bernoulli and tail-dependence compatibilitysas.uwaterloo.ca/~wang/papers/2015ehw(aap).pdf ·...
TRANSCRIPT
0001
0002
0003
0004
0005
0006
0007
0008
0009
0010
0011
0012
0013
0014
0015
0016
0017
0018
0019
0020
0021
0022
0023
0024
0025
0026
0027
0028
0029
0030
0031
0032
0033
0034
0035
0036
0037
0038
0039
0040
0041
0042
0043
0044
0045
0046
0047
0048
0049
0050
0051
0052
0053
0054
0055
0056
0057
0058
0059
0060
0061
0062
0063
0064
Bernoulli and Tail-Dependence Compatibility
Paul Embrechts∗, Marius Hofert†and Ruodu Wang‡
July 30, 2015
Abstract
The tail-dependence compatibility problem is introduced. It raises the question whether
a given d×d-matrix of entries in the unit interval is the matrix of pairwise tail-dependence co-
efficients of a d-dimensional random vector. The problem is studied together with Bernoulli-
compatible matrices, i.e., matrices which are expectations of outer products of random vec-
tors with Bernoulli margins. We show that a square matrix with diagonal entries being 1 is
a tail-dependence matrix if and only if it is a Bernoulli-compatible matrix multiplied by a
constant. We introduce new copula models to construct tail-dependence matrices, including
commonly used matrices in statistics.
Key-words: Tail dependence, Bernoulli random vectors, compatibility, matrices, copulas, in-
surance application.
MSC2010: 60E05, 62H99, 62H20, 62E15, 62H86.
1 Introduction
The problem of how to construct a bivariate random vector (X1, X2) with log-normal
marginals X1 ∼ LN(0, 1), X2 ∼ LN(0, 16) and correlation coefficient Cor(X1, X2) = 0.5 is
well known in the history of dependence modeling, partially because of its relevance to risk
management practice. The short answer is: There is no such model; see Embrechts et al. (2002)
who studied these kinds of problems in terms of copulas. Problems of this kind were brought
to RiskLab at ETH Zurich by the insurance industry in the mid 1990s when dependence was
∗RiskLab and SFI, Department of Mathematics, ETH Zurich, 8092 Zurich, Switzerland.
Email: [email protected]†Department of Statistics and Actuarial Science, University of Waterloo, Canada.
Email: [email protected]‡Department of Statistics and Actuarial Science, University of Waterloo, Canada.
Email: [email protected]
1
0065
0066
0067
0068
0069
0070
0071
0072
0073
0074
0075
0076
0077
0078
0079
0080
0081
0082
0083
0084
0085
0086
0087
0088
0089
0090
0091
0092
0093
0094
0095
0096
0097
0098
0099
0100
0101
0102
0103
0104
0105
0106
0107
0108
0109
0110
0111
0112
0113
0114
0115
0116
0117
0118
0119
0120
0121
0122
0123
0124
0125
0126
0127
0128
thought of in terms of correlation (matrices). For further background to Quantitative Risk Man-
agement, see McNeil et al. (2015). Now, almost 20 years later, copulas are a well established
tool to quantify dependence in multivariate data and to construct new multivariate distribu-
tions. Their use has become standard within industry and regulation. Nevertheless, dependence
is still summarized in terms of numbers (as opposed to (copula) functions), so-called measures
of association. Although there are various ways to compute such numbers in dimension d > 2,
measures of association are still most widely used in the bivariate case d = 2. A popular measure
of association is tail dependence. It is important for applications in Quantitative Risk Manage-
ment as it measures the strength of dependence in either the lower-left or upper-right tail of the
bivariate distribution, the regions Quantitative Risk Management is mainly concerned with.
We were recently asked1 the following question which is in the same spirit as the log-
normal correlation problem if one replaces “correlation” by “tail dependence”; see Section 3.1
for a definition.
For which α ∈ [0, 1] is the matrix
Γd(α) =
1 0 · · · 0 α
0 1 · · · 0 α...
.... . .
......
0 0 · · · 1 α
α α · · · α 1
(1.1)
a matrix of pairwise (either lower or upper) tail-dependence coefficients?
Intrigued by this question, we more generally consider the following tail-dependence compatibility
problem in this paper:
When is a given matrix in [0, 1]d×d the matrix of pairwise (either lower or upper)
tail-dependence coefficients?
In what follows, we call a matrix of pairwise tail-dependence coefficients a tail-dependence
matrix. The compatibility problems of tail-dependence coefficients were studied in Joe (1997).
In particular, when d = 3, inequalities for the bivariate tail-dependence coefficients have been
established; see Joe (1997, Theorem 3.14) as well as Joe (2014, Theorem 8.20). The sharpness of
these inequalities is obtained in Nikoloulopoulos et al. (2009). It is generally open to characterize
the tail-dependence matrix compatibility for d > 3.
Our aim in this paper is to give a full answer to the tail-dependence compatibility problem;
see Section 3. To this end, we introduce and study Bernoulli-compatible matrices in Section 2.
1by Federico Degen (Head Risk Modeling and Quantification, Zurich Insurance Group) and Janusz Milek
(Zurich Insurance Group)
2
0129
0130
0131
0132
0133
0134
0135
0136
0137
0138
0139
0140
0141
0142
0143
0144
0145
0146
0147
0148
0149
0150
0151
0152
0153
0154
0155
0156
0157
0158
0159
0160
0161
0162
0163
0164
0165
0166
0167
0168
0169
0170
0171
0172
0173
0174
0175
0176
0177
0178
0179
0180
0181
0182
0183
0184
0185
0186
0187
0188
0189
0190
0191
0192
As a main result, we show that a matrix with diagonal entries being 1 is a compatible tail-
dependence matrix if and only if it is a Bernoulli-compatible matrix multiplied by a constant. In
Section 4 we provide probabilistic models for a large class of tail-dependence matrices, including
commonly used matrices in statistics. Section 5 concludes.
Throughout this paper, d and m are positive integers, and we consider an atomless prob-
ability space (Ω,A,P) on which all random variables and random vectors are defined. Vectors
are considered as column vectors. For two matrices A,B, B > A and B 6 A are understood as
component-wise inequalities. We let A B denote the Hadamard product, i.e., the element-wise
product of two matrices A and B of the same dimension. The d× d identity matrix is denoted
by Id. For a square matrix A, diag(A) represents a diagonal matrix with diagonal entries equal
to those of A, and A> is the transpose of A. We denote IE the indicator function of an event
(random or deterministic) E ∈ A. 0 and 1 are vectors with all components being 0 and 1
respectively, as long as the dimension of the vectors is clear from the context.
2 Bernoulli compatibility
In this section we introduce and study the Bernoulli-compatibility problem. The results
obtained in this section are the basis for the tail-dependence compatibility problem treated in
Section 3; many of them are of independent interest, e.g., for the simulation of sequences of
Bernoulli random variables.
2.1 Bernoulli-compatible matrices
Definition 2.1 (Bernoulli vector, Vd). A Bernoulli vector is a random vector X supported by
0, 1d for some d ∈ N. The set of all d-Bernoulli vectors is denoted by Vd.
Equivalently, X = (X1, . . . , Xd) is a Bernoulli vector if and only if Xi ∼ B(1, pi) for some
pi ∈ [0, 1], i = 1, . . . , d. Note that here we do not make any assumption about the dependence
structure among the components of X. Bernoulli vectors play an important role in Credit Risk
Analysis; see, e.g., Bluhm and Overbeck (2006) and Bluhm et al. (2002, Section 2.1).
