Split Rank of Triangle and Quadrilateral Inequalities Quentin - - PowerPoint PPT Presentation

split rank of triangle and quadrilateral inequalities
SMART_READER_LITE
LIVE PREVIEW

Split Rank of Triangle and Quadrilateral Inequalities Quentin - - PowerPoint PPT Presentation

Split Rank of Triangle and Quadrilateral Inequalities Quentin Louveaux Universit e de Li` ege - Montefiore Institute January 2009 Joint work with Santanu Dey (CORE) Quentin Louveaux (Universit e de Li` ege - Montefiore Institute) Split


slide-1
SLIDE 1

Split Rank of Triangle and Quadrilateral Inequalities

Quentin Louveaux

Universit´ e de Li` ege - Montefiore Institute

January 2009 Joint work with Santanu Dey (CORE)

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 1 / 1

slide-2
SLIDE 2

Outline

Cuts from two rows of the simplex tableau The different cases to consider Split cuts and split ranks Finiteness proofs for the triangles The ideas for the quadrilaterals Conclusion

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 2 / 1

slide-3
SLIDE 3

Cuts from two rows of the simplex tableau

Consider a mixed-integer program min cTx s.t. Ax = b x ∈ Zn1

+ × Rn2 + .

We consider the problem of finding valid inequalities cutting off the linear relaxation

  • ptimum.

We consider the simplex tableau x1 − ¯ a11s1−· · ·− ¯ a1nsn = ¯ b1 ... . . . xm−¯ am1s1−· · ·−¯ amnsn = ¯ bm.

  • Select two rows
  • Relax the integrality requirements of the non-basic variables
  • Relax the nonnegativity requirements of the basic variables but keeping integrality

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 3 / 1

slide-4
SLIDE 4

Cuts from two rows of the simplex tableau

Consider a mixed-integer program min cTx s.t. Ax = b x ∈ Zn1

+ × Rn2 + .

We consider the problem of finding valid inequalities cutting off the linear relaxation

  • ptimum.

We consider the simplex tableau x1 − ¯ a11s1−· · ·− ¯ a1nsn = ¯ b1 ... . . . xm−¯ am1s1−· · ·−¯ amnsn = ¯ bm.

  • Select two rows
  • Relax the integrality requirements of the non-basic variables
  • Relax the nonnegativity requirements of the basic variables but keeping integrality

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 3 / 1

slide-5
SLIDE 5

Cuts from two rows of the simplex tableau

Consider a mixed-integer program min cTx s.t. Ax = b x ∈ Zn1

+ × Rn2 + .

We consider the problem of finding valid inequalities cutting off the linear relaxation

  • ptimum.

We consider the simplex tableau x1 − ¯ a11s1−· · ·− ¯ a1nsn = ¯ b1 ... . . . xm−¯ am1s1−· · ·−¯ amnsn = ¯ bm.

  • Select two rows
  • Relax the integrality requirements of the non-basic variables
  • Relax the nonnegativity requirements of the basic variables but keeping integrality

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 3 / 1

slide-6
SLIDE 6

The 2 row-model

The model

„ x1 x2 « = „ f1 f2 « +

n

X

j=1

„ r j

1

r j

2

« sj, x1, x2 ∈ Z, sj ∈ R+ Model studied in [Andersen, Louveaux, Weismantel, Wolsey, IPCO2007] (for the finite case) and [Cornu´ ejols, Margot, 2009] (for the infinite case).

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 4 / 1

slide-7
SLIDE 7

The 2 row-model

The model

„ x1 x2 « = „ f1 f2 « +

n

X

j=1

„ r j

1

r j

2

« sj, x1, x2 ∈ Z, sj ∈ R+

The geometry

„ x1 x2 « = „ 1/4 1/2 « + „ 2 1 « s1 + „ 1 1 « s2 + „ −3 2 « s3 + „ −1 « s4 + „ 1 −2 « s5

r r r r 2 3 4 5 f x 1 x 2 r 1 Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 4 / 1

slide-8
SLIDE 8

The geometry

The projection picture

2s1 + 2s2 + 4s3 + s4 + 12 7 s5 ≥ 1 We project the n + 2-dim space onto the x-space The facet is represented by a polygon Lα There is no integer point in the interior of Lα The coefficients are a ratio of distances on the figure α1 α3

r r r r 2 3 4 5 f x 1 x 2 r 1 Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 5 / 1

slide-9
SLIDE 9

The geometry

The projection picture

2s1 + 2s2 + 4s3 + s4 + 12 7 s5 ≥ 1 We project the n + 2-dim space onto the x-space The facet is represented by a polygon Lα There is no integer point in the interior of Lα The coefficients are a ratio of distances on the figure α1 α3

r r r r 2 3 4 5 f x 1 x 2 r 1 Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 5 / 1

slide-10
SLIDE 10

The geometry

The projection picture

2s1 + 2s2 + 4s3 + s4 + 12 7 s5 ≥ 1 We project the n + 2-dim space onto the x-space The facet is represented by a polygon Lα There is no integer point in the interior of Lα The coefficients are a ratio of distances on the figure α1 α3

