You are on page 1of 13

Fractals, Vol. 10, No.

1 (2002) 117–129
c World Scientific Publishing Company

LARGE DEVIATIONS OF
MULTIFRACTAL MEASURES

DANIELE VENEZIANO∗
Department of Civil and Environmental Engineering
Massachusetts Institute of Technology
Cambridge, Massachusetts 02139, USA
venezian@mit.edu

Received December 7, 2000; Accepted January 30, 2001

Abstract
We analyze the extremes of stationary multifractal measures using large deviation theory. We
consider various cases involving discrete multiplicative cascades: scalar or vector cascades with
dependent or independent generators, bare or dressed measures, and marginal (single-point) or
joint (multi-point) extremes. In each case, we obtain the scaling behavior of the probability
of large deviations as the resolution of the cascade diverges. Existing rough exponential limits
for scalar cascades are confirmed, whereas for other cases our scaling relationships differ from
previously published results. For scalar cascades, we refine the rough limits by obtaining the
asymptotic pre-factor to the exponential term. Based on these refined asymptotics, we propose
a variant to the Probability Distribution/Multiple Scaling (PDMS) technique to estimate the
co-dimension function c(γ).

1. INTRODUCTION quality one is concerned with the probability that


the average rainfall intensity or the average concen-
During the past three decades, multifractal pro-
cesses have been used to represent a wide tration of a pollutant in a space-time region of given
range of physical phenomena; see Schertzer and size exceeds some critical high level.
Lovejoy (1996)1 for a review. In many applications, Since multifractal measures are statistically
interest is in the probability of rare events. For invariant under certain renormalization opera-
example, in the case of precipitation and air tions,2 – 4 one may expect that the extremes in

Correspondence to: Department of Civil and Environmental Engineering, MIT, Room 1-348, Cambridge, MA 02139.
Tel: (617) 253-7199, Fax: (617) 253-6044.

117
118 D. Veneziano

regions of different size should display some scaling behavior for large n of exceedance probabilities of
behavior. Indeed, scaling is manifested in the fact the type P [εn ≥ s], where εn is either εn, b or εn, d
that high-level crossing sets are fractal (in a sense and s varies with n in such a way that P [εn ≥ s] →
that will be clarified later) with fractal dimension 0 as n → ∞. One important case that we study in
that depends on the level.5 – 7 detail is when s = mnγ where mn is the resolution
In the past, multifractal extremes have been and γ is a constant larger than E[log m Y ]. For this
examined using rather ad hoc methods. Here, we case, it is well known that5,6,12
use the close analogy between multifractal extremes
and large deviations for sums of random variables to P [εn, b ≥ mnγ ] ∼ m−ncb(γ) (1)
verify, correct and expand existing results on mul-
tifractal measures generated by discrete cascades. where ∼ denotes equality for large n up to a
First, we introduce some notation. Consider a factor g(n, γ) that varies slowly with the resolu-
multiplicative cascade that starts at level 0 as a uni- tion mn . If the factor g is mildly dependent on n,
form unit measure in the unit d-dimensional cube then Eq. (1) allows one to interpret the quantity
Sd . The cascade has integer multiplicity m. This d − cb (γ) as the fractal dimension of the exceedance
means that, after n cascade levels, Sd is partitioned sets Ωγ, n = {Tni : εn, b (Tni ) ≥ mnγ }, where Tni is
into mnd cubic tiles Tni , i = 1, . . . , mnd , of side the generic cascade tile at level n. For this reason,
length m−n . The measure density inside Tni is ob- cb (γ) is known as the co-dimension function of the
measure.5,6 The fact that cb (γ) varies with the ex-
tained by multiplying the measure density in the
ponent γ is the origin of the term “multi-fractality.”
parent tile at level n−1 by a mean-one non-negative
Notice that the exceedance sets Ωγ, n for lower cas-
random variable Yni . The set of variables {Yni } is
cade levels n are not obtained by coarsening the
called the generator of the cascade. In the classic
exceedance sets at higher levels. Therefore, cb (γ) is
cascade scheme, the Yni are independent copies of
not the co-dimension of any single set.
a random variable Y , but one may consider gener-
If Y has finite moments of all orders, then cb (γ)
ators with dependencies both within and between
in Eq. (1) is the Legendre transform of Kb (q) =
levels. Except for Sec. 4.3, we shall assume that
logm E[Y q ],
the cascade generator is independent.
When the cascade process is terminated at a cb (γ) = max{qγ − Kb (q)} . (2)
finite level n, we say that the cascade has been q
developed to level n or equivalently to resolution
r = mn . The multifractal properties of fully The function Kb (q) is sometimes called the (bare)
moment scaling function of the cascade because the
developed (n → ∞) cascades have been extensively
bare moments E[εqn, b ] vary with n as E[εqn, b ] =
studied; see for example, Gupta and Waymire
(1993), Mandelbrot (1974), and Kahane and mnKb(q) .
Peyriere (1976) for the case of independent It is physically impossible to observe bare mea-
generator,7 – 9 and Waymire and Williams (1995 and sures at different levels n. Hence the relations in
1996) for cases with dependent generators.10,11 Eqs. (1) and (2) have limited practical interest.
Our focus is on εn , the average measure den- What is more relevant for application are the cor-
sity inside a cascade tile at level n. We distinguish responding relationships for the dressed measure
between two such average densities: the bare den- densities εn, d . The main novelty when passing
sity εn, b , which refers to the cascade developed to from bare to dressed densities is that, also when
level n, and the dressed density εn, d , which refers Y has finite moments of all orders, the moments of
to the fully developed cascade (for more on the the dressed density εn, d of order higher than some
bare/dressed distinction, see Ref. 6). In the case finite value q ∗ may diverge. The critical order q ∗ is
d Q obtained from the conditions q ∗ > 1 and Kb (q ∗ ) =
of independent generators, εn, b = ni=1 Yi where d(q ∗ − 1), where d is the dimension of the space
the variables Yi are independent copies of Y , and where the cascade is defined.9 When q ∗ exists, the
d
εn, d = εn, b · Z where Z is a “dressing factor” inde- dressing factor Z and the measure density εn, d have

pendent of εn, b . The factor Z has the same distri- algebraic upper tails of the type P [εn, d ≥ s] ∼ s−q ,
bution as the dressed measure in Sd . as s → 8.
We are specifically interested in the large- The moment scaling function Kb (q) may be
deviation properties of εn, b and εn, d , i.e. in the obtained from the bare densities as Kb (q) =
Large Deviations of Multifractal Measures 119

