You are on page 1of 5

Sensitivity of Spectrum Sensing Techniques to RF impairments

Jonathan Verlant-Chenet Julien Renard Jean-Michel Dricot Philippe De Doncker Francois Horlin
Universit e Libre de Bruxelles - OPERA Dpt., Avenue F.D. Roosevelt 50, B-1050 Brussels, Belgium
jverlant@ulb.ac.be
AbstractCognitive radios are devices capable of sensing a large range
of frequencies in order to detect the presence of primary networks and
reuse their bands when they are not occupied. Due to the large spectrum
to be sensed and the high power signal dynamics, low-cost implementations
of the analog front-ends leads to imperfections. Two of them are studied
in this paper: IQ imbalance and sampling clock offset (SCO). Based on a
mathematical system model, we study analytically the impact of the two
imperfections on the sensing performance of the energy detector and of the
cyclostationarity detector. We show that the IQ imbalance does not impact
the performance of the two detectors, and that the SCO only impacts sig-
nicantly the performance of the cyclostationarity detector.
I. INTRODUCTION
In the eld of wireless communications, each device uses a
certain frequency band in order to transmit and receive infor-
mation. Actors involved make a request to federal agencies to
be able to legally communicate over a limited band. This leads
to a shortage of available spectrum and slows down the devel-
opment of other innovative applications. However, it has been
noticed that, in practice, the spectrum use is less than 10% [1],
which leads directly to the interesting idea to reuse the spectrum
by a secondary network when it is not requested by the primary
user ([2] and [3]). Such devices are called cognitive radios be-
cause they are aware of the spectrum availability that surrounds
it. First, they must investigate a certain range of frequencies in
order to discover potential opportunities (that is, free frequency
bands). Then, the devices communicate in the licensed bands,
while avoiding interference with the primary networks.
A standard terminal consists of an analog front-end (for con-
tinuous signal processing as frequency change, ltering and am-
plication) and a digital processor (allowing demodulation and
synchronization algorithms). The design of the analog front-end
is critical in the case of cognitive radios. Indeed, to analyze a
large number of frequencies, they have to work on a large spec-
trum composed of several bands. In addition, they must be able
to detect signals from close and distant devices together. Cogni-
tive radios must therefore cope with high power signal dynamics
([4] and [5]). These two constraints lead to different imperfec-
tions in the implementation of the analog front-end, such as the
IQ imbalance caused by different physical analog paths on the
I and Q branches if a direct conversion receiver is used and the
sampling clock offset (SCO) caused by the non-ideal local os-
cillators.
In this article, we will study the impact of IQ imbalance and
SCO on the sensing performance. In order to do this, we rst
introduce a new system model which includes the two imperfec-
tions. Then, we summarize the sensing algorithms in the case of
the energy detector and the cyclostationarity detector. An an-
alytical study of the impact of each imperfection is performed
and the results are supported with Matlab
c
simulations.
II. SYSTEM MODEL
During the sensing phase, the cognitive radio has to decide
between two hypotheses:
H
0
: there is a communication signal in the bandwidth;
H
1
: no signal is present in the bandwidth.
Fig. 1 illustrates the overall system when a linearly modulated
communication signal is present in the bandwidth (H
1
).
cos(t)
sin(t)
Re
Im
I
n
QAM
(T
symb
)
Transmitter
u(t)
nT
s
(1+)
nT
s
(1+)
j
y
R
(t)
y
I
(t)
Sensing
(1+) cos(t+)
(1) cos(t)
y
n
f(t)
f(t)
Receiver
y(t)
n
RF
(t)
c
RF
(t)
Fig. 1. Bandpass architecture with front-end imperfections.
At the transmitter, independent QAM modulated symbols I
n
of variance
2
I
are transmitted at the rate 1/T
symb
. The sequence
is low-pass ltered by u(t) in order to remove the out-of-band
components (typically a halfroot Nyquist lter is used). The
resulting baseband signal, given by:
s(t) =

