Partitioning the hypercube into smaller hypercubes

arXiv:2401.00299v3 [math.CO] 7 Nov 2024

Noga Alon∗ József Balogh† Vladimir N. Potapov‡

November 8, 2024

Abstract
Denote by Qd the d-dimensional hypercube. We estimate the number of
ways the vertex set of Qd can be partitioned into vertex disjoint smaller cubes.
Among other results, we prove that the asymptotic order of this function is
larger than the number of perfect matchings of Qd by an exponential factor
in the number of vertices, and not by a larger factor. We also describe and
address several new (and old) related questions.

1 The problem and main results

Denote by Qd := {0, 1}d the d-dimensional hypercube, and let f (d) be the number
of partitions (tilings) of the vertex set of Qd in which each of the classes spans
a subhypercube. Let fS (d) denote the number of such partitions where each of the
classes has dimension in S ⊂ {0, 1, . . . , d}. In particular, f≤2 (d) counts the partitions
where the parts are singletons, edges or spanning a 2-cube.
For a graph G, denote by m(G) the number of perfect matchings in G, and by m′ (G)
the number of matchings, and write m(d) := m(Qd ) = f1 (d) and m′ (d) = m′ (Qd ) =
f≤1 (d).
An easy observation is that the number of (perfect) matchings of Qd is a lower
bound on f (d). Determining or estimating the number of (perfect) matchings of
∗
Princeton University, Princeton, NJ 08544, USA and Tel Aviv University, Tel Aviv 69978, Israel.
E-mail: [email protected]. Research supported in part by NSF grant DMS-2154082.
†
Department of Mathematics, University of Illinois at Urbana-Champaign, Urbana, Illinois
61801, USA. E-mail: [email protected]. Research is partially supported by NSF Grant DMS-
1764123, Arnold O. Beckman Research Award (UIUC Campus Research Board RB 24012), and the
Langan Scholar Fund (UIUC).
‡
E-mail: [email protected].

1
graphs is a classical problem, see, e.g., [2, 7, 12, 17]. Although it has been studied
for Qd , besides determining it for small d, only general results are known about its
asymptotics, as we discuss later. Below we collect many problems that might be of
interest; and we will address several of them.

Problem 1.1. Determine or estimate the functions

(i) m(d), (ii) m′ (d), (iii) f (d), (iv) f≤2 (d).

Problem (i) is natural and was studied in the literature, (ii) is a natural extension.
Problem (iii) was raised by Gadouleau at the 29th British Combinatorial Conference
in 2022, motivated by the paper of Bridoux, Gadouleau, and Theyssier [6]. Prob-
lem (iv) is a variant we suggest here, and seems to be closely related to (iii).
As pointed out in [6], the first few terms of f (d) are given in the Online Encyc-
lopedia of Integer Sequences (OEIS) as A018926. Starting from d = 0 these are:
1, 2, 8, 154, 89512, 71319425714.
Graham and Harary [14] studied the number of perfect matchings of the hyper-
cube. They calculated
m(1) = 1, m(2) = 2, m(3) = 9, m(4) = 272, m(5) = 589185.
Using dynamic programming, Östergård and Pettersson [21] determined m(6)
(which was also known earlier) and m(7):
m(6) = 16332454526976, m(7) = 391689748492473664721077609089.
Determining precisely, or even providing satisfactory estimates for the functions
m(d), m′ (d) and fS for some interesting families S seems to be difficult, hence it would
also be interesting to compare the orders of magnitude of some of these functions.
Set
n := 2d−1 and N := (d/e)n .
Note that n is the number of vertices in each vertex class of Qd , and as mentioned
below, N is a rough estimate for the number of perfect matchings in it.
The following hierarchy follows from the definitions:
m(d) ≤ m′ (d) ≤ f≤2 (d) ≤ f (d). (1)
We prove that the ratio f (d)/m(d) is exponential in n, and wonder in which of
the inequalities in (1) there is an exponential ratio.

Problem 1.2. What is the order of magnitude of the following ratios? In particular,
which one of them is an exponential function of n?
(i) m′ (d)/m(d), (ii) f≤2 (d)/m′ (d), (iii) f (d)/f≤2(d), (iv) f (d)/m(d).

Our results and conjectures are summarized below.

