Some multivariate imprecise shock model copulas
David Dolžan, Damjana Kokol Bukovšek, Matjaž Omladič, Damjan Škulj
SSOME MULTIVARIATE IMPRECISE SHOCK MODELCOPULAS
DAVID DOLˇZAN, DAMJANA KOKOL BUKOVˇSEK, MATJAˇZ OMLADI ˇC,AND DAMJAN ˇSKULJ
Abstract
Bivariate imprecise copulas have recently attracted substantial attention.However, the multivariate case seems still to be a “blank slate”. It is thennatural that this idea be tested first on shock model induced copulas, a fam-ily which might be the most useful in various applications. We investigate amodel in which some of the shocks are assumed imprecise and develop thecorresponding set of copulas. In the Marshall’s case we get a coherent set ofdistributions and a coherent set of copulas, where the bounds are naturallycorresponding to each other. The situation with the other two groups ofmultivariate imprecise shock model induced copulas, i.e., the maxmin andthe the reflected maxmin (RMM) copulas, is substantially more involved,but we are still able to produce their properties. These are the main resultsof the paper that serves as the first step into a theory that should developin this direction. In addition, we unfold the theory of bivariate impreciseRMM copulas that has not yet been done before. Introduction
Copulas arising from shock models in the presence of probabilisticuncertainty, which means that probability distributions are not neces-sarily precisely known, have been proposed for the first time by Omladiˇcand ˇSkulj [25] in bivariate setting. The main purpose of this paper isto present some extensions of the results presented there including anexpansion of these notions to the multivariate case.Copulas have been introduced in the precise setting by A. Sklar[32], who considered copulas as functions C ( u ) = C ( u , u , . . . , u n ) Mathematics Subject Classification.
Primary: 62H05, 60A86, Secondary:62H86.
Key words and phrases. imprecise probability; shock model; Marshall’s copula;maxmin copula; reflected maxmin copula.Damjan ˇSkulj acknowledges the financial support from the Slovenian ResearchAgency (research core funding No. P5-0168). David Dolˇzan, Damjana KokolBukovˇsek, and Matjaˇz Omladiˇc acknowledge financial support from the SlovenianResearch Agency (research core funding No. P1-0222). a r X i v : . [ m a t h . P R ] A ug D. DOLˇZAN, D. KOKOL BUKOVˇSEK, M. OMLADI ˇC, AND D. ˇSKULJ satisfying certain conditions. They can be defined equivalently asjoint distribution functions of random vectors with uniform marginaldistributions. He proved a two-way theorem: firstly, given a ran-dom vector X = ( X , X , . . . , X n ) with a vector of marginal proba-bility distributions F = ( F , F , . . . , F n ) and a copula C , the function C ( F ( x )) = C ( F ( x ) , F ( x ) , . . . , F n ( x n )) is a joint distribution of therandom vector X having distributions F as its marginals. Secondly,given a random vector X with joint distribution H ( x ) there exists acopula C ( u ) such that H ( x ) = C ( F ( x )), where F is the vector of themarginal distribution functions of the respective random variables X .Since then, copula models have become popular in various applicationsin view of their ability to describe the relationships among randomvariables in a flexible way and several families of copulas have beenintroduced to this end, motivated by specific needs from the scientificpractice (cf. [11, 20, 6]).Among the first widely studied and applied families of copulas werethe ones arising in shock models. They appear naturally as models ofjoint distributions for random variables representing lifetimes of com-ponents affected by shocks. Two types of shocks are usually consideredin these models, the first type only affects each one of the componentsseparately (the idiosyncratic shocks), while the second one simultane-ously affects all the components (the exogenous shock). In the originalMarshall’s case (cf. [15] based on an earlier work of Marshall and Olkin[14]) both types of shocks cause the component to cease to work im-mediately. Recently a new family of shock induced copulas has beenproposed by Omladiˇc and Ruˇzi´c [22] where the exogenous, i.e., sys-temic, shock has a detrimental effect on some of the components anda beneficial effect on the other ones. A third type of shock modelinduced copulas was introduced by Koˇsir and Omladiˇc [12], the re-flected maxmin copulas, RMM for short. Actually, the two papersintroduce the bivariate version for independent shocks, while an ex-tension to somewhat more general multivariate setting is presented inpapers [8, 13].Quantitative modeling of uncertainty is traditionally based on theuse of precise probabilities: for each event A , a single probability P ( A )is assigned, universally accepted to satisfy Kolmogorov’s axioms. Therehave been many successful applications of this concept, but also somecriticism. The requirement that P be σ -additive should be replaced,as some believe, by a more realistic requirement that it be additive.A more flexible theory of uncertainty that has evolved is the conceptof imprecise probabilities. For an event A , the lower probability P ( A )can informally be interpreted as reflecting the evidence certainly in OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 3 favour of the event A , while the upper probability P ( A ) reflects allevidence possibly in favour of A . So, the imprecise probability of A may be seen as the set of values lying between the two extremes. Acomprehensive study of this notion started by Walley [38], while morerecent development in the area can be found in [1]. It is natural toassume probabilities in these considerations to be finitely additive andnot necessarily σ -additive. The imprecise distribution of a randomvariable then consists of the interval of all distributions between a lowerbound F and an upper bound F ; this set is called a probability box , a p -box for short [9, 35].One can find numerous arguments for imprecision, such as scarcityof available information, costs connected to acquiring precise inputsor even inherent uncertainty related to phenomena under consider-ation. Ignoring imprecision may lead to deceptive conclusions andconsequentially to harmful decisions, especially if the conclusions arebacked by seemingly precise outputs. Methods of imprecise proba-bilities have been applied to various areas of probabilistic modeling,such as stochastic processes [5, 33], game theory [16, 19], reliabilitytheory [2, 21, 36, 39], decision theory [10, 17, 34], financial risk the-ory [26, 37], and others. Perhaps the first application of the theoryof copulas to models of imprecise probabilities has been proposed bySchmelzer [29, 30, 31].A possible definition of an imprecise bivariate p -box was given in Pe-lessoni et al. [28] thus raising the question of the corresponding Sklartype theorem. The first move in this direction was made by Monteset al. [18] proving one half of the imprecise Sklar’s theorem using thedefinition of bivariate p -box introduced in [28]. The same authors in-troduce in an earlier paper [27] an imprecise copula as an interval ofquasi-copulas satisfying certain axioms. The four authors propose acoherence question in these papers that is answered in the negativeby Omladiˇc and Stopar [23]; continuing their work in [24], the sameauthors give a full scale Sklar’s theorem in the bivariate imprecise set-ting using a slightly different notion than the bivariate p -box of [28],i.e., what they call a restricted bivariate p -box . Perhaps an even moreimportant result there is [24, Theorem 4] saying that if a joint distri-bution function emerges on a finitely additive probability space, theresulting copula exists and may be chosen so that it satisfies the usualSklar’s axioms. So, all the possible problems that may arise from re-laxing the Kolmogorov’s σ -additivity axiom, stay exclusively in theunivariate marginal distributions. D. DOLˇZAN, D. KOKOL BUKOVˇSEK, M. OMLADI ˇC, AND D. ˇSKULJ
As explained earlier, the main contribution of this paper is on themultivariate level, where we extend to the imprecise setting all the threetypes of shock model induced copulas: Marshall’s, maxmin and RMM.For the bivariate case the Marshall’s and the maxmin copulas have beenfirst presented in [25], while for the RMM copulas this has not beendone yet, so we have to do it first. The paper is organized as follows.In Section 2 we revisit some information on the imprecise distributionsand p -boxes and in Section 3 we present some details on copulas andshock model induced copulas. Section 4 brings facts on the two knownfamilies of bivariate shock model induced copulas, the Marshall’s andthe maxmin ones. Section 5 unfolds our first main result – the impreciseversion of the bivariate reflected maxmin copulas. In Sections 6, 7, and8 we give our multivariate extensions of the imprecise shock modelbased copulas, namely the Marshall’s, the maxmin, respectively theRMM copulas.2. Theory of imprecise probabilities revisited
Coherent lower and upper probabilities.
We first introducebriefly the basic concepts and ideas of imprecise probability models.For a detailed treatment, the reader is referred to [1, 38]. Let Ω bea possibility space, and A a collection of its subsets, called events .Usually we assume A to be an algebra, but not necessarily a σ -algebra.The concept of precise probability on the measurable space (Ω , A )can be generalised by allowing probabilities of events in A to be givenin terms of intervals [ P ( A ) , P ( A )] rather than precise values. The func-tions P (cid:54) P are mapping events to their lower and upper probabilitybounds and are respectively called lower and upper probabilities . If A is an algebra, then the following conjugacy relation between lower andupper probabilities is usually required:(1) P ( A ) = 1 − P ( A c ) for every A ∈ A . To every pair of lower and upper probabilities P and P we can alsoassociate the set M = { P : P is a finitely additive probability on A , P (cid:54) P (cid:54) P } . It is clear from the above that the set M is non-empty only if P (cid:54) P .Another central question regarding a pair of lower and upper prob-abilities is whether the bounds are pointwise limits of the elements in M : P ( A ) = inf P ∈M P ( A ) , P ( A ) = sup P ∈M P ( A ) for every A ∈ A . OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 5
If the above conditions are satisfied, P and P are said to be coherent lower and upper probabilities respectively. In the case of coherence,the conjugacy Condition (1) is automatically fulfilled, which means inparticular that if a lower probability P is coherent, then it uniquelydetermines the corresponding upper probability. A simple character-ization of coherence in terms of the properties of P and P does notseem to be known in the literature.Instead of the full structure of probability spaces, we are often con-cerned only with the distribution functions of specific random vari-ables. The set of relevant events where the probabilities have to begiven then shrinks considerably. In the precise case, a single distri-bution function F describes the distribution of a random variable X ,which gives the probabilities of the events of the form { X (cid:54) x } . Thus F ( x ) = P ( X (cid:54) x ) for x ∈ R where R = R ∪ {−∞ , + ∞} . Some-times we will also consider the corresponding survival function , whichwe will denote by (cid:98) F ( x ) = 1 − F ( x ) = P ( X > x ), which is decreas-ing and positive. (In the copula theory literature it is more usual todenote the survival function of F by F , but we will reserve this nota-tion for a different meaning.) Notice that the same operator (cid:98) · sends asurvival function back to its distribution function. Observe also thatin the standard probability theory distribution functions are cadlag,i.e., continuous from the right, and survival functions are caglad, i.e.,continuous from the left, while in the finitely additive approach theonly property a distribution function has is monotone increasing, andsurvival function is only monotone decreasing.2.2. Bivariate p -boxes. In the imprecise case, the probabilities ofthe above form are replaced by the corresponding lower (and upper)probabilities, resulting in sets of distribution functions called p -boxes[9, 35]. A p -box is a pair ( F , F ) of distribution functions with F (cid:54) F ,where F ( x ) = P ( X (cid:54) x ) and F ( x ) = P ( X (cid:54) x ). To every p -box weassociate the set of all distribution functions with the values betweenthe bounds: F ( F ,F ) = { F : F is a distribution function , F (cid:54) F (cid:54) F } . Clearly, F is a convex set of distribution functions. Conversely, sincesupremum and infimum of any set of distribution functions are them-selves distribution functions, every set of distribution functions gener-ates a p -box containing the original set.In the theory of imprecise probabilities, precise probability denotesa probability measure that is finitely additive, and not necessarily σ -additive, as is the case in most models using classical probabilities. D. DOLˇZAN, D. KOKOL BUKOVˇSEK, M. OMLADI ˇC, AND D. ˇSKULJ
The general theory of bivariate p -boxes is relatively new [18, 28]. Amapping F : R × R → [0 ,
1] is called standardized if(i) it is componentwise increasing: F ( x , y ) (cid:54) F ( x , y ) and F ( x, y ) (cid:54) F ( x, y ) whenever x (cid:54) x and y (cid:54) y and for all x, y ∈ R ;(ii) F ( −∞ , y ) = F ( x, −∞ ) = 0 for every x, y ∈ R ;(iii) F ( ∞ , ∞ ) = 1.If in addition,(iv) F ( x , y ) − F ( x , y ) − F ( x , y )+ F ( x , y ) (cid:62) x (cid:54) x and y (cid:54) y ,then it is called a bivariate distribution function .A pair ( F , F ) of standardized functions, where F (cid:54) F , is called a bivariate p -box .Observe that (1) neither the infimum nor supremum of a set of bivari-ate distribution functions need to be a bivariate distribution function;(2) the set F ( F ,F ) = { F : R × R → [0 , , F is a bivariate distribution function , F (cid:54) F (cid:54) F } may be empty in general. If it is not empty and its pointwise infimumand supremum equals F and F respectively, then this bivariate p -boxis said to be coherent .2.3. Independent random variables.
