Fix FullInfiniteProjector
#107
Conversation
Codecov ReportAttention: Patch coverage is
|
|
Thanks for the fix! Doing just the SVD simplifies things and perhaps makes the full-infinite scheme more stable in the reverse-mode differentiation. I will quickly adapt the fixes to the code convention (i.e. outsource the contractions to the contraction file and naming). Also I will add some tools for working the the actual full-infinite environment which I coded up for #99 but then didn't use because of the QR. (This will make it possible to use function-handle SVDs for the full-infinite algorithm in the future such that the environments don't have to be constructed as dense tensors.) |
Yes that was the plan. I'll write up tests once we fully implement sparse CTMRG :-) |
In Corboz's original way (PRL 113 046402) to get projectors from the full-infinite environment, he first performs QR on the two halves and then SVD on the two R tensors. However, it is later refined in PRB 98 235148 that there's no need to do QR first, and we can directly do SVD on the two halves. This PR implements the latter method to replace the original (instead of adding an option to switch between the two).
Before SVD, the indices in the lower half of the two half-infinite environments should be contracted instead of multiplied (as two linear maps; see original here) to include fermion signs.
This will also fix the warning message reported in Warning when precompiling: using OhMyThreads.@set in module PEPSKit conflicts with an existing identifier #104.