Abstract
We study the limiting behaviour of the maximum of a bivariate (finite or infinite) moving average model, based on discrete random variables. We assume that the bivariate distribution of the iid innovations belong to the Anderson’s class (Anderson, 1970). The innovations have an impact on the random variables of the INMA model by binomial thinning. We show that the limiting distribution of the bivariate maximum is also of Anderson’s class, and that the components of the bivariate maximum are asymptotically independent.
Similar content being viewed by others
1 Introduction
Hall (2003) studied the limiting distribution of the maximum term \(M_n=\max (X_1,\cdots ,X_n)\) of stationary sequences \(\{X_j\}\) defined by non-negative integer-valued moving average (INMA) sequences of the form
where the innovation sequence \(\{V_i\}\) is an iid sequence of non-negative integer-valued random variables (rvs) with exponential type tails of the form
where \(\xi \in \mathbb {R},\, \lambda >0,\, L(n)\) is slowly varying at \(+\infty\) and \(\alpha_i \circ\) denotes binomial thinning with probability \(\alpha _i \in [0,1]\). Hall (2003) proved that \(\{X_j\}\) satisfies Leadbetter’s conditions \(D(x+b_n)\) and \(D'(x+b_n)\), for a suitable real sequence \(b_n\), and then
for all real x and \(\alpha _{\max }:=\max \{\alpha _i,\, i \in \mathbb{Z}\}\). Note that \(\alpha _{\max }\) plays an important role in this result. This is an extension of Theorem 2 of Anderson (1970), where it is proved that for sequences of iid rvs with an integer-valued distribution function (df) F with infinite right endpoint, the limit
is equivalent to
for all real x.
The class of dfs satisfying (1), which is a particular case of (2) (see, e.g., Hall and Temido (2007)) is called Anderson’s class.
In this paper we extend the result of Hall (2003) for the bivariate case of an INMA model. Concretely, we study the limiting distribution of the maximum term of stationary sequences \(\{(X_{j},Y_{j})\}\) where the two marginals are defined by non-negative integer-valued moving average sequences of the general form
where \(X_j\) and \(Y_j\) are defined as above with respect to a two-dimensional iid innovation sequence \(\{V_i, W_i\}\). The binomial thinning operator \(\beta \circ\), due to Steutel and van Harn (1979), is defined by \(\beta \circ Z=\sum _{s=1}^{Z} B_s(\beta ),\,\, \beta \in [0,1],\) where \(\{B_s(\beta )\}\) is an iid sequence of Bernoulli rvs independent of the positive integer rv Z. The possible class of bivariate discrete distributions \(F_{V,W}\) (see (4)) includes also the bivariate geometric models.
We assume that \(X=\alpha \circ V\) and \(Y=\beta \circ W\) are conditionally independent given (V, W), because the binomial thinning with \(\alpha \circ\) and \(\beta \circ\) are independent, X and Y are binomial rv’s. with parameters \((V, \alpha )\) respectively \((W, \beta )\), i.e.
for all events A and B and for all possible values of v and w. We assume that \(\alpha _i,\beta _i\in [0,1]\) and
for some \(\delta >2\).
We investigate the limiting behaviour of \((M_n^{(1)}, M_n^{(2)})=(\max _{1\le j\le n} X_j, \max _{1\le j\le n} Y_j)\) and want to find out whether the two maxima components are asymptotically dependent, because of the dependence of the innovations \((V_i,W_i)\). However, we will show that this is not occurring because of the independent thinning, as we believe. We investigate the impact of the dependence of \((V_i,W_i)\) on the limiting distribution and the convergence rate.
Following similar ideas of Hall (2003) for the univariate case, we:
-
Define a bivariate model \(F_{V,W}\) which contains the bivariate geometric model;
-
Characterize the tail of \((\alpha \circ V,\beta \circ W)\) and the tail of \((X_j,Y_j)\), in terms of the model \(F_{V,W}\);
-
Establish the limiting behaviour of the bivariate maximum \((M_n^{(1)}, M_n^{(2)})\) of the stationary sequence \(\{(X_{j},Y_{j})\}\) which is defined componentwise; and
-
Investigate the convergence of the joint distribution of the bivariate maximum to the limiting distribution by simulations.
Examples: 1.) We may consider the \(\{(V_i,W_i)\}\) as the number of person newly infected by virus1 (say COVID-19 virus) and virus2 (say the usual seasonal virus) at time i. It is possible that a person is infected by both virus only at the same time point. We count by \(\{(X_{j},Y_{j})\}\) the total number of infected and still contagious persons at time j adding all infected persons before and at time j. After some time these persons are cured (or died) and are no more counted to the number of infected but still contagious persons. Hence the random numbers \((V_i,W_i)\) are thinned at each time point, so the contribution to \(X_j\) is \(\alpha _{j-i}\circ V_i\), and to \(Y_j\) is \(\beta _{j-i}\circ W_i\) for \(i\le j\).
2.) Another example for bivariate integer valued time series is presented in Pedeli and Karlis (2011) who discuss the bivariate INAR(1) model with negative binomial innovations for the application of road accidents at two different time intervals in Schiphol area. However their bivariate negative binomial innovations are in the case of geometric innovations of a different type herein considered. Similar is the situation in the paper of Silva et al. (2020) who discuss inference of such a bivariate time series with different distribution of the innovations. But their bivariate negative binomial distribution is also not of our type.
3.) A further application of a bivariate time series for count data in finance is given by Quoreshi (2006). He did not specify the bivariate distribution. He derived the mean and variance/covariances of this time series.
2 Preliminaries Results for Bivariate Innovations
Let (V, W) be a non-negative random vector with bivariate df \(F_{V, W}\) satisfying
as \(v,w \rightarrow +\infty\), for positive real constants \(\lambda _i>0\), \(i=1,2\), \(\theta >0\) such that \(\theta < \min \{1+\lambda _1, 1+\lambda _2\}\) and \(\theta >1-\lambda _1\lambda _2\), some real constants \(\xi _i\) and slowly varying functions \(L_i\), \(i=1,2,3,4\), and where \(\ell (v,w)\) is a positive bounded (say by \(\vartheta\)) function which converges to a positive constant L as \(v,w \rightarrow \infty\). That \(\ell (v,w)\) converges to L is for simplicity. It has no impact on the results if the limit L would depend on \(v<w, v=w\) or \(v>w\). By [x] we denote the greatest integer not greater than x.
Remark 2.1
The marginal tails of \(F_{V,W}\) are of the form:
for \(v, w \rightarrow +\infty\). Hence, both marginal dfs belong to the Anderson’s class with
From (4), we can derive the probability function (pf) of (V, W). Because the proofs of the following propositions are technical, we move them to Appendix Proofs.
Proposition 2.1
The pf of the random vector (V, W) with df (4) is given by
for v, w large integers, where
and \(\ell (v,w)\ell ^*(v,w)\) is bounded and converges to positive constants.
