12
$\begingroup$

In the book "Chaos, Fractals and Noise - Stochastic Aspects of Dynamics" from Lasota and Mackey the operator $P: L^1[0,1] \to L^1[0,1]$
$$ (Pf)(x) = \frac{1}{4\sqrt{1-x}} \left[ f\left(\frac{1}{2}\left(1-\sqrt{1-x}\right)\right) + f\left(\frac{1}{2}\left(1+\sqrt{1-x}\right)\right)\right]$$ was considered, where $L^1$ denotes the space of Lebesgue integrable functions.

It is stated in the book that Ulam and Neumann showed that $$ f^*(x) =\frac{1}{\pi\sqrt{x(1-x)}} $$ solves $$ Pf^* = f^*,$$ in the article "On combination of stochastic and deterministic processes, Bull. Amer. Math. Soc. vol. 53 (1947) p. 1120.". I went to the library of my university and I actually found the book "vol 53 of Bull. Amer. Math. Soc.", but this book only contains the abstract of the article, not the article itself.

So my question is: Where do I find the article - or how did they find $f^*$ ?

Edit: I am interested how Ulam and Neumann actually constructed $f^*$.

$\endgroup$
  • 6
    $\begingroup$ To answer the question in the title, they were both supersmart. $\endgroup$ – Asaf Karagila Sep 2 '14 at 20:55
  • $\begingroup$ The problem is that the citation is not actually to an article. The page cited falls amid a list of abstracts for the Summer Meeting of the AMS in 1947. So that citation regards Ulam and Neumann's announcement of their work, not the publication itself. $\endgroup$ – Semiclassical Sep 3 '14 at 0:20
  • $\begingroup$ Maybe trial and error? $\endgroup$ – Thomas Sep 3 '14 at 6:18
5
+50
$\begingroup$

This is just a speculation so I'm not sure they found the solution this way, but this is 'doable' in my opinion. The idea comes from observing the $x$ that $S^n(x)$ changes its increasing\decreasing state. ($S(x)=4x(1-x)$ in the book)

$$S(1/2)=1\\S^2(\frac{2-\sqrt{2}}{4})=S^2(\frac{2+\sqrt{2}}{4})=1\\S^3(\frac{2-\sqrt{2-\sqrt{2}}}{4})=S^3(\frac{2+\sqrt{2-\sqrt{2}}}{4})=S^3(\frac{2-\sqrt{2+\sqrt{2}}}{4})=S^3(\frac{2+\sqrt{2+\sqrt{2}}}{4})=1\\\vdots$$

Which can be rewritten as

$$S(\sin^2\frac{\pi}{4})=1\\S^2(\sin^2\frac{\pi}{8})=S^2(\sin^2\frac{3\pi}{8})=1\\S^3(\sin^2\frac{\pi}{16})=S^3(\sin^2\frac{3\pi}{16})=S^3(\sin^2\frac{5\pi}{16})=S^3(\sin^2\frac{7\pi}{16})=1\\\vdots$$

