Revert back to older version of ITensors.jl

ITensor Julia Questions
1

Hi all,

Is there a straightforward way to revert back to (or do a fresh install of) older version of ITensors.jl, say v0.3.43? I have tried ] add ITensors@0.3.43, but then it is not compatible with other packages that are downloaded with ITensors.jl, mostly NDTensors.jl that has gone through huge revamp. So, performing a fresh install or replacing the new version with ITensors@0.3.43 by hand throws lots of errors due to incompatibility. I also tried to dev the downloaded v0.3.43, but it was still downloading the latest NDTensors.

Actually, I was working with a long-range 2D Hamiltonian, where a hand written version of DMRG was working as expected (matched with analytical results, where available). But then I performed the update of ITensors.jl, and in the new version (0.3.52) the same code is giving totally unexpected results. I wanted to revert back to old version to understand the origin of the problem.

Thanks in advance!

Regards,
Titas

2

Sorry to hear you had an issue with the latest version of ITensors.jl.

I would start by making a new local environment (see 4. Working with Environment · Pkg.jl) that has nothing installed, and then try ] add ITensors@0.3.43. That should install all of the dependencies from scratch and avoid conflicts.

3

I tried that. I also tried by removing the entire Julia installation and reinstalling it again (removed hidden cache files by hand). But still such fresh installation was downloading the latest NDTensors.jl.

4

I see, I didn’t realize you also wanted a specific version of NDTensors. Can’t you then just add NDTensors at a specific version as well?

1 Like
5

Actually, older version of ITensors.jl seems incompatible with newer NDTensors.jl.

Let me try ] dev 'ing the older NDTensors as well.

6

Sorry, it did not work. I performed ] dev on NDTensors@0.2.11 that comes with v0.3.43. I am getting following error. It seems there is incompatibility with another package.

ERROR: LoadError: MethodError: no method matching sort(::Tuple{Int64, Int64, Int64})

Closest candidates are:
  sort(::AbstractUnitRange)
   @ Base range.jl:1397
  sort(::AbstractRange)
   @ Base range.jl:1400
  sort(::Dictionaries.AbstractIndices{I}; kwargs...) where I
   @ Dictionaries ~/.julia/packages/Dictionaries/7aBxp/src/AbstractIndices.jl:376
  ...

Stacktrace:
  [1] permutedims_combine_output(T::NDTensors.BlockSparseTensor{Float64, 5, NTuple{5, Index{Vector{Pair{QN, Int64}}}}, NDTensors.BlockSparse{Float64, Vector{Float64}, 5}}, is::Tuple{Index{Vector{Pair{QN, Int64}}}, Index{Vector{Pair{QN, Int64}}}, Index{Vector{Pair{QN, Int64}}}}, perm::NTuple{5, Int64}, combdims::Tuple{Int64, Int64, Int64}, blockperm::Vector{Int64}, blockcomb::Vector{Int64})
    @ NDTensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/NDTensors/src/blocksparse/blocksparsetensor.jl:412
  [2] permutedims_combine(T::NDTensors.BlockSparseTensor{Float64, 5, NTuple{5, Index{Vector{Pair{QN, Int64}}}}, NDTensors.BlockSparse{Float64, Vector{Float64}, 5}}, is::Tuple{Index{Vector{Pair{QN, Int64}}}, Index{Vector{Pair{QN, Int64}}}, Index{Vector{Pair{QN, Int64}}}}, perm::NTuple{5, Int64}, combdims::Tuple{Int64, Int64, Int64}, blockperm::Vector{Int64}, blockcomb::Vector{Int64})
    @ NDTensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/NDTensors/src/blocksparse/blocksparsetensor.jl:440
  [3] contract(tensor::NDTensors.BlockSparseTensor{Float64, 5, NTuple{5, Index{Vector{Pair{QN, Int64}}}}, NDTensors.BlockSparse{Float64, Vector{Float64}, 5}}, tensor_labels::NTuple{5, Int64}, combiner_tensor::NDTensors.Tensor{Number, 4, NDTensors.Combiner, NTuple{4, Index{Vector{Pair{QN, Int64}}}}}, combiner_tensor_labels::NTuple{4, Int64})
    @ NDTensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/NDTensors/src/blocksparse/combiner.jl:72
  [4] _contract(A::NDTensors.BlockSparseTensor{Float64, 5, NTuple{5, Index{Vector{Pair{QN, Int64}}}}, NDTensors.BlockSparse{Float64, Vector{Float64}, 5}}, B::NDTensors.Tensor{Number, 4, NDTensors.Combiner, NTuple{4, Index{Vector{Pair{QN, Int64}}}}})
    @ ITensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/src/tensor_operations/tensor_algebra.jl:3
  [5] _contract(A::ITensor, B::ITensor)
    @ ITensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/src/tensor_operations/tensor_algebra.jl:9
  [6] contract(A::ITensor, B::ITensor)
    @ ITensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/src/tensor_operations/tensor_algebra.jl:104
  [7] _contract(As::Tuple{ITensor, ITensor, ITensor}, sequence::Vector{Int64}; kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ ITensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/src/tensor_operations/tensor_algebra.jl:207
  [8] _contract(As::Tuple{ITensor, ITensor, ITensor}, sequence::Vector{Int64})
    @ ITensors ~/MEGA/projects/FQH_cavity/julia_test/0.3.43/ITensors.jl-0.3.43/src/tensor_operations/tensor_algebra.jl:206
7