Example 2.1
The Bivariate Geometric (BG) distribution is a particular case of the model (4) with margins (5). Consider the bivariate Bernoulli random vector \((B_1,B_2)\) with \(P(B_1=k,B_2=\ell )=p_{k \ell }, (k,\ell ) \in \{0,1\}^2,\) and success marginal probabilities \(p_{+1}=p_{01}+p_{11}\) and \(p_{1+}=p_{10}+p_{11}\). Due to Mitov and Nadarajah (2005), using the construction of a BG, the pf and the df of a random vector (V, W) with BG distribution are given, respectively, by
for \(v,w \in \mathbb {N}_0\), and
for \(v,w \in \mathbb {R}_0^{+}\), assuming that \(0< p_{0+}, p_{+0} <1\). Hence, this df satisfies (4) with the constants \(\lambda _1\), \(\lambda _2\) given by
and the index \(\theta\) associated to the dependence structure of \((B_1,B_2)\) is
The slowly varying functions are constants and \(\xi _i=0\), for \(i=1,2,3,4\). The independence case occurs when \(\theta =1\). For dependence cases, we can have \(0< \theta < 1\) or \(\theta >1\). Finally, we note that \(\ell (v,w)\) is a constant. For instance, take \(L_1(v)=L_3(v)= 1/(1+\lambda _1)\), \(L_2(v)=L_4(v)= 1/(1+\lambda _2)\), we have \(\ell (v,w)=\theta\) with \(\ell ^*(v,w)\) as in (6).
The marginal df of V and W are obviously
which means V and W are geometrically distributed rvs with parameter \(p_{1+}\) and \(p_{+1}\), respectively. \(\Box\)
In order to characterize the df of \((X,Y)=(\alpha \circ V, \beta \circ W)\) we start by establishing the relationship between the probability generating function (pgf) of (V, W) and (X, Y), defined e.g. for (V, W) as
which exists for \((s_1, s_2)\) in the following region \(\mathcal {R}\) (given in Lemma 2.1).
Taking into account Proposition 2.1, the series \(G_{V,W}( s_1 ,s_2 )\) converges obviously for any \(s_i\le 1\). Even for some \(s_i>1\) the series converges because of the assumption (4). By this assumption, we have \(E(s_1^V)< + \infty\) if \(s_1< 1+\lambda _1\) and \(E(s_2^W)< + \infty\) if \(s_2< 1+\lambda _2\). The following lemma gives a condition such that the series \(G_{V,W}(s_1,s_2)\) exists.
Lemma 2.1
The pgf \(G_{V,W}(s_1 ,s_2)= E(s_1^V s_2^W)\) exists for \((s_1, s_2)\) in
Its more technical proof is given also in the appendix. As consequence of this lemma, the pgf \(G_{V,W}(s_1 ,s_2)\) exists for \(s_1, s_2>1\), if \(s_i\le 1+\lambda _i, i=1,2\) in case \(\theta \le 1\), and if \(s_1\le 1+\lambda _1\) and \(s_2\theta \le 1+\lambda _2\) in case of \(\theta >1\). In the following, we use these convenient conditions for the convergence of \(G_{V,W}\).
Now the relationship of the two pgf is the following. It holds as long as the pgf’s exist. For our derivations it is convenient to use in the following the given domain \(\mathcal {R}\). The proof of this relationship is also given in the appendix.
Proposition 2.2
The pgf of \((X,Y)=(\alpha \circ V, \beta \circ W)\) is given in terms of the pgf of (V, W):
for all \((s_1,s_2)\) such that \((\alpha s_1 +1- \alpha ,\beta s_2+1-\beta ) \in \mathcal {R}\).
We want to derive an exact relationship of the two distributions \(F_{V,W}\) and \(F_{X,Y}\) with the help of a suitable transformation, as a modified pgf or a Mellin transform. We define the (bivariate) modified pgf or tail generating function (Sagitov (2017))
and analogously for X, Y. The relationship between \(Q_{V,W}\) and \(G_{V,W}\) is given in the following proposition.
Proposition 2.3
For \((s_1,s_2) \in \mathcal {R}\), we have
Proposition 2.4
The modified pgf of (X, Y) and (V, W) satisfy
if the series converge, i.e. \((\alpha s_1+1-\alpha ,\beta s_2+1-\beta ) \in \mathcal {R}\).
From Propositions 2.2 and 2.4, we can derive now the tail \(1- F_{X,Y}\) in terms of \(1- F_{V,W}\)
Proposition 2.5
The df \(F_{X,Y}\) is given in terms of the df \(F_{V,W}\) with \(x, y \in \mathbb {Z^+}\) :
Hence the tail of \(F_{X,Y}\) can be estimated by the assumption (4).
Proposition 2.6
If the joint df of (V, W) satisfies (4), then for large integers x and y
with
where \(L_i^*\) are slowly varying functions, being
with
and \(\vartheta\) the bound of \(\ell (v,w)\).
Note that \(1<\lambda _{2\theta }< \lambda _2\).
We observe that the stationary bivariate INMA model \(\left( X_j, Y_j\right)\) introduced in our work is an extension of the BINAR model of Pedeli and Karlis (2011) defined by
with an iid innovations sequence \(\{ (R_{1j}, R_{2j})\}\). In their paper it is stated that it has also the representation
Hence, considering \(\left( X_j, Y_j\right)\) with \(\alpha _i=\beta _i=0\) for \(i < 0\), \(\alpha _i=\alpha ^i\) and \(\beta _i=\beta ^i\) for \(i \ge 0\) we obtain \(\left( \widetilde{X}_j, \widetilde{Y}_j\right)\).
3 The Bivariate Stationary Sequence
We consider now the stationary bivariate INMA model \(\{\left( X_j, Y_j\right) \}\) with iid innovations \(\{(V_i,W_i)\}\) with df satisfying (4). We establish first the tail behaviour of \(\left( X_j, Y_j\right)\). The maximal values of \(\alpha _i\) and \(\beta _i\) are most important as in the univariate case. Therefore we write \(\alpha _{\max }=\max \{\alpha _i : \left| i\right| \ge 0\}\) and \(\beta _{\max }=\max \{\beta _i : \left| i\right| \ge 0\}\). We assume that they are unique. It may happen in the bivariate case that \(\alpha _{\max }\) and \(\beta _{\max }\) occurs at the same index or at different ones. We consider both cases. Furthermore, we use that
which holds because of (3).
Suppose first that \(\alpha _{\max }\) and \(\beta _{\max }\) are occuring at different indexes \(i_0\) and \(i_1\), respectively. We write for any j
and
Denote \(S_1=\alpha _{\max }\circ V_{j-i_0}\), \(S_2=\alpha _{i_1} \circ V_{j-i_1}\), \(\displaystyle S_3=\sum _{i \ne i_0,i_1} \alpha _i \circ V_{j-i}\), \(S=S_2+S_3\), \(T_1=\beta _{\max } \circ W_{j-i_1}\), \(T_2=\beta _{i_0}\circ W_{j-i_0}\) and \(\displaystyle T_3=\sum _{i \ne i_0, i_1}\beta _i \circ W_{j-i}\), \(T=T_2+T_3\). Hence, \(X_j =S_1+S_2+S_3=S_1 + S\) and \(Y_j =T_1+T_2+T_3=T_1 + T\). Note that \(S,S_i, T\) and \(T_i\) depend on j.
