Extreme values of a portfolio of Gaussian processes and a trend
Ju¨rg Hu¨sler&Christoph M. Schmid
Received: 24 June 2003 / Revised: 1 December 2005 / Accepted: 2 December 2005
#Springer Science + Business Media, LLC 2006
Abstract We consider the extreme values of a portfolio of independent continuous Gaussian processes Pk
i¼1wiXiðtÞ (wi2R; k2N) which are asymptotically locally stationary, with expectationsE½XiðtÞ ¼0 and variancesVar½XiðtÞ ¼dit2Hi ðdi2Rþ; 0<Hi<1Þ, and a trendctfor some constants;c>0 with >Hi. We derive the probabilityPfsupt>0Pk
i¼1wiXiðtÞ ct>ugfor u! 1, which may be interpreted as ruin probability.
Keywords Gaussian processes, Extreme values, Portfolio of assets, Tail behavior, Ruin probability, Large deviations
AMS 2000 Subject Classification Primary— 60G15, 62G32, 91B28 1. Introduction
The tail behavior of stochastic processes is important e.g., for calculating ruin probabilities in insurance or finance. In this context we consider in this paper particular Gaussian processes, to determine the probability that a Gaussian process YðtÞexceeds a certain boundaryu2Rin an intervalT2R
PðuÞ ¼P sup
t2T
YðtÞ>u
:
In general, it is almost impossible to find the distribution of this supremum.
Precise formulas are only known for a couple of stationary processes in a finite or infinite interval (cf. Adler (1990)). The best we can do in general, is to derive the asymptotic behavior of PðuÞ whenu! 1. This asymptotic behavior is sufficiently
DOI 10.1007/s10687-006-7966-9
J. Hu¨sler (*)
:
C. M. SchmidInstitut fu¨r Mathematische Statistik und Versicherungslehre, Universita¨t Bern, Sidlerstrasse 5,
3012 Bern, Switzerland
e-mail: juerg.huesler@stat.unibe.ch C. Schmid
e-mail: christoph.schmid@stat.unibe.ch
interesting for its own. However, we are going to derive the probability that the Gaussian processYðtÞ ¼XðtÞ ct, with trend, exceeds some boundaryu
P sup
t>0
ðXðtÞ ctÞ>u
;
asu! 1 withc; >0. In insurance we may note thatXðtÞ represents the sum of the claims up to timet,ctrepresents the sum of the premium payments up to timet anduthe initial reserve of the firm. The ruin occurs if at some timetthe sum of the claims is larger than the sum of premium payments and the reserve.
This problem is investigated for a class of Gaussian processesXðtÞ(including frac- tional Brownian motion and self-similar Gaussian processes) in Hu¨sler and Piterbarg (1999) and for integrated Gaussian processes in Debicki (2002) as well as Hu¨sler and Piterbarg (2004). In these cases the probability PðuÞ is approximated by exceedances ofYðtÞin a small neighborhood of a unique point where the boundary ðuþctÞ=ðtÞhas smallest value. This boundary is the result of usual standardization of the processXðtÞbyðtÞwhere2ðtÞdenotes its variance. Often, only the (unique) point of maximal variance plays the important role. Here the trend has to be considered also, which results in the mentioned minimal boundary value.
In this paper we deal with another particular class of Gaussian processes. We think that a portfolio consists of many different processes Xið
Þ which can be modelled e.g., as fractional Brownian motions with E½XiðtÞ ¼0 and Var½XiðtÞ ¼ dit2Hi and possibly different parameters Hi2 ð0;1Þ and di>0. Therefore, we consider XðtÞ ¼Pki¼1wiXiðtÞ as the portfolio of all risks at time t, with wið2RÞ some weights. As mentioned, the biggest liability of the firm after its start of economic activities at timet¼0 is then denoted by supt>0ðXðtÞ ctÞ. Thus
P sup
t>0
Xk
i¼1
wiXiðtÞ ct
!
>u
( )
will be investigated. We assume that the Xið
Þ are independent processes. This probability is well-defined ifHi< for alli. As mentioned above, the minima of the boundary function ðuþctÞ=ðtÞ have to be analyzed together with the path behaviour ofXðtÞin the neighborhood of possible minima.Note thatXðtÞ is thus a centered Gaussian process with variance Var½XðtÞ ¼Xk
i¼1
w2idit2Hi¼Xk
i¼1
Wit2Hi
where Wi¼w2idi. Hence, we might set w.l.o.g. di¼1 or wi¼1, since in the following only the Wi_s are used. It is not necessary to assume that the Gaussian processes Xið
Þ are fractional Brownian motions. But certain regularity conditions will be assumed. E.g., we assume that the Gaussian processes are asymptotically locally stationary (see (8) in Condition (A1)) for largeu. The processes with largest Hiare important. Hence, let w.l.o.g.H¼H1H2. . .Hkand definemð1Þas largest index such thatHm¼H.In the next section we introduce the sufficient conditions on the Gaussian process XðtÞand the main result which is proved in the third section. For its proof we need to investigate the local behaviour of the boundary function in the vicinity of the points with minimum value. This will be combined with the behavior of the weighted sumXðtÞof Gaussian processes in these vicinities.