2
 Classical results easily imply Proposition 1.3, which determines the main term of
the asymptotics of m(d), partially solving Problem 1.1 (i).1 Proposition 1.4 partially
solves Problem 1.1 (ii) and Problem 1.2 (i), showing that m′ (d)/m(d) is subexpo-
nential. For both statements we state general results about regular bipartite graphs,
which imply the required bounds for the hypercube. Proposition 1.5 shows that
f≤2 (d)/m′ (d) and f≤2 (d)/m(d) are (at least) exponential in n, addressing Prob-
lem 1.2 (ii) and (iv). In Proposition 1.6 we show that the ratio f (d)/m(d) is at
most exponential in n, and a modification of its proof implies that allowing to use
2-dimensional subcubes in the partitions has a large impact on their numbers, see
Proposition 1.8.
Our final result, Proposition 1.11, shows that there are many irreducible tight
partitions. The precise definition of this notion appears before the statement of the
result.
We proceed with the formal statements of all the results.
Proposition 1.3. (i) Let G be an a-regular bipartite graph with vertex classes of
size b, where a ≥ 2. Then
b/2a
 a b √ 1/(12a) b/a
 a b
(1+o(1))·log a·(b/2a)
 a b
e · ≤ m(G) ≤ [ 2πa · e ] · =2 · , (2)
e e e
when both a, b → ∞.
(ii) In particular for Qd we have
en/2d · N ≤ m(d) ≤ 2(1+o(1))·log d·(n/2d) · N. (3)
The following result is addressing Problem 1.1 (ii) and Problem 1.2 (i). The lower
bound in Proposition 1.4 (i) also follows from known results in [7], and the upper
bound from √
the ones in [17]. In particular, these results imply that m′ (G)/m(G) =
e(2+o(1))n/ d . Here we describe a short and simple proof, which does not provide the
optimal constant in the exponent obtained in the upper bound.
Proposition 1.4. (i) Let G be an a-regular bipartite graph with vertex classes of
size b, where a ≥ 1 and b → ∞. Then
√  a b √
m′ (G) = m(G) · 2Θ(b/ a) = · 2Θ(b/ a) . (4)
e
(ii) In particular for Qd we have
√ √
m′ (d) = m(d) · 2Θ(n/ d)
= N · 2Θ(n/ d)
. (5)
1
The proof of the upper bound, and that of a slightly weaker lower bound on m(d) as stated in
Proposition 1.3, are posted in several class notes at various course websites, see also [19]. Here we
provide a short proof, for completeness.

3
 Next, we show that the number of cube partitions of Qd is exponentially larger
than the number of matchings.

Proposition 1.5. There exists a constant c > 1 such that for all d ≥ 3

cn · m′ (d) ≤ f≤2 (d) ≤ f (d). (6)

The following result is a simple upper bound on f (d). Note that a sketch of the
proof appeared already in [23].

Proposition 1.6.
f (d) ≤ (d + 1)n ≤ en+n/d · N.

The following Propositions show the effect of allowing 1- and 2-dimensional (and
in general small dimensional) subcubes in a partition.

Proposition 1.7. For every fixed r ≥ 2

r−1 −o(1) r−1 +o(1)
N r/2 ≤ f0,r (d) ≤ f0,r,r+1,r+2,...(d) ≤ N r/2 .

Proposition 1.8.

f0,1,3,4,... (d) ≤ exp(20n/d1/4 ) · N ≤ exp(20n/d1/4 ) · f1 (d) = exp(20n/d1/4 ) · m(d).

It is natural to ask what happens if only 2-dimensional subcubes are allowed to

be in a partition. We believe that the same bound as in Proposition 1.7, with r = 2
holds.

Problem 1.9. Determine the asymptotic behaviour of the number f2 (d).

Additionally, we left open another question addressing Problem 1.2 (iii).

Problem 1.10. Is it true that f (d)/f≤2(d) is subexponential in n?

A subcube partition is an irreducible, if no subcube is spanned by a subfamily

of the partition, and it is tight if the partition is ‘proper’, in the sense that every
coordinate is used, i.e., every coordinate is fixed in at least one subcube of the
partition. For example, the partition of the 2-cube into the two subcubes {0∗}, in
which the first coordinate is fixed to 0, and {1∗}, in which the first coordinate is
fixed to 1, is not a tight partition, as the second coordinate is not used.
Irreducible subcube partitions appear in a work of Kullmann and Zhao [18] and
variants are described in several other papers, see [10] for relevant references. Note

4
that it is not immediately clear that there exists an irreducible tight partition for
every given (large) d.
Peitl and Szeider [22] enumerated all tight irreducible subcube partitions for
d = 3, 4, and asked whether there are infinitely many such partitions. Filmus,
Hirsch, Kurz, Ihringer, Riazanov, Smal, and Vinyals [10] answered this question
in the affirmative, giving many explicit constructions of tight irreducible subcube
partitions. Here we prove the existence of many more tight irreducible partitions.

Proposition 1.11. The number of irreducible tight partitions of the d-dimensional

cube is at least cn for some absolute constant c > 1, where, as before, n = 2d−1 .