1
n logm E[εqn, b ] (independent of n). The cor- determine the asymptotic scaling of P {εn, b ∈
responding asymptotic moment scaling function mnB }, where B is a given region of Rk . The third
for the dressed densities is Kd (q) = limn→∞ extension considers vector-valued cascades with de-
1 q 1 q
n logm E[εn, d ] = Kb (q) + limn→∞ n logm E[Z ]. pendent generators. Section 5 summarizes our main
Since E[Z q ] diverges for q ≥ q ∗ , findings and mentions other implications of large
( deviation theory on multifractal extremes.
Kb (q), q < q∗
Kd (q) = . (3)
∞, q ≥ q∗
2. ROUGH AND REFINED
Schertzer and Lovejoy (1991),13
p. 56, observe that LIMITS FOR LARGE
the dressed densities εn, d satisfy a large-deviation DEVIATIONS
relation analogous to Eq. (1),
A classic result in large deviations is as follows.17 – 21
Let Pn be the distribution of the average Sn =
P [εn, d ≥ mnγ ] ∼ m−ncd (γ) (4)
n (X1 + X2 + · · · + Xn ) of n independent variables
1

with common distribution F . Suppose that the


where the dressed co-dimension function cd (γ) is the
moment generating function M (q) = E[eqX ] is
Legendre transform of Kd (q) in Eq. (3),
finite for all q and that X is unbounded. Then the
cd (γ) = sup(γq − Kd (q)) sequence {Pn } satisfies the so-called large deviation
q principle with rate function
(
cb (γ), γ ≤ γ∗ I(γ) = sup{qγ − ln M (q)} . (6)
= (5)
cb (γ ∗ ) + q ∗ (γ − γ ∗ ), γ > γ∗ q

This means that, for any measurable set A of the


and γ ∗ is the slope of Kb (q) at q ∗ . Equation (5) is real line,
given by Schertzer and Lovejoy without a detailed
derivation. Gupta and Waymire (1993),7 p. 260, ln Pn (A)
lim = − inf I(t) . (7)
used Chernoff’s Theorem of large deviations to n→∞ n t∈A
obtain cd (γ) = cb (γ) for all γ. However, their argu-
Consider in particular the set A = [γ, ∞) where
ment is flawed for γ > γ ∗ . In Sec. 3.2, we shall
γ > E[X]. From monotonicity of the function I(·),
derive refined limits for P [εn, d ≥ mnγ ], which
one obtains inf t≥γ I(t) = I(γ) and Eq. (7) gives
confirm Eqs. (4) and (5) and give the asymptotic
pre-factor to m−ncd (γ) in Eq. (4). The refined ln P [Sn ≥ γ]
asymptotics are used to suggest a variant of the lim = −I(γ) . (8)
n→∞ n
Probability Distribution/Multiple Scaling (PDMS)
method of Lavallee et al. (1991),14 to estimate the Equations (6) and (8) correspond to Cramer’s
co-dimension function cb (γ) from data. Theorem.15
The paper is organized as follows. In Sec. 2, The requirement that K(q) < ∞ for all q can be
we recall some fundamental results in large devi- considerably relaxed. In fact, Eq. (7) is known to
ations due to Cramer (1938)15 and Bahadur and hold under the much milder condition that M (q) is
Ranga Rao (1960).16 In Sec. 3, we use these re- finite in an infinitesimal right and left neighborhood
sults to derive refined limits for P [εn, b ≥ mnγ ] of zero, i.e. under
and P [εn, d ≥ mnγ ]. Section 4 makes three ex- 0 ∈ int(SM ), SM = {q : M (q) < ∞} (9)
tensions. The first extension is to the multi-point
case, for which we determine the asymptotic rough where int(SM ) is the interior of SM ; see for
scaling of P {[εn (x1 ), . . . , εn (xs )]T ∈ mnA }, where example Dembo and Zeitouni (1993), Sec. 2.2.1.18
{x1 , . . . , xs } is any given set of points in Rd , the It is also interesting that, under the condition in
measure density εn (xi ) is either εn, b or εn, d in the Eq. (9), there is a one-to-one correspondence be-
cascade tile of side length m−n that contains xi , A tween the functions M (q) and the distributions of X
is any measurable set in Rs , and r A = {ε : r −ε ∈ [den Hollander (2000), Sec. 1.4].19 Also the con-
A}. The second extension considers multivariate dition that X is unbounded can be relaxed; see
cascades for which the measure density is a vec- Varadhan (1984) for examples with bounded X for
tor ε with components ε1 , . . . , εk . In this case, we which Eqs. (6) and (8) hold.21
120 D. Veneziano

Under Eq. (9), it is possible to refine the expo- I(γ) as


nential limit in Eq. (8). In the case when X is (
continuous, Cramer15 has shown that M (q) = mKb (q/ ln(m))
. (12)
!−1/2 I(γ) = ln(m) · cb (γ)
[I 0 (γ)]2
P [Sn ≥ γ] = 2π n
I 00 (γ) Replacement of I(γ) = ln(m) · cb (γ) into Eq. (11)
gives
× e−nI(γ) (1 + o(1)) (10)
1
where I(γ) is the rate function in Eq. (8), and the lim logm P [εn, b ≥ mnγ ] = −cb (γ) (13)
n→∞ n
0 and 00 signs denote first and second derivative,

respectively. Similar asymptotic expansions for which implies Eq. (1).


cases when X is not continuous have been obtained Similarly, the refined limit in Eq. (10) implies the
by Bahadur and Ranga Rao.16 following asymptotic expansion of P [εn, b ≥ mnγ ]:
! !−1/2
[c0b (γ)]2
3. LARGE DEVIATIONS P [εn, b ≥ m ] = 2π

n ln(m)
c00b (γ)
OF SCALAR CASCADES
× m−ncb (γ) (1 + o(1)) . (14)
Next, we use the results of Sec. 2 to obtain rough
and refined limits for P [εn, b ≥ mnγ ] and P [εn, d ≥ Notice that the factor that multiplies the expo-
mnγ ], where εn, b and εn, d are the bare and dressed nential term is a separable √function of n and γ,
measure densities generated by the cascade model which depends on n as 1/ n. Therefore, for
in Sec. 1. n large, m−ncb(γ) over-estimates P [εn, b ≥ mnγ ].
Equation (14) holds for continuous generators Y .
Similar asymptotic expansions for non-continuous
3.1 Bare Densities generators can be obtained using results in Ref. 16.
d Q d
The bare densities εn, b satisfy εn, b = ni=1 Yi =
mnSn , where Sn = n1 (X1 + X2 + · · · + Xn ) and Xi =
logm Yi . Therefore, using Eq. (8), 3.2 Dressed Densities