2. Weighted sum of Gaussian processes and main result
The portfolioXðtÞ is modelled as weighted sum of centered independent Gaussian processes XiðtÞ with Var½XiðtÞ ¼t2Hi with the mentioned numeration H¼H1 H2. . .Hk. Defining the standardized process
e
XðtÞ ¼XðtÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XðtÞ
p ¼Xk
i¼1
wiXiðtÞ. ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Xk
i¼1
Wit2Hi vu
ut ;
we analyze
P sup
t>0
ðXðtÞ ctÞ>u
¼P 9t>0 :XðtÞe >~ fuðtÞ
n o
; where
~fuðtÞ ¼ uþct ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pk
i¼1
Wit2Hi
s ¼ uþct
ffiffiffiffiffiffiffiffiffiffiffi Wt2H
p ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1þP
j>m
W1Wjt2ðHjHÞ r
is the boundary function and
W¼Xm
i¼1
Wi: ð1Þ
We make an appropriate time transformation such that the points where the boundary values are minimal, remain finite asu! 1. Let for eachu
s¼W2H1u1t: ð2Þ
The transformed centered Gaussian processes, depending onu, are denoted by XiðuÞðsÞ ¼Xiðu1W2H1sÞ; ik; ð3Þ XðuÞðsÞ ¼Xk
i¼1
wiXiðuÞðsÞ ¼Xðu1W2H1sÞ: ð4Þ The time transformation results in the corresponding boundary functionfuðsÞ:
fuðsÞ ¼~
fuðW2H1u1sÞ ¼u1HvðsÞð1þuðsÞÞ ð5Þ consisting of three factors where
vðsÞ ¼1þ~cs
sH ð6Þ
and
uðsÞ ¼ 1þX
j>m
WjWHjHu2ðHjHÞs2ðHjHÞ
!12
1 ð7Þ
with ~c¼cW2H. Note that uðsÞ !1 as u! 1. The boundary fuðsÞ may have several points with minimal value, depending onu. The smallest of these points is denoted by su*¼inffargminfuðsÞg. We will show that these points with minimal
value converge to the point of minimal value ofvðsÞ. Hence we have to investigate the approximation of the probability:
P sup
t>0
ðXðtÞ ctÞ>u
¼Pn9s>0 :XeðuÞðsÞ>fuðsÞo
;
whereXeðuÞðsÞ ¼XðuÞðsÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XðuÞðsÞ
p denotes the standardized processXðuÞðsÞ.
We need the following assumptions for the main result:
(A1) We assume that each XeðuÞi ðsÞ ¼XiðuÞðsÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XiðuÞðsÞ q
is asymptotically locally stationary forik, i.e., there exists a functionK2ið
Þ, regular varying (at 0) with parameteri2 ð0;2Þ, such thatlim
u!1
E½XeðuÞi ðsÞ XeðuÞi ðs0Þ2
Ki2ðjss0jÞ ¼Di ð8Þ uniformly for s;s02Su¼ ½su*ðuÞ;s*uþðuÞ withDi>0, ðuÞ ¼uH1logu and su* denotes the smallest point with minimal boundary value: s*u¼ inffarg minfuðsÞg.
This condition implies that each Gaussian process Xið
Þ has continuous paths in the crucial interval. We could assume this regularity condition to hold for alls, but this is not necessary.For the variance of the increments ofXXeðuÞðsÞ we consider the weighted sum of the regularly varying functionsKið
Þ. By (A1) there existsK2ðÞ, reg-ularly varying at 0 with index¼minfig, such that for some positiveWWe
lim
s!s0
P
im
WiDiKi2ðjss0jÞ K2ðjss0jÞ ¼WWe:
We denote the inverse ofKð
ÞbyK1ðÞwhereK1ðyÞ ¼inffs:KðsÞ yg.Since Kð
Þ is regularly varying at 0 with index =2, K1ðÞ is regularly varying at 0 with index 2=.Note that for fractional Brownian motions with Hurst parameterHi, we havei¼2HiandKi2ðhÞ ¼ jhj2Hi.
(A2) Forj>m, letK2jð
Þbe such that lim suph#0KjðhÞKðhÞ<1, asu! 1.
This condition implies thatj also forj>m. Note that (A2) holds for jmalso by (A1). We use in the following Pickands constant defined by
H¼lim
T!1
1
TE exp max
0tT ðtÞ
;
where ðtÞ,t0, is a fractional Brownian motion with drift E½ðtÞ ¼ t and covariance functionCov½ðsÞ; ðtÞ ¼tþs jtsj, hence with Hurst parameterH¼=2. Note that here the fractional Brownianð
Þmotion hasvariance 2t.
