Journal of Computational and Applied Mathematics 457 (2025) 116259

Contents lists available at ScienceDirect

Journal of Computational and Applied Mathematics

journal homepage: [Link]/locate/cam

Revisiting the region determined by Spearman’s 𝜌 and Spearman’s

footrule 𝜙
Marco Tschimpke a , Manuela Schreyer b , Wolfgang Trutschnig a ,∗
a University of Salzburg, Department for Artificial Intelligence & Human Interfaces, Hellbrunnerstrasse 34, Salzburg, 5020, Salzburg, Austria
b AMAG Austria Metall GmbH, Lamprechtshausenerstr 61, Ranshofen, 5282, Upper Austria, Austria

ARTICLE INFO ABSTRACT

MSC: Kokol and Stopar (2023) recently studied the exact region 𝛺𝜙,𝜌 determined by Spearman’s
62H20 footrule 𝜙 and Spearman’s 𝜌 and derived a sharp lower, as well as a non-sharp upper bound for 𝜌
62H05 given 𝜙. Considering that the proofs for establishing these inequalities are novel and interesting,
60J35
but technically quite involved we here provide alternative simpler proofs mainly building upon
Keywords: shuffles, symmetry, denseness and mass shifting. As a by-product of these proofs we derive
Copulas several additional results on shuffle rearrangements and the interplay between diagonal copulas
Concordance
and shuffles which are of independent interest. Moreover we finally show that we can get closer
Shuffle
to the (non-sharp) upper bound than established in the literature so far.
Markov kernel
Optimization

1. Introduction

A standard approach for quantifying the extent of concordance or, more generally, association of a pair (𝑋, 𝑌 ) of random variables
𝑋, 𝑌 is to consider different (weak) measures of concordance or association such as Spearman’s 𝜌, Kendall’s 𝜏, Gini’s 𝛾, Spearman’s
footrule 𝜙 or Blomqvist’s 𝛽 (see [1,2]). Each of the just mentioned measures only depends on the dependence structure of (𝑋, 𝑌 ), so
in the case of continuous marginals all these measures are functions of the (unique) copula 𝐶 underlying (𝑋, 𝑌 ). Given two (weak)
measures of concordance 𝜅1 and 𝜅2 a seemingly natural question is, how much the value of 𝜅2 can vary given the value of 𝜅1 , or
vice versa. In other words: one might naturally be interested in determining the region

𝛺𝜅1 ,𝜅2 ∶= {(𝜅1 (𝐶), 𝜅2 (𝐶)) ∶ 𝐶 ∈ },

where  denotes the family of all bivariate copulas. The larger the portion of the rectangle

{𝜅1 (𝐶) ∶ 𝐶 ∈ } × {𝜅2 (𝐶) ∶ 𝐶 ∈ }

covered by 𝛺𝜅1 ,𝜅2 , the more different the measures 𝜅1 , 𝜅2 may be considered.
The presumably most well known question in this context was, whether the inequality for Kendall’s 𝜏 and Spearman’s 𝜌 as
established by Durbin and Stuart in [3] is sharp. This very question and some related ones were answered in [4], where 𝛺𝜏,𝜌 was
characterized and shown to be compact but not convex. Since then, various contributions have followed: Article [5] studies the
interrelations between Kendall’s 𝜏 and Gini’s 𝛾 / Spearman’s footrule 𝜙. The lower and upper bound for Spearman’s 𝜌 (Gini’s 𝛾)
given Spearman’s footrule 𝜙 were established in [6,7]. Finally, [2,8] cover the relations between Blomqvist’s 𝛽 and all remaining
(weak) measures of concordance (see Table 1 for a quick overview).

∗ Corresponding author.
E-mail address: wolfgang@[Link] (W. Trutschnig).

[Link]
Received 30 March 2024; Received in revised form 3 August 2024
Available online 7 September 2024
0377-0427/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license
([Link]
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

Table 1
Already studied pairs of mea-
sures of (weak) concordance,
with the following nomencla-
ture:
✓...exactly known region,
−...partially known region,
×...unknown region.
𝜏 𝛾 𝜙 𝛽
𝜌 ✓ × − ✓
𝜏 ✓ ✓ ✓
𝛾 ✓ ✓
𝜙 ✓

In what follows we focus on 𝛺𝜙,𝜌 , the region determined by Spearman’s footrule 𝜙 and Spearman’s 𝜌. According to [6] the
inequality
2√ 2
3 (1 + 2𝜙(𝐶))3∕2 − 1 ≤ 𝜌(𝐶) ≤ 1 − (𝜙(𝐶) − 1)2 (1)
9 3
holds for every copula 𝐶. Furthermore (again see [6]), the lower bound in Eq. (1) is sharp while the upper bound is only known to
be sharp in countably many points (with only accumulation point (1, 1)). Main objective of our contribution is to show that both,
the lower and the upper inequality in Eq. (1) can be established alternatively by proceeding similarly as in [4]. In fact, working
with shuffles, symmetry and continuity, proving the right hand-side of (1) boils down to a straightforward application of classical
Cauchy–Schwarz inequality (which, in turn, even provides a simple characterization for those shuffles, for which the inequality
becomes an equality); and proving the lower (sharp) bound to a rearrangement property of integrals/sums (which holds in a very
general setting). Apart from providing alternative simple proofs for Eq. (1) we also show that we can get closer to the upper bound
than established in [6]. In other words, our main contributions here are the following: (i) Alternative methods for re-establishing
the known inequalities (1) - a simple application of Cauchy–Schwarz for the upper inequality and a seemingly novel observation on
rearrangements of 𝐿2 -functions on [0, 1], which might also prove useful in the context of other problems/inequalities. (ii) Getting
closer to the upper bound than established so far in the literature.
The rest of the contribution is organized as follows: Section 2 introduces the necessary notation and preliminaries used in the
sequel. Section 3 provides two alternative simple proofs for the upper bound: one working with symmetric shuffles and the other
one building upon maximality of diagonal copulas within the family of all symmetric copulas with given diagonal, and the interplay
between diagonal copulas and symmetric shuffles. Section 4 revisits the lower sharp inequality and derives it via the afore-mentioned
novel rearrangement idea. Finally, working with ordinal sums and ‘interpolations’ of copulas, Section 5 extends the known subset
of 𝛺𝜙,𝜌 . Several examples and graphics illustrate the main ideas and chosen approaches.

2. Notation and preliminaries

Given an arbitrary metric space (𝑆, 𝑑), the Borel 𝜎-field on 𝑆 will be denoted by (𝑆). Moreover, the one- and two dimensional
Lebesgue measure (on (R) and (R2 ), respectively) will be denoted by 𝜆 and 𝜆2 , respectively. The class of all bivariate copulas is
denoted by , for every 𝐶 ∈  we will let 𝜇𝐶 denote the corresponding doubly stochastic measure. Prominent examples of copulas
are the independent copula 𝛱 and the lower and upper Frèchet-Hoeffding bounds 𝑊 and 𝑀. In the following 𝐶 𝑡 will denote the
transpose of a copula 𝐶, i.e., 𝐶 𝑡 (𝑢, 𝑣) = 𝐶(𝑣, 𝑢) for all 𝑢, 𝑣 ∈ [0, 1]. A copula 𝐶 is called symmetric if 𝐶 = 𝐶 𝑡 holds. The uniform metric
𝑑∞ on  is defined by

𝑑∞ (𝐶, 𝐷) ∶= max |𝐶(𝑢, 𝑣) − 𝐷(𝑢, 𝑣)|.

(𝑢,𝑣)∈[0,1]2

It is well known that (, 𝑑∞ ) is a compact metric space (see [9]). For every measurable function 𝑓 ∶ R → R we set ‖𝑓 ‖∞ ∶=
sup𝑥∈R |𝑓 (𝑥)|. For background on copulas and doubly stochastic measures we refer to [2,9].
A mapping 𝐾 ∶ R×(R) → [0, 1] is a called a Markov kernel if 𝑥 ↦ 𝐾(𝑥, 𝐹 ) is measurable for every set 𝐹 ∈ (R) and 𝐹 ↦ 𝐾(𝑥, 𝐹 )
is a probability measure for every 𝑥 ∈ R. Given two random variables 𝑋 and 𝑌 on a probability space (𝛺, (𝛺), P) a Markov kernel
is called regular conditional distribution of 𝑌 given 𝑋 if
( )
𝐾(𝑋(𝜔), 𝐹 ) = E 𝟏𝐹 ◦𝑌 |𝑋 (𝜔)

holds for every 𝐹 ∈ (R) and P-almost every 𝜔 ∈ 𝛺. It is well known that for every pair (𝑋, 𝑌 ) the Markov kernel 𝐾(𝑥, ⋅) is unique
for P𝑋 -almost every 𝑥 ∈ R. In the sequel we will write (𝑈 , 𝑉 ) ∼ 𝐶 if 𝐶 is the distribution function of (𝑈 , 𝑉 ) and 𝑈 , 𝑉 are uniform
on [0, 1]. For every 𝐶 ∈  there exists a Markov kernel 𝐾𝐶 satisfying the disintegration property

𝜇𝐶 (𝐺) = 𝐾𝐶 (𝑢, 𝐺𝑢 )𝑑𝜆(𝑢)

∫[0,1]

for every 𝐺 ∈ ([0, 1]2 ), where 𝐺𝑢 ∶= {𝑣 ∈ [0, 1] ∶ (𝑢, 𝑣) ∈ 𝐺}. For more information on disintegration and conditional expectations
see [10,11]; for more background on Markov kernels and their applications in the context of copulas we refer to [12–14].

2
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

A measurable transformation ℎ ∶ [0, 1] → [0, 1] is called 𝜆-preserving if 𝜆ℎ (𝐸) ∶= 𝜆(ℎ−1 (𝐸)) = 𝜆(𝐸) holds for every 𝐸 ∈ ([0, 1]),
i.e., if the push forward 𝜆ℎ of 𝜆 via ℎ coincides with 𝜆. A copula 𝐶 is said to be completely dependent if there exists a 𝜆-preserving
transformation ℎ ∶ [0, 1] → [0, 1] such that 𝐾𝐶 (𝑥, 𝐹 ) = 𝟏𝐹 (ℎ(𝑥)) is a version of the Markov kernel of 𝐶. In other words: A copula
is called completely dependent if it allows a Markov kernel whose conditional distributions are all degenerated. For alternative
equivalent definitions of complete dependence we refer to [12] and the references therein. For every 𝜆-preserving transformation
ℎ ∶ [0, 1] → [0, 1] we will let 𝐶ℎ denote the corresponding (unique) copula and write 𝑑 for the class of all completely dependent
copulas. If a copula 𝐶 fulfills 𝐶 ∈ 𝑑 and 𝐶 𝑡 ∈ 𝑑 we will refer to it as mutually completely dependent. It is straightforward to verify
that the latter is the case if, and only if the corresponding 𝜆-preserving transformation is bijective outside a set of 𝜆-measure zero.
We call a 𝜆-preserving transformation ℎ ∶ [0, 1] → [0, 1] a (classical) equidistant even shuffle (a.k.a. equidistant shuffle of 𝑀)
with 𝑁 ∈ N stripes if, and only if ℎ is linear with slope 1 on each interval 𝐼𝑁 𝑖 ∶= ( 𝑖−1 , 𝑖 ), injective on ⋃𝑁 ( 𝑖−1 , 𝑖 ), and just
𝑁 𝑁 𝑖=1 𝑁 𝑁
1 𝑁
permutes the intervals 𝐼𝑁 , … , 𝐼𝑁 . In the sequel 𝛴𝑁 will denote the set of all permutations of the set {1, … , 𝑁} and 𝑁 the family of
all equidistant even shuffles with 𝑁 stripes. It is well-known and straightforward to check that for every 𝑁 ∈ N there is a one-to-one
correspondence between 𝛴𝑁 and 𝑁 (see [15]). Emphasizing the permutation we will therefore frequently write 𝑆𝜋 ∈ 𝑁 .
In what follows we will work with the subclass of symmetric shuffles: A shuffle 𝑆𝜋 ∈ 𝑁 is called symmetric if, and only if the
corresponding completely dependent copula 𝐶𝑆𝜋 ∈ 𝑑 is symmetric. It is straightforward to verify that 𝑆𝜋 ∈ 𝑁 is symmetric if, and
only if the corresponding permutation 𝜋 ∈ 𝛴𝑁 is an involution (a.k.a. self inverse), i.e., if 𝜋(𝜋(𝑖)) = 𝑖 holds for every 𝑖 ∈ {1, … , 𝑁}. In
accordance with shuffles in the sequel we will simply refer to self inverse permutations as symmetric. The subclass of all symmetric
𝑠𝑦𝑚
elements of 𝑁 will be denoted by 𝑁 . Furthermore, to simplify notation we will write
⋃ ⋃ 𝑠𝑦𝑚
𝑠𝑦𝑚
 ∶= 𝑁 ,  ∶= 𝑁
𝑁∈N 𝑁∈N
and refer to elements of  (or  𝑠𝑦𝑚 ) as shuffles (or symmetric shuffles). As commonly done in the literature we will also refer to
the corresponding mutually completely dependent copula 𝐶ℎ as shuffle (or symmetric shuffle) and write
{ } { }
 ∶= 𝐶ℎ ∶ ℎ ∈  ,  𝑠𝑦𝑚 ∶= 𝐶ℎ ∶ ℎ ∈  𝑠𝑦𝑚 . (2)
According to [16] a mapping 𝜅 ∶  → R is called a measure of concordance if it satisfies the following properties:

