Not really blogging

A. Makelov

Reflections on “Orphans of the sky”

While the book is still fresh in my mind (it’s about 1 hour, or 60 minutes, that is, 3600 seconds, behind me). You know a science fiction book is good (to you) when it constructs curious ideas and situations you haven’t ever imagined before (which are of course made possible by some kind of, well, technology; otherwise it wouldn’t be much of a SF work; or would it?). Another way you know a science fiction book was good to you is when you read it, and then (of course) go to wikipedia, see when the thing was written, and be like “What? I thought it was written in the 70s or something…”.

“Orphans of the sky” by Heinlein was good to me in both respects. If I had to summarize the insight I gained from it in a sentence, it would roughly say this. The concepts of ‘humanity’, ‘human nature’ and ‘common sense’ are highly dependent on, and extremely, short-time-scale flexible with respect to, the knowledge passed from parents to children.

And here’s a quote that is both representative of this idea and a sort of motivation for the study of general topological spaces as opposed to metric spaces (What? What did I just say?…):

Metrical time caused him as much mental confusion as astronomical distances, but no emotional upset The trouble was again the lack of the concept in the Ship. The Crew had the notion of topological time; they understood “now,” “before,” “after,” “has been,” “will be,” even such notions as long time and short time, but the notion of measured time had dropped out of the culture. The lowest of earthbound cultures has some idea of measured time, even if limited to days and seasons, but every earthly concept of measured time originates in astronomical phenomena; the Crew had been insulated from all astronomical phenomena for uncounted generations.


Yossarian lives!

“They’re trying to kill me,” Yossarian told him calmly.
“No one’s trying to kill you,” Clevinger cried.
“Then why are they shooting at me?” Yossarian asked.
“They’re shooting at everyone,” Clevinger answered. “They’re trying to kill everyone.”
“And what difference does that make?”

Joseph Heller, “Catch 22”




It’s not a metaphor for a search engine. Get it? It’s a thing that finds metaphors for you. And it’s called “Yossarian Lives!”. Now how cool is that. I saw this sometime this summer, and recently the alpha’s image search has become operational. I tried it briefly and I have to admit the images it returned were pretty crazy and I could see a meaningful connection to my query in only a couple of them – but still, it was an interesting analogy. I was yossarianlives!ing for “moon” and I got devils. I then realized the shape of their horns resembled that of the moon crescent and was “wow”. Of course, there might have been many other connections I missed, due to the Stephen Fry problem (oh man I love these guys, they like everything I like).

Is the Stephen Fry problem just a convenient excuse? Is the final version going to be much better than the alpha? Are the images returned plain random? The future will show. However, the idea itself is amazing. “Outsourcing our minds”, bla bla bla. Shut up. Nothing (or at least, nothing yet) can outsource your mind, it can only inspire you to think deeper. Yossarian Lives!, if it lives up to its promise, will be a free, automated, non-stop service for blowing people’s minds. Remember the good old hanging around and not thinking about anything, just staring at the emptiness, when suddenly an amazing idea flashes through you mind, and you’re like “OMG THIS IS SO EPIC HOW COULD I NOT SEE THE CONNECTION BEFORE”? No more waiting for it to randomly happen – you just go to Yossarian Lives!, strike a couple of keys and be like “Wow. Wow. Wow. Wow….” for hours. An intellectual bomb. A mind-bender. In terms of a simple metaphor (haha), if the thing works as promised, we’ll have

\frac{\text{Yossarian Lives!}}{\text{hanging around waiting for a flash}}=\frac{\text{taking the roller coaster}}{\text{walking to grandma's house}} \ \ \ (1)

Is (1) good? Is it bad? Is it unnatural? Well, it’s tempting, and it could bring great change to the way people think about the world. And I think this is always good. If you don’t like it you just don’t use it.

Some other arguments the creators bring up (as if (1) is not enough) can be found in this very interesting essay. A great one is the following: search engines nowadays are, by nature, predicting their users and pointing them either to knowledge the majority of other people found useful (like, autocompletion), or to knowledge that is similar to what they searched for before. This may have many obvious advantages, but they come at a price: your knowledge horizon becomes conformist and hard to change. Search engines are trying to kill everyone. So they’re trying to kill you! It’s a pretty unfortunate Catch-22, isn’t it? On the other hand, the more subjective nature of the experience of understanding a metaphor has the potential to turn all this around (and confuse entire nations, I’m guessing, but there’s no other way).

Another reason why I find all this amazing is that this project seems to lie at the intersection of… everything. You notice this post is in almost all categories on my blog, as well as in Meta. Yossarian Lives!’s very heart is an objective process that seeks highly subjective results. The idea relates mathematics with art, determinism and rigorous theoretical ideas with the mind’s inner, sometimes seemingly arbitrary, associations and feelings. All fields of knowledge are basically clustered around these two poles, which often leads to people entirely dedicating themselves to one and forgetting the other, and consequently to lack of communication, understanding, and, I’m pretty sure, many interesting ideas. Well, I believe the poles are much closer than they appear to be to most people, and this project has a great potential to make me more right (:

So, what are you waiting for??? Go ahead and try it!

