Spectra of Graphons:some spectral results from
Laszlo Lovasz’s textbook Large Networks and Graph Limits
Alexander W. N. Riasanovsky
[email protected] Graph Theory (MATH 595)
at Iowa State University
April 19, 2017
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 1 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neuroscience
the Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculus
quasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)
graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Motivation
study large networks
neurosciencethe Internet, social networks
utilize analytic tools in graph theory
continuity, compactness, measure theory, integration, variationalcalculusquasirandomness, Szemeredi’s regularity lemma
compare graphs of different orders to each other
graphs which change slightly over time (Internet, social media)graphs with the same basic shape but vastly different orders
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 2 / 22
Pixel Pictures of Graphs
0 1 1 11 0 1 01 1 0 01 0 0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0 1 1 11 0 1 01 1 0 01 0 0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0 1 1 11 0 1 01 1 0 01 0 0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0
1 1 1
1
0 1 0
1
1 0 0
1
0 0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0 1
1 1
1 0
1 0
1 1
0 0
1 0
0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0 1 1
1
1 0 1
0
1 1 0
0
1 0 0
0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0 1 1 11 0 1 01 1 0 01 0 0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0 1 1 11 0 1 01 1 0 01 0 0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Pixel Pictures of Graphs
0 1 1 11 0 1 01 1 0 01 0 0 0
Table: The Paw graph.
Pixel Picture
The pixel picture WG is the {0, 1}-valued step function found by fittingthe adjacency matrix of G in the unit square.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 3 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 21 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 22 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 23 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 24 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 25 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 26 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 27 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 28 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Two Sequences of Random Graphs
Table: The only probable limits in growing sequences of Erdos-Renyi randomgraphs (p = 1/2) and uniform attachment graphs.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 4 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n],
the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣∣ .Cut Norm of a Matrix [[Frieze, Kannan]’99]
For A some n × n matrix,
||A||� := maxS,T⊆[n]
∣∣∣∣∣∣∑
i∈S ,j∈T
aijn2
∣∣∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS A1Tn2
∣∣∣∣ .
For A = AG , this is different from Cheeger’s constant hG .
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 5 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Cut Norm of a Matrix [[Frieze, Kannan]’99]
For A some n × n matrix,
||A||� := maxS,T⊆[n]
∣∣∣∣∣∣∑
i∈S ,j∈T
aijn2
∣∣∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS A1Tn2
∣∣∣∣ .
For A = AG , this is different from Cheeger’s constant hG .
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 5 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣
= maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Cut Norm of a Matrix [[Frieze, Kannan]’99]
For A some n × n matrix,
||A||� := maxS,T⊆[n]
∣∣∣∣∣∣∑
i∈S ,j∈T
aijn2
∣∣∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS A1Tn2
∣∣∣∣ .
For A = AG , this is different from Cheeger’s constant hG .
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 5 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .
Cut Norm of a Matrix [[Frieze, Kannan]’99]
For A some n × n matrix,
||A||� := maxS,T⊆[n]
∣∣∣∣∣∣∑
i∈S ,j∈T
aijn2
∣∣∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS A1Tn2
∣∣∣∣ .
For A = AG , this is different from Cheeger’s constant hG .
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 5 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Cut Norm of a Matrix [[Frieze, Kannan]’99]
For A some n × n matrix,
||A||� := maxS,T⊆[n]
∣∣∣∣∣∣∑
i∈S ,j∈T
aijn2
∣∣∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS A1Tn2
∣∣∣∣ .
For A = AG , this is different from Cheeger’s constant hG .
