25
$\begingroup$

Problem:

Show by hand without any computer assistance: $$e^{e^2}>1000\phi,$$ where $\phi$ denotes the golden ratio $\frac{1+\sqrt{5}}{2} \approx 1.618034$.

I come across this limit showing: $$\lim_{x\to 0}x!^{\frac{x!!^{\frac{2}{x!!-1}}}{x!-1}}=e^{e^2}.$$ I cannot show it without knowing some decimals and if so using power series or continued fractions.

It seems challenging and perhaps a tricky calculation is needed.

If you have no restriction on the method, how to show it with pencil and paper ?

Some approach:

We have, using the incomplete gamma function and continued fractions: $$\int_{e}^{\infty}e^{-e^{-193/139}x^{193/139+2}}dx=\frac{139}{471}\cdot e\cdot\operatorname{Ei}_{332/471}(e^2)>e^{-e^2},$$ where $\operatorname{Ei}$ denotes the exponential integral.

Finding an integral for the golden ratio $\phi$ is needed now.

Following my comment we have :

$$e^{-e^2}<\int_{e}^{\infty}e^{-e^{-193/139}x^{193/139+2}}dx<\int_{e}^{\infty}\frac{e-2}{1000\phi (x-2)^2}dx$$

Where the function in the integral follow the current order for $x\ge e$.As said before use continued fraction of incomplete Gamma function.To finish take the log.

$\endgroup$
15
  • 5
    $\begingroup$ @DietrichBurde: It's the "golden ratio" $\varphi = \frac{1+\sqrt{5}}{2}$. The bound is very tight; to two decimal places, it is $1618.18 > 1618.03$. $\endgroup$
    – Brian Tung
    Commented Feb 12 at 19:53
  • 16
    $\begingroup$ @BrianTung I know, and you also know. The OP should write this nevertheless, because $\phi$ is also used as totient function and other things. It would help several other readers, I believe (searching titles). $\endgroup$ Commented Feb 12 at 19:54
  • 2
    $\begingroup$ FWIW, the smallest-denominator rational number that's in between the two sides is 9709/6 = 1618.1666666... $\endgroup$
    – Dan
    Commented Feb 12 at 22:20
  • 2
    $\begingroup$ are you allowed to take $\ln$ from both sides ?? $\endgroup$
    – dezdichado
    Commented Feb 13 at 1:16
  • 2
    $\begingroup$ @DesmosTutu is there any reason to believe that this is doable by hand? Like, what is the source? $\endgroup$
    – D S
    Commented Feb 13 at 13:27

7 Answers 7

13
+50
$\begingroup$

Ok, so my answer works but requires you to:

  1. Be insanely good at arithmetic (so maybe does not meet your criteria of doable by hand)
  2. Know that $\ln 10<\color{red}{2.3026}$ (this might be known to you if you frequently change bases between logarithms, or may be calculated by series expansion)

Edit (D S) :

I noticed that getting a better approximation for $\ln 10$ significantly reduces the work required for $\ln(\phi)$. So here it goes: $$\ln(10) = \ln(2)+\ln(5) = 3\ln(2) + \ln(5/4)$$ $$\begin{align}\ln(2) &= \ln((1+1/3)/(1-1/3)) \\ &= 2\sum_{k=0}^\infty \frac{(1/3)^{2k+1}}{2k+1}\\ &< 2\left(\sum_{k=0}^3 \frac{(1/3)^{2k+1}}{2k+1}\right)+ 2\left(\sum_{k=4}^\infty \frac{(1/3)^{2k+1}}{9}\right) \\ &= 2\left(\sum_{k=0}^3 \frac{(1/3)^{2k+1}}{2k+1}\right)+\frac{(1/3)^9}{4} \\ &= \frac{1910051}{2755620} < 0.693148 \end{align}$$and $$\begin{align}\ln(5/4) &= \ln((1+1/9)/(1-1/9)) \\ &= 2\sum_{k=0}^\infty \frac{(1/9)^{2k+1}}{2k+1}\\ &< 2\left(\sum_{k=0}^2 \frac{(1/9)^{2k+1}}{2k+1}\right)+ 2\left(\sum_{k=3}^\infty \frac{(1/9)^{2k+1}}{7}\right) \\ &= 2\left(\sum_{k=0}^2 \frac{(1/9)^{2k+1}}{2k+1}\right)+ \frac{(1/9)^5}{280} \\ &= \frac{3689393}{16533720}<0.223144 \end{align}$$ Hence $\ln10 < 3\times 0.693148+0.223144 = \color{blue}{2.302588}$.


