Checklist

• There are no similar issues or pull requests for this yet.

Problem

Using the same model, implementation of add_loops (addLoopLawConstraints in cobra toolbox) takes significantly longer (30-300 seconds versus <1 second) in cobrapy than its matlab counterpart. Timing each step I find that it is the function nullspace, and specifically inside this function the line: _, s, vh = np.linalg.svd(A) that is taking up the majority if the ocmputation time (98%+).

In the matlab implementation a custom (it seems?) function for the nullspace is utilized, and I wondered if it is possible to implement and add this to how cobrapy handles the nullspace approximation?
https://github.com/opencobra/cobratoolbox/blob/572cad82f50b93d7ddeab5d2b5e5e9fca9da4250/external/base/utilities/sparseNull.m#L4

Solution

Implement matlab's version of approximating the nullspace to achieve similar performance as in MATLAB

Alternatives

I attemtped using scipy's sparse svds function, but this was incompatible (produced no null space at all), as well as scipy's regular svd function (which in the few tests i did is about 2x as slow).

Anything else?

No response