I have the following MATLAB snippet:

>> R = randn(3000,6000);   % build a random 3000 by 6000 matrix
>> tic; norm(R, 1); toc;
Elapsed time is 0.005586 seconds.
>> tic; norm(R, 2); toc;
Elapsed time is 3.019667 seconds.
>> tic; norm(R, inf); toc;
Elapsed time is 0.005393 seconds.
>> 

My question is, why would the L2 norm computation be that much slower than the L1 or L infinity norm? This is a random matrix for testing purposes, of course, but for the actual matrix in my work I can see a similar pattern in terms of elapsed time.

However, on Julia, the results are as follows

julia> @time norm(R, 1);
  0.007156 seconds (1 allocation: 16 bytes)

julia> @time norm(R, 2);
  0.009142 seconds (1 allocation: 16 bytes)

julia> @time norm(R, Inf);
  0.034633 seconds (1 allocation: 16 bytes)

This makes no sense at all. Any help is appreciated!

CodePudding user response：

I invite you to read the documentation for norm. It is a good idea to always read the documentation to a function and not make assumptions about what it does. In short, with a matrix input, norm computes the matrix norm:

1. norm(R,1) is the maximum absolute column sum of R.
2. norm(R,Inf) is the maximum absolute row sum of R.
3. norm(R,2) is approximately max(svd(R)).

The 1-norm and infinity-norm of the matrix are computed in a similar way, and are therefore expected to be similar in cost. Computing the sum over rows or columns, and the max of the result, is quite cheap.

The 2-norm of the matrix in contrast requires a singular value decomposition, which is significantly more expensive.

In Julia, norm computes the vector norm. To compute a matrix norm, use opnorm.

To compute the vector norm of rows or columns of a matrix in MATLAB, use vecnorm (since R2017b). To compute the norm of the vectorized matrix, use norm(R(:)).

PS: The real question is why is the infinity-norm in Julia so slow? It should be cheaper than the 1-norm and much cheaper than the 2-norm!