Decision theory was used to derive point estimators of allelic frequencies with optimal statistical properties. Uniparameter (2 alleles) and multiparameter (multiple alleles) estimation problems were addressed for an arbitrary locus, then results were extended to multiple loci. First, estimators satisfying the Bayes principle of average risk optimality were obtained using a multinomial sampling model, a Beta (biallelic loci) and a Dirichlet (multiallelic loci) prior and 3 different loss functions: Squared Error Loss (SEL), Kullback-Leibler Loss (KLL) and a Quadratic Error Loss (QEL). Second, these Bayes estimators were used to obtain minimax estimators by finding values of the hyperparameters such that the frequentist risk functions were constant, a condition that implies minimaxity. Finally, the admissibility of the estimators was checked using standard theorems from decision theory. The frequentist risk function of the Bayes estimator derived from KLL involved a finite sum without a closed form, hence this risk function could not be written as a simple algebraic expression. However, this does not prevent its computation. Under SEL and QEL it was possible to find Bayes-minimax-admissible estimators (BMAE). Sufficient conditions for the usual maximum likelihood estimator to be BMAE and for the risk functions tending to infinite or converging to zero were also found. In addition to optimal statistical properties, these estimators have the appealing feature of taking into account random variation in allelic frequencies. The impact of using these estimators in different areas of quantitative and population genetics needs be assessed either empirically or theoretically and this poses a problem for further research.