615
Views
1
CrossRef citations to date
0
Altmetric
Articles

Rates of convergence of powered order statistics from general error distribution

, &
Pages 1-29 | Received 03 May 2020, Accepted 04 Nov 2022, Published online: 21 Nov 2022

Abstract

Let {Xn:n1} be a sequence of independent random variables with common general error distribution GED(v) with shape parameter v>0, and let Mn,r denote the r-th largest order statistics of X1,X2,,Xn. With different normalizing constants the distributional expansions and the uniform convergence rates of normalized powered order statistics |Mn,r|p are established. An alternative method is presented to estimate the probability of the r-th extremes. Numerical analyses are provided to support the main results.

1. Introduction

The general error distribution is an extension of the normal distribution. The density of general error distribution, say gv(x), is given by (1) gv(x)=vexp(12|xλ|v)λ21+1vΓ(1v),xR,(1) where v>0 is the shape parameter, and λ=[22/vΓ(1/v)/Γ(3/v)]1/2 with Γ() denoting the Gamma function. It is known that g2(x)=12πex2/2 is the standard normal density and g1(x)=12e2|x| is the density of Laplace distribution.

For the general error distribution with parameter v (denoted by GED(v)), Nelson (Citation1991) used it in volatility models since GED(v) can capture the heavy tailedness of high frequency financial time series. Peng et al. (Citation2009) considered the tail behaviour of GED(v). Peng et al. (Citation2010) established the uniform convergence rates of normalized extremes under optimal normalizing constants. Jia and Li (Citation2014) established the higher-order distributional expansions of normalized maximum. Aforementioned studies show that the optimal convergence rate is proportional to 1/logn, similar to the results of Hall (Citation1979) and Nair (Citation1981) on normal extremes. In order to improve the convergence rate of normal extremes, Hall (Citation1980) studied the asymptotics of |Mn,r|p, the powered r-th largest order statistics, and showed that the distribution of normalized Mn,r2 converges to its limit at the rate of 1/(logn)2 under optimal normalizing constants, while the convergence rates are still the order of 1/logn in the case of p2. For more details, see Hall (Citation1980). For more work on higher-order expansions of powered-extremes from normal samples, see Li and Peng (Citation2018) and Zhou and Ling (Citation2016). For other related work on distributional expansions of extremes, see Liao and Peng (Citation2012) for lognormal distribution, Liao et al. (Citation2014aCitation2014b) for logarithmic general error distribution and skew-normal distribution, and Hashorva et al. (Citation2016), Liao and Peng (Citation2014Citation2015), Liao et al. (Citation2016) and Lu and Peng (Citation2017) for bivariate Hüsler-Reiss models.

In extreme value theory and its applications, it is important to know the convergence rate of distribution of normalized maximum to its ultimate distribution, cf. Cao and Zhang (Citation2021), Hall (Citation1979Citation1980), Leadbetter et al. (Citation1983) and Nair (Citation1981). For more advanced topics related to extreme value theory and its applications, see Zhang (Citation2021) for an excellent review. Motivation of this paper is to consider the higher-order expansions and the uniform convergence rate of powered order statistics from GED(v) sample, and an alternative method to estimate the probability related to powered order statistics, which are the complete extensions the results given by Hall (Citation1980) for GED(2) case. Let X1,X2,,Xn be independent and identically distributed random variables with common distribution function Gv following GED(v) with density given by (Equation1). For positive integer r, let Mn,r denote the r-th largest order statistics of X1,X2,,Xn and Mn=Mn,1=max1kn{Xk} for later use. It is known that the distributional convergence rate of normalized maximum may depend heavily on the normalizing constants, see Hall (Citation1979Citation1980), Leadbetter et al. (Citation1983), Nair (Citation1981) and Resnick (Citation1987) for normal samples. For the GED(v) distribution and positive power index p, it is necessary to discuss how to find the normalizing constants cn>0 and dn such that (2) limn(P{|Mn,r|pcnx+dn}Λr(x))=0,(2) where Λr(x)=Λ(x)j=0r1ejx/j! with Λ(x)=exp(exp(x)),xR, and further work on distributional expansions and uniform convergence rates of |Mn,r|p with different normalizing constants.

For v>0 with normalizing constants αn and βn given by (3) αn=21/vλv(logn)11/v,βn=21/vλ(logn)1/v21/vλ[((v1)/v)loglogn+log{2Γ(1/v)}]v(logn)11/v,(3) Peng et al. (Citation2009) showed that (4) MnβnαndM,(4) where M follows the Gumbel extreme value distribution Λ(x). With normalizing constants αn and βn given by (5) αn=pαnβnp1,βn=βnp,(5) we will show that (Equation2) holds by replacing cn and dn by αn and βn, respectively, and investigate further its higher-order expansions and the uniform convergence rates. Similarly, with pv the optimal convergence rates of |Mn,r|p are derived under the following normalizing constants (6) cn=2pv1λvbnpv,dn=bnp,(6) where constant bn is the solution of the equation (7) 21/vλ1vΓ(1/v)bnv1exp(bnv2λv)=n.(7) Note that for the normal case, it follows from (Equation6) that cn=pbnp2 and dn=bnp since v = 2 and λ=1, which are just the normalizing constants given by Hall (Citation1980).

For the normal case, Hall (Citation1980) showed that the optimal convergence rate of Mn,r2 is the order of 1/(logn)2 if we choose the normalizing constants cn=2(1bn2) and dn=bn22bn2. For the powered r-th largest order statistics |Mn,r|p from the GED(v) sample, it follows from (Equation6) that the convergence rate can be improved if p = v. By Equation (3.1) of Lemma 1 in Jia and Li (Citation2014), for v(0,1)(1,+) and large x we have (8) 1Gv(x)=2λvv{1+2(v11)λvxv+4(v11)(v12)λ2vx2v+8(v11)(v12)(v13)λ3vx3v+O(x4v)}x1vgv(x).(8) Similar to Hall (Citation1980), as p = v we choose the optimal normalizing constants cn and dn as follows. (9) {dn=bnv+4(v11)λ2vbnv,cn=f(bnv)=2λv+4(v11)λ2vbnv,(9) where bn is given by (Equation7). Note that if v = 2 and λ=1, for the normal case, cn=22bn2 and dn=bn22bn2 are just the normalizing constants given by Hall (Citation1980).

The rest of this paper is organized as follows. Section 2 provides the main results and Section 3 presents some numerical analyses. Auxiliary lemmas are deferred to Section 4. Section 5 gives the proofs of the main results.

2. Main results

In this section, we provide the higher-order distributional expansions and the uniform convergence rates of powered extremes under different normalizing constants. Furthermore, we showed a method to estimate the probabilities of the extremes. Throughout this paper, let Λr(x)=Λ(x)j=0r1ejx/j! for positive integer r and Λr(x)=0 for r0. Recall that the power index p is positive.

Theorem 2.1

Let {Xn:n1} be a sequence of independent random variables with common distribution Gv(x),v>0, and Mn,r denotes the r-th maximal term of {X1,X2,,Xn}. Then,

  1. if v = 1 and p = 1, with normalizing constants αn and βn given by αn=212 and βn=212logn2, we have limnn{n[P(|Mn,r|pαnx+βn)Λr(x)]e(r+1)x[(r1)ex1]2(r1)!Λ(x)}=e(r+2)x24(r1)![(3r3+10r29r+2)e2x+(9r211r+2)ex+3ex9r+1]Λ(x);

  2. if v = 1 and p1, with normalizing constants αn and βn given by αn=p2p2(logn2)p1 and βn=(212logn2)p, we have limn(logn){(logn2)[P(|Mn,r|pαnx+βn)Λr(x)](1p)x2erx2(r1)!Λ(x)}=(1p)x3erx24(r1)![4(12p)3(1p)rx+3(1p)xex]Λ(x);

  3. if v(0,1)(1,+), with normalizing constants αn and βn given by (Equation5), we have limn(loglogn){logn(loglogn)2[P(|Mn,r|pαnx+βn)Λr(x)](1v1)3erx2(r1)!Λ(x)}=(1v1)2(1+xlog2Γ(1/v))erx(r1)!Λ(x);