In the case where probabilitydistributions are known imprecisely, several distinct concepts of inde-pendence exist, such as epistemic irrelevance, epistemic independence and strong independence (see e.g. [3, 4]). However, as long as p -boxesare concerned, all these notions result in the factorization property , (cf.[18]): let a pair of p -boxes ( F X , F X ) and ( F Y , F Y ) correspond to thedistributions of random variables X and Y . The bivariate p -box ( F , F )is factorizing if F ( x, y ) = F X ( x ) F Y ( y ) ,F ( x, y ) = F X ( x ) F Y ( y ) . Thus a bivariate p -box corresponding to the bivariate distribution of apair of independent random variables is factorizing, regardless of thetype of independence.3. Copulas and Shock model copulas revisited
Copulas.
Copulas present a very convenient tool for modelingdependence of random variables free of their marginal distributions –only when one inserts these distributions into a copula, it becomes
OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 7 a joint distribution. A function C : [0 , × [0 , → [0 ,
1] is called a(bivariate) copula if it satisfies the following conditions:(C1) C ( u,
0) = C (0 , v ) = 0 for every u, v ∈ [0 , C ( u,
1) = u and C (1 , v ) = v for every u, v ∈ [0 , C ( u , v ) − C ( u , v ) − C ( u , v ) + C ( u , v ) (cid:62) (cid:54) u (cid:54) u (cid:54) (cid:54) v (cid:54) v (cid:54) Theorem 1 (Sklar’s theorem[32]) . Let F : R × R → [0 , be a bivariatedistribution function with marginals F X and F Y . Then there exists acopula C such that (2) F ( x, y ) = C ( F X ( x ) , F Y ( y )) for all ( x, y ) ∈ R × R ; and conversely, given any copula C and a pair of distribution functions F X and F Y , equation (2) defines a bivariate distribution function. It is our goal to show how some important classes of copulas can beextended to the case of imprecise probability models. Our constructionwill spread to the general multivariate case.3.2.
Marshall’s copulas revisited.
Copulas of the form C M ϕ,ψ ( u, v ) = uv min (cid:26) ϕ ( u ) u , ψ ( v ) v (cid:27) if uv > uv = 0 , where(P1) ϕ and ψ are two increasing real valued maps on [0 , ϕ (0) = ψ (0) = 0 and ϕ (1) = ψ (1) = 1;(P3) ϕ ∗ ( u ) = ϕ ( u ) u : (0 , → [1 , ∞ ] and ψ ∗ ( v ) = ψ ( v ) v : (0 , → [1 , ∞ ]are decreasing,were first introduced in [15] and are called (bivariate) Marshall’s cop-ulas . They were historically the first shock model induced copulas.There is an alternative way for writing down this definition which isbetter for generalizing it to more than 2 dimensions, i.e., C M ϕ,ψ ( u, v ) = ϕ ( u ) ψ ( v ) min (cid:26) uϕ ( u ) , vψ ( v ) (cid:27) if ϕ ( u ) ψ ( v ) > . Observe that this definition is equivalent to the previous one as maybe seen via a straightforward consideration. Here is the stochastic
D. DOLˇZAN, D. KOKOL BUKOVˇSEK, M. OMLADI ˇC, AND D. ˇSKULJ interpretation of these copulas and their generating functions ϕ and ψ emerging from [15]. Proposition 2.
Let
X, Y, Z be independent random variables with cor-responding distribution functions F X , F Y and F Z . Define U = max { X, Z } and V = max { Y, Z } and let F U and F V denote their respective distri-bution functions. Furthermore, let H be the bivariate joint distributionfunction of the pair ( U, V ) . Then:(i) F U = F X F Z and F V = F Y F Z .(ii) A pair of functions ϕ and ψ satisfying (P1)–(P3) exists, so that F X ( x ) = ϕ ( F U ( x )) for all x , where F U ( x ) > , and F Y ( y ) = ψ ( F V ( y )) for all y , where F V ( y ) > .(iii) H ( x, y ) = C M ϕ,ψ ( F U ( x ) , F V ( y )) .(iv) ϕ ∗ ◦ F U = ψ ∗ ◦ F V .(v) F Z = F U F X = F U ϕ ( F U ) = F V F Y = F V ψ ( F V ) , where the expressions aredefined. It turns out that the generating functions are necessarily continuouson the interval (0 ,
1] (cf. [22]); however, they are not uniquely deter-mined with the condition (ii) .These copulas were extended to the imprecise setting in [25] for thebivariate case and we will extend this further to the multivariate case.3.3.
Maxmin copulas revisited.
Another family of shock model in-duced copulas are the so called maxmin copulas introduced recently byOmladiˇc and Ruˇzi´c [22]. A maxmin copula depends on two functions ϕ and χ : [0 , → [0 , ϕ (0) = χ (0) = 0 and ϕ (1) = χ (1) = 1;(F2) ϕ and χ are increasing;(F3) ϕ ∗ ( u ) = ϕ ( u ) u : (0 , → [1 , ∞ ] and χ ∗ ( v ) : [0 , → [1 , ∞ ] aredecreasing, where χ ∗ ( v ) = − χ ( v ) v − χ ( v ) if v (cid:54) = χ ( v );+ ∞ if v = χ ( v ) (cid:54) = 1;1 if v = 1 . A maxmin copula is a map C MM : [0 , × [0 , → [0 ,
1] defined by C MM ϕ,χ ( u, v ) = uv + min { u (1 − v ) , ( ϕ ( u ) − u )( v − χ ( v )) } . Here is the stochastic interpretation of these copulas and of functions ϕ and χ . (Observe that the random variable U in the following propo-sition is the same as in Proposition 2.) OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 9
Proposition 3.
Let independent random variables
X, Y and Z begiven with respective distribution functions F X , F Y and F Z . Define U = max { X, Z } and W = min { Y, Z } and let F U , F W denote the dis-tribution functions of U and W respectively. Let H be the joint distri-bution function of ( U, W ) . Then:(i) F U = F X F Z and F W = F Y + F Z − F Y F Z .(ii) A pair of functions ϕ and χ satisfying (F1)–(F3) exists, so that F X ( x ) = ϕ ( F U ( x )) for all x , where F U ( x ) > and F Y ( y ) = χ ( F W ( y )) for all y , where F W ( y ) < .(iii) ϕ ∗ ◦ F U = χ ∗ ◦ F W .(iv) In terms of survival functions instead of distribution functions,the second equation in (i) assumes the following equivalent form (cid:98) F W = (cid:98) F Y (cid:98) F Z .(v) H ( x, y ) = C MM ϕ,χ ( F U ( x ) , F W ( y )) . Observe that, as in the first paragraph after Proposition 2, functions ϕ and χ are necessarily continuous, although not unique. Note also thatthe roles of generating functions ϕ in Marshall’s and maxmin modelsare equivalent, while the roles of ψ and χ may be seen opposite in somesense.Marshall and maxmin copulas were also extended to the imprecisesetting in [25] for the bivariate case. We extend them further to themultivariate case in Sections 6 and 7.3.4. Reflected maxmin copulas revisited.
Let
X, Y, Z be indepen-dent variables with probability distribution functions F X , F Y , and F Z respectively. Define U = max { X, Z } and W = min { Y, Z } as in Sub-section 3.3. Recall the definition of a survival function from Subsection2.1. So, we have F U = F X F Z , and (cid:98) F W = (cid:98) F Y (cid:98) F Z or F W = (cid:92) (cid:98) F Y (cid:98) F Z . From Subsection 3.3 we recall the existence of functions ϕ and χ suchthat ϕ ( F U ) = F X , and χ ( F W ) = F Y , whenever F U > F W <
1, sothat(3) ϕ ( F X F Z ) = F X and χ (cid:18) (cid:92) (cid:98) F Y (cid:98) F Z (cid:19) = F Y . Rewrite functions ϕ and χ into(4) f ( x ) = ϕ ( x ) − x, g ( x ) = 1 − x − χ (1 − x )to get f ( F U ) = f ( F X F Z ) = ϕ ( F U ) − F X F Z = F X − F X F Z = F X (cid:98) F Z if F U > g ( (cid:98) F W ) = F W − χ ( F W ) = F Y + F Z − F Y F Z − F Y = (cid:98) F Y F Z if F W <
1, which summarizes into(5) f ( F X F Z ) = F X (cid:98) F Z and g ( (cid:98) F Y (cid:98) F Z ) = (cid:98) F Y F Z , whenever F X F Z > (cid:98) F Y (cid:98) F Z > y (cid:55)→ − y that turns a general copula C ( x, y ) into y − C ( x, − y ) isused, together with the replacement of the generators ϕ and χ with f and g to transform the class of maxmin copulas into the class of whatthe authors call reflected maxmin copulas , RMM for short. They alsoprove ([12, Lemmas 1&2]): Claim.