Now we can state our main result on the extreme values of a portfolio of the Gaussian processesXiðuÞðsÞði¼1;. . .;kÞas defined in (3). Condition (A1) restricts the behavior of these processes in the small interval Su. For the
other time points s we suppose a quite common Ho¨lder condition for the increments. We assume that for eachik
lim sup
u!1
E Xh iðuÞðsÞ XiðuÞðs0Þi2
Gijss0ji: ð9Þ holds for anys;s0withjss0j C, for someGi; i>0 and some largeC. For the asymptotic expression we use the constantsAandB:
A¼ H
~cðHÞ
H
H
B¼ H
~cðHÞ
Hþ2
H;
which arevðs0Þ and v00ðs0Þ, respectively, withs0¼argminvðsÞ ¼ ðH=ð~cð HÞÞ1=, the unique point of minima ofvð
Þ(see Lemma 3.3 below). We can now state the main theorem.Theorem 2.1: Let XiðtÞ;t>0;ði¼1;. . .;kÞ be independent centered continuous Gaussian processes with variance dit2Hi and ct a trend where;c;di>0 and0<
Hk . . . H1<minf1; g.Let wi2R denote the weights. Assume the conditions (A1) and (A2) with0< i < 2and(9).Then, the tail behavior is given by
P sup
t>0
Xk
i¼1
wiXiðtÞ ct
!
>u
( )
ffiffiffiffiWe
W
q A
2
H21 exp 12fu2ðsu*Þ A ffiffiffiffiffiffiffiffi
pAB
u22HK1uH1 ; as u! 1,where ¼minimi,m the number of Hi¼H,fuðsÞ defined in (5)and su*¼inffargminfuðsÞg.
Remark 2.1: In some particular cases we have explicit expressions for su*and fuðsuÞ. For example, ifk¼1, we have H¼H1,¼1,Kð
Þ ¼K1ðÞ,W¼W1¼d1w21, WeW¼W1D1,A¼ ðH=ð~cðHÞÞÞH=ðHÞ andB¼ ðH=ð~
cðHÞÞÞH2 H. K1ðsÞ is regularly varying at 0 with index2=ð0< <2ÞandfuðsÞreduces tofuðsÞ ¼u1HvðsÞ. Hence fu2ðsu*Þ ¼u22Hv2ðs0Þ ¼u22HA2. We note that~c¼cW2H ¼cðd1w21Þ2H ¼c, by choosingd1¼w1¼1. Hence
P sup
t>0
XðtÞ ct
>u
ffiffiffiffiffiffi D1
p A
2
H21expn12A2u22Ho A ffiffiffiffiffiffiffiffi
pAB
u22HK1uH1 ;
which is the result of Hu¨sler and Piterbarg (1999). But this result holds also for a more general Gaussian process, not only for a fractional Brownian motion, if the stated assumptions (A1) and (9) hold.
Remark 2.2 Further Examples: Let us consider some other simple examples withXðtÞsatisfying the conditions of the theorem. We do not need to assume that XiðtÞare fractional Brownian motions.
We begin with two processes XiðtÞ;i¼1;2. Let H¼H1>H2. It implies that m¼1 andW¼W1. Also Kð
Þ ¼K1ðÞ with¼1 andWWe ¼W1D1 and the same result holds as in Remark 2.1.This simple situation holds also with more than two processes, ifH1¼H2¼. . .¼ Hm (k¼m2) and 1 < j;j>1. Then WWe ¼W1D1, ¼1 and K¼K1 in the asymptotic formula.
If in addition to Hj also some of thej_s are equal to 1¼, then the result depend on the possible domination of one of theKj_s. For example, letHj¼H1;for all jm¼k, and 1¼j for jm0m with 1< j for j>m0. In addition, assume that KjcjK1 for jm0 and some cj0, as h!0. Then WWe ¼ P
jm0WjDjcj with c1¼1, and the result of Theorem 2.1 holds with K¼K1, ¼ 1andW¼P
jWj.
If one of the Kj_s ðjm0Þdominates the others, by renumbering let this beK1, then this would mean thatcj¼0 for all 1<jm0. The result holds then also with suchcj_s.
3. Proof
Idea of the proof:
Applying the time transformation (2), the original problem gets P sup
t>0
ðXðtÞ ctÞ>u
n o
¼P 9t>0 :XðtÞ>uþct
¼P 9t>0 :XðtÞe > ~ fuðtÞ
n o
¼Pn9s>0 :XXeðuÞðsÞ>fuðsÞo :
In Proposition 3.2 we show that all minima of fuðsÞ occur in the interval
½su*ðuÞ;s0, wheresu*¼inffargminfuðsÞg. Therefore we splitPf9s>0 :XXeðuÞðsÞ>
fuðsÞginto the probabilities
Pn9s2Su :XXeðuÞðsÞ>fuðsÞo
and Pn9s62Su :XXeðuÞðsÞ>fuðsÞo :
whereSu¼ ½s*uðuÞ;su*þðuÞ. Then we will show that foru! 1 Pn9s62Su :XXeðuÞðsÞ>fuðsÞo
¼o P n9s2Su :XXeðuÞðsÞ>fuðsÞo :
We choose ðuÞ ¼uH1logu, since we need later that ðuÞu1H ! 1 ðu! 1Þ.