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 5 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Cut Norm of a Matrix [[Frieze, Kannan]’99]
For A some n × n matrix,
||A||� := maxS,T⊆[n]
∣∣∣∣∣∣∑
i∈S ,j∈T
aijn2
∣∣∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS A1Tn2
∣∣∣∣ .For A = AG , this is different from Cheeger’s constant hG .Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 5 / 22
Example of Cut Distance of Graphs
AG − AH
=
1 −1
1 −1
−1 1−1 1
=
1 −1
1 −1
−1 1−1 1
Example
If G has edges 12 and 34 and H has edges 13 and 24,
thenD�(G ,H) = ||AG − AH || = 1/8. Note also that ||AG ||� = ||AH ||� = 1/4.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 6 / 22
Example of Cut Distance of Graphs
AG − AH
=
1 −1
1 −1
−1 1−1 1
=
1 −1
1 −1
−1 1−1 1
Example
If G has edges 12 and 34 and H has edges 13 and 24,
thenD�(G ,H) = ||AG − AH || = 1/8. Note also that ||AG ||� = ||AH ||� = 1/4.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 6 / 22
Example of Cut Distance of Graphs
AG − AH
=
1 −1
1 −1
−1 1−1 1
=
1 −1
1 −1
−1 1−1 1
Example
If G has edges 12 and 34 and H has edges 13 and 24,
thenD�(G ,H) = ||AG − AH || = 1/8. Note also that ||AG ||� = ||AH ||� = 1/4.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 6 / 22
Example of Cut Distance of Graphs
AG − AH
=
1 −1
1 −1
−1 1−1 1
=
1 −1
1 −1
−1 1−1 1
Example
If G has edges 12 and 34 and H has edges 13 and 24,
thenD�(G ,H) = ||AG − AH || = 1/8. Note also that ||AG ||� = ||AH ||� = 1/4.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 6 / 22
Example of Cut Distance of Graphs
AG − AH
=
1 −1
1 −1
−1 1−1 1
=
1 −1
1 −1
−1 1−1 1
Example
If G has edges 12 and 34 and H has edges 13 and 24,
thenD�(G ,H) = ||AG − AH || = 1/8. Note also that ||AG ||� = ||AH ||� = 1/4.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 6 / 22
Example of Cut Distance of Graphs
AG − AH
=
1 −1
1 −1
−1 1−1 1
=
1 −1
1 −1
−1 1−1 1
Example
If G has edges 12 and 34 and H has edges 13 and 24, thenD�(G ,H) = ||AG − AH || = 1/8.
Note also that ||AG ||� = ||AH ||� = 1/4.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 6 / 22
Example of Cut Distance of Graphs
AG − AH
=
1 −1
1 −1
−1 1−1 1
=
1 −1
1 −1
−1 1−1 1
Example
If G has edges 12 and 34 and H has edges 13 and 24, thenD�(G ,H) = ||AG − AH || = 1/8. Note also that ||AG ||� = ||AH ||� = 1/4.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 6 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Quasirandomness
Pretend that H is a p = 1/2 ER random graph: replace “eH(S ,T )” with“|S ||T |/2”, and “AH” with “1
2J”, the all-12 matrix. Then D�(G , 1/2) < εimplies that G is ε-quasirandom.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 7 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Quasirandomness
Pretend that H is a p = 1/2 ER random graph:
replace “eH(S ,T )” with“|S ||T |/2”, and “AH” with “1
2J”, the all-12 matrix. Then D�(G , 1/2) < εimplies that G is ε-quasirandom.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 7 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Quasirandomness
Pretend that H is a p = 1/2 ER random graph: replace “eH(S ,T )” with“|S ||T |/2”
, and “AH” with “12J”, the all-12 matrix. Then D�(G , 1/2) < ε
implies that G is ε-quasirandom.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 7 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Quasirandomness
Pretend that H is a p = 1/2 ER random graph: replace “eH(S ,T )” with“|S ||T |/2”, and “AH” with “1
2J”, the all-12 matrix.
