0% found this document useful (0 votes)
19 views17 pages

2405 20368v1

Uploaded by

gamininganela
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
19 views17 pages

2405 20368v1

Uploaded by

gamininganela
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH

HONGLIN ZHU
arXiv:2405.20368v1 [math.CO] 30 May 2024

Abstract. We introduce a new notion of error-correcting codes on [q]n where a code is a


set of proper q-colorings of some fixed n-vertex graph G. For a pair of proper q-colorings
X, Y of G, we define their distance as the minimum Hamming distance between X and σ(Y )
over all σ ∈ Sq . We then say that a set of proper q-colorings of G is δ-distinct if any pair of
colorings in the set have distance at least δn.
We investigate how one-sided spectral expansion relates to the largest possible set of
δ-distinct colorings on a graph. For fixed (δ, λ) ∈ [0, 1] × [−1, 1] and positive integer d, let
fδ,λ,d (n) denote the maximal size of a set of δ-distinct colorings of any d-regular graph on at
most n vertices with normalized second eigenvalue at most λ. We study the growth of f as
n goes to infinity. We partially characterize regimes of (δ, λ) where f grows exponentially, is
finite, and is at most 1, respectively. We also prove several sharp phase transitions between
these regimes.

1. Introduction
A central topic in coding theory is the study of error-correcting codes of a fixed length
over some alphabet. Given the word length n, the alphabet [q], and the minimum distance
d, the goal is to find a large subset of [q]n where the Hamming distance between any two
codewords is at least d. This can be viewed as a sphere packing problem on [q]n equipped
with the Hamming distance.
We extend this notion to introduce the sphere packing problem of proper q-colorings of a
graph. In particular, given an n vertex graph G, we identify the set of vertex colorings of G
using q colors with [q]n /Sq ; that is, the set [q]n with the equivalence relation (x1 , . . . , xn ) ≡
(σ(x1 ), . . . , σ(xn )) for all σ ∈ Sq . Now, given a minimum distance d, we would like to find a
large set of vertex colorings where the distance between any two colorings is at least d, and
we further require that each coloring is a proper coloring of the graph G.
In this paper, we investigate the relationship between expansion and this new sphere
packing problem. Expander graphs play an important role in combinatorics and theoretical
computer science. They also appear in the study of error-correcting codes, such as in the
seminal work of Sipser and Spielman [10], who construct good error-correcting codes using
expander graphs. By establishing a new connection between these subjects, we relate spectral
expansion, an algebraic quantity, with the geometry of the space of proper q-colorings of a
graph.

1.1. Setup. Fix an integer q ≥ 3. Let G = (V, E) be a graph on n vertices. A proper


q-coloring of G is a map X : V → [q] such that X(u) 6= X(v) whenever uv ∈ E. For the
rest of the paper, unless otherwise specified, we will simply refer to a proper q-coloring as
a coloring. Identifying V with [n], we can view X as an element of [q]n . For two colorings
1
2 HONGLIN ZHU

X, Y of G, we define their distance as


d(X, Y ) = min dH (X, σ(Y )) = min |{i ∈ [n]|X(i) 6= σ(Y (i))}|.
σ∈Sq σ∈Sq

This is simply the minimum Hamming distance between X and σ(Y ). We say that a set of
colorings of G is δ-distinct if any pair of colorings in the set have distance at least δn.
Throughout the rest of this paper, we only work with regular graphs unless otherwise
specified. Given a d-regular graph G, let A denote its normalized adjacency matrix, with
entries given by (
1
d
if uv ∈ E
Auv =
0 otherwise.
Let λ1 ≥ λ2 ≥ · · · ≥ λn denote its eigenvalues. Recall that λ1 = 1 and λn ≥ −1.
For any 0 ≤ δ ≤ 1, let CG,q,δ denote the (finite) collection of δ-distinct sets of q-colorings of
G. For positive integers n and d and real number −1 ≤ λ ≤ 1, let Gn,d,λ denote the (finite)
set of d-regular graphs on at most n vertices with λ2 ≤ λ. For convenience, we allow for the
empty graph on n vertices, which we regard as a 0-regular graph with λ2 = 1.
For any (δ, λ) ∈ [0, 1] × [−1, 1] and d ≥ 0, we define
fδ,λ,d(n) = max max |C|,
G∈Gn,d,λ C∈CG,q,δ

where the maximum outputs 0 if it is taken over an empty set. For d = 0, λ = 1, finding this
quantity can be viewed as a variant of the sphere packing or error-correcting code problem on
[q]n , where we want to find a maximal C ⊂ [q]n such that any X, Y, ∈ C have d(X, Y ) ≥ δn.
In general, C comes with the extra constraint that it must be a valid set of q-colorings of
some d-regular n-vertex graph G with λ2 ≤ λ.
The main subject of study of this paper is the asymptotic behavior of f when we fix δ
and λ. Observe that d(X, Y ) √ ≤ (1 − 1/q)n. Also, the Alon-Boppana bound [9] states that
d-regular graphs have λ2 ≥ 2 d − 1/d + o(1) > 0. Therefore, for the rest of the paper we
restrict our attention to the nontrivial region (δ, λ) ∈ [0, 1 − 1/q] × (0, 1].
Remark 1.1. We focus on one-sided expanders (graphs with λ2 ≤ λ) instead of two sided
expanders (graphs with max{|λ2 |, |λn |} ≤ λ). This is due to Hoffman’s bound [6] on the
chromatic number of a two-sided expander:
1 1
χ(G) ≥ 1 − ≥ 1+ .
λn λ
1
In particular, this implies that a two-sided expander with λ < q−1 cannot be q-colorable.
Nevertheless, we leave as a future direction to investigate the problem for two-sided expanders
above this bound.
Next, we define the relevant regimes of f which we study. We say that (δ, λ) is in the
exponential regime, denoted Rexp , if
log fδ,λ,d (n)
sup lim sup > 0.
d≥0 n→∞ n
We say that (δ, λ) is in the constant regime, denoted Rcon , if
sup lim sup fδ,λ,d (n) < ∞.
d≥0 n→∞
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 3

Furthermore, we say that (δ, λ) is in the unique regime, denoted Runi , if


sup lim sup fδ,λ,d(n) = 1.
d≥0 n→∞

A simple yet useful observation is the monotonicity of f .


