Introduction
In the theory of aggregation functions, the volume that a bivariate function assigns to a rectangle of the unit square measures the interaction between its two arguments on that rectangle; it is non-negative on every rectangle exactly when the function is supermodular (2-increasing), a property shared by copulas and by the functions that generate universal integrals [12 , 13 , 9 ] . Motivated by this role, the average rectangular volume (ARV) of an integrable function F : [ 0 , 1 ] 2 → R F : [0, 1]^2 \to\mathbb{R} F : [ 0 , 1 ] 2 → R was introduced in [3 ] as the average of the F F F -volumes of all axis-parallel rectangles of the unit square. For supermodular aggregation functions it is a global index of the strength of the interaction between the two inputs; for a copula C C C , it is the expected C C C -volume of the rectangle spanned by two independent uniform points.
On the class of copulas, ARV behaves quite differently from the classical measures of concordance. It is invariant under reflections of one argument, so that it does not distinguish positive from negative dependence, and it satisfies 11 120 ≤ ARV ( C ) ≤ 2 15 \frac{11}{120} \leq\operatorname{ARV}(C) \leq\frac{2}{15} 120 11 ≤ ARV ( C ) ≤ 15 2 [3 Proposition 3.8] . It is therefore natural to ask how much information ARV \operatorname{ARV} ARV carries about a measure of concordance such as Spearman’s rho.
For two functionals defined on copulas, this question is encoded in their exact region ; determining such regions is often a challenging problem. A classical example is the relation between Kendall’s tau and Spearman’s rho, for which universal inequalities and sharp bounds have been established [6 , 26 ] . Related bounds and exact-region problems for restricted classes of copulas and for other measures of association have been studied, for instance, in [14 , 16 , 17 , 18 , 28 , 29 , 10 , 2 , 15 , 24 ] .
Partial information on the ( ARV , ρ ) (\operatorname{ARV},\rho) ( ARV , ρ ) region was obtained in [3 , 4 ] : the inequality ∣ ρ ( C ) ∣ ≤ 3 ARV ( C ) + 3 5 \lvert\rho(C)\rvert\leq3\operatorname{ARV}(C)+\frac{3}{5} ∣ ρ ( C )∣ ≤ 3 ARV ( C ) + 5 3 holds for every copula, minimal ARV \operatorname{ARV} ARV forces ∣ ρ ( C ) ∣ ≤ 1 2 \lvert\rho(C)\rvert\leq\frac{1}{2} ∣ ρ ( C )∣ ≤ 2 1 , and a family of shuffles of Min traces two curves joining the extreme parts of the region, which were conjectured to be its boundary.
In this paper we give a complete description of the region (Theorem 4.1 ). Its upper boundary is the graph of a concave, strictly increasing function r ∗ r^* r ∗ and consists of three analytically parametrized arcs. The shuffles of Min considered in [4 ] generate the rightmost arc only for 0 ≤ a ≤ 2 5 0 \leq a \leq\frac{2}{5} 0 ≤ a ≤ 5 2 ; for 2 5 < a < 1 2 \frac{2}{5}<a<\frac{1}{2} 5 2 < a < 2 1 the corresponding curve lies strictly inside the region. The remaining two arcs are generated by singular copulas whose supports contain hyperbola-type curves. The boundary is continuously differentiable away from its left endpoint, and
r ∗ ( α ) = 1 2 + 4 ( α − 11 120 ) 3 / 4 + o ( ( α − 11 120 ) 3 / 4 ) , α ↓ 11 120 , r^*(\alpha)=\frac{1}{2}+4\left(\alpha-\frac{11}{120}\right)^{3/4}+o\left(\left(\alpha-\frac{11}{120}\right)^{3/4}\right),\qquad\alpha\downarrow\frac{11}{120}, r ∗ ( α ) = 2 1 + 4 ( α − 120 11 ) 3/4 + o ( ( α − 120 11 ) 3/4 ) , α ↓ 120 11 ,
so that the boundary leaves ( 11 120 , 1 2 ) \left(\frac{11}{120},\frac{1}{2}\right) ( 120 11 , 2 1 ) with a vertical tangent.
The ( ARV , ρ ) (\operatorname{ARV},\rho) ( ARV , ρ ) region is also of interest in practical applications. In § 5 we show that, for every copula C C C and ( U , V ) ∼ C (U,V)\sim C ( U , V ) ∼ C ,
45 ARV ( C ) − 5 = Corr ( ( 2 U − 1 ) 2 , ( 2 V − 1 ) 2 ) , 45\operatorname{ARV}(C)-5=\operatorname{Corr}\left((2U-1)^2,(2V-1)^2\right), 45 ARV ( C ) − 5 = Corr ( ( 2 U − 1 ) 2 , ( 2 V − 1 ) 2 ) ,
i.e. ARV is an affine transformation of the arachnitude . This index is used in actuarial practice to calibrate the dependence between risk drivers in internal capital models, since it measures whether two variables tend to be simultaneously extreme, in either direction, a feature that Spearman’s rho cannot detect [27 , 1 ] ; it is the second diagonal coefficient of the Legendre expansion of the copula density, Spearman’s rho being the first [20 ] . Consequently, Theorem 4.1 describes exactly which combinations of Spearman’s rho and arachnitude are compatible with some copula, and thus provides a consistency check when both coefficients are estimated from data or specified by expert judgment. For instance, a Spearman’s rho of 0.9 is compatible only with an arachnitude of at least 0.249.
The paper is organized as follows. § 2 recalls aggregation functions, copulas and ARV, introduces the folding transformation and establishes the basic geometry of the region. § 3 constructs the three copula families. § 4 states the main result and its consequences. § 5 establishes the link with arachnitude and discusses examples, while § 6 concludes the paper. The proof of the main result is deferred to Appendix § 11 .
Generative AI tools were used in the research process, for the numerical exploration, for the construction of the extremal families and of the Kantorovich potentials, and for symbolic verifications; all arguments have been checked by the authors. Details are given in the declaration at the end of the paper.
Preliminary definitions and properties
We first recall the basic notions on aggregation functions and copulas and the definition of the average rectangular volume. A (bivariate) aggregation function is a function A : [ 0 , 1 ] 2 → [ 0 , 1 ] A : [0,1]^2 \to[0,1] A : [ 0 , 1 ] 2 → [ 0 , 1 ] that is non-decreasing in each argument and satisfies A ( 0 , 0 ) = 0 A(0,0)=0 A ( 0 , 0 ) = 0 and A ( 1 , 1 ) = 1 A(1,1)=1 A ( 1 , 1 ) = 1 [12 ] . For a function F : [ 0 , 1 ] 2 → R F : [0,1]^2 \to\mathbb{R} F : [ 0 , 1 ] 2 → R and a rectangle
R = [ x 1 , y 1 ] × [ x 2 , y 2 ] , x i ≤ y i , R = [x_1,y_1] \times[x_2,y_2], \qquad x_i \le y_i, R = [ x 1 , y 1 ] × [ x 2 , y 2 ] , x i ≤ y i ,
the F F F -volume of R R R is
V F ( R ) = F ( y 1 , y 2 ) − F ( x 1 , y 2 ) − F ( y 1 , x 2 ) + F ( x 1 , x 2 ) . V_F(R) = F(y_1,y_2) - F(x_1,y_2) - F(y_1,x_2) + F(x_1,x_2). V F ( R ) = F ( y 1 , y 2 ) − F ( x 1 , y 2 ) − F ( y 1 , x 2 ) + F ( x 1 , x 2 ) .
F F F is supermodular (or 2-increasing ) if V F ( R ) ≥ 0 V_F(R) \ge0 V F ( R ) ≥ 0 for every rectangle R R R . A (bivariate) copula is a supermodular aggregation function with uniform margins, C ( u , 1 ) = u C(u,1)=u C ( u , 1 ) = u and C ( 1 , v ) = v C(1,v)=v C ( 1 , v ) = v for u , v ∈ [ 0 , 1 ] u,v \in[0,1] u , v ∈ [ 0 , 1 ] [22 , 9 ] . Let C \mathcal{C} C denote the class of bivariate copulas and let
M ( u , v ) = min { u , v } , W ( u , v ) = max { u + v − 1 , 0 } , Π ( u , v ) = u v . M(u,v) = \min\{u,v\}, \qquad W(u,v) = \max\{u+v-1,0\}, \qquad\Pi(u,v) = uv. M ( u , v ) = min { u , v } , W ( u , v ) = max { u + v − 1 , 0 } , Π ( u , v ) = uv .
We identify a copula with the probability measure it induces on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 . We write ( U , V ) ∼ C (U,V) \sim C ( U , V ) ∼ C to mean that C C C is the joint distribution function of the random pair ( U , V ) (U,V) ( U , V ) , i.e. C ( u , v ) = P ( U ≤ u , V ≤ v ) C(u,v)=\mathbb{P}(U \le u,V \le v) C ( u , v ) = P ( U ≤ u , V ≤ v ) ; in particular, U U U and V V V are uniformly distributed on [ 0 , 1 ] [0,1] [ 0 , 1 ] . In this case V C ( R ) = P ( ( U , V ) ∈ R ) V_C(R)=\mathbb{P}((U,V) \in R) V C ( R ) = P (( U , V ) ∈ R ) for every closed rectangle R R R , since the margins are continuous. Spearman’s rho is
ρ ( C ) = 12 E [ U V ] − 3 = 12 ∫ [ 0 , 1 ] 2 C ( u , v ) d u d v − 3. \rho(C) = 12\mathbb{E}[UV] - 3 = 12 \int_{[0,1]^2} C(u,v)\,\mathrm{d}u\,\mathrm{d}v - 3. ρ ( C ) = 12 E [ U V ] − 3 = 12 ∫ [ 0 , 1 ] 2 C ( u , v ) d u d v − 3.
For x ≤ y \mathbf{x} \le\mathbf{y} x ≤ y in [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 (componentwise), write [ x , y ] = [ x 1 , y 1 ] × [ x 2 , y 2 ] [\mathbf{x},\mathbf{y}] = [x_1,y_1] \times[x_2,y_2] [ x , y ] = [ x 1 , y 1 ] × [ x 2 , y 2 ] .
Definition 2.1 ([3 ] ). The average rectangular volume of an integrable function F : [ 0 , 1 ] 2 → R F : [0,1]^2 \to\mathbb{R} F : [ 0 , 1 ] 2 → R is
ARV ( F ) = 1 ∣ R 4 ∣ ∫ R 4 V F ( [ x , y ] ) d x d y , \operatorname{ARV}(F) = \frac{1}{|\mathcal{R}_4|} \int_{\mathcal{R}_4} V_F([\mathbf{x},\mathbf{y}])\,\mathrm{d}\mathbf{x}\,\mathrm{d}\mathbf{y}, ARV ( F ) = ∣ R 4 ∣ 1 ∫ R 4 V F ([ x , y ]) d x d y , where R 4 = { ( x , y ) ∈ [ 0 , 1 ] 2 × [ 0 , 1 ] 2 : x ≤ y } \mathcal{R}_4 = \{(\mathbf{x},\mathbf{y}) \in[0,1]^2 \times[0,1]^2 : \mathbf{x} \le\mathbf{y}\} R 4 = {( x , y ) ∈ [ 0 , 1 ] 2 × [ 0 , 1 ] 2 : x ≤ y } .
By Fubini’s theorem [3 ] ,
ARV ( F ) = 4 ∫ [ 0 , 1 ] 2 F ( u , v ) ( 1 − 2 u ) ( 1 − 2 v ) d u d v . (1) \operatorname{ARV}(F) = 4 \int_{[0,1]^2} F(u,v)(1-2u)(1-2v)\,\mathrm{d}u\,\mathrm{d}v. \tag*{(1)} ARV ( F ) = 4 ∫ [ 0 , 1 ] 2 F ( u , v ) ( 1 − 2 u ) ( 1 − 2 v ) d u d v . (1)
For a copula, the normalized integral over R 4 \mathcal{R}_4 R 4 equals, by symmetry, the expected C C C -volume of the rectangle R X , Y R_{\mathbf{X},\mathbf{Y}} R X , Y spanned by two independent uniform points X , Y \mathbf{X},\mathbf{Y} X , Y of [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 [3 Remark 2.2] . Since the interval spanned by two independent uniform points of [ 0 , 1 ] [0,1] [ 0 , 1 ] covers u u u with probability 2 u ( 1 − u ) 2u(1-u) 2 u ( 1 − u ) , it holds that
ARV ( C ) = P ( ( U , V ) ∈ R X , Y ) = 4 E [ U ( 1 − U ) V ( 1 − V ) ] , (2) \operatorname{ARV}(C) = \mathbb{P}\left((U,V) \in R_{\mathbf{X},\mathbf{Y}}\right) = 4\mathbb{E}[U(1-U)V(1-V)], \tag*{(2)} ARV ( C ) = P ( ( U , V ) ∈ R X , Y ) = 4 E [ U ( 1 − U ) V ( 1 − V )] , (2)
where ( U , V ) ∼ C (U,V) \sim C ( U , V ) ∼ C is independent of ( X , Y ) (\mathbf{X},\mathbf{Y}) ( X , Y ) . On C \mathcal{C} C , ARV is affine, invariant under the reflection of one argument, and satisfies 11 120 ≤ ARV ( C ) ≤ 1 15 \frac{11}{120} \le\operatorname{ARV}(C) \le\frac{1}{15} 120 11 ≤ ARV ( C ) ≤ 15 1 (see [3 ] ); under independence, ARV ( Π ) = 1 9 \operatorname{ARV}(\Pi)=\frac{1}{9} ARV ( Π ) = 9 1 .
The main tool of this paper is a transformation that exploits the symmetry of ARV with respect to the center of the unit square. Rectangles centered at ( 1 2 , 1 2 ) \left(\frac{1}{2},\frac{1}{2}\right) ( 2 1 , 2 1 ) play a special role: for ( u , v ) ∈ [ 0 , 1 2 ] 2 (u,v) \in\left[0,\frac{1}{2}\right]^2 ( u , v ) ∈ [ 0 , 2 1 ] 2 , let
R u , v = [ u , 1 − u ] × [ v , 1 − v ] R_{u,v}=[u,1-u]\times[v,1-v] R u , v = [ u , 1 − u ] × [ v , 1 − v ]
be the rectangle centered at ( 1 / 2 , 1 / 2 ) (1/2,1/2) ( 1/2 , 1/2 ) .
Let ( U , V ) ∼ C (U,V)\sim C ( U , V ) ∼ C and define
S = ∣ 2 U − 1 ∣ , T = ∣ 2 V − 1 ∣ . S=|2U-1|,\qquad T=|2V-1|. S = ∣2 U − 1∣ , T = ∣2 V − 1∣.
Since U U U and V V V are uniform on [ 0 , 1 ] [0,1] [ 0 , 1 ] , both S S S and T T T are again uniform on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
Definition 2.2. For a given copula C C C , the folded copula of C C C , denoted by C F C^{\mathrm{F}} C F , is the copula of the random pair ( S , T ) = ( ∣ 2 U − 1 ∣ , ∣ 2 V − 1 ∣ ) (S,T)=(|2U-1|,|2V-1|) ( S , T ) = ( ∣2 U − 1∣ , ∣2 V − 1∣ ) .
Notice that this construction is related to the shuffle of copulas via suitable measure-preserving transformations of the unit interval (see [8 ] ).
By flipping one coordinate we mean replacing ( U , V ) (U,V) ( U , V ) by ( 1 − U , V ) (1-U,V) ( 1 − U , V ) or by ( U , 1 − V ) (U,1-V) ( U , 1 − V ) . Such a flip does not change the folded copula; it preserves ARV and changes the sign of ρ \rho ρ .
For a copula D D D , we write E D \mathbb{E}_D E D for expectation with respect to ( S , T ) ∼ D (S,T)\sim D ( S , T ) ∼ D .
Proposition 2.3. For every copula C C C the following statements hold.
(i)
V C ( R u , v ) = C F ( 1 − 2 u , 1 − 2 v ) , ( u , v ) ∈ [ 0 , 1 2 ] 2 . V_C(R_{u,v})=C^{\mathrm{F}}(1-2u,1-2v),\qquad(u,v)\in\left[0,\frac{1}{2}\right]^2. V C ( R u , v ) = C F ( 1 − 2 u , 1 − 2 v ) , ( u , v ) ∈ [ 0 , 2 1 ] 2 . (ii)
ARV ( C ) = ∫ [ 0 , 1 ] 2 s t C F ( s , t ) d s d t = 1 12 + 1 4 E C F [ S 2 T 2 ] . (3) \operatorname{ARV}(C)=\int_{[0,1]^2}stC^{\mathrm{F}}(s,t)\,\mathrm{d}s\,\mathrm{d}t=\frac{1}{12}+\frac{1}{4}\mathbb{E}_{C^{\mathrm{F}}}[S^2T^2]. \tag*{(3)} ARV ( C ) = ∫ [ 0 , 1 ] 2 s t C F ( s , t ) d s d t = 12 1 + 4 1 E C F [ S 2 T 2 ] . (3) (iii)
∣ ρ ( C ) ∣ ≤ 3 E C F [ S T ] . (4) |\rho(C)|\le3\mathbb{E}_{C^{\mathrm{F}}}[ST]. \tag*{(4)} ∣ ρ ( C ) ∣ ≤ 3 E C F [ S T ] . (4) Conversely, for every copula D D D there exists a copula C C C such that
C F = D and ρ ( C ) = 3 E D [ S T ] . C^{\mathrm{F}}=D\qquad\text{and}\qquad\rho(C)=3\mathbb{E}_D[ST]. C F = D and ρ ( C ) = 3 E D [ S T ] .
Proof. Since the marginals of C C C are continuous, V C ( R u , v ) = P ( ( U , V ) ∈ R u , v ) V_C(R_{u,v})=\mathbb{P}((U,V)\in R_{u,v}) V C ( R u , v ) = P (( U , V ) ∈ R u , v ) (boundaries carry no mass). The event ( U , V ) ∈ R u , v (U,V)\in R_{u,v} ( U , V ) ∈ R u , v is equivalent to
S ≤ 1 − 2 u , T ≤ 1 − 2 v , S\le1-2u,\qquad T\le1-2v, S ≤ 1 − 2 u , T ≤ 1 − 2 v ,
because u ≤ x ≤ 1 − u u\le x\le1-u u ≤ x ≤ 1 − u if and only if ∣ 2 x − 1 ∣ ≤ 1 − 2 u |2x-1|\le1-2u ∣2 x − 1∣ ≤ 1 − 2 u . Since S S S and T T T are uniform, the joint distribution function of ( S , T ) (S,T) ( S , T ) coincides with C F C^{\mathrm{F}} C F on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 , and (i) follows.
To prove (ii), recall that the average over R 4 \mathcal{R}_4 R 4 in the definition of ARV equals, by symmetry, the expected C C C -measure of the random rectangle R X , Y R_{X,Y} R X , Y spanned by two independent uniform points X , Y X,Y X , Y of [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 [3 Remark 2.2] . Fix a point ( x , y ) (x,y) ( x , y ) in the unit square. The probability that the random rectangle covers this point is 4 x ( 1 − x ) y ( 1 − y ) 4x(1-x)y(1-y) 4 x ( 1 − x ) y ( 1 − y ) (each coordinate interval covers the corresponding coordinate with probability 2 x ( 1 − x ) 2x(1-x) 2 x ( 1 − x ) , resp. 2 y ( 1 − y ) 2y(1-y) 2 y ( 1 − y ) . It therefore holds that
ARV ( C ) = 4 E [ U ( 1 − U ) V ( 1 − V ) ] = 1 4 E [ ( 1 − S 2 ) ( 1 − T 2 ) ] = 1 12 + 1 4 E [ S 2 T 2 ] . \operatorname{ARV}(C)=4\mathbb{E}[U(1-U)V(1-V)]=\frac{1}{4}\mathbb{E}[(1-S^2)(1-T^2)]=\frac{1}{12}+\frac{1}{4}\mathbb{E}[S^2T^2]. ARV ( C ) = 4 E [ U ( 1 − U ) V ( 1 − V )] = 4 1 E [( 1 − S 2 ) ( 1 − T 2 )] = 12 1 + 4 1 E [ S 2 T 2 ] .
(because 4 U ( 1 − U ) = 1 − S 2 4U(1-U)=1-S^2 4 U ( 1 − U ) = 1 − S 2 and 4 V ( 1 − V ) = 1 − T 2 4V(1-V)=1-T^2 4 V ( 1 − V ) = 1 − T 2 ). Here we used E [ S 2 ] = E [ T 2 ] = 1 / 3 \mathbb{E}[S^2]=\mathbb{E}[T^2]=1/3 E [ S 2 ] = E [ T 2 ] = 1/3 . Also, for any copula D D D ,
∫ [ 0 , 1 ] 2 s t D ( s , t ) d s d t = E D [ ∫ S 1 s d s ∫ T 1 t d t ] = 1 4 E D [ ( 1 − S 2 ) ( 1 − T 2 ) ] . \int_{[0,1]^2}st\,D(s,t)\,ds\,dt=\mathbb{E}_{D}\left[\int_{S}^{1}s\,ds\int_{T}^{1}t\,dt\right]=\frac{1}{4}\mathbb{E}_{D}[(1-S^2)(1-T^2)]. ∫ [ 0 , 1 ] 2 s t D ( s , t ) d s d t = E D [ ∫ S 1 s d s ∫ T 1 t d t ] = 4 1 E D [( 1 − S 2 ) ( 1 − T 2 )] .
This proves both equalities in (3) .
For (iii), uniformity of the marginals gives
ρ ( C ) = 3 E [ ( 2 U − 1 ) ( 2 V − 1 ) ] . \rho(C)=3\mathbb{E}[(2U-1)(2V-1)]. ρ ( C ) = 3 E [( 2 U − 1 ) ( 2 V − 1 )] .
Consequently,
∣ ρ ( C ) ∣ ≤ 3 E [ ∣ 2 U − 1 ∣ ∣ 2 V − 1 ∣ ] = 3 E C F [ S T ] . |\rho(C)|\leq3\mathbb{E}[|2U-1||2V-1|]=3\mathbb{E}_{C^{\mathrm{F}}}[ST]. ∣ ρ ( C ) ∣ ≤ 3 E [ ∣2 U − 1∣∣2 V − 1∣ ] = 3 E C F [ S T ] .
For the converse, let ( S , T ) ∼ D (S,T)\sim D ( S , T ) ∼ D , let ε \varepsilon ε be an independent symmetric random sign, and set
U = 1 + ε S 2 , V = 1 + ε T 2 . U=\frac{1+\varepsilon S}{2},\qquad V=\frac{1+\varepsilon T}{2}. U = 2 1 + εS , V = 2 1 + εT .
Then U U U and V V V are uniform on [ 0 , 1 ] [0,1] [ 0 , 1 ] , the copula C C C of ( U , V ) (U,V) ( U , V ) satisfies C F = D C^{\mathrm{F}}=D C F = D , and
E [ U V ] = 1 4 ( 1 + E D [ S T ] ) . \mathbb{E}[UV]=\frac{1}{4}(1+\mathbb{E}_{D}[ST]). E [ U V ] = 4 1 ( 1 + E D [ S T ]) .
Therefore
ρ ( C ) = 12 E [ U V ] − 3 = 3 E D [ S T ] . \rho(C)=12\mathbb{E}[UV]-3=3\mathbb{E}_{D}[ST]. ρ ( C ) = 12 E [ U V ] − 3 = 3 E D [ S T ] .
We denote this particular unfolding by D U D^{\mathrm{U}} D U . Explicitly, conditioning on ε \varepsilon ε ,
D U ( u , v ) = 1 2 [ D ( ( 2 u − 1 ) + , ( 2 v − 1 ) + ) + D ( ( 1 − 2 u ) + , ( 1 − 2 v ) + ) + 1 − ( 1 − 2 u ) + − ( 1 − 2 v ) + ] . D^{\mathrm{U}}(u,v)=\frac{1}{2}\left[D\left((2u-1)^{+},(2v-1)^{+}\right)+D\left((1-2u)^{+},(1-2v)^{+}\right)+1-(1-2u)^{+}-(1-2v)^{+}\right]. D U ( u , v ) = 2 1 [ D ( ( 2 u − 1 ) + , ( 2 v − 1 ) + ) + D ( ( 1 − 2 u ) + , ( 1 − 2 v ) + ) + 1 − ( 1 − 2 u ) + − ( 1 − 2 v ) + ] .
i.e. D U D^{\mathrm{U}} D U is the ordinal sum of the survival copula of D D D on [ 0 , 1 2 ] 2 [0,\frac{1}{2}]^2 [ 0 , 2 1 ] 2 and of D D D on [ 1 2 , 1 ] 2 [\frac{1}{2},1]^2 [ 2 1 , 1 ] 2 . It uses a common independent sign in the two coordinates, so
( D U ) F = D , ρ ( D U ) = 3 E D [ S T ] . (D^{\mathrm{U}})^{\mathrm{F}}=D,\qquad\rho(D^{\mathrm{U}})=3\mathbb{E}_{D}[ST]. ( D U ) F = D , ρ ( D U ) = 3 E D [ S T ] .
This construction will allow us to specify the extremal copulas on the simpler folded square and then recover their original supports.
With these tools at hand, we can establish the basic geometric properties of the region under study, namely the set of all pairs of values of the average rectangular volume and of Spearman’s rho that are attained by some copula:
A = { ( ARV ( C ) , ρ ( C ) ) : C ∈ C } ⊆ R 2 . \mathcal{A}=\{(\operatorname{ARV}(C),\rho(C)):C\in\mathcal{C}\}\subseteq\mathbb{R}^{2}. A = {( ARV ( C ) , ρ ( C )) : C ∈ C } ⊆ R 2 .
Throughout, we denote by ( α , r ) (\alpha,r) ( α , r ) the generic point of the plane, where the first coordinate α \alpha α refers to ARV and the second coordinate r r r to Spearman’s rho. The class C \mathcal{C} C is compact with respect to the uniform convergence. Both functionals are continuous and affine: this is immediate from their representations as integrals of bounded continuous functions. Thus A \mathcal{A} A is compact and convex. Reflection in one coordinate preserves ARV and reverses the sign of ρ \rho ρ , so A \mathcal{A} A is symmetric with respect to the horizontal axis { ( α , r ) : r = 0 } \{(\alpha,r):r=0\} {( α , r ) : r = 0 } , i.e. ( α , r ) ∈ A (\alpha,r)\in\mathcal{A} ( α , r ) ∈ A if and only if ( α , − r ) ∈ A (\alpha,-r)\in\mathcal{A} ( α , − r ) ∈ A .
For any copula D D D (in the role of a folded copula), the Fréchet-Hoeffding bound inequalities W ≤ D ≤ M W\leq D\leq M W ≤ D ≤ M and (3) give, since 1 12 + 1 4 E D [ S 2 T 2 ] = ∫ s t D ( s , t ) d s d t \frac{1}{12}+\frac{1}{4}\mathbb{E}_{D}[S^{2}T^{2}]=\int st\,D(s,t)\,ds\,dt 12 1 + 4 1 E D [ S 2 T 2 ] = ∫ s t D ( s , t ) d s d t is monotone in D D D ,
11 120 ≤ 1 12 + 1 4 E D [ S 2 T 2 ] ≤ 2 15 . \frac{11}{120}\leq\frac{1}{12}+\frac{1}{4}\mathbb{E}_{D}[S^{2}T^{2}]\leq\frac{2}{15}. 120 11 ≤ 12 1 + 4 1 E D [ S 2 T 2 ] ≤ 15 2 .
