Noncrossing Longest Paths and Cycles∗

Greg Aloupis† Ahmad Biniaz‡ Prosenjit Bose§ Jean-Lou De Carufel¶

David Eppstein‖ Anil Maheshwari§ Saeed Odak¶ Michiel Smid§
Csaba D. Tóth∗∗ Pavel Valtr††
arXiv:2410.05580v1 [cs.CG] 8 Oct 2024

Abstract
Edge crossings in geometric graphs are sometimes undesirable as they could lead to unwanted
situations such as collisions in motion planning and inconsistency in VLSI layout. Short geo-
metric structures such as shortest perfect matchings, shortest spanning trees, shortest spanning
paths, and shortest spanning cycles on a given point set are inherently noncrossing. However,
the longest such structures need not be noncrossing. In fact, it is intuitive to expect many edge
crossings in various geometric graphs that are longest.
Recently, Álvarez-Rebollar, Cravioto-Lagos, Marı́n, Solé-Pi, and Urrutia (Graphs and Com-
binatorics, 2024) constructed a set of points for which the longest perfect matching is non-
crossing. They raised several challenging questions in this direction. In particular, they asked
whether the longest spanning path, on any finite set of points in the plane, must have a pair
of crossing edges. They also conjectured that the longest spanning cycle must have a pair of
crossing edges.
In this paper, we give a negative answer to the question and also refute the conjecture. We
present a framework for constructing arbitrarily large point sets for which the longest perfect
matchings, the longest spanning paths, and the longest spanning cycles are noncrossing.

1 Introduction
Traversing points in the plane by a polygonal path or cycle possessing a desired property has a
rich background. For instance, the celebrated travelling salesperson problem asks for a polygonal
∗
This work has been presented at the 32nd International Symposium on Graph Drawing and Network Visualization
(Vienna, 2024), GD 2024. The main results and ideas have also been reported at https://11011110.github.io/
blog/2024/09/25/long-non-crossing.html.
†
Khoury College of Computer Sciences, Northeastern University, Boston, MA, USA,
[email protected]
‡
School of Computer Science, University of Windsor, Windsor, ON, Canada, [email protected]. Research
supported by NSERC.
§
School of Computer Science, Carleton University, Ottawa, ON, Canada, [email protected],
[email protected], [email protected]. Research supported by NSERC.
¶
School of Electrical Engineering and Computer Science, University of Ottawa, Ottawa, ON, Canada,
[email protected], [email protected]. Research supported by NSERC.
‖
Computer Science Department, University of California, Irvine, CA, USA, [email protected]. Research sup-
ported by NSF grant CCF-2212129.
∗∗
Department of Mathematics, California State University Northridge, Los Angeles, CA; and Department of Com-
puter Science, Tufts University, Medford, MA, USA, [email protected]. Supported by NSF grant DMS-2154347.
††
Department of Applied Mathematics, Charles University, [email protected]. Research supported by Czech
Science Foundation grant GAČR 23-04949X.

1
path or cycle with minimum total edge length [9, 25, 27]. In recent years, there has been increased
interest in paths and cycles with properties such as being noncrossing [2, 17], minimizing the
longest edge length [6, 12, 24], maximizing the shortest edge length [7], minimizing the total or
largest turning angle [1, 14, 19, 22], and minimizing the number of turns [13, 18, 28] to name a
few. The longest cycle—the MaxTSP—is NP-hard in Euclidean spaces of dimension ≥ 3, but the
complexity of the planar MaxTSP is unknown [10, 11, 21].1 Paths and cycles that have combinations
of these properties have also attracted attention. For example, simultaneously being noncrossing
and having maximum total edge length [3, 20] is difficult to satisfy: to achieve a larger length we
typically introduce more crossings.
Edge crossings in geometric graphs are usually undesirable as they have the potential of cre-
ating unwanted situations such as collisions in motion planning and inconsistency in VLSI layout.
They are also undesirable in the context of graph drawing and network visualization as they make
drawings more difficult to read and use. Short geometric structures such as shortest perfect match-
ings, shortest spanning trees, shortest spanning paths, and shortest spanning cycles are inherently
noncrossing. This property, however, does not necessarily hold if the structure is not shortest.
For long structures such as longest perfect matchings, longest spanning trees, longest spanning
paths, and longest spanning cycles—the other end of the spectrum—it seems natural to expect
many crossings. Counting crossings in geometric graphs and finding geometric structures with a
minimum or maximum number of crossings are active research areas in discrete geometry. The
study of this type of problem attracted more attention after the work of Aronov et al. [8] in 1994,
who showed that any set of n points in the plane in general position admits a crossing family (a set
√
of pairwise intersecting segments) of size Ω( n). They also conjectured that the true lower bound
is linear in n. The current best lower bound, n1−o(1) , was established by Pach et al. [26] in 2019.
The noncrossing property of shortest structures is mainly ensured by the triangle inequality.
The triangle inequality, as noted by Alon et al. [3], also implies that the longest structures often
have crossings because a structure usually gets longer by creating more crossings. Alon et al. [3]
studied the problem of finding longest noncrossing structures (such as matchings, paths, or trees).
Some of their initial results have been improved and extended by Dumitrescu and Tóth [20] (for
matchings, paths, and cycles), by Biniaz et al. [15] and by Cabello et al. [16] (for trees). Along
this direction, one might wonder whether a longest structure (defined on an arbitrarily large point
set) is necessarily crossing. This was explicitly asked by Álvarez-Rebollar et al. [4]. Among other
interesting results, they presented arbitrarily large planar point sets for which the longest perfect
matching is noncrossing. They asked the following question and proposed the following conjecture:
Question 1 (Álvarez-Rebollar et al. [4]). For every sufficiently large planar point set, must the
longest spanning path have two edges that cross each other?
Conjecture 1 (Álvarez-Rebollar et al. [4]). The longest spanning cycle on every sufficiently large
set of points in the plane has a pair of crossing edges.
The “sufficiently large” condition in the question and conjecture makes sense, as otherwise one
can take any 3 points in general position, or any 4 points that are not in a convex position—for
such point sets, all spanning paths and cycles are noncrossing.
In the other direction, one might wonder about maximizing the number of crossings in cycles.
Here, we would like to highlight another result of Álvarez-Rebollar et al. [4, 5]. Let C(n) be the
1
It is interesting that the MaxTSP under the L1 -norm for points in the plane can be solved in linear time [11, 21],
in contrast to the fact that MinTSP in this case is NP-hard [23].

2
largest number such that any set of n points in the plane admits a spanning cycle with at least
C(n) pairs of crossing edges. Álvarez-Rebollar et al. [4, 5] established the following lower and upper
bounds: n2 /12 − O(n) < C(n) < 5n2 /18 − O(n). In other words, any set of n points in the plane
admits a spanning cycle with at least n2 /12 − O(n) crossings, and there is a family of point sets
that does not admit any cycle with more than 5n2 /18 − O(n) crossings.

