"Exact and approximate algorithms for resultant polytopes."

1. Exact and approximate algorithms for resultant polytopes Vissarion Fisikopoulos Joint work with I.Z. Emiris and C. Konaxis∗ Dept Informatics & Telecoms, University of Athens ∗ currently with Univeristy of Crete EuroCG 2012, Assisi, Perugia, Italy, 20.March.2012

2. Polynomials and Newton polytopes The support of a polynomial f is the set of exponents of its monomials with non-zero coeﬃcient. The Newton polytope of f is the convex hull of its support. f(x, y) = 8y + xy − 24y2 − 16x2 + 220x2 y − 34xy2 − 84x3 y + 6x2 y2 − 8xy3 + 8x3 y2 + 8x3 + 18y3 3 2 1 1 2 3 5 5

3. Polynomial systems We study polynomials that expresses the solvability of polynomial systems. Given a system of n + 1 linear polynomials f0, f1, . . . , fn, on n variables the determinant is a polynomial on the coeﬃcients which is zero iﬀ the system has a common solution.

4. Polynomial systems Given a system of n + 1 linear polynomials f0, f1, . . . , fn, on n variables the determinant is a polynomial on the coeﬃcients which is zero iﬀ the system has a common solution. f0 = ax + by + c = 0 f1 = dx + ey + f = 0 f2 = gx + hy + i = 0 a b c d e f g h i

5. Polynomial systems Given a system of n + 1 linear polynomials f0, f1, . . . , fn, on n variables the determinant is a polynomial on the coeﬃcients which is zero iﬀ the system has a common solution. f0 = 4x + y + 2 = 0 f1 = x + 2y + 1 = 0 f2 = x + y + 8 = 0 4 1 2 1 2 1 1 1 8 = 51

6. Polynomial systems Given a system of n + 1 linear polynomials f0, f1, . . . , fn, on n variables the determinant is a polynomial on the coefficients which is zero iff the system has a common solution. f0 = ax + by + c = 0 f1 = dx + ey + f = 0 f2 = gx + hy + i = 0 supports a b c d e f g h i Newton polytope of determinant (Birkhoff polytope)

7. Polynomial systems Given a system of n + 1 linear general polynomials f0, f1, . . . , fn, on n variables the determinant resultant is a polynomial on the coeﬃcients which is zero iﬀ the system has a common solution. f0 = 4xy2 +x4 y+2 = 0 f1 = x + 2y = 0 f2 = 3x2 + y + 8 = 0

8. Polynomial systems Given a system of n + 1 linear general polynomials f0, f1, . . . , fn, on n variables the determinant resultant is a polynomial on the coeﬃcients which is zero iﬀ the system has a common solution. f0 = 4xy2 +x4 y+2 = 0 f1 = x + 2y = 0 f2 = 3x2 + y + 8 = 0 hard to compute the resultant

9. Polynomial systems Given a system of n + 1 linear general polynomials f0, f1, . . . , fn, on n variables the determinant resultant is a polynomial on the coefficients which is zero iff the system has a common solution. f0 = 4xy2 +x4 y+2 = 0 f1 = x + 2y = 0 f2 = 3x2 + y + 8 = 0 supports hard to compute the resultant Newton polytope of determinant resultant (Birkhoff resultant polytope Π)

10. Polynomial systems Given a system of n + 1 linear general polynomials f0, f1, . . . , fn, on n variables the determinant resultant is a polynomial on the coefficients which is zero iff the system has a common solution. supports Newton polytope of determinant resultant (Birkhoff resultant polytope Π)

11. The idea of the algorithm The input supports deﬁne a pointset A ∈ Z2n Naive method: compute the secondary polytope Σ(A) to compute Π

12. The idea of the algorithm The input supports deﬁne a pointset A ∈ Z2n Naive method: compute the secondary polytope Σ(A) to compute Π Idea: incrementally construct Π using an oracle that given a direction produces vertices of Π

13. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step Π Q initialization: Q ⊂ Π dim(Q)=dim(Π)

14. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal Π Q 2 kinds of hyperplanes of QH : legal if it supports facet ⊂ Π illegal otherwise

15. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal 3. while ∃ illegal hyperplane H ⊂ QH with outer normal w do call oracle for w and compute v, QV ← QV ∪ {v} Π Q w Extending an illegal facet

16. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal 3. while ∃ illegal hyperplane H ⊂ QH with outer normal w do call oracle for w and compute v, QV ← QV ∪ {v} if v /∈ QV ∩ H then QH ← CH(QV ∪ {v}) else H is legal Π Q Extending an illegal facet

17. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal 3. while ∃ illegal hyperplane H ⊂ QH with outer normal w do call oracle for w and compute v, QV ← QV ∪ {v} if v /∈ QV ∩ H then QH ← CH(QV ∪ {v}) else H is legal Π Q Validating a legal facet

18. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal 3. while ∃ illegal hyperplane H ⊂ QH with outer normal w do call oracle for w and compute v, QV ← QV ∪ {v} if v /∈ QV ∩ H then QH ← CH(QV ∪ {v}) else H is legal Π Q Validating a legal facet

19. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal 3. while ∃ illegal hyperplane H ⊂ QH with outer normal w do call oracle for w and compute v, QV ← QV ∪ {v} if v /∈ QV ∩ H then QH ← CH(QV ∪ {v}) else H is legal Π Q

20. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal 3. while ∃ illegal hyperplane H ⊂ QH with outer normal w do call oracle for w and compute v, QV ← QV ∪ {v} if v /∈ QV ∩ H then QH ← CH(QV ∪ {v}) else H is legal Π Q At any step, Q is an inner approximation . . .

21. Incremental Algorithm Input: A Output: H-rep. QH, V-rep. QV of Q = Π 1. initialization step 2. all hyperplanes of QH are illegal 3. while ∃ illegal hyperplane H ⊂ QH with outer normal w do call oracle for w and compute v, QV ← QV ∪ {v} if v /∈ QV ∩ H then QH ← CH(QV ∪ {v}) else H is legal Π Q Qo At any step, Q is an inner approximation . . . from which we can compute an outer approximation Qo.

36. Complexity Theorem We compute the Vertex- and Halfspace-representations of Π, as well as a triangulation T of Π, in O∗ ( m5 |vtx(Π)| · |T|2 ), where m = dim Π, and |T| the number of full-dim faces of T. Elements of proof Most computation is done in dimension ≤ m. At most ≤ vtx(Π) + fct(Π) oracle calls Beneath-Beyond algorithm for converting V-rep. to H-rep. (bottleneck)

37. ResPol Implementation Tools C++, CGAL, triangulation [Boissonnat,Devillers,Hornus], extreme points d [G¨artner] Experiments (dim(Π) = 4) 0.01 0.1 1 10 100 1000 10000 100000 10 15 20 25 30 35 40 45 time(sec) Number of points Respol-hash Respol-no hash Gfan-TTR Gfan-NFSI

38. Future work approximate resultant polytopes (dim(Π) ≥ 7) preliminary results: input m 3 3 4 4 5 5 |A| 200 490 20 30 17 20 exact #vtx(Π) 98 133 416 1296 1674 5093 time 2.03 5.87 3.72 25.97 51.54 239.96 approx. #vtx(Qin) 15 11 63 121 – – vol(Qin)/vol(Π) 0.96 0.95 0.93 0.94 – – vol(Qout)/vol(Π) 1.02 1.03 1.04 1.03 – – time 0.15 0.22 0.37 1.42 > 10hr > 10hr approximate volume computation [Lovász-Vempala06]

39. References The code http://respol.sourceforge.net The full version of the paper http://arxiv.org/abs/1108.5985v2

40. References The code http://respol.sourceforge.net The full version of the paper http://arxiv.org/abs/1108.5985v2 Thank You !

41. Convex hull implementations From V- to H-rep. of Π. triangulation (on/off-line), polymake beneath-beyond, cdd, lrs 0.01 0.1 1 10 100 0 500 1000 1500 2000 2500 3000 time(sec) Number of points bb cdd lrs triang_oﬀ triang_on dim(Π) = 4

"Exact and approximate algorithms for resultant polytopes."

More Related Content

What's hot (6)

More from Vissarion Fisikopoulos (20)

Recently uploaded (20)

"Exact and approximate algorithms for resultant polytopes."