Escolar Documentos
Profissional Documentos
Cultura Documentos
E
s
N
f
N
f
1
k=0
sp(t kN
f
) (1)
where E
s
is the symbol energy and is related to the average
bit energy E
b
as E
s
= 2E
b
, p(t) represents the ultra-short
pulse with unit energy, i.e.,
_
p
2
(t)dt = 1, and the width
T
p
T
f
.
After passing through the multipath channel, denoted as
h(t), the received signal is given by
r(t) =
E
s
N
f
N
f
1
k=0
sg(t kN
f
) + n(t) (2)
where g(t) p(t) h(t) represents the normalized composite
channel, n(t) denotes the ltered Gaussian noise with two-
sided power spectrum N
0
/2 and bandwidth B. In (1) and (2),
we assume that the frame duration T
f
is chosen such that
theres no inter-frame interference and inter-symbol interfer-
ence. Therefore, we can consider each symbols transmission
and reception separately.
1089-7798/13$31.00 c 2013 IEEE
w
w
w
.
t
a
k
e
o
f
f
p
r
o
j
e
c
t
s
.
c
o
m
For Further Details-A Vinay 9030333433,0877-2261612 0
354 IEEE COMMUNICATIONS LETTERS, VOL. 17, NO. 2, FEBRUARY 2013
A WED receiver divides T
f
into M equi-length sub-
intervals, each of duration T
M
= T
f
/M. Squaring r(t) and
integrating it over each sub-interval, the receiver obtains M-
tuple outputs as
y
m
=
Nf 1
k=0
_
kTf +mTM
kTf +(m1)TM
r
2
(t)dt, m = 1, 2, ..., M (3)
By invoking the sampling theorem [7] and Gaussian ap-
proximation, y
m
s can be shown to be statistically independent
and y
m
can be modeled as a Gaussian variable with mean and
variance as shown below [4][6], [10]
m
= N
f
N
0
BT
M
+ sE
m
2
m
= N
f
N
2
0
BT
M
+ 2sN
0
E
m
(4)
where E
m
= E
s
_
mTM
(m1)TM
g
2
(t)dt. For compactness, we
collect y
m
s into a vector y = [y
1
, ..., y
M
]
T
, whose mean
vector and diagonal covariance matrix, respectively, are
m = [
1
,
2
, ...,
M
]
T
C = diag{
2
1
,
2
2
, ...,
2
M
} (5)
III. OPTIMAL BER-BALANCED WED
The key in WED is the determination of weighting vector
w R
M
and threshold . Once w, are given, the transmitted
symbol s can be recovered as follows
z = w
T
y, s =
_
0, z
1, z >
(6)
Based on (4) and (5), the decision statistic is Gaus-
sian distributed as z N(w
T
m
0
, w
T
C
0
w) or z
N(w
T
m
1
, w
T
C
1
w), depending on s = 0 or s = 1, where
m
0
, C
0
(m
1
, C
1
) are, respectively, the mean vector and the
covariance matrix in (5) corresponding to s = 0 (s = 1).
Accordingly, the BER of the WED using (6) is derived as:
p
b
=
1
2
p
b,0
+
1
2
p
b,1
=
1
2
Q
_
m
T
0
w
_
w
T
C
0
w
_
+
1
2
Q
_
m
T
1
w
_
w
T
C
1
w
_ (7)
where p
b,0
, p
b,1
are the BER expressions for data bits 0 and
1, respectively, Q() is the complementary error function. To
achieve the balanced BER, we set p
b,0
=p
b,1
and thus have
=
m
T
1
w
_
w
T
C
0
w+m
T
0
w
_
w
T
C
1
w
_
w
T
C
0
w+
_
w
T
C
1
w
(8)
p
b
= Q
_
h
T
w
_
w
T
C
0
w+
_
w
T
C
1
w
. .
x
_
(9)
where h m
1
m
0
= [E
1
, E
2
, ..., E
M
]
T
, such channel-
related information can generally be obtained through training
symbols.
