orthogonalize latents using svd(L) from FA. Are our error metrics sensitive to this rotation/scaling? i.e., do relative errors on hypotheses change?
volitional: find volitional control space (in place of Row(M2) from Blk1) using svd(L) from FA (see below). In the future, somehow regress out sensory-components of latents during Blk1.
habitual: why does it fail for certain kinematics? are errors bigger for these? check into mean(abs(angError)) vs. mean(angError).
find minimum mean error, cov using cross-validation
Quick
look at all 4 datasets
check covariance errors
recreate DAP stick plot
make comparison plot that shows all hypotheses as columns, null-space/kinematics as rows