(i) 𝜅(𝑀) = 1;
(ii) 𝜅(𝐶 𝑡 ) = 𝜅(𝐶) for all 𝐶 ∈ ;
(iii) 𝜅(𝐶 𝜈 ) = −𝜅(𝐶) for all 𝐶 ∈  where 𝐶 𝜈 is the reflection of 𝐶 at 𝑢 = 21 , i.e., 𝐶 𝜈 (𝑢, 𝑣) ∶= 𝑣 − 𝐶(1 − 𝑢, 𝑣) for all 𝑢, 𝑣 ∈ [0, 1];
(iv) 𝜅(𝐶) ≤ 𝜅(𝐷) whenever 𝐶 ≤ 𝐷, i.e., whenever 𝐶 and 𝐷 are ordered pointwise;
(v) lim𝑛→∞ 𝜅(𝐶𝑛 ) = 𝜅(𝐶) for any sequence (𝐶𝑛 )𝑛∈N of copulas converging to 𝐶 ∈ .

A mapping 𝜅 ∶  → R only satisfying (i), (ii), (iv) and (v) is called a weak measure of concordance. It is well-known (see [2]) that
Spearman’s 𝜌 and Spearman’s footrule 𝜙 can be expressed in terms of the underlying copula as follows:

𝜌(𝐶) = 12 𝐶(𝑢, 𝑣)𝑑𝜇𝛱 (𝑢, 𝑣) − 3 = 12 𝐶(𝑢, 𝑣)𝑑𝜆2 (𝑢, 𝑣) − 3, (3)

∫[0,1]2 ∫[0,1]2

𝜙(𝐶) = 6 𝐶(𝑢, 𝑣)𝑑𝜇𝑀 (𝑢, 𝑣) − 2 = 6 𝐶(𝑢, 𝑢)𝑑𝜆(𝑢) − 2. (4)

∫[0,1]2 ∫[0,1]

Considering 𝜙(𝑀 𝜈 ) = 𝜙(𝑊 ) = − 21 ≠ −1 = −𝛷(𝑀) Spearman’s footrule 𝜙 is only a weak measure of concordance. To simplify
notation, for 𝑆𝜋 ∈ 𝑁 we will also write 𝜌(𝑆𝜋 ) ∶= 𝜌(𝐶𝑆𝜋 ) as well as 𝜙(𝑆𝜋 ) ∶= 𝜙(𝐶𝑆𝜋 ) in the sequel.
Finally, denoting the diagonal of a copula 𝐶 by 𝛿𝐶 , i.e., 𝛿𝐶 (𝑡) ∶= 𝐶(𝑡, 𝑡) for every 𝑡 ∈ [0, 1], it is well known (see [2,17]) that 𝛿𝐶
satisfies the following properties:

• 𝛿𝐶 (0) = 0 and 𝛿𝐶 (1) = 1,

• 𝛿𝐶 is non-decreasing,
• 𝛿𝐶 is Lipschitz continuous with Lipschitz constant 𝐿 = 2 and
• 𝛿𝐶 (𝑡) ≤ 𝑡 for all 𝑡 ∈ [0, 1].

In the sequel  denotes the family of all diagonals of copulas (which is well known to coincide with the class of all functions
𝛿 ∶ [0, 1] → [0, 1] fulfilling the afore-mentioned four points). For every 𝛿 ∈ , setting 𝛿(𝑡) ̂ ∶= 𝑡 − 𝛿(𝑡) for all 𝑡 ∈ [0, 1] it follows that
both 𝛿 and 𝛿̂ are differentiable 𝜆-almost everywhere (see [18]), hence there exist some measurable functions 𝑤𝛿 ∶ [0, 1] → [0, 2]
and 𝑤̂ 𝛿 ∶ [0, 1] → [−1, 1] with 𝑤𝛿 (𝑥) = 𝛿 ′ (𝑥) as well as 𝑤̂ 𝛿 (𝑡) = 1 − 𝛿̂′ (𝑡) for 𝜆-almost every 𝑡 ∈ [0, 1]. We will refer to 𝑤𝛿 and 𝑤̂ 𝛿 as
measurable versions of the derivative of 𝛿 and 𝛿, ̂ respectively.

3. Novel proofs for the upper bound

We first tackle the upper bound

2
𝜌(𝐶) ≤ 1 − (𝜙(𝐶) − 1)2 (5)
3
going back to [6] and established by working with diagonal copulas, sufficiently smooth diagonals and Bernstein approximations
(the technically quite involved Lemma 9 being key). We provide two simple alternative proofs, one purely based on Cauchy–Schwarz
inequality, and the other one using maximality properties of diagonal copulas.

3
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

3.1. A simple proof via symmetric shuffles and Cauchy–Schwarz inequality

Building upon the fact that for every copula 𝐶 the symmetric copula 𝐶 ∗ ∶= 12 (𝐶 + 𝐶 𝑡 ) fulfills 𝜌(𝐶 ∗ ) = 𝜌(𝐶) and 𝜙(𝐶 ∗ ) = 𝜙(𝐶), it
suffices to prove Eq. (5) for the class  sym of all symmetric copulas. Moreover, using continuity of 𝜌 and 𝜙 w.r.t 𝑑∞ we can further
reduce the problem to any dense subclass of the family  sym . The following lemma will therefore be key.

Lemma 3.1. The family of all symmetric shuffles  𝑠𝑦𝑚 is dense in ( sym , 𝑑∞ ).

Proof. Looking into the proof of the corresponding result for the full class (, 𝑑∞ ) in [15] reveals the fact that, starting with an
arbitrary symmetric copula 𝐶 the constructed approximating shuffle is symmetric as well. In other words: the original proof directly
yields Lemma 3.1. □
𝑠𝑦𝑚
Next we derive handy formulas for Spearman’s 𝜌 and Spearman’s footrule 𝜙 for 𝐶𝑆𝜋 with 𝑆𝜋 ∈ 𝑁 and work with the following
sets:

𝐼𝜋− ∶= {𝑖 ∈ {1, … , 𝑁} ∶ 𝜋(𝑖) < 𝑖}, 𝐼𝜋0 ∶= {𝑖 ∈ {1, … , 𝑁} ∶ 𝜋(𝑖) = 𝑖},

𝐼𝜋+ ∶= {𝑖 ∈ {1, … , 𝑁} ∶ 𝜋(𝑖) > 𝑖}. (6)
𝑠𝑦𝑚
We will only write 𝐼 − , 𝐼 0 and 𝐼 + whenever no confusion can arise. Notice that symmetry of 𝑆𝜋 ∈ 𝑁 implies that (i) 𝑖 ∈ 𝐼 − if, and
only if 𝜋(𝑖) ∈ 𝐼 + and that (ii) 𝑖 ∈ 𝐼 + if, and only if 𝜋(𝑖) ∈ 𝐼 − .

𝑠𝑦𝑚
Lemma 3.2. For every 𝑁 ∈ N and 𝑆𝜋 ∈ 𝑁 the following identities hold:
( )2
12 ∑ 𝑖 − 𝜋(𝑖)
𝜌(𝑆𝜋 ) = 1 − , (7)
𝑁 𝑖∈𝐼 − 𝑁
6 ∑ 𝑖 − 𝜋(𝑖)
𝜙(𝑆𝜋 ) = 1 − . (8)
𝑁 𝑖∈𝐼 − 𝑁

Proof. Fix 𝑁 ∈ N and 𝑆𝜋 ∈ 𝑁 . Then obviously (a version of) the Markov kernel is given by 𝐾𝑆𝜋 (𝑥, 𝐹 ) = 𝟏𝐹 (𝑆𝜋 (𝑥)) with
( )
∑
𝑁
𝑖 − 𝜋(𝑖)
𝑆𝜋 (𝑥) = 𝑥− 𝟏𝐼 𝑖 (𝑥).
𝑖=1
𝑁 𝑁
⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟⏞⏞⏞⏞⏞⏞⏞⏞⏟
=∶ℎ𝑖 (𝑥)

Using disintegration therefore yields (for a justification for the interchange of the order of integration see, e.g., [16])
𝜙(𝑆𝜋 ) + 2
= 𝐶 (𝑢, 𝑣)𝑑𝜇𝑀 (𝑢, 𝑣) = 𝑀(𝑢, 𝑣)𝑑𝜇𝐶𝑆 (𝑢, 𝑣)
6 ∫[0,1]2 𝑆𝜋 ∫[0,1]2 𝜋

∑
𝑁
= 𝑀(𝑢, 𝑆𝜋 (𝑢))𝑑𝜆(𝑢) = 𝑀(𝑢, ℎ𝑖 (𝑢))𝑑𝜆(𝑢)
∫[0,1] ∫𝐼 𝑖
𝑖=1 𝑁
∑ ∑ 𝑖 − 𝜋(𝑖)
= 𝑢𝑑𝜆(𝑢) + 𝑢− 𝑑𝜆(𝑢)
∫𝐼 𝑖 ∫𝐼 𝑖 𝑁
𝑖∶𝜋(𝑖)≥𝑖 𝑁 𝑖∶𝜋(𝑖)<𝑖 𝑁

∑
𝑁
1 ∑ 𝑖 − 𝜋(𝑖)
= 𝑢𝑑𝜆(𝑢) −
∫𝐼 𝑖 𝑁 𝑁
𝑖=1 𝑁 𝑖∶𝜋(𝑖)<𝑖
1 1 ∑ 𝑖 − 𝜋(𝑖)
= − .
2 𝑁 𝑖∈𝐼 − 𝑁

Proceeding analogously for Spearman’s 𝜌 we get

𝜌(𝑆𝜋 ) + 3
= 𝐶 (𝑢, 𝑣)𝑑𝜇𝛱 (𝑢, 𝑣) = 𝛱(𝑢, 𝑣)𝑑𝜇𝐶𝑆 (𝑢, 𝑣)
12 ∫[0,1]2 𝑆𝜋 ∫[0,1]2 𝜋

∑
𝑁
= 𝛱(𝑢, 𝑆𝜋 (𝑢))𝑑𝜆(𝑢) = 𝛱(𝑢, ℎ𝑖 (𝑢))𝑑𝜆(𝑢)
∫[0,1] ∫𝐼 𝑖
𝑖=1 𝑁

∑
𝑁
𝑖 − 𝜋(𝑖)
= 𝑢2 − 𝑢 𝑑𝜆(𝑢)
∫𝐼 𝑖 𝑁
𝑖=1 𝑁

∑𝑁
𝑖 − 𝜋(𝑖) 𝑖2 − (𝑖 − 1)2
= 𝑢2 𝑑𝜆(𝑢) − ⋅
∫[0,1] 𝑁 2𝑁 2
𝑖=1

1 ∑ 𝑖 − 𝜋(𝑖) 2𝑖 − 1
𝑁
= − ⋅ .
3 𝑖=1 𝑁 2𝑁 2

4
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

∑𝑁
Since 𝜋 ∈ 𝛴𝑁 we have 𝑖=1 (𝑖 − 𝜋(𝑖)) = 0, which altogether yields

∑𝑁
𝑖 − 𝜋(𝑖) 2𝑖 − 1 12 ∑ 𝑖(𝑖 − 𝜋(𝑖))
𝑁
𝜌(𝑆𝜋 ) = 1 − 6 ⋅ =1− .
𝑁 𝑁 2 𝑁 𝑖=1 𝑁2
𝑖=1

Finally, using symmetry of 𝑆𝜋 the last summand simplifies to

∑
𝑁 ∑ ∑ ∑
𝑖(𝑖 − 𝜋(𝑖)) = 𝑖(𝑖 − 𝜋(𝑖)) + 𝑖(𝑖 − 𝜋(𝑖)) + 𝑖(𝑖 − 𝜋(𝑖))
𝑖=1 𝑖∈𝐼 − 𝑖∈𝐼 0 𝑖∈𝐼 +
∑ ∑
= 𝑖(𝑖 − 𝜋(𝑖)) + 𝜋(𝑗)(𝜋(𝑗) − 𝜋(𝜋(𝑗)))
𝑖∈𝐼 − 𝑗∈𝐼 −
∑( ) ∑
= 𝑖(𝑖 − 𝜋(𝑖)) + 𝜋(𝑖)(𝜋(𝑖) − 𝑖) = (𝑖 − 𝜋(𝑖))2 ,
𝑖∈𝐼 − 𝑖∈𝐼 −

which completes the proof. □

According to Lemma 3.2, the values of the permutation 𝜋 on the set 𝐼 − contain all relevant information for calculating 𝜌 and
𝜙. This simple observation opens the door for applying Cauchy–Schwarz inequality and deriving a very simple proof for the upper
inequality (#𝐼 − denoting the cardinality of 𝐼 − ):

𝑠𝑦𝑚
Theorem 3.3. For every 𝑁 ∈ N and every symmetric shuffle 𝑆𝜋 ∈ 𝑁 the inequality
2( )2
𝜌(𝑆𝜋 ) ≤ 1 − 1 − 𝜙(𝑆𝜋 ) (9)
3
𝑁
holds. Furthermore we have equality in (9) if, and only if 𝑖 ↦ 𝜋(𝑖) − 𝑖 is constant on 𝐼 − and #𝐼 − = 2
holds.

