29

$f(x)$ is an increasing, differentiable function satisfying $f(x)-f^{-1}(x)=e^{x}-1$ for every real number $x$
I couldn't figure it out whether such function $f(x)$ exists or not.
And if it exists, I want to know the method to find what $f(x)$ is.
Thank you.

Gastly
  • 291
  • 8
    Where does the problem come from? – Martin R Aug 08 '19 at 11:45
  • 2
    Wolfram alpha gives a solution to the differential equation $\frac{dy}{dx} - (\frac{dy}{dx})^{-1}=e^{x}$ in terms of elementary functions, though I'm not sure how to account for when $\frac{dy}{dx}=0$ – BaroqueFreak Aug 08 '19 at 12:03
  • If increasing means strictly increasing i.e. $f'(x)=\frac{dy}{dx}>0$ then you should be be able to convert the ODE into two separable ODEs using the quadratic formula and then take the positive case. – BaroqueFreak Aug 08 '19 at 12:11
  • 3
    @BaroqueFreak: I am not sure if that helps. The derivative of $f^{-1}(x)$ is not $(f'(x))^{-1}$. – Martin R Aug 08 '19 at 12:22
  • 2
    @BaroqueFreak It is increasing and invertible so it must be strictly increasing and thus $f'(x)>0$ for all $x$ – Surb Aug 08 '19 at 12:22
  • Not sure if it helps but it can be proved that $f'(0)-1 = f'(f(0))$ by noting that $f(0)=f^{-1}(0)$. – Surb Aug 08 '19 at 12:32
  • 6
    You have to be more precise: (1) Specify domain and range of $f$. ($f : A \to B$ - what are $A, B$?) (2) I guess $f^{-1}$ is the inverse of $f$. This only makes sense if $f$ is a bijection. ($f^{-1} : B \to A$) (3) Since you form $f(x) - f^{-1}(x)$ you must have $A = B$ (probably $A = B =\mathbb R$). – Paul Frost Aug 08 '19 at 12:39
  • Assuming $f : \mathbb R \to \mathbb R$, both $f, f^{-1}$ must be strictly increasing bijections with positive derivatives. – Paul Frost Aug 08 '19 at 14:27
  • Am I right when thinking that none of the current actually implies the existence of a solution $f\colon\Bbb R\to \Bbb R$? – Surb Aug 08 '19 at 16:07
  • 1
    @Surb If you can prove $f'(0) -1 = f'(f(0))$, then we see that no such $f$ can exist because $f(0) = 0$. – Paul Frost Aug 08 '19 at 22:26
  • I'm betting dollars to donuts that the original question said something like "Suppose that $f$ is increasing, differentiable, and satisfies $f(x) + f^{-1}(x) = e^x - 1$; what can you say about $f'(1)$?", and OP thought "Well, I should find out what $f$ is, differentiate it, and then evaluate the derivative at $1$," rather than, "hunh...maybe there's a way to answer that question without knowing a formula for $f$ itself, perhaps via the derivative-of-the-inverse formula and a little implicit differentiation." – John Hughes Aug 08 '19 at 22:41
  • @PaulFrost $g(x)=f(x)-f^{-1}(x)-e^x$ is a constant function. As $f$ is differentiable, it holds $g'(x)=f'(x)-\frac{1}{f'(f^{-1}(x))}-e^x=0$. In particular $g'(0)=f'(0)-\frac{1}{f'(f^{-1}(0))}-1=0$. Finally, as $f^{-1}(0)=f(0)$, we find that $f'(0)-1=f'(f(0))$. – Surb Aug 08 '19 at 23:57
  • @PaulFrost It seems indeed that there is no solution! You are very welcome to put the pieces together in an answer if you wish. – Surb Aug 09 '19 at 00:00
  • @Surb I can only see that $f'(0) -1 = \frac{1}{f'(0)}$. – Paul Frost Aug 09 '19 at 08:24
  • @PaulFrost you are correct, I realized later in my bed that I made a mistake. Indeed, we can get $f'(0)^2-f'(0)-1=0$ so that $f'(0)=\phi$ but that's all. – Surb Aug 09 '19 at 09:24
  • 2
    @MartinR Well actually I made it myself, to test derivative-of-the-inverse formula. If such function exists, then student can calculate f'(0) via the formula. So I need to confirm the existence of the function, however, I couldn't do it with myself. – Gastly Aug 09 '19 at 16:58
  • 1
    @Gastly Please address the few precisions asked by Paul Frost in this comment about the domain of $f$. I guess this lack of context is motivating some of the (unfortunate) close vote in the question. Note that you don't need to be particularly precise. A simple requirement like: existence of an interval $U$ containing $0$ and $f\colon U \to U$ s.t. ... should be enough. – Surb Aug 10 '19 at 11:39
  • Have you thought about crossposting on MathOverflow? – ViktorStein Aug 11 '19 at 21:25
  • 1
    If we consider a simpler case, e.g. $f(x)-f^{-1}(x)=x$ then is it solvable yet? – C.F.G Aug 13 '19 at 09:52
  • 1
    @Gastly, please include the info from your last comment in the question as it is vital context. – ViktorStein Aug 13 '19 at 15:31
  • @C.F.G I am not sure if this particular case will help much. Anyway, $f(x)=\phi x$ with $\phi=(1+\sqrt{5})/2$ is a differentiable increasing solution over $\Bbb R$. – Surb Aug 14 '19 at 19:12

