Safety, Reliability and Risk Analysis: Beyond the Horizon – Steenbergen et al. (Eds) © 2014 Taylor & Francis Group, London, ISBN 978-1-138-00123-7
A note on the ratio of the extreme to the root of the sum of squares
in sequences of absolute values of Gaussian variables
Michiel A. Odijk
Independent Research, Gouda, The Netherlands
Pieter H.A.J.M. van Gelder
Faculty of Civil Engineering and Geosciences, Delft University of Technology, The Netherlands
ABSTRACT: Let Z1, ...,Zn be i.i.d. standard normal variables, Mn the extreme among their
abso-lute values, and Qn their rooted sum of squares. Here we study the ratio υn=M Qn/ . We show that n
µ υ( )n =µ(Mn)/ ( ) ~µQn bn/ n~ 2log( )/n n, where cn~ denotes that c ddn n/ →1 for n→∞. Moreover, n
convergence happens in a stable manner, i.e. υn~ ( ) in probability. For low dimensions we also derive µ υn
explicit expressions for µ υ( )n and variance σ υ2( )n . These expressions are calculated by using an appeal-ing geometrical interpretation of υn.
the extreme value distributions), it is tempting to believe that υn behaves asymptotically like b nn/ .
In this note we show that this intuition is indeed right and that in fact µ υ( )n =µ(Mn)/ ( ) ~µQn
bn/ n~ 2log( )/n n. Moreover, convergence of
happens in a stable manner, i.e. υn~ ( ) in µ υn
probability.
A noteworthy result, that we use explicitly in this note, is that by Downey & Wright (2007). They show that for non-negative i.i.d. random variables
Xi with finite second moment
µ µ µ max( ) ~ max( ) . X X X X i i i i
∑
∑
(
)
(
)
In particular this implies that µ(M Sn/ ) ~n
µ(Mn)/ ( )µSn , which mimicks µ υ( )n =µ(Mn)/
µ( )Qn , though in a weaker manner.
Central element to the above is the observation that υn has the same probability distribution as
ηn: max{|= X1|, ..., |Xn|}|X=( , ...,X1 Xn) ~ ( )U Bn
where ~ ( )U Bn denotes sampling from the
uni-form distribution U over Bn, being the surface of
the n-dimensional unit hypersphere. This obser-vation follows from the fact that points can be sampled randomly from Bn by sampling from
( , ...,Z1 Zn) and normalize the samples with Qn
(see Muller 1959). In Section 2, this observation allows us to calculate µ υ( )n and σ υ2( )n explicitly
for low values of n, which is another aim of this note.
1 InTRODuCTIOn
Let {Zn} be i.i.d. standard normal variables and
define the entities
Mn i n Zi Sn Zi Q Z i n n i n i = = = = =
∑
=∑
max (| |), | |, ,..., ,..., ,..., 1 1 2 1 andυn n n M Q = .For a random variable X, denote with µ(X) its mean and with σ2( )X its variance. If { }p
n and
{ }qn are sequences, then qn= ( ) means that o pn
q pn n→0 (n→ ∞) . For notational convenience,
when it is clear from the context, we don’t mention explicitly that n → ∞ when we discuss asymptotic behavior. Also, we write pn~ as a shorthand qn
for p qn/ →1 or, equivalently, pn n=qn(1+o( ))1 .
note that the tilde-operator defines an equivalence relation. note that, if pn~ and cqn n~ , then dn
p cn n/ ~ / .q dn n
As Qn~ n in probability (note that for all ε > 0 Chebychev’s inequality implies
Pr(| / | ) Pr(| / )/( / ) | ) Pr(| / | ) Q n Q n Q n Q n n n n n − < = − + < ≥ − < ≥ 1 1 1 1 2 2 ε ε ε σ ((| | )/Z12 ε n→0)
and Mn is attracted by a stable law (meaning
that there are sequences { }an and { }bn such that
Also, by insensitivity of norms to scaling and translation, it follows that υn can be interpreted as
the max-distance relative to the Euclidian-distance of two arbitrary points in n-space, which is an appealing geometrical interpretation of υn. Put
dif-ferently, the ratio υn can be interpreted as the
long-est edge of a hyperrectangle (also called orthotope), normalised by the space diagonal (which is Qn).