Lemma 1.2. For fixed n and d, the function fδ,λ,d(n) is weakly decreasing in δ and weakly
increasing in λ.
As a corollary of Lemma 1.2, if (δ, λ) is in the exponential regime, then any (δ ′ , λ′ ) with
δ ≤ δ and λ′ ≥ λ is in the exponential regime. Similarly, if (δ, λ) is in the constant/unique

regime, then any (δ ′ , λ′ ) with δ ′ ≥ δ and λ′ ≤ λ is in the constant/unique regime.

1.2. Main Results. The main results of this paper establish a landscape for the three
different growth regimes for the function f (see Figure 1). First, we construct an infinite
family of expander graphs that have an exponential number of δ-distinct colorings for any
δ < 1 − 1/q, which is the largest possible.

f decreasing
1
Rexp
Theorem 1.3
0.9999

Rexp
Theorem 1.4 ? λ

1
4

Runi
Rcon Theorem 1.6
Theorem 1.5
0
0 1 1 2
4 δ 2 3

Figure 1. Schematic diagram for the three regimes when q = 3. The curves
bounding the regimes may not be precise. Note that the boundary of the
exponential regime at ( 14 , 0) has a positive slope. The growth of f in the
central region is unknown.

Theorem 1.3. There exists some universal constant κ > 0 such that any (δ, 1 − κ) with
δ < 1 − 1/q is in the exponential regime.
We also show that bipartite expanders with arbitrarily small λ2 can have an exponential
number of δ-distinct colorings for a certain positive value of δ.
4 HONGLIN ZHU
   
Theorem 1.4. For any λ > 0, there exists ǫ > 0 such that 1 − 21 ⌊ 1q ⌋ + 1
⌈ q2 ⌉
+ ǫ, λ is in
2
the exponential regime.
Conversely, we show a sharp phase transition between the exponential regime and the
constant regime at the threshold above for λ → 0. In other words, for graphs with nearly
perfect spectral expansion, increasing δ even slightly above the threshold in Theorem 1.4
bounds the number of possible δ-distinct colorings by a constant.
   
Theorem 1.5. For any ǫ > 0, there exists some λ = λ(ǫ) > 0 such that 1 − 12 ⌊ 1q ⌋ + ⌈ 1q ⌉ + ǫ, λ
2 2
is in the constant regime.
Finally, we also establish sharp boundary points for the unique regime.
 
1
Theorem 1.6. For any ǫ > 0, there exists some λ = λ(ǫ) > 0 such that 1 − q−1 + ǫ, λ
 
1
is in the unique regime. Conversely, 1 − q−1 , λ is not in the unique regime for all λ > 0.
   
1
For any ǫ > 0, 1 − 1q , (q−1)2 − ǫ is in the unique regime. Conversely, 1 − 1
, 1
q (q−1)2
is not
in the constant regime.
1.3. Outline. The rest of the paper is organized as follows. In §2, we recall some preliminary
results and prove Lemma 1.2. In §3, we study the exponential regime and prove Theorems 1.3
and 1.4. In §4, we study the constant regime and prove Theorem 1.5. In §5, we study the
unique regime and prove Theorem 1.6. In §6, we discuss some possible future directions.

Acknowledgements. This work was done as part of a direct funding project in the Un-
dergraduate Research Opportunities Program at MIT. The author would like to thank Yufei
Zhao for suggesting the problem and for his mentorship in this project. We also thank Mi-
tali Bafna for sharing her research which inspired this work. We thank Mitali Bafna and
Yufei Zhao for helpful discussions. We thank Sammy Luo and Yufei Zhao for valuable com-
ments which improved this paper. Finally, we thank the MIT UROP office and the MIT
Department of Mathematics for providing this research opportunity.

2. Preliminaries
In this paper, we deal with graphs with numbers of vertices going to infinity. Sometimes
it is convenient to suppress the size of the graph and measure vertex or edge subsets by
the fraction of vertices/edges they contain. In particular, this makes it conceptually more
intuitive when we disregard small parts of the graph. For vertex subsets A, B of a graph
G = (V, E), write w(A) = |A|/|V |, e(A) = |E(A)|/|E|, and e(A, B) = |E(A, B)|/|E|.
We first prove Lemma 1.2.
Proof of Lemma 1.2. Fixing n, d, λ, we see that f weakly decreases in δ as the collection
CG,q,δ shrinks as δ increases. Similarly, fixing n, d, δ, we see that f weakly increases in λ as
the collection Gn,d,λ grows as λ increases. 
Next, we recall several general preliminary results we use in later sections. First, recall
the Courant-Fischer characterization of the second eigenvalue.
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 5

Lemma 2.1. Let G = (V, E) be a graph and let A be its normalized adjacency matrix. Then
the second eigenvalue λ2 of A satisfies
hx, Axi
λ2 = max .
x6=0 hx, xi
x⊥1
Also recall Cheeger’s inequality, which relates spectral expansion with edge expansion.
Lemma 2.2 ([4]). Let G = (V, E) be a d-regular graph and let
|E(S, V \ S)|
h = h(G) = min
S⊂V |S|
0<|S|≤|V |/2

be its edge expansion ratio. Then


d(1 − λ2 ) p
≤ h ≤ d 2(1 − λ2 ).
2
A main strategy we use in constructions in §3 is random coloring. For this purpose,
we recall McDiarmid’s inequality, also known as the bounded difference inequality, which
provides a concentration bound on Lipschitz functions of independent random inputs.
Lemma 2.3 ([8]). Let X1 ∈ Ω1 , . . . , Xn ∈ Ωn be independent random variables. Suppose
f : Ω1 × · · · × Ωn → R is c-Lipschitz; that is,
|f (x1 , . . . , xn ) − f (x′1 , . . . , x′n )| ≤ c
whenever (x1 , . . . , xn ) and (x′1 , . . . , x′n ) only differ on one coordinate. Then for any t ≥ 0,
−2t2
P[f (X1 , . . . , Xn ) ≤ E[f (X1 , . . . , Xn )] − t] ≤ e c2 n
and
−2t2
P[f (X1 , . . . , Xn ) ≥ E[f (X1 , . . . , Xn )] + t] ≤ e c2 n
The following two results provide us with bipartite d-regular graphs that have “optimal
expansion”. However, for our purposes, we only need graphs with sufficiently good expansion.
We use these results mainly for concreteness in the computation of implicit constants in our
proofs.
Lemma 2.4 ([3]). Let G be a uniformly random d-regular bipartite graph on n vertices (n
is even if d is odd). Then there exist a sequence of ǫn → 0 such that

λ2 ≤ 2 d − 1/d + ǫn
asymptotically almost surely as n → ∞.

