11

Consider the sequence $\{a_n\}$ defined by

$$ a_n = \sqrt{n}\cdot \int_{0}^{\pi} \left( \frac{1 + \cos t}{2} \right)^n dt.$$

An exercise in Rudin, Real and Complex Analysis, requires showing that this sequence is convergent to a real number $a$, with $ a > 0$. I don't have any idea of how to prove this. I only obtained the following estimation

$$ \begin{align*} \int_{0}^{\pi} \left( \frac{1 + \cos t}{2} \right)^n dt &= 2 \int_{0}^{\frac{\pi}{2}} \left( 1 - \sin^2 t \right)^n dt \\ &> 2 \int_{0}^{\frac{1}{\sqrt{n}}} (1 - t^2)^n dt \\ &> 2 \int_{0}^{\frac{1}{\sqrt{n}}} (1 - n t^2) dt \\ & = \frac{4}{3 \sqrt{n}}, \end{align*}$$

which shows that $ a_n > \frac{4}{3}$.

Thank you very much in advance for any help.

Cortizol
  • 3,669

5 Answers5

10

Edited. The original answer is appended.

Put $$c_m:=\int_0^{\pi/2} \cos^m s\ ds\qquad(m\geq0)\ .$$ Then $c_1=1$, $\> c_2={\pi\over4}$, and an easy partial integration shows that $$c_m={m-1\over m} c_{m-2}\qquad(m\geq2)\ .\tag{1}$$ Now $$a_n:=\sqrt{\mathstrut n}\int_0^\pi\left({1+\cos t\over2}\right)^n\ dt=2\sqrt{\mathstrut n}\int_0^{\pi/2}\cos^{2n}s\ ds$$ and therefore $$a^2_n=4 n\> c_{2n}^2\ .$$ I claim that $$a_n^2={c_{2n}\over c_{2n-1}}\pi\qquad(n\geq1)\ .\tag{2}$$ Proof by induction: $(2)$ is true for $n=1$. Assume that it is true for $n-1$. Using $(1)$ we then have $$\eqalign{a_n^2 &= {n\over n-1}\left({c_{2n}\over c_{2n-2}}\right)^2 a_{n-1}^2 = {n\over n-1}{2n-1\over 2n}{c_{2n}\over c_{2n-2}}{c_{2n-2}\over c_{2n-3}}\pi\cr &={n\over n-1}{2n-1\over 2n}{2n-2\over 2n-1}{c_{2n}\over c_{2n-1}}\pi={c_{2n}\over c_{2n-1}}\pi\ .\qquad\square\cr}$$ Since $c_m\leq c_{m-1}\leq c_{m-2}$ it follows easily from $(1)$ that $$\lim_{m\to\infty}{c_m\over c_{m-1}}=1\ .$$ From $(2)$ we therefore conclude that $\lim_{n\to\infty} a_n=\sqrt{\mathstrut \pi}$.

Original answer: Let $$b_n:=\int_0^\pi\left({1+\cos t\over2}\right)^n\ dt=2\int_0^{\pi/2}\cos^{2n}s\ ds\ .$$ Then $$b_n =2\int_0^{\pi/2}\cos^{2n-2}s(1-\sin^2 s)\ ds=b_{n-1}-2\int_0^{\pi/2}\bigl(\cos^{2n-1}s\>\sin s\bigr)\cdot\sin s\ ds\ .$$ We integrate the last integral by parts and obtain $$b_n=b_{n-1}-{1\over 2n-1} b_n\ .$$ Therefore the $b_n$ satisfy the recursion $$b_n={2n-1\over 2n}\>b_{n-1}\qquad(n\geq1)\ ,$$ and the $a_n=\sqrt{\mathstrut n}\>b_n$ satisfy the recursion $$a_n^2={(2n-1)^2\over 2n\>(2n-2)}\>a^2_{n-1}\qquad(n\geq2)\ .$$ As $a_1={\pi\over2}$ we therefore have $$a_n^2=2\cdot\left({1\over2}\cdot{3\over2}\cdot{3\over4}\cdot{5\over4}\cdot{5\over6}\cdots{2n-1\over2n-2}{2n-1\over2n}\right)\cdot{\pi^2\over4}\qquad(n\geq2)\ .$$ In the large parenthesis the reciprocal of Wallis' product appears. It follows that this parenthesis converges to ${2\over\pi}$, and we obtain $$\lim_{n\to\infty}a_n=\sqrt{\mathstrut \pi}\ .$$

