Sie sind auf Seite 1von 7

1

Adaptive Non-Data-Aided Compensation for I/Q Mismatch


and Frequency Offset in Low-IF Receivers
Shafayat Abrar1 , Azzedine Zerguine2 and Asoke Nandi3
1 Department

of Electrical Engineering, COMSATS Institute of Information Technology, Islamabad 44000


of Electrical Engineering, King Fahd University of Petroleum and Minerals, Dhahran 31261
3 Department of Electronic and Computer Engineering, Brunel University, Uxbridge UB8 3PH

2 Department

AbstractIn this work, we present an adaptive non-data-aided


compensator for the in-phase/quadrature-phase (I/Q) mismatch in
low-intermediate frequency (heterodyne) receivers. In particular,
the adaptive I/Q mismatch algorithm is derived by exploiting the
uncorrelatedness between the desired and the image signals, and
is used to compensate for gain imbalance, phase-offset error and
frequency-offset error.

cos(2 fLO t)

Analogue
Processing

Digital
Processing

xn
LPF

LPF

A/D

rn

rRF(t)
LPF

Index TermsI/Q imbalance, adaptive filter

LPF

A/D

exp(+i2 fIF nT)

I. I NTRODUCTION
Due to components mismatches in analogue electronics and
resulting in-phase/quadrature-phase (I/Q) imbalances, the performance of a heterodyne receiver may degrade significantly [1].
To compensate these imbalances and remove image signal from
the desired band, statistical independence based adaptive methods
were introduced (refer to [2] and references therein). Here, in
this work, we discuss a simple algorithm for non-data-aided
I/Q compensation which is derived by exploiting the correlation
properties of desired and image signals.
The paper is organized as follows. In Section II, a general
signal model for an imbalanced analog front-end is presented.
In Section III, a cost function sensitive to imbalance is proposed
and optimized to yield into an adaptive algorithm for imbalance
compensation. In Section IV, an iterative method is derived for the
estimation of frequency-offset. Simulation results are presented in
Section V and conclusion are drawn in Section VI.
II. S YSTEM M ODEL
The system model is shown in Fig. 1. In the absence of additive
noise, the received RF signal is given by
(1)

where fC is the carrier frequency of desired band, fI = 2fLO fC


is the central frequency of the image band, and s(t) and q(t)
are the baseband desired and image signals, respectively. The
quadrature mixture is assumed to suffer with amplitude mismatch
, frequency mismatch f = fC fLO fIF and phase
mismatch . The impairments and are assumed to be frequency
independent. The digital IF signal rn , sampled at the rate of 1/T ,
is expressed as:
rn = (sn +

qn ) e+i2n(fIF T +)

+ (qn + sn ) ei2n(fIF T +) ,
where



= 0.5 1 (1 + )ei ,


= 0.5 1 + (1 + )ei ,

zn

i
(1+ ) sin(2 fLO t+ )

rRF (t) = 2{s(t)ei2fC t } + 2{q(t)ei2fI t },

exp( i2 fIF nT)

(2)

(3a)
(3b)

xn
zn

Adaptive
gain/phase
mismatch
removal

Fig. 1.
sators.

~
sn
~
qn

Iterative
frequency
mismatch
removal

sn

fLO is local oscillator freq.


fIF is desired IF freq.
T is symbol period.

qn

System model of a heterodyne receiver and mismatch compen-

= T f , sn = s(t)|t=nT and qn = q(t)|t=nT . After downconversion and low-pass filtering, we obtain the baseband signals
{xn } and {zn }:
xn = (sn + qn ) e+i2n ,

(4a)

zn = (qn + sn ) ei2n .

(4b)

The digital processor at receiver uses the baseband signals {xn }


and {zn } to estimate the impairments, , and . Assuming
that impairments are perfectly known, then the desired and image
signals are expressed as


1+
sbn = e
sn ei2n = 1||
(xn zn ) ei2n ,
(5a)
2



+i2n
+i2n
1+
,
(5b)
qbn = qen e
= 1||2 (zn xn ) e
where = / . The impairments and are related to , we
can show:




