5

May somebody help me to correctly calculate the dervative of the $n$-th power tower function?

$$ \begin{align} f_1(x)&=x\\ f_n(x)&=x^{f_{n-1}(x)}\\ &=x^{x^{x^{...^x}}}\text{ where }x\text{ occurs }n\text{ times} \end{align} $$

The solution given here is for the infinite case $f_{\infty}=\lim_{n\to\infty}f_n$.

String
  • 18,395
  • 6
    What is wrong with the solution found there? I did not find any step that did not make sense, but I may have overlooked something. Still, it would surprise me why a professor in Mathematics at Clark University should provide an incorrect analysis of the problem. – String Apr 24 '15 at 19:58
  • they say that y=x^ybut in fact if the power ups by n times it will be Y(n)=x^Y(n-1) – Razvan Paraschiv Apr 24 '15 at 20:04
  • It is an infinite power up, that is a limit. You cannot use $n$ here. – Hasan Saad Apr 24 '15 at 20:05
  • 3
    The $\ldots$ implies that this is an infinite tower, which is what David Joyce computed the derivative of. – Robert Israel Apr 24 '15 at 20:05
  • @RazvanParaschiv what is $1 - 0.999...$? – Ben Grossmann Apr 24 '15 at 20:07
  • 1
    What is your concept of infinity, Razvan? What you say is true of any finite number of powers $n$, but not when infinity is involved, since $\infty-1=\infty$ if we are to make sense of such statement. – String Apr 24 '15 at 20:07
  • This is what i was told by maths teacher and im studying at the best high school here. He told me that when i take out an "x " there will remain n-1 powers.Dont know guys. i was just asking :D – Razvan Paraschiv Apr 24 '15 at 20:11
  • If you are asking for the derivative of a finite power tower, that is something different @RazvanParaschiv. Is that actually, what you are after? Then the question should be rephrased to reflect that! – String Apr 24 '15 at 20:12
  • Probably i misunderstood the task.Can you help me with that problem? Thank you and sorry for my mistake! – Razvan Paraschiv Apr 24 '15 at 20:15
  • @RazvanParaschiv: Is my last edit correctly reflecting your question? – String Apr 24 '15 at 20:20
  • yes ,indeed.Thank you for your help sir! – Razvan Paraschiv Apr 24 '15 at 20:23
  • Something fun is to make a DL at 0 with this function ! :P – ParaH2 Apr 24 '15 at 20:23
  • See this answer: http://math.stackexchange.com/a/1032089/166413 – Pauly B Apr 24 '15 at 22:36
  • @RazvanParaschiv: Would you consider accepting one of the answers? Best regards, String. – String Apr 26 '15 at 17:33
  • thank you for the link! It is really useful and ,indeed, the answer which the owner has consider correct is in fact correct! thank you, sir for your help! – Razvan Paraschiv Apr 26 '15 at 20:23

3 Answers3

3

My approach would be to try to find derivatives inductively. First we have $f_1'(x)=1$. Then differentiating $f_2(x)=x^{f_1(x)}$ we have $$ \log(f_2(x))=f_1(x)\log x\implies\frac{f_2'(x)}{f_2(x)}=f_1'(x)\log x +\frac{f_1(x)}{x}=1+\log x $$ so that $f_2'(x)=f_2(x)(1+\log x)=x^x(1+\log x)$ and more generally $$ f_n'(x)=f_n(x)\cdot\left(f_{n-1}'(x)\log x+\frac{f_{n-1}(x)}{x}\right) $$ Then perhaps looking at $f_3',f_4'$ some kind of pattern emerges that can be proven inductively. But perhaps someone comes up with something smarter.


