universidade de coimbrajgouveia/core.pdf · universidade de coimbra 20 dec 2012 - core -...
TRANSCRIPT
![Page 1: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/1.jpg)
Lifts of convex sets and cone factorizations
João Gouveia
Universidade de Coimbra
20 Dec 2012 - CORE - Université Catholique de Louvain
with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)
![Page 2: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/2.jpg)
Lifts of Polytopes
What is a hard domain to do linear programming in?
Firstguess: a polytope with many vertices and facets.
However, polytopes with many facets can be projections ofmuch simpler polytopes. An example is the Parity Polytope:
PPn = conv({x ∈ {0,1}n : x has odd number of 1}).
For every even set A ⊆ {1, . . . ,n},∑i∈A
xi −∑i 6∈A
xi ≤ |A| − 1
is a facet, so we have at least 2n−1 facets.
![Page 3: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/3.jpg)
Lifts of Polytopes
What is a hard domain to do linear programming in? Firstguess: a polytope with many vertices and facets.
However, polytopes with many facets can be projections ofmuch simpler polytopes. An example is the Parity Polytope:
PPn = conv({x ∈ {0,1}n : x has odd number of 1}).
For every even set A ⊆ {1, . . . ,n},∑i∈A
xi −∑i 6∈A
xi ≤ |A| − 1
is a facet, so we have at least 2n−1 facets.
![Page 4: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/4.jpg)
Lifts of Polytopes
What is a hard domain to do linear programming in? Firstguess: a polytope with many vertices and facets.
However, polytopes with many facets can be projections ofmuch simpler polytopes.
An example is the Parity Polytope:
PPn = conv({x ∈ {0,1}n : x has odd number of 1}).
For every even set A ⊆ {1, . . . ,n},∑i∈A
xi −∑i 6∈A
xi ≤ |A| − 1
is a facet, so we have at least 2n−1 facets.
![Page 5: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/5.jpg)
Lifts of Polytopes
What is a hard domain to do linear programming in? Firstguess: a polytope with many vertices and facets.
However, polytopes with many facets can be projections ofmuch simpler polytopes. An example is the Parity Polytope:
PPn = conv({x ∈ {0,1}n : x has odd number of 1}).
For every even set A ⊆ {1, . . . ,n},∑i∈A
xi −∑i 6∈A
xi ≤ |A| − 1
is a facet, so we have at least 2n−1 facets.
![Page 6: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/6.jpg)
Lifts of Polytopes
What is a hard domain to do linear programming in? Firstguess: a polytope with many vertices and facets.
However, polytopes with many facets can be projections ofmuch simpler polytopes. An example is the Parity Polytope:
PPn = conv({x ∈ {0,1}n : x has odd number of 1}).
For every even set A ⊆ {1, . . . ,n},∑i∈A
xi −∑i 6∈A
xi ≤ |A| − 1
is a facet, so we have at least 2n−1 facets.
![Page 7: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/7.jpg)
Parity Polytope
There is a much shorter description.
PPn is the set of x ∈ Rn such that there exists for every odd1 ≤ k ≤ n a vector zk ∈ Rn and a real number αk such that
I∑
k zk = x;I∑
k αk = 1;I ‖ zk ‖1 = k αk ;I 0 ≤ ( zk )i ≤ αk .
O(n2) variables and O(n2) constraints.
![Page 8: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/8.jpg)
Parity Polytope
There is a much shorter description.
PPn is the set of x ∈ Rn such that there exists for every odd1 ≤ k ≤ n a vector zk ∈ Rn and a real number αk such that
I∑
k zk = x;I∑
k αk = 1;I ‖ zk ‖1 = k αk ;I 0 ≤ ( zk )i ≤ αk .
O(n2) variables and O(n2) constraints.
![Page 9: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/9.jpg)
Parity Polytope
There is a much shorter description.
PPn is the set of x ∈ Rn such that there exists for every odd1 ≤ k ≤ n a vector zk ∈ Rn and a real number αk such that
I∑
k zk = x;
I∑
k αk = 1;I ‖ zk ‖1 = k αk ;I 0 ≤ ( zk )i ≤ αk .
O(n2) variables and O(n2) constraints.
![Page 10: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/10.jpg)
Parity Polytope
There is a much shorter description.
PPn is the set of x ∈ Rn such that there exists for every odd1 ≤ k ≤ n a vector zk ∈ Rn and a real number αk such that
I∑
k zk = x;I∑
k αk = 1;
I ‖ zk ‖1 = k αk ;I 0 ≤ ( zk )i ≤ αk .
O(n2) variables and O(n2) constraints.
![Page 11: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/11.jpg)
Parity Polytope
There is a much shorter description.
PPn is the set of x ∈ Rn such that there exists for every odd1 ≤ k ≤ n a vector zk ∈ Rn and a real number αk such that
I∑
k zk = x;I∑
k αk = 1;I ‖ zk ‖1 = k αk ;
I 0 ≤ ( zk )i ≤ αk .
O(n2) variables and O(n2) constraints.
![Page 12: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/12.jpg)
Parity Polytope
There is a much shorter description.
PPn is the set of x ∈ Rn such that there exists for every odd1 ≤ k ≤ n a vector zk ∈ Rn and a real number αk such that
I∑
k zk = x;I∑
k αk = 1;I ‖ zk ‖1 = k αk ;I 0 ≤ ( zk )i ≤ αk .
O(n2) variables and O(n2) constraints.
![Page 13: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/13.jpg)
Parity Polytope
There is a much shorter description.
