Combinatorica 26 (4) (2006) 475487

C OM BIN A TORIC A
Bolyai Society Springer-Verlag

HOMOLOGICAL CONNECTIVITY OF RANDOM 2-COMPLEXES

NATHAN LINIAL*, ROY MESHULAM*

Received May 25, 2003

Let n1 denote the (n 1)-dimensional simplex. Let Y be a random 2-dimensional

subcomplex of n1 obtained by starting with the full 1-dimensional skeleton of n1
and then adding each 2simplex independently with probability p. Let H1 (Y ; F2 ) denote
the rst homology group of Y with mod 2 coecients. It is shown that for any function
(n) that tends to innity

2 log n(n)
0p=
lim Prob[H1 (Y ; F2 ) = 0] = n
2 log n+(n) .
n 1p= n

1. Introduction

Let G(n, p) denote the probability space of graphs on the vertex set [n] =
{1, . . . , n} with independent edge probabilities p. Let log denote the natural
logarithm. A classical result of Erdos and Renyi [2] asserts that the threshold
probability for connectivity of G G(n, p) coincides with the threshold for
the non-existence of isolated vertices in G. In particular, for any function
(n) that tends to innity

log n(n)
0p=
lim Prob[G G(n, p) : G connected] = n
log n+(n) .
n 1p= n

In this paper we study an analogous problem for random 2dimensional

complexes. Unlike the graphical case, there are several distinct notions of
Mathematics Subject Classication (2000): 55U10, 05C80
* Supported by an Israel Science Foundation grant.

02099683/106/$6.00
2006
c Janos Bolyai Mathematical Society and Springer-Verlag
476 NATHAN LINIAL, ROY MESHULAM

1-dimensional connectivity of a (connected) 2-dimensional complex X. The

strongest such notion is simple connectivity, i.e. the triviality of the funda-
mental group 1 (X). Next comes homological 1-connectivity, i.e. the vanish-
ing of the rst integral homology H1 (X; Z). Here we are concerned with the
still weaker notion of F2 -homological 1-connectivity, namely the vanishing of
the rst homology with mod 2 coecients H1 (X; F2 ).
We recall some topological terminology (see e.g. [3]). For a simplicial
complex X, let fk (X) denote the number of k-dimensional simplices in X
and let X (k) = { X : dim k} denote the k-dimensional skeleton of X.
Let C k (X) denote the F2 -vector space of F2 k-cochains on X i.e. the space
of F2 -valued functions on the k-simplices of X. The coboundary operator
dk : C k (X) C k+1 (X) is given as follows. For f C k (X) and a (k + 1)-
dimensional X 
dk (f )() = f ( ) .
, dim =k

Let Z k (X) = ker(dk ) denote the space of k-cocycles of X and let B k (X) =
Im(dk1 ) denote the space of k-coboundaries of X. The k-th cohomology
group of X with F2 coecients is

Z k (X)
H k (X; F2 ) = .
B k (X)

The k-th homology group Hk (X; F2 ) is isomorphic to H k (X; F2 ). We abbre-

viate H k (X) = H k (X; F2 ).
Let n1 denote the (n 1)-dimensional simplex on the vertex set [n].
(1) (2)
Let Y (n, p) denote the probability space of subcomplexes n1 Y n1
with probability measure
n
Pr(Y ) = pf2 (Y ) (1 p)( 3 )f2 (Y ) .
 
An edge ij [n] 2 is isolated in Y if it is not contained in any of the 2-
simplices of Y . If ij is isolated then the indicator function of ij is a non-
trivial 1-cocycle of Y , hence H 1 (Y ) = 0. Our main result is that the threshold
probability for the vanishing of H 1 (Y ) coincides with the threshold for the
non-existence of isolated edges in Y .

Theorem 1.1. Let (n) be any function which satises (n) then

2 log n(n)
0p=
(1) lim Prob[Y Y (n, p) : H (Y ; F2 ) = 0] =
1 n
2 log n+(n) .
n 1p= n
 HOMOLOGICAL CONNECTIVITY OF RANDOM 2-COMPLEXES 477

The case p = 2 log n(n)

n is straightforward: Let g(Y ) denote the number
of isolated edges of Y . Then
 
n
E[g] = (1 p)n2 = (exp((n))) .
2

A standard second moment argument then shows that

Prob[H 1 (Y ) = 0] Prob[g = 0] = o(1) .