За математическите гимназии

Преместено от на 03.11.2012

Както казахме, новият закон за училищното образование е това, което ни подтикна да започнем да се занимаваме с

Проектозаконът привлече вниманието ни още през март и може да бъде намерен ето тук. Частта, която има значение, е в Чл. 36 (2). Забележете, че възможността “V-XII” клас отсъства. Според закона, единствените училища, които могат да обучават точно този интервал от класове, са спортните (виж Чл.37 (3)).

Проблемът с това е, че математическите гимназии в България традиционно предлагат прием след четвърти клас за най-способните деца. Преимуществата на тази система (спрямо прием след седми) са, според нас, големи и лесно видими; за това може да се говори още много в някоя от следващите ни публикации. За нас, като възпитаници на ПМГ-Бургас, е ясно че годините между четвърти и осми клас са незаменими в развитието на основите на математичеко мислене. След много проведени разговори с учители установихме, че ако този закон влезе в сила, математическите гимназии ще трябва или да се откажат от приема след 4-ти клас (вредно), или да въведат прием след първи (трудно и ненужно, а в някои случаи и физически невъзможно), или да сменят името и статута си (пак вредно). Според нас, никой от тези варианти не е приемлив – системата работи добре от десетки години и тези промени със сигурност няма да я подобрят.

Още през февруари/март бяха предприети действия от страна на родители и учители (виж например тук); събра се подписка (ето тази например; на нея има доста полезна информация за това как се развиваха нещата до юни) и ситуацията се успокои. През юли отново стана дума (например тук) и се оказа, че нищо не се е променило. Събрахме много съмишленици от различни общности и организирано изпращахме писма до народните представители, за да изразим несъгласието си; уви, в отговор не получихме почти нищо. Това, което знаем сега, е, че законът е бил приет на първо четене и през септември трябва да се гласува окончателно.

Google Summer of Code 2012: Week 13

Hi all, here’s a brief summary of my 13th (and last) week of GSoC.

  • I continued my work on centralizers, improving normal closure, derived in lower central series, etc. My most recent pull request containing these additions just got merged and can be found here. This week I spent a lot of time on writing better tests and developing some new test practices. The group-theoretical algorithms in the combinatorics module are getting more and more complicated, so better, cleverer and more thorough tests are needed. I came up with the following model for verification:
    – since the results of the tests are very hard to compute by hand, some helper functions are needed that find the wanted object in a brute-force manner using only definitions. For example, we often look for a subgroup with certain properties. The most naive and robust approach to this is to:
    – list all group elements, go over the list and check each element for the given property.
    – Then, make a list of all the “good” elements and compare it (as a set) with the list of all elements of the group the function being tested returns.
    Hence, a new file was created, sympy/combinatorics/, that will host such functions. (Needless to say, they are exponential in complexity, and for example going over all the elements of SymmetricGroup(n) becomes infeasible for n larger than 10.)
    – The presence of functions being used to test other functions gets us in a bit of a Quis custodiet ipsos custodes? situation, but this is not fatal: the functions in are extremely straightforward compared to the functions in that they test, and it’s really obvious what they’re doing, so it’ll take less tests to verify them.
    – In the tests for the new functions from, I introduced some comments to indicate what (and why) I’m testing. Another practice that seems to be good is to verify the algorithms for small groups (degrees 1, 2, 3) since there are a lot of corner cases there that seem to break them.
  • I started work on improving the disjoint cycle notation, namely excluding singleton cycles from the cyclic form; however, there are other changes to handling permutations that are waiting to be merged in the combinatorics module here, so I guess I’ll first discuss my changes with Christopher. Currently, I see the following two possibilities for handling the singleton cycles:
    – add a _size attribute to the Permutation class, and then, when faced with something like Permutation([[2, 3], [4, 5, 6], [8]]), find the maximum index appearing in the permutation (here it’s 8) and assign the size of the permutation to that + 1. Then it remains to adjust some of the other methods in the class (after I adjusted mul so that it treats permutations of different sizes as if they leave all points outside their domain fixed, all the tests passed) so that they make sense with that new approach to cyclic forms.
    – more ambitious: make a new class, ExtendedArrayForm or something, with a field _array_form that holds the usual array form of a permutation. Then we overload the __getitem__ method so that if the index is outside the bounds of self._array_form we return the index unchanged. Of course, we’ll have to overload other things, like the __len__ and __str__ to make it behave like a list. Then instead of using a list to initialize the array form of a permutation, we use the corresponding ExtendedArrayForm. This will make all permutations behave as if they are acting on a practically infinite domain, and if we do it that way, we won’t have to make any changes to the methods in Permutation – everything is going to work as expected, no casework like if len(a) > len(b),... will be needed. So this sounds like a rather elegant approach. On the other hand, I’m not entirely sure if it is possible to make it completely like a list, and also it doesn’t seem like a very performance-efficient decision since ExtendedArrayForm instances will be created all the time. (see the discussion here).
  • Still nothing on a database of groups. I looked around the web for a while but didn’t find any resources… the search continues. Perhaps I should ask someone more knowledgeable.

That’s it for now, and that’s the end of my series of blog posts for the GSoC, but I don’t really feel that something has ended since it seems that my contributions to the combinatorics module will continue (albeit not that regularly : ) ). After all, it’s a lot of fun, and there are a lot more things to be implemented/fixed there! So, a big “Thank you” to everyone who helped me get through (and to) GSoC, it’s been a pleasure and I learned a lot. Goodbye!