Hence it remains to analyze the asymptotic behavior of the leading probability term, wheres2Su. For this proof we need to know the behavior of the portfolio process XðuÞðsÞorXXeðuÞðsÞand the boundary functionfuðsÞ.
3.1. Properties of the portfolio process
By the definition of the portfolio process its behavior can be characterized as follows.
Lemma 3.1: For ik;the means and variances are given by E
XiðuÞðsÞ
¼ E XðuÞðsÞ
¼0;
Var½XiðuÞðsÞ ¼ diu2HW1s2H for im Var½XjðuÞðsÞ ¼ djs2HjWHjHu2Hj ¼oðu2HÞ for j>m
Var½XðuÞðsÞ ¼ u2H=s2Hð1þuðsÞÞ2¼u2H=s2H1þO u 2ðHmþ1HÞ as u! 1.
Proof: The processes are obviously centered. For any im and j>m, the variances are simply
Var½XiðuÞðsÞ ¼ diðu1W2H1sÞ2Hi¼diu2HW1s2H; for im Var½XjðuÞðsÞ ¼ djðu1W2H1sÞ2Hj¼dju2Hj=WHj=Hs2Hj ¼oðu2HÞ;
asu! 1, and
Var½XðuÞðsÞ ¼ P
im
w2iVar½XiðuÞðsÞ þP
j>m
w2jVar½XjðuÞðsÞ
¼ u2Hs2H 1þP
j>m
WjWHjHu2ðHjHÞs2ðHjHÞ
!
¼ u2Hs2Hð1þuðsÞÞ2
¼ u2Hs2H1þO u 2ðHmþ1HÞ
The correlation of XeðuÞðsÞ is given by the regularly varying function K2ð
Þ, inthe intervalSu.
Í
Lemma 3.2: Assume the conditions (A1) and (A2). Then the correlation function of XeðuÞðsÞ ¼XðuÞðsÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Var½XðuÞðsÞ
p is for small lags, for s;s02Su
1Corr½XeðuÞðsÞ;XeðuÞðs0Þ ¼1
2E½XeðuÞðsÞ XeðuÞðs0Þ2 WWe
2WK2ðjss0jÞ:
Proof: Using the fact that XeðuÞðsÞandXeðuÞi ðsÞare standardized, we get
EhXeðuÞðsÞ XeðuÞðs0Þi2
¼E Pk
i¼1
wiXiðuÞðsÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XðuÞðsÞ
p
Pk
i¼1
wiXiðuÞðs0Þ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XðuÞðs0Þ p
2 66 64
3 77 75
2
¼E Xk
i¼1
w2i
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XiðuÞðsÞ Var½XðuÞðsÞ s
e XðuÞi ðsÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XiðuÞðs0Þ Var½XðuÞðs0Þ s
e XðuÞi ðs0Þ 0
@
1 A 2 2
64
3 75:
To derive the claim, we split the sumPk
i¼1 intoP
im andP
j>m. Using Lemma 3.1, we see that forimandu! 1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XiðuÞðsÞ Var½XðuÞðsÞ s
¼ ffiffiffiffiffiffi
di
W r
ð1þuðsÞÞ
and get
E
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XiðuÞðsÞ Var½XðuÞðsÞ s
e XðuÞi ðsÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XiðuÞðs0Þ Var½XðuÞðs0Þ s
e XðuÞi ðs0Þ 2
4
3 5
2
¼ di
WE ð1þuðsÞÞXeðuÞi ðsÞ XeðuÞi ðs0Þ
þ ðuðsÞ uðs0ÞÞXeðuÞi ðs0Þ
h i2
¼ di
Wð1þuðsÞÞ2EhXeðuÞi ðsÞ XeðuÞi ðs0Þi2
þdi
WðuðsÞ uðs0ÞÞ2 þ2di
Wð1þuðsÞÞðuðsÞ uðs0ÞÞE XeðuÞi ðsÞ XeðuÞi ðs0Þ e XðuÞi ðs0Þ
h i
di
WEhXeðuÞi ðsÞ XeðuÞi ðs0Þi2
;
where for the last step we used that XeiðuÞðsÞis asymptotically locally stationary with i2 ð0;2Þ and that uðsÞ uðs0Þ ¼ ðss0Þu0ðÞ ¼ ðss0ÞO u 2ðHmþ1HÞ
for some 2 ðs0;sÞ.
Forj>mwe use Lemma 3.1 again to derive ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Var½XjðuÞðsÞ Var½XðuÞðsÞ vu
ut ¼ ffiffiffiffi
dj
q
sHjWHj2HuHj=.