PPn is the set of x ∈ Rn such that there exists for every odd1 ≤ k ≤ n a vector zk ∈ Rn and a real number αk such that
I∑
k zk = x;I∑
k αk = 1;I ‖ zk ‖1 = k αk ;I 0 ≤ ( zk )i ≤ αk .
O(n2) variables and O(n2) constraints.
![Page 14: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/14.jpg)
Motivation
Polytopes with many facets can be projections of much simplerpolytopes.
Canonical LP LiftGiven a polytope P, a canonical LP lift is a description
P = Φ(Rk+ ∩ L)
for some affine space L and affine map Φ. We say it is a Rk+-lift.
The smallest k such that P has a Rk+-lift is a much better
measure of “LP-complexity” than number of facets and vertices.
![Page 15: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/15.jpg)
Motivation
Polytopes with many facets can be projections of much simplerpolytopes.
Canonical LP LiftGiven a polytope P, a canonical LP lift is a description
P = Φ(Rk+ ∩ L)
for some affine space L and affine map Φ. We say it is a Rk+-lift.
The smallest k such that P has a Rk+-lift is a much better
measure of “LP-complexity” than number of facets and vertices.
![Page 16: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/16.jpg)
Motivation
Polytopes with many facets can be projections of much simplerpolytopes.
Canonical LP LiftGiven a polytope P, a canonical LP lift is a description
P = Φ(Rk+ ∩ L)
for some affine space L and affine map Φ. We say it is a Rk+-lift.
The smallest k such that P has a Rk+-lift is a much better
measure of “LP-complexity” than number of facets and vertices.
![Page 17: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/17.jpg)
Two definitions
Let P be a polytope with facets defined byh1(x) ≥ 0, . . . ,hf (x) ≥ 0, and vertices p1, . . . ,pv .
Slack MatrixThe slack matrix of P is the matrix SP ∈ Rv×f defined by
SP(i , j) = hj(pi).
Nonnegative FactorizationGiven a nonnegative matrix M ∈ Rn×m
+ we say that it has ak -nonnegative factorization, or a Rk
+-factorization if there existmatrices A ∈ Rn×k
+ and B ∈ Rk×m+ such that
M = A · B.
![Page 18: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/18.jpg)
Two definitions
Let P be a polytope with facets defined byh1(x) ≥ 0, . . . ,hf (x) ≥ 0, and vertices p1, . . . ,pv .
Slack MatrixThe slack matrix of P is the matrix SP ∈ Rv×f defined by
SP(i , j) = hj(pi).
Nonnegative FactorizationGiven a nonnegative matrix M ∈ Rn×m
+ we say that it has ak -nonnegative factorization, or a Rk
+-factorization if there existmatrices A ∈ Rn×k
+ and B ∈ Rk×m+ such that
M = A · B.
![Page 19: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/19.jpg)
Two definitions
Let P be a polytope with facets defined byh1(x) ≥ 0, . . . ,hf (x) ≥ 0, and vertices p1, . . . ,pv .
Slack MatrixThe slack matrix of P is the matrix SP ∈ Rv×f defined by
SP(i , j) = hj(pi).
Nonnegative FactorizationGiven a nonnegative matrix M ∈ Rn×m
+ we say that it has ak -nonnegative factorization, or a Rk
+-factorization if there existmatrices A ∈ Rn×k
+ and B ∈ Rk×m+ such that
M = A · B.
![Page 20: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/20.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis 1991)A polytope P has a Rk
+-lift if and only if SP has aRk+-factorization.
Our questions:
I Does it work for other types of lifts?
I Does it work for other types of convex sets?
I Can we include symmetry in the result?
![Page 21: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/21.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis 1991)A polytope P has a Rk
+-lift if and only if SP has aRk+-factorization.
Our questions:
I Does it work for other types of lifts?
I Does it work for other types of convex sets?
I Can we include symmetry in the result?
![Page 22: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/22.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis 1991)A polytope P has a Rk
+-lift if and only if SP has aRk+-factorization.
Our questions:
I Does it work for other types of lifts?
I Does it work for other types of convex sets?
I Can we include symmetry in the result?
![Page 23: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/23.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis 1991)A polytope P has a Rk
+-lift if and only if SP has aRk+-factorization.
Our questions:
I Does it work for other types of lifts?
I Does it work for other types of convex sets?
I Can we include symmetry in the result?
![Page 24: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/24.jpg)
The Hexagon
Consider the regular hexagon.
It has a 6× 6 slack matrix SH .
0 0 1 2 2 11 0 0 1 2 22 1 0 0 1 22 2 1 0 0 11 2 2 1 0 00 1 2 2 1 0
=
1 0 1 0 0
1 0 0 0 1
0 0 0 1 2
0 1 0 0 1
0 1 1 0 0
0 0 2 1 0
0 0 0 1 2 1
1 2 1 0 0 0
0 0 1 1 0 0
0 1 0 0 1 0
1 0 0 0 0 1
![Page 25: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/25.jpg)
The Hexagon
Consider the regular hexagon.
It has a 6× 6 slack matrix SH .
0 0 1 2 2 11 0 0 1 2 22 1 0 0 1 22 2 1 0 0 11 2 2 1 0 00 1 2 2 1 0
=
1 0 1 0 0
1 0 0 0 1
0 0 0 1 2
0 1 0 0 1
0 1 1 0 0
0 0 2 1 0
0 0 0 1 2 1
1 2 1 0 0 0
0 0 1 1 0 0
0 1 0 0 1 0
1 0 0 0 0 1
![Page 26: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/26.jpg)
The Hexagon
Consider the regular hexagon.
It has a 6× 6 slack matrix SH .