In perspective, a vast amount of literature exists that deals with the properties of the ratio
M Sn/ . Application of this ratio can be found in n
the analysis of process speedup and the perform-ance of scheduling (Downey & Wright 2007), on-line bin-batching and wireless link quality (Zhang 2001; nguyen 2011) or in the modelling of high dimensional databases (Brabants 2005). Like with
υn=M Qn/ , the ratio M Sn n/ can be interpreted as n
the longest edge of a hyperrectangle, normalised by the sum of the lengths of its edges (which is Sn,
taking one edge per dimension). 2 ExACT RESuLT
In this section we calculate exact val-ues for µ υ( )n and σ υ2( )n for low
val-ues of n. To do this, like mentioned in the introduction, we use the observation that υn and
ηn: max{= X1, ..., Xn} |X=( , ...,X1 Xn) ~ ( )U Bn
have the same distribution as a consequence of the fact that points can be sampled randomly from unit hypersphere Bn by sampling from Z1, ...,Zn
and normalize the samples with Qn.
Let Bn*: {= x B x∈ n| 1≥x2≥ ≥... xn≥0 , i.e. B} n* is
the part of Bn of which all points are non-negative
and define a non-increasing mapping. It is not difficult to see that by virtue of symmetry prop-erties µ η( ) µ η( | *)
nk = nk Bn , where µ η( nk| .) denotes
the conditional k-th moment of ηn. note that
vol B( )n* =vol B( )/n 2nn!.
When integrating over Bn* we switch to
polar coordinates. That is, if x B∈ n*, then
ϕ=( , ...,ϕ1 ϕn−1) [ , / ]∈0π 2n−1 and x1=cos( )cos( )ϕ1 ϕ2 ⋅⋅⋅cos(ϕn−1)
x2=sin( )cos( ) cos(ϕ1 ϕ2 ⋅⋅⋅ ϕn−1)
x3=sin( )cos( ) cos(ϕ2 ϕ3 ⋅⋅⋅ ϕn−1)
….
xn−1=sin(ϕn−2)cos(ϕn−1)
xn=sin(ϕn−1)
the Jacobian of which is well known and satisfies
Jn( )ϕ =Jn( , ...,ϕ1 ϕn−1) cos (= n−2ϕn−1)⋅Jn−1( , ...,ϕ1 ϕn−2) with J2( )ϕ = . In terms of polar coordinates we have1
Bn* { ( )x Bn| tan( ) ,tan( ) sin( ),
tan( ) sin( ), ... = ∈ ≤ ≤ ≤ ϕ ϕ ϕ ϕ ϕ3 ϕ2 1 2 1 1 ,, tan(ϕn−1) sin(≤ ϕn−2)}. Clearly max{ , ..., }x1 xn = =x1 cos( ) cos(ϕ1 ⋅⋅⋅ ϕn−1) for x B∈ n*. Hence, the k-th moment of υn is
calcu-lated as µ υ µ η ϕ ϕ ϕ ( ) ( ) ! ( ) ... cos( ) cos( ) tan( ) nk nk n n n n vol B = = ⋅⋅⋅
(
≤ −∫
2 1 1 1 1))
−⋅⋅⋅ ≤ −∫
− k n n J d d n n ( ) . tan( ) sin( ) ϕ ϕ ϕ ϕ ϕ 1 1 1 2Evaluating this integral is quite cumbersome for large values of n, even for k = 1 or k = 2 (the latter being required to calculate σ υ2( )
n ). In Appendix A
we derive for lower dimensions
n µ(υn) approx. σ2(υn) approx. 1 1 1 0 0 2 2 2 π 0.900 1 2 1 8 2 + − π π 0.00774 3 3 2 1 2 2 π × arctan 0.831 1 3 2 3 18 1 2 2 2 2 + − π π × arctan 0.0100 4 8 2 32 2 1 2 2 2 π π − × arctan 0.779
For n = 2 the probability distribution has sup-port [ / , ]1 2 1 and reads
F z z z U z 2 2 1 1 0 4 1 4 ( ) Pr( ) Pr(cos( ) | ~ [ , / ]) arccos( ) = ≤ = ≤ = − υ ϕ ϕ π π
for z in the support.
