Beruflich Dokumente
Kultur Dokumente
1
Ladder Heights and the Zeta Function 2
1 Introduction
Let X ; X ; : : : be independent and distributed as N (; 1), the normal distribution
1 2
with mean and variance 1. Consider the random walk fSn : n 0g consisting of the
partial sums Sn = X + + Xn. The rst time = inf fn : Sn > 0g that the random
1
walk is positive is called the rst ladder epoch, and the rst positive value S taken by the
random walk is called the rst ladder height.
The Riemann zeta function is originally dened for Re(z) > 1 by the series
(z ) = 1 ?z
n=1 n . This denition is extended by analytic continuation to the entire
P
complex plane except z = 1, where has a simple pole. For example, the extension to
Re(z) < 0 may be seen from Riemann's functional equation (see (19)), which expresses
(z ) in terms of (1 ? z ). Calculation of the zeta function is routine; it is implemented in
the computer package Mathematica, for example.
E S = p exp p 2
(1)
2 2 n n!(2n + 1) 2
=0
p
in the disk fjj < 2 g.
Ladder Heights and the Zeta Function 3
Previously only the n = 0 term in the expansion of Theorem 1.1 was known explicitly.
We will also determine the extent to which the function 7! E S can be analytically
continued throughout the complex plane. It turns out that the function has branch points
p p
at f 2k i 2k : k 6= 0; k integerg; in particular, it can be analytically continued to a
neighborhood containing the real axis.
Our starting point is a classical expression for E S from the
uctuation theory of
random walks; see (5) below. As # 0, the series in (5) converges ever more slowly,
whereas the convergence in the Taylor expansion (1) that we shall derive from it becomes
more rapid. The behavior of E S as # 0 is of particular signicance for applications
such as heavy trac limit theorems for queues [see, for example, Whitt (1974) and
Asmussen (1987)] and the corrected diusion approximations of Siegmund (1979).
Theorem 1.1 is based on a result (see Theorem 2.1 in Section 2) about power series of
the form P1 ? ?nw , which also arise in other contexts. For each this series
n n e =1
represents an analytic function of w in the right half-plane fRe(w) > 0g. We investigate
the analytic continuation of this function; we are particularly interested in continuing to a
neighborhood of w = 0. Most of the work was in fact done by Hardy (1905), who
investigated certain questions about power series, including the behavior of the series
P 1 n? z n about the point z = 1. So our problem involves a reparametrization of
n=1
Hardy's. This reparametrization turns out to be felicitous, in that the resulting expansion
has a neat and explicit expression in terms of the zeta function: whereas Hardy stopped
with the rst coecient in his expansion, we obtain explicit expressions for all of the
coecients in ours. To keep the present paper self contained, some of our development
Ladder Heights and the Zeta Function 4
For simplicity, our treatment here is conned to random walks with Gaussian
increments, which is the case that has been the most important in applications. The
Gaussian case also plays a central role in the theory; a heuristic sense of this may be
obtained by combining the central limit theorem with the basic identity (5) below, which
holds for general increment distributions. An intricate analysis in this spirit was carried
through by Lai (1976), who analyzed more general increment distributions|although
with the additional restriction of zero mean|by expanding around the Gaussian case.
Likewise, the importance of the series P1 ? n
n n z goes beyond the Gaussian case; indeed
=1
Moments of the rst ladder height have been studied in the general setting of an
exponential family of distributions fF : 2 g, which may be written in the form
F (dx) = ex?
( )
F0 (dx). For example, the normal family fN (; 1) : 2 IRg under
consideration here is a special case having () = =2. Let P and E denote probability
2
as # 0. The question of whether the remainder o() in (2) is actually O( ) was left
2
open by Siegmund's development. Under the slightly strengthened hypothesis that the
distribution F is strongly nonlattice, Chang (1992) replaced the o() by ca + O( ).
0
2 3
Ladder Heights and the Zeta Function 5
Here ca is a constant dened in terms of F whose general form is rather complicated, but
0
it takes a simple form when a = 1 and the distribution F is symmetric about 0. For the
0
p1 exp + O( ) ;
h i
E S = 3
(3)
2
where = E (S )=(2E S ) =: 0:583.