The celebrated Alon-Boppana bound [9] states that asymptotically λ2 ≥ 2 d − 1/d+o(1).
Thus, the above result says that almost all bipartite d-regular graphs approach this bound.
A similar result for random d-regular graphs is proved by Friedman in [5].
Another seminal result by Marcus, Spielman, and Srivastava [7] establishes the existence
of d-regular bipartite graphs that achieve the Alon-Boppana bound, known as bipartite
Ramanujan graphs.
Lemma 2.5 ([7]). For each d ≥ 3, there exists an infinite sequence of d-regular bipartite
graphs with √
λ2 ≤ 2 d − 1/d.
6 HONGLIN ZHU

3. The exponential regime


In this section, we give constructions that prove Theorems 1.3 and 1.4.
3.1. Construction for Theorem 1.3. Fix δ < 1 − 1/q and let ǫ = 1 − 1/q − δ. Let
α = 10−4. To prove Theorem 1.3, we construct an infinite sequence of 3-regular graphs Gn
with λ2 (Gn ) ≤ 1 − α such that for sufficiently large n, each graph Gn has a δ-distinct set
of q-colorings of size at least ec|V (Gn )| , for some constant c = c(ǫ) to be specified. Such a
construction shows that (δ, 1 − α) ∈ Rexp .

Hn Gn

x y x y

Kx,y ∼ −
= K3,3

Figure 2. Illustration of the construction of Gn from Hn .



First, let Hn be a sequence of 3-regular expander graphs with λ2 (Hn ) ≤ 2 2/3 given by

Lemma 2.5. Let K3,3 denote the complete bipartite graph K3,3 with one edge removed and
let u, v be the two endpoints of the removed edge. We construct Gn as follows. Start with
the graph Hn . For each edge xy ∈ E(Hn ), we add a copy of Kxy ∼ −
= K3,3 to the vertex set of
Gn . Remove the edge xy and add edges xu and yv. In other words, we replace each edge of

Hn with a K3,3 “gadget” that is connected to the two endpoints of the edge (see Figure 2).
It is easy to see that Gn is still 3-regular. For each edge of Hn , we added 6 vertices. Thus,
Gn has 10 · |V (Hn )| vertices. Write H = Hn , G = Gn , and let N = |V (Hn )|. We show that
G has the desired properties.
Lemma 3.1. Let G = Gn be the graph defined as above. Then λ2 ≤ 1 − α.
Proof. For the sake of contradiction suppose that there exists some S ⊂ V = V with |S| ≤
N/2 such that
|E(S, V \ S)| √
< 3 2α.
|S|
For each gadget Kxy , √ if 0 < |S ∩ Kxy | < 6, then this gadget contributes at least 2 edges
to E(S, V \ S). Since 3 2α < 2/5, this means in particular that S cannot be a union of
such “incomplete gadgets.” Now let S ′ be obtained from S by deleting vertices of S in each
incomplete gadget. Then S ′ is nonempty and each incomplete gadget reduces E(S ′ , V \ S ′ )
by at least 1 and |S ′ | by at most 5. If k is the number of incomplete gadgets in S, then
|E(S ′, V \ S ′ )| |E(S, V \ S)| − k |E(S, V \ S)|

≤ ≤
|S | |S| − 5k |S|
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 7

since 3 2α < 1/5. Thus, we may assume that S has no incomplete gadgets.
Let k1 be the number of gadgets Kxy ⊂ S such that x, y ∈ S. Let k2 be the number of
gadgets Kxy ⊂ S such that exactly one of x, y is in S. Let k1 be the number of gadgets
Kxy ⊂ S such that x, y ∈
/ S. Then letting SH = S ∩ H,
|E(S, V \ S)| 3|SH | − 2k1 + 2k3 3|SH | − 2k1
= ≥ .
|S| |SH | + 6k1 + 6k2 + 6k3 |SH | + 6k1 + 6k2
Thus, we can assume k3 = 0.
Suppose |SH | ≥ 3N/5. Since |S| = |SH | + 6k1 + 6k2 ≤ |V |/2 = 5N, we have that
10
|E(S, V \ S)| 3
− 53 N
|SH | 1 √
≥ ≥ ≥ 3 2α.
|S| 5N 15
Thus we must have |SH | ≤ 3N/5. Let SH = VH \ SH . Observe that k1 ≤ EH (SH ) and
3|SH | − 2k1 = 2EH (SH ) + EH (SH ) − 2k1 ≥ EH (SH , SH ) ≥ k2 .
Then
|E(S, V \ S)| 3|SH | − 2k1
=
|S| |SH | + 6k1 + 6k2
EH (SH , SH )

19|SH |
 
h(H) |SH |
≥ min 1,
19 |SH |
2h(H)

57
√ √
By our choice of H and Lemma 2.2, this is at least ≥ 2572 ≥ 3 2α. Thus, no such S exists
and √
h(G) ≥ 3 2α.
By Lemma 2.2, this implies that λ2 ≥ 1 − α. 
Lemma 3.2. Let G = Gn be the graph defined as above with |V (G)| = 10N. Then for n
sufficiently large, G has a δ-distinct set of q-colorings of size at least ecN , where c = ǫ2 /8
and ǫ = 1 − 1/q − δ.
S
Proof. Recall that V (G) = V (H) ∪ xy∈E(H) V (Kxy ), where the gadget Kxy is a copy of