1.1 Our contributions

In this paper, we provide negative answers to both Question 1 and Conjecture 1. For any integer
n ≥ 1 we present a set of n points in the plane for which the longest spanning path is unique and
noncrossing. Similarly, for any integer n ≥ 4, we present a set of n points in the plane for which the
longest spanning cycle is unique and noncrossing. To build such point sets, we use the following
framework: First, we choose a set P of points on the x-axis for which the longest structure may not
be unique. Then, we assign new y-coordinates to points in P to obtain a new point set P ′ for which
the longest structure corresponds to one in P and is also unique and noncrossing. In Section 6,
we present some structural properties of longest paths and cycles, which may be of independent
interest.

1.2 Preliminaries
All point sets considered in this paper are in the Euclidean plane. A geometric graph is a graph
with vertices represented by points and edges represented by line segments between the points. Let
P be a finite point set. A spanning path for P is a path drawn with straight-line edges such that
every point in P lies at a vertex of the path and every vertex of the path lies at a point in P . A
spanning cycle is defined analogously. In other words, a spanning path is a Hamiltonian path in
the complete geometric graph on P , and a spanning cycle is a Hamiltonian cycle in this graph.
Consider two line segments, each connecting a pair of points in P . If the interiors of the segments
intersect, then we say that they cross; this configuration is called a crossing. A path or a cycle
is called noncrossing if its edges do not cross each other. We denote the undirected edge between
two points p and q by pq, the directed edge from p towards q by (p, q), and the Euclidean distance
between p and q by |pq|. The length of a geometric graph G is the sum of the lengths of its edges,
and we denote it by |G|.

2 Longest Paths and Cycles on the Real Line

In this section we characterize longest paths and cycles in dimension one. These observations play
a pivotal role in our constructions in the plane (Sections 3 and 4). We say that an edge e intersects
a point p if the intersection of e and p is not empty (the intersection could be an endpoint of e).
For a sorted set of 2k+1 numbers, the median is the number with rank k+1, and for a sorted set
of 2k numbers, the median is the mean of the two numbers with ranks k and k+1.
Lemma 1. Let P be a set with an even number of points in R, i.e., in dimension one. The
endpoints of any longest spanning path on P lie on different sides of the median of P .
Proof. Let P = {p1 , . . . , pn } and assume w.l.o.g. that 0 is the median of P (in particular, 0 ∈
/ P ).
Let H be a longest spanning path on P . Orient the edges of H to make it a directed path. Let ps
and pe be the starting and ending points of H, respectively. For the sake of contradiction, assume

3
that ps and pe have the same sign, which we may assume, due to symmetry, to be positive. Thus
ps , pe > 0. Then, the sum of degrees of vertices in H to the left of the origin is 2 more than the sum
of degrees of vertices to the right. Therefore, H must have a directed edge (pa , pb ) where pa , pb < 0.
If pb < pa , then by replacing (pa , pb ) with the undirected edge ps pb we obtain a longer undirected
path; and if pb > pa by replacing (pa , pb ) with pe pa we obtain a longer undirected path. Both cases
lead to a contradiction.

Lemma 2. Let P be a set with an even number of points in R, i.e., in dimension one. Let H be a
spanning path on P . Then H is a longest spanning path if and only if
(i) every edge of H intersects the median of P , and

(ii) the two endpoints of H are the two points closest to the median of P .
Proof. Let P = {p1 , . . . , pn } so that pi < pj for all i < j ∈ {1, . . . , n}, and assume w.l.o.g. that 0
is the median of P . Note that 0 ∈ / P since n is even. First, we prove by contradiction that if H is
a longest spanning path, then (i) and (ii) hold.
Suppose that (i) does not hold. Orient the edges of H to make it a directed path. Let (pa , pb )
be an edge of H that does not intersect the median. Due to symmetry, assume that pa , pb < 0.
By Lemma 1, the endpoints of H lie on different sides of the median. This implies that both sides
have the same sum of vertex degrees. Thus H must have an edge (pc , pd ) such that c, d > 0. By
replacing these edges with pa pc and pb pd we obtain an (undirected) spanning path that is longer
than H because |pa − pc | + |pb − pd | > |pa − pb | + |pc − pd |. This contradicts H being a longest path.
Now suppose that (ii) does not hold: without loss of generality pn/2 is not an endpoint of H.
(The case for pn/2+1 can be handled symmetrically). Then H has an endpoint pa with a < n/2.
Orient the edges of H so that the path is directed from pa towards the other endpoint. Let (pn/2 , pb )
be the outgoing edge from pn/2 . By part (i), we have pb ≥ 0. By removing (pn/2 , pb ) we obtain two
paths, and pb is an endpoint on one of those paths. Next, join the paths with a new edge (pa , pb ).
Thus we obtain an (undirected) spanning path that is longer than H because |pa − pb | > |pn/2 − pb |.
This contradicts H being longest.
Finally, we prove that any spanning path H that satisfies (i) and (ii) is longest, using a direct
proof. Consider a longest spanning path L on P . By the sufficiency proof, (i) and (ii) hold for L.
This implies that the positive interval [pn/2 , pn/2+1 ] is contained in each of the n−1 edges, hence it
contributes to the length of L with multiplicity n−1. Similarly, for any i ∈ {2, . . . , n/2} the positive
interval [pn/2+i−1 , pn/2+i ] contributes to the length of L by multiplicity n − 2i. A similar argument
holds for negative intervals. See Figure 1. On the other hand, any spanning path (including H)
that satisfies (i) and (ii) receives the exact same multiplicities from the corresponding intervals.
Therefore H and L have the same length, and hence H is also a longest path.

A statement similar to that of Lemma 2 can be proved for paths with an odd number of points
(in this case one endpoint is the median itself and the other endpoint is the closest point to the
median). However, we will not use this in our construction.
Lemma 3. Let P be a finite set in R, i.e., in dimension one.
(i) A spanning cycle on P is longest iff each of its edges intersects the median of P .

(ii) If P contains an odd number of points, then for any longest spanning cycle the two edges
incident to the median lie on opposite sides of it.

4
 p1 p2 ... pn/2 pn/2+1 4 2 pn
n-1 multiplicity

Figure 1: Illustration of a longest path for a point set on a line, for the case where the number
of points, n, is even. Numbers below intervals [pn/2+i , pn/2+i+1 ] represent the multiplicity of the
contribution of the corresponding intervals to the length of the longest path.

(iii) Assume that P contains n = 2k+1 points and there is an interval I of length h > 0
between the leftmost k+1 and the rightmost k points. Then in any longest spanning cycle,
n−1 = 2k edges contain the interval I; and if a spanning cycle has fewer than 2k edges that
contain I, then it is at least 2h shorter than a longest cycle.