Before proceeding to the main part of the paper we briey outline the
analogy between the proof of Theorem 1.1 and that of the classical theo-
rem on the threshold for graph connectivity. From a topological perspective,
graph connectivity is homological 0-connectivity and here we are concerned
with F2 -homological 1-connectivity. The coupon-collector arguments for es-
tablishing the thresholds for the existence of an isolated vertex/edge are
identical in both cases. The proof that p = log n+(n)
n guarantees a.e. con-
nectivity of G G(n, p) may be formulated as follows. For g : [n] F2 , let
B(g) be the number of edges e = uv in the complete graph Kn , such that
d0 (g)(e) = g(u) + g(v) = 1. In other words, B(g) = |g1 (0)||g1 (1)| is the
number of edges in the cut (g1 (0), g1 (1)) of the complete graph Kn . A
graph G = ([n], E) is thus disconnected i there exists a non-constant func-
tion g : V F2 with d0 (g)(e) = 0 for all e E. For a given g, the probability
of this event in the space G(n, p) is (1 p)B(g) . Limiting ourselves, as we
may, to gs with |g1 (1)| n/2, we now apply a union bound to derive the
desired conclusion.
Something similar happens here with p = 2 log n+(n) . For a mapping
[n] n
f : 2 F2 , let B(f ) be the number of triples = uvw for which
(1) (2)
d1 (f )() = f (uv) + f (vw) + f (uw) = 1. A complex n1 Y n1 is not
 
F2 -homologically 1-connected i there exists a mapping f : [n] 2 F2 that
is not of the form d0 (g) for any g, such d1 (f )() = 0 for all 2-dimensional
Y . Again, for a given f , this event has probability (1p)B(f ) in the space
Y (n, p). At this point, the perfect analogy breaks and we need more rened
arguments. In the graphical case, the functions g and 1 g are equivalent,
whence we had the liberty of assuming that |Supp(g)| n/2, by switching,
if necessary from g to 1 g. Likewise, here f is equivalent to any function
of the form f + d0 (h) for any h : [n] F2 . (Note that d0 (h) is simply the
characteristic function of a cut in Kn .) We are thus allowed to consider
only functions f such that |Supp(f )| |Supp(f + d0 (h))| for all h. It turns
out quite easily that we may further restrict ourselves and assume that the
478 NATHAN LINIAL, ROY MESHULAM

graph ([n], Supp(f )) has a single nontrivial connected component. (This sim-
plication has no graphical analogue.) Let Fn be the class of all functions
f satisfying these  two assumptions. As in the graphical case we estimate
Prob[H1 (Y ) = 0] f Fn (1 p)B(f ) . The proof that this sum is o(1) is sig-
nicantly more involved than in the graphical case. There are two main steps
in the proof. First we show (Proposition 2.1) that B(f ) cn|Suppf | for every
f Fn , where c > 0 is an absolute constant. We then derive an upper bound
(Proposition 2.3) on the number of f Fn for which B(f ) = (1)n|Suppf |
for bounded away from zero.
We turn to a more concrete discussion. For an f C 1 (n1 ) denote by
(1)
[f ] the image of f in H 1 (n1 ). Let
  


[n]
B(f ) = : d1 (f )() = 1 .
3
(1)
For any complex Y n1 we identify H 1 (Y ) with its image under the
(1)
natural injection H 1 (Y ) H 1 (n1 ). It follows that for f C 1 (n1 )

Prob[[f ] H 1 (Y )] = (1 p)B(f ) .
[n]
Let Suppf = {e 2 : f (e) = 1}. The Hamming Weight of the cohomology
(1)
class [f ] H 1 (n1 ) is dened by
w([f ]) = min{|Suppf  | : [f  ] = [f ]}
= min{|Supp(f + d0 (h))| : h C 0 (n1 )}.
Let
Fn = {f C 1 (n1 ) : w([f ]) = |Suppf |}.
Associate with any f C 1 (n1 ) the simple undirected graph Gf = ([n], Ef )
with edge set Ef = Suppf . Let Fn consist of all 0 = f Fn such that Gf has
exactly one connected component which is not an isolated point.
If H 1 (Y ) = 0 then any 0 = f Z 1 (Y ) with minimal support must belong
to Fn . Therefore

Prob[H 1 (Y ) = 0] Prob[[f ] H 1 (Y )] .
f Fn

Theorem 1.1 will thus follow from

Theorem 1.2. For p = 2 log n+(n)
n

(2) (1 p)B(f ) = o(1) .
f Fn
 HOMOLOGICAL CONNECTIVITY OF RANDOM 2-COMPLEXES 479

In Section 2 we formulate the problem in graph theoretical terms and

state Propositions 2.1 and 2.3 which are the main ingredients in the proof
of Theorem 1.2. The proofs of these results are given in Sections 3 and 4. In
Section 5 we combine Propositions 2.1 and 2.3 to derive Theorem 1.2. We
conclude in Section 6 with some remarks on possible extensions and open
problems.

