3 ms·
Not sure I 100% agree with (2) forming A^TA. In many real-world use cases A^TA+choleksy is going to be considerably faster than QR on A, and come with few nume
by owlbite 4y ago
Not sure I 100% agree with (2) forming A^TA.
In many real-world use cases A^TA+choleksy is going to be considerably faster than QR on A, and come with few numerical consequences. Even in the numerically challenging cases, pivoted LDL^T on the saddle point system is still going to be faster than the suggested solution of doing QR on A. (Essentially no optimization solver I've ever seen uses QR).