Add `evolve!` for evolution of an `MPS` with an `MPO` #264

jofrevalles · 2024-11-27T10:48:12Z

Summary

This PR enhances the evolve! to handle the evolution of an MPS with an MPO in different canonical forms. Additionally, it includes new test cases to verify the correctness of the evolution functions.

Maybe fix code so the output indices are always like the input mps. This way things are easier to check if works properly.
Refactor and regroup some code so new canonical Forms do not need to change the evolve! funciton, but just extend it as they need.
Add more tests for variable maxdim, threshold? I don't know.

@starsfordummies Please take a look and tell me what do you think!

Also we should take a look at issue #238 (which asks to extend mixed_canonize! so it dispatches on Form). So if we already have an mps that is in MixedCanonical, there is no need to do the full sweep if we just want to change the ortogonality center.

Example

julia> using Tenet; using Test

julia> ψ = MPS([rand(2, 2), rand(2, 2, 2), rand(2, 2, 2), rand(2, 2, 2), rand(2, 2)]); normalize!(ψ)
MPS (inputs=0, outputs=5)

julia> mpo = rand(MPO; n=5, maxdim=4)
MPO (inputs=5, outputs=5)

julia> ϕ_1 = evolve!(canonize(ψ), mpo)
MPS (inputs=0, outputs=5)

julia> ϕ_2 = evolve!(mixed_canonize!(ψ, Site(3)), mpo)
MPS (inputs=0, outputs=5)

julia> size.(ϕ_1)
9-element Vector{Tuple{Int64, Vararg{Int64}}}:
 (2,)
 (2, 2, 4)
 (2, 2)
 (2,)
 (4, 2, 2)
 (2, 2)
 (4,)
 (4, 2, 4)
 (4,)

julia> function create_replacements(ϕ_1_sites::Dict, ϕ_2_sites::Dict) # Function to reindex the output tensor
           # Ensure the keys match in both dictionaries
           if keys(ϕ_1_sites) != keys(ϕ_2_sites)
               throw(ArgumentError("Keys of the dictionaries do not match."))
           end

           # Create a list of replacements
           replacements = [(ϕ_2_sites[key] => ϕ_1_sites[key]) for key in keys(ϕ_1_sites)]

           return replacements
       end
create_replacements (generic function with 1 method)

julia> t1 = contract(ϕ_1)
2×2×2×2×2 Tensor{Float64, 5, Array{Float64, 5}}: ...

julia> t2 = replace(contract(ϕ_2), create_replacements(Quantum(ϕ_1).sites, Quantum(ϕ_2).sites)...)
2×2×2×2×2 Tensor{Float64, 5, Array{Float64, 5}}: ...

julia> @test t1 ≈ t2
Test Passed

codecov · 2024-11-28T09:34:37Z

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 80.47%. Comparing base (0ba1707) to head (4b7bdaf).
Report is 10 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #264      +/-   ##
==========================================
+ Coverage   79.59%   80.47%   +0.88%     
==========================================
  Files          32       34       +2     
  Lines        2068     2182     +114     
==========================================
+ Hits         1646     1756     +110     
- Misses        422      426       +4

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

jofrevalles · 2024-11-28T17:02:49Z

src/MPS.jl

+If `threshold` or `maxdim` are not `nothing`, the tensors are truncated after each sweep at the proper value.
+"""
+function evolve!(ψ::AbstractAnsatz, mpo::AbstractMPO; threshold=nothing, maxdim=nothing, normalize=true)
+    evolve!(form(ψ), ψ, mpo; threshold, maxdim, normalize)


@starsfordummies Tomorrow I will add here the replace! so this evolve! function does not change the output indices, after that, we can merge!

jofrevalles · 2024-11-29T10:48:36Z

@starsfordummies Okay, I added the reset_inded kwarg on evolve! (which defaults to true). If this one is true, the indices after the evolution are reset to the original ones.