Let $x$ be such that $$x^{x^2} = 1000\phi$$ or$$x^2\ln(x) = 3\ln(10)+\ln(\phi)$$ We are interested in showing $$3\ln(10)+\ln(\phi)<e^2 \iff 3\ln(10)+\ln(\phi)- e^2<0$$since $x^2\ln(x)$ is increasing in $x$.


First, we approximate $\ln(\phi)$. Note that the series expansion of $\ln(1+x)$ for $|x|>1$ is given by $$\ln(1+x) = \ln(x) - \sum_{k=1}^\infty(-1)^k\frac{x^{-k}}{k}$$So that $$2\ln(\phi) = \ln(\phi^2) = \ln(1+\phi) = \ln(\phi) + \sum_{k=1}^\infty(-1)^{k+1}\frac{\phi^{-k}}{k}$$ $$\iff \ln(\phi) = \sum_{k=1}^\infty(-1)^{k+1}\frac{\phi^{-k}}{k}< \sum_{k=1}^{17}(-1)^{k+1}\frac{\phi^{-k}}{k}$$since the magnitude of the terms in the sum are decreasing, and we have stopped at a positive term, so this approximation is greater.

Now, to calculate negative powers of $\phi$, you can just extend the Fibonacci sequence backwards, i.e, $$\ldots \color{blue}{-8,5,-3,2,-1,1},0,1,1,2,3 \ldots$$ For example, $\phi^{-1} = \color{blue}1\cdot \phi \color{blue}{- 1}$, $\phi^{-2} = \color{blue}{-1}\cdot \phi +\color{blue}2$, $\phi ^{-3} = \color{blue}2\cdot \phi \color{blue}{- 3}$ and so on.

In the end, you should get $$\sum_{k=1}^{17}(-1)^{k+1}\frac{\phi^{-k}}{k} = \frac{3433663565\sqrt 5 - 7666113149}{24504480}<\color{red}{0.48123}$$

Edit:
With our new approximation of $\ln(10)$, it is enough to take the sum till 13, i.e, $$\ln(\phi) < \sum_{k=1}^{13}(-1)^{k+1}\frac{\phi^{-k}}{k} = \frac{4109827\sqrt 5- 9120481}{144144} < \color{blue}{0.481266}$$


Next, we will approximate $e^2$. $$e^2 = \left(\sum_{k=0}^\infty\frac{1}{k!}\right)^2>\left(\sum_{k=0}^8\frac{1}{k!}\right)^2$$ (We don't use series expansion of $e^z$ as that requires more terms.) $$\left(\sum_{k=0}^8\frac{1}{k!}\right)^2 = \left(\frac{109601}{40320}\right)^2>\color{red}{7.38903}$$


Finally, $$\boxed{3\ln(10)+\ln(\phi)- e^2< 3\cdot2.3026 + 0.48123- 7.38903 = 0 \ \ }$$ Phew!