2. A Graph Theoretic Formulation

The mapping f Gf denes a 1-1 correspondence between the 1-cochains

C 1 (n1 ) and the (simple undirected) graphs on the vertex set [n]. For a
graph G = ([n], E) = Gf we denote B(G) = B(f ). Clearly B(G) is the number
 
of triangles T [n]
3 which contain either one or three edges of G. Let

Gn = {Gf : f Fn } , Gn = {Gf : f Fn } .

Suppose G = ([n], E) = Gf Gn . Then f satises w([f ]) = |Supp(f )|. Hence

for all g C 0 (n1 )

(3) |Supp(f + d0 (g))| |Supp(f )| .

[n]
Let S = Supp(g) = {v [n] : g(v) = 1} then Supp(d0 (g)) 2 consists of all
|S||S| edges of the cut (S, S).
It follows that G Gn if and only if

|S||S|
(4) |E (S, S)|
2

for all cuts (S, S). In particular the maximal degree in G Fn is less then n2 .
Gn consists of all G Gn that have at most one connected component which
is not an isolated point. The proof of Theorem 1.2 depends on two results.
In section 3 we prove the following lower bound on B(G):

Proposition 2.1. There exists a constant c 1

120 such that for any G =
([n], E) Gn
(5) B(G) c|E|n .
Remark. Proposition 2.1 is equivalent to the following result which roughly
(2)
says that if Y is close to n1 then any 1-cohomology class of Y must contain
a cocycle with a small support.
480 NATHAN LINIAL, ROY MESHULAM

Corollary 2.2. Let c be the constant of Proposition 2.1. Then for any sim-
(1) (2)
plicial complex n1 Y n1 and any [f ] H 1 (Y ; F2 )
n
1 f2 (Y )
(6) w([f ]) c 3
.
n
2
Example. Suppose n is divisible by 3. Let [n] = i=0 Vi be a partition of
[n] with |Vi | = n3 and let
  

(2) [n]
Y = n1 : | Vi | = 1 for all 0 i 2 .
3

The characteristic function f = 1E C 1 (Y ) of the set

E = {{u, v} : u V0 , v V1 }

is a 1-cocycle of Y . It can be checked that f is a cocycle of minimal support

in its cohomology class [f ]. Hence
n
n2 f2 (Y )
w([f ]) = |Suppf | = =3 3
.
9 n

It follows that c cannot be replaced by any constant bigger then 13 .

n
For k 2 and 0 1 Let

Gn (k) = {G = ([n], E) Gn : |E| = k}

Gn (k) = {G = ([n], E) Gn : |E| = k}
Gn (k, ) = {G Gn (k) : B(G) = (1 )kn} .

In Section 4 we prove an estimate on the cardinality of Gn (k, ):

Proposition 2.3. For any 0 < ( < 12 there exists a constant C(() such that
for any ( and n5 k n2

 k
(7) |Gn (k, )| C(() n2(1(2 )) .
 HOMOLOGICAL CONNECTIVITY OF RANDOM 2-COMPLEXES 481

3. A Lower Bound on B(G)

Proof of Proposition 2.1. We show (5) with c = 120 1

. For e E let (e)
[n]
denote the number of 2-simplices 3 which contain e but whose two
 
other edges are not in E. Let (e) denote the number of [n] 3 which
contain e and whose two other edges are both in E. Then
 1
(8) B(G) = (e) + (e) .
eE
3 eE

Suppose for contradiction that

(9) B(G) < c|E|n .
Let < 1 be a constant whose value will be assigned later, and let


 1 cn
E = e E : (e) + (e) .
3
Then (8) and (9) imply that |E  | (1 )|E|.
For v [n] let G (v) = {u [n] : uv E} and let degG (v) = |G (v)|. Recall
that
n
(10) degG (v) <
2
for all v [n]. Let e = uv E  then
cn
(11) (e) = n |G (u) G (v)| n (degG (u) + degG (v))

and
3cn
(12) (e) = |G (u) G (v)| .

(10) and (11) imply that degG (u), degG (v) (1 2c n
)2.
 