Proof. Since the result is obvious for 𝐼 − = ∅ it suffices to consider 𝐼 − ≠ ∅. Cauchy–Schwarz inequality implies
( ( ) )1∕2
∑ 𝑖 − 𝜋(𝑖) ∑ 𝑖 − 𝜋(𝑖) 2
− 1∕2
≤ (#𝐼 ) , (10)
𝑖∈𝐼 −
𝑁 𝑖∈𝐼 −
𝑁

so using Eqs. (7) and (8) shows

( )2 ( )
1 − 𝜙(𝑆𝜋 ) 1 − 𝜌(𝑆𝜋 ) 𝑁 1 − 𝜌(𝑆𝜋 )
𝑁 ≤ #𝐼 − 𝑁 ≤ ⋅ 𝑁, (11)
6 12 2 12
( )2
which directly yields the desired inequality 𝜌(𝑆𝜋 ) ≤ 1 − 23 1 − 𝜙(𝑆𝜋 ) .
𝑖−𝜋(𝑖)
Turning toward sharpness, obviously Eq. (10) is sharp if, and only if the mapping 𝑖 ↦ 𝑁
is constant on 𝐼 − . Furthermore the
second part of Eq. (11) is sharp if, and only if #𝐼 − = 𝑁2 . □

For an even 𝑛 ∈ N let 𝑆𝜋 ∗ ∈ 𝑛𝑠𝑦𝑚 denote the shuffle corresponding to 𝜋 ∗ ∈ 𝛴𝑛 fulfilling 𝑖 − 𝜋 ∗ (𝑖) = 1 for every 𝑖 ∈ 𝐼 − = {𝑖 ∈
∗
{1, … , 𝑛} ∶ 𝑖 even}; for the case 𝑛 = 6 see Fig. 1. Then obviously the mapping 𝑖 ↦ 𝑖−𝜋𝑁 (𝑖) is constant on 𝐼 − , so we already know from
Eq. (9) that Eq. (9) becomes an equality (compare with Example 12 in [6]).

Fig. 1. The shuffle 𝑆𝜋 ∗ ∈ 6𝑠𝑦𝑚 for which Eq. (9) becomes an equality.

Combining Theorem 3.3 and Lemma 3.1 already yields the following inequality for all copulas:

Theorem 3.4 ([6]). For every copula 𝐶 ∈  the following inequality holds:
2
𝜌(𝐶) ≤ 1 − (1 − 𝜙(𝐶))2 .
3

5
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

3.2. A second alternative proof via diagonal copulas and their interrelation with shuffles

It is well-known (see [17,19,20]) that, given a diagonal 𝛿 ∈ , the diagonal copula 𝐸𝛿 , given by
{ 𝛿(𝑢) + 𝛿(𝑣) }
𝐸𝛿 (𝑢, 𝑣) = min 𝑢, 𝑣,
2
for all 𝑢, 𝑣 ∈ [0, 1] is the maximal element in the class of all symmetric copulas with diagonal 𝛿, i.e., 𝐶 ≤ 𝐸𝛿 for all copulas 𝐶 ∈  sym
with diagonal 𝛿 ∈ .
Before proceeding with the alternative proof of Theorem 3.4 we recall some properties of diagonal copulas going back to [17],
which will prove useful in the sequel. Obviously the mapping 𝜄 ∶ (, ‖ ⋅ ‖∞ ) → (, 𝑑∞ ), defined by 𝜄(𝛿) = 𝐸𝛿 is continuous. Moreover,
for a given diagonal 𝛿 ∈  set 𝑔(𝑡) ∶= 2𝑡 − 𝛿(𝑡) and define

𝐿(𝑡) ∶= min{𝑧 ∈ [0, 1] ∶ 𝑔(𝑧) ≥ 𝛿(𝑡)} and 𝑈 (𝑡) ∶= min{𝑧 ∈ [0, 1] ∶ 𝛿(𝑧) ≥ 𝑔(𝑡)}.

It is straightforward to verify that both 𝐿 and 𝑈 are non-decreasing and that 𝐿(𝑡) ≤ 𝑡 ≤ 𝑈 (𝑡) holds for every 𝑡 ∈ [0, 1]. Furthermore, it
can be shown that the diagonal copula 𝐸𝛿 distributes its mass on the graphs of the functions 𝐿 and 𝑈 . More precisely, the following
result holds:

Proposition 3.5 ([17]). Suppose that 𝛿 ∈  and let 𝑤𝛿 denote a measurable version of its derivative. Then (a version of) the Markov
kernel 𝐾𝐸𝛿 of 𝐸𝛿 is given by
( )
𝑤 (𝑡) 𝑤 (𝑡)
𝐾𝐸𝛿 (𝑡, 𝐹 ) = 𝛿 𝟏𝐹 (𝐿(𝑡)) + 1 − 𝛿 𝟏𝐹 (𝑈 (𝑡)).
2 2
Moreover, 𝐸𝛿 is (mutually) completely dependent and concentrates its mass on the graph of a 𝜆-preserving bijection ℎ ∶ [0, 1] → [0, 1]
fulfilling ℎ◦ℎ = 𝑖𝑑[0,1] if, and only if for 𝜆-almost every 𝑥 ∈ [0, 1] either 𝛿 ′ (𝑥) ∈ {0, 2} or 𝛿(𝑥) = 𝑥 holds.

For every 𝑁 ∈ N we will let 0,2 𝑁

denote the family of all diagonals 𝛿 ∈  fulfilling that on each open interval 𝐼𝑁 𝑖 we either have

that (i) 𝛿 ′ (𝑥) = 0 for all 𝑥 ∈ 𝐼𝑁𝑖 or that (ii) 𝛿 ′ (𝑥) = 2 for all 𝑥 ∈ 𝐼 𝑖 . Fig. 2 depicts an example of such a diagonal. Obviously 0,2 = ∅
𝑁 ⋃ 𝑁
for odd 𝑁 ∈ N and 0,2 𝑁
≠ ∅ for even 𝑁 ∈ N. To simplify notation set 0,2 ∞ ∶=
0,2
𝑁∈N 𝑁 .
Using this notation Proposition 3.5 opens the door to the following alternative idea of proof for Eq. (5), which we will now
tackle step by step:
(Step 1) Show that the family 0,2 ∞ is dense in (, ‖.‖∞ ).
(Step 2) Show that for every 𝛿 ∈ 0,2 𝑁
the diagonal copula 𝐸𝛿 is a symmetric, equidistant even shuffle, i.e., 𝐸𝛿 ∈  𝑠𝑦𝑚 .
(Step 3) Use the maximality property of diagonal copulas (mentioned before) and apply Theorem 3.3.

Lemma 3.6. 0,2

∞ is dense in (, ‖.‖∞ ).

Proof. Let 𝛿 ∈  and 𝑁 ∈ N be arbitrary but fixed. We construct an element 𝛿̃ ∈ 0,2 2𝑁

as follows: For every 𝑖 ∈ {1, … , 2𝑁} consider
𝑖
𝑦𝑖 ∶= 𝛿( 2𝑁 ) and define
{ }
2𝑘 𝑘
𝑖𝑘 ∶= min 𝑖 ∈ {1, … , 2𝑁} ∶ 𝑦𝑖 ≥ =
2𝑁 𝑁
for every 𝑘 ∈ {1, … , 𝑁}. Setting 𝑖0 ∶= 0 this obviously yields 0 = 𝑖0 < 2 ≤ 𝑖1 < 𝑖2 < ⋯ < 𝑖𝑁−1 < 𝑖𝑁 = 2𝑁. Defining

∑
𝑁
𝑓 (𝑥) ∶= 2 𝟏𝑖 −1 𝑖𝑘 (𝑥),
( 𝑘2𝑁 , 2𝑁 )
𝑘=1

it follows immediately that 𝑓 is a probability density (w.r.t. 𝜆), so the function 𝛿̃ ∶ [0, 1] → [0, 1], defined by

̃ =
𝛿(𝑥) 𝑓 (𝑡) 𝑑𝜆(𝑡)
∫[0,𝑥]
̃ = 0 and 𝛿(1)
obviously is a Lipschitz continuous (Lipschitz constant 𝐿 = 2) piecewise linear distribution function fulfilling 𝛿(0) ̃ = 1.
Furthermore 𝛿̃ ≥ 𝛿𝑊 (otherwise 𝛿(1)
̃ = 1 is impossible) and by construction of 𝑓 we have that
( ) ( )
𝑖𝑘 2𝑘 𝑖𝑘
𝛿̃ = ≤ 𝑦𝑖𝑘 = 𝛿
2𝑁 2𝑁 2𝑁
𝑖 −1 𝑖
holds for every 𝑘, implying 𝛿̃ ≤ 𝛿 on [0, 1] since 𝑓 is 0 outside ∪𝑁 ( 𝑘 , 𝑘 ). Altogether 𝛿̃ ∈ 0,2
𝑘=1 2𝑁 2𝑁 2𝑁
follows.
𝑖1 𝑖2 𝑖𝑁−1 2 1
Finally, considering that on the grid {0, 2𝑁 , 2𝑁 , … , 2𝑁 , 1} the two diagonals have a maximum distance of at most 2𝑁
= 𝑁
,
using monotonicity and the fact that on each interval 𝐼2𝑁 𝑖 the diagonal 𝛿̃ has either slope 0 or 2 it follows that
1
‖𝛿 − 𝛿‖
̃ ∞≤ .
𝑁
1
This completes the proof since for sufficiently large 𝑁 the quantity 𝑁
is smaller than any fixed 𝜀 > 0. □

6
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

We will now clarify under which conditions diagonal copulas are equidistant even shuffles (and vice versa). Doing so we will
call 𝜋 ∈ 𝛴𝑁 bi-monotone if 𝜋 restricted to 𝐼𝜋− is strictly increasing. Notice that if 𝜋 is symmetric and bi-monotone then 𝜋 restricted
to 𝐼𝜋+ is strictly increasing too. The right panel of Fig. 2 depicts a shuffle 𝑆𝜋 corresponding to a bi-monotone symmetric permutation
𝜋 ∈ 𝛴12 .

Fig. 2. Example of a diagonal 𝛿 ∈ 0,2 12

(left panel) and the corresponding diagonal copula/shuffle 𝐸𝛿 = 𝐶𝑆𝜋 with 𝜋 = (3, 5, 1, 6, 2, 4, 8, 7, 11, 12, 9, 10) and
𝐼𝜋− = {3, 5, 6, 8, 11, 12} (right panel).

