See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224386821

Differential Preamble Detection in Packet-Based Wireless Networks

Article in IEEE Transactions on Wireless Communications · March 2009

DOI: 10.1109/TWC.2009.071169 · Source: IEEE Xplore

CITATIONS READS

29 472

4 authors, including:

Christian Schlegel
Dalhousie University
287 PUBLICATIONS 4,503 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Turbo Coding View project

Iterative MIMO receivers View project

All content following this page was uploaded by Christian Schlegel on 15 December 2014.

The user has requested enhancement of the downloaded file.

 2006 IEEE Ninth International Symposium on Spread Spectrum Techniques and Applications

On Preamble Detection in Packet-Based Wireless

Networks
Sumeeth Nagaraj, Sheehan Khan and Christian Schlegel Marat Burnashev
Department of Electrical and Computer Engineering Institute for Problems of Information Transmission
University of Alberta, Bolshoi Karetni 19,
Edmonton, Canada. 101447, Moscow, RUSSIA.
Email:{snagaraj,sheehan,schlegel}@ece.ualberta.ca Email: [email protected]

Abstract— Performance bounds on detecting a preamble em- to achieve the same performance as coherent detection [1]. To
bedded at the start of every packet for communication over an mitigate this performance loss, coherent detection is employed
additive white Gaussian noise channel are derived. The preamble on portions (blocks) of the sequence whose length is no longer
sequence consists of blocks of spreading sequences whose length
is at most the coherence time of the channel. These blocks than the coherence time of the channel [2], [3]. These blocks
are differentially combined. A correlation-based detection is are then combined differentially.
employed to locate the boundaries of the preamble. Simulation
Sync
results closely follow the analysis. Furthermore, the effects of
frequency offset on the system performance are discussed.

I. I NTRODUCTION
Preamble Payload
A significant number of present day wireless systems are (a)
based on packetized communication. For instance, the IEEE W Lb
standards for local, metropolitan and personal area networks,
and several cellular communication standards. In such commu- Sync

nication systems, packet arrival instants are generally random Preamble Payload
and unknown at the receiver. The payload can be successfully (b)
demodulated provided the receiver has symbol timing, fre-
Fig. 1. Packet Formats.
quency and phase offset information, in addition to achieving
frame synchronization. In many practical applications, these
In this paper, we investigate the performance of a new
parameters are not known a priori at the receiver and hence
preamble detection method for an uncoordinated direct se-
have to be estimated from the received signal.
quence spread spectrum system (DSSS) using the modified
To facilitate the process of synchronization, a preamble is
packet format as shown in Fig. 1(b). The preamble is made up
prefixed for each payload transmission. Additionally, a syn-
of a sequence of W blocks, which are differentially encoded
chronization (sync) pattern included in the preamble enables
and modulated by a random DSSS sequence of Lb chips per
frame synchronization. Figure 1(a) shows a typical packet
block. A correlation type preamble detector is employed. The
format with preamble and sync sequence prefixed to the
analysis assumes a Poisson source traffic model with packet
payload, however, the fundamental problem is to achieve frame
inter-arrival instants modeled by an exponential distribution.
synchronization, i.e., finding the exact starting time of the
In [4] the preamble detection problem for a TDMA system
payload and a more general view is given in Figure 1(b).
employing coherent detection was considered. Note that co-
If perfect timing, frequency and phase offset estimates are
herent preamble detection requires frequency and phase offset
available at the receiver1, a typical frame synchronization
estimates and hence the system incurs rate loss. Moreover,
algorithm correlates the known sync word with the received
in a TDMA system, packet arrival instants are known at the
sequence to find the start of the packet. The maximum
receiver to be at the beginning of each time slot. Hence, a
correlation peak determines the location of the sync word.
preamble is either detected or missed. However, in an unco-
On the other hand, without the frequency and phase offset
ordinated multiple access system the packet arrival instants
estimates, the baseband signal rotates thereby degrading the
are random. This leads to an additional event of erroneously
performance of a correlation-based preamble detection.
detecting a preamble.
The effect of carrier phase offset can be overcome by
The rest of the paper is organized as follows. In Section II
using differential encoded phase modulation. This scheme,
we describe the system model and the preamble detection
however, requires nearly 3 dB additional signal-to- noise ratio
algorithm. Based on the proposed system model, performance
1 Estimation of frequency and phase offset require pilot signal. Hence the bounds on the preamble detection algorithm are derived in
system incurs rate loss. Section III. In Section IV, simulation results are provided to

