E-catalog        

Библиогр.: 21 назв.

Изучаются свойства робастных оценок масштабного параметра, который характеризует «разброс» случайной величины. Предложенные оценки асимтотически нормально распределены, имеют ограниченные функции влияния и, следовательно, в отличие от оценки стандартного отклонения, «защищены» от наличия выбросов в выборке. Рассматриваемые оценки вычисляются на основе упорядоченной статистики, из которой предва-рительно удаляется часть наблюдений. Предложен адаптивный вариант оценок, основанный на использова-нии выборочных оценок функционалов, характеризующих степень «затянутости хвостов» распределений. Приводятся результаты сравнения оценок масштабного параметра в условиях различных моделей наблюде-ний, в частности в рамках гауссовской модели с масштабным засорением. In the work, various estimates of the scale parameter characterizing the “spread” of a random variable are studied. It is shown that traditionally used in practice estimates of the scale parameter, such as a sample estimate of the standard deviation, and an esti-mate of the average absolute deviations, have unlimited influence functions, and they are very sensitive to the presence of outliers in the sample. The paper proposes truncated versions of these estimates and, which are calculated not from the initial sample, but based on ordered statistics, from which the smallest and largest ordinal statistics are previously removed. These estimates are “protected” from the presence of outliers in the sample, they have limited influence functions and their characteristics depend significantly on the parameter α, which in practice leads to additional efforts to select this parameter. In the work, adaptive versions of these estimates are proposed, for which the parameter is determined based on the initial sample. The results of comparing estimates under the conditions of various observation models, in particular, under the conditions of the Gaussian model with large-scale contaminating, are presented. The results obtained lead to the following conclusion. In cases where there is no certainty that the initial distribution is Gaussian, or the sample may contain gross errors (outliers), it is more advisable to use the adaptive standard deviation or the estimate in the form of an estimate of the median of the absolute differences. These estimates have limited influence functions and, therefore, are “protected” from the presence of outliers in the sample, and they are preferable in terms of efficiency over the other estimates considered under the conditions of various observation models.