The problem of how to maximize the return on a given portfolio of assets within the theory of Markowitz has been given considerable attention in the literature and improvements of standard methods continues to progress. Recent developments, often based on Stein estimators or other regularized estimators, usually focus on settings when the numbers of assets (say p) is close to the number of observations (n) since this is the scenario met in most real applications. Before any specific method is applied investors would want to know the basic properties and the relative performance of them. The performance of any estimation method, however, depends on which quality criterea of judgement is being used. Proposed methods may be optimal with respect to precision of the parameters involved in the portfolio procedure, on the proximity between estimated vs true global minimum variance portfolio (GMVP) weights, on the out-of-sample performance etc. Moreover, regularized estimators are often associated with very complicated or even unknown sampling distributions, which in turn complicate statistical inference drastically. The extent to which a method allows for statistical inference therefore also becomes an important matter when judging the properties of a data driven GMVP estimator. In this paper we give an in-depth discussion of risk critereas and their impact on GMVP optimization. A Monte Carlo simulation investigating the properties of some common estimators, including a new one proposed by the authors, with respect to several quality critereas is included to compare and contrast recent proposals. An empirical study is also included using Stockholm stock exchange data.