Proof. Let P = {p1 , . . . , pn } so that pi < pj for all i < j ∈ {1, . . . , n}, and assume w.l.o.g. that 0
is the median of P . Note that 0 ∈ / P if n is even, and p⌈n/2⌉ = 0 if n is odd.
First we prove the sufficiency of (i) by contradiction. Let C be a longest cycle on P , and
orient its edges to obtain a directed cycle. Suppose, for the sake of contradiction, that the edge
(pa , pb ) of C does not intersect the median. We may assume w.l.o.g. that pa , pb < 0. The sum of
vertex degrees strictly on the left and right side of the median are the same, and the edges that
contain 0 in their interior contribute 1 to both sums. Consequently, C contains an edge (pc , pd )
with pc , pd > 0; or (when n is odd) there are two edges incident to the median, say (pc , 0) and
(0, pd ) with pc , pd > 0. In the first case, we can replace edges (pa , pb ) and (pc , pd ) with (pa , pc ) and
(pb , pd ). In the second case, replace (pa , pb ) and (pc , 0) with (pa , pc ) and (pb , 0). In both cases, we
obtain a longer (undirected) spanning cycle, contradicting the maximality of C.
The necessity of (i) can be proved by a counting argument similar to that of Lemma 2-(i).
Now, we prove (ii) by contradiction. Without loss of generality, let 0 ∈ P be the median of P .
Suppose that the median is incident to two edges (pc , 0) and (0, pd ) with pc , pd > 0. Then, there is
a point in P to the right of 0 incident to an edge of C that does not contain 0 in its interior. Denote
this edge by (pa , pb ), where pa , pb < 0. We can replace edges (pa , pb ) and (pc , 0) with (pa , pc ) and
(pb , 0) to obtain a longer spanning cycle, contradicting the maximality of C.
To prove the first part of (iii), note that if n = 2k+1, then the median is the (k+1)-st point of
P , that we denote by p0 . Let C be a longest cycle on P . It is implied from (i) and (ii) that exactly
one edge of C (which is incident to p0 ) does not contain I. The remaining n−1 = 2k edges contain
I.
For the second claim in (iii), let C be a spanning cycle on P in which fewer than 2k edges
contain I. Orient the edges of C to obtain a directed cycle. The sum of degrees of the leftmost k+1
(resp., rightmost k) vertices is 2k+2 (resp., 2k), and the edges containing I have fewer than 2k left
(resp., right) endpoints. Consequently, the leftmost k+1 (resp., rightmost k) points in P induce
at least two edges (resp., one edge) of C. Therefore, C contains two edges, (pa , pb ) and (pc , pd ),
such that pa , pb are to the left of I and pb , pd are to the right of I. We can replace these two edges
with (pa , pc ) and (pb , pd ), to obtain a spanning cycle C ′ that traverses I two more times than C. In
particular, we have |C ′ | ≥ |C| + 2 |I| = |C| + 2h, hence |C| ≤ |C ′ | − 2h ≤ |Cmax | − 2h, where Cmax
is a longest cycle on P .

5
3 Noncrossing Longest Paths
Let n ≥ 1 be an integer. In this section, we construct n points for which the longest spanning
path is unique and noncrossing. This can be easily observed for n < 5: For example, for n = 4,
any spanning path of the vertices of a triangle and a point in the interior is noncrossing. Thus,
we will now assume that n ≥ 5. In Section 2, we uncovered some structural properties of longest
paths for n points on a line. Here we show how to construct a 2-dimensional point set starting with
n points on the x-axis and then assigning y-coordinates to the points. We show that the longest
path is unique and noncrossing. We describe our construction for the case where n is even; the
construction for the case where n is odd follows with some minor changes. The following theorem
summarizes our result in this section.

Theorem 1. For every integer n ≥ 1 there exists a set of n points in the plane for which the
longest spanning path is unique and noncrossing.

In Section 3.1 we give an overview of our construction for an even number of points. The details
and proofs are given in Section 3.2. The case of odd paths is considered in Section 3.3.

3.1 A path with an even number of points: An overview

For k ≥ 3, consider a set P of n = 2k points pi on the x-axis such that p1 = (0, 0) and pi = (i, 0)
for i = −1, ±2, . . . , ±k, as illustrated in Figure 2(a). Our construction would work even if we set
p1 = (1, 0); however, for a reason that will become clear in Section 3.3, we set p1 differently. The
longest spanning path for this point set is not unique. In fact, Lemma 2 implies that any spanning
path with endpoints p1 and p−1 and with all edges crossing the y-axis is a longest path. Conversely,
any longest path must have endpoints p1 and p−1 , and its edges must cross the y-axis. Let H be
the set of these paths. Let P ′ be the point set obtained by assigning to each point pi a y-coordinate
yi such that, as illustrated in Figure 2(b), the following holds:
1
= y1 ≫ y−2 ≫ y2 ≫ y−3 ≫ y3 ≫ · · · ≫ y−k ≫ yk ≫ y−1 = 0.
8k
The value y1 is much larger than y−2 , which is in turn much larger than y2 and so on. Notice that
the largest y-coordinate y1 is 1/8k which is much smaller than 1. Due to the small y-coordinates,
a longest path H ′ on P ′ corresponds to a path H ∈ H. The length of H ′ is roughly the length
of H plus a very small value ∆(H ′ ), which depends on the new y-coordinates. Let e1 be the only
edge of H ′ incident to p1 . Since p1 has a very large y-coordinate compared to other points, the
contribution of e1 to ∆(H ′ ) is larger than the contribution of other edges. The contribution of e1
is maximized if it connects p1 to the nearest plausible neighbor, which is p−2 ; this can be observed
from Figure 2(b). Therefore e1 = p1 p−2 . By a similar argument, p−2 gets connected to p2 , and so
on. It follows that the path H ′ is unique and it is p1 , p−2 , p2 , p−3 , p3 , . . . , p−k , pk , p−1 . This path is
y-monotone, and hence noncrossing; see Figure 2(b).

Note. Figures 2(a) and 2(b) are not to scale. The y-coordinates should be small enough so that
all points lie almost on the x-axis (We exaggerated the y-coordinates to facilitate readability).
Moreover, if we orient the path from p1 towards p−1 , then the extension of every directed edge
intersects all edges that follow.

6
 1/8k p1

ℓk ℓ2

p−k p−2 p−1

∆2 p2 pk
∆k
(a)

1/8k p1

p−2
δ
p2
p−k pk
p−1
(b)

Figure 2: Illustration of the construction of a longest path for 2k points. The figure is not to scale
as the real y-coordinates are very small so that the points lie almost on the x-axis. (a) Lifting p1
to the y-coordinate 1/8k. (b) The final longest path.

3.2 A path with an even number of points: Details

Recall the set P of 2k points, k ≥ 3, on the x-axis, described in the previous section and illustrated
in Figure 2(a). We say that an edge e intersects the y-axis if the intersection of e and the y-axis
is not empty (the intersection could be an endpoint of e). The longest paths for points on a line
were characterized in Lemma 2. Denote by H the set of all longest spanning paths on P .
1
Lemma 4. Let 0 ≤ ε ≤ 8k be a real number. Suppose that every point in P is perturbed by a
distance of at most ε. Let P ′ be the new point set after perturbation. Then, the order of the points
along any longest path for P ′ is the same as the order of the points along some path in H.

Proof. The length of any path on P is an integer. Therefore, any path in H is at least 1 unit longer
than any path not in H.
Let H ′ be any longest path on P ′ . The difference between its length and the length of any
path in H is at most (2k−1) · 2ε because H ′ has 2k−1 edges, each edge has 2 endpoints, and each
1
endpoint is at distance at most ε from its corresponding point in P . Since ε ≤ 8k the difference is

7
less than 1/2. Therefore, H ′ cannot correspond to a path that is not in H, so H ′ corresponds to a
path in H with the same order of points.

