Skip to content

Cholesky of a view of a BandedMatrix forgets the structure #450

@DanielVandH

Description

@DanielVandH
julia> using BandedMatrices, LinearAlgebra

julia> L = brand(10, 10, 2, 0); B = BandedMatrix(Symmetric(L * L'));

julia> cholesky(Symmetric(B)) # good
Cholesky{Float64, BandedMatrix{Float64, Matrix{Float64}, Base.OneTo{Int64}}}
U factor:
10×10 UpperTriangular{Float64, BandedMatrix{Float64, Matrix{Float64}, Base.OneTo{Int64}}}:
 0.959058  0.698054  0.15382                                                           
          0.184258  0.432745  0.686227                                                 
                   0.311502  0.977993  0.642646                                        
                            0.148787  0.285363   0.155638                               
                                     0.0493197  0.259591  0.55198                      
                                               0.293538  0.124905  0.113931            
                                                        0.769331  0.620623  0.106919   
                                                                 0.462947  0.723893  0.75173
                                                                          0.714449  0.32878
                                                                                   0.215385

julia> cholesky(Symmetric(B[1:end-1, 1:end-1])) # good
Cholesky{Float64, BandedMatrix{Float64, Matrix{Float64}, Base.OneTo{Int64}}}
U factor:
9×9 UpperTriangular{Float64, BandedMatrix{Float64, Matrix{Float64}, Base.OneTo{Int64}}}:
 0.959058  0.698054  0.15382                                                  
          0.184258  0.432745  0.686227                                        
                   0.311502  0.977993  0.642646                                
                            0.148787  0.285363   0.155638                     
                                     0.0493197  0.259591  0.55198             
                                               0.293538  0.124905  0.113931   
                                                        0.769331  0.620623  0.106919
                                                                 0.462947  0.723893
                                                                          0.714449

julia> cholesky(Symmetric(@view B[1:end-1, 1:end-1])) # bad
Cholesky{Float64, Matrix{Float64}}
U factor:
9×9 UpperTriangular{Float64, Matrix{Float64}}:
 0.959058  0.698054  0.15382   0.0       0.0        0.0       0.0       0.0       0.0
          0.184258  0.432745  0.686227  0.0        0.0       0.0       0.0       0.0
                   0.311502  0.977993  0.642646   0.0       0.0       0.0       0.0
                            0.148787  0.285363   0.155638  0.0       0.0       0.0
                                     0.0493197  0.259591  0.55198   0.0       0.0
                                               0.293538  0.124905  0.113931  0.0
                                                        0.769331  0.620623  0.106919
                                                                 0.462947  0.723893
                                                                          0.714449

Presumably (?) it would be possible for the latter to also return something like a

Cholesky{Float64, SubArray{Float64, 2, BandedMatrix{Float64, Matrix{Float64}, Base.OneTo{Int64}}, Tuple{UnitRange{Int64}, UnitRange{Int64}}, false}}

which knows the structure. Not 100% sure how views propagate in all the code.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions