This is related to discrete dynamical systems, with the initial condition $X_1$ being a random variable with a non singular distribution. The system is driven by the iteration $X_{n+1} = g(X_n)$ for some rather smooth mapping $g$. The purpose here is to find a mapping $g$ so that the invariant distribution (also called attractor or invariant measure or fixed point distribution solution of an integral equation) is pre-specified. This is an inverse problem in the sense that typically $g$ is known and we are looking for the invariant distribution. Here the opposite is true: we assume the invariant distribution is known, and we want to find $g$ that will guarantee that no matter what the non-singular distribution of $X_1$ is, we end up with the the target invariant distribution. There are typically many possible solutions for $g$. The reason I am asking this question is because I believe I got the wrong solution in a simple case, and would like to know how to get it right.

The most basic example is when the invariant distribution is uniform on $[0, 1]$, that is, $F_X(x) = x$. In that case, $g(x) = \{ bx \}$ (the fractional part of $bx$), regardless of $b$ ( an integer $>1$) will work. Here I am trying to solve the case where the invariant distribution (the CDF) is $F_X(x) = x^2$, also with the same support domain being $[0, 1]$. Of course, this is a fist step towards dealing with much more complicated $F_X(x)$.

So, here is my investigation about $F_X(x)=x^2$, and my question is about what is wrong with my analysis. I know something is wrong. First, $g$ must be a many-to-one mapping. The easiest case is when $g$ is a two-to-one mapping, as in the logistic map where $g(x)=4x(1-x)$, with $x\in [0, 1]$. Besides that, if the density $f_X$ is symmetric around $x_0$, then $g(\cdot)$ can not be symmetric around $x_0$. The reason I say this is because I am focusing on mappings $g$ of $[0, 1]$ that are symmetric around some point $x_0$, and in this case, $x_0=\frac{1}{2}$. That is, $g(x)=g(1-x)$.

My wrong solution that needs a fix

Let $g$ be a two-to-one mapping, with $g(x)=g(1-x)$. If the (known) invariant distribution is $F_X(x)$, then $g$ must satisfy the functional equation $f_X(x)= f_X(h_1(x))-f_X(h_2(x))$ with $h_1(x)=g^{-1}(x)$ being one of the two inverses of $g(x)$ when $x<1/2$, and $h_2(x)=1-h_1(x)$ being the other inverse. Based on this, we must have, using the notation $y=g^{-1}(x)$,

$$x^2 = y^2 - (1-y^2)$$

which results in $y = (x^2+1)/2$. Inverting this again, we get $x=\sqrt{2y-1}$, valid if $y\geq \frac{1}{2}$. It follows that $g(x)=\sqrt{2x-1}$ if $\frac{1}{2}\leq x \leq 1$, and $g(x) = \sqrt{1-2x}$ if $0\leq x \leq \frac{1}{2}$. Yet the true mean of the invariant distribution ($F_X(x) = x^2$ on $[0, 1]$) does not seem to match that resulting from $g(\cdot)$ at equilibrium, when $n=\infty$. The system in question is ergodic.

  • $\begingroup$ Shouldn't it be $x^2 = y^2 - (1-y)^2$ in your centered formula ? $\endgroup$
    – EtienneBfx
    Aug 23 at 8:19
  • $\begingroup$ I think the correct formula should be $f_X(x)= f_X(h_1(x))h_1 '(x)-f_X(h_2(x)h_2 '(x))$ $\endgroup$
    – EtienneBfx
    Aug 23 at 10:13
  • $\begingroup$ If you add some noise to $g$ it can be done in some cases. For example, it is known that any Markov chain can be described as iterations of a noisy map. You can design Markov chains with prescribed stationary measure using the Metropolis method $\endgroup$ Aug 23 at 10:21

To have here the invariant distribution with cdf $F$ given by $F(x)=x^2$ for $x\in[0,1]$, all that is needed is a change of variables.

More generally, let $F$ be the cdf of any non-atomic distribution supported on an interval $I$ in $\mathbb R$. Let $F^{-1}$ denote the inverse of the restriction of $F$ to $I$. Let $U$ be a random variable uniformly distributed on $[0,1]$. Then the cdf of the random variable $$X:=F^{-1}(U)$$ is $F$, and $$U=F(X).$$

Suppose that a function $g_*$ is such that the uniform distribution on $[0,1]$ is invariant for the dynamical system $U_{n+1}=g_*(U_n)$ -- that is, $U\overset D=g_*(U)$, where $\overset D=$ means the equality in distribution. Then $F(X)\overset D=g_*(F(X))$, that is, $$X\overset D=g(X),$$ where $$g:=F^{-1}\circ g_*\circ F.$$ So, the distribution with the prescribed cdf $F$ is invariant for the dynamical system $X_{n+1}=g(X_n)$, as desired.

In particular, if $F(x)=x^2$ for $x\in[0,1]$ and $g_*(u)=\{bu\}$ for $u\in[0,1]$, then $$g(x)=\sqrt{\{bx^2\}}$$ for $x\in[0,1]$.

  • $\begingroup$ Thank you! How did you pick up $g_*(u)=\lfloor bu\rfloor$? I assume $b$ is an integer $> 1$. $\endgroup$ Aug 23 at 17:02
  • $\begingroup$ I think $g_*(u) = \{bu\}$. I got it wrong (and you probably used my wrong solution) but I fixed it. So $g(x)=\sqrt{\{bx^2\}}$. $\endgroup$ Aug 23 at 17:28
  • 2
    $\begingroup$ @VincentGranville : Oops! Indeed, I had just thoughtlessly copied the incorrect expression for $g_*$. This is now fixed. $\endgroup$ Aug 23 at 17:34

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.