  4. if v(0,1)(1,+) and pv, with normalizing constants cn and dn given by (Equation6), we have limnbnv{bnv[P(|Mn,r|pcnx+dn)Λr(x)]Λ(x)hv(x)e(r1)x(r1)!}=[qv(x)+(1(r1)ex)hv2(x)2]e(r1)x(r1)!Λ(x),where (10) hv(x)=[v1(vp)λvx22v1(1v)λvx2(v11)λv]ex,(10) and (11) qv(x)=[12λ2vv2(vp)2x4+v2(vp)λ2v(243v43p)x32v2(1v)λ2vx2124(v11)(v12)λ2vx4(v11)(v11)λ2v]ex;(11)

  5. if v(0,1)(1,+) and p = v, with normalizing constants cn and dn given by (Equation9), we have limnbnv{bn2v[P(|Mn,r|vcnx+dn)Λr(x)]Λ(x)Sv(x)e(r1)x(r1)!}=Λ(x)B(x)e(r1)x(r1)!,where (12) Sv(x)=2(v11)λ2v[x22(v12)x(3v15)]ex,(12) and (13) B(x)=43(v11)λ3v[(4v1)(v11)x36(v12)x26(3v15)x+(2v222v1+32)]ex.(13)

Remark 2.1

Theorem 2.1 (i)–(ii) show the difference of the convergence rates for the powered-extremes of the Laplace distribution as p = 1 and p1, respectively. Meanwhile, it follows from (Equation3) and (Equation5)–(Equation7) that cn=αn and dn=βn as v = 1 since λ=23/2, so it is not necessary to consider the case of v = 1 in Theorem 2.1 (iv)–(v).

Remark 2.2

For pv, with normalizing constants cn and dn given by (Equation6), Theorem 2.1 (iv) shows that the convergence rate of P(|Mn|pcnx+dn) to the extreme value distribution Λ(x) is proportional to 1/logn since bnv2λvlogn by (Equation7), while it can be improved to the order of 1/(logn)2 with optimal choice of normalizing constants cn and dn given by (Equation9) as p = 2, which coincides with the normal case studied by Hall (Citation1980).

Theorem 2.2

Let {Xn:n1} be a sequence of independent random variables with common distribution Gv(x),v>0, and Mn,r denotes the r-th maximal term of {X1,X2,,Xn}. The following results hold.

  1. If v = 1 and p = 1, with normalizing constants αn and βn given by αn=212 and βn=212logn2, then supxR|P(|Mn,r|pαnx+βn)Λr(x)| is the order of 1/n.

  2. If v = 1 and p1, with normalizing constants αn and βn given by αn=p2p2(logn2)p1 and βn=(212logn2)p, then supxR|P(|Mn,r|pαnx+βn)Λr(x)| is the order of 1/logn.

  3. If v(0,1)(1,+) and p>0, with normalizing constants αn and βn given by (Equation5), then supxR|P(|Mn,r|pαnx+βn)Λr(x)| is the order of (loglogn)2/logn.

  4. If v(0,1)(1,+) and pv, with normalizing constants cn and dn given by (Equation6), then supxR|P(|Mn,r|pcnx+dn)Λr(x)| is the order of 1/logn.

  5. If v(0,1)(1,+) and p = v, with normalizing constants cn and dn given by (Equation9), then supxR|P(|Mn,r|pcnx+dn)Λr(x)| is the order of 1/(logn)2.

Theorem 2.3

Let bn>0 be defined by (Equation7) and λn(x)=λv121vΓ(1v)nx1vexp{xv2λv}.If v>1 and xbn, then (14) Q1n(x)=exp{λn(x)(12(v1)vλvxv+4(v1)(2v1)v2λ2vx2v+λn(x)2(n1))}<P(Mnx)<Q2n(x)=exp{λn(x)(12(v1)vλvxv)}.(14) If 0<v<1, x>λ[2(1v)v]1v and xbn, then (15) Q1n(x)=exp{zz22n(1zn)1}<P(Mnx)<Q2n(x)=exp{λn(x)},(15) where z=λn(x)(1+2(v1)vλvxv)1.

Similarly, for the bounds of the r-th order statistics with r>1, we have (16) Q1n,r(x)P(Mn,rx)<Q2n,r(x),(16) where (17) {Q1n,r=j=0r1λnj(x)(1+2(v1)λvxvv)jj!exp{(1jn)λn(x)j(j1)n},Q2n,r=j=0r1λnj(x)j!exp{(1jn)λn(x)(12(v1)λvxvv)},(17) for v>1, and for v(0,1), (18) {Q1n,r(x)=j=0r1λnj(x)j!exp{(1jn)λn(x)(1+2(v1)λvxvv)1j(j1)n},Q2n,r(x)=j=0r1λnj(x)(1+2(v1)λvxvv)jj!exp{(1jn)λn(x)}.(18)

3. Numerical analyses

In this section, small numerical analyses are provided to support the main results. We compare the actual values of probability of powered order statistics with its higher order expansions provided by Theorem 2.1, and with two bounds based on Theorem 2.3.

First we compare the actual values of P(|Mn,r|pαnx+βn) with the following asymptotics.

  1. If v = 1 and p = 1, the second-order and the third-order asymptotics are respectively given by Λr(x)+e(r+1)x[(r1)ex1]2n(r1)!Λ(x) and Λr(x)+e(r+1)x[(r1)ex1]2n(r1)!Λ(x)+e(r+2)x24n2(r1)!Λ(x)[(3r3+10r29r+2)e2x+(9r211r+2)ex+3ex9r+1].

  2. If v = 1 and p1, the second-order and the third-order asymptotics are respectively given by Λr(x)+(1p)x2erx2logn2(r1)!Λ(x) and Λr(x)+(1p)x2erx2logn2(r1)!Λ(x)+(1p)x3erx24(r1)!lognlogn2Λ(x)[4(12p)3(1p)rx+3(1p)xex].

  3. If v(0,1)(1,+), the second-order and the third-order asymptotics are respectively given by Λr(x)+(loglogn)2(1v1)3erx2(r1)!lognΛ(x) and Λr(x)+(loglogn)2(1v1)3erx2(r1)!lognΛ(x)(loglogn)logn(1v1)2(1+xlog2Γ(1v))×erx(r1)!Λ(x).

  4. If v(0,1)(1,+) and pv, the second-order and the third-order asymptotics are respectively given by Λr(x)+Λ(x)hv(x)e(r1)xbnv(r1)! and Λr(x)+Λ(x)hv(x)e(r1)xbnv(r1)!+e(r1)xbn2v(r1)!Λ(x)[qv(x)+(1(r1)ex)hv2(x)2].

  5. If v(0,1)(1,+) and p = v, the second-order and the third-order asymptotics are respectively given by Λr(x)+Λ(x)Sv(x)e(r1)xbn2v(r1)! and Λr(x)+Λ(x)Sv(x)e(r1)xbn2v(r1)!+Λ(x)B(x)e(r1)xbn3v(r1)!.

For different parameters v and p, with sample size n=1000, Figure  shows the relationship between the actual values and the three asymptotics mentioned above with parameter r = 1 and given interval x[5,10], which supports our findings. For the case of r>1, it is difficult to calculate the actual value directly, so we estimate the actual values by calculating the empirical distribution function. For given x[6,6], Figure  shows the difference between the estimated actual values with the three asymptotics as r = 2, which may be acceptable.

Figure 1. Actual values and its approximations with n = 1000, r = 1, x[5,10]. The actual values compared with the first-order asymptotics, the second-order asymptotics and the third-order asymptotics.

Figure 1. Actual values and its approximations with n = 1000, r = 1, x∈[−5,10]. The actual values compared with the first-order asymptotics, the second-order asymptotics and the third-order asymptotics.

Figure 2. Actual values and its approximations with r = 2, x[6,6]. The actual values (estimated by empirical distribution) compared with the first-order asymptotics, the second-order asymptotics and the third-order asymptotics.

