Ramanujan Graphs of Every Degree Adam Marcus (Crisply, Yale)
Daniel Spielman (Yale) Nikhil Srivastava (MSR India)
Slide 2
Expander Graphs Sparse, regular well-connected graphs with many
properties of random graphs. Random walks mix quickly. Every small
set of vertices has many neighbors. Pseudo-random generators.
Error-correcting codes. Sparse approximations of complete graphs.
Major theorems in Theoretical Computer Science.
Slide 3
Spectral Expanders Let G be a graph and A be its adjacency
matrix a c d e b 0 1 0 0 1 1 0 1 0 1 0 1 0 1 0 0 0 1 0 1 1 1 0 1 0
Eigenvalues If d -regular (every vertex has d edges), trivial
Slide 4
Spectral Expanders If bipartite (all edges between two
parts/colors) eigenvalues are symmetric about 0 trivial a cd f b e
If d -regular and bipartite, 0 0 0 1 0 1 0 0 0 1 1 0 0 0 0 0 1 1 1
1 0 0 0 0 0 1 1 0 0 0 1 0 1 0 0 0
Slide 5
Spectral Expanders 0 [] -d d G is a good spectral expander if
all non-trivial eigenvalues are small
Slide 6
Bipartite Complete Graph Adjacency matrix has rank 2, so all
non-trivial eigenvalues are 0 a cd f b e 0 0 0 1 1 1 1 1 1 0 0
0
Slide 7
Spectral Expanders 0 [] -d d Challenge: construct infinite
families of fixed degree G is a good spectral expander if all
non-trivial eigenvalues are small
Slide 8
Spectral Expanders G is a good spectral expander if all
non-trivial eigenvalues are small 0 [] -d d Challenge: construct
infinite families of fixed degree Alon-Boppana 86: Cannot beat )
(
Slide 9
G is a Ramanujan Graph if absolute value of non-trivial eigs 0
[] -d d ) ( Ramanujan Graphs:
Slide 10
G is a Ramanujan Graph if absolute value of non-trivial eigs 0
[] -d d ) ( Ramanujan Graphs: Margulis, Lubotzky-Phillips-Sarnak88:
Infinite sequences of Ramanujan graphs exist for
Slide 11
G is a Ramanujan Graph if absolute value of non-trivial eigs 0
[] -d d ) ( Ramanujan Graphs: Margulis, Lubotzky-Phillips-Sarnak88:
Infinite sequences of Ramanujan graphs exist for Can be quickly
constructed: can compute neighbors of a vertex from its name
Slide 12
G is a Ramanujan Graph if absolute value of non-trivial eigs 0
[] -d d ) ( Ramanujan Graphs: Friedman08: A random d -regular graph
is almost Ramanujan :
Slide 13
Ramanujan Graphs of Every Degree Theorem: there are infinite
families of bipartite Ramanujan graphs of every degree.
Slide 14
Theorem: there are infinite families of bipartite Ramanujan
graphs of every degree. And, are infinite families of (c,d)
-biregular Ramanujan graphs, having non-trivial eigs bounded by
Ramanujan Graphs of Every Degree
Slide 15
Find an operation that doubles the size of a graph without
creating large eigenvalues. 0 [] -d d Bilu-Linial 06 Approach (
)
Slide 16
0 [] -d d Bilu-Linial 06 Approach Find an operation that
doubles the size of a graph without creating large eigenvalues. (
)
Slide 17
2-lifts of graphs a c d e b
Slide 18
a c d e b a c d e b duplicate every vertex
Slide 19
2-lifts of graphs duplicate every vertex a0a0 c0c0 d0d0 e0e0
b0b0 a1a1 c1c1 d1d1 e1e1 b1b1
Slide 20
2-lifts of graphs a0a0 c0c0 d0d0 e0e0 b0b0 a1a1 d1d1 e1e1 b1b1
c1c1 for every pair of edges: leave on either side (parallel), or
make both cross
Slide 21
2-lifts of graphs for every pair of edges: leave on either side
(parallel), or make both cross a0a0 c0c0 d0d0 e0e0 b0b0 a1a1 d1d1
e1e1 b1b1 c1c1
Eigenvalues of 2-lifts (Bilu-Linial) Given a 2-lift of G,
create a signed adjacency matrix A s with a -1 for crossing edges
and a 1 for parallel edges 0 -1 0 0 1 -1 0 1 0 1 0 1 0 -1 0 0 0 -1
0 1 1 1 0 1 0 a0a0 c0c0 d0d0 e0e0 b0b0 a1a1 d1d1 e1e1 b1b1
c1c1
Slide 26
Eigenvalues of 2-lifts (Bilu-Linial) Theorem: The eigenvalues
of the 2-lift are the union of the eigenvalues of A (old) and the
eigenvalues of A s (new) 0 -1 0 0 1 -1 0 1 0 1 0 1 0 -1 0 0 0 -1 0
1 1 1 0 1 0 a0a0 c0c0 d0d0 e0e0 b0b0 a1a1 d1d1 e1e1 b1b1 c1c1
Slide 27
