Absorb ITensors.ITensorMPS and ITensorTDVP.jl

.gitignore

@@ -25,3 +25,5 @@ Manifest.toml
 
 # Vi/Vim backup files
 .*.swp
+
+.DS_Store

Project.toml

@@ -1,17 +1,53 @@
 name = "ITensorMPS"
 uuid = "0d1a4710-d33b-49a5-8f18-73bdf49b47e2"
 authors = ["Matthew Fishman <[email protected]>", "Miles Stoudenmire <[email protected]>"]
-version = "0.2.6"
+version = "0.3.0"
 
 [deps]
-ITensorTDVP = "25707e16-a4db-4a07-99d9-4d67b7af0342"
+Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
+Compat = "34da2185-b29b-5c13-b0c7-acf172513d20"
 ITensors = "9136182c-28ba-11e9-034c-db9fb085ebd5"
-Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
+IsApprox = "28f27b66-4bd8-47e7-9110-e2746eb8bed7"
+KrylovKit = "0b1a1467-8014-51b9-945f-bf0ae24f4b77"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
+NDTensors = "23ae76d9-e61a-49c4-8f12-3f1a16adf9cf"
+Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
+Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+SerializedElementArrays = "d3ce8812-9567-47e9-a7b5-65a6d70a3065"
+TupleTools = "9d95972d-f1c8-5527-a6e0-b4b365fa01f6"
+
+[weakdeps]
+ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
+HDF5 = "f67ccb44-e63f-5c2f-98bd-6dc0ccc4ba2f"
+Observers = "338f10d5-c7f1-4033-a7d1-f9dec39bcaa0"
+PackageCompiler = "9b87118b-4619-50d2-8e1e-99f35a4d4d9d"
+ZygoteRules = "700de1a5-db45-46bc-99cf-38207098b444"
+
+[extensions]
+ITensorMPSChainRulesCoreExt = "ChainRulesCore"
+ITensorMPSHDF5Ext = "HDF5"
+ITensorMPSObserversExt = "Observers"
+ITensorMPSPackageCompilerExt = "PackageCompiler"
+ITensorMPSZygoteRulesExt = ["ChainRulesCore", "ZygoteRules"]
 
 [compat]
-ITensorTDVP = "0.4.1"
-ITensors = "0.6.7"
-Reexport = "1"
+Adapt = "4.1.0"
+ChainRulesCore = "1.10"
+Compat = "4.16.0"
+HDF5 = "0.14, 0.15, 0.16, 0.17"
+ITensors = "0.7"
+IsApprox = "2.0.0"
+KrylovKit = "0.8.1"
+LinearAlgebra = "1.10"
+NDTensors = "0.3.46"
+Observers = "0.2"
+PackageCompiler = "1, 2"
+Printf = "1.10"
+Random = "1.10"
+SerializedElementArrays = "0.1.0"
+Test = "1.10"
+TupleTools = "1.5.0"
+ZygoteRules = "0.2.2"
 julia = "1.10"
 
 [extras]

README.md

@@ -1,22 +1,20 @@
 # ITensorMPS.jl
 
-[![Stable](https://img.shields.io/badge/docs-stable-blue.svg)](https://itensor.github.io/ITensorMPS.jl/stable/)
-[![Dev](https://img.shields.io/badge/docs-dev-blue.svg)](https://itensor.github.io/ITensorMPS.jl/dev/)
+[![Docs](https://img.shields.io/badge/docs-latest-blue.svg)](https://itensor.github.io/ITensors.jl/dev/)
 [![Build Status](https://github.com/ITensor/ITensorMPS.jl/actions/workflows/CI.yml/badge.svg?branch=main)](https://github.com/ITensor/ITensorMPS.jl/actions/workflows/CI.yml?query=branch%3Amain)
-[![Coverage](https://codecov.io/gh/ITensor/ITensorMPS.jl/branch/main/graph/badge.svg)](https://codecov.io/gh/ITensor/ITensorMPS.jl)
 [![Code Style: Blue](https://img.shields.io/badge/code%20style-blue-4495d1.svg)](https://github.com/invenia/BlueStyle)
 
-Finite MPS and MPO methods based on ITensor (ITensors.jl).
+Finite MPS and MPO methods based on the Julia version of [ITensor](https://www.itensor.org) ([ITensors.jl](https://github.com/ITensor/ITensors.jl)). See the [ITensors.jl documentation](https://itensor.github.io/ITensors.jl/dev/) for more details.
 