  • r

r r r 2 3 4 5 f x 1 x 2 r 1 Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 5 / 1

slide-11
SLIDE 11

The geometry

The projection picture

2s1 + 2s2 + 4s3 + s4 + 12 7 s5 ≥ 1 We project the n + 2-dim space onto the x-space The facet is represented by a polygon Lα There is no integer point in the interior of Lα The coefficients are a ratio of distances on the figure α1 α3

r r r r 2 3 4 5 f x 1 x 2 r 1 Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 5 / 1

slide-12
SLIDE 12

The geometry

The projection picture

2s1 + 2s2 + 4s3 + s4 + 12 7 s5 ≥ 1 We project the n + 2-dim space onto the x-space The facet is represented by a polygon Lα There is no integer point in the interior of Lα The coefficients are a ratio of distances on the figure α1 α3

r r r r 2 3 4 5 f x 1 x 2 r 1 Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 5 / 1

slide-13
SLIDE 13

The geometry

The projection picture

2s1 + 2s2 + 4s3 + s4 + 12 7 s5 ≥ 1 We project the n + 2-dim space onto the x-space The facet is represented by a polygon Lα There is no integer point in the interior of Lα The coefficients are a ratio of distances on the figure α1 α3

1 r r r r 2 3 4 5 f x 1 x 2 r Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 5 / 1

slide-14
SLIDE 14

Classification of all possible facet-defining inequalities

Theorem : All facets are projected to triangles and quadrilaterals [Andersen et al 2007].

Cook−Kannan−Schrijver Disection Quadrilateral

Quadrilateral Cut Split Cut Triangle Cut Disection Triangle

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 6 / 1

slide-15
SLIDE 15

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P x = f + RS s1 ≥ 0 . . . sn ≥ 0 πTx ≤ π0 The first split closure P1 of P is what you obtain after having applied all possible split disjunctions π. The split rank of a valid inequality is the minimum i such that the inequality is valid for Pi Most inequalities used in commercial softwares are split cuts Question : what is the split rank of the 2 row-inequalities ? In how many rounds of split cuts only can we generate the inequalities ? The Cook-Kannan-Schrijver has infinite rank and we prove that the other triangles have finite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-16
SLIDE 16

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P x = f + RS s1 ≥ 0 . . . sn ≥ 0 πTx ≤ π0 The first split closure P1 of P is what you obtain after having applied all possible

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-17
SLIDE 17

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P x = f + RS s1 ≥ 0 . . . sn ≥ 0 πTx ≤ π0 The first split closure P1 of P is what you obtain after having applied all possible

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-18
SLIDE 18

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P x = f + RS s1 ≥ 0 . . . sn ≥ 0 πTx ≤ π0 The first split closure P1 of P is what you obtain after having applied all possible

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-19
SLIDE 19

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P x = f + RS s1 ≥ 0 . . . sn ≥ 0 πTx ≤ π0 The first split closure P1 of P is what you obtain after having applied all possible

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-20
SLIDE 20

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P The first split closure P1 of P is what you obtain after having applied all possible split disjunctions π. The split rank of a valid inequality is the minimum i such that the inequality is valid for Pi Most inequalities used in commercial softwares are split cuts Question : what is the split rank of the 2 row-inequalities ? In how many rounds of split cuts only can we generate the inequalities ? The Cook-Kannan-Schrijver has infinite rank and we prove that the other triangles have finite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-21
SLIDE 21

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P The first split closure P1 of P is what you obtain after having applied all possible split disjunctions π. The split rank of a valid inequality is the minimum i such that the inequality is valid for Pi Most inequalities used in commercial softwares are split cuts Question : what is the split rank of the 2 row-inequalities ? In how many rounds of split cuts only can we generate the inequalities ? The Cook-Kannan-Schrijver has infinite rank and we prove that the other triangles have finite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-22
SLIDE 22

