2
x
k
h
k
n
k
W
e
k
y
k
b
Fig. 1. System block diagram
B. Signal Model
We consider the case of a linear, time invariant, dispersive,and noisy communication channel. The complex-valued equiv-alent baseband signal model is used. Assuming an oversam-pling factor of
l
, the received samples have the standard form
y
k
=
ν
m
=0
h
m
x
k
−
m
+
n
k
(2)where
h
m
is the CIR whose memory is denoted by
ν
and
n
k
is the additive noise. The quantities
{
y
,
h
m
,
n
}
are
l
×
1
col-umn vectors corresponding to the
l
time (fractionally-spaced)samples per symbol in the assumed temporally oversampledchannel model. Furthermore,
x
k
−
m
is the transmitted symbolat time
(
k
−
m
)
.Grouping
{
y
k
}
over a block of
N
f
symbol periods, theinput-output relation in (2) can be expressed in the followingmatrix notation
⎡⎢⎢⎢⎣
y
k
y
k
−
1
...
y
k
−
N
f
+1
⎤⎥⎥⎥⎦
=
⎡⎢⎢⎢⎣
h
0
h
1
...
h
ν
0
...
00
h
0
h
1
...
h
ν
0
...
.........
0
...
0
h
0
h
1
...
h
ν
⎤⎥⎥⎥⎦
.
⎡⎢⎢⎢⎣
x
k
x
k
−
1
...
x
k
−
N
f
−
ν
+1
⎤⎥⎥⎥⎦
+
⎡⎢⎢⎢⎣
n
k
n
k
−
1
...
n
k
−
N
f
+1
⎤⎥⎥⎥⎦
(3)or more compactly
y
k
:
k
−
N
f
+1
=
Hx
k
:
k
−
N
f
−
ν
+1
+
n
k
:
k
−
N
f
+1
(4)The
(
N
f
+
ν
)
×
(
N
f
+
ν
)
input correlation matrix is definedby
R
xx
≡
E
x
k
:
k
−
N
f
−
ν
+1
x
H k
:
k
−
N
f
−
ν
+1
and the
(
lN
f
)
×
(
lN
f
)
noise correlation matrix is defined by
R
nn
≡
E
n
k
:
k
−
N
f
+1
n
H k
:
k
−
N
f
+1
.
Both
R
xx
and
R
nn
are assumed to be positive-definite (non-singular) correlation matrices. Furthermore, the output-inputcross-correlation and the output auto-correlation are, respec-tively, given by
R
yx
≡
E
y
k
:
k
−
N
f
+1
x
H k
:
k
−
N
f
−
ν
+1
=
HR
xx
(5)
R
yy
≡
E
y
k
:
k
−
N
f
+1
y
H k
:
k
−
N
f
+1
=
HR
xx
H
H
+
R
nn
(6)III. S
PARSE
C
HANNEL
S
HORTENING
A. Algorithm Derivation
The system block diagram is shown in Fig. 1 where ourgoal is to design the length-(
lN
f
) fractionally-spaced CSE,
W
, and the length-(
N
f
+
ν
) TIR,
b
, which minimize the meansquare of the error signal,
e
k
. In other words, the CSE isto be designed such that the overall impulse response of thechannel,
h
, and the CSE,
W
, best approximates a TIR withfew, namely (
N
b
+ 1
), taps. Then, the ML or MAP detectorsare designed based on the new short TIR. Although the lengthof the TIR vector
b
is (
N
f
+
ν
), only (
N
b
+ 1
) of its tapsare designed to be nonzero. The choice of
N
b
represents aperformance-complexity tradeoff. In [5], the nonzero (
N
b
+1
)taps were chosen to be
contiguous
and their location withinthe (
N
f
+
ν
)-span of
b
was optimized. In this paper, we relaxthe contiguousness constraint and allow the nonzero (
N
b
+1
)taps to be anywhere within the (
N
f
+
ν
)-span. Following thisapproach, we show that better performance can be achievedwithout increasing complexity. From Fig. 1, the error sequence
e
k
is given by [5]
e
k
=
W
H
y
k
:
k
−
N
f
+1
−
b
H
x
k
:
k
−
N
f
−
ν
+1
(7)The MSE is formulated asMSE
ξ
=
E
|
e
k
|
2
=
W
H
R
yy
W
−
W
H
R
yx
b
−
b
H
R
H yx
W
+
b
H
R
xx
b
(8)where
E
[
.
]
and
|
.
|
denote the statistical expectation and theabsolute value, respectively. Minimizing the MSE over
W
bydifferentiating it with respect to (w.r.t)
W
and equating theresult to zero, we get
W
opt
=
R
−
1
yy
R
yx
b
(9)Substituting
W
opt
for
W
in (8), we get
ξ
=
b
H
R
⊥
x/y
b
(10)where
R
⊥
x/y
R
xx
−
R
H yx
R
−
1
yy
R
yx
(11)Minimizing
ξ
over
b
yields the trivial solution
b
= 0
. To avoidthis, we perform the minimization subject to the UTC whereone of the (
N
b
+1
) taps is constrained to be unity. The indexof the unit tap is denoted by
i
(0
≤
i
≤
N
f
+
ν
−
1)
. Now, weaim at designing
b
which minimizes
ξ
such that its nonzerotaps are not constrained to be contiguous unlike [5]. Towardsthis end, we define the Cholesky factorization [16] of
R
⊥
x/y
as
R
⊥
x/y
U
H
U
where
U
is an upper-triangular matrix andrewrite (10) as follows
ξ
=
b
H
U
H
Ub
=
Ub
22
=
¯
U
¯
b
+
u
i
22
(12)where
¯
U
is formed by all the columns of
U
except for the
i
th
column,
u
i
is the
i
th
column of
U
, and
¯
b
is formed by all theelements of
b
except for the
i
th
unity element. Observe thatthe length-(
N
f
+
ν
−
1
) vector
¯
b
contains only
N
b
nonzerotaps whose locations and values need to be determined suchthat
ξ
is minimized. To achieve this goal, we formulate thefollowing convex optimization problem
min
¯
b
¯
b
1
subject to
¯
U
¯
b
+
u
i
2
≤
ch
(13)