GSoC with Flint - Week 415 Jun 2014
This week I've been working more on some of the older approaches to computing the HNF. First I finished off the modulo determinant method which works well for square matrices of full rank. One of the issues with this method however is that the determinant needs to be calculated first and can be quite large, in which case there may be little gained by calculating modulo it.
I've also implemented the algorithm described by Kannan and Bachem which takes the principal minors to Hermite form one at a time, as opposed to other methods I've implemented so far which work column by column. This also provides significant improvement on the classical method, and indeed is performing better than the modular approach in many cases.
For the next week I'll be trying to optimise the algorithms implemented so far as much as possible, especially the modular approach of Domich, Kannan and Trotter as this plays a key role in the algorithm of Pernet and Stein. One thing I would like to try is combining the two approaches I worked on this week and making a modulo determinant version of the Kannan-Bachem algorithm. There is also the possibility that factoring the determinant into some relatively prime parts and working modulo each of those, before recombining at the end with the Chinese remainder theorem, of course this involves factoring the determinant which could make the whole procedure less worthwhile. I also need to make the necessary modifications to ensure that all algorithms work in the non-square case as efficiently as possible.