Theorem 3.7. Suppose that 𝑁 ∈ N is even and that 𝛿 ∈ 0,2 𝑁

. Then the diagonal copula 𝐸𝛿 is an equidistant even shuffle,
i.e., 𝐸𝛿 ∈  𝑠𝑦𝑚 ⊆  𝑠𝑦𝑚 and the corresponding permutation 𝜋 ∈ 𝛴𝑁 is symmetric, bi-monotone and fulfills 𝐼𝜋0 = ∅.
𝑁

Proof. We already know from Proposition 3.5 that under the assumptions of the theorem the corresponding diagonal copula 𝐸𝛿
is mutually completely dependent and that there exists some 𝜆-preserving, bijective ℎ ∶ [0, 1] → [0, 1] fulfilling ℎ◦ℎ = 𝑖𝑑 such that
𝑠𝑦𝑚
𝐸𝛿 = 𝐶ℎ . It remains to show that ℎ ∈ 𝑁 and that the corresponding permutation 𝜋 is bi-monotone (the fact that 𝜋◦𝜋 is a direct
consequence of symmetry of 𝐸𝛿 ). Defining

𝐽 0 ∶= {𝑖 ∈ {1, … , 𝑁} ∶ 𝛿 ′ (𝑥) = 0 for every 𝑥 ∈ 𝐼𝑁

𝑖
},
𝐽 2 ∶= {𝑖 ∈ {1, … , 𝑁} ∶ 𝛿 ′ (𝑥) = 2 for every 𝑥 ∈ 𝐼𝑁
𝑖
},
𝑗
we have 𝐽 0 ∪𝐽 2 = {1, … , 𝑁}. Notice that on every interval 𝐼𝑁 with 𝑗 ∈ 𝐽 0 the function 𝐿 is constant, whereas 𝑈 is constant on every
𝑗
interval 𝐼𝑁 with 𝑗 ∈ 𝐽 2 . Furthermore interpreting 𝛿 and 𝑔 as distribution functions with quasi-inverses 𝛿 − and 𝑔 − , respectively, it
follows that 𝐿 = 𝑔 − ◦𝛿 and that
⋃ 𝑗
𝑔 − ([0, 1]) = 𝐼𝑁
𝑗∈𝐽 0
𝑖 𝑖 . As a direct 𝑗
holds, whereby 𝐼 𝑁 denotes the closure of the open interval 𝐼𝑁 consequence, for every 𝑥 ∈ 𝐼𝑁 with 𝑗 ∈ 𝐽 2 we have that
⋃ 𝑗 𝑗 0 ⋃ 𝑗
𝐿(𝑥) ∈ 𝑗∈𝐽 0 𝐼𝑁 . Proceeding analogously for 𝑈 shows that for every 𝑥 ∈ 𝐼𝑁 with 𝑗 ∈ 𝐽 we have that 𝑈 (𝑥) ∈ 𝑗∈𝐽 2 𝐼𝑁 . Finally
considering at every point of differentiability 𝐿 and 𝑈 can only have slope 0 or 1 by the chain rule it follows that 𝐸𝛿 is indeed an
equidistant even shuffle. Letting 𝜋 ∈ 𝛴𝑁 denote the corresponding permutation we get that 𝜋 maps 𝐽 0 (bijectively) to 𝐽 2 and vice
versa and that 𝐼𝜋− = 𝐽 2 holds. Furthermore, using the fact that 𝐿, 𝑈 are non-decreasing, 𝜋 is strictly increasing on 𝐽 0 and on 𝐽 2 ,
i.e., 𝜋 is bi-monotone.
Finally, considering that according to [17] 𝛿(𝑡) < 𝑡 implies 𝐿(𝑡) < 𝑡 and 𝑈 (𝑡) > 𝑡 and that for 𝛿 ∈ 0,2
𝑁
obviously 𝛿(𝑡) < 𝑡 holds for
all but at most 𝑁 points, we conclude that 𝐼𝜋0 = ∅ and the proof is complete. □
Now Step 3 is obvious - combining Lemma 3.6 and Theorem 3.7 directly completes our second alternative proof for the upper
inequality.
Considering that (to the best of our knowledge) the interplay between diagonal copulas and equidistant even shuffles has not be
studied yet we conclude this section with the converse of Theorem 3.7.

Proposition 3.8. Let 𝑁 ∈ N be even, 𝜋 ∈ 𝛴𝑁 be a symmetric, bi-monotone permutation with 𝐼𝜋0 = ∅, and 𝑆𝜋 denote the corresponding
shuffle. Then the shuffle 𝐶𝑆𝜋 is a diagonal copula.

Proof. If 𝜋 ∈ 𝛴𝑁 fulfills the assumptions of the proposition, then letting 𝑆𝜋 denote the corresponding shuffle we have 𝛿 ∶= 𝛿𝐶𝑆 ∈
𝜋
0,2
𝑁
, so according to Theorem 3.7 the induced diagonal copula 𝐸𝛿 is an equidistant shuffle and the corresponding permutation
𝜋 ∗ is symmetric, bi-monotone and fulfills 𝐼𝜋0∗ = ∅. Furthermore (as shown in the proof of Theorem 3.7) we have 𝐼𝜋−∗ = 𝐽 2 = 𝐼𝜋− .

7
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

Considering the facts that 𝜋 is strictly increasing on 𝐼𝜋− , that 𝜋 ∗ is strictly increasing on 𝐼𝜋−∗ , and that the two sets coincide, the
identity 𝜋 = 𝜋 ∗ follows immediately, and the proof is complete. □

4. A novel proof for the sharp lower bound

In this section we focus on the inequality

2√
𝜌(𝐶) ≥ 3 (1 + 2𝜙(𝐶))3∕2 − 1. (12)
9
In the original paper [6] the authors derived this very equality by first showing it for the subclass of copulas assigning full mass to
the main and second diagonal (i.e., P(𝑋 = 𝑌 ) + P(𝑌 = 1 − 𝑋) = 1 with (𝑋, 𝑌 ) ∼ 𝐶) and then extending the results to the full class.
The chosen method of proof is interesting and novel but at the same time technically quite involved. In the sequel we show that
working with shuffles, another (seemingly novel) mass rearrangement idea, and denseness arguments allows for a shorter and less
technical alternative proof.
We start with some first observations on the lower inequality, motivate the mass rearrangement idea, then prove a much more
general mass rearrangement result for 𝐿2 functions, and finally apply it to derive Eq. (12).
As in the previous section we will work with symmetric permutations 𝜋 ∈ 𝛴𝑁 , the shuffles 𝑆𝜋 and the corresponding sets 𝐼𝜋− . To
simplify notation throughout this section we will write 𝑘 ∶= #𝐼𝜋− ≤ 𝑁2 . For 𝑘 = 0 we obviously have 𝑆𝜋 = 𝑖𝑑[0,1] , implying 𝐶𝑆𝜋 = 𝑀,
which yields equality in (12). It therefore suffices to consider 𝑘 ≥ 1.
𝑖 −𝜋(𝑖 )
For 𝑘 ≥ 1 we will let 0 < 𝑝1 ≤ ⋯ ≤ 𝑝𝑘 denote the order statistics of the points 𝑙 𝑁 𝑙 with 𝑙 ∈ {1, … , 𝑘} and 𝐼𝜋− = {𝑖1 , … , 𝑖𝑘 } ⊂
𝑘
{1, … , 𝑁}, and will write 𝐩 = (𝑝1 , … , 𝑝𝑘 ) ∈ [0, 1] . Using this notation the formulas for Spearman’s 𝜌 and Spearman’s footrule 𝜙 (see
Lemma 3.2) simplify to

12 ∑ 2 6 ∑
𝑘 𝑘
𝜌(𝑆𝜋 ) = 1 − 𝑝 , 𝜙(𝑆𝜋 ) = 1 − 𝑝.
𝑁 𝑖=1 𝑖 𝑁 𝑖=1 𝑖

Hence the desired inequality is equivalent to

( )3∕2
2√ 12 ∑ 12 ∑ 2
𝑘 𝑘
3 3− 𝑝𝑖 −1≤1− 𝑝
9 𝑁 𝑖=1 𝑁 𝑖=1 𝑖

which, in turn, simplifies to

( )3∕2
4 ∑ 6 ∑ 2
𝑘 𝑘
1− 𝑝𝑖 ≤1− 𝑝 .
𝑁 𝑖=1 𝑁 𝑖=1 𝑖

In other words, Eq. (12) boils down to showing that

( )3∕2
6 ∑ 2 4 ∑
𝑘 𝑘
𝑚𝜋,𝑘 (𝐩) ∶= 1 − 𝑝𝑖 − 1 − 𝑝𝑖 ≥0 (13)
𝑁 𝑖=1 𝑁 𝑖=1

for all symmetric permutations 𝜋.

The following example illustrates the idea underlying the rearrangement, which we will work with in this section:

Example 4.1. Consider 𝑁 = 8 and the symmetric permutation 𝜋 ∈ 𝛴8 , given by 𝜋 = (4, 7, 8, 1, 6, 5, 2, 3). Then we have
𝑖 −𝜋(𝑖 ) 𝑖 −𝜋(𝑖 ) 𝑖 −𝜋(𝑖 ) 𝑖 −𝜋(𝑖 )
𝐼𝜋− = {4, 6, 7, 8}, 𝑘 = 4, and 1 𝑁 1 = 38 , 2 𝑁 2 = 18 , 3 𝑁 3 = 58 , 4 𝑁 4 = 85 , so the vector 𝐩 is given by 𝐩 = 18 ⋅ (1, 3, 5, 5). The
∑
left panel of Fig. 3 depicts the shuffle 𝐶𝑆𝜋 , the shaded squares illustrate 𝐩 and 𝐼𝜋− , their total number is 𝛥 = 𝑁 ⋅ 𝑘𝑖=1 𝑝𝑖 = 14. Notice
that 𝑆𝜋̂ in the right panel can be constructed from 𝑆𝜋 by shifting the squares to the right below the main diagonal while preserving
the total number squares 𝛥. More precisely, starting (from the right) with the last stripe we move the maximal possible 𝑁 − 1 = 7
squares to [ 87 , 1] × [0, 78 ], then we shift the maximal possible 𝑁 − 3 = 5 squares to [ 68 , 87 ] × [ 82 , 68 ], and finally we move the remaining
14 − 7 − 5 = 2 squares to [ 85 , 68 ] × [ 48 , 58 ] (see Fig. 3). Hence the permutation 𝜋̂ is defined for 𝑖 ∈ {6, 7, 8} and via symmetry extended to
𝑖 ∈ {1, … , 5}. Altogether in this case we obtain that the symmetric permutation 𝜋̂ is given by 𝜋̂ = (8, 7, 3, 6, 5, 4, 2, 1) ∈ 𝛴8 . Denoting all
∑̂
quantities corresponding to 𝜋̂ with a hat, for 𝜋̂ we have 𝐼𝜋−̂ = {6, 7, 8}, 𝑘̂ = 3, and 𝐩̂ = 81 ⋅(2, 5, 7), implying 𝛥̂ = 𝑁 ⋅ 𝑘𝑖=1 𝑝̂𝑖 = 14 = 𝛥. As
5
a direct consequence, 𝜙(𝑆𝜋 ) = 𝜙(𝑆𝜋̂ ) = − 16 holds, i.e., Spearman’s footrule does not change when moving from 𝑆𝜋 to 𝑆𝜋̂ . Spearman’s
𝜌, however does change, it decreases: in fact, we get 𝜌(𝑆𝜋̂ ) = 1 − 23 ⋅ 39 32
= −0.828125 < −0.40625 = 1 − 32 ⋅ 15
16
= 𝜌(𝑆𝜋 ). In terms of 𝑚𝜋,𝑘
this translates to

̂
𝑚𝜋,𝑘 (𝐩) > 𝑚𝜋,̂ 𝑘̂ (𝐩)

and for showing 𝑚𝜋,𝑘 (𝐩) ≥ 0 it suffices to show 𝑚𝜋,̂ 𝑘̂ (𝐩)

̂ ≥ 0.
The rearrangement from 𝑆𝜋 to 𝑆𝜋̂ can be formalized as follows: Defining the functions 𝑓 ∶ [0, 1] → [0, ∞) and 𝑔 ∶ [0, 1] → R by

1∑ 1∑
4 4
𝑓 (𝑥) = 𝑝 𝟏 𝑖−1 𝑖 (𝑥), 𝑔(𝑥) = 𝑞 𝟏 𝑖−1 𝑖 (𝑥),
2 𝑖=1 𝑖 ( 4 , 4 ] 2 𝑖=1 𝑖 ( 4 , 4 ]

8
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

Fig. 3. The mass rearrangement discussed in Example 4.1. Starting from the shuffle 𝑆𝜋 in the left panel we construct the shuffle 𝑆𝜋̂ for which Spearman’s
footrule is the same but Spearman’s 𝜌 is strictly smaller.