0 0 1 2 2 11 0 0 1 2 22 1 0 0 1 22 2 1 0 0 11 2 2 1 0 00 1 2 2 1 0
=
1 0 1 0 0
1 0 0 0 1
0 0 0 1 2
0 1 0 0 1
0 1 1 0 0
0 0 2 1 0
0 0 0 1 2 1
1 2 1 0 0 0
0 0 1 1 0 0
0 1 0 0 1 0
1 0 0 0 0 1
![Page 27: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/27.jpg)
The Hexagon
Consider the regular hexagon.
It has a 6× 6 slack matrix SH .
0 0 1 2 2 11 0 0 1 2 22 1 0 0 1 22 2 1 0 0 11 2 2 1 0 00 1 2 2 1 0
=
1 0 1 0 0
1 0 0 0 1
0 0 0 1 2
0 1 0 0 1
0 1 1 0 0
0 0 2 1 0
0 0 0 1 2 1
1 2 1 0 0 0
0 0 1 1 0 0
0 1 0 0 1 0
1 0 0 0 0 1
![Page 28: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/28.jpg)
The Hexagon
Consider the regular hexagon.
It has a 6× 6 slack matrix SH .
0 0 1 2 2 11 0 0 1 2 22 1 0 0 1 22 2 1 0 0 11 2 2 1 0 00 1 2 2 1 0
=
1 0 1 0 0
1 0 0 0 1
0 0 0 1 2
0 1 0 0 1
0 1 1 0 0
0 0 2 1 0
0 0 0 1 2 1
1 2 1 0 0 0
0 0 1 1 0 0
0 1 0 0 1 0
1 0 0 0 0 1
![Page 29: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/29.jpg)
Hexagon - continued
It is the projection of the slice of R5+ cut out by
y1 + y2 + y3 + y5 = 2, y3 + y4 + y5 = 1.
For irregular hexagons a R6+-lift is the only we can have.
![Page 30: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/30.jpg)
Hexagon - continued
It is the projection of the slice of R5+ cut out by
y1 + y2 + y3 + y5 = 2, y3 + y4 + y5 = 1.
For irregular hexagons a R6+-lift is the only we can have.
![Page 31: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/31.jpg)
Hexagon - continued
It is the projection of the slice of R5+ cut out by
y1 + y2 + y3 + y5 = 2, y3 + y4 + y5 = 1.
For irregular hexagons a R6+-lift is the only we can have.
![Page 32: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/32.jpg)
Generalizing to non-LP
We want to generalize this result to other types of lifts.
K -LiftGiven a polytope P, and a closed convex cone K , a K -lift of Pis a description
P = Φ(K ∩ L)
for some affine space L and affine map Φ.
Important cases are Rn+, PSDn, SOCPn, CPn, CoPn,. . .
We also need to generalize the nonnegative factorizations.
![Page 33: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/33.jpg)
Generalizing to non-LP
We want to generalize this result to other types of lifts.
K -LiftGiven a polytope P, and a closed convex cone K , a K -lift of Pis a description
P = Φ(K ∩ L)
for some affine space L and affine map Φ.
Important cases are Rn+, PSDn, SOCPn, CPn, CoPn,. . .
We also need to generalize the nonnegative factorizations.
![Page 34: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/34.jpg)
Generalizing to non-LP
We want to generalize this result to other types of lifts.
K -LiftGiven a polytope P, and a closed convex cone K , a K -lift of Pis a description
P = Φ(K ∩ L)
for some affine space L and affine map Φ.
Important cases are Rn+, PSDn, SOCPn, CPn, CoPn,. . .
We also need to generalize the nonnegative factorizations.
![Page 35: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/35.jpg)
Generalizing to non-LP
We want to generalize this result to other types of lifts.
K -LiftGiven a polytope P, and a closed convex cone K , a K -lift of Pis a description
P = Φ(K ∩ L)
for some affine space L and affine map Φ.
Important cases are Rn+, PSDn, SOCPn, CPn, CoPn,. . .
We also need to generalize the nonnegative factorizations.
![Page 36: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/36.jpg)
K -factorizations
Recall that if K ⊆ Rl is a closed convex cone, K ∗ ⊆ Rl is itsdual cone, defined by
K ∗ = {y ∈ Rl 〈y , x〉 ≥ 0, ∀x ∈ K}.
K -FactorizationGiven a nonnegative matrix M ∈ Rn×m
+ we say that it has aK -factorization if there exist a1, . . .an ∈ K and b1, . . . ,bm ∈ K ∗
such thatMi,j =
⟨ai ,bj
⟩.
We can now generalize Yannakakis.
Theorem (G-Parrilo-Thomas)A polytope P has a K -lift if and only if SP has a K -factorization.
![Page 37: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/37.jpg)
K -factorizations
Recall that if K ⊆ Rl is a closed convex cone, K ∗ ⊆ Rl is itsdual cone, defined by
K ∗ = {y ∈ Rl 〈y , x〉 ≥ 0, ∀x ∈ K}.
K -FactorizationGiven a nonnegative matrix M ∈ Rn×m
+ we say that it has aK -factorization if there exist a1, . . .an ∈ K and b1, . . . ,bm ∈ K ∗
such thatMi,j =
⟨ai ,bj
⟩.
We can now generalize Yannakakis.
Theorem (G-Parrilo-Thomas)A polytope P has a K -lift if and only if SP has a K -factorization.
![Page 38: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/38.jpg)
K -factorizations
Recall that if K ⊆ Rl is a closed convex cone, K ∗ ⊆ Rl is itsdual cone, defined by
K ∗ = {y ∈ Rl 〈y , x〉 ≥ 0, ∀x ∈ K}.