We shall frequently use the Stirling formula:

√  m m √  m m
2πm · ≤ m! ≤ e1/(12m) · 2πm · .
e e
Furthermore, we use the binary entropy function h(x) := −x log2 x−(1−x) log 2 (1−x)
to estimate the binomial coefficients for 0 < x < 1/2:
X n  
n
= Θ(1) · = 2(1+o(1))h(x)·n . (7)
k≤xn
k xn

All logarithms throughout the paper have base 2, unless otherwise indicated. To
simplify the presentation, we omit all floor and ceiling signs whenever these are not
crucial. We also assume, whenever this is needed, that d (and hence also n) is
sufficiently large.
In Section 2 we prove Propositions 1.3, 1.4, and 1.5. In Section 3 we prove
Propositions 1.6, 1.7, 1.8, and 1.11.
Remark: We mention very briefly two motivations for the study of f (d). First, f (d)
is the number of so-called bijective commutative Boolean networks of dimension d.
See [6] for the definition of this notion and the fact that there is a bijection between
partitions of the d-cube into subcubes and such networks.
Secondly, f (d) is the number of instances of SAT on d Boolean variables such
that any truth assignment fails to satisfy exactly one clause of the instance. Indeed,
to any clause (e.g. x1 ∨ ¬x3 ) we can associate the subcube of assignments that fail
to satisfy it (x1 = 0, x3 = 1 in the example above). Then those subcubes partition
the d-cube if and only if any truth assignment belongs to exactly one of them, i.e., it
fails to satisfy exactly one clause. This is equivalent to the description from OEIS,
which reads: “the number of ways to make a tautology from disjoint terms with d
Boolean variables”.

5
2 Matchings and 2-dimensional subcubes
First, we list some of the classical results that we shall use in our proofs.
Bregman-Minc inequality: The celebrated Bregman-Minc inequality, conjec-
tured by Minc [20] and proved by Bregman [5] (see also [25], [3], [24] for short proofs)
implies that the maximum possible number of perfect matchings in an a-regular bi-
partite graph with b vertices in each vertex class is at most (a!)b/a . Equality is
achieved when b is divisible by a, by a vertex disjoint collection of complete bipartite
graphs. When the degrees of the vertices on one side are a1 , . . . , ab , with average
degree a, the following upper bound holds: Πi (ai !)1/ai ≤ (a!)b/a .
Van der Waerden inequality: In 1926 Van der Waerden [27] conjectured that
the minimum possible value of the permanent of a b-by-b doubly stochastic matrix
is b!/bb , achieved by the matrix in which all entries are 1/b. Proofs of this conjecture
were given in the early 80s by Falikman [9] and by Egorychev [8], see also Gyires [16].
Schrijver’s bound: Schrijver [26] (see also [15] for subsequent work) proved that
every a-regular bipartite graph with b vertices in each class has at least
b
(a − 1)a−1


aa−2
perfect matchings.
Proof of Proposition 1.3 (i). The Bregman-Minc inequality gives that

b/a
√  a a
1/(12a) b/a (1/2+o(1)) log a·(b/a)
 a b
m(G) ≤ (a!) ≤ [ 2πa · ·e ] =2 · .
e e
The Van der Waerden inequality gives the following lower bound on m(G):
 b
b! √ b
m(G) ≥ b · ab ≥ 2πb · · b−b · ab .
b e

Using Schrijver’s bound, and some delicate estimates on e, we obtain the following
improved asymptotics for the lower bound:
b    b
(a − 1)a−1 a b e · (a − 1)a−1

m(G) ≥ =
aa−2 e aa−1
 a b   (a−1)b
1 1 1 1
≥ · 1+ + + 1−
e a − 1 2(a − 1)2 6(a − 1)3 a
 a b   (a−1)b
1 1  a b
= · 1+ + ≥ · eb/2a .
e 2a(a − 1) 6a(a − 1)2 e

6
 The inequality in the second line above is equivalent to the statement that
1 1 1
e1/(a−1) ≥ 1 + + 2
+ ,
a − 1 2(a − 1) 6(a − 1)3

which follows from the fact that the right-hand-side is the sum of the four first terms
in the power series of the left-hand-side, in which all terms are positive. The final
inequality is equivalent to the fact that
1 1
e1/(2a(a−1)) ≤ 1 + + .
2a(a − 1) 6a(a − 1)2

This inequality holds for every a ≥ 2 since the power series of the left-hand-side is
1 1
1+ + 2 + ...
2a(a − 1) 8a (a − 1)2

The first two terms here are equal to the first two terms in the expression above, and
it is easy to see that for a ≥ 2 the sum of all the remaining terms is smaller than
1
6a(a−1)2
.

n
(ii) Follows from (i) by setting a := d, b := n and N = de .