0-7803-9780-0/06/$20.00 ©2006 IEEE 476

show the accuracy of the analysis. The conclusions follow in Assuming perfect chip timing knowledge at the receiver2
Section V. and after chip-matched filtering and sampling, the baseband
output is represented as
II. S YSTEM M ODEL AND A SSUMPTIONS
The packet arrival pattern is modeled as a Poisson process r(k) = s(k)ej(2πΔf Tc k+φ) + n(k). (5)
with intensity λ. The packet inter-arrival instants follow a
truncated exponential distribution in the time window [1, S]. These samples are despread and summed to obtain âm , a
Therefore, the packet inter-arrival time distribution is given by complex valued estimate of the bit am , i.e.,
 (m+1)(Lb −1)
c(S)λe(−λi) , 1≤i≤S 
pi0 (i) = (1) âm (k) = r(l + k)bl for m = 0 . . . W − 1. (6)
0, otherwise
l=mLb
where c(S) is a normalization constant
To detect the boundaries of the preamble we use the fol-
1 1 − e−λ lowing decision rule. The receiver differentially correlates the
c(S) = S = −λ . (2)
i=1 λe−λi λe (1 − e−Sλ ) sequence {âm (k)} with the known preamble. The correlation
The modulator block diagram is shown in Figure 2. The threshold at kth time interval is given by
preamble sequence c = (c0 , c1 , . . . , cW −1 ) composed of W −1
1 
independent and identically distributed binary phase shift ηk = âm (k)â∗m−1 (k)cm , (7)
keying (BPSK) bits, i.e., cm ∈ {1, −1}, is differentially Lb m=0
encoded into a sequence {am }, wherein am = cm am−1
where â∗ denotes the complex conjugate of â and ηk is a
with initial value a0 = c0 . The sequence {am } is then
complex decision variable. The preamble detection algorithm
oversampled by a factor Lb and multiplied by a random but
computes ηk at every chip time interval. Note that only the
known spreading (chip) sequence {bk } resulting in {dk }, for
real part of the decision variable ηk is required to make a
k = 0 . . . Lb W − 1. The block interval is given by Tb = Lb Tc ,
decision. The decision variable Re ηk greater than a threshold
where Tc is the chip interval. The spreading sequence {bk }
G indicates the presence of a preamble, i.e., at the stopping
consists of independent and identically distributed random
moment: min{k ∈ [1, S] : Re ηk ≥ G}.
variables with P(bk = 1) = P (bk = −1) = 1/2. The spread-
ing sequence need not be periodic but periodic sequences lead The threshold G is the correlation peak for a perfect overlap
to significantly less complex hardware implementations. The case with k = i0 for the noise free case. This is defined as
complex envelope of the transmitted signal s(t) has the form def
G = (1 − g)(W − 1)Lb Ec cos(2πΔf Lb Tc ), (8)
 W −1
Lb
s(t) = Ec dk p(t − kTc − i0 Tc ), (3) where g ∈ [0, 1] determines a fraction of the threshold level
k=0 G.
where i0 ∈ [1, S] is the starting moment of the preamble, Ec
denotes energy per chip and p(t) is a pulse shaping function III. P ERFORMANCE A NALYSIS
with support [0, Tc ] and normalized to unit energy. We define two error events conditioned on the presence of
the preamble at moment i0 . First, the decision variable being
cm am dk s(t)
↑Lb p(t) greater than the threshold G, at moment k < i0 , flags the
erroneous presence of a preamble. We call such an event a
Tb wrong detection and denote its probability by P(wrong|i0 ).
bk Secondly, a preamble is missed if the decision variable at
moment i0 , is less than the threshold G. We call such an event
Fig. 2. Modulator block diagram
a miss and denote its probability by P(miss). The probability
of successful preamble detection is then given by
We consider communication over an additive white Gaus-
sian noise (AWGN) channel. The received signal is represented
as P(success) = 1 − P(wrong) − P(miss). (9)
r(t) = e(2πΔf t+φ)j s(t) + n(t), (4)
A. Probability of Missing a Preamble
where Δf and φ denotes carrier frequency and phase offset
Assuming that a preamble starts at i0 , the probability of
respectively. The frequency offset is a constant whereas phase
missing the preamble is given by
offset is modeled as a uniform random variable in the interval
[0, 2π). Additionally, the transmitted signal s(t) is corrupted P(miss) = P(Re ηi0 ≤ G). (10)
by n(t), a zero mean complex valued Gaussian white noise
process with independent real and imaginary components each 2 This assumption leads to an upper bound on the performance of the
with variance N0 /2. preamble detection method for an asynchronous system.

