You are on page 1of 13

Poularikas A. D.

Nonlinear Digital Filtering


The Handbook of Formulas and Tables for Signal Processing.
Ed. Alexander D. Poularikas
Boca Raton: CRC Press LLC,1999

1999 by CRC Press LLC


1999 by CRC Press LLC

41

Nonl i near Di gi tal

Fi l teri ng

41.1 Denitions

41.2 Mean Filter

41.3 Median Filter

41.4 Trimmed Mean Filters

41.5 L-Filters (Order Statistic Filters)

41.6 Weighted Median Filters

41.7 Ranked-Order Filter

41.8 Separable Two-Dimensional Median Filters

41.9 M-Filters

41.10 R-Filters

References

41.1 Denitions

41.1.1 One- and Two-Dimensional Signals

f(t)=continuous, f(n)=discrete; f(x,y)=continuous, f(n,m)=discrete

41.1.2 Filter Length

N=2k+1, N= odd number, range of indices:

i k,i k

+ 1,,

i

+

k



41.1.3 Data Length

x

={x(1),,x(N

d

)}

41.2 Mean Filter

41.2.1 Mean Filter

Example

x

={4,14,18,40,10}. For k=1, N=21+1=3 and y(1)=4,
y i
k
x i j
j k
k
( ) ( )
+
+

1
2 1

1999 by CRC Press LLC

Note:

1. Mean ltering is a linear operation 2. Mean lter attenuates noise

41.2.2 Mean Filter Algorithms

Input: data

x

={x(1),,x(N

d

)} Inputs: No of rows N

R



No of columns N

C

(image)
moving window w

N

, N=2k+1 moving window w

N

=N=2k+1
Output:

y

={y(1),,y(N)} Output: N

R


N

C

(image)
for

i

= 1 to N

d

for i=1 to N

R

set number at i for j=1 to N

C

Sum=0 set window w

N

at (i,j)
for every element x(m) of the Sum=0
image inside the window w

N

for every element x(m,n) of the
Sum=Sum+x(m) image inside w

N

end Sum=Sum+x(m,n)
y(i)=Sum/N end
end y(i,j)=Sum/N
end
end

41.3 Median Filter

41.3.1 Median Filter

y(i)=MED{x(i2), x(i1), x(i), x(i+1), x(i+2)} i=3,4,,18 for data

x

={x(1),x(2),,x(20)}

Example

Let

x

={1,5,3,2,10,2} and k=1 then y(1)=1,y(2)=MED{1,5,3}=MED{1,3,5}=3,
y( 3) =MED{5, 3, 2}=MED{2, 3, 5}=3,
y(4)=MED{3,2,10}=MED{2,3,10}=3,y(4)=MED{3,2,10}=MED{2,3,10}=3,y(5)=MED{2,10,2}=MED
{2,2,10}=2,y(6)=2. Hence,

y

={1,3,3,3,2,2}

Note:

a) the lter is nonlinear, b) attenuates noise to some degree c) does not attenuate noise well in
the case of noisy ramp signal d) noisy step signal is rounded at the step e) eliminates impulse noise
f) if impulse noise is close to an edge it may be removed but the edge moves toward the impulse (edge
jitter).

41.3.2 Median Filter Algorithms

Inputs: data

x

={x(1),,x(N

d

)} Inputs: No of rows N

R



No of columns N

C

moving window w

N

, N=2k+1 moving window w

N

, N=2k+1
Output: data

y

={y(1),,y(N

d

)} Output: N

R


N

C

(image)
for i=1 to N

d

for i=1 to N

R

place window w

N

at i for j=1 to N

C

store the data inside the window w

N

in place window w

N

at (i,j)

x


={

x


(1),,

x


(

N

)} store image values inside w

N

nd the median of

x


, MED{

x


}

x


={x


(1),,x


(N)}
y(i)=MED{

x


}; output nd median of

x


, MED{

x


}
end y(i,j)=MED{

x


}; output
end
end
y x j y y y
j
( ) ( ) ( ) , ( ) ( / )( ) , ( ) , ( ) 2
1
3
2
1
3
4 14 18 12 3 1 3 14 18 40 24 4 68 5 10
1
1
+ + + + +


1999 by CRC Press LLC

41.4 Trimmed Mean Filters

41.4.1 Motivation for Trimmed Mean Filters

The idea of trimmed lters is to reject the most probable outliers some of the very small and the
very large values.