Proof of Proposition 1.4. (i) Note that the second equation instantly follows from
(2), we just need to prove the first equation.
√
To prove the lower bound, set t = b/ a. For every perfect matching M of G,
let F be a random subset of t edges of M, chosen uniformly among all subsets of
cardinality

t of M. Note that M − F is a matching (of size b − t). This provides
m(G) · bt matchings, but the same matching may be obtained multiple times. More
precisely, the number of times such a matching M − F appears is exactly the number
of perfect matchings in the induced subgraph of G on the set of vertices V (F ) covered
by the edges of F . The expected number of edges in this induced subgraph is exactly
t


t + (b · a − b) 2b
< t + b.
2

Indeed, the subgraph contains exactly t edges of M, and each
edge
that does not
t b 2 2
belong to M lies in this induced subgraph with probability 2 / 2 < t /b = 1/a.
The above estimate thus follows from the linearity of expectation. By Markov’s
Inequality it follows that with probability at least, say, 1/3a, the number of edges in
this induced subgraph is smaller than b+ 2t. This gives that with probability at least

7
1/3a the induced subgraph of G on V (F ) has average degree smaller than (b+2t)/t =
√
a + 2. By Minc Conjecture and the fact that as shown in [1], Corollary 2.3 (see
also [3], page 66) the upper bound provided by the Bregman-Minc inequality for the
permanent of a {0, 1}-matrix with a given number of 1 entries is obtained when all
the rows have the same sum, it follows that the number of perfect matchings spanned
by V (F ) is at most
√ √ √
[( a + 2)!]t/( a+2) ≤ [(1 + o(1)) a/e]t .

Therefore,
√ inteach of these cases the matching M −F obtained is counted at most
[(1 + o(1)) a/e] times. It follows that the number of matchings of size b − t in G is
at least
b


1 t
√
2 b/ a
· m(G) · √ = m(G) · [(1 + o(1))e ] ,
3a [(1 + o(1)) a/e]t
implying the desired lower bound.
For the proof of the upper bound for m′ (G) denote by M(t) the set of matchings

2
in G with b − t edges, where 0 ≤ t ≤ b. For a given t, there are bt ways to
choose the set of uncovered vertices, and given this choice, by the Bregman-Minc
inequality, there are at most (a!)(b−t)/a ways to place a perfect matching on the rest
of the vertices. Therefore
 2
b
|M(t)| ≤ · (a!)(b−t)/a .
t
To bound the sum of these terms for all t define x = (a!)1/a and observe that this
sum satisfies
!2
X b2 X b √ √ 2b √ 2b

1
2b
b−t b−t
x ≤ ( x) = (1 + x) = ( x) 1+ √ .
t
t t
t x
Using Stirling’s formula it is easy to check that
2b
√  a b 
2b b/a O(b log a/a) 1 √ √
( x) = (a!) ≤ e and 1+ √ ≤ eO(b/ x) = eO(b/ a) .
e x

b √
This provides the required upper bound ae · 2O(b/ a) .
d n


(ii) Follows from (i) by setting a := d, b := n and N = e
.

Next we describe the proof of Proposition 1.5, starting with an outline of the proof.
We have to prove that there is a constant c > 1 such that f (d) > cn · m′ (d). To do so

8
we fix a small positive constant α and choose randomly and uniformly (1+o(1))αn 2-
dimensional subcubes of Qd . Whenever two such chosen cubes have common vertices,
we remove one of them, noting that typically the number of subcubes removed is
only O(α2 )n. This gives a collection B of some βn = (α − O(α2))n 2-dimensional
subcubes, and by lower bounding the probability that the set B produced is of size
βn we get that there are many distinct choices for such sets.
We complete this partial covering by placing a (nearly) perfect matching on the
rest of the graph, which has (1 − 2β)n vertices in each of the vertex classes, and is
roughly (1 − 2β)d regular. Since the rest of the graph is not exactly regular, we do
not have a good lower bound on the number of its matchings. Therefore, we create a
regular bipartite graph by adding some vertices and removing some edges to make it
regular. This enables us to apply the lower bound in Proposition 1.3 and to estimate
the number of ways to add a matching and single vertices (0-dimensional cubes) for
each collection B of 2-cubes. A careful computation then yields the desired estimate.