477
The decision variable Re ηi0 may be expressed in the quadratic The distribution function of a linear combination of noncentral
form chi-square random variables is calculated using the character-
W −1 istic function (c.f) approach and numerical integration (see
1   
Re ηi0 = cm âm (i0 )â∗m−1 (i0 ) + â∗m (i0 )âm−1 (i0 ) Appendix). The distribution function can also be evaluated
2 m=0 by a series expansion given in [6]. However, the rate of
1 † convergence becomes slower for increasing values of W .
= â Câ, (11)
2 B. Probability of Erroneous Detection
where [·]† denotes conjugate transpose, and The probability of wrong detection conditioned on the
â = (â0 , â2 , . . . , âW −1 )T . Further, the estimates {âm } occurrence of the preamble at moment i0 is given by
are complex Gaussian random variables with expectation μa . i0 −1
The (W × W ) matrix C is symmetric with entries, P(wrong|i0 ) = P Re ηk ≥ G|i0
 k=1
ci+1 , |i − j| = 1
Cmin (i,j) = (12) 0 −1
i
0, otherwise.
≤ P(Re ηk ≥ G|i0 ). (17)
Note that the entries in the vector â are not independent k=1
because of differential encoding. In general, to make the The inequality in (17) is obtained by using the union bound on
entries independent we first apply a whitening transformation. the erroneous detection events. We consider two cases based
The uncorrelated vector is then orthogonally rotated to ob- on the correlation overlap.
tain independent components. In this problem, the variance- (i0 −W Lb )+

covariance matrix is the identity matrix, i.e., P(wrong|i0 ) ≤ P(Re ηk ≥ G|i0 )
Σa = E(â − μa )(â − μa )† = I.
k=1
(13)
0 −1
i

where N0 = 1. Furthermore, there exists a matrix P such that + P(Re ηk ≥ G|i0 ) (18)
k=(i0 −W Lb +1)+
PT CP = Λ, PT P = I, (14) Assuming that i0 − W Lb ≥ 1, the first case involves
correlating the known preamble with the received Gaussian
where Λ is the diagonal Eigenvalue matrix of C. Now,
noise samples. The second case captures partial correlation of
substituting for C in (11) results in
the known preamble sequence with the received preamble.
1 † 1
W −1
1  Case 1. 1 ≤ k ≤ i0 − W Lb : The decision variable, ηk , may
Re ηi0 = â PΛPT â = Z† ΛZ = λi |Zi |2 , (15) be written in the quadratic form
2 2 2 i=0
1
Re ηk = ↠Câ, (19)
where Z = PT â, is an independent complex Gaussian vector 2
with unit variance and expected value E[Z] = PT μa . There- where âT = [n(k), n(k + 1), . . . , n(k + W − 1)]. After mean
fore, |Zi |2 are non central chi-squared distributed random removal, whitening and orthogonal transformation, (19) may
variables with two degrees of freedom and non-centrality be written as
parameter given by δi2 = |E[Zi ]|2 . Hence the quadratic form k+W −1
1 1 
is a linear combination of non central chi-square distributed Re ηk = Z† ΛZ = |Z|2i λi , (20)
random variables. 2 2
i=k
In the sequel the following Lemma will be needed. where Z = P â is a complex standard Gaussian random vec-
T
Lemma 1: The Eigenvalues λi and corresponding Eigen- tor. Therefore, |Z|2i are central chi-square distributed random
vectors pi of the matrix C are given by variables with two degrees of freedom. Equation (20) is a

πi linear combination of central chi-squared distributed random
λi = −2 cos , 1 ≤ i ≤ W. variables. The distribution function of a linear combination
W +1
j−1    of central chi-squared distributed random variables can be
iπ iπ computed in closed form (see Appendix). Therefore,
pij = −2 ck cot sin j π − ,
W +1 W +1 (i0 −W Lb )+
k=1 
1 ≤ i, j ≤ W, P (Re ηk ≥ G|i0 )
k=1
where λ1 < λ2 < . . . < λW . −1

W
Proof: See [5] = (i0 − W Lb )+ P |Z|2i λi ≥ 2G|i0
The probability of miss (10) is given by, i=0
⎛ ⎞

W −1 
W W
1 ⎠ e−G/(λr )
P(miss) = P(Re ηi0 < G) = P λi |Zi |2 < 2G . (16) = (i0 − W Lb )+ ⎝ ,
λr − λi λr
i=0 r=1 i=1,i=r

478
where the W Eigenvalues λi ’s of the matrix C are computed 10
0
Header Length = 30: Spreading Gain = 32 : Detection Level = 0.5G.

using Lemma 1. Numerical

Case 2. i0 −W Lb +1 ≤ k ≤ i0 −1: The decision variable may Simulation

be written in the same form as in (20). Note that, unlike in the Δf = .2

Ŧ1 32Tc
previous case, Z = PT â, is an independent complex Gaussian 10

vector with unit variance and expected value E[Z] = PT μa .