In this section, we investigate the following question which we refer to as the Bernoulli-
compatibility problem:
Question 1. Given a matrix B ∈ [0, 1]d×d, can we find a Bernoulli vector X such that B =
E[XX>]?
For studying the Bernoulli-compatibility problem, we introduce the notion of Bernoulli-
compatible matrices.
3
0193
0194
0195
0196
0197
0198
0199
0200
0201
0202
0203
0204
0205
0206
0207
0208
0209
0210
0211
0212
0213
0214
0215
0216
0217
0218
0219
0220
0221
0222
0223
0224
0225
0226
0227
0228
0229
0230
0231
0232
0233
0234
0235
0236
0237
0238
0239
0240
0241
0242
0243
0244
0245
0246
0247
0248
0249
0250
0251
0252
0253
0254
0255
0256
Definition 2.2 (Bernoulli-compatible matrix, Bd). A d× d matrix B is a Bernoulli-compatible
matrix, if B = E[XX>] for some X ∈ Vd. The set of all d× d Bernoulli-compatible matrices is
denoted by Bd.
Concerning covariance matrices, there is extensive research on the compatibility of covari-
ance matrices of Bernoulli vectors in the realm of statistical simulation and time series analysis;
see, e.g., Chaganty and Joe (2006). It is known that, when d > 3, the set of all compatible d-
Bernoulli correlation matrices is strictly contained in the set of all correlation matrices. Note that
E[XX>] = Cov(X) + E[X]E[X]>. Hence, Question 1 is closely related to the characterization
of compatible Bernoulli covariance matrices.
Before we characterize the set Bd in Section 2.2 and thus address Question 1, we first collect
some facts about elements of Bd.
Proposition 2.1. Let B,B1, B2 ∈ Bd. Then
i) B ∈ [0, 1]d×d.
ii) maxbii + bjj − 1, 0 6 bij 6 minbii, bjj for i, j = 1, . . . , d and B = (bij)d×d.
iii) tB1 + (1− t)B2 ∈ Bd for t ∈ [0, 1], i.e., Bd is a convex set.
iv) B1 B2 ∈ Bd, i.e., Bd is closed under the Hadamard product.
v) (0)d×d ∈ Bd and (1)d×d ∈ Bd.
vi) For any p = (p1, . . . , pd) ∈ [0, 1]d, the matrix B = (bij)d×d ∈ Bd where bij = pipj for i 6= j
and bii = pi, i, j = 1, . . . , d.
Proof. Write B1 = E[XX>] and B2 = E[Y Y >] for X,Y ∈ Vd, and X and Y are independent.
i) Clear.
ii) This directly follows from the Frechet–Hoeffding bounds; see McNeil et al. (2015, Re-
mark 7.9).
iii) Let A ∼ B(1, t) be a Bernoulli random variable independent of X,Y , and let Z = AX +
(1 − A)Y . Then Z ∈ Vd, and E[ZZ>] = tE[XX>] + (1 − t)E[Y Y >] = tB1 + (1 − t)B2.
Hence tB1 + (1− t)B2 ∈ Bd.
iv) Let p = (p1, . . . , pd), q = (q1, . . . , qd) ∈ Rd. Then
(p q)(p q)> = (piqi)d(piqi)>d = (piqipjqj)d×d = (pipj)d×d (qiqj)d×d
= (pp>) (qq>).
4
0257
0258
0259
0260
0261
0262
0263
0264
0265
0266
0267
0268
0269
0270
0271
0272
0273
0274
0275
0276
0277
0278
0279
0280
0281
0282
0283
0284
0285
0286
0287
0288
0289
0290
0291
0292
0293
0294
0295
0296
0297
0298
0299
0300
0301
0302
0303
0304
0305
0306
0307
0308
0309
0310
0311
0312
0313
0314
0315
0316
0317
0318
0319
0320
Let Z = X Y . It follows that Z ∈ Vd and E[ZZ>] = E[(X Y )(X Y )>] = E[(XX>)
(Y Y >)] = E[XX>] E[Y Y >] = B1 B2. Hence B1 B2 ∈ Bd.
v) Consider X = 0 ∈ Vd. Then (0)d×d = E[XX>] ∈ Bd. Similarly for (1)d×d.
vi) Consider X ∈ Vd with independent components and E[X] = p.
2.2 Characterization of Bernoulli-compatible matrices
We are now able to give a characterization of the set Bd of Bernoulli-compatible matrices
and thus address Question 1.
Theorem 2.2 (Characterization of Bd). Bd has the following characterization:
Bd = n∑i=1
aipip>i : pi ∈ 0, 1d, ai > 0, i = 1, . . . , n,
n∑i=1
ai = 1, n ∈ N
; (2.1)
i.e., Bd is the convex hull of pp> : p ∈ 0, 1d. In particular, Bd is closed under convergence
in the Euclidean norm.
Proof. Denote the right-hand side of (2.1) by M. For B ∈ Bd, write B = E[XX>] for some
X ∈ Vd. It follows that
B =∑
p∈0,1dpp>P(X = p) ∈M,
hence Bd ⊆ M. Let X = p ∈ 0, 1d. Then X ∈ Vd and E[XX>] = pp> ∈ Bd. By
Proposition 2.1, Bd is a convex set which contains pp> : p ∈ 0, 1d, hence M ⊆ Bd. In
summary,M = Bd. From (2.1) we can see that Bd is closed under convergence in the Euclidean
norm.
A matrix B is completely positive if B = AA> for some (not necessarily square) matrix
A > 0. Denote by Cd the set of completely positive matrices. It is known that Cd is the convex
cone with extreme directions pp> : p ∈ [0, 1]d; see, e.g., Ruschendorf (1981) and Berman and
Shaked-Monderer (2003). We thus obtain the following result.
Corollary 2.3. Any Bernoulli-compatible matrix is completely positive.
Remark 2.1. One may wonder whether B = E[XX>] is sufficient to determine the distribution
of X, i.e., whether the decomposition
B =
2d∑i=1
aipip>i (2.2)
5
0321
0322
0323
0324
0325
0326
0327
0328
0329
0330
0331
0332
0333
0334
0335
0336
0337
0338
0339
0340
0341
0342
0343
0344
0345
0346
0347
0348
0349
0350
0351
0352
0353
0354
0355
0356
0357
0358
0359
0360
0361
0362
0363
0364
0365
0366
0367
0368
0369
0370
0371
0372
0373
0374
0375
0376
0377
0378
0379
0380
0381
0382
0383
0384
is unique for distinct vectors pi in 0, 1d. While the decomposition is trivially unique for d = 2,
this is in general false for d > 3, since there are 2d − 1 parameters in (2.2) and only d(d+ 1)/2
parameters in B. The following is an example for d = 3. Let
B =1
4
2 1 1
1 2 1
1 1 2
=
1
4
((1, 1, 1)>(1, 1, 1) + (1, 0, 0)>(1, 0, 0) + (0, 1, 0)>(0, 1, 0)
+ (0, 0, 1)>(0, 0, 1))
=1
4
((1, 1, 0)>(1, 1, 0) + (1, 0, 1)>(1, 0, 1) + (0, 1, 1)>(0, 1, 1)
+ (0, 0, 0)>(0, 0, 0)).
Thus, by combining the above two decompositions, B ∈ B3 has infinitely many different de-
compositions of the form (2.2). Note that, as in the case of completely positive matrices, it is
generally difficult to find decompositions of form (2.2) for a given matrix B.
2.3 Convex cone generated by Bernoulli-compatible matrices
In this section we study the convex cone generated by Bd, denoted by B∗d:
B∗d = aB : a > 0, B ∈ Bd. (2.3)
The following proposition is implied by Proposition 2.1 and Theorem 2.2.