with (𝑞1 , 𝑞2 , 𝑞3 , 𝑞4 ) = 18 ⋅ (1, 1, 0, −2), it follows that 𝑝1 − 𝑞1 = 0 and 𝑝𝑖 − 𝑞𝑖 = 𝑝̂𝑖−1 for 𝑖 ∈ {2, 3, 4}. Obviously 𝑓 is non-decreasing whereas
∑
𝑔 is non-increasing and fulfills ∫[0,1] 𝑔(𝑥)𝑑𝜆(𝑥) = 18 4𝑖=1 𝑞𝑖 = 0. Furthermore, letting ‖ ⋅ ‖2 denote the 𝐿2 -norm with respect to 𝜆 on
([0, 1]) we have that

1 ∑ 2 1 ∑
3 4
𝑝̂𝑖 = (𝑝 − 𝑞𝑖 )2 = ‖𝑓 − 𝑔‖22 ,
𝑁 𝑖=1 𝑁 𝑖=1 𝑖

1 ∑ 2
4
𝑝 = ‖𝑓 ‖22 ,
𝑁 𝑖=1 𝑖

which yields the following equivalence:

1 ∑ 2 1 ∑ 2
3 4
𝑝̂𝑖 ≥ 𝑝 if, and only if ‖𝑓 − 𝑔‖22 ≥ ‖𝑓 ‖22 .
𝑁 𝑖=1 𝑁 𝑖=1 𝑖

The last inequality ‖𝑓 − 𝑔‖22 ≥ ‖𝑓 ‖22 turns out to be a (very) special case of a more general observation on 𝐿2 -norms of (almost)
monotone functions 𝑓 , 𝑔. Considering that the result may also be useful in the context of other problems we formulate and prove it
directly for general finite measure spaces and a non-decreasing function 𝑓 in combination with a rearrangement function 𝑔 fulfilling
the following (much weaker monotonicity) property:

Definition 4.2. A function 𝑔 ∶ [0, 1] → R is called block rearrangement if there exists some 𝑥0 ∈ [0, 1] such that 𝑔(𝑥) ≥ 0 holds for
every 𝑥 ∈ [0, 𝑥0 ) and 𝑔(𝑥) ≤ 0 holds for every 𝑥 ∈ (𝑥0 , 1].

Obviously every non-increasing 𝑔 ∶ [0, 1] → R is a block rearrangement. Motivated by the previous example we will work with
block rearrangements also fulfilling ∫[0,1] 𝑔𝑑𝜇 = 0 for some finite measure 𝜇 on ([0, 1]).

Lemma 4.3. Suppose that 𝜇 is a finite measure on ([0, 1]) and that 𝑓 , 𝑔 ∈ 𝐿2 (𝜇) fulfill the following properties:

1. 𝑓 ∶ [0, 1] → [0, ∞) is non-decreasing.

2. 𝑔 ∶ [0, 1] → R is a block rearrangement fulfilling ∫[0,1] 𝑔𝑑𝜇 = 0.

Then the following inequality holds:

‖𝑓 − 𝑔‖22 ≥ ‖𝑓 ‖22 + ‖𝑔‖22 . (14)

Proof. First of all we obviously have 𝑓 − 𝑔 ∈ 𝐿2 (𝜇) as well as

‖𝑓 − 𝑔‖22 = ‖𝑓 ‖22 + ‖𝑔‖22 − 2 𝑓 𝑔𝑑𝜇,

∫[0,1]
⏟⏞⏞⏞⏟⏞⏞⏞⏟
=∶𝐼

so it suffices to show 𝐼 ≤ 0. We will prove the inequality assuming that 𝑔(𝑥0 ) ≤ 0 - the case 𝑔(𝑥0 ) ≥ 0 can be handled in the same
manner. Using 0 = ∫[0,1] 𝑔𝑑𝜇 = ∫[0,𝑥 ) 𝑔𝑑𝜇 + ∫[𝑥 ,1] 𝑔𝑑𝜇 we have
0 0

𝑔𝑑𝜇 = − 𝑔𝑑𝜇 = (−𝑔) 𝑑𝜇. (15)

∫[0,𝑥0 ) ∫[𝑥0 ,1] ∫[𝑥0 ,1] ⏟⏟⏟
≥0

9
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

Therefore, using monotonicity of 𝑓 it follows that

𝑓 𝑔𝑑𝜇 ≤ 𝑓 (𝑥0 ) 𝑔𝑑𝜇 = 𝑓 (𝑥0 ) (−𝑔)𝑑𝜇 ≤ 𝑓 (−𝑔)𝑑𝜇

∫[0,𝑥0 ) ∫[0,𝑥0 ) ∫[𝑥0 ,1] ∫[𝑥0 ,1]

=− 𝑓 𝑔𝑑𝜇,
∫[𝑥0 ,1]
implying 𝐼 ≤ 0, and the proof is complete. □
The previous lemma can be extended to finite sums of block rearrangements 𝑔𝑖 - the following general result holds:

Theorem 4.4. Suppose that 𝜇 is a finite measure on ([0, 1]) and that the functions 𝑓 , 𝑔1 , … , 𝑔𝑛 ∈ 𝐿2 (𝜇) fulfill the following properties:

1. 𝑓 ∶ [0, 1] → [0, ∞) is non-decreasing.

2. Each 𝑔𝑖 ∶ [0, 1] → R is a block rearrangement and fulfills ∫[0,1] 𝑔𝑖 𝑑𝜇 = 0.
∑
Then the following inequality holds for 𝑔 = 𝑛𝑖=1 𝑔𝑖 :

‖𝑓 − 𝑔‖22 ≥ ‖𝑓 ‖22 + ‖𝑔‖22 . (16)

Proof. From the proof of the previous lemma we know that ∫[0,1] 𝑓 𝑔𝑖 𝑑𝜇 ≤ 0 for every 𝑖 ∈ {1, … , 𝑛}. Having that and considering
( )2 ( 𝑛 )2
∑𝑛 ∑
𝑓− 𝑔𝑖 𝑑𝜇 = 𝑓 2 𝑑𝜇 + 𝑔𝑖 𝑑𝜇
∫[0,1] ∫[0,1] ∫[0,1]
𝑖=1 𝑖=1
⏟⏞⏞⏞⏟⏞⏞⏞⏟
=𝑔 2
∑
𝑛
−2 𝑓 𝑔 𝑑𝜇
∫[0,1] 𝑖
𝑖=1
⏟⏞⏞⏞⏞⏟⏞⏞⏞⏞⏟
≤0

2
≥ 𝑓 𝑑𝜇 + 𝑔 2 𝑑𝜇
∫[0,1] ∫[0,1]
yields the desired result. □

Remark 4.5. Theorem 4.4 can further be generalized since sets of 𝜇-measure 0 can be ignored. However, for tackling the lower
inequality we will work with 𝜇 = 𝜆, and the version stated above will suffice.
∑
Before proceeding with the proof of the lower inequality we illustrate with an example, why working with sums 𝑛𝑖=1 𝑔𝑖 is
necessary for proving the lower inequality via rearrangements.

Example 4.6. Consider 𝑁 = 16 and the symmetric permutation 𝜋 = (15, 2, 16, 14, 11, 12, 7, 8, 10, 9, 5, 6, 13, 4, 1, 3) ∈ 𝛴16 , for which
1
we have 𝐼𝜋− = {10, 11, 12, 14, 15, 16}, 𝑘 = 6 and 𝐩 = 16 ⋅ (1, 6, 6, 10, 13, 14). Again following the previously described rearrangement
1
yields the permutation 𝜋̂ = (16, 15, 14, 13, 5, 6, 7, 8, 9, 12, 11, 10, 4, 3, 2, 1) with 𝐼𝜋̂ = {12, 13, 14, 15, 16}, 𝑘̂ = 5 and 𝐩̂ = 16 ⋅ (2, 9, 11, 13, 15).
Fig. 4 illustrates the two shuffles and the mass rearrangement while Fig. 5 depicts the functions 𝑓 and 𝑓 , respectively. ̂

Fig. 4. The original shuffle (left panel) and its rearrangement (right panel) as discussed in Example 4.6.

We show that 𝑓 − 𝑓̂ can be expressed as a sum of block rearrangements fulfilling the properties of the previous theorem and
proceed as follows: Define the functions 𝑔𝑖𝑗 ∶ [0, 1] → R for 𝑖 < 𝑗 and 𝑖, 𝑗 ∈ {1, … , 𝑘} by
𝑔𝑖𝑗 (𝑥) ∶= 1𝐼 𝑖 (𝑥) − 1𝐼 𝑗 (𝑥),
𝑘 𝑘

10
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

Fig. 5. The functions 𝑓 and 𝑓̂ as considered in Example 4.6.

with 𝐼𝑘𝑖 and 𝐼𝑘𝑗 as in Section 2. Obviously each 𝑔𝑖𝑗 is a block rearrangement and satisfies ∫[0,1] 𝑔𝑖𝑗 𝑑𝜆 = 0. Notice that the function
𝑓 −𝓁 ⋅𝑔𝑖𝑗 corresponds to shifting 𝓁 masses from 𝐼𝑘𝑖 to 𝐼𝑘𝑗 (i.e., shifting to the right). Finally, considering 𝑔 ∶= 𝑔56 +𝑔45 +2𝑔34 +5𝑔23 +𝑔12
we have

𝑓̂ = 𝑓 − 𝑔,

so applying Eq. (16) yields ‖𝑓 − 𝑔‖22 ≥ ‖𝑓 ‖22 + ‖𝑔‖22 , which in turn shows that 𝜌(𝑆𝜋 ) > 𝜌(𝑆𝜋̂ ) as well as

̂
𝑚𝜋,𝑘 (𝐩) > 𝑚𝜋,̂ 𝑘̂ (𝐩)

hold. Notice that 𝑔 can alternatively be represented as 𝑔 ∶= 𝑔16 + 3𝑔23 + 𝑔24 , i.e., the representation of 𝑔 = 𝑓 − 𝑓̂ in terms of the
functions 𝑔𝑖𝑗 is not necessarily unique.

We finally return to proving Eq. (12) for symmetric shuffles. To this end, we apply Lemma 4.3 and proceed in two steps: (i) Using
the rearrangement idea reduce the problem to a handy subclass of symmetric shuffles. (ii) Prove the inequality for all elements of
the subclass and again use the fact, that symmetric shuffles are dense (Lemma 3.1).
The subclass consists of all symmetric shuffles/permutations of the type 𝜋̂ considered in the two previous examples. We will
define these permutations 𝜋 only on the sets 𝐼𝜋− since, using symmetry their extension to 𝐼𝜋+ and 𝐼𝜋0 is unique. In what follows,
𝑁 ∈ N will be arbitrary but fixed. For each such 𝑁 define
1
𝛴̂ 𝑁 ∶= {𝜋̂ ∈ 𝛴𝑁 ∶ 𝜋symmetric,
̂ 𝐼𝜋−̂ = {𝑁} and 𝜋(𝑁)
̂ ∈ {1, … , 𝑁 − 1}}
{
2 −
̂ ̂
𝛴𝑁 ∶= 𝜋̂ ∈ 𝛴𝑁 ∶ 𝜋̂ symmetric, 𝑘 ≥ 2, 𝐼𝜋̂ = {𝑁 − 𝑘 + 1, 𝑁 − 𝑘̂ + 2, … , 𝑁}
̂

such that 𝜋(𝑁

̂ − 𝑘̂ + 1) ∈ {𝑘,
̂ … , 𝑁 − 𝑘}̂ and
}
𝜋(𝑁
̂ − 𝑖 + 1) = 𝑖 for 𝑖 ∈ {1, … , 𝑘̂ − 1} ,

set 𝛴̂ 𝑁 ∶= 𝛴̂ 𝑁 ̂ 2 , and let ̂𝑁 , ̂ 1 , ̂ 2 ⊆  𝑠𝑦𝑚 denote the corresponding families of shuffles. The following result formalizes the
1 ∪𝛴
𝑁 𝑁 𝑁 𝑁
afore-mentioned reduction and shows that it suffices to prove Eq. (12) for all shuffles in ̂𝑁 .

𝑠𝑦𝑚
Lemma 4.7. Suppose that 𝑁 ∈ N with 𝑁 ≥ 4 and that 𝑆𝜋 ∈ 𝑁 . Then there exists some shuffle 𝑆𝜋̂ ∈ ̂𝑁 such that 𝑚𝜋,𝑘 (𝐩) ≥ 𝑚𝜋,̂ 𝑘̂ (𝐩)
̂
holds.

