Marginal distributions of the multinomial normal distribution

-

 Marginal distributions of a multivariate normal distribution are also normal distributions. Let's prove this.



See also: Multivariate normal distribution [Wikipedia]

The density function of a multivariate normal distribution is given as

\[ f(\mathbf{x}) = \frac{1}{\sqrt{(2\pi)^n|\Sigma|}}\exp\left[-\frac{1}{2}(\mathbf{x}-\boldsymbol{\mu})^{\top}\Sigma^{-1}(\mathbf{x}-\boldsymbol{\mu})\right] \]

where \(\mathbf{x}\in\mathbb{R}^n\) is the random vector, \(\boldsymbol{\mu}\) is the mean vector and \(\Sigma\) is the covariance matrix. By changing the variables \(\mathbf{x} - \boldsymbol{\mu} \mapsto \mathbf{x}\), we can assume the mean is zero without losing generality. So, in the following, we only consider

\[ f(\mathbf{x}) = \frac{1}{\sqrt{(2\pi)^n|\Sigma|}}\exp\left[-\frac{1}{2}\mathbf{x}^{\top}\Sigma^{-1}\mathbf{x}\right]. \]

We need the following theorems from linear algebra.

Theorem 1

Let \(A\) be an \(n\times n\) regular, \(D\) be \(m\times m\) regular, \(B\) be \(n\times m\), and \(D\) be \(m\times n\) matrices, and \(S = D - CA^{-1}B\) be regular. Then,

\[ \begin{bmatrix} A & B \\ C & D \end{bmatrix}^{-1} = \begin{bmatrix} A^{-1} + A^{-1}BS^{-1}CA^{-1} & -A^{-1}BS^{-1}\\ -S^{-1}CA^{-1} & S^{-1} \end{bmatrix} \tag{eq:1} \]

Remark. Perhaps it is easier to memorize if we write in the following way:

\[ \begin{bmatrix} A_{11} & A_{12} \\ A_{21} & A_{22} \end{bmatrix}^{-1} = \begin{bmatrix} A_{11}^{-1} + A_{11}^{-1}A_{12}S_{22}^{-1}A_{21}A_{11}^{-1} & -A_{11}^{-1}A_{12}S_{22}^{-1}\\ -S_{22}^{-1}A_{21}A_{11}^{-1} & S_{22}^{-1} \end{bmatrix} \]

with \(S_{22} = A_{22} - A_{21}A_{11}^{-1}A_{12}\). □

Proof. Let's define the following matrices:

\[ \begin{align*} F &= \begin{bmatrix} A & B\\ C & D \end{bmatrix},\\ W &= \begin{bmatrix} I_n & O_{n,m}\\ -CA^{-1} & I_m \end{bmatrix},\\ U &= \begin{bmatrix} I_n & -BS^{-1}\\ O_{m,n} & I_m \end{bmatrix}. \end{align*} \]

Here, \(I_n\) is the \(n\times n\) identity matrix and \(O_{n,m}\) is the \(n\times m\) zero matrix. Then,

\[ \begin{align*} G & = WF = \begin{bmatrix} A & B\\ O_{m,n} & S \end{bmatrix},\\ H &= UG = UWF = \begin{bmatrix} A & O_{n,m}\\ O_{m,n} & S \end{bmatrix}. \end{align*} \]

Therefore, 

\[ H^{-1} = F^{-1}W^{-1}U^{-1}. \]

Hence,

\[F^{-1} = H^{-1}UW.\]

But

\[ H^{-1}= \begin{bmatrix} A^{-1} & O_{n,m}\\ O_{m,n} & S^{-1} \end{bmatrix}. \]

Thus, we obtain (eq:1). ■

Theorem 2

With the same setting as Theorem 1, we have the matrix determinant

\[ \begin{align} \begin{vmatrix} A & B\\ C & D \end{vmatrix} &= |A|\cdot |D - CA^{-1}B|\end{align} \]

Proof.   We use the same notation as in the proof of Theorem 1.

