Implement ForwardDiff in master solvers and general DiffEq problems on QO types

Project.toml

@@ -39,10 +39,11 @@ WignerSymbols = "1, 2"
 julia = "1.3"
 
 [extras]
+FiniteDiff = "6a86dc24-6348-571c-b903-95158fe2bd41"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [targets]
-test = ["LinearAlgebra", "SparseArrays", "Random", "Test"]
+test = ["FiniteDiff", "LinearAlgebra", "SparseArrays", "Random", "Test"]

src/master.jl

@@ -14,6 +14,7 @@ function master_h(tspan, rho0::Operator, H::AbstractOperator, J;
     _check_const.(J)
     _check_const.(Jdagger)
     check_master(rho0, H, J, Jdagger, rates)
+    tspan, rho0 = _promote_time_and_state(rho0, H, J, tspan)
     tmp = copy(rho0)
     dmaster_(t, rho, drho) = dmaster_h!(drho, H, J, Jdagger, rates, rho, tmp)
     integrate_master(tspan, dmaster_, rho0, fout; kwargs...)
@@ -41,6 +42,7 @@ function master_nh(tspan, rho0::Operator, Hnh::AbstractOperator, J;
     _check_const.(J)
     _check_const.(Jdagger)
     check_master(rho0, Hnh, J, Jdagger, rates)
+    tspan, rho0 = _promote_time_and_state(rho0, Hnh, J, tspan)
     tmp = copy(rho0)
     dmaster_(t, rho, drho) = dmaster_nh!(drho, Hnh, Hnhdagger, J, Jdagger, rates, rho, tmp)
     integrate_master(tspan, dmaster_, rho0, fout; kwargs...)
@@ -86,6 +88,7 @@ function master(tspan, rho0::Operator, H::AbstractOperator, J;
     _check_const(H)
     _check_const.(J)
     _check_const.(Jdagger)
+    tspan, rho0 = _promote_time_and_state(rho0, H, J, tspan)
     isreducible = check_master(rho0, H, J, Jdagger, rates)
     if !isreducible
         tmp = copy(rho0)
@@ -124,6 +127,7 @@ function master(tspan, rho0::Operator, L::SuperOperator; fout=nothing, kwargs...
     b = GenericBasis(dim)
     rho_ = Ket(b,reshape(rho0.data, dim))
     L_ = Operator(b,b,L.data)
+    tspan, rho_ = _promote_time_and_state(rho_, L_, tspan)
     dmaster_(t,rho,drho) = dmaster_liouville!(drho,L_,rho)
 
     # Rewrite into density matrix when saving

src/schroedinger.jl

@@ -122,22 +122,3 @@ function check_schroedinger(psi::Bra, H)
     check_multiplicable(psi, H)
     check_samebases(H)
 end
-
-
-function _promote_time_and_state(u0, H::AbstractOperator, tspan)
-    Ts = eltype(H)
-    Tt = real(Ts)
-    p = Vector{Tt}(undef,0)
-    u0data_promote = DiffEqBase.promote_u0(u0.data, p, tspan[1])
-    tspan_promote = DiffEqBase.promote_tspan(u0data_promote, p, tspan, nothing, Dict{Symbol, Any}())
-    if u0data_promote !== u0.data
-        u0_promote = rebuild(u0, u0data_promote)
-        return tspan_promote, u0_promote
-    end
-    return tspan_promote, u0
-end
-_promote_time_and_state(u0, f, tspan) = _promote_time_and_state(u0, f(first(tspan), u0), tspan)
-
-rebuild(op::Operator, new_data) = Operator(op.basis_l, op.basis_r, new_data)
-rebuild(state::Ket, new_data) = Ket(state.basis, new_data)
-rebuild(state::Bra, new_data) = Bra(state.basis, new_data)