Indeed, E W [ S 2 T 2 ] = ∫ 0 1 s 2 ( 1 − s ) 2 d s = 1 / 30 \mathbb{E}_{W}[S^{2}T^{2}]=\int_{0}^{1}s^{2}(1-s)^{2}\,ds=1/30 E W [ S 2 T 2 ] = ∫ 0 1 s 2 ( 1 − s ) 2 d s = 1/30 and E M [ S 2 T 2 ] = ∫ 0 1 s 4 d s = 1 / 5 \mathbb{E}_{M}[S^{2}T^{2}]=\int_{0}^{1}s^{4}\,ds=1/5 E M [ S 2 T 2 ] = ∫ 0 1 s 4 d s = 1/5 . Both bounds are attained by unfolding W W W and M M M , respectively; convexity gives every intermediate value. Moreover, equality at either endpoint forces D = W D=W D = W or D = M D=M D = M : the non-negative continuous function D − W D-W D − W or M − D M-D M − D has zero integral against the strictly positive weight s t st s t in the open square, and hence vanishes everywhere by continuity.
Compactness and Proposition 2.3 now give
r ∗ ( α ) : = max { ρ ( C ) : ARV ( C ) = α } = max { 3 E D [ S T ] : D ∈ C , 1 12 + 1 4 E D [ S 2 T 2 ] = α } , (5) \begin{aligned}
r^{*}(\alpha)&:=\max\{\rho(C):\operatorname{ARV}(C)=\alpha\}\\
&=\max\left\{3\mathbb{E}_{D}[ST]:D\in\mathcal{C},\ \frac{1}{12}+\frac{1}{4}\mathbb{E}_{D}[S^{2}T^{2}]=\alpha\right\}, \tag*{(5)}
\end{aligned} r ∗ ( α ) := max { ρ ( C ) : ARV ( C ) = α } = max { 3 E D [ S T ] : D ∈ C , 12 1 + 4 1 E D [ S 2 T 2 ] = α } , (5)
for α ∈ [ 11 / 120 , 2 / 15 ] \alpha\in[11/120,2/15] α ∈ [ 11/120 , 2/15 ] . Here “≤ \leq ≤ ” in the second equality follows from Proposition 2.3 (ii)–(iii) applied to D = C ♭ D=C^{\flat} D = C ♭ , “≥ \geq ≥ ” from the unfolding D ∪ D^{\cup} D ∪ , and both maxima are attained because the constraint sets are closed subsets of the compact set C \mathcal{C} C and the functionals are continuous.
Proposition 2.4. The region is
A = { ( α , r ) : 11 120 ≤ α ≤ 2 15 , ∣ r ∣ ≤ r ∗ ( α ) } . \mathcal{A}=\left\{(\alpha,r):\frac{11}{120}\leq\alpha\leq\frac{2}{15},\ |r|\leq r^{*}(\alpha)\right\}. A = { ( α , r ) : 120 11 ≤ α ≤ 15 2 , ∣ r ∣ ≤ r ∗ ( α ) } . The function r ∗ r^{*} r ∗ is continuous, concave and strictly increasing, with
r ∗ ( 11 / 120 ) = 1 2 , r ∗ ( 2 / 15 ) = 1. r^{*}(11/120)=\frac{1}{2},\qquad r^{*}(2/15)=1. r ∗ ( 11/120 ) = 2 1 , r ∗ ( 2/15 ) = 1.
Proof. Every vertical section of A \mathcal{A} A over [ 11 / 120 , 2 / 15 ] [11/120,2/15] [ 11/120 , 2/15 ] (the projection of A \mathcal{A} A on the first axis, by the preceding paragraph) is a nonempty compact interval, symmetric about zero, which proves the displayed description. Convexity gives concavity of r ∗ r^{*} r ∗ . Compactness gives upper semicontinuity, and concavity gives continuity in the interior. At each endpoint, comparison with a chord to a fixed interior point gives the complementary lower limit inequality, so r ∗ r^{*} r ∗ is continuous on the closed interval as well.
The endpoint characterization above and (5) yield
r ∗ ( 11 / 120 ) = 3 ∫ 0 1 u ( 1 − u ) d u = 1 2 , r ∗ ( 2 / 15 ) = 3 ∫ 0 1 u 2 d u = 1. r^{*}(11/120)=3\int_{0}^{1}u(1-u)\,du=\frac{1}{2},\qquad r^{*}(2/15)=3\int_{0}^{1}u^{2}\,du=1. r ∗ ( 11/120 ) = 3 ∫ 0 1 u ( 1 − u ) d u = 2 1 , r ∗ ( 2/15 ) = 3 ∫ 0 1 u 2 d u = 1.
Finally, C ≤ M C \le M C ≤ M and the integral formula for ρ \rho ρ imply ρ ( C ) ≤ 1 \rho(C) \le1 ρ ( C ) ≤ 1 , with equality only for C = M C = M C = M . Thus r ∗ ( α ) < 1 r^*(\alpha) < 1 r ∗ ( α ) < 1 when α < 2 / 15 \alpha< 2/15 α < 2/15 . For α 1 < α 2 < 2 / 15 \alpha_1 < \alpha_2 < 2/15 α 1 < α 2 < 2/15 , write α 2 = ( 1 − λ ) α 1 + λ ( 2 / 15 ) \alpha_2 = (1-\lambda)\alpha_1 + \lambda(2/15) α 2 = ( 1 − λ ) α 1 + λ ( 2/15 ) with 0 < λ < 1 0 < \lambda< 1 0 < λ < 1 . Concavity gives
r ∗ ( α 2 ) ≥ ( 1 − λ ) r ∗ ( α 1 ) + λ > r ∗ ( α 1 ) . r^*(\alpha_2) \ge(1-\lambda)r^*(\alpha_1) + \lambda> r^*(\alpha_1). r ∗ ( α 2 ) ≥ ( 1 − λ ) r ∗ ( α 1 ) + λ > r ∗ ( α 1 ) .
The comparison with the right endpoint is immediate.
Three families of copulas
We now construct the copulas that will generate the upper boundary of the region A \mathcal{A} A . Family I is a shuffle-of-Min family already considered in [4 ] ; see also [21 , 8 ] for background on shuffles. Families II and III are constructed on the folded square and then unfolded. This makes their marginal constraints and moment formulas transparent. Optimality will be proved in Appendix § 11 .
Family I: shuffles of Min
For a ∈ [ 0 , 1 2 ] a \in[0,\frac{1}{2}] a ∈ [ 0 , 2 1 ] , define
C a I ( u , v ) = { max { u + v − a , 0 } , ( u , v ) ∈ [ 0 , a ] 2 , max { u + v − 1 , 1 − a } , ( u , v ) ∈ [ 1 − a , 1 ] 2 , M ( u , v ) , otherwise . C_a^{\mathrm{I}}(u,v) =
\begin{cases}
\max\{u+v-a,0\}, & (u,v) \in[0,a]^2,\\
\max\{u+v-1,1-a\}, & (u,v) \in[1-a,1]^2,\\
M(u,v), & \text{otherwise}.
\end{cases} C a I ( u , v ) = ⎩ ⎨ ⎧ max { u + v − a , 0 } , max { u + v − 1 , 1 − a } , M ( u , v ) , ( u , v ) ∈ [ 0 , a ] 2 , ( u , v ) ∈ [ 1 − a , 1 ] 2 , otherwise .
Its flipped version is denoted by C ~ a I \widetilde{C}_a^{\mathrm{I}} C a I and is given by
C ~ a I ( u , v ) = u − C a I ( u , 1 − v ) = v − C a I ( 1 − u , v ) . \widetilde{C}_a^{\mathrm{I}}(u,v) = u - C_a^{\mathrm{I}}(u,1-v) = v - C_a^{\mathrm{I}}(1-u,v). C a I ( u , v ) = u − C a I ( u , 1 − v ) = v − C a I ( 1 − u , v ) .
The support of C a I C_a^{\mathrm{I}} C a I is
{ u + v = a } ∩ [ 0 , a ] 2 ∪ { u = v } ∩ [ a , 1 − a ] 2 ∪ { u + v = 2 − a } ∩ [ 1 − a , 1 ] 2 . \{u+v=a\}\cap[0,a]^2 \mathbin{\cup} \{u=v\}\cap[a,1-a]^2 \mathbin{\cup} \{u+v=2-a\}\cap[1-a,1]^2. { u + v = a } ∩ [ 0 , a ] 2 ∪ { u = v } ∩ [ a , 1 − a ] 2 ∪ { u + v = 2 − a } ∩ [ 1 − a , 1 ] 2 .
The central segment is omitted at the endpoint a = 1 / 2 a=1/2 a = 1/2 . The support of C ~ a I \widetilde{C}_a^{\mathrm{I}} C a I is obtained by flipping one coordinate.
In particular,
C 0 I = M , C ~ 0 I = W , C_0^{\mathrm{I}}=M,\qquad\widetilde{C}_0^{\mathrm{I}}=W, C 0 I = M , C 0 I = W ,
whereas, for later reference, we set
C 2 : = C 1 / 2 I , C 3 : = C ~ 1 / 2 I . C_2 := C_{1/2}^{\mathrm{I}},\qquad C_3 := \widetilde{C}_{1/2}^{\mathrm{I}}. C 2 := C 1/2 I , C 3 := C 1/2 I .
The supports of C 2 C_2 C 2 and C 3 C_3 C 3 are shown in Figure 1 .
Figure 1. Supports of the endpoint copulas of Family I: (a) C 2 = C 1 / 2 I C_2=C^{{\rm I}}_{1/2} C 2 = C 1/2 I and (b) C 3 = C ~ 1 / 2 I C_3=\widetilde{C}^{{\rm I}}_{1/2} C 3 = C 1/2 I . The support of C 3 C_3 C 3 is the reflection of that of C 2 C_2 C 2 about v = 1 / 2 v=1/2 v = 1/2 . Each segment carries probability mass 1 / 2 1/2 1/2 .
A direct computation, detailed below, gives
ARV ( C a I ) = 2 15 − 4 3 ( 1 − a ) 2 a 3 , ρ ( C a I ) = 1 − 4 a 3 . (6) \operatorname{ARV}(C_a^{\mathrm{I}})=\frac{2}{15}-\frac{4}{3}(1-a)^2a^3,\qquad\rho(C_a^{\mathrm{I}})=1-4a^3. \tag*{(6)} ARV ( C a I ) = 15 2 − 3 4 ( 1 − a ) 2 a 3 , ρ ( C a I ) = 1 − 4 a 3 . (6)
Since ARV is invariant under flipping, while ρ \rho ρ changes sign,
ARV ( C ~ a I ) = ARV ( C a I ) , ρ ( C ~ a I ) = − ρ ( C a I ) . \operatorname{ARV}(\widetilde{C}_a^{\mathrm{I}})=\operatorname{ARV}(C_a^{\mathrm{I}}),\qquad\rho(\widetilde{C}_a^{\mathrm{I}})=-\rho(C_a^{\mathrm{I}}). ARV ( C a I ) = ARV ( C a I ) , ρ ( C a I ) = − ρ ( C a I ) .
The folded representation of this family is particularly simple. For b ∈ [ 0 , 1 ] b\in[0,1] b ∈ [ 0 , 1 ] , let D b D_b D b be the copula of ( S , T b ( S ) ) (S,T_b(S)) ( S , T b ( S )) , where S S S is uniform and
T b ( s ) = { s , 0 ≤ s ≤ b , 1 + b − s , b < s ≤ 1. T_b(s)=\begin{cases}
s, & 0\le s\le b,\\
1+b-s, & b<s\le1.
\end{cases} T b ( s ) = { s , 1 + b − s , 0 ≤ s ≤ b , b < s ≤ 1.
Thus D b D_b D b is the ordinal sum of a comonotone part on [ 0 , b ] 2 [0,b]^2 [ 0 , b ] 2 and a countermonotone part on [ b , 1 ] 2 [b,1]^2 [ b , 1 ] 2 . Directly from the supports,
( C a I ) F = D 1 − 2 a , C a I = D 1 − 2 a U . (7) (C^{\rm I}_a)^{\rm F}=D_{1-2a},\qquad C^{\rm I}_a=D^{\rm U}_{1-2a}. \tag*{(7)} ( C a I ) F = D 1 − 2 a , C a I = D 1 − 2 a U . (7)
In particular, D 0 = W D_0=W D 0 = W and D 1 = M D_1=M D 1 = M .
The moments of Family I are best computed on the folded square. By (7) , it suffices to compute the moments of D b D_b D b . Since D b D_b D b is concentrated on the diagonal of [ 0 , b ] 2 [0,b]^2 [ 0 , b ] 2 and on the anti-diagonal of [ b , 1 ] 2 [b,1]^2 [ b , 1 ] 2 ,
E D b [ S T ] = b 3 3 + ∫ b 1 s ( 1 + b − s ) d s = 1 + 3 b − 3 b 2 + b 3 6 , \mathbb{E}_{D_b}[ST]=\frac{b^3}{3}+\int_b^1s(1+b-s)\,ds=\frac{1+3b-3b^2+b^3}{6}, E D b [ S T ] = 3 b 3 + ∫ b 1 s ( 1 + b − s ) d s = 6 1 + 3 b − 3 b 2 + b 3 ,
E D b [ S 2 T 2 ] = b 5 5 + ∫ b 1 s 2 ( 1 + b − s ) 2 d s = 1 + 5 b + 10 b 2 − 10 b 3 − 5 b 4 + 5 b 5 30 . \mathbb{E}_{D_b}[S^2T^2]=\frac{b^5}{5}+\int_b^1s^2(1+b-s)^2\,ds=\frac{1+5b+10b^2-10b^3-5b^4+5b^5}{30}. E D b [ S 2 T 2 ] = 5 b 5 + ∫ b 1 s 2 ( 1 + b − s ) 2 d s = 30 1 + 5 b + 10 b 2 − 10 b 3 − 5 b 4 + 5 b 5 .
Substituting b = 1 − 2 a b=1-2a b = 1 − 2 a in (3) and in ρ ( C a I ) = 3 E D b [ S T ] \rho(C^{\rm I}_a)=3\mathbb{E}_{D_b}[ST] ρ ( C a I ) = 3 E D b [ S T ] gives (6) . Moreover,
d d b E D b [ S T ] = ( 1 − b ) 2 2 , \frac{d}{db}\mathbb{E}_{D_b}[ST]=\frac{(1-b)^2}{2}, d b d E D b [ S T ] = 2 ( 1 − b ) 2 ,
d d b E D b [ S 2 T 2 ] = ( 1 − b ) 2 ( 1 + b ) ( 1 + 5 b ) 6 , \frac{d}{db}\mathbb{E}_{D_b}[S^2T^2]=\frac{(1-b)^2(1+b)(1+5b)}{6}, d b d E D b [ S 2 T 2 ] = 6 ( 1 − b ) 2 ( 1 + b ) ( 1 + 5 b ) ,
which are the derivatives used in Remark Appendix § 11 .§ 3 and in the proof of Proposition Appendix § 11 .§ 4 .
Family II: a construction on the folded square
Let 0 < b ≤ 1 / 5 0<b\le1/5 0 < b ≤ 1/5 and 1 / 2 + b ≤ c ≤ 1 1/2+b\le c\le1 1/2 + b ≤ c ≤ 1 , and set
κ = c ( 1 + 2 b − c ) , b ′ = 1 + b − c . \kappa=c(1+2b-c),\qquad b'=1+b-c. κ = c ( 1 + 2 b − c ) , b ′ = 1 + b − c .
For c ≤ t ≤ 1 c \le t \le1 c ≤ t ≤ 1 , define
ξ 1 ( t ) = 1 2 ( κ t + t − 1 ) , ξ 2 ( t ) = 1 2 ( κ t − t + 1 ) , \xi_1(t) = \frac{1}{2}\left(\frac{\kappa}{t}+t-1\right), \qquad\xi_2(t) = \frac{1}{2}\left(\frac{\kappa}{t}-t+1\right), ξ 1 ( t ) = 2 1 ( t κ + t − 1 ) , ξ 2 ( t ) = 2 1 ( t κ − t + 1 ) ,
and
w 1 ( t ) = ξ 1 ′ ( t ) = 1 2 ( 1 − κ t 2 ) , w 2 ( t ) = − ξ 2 ′ ( t ) = 1 2 ( 1 + κ t 2 ) . w_1(t) = \xi_1'(t) = \frac{1}{2}\left(1-\frac{\kappa}{t^2}\right), \qquad w_2(t) = -\xi_2'(t) = \frac{1}{2}\left(1+\frac{\kappa}{t^2}\right). w 1 ( t ) = ξ 1 ′ ( t ) = 2 1 ( 1 − t 2 κ ) , w 2 ( t ) = − ξ 2 ′ ( t ) = 2 1 ( 1 + t 2 κ ) .
The parameter restrictions imply
0 ≤ b ≤ κ / 2 ≤ b ′ ≤ c , w 1 ≥ 0 , w 2 > 0 , w 1 + w 2 = 1 0 \le b \le\kappa/2 \le b' \le c,\qquad w_1 \ge0,\quad w_2 > 0,\quad w_1+w_2=1 0 ≤ b ≤ κ /2 ≤ b ′ ≤ c , w 1 ≥ 0 , w 2 > 0 , w 1 + w 2 = 1
(see (A.14)–(A.15) below: w 1 ≥ 0 w_1 \ge0 w 1 ≥ 0 is equivalent to c 2 ≥ κ c^2 \ge\kappa c 2 ≥ κ , i.e. to c ≥ 1 2 + b c \ge\frac{1}{2}+b c ≥ 2 1 + b ).
The increasing branch ξ 1 \xi_1 ξ 1 joins b b b to κ / 2 \kappa/2 κ /2 , whereas the decreasing branch ξ 2 \xi_2 ξ 2 joins b ′ b' b ′ to κ / 2 \kappa/2 κ /2 .
We define a measure D b , c D_{b,c} D b , c through its integrals against bounded Borel functions f f f :
∫ f d D b , c = ∫ 0 b f ( s , s ) d s + ∫ b c f ( s , 1 + b − s ) d s + ∑ i = 1 2 ∫ c 1 w i ( t ) [ f ( ξ i ( t ) , t ) + f ( t , ξ i ( t ) ) ] d t . (8) \int f\,\mathrm{d}D_{b,c}=\int_0^b f(s,s)\,\mathrm{d}s+\int_b^c f(s,1+b-s)\,\mathrm{d}s+\sum_{i=1}^2\int_c^1 w_i(t)\left[f(\xi_i(t),t)+f(t,\xi_i(t))\right]\,\mathrm{d}t. \tag*{(8)} ∫ f d D b , c = ∫ 0 b f ( s , s ) d s + ∫ b c f ( s , 1 + b − s ) d s + i = 1 ∑ 2 ∫ c 1 w i ( t ) [ f ( ξ i ( t ) , t ) + f ( t , ξ i ( t )) ] d t . (8)
This formula specifies both the support and the mass carried by each branch; the support alone would not determine the measure.
Figure 2. The construction of Family II for b = 0.1 b=0.1 b = 0.1 and c = c ( b ) ≈ 0.8904 c=c(b)\approx0.8904 c = c ( b ) ≈ 0.8904 (the parameters of the boundary arc Γ I I \Gamma^{\mathrm{II}} Γ II ; see § 4 ). (a) Support of the folded copula D b , c D_{b,c} D b , c : comonotone block, anti-diagonal and the two hyperbolic branches with their transposes. (b) Zoom on the region where the pieces are glued together. (c) Support of the unfolded copula C b , c I I C^{\mathrm{II}}_{b,c} C b , c II .
Lemma 3.2. The measure D b , c D_{b,c} D b , c is a symmetric copula measure.
Proof. All terms in (8) are non-negative when f ≥ 0 f\geq0 f ≥ 0 . For the first marginal, the diagonal and anti-diagonal terms give Lebesgue measure on [ 0 , b ] [0,b] [ 0 , b ] and [ b ′ , c ] [b',c] [ b ′ , c ] , respectively. Changing variables s = ξ 1 ( t ) s=\xi_1(t) s = ξ 1 ( t ) and s = ξ 2 ( t ) s=\xi_2(t) s = ξ 2 ( t ) in the untransposed curved terms gives Lebesgue measure on [ b , κ / 2 ] [b,\kappa/2] [ b , κ /2 ] and [ κ / 2 , b ′ ] [\kappa/2,b'] [ κ /2 , b ′ ] . The transposed terms give density w 1 + w 2 = 1 w_1+w_2=1 w 1 + w 2 = 1 on [ c , 1 ] [c,1] [ c , 1 ] . These intervals partition [ 0 , 1 ] [0,1] [ 0 , 1 ] , up to endpoints, so the first marginal is uniform and the total mass is one. The measure is invariant under transposition: this is explicit for the curved terms and follows from s ↦ 1 + b − s s\mapsto1+b-s s ↦ 1 + b − s for the anti-diagonal term. Its second marginal is therefore uniform as well. When c = 1 c=1 c = 1 , the curved integrals vanish and the same argument applies with the corresponding empty intervals omitted.
Define the copula of Family II by
C b , c I I : = D b , c u . C^{\mathrm{II}}_{b,c}:=D^{\mathrm{u}}_{b,c}. C b , c II := D b , c u .
Thus its measure is obtained by mapping ( s , t ) (s,t) ( s , t ) to ( ( 1 + s ) / 2 , ( 1 + t ) / 2 ) ((1+s)/2,(1+t)/2) (( 1 + s ) /2 , ( 1 + t ) /2 ) or ( ( 1 − s ) / 2 , ( 1 − t ) / 2 ) ((1-s)/2,(1-t)/2) (( 1 − s ) /2 , ( 1 − t ) /2 ) , each with probability 1 / 2 1/2 1/2 . The resulting copula is singular and radially symmetric. For nondegenerate parameter values, its support consists of a central diagonal segment, two anti-diagonal segments and eight curvilinear arcs. The latter are images of hyperbolic branches: for example, s = ξ 1 ( t ) s=\xi_1(t) s = ξ 1 ( t ) satisfies 2 s t = t 2 − t + κ 2st=t^2-t+\kappa 2 s t = t 2 − t + κ . When c = 1 c=1 c = 1 , the curved pieces disappear.
For clarity, write
D b = { { ( s , s ) : 0 ≤ s ≤ b } , b > 0 , ∅ , b = 0. D_b=\begin{cases}
\{(s,s):0\leq s\leq b\}, & b>0,\\
\varnothing, & b=0.
\end{cases} D b = { {( s , s ) : 0 ≤ s ≤ b } , ∅ , b > 0 , b = 0.
The support on the folded square, including the extension to b = 0 b=0 b = 0 below, is
Σ b , c = supp ( D b , c ) = D b ∪ { ( s , 1 + b − s ) : b ′ ≤ s ≤ c } ∪ ⋃ i = 1 2 { ( ξ i ( t ) , t ) , ( t , ξ i ( t ) ) : c ≤ t ≤ 1 } . (9) \Sigma_{b,c}=\operatorname{supp}(D_{b,c})=\mathcal{D}_b\cup\{(s,1+b-s):b'\le s\le c\}
\cup\bigcup_{i=1}^{2}\{(\xi_i(t),t),(t,\xi_i(t)):c\le t\le1\}. \tag*{(9)} Σ b , c = supp ( D b , c ) = D b ∪ {( s , 1 + b − s ) : b ′ ≤ s ≤ c } ∪ i = 1 ⋃ 2 {( ξ i ( t ) , t ) , ( t , ξ i ( t )) : c ≤ t ≤ 1 } . (9)
In particular, when b = 0 b=0 b = 0 the origin is not an additional support point. The moment calculation can be performed before unfolding. For n = 1 , 2 n=1,2 n = 1 , 2 ,
E D b , c [ ( S T ) n ] = b 2 n + 1 2 n + 1 + ∫ b ′ c [ s ( 1 + b − s ) ] n d s + 2 ∑ i = 1 2 ∫ c 1 w i ( t ) [ t ξ i ( t ) ] n d t . (10) \mathbb{E}_{D_{b,c}}[(ST)^n]=\frac{b^{2n+1}}{2n+1}+\int_{b'}^{c}[s(1+b-s)]^n\,\mathrm{d}s+2\sum_{i=1}^{2}\int_c^1w_i(t)[t\xi_i(t)]^n\,\mathrm{d}t. \tag*{(10)} E D b , c [( S T ) n ] = 2 n + 1 b 2 n + 1 + ∫ b ′ c [ s ( 1 + b − s ) ] n d s + 2 i = 1 ∑ 2 ∫ c 1 w i ( t ) [ t ξ i ( t ) ] n d t . (10)
The integrals in (10) are elementary. On the curved branches, a direct computation gives
∑ i = 1 2 w i ( t ) t ξ i ( t ) = κ 2 t , ∑ i = 1 2 w i ( t ) [ t ξ i ( t ) ] 2 = t 2 ( 1 − t ) 2 4 + κ 2 ( 2 − t ) 4 t , \sum_{i=1}^{2}w_i(t)t\xi_i(t)=\frac{\kappa}{2t},\qquad\sum_{i=1}^{2}w_i(t)[t\xi_i(t)]^2=\frac{t^2(1-t)^2}{4}+\frac{\kappa^2(2-t)}{4t}, i = 1 ∑ 2 w i ( t ) t ξ i ( t ) = 2 t κ , i = 1 ∑ 2 w i ( t ) [ t ξ i ( t ) ] 2 = 4 t 2 ( 1 − t ) 2 + 4 t κ 2 ( 2 − t ) ,
for c ≤ t ≤ 1 c\le t\le1 c ≤ t ≤ 1 . Hence (10) becomes
E D b , c [ S T ] = b 3 3 + ∫ b ′ c s ( 1 + b − s ) d s − κ log c , \mathbb{E}_{D_{b,c}}[ST]=\frac{b^3}{3}+\int_{b'}^c s(1+b-s)\,\mathrm{d}s-\kappa\log c, E D b , c [ S T ] = 3 b 3 + ∫ b ′ c s ( 1 + b − s ) d s − κ log c ,
E D b , c [ S 2 T 2 ] = b 5 5 + ∫ b ′ c s 2 ( 1 + b − s ) 2 d s + 1 2 ∫ c 1 t 2 ( 1 − t ) 2 d t − κ 2 log c − κ 2 ( 1 − c ) 2 . \mathbb{E}_{D_{b,c}}[S^2T^2]=\frac{b^5}{5}+\int_{b'}^c s^2(1+b-s)^2\,\mathrm{d}s+\frac{1}{2}\int_c^1t^2(1-t)^2\,\mathrm{d}t-\kappa^2\log c-\frac{\kappa^2(1-c)}{2}. E D b , c [ S 2 T 2 ] = 5 b 5 + ∫ b ′ c s 2 ( 1 + b − s ) 2 d s + 2 1 ∫ c 1 t 2 ( 1 − t ) 2 d t − κ 2 log c − 2 κ 2 ( 1 − c ) .