  • This is very, very nice! +1 – Pedro Jul 16 '13 at 01:02
  • @PeterTamaroff I upvoted the answer (+1), but this is just a high school approch with a well-defined pathway. I'm not sure what you refer at when you say "very, very". – user 1591719 Jul 20 '13 at 09:46
  • @PeterTamaroff http://en.wikipedia.org/wiki/Wallis_product – user 1591719 Jul 20 '13 at 09:48
  • @Chris'swisesister "This is just a high school approach..." Are you serious? I find it very, very stupid to underestimate things like that. This answer is clear and to the point. The integral is just another way to encode Wallis' result, and the answer doesn't hide it. – Pedro Jul 20 '13 at 16:40
  • @PeterTamaroff "integration by parts" is a very powerful tool, but I cannot say it's a "very very (many times) nice too" every time I see it because it's just an usual tool. This case above is the same. It's a well-known case of integral that is connected to Wallis product and where integration by parts leads to some recurrence, a well-well known way. The way is definitely nice, but you refer to it as if you saw it for the first time in your life. – user 1591719 Jul 20 '13 at 19:30
  • @Peter Tamaroff: Thanks for the flowers! I have edited my answer so that no reference to Wallis' product needs to be made. – Christian Blatter Jul 22 '13 at 12:10
  • @Chris' sis: I have edited my answer so that no reference to Wallis' product needs to be made. – Christian Blatter Jul 22 '13 at 12:11
  • @ChristianBlatter Heh, every now and then one can rederive Wallis' product, why not? =) – Pedro Jul 22 '13 at 13:43
9

$$ \begin{align} \lim_{n\to\infty}\sqrt{n}\int_0^\pi\left(\frac{1+\cos(t)}{2}\right)^n\,\mathrm{d}t &=\lim_{n\to\infty}\sqrt{n}\int_0^\pi\cos^{2n}(t/2)\,\mathrm{d}t\\ &=\lim_{n\to\infty}2\sqrt{n}\int_0^{\pi/2}\cos^{2n}(t)\,\mathrm{d}t\\ &=\lim_{n\to\infty}2\sqrt{n}\int_0^{n^{-1/3}}\left(1-\sin^2(t)\right)^n\,\mathrm{d}t\\ &=\lim_{n\to\infty}2\sqrt{n}\int_0^{n^{-1/3}}\left(1-t^2+\color{#C00000}{O}\big(t^4\big)\right)^n\,\mathrm{d}t\\ &=\lim_{n\to\infty}2\int_0^{n^{1/6}}\left(1-t^2/n+\color{#C00000}{O}\big(t^4/n^2\big)\right)^n\,\mathrm{d}t\\ &=\lim_{n\to\infty}2\int_0^{n^{1/6}}e^{-t^2+\color{#00A000}{O}(t^4/n)}\,\mathrm{d}t\\ &=2\int_0^\infty e^{-t^2}\,\mathrm{d}t\\[6pt] &=\sqrt\pi \end{align} $$ Where, for $n\ge3$, the red $O$ has constant $1/3$ and the green $O$ has constant $2$.

Note that $$ \begin{align} \lim_{n\to\infty}\sqrt{n}\int_{n^{-1/3}}^{\pi/2}\left(1-\sin^2(t)\right)^n\,\mathrm{d}t &\le\lim_{n\to\infty}\frac\pi2\sqrt{n}\left(1-\frac4{\pi^2}n^{-2/3}\right)^n\\ &\le\lim_{n\to\infty}\frac\pi2\sqrt{n}\,e^{-4n^{1/3}/\pi^2}\\[4pt] &=0 \end{align} $$


Motivation

As Didier points out, this approach is less elegant than those that make use of certain trigonometric identities, and perhaps Rudin posed this problem using trigonometric functions because one can use trigonometric identities to get a more elegant solution. However, I had computed the integral of $\cos^n(t)$ too many times recently, and I wanted to single out key sufficient features of $f$ that allow the limit $$ \lim_{n\to\infty}\sqrt{n}\int_0^\pi f(t)^n\,\mathrm{d}t=\sqrt\pi $$ those being

