2

As I understand it, many people thought I am asking in a question with exactly this title about pairs of (x,y) such that $x^y=y^x$, or that I am looking for a function $f(x)$, such that $x^{f(x)}=(f(x))^x$, but I am not. I am asking for another function $g(x)=x^{f(x)}$, and where is it concave, convex, its minima, maxima, asymptotes etc. And more importantly, there are two such functions f(x) with the above property and their corresponding functions g(x). Because one $f(x)$ is simply $x$, so for $y=f(x)=x$ we get $g(x)=x^x$, but I am interested in the other function $g(x)$ that is yielded by the rather complicated function $f(x)$.

From the first time I asked the question, I already manage through guesses and observation on individual values come up with these possible properties:

So, with more of a thinking and wolframalpha, I found out (hopefuly this time correctly) that the function is defined only on the interval $(1,\infty)$, goes to infinity for $x$ going from the right to $1$ , it is decreasing in a convex fashion all the way to $x=e$, where is global minimum $e^e$, and it is increasing towards infinity in rather a quite slow fashion, with a possible asymptote $x=y$, to which it is approaching from the left (from the above), which would indicate it is concave somewhere after $x=e$. Can anybody confirm or correct that?

Is anyone able to help me with this, now that I hope we understand each other?

EDIT: The asymptote part is incorrect, as proved by Ennar. Convexity and concavity are still debatable.

supinf
  • 13,433
TStancek
  • 1,027
  • 5
  • 14

2 Answers2

1

It's correct.

Let $y=kx$, where $k>1$. Hence, $$x^y=y^x$$ it's $$x^{kx}=(kx)^x$$ or $$x=k^{\frac{1}{k-1}}$$ and from here $$y=k^{\frac{k}{k-1}}.$$

Thus, $$x^y=k^\frac{k^{\frac{k}{k-1}}}{k-1}.$$ Let $$f(k)=k^\frac{k^{\frac{k}{k-1}}}{k-1},$$ where $k>1.$

Hence, $$f'(k)=\frac{f(k)\cdot k^\frac{k}{k-1}}{(k-1)^3}\cdot\left(\frac{k-1}{\sqrt k}+\ln{k}\right)\left(\frac{k-1}{\sqrt k}-\ln k\right)>0$$ and since $\lim\limits_{k\to1+}f(k)=e^e,$ we obtain $f(k)>e^e$ and we are done!

  • Thank you for your answer, though it might be worth noting that the function you are using and my function are not exactly the same. The distribution of values is slightly different, but global minimum should be the same for both functions. And I guess computing other characteristics of "my" function g(x) is already kind of beyond mortal's possibilities? – TStancek Aug 17 '17 at 06:02
  • @TStancek You want a values of $x^y$ and I proved that with your conditions we have $x^y>e^e$. – Michael Rozenberg Aug 17 '17 at 06:28
  • I agree, and I thanked for that and upvoted your answer, it is just that these two functions are not equal, though some might think that they are. – TStancek Aug 17 '17 at 07:22
  • However "my" function approaches the global minimum from both sides and the function also takes values for $0<k<1$ and you provided it only for one-sided limit. Also, google plot utility plots the graph for values $0<k<\infty$, but nevertheless thank you once again for providing at least some proof of one of my guesses, from here I can see how the otherside of that limit is going to work as well. – TStancek Aug 17 '17 at 07:59
1

TL;DR You are correct.

So, in the question you were already given the link to, it is stated that $$f(x) = -\frac x{\ln x}W\left(-\frac{\ln x}x\right),$$

however, it is bit imprecise, because $W$ is multivalued function, with its branches denoted with $W_0$ and $W_{-1}$ and thus, if you were to plot $f$ this way in some software that has Lambert W-function implemented, you would get the same graph as by plotting $x^y-y^x=0.$

Thus, the first step would be to eliminate the line $y=x$ to get your $g$.

We will use identities

\begin{align} W_0(xe^x) = x,&\ x\geq -1,\\ W_{-1}(xe^x) = x,&\ x\leq -1, \end{align}

to identify which branch in the above definition of $f$ we need to use:

\begin{align} f(x) = x &\iff -\frac x{\ln x}W\left(-\frac{\ln x}x\right) = x\\ &\iff W\left(te^t\right) = t,\ x = e^{-t} \end{align}

By the above identities, we need to use $W_{-1}$ when $t\geq -1$ and $W_{0}$ when $t\leq -1$ to avoid $f(x) = x$. Substituting back with $x$, we get the following definition of $f$:

$$f(x) = \begin{cases} -\frac x{\ln x}W_{-1}\left(-\frac{\ln x}x\right),\ 0<x< e,\\ -\frac x{\ln x}W_{0}\left(-\frac{\ln x}x\right),\ x\geq e, \end{cases}$$

but here is the thing, $W_{-1}\left(-\frac{\ln x}x\right)$ is not defined on whole $(0,e)$. The reason is that $W_{-1}$ is only defined on $[-1/e,0)$, so we need to solve $-1/e \leq -\ln x/x < 0$. We get that $x>1$ so the correct definition of $f$ is

$$f(x) = \begin{cases} -\frac x{\ln x}W_{-1}\left(-\frac{\ln x}x\right),\ \color{blue}{1<x< e},\\ -\frac x{\ln x}W_{0}\left(-\frac{\ln x}x\right),\ x\geq e. \end{cases}$$

