Download - Extended Formulations I (BootCamp)
![Page 1: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/1.jpg)
Extended Formulations I
(Boot Camp)
Thomas Rothvoss
UW Seattle
![Page 2: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/2.jpg)
Organization
1. Thomas Rothvoss (1.x lectures): Introduction to LPExtended Formulations
2. Hamza Fawzi (1.x lectures): Introduction to SDP ExtendedFormulations
3. Prasad Raghavendra (1.x lectures): Lower bounds forLP/SDP lifts
![Page 3: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/3.jpg)
Extended formulation
![Page 4: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/4.jpg)
Extended formulation
Given polytope P = x ∈ Rn | Ax ≤ b
P
![Page 5: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/5.jpg)
Extended formulation
Given polytope P = x ∈ Rn | Ax ≤ b
Write P = x ∈ Rn | ∃y : Bx+ Cy ≤ d
P
Q
linearprojection
![Page 6: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/6.jpg)
Extended formulation
Given polytope P = x ∈ Rn | Ax ≤ b→ many inequalities
Write P = x ∈ Rn | ∃y : Bx+ Cy ≤ d→ few inequalities
P
Q
linearprojection
![Page 7: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/7.jpg)
Extended formulation
Given polytope P = x ∈ Rn | Ax ≤ b→ many inequalities
Write P = x ∈ Rn | ∃y : Bx+ Cy ≤ d→ few inequalities
P
Q
linearprojection
The extension complexity of P is
xc(P ) := min
#facets of Q |
Q polyhedronp linear mapp(Q) = P
![Page 8: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/8.jpg)
Motivation
![Page 9: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/9.jpg)
Motivation
xc(conv(Spanning trees)) ≤ O(n3)⇒ optimize over all Spanning trees with LP of size O(n3)
P
Q
linearprojection
![Page 10: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/10.jpg)
Motivation
xc(conv(Spanning trees)) ≤ O(n3)⇒ optimize over all Spanning trees with LP of size O(n3)
P
Q
linearprojection
In reverse: If xc(P ) is high forTSP / MaxCut / Correlation / Matchings, then thoseproblems cannot be solved with a single poly-size LP!
![Page 11: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/11.jpg)
Part I
A non trivial example -
Knapsack
![Page 12: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/12.jpg)
Knapsack
Input: n objects with weight wi ∈ Z+
profit pi ∈ Q+
knapsack size B ∈ Q+
Goal: Find subset of objects, maximizing the profit andnot exceeding the weight bound:
OPT = maxI⊆1,...,n
∑
i∈I
pi |∑
i∈I
wi ≤ B
![Page 13: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/13.jpg)
Knapsack
Input: n objects with weight wi ∈ Z+
profit pi ∈ Q+
knapsack size B ∈ Q+
Goal: Find subset of objects, maximizing the profit andnot exceeding the weight bound:
OPT = maxI⊆1,...,n
∑
i∈I
pi |∑
i∈I
wi ≤ B
Known:
weakly NP-hard
Pseudo-polynomial time algorithm
FPTAS
![Page 14: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/14.jpg)
Knapsack
Input: n objects with weight wi ∈ Z+
profit pi ∈ Q+
knapsack size B ∈ Q+
Goal: Find subset of objects, maximizing the profit andnot exceeding the weight bound:
OPT = maxI⊆1,...,n
∑
i∈I
pi |∑
i∈I
wi ≤ B
Known:
weakly NP-hard
Pseudo-polynomial time algorithm
FPTAS
xc(convx ∈ 0, 1n |∑n
i=1wixi ≤ B) ≤ O(n ·B).
![Page 15: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/15.jpg)
A dynamic program for Knapsack
Lemma
Knapsack can be solved in time O(n ·B).