The notation gets messy, so let us instead write $a_n:=f_n(x)$ and $b_n:=f_n'(x)$ together with $c:=\log x$ and $d:=1/x$ to have $$ b_n=a_n(c\cdot b_{n-1}+d\cdot a_{n-1})\\ $$ so with this we get $$ \begin{align} b_2&=a_2(c\cdot b_1+d\cdot a_1)\\ b_3&=a_3(c\cdot b_2+d\cdot a_2)\\ &=a_3(c\cdot a_2(c\cdot b_1+d\cdot a_1)+d\cdot a_2)\\ &=c^2\cdot a_3a_2b_1+cd\cdot a_3a_2a_1+d\cdot a_3a_2\\ b_4&=a_4(c\cdot b_3+d\cdot a_3)\\ &=a_4(c\cdot(c^2\cdot a_3a_2b_1+cd\cdot a_3a_2a_1+d\cdot a_3a_2)+d\cdot a_3)\\ &=c^3\cdot a_4a_3a_2b_1+c^2d\cdot a_4a_3a_2a_1+cd\cdot a_4a_3a_2+d\cdot a_4a_3 \end{align} $$ and maybe a pattern has emerged. It looks like we have, noting that $b_1=1$ and $d\cdot a_1=1$ and defining $a_0=1$: $$ \begin{align} b_n=&c^{n-1}d\cdot a_na_{n-1}\cdots a_2a_1a_0+c^{n-2}d\cdot a_n\cdots a_1\\ &+c^{n-3}d\cdot a_n\cdots a_2+...+d\cdot a_n a_{n-1}\\ =&\sum_{i=1}^n c^{n-i}d\cdot \prod_{j=i-1}^n a_j \end{align} $$


Let us try to prove this. The base cases $n=1,2,3,4$ have already been computed above. So let us turn to the inductive step: $$ \begin{align} b_{n+1}&=a_{n+1}(c\cdot b_n+d\cdot a_n)\\ &=a_{n+1}\left(c\cdot\left(\sum_{i=1}^n c^{n-i}d\cdot \prod_{j=i-1}^n a_j\right) +d\cdot a_n\right)\\ &=\left(\sum_{i=1}^n c^{n+1-i}d\cdot \prod_{j=i-1}^{n+1} a_j\right)+d\cdot a_{n+1}a_n\\ &=\sum_{i=1}^{n+1} c^{n+1-i}d\cdot \prod_{j=i-1}^{n+1} a_j \end{align} $$ which proves the claim.


Substituting back the values of $a_n,b_n,c$ and $d$ we then have shown that $$ \begin{align} f_n'(x)&=\sum_{i=1}^n (\log x)^{n-i}\frac 1x\cdot \prod_{j=i-1}^n f_j(x)\\ &=\sum_{i=1}^n (\log x)^{n-i}\cdot x^{-1+\prod_{j=i-1}^n f_{j-1}(x)} \end{align} $$ where $f_0(x)=1$ and $f_{-1}(x)=1$ have been added.

String
  • 18,395
  • i have riched this point and he asked me"then ,who is f'(n-1).Can you provide a general form?" – Razvan Paraschiv Apr 24 '15 at 20:32
  • These derivatives get more and more complicated. So this method is not advisable to compute the derivative for $f_\infty$. – GEdgar Apr 24 '15 at 20:34
  • 1
    @GEdgar: But we are after $f_n'$ and not $f_\infty'$. – String Apr 24 '15 at 20:35
  • ... and each $f_n'$ is more complicated than the previous one. You can do as many as you have patience for as shown here. But do not hope for a general formula in terms of $n$. – GEdgar Apr 24 '15 at 20:36
  • i tried to calculate the last 4 terms with the hope that ill find out a general form,but unluckily i got stucked.Might there be any documents with the result for this exercise? – Razvan Paraschiv Apr 24 '15 at 20:36
  • And once again my answer get downvoted although I am more than willing to hear constructive feedback instead. Downvoting should be saved for those who show careless behaviour, not sincere posters, I think. – String Apr 24 '15 at 20:37
  • OK, the downvote disappeared. No harm done! – String Apr 24 '15 at 20:38
  • @GEdgar: I agree! I have not found anything appealing about those expressions yet in terms of some "nice" inductive pattern. – String Apr 24 '15 at 20:39
  • really good!Big up! – Razvan Paraschiv Apr 24 '15 at 22:00
  • I was wondering why did you assume that a0 is equal to 1? – Razvan Paraschiv Apr 26 '15 at 20:18
  • @RazvanParaschiv: To make complete the pattern. We had the first term in the examples $b_2,b_3,b_4$ to be of the form $$c^{n-1}\cdot a_n\cdot a_{n-1}\cdots a_2\cdot b_1$$ but I really wanted it to match the form of the other terms, namely getting rid of $b_1$, adding $d$ as a factor, and adding $a_1,a_0$ as factors. Then all terms follow the same pattern. Luckily $b_1=1$ could just be left out, and $d=1/x$ and $a_1=x$ made it possible to multiply by $d\cdot a_1$ without changing it. Then defining $a_0=1$ made it all fit. – String Apr 26 '15 at 21:03
  • i finally understood.You have defined a0 as 1 to make it legit.Thank you ,indeed! – Razvan Paraschiv Apr 27 '15 at 13:58
2

