141 composite vector space

[eparish1]

addressing #141

[fnrizzi]

Why is a composite vec space not inheriting from vector space?
I think it should

[eparish1]

@fnrizzi I broke up the constructor and made this inherit from VectorSpace. I think these both make sense, although the inheritance will probably go away if we switch to protocols.

[fnrizzi]

IMO we should not allow for a "compact_representation" to be true for these reasons:

• all this does is to get the basis from each individual vec space which one can do outside as easily. Havign this class does not add a layer of abstraction that makes it worth it IMO.

• if we provide a get_compact_basis we kind of break the vector space concept. So this breaks our usage internally because we cannot just rely on the interface of vector space get_basis.

I would suggest to make this strictly a global composite vector space.

[eparish1]

@fnrizzi We can make this a global composite vector space, but then need to have another class that is compact for a variety of reasons. I'm fine with either. Do you think the latter is better? The classes will basically be the same, but the compact one won't have a get_basis.

[fnrizzi]

@fnrizzi We can make this a global composite vector space, but then need to have another class that is compact for a variety of reasons. I'm fine with either. Do you think the latter is better? The classes will basically be the same, but the compact one won't have a get_basis.

can you explain to me later why we need a compact one? the compact one without a get_basis is not usable for any API that requires a vector space

[fnrizzi]

the compact one is just a list of basis that is created by calling get_basis on each

[eparish1]

@fnrizzi If we don't have a compact one, then we need to have completely different notions for what inputs to the algorithms are. Regardless of how we represent the basis, the collection of vector spaces forms the same concept.

What are your thoughts if we move the construction of the dense basis to the "get_basis", and only do it once the first time it is called. This way there is no notion of True or False for compact representation and we meet the vector space API. We can then still offer the get_compact_basis as an extra function.

[fnrizzi]

Can we talk later about this?

[jtencer]

I think we should think/talk about this variable ordering thing. The normal VectorSpace concept doesn't include anything about variable ordering and I'm not 100% sure we need/want to support it here (vs just assuming that they pass in the VectorSpaces in a consistent ordering.

[eparish1]

@jtencer @fnrizzi

I updated John's compromise with some of the stuff we talked about. Can you please see if you agree? Only thing left on my mind: should we add logic so that we only construct the dense basis once if we call get_basis()? (e.g., we construct it and store it as a member in the class and just return it if get_basis gets called multiple times?). What do you think?