[ieee 2012 international conference on wireless communications & signal processing (wcsp 2012) -...
TRANSCRIPT
A Low-Complexity Symbol-Level Differential
Detection Scheme for IEEE 802.15.4 O-QPSK
Signals
Cong Wang∗ Yongpan Liu∗ Rong Luo∗ Huazhong Yang ∗
∗ Electronic Engineering Dept.
Tsinghua University
Beijing, 100084, China
Abstract—This paper presents a low-complexity differential
detection scheme for the IEEE 802.15.4 O-QPSK signals. The
proposed scheme utilizes a novel additive decision metric to
perform symbol-level detection. Corresponding low-complexity
receiver architecture is also developed. Compared with tradi-
tional MBDD chip-level detector, the proposed scheme achieves
a performance improvement of 2.1 to 3.5 dB. In the meantime,
the complexity does not increase much.
I. INTRODUCTION
The IEEE 802.15.4 is a protocol designed for wireless
personal area networks (WPAN), which mainly focuses on
short-range, low-power and low-cost network applications. It
includes the specifications of the physical layer (PHY) and the
media access control layer (MAC) [1].
There are four PHYs that are specified in the IEEE 802.15.4
standard, of which three are in 868/915 MHz band and the
other one is in 2450 MHz ISM band. As the 2450 MHz band
is in the worldwide ISM band which can be used without
a licence, in this paper, we focus on the 2450 MHz half
sine pulse shaped O-QPSK signal, which is equivalent to a
precoded MSK signal [2].
As WPAN devices conforming with the IEEE 802.15.4 are
often battery-powered but expected to operate for months,
low-power receivers with both low complexity and good bit
error rate (BER) performance are highly appreciated. Low
complexity enables lesser processing power of the detector,
while good receiving BER performance allows lesser RF
power at the transmitting-side.
In low-complexity receivers, differential detection is often
used because carrier phase recovery is not needed. Although
traditional differential detector (with observing interval N =
1) is relatively simple, it suffers from performance degradation,
which has been well discussed in [3].
Simon and Divsalar improve the BER performance of
conventional differential detection with a so-called multiple-
bit differential detection (MBDD) technique. By extending
This work was supported in part by the NSFC under grant 60976032, National
Science and Technology Major Project under contract 2010ZX03006-003-
01, and High-Tech Research and Development (863) Program under contract
2009AA01Z130.
the observing interval, the BER performance of the MBDD
detector fills the gap between conventional differential detector
and the coherent detector [4]. Such technique has been used
to detect various kinds of signals including CPM [5, 6]
GMSK [7], and O-QPSK [8–10].
However, the previous works in [5–10] all assume the
phase offset to be constant within the observing window,
which makes them unsuitable for symbol-level detection of the
IEEE 802.15.4 OQPSK signal. Since the length of the spread
spectrum sequence is 32, to perform symbol-level detection of
the IEEE 802.15.4 OQPSK signal, we have to set the observing
interval N to 32. However, the phase offset is unlikely to be
constant during such a long interval of 32 chips, especially
when the carrier frequency offset is up to ±40 ppm as specified
in the standard [11], resulting in heavy performance loss.
Shortening the observing interval, we can perform chip-level
detection for the same signal in a block-by-block manner, due
to the non-additive property of the traditional MBDD decision
metric. Specifically, the MBDD detector outputs a block of N
chips at a time. And several blocks are combined, obtaining
a 32-chip sequence. The 32-chip sequence is then mapped
back to a symbol by comparing hamming distances. However,
such block-by-block chip-level detection scheme also faces
performance loss problem because it ignores the information
between adjacent blocks. Besides, it considers some chip
sequences that are not in the signal set, which indeed increases
the implementation complexity.
In this paper, we concentrate on designing a detector with
both low complexity and good BER performance. We first
develop a Novel Additive Metric (NAM) for the detection of
a chip based on its correlation with the previous N chips.
Then the decision metric of a hypothetical chip sequence is
derived by summing up the NAM of each chip. By calculating
and comparing the decision metric of each of the 16 chip
sequences in the IEEE 802.15.4 signal set, the most likely
symbol is selected. Corresponding simplified implementation
techniques for such symbol-level detector are also presented.