Conditions (F1)–(F3) are satisfied for the original generat-ing functions of the maxmin copula if and only if the following condi-tions are satisfied by functions f and g :(G1) f (0) = g (0) = 0 , f (1) = g (1) = 0 , f ∗ (1) = g ∗ (1) = 0 ,(G2) the functions (cid:98) f ( u ) = f ( u )+ u and (cid:98) g ( w ) = g ( w )+ w are increasingon [0 , ,(G3) the functions f ∗ and g ∗ are decreasing on (0 , . Here we use the notation from [12] for the functions f ∗ ( u ) = f ( u ) u , g ∗ ( w ) = g ( w ) w , for u, w >
0. In addition we define f ∗ (0) = (cid:26) lim u ↓ f ( u ) u if it exists; ∞ otherwise,and similarly for g ∗ (0). Hence f, g : [0 , → [0 ,
1] and f ∗ , g ∗ : [0 , → [0 , ∞ ]. Also, using [12, Theorem 3], we know that the copula corre-sponding to the random vector ( U, W ) with respect to the distributionfunction of U and survival function of W is equal to(6) C RMM f,g ( x, y ) = max { , xy − f ( x ) g ( y ) } . So, clearly, the reflected maxmin copula is the copula obtained fromthe corresponding (maxmin) copula of the random vector (
U, W ) afterapplying the reflection on the second variable, and the functions f and g satisfying Conditions (G1)–(G3) are its generators. OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 11 Imprecise shock-model copulas revisited – Marshall’sand maxmin
Order relations generated by shock models.
The theory of p -boxes, univariate and bivariate, is based on the sets of probabilitydistributions that lie between the boundary distributions F and F . Inorder to transfer the theory of Marshall’s copulas from precise distri-bution functions to the more general case of p -boxes, the critical stepis to determine, whether the order on the set of distribution functionsimposed by p -boxes is preserved on the corresponding copulas. Asshown in [25], the order is indeed preserved, yet in different ways forMarshall’s and maxmin case.From Subsections 3.2 and 3.3 recall the triples of independent dis-tribution functions ( F X , F Y , F Z ) which give rise to the pairs of notnecessarily independent functions ( F U , F V ) in the Marshall’s case, and( F U , F W ) in the maxmin case; and then further to pairs of generatingfunctions ϕ, ψ and ϕ, χ , and to corresponding copulas C M ϕ,ψ and C MM ϕ,χ respectively. We follow [25] to introduce imprecision in these models.We allow F X and F Y to be imprecise, while for technical reasons F Z is assumed precise. Replace F X and F Y with p -boxes ( F X , F X ) and( F Y , F Y ). So, we consider triples ( F X , F Y , F Z ) where F X ≤ F X ≤ F X and F Y ≤ F Y ≤ F Y , or in p -box notation F X ∈ F ( F X ,F X ) and F Y ∈F ( F Y ,F Y ) .We now relate respective copulas C M ϕ,ψ and C MM ϕ,χ to the triple ( F X , F Y , F Z )via distribution functions F U , F V and F W . Let pairs of generating func-tions ϕ, ψ , and ϕ, χ , all mapping [0 , → [0 , ϕ ( F U ) = F X , ψ ( F V ) = F Y and χ ( F W ) = F Y whenever F U > , F V > , F W < ϕ, ψ, χ ) is associated to the triple( F X , F Y , F Z ). In this case we will also say that any of the generat-ing functions ϕ, ψ , or χ is associated to the triple ( F X , F Y , F Z ). Now,these conditions determine the generating functions only on the imagesof the corresponding distribution functions and with these functionschanging within their p -boxes we have to adjust the appropriate exten-sions so that the required order relations are satisfied. For instance, if( F X , F Y , F Z ) and ( F (cid:48) X , F (cid:48) Y , F Z ) are two respective triples with F (cid:48) X ≤ F X and F (cid:48) Y ≤ F Y , relations ϕ (cid:48) ≤ ϕ, ψ (cid:48) ≤ ψ and χ (cid:48) ≤ χ will not be satisfiednecessarily for any pairs of triples of generating functions ( ϕ, ψ, χ ) and( ϕ (cid:48) , ψ (cid:48) , χ (cid:48) ) associated with them. Rather surprisingly, it is possible tofind explicit formulas for the extensions that do preserve the order. Wepresent here the solution of this problem from [25]. Denote by f ( x +), respectively f ( x − ), the right limit, respectivelythe left limit of a monotone (increasing) function f at x ; note that thelimits exist because f is monotone. For distribution functions F X and F Z let F U = F X F Z . Choose a u ∈ (0 ,
1) and let x be any value suchthat F U ( x − ) (cid:54) u (cid:54) F U ( x +). Define(7) ϕ ( u ) = u = 0; F X ( x − ) if u − (cid:54) u (cid:54) u l ; uF Z ( x ) if u l (cid:54) u (cid:54) u u ; F X ( x +) if u u (cid:54) u (cid:54) u + ;1 if u = 1 , where u − = F X ( x − ) F Z ( x − ) = F U ( x − ) , u l = F X ( x − ) F Z ( x ) ,u + = F X ( x +) F Z ( x +) = F U ( x +) , u u = F X ( x +) F Z ( x ) . Furthermore, choose a v ∈ (0 ,
1) and let y be any value such that F W ( y − ) (cid:54) v (cid:54) F W ( y +). The extension of χ at v is defined asfollows:(8) χ ( v ) = v = 0; F Y ( y − ) if v − (cid:54) v (cid:54) v l ; v − F Z ( y )1 − F Z ( y ) if v l (cid:54) v (cid:54) v u ; F Y ( y +) if v u (cid:54) v (cid:54) v + ;1 if v = 1 , where v − = F Y ( y − ) + F Z ( y − ) − F Y ( y − ) F Z ( y − ) = F W ( y − ) ,v l = F Y ( y − ) + F Z ( y ) − F Y ( y − ) F Z ( y ) ,v + = F Y ( y +) + F Z ( y +) − F Y ( y +) F Z ( y +) = F W ( y +) ,v u = F Y ( y − ) + F Z ( y ) − F Y ( y − ) F Z ( y ) . The generating functions obtained using the extension (7) for ϕ , andappropriately adjusted for ψ , and the extension (8) for χ , are associatedwith the triple ( F X , F Y , F Z ). Moreover, the following lemmas hold (cf.[25]). Lemma 4.
Let F (cid:48) X (cid:54) F X and F Z be given, and let F U = F X F Z and F (cid:48) U = F (cid:48) X F Z . Then ϕ (cid:48) (cid:54) ϕ , where ϕ (cid:48) and ϕ are defined by applying (7) to F (cid:48) X and F X respectively. OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 13
Lemma 5.
Let F (cid:48) Y (cid:54) F Y and F Z be given, and let F W = F Y + F Z − F Y F Z and F (cid:48) W = F (cid:48) Y + F (cid:48) Z − F (cid:48) Y F (cid:48) Z . Then χ (cid:48) (cid:54) χ , where χ (cid:48) and χ aredefined by applying (8) to F (cid:48) Y and F Y respectively. Imprecise Marshall’s copulas and imprecise maxmin cop-ulas.
Based on the results in the previous subsection, we can nowdefine the imprecise version of the Marshall’s and maxmin copulas.The family of copulas(9) C M = { C M ϕ,ψ : ϕ (cid:54) ϕ (cid:54) ϕ, ψ (cid:54) ψ (cid:54) ψ } , where ϕ (cid:54) ϕ and ψ (cid:54) ψ , and all ϕ and ψ , including the bounds,satisfy Conditions (P1)–(P3), is called an imprecise Marshall’s copula .The family of copulas(10) C MM = { C MM ϕ,χ : ϕ (cid:54) ϕ (cid:54) ϕ, χ (cid:54) χ (cid:54) χ } , where ϕ (cid:54) ϕ and χ (cid:54) χ , and all ϕ and χ , including the bounds, satisfyConditions (F1)–(F3), is called an imprecise maxmin copula .Here is the stochastic interpretation of the imprecise shock modelcopulas. Let X and Y be random variables, whose distributions aregiven in terms of p -boxes ( F X , F X ) and ( F Y , F Y ), and Z a random vari-able with a precise distribution function F Z . To every triple ( F X , F Y , F Z )where F X ∈ F ( F X ,F X ) and F Y ∈ F ( F Y ,F Y ) , there exists a Marshall’scopula C M ϕ,ψ with generating functions ϕ, ψ given by (7), such that F U and F V are the respective distribution functions of random variables U = max { X, Z } and V = max { Y, Z } , and C M ϕ,ψ ( F U , F V ) is their jointdistribution function. Next, we will denote the minimal generatingfunctions associated to the triple ( F X , F Y , F Z ) by ϕ and ψ , and thecorresponding maximal generating functions associated to the triple( F X , F Y , F Z ) by ϕ and ψ . (Note that ϕ, ψ, ϕ, ψ themselves are notnecessarily constructed by Equation (7). The existence of these func-tions was proven in [25, Proposition 4].) Moreover, we will denoteby F U and F V the infimum of the distribution functions of U and V respectively, and by F U and F V the supremum of the distributionfunctions of U and V respectively.Similarly, for the maxmin case, given a triple ( F X , F Y , F Z ) where F X ∈ F ( F X ,F X ) and F Y ∈ F ( F Y ,F Y ) , there exists a maxmin copula C MM ϕ,χ with generating functions ϕ and χ given by (7) and (8), re-spectively, such that F U and F W are the respective distribution func-tions of random variables U = max { X, Z } and W = min { Y, Z } , and C MM ϕ,χ ( F U , F W ) is their joint distribution function. Next, let ϕ and χ bethe minimal generating functions associated to the triple ( F X , F Y , F Z ),not necessarily constructed by Equations (7)&(8). Also, let ϕ and χ be the maximal generating functions associated to the triple ( F X , F Y , F Z ),not necessarily constructed by Equations (7)&(8). The existence ofthese functions was likewise proven in [25, Proposition 4]. Finally, thesupremum and infimum of the distribution functions F U , F W will bedenoted by F U , F W and F U , F W respectively.The following theorems describe the properties of the imprecise Mar-shall’s and maxmin copulas. Recall that the definitions of functions ϕ ∗ , ϕ ∗ , ψ ∗ , ψ ∗ and χ ∗ , χ ∗ are exhibited in Conditions (P3) and (F3),respectively. Theorem 6 (Properties of imprecise Marshall’s copulas [25, Theorem3]) . In the situation described above we have:(i) ϕ (cid:54) ϕ and ψ (cid:54) ψ .