Edit:
Armed with our new bounds, we have $$\boxed{3\ln(10)+\ln(\phi) - e^2<3\cdot 2.302588+0.481266-7.38903 = 0}$$

Update DesmosTutu :

We have a trick to evaluate $\ln 10$

$$\operatorname{arcsinh}(10)=\ln(20)+\sum_{n=1}^{\infty}(-1)^{n-1}\frac{(2n-1)!!}{2n(2n)!!(10)^{2n}},\operatorname{arcsinh}(10)\simeq 3$$

Now use expansion of $\operatorname{arcsinh(x)}$ and Engels expansion of $\ln 2$

Source: Wikipedia

$\endgroup$
11
  • 7
    $\begingroup$ Woah, you have high standards for good arithmetic, such as $\frac{3433663565\sqrt 5 - 7666113149}{24504480}<\color{red}{0.48123}$. +1 though. $\endgroup$
    – Sahaj
    Commented Feb 13 at 17:52
  • 7
    $\begingroup$ @Sahaj I did say "insanesly good arithmetic" :) $\endgroup$
    – D S
    Commented Feb 13 at 18:06
  • 4
    $\begingroup$ +1. These kinds of questions are always a bit ambiguous on the requirement, but this is certainly doable. People multiply ten digit numbers in their head (I did 4 digits at my best), so this one is routine with pen and paper. $\endgroup$
    – dezdichado
    Commented Feb 13 at 23:04
  • 2
    $\begingroup$ Answer accepted!!! I open a bounty for you. $\endgroup$ Commented Feb 15 at 11:05
  • $\begingroup$ Using the integral in the comment section we have over $(0,1)$ : $$1-\frac{1/2×x^2}{1/2×x^2+1}+1/7×x^4-2/3×x^6×(1-x)^2-1/3×x^8(1-x)^2-1/13x^8(1-x^2)>1/\sqrt{x^2+1}$$ if there is no sign mistake $\endgroup$ Commented Feb 15 at 12:53
8
$\begingroup$

A very small idea which may help someone start towards a solution:

Let $\lambda = e^{e^2}/1000$; we want to show that $\lambda > \phi$. We see directly that

$$e^2 = \sum_{n=0}^\infty \frac{2^n}{n!} > \sum_{n=0}^4 \frac{2^n}{n!} = 7$$

and

$$e^{7/6} = \sum_{n=0}^\infty \frac{(7/6)^n}{n!} > \sum_{n=0}^3 \frac{(7/6)^n}{n!} = \frac{4033}{1296} > 3,$$

so

$$e^{e^2} > e^7 > 3^6.$$

Thus,

$$\lambda > 3^6/10^3 > 2/3 > 2/(1+\sqrt{5}) = 1/\phi.$$

The argument up to this point surely can be improved!

Since the roots of $x^2-x-1$ are $\phi$ and $1/\phi$, it now suffices to show that $\lambda^2-\lambda-1 > 0$, i.e. that

$$(e^{e^2})^2 > 10^3 e^{e^2} + 10^6.$$

I'm not sure yet how to do this without a computer, but there is more room between the LHS and RHS here than between $\lambda$ and $\phi$, so hopefully this is a helpful reduction. Basically we are using the fact that $x \mapsto x^2-x-1$ has slope $\approx 2\phi-1 = \sqrt{5} > 2$ near $x = \phi$ to amplify the gap between the two things we want to compare.

$\endgroup$
1
  • $\begingroup$ If you look to my approach it should be a sufficient bound maybe. $\endgroup$ Commented Feb 17 at 18:38
8
$\begingroup$

Although this calculation is long, no special skill, nor even special patience, is required to carry it out entirely by hand.

