Improving the performance of a package for working with rational functions

As Mathematica gets slow for large symbolic calculations, the cost of putting terms over a common denominator (Together), in particular, gets too high. It occurred to me that, if one has a small number of variables, it should be much faster if we represent our expressions by arrays of their coefficients.

I have implemented a small package that does this with an improvement of a factor of 10 in both speed and memory (in putting Together a sum of rational terms, more or less independent of the size of the sum), but it seems to me that I should be able to do better. I post my package here with the hope that someone might point out glaring performance bottlenecks.

First, the public initializations, with their usability comments.

AA::usage = "DATA STRUCTURE AA[num,den] that stores polynomials as their tensor coefficients of a basis, multiplication is the tensor product"
PolytoAA::usage = "PolytoAA[poly_,vars_List] returns an AA[num_SparseArray,1]"
AAtoPoly::usage = "AAtoPoly[aa_AA,vars_] gives back a rational function of the polynomials num/den"


The actual code begins here, PolProd takes care of multiplying polynomials

SetAttributes[PolProd, Orderless]
PolProd[i__Integer] := Times @ i;
PolProd[sa__SparseArray] := Outer[Times, sa];
(*multiplying 2 polynomials is the tensor product in their vector space*)
PolProd[i__Integer, sa__SparseArray] := Times[i] Outer[Times, sa];

(*It would be cleaner and more object oriented to call two instances of polprod
on i and sa but it would force more pattern matching so it would be less efficient*)


This code performs the arithmetic operations on the AA[num,den] data structure

AA /: i_Integer*AA[num_, den_] := AA[i*num, den]
(*absorb all integer into AA, this means that the sum rule has to do less
pattern matching and is more efficient, overall more efficient?*)
AA /: AA[a_, a_] := 1
AA /: AA[0, _] := 0
AA /: AA[num1_, den1_]*AA[num2_, den2_] :=
AA @@ (PolProd[#1, #2]& @@@ {{num1, num2}, {den1, den2}})
AA /: AA[num1_, den_] + AA[num2_, den_] :=
AA[num1 + num2, den]
(*can't check for a SparseArray cos it might be an Integer,
by dimensional analysis we should get two arrays here though*)
AA /: AA[num1_, den1_] + AA[num2_, den2_] :=
AA[PolProd[num1, den2] + PolProd[num2, den1], PolProd[den1, den2]]
(*together rule, this a recursive implementation that doesn't treat the
sum of many AA all at once, this is intentional cause it minimises the
number of array outer products that are calculated and I believe it's
the more efficient but maybe not*)
AA /: aa_AA^n_Integer :=
If[n < 0, Reverse @ #, #]& @ (PolProd[Sequence @@ ConstantArray[#, Abs @ n]]& /@ aa)
(*this takes care of both efficiency of recursively pattern matching the
product rule on large n and dividing by AA*)
(*NOTICE THE WEIRD SYNTAX/BUG If[n<0, Reverse @ #, #]& @ THERE ARE TWO INTANCES OF @*)


I also post the the functions that go between AA and the polynomials for the sake of clarity and completeness

AAtoPoly[aa_AA,vars_] := #1/#2& @@ (aa /. l_SparseArray :>
Dot[l, Sequence @@ ConstantArray[vars, Depth[l] - 1 (*gives rank of the tensor*)]])
PolytoAA[expr_^n_, vars_List] := PolytoAA[expr, vars]^n
PolytoAA[expr : (_Times | _Plus), vars_List] := PolytoAA[#, vars]& /@ expr
PolytoAA[expr_,vars_List] /; PolynomialQ[expr, vars] :=
AA[Last[CoefficientArrays[expr, vars]](*num*), 1 (*den*)]


BENCHMARKS

I define some random, small polynomials

vars = a /@ Range[10]
pols = Table[
a[RandomInteger[{1, 10}]] + a[RandomInteger[{1, 10}]], {20}] //
DeleteDuplicates
% // Length


{a[1], a[2], a[3], a[4], a[5], a[6], a[7], a[8], a[9], a[10]}

{a[1] + a[7], a[4] + a[6], a[2] + a[4], a[5] + a[9], 2 a[8], a[5] + a[6], a[2] + a[9], a[4] + a[7], a[4] + a[9], a[7] + a[9], a[3] + a[7], a[2] + a[6], a[1] + a[9], a[2] + a[3], a[7] + a[10], 2 a[2], a[3] + a[9]}

17

First I put them Together using my code

aa = PolytoAA[#, vars] & /@ pols;
togaa = Sum[aa[[i]]/aa[[i + 1]], {i, Length@pols - 1}] //
AbsoluteTiming
% // ByteCount


{0.8960513,AA[SparseArray[<274432>,{10,10,10,10,10,10,10,10,10,10,10,10,10,10,10,10}],SparseArray[<16384>,{10,10,10,10,10,10,10,10,10,10,10,10,10,10,10,10}]]}

18614128

and now without my implementation

togpols = (Sum[pols[[i]]/pols[[i + 1]], {i, Length@pols - 1}] //
Together) // AbsoluteTiming//
% // ByteCount


{8.2054693,...}

34567480

I check that both give the same result

togpols2 = AAtoPoly[togaa // Last, vars](*//Simplify*);
Last[togpols] - togpols2 // Together


0

-
Could you add your benchmarking results as well? – jVincent Aug 25 '13 at 11:21
Just added benchmarks, sorry for the delay, was travelling. – Eduardo Serna Aug 27 '13 at 11:38
I think the Together relative speed loss is from it pulling out common denominators. Just checking for these can be costly since it involves multivariate polynomial gcd extraction. – Daniel Lichtblau Aug 28 '13 at 14:41
@DanielLichtblau I wasn't aware of this. My code works with polynomials in Expanded form effectively so it bypasses this at the cost of having common factors in both numerator and denominar, this could make it more costly in memory for certain types of exppressions but I think it should be faster anyways most of the time – Eduardo Serna Aug 30 '13 at 9:26
Any particular reason you're restricting to integer coefficients? Just efficiency? – Matt Nov 19 '13 at 19:20