Proposition 2.4. B∗d is the convex cone with extreme directions pp> : p ∈ 0, 1d. Moreover,
B∗d is a commutative semiring equipped with addition (B∗d,+) and multiplication (B∗d, ).
It is obvious that B∗d ⊆ Cd. One may wonder whether B∗d is identical to Cd, the set of
completely positive matrices. As the following example shows, this is false in general for d > 2.
Example 2.1. Note that B ∈ B∗d also satisfies Proposition 2.1, Part ii). Now consider p =
(p1, . . . , pd) ∈ (0, 1)d with pi > pj for some i 6= j. Clearly, pp> ∈ Cd, but pipj > p2j = minp2
i , p2j
contradicts Proposition 2.1, Part ii), hence pp> 6∈ B∗d.
For the following result, we need the notion of diagonally dominant matrices. A matrix
A ∈ Rd×d is called diagonally dominant if, for all i = 1, . . . , d,∑j 6=i |aij | 6 |aii|.
Proposition 2.5. Let Dd be the set of non-negative, diagonally dominant d× d-matrices. Then
Dd ⊆ B∗d.
6
0385
0386
0387
0388
0389
0390
0391
0392
0393
0394
0395
0396
0397
0398
0399
0400
0401
0402
0403
0404
0405
0406
0407
0408
0409
0410
0411
0412
0413
0414
0415
0416
0417
0418
0419
0420
0421
0422
0423
0424
0425
0426
0427
0428
0429
0430
0431
0432
0433
0434
0435
0436
0437
0438
0439
0440
0441
0442
0443
0444
0445
0446
0447
0448
Proof. For i, j = 1, . . . , d, let p(ij) = (p(ij)1 , . . . , p
(ij)d ) where p
(ij)k = Ik=i∪k=j. It is straightfor-
ward to verify that the (i, i)-, (i, j)-, (j, i)- and (j, j)-entries of the matrix M (ij) = p(ij)(p(ij))>
are 1, and the other entries are 0. For D = (dij)d×d ∈ Dd, let
D∗ = (d∗ij)d×d =
d∑i=1
d∑j=1,j 6=i
dijM(ij).
By Proposition 2.4, D∗ ∈ B∗d. It follows that d∗ij = dij for i 6= j and d∗ii =∑dj=1,j 6=i dij 6 dii.
Therefore, D = D∗ +∑di=1(dii − d∗ii)M (ii), which, by Proposition 2.4, is in B∗d.
For studying the tail-dependence compatibility problem in Section 3, the subset
BId = B : B ∈ B∗d, diag(B) = Id
of B∗d is of interest. It is straightforward to see from Proposition 2.1 and Theorem 2.2 that BIdis a convex set, closed under the Hadamard product and convergence in the Euclidean norm.
These properties of BId will be used later.
3 Tail-dependence compatibility
3.1 Tail-dependence matrices
The notion of tail dependence captures (extreme) dependence in the lower-left or upper-
right tails of a bivariate distribution. In what follows, we focus on lower-left tails; the problem
for upper-right tails follows by a reflection around (1/2, 1/2), i.e., studying the survival copula
of the underlying copula.
Definition 3.1 (Tail-dependence coefficient). The (lower) tail-dependence coefficient of two
continuous random variables X1 ∼ F1 and X2 ∼ F2 is defined by
λ = limu↓0
P(F1(X1) 6 u, F2(X2) 6 u)
u, (3.1)
given that the limit exists.
If we denote the copula of (X1, X2) by C, then
λ = limu↓0
C(u, u)
u.
Clearly λ ∈ [0, 1], and λ only depends on the copula of (X1, X2), not the marginal distributions.
For virtually all copula models used in practice, the limit in (3.1) exists; for how to construct an
example where λ does not exist, see Kortschak and Albrecher (2009).
7
0449
0450
0451
0452
0453
0454
0455
0456
0457
0458
0459
0460
0461
0462
0463
0464
0465
0466
0467
0468
0469
0470
0471
0472
0473
0474
0475
0476
0477
0478
0479
0480
0481
0482
0483
0484
0485
0486
0487
0488
0489
0490
0491
0492
0493
0494
0495
0496
0497
0498
0499
0500
0501
0502
0503
0504
0505
0506
0507
0508
0509
0510
0511
0512
Definition 3.2 (Tail-dependence matrix, Td). Let X = (X1, . . . , Xd) be a random vector with
continuous marginal distributions. The tail-dependence matrix of X is Λ = (λij)d×d, where λij
is the tail-dependence coefficient of Xi and Xj , i, j = 1, . . . , d. We denote by Td the set of all
tail-dependence matrices.
The following proposition summarizes basic properties of tail-dependence matrices. Its
proof is very similar to that of Proposition 2.1 and is omitted here.
Proposition 3.1. For any Λ1,Λ2 ∈ Td, we have that
i) Λ1 = Λ>1 .
ii) tΛ1 + (1− t)Λ2 ∈ Td for t ∈ [0, 1], i.e., Td is a convex set.
iii) Id 6 Λ1 6 (1)d×d with Id ∈ Td and (1)d×d ∈ Td.
As we will show next, Td is also closed under the Hadamard product.
Proposition 3.2. Let k ∈ N and Λ1, . . . ,Λk ∈ Td. Then Λ1 · · · Λk ∈ Td.
Proof. Note that it would be sufficient to show the result for k = 2, but we provide a general
construction for any k. For each l = 1, . . . , k, let Cl be a d-dimensional copula with tail-
dependence matrix Λl. Furthermore, let g(u) = u1/k, u ∈ [0, 1]. It follows from Liebscher (2008)
that C(u1, . . . , ud) =∏kl=1 Cl(g(u1), . . . , g(ud)) is a copula; note that
(g−1( max
16l6kUl1), . . . , g−1( max
16l6kUld)
)∼ C (3.2)
for independent random vectors (Ul1, . . . , Uld) ∼ Cl, l = 1, . . . , k. The (i, j)-entry λij of Λ
corresponding to C is thus given by
λij = limu↓0
∏kl=1 Cl,ij(g(u), g(u))
u= lim
u↓0
k∏l=1
Cl,ij(g(u), g(u))
g(u)
=
k∏l=1
limu↓0
Cl,ij(g(u), g(u))
g(u)=
k∏l=1
limu↓0
Cl,ij(u, u)
u=
k∏l=1
λl,ij ,
where Cl,ij denotes the (i, j)-margin of Cl and λl,ij denotes the (i, j)th entry of Λl, l = 1, . . . , k.
3.2 Characterization of tail-dependence matrices
In this section, we investigate the following question:
Question 2. Given a d× d matrix Λ ∈ [0, 1]d×d, is it a tail-dependence matrix?
8
0513
0514
0515
0516
0517
0518
0519
0520
0521
0522
0523
0524
0525
0526
0527
0528
0529
0530
0531
0532
0533
0534
0535
0536
0537
0538
0539
0540
0541
0542
0543
0544
0545
0546
0547
0548
0549
0550
0551
0552
0553
0554
0555
0556
0557
0558
0559
0560
0561
0562
0563
0564
0565
0566
0567
0568
0569
0570
0571
0572
0573
0574
0575
0576
The following theorem fully characterizes tail-dependence matrices and thus provides a
theoretical (but not necessarily practical) answer to Question 2.
Theorem 3.3 (Characterization of Td). A square matrix with diagonal entries being 1 is a tail-
dependence matrix if and only if it is a Bernoulli-compatible matrix multiplied by a constant.