3 MOMEnT COnvERGEnCE AnD STABILITy OF υn
Following the previous section, since
vol B( )n = 2πn/2/ ( / )Γn 2 , Qn is chi-squared
distrib-uted with n degrees of freedom and density Χ( , ),t n
and Mnk Z Z
k n
k
µ
ϕ
ϕ ϕ ϕ
( ) ! ...
cos( )
tan( ) tan( ) sin( )
M n r nk n n n = ⋅⋅⋅ ≤ ≤ ∞
∫
∫
∫
− − 2 1 1 1 2 0 1 ccos( ) ( ) ( ) ϕ ϑ ϕ ϕ ϕ n k n n n n r J r drd d − − −(
)
⋅⋅⋅ 1 1 1 1 = ⋅⋅⋅(
)
⋅ ≤ − −∫
2 1 1 1 1 1 n n n k n n n vol B J d ! ( ) ... cos( ) cos( ) ( ) tan( )ϕ ϕ ϕ ϕ ϕ ⋅⋅⋅ × − ≤ −∫
d n n ϕ ϕ 1 ϕ 2 1 tan( ) sin( ) ... ... ( ) ( )/ / × ∞ − − +∫
vol Bn 1n e r rn kdr 2 2 2 1 0 2 π = ∞ − − +∫
µ υ π ( ) ( ) ( )/ ( ) ( / ) / ( )/ nk vol Bn n e r r n k d r 1 2 2 2 2 2 2 2 2 0 2 = ∞ − −∫
µ υ π ( ) ( ) ( / ) ( / ) / / / / / nk nn k n t n vol B n t n e t dt Γ Γ 2 2 1 2 2 2 2 2 2 2 1 0 =µ υ( )nk ∞∫
tk/2 ( , )t n dt 0 Χ =µ υ µ( ) (nk Qnk)where Γ is the Gamma function, for which it is known that Γ( ) (k = k−1 and )! Γ(k+1 2/ )= π(2k−1 2)!!/ k.
Furthermore, it is well-known in literature that
µ(Mn) ~ ~bn 2log( )n , where bn is defined in the
introduction and more precise in Appendix B. As to µ( )Qn , note that µ( ) ( / ) (( )/ ) ( / ) / / / Q xx e n dx n n n n x n =∞
∫
22 1− − 2 = + 0 2 2 2 1 2 2 Γ Γ Γ By observing that µ(Qn−1)<µ( )Qn <µ(Qn+1) , we calculate straightforwardly thatn− =1 µ( ) (Qn µQn−1)<µ2( )Qn <µ( ) (Qn µQn+1)=n, so that n− <1 µ( )Qn < n. Hence µ( ) ~Qn n, implying µ υ µ µ ( ) ( ) ( ) ~ ~ log( ) n n n n M Q b n n n = 2 .
Also, from Appendix it follows directly (note that µ( )Qn2 = ) thatn σ υ( )n π n nb 2 2 2 6 0 − → so that σ υ( )n2 → .0
From Appendix B it also follows straightfor-wardly that σ υ µ υ µ υ µ υ µ µ µ µ 2 2 2 2 2 2 2 2 1 1 ( ) ( ) ( ) ( ) ( )/ ( ) ( )/ ( ) ( n n n n n n n n M Q M Q = − = − = bb n o b n o n n 2 211 11 1 0 / )( ( )) ( / ) ( ( )) + + − →
and so (applying Chebychev’s inequality), for all
ε > 0 Pr ( ) ( ) Pr ( ) ( ) ( ) ( ) υ µ υ µ υ ε υ µ υ ε µ υ σ υ σ υ n n n n n n n n − ≤ = − ≤ ≥ 11−σ υ22 12→1 µ υ ε ( ) ( )nn
Thus, we conclude that υn is relatively
sta-ble, i.e. υn~ ( ) in probability. note that since µ υn µ υ( )n → 0 this is a stronger concept than absolute
stability.
4 SIMuLATIOn OF υn
Following the previous elaboration, a Monte Carlo simulation is conducted to verify the asymptotic behaviour towards 2log( )/n n . The simulation
code as included in Appendix C is used for this purpose. Figure 1 shows the result of 1000 runs. It is seen that the realisations are symmetrically dis-tributed around the mean value and that the uncer-tainty vanishes large n.
REFEREnCES
Brabants, M. 2005, Introduction to Multimedia Data-bases, PhD thesis university of Hasselt, Belgium, (in Dutch).
Cramer, H.1946, Mathematical methods in statistics, Princeton university.
