As you might know, SageMath is a software system for mathematical computation. Built on Python, it has extensive libraries for numerous areas of mathematics. One of these areas is Matroid Theory, as has been exhibited several times on this blog.

Google Summer of Code is a program where Google pays students to work on open-source software during the summer.

Once again, SageMath has been selected as a mentoring organization for the Google Summer of Code. We’ve had students work on aspects of the Matroid Theory functionality for the past four years. Maybe this year, YOU can join those illustrious ranks! Check out the call for proposals and ideas list. Read the instructions on both pages carefully. Applications open on March 12, so it’s a good idea to start talking to potential mentors and begin writing your proposal!

Guest post by Chao Xu

In the summer, I have extended the SAGE code base for matroids for Google Summer of Code. This post shows a few example of it’s new capabilities.

Connectivity

Let $M$ be a matroid with groundset $E$ and rank function $r$. A partition of the groundset $\{E_1,E_2\}$ is a $m$-separation if $|E_1|,|E_2|\geq m$ and $r(E_1)+r(E_2)-r(E)\leq m-1$. $M$ is called $k$-connected if there is no $m$-separation for any $m < k$. The Fano matroid is an example of $3$-connected matroid.

The Fano matroid is not $4$-connected. Using the certificate=True field, we can also output a certificate that verify its not-$4$-connectness. The certificate is a $m$-separation where $m < 4$. Since we know Fano matroid is $3$-connected, we know the output should be a $3$-separation.

We also have a method for deciding $k$-connectivity, and returning a certificate.

There are 3 algorithms for $3$-connectivity. One can pass it as a string to the algorithm field of is_3connected.

1. "bridges": The $3$-connectivity algorithm Bixby and Cunningham. [BC79]
2. "intersection": the matroid intersection based algorithm
3. "shifting": the shifting algorithm. [Raj87]

The default algorithm is the bridges based algorithm.

The following is an example to compare the running time of each approach.

The new bridges based algorithm is much faster than the previous algorithm in SAGE.

For $4$-connectivity, we tried to use the shifting approach, which has an running time of $O(n^{4.5}\sqrt{\log n})$, where $n$ is the size of the groundset. The intuitive idea is fixing some elements and tries to grow a separator. In theory, the shifting algorithm should be fast if the graph is not $4$-connected, as we can be lucky and find a separator quickly. In practice, it is still slower than the optimized matroid intersection based algorithm, which have a worst case $O(n^5)$ running time. There might be two reasons: the matroid intersection actually avoids the worst case running time in practice, and the shifting algorithm is not well optimized.

Matroid intersection and union

There is a new implementation of matroid intersection algorithm based on Cunningham’s paper [Cun86]. For people who are familiar with blocking flow algorithms for maximum flows, this is the matroid version. The running time is $O(\sqrt{p}rn)$, where $p$ is the size of the maximum common independent set, $r$ is the rank, and $n$ is the size of the groundset. Here is an example of taking matroid intersection between two randomly generated linear matroids.

Using matroid intersection, we have preliminary support for matroid union and matroid sum. Both construction takes a list of matroids.

The matroid sum operation takes disjoint union of the groundsets. Hence the new ground set will have the first coordinate indicating which matroid it comes from, and second coordinate indicate the element in the matroid.

Here is an example of matroid union of two copies of uniform matroid $U(1,5)$ and $U(2,5)$. The output is isomorphic to $U(4,5)$.

One of the application of matroid union is matroid partitioning, which partitions the groundset of the matroid to minimum number of independent sets. Here is an example that partitions the edges of a graph to minimum number of forests.

Acknowledgements

I would like to thank my mentors Stefan van Zwam and Michael Welsh for helping me with the project. I also like to thank Rudi Pendavingh, who have made various valuable suggestions and implemented many optimizations himself.

References

[BC79] R.E Bixby, W.H Cunningham. Matroids, graphs and 3-connectivity, J.A Bondy, U.S.R Murty (Eds.), Graph Theory and Related Topics, Academic Press, New York (1979), pp. 91-103.

[Raj87] Rajan, A. (1987). Algorithmic applications of connectivity and related topics in matroid theory. Northwestern university.

[Cun86] William H Cunningham. 1986. Improved bounds for matroid partition and intersection algorithms. SIAM J. Comput. 15, 4 (November 1986), 948-957.