Now we are in the position to determine the optimal
weighting vector (denoted as w
o
) which achieves the min-
imal p
b
, denoted as p
b,o
. Since Q function is monotonically
decreasing, minimizing p
b
in (9) is equivalent to the following
minimization problem:
min
w
f(w) =
_
w
T
C
0
w+
_
w
T
C
1
w
h
T
w
(P1) (10)
Note that w and its scaling w with being an arbitrary
positive scalar, lead to the same f(w). Therefore, P1 can be
transformed into
min
w
g(w) =
_
w
T
C
0
w+
_
w
T
C
1
w (P2)
subject to h
T
w = 1
(11)
Since square root function is increasing in the argument, and
its argument is convex with respect to w, so the function g(w)
is convex. As a result, (P2) is a convex minimization problem
and can be easily solved for w
o
[8].
To bypass the computation burden required by (11), can we
nd a simple but near-optimal weighting vector? To answer
this question, we want to derive lower and upper bounds on
p
b,o
, from which we will gain some hints.
Lemma 1: The minimal BER can be bounded as:
Q(
1
2
C
1/2
0
h
2
) p
b,o
Q(
1
2
C
1/2
1
h
2
) .
Proof: For a diagonal matrix, the diagonal elements are
its eigenvalues. Exploiting this fact and (4), (5), we have
min
(C
1
) = min
m
(N
f
N
0
BT
M
+ E
m
)
max
(C
0
) = N
f
N
0
BT
M
(12)
where
max
(),
min
() represent the maximum and minimum
eigenvalues of a matrix, respectively. Therefore, we can readily
obtain w
T
C
1
w w
T
C
0
w, which leads to
h
T
w
2
_
w
T
C
1
w
x
h
T
w
2
_
w
T
C
0
w
(13)
Let us rst consider the left-hand inequality
max
w
x max
w
h
T
w
2
_
w
T
C
1
w
=
1
2
1/2
max
(C
T/2
1
hh
T
C
1/2
1
. .
A
)
=
1
2
C
1/2
1
h
2
(14)
where the inequality in the rst step is obvious, in the
second step we use the Rayleigh Ritz inequality, and the
equality is achieved when w
= C
1
2
1
w is the principal
eigenvector corresponding to
max
of A and h
T
w
> 0.
Considering that A is rank-one, its maximal eigenvalue is
C
T/2
1
h
2
2
= C
1/2
1
h
2
2
, which leads to the third step,
and its principal eigenvector is C
T/2
1
h = C
1/2
1
h with
denoting an arbitrary non-zero scalar. To satisfy h
T
w
> 0,
> 0 is required. Therefore, the weighting vector enabling
the second and third steps is w
l
C
1/2
1
w
= C
1
1
h with
being an arbitrary positive scalar. Note that w
l
with different
positive s lead to the same result, which can be veried by
substituting w
l
into (14) or (10). So different s make no
difference for data demodulation.
For the right-hand inequality, in a similar way we can arrive
at
max
w
x
1
2
C
1/2
0
h
2
(15)
w
w
w
.
t
a
k
e
o
f
f
p
r
o
j
e
c
t
s
.
c
o
m
For Further Details-A Vinay 9030333433,0877-2261612 1
CHENG et al.: OPTIMAL BER-BALANCED COMBINING FOR WEIGHTED ENERGY DETECTION OF UWB OOK SIGNALS 355
where the weighting vector used is w
r
C
1
0
h = C
1
0
h
with being an arbitrary positive scalar (the explanation is
the same as ). Combining (14) and (15) leads to Lemma 1.
Remark 1: If C
0
C
1
, the bounds on max
w
x will be tight.
In this case, w
l
, w
r
will be close to the optimal weighting
vector w
o
and thus are good candidates for being a near-
optimal weighting vector. Fortunately, it is veried through
simulations that at the signal-to-noise ratios (SNRs) of interest,
C
0
, C
1
are close. Thus we can choose w
l
, w
r
as near-optimal
alternatives.
1
In particular, w
r
is preferred since it bears a
simpler form w
r
= h.
2
Note that w
r
is also presented in [6],
where, however, it is derived by maximizing the deection
coefcient and is used for timing synchronization.
For data demodulation of UWB OOK signals, one counter-
part scheme is used in [6], which works as shown below
z = a
T
y, s =
z N
f
N
0
BT
M
(a
T
1
M
)
a
T
h
s = sign( s 0.5)
(16)
where a represents the weighting vector, 1
M
represents a
column vector of M 1s. In particular, the following MMSE-
based weighting vector is recommended in [6]:
a
mmse
=
M
1
T
M
(C
0
+C
1
)
1
1
M
(C
0
+C
1
)
1
1
M
(17)
where a
T
1
M
= M.
