sedumipy.getada¶
Port of getada.m: builds ADA = A*P(d)*A’ for LP(+Lorentz) problems directly via sparse linear algebra – no C kernels, pure NumPy/SciPy. Used both by optstep.m’s LP-optimality-projection path (dead in practice there, since sedumi.m only calls optstep.m when lponly, which forces K.q and K.s empty too) AND, for real, by this port’s own sedumi.py main loop whenever has_psd is false – i.e. every LP+SOCP problem with no PSD blocks (K.s empty), where the Lorentz-block loop below is very much live code.
DAt[“q”] (lorN x m) and the ADA it feeds into can be either sparse or dense, picked once per solve by sedumi.py’s is_dense hint (from getsymbada()’s structural ADA pattern density) and threaded through via DAt[“q”]’s own representation – this function just follows whatever form getDAtm() handed it. Sparse is required on large-lorN problems like DIMACS’s nql180/qssp180 (m ~ 1.3e5, lorN ~ 3.2e4 – a dense DAt_q alone is >30 GiB, and the dense ADA = DAt_q.T @ DAt_q this function used to build unconditionally would be m x m, far bigger still); dense is faster whenever ADA comes out dense/near-dense anyway (e.g. nb.mat, m=123), where sparse @ sparse pays real bookkeeping overhead a plain BLAS matmul doesn’t. Both branches build bug-for-bug identical values; ADA is always handed back as sp.csc_matrix regardless (numeric_cholesky needs that either way).
Functions
|
absd = getada(A,K,d,DAt): also returns the freshly built ADA matrix (unlike the .m file, which stores it into the global ADA_sedumi_ instead of returning it) -- callers here get it back as a normal return value: (ADA, absd). |