4 Answers4

23

A general method for such problems is to use power series with undetermined coefficients. We decide to expand the function $f(x)$ as a power series around $x=0$. The equation to solve for $f$ is $$ f(x) - f^{-1}(x) = e^x-1. \tag1 $$ If $x=0$ then the right side is $0$ and thus $f(0)=f^{-1}(0).$ We are given that $f$ is increasing, thus $f(0)=0.$ Our Ansatz now becomes $$ f(x) = a_1 x + a_2 x^2 + a_3 x^3 + \cdots,\quad g(x) := f^{-1}(x) = b_1 x + b_2 x^2 + b_3 x^3 + \cdots. \tag2$$ We can find the power series for $g(x)$ using the Lagrange inversion theorem or by using the identity $f(g(x)) = x$ and solving for the coefficients of $g(x).$ By substituting equation $(2)$ into equation $(1)$ using the expansions of $f$ and $g$ we can solve for $a_1,a_2,a_3,\dots.$ The equation to solve for $a_1$ is $a_1-1/a_1=1$ whose positive solution is $a_1 = \phi := (1+\sqrt{5})/2.$ The rest of the coefficients are solutions of linear equations and the result is $$ f(x) = \phi\, x + \frac{\phi}4 x^2 + \frac{1+7\phi}{72} x^3 + \frac{13+170\phi}{6336} x^4 + \frac{-1279+5003\phi}{950400} x^5 + \cdots. \tag3 $$ There seems to be no obvious pattern to the coefficients. Of course, the radius of convergence is not yet known, and also if the power series is increasing. This is one of the limitations of this method. If the function has a power series, then we can find it, but we don't know much else about the function.


A general method which proves existence uses ideas from a solution to the Grossman's Constant problem. Specifically, the solution by Gabor Nyerges from 2000 in his "The Solution of the Functional Equation $x = (1+F(x))F^2(x)$" available from the Internet Archive Wayback Machine. His solution can be easily generalized to the functional equation $(1)$.

