[petsc-users] HermitianTranspose version of MatCreateTranspose.

[Andrew Spott]

So, I’ve got a MatCreateHermitianTranspose function that has close to the same functionality as the MatCreatTranspose version.  So I’m getting ready to send a pull request.




A few questions:




What branch should I add my changes to?




Should I create tests for this?  Where should I put them?




On Monday, Feb 23, 2015 at 9:02 PM, Barry Smith <bsmith at mcs.anl.gov>, wrote:


   We've had a small amount of debate over the years on how to handle the Hermitian transpose and non-Hermitian transpose that never got fully resolved.


Approach 1) Each (complex) matrix has a full set of transpose and Hermitian transpose operations (MatTranspose(), MatHermitianTranspose(), MatMultTranspose()), MatMultHermitianTranspose(), MatSolveTranspose(), MatSolveHermitianTranspose(), MatMatMultTranspose(), MatMatMultHermitianTranspose(), MatTranposeMatMult(), MatHermitianTransposeMatMult().......)  plus there are two vector "inner" products; VecDot() and VecTDot(). 


Approach 2) Consider a (complex) vector (and hence the associated matrix operators on it) to live in the usual Hermitian inner product space or the non-Hermitian "inner product space". Then one only needs a single VecDot() and MatTranspose(), MatMultTranspose() ... that just "does the right thing" based on what space the user has declared the vectors/matrices to be in. 


Approach 2) seems nicer since it only requires 1/2 the functions :-) and so long as the two vector "spaces" never interact directly (for example what would be the meaning of the "inner" product of a vector in the usual Hermitian inner product space with a vector from the non-Hermitian "inner product space"?) certain seems simpler.  Approach 1) might be simpler for some people who like to always see exactly what they are doing. 


I personally wish I had started with Approach 2 (but I did not), but there could be some flaw with it I am not seeing.


  Barry








> On Feb 23, 2015, at 6:50 PM, Andrew Spott <ansp6066 at colorado.edu> wrote:

> 

> I’m definitely willing to submit it as a pull request.

> 

> Also, while I’m at it, I’m going to write a “duplicate” function for transpose and hermitian_transpose.  Just because this seems 1) easy ( MatHermitianTranspose can return a new copy, as well as MatTranspose), and 2) necessary to use these for EPS.

> 

> Also, is “transpose” a good enough MatType?  Or does a new one need to be written?

> 

> -Andrew

> 

> 

> 

> On Mon, Feb 23, 2015 at 3:12 PM, Jed Brown <jed at jedbrown.org> wrote:

> 

> <signature.asc>

>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150318/c57d2e04/attachment.html>