ufl team mailing list archive

["Martin Sandve Alnæs" <martinal@xxxxxxxxx>]

2008/5/6 Anders Logg <logg@xxxxxxxxx>:
> On Tue, May 06, 2008 at 10:26:27AM +0200, Martin Sandve Alnæs wrote:
>> 2008/5/6 Anders Logg <logg@xxxxxxxxx>:
>> > On Mon, May 05, 2008 at 10:45:25PM +0200, Martin Sandve Alnæs wrote:
>> >> 2008/5/5 Anders Logg <logg@xxxxxxxxx>:
>> >> > On Mon, May 05, 2008 at 03:34:41PM +0200, Martin Sandve Alnæs wrote:
>> >> >> 2008/5/5 Anders Logg <logg@xxxxxxxxx>:
>> >> >> > On Mon, May 05, 2008 at 03:05:24PM +0200, Martin Sandve Alnæs wrote:
>> >> >> >> 2008/5/5 Anders Logg <logg@xxxxxxxxx>:
>> >> >> >> > On Mon, May 05, 2008 at 02:02:39PM +0200, Martin Sandve Alnæs wrote:
>> >> >> >> >> 2008/5/2 UFL <ufl@xxxxxxxxxx>:
>> >> >> >> >> > One or more new changesets pushed to the primary ufl repository.
>> >> >> >> >> > A short summary of the last three changesets is included below.
>> >> >> >> >> >
>> >> >> >> >> > changeset:   99:ab1f19b9eb5545605d2651bf95508956aa27ad0b
>> >> >> >> >> > tag:         tip
>> >> >> >> >> > user:        Anders Logg <logg@xxxxxxxxx>
>> >> >> >> >> > date:        Fri May 02 23:16:10 2008 +0200
>> >> >> >> >> > files:       ufl/basisfunctions.py ufl/finiteelement.py
>> >> >> >> >> > description:
>> >> >> >> >> > Implement recursive initialization of tensor elements.
>> >> >> >> >> > Unit tests pass, but see what you think of it.
>> >> >> >> >>
>> >> >> >> >> I don't agree with this, we want to have e.g. symmetric
>> >> >> >> >> second order tensor elements, which do not fit into this.
>> >> >> >> >
>> >> >> >> > In what way do they not fit?
>> >> >> >> >
>> >> >> >> > The flag is_symmetric() is still there.
>> >> >> >> >
>> >> >> >>
>> >> >> >> For a symmetric tensor in 3D, there should be only 6 subelements,
>> >> >> >> and this recursive initialization will create 9.
>> >> >> >
>> >> >> > Does that matter? Creating an UFL element is very lightweight so it
>> >> >> > shouldn't matter that we have a few extra elements. The important
>> >> >> > thing is for algorithms (and form compilers) to look at is_symmetric
>> >> >> > and use that to optimize.
>> >> >>
>> >> >> I'm not sure, but I feel that this should closely match the resulting
>> >> >> UFC element hierarchy that comes out at the other side.
>> >> >
>> >> > Then what should the hierarchy look like for a symmetric 2x2 tensor
>> >> > element?
>> >> >
>> >> > We could do
>> >> >
>> >> >  [[e, e], [None, e]]
>> >> >
>> >> > if we like.
>> >> >
>> >>
>> >> Basically, yes, but I'm not sure what you mean with the []'s?
>> >
>> > I mean that we have a mixed element which has two sub elements,
>> > each consisting of two (scalar) elements.
>> >
>> >> But if I interpret this correctly as Python list, I don't see the
>> >> advantage of storing this kind of list with None values.
>> >
>> > No, it's nested classes. The list notation was just shorthand (and
>> > confusing).
>> >
>> >> I'd rather store a single list of subelements, which can
>> >> be handled like in any MixedElement most of the time,
>> >> and store the subelement indices like:
>> >
>> > How would you store a simple Taylor-Hood element that way?
>>
>> Irrelevant. A Taylor-Hood element is not a TensorElement.
>
> No, it's not irrelevant if a TensorElement should be a subclass of
> MixedElement (which I think it should since it is a special case).
>
>> > I think the natural thing would be to have a MixedElement which has
>> > two sub elements. The first one of these (for u) will be a
>> > VectorElement (which is a special case of a MixedElement) and that
>> > element will have a list of two scalar sub elements. The second one
>> > (for p) will be a FiniteElement (a scalar).
>> >
>> > Similarly, one can take a vector-valued BDM-element for u and a scalar
>> > element for p and combine them into a MixedElement. In that case, each
>> > of the two sub elements will be a FiniteElement, but one will be
>> > vector-valued and one scalar.
>>
>> I completely agree with that for a general MixedElement.
>>
>> I just don't agree that a TensorElement should be
>> represented as a nested hierarchy of MixedElements,
>> since this will not be simple as intended for symmetric elements.
>>
>> Actually, I'm still not quite comfortable with VectorElement meaning
>> "vector of element" instead of "vector valued element", and I don't
>> think that matches what people would expect. Our different opinions
>> here may be related to this.
>
> It's very convenient if one can take any scalar element and use that
> to create a vector-valued element, and it's also convenient if one can
> have a single base class (MixedElement) which takes care of it all
> and have VectorElement and TensorElement as special cases.
>
> --
> Anders

I never said TensorElement shouldn't be a MixedElement.
I just want it to be a single MixedElement, instead of
containing a nested hierarchy of MixedElements.
The reason is that I don't think symmetric elements will fit
nicely into the hierarchic approach. I suspect using "None"
as you suggested above will lead to all sorts of special cases.

--
Martin