Eigenvalues of 2-lifts (Bilu-Linial) Conjecture: Every d
-regular graph has a 2-lift in which all the new eigenvalues have
absolute value at most Theorem: The eigenvalues of the 2-lift are
the union of the eigenvalues of A (old) and the eigenvalues of A s
(new)
Slide 28
Eigenvalues of 2-lifts (Bilu-Linial) Conjecture: Every d
-regular graph has a 2-lift in which all the new eigenvalues have
absolute value at most Would give infinite families of Ramanujan
Graphs: start with the complete graph, and keep lifting.
Slide 29
Eigenvalues of 2-lifts (Bilu-Linial) Conjecture: Every d
-regular graph has a 2-lift in which all the new eigenvalues have
absolute value at most We prove this in the bipartite case. a
2-lift of a bipartite graph is bipartite
Slide 30
Eigenvalues of 2-lifts (Bilu-Linial) Theorem: Every d -regular
graph has a 2-lift in which all the new eigenvalues have absolute
value at most Trick: eigenvalues of bipartite graphs are symmetric
about 0, so only need to bound largest
Slide 31
Eigenvalues of 2-lifts (Bilu-Linial) Theorem: Every d -regular
bipartite graph has a 2-lift in which all the new eigenvalues have
absolute value at most
Slide 32
First idea: a random 2-lift Specify a lift by Pick s uniformly
at random
Slide 33
First idea: a random 2-lift Specify a lift by Pick s uniformly
at random Are graphs for which this usually fails
Slide 34
First idea: a random 2-lift Specify a lift by Pick s uniformly
at random Are graphs for which this usually fails Bilu and Linial
proved G almost Ramanujan, implies new eigenvalues usually small.
Improved by Puder and Agarwal-Kolla-Madan
Slide 35
The expected polynomial Consider
Slide 36
The expected polynomial Consider Conclude there is an s so that
Prove Prove is an interlacing family
Slide 37
The expected polynomial Theorem (Godsil-Gutman 81): the
matching polynomial of G
Slide 38
The matching polynomial (Heilmann-Lieb 72) m i = the number of
matchings with i edges
Slide 39
Slide 40
one matching with 0 edges
Slide 41
7 matchings with 1 edge
Slide 42
Slide 43
Slide 44
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations
Slide 45
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations same edge: same
value
Slide 46
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations same edge: same
value
Slide 47
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations Get 0 if hit any
0s
Slide 48
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations Get 0 if take just
one entry for any edge
Slide 49
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations Only permutations
that count are involutions
Slide 50
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations Only permutations
that count are involutions
Slide 51
Proof that x 1 0 0 1 1 1 x 1 0 0 0 0 1 x 1 0 0 0 0 1 x 1 0 1 0
0 1 x 1 1 0 0 0 1 x Expand using permutations Only permutations
that count are involutions Correspond to matchings
Slide 52
The matching polynomial (Heilmann-Lieb 72) Theorem
(Heilmann-Lieb) all the roots are real
Slide 53
The matching polynomial (Heilmann-Lieb 72) Theorem
(Heilmann-Lieb) all the roots are real and have absolute value at
most
Slide 54
The matching polynomial (Heilmann-Lieb 72) Theorem
(Heilmann-Lieb) all the roots are real and have absolute value at
most Implies
Slide 55
Interlacing Polynomial interlaces if
Slide 56
Common Interlacing and have a common interlacing if can
partition the line into intervals so that each interval contains
one root from each poly
Slide 57
Common Interlacing and have a common interlacing if can
partition the line into intervals so that each interval contains
one root from each poly ) ) ) ) ( (( (
Slide 58
If p 1 and p 2 have a common interlacing, for some i. Largest
root of average Common Interlacing
Slide 59
If p 1 and p 2 have a common interlacing, for some i. Largest
root of average Common Interlacing
