WATER RESOURCES RESEARCH, VOL. 33, NO. 4, PAGES 907-908, APRIL 1997
Comment on "Evidence for non-Gaussian scaling behavior
in heterogeneous sedimentary formations" by Scott Painter
Hui Hai Liu and Fred J. Molz
Environmental Systems Engineering Department,Clemson University, Anderson, SouthCarolina
For some time hydrologists haveaccepted, more or less asa
rule of thumb rather than as an absolute fact, that log (K)
measurements in manysedimentary formations are distributed
lognormally.Now that more detailed K measurements are
becomingavailable through the use of borehole flowmeters
and other geophysical loggingequipment,it is important to
examine the lognormal assumption in greaterdetail.This ques-
tion is important also whenselecting a fractal-based procedure
for generating K or other property distributions based on a set
of relatively small-scale measurements [Hewett, 1986; Molz and
Boman, 1993, 1995;Painter and Paterson,1994; Painter, 1996b;
Liu and Molz, 1996].
Painter[1996a]is to be commended for his excellent paper
wherein he describes evidence for non-Gaussian behavior in
hydraulic conductivity and porosity data. As an alternative to
the Gaussian model he proposes the L6vy-stable probability
distribution, a distribution with tailsthat decrease according to
P(Ixl >- u) • u-", which is much more slowly than the
exponentialdecrease associated with the Gaussiandistribu-
tion. This causes the L6vy-stable distribution to have second
and higher moments that are infinite [Samorodnitsky and
Taqqu, 1994]. As shown in (1),
i f0 © (ax) da
-- exp(-Iekl cos
f(x) = rr
the L6vy-stable process is characterized by two parameters, a
so-called L6vy index, a, and a width parameter,C. The distri-
bution converges to a Gaussian distribution for a = 2 [Man-
tegna, 1994].
The purposeof this commentis not to disagree with the
main conclusions of Painter[1996a]concerning non-Gaussian
scaling.Rather, we would like to present some interesting
results from a brief analysis of the macrodispersion experiment
(MADE) dataset,a collection of several thousand K measure-
ments made using a borehole flowmeter at ColumbusAir
Force Base, Mississippi [Boggs et al., 1992, 1993].This data set
is more extensive than that examined by Painter[1996a].
Two methods were used to estimate a values for probability
density functions of increments of In (K). The first is a curve-
fitting method basedon (1), wherein a and C valueswere
estimated by numerically minimizing the following mean
square deviation:
= • [f(xj) -- f*(xj)] 2
j=l
(2)
where f* (x i) is the observed histogram value when an incre-
mentof In (K) is equal to xi. The second is the method of
Copyright 1997by the AmericanGeophysical Union.
Paper number96WR03788.
0043-1397/97/96WR-03788509.00
Fama and Roll [1972],used by Painter[1996a]. Shown in Fig-
ures1 and 2 are comparisons between the observed histograms
and L6vy-stable distributions with parametervalues estimated
from the firstmethod. Figure 1 is for a lag, Az, of 0.15 m, which
is the same as the measurement interval at the MADE site.
Figure 2 corresponds to a Az of 2.4 m. Apparently, the calcu-
lated L6vy-stable distributions matchthe observed histograms
very well. However, we found that a values, estimatedwith
both methods, are not constant but are lag-dependent (Figure
3). The a value consistently increases with lag and converges
approximately to 2, corresponding to the Gaussian distribu-
tion, when the lag getslarge.
The scale-variant behavior of a may be a very important
phenomenon in heterogeneous sedimentary formations, al-
thoughfurther research is needed to confirm this finding. It
implies that a distribution of In (K) at a smaller scale is more
heterogeneous relativeto that at a larger scale. Physically, this
makes sense, because an abruptchange in a subsurface prop-
erty often occurs over a smallvertical interval. The observed
behavior also seems to indicate that the Gaussian distribution
may be applicable abovea certainscale.
Mechanisms behind this scale-variant behavior of a are not
very clear at this stage. However, the followinganalysis of an
idealized case, in which increments of In (K) are assumed to be
independent for smalllags, may shedsome light on the mech-
anisms. An increment of W -- In (K) at a relatively large lag
nAz, where n is an integerlarger than 1, can be expressed as
n
•)n • W(Z1 q- I'IAZ) -- W(z1)= Z [W(zi) - W(zi-1)]
i=2
(3)
10 0
0. t
1
10-2
o
oo
10-•
-10 ß -5 0 5 10
increment of In(K)
Figure 1. Comparison betweenthe histogram of increments
of In (K) for lag Az = 0.15 m and a L6vy-stable distribution
with a = 0.8 and C = 0.36, estimated with the curve-fitting
method.
907