Take the 2-minute tour ×
Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It's 100% free, no registration required.

I read that if you have a quadratic $f(x) = ax^2+bx+c$ and providing the leading coefficient $a < 0$, then $e^{f(x)}$ is the pdf of a normal distribution with mean $\mu = -\frac{b}{2a}$ and $\sigma^2 = -\frac{1}{2a}$.

Does this generalise to $e^{f(\mathbf{x})}$ defining a multivariate normal distribution where $f(\mathbf{x}) = \mathbf{x}'\mathbf{A}\mathbf{x} + \mathbf{x}'\mathbf{b} + \mathbf{c}$. If so what are the conditions for this to be true and what is the mean and covariance?

Thanks!

share|improve this question
1  
Just a little clarification, $e^{f(x)}$ is proportional to the normal pdf. –  user10525 Jul 5 '12 at 14:37
2  
All these questions are answered near the top of the Wikipedia article on the multivariate normal distribution. –  whuber Jul 5 '12 at 15:39
    
Thanks for the link –  NapierL Jul 5 '12 at 16:02

1 Answer 1

up vote -1 down vote accepted

The result in 1 dimension that you gave above comes about by equating exponents namely ax$^2$ +bx +c =-(x-μ)$^2$/(2σ$^2$)= -1/(2σ$^2$) x$^2$ +μx/(σ$^2$)-μ$^2$/(2σ$^2$) which gives

a=-1/(2σ$^2$)

b=μ/(σ$^2$) and

c=-μ$^2$/(2σ$^2$) Now solving for μ and σ$^2$ we get

μ =-b/(2a) and

σ$^2$= -1/(2a) The equations involving a and b were all that was needed.

Now similarly for the multivariate normal the quadratic form in the exponent is

-(x-μ)' ∑$^-$ $^1$ (x-μ)/2 and you equate it to

x′Ax+x′b+c and solve for ∑ and μ in terms of the given A, b and c.

share|improve this answer
    
Thanks, seems obvious once it's been explained! –  NapierL Jul 5 '12 at 15:40
2  
Perhaps obvious but still incomplete, because (1) you cannot always find solutions $\Sigma$ and $\mu$ and (2) even when you do they do not necessarily determine a valid distribution. Isn't that what the question asks, though: what are the conditions for $f(\mathbf{x})$ to determine a multivariate normal? –  whuber Jul 5 '12 at 17:24
    
@whuber No based on the OPs satisfaction with my answer. I think he was only asking how to determine the mean and covariance of an MVN if it is expressed in the form exp(f(x)) where f is given in the form x′Ax+x′b+c. In some respects the question seems to be consistent with your interpretation excpet that in the 1 D case he mentioned the restriction that a must be less than 0. So he probably recognizes that there may be restrictions requires on A b and c for Σ and μ to exist. –  Michael Chernick Jul 5 '12 at 17:32

Your Answer

 
discard

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.