Proof of Proposition 1.5. The goal is to prove that there is a constant c > 1 such
that f (d) > cn · m′ (d). Let α > 0 be a sufficiently small constant, c > 1 will depend
on the choice of α. Consider the hypercube Qd as a bipartite graph with vertex
classes (U, V ), (the even vertices and the odd ones). Let A ⊂ V be a random subset
of V obtained by picking each vertex of V , randomly and independently, to lie in A
with probability α. Therefore the size of A is close to αn with high probability. For
each ui ∈ A choose a vertex vi ∈ V uniformly and randomly among the d2 choices

so that {ui , vi } together with two additional vertices from U span a copy of C4 .
This way we get a partial covering of Qd by the 2-cubes spanned by {(ui , vi ) :
ui ∈ A}. We try to complete this covering by placing a (nearly) perfect matching
on the rest of the graph, which has about (1 − 2α)n vertices in each of the vertex
classes, and is roughly (1 − 2α)d regular. Our gain on the number of cube partitions
will come from the number of ways of choosing the set of 2-cubes. Unfortunately, the
rest of the graph is not exactly regular, hence we do not have a good lower bound on
the number of its matchings. To go around this we create a regular bipartite graph
by adding some vertices and removing some edges to make it regular.
In case two 2-cubes are overlapping, we remove one of them. With high probabi-
lity, the number of removed vertices is Θ(α2 n). We thus assume this is the case, (and
count only partitions in which this holds). This way we obtain a partial C4 -covering
B of Qd (that is, B is a family of vertex disjoint C4 ’s), with |B| = (α−Θ(α2 ))n =: βn,
i.e., (α −β)n = Θ(α2 )n. When α > 0 is sufficiently small, then the number of choices
for B is at least

9
    αn  −1  −(α−β)n
1 n d n d
2
· · · ≥ 2(h(α)/2)n · d2βn . (8)
n αn 2 (α − β)n 2
The first 1/n2 factor above is for considering only random choices of A in which
|A| = αn and |B| = βn, meaning that B is
obtained from A by removing exactly
1 n d αn


(α − β)n 2-cubes. There are at least n2 αn · 2 ways to choose the collection A
so that this holds, and each such A produces a collection B. The number of times a
fixed collection B can be obtained this way is the number of ways to add (α − β)n
2-cubes to this fixed collection, where these added subcubes contain (α−β)n distinct
vertices ui in V and the subcube is determined
by ui and a vertex vi of Hamming
n
distance 2 from ui . There are at most (α−β)n ways to choose the vertices ui and

(α−β)n
given those, at most d2 to choose the corresponding vertices vi . Therefore,

(α−β)n
dividing by the product (α−β)n · d2
n


ensures that each partial collection B
of pairwise disjoint 2-subcubes is counted at most once. The last inequality follows
from the fact that for small fixed α > 0
   
n (1+o(1))h(α)n 2 αn h(O(α2 ))n 2 αn n
= 2 ≫n 2 ·2 = n ·2 · .
αn (α − β)n

Let H be the (random) graph spanned by Qd −V (B). Observe that if u, v ∈ V (H)

and the distance between them in Qd is a least, say, 10, then their degrees in this
graph are independent.

Claim 2.1. With high probability, all but at most n/d vertices in H have degree in
the interval J := [(1 − 2β)d − d2/3 , (1 − 2β)d + d2/3 ].

Proof. It is easy to see that for each fixed vertex of H, its degree in H lies in the
interval J with high probability. This follows, for example, from Azuma’s martingale
inequality (c.f., e.g., [3]). Note, however, that this does not suffice to imply the claim,
as the events corresponding to distinct vertices are not independent.
To complete the proof of the claim partition V (H) into d10 classes, so that in
each class the vertices are at distance at least 10 from each other. In each class,
the degrees of the vertices in H are independent. Applying the Chernoff bound to
each class, we obtain that each class contains about the expected number of vertices
whose degrees are not in the interval. The claimed result follows by the union bound,
with room to spare.
Returning to the proof of the proposition, we assume that the assertion of the
claim holds (and only count such partitions). We thus assume that there are at most

10
n/d vertices in H with degrees not in J. As long as there is a vertex with degree larger
than (1 −2β)d + d2/3, we remove an arbitrary subset of its edge set to make its degree
(1 − 2β)d + d2/3 . This way we remove a total of at most n/d · 2β · d = 2βn edges. The
number of edges missing in each vertex class in order to get a ((1−2β)d+d2/3)-regular
graph is smaller than

n/d · ((1 − 2β)d + d2/3 ) + n · 2d2/3 + 2βn < 2.5d2/3 · n.