Algorithm
(1) Compute table entries
T (i,W ) = maxI⊆1,...,i
∑
j∈I
pj |∑
j∈I
wj = W
= max. profit of weight W subsets of first i items
using dynamic programming
T (i,W ) = maxT (i− 1,W )︸ ︷︷ ︸
don’t take i
, T (i− 1,W − wi) + pi︸ ︷︷ ︸
take i
∀i ∀W ≤ B
![Page 16: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/16.jpg)
Knapsack (3)
T (0, ∗) T (1, ∗) T (2, ∗) T (3, ∗)
T (∗, 0)
T (∗, 1)
T (∗, 2)
T (∗, 3) Instance:
w1 = 2
w2 = 1
w3 = 2
Create a network with nodes (i,W ) and
![Page 17: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/17.jpg)
Knapsack (3)
T (0, ∗) T (1, ∗) T (2, ∗) T (3, ∗)
T (∗, 0)
T (∗, 1)
T (∗, 2)
T (∗, 3) Instance:
w1 = 2
w2 = 1
w3 = 2
Create a network with nodes (i,W ) and “Don’t take i”-edge from (i− 1,W )→ (i,W ) at cost 0
![Page 18: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/18.jpg)
Knapsack (3)
T (0, ∗) T (1, ∗) T (2, ∗) T (3, ∗)
T (∗, 0)
T (∗, 1)
T (∗, 2)
T (∗, 3) Instance:
w1 = 2
w2 = 1
w3 = 2
Create a network with nodes (i,W ) and “Don’t take i”-edge from (i− 1,W )→ (i,W ) at cost 0 “Take i”-edge from (i− 1,W − wi)→ (i,W ) at cost pi
![Page 19: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/19.jpg)
Knapsack (3)
s t
1 2objects
3
1
B
weigh
ts
Instance:
w1 = 2
w2 = 1
w3 = 2
Create a network with nodes (i,W ) and “Don’t take i”-edge from (i− 1,W )→ (i,W ) at cost 0 “Take i”-edge from (i− 1,W − wi)→ (i,W ) at cost pi
![Page 20: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/20.jpg)
Knapsack (3)
s t
1 2objects
3
1
B
weigh
ts
Instance:
w1 = 2
w2 = 1
w3 = 2
Create a network with nodes (i,W ) and “Don’t take i”-edge from (i− 1,W )→ (i,W ) at cost 0 “Take i”-edge from (i− 1,W − wi)→ (i,W ) at cost pi
Observations:
s-t path ←→ Knapsack solution
![Page 21: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/21.jpg)
Knapsack (3)
s t
1 2objects
3
1
B
weigh
ts
Instance:
w1 = 2
w2 = 1
w3 = 2
Create a network with nodes (i,W ) and “Don’t take i”-edge from (i− 1,W )→ (i,W ) at cost 0 “Take i”-edge from (i− 1,W − wi)→ (i,W ) at cost pi
Observations:
s-t path ←→ Knapsack solution
max cost s-t path = max profit packing
![Page 22: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/22.jpg)
Knapsack (4)
Let G = (V,E) be network.
Let Ei = take item i edges
![Page 23: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/23.jpg)
Knapsack (4)
Let G = (V,E) be network.
Let Ei = take item i edges
Observation
The Knapsack polytope is the projection of Q with
xi =∑
e∈Ei
y(e) ∀i ∈ [n]
y(δ+(v))− y(δ−(v)) =
1 v = s
−1 v = t
0 otherwise
∀v ∈ V
x, y ≥ 0
![Page 24: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/24.jpg)
Knapsack (4)
Let G = (V,E) be network.
Let Ei = take item i edges
Observation
The Knapsack polytope is the projection of Q with
xi =∑
e∈Ei
y(e) ∀i ∈ [n]
y(δ+(v))− y(δ−(v)) =
1 v = s
−1 v = t
0 otherwise
∀v ∈ V
x, y ≥ 0
Corollary
xc(Knapsack polytope) ≤ O(n ·B).
![Page 25: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/25.jpg)
Knapsack (5)
Open problem
Consider a Knapsack polytope
P = conv
x ∈ 0, 1n |n∑
i=1
wixi ≤ B
.
P
![Page 26: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/26.jpg)
Knapsack (5)
Open problem
Consider a Knapsack polytope
P = conv
x ∈ 0, 1n |n∑
i=1
wixi ≤ B
.
Is there is always a polytope K with
K ⊆ P ⊆ (1 + ε)K
xc(K) ≤ poly(n, 1ε )?
K
(1 + ε)KP
![Page 27: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/27.jpg)
Knapsack (5)
Open problem
Consider a Knapsack polytope
P = conv
x ∈ 0, 1n |n∑
i=1
wixi ≤ B
.
