rename "ExpandedILR" to "ILR" since no parameter expansion happens here
simplify computation of semi-orthogonal matrix. In particular, we now construct the matrix directly as in the original paper instead of 1) constructing contrasts, 2) making matrix full-rank, 3) inverting the matrix.
bring notation in-line with that in the text
Numerically, I believe the previous implementation was equivalent to this one.
This PR updates the ILR implementations:
Numerically, I believe the previous implementation was equivalent to this one.
Supersedes #64