The low decoding complexity structure of Linear Dispersion Space Time Block Codes (LDSTBCs) with unitary weight matrices is analyzed. It is shown that given n = 2alpha, the maximum number of groups in which the information symbols can be separated and decoded independently is (2a + 2), and as we lower the number of different groups to (2k + 2), 0 les k les alpha, we get higher rate codes. We also find the analytic expression for rates that such codes can achieve for any chosen group number, thus completely characterizing the rate-ML-decoding-complexity tradeoff for this class of codes. The proof of the result also includes a method for constructing such optimal rate achieving codes. Interestingly, this analysis produces some low decoding complexity codes with rate greater than one.