1 For the case when the cascade generator is contin-


lim ln P [εn, b ≥ mnγ ] = −I(γ) (11) uous, refined limiting results for P [εn, d ≥ mnγ ] =
n→∞ n
P [εn, b · Z ≥ mnγ ] are derived in the Appendix.
where I(γ) is given by Eq. (6). The functions Kb (q) Using Eq. (14) and the upper tail of the distribu-
and cb (γ) defined in Sec. 1 are related to M (q) and tion of Z, we find [see Eqs. (A.5) and (A.14)]

 !−1/2

 [c0 (γ)]2
 0
 E[Z cb (γ) ] 2π b00 n ln(m) m−ncb(γ) (1 + o(1)), for γ < γ ∗

 cb (γ)
P [εn, d ≥ mnγ ] = (15)



 f 0 (n(γ − γ ∗ )) −ncb (γ ∗ )−q∗ (γ−γ ∗ )

 Z m (1 + o(1)) , for γ > γ ∗
q∗


where fZ 0 is the probability density function of ∗ ≥mnγ ]
m−nc(γ ) with limn→∞ mn,−nc(γ
P [ε
0
d
∗) = ∞, but we
Z 0 = logm (Z). Except for the factor E[Z cb (γ) ], the do not obtain an asymptotic expansion compara-
expansion in Eq. (15) for γ < γ ∗ is the same as that ble to Eq. (15). The above results confirm and for
for bare densities in Eq. (14). For γ = γ ∗ , we find γ 6= γ ∗ refine the rough exponential limits in

in the Appendix the rough limit P [εn, d ≥ mnγ ] ∼ Eqs. (4) and (5).
Large Deviations of Multifractal Measures 121

3.3 Estimation of cb (γ) Evaluation of the estimators ĉb1 (γ) and ĉb2 (γ) is not
easy due to the complicated effects of the dress-
There are two main methods to estimate the co- ing factor Z on the probability P [εn, d ≥ mnγ ].
dimension function cb (γ). One is to first estimate However, one may get an indication of their perfor-
the moment scaling function Kb (q) and then take mance by applying the same estimation procedures
the Legendre transform of Kb (q). The other is to the bare probabilities in Eqs. (1) and (14). In
to use the PDMS method of Lavallee et al.14 The this case, one considers
PDMS method uses Eqs. (4) and (5) to find cb (γ),
γ < γ ∗ , as the slope of logm {P [εn, d ≥ mnγ ]} against 
 d
n, i.e. 
 ĉb1 (γ) = − logm {P [εn, b ≥ mnγ ]}
dn
. (18)
d 
 √
 ĉb (γ) = − d log { nP [εn, b ≥ mnγ ]}
ĉb1 (γ) = − logm {P [εn, d ≥ mnγ ]} . (16) 2 m
dn dn
In practice, the derivative is replaced with the slope Suppose for example that X = logm (Y ) has nor-
of the straight-line regression √ of logm {P [εn, d ≥ mal distribution with variance σ 2 and√mean value

m ]} against n. The factor 1/ n in Eq. (15) for −0.5σ 2 . Then P [εn, b ≥ mnγ ] = Φ(− n( σ2 + σγ ))
γ < γ ∗ suggests the alternative estimator where Φ is the standard normal cumulative distri-
d √ bution function and the derivatives in Eq. (18) have
ĉb2 (γ) = − logm { nP [εn, d ≥ mnγ ]} . (17) theoretical values
dn

     

 d d √ σ γ 1 xφ(x)

 − logm {P [εn, b ≥ m ]} = −

logm Φ − n + =
 dn dn 2 σ 2n ln(m) Φ(−x)
  (19)

 d √ 1 xφ(x)


− logm { nP [εn, b ≥ mnγ ]} = − −1
dn 2n ln(m) Φ(−x)

where φ is the standard normal density function and x = n( σ2 + σγ ). The exact co-dimension function of
x2
1
the cascade is cb (γ) = 2 ln(m) ( σγ + σ2 )2 = 2n ln(m) . Hence the bias factors of the estimators in Eq. (18) are


 d φ(x)

 B (γ, n) = − logm {P [εn, b ≥ mnγ ]}/cb (γ) =
 1 dn xΦ(−x)
. (20)

 d √ φ(x) 1


 B2 (γ, n) = − logm { n P [εn, b ≥ mnγ ]}/cb (γ) = − 2
dn xΦ(−x) x

Figure √ 1 shows plots of these bias factors against


x = n( σ2 + σγ ). For x < 1, both estimators are
highly biased. This is especially true for ĉb2 (γ).
However, convergence to 1 is faster for B2 (γ, n)
than for B1 (γ, n). For example, when x = 3,
ĉb2 (γ) is practically unbiased, whereas ĉb1 (γ) over-
estimates cb (γ) by about 10%.

4. EXTENSIONS
In this section, we extend the rough limits of the
previous section to multi-point and multivariate ex-
tremes and to extremes of cascades with dependent
√ σ γ
Fig. 1. Bias factors in Eq. (20) for different x = n( 2 + σ ). generators.
122 D. Veneziano