Experiments showed that the proposed symbol-level detec-
tion scheme achieves 2.4-3.5 dB performance gain compared
978-1-4673-5829-3/12/$26.00 ©2012 IEEE
Fig. 1. Reference Modulation Diagram
with conventional chip-level MBDD detector (N > 1), while
the implementation complexity does not increase much.
The rest of the paper is organized as follows. Section II
describes the mathematical representations of the transmitted
and received signal. In Section III, we review the traditional
chip-level detection scheme, and discuss the pros and cons of
it. The proposed symbol-level detection scheme is developed
and discussed in Section IV. BER performance and implemen-
tation complexity of different detection schemes are compared
in Section V. Conclusion is drawn in Section VI.
II. SIGNAL MODEL
Figure 1 illustrates the reference modulation flow for the
2450 MHz O-QPSK signal specified in [11]. First, every four
bits of the binary data are mapped into a symbol (0-15). Then
each of the symbol is mapped into a 32-chip sequence c
according to Table I. Next, in the O-QPSK modulator, the
chip sequences are modulated onto the carrier using Offset
Quadrature Phase Shift Keying with half sine pulse shaping.
Even-indexed chips are modulated onto the in-phase (I) carrier
and odd-indexed chips are modulated onto the quadrature-
phase (Q) carrier [11]. Mathematically, the complex baseband
TABLE I
SYMBOL TO CHIP MAPPING TABLE
Binary Data Symbol Chip Sequence c
[b0b1b2b3]) (Decimal) [c0c1 . . . c31]
0000 0 11011001110000110101001000101110
0001 1 11101101100111000011010100100010
0010 2 00101110110110011100001101010010
0011 3 00100010111011011001110000110101
0100 4 01010010001011101101100111000011
0101 5 00110101001000101110110110011100
0110 6 11000011010100100010111011011001
0111 7 10011100001101010010001011101101
1000 8 10001100100101100000011101111011
1001 9 10111000110010010110000001110111
1010 10 01111011100011001001011000000111
1011 11 01110111101110001100100101100000
1100 12 00000111011110111000110010010110
1101 13 01100000011101111011100011001001
1110 14 10010110000001110111101110001100
1111 15 11001001011000000111011110111000
signal can be represented as
si(t) = Ii(t) + jQi(t), i = 0, 1, . . . , 15 (1)
where
Ii(t) =15∑
n=0
C(i)2ng(t− 2nTc), i = 0, 1, . . . , 15 (2)
Fig. 2. Demodulation Diagram
Qi(t) =15∑
n=0
C(i)2n+1g(t− (2n+1)Tc), i = 0, 1, . . . , 15 (3)
and g(t) is the half sine pulse
g(t) =
{sin( π
2Tct) , 0 ≤ t ≤ 2Tc
0 ,else(4)
In (2) to (3), C(i)k = 2c
(i)k − 1, and c
(i)k is the kth chip of
Symbol i in Table I. Tc is the inverse of the chip rate, which
equals 0.5μs, since the chip rate is 2 Mcps in 2450 MHz band.
It had been proven that the O-QPSK signal with half sine
pulse shaping is equivalent to precoded MSK [3]. Along with
the representations in [10], the MSK signal can be written as
a special case of CPM signal
s(t) = Aejφ(t,p) (5)
where during the time interval kTc ≤ t ≤ (k + 1)Tc,
φ(t,p) = πpkq(t− kTc) +π
2
k−1∑i=−∞
pi (6)
and q(t) is the phase response with linear property.
q(t) =
{t
2Tc, 0 ≤ t ≤ Tc
0 ,else(7)
pi is the ith chip of the sequence of p, which determines
the moving direction of phase during the chip interval. The
relation between sequence p and c is shown as follows.
pi = −ai−1ai
a2i = (−1)iC2i, a2i+1 = (−1)i+1C2i+1
(8)
where Ci, ai, pi ∈ {1,−1} and i ∈ Z.
Therefore, we can demodulate the O-QPSK signal with an
MSK demodulator as shown in Figure 2, where the p to
symbol mapping table can be calculated in advance with (8)
and Table I.
We consider the modulated signal transmitted over an
AWGN channel, and the carrier frequency and phase are not
perfectly recovered at the receiving end. Thus the received
signal has the form
r(t) = s(t)ej(ωct+γ(t)) + n(t) (9)
where n(t) is complex additive white Gaussian noise with
single-sided power spectral density n0, ωc is the carrier
frequency offset and γ(t)) is the random phase offset.
