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

Given the following joint normal PDF of $X \in \mathcal{R}^K$ and $Y,Z \in \mathcal{R}$

$p(\begin{bmatrix} X \\\ Y \\\ Z \end{bmatrix}) = \mathcal{N}(\begin{bmatrix} \mu_X \\\ \mu_Y \\\ \mu_Z \end{bmatrix},\begin{bmatrix} \Sigma_{XX} \ \Sigma_{XY} \ \Sigma_{XZ} \\\ \Sigma_{YX} \ \Sigma_{YY} \ \Sigma_{YZ} \\\ \Sigma_{ZX} \ \Sigma_{ZY} \ \Sigma_{ZZ} \end{bmatrix})$

How can we derive the closed form expression for the following PDF?

$P(X|A)$ (or equivalently $P(X|A^2)$)

where, $A = \sqrt{Y^2+Z^2}$

share|cite|improve this question
Duplicate of – Dilip Sarwate Apr 30 '12 at 21:22
The closed form is quite unlikely. One is first to consider $X|(Y,Z)$, then represent $(Y,Z) = A (\cos(\Phi), \sin(\Phi))$ and then average out $\Phi$. I would expect appearance of some matrix Bessel functions. Has this example its origin in wireless communication theory, by chance? – Sasha Apr 30 '12 at 21:37
Thanks for the hints. Actually, the question is originated from speech enhancement domain. – smo Apr 30 '12 at 21:58
smo: Any luck with the answer below? – Did May 7 '12 at 12:12

In theory, this is straightforward: the $(K+2)$-dimensional vector $(X,Y,Z)$ is jointly normal, hence there exists some deterministic $x$, $\eta$ and $\zeta$ in $\mathbb R^K$ and a centered $K$-dimensional normal random vector $T$, with variance-covariance matrix $C$, independent of $(Y,Z)$, such that conditionally on $(Y,Z)$, $X$ is distributed as $$ U=x+Y\eta+Z\zeta+T. $$ Thus, conditionally on $A^2=Y^2+Z^2$, $X$ is distributed as $x+T+V$, where $V$ follows the conditional distribution of $Y\eta+Z\zeta$ conditionally on $A^2$.

First step: To find $(x,\eta,\zeta,C)$, one solves the system $$ (1)\ \mathrm E(U)=\mathrm E(X),\quad (2)\ \mathrm E(UY)=\mathrm E(XY),\quad (3)\ \mathrm E(UZ)=\mathrm E(XZ), $$ and $$ (4)\ \mathrm{var}(Y\eta+Z\zeta)+C=\mathrm{var}(X). $$ These four equations are equivalent to $(1)\ x+\mu_Y\eta+\mu_Z\zeta=\mu_X$, $$ (2)\ x\mu_Y+\Sigma_{YY}\eta+\Sigma_{YZ}\zeta=\Sigma_{XY},\qquad (3)\ x\mu_Z+\Sigma_{YZ}\eta+\Sigma_{ZZ}\zeta=\Sigma_{XZ}, $$ and $$ (4)\ \Sigma_{YY}\eta\eta^*+\Sigma_{YZ}(\eta\zeta^*+\zeta\eta^*)+\Sigma_{ZZ}\zeta\zeta^*+C=\Sigma_{XX}, $$ which, in the general case, determine uniquely the $K$-dimensional vectors $x$, $\eta$ and $\zeta$ and the $K\times K$ matrix $C$. (Note that $\mu_Y$, $\mu_Z$, $\Sigma_{YY}$, $\Sigma_{YZ}=\Sigma_{ZY}$ and $\Sigma_{ZZ}$ are real numbers while $x$, $\eta$, $\zeta$, $\Sigma_{XY}$ and $\Sigma_{XZ}$ are $K$-dimensional vectors and $\Sigma_{XX}$, $\eta\eta^*$, $\eta\zeta^*$, $\zeta\eta^*$, $\zeta\zeta^*$ and $C$ are $K\times K$ matrices.)

Second step: Basically, a two dimensional gaussian vector $(Y,Z)$ with prescribed distribution is given and one looks for the distribution of the vector combination $Y\eta+Z\zeta$ conditionally on $A^2=Y^2+Z^2$, for possibly any $\eta$ and $\zeta$ in $\mathbb R^K$.

There exists some i.i.d. standard gaussian vector $(Y_0,Z_0)$ and a $2\times2$ matrix $B$ such that $(Y,Z)=M+(Y_0,Z_0)B$, where $M=(\mu_Y,\mu_Z)$ and $B^*B$ is the variance-covariance matrix of $(Y,Z)^*$. Thus, $A^2=a^2$ corresponds to an ellipse in the $(Y_0,Z_0)$ plane and one seeks the distribution of some ($K$-dimensional) affine combination of $(Y_0,Z_0)$ conditionally on the fact that $(Y_0,Z_0)$ is on this ellipse.

I doubt that some convenient formula exists in the general case, but maybe some more specific hypotheses are made.

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.