4.1 Multi-Point Extremes where A is a measurable set of Rs and mnA =


{mnε : ε ∈ A}. A similar generalization holds for
Schertzer and Lovejoy,13 p. 51, have generalized dressed densities.
Eq. (1) to multi-point extremes. Let {x1 , . . . , xs }
be any given set of points in Rd and denote by
εr, b (xi ) the bare measure density at location xi 4.2 Multivariate Extremes
when the cascade is developed to resolution r.
Schertzer and Lovejoy13 state that A theory of vector-valued multifractal measures
( ) based on Lie cascades has been developed by
\
s
P [εr, b (xi ) ≥ r γi ] ∼ r −cb, s (γ) (21) Schertzer and Lovejoy.1,3,13,22 These measures have
i=1 a renormalization property that in the most general
case involves groups of random transformations of
where γ is the vector of exponents γi and cb, s (γ) both the space and the vector field. Here, we con-
is the Legendre transform of the moment scaling sider the case of vector-valued measure densities
function Kb, s (q) that satisfies ε(B) = [ε1 (B), . . . , εk (B)]T , B ⊂ Rd , generated
( ) ( ) by discrete multiplicative cascades. At level 0, the
Y
s Y
s
E [εr, b (xi )] qi
=r Kb, s (q)
·E [ε1, b (xi )] qi
. measure densities εi are deterministic and uniform
i=1 i=1 inside the unit cube of Rd . We further assume that
(22) all the components of the cascade have multiplicity
m and follow the cascade scheme described in Sec. 1
Schertzer and Lovejoy13 further notice that the with generator {Y ni }, where the Y ni are indepen-
functions Kb, s (q) and cb, s (γ) in Eqs. (21) and (22) dent copies of a non-negative random vector Y . The
are sensitive to r and to the distances between the components Y1 , . . . , Yk of Y have mean value 1 and
points xi . are possibly dependent. In analogy with the scalar
We evaluate the exponent cb, s (γ) in Eq. (21) case, we are interested in the large deviations of the
using a more direct approach. Let no be the small- bare density vector εn, b as n → ∞.
est cascade level at which each point xi belongs to For this case, the analysis of Schertzer and
a different cascade tile and put n0 = n − no . Then Lovejoy,22 p. 33, uses a diagonal matrix-valued
the densities εn, b (xi ) are obtained as the product of moment scaling function K b (q), defined as
the bare densities at level no , εno ,b (xi ), and indepen-
dent factors εn0 ,i all with the distribution of εn0 (x). K b (q) = logm E{[diag(Yi )]q } = diag(Kb,i (q)) .
As n → ∞, the variables εno ,b (xi ) make a negligi- (25)
ble contribution to the scaling of P {∩si=1 [εn, b (xi ) ≥
mnγi ]}, and one obtains Notice that K b (q) in Eq. (25) does not reflect any
( ) P
s dependence that might exist among the variables
\
s −(n−no ) cb (γi )
Yj . Schertzer and Lovejoy (1995)22 also introduce
P [εn, b (xi ) ≥ mnγi ] ∼m i=1
(but do not explicitly define) a vector of singulari-
i=1
ties γ and a co-dimension function cb (γ) and state
P
s
−n cb (γi ) [see their Eq. (22)] that
∼m i=1 . (23)
P {εn, b ∈ Sn (γ)} ∼ m−ncb (γ) (26)
Hence,
Ps
Eq. (21) holds with exponent cb, s (γ) =
i=1 cb (γi ). The
P
Legendre transform of this expo- where Sn (γ) is the quadrant Sn (γ) = {ε: εi ≥
nent is Kb (q) = si=1 Kb (qi ), which is also the limit mnγi }. Next, we use large deviation theory to show
of Kb, s (q) in Eq. (22) for n → ∞. Through a sim- that Eq. (26) applies and evaluate the exponent
ilar argument one finds that, in the dressed case, cb (γ). We also show that a definition of the moment
the multi-point exceedance
P
probabilities scale with scaling function different from Eq. (25) is appropri-
exponent cd,s (γ) = si=1 cd (γi ). A generalization of ate for vectors. Finally, we consider the probability
Eq. (23) is content of regions of general shape, not necessarily
P {[εn, b (x1 ), . . . , εn, b (xs )]T ∈ mnA } quadrants.
  Cramer’s Theorem has the following multi-
P
s
−n inf cb (γi ) dimensional extension (e.g. Ref. 21, Sec. 4; or
∼m γ∈A i=1 (24) Ref. 19, Chap. 5). Let X 1 , X 2 , . . . , X n be
Large Deviations of Multifractal Measures 123

independent copies of a random vector X with k I(γ) = ln(m) · cb (γ) . (31)


components and denote by Pn the joint distribu-
tion of the average S n = n1 (X 1 + X 2 + · · · + X n ). It also follows from S n = n1 logm εn, b that Pn (A) =
Suppose that the origin is an interior point of the P [Sn ∈ A} = P [εn, b ∈ mnA ]. Using this last rela-
T
set SM = {q : M (q) = E[eq X ] < ∞}. If X is tion and Eq. (31), Eq. (28) gives
unbounded, then the sequence {Pn } satisfies the
large deviation principle with rate function 1
lim logm P [εn, b ∈ mnA } = − inf cb (γ) . (32)
n→∞ n γ∈A
I(γ) = sup{q T γ − ln M (q)} . (27)
q To exemplify, we consider multivariate extensions
of the log-normal and beta cascades.
This means that, for any measurable set A ⊂ Rk ,
ln Pn (A) Example 4.1. (multivariate log-normal cas-
lim = − inf I(γ) . (28) cades). In this case, Y = eZ , where Z ∼
n→∞ n γ∈A
N (µ, Σ) and µi = − 12 Var[Zi ], so that E[Yi ] = 1.
Equation (28) is a multivariate extension of Eq. (7). Using properties of the multivariate normal distri-
Now we use Eq. (28) to obtain a scaling relation bution (e.g. Johnson and Kotz (1983),23 p. 39; or
for P {εn, b ∈ mnA }. We set X = logm Y where Varadhan (1984),12 p. 13), the functions M (q),
Y = [Y1 , . . . , Yk ]T is the random generator vector Kb (q) and cb (γ) are given by
of the cascade. For q a vector of moment orders   
 1 T
q1 , . . . , qk and γ a vector of components γ1 , . . . , γk , 

T
M (q) = exp q µ + (q Σq)

 2
we define the moment scaling function Kb (q) as 


  
 1 1 T
" # Kb (q) = logm M (q) = T
q µ + (q Σq) .
Y
k
 ln(m) 2
Yiqi 

Kb (q) = logm E = logm M (q ln(m)) 



i=1  1

 cb (γ) = (γ − µ)T Σ−1 (γ − µ)
(29) 2 ln(m)
(33)
and cb (γ) as the Legendre transform of Kb (q),
Example 4.2. (multivariate beta cascades).
cb (γ) = sup{γ T q − Kb (q)} . (30)
q In this case, Yi attains value zero with probabil-
ity 1 − Pi and value 1/Pi with probability Pi . Let
Then, in analogy with Eq. (13), P + be the probability that all Yi are nonzero. Then



 Xk

 K (q) = log P +− qi logm Pi

 b m
i=1
( (34)


 c (γ) = logm P , for γi ≤ logm Pi , i = 1, . . . , k
+


 b
∞, otherwise .