1
1
1.
= angle
(6)
, and =
1+
1+
III. E STIMATION OF G AIN I MBALANCE

Exploiting the fact that the desired and image signals {sn } and
{qn } are mutually uncorrelated, optimum estimates were obtained
in [3] as follows:
p
B B 2 4|A|2
(1)
,
(7a)
opt =

p2A
2
2
B + B 4|A|
(2)
opt =
,
(7b)
2A


where A := Exn zn , and B := E |xn |2 + |zn |2 . For vanishing
(1)
(2)
imbalance, i.e., A 0, we have opt 0 and opt .
Note that authors in [3] preferred to use the root with smaller
(1)
magnitude, i.e., opt .
In this work, we propose to obtain the value of adaptively by
minimizing a cost which is measure of the correlation between
the estimated signals, sbn and qbn , mathematically it is expressed
as
2



= arg min E (xn zn ) (zn xn ) ,
(8)

Note that this cost is insensitive to frequency offset error, , which


facilitates separate estimation of . To obtain a gradient-based
adaptive algorithm for , we use


bn = b
n1 |C|2 ,
(9)

for = n1 and C := E(xn b


n1 zn )(b
n1 xn zn ). Note that
the auxiliary variable C can be expressed as C = A B
bn1 +
2
A
bn1
, where A and B are as specified in (7); next, we find
|C|2 C
|C|2
=
= C (2A
bn1 B) ,
b
n1
C b
n1

where N denotes number of symbols [4]. Note that this estimator


assumes that the signal has constant modulus; in the presence of
gain imbalance, however, we would need gain normalization to
b n1 , and assuming
ensure this property. Denoting n :=
no additive noise, note that
!m

m
b
b

sbn1
sbn
se ei2(n1)n2 sen ei2nn1
= n1
|b
sn1 | |b
sn |
|b
sn1 | |b
sn |

m
sn1 ei2(n1)n1 sn ei2nn
= ei2m(n1 +n(n n1 )) ,

(15)
Further assuming n n1 , we obtain

m 

sbn1
sn
b
1
n
angle
,
2m
|b
sn1 | |b
sn |

b n is obtained as
With the aid of (16), an iterative estimate of

Replacing the statistics A, B and C with their respective estimates, we get the following gradient-based algorithm:

Bn = g Bn1 + (1 g ) |xn |2 + |zn |2 ,
2
Cn = An Bn b
n1 + An
bn1
,

where

bn = b
n1 + g Cn (Bn

(11)

2An
bn1
),

where g is a positive step-size and 0 < g < 1 is a forgetting(2)


(1)
factor. Substituting An and Bn in (7), we obtain n1 and n1
(1)
(2)
(1)
as the estimates of opt and opt , respectively. Using n1 and
(2)
n1 , we can express (11) as follows:
(2) 
(1) 
bn1 n1

bn =
bn1 b
n1 n1

(2) 
(1)
(12)
+ n1


bn1 n1
,
2
(1)

(2)

where = 2|A|2 and 0.5(n1 + n1 ) is the estimate of


saddle point (see Fig. 2(a)). This implies that, depending on
(2)
(1)
initialization, the update may converge either to opt and opt .
Under no imbalance condition, however, as one of the roots is
required to be zero, the update has a natural tendency to converge
to the root with smaller magnitude provided that the step-size is
large enough to help escape the other root (see Fig. 2(b)-(c)).
IV. E STIMATION OF F REQUENCY O FFSET
A. For PSK Signals:

The presence of frequency-offset error contaminates the estib n1 is the


mated signal sbn by the factor e+i2n . Suppose
available estimate of , then sbn is expressed as
b

sbn = sen ei2nn1

1+
bn1
b
=
(xn
bn1 zn ) ei2nn1 ,
1 |b
n1 |2

(13)

If b
sn is an m-PSK, then the maximum likelihood approach
estimates , as given by,
m
 PN

snk b
sn1k
angle
k=0 b
b
,
(14)
=
2m

m

sbn1
sbn
,
|b
sn1 sbn |
n o
bn
angle
+ (1 o )
,
2m

b n = d
b n1 + (1 d )

(10)

An = g An1 + (1 g ) xn zn ,

(16)

b n1
b n = o

and

(17)

are positive forgetting factors.