Google Summer of Code 2012: Week 12

Hi all, here’s a brief summary of the 12th week of my GSoC:

  • Centralizers got some more attention since there were several bugs in the implementation from last week; this also exposed a bug in .subgroup_search() as it is on sympy/master right now. Fortunately, I located it and fixed it earlier today, so the fix for .subgroup_search() will be contained in my next pull request. In fact, it is just three more lines that should be added. Namely,
    # line 29: set the next element from the current branch and update
    # accorndingly
    c[l] += 1
    element = ~(computed_words[l - 1])

    should be replaced with

    # line 29: set the next element from the current branch and update
    # accorndingly
    c[l] += 1
    if l == 0:
        element = identity
        element = ~(computed_words[l - 1])

    since we might be at the bottom level with l=0. In this case, python doesn’t yell at you for looking up computed_words[-1] since negative indices wrap around the list in python. Yet another silly mistake that’s incredibly hard to track down! I hope that it will work properly from now on, and I’ll have to include some more tests to it.

  • The description of the algorithm for finding the center in polynomial time given in [1] didn’t really make sense to me, so instead a straightforward one,
    def center(self):
        return self.centralizer(self)

    was used. This can be updated later when I (or someone else) figures out the polynomial-time algorithm.

  • A new, faster algorithm for finding normal closures: this one uses the incremental version of Schreier-Sims, and some randomization. It’s described in [1].
  • Some applications of normal closure: the derived series, lower cenral series, the commutator of two subgroups of a group, nilpotency testing. Now we have things like this:
    In [68]: from sympy.combinatorics.named_groups import *
    In [69]: S = SymmetricGroup(4)
    In [70]: ds = S.derived_series()
    In [71]: len(ds)
    Out[71]: 4
    In [72]: ds[1] == AlternatingGroup(4)
    Out[72]: True
    In [73]: ds[2] == DihedralGroup(2)
    Out[73]: True
    In [74]: ds[3] == PermutationGroup([Permutation([0, 1, 2, 3])])
    Out[74]: True

    demonstrating the well-known normal series of groups e < K_4 < A_4 < S_4 that solves the symmetric group on 4 letters. Note that the normal closure algorithm was already there thanks to the work of Mario, I just improved it a bit and added some applications.

  • Moved DirectProduct() to a new file,, that is planned to hold functions that treat several groups equally (for one other example, the commutator of two groups in the full symmetric group) rather than treating them in some sort of subgroup-supergroup relationship (such as .centralizer()).

I wrote docstrings for the new stuff, and my current work can be found on my week10 branch. There will be some comprehensive test following the new additions (and I’ll need GAP to verify the results of some of them, probably). It seems that Todd-Coxeter won’t happen during GSoC since there’s just one more week; instead, I plan to focus on improving disjoint cycle notation and group databases.

[1] Derek F. Holt, Bettina Eick, Bettina, Eamonn A. O’Brien, “Handbook of computational group theory”, Discrete Mathematics and its Applications (Boca Raton). Chapman & Hall/CRC, Boca Raton, FL, 2005. ISBN 1-58488-372-3

Google Summer of Code 2012: Week 11

Hi all, here’s a brief summary of the 11th week of my GSoC.

  • Yay! Subgroup searching now works with the use of .stabilizer(), as I discussed in my previous blog post. Surprisingly, the running time is similar to that of the flawed version using .baseswap() (whenever the one using .baseswap() works), you can play around with the two versions on my week6 (has a bug, using .baseswap()) and week9 (seems to work, using .stabilizer()) branches.
  • Consequently, I made a new pull request containing the incremental version of Schreier-Sims, the remove_gens utility for getting rid of redundant generators in a strong generating set, and the new (working) subgroup_search algorithm. You’re most welcome to help with the review!
  • I worked on several applications of subgroup_search() and the incremental Schreier-Sims algorithm. Namely, the pointwise stabilizer of a set of points (via the incremental Schreier-Sims algorithm):
In [4]: from sympy.combinatorics.named_groups import *
In [5]: A = AlternatingGroup(9)
In [6]: G = A.pointwise_stabilizer([2, 3, 5])
In [7]: G == A.stabilizer(2).stabilizer(3).stabilizer(5)
Out[7]: True

(this is much faster than the naive implementation using .stabilizer() repeatedly), and the centralizer of a group H inside a group G:

In [11]: from sympy.combinatorics.named_groups import *
In [12]: S = SymmetricGroup(6)
In [13]: A = AlternatingGroup(6)
In [14]: C = CyclicGroup(6)
In [15]: S_els = list(S.generate())
In [16]: G = S.centralizer(A)
In [17]: G.order()
Out[17]: 1
In [18]: temp = [[el*gen for gen in A.generators] == [gen*el for gen in A.generators] for el in S_els]
In [19]: temp.count(False)
Out[19]: 719
In [20]: temp.count(True)
Out[20]: 1
In [21]: G = S.centralizer(C)
In [22]: G == C
Out[22]: True
In [23]: temp = [[el*gen for gen in C.generators] == [gen*el for gen in C.generators] for el in S_els]
In [24]: temp.count(True)
Out[24]: 6

(it takes some effort to see that these calculations indeed prove that .centralizer() returned the needed centralizer). The centralizer algorithm uses a pruning criterion described in [1], and even though it’s exponential in complexity, it’s fast for practical purposes. Both of the above functions are available (albeit not documented yet) on my week10 branch.

