log(\cos^2x)\mathrm dx=\frac{J}{1-\mathrm e^{-\pi}},\qquad J=\int_0^\pi\mathrm e^{-x}\log(\cos^2x)\mathrm dx. $$ Since $\mathrm e^{-\pi}\leqslant\mathrm e^{-x}\leqslant1$ on $(0,\pi)$, $J$ converges if and only if $K$ converges, where $$ K=\int_0^\pi\log(\cos^2x)\mathrm dx=2\int_0^{\pi/2}\log(\cos^2x)\mathrm dx=2\int_0^{\pi/2}\log(\sin^2x)\mathrm dx. $$ The last integral has a singularity at $x=0$, $x\gt0$, where $\log(\sin^2x)\sim2\log x$. Since $x\mapsto\log x$ is integrable at $x=0$, $x\gt0$, $K$ is finite. Finally, $I$ converges.

Did
  • 279,727
0

CORRECTION
Well, I guess I should admit that my first attempt contained an error. So may be I should try the second time.
But I still will try to use the same series expansion as in the first time. One can utilize the idea of splitting the integral over periods of $\ln(\cos^2(x))$ (and the property of logarythms $\ln(\cos(x)^2)=2 \ln(\cos(x))$) like: $$2\int_{0}^{\pi \over 2}\mathrm e^{-x}\log(\cos(x))\mathrm dx+2\sum_{n=1}^\infty\int_{\frac{\pi}{2}+n}^{\frac{3\pi}{2}+n}\mathrm e^{-(x+n\pi)}\log(\cos(x))\mathrm dx$$ Then for $ x^2\leq \frac{\pi^2}{4}$ one can use expansion: $$\ln(\cos(x))=-\frac{1}{2}\sum_{k=1}^\infty \frac{\sin^{2k}(x)}{k},$$ And $$-\sum_{k=1}^\infty \frac{1}{k}\int_{0}^{\pi \over 2}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx-\sum_{k=1}^\infty \frac{1}{k}\sum_{n=1}^\infty \mathrm e^{-n\pi}\int_{\frac{\pi}{2}+n}^{\frac{3\pi}{2}+n}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx$$ Then one can use the fact that the functions $\mathrm e^{-x}\sin^{2k}(x)$ within the integration limits, so: $$\int_{0}^{\pi \over 2}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx\leq \int_{0}^{\infty}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx$$ And the function under integral $\int_{\frac{\pi}{2}+n}^{\frac{3\pi}{2}+n}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx$ gets smaller on every sequential "quasi-period" $n$, so: $$\sum_{n=1}^\infty \mathrm e^{-n\pi}\int_{\frac{\pi}{2}+n}^{\frac{3\pi}{2}+n}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx\leq \sum_{n=1}^\infty \mathrm e^{-n\pi} n \int_{\frac{\pi}{2}}^{\frac{3\pi}{2}}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx.$$ One can again bound the obtained integral: $$\int_{\frac{\pi}{2}}^{\frac{3\pi}{2}}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx\leq \int_{0}^{\infty}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx.$$ Assuming, that: $$\sum_{n=1}^\infty \mathrm e^{-n\pi} n=\frac{e^{\pi }}{\left(e^{\pi }-1\right)^2}$$ one will finally gets: $$\int_0^\infty e^{-x}\log(\cos^2x)dx\leq -\bigg(\!1+\frac{e^{\pi }}{\left(e^{\pi }-1\right)^2}\!\!\bigg)\sum_{k=1}^\infty \frac{1}{k}\int_{0}^{\infty}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx $$ At last one can use the following equality (see: "Integrals and Series: Elementary Functions (Vol. 1)" A. P. Prudnikov, Y.A. Brychkov, O.I. Marichev): $$\int_{0}^{\infty}\mathrm e^{-x}\sin^{2k}(x)\mathrm dx=(2k)!\prod_{l=0}^k\frac{1}{1+4l^2}$$ And $$\prod_{l=0}^k\frac{1}{1+4l^2}=\frac{\pi 2^{-2 k-1} \mathrm{csch\big(\frac{\pi}{2}\big)}}{\left(\Gamma \left(\left(1-\frac{i}{2}\right)+k\right) \Gamma \left(\left(1+\frac{i}{2}\right)+k\right)\right)}$$ The resultant sum is convergent (here I used Wolfram Mathematica): $$\frac{2 \ \mathrm{csch\big(\frac{\pi}{2}\big)}}{\pi }\sum _{k=1}^{\infty }\frac{1}{k} \frac{2^{-2 k} (2 k)!}{\Gamma \left(\left(1-\frac{i}{2}\right)+k\right) \Gamma \left(\left(1+\frac{i}{2}\right)+k\right)}=\frac{\pi}{4}\mathrm{csch\big(\frac{\pi}{2}\big)}\frac{\, _3F_2\left(1,1,\frac{3}{2};2-\frac{i}{2},2+\frac{i}{2};1\right)}{\Gamma \left(2-\frac{i}{2}\right) \Gamma \left(2+\frac{i}{2}\right)}.$$ So finally: $$\int_0^\infty e^{-x}\log(\cos^2x)dx\leq -\bigg(\!1+\frac{e^{\pi }}{\left(e^{\pi }-1\right)^2}\!\!\bigg)\frac{\pi}{4}\mathrm{csch\big(\frac{\pi}{2}\big)}\frac{\, _3F_2\left(1,1,\frac{3}{2};2-\frac{i}{2},2+\frac{i}{2};1\right)}{\Gamma \left(2-\frac{i}{2}\right) \Gamma \left(2+\frac{i}{2}\right)}$$ or numerically it equals to $-1.081$. Numerical integration of the initial integral gives the answer $-1.032$. So: 1) the initial integral is convergent, 2) the relative error of the obtained estimate from the numeric integration of the integral is about $4.75 \%$. P.S.: by the way, if we neglect the second integral (which includes the periods $n\geq 1$) then the relative error will drop dramatically to $10^{-12}\%$.