Now assume $Pf=f$. From an intuitive viewpoint, the transformation $P$ sends $S^{-1}(x)$ to $x$. So we get the idea to define a function $g : [0,1]\rightarrow [0,1]$ as$$g(x)=f(\sin^2\frac{\pi x}{2})$$ Then the functional equation of $f$ becomes much simpler. $$g(x)=f(\sin^2\frac{\pi x}{2})=(Pf)(\sin^2\frac{\pi x}{2})=\frac{f(\sin^2\frac{\pi x}{4})+f(\sin^2(\frac{\pi}{2}-\frac{\pi x}{4}))}{4\cos\frac{\pi x}{2}}=\frac{g(\frac{x}{2})+g(1-\frac{x}{2})}{4\cos\frac{\pi x}{2}}$$ It is now easy to see $$g(\frac{x}{2})\sin\frac{\pi x}{2}+g(1-\frac{x}{2})\sin\pi(1-\frac{x}{2})=4g(x)\sin\frac{\pi x}{2}\cos\frac{\pi x}{2}=2g(x)\sin\pi x$$ Putting $h(x)=g(x)\sin\pi x$, we get $$\frac{1}{2}(h(\frac{x}{2})+h(1-\frac{x}{2}))=h(x)$$ We will now see that the only possible continuous function $h$ is the constant function. Using this equation repetitively, $$h(x)=\frac{1}{2^n}\Big(\sum_{k=1}^{2^{n-1}-1}\big(h(\frac{k}{2^{n-1}}+\frac{x}{2^n})+h(\frac{k}{2^{n-1}}-\frac{x}{2^n})\big)+h(\frac{x}{2^n})+h(1-\frac{x}{2^n})\Big)$$ If x is irrational, each intervals $(\frac{t}{2^n}, \frac{t+1}{2^n})$ contain exactly one of the $\frac{k}{2^{n-1}}\pm\frac{x}{2^n}$ or $\frac{x}{2^n}$ or $1-\frac{x}{2^n}$. (More specifically, $\frac{k}{2^{n-1}}+\frac{x}{2^n}\in(\frac{2k}{2^n}, \frac{2k+1}{2^n})$ and similarly for others.) So consindering the partition $0, \frac{1}{2^n}, \frac{2}{2^n}, \cdots, 1$ the above is a riemann sum. Because $h$ is a continous function, the sum converges to the Riemann integral which makes $$h(x)=\int_{0}^{1}h(t)dt=c$$ for all irrational $x$. Since the irrational numbers in $[0,1]$ are dense in $[0,1]$, $h(x)=c$ for all $[0,1]$ by the continuity of $h$.

So we finally have $h(x)=c$ and from this we can get $$f(x)=\frac{c}{\sqrt{x(1-x)}}$$

Putting $c=1/\pi$ makes the function $f$ a proabability density function on $[0,1]$.

Update. I've just found out that this was actually a method called the 'change of variables'. It is illustrated Ch6.5 in the OP's book and the idea is to transform $S$ to an exact function $T=g\circ S\circ g^{-1}$ in order to make it more easy to manipulate.(in this case, to the tent function) According to the book's theorem 6.5.2, if we let $T:(0,1)\rightarrow (0,1)$ a measurable nonsingular transformation, $\phi\in D((a,b))$ a positive density, and $S:(a,b)\rightarrow (a,b)$ defined as $S=g^{-1}\circ S\circ g$ with $$g(x)=\int_{a}^{x}\phi (y)dy$$ then $T$ is exact iff $S$ is statistically stable and $\phi$ is the measure invariant to $S$. So the method only requires to find such $\phi$.

The problem is that this idea was originally due to Ruelle[1977] and Pianigiani[1983] about 30 years later which means that it's not likely for this method that was used by Ulam and Neumann to find $f^*$. So I guess my answer becomes "Well, there is a method for finding such function. But I'm not sure how they found it...".

