Dirac’s Theorem for hamiltonian Berge cycles in uniform

hypergraphs
Alexandr Kostochka∗ Ruth Luo† Grace McCourt‡

September 28, 2021

arXiv:2109.12637v1 [math.CO] 26 Sep 2021

Abstract
The famous Dirac’s Theorem gives an exact bound on the minimum degree of an n-vertex
graph guaranteeing the existence of a hamiltonian cycle. We prove exact bounds of similar
type for hamiltonian Berge cycles in r-uniform, n-vertex hypergraphs for all 3 ≤ r < n. The
bounds are different for r < n/2 and r ≥ n/2, and the proofs are different for r < (n − 2)/2 and
r ≥ (n − 2)/2.
Mathematics Subject Classification: 05D05, 05C65, 05C38, 05C35.
Keywords: Berge cycles, extremal hypergraph theory, minimum degree.

1 Introduction and Results

Terminology. A hypergraph H is a family of subsets of a ground set. We refer to these subsets as
the edges of H and the elements of the ground set as the vertices of H. We use E(H) and V (H) to
denote the set of edges and the set of vertices of H respectively. We say H is r-uniform (r-graph,
for short) if every edge of H contains exactly r vertices. A graph is a 2-graph.
The degree dH (v) of a vertex v in a hypergraph H is the number of edges containing v. The
minimum degree, δ(H), is the minimum over degrees of all vertices of H.
A hamiltonian cycle in a graph is a cycle which visits every vertex. Sufficient conditions for
existence of hamiltonian cycles in graphs have been well-studied. In particular, the first extremal
result of this type was due to Dirac in 1952.

Theorem 1 (Dirac [4]). Let n ≥ 3. If G is an n-vertex graph with minimum degree δ(G) ≥ n/2,
then G has a hamiltonian cycle.

Recently, a series of results of the same flavor was proved for Berge cycles in hypergraphs.

Definition 1.1. A Berge cycle of length ℓ in a hypergraph is a list of ℓ distinct vertices and ℓ
distinct edges (v1 , e1 , v2 , . . . , eℓ−1 , vℓ , eℓ , v1 ) such that {vi , vi+1 } ⊆ ei for all 1 ≤ i ≤ ℓ (here we take
indices modulo ℓ). Similarly, a Berge path of length ℓ is a list of ℓ + 1 distinct vertices and ℓ
distinct edges (v1 , e1 , v2 , . . . , eℓ , vℓ+1 ) such that {vi , vi+1 } ⊆ ei for all 1 ≤ i ≤ ℓ.
∗
University of Illinois at Urbana–Champaign, Urbana, IL 61801 and Sobolev Institute of Mathematics, Novosibirsk
630090, Russia. E-mail: [email protected]. Research is supported in part by NSF RTG Grant DMS-1937241
and grant 19-01-00682 of the Russian Foundation for Basic Research.
†
University of California, San Diego, La Jolla, CA 92093. E-mail: [email protected]. Research is supported in part
by NSF grant DMS-1902808.
‡
University of Illinois at Urbana–Champaign, Urbana, IL 61801, USA. E-mail: [email protected]. Research
is supported in part by NSF RTG grant DMS-1937241.

1
 Although the edges in a Berge cycle may contain other vertices, we say V (C) = {v1 , . . . , vℓ },
and E(C) = {e1 , . . . , eℓ }. The vertices in V (C) are called key vertices. Notation for Berge paths is
similar.
An analogue of Dirac’s Theorem for non-uniform hypergraphs was given in [7]. For r-uniform
hypergraphs, Coulson and Perarnau proved the following result for n much larger than r.
√
Theorem 2 (Coulson and Perarnau [3]). Let H be an r-graph on n vertices such that r = o( n).
If H has minimum degree δ(H) ≥ ⌊(n−1)/2⌋


r−1 + 1, then H contains a hamiltonian Berge cycle.
This bound in Theorem 2 on δ(H) is sharp. Moreover, Coulson and Perarnau’s result is a
√
corollary of their stronger result for rainbow cycles in graphs. With their methods, r = o( n)
√
is best possible, so the case r = Ω( n) remained open. Various results by Bermond, Germa,
Heydemann, and Sotteau [1]; Clemens, Ehrenmüller, and Person [2]; and Ma, Hou, and Gao [9]
√
provided upper bounds for the minimum degree when r = Ω( n). The best bound to date was
due to Ma, Hou and Gao.
Theorem 3 (Ma, Hou and Gao [9]). Let r ≥ 4 and
n ≥ 2r + 4, and let H be an r-graph on n
vertices. If H has minimum degree δ(H) ≥ ⌊(n−1)/2⌋
r−1 + ⌈(n − 1)/2⌉, then H contains a hamiltonian
Berge cycle.
In this paper we derive exact bounds for all possible 3 ≤ r < n, improving the aforementioned
theorems.
Theorem 4. Let t = t(n) = ⌊(n − 1)/2⌋, and suppose 3 ≤ r < n. Let H be an r-graph with
 t

δ(H) ≥ r−1 + 1 when r ≤ t,
r when r ≥ n/2.

Then H contains a hamiltonian Berge cycle.

This bound is best possible due to the following constructions.
(r)
Construction 1. Suppose r ≤ t. If n is odd, let H1 consist of two copies of K(n+1)/2 that share
exactly one vertex. If n is even, let H1 consist of two disjoint Kn/2 r and a single edge intersecting
both cliques.
Construction 2. Suppose r ≤ t. Let H2 have vertex set X ∪ Y such that |X| = t and
|Y | = n − t. The edge set of H2 consists of every edge with at most one vertex in Y .
Construction 3. Suppose r ≥ n/2. Let H3 be obtained by removing a single edge from an
r-uniform tight cycle on n vertices.
t

It is easy to check that both H1 and H2 have minimum degree r−1 . Observe that neither H1
nor H2 have a hamiltonian Berge cycle: H1 has a either a cut vertex or a cut edge, and in H2 a
hamiltonian Berge cycle must visit two vertices in Y consecutively, but no edge of H2 contains any
pair of vertices from Y .
Since an r-uniform tight cycle is r-regular, δ(H3 ) = r − 1. Also, H3 does not have a hamiltonian
Berge cycle because |E(H3 )| = n − 1.
Note that the length of the longest cycle in Construction 1 is ⌈n/2⌉. Thus Theorem 4 yields
exact bounds on the minimum degree guaranteeing the existence of any cycle of length at least k
in n-vertex r-uniform hypergraphs for all r ≤ t and all k ≥ 1 + n/2.
The proofs for r < t and for r ≥ t are quite different. For r < t we use the original approach
of Dirac using lollipops. In Section 2 we set up the proof for this case and prove that our r-graph

2
must have a cycle of length at least t + 2. In Section 3 we analyze lollipops with a long (but
not hamiltonian) Berge cycle, and in all cases get a contradiction. In Section 4 we start the case
r ≥ t, introduce suitable ends and state the main lemmas. In subsequent sections we prove the
lemmas, and in the last section finish the whole proof. The case r = t is especially difficult and
adds substantially to the length of the sections in which we consider r ≥ t.
Also, since we always consider only Berge paths and cycles, below we drop the word “Berge”.

2 Setup for r < t

A lollipop is a pair (C, P ) such that C is a cycle, P is a path, E(C) ∩ E(P ) = ∅, |V (C) ∩ V (P )| = 1,
and the shared vertex of V (C) and V (P ) is one of the endpoints of P . If E(P ) 6= ∅, then the end
of the lollipop (C, P ) is the endpoint vertex of P that is not in C. If E(P ) = ∅, then (C, P ) has no
end.
A lollipop (C, P ) is better than a lollipop (C ′ , P ′ ) if
(i) |E(C)| > |E(C ′ )|, or
(ii) |E(C)| = |E(C ′ )| and |E(P )| > |E(P ′ )|, or
(iii) |E(C)| = |E(C ′ )|, |E(P )| = |E(P ′ )| ≥ 1 and the number of edges in C containing the end of
(C, P ) is greater than the number of edges in C ′ containing the end of (C ′ , P ′ ).
The original proof [4] of Dirac’s Theorem considered a best lollipop and showed that it could
be only a hamiltionian cycle. We use the same strategy for r < t.
t


Let r < t and let H be an n-vertex hypergraph with minimum degree at least r−1 +1. Suppose
H does not have a hamiltonian cycle. Let (C, P ) be a best lollipop in H.
Let C = (v1 , e1 , v2 , . . . , ek−1 , vk , ek , v1 ) with V (C)∩V (P ) = {v1 }, and P = (u1 , f1 , u2 , . . . , fℓ , uℓ+1 )
with u1 = v1 and uℓ+1 = x. So, C has length k, P has length ℓ and x is the end of (C, P ) when
ℓ ≥ 1.
Define H ′ = (V (H), E(H)−(E(C)∪E(P ))). That is, E(H ′ ) is the set of edges not in the lollipop.
For a vertex v, define Ev = {e ∈ E(H ′ ) : v ∈ e} and NH ′ (v) = {v ∈ V (H) : v ∈ e for some e ∈ Ev }.
We call a vertex u an H ′ -neighbor of v if u ∈ NH ′ (v).
Let B = (V (P ) − {x}) ∪ (NH ′ (x) ∩ V (C)) and set b = |B|. In this and next sections, we will
provide better and better bounds on k and ℓ. We begin from a lemma providing that k ≥ t + 2.
Lemma 2.1. |V (C)| ≥ t + 2 = ⌊(n + 3)/2⌋.
Proof. Suppose H has no cycles of length at least t + 2. Let Q be a longest path in H, say
Q = (v1 , e1 , v2 , . . . , es−1 , vs ). Let q = min{t + 1, s}, V (q) = {v1 , . . . , vq } and let Q(q) denote the
subpath of Q with vertex set V (q) and edge set E(q) = {e1 , . . . , eq−1 }. Among such paths Q,
choose one in which
(a) the most edges in E(q) are contained in V (q), and
(1)
(b) modulo (a), the fewest edges in E(q) ∪ {eq } contain v1 .
Let H1 = H − E(Q). Since H has no cycles of length at
least t + 2 and Q is a longest path, all
q−1
neighbors of v1 in H1 are in V (q). Thus dH1 (v1 ) ≤ r−1 . By the same reason, the edges ei for
q + 1 ≤ i ≤ s − 1 must not contain v1 . So
 
q−1
dH (v1 ) ≤ dH1 (v1 ) + min{q, s − 1} ≤ + min{q, s − 1}. (2)
r−1

3
 t−1
t

If q = s ≤ t, then since 3 ≤ r ≤ t − 1, this is at most r−1 + t − 1 ≤ r−1 , contradicting the
′
minimum degree condition. Hence s ≥ t + 1 and q = t + 1. Let E (q) = E(q) ∪ {eq } if eq exists,
and E ′ (q) = E(q) otherwise.
Let E0 be the set of edges in E ′ (q) not containing v1 , E1 be the set of edges in E ′ (q) containing
v1 and contained in V (q), and E2 = E ′ (q) − E0 − E1 . In particular, eq ∈ E0 ∪ E2 .
Let us show that
|E1 ∪ E2 | ≤ t − 1. (3)
Indeed, suppose |E1 ∪ E2 | = m. For every 2 ≤ i ≤ t + 1, we can consider the path Qi
from vi to vs obtained from Q by replacing the subpath (v1 , e1 , v2 , . . . , ei , vi+1 ) with the subpath
(vi , ei−1 , vi−1 , . . . , e1 , v1 , ei , vi+1 ). This path uses the same edges as Q, so by Rule (a) in (1) it
is also a valid choice for a best path, and if vi is in fewer than m edges in E ′ (q), then Qi is
better by Rule (b). Hence each vi such that ei ∈ E1 ∪ E2 is in at least m edges in E ′ (q). Thus,
m2 ≤ r(t + 1) ≤ (t − 1)(t + 1), and so m < t. This proves (3).
Let R = R(v1 ) be the set of r-tuples contained in V (q) that contain v1 and are not edges of H.
Since the only edges containing v1 and not contained in V (q) are those in E2 ,
 
t
dH (v1 ) = + |E2 | − |R|. (4)
r−1
t

So, if E2 = ∅, then dH (v1 ) ≤ r−1 , a contradiction. Hence for some j ∈ [t + 1], ej ∈ E2 , i.e., x ∈ ej
but ej * V (q). Choose the smallest such j.
Case 1: j = 1. If there is an edge g ⊂ V (q) in E(H) − E ′ (q) containing {v1 , v2 }, then by
t−1

replacing e1 with g we get a contradiction to (1)(a). Thus each of the r−2 r-tuples g ⊂ V (q)
containing {v1 , v2 } is in R ∪ E1 .
Case 1.1: r = 3. For any edge ei containing v1 , {vi , vi+1 , v1 } ⊆ ei . Then only e2 may contain
{v1 , v2 } and be contained in V (q). Moreover for 2 ≤ i ≤ t, if x ∈ ei , then ei = {v1 , vi , vi+1 } ⊆ V (q),
so |E2 | ≤ 1. Hence
       
t t t−1 t
dH (v1 ) ≤ − |R| + |{e2 , eq }| ≤ − +2≤ ,
2 2 1 r−1

a contradiction.
Case 1.2: r ≥ 4. It follows that
        
t t−1 t t−1
dH (v1 ) ≤ |E2 | + − − |E1 | = |E1 ∪ E2 | + − .
r−1 r−2 r−1 r−2
t
t−1