Figure 2. Actual values and its approximations with r = 2, x∈[−6,6]. The actual values (estimated by empirical distribution) compared with the first-order asymptotics, the second-order asymptotics and the third-order asymptotics.

To end this section, we compare the two bounds given by Theorem 2.3 with the actural values of extremes.

For Mn, calculate Q1n(x), Q2n(x), P(Mnx) and relative error en(x)=[Q2n(x)Q1n(x)]/[1Q2n(x)], where Q1n(x) and Q2n(x) are given by (Equation14) and (Equation15), respectively. Tables  show the results. For the case of 0<v<1, Table  shows that for given v = 0.2, 0.5, with increasing sample size n, en(x) decreases; and for given n, with increasing v, en(x) decreases. For given n and v, en(x) is a decreasing function of x. Table  shows the fact that for given x = 5, en(x) is a decreasing function of n and v. For similar facts for the case of v>1, see Tables  and  .

Table 1. Comparison of Q1n(x), Q2n(x) and P(Mnx) along with n and x for given v = 0.2 and v = 0.5, respectively.

Table 2. Comparison of Q1n(x), Q2n(x) and P(Mnx) along with n and 0<v<1 for given x = 5.

Table 3. Comparison of Q1n(x), Q2n(x) and P(Mnx) along with n and x for given v = 2 and v = 5, respectively.

Table 4. Comparison of Q1n(x), Q2n(x) and P(Mnx) along with n and v>1 for given x = 3.5.

For the r-th order statistics Mn,r, calculate Q1n,r(x), Q2n,r(x), P(Mn,rx) and en,r(x)=[Q2n,r(x)Q1n,r(x)]/[1Q2n,r(x)] with Q1n,r(x) and Q2n,r(x) given by (Equation17) for v>1, and (Equation18) for 0<v<1. Table  shows that en,r(x) is an increasing function of n and x for given v = 0.5 and r = 2, and it is irregular for the case of v = 5; the two bounds are closer to the actual values. For given x = 2.8 and n = 30, Table  shows that en,r(x) is an increasing function of r for given v. Tables  also show the fact that two relative accurate bounds control the range of the probability of the extremes, and provide an alternative to linear interpolation. It is an effective method to estimate the probability of the extremes.

Table 5. Comparison of Q1n,r(x), Q2n,r(x) and P(Mn,rx) along with n and x for given r = 2, and v = 0.5 and v = 5, respectively.

Table 6. Comparison of Q1n,r(x), Q2n,r(x) and P(Mn,rx) along with v and r2 for given x = 2.8 and n = 30.

4. Auxiliary lemmas

Let C be a positive constant with values varying from place to place. In order to prove the main results, we need some auxiliary lemmas.

Lemma 4.1

Let Gv(x) denote the GED(v) distribution function with parameter v>0. We have the following results.

  1. If v = 1 and p = 1, with normalizing constants αn and βn given by αn=212 and βn=212logn2, (19) 1G1(αnx+βn)=n1ex.(19)

  2. If v = 1 and p1, with normalizing constants αn and βn given by αn=p2p2(logn2)p1 and βn=(212logn2)p, (20) 1G1((αnx+βn)1p)=n1ex{1(1p)x22logn2+(1p)[3(1p)x4(12p)]x324(logn2)2+o(1(logn2)2)}.(20)

  3. If v(0,1)(1,+), with normalizing constants αn and βn given by (Equation5), (21) 1Gv((αnx+βn)1p)=n1ex{(1v1)2(1+xlog{2Γ(1v)})loglognlogn1(1v1)3(loglogn)22logn+(1v1)2(1+xlog{2Γ(1v)})loglognlogn+o(loglognlogn)}.(21)

  4. If v(0,1)(1,+) and pv, with normalizing constants cn and dn given by (Equation6), (22) 1Gv((cnx+dn)1p)=n1ex{431[v1(vp)λvx22v1(1v)λvx2(v11)λv]bnv+[12λ2vv2(vp)2x4v2(vp)λ2v(243v43p)x3+2v2(1v)λ2vx2+124(v11)(v12)λ2vx+4(v11)(v12)λ2v]bn2v+o(bn2v)}.(22)

  5. If v(0,1)(1,+) and p = v, with normalizing constants cn and dn given by (Equation9), (23) 1Gv((cnx+dn)1p)=n1ex{4312(v11)λ2v[x22(v12)x3v1+5]bn2v+(v11)λ3v[43(v11)(4v1)x38(v12)x28(3v15)x]bn3v+83(v11)(v211v1+16)λ3vbn3v+o(bn3v)}.(23)

Proof.

(i) If v = 1 and p = 1, by (Equation1) the Laplace density g1(x) is given by (24) g1(x)=12exp(2|x|).(24) By (Equation24) and the values of αn and βn given by Theorem 2.1(i), we get 1G1(αnx+βn)=12212x+212logn2exp(2x)dx=n1ex.(ii) If v = 1 and p1, note that (25) (αnx+βn)1p=logn22(1+xlogn2+(1p)x22(logn2)2+(1p)(12p)x36(logn2)3+o(1(logn2)3)).(25) The claimed result (Equation20) follows from (Equation24) and (Equation25).

(iii) If v(0,1)(1,+), with normalizing constants αn and βn given by (Equation5), zn,p(x)=(αnx+βn)1p and zv,n=v1vloglogn+log2Γ(1/v), we have (26) zn,p1v(x)=21vvλ1v(logn)1vv{1+(v11)[xlog2Γ(1v)]+(v11)2(loglogn)logn(v11)3(loglogn)22(logn)2+o((loglognlogn)2)},(26) and gv(zn,p(x))=vλ21+1vΓ(1v)exp[(logn)(1zv,nvlogn)v(1+pxvlognzv,n)vp]=vexλ21vn(logn)v1vexp[(1v1)32(loglogn)2+(1v1)2(log2Γ(1v)x)loglognlogn+o(loglognlogn)]=vexλ21vn(logn)v1v[1(1v1)32(loglogn)2+(1v1)2(log2Γ(1v)x)loglognlogn+o(loglognlogn)].Further, (27) 1+2(v11)λv(zn,p(x))v+4(v11)(v12)λ2v(zn,p(x))2v=1+v11logn(v11)2loglognlog2n+o(loglognlog2n).(27) Combining (Equation8) and (Equation26)–(Equation27), we derive (Equation21).

(iv) If v(0,1)(1,+) and pv, with normalizing constants cn and dn, we have zn,p(x)=(cnx+dn)1p=bn(1+2pv1λvxbnv)1p.By using arguments similar to (Equation26)–(Equation27), we have (28) zn,p1v(x)=bn1v(1+2v1(1v)λvxbnv+2(1v)(1vp)v2λ2vx2bn2v+o(1bn2v)),(28) and by (Equation7), (29) gv(zn,p(x))=n1exv2λvbnv1{1v1(vp)λvx2bnv4λ2vv2(vp)(v2p)x33λ2vv2(vp)2x46bn2v+o(1bn2v)}.(29) Further, (30) 1+2(v11)λv(zn,p(x))v+4(v11)(v12)λ2v(zn,p(x))2v=1+2(v11)λvbnv+4(v11)(v12)λ2v4(v11)λ2vxbn2v+o(1bn2v).(30) Combining (Equation8) and (Equation28)–(Equation30), we derive the desired result (Equation22).

(v) If v(0,1)(1,+) and p = v, with normalizing constants cn and dn given by (Equation9), let zn,v(x)=(cnx+dn)1v=bn(1+2λvbnvx+4(v11)λ2v(x+1)bn2v)1v.By arguments similar to (Equation26)–(Equation27), we can get (31) (zn,v(x))1v=bn1v(1+2(v11)λvxbnv+4(v11)2λ2v(x+1)+2(v11)(v12)λ2vx2bn2v+8(v11)2(v12)λ3vx(x+1)bn3v+o(1bn3v)),(31) and by (Equation7), (32) gv(zn,v(x))=vex2λvn1bnv1{12(v11)λv(x+1)bnv+2(v11)2λ2v(x+1)2bn2v4(v11)3λ3v(x+1)33bn3v+o(1bn3v)}.(32) Further, (33) 1+2(v11)λv(zn,v(x))v+4(v11)(v12)λ2v(zn,v(x))2v+8(v11)(v12)(v13)λ3v(zn,v(x))3v=1+2(v11)λvbnv+4(v11)(v12)λ2v4(v11)λ2vxbn2v+8(v11)λ3vx28(v11)(3v15)λ3vxbn3v+8(v11)λ3v(v26v1+7)bn3v+o(1bn3v).(33) Combining (Equation8) and (Equation31)–(Equation33), we derive (Equation23).