In more detail, consider the Fibonacci sequence $F_n$. The forward recurrence is $\,F_{n+2} = F_{n+1} + F_n\,$ for all integer $n.$ The backward recurrence is $\,F_{n-2} = F_n - F_{n-1}.\,$ There are two linearly independent solutions to the recurrence $\,u_{n+2} = u_{n+1} + u_n.\,$ One for each root of the characteristic polynomial $\,x^2 - x - 1.\,$ One root is the golden ratio $\,\phi := (1+\sqrt{5})/2 > 1\,$ the other is $\,0 > -1/\phi > -1.\,$ The solution for $\phi$ is exponentially increasing as $\,n \to \infty\,$ and also exponentially decreasing to $0$ as $n \to -\infty.\,$ The solution for the other root is exponentially alternating decreasing to $0$ as $\,n \to \infty\,$ and also exponentially alternating increasing as $\,n \to -\infty.\,$ Given two initial values $a_0$ and $a_1,$ the sequence is uniquely determined for all integer $n$ and is a linear combination of the two exponential solutions. As $\,n \to -\infty\,$ and if $\,u_0\,\phi = u_1,\,$ then $u_n$ decreases exponentially to $0,$ but if $\,u_0\,\phi \ne u_1,\,$ then the other solution is dominant and $u_n$ will eventually alternate in sign and become unbounded.

A similar situation arises in this question. Let us generalize the problem. Suppose that we want to construct $f(x)$ which is an increasing bijection on the reals that satisfies $$ f(x) - f^{-1}(x) = s(x) \tag{4} $$ where $\,s(x)\,$ is a differentiable function that satisfies $$ s(0) = 0 \quad \text{ and } \quad s'(x) > 0 \;\forall x. \tag{5} $$ Given $\,0 < x < y\,$ construct a sequence $\,u_n\,$ such that $$ u_0 = x,\quad u_1 = y,\quad u_{n+2} = s(u_{n+1}) + u_n\;\; \forall n\in\mathbb{Z}.\tag{6} $$ Notice that this is the Fibonacci recurrence if $\,s(x) = x.\,$ Notice that $\,u_{-1} = u_1 - s(u_0),\;\;$ $u_{-2} = u_0 - s(u_{-1}),\,$ and so on where we stop the backwards recursion if $\,u_n\,$ ever goes negative. Notice also that $\,u_{2k}\,$ and $\,u_{2k+1}\,$ are increasing sequences iff $\,u_n > 0\;\; \forall n \in\mathbb{Z},\,$ but the sequence $\,u_n\,$ itself may not be. We need to find out exactly when it is monotone increasing.

Now suppose that we know $\,s(x) = c_1\,x + O(x^2)\,$ where $\,c_1 > 1.\,$ By using arguments similar to the Fibonacci recurrence, we find that given $x$ there is a unique value of $\,y\,$ such that the sequence $\,u_n\,$ is such that $\,u_n \to 0\,$ monotonically as $\,n \to -\infty.\,$ Now define $\,f(x)\,$ to be this unique $\,y.\,$ This implies that $\, u_{n+1} = f(u_n)\;\; \forall n\in\mathbb{Z}.\,$ A similar argument holds for $\,x < 0\,$ using the forward recurrence. Thus $\,f(x)\,$ is defined for all reals. Because of equations $(5)$ and $(6)$ it is monotone increasing. Suppose that $\,s(x)\, > L > 0\,$ for $\,x > K > 0\,$ for some $\,L,K\,$ and the same with $>$ replaced with $<$. Then $\,f(x)\,$ is unbounded and hence it is a bijection. Standard $\delta-\epsilon$ methods can prove that $\,f(x)\,$ is differentiable.

Notice that similar arguments construct $\,f(x)\,$ if it is required to be a decreasing function instead of increasing. Notice that if $\,s(x) = c_1\,x\;\; \forall x\in\mathbb{R}\,$ then we have the linear recursion case and $\,f(x) = a_1\,x\,$ where $\,a_1\,$ is a positive root of the characteristic polynomial $\, x^2 - c_1\,x - 1.$

Somos
  • 37,457
  • 3
  • 35
  • 85
  • 1
    I have the impression that one can show that the $a_n$ are positive and $< 1/n!$ (perhaps by induction). This would show convergence. – Paul Frost Aug 10 '19 at 13:42
  • 2
    w00t? The golden ratio surfaces in a problem that isn't just a trivial quadratic equation problem? Wow. :D – The_Sympathizer Aug 18 '19 at 04:00
  • So, if I understand correctly: you define $f$ pointwise. At each point $x$ the value of $f$ is set to be the $y$ for which this "generalized Fibonacci sequence" converges. Then, $f$ satisfies $f^2(x)=s(f(x))+f(x)$ which is indeed equivalent to OP if $f$ is invertible. But it not clear to me how does your method guarantee, differentiability, invertibility and monotonicity? – Surb Aug 19 '19 at 08:02
  • @Surb I added a few more details of the proof. I hope it is enough for you. – Somos Aug 19 '19 at 11:41
14

This is not an answer but only collects some properties which $f$ must neccesarily have.