For the proof of the main proposition of this section we need the following lemma.
Lemma 3.1
-
a)
If the rv V belongs to the Anderson’s class, then \(E(1+h)^V= 1+hE(V)(1+o_h(1)),\quad \mathrm{as} \;\; h\rightarrow 0^+.\)
-
b)
For any set I of integers with \(\alpha _I=\max \{\alpha _i, i \in I \}\), consider the rv \(Z=\sum _{i \in I} \alpha _i \circ V_{-i}.\) Then \(E(1+h)^Z\) is finite for any \(0< h <\frac{\lambda _1}{\alpha _I}\).
The proof of this lemma is given in the appendix. We deal now with the limiting behaviour of the tail of \((X_j, Y_j)\). Besides of the univariate tail distributions we derive only an appropriate positive upper bound \(H^*(x,y)\) for the joint tail which is sufficient for the asymptotic limit distribution of the maxima. We will see that we get asymptotic independence of the components of the bivariate maxima \((M_n^{(1)}, M_n^{(2)})\), since this normalized \(H^*(x,y)\) is vanishing, not contributing to the limit.
For the asymptotic behaviour of the tail of the stationary distribution of the sequence \(\{(X_{j},Y_{j})\}\), we write simply (X, Y) for any \((X_j,Y_j)\). As mentioned we deal with the two cases that \(\alpha _{\max }\) and \(\beta _{\max }\) are occurring at different indexes or at the same one. We start with the first case and the above defined \(S, S_i, T, T_i\).
For this derivation, we use \(\psi , \rho \in (0,1)\) and \(\lambda >0\) such that \(\frac{\lambda _1}{\alpha _{\max }}<\lambda < \frac{\lambda _1}{\alpha ^*}\), with \(\alpha ^*=\max \{ \alpha _i, i\ne i_0\}\), and \(\lambda _{2\theta }\) given in (9),
and
Proposition 3.1
If \(\left( V, W\right)\) satisfies (4) and \(\alpha _{\max }\) and \(\beta _{\max }\) are unique and taken at different indexes, then
-
(i)
for the marginal dfs
$$\begin{aligned} 1- F_{X}(x) \sim x^{\xi _1}\left( 1+\frac{\lambda _1}{\alpha _{\max }}\right) ^{-\left[ x\right] }L_1^{**}(x),\,x \rightarrow +\infty , \end{aligned}$$and
$$\begin{aligned} 1- F_{Y}(y) \sim y^{\xi _2}\left( 1+\frac{\lambda _2}{\beta _{\max }}\right) ^{-\left[ y\right] }L_2^{**}(y),\,y \rightarrow +\infty , \end{aligned}$$ -
(ii)
for the joint df with \(\psi , \rho , \lambda\) satisfying (12) and (13)
as \(\,x,y \rightarrow +\infty\), where
and
for some constant \(C>0\).
We show also that \(P(S>\psi x)=o_x(P(S_1 > x))\).
Proof
In fact
We deal with the three terms in (16), separately.
-
(i)
Since \(\frac{\lambda _1}{\alpha _{\max }} < \frac{\lambda _1}{\alpha ^*}\), taking the sum \(S=Z\) in Lemma 3.1, we conclude that \(E\left( 1+\frac{\lambda _1}{\alpha _{\max }}\right) ^{S}\) is finite. Similarly \(E\left( 1+\frac{\lambda _2}{\beta _{\max }}\right) ^{T}\) is finite since \(\frac{\lambda _2}{\beta _{\max }} < \frac{\lambda _2}{\beta ^*}\) with \(\beta ^*=\max \{\beta _i, {i\ne i_1}\}\). The tail function of X is given, with \(\psi _x=[\psi x]\), by
$$\begin{aligned} \begin{array}{lll} 1- F_{X}(x) &{} = &{} P\left( S_1 + S> x\right) =\displaystyle {\sum _{k=0}^{+ \infty }P(S_1> x-k)P(S=k)}\\ &{} &{} \\ &{} = &{} P\left( S_1> x\right) \displaystyle { \sum _{k=0}^{\psi _x}} \,\,\frac{P(S_1> x-k)}{P(S_1>x)}P(S=k)+ \\ &{}&{} \\ &{} &{} \quad +\displaystyle { \sum _{k=\psi _x+1}^{+ \infty } P(S_1> x-k)P(S=k)}.\\ \end{array} \end{aligned}$$(17)For the first sum of (17), we get by applying Proposition 2.6 with \(\alpha =\alpha _{\max }\) for the marginal distribution
$$\begin{aligned} \begin{aligned}&\sum _{k=0}^{\psi _x} \frac{P(S_1> x-k)}{P(S_1>x)}P(S=k) = \sum _{k=0}^{\psi _x}\left( 1+ \frac{\lambda _1}{\alpha _{\max }}\right) ^k (1+o_x(1))P(S=k)\\&\quad \rightarrow \displaystyle {\sum _{k=0}^{+ \infty }}\left( 1+ \frac{\lambda _1}{\alpha _{\max }}\right) ^k P(S=k)\\&\quad = E\left( 1+\frac{\lambda _1}{\alpha _{\max }}\right) ^{S}, \, x \rightarrow +\infty ,\\ \end{aligned} \end{aligned}$$by dominated convergence. For the second sum in (17), we get for x large
$$\begin{aligned} \begin{aligned}&\sum _{k=\psi _x+1}^{+ \infty } P(S_1> x-k)P(S=k) \le P(S>\psi _x)\\&= P\left( (1+\lambda )^{S} > (1+ \lambda )^{\psi _x}\right) \le \frac{E\left( 1+\lambda \right) ^{S}}{(1+ \lambda )^{\psi _x}},\\ \end{aligned} \end{aligned}$$(18)using the Markov inequality, since \(E\left( 1+\lambda \right) ^{S}\) is finite for \(\lambda <\lambda _1/\alpha ^*\). Since \((1+\lambda )^\psi > 1+\frac{\lambda _1}{\alpha _{\max }},\) we get by Theorem 4 of Hall (2003)
$$\begin{aligned} \begin{aligned} \frac{(1+ \lambda )^{-\psi _x}}{P(S_1>x)} \rightarrow 0,\, x \rightarrow + \infty ,\end{aligned} \end{aligned}$$(19)and thus together
$$\begin{aligned} \begin{aligned} 1- F_{X}(x)&= P(S_1>x)\left[ E\left( 1+\frac{\lambda _1}{\alpha _{\max }}\right) ^{S} + O_x\left( \frac{(1+\lambda )^{-\psi _x}}{P(S_1>x)}\right) \right] \\&= P(S_1>x)E\left( 1+\frac{\lambda _1}{\alpha _{\max }}\right) ^{S}(1+o_x(1)).\\ \end{aligned} \end{aligned}$$With the same arguments we characterize the tail \(1- F_{Y}\). Hence, the statements on the marginal dfs are shown.