Therefore, |Zi |2 ’s are non central chi-square distributed ran-

P(miss)
Δf = 0 Δf = .1
Ŧ2 32Tc
dom variables with two degrees of freedom and non-centrality 10

parameters given by δi2 = |E[Zi ]|2 , (i = 0 . . . W − 1). The

required probabilities, i.e., P(Re ηk ≥ G|i0 ), for i0 − W Lb +
1 ≤ k ≤ i0 − 1 are calculated from the characteristic function 10
Ŧ3

approach and numerical integration (see Appendix).

Using the above two cases in (18) we obtain an upper bound
on P(wrong|i0 ). The probability of wrong detection is then 10
Ŧ4

Ŧ25 Ŧ20 Ŧ15 Ŧ10 Ŧ5

given by the formula of total probability as Ec/N0 (dB)


S
P(wrong) = P(wrong|i)P(i0 = i) (21) Fig. 3. Probability of missing a packet for the proposed preamble detection
i=1
algorithm is shown for different SNR and frequency offset values.


S 
i−1
≤ P(Re ηk ≥ G|i)P(i0 = i) (22)
i=1 k=1 probability of erroneous detection given in (23) matches
1 closely with the upper bound.
≈ ( − 1)P(Re η1 ≥ G), (23)
λ
Header Length = 30: Spreading Gain = 32 : Detection Level = 0.5G.
where the factor 1/λ − 1 accounts for the traffic intensity 10
0

and approximation in (23) is permissible if Ec  N0 where .2

Δf =
32Tc
the self-noise is negligible. Further, the probability in (23) is
evaluated for Case 1. Ŧ1
10
A lower bound on the probability of successful detection
P(success) is obtained by substituting (21) and (16) in (9).
P(wrong)

Δf = 0 Δf = .1
Ŧ2 32Tc
IV. A NALYTICAL AND S IMULATION R ESULTS 10

The performance of the proposed preamble detection al-

gorithm is evaluated in terms of probability of miss and
Ŧ3
probability of erroneously detecting a packet versus signal to 10
noise ratio. Numerical
The probability of packet miss is shown in Figure 3 for Simulation
Approximation
different SNR and frequency offset values. Here, we consider 10
Ŧ4

a preamble sequence of length W = 30 bits. These bits are Ŧ24 Ŧ22 Ŧ20 Ŧ18 Ŧ16 Ŧ14 Ŧ12 Ŧ10 Ŧ8
SNR (dB)
differentially encoded and then multiplied by a pseudo noise
code of spreading factor Lb = 32. The threshold level g is
Fig. 4. Probability of erroneous packet detection for the proposed preamble
set to 0.5 and (16) is used to plot the probability of miss. detection algorithm is shown for different SNR and frequency offset values
From Figure 3 we can observe that the simulation results without using the traffic intensity factor.
closely follow the analysis. Further, note that a frequency
offsets, Δf = 0.1/(32Tc) and Δf = 0.2/(32Tc), results in The receiver operating curves are given in Figure 5 for
performance degradation of 0.7dB and 4.1 dB respectively. different detection threshold values. It can be seen from the
This is because, presence of frequency offset rotates the signal figure that as the detection threshold is lowered (increasing
and degrades the system performance. To compensate for the g) the probability of miss decreases while probability of false
performance degradation, the block length Lb is made at most alarm increases. On the other hand, increasing the detection
equal to the coherence time of the channel. threshold (decreasing g) leads to lower false alarms but higher
Figure 4 shows an upper bound on the probability of probability of missing a packet.
erroneous detection (18) (17) for different SNR (Ec /N0 ) and
frequency offset values. From this figure, we can observe a V. C ONCLUSION
good agreement between the analysis and simulation results. We derived performance bounds on detecting a preamble
Further, the performance degradation due to frequency offset sequence embedded at the start of each packet transmission
is shown. At low Ec /N0 values, an approximation for the over an AWGN channel. Exact results were provided for the

479
 0
Header Length = 30 : Spreading Gain = 32 The distribution function is obtained using Gil-Pelaez’s inver-
10
sion formula [7], [8]
∞ jtx
Ŧ2
10 1 1 e Φ(−t) − e−jtx Φ(t)
F(x) = + dt (25)
2 2π jt
Ŧ4 0
10
The c.f. inversion formula is evaluated numerically using the
P(miss)

Ŧ6 trapezoidal method to obtain the probabilities.