With (3) , ρ ( C b , c I I ) = 3 E D b , c [ S T ] \rho(C_{b,c}^{\mathrm{II}})=3\mathbb{E}_{D_{b,c}}[ST] ρ ( C b , c II ) = 3 E D b , c [ S T ] and κ = c ( 1 + 2 b − c ) \kappa=c(1+2b-c) κ = c ( 1 + 2 b − c ) , the remaining polynomial integrals give
ARV ( C b , c I I ) = 19 240 + b 5 24 − b 4 24 − b 3 12 − b 2 12 − b 24 + 2 b 2 c 3 3 − b 2 c 2 2 − 3 b c 4 4 + 4 b c 3 3 − b c 2 2 + c 5 5 − 9 c 4 16 + c 3 2 − c 2 8 − c 2 4 ( c − 1 − 2 b ) 2 log c . (11) \operatorname{ARV}(C_{b,c}^{\mathrm{II}})=\frac{19}{240}+\frac{b^5}{24}-\frac{b^4}{24}-\frac{b^3}{12}-\frac{b^2}{12}-\frac{b}{24}
+\frac{2b^2c^3}{3}-\frac{b^2c^2}{2}-\frac{3bc^4}{4}+\frac{4bc^3}{3}-\frac{bc^2}{2}
+\frac{c^5}{5}-\frac{9c^4}{16}+\frac{c^3}{2}-\frac{c^2}{8}-\frac{c^2}{4}(c-1-2b)^2\log c. \tag*{(11)} ARV ( C b , c II ) = 240 19 + 24 b 5 − 24 b 4 − 12 b 3 − 12 b 2 − 24 b + 3 2 b 2 c 3 − 2 b 2 c 2 − 4 3 b c 4 + 3 4 b c 3 − 2 b c 2 + 5 c 5 − 16 9 c 4 + 2 c 3 − 8 c 2 − 4 c 2 ( c − 1 − 2 b ) 2 log c . (11)
ρ ( C b , c I I ) = b 3 2 − 3 b 2 2 − 3 b 2 + 3 b c 2 − 2 c 3 + 3 c 2 − 1 2 + 3 c ( c − 1 − 2 b ) log c . (12) \rho(C_{b,c}^{\mathrm{II}})=\frac{b^3}{2}-\frac{3b^2}{2}-\frac{3b}{2}+3bc^2-2c^3+3c^2-\frac{1}{2}+3c(c-1-2b)\log c. \tag*{(12)} ρ ( C b , c II ) = 2 b 3 − 2 3 b 2 − 2 3 b + 3 b c 2 − 2 c 3 + 3 c 2 − 2 1 + 3 c ( c − 1 − 2 b ) log c . (12)
Family III
The measure formula (8) remains valid at b = 0 b=0 b = 0 : the diagonal integral disappears, while the marginal calculation is unchanged. Its integrals against continuous functions depend continuously on the parameters, so this extension also agrees with the weak limit of Family II. For c ∈ [ 1 2 , 1 ] c\in[\frac{1}{2},1] c ∈ [ 2 1 , 1 ] , we define
C c I I I : = C 0 , c I I . (13) C_c^{\mathrm{III}}:=C_{0,c}^{\mathrm{II}}. \tag*{(13)} C c III := C 0 , c II . (13)
In this case
κ = c ( 1 − c ) , b ′ = 1 − c , \kappa= c(1-c), \qquad b' = 1-c, κ = c ( 1 − c ) , b ′ = 1 − c ,
and the central diagonal segment disappears. For 1 / 2 < c < 1 1/2 < c < 1 1/2 < c < 1 , its support consists of two anti-diagonal segments and eight curvilinear arcs; at the endpoints, the degenerate pieces are omitted.
For c = 1 c = 1 c = 1 , the curvilinear pieces disappear and
C 1 I I I = C 2 = C 1 / 2 I . C^{\mathrm{III}}_1 = C_2 = C^{\mathrm{I}}_{1/2}. C 1 III = C 2 = C 1/2 I .
(here κ = 0 \kappa= 0 κ = 0 ; the measure (8) is still well defined, since the curved integrals are over [ 1 , 1 ] [1,1] [ 1 , 1 ] , and D 0 , 1 = W D_{0,1} = W D 0 , 1 = W ).
Setting b = 0 b = 0 b = 0 in (11) –(12) gives
ARV ( C c I I I ) = 19 240 + c 5 5 − 9 c 4 16 + c 3 2 − c 2 8 − c 2 ( 1 − c ) 2 4 log c , (14) \operatorname{ARV}\left(C^{\mathrm{III}}_c\right) = \frac{19}{240} + \frac{c^5}{5} - \frac{9c^4}{16} + \frac{c^3}{2} - \frac{c^2}{8} - \frac{c^2(1-c)^2}{4}\log c, \tag*{(14)} ARV ( C c III ) = 240 19 + 5 c 5 − 16 9 c 4 + 2 c 3 − 8 c 2 − 4 c 2 ( 1 − c ) 2 log c , (14)
ρ ( C c I I I ) = − 2 c 3 + 3 c 2 − 1 2 − 3 c ( 1 − c ) log c . (15) \rho\left(C^{\mathrm{III}}_c\right) = -2c^3 + 3c^2 - \frac{1}{2} - 3c(1-c)\log c. \tag*{(15)} ρ ( C c III ) = − 2 c 3 + 3 c 2 − 2 1 − 3 c ( 1 − c ) log c . (15)
Family III is thus obtained from Family II by removing the comonotone block: on the folded square, small values of S S S are no longer coupled with small values of T T T , but are sent along the increasing branch ξ 1 \xi_1 ξ 1 to the top strip [ c , 1 ] [c,1] [ c , 1 ] . In the original square this means that no mass is placed close to the center along the main diagonal, which explains why these copulas attain the smallest values of ARV \operatorname{ARV} ARV . As c c c decreases from 1 1 1 to 1 / 2 1/2 1/2 , the anti-diagonal piece [ 1 − c , c ] [1-c,c] [ 1 − c , c ] shrinks and the curved branches grow; for c = 1 c = 1 c = 1 we recover C 2 C_2 C 2 , whereas for c = 1 / 2 c = 1/2 c = 1/2 we have κ = 1 / 4 \kappa= 1/4 κ = 1/4 , the anti-diagonal piece reduces to a point, and the support of D 0 , 1 / 2 D_{0,1/2} D 0 , 1/2 consists only of the two branches and their transposes. Only the range c ∗ ≤ c ≤ 1 c^* \le c \le1 c ∗ ≤ c ≤ 1 will be relevant for the boundary.
At the other boundary of Family II, namely c = 1 c = 1 c = 1 , we have
C b , 1 I I = C ( 1 − b ) / 2 I . (16) C^{\mathrm{II}}_{b,1} = C^{\mathrm{I}}_{(1-b)/2}. \tag*{(16)} C b , 1 II = C ( 1 − b ) /2 I . (16)
At the folded level, these identities read
( C c I I I ) F = D 0 , c , D b , 1 = D b . (17) \left(C^{\mathrm{III}}_c\right)^{\mathrm{F}} = D_{0,c}, \qquad D_{b,1} = D_b. \tag*{(17)} ( C c III ) F = D 0 , c , D b , 1 = D b . (17)
Thus Family II provides a bridge between Families III and I. Figure Figure 3 shows representative supports of the three families.
Figure 3. Representative supports of the three families. Left: Family I with a = 0.30 a = 0.30 a = 0.30 . Center: Family II with b = 0.10 b = 0.10 b = 0.10 and c ≈ 0.890363 c \approx0.890363 c ≈ 0.890363 . Right: Family III with c = 0.85 c = 0.85 c = 0.85 .
The exact boundary and the main result
We now identify the portions of the three families introduced in § 3 that form the upper boundary of the ( ARV , ρ ) (\operatorname{ARV},\rho) ( ARV , ρ ) region.
The three boundary arcs
For Family I, define
Γ I = { ( ARV ( C a I ) , ρ ( C a I ) ) : 0 ≤ a ≤ 2 5 } . (18) \Gamma^{\mathrm{I}} = \left\{ \left(\operatorname{ARV}(C_a^{\mathrm{I}}), \rho(C_a^{\mathrm{I}})\right) : 0 \leq a \leq\frac{2}{5} \right\}. \tag*{(18)} Γ I = { ( ARV ( C a I ) , ρ ( C a I ) ) : 0 ≤ a ≤ 5 2 } . (18)
For Family II, introduce the polynomial
P ( b , c ) = 6 c 4 − 20 c 3 + 18 c 2 − 6 c + 1 − 8 b c 3 + 36 b c 2 − 24 b c + 4 b − 6 b 2 c 2 − 18 b 2 c + 6 b 2 + 12 b 3 c + 4 b 3 − 5 b 4 . (19) \begin{aligned}
P(b,c) ={}& 6c^4 - 20c^3 + 18c^2 - 6c + 1 - 8bc^3 + 36bc^2 - 24bc + 4b \\
&- 6b^2c^2 - 18b^2c + 6b^2 + 12b^3c + 4b^3 - 5b^4. \tag*{(19)}
\end{aligned} P ( b , c ) = 6 c 4 − 20 c 3 + 18 c 2 − 6 c + 1 − 8 b c 3 + 36 b c 2 − 24 b c + 4 b − 6 b 2 c 2 − 18 b 2 c + 6 b 2 + 12 b 3 c + 4 b 3 − 5 b 4 . (19)
Let c ∗ c^* c ∗ be the unique root in ( 1 / 2 , 1 ) (1/2,1) ( 1/2 , 1 ) of
P ( 0 , c ) = 6 c 4 − 20 c 3 + 18 c 2 − 6 c + 1 = 0 , (20) P(0,c) = 6c^4 - 20c^3 + 18c^2 - 6c + 1 = 0, \tag*{(20)} P ( 0 , c ) = 6 c 4 − 20 c 3 + 18 c 2 − 6 c + 1 = 0 , (20)
so that
c ∗ ≈ 0.7803501. c^* \approx0.7803501. c ∗ ≈ 0.7803501.
For b ∈ ( 0 , 1 / 5 ] b \in(0,1/5] b ∈ ( 0 , 1/5 ] , let c ( b ) c(b) c ( b ) denote the unique solution of
P ( b , c ) = 0 , 1 2 + b < c ≤ 1. (21) P(b,c) = 0, \qquad\frac{1}{2} + b < c \leq1. \tag*{(21)} P ( b , c ) = 0 , 2 1 + b < c ≤ 1. (21)
We set
c ( 0 ) : = c ∗ . c(0) := c^*. c ( 0 ) := c ∗ .
As will be proved later, c ( b ) c(b) c ( b ) is well defined on [ 0 , 1 / 5 ] [0,1/5] [ 0 , 1/5 ] , satisfies
lim b ↓ 0 c ( b ) = c ∗ , c ( 1 5 ) = 1 , \lim_{b \downarrow0} c(b) = c^*, \qquad c\left(\frac{1}{5}\right) = 1, b ↓ 0 lim c ( b ) = c ∗ , c ( 5 1 ) = 1 ,
and selects the relevant one-dimensional subfamily of Family II.
We then define
Γ I I = { ( ARV ( C b , c ( b ) I I ) , ρ ( C b , c ( b ) I I ) ) : 0 ≤ b ≤ 1 5 } , (22) \Gamma^{\mathrm{II}}=\left\{\left(\operatorname{ARV}\left(C^{\mathrm{II}}_{b,c(b)}\right),\rho\left(C^{\mathrm{II}}_{b,c(b)}\right)\right):0\le b\le\frac{1}{5}\right\}, \tag*{(22)} Γ II = { ( ARV ( C b , c ( b ) II ) , ρ ( C b , c ( b ) II ) ) : 0 ≤ b ≤ 5 1 } , (22)
where at b = 0 b=0 b = 0 we use the extension C 0 , c I I = C c I I I C^{\mathrm{II}}_{0,c}=C^{\mathrm{III}}_{c} C 0 , c II = C c III introduced in § 3.3 .
Finally, for Family III, define
Γ I I I = { ( ARV ( C c I I I ) , ρ ( C c I I I ) ) : c ∗ ≤ c ≤ 1 } . (23) \Gamma^{\mathrm{III}}=\left\{\left(\operatorname{ARV}\left(C^{\mathrm{III}}_{c}\right),\rho\left(C^{\mathrm{III}}_{c}\right)\right):c^{*}\le c\le1\right\}. \tag*{(23)} Γ III = { ( ARV ( C c III ) , ρ ( C c III ) ) : c ∗ ≤ c ≤ 1 } . (23)
The three arcs join continuously. The arcs Γ I I I \Gamma^{\mathrm{III}} Γ III and Γ I I \Gamma^{\mathrm{II}} Γ II meet at
( ARV ( C c ∗ I I I ) , ρ ( C c ∗ I I I ) ) ≈ ( 0.0917551 , 0.5039872 ) . \left(\operatorname{ARV}\left(C^{\mathrm{III}}_{c^{*}}\right),\rho\left(C^{\mathrm{III}}_{c^{*}}\right)\right)\approx(0.0917551,0.5039872). ( ARV ( C c ∗ III ) , ρ ( C c ∗ III ) ) ≈ ( 0.0917551 , 0.5039872 ) .
The arcs Γ I I \Gamma^{\mathrm{II}} Γ II and Γ I \Gamma^{\mathrm{I}} Γ I meet at
( 962 9375 , 93 125 ) , \left(\frac{962}{9375},\frac{93}{125}\right), ( 9375 962 , 125 93 ) ,
corresponding to
b = 1 5 , c = 1 , a = 2 5 . b=\frac{1}{5},\qquad c=1,\qquad a=\frac{2}{5}. b = 5 1 , c = 1 , a = 5 2 .
Thus
Γ I I I , Γ I I , Γ I \Gamma^{\mathrm{III}},\qquad\Gamma^{\mathrm{II}},\qquad\Gamma^{\mathrm{I}} Γ III , Γ II , Γ I
join to form a continuous curve connecting
( 11 120 , 1 2 ) and ( 2 15 , 1 ) . \left(\frac{11}{120},\frac{1}{2}\right)\qquad\text{and}\qquad\left(\frac{2}{15},1\right). ( 120 11 , 2 1 ) and ( 15 2 , 1 ) .
Figure 4 shows the three arcs in the ( ARV , ρ ) (\operatorname{ARV},\rho) ( ARV , ρ ) -plane.
Figure 4. The three boundary arcs Γ I I I \Gamma^{\mathrm{III}} Γ III , Γ I I \Gamma^{\mathrm{II}} Γ II and Γ I \Gamma^{\mathrm{I}} Γ I in the ( ARV , ρ ) (\operatorname{ARV},\rho) ( ARV , ρ ) -plane. The inset resolves the left endpoint and the junction c = c ∗ c=c^* c = c ∗ ; its horizontal coordinate is 10 5 ( ARV − 11 / 120 ) 10^5(\operatorname{ARV}-11/120) 1 0 5 ( ARV − 11/120 ) . Distinct line styles identify the arcs in monochrome.
Main result
We can now state the main result.
Theorem 4.1. Let
A = { ( ARV ( C ) , ρ ( C ) ) : C ∈ C } \mathcal{A}=\left\{\left(\operatorname{ARV}(C),\rho(C)\right):C\in\mathcal{C}\right\} A = { ( ARV ( C ) , ρ ( C ) ) : C ∈ C } be the ( ARV , ρ ) (\operatorname{ARV},\rho) ( ARV , ρ ) region of bivariate copulas. Then
A = { ( α , r ) ∈ R 2 : 11 120 ≤ α ≤ 2 15 , ∣ r ∣ ≤ r ∗ ( α ) } , \mathcal{A}=\left\{(\alpha,r)\in\mathbb{R}^{2}:\frac{11}{120}\le\alpha\le\frac{2}{15},\ |r|\le r^{*}(\alpha)\right\}, A = { ( α , r ) ∈ R 2 : 120 11 ≤ α ≤ 15 2 , ∣ r ∣ ≤ r ∗ ( α ) } , where r ∗ r^{*} r ∗ is concave and strictly increasing and
graph ( r ∗ ) = Γ I I I ∪ Γ I I ∪ Γ I . \operatorname{graph}(r^{*})=\Gamma^{\mathrm{III}}\cup\Gamma^{\mathrm{II}}\cup\Gamma^{\mathrm{I}}. graph ( r ∗ ) = Γ III ∪ Γ II ∪ Γ I . The arcs are generated by the following copulas (in order from left to right):
The lower boundary is obtained by flipping one coordinate, which preserves ARV and changes the sign of ρ \rho ρ .
The region is represented in Figure 5 , together with the supports of some of the copulas attaining its boundary; an enlargement near the left endpoint C 2 C_2 C 2 is shown in Figure 6 . Theorem 4.1 can be stated as a sharp inequality between the two functionals.
Figure 5. The ( A R V , ρ ) (\mathrm{ARV},\rho) ( ARV , ρ ) region A \mathcal{A} A (shaded). The upper and lower boundaries are formed by the arcs Γ I I I \Gamma^{\mathrm{III}} Γ III , Γ I I \Gamma^{\mathrm{II}} Γ II and Γ I \Gamma^{\mathrm{I}} Γ I and by their flipped versions; the vertical segment joins C 2 C_2 C 2 and C 3 C_3 C 3 . Open circles mark the junctions of the arcs. The top row shows the supports of the copulas C 2 C_2 C 2 , C 0.9 I I I C^{\mathrm{III}}_{0.9} C 0.9 III , C c ∗ I I I C^{\mathrm{III}}_{c^*} C c ∗ III , C 0.1 , c ( 0.1 ) I I C^{\mathrm{II}}_{0.1,c(0.1)} C 0.1 , c ( 0.1 ) II , C 2 / 5 I C^{\mathrm{I}}_{2/5} C 2/5 I , C 1 / 4 I C^{\mathrm{I}}_{1/4} C 1/4 I and M M M , which attain the upper boundary; the bottom row shows the reflected supports, ( u , v ) ↦ ( u , 1 − v ) (u,v) \mapsto(u,1-v) ( u , v ) ↦ ( u , 1 − v ) , of the copulas attaining the lower boundary, from C 3 C_3 C 3 to W W W . Since the points of Γ I I I \Gamma^{\mathrm{III}} Γ III are within 10 − 4 10^{-4} 1 0 − 4 of C 2 C_2 C 2 in the ARV coordinate, they are not resolved at this scale (see Figure 6 ). The dashed lines show the earlier bound ∣ ρ ∣ ≤ 3 A R V + 3 5 |\rho| \le3\,\mathrm{ARV} + \frac{3}{5} ∣ ρ ∣ ≤ 3 ARV + 5 3 of [3 ] .
Figure 6. Enlargement of the upper boundary of A \mathcal{A} A near its left endpoint C 2 = ( 11 120 , 1 2 ) C_2 = \left(\frac{11}{120}, \frac{1}{2}\right) C 2 = ( 120 11 , 2 1 ) ; the horizontal coordinate is ARV − 11 120 \operatorname{ARV} - \frac{11}{120} ARV − 120 11 . The arc Γ I I I \Gamma^{\mathrm{III}} Γ III leaves C 2 C_2 C 2 with a vertical tangent (Corollary 4.4) and meets Γ I I \Gamma^{\mathrm{II}} Γ II at c = c ∗ c=c^* c = c ∗ .
Corollary 4.2. For every copula C C C ,
∣ ρ ( C ) ∣ ≤ r ∗ ( ARV ( C ) ) , \lvert\rho(C)\rvert\leq r^*(\operatorname{ARV}(C)), ∣ ρ ( C )∣ ≤ r ∗ ( ARV ( C )) ,
and the inequality is sharp for every value of ARV ( C ) \operatorname{ARV}(C) ARV ( C ) . Equivalently, ARV ( C ) ≥ ( r ∗ ) − 1 ( ∣ ρ ( C ) ∣ ) \operatorname{ARV}(C) \geq(r^*)^{-1}(\lvert\rho(C)\rvert) ARV ( C ) ≥ ( r ∗ ) − 1 (∣ ρ ( C )∣) whenever ∣ ρ ( C ) ∣ ≥ 1 2 \lvert\rho(C)\rvert\geq\frac{1}{2} ∣ ρ ( C )∣ ≥ 2 1 , while every value ARV ∈ [ 11 120 , 2 15 ] \operatorname{ARV} \in\left[\frac{11}{120},\frac{2}{15}\right] ARV ∈ [ 120 11 , 15 2 ] is compatible with ∣ ρ ∣ ≤ 1 2 \lvert\rho\rvert\leq\frac{1}{2} ∣ ρ ∣ ≤ 2 1 .
The boundary is smooth away from its left endpoint, where it has a vertical tangent. The proofs of the following two results are given in Section A.3.7.
Corollary 4.3. The function r ∗ r^* r ∗ is continuously differentiable on ( 11 / 120 , 2 / 15 ] (11/120,2/15] ( 11/120 , 2/15 ] , with a one-sided derivative at 2 / 15 2/15 2/15 . Its derivative is determined by the parameter κ \kappa κ of the supporting cost:
( r ∗ ) ′ ( ARV ( C a I ) ) = 6 κ 1 − 2 a = 9 ( 1 − a ) ( 3 − 5 a ) , 0 ≤ a ≤ 2 5 , (r^*)'\left(\operatorname{ARV}\left(C_a^{\mathrm{I}}\right)\right)=\frac{6}{\kappa_{1-2a}}=\frac{9}{(1-a)(3-5a)}, \qquad0\leq a\leq\frac{2}{5}, ( r ∗ ) ′ ( ARV ( C a I ) ) = κ 1 − 2 a 6 = ( 1 − a ) ( 3 − 5 a ) 9 , 0 ≤ a ≤ 5 2 ,
( r ∗ ) ′ ( ARV ( C b , c ( b ) I I ) ) = 6 c ( b ) ( 1 + 2 b − c ( b ) ) , 0 ≤ b ≤ 1 5 , (r^*)'\left(\operatorname{ARV}\left(C_{b,c(b)}^{\mathrm{II}}\right)\right)=\frac{6}{c(b)\left(1+2b-c(b)\right)}, \qquad0\leq b\leq\frac{1}{5}, ( r ∗ ) ′ ( ARV ( C b , c ( b ) II ) ) = c ( b ) ( 1 + 2 b − c ( b ) ) 6 , 0 ≤ b ≤ 5 1 ,
( r ∗ ) ′ ( ARV ( C c I I I ) ) = 6 c ( 1 − c ) , c ∗ ≤ c < 1. (24) (r^*)'\left(\operatorname{ARV}\left(C_c^{\mathrm{III}}\right)\right)=\frac{6}{c(1-c)}, \qquad c^*\leq c<1. \tag*{(24)} ( r ∗ ) ′ ( ARV ( C c III ) ) = c ( 1 − c ) 6 , c ∗ ≤ c < 1. (24)
In particular, ( r ∗ ) ′ ( 2 / 15 ) = 3 (r^*)'(2/15)=3 ( r ∗ ) ′ ( 2/15 ) = 3 , the three arcs join without corners, with slopes 15 at Γ I I ∩ Γ I \Gamma^{\mathrm{II}} \cap\Gamma^{\mathrm{I}} Γ II ∩ Γ I and 6 / ( c ∗ ( 1 − c ∗ ) ) ≈ 35.005 6/(c^*(1-c^*)) \approx35.005 6/ ( c ∗ ( 1 − c ∗ )) ≈ 35.005 at Γ I I I ∩ Γ I I \Gamma^{\mathrm{III}} \cap\Gamma^{\mathrm{II}} Γ III ∩ Γ II , and ( r ∗ ) ′ ( α ) → + ∞ (r^*)'(\alpha) \to+\infty ( r ∗ ) ′ ( α ) → + ∞ as α ↓ 11 / 120 \alpha\downarrow11/120 α ↓ 11/120 .
Corollary 4.4. As α ↓ 11 / 120 \alpha\downarrow11/120 α ↓ 11/120 ,
r ∗ ( α ) = 1 2 + 4 ( α − 11 120 ) 3 / 4 + o ( ( α − 11 120 ) 3 / 4 ) . r^*(\alpha) = \frac{1}{2} + 4\left(\alpha-\frac{11}{120}\right)^{3/4} + o\left(\left(\alpha-\frac{11}{120}\right)^{3/4}\right). r ∗ ( α ) = 2 1 + 4 ( α − 120 11 ) 3/4 + o ( ( α − 120 11 ) 3/4 ) .
In particular, the upper boundary has a vertical tangent at ( 11 / 120 , 1 / 2 ) (11/120,1/2) ( 11/120 , 1/2 ) ; by symmetry, the lower boundary has a vertical tangent at ( 11 / 120 , − 1 / 2 ) (11/120,-1/2) ( 11/120 , − 1/2 ) .
The proof of Theorem 4.1 , together with those of Corollaries 4.3 and 4.4, is given in Appendix § 11 . It relies on a reformulation of the boundary problem as an optimal transport problem on the folded square and on the construction of explicit Kantorovich potentials for the three extremal families.
Average rectangular volume and arachnitude
In this section we show that, on copulas, ARV is an affine function of the arachnitude, a second-order rank correlation used in actuarial practice, and we restate Theorem 4.1 as the exact joint region of Spearman’s rho and arachnitude. Spearman’s rho measures the direction of dependence; arachnitude measures its magnitude , i.e. whether the two ranks tend to be simultaneously extreme, in either tail, or simultaneously central. It was introduced to calibrate the dependence between risk drivers in internal capital models [27 , 1 ] , and is a prominent example of the non-monotonic rank correlations studied in [20 ] .
We first recall the Legendre rank correlations. Let C \mathcal{C} C be the class of bivariate copulas and ( U , V ) ∼ C (U,V) \sim\mathcal{C} ( U , V ) ∼ C . For non-constant g , h ∈ L 2 ( 0 , 1 ) g,h \in L^2(0,1) g , h ∈ L 2 ( 0 , 1 ) , the generalized Spearman correlation of [20 ] is Corr ( g ( U ) , h ( V ) ) \operatorname{Corr}(g(U),h(V)) Corr ( g ( U ) , h ( V )) ; if g g g and h h h are centered and normalized in L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) , it equals E [ g ( U ) h ( V ) ] \mathbb{E}[g(U)h(V)] E [ g ( U ) h ( V )] . Let P j P_j P j be the Legendre polynomial of degree j j j on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] and
Λ j ( u ) = 2 j + 1 P j ( 2 u − 1 ) , j ≥ 0 , \Lambda_j(u) = \sqrt{2j+1}P_j(2u-1), \qquad j \ge0, Λ j ( u ) = 2 j + 1 P j ( 2 u − 1 ) , j ≥ 0 ,
the orthonormal shifted Legendre basis of L 2 ( 0 , 1 ) L^2(0,1) L 2 ( 0 , 1 ) . The Legendre rank correlations of C \mathcal{C} C are ρ j k Λ ( C ) = E [ Λ j ( U ) Λ k ( V ) ] \rho^\Lambda_{jk}(\mathcal{C}) = \mathbb{E}[\Lambda_j(U)\Lambda_k(V)] ρ j k Λ ( C ) = E [ Λ j ( U ) Λ k ( V )] , j , k ≥ 1 j,k \ge1 j , k ≥ 1 ; if C \mathcal{C} C has a square-integrable density c c c , then c ( u , v ) = 1 + ∑ j , k ≥ 1 ρ j k Λ ( C ) Λ j ( u ) Λ k ( v ) c(u,v) = 1 + \sum_{j,k\ge1}\rho^\Lambda_{jk}(\mathcal{C})\Lambda_j(u)\Lambda_k(v) c ( u , v ) = 1 + ∑ j , k ≥ 1 ρ j k Λ ( C ) Λ j ( u ) Λ k ( v ) in L 2 L^2 L 2 . Since
Λ 1 ( u ) = 3 ( 2 u − 1 ) , Λ 2 ( u ) = 5 ( 6 u 2 − 6 u + 1 ) = 5 2 ( 3 ( 2 u − 1 ) 2 − 1 ) , (25) \Lambda_1(u) = \sqrt{3}(2u-1), \qquad\Lambda_2(u) = \sqrt{5}(6u^2-6u+1) = \frac{\sqrt{5}}{2}\left(3(2u-1)^2-1\right), \tag*{(25)} Λ 1 ( u ) = 3 ( 2 u − 1 ) , Λ 2 ( u ) = 5 ( 6 u 2 − 6 u + 1 ) = 2 5 ( 3 ( 2 u − 1 ) 2 − 1 ) , (25)
we have
ρ 11 Λ ( C ) = 3 E [ ( 2 U − 1 ) ( 2 V − 1 ) ] = 12 E [ U V ] − 3 = ρ ( C ) , (26) \rho^\Lambda_{11}(\mathcal{C}) = 3\mathbb{E}[(2U-1)(2V-1)] = 12\mathbb{E}[UV]-3 = \rho(\mathcal{C}), \tag*{(26)} ρ 11 Λ ( C ) = 3 E [( 2 U − 1 ) ( 2 V − 1 )] = 12 E [ U V ] − 3 = ρ ( C ) , (26)
Spearman’s rho, and, since Λ 2 \Lambda_2 Λ 2 is the standardized version of ( 2 u − 1 ) 2 (2u-1)^2 ( 2 u − 1 ) 2 ,
ρ 22 Λ ( C ) = Corr ( ( 2 U − 1 ) 2 , ( 2 V − 1 ) 2 ) , (27) \rho^\Lambda_{22}(\mathcal{C}) = \operatorname{Corr}((2U-1)^2,(2V-1)^2), \tag*{(27)} ρ 22 Λ ( C ) = Corr (( 2 U − 1 ) 2 , ( 2 V − 1 ) 2 ) , (27)
the arachnitude [27 , 20 ] . A flip u ↦ 1 − u u \mapsto1-u u ↦ 1 − u changes the sign of Λ 1 \Lambda_1 Λ 1 and leaves Λ 2 \Lambda_2 Λ 2 unchanged; hence flipping one coordinate changes the sign of ρ 11 Λ \rho^\Lambda_{11} ρ 11 Λ and preserves ρ 22 Λ \rho^\Lambda_{22} ρ 22 Λ . In particular ρ 22 Λ ( M ) = ρ 22 Λ ( W ) = 1 \rho^\Lambda_{22}(M) = \rho^\Lambda_{22}(W) = 1 ρ 22 Λ ( M ) = ρ 22 Λ ( W ) = 1 .
