Transformation of a random variable vs. joint transformation of several random variables

89 Views Asked by At

Let $X \sim f_X(x)$ and $Y = g(X)$. If $g(X)$ is a differentiable, monotonic function with inverse such that $X = g^{-1}(Y)$ then the PDF of $Y$ can be described:

$$ f_Y(y) = f_X(g^{-1}(y)) \bigg| \dfrac{d}{dy}g^{-1}(y) \bigg| $$

In the multidimensional case, I have seen a very similar expression where the absolute value of the derivative is replaced by the absolute value of the determinant of the Jacobian:

$f_Y(\pmb{y}) = f_X(g^{-1}(\pmb{y})) |\bigg| \dfrac{\partial \pmb{x}}{\partial \pmb{y}} \bigg| |$

However, I have not seen a qualifier that in the multivariate case the function $\pmb{y} =g(\pmb{x})$ must be monotonic in all variables. Is this assumed or does this requirement not apply in this case? If not why?

1

There are 1 best solutions below

2
On BEST ANSWER

The reason is that it is assumed that the system $$\pmb{y} =g(\pmb{x})$$ has a unique solution $x=g^{-1}(\pmb{y})$, i.e., $g$ is invertible when using the formula. In the univariate case, the function $g$ is invertible if and only if $g$ is strictly monotone, which is not the case in the multivariate case ($y_1=x_1, y_2=-x_2$ is a simple example where $g$ is not monotone). Therefore, the condition of strict monotonicity of $g$ used in univariate case is replaced with other equivalent conditions holding in the multivariate case (i.e., the system has a unique solution or $g$ is invertible).

Suppose that the system has multiple solutions:

$$\pmb{x}=h^j(\pmb{y}), j=1,\dots, k$$

such that the union of the sets $h^j(S_Y), j=1,\dots,k$ is $S_X$ with null intersections ($S_X$ and $S_Y$ denote the support of $X$ and $Y$), that is, $h^j(S_Y)\cap h^k(S_Y)$ has Lebesgue measure zero for each $j \neq k$, and $\cup_{j=1}^k h^j(S_Y)=S_X$. Then, the following

$$\color{blue}{f_Y(\pmb{y}) = \sum_{j=1}^kf_X(h^j(\pmb{y})) \left | \det \left (\dfrac{\partial h^j(\pmb{y})}{\partial \pmb{y}} \right ) \right |}$$

is a useful formula. For example, for $y=x^2$, $h^1(y)=\sqrt{x}$ and $h^2(y)=-\sqrt{x}$ with $S_X=\mathbb R$ and $S_Y=\mathbb R_{\, \ge0}$. You can see $$h^1(\mathbb R_{\, \ge0})=\mathbb R_{\, \ge0}, h^2(\mathbb R_{\, \ge0})=\mathbb R_{\, \le0},$$

with null intersection $\{ 0\}$ and union $\mathbb R$. Thus, we have for $y\ge0$:

$$f_Y(y)=\frac{1}{\sqrt{y}}f_X(\sqrt{y}) +\frac{1}{\sqrt{y}}f_X(-\sqrt{y}).$$

For more details and examples, you can see pages 6-10 of this note.