41.4.2 (r,s)-Fold Trimmed Mean Filters

Short the samples in the window and omit r+s samples x

(1)

,,x

(r)

and x

(Ns+1)

,x

(Ns+2)


41.4.2.1 (r,s)-Fold Trimmed Mean Filters Algorithm

Inputs: data

x

={x(1),,x(N

d

)} Inputs: No of rows N

R



No of columns N

C

(image)
moving window w

N

, N=2k+1, r and s moving window w

N

, N=2k+1, integer numbers
integer numbers r,s
Output: y(i) Output: N

R


N

C

(image)
for i=1 to N

d

for i=1 to N

R

place window w

N

at i for j=1 to N

C

store the values inside the window into place window w

N

at (i,j)

x


={x


(1),x


(2),,x


(N)} store data inside window into
sort

x


to

x




=

x

={x(1),,x(N)}
Sum=0 sort

x

into x =
for m=r+1 to Ns Sum=0
Sum=Sum+ for m=r+s to Ns
end Sum=Sum+
y(i)=Sum/(Nrs); output end
end y(i,j)=Sum/(Nrs); output
end end
Note: x
(i)
x
(j)
for all i < j end
Example
x=(1,4,2,10,5,7,3,11,15), x
( )
={1,2,3,4,5,7,10,11,15}, N=5, r=1, s=2, y(1)=1, y(2)=2, y(3)=(2+3)/2,
y(4)=(3+4)/2 etc.
41.4.3 (r,s)-Fold Winsorized Filters
The r smallest values inside the window are replaced by x
(r+1)
and the s largest values inside the window
are replaced by x
(Ns)
Example (same as in 41.4.2.1)
y(1)=1, y(2)=2, y(3)=(1/5)(12+2+3+23)=(13/5), y(4)=(1/5)(13+3+4+24)=18/5 etc.

TrMean( x x x(N r,s)
d
x

+

{ ( ), ( ), , )};
( )
1 2
1
1
K
N r s
x
i
i r
N s

{ , , , }
( ) ( ) ( )


x x x
N 1 2
L

{ , , , }
( ) ( ) ( )
x x x
N 1 2
L
x
m ( )
x
m ( )

WinMean( ( r,s) x x x N
N
rx x sx
d r i
i r
N s
N s
+ +
+
+


{ ( ), , )}; (
( ) ( ) ( )
1
1
1
1
K
1999 by CRC Press LLC
41.4.4 -Trimmed Mean Filters
If r = s then = j/N, 0 j N/2 is an integer, and the formula is
41.4.5 -WinMean Filters (see 41.4.4)
Note: a) (0,0)-fold trimmed, (0,0)-fold Winsorized, 0-trimmed, and 0-Winsorized mean lters are
the same as the mean lter of the same window size; b) (k,k)-fold trimmed, (k,k)-fold Winsorized, 0.5-
trimmed and 0.5-Winsorized mean lters are the same as the median lter of the same window size
Example
x=(4,14,18,40,10), N=5, x
( )
=(4,10,14,18,40) a) (2,1)-fold trimmed mean lter (14+18)/2=16, b) (1,2)-
Winsorized mean lter (210+314)/5=12.4, c) 0.2-trimmed mean (14+18+40)/3=14, d) 0.2-Winsorized
mean lter (210+14+218)/5=14
41.4.6 Modied Trimmed Mean Filters
A real valued constant q is rst xed and then all the samples in the range [x
(k+1)
q, x
(k+1)
+ q] are
averaged, and the results constitute the output
Example
x={4,14,18,40,10}, x
( )
={4,10,14,18,40}, a) for 0 q < 4 med is 14, b) for 4 q < 10, (10 + 14 +
18)/3=14, c) for 10 q < 26, (4 + 10 + 14 + 18)/4=11.5, d) mean (4 + 10 + 14 + 18 + 40)/5=17.2
41.4.6.1 Modied Trimmed Mean Algorithm
Inputs: data x={x(1),,x(N
d
)} Inputs: No of rows N
R
No of columns N
C
(image)
moving window w
N
, N=2k+1 moving window w
N
, N=2k+1
positive real number q positive real number q
Output: y={y(1),,(N
d
)} Output: N
R
N
C
(image)
for i=1 to N
d
for i=1 to N
R
place window w
N
at i for j=1 to N
C
store the values in the window in place window w
N
at (i,j)
x={x(1),,x(N)} store the values inside the window w
N
in
nd the median x
(k+1)
of x x={x(1),,x(N)}
Sum1=0 nd the median x
(k+1)
of x
Sum2=0 Sum1=0
for m=1 to N Sum2=0
if abs(x(m)x
(k+1)
)q then for m=1 to N

