43 Multivariate Transformations
$$
$$
Given i.i.d. data \(X_1, \dots, X_n\), we know how to determine the distributions of statistics such as
- the sample mean \(\bar X\), and more generally, functions of means \(g(\bar X)\), and
- order statistics \(X_{(k)}\), such as the sample median.
A general statistic is of the form \(g(X_1, \dots, X_n)\). In this chapter, we will learn a general strategy for deriving the distribution of a general statistic when the random variables are continuous and \(g\) is differentiable.
43.1 Univariate Transformations
As a warm-up, consider the case of a single continuous random variable \(X\) with PDF \(f_X(x)\). What is the PDF of \(Y = g(X)\)?
In Chapter 20, we presented a general strategy for deriving the PDF of \(Y\). First, find the CDF of \(Y\), then differentiate to obtain the PDF. However, when \(g(x)\) is a differentiable function that is strictly increasing (or decreasing), there is a simple recipe.
One natural application of Theorem 43.1 is to location-scale transformations.
43.2 Multivariate Transformations
Now, we consider the case where we have \(n\) random variables \(X_1, \dots, X_n\), and \(g\) is a differentiable one-to-one function from \(A \subset \mathbb{R}^n\) to \(B \subset \mathbb{R}^n\). That is, \[ (Y_1, \dots, Y_n) = g(X_1, \dots, X_n). \] Note that the “derivative” of a function from \(\mathbb{R}^n\) to \(\mathbb{R}^n\) is an \(n\times n\) matrix of partial derivatives, \[ \left[\frac{\partial (y_1, \dots, y_n)}{\partial (x_1, \dots, x_n)}\right] \overset{\text{def}}{=}\begin{bmatrix} \frac{\partial y_1}{\partial x_1} & \frac{\partial y_1}{\partial x_2} & \dots & \frac{\partial y_1}{\partial x_n} \\ \frac{\partial y_2}{\partial x_1} & \frac{\partial y_2}{\partial x_2} & \dots & \frac{\partial y_2}{\partial x_n} \\ \vdots & \vdots & \ddots & \vdots \\ \frac{\partial y_n}{\partial x_1} & \frac{\partial y_n}{\partial x_2} & \dots & \frac{\partial y_n}{\partial x_n} \end{bmatrix}, \] called the Jacobian matrix.
Then, the natural generalization of Theorem 43.1 is the following.
Let’s apply Theorem 43.2 to some examples.
43.2.1 Simulating normal random variables
Theorem 43.2 provides a basis for simulating normal random variables. We begin by discussing how it can be used to simulate a single normal random variable, which is harder than it sounds.
With the Box-Muller transform, we get two independent normal random variables for the price of one. We now discuss how to transform these into correlated normal random variables.
Notice that we can further apply a location-scale transformation (Section 20.2) to \(X\) and \(Y\) to obtain a bivariate normal distribution with any mean and variance.
43.3 Exercises
Exercise 43.1 (Polar coordinates of a random point in the unit circle) Let \(X\) and \(Y\) denote the coordinates of a point chosen uniformly in the unit circle. That is, the joint density is \[ f_{X,Y}(x,y) = \frac{1}{\pi}, \qquad x^2 + y^2 \leq 1. \] Find the joint density of the polar coordinates \(R = \sqrt{X^2+Y^2}\) and \(\Theta = \tan^{-1}(Y/X)\).
Exercise 43.2 (Polar coordinates of independent standard normals) Let \((X,Y)\) denote a random point in the plane, and assume the rectangular coordinates \(X\) and \(Y\) are independent standard normal random variables. Find the joint distribution of \(R\) and \(\Theta\), the polar coordinates.
Exercise 43.3 (Rederiving the distribution of a sum of i.i.d. exponentials) Let \(X_1, \dots, X_n\) be i.i.d. \(\text{Exponential}(\lambda)\). Define \[ Y_k = X_1 + \cdots + X_k \] for \(1 \leq k \leq n\).
- Find the joint density of \(Y_1, \dots, Y_n\).
- Use part (a) to find the density of \(Y_n\). Does your result make sense?
- Find the conditional density of \(Y_1, \dots, Y_{n-1}\) given \(Y_n = t\).
Exercise 43.4 (Transforming standard normals) Let \(Z_1\) and \(Z_2\) be i.i.d. \(\textrm{Normal}(\mu= 0, \sigma^2= 1)\). Let \(X = Z_1 + Z_2\) and \(Y = Z_1 - Z_2\). We know that both are \(\textrm{Normal}(\mu= 0, \sigma^2= 2)\). But are they independent?
Exercise 43.5 (Transforming exponentials) Let \(X\) and \(Y\) be i.i.d. \(\text{Exponential}(\lambda)\). Let \(T = X + Y\) and \(\displaystyle W = \frac{X}{Y}\). Are \(T\) and \(W\) independent? Find the marginal densities of \(T\) and \(W\).