K3,3 . Consider the following random q-coloring of G. First color each vertex in V (H)
independently and uniformly at random. For each xy ∈ E(H), we have two cases. If x and y
are assigned the same color, say i, then color one part of Kxy by i + 1 mod q and the other
part by i + 2 mod q. If x and y are assigned different colors i and j respectively, then color
the part of Kxy with an edge to x by j and the other part by i.
Observe that under this random coloring, the marginal distribution of the color of any
vertex is uniform. Thus, for X, Y independent samples of this random coloring, we have
 
1
E[d(X, Y )] = min E[dH (X, σ(Y ))] = 1 − |V (G)|.
σ∈Sq q
8 HONGLIN ZHU

For any σ ∈ Sq , let x1 , . . . , xN be the colors of the vertices in H under X and let y1 , . . . , yN
be the colors of the vertices in H under σ(Y ). Since the coloring is uniquely determined by
the colors of vertices in H, we can write
dH (X, σ(Y )) = fσ (x1 , . . . , xN , y1, . . . , yN ).
Observe that changing one input of f changes the colors of at most 1 + 3 · 6 = 19 vertices;
namely, one vertex in H and the vertices of the three gadgets adjacent to it. Thus, fσ is
19-Lipschitz. By Lemma 2.3,
   2 
−2(10ǫN)2 −ǫ N
P[dH (X, σ(Y )) ≤ δ|V (G)|] ≤ exp 2
≤ exp .
19 · 2N 3.9
By union bound, we have that
 
−ǫ2 N
P[d(X, Y ) ≤ δ|V (G)|] ≤ q! exp .
3.9
By union bound again, for a collection C of ecN i.i.d. samples of the random colorings where
c = ǫ2 /8,
 2 
2 −ǫ N
P[C is δ-distinct] ≥ 1 − |C| q! exp >0
3.9
for N sufficiently large. 
Proof of Theorem 1.3. Lemmas 3.1 and 3.2 together show that (δ, 1 − α) ∈ Rexp for α =
10−4 . 

3.2. Construction for Theorem 1.4. For any λ > 0, choose d with 2/ d ≤ λ. Let
G = Gn be a sequence of uniformly random bipartite d-regular graphs on 2n vertices. By
Lemma 2.4, with probability 1 − o(1), each Gn has λ2 ≤ λ.
Consider the following random coloring of G. Let τ > 0 be a parameter to be determined.
Let V (G) = [2n], where [n] is one part of the graph. For each i ∈ [n], independently color
i by color q with probability τ , and by each color c ≤ ⌊ 2q ⌋ with the same probability. For
each i ∈ [2n] \ [n], color i by color q − 1 if it is a neighbor of a vertex colored q. Otherwise,
color i by a color c > ⌊ 2q ⌋ uniformly and independently at random. Let Q ∈ [n] be the set
of vertices in the first part colored q. If we disregard Q and N(Q), the coloring is just the
uniform random coloring of one part with the first ⌊ 2q ⌋ colors and the other part with the
remaining colors. 
1 1 1
Let δ = 1 − 2 ⌊ q ⌋ + ⌈ q ⌉ . We show that two independent samples of this random coloring
2 2
have distance exceeding δ with exponentially high probability.
Lemma 3.3. Let G be a d-regular bipartite graph on 2n vertices. Let X and Y be independent
samples of the random coloring of G described above. Then there exists some parameter τ > 0
for the random coloring and constant c = c(d) > 0 such that
   
1
P d(X, Y ) < δ + 2
2n ≤ e−cn
32d
for sufficiently large n.
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 9

Proof. let QX and QY denote the sets of vertices in [n] colored q by X and Y respectively.
Let E1 denote the event that |QX ∩ QY | ≥ 2τ 2 n. Then by Lemma 2.3,
4
P[E1 ] ≤ e−2τ n .
Let A = N(QX ) ∩ N(QY ) and B = N(QX ∩ QY ). Each vertex i ∈ [2n] \ [n] with i ∈ / B is
2
in A with probability at most (dτ ) . Indeed, i ∈ N(QX ) \ B and i ∈ N(QY ) \ B both occur
with probability at most dτ and these events are negatively correlated. Now let E2 denote
the event that |A| ≥ |B| + 2d2 τ 2 n. Then by Lemma 2.3,
4τ 4n
P[E2 ] ≤ e−2d .
Now we break down dH (X, Y ). A vertex in [n] \ QX ∩ QY is assigned different colors by X
and Y with probability 1 − ⌊ 1q ⌋ . A vertex in ([2n] \ [n]) \ A is assigned different colors by
2
1
X and Y with probability 1 − ⌈ q2 ⌉
. A vertex in QX △QY is always assigned different colors.
Therefore,
   
1 1
E[dH (X, Y )] ≥ (n − |QX ∩ QY |) 1 − q + (n − |A|) 1 − q + 2ǫn − |QX ∩ QY |.
⌊2⌋ ⌈2⌉
Now if E1 and E2 both do not occur, then |QX ∩QY | ≤ 2τ 2 n and |A| ≤ |B|+2d2τ 2 n ≤ 3d2 τ 2 n.
Then for τ = 8d12 ,

E[dH (X, Y )|E1 , E2 ] ≥ δ · 2n + τ · 2n 1 − 2τ − 3d2 τ
 
1
≥ δ+ · 2n.
16d2
By Lemma 2.3 again,
   
1 n n
P dH (X, Y ) ≥ δ + 2n ≤ P[E1 ] + P[E2 ] + e− 128d2 ≤ e− 212 d8 .
32d2
Observe that the same argument applies to dH (X, σ(Y )) for all σ. Thus, by union bound,
   
1 n n
P d(X, Y ) ≥ δ + 2
2n ≤ q!e− 212 d8 ≤ e− 213 d8 .
32d
for sufficiently large n. 

Now we are in a position to prove Theorem 1.4.