(ii) C M ϕ,ψ (cid:54) C M ϕ,ψ (cid:54) C M ϕ,ψ , where C M ϕ,ψ is the Marshall’s copula corre-sponding to some triple ( F X , F Y , F Z ) , where F X ∈ F ( F X ,F X ) and F Y ∈ F ( F Y ,F Y ) .(iii) F U = F X F Z F V = F Y F Z F U = F X F Z F V = F Y F Z . (iv) F X ( x ) = ϕ ( F U ( x )) , if F U ( x ) > F X ( x ) = ϕ ( F U ( x )) , if F U ( x ) > F Y ( y ) = ψ ( F V ( y )) , if F V ( y ) > F Y ( y ) = ψ ( F V ( y )) , if F V ( y ) > . (v) ϕ ∗ ◦ F U = ψ ∗ ◦ F V = ϕ ∗ ◦ F U = ψ ∗ ◦ F V if F U , F V , F U , F V > .(vi) F U (cid:54) F U and F V (cid:54) F V .(vii) The distributions of the random variables U = max { X, Z } and V = max { Y, Z } are described with the p -boxes ( F U , F U ) and ( F V , F V ) respectively.(viii) C M ϕ,ψ ( F U , F V ) (cid:54) C M ϕ,ψ ( F U , F V ) .(ix) The joint distribution of ( U, V ) is described by a bivariate p -box ( H, H ) = ( C M ϕ,ψ ( F U , F V ) , C M ϕ,ψ ( F U , F V )) . Theorem 7 (Properties of imprecise maxmin copulas [25, Theorem4]) . In the above situation we have:(i) ϕ (cid:54) ϕ and χ (cid:54) χ .(ii) C MM ϕ,χ (cid:54) C MM ϕ,χ (cid:54) C MM ϕ,χ where C MM ϕ,χ is a maxmin copula corre-sponding to some triple ( F X , F Y , F Z ) , where F X ∈ F ( F X ,F X ) and F Y ∈ F ( F Y ,F Y ) . OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 15 (iii) F U = F X F Z , F W = F Y + F Z − F Y F Z ,F U = F X F Z , F W = F Y + F Z − F Y F Z . (iv) F X ( x ) = ϕ ( F U ( x )) , if F U ( x ) > F X ( x ) = ϕ ( F U ( x )) , if F U ( x ) > F Y ( y ) = χ ( F W ( y )) , if F W ( y ) < F Y ( y ) = χ ( F W ( x )) , if F W ( y ) < . (v) ϕ ∗ ◦ F U = χ ∗ ◦ F W = ϕ ∗ ◦ F U = χ ∗ ◦ F W if F U , F U > and F W , F W < .(vi) F U (cid:54) F U and F W (cid:54) F W .(vii) The distributions of the random variables U = max { X, Z } and W = min { Y, Z } are described with the p -boxes ( F U , F U ) and ( F W , F W ) respectively.(viii) C MM ϕ,χ ( F U , F W ) (cid:54) C MM ϕ,χ ( F U , F W ) .(ix) The joint distribution of ( U, W ) is described by a bivariate p -box (11) ( H, H ) = ( C MM ϕ,χ ( F U , F W ) , C MM ϕ,χ ( F U , F W )) . Bivariate imprecise reflected maxmin copulas
In this section we present one of our main results, the impreciseversion of the bivariate reflected maxmin copulas. Observe that re-flected maxmin copulas were first introduced in [12] as a simplificationof the maxmin copulas first introduced in [22], and they are revisitedin Subsection 3.4. The imprecise version of the maxmin copulas are in-troduced in [25] and revisited in Subsection 4.2. In case of a conflictingnotation of the two sources we prefer to use the notation of Subsection3.4.We assume that the distribution functions of X and Y are imprecise,i.e. they are all obtained via finitely additive measures and live in their p -boxes F X ∈ F ( F X ,F X ) and F Y ∈ F ( F Y ,F Y ) . However, for technical reasons we assume that the distribution func-tion F Z is precise. Following Subsection 3.4 we associate to any triple( F X , F Y , F Z ) the corresponding generating functions ϕ and χ satisfyingthe Conditions (F1)–(F3). Observe that the existence of the kind ofgenerating functions satisfying the defining Relations (3) and such thatorder relations on the respective generating functions ϕ and χ are inaccordance with order relations on the respective distribution functions F X and F Y was presented in Subsection 4.2 via Equations (7) and (8). We now transform these generating functions into functions f and g us-ing Equation (4). In this sense functions f and g are defined indirectlyvia Equations (7) and (8) and satisfy Conditions (G1)–(G3). Moreover,the order is preserved for functions f and reversed for functions g . Wewill say that they are associated to the triple ( F X , F Y , F Z ).Based on the results presented in Subsection 3.4 we can now definethe imprecise version of the reflected maxmin copulas. Following alsothe ideas presented in Subsection 4.2 we introduce the family of cop-ulas depending on two pairs of functions f (cid:54) f and g (cid:54) g satisfyingConditions (G1)–(G3). We let(12) C RMM = { C RMM f,g : f (cid:54) f (cid:54) f , g (cid:54) g (cid:54) g } , where all f and g also satisfy Conditions (G1)–(G3). We call this familyan imprecise reflected maxmin copula . We can expect a possible sto-chastic interpretation only if in this definition we let f be the infimumof all functions f associated to the triple ( F X , F Y , F Z ) and we let f bethe supremum of all functions f associated to the triple ( F X , F Y , F Z ).Furthermore, let g be the infimum of all functions g associated to thetriple ( F X , F Y , F Z ) and let g be the supremum of all functions g asso-ciated to the triple ( F X , F Y , F Z ). Note that g ( x ) = 1 − x − χ (1 − x )and g ( x ) = 1 − x − χ (1 − x ). Proposition 8.
For every F X ∈ F ( F X ,F X ) and every F Y ∈ F ( F Y ,F Y ) there exist functions f and g associated to the triple ( F X , F Y , F Z ) suchthat f (cid:54) f (cid:54) f and g (cid:54) g (cid:54) g .Proof. By Lemmas 4&5 we get ϕ (cid:54) ϕ and χ (cid:54) χ . Using (4) we geteasily the desired result for f and g . (cid:3) We need another fact, namely that operator (cid:98) · reverses the order.Let us combine all these facts in Equation (5) to get for the functions f , g, f ,and g, f ( F X F Z ) = F X (cid:98) F Z , f ( F X F Z ) = F X (cid:98) F Z ,g ( (cid:98) F Y (cid:98) F Z ) = (cid:98) F Y F Z , g ( (cid:98) F Y (cid:98) F Z ) = (cid:98) F Y F Z , (13)whenever F X F Z > F X F Z > (cid:98) F Y (cid:98) F Z > (cid:98) F Y (cid:98) F Z >
0. Here and inthe sequel we denote (cid:98) F = 1 − F and (cid:98) F = 1 − F . Out of these fourequations let us show, say, the southwest one. The others go similarly.When seeking the infimum of the left hand side of the second equationof (5), function g reaches g , the infimum of the value of (cid:98) F Y becomes (cid:98) F Y OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 17 and (cid:98) F Z remains unchanged; similar considerations apply to the righthand side of the equation.Before we summarize these observations let us also introduce H σ ( x, y ) = P ( U (cid:54) x, W > y ) which is playing the role of the combined jointdistribution-survival function we need in relation with RMM copulas. Theorem 9 (Properties of the imprecise RMM copulas) . It holds that:(i) f (cid:54) f , g (cid:54) g .(ii) C RMM f,g (cid:62) C RMM f,g (cid:62) C RMM f,g .(iii) F U = F X F Z , F U = F X F Z , (cid:98) F W = (cid:98) F Y (cid:98) F Z , (cid:98) F W = (cid:98) F Y (cid:98) F Z . (iv) F U (cid:54) F U (cid:54) F U , (cid:98) F W (cid:62) (cid:98) F W (cid:62) (cid:98) F W .(v) f ( F U ) = F X − F U , f ( F U ) = F X − F U ,g ( (cid:98) F W ) = (cid:98) F Y − (cid:98) F W , g ( (cid:98) F W ) = (cid:98) F Y − (cid:98) F W , whenever F U , F U , (cid:98) F W , (cid:98) F W > .(vi) f ∗ ( F U ) g ∗ ( (cid:98) F W ) = f ∗ ( F U ) g ∗ ( (cid:98) F W ) = f ∗ ( F U ) g ∗ ( (cid:98) F W ) = f ∗ ( F U ) g ∗ ( (cid:98) F W ) =1 , if F U , F U , (cid:98) F W , (cid:98) F W > .(vii) H σ ( x, y ) = F X ( x ) (cid:98) F Y ( y ) max { , F Z ( x ) − F Z ( y ) } = C RMM f,g ( F U ( x ) , (cid:98) F W ( y )) .(viii) H σ ( x, y ) = C RMM f,g ( F U ( x ) , (cid:98) F W ( y )) and H σ ( x, y ) = C RMM f,g ( F U ( x ) , (cid:98) F W ( y )) .(ix) C RMM f,g ( F U ( x ) , (cid:98) F W ( y )) (cid:54) C RMM f,g ( F U ( x ) , (cid:98) F W ( y )) (cid:54) C RMM f,g ( F U ( x ) , (cid:98) F W ( y )) .Proof. (i) follows by Proposition 8. (ii) : Choose arbitrary u, v ∈ [0 , f ( u ) (cid:54) f (cid:48) ( u ) and g ( v ) (cid:54) g (cid:48) ( v ) we geteasily f ( u ) g ( v ) (cid:54) f (cid:48) ( u ) g (cid:48) ( v ), so that uv − f ( u ) g ( v ) (cid:62) uv − f (cid:48) ( u ) g (cid:48) ( v ),consequently C RMM f,g (cid:62) C RMM f (cid:48) ,g (cid:48) , and the desired conclusion follows. (iii) :Clear. Now, we have F U = F X F Z by definition, so that F X (cid:54) F (cid:48) X implies F U = F X F Z (cid:54) F (cid:48) X F Z = F (cid:48) U yielding the first two relations of (iv) . The other two relations follow from the fact that F Y (cid:54) F (cid:48) Y implies (cid:98) F W = (cid:98) F Y (cid:98) F Z (cid:62) (cid:99) F (cid:48) Y (cid:98) F Z = (cid:99) F (cid:48) W . (v) : These are Relations (13) rewritten. (vi) : Follows directly by using (iii) and (v) . (vii) : Write H σ ( x, y ) = P (max( X, Z ) (cid:54) x, min( Y, Z ) > y )= P ( X (cid:54) x, Y > y, y < Z (cid:54) x )= F X ( x ) (cid:98) F Y ( y )( F Z ( x ) − F Z ( y ))(14) if x > y and zero otherwise. On the other hand C RMM f,g ( F U ( x ) , (cid:98) F W ( y )) = F X ( x ) F Z ( x ) (cid:98) F Y ( y ) (cid:98) F Z ( y ) − F X ( x ) (cid:98) F Z ( x ) (cid:98) F Y ( y ) F Z ( y )whenever this expression is positive, and zero otherwise; and this amountsto the same as in (14). (viii) : Compute infimum of the leftmost sideand on the rightmost side of (14) to get H σ ( x, y ) = F X ( x ) (cid:98) F Y ( y ) max { , F Z ( x ) − F Z ( y ) } which implies the first desired relation using (iii) and (v) . The secondone goes similarly. (ix) : Follows from (vii) and (viii) . (cid:3) Remark.
