Strange behaviour computing l1 norm of rows of a matrix

I have a (1102,8232)-matrix B (from Data that I don't have control over).

The query

B(1102,:) 

yields

ans =  All zero sparse: 1×8232 

However, when I compute

beta = sum(abs(B),2);

or equivalently

beta = vecnorm(B,1,2) 

I get

beta(1102)
ans =   (1,1)       55

On the other hand, the command

norm(B(1102,:),1)
ans =   0

works as expected.

I'm really confused what's happening here. Any help is appreciated.

Answer 1

[Bruno Luong]

Who build this sparse matrix B?

If it's build from external source it can be not-valid. I believe there is a tool on FileExchange to validate the sparse matrix, the name escapes me now.

EDIT: this command crahes my MATLAB

B = B+0

indicating this matrix has serious building flaw.

[Christoph Neumann]

Yeah, it was transferred to Matlab via the function gurobi_read(), reading an .mps file that was previously written out by a python function, also using a Gurobi-interface.

EDIT: Thanks - that explains where the problem is and indicates why it only occurred when I applied the presolving step.

[Bruno Luong]

Here we go, this B is not valid MATLAB sparse, and detected by spok tool

>> run('C:\Users\Bruno\Documents\MATLAB\misc\spok\spok_install.m')
Compiling spok (64-bit)
Building with 'Microsoft Visual C++ 2017 (C)'.
MEX completed successfully.
>> load('C:\Users\Bruno\Documents\MATLAB\misc\spok\B.mat')
>> spok(B)
Warning: 59 jumbled row indices, 0 explicit zeros 

ans =

     0

>> 

[Bruno Luong]

Quick fix of error:

>> B=sparse(full(B)); % fix error, brute force
>> B(1102,:)

ans =

   (1,6)        1
   (1,199)      1
   (1,203)      1
   (1,247)      1
   (1,281)      1
   (1,332)      1
   (1,755)      1
   (1,756)      1
   (1,757)      1
   (1,758)      1
   (1,1423)     1
   (1,1635)     1
   (1,1636)     1
   (1,1637)     1
   (1,1638)     1
   (1,1639)     1
   (1,1640)     1
   (1,1888)     1
   (1,2281)     1
   (1,2282)     1
   (1,2283)     1
   (1,2284)     1
   (1,2285)     1
   (1,2286)     1
   (1,2576)     1
   (1,2577)     1
   (1,2578)     1
   (1,2579)     1
   (1,2580)     1
   (1,2891)     1
   (1,3317)     1
   (1,4328)     1
   (1,4329)     1
   (1,4330)     1
   (1,4331)     1
   (1,4332)     1
   (1,4333)     1
   (1,4334)     1
   (1,4335)     1
   (1,4336)     1
   (1,4337)     1
   (1,4338)     1
   (1,5898)     1
   (1,5899)     1
   (1,5900)     1
   (1,5901)     1
   (1,5902)     1
   (1,5903)     1
   (1,5980)     1
   (1,5981)     1
   (1,5982)     1
   (1,5983)     1
   (1,5984)     1
   (1,5985)     1
   (1,8232)    -1

[Christoph Neumann]

Cool, thanks!

[Bruno Luong]

Yeah but be aware the fixed matrix might not be correct. You better try to fix at gurobi_read or python, whichever step that is responsible for the error.

[James Tursa]

Note that spok has been removed from the FEX. You can use my replacement function sarek instead:

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