B. For QAM Signals:


The estimator (17) is not useful for frequency-offset estimation
in quadrature amplitude modulation due to its multi-modulus
constellation. Assuming that the gain imbalance has been compensated and denoting = 2n, we have
sbn = e
sn ei = sn ei()
b

(18)

b we can show that the fourth-order


Denoting e := ,
statistics of sbn contains the information of unknown e

 1

E sb4n,I + sb4n,Q = E s4n,I + s4n,Q 6s2n,I s2n,Q cos(4e )
4
+constant
(19)
Note that cos(4e ) is maximum (that is equal to +1) when e =
0 and it is minimum (that is equal to 1) when e = /4.
So the unknown phase is compensated if it is between /4 and
+/4. For phase ambiguity due to the multiples of 90 degree may
be compensated using differential encoding. Further note that, for
QAM signals, E s4n,I + s4n,Q 6s2n,I s2n,Q is a negative quantity
which helps us formulate minimization of the following cost for
the recovery of unknown phase:
4
4
min E sbn,I
+ sbn,Q

(20)

b + sen,Q sin
b and b
Notice that sbn,I = [b
sn ] = sen,I cos
sn,Q =
b + sen,Q cos ,
b these relations help us obtain
[b
sn ] = e
sn,I sin
the following:

3
Eb
s 4 = +4Eb
sn,I
sbn,Q ,
b n,I

3
Eb
s 4 = 4Eb
sn,Q
sbn,I ,
b n,Q

(21a)
(21b)

These statistics may be computed iteratively and lead to the


following gradient-based algorithm:
3
sbn,I ,
Gn = t Gn1 + (1 t ) sbn,Q

3
Hn = t Hn1 + (1 t ) sbn,I
sbn,Q ,
b
b
n = n1 + t n , (n := Gn Hn ),

Taking the z-transform of (27), we get


1
b
b
(z)
= (z)z
+ 2(z)z 1 + t (z),

Combining the two expressions in (28), we obtain


(22)

where t is a positive step-size and t is a positive forgettingfactor less that one. Note that the algorithm (22) does not (explicity) exploit the fact that = 2n . Exploiting this information,
we modify the problem (20) as follows:

2 s (z)z 1
1
b
b
(z)
= (z)z
+ t (z) +
1 z 1

J := min E
b ,
b

4
sn,Q
b

b = 2n
b
, s.t.

(23)

b is known, a fine estimate of


b n is
bn =
b n /(2n);
Once
b
however, in practice, the n is not explicitly required to be
b n is equivalently sufficient for
computed as the knowledge of
b n = n , where the statistical
the purpose. Note that J /
error quantity n is as specified in (27). The derivative of cost J
w.r.t. n requires attention; note that
(25)

bn
b n1 +
The constraint in (23) allows us to express
2n1 , which gives
bn
b n1
b0

+ 2
+ 2n 2n,
n1
n1
n1

(26)

b
Note that the gradient /
is growing linearly in time which
is analytically correct but its use in the update expression may
cause divergence. One possible way to handle this situation is
to use a diminishing step-size to overcome the linear growth of
b
/.
However, a diminishing step size usually leads to slow
convergence and requires exhaustive experimentation to determine
how rapidly the step-size must decrease in order to prevent
scenarios in which it (the step-size) becomes too small when the
iterates are far from the required estimate. The other solution is
to simply drop this gradient factor as it is always positive and
has no role in determining the direction of the update. We prefer
to adopt the latter proposal while using a fixed but very small
step-size s for n to ensure the stability and low jitter.
bn =
b n1 + 2n1 + t n ,

n = n1 + s n ,

n = n1 + n1 ,

The optimization of (23) may be realized as separate minib and ;


b however, the resulting two
mizations with respect to
updates must also satisfy the constraint in (23). To realize such an
optimization, we introduce an auxiliary (or intermediate) variable
n , minimize the cost w.r.t. it, and obtain a coarse (but gradientb Since the relation = 2n can
based adaptive) estimate of .
equivalently be expressed as n = n1 +2, where n is the
true value of at time n. With these considerations, we suggest
to solve
bn =
b n1 + 2n1 t J ,

bn
4
(24)
s J
,
n = n1
bn
4

b J
J

,
=
b

(29)

Denoting (z) := (z)z 1 /(1 z 1 ), we obtain an alternate


form of (28) as follows:
o

4
sbn,I

(28)

(z) = (z)z 1 + s (z),

(27)

bn =
b n1 + t n + 2 s n ,

Experimentally, we have found that a suitable value of


to the square of t , i.e.,

(30)
s

is close

( t ) .

