[Background: Trying to build up my math base knowledge (self-taught) to follow an explanation on page 291 of this document, dealing with spherical parametrization to estimate unknown variance-covariance matrices. The author starts by giving a concrete covariance matrix].

We have the following matrix corresponding to a variance-covariance matrix:

$$\begin{bmatrix}1&1&1\\1&5&5\\1&4&14\end{bmatrix}$$ with a Cholesky decomposition,


I need help with an explanation for the step that follows: The spherical coordinates of the elements of the Cholesky matrix are given as $l_i$, with $i$ corresponding to the first $i$ elements of the column $L_i$ of the Cholesky:

$[L_i]_1 = [l_i]_1 \,\,cos([l_i]_2)$

$[L_i]_2 = [l_i]_1\,\, sin([l_i]_2) \,\,cos([l_i]_3)$


$[L_i]_{i-1} = [l_i]_1 \,\,sin([l_i]_2)\,...\, cos([l_i]_i)$

$[L_i]_i = [l_i]_1 \,\,sin([l_i]_2)\,...\,sin([l_i]_i) $

What is the meaning of the "spherical coordinates"? And if at all possible, why is then that $\Sigma_{ii} = [l_i]_1^2$ and $\rho_{1i} = cos([l_i]_2)$, $i= 2,...,\,n$, where $\rho_{ij}$ denotes the correlation coefficient between $X_i$ and $X_j$?

This last, nested question may become self-evident after understanding spherical coordinates in this context.


OK... the symbols are more disorienting than what it really is. The spherical coordinates of the Cholesky upper triangular matrix, $L$ can be understood in 3-dimensional space and then extrapolated to higher dimensions.

Let's look at the $L$ matrix as bound column vectors $L_i$ so that in the example above, there would be three column vectors ($L_1$, $L_2$ and $L_3$). The spherical coordinates in this case would require three parameters ($l_i$) for each of the column vectors: $[l_i]_1$, $[l_i]_2$ and $[l_i]_3$, denoting respectively: 1. The length of the column vector or norm ($||v||$ or $|v|)$ symbolized by $[l_i]_1$; 2. The angle between the column vector and the $x$ axis $[l_i]_2$ (named $\Omega$ on the illustration below); and 3. The angle between the projection of the column vector on the $(x, y)$ plane and the $y$ axis (name $\Phi$ in the illustration):

enter image description here

In this way the spherical projections boil down to basic trigonometry. The missing part was the "mystery" surrounding the $l_i$ parameters.

Solving it now, the first column of $L$, $L_1$, is simply $[1\,0\,0]^T$ and the parameters are $[l_1]_1 = 1$ because this is clearly the length of the vector; and $[l_1]_2 =\Omega= 0$ , because the vector lies in the $x$ axis. Therefore, the spherical coordinates are: $[L_1]_1= 1 \,\cdot cos([l_1]_2)\, = 1\,\cdot cos(0) =1$.

The second column vector has a norm of $\sqrt {1^2+2^2}$, which is $[l_2]_1$. In this case, $\Omega = [l_2]_2=arccos(\frac{1}{\sqrt{5}})=1.107$; and $\Phi = [l_2]_3=0$, because the vector is not in the $(y,z)$ plane. Just to check, the first entry of the second column vector, $[L_2]_1 = [l_2]_1\cos([l_2]_2)= \sqrt{5}\cdot\frac{1}{\sqrt{5}}=1$, and the second entry is, $[L_2]_2 = [l_2]_1\sin([l_2]_2\cos([l_2]_3)= \sqrt{5}\cdot sin(1.107)\cdot cos(0)=2.$

The third column vector, $L_3$, has parameters: $[l_3]_1=\sqrt{14}$; $[l_3]_2=\Omega=arccos\frac{1}{\sqrt{14}}=1.3$; and $[l_3]_3=\Phi=arctg\frac{3}{2}=0.983$. Verifying, the spherical coordinates would be: $[L_3]_1=\sqrt{14}\cdot cos(1.3)=\sqrt{14}\frac{1}{sqrt{14}}=1$; $[L_3]_2=\sqrt{14}\cdot sin(1.3)\cdot cos(0.983)=2$; and $[L_3]_3=\sqrt{14}\cdot sin(1.3)\cdot sin(0.983) =3.$

Therefore, the spherical coordinates of $L$ are:

$\begin{bmatrix}1&\sqrt{5}&\sqrt{14}\\ 0&1.107&1.3\\ 0&0&0.983\end{bmatrix}.$

As for the second part of the question, $\Sigma_{ii} = [l_i]_1^2$ because it is the square of the norm of the column vector in the Cholesky matrix, which is simply a dot product of any column vector of the Cholesky with itself.

$\rho_{1i} = cos([l_i]_2)$, $i= 2,...,\,n$, where $\rho_{ij}$ because the first column vector in the Cholesky is of the form $[[l_1]_1,0,0,...,n]$ and hence, the value of the angle $\Omega$ is zero, meaning that the angles of the cosines of the $\Omega$ angles of the other columns are really a projection of the any other column vector on the first one, or a dot product.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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