Observe that, somewhat surprisingly, Relation (ix) holdsin spite of the fact that C RMM f,g (cid:62) C RMM f,g as implied by (ii) . Example.
Suppose the occurrence of endogenous shocks in themodel is governed by independent Poisson processes and exogenousshock comes at a fixed future time. Then
X, Y and Z are independentrandom variables with distribution functions: F X ( x ) = 1 − e − λx , for x (cid:62) x < F Y ( y ) = 1 − e − µy , for y (cid:62) y < F Z ( x ) = (cid:40) x < x (cid:62) , where λ and µ are some positive constants, actually they are the pa-rameters of the underlying Poisson processes. We are normalizing theparameters so that shock Z comes at time 1. Further, the distributionfunctions of U = max { X, Z } and W = min { Y, Z } are equal to F U ( x ) = (cid:40) − e − λx if x (cid:62) F W ( y ) = y < − e − µy if 0 (cid:54) y < (cid:54) y. Reflected maxmin copula C RMM f,g modeling the dependence between U and W is generated by the functions f ( u ) = max { − e − λ − u, } ,g ( w ) = max { e − µ − w, } OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 19 for u, w ∈ (0 , f (0) = g (0) = 0. It is equal to C RMM f,g ( u, w ) = e − µ u + (1 − e − λ ) w (cid:54) e − µ (1 − e − λ ); e − µ u + (1 − e − λ ) w − e − µ (1 − e − λ ) if e − µ u + (1 − e − λ ) w > e − µ (1 − e − λ ) ,u (cid:54) − e − λ , w (cid:54) e − µ ; uw if u > − e − λ or w > e − µ . Suppose now that we cannot assume precisely given parameters, butinstead we consider the p -boxes ( F X , F X ) and ( F Y , F Y ), where F X ( x )is an exponential distribution with parameter λ and F X ( x ) with someparameter λ > λ . It is immediate that F X (cid:54) F X holds. Similarly,let F Y and F Y be exponential with parameters µ < µ respectively.It is easy to check that f ( u ) = max { − e − λ − u, } ,f ( u ) = max { − e − λ − u, } ,g ( w ) = max { e − µ − w, } ,g ( w ) = max { e − µ − w, } for u, w ∈ (0 ,
1] and 0 otherwise are the generating functions of copulas C RMM f,g and C RMM f,g . Notice that the order of functions g is reversed withrespect to the order of the parameters µ . In Figure 1 we give the 3Dgraphs of copulas C RMM f,g and C RMM f,g for the parameters λ = µ = 1and λ = µ = 2, where the relation C RMM f,g (cid:62) C RMM f,g can be seen.6.
Multivariate imprecise Marshall’s copulas
In this section we extend the bivariate imprecise Marshall’s cop-ulas described in Subsection 4.2 to the multivariate case. Start byrevisiting the precise case (see for example [7] and references therein).Let X , . . . , X n , Z be independent variables with respective distributionfunctions F , . . . , F n , F Z . Define(15) U i = max { X i , Z } , for i = 1 , . . . , n. Then for the respective distribution functions G , . . . , G n of U , . . . , U n we have clearly G i = F i F Z for i = 1 , , . . . , n. Denote by H ( x , . . . , x n ) the joint distribution function of the randomvector ( U , U , . . . , U n ). Following the notation of the bivariate casepresented in Subsection 3.2 we introduce the generating functions so Figure 1.
3D graphs of copulas C RMM f,g and C RMM f,g forthe parameters λ = 1 , µ = 2 (left) and λ = 2 , µ = 1(right).that they satisfy Conditions (F1)–(F3) and the defining relations ϕ i ( G i ) = F i if G i > i = 1 , , . . . , n. Note that these relations do not determine the generating functionsuniquely. Recall that the joint distribution function equals H ( x , . . . , x n ) = F ( x ) · · · F n ( x n ) F Z (min { x , . . . , x n } ) , and that(16) C ( u , . . . , u n ) = ϕ ( u ) · · · ϕ n ( u n ) min (cid:26) u ϕ ( u ) , . . . , u n ϕ n ( u n ) (cid:27) , if all ϕ i ( u i ) > C M ϕϕϕ for this copula. After a straightforward computation, one concludesthat C is a copula such that(17) H ( x , . . . , x n ) = C ( G ( x ) , . . . , G n ( x n )) . Observe that we have thus extended an alternative version of the usu-ally preferred Marshall’s formula from the bivariate case (we presentedboth versions in Subsection 3.2) to the n -variate case.In the imprecise setting we need to be more careful. From now on inthis subsection all our distribution functions are assumed to come froma finitely-additive probability space meaning that they are monotoneonly. So, the random variables representing endogenous shocks are as-sumed to be given by distributions F i ∈ F ( F i ,F i ) for i = 1 , , . . . , n , and OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 21 by the Marshall’s assumption (15) the random variables correspondingto the lives of the components satisfy G i = F i F Z for i = 1 , , . . . , n .For any choice of marginal distributions we define the joint distribu-tion function H ( x , . . . , x n ) = F ( x ) · · · F n ( x n ) F Z (min { x , . . . , x n } ) byanalogy with the above. This implies that the minimal and maximaljoint distribution functions H ( x , . . . , x n ) = min { H ( x , . . . , x n ) | F i ∈ F ( F i ,F i ) for i = 1 , , . . . , n } and H ( x , . . . , x n ) = max { H ( x , . . . , x n ) | F i ∈ F ( F i ,F i ) for i = 1 , , . . . , n } are clearly equal to H ( x , . . . , x n ) = F ( x ) · · · F n ( x n ) F Z (min { x , . . . , x n } ) and H ( x , . . . , x n ) = F ( x ) · · · F n ( x n ) F Z (min { x , . . . , x n } ) . (18)We define the corresponding generating functions ϕ i using Equation(7) in which we are consecutively replacing F X by F i for i = 1 , , . . . , n .By Proposition 2 (ii) we have ϕ i ( G i ) = F i , if G i > , for i = 1 , , . . . , n, in agreement with the defining relations above. By Lemma 4 we deducethat F (cid:48) i (cid:54) F i implies ϕ (cid:48) i (cid:54) ϕ i . Introduce the vectors ϕϕϕ = ( ϕ , ϕ , . . . , ϕ n ), ϕϕϕ = ( ϕ , ϕ , . . . , ϕ n ) and ϕϕϕ = ( ϕ , ϕ , . . . , ϕ n ). For the copula of Equation (16) introduce nota-tion C M ϕϕϕ ( uuu ) = n (cid:89) i =1 ϕ i ( u i ) min i =1 ,...,n (cid:26) u i ϕ i ( u i ) (cid:27) if ϕ i ( u i ) > i = 1 , , . . . , n , and zero otherwise. We rewrite alsothe Equation (17) into(19) H ( x , . . . , x n ) = C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) . Given two vectors of functions ϕϕϕ (cid:54) ϕϕϕ (here and in what follows therelation “less than or equal to” is meant componentwise) such thateach of their components satisfies Conditions (P1)–(P3), we let C M = { C M ϕϕϕ : ϕϕϕ (cid:54) ϕϕϕ (cid:54) ϕϕϕ } , where each component of ϕϕϕ also satisfies Conditions (P1)–(P3), andcall this set of copulas an n -variate imprecise Marshall’s copula .Using the ideas of Subsection 4.2 let ϕ i respectively ϕ i be the minimalrespectively the maximal function satisfying Conditions (P1)–(P3) and ϕ i ( G i ) = F i if G i > ϕ i ( G i ) = F i if G i > for i = 1 , , . . . , n . Define also functions ϕ i ∗ and ϕ i ∗ as in Condition(P3).Let us summarize. Theorem 10 (Properties of multivariate imprecise Marshall’s copu-las) . In the situation described above we have:(i) ϕϕϕ (cid:54) ϕϕϕ .(ii) C M ϕϕϕ (cid:54) C M ϕϕϕ (cid:54) C M ϕϕϕ .(iii) G i = F i F Z , G i = F i F Z , G i = F i F Z , for i = 1 , , . . . , n .(iv) ϕ i ( G i ) = F i if G i > ,ϕ i ( G i ) = F i if G i > ,ϕ i ( G i ) = F i if G i > , for i = 1 , , . . . , n .(v) C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) (cid:54) C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) (cid:54) C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) .(vi) H ( x , . . . , x n ) = C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) ,H ( x , . . . , x n ) = C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) ,H ( x , . . . , x n ) = C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) . (vii) For all i, j = 1 , . . . , n we have ϕ i ∗ ( G i ) = ϕ j ∗ ( G j ) if G i , G j > .Proof. (i) follows by the observations above, (ii) follows from (i) andEquation (16). (iii) and (iv) follow by definition. (v) First, C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) (cid:54) C M ϕϕϕ ( G ( x ) , . . . , G n ( x n ))since G i = F i F Z (cid:54) F i F Z = G i and C M ϕϕϕ is monotone. Next, by (ii) weget C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) (cid:54) C M ϕϕϕ ( G ( x ) , . . . , G n ( x n )) . The other side of the inequality goes similarly. Point (vi) follows fromEquations (17) and (18) and point (vii) follows from (iv) . (cid:3) OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 23 Multivariate imprecise maxmin copulas