Lemma 4.2

Let {Xn:n1} be a sequence of i.i.d. random variables with common distribution Gv(x) with parameter v>0 and Mn,r denotes the r-th maximal term of {X1,X2,,Xn} for 1rn. Assume that there exists positive constant zn(x) such that when n(1Gv(zn(x))ex, (34) P(|Mn,r|pznp(x))Λr(x)=Λ(x)[112(1θn,v(x))(r1ex)](1θn,v(x))erx(r1)!+O(n1),(34) where θn,v(x)=nex(1Gv(zn(x))).

Proof.

By arguments similar to Hall (Citation1980), we have P(|Mn,r|pznp(x))Λr(x)=j=0r1(nj)[1n1exθn,v(x)]nj[n1exθn,v(x)]jΛr(x)+O(nr12n)=j=0r1[1n1exθn,v(x)]nθn,vj(x)ejxj!Λr(x)+O(n1)=Λ(x)j=0r1{exp[(1θn,v(x))ex]}[1(1θn,v(x))]jejxj!Λr(x)+O(n1)=Λ(x)j=0r1[1+(1θn,v(x))ex+(1θn,v(x))22e2x]×[1j(1θn,v(x))+j(j1)2(1θn,v(x))2]ejxj!Λr(x)+O(n1)=Λ(x)(1θn,v(x))(Λr(x)Λr1(x))ex12Λ(x)(1θn,v(x))2[Λr1(x)Λr2(x)(Λr(x)Λr1(x))]e2x+O(n1)=Λ(x)[112(1θn,v(x))(r1ex)](1θn,v(x))erx(r1)!+O(n1)since (35) j=0r1jejxj!=exΛr1(x),j=0r1j2ejxj!=e2xΛr2(x)+exΛr1(x).(35) The proof is completed.

Lemma 4.3

Let {Xn:n1} satisfy the assumptions of Lemma 4.2 and Mn,r denotes the r-th largest order statistics of X1,X2,,Xn. The following results hold.

  1. If v = 1 and p1, with normalizing constants αn and βn given by αn=p2p2(logn2)p1 and βn=(212logn2)p, let βn=2loglogn2, and then (36) supxβn|P(|Mn,r|pαnx+βn)Λr(x)|Clogn.(36)

  2. If v(0,1)(1,+), with normalizing constants αn and βn given by (Equation5), let βn=2log(logn(loglogn)2), and then (37) supxβn|P(|Mn,r|pαnx+βn)Λr(x)|C(loglogn)2logn.(37)

  3. If v(0,1)(1,+) and pv, with normalizing cn and dn given by (Equation6), let βn=2log(an1bnv1), and then (38) supxβn|P(|Mn,r|pcnx+dn)Λr(x)|Clogn.(38)

  4. If v(0,1)(1,+) and p = v, with normalizing cn and dn given by (Equation9), let βn=4log(an1bnv1), and then (39) supxβn|P(|Mn,r|pcn+dn)Λr(x)|C(logn)2.(39)

Proof.

Note that (40) Δn(x)=|P(|Mn,r|pcnx+dn)Λr(x)|=|j=1r1(nj)Gnj(zn)[1G(zn)]jΛr(x)+O(nr12n)|rnr1(12)nr+Λr(βn),(40) and (41) Λr(βn)=Λ(βn)j=0r1ejβnj!=CΛ(βn)1erβn1rβnCΛ(βn)e(r1)βnCeβn.(41) If v = 1 and pv, note that βn=2loglogn2 and rnr1(12)nr=o(nα) for any α>0. It follows from (Equation40) and (Equation41) that (Equation36) holds. If v(0,1)(1,+), and βn=2log(logn(loglogn)2), from (Equation40) and (Equation41) we can get (Equation37). If v(0,1)(1,+) and pv, recall that anbn1v1logn and βn=2log(an1bnv1), and then (Equation38) follows. If v(0,1)(1,+), p = v, and βn=4log(an1bnv1), from (Equation40) and (Equation41) we can get (Equation39).

The following is about the Mills' type inequalities of GED.

Lemma 4.4

  1. For 0<v<1, as x>[2(1v)v]1v we have 2λvvx1v<1G(x)gv(x)<2λvvx1v(1+2(v1)λvvxv)1.

  2. For v>1, for all x>0 we have 2λvvx1v(1+2(v1)λvvxv)1<1Gv(x)gv(x)<2λvvx1v.

Proof.

Note that assertion (ii) is just Lemma 2.2 of Peng et al. (Citation2009). We only show that assertion (i) holds. For x>[2(1v)v]1v as 0<v<1, 1xvxetv2dt>x1tvetv2dt=v2(1v)xetv2dtx1v1vexv2implies that (42) xetv2dt<(v1x+vxv12)1exp(xv2)(42) and (43) xetv2dt>(vxv12)1exp(xv2).(43) It follows from (Equation42) that 1Gv(x)=v21+1vΓ(1v)x/λetv2dt>v21+1vΓ(1v)(v2(xλ)v1)1exp(xv2λv)=gv(x)2λvvx1v.Similarly, by (Equation43), 1Gv(x)<gv(x)2λvvx1v(1+2(v1)λvvxv)1.Hence, the assertion follows.

Recall that zn,p(x)=(cnx+dn)1p and for large enough n, zn,p(x)>0 and zn,p(x)>λ[2(1v)v]1v.

Lemma 4.5

  1. If v = 1 and p1, let αn=logn24(p+1) and |x|αn, and then |exn(1Gv(zn,p(x)))|Clogn.

  2. If v(0,1)(1,+) and p>0, let αn=an1bn4(v+2)(1+δnanbn1)rn and |x|αn, and then |exn(1Gv(zn,p(x)))|C(x2+|x|+1)δnexp(14|x|x),where δn=βnbnan(v1)3v2(loglogn)2logn,rn=αn/an, and rn1(v1)3v2loglognlogn.

  3. If v(0,1)(1,+) and pv, let αn=an1bn4(p+2v) and |x|αn, and then |exn(1Gv(zn,p(x)))|C(x2+|x|+1)anbn1vexp(14|x|x).

  4. If v(0,1)(1,+) and p = v, let αn=an1bn4(v+2) and |x|αn, and then |exn(1Gv(zn,p(x)))|C(|x|+1)(anbn1v)2ex.

Proof.

(i) If v = 1 and p1, then zn,p(x)=logn22(1+pxlogn2)1p. It follows from Lemma 1 of Hall (Citation1980) and (Equation20) that exn[1G(zn,p(x))]exn2exp{logn2(1+1logn2x+2(logn2)2x2)}2x2exlogn2Clogn,and n[1G(zn,p(x))]exn2exp{logn2(1+xlogn2px2(logn2)2)}exexpx2logn2exp{px2logn2}e34xpx2logn2Clogn.Combining the above results, we can get (i).

(ii) In case of v>1, zn,p(x)=bn(1+panbn1rn1+δnanbn1)1p(1+δnanbn1). By Lemma 1 of Hall (Citation1980), Lemma 4.4 and some tedious calculation, we have exn[1G(zn,p(x))]2exanbn1v(|x|+1+x2)+exδnand n[1G(zn,p(x))]exn21vλv1Γ(1v)zn,p1v(x)exp(zn,pv(x)2λv)ex[(v1)anbn1|x|+panbn1x2]exp{(v1)anbn1|x|+panbn1x2x}vanbn1(v|x|+px2)exp(14+14|x|x),which implies (ii) for the case v>1.