Is there is always a polytope K with
K ⊆ P ⊆ (1 + ε)K
xc(K) ≤ poly(n, 1ε )?
Known:
xc(K) ≤ nO(1/ε) possible(Bienstock)
K
(1 + ε)KP
![Page 28: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/28.jpg)
Part II
Slack-matrices, Yannakakis’
Theorem and Communication
Complexity
![Page 29: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/29.jpg)
Slack-matrix
Write: P = conv(x1, . . . , xv) = x ∈ Rn | Ax ≤ b
S# facets
# vertices
SijSij = bi −AT
i xj
slack-matrix
Pb
b b
b
b
![Page 30: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/30.jpg)
Slack-matrix
Write: P = conv(x1, . . . , xv) = x ∈ Rn | Ax ≤ b
S# facets
# vertices
facet i
vertexj
SijSij = bi −AT
i xj
slack-matrix
Pb
b b
b
bAix = bi
bxj
Sij
![Page 31: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/31.jpg)
Slack-matrix
Write: P = conv(x1, . . . , xv) = x ∈ Rn | Ax ≤ b
S# facets
# vertices
U≥0
V ≥ 0rr
SijSij = bi −AT
i xj
slack-matrix
Pb
b b
b
bAix = bi
bxj
Sij
Non-negative rank:
rk+(S) = minr | ∃U ∈ Rf×r≥0 , V ∈ Rr×v
≥0 : S = UV
![Page 32: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/32.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
![Page 33: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/33.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Factorization S = UV ⇒ extended formulation: Let P = x ∈ Rn | ∃y ≥ 0 : Ax+ Uy = b
![Page 34: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/34.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Factorization S = UV ⇒ extended formulation: Let P = x ∈ Rn | ∃y ≥ 0 : Ax+ Uy = b For vertex xj : Aix
j + UiVj = bi.
![Page 35: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/35.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Factorization S = UV ⇒ extended formulation: Let P = x ∈ Rn | ∃y ≥ 0 : Ax+ Uy = b For vertex xj : Aix
j + UiVj = bi.
Aix > bi =⇒ Aix+ Uiy︸︷︷︸
≥0
> bi.
![Page 36: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/36.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Extended form. ⇒ factorization:
Given an extensionQ = (x, y) | Bx+ Cy ≤ d Qb
bb
b
b
b
b
b
b
b
b
b
P
![Page 37: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/37.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Extended form. ⇒ factorization:
Given an extensionQ = (x, y) | Bx+ Cy ≤ d Q
Aix+ 0y ≤ bi
b
bb
b
b
b
b
b
b
b
b
b
xjb
P
〈u(i), v(j)〉 = Sij
![Page 38: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/38.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Extended form. ⇒ factorization:
Given an extensionQ = (x, y) | Bx+ Cy ≤ d
For facet i:u(i) := conic comb of i
Q
Aix+ 0y ≤ bi
b
bb
b
b
b
b
b
b
b
b
b
xjb
P
〈u(i), v(j)〉 = Sij
![Page 39: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/39.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Extended form. ⇒ factorization:
Given an extensionQ = (x, y) | Bx+ Cy ≤ d
For facet i:u(i) := conic comb of i
For vertex xj :v(j) := d−Bxj − Cyj = slack of (xj , yj)
Q
Aix+ 0y ≤ bi
b
bb
b
b
b
b
b
b
b
b
b
xjb
(xj , yj)b
P
〈u(i), v(j)〉 = Sij
![Page 40: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/40.jpg)
Yannakakis’ Theorem
Theorem (Yannakakis ’88)
If S is the slack-matrix for P = x ∈ Rn | Ax ≤ b, thenxc(P ) = rk+(S).
Extended form. ⇒ factorization:
Given an extensionQ = (x, y) | Bx+ Cy ≤ d
For facet i:u(i) := conic comb of i
For vertex xj :v(j) := d−Bxj − Cyj = slack of (xj , yj)
Q
Aix+ 0y ≤ bi
b
bb
b
b
b
b
b
b
b
b
b
xjb
(xj , yj)b
P
〈u(i), v(j)〉 = u(i)Td︸ ︷︷ ︸
=bi
−u(i)B︸ ︷︷ ︸
=Ai
xj − u(i)C︸ ︷︷ ︸
=0
yj = Sij
![Page 41: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/41.jpg)
Rectangle covering lower bound
Observation
rk+(S) ≥ rectangle-covering-number(S).