Let us assume $f : \mathbb R \to \mathbb R$. Then

  1. $f$ is a strictly increasing differentiable bijection.

  2. $f'(x) > 0$ for all $x \in \mathbb R$. [Note that this does not follow from 1. as the example $f(x) = x^3$ shows.]

  3. $f^{-1}$ is a strictly increasing differentiable bijection. We have $$(f^{-1})'(x) = \frac{1}{f'(f^{-1}(x))} > 0 .$$ Note that here it is essential to know that $f'$ does not have zeros.

  4. One has the equation $$f'(x) - \frac{1}{f'(f^{-1}(x))} = e^x .$$

  5. $f(0) = 0$.

  6. $f'(0) = \frac{1+\sqrt{5}}{2} = \phi$, $(f^{-1})'(0) = \frac{1}{\phi} = \frac{-1+\sqrt{5}}{2}$.

  7. $\lim_{x \to \infty}f'(x) = \infty$, $\lim_{x \to \infty}(f^{-1})'(x) = 0$.

  8. If $\lim_{x \to -\infty}f'(x)$ exists, then it has the value $1$.

  9. If $f$ has higher derivatives, we get additional functional equations. They can easily be computed recursively. Let us write $y = f^{-1}(x)$. Then the basic functional equation is $$f(x) - y = e^x -1 .$$ For $n > 0$ we get $$f^{(n)}(x) - y^{(n)} = e^x .$$ The $y^{(n)}$ can be computed recursively. We have $$y' = \frac{1}{f'(y)} ,$$ $$y'' = -\frac{f''(y) \cdot y'}{(f'(y))^2} = -\frac{f''(y)}{(f'(y))^3} ,$$ $$y''' = -\frac{f'''(y) \cdot y' \cdot (f'(y))^3 - f''(y)\cdot 3(f'(y))^2 \cdot f''(y) \cdot y'}{(f'(y))^6} \\ = - \frac{f'''(y) \cdot f'(y) - 3(f''(y))^2 }{(f'(y))^5}$$ etc. In this context see http://vixra.org/pdf/1703.0295v1.pdf which shows that there seems to exist no simple formula for $y^{(n)}$. Anyway, this allows to compute $f^{(n)}(0)$ and $(f^{-1})^{(n)}(0)$ recursively by noting that for $x= 0$ we have $y = 0$. This gives a linear equation for $f^{(n)}(0)$ whose solution can be expressed by the collection of $f^{(i)}(0)$ with $i < n$. Once we have determined the $f^{(i)}(0)$ with $i \le n$, we get $(f^{-1})^{(n)}(0)$ by inserting these values into the above equation for $y^{(n)}$. For example we get $$f''(0) = \frac{\phi^3}{1+\phi^3} = \frac{\phi}{2} , (f^{-1})''(0) = -\frac{\frac{\phi}{2}}{\phi^3} = -\frac{1}{2\phi^2} .$$

Let us give proofs.

  1. In order to have an inverse $f^{-1}$ the function $f$ must be injective (in which case $f^{-1}$ is defined on $f(\mathbb R)$) and in order that $f(x) + f^{-1}(x)$ is defined for all $x \in \mathbb R$ the function $f$ must be surjective.

  2. Since $f$ is increasing, we have $f'(x) \ge 0$ for all $x \in \mathbb R$. Replacing $x$ by $f(x)$ the functional equation yields $$f(f(x)) - x = f(f(x)) - f^{-1}(f(x)) = e^{f(x)} -1 .$$ Differentiating gives $$f'(x) \cdot f'(f(x)) - 1 = f'(x)\cdot e^{f(x)} .$$ This shows that $f'(x) \ne 0$ for all $x \in \mathbb R$.