-
(ii)
Now we deal with the third term in (16). Note that \((S_1,T_2)\), \((S_2,T_1)\) and \((S_3, T_3)\) in the representation of X and Y are independent. For any \(\psi \in (0,1)\) and \(\lambda >0\) satisfying (12), we use that (18) and (19) imply
$$\begin{aligned} \begin{aligned}&P(S_2 + S_3>\psi x) = P\left( S > \psi x\right) = O_x((1+ \lambda )^{-\psi x})\ \end{aligned} \end{aligned}$$(20)and
$$\begin{aligned} \begin{aligned}&P(S> \psi x) = o_x (P\left( S_1 > x\right) ).\ \end{aligned} \end{aligned}$$(21)The probability in the third term of (16) is split into four summands with \(\psi <1\) satisfying (12), \(\psi _x=[\psi x]\) and \(\delta _y=[y-(\log y)^2]\). We get for x and y large,
$$\begin{aligned} \begin{aligned}&P(X>x,\ Y>y)=P(S_1+S_2+S_3> x,\ T_1+T_2+T_3>y)\\&= \sum _{k=0}^{\psi _x}\sum _{\ell =0}^{ \delta _y} P(S_1> x-k, T_2> y-\ell ) P(S_2+S_3=k,T_1+ T_3=\ell )+ \\&{+\sum _{k=0}^{\psi _x}\sum _{\ell =\delta _y+1}^{+ \infty } P(S_1> x-k, T_2> y-\ell ) P(S_2+S_3=k,T_1+ T_3=\ell ) +} \\&{+ \sum _{k=\psi _x+1}^{+ \infty }\sum _{\ell =0}^{ \delta _y} P(S_1> x-k, T_2> y-\ell ) P(S_2+S_3=k,T_1+ T_3=\ell ) +}\\&{+ \sum _{k=\psi _x+1}^{+ \infty }\sum _{\ell =\delta _y+1}^{+ \infty } P(S_1> x-k, T_2 > y-\ell ) P(S_2+S_3=k,T_1 +T_3=\ell )}\\&=: \sum _{m=1}^4 S_m(\psi _x,\delta _y)\\ \end{aligned} \end{aligned}$$(22)to simplify the proof. The last sum \(S_4(\psi _x,\delta _y)\) is bounded by \(P(S_2+S_3> \psi _x, T_1+T_3>\delta _y)\le P(S_2+S_3 > \psi _x)=O_x((1+\lambda )^{-\psi _x})\) by (20). For the first sum \(S_1(\psi _x,\delta _y)\) of (22) we use Proposition 2.6 and obtain with \(\rho <1\) such that (13) holds,
$$\begin{aligned} \begin{aligned}&\sum _{k=0}^{\psi _x}\sum _{\ell =0}^{\delta _y} P(S_1> x-k, T_2 > y-\ell ) P(S_2+S_3=k,T_1 +T_3=\ell )\\&\le \vartheta \sum _{k=0}^{\psi _x}\sum _{\ell =0}^{\delta _y} \left( [x]-k\right) ^{\xi _{3}}\left( [y]-\ell \right) ^{\xi _{4}}\left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{- ([x]- k)}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{-([y]-\ell )}\times \\& \times L_3^{*}([x]-k) \ L_4^{*}( [y]- \ell ) P\left( S_2+S_3=k, T_1+ T_3=\ell \right) \\&\le \vartheta \sum _{k=0}^{\psi _x}\sum _{\ell =0}^{\delta _y} \left( [x]-k\right) ^{\xi _{3}}\left( [y]-\ell \right) ^{\xi _{4}}\left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{- ([x]- k)}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{-((1-\rho )+\rho )([y]-\ell )}\times \\& \times L_3^{*}([x]-k) L_4^{ *}( [y]- \ell ) P\left( S_2+S_3=k, T_1+ T_3=\ell \right) . \end{aligned} \end{aligned}$$Note that \(\left( [y]-\ell \right) ^{\xi _{4}}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{-(1-\rho )([y]-\ell )} L_4^{ *} [y]- \ell =o_y(1)\) uniformly for \(\ell \le \delta _y\), i.e. \(y-\ell > (\log y)^2\rightarrow \infty\). Hence the sum is bounded above by
$$\begin{aligned} \begin{aligned}&o_y(1) x^{\xi _{3}}L_3^{ *}(x) \left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{- x}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{-\rho y}\times \\&\times \sum _{k=0}^{\psi _x}\sum _{\ell =0}^{\delta _y} \left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{k} \left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{\rho \ell } P\left( S_2+S_3=k, T_1+ T_3=\ell \right) \end{aligned} \end{aligned}$$$$\begin{aligned} \begin{aligned}&\le o_y(1) x^{\xi _{3}}L_3^{ *}(x) \left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{- x}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{-\rho y} \times \\& \times E\left( \left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{ (S_2+ S_3)}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{\rho (T_1 + T_3)}\right) \\&\le o_y(1) x^{\xi _{3}}L_3^{*}(x) \left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{- x}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{-\rho y}, \\ \end{aligned} \end{aligned}$$since the last pgf exists due to Lemma 3.1 and (13) Note that
$$\begin{aligned}&E\left( \left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{ (S_2+ S_3)}\left( 1+\frac{\lambda _{2\theta }}{\beta _{i_0}}\right) ^{\rho (T_1 + T_3)}\right) \\&= E\left( \prod _{i\neq i_0}(1+\frac{\lambda _{1}}{\alpha _{\max }})^{\alpha _i \circ V_{-i}}\left( (1+\frac{\lambda _{2\theta }}{\beta _{i_0}})^{\rho }\right) ^{\beta _i \circ W_{-i}}\right) \\&= \prod _{i\neq i_0} E\left( \left(1+ \frac{\alpha _i \lambda _{1}}{\alpha _{\max }}\right)^ {V_{-i}} \left(1+\beta _i([1+\frac{\lambda _{2\theta }}{\beta _{i_0}}]^\rho -1)\right)^ {W_{-i}}\right) . \end{aligned}$$The expectations exist by assumption (4) since \(1 + \frac{\alpha _i\lambda _{1}}{\alpha _{\max }} < 1 + \lambda _1\), and also \(1+\beta _i([1+\frac{\lambda _{2\theta }}{\beta _{i_0}}]^\rho -1) \le 1+\beta _{\max }([1+\frac{\lambda _{2\theta }}{\beta _{i_0}}]^\rho -1) < 1 + \lambda _2\), for all i, by the choice of \(\rho\) in (13), by using the arguments of Lemma 3.1. We consider now the approximation of the second sum \(S_2(\psi _x,\delta _y)\) in (22). We have with some positive constant C
$$\begin{aligned} S_2(\psi _x,\delta _y)\le & {} \displaystyle { \sum _{k=0}^{\psi _x}\sum _{\ell =\delta _y+1}^{+ \infty }} P(S_1> x-k) P(T_1+ T_3=\ell )\nonumber \\\le & {} C x^{\xi _{1}+1}L_1^{*}(x) \left( \displaystyle 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{- (1-\psi ) x} P(T_1+ T_3>\delta _y).\end{aligned}$$(23)By the arguments used to approximate \(P(X>x)=P(S_1+S_2+S_3>x)\) in (i), we also obtain