![Page 42: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/42.jpg)
Rectangle covering lower bound
U
V
S
3
1
0
0
2
0 0 2 1 0
2
1
2
0
0
0 2 2 0 3
0 4 10 3 5
0 2 4 1 3
0 4 4 0 6
0 0 0 0 0
0 0 4 2 0
Observation
rk+(S) ≥ rectangle-covering-number(S).
![Page 43: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/43.jpg)
Rectangle covering lower bound
U
V
S
+
+
0
0+
0 0 + + 0
+
+
+
0
0
0 + + 0 +
0 + + + +
0 + + + +
0 + + 0 +
0 0 0 0 0
0 0 + + 0
Observation
rk+(S) ≥ rectangle-covering-number(S).
![Page 44: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/44.jpg)
Rectangle covering lower bound
U
V
S
+
+
0
0+
0 0 + + 0
+
+
+
0
0
0 + + 0 +
0 + + + +
0 + + + +
0 + + 0 +
0 0 0 0 0
0 0 + + 0
Observation
rk+(S) ≥ rectangle-covering-number(S).
![Page 45: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/45.jpg)
Rectangle covering lower bound
U
V
S
+
+
0
0+
0 0 + + 0
+
+
+
0
0
0 + + 0 +
0 + + + +
0 + + + +
0 + + 0 +
0 0 0 0 0
0 0 + + 0
Observation
rk+(S) ≥ rectangle-covering-number(S).
![Page 46: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/46.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Alice Bob
![Page 47: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/47.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Players Alice and Bob agree apriori on a deterministiccommunication protocoll
Alice Bob
![Page 48: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/48.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Players Alice and Bob agree apriori on a deterministiccommunication protocoll
Alice receives x ∈ X, Bob receives y ∈ X
Alicex ∈ X
Boby ∈ Y
![Page 49: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/49.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Players Alice and Bob agree apriori on a deterministiccommunication protocoll
Alice receives x ∈ X, Bob receives y ∈ X
They exchange messages to compute f(x, y)
Alicex ∈ X
Boby ∈ Y1
![Page 50: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/50.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Players Alice and Bob agree apriori on a deterministiccommunication protocoll
Alice receives x ∈ X, Bob receives y ∈ X
They exchange messages to compute f(x, y)
Alicex ∈ X
Boby ∈ Y1
1, 0
![Page 51: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/51.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Players Alice and Bob agree apriori on a deterministiccommunication protocoll
Alice receives x ∈ X, Bob receives y ∈ X
They exchange messages to compute f(x, y)
Alicex ∈ X
Boby ∈ Y1
1, 0
0
![Page 52: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/52.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Players Alice and Bob agree apriori on a deterministiccommunication protocoll
Alice receives x ∈ X, Bob receives y ∈ X
They exchange messages to compute f(x, y)
Alicex ∈ X
Boby ∈ Y1
1, 0
0
1
both know f(x, y)
![Page 53: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/53.jpg)
Communication complexity
Setting:
Function f : X × Y → R
Players Alice and Bob agree apriori on a deterministiccommunication protocoll
Alice receives x ∈ X, Bob receives y ∈ X
They exchange messages to compute f(x, y)
Alicex ∈ X
Boby ∈ Y1
1, 0
0
1
both know f(x, y)
CC(f) = minprotocoll
maxx∈X,y∈Y
bits to compute f(x, y)
![Page 54: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/54.jpg)
Communication complexity (3)
decision tree:0 1
0 1 0 1
2222000000
222
222
222
111
111
![Page 55: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/55.jpg)
Communication complexity (3)
decision tree:0 1
0 1 0 1
2222
000000
222
222
222
111
111
2222000000
222
222
222
111
111
![Page 56: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/56.jpg)
Communication complexity (3)
decision tree:0 1
0 1 0 1
2222
000000
222
222
222
111
111
2222000000
222
222
222
111
111
v
Observations:
For a leave v of tree, Rv := (x, y) : protocoll ends in v isa monochromatic rectangle