TrMean( ( ) x { ( ), , )};
( )
x x N
N N
x
d i
i N
N N
1
1
2
1
K

+ +

_
,

+
+


WinMean( ( ) x { ( ), , )};
( ) ( ) ( )
x x N
N
Nx x Nx
d N i
i N
N N
N N
1
1
1
1
K

MTM x x N
a x i
a
a
x i x k q
d
i
i
N
i
i
N i
( ( )
if
otherwise
x
+

'

{ ( ), , )}
( )
,
( ) ( )
1
1 1
0
1
1
K
1999 by CRC Press LLC
Sum1=Sum1+x(m) if abs(x(m)x
(k+1)
)q then
Sum2=Sum2+1 Sum1=Sum1+x(m)
end Sum2=Sum2+1
y(i)=Sum1/Sum2; output end
end y(i,j)=Sum1/Sum2; output
end end
end
41.4.7 Double Window Modied Trimmed Mean Filters
The idea behind this lter is to suppress additive Gaussian noise with the large window and preserve
the details by rejecting pixels that are far away from the median of the smaller window.
41.4.7.1 Double Window Modied Trimmed Mean Filters Algorithm
Inputs: data x={x(1),,x(N
d
)} Inputs: No of rows N
R
No of columns N
C
(image)
moving window w
N
, N=2k+1 moving window w
N
, N=2k+1
moving window w
1
, M>N moving window w
1
, M>N
positive real number q positive real number q
Output: y={y(1),,y(N
d
)} Output: N
R
N
C
(image)
for i=1 to N
d
for i=1 to N
R
place window w
N
and w
1
at i for j=1 to N
C
store values inside w
N
in place windows w
N
and w
1
at (i,j)
x={x(1),,x(N)} store the values inside w
N
in x={x(1),,x(N)}
nd the median x
(k+1)
of x nd the median x
(k+1)
of x
sum1=0, sum2=0 sum1=0, sum2=0
store values inside w
1
in store values inside w
1
in
z={x(1),,x(M)} z={z(1),,z(M)}
for m=1 to M for m=1 to M
if abs(x(m)x
(k+1)
)q then if abs(x(m)x
(k+1)
)q then
sum1=sum1+x(m) sum1=sum1+x(m)
sum2=sum2+1 sum2=sum2+1
end end
y(i)=sum1/sum2 y(i,j)=sum1/sum2
end end
end end
41.4.8 K-Nearest Neighbor Filter
The output is given by the mean of K, 1KN, samples whose values are closest to the value of the
central value x
*
inside the lter window.
Example
x={4,14,18,40,10}. Center sample is x
*
=18 a) for K=1, 18 (itself) is the closest sample to 18 and output
is 18/1=18 b) for K=2, 18 and 14 must be used and the output is (18+14)/2=16 c) for K=3, 18, 14 and
10 are closest and output is (18+14+10)/3=14.
41.4.8.1 K-Nearest Neighbor Filter Algorithm
Inputs: data x={x(1),,x(N
d
)} Inputs: No of rows N
R
No of columns N
C
(image)
moving window w
N
, N=2k+1 moving window w
N
, N=2k+1
integer number K, 1KN integer number K, 1KN
1999 by CRC Press LLC
Output: y={y(1),,y(N
d
)} Output: N
R
N
C
(image)
for i=1 to N
d
for i=1 to N
R
place window w
N
at i for j=1 to N
C
store values inside the window w
N
place windows w
N
at (i,j)
and their differences from x
*
in store values inside w
N
and their differences
z = from x
*
(pixel (i,j)) in
sort z with respect to and store in z =
t = sort z with respect to abs(x(i)x
*
), and store
sum=0 the results in
for m=1 to K t =
sum=sum+x
(m)
sum=0
end for m=1 to K
y(i)=sum/K sum=sum+x
(m)
end end
end y(i,j)=sum/K
end
end
41.4.8.2 Modied K-Nearest Neighbor Filter
If we change the rule for a
i
in 41.4.6 to
If q is chosen to be twice the standard deviation of the noise, the lter is known as the sigma lter.
41.5 L-Filters (Order Statistic Filters)
41.5.1 Purpose of the Estimator
L-lters are running estimators making a compromise between a pure nonlinear operation (ordering)
and pure operation (weighting).
41.5.2 L-Filters
41.5.3 L-Filters Algorithms
Inputs: data x={x(1),,x(N
d
)} Inputs: No of rows N
R
No of columns N
C
(image)
moving window w
N
, N=2k+1 moving window w
N
, N=2k+1
weight vector a={a
1
,,a
N
} weight vector a={a
1
,,a
N
}
Output: y={y(1),,y(N
d
)} Output: N
R
N
C
(image)
for i=1 to N
d
for i=1 to N
R