  • The next steps are an algorithm for the centre in polynomial time, and an algorithm to find the intersection of two subgroups! And after that, I hope to be able to implement the Todd-Coxeter algorithm…

That’s it for now!

[1] Derek F. Holt, Bettina Eick, Bettina, Eamonn A. O’Brien, “Handbook of computational group theory”, Discrete Mathematics and its Applications (Boca Raton). Chapman & Hall/CRC, Boca Raton, FL, 2005. ISBN 1-58488-372-3

Google Summer of Code 2012: Week 10

Hi all,

here’s a brief summary of what I’ve been doing during the 10th week of my GSoC.

  • Though I fixed a bug in the SUBGROUPSEARCH function during the week, I ran some more comprehensive tests as I had planned to, and some of them broke the function. If you’re particularly interested, something like that will work:
    In [87]: S = SymmetricGroup(5)
    In [88]: prop_fix_3 = lambda x: x(3) == 3
    In [89]: %autoreload
    In [90]: S.subgroup_search(prop_fix_3)
    StopIteration                             Traceback (most recent call last)
    <ipython-input-90-6b85aa1285b8> in <module>()
    ----> 1 S.subgroup_search(prop_fix_3)
    /home/alexander/workspace/sympy/sympy/combinatorics/ in subgroup_search(self, prop, base, strong_gens, tests, init_subgroup)
    2661                 # this function maintains a partial BSGS structure up to position l
    -> 2662                 _insert_point_in_base(res, res_base, res_strong_gens, l, new_point, distr_gens=res_distr_gens, basic_orbits=res_basic_orbits, transversals=res_transversals)
    2663                 # find the l+1-th basic stabilizer
    2664                 new_stab = PermutationGroup(res_distr_gens[l + 1])
    /home/alexander/workspace/sympy/sympy/combinatorics/ in _insert_point_in_base(group, base, strong_gens, pos, point, distr_gens, basic_orbits, transversals)
    423     # baseswap with the partial BSGS structures. Notice that we need only
    424     # the orbit and transversal of the new point under the last stabilizer
    --> 425     new_base, new_strong_gens = group.baseswap(partial_base, strong_gens, pos, randomized=False, transversals=partial_transversals, basic_orbits=partial_basic_orbits, distr_gens=partial_distr_gens)
    426     # amend the basic orbits and transversals
    427     stab_pos = PermutationGroup(distr_gens[pos])
    /home/alexander/workspace/sympy/sympy/combinatorics/ in baseswap(self, base, strong_gens, pos, randomized, transversals, basic_orbits, distr_gens)
    2472             # ruling out member of the basic orbit of base[pos] along the way
    2473             while len(current_group.orbit(base[pos])) != size:
    -> 2474                 gamma = iter(Gamma).next()
    2475                 x = transversals[pos][gamma]
    2476                 x_inverse = ~x

    The reason is certainly the change of base performed on line 11 in the pseudocode (this is also indicated in my code on my local week6 branch here ). The use of the function BASESWAP there is what gets us into trouble. It is meant to be applied to  base and a strong generating set relative to it, switch two consecutive base points and change the generating set accordinly.  However, in subgroup_search the goal is to change a base (b_1, b_2, \ldots, b_l, \ldots, b_k) to (b_1, b_2, \ldots, b_l', \ldots, b_k) where b_l' is a new point. The book ([1]) mentions that this is done by using BASESWAP but doesn’t provide any details. My strategy is the following: I cut the base so that it becomes (b_1, b_2,\ldots, b_l) and cut the correponding data structures – the strong generators strong_gens, the basic_orbits,  the transversals, and the strong generators distributed according to membership in basic stabilizers distr_gens (I know, I still have to rename this to strong_gens_distr). Then I append the point b_l' so that the base is (b_1, b_2, \ldots, b_l, b_l') and calculate an orbit and transversal for $b_l’$ under the stabilzier of b_1, b_2, \ldots, b_l. Finally I apply BASESWAP to this new base in order to switch the two rightmost points. Then I go back to (b_1, b_2, \ldots, b_l', \ldots, b_k) by appending what I had cut in the start and calculating a transversal/orbit for b_{l+1} under the stabilizer just found, that of b_1, \ldots, b_l'. Obviously, the resulting BSGS structures are valid only up to position l, and that’s all the information we can acquire without another application of baseswap or finding another stabilizer ( and in general, finding a stabilizer is a computationally hard task relative to calculating orbits/transversals). The entire purpose of this use of BASESWAP in SUBGROUPSEARCH is to obtain generators for the stabilizer of b_1, b_2, \ldots, b_l' and maintain a base/strong generating set that are valid up to a certain position. There are many such base changes performed on the same base throughout the course of the function and something goes wrong along the way. I still have to figure out why and where.