![Page 57: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/57.jpg)
Communication complexity (3)
decision tree:0 1
0 1 0 1
2222
000000
222
222
222
111
111
2222000000
222
222
222
111
111
v
Observations:
For a leave v of tree, Rv := (x, y) : protocoll ends in v isa monochromatic rectangle
A protocol for matrix S exchanging k bits⇒ S can be partitioned into 2k monochromatic rectangles⇒ S is sum of 2k rank-1 matrix
![Page 58: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/58.jpg)
Communication complexity (3)
decision tree:0 1
0 1 0 1
2222
000000
222
222
222
111
111
2222000000
222
222
222
111
111
v
Observations:
For a leave v of tree, Rv := (x, y) : protocoll ends in v isa monochromatic rectangle
A protocol for matrix S exchanging k bits⇒ S can be partitioned into 2k monochromatic rectangles⇒ S is sum of 2k rank-1 matrix
xc(polytope) ≤ 2CC(slack matrix)
![Page 59: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/59.jpg)
An exact model (1)
We have seen:
2CCnondet(supp. of slack matrix) ≤ xc(polytope) ≤ 2CC(slack matrix)
![Page 60: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/60.jpg)
An exact model (1)
We have seen:
2CCnondet(supp. of slack matrix) ≤ xc(polytope) ≤ 2CC(slack matrix)
Exact communication model (Faenza, Fiorini, Grappe,Tiwary & Zhang):
![Page 61: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/61.jpg)
An exact model (1)
We have seen:
2CCnondet(supp. of slack matrix) ≤ xc(polytope) ≤ 2CC(slack matrix)
Exact communication model (Faenza, Fiorini, Grappe,Tiwary & Zhang):
Alice receives row index i. Bob receives col index j
![Page 62: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/62.jpg)
An exact model (1)
We have seen:
2CCnondet(supp. of slack matrix) ≤ xc(polytope) ≤ 2CC(slack matrix)
Exact communication model (Faenza, Fiorini, Grappe,Tiwary & Zhang):
Alice receives row index i. Bob receives col index j
Alice and Bob have internal randomness. They exchange0/1 bits.
![Page 63: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/63.jpg)
An exact model (1)
We have seen:
2CCnondet(supp. of slack matrix) ≤ xc(polytope) ≤ 2CC(slack matrix)
Exact communication model (Faenza, Fiorini, Grappe,Tiwary & Zhang):
Alice receives row index i. Bob receives col index j
Alice and Bob have internal randomness. They exchange0/1 bits.
At the end one of them outputs a non-negative numberwith Sij = Erandomness[protocoll(i, j)].
![Page 64: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/64.jpg)
An exact model (1)
We have seen:
2CCnondet(supp. of slack matrix) ≤ xc(polytope) ≤ 2CC(slack matrix)
Exact communication model (Faenza, Fiorini, Grappe,Tiwary & Zhang):
Alice receives row index i. Bob receives col index j
Alice and Bob have internal randomness. They exchange0/1 bits.
At the end one of them outputs a non-negative numberwith Sij = Erandomness[protocoll(i, j)].
Let CCRAND(S) be min. # bits. Then
xc(polytope) = 2CCRAND(slack matrix S)
![Page 65: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/65.jpg)
An exact model (2)
Let ui, vj ∈ Rr≥0 with Sij = 〈ui, vj〉.
Sij
S
i
j
Alice
Bob
![Page 66: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/66.jpg)
An exact model (2)
Let ui, vj ∈ Rr≥0 with Sij = 〈ui, vj〉.
Sij
S
i
j
Alice
Bob
Protocoll with log2(r) bits:
![Page 67: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/67.jpg)
An exact model (2)
Let ui, vj ∈ Rr≥0 with Sij = 〈ui, vj〉.
Sij
S
i
j
Alice
Bob
Protocoll with log2(r) bits:
Alice picks k ∈ [r] with probability vi(k) (ass. ‖vi‖1 ≤ 1)
![Page 68: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/68.jpg)
An exact model (2)
Let ui, vj ∈ Rr≥0 with Sij = 〈ui, vj〉.
Sij
S
i
j
Alice
Bob
Protocoll with log2(r) bits:
Alice picks k ∈ [r] with probability vi(k) (ass. ‖vi‖1 ≤ 1)
Bob receives k from Alice and outputs uj(k)
![Page 69: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/69.jpg)
An exact model (2)
Let ui, vj ∈ Rr≥0 with Sij = 〈ui, vj〉.