{( ( ), ( ) ), ,( ( ), ( ) )}
* *
x x x x N x N x 1 1 L
x i x ( )
*


{( ( ), ( ) ), ,( ( ), ( ) )}
* *
x x x x N x N x 1 1 L

{( , ), ,( , )}
( ) ( )
*
( ) ( )
*
x x x x x x
N N 1 1
L

{( , ), ,( , )}
( ) ( )
*
( ) ( )
*
x x x x x x
N N 1 1
L
a
x i x q
i

'

1
0
if
otherwise
( )
*

L x x N a x a
i i
i
N
i
i
N
( ( ), , ); ) ,
( )
1 1
1 1
K ( a


1999 by CRC Press LLC
place window w
N
at i for j=1 to N
C
store the data inside the window place window w
N
at (i,j)
w
N
in z={x(1),,x(N)} store values inside the window w
N
in
sort z and store in the vector q z={x(1),,x(N)}
q={x
(1)
,,x
(N)
} sort z and store in q={x
(1)
,,x
(N)
}
sum=0 sum=0
for m=1 to N for m=1 to N
sum=sum+a
m
x
(m)
sum=sum+a
m
x
(m)
end end
y(i)=sum; output y(i,j)=sum; output
end end
end end
41.6 Weighted Median Filters
41.6.1 Purpose
With this lter we may emphasize the samples that for some reason are supposed to be more reliable,
e.g., center sample x
*
, and the emphasis is obtained by weighing them more heavily.
41.6.2 Repetition or Duplication Operator &
3&x=x,x,x
41.6.3 Multiset
A multiset is a collection of objects, where the repetition of objects is permitted, e.g.,
{3&2,1,2&4}={2,2,2,1,4,4}
41.6.4 Weighted Median Filters
WeightMed{x(1),,x(N);a}=Med{a
1
&x(1),,a
N
&x(N)}
41.6.5 Weighted Median Algorithm
Inputs: data x={x(1),,x(N
d
)} Inputs: No of rows N
R
No of columns N
C
(image)
moving window w
N
, N=2k+1 moving window w
N
, N=2k+1
weight vector a={a
1
,,a
N
} weight vector a={a
1
,,a
N
}
Output: y={y(1),,y(N
d
)} Output: N
R
N
C
(image)
halfsum= halfsum=
for i=1 to N
d
for i=1 to N
R
place window w
N
at i for j=1 to N
C
store values inside the window and weight place window w
N
and corresponding weighs in
in z={(x
1
,a
1
),,(x
N
,a
N
)} z={(x
1
,a
1
),,(x
N
,a
N
)}
sort z with respect to x
i
s and store in q sort z with respect to x
i
s and store in
q={(x
(1)
a
(1)
),,(x
(N)
,a
(N)
)} q={(x
(1)
,a
(1)
),,(x
(N)
,a
(N)
)}
sum=0, m=1 sum=0, m=1
repeat repeat
sum=sum+a
(m)
sum=sum+a
(m)
m=m+1 m=m+1
a
i
i
N