  • The good news: There is a straightforward alternative to using BASESWAP: maintain a list of generators for each of the basic stabilizers in (b_1, b_2, \ldots, b_k) and change it accordingly as the base is changed, using the function stabilizer() in sympy/combinatorics/ For each base change we have to calculate one more stabilizer, so that’s not terrible. It is also sort of suggested in “Notes on Computational Group Theory” by Alexander Hulpke (page 34). The problem with this approach is that stabilizer() tends to return a group with many generators, and repeated applications keep increasing this number. However, using this removed the bug from SUBGROUPSEARCH. As before, more comprehensive tests are on the way : )
  • Yet another alternative : we can use the incremental Schreier-Sims algorithm with the new base (b_1, \ldots, b_l', \ldots, b_k) and the strong generating set for (b_1, \ldots, b_l, \ldots, b_k). There will likely be redundant generators after that, and it will probably involve more computation than finding a single stabilizer. However, in the long run (since there are many base changes performed) this might perform faster (due to the increasing number of generators that stabilizer() tends to create). I have not tried that approach yet.
  • Other than that, I had my latest major pull request merged! Thanks a lot to Stefan and my mentor David for the review! That was the largest one so far…
  • I started reading about some of the applications of subgroup search; subgroup intersection seems to be the easiest to implement, so I’ll probably go for it first.

That’s it for now : )

Google Summer of Code 2012: Week 9

Hi all, here’s a brief summary of what I’ve been doing for the 9th week of my GSoC.

This week saw (and still has to see) some exciting new additions:

I. The incremental Schreier-Sims algorithm.

This is a version of the Schreier-Sims algorithm that takes a sequence of points B and a generating set S for a group G as input, and extends B to a base and S to a strong generating set relative to it. It is described in [1], pp.87-93. The default value of B is [], and that of S is \text{G.generators}. Here’s an example:

In [41]: S = SymmetricGroup(5)
In [42]: base = [3, 4]
In [43]: gens = S.generators
In [44]: x = S.schreier_sims_incremental(base, gens)
In [45]: x
([3, 4, 0, 1],
[Permutation([1, 2, 3, 4, 0]),
Permutation([1, 0, 2, 3, 4]),
Permutation([4, 0, 1, 3, 2]),
Permutation([0, 2, 1, 3, 4])])
In [46]: from sympy.combinatorics.util import _verify_bsgs
In [47]: _verify_bsgs(S, x[0], x[1])
Out[47]: True

The current implementation stores the transversals for the basic orbits explicitly (the alternative is to use Schreier vectors to describe the orbits – this saves a lot of space, but requires more time in order to compute transversal elements whenever they are needed. This feature is still to be implemented, and this probably won’t happen in this GSoC). The current implementation of the Schreier-Sims algorithm on the master branch uses Jerrum’s filter (for more details and comparisons of the incremental version and the one using Jerrum’s filter, go here) as an optimization, and also stores the transversals explicitly. The incremental version seems to be asymptotically faster though. Here’s several comparisons of the current version on the master branch and the incremental one which can be found on a local branch of mine which is somewhat inadequately called week6):

For symmetric groups:

In [50]: groups = []
In [51]: for i in range(20, 30):
....:     groups.append(SymmetricGroup(i))
In [52]: for group in groups:
....:     %timeit -r1 -n1 group.schreier_sims()
1 loops, best of 1: 590 ms per loop
1 loops, best of 1: 719 ms per loop
1 loops, best of 1: 981 ms per loop
1 loops, best of 1: 1.35 s per loop
1 loops, best of 1: 1.66 s per loop
1 loops, best of 1: 2.19 s per loop
1 loops, best of 1: 2.74 s per loop
1 loops, best of 1: 3.37 s per loop
1 loops, best of 1: 4.28 s per loop
1 loops, best of 1: 5.37 s per loop
In [53]: for group in groups:
....:     %timeit -r1 -n1 group.schreier_sims_incremental()
1 loops, best of 1: 612 ms per loop
1 loops, best of 1: 737 ms per loop
1 loops, best of 1: 927 ms per loop
1 loops, best of 1: 1.15 s per loop
1 loops, best of 1: 1.41 s per loop
1 loops, best of 1: 1.72 s per loop
1 loops, best of 1: 2.1 s per loop
1 loops, best of 1: 2.52 s per loop
1 loops, best of 1: 3.02 s per loop
1 loops, best of 1: 3.58 s per loop

For alternating groups:

In [54]: groups = []
In [55]: for i in range(20, 40, 2):
....:     groups.append(AlternatingGroup(i))
In [56]: for group in groups:
%timeit -r1 -n1 group.schreier_sims()
1 loops, best of 1: 613 ms per loop
1 loops, best of 1: 1.03 s per loop
1 loops, best of 1: 1.77 s per loop
1 loops, best of 1: 2.65 s per loop
1 loops, best of 1: 3.51 s per loop
1 loops, best of 1: 5.31 s per loop
1 loops, best of 1: 7.71 s per loop
1 loops, best of 1: 11.1 s per loop
1 loops, best of 1: 15.3 s per loop
1 loops, best of 1: 19.1 s per loop
In [57]: for group in groups:
%timeit -r1 -n1 group.schreier_sims_incremental()
1 loops, best of 1: 504 ms per loop
1 loops, best of 1: 787 ms per loop
1 loops, best of 1: 1.23 s per loop
1 loops, best of 1: 1.9 s per loop
1 loops, best of 1: 2.8 s per loop
1 loops, best of 1: 3.99 s per loop
1 loops, best of 1: 5.48 s per loop
1 loops, best of 1: 7.45 s per loop
1 loops, best of 1: 10 s per loop
1 loops, best of 1: 13.2 s per loop