Slide 60
is an interlacing family If the polynomials can be placed on
the leaves of a tree so that when put average of descendants at
nodes siblings have common interlacings Interlacing Family of
Polynomials
Slide 61
is an interlacing family Interlacing Family of Polynomials If
the polynomials can be placed on the leaves of a tree so that when
put average of descendants at nodes siblings have common
interlacings
Slide 62
Interlacing Family of Polynomials There is an s so that
Theorem:
Slide 63
An interlacing family Theorem: Let is an interlacing
family
Slide 64
Interlacing and have a common interlacing iff is real rooted
for all
Slide 65
To prove interlacing family Let
Slide 66
To prove interlacing family Let Need to prove that for all, is
real rooted
Slide 67
To prove interlacing family Let are fixed is 1 with
probability, -1 with are uniformly Need to prove that for all, is
real rooted
Slide 68
Generalization of Heilmann-Lieb We prove is real rooted for
every independent distribution on the entries of s
Slide 69
Generalization of Heilmann-Lieb We prove is real rooted for
every independent distribution on the entries of s By using mixed
characteristic polynomials
Slide 70
Mixed Characteristic Polynomials For independently chosen
random vectors is their mixed characteristic polynomial. Theorem:
Mixed characteristic polynomials are real rooted. Proof: Using
theory of real stable polynomials.
Slide 71
Mixed Characteristic Polynomials For independently chosen
random vectors is their mixed characteristic polynomial. Obstacle:
our matrix is a sum of random rank-2 matrices 0 -1 -1 0 0 1 1 0
or
Slide 72
Mixed Characteristic Polynomials For independently chosen
random vectors is their mixed characteristic polynomial. Solution:
add to the diagonal 1 -1 -1 1 1 1 or
Slide 73
Generalization of Heilmann-Lieb We prove is real rooted for
every independent distribution on the entries of s Implies is an
interlacing family
Slide 74
Generalization of Heilmann-Lieb We prove is real rooted for
every independent distribution on the entries of s Conclude there
is an s so that Implies is an interlacing family
Slide 75
Universal Covers The universal cover of a graph G is a tree T
of which G is a quotient. vertices map to vertices edges map to
edges homomorphism on neighborhoods Is the tree of non-backtracking
walks in G. The universal cover of a d -regular graph is the
infinite d -regular tree.
Slide 76
Quotients of Trees d -regular Ramanujan as quotient of infinite
d -ary tree Spectral radius and norm of inf d -ary tree are
Slide 77
Godsils Proof of Heilmann-Lieb T(G,v) : the path tree of G at v
vertices are paths in G starting at v edges to paths differing in
one step
Slide 78
Godsils Proof of Heilmann-Lieb a c d e b a a b e a e a b e a b
c a b d e a
Slide 79
T(G,v) : the path tree of G at v vertices are paths in G
starting at v edges to paths differing in one step Theorem: The
matching polynomial divides the characteristic polynomial of
T(G,v)
Slide 80
Godsils Proof of Heilmann-Lieb T(G,v) : the path tree of G at v
vertices are paths in G starting at v edges to paths differing in
one step Theorem: The matching polynomial divides the
characteristic polynomial of T(G,v) Is a subgraph of infinite tree,
so has smaller spectral radius
Slide 81
Quotients of Trees (c,d) -biregular bipartite Ramanujan as
quotient of infinite (c,d) -ary tree Spectral radius For (c,d)
-regular bipartite Ramanujan graphs
Slide 82
Irregular Ramanujan Graphs (Greenberg-Lubotzky) Def: G is
Ramanujan if its non-trivial eigenvalues have abs value less than
the spectral radius of its cover Theorem: If G is bipartite and
Ramanujan, then there is an infinite family of Ramanujan graphs
with the same cover.
Slide 83
Questions Non-bipartite Ramanujan Graphs of every degree?
Efficient constructions? Explicit constructions?