We often think of graphs and of their global properties in a very visual way. This works fine for smaller graphs, but for most real-world applications, the underlying graphs are very large---sometimes so large that their data cannot fit on a computer. In that case, how can one understand such a graph and its global properties? One way is to think about it locally. Given a large graph G, create a vector recording the induced density in G of every small graph in some fixed finite list of graphs. We can think of this vector as the coordinates of G in the space of the smaller graphs. This way of thinking about large graphs immediately raises two questions. First, how are local and global properties related? In other words, given the coordinates of G, what are global properties of G? Second, what is even possible locally? Say you want to create a graph with certain local distributions, can it be done? I am interested in the second question and, in this talk, I will explore its connections to extremal graph theory and symmetric sums of squares. This is joint work with Greg Blekherman, Mohit Singh and Rekha Thomas.
2018 Modern Math Workshop - Symmetry and Graph Profiles - Annie Raymond, October 10, 2018
1. Symmetry and Graph Profiles
Annie Raymond
(joint with Greg Blekherman, Mohit Singh and Rekha Thomas)
University of Massachusetts, Amherst
October 10, 2018
2. An example
Theorem (Mantel, 1907)
The maximum number of edges in a graph on n vertices with no triangles
is n2
4 . In particular, as n → ∞, the maximum edge density goes to 1
2.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 2 / 18
3. An example
Theorem (Mantel, 1907)
The maximum number of edges in a graph on n vertices with no triangles
is n2
4 . In particular, as n → ∞, the maximum edge density goes to 1
2.
The maximum is attained on K n
2
, n
2
:
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 2 / 18
4. An example
Theorem (Mantel, 1907)
The maximum number of edges in a graph on n vertices with no triangles
is n2
4 . In particular, as n → ∞, the maximum edge density goes to 1
2.
The maximum is attained on K n
2
, n
2
:
n
2 · n
2 edges out of n
2 potential edges, no triangles.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 2 / 18
5. Other triangle densities
What if I want to know the maximum edge density in a graph on n
vertices with a triangle density of y, for some 0 ≤ y ≤ 1?
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 3 / 18
6. Other triangle densities
What if I want to know the maximum edge density in a graph on n
vertices with a triangle density of y, for some 0 ≤ y ≤ 1?
Example
Let G = ,
then (d( , G), d( , G)) = 9
(7
2)
, 2
(7
3)
≈ (0.43, 0.06).
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 3 / 18
7. Other triangle densities
What if I want to know the maximum edge density in a graph on n
vertices with a triangle density of y, for some 0 ≤ y ≤ 1?
Example
Let G = ,
then (d( , G), d( , G)) = 9
(7
2)
, 2
(7
3)
≈ (0.43, 0.06).
Is that the max edge density among graphs on 7 vertices with 2 triangles?
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 3 / 18
8. Other triangle densities
What if I want to know the maximum edge density in a graph on n
vertices with a triangle density of y, for some 0 ≤ y ≤ 1?
Example
Let G = ,
then (d( , G), d( , G)) = 9
(7
2)
, 2
(7
3)
≈ (0.43, 0.06).
Is that the max edge density among graphs on 7 vertices with 2 triangles?
What can (d( , G), d( , G)) be if G is any graph on 7 vertices?
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 3 / 18
9. All density vectors for graphs on 7 vertices
(d( , G), d( , G)) for any graph G on 7 vertices
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 4 / 18
10. All density vectors for graphs on n vertices as n → ∞
(d( , G), d( , G)) for any graph G on n vertices as n → ∞
(Razborov, 2008)
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 5 / 18
11. Why care?
Large graphs are everywhere!
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 6 / 18
12. Why care?
Large graphs are everywhere!
Biology
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 6 / 18
13. Why care?
Large graphs are everywhere!
Biology
Facebook graph
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 6 / 18
14. More reasons to care!
Google Maps
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 7 / 18
15. More reasons to care!
Google Maps
Alfred Pasieka/Science Photo Library/Getty Images
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 7 / 18
16. Problem
Adjacency matrix A(G) of a graph G:
Aij (G) =
1 if {i, j} ∈ E(G)
0 if {i, j} ∈ E(G)
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 8 / 18
17. Problem
Adjacency matrix A(G) of a graph G:
Aij (G) =
1 if {i, j} ∈ E(G)
0 if {i, j} ∈ E(G)
Sometimes too large for computer!
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 8 / 18
18. Problem
Adjacency matrix A(G) of a graph G:
Aij (G) =
1 if {i, j} ∈ E(G)
0 if {i, j} ∈ E(G)
Sometimes too large for computer!
Idea: understand the graph locally
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 8 / 18
19. Problem
Adjacency matrix A(G) of a graph G:
Aij (G) =
1 if {i, j} ∈ E(G)
0 if {i, j} ∈ E(G)
Sometimes too large for computer!
Idea: understand the graph locally
This raises immediately two questions:
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 8 / 18
20. Problem
Adjacency matrix A(G) of a graph G:
Aij (G) =
1 if {i, j} ∈ E(G)
0 if {i, j} ∈ E(G)
Sometimes too large for computer!