We now add auxiliary vertices to H, to create a ((1 − 2β)d + d2/3 )-regular graph
F . This can be done by adding at most 2.5d2/3 · n/(d(1 − 2β)) ≤ 3d−1/3 · n vertices
to each class.
Consider an arbitrary fixed perfect matching M of F . By the lower bound in
Proposition 1.3, the number of such M is at least
(1−2β)n  (1−2β)n
(1 − 2β)d + d2/3

(1−2β)n d
≥ (1 − 2β) · .
e e
Using M, we define the following cube partition of Qd − B. We let the edges of
M be the 1-dimensional cubes, with the (obvious) restriction that if an endpoint of
an edge of M is not in Qd , then it is not used, and if exactly one end point of the
edge is in V (Qd ), then it is considered as a 0-dimensional subcube of the partition.
Note that the number of such 0-cubes is at most 3n/d1/3 in each class.
This way, we obtain a cube covering using subcubes of dimensions 0, 1, 2; but
we might obtain the same covering from different matchings M. The overcounting
is at most the number of ways a perfect matching could be placed on the vertices
of V (F ) − V (Qd ) and the vertices of the 0-dimensional cubes. The number of such
vertices in V is at most 6n/d1/3 . Hence, by the Bregman-Minc inequality, the number
4/3
of ways to cover them with a perfect matching is at most (d!)6n/d . To summarize,
the number of tilings using subcubes of dimensions {0, 1, 2} is at least
 (1−2β)n
1 (h(α)/2)n 2βn (1−2β)n d 4/3
·2 · d · (1 − 2β) · · (d!)−6n/d > 2(h(α)/3)n · N > cn · m′ (d).
2 e

The first 1/2 factor takes care of the fact that we count only partitions in which
the required high probability events hold. The constant c can be chosen, for exam-
ple, to be c = 2(h(α)/4) > 1, where α > 0 is a sufficiently small absolute constant.

11
3 Upper bound on f (d) and related estimates
Proof of Proposition 1.6. We need to show that f (d) ≤ (d+1)n . Let v1 , v2 , . . . , vn
be an enumeration of all the n even vertices of the hypercube Qd (for example, lexico-
graphically), and fix a similar enumeration of all odd vertices. For each partition P of
the hypercube into subcubes we construct a sequence S = S(P ) = (s1 , s2 , . . . , sn ) of
length n over the alphabet {0, 1, . . . , d} so that S(P ) completely defines the partition
P . This is done as follows. Each element si of S(P ) corresponds to the vertex vi . If
vi lies in a subcube D of the partition P of positive dimension, and it is a neighbor of
the lexicographically first odd vertex u of D, then si = j, where j is the coordinate
in which vi and u differ. If vi lies in such a subcube D and is not a neighbor of u
then si = 0, and this is also the case if vi forms a subcube of P of dimension 0. This
completes the definition of S = S(P ). It is easy to see that given S = S(P ), we can
reconstruct all the odd vertices which are the lexicographically smallest ones in sub-
cubes in the partition P of positive dimension. For each of those, we can reconstruct
the corresponding subcubes, and then we also get all the remaining vertices which
are subcubes of dimension 0. This shows that f (d) ≤ (d + 1)n and completes the
proof of the proposition.
Proof of Proposition 1.7. First we prove the lower bound. A slight modification of
the Frankl-Rödl [11] nibble method, as pointed out by Grable and Phelps [13], gives
an estimate on the number of almost perfect matchings in r-uniform hypergraphs.
We use a version from Asratian and Kuzjurin [4]:
Theorem 3.1. The following holds for every fixed small δ > 0. Let r be fixed and
let H2n be an r-uniform D-regular hypergraph on 2n vertices, where both D and n
are tending to infinity. Furthermore, assume that the maximum codegree is o(D),
where the codegree of a pair of vertices is the number of hyperedges containing both.
Then, the number of matchings of H2n with at least (1 − δ)2n/r hyperedges is at least
D (1−2δ)(2n/r) .
To prove the proposition, define a 2r -uniform hypergraph H2n on the vertex set
of Qd ,
where the hyperedges are the r-subcubes of Qd . The hypergraph H2n is
d d−1
D = r -regular, with maximum codegree r−1 . A matching M of size (1 − δ)n/2r−1

corresponds to a {0, r}-covering of Qd , where if a vertex is not covered by M, then

we cover it by a 0-cube, and the hyperedges of M correspond to r-cubes. As δ can
be chosen to be arbitrarily small, this implies that
 (21−r −o(1))n
d r−1
f0,r (d) ≥ = N r/2 +o(1) .
r

12
 To prove the upper bound we apply the same encoding that appears in the proof
of Proposition 1.6 to each of the relevant partitions here. The improved bound is
obtained since every sequence that appears in this encoding contains only a small
number of nonzero entries. Indeed, note that each partition here contains no sub-
cubes of dimensions 1, 2, . . . , r−1, and for each subcube D of dimension k > 0 (which
must be at least r) we get exactly k nonzero elements of the sequence, and 2k−1 − k
zeros. All other elements of the sequence that correspond to subcubes of dimension
0 (consisting of an even vertex) are also 0. This gives sequences of length n in which
the fraction of non-zeros is at most r/2r−1 providing an upper bound of
X n r−1
dj = N r/2 +o(1) ,
r−1
j
j≤rn/2

that provides the required result.