Finally, we can now give precise definition of $g$ as well (using $x^y = e^{y\ln x}$):

$$g(x) = \begin{cases} \exp\left(-xW_{-1}\left(-\frac{\ln x}x\right)\right),\ 1<x< e,\\ \exp\left(-xW_{0}\left(-\frac{\ln x}x\right)\right),\ x\geq e. \end{cases}$$


Let us analize what is happening on $(1,e)$. We want to prove that $-xW_{-1}\left(-\frac{\ln x}x\right)$ is strictly decreasing on $(1,e)$. Let $u(x) = W_{-1}\left(-\frac{\ln x}x\right)$. By differentiating $-xu(x)$ we get $$-u(x) + \frac{(\ln x - 1)u(x)}{(1+u(x))\ln x}$$ and we want to prove it is strictly negative on $(1,e)$. Having in mind that $0<\ln x< 1$ and $u(x)\leq -1$ on $(1,e)$, we can manipulate the expression to get

$$-u(x) + \frac{(\ln x - 1)u(x)}{(1+u(x))\ln x} < 0\iff u(x)+\frac1{\ln x} > 0.$$

Now, to see this last thing is true involves some trickery which I will skip. It is not as hard as is very tedious. I can write it down if somebody would be interested in that.

We conclude that $-xu(x)$ is strictly decreasing on $(1,e)$ and so is $g(x)$.


Now, the analysis on $[e,+\infty)$ is much easier. If we denote with $v(x) = W_0\left(-\frac{\ln x}{x}\right)$, then $-1 \leq v(x) < 0$ and $v(x)$ is strictly increasing on $(e,+\infty)$ since both $W_0$ and $-\ln x/x$ are. Calculating derivative of $-xv(x)$ we get

$$-v(x) + \frac{(\ln x - 1)v(x)}{\ln x(1+v(x))}$$

and like the above case, we get that $$-v(x) + \frac{(\ln x - 1)v(x)}{\ln x(1+v(x))}> 0 \iff v(x)\ln x > -1$$ which is true on $(e,+\infty)$ since $v(x)> -1$ and $\ln x> 1$. Thus $-xv(x)$ is strictly increasing on $[e,+\infty)$ and so is $g$.


Having done that, it immediately follows that $g$ has minimum at $x = e$, which is easy to calculate: $g(e) = \exp(-e W_0(-e^{-1})) = \exp(-e\cdot (-1)) = e^e.$

Finally, let us check the limits at boundaries:

$$\lim_{x\to 1^+} g(x) = \exp(\lim_{x\to 1^+} (-xu(x))) = \exp(\lim_{x\to 1^+} (-xW_{-1}(-\frac{\ln x}x))) = +\infty$$ because $\lim_{x\to 0^-}W_{-1}(x) = - \infty.$

To verify that $g(x)\sim x$ at $+\infty$:

$$\lim_{x\to+\infty}\frac{g(x)}{x} = \lim_{x\to+\infty}\frac{\exp(-xv(x))}{x} = \exp(\lim_{x\to+\infty}(-xv(x) - \ln x))$$

and to calculate last limit, we can use Taylor expansion of $W_0$ at $0$:

$$\exp(\lim_{x\to+\infty}(-xv(x) - \ln x)) = \exp(\lim_{x\to+\infty}(-x(-\frac{\ln x}{x}-\frac{\ln^2x}{x^2}+\ldots) - \ln x)) = e^0 = 1.$$

However, $\lim_{x\to +\infty}(g(x)-x) = +\infty$.

Ennar
  • 23,082
  • I know I am not supposed to say things like "Oh boy!", however that is really astonishing from my point of view. But, I believe it is correct. And I once again apologize for earlier, like I said, I did not realize others were confused from the definition stated in the first question, but on the other, how could I know, when they just marked the question as duplicate and didn't say or ask a thing? – TStancek Aug 17 '17 at 14:04
  • @TStancek, no apology needed, it was misunderstanding. Hopefully, this answer of mine shows my good will. Everything should be correct, up to typos, as I've relied heavily on software to do derivations for me. I do not recommend trying this by hand. The only part that needed working is the part that $g$ is strictly decreasing on $(1,e)$. It's surprisingly elusive, compared to its counterpart on $(e,+\infty)$, and I've decided not to write that part, as this is lots of stuff already. But it can be done, and I'll write it if you are interested. – Ennar Aug 17 '17 at 14:15
  • Well, no need adding anything, it is very nice in its current form, not too long and yet enough descriptive. Except maybe, if you were using software, was it able to generate the graph? – TStancek Aug 17 '17 at 14:40
  • 1
    @TStancek, yes, go to https://sandbox.open.wolframcloud.com/ and type Plot[ If[ x < E, Exp[ -x ProductLog[-1,-Log@x/x] ], Exp[ -x ProductLog[0,-Log@x/x] ] ] , {x,1,10}, AxesOrigin->{0,0}]. Also, I've added slight correction at the end. – Ennar Aug 17 '17 at 14:50
  • Oh, I see, no asymptote, but does the $1$ in the ratio limit mean that line $y=ax$ eventually crosses the graph for any $a>1$? – TStancek Aug 17 '17 at 15:11
  • 1
    @TStancek, yes. You can clearly see it this way: $$g(x) < ax \iff \ln a+\ln x+xv(x) > 0,$$ and here's the deal, $\ln x +xv(x) < 0$ on $(e,+\infty)$, but it also approaches $0$, so for any $a>1$, $\ln a>0$ and so $\ln a+\ln x+xv(x)$ will eventually be positive. You can clearly see that it won't happen for $a = 1$. – Ennar Aug 17 '17 at 15:49