Sij
S
i
j
Alice
Bob
Protocoll with log2(r) bits:
Alice picks k ∈ [r] with probability vi(k) (ass. ‖vi‖1 ≤ 1)
Bob receives k from Alice and outputs uj(k)
Then
E[protocoll(i, j)] =r∑
k=1
vi(k) · uj(k) = 〈ui, vj〉
![Page 70: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/70.jpg)
An exact model (3)
Consider depth c protocoll tree:
Alice
Bob
Alice
Bob
leaf w
c
![Page 71: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/71.jpg)
An exact model (3)
Consider depth c protocoll tree:
Alice
Bob
Alice
Bob
leaf w
c
Sij =∑
leaves w
value(w) · Pr[protocoll(i, j) ends at w]
![Page 72: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/72.jpg)
An exact model (3)
Consider depth c protocoll tree:
Alice
Bob
Alice
Bob
leaf w
c
Sij =∑
leaves w
value(w) · Pr[protocoll(i, j) ends at w] =
∑
leaves w
value(w)︸ ︷︷ ︸
≥0
·Pr[Alice stays on w path | i] · Pr[Bob stays on w path | j]
![Page 73: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/73.jpg)
An exact model (3)
Consider depth c protocoll tree:
Alice
Bob
Alice
Bob
leaf w
c
Sij =∑
leaves w
value(w) · Pr[protocoll(i, j) ends at w] =
∑
leaves w
value(w)︸ ︷︷ ︸
≥0
·Pr[Alice stays on w path | i] · Pr[Bob stays on w path | j]
⇒ 2c-size non-neg factorization
![Page 74: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/74.jpg)
Part III
The Lower Bound on the
Correlation Polytope
![Page 75: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/75.jpg)
Correlation polytope (1)
The correlation polytope is
COR = convbbT : b ∈ 0, 1n
![Page 76: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/76.jpg)
Correlation polytope (1)
The correlation polytope is
COR = convbbT : b ∈ 0, 1n
Example: For n = 2,
COR = conv
(0 00 0
)
,
(1 00 0
)
,
(0 00 1
)
,
(1 11 1
)
![Page 77: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/77.jpg)
Correlation polytope (1)
The correlation polytope is
COR = convbbT : b ∈ 0, 1n
Example: For n = 2,
COR = conv
(0 00 0
)
,
(1 00 0
)
,
(0 00 1
)
,
(1 11 1
)
Observation: The polytope is NP-hard.For graph G = ([n], E) with adjacency matrix AG
maxcut(G) = maxx∈0,1n
(DG−AG)•xxT =
∑
(i,j)∈E
(xi + xj − 2xixj)︸ ︷︷ ︸
=1 if xi 6=xj ,0 o.w.
![Page 78: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/78.jpg)
Correlation polytope (1)
The correlation polytope is
COR = convbbT : b ∈ 0, 1n
Example: For n = 2,
COR = conv
(0 00 0
)
,
(1 00 0
)
,
(0 00 1
)
,
(1 11 1
)
Observation: The polytope is NP-hard.For graph G = ([n], E) with adjacency matrix AG
maxcut(G) = maxx∈0,1n
(DG−AG)•xxT =
∑
(i,j)∈E
(xi + xj − 2xixj)︸ ︷︷ ︸
=1 if xi 6=xj ,0 o.w.
Theorem (Fiorini, Massar, Pokutta, Tiwary, de Wolf ’12)
xc(COR) ≥ 2Ω(n).
Here: Simplified proof by Kaibel and Weltge
![Page 79: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/79.jpg)
Correlation polytope (2)
Lemma
For all a ∈ 0, 1n, (2diag(a)− aaT ) • Y ≤ 1 is a feasibleinequality for Y ∈ COR.
![Page 80: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/80.jpg)
Correlation polytope (2)
Lemma
For all a ∈ 0, 1n, (2diag(a)− aaT ) • Y ≤ 1 is a feasibleinequality for Y ∈ COR.
Suffices to check slack for Y = bbT .
