How to use

Compression Features

• Compression with SVD
• Bond dimension determined by fixed value or SVD threshold or a mixed approach
• For fixed value, uniform distribution and Gaussian distribution are supported
• use CompressConfig to control the behavior

Ground State Optimization Features

• Ground state or several lowest excited states
• One-site or two-site optimization
• Mix sub-optimal states by procedure to avoid local minimum
• Control bond dimension by procedure to accelerate computation at an early stage and achieve high accuracy in the end.
• TDH hybrid optimization configs
• use OptmizeConfig to control the behavior
• No convergence on the optimization is guaranteed and the user should check the outcome before using it

Evolution Features

• Support P&C, TDVP_MU, TDVP_PS as evolution method.
• All schemes could be adaptive
• TDVP_PS support Krylov subspace method
• TDVP_MU_VMF: variable mean-field with RKF45 time integration (VMF could automatically determine the time step dt if the rtol/atol and the regularization parameters are given.)
• use EvolveConfig to control the behavior

More details about the evolution schemes, please refer to arXiv:1907.12044.

Parallelism

We set environment variables to limit NumPy cpu usage when Renormalizer is imported because the parallel efficiency is poor. Preliminary tests show that using 4 cores can hopefully double the speed, while devoting more resources has little benefit.

Related code can be found in Renormalizer/__init__.py.

for env in ["MKL_NUM_THREADS", "NUMEXPR_NUM_THREADS", "OMP_NUM_THREADS"]:
    os.environ[env] = "1"
    del env

Tips

• use -O flag to enable optimizations (disable asserts) for Python can typically speed things up for 50% (time cost drops to 66%).
• use float32/complex64 as backend (rather than float64/complex128) can speed things up for 50%. Although all test cases can pass in such configuration, care should be taken because the precision is significantly lower.