The mixed linear model is characterized using the classic linear model of Gauss-Markov. The multipliers of Lagrange are a tool to obtain the best lineal predictors (BLUP), we shown the results of Searle (1997), where some sums of the best linear unbiased predictors of random effects are zero. This characteristic is similar with the reparametrization ∑-restriction in the fixed linear models. We present an illustration based on results of Gaona (2000) in crossed classification with the data measured in young bulls sanmartiniano, and other example in hierarchical models with the results presented in Harville & Fenech (1985) corresponding to mensurations of weight of a group of male sheep. In the usual model of analysis of variance for mixed models, some sums of the unbiased lineal predictors (BLUP) associated to random effects are zero when the model has a single variable answer, however, this property does not work in cases in which there are different evaluations in the same experimental unit, which will be correlated.