3
$\begingroup$

Let $X \in \mathbb{R}^d$ be a random variable with probability distribution $P$. Let $f:\mathbb{R}^d \to \mathbb{R}^d$ be an invertible function and let $P_{f}$ be the distribution of random variable $f(X)$. Suppose $P$ and $P_f$ are such that the mixture distribution $(1-\alpha)P + \alpha P_f$, for some $\alpha \in [0,1]$, is equal to the standard normal distribution $\mathcal{N}(0, I_{d\times d})$, where $I_{d\times d}$ is the $d\times d$ identity matrix.

Given $\alpha, f$, I'm interested in finding the distribution $P$ for which the above condition holds, i.e., $(1-\alpha)P + \alpha P_f$ is equal to $\mathcal{N}(0, I_{d\times d})$. (By finding a distribution what I mean is that, at any given $x \in \mathbb{R}^d$, I would like to compute the probability density of $P$ at $x$.)

Here are my questions:

1) For any given $\alpha, f$, is there a unique $P$ which satisfies this i.e., is the problem identifiable? Or can there be multiple distributions? Of course, when $\alpha = 1/2$ the problem is clearly not identifiable. Assuming $\alpha \neq 1/2$, is the problem identifiable?

2) Assuming the problem is identifiable, how can I compute the density of $P$ at any given $x \in \mathbb{R}^d$? We have the following relation between the densities of $P, P_f$ and $\mathcal{N}(0, I_{d\times d})$

$$\frac{1}{\sqrt{2\pi}} e^{-x^2/2} = (1-\alpha)p(x) + \alpha |\det{J_{f^{-1}}(x)}| p(f^{-1}(x)),$$

where $p(x)$ is the probability density of $P$ at $x$ and $\det{J_{f^{-1}}(x)}$ is the determinant of the Jacobian of $f^{-1}$ evaluated at $x$. Is there a way to compute $p(x)$ from this equation?

$\endgroup$

1 Answer 1

1
$\begingroup$

$\newcommand{\al}{\alpha} \newcommand{\R}{\mathbb{R}} \renewcommand{\x}{\bar x} \newcommand{\ga}{\gamma}$ Welcome to MathOverflow! If $\al\ne1/2$ and $f$ is an involution of $\R^d$ (so that $f=f^{-1}$), then $p$ is identifiable and can be found as follows. We have \begin{equation*} \phi(x):=\frac{1}{\sqrt{2\pi}} e^{-x^2/2} = (1-\al)p(x) + \al |\det{J_{f^{-1}}(x)}| p(f^{-1}(x)) \end{equation*} for all $x\in\R^d$, which we can rewrite as \begin{equation*} (1-\al)p(x) + \al p(\x)j(x)=\phi(x), \tag{1} \end{equation*} where $j(x):=|\det{J_f(x)}|$ and $\x:=f(x)$. Replacing here $x$ by $\x$, we have \begin{equation*} (1-\al)p(\x) + \al p(x)/j(x) =\phi(\x), \tag{2} \end{equation*} since $j(\x)=1/j(x)$. Solving the system of linear equations (1)--(2) for $p(x),p(\x)$, we find \begin{equation*} p(x)=\frac{(1-\al) \phi (x)-\al j(x) \phi (\x)}{2 \al-1}. \end{equation*}

$\endgroup$
3
  • $\begingroup$ Thanks. This is interesting. Can something be said when f is not an involution? $\endgroup$
    – v0511
    Jul 17, 2019 at 13:52
  • $\begingroup$ One way I can think of extending your approach is by assuming that $f^{-1}$ is contractive (which also means it has a fixed point). Then one can make use of the expressions for $\phi(x), \phi(f^{-1}(x)), \phi(f^{-1}\circ f^{-1}(x)) \dots$ to at least approximately evaluate $p(x)$ $\endgroup$
    – v0511
    Jul 17, 2019 at 14:13
  • $\begingroup$ More generally, if $f$ is not an involution, this suggests that the answer will depend on the structure of the orbits $\{x,f(x),f(f(x)),\dots\}$. If the orbits are finite, then I think we can reason similarly. $\endgroup$ Jul 17, 2019 at 14:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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