In order to have dH (v1 ) ≥ 1 + r−1 , we need r−2 ≤ |E1 ∪ E2 | − 1. Since r − 2 ≥ 2 and
t−1 (t−1)(t−2)
≥ t−1



(t − 1) − (r − 2) ≥ 2, we have r−2 2 . So, using (3), we need 2 ≤ t − 2, which does
not hold for integer t ≥ 4.
Case 2: 2 ≤ j ≤ t. In order for ej to contain v1 , vj , vj+1 and a vertex outside of V (q), we need
r ≥ 4. Similarly to Case 1, if there is an edge g ⊂ V (q) in E(H) − E ′ (q) containing {v1 , vj+1 }, then
the path
(vj , ej−1 , vj−1 , . . . , e1 , v1 , g, vj+1 , ej+1 , vj+2 , . . . , en−1 , vn )

4
 t−1

contradicts (1)(a). Hence each of the r−2 r-tuples g ⊂ V (q) containing {v1 , vj+1 } is in R ∪ E1 .
So, now we repeat the argument of Case 1.2 word by word.
Case 3: j = t +

1. This means all edges containing v1 apart from et+1 are contained in V (q).
t
Then dH (x) ≤ r−1 − |R| + 1, so we may assume |R| = 0. In other words,

all r-tuples contained in V (q) and containing v1 are edges of H. (5)

Since r < t, there is i ≤ t such that vi ∈ / et+1 . By (5), we can construct a path on the vertices
(vi , vi−1 , . . . , v1 , vi+1 , vi+2 , . . . , vt+1 ) all edges of which are contained in V (q). So, we will have no
edges containing vi and not contained in V (q), a contradiction.

Our next step is to show that P is nontrivial.

Lemma 2.2. ℓ ≥ 1. On the other hand, if n is even, then ℓ ≤ t, and if n is odd, then ℓ ≤ t − 1.

Proof. The upper bound follows from the fact that k + ℓ ≤ n and by Lemma 2.1.
For the lower bound, suppose ℓ = 0, that is, P has no edges. Fix any v ∈ V (H) − V (C). If there
exists an edge e ∈ H ′ such that e∩V (C) 6= ∅ and v ∈ e, then the path P ′ with only the edge e forms
together with C a lollipop better than (C, P ). So we may assume that for all v ∈ V (H) − V (C),
every edge e ∈ H ′ containing v does not intersect V (C). Symmetrically, every edge f containing a
vertex u ∈ V (C) either is in E(C) is contained in V (C).
Case 1: There are two consecutive edges ei , ei+1 in C containing v. If v ∈ e for some e ∈ H ′ ,
then we can replace vi+1 with v in C to obtain a cycle C ′ with the same length and a path P ′ with
the unique edge e, so that (C ′ , P ′ ) is better than (C, P ). Thus,
v cannot be belong to any edge of
t
H ′ , hence dH (v) ≤ |C| ≤ n − 1. If n ≥ 11, this is at most r−1 , a contradiction. Suppose n ≤ 10.
Then t ≤ 4. Since r ≥ 3, this means r = 3, t = 4, δ(H) ≥ 7, and n ∈ {9, 10}. So each of v and
vi+1 belongs to at least 7 edges of C, and since r = 3, only ei and ei+1 may contain both. Thus,
|C| ≥ 2 · 7 − 2 = 12 > n, a contradiction.
Case 2: v is not contained in two consecutive edges in C. Then v is contained in at most
n−k−1

r−1 edges in H ′ and at most ⌊k/2⌋ edges in C. By Lemma 2.1 k ≥ t + 2, so |V (H) − V (C)| ≤
n − (t + 2) ≤ t. Thus whenever k ≤ n − 4,

n−k−1 t−1 t−1 t−1

         
t
dH (v) ≤ + ⌊k/2⌋ ≤ + ⌊(n − 4)/2⌋ ≤ + = ,
r−1 r−1 r−1 r−2 r−1

a contradiction.
If k > n − 4, then
 
2 n−1 n+1
dH (v) ≤ + ⌊k/2⌋ ≤ 1 + = .
r−1 2 2
n+1 t

Notice that 2 ≤ r−1 for all n ≥ 9, a contradiction.

Based on these two lemmas, in the next section we prove the bound in a sequence of claims.

5
3 Proof of the bound for r < t
Let W = W (C) = {v2 , . . . , vℓ+1 } ∪ {vk , . . . , vk−ℓ+1 }. When k ≤ 2ℓ + 1, W = V (C) − {v1 }.
Claim 3.1. Suppose x ∈ e ∈ E(H ′ ). Then e ⊆ V (C) ∪ V (P ) − W .
Proof. Suppose first there is a vertex y ∈ V (H) − (V (C) ∪ V (P )) such that y ∈ e. Let P ′ be the
path on V (P ) ∪ {y} obtained by joining the edge e to the end of P . Then (C, P ′ ) is a lollipop with
|V (P ′ )| > |V (P )|, a contradiction.
Now suppose e contains vi for some i ∈ {2, . . . , ℓ + 1}. Recall that x = uℓ+1 and v1 = u1 .
Let C ′ = (v1 , f1 , u2 , . . . , fℓ , uℓ+1 , e, vi , ei , vi+1 , . . . , ek−1 , vk , ek , v1 ) and P ′ be the 1-vertex path.
Then lollipop (C ′ , P ′ ) is better than (C, P ) since |V (C ′ )| ≥ |V (C)| + 1.
The proof for i ∈ {k, . . . , k − ℓ + 1} is similar, so we omit it.

The next two claims have a similar flavor.

Claim 3.2. If uj is an H ′ -neighbor of x in V (P ), then fj does not contain a vertex outside of
V (P ) ∪ V (C) − W .
Proof. Let e be any edge in H ′ containing x and uj .
Suppose first fj contains a vertex y ∈
/ V (P ) ∪ V (C) (so {y, uj , uj+1 } ⊂ fj ). The path

P ′ = (u1 , f1 , u2 , . . . , fj−1 , uj , e, x, fℓ , uℓ , fℓ−1 , uℓ−1 , . . . , fj+1 , uj+1 , fj , y)

is a longer than P ; so lollipop (C, P ′ ) is better than (C, P ), a contradiction.

Suppose now there is an i ∈ {2, . . . , ℓ + 1} such that vi ∈ fj . Let

C ′ = (v1 , f1 , u2 , . . . , fj−1 , uj , e, uℓ+1 , fℓ , uℓ , . . . , fj+1 , uj+1 , fj , vi , ei , vi+1 , . . . , ek−1 , vk , ek , v1 )

and P ′ be the 1-vertex path. Then lollipop (C ′ , P ′ ) is better than (C, P ) since |V (C ′ )| ≥ |V (C)|+1.
The proof for i ∈ {k, . . . , k − ℓ + 1} is similar, so we omit it.

Claim 3.3. x does not belong to any edge in the set {e1 , e2 , . . . , eℓ } ∪ {ek , ek−1 , . . . , ek−ℓ+1 }.
Proof. Suppose i ≤ ℓ and suppose x ∈ ei . The cycle obtained by replacing the segment from v1
to vi+1 with the path (v1 , f1 , u2 , . . . , fℓ , x, ei , vi+1 ) has length at least |V (C)| + 1, contradicting the
choice of C. The argument for ek−i+1 is symmetric, so we omit the proof.

Claim 3.4. For each 1 ≤ i ≤ k, if some edge g ∈ / E(C) contains {x, vi }, then
(a) neither of ei−1 and ei contains x, and
(b) no edge in H ′ contains {x, vi−1 } or {x, vi+1 } (indices count modulo k).
Proof. If x ∈ ei−1 , consider

C ′ = (v1 , e1 , v2 , . . . , ei−2 , vi−1 , g, x, ei , vi , ei+1 , vi+1 , . . . , ek−1 , vk , ek , v1 ).

Then C ′ is longer than C, contradicting the choice of C. If x ∈ ei , the proof is symmetric. This
proves (a).
Suppose now some e ∈ E(H ′ ) contains {x, vi−1 } (the case in which e ∈ E(H ′ ) contains {x, vi−1 }
is symmetric). If e 6= g, then the cycle

(v1 , e1 , v2 , . . . , ei−2 , vi−1 , e, x, g, vi+1 , ei , vi , . . . , ek−1 , vk , ek , v1 )

6
is longer than C, a contradiction.
If e = g, then by (a), x ∈
/ ei−1 . Let

C ′′ = (v1 , e1 , v2 , . . . , ei−2 , vi−1 , e, vi , ei , vi+1 , . . . , ek−1 , vk , ek v1 ).

Cycle C ′′ has k edges, but more edges containing x than C, contradicting condition (iii) in the
choice of (C, P ).

In the remainder of the proof, we let C(x) = {ei : x ∈ ei }, P (x) = {fj : x ∈ fj }, c = c(x) =
|C(x)| and p = p(x) = |P (x)|. We have

dH (x) = |Ex | + c + p. (6)

b

By Claim 3.1, each edge in Ex is contained in V (C)∪V (P )−W , and so |Ex | ≤ r−1 . By Claim 3.3,
c ≤ k − 2ℓ. Since p ≤ ℓ and k + ℓ ≤ n, these facts give
   
b b
dH (x) ≤ + c + ℓ and dH (x) ≤ + n − 2ℓ. (7)
r−1 r−1
Claim 3.5. k > 2ℓ + 1.

