I wonder if someone can please help me with a passage on the article by José Pinheiro and Douglas Bates on unconstrained matrix parametrization. hat ties directly into the question.

Although the authors start with a given variance-covariance matrix, the idea is to parametrize an unknown matrix of random effects to produce positive semi-definite matrices that can be tested for likelihood.

The basic algebra is upper-case $L_i$ symbolizing the $i$ columns of the upper triangular Cholesky with lower-case $l_i$ denoting the parameters needed to represent the Cholesky matrix $L$ in spherical coordinates. As I understand it, $[l_i]_1$ corresponds to the norm of the column vector $L_i$ within the Cholesky $L$, while $[l_i]_j$ with $j>1$ are actual angles in counterclockwise Jacobi rotations (page 4), which are the basis for the geometric interpretation of the dot products resulting in scalar values $(\rho_{ij}$ - variances and covariances).

Since we don't know these parameters, the authors propose the following parametrization:

$$\theta_i=\log\left([l_i]_1\right),\quad i= 1,...,n$$


$$\theta _{n+(i-2)(i-1)/2+(j-1)} = \log\ \left( \frac{[l_i]_j}{\pi-[l_i]_j}\right), \quad i=2,...,n, \,j= 2,...,i. $$

The first parameter definition is clear - it ensures that the vectors will be positive.

As for the second rule, I am lost as to the subscripts of $\theta$. I understand that they apply to the angles of the spherical coordinates securing that they will be positive and bounded $(0,\pi]),$ but I don't see the subscripts in the $\theta _{n+(i-2)(i-1)/2+(j-1)}$ and would really appreciate some pointers.

If it helps the parameters that would correspond to a variance-covariance matrix, such in the post:

$$\begin{align} A&=\begin{bmatrix}1&1&1\\1&5&5\\1&4&14\end{bmatrix}=L^{T}L=\begin{bmatrix}1&0&0\\1&2&0\\1&2&3\end{bmatrix}\begin{bmatrix}1&1&1\\0&2&2\\0&0&3\end{bmatrix} \end{align}$$

with spherical coordinates,

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

would be, as given by the authors, $$\theta = (0, \log(5)/2, \log(14)/2, -0.608, -0.348, -0.787)^\top.$$ These values correspond to the $\ln(\cdot)$ of the first three $[l_i]_1$ values, and the $\ln\left(\frac{[l_i]_j}{\pi-l_i]_j}\right)$ of all the other values.

Can the subscripts simply be a typo?


According to the article $θn+(i−2)(i−1)/2+(j−1)=log([li]jπ−[li]j)$

This means for example that for the element 2,2 (1.107),

The folowing is true: $-0.608=log(1.107/(π-1.107))$

also is true that $1.107=exp(-0.608)*π/(1+exp(-0.608))$

  • $\begingroup$ Thank you for your answer. I +1 honestly by mistake while reading your answer on my phone. I am sure it is more than deserved, even if it were just for the intention to help, but I don't follow your response. Do you go from $log([l_i]_j\pi - [l_i]_j) = log(1.107/(\pi - 1.107))$? I guess not, and I'm just stuck with the indexes not making sense. I'd appreciate it if you could be more descriptive, and I'd happily accept the answer once I understand it. $\endgroup$ – Antoni Parellada Sep 15 '15 at 12:21

I'm not sure what you think the subscripts should be, but they're really just a way to map the (i,j) location of the index in L into the linear array $\theta$. One thing to note is the authors don't mention that $\theta$ in this section is row-wise, not column-wise as defined earlier.


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.