V. S IMULATION R ESULTS
We carry out simulations to evaluate the performance of the
proposed estimators. The baseband signals in the desired and
image bands are expressed as sn = an + wn and qn = bn + vn ,
respectively, where {an } and {bn } are transmitted quadrature
phase-shift keying (QPSK) symbols, and {wn } and {vn } denote
additive white Gaussian noise. The signal-to-noise ratios (SNRs)
of the received signals {sn } and {qn } are taken as 30 dB. The
forgetting factors were selected as g = d = o = 0.998
and the step-size = 6 104 . At time zero, adaptive/iterative
b 0 = 1,
parameters were initialized as A0 = 1, B0 = 2,
b0 = 0,
b 0 = 0. The frequency offset = 1 104 , the amplitude
and
mismatch = 0.8, and the phase mismatch = 10 (this gives
= 0.3863 i0.1563, = 1.3863 i0.1563 resulting in
(1)
(2)
opt = 0.2877 i0.0803 and opt = 3.2245 i0.8999).
Experiment 1: In this experiment, we study convergence
behaviour of update (11) for small and relatively large step-sizes
(for QPSK signal).
Refer to Fig. 2(a) for the contour plot of the cost where the
(1)
(2)
(1)
global minima, opt and opt , and the saddle point 0.5(opt +
(2)
opt ) are labeled. Next in Fig. 2(a) and (b), we provide traces of
convergence for small and relatively large step sizes, respectively.
It can be noticed that for small step-size (i.e., g = 5 105 ),
(2)
(2)
when
bn is initialized near opt , it converged to opt ; however,
4
for relatively large step-size (i.e., g = 1 10 ), regardless of
the initialization,
bn is found to be always converging to the root
(1)
with smaller magnitude, i.e., opt .
4
Further, with g = 1 10 , refer to Fig. 3(a)-(d) and Fig.
3(e)-(f) for scatter plots and convergence traces, respectively;
both estimators can be noticed to be converging steadily to
true values. Refer to Fig. 3(g) for the traces of empirically
obtained mean square error E|b
sn sn |2 and squared absolute
2
correlation |Eb
sn qbn | . Both indices are decreasing along iteration
and attaining a stable floor in steady-state; this means that, as a
result of successful convergence, estimated signal sbn is getting
close to desired signal sn and image qbn is rejected from sbn .
Note that 1000 symbol points are used in each scatter plot (for a
single realization) and traces (in Fig. 3) were averaged over 500
independent realizations.
Experiment 2:

VI. C ONCLUSIONS

(a) sn

In this work, an adaptive non-data aided in-phase / quadraturephase imbalance compensator for heterodyne receiver was developed. Simulation results showed that the proposed adaptive
scheme can successfully compensate for frequency-independent
imbalances.

(2)
opt

0
1
1

o and + indicate saddle point and minima, resp.

(c) sen

(1)
opt

(b) xn
2

2
2

(d) sbn

[]

o
+

1
(1)
0.5(opt

[]

= 5x10

(2)
opt )

1
1

(e) |b
n |

x 10

0.3

and Iterations = 30000

0
4

bn
(f )

0.2

0.5
0.1

Simulated
Simulated
True value
True value
0
2000 4000 6000
2000 4000 6000 0
(g) MSE and SC traces

[]

0
0

10

3
4

= 1x10

[]

and Iterations = 8000

[dB]

|Eb
sn qbn |2
E|b
sn sn |2

10
20

30

[]

1000

2000

3000

4000

5000

6000

Iterations

Fig. 3.

2
4

[]

Scatter plots and convergence traces for QPSK.

(a) sn

Fig. 2. (a) Contour plot of cost function for = 0.8 and = 10 ,


and (b)-(c) convergence trajectories of
bn for small and relatively large
step-sizes.

0
1
1

R EFERENCES
[1] S. Mirabbasi and K. Martin, Classical and modern receiver architectures, IEEE Commun. Mag., vol. 38, pp. 132139, Nov. 2000.
[2] M. Valkama, M. Renfors, and V. Koivunen, Advanced methods for
I/Q imbalance compensation in communication receivers, IEEE Trans.
Signal Process., vol. 49, no. 10, pp. 23352344, Oct. 2001.
[3] G.-T. Gil, Y.-D. Kim and Y.H. Lee, Non-data-aided approach to
I/Q mismatch compensation in low-IF receivers, IEEE Trans. Signal
Process., vol.55, no.7, pp.33603365, July 2007.
[4] U. Mengali, Synchronization Techniques for Digital Receivers,
Springer 1997.

(b) xn
2

(c) sen

1
0

(d) sbn

1
1

Fig. 4.