src/timeevolution_base.jl

@@ -121,3 +121,47 @@ macro skiptimechecks(ex)
 end
 
 Base.@pure pure_inference(fout,T) = Core.Compiler.return_type(fout, T)
+
+function _promote_time_and_state(u0, H::AbstractOperator, tspan)
+    Ts = eltype(H)
+    Tt = real(Ts)
+    p = Vector{Tt}(undef,0)
+    u0_promote = DiffEqBase.promote_u0(u0, p, tspan[1])
+    tspan_promote = DiffEqBase.promote_tspan(u0_promote.data, p, tspan, nothing, Dict{Symbol, Any}())
+    return tspan_promote, u0_promote
+end
+function _promote_time_and_state(u0, H::AbstractOperator, J, tspan)
+    Ts = DiffEqBase.promote_dual(eltype(H), DiffEqBase.anyeltypedual(J))
+    Tt = real(Ts)
+    p = Vector{Tt}(undef,0)
+    u0_promote = DiffEqBase.promote_u0(u0, p, tspan[1])
+    tspan_promote = DiffEqBase.promote_tspan(u0_promote.data, p, tspan, nothing, Dict{Symbol, Any}())
+    return tspan_promote, u0_promote
+end
+
+_promote_time_and_state(u0, f, tspan) = _promote_time_and_state(u0, f(first(tspan)..., u0), tspan)
+
+@inline function DiffEqBase.promote_u0(u0::Ket, p, t0)
+    u0data_promote = DiffEqBase.promote_u0(u0.data, p, t0)
+    if u0data_promote !== u0.data
+        u0_promote = Ket(u0.basis, u0data_promote)
+        return u0_promote
+    end
+    return u0
+end
+@inline function DiffEqBase.promote_u0(u0::Bra, p, t0)
+    u0data_promote = DiffEqBase.promote_u0(u0.data, p, t0)
+    if u0data_promote !== u0.data
+        u0_promote = Bra(u0.basis, u0data_promote)
+        return u0_promote
+    end
+    return u0
+end
+@inline function DiffEqBase.promote_u0(u0::Operator, p, t0)
+    u0data_promote = DiffEqBase.promote_u0(u0.data, p, t0)
+    if u0data_promote !== u0.data
+        u0_promote = Operator(u0.basis_l, u0.basis_r, u0data_promote)
+        return u0_promote
+    end
+    return u0
+end

test/test_ForwardDiff.jl

@@ -1,6 +1,7 @@
 using Test
 using OrdinaryDiffEq, QuantumOptics
 import ForwardDiff
+import FiniteDiff
 
 # for some caese ForwardDiff.jl returns NaN due to issue with DiffEq.jl. see https://github.com/SciML/DiffEqBase.jl/issues/861
 # Here we test;
@@ -12,6 +13,27 @@ import ForwardDiff
 # here we partially control the gradient error by limiting step size (dtmax)
 
 
+@testset "ForwardDiff on ODE Problems" begin
+
+# schroedinger equation
+b = SpinBasis(10//1)
+psi0 = spindown(b)
+H(p) = p[1]*sigmax(b) + p[2]*sigmam(b)
+f_schrod!(dpsi, psi, p, t) = timeevolution.dschroedinger!(dpsi, H(p), psi)
+function cost_schrod(p)
+    prob = ODEProblem(f_schrod!, psi0, (0.0, pi), p)
+    sol = solve(prob, DP5(); save_everystep=false)
+    return 1 - norm(sol[end])
+end
+
+p = [rand(), rand()]
+fordiff_schrod = ForwardDiff.gradient(cost_schrod, p)
+findiff_schrod = FiniteDiff.finite_difference_gradient(cost_schrod, p)
+
+@test isapprox(fordiff_schrod, findiff_schrod; atol=1e-2)
+
+end
+
 @testset "ForwardDiff with schroedinger" begin
 
 # system
@@ -73,3 +95,37 @@ Ftdop(1.0)
 @test ForwardDiff.derivative(Ftdop, 1.0) isa Any
 
 end # testset
+
+
+@testset "ForwardDiff with master" begin
+
+b = SpinBasis(1//2)
+psi0 = spindown(b)
+rho0 = dm(psi0)
+params = [rand(), rand()]
+
+for f in (:(timeevolution.master), :(timeevolution.master_h), :(timeevolution.master_nh))
+    # test to see if parameter propagates through Hamiltonian
+    H(p) = p[1]*sigmax(b) + p[2]*sigmam(b)  # Hamiltonian
+    function cost_H(p) #
+        tf, psif = eval(f)((0.0, pi), rho0, H(p), [sigmax(b)])
+        return 1 - norm(psif)
+    end
+
+    forwarddiff_H = ForwardDiff.gradient(cost_H, params)
+    finitediff_H = FiniteDiff.finite_difference_gradient(cost_H, params)
+    @test isapprox(forwarddiff_H, finitediff_H; atol=1e-2)
+
+    # test to see if parameter propagates through Jump operator
+    J(p) = p[1]*sigmax(b) + p[2]*sigmam(b)  # jump operator
+    function cost_J(p)
+        tf, psif = eval(f)((0.0, pi), rho0, sigmax(b), [J(p)])
+        return 1 - norm(psif)
+    end
+
+    forwarddiff_J = ForwardDiff.gradient(cost_J, params)
+    finitediff_J = FiniteDiff.finite_difference_gradient(cost_J, params)
+    @test isapprox(forwarddiff_J, finitediff_J; atol=1e-2)
+end
+
+end