1

Proof of Theorem 1.4. Let ǫ = 32d 2 where we chose d such that 2/ d ≤ λ. Let τ be the
parameter such that Lemma 3.3 holds. For a collection C of i.i.d. samples of the random
colorings of G as described above, union bound and Lemma 3.3 gives
P[C is (δ + ǫ)-distinct] ≥ 1 − |C|2 e−cn .
We also need G to have λ2 ≤ λ. This happens with probability 1 − o(1) by Lemma 2.4.
Therefore, for all sufficiently large n, we can find some bipartite d-regular graph G on 2n
vertices with λ2 ≤ λ that has a (δ + ǫ)-distinct set of colorings of exponential size. 
10 HONGLIN ZHU

4. The constant regime


Theorem
 1.4 shows that the exponential regime covers all points (δ, λ) satisfying δ ≤
1 − 2 ⌊ q ⌋ + ⌈ 1q ⌉ + ǫ for some ǫ = ǫ(λ). The threshold here is roughly achieved by bipartite
1 1
2 2
graphs, coloring one part randomly using half of the colors and the other part randomly
using the rest of the colors. We show in this section that this threshold is essentially tight,
and furthermore that there is a sharp phase transition at λ = 0 from the exponential regime
to the constant regime. However, Theorem 1.4 and Theorem 1.5 still leave a gap (which
vanishes as λ → 0) between the exponential regime and the constant regime and it is not
known whether a different growth behavior occurs within this gap (see Figure 1).
In fact, we prove the following stronger result, which implies Theorem 1.5.
Theorem 4.1. There exists a constant K = K(q) such thatthe following  holds. For any
1 1 1
ǫ > 0, there exists some λ = λ(ǫ) > 0 such that for δ = 1 − 2 ⌊ q ⌋ + ⌈ q ⌉ + ǫ, we have
2 2

sup lim sup fδ,λ,d(n) ≤ K.


d≥0 n→∞

In particular, K(3) = 2.
In other words, the sharp phase transition occurs between the exponential regime and the
regime where graphs cannot have more than K δ-distinct colorings. For 3-colorings, there is
a “jump” from having at most 2 distinct colorings to having exponentially many!
Observe that the theorem is about graphs with very small λ2 . A crucial property of such
graphs, as shown in Lemma 4.2 below, is that the union of any two sufficiently large vertex
sets with few edges between them is a nearly independent set.
Lemma 4.2. Let G = (V, E) be a graph with second eigenvalue λ2 . For disjoint vertex
subsets A, B with w(A), w(B) ≥ γ and e(A, B) ≤ ξ, we have
3 max{λ2 , ξ}
e(A ∪ B) ≤ .
γ
Proof. Let x ∈ RV be given by


w(B) v∈A
xv = −w(A) v∈B

0 otherwise.
Then xv ⊥ 1. By Lemma 2.1,
1
d
(2|E(A)|w(B)2 + 2|E(B)|w(A)2 − 2|E(A, B)|w(A)w(B))
λ2 ≥
w(B)2 |A| + w(A)2 |B|
e(A)w(B)2 + e(B)w(A)2 − e(A, B)w(A)w(B)
=
(w(A) + w(B))w(A)w(B)
(e(A) + e(B))γ − ξ

(w(A) + w(B))
≥ (e(A) + e(B))γ − ξ.
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 11

As e(A ∪ B) = e(A) + e(B) + e(A, B), we have


λ2 + ξ 3 max{λ2 , ξ}
e(A ∪ B) ≤ +ξ ≤ .
γ γ

Given a set of δ-distinct colorings of G, we use Lemma 4.2 to partition V into nearly
independent sets such that no coloring gives the same color to vertices of different nearly
independent sets. This allows us to upper bound the distance between colorings on each of
these nearly independent sets.
To find these vertex
 subsets, we group the vertices of G using the colorings. Fix ǫ > 0
1 1
and let δ = 1 − 2 ⌊ q ⌋ + ⌈ 1q ⌉ + ǫ. Suppose that C = {X1 , . . . , XK } is a δ-distinct set of q
2 2
colorings of G. For α = (a1 , . . . , aK ) ∈ [q]K , let
Sα = {v ∈ V |Xi (v) = ai for i ∈ [K]}
be the set of vertices that are colored according to α by the colorings in C.. Observe that
for any α, β ∈ [q]K that agree on at least one coordinate, there are no edges between Sα and
Sβ .
Another way to view the above setup is a homomorphism from G to Kq⊗K , the graph
on [q]K where (a1 , . . . , aK ) and (b1 , . . . , bK ) are adjacent if and only if ai 6= bi for each i.
The homomorphism is given by v 7→ (X1 (v), . . . , XK (v)). Then the preimage of a vertex α
of Kq⊗K is precisely Sα . We write α ∼ β if they are adjacent; that is, if they don’t agree
on any coordinate. From now on, we will view G and the corresponding weighted Kq⊗K
interchangeably, with w(α) = w(Sα), e(α) = e(Sα ), and e(α, β) = e(Sα , Sβ ) and extending
to subsets of [q]K accordingly.
Let γ = γ(ǫ) > 0 be a constant to be determined later. For now, we disregard those α
with w(α) < γ and build nearly independent sets by combining the remaining vertices of
[q]K . In particular, suppose that S ⊂ [q]K and each element in S has weight at least γ. Now
for any β ∈/ S with w(β) ≥ γ, let
A = {β} ∪ {α ∈ S|α ∼ β}
and
B = {α ∈ S|α ≁ β}.
Then since there are no edges between β and B, we have that e(A, B) ≤ e(S). As long
as there is some α ∈ S such that α ≁ β, B has weight at least γ. Writing S ′ = S ∪ {β},
Lemma 4.2 then gives e(S ′ ) ≤ 3 max{λγ2 ,e(S)} . With the base case e({α}, {β}) = 0 for any pair
α ≁ β and w({α}), w({α}) ≥ γ, induction gives
 |S|  qK
3 3
e(S) ≤ λ2 ≤ λ2 .
γ γ
We form a graph Hγ on the set of α ∈ [q]K with w(α) ≥ γ by joining each pair α ≁ β. Let
S1 , . . . , St be the connected components of Hγ . Then each Si is a nearly independent set as
above. This results in the following lemma.
12 HONGLIN ZHU

Lemma 4.3. Let Vγ = {α ∈ [q]K |w(α) ≥ γ}. Then there is a partition of Vγ into subsets
S1 , . . . , St such that
 qK
3
e(Si ) ≤ λ2 ,
γ
and for each α ∈ Si , β ∈ Sj with i 6= j, α ∼ β.
One simple constraint on nearly independent sets is that they cannot get much larger than
half of |V |.
Lemma 4.4. Let G = (V, E) be a graph and let A ⊂ V . For any ǫ > 0, if e(A) ≤ ǫ, then
w(A) ≤ 1+ǫ
2
.
Proof. Since G is a regular graph,
2w(A) = 2e(A) + e(A, V \ A) ≤ 1 + e(A) ≤ 1 + ǫ.