2
2

1
1

Scatter plots for 8PSK.

VII. ACQUISITION A BILITY OF DMD S YNCHRONIZER


A. Steady-state mean square deviation
A good measure for the performance of synchronizer is the
steady-state error deviation (or variance). We define mean square

2
2
deviation as MSD =
= E[
] (rad2 ) and compute it for the
three algorithms. Let k = k be the parameter error at time
instant k. Using a generalized form of adaptive phase estimator,

(a) sn

where Ez is a nonlinear function of zk . Squaring and averaging


the update (31), we get
 2 
 
 
(32)
E k+1
= E k2 + 2 E Ez2 2 E [k Ez ]

(b) xn

4
5

2
0

2
2
where Ez = (zk,I
zk,R
)zk,I zk,R . Using the following approximations for some angle 1: sin() , and cos()
1 0.52 , and some simple algebra, it is possible to show that


E Ez2 |k c2 k2 + c3 ,
(33)
c1
E [Ez |k ] k ,
2

4
4 2

(c) sen

(d) sbn

4
2

which make it further possible to rewrite (32) as


 2 
 
E k+1
AE k2 + B,
.
where A = 1 c1 + c2 2 ,
.
and B = c3 2 ,

2
5
5

Fig. 5.

4
4 2

Scatter plots for 16QAM.

while parameters c1 , c2 and c3 are obtained as follows:




.
c1 = 4E 3x2I x2R x4R ,


.
c2 = 2E x8R 28x6R x2I + 35x4R x4I ,


.
c3 = 2E x6R x2I x4R x4I ,

Frequencyoffset estimates

1.5

x 10

1
0.5

Upon
successful
convergence,
limk E[k2 ] is true, which yields

0
0.5

True = 10
b n for t = 0.996

b n for t = 0.999

1
1.5
0

1000

2000
3000
Iterations

4000

5000

bn =
b n /(2n) for 16QAM;
b n is
Fig. 6. Convergence traces:
estimated as specified in (22). Smaller values of t can ensure shorter
settling time while causing relatively large over/undershoot.

10

2
b n/ 1
NMSE: E

10

6000

8000

10000

Iterations

Fig. 7.

Frequency-offset recovery: normalized MSE traces for 16QAM.

we can obtain
k+1 = k+1 = k Ez

(36)

By rearranging Equation (36), we can obtain the value of loopgain for the required MSD and given signal statistics:
=

c1 MSD
c3 + c2 MSD

(37)

(38)

where MSD is obtained from (36), while and are obtained


as = ( 0 )2 and = ln (A). The is the initial value
of MSD and determines the convergence speed of the adaptive
synchronizer (in the mean square error sense). Similarly, we can
obtain

40

4000

B. Loop gain for given MSD

 
E k2 = MSD + ( MSD) ek

2000

(35c)

We denote E[k2 ] as the (instantaneous) mean square deviation


at time index k. From the expression (34), it is easily verified that
we have,

30

50
0

(35b)

C. Instantaneous MSD and Phase-Error Trajectories

Adaptive Frequency Offset Recovery


16QAM : = 0.8, = 10 ,
= 1 104 , SNR = 30 dB,
t = 0.98, g = 0.998, g =1104

20

c3
B
=
(rad2 )
1A
c1 c2

(35a)

Equation (37) is very useful for simulation study that we dont


need to make trial-and-error guesses to obtain proper values of
loop-gains for fair comparison.

One-Step
Solution
t=
210 4
.
.
1-step

Two-Step
Solution
t = 1.510 4
s = 2t
2-step

2
MSD =
=

2
limk E[k+1
]

(34)

(31)

E [k ] = ( 0 )ek
(39)

c
where = ln 1 2 1 and it helps us to readily obtain


E [k ] = 1 ek + 0 ek
(40)

D. Convergence time for given target


We define the convergence time as the number of iterations, K,
that is needed for its E[k2 ] to reach (1 + ) times its steady-state
2
value MSD = E[
], for some given > 0. That is, it is the
time k = K at which we have
B
2
(41)
E[K
] = (1 + )MSD = (1 + )
1A

We assume that the estimator is initialized with 0 , which


provides E[02 ] = E[(0 )2 ] = (0 )2 . Next, we rewrite
expression (34) more conveniently as
 
 2 
E k2 MSD = AE k1
+ B MSD
 2 

= A E k1
MSD
(42)
 2

k
= A E 0 MSD .