Deo, C.M. 1972, Some limit theorems for maxima of absolute values of Gaussian sequences, Indian Journal
of Stat. (Series A) 34(3): 289–292.
Downey, P.J. & Wright, P.E. 2007, The ratio of the extreme to the sum in a random sequence, Extremes 10(4): 249–266, Alternatively, see working paper TR 94-18 by the same authors and with the same title from the Dep. of Comp.Sc., university of Arizona, 1994.
Falk, M. & Marohn, F. 1993, von Mises conditions revisited, Ann. Prob. 21(3): 1310–1328.
Hoorfar, A. & Hassani, M. 2008, Inequalities on the Lam-bert W function and hyperpower function, Journal of
Inequalities in Pure and Appl. Math. 9(2), Article 51.
Muller, M.E. 1959, A note on a method for generating points uniformly on n-dimensional spheres, Communications of the ACM 2(4): 19–20.
nair, K.A. 1981, Asymptotic distribution and moments of normal extremes, Ann. Prob. 9(1): 150–153. nguyen v.M. 2011, Wireless Link Quality Modelling and
Mobility Management for Cellular networks, Report
Bell Labs and TREC Collaboration(s).
Pickands III, J. 1968, Moment convergence of sample extremes, Ann. of Math. Stat. 39(3): 881–889. Smith, R.L. 1987, Approximations in extreme value
theory, Technical Report 205, Dept. Statistics, univ. north Carolina.
von Mises, R. 1954, La distribution de la plus grande de n valeurs, Amer. Math. Soc., reprinted in: Selected Papers II: 271–294.
Zhang G. 2001, An on-line bin-batching problem, Discrete Appl. Math. 108(3): 329–333.
APPEnDIx A: DERIvATIOnS OF ExACT RESuLTS For n =2 we have µ υ ϕ ϕ π π π ( ) ! ( ) cos( ) , / 2 2 2 0 1 4 1 2 2 8 2 1 2 2 2 2 0 900 = = = ≈
∫
vol B dLikewise, for n = 3 we calculate
µ υ π ϕ ϕ ϕ ϕ ϕ ( ) ! ( ) cos( ) cos ( ) / arctan(sin( 3 3 3 0 4 1 2 2 2 1 0 3 2 1 =vol B
∫
⋅ d d )))∫
= ⋅ 48 4 1 2 1 1 1 0 4 π ϕ ϕ ϕ π cos( ) [cos(arctan(sin( )) sin(arctan(sin( )) /∫∫
+ +... arctan(sin( ))]ϕ1 dϕ1 = + ⋅ + + 6 1 1 1 1 2 1 1 2 1 1 π ϕ ϕ ϕ ϕ ϕ cos( ) sin ( ) sin( ) sin ( ) arctan(sin( )) ∫
dϕ π 1 0 4 / = + +∫
6 1 2 0 1 2 π y y arctan( )y dy / =6[
]
1 20 π yarctan( )y / =3 2 1 ≈ 2 2 0 831 π arctan( ) , and for n = 4 µ υ ϕ ϕ π ϕ ( ) ! ( )cos( )cos ( )cos
/ arctan(sin( )) 4 4 4 0 4 0 1 2 2 4 2 1 = ⋅
∫
∫
vol B 33 3 3 2 1 0 2 ( ) arctan(sin( )) ϕ ϕ ϕ ϕ ϕ d d d∫
=384∫
2 2 3 2 0 4 1 2 2 2 2 1 0 π ϕ ϕϕ ϕ ϕ π / ar cos( )cos ( ) sin(arctan(sin( )))d d cctan(sin( )) ... ϕ1∫
+ ... cos( )cos ( ) sin(arctan(sin( ))) / ar + 384∫
⋅⋅⋅ 2 2 13 0 4 1 2 2 2 0 π ϕ ϕ ϕ π cctan(sin( ))ϕ1∫
⋅⋅⋅ cos (arctan(sin( )))2 2 2 1 ϕ d dϕ ϕ = +∫
128 1 2 0 4 1 2 2 22 2 2 1 0 π ϕ ϕ ϕ ϕ ϕ ϕ π / arctan(scos( )cos ( ) sin( ) sin ( )d d iin( )) ... ϕ1
∫
+ ...cos( )cos ( ) sin( )
sin ( ) / / a + +
(
)
∫
64 1 2 0 4 1 2 2 2 2 2 3 2 2 1 0 π ϕ ϕ ϕ ϕ ϕ ϕ π d d rrctan(sin( ))ϕ1∫
= + +∫
∫
128 1 2 0 4 1 2 2 2 1 0 1 π ϕ ϕ ϕ ϕ π ϕ / arctan(sin( ))cos( )cos( )d sin ( )d ....