For me, this PR is ready to merge.

starsfordummies

Awesome! If you haven't yet, you could add an evolve() function so that you're not forced to do it inplace, other than that looks great to me!

mofeing

great! thanks! this was very needed

src/Ansatz.jl

mofeing · 2024-11-29T11:11:00Z

src/MPS.jl

+    old_norm = norm(ψ)
    if isnothing(bond) # Normalize all λ tensors
        for i in 1:(nsites(ψ) - 1)
            λ = tensors(ψ; between=(Site(i), Site(i + 1)))
-            replace!(ψ, λ => λ ./ norm(λ)^(1 / (nsites(ψ) - 1)))
+            replace!(ψ, λ => λ ./ old_norm^(1 / (nsites(ψ) - 1)))
        end
    else
        λ = tensors(ψ; between=bond)
-        replace!(ψ, λ => λ ./ norm(λ))
+        replace!(ψ, λ => λ ./ old_norm)


so in reality, there's no need for this because norm(λ) will only be computed once as it already is

Nop, you need the norm of the full mps.

I already tried that but fails on some truncate tests.

ahh i see the difference now, I thought it was the same norm call

src/MPS.jl

mofeing · 2024-11-29T11:57:30Z

src/MPS.jl

+    for i in nsites(ψ):-1:2
+        canonize_site!(ψ, Site(i); direction=:left, method=:svd)
+    end
+
+    # left-to-right SVD sweep, get left-canonical tensors and singular values and truncate
+    for i in 1:(nsites(ψ) - 1)
+        canonize_site!(ψ, Site(i); direction=:right, method=:svd)
+        (!isnothing(threshold) || !isnothing(maxdim)) &&
+            truncate!(ψ, [Site(i), Site(i + 1)]; threshold, maxdim, normalize, compute_local_svd=false)
+    end


I guess you can replace this with mixed_canonize! to the left-most site and reuse the method for MixedCanonical

Suggested change

for i in nsites(ψ):-1:2

canonize_site!(ψ, Site(i); direction=:left, method=:svd)

end

# left-to-right SVD sweep, get left-canonical tensors and singular values and truncate

for i in 1:(nsites(ψ) - 1)

canonize_site!(ψ, Site(i); direction=:right, method=:svd)

(!isnothing(threshold) || !isnothing(maxdim)) &&

truncate!(ψ, [Site(i), Site(i + 1)]; threshold, maxdim, normalize, compute_local_svd=false)

end

mixed_canonize!(ψ, site"1")

truncate_sweep(ψ; threshold, maxdim, normalize)

No you can't do this. I specifically created this function because in the Canonical form you do not contract the lambdas but still truncate them. This is different from the other methods.

okay, i want to better understand this but let's merge it and refactor it later (if we need)

mofeing · 2024-11-29T12:02:08Z

src/MPS.jl

+    if !isnothing(threshold) || !isnothing(maxdim)
+        truncate_sweep!(Canonical(), ψ; threshold, maxdim, normalize)
+    else
+        normalize && canonize!(ψ; normalize)
+    end


do you need to normalize if you don't truncate?
also, for this case (i.e. Canonical form) you will canonize 2 times if you truncate and normalize (1 in truncate_sweep for recanonization and another here for normalization)

Suggested change

if !isnothing(threshold) || !isnothing(maxdim)

truncate_sweep!(Canonical(), ψ; threshold, maxdim, normalize)

else

normalize && canonize!(ψ; normalize)

end

if !isnothing(threshold) || !isnothing(maxdim)

truncate_sweep!(Canonical(), ψ; threshold, maxdim, normalize)

normalize && canonize!(ψ; normalize)

end

I don't understand this. This is not the case, you don't need to normalize if you already normalize the bond that you truncate.

And yes, you might want to normalize if you don't truncate if the mpo is not unitary and evolves the ψ out of norm.

why would you want to normalize here a TN that it's not already normalized? but ok, I see that both things could be understandable from a semantics point of view

let's merge, discuss on Monday and refactor it if we need

jofrevalles · 2024-11-29T12:27:56Z

Okay, @mofeing I removed some unnecessary Quantum calls. This is ready to merge.