$$P(T_1+ T_3>\delta _y) \sim C y^{\xi _{2}} \ L_2^{*}( y)E\left( 1+ \frac{\lambda _2}{\beta _{\max }}\right) ^{T_3}\left( 1+\frac{\lambda _{2}}{\beta _{\max }}\right) ^{-\delta _y},\, y \rightarrow +\infty ,$$with some generic constant C. Hence, it implies together with (23)
$$\begin{aligned} \begin{aligned}&\sum _{k=0}^{\psi _x}\sum _{\ell =\delta _y+1}^{+ \infty } P(S_1> x-k, T_2 > y-\ell ) P(S_2+S_3=k,T_1+ T_3=\ell ) \\&\le C x^{\xi _{1}+1}y^{\xi _{2}}L_1^{*}(x) \ L_2^{*}( y) \left( 1+\frac{\lambda _{1}}{\alpha _{\max }}\right) ^{-(1-\psi ) x}\left( 1+\frac{\lambda _{2}}{\beta _{\max }}\right) ^{-y+ (\log y)^2},\, y \rightarrow + \infty .\\ \end{aligned} \end{aligned}$$For the third sum \(S_3(\psi _x,\delta _y)\) in (22), we get analogously to the derivation of the second sum
$$\begin{aligned} \begin{aligned} S_3(\psi _x,\delta _y)&\le \delta _y P(T_2> y-\delta _y) P(S_2+S_3> \psi _x)\\&\le y (\log y)^{2\xi _2} L_2^*((\log y)^2)(1+\frac{\lambda _2}{\beta _{i_0}})^{-(\log y)^2}P(S_2+S_3> \psi _x) \\&=o_y(1)P(S_2+S_3> \psi _x) = o_x(P(S>\psi x)). \\ \end{aligned} \end{aligned}$$
Combining now the bounds of the four terms \(S_i(\psi _x,\delta _y)\), we get the upper bound for \(H^*(x,y)\) which shows our statement. \(\square\)
Suppose now the case that the unique \(\alpha _{\max }\) and \(\beta _{\max }\) are taken at the same index \(i_0\), say. Write for any j
and
Denote \(S_1=\alpha _{\max }\circ V_{j-i_0}\), \(\displaystyle S=\sum _{i \ne i_0} \alpha _i \circ V_{j-i}\), \(T_1=\beta _{\max } \circ W_{j-i_0}\), and \(\displaystyle T=\sum _{i \ne i_0}\beta _i \circ W_{j-i}\), as used for Proposition 3.1. Observe that \((S_1,T_1)\) and (S,T) are independent. Then the corresponding statement of Proposition 3.1 holds for this case (letting \(\beta _{i_0}=\beta _{\max }\)) which is given in Proposition 3.2. We omit the proof since it is very similar to the given one with a few obvious changes.
Proposition 3.2
If \(\left( V, W\right)\) satisfies (4) and \(\alpha _{\max }\) and \(\beta _{\max }\) are unique, occurring at the same index, then the stationary distribution satisfies
as \(x,y \,\rightarrow +\infty\), where
and
for some constant \(C>0\) and \(\psi \in (0,1)\) satisfying (12).
Now we investigate the limiting behaviour for the bivariate maxima, in case of an iid sequence \(\{(X_j, Y_j)\}\).
Theorem 3.1
Let \(\left( V, W\right)\) be such that (4) holds and \(\alpha _{\max }\) and \(\beta _{\max }\) are unique, occurring either at the same or not the same index. Let
Define the normalizations
and
Then, for x, y real,
Proof
The convergence for the marginal distributions holds by applying Proposition 3.1 or 3.2 with the chosen normalization sequences. Since \(u_n(x)\) and \(v_n(y)\) are similar in type, we only show the derivation of the first marginal. Because the normalization \(u_n(x)\) is not always an integer, we have to consider \(\limsup\) and \(\liminf\). Let us deal with the \(\limsup\) case. Note that
and
For the normalization we get
So
The derivation of the \(\liminf\) is similar using \([u_n(x)]\le u_n(x)\).
Now for the joint distribution we use the bounds of \(H^*(u_n,v_n)\) of the two propositions. First we consider the case of Proposition 3.1 with \(\alpha _{max}\) and \(\beta _{max}\) at different indexes. We have to derive the limits of three boundary terms of \(H^*(u_n,v_n)\) given in Proposition 3.1 multiplied by n. The last of these terms tends to 0 because (21) holds and due to the fact that from (14), we get
which is bounded.
The first of the three boundary terms of \(H^*(u_n,v_n)\) is smaller than
because \(\rho /B>0\) with B given by (13).
The second boundary term of \(H^*(u_n,v_n)\) is smaller than
since \(1-\psi >0\) and where \(C_1\) represents a generic positive constant.
Thus the limiting distribution is proved in case of Proposition 3.1.
Now let us consider the changes of the proof for the case of Proposition 3.2. Again we have to deal with the three boundary terms of \(H^*(u_n,v_n)\) where the last two are as in Proposition 3.1. In the first of these terms we have similarly
since \(d_2\log \left( 1+ \frac{\lambda _{2\theta }}{\beta _{\max }} \right) >0\). Thus the statements are shown. \(\square\)
4 Main result
We consider now the stationary sequence \(\{(X_{j},Y_{j})\}\). From extreme value theory it is known that the behaviour of their extremes is as in the case of an iid sequence \(\{(X_{j},Y_{j})\}\) if the following two conditions hold: a mixing condition, called \(D(u_n,v_n)\), and a local dependence condition, called \(D'(u_n,v_n)\). In our bivariate extreme value case we consider the conditions \(D(u_n,v_n)\) and \(D'(u_n,v_n)\) of Hüsler (1990) (see also Hsing (1989) and Falk et al. (1990)). The condition \(D(u_n,v_n)\) is a long range mixing one for extremes and means that extreme values occurring in largely separated (by \(\ell _n\)) intervals of positive integers are asymptotically independent. The condition \(D'(u_n,v_n)\) considers the local dependence of extremes and excludes asymptotically the occurrences of local clusters of extreme or large values in each individual margin of \(\{(X_{j},Y_{j})\}\) as well as jointly in the two components. We write \(u_n, v_n\) for short because \(x, y\) do not play a role in the following proofs.
Definition 4.1
The sequence \(\{(X_{j},Y_{j})\}\) satisfies the condition \(D(u_{n},v_n)\) if for any integers \(1\le i_{1}<...<i_{p}<j_{1}<...<j_{q}\le n,\) for which \(j_{1}-i_{p}>\ell _{n},\) we have
for some \(\alpha _{n,{\ell _{n}}}\) with \(\displaystyle {\lim _{n\rightarrow +\infty } \alpha _{n,{\ell _{n}}} =0}\), for some integer sequence \(\ell _{n}=o(n)\).