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Very nice work! Can you elaborate a bit why $h(x)$ must be a constant function? (It obviously works, but the implied uniqueness isn't so evident.) $\endgroup$ – Semiclassical Sep 4 '14 at 12:24
  • $\begingroup$ Thank you for your answer. Could you explain how you got $1+ \sqrt{1-sin^2( \frac{2x\pi}{4})} = sin^2( \frac{\pi}{2} -\frac{\pi x}{4})$? When I follow your explenation I end up with $f(sin^2(\frac{\pi x}{2}))=\frac{c}{\sin(\pi x)}$ - how did you suddenly arrive at the correct $f$? For the density argument: Each number in [0,1] is irrational - why do you need the density argument? $\endgroup$ – Adam Sep 8 '14 at 12:05
  • 1
    $\begingroup$ @Adam For the first one, I assume you refer to the place where I use the condition of $Pf$. I think you forgot to include the 1/2. Anyway,$$(1+ \sqrt{1-\sin^2( \frac{2x\pi}{4})})/2 = (1+\cos(\frac{2x\pi}{4}))/2=\cos^2(\frac{x\pi}{4})=\sin^2(\frac{\pi}{2}-\frac{x \pi}{4})$$. For the second one, $$f(\sin^2(\frac{x\pi}{2}))=\frac{c}{\sin(x\pi)}=\frac{c}{2\sin(\frac{x\pi}{2}) \cos (\frac{x\pi}{2})}=\frac{c}{2\sqrt{\sin^2(\frac{x\pi}{2})(1-\sin^2(\frac{x\pi}{2} ))}}$$ So seems like I forgot to include the 1/2 here. haha $\endgroup$ – karvens Sep 8 '14 at 12:27
  • $\begingroup$ @Adam Sorry for the confusion of the comments. I am really not used to Latex and I needed to see the formated form of the equations so I kept on editing and changed the equations... For the density thing, what I showed wanted to show that $h$ is a constant function. For that, I managed to show that the function $h$ has the same value($c$) for all irrational in $[0,1]$. Then I used the continuity of $h$. $\endgroup$ – karvens Sep 8 '14 at 12:34
  • $\begingroup$ @karvens do you mean x which is irrational but not rational? $\endgroup$ – Adam Sep 8 '14 at 12:35
0
$\begingroup$

I don't have the reference for the derivation, but I can proof, that $f^*$ is indeed an eigenfunction of $P$ with eigenvalue $1$, i.e. $Pf^*=f^*$.

Proof

\begin{align} (Pf^*)(x)&= \frac{1}{4\sqrt{1-x}} \Bigl( f\Bigl(\frac12 (1-\sqrt{1-x})\Bigr)+ f\Bigl(\frac12 (1+\sqrt{1-x})\Bigr)\Bigr) \\ &=\frac{1}{4\sqrt{1-x}} \Bigl( \frac1\pi \frac{1}{\sqrt{\frac12 (1-\sqrt{1-x})(1-\frac12 (1-\sqrt{1-x})}}\Bigr) + \frac1\pi \frac{1}{\sqrt{\frac12 (1+\sqrt{1-x})(1-\frac12 (1+\sqrt{1-x})}}\Bigr) \Bigr) \\ &=\frac{1}{4\pi\sqrt{1-x}}\Bigl( \frac{1}{\sqrt{\frac12 -\frac12 \sqrt{1-x})(\frac12 +\frac12 \sqrt{1-x})}}\Bigr) +\frac{1}{\sqrt{\frac12 +\frac12 \sqrt{1-x})(\frac12 -\frac12 \sqrt{1-x})}}\Bigr) \\ &= \frac{1}{4\pi\sqrt{1-x}}\Bigl( \frac{1}{\sqrt{(\frac12)^2 -(\frac12 \sqrt{1-x})^2}}\Bigr) + \frac{1}{\sqrt{(\frac12)^2 -(\frac12 \sqrt{1-x})^2}}\Bigr) \\ &= \frac{2}{4\pi\sqrt{1-x}}\Bigl( \frac{1}{\sqrt{(\frac12)^2 -(\frac12 \sqrt{1-x})^2}}\Bigr) \\ &= \frac{2}{4\pi\sqrt{1-x}}\Bigl( \frac{1}{\sqrt{\frac14 -\frac14({1-x})}}\Bigr) \\ &= \frac{2}{\pi\sqrt{1-x}}\Bigl( \frac{1}{\sqrt{1-({1-x})}}\Bigr) \\ &= \frac{2}{2\pi\sqrt{1-x}}\Bigl( \frac{1}{\sqrt{x}}\Bigr)=f^* \end{align}

Okay there are a lot of brackets missing/wrong and maybe this is of no use. The message was only, that it is trivial to proof, that $Pf^*=f^*$. Simply plug in the definition of $f^*$ and use binomial expansion for $(a-b)(a+b)$.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks for your note. But I am actually more interested how one derives $f^*$. Are there any tools for it? I guess my question was a bit unclear. Sorry for that. $\endgroup$ – Adam Sep 2 '14 at 20:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.