Design Notes: Evaluating expr that results in `nan` or `oo`

Jim Crist edited this page Jul 7, 2014 · 2 revisions

These are notes on possible solutions to a problem I'm having in evaluating expressions in sympy.physics.mechanics. They are here for review by the community, and possible suggestions.

The Problem:

When evaluating symbolic expressions at points, sometimes the result is nan or oo unless the expression is fully simplified:

>>> expr = sin(x)/tan(x)
>>> expr.subs(x, 0)
>>> simplify(expr).subs(x, 0)

For small expressions this simplification is acceptable, but for large expressions (like the kind often generated in sympy.physics.mechanics this operation is too time consuming. I have come up with a few proposed solutions, which I will outline here:

Selective Simplification

Only apply simplify to expressions that would evaluate to nan. This is what I tried first. The algorithm is fairly simple, and can be done in two traverses of the expression tree:

  1. Replace all tan with sin/cos. This will allow us to only have to check for denominator = 0 conditions.

  2. For nodes that are fractions, check if the denominator evaluates to 0. If so, apply simplify to the fraction, then evaluate at the subs_dict.

This has been written up and tested. You can check out the code in this blog post (scroll down to the simplification section).

After some testing, it was found to be considerably faster than applying simplify, and then subs. However, for large expressions it consumed 4 GB of my RAM and ran for 6 hours before I had to kill it. This is due to the size of the expressions that need to be simplified. One possible solution is to run expand on the expression beforehand to reduce the size of each fraction, but this is also time consuming.

CSE subs

I haven't tried this, but it's a thought. Perform cse on the expression, then evaluate each sub-expression. If it evaluates to nan simplify it, then repeat. I have doubts that this will work though, as the parts that could be simplified out will probably already be numbers due to the previous subs operations. By this I mean:

>>> expr = cos(x + y)*tan(x + y)
>>> cse(expr)
([(x0, x(t) + y(t))], [cos(x0)*tan(x0)])
>>> sub_dict = {x: 0, y: pi/2}

In this case, x0 would evaluate fine, but then cos(pi/2)*tan(pi/2) = nan. In this case applying simplify to the final expression before subbing in x0 would work, but my gut feeling is that there will be expressions where this won't work out. I still plan on trying this.

Use limit to evaluate the troublesome expressions

Similar to the selective simplification, but instead take the limit of the troubling sub-expression. This would probably work, but there are some issues:

  • Multivariable sub_dict:

For most cases, the sub_dict used contains many variables. These could each be evaluated individually using normal subs, and if it evaluates to nan then take the limit for that variable. For example (by hand):

>>> expr = cos(x + y)*tan(x + y)
>>> expr1 = expr.subs(x, 0)
>>> expr1
>>> expr2 = expr.subs(y, pi/2)
>>> expr2
>>> expr3 = limit(expr2, y, pi/2)
>>> expr3

This would change the algorithmic complexity, as it would require a traversal for each key in the sub_dict (where before it was done in 1 traversal by msubs). However, I think this may be the best option for finding a correct solution everytime.

  • limit doesn't work with dynamicsymbols (sympy type AppliedUndef)

In mechanics the sub_dict is almost always composed of {dynamicsymbol: value}. This is less of an issue. Perhaps limit could be modified to work with AppliedUndef, or a Dummy pre-subbed in for the dynamicsymbol before taking the limit. This would add another traversal though, increasing runtime.

Calculate the value by perturbation

Similar to above (check if nan, then apply select rule to sub-expression). Instead of applying limit, perturb from the sub_dict value by epsilon (some small value), then check if expr.subs(sub_dict_plus_epsilon) == expr.subs(sub_dict_minus_epsilon) within some specified tolerance.

This is probably a bad way to do it, because (if I understand correctly) mathematically we'd need to check all possible paths to the setpoint. For example, suppose we have some expression f(x, y, z), with a setpoint setpoint=[a, b, c] that we'd like to evaluate it at. If we have epsilon=1e-10, then we'd need to check all possible combinations of [a +/- epsilon, b +/- epsilon, c +/- epsilon]. Also, this is more numerical, and may not be proper for sympy.

The benefit is that for we wouldn't have to traverse the expression for each key in the sub_dict, as this could be evaluated for all keys at the same time.

Note that I'm not sure if this is numerically/mathematically correct. In my mind it makes sense, but my preliminary research hasn't turned up any algorithms for calculating limits numerically.

Clone this wiki locally
You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.
Press h to open a hovercard with more details.