If 0<v<1, it follows from Lemma 1 of Hall (Citation1980) and Lemma 4.4 that exn[1G(zn,p(x))]exn21vλv1Γ(1v)zn,p1v(x)exp(zn,pv(x)2λv)ex{1exp{2vanbn1x2}(1+(1v)δnanbn1)}ex{2vanbn1x2+2v(1v)δn(anbn1x)2}2x2anbn1vex+2v(1v)16δnex,and n[1G(zn,p(x))]exn21vλv1Γ(1v)zn,p1v(x)exp(zn,pv(x)2λv)(1+2(v1)λvvzn,pv(x))1exex{exp[(1v)anbn1xanbn1(v1)(1vanbn1x)+panbn1x2δn]1}Canbn1v(|x|+1+x2)exp(14+14|x|x)+|δn|exp(14+14|x|x),which implies (ii) for the case 0<v<1.

(iii) If pv, in case of v>1, zn,p(x)=bn(1+anbn1px)1p. It follows from Lemma 1 of Hall (Citation1980) and Lemma 4.4 that exn[1G(zn,p(x))]exn21vλv1Γ(1v)zn,p1v(x)exp{zn,pv(x)2λv}(12(v1)λvvzn,pv(x))ex(1(v1)anbn1x)exp{x2anbn1x2}×[12(v1)λvvbnv(1anbn1vx+x2(anbn1)2(8v2+pv))]ex{(8+pv)1(1(v1)anbn1x)exp{2anbn1x2}×[1(v1)anbn1(1anbn1vx+(anbn1v)2x2(8+pv))]}ex{1[1anbn1v(|x|+2)]exp(2anbn1x2)}ex{1[1anbn1v(|x|+2)](12anbn1x2)}2exanbn1v(|x|+1+x2)since (1(v1)anbn1x)[1(v1)anbn1(1anbn1vx+(anbn1v)2x2(8+pv))]1anbn1v(|x|+2)as |x|αn. Next, n[1G(zn,p(x))]exn21vλv1Γ(1v)zn,p1v(x)exp(zn,pv(x)2λv)ex(1+panbn1x)1vpexp(x+panbn1x2)ex[(v1)anbn1|x|+panbn1x2]exp{(v1)anbn1|x|+panbn1x2x}vanbn1(v|x|+px2)exp(14+14|x|x),which implies (iii) for the case v>1.

If pv, in case of 0<v<1, it follows from Lemma 1 of Hall (Citation1980) and Lemma 4.4 that exn[1G(zn,p(x))]exn21vλv1Γ(1v)zn,p1v(x)exp(zn,pv(x)2λv)ex[1+(1v)anbn1x(1v)(anbn1)2px2]exp(x2vanbn1x2)ex[1exp(2vanbn1x2)]2x2anbn1vex,and n[1G(zn,p(x))]exn21vλv1Γ(1v)zn,p1v(x)exp(zn,pv(x)2λv)(1+2(v1)λvvzn,pv(x))1exex{exp[(1v)anbn1anbn1(v1)(1vanbn1x)+panbn1x2]1}[(1v)anbn1|x|+(1v)anbn1(1v)v(anbn1)2|x|+panbn1x2]×exp{(1v)anbn1|x|+(1v)anbn1(1v)v(anbn1)2|x|+panbn1x2x}[anbn1|x|+anbn1v(anbn1)2|x|+panbn1x2]×exp{anbn1|x|+anbn1v(anbn1)2|x|+panbn1x2x}anbn1v(1v|x|+1vanbn1|x|+pvx2)exp(14+14|x|x)Canbn1v(|x|+1+x2)exp(14+14|x|x),which implies (iii) for the case of 0<v<1.

(iv) If p = v, the proof is similar to the case of (iii). If v>1 and zn,p(x)=bn[1+vanbn1x+v(1v)(anbn1)2(1+x)]1v, we have exn[1G(zn,p(x))]ex[1(v1)anbn1x+(v1)2(anbn1)2(x+1)]×exp{x+(v1)(anbn1)(1+x)}[1(v1)anbn1]ex{1[1vanbn1xvanbn1(vanbn1)3(x+1)][1+vanbn1(1+x)]}2(anbn1v)2ex(x2+|x|+1),and n[1G(zn,p(x))]exex{exp{(1v)anbn1x+(1v)2(anbn1)2(1+x)+(v1)anbn1(1+x)}1}[(v1)anbn1+(v1)2(anbn1)2(1+|x|)]exp{(v1)anbn1+(v1)2(anbn1)2(1+|x|)x}C(|x|+1)(anbn1v)2ex.Therefore, the result in (iv) for the case of v>1 can be proved.

If p = v, in case of 0<v<1, exn[1G(zn,p(x))]ex{1exp[(1v)anbn1(1+x)]}Cexanbn1(1+|x|),and n[1G(zn,p(x))]exex{exp[(1v)anbn1|x|+(1v)2(anbn1)2(1+x)(1v)anbn1(1+x)+anbn1(1v)(1vanbn1xv(1v)(anbn1)2(1+x))]1}ex{exp[(1v)2(anbn1)2(1+x)v(1v)(anbn1)2x+v(1v)2(anbn1)3(1+x)]1}[(1v)2(anbn1)2(1+|x|)v(1v)(anbn1)2|x|+v(1v)2(anbn1)3(1+|x|)]×exp{(1v)2(anbn1)2(1+|x|)v(1v)(anbn1)2|x|+v(1v)2(anbn1)3(1+|x|)}C(|x|+1)(anbn1v)2ex.Hence the result of (iv) can be proved for 0<v<1.

Lemma 4.6

  1. If v = 1 and p1, let αn=logn24(p+1) and xαn, and then |exn[1Gv(zn,p(x))]|exp{Clogn2}.

  2. If v(0,1)(1,+), let αn=an1bn4(p+2v)1+δnanbn1rn and xαn, and then |exn[1Gv(zn,p(x))]|Cexp{Can1bnv1}.

  3. If v(0,1)(1,+) and pv, let αn=an1bn4(p+2v) and xαn, and then |exn[1Gv(zn,p(x))]|Cexp{Can1bnv1}.

  4. If v(0,1)(1,+) and p = v, let αn=an1bn4(v+2) and xαn, and then |exn[1Gv(zn,p(x))]|Cexp{C(an1bnv1)2}.

Proof.

(i) We can get the assertion of (i) since n[1Gv(zn,p(αn))]n2exp{logn2(1+1logn2αnp(logn2)2(αn)2)}exp{34αn}exp{Clogn2}.(ii) In case of v>1, n[1Gv(zn,p(αn))](1+panbn1αnrn1+δnanbn1)1vp×exp{bnv2λv[1(1+anbn1pαnrn1+δnanbn1)vp(1+δnanbn1)v]}exp{Can1bnv1},and for the case 0<v<1, n[1Gv(zn,p(αn))](1+panbn1αnrn1+δnanbn1)1vp×exp{bnv2λv[1(1+panbn1αnrn1+δnanbn1)vp(1+δnanbn1)v]}×(1+2(v1)λvvbnv(1+panbn1αnrn1+δnanbn1)vp(1+δnanbn1)v)1Cexp{bnv2λv[1(5p+8v4p+8v)vp]}Cexp{Can1bnv1}.Hence, we can get (ii).

(iii) If pv, in case of v>1, n[1Gv(zn,p(αn))](1+panbn1αn)1vpexp{bnv2λv[1(1+anbn1pαn)vp]}exp{Can1bnv1},and for the case 0<v<1, n[1Gv(zn,p(αn))](1+anbn1pαn)1vpexp{bnv2λv[1(1+anbn1pαn)vp]}×(1+2(v1)λvvbnv(1+panbn1αn)vp)1Cexp{bnv2λv[1(5p+8v4p+8v)vp]}Cexp{Can1bnv1}.Therefore, we can get (iii).