Proof. Suppose first that k = 2ℓ + 1. Since n ≥ k + ℓ = 3ℓ + 1, ℓ ≤ (n − 1)/3. By Claims 3.1

and 3.3, B = V (P ) − {x} and the only edge in C that may contain x is eℓ+1 . Therefore by (7),

⌊(n − 1)/3⌋ t−2

       
ℓ t
dH (x) ≤ +1+ℓ≤ + 1 + ⌊(n − 1)/3⌋ ≤ +t−1≤ ,
r−1 r−1 r−1 r−1

a contradiction.
Suppose now that k < 2ℓ + 1. By Claims 3.1 and 3.3, B
= V (P ) − {x}, and moreover x cannot
ℓ
be contained in any edge of C. Then by (6), dH (x) ≤ r−1 + ℓ.
t

If ℓ ≤ t − 1, then this is at most r−1 . So suppose ℓ ≥ t. By Lemma 2.2, ℓ = t and
n = 2t + 2. Suppose that exactly s edges fj1 , . . . , fjs containing x are not contained in V (P ).
ℓ
t

If s = 0, then dH (x) ≤ r−1 ≤ r−1 , a contradiction. Let s ≥ 1. By Claim 3.2, none of
′ ℓ−s

uj1 , . . . , ujs is an H -neighbor of x. Therefore |NH ′ (x)| ≤ ℓ − s, so dH (x) ≤ r−1 + ℓ. Since
ℓ

r − 1 ≤ ℓ − 2, the last expression can exceed r−1 only if s = 1 and r = 3. If s = 1, r = 3
and at least one 3-tuple A ⊂ B ∪ {x} containing x and uj1 is not in {f1 , . . . , fj1 −1 , fj1 +1 , . . . , fℓ },
ℓ


then dH (x) ≤ r−1 − 1 + 1, a contradiction again. Otherwise, for each i ∈ {1, . . . , ℓ} − {j1 }, the
triple {ui , uj1 , x} is in {f1 , . . . , fj1 −1 , fj1 +1 , . . . , fℓ }. For this, at least two of i, j1 and ℓ + 1 must be
consecutive integers. Then j1 ≤ 2 and ℓ + 1 − j1 ≤ 3. Moreover, if ℓ + 1 − j1 = 3, then we know
that fi 6= {uj1 , uj1 +1 , x}, so ℓ + 1 − j1 ≤ 2, and hence ℓ ≤ −1 + j1 + 2 ≤ −1 + 2 + 2 = 3. But
ℓ = t ≥ r + 1 ≥ 4, a contradiction.

From now on, we heavily use the fact that k > 2ℓ + 1. The following technical fact will also be
helpful.

Claim 3.6. If Q = w1 , . . . , wq is a graph path, A is any set of c edges of Q and I is an independent

subset of {w2 , . . . , wq−1 } disjoint from all edges in A, then |I| ≤ ⌈ q−1−c
2 ⌉.

7
Proof. We show the claim by induction on q. Since c ≤ q − 1 and w1 , wq ∈ / I, the claim holds for
q ≤ 2. For the induction step, let Qi denote the subpath w1 , . . . , wi of Q. If wq−1 ∈ / I, then since
(q−1)−1−(c−1) q−1−c
at least c − 1 edges of A are in Qq−1 , by induction |I| ≤ ⌈ 2 ⌉ = ⌈ 2 ⌉, as claimed.
So suppose wq−1 ∈ I. Then wq−2 ∈ / I and all c edges of A are in Qq−2 . Again by induction,
|I ∩ Qq−2 | ≤ ⌈ (q−2)−1−c
2 ⌉, and so |I| ≤ 1 + |I ∩ Qq−2 | ≤ ⌈ q−1−c
2 ⌉. This proves the claim.

Claim 3.7. b = |B| ≤ ⌈ k−1−c

2 ⌉.

Proof. By Claims 3.2 and 3.1, if v is an H ′ -neighbor of x, then v ∈ (V (P )−{x})∪{vℓ+2 , vℓ+3 , . . . , vk−ℓ }.
Since B contains exactly ℓ vertices in P , we need to prove that

k−1−c
 
|B ∩ {vℓ+2 , vℓ+3 , . . . , vk−ℓ }| ≤ − ℓ. (8)
2

By Claim 3.4, no two H ′ -neighbors of x in C may be consecutive. By Claim 3.3, if x ∈ ei then

ei ∈ {eℓ+1 , . . . , ek−ℓ }. Finally, by Claim 3.4, if some vi is an H ′ -neighbor of x, then neither of ei−1
and ei contains x. Thus, since |{vℓ+1 , vℓ+2 , . . . , vk−ℓ+1 }| = k − 2ℓ + 1, applying Claim 3.6 with with
Q = vℓ+1 , . . . , vk−ℓ+1 , A = C(x), and I = B ∩ {vℓ+1 , vℓ+2 , . . . , vk−ℓ+1 } yields our claim.

Claim 3.8. If n is odd, then k = n − 1, and if n is even, then k ∈ {n − 2, n − 1}.

t
Proof. By Claim 3.7, b ≤ ⌈ k−1−c k−1−c


2 ⌉. If ⌈ 2 ⌉ < r − 1, then by (7), dH (x) ≤ n − 2ℓ ≤ n − 2 ≤ r−1
when n ≥ 11. Suppose n ≤ 10, then we must have r = 3, t = 4, and n ∈ {9, 10}.
Additionally, the
t
statement holds if k ≥ 8, so we have k ≤ 7. If ℓ ≥ 2, then dH (x) ≤ n − 2ℓ ≤ r−1 = 6, so we may
assume ℓ = 1. We also know dH (x) ≤ c + ℓ ≤ c + 1, so we reach a contradiction if c ≤ 5. Since
c ≥ 6 and k ≤ 7, x must be contained in all but at most one edge of C. But by Claim 3.3, x is not
contained in e1 or ek , a contradiciton. Therefore we may assume ⌈ k−1−c
2 ⌉ ≥ r − 1.
Suppose first that either n is odd and k ≥ t + 3 or n is even and k ≥ t + 4. Then since
|E(C) ∪ E(P )| ≤ n, we have that ℓ = |E(P )| ≤ n − k ≤ t − 2. Therefore if k ≤ n − 2 and n is odd
or k ≤ n − 3 and n is even,  k−1−c 
⌈ 2 ⌉
dH (x) ≤ + c + t − 2.
r−1
This quantity is maximized when c is either 0 or 1, depending on the parity of k. Thus,
       
⌊(n − 3)/3⌋ t−1 t−1 t
dH (x) ≤ +1+t−2≤ + ≤ ,
r−1 r−1 r−2 r−1

a contradiction.
Otherwise, by Lemma 2.1, either n is odd and k = t + 2 or n is even and k ∈ {t + 2, t + 3}.
If n is odd and k = t + 2 or n is even and k = t + 3, then k ≤ n − 3. Moreover, if n is even and
k = t + 2, then k ≤ n − 4. We have that ℓ ≤ n − k = t − 1 if n is odd or n is even and k = t + 3,
and ℓ = t otherwise.
If n is odd, then by (7),
 n−4−c   n−4−1     
⌈ 2 ⌉ ⌈ 2 ⌉ t−2 t
dH (x) ≤ +c+ℓ≤ +1+t−1 ≤ +t≤ .
r−1 r−1 r−1 r−1

8
Suppose n is even. If k = t + 3, then
 n−4−c   n−4−0   
⌈ 2 ⌉ ⌈ 2 ⌉ t
dH (x) ≤ +c+ℓ≤ +0+t−1≤ .
r−1 r−1 r−1
Otherwise, if k = t + 2, we instead get
 n−5−c   n−4   
⌈ 2 ⌉ 2 t−1
dH (x) ≤ +c+ℓ≤ +t+0= + t.
r−1 r−1 r−1
t

If r ≥ 4, then this quantity is at most r−1 . If r = 3, then for every fi ∈ E(P ) such that x ∈ fi
and i ≤ ℓ − 1, fi = {ui , ui+1 , x}, i.e., fi ⊆ B ∪ {x}. So instead we have
 
b
dH (x) ≤ |{e : x ∈ e, e ⊆ B ∪ {x}}| + |{ei : x ∈ ei }| + |{fℓ }| ≤ +c+1
r−1
 n−5−c   n−4     
⌈ 2 ⌉ 2 t−1 t
≤ +c+1= +1≤ +1< .
r−1 r−1 r−1 r−1
In either case, we get a contradiction.

Claim 3.9. If uj is an H ′ -neighbor of x and x ∈ fj , then fj − {x} ⊆ B.

Proof. Suppose that there exists some v ∈ fj − B. By Claim 3.2, v ∈ V (C) − W , say v = vi , where
ℓ + 2 ≤ i ≤ k − ℓ.
By Claim 3.4, vi ∈ / B + ∪ B − because {vi , x} is contained in fj ∈ / E(C). Thus we must
have vi−1 , vi , vi+1 ∈/ B. By Claim 3.4, we know x ∈ / ei , ei−1 . Thus, taking in Claim 3.6 I =
{vi } ∪ (B ∩ {vℓ+1 , vℓ+2 , . . . , vk−ℓ+1 }) yields b ≤ ⌈ k−1−c
2 ⌉ − 1.
Case 1: n is odd and k = n − 1. Then
 k−1−c   
⌈ 2 ⌉−1 (n − 3)/2
dH (x) ≤ + c + ℓ ≤ max{ + 1, c + ℓ}.
r−1 r−1

We have that (n−3)/2 t

r−1 + 1 ≤ r−1 , and by Claim 3.4 and Claim 3.3, c ≤ k − 3. Hence c + ℓ ≤
t


n − 3 ≤ r−1 as well, a contradiction.
Case 2: n is even and k ∈ {n − 2, n − 1}. Then as in the previous case,
 k−1−c   
⌈ 2 ⌉−1 (n − 4)/2
dH (x) ≤ + c + ℓ ≤ max{ + 2, c + ℓ}.
r−1 r−1

We have that (n−4)/2 t

r−1 + 2 ≤ r−1 , and by Claim 3.4 and Claim 3.3, c ≤ k − 3. Hence c + ℓ ≤
t

n − 3 ≤ r−1 , which gives a contradiction unless n = 10, k = 9, c = 6, and ℓ = 1.
In this remaining case, we get a contradiction if i 6= 2, since by Claim 3.4 this will give c ≤ k − 4.
Also since n = 10, we have r = 3, t = 4. Then f1 = {x, v1 , v2 } and e1 = {v1 , v2 , vi′ } for some i′ . We
may assume 3 ≤ i′ ≤ 6 without loss of generality. Since c = 6, we know x ∈ e3 , e4 , . . . , e8 . Then we
have the cycle

(v2 , f1 , x, ei′ , vi′ +1 , ei′ +1 , vi′ +2 , . . . , e9 , v1 , e1 , vi′ , ei′ −1 , vi′ −1 , . . . , e2 , v2 ).

9
 b

Claim 3.10. If every ui in V (P ) − {x} is an H’-neighbor of x, then dH (x) ≤ r−1 + c. Otherwise
′ b−s

if there exists s vertices in V (P ) − {x} that are not H -neighbors of x, then dH (x) ≤ r−1 + ℓ + c.
Proof. By Claim 3.9, if every ui is an H ′ -neighbor of x, then for any edge fi ∈ E(P ) containing x,
fi ⊆ B ∪ {x}. It follows that the only edges containing x that may also contain vertices outside of
B ∪ {x} are the edges in C. Therefore we have
 
b
dH (x) ≤ |{e : x ∈ e, e ⊆ B ∪ {x}}| + c ≤ + c.
r−1

Next let uj1 , . . . , ujs be non-H ′ -neighbors of x in V (P ) − {x}. By definition, for all e ∈ Ex , e ⊆
b−s

(B − {uj1 , . . . , ujs }) ∪ {x}. We start from (6) and obtain dH (x) ≤ r−1 + ℓ + c.

Claim 3.11. Every vertex in V (P ) − {x} is an H ′ -neighbor of x.

Proof. By Claim 3.8, we only need to consider the cases where n is odd and k = n − 1 or n is even
and k ∈ {n − 1, n − 2}. In all cases, ℓ ≤ 2. By Claim 3.10, if V (P ) − {x} contains a non-H ′ -neighbor
of x, then
 k−1−c  n−2−c
b−1 ⌈ 2 ⌉−1 ⌈ 2 ⌉−1
   
dH (x) ≤ +ℓ+c= +2+c≤ + 2 + c.
r−1 r−1 r−1
This quantity is maximized when c = 0 for
n odd and c
= 1 for n even. In both cases, we get
t−1 t
⌈ n−2−c
2 ⌉ − 1 = t − 1. Therefore dH (x) ≤ r−1 + 3 ≤ r−1 .

Claim 3.12. |V (C)| = n − 1 and n is even.