Write $$ f_{-1}(x)=0;\qquad f_n(x) = x^{f_{n-1}(x)} $$ so that $f_0(x)=1$ and the higher ones are the same as before. Then $$ \frac{d}{dx} f_n(x) = \sum_{k=1}^n x^{\left(-1+\sum_{j=n-1-k}^{n-1}f_j(x)\right)}\log^{k-1}(x) $$

GEdgar
  • 111,679
  • your statement from the first part might be wrong as there cant be x^x^x... by ba negative number of times – Razvan Paraschiv Apr 24 '15 at 21:05
  • @RazvanParaschiv: $f_{-1}(x)=0$ gives $f_0(x)=x^0=1$ and so $f_1(x)=x^1=x$ such that $f_2(x)=x^x$ and so on, so it works just fine. – String Apr 24 '15 at 21:06
  • Does $\log^{k-1}(x)$ mean $(\log x)^{k-1}$ rather than $\log\log\log...\log x$ in your formula. I suppose it does. Could you provide any details as to how you derived it? – String Apr 24 '15 at 21:08
  • Yes, $\log^{k-1}(x) = (\log x)^{k-1}$. How did I do it? I had Maple do the case $n=6$ and observed the pattern. Given the formula, can you prove it by induction? – GEdgar Apr 24 '15 at 21:16
  • @GEdgar: I will have a look :o) Great work! – String Apr 24 '15 at 21:19
  • @GEdgar: I think I managed to do it. Would you mind to have a quick look? – String Apr 24 '15 at 21:57
2

let ${}^nx$ = $x\uparrow\uparrow n =\underbrace {x^{x^{\cdot ^{\cdot ^{x}}}}}_{n}$ the nth tetration of x

Our goal: Is to find $\displaystyle{d\over dx}(^nx)$

Calculating some derivatives at first to try understand the patterns:

$\displaystyle{d\over dx}(^2x) = x^x(1+\ln x)$

$\displaystyle{d\over dx}(^3x) = (^3x)(^2x)\Big({1\over x}+\ln x+(\ln x)^2\Big)$

$\displaystyle{d\over dx}(^4x) = (^4x)(^3x)\Big({1\over x}+{^2x\over x}\ln x+{}^2x(\ln x)^2+{}^2x(\ln x)^3\Big)$

Now, we begin to notice a pattern involving $\ln x$. Just computing one more expression

$\displaystyle{d\over dx}(^5x) = (^5x)(^4x)\Big({1\over x}+{^3x\over x}\ln x+{({}^3x)({}^2x)\over x}(\ln x)^2+({}^3x)({}^2x)(\ln x)^3+({}^3x)({}^2x)(\ln x)^4\Big)$


Computing the general derivative

Since $\displaystyle {}^nx = x^{\big({}^{(n-1)}x\big)} = e^{{}^{(n-1)}x\ln x}$

Therefore $${d\over dx}({}^nx) = {d\over dx}(e^{{}^{(n-1)}x\ln x}) = e^{{}^{(n-1)}x\ln x}{d\over dx}\big({}^{(n-1)}x\ln x\big) = {}^nx\Big({{}^{(n-1)}x\over x}+\ln x {d\over dx}\big({}^{(n-1)}x\big)\Big)$$

Now try to cast the general form of the derivative to resemble the pattern in the calculated derivatives, you can see they come in the form: $${d\over dx}({}^nx) = ({}^nx)({}^{(n-1)}x)\Big({1\over x} + \cdots \Big)$$

Now, we could make $${d\over dx}({}^nx) = ({}^nx)({}^{(n-1)}x)\Big({1\over x} + {{d\over dx}\big({}^{(n-1)}x\big)\over {}^{(n-1)}x}\ln x\Big)$$

Notice that the derivative always starts with two tetration factors; so that $${{d\over dx}\big({}^{(n-1)}x\big)\over {}^{(n-1)}x{}^{(n-2)}x} = B(n) = \text{The big bracket}$$

Now, we could rewrite the derivative as $${d\over dx}({}^nx) = ({}^nx)({}^{(n-1)}x)\Big({1\over x} + {}^{(n-2)}x\ln x B(n-1)\Big)$$