ðuH=sHð1þuðsÞÞ1Þ
¼ ffiffiffiffi dj
q
W2HHju1ðHjHÞsHjHð1þuðsÞÞ ¼:gjðu;sÞ and
E
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XjðuÞðsÞ Var½XðuÞðsÞ vu
ut XeðuÞj ðsÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Var½XjðuÞðs0Þ Var½XðuÞðs0Þ vu
ut XeðuÞj ðs0Þ 2
64
3 75
2
¼E gjðu;sÞXeðuÞj ðsÞ XeðuÞj ðs0Þ
þgjðu;sÞ gjðu;s0ÞXeðuÞj ðs0Þ
h i2
¼g2jðu;sÞEhXeðuÞj ðsÞ XeðuÞj ðs0Þi2
þgjðu;sÞ gjðu;s0Þ2
þ2gjðu;sÞ gjðu;sÞ gjðu;s0Þ
E XeðuÞj ðsÞ XeðuÞj ðs0Þ e XðuÞj ðs0Þ
h i
¼O g 2jðu;sÞEhXeðuÞj ðsÞ XeðuÞj ðs0Þi2
¼O u 2ðHjHÞKj2ðjss0jÞ
;
as u! 1, where we used that gjðu;sÞ ¼Oðu1ðHjHÞÞ and gjðu;sÞ gjðu;s0Þ ¼ ðss0Þgj0ðu; Þ ¼ ðss0ÞOðu1ðHjHÞÞ for some2 ðs0;sÞ, as well as
2gjðu;sÞðgjðu;sÞ gjðu;s0ÞÞE½ðXeðuÞj ðsÞ XeðuÞj ðs0ÞÞXeðuÞj ðs0Þ
¼O u 2ðHjHÞjss0jKj2ðjss0jÞ Putting the various terms together results in
EhXeðuÞðsÞ XeðuÞðs0Þi2
¼ ð1þoð1ÞÞX
im
w2i di
WEhXeðuÞi ðsÞ XeðuÞi ðs0Þi2
þX
j>m
O u 2ðHjHÞK2jðjss0jÞ WWe
WK2ðjss0jÞ;
using that for allj>m
O u 2ðHjHÞKj2ðjss0jÞ
¼o K 2ðjss0jÞ sinceHj<HandKjðhÞ=KðhÞ ¼Oð1Þforhsmall.
Hence forularge, the correlation function ofXeðuÞðsÞbehaves fors!s0as 1Corr½XeðuÞðsÞ;XeðuÞðs0Þ ¼1
2EhXeðuÞðsÞ XeðuÞðs0Þi2
WWe
2WK2ðjss0jÞ:
3.2. Behavior of the boundary function
Í
We need to analyze the behavior of the boundary function. Examining uðsÞ, we see that lims!0uðsÞ ¼ 1, lims!1uðsÞ ¼0 ands2ðHjHÞstrictly decreases to zero, ass! 1, since Hj<Hforj>m. HenceuðsÞstrictly increases from1 to 0. With s0¼argmin vðsÞ, it is straightforward to prove the following lemma (given also in Hu¨sler and Piterbarg (2004)) which implies the next proposition by (5).
Lemma 3.3: We get s0¼ ðH
ð~cðHÞÞÞ1 as well as v0ðsÞ<0 for s<s0 and v0ðsÞ>0for s>s0. Hence s0is unique as point of minimal value of vð
Þ.Furthervðs0Þ ¼ H
~cðHÞ
H
H¼:A and
v00ðs0Þ ¼ H
~cðHÞ
Hþ2
H¼:B:
Proposition 3.1: For u! 1we get
fuðs0Þ ¼u1HAð1þuðs0ÞÞ ¼u1HA1þO u 2ðHmþ1HÞ
; fu00ðs0Þ ¼u1HB1þO u 2ðHmþ1HÞ
; taking the derivative w.r.t. s.
Proof: Taking derivatives (w.r.t. s), we get ð1þuðsÞÞ0¼Oðu2ðHmþ1HÞÞ and ð1þuðsÞÞ00¼Oðu2ðHmþ1HÞÞforson a compact interval ð0;1Þ. Thus
fu00ðsÞ ¼u1H½v00ðsÞð1þuðsÞÞ þ2v0ðsÞð1þuðsÞÞ0þvðsÞð1þuðsÞÞ00
¼u1Hv00ðsÞh1þO u 2ðHmþ1HÞi
;
asu! 1, fors2 ðs0;^ s0þÞ^ with >^ 0 such thatv00ðsÞ>0. Together with Lem-
ma 3.3 the statements follows.
Í
Remark 3.1: The boundary function fuðsÞ is continuous, has the limits lims!0 fuðsÞ ¼lims!1fuðsÞ ¼ 1 and at least one minimum, tending to 1, as u! 1.
Unfortunately,fuðsÞ is a non-algebraic function and thus explicit solutions for the minimum points offuðsÞdo not exist in the general casek>1. Further it is unclear whether the global minimum is unique.
