Appendix B: Multivariate Normal Distributions
A k-dimensional random vector follows a multivariate normal distribution with mean and positive-definite covariance matrix if its probability density function (pdf) is
8.47
We use the notation to denote that follows such a distribution. This normal distribution plays an important role in multivariate statistical analysis and it has several nice properties. Here we consider only those properties that are relevant to our study. Interested readers are referred to Johnson and Wichern (1998) for details.
To gain insight into multivariate normal distributions, consider the bivariate case (i.e., k = 2). In this case, we have
Using the correlation coefficient ρ = σ12/(σ1σ2), where is the standard deviation of xi, we have and . The pdf of then becomes
where
Chapter 4 of Johnson and Wichern (1998) contains some plots of this pdf function.
Let be a nonzero k-dimensional vector. Partition the random vector as , where and with 1 ≤ p < k. Also partition and accordingly as
Some properties of are as follows:
1. . That is, any nonzero linear combination of is univariate normal. The inverse of this property also holds. Specifically, if is univariate normal for any nonzero vector , then is multivariate normal.
2. The marginal distribution of is normal. In fact, for i = 1 and 2, where k1 = p and k2 = k − p.
3. if and only if and are independent.
4. The random variable follows a chi-squared distribution with m degrees of freedom.
5. The conditional distribution of given = is also normally distributed as
The last property is useful in many scientific areas. For instance, it forms the basis for time series forecasting under the normality assumption and for recursive least-squares estimation.