From e Continued Fraction --- from Wolfram MathWorld, \begin{multline*} e > 2 + \cfrac{1}{1 + \cfrac{1}{2 + \cfrac{1}{1 + \cfrac{1}{1 + \cfrac{1}{4 + \cfrac{1}{1 + \cfrac{1}{1 + \cfrac{1}{6}}}}}}}} = 2 + \cfrac{1}{1 + \cfrac{1}{2 + \cfrac{1}{1 + \cfrac{1}{1 + \cfrac{1}{4 + \cfrac{1}{1 + \cfrac{6}{7}}}}}}} \\ = 2 + \cfrac{1}{1 + \cfrac{1}{2 + \cfrac{1}{1 + \cfrac{1}{1 + \cfrac{1}{4 + \cfrac{7}{13}}}}}} = 2 + \cfrac{1}{1 + \cfrac{1}{2 + \cfrac{1}{1 + \cfrac{1}{1 + \cfrac{13}{59}}}}} \\ = 2 + \cfrac{1}{1 + \cfrac{1}{2 + \cfrac{1}{1 + \cfrac{59}{72}}}} = 2 + \cfrac{1}{1 + \cfrac{1}{2 + \cfrac{72}{131}}} = 2 + \cfrac{334}{465} = \cfrac{1264}{465}, \end{multline*} From $465 \times 271{,}827 = 126{,}399{,}555,$ we get $$ e > \frac{1264}{465} > 2.71827. $$ Also, from $738{,}904 \times 465^2 = 184{,}726 \times 8{,}649 \times 100 = 159{,}769{,}517{,}400,$ we get $$ e^2 > \frac{1264^2}{465^2} = \frac{1597696}{465^2} > 7.38904. $$ From $22{,}361^2 = 500{,}014{,}321,$ we get $\sqrt5 < 2.2361,$ whence $$ \phi^2 = \frac{3 + \sqrt5}2 < 2.61805. $$ Therefore $$ e-\phi^2 > 2.71827 - 2.61805 = 0.10022 > 0.1002 = 6 \times 0.0167. $$ From $157 \times 167 = 26{,}219,$ we get $$ \frac{6\phi^2}{157} < \frac{6 \times2.61805}{157} < \frac{6 \times2.6219}{157} = 6 \times 0.0167 < e - \phi^2. $$ Recall the inequality (for a proof, see e.g. this answer, or this): $$ e^x < \frac{2 + x}{2 - x} \quad (0 < x < 2). $$ This gives $$ \frac{e}{\phi^2} > \frac{163}{157} = \frac{2 + \tfrac{3}{80}}{2 - \tfrac{3}{80}} > e^{3/80}, $$ whence $$ \log\left(\frac{e}{\phi^2}\right) > \frac{3}{80}, $$ whence $$ \log\phi < \frac{77}{160} = 0.48125. $$ To estimate $\log{10}$ accurately, we need to take an extra term in the power series used in this answer, thus: \begin{gather*} 10 = \frac{65536}{6561}\times\frac{65610}{65536} = \left(\frac43\right)^8\!\!\times\frac{65573 + 37}{65573 - 37} \\ \therefore\ \log10 = 8\log\left(\frac{1 + 1/7}{1 - 1/7}\right) + \log\left(\frac{1 + 37/65573}{1 - 37/65573}\right) \\ < 16\left(\frac17 + \frac{1}{3\cdot7^3} + \frac{1}{5\cdot7^5\cdot(1 - (1/7)^2)}\right) + 2\left(y + \frac{y^3}{3(1 - y^2)}\right), \end{gather*} where $y = 37/65573.$ We have: \begin{multline*} 16\left(\frac17 + \frac{1}{3\cdot7^3} + \frac{1}{5\cdot7^5\cdot(1 - (1/7)^2)}\right) = \frac{16}7 + \frac{16}{3\cdot7^3} + \frac1{3\cdot5\cdot7^3} \\ = \frac{16}7 + \frac{27}{5\cdot7^3} = \frac{3947}{1715}, \end{multline*} and $y < 1/1000,$ therefore $1 - y^2 > 2/3,$ therefore $$ 2\left(y + \frac{y^3}{3(1 - y^2)}\right) < 2y + y^3 < 2y + 10^{-9}. $$ Therefore $$ \log10 < \frac{3947}{1715} + 2y + 10^{-9}. $$

