fenics team mailing list archive

[Hans Petter Langtangen <hpl@xxxxxxxxx>]

Tue, 04 May Marie Rognes wrote:
> Anders Logg wrote:
> >On Mon, May 03, 2010 at 11:08:13AM +0200, Kent Andre wrote:
> >>On Mon, 2010-05-03 at 09:36 +0100, Garth N. Wells wrote:
> >>>On 03/05/10 08:48, Kent Andre wrote:
> >>>>I am in favour of a(u,v), simply because it is the way I am used to
> >>>>read and write it. Also, very few users need to care about this due to
> >>>>the high-level syntax in UFL which implicitly takes care of the
> >>>>ordering.
> >>>>
> >>>They do need to worry about it on the DOLFIN side with the creation of a
> >>>Form:
> >>>
> >>>    Poissin::Bilinearform a(V0, V1);
> >>>
> >>>Garth
> >>>
> >>Yes, of course, sorry, in C++ this is important,
> >>
> >>Anyway, I think both alternatives have their advantages.
> >>But, as Godel put it, you must choose between consistency and
> >>conformity :)
> >
> >I've been thinking some more, and I'm getting more convinced that the
> >problem is really the ordering of rows and columns in matrices, which
> >should be (column, row) rather than (row, column) because a matrix is
> >really a set of columns that span the range of A, so A_i should be
> >column i and A_ij should be element j of that column.
> >
> >So then a(u, v) would be consistent with *that* notation but it would
> >be a pain for us to have an unorthodox matrix index notation.
> >
> >Which leaves us with the following options:
> >
> >1. Keep a(v, u) as we have now and not change anything.
> >
> >2. Change to a(u, v) and comment on the need for swapping the indices
> >when we define the matrix (tensor) in the text (book and manual).
> >
> >Alternative (2) would imply the following changes:
> >
> >(i) Redefine the UFC interface so that the test space is the last
> >space, not the first.
> >
> >(ii) Swapping of indices inside Assembler.cpp.
> >
> >Further comments?
> >
> 
> 
> Vote in favor of (1).

I also vote in favor of (1) for the source code and technical documentation.

However, I still repeat my arguments for a(u, v) in the FEniCS tutorial,
at least in the beginning. We may well introduce a(v, u) later in the tutorial
and explain the u,v-v,u problem, or postpone that to the introductory chapter
on the finite element method. 

Regardless of what is mostly used in the FEniCS book and technical
documentation, there will probably be a lot of application
contributions over the next years where people write a(u,v).
Therefore I think we should have a small section in the book explaining
the issue of a(u,v) vs a(v,u), so we have a reference for resolving
potential confusion.

Hans Petter