  1. $f(t)=1-t^2+O(t^4)$

  2. $f(t)\le1-kt^2$ for some $k\gt0$

robjohn
  • 345,667
  • And of course the previous-to-last equal sign is where everything happens (so much so that, with no further justification, it could cause violent reactions from a susceptible teacher...). – Did Jul 15 '13 at 10:48
  • @Did: yes, I noticed that and I am working on it. – robjohn Jul 15 '13 at 10:49
  • (+1) although the equality $1-\sin^2(t)=1-t^2+O(t^4)$ is valid only on a neighborhood of $0$. –  Jul 15 '13 at 10:58
  • @SamiBenRomdhane: It is certainly valid for all $t$ since $1-\sin^2(t)\le1$. – robjohn Jul 15 '13 at 11:01
  • @Did: hopefully, the violence can be avoided now. – robjohn Jul 15 '13 at 11:20
  • Still no idea about the precise result (you mention none) which would allow this step. – Did Jul 16 '13 at 00:01
  • @Did:$$ \begin{align} \lim_{n\to\infty}2\int_0^{n^{1/6}}e^{-t^2+O(t^4/n)},\mathrm{d}t &=\lim_{n\to\infty}2\int_0^{n^{1/6}}\left(e^{-t^2} +e^{-t^2}O(n^{-1/3})\right),\mathrm{d}t\ &=\lim_{n\to\infty}2\int_0^{n^{1/6}}e^{-t^2},\mathrm{d}t +\lim_{n\to\infty}2\int_0^{n^{1/6}}e^{-t^2}O(n^{-1/3}),\mathrm{d}t\ &=2\int_0^\infty e^{-t^2},\mathrm{d}t+\lim_{n\to\infty}O(n^{-1/3})\ &=2\int_0^\infty e^{-t^2},\mathrm{d}t \end{align} $$ – robjohn Jul 16 '13 at 00:34
  • This approach seems overcomplicated and what it really achieves is unclear to me. What is the precise status of the $O(\ )$ terms? How comes one can integrate them at will? Is the result another $O(\ )$ term? Since the answer to these questions in general is a resounding "no", the answer to these questions in the specific case of your post can at best be a "maybe, who knows". – Did Jul 16 '13 at 08:12
  • @Did: I was attempting to show that the action happens very close to $0$ by showing the integral away from $0$ vanishes, and approximating the function near $0$ by series. The absolute value of each $O(x)$ term is bounded by a constant, independent of $n$, times $x$ on the given range. – robjohn Jul 16 '13 at 13:07
  • @Did: For example, for all $t$, $$ \sin^2(t)=t^2+O(t^4)$$ where $|O(t^4)|\le\frac13|t|^4$

    For $x\ge-\frac23$, $x-x^2\le\log(1+x)\le x$. Therefore, if $|t|\le n^{1/6}$ and $n\ge3$, then $$ \begin{align} e^{-t^2+\frac13t^4/n} &\ge\left(1-\frac{t^2}n\pm\frac{\frac13t^4}{n^2}\right)^n\ &\ge e^{-t^2-\frac13t^4/n-(t^4/n+\frac23t^6/n^2+\frac19t^8/n^3)}\ &\ge e^{-t^2-\frac43t^4/n-\frac23t^4/n^{5/3}-\frac19t^4/n^{7/3}}\ &\ge e^{-t^2-2t^4/n}\ \end{align} $$

    – robjohn Jul 16 '13 at 13:07
  • If these considerations are necessary to turn your post into a proof, this suggests to avoid the slippery $O(\ )$s altogether and to present from the start the nonasymptotic inequalities needed to reach the conclusion. For example, the conclusion of your last comment, if needed in your post, could be stated separately as $\exp(-s-2s^2)\leqslant1-s\pm s^2/3\leqslant\exp(-s+s^2/3)$ for every $s$ in a suitable range, proven, and suitably applied. Likewise for the convergences of integrals on expanding intervals (for which a theorem, with conditions, is definitely needed). – Did Jul 17 '13 at 08:43
  • Thanks for writing this answer. I learned a new trick. – Potato Aug 23 '13 at 20:08
  • @robjohn That is essentially Laplace Method. +1. – Felix Marin Mar 30 '14 at 10:08
7

