Allow more elementwise simplifications for symbolic matrices

[kcrisman]

As with several questions at ask.sagemath. Mike Hansen's answer at the first one seems like a good start:

age: m = matrix([[sin(x), cos(x)], [sin(x), cos(x)]]); m
[sin(x) cos(x)]
[sin(x) cos(x)]
sage: o = m*m.transpose(); o
[sin(x)^2 + cos(x)^2 sin(x)^2 + cos(x)^2]
[sin(x)^2 + cos(x)^2 sin(x)^2 + cos(x)^2]
sage: o.apply_map(lambda x: x.trig_reduce())
[1 1]
[1 1]

but it seems reasonable for matrices with symbolic elements to have some of these methods (also vectors, I suppose) without having to use any special terminology.

Open to suggestions on how that might be accomplished without creating a myriad of special methods, but by somehow piggybacking on symbolic.expression.Expression methods done elementwise...

Putting this under symbolics because it isn't really linear algebra, but that doesn't seem right either.

CC: @jasongrout @rbeezer

Component: symbolics

Keywords: matrix symbolic simplify expand

Author: Joris Vankerschaver

Reviewer: Karl-Dieter Crisman

Issue created by migration from https://trac.sagemath.org/ticket/10552

[jvkersch]

comment:1

In the code written for #10132 (metric surfaces in 3D), we compute a lot of symbolic matrices with complicated entries which eventually need to be simplified. The current patch adds a method simplify_full to the symbolic matrix class, which simply calls the same method on each of the matrix elements.

[kcrisman]

Author: Joris Vankerschaver

[kcrisman]

comment:2

Thanks for the patch! I am not in a position to review it right now, but it seems fine at a glance.

Two comments:

• The ticket actually asks for access to all simplification methods, if possible. Possibly also for vectors. It's okay if you are aiming a little lower, but then we'll want to open another ticket for the more general thing. (Or that new ticket could be connected to this one so that people fixing this one knew about it. Or whatever.)
• The commit message is a little long for that first line; we want it to be about 80 characters or less so that it looks nice in a terminal. You can add additional lines that are as long as you like, of course.

[jvkersch]

comment:3

Ah, sorry to overlook the other simplification methods! I'll get onto it. Also, the code for vectors is now #11335, and I'll update that patch in tandem with this one.

One thing that I noticed in the symbolic matrix class is that (for instance) the simplify_trig method duplicates the code from the elementwise trig simplification. Presumably we want the matrix method to call the simplification method on the elements? This doesn't result too much of a performance loss (47.4ms for elementwise calls, 32.7ms for the duplicate code), so I'll just go ahead and rewrite all of the simplification methods. However, if there's a good reason for doing things the way they are, please let me know and I'll revert.

[kcrisman]

comment:4

Huh, I didn't even notice that these already existed - I should have, and probably even helped review it... sigh.

I think the difference here is that we're using Maxima to simplify the whole matrix, instead of doing it elementwise. Maybe that would be better, since Maxima might be faster (especially once we are using the library interface). So the real question is whether there are any simplifications that are better done elementwise than all at once by Maxima.

In fact, my guess is that Maxima just does them elementwise in any case, in which case they should all look like they already do - it would be faster, as you point out.

So this ticket really is asking to increase the number of methods available. simplify_full is a notable missing method, but notice that while simplify_trig is available, trig_reduce and reduce_trig are not. Though I guess those are technically different, looking at the underlying code...

[jvkersch]

comment:5

Replying to @kcrisman:

I think the difference here is that we're using Maxima to simplify the whole matrix, instead of doing it elementwise. Maybe that would be better, since Maxima might be faster (especially once we are using the library interface). So the real question is whether there are any simplifications that are better done elementwise than all at once by Maxima.

Good point. I did some comparisons for various matrix sizes of simplify_trig elementwise versus by passing the whole matrix to Maxima, and it seems passing the matrix directly to maxima is about 50% faster. For large-ish matrices (say, 10x10) this is already considerable (1 sec. vs 450 ms. for a random matrix involving sines and cosines). I will therefore use the matrix version wherever available.

[jvkersch]

comment:6

Robert Bradshaw proposed a nice way at #11381 of implementing elementwise methods by using the corresponding functionality in Expression but I don't think this trick will work for matrices, as they are cdef classes. So I went ahead and just implemented the elementwise methods piece by piece. It's not exactly rocket science, but we can always replace these elementwise methods by something more clever later on...