Proof. Suppose first that n is odd and k = n − 1. By Claims 3.11 and 3.10,
 n−2−c 
⌈ 2 ⌉
 
b
dH (x) ≤ +c≤ + c,
r−1 r−1

which is maximized when c = 0. Thus dH (x) ≤ (n−1)/2 t

r−1 = r−1 , a contradiction.
Now we may assume n is even and k = n − 2. Similarly,
 n−3−c 
⌈ 2 ⌉
 
b
dH (x) ≤ +c≤ + c,
r−1 r−1

which is maximized when c = 0. Thus dH (x) ≤ (n−2)/2 t

r−1 = r−1 , a contradiction.

Finally, we handle the case that n is even and |V (C)| = n − 1 (and so ℓ = 1). If c ≥ 2, then by
Claims 3.11 and 3.10,  n−4   n−2   
2 2 t
dH (x) ≤ +2≤ = ,
r−1 r−1 r−1
n−2

t


a contradiction. If c = 0, then we also have dH (x) ≤ r−12 = r−1 . Thus the only remaining case
is when c = 1.
We now have    n−2   
b 2 t
dH (x) ≤ +c≤ +1= +1
r−1 r−1 r−1

10
with equality if and only if every subset consisting of x along with r − 1 vertices in B is an edge of
H, and x is contained in exactly one edge of C, say ei . By Claim 3.11, u1 = v1 , the unique vertex
in V (P ) − {x}, is an H ′ -neighbor of x. By Claim 3.4, vi and vi+1 are not H ′ -neighbors of x.
Hence we may assume there are exactly (n − 2)/2 = t H ′ -neighbors of x. By Claim 3.4, all
vertices in {v1 , v3 , . . . , vi−1 } ∪ {vi+2 , vi+4 , . . . , vn−2 } are H ′ -neighbors.
Let vj ∈ V (C)−ei be a vertex that is not an H ′ -neighbor of x. By Claim 3.9, vj is not contained
in the unique edge of P . Let e∗ be any edge containing vj . Note that x ∈ / e∗ since vj is not an
H -neighbor and ei is the only edge in C containing x. We claim that all vertices of e∗ are H ′ -
′

neighbors of x except for vj . Suppose for contradiction that vr ∈ e∗ and vr is not an H ′ -neighbor.
Without loss of generality, assume j < i.
First suppose vr 6= vi . So vr+1 is an H ′ -neighbor. Let e and e′ be distinct edges of Ex such that
vj+1 ∈ e and vr+1 ∈ e′ . Then the cycle

(v1 , e1 , v2 , . . . , ej−1 , vj , e∗ , vr , er−1 , vr−1 , . . . , ej+1 , vj+1 , e, x, e′ , vr+1 , er+1 , vr+2 , . . . , ek−1 , vk , ek , v1 )

is longer than C, a contradiction.

Otherwise if vr = vi , let e ∈ Ex be any edge containing vj+1 . Then we instead take the cycle

(v1 , e1 , v2 , . . . , ej−1 , vj , e∗ , vi , ei−1 , vi−1 , . . . , ej+1 , vj+1 , e, x, ei , vi+1 , ei+1 , vi+2 , . . . , ek−1 , vk , ek , v1 ).

Therefore all vertices of e∗ are H ′ -neighbors except for vj , and the argument applies to every
t


edge containing vj . Thus dH (vj ) ≤ r−1 , a contradiction. This proves Theorem 4 for r < t.

4 Setup for r ≥ t
In order to prove Theorem 4 for r ≥ t, among non-hamiltonian r-uniform hypergraphs with δ(H) ≥
max{t + 1, r}, we choose H with the maximum number of edges. In particular, adding any edge to
H creates a hamiltonian cycle. Therefore H contains a hamiltonian path.
For each such path P = (v1 , e1 , v2 , . . . , en−1 , vn ), an end v1 (respectively, vn ) is suitable if it
belongs to at least two edges in E(H) − E(P ), or belongs to some edge f ∈ E(H) − E(P ) that
contains v2 (respectively, vn−1 ). This notion will be quite useful in our proofs.
For s ∈ {1, n} and a suitable end vs of P above, let Es (P ) denote the set of the edges of H
containing vs that are not in E(P ). If none of the edges in E1 (P ) (respectively, in En (P )) contains
v2 (respectively, vn−1 ), then we define the set A1 = A1 (P ) = E1 (P ) (respectively, An = An (P ) =
En (P )); otherwise we let A1 = A1 (P ) = E1 (P ) ∪ e1 (respectively, An = An (P ) = En (P ) ∪ en−1 ).
By the definition of a suitable end, |As | ≥ 2 in all cases.
Denote S1 = S1 (P ) = {vi ∈ V (P ) : v1 ∈ ei } and Sn = Sn (P ) = {vi+1 ∈ V (P ) : vn ∈ ei }. For a
set of vertices S, set S − = {vi : vi+1 ∈ S} and S + = {vi : vi−1 ∈ S}.
Our plan is to prove in the four subsequent sections the series of four lemmas below and use
the last of them to prove the nonexistence of the above counterexamples H to Theorem 4 in the
final section.
Lemma 4.1. Let P = (v1 , e1 , v2 , . . . , en−1 , vn ) be any hamiltonian path in H. Then we may find
another hamiltonian path P ′ = (w1 , f1 , w2 , . . . , fn−1 , wn ) such that vn = wn , w1 belongs to at least
one edge outside of P ′ , and if vn is suitable or belongs to an edge not in P , then wn is suitable or
belongs to an edge not in P ′ , respectively.

11
Lemma 4.2. H has a hamiltonian path P = (v1 , e1 , v2 , . . . , en−1 , vn ) with at least one suitable end.

Lemma 4.3. H has a hamiltonian path P = (v1 , e1 , v2 , . . . , en−1 , vn ) with both ends suitable.

Lemma 4.4. If H has a hamiltonian path P = (v1 , e1 , v2 , . . . , en−1 , vn ) with both ends suitable,
then either H is hamiltonian or r = t and H has a cycle C of length n − 1 such that the vertex
v ∈ V (H) − V (C) belongs to at most 2 edges in C. Moreover, these edges are either consecutive on
C or separated by 1 edge.

But first we make two useful observations.

Remark 1. If r = t, then H has at least n + 3 edges, and if r ≥ t + 1, then H has at least n
edges.
Indeed, for r = t the sum of degrees of H is at least (t + 1)n, so
   
(t + 1)n (n + 1)n
|E(H)| ≥ = = n + 3.
t n−1

Similary, for r ≥ t + 1, |E(H)| ≥ nr/r = n.

Claim 4.1. Let P = (v1 , e1 , v2 , . . . , en−1 , vn ) be a path. For any f1 ∈ A1 (P ) and/or fn ∈ An (P ),

(i). (f1 − {v1 })− ∩ fn = ∅ and f1 ∩ (fn − {vn })+ = ∅, and

(ii). S1 (P ) ∩ fn = ∅ and Sn (P ) ∩ f1 = ∅.

Proof. For (i), let vi ∈ (f1 − {v1 })− ∩ fn . Suppose first that f1 ∈ E1 (P ) and fn ∈ En (P ), i.e.,
f1 , fn ∈
/ E(P ). Then there exists a hamiltonian cycle

C = (v1 , e1 , v2 , . . . , ei−1 , vi , fn , vn , en−1 , vn−1 , . . . , ei+1 , vi+1 , f1 , v1 ).

Recall that the only edge outside P that can possibly be contained in A1 is e1 . Similarly, the only
edge outside of P that can be contained in An is en−1 . If f1 = e1 and/or fn = en−1 , there exists
some edge f1′ ∈ A1 and/or fn′ ∈ An such that f1′ 6= e1 , v2 ∈ f1′ and/or fn′ 6= en−1 , vn ∈ fn′ . Then we
simply replace e1 and/or en−1 with f1′ and/or fn′ respectively in C to get a hamiltonian cycle.
The proof for f1 ∩ (fn − {vn })+ is similar. This proves (i).
We now prove (ii). If vj ∈ S1 ∩ fn and fn 6= en−1 , then we get the hamiltonian cycle

(v1 , e1 , v2 , . . . , ej−1 , vj , fn , vn , en−1 , vn−1 , . . . , ej+1 , vj+1 , ej , v1 ).

If fn = en−1 , then An contains an edge fn′ such that {vn , vn−1 } ⊆ fn′ . Then we may replace
en−1 in the previous cycle with fn′ to get a hamiltonian cycle. The proof for Sn ∩ f1 is similar.

5 Proof of Lemma 4.1

Let P = (v1 , e1 , v2 , . . . , en−1 , vn ) be a hamiltonian path in H and suppose v1 is not contained in
any edges of E(H) − E(P ). Then

|S1 | ≥ dH (v1 ) ≥ t + 1. (9)

12
 S
Let F := f ∈E(H)−E(P ) f . Note that v1 ∈
/ F . By Remark 1 we have that

|F | ≥ r, and if r = t, then |F | ≥ t + 1 with equality if and only if |f ∩ f ′ | = t − 1 for all f, f ′ ∈

/ E(P ).
(10)
If there is vi ∈ F ∩ S1 , then the end vi of the hamiltonian path

P0 = (vi , ei−1 , vi−1 , . . . , e1 , v1 , ei , vi+1 , ei+1 , vi+2 , . . . , en−1 , vn )

belongs to at least one edge of E(H) − E(P0 ). Since E(P0 ) = E(P ), we are done. Thus we may
assume that S1 ∩ F = ∅. It follows that |S1 | + |F | ≤ n ∈ {2t + 1, 2t + 2}. By (10) and (9),
|F | + |S1 | ≥ (t + 1) + (t + 1). If |F | + |S1 | > 2(t + 1), then we obtain a contradiction. Therefore
we may assume that n = 2t + 2 and |F | = |S1 | = t + 1. In particular, r ∈ {t, t + 1}. The next fact
follows.
Each vertex in P is in exactly one of S1 or F . (11)
Let s be the largest index such that vs ∈ S1 . Fix any vi ∈ S1 − {v1 , vs } and let P0 be as above.
For simplicity, let us rename the vertices and edges of P0 = (u1 , g1 , u2 , . . . , gn−1 , un ) where u1 = vi
and un = vn . If u1 belongs to an edge of E(H) − E(P0 )(= E(H) − E(P )), then we are done.
Otherwise we observe that since E(P0 ) = E(P ), dP0 (u1 ) = dH (vi ) ≥ t + 1. Moreover, (11) applies
to S1 (P0 ) and F . Therefore S1 (P0 ) = V (H) − F = S1 (P ). In particular, since vs ∈ S1 (P ) and
P0 and P are identical starting with the vertex vi+1 (ui+1 in P0 ), by (11), we have that u1 = vi
belongs to the edge gs = es . Since vi was an arbitrary vertex in S1 − {vs }, the same holds for all t
other vertices in S1 , i.e., the edge es contains all of S1 as well as vs+1 . Then |es | ≥ t + 1 + 1 > r, a
contradiction.

6 Proof of Lemma 4.2

Suppose the lemma fails. Let P = (v1 , e1 , v2 , . . . , en−1 , vn ) be a hamiltonian path in H such that
there exist f1 , fn ∈ E(H)−E(P ) with v1 ∈ f1 , vn ∈ fn . We may assume f1 and fn exist by applying
the previous lemma twice (once for each endpoint of P ). If neither v1 nor vn are suitable, then
v2 ∈
/ f1 , vn−1 ∈ / fn , and v1 and vn each are contained in at least max{t + 1, r} − 1 edges in P . By
Claim 4.1, (f1 − {v1 })− ∩ fn = ∅ and S1 ∩ fn = ∅.
Case 1: r ≥ t + 1. We have |fn ∪ (f1 − {v1 })− | ≥ 2r − 1. Since vn ∈ / f1 and vn−1 ∈ / fn ,
vn−1 ∈ −
/ fn ∪ (f1 − {v1 }) . Similarly, since v1 ∈ / fn and v2 ∈ / f 1 , v1 ∈ −
/ fn ∪ (f1 − {v1 }) . Thus
|fn ∪ (f1 − {v1 })− | ≤ n − 2 < 2r − 1, a contradiction.
Case 2: r = t. By Remark 1, there are distinct edges f3 and f4 in E(H) − E(P ) − f1 − fn . Let
F = f3 ∪ f4 . Since f4 6= f3 ,

|F | ≥ t + 1, and if |F | = t + 1, then |f3 ∩ f4 | = t − 1. (12)

If there is vi ∈ F ∩ (f1 − {v1 })− , then the end vi of the hamiltonian path

P1 = (vi , ei−1 , vi−1 , . . . , e1 , v1 , f1 , vi+1 , ei+1 , vi+2 , . . . , en−1 , vn )

belongs to at least one of f3 and f4 and to ei that are not in E(P1 ), and hence is a suitable end of
P1 , as claimed.