And for some dihedral groups of large degree (to illustrate the case of small-base groups of large degrees):

In [58]: groups = []
In [59]: for i in range(100, 2000, 200):
....:     groups.append(DihedralGroup(i))
In [60]: for group in groups:
%timeit -r1 -n1 group.schreier_sims()
1 loops, best of 1: 29.6 ms per loop
1 loops, best of 1: 108 ms per loop
1 loops, best of 1: 278 ms per loop
1 loops, best of 1: 527 ms per loop
1 loops, best of 1: 861 ms per loop
1 loops, best of 1: 1.29 s per loop
1 loops, best of 1: 1.83 s per loop
1 loops, best of 1: 2.39 s per loop
1 loops, best of 1: 3.06 s per loop
1 loops, best of 1: 3.83 s per loop
In [61]: for group in groups:
%timeit -r1 -n1 group.schreier_sims_incremental()
1 loops, best of 1: 20.8 ms per loop
1 loops, best of 1: 52.8 ms per loop
1 loops, best of 1: 121 ms per loop
1 loops, best of 1: 223 ms per loop
1 loops, best of 1: 365 ms per loop
1 loops, best of 1: 548 ms per loop
1 loops, best of 1: 766 ms per loop
1 loops, best of 1: 1 s per loop
1 loops, best of 1: 1.25 s per loop
1 loops, best of 1: 1.51 s per loop

In addition to this algorithm I implemented a related function _remove_gens in sympy.combinatorics.util which removes redundant generators from a strong generating set (since there tend to be some redundant ones after schreier_sims_incremental() is run):

In [68]: from sympy.combinatorics.util import _remove_gens
In [69]: S = SymmetricGroup(6)
In [70]: base, strong_gens = S.schreier_sims_incremental()
In [71]: strong_gens
[Permutation([1, 2, 3, 4, 5, 0]),
Permutation([1, 0, 2, 3, 4, 5]),
Permutation([0, 5, 1, 2, 3, 4]),
Permutation([0, 1, 2, 3, 5, 4]),
Permutation([0, 1, 2, 4, 3, 5]),
Permutation([0, 1, 3, 2, 4, 5]),
Permutation([0, 1, 2, 5, 4, 3]),
Permutation([0, 1, 5, 3, 4, 2])]
In [72]: new_gens = _remove_gens(base, strong_gens)
In [73]: new_gens
[Permutation([1, 0, 2, 3, 4, 5]),
Permutation([0, 5, 1, 2, 3, 4]),
Permutation([0, 1, 2, 4, 3, 5]),
Permutation([0, 1, 2, 5, 4, 3]),
Permutation([0, 1, 5, 3, 4, 2])]
In [74]: _verify_bsgs(S, base, new_gens)
Out[74]: True

II. Subgroup search.
This is an algorithm used to find the subgroup K of a given group G of all elements of G satisfying a given property P. It is described in [1], pp.114-118 and is quite sophisticated (the book is right when it says “The function SUBGROUPSEARCH is rather complicated and will require careful study by the reader.”). On the other hand, it is one of the most interesting additions to the groups module to date since it can do so much. The idea is to do a depth-first search over all group elements and prune large parts of the search tree based on several different criteria. It’s currently about 150 lines of code and works in many cases but still needs debugging. It can currently do some wonderful stuff like this:

In [77]: S = SymmetricGroup(6)
In [78]: prop = lambda g: g.is_even
In [79]: G = S.subgroup_search(prop)
In [80]: G == AlternatingGroup(6)
Out[80]: True

to find the alternating group as a subgroup of the full symmetric group by the defining property that all its elements are the even permutations, or this:

In [81]: D = DihedralGroup(10)
In [82]: prop_true = lambda g: True
In [83]: G = D.subgroup_search(prop_true)
In [84]: G == D
Out[84]: True

to find the dihedral group D_{10} as a subgroup of itself using the trivial property that always returns \text{True}; or this:

In [106]: A = AlternatingGroup(4)
In [107]: G = A.subgroup_search(prop_fix_23)
In [108]: G == A.stabilizer(2).stabilizer(3)
Out[108]: True

to find the pointwise stabilizer of \{2,3\}. And so on and so on. What is more wonderful is that you can specify the base used for G in advance, and the generating set returned for K will be a strong generating set with respect to that base!

In [119]: A = AlternatingGroup(5)
In [120]: base, strong_gens = A.schreier_sims_incremental()
In [121]: G = A.subgroup_search(prop_fix_1, base=base, strong_gens=strong_gens)
In [122]: G == A.stabilizer(1)
Out[122]: True
In [123]: _verify_bsgs(G, base, G.generators)
Out[123]: True

The bad news is that the function breaks somewhere. For example:

In [125]: S = SymmetricGroup(7)
In [126]: prop_true = lambda g: True
In [127]: G = S.subgroup_search(prop_true)
In [128]: G == S
Out[128]: False

This needs some really careful debugging, but overall it looks promising since it works in so many cases – so the bug is hopefully small : ).

So, that’s it for now!

