NNLS

A non-negative least squares (NNLS) solver in Fortran. Modern examples exist for Python and Julia at least (I'm sure there are others) but the original Fortran code they're based on is from the 80s and fixed format hurts me to read so I wrote this one.

compile with make all as usual; requires BLAS and LAPACK. Otherwise all strictly fortran 2018. Only tested on linux and with gfortran.

set SHARED = 1 in the makefile to build as a shared library; it uses iso_c_binding and so should be straightforwardly callable from python using ctypes.

usage: solve(A, b, x, m, n, mode, res, maxiter, tol). when the subroutine returns x will be overwritten with the result, mode will be 0 on success and -1 on failure and res will contain the norm $`\left|Ax - b\right|`$.

set SHARED = 0 to compile standalone with main.f and then run ./test to check it works. this generates a set of 1000 matrices $A$ of random dimensions $`m \times n \text{ for } m, n \in [2, 30] `$ along with solution vectors $`x_{\text{ref}}`$, does $`A x_{\text{ref}} = b`$ and then runs solve as above with $A$ and $b$ to get a solution $x$.

note that there might be multiple solutions of the equations it generates; for each iteration it prints a.) res $`= \left|Ax - b\right|`$ and b.) $`\left|x_{\text{ref}} - x\right|`$ to stdout so you can check whether a.) the solution is reasonable and b.) whether it's the same solution or not.

any set of equations for which the maximum number of iterations is reached (currently set to $3n$as in the scipy implementation) or the norm$\left|Ax - b\right| > \epsilon = 1\times10^{-6}$ will be counted as a failure; the program will attempt to write an output file with a load of relevant details. at the end it will print the number of failures and the time taken.