13
 Thus F ∩ (f1 − {v1 })− = ∅. Similarly, F ∩ (fn − {vn })+ = ∅. Also by definition, vn ∈ / F ∪ (f1 −
−
{v1 }) . Since v2 ∈/ f1 , also v1 ∈ − −
/ F ∪ (f1 − {v1 }) . It follows that |(f1 − {v1 }) | + |F | ≤ n − 2 = 2t.
Since |(f1 − {v1 })− | = t − 1, we must have that the second part of (12) holds. In particular,

|F | = t + 1 and every vertex in {v2 , . . . , vn−1 } must belong to exactly one of F or (f1 − {v1 })− .
(13)
By (13), we conclude that (fn − {vn })+ = (f1 − {v1 })− . That is, fn = {vi : vi+2 ∈ f1 } ∪ {vn }.
Denote B = (fn − {vn })+ = (f1 − {v1 })− .
By Claim 4.1, B ∩ fn = ∅. By (13), fn − {vn } ⊂ F . Moreover, because (fn − {vn })+ ∩ F = ∅,
fn does not contain a pair of consecutive vertices in P . Symmetrically, we get f1 − {v1 } ⊆ F and
f1 does not contain a pair of consecutive vertices. Since v1 ∈ / fn and B = (fn − {vn })+ , v1 , v2 ∈
/ B.
Symmetrically, vn , vn−1 ∈
/ B. Since |B| = t − 1,

|B − ∪ B + | ≥ t, with equality if and only if B = {vi , vi+2 , vi+4 , . . . , vi+2(t−2) } for some i. (14)

Case 2.1: There exists 2 ≤ i ≤ n − 1 such that v1 ∈ ei and vi ∈ (f1 ∪ fn ) − {v1 , vn }. Then the
hamiltonian path

P2 = (vi , ei−1 , vi−1 , . . . , e1 , v1 , ei , vi+1 , ei+1 , vi+2 , . . . , en−1 , vn )

begins with the suitable vertex vi , since vi belongs to f1 ∪ fn and (f1 ∪ fn ) − {v1 , vn } ⊆ F , it also
belongs to at least one of {f3 , f4 }.
Case 2.2: There exists 2 ≤ i ≤ n − 1 such that vn ∈ ei−1 and vi ∈ (f1 ∪ fn ) − {v1 , vn }. This is
symmetric to Case 2.1.
Case 2.3: For all 2 ≤ i ≤ n−1 such that v1 ∈ ei , vi ∈ / (f1 ∪fn)−{v1 , vn } and for all 2 ≤ i ≤ n−1
/ (f1 ∪ fn ) − {v1 , vn }. Recall that (f1 ∪ fn ) − {v1 , vn } = B + ∪ B − . Let
such that vn ∈ ei−1 , vi ∈
I1 = S1 − {v1 } and I2 = Sn − {vn }. We know that |I1 | = |I2 | ≥ t − 1. By the case,

I1 , I2 ⊆ {2, . . . , n − 1} − (B − ∪ B + ). (15)

By (14), the RHS of (15) has size at most 2t − t = t. So there is j ∈ I1 ∩ I2 . Note that
B ∩ (B + ∪ B − ) = ∅ since B + ∪ B − ⊆ F and by Claim 4.1. Thus by the structure of B, either
vj ∈ B or j ∈ {2, n − 1}.
Case 2.3.1: vj ∈ B. Then vj−1 ∈ fn and vj+1 ∈ f1 . Consider the hamiltonian path

P3 = (vj+1 , ej+1 , vj+2 , . . . , en−1 , vn , fn , vj−1 , ej−2 , vj−2 , . . . , e1 , v1 , ej , vj ).

The end vj+1 of this path is suitable, since vj+1 belongs to at least one of f3 and f4 (because
B + ⊆ F ) and in ej that was not used in this path.
Case 2.3.2: j = n − 1 (the case j = 2 is symmetric to this). Then the only possible B is
{v3 , v5 , . . . , vn−3 }, so f1 = {v1 , v4 , v6 , . . . , vn−2 } and fn = {v2 , v4 , . . . , vn−4 , vn }. Since t − 1 ≥ 2,
there is i ∈ I1 − {j}. Consider very similar to P3 hamiltonian path

P4 = (vi+1 , ei+1 , vi+2 , . . . , en−1 , vn , fn , vi−1 , ei−2 , vi−2 , . . . , e1 , v1 , ei , vi ).

By the case, vi+1 ∈ f1 and hence it also belongs to some of f3 and f4 which are not used in P4 .
Thus, vi+1 is suitable for P4 , as claimed.

14
7 Proof of Lemma 4.3
Suppose the lemma fails. By Lemmas 4.1 and 4.2, H has a hamiltonian path P = (v1 , e1 , v2 , . . . , en−1 , vn )
with suitable S
end vn .
Let Fn = f ∈An f . Since |An | ≥ 2,

|Fn | ≥ r + 1, and if |Fn | = t + 1, then r = t and |f ∩ f ′ | = t − 1 for all distinct f, f ′ ∈ An . (16)

Recall S1 = {vi ∈ V (P ) : v1 ∈ ei }. Similarly to (9), |S1 | ≥ dH (v1 ) − 1 ≥ t. By Claim 4.1(ii),

Fn ∩ S1 = ∅, and hence |Fn | + |S1 | ≤ n ≤ 2t + 2. If |Fn | + |S1 | ≤ 2t, then we get a contradiction
since |Fn | ≥ t + 1 and |S1 | ≥ t.
From now on, we may assume by Lemma 4.1 that v1 belongs to exactly one edge f1 ∈ E(H) −
E(P ).
Case 1: |Fn | + |S1 | = n. In this case, each vi is in exactly one of Fn and S1 .
Let j = j(P ) be the smallest index larger than 1 such that vj ∈ f1 . Since v1 is not suitable,
j ≥ 3. Among such paths P with one end suitable, choose one with the smallest value of j(P ).
Then vj−1 is in exactly one of Fn and S1 . By Claim 4.1, there cannot be an edge f ∈ An such that
vj−1 ∈ f , i.e., vj−1 ∈
/ Fn . So, we may assume vj−1 ∈ S1 . This means v1 ∈ ej−1 . Recall vj ∈ f1 . If
f1 6= e1 , then consider path

P1 = (vj−1 , ej−2 , vj−2 , . . . , e1 , v1 , f1 , vj , ej , vj+1 , . . . , en−1 , vn ).

Observe that vn is still a suitable end of P1 and that ej−1 ∈ E(H) − E(P1 ). Moreover, the next
vertex of ej−1 on P1 is v1 which is closer than j−1 to vj−1 . This means j(P1 ) < j(P ), a contradiction
to the choice of P . If f1 = e1 , then we may replace e1 in P1 with the edge f ′ ∈ A1 − {f1 } and
obtain the same conclusion.
Case 2: |Fn | + |S1 | = n − 1. In this case, |S1 | = t, |Fn | = t + 1, and all but one vi are contained
in exactly one of S1 or Fn . By (16), r = t. Call a vertex vi bad if it is not contained in S1 ∪ Fn .
Let j = j(P ) be as defined in the previous case. If vj−1 is not bad then we are finished as before.
Moreover, if for any vi ∈ f1 − {v1 }, vi−1 ∈ Fn , then we may find a hamiltonian cycle as in the
previous case. So we may assume for all vi ∈ f1 − {v1 , vj }, vi−1 ∈ S1 . These vertices make up t − 2
of the t vertices in S1 .
Let s be the largest index such that vs ∈ S1 . Fix any vi ∈ S1 − {v1 , vs }. So v1 ∈ ei . Consider
the path
Pi = (vi , ei−1 , vi−1 , . . . , e1 , v1 , ei , vi+1 , ei+1 , vi+2 , . . . , en−1 , vn ).
For simplicity, rename Pi = (u1 , g1 , u2 , . . . , gn−1 , un ). Observe that u1 = vi and for all j ≥ i + 1,
uj = vj , gj = ej . If u1 ∈ Fn , then we may find a hamiltonian cycle by extending Pi (possibly
switching gn−1 with another edge in E(H) − E(Pi ). Therefore we again have that S1 (Pi ) ∩ Fn = ∅.
We will show that es contains at least t + 1 > r vertices in S1 ∪ {vs+1 }, giving us a contradiction.
If |S1 (Pi )| ≥ t + 1, then either |S1 | + |Fn | ≥ n + 1 and we can find a hamiltonian cycle
or |S1 (Pi )| = t + 1 and we have that S1 (Pi ) and Fn form a partition of V (P ). In particular,
S1 (P ) ⊂ S1 (Pi ), and so vs ∈ S1 (Pi ). This implies that vi ∈ gs = es .
Now suppose |S1 (Pi )| = t. We have that S1 (Pi ), S1 (P ) ⊂ V (H) − Fn . Since dH (u1 ) ≥ t + 1,
u1 belongs to some edge f ′ ∈ E(H) − E(Pi ). We may assume that uj(Pi )−1 is bad with respect
to S1 (Pi ), otherwise we get a contradiction as in Case 1. Similarly, we may assume that for every
uk ∈ f ′ with k > j(Pi ), uk−1 ∈ S1 (Pi ). Then S1 (Pi ) = (S1 (P ) − {uj(Pi )−1 }) ∪ {vj(P )−1 }. This

15
implies that uj(Pi )−1 6= us since us is the last appearing vertex in S1 (P ) and therefore S1 (Pi ). We
conclude that us ∈ S1 (Pi ) and hence u1 = vi ∈ gs = es .
Altogether, we have that es contains every vertex in S1 (P ) as well as the vertex vs+1 . Therefore
|es | ≥ t + 1 > r, a contradiction.