Notice that, if the set A does not intersect the to the exponential term depends on the curvature
quadrant {ε : εi ≤ logm Pi }, then inf γ∈A cb (γ) = ∞ of the set A near the dominating point.
and P [εr, b ∈ r A ] = 0. The previous results are for bare densities. A
general scaling theory for multivariate dressed den-
An extension to the multivariate case of the refine- sities appears difficult. In the trivial case when
ment in Ref. 16 has been made by Ney (1983 and the cascade components are independent, cd (γ) =
Pk
1984).24,25 A complication that arises in spaces of i=1 cdi (γi ), where the functions cdi (·) are given by
dimension d > 1 is that the asymptotic pre-factor Eq. (5) with (γi , γi∗ , qi∗ ) in place of (γ, γ ∗ , q ∗ ).
124 D. Veneziano

4.3 Extremes of Cascades with Notice that ±∞ are allowed in Eq. (35) both as
Dependent Generators limits and as elements of the sequence
{ n1 logm Mn (q)}. If the origin is an interior point
Up to now we have considered cascades with in-
of SK = {q : Kb (q) < ∞}, then by the Gartner-
dependent generators. Multifractal measures may
Ellis Theorem, Eq. (32) holds and the exponent
in general arise also from cascade constructions
cb (γ) in that equation is the Legendre transform
in which the variables Yni are dependent.10,11,26
of Kb (q) in Eq. (35). Although this result holds
For example, Carsteanu and Foufoula-Georgiou
under no further restriction on the dependence
(1996)27 have found evidence of such dependence in
structure of the generator, in the case of cascades
temporal rainfall. In the dependent case, one may
it is often natural to assume that Y 1 , Y 2 , . . . is
use the Gartner-Ellis Theorem28,29 to obtain large
a stationary Markov sequence. Next we give two
deviation properties of the cascade, as follows.
examples. In the first case, the sequence {X i }
Consider in general a vector-valued cascade with
is stationary Gaussian with no special correlation
generator {Y 1 , Y 2 , . . .} at a generic point x of the
unit cube Sd . The sequence {Y n } is the subset of structure. In the second example, we specialize the
the complete generator {Y ni } obtained by selecting results for the case when {X i } is an autoregressive
for each level n the index i of the tile that contains x. sequence.
P
Let X i = ln Y i and Mn (q) = E[exp{q T ni=1 X i }]
Example 4.3. ({X i } stationary Gaussian).
for n = 1, 2, . . . . The vectors X i are assumed to be
Suppose that {X i } is stationary Gaussian with
weakly dependent in the sense that
mean value µ and covariance function C i =
1
Kb (q) = lim logm Mn (q) ∈ [−∞, ∞] exists for all q . E[(X 0 −µ)(X i −µ)T ], i = 0, ±1, . . . , such that C =
n→∞ n P∞ Pn Pn
Pni=−∞ C i is finite. Then i=1 X i ∼ N (nµ, i=1
(35)
j=1 C i−j ) and

       
 1 X n X n   1 X
n−1 
Mn (q) = exp nqT µ + qT  C i−j  q = exp nq T µ + q T  (n − |i|)C i  q . (36)
 2   2 
i=1 j=1 i=−(n−1)

Hence
 
1 1 1 1
Kb (q) = lim ln Mn (q) = q T µ + q T Cq . (37)
ln(m) n→∞ n ln(m) 2

For conservation of the cascade mean as n → ∞, Example 4.4. ({Xi } autoregressive Gaussian).
it must be µ = − 12 v, where v is the vector whose As an example of Markov generator, suppose that
ith component is the ith diagonal element of C. {X i } satisfies X i = AX i−1 + BZ i , i = 1, 2, . . .
Hence where A is a k × k full rank matrix with absolute
 eigenvalues strictly less than 1 and the Z i are iid
 1

 K (q) = (q T v + q T Cq) normal vectors with mean value µ and covariance
 b 2 ln(m)
    matrix Σ. A restriction on µ and Σ to make the

 1 1 T −1 1 cascade (asymptotically) conservative in the mean

 cb (q) = γ− v C γ− v .
2 ln(m) 2 2 will be imposed later. The moment generating func-
(38) tion of Z is

Comparison with Eq. (33) shows that these are the  


TZ 1
same moment scaling and co-dimension functions as MZ (q) = E[eq ] = exp q T µ + qT Σq .
2
those of a cascade with independent generator and
ln Y ∼ N (− 12 v, C). (39)
Large Deviations of Multifractal Measures 125

Following Bucklew (1990),17 p. 22, This implies that the asymptotic moment scaling
function Kb (q) and co-dimension function cb (γ) of
Y
n−1
the cascade are Kb (q) = logm (λ(q)) and cb (γ) =
Mn (q) = MZ ([(I −An−i )(I −A)−1 B]T q ). (40)
supq {qγ − Kb (q)} = ln(m) 1
I(γ ln(m)), respectively.
i=0
For the present case, Waymire and Williams10
Therefore, Kb (q) in Eq. (35) is (Theorem 2.1) have shown that the fully developed
cascade is non-degenerate if and only if Kb0 (1− ) < 1
Kb (q) = logm MZ ([(I − A)−1 B]T q) and that the dressing factor Z has finite moment
of order q > 1 if and only if q < q ∗ = inf{q ≥
= logm MZ (QT q) (41) 1 : Kb (q) < q − 1}. These results are formally
identical to those for independent cascades. There-
where I is the identity matrix and Q = (I − A)−1 B. fore, the considerations in Sec. 3 on the extremes of
Substitution of MZ in Eq. (39) gives dressed independent cascades apply without modi-
  fication also to the present Markov case.
1 1
Kb (q) = q T Qµ + q T QΣQT q . (42) Miller’s Theorem allows one to extend the pre-
ln(m) 2
vious results to cascade generators Yi given by
For asymptotic conservation of the cascade mean as Yi = ef (Xi ) , where the sequence {Xı } is Markovian
n → ∞, one must have Qµ = − 12 v where v is the as above and f (·) is any given real-valued function.
vector with ith component equal to the ith diagonal Further extension to Yi = ef (Xi+1 , Xi ) is also possi-
element of the matrix QΣQT . Hence, ble. In the latter case, λ(q) is the largest eigenvalue
of the k × k matrix with (i, j)th coefficient given by
 pij eqf (uj , ui ) ; see for example Ref. 17, Chap. 5.
 1