𝑠𝑦𝑚
Proof. Let 𝑁 be as in the theorem, 𝜋 ∈ 𝛴𝑁 be arbitrary but fixed, and 𝑘 = #𝐼𝜋− . Setting

𝑘 ∑
𝑘
𝑓 (𝑥) = 𝑝 𝟏 𝑖−1 𝑖 (𝑥)
𝑁 𝑖=1 𝑖 ( 𝑘 , 𝑘 ]
∑𝑘
with 𝐩 ∈ [0, 1]𝑘 corresponding to the shuffle 𝑆𝜋 , it follows that 𝑓 is non-decreasing. Writing 𝛥 = 𝑁 𝑖=1 𝑝𝑖 define
{
0, if 𝛥≤𝑁 − 1,
𝓁 ∶= ∑ ( )
max{𝑗 ∈ {1, … , 𝑘} ∶ 𝑗𝑖=1 𝑁 − (2𝑖 − 1) ≤ 𝛥}, else.
𝛥
(i) In case of 𝓁 = 0 we consider the symmetric permutation 𝜋̂ ∈ 𝛴̂ 𝑁
1 with 𝐼 − = {𝑁} and 𝜋(𝑁)
𝜋̂
̂ ∶= 𝑁 − 𝛥. Then 𝑝̂ = 𝑁
and setting
𝑘
𝑓̂(𝑥) = 𝑝̂ 𝟏 𝑘−1 (𝑥)
𝑁 (
𝑘
,1]

11
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

yields that 𝑔 ∶= 𝑓 − 𝑓̂ satisfies ∫[0,1] 𝑔𝑑𝜆 = 0 and 𝑔 is a block rearrangement. Applying Lemma 4.3 yields ‖𝑓 − 𝑔‖22 ≥ ‖𝑓 ‖22 or,
equivalently, 𝜌(𝑆𝜋 ) ≥ 𝜌(𝑆𝜋̂ ). By construction we have 𝜙(𝑆𝜋 ) = 𝜙(𝑆𝜋̂ ) and therefore 𝑚𝜋,𝑘 (𝐩) ≥ 𝑚𝜋,1 ̂ holds.
̂ (𝑝)
∑
(ii) Let 𝓁 ≥ 1 and set 𝛥∗ = 𝓁𝑖=1 𝑁 − (2𝑖 − 1). (a) Suppose that 𝛥 − 𝛥∗ > 0. Setting 𝑘̂ ∶= 𝓁 + 1 ≤ 𝑘 we consider 𝜋̂ ∈ 𝛴𝑁
⋃̂
with(𝐼𝜋−̂ = 𝑘𝑖=1 {𝑁 − 𝑖 + 1},)𝜋(𝑁̂ − 𝑘̂ + 1) = 𝑁 − 𝑘̂ + 1 − (𝛥 − 𝛥∗ ) and 𝜋(𝑁
̂ − 𝑖 + 1) = 𝑖 for 𝑖 ∈ {1, … , 𝑘̂ − 1}. Then 𝜋̂ ∈ 𝛴̂ 𝑁
2 with
∗
𝛥−𝛥 𝑁−(2𝓁−1) 𝑁−1
𝐩̂ = 𝑁
, 𝑁
,…, 𝑁 and, by construction, we have 𝜙(𝑆𝜋 ) = 𝜙(𝑆𝜋̂ ). Setting

𝑘 ∑
𝑘
𝑓̂(𝑥) = 𝑝̂𝑖 𝟏 𝑖−1 𝑖 (𝑥)
𝑁 ̂
( , ]
𝑘 𝑘
𝑖=𝑘−𝑘+1

yields that 𝑔 ∶= 𝑓 − 𝑓̂ satisfies ∫[0,1] 𝑔𝑑𝜆 = 0. Considering that 𝑔 can be expressed as finite sum of bock rearrangements 𝑔𝑖 of
satisfying ∫[0,1] 𝑔𝑖 𝑑𝜆 = 0 (see Example 4.6 for illustration), applying Theorem 4.4 yields ‖𝑓 − 𝑔‖22 ≥ ‖𝑓 ‖22 , i.e., 𝜌(𝑆𝜋 ) ≥ 𝜌(𝑆𝜋̂ ) and
𝑚𝜋,𝑘 (𝐩) ≥ 𝑚𝜋,̂ 𝑘̂ (𝐩).
̂
⋃𝑘̂
(b) Finally, suppose that( 𝛥 − 𝛥∗ = 0. Set 𝑘̂ ∶= −
) 𝓁 and consider 𝜋̂ ∈ 𝛴𝑁 with 𝐼𝜋̂ = 𝑖=1 {𝑁 − 𝑖 + 1} and 𝜋(𝑁
̂ ̂
− 𝑖 + 1) = 𝑖 for 𝑖 ∈ {1, … , 𝑘}.
𝑁−(2𝑘−1) 𝑁−1
Then 𝜋̂ ∈ 𝛴̂ 𝑁 with 𝐩̂ =
2
𝑁
, … , 𝑁 . Setting

𝑘 ∑
𝑘
𝑓̂(𝑥) = 𝑝̂ 𝟏 𝑖−1 𝑖 (𝑥)
𝑁 𝑖=1 𝑖 ( 𝑘 , 𝑘 ]

as well as 𝑔 ∶= 𝑓 − 𝑓̂ and proceeding analogously to the previous case concludes the proof. □

As second and ultimate step we show that Eq. (12) holds for all shuffles within the subclass ̂𝑁 . Building upon the fact that
∑ ∑
𝜙(𝑆𝜋 ) ≥ − 21 implies 𝑘𝑖=1 𝑝𝑖 ≤ 1∕4 for any symmetric shuffle 𝑆𝜋 , we define 𝑀𝑁,𝑘 ∶= {𝐱 ∈ [0, 1]𝑘 ∶ 𝑁1 𝑘𝑖=1 𝑥𝑖 ≤ 1∕4} and interpret
𝑚𝜋,𝑘 as function on 𝑀𝑁,𝑘 .

Theorem 4.8. For every shuffle 𝑆𝜋̂ ∈ ̂𝑁 with resolution 𝑁 ∈ N and 𝑁 ≥ 4 the following inequality holds:
2√ ( )3∕2
𝜌(𝑆𝜋̂ ) ≥ 3 1 + 2𝜙(𝑆𝜋̂ ) − 1.
9

Proof. We show that under the assumptions of the theorem 𝑚𝜋,̂ 𝑘̂ (𝐩) ̂ ≥ 0 holds and proceed as follows: For 𝐱 ∈ 𝑀𝑁,𝑘̂ , calculating
the first and second partial derivative w.r.t. 𝑥𝑗 yields
√ √
√ ⎛√ ⎞
√ ̂ √ 𝑘̂
6√ 4 ∑ 6 ⎜√ ∑
𝑘
𝜕 12 √1 − 4 ⎟
𝑚𝜋,̂ 𝑘̂ (𝐱) = − 𝑥𝑗 + √1 −
𝑁 ⎜⎜ 𝑗⎟
𝑥𝑖 = 𝑥 − 2𝑥
𝜕𝑥𝑗 𝑁 𝑁 𝑁 𝑖=1 𝑁 𝑖=1 𝑖 ⎟
⎝ ⎠
as well as
⎛ ⎞
𝜕2 12 ⎜ 1 ⎟
𝑚𝜋,̂ 𝑘̂ (𝐱) = − ⎜ √ + 1⎟ < 0
𝜕 2 𝑥𝑗 𝑁 ⎜ 4 ∑𝑘̂ ⎟
⎝ 𝑁 1 − 𝑁 𝑖=1 𝑥𝑖 ⎠
̂ As a direct consequence, the mapping
for every 𝑗 ∈ {1, … , 𝑘}.

𝑡 ↦ 𝑚𝜋,̂ 𝑘̂ (𝑥1 , … , 𝑥𝑗−1 , 𝑡, 𝑥𝑗+1 , … , 𝑥𝑘̂ )

is concave on the convex polytope 𝑀𝑁,𝑘̂ .

We again distinguish two cases: (i) if 𝑆𝜋̂ ∈ ̂𝑁 1 then 𝑘 ̂ = 1 and concavity of 𝑚𝜋,1
̂ implies that 𝑚𝜋,1 ̂ ≥ 𝑚𝜋,1
̂ (𝑝) ̂ (0) = 0 or
𝑁−1
𝑚𝜋,1
̂ ( 𝑝)
̂ ≥ 𝑚𝜋,1
̂ ( 𝑁
). A straightforward calculation yields
( ( )3∕2 )
3 𝑁−1 3 6(𝑁 − 1)2 4(𝑁 − 1)
𝑁 𝑚𝜋,1̂ ( 𝑁 )=𝑁 1− − 1−
𝑁3 𝑁2
( )3∕2
= 𝑁 3 − 6(𝑁 − 1)2 − 𝑁 2 − 4(𝑁 − 1)
= 𝑁 3 − 6(𝑁 − 1)2 − (𝑁 − 2)3 = 2,

implying the assertion. ( )

̂
(ii) In case that 𝑆𝜋̂ ∈ ̂𝑁 ̂ = 𝑗,
2 we have 𝐩 𝑁−(2(𝑘−1)−1)
, … , 𝑁−1 with 𝑗 ∈ {1, … , 𝑁 − (2𝑘̂ − 1)}. Now, setting
𝑁 𝑁 𝑁
( ̂
)
𝐩̂ 0 ∶= 0, 𝑁−(2(𝑁
𝑘−1)−1)
, … , 𝑁−1
𝑁
∈ 𝑀𝑁,𝑘̂ ,
( ̂
)
𝐩̂ 1 ∶= 𝑁−(2𝑁
𝑘−1) 𝑁−(2(𝑘−1)−1)
, 𝑁
, … , 𝑁−1
𝑁
∈ 𝑀𝑁,𝑘̂ ,

12
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

̂ ≥ 𝑚𝜋,̂ 𝑘̂ (𝐩̂ 𝑗 ) holds for at least one 𝑗 ∈ {0, 1}. Therefore, considering
concavity of 𝑚𝜋,̂ 𝑘̂ in each coordinate implies that 𝑚𝜋,̂ 𝑘̂ (𝐩)
𝑘̂ 𝑗 ∶= 𝑘̂ + 𝑗 − 1 and 𝑘̂ 𝑗 ≤ 𝑁∕2 it follows that

𝑘̂ 𝑘𝑗 ̂
4 ∑ 4 ∑ 𝑁 − (2𝑖 − 1) (𝑁 − 2𝑘̂ 𝑗 )2
1− 𝑝̂𝑗𝑖 = 1 − = ,
𝑁 𝑖=1 𝑁 𝑖=1 𝑁 𝑁2
as well as
𝑘̂ 𝑘𝑗 ̂
6 ∑ 2 6 ∑ (𝑁 − (2𝑖 − 1))2 2𝑘𝑗
̂ (𝑁 − 2𝑘̂ 𝑗 )3
1− 𝑝̂𝑗𝑖 > 1 − − = .
𝑁 𝑖=1 𝑁 𝑖=1 𝑁 2 𝑁 3 𝑁3
Therefore 𝑚𝜋,̂ 𝑘̂ (𝐩̂ 𝑗 ) > 0 and the proof is complete. □

Following the idea of the proof of Theorem 4.8 the copula 𝐶𝛼 depicted in Fig. 6 for 𝛼 ∈ [0, 12 ] seems a very natural suspect for
attaining the lower bound.

Fig. 6. The copula 𝐶𝛼 (left panel) for which the lower inequality is sharp, and its corresponding diagonal 𝛿𝛼 (right panel).

In fact, it was already shown in [6] that for this copula

𝜙(𝐶𝛼 ) = 6𝛼 2 − 6𝛼 + 1,
𝜌(𝐶𝛼 ) = −16𝛼 3 + 24𝛼 2 − 12𝛼 + 1,
√ ( )3∕2
holds, so 𝜌(𝐶𝛼 ) = 92 3 1 + 2𝜙(𝐶𝛼 ) − 1 for every 𝛼 ∈ [0, 12 ].