Let G = ([n], E ) then
 
(1 ) degG (u) = 2(1 )|E| 2|E  | = degG (u) .
u[n] u[n]

Hence there exists a u [n] such that

0 < (1 ) degG (u) degG (u) .
Since u is incident with at least one edge in E  it follows that degG (u)
(1 2c n
) 2 . Thus  
2c n
degG (u) (1 ) 1 .
2
482 NATHAN LINIAL, ROY MESHULAM

Let S = G (u) then (12) implies that for each v S

3cn
|G (v) S| |G (v) G (u)| .

It follows that
   
3cn 2c n 3cn 8c n
|G (v) S| degG (v) 1 = 1 .
2 2
Therefore

|E (S, S)| = |G (v) S|
vS
     
2c n 8c n 10c n2
(1 ) 1 1 (1 ) 1 .
2 2 4

Taking = 10c we obtain
1
|E (S, S)| (1 10c)2 |(S, S)| > |(S, S)|
2
in contradiction with (4).

4. Near Transversals

The proof of Proposition 2.3 depends on the following

Claim 4.1. For any 0 < ( < 12 there exist constants C1 ((), n0 (() such that
for any ( 1 and for any graph G = ([n], E) with n n0 (() and degree
sequence d1 , . . . , dn n2 which satises

n 
n
(13) |E| = k n2 and d2i kn
5 i=1

there exists a set of vertices S [n] such that

C1 (() k
(14) |S| and |{e E : S e = }| (2 ()k .
n

Proof. Pick S at random with Prob[i S] = 2dni and let

X = |{e E : S e = }| .
 HOMOLOGICAL CONNECTIVITY OF RANDOM 2-COMPLEXES 483

Then
  
2di

2dj

E[X] = 1 1 1 =
ijE
n n
 n 2
2  4  2 n
2 
(di + dj ) 2 (di dj ) d2i 2 di
n ijE n ijE n i=1 n i=1
8k 2
2k (2 8n )k .
n2
1
It follows that for n n0 (() = ( 100
2
)
   
(
Prob[X (2 ()k] Prob X 1 E[X]
3
( E[X] ( (2
(15) .
3 k 3 3
Since

n
2di 4k
E[|S|] = =
i=1
n n

it follows by a large deviation estimate (see e.g. Theorem A.1.12 in [1]) that
for > e
    4k   4
4k e n e 5
(16) Prob |S| > < .
n
Hence for a suciently large C1 (()
 
C1 (() k (2
(17) Prob |S| > < .
n 3
(15) and (17) imply that there exists an S [n] which satises (14).
Remark. Claim 4.1 is not far from optimal in the sense that there exist
graphs which satisfy (13), but such that any set of O( nk ) vertices intersects at
most 2k(1+ o(1)) edges. Indeed let < 12 be xed and let (n) k o(n2 ).
Let A, B, C be three disjoint subsets of [n] such that |A| = 4k n , |B| = 2 , and
n
t
|C| = t where 2 = (12)k. Let G = ([n], E) consist of the complete bipartite
graph with sides Aand B, together with the complete graph on C. Then
|E| = k, di n2 and ni=1 d2i kn. For any S [n] such that |S| = O( nk )

|{e E : S e = }| 2k + |S| t = 2k(1 + o(1)) .

484 NATHAN LINIAL, ROY MESHULAM

Proof of Proposition 2.3. It suces to consider n n0 ((). Let G =

([n], E) Gn (k, ). Then di = degG (i) < n2 and
  
n
(1 )kn = B(G) (e) (n degG (i) degG (j)) = kn d2i .
eE ijE i=1
n
i=1 di kn
hence by Claim 4.1 there exists a subset S [n]
It follows that 2

which satises (14). We now estimate |Gn (k, )| as follows: The number of
possible subsets S is at most 2n . The number of choices for the edges of G
which are incident with a xed S is at most
 |S|+1    k

k
|S|n |S|n e|S|n
2 <k <k < k(eC1 (())k .
=(2 )k
0 k k

The

number of choices for the edges that are not covered by S is at most
(n2 ) . It follows that
k(1(2 ))
 k
|Gn (k, )| 2n k (eC1 (())k n2k(1(2 )) C(() n2(1(2 )) .