We are now in a position to prove Theorem 4.1. We first take the partition of Vγ into
S1 , . . . , St provided by Lemma 4.3. Using a pigeonhole argument, we show that if G has
too many δ-distinct colorings, then we can find two colorings such that for each Si , the two
colorings use the same number of colors. Then we deduce that these two colorings cannot
be δ-distinct to obtain a contradiction.

Proof of Theorem 4.1. We claim that K(q) ≤ ⌊q−1 q−1

, so in particular K(3) ≤ 2. Let K =
q−1  S 2
t
⌊ q−1 ⌋
+ 1. Let γ = ǫ/(2q K ). Let Vγ = i=1 Si be the partition given by Lemma 4.3. Then
2
w(Vγ ) ≥ 1 − q K γ = 1 − ǫ/2. By Lemma 4.4, we have that t ≥ 2. Let λ = λ(ǫ) be such that
e(Si ) ≤ ǫ for any λ2 ≤ λ.
Let Sij ⊂ [q] be the projection of Si onto the j-th coordinate for i ∈ [t], j ∈ [K]. Since no
elements of different parts share a coordinate, for each j, the sets S1j , . . . , Stj are disjoint.
We add the remaining elements  of [q] to S1j such that S1j , . . . , Stj partition [q] for each j.
q−1
Since there are at most t−1 ≤ K − 1 distinct such partitions, the pigeonhole principle
implies that there exist two coordinates a, b ∈ [K] for which |Sia | = |Sib | for each i ∈ [t].
Without loss of generality suppose a = 1 and b = 2. For convenience, write |Si1 | = |Si2 | = si .
We claim that d(X1 , X2 )/n < δ. This implies that we cannot have K δ-distinct colorings,
which proves the theorem. Since d(X1 , X2 ) = d(σ1 (X1 ), σ2 (X2 )) for any σ1 , σ2 ∈ Sq , we can
arbitrarily apply a permutation matrix to [q]2 . In particular, we can do it so that the square
boxes Bi = Si1 × Si2 ⊂ [q]2 line up on the diagonal in order and that the sum of weights of
the elements of Bi on the main diagonal is at least w(Bi )/si , where w(Bi ) is the total weight
of the elements in [q]K that project into the box Bi . In particular, w(Bi ) ≥ w(Si ). Since
the sum of the weights of the elements on the main diagonal is precisely the weight of the
vertices on which X1 and X2 (after the above permutations) agree, we have
t
X w(Si )
d(X1 , X2 )
≤1− .
n i=1
si
P P
We also have the constraints 2 ≤ t ≤ q, si ∈ [q], ti=1 si ≤ q, ti=1 w(Si ) = w(Vγ ) ≥ 1 − ǫ/2,
and w(Sj ) ≤ 1+ǫ
2
(by Lemma 4.4).
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 13
P P
We wish to lower bound the quantity ti=1 w(S i)
. Thus, we may assume that si = q
P si
and w(Si ) = 1 − ǫ/2. Suppose that s1 ≥ · · · ≥ st . Then the quantity is minimized by
maximizing w(S1 ) and then w(S2 ). Therefore, we have
t
X 1+ǫ 1−2ǫ 1+ǫ 1−2ǫ
w(Si ) 2 2 2 2
≥ + ≥ +
i=1
si s1 s2 ⌈ 2q ⌉ ⌊ 2q ⌋

for sufficiently small ǫ. Finally, we have


 1+ǫ 1−2ǫ 
d(X1 , X2 ) 2
≤ 1− + 2q
n ⌈ 2q ⌉ ⌊2⌋
   
1 1 1 1 1 1
≤ 1− + +ǫ −
2 ⌊ 2q ⌋ ⌈ 2q ⌉ ⌊ 2q ⌋ 2 ⌈ 2q ⌉
 
1 1 1
<1− + +ǫ
2 ⌊ 2q ⌋ ⌈ 2q ⌉
= δ.

Remark 4.5. The constant K = K(q) is likely not tight for q > 3. (It is tight for q = 3
by Theorem 1.6, which shows that Rconst strictly contains Runi for small λ.) We have that
K(q) ≥ q − 1 by the following construction. Let G be a bipartite graph. For 1 ≤ i ≤ q − 1,
let Xi be a coloring that colors one part with the first i colors randomly and the other part
with the remaining  q − i colors
 randomly. A computation shows that, as the size of G goes
1 1 1
to infinity, this is a 1 − 2 ⌊ q ⌋+1 + ⌈ q ⌉ -distinct set of colorings.
2 2

5. The unique regime


We prove Theorem 1.6 in this section. We first show the inclusion results for the unique
regime. This extends Lemma 5.7 of [1] which focuses on the case q = 3.
Theorem 5.1. Any (δ, λ) ∈ [1 − 1/(q − 1), 1 − 1/q] × (0, 1] satisfying
1
2 2
1− q−1
(q − 1)(1 − δ) + (1 − (q − 1)(1 − δ)) < 1 −
1−λ
 
is in the unique regime. In particular, for any ǫ > 0, 1 − 1q , (q−1)
1
2 − ǫ is in the unique
 
1
regime, and there exists some λ = λ(ǫ) > 0 such that 1 − q−1 + ǫ, λ is in the unique
regime.
Proof. Suppose that G = (V, E) is an n vertex d-regular graph with λ2 ≤ λ and q-colorings
X and Y satisfying d(X, Y ) ≥ δn. Furthermore, assume that δ ≥ 1 − 1/(q − 1). For each
σ ∈ Sq , let
Vσ = {v ∈ V |X(v) = σ(Y (v))}.
14 HONGLIN ZHU