Idea: understand the graph locally
This raises immediately two questions:
1 How do global and local properties relate?
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 8 / 18
21. Problem
Adjacency matrix A(G) of a graph G:
Aij (G) =
1 if {i, j} ∈ E(G)
0 if {i, j} ∈ E(G)
Sometimes too large for computer!
Idea: understand the graph locally
This raises immediately two questions:
1 How do global and local properties relate?
2 What is even possible locally?
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 8 / 18
22. Valid linear inequality involving graph densities
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 9 / 18
23. Valid linear inequality involving graph densities
− ≥ 0, −4
3 + + 2
3 ≥ 0, . . .
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 9 / 18
24. Valid linear inequality involving graph densities
− ≥ 0, −4
3 + + 2
3 ≥ 0, . . .
Consider any linear inequality involving any graphs (not just edges and
triangles) and any number of graphs (not just two).
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 9 / 18
26. Certifying polynomial inequalities
A polynomial p ∈ R[x1, . . . , xn] =: R[x]
is nonnegative if p(x1, . . . , xn) ≥ 0 for all (x1, . . . , xn) ∈ Rn
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 10 / 18
27. Certifying polynomial inequalities
A polynomial p ∈ R[x1, . . . , xn] =: R[x]
is nonnegative if p(x1, . . . , xn) ≥ 0 for all (x1, . . . , xn) ∈ Rn
p sum of squares (sos), i.e., p = l
i=1 f 2
i where fi ∈ R[x] ⇒ p ≥ 0
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 10 / 18
28. Certifying polynomial inequalities
A polynomial p ∈ R[x1, . . . , xn] =: R[x]
is nonnegative if p(x1, . . . , xn) ≥ 0 for all (x1, . . . , xn) ∈ Rn
p sum of squares (sos), i.e., p = l
i=1 f 2
i where fi ∈ R[x] ⇒ p ≥ 0
Hilbert (1888): Not all nonnegative polynomials are sos.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 10 / 18
29. Certifying polynomial inequalities
A polynomial p ∈ R[x1, . . . , xn] =: R[x]
is nonnegative if p(x1, . . . , xn) ≥ 0 for all (x1, . . . , xn) ∈ Rn
p sum of squares (sos), i.e., p = l
i=1 f 2
i where fi ∈ R[x] ⇒ p ≥ 0
Hilbert (1888): Not all nonnegative polynomials are sos.
Artin (1927): Every nonnegative polynomial can be written as a (finite)
sum of squares of rational functions.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 10 / 18
30. Certifying polynomial inequalities
A polynomial p ∈ R[x1, . . . , xn] =: R[x]
is nonnegative if p(x1, . . . , xn) ≥ 0 for all (x1, . . . , xn) ∈ Rn
p sum of squares (sos), i.e., p = l
i=1 f 2
i where fi ∈ R[x] ⇒ p ≥ 0
Hilbert (1888): Not all nonnegative polynomials are sos.
Artin (1927): Every nonnegative polynomial can be written as a (finite)
sum of squares of rational functions.
Motzkin (1967, with Taussky-Todd): M(x, y) = x4y2 + x2y4 + 1 − 3x2y2
is a nonnegative polynomial but is not a sos.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 10 / 18
31. How do sums of squares fare with graph densities?
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 11 / 18
32. How do sums of squares fare with graph densities?
Hatami-Norine (2011): Not every valid inequality involving graph densities
can be written as a sum of squares of graph densities
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 11 / 18
33. How do sums of squares fare with graph densities?
Hatami-Norine (2011): Not every valid inequality involving graph densities
can be written as a sum of squares of graph densities or even as a rational
sum of squares of graph densities.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 11 / 18
34. How do sums of squares fare with graph densities?
Hatami-Norine (2011): Not every valid inequality involving graph densities
can be written as a sum of squares of graph densities or even as a rational
sum of squares of graph densities.
Lov´asz-Szegedy (2006) + Netzer-Thom (2015): Every valid inequality
involving graph densities plus any > 0 can be written as a sum of squares
of graph densities.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 11 / 18
35. How do sums of squares fare with graph densities?
Hatami-Norine (2011): Not every valid inequality involving graph densities
can be written as a sum of squares of graph densities or even as a rational
sum of squares of graph densities.
Lov´asz-Szegedy (2006) + Netzer-Thom (2015): Every valid inequality
involving graph densities plus any > 0 can be written as a sum of squares
of graph densities.
BRST (2018): − ≥ 0 is not a sum of squares.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 11 / 18
36. How do sums of squares fare with graph densities?
Hatami-Norine (2011): Not every valid inequality involving graph densities
can be written as a sum of squares of graph densities or even as a rational
sum of squares of graph densities.
Lov´asz-Szegedy (2006) + Netzer-Thom (2015): Every valid inequality
involving graph densities plus any > 0 can be written as a sum of squares
of graph densities.
BRST (2018): − ≥ 0 is not a sum of squares.