In this section we extend the bivariate imprecise maxmin copulasdescribed in Subsection 4.2 to the multivariate case. Start again by re-visiting the precise case (see [8]). We let X , . . . , X p , X p +1 , . . . , X n , Z beindependent variables with respective distribution functions F , . . . , F p ,F p +1 , . . . , F n , F Z and define U i = max { X i , Z } , for i = 1 , . . . , p, and U j = min { X j , Z } , for j = p + 1 , . . . , n. (20)So, for the respective distribution functions G , . . . , G n of U , . . . , U n we have G i = F i F Z for i = 1 , . . . , p, and (cid:98) G j = (cid:98) F j (cid:98) F Z for j = p + 1 , . . . , n. Following the elaboration of the bivariate case presented in Subsec-tion 3.3 (and in particular Proposition 3) we deduce that this timethe defining relations of the generating functions, necessarily satisfyingConditions (F1)–(F3), are given by ϕ i ( G i ) = F i if G i > i = 1 , . . . , p, and χ j ( G j ) = F j if G j < j = p + 1 , . . . , n. For a vector ( x , . . . , x n ) ∈ R n define H ( x , . . . , x n ) = P ( U (cid:54) x , . . . , U n (cid:54) x n )and exploit Formula (4.3) of [8] to get for the independent case H ( x , . . . , x n ) = (cid:88) K ⊆ S (cid:89) i ∈ T ∪ ( S \ K ) F i ( x i ) × max (cid:26) , F Z (cid:18) min i ∈ T ∪ K x i (cid:19) − F Z (cid:18) max j ∈ S \ K x j (cid:19)(cid:27) , (21)where T = { , , . . . , p } , S = { p + 1 , p + 2 , . . . , n } , and K runs throughall the subsets of S . If we want to do the imprecise case, we need to havea deeper understanding of this formula. We first write H ( x , . . . , x n ) = P ( A ) where A = (cid:84) ni =1 A i and A i = ( U i (cid:54) x i ), for i ∈ T ∪ S . Moreover,for j ∈ S define B j = A c j = Ω \ A j so that B j = ( U j > x j ). It followsthat A = (cid:32)(cid:92) i ∈ T A i (cid:33) ∩ (cid:32)(cid:91) j ∈ S B j (cid:33) c , and by the inclusion-exclusion principle P ( A ) = P (cid:32)(cid:92) i ∈ T A i (cid:33) − P (cid:32)(cid:32)(cid:92) i ∈ T A i (cid:33) ∩ (cid:32)(cid:91) j ∈ S B j (cid:33)(cid:33) = P (cid:32)(cid:92) i ∈ T A i (cid:33) − (cid:88) ∅(cid:54) = K ⊆ S ( − | K | +1 P (cid:32)(cid:32)(cid:92) i ∈ T A i (cid:33) ∩ (cid:32) (cid:92) j ∈ K B j (cid:33)(cid:33) . Observe that A i = (max { X i , Z } (cid:54) x i ) = ( X i (cid:54) x i ) ∩ ( Z (cid:54) x i ) for i ∈ T and that B j = (min { X j , Z } > x j ) = ( X j > x j ) ∩ ( Z > x j ) for j ∈ S ,so that, using the independence assumption P ( A ) = (cid:89) i ∈ T F i ( x i ) P (cid:18) Z (cid:54) min i ∈ T x i (cid:19) − (cid:88) ∅(cid:54) = K ⊆ S ( − | K | +1 P (cid:32) (cid:92) j ∈ K B j (cid:33) = (cid:89) i ∈ T F i ( x i ) P (cid:18) Z (cid:54) min i ∈ T x i (cid:19) (cid:32) (cid:88) K ⊆ S ( − | K | P (cid:32) (cid:92) j ∈ K B j (cid:33)(cid:33) = (cid:89) i ∈ T F i ( x i ) (cid:32) (cid:88) K ⊆ S ( − | K | (cid:89) j ∈ K (cid:98) F j ( x j ) P (cid:18) max j ∈ K x j < Z (cid:54) min i ∈ T x i (cid:19)(cid:33) . By reordering variables X p +1 , . . . , X n , if necessary, let us order themembers of the set { x p +1 , . . . , x n } so that x p +1 (cid:54) x p +2 (cid:54) . . . (cid:54) x n . Wealso introduce y = min i ∈ T x i and choose α to be the largest index with p < α (cid:54) n and x α (cid:54) y . In case that y < x p +1 we choose α = p . For a K ⊆ S, K (cid:54) = ∅ , denote by r the greatest index contained in K and let K (cid:48) ⊆ { p + 1 , . . . , r − } be such that K = K (cid:48) ∪ { r } . If α < r then theterm of the above sum corresponding to K is zero, so that we can keeponly those terms for which α (cid:62) r . Therefore, P ( A ) = (cid:89) i ∈ T F i ( x i ) (cid:32) P ( Z (cid:54) y ) + α (cid:88) r = p +1 E r (cid:98) F r ( x r ) P ( x r < Z (cid:54) y ) (cid:33) , where E r = 1 if r = p + 1, and E r stands for the expression E r = (cid:88) K (cid:48) ⊆{ p +1 ,...,r − } ( − | K (cid:48) | +1 (cid:89) j ∈ K (cid:48) (cid:98) F j ( x j ) , otherwise. Clearly, this is an alternating sum of the elementary sym-metric polynomials which is known to be equal to E r = − r − (cid:89) j = p +1 (1 − (cid:98) F j ( x j )) = − r − (cid:89) j = p +1 F j ( x j ) , OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 25 so that P ( A ) = (cid:89) i ∈ T F i ( x i ) (cid:32) P ( Z (cid:54) y ) − α (cid:88) r = p +1 r − (cid:89) j = p +1 F j ( x j )(1 − F r ( x r )) P ( x r < Z (cid:54) y ) (cid:33) = (cid:89) i ∈ T F i ( x i ) (cid:32) P ( Z (cid:54) x p +1 ) + α − (cid:88) r = p +1 r (cid:89) j = p +1 F j ( x j ) P ( x r < Z (cid:54) x r +1 )+ α (cid:89) j = p F j ( x j ) P ( x α < Z (cid:54) y ) (cid:33) ;in order to get the sum in the second expression, we subtract the sub-trahend of a certain term from the minuend of the previous term of thesum in the first expression, going through all terms. We now rewritethis sum and compute further P ( A ) = (cid:89) i ∈ T F i ( x i ) (cid:32) F Z ( x p ) + α (cid:88) r = p r (cid:89) j = p F j ( x j ) ×× (cid:18) F Z (cid:18) min j ∈ T ∪{ r +1 ,...,n } x j (cid:19) − F Z (cid:18) max j ∈{ ,...,r } x j (cid:19)(cid:19)(cid:19) = (cid:89) i ∈ T F i ( x i ) (cid:32) (cid:88) K ⊆ S (cid:89) j ∈ K F j ( x j ) max (cid:26) , F Z (cid:18) min j ∈ T ∪ S \ K x j (cid:19) − F Z (cid:18) max j ∈ K x j (cid:19)(cid:27)(cid:33) . Now, the last sum above is written independently of the ordering of thecomponents x j and this brings us immediately to the desired formula(21).We now introduce the auxiliary generating functions. Here we arefacing the dilemma that two ways of defining ϕ ∗ have been used in theliterature. To avoid confusion we introduce a new notation ϕ † i ( u i ) = u i ϕ i ( u i ) for i = 1 , . . . , p, u i > ϕ † j ( u j ) = u j − χ j ( u j )1 − χ j ( u j ) for j = p + 1 , . . . , n, u j < . With this notation of ϕ † we follow the definition of ϕ ∗ and χ ∗ of [8] and[13], while the according definitions of these auxiliary functions weresomewhat different in [22]. Next we define ϕϕϕ = ( ϕ , . . . , ϕ p , χ p +1 , . . . , χ n ) and ϕϕϕ † = ( ϕ † , . . . , ϕ † n ) . Formula (21) now yields (cf. also [8, (4.4)]) C MM ϕϕϕ ( u ) = (cid:32)(cid:89) i ∈ T ϕ i ( u i ) (cid:33) × (cid:88) K ⊆ S (cid:89) j ∈ S \ K χ j ( u j ) max (cid:26) , min i ∈ T ∪ K ϕ † i ( u i ) − max j ∈ S \ K ϕ † j ( u j ) (cid:27) , (22)and H ( x , . . . , x n ) = C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )).In the imprecise setting we work in a finitely-additive probabilityspace. Endogenous shocks are given by random variables whose dis-tributions F i belong to F ( F i ,F i ) . We define the generating functionsusing the ideas of Equations (7) and (8) so that, in particular, they dosuffice the above defining relations. In addition, by Lemmas 4 and 5we deduce that F (cid:48) i (cid:54) F i implies ϕ (cid:48) i (cid:54) ϕ i for i = 1 , . . . , p, and F (cid:48) j (cid:54) F j implies χ (cid:48) j (cid:54) χ j for j = p + 1 , . . . , n. As in Section 6 we introduce the minimal and the maximal joint dis-tribution functions H = min { H ( x , . . . , x n ) | F i ∈ F ( F i ,F i ) for i = 1 , , . . . , n } and H = max { H ( x , . . . , x n ) | F i ∈ F ( F i ,F i ) for i = 1 , , . . . , n } . Given two vectors of functions ϕϕϕ (cid:54) ϕϕϕ such that each of their com-ponents satisfies Conditions (F1)–(F3), we let C MM = { C MM ϕϕϕ : ϕϕϕ (cid:54) ϕϕϕ (cid:54) ϕϕϕ } , where each component of ϕϕϕ also satisfies Conditions (F1)–(F3), andcall this set of copulas an n -variate imprecise maxmin (MM for short)copula . In Condition (F3) we apply the respective definitions of ϕ † i and ϕ † j given above for i = 1 , . . . , p, and for j = p + 1 , . . . , n instead ofstarred functions of Subsection 3.3.We continue to use the ideas of Subsection 3.3 by letting ϕ i and χ j respectively ϕ i and χ j be the minimal respectively the maximalfunction satisfying Conditions (F1)–(F3) and ϕ i ( G i ) = F i if G i > ϕ i ( G i ) = F i if G i > i = 1 , . . . , p, and χ j ( G j ) = F j if G j < χ j ( G j )= F j if G j < j = p + 1 , . . . , n. Let us summarize.
Theorem 11 (Properties of multivariate imprecise MM copulas) . Inthe situation described above we have:
OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 27 (i) ϕϕϕ (cid:54) ϕϕϕ .(ii) G i = F i F Z , G i = F i F Z , G i = F i F Z , for i = 1 , . . . , p, and (cid:98) G j = (cid:98) F j (cid:98) F Z , (cid:98) G j = (cid:98) F j (cid:98) F Z , (cid:98) G j = (cid:98) F j (cid:98) F Z , for j = p + 1 , . . . , n. (iii) G i (cid:54) G i (cid:54) G i for i = 1 , . . . , p, and G j (cid:54) G j (cid:54) G j for j = p + 1 , . . . , n. (iv) ϕ i ( G i ) = F i if G i > resp. ϕ i ( G i ) = F i if G i > for i = 1 , . . . , p, and χ j ( G j ) = F j if G j < resp. χ j ( G j )= F j if G j < for j = p + 1 , . . . , n. (v) ϕ † i ( G i ) = ϕ † j ( G j ) = ϕ † k ( G k ) = F Z for all i, j, k = 1 , . . . , n .(vi) H ( x , . . . , x n ) = C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) . (vii) C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) (cid:54) C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) (cid:54) C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) . (viii) H ( x , . . . , x n ) = C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) ,H ( x , . . . , x n ) = C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) . Proof.