K -FactorizationGiven a nonnegative matrix M ∈ Rn×m
+ we say that it has aK -factorization if there exist a1, . . .an ∈ K and b1, . . . ,bm ∈ K ∗
such thatMi,j =
⟨ai ,bj
⟩.
We can now generalize Yannakakis.
Theorem (G-Parrilo-Thomas)A polytope P has a K -lift if and only if SP has a K -factorization.
![Page 39: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/39.jpg)
The Square
The 0/1 square is the projectiononto x and y of the slice of PSD3given by 1 x y
x x zy z y
� 0.
Its slack matrix is given by
SP =
0 0 1 10 1 1 01 1 0 01 0 0 1
,and should factorize in PSD3.
![Page 40: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/40.jpg)
The Square
The 0/1 square is the projectiononto x and y of the slice of PSD3given by 1 x y
x x zy z y
� 0.
Its slack matrix is given by
SP =
0 0 1 10 1 1 01 1 0 01 0 0 1
,and should factorize in PSD3.
![Page 41: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/41.jpg)
The Square
The 0/1 square is the projectiononto x and y of the slice of PSD3given by 1 x y
x x zy z y
� 0.
Its slack matrix is given by
SP =
0 0 1 10 1 1 01 1 0 01 0 0 1
,and should factorize in PSD3.
![Page 42: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/42.jpg)
Square - continued
SP =
0 0 1 10 1 1 01 1 0 01 0 0 1
,is factorized by
0 0 00 1 00 0 0
,
0 0 00 0 00 0 1
,
1 −1 0−1 1 00 0 0
,
1 0 −10 0 0−1 0 1
,
for the rows and 1 0 00 0 00 0 0
,
1 0 10 0 01 0 1
,
1 1 11 1 11 1 1
,
1 1 01 1 00 0 0
,for the columns.
![Page 43: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/43.jpg)
Slack Operator
To further generalize Yannakakis to other convex sets, we haveto introduce a slack operator.
Given a convex set C ⊆ Rn, consider its polar set
C◦ = {x ∈ Rn : 〈x , y〉 ≤ 1, ∀y ∈ C},
and define the slack operator SC : ext(C)× ext(C◦)→ R+ as
SC(x , y) = 1− 〈x , y〉 .
Note that this generalizes the slack matrix.
![Page 44: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/44.jpg)
Slack Operator
To further generalize Yannakakis to other convex sets, we haveto introduce a slack operator.
Given a convex set C ⊆ Rn, consider its polar set
C◦ = {x ∈ Rn : 〈x , y〉 ≤ 1, ∀y ∈ C},
and define the slack operator SC : ext(C)× ext(C◦)→ R+ as
SC(x , y) = 1− 〈x , y〉 .
Note that this generalizes the slack matrix.
![Page 45: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/45.jpg)
Slack Operator
To further generalize Yannakakis to other convex sets, we haveto introduce a slack operator.
Given a convex set C ⊆ Rn, consider its polar set
C◦ = {x ∈ Rn : 〈x , y〉 ≤ 1, ∀y ∈ C},
and define the slack operator SC : ext(C)× ext(C◦)→ R+ as
SC(x , y) = 1− 〈x , y〉 .
Note that this generalizes the slack matrix.
![Page 46: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/46.jpg)
Slack Operator
To further generalize Yannakakis to other convex sets, we haveto introduce a slack operator.
Given a convex set C ⊆ Rn, consider its polar set
C◦ = {x ∈ Rn : 〈x , y〉 ≤ 1, ∀y ∈ C},
and define the slack operator SC : ext(C)× ext(C◦)→ R+ as
SC(x , y) = 1− 〈x , y〉 .
Note that this generalizes the slack matrix.
![Page 47: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/47.jpg)
Generalized Yannakakis for convex sets
We can then define a K -factorization of SC as a pair of maps
A : ext(C)→ K B : ext(C◦)→ K ∗
such that〈A(x),B(y)〉 = SC(x , y)
for all x , y .
Theorem (G-Parrilo-Thomas)A convex set C has a K -lift if and only if SC has aK -factorization.
![Page 48: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/48.jpg)
Generalized Yannakakis for convex sets
We can then define a K -factorization of SC as a pair of maps
A : ext(C)→ K B : ext(C◦)→ K ∗
such that〈A(x),B(y)〉 = SC(x , y)
for all x , y .
Theorem (G-Parrilo-Thomas)A convex set C has a K -lift if and only if SC has aK -factorization.
![Page 49: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/49.jpg)
The Disk
The unit disk D is the projection onto xand y of the slice of PSD2 given by[
1 + x yy 1− x
]� 0.
D◦ = D, there must be A : S1 → PSD2 and B : S1 → PSD2such that 〈A(x),B(y)〉 = 1− 〈x , y〉
A(x , y) =
[1 + x y
y 1− x
], B(x , y) =
[1− x −y−y 1 + x
].
![Page 50: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/50.jpg)
The Disk
The unit disk D is the projection onto xand y of the slice of PSD2 given by[
1 + x yy 1− x
]� 0.
D◦ = D, there must be A : S1 → PSD2 and B : S1 → PSD2such that 〈A(x),B(y)〉 = 1− 〈x , y〉
A(x , y) =
[1 + x y
y 1− x
], B(x , y) =
[1− x −y−y 1 + x
].
![Page 51: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/51.jpg)
The Disk
The unit disk D is the projection onto xand y of the slice of PSD2 given by[
1 + x yy 1− x
]� 0.