Because of Remark 3.1 we need an upper and a lower simple approximation function offuðsÞ. Let us define
u ¼X
j>m
WjWHjHu2ðHjHÞ: ForðuÞ>0, we introduce the functions
fuþðsÞ ¼u1HvðsÞð1þuðs0þðuÞÞÞ ðs>0Þ and for some smalls1<minf1;s0ðuÞg
fuðsÞ ¼
fu;1ðsÞ ¼u1HvðsÞð1þ us2ðHkHÞÞ12 ð0<ss1Þ fu;2ðsÞ ¼u1HvðsÞð1þuðs1ÞÞ ðs1<ss0ðuÞÞ fu;3ðsÞ ¼u1HvðsÞð1þuðs0ðuÞÞÞ ðs>s0ðuÞÞ:
8>
>>
<
>>
>:
The constant s1 is chosen such that fu;1ðsÞ is strictly decreasing in ð0;s1Þ. This holds because
@
@sfu;1ðsÞ ¼ u1H sðs2Hþ us2HkÞ32
~csþ2Hþ~c usþ2HkHs2HHk us2Hk~cHsþ2H~cHk usþ2Hk is negative if~cðHÞsþ2ðHHkÞ þ ~c uðHkÞsHs2ðHHkÞHk u < 0, which is true ifs<s1 for somes1>0 small enough.
Since 1þuðsÞis strictly increasing, we have
fuðsÞ < fuðsÞ < fuþðsÞ if s2 ð0;s0þðuÞÞ ð10Þ for anyðuÞ 0, and
fuðsÞ fuþðsÞ>fuðsÞ if s2 ½s0þ 2 ðuÞ;1Þ; ð11Þ with equality fors¼s0þðuÞ.
Lemma 3.4: The global minimum/minima of fuðsÞis/are in the interval sð 1;s0. Proof: Since fuðsÞis strictly increasing for s>s0, we have mins>s0fuðsÞ ¼fuðs0Þ hence the minima are smaller thans0. Nowfu;1ðsÞis strictly decreasing inð0;s1Þand fu;1ðsÞ ! 1fors!0. Since s1<s0 we have with (10)
min
ss0
fuðsÞ f g min
ss1
fu;1ðsÞ
n o
< min
ss1
fuðsÞ
f g
for anyu, which finishes the proof.
Í
Even if there are more than one point with minimal boundary value, they all converge tos0 which is shown next since this holds for the smallest of these points, denoted bys*u.
Proposition 3.2: We have su*!s0for u! 1.Hence the global minimum/minima points are inðsu*;s0Þ.
Proof: Lemma 3.4 guarantees that su*s0. It remains to show for any >0 that if there exists a sequence uðiÞ ! 1such thatsuðiÞ* < s0leads to a contra- diction. Because s0 minimizes vðsÞ and sinces*uðiÞ< s0 we havevðs*uðiÞÞ=vðs0Þ>
1þDðÞfor some DðÞ>0. Further we know that ð1þuðiÞðs*uðiÞÞÞ=ð1þuðiÞðs0ÞÞ ¼ 1þoð1ÞsinceuðiÞð
Þ ¼oð1Þ, asuðiÞ ! 1. HencefuðiÞðs*uðiÞÞ ¼fuðiÞðs0Þvðs*uðiÞÞð1þuðiÞðs*uðiÞÞÞ vðs0Þð1þuðiÞðs0ÞÞ
>fuðiÞðs0Þð1þDðÞÞð1þoð1ÞÞ
>fuðiÞðs0Þ:
So we get fuðiÞðs*uðiÞÞ>fuðiÞðs0Þ, which cannot be true sinces*uðiÞis the smallest of
the possible minimal points.
Í
Combining the proof of Proposition 3.1 with the above Proposition 3.2 gives Proposition 3.3: For u! 1we get
fuðsu*Þ u1HA fu00ðsu*Þ u1HB:
3.3. Tail behavior ofXðuÞðsÞfors2Su
We derive the probability thatXðuÞðsÞexceedsfuðsÞfors2Suby applying a result of Bra¨ker (1993a), given also in Bra¨ker (1993b), similar to Hu¨sler and Piterbarg (1999).
This probability will be the major contribution to the investigated probability, asymptotically. Bra¨ker’s result (formulated below) is given for locally stationary Gaussian processes, being not dependent onu. So a further approximation step is necessary sinceXðuÞðsÞdepends onu.
Proposition 3.4: Assume (A1) and (A2). Then with the correlation function Kð
Þof XðuÞðsÞwe have for u! 1Pn9s2Su:XXeðuÞðsÞ>fuðsÞo
ffiffiffiffiWe
W W W
q A
2
H21exp12fu2ðs*uÞ A ffiffiffiffiffiffiffiffi
pAB
u22HK1uH1 :
Proof: To apply Bra¨ker’s Theorem (Bra¨ker, 1993a), we need to approximate e
XðuÞðsÞ by Gaussian processes UþðsÞ and UðsÞ which are independent of u. The original probability will then be estimated applying Slepian’s inequality (Adler,1990).