Our plan is to assign new y-coordinates to the points of P to obtain a point set P ′ for which the
1
longest path is y-monotone and unique. The new y-coordinates will be at most 8k , and thus, by
Lemma 4, the longest path H of P will correspond to a path in H. We will make H ′ correspond to
′ ′

the path p1 , p−2 , p2 , p−3 , p3 , . . . , pk , p−1 , which is in H (by Lemma 2) and depicted in Figure 2(b).
We assign to each point pi the y-coordinate yi such that the following holds:

y1 ≫ y−2 ≫ y2 ≫ y−3 ≫ y3 ≫ · · · ≫ y−k ≫ yk ≫ y−1 .

1
We set y1 = 8k , y−1 = 0, and use the following lemma to identify the remaining y-coordinates.
Lemma 5. There exists a real number δ, with 0 < δ < y1 , such that if 0 ≤ yi ≤ δ for each i ̸= 1
then the longest path on P ′ connects p1 to p−2 .
Proof. Since each yi is at most 1/8k, Lemma 4 implies that any longest path H ′ on P ′ corresponds
to a path H in H. Due to small y-coordinates, we have |H ′ | = |H| + ∆(H ′ ) for some small value
∆(H ′ ) ≥ 0 which depends on the new y-coordinates. Specifically, we have
X X q
|H ′ | = |pi pj | = |i − j|2 + |yi − yj |2
(pi ,pj )∈E(H ′ ) (pi ,pj )∈E(H ′ )
X q 
= |H| + |i − j|2 + |yi − yj |2 − |i − j| = |H| + ∆(H ′ ),
(pi ,pj )∈E(H ′ )
p
where 0 ≤ |i − j|2 + |yi − yj |2 − |i − j| ≤ |yi − yj | ≤ max{yi , yj }.
Recall from Lemma 2 that p1 is an endpoint of any longest path in H. Moreover, p1 is connected
to a point (different from p−1 ) to the left of the y-axis. For j ∈ {2, . . . , k} let ℓj be the Euclidean
distance between p1 and the point (−j, 0), and let ∆j be the difference of their x-coordinates as in
Figure 2(a). The contribution of p1 p−j to |H ′ | would be at least ℓj − δ (when p−j has y-coordinate
δ) and at most ℓj (when p−j has y-coordinate 0). The contribution of the corresponding edge to
|H| would be ∆j . Hence the contribution of p1 p−j to ∆(H ′ ) would be at least ℓj − δ − ∆j and at
most ℓj − ∆j . An easy calculation shows that

ℓ2 − ∆2 > ℓ3 − ∆3 > · · · > ℓk − ∆k ;

this is also implied by the fact that ∆i+1 −∆i = 1 while ℓi+1 −ℓi < 1. If we set δ < (ℓ2 − ∆2 ) − (ℓ3 −
∆3 ), then the contribution of p1 p−2 to ∆(H ′ ) is at least

ℓ2 − δ − ∆2 > ℓ2 − ∆2 − ((ℓ2 − ∆2 ) − (ℓ3 − ∆3 )) = ℓ3 − ∆3 ,

which is larger than the contribution of any other plausible edge p1 p−j . Since the y-coordinates of
all other points are less than δ, any other edge of H ′ contributes less than δ to ∆(H ′ ). By setting
(ℓ2 − ∆2 ) − (ℓ3 − ∆3 )
δ= ,
2k − 1
the contribution of p1 p−2 exceeds the sum of the contributions of the remaining 2k − 2 edges of H ′ .
Thus, for this choice of δ the longest path H ′ connects p1 to p−2 .

8
 By Lemma 5, we have a specific value δ such that the longest path includes edge (p1 , p−2 ). Now
we set y−2 = δ and repeat the arguments of Lemma 5, with y−2 and p−2 (instead of y1 and p1 ).
This implies that the next edge of the longest path will connect p−2 to p2 . Repeating this 2k−5
more times, we obtain the unique longest path p1 , p−2 , p2 , p−3 , p3 , . . . , pk , p−1 , as in Figure 2(b); in
each of the last two steps, there is only one remaining plausible edge (namely, p−k pk from p−k , and
pk p−1 from pk ). This path is y-monotone and hence is noncrossing.

3.3 A path with an odd number of points

In this section, we obtain a noncrossing longest path with an odd number of points. Here is the
place where we use the coordinate (0, 0) of the point p1 ∈ P . We show that our construction for
even paths leads to a construction for odd paths by simply removing p1 . Thus we do not need to
repeat the lemmas of Section 3.2 for the odd case.
We claim that if we remove the point p1 from the path H ′ constructed on P ′ in the previous
section, the remaining path, i.e., H ′′ = p−2 , p2 , p−3 , p3 , . . . , pk , p−1 , is the longest path for the
remaining 2k−1 points. By construction, |H ′ | = |H ′′ | + |p1 p−2 |. Assume, for the sake of proof by
contradiction, that the longest path L for the remaining points is longer than H ′′ . Among the two
endpoints of L, let pi be an endpoint that is not p−1 . Due to our choices of the x- and y-coordinates
we have |p1 pi | ≥ |p1 p−2 |. Therefore the concatenation of L and p1 pi would give a path on P ′ of
length |L| + |p1 p−2 | which is larger than |H ′ |. This contradicts H ′ being the longest path on P ′ .

4 Noncrossing Longest Cycles

Let n ≥ 3 be an integer. In this section, we construct a set of n points for which the longest
spanning cycle is unique and noncrossing. For n = 3, every spanning cycle is noncrossng. For
n = 4, we take three vertices of a triangle and a point in the interior. Thus, we assume that n ≥ 5.

Theorem 2. For every integer n ≥ 3 there exists a set of n points in the plane for which the
longest spanning cycle is unique and noncrossing.

In Section 4.1 we give an overview of our construction for an even number of points. The
details and proofs are given in Section 4.2. For an odd number of points we sketch a construction
in Section 4.3.

4.1 A cycle with an even number of points: An overview

Let n ≥ 6 be an even number. Then either n = 4k or n = 4k−2 for some integer k. To simplify the
indexing (of points and y-coordinates) in our construction, from now on we assume that n = 4k−2.
Let P be a set of n points, consisting of 2k points pi = (i, 0) for i = ±1, ±2, . . . , ±k and 2k−2
points p′i = (i+ϵ, 0) for i = −1, ±2, . . . , ±(k−1), k, where ϵ > 0 is a small value to be determined;
see Figure 3. (The construction for n = 4k is similar; it consists of P and two additional points
pk+1 = (k+1, 0) and p′−k = (−k+ϵ, 0).) Our construction for cycles is somewhat similar to that of
paths in the sense that our cycle consists of two y-monotone interior-disjoint paths between p1 and
p−k (or between p1 and pk+1 when n is a multiple of 4). Although the main idea sounds simple,
the noncrossing property of the longest cycle is not straightforward and involves a more detailed
analysis.