D◦ = D, there must be A : S1 → PSD2 and B : S1 → PSD2such that 〈A(x),B(y)〉 = 1− 〈x , y〉
A(x , y) =
[1 + x y
y 1− x
], B(x , y) =
[1− x −y−y 1 + x
].
![Page 52: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/52.jpg)
Cone ranks of polytopes
Recall - rank+(M) is the smallest k such that M has anRk+-factorization. rank+(P) := rank+(SP)
Given K = {K1,K2, · · · }, (e.g. Rk+, PSDk , CPk , CoPk ,. . . )
rankK(M) is the smallest i such that M has a Ki -factorization.Again rankK(P) := rankK(SP).
We are specially interested in rankpsd(M).
![Page 53: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/53.jpg)
Cone ranks of polytopes
Recall - rank+(M) is the smallest k such that M has anRk+-factorization. rank+(P) := rank+(SP)
Given K = {K1,K2, · · · }, (e.g. Rk+, PSDk , CPk , CoPk ,. . . )
rankK(M) is the smallest i such that M has a Ki -factorization.
Again rankK(P) := rankK(SP).
We are specially interested in rankpsd(M).
![Page 54: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/54.jpg)
Cone ranks of polytopes
Recall - rank+(M) is the smallest k such that M has anRk+-factorization. rank+(P) := rank+(SP)
Given K = {K1,K2, · · · }, (e.g. Rk+, PSDk , CPk , CoPk ,. . . )
rankK(M) is the smallest i such that M has a Ki -factorization.Again rankK(P) := rankK(SP).
We are specially interested in rankpsd(M).
![Page 55: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/55.jpg)
Cone ranks of polytopes
Recall - rank+(M) is the smallest k such that M has anRk+-factorization. rank+(P) := rank+(SP)
Given K = {K1,K2, · · · }, (e.g. Rk+, PSDk , CPk , CoPk ,. . . )
rankK(M) is the smallest i such that M has a Ki -factorization.Again rankK(P) := rankK(SP).
We are specially interested in rankpsd(M).
![Page 56: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/56.jpg)
Bounds for matrices
For M ∈ Rp×q+ .
I rank(M) ≤ rank+(M) ≤ min{p,q}.
I rank(M) ≤(rankpsd(M)+1
2
).
I rankpsd(M) ≤ rank+(M).
PropositionIf M ∈ Rn×n
+ is zero on the diagonal and positive everywhereelse then rank+(M) ≥ k, where k is the smallest integer suchthat n ≤
( kbk/2c
).
![Page 57: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/57.jpg)
Bounds for matrices
For M ∈ Rp×q+ .
I rank(M) ≤ rank+(M) ≤ min{p,q}.I rank(M) ≤
(rankpsd(M)+12
).
I rankpsd(M) ≤ rank+(M).
PropositionIf M ∈ Rn×n
+ is zero on the diagonal and positive everywhereelse then rank+(M) ≥ k, where k is the smallest integer suchthat n ≤
( kbk/2c
).
![Page 58: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/58.jpg)
Bounds for matrices
For M ∈ Rp×q+ .
I rank(M) ≤ rank+(M) ≤ min{p,q}.I rank(M) ≤
(rankpsd(M)+12
).
I rankpsd(M) ≤ rank+(M).
PropositionIf M ∈ Rn×n
+ is zero on the diagonal and positive everywhereelse then rank+(M) ≥ k, where k is the smallest integer suchthat n ≤
( kbk/2c
).
![Page 59: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/59.jpg)
Bounds for matrices
For M ∈ Rp×q+ .
I rank(M) ≤ rank+(M) ≤ min{p,q}.I rank(M) ≤
(rankpsd(M)+12
).
I rankpsd(M) ≤ rank+(M).
PropositionIf M ∈ Rn×n
+ is zero on the diagonal and positive everywhereelse then rank+(M) ≥ k, where k is the smallest integer suchthat n ≤
( kbk/2c
).
![Page 60: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/60.jpg)
Bounds for matrices - 2
PropositionIf M ∈ Rp×q has rank k, then the matrix M ′ obtained bysquaring each entry of M has psd-rank at most k.
Consider the matrix A ∈ Rn×n defined by ai,j = (i − j)2.
I rank(A) = 3;
I rankpsd(A) = 2;
I rank+(A) ≥ log2(n) grows with n.
rank+ can be arbitrarily larger than rank and rankpsd.
![Page 61: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/61.jpg)
Bounds for matrices - 2
PropositionIf M ∈ Rp×q has rank k, then the matrix M ′ obtained bysquaring each entry of M has psd-rank at most k.
Consider the matrix A ∈ Rn×n defined by ai,j = (i − j)2.
I rank(A) = 3;
I rankpsd(A) = 2;
I rank+(A) ≥ log2(n) grows with n.
rank+ can be arbitrarily larger than rank and rankpsd.
![Page 62: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/62.jpg)
Bounds for matrices - 2
PropositionIf M ∈ Rp×q has rank k, then the matrix M ′ obtained bysquaring each entry of M has psd-rank at most k.
Consider the matrix A ∈ Rn×n defined by ai,j = (i − j)2.
I rank(A) = 3;
I rankpsd(A) = 2;
I rank+(A) ≥ log2(n) grows with n.
rank+ can be arbitrarily larger than rank and rankpsd.
![Page 63: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/63.jpg)
Bounds for matrices - 2
PropositionIf M ∈ Rp×q has rank k, then the matrix M ′ obtained bysquaring each entry of M has psd-rank at most k.
Consider the matrix A ∈ Rn×n defined by ai,j = (i − j)2.