@show

* Add evolve! function for an MPS with an MPO * Add tests for MPS-MPO evolution * Remove unnecessary @show * Add comment * Fix kwarg handling on truncate! function * Fix normalize! for Canonical form and small fixes on evolve! with MPO * Extend evolve!(mps, mpo) tests * Format code * Remove unnecessary kwarg * Refactor code so it is easier to extend for other canonical forms * Fix comment * Enhance tests * Add reset_index kwarg * Update tests * Format code * Remove unnecessary Quantum functions, add docstring for truncate_sweep! * Format code * Remove stale function

jofrevalles added 3 commits November 27, 2024 11:38

Add evolve! function for an MPS with an MPO

1a34d96

Add tests for MPS-MPO evolution

8edd30d

Remove unnecessary @show

d61adf7

jofrevalles requested a review from starsfordummies November 27, 2024 10:48

jofrevalles added 6 commits November 27, 2024 11:50

Add comment

d2d08c5

Fix kwarg handling on truncate! function

1c949d2

Fix normalize! for Canonical form and small fixes on evolve! with MPO

4e58415

Extend evolve!(mps, mpo) tests

aa082cb

Format code

fe88489

Remove unnecessary kwarg

4b7bdaf

jofrevalles added 3 commits November 28, 2024 17:48

Refactor code so it is easier to extend for other canonical forms

2111b93

Fix comment

e256ede

Enhance tests

07534eb

jofrevalles commented Nov 28, 2024

View reviewed changes

jofrevalles added 2 commits November 29, 2024 11:45

Add reset_index kwarg

231bcc5

Update tests

a2ca518

jofrevalles requested a review from mofeing November 29, 2024 10:48

Format code

d8688e8

starsfordummies approved these changes Nov 29, 2024

View reviewed changes

mofeing requested changes Nov 29, 2024

View reviewed changes

jofrevalles added 3 commits November 29, 2024 13:21

Remove unnecessary Quantum functions, add docstring for truncate_sweep!

98018c3

Format code

f1d7fd8

Remove stale function

ce30ea1

mofeing approved these changes Nov 29, 2024

View reviewed changes

mofeing merged commit 8ac83f4 into master Nov 29, 2024
5 checks passed

mofeing deleted the feature/evolve-mps-with-mpo branch November 29, 2024 13:41

jofrevalles mentioned this pull request Dec 2, 2024

Extend evolve of an MPS with an MPO #169

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add `evolve!` for evolution of an `MPS` with an `MPO` #264

Add `evolve!` for evolution of an `MPS` with an `MPO` #264

jofrevalles commented Nov 27, 2024 •

edited

Loading

codecov bot commented Nov 28, 2024 •

edited

Loading

jofrevalles Nov 28, 2024

jofrevalles commented Nov 29, 2024

starsfordummies left a comment

mofeing left a comment

mofeing Nov 29, 2024

jofrevalles Nov 29, 2024

jofrevalles Nov 29, 2024

mofeing Nov 29, 2024

mofeing Nov 29, 2024

jofrevalles Nov 29, 2024

mofeing Nov 29, 2024

mofeing Nov 29, 2024

jofrevalles Nov 29, 2024

mofeing Nov 29, 2024

jofrevalles commented Nov 29, 2024

Add evolve! for evolution of an MPS with an MPO #264

Add evolve! for evolution of an MPS with an MPO #264

Conversation

jofrevalles commented Nov 27, 2024 • edited Loading

Summary

Example

codecov bot commented Nov 28, 2024 • edited Loading

Codecov Report

Choose a reason for hiding this comment

jofrevalles commented Nov 29, 2024

starsfordummies left a comment

Choose a reason for hiding this comment

mofeing left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jofrevalles commented Nov 29, 2024

Add `evolve!` for evolution of an `MPS` with an `MPO` #264

Add `evolve!` for evolution of an `MPS` with an `MPO` #264

jofrevalles commented Nov 27, 2024 •

edited

Loading

codecov bot commented Nov 28, 2024 •

edited

Loading