Equivalently, Td = BId.
Proof. We first show that Td ⊆ BId. For each Λ = (λij)d×d ∈ Td, suppose that C is a copula
with tail-dependence matrix Λ and U = (U1, . . . , Un) ∼ C. Let Wu = (IU16u, . . . , IUd6u).
By definition,
λij = limu↓0
1
uE[IUi6uIUj6u]
and
Λ = limu↓0
1
uE[WuW
>u ].
Since BId is closed and E[WuW>u ]/u ∈ BId, we have that Λ ∈ BId.
Now consider BId ⊆ Td. By definition of BId, each B ∈ BId can be written as B = E[XX>]/p
for an X ∈ Vd and E[X] = (p, . . . , p) ∈ (0, 1]d. Let U, V ∼ U[0, 1], U, V,X be independent and
Y = XpU + (1−X)(p+ (1− p)V ). (3.3)
We can verify that for t ∈ [0, 1] and i = 1, . . . , d,
P(Yi 6 t) = P(Xi = 1)P(pU 6 t) + P(Xi = 0)P(p+ (1− p)V 6 t)
= pmint/p, 1+ (1− p) max(t− p)/(1− p), 0 = t,
i.e., Y1, . . . , Yd are U[0, 1]-distributed. Let λij be the tail-dependence coefficient of Yi and Yj ,
i, j = 1, . . . , d. For i, j = 1, . . . , d we obtain that
λij = limu↓0
1
uP(Yi 6 u, Yj 6 u) = lim
u↓0
1
uP(Xi = 1, Xj = 1)P(pU 6 u)
=1
pE[XiXj ].
As a consequence, the tail-dependence matrix of (Y1, . . . , Yd) is B and B ∈ Td.
It follows from Theorem 3.3 and Proposition 2.4 that Td is the “1-diagonals” cross-section
of the convex cone with extreme directions pp> : p ∈ 0, 1d. Furthermore, the proof of
Theorem 3.3 is constructive. As we saw, for any B ∈ BId, Y defined by (3.3) has tail-dependence
matrix B. This interesting construction will be applied in Section 4 where we show that com-
monly applied matrices in statistics are tail-dependence matrices and where we derive the copula
of Y .
9
0577
0578
0579
0580
0581
0582
0583
0584
0585
0586
0587
0588
0589
0590
0591
0592
0593
0594
0595
0596
0597
0598
0599
0600
0601
0602
0603
0604
0605
0606
0607
0608
0609
0610
0611
0612
0613
0614
0615
0616
0617
0618
0619
0620
0621
0622
0623
0624
0625
0626
0627
0628
0629
0630
0631
0632
0633
0634
0635
0636
0637
0638
0639
0640
Remark 3.1. From the fact that Td = BId and BId is closed under the Hadamard product (see
Proposition 2.1, Part iv)), Proposition 3.2 directly follows. Note, however, that our proof of
Proposition 3.2 is constructive. Given tail-dependence matrices and corresponding copulas, we
can construct a copula C which has the Hadamard product of the tail-dependence matrices as
corresponding tail-dependence matrix. If sampling of all involved copulas is feasible, we can
sample C; see Figure 1 for examples2.
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
U1
U2
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
U1
U2
Figure 1: Left-hand side: Scatter plot of 2000 samples from (3.2) for C1 being a Clayton copula
with parameter θ = 4 (λ1 = 2−1/4 ≈ 0.8409) and C2 being a t3 copula with parameter ρ = 0.8
(tail-dependence coefficient λ2 = 2t4(−2/3) ≈ 0.5415). By Proposition 3.2, the tail-dependence
coefficient of (3.2) is thus λ = λ1λ2 = 23/4t4(−2/3) ≈ 0.4553. Right-hand side: C1 as before,
but C2 is a survival Marshall–Olkin copula with parameters α1 = 2−3/4, α2 = 0.8, so that
λ = λ1λ2 = 1/2.
Theorem 3.3 combined with Corollary 2.3 directly leads to the following result.
Corollary 3.4. Every tail-dependence matrix is completely positive, and hence positive semi-
definite.
Furthermore, Theorem 3.3 and Proposition 2.5 imply the following result.
Corollary 3.5. Every diagonally dominant matrix with non-negative entries and diagonal en-
tries being 1 is a tail-dependence matrix.
Note that this result already yields the if-part of Proposition 4.7 below.
2All plots can be reproduced via the R package copula (version > 0.999-13) by calling
demo(tail compatibility).
10
0641
0642
0643
0644
0645
0646
0647
0648
0649
0650
0651
0652
0653
0654
0655
0656
0657
0658
0659
0660
0661
0662
0663
0664
0665
0666
0667
0668
0669
0670
0671
0672
0673
0674
0675
0676
0677
0678
0679
0680
0681
0682
0683
0684
0685
0686
0687
0688
0689
0690
0691
0692
0693
0694
0695
0696
0697
0698
0699
0700
0701
0702
0703
0704
4 Compatible models for tail-dependence matrices
4.1 Widely known matrices
We now consider the following three types of matrices Λ = (λij)d×d which are frequently
applied in multivariate statistics and time series analysis and show that they are tail-dependence
matrices.
a) Equicorrelation matrix with parameter α ∈ [0, 1]: λij = Ii=j + αIi6=j, i, j = 1, . . . , d.
b) AR(1) matrix with parameter α ∈ [0, 1]: λij = α|i−j|, i, j = 1, . . . , d.
c) MA(1) matrix with parameter α ∈ [0, 1/2]: λij = Ii=j + αI|i−j|=1, i, j = 1, . . . , d.
Chaganty and Joe (2006) considered the compatibility of correlation matrices of Bernoulli vectors
for the above three types of matrices and obtained necessary and sufficient conditions for the
existence of compatible models for d = 3. For the tail-dependence compatibility problem that
we consider in this paper, the above three types of matrices are all compatible, and we are able
to construct corresponding models for each case.
Proposition 4.1. Let Λ be the tail-dependence matrix of the d-dimensional random vector
Y = XpU + (1−X)(p+ (1− p)V ), (4.1)
where U, V ∼ U[0, 1], X ∈ Vd and U, V,X are independent.
i) For α ∈ [0, 1], if X has independent components and E[X1] = · · · = E[Xd] = α, then Λ is
an equicorrelation matrix with parameter α; i.e., a) is a tail-dependence matrix.
ii) For α ∈ [0, 1], if Xi =∏i+d−1j=i Zj, i = 1, . . . , d, for independent B(1, α) random variables
Z1, . . . , Z2d−1, then Λ is an AR(1) matrix with parameter α; i.e., b) is a tail-dependence
matrix.
iii) For α ∈ [0, 1/2], if Xi = IZ∈[(i−1)(1−α),(i−1)(1−α)+1], i = 1, . . . , d, for Z ∼ U[0, d], then
Λ is an MA(1) matrix with parameter α; i.e., c) is a tail-dependence matrix.
Proof. We have seen in the proof of Theorem 3.3 that if E[X1] = · · · = E[Xd] = p, then Y
defined through (4.1) has tail-dependence matrix E[XX>]/p. Write Λ = (λij)d×d and note that
λii = 1, i = 1, . . . , d, is always guaranteed.
i) For i 6= j, we have that E[XiXj ] = α2 and thus λij = α2/α = α. This shows that Λ is an
equicorrelation matrix with parameter α.