The second of these coefficients is an affine function of the average rectangular volume.
Theorem 5.1. For every bivariate copula C \mathcal{C} C ,
ρ 22 Λ ( C ) = 45 ARV ( C ) − 5 , equivalently ARV ( C ) = 1 9 + 1 45 ρ 22 Λ ( C ) . (28) \rho^\Lambda_{22}(\mathcal{C}) = 45\operatorname{ARV}(\mathcal{C})-5,\qquad\text{equivalently}\qquad\operatorname{ARV}(\mathcal{C}) = \frac{1}{9}+\frac{1}{45}\rho^\Lambda_{22}(\mathcal{C}). \tag*{(28)} ρ 22 Λ ( C ) = 45 ARV ( C ) − 5 , equivalently ARV ( C ) = 9 1 + 45 1 ρ 22 Λ ( C ) . (28)
Proof. By (25) , u ( 1 − u ) = 1 6 − 1 6 5 Λ 2 ( u ) u(1-u) = \frac{1}{6}-\frac{1}{6\sqrt{5}}\Lambda_2(u) u ( 1 − u ) = 6 1 − 6 5 1 Λ 2 ( u ) . Since E [ Λ 2 ( U ) ] = E [ Λ 2 ( V ) ] = 0 \mathbb{E}[\Lambda_2(U)] = \mathbb{E}[\Lambda_2(V)] = 0 E [ Λ 2 ( U )] = E [ Λ 2 ( V )] = 0 , (2) gives
ARV ( C ) = 4 E [ ( 1 6 − Λ 2 ( U ) 6 5 ) ( 1 6 − Λ 2 ( V ) 6 5 ) ] = 1 9 + 1 45 E [ Λ 2 ( U ) Λ 2 ( V ) ] . \operatorname{ARV}(\mathcal{C}) = 4\mathbb{E}\left[\left(\frac{1}{6}-\frac{\Lambda_2(U)}{6\sqrt{5}}\right)\left(\frac{1}{6}-\frac{\Lambda_2(V)}{6\sqrt{5}}\right)\right] = \frac{1}{9}+\frac{1}{45}\mathbb{E}[\Lambda_2(U)\Lambda_2(V)]. ARV ( C ) = 4 E [ ( 6 1 − 6 5 Λ 2 ( U ) ) ( 6 1 − 6 5 Λ 2 ( V ) ) ] = 9 1 + 45 1 E [ Λ 2 ( U ) Λ 2 ( V )] .
Under independence, ARV ( Π ) = 1 9 \operatorname{ARV}(\Pi) = \frac{1}{9} ARV ( Π ) = 9 1 and ρ 22 Λ ( Π ) = 0 \rho^\Lambda_{22}(\Pi) = 0 ρ 22 Λ ( Π ) = 0 . Positive arachnitude means that random rectangles capture more copula mass than under independence, i.e. that U ( 1 − U ) U(1-U) U ( 1 − U ) and V ( 1 − V ) V(1-V) V ( 1 − V ) are positively correlated: the two ranks tend to be simultaneously central or simultaneously extreme. Negative arachnitude means that one rank tends to be extreme when the other is central.
Spearman’s rho is also an expected volume of random rectangles, but of rectangles attached to the corners of the square. For x ∈ [ 0 , 1 ] 2 \mathbf{x} \in[0,1]^2 x ∈ [ 0 , 1 ] 2 and a corner e ∈ { 0 , 1 } 2 \mathbf{e} \in\{0,1\}^2 e ∈ { 0 , 1 } 2 , let R e ( x ) R_{\mathbf{e}}(\mathbf{x}) R e ( x ) be the rectangle with opposite vertices x \mathbf{x} x and e \mathbf{e} e .
Proposition 5.2. Let X \mathbf{X} X be uniform on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 . For every copula C \mathcal{C} C ,
ρ 11 Λ ( C ) = 3 E [ V C ( R ( 0 , 0 ) ( X ) ) + V C ( R ( 1 , 1 ) ( X ) ) − V C ( R ( 1 , 0 ) ( X ) ) − V C ( R ( 0 , 1 ) ( X ) ) ] . \rho^\Lambda_{11}(\mathcal{C}) = 3\mathbb{E}\left[V_{\mathcal{C}}(R_{(0,0)}(\mathbf{X}))+V_{\mathcal{C}}(R_{(1,1)}(\mathbf{X}))-V_{\mathcal{C}}(R_{(1,0)}(\mathbf{X}))-V_{\mathcal{C}}(R_{(0,1)}(\mathbf{X}))\right]. ρ 11 Λ ( C ) = 3 E [ V C ( R ( 0 , 0 ) ( X )) + V C ( R ( 1 , 1 ) ( X )) − V C ( R ( 1 , 0 ) ( X )) − V C ( R ( 0 , 1 ) ( X )) ] .
Proof. The four rectangles cover ( u , v ) (u,v) ( u , v ) with probabilities ( 1 − u ) ( 1 − v ) (1-u)(1-v) ( 1 − u ) ( 1 − v ) , u v uv uv , u ( 1 − v ) u(1-v) u ( 1 − v ) and ( 1 − u ) v (1-u)v ( 1 − u ) v , whose signed sum is ( 1 − 2 u ) ( 1 − 2 v ) (1-2u)(1-2v) ( 1 − 2 u ) ( 1 − 2 v ) . As in (2) , the right-hand side equals 3 E [ ( 1 − 2 U ) ( 1 − 2 V ) ] = ρ 11 ( C ) 3\mathbb{E}[(1-2U)(1-2V)]=\rho_{11}(C) 3 E [( 1 − 2 U ) ( 1 − 2 V )] = ρ 11 ( C ) .
A flip exchanges the corners ( 0 , 0 ) ↔ ( 1 , 0 ) (0,0)\leftrightarrow(1,0) ( 0 , 0 ) ↔ ( 1 , 0 ) and ( 0 , 1 ) ↔ ( 1 , 1 ) (0,1)\leftrightarrow(1,1) ( 0 , 1 ) ↔ ( 1 , 1 ) , so the anchored representation changes sign, whereas the law of the spanned rectangle R X , Y R_{X,Y} R X , Y is invariant. Spearman’s rho measures the direction of dependence because its rectangles are anchored; arachnitude measures its magnitude because its rectangles are attached to nothing.
As a consequence, every result on ARV can be translated into a result on arachnitude.
By Theorem 5.1 , the map
Φ ( α , r ) = ( 45 α − 5 , r ) (29) \Phi(\alpha,r)=(45\alpha-5,r) \tag*{(29)} Φ ( α , r ) = ( 45 α − 5 , r ) (29)
is an increasing affine bijection of the plane that carries the (ARV, ρ \rho ρ ) region onto the joint region of arachnitude and Spearman’s rho:
A Λ : = { ( ρ 22 Λ ( C ) , ρ 11 Λ ( C ) ) : C ∈ C } = Φ ( A ) . (30) \mathcal{A}^{\Lambda}:=\left\{(\rho_{22}^{\Lambda}(C),\rho_{11}^{\Lambda}(C)):C\in\mathcal{C}\right\}=\Phi(\mathcal{A}). \tag*{(30)} A Λ := { ( ρ 22 Λ ( C ) , ρ 11 Λ ( C )) : C ∈ C } = Φ ( A ) . (30)
Boundaries are mapped onto boundaries, slopes are divided by 45, and the first coordinate is translated according to the following dictionary.
Table 1. Correspondence between ARV and arachnitude, ρ 22 Λ = 45 ARV − 5 \rho_{22}^{\Lambda}=45\operatorname{ARV}-5 ρ 22 Λ = 45 ARV − 5 .
In particular, the range 11 120 ≤ ARV ≤ 2 15 \frac{11}{120}\leq\operatorname{ARV}\leq\frac{2}{15} 120 11 ≤ ARV ≤ 15 2 of [3 ] gives the sharp range of arachnitude,
− 7 8 ≤ ρ 22 Λ ( C ) ≤ 1 , -\frac{7}{8}\leq\rho_{22}^{\Lambda}(C)\leq1, − 8 7 ≤ ρ 22 Λ ( C ) ≤ 1 ,
a value also obtained in [20 ] . By the discussion preceding Proposition 2.4 , the lower bound is attained exactly by the copulas whose folded copula is W W W and the upper bound exactly by those whose folded copula is M M M .
In particular, this applies to the main result of this paper.
By Theorem 5.1 , Theorem 4.1 and corollaries 4.3 and 4.4 translate into statements on the joint region A Λ \mathcal{A}^{\Lambda} A Λ of (30) . For η ∈ [ − 7 8 , 1 ] \eta\in[-\frac{7}{8},1] η ∈ [ − 8 7 , 1 ] put
R ∗ ( η ) = r ∗ ( η + 5 45 ) = max { ρ 11 Λ ( C ) : C ∈ C , ρ 22 Λ ( C ) = η } . (31) R^{*}(\eta)=r^{*}\left(\frac{\eta+5}{45}\right)=\max\{\rho_{11}^{\Lambda}(C):C\in\mathcal{C},\ \rho_{22}^{\Lambda}(C)=\eta\}. \tag*{(31)} R ∗ ( η ) = r ∗ ( 45 η + 5 ) = max { ρ 11 Λ ( C ) : C ∈ C , ρ 22 Λ ( C ) = η } . (31)
Theorem 5.3. The joint region of arachnitude and Spearman’s rho is
A Λ = { ( η , r ) ∈ R 2 : − 7 8 ≤ η ≤ 1 , ∣ r ∣ ≤ R ∗ ( η ) } . \mathcal{A}^{\Lambda}=\left\{(\eta,r)\in\mathbb{R}^{2}:-\frac{7}{8}\leq\eta\leq1,\ |r|\leq R^{*}(\eta)\right\}. A Λ = { ( η , r ) ∈ R 2 : − 8 7 ≤ η ≤ 1 , ∣ r ∣ ≤ R ∗ ( η ) } . where R ∗ R^{*} R ∗ is continuous, concave and strictly increasing, with R ∗ ( − 7 8 ) = 1 2 R^{*}\left(-\frac{7}{8}\right)=\frac{1}{2} R ∗ ( − 8 7 ) = 2 1 and R ∗ ( 1 ) = 1 R^{*}(1)=1 R ∗ ( 1 ) = 1 . The graph of R ∗ R^{*} R ∗ is the union of the three arcs Φ ( Γ I I I ) \Phi(\Gamma^{\mathrm{III}}) Φ ( Γ III ) , Φ ( Γ I I ) \Phi(\Gamma^{\mathrm{II}}) Φ ( Γ II ) and Φ ( Γ I ) \Phi(\Gamma^{\mathrm{I}}) Φ ( Γ I ) , generated by the copulas of Theorem 4.1 :
(a) for c ∗ ≤ c ≤ 1 c^{*}\leq c\leq1 c ∗ ≤ c ≤ 1 , the points ( ρ 22 Λ ( C c I I I ) , ρ 11 Λ ( C c I I I ) ) \left(\rho_{22}^{\Lambda}(C_{c}^{\mathrm{III}}),\rho_{11}^{\Lambda}(C_{c}^{\mathrm{III}})\right) ( ρ 22 Λ ( C c III ) , ρ 11 Λ ( C c III ) ) , with
ρ 22 Λ ( C c I I I ) = − 23 16 + 9 c 5 − 405 c 4 16 + 45 c 3 2 − 45 c 2 8 − 45 c 2 ( 1 − c ) 2 4 log c , \rho_{22}^{\Lambda}(C_{c}^{\mathrm{III}})=-\frac{23}{16}+9c^{5}-\frac{405c^{4}}{16}+\frac{45c^{3}}{2}-\frac{45c^{2}}{8}-\frac{45c^{2}(1-c)^{2}}{4}\log c, ρ 22 Λ ( C c III ) = − 16 23 + 9 c 5 − 16 405 c 4 + 2 45 c 3 − 8 45 c 2 − 4 45 c 2 ( 1 − c ) 2 log c , and ρ 11 Λ ( C c I I I ) = ρ ( C c I I I ) \rho_{11}^{\Lambda}(C_{c}^{\mathrm{III}})=\rho(C_{c}^{\mathrm{III}}) ρ 11 Λ ( C c III ) = ρ ( C c III ) given by (15) ;
(b) for 0 ≤ b ≤ 1 5 0\leq b\leq\frac{1}{5} 0 ≤ b ≤ 5 1 , the points ( ρ 22 Λ ( C b , c ( b ) I I ) , ρ 11 Λ ( C b , c ( b ) I I ) ) \left(\rho_{22}^{\Lambda}(C_{b,c(b)}^{\mathrm{II}}),\rho_{11}^{\Lambda}(C_{b,c(b)}^{\mathrm{II}})\right) ( ρ 22 Λ ( C b , c ( b ) II ) , ρ 11 Λ ( C b , c ( b ) II ) ) , with ρ 22 Λ = 45 ARV − 5 \rho_{22}^{\Lambda}=45\operatorname{ARV}-5 ρ 22 Λ = 45 ARV − 5 and ARV , ρ \operatorname{ARV},\rho ARV , ρ given by (11) –(12) ;
(c) for 0 ≤ a ≤ 2 5 0\leq a\leq\frac{2}{5} 0 ≤ a ≤ 5 2 , the points
( ρ 22 Λ ( C a I ) , ρ 11 Λ ( C a I ) ) = ( 1 − 60 ( 1 − a ) 2 a 3 , 1 − 4 a 3 ) . \left(\rho_{22}^{\Lambda}(C_{a}^{\mathrm{I}}),\rho_{11}^{\Lambda}(C_{a}^{\mathrm{I}})\right)=\left(1-60(1-a)^{2}a^{3},1-4a^{3}\right). ( ρ 22 Λ ( C a I ) , ρ 11 Λ ( C a I ) ) = ( 1 − 60 ( 1 − a ) 2 a 3 , 1 − 4 a 3 ) . The arcs meet at ( ρ 22 Λ , ρ 11 Λ ) ≈ ( − 0.871022 , 0.503987 ) \left(\rho_{22}^{\Lambda},\rho_{11}^{\Lambda}\right)\approx(-0.871022,0.503987) ( ρ 22 Λ , ρ 11 Λ ) ≈ ( − 0.871022 , 0.503987 ) and at ( − 239 625 , 93 125 ) \left(-\frac{239}{625},\frac{93}{125}\right) ( − 625 239 , 125 93 ) , and join ( − 7 8 , 1 2 ) \left(-\frac{7}{8},\frac{1}{2}\right) ( − 8 7 , 2 1 ) to ( 1 , 1 ) (1,1) ( 1 , 1 ) . The lower boundary is obtained by flipping one coordinate, which preserves ρ 22 Λ \rho_{22}^{\Lambda} ρ 22 Λ and changes the sign of ρ 11 Λ \rho_{11}^{\Lambda} ρ 11 Λ .
Proof. By (30) , A Λ = Φ ( A ) \mathcal{A}^{\Lambda}=\Phi(\mathcal{A}) A Λ = Φ ( A ) , and since Φ \Phi Φ only rescales and translates the first coordinate, Theorem 4.1 gives the description of A Λ \mathcal{A}^{\Lambda} A Λ with R ∗ = r ∗ ∘ Φ 1 − 1 R^{*}=r^{*}\circ\Phi_{1}^{-1} R ∗ = r ∗ ∘ Φ 1 − 1 , where Φ 1 ( α ) = 45 α − 5 \Phi_{1}(\alpha)=45\alpha-5 Φ 1 ( α ) = 45 α − 5 ; continuity, concavity and strict monotonicity are preserved. The endpoint values follow from Proposition 2.4 . The parametrizations follow from (6) , (14) and (11) by applying Φ 1 \Phi_{1} Φ 1 , and the junction points from those of § 4.1 .
Corollary 5.4. For every copula C C C ,
∣ ρ 11 Λ ( C ) ∣ ≤ R ∗ ( ρ 22 Λ ( C ) ) , |\rho_{11}^{\Lambda}(C)|\leq R^{*}\left(\rho_{22}^{\Lambda}(C)\right), ∣ ρ 11 Λ ( C ) ∣ ≤ R ∗ ( ρ 22 Λ ( C ) ) , and the inequality is sharp for every value of ρ 22 Λ ( C ) \rho_{22}^{\Lambda}(C) ρ 22 Λ ( C ) . Equivalently, ρ 22 Λ ( C ) ≥ ( R ∗ ) − 1 ( ∣ ρ 11 Λ ( C ) ∣ ) \rho_{22}^{\Lambda}(C)\geq(R^{*})^{-1}(|\rho_{11}^{\Lambda}(C)|) ρ 22 Λ ( C ) ≥ ( R ∗ ) − 1 ( ∣ ρ 11 Λ ( C ) ∣ ) whenever ∣ ρ 11 Λ ( C ) ∣ ≥ 1 2 |\rho_{11}^{\Lambda}(C)|\geq\frac{1}{2} ∣ ρ 11 Λ ( C ) ∣ ≥ 2 1 , while every value ρ 22 Λ ∈ [ − 7 8 , 1 ] \rho_{22}^{\Lambda}\in\left[-\frac{7}{8},1\right] ρ 22 Λ ∈ [ − 8 7 , 1 ] is compatible with ∣ ρ 11 Λ ∣ ≤ 1 2 |\rho_{11}^{\Lambda}|\leq\frac{1}{2} ∣ ρ 11 Λ ∣ ≤ 2 1 .
Corollary 5.4 is a consistency condition for first- and second-order rank dependence: a pair of estimated or expert-specified values ( ρ 22 Λ , ρ 11 Λ ) (\rho_{22}^{\Lambda},\rho_{11}^{\Lambda}) ( ρ 22 Λ , ρ 11 Λ ) outside A Λ A^{\Lambda} A Λ cannot be produced by any copula. For instance, Table 2 shows that a strongly concordant pair of risks cannot be strongly anti-cruciform.
Table 2. Smallest arachnitude compatible with a given value of Spearman’s rho, ( R ∗ ) − 1 ( ∣ ρ 11 Λ ∣ ) (R^{*})^{-1}(|\rho_{11}^{\Lambda}|) ( R ∗ ) − 1 ( ∣ ρ 11 Λ ∣ ) , computed from the arcs of Theorem 5.3 .
We conclude the section with the values of the two coefficients for some copulas and classical families, which are represented in Figure 7 . These values were computed by numerical integration of ρ 11 Λ ( C ) \rho_{11}^{\Lambda}(C) ρ 11 Λ ( C ) and of ρ 22 L a m b d a ( C ) \rho_{22}^{\\Lambda}(C) ρ 22 L amb d a ( C ) , with ARV given by (1) , for copulas with a closed form, and of E [ Λ j ( U ) Λ j ( V ) ] \mathbb{E}[\Lambda_j(U)\Lambda_j(V)] E [ Λ j ( U ) Λ j ( V )] for elliptical copulas. We start from the extreme values of arachnitude.
Figure 7. The region A Λ A^{\Lambda} A Λ of (arachnitude, Spearman’s rho) (shaded; see Theorem 5.3 ) with parametric families. (a) Gaussian and Student t t t copulas, r ∈ ( − 1 , 1 ) r \in(-1,1) r ∈ ( − 1 , 1 ) . (b) Clayton, Frank and Gumbel copulas, and Cuadras–Augé copulas. Points: M M M , W W W , independence Π \Pi Π , the cross copula X X X and the minimum-arachnitude copulas C 2 , C 3 C_2,C_3 C 2 , C 3 .
Example 5.5 (Extreme values of arachnitude). M M M , W W W and the cross copula X = ( M + W ) / 2 X=(M+W)/2 X = ( M + W ) /2 have folded copula M M M , hence ρ 22 Λ = 1 \rho_{22}^{\Lambda}=1 ρ 22 Λ = 1 , while ρ 11 Λ = 1 , − 1 , 0 \rho_{11}^{\Lambda}=1,-1,0 ρ 11 Λ = 1 , − 1 , 0 , respectively. At the other extreme, the shuffles of Min C 2 C_2 C 2 and C 3 C_3 C 3 of § 3.1 have folded copula W W W , hence ρ 22 Λ = − 7 8 \rho_{22}^{\Lambda}=-\frac{7}{8} ρ 22 Λ = − 8 7 , while ρ 11 Λ = 1 2 \rho_{11}^{\Lambda}=\frac{1}{2} ρ 11 Λ = 2 1 and ρ 11 Λ = − 1 2 \rho_{11}^{\Lambda}=-\frac{1}{2} ρ 11 Λ = − 2 1 ; for these copulas one rank is extreme exactly when the other is central. The mixtures ( 1 − p ) Π + p X (1-p)\Pi+pX ( 1 − p ) Π + pX and ( 1 − p ) Π + p C 2 + C 3 2 (1-p)\Pi+p\frac{C_2+C_3}{2} ( 1 − p ) Π + p 2 C 2 + C 3 , p ∈ [ 0 , 1 ] p\in[0,1] p ∈ [ 0 , 1 ] , have ρ 11 Λ = 0 \rho_{11}^{\Lambda}=0 ρ 11 Λ = 0 and ρ 22 Λ = p \rho_{22}^{\Lambda}=p ρ 22 Λ = p , respectively ρ 22 Λ = − 7 8 p \rho_{22}^{\Lambda}=-\frac{7}{8}p ρ 22 Λ = − 8 7 p ; they fill the horizontal axis of the region.
Among elliptical copulas, the Gaussian family admits an explicit description.
Proposition 5.6. Let C r C_r C r be the Gaussian copula with correlation r ∈ [ − 1 , 1 ] r\in[-1,1] r ∈ [ − 1 , 1 ] . Then ρ 11 Λ ( C r ) = 6 π arcsin r 2 \rho_{11}^{\Lambda}(C_r)=\frac{6}{\pi}\arcsin\frac{r}{2} ρ 11 Λ ( C r ) = π 6 arcsin 2 r and
ρ 22 Λ ( C r ) = ∑ k ≥ 1 β 2 k r 2 k , β 2 k = ( E [ Λ 2 ( Φ ( Z ) ) H e 2 k ( Z ) ] ) 2 ( 2 k ) ! ≥ 0 , \rho_{22}^{\Lambda}(C_r)=\sum_{k\geq1}\beta_{2k}r^{2k}, \qquad\beta_{2k}=\frac{\left(\mathbb{E}[\Lambda_2(\Phi(Z))H\!e_{2k}(Z)]\right)^2}{(2k)!}\geq0, ρ 22 Λ ( C r ) = k ≥ 1 ∑ β 2 k r 2 k , β 2 k = ( 2 k )! ( E [ Λ 2 ( Φ ( Z )) H e 2 k ( Z )] ) 2 ≥ 0 ,
with ∑ k ≥ 1 β 2 k = 1 \sum_{k\geq1}\beta_{2k}=1 ∑ k ≥ 1 β 2 k = 1 , where Z Z Z is standard normal, Φ \Phi Φ its distribution function and H e m He_m H e m the probabilists’ Hermite polynomials. In particular β 2 = 15 2 π 2 ≈ 0.760 \beta_2=\frac{15}{2\pi^2}\approx0.760 β 2 = 2 π 2 15 ≈ 0.760 , so ρ 22 Λ ( C r ) = 15 2 π 2 r 2 + O ( r 4 ) \rho^\Lambda_{22}(C_r)=\frac{15}{2\pi^2}r^2+O(r^4) ρ 22 Λ ( C r ) = 2 π 2 15 r 2 + O ( r 4 ) , and ρ 22 Λ ( C r ) \rho^\Lambda_{22}(C_r) ρ 22 Λ ( C r ) is even, non-negative and increasing in ∣ r ∣ |r| ∣ r ∣ .
Proof. The formula for ρ 11 Λ \rho^\Lambda_{11} ρ 11 Λ is classical. Write f = Λ 2 ∘ Φ f=\Lambda_2\circ\Phi f = Λ 2 ∘ Φ , so E f ( Z ) = 0 \mathbb{E}f(Z)=0 E f ( Z ) = 0 and E f ( Z ) 2 = 1 \mathbb{E}f(Z)^2=1 E f ( Z ) 2 = 1 . Its Hermite expansion f = ∑ m c m H e m / m ! , c m = E [ f ( Z ) H e m ( Z ) ] f=\sum_m c_m He_m/m!,\ c_m=\mathbb{E}[f(Z)He_m(Z)] f = ∑ m c m H e m / m ! , c m = E [ f ( Z ) H e m ( Z )] , has only even terms because f f f is even. By Mehler’s formula (see, e.g., [19 ] ), for ( Z 1 , Z 2 ) (Z_1,Z_2) ( Z 1 , Z 2 ) standard bivariate normal with correlation r r r , E [ f ( Z 1 ) f ( Z 2 ) ] = ∑ m c m 2 r m / m ! \mathbb{E}[f(Z_1)f(Z_2)]=\sum_m c_m^2r^m/m! E [ f ( Z 1 ) f ( Z 2 )] = ∑ m c m 2 r m / m ! , and ∑ m c m 2 / m ! = E f ( Z ) 2 = 1 \sum_m c_m^2/m!=\mathbb{E}f(Z)^2=1 ∑ m c m 2 / m ! = E f ( Z ) 2 = 1 . For m = 2 m=2 m = 2 , Stein’s identity gives c 2 = E [ f ( Z ) ( Z 2 − 1 ) ] = E [ f ′ ′ ( Z ) ] c_2=\mathbb{E}[f(Z)(Z^2-1)]=\mathbb{E}[f''(Z)] c 2 = E [ f ( Z ) ( Z 2 − 1 )] = E [ f ′′ ( Z )] , where, with ϕ = Φ ′ \phi=\Phi' ϕ = Φ ′ ,
f ′ ′ ( z ) = 5 ( 12 ϕ ( z ) 2 − 12 z Φ ( z ) ϕ ( z ) + 6 z ϕ ( z ) ) f''(z)=\sqrt{5}\left(12\phi(z)^2-12z\Phi(z)\phi(z)+6z\phi(z)\right) f ′′ ( z ) = 5 ( 12 ϕ ( z ) 2 − 12 z Φ ( z ) ϕ ( z ) + 6 z ϕ ( z ) )
Now E [ Z ϕ ( Z ) ] = 0 \mathbb{E}[Z\phi(Z)]=0 E [ Z ϕ ( Z )] = 0 by symmetry, E [ Z Φ ( Z ) ϕ ( Z ) ] = 1 2 E [ ϕ ( Z ) 2 ] \mathbb{E}[Z\Phi(Z)\phi(Z)]=\frac{1}{2}\mathbb{E}[\phi(Z)^2] E [ Z Φ ( Z ) ϕ ( Z )] = 2 1 E [ ϕ ( Z ) 2 ] by Stein’s identity E [ Z k ( Z ) ] = E [ k ′ ( Z ) ] \mathbb{E}[Zk(Z)]=\mathbb{E}[k'(Z)] E [ Z k ( Z )] = E [ k ′ ( Z )] applied to k = Φ ϕ k=\Phi\phi k = Φ ϕ , and E [ ϕ ( Z ) 2 ] = 1 2 π 3 \mathbb{E}[\phi(Z)^2]=\frac{1}{2\pi\sqrt{3}} E [ ϕ ( Z ) 2 ] = 2 π 3 1 . Hence c 2 = 6 5 E [ ϕ ( Z ) 2 ] = 15 / π c_2=6\sqrt{5}\mathbb{E}[\phi(Z)^2]=\sqrt{15}/\pi c 2 = 6 5 E [ ϕ ( Z ) 2 ] = 15 / π and β 2 = c 2 2 / 2 = 15 / ( 2 π 2 ) \beta_2=c_2^2/2=15/(2\pi^2) β 2 = c 2 2 /2 = 15/ ( 2 π 2 ) .
