Fisher Information in Single Factor Analysis


The single factor analysis model for data \{ {\bf x}_1, {\bf x}_2, \ldots, {\bf x}_n \} is

 {\bf x}_i  = {\bf m} + v_i {\bf a} + \boldsymbol{\epsilon}_i \quad \; (i=1,2,\ldots,n),

where {\bf x}_i \in \mathbb{R}^k, \boldsymbol{\epsilon}_i is an i.i.d. zero-mean multivariate Gaussian random variable with a symmetric, positive definite variance-covariance matrix \boldsymbol{\Sigma} \in \mathbb{R}^{k \times k}; that is, \boldsymbol{\epsilon}_i \sim {\rm N}_k({\bf 0}, \boldsymbol{\Sigma}) for all i=1,2,\ldots,n. It is thus assumed that each data vector {\bf x}_i follows a k-variate Gaussian distribution with moments

	{\rm E} \{ {\bf x}_i \} = {\bf m} + v_i {\bf a},

	{\rm Var} \{ {\bf x} \} = \boldsymbol{\Sigma}.

The aim of single factor analysis is to estimate the unknown factor scores {\bf v} = (v_1, v_2, \ldots, v_n)^\prime \in \mathbb{R}^n, the mean {\bf m} = (m_1,m_2, \ldots, m_k)^\prime \in \mathbb{R}^k, the variance-covariance matrix \boldsymbol{\Sigma} and the factor loadings {\bf a} = (a_1, a_2, \ldots, a_k)^\prime \in \mathbb{R}^k.

I’ve written this note to show how one may derive the Fisher information for the single factor model. This turns out to be a fairly involved task requiring a lot of linear algebra and matrix differential calculus identities. I’ve uploaded a document describing all the necessary mathematical steps to the Publications section. Anyone interested in matrix differential calculus and linear algebra should find it useful.

  1. No comments yet.
(will not be published)