III. TRADITIONAL MBDD DETECTOR
An MBDD detector improves the BER performance of a
conventional differential detector by extending the observing
interval from one chip to N chips. With the representation in
[10], the branch metric of the kth chip is defined as
z (k, p̃k) = e−jθ̃k
∫ (k+1)Tc
kTc
r(t)e−jπp̃kq(t−kTc)dt (10)
The decision metric for the N -chip sequence during time
period nTc ≤ t ≤ (n+N)Tc is
η(p̃n,N ) =
∣∣∣∣∣n+N−1∑k=n
z(k, p̃k)
∣∣∣∣∣ (11)
where p̃n,N is the hypothetical N -chip sequence indexed from
n to n + N − 1, i.e., p̃n,N = (p̃n, p̃n+1, . . . p̃n+N−1). θ̃k is
defined in a recursive way as follows:
θ̃k = θ̃k−1 +π
2p̃k−1 (12)
With the metric defined in (11), we can demodulate a block
of N chips according to the following decision rule.
p̂n,N = argmaxp̃n,N
{η(p̃n,N )
}(13)
By setting N = 32, we can get an ideal symbol-level MB-
DD detector for the IEEE 802.15.4 OQPSK signal. However,
the decision rule in (13) requires that ωc = 0 and γ(t) remains
unchanged during the N -chip time interval. Otherwise only a
small N can be used. Experiments in section V shows that
with ωc being 120 kHz, the BER performance of an MBDD
detector starts to fall when N grows to larger than 2.
Therefore, with N being far less than 32 and that the
metric in (11) is not additive, we can only use an MBDD
detector to perform chip-level detection for the IEEE 802.15.4
OQPSK signal in a block-by-block manner. For example, if
N = 4, then eight blocks of length-4 sequence are cascaded,
generating a 32-chip sequence p̂, which is mapped back to a
symbol by comparing hamming distances with each of the p
sequences in the p to symbol mapping table. But such chip-
level detection scheme ignores the relation between adjacent
blocks. We can further improve its performance with a symbol-
level detection scheme.
IV. PROPOSED DETECTION SCHEME
In this section, we present a novel detection scheme with
both good BER performance and low implementation com-
plexity. It is a symbol-level scheme with a Novel Additive
Metric (NAM) where soft information of a 32-chip sequence
is used to extract one symbol at a time, which enhances
the error probability performance. In the meantime, measures
are incorporated to restrict the growth in implementation
complexity. First, in a symbol-level detection scheme, we only
select the most likely symbol among the 16 symbols instead
of searching for the maximal likelihood chip sequence among
2N sequences. Second, with distributive law of multiplication,
number of multiplications required in decision metric calcula-
tion can be reduced.
IV.A. Detection Based on NAM
We start by introducing the correlation between the kth and
lth chip as
L(l, k, p̃n,N ) =
∫ (k+1)Tc
kTc
r(t)V ∗(t, l, k, p̃n,N )dt (14)
where
V (t, l, k, p̃n,N ) ={r(t+ (l − k)Tc)e
j(−θ̃l+θ̃k) , p̃l = p̃k
r(−t+ (l + k + 1)Tc)ej(−θ̃l+1+θ̃k) , otherwise
(15)
Again, p̃n,N is a hypothetical N -chip sequence indexed from
n to n+N − 1 and p̃k represents the k-indexed element of it.
θk has been defined in (12). V ∗(t, l, k, p̃n,N ) is the conjugate
of V (t, l, k, p̃n,N ). In addition, n ≤ l < k ≤ n + N − 1 is
imposed.
It can be seen in (15) that in the case of p̃l = p̃k, the received
signal r(t+ (l− k)Tc) is rotated by ej(−θ̃l+θ̃k), attempting to
align with r(t). The correlation in (14) can be viewed as a
pseudo match filter, where the rotated signal of the lth chip
serves as a correlation reference. In the case of p̃l �= p̃k, the
operation is pretty the same except that the rotation factor is
different and that the received signal during lT ≤ t ≤ (l+1)T
is reversed before rotation to generate a pseudo match filter
reference.
As the rotation factors ej(−θ̃l+θ̃k) and ej(−θ̃l+1+θ̃k) in (15)
are irrelevant to ωc and γ(t), the existence of frequency
and phase offset does have some negative impact on the
correlation performance because the alignment is not perfect.