  3. $f^{-1}$ is trivially a bijection. Since $f'$ does not have zeros, it is differentiable with $(f^{-1})'(x) = \frac{1}{f'(f^{-1}(x))} > 0$. Hence $f^{-1}$ is increasing. Note that this also follows easily from the fact that $f$ is increasing (without using differentiabilty).

  4. This is obtained by differentiating the functional equation.

  5. We have $f(0) = f^{-1}(0)$. Assume $f(0) > 0$. Then $0 = f^{-1}(f(0)) > f^{-1}(0)$ which is impossible. Similarly $f(0) < 0$ is impossible.

  6. By 4. $$f'(0) - \frac{1}{f'(0)} = f'(0) - \frac{1}{f'(f^{-1}(0))} = e^0 = 1 .$$ This implies (note $f'(0) > 0$) $f'(0) = \frac{1+\sqrt{5}}{2} = \phi$ and $(f^{-1})'(0) = \frac{1}{f'(0))} = \frac{1}{\phi}$.

  7. This follows from $f'(x) = e^x + \frac{1}{f'(f^{-1}(x))}$ and 3.

  8. We have $\lim_{x \to -\infty}(f'(x) - \frac{1}{f'(f^{-1}(x))}) = 0$. Assume that $a = \lim_{x \to -\infty}f'(x)$ exists, $0 \le a \le \infty$. Then also $\lim_{x \to -\infty}f'(f^{-1}(x)) = a$ since $f^{-1}$ is an increasing bijection. $a = 0$ is impossible since then $\lim_{x \to -\infty}(f'(x) - \frac{1}{f'(f^{-1}(x))}) = -\infty$. $a = \infty$ is impossible since then $\lim_{x \to -\infty}(f'(x) - \frac{1}{f'(f^{-1}(x))}) = \infty$. Hence $0 < a < \infty$ and $a - \frac{1}{a}= 0$ which yields $a= 1$.

Paul Frost
  • 87,968
  • 1
    I like very much this post. If I may suggest a small improvement. Could you please collect all properties in a small list at the beginning? (Something like a Proposition and then a proof). Makes it easier to have a clear idea of what is currently know. – Surb Aug 10 '19 at 00:18
  • 1
    @Surb Okay, done! – Paul Frost Aug 10 '19 at 09:35
  • It seems that by continuing to differentiate we could get $f^{(n)}(0)$ for all $n$. But I can't see a pattern to be honest. I get $f'''(0)=\frac{1}{24}(9 + 7 \sqrt{5})$. – Surb Aug 12 '19 at 22:38
  • @Surb You are right. That is what Somos noticed in his answer. But who knows whether a solution, if it exists, is $C^\infty$? – Paul Frost Aug 13 '19 at 07:42
  • I see, but as you noted under Somos post, if it can be shown that the series converge on a neighborhood of $0$, then this would show existence of a solution, right? – Surb Aug 13 '19 at 08:14
  • 1
    @Surb This would be only the first step. If we have shown convergence, we only get a candidate for a solution. We have to show that the resulting function is increasing and satisfies the functional equation (which is certainly non-trivial for $x \ne 0$). – Paul Frost Aug 13 '19 at 08:33
  • @Somos I corrected it. 7. was in fact wrong, I wanted to consider $x \to \infty$. – Paul Frost Aug 15 '19 at 15:34
8

If $f(x) = x+k (e^x-1)$, $0<k<1$, then $f^{-1}(x) = k+x-W(ke^{k+x})$, $W$ is the Lambert W function.

enter image description here

If $f(x)- f^{-1}(x)= e^x-1$, then $W(ke^{k+x})=k+(1-k)(e^x-1)$. Then plugging into $f^{-1}(x)$ gives $f^{-1}(x)= k+x-W(ke^{k+x})=k+x-(k+(1-k)(e^x-1))=x-(1-k)(e^x-1)$, which doesn't seem to be right because now $f^{-1}(f(x))\neq f(f^{-1}(x))\neq x$. I think a closed form for the solution is difficult to find, but maybe (?) you can choose a $k$, generate some points and interpolate. Good luck, my friend!