Then D�(G , 1/2) < εimplies that G is ε-quasirandom.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 7 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Quasirandomness
Pretend that H is a p = 1/2 ER random graph: replace “eH(S ,T )” with“|S ||T |/2”, and “AH” with “1
2J”, the all-12 matrix. Then D�(G , 1/2) < εimplies that G is ε-quasirandom.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 7 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the (unlabeled) graph cutdistance d�(G ,H) of G and H is given by
d�(G ,H) := minσ∈Sn{D�(G ,Hσ)} .
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 8 / 22
Cut Distance of Graphs
Labeled Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the labeled graph cutdistance D�(G ,H) ∈ [0, 1] of G and H is given by
D�(G ,H) := maxS ,T⊆[n]
∣∣∣∣eG (S ,T )− eH(S ,T )
n2
∣∣∣∣ = maxS ,T⊆[n]
∣∣∣∣1TS (AG − AH)1Tn2
∣∣∣∣ .Cut Distance of Graphs
Given two graphs G and H with vertex sets [n], the (unlabeled) graph cutdistance d�(G ,H) of G and H is given by
d�(G ,H) := minσ∈Sn{D�(G ,Hσ)} .
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 8 / 22
Cut Distance and Graphons
Definition
If W : [0, 1]2 → [−1, 1] is symmetric and integrable, it is a signed graphon.If 0 ≤W ≤ 1, it is a graphon. For any signed graphon, let its cut norm
||W ||� of W be ||W ||� := supS,T⊆[0,1]
∣∣∣∫S×T W (x , y)dxdy∣∣∣ . Also,
D�(W ,X ) := ||W − X ||� is the labeled cut distance.
− =
Table: WG −WH with G and H from before. Here, ||WG ||� = ||WH ||� = 1/4.Also, D�(WG ,WH) = ||WG −WH ||� = 1/8.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 9 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H,
therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ) and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24
and σ swaps [1/4, 1/2]↔ [1/2, 3/4],then W σ
G = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H, therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ) and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24
and σ swaps [1/4, 1/2]↔ [1/2, 3/4],then W σ
G = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H, therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ) and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24
and σ swaps [1/4, 1/2]↔ [1/2, 3/4],then W σ
G = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H, therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ)
and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24
and σ swaps [1/4, 1/2]↔ [1/2, 3/4],then W σ
G = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H, therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ) and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24
and σ swaps [1/4, 1/2]↔ [1/2, 3/4],then W σ
G = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H, therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ) and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24
and σ swaps [1/4, 1/2]↔ [1/2, 3/4],then W σ
G = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H, therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ) and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24 and σ swaps [1/4, 1/2]↔ [1/2, 3/4],
then W σG = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance of Graphs and Graphons
Cut Distances
For any graphons W ,X : [0, 1]2 → [0, 1] W ,X and graphs G ,H, therespective cut distances are:
δ�(W ,X ) := minσ∈S [0,1]
D�(W ,X σ) and δ�(G ,H) := δ�(WG ,WH).
σ−→
Table: If G = e12 + e34 and H = e13 + e24 and σ swaps [1/4, 1/2]↔ [1/2, 3/4],then W σ
G = WH . So δ�(G1,G2) = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 10 / 22
Cut Distance Examples
Table: Complete graphsand J, the all-onesgraphon:δ�(WKn , J) = 1/n→ 0.
Table: Completebipartite graphs:δ�(K2,Kdn/2e,bn/2c)→0.
Question
How do the spectra compare?
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 11 / 22
Cut Distance Examples
Table: Complete graphsand J, the all-onesgraphon:δ�(WKn , J) = 1/n→ 0.
Table: Completebipartite graphs:δ�(K2,Kdn/2e,bn/2c)→0.
Question
How do the spectra compare?
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 11 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 21 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 22 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 23 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 24 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 25 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 26 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 27 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: An Erdos-Renyi random graph (p = 1/2) and a uniform attachmentrandom graph, both on 28 vertices.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Revisiting Convergence
Table: The only probable limits in growing sequences of Erdos-Renyi randomgraphs (p = 1/2) and uniform attachment graphs.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 12 / 22
Density and Completeness
TheoremLet W be the metric space ofgraphons modded out by δ� = 0.Then W is compact and the subsetG ⊆ W of graph graphons (those ofthe form WG ) is dense.
