Research Seminar
Visual representation of distributions of covariance matricesTomoki TokudaKU Leuven | |
| Abstract: | Distributions of covariance matrices play an important role in Bayesian statistics. For instance, when we deal with a model that includes a multivariate normal distribution, we must assume some prior distribution for the covariance matrix. For this purpose, one may typically opt for an inverse Wishart distribution with a particular setting of tuning parameters. This choice, however, is usually made 'blindly' without taking into account the 'shape' or the properties of the distribution in question. The major reason for such a blind attitude can be attributed to the difficulty of visualizing a distribution of covariance matrices due to its high dimensionality. As a remedy for this problem, in this talk I will propose two types of visualization methods. It will be shown that these methods may unveil properties of the inverse Wishart distribution that are less obvious and that do not necessarily follow from known analytical results on characteristics of the distribution. Furthermore, the methods will also be applied to other possible distributions of covariance matrices such as the scaled inverse Wishart and Jeffreys' prior, and the differences between all distributions involved will be discussed. |
| Date: | Tue Mar 9, 12:15 pm - 1:15 pm |
| Place: | room 02.51 (Department of Psychology, Tiensestraat 102, 3000 Leuven) |