Points from (i) to (iv) follow by the observations above. (v) :Let us compute only one of the three cases that go in a similar way: ϕ † i ( G i ) = G i ϕ i ( G i ) = G i F i = F Z , for i = 1 , . . . , p, and ϕ † j ( G j ) = G j − χ j ( G j )1 − χ j ( G j ) = G j − F j − F j = F Z , for j = p + 1 , . . . , n. Point (vi) amounts to the same as Equation (22). To get (vii) observethat C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) = (cid:88) K ⊆ S (cid:89) i ∈ T ∪ ( S \ K ) F i ( x i ) max (cid:26) , min i ∈ T ∪ K ϕ † i ( G i ( x i )) − max j ∈ S \ K ϕ † j ( G j ( x j )) (cid:27) = (cid:88) K ⊆ S (cid:89) i ∈ T ∪ ( S \ K ) F i ( x i ) max (cid:26) , min i ∈ T ∪ K F Z ( x i ) − max j ∈ S \ K F Z ( x j ) (cid:27) (cid:54) (cid:88) K ⊆ S (cid:89) i ∈ T ∪ ( S \ K ) F i ( x i ) max (cid:26) , min i ∈ T ∪ K F Z ( x i ) − max j ∈ S \ K F Z ( x j ) (cid:27) = C MM ϕϕϕ ( G ( x ) , . . . , G n ( x n )) . and the first desired inequality follows. Considerations of the samekind yield the second one. Finally, Point (viii) follows from Points (vi) and (vii) . (cid:3) Multivariate imprecise RMM copulas
Finally, we extend the imprecise reflected maxmin copulas from Sec-tion 5 to the multivariate case as well. For the third time we start byrevisiting the precise case. As in Section 7 we let X , . . . , X p , X p +1 , . . . , X n , Z be independent variables with respective distribution functions F , . . . , F p , F p +1 , . . . , F n , F Z and define U i = max { X i , Z } , for i = 1 , . . . , p, and U j = min { X j , Z } , for j = p + 1 , . . . , n. (23)So, for the respective distribution functions G , . . . , G n of U , . . . , U n we have again G i = F i F Z for i = 1 , . . . , p, and (cid:98) G j = (cid:98) F j (cid:98) F Z for j = p + 1 , . . . , n. Following the notation of the bivariate case presented in Subsection 3.4we determine that the generating functions should suffice ϕ i ( G i ) = F i , f i ( G i ) = F i (cid:98) F Z if G i > i = 1 , . . . , p, and χ j ( G j ) = F j , f j ( (cid:98) G j )= (cid:98) F j F Z if G j < j = p + 1 , . . . , n. Following the notation of the bivariate case write H σ ( x , . . . , x n ) forthe joint distribution function of the random vector ( U , U , . . . , U n ) inwhich the last n − p entries are reflected, i.e., H σ ( x , . . . , x n ) = P ( U (cid:54) x , . . . , U p (cid:54) x p , U p +1 > x p +1 , . . . , U n > x n ) . Recall [13, Theorem 14] to get(24) H σ ( x , . . . , x n ) = C RMM f ( G ( x ) , . . . , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n )) . OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 29
There the authors introduced notation(25) C RMM f ( u ) = max , min i ∈{ ,...,p } j ∈{ p +1 ,...,n } ( u i u j − f i ( u i ) f j ( u j )) (cid:89) l ∈{ ,...,n } l (cid:54) = i,j ( u l + f l ( u l )) , where f = ( f , f , . . . , f n ) and u = ( u , u , . . . , u n ).In the imprecise setting we work in a finitely-additive probabilityspace. Endogenous shocks are given by random variables whose distri-butions F i belong to F ( F i ,F i ) . We define the generating functions usingthe ideas of Equations (7) and (8) so that, in particular, they do sufficethe above defining relations. In addition, by Lemmas 4 and 5, and byEquation (4) we deduce that F (cid:48) i (cid:54) F i implies ϕ (cid:48) i (cid:54) ϕ i and f (cid:48) i (cid:54) f i for i = 1 , . . . , p, and F (cid:48) j (cid:54) F j implies χ (cid:48) j (cid:54) χ j and f j (cid:54) f (cid:48) j for j = p + 1 , . . . , n. As in Sections 6 and 7 we introduce the minimal and the maximal jointdistribution functions H σ = min { H σ ( x , . . . , x n ) | F i ∈ F ( F i ,F i ) for i = 1 , , . . . , n } and H σ = max { H σ ( x , . . . , x n ) | F i ∈ F ( F i ,F i ) for i = 1 , , . . . , n } . Given two vectors of functions f (cid:54) f such that each of their compo-nents satisfies Conditions (G1)–(G3), we let C RMM = { C RMM f : f (cid:54) f (cid:54) f } , where each component of f also satisfies Conditions (G1)–(G3), andcall this set of copulas an n -variate imprecise reflected maxmin (RMMfor short) copula . In Condition (G3) we apply the respective definitionsof f ∗ and g ∗ given in Subsection 3.4 in an obvious way to introduce f ∗ i and f ∗ j for i = 1 , . . . , p, and for j = p + 1 , . . . , n .We continue to use the ideas of Subsection 3.4 by letting f i respec-tively f i be the minimal respectively the maximal function satisfyingConditions (G1)–(G3) and f i ( G i ) = F i (cid:98) F Z if G i > f i ( G i ) = F i (cid:98) F Z if G i > i = 1 , . . . , p, and f j ( (cid:98) G j ) = (cid:98) F j F Z if G j < f j ( (cid:98) G j )= (cid:98) F j F Z if G j < j = p + 1 , . . . , n. Let us summarize.
Theorem 12 (Properties of multivariate imprecise RMM copulas) . Inthe situation described above we have:(i) f (cid:54) f . (ii) For C RMM f = inf C RMM and C RMM f = sup C RMM we have that C RMM f = (cid:94) (cid:101) f i ∈{ f i ,f i } C RMM (cid:101) f = (cid:94) i ∈{ ,...,p } j ∈{ p +1 ,...,n } (cid:101) f i = f i , (cid:101) f j = f j , (cid:101) f l = f l ,l (cid:54) = i,j C RMM (cid:101) f and C RMM f = (cid:95) (cid:101) f i ∈{ f i ,f i } C RMM (cid:101) f = (cid:95) i ∈{ ,...,p } j ∈{ p +1 ,...,n } (cid:101) f i = f i , (cid:101) f j = f j , (cid:101) f l = f l ,l (cid:54) = i,j C RMM (cid:101) f , where the infimum and supremum are attained pointwise.(iii) G i = F i F Z , G i = F i F Z , G i = F i F Z , for i = 1 , . . . , p, and (cid:98) G j = (cid:98) F j (cid:98) F Z , (cid:98) G j = (cid:98) F j (cid:98) F Z , (cid:98) G j = (cid:98) F j (cid:98) F Z , for j = p + 1 , . . . , n. (iv) G i (cid:54) G i (cid:54) G i for i = 1 , . . . , p, and (cid:98) G j (cid:62) (cid:98) G j (cid:62) (cid:98) G j for j = p + 1 , . . . , n. (v) f i ( G i ) = F i − G i , f i ( G i ) = F i − G i , if G i , G i > resp. for i = 1 , . . . , p, and f j ( (cid:98) G j ) = (cid:98) F j − (cid:98) G j , f j ( (cid:98) G j ) = (cid:98) F j − (cid:98) G j , if G j , G j < resp. for j = p + 1 , . . . , n. (vi) f ∗ i ( G i ) f ∗ j ( (cid:98) G j ) = 1 , f ∗ i ( G i ) f ∗ j ( (cid:98) G j ) = 1 , f ∗ i ( G i ) f ∗ j ( (cid:98) G j ) = 1 ,f ∗ i ( G i ) f ∗ j ( (cid:98) G j ) = 1 , for i = 1 , . . . , p and j = p + 1 , . . . , n .(vii) H σ ( x , . . . , x n ) = C RMM f ( G ( x ) , . . . , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n )) . (viii) C RMM f ( G ( x ) , · · · , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n )) (cid:54) C RMM f ( G ( x ) , . . . , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n )) (cid:54) C RMM f ( G ( x ) , . . . , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n )) . (ix) H σ ( x , . . . , x n ) = C RMM f ( G ( x ) , . . . , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n )) ,H σ ( x , . . . , x n ) = C RMM f ( G ( x ) , . . . , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n )) . OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 31
Proof. (i) follows by the observations above, (ii) follows from Equation(25) after a short computation. (iii) , (iv) and (v) are immediate fromthe above. (vi) Let us compute only one of the four cases that go in asimilar way: f ∗ i ( G i ) = f i ( G i ) G i = F i − G i G i = 1 F Z − (cid:98) F Z F Z , for i = 1 , . . . , p, and f ∗ j ( (cid:98) G j ) = f j ( (cid:98) G j ) (cid:98) G j = (cid:98) F j − (cid:98) G j (cid:98) G j = 1 (cid:98) F Z − F Z (cid:98) F Z , for j = p + 1 , . . . , n. Point (vii) amounts to the same as Equation (24). To get (viii) applyEquation (25) to the formula C = C RMM f ( G ( x ) , · · · , G p ( x p ) , (cid:98) G p +1 ( x p +1 ) , . . . , (cid:98) G n ( x n ))and write the expression under the min operator as a product of threefactors A ij = G i ( x i ) (cid:98) G j ( x j ) − f i ( G i ( x i )) f j ( (cid:98) G j ( x j )) B i = p (cid:89) l =1 l (cid:54) = i ( G l ( x l ) + f l ( G l ( x l ))) and D j = n (cid:89) l = p +1 l (cid:54) = j ( (cid:98) G l ( x l ) + f l ( (cid:98) G l ( x l ))) , for i = 1 , . . . , p, and j = p + 1 , . . . , n . A straightforward computationusing (v) yields A ij = F i ( x i ) (cid:98) F j ( x j )( (cid:98) F Z ( x j ) − (cid:98) F Z ( x i )) so that A ij (cid:54) F i ( x i ) (cid:98) F j ( x j )( (cid:98) F Z ( x j ) − (cid:98) F Z ( x i )) = G i ( x i ) (cid:98) G j ( x j ) − f i ( G i ( x i )) f j ( (cid:98) G j ( x j )) . In a similar way we get B i (cid:54) p (cid:89) l =1 l (cid:54) = i ( G l ( x l ) + f l ( G l ( x l ))) and D j (cid:54) n (cid:89) l = p +1 l (cid:54) = j ( (cid:98) G l ( x l ) + f l ( (cid:98) G l ( x l ))) , and the first desired inequality follows. Considerations of the samekind yield the second one. Finally, Point (ix) follows from Points (vii) and (viii) . (cid:3) Remark.