...
cos( )cos( )sin( ) sin( )
sin ( ) / ar + +
∫
64 1 2 0 4 1 2 2 22 2 1 0 π ϕ ϕ ϕ ϕ ϕ ϕ π d d cctan(sin( ))ϕ1∫
= + + − −∫
128 1 2 1 1 2 2 2 1 2 1 2 1 1 0 4 π ϕ ϕ ϕ ϕ π cos( ) sin ( ) sin ( ) ... / d...−1282[ cos( )/ 1 arctan(sin( ))
∫
2−cos ( ) cos( )2 2 2 1 0 0 1 π ϕ ϕ ϕ ϕ ϕ π d d 44∫
+ ] ... ... cos( ) sin ( ) ( sin ( )) sin ( ) + + + − − − 64 1 1 2 2 1 2 2 1 2 1 2 1 2 1 2 2 π ϕ ϕ ϕ ϕ y y dy 11 1 1 0 4 1 2 1 / sin ( ) / +∫
∫
ϕ π ϕ d =1282 +642 π A π B. where A y y dy y dyd = + + − + −∫
∫
∫
+ 1 2 1 1 2 2 2 2 2 0 1 2 1 2 1 1 1 1 0 4 2 1 / / sin ( ) / cos( )ϕ ϕ ϕ πand B y y dy = + + − − − + cos( ) sin ( ) ( sin ( )) sin ( ) / si ϕ ϕ ϕ ϕ 1 2 1 2 1 2 1 2 2 1 1 1 1 2 2 1 2 nn ( ) / 2 1 1 0 4 1 ϕ π ϕ
∫
∫
d = + + − − −∫
∫
+ y y y dy y y dy 2 2 2 0 1 2 1 2 2 1 1 1 0 41 1 2 2 1 2 2 1 ( ) cos( ) / / sin ( ) / ϕ ϕ π∫∫
dϕ1 Since 2 1 2 2 12 2 2 1 1 1 2 1 1 2 1 − = − + + +∫
y dy y y y / sin ( ) / sin arcsin ϕ 22 1 1 ( )ϕ = + − + + − + 1 2 1 4 1 2 2 2 1 2 2 2 1 2 1 2 1 π ϕ ϕ ϕ sin ( )sin ( ) arcsin sin ( )
we find cos( ) / sin ( ) / ϕ ϕ π ϕ π 1 2 1 1 1 1 0 4 2 2 0 2 1 4 2 18 2 1 2 2 2 2 1 − = + − + + +
∫
∫
y dyd y y dy 11 2 2 0 1 2 1 2 2 / / arcsin .∫
−∫
+ y dy Also, we have arcsin arcsin ( / 1 2 2 1 2 2 2 2 2 2 0 1 2 2 + = + + y dy y y y ++ +∫
∫
2 2 1 2 2 0 1 2 0 1 2 y ) y dy / / = + + +∫
1 2 2 1 3 3 2 2 2 1 2 2 2 2 0 1 2 arcsin ( ) / y y y dy so that A y y dy = − + − + + +∫
1 4 2 18 2 12 2 13 3 1 2 2 2 1 2 2 0 1 2 π arcsin ( ) / . Furthermore, 2 1 2 2 4 2 3 2 2 2 1 1 1 2 2 1 1 1 2 2 1 2 1 y y dy y y dy y d − − = − − + − +∫
+∫
/ sin ( )ϕ / sin ( )ϕ yy 1 1 1 2 1 / +sin ( )∫
ϕ = − − + − +∫
+∫
2 2 3 1 2 2 1 1 1 2 1 1 1 2 1 2 1 y dy y dy / sin ( )ϕ / sin ( )ϕ = − + − + + − + 2 1 2 1 4 1 2 2 2 1 2 2 2 1 2 1 2 1 π ϕ ϕ ϕ sin ( ) sin ( ) arcsin sin ( ) + ... = − + + + + − + 1 1 4 1 2 1 1 2 2 2 1 2 1 2 1 π ϕ ϕ ϕ sin ( ) sin ( ) arcsin sin ( ) so that B y y y y dy = + − + + + −∫
(( )) ( ) ... / 2 1 1 1 1 2 2 2 2 2 0 1 2 ... sin ( ) sin ( ) arcsin sin ( ) − − + + + + − + 1 1 4 1 2 1 1 2 2 2 1 2 1 2 1 π ϕ ϕ ϕ ∫
0 4 1 π ϕ / sin( ) d = + − + + + − − + + + + −∫