I’m back with more on matroid computation in Sage. Previous installments are here, here, and here. As in the previous post, clicking the “Evaluate” buttons below will execute the code and return the answer. The various computations are again linked, so execute them in the right order.

Today we will look at various ways to ask (and answer) the question “Are these two matroids equal?” There are some subtle aspects to this, since we had to balance efficiency of the code and mathematical interpretation. The result is that we have various ways to compare matroids.

Isomorphism

Matroids $M$ and $N$ are isomorphic if there is a bijection $\phi: E(M) \to E(N)$ mapping bases to bases and nonbases to nonbases. In Sage, we check this as follows:

In other words, every matroid in Sage has a method .is_isomorphic() taking another matroid as input and returning True or False. Currently there is no way to extract the actual isomorphism (it is on our wish list), but if you guessed one, you can check its correctness:

We defined $\phi$ using a dictionary: for instance, phi['c'] equals 2. If you defined your map differently (e.g. as a function or a permutation), Sage will probably understand that too.

Equality

Matroids $M$ and $N$ are equal if $E(M) = E(N)$ and the identity map is an isomorphism. We can check for equality as follows:

==

The standard way to compare two objects in Sage is through the comparison operator ==. For instance,

When writing the matroid code, we chose to interpret the question M == N to mean “Is the internal representation of the matroid $M$ equal to the internal representation of $N$?” This is a very restricted view: the only time M == N will return True is if

• $M$ and $N$ have the same internal data structure (i.e. both are of type BasisMatroid or both are of type LinearMatroid or …)
• $M$ and $N$ are equal as matroids
• The internal representations are equivalent (this is at the moment only relevant for linear matroids).

Let’s consider a few examples:

So only matroids $M_1$, $M_2$, and $M_4$ pass the equality test. This is because they are all linear matroids over the field $\mathrm{GF}(2)$, have the same ground set, and the matrices are projectively equivalent: one can be turned into the other using only row operations and column scaling. Matrix $M_3$ is isomorphic to $M_1$ but not equal; matroid $M_5$ is represented over a different field; matroid $M_6$ is represented by a list of bases, and matroid $M_7$ is represented by a list of “circuit closures”.

This notion of equality has consequences for programming. In Python, a set is a data structure containing at most one copy of each element.

So $S$ actually contains $M_3, M_5, M_6, M_7$, and one copy of $M_1, M_2, M_4$.

This means, unfortunately, that you cannot rely on Python’s default filtering tools for sets if you want only matroidally equal objects, or only isomorphic objects. But those equality tests are way more expensive computationally, whereas in many applications the matroids in a set will be of the same type anyway.

Inequivalent representations

As mentioned above, each instance of the LinearMatroid class is considered a represented matroid. There are specialized methods for testing projective equivalence and projective isomorphism. Recall that matroids are projectively equivalent (in Sage’s terminology, field-equivalent) if the representation matrices are equal up to row operations and column scaling. So below, matroids $M_1$ and $M_3$ are field-equivalent; matroids $M_1$ and $M_4$ are field-isomorphic; and matroid $M_2$ has an inequivalent representation:

I probably caused a lot of confusion, so feel free to ask questions in the comments below. Also, the documentation for the functions discussed above gives more explanations and examples.

Google’s Summer of Code is an annual event in which Google will pay a bunch of students to spend the summer writing code for open-source projects.

As last year, Sage is one of the mentoring organizations, and as last year, we are hoping to have one of about 4 students work on Sage’s matroid capabilities. We are looking for an undergraduate or graduate student who

• Knows a bit of matroid theory (some possible topics require more knowledge than others)
• Has some programming experience, preferably in Python
• Wants to work on building exciting mathematics software for 3 months during the summer
• Would like the word “Google” on his or her CV
• Is interested in earning 5500 US dollars and a t-shirt

On the ideas page, a list of possible features to work on can be found. Students must produce a detailed proposal including a timeline with milestones. Early interaction with the mentors is important: we want to know you’re serious about this, and able to deliver. The place for this is the sage-gsoc Google group.

If you know an interested student, send her or him to this post. If you are an interested student, study the links above to figure out the requirements, timeline, etc. One important date: applications must be in on

MARCH 21

and we’d like to start working on the proposal with you well before then.