Ahmed Hossam
  • 1,171
  • 2
    If there is a solution of this form, then neccesarily $k = \phi -1 = \frac{-1+\sqrt{5}}{2}$ since $f'x) = 1 +ke^x$ and $f'(0) = \phi$ (see my answer). – Paul Frost Aug 10 '19 at 10:34
  • 2
    The approach fails. You must have $k = \phi -1$. But you also get $f''(0)=k = \phi-1$ which differs from the value $\frac{\phi^3}{1+\phi^3}$ (see my answer). – Paul Frost Aug 10 '19 at 12:57
  • Yes, I saw that already. If there is a $k$, then it can't be a constant. – Ahmed Hossam Aug 10 '19 at 17:00
4

Calling $f_n(x) = \sum_{k=0}^n a_k x^k$ as an $n-$degree approximation we have

$$ f^{-1}(x) = f(x) -e^x+1 $$

or

$$ x = f\left(f(x)-e^x+1\right) $$

or putting the approximation

$$ x = f_n\left(f_n(x)-\sum_{j=1}^n\frac{x^j}{j!}\right) $$

now expanding and equating to zero the polynomial coefficients we have

$$ \left\{ \begin{array}{rcl} -a_1^2+a_1+1 &=&0\\ -a_2 a_1^2+a_2 a_1+\frac{a_1}{2}-a_2 &=&0\\ -a_3 a_1^3+3 a_3 a_1^2-2 a_2^2 a_1+a_2 a_1-4 a_3 a_1+\frac{a_1}{6}+2 a_2^2-a_2+a_3 &=&0\\ \vdots& = &0 \end{array} \right. $$

Attached a MATHEMATICA script to determine the polynomial coefficients

f[x_, n_] := Sum[Subscript[a, k] x^k, {k, 1, n}]
n = 5;
pol = x - f[f[x, n] - Sum[x^k/k!, {k, 1, n}], n];
coefs = Take[CoefficientList[pol, x], {2, n + 1}];
For[k = 1; sols = {}; j = 2, k <= n, k++,
  sol = Solve[(coefs[[k]] /. sols) == 0, Subscript[a,k]][[j]];
  j = 1;
  AppendTo[sols, sol];
  sols = Flatten[sols]
]
A = Table[Subscript[a, k], {k, 1, n}];
A /. sols

NOTE

For $n=2$ we have

$$ x-a_2 \left(a_2 x^2+a_1 x-\frac{x^2}{2}-x\right){}^2-a_1 \left(a_2 x^2+a_1 x-\frac{x^2}{2}-x\right)=0 $$

or

$$ 0\times 1+(1+a_1-a_1^2)x + (-a_2 a_1^2+a_2 a_1+\frac{a_1}{2}-a_2)x^2 + (-2 a_1 a_2^2+2 a_2^2+a_1 a_2-a_2)x^3+(-a_2^3+a_2^2-\frac{a_2}{4})x^4 \equiv 0 $$

but for $n=2$ we collect only the $n+1$ first coefficients which are

$$ \{0,1+a_1-a_1^2,-a_2 a_1^2+a_2 a_1+\frac{a_1}{2}-a_2\}= 0 $$

enter image description here

$$ f(x) = \frac{1}{2} \left(1+\sqrt{5}\right) x+\frac{1}{8} \left(1+\sqrt{5}\right) x^2+\frac{1}{144} \left(9+7 \sqrt{5}\right) x^3+\frac{\left(98+85 \sqrt{5}\right) x^4}{6336}+\frac{\left(2445+5003 \sqrt{5}\right) x^5}{1900800}+\frac{\left(18257 \sqrt{5}-2260\right) x^6}{30067200}+\frac{\left(192234608 \sqrt{5}-32376225\right) x^7}{1069610572800}+\frac{\left(15850653103 \sqrt{5}-45039491325\right) x^8}{650323228262400}+\cdots $$

Cesareo
  • 36,341