Proof of Proposition 1.8. We estimate the number of cube partitions which do
not use 2-dimensional subcubes. To do so we bound the number of partitions in
which exactly T vertices in each of the two vertex classes, the even (U) and the odd
(V ), are covered by cubes
of dimension at least 3, where here 0 ≤ T ≤ n:
n
– there are at most T ways to choose a set A of T even vertices covered by cubes
of dimension at least 3,
– given such choices, we can construct a sequence of length T over the alphabet
{0, 3, 4, . . . , d} by following the procedure in the previous proofs, but here it is applied
only to the set A of T chosen even vertices. More precisely, the sequence produced
corresponds to the enumeration of the vertices of A induced by the lexicographic
order. For each subcube D of the partition of dimension at least 3 let u be its
lexicographically first odd vertex. If a vertex v of A belongs to D and is a neighbor
of u then the corresponding element of the sequence is j, where u and v differ in
coordinate j, and if v is in D but is not a neighbor of u then the corresponding
element is 0. This produces a sequence of length T over {0, 1, 2, . . . , d} in which the
fraction of nonzero elements is at most 3/4. Note that the sequence enables one
to reconstruct all the subcubes of dimension at least 3 in the partition. Therefore,
for each fixed choice of A there are less than 2T · d3T /4 ways to place the cubes of
dimension at least 3.
– Using Proposition 1.4 (the statement holds for maximum degree a, instead of a-
regular) with a =√ d and b = n − T , given the above choices, there are at most
d n−T
· 2Θ((n−T )/ d) ways to place a perfect matching on the rest of the vertices.


e

13
Hence, for a fixed T the number of these cube partitions is at most
   n−T
n T 3T /4 d √
·2 ·d · · 2Θ((n−T )/ d) .
T e

Summing up over all choices of T , and using the binomial theorem, we obtain the
upper bound √
(1 + 2ed−1/4 )n · 2Θ(n/ d) · N ≤ exp(20n/d1/4 ) · N.

Proof of Proposition 1.11. The construction is recursive. By induction on the

dimension d we construct many irreducible tight partitions of Qd , all containing one
specific subcube. We start in some fixed dimension d with at least 3 irreducible tight
partitions of Qd all containing some subcube. Given two distinct partitions B1 , B2
among those consider the partition B = B1 0 ∪ B2 1 of Qd+1 in which the partition B1
is used in the first hyperplane in which the last coordinate is 0 and the partition B2
is used in the second hyperplane in which the last coordinate is 1. Since B1 and B2
have the same subcube D we replace D0 and D1 in B by their union D ′ = D0 ∪ D1 .
By replacing all (the nonzero number) of such D we obtain this way an irreducible
tight partition of Qd+1 from each ordered pair of partitions B1 , B2 we had, and all
these new partitions contain the same subcube D ′ . It is not difficult to see that all
the partitions obtained are tight since B1 , B2 are tight and different. Each partition
obtained is also irreducible since no subcube in which the last coordinate is fixed
can be spanned by a subfamily, as B1 , B2 are irreducible. Similarly, if a subcube in
which the last coordinate is not fixed is spanned by a subfamily, then the construction
ensures that its two subcubes consisting of all vertices with a fixed last coordinate
are also spanned by a subfamily of B1 and a subfamily of B2 . As all the partitions
obtained are distinct, this shows that if the number of irreducible tight partitions we
get for dimension d in this way is xd , then xd+1 = xd (xd − 1), implying the desired
lower bound.
Acknowledgment. We are grateful to Maximilien Gadouleau for pointing out
reference [6], and for explaining the motivation. We also thank Artur Riazanov and
Yuval Filmus for useful comments. Part of this work was completed while the second
author visited Tel Aviv University, partially supported by the Simons Foundation, by
the Mathematisches Forschungsinstitut Oberwolfach, and by the ERC Consolidator
Grant 101044123 (RandomHypGra) of Wojciech Samotij. Finally, we thank the
anonymous referees for many useful comments.

14
References
[1] N. Alon: The maximum number of Hamiltonian paths in tournaments. Combi-
natorica, 10 (1990), 319–324.

[2] N. Alon, S. Friedland: The Maximum Number of Perfect Matchings in Graphs

with a Given Degree Sequence. The Electronic J. Combinatorics, 15 (2008),
N13.