0
2
0
The quantity has an interesting history. It arose in a paper of Cherno (1965), who
studied discrete and continuous versions of the problem of sequentially testing whether
the drift of a Brownian motion is positive. In the discrete version corresponding to a
given > 0, the Brownian motion is observed only at times 0; ; 2; : : :. Both the discrete
and continuous versions have optimal solutions of the form: stop the Brownian motion at
the rst observation that crosses a certain boundary. Denoting the optimal boundaries for
the discrete problem and the continuous problem by x = x (t) and x = x(t), respectively,
Cherno (1965) showed, in his notation, that
p p
x (t) = x(t) + z^ + o( ) as # 0: (4)
It turns out that the number z^ is in fact ?. A more fundamental probabilistic
interpretation of is as a limiting expected overshoot: dening the overshoot
Rb = S (b) ? b, where (b) is the rst passage time inf fn : Sn > bg, standard results from
renewal theory say that Rb converges in distribution to a random variable R1, and
E0 R1 = E0 (S2 )=(2E0 S ). Siegmund's result (2) with a = 1 shows that the limiting
expected overshoot is also the coecient of in the expansion of E S =E S as # 0.
0
done, Lai and Siegmund were not aware of the connection to the work of Cherno; Hogan
(1986) explained the connection between the solution of Cherno's problem and the
limiting expected overshoot.
As far as we are aware, the only previous observation that relates Gaussian random
walks and the zeta function is the tantalizing nal sentence of the following quotation
from Cherno (1965) about his result (4):
[References omitted. Also, the quoted numerical value is slightly incorrect; to 4 decimal
places it should be ?:5826.] That = ? (1=2)=(2) = can be seen by comparing (3) and
1 2
the expansion in Theorem 1.1. Lai (1976) partially rediscovered the expansion of Hardy
by a dierent method. However, Lai's development recovered only the same terms that
Hardy gave explicitly, and did not establish analyticity or relate the results to the zeta
function.
The rst ladder height plays a fundamental role in the theory and applications of
random walks and renewal theory; see, for example, the treatises of Asmussen (1987),
Feller (1971), Prabhu (1980), and Siegmund (1985), all of which feature ladder variables
prominently. The particular area of application that originally motivated this
Ladder Heights and the Zeta Function 7
for the overshoot as ! 0, the behavior of moments of the rst ladder height as ! 0
must be understood.
The analytic continuation and Taylor expansion of log E S are developed in the
next section. In the nal section we discuss some simple applications to the study of
boundary crossing probabilities.
Ladder Heights and the Zeta Function 8
P fSn 0g :
X
E S = exp (5)
n=1 n
log E S = ? p pn
X
(6)
d 2 n=1
for > 0. This motivates our study of series of the form P n?e?nw ; we are particularly
interested in the case = 1=2 and w = =2, and we would like to determine the behavior
2
which is analytic in the half-plane fRe(w) > 0g, may be analytically continued to the disk
fjwj < 2g. In that disk, for each 2 C other than the positive integers, H has the
Taylor series expansion
1 (?w)n
( ? n)
X
H (w) = c() : (8)
n=0 n!
Remark: With the appropriate interpretation, (8) holds also when is a positive
integer. In this case, the product c() ( ? n) should be interpreted as 0, except for the
case n = ? 1, where one uses that fact that c() ( ? n) has a removable singularity at
= n + 1.
Theorem 1.1 is an easy consequence: taking = 1=2 and w = =2 and using the fact 2
p p
that ?(1=2) = (so that c(1=2) = 2i ) gives
1 (e?2 =2 )n p ?=
1 2 1 (1=2 ? n) ? 2 n
!
pn ? 2
X X
2
= :
n=1 n=0 n! 2
Combining this with (6) we get
log E S = p?1
d X1 (1=2 ? n) ? 2 n
!
:
d 2 n=0 n! 2
Theorem 1.1 is obtained by integrating, exponentiating, and using the fact that
p
E0 S = 1= 2.
Figure 1.
Figure 2.
where ? < arg(u) < , say. We then adopt the denition uz = ez log u; for each z 2 C this
is also an analytic function of u for u 2 C n IR?. Having xed a branch of the logarithm,
certain operations require a bit of care; for example, a fact that we will use below is
traversed as shown in the dashed curve in Figure 1. Here the addition of contours
Ladder Heights and the Zeta Function 11
indicates successive traversal in the positive direction (i.e. whatever direction the contour
had as originally dened). Taking a > 0 and b 2 (0; 2), let
+ fa + yi : b ? 2 y bg + fx + bi : a x ? ? g + f?? + iy : b y g;
1 1
and let L0 denote the dierence L ? L , as shown in Figure 2. Finally, let D denote the
domain
D = fu : Re(u) < a and b ? 2 < Im(u) < bg
bounded by L, and let D be a bounded domain containing the point 0 such that the
0
closure of D is contained in D.