That’s an issue that Julia temporarily define sort(::Tuple) but then removed it. You could try ] add Compat@4.10.0.

8

Great!!! It worked like a charm !!! Thanks !!!

1 Like
9

Glad that worked, please keep us updated about any issues you are finding in more recent versions of NDTensors/ITensors.

10

Hi again,

The problem I am facing starts from v0.3.47, and till v0.3.46 it was working well.

In my case, I have a long-range Hamiltonian, and I am doing DMRG to get the GS. In the solvable limit, I know the exact energy to be zero. Since the Hamiltonian is long-range, DMRG has a tendency to get stuck in some local minima (there are lots of them). Till v0.3.46, a noise level of 1e-8 was enough to get the energy til E~1e-7 (this is the limit I can get considering all the approximations used in construction of the MPO). But from v0.3.47, smaller noise like 1e-8 gets stuck at E~1.1…, and with larger noise > 1e-3 it gets stuck at E~1e-4. There is no way I could reduce the energy to E~1e-7 using noise. Only by doing a global subspace expansion (ala TDVP), I could get again E~1e-7.

I will try to figure out the exact causes for such behavior.

11

Perhaps related to this change: Fix behavior of factorize mindim by ryanlevy · Pull Request #1214 · ITensor/ITensors.jl · GitHub? That was included in the v0.3.47 release.

13

Hi Matt,

Sorry for the late reply. I deleted the previous post, as I was wrong.
You are right, the cause of the problem is that:

  Lis = indices(Linds...)
  dL, dR = dim(Lis), dim(indices(setdiff(inds(A), Lis)))
  maxdim = get(kwargs, :maxdim, min(dL, dR))

was replaced by

  Lis = commoninds(A, indices(Linds...))
  Ris = uniqueinds(A, Lis)
  dL, dR = dim(Lis), dim(Ris)
  # maxdim is forced to be at most the max given SVD
  if isnothing(maxdim)
    maxdim = min(dL, dR)
  end
  maxdim = min(maxdim, min(dL, dR))

in src/tensor_operations/matrix_decomposition.jl.

While the change in the PR #1214 is justified for the situations where noise is zero. But in presence of perturbations, the restriction of maxdim in factorize_eigen is strongly affecting the convergence of DMRG.

For example, even presence of noise, the bond dimension in the bulk is getting bounded by 16 after one fullsweep (left-to-right and right-to-left) for spin=1/2 systems.

14

I guess that makes sense, since the noise term can make those eigenvalues nonzero.

Maybe we can just remove that maxdim restriction if !isnothing(eigen_perturbation), or set it to max(dL, dR). @ryanlevy, any comments on this?

Probably the noise term should not be handled inside of factorize since it doesn’t seem like the right place for it, but that’s a separate issue.

15

@titaschanda feel free to make a PR to fix this, not sure I’ll be able to get to it soon.

16

Sorry about my delayed reply. max(dL, dR) is still a bit strange to me, since factorize as a standalone object I would imagine shouldn’t be allowed to grow the rank. But for just eigen I think the max restriction is reasonable for the moment

17

Right, but when you pass a noise term, you are actually factorizing a sum of two tensors (based on the original one and the noise term), so I think the issue is that the rank of that sum of tensors can exceed the maximum dimension set in Fix behavior of factorize mindim by ryanlevy · Pull Request #1214 · ITensor/ITensors.jl · GitHub.

18

@titaschanda I proposed a fix in [ITensors] Fix `eigen_perturbation` rank by mtfishman · Pull Request #1302 · ITensor/ITensors.jl · GitHub, could you test it out and see if it fixes the issue you are seeing?

19

Hi @mtfishman, great! I will test this over the weekends, and get back to you. Enormously grateful, as always!

20

Hi Matt, it works!! :smiley:

1 Like
21

Glad to hear it, I merged the PR and it will be included in the next release.