(iv) If p = v, in case of v>1 n[1Gv(zn,p(αn))][1+vanbn1αnv(v1)(anbn1)2(1+αn)]1vv×exp{bnv2λv[1(1+vanbn1αn+v(1v)(anbn1)2(1+αn))]}exp{bnv2λv[1(1+vanbn1αn+v(1v)(anbn1)2(1+αn))]}exp{bnv2λv[vanbn1αn+v(v1)(anbn1)2(1+αn)]}Cexp{bnv2λvαn+(v1)anbn1(1+αn)}Cexp{C(an1bnv1)2},and for the case of 0<v<1, n[1Gv(zn,p(αn))][1+vanbn1αn+v(1v)(anbn1)2(1+αn)]1vv×exp{bnv2λv[1(1+vanbn1αn+v(1v)(anbn1)2(1+αn))]}×{1+2(v1)λvvbnv[1+vanbn1αn+v(1v)(anbn1)2(1+αn)]1}1Cexp{bnv2λv[vanbn1αnv(1v)(anbn1)2(1+αn))]}Cexp{C(an1bnv1)2}.Then, we complete the proof of (iv).

Let An=exp{n(1Gv(zn,p(x)))+ex},Bn(x)=exp{Rn(x)},where Rn(x)=n[1Gv(zn,p(x))]nlogG(zn) and 0<Rn(x)n(1Gv(zn,p(x)))22Gv(zn,p(x)).

Lemma 4.7

  1. If v = 1 and p1, let αn=logn22(p+1) and xαn, and then (1Gv(zn,p(x)))Cn1116,|Bn(x)1|Cn14.

  2. If v(0,1)(1,+) and p>0, let αn=an1bn4(p+2v)1+δnanbn1rn and xαn, and then (1Gv(zn,p(x)))Cn1516,|Bn(x)1|Cn34.

  3. If v(0,1)(1,+) and pv, let αn=an1bn4(p+2v) and xαn, and then (1Gv(zn,p(x)))Cn1516,|Bn(x)1|Cn34.

  4. If v(0,1)(1,+) and p = v, let αn=an1bn4(v+2) and xαn, and then (1Gv(zn,p(x)))Cn34,|Bn(x)1|Cn14.

Proof.

(i) If v = 1 and p1, since xαn, then 1Gv(zn,p(x))1Gv(zn,p(αn))12exp{logn2(11logn2αnp(logn2)2(αn)2)}exp{516logn2}Cn1116,and |Bn(x)1|=|exp{Rn(x)}1|Rn(x)n[1Gv(zn,p(x))]22Gv(zn,p(x))Cn38.Therefore, we can get (i).

(ii) If v>1, since an1bnvlogn, 1Gv(zn,p(x))1Gv(zn,p(αn))n121vλv1Γ(1v)bn1v(1panbn1αnrn1+δnanbn1)1vp×exp{bnv2λv(1panbn1αnrn1+δnanbn1)vp}n1(1panbn1αnrn1+δnanbn1)1vpexp(rn1+δnanbn1αn+panbn1(αn)2)Cn1exp(rn1+δnanbn1αn+panbn1(αn)2)Cn1exp{an1bn16v}Cn1516,and |Bn(x)1|n[1Gv(zn,p(x))]22Gv(zn,p(x))<Cn34.In case of 0<v<1, noting that an1bnvlogn, we have 1Gv(zn,p(x))1Gv(zn,p(αn))n121vλv1Γ(1v)bn1v(1panbn1αnrn1+δnanbn1)1vp×exp[bnv(1panbn1αnrn1+δnanbn1)vp2λv(1+δnanbn1)v]×(1+2(v1)λvvbnv(1panbn1αnrn1+δnanbn1)vp(1+δnanbn1)v)1Cn1exp(αn+panbn1(αn)2)Cn1exp{an1bn16v}Cn1516,and |Bn(x)1|n[1Gv(zn,p(x))]22Gv(zn,p(x))<Cn34.Therefore, we can get (ii).

(iii) If pv, in case of v>1 and noting that an1bnvlogn, we have 1Gv(zn,p(x))1Gv(zn,p(αn))n121vλv1Γ(1v)bn1v(1panbn1αn)1vpexp[bnv(1panbn1αn)vp2λv]n1(1panbn1αn)1vpexp(αn+panbn1(αn)2)Cn1exp(αn+panbn1(αn)2)Cn1exp{an1bn16v}Cn1516,and |Bn(x)1|n[1Gv(zn,p(x))]22Gv(zn,p(x))<Cn34.If pv, in case of 0<v<1, noting that an1bnvlogn, we have 1Gv(zn,p(x))1Gv(zn,p(αn))n21vλv1Γ(1v)bn1v(1panbn1αn)1vpexp[bnv(1panbn1αn)vp2λv]×(1+2(v1)λvvbnv(1panbn1αn)vp)1Cn1exp(αn+panbn1(αn)2)Cn1exp{an1bn16v}Cn1516,and |Bn(x)1|n[1Gv(zn,p(x))]22Gv(zn,p(x))<Cn34.Therefore, we can get (iii).

(iv) If p = v, in case of v>1, noting that an1bnvlogn, we have 1Gv(zn,p(x))1Gv(zn,p(αn))n1[1vanbn1αnv(v1)(anbn1)2(1αn)]1vvexp{αn+(v1)anbn1(1αn)}Cn1exp{αn+(v1)anbn1(1αn)}Cn1exp{an1bn4v}Cn34,and |Bn(x)1|n[1Gv(zn,p(x))]22Gv(zn,p(x))<Cn14.If p = v, in case of 0<v<1, noting that an1bnvlogn, we have 1Gv(zn,p(x))1Gv(zn,p(αn))n1[1vanbn1αn+v(1v)(anbn1)2(1αn)]1vv×exp{αn+(1v)anbn1(1αn))}×[1+2(v1)λvvbnv[1vanbn1αn+v(1v)(anbn1)2(1αn)]1]1Cn1exp{αn(1v)anbn1(1αn)}Cn1exp{an1bn4v}Cn34,and |Bn(x)1|n[1Gv(zn,p(x))]22Gv(zn,p(x))<Cn14.The desired result (iv) follows.

5. Proofs

Proof

Proof of Theorem 2.1

(i) Let zn,1(x)=(αnx+βn) with αn and βn given by Theorem 2.1 (i). By using (Equation19) and (Equation35), and some tedious calculation we have (44) P(|Mn,r|zn,1(x))Λr(x)=Λ(x)(r1)ex12n{(r1)!}e(r+1)x+e(r+2)x24n2{(r1)!}[(3r3+10r29r+2)e2x+(9r211r+2)ex+3ex9r+1]Λ(x)+o(n2).(44) Hence, the desired results follow from (Equation44).

(ii) Let zn,p(x)=(αnx+βn)1/p with αn and βn given by Theorem 2.1(ii). By using (Equation20) we have (45) 1θn,1(x)=(1p)x22logn2(1p)[3(1p)x4(12p)]x324(logn2)2+o((logn)2).(45) It follows from Lemma 4.3 and (Equation45) that (46) P(|Mn,r|pαnx+βn)Λr(x)=Λ(x)[112(1θn,1(x))(r1ex)](1θn,1(x))erx(r1)!+O(n1)=Λ(x){(1p)x22logn2+(1p)x324(logn2)2[4(12p)3(1p)rx+3(1p)xex]}erx(r1)!+o((logn)2).(46) Hence, following (Equation60) we get the desired results.

(iii) Let zn,p(x)=(αnx+βn)1/p with αn and βn given by (iii) of Theorem 2.1. By using (Equation21) we have (47) 1θn,v(x)=(1v1)3(loglogn)22logn(1v1)2(1+xlog2Γ(1v))loglognlogn+o(loglognlogn).(47) It follows from Lemma 4.3 and (Equation47) that (48) P(|Mn,r|pαnx+βn)Λr(x)=Λ(x)[112(1θn,v(x))(r1ex)](1θn,v(x))erx(r1)!+O(n1)=Λ(x)[(1v1)3(loglogn)22logn(1v1)2(1+xlog2Γ(1v))loglognlogn]erx(r1)!+o(loglognlogn).(48) Therefore, following (Equation61), we get the desired results.