In other words...
1 Convergent sequences ofgraphs are graphons
2 Graphons are convergentsequences of graphs
3 Any sequence of graph(on)shas a limit graphon
4 extremal constructions haveone or more limts
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 13 / 22
Density and Completeness
TheoremLet W be the metric space ofgraphons modded out by δ� = 0.Then W is compact and the subsetG ⊆ W of graph graphons (those ofthe form WG ) is dense.
In other words...
1 Convergent sequences ofgraphs are graphons
2 Graphons are convergentsequences of graphs
3 Any sequence of graph(on)shas a limit graphon
4 extremal constructions haveone or more limts
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 13 / 22
Density and Completeness
TheoremLet W be the metric space ofgraphons modded out by δ� = 0.Then W is compact and the subsetG ⊆ W of graph graphons (those ofthe form WG ) is dense.
In other words...
1 Convergent sequences ofgraphs are graphons
2 Graphons are convergentsequences of graphs
3 Any sequence of graph(on)shas a limit graphon
4 extremal constructions haveone or more limts
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 13 / 22
Density and Completeness
TheoremLet W be the metric space ofgraphons modded out by δ� = 0.Then W is compact and the subsetG ⊆ W of graph graphons (those ofthe form WG ) is dense.
In other words...
1 Convergent sequences ofgraphs are graphons
2 Graphons are convergentsequences of graphs
3 Any sequence of graph(on)shas a limit graphon
4 extremal constructions haveone or more limts
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 13 / 22
Density and Completeness
TheoremLet W be the metric space ofgraphons modded out by δ� = 0.Then W is compact and the subsetG ⊆ W of graph graphons (those ofthe form WG ) is dense.
In other words...
1 Convergent sequences ofgraphs are graphons
2 Graphons are convergentsequences of graphs
3 Any sequence of graph(on)shas a limit graphon
4 extremal constructions haveone or more limts
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 13 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn
0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W
and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally,
TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
AG versus TW
AG
Given |V (G )| = n and v ∈ Rn
(AGv)k =n∑
i=1
akivi
AG : Rn → Rn0 1 1 11 0 1 01 1 0 01 0 0 0
πξχρ
=
π + ξ
TW
Given a graphon W and f a“nice” f : [0, 1]→ R
(TW f )(x) =
∫ 1
0W (x , y)f (y)dy
Formally, TW : nice→ nice
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 14 / 22
Spectral Decomposition of TW
Theorem
If W is a graphon, then its eigenvalues can be ordered as1 ≥ |λ1| ≥ |λ2| ≥ · · · so that λk → 0 as k →∞ and
TW ∼∞∑k=1
λk fk(x)fk(y) treating TW : L2[0, 1]→ L2[0, 1]
where the inner product∫ 10 fk(x)fl(y) = 1 if k = l and 0 otherwise.
Example
If W (x , y) = xy , λ1 = 1/4 andf1(x) = 2x and all other λk = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 15 / 22
Spectral Decomposition of TW
Theorem
If W is a graphon, then its eigenvalues can be ordered as1 ≥ |λ1| ≥ |λ2| ≥ · · · so that λk → 0 as k →∞ and
TW ∼∞∑k=1
λk fk(x)fk(y) treating TW : L2[0, 1]→ L2[0, 1]
where the inner product∫ 10 fk(x)fl(y) = 1 if k = l and 0 otherwise.