However, we can shorten the correlation interval (smaller
interval between k and l) to reduce the impact, making the
detection more robust to large frequency offset and fast varing
phase offset.
Like a traditional MBDD detector, consider an observing
window of N . We define the metric for the k-indexed chip as
follows:
M(k, p̃k−N,N+1) = �
{k−1∑
l=k−N
L(l, k, p̃k−N,N+1)
}(16)
namely NAM. Substituting (14) into (16) and with the dis-
tributive law of multiplication, we can write the NAM as:
M(k, p̃k−N,N+1) =
�
{∫ (k+1)Tc
kTc
r(t)k−1∑
l=k−N
V ∗(t, l, k, p̃n,N )dt
}(17)
By summing the rotated signal of the previous chips, noises
in the observing interval are averaged, generating a better
reference for the detection of pk (when hypothesis on sequence
pk−N,N+1 is correct). Besides, writing in the form of (17)
instead of (16) reduces the number of correlations needed from
N to 1.
More importantly, NAM has the form of match filter output.
It’s intuitive to sum the NAM of 32 chips to perform symbol-
level detection for the IEEE 802.15.4 2450 MHz signal.
Therefore, decision metric for a symbol can be written as
w(p̃(i)n,32) = �
{n+31∑k=n
M(k, p̃k−N,N+1)
}(18)
In the right side of (18), the chips before p̃n, i.e.,
p̃n−N , p̃n−N+1, . . . , p̃n−1, are known chips, which are the last
few ones of the previous symbol. If we want to restrict the
correlations only within the symbol, an alternative decision
metric can be used:
w(p̃(i)n,32) = �
⎧⎨⎩
n+31∑k=n+1
∑n≤l<k
L(l, k, p̃(i)n,32)
⎫⎬⎭ (19)
To detect the modulated symbol, we just need to examine
w(p̃(i)n,32) for each of the chip sequences in the signal set, i.e.,
p̃(i)n,32 = p
(i), i = 0, 1, . . . 15, and select the one that maximize
the decision metric, which is
symbol = argmaxi
{w(p(i))
}(20)
IV.B. Implementation
The receiver architecture corresponding to the detection
scheme described above is shown in Figure 3. Due to that
reverse operation on received signal is required, we implement
the proposed scheme discretely, where received signals are
sampled with over sampling rate of β, i.e., r[n] = r(nTs) and
Ts = 1βTc. In the reference generation block V (n, ∗, ∗,p),
received signal of previous chips are delayed (p̃l = p̃k) or
reversed (p̃l �= p̃k), and then rotated to align with the current
chip. Integration in (17) is approximated by sum of discrete
samples:∫ (k+1)Tc
kTc
f(t)dt ≈
Ts
⎡⎣1
2f(kTc) +
((k+1))β−1∑n=kβ+1
f(nTs) +1
2f((k + 1)Tc)
⎤⎦(21)
which is the pseudo match filter output of the current chip.
Next pseudo match filter output of 32 chips are summed,
generating a decision metric for a chip sequence p(i) in the
signal set.
Figure 4 shows a more detailed structure of the reference
generation block. The reverse operation can be implemented
with a random access memory by writing the received samples
of a chip in one direction but read them out in an opposite
direction. Such operation may be guided by a controller.
Additionally, the controller should also set the correct rotation
factor as well as correct control command for the multiplexer
Fig. 3. Receiver Architecture of the Proposed Detection Scheme
Fig. 4. Reference Generation Block V (n, ∗, ∗, p̃)
(MUX) [12]. As the hypothetical sequence are known, which
is the 16 sequences in the signal set, these control words can
be calculated in advance and saved in a table. Furthermore,
the rotation factor ej(−θ̃l+θ̃k) or ej(−θ̃l+1+θ̃k) only takes one
of the four values: 1, j,−1,−j, the multiplication of the can
be simplified to manipulations to the real and image part of
the received samples, i.e., let r[n] = I[n] + jQ[n], we have
r[n] · 1 = I[n] + jQ[n]; r[n] · (−1) = −I[n]− jQ[n]
r[n] · j = −Q[n] + jI[n]; r[n] · (−j) = Q[n]− jI[n](22)
V. NUMERICAL RESULTS
V.A. Experiment Configuration
In this section, BER performance of the following four
detectors are compared. 1) Proposed detector. 2) Traditional
MBDD detector. 3) Coherent soft-decision O-QPSK detector.