-This package currently re-exports the MPS and MPO functionality of the [ITensors.jl](https://github.com/ITensor/ITensors.jl), including functionality like DMRG, applying MPO to MPS, applying gates to MPS and MPO, etc. See the [ITensor documentation](https://itensor.github.io/ITensors.jl/dev) for guides and examples on using this package.
+## News
 
-Additionally, it re-exports the functionality of the [ITensorTDVP.jl](https://github.com/ITensor/ITensorTDVP.jl) package, which provides other DMRG-like MPS solvers such as TDVP and MPS linear equation solving.
+### ITensorMPS.jl v0.3 release notes
 
-## Upgrade guide
+All MPS/MPO code from [ITensors.jl](https://github.com/ITensor/ITensors.jl) and [ITensorTDVP.jl](https://github.com/ITensor/ITensorTDVP.jl) has been moved into this repository and this repository now relies on ITensors.jl v0.7 and above. All of the MPS/MPO functionality that was previously in ITensors.jl and ITensorTDVP.jl will be developed here from now on. For users of this repository, this change should not break any code, though please let us know if you have any issues.
 
-The goal will be to move the MPS and MPO code from the ITensors.jl package, along with all of the code from the [ITensorTDVP.jl](https://github.com/ITensor/ITensorTDVP.jl) package, into this repository. If you are using any MPS/MPO functionality of ITensors.jl, such as the `MPS` and `MPO` types or constructors thereof (like `randomMPS`), `OpSum`, `siteinds`, `dmrg`, `apply`, etc. you should install the ITensorMPS.jl package with `import Pkg; Pkg.add("ITensorMPS")` and add `using ITensorMPS` to your code. Additionally, if you are currently using [ITensorTDVP.jl](https://github.com/ITensor/ITensorTDVP.jl), you should replace `using ITensorTDVP` with `using ITensorMPS` in your codes.
+#### Upgrade guide
 
-## News
+If you are using any MPS/MPO functionality of ITensors.jl, such as the `MPS` and `MPO` types or constructors thereof (like `random_mps`), `OpSum`, `siteinds`, `dmrg`, `apply`, etc. you should install the ITensorMPS.jl package with `import Pkg; Pkg.add("ITensorMPS")` and add `using ITensorMPS` to your code. Additionally, if you are currently using [ITensorTDVP.jl](https://github.com/ITensor/ITensorTDVP.jl), you should replace `using ITensorTDVP` with `using ITensorMPS` in your code.
 
 ### ITensorMPS.jl v0.2.1 release notes
 

examples/autodiff/circuit_optimization/op.jl

@@ -0,0 +1,52 @@
+using ITensors, ITensorMPS
+using Zygote
+
+s = siteind("Qubit")
+
+f(x) = op("Ry", s; θ=x)[1, 1]
+
+x = 0.2
+@show f(x), cos(x / 2)
+@show f'(x), -sin(x / 2) / 2
+
+# Simple gate optimization
+ψ0 = state(s, "0")
+ψp = state(s, "+")
+
+function loss(x)
+  U = op("Ry", s; θ=x)
+  Uψ0 = replaceprime(U * ψ0, 1 => 0)
+  return -(dag(ψp) * Uψ0)[]
+end
+
+# Extremely simple gradient descent implementation,
+# where gradients are computing with automatic differentiation
+# using Zygote.
+function gradient_descent(f, x0; γ, nsteps, grad_tol)
+  @show γ, nsteps
+  x = x0
+  f_x = f(x)
+  ∇f_x = f'(x)
+  step = 0
+  @show step, x, f_x, ∇f_x
+  for step in 1:nsteps
+    x -= γ * ∇f_x
+    f_x = f(x)
+    ∇f_x = f'(x)
+    @show step, x, f_x, ∇f_x
+    if norm(∇f_x) ≤ grad_tol
+      break
+    end
+  end
+  return x, f_x, ∇f_x
+end
+
+x0 = 0
+γ = 2.0 # Learning rate
+nsteps = 30 # Number of steps of gradient descent
+grad_tol = 1e-4 # Stop if gradient falls below this value
+x, loss_x, ∇loss_x = gradient_descent(loss, x0; γ=γ, nsteps=nsteps, grad_tol=grad_tol)
+
+@show x0, loss(x0)
+@show x, loss(x)
+@show π / 2, loss(π / 2)

