i am partway through addressing the imperfections. i got the data much closer by prefixing the subrange with the DC offset. obviously, it would never have matched without that. there are further issues that also have that property, that i haven't thought of. before this change, the data compared like this: (Pdb) p reconstructed_from_original array([263.0608397 , 0.95268138, 0.78292761, ..., 1.21166631, 0.78292761, 0.95268138]) (Pdb) p reconstructed_from_reconstructed array([ 2.53569771e+02, 6.97125806e-02, -2.41580764e-01, ..., -1.75374096e-01, -2.41580764e-01, 6.97125806e-02]) after the change, it now compares like this: (Pdb) p reconstructed_from_reconstructed array([245.42651702, 0.86299853, 1.39828105, ..., 0.86886888, 1.39828105, 0.86299853]) (Pdb) p reconstructed_from_original array([262.83951251, 1.20706425, 0.73003879, ..., 0.84832615, 0.73003879, 1.20706425]) the shapes are roughly the same, and the orders of magnitude of every sample are all the same. the data is still significantly distant, and it's notable that removing the frequency information collected a ton of energy at t=0, removing the original shape of the signal.