[1] Derek F. Holt, Bettina Eick, Bettina, Eamonn A. O’Brien, “Handbook of computational group theory”, Discrete Mathematics and its Applications (Boca Raton). Chapman & Hall/CRC, Boca Raton, FL, 2005. ISBN 1-58488-372-3

Google Summer of Code 2012: Week 8

Hi everyone, here’s a brief summary of what I’ve been doing for the 8th week of my GSoC:

  • The issue with the BASESWAP function on page 103 of [1] that I discussed here is now resolved: one of the authors, Professor Derek Holt at Warwick, replied to me that this is indeed a typo and added it to the errata page here.
  • I studied the SUBGROUPSEARCH algorithm described in [1] in more depth. It takes as input a group G with a BSGS, a subgroup K < G with a BSGS having the same base as that of G, a property P such that P(g) for g \in G is either true or false, P(g) is always true for $g \in K$, and the elements of G satisfying P form a subgroup H, and tests \text{TEST}(g, l) used to rule out group elements (i.e., make sure they don’t satisfy P) based on the image of the first l base points of G, the so-called partial base image. It modifies K by adding generators until K = H, and returns a strong generating set for H. It performs a depth-first search over all possible base images (which by the definition of a base determine uniquely every element of G), but uses several conditions to prune the search tree and is said to be fast in practice. This algorithm is the basis for finding normalizers and centralizers and intersections of subgroups, so it’s pretty fundamental. One of its features is the frequent change of base for K: at level l in the search tree we want to make sure that the base for K starts with the current partial base image (i.e., the image of the first l points in the base). In [1] it is said that this requires only one application of BASESWAP (which swaps two neighbouring base points). This was confusing me for a while. However, since we want to only change the l-th base point at any base change, and the base after the l-th point doesn’t matter at level l, it seems that we can do the following. Treat the partial base image, denote it by c_1 c_2 \ldots c_l, as a base, and then run BASESWAP on c_1 c_2 \ldots c_l c, interchanging the last two elements, where c is the new l-th point in the base. Now I’m more confident that I can implement SUBGROUP search (the other parts of the procedure are easily approachable). But there is one other problem with it:
  • We want K, the group we initialize H with, to have the same base as G. The current deterministic implementation of the Schreier-Sims algorithm (using Jerrum’s filther) always produces a BSGS from scratch, and therefore we can’t tell it to make a BSGS for K with respect to some particular base. Hence we need an implementation of the so-called “incremental” Schreier-Sims algorithm, which takes a sequence of points and a generating set and extends them to a BSGS. This is also described in [1], together with some optimizations, and it won’t be very hard to go through the pseudocode and implement it – so that’ going to be the next step. It would also be a useful addition to the entire group-theoretical module since often in algorithms we want a BSGS with respect to some convenient base.

More or less, that’s it for now. In the next few days I’ll try to write some actual code implementing the above two bullets and get some more reviewing for my most recent pull request.

[1] Derek F. Holt, Bettina Eick, Bettina, Eamonn A. O’Brien, “Handbook of computational group theory”, Discrete Mathematics and its Applications (Boca Raton). Chapman & Hall/CRC, Boca Raton, FL, 2005. ISBN 1-58488-372-3


Google Summer of Code 2012: Week 7

Hi all,

here’s a brief summary of what I’ve been doing during the 7th week of my GSoC, as well as a general overview of what’s going on and where things are going with computational group theory in sympy.

Things I did during the week.

This week I focused on:

  • improving the existing code for the functions I recently added – the randomized Schreier-Sims algorithm, the function BASESWAP that changes two points in a base, and the PRINTELEMENTS function (I talk about these here and here). I included some comments in the bodies of the functions since these tend to be quite long. Also, I adopted some new naming conventions for handling all the structures related to a base and a strong generating set. It’d be nice if this naming convention is used throughout the combinatorics module (which for now depends mostly on me, as it seems 🙂 ), and it’d be nice if people provide some feedback on the names I chose. So here we go:
  • making possible the interaction with the deterministic Schreier-Sims algorithm. After some insights from Mario on the values returned by his implementation, I extracted from it the data necessary to make the algorihtms described in [1] that use a base and strong generating set possible.
  • splitting the code further, with the sympy.combinatorics.util file which now holds the internal functions used to handle permutaion groups (this can be later expanded with other internal functions across the combinatorics module).
  • Finally, adding docstrings, tests and making a pull request which is available here . It’s about 1300 lines of code, which is sort of bad, but I can remove some of the stuff and keep it for a future pull request.

So here are the naming conventions for working with a BSGS:

degree – the degree of the permutation group.

base – This is sort of obvious. A base for a permutation group G is an ordered tuple of points (b_1, b_2,\ldots, b_k) such that no group element g \in G fixes all the points b_1, b_2, \ldots, b_k (the significance of the ordering will become apparent later). This is implemented as a list.

base_len – the number of elements in a base.

strong_gens – the strong generating set (relative to some base). This is implemented as a list of Perm objects.

basic_stabilizers – For a base (b_1, b_2,\ldots, b_k) , the basic stabilizers are defined as G^{(i)} = G_{b_1, \ldots, b_{i-1}} := \{ g \in G | g(b_1) = b_1, \ldots, g(b_{i-1}) = b_{i-1}\} for i \in \{1, 2, \ldots, k\} so that we have G^{(1)} = G. This is implemented as a list of permutation groups.