5. Proof of Theorem 1.2

Proof of Theorem 1.2. Let (n) and let p = 2 log n+(n)

n . Writing (2)
in the equivalent graph theoretic formulation, we have to show that
  
(18) (1 p)B(G) = (1 p)B(G) = o(1) .
 (k)
k1 GGn 
GGn

Let c 120
1
denote the constant of Proposition 2.1. We deal separately with
three intervals of k:
(i) 1 k n5 . Recall that G Gn (k) has exactly one connected component
which is not an isolated point. It follows that
  
k+1  en k+1  e(k + 1) k
n
|Gn (k)| 2 < (10n)k+1 .
k+1 k k+1 2
Next note that (e) nk1 for all edges e of G, hence B(G) k(nk1).
It follows that
n/5 n/5
  
(1 p)B(G) |Gn (k)|(1 p)k(nk1)
 (k)
k=2 GGn k=2
n/5
  4k n/5

(10n)k+1 n2 e(n) (10n)k+1 n 5 = O(n 5 ).
8k 1
5

k=2 k=2
 HOMOLOGICAL CONNECTIVITY OF RANDOM 2-COMPLEXES 485

Hence
n/5  
  n
(1 p)n2 + O(n 5 )
1
(1 p) B(G)

 (k)
k=1 GGn
2
 
(n)
+ O(n 5 ) .
1
(19) exp
2

(ii) n
5 k n2c . We need the following:
Claim 5.1. For any n
5 k n2c and 0 1

(1 p)B(G) = O(n 4 ) .
k
(20)
GGn (k,)

Proof. If 0 13 then
 
 n
(1 p) B(G)
2 (1 p)(1)kn
GGn (k,)
k
 k
en2
n2(1)k (10n21 )k (10n 3 )k .
1

2k

Suppose now that 1

3 1. Applying Proposition 2.3 with ( = c we obtain

(1 p)B(G) |Gn (k, )| (1 p)(1)kn
GGn (k,)
 k
C(() n2(1(2 )) n2(1)k
 k
= C(() n2(1 )
 k
2(1c)

C(c) n 3 .

For each k the number of s for which Gn (k, ) is non-empty is at most n3 .

It follows by Claim 5.1 that
2c
n 2c
 n  
(1 p)B(G) = (1 p)B(G)
k= n
5
GGn (k) k= n
5
GGn (k,)
2c
n 
O n 4 = n(n) .
k
(22) n3
k= n
5
486 NATHAN LINIAL, ROY MESHULAM

(iii) k n2c . By Proposition 2.1

 
   n
(1 p) B(G)
2 (1 p)ckn
k
kn2c GGn (k) kn2c
 k
 en2 
(23) n2ck (2nc )k = n(n) .
2k
kn2c kn2c

Finally (18) follows from (19), (22) and (23).

6. Concluding Remarks

We have shown that in the model Y (n, p) of random 2-complexes on n

vertices, the threshold for the vanishing of H1 (Y ; F2 ) occurs at p = 2 log n
n . A
straightforward extension of the proof shows that the same result holds for
homology with coecients in any xed nite abelian group. We still do not
know the answer to the next set of naturally-arising questions: Where is the
threshold for the vanishing of H1 (Y, Z)? For being simply connected? What
happens in the higher-dimensional situation?
We believe the methods of this paper will be relevant to the questions
of homological connectivity of random complexes in higher dimensions. The
problem of simple connectivity will probably require a dierent approach
and is particularly intriguing, since this property is, in general, undecidable.
One point worth mentioning here is this. The three conditions: Vanishing of
the rst homology over F2 , over Z, and simple connectivity, are progressively
stronger, in this order. If the thresholds for their occurrence dier, this will
supply us with a large set of instances where some, but not all of these
conditions hold. This may be of interest also outside of combinatorics. If the
thresholds coincide, it will be of interest to develop more rened probabilistic
models which do dierentiate between these criteria for connectivity.
At any event, we believe that further study of topological properties of
random complexes will prove both interesting and useful.

Acknowledgement. We would like to thank Gil Kalai for helpful discus-

sions.

References

[1] N. Alon and J. Spencer: The Probabilistic Method, 2nd Edition, Wiley-Intescience,
2000.
 HOMOLOGICAL CONNECTIVITY OF RANDOM 2-COMPLEXES 487

[2] P. Erdos and A. Renyi: On the evolution of random graphs, Publ. Math. Inst.
Hungar. Acad. Sci. 5 (1960), 1761.
[3] J. Munkres: Elements of Algebraic Topology, Addison-Wesley, 1984.

Nathan Linial Roy Meshulam

Department of Computer Science Department of Mathematics
Hebrew University Technion
Jerusalem 91904 Haifa 32000
Israel Israel
nati@cs.huji.ac.il meshulam@math.technion.ac.il