Combining the two leading terms, Gaussian copulas satisfy ρ 22 Λ ≈ 5 6 ( ρ 11 Λ ) 2 \rho^\Lambda_{22}\approx\frac{5}{6}(\rho^\Lambda_{11})^2 ρ 22 Λ ≈ 6 5 ( ρ 11 Λ ) 2 for weak dependence. A sample with small Spearman’s rho and clearly positive or negative arachnitude is therefore not Gaussian. Numerically, ρ 22 Λ ( C 0.5 ) ≈ 0.202 \rho^\Lambda_{22}(C_{0.5})\approx0.202 ρ 22 Λ ( C 0.5 ) ≈ 0.202 and ρ 22 Λ ( C 0.9 ) ≈ 0.763 \rho^\Lambda_{22}(C_{0.9})\approx0.763 ρ 22 Λ ( C 0.9 ) ≈ 0.763 . Student t t t copulas behave differently, since they exhibit joint extremes even without correlation.
Proposition 5.7. Let C r , ν C_{r,\nu} C r , ν be the Student t t t copula with correlation parameter r r r and ν > 0 \nu>0 ν > 0 degrees of freedom. Then ρ 22 Λ ( C 0 , ν ) > 0 \rho^\Lambda_{22}(C_{0,\nu})>0 ρ 22 Λ ( C 0 , ν ) > 0 , although ρ 11 Λ ( C 0 , ν ) = 0 \rho^\Lambda_{11}(C_{0,\nu})=0 ρ 11 Λ ( C 0 , ν ) = 0 .
Proof. Write ( X , Y ) = G ( Z 1 , Z 2 ) (X,Y)=\sqrt{G}(Z_1,Z_2) ( X , Y ) = G ( Z 1 , Z 2 ) , with Z 1 , Z 2 Z_1,Z_2 Z 1 , Z 2 independent standard normal and G = ν / χ ν 2 G=\nu/\chi^2_\nu G = ν / χ ν 2 independent of them. By symmetry of the t t t distribution, ( 2 U − 1 ) 2 = k ( ∣ X ∣ ) (2U-1)^2=k(|X|) ( 2 U − 1 ) 2 = k ( ∣ X ∣ ) and ( 2 V − 1 ) 2 = k ( ∣ Y ∣ ) (2V-1)^2=k(|Y|) ( 2 V − 1 ) 2 = k ( ∣ Y ∣ ) with k k k strictly increasing. Conditionally on G G G , ∣ X ∣ |X| ∣ X ∣ and ∣ Y ∣ |Y| ∣ Y ∣ are independent, so
Cov ( k ( ∣ X ∣ ) , k ( ∣ Y ∣ ) ) = Var ( m ( G ) ) , m ( g ) = E [ k ( g ∣ Z 1 ∣ ) ] \operatorname{Cov}\bigl(k(|X|),k(|Y|)\bigr)=\operatorname{Var}(m(G)),\qquad m(g)=\mathbb{E}\left[k(\sqrt{g}|Z_1|)\right] Cov ( k ( ∣ X ∣ ) , k ( ∣ Y ∣ ) ) = Var ( m ( G )) , m ( g ) = E [ k ( g ∣ Z 1 ∣ ) ]
and m m m is strictly increasing. Since G G G is not degenerate, the variance is positive. Finally ρ 11 Λ = 0 \rho^\Lambda_{11}=0 ρ 11 Λ = 0 by the symmetry ( X , Y ) ↦ ( X , − Y ) (X,Y)\mapsto(X,-Y) ( X , Y ) ↦ ( X , − Y ) .
The mixing variable G G G inflates and deflates both coordinates together, which produces the cross-shaped clustering in all four corners that arachnitude detects. Numerically ρ 22 Λ ( C 0 , 3 ) ≈ 0.241 \rho^\Lambda_{22}(C_{0,3})\approx0.241 ρ 22 Λ ( C 0 , 3 ) ≈ 0.241 and ρ 22 Λ ( C 0 , 10 ) ≈ 0.076 \rho^\Lambda_{22}(C_{0,10})\approx0.076 ρ 22 Λ ( C 0 , 10 ) ≈ 0.076 . For fixed r r r the t t t curves lie to the right of the Gaussian one (Figure 7 a): at the same Spearman’s rho, a t t t copula has larger arachnitude.
We now turn to Archimedean copulas.
For the Clayton, Frank and Gumbel families (Figure 7 b) the curves start at Π \Pi Π and end at M M M (and at W W W for Clayton and Frank). All three stay in, or extremely close to, the half-plane ρ 22 Λ ≥ 0 \rho^\Lambda_{22}\geq0 ρ 22 Λ ≥ 0 ; the smallest value we found is ρ 22 Λ ≈ − 0.009 \rho^\Lambda_{22}\approx-0.009 ρ 22 Λ ≈ − 0.009 for Clayton copulas with negative parameter. At equal Spearman’s rho, Frank copulas, which have no tail dependence, have the smallest arachnitude, while Clayton and Gumbel copulas, with lower and upper tail dependence respectively, have larger values.
A class of copulas for which the two coefficients can be computed in closed form is that of lower semilinear copulas [7 ] . These copulas are linear on each segment joining a point of the main diagonal to the lower or to the left boundary of the unit square; equivalently, they can be written as
C f ( u , v ) = min { u , v } f ( max { u , v } ) , C_f(u,v)=\min\{u,v\}f(\max\{u,v\}), C f ( u , v ) = min { u , v } f ( max { u , v }) ,
where f : ( 0 , 1 ] → R f:(0,1]\to\mathbb{R} f : ( 0 , 1 ] → R is non-decreasing, t ↦ f ( t ) / t t\mapsto f(t)/t t ↦ f ( t ) / t is non-increasing and f ( 1 ) = 1 f(1)=1 f ( 1 ) = 1 (here f ( t ) = δ ( t ) / t f(t)=\delta(t)/t f ( t ) = δ ( t ) / t , δ \delta δ being the diagonal section). Since t ≤ f ( t ) ≤ 1 t\leq f(t)\leq1 t ≤ f ( t ) ≤ 1 , every lower semilinear copula satisfies Π ≤ C f ≤ M \Pi\leq C_f\leq M Π ≤ C f ≤ M .
Proposition 5.8. Let C f C_f C f be a lower semilinear copula and let μ f \mu_f μ f be the non-negative Borel measure on ( 0 , 1 ] (0,1] ( 0 , 1 ] such that f ( t ) / t − 1 = μ f ( ( t , 1 ] ) f(t)/t-1=\mu_f((t,1]) f ( t ) / t − 1 = μ f (( t , 1 ]) for all but countably many t t t . Then
ρ 11 Λ ( C f ) = 12 ∫ 0 1 t 2 f ( t ) d t − 3 = 3 ∫ ( 0 , 1 ] s 4 d μ f ( s ) . \rho^\Lambda_{11}(C_f)=12\int_0^1t^2f(t)\,\mathrm{d}t-3=3\int_{(0,1]}s^4\,\mathrm{d}\mu_f(s). ρ 11 Λ ( C f ) = 12 ∫ 0 1 t 2 f ( t ) d t − 3 = 3 ∫ ( 0 , 1 ] s 4 d μ f ( s ) . ρ 22 Λ ( C f ) = 45 ∫ ( 0 , 1 ] s 4 ( 1 − 4 s 3 ) 2 d μ f ( s ) . \rho^\Lambda_{22}(C_f)=45\int_{(0,1]}s^4\left(1-\frac{4s}{3}\right)^2\,\mathrm{d}\mu_f(s). ρ 22 Λ ( C f ) = 45 ∫ ( 0 , 1 ] s 4 ( 1 − 3 4 s ) 2 d μ f ( s ) . Consequently, 0 ≤ ρ 22 Λ ( C f ) ≤ 15 ρ 11 Λ ( C f ) 0\leq\rho^\Lambda_{22}(C_f)\leq15\rho^\Lambda_{11}(C_f) 0 ≤ ρ 22 Λ ( C f ) ≤ 15 ρ 11 Λ ( C f ) , and ρ 22 Λ ( C f ) = 0 \rho^\Lambda_{22}(C_f)=0 ρ 22 Λ ( C f ) = 0 or ρ 11 Λ ( C f ) = 0 \rho^\Lambda_{11}(C_f)=0 ρ 11 Λ ( C f ) = 0 only if C f = Π C_f=\Pi C f = Π .
Proof. Since C f C_f C f is symmetric, ∫ C f = 2 ∫ 0 1 f ( y ) ∫ 0 y x d x d y = ∫ 0 1 y 2 f ( y ) d y \int C_f=2\int_0^1f(y)\int_0^yx\,\mathrm{d}x\,\mathrm{d}y=\int_0^1y^2f(y)\,\mathrm{d}y ∫ C f = 2 ∫ 0 1 f ( y ) ∫ 0 y x d x d y = ∫ 0 1 y 2 f ( y ) d y , and, by (1) ,
ARV ( C f ) = 8 ∫ 0 1 f ( y ) ( 1 − 2 y ) ∫ 0 y x ( 1 − 2 x ) d x d y = 4 3 ∫ 0 1 y 2 ( 1 − 2 y ) ( 3 − 4 y ) f ( y ) d y . \operatorname{ARV}(C_f)=8\int_0^1f(y)(1-2y)\int_0^yx(1-2x)\,\mathrm{d}x\,\mathrm{d}y=\frac{4}{3}\int_0^1y^2(1-2y)(3-4y)f(y)\,\mathrm{d}y. ARV ( C f ) = 8 ∫ 0 1 f ( y ) ( 1 − 2 y ) ∫ 0 y x ( 1 − 2 x ) d x d y = 3 4 ∫ 0 1 y 2 ( 1 − 2 y ) ( 3 − 4 y ) f ( y ) d y .
Write f ( t ) = t + t g ( t ) f(t)=t+tg(t) f ( t ) = t + t g ( t ) with g ( t ) = μ f ( ( t , 1 ] ) g(t)=\mu_f((t,1]) g ( t ) = μ f (( t , 1 ]) . The term f ( t ) = t f(t)=t f ( t ) = t corresponds to Π \Pi Π , for which both coefficients vanish. For the second term, Fubini’s theorem gives ∫ 0 1 w ( t ) t g ( t ) d t = ∫ ( 0 , 1 ] ∫ 0 s w ( t ) t d t d μ f ( s ) \int_0^1w(t)tg(t)\,\mathrm{d}t=\int_{(0,1]}\int_0^sw(t)t\,\mathrm{d}t\,\mathrm{d}\mu_f(s) ∫ 0 1 w ( t ) t g ( t ) d t = ∫ ( 0 , 1 ] ∫ 0 s w ( t ) t d t d μ f ( s ) , and
∫ 0 s 12 t 3 d t = 3 s 4 , ∫ 0 s 4 3 t 3 ( 1 − 2 t ) ( 3 − 4 t ) d t = s 4 ( 1 − 4 s 3 ) 2 . \int_0^s12t^3\,\mathrm{d}t=3s^4,\qquad\int_0^s\frac{4}{3}t^3(1-2t)(3-4t)\,\mathrm{d}t=s^4\left(1-\frac{4s}{3}\right)^2. ∫ 0 s 12 t 3 d t = 3 s 4 , ∫ 0 s 3 4 t 3 ( 1 − 2 t ) ( 3 − 4 t ) d t = s 4 ( 1 − 3 4 s ) 2 .
With Theorem 5.1 , this gives the two formulas. The inequalities follow from 0 ≤ ( 1 − 4 s 3 ) 2 ≤ 1 0\leq\left(1-\frac{4s}{3}\right)^2\leq1 0 ≤ ( 1 − 3 4 s ) 2 ≤ 1 on ( 0 , 1 ] (0,1] ( 0 , 1 ] . If ρ 11 Λ ( C f ) = 0 \rho^\Lambda_{11}(C_f)=0 ρ 11 Λ ( C f ) = 0 , then μ f = 0 \mu_f=0 μ f = 0 ; if ρ 22 Λ ( C f ) = 0 \rho^\Lambda_{22}(C_f)=0 ρ 22 Λ ( C f ) = 0 , then μ f \mu_f μ f is concentrated at s = 3 4 s=\frac{3}{4} s = 4 3 , which is compatible with the monotonicity of f f f only if μ f = 0 \mu_f=0 μ f = 0 . In both cases f ( t ) = t f(t)=t f ( t ) = t and C f = Π C_f=\Pi C f = Π .
Thus lower semilinear copulas, which are positively quadrant dependent, always have non-negative arachnitude: they cannot be anti-cruciform. The most prominent example is the Cuadras–Augé family [5 ] , which is also a family of extreme-value copulas.
Example 5.9 (Cuadras–Augé copulas). For θ ∈ [ 0 , 1 ] \theta\in[0,1] θ ∈ [ 0 , 1 ] , the Cuadras–Augé copula C θ ( u , v ) = min { u , v } ( u v ) 1 − θ C_\theta(u,v)=\min\{u,v\}(uv)^{1-\theta} C θ ( u , v ) = min { u , v } ( uv ) 1 − θ is lower semilinear with f ( t ) = t 1 − θ f(t)=t^{1-\theta} f ( t ) = t 1 − θ . By Proposition 5.8 , or by direct integration,
ρ 11 Λ ( C θ ) = 3 θ 4 − θ , ρ 22 Λ ( C θ ) = 5 θ ( θ 2 − 3 θ + 14 ) ( 4 − θ ) ( 5 − θ ) ( 6 − θ ) , \rho^\Lambda_{11}(C_\theta)=\frac{3\theta}{4-\theta},\qquad\rho^\Lambda_{22}(C_\theta)=\frac{5\theta(\theta^2-3\theta+14)}{(4-\theta)(5-\theta)(6-\theta)}, ρ 11 Λ ( C θ ) = 4 − θ 3 θ , ρ 22 Λ ( C θ ) = ( 4 − θ ) ( 5 − θ ) ( 6 − θ ) 5 θ ( θ 2 − 3 θ + 14 ) , ARV ( C θ ) = 1 9 θ ( θ 2 − 3 θ + 14 ) ( 4 − θ ) ( 5 − θ ) ( 6 − θ ) . \operatorname{ARV}(C_\theta)=\frac{1}{9}\frac{\theta(\theta^2-3\theta+14)}{(4-\theta)(5-\theta)(6-\theta)}. ARV ( C θ ) = 9 1 ( 4 − θ ) ( 5 − θ ) ( 6 − θ ) θ ( θ 2 − 3 θ + 14 ) . The curve joins Π \Pi Π ( θ = 0 ) (\theta=0) ( θ = 0 ) to M M M ( θ = 1 ) (\theta=1) ( θ = 1 ) . For small θ \theta θ , ρ 22 Λ ( C θ ) = 7 12 θ + O ( θ 2 ) \rho^\Lambda_{22}(C_\theta)=\frac{7}{12}\theta+O(\theta^2) ρ 22 Λ ( C θ ) = 12 7 θ + O ( θ 2 ) and ρ 11 Λ ( C θ ) = 3 4 θ + O ( θ 2 ) \rho^\Lambda_{11}(C_\theta)=\frac{3}{4}\theta+O(\theta^2) ρ 11 Λ ( C θ ) = 4 3 θ + O ( θ 2 ) , so that ρ 22 Λ ≈ 7 9 ρ 11 Λ \rho^\Lambda_{22}\approx\frac{7}{9}\rho^\Lambda_{11} ρ 22 Λ ≈ 9 7 ρ 11 Λ : arachnitude grows linearly with Spearman’s rho, whereas for Gaussian copulas it grows quadratically (Proposition 5.6 ). This reflects the singular component of C θ C_\theta C θ on the main diagonal, which places mass at points where both ranks are simultaneously extreme (Figure 7 b).
Concluding remarks
We have determined the exact ( ARV , ρ ) (\operatorname{ARV},\rho) ( ARV , ρ ) region of bivariate copulas. Equivalently, the sharp inequality
∣ ρ ( C ) ∣ ≤ r ∗ ( ARV ( C ) ) , C ∈ C , \lvert\rho(C) \rvert\le r^{*}(\operatorname{ARV}(C)), \qquad C \in\mathcal{C}, ∣ ρ ( C )∣ ≤ r ∗ ( ARV ( C )) , C ∈ C ,
holds, with equality on the upper boundary for C a I C^{\mathrm{I}}_{a} C a I (0 ≤ a ≤ 2 5 0 \le a \le\frac{2}{5} 0 ≤ a ≤ 5 2 ), C b , c ( b ) I I C^{\mathrm{II}}_{b,c(b)} C b , c ( b ) II (0 ≤ b ≤ 1 5 0 \le b \le\frac{1}{5} 0 ≤ b ≤ 5 1 ) and C c I I I C^{\mathrm{III}}_{c} C c III (c ∗ ≤ c ≤ 1 c^{*} \le c \le1 c ∗ ≤ c ≤ 1 ), and on the lower boundary for their flipped versions. Since r ∗ r^{*} r ∗ is continuous and strictly increasing from 1 2 \frac{1}{2} 2 1 to 1 1 1 , the same result can be read as a sharp lower bound for ARV in terms of Spearman’s rho:
ARV ( C ) ≥ ( r ∗ ) − 1 ( ∣ ρ ( C ) ∣ ) whenever ∣ ρ ( C ) ∣ ≥ 1 2 , \operatorname{ARV}(C) \ge(r^{*})^{-1}(\lvert\rho(C) \rvert) \qquad\text{whenever } \lvert\rho(C) \rvert\ge\frac{1}{2}, ARV ( C ) ≥ ( r ∗ ) − 1 (∣ ρ ( C )∣) whenever ∣ ρ ( C )∣ ≥ 2 1 ,
while for ∣ ρ ( C ) ∣ ≤ 1 2 \lvert\rho(C) \rvert\le\frac{1}{2} ∣ ρ ( C )∣ ≤ 2 1 only the global bound ARV ( C ) ≥ 11 120 \operatorname{ARV}(C) \ge\frac{11}{120} ARV ( C ) ≥ 120 11 is available, and it is attained. The upper boundary is continuously differentiable away from its left endpoint (Corollary 4.3) and has a vertical tangent there (Corollary 4.4).
By § 5 , on copulas the average rectangular volume is an affine function of arachnitude, ρ 22 Λ = 45 ARV − 5 \rho_{22}^{\Lambda}=45\operatorname{ARV}-5 ρ 22 Λ = 45 ARV − 5 . An index introduced for aggregation functions through rectangular volumes thus turns out to be a second-order rank correlation, the natural companion of Spearman’s rho, and Theorem 4.1 is equivalently a sharp compatibility condition between first- and second-order rank dependence (Theorem 5.3 ).
Two ingredients of the proof may be useful beyond the present problem. First, the folding transformation turns ARV into a moment of the folded copula and gives a sharp upper bound for ∣ ρ ∣ \lvert\rho\rvert ∣ ρ ∣ in terms of another moment; the same reduction applies to every pair of functionals that, after folding, become moments of the form E [ g ( S ) h ( T ) ] \mathbb{E}[g(S)h(T)] E [ g ( S ) h ( T )] . Second, the supporting costs c κ c_{\kappa} c κ are neither supermodular nor submodular for 0 < κ < 2 0 < \kappa< 2 0 < κ < 2 , so the optimal couplings are not monotone rearrangements. They were obtained by constructing explicit Kantorovich potentials, and the algebraic parameter constraint P ( b , c ) = 0 P(b,c)=0 P ( b , c ) = 0 appeared as the continuity condition of the potential.
Declaration of generative AI and AI-assisted technologies in the manuscript preparation process
During the preparation of this work the authors used Claude (Anthropic) in order to explore the problem numerically, to propose and check constructions and proofs (in particular, related to Families II and III with the associated Kantorovich potentials), to produce the figures, and to improve the organization and the language of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data availability
No data were used for the research described in the article. The Python code that reproduces all figures is provided as supplementary material.
Acknowledgments
The authors are members of the group GNAMPA of INdAM, which is acknowledged.
Proof of the main result
This section contains the proof of Theorem 4.1 and of corollaries 4.3 and 4.4. By the folding transformation of Proposition 2.3 , the upper boundary is described by the maximization of E D [ S T ] \mathbb{E}_{D}[ST] E D [ S T ] over copulas D D D with a prescribed value of E D [ S 2 T 2 ] \mathbb{E}_{D}[S^{2}T^{2}] E D [ S 2 T 2 ] ; supporting lines lead to a family of optimal transport problems. For each family we exhibit a Kantorovich potential, which proves optimality. For Families II and III the potential is continuous exactly when P ( b , c ) = 0 P(b,c)=0 P ( b , c ) = 0 , which selects the parameter curve c = c ( b ) c=c(b) c = c ( b ) .
Supporting costs and Kantorovich duality
For κ > 0 \kappa>0 κ > 0 , define
h κ ( z ) = z − z 2 2 κ , c κ ( s , t ) = h κ ( s t ) = s t − s 2 t 2 2 κ . (32) h_{\kappa}(z)=z-\frac{z^{2}}{2\kappa},\qquad c_{\kappa}(s,t)=h_{\kappa}(st)=st-\frac{s^{2}t^{2}}{2\kappa}. \tag*{(32)} h κ ( z ) = z − 2 κ z 2 , c κ ( s , t ) = h κ ( s t ) = s t − 2 κ s 2 t 2 . (32)
Write x = E D [ S 2 T 2 ] x=\mathbb{E}_{D}[S^{2}T^{2}] x = E D [ S 2 T 2 ] and y = E D [ S T ] y=\mathbb{E}_{D}[ST] y = E D [ S T ] . The set of pairs ( x , y ) (x,y) ( x , y ) , D ∈ C D\in\mathcal{C} D ∈ C , is compact and convex, and by (5) its upper boundary is mapped onto the graph of r ∗ r^{*} r ∗ by the affine map ( x , y ) ↦ ( 1 12 + x 4 , 3 y ) (x,y)\mapsto\left(\frac{1}{12}+\frac{x}{4},3y\right) ( x , y ) ↦ ( 12 1 + 4 x , 3 y ) . A supporting line of positive finite slope 1 / ( 2 κ ) 1/(2\kappa) 1/ ( 2 κ ) in the ( x , y ) (x,y) ( x , y ) -plane is obtained by maximizing y − x / ( 2 κ ) y-x/(2\kappa) y − x / ( 2 κ ) , that is, by solving
sup D ∈ C E D [ c κ ( S , T ) ] . (33) \sup_{D\in\mathcal{C}}\mathbb{E}_{D}[c_{\kappa}(S,T)]. \tag*{(33)} D ∈ C sup E D [ c κ ( S , T )] . (33)
In the ( A R V , ρ ) (\mathrm{ARV},\rho) ( ARV , ρ ) -plane the corresponding slope is 6 / κ 6/\kappa 6/ κ . We shall construct optimizers for suitable values of κ \kappa κ and then show that their moment pairs cover the upper boundary. See, for example, [23 , 30 ] for the general theory of optimal transport. For our purposes, the following elementary form of Kantorovich duality is sufficient.
Lemma Appendix A.1 (Kantorovich duality: a sufficient optimality condition). Let D ∈ C D\in\mathcal{C} D ∈ C and κ > 0 \kappa>0 κ > 0 . Suppose that there exists a bounded Borel function φ : [ 0 , 1 ] → R \varphi:[0,1]\to\mathbb{R} φ : [ 0 , 1 ] → R such that
G ( s , t ) : = φ ( s ) + φ ( t ) − c κ ( s , t ) ≥ 0 for all ( s , t ) ∈ [ 0 , 1 ] 2 , G(s,t):=\varphi(s)+\varphi(t)-c_{\kappa}(s,t)\geq0 \qquad\text{for all }(s,t)\in[0,1]^{2}, G ( s , t ) := φ ( s ) + φ ( t ) − c κ ( s , t ) ≥ 0 for all ( s , t ) ∈ [ 0 , 1 ] 2 , and
G = 0 D -almost surely . G=0 \qquad D\text{-almost surely}. G = 0 D -almost surely . Then D D D solves (A.2). In particular, if
α D = 1 12 + 1 4 E D [ S 2 T 2 ] , \alpha_{D}=\frac{1}{12}+\frac{1}{4}\mathbb{E}_{D}[S^{2}T^{2}], α D = 12 1 + 4 1 E D [ S 2 T 2 ] , then
r ∗ ( α D ) = 3 E D [ S T ] . r^{*}(\alpha_{D})=3\mathbb{E}_{D}[ST]. r ∗ ( α D ) = 3 E D [ S T ] .
Proof. For every D ′ ∈ C D' \in\mathcal{C} D ′ ∈ C ,
E D ′ [ c κ ( S , T ) ] ≤ E D ′ [ ϕ ( S ) + ϕ ( T ) ] = 2 ∫ 0 1 ϕ ( u ) d u . \mathbb{E}_{D'}[c_\kappa(S,T)] \le\mathbb{E}_{D'}[\phi(S)+\phi(T)] = 2\int_0^1 \phi(u)\,\mathrm{d}u. E D ′ [ c κ ( S , T )] ≤ E D ′ [ ϕ ( S ) + ϕ ( T )] = 2 ∫ 0 1 ϕ ( u ) d u .
Since equality holds for D D D , this proves the optimality of D D D . If D ′ D' D ′ has the same value of E [ S 2 T 2 ] \mathbb{E}[S^2T^2] E [ S 2 T 2 ] as D D D , the inequality
E D ′ [ c κ ] ≤ E D [ c κ ] \mathbb{E}_{D'}[c_\kappa] \le\mathbb{E}_{D}[c_\kappa] E D ′ [ c κ ] ≤ E D [ c κ ]
reduces to
E D ′ [ S T ] ≤ E D [ S T ] . \mathbb{E}_{D'}[ST] \le\mathbb{E}_{D}[ST]. E D ′ [ S T ] ≤ E D [ S T ] .
The conclusion follows from (5) . □
Terminology. Lemma Appendix § 11 .§ 1 is the elementary half of Kantorovich duality; see [30, Ch. 5 and [25 Ch. 1] . The dual problem associated with (A.2) is
inf { ∫ 0 1 ϕ ( u ) d u + ∫ 0 1 ψ ( u ) d u : ϕ ( s ) + ψ ( t ) ≥ c κ ( s , t ) for all ( s , t ) ∈ [ 0 , 1 ] 2 } . \inf\left\{\int_0^1 \phi(u)\,\mathrm{d}u+\int_0^1 \psi(u)\,\mathrm{d}u:\ \phi(s)+\psi(t)\ge c_\kappa(s,t)\text{ for all }(s,t)\in[0,1]^2\right\}. inf { ∫ 0 1 ϕ ( u ) d u + ∫ 0 1 ψ ( u ) d u : ϕ ( s ) + ψ ( t ) ≥ c κ ( s , t ) for all ( s , t ) ∈ [ 0 , 1 ] 2 } .
and, since c κ c_\kappa c κ is symmetric and both marginals are uniform, it suffices to consider symmetric pairs ψ = ϕ \psi=\phi ψ = ϕ . A function ϕ \phi ϕ as in Lemma Appendix § 11 .§ 1 is called a Kantorovich potential (for D D D and the cost c κ c_\kappa c κ ): the pair ( ϕ , ϕ ) (\phi,\phi) ( ϕ , ϕ ) solves the dual problem and the primal and dual values coincide. We call G G G the duality slack ; the requirement G = 0 G=0 G = 0 D D D -almost surely is the complementary slackness condition . The set
Γ ϕ = { ( s , t ) ∈ [ 0 , 1 ] 2 : G ( s , t ) = 0 } \Gamma_\phi=\{(s,t)\in[0,1]^2:G(s,t)=0\} Γ ϕ = {( s , t ) ∈ [ 0 , 1 ] 2 : G ( s , t ) = 0 }
is the contact set of ϕ \phi ϕ , and its section
∂ c ϕ ( t ) = { s ∈ [ 0 , 1 ] : ( s , t ) ∈ Γ ϕ } \partial^c\phi(t)=\{s\in[0,1]:(s,t)\in\Gamma_\phi\} ∂ c ϕ ( t ) = { s ∈ [ 0 , 1 ] : ( s , t ) ∈ Γ ϕ }
is the c c c -superdifferential of ϕ \phi ϕ at t t t (the convention of [30 Ch. 5] , adapted to a maximization problem). If ϕ \phi ϕ is continuous, Γ ϕ \Gamma_\phi Γ ϕ is closed and complementary slackness is equivalent to supp ( D ) ⊆ Γ ϕ \operatorname{supp}(D)\subseteq\Gamma_\phi supp ( D ) ⊆ Γ ϕ .