Example
If W (x , y) = xy ,
λ1 = 1/4 andf1(x) = 2x and all other λk = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 15 / 22
Spectral Decomposition of TW
Theorem
If W is a graphon, then its eigenvalues can be ordered as1 ≥ |λ1| ≥ |λ2| ≥ · · · so that λk → 0 as k →∞ and
TW ∼∞∑k=1
λk fk(x)fk(y) treating TW : L2[0, 1]→ L2[0, 1]
where the inner product∫ 10 fk(x)fl(y) = 1 if k = l and 0 otherwise.
Example
If W (x , y) = xy , λ1 = 1/4 andf1(x) = 2x
and all other λk = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 15 / 22
Spectral Decomposition of TW
Theorem
If W is a graphon, then its eigenvalues can be ordered as1 ≥ |λ1| ≥ |λ2| ≥ · · · so that λk → 0 as k →∞ and
TW ∼∞∑k=1
λk fk(x)fk(y) treating TW : L2[0, 1]→ L2[0, 1]
where the inner product∫ 10 fk(x)fl(y) = 1 if k = l and 0 otherwise.
Example
If W (x , y) = xy , λ1 = 1/4 andf1(x) = 2x and all other λk = 0.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 15 / 22
Exercise: From spec(G ) to spec(W )
To turn a spectral decomposition of AG into one of TW ...
1 Write AG out spectrally.
2 Replace the standard basis of Rn with weighted inticator functions.(Hint: consider indicator functions of intervals. Remember thatdouble integration over line segments produces 0. )
3 Justify that there are no more eigenvalues. (Hint: the “all zeros”graphon 0 · J has all eigenvalues 0 and hence spectral decompositionT0·J = 0.)
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 16 / 22
Exercise: From spec(G ) to spec(W )
To turn a spectral decomposition of AG into one of TW ...
1 Write AG out spectrally.
2 Replace the standard basis of Rn with weighted inticator functions.(Hint: consider indicator functions of intervals. Remember thatdouble integration over line segments produces 0. )
3 Justify that there are no more eigenvalues. (Hint: the “all zeros”graphon 0 · J has all eigenvalues 0 and hence spectral decompositionT0·J = 0.)
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 16 / 22
Exercise: From spec(G ) to spec(W )
To turn a spectral decomposition of AG into one of TW ...
1 Write AG out spectrally.
2 Replace the standard basis of Rn with weighted inticator functions.
(Hint: consider indicator functions of intervals. Remember thatdouble integration over line segments produces 0. )
3 Justify that there are no more eigenvalues. (Hint: the “all zeros”graphon 0 · J has all eigenvalues 0 and hence spectral decompositionT0·J = 0.)
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 16 / 22
Exercise: From spec(G ) to spec(W )
To turn a spectral decomposition of AG into one of TW ...
1 Write AG out spectrally.
2 Replace the standard basis of Rn with weighted inticator functions.(Hint: consider indicator functions of intervals. Remember thatdouble integration over line segments produces 0. )
3 Justify that there are no more eigenvalues. (Hint: the “all zeros”graphon 0 · J has all eigenvalues 0 and hence spectral decompositionT0·J = 0.)
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 16 / 22
Exercise: From spec(G ) to spec(W )
To turn a spectral decomposition of AG into one of TW ...
1 Write AG out spectrally.
2 Replace the standard basis of Rn with weighted inticator functions.(Hint: consider indicator functions of intervals. Remember thatdouble integration over line segments produces 0. )
3 Justify that there are no more eigenvalues.
(Hint: the “all zeros”graphon 0 · J has all eigenvalues 0 and hence spectral decompositionT0·J = 0.)
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 16 / 22
Exercise: From spec(G ) to spec(W )
To turn a spectral decomposition of AG into one of TW ...
1 Write AG out spectrally.
2 Replace the standard basis of Rn with weighted inticator functions.(Hint: consider indicator functions of intervals. Remember thatdouble integration over line segments produces 0. )
3 Justify that there are no more eigenvalues. (Hint: the “all zeros”graphon 0 · J has all eigenvalues 0 and hence spectral decompositionT0·J = 0.)