I rank(A) = 3;
I rankpsd(A) = 2;
I rank+(A) ≥ log2(n) grows with n.
rank+ can be arbitrarily larger than rank and rankpsd.
![Page 64: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/64.jpg)
Bounds for matrices - 2
PropositionIf M ∈ Rp×q has rank k, then the matrix M ′ obtained bysquaring each entry of M has psd-rank at most k.
Consider the matrix A ∈ Rn×n defined by ai,j = (i − j)2.
I rank(A) = 3;
I rankpsd(A) = 2;
I rank+(A) ≥ log2(n) grows with n.
rank+ can be arbitrarily larger than rank and rankpsd.
![Page 65: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/65.jpg)
Bounds for polytopes - LP
PropositionAn Rk
+-lift of P induces an embedding from the lattice of facesof P, L(P), to the boolean lattice 2[k ]. In particular:
I If p is the size of the largest antichain in L(P), thenrank+(P) ≤ k where k is the smallest integer such thatp ≤
( kbk/2c
).
I [Goemans] If nP is the number of faces of P,rank+(P) ≥ log2(nP).
P = 3-cube: rank+(P) ≤ 6.I nP = 28⇒ rank+(P) ≥ log2(28) ≈ 4.807.I nedges = 12,
(52
)= 10,
(63
)= 20, hence rank+(P) ≥ 6.
![Page 66: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/66.jpg)
Bounds for polytopes - LP
PropositionAn Rk
+-lift of P induces an embedding from the lattice of facesof P, L(P), to the boolean lattice 2[k ]. In particular:
I If p is the size of the largest antichain in L(P), thenrank+(P) ≤ k where k is the smallest integer such thatp ≤
( kbk/2c
).
I [Goemans] If nP is the number of faces of P,rank+(P) ≥ log2(nP).
P = 3-cube: rank+(P) ≤ 6.I nP = 28⇒ rank+(P) ≥ log2(28) ≈ 4.807.I nedges = 12,
(52
)= 10,
(63
)= 20, hence rank+(P) ≥ 6.
![Page 67: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/67.jpg)
Bounds for polytopes - LP
PropositionAn Rk
+-lift of P induces an embedding from the lattice of facesof P, L(P), to the boolean lattice 2[k ]. In particular:
I If p is the size of the largest antichain in L(P), thenrank+(P) ≤ k where k is the smallest integer such thatp ≤
( kbk/2c
).
I [Goemans] If nP is the number of faces of P,rank+(P) ≥ log2(nP).
P = 3-cube: rank+(P) ≤ 6.I nP = 28⇒ rank+(P) ≥ log2(28) ≈ 4.807.I nedges = 12,
(52
)= 10,
(63
)= 20, hence rank+(P) ≥ 6.
![Page 68: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/68.jpg)
Bounds for polytopes - LP
PropositionAn Rk
+-lift of P induces an embedding from the lattice of facesof P, L(P), to the boolean lattice 2[k ]. In particular:
I If p is the size of the largest antichain in L(P), thenrank+(P) ≤ k where k is the smallest integer such thatp ≤
( kbk/2c
).
I [Goemans] If nP is the number of faces of P,rank+(P) ≥ log2(nP).
P = 3-cube: rank+(P) ≤ 6.
I nP = 28⇒ rank+(P) ≥ log2(28) ≈ 4.807.I nedges = 12,
(52
)= 10,
(63
)= 20, hence rank+(P) ≥ 6.
![Page 69: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/69.jpg)
Bounds for polytopes - LP
PropositionAn Rk
+-lift of P induces an embedding from the lattice of facesof P, L(P), to the boolean lattice 2[k ]. In particular:
I If p is the size of the largest antichain in L(P), thenrank+(P) ≤ k where k is the smallest integer such thatp ≤
( kbk/2c
).
I [Goemans] If nP is the number of faces of P,rank+(P) ≥ log2(nP).
P = 3-cube: rank+(P) ≤ 6.I nP = 28⇒ rank+(P) ≥ log2(28) ≈ 4.807.
I nedges = 12,(5
2
)= 10,
(63
)= 20, hence rank+(P) ≥ 6.
![Page 70: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/70.jpg)
Bounds for polytopes - LP
PropositionAn Rk
+-lift of P induces an embedding from the lattice of facesof P, L(P), to the boolean lattice 2[k ]. In particular:
I If p is the size of the largest antichain in L(P), thenrank+(P) ≤ k where k is the smallest integer such thatp ≤
( kbk/2c
).
I [Goemans] If nP is the number of faces of P,rank+(P) ≥ log2(nP).
P = 3-cube: rank+(P) ≤ 6.I nP = 28⇒ rank+(P) ≥ log2(28) ≈ 4.807.I nedges = 12,
(52
)= 10,
(63
)= 20, hence rank+(P) ≥ 6.
![Page 71: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/71.jpg)
Bounds for polytopes - SDP
TheoremIf a polytope P in Rn has rankpsd = k than it has at most kO(k2n)
facets.
For Pn = n-gon, rank+(Pn) and rankpsd(Pn) grow to infinity as ngrows, despite rank(SPn ) = 3.
Open questions:I Can we find a separation between rankpsd and rank+ for
polytopes?I Recently, [Fiorini-Massar-Pokutta-Tiwary-de Wolf] proved
rank+(TSP) grows exponentially. What about rankpsd?
![Page 72: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/72.jpg)
Bounds for polytopes - SDP
TheoremIf a polytope P in Rn has rankpsd = k than it has at most kO(k2n)
facets.
For Pn = n-gon, rank+(Pn) and rankpsd(Pn) grow to infinity as ngrows, despite rank(SPn ) = 3.
