18

Experimenting with WA I noticed that the function $1- \cos^{\frac{2}{3}}x$ has the Maclaurin expansion with all coefficients positive ( works for any exponent in $[0, \frac{2}{3}]$). A trivial conclusion from this is $|\cos x|\le 1$, but it implies more than that, for instance see this. Maybe some ''natural'' proofs are available. Thank you for your interest!

Note: an attempt used a differential equation satisfied by the function. But the answer by @metamorphy just solved it the right way.

$\bf{Added:}$ Some comments about series with positive coefficients.

By $P$ we denote a series with positive coefficients ( no free term),

  1. If $a>0$ then $\frac{1}{(1-P)^a} = 1+P$ (moreover, the positive expression on RHS is a polynomial in $a$ with positive coefficients

  2. If $0<a < 1$ then $(1-P)^a = 1-P$. Similarly the expressions for $a = \frac{t}{t+1}$ are positive in $t$.

2'. If $1-f= P$ then $1- f^{a} =P$ for any $0 < a < 1$, and similar with above.

$\bf{Added:}$ It turns out that the function $\cos^{2/3} x$ has a continued fraction (an $S$-fraction, from Stieltjes) that is "positive" ( similar to the continued fraction for $\tan x$). This is a stronger statement than the one before. Maybe there is some approach using hypergeometric functions.

orangeskid
  • 53,909
  • 2
    Nice and interesting observation. $\to +1$ – Claude Leibovici May 05 '22 at 06:15
  • 1
    I'm sure there's a clever proof that obtains for $n\ge1$ results of the form$$\frac{d^n}{dx^n}(1-\cos^{2/3}x)=\sum_kf_{n,,k}(\cos x)\sin^kx,,f_{n,,0}(1)\ge0.$$ – J.G. May 05 '22 at 07:23
  • @J.G.: Would that imply the positivity of the coefficients? – orangeskid May 05 '22 at 08:21
  • 1
    Well, we expect a Taylor series for $1-\cos^{2/3}x$ whose $x^n$ coefficient is $1/n!$ times the $n$th derivative at $x=0$. @metamorphy's recent answer is similar to what I'd hoped for: actually, their approach is even nicer. – J.G. May 05 '22 at 09:00
  • @Claude Leibovici: I got into thinking about expansions with positive coefficients after reading an answer of yours some time ago! – orangeskid May 06 '22 at 11:34
  • 1
    Although the coefficient of $x^{2m}$ can be written as $c(m) = \frac{(-1)^{m-1}}{2^{a}(2m)!} \sum_{k\ge 0} \binom{a}{k} (2k-a)^{2m}$ it seems hard to make a proof out of it. – Dr. Wolfgang Hintze May 13 '22 at 09:49
  • @Dr. Wolfgang Hintze: Directly it seems difficult. One approach for this question and the related one uses a differential equation and a "helper" series that turns out to be positive by other means ( being understood better). – orangeskid May 13 '22 at 10:08
  • @Dr. Wolfgang Hintze: I wonder if you can prove in this way that $\cos^a x$ has a positive expansion for $a< 0$ ? ( that was known already from the product expansion for $\cos x$). – orangeskid May 13 '22 at 10:53
  • @orangeskid Interesting question. I have checked that at least for some small values of $m$ all coefficients are negative for $a\lt0$ close to $0$. – Dr. Wolfgang Hintze May 13 '22 at 11:04
  • 1
    Replacing $cos(x)$ by $1-\frac{x^2}{2}$ gives for the coefficient of $x^{2k}$ the expression $c(k) = -(1/2^k) \binom{k - a - 1}{k}$ which has the sign of $a$. – Dr. Wolfgang Hintze May 13 '22 at 11:15
  • 1
    Just one comment more: The replacement $\cos\to 1-\frac{x^2}{2}+\frac{x^4}{24}$ is enough to generate the "critical" value $a=\frac{2}{3}$, the sign question and the "nicer" roots for $a\ge 1$. – Dr. Wolfgang Hintze May 13 '22 at 11:31
  • @Dr. Wolfgang Hintze: A wonderful observation! Now, can we prove this for the substitute of $\cos x$? – orangeskid May 13 '22 at 17:01
  • @Dr.Wolfgang Hintze: I posted a question based on your great observation ! – orangeskid May 17 '22 at 14:44
  • 1
    @orangeskid thank you for your kind words. I'm happy that my observation has already blossomed. – Dr. Wolfgang Hintze May 18 '22 at 16:04
  • Look at the update. – Claude Leibovici May 20 '22 at 03:22
  • @Dr. Wolfgang Hintze: Indeed it had! :-) Now...the higher truncation seem to work too, right?. Every case Is reducible to showing that certain rational fraction have positive series. The quadratic polynomial case is simpler. Now, there are poly of deg $3$ for which the optimal exponent is not given by the quadratic part. So these are very particular. Also, the analogue with truncations with $\sin x/x$ does not work(!). Very interesting – orangeskid May 20 '22 at 15:16
  • 1
    @orangeskid Very interesting. I also considered the general question of the power series of the composition of two functions. This led me to this possibly helpful article on Faà di Bruno's formula: https://en.wikipedia.org/wiki/Fa%C3%A0_di_Bruno%27s_formula#Combinatorics_of_the_Fa%C3%A0_di_Bruno_coefficients – Dr. Wolfgang Hintze May 21 '22 at 10:27
  • @Dr. Wolfgang Hintze: Yes, that seems to be very interesting. So you consider the outside series as $1-(1-t)^{\alpha}$ if I understand correctly. – orangeskid May 27 '22 at 09:42
  • 1
    @orangeskd I mean f(t(x)) = A + B t + C t^2 + ... and t=t(x) = x+a x + b x^2 +... = u + v x + w x^2 + ... – Dr. Wolfgang Hintze May 28 '22 at 13:58