11
0705
0706
0707
0708
0709
0710
0711
0712
0713
0714
0715
0716
0717
0718
0719
0720
0721
0722
0723
0724
0725
0726
0727
0728
0729
0730
0731
0732
0733
0734
0735
0736
0737
0738
0739
0740
0741
0742
0743
0744
0745
0746
0747
0748
0749
0750
0751
0752
0753
0754
0755
0756
0757
0758
0759
0760
0761
0762
0763
0764
0765
0766
0767
0768
ii) For i < j, we have that
E[XiXj ] = E[ i+d−1∏
k=i
Zk
j+d−1∏l=j
Zl
]= E
[ j−1∏k=i
Zk
]E[ i+d−1∏
k=j
Zk
]E[ j+d−1∏k=i+d
Zk
]= αj−iαi+d−jαj−i = αj−i+d
and E[Xi] = E[X2i ] = αd. Hence, λij = αj−i+d/αd = αj−i for i < j. By symmetry,
λij = α|i−j| for i 6= j. Thus, Λ is an AR(1) matrix with parameter α.
iii) For i < j, note that 2(1− α) > 1, so
E[XiXj ] = P(Z ∈ [(j − 1)(1− α), (i− 1)(1− α) + 1])
= Ij=i+1P(Z ∈ [i(1− α), (i− 1)(1− α) + 1]) = Ij=i+1α
d
and E[Xi] = E[X2i ] = 1
d . Hence, λij = αIj−i=1 for i < j. By symmetry, λij = αI|i−j|=1
for i 6= j. Thus, Λ is an MA(1) matrix with parameter α.
4.2 Advanced tail-dependence models
Theorem 3.3 gives a characterization of tail-dependence matrices using Bernoulli-compatible
matrices and (3.3) provides a compatible model Y for any tail-dependence matrix Λ(= E[XX>]/p).
It is generally not easy to check whether a given matrix is a Bernoulli-compatible matrix or
a tail-dependence matrix; see also Remark 2.1. Therefore, we now study the following question.
Question 3. How can we construct a broader class of models with flexible dependence structures
and desired tail-dependence matrices?
To enrich our models, we bring random matrices with Bernoulli entries into play. For
d,m ∈ N, let
Vd×m =X = (Xij)d×m : P(X ∈ 0, 1d×m) = 1,
m∑j=1
Xij 6 1, i = 1, . . . , d,
i.e., Vd×m is the set of d × m random matrices supported in 0, 1d×m with each row being
mutually exclusive; see Dhaene and Denuit (1999). Furthermore, we introduce a transformation
L on the set of square matrices, such that, for any i, j = 1, . . . , d, the (i, j)th element bij of L(B)
is given by
bij =
bij , if i 6= j,
1, if i = j;
(4.2)
i.e., L adjusts the diagonal entries of a matrix to be 1, and preserves all the other entries. For a
set S of square matrices, we set L(S) = L(B) : B ∈ S. We can now address Question 3.
12
0769
0770
0771
0772
0773
0774
0775
0776
0777
0778
0779
0780
0781
0782
0783
0784
0785
0786
0787
0788
0789
0790
0791
0792
0793
0794
0795
0796
0797
0798
0799
0800
0801
0802
0803
0804
0805
0806
0807
0808
0809
0810
0811
0812
0813
0814
0815
0816
0817
0818
0819
0820
0821
0822
0823
0824
0825
0826
0827
0828
0829
0830
0831
0832
Theorem 4.2 (A class of flexible models). Let U ∼ CU for an m-dimensional copula CU with
tail-dependence matrix Λ and let let V ∼ CV for a d-dimensional copula CV with tail-dependence
matrix Id. Furthermore, let X ∈ Vd×m such that X,U ,V are independent and let
Y = XU + Z V , (4.3)
where Z = (Z1, . . . , Zd) with Zi = 1 −∑mk=1Xik, i = 1, . . . , d. Then Y has tail-dependence
matrix Γ = L(E[XΛX>]).
Proof. Write X = (Xij)d×m, U = (U1, . . . , Um), V = (V1, . . . , Vd), Λ = (λij)d×d and Y =
(Y1, . . . , Yd). Then, for all i = 1, . . . , d,
Yi =
m∑k=1
XikUk + ZiVi =
Vi, if Xik = 0 for all k = 1, . . . ,m, so Zi = 1,
Uk, if Xik = 1 for some k = 1, . . . ,m, so Zi = 0.
Clearly, Y has U[0, 1] margins. We now calculate the tail-dependence matrix Γ = (γij)d×d of Y
for i 6= j. By our independence assumptions, we can derive the following results:
i) P(Yi 6 u, Yj 6 u, Zi = 1, Zj = 1) = P(Vi 6 u, Vj 6 u, Zi = 1, Zj = 1) = CVij (u, u)P(Zi =
1, Zj = 1) 6 CVij (u, u), where CV
ij denotes the (i, j)th margin of CV . As V has tail-
dependence matrix Id, we obtain that
limu↓0
1
uP(Yi 6 u, Yj 6 u, Zi = 1, Zj = 1) = 0.
ii) P(Yi 6 u, Yj 6 u, Zi = 0, Zj = 1) =∑mk=1 P(Uk 6 u, Vj 6 u,Xik = 1, Zj = 1) =∑m
k=1 P(Uk 6 u)P(Vj 6 u)P(Xik = 1, Zj = 1) 6 u2 and thus
limu↓0
1
uP(Yi 6 u, Yj 6 u, Zi = 0, Zj = 1) = 0.
Similarly, we obtain that
limu↓0
1
uP(Yi 6 u, Yj 6 u, Zi = 1, Zj = 0) = 0.
iii) P(Yi 6 u, Yj 6 u, Zi = 0, Zj = 0) =∑mk=1
∑ml=1 P(Uk 6 u, Ul 6 u,Xik = 1, Xjl = 1) =∑m
k=1
∑ml=1 C
Ukl (u, u)P(Xik = 1, Xjl = 1) =
∑mk=1
∑ml=1 C
Ukl (u, u)E[XikXjl] so that
limu↓0
1
uP(Yi 6 u, Yj 6 u, Zi = 0, Zj = 0)
=
m∑k=1
m∑l=1
λklE[XikXjl] = E[ m∑k=1
m∑l=1
XikλklXjl
]=(E[XΛX>]
)ij.
By the Law of Total Probability, we thus obtain that
γij = limu↓0
P(Yi 6 u, Yj 6 u)
u= lim
u↓0
P(Yi 6 u, Yj 6 u, Zi = 0, Zj = 0)
u
=(E[XΛX>]
)ij.
13
0833
0834
0835
0836
0837
0838
0839
0840
0841
0842
0843
0844
0845
0846
0847
0848
0849
0850
0851
0852
0853
0854
0855
0856
0857
0858
0859
0860
0861
0862
0863
0864
0865
0866
0867
0868
0869
0870
0871
0872
0873
0874
0875
0876
0877
0878
0879
0880
0881
0882
0883
0884
0885
0886
0887
0888
0889
0890
0891
0892
0893
0894
0895
0896
This shows that E[XΛX>] and Γ agree on the off-diagonal entries. Since Γ ∈ Td implies that
diag(Γ) = Id, we conclude that L(E[XΛX>]) = Γ.
A special case of Theorem 4.2 reveals an essential difference between the transition rules
of a tail-dependence matrix and a covariance matrix. Suppose that for X ∈ Vd×m, E[X] is a
stochastic matrix (each row sums to 1), and U ∼ CU for an m-dimensional copula CU with
tail-dependence matrix Λ = (λij)d×d. Now we have that Zi = 0, i = 1, . . . , d in (4.3). By
Theorem 4.2, the tail dependence matrix of Y = XU is given by L(E[XΛX>]). One can check
the diagonal terms of the matrix Λ∗ = (λ∗ij)d×d = XΛX> by
λ∗ii =
m∑j=1
m∑k=1
XikλkjXij =
m∑k=1
Xikλkk = 1, i = 1, . . . ,m.
