I have a random matrix that we suppose follows N(m,C). The problem is that m is also a matrix and then C must be a tensor. How to deal with partial derevatives in MLE minimization ?
the matrix normal distribution is a generalization of the multivariate normal in cases of random matrices. See chapter 2 of Gupta and Nagar (1999), Matrix Variate Distributions. A useful property is that if a matrix X follows a matrix normal distribution, then the vector which arises by stacking all columns of X together follows a multivariate normal distribution. So you can work out the ML procedure using the vectorized representation.