MATLAB sparse matrix: vec==0 vs. length(find(vec))==0

Remember that in Matlab sparse matrices are stored column-wise. Hence it is much more efficient to do column operations than to do row operations on a sparse matrix. When possible, always have two copies of the matrix A: A and AT=A’. This is important for time critical experiments.

Now let’s restrict to column operations. How do we find zero columns of a sparse matrix? There are two (or many more?) possible ways: (1) Use the A(:,i) == 0 test; (2) Use the length(find(A(:,i))) == 0 test. It turns out that the former method is way slower than the latter. Try it! Use the following codes:

m = 10000;
n = 10000;
density = 1e-4;
n_zero_col = 100;


% generate sparse matrix A
A = sprand(m, n, density);



% zero out some columns
idx = ceil(rand(n_zero_col,1)*n);
A(:,idx) = 0;



% two ways to find zero columns
% way 1: the `==0' construct
tic;
zero_cols = [];
for i = 1:n
if A(:,i)==0
zero_cols = [zero_cols i];
end
end
toc



% way 2: the `find' construct
tic;
zero_cols = [];
for i = 1:n
if length(find(A(:,i)))==0
zero_cols = [zero_cols i];
end
end
toc