Open questions:
I Can we find a separation between rankpsd and rank+ forpolytopes?
I Recently, [Fiorini-Massar-Pokutta-Tiwary-de Wolf] provedrank+(TSP) grows exponentially. What about rankpsd?
![Page 73: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/73.jpg)
Bounds for polytopes - SDP
TheoremIf a polytope P in Rn has rankpsd = k than it has at most kO(k2n)
facets.
For Pn = n-gon, rank+(Pn) and rankpsd(Pn) grow to infinity as ngrows, despite rank(SPn ) = 3.
Open questions:I Can we find a separation between rankpsd and rank+ for
polytopes?
I Recently, [Fiorini-Massar-Pokutta-Tiwary-de Wolf] provedrank+(TSP) grows exponentially. What about rankpsd?
![Page 74: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/74.jpg)
Bounds for polytopes - SDP
TheoremIf a polytope P in Rn has rankpsd = k than it has at most kO(k2n)
facets.
For Pn = n-gon, rank+(Pn) and rankpsd(Pn) grow to infinity as ngrows, despite rank(SPn ) = 3.
Open questions:I Can we find a separation between rankpsd and rank+ for
polytopes?I Recently, [Fiorini-Massar-Pokutta-Tiwary-de Wolf] proved
rank+(TSP) grows exponentially. What about rankpsd?
![Page 75: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/75.jpg)
Symmetric Lifts
In the LP case there has been much interest in symmetric lifts.[Kaibel-Pashkovich-Theis]
Symmetric liftsLet P be a polytope and P = Φ(K ∩ L) a lift of P. We say the liftis symmetric if there exists a group homomorphism sending g∈ Aut(P) to ψg ∈ Aut(K ) such that ψg(L) = L and Φ ◦ ψg = g.
Symmetric lift preserves symmetries of the lifted objects.
Common lift-and-project methods are symmetric (w.r.t.permutation of variables): LS, SA, Las...
![Page 76: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/76.jpg)
Symmetric Lifts
In the LP case there has been much interest in symmetric lifts.[Kaibel-Pashkovich-Theis]
Symmetric liftsLet P be a polytope and P = Φ(K ∩ L) a lift of P. We say the liftis symmetric if there exists a group homomorphism sending g∈ Aut(P) to ψg ∈ Aut(K ) such that ψg(L) = L and Φ ◦ ψg = g.
Symmetric lift preserves symmetries of the lifted objects.
Common lift-and-project methods are symmetric (w.r.t.permutation of variables): LS, SA, Las...
![Page 77: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/77.jpg)
Example:The square
Recall the lift of the 0/1 square[1 x yx x zy z y
]� 0.
Aut(g) = 〈g(x , y) = (y , x),h(x , y) = (1− x , y)〉.
φg(A) =[
1 y xy y zx z x
]= P23AP23,
φh(A) =[
1 1 − x y1 − x 1 − x y − z
y y − z y
]=[
1 0 01 −1 00 0 1
]A[
1 1 00 −1 00 0 1
],
generate a homomorphism so the lift is symmetric.
![Page 78: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/78.jpg)
Example:The square
Recall the lift of the 0/1 square[1 x yx x zy z y
]� 0.
Aut(g) = 〈g(x , y) = (y , x),h(x , y) = (1− x , y)〉.
φg(A) =[
1 y xy y zx z x
]= P23AP23,
φh(A) =[
1 1 − x y1 − x 1 − x y − z
y y − z y
]=[
1 0 01 −1 00 0 1
]A[
1 1 00 −1 00 0 1
],
generate a homomorphism so the lift is symmetric.
![Page 79: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/79.jpg)
Example:The square
Recall the lift of the 0/1 square[1 x yx x zy z y
]� 0.
Aut(g) = 〈g(x , y) = (y , x),h(x , y) = (1− x , y)〉.
φg(A) =[
1 y xy y zx z x
]= P23AP23,
φh(A) =[
1 1 − x y1 − x 1 − x y − z
y y − z y
]=[
1 0 01 −1 00 0 1
]A[
1 1 00 −1 00 0 1
],
generate a homomorphism so the lift is symmetric.
![Page 80: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/80.jpg)
Example 2 : Regular n-gons
PropositionFor p prime the smallest k for which there exists a symmetricRk+-lift of the p-gon is p.
Note that we know that there are actually O(log(n))dimensional lifts of these polytopes [Ben-Tal, Nemirovski].
Open (small) problem: prove that the smallest symmetric lift ofan n-gon is to Rn
+.
![Page 81: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/81.jpg)
Example 2 : Regular n-gons
PropositionFor p prime the smallest k for which there exists a symmetricRk+-lift of the p-gon is p.
Note that we know that there are actually O(log(n))dimensional lifts of these polytopes [Ben-Tal, Nemirovski].
Open (small) problem: prove that the smallest symmetric lift ofan n-gon is to Rn
+.
![Page 82: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/82.jpg)
Example 2 : Regular n-gons
PropositionFor p prime the smallest k for which there exists a symmetricRk+-lift of the p-gon is p.
Note that we know that there are actually O(log(n))dimensional lifts of these polytopes [Ben-Tal, Nemirovski].
Open (small) problem: prove that the smallest symmetric lift ofan n-gon is to Rn
+.
![Page 83: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/83.jpg)
Symmetric Yannakakis
K -FactorizationGiven a polytope P and its slack matrix SP ∈ Rn×m
+ and itsK -factorization given by a1, . . .an ∈ K , b1, . . . ,bm ∈ K ∗, we saythat it is symmetric if there is an homomorphismφ : Aut(P)→ Aut(K ) such that if g send the i-th vertex to thej-th vertex, φ(ai) = aj .