8 Proof of Lemma 4.4

Suppose towards contradiction that H contains a hamiltonian path with 2 suitable ends. We
begin by recalling some definitions. If P = (v1 , e1 , v2 , . . . , en−1 , vn ) is such a path, for j ∈ {1, n},
define Ej = Ej (P ) be the set of edges in E(H) − E(P ) containing vj . If there exists f ∈ E1
(resp. En ) such that v2 ∈ f (resp. vn−1 ∈ f ), then we set A1 = A1 (P ) = E1 ∪ {e1 } (resp.
An = An (P ) = En ∪ {en−1 }). Otherwise, we set E1 = A1 (resp. En = An ).
Since v1 and vn are suitable,

|Fj | ≥ r + 1, and if |Fj | = r + 1, then |f ∩ f ′ | = r − 1 for all distinct f, f ′ ∈ Aj . (17)

Let h = h(P ) be the minimum index such that vh ∈ Fn and h′ = h′ (P ) be the maximum index
such that vh′ ∈ F1 .
A crossing pair in P is a pair of vertices (vi , vj ) such that i < j, vi ∈ Fn , vj ∈ F1 , and for
all i < k < j, vk ∈ / F1 ∪ Fn . For a crossing pair (vi , vj ), the gap of (vi , vj ) is the set of vertices
{vi+1 , . . . , vj−1 }.
We will show that P contains at least one crossing pair with a gap of size 1 which will be used
to find a cycle of length n − 1.
Let P and P ′ be hamiltonian paths with two suitable ends. We say that P is better than P ′ if
(i) P has a crossing pair and P ′ does not, or

(ii) P and P ′ both have a crossing pair, but the smallest gap of a crossing pair in P has fewer
vertices than that of P ′ , or

(iii) P and P ′ both do not have a crossing pair, but h(P ) − h′ (P ) < h(P ′ ) − h′ (P ′ ).
Among all hamiltonian paths with two suitable ends, choose P to be a best one.
As before, by Claim 4.1(i), we have

Fn ∩ (F1 − {v1 })− = ∅. (18)

If r > t, then by (17), |(F1 − {v1 })− | + |Fn | ≥ r + r + 1 ≥ (t + 1) + (t + 2) > n, contradicting

(18). Therefore r = t and either n = 2t + 1 and |F1 | = |Fn | = t + 1, or n = 2t + 2 and
{|F1 |, |Fn |} ⊆ {t + 1, t + 2}.
Call a vertex typical (with respect to P ) if it belongs to (F1 − {v1 })− ∪ Fn , and atypical if it
is not. If n = 2t + 1, or n = 2t + 2 and {|F1 |, |Fn |} = {t + 1, t + 2}, then every vertex is typical.
Otherwise, there is exactly one atypical vertex, |F1 | = |Fn | = t + 1, and n = 2t + 2.
By the definitions of vh , vh′ , {v1 , . . . , vh−1 } ∩ Fn = ∅, and {vh′ +1 , . . . , vn } ∩ F1 = ∅. Moreover,
since H contains at most one atypical vertex, {v1 , . . . , vh } ⊆ F1 if and only if the atypical vertex is
not contained in {v1 , . . . , vh−1 }, and the same holds for Fn .
Claim 8.1. For k ≤ h, if vk ∈ F1 , then for each vi ∈ ek−1 , vi−1 ∈
/ Fn . In particular, if vi−1 is
typical, then vi ∈ F1 .

16
Proof. Suppose vk ∈ f1 ∈ A1 , vi ∈ ek−1 and vi−1 ∈ fn ∈ An . Without loss of generality, we may
assume f1 6= e1 and fn 6= en−1 . Then we have the hamiltonian cycle

(vn , fn , vi−1 , ei−2 , vi−2 , . . . , ek , vk , f1 , v1 , e1 , v2 , . . . , ek−2 , vk−1 , ek−1 , vi , ei , vi+1 , . . . , en−1 , vn ),

a contradiction.

Claim 8.2. Suppose there exists j such that {vj , vj+1 } ⊂ F1 . If (a) |A1 − {e1 }| ≥ 2, (b) vj−1 ∈ ej ,
or (c) ej−1 ⊂ F1 and v1 ∈ ej , then there exists a hamiltonian path Pj starting at vj and ending at
vn such that

(1) both ends are suitable,

(2) ej ∈
/ E(Pj ),

(3) E(Pj ) ⊆ E(P ) ∪ A1 , and

(4) Pj and P are identical starting at vj+1 .

Proof. Let f ∈ A1 be an edge containing vj+1 . Without loss of generality, f 6= e1 . Consider the
hamiltonian path

P ′ = (vj , ej−1 , vj−1 , . . . , e1 , v1 , f, vj+1 , ej+1 , vj+2 , . . . , en−1 , vn ).

Note that ej ∈/ E(P ′ ) and vj−1 is the second vertex of P ′ . If |A1 (P ) − {e1 }| ≥ 2 or if vj−1 ∈ ej ,
then P ′ has two suitable ends and satisfies (1)-(4).
Otherwise, |A1 − {e1 }| = 1, v1 ∈ ej and ej−1 ⊆ F1 . If vj+1 ∈ ej−1 , let f1 = ej−1 and let f2 ∈ A1
be an edge containing vj . Without loss of generality, f2 6= e1 . Otherwise, if vj+1 ∈ / ej−1 , then
ej−1 = F1 − {vj+1 }, and every edge of A1 contains vj+1 . Let f1 ∈ A1 be an edge containing vj−1
(and vj+1 ) and let f2 = ej−1 . Again, we may assume f1 6= e1 . We obtain the hamiltonian path

P ′′ = (vj , f2 , v1 , e1 , v2 , . . . , ei−2 , vj−1 , f1 , vj+1 , ej+1 , vj+2 , . . . , en−1 , vn ).

Similarly, ej ∈/ E(P ′′ ) and v1 is the second vertex of P ′′ . Therefore P ′′ has both ends suitable
and satisfies (1)-(4).

Claim 8.3. H contains at least one crossing pair.

Proof. It suffices to prove that h < h′ . Suppose not. Since |F1 |, |Fn | ≥ t + 1 and h′ ≥ t + 1, we have
h ≤ n − t ≤ 2t + 2 − t = t + 2. Therefore either h = h′ or h = h′ + 1, |F1 | = |Fn | = t + 1, and there is
exactly one atypical vertex.
Since {|F1 |, |Fn |} ⊆ {t + 1, t + 2}, we may assume F1 = {v1 , . . . , vt+1 }.
t
There are at most t−1 = t edges within F1 that contain v1 , so v1 must belong to at least
dH (v1 ) − t ≥ 1 edges which are not contained entirely in F1 . By the definition of F1 , these edges
appear in P . Let ej be such an edge with j minimized, and let vi ∈ ej ∩ {vt+2 , . . . , vn }. By
Claim 4.1, we have that ej ∈ / Fn . Therefore ej ∈ {e1 , . . . , et } ∪ {es } where vs is the unique atypical
vertex (if it exists).
Case 1: j ≤ t. If j satisfies (a), (b), or (c) of Claim 8.2, then let Pj be a path guaranteed by
/ E(Pj ), we have h′ (Pj ) ≥ i ≥ t + 2 > h′ (P ) and h(Pj ) = h(P ).
the claim starting with vj . Since ej ∈
Therefore either h′ (Pj ) > h(Pj ) or |h′ (Pj ) − h(Pj )| < |h′ (P ) − h(P )|. In either case, Pj is better
than P , a contradiction.

17
 Therefore we may assume |A1 − {e1 }| = 1, vj−1 ∈ / ej , and ej−1 * F1 . Note that this third
condition implies v1 ∈ / ej−1 , by the choice of j. In particular, since dH (v1 ) ≥ t+1, and |A1 −E(P )| =
1, ej−1 is the unique edge in {e1 , . . . , et } that does not contain v1 . So j−1 ≥ 2. Let k ∈ {t+2, . . . , n}
be such that vk ∈ ej−1 . By Claim 8.1, vk−1 ∈ / Fn , so vk−1 must be the unique atypical vertex vs .
This shows that the only vertex in ej−1 outside of {v1 , . . . , vt+1 } is vs+1 . Since v1 ∈ / ej−1 , we have
that ej−1 omits exactly one vertex in {v2 , . . . , vt+1 }.
If vj+1 ∈ ej−1 , then we consider the hamiltonian path

P ′ = (vj , f1 , v1 , e1 , v2 , . . . , ej−2 , vj−1 , ej−1 , vj+1 , ej+1 , vj+2 , . . . , en−1 , vn ),

where f1 6= e1 is an edge of A1 containing vj . We have h′ (P ′ ) ≥ i, and as before this contradicts

the choice of P .
Thus, if j − 1 ≥ 3, then vj−2 ∈ ej−1 . Let f1 ∈ A1 contain vj where without loss of generality,
f1 6= e1 . We obtain the hamiltonian path

P ′′ = (vj−1 , ej−2 , vj−2 , . . . , e1 , v1 , f1 , vj , ej , vj+1 , . . . , en−1 , vn ).

which does not contain the edge ej−1 . Because {vj−1 , vj−2 } ⊂ ej−1 , both ends of P ′′ are suitable,
and h′ (P ′′ ) ≥ s + 1. Similar to before, P ′′ contradicts the choice of P .
Finally, if j − 1 = 2, recall that v1 ∈ e3 and {v1 , v2 } ⊆ F1 . We may assume that v3 ∈ f1 ∈ A1
and f1 6= e1 . Then we have path

P ′′′ = (v3 , f1 , v2 , e1 , v1 , e3 , v4 , e4 , v5 , . . . , en−1 , vn ).

Again, e2 ∈ / E(P ′′′ ) and the first two vertices of P ′′′ belong to e2 , so both ends of P ′′′ are suitable
′′′
and P is better than P .
Case 2: j ≥ t + 1. Then vj = vs is the unique atypical vertex. In particular, ej is the unique
edge of P which contains v1 and intersects {vt+2 , . . . , vn }. If |A1 −{e1 }| ≥ 2, then for any k ≤ t, fix a
path Pk given by Claim 8.2. Either h′ (Pk ) > h′ (P ), contradicting the choice of P , or h′ (Pk ) = h′ (P )
and F1 (Pk ) = F1 (P ) = {v1 , . . . , vt+1 }. Then repeating the arguments for Pk instead, we have that
its first vertex vk must also be contained in ej . We conclude that {v1 , . . . , vt } ∪ {vj , vj+1 } ⊆ ej .
But then |ej | ≥ t + 2, a contradiction.
So suppose |A1 − {e1 }| = 1. Then v1 belongs to at least dH (v1 ) − 1 ≥ t edges in P , and at least
t − 1 of them must be in {e1 , . . . , et }. If it exists, let em be the unique edge of {e1 , . . . , et } that does
not contain v1 . Then for any 2 ≤ k ≤ t such that k 6= m, let Pk be a path guaranteed by Claim 8.2
starting with vk . Then as before, we have {v1 , . . . , vt } ∪ {vj , vj+1 } − {vm } ⊆ ej . So |ej | ≥ t + 1.

Claim 8.4. Suppose (i, i + 2) is a crossing pair with a gap of size 1. If the vertex vi+1 belongs to
an edge ej with j ≤ i − 1, then vj+1 ∈
/ F1 ∪ Fn . Similarly, if vi+1 belongs to ej with j ≥ i + 2, then
vj ∈
/ F1 ∪ Fn .

Proof. We will prove the claim for the case j ≤ i − 1. The other case j ≥ i + 2 is symmetric by
considering P in reverse order. Suppose vi+1 ∈ ej . Since vi+1−1 ∈ Fn , by Claim 8.1, we cannot
have vj+1 ∈ F1 .
If vj+1 ∈ Fn , let f ∈ A1 contain vi+2 and let f ′ ∈ An contain vj+1 . Without loss of generality,
f 6= e1 , f ′ 6= en−1 . We instead get the hamiltonian cycle

18
 (v1 , f, vi+2 , ei+2 , vi+3 , . . . , en−1 , vn , f ′ , vj+1 , ej+1 , vj+2 , . . . , ei , vi+1 , ej , vj , ej−1 , vj−1 , . . . , e1 , v1 ).

Claim 8.5. Suppose vi−1 and vi are typical. If vi ∈ / F1 ∪ Fn , then vi belongs to the gap of the
crossing pair (vi−1 , vi+1 ). In particular, this pair has a gap of size 1.

Proof. Since vi ∈ / Fn , vi ∈ (F1 − {v1 })− , and hence vi+1 ∈ F1 . Because vi ∈ / (F1 − {v1 })− ,
/ F1 , vi−1 ∈
i.e., vi−1 ∈ Fn . This proves that (vi−1 , vi+1 ) is a crossing pair whose gap is {vi }.

Claim 8.6. Suppose |F1 | = t + 1 and {vh , vh−1 } ⊂ F1 . If (vi , vj ) is a crossing pair and vj is typical,
then vj ∈ Fn .