1− 2·
11110000
supp(a)
•
1111
1111
1111
1111
supp(b)
+ 1111
1111
1111
1111
supp(a)
•
1111
1111
1111
1111
supp(b)
![Page 81: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/81.jpg)
Correlation polytope (2)
Lemma
For all a ∈ 0, 1n, (2diag(a)− aaT ) • Y ≤ 1 is a feasibleinequality for Y ∈ COR.
Suffices to check slack for Y = bbT .
1− 2·
11110000
supp(a)
•
1111
1111
1111
1111
supp(b)
+ 1111
1111
1111
1111
supp(a)
•
1111
1111
1111
1111
supp(b)
= 1− 2|a ∩ b|+ |a ∩ b|2 = (1− |a ∩ b|)2 ≥ 0
![Page 82: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/82.jpg)
Correlation polytope (3)
a
b
slack matrix S
(1− |a ∩ b|)2 (2diag(a)− aaT ) • Y ≤ 1
P
b
b b
b
COR
bbT
![Page 83: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/83.jpg)
Correlation polytope (3)
a
b
slack matrix S
(1− |a ∩ b|)2 (2diag(a)− aaT ) • Y ≤ 1
P
b
b b
b
COR
bbT
Observations:
S is a submatrix of the “real” slack-matrix
![Page 84: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/84.jpg)
Correlation polytope (3)
a
b
slack matrix S
(1− |a ∩ b|)2 (2diag(a)− aaT ) • Y ≤ 1
P
b
b b
b
COR
bbT
Observations:
S is a submatrix of the “real” slack-matrix
We have
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
![Page 85: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/85.jpg)
Incomplete slack matrices
Lemma
For a polytope P = x | Ax ≤ b and X = x1, . . . , xv ⊆ Pdefine a matrix S with Si,j := bi −Aixj . Then
rk≥0(S) = minxc(Q) : X ⊆ Q ⊆ P
P
b
b b
b
conv(X)
![Page 86: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/86.jpg)
Incomplete slack matrices
Lemma
For a polytope P = x | Ax ≤ b and X = x1, . . . , xv ⊆ Pdefine a matrix S with Si,j := bi −Aixj . Then
rk≥0(S) = minxc(Q) : X ⊆ Q ⊆ P
Q
P
b
b b
b
conv(X)
![Page 87: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/87.jpg)
Correlation polytope (3)
Sa
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
![Page 88: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/88.jpg)
Correlation polytope (3)
S1
11
11
a
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
Define disjoint pairs P0 := (a, b) : |a ∩ b| = 0
![Page 89: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/89.jpg)
Correlation polytope (3)
S1
11
11
a
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
Define disjoint pairs P0 := (a, b) : |a ∩ b| = 0
Claim
|P0| = 3n.
![Page 90: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/90.jpg)
Correlation polytope (3)
S1
11
11
a
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
Define disjoint pairs P0 := (a, b) : |a ∩ b| = 0
Claim
|P0| = 3n.
For disjoint pair (a, b), for coordinate i there are 3 options ai = 0, bi = 0 ai = 1, bi = 0 ai = 0, bi = 1
![Page 91: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/91.jpg)
Correlation polytope (4)
S
1
11
11
a
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
disjoint pairs P0 := (a, b) : |a ∩ b| = 0
![Page 92: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/92.jpg)
Correlation polytope (4)
S
1
11
11
0
0
0
00
a
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
disjoint pairs P0 := (a, b) : |a ∩ b| = 0 forbidden pairs P1 := (a, b) : |a ∩ b| = 1
![Page 93: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/93.jpg)
Correlation polytope (4)
S
1
11
11
0
0
0
00
Ra
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
disjoint pairs P0 := (a, b) : |a ∩ b| = 0 forbidden pairs P1 := (a, b) : |a ∩ b| = 1
![Page 94: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/94.jpg)
Correlation polytope (4)
S
1
11
11
0
0
0
00
Ra
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
disjoint pairs P0 := (a, b) : |a ∩ b| = 0 forbidden pairs P1 := (a, b) : |a ∩ b| = 1
Lemma
Any rectangle R without forbidden pairs has |R ∩ P0| ≤ 2n.