Hence, the tail-dependence matrix of Y is indeed E[XΛX>].
Remark 4.1. In summary:
i) If an m-vector U has covariance matrix Σ, then XU has covariance matrix E[XΣX>] for
any d×m random matrix X independent of U .
ii) If an m-vector U has uniform [0,1] margins and tail-dependence matrix Λ, then XU has
tail-dependence matrix E[XΛX>] for any X ∈ Vd×m independent of U such that each row
of X sums to 1.
It is noted that the transition property of tail-dependence matrices is more restricted than that
of covariance matrices.
The following two propositions consider selected special cases of this construction which are
more straightforward to apply.
Proposition 4.3. For any B ∈ Bd and any Λ ∈ Td we have that L(B Λ) ∈ Td. In particular,
L(B) ∈ Td and hence L(Bd) ⊆ Td.
Proof. Write B = (bij)d×d = E[WW>] for some W = (W1, . . . ,Wd) ∈ Vd and consider X =
diag(W ) ∈ Vd×d. As in the proof of Theorem 4.2 (and with the same notation), it follows that
for i 6= j, γij = E[XiiλijXjj ] = E[WiWjλij ]. This shows that E[XΛX>] = E[WW> Λ] and
B Λ agree on off-diagonal entries. Thus, L(B Λ) = Γ ∈ Td. By taking Λ = (1)d×d, we obtain
L(B) ∈ Td.
The following proposition states a relationship between substochastic matrices and tail-
dependence matrices. To this end, let
Qd×m =
Q = (qij)d×m :
m∑j=1
qij 6 1, qij > 0, i = 1, . . . , d, j = 1, . . . ,m
,
14
0897
0898
0899
0900
0901
0902
0903
0904
0905
0906
0907
0908
0909
0910
0911
0912
0913
0914
0915
0916
0917
0918
0919
0920
0921
0922
0923
0924
0925
0926
0927
0928
0929
0930
0931
0932
0933
0934
0935
0936
0937
0938
0939
0940
0941
0942
0943
0944
0945
0946
0947
0948
0949
0950
0951
0952
0953
0954
0955
0956
0957
0958
0959
0960
i.e., Qd×m is the set of d×m (row) substochastic matrices; note that the expectation of a random
matrix in Vd×m is a substochastic matrix.
Proposition 4.4. For any Q ∈ Qd×m and any Λ ∈ Tm, we have that L(QΛQ>) ∈ Td. In
particular, L(QQ>) ∈ Td for all Q ∈ Qd×m and L(pp>) ∈ Td for all p ∈ [0, 1]d.
Proof. Write Q = (qij)d×m and let Xik = IZi∈[∑k−1
j=1 qij ,∑k
j=1 qij) for independent Zi ∼ U[0, 1],
i = 1, . . . , d, k = 1, . . . ,m. It is straightforward to see that E[X] = Q, X ∈ Vd×m with
independent rows, and∑mk=1Xik 6 1 for i = 1, . . . , d, so X ∈ Vd×m. As in the proof of
Theorem 4.2 (and with the same notation), it follows that for i 6= j,
γij =
m∑l=1
m∑k=1
E[Xik]E[Xjl]λkl =
m∑l=1
m∑k=1
qikqjlλkl.
This shows that QΛQ> and Γ agree on off-diagonal entries, so L(QΛQ>) = Γ ∈ Td. By taking
Λ = Id, we obtain L(QQ>) ∈ Td. By taking m = 1, we obtain L(pp>) ∈ Td.
4.3 Corresponding copula models
In this section, we derive the copulas of (3.3) and (4.3) which are able to produce tail-
dependence matrices E[XX>]/p and L(E[XΛX>]) as stated in Theorems 3.3 and 4.2, respec-
tively. We first address the former.
Proposition 4.5 (Copula of (3.3)). Let X ∈ Vd, E[X] = (p, . . . , p) ∈ (0, 1]d. Furthermore, let
U, V ∼ U[0, 1], U, V,X be independent and
Y = XpU + (1−X)(p+ (1− p)V ).
Then the copula C of Y at u = (u1, . . . , ud) is given by
C(u) =∑
i∈0,1dmin
minr:ir=1urp
, 1
maxminr:ir=0ur − p
1− p, 0P(X = i),
with the convention min ∅ = 1.
Proof. By the Law of Total Probability and our independence assumptions,
C(u) =∑
i∈0,1dP(Y 6 u,X = i)
=∑
i∈0,1dP(pU 6 min
r:ir=1ur, p+ (1− p)V 6 min
r:ir=0ur,X = i)
=∑
i∈0,1dP(U 6
minr:ir=1urp
)P(V 6
minr:ir=0ur − p1− p
)P(X = i);
the claim follows from the fact that U, V ∼ U[0, 1].
15
0961
0962
0963
0964
0965
0966
0967
0968
0969
0970
0971
0972
0973
0974
0975
0976
0977
0978
0979
0980
0981
0982
0983
0984
0985
0986
0987
0988
0989
0990
0991
0992
0993
0994
0995
0996
0997
0998
0999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
For deriving the copula of (4.3), we need to introduce some notation; see also Example 4.1
below. In the following theorem, let supp(X) denote the support of X. For a vector u =
(u1, . . . , ud) ∈ [0, 1]d and a matrix A = (Aij)d×m ∈ supp(X), denote by Ai the sum of the
i-th row of A, i = 1, . . . , d, and let uA = (u1IA1=0 + IA1=1, . . . , udIAd=0 + IAd=1), and
u∗A = (minr:Ar1=1ur, . . . ,minr:Arm=1ur), where min ∅ = 1.
Proposition 4.6 (Copula of (4.3)). Suppose that the setup of Theorem 4.2 holds. Then the
copula C of Y in (4.3) is given by
C(u) =∑
A∈supp(X)
CV (uA)CU (u∗A)P(X = A). (4.4)
Proof. By the Law of Total Probability, it suffices to verify that P(Y 6 u |X = A) = CV (uA)CU (u∗A).
This can be seen from
P(Y 6 u |X = A)
= P( m∑k=1
AjkUk + (1−Aj)Vj 6 uj , j = 1, . . . , d
)= P(UkIAjk=1 6 uj , VjIAj=0 6 uj , j = 1, . . . , d, k = 1, . . . ,m)
= P(Uk 6 min
r:Ark=1ur, Vj 6 ujIAj=0 + IAj=1, j = 1, . . . , d, k = 1, . . . ,m
)= P
(Uk 6 min
r:Ark=1ur, k = 1, . . . ,m
)P(Vj 6 ujIAj=0 + IAj=1, j = 1, . . . , d
)= CU (u∗A)CV (uA).
As long as CV has tail-dependence matrix Id, the tail-dependence matrix of Y is not
affected by the choice of CV . This theoretically provides more flexibility in choosing the body
of the distribution of Y while attaining a specific tail-dependence matrix. Note, however, that
this also depends on the choice of X; see the following example where we address special cases
which allow for more insight into the rather abstract construction (4.4).