(( )) ( ) arcsin / 2 1 1 1 1 2 1 1 4 1 2 1 1 2 2 2 2 0 1 2 2 2 y y y y dy y y π 22 2 2 0 1 2 + ∫
y dy / = − + + − + + ∫
∫
1 1 2 1 1 4 1 2 2 2 0 1 2 2 0 1 2 y dy y dy / / arcsin π = − + + − +( )
+ + +∫
1 21 12 2 1 8 2 12 2 13 3 2 2 2 1 2 2 0 1 2 2 2 2 0 y dy y y y dy / arcsin ( ) π 11 2/∫
= − + + − + +
∫
1 2 2 18 2 12 2 1 3 3 2 2 2 2 1 2 2 0 1 2 π arcsin ( ) / y y dy . Hence µ υ π π ( ) arcsin( ) ( ) / 4 2 2 2 0 1 2 128 1 4 2 18 2 12 2 1 3 3 1 2 2 1 2 = − + − + + + ∫
y y dy + ... ... arcsin ( ) / + − + + − + + ∫
64 1 2 2 18 2 12 2 13 3 2 2 2 1 2 2 2 2 0 1 2 π π y y dy = − = − ≈ 8 2 32 2 1 3 3 8 2 32 2 1 2 2 0 779 2 2 π π π π arcsin arctan . As to the variance σ υ2 2( ) we calculate the sec-ond moment of µ υ2( ) of υ2 2 as µ υ ϕ ϕ π π π π ( ) ! ( ) cos ( ) , / 22 2 2 2 1 0 4 1 2 2 8 2 1 4 8 1 1 2 0 = = + = + ≈
∫
vol B d 8818 . so that σ υ π π 2 2 1 12 82 0 00774 ( )= + − ≈ , . Also µ υ ϕ ϕ ϕ ϕ π ( ) ! ( ) cos ( ) cos ( ) / arctan(sin( 32 3 3 0 4 2 1 3 2 2 1 0 3 2 = ⋅∫
vol B d d ϕϕ1))∫
=48 − 4 13 2 1 1 3 1 π ϕ ϕ ϕcos ( ) sin(arctan(sin( )))sin (arctan(sin( )))
∫
dϕ π 1 0 4 / = + − + 12 1 3 1 2 1 1 2 1 3 1 2 1 3 2 π ϕ ϕ ϕ ϕ ϕ cos ( ) sin( ) sin ( ) sin ( ) ( sin ( )) / ∫
dϕ π 1 0 4 / = − − +∫
12 2 2 1 2 1 1 0 4 π ϕ ϕ ϕ π cos ( ) cos ( ) cos( ) ... / d ... cos ( )( cos ( )) ( cos ( ))/ cos( ) / + − −∫
12 1 3 2 2 1 2 1 2 1 3 2 1 0 4 π ϕ ϕ ϕ ϕ π d = − − − + −∫
∫
12 2 4 2 4 2 2 2 1 2 1 2 2 3 2 1 2 1 4 2 3 2 1 2 1 π π π y y dy y y dy y y dy / / / / / ( ) ( )∫∫
= − − − 12 2 1 2 2 2 1 2 1 π arcsin y y y / ... ... arcsin ... / − − − + 4 2 2 2 1 2 1 π yy y ... arcsin / + − + − − 4 1 2 2 2 2 3 2 2 2 1 2 1 π y y yy y = − − + − 4 2 2 2 2 2 1 2 1 π arcsin( ) / y y y y y = +1 3 2 3 π . so that σ υ π π 2 3 13 23 182 2 12 2 0 0100 ( )= + − arctan ≈ . .APPEnDIx B: PROOF THAT,
µ(Mn) ~ , (bn µMn2) ~bn2 AnD bn~ 2log( ))n
Let { }Zn and Z be i.i.d. standard normal
vari-ables (with distribution Φ and density ϑ ). Let Φabs( ) Pr(x = Z x≤ and denote its density with )
ϑabs. Also, we write Φ and Φabs as a shorthand
for tail probabilities 1− Φ and 1− Φabs respectively.