distr_gens – the strong generators distributed according to the basic stabilizers. This means: for a base (b_1, b_2,\ldots, b_k) and a strong generating set S= \{ g_1, g_2, \ldots, g_t\}, distribute the g_i in sets S^{(i)} = G^{(i)} \cap S for i \in \{1, 2,\ldots, k\} where the G^{(i)} are defined as above. This is implemented as a list of lists holding the elements of the S^{(i)}

basic_orbits – these are the orbits of b_i under G^{(i)}. These are implemented as a list of lists, being the list of lists of keys for the basic transversals, see below.

basic_transversals – these are transversals for the basic orbits. Notice that the choice for these may not (and in most cases won’t be) unique. For one thing, it depends on the set of strong generators present (which is also not uniquely determined for a given base). They are implemented as a list of dictionaries indexed according to the base (b_1, b_2,\ldots, b_k) , with keys – the elements of the basic orbits, and values – transversal elements sending the current b_i to the key.

I wrote functions extracting basic_orbits, basic_transversals, basic_stabilizers, distr_gens from only a base and strong generating set, as well as functions for extracting all of them from a base, strong generating set, and a part of them, so that if any of them is available, it can be supplied in order to avoid recalculations.

Also, there is a straightforward test _verify_bsgs in sympy.combinatorics.util that tests a sequence of points and group elements for being a base and strong generating set. It simply verifies the definition of a base and strong generating set relative to it. There will likely be other ways to do that in the future – more effective, but surely more complicated and thus error-prone. This will serve as a robust testing tool

Where we are.

So, here’s a checklist of what I’ve promised in my proposal on the melange website, and which parts of it have already been implemented. This is reading the optimistic timeline. This all pertains to permutation groups, unless specified:

  • handling different representations – NO
  • excluding singleton cycles from the cycle decomposition – NO
  • powers and orders of elements – YES. This was actually already there for permutations.
  • orbits – YES.
  • stabilizers – YES.
  • schreier vectors – YES.
  • randomized Schreier-Sims algorithm – YES
  • handling bases and strong generating sets – YES
  • membership testing – YES (the function _strip in sympy.combinatorics.util)
  • rewriting algorithm – NO.
  • actions on cosets – NO.
  • quotient groups – NO.
  • order of a group – YES. This was already there.
  • subgroup testing – NO.
  • coset enumeration by the Todd-Coxeter algorithm & consequences – NO.
  • primitivity testing – YES.
  • finding (minimal) block systems – YES.
  • general backtrack search for a certain property – No, however easy to do by modifying PRINTELEMENTS.
  • outputting all group elements – YES. This was already there, however PRINTELEMENTS does it in lexicographical order according to a base.
  • Sylow subgroups – NO.
  • calculating the center – NO.
  • pointwise stabilizers (of more than one point, see above) – NO.
  • change of base – YES.
  • product groups – YES.
  • more on finitely presented groups (…) – NO.
  • the p-core – NO.
  • the solvable radical – NO.
  • database of known groups – NO.

Things yet to be done.

Apart from the things that got a “NO” on the list above, the following currently come to mind (I’ll update this list periodically):

  • Work on removing redundant generators from a strong/any generating set, as described in [1].
  • Precompute more properties for the groups in the named groups module (transitivity degrees, bases and strong generating sets, etc.)
  • Add more groups to the named groups module.
  • Fix the issues pointed out in the review of my second pull request.
  • Finally do something for handling representations of finite groups over vector spaces, like working with character tables. It’d be cool to have a function that computes the conjugacy classes for a given group, but I don’t know right now how possible that is.
  • Finally implement the group intersection algorithm… I’m currently starting to work my way through the SUBGROUPSEARCH function which is fundamental for implementing backtracking algorithms for group intersection, centralizers, etc.
  • Upgrade the randomized version of Schreier-Sims to Las Vegas type in the case when the order of the group is known.
  • Currently, transversal elements for the basic orbits for a stabilizer chain are stored explicitly. This requires too much memory for large groups. An alternative solution (which slows down execution) is to use Schreier vectors to describe the orbits. This means supplying some more arguments and adding code to many of the functions already present, and is a significant challenge by itself. The good news is that it can be carried out without modifying what is already there.
  • Come up with a more concise functionality to relate the different structures used to describe a base and strong generating set: the generators for basic stabilizers, the basic orbits, the basic transversals… There are many situations in which some of these are given and we need some of the other ones; sometimes it’s more convenient to get the orbits as sets, and sometimes as lists, and so on… the current approach is to write a new utility function whenever the present ones don’t suffice.
  • Handle the case when the identity element is provided as a generator for a permutation group – this can make some algorithms less efficient.
  • Optimize the behavior of BASESWAP so that only the i-th and i+1-th transversals are calculated.
  • Reduce side effects as much as possible (let’s be pythonic!)
  • Improve the docstring quality: it might be reasonable to lay out the theory/notation/definitions behind the Schreier-Sims algorithm in one place in some of the files and then simply refer to it as necessary. Otherwise the descriptions get unnecessarily long.

Well, that’s it for now it seems. If anything else pops up soon, I’ll add it here!


Exploring and venting about quantitative issues

Rational Altruist

Adventures of a would-be do-gooder.

Gowers's Weblog

Mathematics related discussions

What's new

Updates on my research and expository papers, discussion of open problems, and other maths-related topics. By Terence Tao

Stefan's Blog (archive, for news see

Most of the blog moved to