A training sequence is created for space-time diversity arrangement, having any training sequence length, while limiting the training sequence to a standard constellation. Given a number of channel unknowns that need to be estimated, L, a training sequence can be creates that yields minimum means squared estimation error for lengths Nt=kNPRUS+L−1, for any positive integer k≧1, where NPRUS is a selected perfect roots-of-unity sequence (PRUS) of length N. The training sequence is created by concatenating k of the length N perfect roots-of-unity sequences, followed by L−1 initial symbols of that same PRUS. Good training sequences can be created for lengths Nt that cannot be obtained through the above method by concatenating a requisite number of symbols found through an exhaustive search.