negative Grassmann inner product #53

Activity

1. lkdvos

   added

   bugSomething isn't workingSomething isn't working

   on Oct 16, 2023
2. DaanMaertens commented on Dec 6, 2023

   

   DaanMaertens

   on Dec 6, 2023

   Collaborator

   More actions

   So I've been trying to fix this bug by implementing what we discussed together, namely:

   function inner(g1::PrecGrad, g2::PrecGrad, rho=one(g1.rho))
      
       if g1 === g2
           max(real(unsafe_inner(g1,g2,rho)),eps())
       else   
           unsafe_inner(g1,g2,rho)
       end
   end
   function unsafe_inner(g1::PrecGrad, g2::PrecGrad, rho=one(g1.rho))
       Grassmann.base(g1) == Grassmann.base(g2) || throw(ArgumentError("incompatible base"))
       if g1.rho == rho
           Grassmann.dot(g1.g.Z, g2.Pg.Z)
       elseif g2.rho == rho
           Grassmann.dot(g1.Pg.Z, g2.g.Z)
       else
           Grassmann.dot(g1.Pg.Z, g2.Pg.Z * rho)
       end
   
   end

   This fixes the bug for the sqrt but introduces another later down the line

   ERROR: LoadError: linesearch was not given a descent direction!
   Stacktrace:
    [1] error(s::String)
      @ Base ./error.jl:35
    [2] (::OptimKit.HagerZhangLineSearch{Rational{Int64}})(fg::Function, x₀::MPSKit.GrassmannMPS.ManifoldPoint{InfiniteMPS{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, MPSKit.MPOHamInfEnv{MPOHamiltonian{TensorKit.ComplexSpace, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 2, Matrix{ComplexF64}}, ComplexF64}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, InfiniteMPS{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, KrylovKit.GMRES{KrylovKit.ModifiedGramSchmidt2, Float64}}, PeriodicArray{TensorKitManifolds.Grassmann.GrassmannTangent{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{Float64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, 1}, Vector{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}}, η₀::Vector{MPSKit.GrassmannMPS.PrecGrad{TensorKitManifolds.Grassmann.GrassmannTangent{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{Float64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}}, fg₀::Tuple{Float64, Vector{MPSKit.GrassmannMPS.PrecGrad{TensorKitManifolds.Grassmann.GrassmannTangent{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{Float64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}}}; retract::Function, inner::typeof(MPSKit.GrassmannMPS.inner), initialguess::Float64, acceptfirst::Bool)
      @ OptimKit ~/.julia/packages/OptimKit/xpmbV/src/linesearches.jl:255
    [3] optimize(fg::typeof(MPSKit.GrassmannMPS.fg), x::MPSKit.GrassmannMPS.ManifoldPoint{InfiniteMPS{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, MPSKit.MPOHamInfEnv{MPOHamiltonian{TensorKit.ComplexSpace, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 2, Matrix{ComplexF64}}, ComplexF64}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, InfiniteMPS{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, KrylovKit.GMRES{KrylovKit.ModifiedGramSchmidt2, Float64}}, PeriodicArray{TensorKitManifolds.Grassmann.GrassmannTangent{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{Float64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, 1}, Vector{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}}, alg::OptimKit.ConjugateGradient{OptimKit.HagerZhang{Rational{Int64}}, Float64, OptimKit.HagerZhangLineSearch{Rational{Int64}}}; precondition::typeof(OptimKit._precondition), finalize!::typeof(OptimKit._finalize!), retract::Function, inner::typeof(MPSKit.GrassmannMPS.inner), transport!::typeof(MPSKit.GrassmannMPS.transport!), scale!::typeof(MPSKit.GrassmannMPS.scale!), add!::typeof(MPSKit.GrassmannMPS.add!), isometrictransport::Bool)
      @ OptimKit ~/.julia/packages/OptimKit/xpmbV/src/cg.jl:68
    [4] find_groundstate(ψ::InfiniteMPS{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, H::MPOHamiltonian{TensorKit.ComplexSpace, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 2, Matrix{ComplexF64}}, ComplexF64}, alg::GradientGrassmann, envs::MPSKit.MPOHamInfEnv{MPOHamiltonian{TensorKit.ComplexSpace, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 2, Matrix{ComplexF64}}, ComplexF64}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, InfiniteMPS{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, KrylovKit.GMRES{KrylovKit.ModifiedGramSchmidt2, Float64}})
      @ MPSKit ~/.julia/packages/MPSKit/atykv/src/algorithms/groundstate/gradient_grassmann.jl:54
    [5] find_groundstate
      @ ~/.julia/packages/MPSKit/atykv/src/algorithms/unionalg.jl:25 [inlined]
    [6] #find_groundstate#380
      @ ~/.julia/packages/MPSKit/atykv/src/algorithms/groundstate/find_groundstate.jl:40 [inlined]
    [7] find_groundstate
      @ ~/.julia/packages/MPSKit/atykv/src/algorithms/groundstate/find_groundstate.jl:19 [inlined]
    [8] find_groundstate(ψ::InfiniteMPS{TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 1, Matrix{ComplexF64}}, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 1, 1, Matrix{ComplexF64}}}, H::MPOHamiltonian{TensorKit.ComplexSpace, TensorKit.TrivialTensorMap{TensorKit.ComplexSpace, 2, 2, Matrix{ComplexF64}}, ComplexF64})
      @ MPSKit ~/.julia/packages/MPSKit/atykv/src/algorithms/groundstate/find_groundstate.jl:19
    [9] top-level scope
      @ ~/Desktop/GitHub/MyGradientFix.jl/test/runtests.jl:12
   in expression starting at /Users/daan/Desktop/GitHub/MyGradientFix.jl/test/runtests.jl:12

   I made a small package called MyGradientFix.jl that overloads the inner function in GrasmannMPS, hence this stacktrace. The line in question inside OptimKit that throws the error message is

   df₀ = inner(x₀, g₀, η₀)
       if df₀ >= zero(df₀)
           error("linesearch was not given a descent direction!")
       end

   I do not understand OptimKit well enough to know why this check is required and why it fails.
3. maartenvd commented on Dec 7, 2023

   

   maartenvd

   on Dec 7, 2023

   Collaborator

   More actions

   optimkit needs to be given a gradient and a descent direction (in gradient descent you would make the descent direction be - gradient). A descent direction has to satsify that inner(x,gradient,direction) is smaller than zero, which is what optimkit tests. Honestly I wouldn't mind if optimkit turns it into a warning ,and just flips the sign.

   Gradientgrassman does something wrong in the case that C is numerically degenerate. A real fix would probably entail going through the derivation again, and making sure that everything is defined when C is no longer invertible.
4. lkdvos

   linked a pull request that will close this issueFix indefinite GradientGrassmann preconditioner on nearly rank-deficient bonds #532on Oct 1, 2026
5. lkdvos

   closed this as completedin #532on Oct 2, 2026