5 Answers5

15

(A proof, not very "natural" though.) $f(x)=1-\cos^{2/3}x$ satisfies $xf''(x)=g(x)f'(x)$, where $$g(x)=x\cot x+\frac{x}3\tan x=1+\sum_{n=1}^\infty g_n x^n$$ with $g_n=0$ for odd $n$, and $3g_{2n}=(-1)^n 2^{2n}(4-2^{2n})B_{2n}/(2n)!\geqslant 0$ using Bernoulli numbers (and the alternating-sign property of these).

Now $f'(x)=\sum_{n=1}^\infty f_n x^n$ implies $(n-1)f_n=\sum_{k=1}^{n-1}g_{n-k}f_k$, giving $f_n\geqslant 0$ by induction.

metamorphy
  • 39,111
5

This is not an answer but it is too long for a comment.

This is an interesting observation. I think that we just need to look at the coefficient of $x^4$; as soon as it is positive, all the next coefficients are positive too.

There are amazing patterns.

Let $a=\frac {2 \times 10^k+1}{3 \times 10^k}$. The coefficient of $x^4$ in the expansion of $\big[1-\cos^a(x)\big]$ write $$c_4(k)=-\frac {\alpha_k}{24\times 10^{2k}}$$ and the $\alpha_k$ form the sequence $$\{1,7,67,667,6667,66667,666667,\cdots\}$$

Let $b=\frac {2 \times 10^k-1}{3 \times 10^k}$. The coefficient of $x^4$ in the expansion of $\big[1-\cos^b(x)\big]$ write $$c_4(k)=\frac {\beta_k}{72\times 10^{2k}}$$ and the $\beta_k$ form the sequence $$\{1,19,199,1999,19999,199999,1999999,\cdots\}$$

3

As a series, we have $$1-\cos^a(x)=a\sum_{n=1}^\infty(-1)^{n+1}\,\frac {P_n(a)}{(2n)!}\, x^{2n}$$ and the first polynomials are $$\left( \begin{array}{cc} n & P_n(a) \\ 1 & 1 \\ 2 & 3 a-2 \\ 3 & 15 a^2-30 a+16 \\ 4 & 105 a^3-420 a^2+588 a-272 \\ 5 & 945 a^4-6300 a^3+16380 a^2-18960 a+7936 \\ 6 & 10395 a^5-103950 a^4+429660 a^3-893640 a^2+911328 a-353792 \end{array} \right)$$

Explored up to $n=100$, none of the $P_n(a)$ is factorable and none of them shows rational solution (except for $n=2$ !)

