Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

If $X_1, ..., X_n$ are normal random variables, will $[X_1, ..., X_n]$ be a normal random vector?

I know if $X_1, ..., X_n$ are further independent, then $[X_1, ..., X_n]$ is a normal random vector. Can there be more relaxed condition on $X_1, ..., X_n$ than being independent?

Thanks in advance!

share|cite|improve this question
up vote 3 down vote accepted

$[X_1,...,X_n]$ is a normal (or Gaussian) random vector is the same as saying that those variables are jointly normal -or that it is a multivariate Gaussian. That each $X_i$ is normal (i.e., the marginal distributions are normal) is necessary but not sufficient (easy to give counterexamples). That each $X_i$ is normal and they are independent is -as you say- sufficient, but not necessary. A general (necessary and sufficient) condition can be expressed in terms of a linear combination of normal iid scalar variables.

Specifically: given $Z_1 ... Z_n$ iid standard normals (mean 0 and variance 1), then $X = A Z + b$ (with A any square nonsingular matrix and b any fixed vector) is jointly normal - and this is fully general. From this comes the formula of the (general) multivariate Gaussian variable. Conversely, if $X$ is a normal random vector then one can find a matrix C and a vector d such that $Z = C X + d$ is a vector of iid standard normal variables. This comes to be a multivariate generalization of the well known formula to standardize a Gaussian: $z = \frac{x-\mu}{\sigma}$

share|cite|improve this answer
leonbloy: Thanks! I wonder why if each marginal distribution is normal, the joint distribution may not be normal? Could you give a counterexample? – Ethan Nov 9 '11 at 20:29
@Ethan: Let $X_1 $ $X_2$ iid standard normal (so that they are jointly normal). Define $Y_1 Y_2$ with $Y_1 = X_1$ $Y_2 = |X_2| sign(X_1)$, then $Y_1 Y_2$ will have the same sign, each will marginally be a gaussian, but jointly not (graphically: take the 3D bell-shaped joint gaussian and remove the 2nd and 4th cuadrant). – leonbloy Nov 10 '11 at 1:35

$(X_1,\ldots,X_n)$ is a normal vector if and only if any linear combination $\sum\nolimits_{i = 1}^n {a_i X_i }$, $a_i \in \mathbb{R}$, is a univariate normal variable (here, constants are also regarded as normal variables).

Concerning the first question, let $X$ be an arbitrary Gaussian process indexed by $T$ (for example a Brownian motion or bridge, where $T=[0,\infty)$ or $T=[0,1]$, respectively). Then, by definition, $(X_{t_1},\ldots,X_{t_n})$ is a normal vector, for any $n \geq 1$ and $t_1,\ldots,t_n \in T$, but the components are (in general) dependent random variables.

share|cite|improve this answer
Thanks! How can it be proved? – Ethan May 6 '11 at 2:47
It is sometimes the definition, but I'll look for a proof. – Shai Covo May 6 '11 at 2:59
Thanks! My definition for normal vector is by its density function. So can we prove equivalence between its density function and any linear combination being normal? – Ethan May 6 '11 at 7:35
@Ethan, for the proof, just use the moment-generating function. One direction is trivial and the other one is not much harder. – cardinal May 6 '11 at 12:03
@Ethan: Fourier transform. – Did Jun 29 '11 at 18:39

To add a little to leonbloy's excellent answer, I would say that $\vec{X} = (X_1, \ldots, X_n)$ is called a Gaussian vector, equivalently that $X_1, \ldots, X_n$ are called jointly Gaussian random variables if $\vec{X} = \vec{Z}A + \vec{\mu}_X$ where $\vec{Z} = (Z_1, \ldots, Z_k)$ is a vector of standard (zero-mean, unit variance) independent Gaussian random variables, $1 \leq k \leq n$, $A$ is a $k \times n$ matrix, and $\vec{\mu} = E[\vec{X}]$ is the mean vector of $\vec{X}$. This allows us to call $(X, Y) = (aZ+b, cZ+d)$ a Gaussian vector even though the joint density function $f_{X,Y}(x,y)$ cannot be expressed as the usual formula for the bivariate Gaussian density function, because this formula involves quantities of the form $(1 - \rho^2)^{-1}$ while here $\vert \rho\vert = 1$ since $X$ and $Y$ are perfectly correlated. As leonbloy points out, when $k = n$, the $n$-variate density formula for jointly Gaussian random variables can be obtained straightforwardly from this description.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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