At k = K, solving (41) and (42) together we get




B
1
ln
K=
ln(A)
(1 A)(0 )2 B

(43)

Using the definitions of and , we can rewrite (43) in a much


simpler form as follows:


 1
MSD
(iter.)
(44)
K = ln

MSD
Substituting the values of A and B from (34), (??), (??) and (??)
into (43) yields the convergence time for Chung-, Mathis- and
proposed (??) algorithm.
E. Simulation Evidences of Acquisition Ability
1) Acquisition ability and steady-state behavior: In
Fig. 8,
we compute the steady-state mean deviation =
MSD
versus the loop gain (as specified in (36)) for a 32-QAM
source ( = 18.7), in a noise free environment, and compare
them with those obtained from simulation; the values of were
taken to be 5 , 15 and 30 . The simulation results are
obtained by averaging 1000 Monte-Carlo (MC) experiments each
of 50,000 iterations. Observe that the analytical and simulation
results are conforming with each other, and the values of MSD
are independent of the target value .
Fig. 8.

Analysis vs simulation: values of for 32-QAM

2) Convergence behavior at same MSD floor: In Fig.


9, we compare the theory and simulation convergence-time
for a 32-QAM source with unknown target = 10
in a noise free environment. The step-sizes ( ) are selected such that the steady-state phase deviation ( ) is 2
(MSD = 20 log10 (2/180) 29dB). Observe that 1) the convergence behavior of Mathis algorithm is slightly better than that
of Chung algorithm, 2) the proposed update has a convergence
speed significantly faster than Chung and Mathis algorithms,
and 3) the theoretical values of K are in good agreement with
simulation values for all three adaptive estimators.
Fig. 9.
The convergence behavior of Chung, Mathis and proposed
algorithms, acquiring = 5 .

3) Steady-state behavior at same convergence speed: In Fig.


10, we compare the steady-state MSD while keeping the same
convergence-time in a noisy condition. Notice that the MSD floor
of the proposed update is significantly lower than those of Chung
and Mathis algorithms.
Fig. 10.
The steady-state behavior of Chung, Mathis and proposed
algorithms, acquiring = 10 .

Fig. 11. The convergence behavior of the proposed algorithms, acquiring


= 20 with different initialization.

4) Initialization strategy:
VIII. T RACKING ABILITY OF

THE

DMD SYNCHRONIZER

A. Steady-State Tracking Error


Due to the presence of frequency offset of the carrier loop, the
resulting phase offset drifts over time. Assume the phase offset
is drifting linearly at the (normalized) rate , i.e., the true phase
offset at the kth update is given by1 k = + k. The deviation
of the estimated parameter k from the true phase offset is thus
given by:
k = + k k

(45)

Using (45), we modify (31) to obtain


k+1 = k + Ez

(46)

Taking the ensemble average, the steady-state system satisfies


limk E[k+1 ] = limk E[k ]. Using the approximations
(??), we obtain the tracking error (TE) as given by
2
.
TE = E[ ]
(rad)
c1

(47)

B. Total MSD and Optimal Loop Gain


In the presence of frequency offset, the total mean square
deviation (TMSD) can be obtained by combining the distortion
contributed by acquisition and tracking phenomenon, which gives
.
TMSD = MSD + TE2 =

42
c3
+ 2 2 (rad2 )
c1 c2
c1

(48)

Notice that the first and the second term in (48) increases and
decreases monotonically with the loop-gain , respectively. The
optimal loop gain, opt
is obtained by seeking the minimum of
(48). We need to solve TMSD/ = 0, which gives
c31 c3 3 8c22 2 2 + 16c1 c2 2 = 8c21 2 .

(49)

Solving for the positive (real-valued) root of (49), we get the


following optimum value of :

5
1
2 c2 4 3 3c3 c41 42 c32
(4c4 ) 3
82 c22
opt
=
+

(50)
1
3c3 c31
3c3 c31
3c c3 c 3
3 1 4

f
, where
Rs
Rs is the symbol-rate and f is the difference between transmitter and
receiver local-oscillator frequencies.
1 The

normalized frequency offset is defined as = 2

where c4 is obtained as
c4 = 1444 c32 c3 c41 + 27c23 c81 2 + 1286 c62
s
3
27c3 c41 322 c32
2 2 6
2
+ 3 c3 c1
c3

Fig. 15. Tracking behavior of Chung, Mathis and proposed algorithms


for 0 = 0.5 rad, = 1 103 for 128-QAM.