Edit

After @orangeskid's comment, let $a=\frac{2 t}{3 (t+1)}$ $$1-\cos ^{\frac{2 t}{3 (t+1)}}(x)=\frac t 6\sum_{n=1}^\infty \frac {Q_n(t)} {b^n\,(1+t)^n}\,x^{2n}$$ where the $b_n$ form the sequence $$\{1,72,97200,457228800,617258880000,16132678087680000,\cdots\}$$ which is quite interesting; for example, looking at $\frac{b_{n+1}}{b_n}$ gives $$\{72,1350,4704,1350,26136,49686,28800,140454,216600,106722\}$$

and the first $Q_n(t)$ are $$\left( \begin{array}{cc} n & Q_n(t) \\ 1 & 1 \\ 2 & 1 \\ 3 & 2 t^2+9 t+12 \\ 4 & 40 t^3+246 t^2+477 t+306 \\ 5 & 168 t^4+1222 t^3+3183 t^2+3582 t+1488 \\ 6 & 9920 t^5+83304 t^4+269106 t^3+423249 t^2+326646 t+99504 \end{array} \right)$$

Continuing for $$1-\cos ^{\frac{2 a}{3 (a+1)}}\left(x\sqrt{a+1} \right)=\frac a 6\sum_{n=1}^\infty \frac {R_n(a)} {b^n\,(1+t)^n}\,x^{2n}$$ the first $R_n(a)$ being $$\left( \begin{array}{cc} 1 & 1 \\ 2 & 1 \\ 3 & 2 a^2+9 a+12 \\ 4 & 40 a^3+246 a^2+477 a+306 \\ 5 & 168 a^4+1222 a^3+3183 a^2+3582 a+1488 \\ 6 & 9920 a^5+83304 a^4+269106 a^3+423249 a^2+326646 a+99504 \end{array} \right)$$ Compare $Q_n(t)$ and $R_n(a)$.

Amazing problem !

Update

$$P_3(a)=15 a^2-30 a+16=0 \quad \implies \quad a_\pm=1 \pm \frac{i}{\sqrt{15}}$$

$$1-\cos^{a_-}(x)=\frac{x^2}2-$$ $$\frac{x^4}{30}\Bigg[1+\frac{x^4}{315}+\frac{x^6}{945}+\frac{37 x^8}{111375}+\frac{22996 x^{10}}{212837625}+\frac{139 x^{12}}{3869775}+\frac{3109 x^{14}}{255605625}+O\left(x^{16}\right) \Bigg]-$$ $$-i\frac{x^2}{2 \sqrt{15}}+$$ $$i\frac{x^4}{6 \sqrt{15}}\Bigg[1+\frac{x^4}{105}+\frac{53 x^6}{23625}+\frac{23 x^8}{37125}+\frac{63964 x^{10}}{354729375}+\frac{353 x^{12}}{6449625}+\frac{153547 x^{14}}{8946196875}+O\left(x^{16}\right) \Bigg]$$