9
 Lemma 3 implies that a spanning cycle on P is longest if and only if each of its edges intersects
the y-axis. Let C be the set of all longest spanning cycles on P . As illustrated in Figure 3, we
obtain a point set P ′ by assigning to each point pi and p′i the respective y-coordinates yi and yi′
such that:
1
= y1 ≫ y−1 ≫ y2′ ≫ y−2 ≫ y3′ ≫ · · · ≫ yk′ ≫ y−k = 0.
16k
For each i ∈ {2, 3, . . . , k} we choose yi such that pi lies just below (almost on) the segment p′−i+1 p′i ,
and for each i ∈ {−1, −2, . . . , −(k−1)} we choose yi′ such that p′i lies just below (almost on) the
segment p−i pi .
Due to the small y-coordinates, any longest cycle C ′ on P ′ corresponds to a cycle C ∈ C.
Moreover |C ′ | = |C| + ∆(C ′ ) for some small value ∆(C ′ ) which depends on the new y-coordinates.
Since p1 has the largest y-coordinate, the contribution of the two edges of C ′ that are incident to
p1 (say e1 and e2 ) is maximized when they are connected to the nearest plausible neighbors which
are p−1 and p′−1 . We will choose the y-coordinates in such a way that the contribution of e1 and
e2 is larger than the sum of the contributions of the remaining edges of the cycle. Thus C ′ must
connect p1 to p−1 and p′−1 . Similarly, by a suitable choice of y-coordinates, we enforce C ′ to connect
p−1 and p′−1 to the nearest plausible neighbors which are p2 and p′2 , and so on. By repeating this
process, the longest cycle C ′ would be the concatenation of two paths p1 , p−1 , p2 , p−2 , . . . , p−k and
p1 , p′−1 , p′2 , p′−2 , . . . , p′k , p−k .

y1 = 1/16k p1

p′−1
p−1 <ϵ
y−1
p2
p′2
′
p−2 p−2 y2′
y−2 pk p′k
p−k yk′
ϵ ϵ ϵ ϵ
y−k = 0 −2 −1 1 2 k

Figure 3: Illustration of the construction of a longest cycle for 4k−2 points. The figure is not to
scale. The y-coordinates should be small enough so that all points lie almost on the x-axis.

4.2 A cycle with an even number of points: Details

Recall the point set P from the previous section (the y-coordinates and the value of ϵ > 0 will be
determined in this section). The longest cycles for points on a line were characterized in Lemma 3.
Let C be the set of all longest cycles on P .

Lemma 6. Any cycle in C is at least 1 unit longer than any cycle not in C.

10
Proof. Consider any cycle D that is not in C. Lemma 3 implies that D has an edge that does not
intersect the y-axis. Orient the edges of D to make it a directed cycle. Since the number of points
to the left of the y-axis is the same as the number of points to its right, D has two directed edges
(pa , pb ) and (pc , pd ) such that a, b ≤ −1 and c, d ≥ 1. By replacing these edges with pa pc and pb pd
we obtain an (undirected) spanning cycle D′ such that
|D′ | − |D| = (|pa pc | + |pb pd |) − (|pa pb | + |pc pd |) ≥ 2|p1 p′−1 | = 2(2 − ϵ) > 1.
Since the length of any cycle C in C is at least |D′ |, we get |C| > |D| + 1.

Lemma 7. Let 0 ≤ ε ≤ 1/16k be a real number. Suppose that every point of P is perturbed by a
distance of at most ε. Then the order of the points along any longest cycle of the new point set is
the same as the order of the points along some cycle in C.
Proof Sketch. The proof is similar to that of Lemma 4 and uses Lemma 6. The parameter ε is
small enough such that the total change in the length of any spanning cycle on P is less than 1/2.
Together with Lemma 6, this implies that any longest cycle on the perturbed points corresponds
to a cycle in C.

To obtain P ′ we only need to describe the following y-coordinates:

y1 ≫ y−1 ≫ y2′ ≫ y−2 ≫ y3′ ≫ · · · ≫ yk′ ≫ y−k .
The y-coordinates of the remaining points would then follow as outlined in the previous section
1
(more details are given after Lemma 9). We set y1 = 16k and y−k = 0. We use the following lemma
(which can be proven similarly to Lemma 5) to assign the y-coordinates.
Lemma 8. There exists a real number δ, ϵ ≤ δ < y1 , such that if 0 ≤ yi ≤ δ for i ̸= 1 and
0 ≤ yi′ ≤ δ for i ̸= −1, then every longest cycle of P ′ connects p1 to p−1 and p′−1 .
Proof. Lemma 7 implies that any longest cycle C ′ on P ′ corresponds to a cycle C in C. Due to
small y-coordinates, we have |C ′ | = |C| + ∆(C ′ ) for some small value ∆(C ′ ) ≥ 0 which depends on
the new y-coordinates. Lemma 3 implies that C ′ connects p1 to two points to the left of the y-axis.
Similar to Lemma 5, for j ∈ {1, . . . , k} define ℓj as the Euclidean distance between p1 and the point
(−j, 0), and define ∆j as the difference of their x-coordinates. Analogously, for j ∈ {1, . . . , k−1}
define ℓ′j and ∆′j for p1 and the point (0, −j + ϵ). Every edge that connects p1 to a point to the left
of the y-axis has the following contributions to |C|, |C ′ | and ∆(C ′ ).

• For j ∈ {1, . . . , k} the contribution of p1 p−j to |C ′ | is at least ℓj − δ and at most ℓj . The

contribution of the corresponding edge to |C| is ∆j . Hence the contribution of p1 p−j to ∆(C ′ )
is at least ℓj − δ − ∆j and at most ℓj − ∆j .
• For j ∈ {2, . . . , k−1} the contribution of p1 p′−j to |C ′ | is at least ℓ′j − δ and at most ℓj . The
contribution of the corresponding edge to |C| is ∆′j . Thus the contribution of p1 p′−j to ∆(C ′ )
is at least ℓ′j − δ − ∆′j and at most ℓ′j − ∆′j .
• The contribution of p1 p′−1 to |C ′ | is at least ℓ′1 − δ − ϵ because the y-coordinate of p′−1 is at
most δ + ϵ; to verify this observe that y−1 ≤ δ and y−1 ′ −y ′
−1 < ϵ because p−1 is almost on
p1 p−1 whose slope is less than 1; also see Figure 3 (recall that the figure is not to scale). The
contribution of the corresponding edge to |C| is ∆′1 . Therefore the contribution of p1 p′−1 to
∆(C ′ ) is at least ℓ′1 − δ − ϵ − ∆′1 and at most ℓ′1 − ∆′1 .

11
 Observe that

ℓ′1 − ∆′1 > ℓ1 − ∆1 > ℓ′2 − ∆′2 > ℓ2 − ∆2 > · · · > ℓk − ∆k .

