A practical method that reduces the dimensions of a high dimensional random vector X=(x₁, x₂,…,x_p)′_p×1 is put forward. Its fundamental idea is, with the sweep operation o f matrix , to structure a few synthetical indexes (called principal variance variables) of X to depict X 's statistical feature. The theoretical foundation, audio-visual explanation and algorithm of the method are given. The method is markedly superior to of principal component analysis especially when X has serious multi-correlation.