Obviously, for x ≥ 0 we have ϑabs( )x = 2ϑ( )x and Φabs( ) = 2Φ( )x x = erfc x( / 2 , where erfc is the )
complementary error function defined by and having first order Laurent expansion
erfc x e dt e x o t x x ( ) := ∞ − = − ( + ( ))
∫
2 2 2 1 1 π π .For the moments of Z we have µ π π π ( ) ( ) / / / Z t e dt t de k t e d k k t k t k t = = − = − − ∞ − − ∞ − −
∫
∫
2 2 1 2 2 2 2 2 0 1 2 0 2 2 tt k Zk 0 2 1 ∞ −∫
=( − )µ( )
so that µ(Z2i) (= 2i−1)!! and µ(Z2 1i+ )= 2/ ( )!!π i . notably, moments exist and are finite. 2 Also, the inverse Φabs← of Φabs will be used belowand follows from
1 1 2 2 1 1 1 2 2 2 2 2 Φabs x x erfc x x e o ( )= ( / )= ( + ( )) π yielding Φabs x W x o ← = + ( ) 2 1 12( ( )) π
where W is the Lambert W function, implicitly defined by W x e( )W x( )= .x
With Λ we refer to the Gumbel distribution (also known as the doubly exponential distribu-tion and being one of the three limiting cases of the GEv distribution, well known in extreme value theory) defined by Λ( ) exp(x = −e−x), having mean γ and variance π2/ , where γ ≈ 0 5776 . is the
Euler-Mascheroni constant. It is common knowl-edge that Φ satisfies the von Mises condition for being in the domain of attraction of Λ (von Mises 1936; Falk 1993), i.e. ϑ( ) ( ) ( ) ( ) x x x t dt x Φ2 Φ 1 ∞
∫
→ → ∞ .It is easily seen that Φabs also satisfies this condi-tion and is therefore also in the domain of attrac-tion of Λ. Say that { }an and { }bn are sequences for
which (Mn−b an)/ n→ Λ in distribution. Intuitively
we would expect that µ(Mn) (− bn+anγ)→ 0 and that µ(Mn2) ( (− a2n π2/6+γ2)+bn2+2a bn nγ)→ . 0 From Pickands (1968) (see also nair 1981) we know that this is indeed the case. Furthermore, it is widely known that { }an and { }bn are solutions to bn= abs n
←
Φ ( / )1 and an= ( ), where r xr bn ( )= Φabs( )/x ϑabs( )x is the reciprocal hazard function (Smith 1987). Hence
bn= W n +o 2 2 1 1 π ( ( )) and a n e b n b e b n b n n b n n n = 1 − = 2 1 2 1 2 2 2 2 2 / /π / ~ π .
Hence (note that µ(Mn)= + 1 and bbn o( ) n→∞),
µ(Mn) ~bn and µ(Mn2) ~bn2.
Finally, as log( ) log log( )x − x W x≤ ( )≤ log( ) log log( )/x − x 2 for x e≥ (see Hoorfar 2008), we see that in terms of elementary functions bn~ 2log( )n . Consistently, Cramer
(1946) and Deo (1972) approximate the loca-tion and scaling parameters { }an and { }bn by
bn= −cn log( log( ))/4π n 2cn+2/cn and an= 1/ , cn
where cn= 2log( ) .n
APPEnDIx C: MATLAB CODE
FOR SIMuLATIOn OF υn Clear clg n=1000; hold on for j=1:n, z(1)=normrnd(0,1); m(1)=abs(z(1)); s(1)=m(1); q2(1)=z(1)^2; for i=2:n, z(i)=normrnd(0,1); m(i)=max(m(i-1),abs(z(i))); s(i)=s(i-1)+abs(z(i)); q2(i)=q2(i-1)+z(i)^2; q(i)=sqrt(q2(i)); nu(i)=m(i)/q(i); ex(i)=sqrt(2*log(i)/log(2.7181)/i); end loglog([1:n],nu,′b′) end loglog([2:n],ex(2:n),′r′) xlabel(′Sample size n′) ylabel(′nu′) grid
title(′Results of 1000 simulations (blue lines) and theoretical result (red line)′)