$B(n)$: the big bracket can also be defined recursively as $$B(n) = {1\over x} + {}^{(n-2)}x\ln x B(n-1), \text{ if } n>3$$ $$B(3) = {1\over x} + \ln x + (\ln x)^2$$


$$\begin{align}{d\over dx}({}^nx) &= ({}^nx)({}^{(n-1)}x)B(n) \\ &= ({}^nx)({}^{(n-1)}x)\Big({1\over x} + {}^{(n-2)}x\ln x B(n-1)\Big) \\ &= ({}^nx)({}^{(n-1)}x)\Bigg({1\over x} + {}^{(n-2)}x\ln x \Big({1\over x} + {}^{(n-3)}x\ln x B(n-2)\Big)\Bigg) \\ &= ({}^nx)({}^{(n-1)}x)\Bigg({1\over x} + {{}^{(n-2)}x\over x}\ln x + {}^{(n-2)}x{}^{(n-3)}x(\ln x)^2 B(n-2)\Big)\Bigg) \\ &= ({}^nx)({}^{(n-1)}x)\Bigg({1\over x} + {{}^{(n-2)}x\over x}\ln x + {}^{(n-2)}x{}^{(n-3)}x(\ln x)^2 \Big({1\over x} + {}^{(n-4)}x\ln x B(n-3)\Big)\Bigg) \\ &= ({}^nx)({}^{(n-1)}x)\Bigg({1\over x} + {{}^{(n-2)}x\over x}\ln x + {{}^{(n-2)}x{}^{(n-3)}x\over x}(\ln x)^2 + {}^{(n-2)}x{}^{(n-3)}x{}^{(n-4)}x(\ln x)^3 B(n-3)\Bigg) \\ &= ({}^nx)({}^{(n-1)}x)\Bigg({1\over x} + {{}^{(n-2)}x\over x}\ln x + {{}^{(n-2)}x{}^{(n-3)}x\over x}(\ln x)^2 + \cdots + \Big(\prod^i_{j=2}{{}^{n-j}x}\Big){(\ln x)^{i-1}\over x} + \cdots + \Big(\prod^{n-2}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-3)}}B(3)\Bigg) \\ &= ({}^nx)({}^{(n-1)}x)\Bigg({1\over x} + {{}^{(n-2)}x\over x}\ln x + {{}^{(n-2)}x{}^{(n-3)}x\over x}(\ln x)^2 + \cdots + \Big(\prod^i_{j=2}{{}^{n-j}x}\Big){(\ln x)^{i-1}\over x} + \cdots + \Big(\prod^{n-2}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-3)}} \Big({1\over x} + \ln x + (\ln x)^2\Big) \Bigg) \\ \end{align}$$

Now, all terms in $B(n)$ follow the same general form: $\displaystyle\Big(\prod^i_{j=2}{{}^{n-j}x}\Big){(\ln x)^{i-1}\over x}$, where $i$ is the index of the term, except the last 3 terms but notice the following similarity:

The third last term equals $\displaystyle\Big(\prod^{n-2}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-3)}\over x}$ which is the same as its general formula

The second last term equals $$\Big(\prod^{n-2}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-3)}}\cdot \ln x = ({}^1x)\Big(\prod^{n-2}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-2)}\over ({}^1x)} = \Big(\prod^{n-1}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-2)}\over x}$$ again it is the same as its general formula

The last term equals $$\Big(\prod^{n-2}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-3)}}\cdot (\ln x)^2 = ({}^0x)({}^1x)\Big(\prod^{n-2}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-1)}\over ({}^0x)({}^1x)} = \Big(\prod^{n}_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(n-1)}\over x}$$ again it is the same as its general formula

Note that $\displaystyle ({}^0x) = 1, ({}^1x) = x$


Now we deduce that

$$\begin{align}{d\over dx}({}^nx) &= ({}^nx)({}^{(n-1)}x)\sum_{i=1}^{n}{\Bigg[\Big(\prod^i_{j=2}{{}^{n-j}x}\Big){(\ln x)^{(i-1)}\over x}\Bigg]} \\ &= {({}^nx)({}^{(n-1)}x)\over x}\sum_{i=1}^{n}{\Bigg[{(\ln x)^{(i-1)}}\prod^i_{j=2}{{}^{n-j}x}\Bigg]}\blacksquare \end{align}$$