If we set δ < 12 ((ℓ1 − ∆1 ) − (ℓ′2 − ∆′2 )), then the contributions of p1 p−1 and p1 p′−1 to ∆(C ′ ) would
respectively be at least

ℓ1 − δ − ∆1 > ℓ1 − 2δ − ∆1 > ℓ1 − ∆1 − ((ℓ1 − ∆1 ) − (ℓ′2 − ∆′2 )) = ℓ′2 − ∆′2 , and

ℓ′1 − δ − ϵ − ∆′1 ≥ ℓ′1 − 2δ − ∆′1 > ℓ′1 − ∆′1 − ((ℓ1 − ∆1 ) − (ℓ′2 − ∆′2 )) > ℓ′2 − ∆′2 ,
which are larger than the contribution of any other edge p1 p−j and p1 p′−j . By setting

1 (ℓ1 − ∆1 ) − (ℓ′2 − ∆′2 )

δ=
2 4k − 2
the contribution of each of p1 p−1 and p1 p′−1 would be even larger than the sum of the contributions
of the remaining 4k−4 edges of C ′ . Thus, for this choice of δ the longest cycle C ′ connects p1 to
p−1 and p′−1 .

We choose δ as in the proof of Lemma 8, and set y−1 = δ. Then we set y−1′ so that p′
−1 lies just
′
below (almost on) the segment p1 p−1 , as in Figure 3. Notice that δ < y−1 < δ + ϵ = y−1 + ϵ. Then,
by Lemma 8 the longest cycle connects p1 to p−1 and p′−1 . By Lemma 3, the other edges incident
to p−1 and p′−1 must cross the y-axis.

Lemma 9. There exists a real δ, ϵ ≤ δ < y−1 , such that if 0 ≤ yi ≤ δ for i ̸= −1, 1, 2 and 0 ≤ yi′ ≤ δ
for i ̸= −1, then every longest cycle of P ′ connects p−1 to p2 and p′−1 to p′2 .

Proof. Recall the longest cycle C ′ from the proof of Lemma 8. We choose δ small enough such that
the contribution of each of p−1 p2 , p−1 p′2 , p′−1 p2 , and p′−1 p′2 to ∆(C ′ ) is larger than the sum of the
contributions of the remaining 4k−6 edges of C ′ . This would force C ′ to connect p−1 and p′−1 to
p2 and p′2 .
By an argument similar to that of Lemma 8 we can find a parameter δ1 that forces C ′ to connect
p−1 to p2 or p′2 (δ1 , y−1 , p−1 , p2 , and p′2 play the roles of δ, y1 , p1 , p′−1 , and p−1 , respectively).
Similarly, we can find a parameter δ1′ that forces C ′ to connect p′−1 to p2 or p′2 (where δ1′ , y−1 ′ , p′ ,
−1
p2 , and p2 play the roles of δ, y1 , p1 , p−1 , and p−1 , respectively). Then we choose δ = min{δ1 , δ1′ }.
′ ′

Our choice of δ ensures that C ′ connects p−1 and p′−1 to p2 and p′2 . Notice that p−1 and p′−1
cannot both connect to p2 or to p′2 because it closes the cycle. Thus C ′ must use p−1 p2 and p′−1 p′2
or p−1 p′2 and p′−1 p2 . We show that C ′ uses p−1 p2 and p′−1 p′2 . See Figure 4. Recall that p2 is
almost on the edge p′−1 p′2 , and hence |p′−1 p′2 | ≈ |p′−1 p2 | + |p2 p′2 |. By the triangle inequality we get
|p−1 p2 | + |p2 p′2 | > |p−1 p′2 |. Adding these two yields

|p−1 p2 | + |p′−1 p′2 | > |p−1 p′2 | + |p′−1 p2 |, (1)

which means that C ′ connects p−1 to p2 and p′−1 to p′2 .

12
 p′−1

p−1
p2
p′2

Figure 4: The longest cycle connects p−1 to p2 and p′−1 to p′2

We choose our new δ as in the proof of Lemma 9, and set y2′ = δ. Now that the point p′2 is fixed
we can choose the y-coordinate of p2 in the triangle △p−1 p′−1 p′2 and very close to the segment p′−1 p′2
such that (1) holds. This forces the longest cycle to use p−1 p2 and p′−1 p′2 . By repeatedly applying
Lemma 9, the longest cycle will use the edges pi p−i and p′i p′−i (for i > 0) and the edges pi p−i+1 and
p′i p′−i+1 (for negative i < 0). Therefore the longest cycle on P ′ is the concatenation of two paths:
p1 , p−1 , p2 , p−2 , . . . , p−k and p1 , p′−1 , p′2 , p′−2 , . . . , p′k , p−k . This cycle is unique and noncrossing.
Each time we apply Lemma 9 we obtain a new value for δ. In each application we need δ to be
greater than or equal to our fixed parameter ϵ. For this purpose, we choose ϵ to be the parameter
δ that is obtained in the last application of Lemma 9, i.e., δ = yk′ .

4.3 A cycle with an odd number of points: An overview

Our construction uses the longest paths of Section 3.2. First we observe that our path construction
can be generalized to any set of x-coordinates.

Lemma 10. For every even integer n ≥ 4, every set P of n real numbers, and every δ > 0 such
that the δ-neighborhood of the median of P does not contain any points in P , there exists a set P ′
of n points in the plane with the following properties:

1. the x-projection of P ′ is P ;

2. all y-coordinates are in the interval [0, δ];

3. the x-projection of any longest path on P ′ is a longest path on P ;

4. the longest spanning path on P ′ is unique and noncrossing; and

5. the y-coordinates of the two endpoints of the longest path are 0 and δ.

Proof sketch. We choose the points in P ′ such that their x-coordinates are the same as the numbers
in P and their y-coordinates are in [0, δ], and thus (1) and (2) follow.
By an argument similar to the proof of Lemma 2(i) one can show that the difference of lengths
of a longest and a non-longest path on P is at least 2δ. Therefore Lemma 4 would imply that
by choosing the y-coordinates in the interval [0, 2δ/8k], any longest path on P ′ corresponds to a
longest path on P , and thus (3) follows. Items (4) and (5) follow by proper choices of y-coordinates
similar to that of Lemma 5.

We can now outline the construction; see Figure 5 for an illustration. Let n = 2k+1, for
k ≥ 2. We choose a set of x-coordinates as P = {−k, −(k−1)ϵ, −(k−2)ϵ, . . . , −ϵ, 0, 1, 2, . . . , k},

13
where ϵ ∈ (0, 1/16k 2 ) will be specified later. Note that 0 is the median of P , and the set A =
{−i · ϵ : i = 0, 1, . . . , k − 1} ⊂ [−1/16k, 0] forms a small cluster. By Lemma 3(ii), all edges of any
longest cycle on P intersect the y-axis; and Lemma 3(iii) implies the following.
Observation 1. The length of any cycle on P that connects p−k to two points in A is at least 2
units shorter than a longest cycle on P .

y0 = δ p0

p2

p3 pk
ϵ ϵ ϵ
0
−k −ϵ 1 2 3 k
−(k − 1)ϵ

p−k y−k = 1/8k

Figure 5: Illustration of the construction of a longest cycle for 2k+1 points.