_
,

1
2 a
i
i
N

_
,

1
2
1999 by CRC Press LLC
until sumhalfsum until sumhalfsum
y(i)=x
(m1)
y(i,j)=x
(m1)
end end
end end
Example
x={4,14,18,40,10}, a={0.05,0.1,0.15,0.1,0.05}. After ordering
z={(4,0.05),(10,0.05),(14,0.1),(18,0.15),(40,0.1)}. Halfsum=(0.05+0.05+0.1+0.15+0.1)/2=0.225. But
(0.05+0.05+0.1+0.15)=0.35>0.225 and the lter output is 18.
41.7 Ranked-Order Filter
41.7.1 Purpose
These lters can be used in situations where the noise distribution is not symmetric, e.g., where there
are more positive than negative impulses.
41.7.2 Ranked-Order Filter
RO(x(1),,x(N);r)=x
(r)
, N=2k+1
Note: When r=k+1 we obtain the median lter. If r<k+1 the lter introduces bias toward small values
and toward large values if r>k+1.
Example
If x={4,14,18,40,10} and the ordered x-ordered={4,10,14,18,40}. Here N=5=2k+1 and k=2. The r
th
ranked-order lter gives the outputs: 2 for r=1, 10 for r=2, 14 for r=3, 18 for r=4, and 40 for r=5.
41.7.3 Ranked-Order Filter Algorithms
Inputs: data x={x(1),,x(N
d
)} Inputs: No. of rows N
R
No. of columns N
C
(image)
moving window w
N
, N=2k+1 moving window w
N
, N=2k+1
rank r rank r
Output: y={y(1),,y(N
d
)} Output: N
R
N
C
(image)
for i=1 to N
d
for i=1 to N
R
place window w
N
at i for j=1 to N
C
store the values in the window w
N
place window w
N
at (i,j)
in z={x(1),,x(N)} store values inside the window w
N
in
sort z to q={x
(1)
,,x
(N)
} z={x(1),,x(N)}
y(i)=x
(r)
; output sort z to q={x
(1)
,,x
(N)
}}
end y(i,j)=x
(r)
end end
end
41.8 Separable Two-Dimensional Median Filters
41.8.1 Purpose
The lter has the same properties as the median lter but is much faster. It is accomplished by using
two successive one-dimensional median lters of window size M=2l+1 (MM=N), the rst along the
rows and the second one along the columns of the so-obtained image.
1999 by CRC Press LLC
41.8.2 Separable Two-Dimensional Median Filters
Med{x(M,1),,x(M,M)}}
41.8.3 Algorithm
Inputs: No. of rows N
R
time No. of columns N
C
; (image)
Moving horizontal window w
1,M
, M=2l+1
Moving vertical window w
2,M
, M=2l+1
Output: N
R
N
C
; (image)
Median lter the image by using horizontal moving window w
1,M
for each row
Store the resulting values as they are found in a column vector
Median lter the column vector using vertical moving window w
2,M
41.9 M-Filters
41.9.1 Purpose
M-estimators are generalizations of maximum likelihood estimators
41.9.2 M-Filters
M(x(1),,x(N
d
); ) =arg
41.9.3 M-Filter by the Partial Derivative (with respect to ) of
The estimate M(x(1),,x(N
d
)) satises the equation
41.9.4 Function in Use
1. Skipped median,
2. Andrews sine,
3. Tukeys biweight,
4. Standard M-lters,

SepMed Med{Med{
x x x M
x x x M
x M x M x M M
x x M
( , ) ( , ) ( , )
( , ) ( , ) ( , )
( , ) ( , ) ( , )
( , ), , ( , )},
1 1 1 2 1
2 1 2 2 2
1 2
1 1 1
L
L
M M M M
L
L L

1
]
1
1
1
1

min ( ( ) )

i
N
x i
1
( ( ) )
i
N
x i


1
0

med r
x
x x r
r x
( )
( )
sin( )

<

'

0
0

sin( )
( )
sin( )
a
x
xa x a
a x

<

'

0
0

bi r
x
x r x x r
r x
( )
( )
( )

<

'

2 2 2
0
0

st
x
ap x p
ax x p
ap x p
( )
>

<

'