Remark 2: Using s(i) to represent s in (16) corresponding
to s = i, we have s(0) N(
0
= 0,
2
0
=
a
T
C0a
(a
T
h)
2
) and
s(1) N(
1
= 1,
2
1
=
a
T
C1a
(a
T
h)
2
). Since s(0) and s(1) have
different variances, the threshold 0.5 will lead to different
BERs for bits 0 and 1. To minimize the BER of s = 0,
we should minimize
2
0
, which can be achieved by using the
weighting vector a = C
1
0
h.
3
Similarly for s = 1, the BER-
minimizing weighting vector is a = C
1
1
h. It is easily seen
that both a and a impose larger (smaller) weight on y
m
with
larger (smaller) signal energy E
m
. However, a
mmse
in (17)
behaves in the opposite way. In other words, a
mmse
weights
y
m
of larger (smaller) E
m
with smaller (larger) weighting
coefcient. Therefore, a
mmse
is an improper weighting vector.
Simulations in next section show that a
mmse
leads to poor
BER performance.
IV. SIMULATION AND DISCUSSION
The second derivative Gaussian pulse with width T
p
= 1ns
is used. Multipath channels are generated according to the
CM1 model in [9] with 1ns resolution and maximum spread
truncated to be 100ns. The other parameters are: B = 5GHz,
T
f
= 100ns and N
f
= 10. Besides, we assume h is per-
fectly known for all WED schemes. The information symbol
s {0, 1} is randomly generated with equal probability. For
clarity, in the gures, we use the following notations:
1
In fact, these two vectors performance curves merge together.
2
Since stretching wr makes no difference for the demodulation, we omit
the scalar:/(N
f
N
2
0
BT
M
).
3
This can be derived by using the Rayleigh Ritz inequality, as in the proof
of Lemma 1.
12 13 14 15 16 17 18 19 20 21 22
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
1.02
1.04
E
b
/N
0
(dB)
B
E
R
(
0
/
B
E
R
(
1
)
MMSE
ML
BERbalanced (optimal)
BERbalanced (nearoptimal)
Fig. 1. BER ratios of different WED schemes with BER(i) representing the
BER conditioned on s = i, M = 10, T
M
= 10ns.
CED represents the conventional energy detection in
[3].
BER-balanced (optimal) represents the proposed BER-
balanced WED using w
o
in (11) as the weighting vector.
BER-balanced (near-optimal) represents the proposed
BER-balanced WED using w
r
= h as the weighting
vector.
MMSE represents the MMSE WED in [6], which uses
(16) and (17)) for data modulation
ML represents the non-linear ML WED [10], which is
derived using the fact that y
m
s are Gaussian distributed
(see (4)). The ML WED achieves the minimum average
BER and thus serves as the performance benchmark.
For fair comparison, all WED schemes use the same M in the
simulations.
Fig. 1 compares the ratios of BER(0) to BER(1) of different
WED schemes, where BER(i) represents the BER conditioned
on that s = i is transmitted. The number of sub-intervals is
M = 10 and thus T
M
= T
f
/M = 10ns. Its shown that
the ratio curves corresponding to BER-balanced (optimal)
and BER-balanced (near-optimal) merge with the line 1
over varying SNRs. This means that the proposed schemes
can provide balanced BERs for bits 0 and 1. In constrast,
the curves of ML and MMSE schemes deviate from the
constant 1, which means that their BERs are unbalanced.
Moreover, when increasing SNR, ML and MMSE incur
more unbalance. This is because for these two schemes, the
difference of the decision statistics variances for bits 0 and
1 becomes larger with increasing SNR.
Fig. 2 depicts the BER curves of the CED and different
WED schemes , together with the lower and upper bounds
derived in Lemma 1. The curves are obtained by averaging
over the results of bits 0 and 1. Note that the three curves
corresponding to ML, BER-balanced (optimal) and BER-
balanced (near-optimal) merge together. This means that
the proposed BER-balanced WED schemes achieve the best
average BER performance. While for MMSE, it performs
obviously worse than the CED and the other WED schemes,
w
w
w
.
t
a
k
e
o
f
f
p
r
o
j
e
c
t
s
.
c
o
m
For Further Details-A Vinay 9030333433,0877-2261612 2
356 IEEE COMMUNICATIONS LETTERS, VOL. 17, NO. 2, FEBRUARY 2013
14 16 18 20 22 24 26
10
5
10
4
10
3
10
2
10
1
10
0
E
b
/N
0
(dB)
B
E
R
MMSE
BERbalanced (optima)l
BERbalanced (nearoptimal)
Lower bound
Upper bound
ML
CED
Fig. 2. BER performance of the CED and different WED schemes, M =
10, T
M
= 10ns for WED schemes.
