> < ^ Date: Mon, 15 Mar 1993 14:49:21 +0100
> < ^ From: Alexander Hulpke <hulpke@math.colostate.edu >
< ^ Subject: Re: character tables

Dear GAP-Forum,
in his letter, David sibley wrote:

gap> g:=ThreeGroup(243,1);;
gap> t:=CharTablePGroup(g);;

It took just about an hour to construct the character table. This is a
lot more than I am used to for groups of this size (roughly) from
version 3.1. However, I was appalled when I looked more closely and
discovered this group is the cyclic group of order 243. An hour to find
the character table of a cyclic group? (This is actual running time,
according to ps.)

Naturally, it is disappointing, if an algorithm shows poor performance in
trivial cases. However, this special example will also run as fast (or better:
as slow) in GAP-3.1.
To explain this seemingly strange behaviour, one has to look
a bit inside the algorithms:

Computing character tables can be a very tedious task. Especially, there are
some time consuming subroutines, that have to be called very often (like
identification of a class). Thus, the routines do some pre-computation to
make things faster afterwards. However, if you are dealing with an abelian or
even a cyclic group, these computations, do not need to be done, as one can
write down the appropriate character table from scratch. The only real
'task' to be done is attaching the classes of the group to the appropriate
column of the character table.

Thus the question would be: Why doesn't the routine detect, that computation
will be that simple?
CharTablePGroup is a routine, that computes the character table, using the
algorithm of U. Baum as explained by Thomas Breuer in some previous mail to
the forum. In the next version (hopefully, this was left out of 3.2 due to time
reasons and since the computation of character tables of cyclic groups was
not very high on our schedule), the general dispatcher CharTable will provide
an all-purpose routine for computing character tables. This implies, that
computations will be shortened for abelian groups. A routine like
CharTablePGroup still will exist, and provide the specific algorithm, which
might result in poor performance for specific examples.

Is there some flaw in the method being used here?

I would not call this behaviour a 'flaw'. CharTablePGroup is a specific
routine. Adding examinations of special cases to this routine will result in
poorer performance in all other cases, since the test for this special case
must be run always. The handling of these special cases is a task of the
more general dispatcher. The user may then choose, if he uses the general,
or a specific routine: The general routine is slower, but takes the task of
selecting a specific routine itself (though this selection might be not
as perfect, as the one done by a human). It is just a question, how thinking
is divided between human and computer.

I am a bit worried
that there might be similar performance problems if a group happens to
have a large cyclic subgroup or quotient group in the wrong place.

This behaviour most likely will reappear, if the group has a large cyclic
factorgroup. To cope with this beaviour surely is part of the DWIM (Do What
I Mean) problem (baptized this way by Charles Leedham-Green):
If I ask the computer to compute the character table of a large abelian
group, most likely, I'm not interested in a long list of irreducible
characters (unless I want to check memory allocation routines...).
The information, I can make use of, would be, that the group *is* abelian,
that the structure is Z_a x Z_b x ..., and maybe in *some* irreducible
characters, from which the other ones can be obtained by tensor products.
If I have a group with a large cyclic factorgroup, either the group itself
is very large --- then the cyclic factor will not count --- or the cyclic
factor is of disproportional size, compared to the normal subgroup. Then the
group most likely is of some special structure, and much more information
can be gained, if I use this special structure.
Of course this implies, that special algorithms for special structures
should be available, i.e. the program should behave 'intelligent'.
Reaching that point will mean *a lot* of work, I do not even dare to
estimate.

Finally, I still did not answer, how to cope with this behaviour (unless
one wants to wait centuries for the finished GAP 3.$\infty$). I can only
give some suggestions for coping with it:
Before computing a character table, I would check, how many conjugacy classes
the group has. In the problematic cases, the group will have a lot of classes.
Afterwards one can decide, if one *really* wants the character table of that
kind of group. (In case if one wants it, there still is the question,
whether the table should be computed from the group --- which is only
necessary, if one wants a correspondence of class representatives and columns
--- or whether one will compute the table itself by CharTableDirectProduct
or similar.)
In case, this does not help for your particular problem, feel free to ask;
maybe there is a problem, we never thought about.

I hope this rather lengthy pamphlet has shed some light on this problem.

Alexander Hulpke


> < [top]