5. Getting closer to the non-sharp upper bound

We now return to the upper bound for 𝛺𝜙,𝜌 which was already shown not to be (globally) sharp by the authors in [6]. The
natural question therefore is, how non-sharp it is, i.e., how close one can get to the upper bound. In [6] the authors also tackled
this question, derived the function 𝑟 ∶ [− 12 , 1] → [−1, 1], defined by
√
3
⎧2𝑥 + 1 − 3 (1 + 2𝑥) 2 , if 𝑥 ∈ [− 21 , − 18 ],
⎪4 2
7
9
⎪ 3 𝑥 + 24 , if 𝑥 ∈ [− 81 , 14 ],
𝑟(𝑥) = ⎨ 2𝑛+1 2𝑛2 −2𝑛+1 3 3
⎪ 𝑛2 +𝑛 𝑥 + 2𝑛2 +2𝑛 , if 𝑥 ∈ [1 − 2𝑛
, 1 − 2(𝑛+1) ],
⎪
⎩1, if 𝑥 = 1,
and showed that for every point (𝑥, 𝑟(𝑥)) there exists some copula 𝐶 with 𝜙(𝐶) = 𝑥 and 𝜌(𝐶) = 𝑟(𝑥). Notice that 𝑟 is piecewise linear
on the interval [− 18 , 1] and is, for arbitrary 𝑥 ∈ [− 21 , 1] quite close to the upper bound given by 1 − 23 (1 − 𝑥)2 .
The goal of this section is to prove that on the interval [− 81 , 1] - outside the countably many points on which the upper inequality
is known to be sharp - the function 𝑟 can be exceeded. Using convexity of 𝛺𝜙,𝜌 we will work with even 𝑛 ∈ N and the shuffles
𝑆𝑛∗ ∶= 𝑆𝜋 ∗ with 𝜋 ∗ ∈ 𝛴𝑛 fulfilling 𝑖 − 𝜋 ∗ (𝑖) = 1 for every 𝑖 ∈ 𝐼𝜋−∗ = {𝑖 ∈ {1, … , 𝑛} ∶ 𝑖 even}. As mentioned in Section 3, these shuffles
constitute points at which the upper inequality is sharp (see Fig. 1 for an example). Our idea consists in ‘interpolating’ between two
consecutive shuffles 𝑆𝑛∗ and 𝑆𝑛+2 ∗ . Illustrating the approach we start with the pair 𝑆 ∗ and 𝑆 ∗ and then extend to 𝑆 ∗ and 𝑆 ∗ for
2 4 𝑛 𝑛+2
arbitrary even 𝑛 ∈ N.

5.1. Special case: Interpolating between 𝑆2∗ and 𝑆4∗

We show that for every 𝑥 ∈ (− 81 , 14 ) there exists some copula 𝐶 with 𝜙(𝐶) = 𝑥 fulfilling that 𝜌(𝐶) > 𝑟(𝑥) and proceed as follows:

13
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

Given 𝑎 ∈ [ 14 , 21 ] and 𝑏 ∈ [0, 14 ] define the two diagonals 𝛿𝑎↑ and 𝛿𝑏↓ by (see Figs. 7 and 8 for an illustration)

⎧0, if 𝑥 ∈ [0, 𝑎],

⎪
𝛿𝑎↑ (𝑥) = ⎨𝑥 − 𝑎, if 𝑥 ∈ [𝑎, 1 − 𝑎],
⎪
⎩2𝑥 − 1, if 𝑥 ∈ [1 − 𝑎, 1],
and
⎧0, if 𝑥 ∈ [0, 41 ],
⎪
⎪2(𝑥 − 1 ), if 𝑥 ∈ [ 14 , 1
+ 𝑏],
↓ ⎪ 4 4
𝛿𝑏 (𝑥) = ⎨𝑥 + 𝑏 − 1 , if 𝑥 ∈ [ 4 + 𝑏, 43 − 𝑏],
1
4
⎪1
⎪2, if 𝑥 ∈ [ 34 − 𝑏, 43 ],
⎪2𝑥 − 1, if 𝑥 ∈ [ 34 , 1],
⎩
and consider the corresponding diagonal copulas 𝐸𝛿 ↑ and 𝐸𝛿 ↓
𝑎 𝑏

Fig. 7. The diagonal 𝛿𝑎↑ (left panel) and the support of the corresponding diagonal copula 𝐸𝛿𝑎↑ (right panel).

Fig. 8. The diagonal 𝛿𝑏↓ (left panel) and the support of the induced diagonal copula 𝐸𝛿↓ (right panel).
𝑏

Note that 𝛿𝑎↑ coincides with the diagonal 𝛿𝑎 studied in Example 14 in [6] (and the copula 𝐾𝛿𝑎 considered there with the diagonal
copula 𝐸𝛿 ↑ ), so we obtain
𝑎

𝜙(𝐸𝛿 ↑ ) = 6𝑎2 − 6𝑎 + 1,
𝑎
3
𝜌(𝐸𝛿 ↑ ) = 8𝑎3 − 6𝑎 + ,
𝑎 2
√
1 3
as well as 𝜌(𝐸𝛿↑ ) = 2𝜙(𝐸𝛿 ↑ ) + 2
− 9
(1 + 2𝜙(𝐸𝛿 ↑ ))3∕2 . Furthermore, tedious but simple calculations (see Appendix) yield
𝑎 𝑎 𝑎

1
𝜙(𝐸𝛿↓ ) = −6𝑏2 + 3𝑏 − ,
𝑏 8
9 1
𝜌(𝐸𝛿 ↓ ) = 8𝑏3 − 12𝑏2 + 𝑏 + ,
𝑏 2 8
√
3 6
which altogether implies 𝜌(𝐸𝛿 ↓ ) = 𝜙(𝐸𝛿↓ ) + 8
− 36
(1 − 4𝜙(𝐸𝛿 ↓ ))3∕2 .
𝑏 𝑏 𝑏

14
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

Hence, defining the concave function ℎ ∶ [− 18 , 14 ] ↦ [0, 1] by

√
3 6
ℎ(𝑥) ∶= + 𝑥 − (1 − 4𝑥)3∕2 ,
8 36
we obtain 𝜌(𝐸𝛿 ↓ ) = ℎ(𝜙(𝐸𝛿↓ )). Varying the values of 𝑎 ∈ [ 41 , 21 ] and 𝑏 ∈ [0, 14 ], the copulas 𝐸𝛿 ↑ and 𝐸𝛿 ↓ yield an ‘interpolation’ from
𝑏 𝑏 𝑎 𝑏
the copula 𝑆2∗ to 𝑆4∗ in the following sense:
𝑎→1∕4 𝑏→1∕4
𝑆2∗ = 𝐸𝛿 ↑ ⟶ 𝐸𝛿 ↑ = 𝐸𝛿 ↓ ⟶ 𝐸𝛿 ↓ = 𝑆4∗
1∕2 1∕4 0 1∕4

Recall that for 𝑥 ∈ [− 81 , 41 ] the graph of function 𝑟 is the linear interpolation between (𝜙(𝐸𝛿 ↓ ), 𝜌(𝐸𝛿↓ )) and (𝜙(𝑆4∗ ), 𝜌(𝑆4∗ )). Considering
0 0
𝜙(𝐸𝛿 ↓ ) = − 18 and 𝜙(𝐸𝛿 ↓ ) = 1
4
, strict concavity of ℎ implies that ℎ(𝑥) > 𝑟(𝑥) for 𝑥 ∈ (− 18 , 41 ), so the afore-mentioned construction
0 1∕4
1 1
exceeds 𝑟 on the interval (− 8 , 4 ).

∗
5.2. General case: Interpolating between 𝑆𝑛∗ and 𝑆𝑛+2

We show that for every 𝑥 ∈ (1 − 3𝑛 , 1 − 𝑛+2 3

) , where 𝑛 ≥ 4 is even, there exists some copula 𝐶 with 𝜙(𝐶) = 𝑥 fulfilling that
𝜌(𝐶) > 𝑟(𝑥). Doing so we again consider the copulas 𝐸𝛿↑ from the previous subsection, work with finite ordinal sums and start with
𝑎
quickly recalling their construction (for more information on ordinal sums and patchworks see [9] and the references therein).
Suppose that 𝑛 ∈ N, that (𝑎1 , 𝑏1 ), (𝑎2 , 𝑏2 ), … , (𝑎𝑛 , 𝑏𝑛 ) are pairwise disjoint, non-degenerated intervals in [0, 1] and that 𝐶1 , … , 𝐶𝑛
( )𝑛
are copulas. Then the ordinal sum 𝑂 = ⟨(𝑎𝑘 , 𝑏𝑘 ), 𝐶𝑘 ⟩ 𝑘=1 is the copula defined by

⎧ ( )
𝑢−𝑎 𝑣−𝑎𝑘
⎪𝑎𝑘 + (𝑏𝑘 − 𝑎𝑘 )𝐶𝑘 𝑏 −𝑎𝑘 , 𝑏𝑘 −𝑎𝑘
, if (𝑢, 𝑣) ∈ (𝑎𝑘 , 𝑏𝑘 )2 ,
𝑂(𝑢, 𝑣) ∶= ⎨ 𝑘 𝑘
⎪𝑀(𝑢, 𝑣), elsewhere.
⎩
It is straightforward to express 𝜙(𝑂) and 𝜌(𝑂) in terms of the corresponding values of the copulas 𝐶𝑘 .
( )𝑛
Lemma 5.1. Suppose that 𝑂 = ⟨(𝑎𝑘 , 𝑏𝑘 ), 𝐶𝑘 ⟩ 𝑘=1 is the ordinal sum of 𝐶1 , … , 𝐶𝑛 with respect to (𝑎1 , 𝑏1 ), (𝑎2 , 𝑏2 ), … , (𝑎𝑛 , 𝑏𝑛 ). Then the
following formulas hold:
∑
𝑁
𝜌(𝑂) = 1 − (𝑏𝑘 − 𝑎𝑘 )3 (1 − 𝜌(𝐶𝑘 )),
𝑘=1
∑
𝑁
( )
𝜙(𝑂) = 6𝑎𝑘 (𝑏𝑘 − 𝑎𝑘 ) + (𝑏𝑘 − 𝑎𝑘 )2 (𝜙(𝐶𝑘 ) + 2) − 2.
𝑘=1

Proof. The first identity was already shown in [6]. Concerning the second one, using change of coordinates, yields
𝜙(𝐶) + 2
= 𝐶(𝑢, 𝑢)𝑑𝜆(𝑢)
6 ∫[0,1]
( )
∑
𝑁
𝑢 − 𝑎𝑘 𝑢 − 𝑎𝑘
= 𝑎𝑘 (𝑏𝑘 − 𝑎𝑘 ) + (𝑏𝑘 − 𝑎𝑘 )𝐶𝑘 , 𝑑𝜆(𝑢)
∫[𝑎𝑘 ,𝑏𝑘 ] 𝑏 𝑘 − 𝑎𝑘 𝑏 𝑘 − 𝑎𝑘
𝑘=1
∑𝑁
= 𝑎𝑘 (𝑏𝑘 − 𝑎𝑘 ) + (𝑏𝑘 − 𝑎𝑘 )2 𝐶𝑘 (𝑢, 𝑢)𝑑𝜆(𝑢)
∫[0,1]
𝑘=1
∑𝑁
𝜙(𝐶𝑘 ) + 2
= 𝑎𝑘 (𝑏𝑘 − 𝑎𝑘 ) + (𝑏𝑘 − 𝑎𝑘 )2 .
𝑘=1
6

This completes the proof. □

( )𝑁
Suppose now that 𝑛 ≥ 4 is even and set 𝑁 = 𝑛∕2. Then according to Lemma 5.1 for the ordinal sum 𝑂𝑎𝑁 ∶= ⟨( 𝑘−1
𝑁
, 𝑘
𝑁
), 𝐸𝛿↑ ⟩
𝑎 𝑘=1
we get
1
𝜌(𝑂𝑎𝑁 ) = 1 − (1 − 𝜌(𝐸𝛿 ↑ )),
𝑁2 𝑎
1
𝜙(𝑂𝑎𝑁 ) = 1 − (1 − 𝜙(𝐸𝛿 ↑ )).
𝑁 𝑎

∗
In order to ‘interpolate’ between 𝑆𝑛∗ and 𝑆𝑛+2 we will work with the ordinal sum 𝑂𝑎𝑁 with 𝑎𝑁 ∶= 𝑁
2𝑁+2
∈ ( 14 , 12 ). The following
𝑁
result holds:

Lemma 5.2. For every even 𝑛 ≥ 4 we have 𝜌(𝑂𝑎𝑁 ) > 𝑟(𝜙(𝑂𝑎𝑁 )).
𝑁 𝑁

15
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

3 3
Proof. As before we set 𝑁 = 𝑛∕2. According to [6], on the compact interval [1 − 2𝑁
,1 − 2(𝑁+1)
] the function 𝑟 is given by
2𝑁+1 2𝑁 2 −2𝑁+1
𝑟(𝑥) = 𝑁 2 +𝑁
𝑥+ 2(𝑁 2 +𝑁)
. Considering