0
Then for each a 2 C, H (w) is an analytic function of w for Re(w) > 0, and I (w) is
analytic for w 2 D. For all w 2 D \ fRe(w) > 0g and Re() > 0 we have H (w) = I (w).
is easily seen to be an analytic function of w for w 2 D ; for example, one could use a
0
u?1 ju?1j
Z
u?1 Z
u?1
lim du = du (12)
!0 L e?u ? e?w L e?u ? e?w
gives
n? =
1 Z juj? enu du
0
1
?() ?1
for n 1. From this, for Re(w) > 0 we obtain
1 e?w Z 0 juj?1
n? e?nw
X
= du; (13)
n=1 ?() ?1 e?u ? e?w
by Fubini's theorem.
This can be seen by integrating over a small circle about w, or by changing variables and
using Cauchy's integral formula. Therefore, for each w 2 D n IR? , (14) holds for
suciently small positive .
lim
Z
u?1
du = ?2i sin( )
Z 0
juj?1 du for w 2 D n IR : (15)
!0 C e?u ? e?w
?
?1 e?u ? e?w
Indeed, since (u i)? ! juj? exp[i( ? 1)] for u < 0 as # 0, routine dominated
1 1
lim
Z 0
(u i)? du = ?ei Z
1 0 juj? du:1
= ?2i sin()
Z 0
juj?1
du +
Z
u?1
du
?1 e?u ? e?w L e?u ? e?w
for w 2 D n IR? and for Re() > 0. Combining this with (13) gives H(w) = I(w) for
w 2 D \ fRe(w) > 0g provided that Re() > 0.
A key idea, also used by Hardy, is to consider the analyticity of various expressions
as functions of the variable as well as w. In the proof of the theorem, we will establish
Ladder Heights and the Zeta Function 14
certain identities rst for values of with large real part, then argue by analytic
continuation that the identities hold for other values of . The next lemma prepares the
way.
Lemma 2.3 Let H and I be as dened in (7) and (10). Then for each
w 2 D \ fRe(w) > 0g, the functions 7! H (w) and 7! I (w) are both entire. Also, for
each nonnegative integer k, the kth derivative
dk
I(k) (0) = k I (w)
dw w=0
is an entire function of .
Proof. For Re(w) > 0, the sum 1 ? e?nw converges uniformly over in
P
1 n
bounded sets, so the sum is an entire function of , and hence so is 7! H (w). For
w 2 D, the integral (11) is an analytic function of for small enough positive , and the
convergence in (12) is uniform over in bounded sets. Therefore, 7! I(w) is entire in
.
To prove the nal assertion of the lemma we will use the expansion
e?w 1 +X 1 e?u
( )
L
ju? j
1
(e?u ? 1)n
+1
jduj =
L je?u ? 1j n
e?u ? 1
=0 n =0
1
X
rn
Z
ju? j jduj < 1:
1
n =0 L je?u ? 1j
Ladder Heights and the Zeta Function 15
!
dk ?w n n
? (?j )k (?1)n?j ;
X
(e 1) n =
dwk
w=0 j =0 j
this is 0 if n > k. Since (16) expresses I (w) as a uniformly convergent series of analytic
functions in the neighborhood N , we may dierentiate term by term to obtain
u?1 e?u
( ) !
dk k Z n n
(?j )k (?1)n?j :
X X
I (w )
= du (17)
dwk w=0 n=1 (e?u ? 1)n
L +1
j =0 j
However, for each n > 0, the integral
Z
u?1 e?u
du (18)
L (e?u ? 1)n +1
To complete the proof of Theorem 2.1, by Lemma 2.2, for each 2 C we have
X1 wk
I (w) = I(k) (0)
k=0 k!
for all w in a neighborhood of 0. The idea will be to evaluate the derivatives Ik (0) by ( )
Let w 2 D \ fRe(w) > 0g. Then Lemma 2.2 gives H(w) = I(w) for all with
Re() > 0. Therefore, by Lemma 2.3 and analytic continuation, H (w) = I (w) holds for
all a 2 C.
Ladder Heights and the Zeta Function 16
Next let be arbitrary. Since we have just shown that H(w) = I(w) for all
w 2 D \ fRe(w) > 0g, and since both H () and I () are analytic on D \ fRe(w) > 0g,
for each nonnegative integer k we have the equality of the kth derivatives
as w decreases to 0 along the real axis. Performing the last dierentiation gives
(
1 )
which, for Re() > k + 1, is obviously (?1)k c() ( ? k). Finally, by Lemma 2.3 and
analytic continuation, the equality Ik (0) = (?1)k c() ( ? k) holds for throughout
( )
By the functional equation of Riemann (see, e.g., section 4.44 of Titchmarsh, 1939),
which expresses (1 ? z) in terms of (z) for z 6= 1,
( ? k)
= 2?k ?k?1 cos[(? + k + 1)=2] ?(k + 1 ? ) (k + 1 ? ): (19)
k! k!