Then d(X, Y ) = n − maxσ |Vσ |. Rearranging, we have w(Vσ ) ≤ 1 − δ ≤ 1/(q − 1) for all σ.
Consider the test vector x ∈ RV given by
(
w(Vσ ) − 1 v ∈ Vσ
xv =
w(Vσ ) v ∈ Vσ .
Then x ⊥ 1. Recall that for any A ⊂ V , e(A) + 12 e(A, A) = w(A). By Lemma 2.1, we have
1
d
(2|E(Vσ )|(w(Vσ ) − 1)2 + 2|E(Vσ )|w(Vσ )2 + 2|E(Vσ , Vσ )|(w(Vσ ) − 1)w(Vσ ))
λ ≥ λ2 ≥
(w(Vσ ) − 1)2 |Vσ | + w(Vσ )2 |Vσ |
e(Vσ )(1 − w(Vσ ))2 + e(Vσ )w(Vσ )2 − e(Vσ , Vσ )(1 − w(Vσ ))w(Vσ )
=
w(Vσ )(1 − w(Vσ ))
w(Vσ )(1 − w(Vσ ))2 + (1 − w(Vσ ))w(Vσ )2 − 12 e(Vσ , Vσ )
=
w(Vσ )(1 − w(Vσ ))
e(Vσ , Vσ )
=1− .
2w(Vσ )(1 − w(Vσ ))
Rearranging, we have
e(Vσ , Vσ ) ≥ 2(1 − λ)(w(Vσ ) − w(Vσ )2 ).
On the other hand, each vertex v ∈ V appears in Vσ for exactly (q − 1)! different σ ∈ Sq
and each edge uv ∈ E appears in E(Vσ , Vσ ) for exactly 2(q − 2)(q − 2)! different σ ∈ Sq .
Therefore, summing over all σ ∈ Sq , we have
 
X
2(q − 2)(q − 2)! ≥ 2(1 − λ) (q − 1)! − w(Vσ )2  .
σ∈Sq

Rearranging, we have
X 1
1 2
1 − q−1
w(Vσ ) ≥ 1 − .
(q − 1)! σ∈S 1−λ
q

Let π = (1 2 · · · q) ∈ Sq be the cyclic shift. Then for any σ ∈ Sq ,


q−1
[
V = Vπ i σ .
i=0

Then since w(Vσ ) ≤ 1 − δ ≤ 1/(q − 1),


q−1
X
w(Vσ )2 ≤ (q − 1)(1 − δ)2 + (1 − (q − 1)(1 − δ))2 .
i=0

Since Sq can be partitioned into (q − 1)! such orbits, we have that


1
2 2
1− q−1
(q − 1)(1 − δ) + (1 − (q − 1)(1 − δ)) ≥ 1 − .
1−λ
Thus, if the inequality does not hold for (δ, λ), it means that G cannot have two q-colorings
that are δ-distinct. 
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 15

Next, we give constructions for the two negative results in Theorem 1.6.
 
1
Lemma 5.2. For all λ > 0, 1 − q−1 , λ is not in the unique regime.

Proof. Fix λ > 0 and choose d such that 2/ d ≤ λ. Let G = Gn be a sequence of d-
regular bipartite
√ graphs on 2(q − 1)n vertices. By Lemma 2.4, we can choose Gn such that
λ2 (Gn ) ≤ 2/ d ≤ λ for sufficiently large n.
We now color G in two ways. Let X be a coloring that colors the first part of G by 1 and
the second part by 2, . . . , q such that each color is used on exactly n vertices. Similarly, let
Y be a coloring that colors the second part of G by q and the first part of G by 1, . . . , q − 1
such that each color is used on exactly n vertices. Then
 
1
d(X, Y ) = 2(q − 2)n = 1 − |V (G)|.
q−1
 
1
Therefore, 1 − q−1 , λ is not in the unique regime. 
 
1 1
Lemma 5.3. The point 1 − q , (q−1)2 is not in the constant regime.
1
To prove Lemma 5.3, we construct a family of graphs G = Gn with λ2 = (q−1) 2 such that
 
G has a 1 − 1q -distinct set of colorings of arbitrary size. We obtain the sequence Gn by
constructing repeated 2-lifts of G1 . Recall that a 2-lift of a graph is defined in the following
way.
Let G = (V, E) be an n-vertex graph. A signing of the edges of G is a function s : E →
{−1, 1}. The 2-lift of G associated with signing s is a graph G b = (Vb , E)
b on 2n vertices given
as follows. For each v ∈ V , we have two vertices v1 , v2 ∈ Vb associated to v. For each uv ∈ E,
we have two edges in E b associated to it. If s(uv) = 1, then the two edges are u1 v1 and u2 v2 .
If s(uv) = −1, then the two edges are u1 v2 and u2 v1 .
Observe that if G is d-regular, then any 2-lift of G is also d-regular. The following result of
Bilu and Linial [2] shows that we can always find a 2-lift that preserves the second eigenvalue
of G.
Lemma 5.4. [2] There exists a constant C > 0 such that the following holds. Let G be
a d-regular graph with normalized second eigenvalue λ2 . There exists a 2-lift of G whose
normalized second eigenvalue is
 s 
 log d 
3
max λ2 , C  .
 d 

Now we prove Lemma 5.3.


Proof of Lemma 5.3. Let N ≥ 2 be any sufficiently large positive integer. Let G1 = Kq⊗N be
the graph on [q]N where (a1 , . . . , aN ) and (b1 , . . . , bN ) are adjacent if and only if ai 6= bi for
all i ∈ [N]. Then G1 is (q − 1)N -regular. Furthermore, the normalized adjacency matrix of
G1 is
A = A⊗N Kq ,
16 HONGLIN ZHU

where AKq is the normalized adjacency matrix of Kq and ⊗ is the Kronecker product. Since
1
the spectrum of Kq is given by λ1 = 1 and λ2 = · · · = λq = − q−1 , we have that λ2 (G1 ) =
1
(q−1)2
. By Lemma 5.4, for N sufficiently large, we can find a sequence of repeated 2-lifts of
1
G1 to form a sequence G = Gn of (q − 1)N -regular graphs with λ2 = (q−1) 2.
 