We use the following \(D'(u_n,v_n)\) condition.
Definition 4.2
Let \(\{s_n\}\) be a sequence of positive integers such that \(s_n \rightarrow + \infty\). The sequence \(\{(X_{j},Y_{j})\}\) satisfies the condition \(D'(u_n,v_n)\) if
In the following we use the sequences \(\{s_n\}, \{\ell _n\}\) and \(\alpha _{n,\ell _{n}}\) such that
Such a sequence \(\{s_n\}\) in (26) exists always. Take e.g. for the given \(\ell _n\) and \(\alpha _{n,\ell _n}\) in condition \(D(u_n,v_n)\) the sequence \(s_n=\min (\sqrt{n/\ell _n}, 1/\sqrt{\alpha _{n,\ell _n}})\rightarrow +\infty\). In our proof we use simpler sequences.
Write \(M_{n}^{(1)}= \max \{X_1,\cdots ,X_{n}\}\) and \(M_{n}^{(2)}= \max \{Y_1,\cdots ,Y_{n}\}\). For the stationary sequence \(\{(X_{j},Y_{j})\}\) satisfying \(D(u_{n},v_n)\) and \(D'(u_{n},v_n)\), the limiting behaviour of the bivariate maxima \(\left( M_{n}^{(1)}, M_{n}^{(2)}\right)\), under linear normalization, is given in Theorem 3.1, as if the sequence \(\{(X_{j},Y_{j})\}\) would be a sequence of independent \((X_j,Y_j)\).
In Theorem 3.1 we derived upper and lower bounds of the limiting distribution of the maximum term of non-negative integer-valued moving average sequences which leads to a “quasi max-stable” limiting behavior of the bivariate maximum in the sense of Anderson’s type. So the main result of the maximum of this bivariate discrete random sequence is the following.
Theorem 4.1
Consider the stationary sequences \(\{(X_{j},Y_{j})\}\) defined by
Suppose that the innovation sequence \(\{(V_i,W_i)\}\) is an iid sequence of non-negative integer-valued random vectors with df of the form (4), the sequences of \(\{\alpha _i\}\) and \(\{\beta _i\}\) satisfy (3) and \(\alpha _{\max }\) and \(\beta _{\max }\) are unique. Then,
for all real x and y and where \(u_n(x)\) and \(v_n(y)\) are defined by (24) and (25).
To prove this theorem, it remains to show that the conditions \(D(u_n, v_n)\) and \(D'(u_n, v_n)\) hold with \(u_n\) and \(v_n\) given by (24) and (25).
Proof of \(D(u_n, v_n)\):
Let \(1\le i_1 \le \cdots \le i_p < j_1 \le \cdots \le j_q \le n\) with \(j_1-i_p> 2 \ell _n\), with separation \(\ell _n=n^\phi\), where \(\phi <1\). We select \(\phi\) later. We use the following notation:
and
Note that
and
are independent.
a) We have as upper bound
where \(M^{(1,1)}_n=\max _{0\le j\le n}X_j^{\star \star }\), \(\,\,M^{(1,2)}_n=\max _{0\le j\le n}Y_j^\star\), \(M^{(2,1)}_n= \max _{0\le j\le n}X_j^{\star }\), and \(M^{(2,2)}_n=\max _{0\le j\le n}Y_j^{\star \star }\).
We split furthermore this upper bound.
The last four terms in (29) tend to 0 as it is proved in Hall (2003) depending on \(\ell _n\). We show it for one term.
for some generic constant C and \(\{\alpha _k\}\) satisfying (3) with \(\delta >2\). Selecting \(\phi >1/(\delta -1)\), this bound tends to 0. The sum of the bounds of the last four terms in (29) gives the bound \(\alpha _{n,\ell _n}=Cn\ell _n^{1-\delta }\), which tends to 0.
b) In the same way we establish the lower bound of (28). In fact, using again the independence mentioned in (27), we get
using (29) and (30). Hence the condition \(D(u_n,v_n)\) holds.
In the proof of \(D^\prime (u_n,v_n)\), we need also that \(s_n\alpha _{n,\ell _n}\rightarrow 0\). With \(s_n=n^\zeta\) we select \(\zeta\) such that \(s_n n \ell _n^{1-\delta }= n^{1+\zeta - \phi (\delta -1) }\rightarrow 0\), which holds for \(1+\zeta < \phi (\delta -1)\).
Proof of \(D^\prime (u_n,v_n)\):
We have to consider first the sums on the terms \(P\left( X_0>u_n,Y_j>v_n\right)\) and on the terms \(P\left( Y_0>v_n,X_j>u_n\right)\).
We show it for the sum of the first terms, since for the second one the proof follows in the same way. Let \(\gamma _n=n^{\nu }\) with \(\nu < 1-\zeta\), which implies that \(\gamma _n =o(n/s_n) =o(n^{1-\zeta })\). For \(j<2\gamma _n\), we write
Note that \(\alpha _{i_0}=\alpha _{\max }\) for some \(i_0\) and \(\beta _{j_0}=\beta _{\max }\) for some \(j_0\). For one j we have \(i_0+j=j_0\), i.e. \(j= j_0-i_0\). Hence the maximum terms occur at the same index for \(V_{-i_0}\) and \(W_{-i_0}\) if \(j=j_0-i_0\). If \(j_0=i_0\), hence \(j=0\), but this case does not occur in the sum. For all other j’s the maxima is occurring at different indexes. We consider the bound established in Proposition 3.1 and 3.2 for \(H^{*}\).
For \(j=j_0-i_0\), we showed in the proof of Theorem 3.1 that \(nH^*(u_n,v_n) \rightarrow 0.\)
For \(j \ne j_0-i_0\), we have \(\beta _{i_0+j}< \beta _{\max }\) for the terms \(P(X_0>u_n, Y_j> v_n)\) and deduce from Proposition 3.1 the following upper bound for \(H^*(u_n, v_n)\)
with \(\rho ,\psi \in (0,1)\) defined in (12) and (13). Note that \(\rho =\rho (j)\) should be such that \(\left( 1+ \frac{\lambda _{2\theta }}{\beta _{i_0+j}}\right) ^{\rho (j)} < 1+ \frac{\lambda _{2}}{\beta _{\max }}\), for all \(j\ne j_0-i_0\) that (13) is satisfied. It means that the term B in (13) depends on j, i.e. \(B=B_j\). Note that \(B_j\) may be larger or smaller than 1, but is bounded above by \(\log (1+\lambda _2/\beta _{\max })/\log (1+\lambda _{2\theta }/\beta _{\max })=B^*\). For \(B_j>1\), we select \(\epsilon <1\) large such that \((1-\epsilon )B^*<1\), which implies that \((1-\epsilon )B_j <1\), thus we select \(\rho(j)> (1-\epsilon )B_j\). In case \(B_j \le 1\), we select also \(\rho(j)> (1-\epsilon )B_j\).