![Page 95: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/95.jpg)
Correlation polytope (4)
S
1
11
11
0
0
0
00
Ra
b
Sab =
1 |a ∩ b| = 0
0 |a ∩ b| = 1
disjoint pairs P0 := (a, b) : |a ∩ b| = 0 forbidden pairs P1 := (a, b) : |a ∩ b| = 1
Lemma
Any rectangle R without forbidden pairs has |R ∩ P0| ≤ 2n.
By rectangle covering lower bound
xc(COR) ≥|P0|
2n=
(3
2
)n
![Page 96: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/96.jpg)
Correlation polytope (5)
R
![Page 97: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/97.jpg)
Correlation polytope (5)
R
a : 1 ∈ a
a : 1 /∈ a
b : 1 /∈ b b : 1 ∈ b
![Page 98: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/98.jpg)
Correlation polytope (5)
R
a : 1 ∈ a
a : 1 /∈ a
b : 1 /∈ b b : 1 ∈ b
R1
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b
![Page 99: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/99.jpg)
Correlation polytope (5)
R
a : 1 ∈ a
a : 1 /∈ a
b : 1 /∈ b b : 1 ∈ b
R1
R2
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b andR2 = (a, b) ∈ R : 1 /∈ a
![Page 100: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/100.jpg)
Correlation polytope (5)
R
a : 1 ∈ a
a : 1 /∈ a
b : 1 /∈ b b : 1 ∈ b
R1
R2
1
1
1
1
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b andR2 = (a, b) ∈ R : 1 /∈ a
All entries in P0 covered
![Page 101: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/101.jpg)
Correlation polytope (5)
R
a : 1 ∈ a
a : 1 /∈ a
b : 1 /∈ b b : 1 ∈ b
R1
R2
1
1
1
1
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b andR2 = (a, b) ∈ R : 1 /∈ a
All entries in P0 covered
Delete symbol 1 from all tuples and apply induction
|P0∩R| ≤ |P0(n−1)∩R1|+ |P0(n−1)∩R2| ≤ 2 ·2n−1 = 2n
![Page 102: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/102.jpg)
Correlation polytope (5)
R
a : 1 ∈ a
a : 1 /∈ a
b : 1 /∈ b b : 1 ∈ b
R1
R2
1
1
1
1
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b andR2 = (a, b) ∈ R : 1 /∈ a
All entries in P0 covered
Delete symbol 1 from all tuples and apply induction
|P0∩R| ≤ |P0(n−1)∩R1|+ |P0(n−1)∩R2| ≤ 2 ·2n−1 = 2n
Good news: No forbidden pairs created!
![Page 103: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/103.jpg)
Correlation polytope (5)
R
R1
R2a
a ∪ 1 1
1
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b andR2 = (a, b) ∈ R : 1 /∈ a
All entries in P0 covered
Delete symbol 1 from all tuples and apply induction
|P0∩R| ≤ |P0(n−1)∩R1|+ |P0(n−1)∩R2| ≤ 2 ·2n−1 = 2n
Good news: No forbidden pairs created!
Potential problem: Diff. tuples collapse to the same one
![Page 104: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/104.jpg)
Correlation polytope (5)
R
R1
R2a
a ∪ 1 1
1
b b ∪ 1
1
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b andR2 = (a, b) ∈ R : 1 /∈ a
All entries in P0 covered
Delete symbol 1 from all tuples and apply induction
|P0∩R| ≤ |P0(n−1)∩R1|+ |P0(n−1)∩R2| ≤ 2 ·2n−1 = 2n
Good news: No forbidden pairs created!
Potential problem: Diff. tuples collapse to the same one
![Page 105: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/105.jpg)
Correlation polytope (5)
R
R1
R2a
a ∪ 1 1
1
b b ∪ 1
1
0
Define rectangles R1 = (a, b) ∈ R : 1 /∈ b andR2 = (a, b) ∈ R : 1 /∈ a
All entries in P0 covered
Delete symbol 1 from all tuples and apply induction
|P0∩R| ≤ |P0(n−1)∩R1|+ |P0(n−1)∩R2| ≤ 2 ·2n−1 = 2n
Good news: No forbidden pairs created!
Potential problem: Diff. tuples collapse to the same one
![Page 106: Extended Formulations I (BootCamp)](https://reader031.vdocuments.site/reader031/viewer/2022012102/616a020a11a7b741a34dcceb/html5/thumbnails/106.jpg)
The end
Thanks for your attention