The split rank question

Split cut : applying a disjunction πTx ≤ π0 ∨ πTx ≥ π0 + 1 to a polyhedron P The first split closure P1 of P is what you obtain after having applied all possible split disjunctions π. The split rank of a valid inequality is the minimum i such that the inequality is valid for Pi Most inequalities used in commercial softwares are split cuts Question : what is the split rank of the 2 row-inequalities ? In how many rounds of split cuts only can we generate the inequalities ? The Cook-Kannan-Schrijver has infinite rank and we prove that the other triangles have finite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 7 / 1

slide-23
SLIDE 23

Useful properties of the split rank

The split rank is invariant up to integer translation and unimodular transformation (Lifting) Consider a triangle (or quadrilateral) inequality for a 3-variable problem. If we keep the same shape of the polygon and consider an n-variable problem, the split rank does not increase. It allows us to work with 3 variables only when trying to find the split rank of triangles. (Projection) Let Pn

i=1 αisi ≥ 1 be an inequality with split rank η. Then the projected

inequality Pn−1

i=1 αisi ≥ 1 has a split rank of at most η for the projected problem.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 8 / 1

slide-24
SLIDE 24

Useful properties of the split rank

The split rank is invariant up to integer translation and unimodular transformation (Lifting) Consider a triangle (or quadrilateral) inequality for a 3-variable problem. If we keep the same shape of the polygon and consider an n-variable problem, the split rank does not increase. It allows us to work with 3 variables only when trying to find the split rank of triangles. (Projection) Let Pn

i=1 αisi ≥ 1 be an inequality with split rank η. Then the projected

inequality Pn−1

i=1 αisi ≥ 1 has a split rank of at most η for the projected problem.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 8 / 1

slide-25
SLIDE 25

Useful properties of the split rank

The split rank is invariant up to integer translation and unimodular transformation (Lifting) Consider a triangle (or quadrilateral) inequality for a 3-variable problem. If we keep the same shape of the polygon and consider an n-variable problem, the split rank does not increase. It allows us to work with 3 variables only when trying to find the split rank of triangles. (Projection) Let Pn

i=1 αisi ≥ 1 be an inequality with split rank η. Then the projected

inequality Pn−1

i=1 αisi ≥ 1 has a split rank of at most η for the projected problem.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 8 / 1

slide-26
SLIDE 26

Useful properties of the split rank

The split rank is invariant up to integer translation and unimodular transformation (Lifting) Consider a triangle (or quadrilateral) inequality for a 3-variable problem. If we keep the same shape of the polygon and consider an n-variable problem, the split rank does not increase. It allows us to work with 3 variables only when trying to find the split rank of triangles. (Projection) Let Pn

i=1 αisi ≥ 1 be an inequality with split rank η. Then the projected

inequality Pn−1

i=1 αisi ≥ 1 has a split rank of at most η for the projected problem.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 8 / 1

slide-27
SLIDE 27

The triangle case

Several cases to consider, after suitable unimodular transformation An illustration of the proof in this talk

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 9 / 1

slide-28
SLIDE 28

The triangle case

Several cases to consider, after suitable unimodular transformation An illustration of the proof in this talk

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 9 / 1

slide-29
SLIDE 29

The triangle case

Several cases to consider, after suitable unimodular transformation An illustration of the proof in this talk

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 9 / 1

slide-30
SLIDE 30

The triangle case

Several cases to consider, after suitable unimodular transformation An illustration of the proof in this talk

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 9 / 1

slide-31
SLIDE 31

The triangle case

Several cases to consider, after suitable unimodular transformation An illustration of the proof in this talk

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 9 / 1

slide-32
SLIDE 32

The triangle case

Several cases to consider, after suitable unimodular transformation An illustration of the proof in this talk

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 9 / 1

slide-33
SLIDE 33

The triangle case

Several cases to consider, after suitable unimodular transformation An illustration of the proof in this talk

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 9 / 1

slide-34
SLIDE 34

Idea of the proof of upper bounds

We prove an upper bound on the split rank. Procedure : We apply a sequence of two split disjunctions. Successively : x1 ≤ 0 ∨ x1 ≥ 1 and x2 ≤ 0 ∨ x2 ≥ 1 At step i , we keep one inequality of rank at most i and proceed to the next disjunction. We prove that this procedure converges in finite time to the desired inequality.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 10 / 1

slide-35
SLIDE 35

Idea of the proof of upper bounds

We prove an upper bound on the split rank. Procedure : We apply a sequence of two split disjunctions. Successively : x1 ≤ 0 ∨ x1 ≥ 1 and x2 ≤ 0 ∨ x2 ≥ 1 At step i , we keep one inequality of rank at most i and proceed to the next disjunction. We prove that this procedure converges in finite time to the desired inequality.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 10 / 1