14 16 18 20 22 24 26
10
4
10
3
10
2
10
1
10
0
E
b
/N
0
(dB)
B
E
R
CED
BERbalanced (nearoptimal). M=5
BERbalanced (nearoptimal). M=10
BERbalanced (nearoptimal). M=20
Fig. 3. BER performance of the proposed near-optimal BER-balanced WED
using different Ms.
due to its improper weighting vector a
mmse
. Its worthy to
point out that due to different channel model and simulation
parameters used, here MMSE exhibits different performance
from that in Fig. 11 of [6]. Moreover, we also conducted the
simulations of the BER-balanced WED using w
l
, the curve
of which merges with that of BER-balanced (near-optimal)
and so is omitted for clarity.
Fig 3 illustrates the effect of M on the performance
of the proposed near-optimal BER-balanced scheme. Since
BER-balanced (optimal) and BER-balanced (near-optimal)
exhibit indistinguishable curves (as seen in Fig. 2), only
the latters results are shown here for clarity. It can be
seen that when enlarging M, the near-optimal scheme enjoys
increasingly more performance gain over the CED, but the
gain increment is decreasing gradually. Note that larger M
means higher system complexity. So it is necessary to draw
an proper trade-off between performance and complexity.
V. CONCLUSION
To counteract the enhanced noise in CED, weighted pro-
cessing is often suggested. However, for the demodulation
of UWB OOK signals, the existing WED schemes fail to
provide balanced BERs for different bits. In view of this, we
propose an optimal BER-balanced WED, which achieves the
BER balance by setting a proper threshold and determines the
optimal weighting vector by solving a convex optimization
problem. Furthermore, a simple near-optimal weighting vector
is derived to avoid the computation required by the optimal
detector. Simulations show that the proposed WEDs can
achieve the best performance and provide desirable balanced
BERs for different bits.
REFERENCES
[1] K. Witrisal, G. Leus, G. Janssen, et al., Noncoherent ultra-wideband
systems, IEEE Signal Process. Mag., pp. 4866, July 2009.
[2] M. E. Sahin, I. Guvenc, and H. Arslan, Optimization of energy detector
receivers for UWB systems, in Proc. 2005 IEEE Veh. Technol. Conf.
Spring, pp. 13861390.
[3] X. Cheng, Y. L. Guan, and Y. Gong, Thresholdless energy detection of
ultra-wideband block-coded OOK signals, Electron. Lett., vol. 44, pp.
755756, June 2008.
[4] J. Wu, H. Xiang, and Z. Tian, Weighted noncohernet receivers for UWB
PPM siganls, IEEE Commun. Lett., vol. 10, pp. 655657, Sep. 2006.
[5] X. Cheng and Y. L. Guan, Mitigation of cross-modulation interference in
UWB energy detector receivers, IEEE Commun. Lett., vol. 13, pp. 375
377, June 2009.
[6] F. Wang, Z. Tian, and B. M. Sadler, Weighted energy detection for
noncoherent ultra-wideband receiver design, IEEE Trans. Wireless Com-
mun., vol. 10, pp. 710720, Feb. 2011.
[7] H. Urkowitz, Energy detection of unknown deterministic signals, Proc.
IEEE, vol. 55, pp. 523531, Apr. 1967.
[8] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge Univer-
sity Press, 2003.
[9] A. F. Molisch, IEEE 802.15.4a Channel ModelFinal Report IEEE, 2004.
Available: http://www.ieee802.org/15/pub/TG4a.html.
[10] D. Mu and Z. Qiu, Weighted non-coherent energy detection receiver
for UWB OOK systems, Proc. 2008 ICSP Conf.
w
w
w
.
t
a
k
e
o
f
f
p
r
o
j
e
c
t
s
.
c
o
m
For Further Details-A Vinay 9030333433,0877-2261612 3