Since ?(k + 1 ? )=(k!) k? and (k + 1 ? ) ! 1 as k ! 1, it is easy to see that the
radius of convergence of the power series P ( ? k)(?w)k =(k!) is 2. This proves
Theorem 2.1.
Ladder Heights and the Zeta Function 17
n=1 j =?J
which is analytic in the half-plane fRe(w) > 0g, may be analytically continued to the disk
fjwj < 2(J + 1)g, where it has the Taylor expansion
1 ?1
!
w k
sin[( ? k)=2]J;(k)
X
2i(2) ; (20)
k=0 k 2
where J; (k) = (k + 1 ? ) ? J j ?(k+1?) .
P
j =1
so that
?(k + 1 ? ) = (?1)k ? 1 ?(1 ? ) = (?1)k ? 1 2i :
! !
k! k k c()
Further substituting [ei k ? = + e?i k ? = ] for cos[(? + k + 1)=2] in (19), we
1
2
( +1 ) 2 ( +1 ) 2
obtain
1 ( ? k)
(?w)k = (2) fi T? ? (?i) T g
X
c() (21)
k=0 k! +
where
1 ?1 iw k (k + 1 ? ):
!
X
T =
k=0 k 2
Ladder Heights and the Zeta Function 18
j =1
In the last equality we have used (9) together with the assumption that jwj < 2. Thus,
for jwj < 2,
1 ( ? k) J J
(?w)k ? 2i (w + 2ji)? ? 2i (w ? 2ji)?
X X X
c() 1 1
k=0 k! j =1 j =1
i=2 X ? 1 ?iw k
(
1 !
1 ? 1! iw k )
(k) ? e ? X
= (2) e J; i=2
J; (k )
k=0 k 2 k=0 k 2
1 ?1
!
w k
J; (k) sin[( ? k)=2];
X
= (2)2i
k=0 2 k
k=0 j =?J
1 ?1
!
X w
k
holds for fjwj < 2; Re(w) > 0g. For xed J , simple estimates show that
so that the last power series has radius of convergence 2(J + 1).
Ladder Heights and the Zeta Function 19
the complex plane except for branch point singularities at the points 2ki for integers k
(and so we must remove a ray, say Rk = fx + 2ki : x 0g, from the plane for each
branch point 2ki). Since G is periodic (with period 2i) in the right half-plane
fRe(w) > 0g, its analytic continuation is clearly periodic in the domain C n 1k ?1 Rk . S
=
The branch point singularities of G are of the form B;k (w) = 2i(w ? 2ki)? ; that is, 1
2
?B = ; 2 :
d 2 2 1 20
However, we know that G = (w) ? B = ; (w) may be analytically continued to all w in the
1 2 1 20
domain C n Sk6 Rk . Thus, E S may be continued for throughout the complex plane,
=0
p
except for branch point singularities at the points = 2 kei= for nonzero integers k
4
Finally, we remark that in any xed disk fjwj < cg, as J increases, the series in (20)
converges more and more rapidly, so that Theorem 2.4 provides a computationally
ecient method of calculating G(w).
3 Applications
Perhaps the most fundamental of boundary crossing problems for random walks is
determining the probability that a random walk with negative drift ever attains a positive
Ladder Heights and the Zeta Function 20
level b. Letting (b) denote the rst passage time inf fn : Sn > bg, we want to
approximate the probability P? f (b) < 1g, where > 0.
Corollary 3.1 There exists r > 0 such that, for 0 < 2p,
(
1 1 ( ? n) ? 1 2 n
! )
2 n n!(2n + 1) 2
=0
as b ! 1.