n
I
n
u(t nT
symb
), (1)
has a variance equal to
2
s
=
2
I

|u(t)|
2
dt. The radio-
frequency signal s
RF
(t) is obtained after up-conversion to the
carrier pulsation (corresponding to the carrier frequency f)
and transmitted through a frequency selective channel c
RF
(t).
Additive white Gaussian noise n
RF
(t), of one-sided power spec-
tral density (PSD) equal to N
0
, corrupts the received signal.
At the receiver, the RF signal is down-converted to the base-
band domain for complex operation and low-pass ltered by
f(t) in order to limit the observation to the desired bandwidth
B
f
. The baseband signal y(t) can be expressed as:
y(t) = s(t) +n(t) (2)
where n(t) is the baseband equivalent noise of variance

2
n
= 2N
0
B
f
. After sampling at the rate 1/T
s
(ofter a multiple
of the symbol rate 1/T
symb
), the obtained received sequence y
n
is used for the spectrum sensing.
When there is only noise in the system (H
0
), the baseband
signal y(t) is given by:
y(t) = n(t) (3)
Because the implementation of the analog front-end is dif-
cult, non-idealities are introduced during the RF-to-baseband
conversion. First, the IQ imbalance is due to the difference of
physical length between the I and Q branches of the receiver.
It is represented by the errors and , respectively changing
the amplitude and the phase of the local oscillators. Second, the
SCO is a difference between the transmitter and receiver symbol
rate reference T
symb
. It is represented by the error changing the
sampling frequency.
III. SENSING ALGORITHMS
The two sensing algorithms considered in this paper rely on
time averages dened as:
x(t) = lim
t
t
2

t
2
x(t)dt (4)
where t is the observation window, nite in practice.
The energy detector estimates the variance
2
y
of the received
signal y(t) given by the integral of the PSD over the frequency
bandwidth. If the signal is ergodic, the variance can be estimated
based on time averages [6]:

2
y
= y(t)y

(t) (5)
The cyclostationarity detector estimates the spectral correla-
tion density (SCD) dened as the Fourier transform of the cyclic
autocorrelation function [7]. If the signal is cyclo-ergodic, the
SCD at cycle frequency c (equal to the inverse of T
symb
in our
system) can be evaluated based on time averages [7]:
S
c
y
(f) = lim
T
1
T

Y
T
(t, f +
c
2
)Y

T
(t, f
c
2
)

(6)
where Y
T
(t, f) =

t+
T
2
t
T
2
y(u)e
j2ct
du is the Fourier transform
of y(t) on an interval T centered on t. In practice, algorithms
will work with nite T (inverse of the frequency resolution f)
and t (temporal window).
To compare the performance of the two detectors, we illus-
trate the PSD (in case of the energy detector) and the SCD (in
case of the cyclostationarity detector) of the received baseband
signal when the channel is ideal (frequency at). As we can see
on Fig. 2, the PSD is a stronger metric than the SCD explaining
why the energy detector usually outperforms the cyclostationar-
ity detector.
This is conrmed by evaluating the probability of misdetec-
tion (p
MD
) and the probability of false alarm (p
FA
), dened as:
p
MD
= Prob {decision is H
0
|H
1
} (7)
p
FA
= Prob {decision is H
1
|H
0
} (8)
Fig. 3 illustrates the probability of misdetection as a function of
the signal-to-noise ratio (SNR) for a xed probability of false
alarm equal to 0.1. Clearly the energy detector outperforms the
cyclostationarity detector because it collects more signal energy.
However we have to keep in mind that it suffers from the noise
uncertainty that makes it impractical at low SNR [8].
Fig. 2. Power spectral density versus spectral correlation density.
Fig. 3. Probability of misdetection with respect to the SNR (p
FA
= 0.1).
IV. IMPACT OF IQ IMBALANCE ON PERFORMANCE
When there is IQ imbalance in the system, the baseband re-
ceived signal y(t) is expressed as a linear combination of the
ideal signal y(t) and its conjugate y

(t) as follows [9]:


y(t) = y(t) +y

(t) (9)
with:
= cos () j sin () (10)
= cos () +j sin () (11)
A. Impact on energy detector
The energy detector evaluates the signal variance:

2
y
= E[y(t)y

(t)] (12)
= E

[y(t) +y

(t)][y(t) +y

(t)]

(13)
=

||
2
+||
2

2
y
+ 2Re

E[y(t)y(t)]

(14)
where only the second term depends on the hypotheses. If there
is only the noise in the system, the expectation in the second
term is equal to:
E[y(t)y(t)] = E[n(t)n(t)] (15)
which is null because the real and imaginary parts of the noise
are uncorrelated. If there is only the signal in the system, the
expectation in the second term is equal to:
E[s(t)s(t)] =

E[I
n
I
n
]u(t nT)u(t n

T) (16)
which is null because the real and imaginary parts of the sym-
bols are uncorrelated.
Therefore, the estimated metric becomes:

2
y
=

| |
2
+| |
2


2
n
if H
0
(17)

2
y
=

| |
2
+| |
2

2
s
+
2
n

if H
1
(18)
and it is possible to derive the probability of misdetection and
the probability of false alarm.
Since the metric is affected by the same power gain
| |
2
+| |
2
under the two hypotheses, the probability of
false alarm and the probability of misdetection are unchanged
due to the IQ imbalance if the detector threshold is multiplied
by the same power gain.
B. Impact on cyclostationarity detector
Since the cyclostationarity detector evaluates the SCD of the
received signal, let us develop its expression in presence of IQ
imbalance to nd out how the detector performances will be af-
fected. The expression (6) of the SCD becomes:
S
c
y
(f) = lim
T
1
T

Y
T

t, f +
c
2

t, f
c
2

(19)
where:
Y
T
(t, f) =

t+
T
2
t
T
2
y (u) e
j2fu
du (20)
With the properties of linearity of the Fourier transform, we ob-
tain Y
T
(t, f) = Y
T
(t, f) + Y

T
(t, f) since we know
that if g(t) has the Fourier transform G(f), then g

(t) has the


Fourier transform G

(f). The estimated SCD becomes:


S
c
y
(f) = | |
2
S
c
y
(f) +| |
2
S
c
y
(f) + 2Re

S
c
yy
(f)

(21)
where:
S
c
yy
(f) lim
T
1
T

Y
T

t, f +
c
2

Y
T

t, f +
c
2

(22)
The rst two terms in the expression of S
c
y
(f) represent the
SCD of the ideal signal y(t) and the mirrored SCD. The third
term depends on the hypotheses.
Let us rst assume that there is only noise in the system:
y(t) = n(t). The rst two terms are the SCD of Gaussian noise,
and are therefore null by denition [7]. The term S
c
yy
(f) repre-
sents the spectral correlation between the points
c
2
f when f is
varying. The only value of f where this spectral correlation can
be different from zero is f = 0 (correlation between the same
two points in the spectrum). We get:
S
c
yy
(0) = lim
T
1
T

Y
T

t,
c
2

Y
T

t,
c
2

(23)
= lim
T
1
T

Y
R
T

t,
c
2

Y
I
T

t,
c
2

(24)
where Y
R
T

t,
c
2

and Y
I
T

t,
c
2

are respectively the real and


imaginary parts of Y
T

t,
c
2

In the last expression, we omitted the term:


2 j Y
R
T

t,
c
2

Y
I
T

t,
c
2

(25)
because it disappears when we later take the real part of the
whole expression. Since the two terms have the same variance,
they will compensate for each other, eventually canceling the
whole integral. As a conclusion, the three terms in (21) are null
when there is only noise in the system, so that the SCD S
c
y
(f)
is equal to zero.
Let us secondly assume that y(t) only contains the signal:
y (t) = s (t). The rst two terms of (21) are different from zero
for certain values of c by denition [7]. In the third term, the
only values of f where the spectral correlation S
c
yy
(f) can be
different from zero are f = 0 (correlation between the same
two points in the spectrum) and f =
1
2T
symb
(since two points
separated by a cyclic frequency c =
1
T
symb
are correlated). The
term vanishes for the case f = 0 for the same reasons as for
the noise. The term also vanishes for the case f =
1
2T
symb
be-
cause it is the correlation between the frequencies 0 and 1/T
symb
where the signal is null. Thus, the third term is null and we get:
S
c
y
(f) = | |
2
S
c
s
(f)+ | |
2
S
c
s
(f) where S
c
s
(f) is the SCD
of the signal s(t).
The resulting values of the metric are:
S
c
y
(f) = 0 if H
0
(26)
S
c
y
(f) = | |
2
S
c
s
(f) + | |
2
S
c
s
(f) if H
1
(27)
Fig. 4 illustrates the expression (27) for the ideal channel and
for a two-paths time dispersive channel (the second path being
delayed by 3 times the symbol period and attenuated by a fac-
tor 0.1). When the channel is ideal, the SCD is only multiplied
by the power gain . When the channel is time dispersive, the
form of the SCD may be signicantly modied. However, the
metric used for detection is the integral of the SCD which is
also only affected by the power gain , even when the channel
is time dispersive. Therefore, the probability of false alarm and
the probability of misdetection are unchanged due to the IQ im-
balance if the detector threshold is multiplied by the same power
gain.
In conclusion, the IQ imbalance has no impact on the perfor-
mance of both detectors.
V. SAMPLING CLOCK OFFSET (SCO)
In the baseband equivalent architecture, the SCO is modeled
by a factor 1 + that multiplies the sampling rate T
s
.
A. Impact on energy detector
Since the variance of the received signal samples does not de-
pend on the sampling rate (it only depends on the receiver lter
bandwidth), the metric of the energy detector is not modied:

2
y
=
2
n
if H
0
(28)

2
y
=
2
s
+
2
n
if H
1
(29)
The probability of false alarm and the probability of misdetec-
tion are therefore also unchanged in the presence of SCO.
Fig. 4. SCD module for the two possible situations in the case of IQ imbalance.
B. Impact on cyclostationarity detector
Since the cyclostationarity detector estimates the SCD of the
received signal, let us develop its expression in the presence of
SCO to nd out how the detector performances will be affected.
In order to assess the impact of the SCO correctly, we need to
take the nite observation window into account. We assume
therefore that the received signal y(t) results from the multipli-
cation of an innite signal x(t) with a nite rectangular window
w
T
(t) dened as:
w
T
(t) =

1
T
2
t
T
2
0 else
(30)
As we know, a product in the temporal domain corresponds to a
convolution in the frequency domain:
Y
T
(t, f) = X (t, f) W
T
(f) (31)
=

X (t, f s) W
T
(s) ds (32)
where W
T
(f) is the Fourier transform of w
T
(t) given by
W
T
(f) = T
sin(Tf)
Tf
. The principle is illustrated in Fig. 5.
We rst derive the SCD in the ideal case, then we add the
SCO effect. By using the denition (6), we obtain:
S
c
y
(f) =

S
c

f
s +s


W
T
(s) W

T
(s

) ds ds

(33)
where c

= c (s s

) and S
c
x
(f) is the SCD of the innite
signal x(t).
The term S
c

f
s+s

is different from zero when c

= c
or when c

= 0. Therefore, we have:
S
c

f
s +s

= S
c

f
s +s

(s s

)
+S
c

f
s +s

(s s

c) (34)
Fig. 5. Windowing of x(t).
and:
S
c
y
(f) =

S
c
x
(f s) | W
T
(s) |
2
ds
+

S
0
x

f +
c
2
s

W
T
(s) W

T
(s c) ds (35)
= S
c
x
(f) | W
T
(f) |
2
+

S
0
x

f +
c
2
s

W
T
(s) W

T
(s c) ds (36)
The rst term corresponds to the convolution between the
SCD of the innite signal x(t) and the square modulus of the
frequency response of the rectangular window w
T
(t). Since
T T
symb
and thus
1
T
c, the product W
T
(s) W

T
(s c)
in the second term is negligible because the value of W
T
(f)
tends to zero when f
1
T
.
If we want to interpret the effect of the SCO, we have to con-
sider the following SCD:
S
c(1+)
y
(f) =