Example 4.1. 1. For m = 1, the copula C in (4.4) is given by
C(u) =∑
A∈0,1dCV (uA)CU (u∗A)P(X = A); (4.5)
note that X,A in Equation 4.4 are indeed vectors in this case. For d = 2, we obtain
C(u1, u2) = M(u1, u2)P(X = ( 11 )) + CV (u1, u2)P(X = ( 0
0 ))
+ Π(u1, u2)P(X = ( 10 ) or X = ( 0
1 )),
16
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
and therefore a mixture of the Frechet–Hoeffding upper bound M(u1, u2) = minu1, u2,
the copula CV and the independence copula Π(u1, u2) = u1u2. If P(X = ( 00 )) = 0 then C
is simply a mixture of M and Π and does not depend on V anymore.
Now consider the special case of (4.5) where V follows the d-dimensional independence cop-
ula Π(u) =∏di=1 ui and X = (X1, . . . , Xd−1, 1) is such that at most one of X1, . . . , Xd−1
is 1 (each randomly with probability 0 6 α 6 1/(d− 1) and all are simultaneously 0 with
probability 1− (d− 1)α). Then, for all u ∈ [0, 1]d, C is given by
C(u) = α
d−1∑i=1
(minui, ud
d−1∏j=1,j 6=i
uj
)+ (1− (d− 1)α)
d∏j=1
uj . (4.6)
This copula is a conditionally independent multivariate Frechet copula studied in Yang
et al. (2009). This example will be revisited in Section 4.4; see also the left-hand side of
Figure 3 below.
2. For m = 2, d = 2, we obtain
C(u1, u2) = M(u1, u2)P(X = ( 1 01 0 ) or X = ( 0 1
0 1 ))
+ CU (u1, u2)P(X = ( 1 00 1 )) + CU (u2, u1)P(X = ( 0 1
1 0 ))
+ CV (u1, u2)P(X = ( 0 00 0 ))
+ Π(u1, u2)P(X = ( 0 01 0 ) or ( 0 0
0 1 ) or ( 1 00 0 ) or ( 0 1
0 0 )). (4.7)
Figure 2 shows samples of size 2000 from (4.7) for V ∼ Π and two different choices of U
(in different rows) and X (in different columns). From Theorem 4.2, we obtain that the
off-diagonal entry γ12 of the tail-dependence matrix Γ of Y is given by
γ12 = p(1,2)(1,1) + p(1,2)(2,2) + λ12(p(1,2)(2,1) + p(1,2)(1,2)),
where λ12 is the off-diagonal entry of the tail-dependence matrix Λ of U .
4.4 An example from risk management practice
Let us now come back to Problem (1.1) which motivated our research on tail-dependence
matrices. From a practical point of view, the question is whether it is possible to find one
financial position, which has tail-dependence coefficient α with each of d − 1 tail-independent
financial risks (assets). Such a construction can be interesting for risk management purposes,
e.g., in the context of hedging.
Recall Problem (1.1):
17
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Y1
Y2
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Y1
Y2
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Y1
Y2
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Y1
Y2
Figure 2: Scatter plots of 2000 samples from Y for V ∼ Π and U following a bivariate (m = 2)
t3 copula with Kendall’s tau equal to 0.75 (top row) or a survival Marshall–Olkin copula with
parameters α1 = 0.25, α2 = 0.75 (bottom row). For the plots on the left-hand side, the number
of rows of X with one 1 are randomly chosen among 0, 1, 2(= d), the corresponding rows and
columns are then randomly selected among 1, 2(= d) and 1, 2(= m), respectively. For the
plots on the right-hand side, X is drawn from a multinomial distribution with probabilities 0.5
and 0.5 such that each row contains precisely one 1.
18
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
For which α ∈ [0, 1] is the matrix
Γd(α) =
1 0 · · · 0 α
0 1 · · · 0 α...
.... . .
......
0 0 · · · 1 α
α α · · · α 1
(4.8)
a matrix of pairwise (either lower or upper) tail-dependence coefficients?
Based on the Frechet–Hoeffding bounds, it follows from Joe (1997, Theorem 3.14) that for
d = 3 (and thus also d > 3), α has to be in [0, 1/2]; however this is not a sufficient condition
for Γd(α) to be a tail-dependence matrix. The following proposition not only gives an answer
to (4.8) by providing necessary and sufficient such conditions, but also provides, by its proof, a
compatible model for Γd(α).
Proposition 4.7. Γd(α) ∈ Td if and only if 0 6 α 6 1/(d− 1).
Proof. The if-part directly follows from Corollary 3.5. We provide a constructive proof based on
Theorem 4.2. Suppose that 0 6 α 6 1/(d−1). Take a partition Ω1, . . . ,Ωd of the sample space
Ω with P(Ωi) = α, i = 1, . . . , d− 1, and let X = (IΩ1, . . . , IΩd−1
, 1) ∈ Vd. It is straightforward to
see that
E[XX>] =
α 0 · · · 0 α
0 α · · · 0 α...
.... . .
......
0 0 · · · α α
α α · · · α 1
.
By Proposition 4.3, Γd(α) = L(E[XX>]) ∈ Td.
For the only if part, suppose that Γd(α) ∈ Td; thus α > 0. By Theorem 3.3, Γd(α) ∈ BId.
By the definition of BId, Γd(α) = Bd/p for some p ∈ (0, 1] and a Bernoulli-compatible matrix Bd.
Therefore,
pΓd(α) =
p 0 · · · 0 pα
0 p · · · 0 pα...
.... . .
......
0 0 · · · p pα
pα pα · · · pα p
19
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
is a compatible Bernoulli matrix, so pΓd(α) ∈ Bd. Write pΓd(α) = E[XX>] for some X =
(X1, . . . , Xd) ∈ Vd. It follows that P(Xi = 1) = p for i = 1, . . . , d, P(XiXj = 1) = 0 for
i 6= j, i, j = 1, . . . , d − 1 and P(XiXd = 1) = pα for i = 1, . . . , d − 1. Note that XiXd = 1,
i = 1, . . . , d − 1, are almost surely disjoint since P(XiXj = 1) = 0 for i 6= j, i, j = 1, . . . , d − 1.
As a consequence,
p = P(Xd = 1) > P( d−1⋃i=1
XiXd = 1)
=
d−1∑i=1
P(XiXd = 1) = (d− 1)pα,
and thus (d− 1)α 6 1.
It follows from the proof of Theorem 4.2 that for α ∈ [0, 1/(d − 1)], a compatible copula
model with tail-dependence matrix Γd(α) can be constructed as follows. Consider a partition
Ω1, . . . ,Ωd of the sample space Ω with P(Ωi) = α, i = 1, . . . , d−1, and let X = (X1, . . . , Xd) =
(IΩ1 , . . . , IΩd−1, 1) ∈ Vd; note that m = 1 here. Furthermore, let V be as in Theorem 4.2,
U ∼ U[0, 1] and U,V ,X be independent. Then,
Y = (UX1 + (1−X1)V1, . . . , UXd−1 + (1−Xd−1)Vd−1, U)
has tail-dependence matrix Γd(α). Example 4.1, Part 1 provides the copula C of Y in this case.
It is also straightforward to verify from this copula that Y has tail-dependence matrix Γd(α).
Figure 3 displays pairs plots of 2000 realizations of Y for α = 1/3 and two different copulas for
V .
Remark 4.2. Note that Γd(α) is not positive semidefinite if and only if α > 1/√d− 1. For
d < 5, element-wise non-negative and positive semidefinite matrices are completely positive; see
Berman and Shaked-Monderer (2003, Theorem 2.4). Therefore Γ3(2/3) is completely positive.
However, it is not in T3. It indeed shows that the class of completely positive matrices with
diagonal entries being 1 is strictly larger than Td.