It implies that there exists an \(\epsilon >0\) to select \(\rho (j)\) for every \(j\ne j_0-i_0\) such that
a) Now the sum of the first term in the bound (31) of \(H^*(u_n,v_n)\) multiplied by n, for \(\{j\le 2\gamma _n, j \ne j_0\}\), is bounded by
if also \(\nu\) is such that \(\nu < 1-\epsilon\).
The sum of the second term in (31) multiplied by n, for \(\{j\le 2\gamma _n, j\ne j_0\}\), tends to 0 because
if also \(\nu < 1-\psi\). Hence we choose \(\nu < \min \{1-\epsilon , 1-\zeta , 1-\psi \}\).
It remains to deal with the sum of the third terms in (31) for \(\{j\le 2\gamma _n, j\ne j_0\}\). We showed that \(P(S>\psi x)= O(( 1+\lambda )^{-\psi x})\) in (20) with \((1+\lambda )^\psi > 1+\frac{\lambda _1}{\alpha _{\max }}\) in (12). Let \(\tilde{\psi }>1\) such that \((1+\lambda )^{\psi /\tilde{\psi }} = 1+\frac{\lambda _1}{\alpha _{\max }}\). This sum on \(\{j\le 2\gamma _n, j\ne j_0\}\) multiplied with n is bounded by
if also \(\nu < \tilde{\psi }-1\) and C is a generic positive constant.
Thus combining these three bounds it shows that
if \(\nu < \min \{1-\epsilon , 1-\zeta , 1-\psi , \tilde{\psi }-1\}\).
b) We consider now the sum on j with \(2\gamma _n<j \le n/s_n\) and write
and
Note that \(X_0^\prime\) and \(Y_j^\prime\) are independent. We have, for \(j>2\gamma _n\) and some \(k>1\) (chosen later, not depending on n),
Similar to Hall (2003), the last two probabilities are sufficiently fast tending to 0. For, we have
We select \(h_n\) such that \(h_n\gamma _n^{1-\delta } =C>0,\) for some constant C. For \(i\le -\gamma _n -1\) and \(\delta >2\) and some positive constant \(C^*\), it follows that
by the assumption (3) on the sequence \(\{\alpha _i\}\). It implies again that
where the expectations exist, and, due to Lemma 3.1,
by the choice of \(h_n\). Note that \(h_n=C\gamma _n^{\delta -1}=Cn^{\nu (\delta -1)}\rightarrow +\infty\). Now, select k depending on \(\delta\), \(\nu\) and \(\zeta\) such that \(n^2/((1+h_n)^{k}s_n)\sim n^2/(C^kn^{k\nu (\delta -1)}n^\zeta )=o(1)\) which holds for \(k> (2-\zeta )/(\nu (\delta -1))\). This choice implies that \((n^2/s_n) P\left( X_0^{\prime \prime }\ge k\right) \rightarrow 0\). In the same way we can show that also \(n \sum _{j\le n/s_n} P\left( Y_j^{\prime \prime }\ge k\right) \rightarrow 0\) for such a k, since also \(\beta _i\le C\,|i|^{-\delta }\) for \(|i|\ge \gamma _n\) and some constant \(C>0\).
c) In order to deduce
we use the same arguments as for \(P\left( X_0> u_n, Y_j>v_n\right)\). In this case, since \(X_0^{\prime }\) and \(X_j^{\prime }\) are independent, we get for some positive k
As above we can show that \(n \sum _{j\le n/s_n} P\left( X_j^{\prime \prime }\ge k\right) \rightarrow 0\) and \((n^2/s_n) P\left( X_0^{\prime \prime }\ge k\right) \rightarrow 0\). In the same way it follows also that
Hence condition \(D'(u_n, v_n)\) holds.
5 Simulations
We investigate the convergence of the distribution of the bivariate maxima \((M_n^{(1)}, M_n^{(2)})\) to the limiting distribution as given in Theorem 4.1. We notice that the thinning coefficients \(\alpha _i\) and \(\beta _i\) have an impact on the norming values of the bivariate maxima, besides of the distribution of the \((V_i,W_i)\).
Let us consider the bivariate geometric distribution for \((V_i,W_i)\) mentioned in Example 2.1 and a finite number of positive values \(\alpha _i\) and \(\beta _i\). As mentioned, the bivariate geometric distribution satisfies the conditions of the general assumptions of the joint distribution of \((V_i,W_i)\). We assumed a strong dependence with \(p_{00}=0.85, p_{01}=0.03, p_{10}=0.02\) and \(p_{11}=0.1\).
We consider quite different models with different \(\alpha _i\) and \(\beta _i\) to investigate the convergence rate. Let in the first case \(\alpha _1=0.8, \alpha _2=0.6, \alpha _3=0.4, \beta _1=0.6, \beta _2=0.45, \beta _3=0.3\) and \(\alpha _i=0=\beta _i\) for \(i>3,\) and in the second case \(\alpha _1=0.6, \alpha _2=0.35, \alpha _3=0.1, \beta _1=0.5, \beta _2=0.3, \beta _3=0.1\) and \(\alpha _i=0=\beta _i\) for \(i>3.\)
For each of these first two models we simulated 10’000 time series, selected \(n=100\) and 500 and derived the bivariate maxima \((M_n^{(1)},M_n^{(2)}\)). Thus we compared the empirical (simulated) distribution functions (cdf) with the asymptotic cdf.
We plotted two cases with \(P(M_n^{(1)}-\tilde{u}_n\le x ,M_n^{(2)}-\tilde{v}_n\le x+\delta )\) where \(\tilde{u}_n= u_n-x\) and \(\tilde{v}_n=v_n -y\) with \(u_n, v_n\) given in (24) and (25), respectively, using \(\delta =0\) and 2 (see Figs. 1 and 2).
We notice from these simulations that the convergence rate is quite good, but it depends on the dependence, which is given by the thinning factors \(\alpha _i\) and \(\beta _i\). We find that the convergence rate is slower for the more dependent time series (the first case, Fig. 1) and that the factor \(\delta\) has a negligible impact. This is even more clear in the second cases shown in Fig. 2.
In some additional models we considered larger and more thinning factors different from 0. We show the simulations of the cases with \(\alpha _i=(0.7)^i, \beta _i=(0.6)^i\), for \(i\le 25\), and also with \(\alpha _i=(0.9)^i, \beta _i=(0.8)^i\), for \(i\le 40\). These cases are close to a infinite MA series, since \(\alpha _i,\beta _i\) are very small for \(i> 26\) or \(i>41\), respectively. It means that such small values have an impact on the maxima. We figured out that the number of positive values is not so important. However, in these cases the second largest value of \(\alpha _i\) or \(\beta _i\) is closer to the maximal value (=1), in particular in the second of these additional models. Considering the results of again 10’000 simulations (Fig. 3), we show that the convergence rates are quite slower than in the first two models (Figs. 1 and 2). We show the results of the two cases with \(n=100\) and 500 with \(\delta =0\) only. We also figured out from the simulations of other models and distributions that if the correlation of the two components of the sequence is stronger, then the convergence to the limiting distribution (with asymptotic independence) is slower.