S
c

f
s +s


W
T
(s) W

T
(s

) ds ds

(37)
where c

= c

(1 + ). The term S
c

f
s+s

is different
from zero when c

= c or when c

= 0. Therefore, we have:
S
c(1+)
y
(f) (38)
=

S
c
x

f s +
c
2

W
T
(s) W
T
(s c) ds (39)
+

S
0
x

f s +
c
2
(1 +)

W
T
(s) W

T
[s c (1 +)] ds
(40)
= S
c
x

f +
c
2

[W
T
(f) W

T
(f c)] (41)
+

S
0
x

f s +
c
2
(1 +)

W
T
(s) W

T
[s c (1 +)] ds
(42)
In conclusion, the SCD will decrease when or T is grow-
ing. We know that the longer the sampling vector, the better
SCD estimate. But now we also know that if T is growing, the
SCO effect will increase. It automatically implies a tradeoff that
we have to make in practice between computing a precise SCD
and avoiding too much SCO effect on it. If we draw the SCD,
we see on Fig. 6 that there is a loss of performance (identical in
both AWGN and dispersive channel) when T is growing (as ex-
pected). Thus, the channel nature does not change the effect of
the SCO. We also observe that performance is decreasing when
is increasing (see Fig. 7). In both two graphs, the p
MD
grows
as is increasing (meaning that the performance is decreasing).
If we increase the window length, the performance loss is more
important (see Fig. 7b).
In conclusion, the SCO does not affect the energy detector
performance but the cyclostationarity detector is less efcient
when and T are growing.
Fig. 6. SCD module for the two possible situations in the case of SCO.
VI. CONCLUSION
This paper evaluates the impact of two analog front-end im-
perfections, IQ imbalance and sample clock offset (SCO), on
the performance of the spectrum sensing. We compare the sen-
sitivity of the energy detector and of the cyclostationarity de-
tector. It is shown that IQ imbalance does not cause any per-
formance degradation so that the low-cost direct conversion ar-
chitecture incurring IQ imbalance can be recommended for the
Fig. 7. Probability of misdetection in the case of SCO ( and T are varying). T
is 10 times bigger in b) than in a).
spectrum sensing. While the energy detector is not sensitive to
SCO, the cyclostationarity detector suffers from a high perfor-
mance degradation when there is SCO in the system. Therefore
high quality of the local oscillators should be integrated on the
device when the cyclostationarity detector is used for spectrum
sensing.
REFERENCES
[1] Federal Communications Commission, Facilitating opportunities for ex-
ible, efcient, and reliable spectrum use employing cognitive radio tech-
nologies, in FCC-03-322, 2003.
[2] F. K. Jondral T. A. Weiss, Spectrum pooling: An innovative strategy for
the enhancement of spectrum efciency, IEEE Radio Communications,
vol. 42, no. 3, pp. S8S14, March 2004.
[3] I. F. Akyildiz W. Y. Lee M. C. Vuran S. Mohanty, Next generation /
dynamic spectrum access / cognitive radio wireless networks: A survey,
Computer Networks Journal (Elsevier), vol. 50, pp. 21272159, September
2006.
[4] R. W. Brodersen D. Cabric, S. M. Mishra, Implementation issues in spec-
trum sensing for cognitive radios, in Asilomar Conference on Signals,
Systems, and Computers, 2004.
[5] B. Natarajan H. Zamat, Use of dedicated broadband sensing receiver in
cognitive radio, in IEEE Proceedings of ICC, June 2008.
[6] Lee D. Davisson Robert M. Gray, An Introduction to Statistical Signal
Processing, Cambridge University Press, 2004.
[7] WilliamA. Gardner, Statistical Spectral Analysis: A Non-Probabilistic The-
ory, Prentice Hall Information and System Sciences Series, 1988.
[8] C.M. Spooner W.A. Gardner, Signal interception: performance advantages
of cyclic-feature detectors, IEEE Transactions on Communications, vol.
40, no. 1, pp. 149159, January 1992.
[9] Andr e Bourdoux Francois Horlin, Digital Compensation for Analog Front-
Ends: A New Approach to Wireless Transceiver Design, Wiley, 2008.

You might also like