4) Coherent hard-decision O-QPSK detector. For the proposed
detector, decision metric in (19) is applied. The traditional
MBDD detector in the simulation is a chip-level detector
in a block-by-block manner. All the BER statistics are of
the data bit after despreading. Over sampling rate β is set
to 4. For simplicity, we assume the carrier phase offset
γ(t) to be constant over the observing N chips while the
carrier frequency offset ωc exists, which can be viewed as
the time varying phase offset case in previous works [5–10].
Additionally, perfect sample timing is assumed.
Implementation complexity of the proposed detector and
a traditional MBDD detector is compared by computing the
4 6 8 10 12 14 16 1810
−4
10−3
10−2
10−1
100
Eb/N
0 in dB
Bit E
rror
Rate
Bit Error Rate Comparison of Different Detectors
CD Hard
CD Soft
Proposed N=1
Proposed N=2
Proposed N=3
Proposed N=4
Trad−MBDD N=1
Trad−MBDD N=2
Trad−MBDD N=4
Fig. 5. Performance comparison between proposed detector and traditional
detectors (ωc = 0)
6 8 10 12 14 16 1810
−4
10−3
10−2
10−1
100
Eb/N
0 in dB
Bit E
rror
Rate
Bit Error Rate Comparison of Different Detectors
Proposed N=1
Proposed N=2
Proposed N=3
Proposed N=4
Trad−MBDD N=1
Trad−MBDD N=2
Trad−MBDD N=4
Fig. 6. Performance comparison between proposed detector and traditional
detectors (ωc=120 kHz)
number of additions and multiplications required in calculation
of their decision metrics.
V.B. BER Performance
Figure 5 shows the simulation results of the BER perfor-
mance of different detectors when frequency offset ωc = 0.
At BER= 10−3, the proposed symbol-level detector achieves
a performance advantage of 2.4 (N = 4) to 3.5 dB (N = 1)
with respect to the conventional chip-level MBDD detector.
Figure 6 is the BER performance of different detectors
when frequency offset ωc = 120 kHz (50 ppm). The optimal
observing interval for a conventional MBDD has been reduced
to 2. As illustrated, the proposed symbol-level detector again
shows a 2.5 dB advantage against traditional MBDD detector
(both N = 2).
In Table II, how the optimal observing window size N
changes with frequency offset ωc is listed. As expected, when
ωc becomes larger, chips that are far ahead cannot be used
as accurate references and the observing interval N should
be shorten. The BER advantages of the proposed detector
(at optimal window size) over the traditional MBDD detector
are also listed in Table II, which are around 2.4 dB at BER
= 10−3.
TABLE II
OPTIMAL OBSERVING WINDOW VS. CARRIER FREQUENCY OFFSET
Frequency Offset0 10 20 30 40 50
(ppm)
Optimal N of32 11 7 4 4 2
Trad. MBDD
Optimal N of32 11 6 3 2 2
Proposed Detector
BER Advantage 2.4(N = 4)2.4 2.5 2.5 2.1 2.4
over MBDD (dB) 3.5(N = 1)
V.C. Implementation Complexity
In an MBDD detector, to calculate the decision metric
in (11), approximately βN complex additions and βN
complex multiplications are required for each hypothetical
sequence. To demodulate an N -chip sequence, we calculate
and compare the decision metrics of all the 2N hypothetical
sequences. And approximately 32/N blocks of such N -chip
sequences are cascaded to demodulate one IEEE 802.15.4
OQPSK symbol. Therefore, total number of additions and
multiplications required to demodulate one symbol can be
represented as:
MBDD additions ≈ 32β · 2N (23)
MBDD multiplications ≈ 32β · 2N (24)
In the proposed detector, for each spread sequence in the
signal set, we need to perform approximately 32Nβ complex
additions and 32β complex multiplications to calculate the
decision metric described in (18). To demodulate one symbol,
decision metrics of the 16 spread sequences are calculated
and compared. Therefore, total number of additions and
multiplications required to demodulate one symbol can be
represented as:
Proposer Detector additions ≈ 32Nβ × 16 (25)
Proposed Detector multiplications ≈ 32β × 16 (26)
Using Table II and Equations (23) to (26), we can compare
the implementation complexity of both detectors at optimal
observing window size under different values of ωc. Table
III summarizes the number of additions and multiplications
required in calculation of the decision metric of one IEEE
802.15.4 2450 MHz symbol. It can be seen that the number of
multiplications remains unchanged when N is growing larger.
