Currently there are two truncation policies: chi and truncation_fidelity. We could add an additional one max_mem that attempts to keep the state representation below a certain memory threshold. Note that there are multiple challenges here that would need to be tackled:
How do we decide which bonds to truncate?
Can we guarantee that there will be no OOM errors? Notice that the memory required by methods such as SVD is much larger than the memory required just to store the MPS. Can we reliably predict the memory used?
Can we combine this policy with the other two? So that the user sets a desired policy, but may set a flag prevent_oom to allow the library to apply extra truncations in an attempt to avoid an OOM error.
Currently there are two truncation policies:
chiandtruncation_fidelity. We could add an additional onemax_memthat attempts to keep the state representation below a certain memory threshold. Note that there are multiple challenges here that would need to be tackled:prevent_oomto allow the library to apply extra truncations in an attempt to avoid an OOM error.