We show that each Gn in this sequence has a 1 − 1q -distinct set of colorings of size N.
First, consider the colorings
 X1 , . . . , XN of G1 where Xi (a1 , . . . aN ) = ai . It is easy to verify
1
that d(Xi , Xj ) = 1 − q q K .
Now it suffices to show that a δ-distinct set of colorings of a graph can be lifted to a
δ-distinct set of colorings of its 2-lift. Indeed, if X1 , . . . , XN are colorings of G = (V, E),
then consider colorings X c1 , . . . , X
d c c
N where Xi (v1 ) = Xi (v2 ) = Xi (v) for all v ∈ V . It is clear
that the distance between any pair of colorings is doubled, so the distance parameter δ is
preserved.
1
Therefore, there exists a sequence G = Gn of (q − 1)N -regular graphs with λ2 = (q−1) 2
 
such that each Gn has a 1 − 1q -distinct set of colorings of size N. Taking N to infinity, we
 
1 1
have that 1 − q , (q−1)2 is not in the constant regime. 

6. Conclusion and future directions


We believe that the sphere packing problem for proper colorings provide a new and and
interesting way to study expander graphs which can lead to useful applications. For example,
Bafna et al. [1] proved a result analogous to Theorem 5.1 in the case q = 3 to construct an
efficient algorithm for finding a large independent set in a 3-colorable expander, while it is
known to be NP-hard (assuming the unique-games conjecture) to find a large independent
set in a general 3-colorable graph.
A natural future direction is to understand the landscape of the regimes in Figure 1 better.
In particular, does f have an intermediate growth behavior for some (δ, λ)?
Question 6.1. Is there some (δ, λ) ∈ [0, 1 − 1/q] × (0, 1] that is neither in the exponential
regime nor the constant regime? If so, what growth behavior does f exhibit as n → ∞?
It appears that when q is large, the exponential regime should be much larger than what
Theorems 1.3 and 1.4 guarantee. Indeed, for d < q − 1, any d-regular graph on n vertices
has at least (q − 1 − d)n proper q-colorings. We expect that there are good expanders for
which these colorings are distributed sufficiently “uniformly” over [q]n .
Conjecture 6.2. For √ all q sufficiently large, there exists an infinite sequence of d-regular
graphs with λ2 = Θ(1/ d) for some d = Θ(q) such that each graph has a δ-distinct set of
proper q-colorings of exponential size for any δ < 1 − 1/q. In particular, this implies that

(δ, λ) is in the exponential regime for all δ < 1 − 1/q and some λ = O(1/ q).
This intuition suggests that only expansion on the order of oq (1) forces the sizes of δ-
distinct sets of colorings to be sub-exponential, and the predominant behavior of the count
function fδ,λ,d (n) is exponential. To gain more refined understanding of the exponential
regime, it is thus natural to study the rate question as in the theory of error-correcting
codes.
SPHERE PACKING PROPER COLORINGS OF AN EXPANDER GRAPH 17

Question 6.3. For fixed q ≥ 3 and (δ, λ, d), give bounds for
logq fδ,λ,d (n)
R(δ, λ, d) := lim sup .
n→∞ n
In particular, we always have that 0 ≤ R(δ, λ, d) ≤ 1.
When q −1 > d, q-colorability is guaranteed. Thus, we may also consider the rate question
for two-sided expanders as suggested in Remark 1.1. In particular, for two-sided expanders,
we have access to more powerful tools such as the expander mixing lemma.
Question 6.4. For appropriate values of q and (δ, λ, d), define R2 (δ, λ, d) as the rate for
the count function of δ-distinct sets of q-colorings of two-sided λ-expanders. What are some
bounds for R2 ? How does R2 relate to R1 ? (Trivially, R2 (δ, λ, d) ≤ R(δ, λ, d) since every
two-sided expander is a one-sided expander.)
Another potential line of inquiry is to directly study graph-theoretic constraints that
having many δ-distinct colorings put on an expander graph. For example, the main idea
behind proving Theorem 1.5 in §4 is to show that having many δ-distinct colorings for δ as in
the statement of Theorem 1.5 along with being a good expander impose very rigid conditions
on the independent sets of the graph. We believe that analogous results relating the geometry
of proper colorings to structural properties of expander graphs hold independent interest.
References
[1] Mitali Bafna, Jun-Ting Hsieh, and Pravesh K. Kothari, Rounding large independent sets on expanders,
2024, arXiv:2405.10238.
[2] Yonatan Bilu and Nathan Linial, Lifts, discrepancy and nearly optimal spectral gap, Combinatorica 26
(2006), 495–519.
[3] Gerandy Brito, Ioana Dumitriu, and Kameron Decker Harris, Spectral gap in random bipartite biregular
graphs and applications, Combin. Probab. Comput. 31 (2022), 229–267.
[4] Jeff Cheeger, A lower bound for the smallest eigenvalue of the laplacian, pp. 195–200, Princeton Uni-
versity Press, Princeton, 1971.
[5] Joel Friedman, A proof of Alon’s second eigenvalue conjecture and related problems, Mem. Amer. Math.
Soc. 195 (2008), viii+100.
[6] Alan J. Hoffman, On eigenvalues and colorings of graphs, Graph Theory and its Applications (Proc.
Advanced Sem., Math. Research Center, Univ. of Wisconsin, Madison, Wis., 1969), Academic Press,
New York-London, 1970, pp. 79–91.
[7] Adam W. Marcus, Daniel A. Spielman, and Nikhil Srivastava, Interlacing families I: Bipartite Ramanu-
jan graphs of all degrees, Annals of Mathematics 182 (2015), 307–325.
[8] Colin McDiarmid, On the method of bounded differences, Surveys in combinatorics, 1989 (Norwich,
1989), London Math. Soc. Lecture Note Ser., vol. 141, Cambridge Univ. Press, Cambridge, 1989,
pp. 148–188.
[9] A. Nilli, On the second eigenvalue of a graph, Discrete Math. 91 (1991), 207–210.
[10] Michael Sipser and Daniel A. Spielman, Expander codes, IEEE Trans. Inform. Theory 42 (1996), 1710–
1722, Codes and complexity.

You might also like