Below, we will specify a y-coordinate for each element in P . This will result in the point set
P ′ for which the longest spanning cycle is unique and noncrossing. We will denote by A′ the set of
points in P ′ corresponding to A.
It remains to specify the y-coordinates of the points in P ′ and the parameter ϵ. Let px denote
the point in P ′ with x-coordinate x ∈ P . We first choose the y-coordinate for the leftmost point:
Let y−k = −1/16k; this is the only negative y-coordinate. We assume that |yi | ≪ 1/16k for all
other points. This ensures that the longest cycle on P ′ corresponds to a longest cycle on the 1-
dimensional multiset where 0 represents the entire cluster A (cf. Lemma 3(iii) and Lemma 7). By
Lemma 3(ii), for any longest cycle on P ′ , the two edges incident to p−k intersect the y-axis (i.e.,
the median). Furthermore, there is a threshold δ > 0 such that if 0 ≤ yi ≤ δ for all remaining
points, then p−k must be adjacent to the two closest points on or to the right of the y-axis: That
is, p−k is adjacent to a point in cluster A′ and to p1 (cf. Observation 1 and Lemma 7). Next, we
set y0 = δ and find a threshold δ1 ∈ (0, δ) such that if 0 ≤ yi ≤ δ1 for all remaining points and
0 < ϵ < δ1 , then the contribution of edge p−k p0 exceeds the sum of contributions of all remaining

14
edges of a spanning cycle. Consequently, the longest cycle must include the edge p−k p0 . Now both
p−k and p0 are fixed, and we choose a sufficiently small ϵ ∈ (0, δ1 ) such that all remaining points
in the cluster A′ are below p−k p0 for all possible y-coordinates.
A longest cycle on P ′ comprises of p−k p0 , p−k p1 , and the longest path H ′ on P ′ \ {p−k } (from p0
to p1 cf. Lemma 2). By Lemma 10, we can choose y-coordinates for the remaining points such that
H ′ is unique and noncrossing; and y1 = 0. In particular, edge p−k p1 lies below the x-axis, hence
below the entire path H ′ ; and P ′ \ {p−k , p0 } lies below the supporting line of p−k p0 . Consequently,
the concatenation of p−k p0 , p−k p1 and H ′ is noncrossing.

Remark. Our construction in this section suggests an alternative construction for even cycles that
can be obtained by connecting a point to both endpoints of an odd path.

5 Noncrossing Longest Matchings

Álvarez-Rebollar et al. [4] showed that there exist point sets for which the longest perfect matchings
are noncrossing. Their example is attributed to Kåra P. Villanger in a paper by Tverberg [29]. As
illustrated in Figure 6, it consists of a set S of k segments with endpoints in A = {a1 , . . . , ak }
and B = {b1 , . . . , bk }. The distance between any two points ai ∈ A and bj ∈ B is larger than the
distance between any two points in A, or the distance between any two points in B. The points in
B are roughly on a vertical line. Álvarez-Rebollar et al. [4] have provided a precise description of
the construction along with a detailed proof that S is a longest matching for A ∪ B.
b4

b3
b2
a4
a3
a1 a2 b1

Figure 6: Villanger’s configuration as illustrated in [4].

Here, we exhibit an alternative point set for which the longest perfect matching is noncrossing.
Our construction follows the same framework as for paths and cycles. Let P be a set of 2k points
pi = (i, 0) for i = ±1, ±2, . . . , ±k. One can verify that a perfect matching on P is longest if and only
if all edges cross the y-axis. One such matching is M = {p−i pi : i = 1, . . . , k}. Using ideas similar
to those used for paths and cycles, one can assign to each pi a new y-coordinate yi to make M
longest and noncrossing at the same time; see Figure 7. The new y-coordinates are of the following
form: y1 ≫ y−1 = y2 ≫ y−2 = y3 ≫ · · · = yk ≫ y−k .

6 Some Properties of Longest Paths and Cycles

In this section we give some structural properties of longest paths and cycles, possibly of inde-
pendent interest. We state these properties only for cycles, but they hold for paths as well. Two
edges are in convex position if they are edges of their convex hull. Two directed edges in convex

15
 p−k p−2 p−1 p1 p2 pk

Figure 7: Illustration of our construction of a longest matching.

position have the same orientation if they are both directed clockwise or counterclockwise along
their convex hull.

Observation 2. Suppose that we orient the edges of a longest cycle C to make it a directed cycle.
Then C cannot have pair of non-adjacent edges that are in convex position and have the same
orientation along their convex hull.

To verify this, note that if C has two such edges, say e1 and e2 , then flipping them (replacing
e1 and e2 by the two diagonals of the convex hull of e1 and e2 ) would produce a longer undirected
cycle as in Figure 8(a). Since e1 and e2 have the same orientation along their convex hull, the
flip does not break the cycle into two components. If every directed simple polygon S contained
a pair of non-adjacent edges in convex position with the same orientation along their convex hull,
Observation 2 would imply Conjecture 1. However, some simple polygons do not have edges that
can be flipped in this way; see e.g., Figure 8(b).

e1

e2

(a) (b)

Figure 8: (a) Flipping two edges in convex position. (b) A simple polygon with no pair of edges in
convex position that have the same orientation, no matter how we direct the polygon.

Observation 3. The longest cycle need not contain an edge between diametric points.

To verify this observation consider an isosceles right triangle abc whose right angle is at b. Place
one point at a, one point at c, and two or more points very close to b. Then, the longest cycle does
not contain the diametric point pair {a, c}. This observation implies that a longest cycle may not
be achieved by greedily choosing longest edges.

16
 The following proposition implies that if the longest cycle is noncrossing, it contains some edge
whose length is among the smallest three-quarters of all distances defined by its vertices.
Proposition 1. Let S be a simple polygon (a noncrossing cycle) on
n points. Then S has an edge
whose length is among the smallest 3n2 /8 + n/8 distances of the n2 point pairs.
Proof. Let e and e′ be two edges of S such that their distance along S (in terms of the number of
edges) is at least 2. Since S is a simple polygon, e and e′ do not cross. Thus, there is an endpoint
p of e and an endpoint p′ of e′ such that |pp′ | is larger than the length of the shorter of e and e′ ,
and pp′ is not an edge of S. The number of pairs of edges at distance 0 is n, and the number of
pairs of edges
at distance 1 is also n. Thus, the total number′ of pairs of edges at a distance at
n ′
least 2 is 2 − 2n. Each such pair of edges yields a pair {p, p }. Each {p, p } can be counted for
4 different pairs of edges that are obtained by combining the two edges incident to p and
the two
n
edges incident to p′ . Therefore the total number ′ } is at least 1


of distinct pairs {p, p 4 2 − 2n .
Subtracting this from the total number n2 of point pairs yields the claimed bound.

Acknowledgements. This work was initiated at the 10th Annual Workshop on Geometry and
Graphs, held at Bellairs Research Institute in Barbados in February 2023. We thank the organizers
and the participants.

