sparse matrix entries are not displayed correctly

A sparse matrix variable has a very strange behavior.

A minimal example is given as follows: this is a 2 by 4 matrix, and I display them in both full and sparse ways for clarification

>> full(A)

ans =

0 0 0 198844

0 0 0 23672

>> A

A =

(2,4) 23672

(1,4) 198844

The problem is: if we print the first row, then it shows tht the first row contains zeros, which is not true.

>> A(1,:)

ans =

All zero sparse: 1×4

However, the second rows are printed correctly, and if we print both first&second rows and fouth column, it is also correct.

>> A(2,:)

ans =

(1,4) 23672

>> A(1:2,4)

ans =

(2,1) 23672

(1,1) 198844

remarks:

1. These commands are entered without doing anything in between.
2. A(1,4) is also considered as zero if I make any calculations, for example, A(1,4)*5 gives zero.

[Walter Roberson]

 A = sparse([          0           0           0      198844
            0           0           0       23672])

A = 
   (1,4)         198844
   (2,4)          23672

 A(1,:)

ans = 
   (1,4)         198844

 A(2,:)

ans = 
   (1,4)          23672

Could you save the matrix to a .mat file and attach the mat file for testing?

Also, please fill out which Release you are using.

[Hao Hu]

Thanks for your response. I have attached the .mat file containg the minimal example, and the release is "MATLAB Version: 9.11.0.1809720 (R2021b) Update 1"

After loading the mat file, we can see the bug by typing

>> load('example_sp_A.mat')

>> A

A =

(2,4) 23672

(1,4) 198844

>> A(1,4)

ans =

All zero sparse: 1×1

Answer 1

[Walter Roberson]

Notice that the matrix is displayed out of order -- the normal order is linear indexing, so (1,4) would have appeared before (2,4)

I suspect that the matrix was created by third-party software that (somehow) did not follow internal rules about constructing sparse matrices.

load('example_sp_A.mat')
A

A = 
   (2,4)          23672
   (1,4)         198844

B = sparse(full(A))

B = 
   (1,4)         198844
   (2,4)          23672

A(1,:)

ans = 
   All zero sparse: 1×4

B(1,:)

ans = 
   (1,4)         198844

mask = A ~= B

mask = 2×4 sparse logical array
   (1,4)      1
   (1,4)      1

full(mask)

ans = 2×4 logical array
   0   0   0   1
   0   0   0   0

nnz(mask)

ans = 2

So we can construct a "repaired" version of the matrix by taking it full, but in the meantime working with A gets us messed up results, such as a sparse array that has two elements both at the same location.

In summary, I think A was constructed corrupted, either by a third party routine writing a corrupted .mat or else by a malfunctioning mex routine.

[Matt J]

Using TIm Davis's spok shows that it is indeed corrupted:

>> spok(A)

Warning: 1 jumbled row indices, 0 explicit zeros

ans =

0

[Hao Hu]

It is indeed produced by a third party routine. More specifically, it is the a subroutine called "Gurobi_read()" from gurobi. It is used to read a mathematical programming file in .mps format. In this case, the problem 'app2-2.mps' is being read. The trick "convert to full and back to sparse" works very well. Hopefully this information is useful for anyone who encounter the same problem.

Thanks, Walter. This helps a lot.

[Matt J]

@Hao Hu You should Accept-click the answer to indicate that it resolved your question.

[Hao Hu]

Accepted. Thanks again. It is so good to know that sparse matrix can be corrupted.

Answer 2

[James Tursa]

I finally finished my sparse matrix integrity checker. You can find it in the FEX under the name SAREK (I think you will be able to figure out the inside joke):

https://www.mathworks.com/matlabcentral/fileexchange/158161-sarek-sparse-analyzer-real-et-komplex?s_tid=srchtitle

Here is what is reports for the posted matrix:

>> sarek(A)

SAREK -- Sparse Analyzer Real Et Komplex , by James Tursa

Compiled in version R2023a (with -R2018a option)
Running  in version R2023a

Matrix is double ...
Matrix is real ...
M = 2 >= 0 OK ...
N = 4 >= 0 OK ...
Nzmax = 2 >= 1 OK ...
Jc[0] == 0 OK ...
Jc[N] = 2 <= Nzmax = 2 OK ...
Jc array OK ...
ERROR: There are 1 Ir entries out of order or out of range
TO FIX: [M,N] = size(A); [I,J,V] = find(A); B = sparse(I,J,V,max(max(I),M),N);
All stored elements nonzero OK ...
    
There were ERRORS found in matrix!

ans =
1