  Noting \(G = WF\) and \(|W| = 1\), we have \(|F| = |G|\). But \(G\) is block triangular, so

\[ |G| = |A|\cdot|S| = |A|\cdot|D - CA^{-1}B|. \]

■

Now, let's go back to our problem. First, split the variable \(\mathbf{x}\) into two blocks:

\[ \mathbf{x} = \begin{bmatrix} \mathbf{x}_1\\ \mathbf{x}_2 \end{bmatrix} \]

where \(\mathbf{x}_1\in \mathbb{R}^p\) and \(\mathbf{x}_2 \in \mathbb{R}^q\) with \(p + q = n\). We consider integrating \(\mathbf{x}_2\) out to get the marginal distribution

\[ f(\mathbf{x}_1) = \int_{\mathbb{R}^q} f(\mathbf{x}_1, \mathbf{x}_2)\,d\mathbf{x}_2. \]

We also split the covariance matrix into blocks as

\[ \Sigma = \begin{bmatrix} \Sigma_{11} & \Sigma_{12}\\ \Sigma_{21} & \Sigma_{22} \end{bmatrix} \]

where \(\Sigma_{11}\) is \(p\times p\),  \(\Sigma_{22}\) is \(q\times q\), \(\Sigma_{12} = \Sigma_{21}^\top\) are \(p\times q\) matrices.

By Theorem 1, we have

\[\Sigma^{-1} = \begin{bmatrix} \Sigma_{11}^{-1} + \Sigma_{11}^{-1}\Sigma_{12}S^{-1}\Sigma_{12}^{\top}\Sigma_{11}^{-1} & - \Sigma_{11}^{-1}\Sigma_{12}S^{-1}\\ -S^{-1}\Sigma_{12}^{\top}\Sigma_{11}^{-1} & S^{-1} \end{bmatrix}  \]

where \(S = \Sigma_{22} - \Sigma_{12}^{\top}\Sigma_{11}^{-1}\Sigma_{12}\). Thus,

\[ \begin{align*} \mathbf{x}^{\top}\Sigma^{-1}\mathbf{x} &= \mathbf{x}_1^\top(\Sigma_{11}^{-1} + \Sigma_{11}^{-1}\Sigma_{12}S^{-1}\Sigma_{12}^{\top}\Sigma_{11}^{-1})\mathbf{x}_{1} -2\mathbf{x}_1^{\top}\Sigma_{11}^{-1}\Sigma_{12}S^{-1}\mathbf{x}_2 + \mathbf{x}_2^{\top}S^{-1}\mathbf{x}_2\\ &= \mathbf{x}_1^\top\Sigma_{11}^{-1}\mathbf{x}_1\\ & ~~ + \mathbf{x}_1^\top\Sigma_{11}^{-1}\Sigma_{12}S^{-1}\Sigma_{12}^{\top}\Sigma_{11}^{-1}\mathbf{x}_1 - \mathbf{x}_1^{\top}\Sigma_{11}^{-1}\Sigma_{12}S^{-1}\mathbf{x}_2\\ & ~~ -\mathbf{x}_2^{\top}S^{-1}\Sigma_{12}^{\top}\Sigma_{11}^{-1}\mathbf{x}_1 + \mathbf{x}_2^{\top}S^{-1}\mathbf{x}_2\\ &= \mathbf{x}_1^\top\Sigma_{11}^{-1}\mathbf{x}_1\\ & ~~ -\mathbf{x}_1^\top\Sigma_{11}^{-1}\Sigma_{12}S^{-1}(\mathbf{x}_2 - \Sigma_{12}^{\top}\Sigma_{11}^{-1}\mathbf{x}_1)\\ & ~~ + \mathbf{x}_2^{\top}S^{-1}(\mathbf{x}_2 - \Sigma_{12}^{\top}\Sigma_{11}^{-1}\mathbf{x}_1)\\ &=\mathbf{x}_1^\top\Sigma_{11}^{-1}\mathbf{x}_1 +(\mathbf{x}_2 - \Sigma_{12}^{\top}\Sigma_{11}^{-1}\mathbf{x}_1)^{\top}S^{-1}(\mathbf{x}_2 - \Sigma_{12}^{\top}\Sigma_{11}^{-1}\mathbf{x}_1) \end{align*} \]

By Theorem 2, 

\[|\Sigma|=|\Sigma_{11}|\cdot|S|\]

Thus,

\[ \begin{align} \frac{1}{\sqrt{(2\pi)^n|\Sigma|}}\int_{\mathbb{R}^q}\exp\left(-\frac{1}{2}\mathbf{x}^{\top}\Sigma^{-1}\mathbf{x}\right)\,d\mathbf{x}_2 &= \frac{\sqrt{(2\pi)^q|S|}}{\sqrt{(2\pi)^n|\Sigma|}}\exp\left(-\frac{1}{2}\mathbf{x}_1^\top\Sigma_{11}^{-1}\mathbf{x}_1\right)\\ &= \frac{1}{\sqrt{(2\pi)^p|\Sigma_{11}|}} \exp\left(-\frac{1}{2}\mathbf{x}_1^\top\Sigma_{11}^{-1}\mathbf{x}_1\right). \end{align} \]

This is the multivariate normal distribution of \(\mathbf{x}_1\) with the covariance matrix \(\Sigma_{11}\) that is obtained from \(\Sigma\) by discarding the terms involving elements related to \(\mathbf{x}_2\). Therefore, marginal distributions of a multivariate normal distribution are normal distributions.

Comments

Post a Comment

Popular posts from this blog

Open sets and closed sets in \(\mathbb{R}^n\)

-
Image
Open sets In \(\mathbb{R}\), we have the notion of an open interval such as \((a, b) = \{x \in \mathbb{R} | a < x < b\}\). We want to extend this idea to apply to \(\mathbb{R}^n\). We also introduce the notions of bounded sets and closed sets in \(\mathbb{R}^n\). Recall that the \(\varepsilon\)-neighbor of a point \(x\in\mathbb{R}^n\) is defined as \(N_{\varepsilon}(x) = \{y \in \mathbb{R}^n | d(x, y) < \varepsilon \}\) where \(d(x,y)\) is the distance between \(x\) and \(y\). Definition (Open set) A subset \(U\) of \(\mathbb{R}^n\) is said to be an  open set  if the following holds: \[\forall x \in U ~ \exists \delta > 0 ~ (N_{\delta}(x) \subset U).\tag{Eq:OpenSet}\] That is, for every point in an open set \(U\), we can always find an open ball centered at that point, that is included in \(U\). See the following figure. Perhaps, it is instructive to see what is  not  an open set. Negating (Eq:OpenSet), we have \[\exists x \in U ~ \forall \delta > 0 ~ (N_{\delta}(x) \not
Read more

Euclidean spaces

-
Image
We would like to study multivariate functions (i.e., functions of many variables), continuous multivariate functions in particular. To define continuity, we need a measure of "closeness" between points. One measure of closeness is the Euclidean distance. The set \(\mathbb{R}^n\) (with \(n \in \mathbb{N}\)) with the Euclidean distance function is called a Euclidean space. This is the space where our functions of interest live. The real line is a geometric representation of \(\mathbb{R}\), the set of all real numbers. That is, each \(a \in \mathbb{R}\) is represented as the point \(a\) on the real line. The coordinate plane , or the \(x\)-\(y\) plane , is a geometric representation of \(\mathbb{R}^2\), the set of all pairs of real numbers. Each pair of real numbers \((a, b)\) is visualized as the point \((a, b)\) in the plane. Remark . Recall that \(\mathbb{R}^2 = \mathbb{R}\times\mathbb{R} = \{(x, y) | x, y \in \mathbb{R}\}\) is the Cartesian product of \(\mathbb{R}\) with i
Read more

Newton's method

-
Image
Newton's method (the Newton-Raphson method) is a very powerful numerical method for solving nonlinear equations.  Suppose we'd like to solve a non-linear equation \(f(x) = 0\), where \(f(x)\) is a (twice) differentiable non-linear function. Newton's method generates a sequence of numbers \(c_1, c_2, c_3, \cdots\) that converges to a solution of the equation. That is, if \(\alpha\) is a solution (i.e., \(f(\alpha) = 0\)) then,  \[\lim_{n\to\infty}c_n = \alpha,\] and this sequence \(\{c_n\}\) is generated by a series of linear approximations of the function \(f(x)\). Theorem (Newton's method) Let \(f(x)\) be a function that is twice differentiable on an open interval \(I\) that contains the closed interval \([a, b]\) (i.e., \([a,b]\subset I\)) and satisfy the following conditions: \(f(a) < 0\) and \(f(b) > 0\); For all \(x\in [a, b]\), \(f'(x) > 0\) and \(f''(x) > 0\). Let us define the sequence \(\{c_n\}\) by \[ \begin{eqnarray} c_1 &
Read more