References
[1] Alok Aggarwal, Don Coppersmith, Sanjeev Khanna, Rajeev Motwani, and Baruch Schieber.
The angular-metric traveling salesman problem. SIAM Journal on Computing, 29(3):697–711,
1999. doi:10.1137/S0097539796312721.
[2] Oswin Aichholzer, Sergio Cabello, Ruy Fabila Monroy, David Flores-Peñaloza, Thomas Hackl,
Clemens Huemer, Ferran Hurtado, and David R. Wood. Edge-removal and non-crossing config-
urations in geometric graphs. Discrete Mathematics & Theoretical Computer Science, 12(1):75–
86, 2010. doi:10.46298/dmtcs.525.
[3] Noga Alon, Sridhar Rajagopalan, and Subhash Suri. Long non-crossing configurations in the
plane. Fundam. Inform., 22(4):385–394, 1995. doi:10.3233/FI-1995-2245.
[4] Jose Luis Álvarez-Rebollar, Jorge Cravioto-Lagos, Nestaly Marı́n, Oriol Andreu Solé-Pi, and
Jorge Urrutia. Crossing and intersecting families of geometric graphs on point sets. Graphs
Comb., 40(1):17, 2024. doi:10.1007/S00373-023-02734-9.
[5] José Luis Álvarez-Rebollar, Jorge Cravioto-Lagos, and Jorge Urrutia. Crossing families and
self crossing Hamiltonian cycles. In Abstracts of XVI Encuentros de Geometrı́a Computacional,
page 13, 2015.
[6] Hyung-Chan An, Robert Kleinberg, and David B. Shmoys. Approximation algorithms for the
bottleneck asymmetric traveling salesman problem. ACM Trans. Algorithms, 17(4):35:1–35:12,
2021. doi:10.1145/3478537.
[7] Esther M. Arkin, Yi-Jen Chiang, Joseph S. B. Mitchell, Steven Skiena, and Tae-Cheon
Yang. On the maximum scatter traveling salesperson problem. SIAM Journal on Computing,
29(2):515–544, 1999. doi:10.1137/S0097539797320281.

17
 [8] Boris Aronov, Paul Erdős, Wayne Goddard, Daniel J. Kleitman, Michael Klugerman, János
Pach, and Leonard J. Schulman. Crossing families. Combinatorica, 14(2):127–134, 1994.
doi:10.1007/BF01215345.

[9] Sanjeev Arora. Polynomial time approximation schemes for Euclidean traveling salesman and
other geometric problems. Journal of the ACM, 45(5):753–782, 1998. doi:10.1145/290179.
290180.

[10] Alexander Barvinok, Edward Kh. Gimadi, and Anatoliy I. Serdyukov. The maximum TSP.
In Gregory Gutin and Abraham P. Punnen, editors, The Traveling Salesman Problem and Its
Variations, pages 585–607. Springer, Boston, MA, 2007. doi:10.1007/0-306-48213-4_12.

[11] Alexander I. Barvinok, Sándor P. Fekete, David S. Johnson, Arie Tamir, Gerhard J. Woeginger,
and Russell Woodroofe. The geometric maximum traveling salesman problem. J. ACM,
50(5):641–664, 2003. doi:10.1145/876638.876640.

[12] Ahmad Biniaz. Euclidean bottleneck bounded-degree spanning tree ratios. Discrete & Com-
putational Geometry, 67(1):311–327, 2022. doi:10.1007/s00454-021-00286-4.

[13] Ahmad Biniaz. Improved bounds for covering paths and trees in the plane. In Proc. 39th
International Symposium on Computational Geometry (SoCG), volume 258 of LIPIcs, pages
19:1–19:15. Schloss Dagstuhl, 2023. doi:10.4230/LIPICS.SOCG.2023.19.

[14] Ahmad Biniaz. Acute tours in the plane. Discrete & Computational Geometry, 72:665–673,
2024. doi:10.1007/s00454-023-00486-0.

[15] Ahmad Biniaz, Prosenjit Bose, Kimberly Crosbie, Jean-Lou De Carufel, David Eppstein, Anil
Maheshwari, and Michiel H. M. Smid. Maximum plane trees in multipartite geometric graphs.
Algorithmica, 81(4):1512–1534, 2019. doi:10.1007/S00453-018-0482-X.

[16] Sergio Cabello, Michael Hoffmann, Katharina Klost, Wolfgang Mulzer, and Josef Tkadlec.
Long plane trees. In Proc. 38th International Symposium on Computational Geometry (SoCG),
volume 224 of LIPIcs, pages 23:1–23:17. Schloss Dagstuhl, 2022. doi:10.4230/LIPICS.SOCG.
2022.23.

[17] Jakub Cerný, Zdenek Dvorák, Vı́t Jelı́nek, and Jan Kára. Noncrossing Hamiltonian paths in
geometric graphs. Discrete Applied Mathematics, 155(9):1096–1105, 2007. doi:10.1016/j.
dam.2005.12.010.

[18] Adrian Dumitrescu, Dániel Gerbner, Balázs Keszegh, and Csaba D. Tóth. Covering paths for
planar point sets. Discrete & Computational Geometry, 51(2):462–484, 2014. doi:10.1007/
s00454-013-9563-4.

[19] Adrian Dumitrescu, János Pach, and Géza Tóth. Drawing Hamiltonian cycles with no large
angles. The Electronic Journal of Combinatorics, 19(2):P31, 2012. doi:10.37236/2356.

[20] Adrian Dumitrescu and Csaba D. Tóth. Long non-crossing configurations in the plane. Discrete
& Computational Geometry, 44(4):727–752, 2010. doi:10.1007/s00454-010-9277-9.

18
[21] Sándor P. Fekete. Simplicity and hardness of the maximum traveling salesman problem under
geometric distances. In Proc. 10th ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 337–345, 1999. URL: http://dl.acm.org/citation.cfm?id=314500.314586.

[22] Sándor P. Fekete and Gerhard J. Woeginger. Angle-restricted tours in the plane. Computa-
tional Geometry: Theory and Applications, 8:195–218, 1997. doi:10.1016/S0925-7721(96)
00012-0.

[23] Michael R. Garey, Ronald L. Graham, and David S. Johnson. Some NP-complete geometric
problems. In Proceedings of the 8th Annual ACM Symposium on Theory of Computing (STOC),
pages 10–22. ACM, 1976. doi:10.1145/800113.803626.

[24] Ming-Yang Kao and Manan Sanghi. An approximation algorithm for a bottleneck traveling
salesman problem. J. Discrete Algorithms, 7(3):315–326, 2009. doi:10.1016/J.JDA.2008.
11.007.

[25] Joseph S. B. Mitchell. Guillotine subdivisions approximate polygonal subdivisions: A simple

polynomial-time approximation scheme for geometric TSP, k-MST, and related problems.
SIAM J. Comput., 28(4):1298–1309, 1999. doi:10.1137/S0097539796309764.

[26] János Pach, Natan Rubin, and Gábor Tardos. Planar point sets determine many pairwise
crossing segments. Advances in Mathematics, 386:107779, 2021. doi:10.1016/j.aim.2021.
107779.

[27] Christos H. Papadimitriou. The Euclidean traveling salesman problem is NP-complete. The-
oretical Computer Science, 4(3):237–244, 1977. doi:10.1016/0304-3975(77)90012-3.

[28] Clifford Stein and David P. Wagner. Approximation algorithms for the minimum bends trav-
eling salesman problem. In Proceedings of the 8th International Conference on Integer Pro-
gramming and Combinatorial Optimization (IPCO), volume 2081 of LNCS, pages 406–422.
Springer, 2001. doi:10.1007/3-540-45535-3_32.

[29] Helge Tverberg. A seperation property of plane convex sets. Mathematica Scandinavica,
45:255–260, 1979. doi:10.7146/math.scand.a-11840.

19