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 16 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and
spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} and
spec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} andspec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} and
spec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence
Table: WKn → J asn→∞.
Table:WKdn/2e,bn/2c →WK2 asn→∞.
Spectra?
spec(Kn) = {n − 1,−1, . . . ,−1} and spec(J) = {1, 0, . . . , 0} andspec(Kdn/2e,bn/2c) = {n/2, 0, . . . , 0,−n/2} andspec(WK2) = {1/2, 0, . . . , 0,−1/2}.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 17 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle.
Viewspec as a two sequences: nonnegative and nonpositive.
Example
spec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Example
spec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) =
{3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1,
0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1,
0, . . . , 0,
− 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�.
Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{
λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}
↓ ↓ ↓ ↓ ↓
{
λ1(W )
,
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{ λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )|
}↓
↓ ↓ ↓ ↓
{ λ1(W ),
λ2(W )
, · · · ,
λ−2(W ) λ−1(W )
}
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{ λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| ,
· · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )| }
↓
↓ ↓ ↓
↓{ λ1(W ),
λ2(W )
, · · · ,
λ−2(W )
λ−1(W ) }
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{ λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| , · · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )| }
↓ ↓
↓ ↓
↓{ λ1(W ), λ2(W ), · · · ,
λ−2(W )
λ−1(W ) }
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{ λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| , · · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )| }
↓ ↓
↓
↓ ↓{ λ1(W ), λ2(W ), · · · , λ−2(W ) λ−1(W ) }
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
Spectral Convergence Theorem
“Padded” Bi-Infinite Spectra
Pad every spectrum with an infinite number of “0”s in the middle. Viewspec as a two sequences: nonnegative and nonpositive.
Examplespec(Petersen) = {3, 1, 1, 1, 1, 1, 0, . . . , 0, − 2,−2,−2,−2}
Theorem [Borgs, Chayes, Lovasz, Sos, Vesztergombi ’12]
Suppose Gn →W in δ�. Then as k →∞
{ λ1(Gk )|V (Gk )| ,
λ2(Gk )|V (Gk )| , · · · ,
λ−2(Gk )|V (Gk )| ,
λ−1(Gk )|V (Gk )| }
↓ ↓ ↓ ↓ ↓{ λ1(W ), λ2(W ), · · · , λ−2(W ) λ−1(W ) }
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 18 / 22
The End
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 19 / 22
Applications
Triangle Removal Lemma
For all ε > 0, there exists some ε′ > 0 so that if G a graph on n verticeshas at most ε′n3 triangles, then there exists some triangle-free G ′ ⊆ Gwith e(G )− e(G ′) ≤ εn2.
Quasirandomness [Chung, Graham, Wilson ’89]
If G1,G2, . . . are graphs where Gn is εn-quasirandom (εn as small aspossible) and |Gn| → ∞, then∑
k
λk(Gn)2 → 1/2 and∑k
λk(Gn)4 → 1/16.
if and only if εn →∞.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 20 / 22
More Applications
Many Strong Szemeredi Regularity Lemmas for Graphs and Graphons
All large graphs (all graphons) may be approxmated in cut distance witharbitrary pre-specified precision by a random graph (graphon) which isgiven by an equipartition.
“Disguises” of Graphons
The following models are cryptomorphic (i.e., the same information):
1 a graphon, up to weak isomorphism
2 a multiplicative, normalized simple graph parameter that isnonnegative on signed graphs
3 a consistent and local graph model
4 a local random countable graph model
5 a point in the completion of the space of finite graphs with the cutdistance
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 21 / 22
References
A. Frieze, R. Kannan (1999)
Quick approximation to matrices and applications
Combinatorica 19(3), 175 – 220.
Lovasz (2012)
Large networks and graph limits
Providence: American Mathematical Society 60.
Alexander W. N. Riasanovsky (ISU) Spectra of Graphons April 19, 2017 22 / 22