From $1{,}129 \times 65{,}573 = 74{,}031{,}917,$ we have $2y < 0.001129,$ whence $2y + 10^{-9} < 2y + 10^{-6} < 0.00113.$ From $1{,}715 \times 230{,}146 = 394{,}700{,}390,$ we have $$ \frac{3947}{1715} < 2.30146, $$ whence $$ \log{10} < 2.30146 + 0.00113 = 2.30259. $$ Finally, $$ 3\log{10} + \log\phi < 3 \times 2.30259 + \frac{77}{160} = 6.90777 + 0.48125 = 7.38902 < 7.38904 < e^2, $$ whence $$ e^{e^2} > 1000\phi. $$

$\endgroup$
2
  • 2
    $\begingroup$ This one wins for me. Excellent answer. $\endgroup$
    – IAAW
    Commented Feb 17 at 3:53
  • $\begingroup$ Alternative proof of the inequality $e^x < \frac{2 + x}{2 - x}$ (corrected version of an earlier comment): $$ \log\frac{2 + x}{2 - x} = \log\frac{1 + x/2}{1 - x/2} = \sum_{n=0}^\infty\frac2{2n+1}\left(\frac{x}{2}\right)^{2n+1} > x \qquad (0 < x < 2). $$ $\endgroup$ Commented Feb 17 at 11:17
4
$\begingroup$

Since $1.6\color{red}{1}81>\phi$, to show $e^{e^2}>1000\phi$ it is enough to show $$e^2>3\ln10+\ln2+\ln(1-0.19095)$$ We have $$e^2>(2.71828)^2>7.38904$$ $$6.90776>3\ln10$$ $$0.69315>\ln2.$$ And utilising the fifth degree Maclaurin polynomial of $\ln(1-x)$ $$-0.21188>\ln(1-0.19095).$$ Check: $$7.38904\stackrel{?}{>}6.90776+0.69315-0.21188$$

$\endgroup$
4
  • 1
    $\begingroup$ What is the source of the constraint $1.681>\phi?$ $\endgroup$ Commented Feb 19 at 5:38
  • $\begingroup$ tr.m.wikipedia.org/wiki/Alt%C4%B1n_oran @YuriNegometyanov $\endgroup$
    – Bob Dobbs
    Commented Feb 19 at 5:53
  • $\begingroup$ @D.S. Thanks. I did not do it for much attention but challange. Why dont they edit but do jokes when they notice mouse slips? (I am on the phone tough...) $\endgroup$
    – Bob Dobbs
    Commented Feb 19 at 7:38
  • $\begingroup$ @BobDobbs Hmm, I usually don't edit posts myself if the poster is an active user. And when you write @D.S. it doesn't send a notification to me. Instead, @DS works $\endgroup$
    – D S
    Commented Feb 20 at 8:47
4
$\begingroup$

$\color{green}{\textbf{Full Solution.}}$

Should be proven the inequality $$e^2>\ln(1000\varphi),$$ wherein $$10^{\frac16}<1.4678=\dfrac{1+y}{1-y},\quad y=\dfrac {2339}{12339};$$ $$\ln(1000)<18\ln\,\dfrac{1+y}{1-y}<36\left(y+\dfrac13y^3+\dfrac15y^5\dfrac1{1-y^2}\right)<6.907784,$$ $$\,\dfrac{\varphi^2}{e}=\dfrac{\sqrt5+3}{2e}<\dfrac{17393}{18059}=\dfrac{1-z}{1+z},$$ where $z=\dfrac{333}{17726},$ $$2\ln\varphi-1<\ln\,\dfrac{1-z}{1+z}< -2\left(z+\dfrac13z^3+\dfrac15z^5\right)<-0.037576348,$$ and then $$\ln(1000\varphi)<6.90776427+\dfrac12(1-0.037576348) < 7.3889761<2.71827^2<e^2.$$ Therefore, the given inequality is correct too.

$\endgroup$
0
3
$\begingroup$

Here's a very tedious calculation using repeated squaring. I did use a computer to automate and typeset, but in principle it could be performed by hand.