For a copula D D D and t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] we write
Γ D ( t ) = { p ∈ [ 0 , 1 ] : ( p , t ) ∈ supp ( D ) } . \Gamma_D(t)=\{p\in[0,1]:(p,t)\in\operatorname{supp}(D)\}. Γ D ( t ) = { p ∈ [ 0 , 1 ] : ( p , t ) ∈ supp ( D )} .
for the set of points coupled with t t t by D D D . Under the assumptions of Lemma Appendix § 11 .§ 1 with ϕ \phi ϕ continuous, Γ D ( t ) ⊆ ∂ c ϕ ( t ) \Gamma_D(t)\subseteq\partial^c\phi(t) Γ D ( t ) ⊆ ∂ c ϕ ( t ) .
If ( p , t ) ∈ Γ ϕ (p,t)\in\Gamma_\phi ( p , t ) ∈ Γ ϕ with 0 < t < 1 0<t<1 0 < t < 1 and ϕ \phi ϕ is differentiable at t t t , then t t t minimizes the non-negative function G ( p , ⋅ ) G(p,\cdot) G ( p , ⋅ ) , hence ∂ t G ( p , t ) = 0 \partial_tG(p,t)=0 ∂ t G ( p , t ) = 0 , that is,
ϕ ′ ( t ) = ∂ t c κ ( p , t ) = p h κ ′ ( p t ) . \phi'(t)=\partial_tc_\kappa(p,t)=p h_\kappa'(pt). ϕ ′ ( t ) = ∂ t c κ ( p , t ) = p h κ ′ ( pt ) .
This first-order condition is the one-dimensional counterpart of the relation between the gradient of a Kantorovich potential and the optimal transport map; cf. [11 ] .
We shall repeatedly use the following factorization of ∂ t G \partial_t G ∂ t G . For t > 0 t>0 t > 0 , let p ∈ Γ D ( t ) p\in\Gamma_D(t) p ∈ Γ D ( t ) , assume that ϕ \phi ϕ is differentiable at t t t and
ϕ ′ ( t ) = p h κ ′ ( p t ) , \phi'(t)=p h_\kappa'(pt), ϕ ′ ( t ) = p h κ ′ ( pt ) ,
then
∂ G ∂ t ( s , t ) = p h κ ′ ( p t ) − s h κ ′ ( s t ) = ( p − s ) ( 1 − t ( p + s ) κ ) = t κ ( s − p ) ( s − q ) , q = κ t − p . (34) \frac{\partial G}{\partial t}(s,t)=p h_\kappa'(pt)-s h_\kappa'(st)
=(p-s)\left(1-\frac{t(p+s)}{\kappa}\right)
=\frac{t}{\kappa}(s-p)(s-q),\qquad q=\frac{\kappa}{t}-p. \tag*{(34)} ∂ t ∂ G ( s , t ) = p h κ ′ ( pt ) − s h κ ′ ( s t ) = ( p − s ) ( 1 − κ t ( p + s ) ) = κ t ( s − p ) ( s − q ) , q = t κ − p . (34)
Thus the sign of ∂ t G \partial_tG ∂ t G is determined by the position of s s s relative to the two points p p p and q q q (the second root q q q need not be coupled with t t t ).
Extremality of Family I
We first prove that Family I generates the rightmost arc Γ I \Gamma^{\mathrm I} Γ I of the upper boundary. By (7) , the folded copula of C a I C_a^{\mathrm I} C a I is D b D_b D b with
b = 1 − 2 a . (35) b=1-2a. \tag*{(35)} b = 1 − 2 a . (35)
Thus the range 0 ≤ a ≤ 2 / 5 0\leq a\leq2/5 0 ≤ a ≤ 2/5 corresponds to
1 5 ≤ b ≤ 1. \frac{1}{5}\leq b\leq1. 5 1 ≤ b ≤ 1.
Optimality for b ≥ 1 / 5 b\geq1/5 b ≥ 1/5
For b ∈ [ 1 / 5 , 1 ] b\in[1/5,1] b ∈ [ 1/5 , 1 ] , set
κ b = ( 1 + b ) ( 1 + 5 b ) 6 . (36) \kappa_b=\frac{(1+b)(1+5b)}{6}. \tag*{(36)} κ b = 6 ( 1 + b ) ( 1 + 5 b ) . (36)
The corresponding transport cost is
c κ b ( s , t ) = h κ b ( s t ) = s t − s 2 t 2 2 κ b . c_{\kappa_b}(s,t)=h_{\kappa_b}(st)=st-\frac{s^2t^2}{2\kappa_b}. c κ b ( s , t ) = h κ b ( s t ) = s t − 2 κ b s 2 t 2 .
Theorem A.2. For every a ∈ [ 0 , 2 / 5 ] a\in[0,2/5] a ∈ [ 0 , 2/5 ] ,
( ARV ( C a I ) , ρ ( C a I ) ) \left(\operatorname{ARV}(C_a^{\mathrm I}),\rho(C_a^{\mathrm I})\right) ( ARV ( C a I ) , ρ ( C a I ) ) lies on the upper boundary of A \mathcal{A} A . Equivalently, Γ I \Gamma^{\mathrm I} Γ I is a boundary arc.
Proof. The case a = 0 a=0 a = 0 is immediate, since C 0 I = M C_0^{\mathrm I}=M C 0 I = M . Assume therefore a > 0 a>0 a > 0 and let b = 1 − 2 a ∈ [ 1 / 5 , 1 ) b=1-2a\in[1/5,1) b = 1 − 2 a ∈ [ 1/5 , 1 ) .
We construct a Kantorovich potential for D b D_b D b . Define ϕ b : [ 0 , 1 ] → R \phi_b:[0,1]\to\mathbb{R} ϕ b : [ 0 , 1 ] → R by
ϕ b ( s ) = 1 2 h κ b ( s 2 ) , 0 ≤ s ≤ b , (37) \phi_b(s)=\frac{1}{2}h_{\kappa_b}(s^2),\qquad0\leq s\leq b, \tag*{(37)} ϕ b ( s ) = 2 1 h κ b ( s 2 ) , 0 ≤ s ≤ b , (37)
and
ϕ b ( s ) = ϕ b ( b ) + ∫ b s ( 1 + b − y ) h κ b ′ ( y ( 1 + b − y ) ) d y , b ≤ s ≤ 1. (38) \phi_b(s)=\phi_b(b)+\int_b^s (1+b-y)h'_{\kappa_b}(y(1+b-y))\,\mathrm{d}y,\qquad b\le s\le1. \tag*{(38)} ϕ b ( s ) = ϕ b ( b ) + ∫ b s ( 1 + b − y ) h κ b ′ ( y ( 1 + b − y )) d y , b ≤ s ≤ 1. (38)
Set
G b ( s , t ) = ϕ b ( s ) + ϕ b ( t ) − c κ b ( s , t ) . G_b(s,t)=\phi_b(s)+\phi_b(t)-c_{\kappa_b}(s,t). G b ( s , t ) = ϕ b ( s ) + ϕ b ( t ) − c κ b ( s , t ) .
On the diagonal piece, G b = 0 G_b=0 G b = 0 by definition. On the anti-diagonal, the derivative of s ↦ G b ( s , 1 + b − s ) s\mapsto G_b(s,1+b-s) s ↦ G b ( s , 1 + b − s ) is zero. Evaluating the integral in (A.7) at s = 1 s=1 s = 1 gives
ϕ b ( b ) + ϕ b ( 1 ) − h κ b ( b ) = ( 1 − b ) 2 12 κ ( 6 κ − ( 1 + b ) ( 1 + 5 b ) ) . \phi_b(b)+\phi_b(1)-h_{\kappa_b}(b)=\frac{(1-b)^2}{12\kappa}\left(6\kappa-(1+b)(1+5b)\right). ϕ b ( b ) + ϕ b ( 1 ) − h κ b ( b ) = 12 κ ( 1 − b ) 2 ( 6 κ − ( 1 + b ) ( 1 + 5 b ) ) .
For κ = κ b \kappa=\kappa_b κ = κ b , the right-hand side vanishes by (A.5). Hence
ϕ b ( b ) + ϕ b ( 1 ) = h κ b ( b ) . \phi_b(b)+\phi_b(1)=h_{\kappa_b}(b). ϕ b ( b ) + ϕ b ( 1 ) = h κ b ( b ) .
Hence the constant on the anti-diagonal is also zero. Thus G b G_b G b vanishes on
supp ( D b ) = { ( s , s ) : 0 ≤ s ≤ b } ∪ { ( s , t ) ∈ [ b , 1 ] 2 : s + t = 1 + b } . \operatorname{supp}(D_b)=\{(s,s):0\le s\le b\}\cup\{(s,t)\in[b,1]^2:s+t=1+b\}. supp ( D b ) = {( s , s ) : 0 ≤ s ≤ b } ∪ {( s , t ) ∈ [ b , 1 ] 2 : s + t = 1 + b } .
It remains to prove that G b ≥ 0 G_b\ge0 G b ≥ 0 on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 .
The block [ 0 , b ] 2 [0,b]^2 [ 0 , b ] 2 . A direct expansion gives
G b ( s , t ) = ( s − t ) 2 2 ( 1 − ( s + t ) 2 2 κ b ) . (39) G_b(s,t)=\frac{(s-t)^2}{2}\left(1-\frac{(s+t)^2}{2\kappa_b}\right). \tag*{(39)} G b ( s , t ) = 2 ( s − t ) 2 ( 1 − 2 κ b ( s + t ) 2 ) . (39)
Since
( s + t ) 2 ≤ 4 b 2 ≤ 2 κ b , (s+t)^2\le4b^2\le2\kappa_b, ( s + t ) 2 ≤ 4 b 2 ≤ 2 κ b ,
we obtain G b ( s , t ) ≥ 0 G_b(s,t)\ge0 G b ( s , t ) ≥ 0 . Indeed,
4 b 2 ≤ 2 κ b 4b^2\le2\kappa_b 4 b 2 ≤ 2 κ b
is equivalent to
12 b 2 ≤ ( 1 + b ) ( 1 + 5 b ) , 12b^2\le(1+b)(1+5b), 12 b 2 ≤ ( 1 + b ) ( 1 + 5 b ) ,
which holds for 0 ≤ b ≤ 1 0\le b\le1 0 ≤ b ≤ 1 .
The block [ b , 1 ] 2 [b,1]^2 [ b , 1 ] 2 . For ( s , t ) ∈ [ b , 1 ] 2 (s,t)\in[b,1]^2 ( s , t ) ∈ [ b , 1 ] 2 one obtains
G b ( s , t ) = − ( s + t − 1 − b ) 2 2 ( 1 + b ) ( 1 + 5 b ) Q b ( s , t ) , (40) G_b(s,t)=-\frac{(s+t-1-b)^2}{2(1+b)(1+5b)}Q_b(s,t), \tag*{(40)} G b ( s , t ) = − 2 ( 1 + b ) ( 1 + 5 b ) ( s + t − 1 − b ) 2 Q b ( s , t ) , (40)
where
Q b ( s , t ) = 3 ( s − t ) 2 − 2 ( 1 + b ) ( s + t ) + 4 b ( 1 + b ) . (41) Q_b(s,t)=3(s-t)^2-2(1+b)(s+t)+4b(1+b). \tag*{(41)} Q b ( s , t ) = 3 ( s − t ) 2 − 2 ( 1 + b ) ( s + t ) + 4 b ( 1 + b ) . (41)
Write
x = s + t , y = s − t . x=s+t,\qquad y=s-t. x = s + t , y = s − t .
If x ≤ 1 + b x\le1+b x ≤ 1 + b , then ∣ y ∣ ≤ x − 2 b |y|\le x-2b ∣ y ∣ ≤ x − 2 b , and therefore
Q b ( s , t ) ≤ 3 ( x − 2 b ) 2 − 2 ( 1 + b ) ( x − 2 b ) = ( x − 2 b ) [ 3 ( x − 2 b ) − 2 ( 1 + b ) ] ≤ 0 , Q_b(s,t)\le3(x-2b)^2-2(1+b)(x-2b)
=(x-2b)\left[3(x-2b)-2(1+b)\right]\le0, Q b ( s , t ) ≤ 3 ( x − 2 b ) 2 − 2 ( 1 + b ) ( x − 2 b ) = ( x − 2 b ) [ 3 ( x − 2 b ) − 2 ( 1 + b ) ] ≤ 0 ,
because
0 ≤ x − 2 b ≤ 1 − b 0 \le x - 2b \le1 - b 0 ≤ x − 2 b ≤ 1 − b
and
3 ( 1 − b ) ≤ 2 ( 1 + b ) ⟺ b ≥ 1 5 . 3(1-b) \le2(1+b) \Longleftrightarrow b \ge\frac{1}{5}. 3 ( 1 − b ) ≤ 2 ( 1 + b ) ⟺ b ≥ 5 1 .
If x ≥ 1 + b x \ge1+b x ≥ 1 + b , then ∣ y ∣ ≤ 2 − x |y| \le2-x ∣ y ∣ ≤ 2 − x . Setting z = 2 − x z=2-x z = 2 − x , so that 0 ≤ z ≤ 1 − b 0 \le z \le1-b 0 ≤ z ≤ 1 − b , we obtain
Q b ( s , t ) ≤ 3 z 2 + 2 ( 1 + b ) z − 4 ( 1 − b 2 ) . Q_b(s,t) \le3z^2 + 2(1+b)z - 4(1-b^2). Q b ( s , t ) ≤ 3 z 2 + 2 ( 1 + b ) z − 4 ( 1 − b 2 ) .
The right-hand side is increasing in z ≥ 0 z \ge0 z ≥ 0 , and at z = 1 − b z=1-b z = 1 − b it equals
( 1 − b ) ( 1 − 5 b ) ≤ 0. (1-b)(1-5b) \le0. ( 1 − b ) ( 1 − 5 b ) ≤ 0.
Hence Q b ≤ 0 Q_b \le0 Q b ≤ 0 also in this case, and (A.9) yields G b ≥ 0 G_b \ge0 G b ≥ 0 on [ b , 1 ] 2 [b,1]^2 [ b , 1 ] 2 .
The mixed blocks. By symmetry it is enough to consider $0 \le s \le b \le t \le 1$. For fixed t t t , let
F ( s ) = G b ( s , t ) . F(s)=G_b(s,t). F ( s ) = G b ( s , t ) .
From (A.6),
F ′ ′ ( s ) = 1 − 3 s 2 κ b + t 2 κ b ≥ 1 − 2 b 2 κ b ≥ 0. F''(s)=1-\frac{3s^2}{\kappa_b}+\frac{t^2}{\kappa_b}\ge1-\frac{2b^2}{\kappa_b}\ge0. F ′′ ( s ) = 1 − κ b 3 s 2 + κ b t 2 ≥ 1 − κ b 2 b 2 ≥ 0.
Moreover,
F ′ ( b − ) = b h κ b ′ ( b 2 ) − t h κ b ′ ( b t ) = ( b − t ) ( 1 − b ( b + t ) κ b ) ≤ 0 , \begin{aligned}
F'(b^-)=b h'_{\kappa_b}(b^2)-t h'_{\kappa_b}(bt) \\
&=(b-t)\left(1-\frac{b(b+t)}{\kappa_b}\right)\le0,
\end{aligned} F ′ ( b − ) = b h κ b ′ ( b 2 ) − t h κ b ′ ( b t ) = ( b − t ) ( 1 − κ b b ( b + t ) ) ≤ 0 ,
because
κ b ≥ b ( 1 + b ) . \kappa_b\ge b(1+b). κ b ≥ b ( 1 + b ) .
Since F F F is convex, F ′ F' F ′ is non-decreasing on [ 0 , b ] [0,b] [ 0 , b ] , hence F ′ ( s ) ≤ F ′ ( b − ) ≤ 0 F'(s)\le F'(b^-)\le0 F ′ ( s ) ≤ F ′ ( b − ) ≤ 0 for s ≤ b s\le b s ≤ b . Therefore
G b ( s , t ) = F ( s ) ≥ F ( b ) = G b ( b , t ) ≥ 0 , G_b(s,t)=F(s)\ge F(b)=G_b(b,t)\ge0, G b ( s , t ) = F ( s ) ≥ F ( b ) = G b ( b , t ) ≥ 0 ,
where the last inequality follows from the already treated block [ b , 1 ] 2 [b,1]^2 [ b , 1 ] 2 .
We have thus proved
G b ≥ 0 on [ 0 , 1 ] 2 , G_b\ge0 \quad\text{on } [0,1]^2, G b ≥ 0 on [ 0 , 1 ] 2 ,
with equality D b D_b D b -almost surely. Hence ϕ b \phi_b ϕ b is a Kantorovich potential for D b D_b D b . Lemma Appendix § 11 .§ 1 shows that D b D_b D b is optimal for the corresponding transport problem. Since
( C a l ) F = D b , \left(C_a^l\right)^F=D_b, ( C a l ) F = D b ,
Lemma Appendix § 11 .§ 1 , (7) and ρ ( D b u ) = 3 E D b [ S T ] \rho(D_b^u)=3\mathbb{E}_{D_b}[ST] ρ ( D b u ) = 3 E D b [ S T ] give
r ∗ ( ARV ( C a l ) ) = ρ ( C a l ) , r^*\left(\operatorname{ARV}(C_a^l)\right)=\rho(C_a^l), r ∗ ( ARV ( C a l ) ) = ρ ( C a l ) ,
as claimed.
Remark Appendix A.3. The value of κ b \kappa_b κ b is determined by the tangent to the curve generated by Family I. More precisely, for 0 ≤ b < 1 0 \le b < 1 0 ≤ b < 1 ,
1 2 κ b = 3 ( 1 + b ) ( 1 + 5 b ) = d d b E D b [ S T ] d d b E D b [ S 2 T 2 ] . \frac{1}{2\kappa_b} = \frac{3}{(1+b)(1+5b)} = \frac{\frac{d}{db}\mathbb{E}_{D_b}[ST]}{\frac{d}{db}\mathbb{E}_{D_b}[S^2T^2]}. 2 κ b 1 = ( 1 + b ) ( 1 + 5 b ) 3 = d b d E D b [ S 2 T 2 ] d b d E D b [ S T ] .
Thus the cost used in the proof is the supporting cost associated with the tangent to Γ I \Gamma^{\mathrm{I}} Γ I .
Sharpness of the threshold
The restriction a ≤ 2 / 5 a \le2/5 a ≤ 2/5 in Theorem A.2 is sharp.
Proposition Appendix A.4. For
2 5 < a < 1 2 , \frac{2}{5} < a < \frac{1}{2}, 5 2 < a < 2 1 , the point
( ARV ( C a I ) , ρ ( C a I ) ) \left(\operatorname{ARV}(C_a^{\mathrm{I}}),\rho(C_a^{\mathrm{I}})\right) ( ARV ( C a I ) , ρ ( C a I ) ) does not lie on the upper boundary of A \mathcal{A} A .
Proof. Let
b = 1 − 2 a , b = 1 - 2a, b = 1 − 2 a , so that 0 < b < 1 / 5 0 < b < 1/5 0 < b < 1/5 , and let κ b \kappa_b κ b be given by (A.5). The mixed derivative of the corresponding cost is
∂ 2 c κ b ∂ s ∂ t ( s , t ) = 1 − 2 s t κ b . \frac{\partial^2 c_{\kappa_b}}{\partial s\,\partial t}(s,t) = 1 - \frac{2st}{\kappa_b}. ∂ s ∂ t ∂ 2 c κ b ( s , t ) = 1 − κ b 2 s t . At the corner ( b , 1 ) (b,1) ( b , 1 ) of the anti-diagonal part of supp ( D b ) \operatorname{supp}(D_b) supp ( D b ) ,
κ b − 2 b = ( 1 − b ) ( 1 − 5 b ) 6 > 0. \kappa_b - 2b = \frac{(1-b)(1-5b)}{6} > 0. κ b − 2 b = 6 ( 1 − b ) ( 1 − 5 b ) > 0. Hence
∂ 2 c κ b ∂ s ∂ t ( b , 1 ) > 0 , \frac{\partial^2 c_{\kappa_b}}{\partial s\,\partial t}(b,1) > 0, ∂ s ∂ t ∂ 2 c κ b ( b , 1 ) > 0 , and the cost is strictly supermodular in a neighborhood of ( b , 1 ) (b,1) ( b , 1 ) .
We make the local improvement explicit. Choose two disjoint intervals I I I and I + d I+d I + d of equal positive length in a sufficiently small right neighborhood of b b b . On these intervals the original map is T ( s ) = 1 + b − s T(s)=1+b-s T ( s ) = 1 + b − s . Replace it by T ( s + d ) T(s+d) T ( s + d ) for s ∈ I s \in I s ∈ I and by T ( s − d ) T(s-d) T ( s − d ) for s ∈ I + d s \in I+d s ∈ I + d , leaving it unchanged elsewhere. This exchanges the two image intervals, so both uniform marginals are preserved.
For s ∈ I s \in I s ∈ I , put x 1 = s x_1=s x 1 = s , x 2 = s + d x_2=s+d x 2 = s + d , y 1 = T ( s + d ) y_1=T(s+d) y 1 = T ( s + d ) and y 2 = T ( s ) y_2=T(s) y 2 = T ( s ) . The new matching pairs ( x 1 , y 1 ) (x_1,y_1) ( x 1 , y 1 ) and ( x 2 , y 2 ) (x_2,y_2) ( x 2 , y 2 ) , whereas the old matching pairs ( x 1 , y 2 ) (x_1,y_2) ( x 1 , y 2 ) and ( x 2 , y 1 ) (x_2,y_1) ( x 2 , y 1 ) . Its gain is
c κ b ( x 1 , y 1 ) + c κ b ( x 2 , y 2 ) − c κ b ( x 1 , y 2 ) − c κ b ( x 2 , y 1 ) = ∫ x 1 x 2 ∫ y 1 y 2 ∂ 2 c κ b ∂ x ∂ y ( x , y ) d y d x > 0. c_{\kappa_b}(x_1,y_1) + c_{\kappa_b}(x_2,y_2) - c_{\kappa_b}(x_1,y_2) - c_{\kappa_b}(x_2,y_1)
= \int_{x_1}^{x_2}\int_{y_1}^{y_2}\frac{\partial^2 c_{\kappa_b}}{\partial x\,\partial y}(x,y)\,\mathrm{d}y\,\mathrm{d}x > 0. c κ b ( x 1 , y 1 ) + c κ b ( x 2 , y 2 ) − c κ b ( x 1 , y 2 ) − c κ b ( x 2 , y 1 ) = ∫ x 1 x 2 ∫ y 1 y 2 ∂ x ∂ y ∂ 2 c κ b ( x , y ) d y d x > 0.
The intervals can be chosen so that every such rectangle lies in the neighborhood where the mixed derivative is positive. Integrating over I I I proves that D b D_b D b does not maximize E D b [ c κ b ( S , T ) ] \mathbb{E}_{D_b}[c_{\kappa_b}(S,T)] E D b [ c κ b ( S , T )] .
It remains to connect this fact with the boundary problem. In the moment plane, put
x ( b ) = E D b [ S 2 T 2 ] , y ( b ) = E D b [ S T ] . x(b)=\mathbb{E}_{D_b}[S^2T^2], \qquad y(b)=\mathbb{E}_{D_b}[ST]. x ( b ) = E D b [ S 2 T 2 ] , y ( b ) = E D b [ S T ] .
For 0 < b < 1 / 5 0<b<1/5 0 < b < 1/5 the parameter value is interior and
x ′ ( b ) = ( 1 − b ) 2 ( 1 + b ) ( 1 + 5 b ) 6 > 0. x'(b)=\frac{(1-b)^2(1+b)(1+5b)}{6}>0. x ′ ( b ) = 6 ( 1 − b ) 2 ( 1 + b ) ( 1 + 5 b ) > 0.
If ( x ( b ) , y ( b ) ) (x(b),y(b)) ( x ( b ) , y ( b )) were on the upper boundary of the convex feasible moment region, any supporting line there would have to be tangent to the differentiable feasible curve b ↦ ( x ( b ) , y ( b ) ) b\mapsto(x(b),y(b)) b ↦ ( x ( b ) , y ( b )) . Its slope would therefore be
y ′ ( b ) x ′ ( b ) = 1 2 κ b \frac{y'(b)}{x'(b)}=\frac{1}{2\kappa_b} x ′ ( b ) y ′ ( b ) = 2 κ b 1
by Remark Appendix A.3. Thus D b D_b D b would maximize
y − 1 2 κ b x = E D b [ c κ b ( S , T ) ] , y-\frac{1}{2\kappa_b}x=\mathbb{E}_{D_b}[c_{\kappa_b}(S,T)], y − 2 κ b 1 x = E D b [ c κ b ( S , T )] ,
contradicting the preceding local improvement. Consequently
( ARV ( C a I ) , ρ ( C a I ) ) \left(\operatorname{ARV}(C_a^{\mathrm{I}}),\rho(C_a^{\mathrm{I}})\right) ( ARV ( C a I ) , ρ ( C a I ) )
cannot lie on the upper boundary.
Extremality of Families II and III
We now prove the extremality of the two new families introduced in § 3 . After folding, both are described by the same family D b , c D_{b,c} D b , c : Family II corresponds to b > 0 b>0 b > 0 , whereas Family III is the limiting case b = 0 b=0 b = 0 .
The proof has three ingredients. First we record the geometry of the support of D b , c D_{b,c} D b , c . We then construct a candidate Kantorovich potential and show that the polynomial P ( b , c ) P(b,c) P ( b , c ) introduced in § 4 is precisely the condition that makes this potential continuous. Finally, a monotonicity argument reduces the global inequality for the Kantorovich slack to its values on the support and at one junction point.
Geometry of the folded family
Throughout this section, let
Δ = { ( b , c ) : 0 ≤ b ≤ 1 5 , 1 2 + b ≤ c ≤ 1 } ∖ { ( 0 , 1 ) } . (42) \Delta= \left\{(b,c): 0 \le b \le\frac{1}{5},\ \frac{1}{2}+b \le c \le1\right\}\setminus\{(0,1)\}. \tag*{(42)} Δ = { ( b , c ) : 0 ≤ b ≤ 5 1 , 2 1 + b ≤ c ≤ 1 } ∖ {( 0 , 1 )} . (42)
and set
κ = c ( 1 + 2 b − c ) , b ′ = 1 + b − c . (43) \kappa= c(1+2b-c),\qquad b' = 1+b-c. \tag*{(43)} κ = c ( 1 + 2 b − c ) , b ′ = 1 + b − c . (43)
Then κ > 0 \kappa>0 κ > 0 on Δ \Delta Δ , and
κ − b ( 1 + b ) = ( c − b ) ( 1 + b − c ) ≥ 0 , (44) \kappa-b(1+b)=(c-b)(1+b-c)\ge0, \tag*{(44)} κ − b ( 1 + b ) = ( c − b ) ( 1 + b − c ) ≥ 0 , (44)
κ − 2 b = ( 1 − c ) ( c − 2 b ) ≥ 0 , (45) \kappa-2b=(1-c)(c-2b)\ge0, \tag*{(45)} κ − 2 b = ( 1 − c ) ( c − 2 b ) ≥ 0 , (45)
c 2 − κ = c ( 2 c − 1 − 2 b ) ≥ 0 , (46) c^2-\kappa=c(2c-1-2b)\ge0, \tag*{(46)} c 2 − κ = c ( 2 c − 1 − 2 b ) ≥ 0 , (46)
u ( 1 − u ) ≤ c ( 1 − c ) ≤ κ , c ≤ u ≤ 1. (47) u(1-u)\le c(1-c)\le\kappa,\qquad c\le u\le1. \tag*{(47)} u ( 1 − u ) ≤ c ( 1 − c ) ≤ κ , c ≤ u ≤ 1. (47)
Recall that
ξ 1 ( t ) = 1 2 ( κ t + t − 1 ) , ξ 2 ( t ) = 1 2 ( κ t − t + 1 ) , c ≤ t ≤ 1. \xi_1(t)=\frac{1}{2}\left(\frac{\kappa}{t}+t-1\right),\qquad
\xi_2(t)=\frac{1}{2}\left(\frac{\kappa}{t}-t+1\right),\qquad c\le t\le1. ξ 1 ( t ) = 2 1 ( t κ + t − 1 ) , ξ 2 ( t ) = 2 1 ( t κ − t + 1 ) , c ≤ t ≤ 1.