Proof. The overshoot (or residual ) Rb over the level b is dened by Rb = S (b) ? b
when (b) is nite. By Wald's likelihood ratio identity (e.g. Siegmund (1985) page 13),
Under P , we have (b) < 1 with probability 1, and, as b tends to innity, Rb converges
in distribution to a random variable R1 having probability density P fS > xg=(E S )
with respect to Lebesgue measure for x > 0. The expectation E (e? Rb ) converges to
2
() := E (e?2R1 ), where we have adopted the notation used by Siegmund. Corollary
2.3 of Chang (1992) shows that this convergence is exponentially fast: for > 0, there
exist C < 1 and r > 0 such that
for all 2 [0; ] and all b 0. The fact that > 0 may be chosen arbitrarily follows
from the fact that the family of distributions fN (; 1) : 0 g is uniformly strongly
nonlattice, that is,
0
inf inf j1 ? E (eitX )j > 0 for all > 0:
jtj
Ladder Heights and the Zeta Function 21
We have
1 1 ?2x
Z
1 1 [1 ? E (e?2S )]:
() = e P fS > xg dx =
E S 0 E S 2
However, letting ? denote the rst descending ladder epoch inf fn : Sn 0g, Wald's
likelihood ratio identity, symmetry of the normal distribution, and duality (e.g. Corollary
8.39 of Siegmund, 1985) give
so that
1 1
() = = :
(E S )(2E ) 2(E S )2
From this, our Taylor series for log E S may be transformed into a Taylor series for
log () about = 0, resulting in
8 s 9
2X 1 ( ? n) ?
< 1 2
! n=
() = exp : 2
: (24)
n n!(2n + 1) 2
=0
;
then
(
1 X 1 ( ? n) ?
1 2 n )
!
The meaning of the last display is as an asymptotic expansion in the sense of Poincare:
that is, for each N ,
(
1 N ( ? n)
? 1 2 n
! )
Other results that follow easily from Theorem 2.1 are explicit expressions for the
moments E (Sp). These may be found by calculating along the lines shown by Lai (1976),
0
using our expansion of the series P n?e?nw about w = 0; here the series with values of
other than = 1=2 arise. For example, the results for the third and fourth moments are
!
E (S ) = p
3 3 1 + (1=2) = 1:25035
2
0
2 4 2
and
(3=2) 3(1=2) 3 (1=2)
E0 (S4 ) =
3=2
? 3=2 ? 2p = 2:26433:
For another example of information that can be extracted from Theorem 1.1, dene
p
the function () = E R1; note that the value of (0) is the number ? (1=2)= 2 we
have been calling . The rst derivative 0(0) = 1=4 is known. The second derivative
00 (0) can be obtained by equating the expansion (24) for the function () with its
denition E (e? R1 ), writing the Taylor expansion of the exponential up to order , and
2 2
using (2). The result of this calculation, which also makes use of the moments E (S ) and 0
3
Concluding remarks.
1. The associate editor raised a question that we had also puzzled over ourselves: For
negative , is there a nice probabilistic interpretation of the analytic continuation of
E S ? In particular, one might entertain E (S ; < 1) and E (S j < 1) as
plausible candidates. In fact neither is correct. By inspection of (1), the analytic
continuation may be expressed as 1=(2E? S ) for 0. Dierentiating the
Wiener-Hopf factorization
[1 ? E (eiS ; < 1)][1 ? E (eiS? ; ? < 1)] = 1 ? E (eiX )
1
for 0. This is easily seen to lie strictly between E (S ; < 1) and
E (S j < 1) as " 0.
2. After this work was done we learned of another occurrence of (1=2) in a related
problem; see Asmussen, Glynn, and Pitman (1995).
REFERENCES
Asmussen, S., Glynn, P., and Pitman, J. (1995). Discretization error in simulation of
one-dimensional re
ecting Brownian motion. Annals of Applied Probability 5
875{896.
Chang, J. T. (1992). On moments of the rst ladder height of random walks with small
drift. Annals of Applied Probability 2 714{738.
Cherno, H. (1965). Sequential test for the mean of a normal distribution IV (discrete
case). Ann. Math. Statist. 36 55{68.
Feller, W. (1971). An Introduction to Probability Theory and its Applications, Vol. II,
second edition. Wiley, New York.
Hardy, G. H. (1905). A method for determining the behavior of certain classes of power
series near a singular point on the circle of convergence. Proc. London Math. Soc.
2 381{389.
Prabhu, N. U. (1980). Stochastic Storage Processes: Queues, Insurance Risk, and Dams.
Springer, New York.
Siegmund, D. (1985). Sequential Analysis: Tests and Condence Intervals. Springer, New
York.
Whitt, W. (1974). Heavy trac limit theorems for queues: a survey. In Mathematical
Methods in Queueing Theory , A. B. Clarke, Ed., Lecture Notes in Economics and
Mathematical Systems 98 307{350. Springer, New York.
Dept. of Statistics, Yale University, Box 208290 Yale Station, New Haven, CT 06520
chang@stat.yale.edu
Dept. of Statistics, 367 Evans Hall, University of California, Berkeley, CA 94720 and
Mathematics Institute, The Hebrew University, Givat Ram, Jerusalem 91904, Israel
peres@stat.berkeley.edu