Proof. Suppose vj ∈ / Fn . Then vj+1 ∈ F1 . Because (vi , vj ) is a crossing pair, vj ∈ F1 . Since all but
at most one vertex in H is typical, by Claim 8.1, all but at most one vi ∈ eh−1 belong to F1 .
We will show first that there exist three distinct edges f1 , f2 , f3 ∈ F1 ∪ {eh−1 , e1 } such that (a)
vh−1 ∈ f1 and (b) v1 ∈ f2 , (c) {v1 , v2 } ⊆ f3 , and (d) f1 and f2 each contain a different vertex in
{vj , vj+1 }.
Case 1: Every vertex of H is typical. Observe that eh−1 is a subset of size t in F1 that
contains vh−1 . Therefore it must contain at least one of {vj , vj+1 }. Set f1 = eh−1 and suppose for
a ∈ {j, j + 1}, va ∈ eh−1 . Let {b} = {j, j + 1} − {a}. Since |A1 | ≥ 2 and at most one edge of A1 does
not contain any given vertex, we may choose some f2 ∈ A1 containing vb and f3 ∈ A1 containing
v2 .
Case 2: There exists an atypical vertex vs . If vs ∈
/ eh−1 , then we are done as in the previous case.
Therefore since |F1 ∪vs | = t+2, we have either {vj , vj+1 }∩eh−1 6= ∅ or eh−1 = F1 ∪{vs }−{vj , vj+1 }.
In the first case, we take f1 = eh−1 , let f2 be any edge of A1 containing vb (as defined in the previous
case), and f3 any edge of A1 − {f2 } containing v2 . In the latter case, set f3 = eh−1 . Similar to
before, we can find an edge f ∈ A1 such that vh−1 ∈ f and f ∩ {vj , vj+1 } = 6 ∅. Defining va , vb as
before, at most one edge does not contain vb . If we can find f2 ∈ A1 − {f1 } containing vb , then
set f1 = f and f2 = f ′ . So assume |A1 | = 2, {f ′ } = A1 − {f }, and f is the only edge of A1 that
contains vb . Then f ′ must contain va and vh−1 . Set f1 = f ′ and f2 = f .
Finally we will show that there exists a hamiltonian cycle using f1 , f2 , and f3 . Let f ′ ∈ An
contain vh . Without loss of generality, f ′ 6= en−1 . If vj+1 ∈ f1 and vj ∈ f2 , then we take the cycle

(vn , f ′ , vh , eh , vh+1 , . . . , ej−1 , vj , f2 , v1 , f3 , v2 , e2 , v3 , . . . , eh−2 , vh−1 , f1 , vj+1 , ej+1 , vj+2 , . . . , en−1 , vn ),

and if vj ∈ f1 and vj+1 ∈ f2 , then we take the cycle

(vn , f ′ , vh , eh , vh+1 , . . . , ej−1 , vj , f1 , vh−1 , eh−2 , vh−2 , . . . , e2 , v2 , f3 , v1 , f2 , vj+1 , ej+1 , vj+2 , . . . , en−1 , vn ).

Claim 8.7. P has a crossing pair (i, i + 2) with gap of size 1.

Proof. Suppose not. By Claim 8.3, there exists at least one crossing pair. If there are no atypical
vertices, then by Claim 8.5, any crossing pair has a gap of size 1.

19
 Suppose first that P has at least two crossing pairs (i1 , j1 ), (i2 , j2 ) where j1 ≥ i1 + 3, and
j2 ≥ i2 + 3. Without loss of generality, vs ∈ / {vi1 , vi1 +1 }. Since vi+1 ∈ / F1 ∪ Fn , by Claim 8.5,
(vi1 , vi1 +2 ) is a crossing pair with a gap of size 1.
Therefore we may assume P contains a single crossing pair (vi , vj ) where j ≥ i + 3. The vertices
vi+1 , . . . , vj−2 are contained in neither Fn nor (F1 − {v1 })− , and hence are all atypical. It follows
that there may only be one such vertex, i.e., vi+1 is the unique atypical vertex, j = i + 3, and the
only crossing pair in P has a gap of size 2. By Claim 8.6 applied to both P and P in reverse order,
we obtain that vi ∈ F1 and vj ∈ Fn (here we use the fact that h ≤ i and h′ ≥ j).
Next we will show that {v1 , . . . , vi−1 } ∩ Fn = ∅. If there exists some k < i such that vk ∈ Fn ,
then P would have another crossing pair contained between vk and vi . Similarly, we obtain that
F1 ∩ {vj+1 , . . . , vn } = ∅. Since all vertices of {v1 , . . . , vi } ∪ {vj , . . . , vn } are typical, it follows that
F1 = {v1 , . . . , vi } ∪ {vj } and Fn = {vi } ∪ {vj , . . . , vn }. Moreover, since |F1 | = |Fn | = t + 1, we have
i = t and j = t + 3.
We will show that

ei+1 is the only edge containing v1 that is not a subset of F1 . (19)

Indeed, because dH (v1 ) ≥ t + 1, there exists an edge ek ∈ E(P ) such that v1 ∈ ek and ek 6⊆ F1 .
We cannot have k ≥ i + 3 or k = i, since vk ∈ Fn , contradicting Claim 4.1(ii).
If k = i + 2, then if there exists f ∈ A1 such that {vi+3 , vi−1 } ⊆ f , let f ′ ∈ An contain vi .
Without loss of generality, f 6= e1 and f ′ 6= en−1 . Then we have the hamiltonian cycle

(v1 , e1 , v2 , . . . , ei−2 , vi−1 , f, vi+3 , ei+3 , vi+4 , . . . , en−1 , vn , f ′ , vi , ei , vi+1 , ei+1 , vi+2 , ei+2 , v1 ).

Otherwise, if no edge in A1 contains vi+3 and vi−1 , then by (17), A1 = {e1 , f1 } for some edge
f1 . Without loss of generality, vi−1 ∈ f1 and vi+3 ∈ e1 . Let fn ∈ An contain vi (without loss of
generality, fn 6= en−1 . We get the hamiltonian cycle

(v1 , f1 , vi−1 , ei−2 , vi−2 , . . . , e2 , v2 , e1 , vi+3 , ei+3 , vi+4 , . . . , en−1 , vn , fn , vi , ei , vi+1 , ei+1 , vi+2 , ei+2 , v1 ).

Finally suppose k ≤ i − 1. Moreover, among all edges containing v1 that are not subsets of F1 ,
suppose we choose ek to have the smallest index. Let vm ∈ ek − F1 . If |A1 − {e1 }| ≥ 2, then let
Pk be a path guaranteed by Claim 8.2, say Pk = (u1 , g1 , u2 , . . . , gn−1 , un ) where u1 = vk . Then
ek ∈ A1 (Pk ), and so vm ∈ F1 (Pk ). By Claim 4.1(ii), vm−1 ∈ / Fn (Pk ) = Fn (P ). Therefore m = i + 2.
We have that (i, i + 2) is a crossing pair of Pk with size 1. This violates the choice of P as a best
path.
So suppose |A1 − {e1 }| = 1, i.e., v1 belongs to exactly one edge outside of P . It follows that
v1 belongs to at least dH (v1 ) − 1 ≥ t edges of P , and we have shown that these edges cannot be
in {ei+2 , ei+3 , . . . , en−1 } ∪ {ei }. Since i = t, it follows that v1 belongs to every edge in the set
{e1 , . . . , et−1 } ∪ {ei+1 }. In particular, by the choice of ek , we have ek−1 ⊆ F1 . Therefore we may
apply Claim 8.2 again to get a path Pk and similarly obtain
a contradiction. This proves (19) which
t
in turn implies that v belongs to dH (v) − 1 ≥ t = t−1 edges which are subsets of F1 . That is,
every t-subset of F1 that contains v1 is an edge of H.
We symmetrically obtain similar results for vn . Without loss of generality, assume that |et+1 ∩
F1 | ≤ |et+1 ∩ Fn |. Then since {v1 , vn , vt+1 , vt+2 } ⊂ et+1 , there is ℓ ∈ {2, 3, . . . , t − 2} such that
vℓ ∈/ et+1 . Since dH (vℓ ) ≥ t + 1, there is an edge fℓ containing vℓ and not contained in F1 . For

20
s ∈ {2, 3, . . . , t, t + 3}, let gs = F1 − {vs }. We have shown that gs ∈ E(H). Consider the new
hamiltonian path

P ′ = (vℓ , ft , v1 , gℓ , v2 , gt , v3 , g2 , v4 , g3 , v5 . . . , gℓ−3 , vℓ−1 , gℓ−2 ,

vℓ+1 , gℓ−1 , vℓ+2 , gℓ+1 , vℓ+3 , . . . , gt−2 , vt , et , vt+1 , . . . , en−1 , vn ).

All edges in Fn (P ) are also in Fn (P ′ ), so vn is suitable for P ′ . The set A1 (P ′ ) contains fℓ and
gt+3 that are not in E(P ′ ). So, vℓ is also a suitable end of P ′ . Moreover, gt+3 contains the second
vertex, v1 , of P ′ . Hence gt−1 also is in A1 (P ′ ). It follows that F1 (P ′ ) contains F1 (P ) and in
addition contains a vertex from fℓ that is not in F1 . Thus, |F1 (P ′ )| ≥ t + 2, and still vt ∈ Fn (P ′ )
and vt+3 ∈ F1 (P ′ ). Since |Fn (P ′ )| ≥ t + 1, this implies that every vertex is typical with respect to
P ′ . Then Claims 8.3 and 8.5 imply that P ′ has a crossing pair with a gap of size 1, contradicting
the choice of P .

Call a crossing pair (i, j) tight if j = i + 2 and vj ∈ Fn . By Claim 8.5 if (vi , vj ) is a crossing pair
such that vi and vi+1 are both typical, then j = i+2. Moreover by Claim 8.6, if also {vh−1 , vh } ∈ F1
and vi+2 is typical, then (vi , vj ) is a tight crossing pair. In particular, this shows that if all vertices
of H are typical, then every crossing pair is tight.

Claim 8.8. Suppose H contains a tight crossing pair, and let (i, i + 2) be the earliest occurring
one. If H contains no atypical vertices, then vi+1 is not contained in any edge of E(P ) − {ei , ei+1 }.
Otherwise, if H contains one atypical vertex vs with s ≥ h, then vi+1 belongs to at most two edges
of E(P ) − {ei , ei+1 }, and further those two edges must either be consecutive or separated by only
one edge.

Proof. Suppose ej contains vi+1 . By Claim 8.4, if j ≤ i − 1 then vj+1 ∈

/ F1 ∪ Fn , and if j ≥ i + 2,
then vj ∈
/ F1 ∪ Fn . We will show that

If j ≥ i + 2 then (vj−1 , vj+1 ) is not a crossing pair. (20)

Indeed, suppose not. Recall that vi+2 ∈ Fn since (vi , vi+2 ) is tight. Let f ∈ A1 contain vj+1
and let f ′ ∈ An contain vi+2 . Without loss of generality, f 6= e1 , f ′ 6= en−1 . Then we have the
hamiltonian cycle

(v1 , e1 , v2 , . . . , ei , vi+1 , ej , vj , ej−1 , vj−1 , . . . , ei+2 , vi+2 , f ′ , vn , en−1 , vn−1 , . . . , ej+1 , vj+1 , f, v1 ).

This proves (20).

Case 1: H contains no atypical vertices. Then every crossing is tight, and hence (i, i + 2) is the
first occuring crossing pair in P . Suppose j ≤ i − 1. By Claim 8.4, vj+1 ∈ / F1 ∪ Fn . By Claim 8.5,
vj+1 is in the gap of a crossing pair that occurs before (vi , vi+2 ), a contradiction. Therefore j ≥ i+2.
Similarly, we have that vj ∈ / F1 ∪ Fn , so it must be in the gap of the crossing pair (vj−1 , vj+1 ), but
this is not possible by (20).
Case 2: H contains one atypical vertex, say vs . Without loss of generality, we may assume
that s ≥ h (otherwise we can reorient P ). Therefore {vh−1 , vh } ∈ F1 .
Suppose first that j ≤ i − 1. Then by Claim 8.4, vj+1 ∈ / F1 ∪ Fn . If vj , vj+1 , and vj+2
are all typical, then by Claims 8.5 and 8.6, (vj , vj+2 ) is a tight crossing pair, contradicting the

21
choice of (vi , vi+2 ). Therefore vs must be one of these vertices. But there are at most 3 edges in
E(P ) − {ei , ei+1 } containing vi+1 in this case, namely, es−2 , es−1 , and es . If vi+1 belongs to each of
es−2 , es−1 , and es . By Claim 8.4 applied to each of these edges, we have vs−1 , vs , vs+1 ∈ / F1 ∪ Fn .
In particular, both vs−1 and vs are atypical, a contradiction. Therefore vi+1 belongs to at most 2
of these 3 edges, and the result follows.
Next suppose j ≥ i+2. Then vj ∈ / F1 ∪Fn . If vj−1 , vj are typical, then by Claim 8.5, (vj−1 , vj+1 )
is a crossing pair, which is not possible by (20). Therefore vs ∈ {vj−1 , vj }, but there are at most
two such ej ’s where this can occur, namely es and es+1 .