Their endpoint values are
ξ 1 ( c ) = b , ξ 2 ( c ) = b ′ , ξ 1 ( 1 ) = ξ 2 ( 1 ) = κ 2 . (48) \xi_1(c)=b,\qquad\xi_2(c)=b',\qquad\xi_1(1)=\xi_2(1)=\frac{\kappa}{2}. \tag*{(48)} ξ 1 ( c ) = b , ξ 2 ( c ) = b ′ , ξ 1 ( 1 ) = ξ 2 ( 1 ) = 2 κ . (48)
Moreover, ξ 1 \xi_1 ξ 1 is increasing (since ξ 1 ′ ( t ) = 1 2 ( 1 − κ / t 2 ) ≥ 0 \xi_1'(t)=\frac{1}{2}(1-\kappa/t^2)\ge0 ξ 1 ′ ( t ) = 2 1 ( 1 − κ / t 2 ) ≥ 0 by (A.15)), ξ 2 \xi_2 ξ 2 is decreasing, and
0 ≤ b ≤ κ 2 ≤ b ′ ≤ c . 0\le b\le\frac{\kappa}{2}\le b'\le c. 0 ≤ b ≤ 2 κ ≤ b ′ ≤ c .
Let
τ : [ b , b ′ ] ⟶ [ c , 1 ] \tau:[b,b']\longrightarrow[c,1] τ : [ b , b ′ ] ⟶ [ c , 1 ]
denote the inverse of ξ 1 \xi_1 ξ 1 on [ b , κ / 2 ] [b,\kappa/2] [ b , κ /2 ] and of ξ 2 \xi_2 ξ 2 on [ κ / 2 , b ′ ] [\kappa/2,b'] [ κ /2 , b ′ ] . Explicitly,
τ ( s ) = 1 2 ( 1 + 2 s + ( 1 + 2 s ) 2 − 4 κ ) , b ≤ s ≤ κ 2 , \tau(s)=\frac{1}{2}\left(1+2s+\sqrt{(1+2s)^2-4\kappa}\right),\qquad b\le s\le\frac{\kappa}{2}, τ ( s ) = 2 1 ( 1 + 2 s + ( 1 + 2 s ) 2 − 4 κ ) , b ≤ s ≤ 2 κ ,
and
τ ( s ) = 1 2 ( 1 − 2 s + ( 1 − 2 s ) 2 + 4 κ ) , κ 2 ≤ s ≤ b ′ . \tau(s)=\frac{1}{2}\left(1-2s+\sqrt{(1-2s)^2+4\kappa}\right),\qquad\frac{\kappa}{2}\le s\le b'. τ ( s ) = 2 1 ( 1 − 2 s + ( 1 − 2 s ) 2 + 4 κ ) , 2 κ ≤ s ≤ b ′ .
Along the two curved branches,
t ξ 1 ( t ) = 1 2 ( κ − t ( 1 − t ) ) ≤ κ 2 , t ξ 2 ( t ) = 1 2 ( κ + t ( 1 − t ) ) ≥ κ 2 . (49) t\xi_1(t)=\frac{1}{2}\left(\kappa-t(1-t)\right)\le\frac{\kappa}{2},\qquad t\xi_2(t)=\frac{1}{2}\left(\kappa+t(1-t)\right)\ge\frac{\kappa}{2}. \tag*{(49)} t ξ 1 ( t ) = 2 1 ( κ − t ( 1 − t ) ) ≤ 2 κ , t ξ 2 ( t ) = 2 1 ( κ + t ( 1 − t ) ) ≥ 2 κ . (49)
As in Section A.1, we write Γ ( t ) = Γ D b , c ( t ) \Gamma(t)=\Gamma_{D_{b,c}}(t) Γ ( t ) = Γ D b , c ( t ) for the set of points coupled with t t t (the relation is symmetric because D b , c D_{b,c} D b , c is). Away from the junction points, these sets are
Table 3.
The candidate potential and the role of P ( b , c ) P(b,c) P ( b , c )
Fix ( b , c ) ∈ Δ (b,c)\in\Delta ( b , c ) ∈ Δ and write
h = h κ , m = 1 + b 2 . h=h_\kappa,\qquad m=\frac{1+b}{2}. h = h κ , m = 2 1 + b .
We define ϕ = ϕ b , c : [ 0 , 1 ] → R \phi=\phi_{b,c}:[0,1]\to\mathbb{R} ϕ = ϕ b , c : [ 0 , 1 ] → R so that the first-order condition holds along the support:
ϕ ( s ) = 1 2 h ( s 2 ) , 0 ≤ s < b , (50) \phi(s)=\frac{1}{2}h(s^2),\qquad0\le s<b, \tag*{(50)} ϕ ( s ) = 2 1 h ( s 2 ) , 0 ≤ s < b , (50)
ϕ ( s ) = 1 2 h ( m 2 ) + ∫ m s ( 1 + b − y ) h ′ ( y ( 1 + b − y ) ) d y , b ′ ≤ s ≤ c , (51) \phi(s)=\frac{1}{2}h(m^2)+\int_m^s(1+b-y)h'\bigl(y(1+b-y)\bigr)\,dy,\qquad b'\le s\le c, \tag*{(51)} ϕ ( s ) = 2 1 h ( m 2 ) + ∫ m s ( 1 + b − y ) h ′ ( y ( 1 + b − y ) ) d y , b ′ ≤ s ≤ c , (51)
ϕ ( t ) = ϕ ( c ) + ∫ c t u ξ 1 ( u ) ξ 2 ( u ) κ d u , c ≤ t ≤ 1 , (52) \phi(t)=\phi(c)+\int_c^t\frac{u\,\xi_1(u)\xi_2(u)}{\kappa}\,du,\qquad c\le t\le1, \tag*{(52)} ϕ ( t ) = ϕ ( c ) + ∫ c t κ u ξ 1 ( u ) ξ 2 ( u ) d u , c ≤ t ≤ 1 , (52)
and
ϕ ( s ) = h ( s τ ( s ) ) − ϕ ( τ ( s ) ) , b ≤ s ≤ b ′ . (53) \phi(s)=h\bigl(s\tau(s)\bigr)-\phi\bigl(\tau(s)\bigr),\qquad b\le s\le b'. \tag*{(53)} ϕ ( s ) = h ( s τ ( s ) ) − ϕ ( τ ( s ) ) , b ≤ s ≤ b ′ . (53)
Define the corresponding slack
G ( s , t ) = ϕ ( s ) + ϕ ( t ) − h ( s t ) . (54) G(s,t)=\phi(s)+\phi(t)-h(st). \tag*{(54)} G ( s , t ) = ϕ ( s ) + ϕ ( t ) − h ( s t ) . (54)
The construction is designed so that G G G vanishes on the support. The only non-automatic compatibility condition occurs at s = b s=b s = b .
Lemma Appendix A.7. Let ( b , c ) ∈ Δ (b,c)\in\Delta ( b , c ) ∈ Δ .
(a) Away from b b b , ϕ \phi ϕ is continuously differentiable on ( 0 , 1 ) (0,1) ( 0 , 1 ) and, for every p ∈ Γ ( t ) p\in\Gamma(t) p ∈ Γ ( t ) ,
ϕ ′ ( t ) = p h ′ ( p t ) . (55) \phi'(t)=p h'(pt). \tag*{(55)} ϕ ′ ( t ) = p h ′ ( pt ) . (55) (b)
for every ( s , t ) ∈ supp ( D b , c ) (s,t)\in\operatorname{supp}(D_{b,c}) ( s , t ) ∈ supp ( D b , c ) with s , t ≠ b s,t\ne b s , t = b .
(c) The mismatch at b b b is
ϕ ( b + ) − 1 2 h ( b 2 ) = − P ( b , c ) 24 κ . (56) \phi(b+) - \frac{1}{2}h(b^2) = -\frac{P(b,c)}{24\kappa}. \tag*{(56)} ϕ ( b + ) − 2 1 h ( b 2 ) = − 24 κ P ( b , c ) . (56)
If b = 0 b=0 b = 0 , this reads
ϕ ( 0 ) = − P ( 0 , c ) 24 κ . (57) \phi(0) = -\frac{P(0,c)}{24\kappa}. \tag*{(57)} ϕ ( 0 ) = − 24 κ P ( 0 , c ) . (57)
Proof. On the diagonal and anti-diagonal pieces, (A.24) follows directly from (A.19)–(A.20). For t ∈ [ c , 1 ] t \in[c,1] t ∈ [ c , 1 ] and p ∈ { ξ 1 ( t ) , ξ 2 ( t ) } p \in\{\xi_1(t),\xi_2(t)\} p ∈ { ξ 1 ( t ) , ξ 2 ( t )} , the other branch is q = κ / t − p q=\kappa/t-p q = κ / t − p , so
p h ′ ( p t ) = p ( κ − p t ) κ = p q t κ = t ξ 1 ( t ) ξ 2 ( t ) κ , p h'(pt) = \frac{p(\kappa-pt)}{\kappa} = \frac{pqt}{\kappa} = \frac{t \xi_1(t)\xi_2(t)}{\kappa}, p h ′ ( pt ) = κ p ( κ − pt ) = κ pq t = κ t ξ 1 ( t ) ξ 2 ( t ) ,
which is exactly the derivative in (A.21). Finally, differentiating (A.22) gives
ϕ ′ ( s ) = τ ( s ) h ′ ( τ ( s ) s ) , \phi'(s) = \tau(s)h'(\tau(s)s), ϕ ′ ( s ) = τ ( s ) h ′ ( τ ( s ) s ) ,
because s ∈ Γ ( τ ( s ) ) s \in\Gamma(\tau(s)) s ∈ Γ ( τ ( s )) .
The expressions defining ϕ \phi ϕ agree at c c c by construction and at κ / 2 \kappa/2 κ /2 because both inverse branches give τ ( κ / 2 ) = 1 \tau(\kappa/2)=1 τ ( κ /2 ) = 1 . At b ′ b' b ′ , agreement follows from the anti-diagonal identity proved below:
ϕ ( b ′ ) + ϕ ( c ) = h ( b ′ c ) . \phi(b')+\phi(c)=h(b'c). ϕ ( b ′ ) + ϕ ( c ) = h ( b ′ c ) .
At b ′ b' b ′ and κ / 2 \kappa/2 κ /2 , the coupled points given by the adjacent formulas coincide. At c c c , the point coupled with c c c along the anti-diagonal is b ′ = ξ 2 ( c ) b'=\xi_2(c) b ′ = ξ 2 ( c ) , and the two formulas along the curved branches give the same derivative by the identity above. Hence ϕ \phi ϕ is continuously differentiable away from b b b , proving (a).
Part (b) is immediate on the diagonal and on the curved branches. On the anti-diagonal, the derivative of
ϕ ( s ) + ϕ ( 1 + b − s ) − h ( s ( 1 + b − s ) ) \phi(s)+\phi(1+b-s)-h\left(s(1+b-s)\right) ϕ ( s ) + ϕ ( 1 + b − s ) − h ( s ( 1 + b − s ) )
vanishes identically, and the expression itself vanishes at s = m s=m s = m .
For (c), note that τ ( b ) = c \tau(b)=c τ ( b ) = c . Hence
ϕ ( b + ) = h ( b c ) − ϕ ( c ) . \phi(b+) = h(bc)-\phi(c). ϕ ( b + ) = h ( b c ) − ϕ ( c ) .
Substituting (A.20), and using κ = c ( 1 + 2 b − c ) \kappa=c(1+2b-c) κ = c ( 1 + 2 b − c ) , direct integration gives
24 κ ϕ ( c ) = 6 c 4 − 20 c 3 + 18 c 2 − 6 c + 1 − 32 b c 3 + 60 b c 2 − 24 b c + 4 b + 42 b 2 c 2 − 30 b 2 c + 6 b 2 − 12 b 3 c + 4 b 3 + b 4 , 24\kappa\phi(c) = 6c^4-20c^3+18c^2-6c+1-32bc^3+60bc^2-24bc+4b
+42b^2c^2-30b^2c+6b^2-12b^3c+4b^3+b^4, 24 κ ϕ ( c ) = 6 c 4 − 20 c 3 + 18 c 2 − 6 c + 1 − 32 b c 3 + 60 b c 2 − 24 b c + 4 b + 42 b 2 c 2 − 30 b 2 c + 6 b 2 − 12 b 3 c + 4 b 3 + b 4 ,
whereas
24 κ ( h ( b c ) − 1 2 h ( b 2 ) ) = 24 κ b c − 12 b 2 c 2 − 12 κ b 2 + 6 b 4 . 24\kappa\left(h(bc)-\frac{1}{2}h(b^2)\right)=24\kappa bc-12b^2c^2-12\kappa b^2+6b^4. 24 κ ( h ( b c ) − 2 1 h ( b 2 ) ) = 24 κb c − 12 b 2 c 2 − 12 κ b 2 + 6 b 4 .
Subtracting the two expressions and simplifying yields
24 κ ( ϕ ( b + ) − ϕ ( b − ) ) = − P ( b , c ) , 24\kappa\left(\phi(b+)-\phi(b-)\right)=-P(b,c), 24 κ ( ϕ ( b + ) − ϕ ( b − ) ) = − P ( b , c ) ,
which proves (A.25).
Thus, for b > 0 b>0 b > 0 ,
P ( b , c ) = 0 ⟺ ϕ b , c is continuous at b . P(b,c)=0 \Longleftrightarrow\phi_{b,c}\text{ is continuous at }b. P ( b , c ) = 0 ⟺ ϕ b , c is continuous at b .
This is the reason why the polynomial P P P appears in the parametrization of Γ I I \Gamma^{\mathrm{II}} Γ II .
A monotonicity lemma for the slack
The remaining task is to show that G ≥ 0 G \ge0 G ≥ 0 away from the support. The factorization (A.3) turns this into a one-dimensional monotonicity problem.
If p = p ( t ) ∈ Γ ( t ) p=p(t)\in\Gamma(t) p = p ( t ) ∈ Γ ( t ) , put
q ( t ) = κ t − p ( t ) . q(t)=\frac{\kappa}{t}-p(t). q ( t ) = t κ − p ( t ) .
Then
∂ G ∂ t ( s , t ) = t κ ( s − p ( t ) ) ( s − q ( t ) ) . (58) \frac{\partial G}{\partial t}(s,t)=\frac{t}{\kappa}(s-p(t))(s-q(t)). \tag*{(58)} ∂ t ∂ G ( s , t ) = κ t ( s − p ( t )) ( s − q ( t )) . (58)
Hence, for fixed s s s , the function t ↦ G ( s , t ) t\mapsto G(s,t) t ↦ G ( s , t ) is strictly decreasing exactly when s s s lies strictly between the two roots p ( t ) p(t) p ( t ) and q ( t ) q(t) q ( t ) .
The following lemma contains the geometric information needed below. Local minima at the endpoints of [ 0 , 1 ] [0,1] [ 0 , 1 ] are understood in the one-sided sense.
Lemma Appendix A.8. Assume that ϕ b , c \phi_{b,c} ϕ b , c is continuous on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Fix s ∈ [ b , 1 ] s\in[b,1] s ∈ [ b , 1 ] . Every local minimum of
t ↦ G ( s , t ) t\mapsto G(s,t) t ↦ G ( s , t ) occurs either at t = b t=b t = b or at a point of Γ ( s ) \Gamma(s) Γ ( s ) , i.e. on the contact set.
Proof. We examine the two roots in (A.27) on the successive intervals determined by the support. Empty intervals are omitted; in particular, when c = 1 c=1 c = 1 we have b = κ / 2 = b ′ b=\kappa/2=b' b = κ /2 = b ′ .
The interval 0 < t < b 0<t<b 0 < t < b . Here p ( t ) = t p(t)=t p ( t ) = t . By (A.13),
q ( t ) = κ t − t > κ b − b ≥ 1. q(t)=\frac{\kappa}{t}-t>\frac{\kappa}{b}-b\ge1. q ( t ) = t κ − t > b κ − b ≥ 1.
Thus, for s ≥ b s\ge b s ≥ b , the function G ( s , ⋅ ) G(s,\cdot) G ( s , ⋅ ) is decreasing throughout ( 0 , b ) (0,b) ( 0 , b ) . This interval can therefore contribute only the endpoint t = b t=b t = b as a local minimum.
The interval b < t < κ / 2 b<t<\kappa/2 b < t < κ /2 . Here p ( t ) = τ ( t ) p(t)=\tau(t) p ( t ) = τ ( t ) and t = ξ 1 ( p ( t ) ) t=\xi_1(p(t)) t = ξ 1 ( p ( t )) . By (A.18),
t p ( t ) ≤ κ 2 , t\,p(t)\le\frac{\kappa}{2}, t p ( t ) ≤ 2 κ ,
and hence
q ( t ) = κ t − p ( t ) ≥ κ 2 t ≥ 1. q(t)=\frac{\kappa}{t}-p(t)\ge\frac{\kappa}{2t}\ge1. q ( t ) = t κ − p ( t ) ≥ 2 t κ ≥ 1.
Moreover p ( t ) = τ ( t ) p(t)=\tau(t) p ( t ) = τ ( t ) is increasing on this interval, because it is the inverse of the increasing branch ξ 1 \xi_1 ξ 1 . Consequently the set where G ( s , ⋅ ) G(s,\cdot) G ( s , ⋅ ) decreases is either empty or an interval whose right endpoint is characterized by
At that endpoint, t ∈ Γ ( s ) t\in\Gamma(s) t ∈ Γ ( s ) . (Since q ≥ 1 ≥ s q\ge1\ge s q ≥ 1 ≥ s , with equality only at t = κ / 2 t=\kappa/2 t = κ /2 and s = 1 s=1 s = 1 , the decreasing set is exactly { t ∈ ( b , κ / 2 ) : p ( t ) < s } \{t\in(b,\kappa/2):p(t)<s\} { t ∈ ( b , κ /2 ) : p ( t ) < s } ; for s = 1 s=1 s = 1 it is the whole interval and its right end-point satisfies κ / 2 ∈ Γ ( 1 ) \kappa/2\in\Gamma(1) κ /2 ∈ Γ ( 1 ) .)
The interval κ / 2 < t < 1 \kappa/2<t<1 κ /2 < t < 1 . Set
L ( t ) = min { p ( t ) , q ( t ) } , U ( t ) = max { p ( t ) , q ( t ) } . L(t)=\min\{p(t),q(t)\},\qquad U(t)=\max\{p(t),q(t)\}. L ( t ) = min { p ( t ) , q ( t )} , U ( t ) = max { p ( t ) , q ( t )} .
We first record the shape of these two functions.
For κ / 2 < t < b ′ \kappa/2 < t < b' κ /2 < t < b ′ , one has p ( t ) = τ ( t ) p(t) = \tau(t) p ( t ) = τ ( t ) with t = ξ 2 ( p ( t ) ) t = \xi_2(p(t)) t = ξ 2 ( p ( t )) . Hence (A.18) gives t p ( t ) ≥ κ / 2 t\,p(t) \ge\kappa/2 t p ( t ) ≥ κ /2 , so that q ( t ) ≤ p ( t ) q(t) \le p(t) q ( t ) ≤ p ( t ) . Thus
U ( t ) = p ( t ) = τ ( t ) , L ( t ) = q ( t ) . U(t) = p(t) = \tau(t), \qquad L(t) = q(t). U ( t ) = p ( t ) = τ ( t ) , L ( t ) = q ( t ) .
Since ξ 2 \xi_2 ξ 2 is decreasing, U U U is decreasing. To see that L L L is also decreasing, write u = p ( t ) u = p(t) u = p ( t ) . Then
t = ξ 2 ( u ) = κ + u − u 2 2 u . t = \xi_2(u) = \frac{\kappa+ u - u^2}{2u}. t = ξ 2 ( u ) = 2 u κ + u − u 2 .
and
q = 2 κ u κ + u − u 2 − u . q = \frac{2\kappa u}{\kappa+ u - u^2} - u. q = κ + u − u 2 2 κ u − u .
A direct differentiation shows that
d q d u ≥ 0 ⟺ 2 κ ( κ + u 2 ) ≥ ( κ + u ( 1 − u ) ) 2 . \frac{dq}{du} \ge0 \quad\Longleftrightarrow\quad2\kappa(\kappa+ u^2) \ge\left(\kappa+ u(1-u)\right)^2. d u d q ≥ 0 ⟺ 2 κ ( κ + u 2 ) ≥ ( κ + u ( 1 − u ) ) 2 .
The difference between the two sides is
κ 2 − u 2 ( 1 − u ) 2 + 2 κ u ( 2 u − 1 ) , \kappa^2 - u^2(1-u)^2 + 2\kappa u(2u-1), κ 2 − u 2 ( 1 − u ) 2 + 2 κ u ( 2 u − 1 ) ,
which is non-negative by (A.16) and u ≥ c ≥ 1 / 2 u \ge c \ge1/2 u ≥ c ≥ 1/2 . Since u = τ ( t ) u = \tau(t) u = τ ( t ) decreases with t t t , L ( t ) = q ( t ) L(t) = q(t) L ( t ) = q ( t ) is decreasing.
For b ′ < t < c b' < t < c b ′ < t < c , Γ ( t ) = { p ( t ) } \Gamma(t) = \{p(t)\} Γ ( t ) = { p ( t )} with
p ( t ) = 1 + b − t . p(t) = 1 + b - t. p ( t ) = 1 + b − t .
The function t ↦ t ( 1 + b − t ) t \mapsto t(1+b-t) t ↦ t ( 1 + b − t ) is concave, and its values at both endpoints are b ′ c b'c b ′ c . Moreover
2 b ′ c − κ = c ( 1 − c ) ≥ 0. 2b'c - \kappa= c(1-c) \ge0. 2 b ′ c − κ = c ( 1 − c ) ≥ 0.
Hence t p ( t ) ≥ κ / 2 t\,p(t) \ge\kappa/2 t p ( t ) ≥ κ /2 , so again q ( t ) ≤ p ( t ) q(t) \le p(t) q ( t ) ≤ p ( t ) . Therefore
U ( t ) = 1 + b − t U(t) = 1 + b - t U ( t ) = 1 + b − t
is decreasing, while
L ( t ) = κ t − ( 1 + b − t ) L(t) = \frac{\kappa}{t} - (1+b-t) L ( t ) = t κ − ( 1 + b − t )
satisfies
L ′ ′ ( t ) = 2 κ t 3 > 0 , L ( c ) = b , L ′ ( c ) = 1 − κ c 2 ≥ 0 L''(t) = \frac{2\kappa}{t^3} > 0, \qquad L(c) = b, \qquad L'(c) = 1 - \frac{\kappa}{c^2} \ge0 L ′′ ( t ) = t 3 2 κ > 0 , L ( c ) = b , L ′ ( c ) = 1 − c 2 κ ≥ 0
by (A.15). Thus on this interval L L L can change monotonicity at most once, from decreasing to increasing.
Finally, for c < t < 1 c < t < 1 c < t < 1 ,
L ( t ) = ξ 1 ( t ) , U ( t ) = ξ 2 ( t ) , L(t) = \xi_1(t), \qquad U(t) = \xi_2(t), L ( t ) = ξ 1 ( t ) , U ( t ) = ξ 2 ( t ) ,
so L L L is increasing and U U U is decreasing.
The pieces match continuously at κ / 2 \kappa/2 κ /2 , b ′ b' b ′ , and c c c . We have therefore proved that U U U is continuous and decreasing on [ κ / 2 , 1 ] [\kappa/2,1] [ κ /2 , 1 ] , whereas L L L is continuous and first decreasing and then increasing. In addition,
L ( c ) = b , L ( t ) = ξ 1 ( t ) ( c ≤ t ≤ 1 ) , L(c)=b,\qquad L(t)=\xi_1(t)\quad(c\le t\le1), L ( c ) = b , L ( t ) = ξ 1 ( t ) ( c ≤ t ≤ 1 ) ,
and U ( t ) ∈ Γ ( t ) U(t)\in\Gamma(t) U ( t ) ∈ Γ ( t ) throughout this interval.
For fixed s ≥ b s\ge b s ≥ b , the decreasing set of G ( s , ⋅ ) G(s,\mathord\cdot) G ( s , ⋅ ) in ( κ / 2 , 1 ) (\kappa/2,1) ( κ /2 , 1 ) is
J s = { t : L ( t ) < s < U ( t ) } . J_s=\{t:L(t)<s<U(t)\}. J s = { t : L ( t ) < s < U ( t )} .
Since L L L is quasi-convex and U U U is decreasing, J s J_s J s is an interval (possibly empty). Its right endpoint can occur only when
U ( t ) = s or L ( t ) = s U(t)=s\qquad\text{or}\qquad L(t)=s U ( t ) = s or L ( t ) = s
on the increasing branch of L L L .
If U ( t ) = s U(t)=s U ( t ) = s , then t ∈ Γ ( s ) t\in\Gamma(s) t ∈ Γ ( s ) , because U ( t ) ∈ Γ ( t ) U(t)\in\Gamma(t) U ( t ) ∈ Γ ( t ) and D b , c D_{b,c} D b , c is symmetric. If L ( t ) = s L(t)=s L ( t ) = s , note that the increasing part of L L L before c c c takes values at most L ( c ) = b ≤ s L(c)=b\le s L ( c ) = b ≤ s . Hence such a right endpoint is either t = c t=c t = c with s = b s=b s = b , or lies in [ c , 1 ] [c,1] [ c , 1 ] , where L = ξ 1 L=\xi_1 L = ξ 1 . In both cases ( s , t ) (s,t) ( s , t ) belongs to the support.
It remains to check the endpoints and junctions. If b > 0 b>0 b > 0 , the function decreases immediately to the right of 0 0 0 , so 0 0 0 is not a local minimum; if b = 0 b=0 b = 0 , this endpoint is already allowed. At t = 1 t=1 t = 1 the left derivative is
∂ t G ( s , 1 − ) = 1 κ ( s − κ / 2 ) 2 . \partial_tG(s,1^-)=\frac{1}{\kappa}(s-\kappa/2)^2. ∂ t G ( s , 1 − ) = κ 1 ( s − κ /2 ) 2 .
Unless s = κ / 2 s=\kappa/2 s = κ /2 , it is strictly positive near 1 1 1 , excluding a one-sided minimum there. When s = κ / 2 s=\kappa/2 s = κ /2 , 1 ∈ Γ ( s ) 1\in\Gamma(s) 1 ∈ Γ ( s ) .
At an interior junction t = κ / 2 > b t=\kappa/2>b t = κ /2 > b , both roots equal 1 1 1 . If s < 1 s<1 s < 1 , the derivative is positive on both sides sufficiently close to the junction; if s = 1 s=1 s = 1 , the junction itself belongs to Γ ( 1 ) \Gamma(1) Γ ( 1 ) . At b ′ b' b ′ and c c c , continuity of the roots shows that a sign change from negative to positive is one of the right endpoints already classified. Coincident junctions are covered by the same one-sided argument. The possible junction t = b t=b t = b is allowed separately.