slide-36
SLIDE 36

Idea of the proof of upper bounds

We prove an upper bound on the split rank. Procedure : We apply a sequence of two split disjunctions. Successively : x1 ≤ 0 ∨ x1 ≥ 1 and x2 ≤ 0 ∨ x2 ≥ 1 At step i , we keep one inequality of rank at most i and proceed to the next disjunction. We prove that this procedure converges in finite time to the desired inequality.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 10 / 1

slide-37
SLIDE 37

Idea of the proof of upper bounds

We prove an upper bound on the split rank. Procedure : We apply a sequence of two split disjunctions. Successively : x1 ≤ 0 ∨ x1 ≥ 1 and x2 ≤ 0 ∨ x2 ≥ 1 At step i , we keep one inequality of rank at most i and proceed to the next disjunction. We prove that this procedure converges in finite time to the desired inequality.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 10 / 1

slide-38
SLIDE 38

One proof for a non-degenerate non-maximal triangle

Rank 0

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-39
SLIDE 39

One proof for a non-degenerate non-maximal triangle

Rank 0

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-40
SLIDE 40

One proof for a non-degenerate non-maximal triangle

Rank 1

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-41
SLIDE 41

One proof for a non-degenerate non-maximal triangle

Rank 1

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-42
SLIDE 42

One proof for a non-degenerate non-maximal triangle

Rank 1

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-43
SLIDE 43

One proof for a non-degenerate non-maximal triangle

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-44
SLIDE 44

One proof for a non-degenerate non-maximal triangle

Rank 2

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-45
SLIDE 45

One proof for a non-degenerate non-maximal triangle

Rank 2

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-46
SLIDE 46

One proof for a non-degenerate non-maximal triangle

Rank 2

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-47
SLIDE 47

One proof for a non-degenerate non-maximal triangle

Rank 3

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-48
SLIDE 48

One proof for a non-degenerate non-maximal triangle

Rank 3

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-49
SLIDE 49

One proof for a non-degenerate non-maximal triangle

Rank 3

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-50
SLIDE 50

One proof for a non-degenerate non-maximal triangle

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-51
SLIDE 51

One proof for a non-degenerate non-maximal triangle

Rank 4

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-52
SLIDE 52

One proof for a non-degenerate non-maximal triangle

Rank 4

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-53
SLIDE 53

One proof for a non-degenerate non-maximal triangle

Rank 4

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-54
SLIDE 54

One proof for a non-degenerate non-maximal triangle

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-55
SLIDE 55

One proof for a non-degenerate non-maximal triangle

Rank 5

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-56
SLIDE 56

One proof for a non-degenerate non-maximal triangle

Rank 5

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-57
SLIDE 57

One proof for a non-degenerate non-maximal triangle

Rank 5

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-58
SLIDE 58

One proof for a non-degenerate non-maximal triangle

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-59
SLIDE 59

One proof for a non-degenerate non-maximal triangle

The goal inequality has a rank of at most 6

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 11 / 1

slide-60
SLIDE 60

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-61
SLIDE 61

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-62
SLIDE 62

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-63
SLIDE 63

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-64
SLIDE 64

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-65
SLIDE 65

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-66
SLIDE 66

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-67
SLIDE 67

The geometry behind the convergence

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 12 / 1

slide-68
SLIDE 68

Assumptions for the following

We have “proven” that a non-maximal triangle where the upward ray points to the left has a finite rank. We can prove that the constructed bound is logarithmic in the number of bits of the input. The proof for the upward ray pointing to the right works similarly (but not identically). In the following, we assume that any non-maximal triangle has a finite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 13 / 1

slide-69
SLIDE 69

The maximal triangles

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 14 / 1

slide-70
SLIDE 70

The maximal triangles

This inequality is a non-maximal triangle ⇒ finite rank !

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 14 / 1

slide-71
SLIDE 71

The maximal triangles

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 14 / 1

slide-72
SLIDE 72

The maximal triangles

The goal inequality is valid for the disjunction.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 14 / 1

slide-73
SLIDE 73

The maximal triangles

The goal inequality has a finite rank

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 14 / 1

slide-74
SLIDE 74

The disection triangle

Disection ≡ each side is tight at exactly one integer point

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-75
SLIDE 75

The disection triangle

Disection ≡ each side is tight at exactly one integer point This inequality is a non-maximal triangle ⇒ finite rank !

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-76
SLIDE 76

The disection triangle

Disection ≡ each side is tight at exactly one integer point Similarly this inequality has a finite rank !

