This project has moved and is read-only. For the latest updates, please go here.

MSR Sho

Feb 8, 2011 at 2:11 PM

Hi.

After MSR Sho (http://research.microsoft.com/en-us/projects/sho/) has been exposes to the public, I think this library should focus on bringing added value that is beyond basic linear algebra, as it is covered efficiently by Sho. Do you aggree? If so, what added value should Meta Numerics strive to provide?

Alex.

Feb 13, 2011 at 11:56 PM

Hi Alex,

I'm happy to see SHO released, and I do agree that its linear algebra support is excellent. Its particular better-than-eveyone-else feature is slicing, which allows parts of one matrix to be manipulated as independent matrices.

If you are asking what Meta.Numerics better-than-everyon-else features are, I would say advanced functions and, to a slightly lesser extent, statistical analysis.

If you are asking wheter Meta.Numerics will abandon linear algebra functionality and just tell everyone to use SHO, the answer is no. Our linear algebra support will continue and grow. We already have support for tridiagonal matrix operations, which SHO does not, and are adding support for SVD in the next release. Our intent is not to match SHO feature-for-feature, but to add those features which are needed to support our users and to support other parts of the library.

If you are asking what entirely new areas Meta.Numerics intends to cover in the future, there are lots of ideas on the table: time series analysis, geographic statistics, ordinary and partial differential equations, constrained and combinatorial optimization, and financial modeling have all been suggested. Feel free to respond with your own suggestions.

Thanks for your interest!

Feb 14, 2011 at 9:53 AM

I think Meta.Numerics linear algebra should extend SHO (use SHO's matrix/vector types) and not replace it. It believe it allows better integration of both libraries and allows users to easily enjoy both worlds. For example, why not implement SVD on using SHO's matrix types? Is it a technical problem? Efficiency? Something else? (I don't know because didn't look at the SVD implementation in Meta.Numerics).