2

Problem:In the mapping $T:\mathbb{R} \rightarrow \mathbb{R}$ defined by $Tx=\begin{cases} x-\frac{1}{2}e^x & \forall x \leq 0 \\ -1/2+1/2x & \forall x>0 \end{cases}$ a contraction?

Definition: A mapping f from a subset A of a normed space E into E is a contraction mapping if there exists a positive number $\alpha<1$ such that $\Vert f(x)-f(y) \Vert \leq \alpha \Vert x-y \Vert \forall x,y \in A$.

Example: Let $f(x)=x+e^{-x}$ as a mapping from $\mathbb{R}^+$ into $\mathbb{R}^+$. For $x,y \in \mathbb{R}^+$, by mean value theorem, $|f(x)-f(y)|<|x-y|$, since $|f'(\xi)|<1$ for all $\xi \in \mathbb{R}^+$. f is not a contraction, since there is no $\alpha<1$ such that $|f(x)-f(y)| \leq \alpha |x-y| \forall x,y \in \mathbb{R}^+$. f does not have a fixed point.

Discussion:This is a piecewise function,so mean value theorem cannot apply here. The traditional method does not work. Cannot make further observations here, so may someone please provide more insights on this problem, thanks!

Update: From @Sven Pistre, For Case 3, by inequality $e^x \geq x+1$ and set $x<0<y$, it's a smooth function, apply mean value theorem, I get $\Vert T(x)-T(y) \Vert= \Vert -1/2+1/2y-x+1/2e^x \Vert \geq 1/2 \Vert (x-y) \Vert$. To show not a contraction, I believe must show $ > \Vert x-y \Vert$. Still unable to figure out. Thanks for any help.

RobPratt
  • 45,619
  • 3
    Can you figure out what happens when you consider three cases? Case 1: $x,y\leq0$; Case 2: $x,y>0$; Case 3: $x<0<y$ In Case 1 and 2 you have a smooth function on the subintervals. In Case 3 you need to do some work. – Sven Pistre Sep 17 '22 at 12:47
  • @SvenPistre I was thinking about that, but was not able to think about how to divide into cases. I will try now. Thanks! – neveryield Sep 17 '22 at 14:26
  • 1
    This should be helpful to you for Case 3 https://math.stackexchange.com/q/504663/243413 Let me know which $\alpha$ you obtain. – Sven Pistre Sep 17 '22 at 23:02
  • @SvenPistre Thanks a lot! Still working on it. – neveryield Sep 18 '22 at 02:05
  • @SvenPistre Still can't get it, I updated my progress. Thanks for your time. – neveryield Sep 18 '22 at 02:59
  • I'm not sure I can supply a "canonical answer" whatever that is. But suppose that for any $\alpha\in(0,1)$ you know how to find $x$ and $y$ with $T$ continuous on $[x,y]$ and differentiable on $(x,y)$ and $T'(z)>\alpha$ for all $z\in (x,y)$. Then you can apply the mean value theorem and conclude that $T$ is not a contraction. – Ali Sep 18 '22 at 14:29

2 Answers2

1

To see that $T$ is not a contraction, consider $y=0$ and $x<0$. Then by l'Hôpital: $$ \begin{align*} \lim_{x\to-\infty}\tfrac{|T(x)-T(0)|}{|x|} &= \lim_{x\to-\infty}\frac{\big|x-\tfrac{1}{2}e^x+\tfrac{1}{2}\big|}{|x|} = \lim_{x\to-\infty}\big|1 + \tfrac{1}{2x}(1-e^x)\big| \\ &= \lim_{x\to-\infty}\big|1 + \tfrac{-e^x}{2}\big| = 1. \end{align*} $$ Therefore, for any $\alpha\in (0,1)$ there is $x_0\in(-\infty,0)$ such that $$ 1 - \tfrac{|T(x_0)-T(0)|}{|x_0|} < 1 - \alpha $$ or, in other words, $$ |T(x_0)-T(0)| > \alpha|x_0|. $$

However, $T$ is still $1$-Lipschitz. To see this you don't actually need to consider three cases since $T$ is, in fact, differentiable at the origin (use the Taylor expansion of $e^x$ around the origin) and its derivative is bounded by $1$ everywhere. So you do get $|T(x)-T(y)|\leq |x-y|$ but as seen above the Lipschitz constant cannot be improved.


The following animation is similar to the one on wikipedia but for your specific function and Lipschitz constant $\alpha=1$. You can see that $T$ is $1$-Lipschitz because its graph stays inside the shaded area.

Graph of 1-Lipschitz function stays inside a cone


For any $\alpha$ strictly smaller than $1$ the graph will always leave the shaded area at some point $x_0$ "if you sail far enough to the west" as Tolkien would say. You can see this in the following animation for $\alpha=0.99$ on the interval $(-700,-690)$.

Graph of function leaves cone

Sven Pistre
  • 1,320
  • I was not able to think about what l’Hopital could bring and the discussion of Lipschitz continuity, thanks a lot! – neveryield Sep 18 '22 at 18:40
  • L'Hôpital and Lipschitz are really unnecessary to answer the question. As was the discussion of case analysis and the inequality linked to. – Ali Sep 19 '22 at 09:33
1

Assume $\alpha\in(0,1)$. For any $z\le\log\left(1-\alpha\right)<0$ we have $\alpha <1-\frac{1}{2}e^{z}=T'(z)$. So let $y=\log\left(1-\alpha\right)$ and $x=y-1$. Then $T$ is continuous on $[x,y]$ and differentiable on $(x,y)$. If it were Lipschitz with constant $\alpha$ then the mean value theorem would imply a point $z$ in $(x,y)$ with $T'(z)\le \alpha$. But we have chosen $x$ and $y$ so that no such point exists. Therefore $T$ cannot be a contraction.

Ali
  • 808