# Matrix Derivatives

I'm trying to calculate a MLE and was wondering if anyone had a reference or could help me with derivatives involving matrices:

$$\frac{\partial}{\partial \sigma^2} \sum\limits_{k = 1}^{K} (Y_k - \mu)^T\frac{1}{\sigma^2}(I_n - \frac{\tau^2}{\sigma^2 + n\tau^2}1_n)(Y_k - \mu)$$ or $$\frac{\partial}{\partial \sigma^2} \sum\limits_{k = 1}^{K} (Y_k - \mu)^T\frac{1}{\sigma^2}(I_n - \frac{\tau^2}{\lambda}1_n)(Y_k - \mu)$$ where $$\lambda = \sigma^2 + n\tau^2$$

Basically this is the last term in a log-likelihood of a multivariate normal, $$MVN(\mu, \sigma^2 I_n + \tau^2 1_n)$$

EDIT: $$Y,\mu, I_n, 1_n \text{ are matrices.}$$ $$\sigma^2, \tau^2\text{ are scalars.}$$

• I don't see any matrix variables anywhere in this post: could you please explain what your notation means? – whuber Feb 19 '16 at 22:01
• Since you have revealed that $\sigma^2$ is a scalar, then there are no matrix derivatives anywhere in this question. In fact, once you simplify the sums you will discover that there aren't any matrices involved in them anyway. Ordinary (one-variable!) Calculus is all you need to know. – whuber Feb 19 '16 at 22:23