Caran-d'Ache
  • 3,564
  • One should replace every $(-1)^k$ by $-1$. Then the convergence of the series is an issue. – Did Apr 03 '13 at 10:27
  • Great, @Caran. Thank you so much! – Sam Apr 03 '13 at 10:33
  • @Did If it was so, then the series wouldn't be convergent. But why should we change $(-1)^k$ to $(-1)$? – Caran-d'Ache Apr 03 '13 at 10:36
  • If it was so, then the series wouldn't be convergent... Why do you think so? The argument is wrong. why should we change $(−1)^k$ to $(−1)$? Because the former is wrong and the latter is right. – Did Apr 03 '13 at 10:39
  • @Sam The decision to accept an answer is entirely yours, of course, but if ever this was part of a homework you are asked to do and you choose to reproduce the answer above, be prepared for some surprises when they hand it back to you. – Did Apr 03 '13 at 10:41
  • @Did If it was so, then the series wouldn't be convergent... Why do you think so? The argument is wrong Because the series $\sum_{k=1}^\infty \frac{1}{k}$ does not converge. Or you think different? – Caran-d'Ache Apr 03 '13 at 10:46
  • Read your own answer more carefully: $\frac1k$ is multiplied by an integral depending on $k$. Your argument would be valid if this integral was $\Theta(1)$. It is not. (Note that all these are refinements compared to the earlier (deadly) problem that the expansion of $\ln(\cos^2x)$ is wrong.) – Did Apr 03 '13 at 10:50
  • Let me mention another blunder, later on in the answer (hence irrelevant to the OP since the answer errs starting at the expansion of $\log(\cos^2)$, but the error might be interesting to ponder): the upper bound by $\frac25$ of each integral cannot imply an upper bound of (what you think is) the alternating sum on the last line since its terms are alternatively (you think) positive and negative. – Did Apr 03 '13 at 10:57
  • @Did well, I guess the obtained series with $\frac{1}{k}$ is convergent. – Caran-d'Ache Apr 03 '13 at 20:36