As in Hu¨sler and Piterbarg (1999), standardized Gaussian processes UþðsÞ and UðsÞ exist by the assumptions in Lemma 3.2 such that
lim
s!s0
E½UðsÞ Uðs0Þ2 K2ðjss0jÞ
" #
¼WWe Wð1Þ for any >0, with correlation function given by
1Corr½UðsÞ;Uðs0Þ WWe
2Wð1ÞK2ðjss0jÞ;
fors!s0. By construction we have fors;s02Su andularge Corr½UðsÞ;Uðs0ÞyCorr½XeðuÞðsÞ;XeðuÞðs0Þ
for any >0, sinceðuÞ !0 foru! 1. Applying Slepian’s inequality we get Pn9s2Su :XeðuÞðsÞ>fuðsÞo
Pf9s2Su :UþðsÞ>fuðsÞg Pn9s2Su :XeðuÞðsÞ>fuðsÞo
Pf9s2Su :UðsÞ>fuðsÞg:
We now calculate the two probabilities
PðuÞ ¼Pf9s2Su :UðsÞ>fuðsÞg
for s;s02Su and show that PþðuÞ ¼ ð1þOðÞÞPðuÞ for u! 1. Hence the probabilityPf9s2Su :XeðuÞðsÞ>fuðsÞgis asymptotically equal to PþðuÞ or PðuÞ,
letting!0.
Í
3.3.1. The calculation of PþðuÞ andPðuÞ
We have to verify thatfuðsÞ satisfies the assumptions (f1), . . ., (f5) of Bra¨ker’s Theorem (Bra¨ker,1993a) (or Bra¨ker,1993b), for the derivations ofPþðuÞandPðuÞ.
We consider onlyPþðuÞ, since the other term is derived in the same way. Bra¨ker’s result states that if the following conditions (f1), (f2), (f3), (f4) and (f5) hold, then thePþðuÞcan be asymptotically approximated by the expression given bellow in 12.
(f1): Being an elementary function,fuðsÞis continuous, which is the condition (f1).
(f2): Since limu!1fuðsÞ ¼ 1for anys>0, we have limu!1infs2SufuðsÞ ¼ 1which is condition (f2).
(f3): SetGðxÞ ¼K1ð1=xÞ; ðx>0Þ, and
DuðsÞ ¼G
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi We Wð1þÞ
W s
fuðsÞ 0
@
1
A for s2Su:
With ðxÞ ¼expðx2=2Þ=pffiffiffiffiffiffi2
xYðxÞ ¼1FðxÞ as x! 1, condition (f3) assumes that for any >0
Z
Su
ðfuðsÞÞ DuðsÞ ds!0
foru! 1. BecauseSuis bounded, (f2) holds andK1ð
Þis regularly varying at 0 with index 2=, as mentioned, hence (f3) follows.(f4): Withuðs; Þ ¼ ½fuðsþDuðsÞÞ fuðsÞfuðsÞ, condition (f4) states that uðs; Þ converges uniformly to a function ðs; Þ for s2Su and , some >0.
This holds since we show that foru! 1;jj ;0 < 1ands2Su
juðs; Þj!0:
For some2 ðs;sþDuðsÞÞwe have
fu0ðÞ ¼ fuðsþDuðsÞÞ fuðsÞ DuðsÞ
and thusuðs; Þ ¼DuðsÞfu0ðÞfuðsÞ:Now forDuðsÞwe get DuðsÞ ¼ W
We Wð1þÞ
!2 1 fuðsÞ
2
e L L
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi W We Wð1þÞ
s 1
fuðsÞ
!
¼O u 2ðH1Þð1þuðsÞÞ2LLeð1=fuðsÞÞ
; asu! 1withLLðe
Þa slowly varying function.To estimatefu0ðÞwe use thatfu0ðs*uÞ ¼0 and Proposition 3.3:
fu0ðÞ ¼fu0ðsu*Þ þfu00ðsu*Þðsu*Þ þoðsu*Þ
¼O u 1Hðsu*Þ
¼OðloguÞ;
sincesu*ðuÞfor2Su, by the choice ofðuÞ.
Putting together the estimations ofDuðsÞandfu0ðsÞwe have juðs; Þj ¼OðjDuðsÞfu0ðÞfuðsÞjÞ
¼O u ð1HÞð12Þð1þuðsÞÞ12LLeð1=fuðsÞÞlogu
¼oð1Þ:
since <2 and >H. Hence the sequence uðs; Þ converges to ðs; Þ ¼0 foru! 1, for any bounded.