  • Interestingly, the OP's choice $a=2/3$ is the root of $P_2$. – Gary May 07 '22 at 03:11
  • @Gary. I was hpoing to find another one. – Claude Leibovici May 07 '22 at 03:22
  • it is fascinating. So we know now that the polynomials $P_n$ are positive on $[0, 2/3]$. That would mean $P_n(\frac{2}{3}\cdot \frac{t}{t+1})$ is positive for $t\ge 0$. It seems that the numerator after that substitution again has positive coefficients... – orangeskid May 07 '22 at 03:27
  • I considered the expansion of $1- \cos( \sqrt{a+1} x )^{\frac{2}{3} \cdot \frac{a}{a+1}}$ in $x$ and got positive coefficients as far as I can see – orangeskid May 07 '22 at 04:16
  • @orangeskid. Look at my edit after your first comment. There are interesting things. – Claude Leibovici May 07 '22 at 04:22
  • What is $t$ in the denominator of the expansion involving the $R_n$'s? – Gary May 07 '22 at 09:28
  • Very interesting! Also, I added some details in the post. I wonder if it could explain some of the patterns. – orangeskid May 07 '22 at 09:38
3

Start with the positive series $$\csc x = \frac{1}{x} + \frac{1}{6} x + \frac{7}{360} x^3 + \frac{31}{15120} x^5 + \cdots $$

Take the square and obtain the positive series $$\csc^2 x =\frac{1}{x^2} + \frac{1}{3} + \frac{1}{15} x^2 + \frac{2}{189} x^4 + \cdots $$

Now take the derivative and get the series $$- \frac{2\cos x}{\sin^3 x} = -\frac{2}{x^3} + \frac{2}{15} x + \frac{8}{189} x^3 + \cdots $$ where all the coefficients after the first are positive. We conclude that the series $$ 1 - \frac{x^3 \cos x}{ \sin ^3 x}=\frac{1}{15}x^4 + \frac{4}{189}x^6 + \frac{1}{225}x^8 + \cdots $$ is positive. Now, since $(1-t)^{-1/3}$ is a positive series in $t$, if we substitute for $t$ a positive series in $x$ without free term we get again a positive series in $x$. We conclude that $$\frac{\sin x}{x \cos^{1/3} x}= 1 + \frac{1}{45}x^4 + \frac{4}{567}x^6 + \frac{1}{405} x^8+ \cdots $$ is a positive series. Now multiply by $\frac{2}{3} x$ and integrate and get the positive $1- \cos^{2/3} x$ ( all coefficients are, except perhaps the free term, which turns out to be $0$).

$\bf{Note:}$ In fact we can prove that the series for $$\sin x - \frac{x^3 \cos x}{ \sin^2 x}$$ is positive. For this, we start with the expansion $$\csc x= x^{-1} + \frac{1}{6} x + \frac{7}{360}x^3 + \frac{31}{15120} x^5 + \cdots = \sum_{n=0}^{\infty} \frac{(-1)^{n+1} 2(2^{2n-1} -1) B_{2n}}{(2n)!}x^{2n-1}$$

Taking the derivative with respect to $x$ we get the expansion of $\frac{\cos x}{\sin^2 x}$. From here we conclude that $\sin x- \frac{x^3 \cos x}{\sin^2 x}$ is positive. Multiplying by the positive $\csc x$ we conclude that $1 - \frac{x^3 \cos x}{\sin^3 x}$ is positive.

orangeskid
  • 53,909
1

Some thoughts:

Let $f(x) := \cos^{2/3} x$.

We have $$f'(x) = \frac23 \cos^{-1/3} x\, (-\sin x) = -\frac23 f(x)\tan x $$ and $$f''(x) = -\frac23 f'(x) \tan x - \frac23 f(x) (\tan x)' = - \frac23 f(x) - \frac29 f(x)\tan^2 x .$$

Conjecture 1: For $n = 1, 2, \cdots$, $$f^{(2n - 1)}(x) = f(x) \sum_{k=1}^n c_{2n-1, k} \tan^{2k-1} x$$ and $$f^{(2n)}(x) = f(x) c_{2n, 0} + f(x) \sum_{k=1}^n c_{2n, k} \tan^{2k} x$$ where $c_{2n-1, k} \le 0$ for all $n\ge 1,\, 1\le k \le n$, and $c_{2n, k} \le 0$ for all $n\ge 1,\, 0 \le k \le n$.

I think we can use Mathematical Induction or Strong Induction to prove it.

River Li
  • 37,323