nncg¶

Non-negative conjugate gradients: the strictly convex bound-constrained quadratic min_{x>=0} 1/2 x'Ax - b'x (and its equality-augmented variant Bx = c) solved by matrix-free CG inside a primal-dual active-set loop with an unconditional finite-termination guarantee.

Reference implementation of the paper Non-Negative Conjugate Gradients (Schmelzer & Stoll) — see the paper repository.

Install¶

pip install nncg

API¶

The public API is solve_nnqp, solve_nnqp_eq, kkt_violation, and the Result dataclass. The matrix-free CG/PCG inner solvers live in nncg.krylov — the package's core contribution, driven by the solvers rather than imported directly. The planted-problem generators used by the numerical study live in the repository's tests/problems.py, outside the installed package. See the README for a quickstart.