A method and system are directed to enabling a detector to perform relatively accurate channel estimates on blocks of data that include relatively short training sequences in wireless communication signals. Channel re-estimation is performed by transferring the estimation calculations from the time domain into a channel state domain. In one embodiment, information obtained from known training sequences and unknown data may be combined to generate an initial channel estimate. An embodiment of a new recursive Expectation Maximization (EM) process generally i) initializes the EM process with the initial channel estimate; ii) determines updates for the channel estimates; iii) produces a reduced observation vector; iv) generates a special sequence of channel states; v) rearranges each channel state estimate in the reduced observation vector; vi) generates a low-dimensional channel model matrix inversion based on the generated special sequence of channel states; and vii) generates channel estimates based on the generated matrix inversion.