为什么需要正定呢？不解。。。不过相关系数矩阵应该正定了吧！

covariance matrices are generally positive semidefinite. in realcomputation involving sample data, however, numerical errors can causethe sample covariance matrix to be negative semidefinite. In that case,you can calculate the square root of the sample covariance matrixdirectly first and then compute the sample covariance matrix, tomaintain the positivity.

however if you are concerned about rank-deficiency of the covariancematrix, there are quite a few reduced-rank subspace methods (i.e.eigenvalue decomposition) that can be helpful, so that you only dealwith the principle components ( this can be found in any multivariablestatistical analysis textbooks).

hope this is helpful.

我也遇到同样的问题,二楼的答复我不太明白,能进一步解释解释吗?