(51)

Notice that the optimum loop-gain depends only on the statistics


of the QAM signal and the frequency offset ; moreover, it
increases with the .
C. Simulation Evidences for Tracking Performance
1) Validating (48) and (50): Due to the space limitation, we
validate Expressions (48) and (50) for the proposed algorithm
only. Refer to Figure 12, where the TMSD is obtained by
simulation and compared with our analytical result (48) for 32QAM with three different values of . Since (48) has been derived
for an ideal noise-free condition, we have not used additive noise
in the simulation as well. It is pretty clear that Expression (48)
coincide with simulation results for all three values of . It
is found that the TMSD seeks its minimum at equals to
2.965 1014 , 4.672 1014 and 7.331 1014 for equals
to 5 105 , 1 104 and 2 104 , respectively which are
exactly the same as obtained from Expression (50).
Fig. 12. Plots of analytical and simulation TMSD with three different
for proposed algorithm.

To find out the analytical performance of the proposed algorithm in relation to Chung and Mathis algorithms, we have plotted
the TMSD versus (that is Expression (48)) in Figure 13 for 32QAM using optimal loop-gain (50). Notice that, for all values of
Fig. 13.
TMSD of Chung, Mathis and proposed algorithms versus
(normalized) frequency offset with optimal loop-gain for 32-QAM.

, the TMSD of the proposed algorithm is substantially smaller


than those of others. For example, at = 105 , the TMSD of
Chung and Mathis algorithms are approximately 103 ; while the
TMSD of the proposed algorithm reaches this level (that is 103 )
at 5 104 . It clearly indicates that the frequency-offset
tolerance of the proposed algorithm is about fifty (50) times more
than those of Chung or Mathis algorithms for 32-QAM (in an
ideal noise-free condition).
2) Tracking the frequency offset with phase wrapping: Here
we present an experiment to demonstrate the tracking behavior
of the three algorithms in a more realistic scenario. We have set
the SNR level of 30 dB, a constant (initial) carrier phase error
0 = 0.5 radian, and a considerably high frequency offset, =
1 103 . The actual and estimated phase are assumed to be
wrapped before and after the synchronizer, respectively, such that
|k |, |k | /2 radian. Figures 14-16 show the carrier tracking
for 32-, 128- and 512-QAM signals, respectively. In these figures,
the dashed line and the full line indicate, respectively, the original
and the estimated carrier phase. From these results, it appears that
clearly, our proposed method performs better than the classical
techniques, which show, in these cases, very large tracking errors.

Fig. 14. Tracking behavior of Chung, Mathis and proposed algorithms


for 0 = 0.5 rad, = 1 103 for 32-QAM.

Fig. 16. Tracking behavior of Chung, Mathis and proposed algorithms


for 0 = 0.5 rad, = 1 103 for 512-QAM.

3) Tracking a random slow-varying phase: Here we present


an experiment to demonstrate the tracking behavior of the three
algorithms in a (random) slow-varying phase environment for 32QAM. To generate a slow-varying phase, we used the following
piece of MATLAB code:
function theta = random_phase(N)
B = fir1(1e3,1e-4);
Y = filter(B,1,randn(1,N)); Y = Y-mean(Y);
theta = Y/max(abs(Y))*pi/6;
For a fair comparison, we first obtained the optimum value of
loop-gain ( ) for the three algorithms. It was done by calculating
the TMSD for several different values of ; the value of
which resulted in the lowest TMSD was selected as an optimum.
Refer to Figure 17, where TMSD has been obtained for Chung-,
Mathis- and proposed algorithm. The TMSD curves have been
obtained by averaging 1000 independent phase initialization for
each value of , while 30 different values of have been used
for each algorithm. To further assist the fairness, in each independent trial, the same value of phase generated by the MATLAB
function random_phase.m is used for all algorithms. Figure
Fig. 17. The tracking behavior of Chung, Mathis and proposed algorithms, acquiring a random slow-varying phase |k | /6 at SNR= 35
dB.

17 also depicts the random (reference) phase k , estimated phase


k , and the phase estimation error k = k k . Notice that
the value of k for the proposed algorithm is considerably small
as compared to the others. This experiment clearly indicates that
the proposed algorithm has a good tracking capability and it can
outperform other existing adaptive synchronizers.

Das könnte Ihnen auch gefallen