(iv) Let zn,p(x)=(cnx+bn)1/p with cn and dn given by (iv) of Theorem 2.1 and hv(x), qv(x) are given by (Equation10) and (Equation11). By using (Equation22) we have (49) 1θn,v(x)=hv(x)exbnv+qv(x)exbn2v+o(bn2v).(49) It follows from Lemma 4.3 and (Equation49) that (50) P(|Mn,r|pαnx+βn)Λr(x)=Λ(x)[112(1θn,v(x))(r1ex)](1θn,v(x))erx(r1)!+O(n1)=Λ(x)[hv(x)bnv+(qv(x)+1(r1)ex2hv2)bn2v]e(r1)x(r1)!+o(bn2v),(50) implying the desired results.

(v) Let zn,p(x)=(cnx+dn)1/v with cn and dn given by Theorem 2.1(v) and Sv(x) and B(x) be given by (Equation12) and (Equation13). By using (Equation23) we have (51) 1θn,v(x)=Sv(x)exbn2v+B(x)exbn3v+o(bn3v).(51) It follows from Lemma 4.3 and (Equation51) that (52) P(|Mn,r|pαnx+βn)Λr(x)=Λ(x)[112(1θn,v(x))(r1ex)](1θn,v(x))erx(r1)!+O(n1)=Λ(x)[Sv(x)bn2v+B(x)bn3v]e(r1)x(r1)!+o(bn3v).(52) Hence, by using (Equation52), we derive the desired results.

Proof

Proof of Theorem 2.2

The lower bounds are from Theorem 2.1. The rest is to derive the upper bounds. By the arguments similar to Hall (Citation1980) and some tedious calculations, we have (53) Δn(x)=|P(|Mn,r|zn,p(x))Λr(x)|=|j=0r1Gvnj(zn,p(x))[n(1Gv(zn,p(x)))]jj!Λr(x)+o(n1)||j=0r1Gvnj(zn,p(x))[n(1Gvnj(zn,p(x)))]jΛ(x)ejx|+C(n1)j=0r1|Gvnj(zn,p(x))Λ(x)|[n(1Gvnj(zn,p(x)))]j+Λ(x)j=0r1|[n(1Gvnj(zn,p(x)))]jejx|+C(n1).(53) For 0j<r, (54) |Gvnj(zn,p(x))Λ(x)||Gvnj(zn,p(x))Gvn(znp(x))|+|Gvn(zn,p(x))Λ(x)||Gvn(zn,p(x))Λ(x)|+o(1n).(54) Note that (55) |Gvn(zn,p(x))Λ(x)|<Λ(x)|An(x)1|+Λ(x)|Bn(x)1|<|exn(1Gv(zn,p(x)))|exp{ex+|exn(1Gv(zn,p(x)))|}+Λ(x)|Bn(x)1|.(55) (i) If v = 1 and p = 1, by (Equation19) and (Equation53)–(Equation55), we have Δn(x)j=0r1ejxΛ(x)|Bn(x)1|+C(n1)and |Bn(x)1|e2x/n. Therefore, combining Theorem 2.1 (i), the assertion of (i) can be proved.

(ii) If v = 1 and p1, by Lemma 4.5 (i) for |x|αn we have (56) Λ(x)j=0r1|[n(1Gv(zn,p(x)))]jejx|<Λ(x)j=1r1|n(1Gv(zn,p(x)))ex|×|ex+(n(1Gv(zn,p(x)))ex)θ|j1(0<θ<1)CΛ(x)|n(1Gv(zn,p(x)))ex|[ex+|n(1Gv(zn,p(x)))ex|+1]r2CΛ(x)|n(1Gv(zn,p(x)))ex|[e(r2)x+|n(1Gv(zn,p(x)))ex|r2+1]Clogn.(56) For xαn, by the arguments similar to those used in (Equation56), it follows from Lemma 4.6(i) that Λ(x)j=0r1|[nΨn(x)]jejx|Clogn.It follows from (Equation53)–(Equation55) that Δn(x)Clogn as x0.

Suppose that βnx0, in view of (Equation54) and Lemma 4.7(i), |Gvn(zn,p(x))Λ(x)|C{1logn2exp(18e|x|)+exp(e|x|)n34}.It follows from (3.19) that (57) j=0r1|Gvnj(zn,p(x))Λ(x)|[n(1Gv(zn,p(x)))]jClogn.(57) Combining with Lemma 4.3(i) and Theorem 2.1(ii), we complete the proof of (ii).

(iii) If v(0,1)(1,+), from Lemma 4.1 (iii), Lemma 4.5 (ii) and the similar arguments used in (Equation56), for |x|αn recall that δn(v1)3v2(loglogn)2logn, and we have Λ(x)j=0r1|[n(1Gv(zn,p(x)))]jejx|Cδn.Next, from Lemma 4.6(ii), for xαn, Λ(x)j=0r1|[n(1Gv(zn,p(x)))]jejx|Canbn1v.Therefore for x0, it follows from (Equation53)–(Equation55) that Δn(x)Cδn.

Note that for the case βnx0, Lemma 4.5 (ii) also implies that |exn(1Gv(zn,p(x)))|C(x2+|x|+1)exp(54|x|)δnC(x2+|x|+1)exp(34|x|)Cexp(78|x|),and in view of (Equation55) and Lemma 4.7(ii), we can get |Gvn(zn,p(x))Λ(x)|C{(x2+|x|+1)δnexp[5|x|4e|x|+C1exp(78|x|)]+Λ(x)n34}C{δnexp(18e|x|)+exp(e|x|)n34}.Now, from (Equation54) and by the similar arguments used in (Equation57), we have j=0r1|Gnj(zn)Λ(x)|[nΨn(x)]jCδn.Hence, combining with Lemma 4.3(ii), Theorem 2.1(iii) and (Equation53)–(Equation55), (iii) can be proved.

(iv) If v(0,1)(1,+) and pv, by Lemma 4.5 (iii) and the similar arguments used in (Equation56), for |x|αn, we have Λ(x)j=0r1|[n(1Gv(zn,p(x)))]jejx|Canbn1v.Next, from Lemma 4.6 (iii), for xαn, Λ(x)j=0r1|[n(1Gv(zn,p(x)))]jejx|Canbn1v.Therefore for x0, it follows from (Equation53)–(Equation55) that Δn(x)Canbn1v.

Note that for the case βnx0, Lemma 4.5 (iii) also implies that |exn(1Gv(zn,p(x)))|C(x2+|x|+1)exp(54|x|)anbn1vC(x2+|x|+1)exp(34|x|)Cexp(78|x|),and in view of (Equation55) and Lemma 4.7 (iii), we can get |Gvn(zn,p(x))Λ(x)|C{(x2+|x|+1)anbn1vexp[5|x|4e|x|+C1exp(78|x|)]+Λ(x)n34}C{anbn1vexp(18e|x|)+exp(e|x|)n34}.Now, from (Equation54) and by the similar arguments used in (Equation57), we have j=0r1|Gnj(zn)Λ(x)|[nΨn(x)]jCanbn1v.Hence, combining with Lemma 4.3 (iii), Theorem 2.1 (iv) and (Equation53)–(Equation55), (iv) can be proved.

(v) If v(0,1)(1,+) and p = v, for x0, it follows from Lemmas 4.5-4.6 (iv) and (Equation53) that Δn(x)C[j=0r1|Gvnj(zn,p(x))Λ(x)|+(anbn1v)2],and combining (Equation54)–(Equation55), we can get Δn(x)C(anbn1v)2.

Note that for βnx0, by Lemma 4.5 (iv) and the similar arguments used in (Equation56), Λ(x)j=0r1|[n(1Gv(zn,p(x)))]jejx|C(anbn1v)2,and in view of (Equation55) and Lemma 4.7 (iv), |Gvn(zn,p(x))Λ(x)|C{(|x|+1)(anbn1v)2exp[|x|e|x|+C]+Λ(x)n14}C{(anbn1v)2exp(e|x|)+exp(e|x|)n14}.Finally, by (Equation54) and the similar arguments used in (Equation57), we have j=0r1|Gvnj(zn,p(x))Λ(x)|[n(1Gv(zn,p(x)))]jC(anbn1v)2.Combining with Lemma 4.3(iv) and Theorem 2.1(v), we finish the proof of (v).

