KrylovKit svdsolve not working

New to Julia
1

Hi! I’m having a little trouble understanding why svdsolve isn’t working- I can’t seem to find my error. I’m running

using SparseArrays
using LinearAlgebra
import LinearMaps
import KrylovKit

A = [1  0; 0 1]
Lambda,Phi,Gamma,info = KrylovKit.svdsolve(A, 1, krylovdim=3, tol=1E-3, maxiter=32);

and getting

ERROR: LoadError: ArgumentError: operator and its adjoint are not compatible
Stacktrace:
  [1] initialize(iter::KrylovKit.GKLIterator{Matrix{Int64}, Vector{Int64}, KrylovKit.ModifiedGramSchmidt2}; verbosity::Int64)
    @ KrylovKit ~/.julia/packages/KrylovKit/xccMN/src/factorizations/gkl.jl:195

Sorry for the basic question! My eventual hope is to be able to use svdsolve for sparse matrices with bigfloat numbers.

Thanks!

2

I think your call to svdsolve is wrong. I think you intended the 1 to mean howmany but instead you match this signature:
svdsolve(f, m::Int, [howmany = 1, which = :LR, T = Float64]; kwargs...)
Where this m denotes the number of rows of the operator. Since this is obviously 2 in your case, this likely explains the error you are seeing.

Try:
KrylovKit.svdsolve(A, 2, 1, krylovdim=3, tol=1E-3, maxiter=32)

3

That fixed it. Thanks for your answer!

I did a different thing and got an ArgumentError again, so I’m just trying to understand the situation/ figure out why I keep on making this error.

I ran

B = BigFloat(1)*sparse([1 0; 0 1])
Lambda,Phi,Gamma,info = KrylovKit.svdsolve(B, 2, 1,:SR, krylovdim=3, tol=1e-3, maxiter=32);

which gives the error

julia> ERROR: ArgumentError: operator and its adjoint are not compatible
Stacktrace:
 [1] initialize(iter::KrylovKit.GKLIterator{SparseMatrixCSC{…}, Vector{…}, KrylovKit.ModifiedGramSchmidt2}; verbosity::Int64)
   @ KrylovKit ~/.julia/packages/KrylovKit/xccMN/src/factorizations/gkl.jl:195
 [2] initialize
   @ ~/.julia/packages/KrylovKit/xccMN/src/factorizations/gkl.jl:186 [inlined]
 [3] svdsolve(A::SparseMatrixCSC{…}, x₀::Vector{…}, howmany::Int64, which::Symbol, alg::KrylovKit.GKL{…}; alg_rrule::KrylovKit.Arnoldi{…})
   @ KrylovKit ~/.julia/packages/KrylovKit/xccMN/src/eigsolve/svdsolve.jl:156
 [4] svdsolve(A::SparseMatrixCSC{…}, x₀::Vector{…}, howmany::Int64, which::Symbol, alg::KrylovKit.GKL{…})
4

If you use an n::Int in second position to set the number of rows, then KrylovKit uses rand(Float64, n) as starting point. However the application of the matrix then gives a vector with different element type. I suspect that this causes the issue. Try:

KrylovKit.svdsolve(B, big(rand(2)), 1,:SR, krylovdim=3, tol=1e-3, maxiter=32)
Or specify the eltype:
KrylovKit.svdsolve(B, 2, 1,:SR, krylovdim=3, tol=1e-3, maxiter=32, T=BigFloat)