1999 by CRC Press LLC


Note: The function is not strictly monotone and 41.9.3 might not have a unique solution. This can
be handled in several ways:
1. Find the global minimum of
2. Select the solution of nearest to the sample median
3. Use Newtons method, starting from the sample median
4. Use one-step method, starting from the sample median
Example
x={2,7,9,20,5} a) Consider standard M-lter with p=5 and a=1/5. b) Consider Tukeys beweight M-lter
with r=5. If we plot the function
we nd that the standard lter has a zero at =7 which is the output. However, we have various candidates
for the output of the Tukeys biweight M-lter. Which one is found depends on the method used.
41.9.5 M-Filter Algorithm
Inputs: No. of rows N
R
No. of columns N
C
; (image)
moving window w
N
, N=2k+1
function (x)
constant real number c0
number of iterations N
I
Output: N
R
N
C
; (image)
for i=1 to N
R
for i=1 to N
C
place window w
N
at (i,j)
store values inside the window w
N
in
z={x(1),x(2),,x(N)}
nd the median (0) of the value inside the window w
N
for k=1 to N
I
sum1=0, sum2=0
for l=1 to N
if x(l) (k 1)
sum1=sum1 + x(l)((x(l) (k 1)))/(x(l) (k 1))
sum2=sum2 + ((x(l) (k 1)))/(x(l) (k 1))
else
sum1=sum1+cx(l)
sum2=sum2+c
end
(k)=sum1/sum2
end
y(i,j)=(k)
end
end
( ( ) )
i
N
x i


1
( ( ) )
i
N
x i


1
0
( ( ) ) ( ) ( ) ( ) ( ) ( )
i
N
x i

+ + + +
1
2 7 9 20 5 0 versus
1999 by CRC Press LLC
41.10 R-Filters
41.10.1 Purpose
R-estimators are known to be robust.
41.10.2 Wilcoxon Filter (R-Filter) with Walsh Averages
41.10.3 Wilcoxon Filter (R-Filter) with Ordered Walsh Averages
Example
x={4,14,18,40,10}. The averages are Med{(4+4)/2, (4+14)/2, (4+18)/2, (4+40)/2, (4+10)/2, (14+14)/2,
(14+18)/2, (14+40)/2, (14+10)/2, (18+18)/2, (18+40)/2, (18+10)/2, (40+40)/2, (40+10)/2, (10+10)/2}=14
41.10.4 Wilcoxon Filter Algorithms
Inputs: x={x(1),,x(N
d
)}; data Inputs: No. of rows N
R
No. of columns N
C
(image)
moving window w
N
, N=2k+1 moving window w
N
, N=2k+1
Output: y={y(1),,y(N
d
)} Output: N
R
N
C
(image)
for i=1 to N
d
for i=1 to N
R
place window w
N
at i for j=1 to N
C
store values inside the window in place the window w
N
at (i,j)
z={x(1),,x(N)} store values inside the window in
m=1 z={x(1),,x(N)}
for s=1 to N m=1
for t=s to N for s=1 to N
Ave(m)=(x(s)+x(t))/2 for t=s to N
m=m+1 Ave(m)=(x(s)+x(t))/2
end m=m+1
end end
Med{Ave(1),Ave(2),,Ave(m1)} end
y(i)=Med Med{Ave(1),,Ave(m1)}
end y(i,j)=Med
end end
end
41.10.5 Winsorized Wilcoxon Filters

Wil ( ( Med x x x N
x i x j
i j N
d
( ), ( ), , ))
( ) ( )
: 1 2
2
1 L
+

'


Wil ( ( Med x x x N
x x
i j N
d
i j
( ), ( ), , )) :
( ) ( )
1 2
2
1 L
+

'


Wins Wil ( ( Med
+
<

'

x x N r
x x
j i r i j N r N
d
i j
( ), , ); ) : , ,
( ) ( )
1
2
1 1 K
1999 by CRC Press LLC
Example
x={4,14,18,40,10}, r=2. Hence x
( )
={4,10,14,18,40} and the average values are: (4+4)/2, (4+10)/2,
(10+10)/ 2, (10+14)/ 2, (14+14)/ 2, (14+18)/ 2, (18+18)/ 2, (18+40)/ 2, (40+40)/ 2, or
Med{4,7,10,12,14,16,18,28,40)=14.
41.10.6 HodgeLehmann D-Filter (related to Wilcoxon)
References
Astola, J. and P. Kuosmanen, Fundamentals of Nonlinear Digital Filtering, CRC Press, Boca Raton,
FL, 1997.
Huber P. J., Robust Statistics, John Wiley & Sons, New York, NY, 1981.
Pitas, I. and A. N. Venetsanopoulos, Nonlinear Digital Filters: Principles and Applications, Kluwer
Academic Publishers, Boston, MA, 1990.

Ho Leh( ( Med
+
+ +

'

+
x x N
x x
i k N k
d
i N i
( ), , )) : ,
( ) ( )
1
2
1 1 2 1
1
L

You might also like