 K (q) = (q T v+qT QΣQT q)
 b 2 ln(m)
   T

 1 1 1 5. CONCLUSIONS

 cb (γ) = γ + v (QΣQT )−1 γ + v .
2 ln(m) 2 2
We have obtained large deviation properties of
(43) multifractal measures generated by multiplicative
cascades. In particular, we have analyzed the be-
havior of P [εn ≥ mnγ ] for n → ∞, where m is
In the scalar case, replacement of (A, B, Σ) with the multiplicity of the cascade, εn is the bare or
σ2 b 2 2
(a, b, σ 2 ) gives Kb (q) = 2 ln(m) ( 1−a ) (q + q). dressed measure density at level n, and γ is a con-
stant. By using large deviation results for sums of
As a final class of multifractal models, we men- iid variables, we have sharpened existing limits for
tion the case of Markov cascades with finite state P [εn ≥ mnγ ]; see Eqs. (14) and (15). We have
space, which have been studied by Waymire and used Eq. (15) to modify the PDMS estimator of the
Williams10 relative to the problems of existence, co-dimension function c(γ) based on exceedance fre-
divergence of the moments, and carrying dimen- quency data; see Eq. (17) and the example in Fig. 1.
sion. Here we look at the problem of extremes. For multi-point extremes, our results [Eqs. (23) and
Specifically, consider a scalar cascade with gener- (24)] correspond to the large-resolution limit of a
ator {Y1 , Y2 , . . .} at a generic point x of the sup- formula proposed by Schertzer and Lovejoy.13 For
port. Suppose that {Xi } = {ln Yi } is an irreducible multivariate extremes, our co-dimension function
Markov chain with finite state space {u1 , . . . , uk } [Eqs. (29) and (30)] differs from previous results.
and transition probabilities pij = P [Xn+1 = Finally, we have used results from large deviation
uj |Xn = ui ] and let Sn = n1 (X1 + · · · + Xn ). Then, theory to find the moment scaling and co-dimension
by Miller’s Theorem (1961),30 Sn obeys the large functions of cascades with dependent generators.
deviation principle with rate function Our focus on the very large deviations P [εn ≥
mnγ ] is due to the fact that traditionally, these
I(γ) = sup(qγ − ln λ(q)} (44) are the extremes considered in the multifractal
q
literature. An example application of the large-n
where λ(q) is the largest eigenvalue of the k × k behavior of P [εn ≥ mnγ ] is the scaling of the so-
matrix with (i, j)th coefficient given by pij equj . called intensity-duration-frequency (IDF) curves in
126 D. Veneziano

hydrology. Let I(D, T ) be the average precipita- 2. V. K. Gupta and E. C. Waymire, J. Geophys. Res.
tion intensity in an interval of duration D that is 95, 1999 (1990).
exceeded with a return period of T years, meaning 3. D. Schertzer and S. Lovejoy, in Space/Time Vari-
that I(D, T ) is exceeded with probability 1/T in ability and Interdependence for Various Hydrological
one year. The IDF curves are plots of I versus D Processes, ed. R. A. Feddes (Cambridge University
for given T . It has long been noticed that the IDF Press, 1995), pp. 153–173.
4. D. Veneziano, Fractals 7, 59 (1999).
curves have a power decay with duration D, and
5. U. Frish and G. Parisi, in Turbulence and Predictabil-
that the exponent is approximately independent of
ity in Geophysical Fluid Dynamics and Climate
T.31 The exponent can be predicted using a multi- Dynamics, eds. M. Ghil, R. Benzi and G. Parisi
fractal characterization of temporal rainfall and the (Elsevier North-Holland, 1985), pp. 84–88.
scaling properties of P [εn ≥ mnγ ].32 – 34 6. D. Schertzer and S. Lovejoy, J. Geophys. Res. D92,
In Sec. 3, we have obtained the large-n behav- 9693 (1987).
ior of P [εn, b ≥ mnγ ] from the large-n properties 7. V. K. Gupta and E. C. Waymire, J. Appl. Meteor.
of P [Sn > γ] where Sn = n1 (X1 + X2 + · · · + Xn ) 32, 251 (1993).
and the variables Xi are iid. Some of the liter- 8. B. B. Mandelbrot, J. Fluid Mech. 62, 331 (1974).
ature on large deviations is concerned with less 9. J.-P. Kahane and J. Peyriere, Advs. Math. 22, 131
extreme deviations, of the type P [Sn ≥ n−β γ] (1976).
where 0 < β < 0.5.35 For example, when the vari- 10. E. C. Waymire and S. C. Williams, “Branching
ables Xi have√zero mean and unit variance, the dis- Processes and Applications,” in IMA Volume in
tribution of n · Sn is attracted to the standard Applied Mathematics, eds. K. Athreya and P. Jagers
(Springer-Verlag, 1995).
normal CDF, Φ, and interest is in the behavior of
P [Sn ≥n−β γ] 11. E. C. Waymire and S. C. Williams, Trans. Am.
1−Φ(n0.5−β γ)
as n → ∞. To apply these results to Math. Soc. 348, 585 (1996).
cascades, one may set X = logm Y. If X has finite 12. T. C. Halsey, M. H. Jensen, L. P. Kadanoff, I.
mean value µ and finite variance σ 2 , then the bare Procaccia and B. Shraiman, Phys. Rev. A33, 1141
density εn, b has the distribution of mn(µ+σSn ) , and (1986).
n o 13. D. Schertzer and S. Lovejoy, in Non-Linear Vari-
−β+1
P [Sn ≥ n−β γ] = P εn, b ≥ m(µ+σγ)n . ability in Geophysics: Scaling and Fractals, eds. D.
Schertzer and S. Lovejoy (Kluwer, 1991), pp. 41–82.
(45) 14. D. Lavallee, S. Lovejoy and D. Schertzer, in
Non-Linear Variability in Geophysics: Scaling and
Hence, results on P [Sn ≥ n−β γ] can be used Fractals, eds. D. Schertzer and S. Lovejoy (Kluwer,
to characterize the non-extreme large deviations 1991), pp. 99–110.
of multifractal measures. The case β = 0 was 15. H. Cramer, in Actualites Scientifiques et Indus-
studied in Sec. 3. trielles, No. 736, Colloque consacre a la theorie des
probabilites (Hermann, 1938), pp. 5–23.
16. R. R. Bahadur and R. Ranga Rao, Ann. Math. Stat.
ACKNOWLEDGMENTS 31, 1015 (1960).
17. J. A. Bucklew, Large Deviation Techniques in
The author is grateful to Daniel Stroock for Decision, Simulation, and Estimation (John Wiley
useful discussions on the refined limits obtained and Sons, 1990).
in the Appendix. This study was supported by 18. A. Dembo and O. Zeitouni, Large Deviations Tech-
the Italian National Research Council under a niques and Applications (Jones and Bartlett, 1993).
Cooperative Agreement with MIT for the Study 19. F. den Hollander, Large Deviations (American
of Climatic Changes and Hydrogeologic Risks and Mathematical Society, 2000).
by the National Science Foundation under Grant 20. D. W. Stroock, Probability Theory: An Analytical
CMS-9612531. View (Cambridge University Press, 1999).
21. S. R. S. Varadhan, Large Deviations and
Applications (Society for Industrial and Applied
Mathematics, 1984).
REFERENCES
22. D. Schertzer and S. Lovejoy, in Fractals in
1. D. Schertzer and S. Lovejoy, Resolution Dependence Geoscience and Remote Sensing, Image Under-
and Multifractals in Remote Sensing and Geograph- standing Research, Vol. 1, eds. G. Wilkinson, I.
ical Information Systems, Lecture Notes (McGill Kanellopoulos and J. Megier (Institute for Remote
University, 1996), pp. 390. Sensing Applications, 1995), pp. 11–40.
Large Deviations of Multifractal Measures 127