Finally we are ready to prove the Lemma 4.4.

Without loss of generality, assume |F1 | = t + 1 and if there exists an atypical vertex vs , then
s ≥ h. In particular, this implies that {vh , vh−1 } ⊂ F1 . Suppose first that P has a tight crossing
pair and let (vi , vi+2 ) be the first occurring one. Let f1 ∈ A1 contain vi+2 and let fn ∈ An contain
vi . Without loss of generality, f1 6= e1 and fn 6= en−1 . Consider the cycle

C = (v1 , e1 , v2 , . . . , ei−1 , vi , fn , vn , en−1 , vn−1 , . . . , ei+2 , vi+2 , f1 , v1 ).

Note that C contains neither ei nor ei+1 . By Claim 8.8, vertex vi+1 ∈ / V (C) belongs to at most 2
edges in C, and these edges are either consecutive in C or are separated by 1 edge, hence we are
done.
Therefore by Claim 8.7, there exists a crossing pair (vi , vi+2 ) that is not tight. That is, vi+2 ∈ / Fn .
By Claim 8.6, vi+2 is the unique atypical vertex.
Suppose that P contains another crossing pair (i′ , j ′ ). Then i 6= i′ and i + 2 6= j ′ . Moreover,
′
i 6= i + 2, otherwise (i, i + 2) would be tight. Therefore the atypical vertex vi+2 is not contained in
{vi′ , . . . , vj ′ }. By Claims 8.5 and 8.6, (vi′ , vj ′ ) is tight, a contradiction. This proves that (vi , vi+2 )
is the unique crossing pair of P .
By Claim 8.5, for any k ∈ / {i + 1, i + 2, i + 3}, vk ∈ F1 ∪ Fn (otherwise there would be an-
other crossing pair). Then Claim 8.4 implies that the only edges in P that may contain vi+1 are
ei , ei+1 , ei+2 , and ei+3 . Therefore the same cycle C as above satisfies the Lemma.

9 Proof of Theorem 4 for r ≥ t

We first prove two useful claims.

Claim 9.1. Let n ∈ {2t + 1, 2t + 2}. Fix n − 1 vertices V = {v1 , . . . , vn−1 } and let f, g be distinct
subsets of V of size at least t. Then there exist vertices x ∈ f, y ∈ g such that {x, y} = {vi , vi+1 }
for some i with indices modulo n − 1.

Proof. Suppose not. Then for all vi ∈ f, vi+1 ∈ / g. Let f + = {vi+1 : vi ∈ f }.

If f contains a pair of consecutive vertices, choose such a pair vj , vj+1 ∈ f such that vj−1 ∈ / f.
Then vj , vj+1 ∈ + +
/ g, and hence g ∩ (f ∪ {vj }) = ∅. We have |V − (f ∪ {vj })| ≤ (n − 1) − (t + 1) ≤
(2t + 1) − (t + 1) = t. Therefore we must have |f | = |g| = t. Because |f | = t, there exists a pair of
consecutive vertices that both are not in f , say vi and vi+1 . Let k be the first index that appears
after i + 1 (modulo n − 1) such that vk ∈ f (possibly k = j). Then g ∩ (F + ∪ {vj , vk−1 }) = ∅, but
|V − (f + ∪ {vj , vk−1 })| ≤ 2t + 1 − (t + 2) = t − 1, contradicting that |g| = t.
Otherwise, f does not contain any two consecutive vertices. Without loss of generality, f =
{v1 , v3 , . . . , v2t−1 }. Then it is easy to check that g = f , a contradiction.

22
Claim 9.2. Let n ∈ {2t + 1, 2t + 2} and let H be a t-uniform hypergraph. Suppose there exists
a cycle C = (v1 , e1 , v2 , . . . , en−2 , vn−1 , en−1 , v1 ) of length n − 1 such that v ∈ V (H) − V (C). Let
Ev ⊆ E(H) − E(C) be the set of edges S outside of the cycle incident to v. If S |Ev | ≥ 3, then either H
contains a hamiltonian cycle or | e∈Ev e| = t + 1 and no two vertices in e∈Ev e are consecutive
in C.

Proof. Fix any f1 , f2 , f3 ∈ Ev . If there exists u ∈ f3 such that u ∈ / f1 ∪ f2 , then we set f =

f1 ∪{u}−{v}, and g = f2 ∪{u}−{v}. Then f and g are subsets of size t and so by Claim 9.1, without
loss of generality there exist vi ∈ f and vi+1 ∈ g. Note that at most one of these vertices (u) may
belong to f3 . Thus for some j, k ∈ {1, 2, 3}, j 6= k, we have that vi ∈ fj and vi+1 ∈ fk . We obtain
a hamiltonian cycle by removing the edge from vi to vi+1 and inserting the path (vi , fj , v, fk , vi+1 ).
Therefore we may assume f3 ⊆ f1 ∪ f2 . But this holds for any arbitrary 3 edges.
Suppose there exist edges f1 and f2 such that f1 ∪ f2 contains some pair {vi , vi+1 }. If both f1
and f2 contain at least one of these vertices, then we can construct a hamiltonian cycle as before.
Otherwise, say {vi , vi+1 } ⊂ f1 and {vi , vi+1 } ∩ f2 = ∅. For any f3 , we have f1 ⊆ f3 ∪ f2 . Therefore
{vi , vi+1 } ⊂ f3 and we may again find a hamiltonian cycle.
Therefore, for all f1 , f2 ∈ Ev , t + 1 ≤ |f1 ∪ f2 | ≤ ⌊|C|/2⌋ + 1 = t + 1. Moreover, each f3 6= f1 , f2
is contained in these t + 1 vertices.

Proof of Theorem 4. By Lemma 4.4, there exists a cycle C of length n−1 such that v ∈ V (H)−V (C)
belongs to at most 2 edges of C, and moreover the edges in C that contain v are separated by at
most one edge. Among all such cycles C and vertices v outside of the cycle, choose C and v such
that v belongs to as few edges of C as possible.
Say C = (u1 , f1 , u2 S , . . . , fn−2 , un−1 , fn−1 , u1 ). Let Ev be the set of edges outside of C that
contain v, and let X = e∈Ev e. By Claim 8.8, |Ev | ≥ dH (v) − dC (v) ≥ t + 1 − dC (v).
Case 1. |Ev | ≥ 3. By Claim 9.2 since H is not hamiltonian, |X| = t + 1, and hence for all
f, f ′ ∈ Ev , |f ∩ f ′ | = t − 1. Moreover, because no two vertices in X are consecutive in C, we may
assume without loss of generality X = {u1 , u3 , u5 , . . . , un−3 }.
Since dH (v) ≥ t + 1, v belongs to at least one edge fj of C. If uj+1 ∈ X, then let f ∈ Ev be
any edge containing uj+1 . We obtain a hamiltonian cycle by replacing the edge fj between uj and
uj+1 with the path (uj , fj , v, f, uj+1 ).
Similarly, by reorienting C, we cannot have uj ∈ X. Therefore the only edge of C that contains
v is the edge fn−2 .
Let uk ∈ V (C) be an even-indexed vertex with k ≤ n − 4 (i.e., {uk−1 , uk+1 } ⊆ X). We will show
also that uk ∈ / fj for any j 6= n − 2. Indeed, if uj+1 ∈ X, then let f, f ′ ∈ Ev be edges containing
uk+1 and uj+1 respectively. Then we take the cycle

(uk+1 , fk+1 , uk+2 , . . . , fj−1 , uj , fj , uk , fk−1 , uk−1 , . . . , fj+1 , uj+1 , f ′ , v, f, uk+1 ).

The argument for if vj ∈ X is symmetric. It follows that the only edge in C ′ that may contain uk
is fn−2 .
Observe that we may construct a new cycle by swapping uj with v and fj−1 , fj with some edges
f, f ′ ∈ Ev that contain uj−1 , uj+1 respectively. Note that uj ∈
/ f, f ′ since uj ∈
/ Vv . Therefore, by the
choice of C and v, uj must be contained in fn−2 . This holds for all such uj , so {u2 , u4 , . . . , un−4 } ∪
{un−2 , un−1 } ∪ {v} ⊆ fn−2 . Therefore |fn−2 | ≥ t + 2, contradicting that H is t-uniform.

23
 Case 2. |Ev | ≤ 2. Since |Ev | ≥ dH (v) − dC (v) ≥ t + 1 − dC (v), we must have dC (v) = 2,
t = 3, and therefore |Ev | = 2. Let fi , fj ∈ V (C) contain v. Recall that by Lemma 4.4, fi and
fj are separated by at most one edge in C. Therefore the set Y = {ui , ui+1 , uj , uj+1 } consists of
consecutive vertices (possibly ui+1 = uj ). As in the previous case, if Y ∩ X 6= ∅, then we can find
a hamiltonian cycle in H.
Set V ′ = {u1 , . . . , ui−1 , u∗ , uj+2 , . . . , un−1 }, i.e., V ′ is obtained by identifying the vertices of
Y . If Ev = {f, g}, then let f ′ = f − {v}, and g ′ = g − {v}. Note that u∗ ∈ / f ′ , g′ . We have
′ ′ ′ ′ ′
|f | = |g | = t − 1 and |V | ≤ 2t + 2 − 3 + 1 = 2t. So |f | ≥ ⌊(|V | − 1)/2⌋, and we may apply
Claim 9.2 to find some uk ∈ f ′ and uk+1 ∈ g ′ . We obtain a hamiltonian cycle by replacing the edge
in C between uk and uk+1 with the path (uk , f, v, g, uk+1 ).

References
[1] J.-C. Bermond, A. Germa, M.-C. Heydemann, D. Sotteau: Hypergraphes Hamiltoniens, in Problèmes
combinatoires et thèorie des graphes (Colloq. Internat. CNRS, Univ. Orsay, Orsay, 1976). Colloq. In-
ternat. CNRS, 260 (CNRS, Paris, 1978), 39–43.

[2] D. Clemens, J. Ehrenmüller, and Y. Person, A Dirac-type theorem for Hamilton Berge cycles in random
hypergraphs, Electron. Notes Discrete Math., 54 (2016), 181–186.

[3] M. Coulson, G. Perarnau: A Rainbow Dirac’s Theorem, arXiv:1809.06392, (2018), 25 pp.

[4] G. A. Dirac: Some theorems on abstract graphs, Proc. London Math. Soc. (3) 2 (1952), 69–81.

[5] P. Erdős, T. Gallai: On maximal paths and circuits of graphs, Acta Math. Acad. Sci. Hungar. 10
(1959), 337–356.

[6] B. Ergemlidze, E. Győri, A. Methuku, N. Salia, C. Tompkins, and O. Zamora, Avoiding long Berge
cycles: the missing cases k = r + 1 and k = r + 2. Combin. Probab. Comput. 29 (2020), 423–435.

[7] Z. Füredi, A. Kostochka, R. Luo, Berge cycles in non-uniform hypergraphs, Electronic JC 27 (2020),
Paper No. 3.9, 13 pp.

[8] A. Kostochka and R. Luo, On r-uniform hypergraphs with circumference less than r, Discrete Appl.
Math. 276 (2020), 69–91.

[9] Y. Ma, X. Hou, J. Gao: A Dirac-type theorem for uniform hypergraphs, arXiv:2004.05073, (2020),
17 pp.

24