How? We characterize exactly all valid inequalities involving only
graphs with three edges, and which of these are sums of squares.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 11 / 18
37. Tools to work on such problems
Graphs on n vertices ←→ subsets of {0, 1}(n
2)
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 12 / 18
38. Tools to work on such problems
Graphs on n vertices ←→ subsets of {0, 1}(n
2)
1
2
3 4
←→
12 13 14 23 24 34
( 1, 1, 1, 0, 0, 0 )
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 12 / 18
39. Tools to work on such problems
Graphs on n vertices ←→ subsets of {0, 1}(n
2)
1
2
3 4
←→
12 13 14 23 24 34
( 1, 1, 1, 0, 0, 0 )
Variables xij
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 12 / 18
40. Tools to work on such problems
Graphs on n vertices ←→ subsets of {0, 1}(n
2)
1
2
3 4
←→
12 13 14 23 24 34
( 1, 1, 1, 0, 0, 0 )
Variables xij → transform polynomials into pictures!
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 12 / 18
41. Tools to work on such problems
Graphs on n vertices ←→ subsets of {0, 1}(n
2)
1
2
3 4
←→
12 13 14 23 24 34
( 1, 1, 1, 0, 0, 0 )
Variables xij → transform polynomials into pictures!
x12 = 1
2
and x12x13x23 = 1
2 3
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 12 / 18
42. Tools to work on such problems
Graphs on n vertices ←→ subsets of {0, 1}(n
2)
1
2
3 4
←→
12 13 14 23 24 34
( 1, 1, 1, 0, 0, 0 )
Variables xij → transform polynomials into pictures!
x12 = 1
2
and x12x13x23 = 1
2 3
x12(G) = 1
2
(G) gives 1 if {1, 2} ∈ E(G), and 0 otherwise
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 12 / 18
43. Tools to work on such problems
Graphs on n vertices ←→ subsets of {0, 1}(n
2)
1
2
3 4
←→
12 13 14 23 24 34
( 1, 1, 1, 0, 0, 0 )
Variables xij → transform polynomials into pictures!
x12 = 1
2
and x12x13x23 = 1
2 3
x12(G) = 1
2
(G) gives 1 if {1, 2} ∈ E(G), and 0 otherwise
x12x13x23(G) = 1
2 3
(G) gives 1 if the vertices 1,2, and 3 form a
triangle in G, and 0 otherwise
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 12 / 18
44. Symmetrization
Example (Definition by example)
Let = symn( 1
2 3
) = 1
n! σ∈Sn
σ( 1
2 3
).
(G) returns the triangle density of G.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 13 / 18
45. Symmetrization
Example (Definition by example)
Let = symn( 1
2 3
) = 1
n! σ∈Sn
σ( 1
2 3
).
(G) returns the triangle density of G.
Example (Crucial definition by example: using only a subgroup of Sn)
Let 1 = symσ∈Sn:σ fixes 1( 1
2
) = 1
n−1 j≥2 x1j
1 (G) returns the relative degree of vertex 1 in G.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 13 / 18
46. Symmetrization
Example (Definition by example)
Let = symn( 1
2 3
) = 1
n! σ∈Sn
σ( 1
2 3
).
(G) returns the triangle density of G.
Example (Crucial definition by example: using only a subgroup of Sn)
Let 1 = symσ∈Sn:σ fixes 1( 1
2
) = 1
n−1 j≥2 x1j
1 (G) returns the relative degree of vertex 1 in G.
Example (One more example to clarify)
1
2 1
3
4 2
5
6
= 2
4
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 13 / 18
50. Certifying a valid inequality with a sum of squares
Show that − ≥ 0.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 15 / 18
51. Certifying a valid inequality with a sum of squares
Show that − ≥ 0.
1
2
symn ( 1 − 2 )2
=
1
2
symn( 1 1 − 2 1 2 + 2 2 )
=
1
2
symn( 1 − 2 1 2 + 2 )
=
1
2
(2 − 2 )
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 15 / 18
52. Certifying a valid inequality with a sum of squares
Show that − ≥ 0.
1
2
symn ( 1 − 2 )2
=
1
2
symn( 1 1 − 2 1 2 + 2 2 )
=
1
2
symn( 1 − 2 1 2 + 2 )
=
1
2
(2 − 2 )
By using symmetry-reduction, we fully characterize all homogeneous
sums of squares of degree 3, 4, and 5.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 15 / 18
53. 3-profiles of graphs
BRST(2018):
(d( , G), d( , G), d( , G), d( , G)) is contained in
B = {x ∈ R4
: x0 + x1 + x2 + x3 = 1,
x0, x1, x2, x3 ≥ 0
3x0 + x1 x1 + x2
x1 + x2 x2 + 3x3
0}
which looks like...
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 16 / 18
54. Convex relaxation for 3-profiles of graphs
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 17 / 18
55. Convex relaxation for 3-profiles of graphs
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 17 / 18
56. Actual 3-profiles of graphs
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 17 / 18
57. Actual 3-profiles of graphs
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 17 / 18
58. Thank you!
Also follow forall on instagram
or check out www.instagram.com/_forall.
Annie Raymond (UMass) Symmetry and Graph Profiles October 10, 2018 18 / 18