Excited State DMRG with saving to disk

ITensor Julia Questions
,
1

Hi there,

I am trying to target the 1st excited state through

energy,psi = dmrg(H,[psi0],psix,sweeps; write_when_maxdim_exceeds=500)

However, I got the following error message:

ERROR: LoadError: MethodError: no method matching disk(::ProjMPO_MPS; path=“/scratch/feng/tmp”)
Closest candidates are:
disk(::SerializedElementArrays.SerializedElementArray; kw…) at ~/.julia/packages/SerializedElementArrays/cdFxy/src/SerializedElementArrays.jl:153
disk(::AbstractArray; kw…) at ~/.julia/packages/SerializedElementArrays/cdFxy/src/SerializedElementArrays.jl:151
disk(::ITensors.AbstractSum; disk_kwargs…) at ~/.julia/dev/ITensors/src/mps/projmposum.jl:110
…
Stacktrace:
[1] dmrg(PH::ProjMPO_MPS, psi0::MPS, sweeps::Sweeps; kwargs::Base.Pairs{Symbol, Any, NTuple{6, Symbol}, NamedTuple{(:start_disk, :write_when_maxdim_exceeds, :write_step, :eigsolve_tol, :eigsolve_krylovdim, :weight), Tuple{Bool, Int64, String, Float64, Int64, Int64}}})
@ ITensors ~/.julia/dev/ITensors/src/mps/dmrg.jl:225
[2] dmrg(H::MPO, Ms::Vector{MPS}, psi0::MPS, sweeps::Sweeps; kwargs::Base.Pairs{Symbol, Any, NTuple{6, Symbol}, NamedTuple{(:start_disk, :write_when_maxdim_exceeds, :write_step, :eigsolve_tol, :eigsolve_krylovdim, :weight), Tuple{Bool, Int64, String, Float64, Int64, Int64}}})
@ ITensors ~/.julia/dev/ITensors/src/mps/dmrg.jl:55

It seems that disk() cannot treat ProjMPO_MPS type object. Can you help?

Best,
Simon

2

Thanks for the report, that will be simple to fix.

3

Hi Matthew,

Thanks for the reply. Can you explain how?

Best,
Simon

4

If you want something working for yourself, you could look at the design of DiskProjMPO: ITensors.jl/src/mps/diskprojmpo.jl at main · ITensor/ITensors.jl · GitHub.

The conceptually simplest strategy would be to make a DiskProjMPS type, based on combining the design of DiskProjMPO and ProjMPS (ITensors.jl/src/mps/projmps.jl at main · ITensor/ITensors.jl · GitHub). Then you could create a DiskProjMPO_MPS similar to the ProjMPO_MPS type (ITensors.jl/src/mps/projmpo_mps.jl at main · ITensor/ITensors.jl · GitHub).

If I were to do it myself, I think I would do a bigger rewrite of those types to share more code across the projected MPO/MPS types types that are stored in memory and on disk, so for example redesign ProjMPO, ProjMPS, and ProjMPO_MPS to be parametrized types that either can be fully in memory or partially on disk depending on what objects they are storing. I may go ahead and do that but if you want something done faster for yourself, you could take the strategy I outlined above.

1 Like
5

An update that @miles will be looking into implementing this.

6

Sounds great! Thank you!