10

Distribution function of a linear combination of central

Ŧ8
10 chi-Square random variables. Let Zi ∼ CN (0, 1). The
g=0.9
g=0.8
characteristic function of the random variable λi |Zi |2 is given
Ŧ10 g=0.7 by
10
g=0.6    1
g=0.5 Φi (t) = E exp jtλi |Zi |2 = (26)
g=0.3 (1 − 2jλi t)
Ŧ12
10
10
Ŧ12 Ŧ10
10
Ŧ8
10 10
Ŧ6
10
Ŧ4 Ŧ2
10
0
10 Therefore, the c.f of a linear combination of central chi-square
P(wrong)
distributed random variables is
W
Fig. 5. Receiver operating curves is shown for different detection threshold 1
levels. SNR values Ec /N0 ∈ [−23, −5]dB are plotted at 1 dB interval. Φ(t) = . (27)
i=1
1 − 2jλit

After partial fraction expansion, (27) is written as [9],

⎛ ⎞
probability of missing a packet. Further, an upper bound on W 
W
the probability of erroneous detection closely matches with the ⎝ λr ⎠ 1
Φ(t) = . (28)
simulation results. Due to block differential encoding, the pro- λ r − λi 1 − 2jλrt
r=1 i=1,i=r
posed preamble detection method does not require phase offset
estimates. However, it is shown that for frequency offsets Now, the probability density function is obtained using the
Δf = 0.1/(32Tc) and Δf = 0.2/(32Tc), the system requires inversion formulae,
⎛ ⎞
0.7 dB and 4.1 dB additional Ec /N0 over the zero frequency W W −x/(2λr )
⎝ λ ⎠e
offset case to achieve a P(miss) = 10−4 . The performance F  (x) =
r
(29)
λr − λi λr
loss is somewhat significant for P(wrong). Additionally, an r=1 i=1,i=r
approximation for the probability of erroneous detection is
R EFERENCES
provided for low Ec /N0 values; the approximation closely
[1] J. Proakis, Digital Communications. New York: McGraw-Hill, 2001.
follows with the upper bound results. [2] D. Divsalar and M. Simon, “Multiple-symbol differential detection of
MPSK,” Communications, IEEE Transactions on, vol. 38, no. 3, pp. 300–
308, March 1990.
A PPENDIX [3] G. Colavolpe and R. Raheli, “Improved differential detection of chip-level
differentially encoded direct-sequence spread-spectrum signals,” Wireless
Communications, IEEE Transactions on, vol. 1, no. 1, pp. 125–133, Jan.
Distribution function of a linear combination of non- 2002.
central chi-square random variables. Let Zi ∼ CN (μi , 1). [4] M. R. Soleymani and H. Girard, “The effect of the frequency offset on
The characteristic function of the random variable λi |Zi |2 is the probability of miss in a packet modem using cfar detection method,”
Communications, IEEE Transactions on, vol. 40, no. 7, pp. 1205 – 1211,
given by Jul. 1992.
[5] S. Nagaraj, S. Khan, C. Schlegel, and M. V. Burnashev, “On preamble
  detection in packet-based wireless networks,” submitted to Communica-
   exp jλi δi t/(1 − 2jλi t) tions, IEEE Transactions on.
Φi (t) = E exp jtλi |Zi |2 = [6] D. Raphaeli, “Distribution of noncentral indefinite quadratic forms in
(1 − 2jλi t) complex normal variables,” IEEE Transactions on Information Theory,
(24) vol. 42, no. 3, pp. 1002–1006, May. 1996.
Therefore, the c.f of a linear combination of non central chi- [7] R. B. Davies, “Numerical inversion of characteristic function,”
square distributed random variables is Biometrika, vol. 60, no. 2, pp. 415–417, 1973.
[8] J. Gil-Pelaez, “Note on the inversion theorem,” Biometrika, vol. 38, pp.
 −1
 481–482, 1951.

W [9] C. Schlegel, “Trellis coded modulation on time-selective fading channels,”
2
Φ(t) = E exp jt λi |Zi | Communications, IEEE Transactions on, vol. 42, p. 16171627, April.
i=0
1994.
 
W
−1
exp jt λi δi /(1 − 2jtλi )
i=0
= W
−1
(1 − 2jtλi )
i=0

480

View publication stats