First:

$$e^z = \left(e^{\frac{z}{2^{m}}}\right)^{2^{m}} $$ and (for $y > 0$) $$e^y > 1 + y + \frac{y^2}{2}$$ so working with 24 bits after the point (truncating towards zero):

$$\begin{aligned} e^2 &> \left(1 + \frac{2}{2^{9}} + \frac{1}{2}\left(\frac{2}{2^{9}}\right)^2\right)^{2^{9}} \\ &= \left(1.000000010000000010000000_2\right)^{2^{9}} \\ &\ge \left(1.000000100000001000000001_2\right)^{2^{8}} \\ &\ge \left(1.000001000000100000001010_2\right)^{2^{7}} \\ &\ge \left(1.000010000010000001010100_2\right)^{2^{6}} \\ &\ge \left(1.000100001000001010110001_2\right)^{2^{5}} \\ &\ge \left(1.001000100001010111111010_2\right)^{2^{4}} \\ &\ge \left(1.010010001011010111001100_2\right)^{2^{3}} \\ &\ge \left(1.101001100001001001011011_2\right)^{2^{2}} \\ &\ge \left(10.101101111110000010000101_2\right)^{2^{1}} \\ &\ge \left(111.011000111001010011010111_2\right)^{2^{0}} \\ &= x \\ \end{aligned}$$

Now working with 32 bits after the point:

$$\begin{aligned} e^{e^2} &> e^x \\ &> \left(1 + \frac{x}{2^{12}} + \frac{1}{2}\left(\frac{x}{2^{12}}\right)^2\right)^{2^{12}} \\ &> \left(1.00000000011101100101010010011001_2\right)^{2^{12}} \\ &\ge \left(1.00000000111011001101111111100100_2\right)^{2^{11}} \\ &\ge \left(1.00000001110110101001101011110101_2\right)^{2^{10}} \\ &\ge \left(1.00000011101110001010010111001100_2\right)^{2^{9}} \\ &\ge \left(1.00000111011111110010010010101001_2\right)^{2^{8}} \\ &\ge \left(1.00001111001101100111110001111000_2\right)^{2^{7}} \\ &\ge \left(1.00011111010101000110011100011110_2\right)^{2^{6}} \\ &\ge \left(1.01000010011111100101101100001001_2\right)^{2^{5}} \\ &\ge \left(1.10010110010000100001101101100000_2\right)^{2^{4}} \\ &\ge \left(10.10000100101101011011111111100110_2\right)^{2^{3}} \\ &\ge \left(110.01010111101000101110111010000110_2\right)^{2^{2}} \\ &\ge \left(101000.00111001101000110101010000011001_2\right)^{2^{1}} \\ &\ge \left(11001010010.00010000000001000110101101110111_2\right)^{2^{0}} \\ &> 11001010010.0001_2 \\ &= 1618 + \frac{1}{16} \end{aligned}$$

Using the continued fraction convergents as bounds for $\phi$, and long division, get

$$\begin{aligned} 1000 \phi &< 1000 \times \frac{233}{144} \\&= \frac{233000}{144} \\&= 1618 + \frac{1}{18} \end{aligned}$$

That is,

$$e^{e^2} > 1618 + \frac{1}{16} > 1618 + \frac{1}{18} > 1000 \phi$$

$\endgroup$
1
  • $\begingroup$ Although the calculation is long, it is certainly creative (+1) $\endgroup$
    – D S
    Commented Feb 20 at 8:44
0
$\begingroup$

$\color{brown}{\textbf{What do we prove?}}$

Really, our goal is to prove the inequality $$\ln(\ln(500\cdot(\sqrt5+1)))<2.$$

We can get the intermediate levels from tables:

$$e^3\approx 20.085\,536\,923\,19,$$ $$e^6\approx 403.428\,793\, 432\,7,$$ $$500\cdot(\sqrt5+1)\approx 1618.033\,988\,75,$$ $$\ln(500(\sqrt5+1))\approx 7.388\,967\,104\,042,$$

But these only look like a magician’s passes, which only distract the audience’s attention from the main process.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .