1

M3S3/M4S3 STATISTICAL THEORY II

WORKED EXAMPLE: EFFICIENT INFERENCE FOR THE CAUCHY DISTRIBUTION
Suppose that X
1
, . . . , X
n
are i.i.d random variables having a Cauchy distribution with pdf
f
X|
(x|) =
1

1
1 + (x )
2
x R
for parameter R.
(i) Show that the Fisher Information for is I() = 1/2.
(ii) Find the an asymptotic approximation to the distribution of the sample median, M
n
, that is the
quantile corresponding to p = 1/2, and hence comment on the eciency of M
n
in estimating .
(iii) Using the one-step estimation approach, construct an ecient estimator of , stating its asymp-
totic variance.
(iv) Find the an asymptotic approximation to the distribution of the length, L
n
, of the central 95 %
sample interval, that is
L
n
= X
(k
2
)
X
(k
1
)
where k
1
= np
1
, k
2
= np
2
for p
1
= 0.025 and p
2
= 0.975, x denotes the smallest integer not
less than x, and X
(k)
denotes the kth order statistic.
SOLUTION (i) The Fisher information is computed in the usual way.
l() = log log(1 + (x )
2
)

l() =
2(x )
1 + (x )
2

l() = 2
1 (x )
2
(1 + (x )
2
)
2
Thus
I() = E
f
X|
[

l()] = 2
_

1 (x )
2
(1 + (x )
2
)
2
1

1
1 + (x )
2
dx
=
2

1 (x )
2
(1 + (x )
2
)
3
dx =
2

1 x
2
(1 + x
2
)
3
dx
=
2

_
/2
/2
1 tan
2
t
(1 + tan
2
t)
3
dx
dt
dt x = tant
=
2

_
/2
/2
1 tan
2
t
(1 + tan
2
t)
3
sec
2
t dt =
2

_
/2
/2
1 tan
2
t
(sec
2
t)
3
sec
2
t dt
=
2

_
/2
/2
1 tan
2
t
(sec
2
t)
2
t dt =
2

_
/2
/2
_
1
sin
2
t
cos
2
t
_
cos
4
t dt
=
2

_
/2
/2
_
cos
4
t sin
2
t cos
2
t
_
dt
=
8

_
/2
0
cos
4
t dt
4

_
/2
0
cos
2
t dt
2
By a standard result
C(k) =
_
/2
0
cos
k
t dt =
_
/2
0
cos t cos
(k1)
t dt
=
_
sint cos
(k1)
t
_
/2
0
+ (k 1)
_
/2
0
sin
2
t cos
(k2)
t dt
= 0 + (k 1)
_
/2
0
(1 cos
2
t) cos
(k2)
t dt
= (k 1)
_
/2
0
cos
(k2)
t dt (k 1)
_
/2
0
cos
k
t dt
= (k 1)C(k 2) (k 1)C(k)
and hence
C(k) =
_
k 1
k
_
C(k 2)
Thus, as C(0) = /2, C(2) = /4 and C(4) = 3/16, so
I() =
8

3
16

4

4
=
1
2
as required.
(ii) M
n
is the p = 1/2 quantile, and by the standard result for sample quantiles

n(M
n
)
L
N(0,
2

)
as the (true) distribution median is (as the pdf is symmetric around ), and where

=
p(1 p)
{f
X|
()}
2
=
1/4
{1/}
2
=

2
4
2.467
Thus the estimator M
n
is inecient, as the lowest possible variance given by I() = 1/2 is I()
1
= 2.
Note that an ecient estimator can be found as the solution to the likelihood equations

l
n
() = 0, or
equivalently
n

i=1
(x
i
)
1 + (x
i
)
2
= 0
but this is not straightforward. Note also that the other natural estimator, the sample mean, is not useful,
as it has intractable asymptotic behaviour (the expectation of the Cauchy does not exist).
(iii) The one-step approach to ecient estimation computes one of the following two estimators

(1)
=

n

_

l
n
(

n
)
_
1

l
n
(

n
) (N)
and

n
+
_
I(

n
)
_
1
1
n

l
n
(

n
) (S)
based on the consistent estimators

n
. The one-step estimators are asymptotically equivalent to the ecient
estimator, and thus have asymptotic variance 2.
3
Now M
n
is a suitable consistent estimator of , and

l
n
() = 2
n

i=1
(x
i
)
1 + (x
i
)
2

l
n
() = 2
n

i=1
1 (x
i
)
2
(1 + (x
i
)
2
)
2
so therefore the one-step estimators are

(1)
= M
n
+ 4
_
n

i=1
1 (x
i
M
n
)
2
(1 + (x
i
M
n
)
2
)
2
_
1
n

i=1
(x
i
M
n
)
1 + (x
i
M
n
)
2
and

= M
n
+
4
n
n

i=1
(x
i
M
n
)
1 + (x
i
M
n
)
2
(iv) Using the standard result from lectures, we have that

n
__
X
(k
1
)
X
(k
2
)
_

_
x
p
1
x
p
2
__
L
N
_
_
_
_
_
_
_
0,
_

_
p
1
(1 p
1
)
_
f
X|
(x
p
1
)
_
2
p
1
(1 p
2
)
f
X|
(x
p
1
) f
X|
(x
p
2
)
p
1
(1 p
2
)
f
X|
(x
p
1
) f
X|
(x
p
2
)
p
2
(1 p
2
)
_
f
X|
(x
p
2
)
_
2
_

_
_
_
_
_
_
_
_
N(0, ),
say, where here, p
1
= 0.025, p
2
= 0.975, and x
p
1
and x
p
2
are the corresponding true distribution quantiles.
Now, here
F
X|
(x) =
_
x

f
X|
(t) dt =
_
x

1
1 + (t )
2
dt =
1

[arctant]
x

=
1

(arctan(x ) + /2)
so
1

(arctan(x
p
1
) + /2) = 0.025 and
1

(arctan(x
p
2
) + /2) = 0.975
and hence
x
p
1
= + tan((0.025) /2) = 12.706
x
p
2
= + tan((0.975) /2) = + 12.706.
Hence
=
2
_
0.025 (1 0.025) (1 + 12.706
2
)
2
0.025 (1 0.975) (1 + 12.706
2
)(1 + 12.706
2
)
0.025 (1 0.975) (1 + 12.706
2
)(1 + 12.706
2
) 0.975 (1 0.975) (1 + 12.706
2
)
2
_
=
2
(1 + 12.706
2
)
2
_
0.025 (1 0.025) 0.025 (1 0.975)
0.025 (1 0.975) 0.975 (1 0.975)
_
=
_
6348.501 162.782
162.782 6348.501
_
Finally, as L
n
= X
k
2
X
k
1
, using the Delta method with function g(t
1
, t
2
) = t
2
t
1
, we have

n(L
n
)
L
N(0, g(x
p
1
, x
p
2
) g
T
)
where is the true length, that is 2 12.706 = 25.412, and
g(t) =
_
g(t)
t
1
g(t)
t
2
_
= [1 1] .
Thus
g(x
p
1
, x
p
2
) g
T
= 12371.44.
and hence

n(L
n
25.412)
L
N(0, 12371.44).