examples/autodiff/circuit_optimization/state_preparation.jl

@@ -0,0 +1,65 @@
+using ITensors, ITensorMPS
+using OptimKit
+using Random
+using Zygote
+
+nsites = 20 # Number of sites
+nlayers = 3 # Layers of gates in the ansatz
+gradtol = 1e-4 # Tolerance for stopping gradient descent
+
+# A layer of the circuit we want to optimize
+function layer(nsites, θ⃗)
+  RY_layer = [("Ry", (n,), (θ=θ⃗[n],)) for n in 1:nsites]
+  CX_layer = [("CX", (n, n + 1)) for n in 1:2:(nsites - 1)]
+  return [RY_layer; CX_layer]
+end
+
+# The variational circuit we want to optimize
+function variational_circuit(nsites, nlayers, θ⃗)
+  range = 1:nsites
+  circuit = layer(nsites, θ⃗[range])
+  for n in 1:(nlayers - 1)
+    circuit = [circuit; layer(nsites, θ⃗[range .+ n * nsites])]
+  end
+  return circuit
+end
+
+Random.seed!(1234)
+
+θ⃗ᵗᵃʳᵍᵉᵗ = 2π * rand(nsites * nlayers)
+𝒰ᵗᵃʳᵍᵉᵗ = variational_circuit(nsites, nlayers, θ⃗ᵗᵃʳᵍᵉᵗ)
+
+s = siteinds("Qubit", nsites)
+Uᵗᵃʳᵍᵉᵗ = ops(𝒰ᵗᵃʳᵍᵉᵗ, s)
+
+ψ0 = MPS(s, "0")
+
+# Create the random target state
+ψᵗᵃʳᵍᵉᵗ = apply(Uᵗᵃʳᵍᵉᵗ, ψ0; cutoff=1e-8)
+
+#
+# The loss function, a function of the gate parameters
+# and implicitly depending on the target state:
+#
+# loss(θ⃗) = -|⟨θ⃗ᵗᵃʳᵍᵉᵗ|U(θ⃗)|0⟩|² = -|⟨θ⃗ᵗᵃʳᵍᵉᵗ|θ⃗⟩|²
+#
+function loss(θ⃗)
+  nsites = length(ψ0)
+  s = siteinds(ψ0)
+  𝒰θ⃗ = variational_circuit(nsites, nlayers, θ⃗)
+  Uθ⃗ = ops(𝒰θ⃗, s)
+  ψθ⃗ = apply(Uθ⃗, ψ0)
+  return -abs(inner(ψᵗᵃʳᵍᵉᵗ, ψθ⃗))^2
+end
+
+θ⃗₀ = randn!(copy(θ⃗ᵗᵃʳᵍᵉᵗ))
+
+@show loss(θ⃗₀), loss(θ⃗ᵗᵃʳᵍᵉᵗ)
+
+loss_∇loss(x) = (loss(x), convert(Vector, loss'(x)))
+algorithm = LBFGS(; gradtol=gradtol, verbosity=2)
+θ⃗ₒₚₜ, lossₒₚₜ, ∇lossₒₚₜ, numfg, normgradhistory = optimize(loss_∇loss, θ⃗₀, algorithm)
+
+@show loss(θ⃗ₒₚₜ), loss(θ⃗ᵗᵃʳᵍᵉᵗ)
+
+nothing