This is expected as we only calculate the decision metric for
the 16 sequences in the signal set. In addition, compared to the
traditional MBDD detector, implementation complexity of the
proposed detector is lower when ωc is small, but a bit higher
with large ωc.
TABLE III
IMPLEMENTATION COMPLEXITY COMPARISON
Frequency Offset10 20 30 40 50
(ppm)
Complex Addition65536β 4096β 512β 512β 128β
Trad. MBDD
Complex Addition5632β 3072β 1536β 1024β 1024β
Proposed Detector
Complex Multiplication65536β 4096β 512β 512β 128β
Trad. MBDD
Complex Multiplication512β 512β 512β 512β 512β
Proposed Detector
VI. CONCLUSION
We develop a symbol-level detector for the IEEE 802.15.4
OQPSK signal based on a novel additive metric. It has been
shown that the proposed detector achieves a performance
gain of 2.1 to 3.5 dB compared with traditional MBDD
detector. Besides, simplified implementation techniques are
also introduced, which keeps the implementation complexity
low. Future works may include methods to compensate the
carrier frequency offset and detection schemes incorporating
low-cost synchronization modules.
REFERENCES
[1] N.Salman, I. Rasool, and A.H.Kemp, “Overview of the
IEEE 802.15.4 standards family for low rate wireless per-
sonal area networks,” in 7th Int. Symposium on Wireless
Commun. Systems (ISWCS), 2010, pp. 701–705.
[2] S. Lanzisera and K. S. J. Pister, “Theoretical and practical
limits to sensitivity in IEEE 802.15.4 receivers,” in 14th
IEEE Int. Conf. on Electronics, Circuits and Systems
(ICECS), 2007, pp. 1344–1347.
[3] M. K. Simon and M. S. Alouini, Digital Communication
over Fading Channels. Second Edition, Wiley, 2005.
[4] D. Divsalar and M. K. Simon, “Multiple-symbol differ-
ential detection of MPSK,” IEEE Trans. on Commun.,
vol. 38, pp. 300–308, 1990.
[5] M. K. Simon and D. Divsalar, “Maximum-likelihood
block detection of noncoherent continuous phase mod-
ulation,” IEEE Trans. on Commun., vol. 41, pp. 90–98,
1993.
[6] A. Abrardo, F. Bencivenni, and G. Benelli, “Multiple-
symbols differential demodulation and decoding of CP-
M,” in ICC ’95: Proc. of IEEE Int. Conf. on Commun.,
1995, pp. 688–691.
[7] A. Abrardo, G. Benelli, and G. R. Cau, “Multiple-symbol
differential detection of GMSK for mobile communica-
tions,” IEEE Trans. on Vehicular Technology, vol. 44, pp.
379–389, 1995.
[8] W. G. Phoel, “Improved performance of multiple-symbol
differential detection for O-QPSK,” in WCNC ’2004:
Proc. IEEE Wireless Commun. and Networking Conf.,
2004, pp. 548–553.
[9] R. Schober, I. Ho, and L. Lampe, “Enhanced multiple-
bit differential detection of DOQPSK,” IEEE Trans. on
Commun., vol. 53, pp. 1490–1497, 2005.
[10] E. Perrins, R. Schober, M. Rice, and M. K. Simon,
“Multiple-bit differential detection of shaped-offset QP-
SK,” IEEE Trans. on Commun., vol. 55, pp. 2328–2340,
2007.
[11] IEEE Standard for Information TechnologyłPart 15.4:
Wireless Medium Access Control (MAC) and Physical
Layer (PHY) Specifica-tions for Low-Rate Wireless Per-
sonal Area Networks (WPANs), IEEE 802.15.4-2006 Std.,
September 2006.
[12] H.-J. Jeon, T. Demeechai, W.-G. Lee, D.-H. Kim, and
T.-G. Chang, “IEEE 802.15.4 BPSK receiver architecture
based on a new efficient detection scheme,” IEEE Trans.
on Signal Processing, vol. 58, pp. 4711–4719, 2010.