slow down when sweep from right to left site

Question

I use dmrg function and set "Quite" to false to see the detail output. And I find that sweep from left to right is quickl while the opposite direction is much slower. Here I use a self-defined site set file for spin-1 Bose Hubbard model with total number and total magnetic moment conserved. Is there any reason for this result ? Or may be I should post my code later.

Answer 1

Hi Junjie,
The reason for this slowdown could depend on many things. On the face of it, it's not too surprising if the initial MPS has a very small bond dimension. This would be because in the first half sweep all of the right-most bonds of each MPS "super block" configuration are very small (super block = the basis used when doing the Davidson optimization with two center sites).

But on the second half sweep, both of the virtual indices sticking out of the super block wavefunction could be large, so one would expect this second half sweep to be slower.

However, if it's slower by an extremely large amount, there is cause for concern. In the worst case there could be a bug causing the slowdown. But unless one has evidence for a bug, it's more likely that the accuracy parameters you are using are just set to aggressively. A good idea for the first few sweeps of DMRG is to limit the maximum bond dimension (maxm) to a rather small value, since DMRG can already get good energies with a small bond dimension, and you only need a large bond dimension toward the end of the calculation if you want very high accuracy.