## More information on my Google Summer of Code project this year

May 26, 2010

So, as I noted here, I have been accepted into the Google Summer of Code program again this year. I mentioned that my project involved improving the integrator, but I didn’t say much other than that. So here I plan on saying a bit more. If you want more details, you can read my application on the SymPy wiki.

My goal is to improve the integrator in SymPy, in other words, the back end to the integrate() function. This is no easy task. Currently, SymPy has a pretty decnet integration engine. It is even able to solve some integrals that no other system is known to be able to solve (the second integral here). But, as I discovered often many times throughout my work on ODEs last year, the integrator can often leave something to be desired. There are two problems that I hope to address.

First, the integrator often fails on elementary integrals. This is because all of the integration in SymPy is based on a heuristic called the Risch-Norman algorithm. Symbolic integration has been completely solved in the form of the Risch algorithm, meaning that there exists an algorithm to determine if an elementary function has an elementary antiderivative or not, and to find it if it does. This algorithm, called the Risch algorithm, is extremely complicated, to the extent that no computer algebra system has ever completely implemented all the parts of it. My plan is to begin implementing the full algorithm in SymPy. I don’t expect to finish the whole thing — as I said no one ever has. Rather, I hope to make a good headway into what is known as the transcendental part. The Risch algorithm is broken up into four parts: rational part, the transcendental part, the algebraic part, and the mixed part.

The rational part is involves integrating rational functions (functions of the form $\frac{a_nx^n + a_{n-1}x^{n-1} + \cdots + a_2x^2 + a_1x + a_0}{b_nx^n + b_{n-1}x^{n-1} + \cdots + b_2x^2 + a_1x + a_0}$). The rational part is the easiest part in the sense that the algorithm is the simplest, and also that all rational function integrals are elementary (a term that I will define later). Rational function integration is already implemented in sympy in full, though I may give a brief outline of how it works in a later post.

The transcendental part is the part that I will be implementing this summer. My guide will be Symbolic Integration I: Transcendental Functions by Manuel Bronstein, which describes and proves the transcendental part of the algorithm in some 300+ pages. I will try to explain a little of how the algorithm works in some blog posts, but understand that it is very complex. Therefore, I will probably explain it without proving things. If you are interested in buying the book and learning the algorithm rigorously, the only prerequisites that I can tell are calculus (so you know what an integral and a derivative are), and a semester of abstract algebra (you need to know about rings, fields, ideals, homomorphisms, etc., as well as the various theorems relating them).

In the book, I am still in the part that develops the theory called differential algebra necessary to prove the integration algorithm correct. So to begin the GSoC program, I am working on learning the polys module in sympy. My method of doing this is to write doctests for all the functions in the module. It’s a daunting task, but it’s been probably the best way of learning how a computer module works that I have ever tried. You really have to understand all aspects of a function to write a doctest for it, the types of the parameters and return value, as well as what the algorithm is actually doing. It’s especially helpful that the code for the functions is right below the docstring for each function, so I can see how it really works on the inside, removing the mystery of the module. Furthermore, it will serve as a reference for me for the remainder of the summer, as well for anyone else who wants to learn the polys module, or just needs to debug it. I’ve also ran into several bugs and inefficiencies in the module that I have taken the liberty of fixing.

Well that’s it for this post. If you want to follow my progress on the doctests, my branch is http://github.com/asmeurer/sympy/tree/polydocs-polys9. Note that the branch will be very unstable until I finish at some point at the end of this week or the beginning of the next.