[3] N. Alon, J.H. Spencer: The Probabilistic Method. Fourth Edition. John Wiley
& Sons, Inc. (2016).

[4] A.S. Asratian, N.N. Kuzjurin: On the number of nearly perfect matchings in
almost regular uniform hypergraphs. Discrete Math., 207 (1-3), (1999), 1–8.

[5] L.M. Bregman: Some properties of nonnegative matrices and their permanents.
Soviet Math. Dokl. 14 (1973), 945–949.

[6] F. Bridoux, M. Gadouleau, G. Theyssier: Commutative automata networks.

Proc. international workshop on cellular automata and discrete complex systems,
Stockholm, Sweden, August 2020, pp. 43–58.

[7] P. Csikvári: Lower matching conjecture, and a new proof of Schrijver’s and
Gurvits’s theorems. J. of the European Math. Soc., 19 (2017), 1811–1844.

[8] G.P. Egorycev: Reshenie problemy van-der-Vardena dlya permanentov (in Rus-
sian). Krasnoyarsk: Akad. Nauk SSSR Sibirsk. Otdel. Inst. Fiz., (1980), p. 12.
G.P. Egorychev: Proof of the van der Waerden conjecture for permanents.
Akademiya Nauk SSSR (in Russian), 22 (6), (1981), 65–71, 225.
G.P. Egorychev: The solution of van der Waerden’s problem for permanents.
Advances in Mathematics, 42 (3), (1981), 299–305.

[9] D.I. Falikman: Proof of the van der Waerden conjecture on the permanent of
a doubly stochastic matrix. Akademiya Nauk Soyuza SSR, (in Russian), 29 (6),
(1981), 931–938, 957.

[10] Y. Filmus, E. Hirsch, S. Kurz, F. Ihringer, A. Riazanov, A. Smal, M. Vinyals:

Irreducible subcube partitions. Electron. J. Combin., 30 (2023), Paper 3.29, 51
pp.

[11] P. Frankl, V. Rödl: Near perfect coverings in graphs and hypergraphs. European
J. Combin., 6 (4), (1985), 317–326.

15
[12] S. Friedland, E. Krop, K. Markström: On the Number of Matchings in Regular
Graphs. The Electron. J. Combin., Volume 15 (2008), R110.

[13] D.A. Grable, K.T. Phelps: Random methods in design theory: a survey. J. Com-
bin. Des., 4 (1996), 255–273.

[14] N. Graham, F. Harary: The number of perfect matchings in a hypercube. Appl.

Math. Lett. 1 (1988), 45–48.

[15] L. Gurvits: Van der Waerden/Schrijver-Valiant like conjectures and stable

(aka hyperbolic) homogeneous polynomials: One theorem for all. Electron. J.
Combin. 15 (2008) RP 66, pp. 1–26.

[16] B. Gyires: The common source of several inequalities concerning doubly stochas-
tic matrices. Publicationes Mathematicae Institutum Mathematicum Universi-
tatis Debreceniensis, 27 (3–4), (1980), 291–304.

[17] M. Jenssen, W. Perkins: A proof of the upper matching conjecture for large
graphs. J. Combin. Theory Ser. B 151 (2021), 393–416.

[18] O. Kullmann, X. Zhao: Unsatisfiable hitting clause-sets with three more clauses
than variables. CoRR, abs/1604.01288, 2016.

[19] L. Lovász, M. D. Plummer: Matching Theory. AMS Chelsea Publishing, Provi-

dence, RI, 2009. 554 pp.

[20] H. Minc: Upper bounds for permanents of (0, 1)-matrices. Bull. Amer. Math.
Soc., 69 (1963), 789–791.

[21] P.R.J. Östergård, V.H. Pettersson: Enumerating perfect matchings in n-cubes.

Order, 30 no. 3 (2013), 821–835.

[22] T. Peitl, S. Szeider: Are hitting formulas hard for resolution? Discrete Applied
Mathematics, 337 (2023), 173–184.

[23] V. N. Potapov: A lower bound on the number of Boolean functions with median
correlation immunity. Proc. 2019 XVI International Symposium “Problems of
Redundancy in Information and Control Systems”, pp. 45–46.

[24] J. Radhakrishnan: An entropy proof of Bregman’s theorem. J. Combin. Theory

Ser. A 77 (1997) 161–164.

16
[25] A. Schrijver: A short proof of Minc’s conjecture. J. Comb. Theory Ser. A, 25
(1978), 80–83.

[26] A. Schrijver: Counting 1-factors in regular bipartite graphs. Journal of Combi-

natorial Theory, Series B, 72(1) (1998), 122–135.

[27] B.L. van der Waerden: Aufgabe 45. Jber. Deutsch. Math.-Verein., 35 (1926),
117.

17