Related problem. One can evaluate the integral using the beta function

$$ \beta(x,y) = 2\int_0^{\pi/2}(\sin\theta)^{2u-1}(\cos\theta)^{2v-1}\,d\theta, \qquad \mathrm{Re}(u)>0,\ \mathrm{Re}(v)>0. $$

First, we use the identity $1+\cos(x)=2\cos^2(x/2)$

$$ a_n=n^{1/2}\int_{0}^{\pi}\left(\frac{1+\cos(x)}{2}\right)^n dx = n^{1/2}\int_{0}^{\pi}{\cos^{2n}(x/2)} dx. $$

Then using the substitution $y=\frac{x}{2}$, we have

$$ a_n= 2 n^{1/2}\int_{0}^{\pi/2}{\cos^{n/2}(y)} dy= \sqrt {\pi }{\frac {\sqrt {n}\,\Gamma \left( n+1/2 \right)}{\Gamma \left( n+1 \right) }}.$$

Taking the limit as $n\to \infty$, we get

$$ \lim_{n\to \infty} a_n = \sqrt{\pi}. $$

Note that, you can use Stirling approximation for $n!=\Gamma(n+1)\sim \left(\frac{n}{e}\right)^n\sqrt{ \pi n}. $ to evaluate the limit.

  • Maybe $\frac{\Gamma(n+\tfrac{1}{2})}{\Gamma(n+1)} = \sqrt{\pi}\frac{(2n)!}{(n!)^2,4^n}$ is helpful too. – Nikolaj-K Jul 15 '13 at 11:43
4

Note that $a_n=\int\limits_0^\infty u_n(t)\mathrm dt$, where $$ u_n(t)=2^{-n}(1+\cos(t/\sqrt{n}))^n\,\mathbf 1_{0\leqslant t\leqslant\pi\sqrt{n}}. $$ It happens that $u_n\to u$ pointwise (can you show this?), where $$ u(t)=\mathrm e^{-t^2/4}, $$ hence a natural conjecture is that $a_n\to a$, where $$ a=\int_0^\infty u(t)\mathrm dt=\sqrt\pi. $$ A tool to make sure this convergence happens is Lebesgue dominated convergence theorem, which requires to find some integrable $v$ such that $|u_n|\leqslant v$ for every $n$. It happens that $v=u$ fits the bill.

To see why, note that $\cos(2t)+1=2\cos^2(t)$ hence $|u_n|\leqslant u$ for every $n$ as soon as, for every $x$ in $(0,\pi/2)$, $\cos(x)\leqslant\mathrm e^{-x^2/2}$. Any idea to show this last (classical) inequality?

Did
  • 279,727
1

(too long to be a comment)

The integrals can be calculated exactly:

$$a_n=\frac{n^{1/2}}{2^n} \int_0^\pi(1+\cos t)^n dt. $$ For starters, we can use Newton's binomial: $$a_n=\frac{n^{1/2}}{2^n}\int_0^\pi \sum_{k=0}^n \binom{n}{k}\cos^ktdt. $$ Use the even-ness of the cosine around $t=\pi$: $$a_n=\frac{n^{1/2}}{2^{n+1}}\sum_{k=0}^n \binom{n}{k} \int_0^{2\pi} \cos^k tdt $$

The integrals $\int_0^{2\pi} \cos^k t dt$ can be calculated using Euler's formula $$\int_0^{2\pi} \cos^k t dt=\begin{cases} 0 & k=2m+1 \\ \frac{2\pi}{2^{2m}} \binom{2m}{m} & k=2m \end{cases}. $$ Therefore the summation is only needed for even indices $k=2m$, and $$a_n=\frac{n^{1/2}}{2^{n+1}} \sum_{m=0}^{\lfloor{\frac{n}{2}} \rfloor} \binom{n}{2m} \frac{2\pi}{2^{2m}} \binom{2m}{m} $$

user1337
  • 24,381