1 ( )
𝜙(𝑂𝑎𝑁 ) = 1 − 1 − (6𝑎2𝑁 − 6𝑎𝑁 + 1)
𝑁 𝑁
6 2𝑁 2 + 𝑁 − 4
= 1 − 𝑎𝑁 (1 − 𝑎𝑁 ) =
𝑁 2(𝑁 + 1)2
it follows that
3 3
1− − 𝜙(𝑂𝑎𝑁 ) = >0
2(𝑁 + 1) 𝑁 2(𝑁 + 1)2
and
3 3
𝜙(𝑂𝑎𝑁 ) − (1 − )= >0,
𝑁 2𝑁 2𝑁(𝑁 + 1)2
( 3 3 )
which directly yields 𝜙(𝑂𝑎𝑁 ) ∈ 1 − 2𝑁 ,1 − 2(𝑁+1)
. Furthermore a straightforward calculation shows
𝑁

2𝑁 4 + 6𝑁 3 + 3𝑁 2
− 7𝑁 − 3
𝑟(𝜙(𝑂𝑎𝑁 )) = ,
𝑁 2𝑁(𝑁 + 1)3
2𝑁 5 + 6𝑁 4 + 3𝑁 3 − 7𝑁 2 − 3𝑁 + 1
𝜌(𝑂𝑎𝑁 ) = ,
𝑁 2𝑁 2 (𝑁 + 1)3
1
implying 𝜌(𝑂𝑎𝑁 ) − 𝑟(𝜙(𝑂𝑎𝑁 )) = 2𝑁 2 (𝑁+1)3
> 0. □
𝑁 𝑁

Summing up, and using convexity of 𝛺𝜙,𝜌 we have shown the following result for the function 𝑠 ∶ [− 21 , 1] → [−1, 1], defined by
√
⎧2𝑥 + 1∕2 − 3 (1 + 2𝑥)3∕2 , if 𝑥 ∈ [− 21 , − 18 ],
⎪ √ 9
⎪𝑥 + 3 − 6 (1 − 4𝑥)3∕2 , if 𝑥 ∈ [− 81 , 14 ],
⎪ 𝑁8 36
∗
⎪ 𝜌(𝑂𝑎𝑁 )−𝜌(𝑆2𝑁 ) ∗ ∗ ∗ ), 𝜙(𝑂𝑁 )] , 𝑁 ≥ 2,
𝑠(𝑥) = ⎨ 𝜙(𝑂𝑁 )−𝜙(𝑆 ∗ ) (𝑥 − 𝜙(𝑆2𝑁 )) + 𝜌(𝑆2𝑁 ), if 𝑥 ∈ [𝜙(𝑆2𝑁 𝑎
𝑎𝑁 𝑁
⎪ 𝜌(𝑆 ∗ )−𝜌(𝑂𝑁 )
2𝑁

⎪ 2𝑁+2 𝑎𝑁
(𝑥 − 𝜙(𝑂𝑎𝑁 )) + 𝜌(𝑂𝑎𝑁 ), if 𝑥 ∈ ∗
[𝜙(𝑂𝑎𝑁 ), 𝜙(𝑆2𝑁+2 )] , 𝑁 ≥ 2,
⎪ 𝜙(𝑆2𝑁+2
∗ )−𝜙(𝑂𝑎𝑁 )
𝑁
𝑁 𝑁 𝑁
⎪1, if 𝑥 = 1.
⎩

3
Theorem 5.3. For every 𝑥 ∈ (− 18 , 1) with 𝑥 ∉ {1 − 2𝑁 ∶ 𝑁 ≥ 2} the function 𝑠 fulfills 𝑠(𝑥) > 𝑟(𝑥) and there exists some copula 𝐶𝑥 fulfilling
𝜙(𝐶𝑥 ) = 𝑥 and 𝜌(𝐶𝑥 ) = 𝑠(𝑥).
We conjecture that the upper bound 𝑠 is not best possible, i.e., that for every 𝑥 ∈ (− 18 , 1) with 𝑥 ∉ {1 − 2𝑁
3
∶ 𝑁 ≥ 2} there exists
some copula 𝐵𝑥 fulfilling 𝜙(𝐵𝑥 ) = 𝑥 and 𝜌(𝐵𝑥 ) > 𝑠(𝑥), so deriving a globally sharp upper inequality is still an open question.

Data availability

No data was used for the research described in the article.

Acknowledgments

The first and the second author gratefully acknowledge the financial support from AMAG Austria Metall AG within the project
‘ProSa’. The third author gratefully acknowledges the support of the WISS 2025 project ‘IDA-lab Salzburg’ (20204-WISS/225/197-
2019 and 20102-F1901166-KZP).

Appendix. Complementary calculations for Section 5

In the sequel we derive the formulas for 𝜙(𝐸𝛿↓ ) and 𝜌(𝐸𝛿 ↓ ). The definition of 𝛿𝑏↓ implies
𝑏 𝑏

𝛿𝑏↓ (𝑢)𝑑𝜆(𝑢) = 2(𝑢 − 14 )𝑑𝜆(𝑢) + (𝑢 + 𝑏 − 14 )𝑑𝜆(𝑢)

∫[0,1] ∫[1∕4,1∕4+𝑏] ∫[1∕4+𝑏,3∕4−𝑏]
1
+ 𝑑𝜆(𝑢) + (2𝑢 − 1)𝑑𝜆(𝑢)
∫[3∕4−𝑏,3∕4] 2 ∫[3∕4,1]
2 1 2 1 3
=𝑏 + 8
− 2𝑏 + 2
𝑏 + 16
,

so 𝜙(𝐸𝛿↓ ) = −6𝑏2 + 3𝑏 − 18 . The diagonal copula 𝐸𝛿 ↓ distributes its mass as depicted in Fig. 9, we therefore obtain
𝑏 𝑏

1 1
𝛱(𝑢, 𝑣)𝑑𝜇𝐸 ↓ (𝑢, 𝑣) = 𝑢(𝑢 + )𝑑𝜆(𝑢) + 𝑢(2𝑢 + − 𝑏)𝑑𝜆(𝑢)
∫[0,1]2 𝛿𝑏 ∫[0,𝑏] 4 ∫[𝑏,1∕4] 4

16
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

+ 𝑢(𝑢 − 41 )𝑑𝜆(𝑢) + 1
𝑢( 21 𝑢 + 5
− 12 𝑏)𝑑𝜆(𝑢)
∫[1∕4,1∕4+𝑏] 2 ∫[1∕4+𝑏,3∕4−𝑏] 8

1
+ 𝑢( 12 𝑢 − 1
+ 21 𝑏)𝑑𝜆(𝑢) + 𝑢(𝑢 + 14 )𝑑𝜆(𝑢)
2 ∫[1∕4+𝑏,3∕4−𝑏] 8 ∫[3∕4−𝑏,3∕4]
5
+ 𝑢(2𝑢 − + 𝑏)𝑑𝜆(𝑢) + 𝑢(𝑢 − 14 )𝑑𝜆(𝑢)
∫[3∕4,1−𝑏] 4 ∫[1−𝑏,1]
1 2 1
=
12
𝑏 (8𝑏 + 3) + 192 (−64𝑏3 − 48𝑏2 − 12𝑏 + 7)+
1 1
192
(−64𝑏3 + 144𝑏2 − 204𝑏 + 43) + 12 𝑏(8𝑏2 − 21𝑏 + 18)
1 3 2
= 96 (64𝑏 − 96𝑏 + 36𝑏 + 25).
( )3∕2
9 1 8 2 1 2
Altogether, this yields 𝜌(𝐸𝛿 ↓ ) = 8𝑏3 − 12𝑏2 + 2
𝑏 + 8
. Finally, considering 3
𝜙(𝐸𝛿 ↓ ) − 3
= −(1 − 4𝑏)2 and 8 3
− 83 𝜙(𝐸𝛿↓ ) =
𝑏 𝑏 𝑏
3
𝜙(𝐸𝛿 ↓ ) − 𝜌(𝐸𝛿 ↓ ) + 8
it follows that
𝑏 𝑏
( )3∕2
3 1 2
𝜌(𝐸𝛿↓ ) = 𝜙(𝐸𝛿 ↓ ) + 8
− 8 3
− 83 𝜙(𝐸𝛿 ↓ )
𝑏 𝑏
√ ( 𝑏
)3∕2
3 6
= 𝜙(𝐸𝛿 ↓ ) + 8
− 36
1 − 4𝜙(𝐸 𝛿𝑏↓
) .
𝑏

Fig. 9. Mass distribution of the diagonal copula 𝐸𝛿↓ .

𝑏

References

[1] M. Scarsini, On measures of concordance, Stochastica 8 (1984) 201–218, 796650.

[2] R.B. Nelson, An Introduction to Copulas, Springer, New York, 2007, 2197664.
[3] J. Durbin, A. Stuart, Inversions and rank correlation coefficients, J. R. Stat. Soc. Ser. B. Stat. Methodol. 13 (1951) 303–309, 47994.
[4] M. Schreyer, R. Paulin, W. Trutschnig, On the exact region determined by Kendall’s 𝜏 and Spearman’s 𝜌, J. R. Stat. Soc. Ser. B. Stat. Methodol. 79 (2017)
613–633, 3611762.
[5] D. Kokol Bukovšek, N. Stopar, On the exact regions determined by Kendall’s tau and other concordance measures, Mediterr. J. Math. 20 (2023) 147,
4554107.
[6] D. Kokol Bukovšek, N. Stopar, On the exact region determined by Spearman’s rho and Spearman’s footrule, J. Comput. Appl. Math. 437 (2024) 115463,
4622519.
[7] D. Kokol Bukovšek, B. Mojškerc, On the exact region determined by Spearman’s footrule and Gini’s gamma, J. Comput. Appl. Math. 410 (2022) 141212,
4395320.
[8] D. Kokol Bukovšek, T. Košir, B. Mojškerc, M. Omladič, Spearman’s footrule and Gini’s gamma: local bounds for bivariate copulas and the exact region
with respect to Blomqvist’s beta, J. Comput. Appl. Math. 390 (2021) 113385, 4205170.
[9] F. Durante, C. Sempi, Principles of Copula Theory, CRC Press, 2015, 3443023.

17
M. Tschimpke et al. Journal of Computational and Applied Mathematics 457 (2025) 116259

[10] O. Kallenberg, Foundations of Modern Probability, Springer, New York, 2002, 4226142.
[11] A. Klenke, Wahrscheinlichkeitstheorie, Springer, Heidelberg, 2008, 3112259.
[12] W. Trutschnig, On a strong metric on the space of copulas and its induced dependence measure, J. Math. Anal. Appl. 384 (2011) 690–705, 2825218.
[13] T.M. Kasper, S. Fuchs, W. Trutschnig, On weak conditional convergence of bivariate Archimedean and Extreme Value copulas, and consequences to
nonparametric estimation, Bernoulli 27 (2021) 2217–2240, 4303881.
[14] T. Mroz, S. Fuchs, W. Trutschnig, How simplifying and flexible is the simplifying assumption in pair-copula constructions – analytic answers in dimension
three and a glimpse beyond, Electron. J. Stat. 15 (2021) 1951–1992, 4255318.
[15] P. Mikusinski, H. Sherwood, M.D. Taylor, Shuffles of min, Stochastica 13 (1992) 61–74, 1197328.
[16] H.H. Edwards, P. Mikusiński, M.D. Taylor, Measures of concordance determined by 𝐷4 –invariant copulas, Int. J. Math. Math. Sci. 70 (2004) 3867–3875,
2129414.
[17] J. Fernández Sánchez, W. Trutschnig, Some members of the class of (quasi-) copulas with given diagonal from the Markov kernel perspective, Comm.
Statist. Theory Methods 45 (2016) 1508–1526, 3462163.
[18] W. Rudin, Real and Complex Analysis, McGram-Hill, New York, 1987, 924157.
[19] M. Úbeda-Flores, On the best-possible upper bound on sets of copulas with given diagonal sections, Soft Comput. 12 (2008) 1019–1025, http:
//[Link]/10.1007/s00500-007-0269-5.
[20] R.B. Nelson, J.J. Quesada Molina, J.A. Rodríguez Lallena, M. Úbeda-Flores, Best-possible bounds on sets of bivariate distribution functions, J. Multivariate
Anal. 90 (2004) 348–358, 2081783.

18