Open huangrzh opened 2 weeks ago
In the two versions of CTMRG in obtaining the projectors/isometries: (1) use the whole 4x4 tensors (2) use half of it (4x2), the method (2) is more stable than (1) in AD is probably becuase of the too large element in sdiag_inv_sqrt or its rrule in method (1).
Thanks for the pseudo-inverse suggestion, I opened a PR!
(1) use the whole 4x4 tensors (2) use half of it (4x2), the method (2) is more stable than (1) in AD is probably becuase of the too large element in sdiag_inv_sqrt or its rrule in method (1)
I'm not super sure what you suggest here; could give me some more details or a reference?
For large bound dimension CTM environment, the inverse of small svd spectrum can be very large and finally may cause an instability to an inf gradient. It would be helpful to set a cutoff in the sdiag_inv_sqrt and the rrule of it.