examples/autodiff/circuit_optimization/vqe.jl

@@ -0,0 +1,81 @@
+using ITensors, ITensorMPS
+using OptimKit
+using Random
+using Zygote
+
+nsites = 4 # Number of sites
+nlayers = 2 # Layers of gates in the ansatz
+gradtol = 1e-4 # Tolerance for stopping gradient descent
+
+# The Hamiltonian we are minimizing
+function ising_hamiltonian(nsites; h)
+  ℋ = OpSum()
+  for j in 1:(nsites - 1)
+    ℋ -= 1, "Z", j, "Z", j + 1
+  end
+  for j in 1:nsites
+    ℋ += h, "X", j
+  end
+  return ℋ
+end
+
+# A layer of the circuit we want to optimize
+function layer(nsites, θ⃗)
+  RY_layer = [("Ry", (n,), (θ=θ⃗[n],)) for n in 1:nsites]
+  CX_layer = [("CX", (n, n + 1)) for n in 1:2:(nsites - 1)]
+  return [RY_layer; CX_layer]
+end
+
+# The variational circuit we want to optimize
+function variational_circuit(nsites, nlayers, θ⃗)
+  range = 1:nsites
+  circuit = layer(nsites, θ⃗[range])
+  for n in 1:(nlayers - 1)
+    circuit = [circuit; layer(nsites, θ⃗[range .+ n * nsites])]
+  end
+  return circuit
+end
+
+s = siteinds("Qubit", nsites)
+
+h = 1.3
+ℋ = ising_hamiltonian(nsites; h=h)
+H = MPO(ℋ, s)
+ψ0 = MPS(s, "0")
+
+#
+# The loss function, a function of the gate parameters
+# and implicitly depending on the Hamiltonian and state:
+#
+# loss(θ⃗) = ⟨0|U(θ⃗)† H U(θ⃗)|0⟩ = ⟨θ⃗|H|θ⃗⟩
+#
+function loss(θ⃗)
+  nsites = length(ψ0)
+  s = siteinds(ψ0)
+  𝒰θ⃗ = variational_circuit(nsites, nlayers, θ⃗)
+  Uθ⃗ = ops(𝒰θ⃗, s)
+  ψθ⃗ = apply(Uθ⃗, ψ0; cutoff=1e-8)
+  return inner(ψθ⃗, H, ψθ⃗; cutoff=1e-8)
+end
+
+Random.seed!(1234)
+θ⃗₀ = 2π * rand(nsites * nlayers)
+
+@show loss(θ⃗₀)
+
+println("\nOptimize circuit with gradient optimization")
+
+loss_∇loss(x) = (loss(x), convert(Vector, loss'(x)))
+algorithm = LBFGS(; gradtol=1e-3, verbosity=2)
+θ⃗ₒₚₜ, lossₒₚₜ, ∇lossₒₚₜ, numfg, normgradhistory = optimize(loss_∇loss, θ⃗₀, algorithm)
+
+@show loss(θ⃗ₒₚₜ)
+
+println("\nRun DMRG as a comparison")
+
+e_dmrg, ψ_dmrg = dmrg(H, ψ0; nsweeps=5, maxdim=10)
+
+println("\nCompare variational circuit energy to DMRG energy")
+@show loss(θ⃗ₒₚₜ), e_dmrg
+
+nothing

examples/autodiff/mps_autodiff.jl

@@ -0,0 +1,49 @@
+using ITensors, ITensorMPS
+using OptimKit
+using Zygote
+
+function ising(n; J, h)
+  os = OpSum()
+  for j in 1:(n - 1)
+    os -= J, "Z", j, "Z", j + 1
+  end
+  for j in 1:n
+    os -= h, "X", j
+  end
+  return os
+end
+
+function loss(H, ψ)
+  n = length(ψ)
+  ψHψ = ITensor(1.0)
+  ψψ = ITensor(1.0)
+  for j in 1:n
+    ψHψ = ψHψ * dag(ψ[j]') * H[j] * ψ[j]
+    ψψ = ψψ * replaceinds(dag(ψ[j]'), s[j]' => s[j]) * ψ[j]
+  end
+  return ψHψ[] / ψψ[]
+end
+
+n = 10
+s = siteinds("S=1/2", n)
+J = 1.0
+h = 0.5
+
+# Loss function only works with `Vector{ITensor}`,
+# extract with `ITensors.data`.
+ψ0 = ITensors.data(random_mps(s; linkdims=10))
+H = ITensors.data(MPO(ising(n; J, h), s))
+
+loss(ψ) = loss(H, ψ)
+
+optimizer = LBFGS(; maxiter=25, verbosity=2)
+function loss_and_grad(x)
+  y, (∇,) = withgradient(loss, x)
+  return y, ∇
+end
+ψ, fs, gs, niter, normgradhistory = optimize(loss_and_grad, ψ0, optimizer)
+Edmrg, ψdmrg = dmrg(MPO(H), MPS(ψ0); nsweeps=10, cutoff=1e-8)
+
+@show loss(ψ0), norm(loss'(ψ0))
+@show loss(ψ), norm(loss'(ψ))
+@show loss(ITensors.data(ψdmrg)), norm(loss'(ITensors.data(ψdmrg)))