Approximations using Dawsons integral

(by Chris H. Rycrof .) t

Consider a random walk following a fat-tailed distribution with PDF
2
p(x) =
(1+x
4
)
.
ThisPDFhaszeromeanandunitvariance,buttheabsolutemomentsoftheform|x|
m
diverge
for m3. Near the origin, the Fourier transform is not analytic, and has the expansion
k
2
3
k
4
p(k) = 1
2
+
3
|k

|
2

24
+O(|k|
5
).
The cumulant generating function has the form
(k) =

2
k
2
+
3
|k

|
3
2
+O(k
4
)
whichwecanusetondapproximationsafterN steps. Ifweconsiderjusttherstterm,weobtain
the Central Limit Theorem (CLT):
P
N
(x)
2
1
N
e
x
2
/2N
However, if we include the next term, we nd that
1
e
x
2
/2N
+
D

(x/

2N)
P
N
(x)
2N 6N

2
where D(x) is Dawsons integral, dened as
x
D(x) =e
x
2
e
t
2
dt.
0
DawsonsintegralsatisesD

(x)=4(x
2
1) +(12x8x
3
)D(x),anditcanbeshownthatD(x)
6x
4
asx. ThewidthofthecentralregionisdeterminedbythevalueofxwheretheDawson
correction has a comparable size to the CLT. We nd the width is

NlogN which is not much

wider than the minimum required by the CLT, and much less than N
3/4
when all the cumulants
are nite.
Simulations
Figure 1 shows a semi-log plot of the PDF. To compare the above predictions to simulation, we
must rst nd a way to simulate steps from this distribution. Unlike the Cauchy distribution, a
simpletransformationcannotbeusedinthiscase. LetX bearandomvariablewiththeone-sided
PDF
4
p
X
(x) =
(1+x
4
)
for x >0.
0.0001
0.001
0.01
0.1
1
10
-8 -6 -4 -2 0 2 4 6 8
P
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

x
Figure 1: Semi-log plot of the PDF under consideration.
If we consider the sequence of substitutions Z = X
2
, = tan
1
Z, =

then we nd that
has PDF
4

() = for 0< < /2.

tan
2
Althoughthisisacomplicatedfunctionitissmoothoverthegivendomain,andboundedbetween
0and4/

. Tosamplefromthisdistribution,wechooserandompairs(A, B)whereAisuniform
on [0,

/2], and B is uniform on [0,4/

]. If we keep sampling pairs until we nd B < p

(A),
and then let = A, then will follow the above PDF. Transforming back allows us to nd X.
MultiplyingX by-1withprobability1/2letsussamplefromthePDFwewishtostudy. Thiscan
be coded eciently in C using the function
double fatrand () {
double b;
while (usq()tan(b=usq()pi/2)>b);
b=sqrt(tan(b));return rand()%2==1?b:b;
}
where usq() returns the square of a number uniformly distributed on [0,1] and rand() returns a
random integer. Figure 2 shows four sample random walks from this distribution, and gures 3 to
6 show the accuracy of the Dawson correction for N =3, N =10, and N =100.
-20
-15
-10
-5
0
5
10
15
20
0 50 100 150 200 250
x

N
Figure 2: Four sample random walks using the PDF under consideration.
0
0.05
0.1
0.15
0.2
0.25
0.3
-10 -5 0 5 10
P
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

x
Simulation
CLT
Dawson
Figure 3: Comparisons between the Central Limit Theorem with and without the Dawson correc-
tion,comparedtosimulationsof1.610
10
walkswithN =3steps. Eveninthecentralregion,the
Dawson curve is a better t to the data.
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0.1
1
-20 -15 -10 -5 0 5 10 15 20
P
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

x
Simulation
CLT
Dawson
Figure 4: A semi-log plot of the N =3 data shows that the Dawson correction matches the tail of
the distribution to a high level of accuracy, although some small deviations are seen near x=5.
1e-08
1e-07
1e-06
1e-05
0.0001
0.001
0.01
0.1
1
0 5 10 15 20
P
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

x
Simulation
CLT
Dawson
Figure 5: Comparisons between the Central Limit Theorem with and without the Dawson correc-
tion,comparedtosimulationsof1.610
10
walkswithN =10steps. Smalldeviationsbetweenthe
Dawson curve and the simulated PDF are apparent.
1e-12
1e-10
1e-08
1e-06
0.0001
0.01
1
0 50 100 150 200
P
r
o
b
a
b
i
l
i
t
y

d
e
n
s
i
t
y

x
Simulation
CLT
Dawson
Figure 6: Comparisons between the Central Limit Theorem with and without the Dawson correc-
tion, compared to simulations of 1.610
9
walks with N =100 steps. The Dawson curve matches
the simulated data to a high level of accuracy.