MathOverflow is a question and answer site for professional mathematicians. 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


A part of this answer says that the inner product $\left( \frac{X}{|X|} \right)^T Y$, where X and Y are vectors with i.i.d. zero-mean Gaussian elements, is independent on X and is again Gaussian distributed. I have found the same claim here, but no reasoning, and my skills are apparently not sufficient to see it. Where should I look? Is $\frac{X}{|X|}$, a vector with N complex elements, somehow just a rotation?

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

$X/|X|$ is almost surely a uniformly random element on the unit sphere of dimension $n-1$; this is not the same thing as a rotation, which is a matrix. Since a multivariate standard Gaussian vector is invariant in law under a fixed rotation, its law is certainly invariant under a random rotation as well: thus if $A$ is a uniformly random orthogonal matrix, then $A Y$ is again standard Gaussian. Now let $Z = (X/|X|)^T Y$. Then $Z$ has the same law as the first component of $AY$, which is clearly univariate standard Guassian. To verify its independence with $X$: if you rotate $X$ by an orthogonal matrix $A$, you can absorb $A$ into $Y$, whose law is invariant under rotation. So conditional law of $Z$ under two different $X$ values differing by a rotation stays the same. More obviously, if you scale $X$, the conditional law of $Z$ remains the same.

Another thing I noticed is that $Z$ and $Y$ are not jointly normal! Notice that conditioning on $Y$ clearly has an effect on $Z$. Now assuming they are jointly normal, we can show $E Z v^T Y = 0$ for all vector $v$. In fact we can show $E (X/|X|)^T u v^T u = 0$ for fixed $u,v$. In fact, this follows simply from $E (X/|X|)^T u = E (-X/|-X|)^T u = 0$. This is a contradiction. This is similar to the situation $B X$ and $X$, where $X$ is standard 1d gaussian and $B$ is an independent Bernoulli $\pm 1$ variable. They are not jointly normal either.

share|cite|improve this answer
About your answer, I still have several questions: 1.Is $\mathbf{A}$ must be the uniformly distributed matrix? What about the unitary matrix with some other distribution? 2.If $\mathbf{Y}$ is not the standard Gaussian vector, i.e. its mean is not zero or its variance is not $1$, is $\mathbf{Y}$ still invariant under a rotation? Looking forward to your reply – user47169 Feb 19 '14 at 12:52

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.