References
Anderson C (1970) Extreme value theory for a class of discrete distributions with applications to some stochastic processes. J Appl Prob 7:99–113
Falk M, Hüsler J, Reiss R (1990) Laws of Small Numbers: Extremes and Rare Events. Birkhäuser, Basel, Boston
Hall A (2003) Extremes of integer-valued moving averages models with exponential type-tails. Extremes 6:361–379
Hall A, Temido MG (2007) On the maximum term of MA and max-AR models with margins in Anderson’s class. Theory Probab Appl 51:291–304
Hsing T (1989) Extreme value theory for multivariate stationary sequences. J Multi Analysis 29:274–291
Hüsler J (1990) Multivariate extreme values in stationary random sequences. Stoch Processes Appl 35:99–108
Mitov K, Nadarajah S (2005) Limit distributions for the bivariate geometric maxima. Extremes 80:357–370
Pedeli X, Karlis D (2011) A bivariate INAR(1) process with application. Stat Model 11:325–349
Quoreshi AMMS (2006) Bivariate time series modeling of financial count data. Commun Stat Theory Methods 35(7):1343–1358
Sagitov S (2017) Tail generating functions for extendable branching processes. Stoch Proc Appl 127:1649–1675
Silva I, Silva ME, Torres C (2020) Inference for bivariate integer-valued moving average models based on binomial thinning operation. J Appl Stat 47:2546–2564
Steutel F, van Harn K (1979) Discrete analogues of selfdecomposability and stability. Ann Probab 7:893–899
Funding
The work of the second author was partially supported by the Centre for Mathematics of the University of Coimbra - UIDB/00324/2020, funded by the Portuguese Government through FCT/MCTES. The work of the third author was supported by the Center for Research and Development in Mathematics and Applications (CIDMA, University of Aveiro) through the Portuguese Foundation for Science and Technology (FCT - Fundação para a Ciência e a Tecnologia), reference UIDB/04106/2020. The work of the first author was partially supported by both projects.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Proofs
Proofs
Proof of Proposition 2.1:
Observe first that
By using the representation (4) for v and w large, and \(w>v> A\) (some large constant A) we deduce, for \(v\le w-1\),
where \(\ell ^*(v,w)\longrightarrow \lambda _2(1+ \lambda _1 -\theta )\), as \(v, w \rightarrow +\infty\).
For \(v \ge w+1\), with \(v>w> A\) (some large constant A), the steps are similar, with \(\ell ^*(v,w)\) such that
For \(v= w >A, \;A\) large, we get with similar steps as above
where \(\ell ^*(v,v)\longrightarrow {\lambda _1\lambda _2 + \theta -1}>0\) as \(v\rightarrow +\infty\), a positive constant by assumption. \(\Box\)
Proof of Proposition 2.2:
Denoting B(n, p) a binomially distributed random variable with parameters n and p, we have for the pgf of (X, Y)
Proof of Lemma 2.1:
We have
The first partial sum is finite. The second one is bounded by \(\frac{(s_1^{m+1}-1)}{s_1-1} \sum _{\ell =n+1}^{+ \infty } s_2^\ell P(W=\ell )\) which is finite for \(s_2 < 1+\lambda _2\) due to (5). Analogously, the third one is bounded by \(\frac{(s_2^{n+1}-1)}{s_2-1} \sum _{k=m+1}^{+ \infty } s_1^k P(V=k)\) which is finite for \(s_1 < 1+\lambda _1\).
Finally, for the last partial sum we use Proposition 2.1. For simplicity we write \(g_i(k)=s_i^k(1+\lambda _{i})^{-k}k^{\xi _{i+2}}L_{i+2}(k)\), for \(k \in \mathbb {N}\), \(i=1,2\). Then, for large \(k,\ell\) and some positive constants \(C_1,C_2\), we get that this sum is bounded by
The sums \(\sum _{\ell \ge k} g_2(\ell )\) and \(\sum _{k\ge \ell } g_1(k)\) are finite by applying the ratio criterium for \(g_2(\ell )\) and \(g_1(k)\). These sums are bounded by \(C g_2(k)\) and \(C g_1(\ell )\) , respectively, with C a generic constant if \(s_i < (1+\lambda _i)\). Then the convergence of the last sum is obtained for \(s_1s_2 < \frac{(1+\lambda _1)(1+\lambda _2)}{\theta }\).
Proof of Proposition 2.3:
Let \(q(k,\ell )=1-F_{(V,W)}(k,\ell )\) and \(p(k,\ell )=P(V=k, W=\ell )\). Then,
Proof of Proposition 2.4:
Write \(a_1=\alpha s_1+1-\alpha\) and \(a_2=\beta s_2 + 1-\beta\). By Proposition 2.2, for \(s_i \ne 1, i =1,2\),
we have
\(\Box\)
Proof of Proposition 2.5:
By Proposition 2.4, and the definition of \(Q_{V,W}\) we have
Proof of Proposition 2.6:
Using Proposition 2.5, \(1-F_{X,Y}(x,y)\) is given by the sum of three terms due to the assumption (4). Each term, defined by double sums, can be determined or bounded by (unique) sums associated to univariate tail functions satisfying Theorem 4 of Hall (2003), see also Hall and Temido (2007). The first sum can be approximated for x and y large, as
The second sum can be dealt with in the same way.
For the third term observe that due to the fact \(\ell (v,w)\) is a bounded function, with bound \(\vartheta\), we get, for large integers x and y,
Proof of Lemma 3.1:
a.) All moments of V exist, since the moment generating function of V exists for small positive values. Applying Taylor’s expansion to the function \(f(1+h)=(1+h)^k,\,h>0\), we get, for \(k \ge 2\),
for some \(0<h_1<h < h^*\), \(h^*\) not depending on k. The expectation \(E(V^2(1+h^*)^{V})\) is finite for \(h^*<\lambda\). Thus \(E(1+h)^V \le 1+hE(V) +h^2E(V^2(1+h^*)^{V})\). Due to the fact that \((1+h)^k > 1+hk\) the proof of the first claim is complete.
b.) Similarly we have
for some \(h^*\) such that \(0<h<h^*<\lambda _1/\alpha _I.\) Then
where \(C_1=E(V)\sum _{i\in I } \alpha _i\) and \(C_2=E(V^2(1+\alpha _Ih^*)^{V})\sum _{i\in I } \alpha _i^2\). \(\Box\)
Rights and permissions
About this article
Cite this article
Hüsler, J., Temido, M.G. & Valente-Freitas, A. On the Maximum of a Bivariate INMA Model with Integer Innovations. Methodol Comput Appl Probab 24, 2373–2402 (2022). https://doi.org/10.1007/s11009-021-09920-3
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11009-021-09920-3