Proof

Proof of Theorem 2.3

For the r-th largest order statistics, we have (58) P(Mn,rx)=j=0r1n!j!(nj)!Gv(nj)(x)(1Gv(x))j,(58) and for any 0<z<n, (59) exp[zz22n(1zn)1]<(1zn)n<ez.(59) We first consider the bounds of P(Mnx) as v>1. It follows from (Equation59) and Lemma 4.4(ii) that P(Mnx)<[121vλv1Γ(1v)x1vexp(xv2λv)(12(v1)λvvxv)]n<exp{λn(x)(12(v1)vλvxv)}.For the lower bound, note that (60) 1Gv(x)=21vλv1Γ(1v)x1vexp(xv2λv)[12(v1)vλvxv+4(v1)(2v1)v2λ2vx2vg(x)](60) with 12(v1)vλvxv+4(v1)(2v1)v2λ2vx2vg(x)<1.By (Equation59) and (Equation60) we have (61) P(Mnx)={1λn(x)[12(v1)vλvxv+4(v1)(2v1)v2λ2vx2vg(x)]n}n>exp{zz22n(1zn)1},(61) where z=λn(x)[12(v1)vλvxv+4(v1)(2v1)v2λ2vx2vg(x)]. Noting that for xbn and 0<z<λn(x)1, we have (62) z22n(1zn)1z22n(11n)1=z22(n1).(62) It follows from (Equation61) and (Equation62) that (63) P(Mnx)>exp{λn(x)(12(v1)vλvxv+4(v1)(2v1)v2λ2vx2v+λn(x)2(n1))},(63) which is the desired lower bound.

For the bounds of the r-th order statistics as v>1, by (Equation58), (Equation59) and Lemma 4.4(ii), we have P(Mn,rx)<j=0r1n!j!(nj)![1λn(x)n(12(v1)λvvxv)]nj(λn(x)n)j<j=0r1λnj(x)j!exp{(1jn)λn(x)(12(v1)λvxvv)}.For the lower bound, by (Equation63), (Equation59) and Lemma 4.4(ii), we have P(Mn,rx)>j=0r1n!j!(nj)!exp{(1jn)λn(x)}[λn(x)n(1+2(v1)λvvxv)1]j>j=0r1λnj(x)(1+2(v1)λvxvv)jj!exp{(1jn)λn(x)j(j1)n}.The remainder is to consider the bounds as 0<v<1. For the bounds of P(Mnx) as 0<v<1, it follows from (Equation59) and Lemma 4.4(i) that P(Mnx)<[121vλv1Γ(1v)x1vexp(xv2λv)]n<exp{λn(x)}and P(Mnx)>[121vλv1Γ(1v)x1vexp(xv2λv)(1+2(v1)vλvxv)1]n>exp{zz22n(1zn)1},where z=λn(x)(1+2(v1)vλvxv)1.

For the bounds of the r-th order statistics, it follows from (Equation59) and Lemma 4.4(i) that P(Mn,rx)<j=0r1n!j!(nj)![1λn(x)n]nj(λn(x)n(1+2(v1)vλvxv)1)j<j=0r1λnj(x)(1+2(v1)λvxvv)jj!exp{(1jn)λn(x)}.By similar arguments used in (Equation63), one can show that P(Mn,rx)>j=0r1n!j!(nj)!exp((1jn)λn(x)(1+2(v1)vλvxv)1)(λn(x)n)j>j=0r1λnj(x)j!exp{(1jn)λn(x)(1+2(v1)λvxvv)1j(j1)n}.Thus, the proof is completed.

Acknowledgements

The authors would like to thank the Editor-in-Chief, the Associated Editor and the two referees for careful reading and comments which greatly improved the paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

References

  • Cao, W., & Zhang, Z. (2021). New extreme value theory for maxima of maxima. Statistical Theory and Related Fields, 5(3), 232–252. https://doi.org/10.1080/24754269.2020.1846115
  • Hall, P. (1979). On the rate of convergence of normal extremes. Journal of Applied Probability, 16(2), 433–439. https://doi.org/10.2307/3212912
  • Hall, P. (1980). Estimating probabilities for normal extremes. Advances in Applied Probability, 12(2), 491–500. https://doi.org/10.2307/1426608
  • Hashorva, E., Peng, Z., & Weng, Z. (2016). Higher-order expansions of distributions of maxima in a Hüsler-Reiss model. Methodology and Computing in Applied Probability, 18(1), 181–196. https://doi.org/10.1007/s11009-014-9407-6
  • Jia, P., & Li, T. (2014). Higher-order expansions for distributions of extremes from general error distribution. Journal of Inequalities and Applications, 2014, 213. https://doi.org/10.1186/1029-242X-2014-213
  • Leadbetter, M. R., Lindgren, G., & Rootzén, H. (1983). Extremes and related properties of random sequences and processes. Springer Verlag.
  • Li, T., & Peng, Z. (2018). Moment convergence of powered normal extremes. Communication in Statistics-Theory and Methods, 47(14), 3453–3463. https://doi.org/10.1080/03610926.2017.1359294
  • Liao, X., Peng, L., Peng, Z., & Zheng, Y. (2016). Dynamic bivariate normal copula. Science China Mathematics, 59(5), 955–976. https://doi.org/10.1007/s11425-015-5114-1
  • Liao, X., & Peng, Z. (2012). Convergence rates of limit distribution of maxima of lognormal samples. Journal of Mathematical Analysis and Applications, 395(2), 643–653. https://doi.org/10.1016/j.jmaa.2012.05.077
  • Liao, X., & Peng, Z. (2014). Convergence rate of maxima of bivariate gaussian arrays to the Hüsler–Reiss distribution. Statistics and Its Interface, 7(3), 351–362. https://doi.org/10.4310/SII.2014.v7.n3.a5
  • Liao, X., & Peng, Z. (2015). Asymptotics for the maxima and minima of Hüsler–Reiss bivariate gaussian arrays. Extremes, 18, 1–14. https://doi.org/10.1007/s10687-014-0196-7
  • Liao, X., Peng, Z., & Nadarajah, S. (2014a). Tail behavior and limit distribution of maximum of logarithmic general error distribution. Communications in Statistics-Theory and Methods, 43(24), 5276–5289. https://doi.org/10.1080/03610926.2012.730168
  • Liao, X., Peng, Z., Nadarajah, S., & Wang, X. (2014b, January). Rates of convergence of extremes from skew-normal samples. Statistics and Probability Letters, 84, 40–47. https://doi.org/10.1016/j.spl.2013.09.027
  • Lu, Y., & Peng, Z. (2017). Maxima and minima of independent and non-identically distributed bivariate Gaussian triangular arrays. Extremes, 20, 187–198. https://doi.org/10.1007/s10687-016-0263-3
  • Nair, K. A. (1981). Asymptotic distribution and moments of normal extremes. Annals of Probability, 9(1), 150–153. https://doi.org/10.1214/aop/1176994515
  • Nelson, D. B. (1991). Conditional heteroskedasticity in asset returns: A new approach. Econometrica, 59(2), 347–370. https://doi.org/10.2307/2938260
  • Peng, Z., Nadarajah, S., & Lin, F. (2010). Convergence rate of extremes for the general error distribution. Journal of Applied Probability, 47(3), 668–679. https://doi.org/10.1239/jap/1285335402
  • Peng, Z., Tong, B., & Nadarajah, S. (2009). Tail behavior of the general error distribution. Communications in Statistics-Thoery and Methods, 38(11), 1884–1892. https://doi.org/10.1080/03610920802478367
  • Resnick, S. I. (1987). Extreme values, regular variation and point processes. Springer-Verlag.
  • Zhang, Z. (2021). On studying extreme values and systematic risks with nonlinear time series models and tail dependence measures. Statistical Theory and Related Fields, 5(1), 1–25. https://doi.org/10.1080/24754269.2020.1856590
  • Zhou, W., & Ling, C. (2016, April). Higher-order expansions of powered extremes of normal samples. Statistics and Probability Letters, 111, 12–17. https://doi.org/10.1016/j.spl.2016.01.003