(f5): (f4) implies that the juðs; Þj !0, hence is finite for all 2R, which is condition (f5): supsjðs; Þj < 1for any:
Thus we can apply Bra¨ker’s Theorem since the considered stochastic process UþðsÞ is a locally stationary Gaussian process with index2 ð0;2Þ, is independent of the parameter u and the sequence of boundary functions fuðsÞ satisfies the conditions (f1),. . .,(f5). Bra¨kers’ result states that
lim
u!1
1 Lu
Pf9s2Su :UþðsÞ>fuðsÞg ¼lim
u!1
PþðuÞ Lu
¼1; ð12Þ
whereLu¼R
SuuðsÞdssincegðs; Þ ¼0 and where fors2Su
uðsÞ ¼ H21 ðfuðsÞÞ G
ffiffiffiffiffiffiffiffiffiffiffiffiffi~
WWð1þÞ W
q
fuðsÞ
¼ H21 ðfuðsÞÞ
K1 1 ffiffiffiffiffiffiffiffiffiffiffiffiffi~
WWð1þÞ W
q
fuðsÞ
:
We derive the behavior of the integral Lu asu! 1. Since K1ð
Þ is regularly varying with index 2=, it follows uniformly fors2Su1=K1 1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi e W Wð1þÞ
W s
fuðsÞ 0
@
1 A 0
@
1
A
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi e WWð1þÞ
W s
A 0
@
1 A
2 .
K1ðuH1Þ:
Since fuðsÞ=fuðs0Þ !1, asu! 1, uniformly for s2Su, we derive with Proposi- tion 3.1
Z
Su
uðsÞds¼H21
Z s*uþðuÞ s*uðuÞ
ðfuðsÞÞ=K1 1 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi e
WWð1þÞ=W q
fuðsÞ
ds
H21 ffiffiffiffiffiffiffiffiffiffiffiffiffi e
WWð1þÞ W
r
A2 ffiffiffiffiffiffi
p2
Au1HK1ðuH1Þ
Z s*uþðuÞ s*uðuÞ
e12fu2ðsÞds:
We expand the exponent of the integrand fors!su* fu2ðsÞ ¼ fuðs*uÞ þfu0ðs*uÞðss*uÞ þ1
2fu00ðs*uÞðss*uÞ2þoðss*uÞ2
2
¼fu2ðs*uÞ þfuðs*uÞfu00ðs*uÞðss*uÞ2ð1þoð1ÞÞ:
Then we change the variable x¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi fuðsu*Þfu00ðsu*Þ
p ðssu*Þ. The bounds su*ðuÞ are replaced by
ðuÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi fuðsu*Þfu00ðsu*Þ q
ðloguÞ ffiffiffiffiffiffiffiffi pAB
! 1 ðu! 1Þ:
Hence, the integral is by this transformation Z s*uþðuÞ
s*uðuÞ e12fu2ðsÞds e12fu2ðs*uÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi fuðsu*Þfu00ðsu*Þ p
Z 1 1
e12x2ð1þoð1ÞÞdx ffiffiffiffiffiffi p2
e12fu2ðs*uÞ
ffiffiffiffiffiffiffiffi pAB
u1H ;
and we get forPþðuÞthe approximations
PþðuÞ Z
Su
uðsÞds
H21 ffiffiffiffiffiffiffiffiffiffiffiffiffi~ WWð1þÞ
W
q
A2 e12fu2ðs*uÞ
A ffiffiffiffiffiffiffiffi pAB
u22HK1ðuH1Þ :
In an analogous way we get the same approximation forPðuÞ, replacingþby . Taking the limit!0, finishes the proof of Proposition 3.4.
Í
3.4. Tail behavior ofXðuÞðsÞfors62Su
The probability of an exceedance outside of Su is bounded by the sum of the following four terms:
Pn9s62Su :XXeðuÞðsÞ>fuðsÞo
Pn9s2 ð0;s1:XXeðuÞðsÞ>fuðsÞo
þPn9s2 ðs1;su*ðuÞÞ:XXeðuÞðsÞ>fuðsÞo þPn9s2 ðsu*þðuÞ;s2Þ:XXeðuÞðsÞ>fuðsÞo
þPn9ss2:XXeðuÞðsÞ>fuðsÞo :
for some larges2.
Now we make use of the Ho¨lder condition (9) which holds for eachXiðuÞð
Þ. Thisimplies by applying similar derivations as in the proof of Lemma 3.2 that also lim sup
u!1E Xh ðuÞðsÞ XðuÞðs0Þi2
Gjss0j;
for anys;s0withjss0j Cwhere¼minfig>0 andG>0 a suitable constant. It means that the Ho¨lder condition (9) holds also forXðuÞð
Þ.Proposition 3.5: For s1 and the lower bound fu;1ðsÞ of the boundary function introduced in Section3.2,we get
Pn9s2 ð0;s1:XXeðuÞðsÞ>fuðsÞo
C0s1uð1HÞ2Yfu;1ðs1Þ
;
as u! 1with C0 only depending on and G.
Proof: For 0<ss1 we have minffuðsÞg>minffu;1ðsÞg ¼fu;1ðs1Þ sincefu;1ðsÞ is strictly decreasing onð0;s1. By Theorem 8.1 of Piterbarg (1996) we get
Pn9s2 ð0;s1:XXeðuÞðsÞ>fuðsÞo
P sup
s2ð0;s1
e X
XðuÞðsÞ>fuðs1Þ
( )
C0s1uð1HÞ2Yfu;1ðs1Þ
for someC0only depending on andG.