23. N. L. Johnson and S. Kotz, Distributions in lity P [εn,b · Z ≥ mnγ ] as n → ∞ when the cascade
Statistics: Continuous Multivariate Distributions generator Y is a continuous variable. For the
(John Wiley and Sons, 1983). cascade to be non-degenerate, the function K(q) =
24. P. Ney, Ann. Prob. 11, 158 (1983). logm E[Y q ] must satisfy K 0 (1) < 1. If in addition
25. P. Ney, Ann. Prob. 12, 903 (1984). K(q) < q − 1 for all q > 1, then all the positive
26. E. C. Waymire and S. C. Williams, Bull. AMS 31, moments of Z are finite. Otherwise E[Z q ] diverges
216 (1994).
for q ≥ q ∗ , where q ∗ > 1 satisfies K(q ∗ ) = q ∗ − 1.9
27. A. Carsteanu and E. Foufoula-Georgiou, J. Geophys.
Res. D101(26), 26363 (1996).
In the latter case, the probability density function

28. J. Gartner, Th. Prob. Appl. 22, 24 (1977). of Z behaves like fZ (z) ∼ z −q −1 for large z and the
29. R. S. Ellis, Ann. Prob. 12, 1 (1984). probability density function of Z 0 = logm Z behaves
0 ∗
30. H. Miller, Ann. Math. Stat. 32, 1260 (1961). like fZ 0 (z 0 ) ∼ m−z q for large z 0 . The slope of K(q)
31. P. Burlando and R. Rosso, J. Hydrol. 187, 45 (1996). at q ∗ , γ∗ = K 0 (q ∗ ), plays an important role in the
32. H. Benjoudi, P. Hubert, D. Schertzer and S. Lovejoy, analysis that follows. We set γ ∗ = ∞ if q ∗ does not
Geosci. Surface (Hydrol.)/Surface Geosci. (Hydrol.) exist.
325, 323 (1997). We start from the relation in Eq. (14) which,
33. H. Benjoudi, P. Hubert, D. Schertzer and S. omitting for simplicity the subscript b for bare, we
Lovejoy, EGS, 24th General Assembly (The Hague, write as
The Netherlands, 1999).
34. P. Furcolo and D. Veneziano, AGU Fall Meeting (San !−1/2
[c0 (γ)]2
Francisco, 1999). P [εn ≥ m ] ≈ 2π 00

n ln(m) m−nc(γ) .
35. I. A. Ibragimov and Yu. V. Linnik, Independent and c (γ)
Stationary Sequences of Random Variables (Wolters- (A.1)
Noordhoff, 1971).
Here and in what follows, the symbol ≈ stands for
equality up to a factor that approaches 1 as n → ∞.
APPENDIX: REFINED LIMITS
We obtain the probability P [εn · Z ≥ mnγ ] by first
FOR DRESSED CASCADES
conditioning on Z and then taking expectation with
We derive asymptotic expressions for the probabi- respect to Z. This gives

z0
P [εn · Z ≥ mnγ ] = P [εn ≥ mn(γ− n ) ]
    −1/2
z0 2
Z
∞  γ− c0 
 n  z0
≈ 2π  0  n ln(m) m−nc(γ− n ) fZ 0 (z 0 )dz 0 (A.2)
−∞  00
z 
c γ−
n
where Z 0 = logm Z. To obtain more explicit results, we consider separately the cases γ < γ ∗ , γ = γ ∗ and
γ > γ∗.

(a) γ < γ ∗
In this case, we expect the dressed probability P [εn · Z ≥ mnγ ] to have the same rough exponential limit
as the bare probability P [εn ≥ mnγ ]. Therefore, it makes sense to multiply and divide the right-hand-side
of Eq. (A.2) by P [εn ≥ mnγ ]. Using Eq. (A.1), we obtain
P [εn · Z ≥ mnγ ] ≈ P [εn ≥ mnγ ]
  
z0 1/2
Z c00 γ−
∞ c0 (γ)  n  z0
×    m−n[c(γ− n )−c(γ)] fZ 0 (z 0 )dz 0
−∞ z 
0
c00 (γ) 
c0 γ−
n
Z ∞ z0
≈ P [εn ≥ m ] nγ
m−n[c(γ− n )−c(γ)] fZ 0 (z 0 )dz 0 (A.3)
−∞
128 D. Veneziano
0 0 0
Next, we replace c(γ − zn ) with the Taylor series expansion c(γ) − c0 (γ) zn + R( zn ) where R(x) = 0(x2 ) is a
residual term. This gives
Z ∞ 0 0 0
P [εn · Z ≥ m ] ≈ P [εn ≥ m ] nγ nγ
mc (γ)z −nR(z /n) fZ 0 (z 0 )dz 0 . (A.4)
−∞

As n → ∞, nR(z 0 /n) → 0 for any given z 0 . Moreover, it Rfollows from convexity of c(γ) that c0 (γ) <
0 ∞ 0 0
c0 (γ ∗ ) = q ∗ when γ < γ ∗ . Hence, the moment E[Z c (γ) ] = −∞ mc (γ)z fZ 0 (z 0 )dz 0 is finite and Eq. (A.4)
gives
0
P [εn · Z ≥ mnγ ] ≈ E[Z c (γ) ]P [εmn ≥ mnγ ]
!−1/2
c0 (γ) [c0 (γ)]2
≈ E[Z ] 2π 00 n ln(m) m−nc(γ) . (A.5)
c (γ)

(b) γ = γ ∗

0 ∗ ∗ P [εn ·Z≥mnγ ]
For γ = γ ∗ , the moment E[Z c (γ ) ] = E[Z q ] diverges. Therefore, using Eq. (A.5), P [εn ≥mnγ ∗ ]
→ ∞ as
P [εn ·Z≥m nγ ∗
n → ∞. Next we show that, for any γ < γ ∗ , P [εn ≥mnγ ]
]
→ 0. From this, we conclude that P [εn · Z ≥
∗ ∗
mnγ ] ∼ m−nc(γ ) ,where ∼ denotes asymptotic equality up to a slowly varying function of mn .