On each remaining open piece the derivative has only isolated zeros: the branches of the support are strictly monotone, and the complementary root is monotone or strictly convex on the pieces considered above. Thus there are no constant intervals that could produce additional local minima. Every interior local minimum must end an interval of decrease, and we have shown that each such endpoint is either b b b or a point of Γ ( s ) \Gamma(s) Γ ( s ) .
This lemma replaces the full two-dimensional sign analysis by a single principle: for s ≥ b s\ge b s ≥ b , the slack can attain a new minimum only on the support or at the junction t = b t=b t = b .
Optimality of the candidate potential
Theorem A.9. Let ( b , c ) ∈ Δ (b,c)\in\Delta ( b , c ) ∈ Δ . Assume that either
b > 0 and P ( b , c ) = 0 , b>0\quad\text{and}\quad P(b,c)=0, b > 0 and P ( b , c ) = 0 ,
or
b = 0 b = 0 b = 0 and P ( 0 , c ) ≤ 0 P(0,c) \le0 P ( 0 , c ) ≤ 0 .
Then ϕ b , c \phi_{b,c} ϕ b , c is a Kantorovich potential for D b , c D_{b,c} D b , c . Hence
r ∗ ( 1 12 + 1 4 E D b , c [ S 2 T 2 ] ) = 3 E D b , c [ S T ] . (59) r^{*}\left(\frac{1}{12}+\frac{1}{4}\mathbb{E}_{D_{b,c}}[S^{2}T^{2}]\right)=3\mathbb{E}_{D_{b,c}}[ST]. \tag*{(59)} r ∗ ( 12 1 + 4 1 E D b , c [ S 2 T 2 ] ) = 3 E D b , c [ S T ] . (59)
Proof. By Lemma Appendix § 11 .7, when b > 0 b>0 b > 0 the condition P ( b , c ) = 0 P(b,c)=0 P ( b , c ) = 0 makes ϕ \phi ϕ continuous at b b b and gives
G ( b , b ) = 0. G(b,b)=0. G ( b , b ) = 0.
When b = 0 b=0 b = 0 , the condition P ( 0 , c ) ≤ 0 P(0,c)\le0 P ( 0 , c ) ≤ 0 gives
G ( 0 , 0 ) = 2 ϕ ( 0 ) ≥ 0. G(0,0)=2\phi(0)\ge0. G ( 0 , 0 ) = 2 ϕ ( 0 ) ≥ 0.
Under the assumptions of the theorem, ϕ \phi ϕ is continuous on [ 0 , 1 ] [0,1] [ 0 , 1 ] (at the endpoint 0 0 0 when b = 0 b=0 b = 0 , no matching condition is needed). Hence, for fixed s s s , the function G ( s , ⋅ ) G(s,\cdot) G ( s , ⋅ ) is continuous on [ 0 , 1 ] [0,1] [ 0 , 1 ] and piecewise continuously differentiable.
We first prove
G ( s , t ) ≥ 0 for s ∈ [ b , 1 ] . G(s,t)\ge0 \qquad\text{for }s\in[b,1]. G ( s , t ) ≥ 0 for s ∈ [ b , 1 ] .
Take first s = b s=b s = b . By Lemma Appendix § 11 .8, every local minimum of G ( b , ⋅ ) G(b,\cdot) G ( b , ⋅ ) occurs either at t = b t=b t = b or at a point of Γ ( b ) \Gamma(b) Γ ( b ) . If b > 0 b>0 b > 0 , then G ( b , b ) = 0 G(b,b)=0 G ( b , b ) = 0 because P ( b , c ) = 0 P(b,c)=0 P ( b , c ) = 0 . If b = 0 b=0 b = 0 , we already know that G ( 0 , 0 ) ≥ 0 G(0,0)\ge0 G ( 0 , 0 ) ≥ 0 . The other relevant point of Γ ( b ) \Gamma(b) Γ ( b ) is t = c t=c t = c (the point ( b , c ) = ( ξ 1 ( c ) , c ) (b,c)=(\xi_{1}(c),c) ( b , c ) = ( ξ 1 ( c ) , c ) lies on the curved branch), and (A.22) gives directly
G ( b , c ) = 0. G(b,c)=0. G ( b , c ) = 0.
Thus every local minimum of G ( b , ⋅ ) G(b,\cdot) G ( b , ⋅ ) is non-negative. Since the function is continuous on the compact interval [ 0 , 1 ] [0,1] [ 0 , 1 ] ,
G ( b , t ) ≥ 0 0 ≤ t ≤ 1. G(b,t)\ge0 \qquad0\le t\le1. G ( b , t ) ≥ 0 0 ≤ t ≤ 1.
By symmetry,
G ( s , b ) ≥ 0 0 ≤ s ≤ 1. G(s,b)\ge0 \qquad0\le s\le1. G ( s , b ) ≥ 0 0 ≤ s ≤ 1.
Now fix any s ∈ [ b , 1 ] s\in[b,1] s ∈ [ b , 1 ] . Again by Lemma Appendix § 11 .8, every local minimum of G ( s , ⋅ ) G(s,\cdot) G ( s , ⋅ ) occurs either at a point of Γ ( s ) \Gamma(s) Γ ( s ) , where G = 0 G=0 G = 0 , or at t = b t=b t = b , where G ( s , b ) ≥ 0 G(s,b)\ge0 G ( s , b ) ≥ 0 . Therefore
G ( s , t ) ≥ 0 b ≤ s ≤ 1 , 0 ≤ t ≤ 1. G(s,t)\ge0 \qquad b\le s\le1,\quad0\le t\le1. G ( s , t ) ≥ 0 b ≤ s ≤ 1 , 0 ≤ t ≤ 1.
It remains to consider 0 ≤ s < b 0\le s<b 0 ≤ s < b , which occurs only when b > 0 b>0 b > 0 . For t < b t<b t < b ,
G ( s , t ) = ( s − t ) 2 2 ( 1 − ( s + t ) 2 2 κ ) ≥ 0 , G(s,t)=\frac{(s-t)^{2}}{2}\left(1-\frac{(s+t)^{2}}{2\kappa}\right)\ge0, G ( s , t ) = 2 ( s − t ) 2 ( 1 − 2 κ ( s + t ) 2 ) ≥ 0 ,
because 4 b 2 ≤ 2 κ 4b^{2}\le2\kappa 4 b 2 ≤ 2 κ by (A.13) (indeed κ ≥ b ( 1 + b ) ≥ 2 b 2 \kappa\ge b(1+b)\ge2b^{2} κ ≥ b ( 1 + b ) ≥ 2 b 2 ).
Now fix t ≥ b t\ge b t ≥ b and let
F ( s ) = G ( s , t ) , 0 ≤ s ≤ b . F(s)=G(s,t),\qquad0\le s\le b. F ( s ) = G ( s , t ) , 0 ≤ s ≤ b .
Then
F ′ ′ ( s ) = 1 − 3 s 2 κ + t 2 κ ≥ 1 − 2 b 2 κ ≥ 0 , F''(s)=1-\frac{3s^2}{\kappa}+\frac{t^2}{\kappa}\ge1-\frac{2b^2}{\kappa}\ge0, F ′′ ( s ) = 1 − κ 3 s 2 + κ t 2 ≥ 1 − κ 2 b 2 ≥ 0 ,
while
F ′ ( b − ) = ( b − t ) ( 1 − b ( b + t ) κ ) ≤ 0 F'(b^-)= (b-t)\left(1-\frac{b(b+t)}{\kappa}\right)\le0 F ′ ( b − ) = ( b − t ) ( 1 − κ b ( b + t ) ) ≤ 0
by (A.13). Since F F F is convex, F ′ F' F ′ is non-decreasing, so F ′ ( s ) ≤ F ′ ( b − ) ≤ 0 F'(s)\le F'(b^-)\le0 F ′ ( s ) ≤ F ′ ( b − ) ≤ 0 on [ 0 , b ] [0,b] [ 0 , b ] . Hence F F F is non-increasing on [ 0 , b ] [0,b] [ 0 , b ] , and
G ( s , t ) = F ( s ) ≥ F ( b ) = G ( b , t ) ≥ 0. G(s,t)=F(s)\ge F(b)=G(b,t)\ge0. G ( s , t ) = F ( s ) ≥ F ( b ) = G ( b , t ) ≥ 0.
Thus G ≥ 0 G\ge0 G ≥ 0 on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 and G = 0 G=0 G = 0 on supp ( D b , c ) \operatorname{supp}(D_{b,c}) supp ( D b , c ) . For b > 0 b>0 b > 0 the diagonal junction ( b , b ) (b,b) ( b , b ) satisfies equality by the matching condition; for b = 0 b=0 b = 0 the point ( 0 , 0 ) (0,0) ( 0 , 0 ) is absent from the support. Therefore ϕ b , c \phi_{b,c} ϕ b , c is a Kantorovich potential, and (A.28) follows from Lemma Appendix § 11 .§ 1 .
Selection of the boundary parameters
It remains to determine when the assumptions of Theorem A.9 hold.
Lemma Appendix A.10. For
0 ≤ b ≤ 1 5 , 1 2 + b ≤ c ≤ 1 , 0\le b\le\frac{1}{5},\qquad\frac{1}{2}+b\le c\le1, 0 ≤ b ≤ 5 1 , 2 1 + b ≤ c ≤ 1 , the following statements hold.
(i) In the interior of Δ \Delta Δ ,
∂ P ∂ c ( b , c ) = 6 ( 2 c − 1 − 2 b ) ( 2 ( 1 − c ) 2 − ( 1 − b ) 2 ) < 0. (60) \frac{\partial P}{\partial c}(b,c)=6(2c-1-2b)\left(2(1-c)^2-(1-b)^2\right)<0. \tag*{(60)} ∂ c ∂ P ( b , c ) = 6 ( 2 c − 1 − 2 b ) ( 2 ( 1 − c ) 2 − ( 1 − b ) 2 ) < 0. (60) (ii)
P ( b , 1 2 + b ) = 1 8 ( 3 − 12 b 2 + 16 b 3 − 8 b 4 ) > 0 , (61) P\left(b,\frac{1}{2}+b\right)=\frac{1}{8}\left(3-12b^2+16b^3-8b^4\right)>0, \tag*{(61)} P ( b , 2 1 + b ) = 8 1 ( 3 − 12 b 2 + 16 b 3 − 8 b 4 ) > 0 , (61) whereas
P ( b , 1 ) = ( 1 − b ) 3 ( 5 b − 1 ) . (62) P(b,1)=(1-b)^3(5b-1). \tag*{(62)} P ( b , 1 ) = ( 1 − b ) 3 ( 5 b − 1 ) . (62) (iii) For every b ∈ [ 0 , 1 / 5 ] b\in[0,1/5] b ∈ [ 0 , 1/5 ] there is a unique
c ( b ) ∈ ( 1 2 + b , 1 ] c(b)\in\left(\frac{1}{2}+b,1\right] c ( b ) ∈ ( 2 1 + b , 1 ] such that P ( b , c ( b ) ) = 0 P(b,c(b))=0 P ( b , c ( b )) = 0 . Moreover,
c ( 0 ) = c ∗ , c ( 1 5 ) = 1 , c(0)=c^*,\qquad c\left(\frac{1}{5}\right)=1, c ( 0 ) = c ∗ , c ( 5 1 ) = 1 , and
P ( 0 , c ) ≤ 0 ⟺ c ∗ ≤ c ≤ 1. P(0,c)\le0\quad\Longleftrightarrow\quad c^*\le c\le1. P ( 0 , c ) ≤ 0 ⟺ c ∗ ≤ c ≤ 1.
Proof. The identities in (i) and (ii) follow by direct factorization. In the interior of Δ \Delta Δ ,
2 c − 1 − 2 b > 0 , 2c-1-2b>0, 2 c − 1 − 2 b > 0 ,
while
2 ( 1 − c ) 2 ≤ 1 2 ( 1 − 2 b ) 2 < ( 1 − b ) 2 . 2(1-c)^2\leq\frac{1}{2}(1-2b)^2<(1-b)^2. 2 ( 1 − c ) 2 ≤ 2 1 ( 1 − 2 b ) 2 < ( 1 − b ) 2 .
Hence ∂ c P < 0 \partial_cP<0 ∂ c P < 0 .
For b ≤ 1 5 b\leq\frac{1}{5} b ≤ 5 1 we have 16 b 3 − 8 b 4 = 8 b 3 ( 2 − b ) ≥ 0 16b^3-8b^4=8b^3(2-b)\geq0 16 b 3 − 8 b 4 = 8 b 3 ( 2 − b ) ≥ 0 and 12 b 2 ≤ 12 25 < 3 12b^2\leq\frac{12}{25}<3 12 b 2 ≤ 25 12 < 3 , so P ( b , 1 2 + b ) > 0 P(b,\frac{1}{2}+b)>0 P ( b , 2 1 + b ) > 0 , while P ( b , 1 ) ≤ 0 P(b,1)\leq0 P ( b , 1 ) ≤ 0 with equality only for b = 1 5 b=\frac{1}{5} b = 5 1 . The endpoint values in (ii), together with the strict monotonicity in c c c , give existence and uniqueness of the zero. The formula for ∂ c P \partial_cP ∂ c P is strictly negative along the entire zero curve, including b = 0 b=0 b = 0 and b = 1 / 5 b=1/5 b = 1/5 . The implicit function theorem therefore makes b ↦ c ( b ) b\mapsto c(b) b ↦ c ( b ) continuous and real analytic locally, with analytic extensions across its two endpoints. The remaining statements follow by evaluating the endpoint cases b = 0 b=0 b = 0 and b = 1 / 5 b=1/5 b = 1/5 .
Corollary Appendix A.11. For every c ∈ [ c ∗ , 1 ] c\in[c^*,1] c ∈ [ c ∗ , 1 ] ,
r ∗ ( ARV ( C c I I I ) ) = ρ ( C c I I I ) , r^*\left(\operatorname{ARV}(C_c^{\mathrm{III}})\right)=\rho(C_c^{\mathrm{III}}), r ∗ ( ARV ( C c III ) ) = ρ ( C c III ) , and for every b ∈ [ 0 , 1 / 5 ] b\in[0,1/5] b ∈ [ 0 , 1/5 ] ,
r ∗ ( ARV ( C b , c ( b ) I I ) ) = ρ ( C b , c ( b ) I I ) , r^*\left(\operatorname{ARV}(C_{b,c(b)}^{\mathrm{II}})\right)=\rho(C_{b,c(b)}^{\mathrm{II}}), r ∗ ( ARV ( C b , c ( b ) II ) ) = ρ ( C b , c ( b ) II ) , where at b = 0 b=0 b = 0 we use
C 0 , c ∗ I I = C c ∗ I I I . C_{0,c^*}^{\mathrm{II}}=C_{c^*}^{\mathrm{III}}. C 0 , c ∗ II = C c ∗ III . Consequently, Γ I I I \Gamma^{\mathrm{III}} Γ III and Γ I I \Gamma^{\mathrm{II}} Γ II are upper-boundary arcs.
Proof. If b = 0 b=0 b = 0 and c ∈ [ c ∗ , 1 ) c\in[c^*,1) c ∈ [ c ∗ , 1 ) , then P ( 0 , c ) ≤ 0 P(0,c)\leq0 P ( 0 , c ) ≤ 0 , so Theorem A.9 applies to D 0 , c D_{0,c} D 0 , c . The endpoint c = 1 c=1 c = 1 is C 1 I I I = C 2 C_1^{\mathrm{III}}=C_2 C 1 III = C 2 . By Proposition 2.4 , r ∗ ( 11 / 120 ) = 1 / 2 r^*(11/120)=1/2 r ∗ ( 11/120 ) = 1/2 , and C 2 C_2 C 2 attains this value. Hence this point belongs to the upper boundary, as required.
If 0 < b ≤ 1 / 5 0<b\leq1/5 0 < b ≤ 1/5 , then P ( b , c ( b ) ) = 0 P(b,c(b))=0 P ( b , c ( b )) = 0 , and the same theorem applies to D b , c ( b ) D_{b,c(b)} D b , c ( b ) . The case b = 0 b=0 b = 0 is the common endpoint of the two arcs.
Completion of the proof of the main theorem
Proof of Theorem 4.1 . Theorem A.2 shows that Γ I \Gamma^{\mathrm{I}} Γ I lies on the upper boundary, while Corollary Appendix A.11 gives the same conclusion for Γ I I \Gamma^{\mathrm{II}} Γ II and Γ I I I \Gamma^{\mathrm{III}} Γ III . The three arcs join continuously because
C 0 , c ∗ I I = C c ∗ I I I , C 1 / 5 , 1 I I = C 2 / 5 I . C_{0,c^*}^{\mathrm{II}}=C_{c^*}^{\mathrm{III}},\qquad C_{1/5,1}^{\mathrm{II}}=C_{2/5}^{\mathrm{I}}. C 0 , c ∗ II = C c ∗ III , C 1/5 , 1 II = C 2/5 I .
Moreover,
C 1 I I I = C 2 , C 0 I = M . C_1^{\mathrm{III}}=C_2,\qquad C_0^{\mathrm{I}}=M. C 1 III = C 2 , C 0 I = M .
Hence their union is a continuous subset of the upper boundary joining
( 11 120 , 1 2 ) to ( 2 15 , 1 ) . \left(\frac{11}{120},\frac{1}{2}\right)\ \text{to}\ \left(\frac{2}{15},1\right). ( 120 11 , 2 1 ) to ( 15 2 , 1 ) .
Its first coordinate therefore takes every value in [ 11 / 120 , 2 / 15 ] [11/120, 2/15] [ 11/120 , 2/15 ] . Since the upper boundary is the graph of the function r ∗ r^* r ∗ defined in (5) , we conclude that
graph ( r ∗ ) = Γ I I I ∪ Γ I I ∪ Γ I . \operatorname{graph}(r^*) = \Gamma^{\mathrm{III}} \cup\Gamma^{\mathrm{II}} \cup\Gamma^{\mathrm{I}}. graph ( r ∗ ) = Γ III ∪ Γ II ∪ Γ I .
Proposition 2.4 supplies the description of the full region, its symmetry, and the continuity, concavity and strict monotonicity of r ∗ r^* r ∗ . The first and third arcs have analytic parametrizations by their displayed moment formulas, while the analyticity of the middle parametrization follows from Lemma Appendix A.10. This completes the proof.
Regularity of the boundary and behavior near the minimum
We now prove corollaries 4.3 and 4.4. The optimality proofs provide, at every point of the three arcs, a supporting line with explicit slope, which yields the regularity of the boundary.
Proof of corollary 4.3. Step 1: supporting lines. Let D D D be the folded copula of a member of one of the three families, with associated κ > 0 \kappa> 0 κ > 0 (that is, κ = κ b \kappa= \kappa_b κ = κ b for Family I, and κ = c ( 1 + 2 b − c ) \kappa= c(1 + 2b - c) κ = c ( 1 + 2 b − c ) for Families II and III, with c < 1 c < 1 c < 1 when b = 0 b = 0 b = 0 ). By Theorem A.2 and Theorem A.9, D D D maximizes E D ′ [ c κ ( S , T ) ] = E D ′ [ S T ] − E D ′ [ S 2 T 2 ] / ( 2 κ ) \mathbb{E}_{D'}[c_\kappa(S,T)] = \mathbb{E}_{D'}[ST] - \mathbb{E}_{D'}[S^2T^2]/(2\kappa) E D ′ [ c κ ( S , T )] = E D ′ [ S T ] − E D ′ [ S 2 T 2 ] / ( 2 κ ) over D ′ ∈ C D' \in\mathcal{C} D ′ ∈ C . By (3) , (4) and ρ ( D U ) = 3 E D [ S T ] \rho(D^{\mathrm{U}}) = 3\mathbb{E}_{D}[ST] ρ ( D U ) = 3 E D [ S T ] , for every copula C C C ,
ρ ( C ) − 6 κ ARV ( C ) ≤ 3 E C F [ c κ ( S , T ) ] − 1 2 κ ≤ 3 E D [ c κ ( S , T ) ] − 1 2 κ = ρ ( D U ) − 6 κ ARV ( D U ) . \rho(C) - \frac{6}{\kappa}\operatorname{ARV}(C) \le3\mathbb{E}_{C^{\mathrm{F}}}[c_\kappa(S,T)] - \frac{1}{2\kappa} \le3\mathbb{E}_{D}[c_\kappa(S,T)] - \frac{1}{2\kappa} = \rho(D^{\mathrm{U}}) - \frac{6}{\kappa}\operatorname{ARV}(D^{\mathrm{U}}). ρ ( C ) − κ 6 ARV ( C ) ≤ 3 E C F [ c κ ( S , T )] − 2 κ 1 ≤ 3 E D [ c κ ( S , T )] − 2 κ 1 = ρ ( D U ) − κ 6 ARV ( D U ) .
Hence the line of slope 6 / κ 6/\kappa 6/ κ through the corresponding boundary point supports A \mathcal{A} A .
Step 2: slopes. For a = 0 a = 0 a = 0 (D = M D = M D = M , κ = 2 \kappa= 2 κ = 2 ), Step 1 gives r ∗ ( α ) ≤ 1 + 3 ( α − 2 / 15 ) r^*(\alpha) \le1 + 3(\alpha- 2/15) r ∗ ( α ) ≤ 1 + 3 ( α − 2/15 ) ; hence the left derivative of r ∗ r^* r ∗ at 2 / 15 2/15 2/15 is at least 3 3 3 and, by concavity, every supergradient of r ∗ r^* r ∗ at an interior point is at least 3 3 3 .
Step 3: no corners. Suppose that, for some α 0 ∈ ( 11 / 120 , 2 / 15 ) \alpha_0 \in(11/120, 2/15) α 0 ∈ ( 11/120 , 2/15 ) , the one-sided derivatives satisfy ( r ∗ ) + ( α 0 ) < ( r ∗ ) − ( α 0 ) (r^*)_+(\alpha_0) < (r^*)_-(\alpha_0) ( r ∗ ) + ( α 0 ) < ( r ∗ ) − ( α 0 ) . Choose a nonempty open interval J J J of slopes strictly between them. Shrinking J J J if necessary, we may assume that K = { 6 / m : m ∈ J } K = \{6/m : m \in J\} K = { 6/ m : m ∈ J } is contained in one of the intervals ( 0 , κ ∗ ) (0,\kappa^*) ( 0 , κ ∗ ) , ( κ ∗ , 2 / 5 ) (\kappa^*,2/5) ( κ ∗ , 2/5 ) , ( 2 / 5 , 2 ) (2/5,2) ( 2/5 , 2 ) , where κ ∗ = c ∗ ( 1 − c ∗ ) \kappa^* = c^*(1-c^*) κ ∗ = c ∗ ( 1 − c ∗ ) . For m ∈ J m \in J m ∈ J , concavity shows that the line of slope m m m through ( α 0 , r ∗ ( α 0 ) ) (\alpha_0,r^*(\alpha_0)) ( α 0 , r ∗ ( α 0 )) meets A \mathcal{A} A only at this point. Along the parameter intervals c ∈ ( c ∗ , 1 ) c \in(c^*,1) c ∈ ( c ∗ , 1 ) (Family III), b ∈ ( 0 , 1 / 5 ) b \in(0,1/5) b ∈ ( 0 , 1/5 ) (Family II, with c = c ( b ) c = c(b) c = c ( b ) ) and a ∈ ( 0 , 2 / 5 ) a \in(0,2/5) a ∈ ( 0 , 2/5 ) (Family I), κ \kappa κ depends continuously on the parameter and tends to the endpoint values 0 , κ ∗ 0,\kappa^* 0 , κ ∗ , respectively κ ∗ , 2 / 5 \kappa^*,2/5 κ ∗ , 2/5 , respectively 2 / 5 , 2 2/5,2 2/5 , 2 . By the intermediate value theorem, the parameters of the corresponding family with κ ∈ K \kappa\in K κ ∈ K form a nonempty open set, which contains an open interval I I I . By Step 1 every parameter in I I I generates the point ( α 0 , r ∗ ( α 0 ) ) (\alpha_0,r^*(\alpha_0)) ( α 0 , r ∗ ( α 0 )) , so ARV is constant on I I I . This is impossible: along each family, ARV is a real-analytic function of the parameter (by (6) , (11) , (14) and Lemma Appendix A.10) that takes different values at the two endpoints of the parameter interval, hence it is not constant on any open subinterval.
Step 4: conclusion. By Step 3, r ∗ r^* r ∗ is differentiable on ( 11 / 120 , 2 / 15 ) (11/120,2/15) ( 11/120 , 2/15 ) ; a concave function that is differentiable on an open interval is continuously differentiable there. At a point generated by a family member with parameter κ \kappa κ , the supporting line of Step 1 must then be the tangent line, which gives (24) ; for Family I, κ 1 − 2 a = 2 3 ( 1 − a ) ( 3 − 5 a ) \kappa_{1-2a} = \frac{2}{3}(1-a)(3-5a) κ 1 − 2 a = 3 2 ( 1 − a ) ( 3 − 5 a ) . As a ↓ 0 a \downarrow0 a ↓ 0 the derivative tends to 3 3 3 , which is therefore the left derivative of the concave function r ∗ r^* r ∗ at 2 / 15 2/15 2/15 . The values at the junctions follow from κ 1 / 5 = 2 5 \kappa_{1/5} = \frac{2}{5} κ 1/5 = 5 2 and c ( 0 ) = c ∗ c(0) = c^{*} c ( 0 ) = c ∗ . Finally, as c ↑ 1 c \uparrow1 c ↑ 1 , ARV ( C c I I I ) ↓ 11 / 120 \operatorname{ARV}(C^{\mathrm{III}}_c) \downarrow11/120 ARV ( C c III ) ↓ 11/120 and 6 / ( c ( 1 − c ) ) → + ∞ 6/(c(1-c)) \to+\infty 6/ ( c ( 1 − c )) → + ∞ .
Proof of corollary 4.4. Near the left endpoint the upper boundary is generated by C c I I I C^{\mathrm{III}}_c C c III : by the expansion below, c ↦ ARV ( C c I I I ) c \mapsto\operatorname{ARV}(C^{\mathrm{III}}_c) c ↦ ARV ( C c III ) is continuous and strictly decreasing near c = 1 c=1 c = 1 , with value 11 120 \frac{11}{120} 120 11 at c = 1 c=1 c = 1 , so every α \alpha α close to 11 120 \frac{11}{120} 120 11 equals ARV ( C c I I I ) \operatorname{ARV}(C^{\mathrm{III}}_c) ARV ( C c III ) for a unique c c c close to 1 1 1 , and then r ∗ ( α ) = ρ ( C c I I I ) r^{*}(\alpha)=\rho(C^{\mathrm{III}}_c) r ∗ ( α ) = ρ ( C c III ) . Set c = 1 − δ c=1-\delta c = 1 − δ , with δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Expanding (15) and (14) gives
ρ ( C 1 − δ I I I ) = 1 2 + δ 3 2 − δ 4 2 + O ( δ 5 ) . \rho(C^{\mathrm{III}}_{1-\delta})=\frac{1}{2}+\frac{\delta^{3}}{2}-\frac{\delta^{4}}{2}+O(\delta^{5}). ρ ( C 1 − δ III ) = 2 1 + 2 δ 3 − 2 δ 4 + O ( δ 5 ) .
ARV ( C 1 − δ I I I ) = 11 120 + δ 4 16 − 7 δ 5 60 + O ( δ 6 ) . \operatorname{ARV}(C^{\mathrm{III}}_{1-\delta})=\frac{11}{120}+\frac{\delta^{4}}{16}-\frac{7\delta^{5}}{60}+O(\delta^{6}). ARV ( C 1 − δ III ) = 120 11 + 16 δ 4 − 60 7 δ 5 + O ( δ 6 ) .
Hence
δ = 2 ( α − 11 120 ) 1 / 4 + o ( ( α − 11 120 ) 1 / 4 ) , \delta=2\left(\alpha-\frac{11}{120}\right)^{1/4}+o\left(\left(\alpha-\frac{11}{120}\right)^{1/4}\right), δ = 2 ( α − 120 11 ) 1/4 + o ( ( α − 120 11 ) 1/4 ) ,
and substitution into the expansion of ρ \rho ρ yields the stated formula.