Theorem (G-Parrilo-Thomas)A convex set C has a symmetric K -lift if and only if SC has asymmetric K -factorization.
![Page 84: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/84.jpg)
Symmetric Yannakakis
K -FactorizationGiven a polytope P and its slack matrix SP ∈ Rn×m
+ and itsK -factorization given by a1, . . .an ∈ K , b1, . . . ,bm ∈ K ∗, we saythat it is symmetric if there is an homomorphismφ : Aut(P)→ Aut(K ) such that if g send the i-th vertex to thej-th vertex, φ(ai) = aj .
Theorem (G-Parrilo-Thomas)A convex set C has a symmetric K -lift if and only if SC has asymmetric K -factorization.
![Page 85: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/85.jpg)
Matchings
MatchingsGiven the complete graph Kn = ([n],En) a matching is acollection M of edges such that there’s one and only one edgeincident to each vertex.
χM ∈ {0,1}En is the indicator vector of M.For this example χM = (0,0,0,0,1,1).
![Page 86: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/86.jpg)
Matchings
MatchingsGiven the complete graph Kn = ([n],En) a matching is acollection M of edges such that there’s one and only one edgeincident to each vertex.
χM ∈ {0,1}En is the indicator vector of M.For this example χM = (0,0,0,0,1,1).
![Page 87: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/87.jpg)
Matchings
MatchingsGiven the complete graph Kn = ([n],En) a matching is acollection M of edges such that there’s one and only one edgeincident to each vertex.
χM ∈ {0,1}En is the indicator vector of M.For this example χM = (0,0,0,0,1,1).
![Page 88: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/88.jpg)
Matchings
MatchingsGiven the complete graph Kn = ([n],En) a matching is acollection M of edges such that there’s one and only one edgeincident to each vertex.
χM ∈ {0,1}En is the indicator vector of M.For this example χM = (0,0,0,0,1,1).
![Page 89: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/89.jpg)
The matching Polytope
MaxMatchGiven a complete graph K2n with edge weights ω : En → R, findthe matching with maximum weight.
This has a geometrical version.
MaxMatchMaximize 〈ω, x〉 over the polytope
conv({χM : M is a matching}).
This polytope is the Matching Polytope, denoted PMatch2n.
![Page 90: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/90.jpg)
The matching Polytope
MaxMatchGiven a complete graph K2n with edge weights ω : En → R, findthe matching with maximum weight.This has a geometrical version.
MaxMatchMaximize 〈ω, x〉 over the polytope
conv({χM : M is a matching}).
This polytope is the Matching Polytope, denoted PMatch2n.
![Page 91: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/91.jpg)
The matching Polytope
MaxMatchGiven a complete graph K2n with edge weights ω : En → R, findthe matching with maximum weight.This has a geometrical version.
MaxMatchMaximize 〈ω, x〉 over the polytope
conv({χM : M is a matching}).
This polytope is the Matching Polytope, denoted PMatch2n.
![Page 92: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/92.jpg)
Symmetric lifts of matching polytope
YannakakisAlthough the max-matching problem is polynomial timesolvable, there is no polynomial size linear symmetric lift for thematching polytope.
What about non-symmetric?
With other versions of the matching polytope, Kaibel,Pashkovich and Theis show that symmetry does matter, but thegeneral question is still open.
![Page 93: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/93.jpg)
Symmetric lifts of matching polytope
YannakakisAlthough the max-matching problem is polynomial timesolvable, there is no polynomial size linear symmetric lift for thematching polytope.
What about non-symmetric?
With other versions of the matching polytope, Kaibel,Pashkovich and Theis show that symmetry does matter, but thegeneral question is still open.
![Page 94: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/94.jpg)
Symmetric lifts of matching polytope
YannakakisAlthough the max-matching problem is polynomial timesolvable, there is no polynomial size linear symmetric lift for thematching polytope.
What about non-symmetric?
With other versions of the matching polytope, Kaibel,Pashkovich and Theis show that symmetry does matter, but thegeneral question is still open.
![Page 95: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/95.jpg)
Symmetry in SDP
Further thoughts:
I SDP lift-and-project algorithms don’t work polynomially formatchings [Tuncel].
I Do all polynomial sized symmetric SDP lifts fail?
I What about non-symmetric?
![Page 96: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/96.jpg)
Symmetry in SDP
Further thoughts:
I SDP lift-and-project algorithms don’t work polynomially formatchings [Tuncel].
I Do all polynomial sized symmetric SDP lifts fail?
I What about non-symmetric?
![Page 97: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/97.jpg)
Symmetry in SDP
Further thoughts:
I SDP lift-and-project algorithms don’t work polynomially formatchings [Tuncel].
I Do all polynomial sized symmetric SDP lifts fail?
I What about non-symmetric?
![Page 98: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/98.jpg)
Symmetry in SDP
Further thoughts:
I SDP lift-and-project algorithms don’t work polynomially formatchings [Tuncel].
I Do all polynomial sized symmetric SDP lifts fail?
I What about non-symmetric?
![Page 99: Universidade de Coimbrajgouveia/CORE.pdf · Universidade de Coimbra 20 Dec 2012 - CORE - Université Catholique de Louvain with Pablo Parrilo (MIT) and Rekha Thomas (U.Washington)](https://reader033.vdocument.in/reader033/viewer/2022050308/5f7081bed6f18468320f4dfc/html5/thumbnails/99.jpg)
The end
Thank You