All numbers are stored and all arithmetic is done in double precision. The algorithms used are generally very accurate numerically. However, when many operations are performed or when the matrices are ill-conditioned, matrix operations should be used in a numerically responsible way because numerical errors add up.