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-77
SLIDE 77

The disection triangle

Disection ≡ each side is tight at exactly one integer point

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-78
SLIDE 78

The disection triangle

Disection ≡ each side is tight at exactly one integer point Brown line : set of points with a representation that satisfy both inequalities with equality

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-79
SLIDE 79

The disection triangle

Disection ≡ each side is tight at exactly one integer point

  • Quentin Louveaux (Universit´

e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-80
SLIDE 80

The disection triangle

Disection ≡ each side is tight at exactly one integer point

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-81
SLIDE 81

The disection triangle

Disection ≡ each side is tight at exactly one integer point The disection cut has a finite rank

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 15 / 1

slide-82
SLIDE 82

The quadrilateral cuts

Two cases : non-maximal quadrilateral and disection quadrilateral. By the projection Lemma, we can deal with most non-maximal quadrilaterals One exception : if the lifted triangle has infinite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 16 / 1

slide-83
SLIDE 83

The quadrilateral cuts

Two cases : non-maximal quadrilateral and disection quadrilateral. By the projection Lemma, we can deal with most non-maximal quadrilaterals One exception : if the lifted triangle has infinite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 16 / 1

slide-84
SLIDE 84

The quadrilateral cuts

Two cases : non-maximal quadrilateral and disection quadrilateral. By the projection Lemma, we can deal with most non-maximal quadrilaterals One exception : if the lifted triangle has infinite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 16 / 1

slide-85
SLIDE 85

The quadrilateral cuts

Two cases : non-maximal quadrilateral and disection quadrilateral. By the projection Lemma, we can deal with most non-maximal quadrilaterals One exception : if the lifted triangle has infinite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 16 / 1

slide-86
SLIDE 86

The quadrilateral cuts

Two cases : non-maximal quadrilateral and disection quadrilateral. By the projection Lemma, we can deal with most non-maximal quadrilaterals One exception : if the lifted triangle has infinite rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 16 / 1

slide-87
SLIDE 87

Conclusion

All triangles except the Cook-Kannan-Schrijver have a finite rank. We provide a constructive split proof of that fact. Split cuts can essentially achieve all triangles in relatively few rounds. In constrast with the results of Basu et al. on the triangle closure compared to the split closure. Ongoing work : (almost ?) all quadrilaterals have a finite rank. Open (and difficult) question : lower bounds on the split rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 17 / 1

slide-88
SLIDE 88

Conclusion

All triangles except the Cook-Kannan-Schrijver have a finite rank. We provide a constructive split proof of that fact. Split cuts can essentially achieve all triangles in relatively few rounds. In constrast with the results of Basu et al. on the triangle closure compared to the split closure. Ongoing work : (almost ?) all quadrilaterals have a finite rank. Open (and difficult) question : lower bounds on the split rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 17 / 1

slide-89
SLIDE 89

Conclusion

All triangles except the Cook-Kannan-Schrijver have a finite rank. We provide a constructive split proof of that fact. Split cuts can essentially achieve all triangles in relatively few rounds. In constrast with the results of Basu et al. on the triangle closure compared to the split closure. Ongoing work : (almost ?) all quadrilaterals have a finite rank. Open (and difficult) question : lower bounds on the split rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 17 / 1

slide-90
SLIDE 90

Conclusion

All triangles except the Cook-Kannan-Schrijver have a finite rank. We provide a constructive split proof of that fact. Split cuts can essentially achieve all triangles in relatively few rounds. In constrast with the results of Basu et al. on the triangle closure compared to the split closure. Ongoing work : (almost ?) all quadrilaterals have a finite rank. Open (and difficult) question : lower bounds on the split rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 17 / 1

slide-91
SLIDE 91

Conclusion

All triangles except the Cook-Kannan-Schrijver have a finite rank. We provide a constructive split proof of that fact. Split cuts can essentially achieve all triangles in relatively few rounds. In constrast with the results of Basu et al. on the triangle closure compared to the split closure. Ongoing work : (almost ?) all quadrilaterals have a finite rank. Open (and difficult) question : lower bounds on the split rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 17 / 1

slide-92
SLIDE 92

Conclusion

All triangles except the Cook-Kannan-Schrijver have a finite rank. We provide a constructive split proof of that fact. Split cuts can essentially achieve all triangles in relatively few rounds. In constrast with the results of Basu et al. on the triangle closure compared to the split closure. Ongoing work : (almost ?) all quadrilaterals have a finite rank. Open (and difficult) question : lower bounds on the split rank.

Quentin Louveaux (Universit´ e de Li` ege - Montefiore Institute) Split Rank of Triangles and Quadrilaterals January 2009 17 / 1