5 Conclusion and discussion
Inspired by the question whether a given matrix in [0, 1]d×d is the matrix of pairwise tail-
dependence coefficients of a d-dimensional random vector, we introduced the tail-dependence
compatibility problem. It turns out that this problem is closely related to the Bernoulli-
compatibility problem which we also addressed in this paper and which asks when a given matrix
in [0, 1]d×d is a Bernoulli-compatible matrix (see Question 1 and Theorem 2.2). As a main find-
ing, we characterized tail-dependence matrices as precisely those square matrices with diagonal
entries being 1 which are Bernoulli-compatible matrices multiplied by a constant (see Question 2
20
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
Y1
0.0 0.4 0.8
0.0 0.4 0.8
0.0
0.4
0.8
0.0
0.4
0.8
Y2
Y3
0.0
0.4
0.8
0.0 0.4 0.8
0.0
0.4
0.8
0.0 0.4 0.8
Y4
Y1
0.0 0.4 0.8
0.0 0.4 0.8
0.0
0.4
0.8
0.0
0.4
0.8
Y2
Y3
0.0
0.4
0.8
0.0 0.4 0.8
0.0
0.4
0.8
0.0 0.4 0.8
Y4
Figure 3: Pairs plot of 2000 samples from Y ∼ C which produces the tail dependence matrix
Γ4(1/3) as given by (1.1). On the left-hand side, V ∼ Π (α determines how much weight is on
the diagonal for pairs with one component being Y4; see (4.6)) and on the right-hand-side, V
follows a Gauss copula with parameter chosen such that Kendall’s tau equals 0.8.
and Theorem 3.3). Furthermore, we presented and studied new models (see, e.g., Question 3
and Theorem 4.2) which provide answers to several questions related to the tail-dependence
compatibility problem.
The study of compatibility of tail-dependence matrices is mathematically different from
that of covariances matrices. Through many technical arguments in this paper, the reader may
have already realized that the tail-dependence matrix lacks a linear structure which is essential
to covariance matrices based on tools from Linear Algebra. For instance, let X be a d-random
vector with covariance matrix Σ and tail-dependence matrix Λ, and A be an m× d matrix. The
covariance matrix of AX is simply given by AΣA>, however, the tail-dependence matrix of AX
is generally not explicit (see Remark 4.1 for special cases). This lack of linearity can also help
to understand why tail-dependence matrices are realized by models based on Bernoulli vectors
as we have seen in this paper, in contrast to covariance matrices which are naturally realized by
Gaussian (or generally, elliptical) random vectors. The latter have a linear structure, whereas
Bernoulli vectors do not. It is not surprising that most classical techniques in Linear Algebra
such as matrix decomposition, diagonalization, ranks, inverses and determinants are not very
helpful for studying the compatibility problems we address in this paper.
Concerning future research, an interesting open question is how one can (theoretically or nu-
merically) determine whether a given arbitrary non-negative, square matrix is a tail-dependence
21
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
or Bernoulli-compatible matrix. To the best of our knowledge there are no corresponding algo-
rithms available. Another open question concerns the compatibility of other matrices of pairwise
measures of association such as rank-correlation measures (e.g., Spearman’s rho or Kendall’s
tau); see (Embrechts et al., 2002, Section 6.2). Recently, Fiebig et al. (2014) and Strokorb et al.
(2015) studied the concept of tail-dependence functions of stochastic processes. Similar results
to some of our findings were found in the context of max-stable processes.
From a practitioner’s point-of-view, it is important to point out limitations of using tail-
dependence matrices in quantitative risk management and other applications. One possible such
limitation is the statistical estimation of tail-dependence matrices since, as limits, estimating tail
dependence coefficients from data is non-trivial (and typically more complicated than estimation
in the body of a bivariate distribution).
After presenting the results of our paper at the conferences “Recent Developments in De-
pendence Modelling with Applications in Finance and Insurance – 2nd Edition, Brussels, May
29, 2015” and “The 9th International Conference on Extreme Value Analysis, Ann Arbor, June
15–19, 2015”, the references Fiebig et al. (2014) and Strokorb et al. (2015) were brought to our
attention (see also Acknowledgments below). In these papers, a very related problem is treated,
be it from a different, more theoretical angle, mainly based on the theory of max-stable and
Tawn-Molchanov processes as well as results for convex-polytopes. For instance, our Theorem
3.3 is similar to Theorem 6 c) in Fiebig et al. (2014).
Acknowledgments
We would like to thank Federico Degen for raising this interesting question concerning tail-
dependence matrices and Janusz Milek for relevant discussions. We would also like to thank the
Editor, two anonymous referees, Tiantian Mao, Don McLeish, Johan Segers, Kirstin Strokorb and
Yuwei Zhao for valuable input on an early version of the paper. Furthermore, Paul Embrechts
acknowledges financial support by the Swiss Finance Institute, and Marius Hofert and Ruodu
Wang acknowledge support from the Natural Sciences and Engineering Research Council of
Canada (NSERC Discovery Grant numbers 5010 and 435844, respectively). Ruodu Wang is
grateful to the Forschungsinstitut fur Mathematik (FIM) at ETH Zurich for financial support
during his visits in 2013 and 2014 when this topic was brought to his attention.
References
Berman, A. and Shaked-Monderer, N. (2003). Completely Positive Matrices. World Scientific.
22
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
Bluhm, C. and Overbeck, L. (2006). Structured Credit Portfolio Analysis, Baskets & CDOs.
Chapman & Hall.
Bluhm, C., Overbeck, L. and Wagner, C. (2002). An Introduction to Credit Risk Modeling.
Chapman & Hall.
Chaganty, N. R. and Joe, H. (2006). Range of correlation matrices for dependent Bernoulli
random variables. Biometrika, 93(1), 197–206.
Dhaene, J. and Denuit, M. (1999). The safest dependence structure among risks. Insurance:
Mathematics and Economics, 25(1), 11–21.
Embrechts, P., McNeil, A. and Straumann, D. (2002). Correlation and dependence in risk man-
agement: properties and pitfalls. In: Risk Management: Value at Risk and Beyond, ed. M.A.H.
Dempster, Cambridge University Press, Cambridge, 176–223.
Fiebig, U., Strokorb, K. and Schlather, M. (2014). The realization problem for tail correlation
functions. arXiv : http://arxiv.org/abs/1405.6876.
Joe, H. (1997). Multivariate Models and Dependence Concepts. Chapman & Hall.
Joe, H. (2014). Dependence Modeling with Copulas. London: Chapman & Hall.
Kortschak, D. and Albrecher, H. (2009). Asymptotic Results for the Sum of Dependent Non-
identically Distributed Random Variables. Methodology and Computing in Applied Probability,
11, 279–306.
Liebscher, E. (2008). Construction of asymmetric multivariate copulas. Journal of Multivariate
Analysis, 99(10), 2234–2250.
McNeil, A. J., Frey, R. and Embrechts, P. (2015). Quantitative Risk Management: Concepts,
Techniques and Tools. Revised Edition. Princeton University Press.
Nikoloulopoulos, A. K., Joe, H. and Li, H. (2009). Extreme value properties of multivariate t
copulas. Extremes, 12(2), 129–148.
Ruschendorf, L. (1981). Characterization of dependence concepts in normal distributions. Annals
of the Institute of Statistical Mathematics, 33(1), 347–359.
Strokorb, K., Ballani, F. and Schlather, M. (2015). Tail correlation functions of max-stable
processes. Extremes, 18(2), 241–271.
Yang, J., Qi, Y. and Wang, R. (2009). A class of multivariate copulas with bivariate Frechet
marginal copulas. Insurance: Mathematics and Economics, 45(1), 139–147.
23