Note that in Theorem 11, there is no statement that isequivalent to Point (ii) of Theorem 12, because it appears that no suchstatement can be proven for multivariate imprecise MM copulas. Thusin a sense, the multivariate imprecise RMM copulas behave nicer fromthe point of view of pointwise order than multivariate imprecise MMcopulas. Conclusion
The uncertainty of the final outcome of the rules of modeling de-pendencies in the bivariate imprecise setting issues a warning that oneshould address this issue on the multivariate imprecise level with ut-most caution. A view on this problem was presented in the last but oneparagraph of our introduction. Therefore, the paper [25] was helpfulto the specialists in the area giving two important examples of bi-variate imprecise copulas, Marshall’s and maxmin copulas; there, thebackground assumptions of the shock model inducing each of the twofamilies of copulas were assumed imprecise thus leading to a naturallydefined imprecise copulas that have many interesting additional prop-erties including coherence.There is an important fact, namely [24, Theorem 4], saying that thecopulas obtained via the Sklar’s theorem from bivariate distributions onfinitely additive probability spaces are the same as the ones obtainedon the standard probability spaces. This means that whenever thecontroversy of the bivariate imprecise dependence is resolved, it willbe resolved both for the standard and for the non-standard approachsimultaneously.This encourages us to present an investigation of three major mul-tivariate cases of shock model induced copulas: the Marshall’s, themaxmin, and the reflected maxmin copulas (RMM). We believe thatthe properties of these objects, no matter what they will be called inthe end, will help further investigations in the area. Here are somequick findings of ours. In all the three cases the extreme values ofthe generators lead us to extreme values of the set of copulas and toextreme values of the set of distributions. In the case of Marshall’scopulas this correspondence is simple and expected, a direct extensionof the bivariate case. The set of copulas (a possible candidate for an n -variate imprecise copula) is coherent (Theorem 10 (ii) ), the set of thecorresponding joint distributions is coherent (Theorem 10 (v) & (vi) ) andthe lower and upper bounds correspond to each other.The maxmin copulas behave somewhat differently. The set of jointdistributions is coherent (Theorem 11 (vii) & (viii) ); however, the ques-tion of coherence of the (possible candidate for an) n -variate imprecisecopula is left open. In the bivariate case one is able to make this setcoherent as well, although the extremes were not correspondent to theaccording extreme distributions. The case of RMM copulas is moreinvolved but in some sense clearer. The set of joint distributions is OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 33 coherent (Theorem 12 (viii) & (ix) ) and we can express the lower andthe upper bound of the (possible candidate for an) n -variate impre-cise copula as a minimum, respectively maximum of a finite number ofcopulas that belong to a specific set (Theorem 12 (ii) ). The obtainedbounds are quasi-copulas in general and the solution to the question ofcoherence of this set needs methods that are yet to be discovered. (Forthe bivariate case the kind of methods were developed in [23].) Sincethe maxmin and RMM copulas are obtainable from each other througha number of reflections, it is possible that a similar conclusion as theone exhibited in Theorem 12 (ii) exists for maxmin copulas as well, butin view of the last remark in the paper, this looks like a nontrivial taskfor further investigations.Consequently, our paper opens a number of questions left to the com-munity of experts on imprecise copulas to solve. These are primarilytasks in the multivariate imprecise setting:(1) How to define a p -box of multivariate joint distributions?(2) What to adopt as an imprecise multivariate copula?(3) One needs a Sklar type theorem connecting the two notionsabove.(4) One needs to develop an n -variate coherence testing algorithmfor a set of (quasi)copulas extended from the bivariate case (cf.[23]).(5) One needs to develop an n -variate coherence testing algorithmfor a set of (quasi)distributions extended from the bivariate case(cf. [24]). References [1] T. Augustin, F. P. A. Coolen, G. de Cooman, M. C. M. Troffaes (editors),
Introduction to imprecise probabilities . John Wiley & sons, Chichester (2014).[2] F. P. A. Coolen,
On the Use of Imprecise Probabilities in Reliability , Qualityand Reliability Engineering International. (2004), 193–202.[3] I. Couso, S. Moral, P. Walley, A survey of concepts of independence for impre-cise probabilities , Risk, Decision and Policy, (2000), 165–181.[4] I. Couso, S. Moral, Independence concepts in evidence theory , InternationalJournal of Approximate Reasoning, (2010), 748–758.[5] G. de Cooman, F. Hermans, E. Quaeghebeur, Imprecise Markov chains andtheir limit behavior. Probability in the Engineering and Informational Sciences , (2009), 597–635. Here we understand the term coherent in the usual way, namely, it means thatthe value of the lower respectively upper bound can be approximated at any fixedpoint of the unit square by values of the copulas from the set; actually in our casethis value is even attained. [6] F. Durante, C. Sempi.
Principles of Copula Theory . CRC/Chapman & Hall,Boca Raton (2015).[7] F. Durante, S. Girard, G. Mazo,
MarshallOlkin type copulas generated by aglobal shock , Journal of Computational and Applied Mathematics, (2016),638–648.[8] F. Durante, M. Omladiˇc, L. Oraˇzem, N. Ruˇzi´c,
Shock models with dependenceand asymmetric linkages , Fuzzy Sets and Systems, (2017), 152–168.[9] S. Ferson, V. Kreinovich, L. Ginzburg, D. S. Myers, K. Sentz,
Con-structing probability boxes and Dempster-Shafer structures . Technical reportSAND2002-4015, 2003.[10] C. Jansen, G. Schollmeyer, T. Augustin,
Concepts for decision making un-der severe uncertainty with partial ordinal and partial cardinal preferences ,International Journal of Approximate Reasoning, (2018), 112–131.[11] H. Joe. Dependence Modeling with Copulas.
Chapman & Hall/CRC, London,2014.[12] T. Koˇsir, M. Omladiˇc,
Reflected maxmin copulas and modeling quadrantsubindependence , Fuzzy Sets and Systems (2020), 125–143.[13] D. Kokol Bukovˇsek, T. Koˇsir, B. Mojˇskerc, M. Omladiˇc,
Asymmetric linkages:maxmin vs. reflected maxmin copulas , Fuzzy sets and systems, (2020),75–95.[14] A. W. Marshall, I. Olkin,
A multivariate exzponential distributions ,J. Amer. Stat. Assoc., , (1967), 30–44.[15] A. W. Marshall, Copulas, marginals, and joint distributions , in:L. R¨uschendorf, B. Schweitzer, M. D. Taylor (eds.), Distributions with FixedMarginals and Related Topics in LMS, Lecture Notes – Monograph Series, (1996), 213–222.[16] E. Miranda, I. Montes, Shapley and Banzhaf values as probability transforma-tions , International Journal of Uncertainty, Fuzziness and Knowledge-BasedSystems. (2018), 917–947.[17] I. Montes, E. Miranda, S. Montes, Decision making with imprecise probabili-ties and utilities by means of statistical preference and stochastic dominance .European Journal of Operational Research, (2014), 209–220.[18] I. Montes, E. Miranda, R. Pelessoni, P, Vicig,
Sklar’s theorem in an imprecisesetting , Fuzzy Sets and Systems, (2015), 48–66.[19] R. Nau,
Imprecise probabilities in Non-cooperative games . In Proceedings ofISIPTA (2011), 297–306.[20] R. B. Nelsen.
An introduction to copulas . 2nd edition, Springer-Verlag, NewYork (2006).[21] M. Oberguggenberger, J. King, B. Schmelzer,
Classical and imprecise probabil-ity methods for sensitivity analysis in engineering: a case study , InternationalJournal of Approximate Reasoning, (2009), 680–693.[22] M. Omladiˇc, N. Ruˇzi´c, Shock models with recovery option via the maxmincopulas , Fuzzy Sets and Systems, (2016), 113–128.[23] M. Omladiˇc, N. Stopar,
Final solution to the problem of relating a true copulato an imprecise copula , Fuzzy sets and systems, (2020), 96–112.[24] M. Omladiˇc, N. Stopar,
A full scale Sklar’s theorem in the imprecise setting ,Fuzzy sets and systems, (2020), 113–125.
OME MULTIVARIATE IMPRECISE SHOCK MODEL COPULAS 35 [25] M. Omladiˇc, D. ˇSkulj,
Constructing copulas from shock models with imprecisedistributions , International Journal of Approximate Reasoning, (2020),27–46[26] R. Pelessoni, P. Vicig,
Convex Imprecise Previsions , Reliable Computing, (2003), 465–485.[27] R. Pelessoni, P. Vicig, I. Montes, and E. Miranda, Imprecise copulas and bi-variate stochastic orders . In: Proc. EUROFUSE 2013, Oviedo 2013, 217–224.[28] R. Pelessoni, P. Vicig, I. Montes, E. Miranda,
Bivariate p -boxes , Interna-tional Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, (2016), 229–263.[29] B. Schmelzer, Joint distributions of random sets and their relation to copulas ,International Journal of Approximate Reasoning, (2015), 59–69.[30] B. Schmelzer, Sklar’s theorem for minitive belief functions , International Jour-nal of Approximate Reasoning, (2015), 48–61.[31] B. Schmelzer, Multivariate capacity functional vs. capacity functionals onproduct spaces . Fuzzy Sets and Systems, (2019), 1–35.[32] A. Sklar,
Fonctions de r´epartition `a n dimensions et leurs marges , Publ. Inst.Stat. Univ. Paris (1959), 229–231.[33] D. ˇSkulj, Discrete time Markov chains with interval probabilities.
InternationalJournal of Approximate Reasoning, (2009), 1314–1329.[34] M. C. M. Troffaes,
Decision making under uncertainty using imprecise proba-bilities , International Journal of Approximate Reasoning, (2007), 17–29.[35] M. C. M. Troffaes, S. Destercke Probability boxes on totally preordered spacesfor multivariate modelling , International Journal of Approximate Reasoning, (2011), 767–791.[36] L. V. Utkin, F. P. A. Coolen, Imprecise Reliability: An Introductory Overview.
In: Levitin G. (eds) Computational Intelligence in Reliability Engineering.Studies in Computational Intelligence, (2007). Springer, Berlin, Heidelberg.[37] P. Vicig, Financial risk measurement with imprecise probabilities , Interna-tional Journal of Approximate Reasoning, (2008), 159–174.[38] P. Walley, Statistical Reasoning with Imprecise Probabilities.
Chapman andHall, London, 1991.[39] L. Yu, S. Destercke, M. Sallak, W. Schon,
Comparing system reliability withill-known probabilities , Proceedings of IPMU 2016, 619–629.
David Dolˇzan, Faculty of Mathematics and Physics, University ofLjubljana, and Institute of Mathematics, Physics and Mechanics, Ljubl-jana, Slovenia
E-mail address : [email protected] Damjana Kokol Bukovˇsek, School of Economics and Business, Uni-versity of Ljubljana, and Institute of Mathematics, Physics and Me-chanics, Ljubljana, Slovenia
E-mail address : [email protected] Matjaˇz Omladiˇc, Institute of Mathematics, Physics and Mechanics,Ljubljana, Slovenia
E-mail address : [email protected] Damjan ˇSkulj, Faculty of Social Sciences, University of Ljubljana,Slovenia
E-mail address ::