Let γ = γ − ∆, where ∆ > 0. Multiplication and division of the right-hand-side of Eq. (A.2) by

P [εn ≥ mn(γ −∆) ] and use of Eq. (A.1) gives
∗ ∗ −∆)
P [εn · Z ≥ mnγ ] ≈ P [εn ≥ mn(γ ]
  
z0 1/2
Z − c00 γ∗
∞ c0 (γ ∗ − ∆)   n 
0
−n[c(γ ∗ − zn )−c(γ ∗ −∆)]
×  0   c00 (γ ∗ − ∆)  m fZ 0 (z 0 )dz 0 . (A.6)
−∞ z
c0 γ ∗ −
n
z0
We expand c(γ ∗ − n) in Taylor series around γ ∗ − ∆. Retaining terms up to the second order,
∗ − z 0 )−c(γ ∗ −∆)] 0 ∗ −∆)(∆− z 0 )+ 1 c00 (γ ∗ −∆)(∆− z 0 )2 +R(∆− z 0 )]
m−n[c(γ n = m−n[c (γ n 2 n n (A.7)

where R(x) = 0(x3 ) is a residual term. Using Eq. (A.7), the integral in Eq. (A.6) becomes
  
z0 1/2
Z ∞ 0 ∗ − c00 γ∗
0 ∗ −∆)∆ c (γ − ∆)  n 
m−nc (γ  



0 ∗ 0
c00 (γ ∗ − ∆)
−∞ c γ − z
n

0 ∗ −∆)z 0 − n c00 (γ ∗ −∆)(∆− z 0 )2 −nR(∆− z 0 )


×mc (γ 2 n n fZ 0 (z 0 )dz 0 (A.8)
0 0
Since c(γ) is convex, the term − n2 c00 (γ ∗ − ∆)(∆ − zn )2 − nR(∆ − zn ) is negative for all z 0 6= n∆. Therefore,
neglecting this term in Eq. (A.8) produces the upper bound
∗ ∗ −∆) 0 ∗ −∆)∆
P [εn · Z ≥ mnγ ] < P [εn ≥ mn(γ ]mc (γ
   1/2
z0
Z − c00 γ∗
∞ c0 (γ ∗ − ∆)   n 
 0 ∗ −∆)z 0
×  0   00 ∗  mc (γ fZ 0 (z 0 )dz 0 . (A.9)
−∞ z c (γ − ∆)
c0 γ ∗ −
n
Large Deviations of Multifractal Measures 129
0 ∗
Using again convexity of c(.), it is c0 (γ ∗ − ∆) < q ∗ and since fZ 0 (z 0 ) ∼ m−z q for large z 0 , the inte-
P [εn ·Z≥m ] nγ ∗
gral in Eq. (A.9) is finite. Hence, limn→∞ ∗
P [εn ≥mn(γ −∆) ]
< ∞ for any given ∆ > 0. This relation and

P [εn ·Z≥mnγ ] ∗ ∗

P [εn ≥mnγ ]
→ ∞ imply the rough exponential limit P [εn · Z ≥ mnγ ] ∼ m−nc(γ ) .

(c) γ > γ ∗
Let ∆ = γ ∗ − γ > 0. We proceed as for γ < γ ∗ , but since in this case we expect the rough asymptotics of

P [εn · Z ≥ mnγ ] to be those of P [εn ≥ mnγ ], we multiply and divide the right-hand-side of Eq. (A.2) by

P [εn ≥ mnγ ]. Using Eq. (A.1), we obtain
∗ +∆) ∗
P [εn · Z ≥ mn(γ ] ≈ P [εn ≥ mnγ ]
  
00 ∗ z0 1/2

c γ + ∆ −
Z ∞ c0 (γ ∗ ) n  ∗ +∆− z 0 )−c(γ ∗ )]
×    m−n[c(γ fZ 0 (z 0 )dz 0
z0  
n
−∞ c00 (γ ∗ )
c0 γ ∗ + ∆ −
n
(A.10)
z0
We now take the Taylor series expansion of c(γ ∗ + ∆ − n) around γ ∗ . Using q ∗ = c0 (γ ∗ ), this gives
     2  
z0 z0 1 z0 z0
c γ∗ + ∆ − − c(γ ∗ ) = q ∗ ∆ − + c00 (γ ∗ ) ∆ − +R ∆− (A.11)
n n 2 n n
∗ z0
where R(x) = 0(x3 ). Also let fZ 0 (z 0 ) = h(z 0 )m−q where h(z 0 ) is slowly varying at infinity. Then,
Eq. (A.10) becomes
∗ +∆) ∗ ∗
P [εn · Z ≥ mn(γ ] ≈ P [εn ≥ mnγ ] · m−n∆q
  
z0 1/2
Z c00 γ∗ +∆−
∞ c0 (γ ∗ ) 
 n 
×    
−∞ z0 00 ∗
c (γ )
c0 γ ∗ + ∆ −
n
0 0
−n c00 (γ ∗ )(∆− zn )2 −nR(∆− zn )
×m 2 h(z 0 )dz 0 . (A.12)

For large n, the integral in Eq. (A.12) is contributed mainly by values of z 0 /n close to ∆. Then, using the
fact that h is slowly varying, that integral may be written as
Z s
∞ 0 n
−n c00 (γ ∗ )(∆− zn )2 0 n∆q ∗
h(n∆) m 2 dz = m fZ 0 (n∆) 2π . (A.13)
−∞ c00 (γ ∗ )

From Eqs. (A.1), (A.12) and (A.13) and the relation c0 (γ ∗ ) = q ∗ , we obtain
s
n(γ ∗ +∆) nγ ∗ −nq ∗ ∆ n
P [εn · Z ≥ m ] ≈ P [εn ≥ m ]·m · h(n∆) 2π
c00 (γ ∗ )

fZ 0 (n∆) −n[c(γ ∗ )+q∗ ∆]


≈ m . (A.14)
q∗

You might also like