Beruflich Dokumente
Kultur Dokumente
7, JULY 2012
2353
of the low-pass filters. They generate conjugate-image interference after frequency down-conversion. As a result, both
the CFO and I/Q imbalance can significantly deteriorate the
system performance [1], [10]. Furthermore, they can be quite
severe in low-cost DCRs [2].
Manuscript received May 24, 2010; revised October 25, 2010, May 4 and
September 29, 2011, and January 4, 2012; accepted March 26, 2012. The
associate editor coordinating the review of this paper and approving it for
publication was M. Torlak.
The authors are with the Graduate Institute of Communication Engineering,
Dept. of Electrical Engineering, National Taiwan University, Taipei, Taiwan
106, R.O.C. (e-mail: d97942022@ntu.edu.tw, smp@cc.ee.ntu.edu.tw).
This work is supported by the National Science Council, Taiwan, R.O.C.,
under Grant NSC 100-2221-E-002-201-MY3.
Digital Object Identifier 10.1109/TWC.2012.051412.100882
Index TermsMulti-carrier modulation, carrier frequency offset (CFO), frequency-dependent I/Q imbalance, direct-conversion
receivers.
I. I NTRODUCTION
N recent years, multi-carrier modulation (MCM) has been
adopted by many broadband wireless standards due to its
high spectrum efficiency, ease of channel equalization, and
the flexible data rate. The direct-conversion receiver (DCR)
has become a major approach to the implementation of the
systems employing MCM due to its small circuit-size and
low power consumption. Although MCM systems are robust
against frequency-selective channels, they are sensitive to
synchronization errors, such as carrier frequency offset (CFO),
and to analog front-end non-idealities, such as I/Q imbalance.
CFO is caused by the frequency difference between the local
oscillators and/or by the Doppler effect in mobile systems.
It will induce the inter-carrier interference (ICI) between
the subcarriers. On the other hand, for wideband DCRs, the
I/Q imbalance consists of two parts. One is the frequencyindependent imbalance resulting from the oscillator gain and
phase mismatches between the in-phase (I-) and quadraturephase (Q-) branches. The other is the frequency-dependent
imbalance resulting from filter (frequency response) mismatch
c 2012 IEEE
1536-1276/12$31.00
2354
(1)
(2)
PAN and PHOONG: A TIME-DOMAIN JOINT ESTIMATION ALGORITHM FOR CFO AND I/Q IMBALANCE IN WIDEBAND DIRECT-CONVERSION . . .
GI f
Re 2 r (t ) e j 2Sf c t
cos2S ( f c 'f )t
1 H sin 2S ( f c 'f )t I
GQ f
Fig. 1.
Ts
Ts
yI n
yQ n
A. Cosine estimator
Like [12], [13], [15], [22], we assume that P 4 and the
channel is stationary during the transmission of the training
sequence. The first period of the received signal (0 n <
M ) is dropped since it is considered as a part of the guard
interval. We assume the order of h+ (n) (as well as h (n))
in (5) is M . Because of the phase difference in (1) and the
assumption of stationary channel, it can be verified that for
M n (P 1)M 1, we have
(7)
Let y(n) be the received signal given in (5) and define the
31 vector1
y(n)
y(n) = y(n + M ) for M n < (P 2)M. (8)
y(n + 2M )
Using (5) and (7), we can write y(n) as
y(n) = A()
with
1
A() = ej(M+)
ej2(M+)
q(n)
+ q(n + M )
q(n + 2M )
(9)
1
ej(M+) .
ej2(M+)
Note that the 32 matrix A() is independent of I/Q imbalance parameters and the time index n. Collecting all the y(n)
vectors in (8), we form the 3(P 3)M matrix
q(M )
q(M + 1)
Q = q(2M ) q(2M + 1)
q(3M ) q(3M + 1)
(11)
form
. . . q((P 2)M 1)
. . . q((P 1)M 1) .
. . . q(P M 1)
(12)
Computing YY H , we have a 33 matrix
H
H
H
YY H A() XX
A () + QQ ,
(13)
2355
(14)
(16)
1
0
2
u0
. (17)
v = Bu with B = 0 1 and u =
u1
1
0
2
The columns of B are normalized so that BH B = I2 . Because
there exists a nonzero vector v such that AH ()v = 0, which
implies that we can also find a nonzero vector u such that
AH ()Bu = 0. The objective becomes finding a nonzero
21 vector u such that
uH BH A()AH ()Bu = 0.
(18)
From u,
we obtain the cosine estimator as cos(M + ) =
u1 /( 2u0 ).
In practice, YY H contains the noise term QQH . Because
qch (n) is white (by assumption), it can be shown by direct
calculation that E{q(n)q (n+M )} = E{q(n)q (n+2M )} =
0. Using this fact and (12), it can be verified that when the
length of the training sequence satisfies P M 1, QQH can
2
be approximated by a diagonal matrix QQH Q
I3 , where
2356
2
Q
= (P 3)M N0 and N0 is the variance of q(n). Using
this approximation, we have
2
YY H A()AH () + Q
I3
(19)
yI n
B YY B = B A()A ()B +
= BH A()AH ()B +
H
2
Q
BH B
2
Q
I2 .
yQ n
cos <
+
.
M
M
M
B. Tangent estimator
In the previous CFO estimation, is obtained by inverting
a cosine function. It is known that the performance of the
CFO estimator based on the cosine function degrades when
M + is close to a multiple of [16]. Moreover, since the
range of inverse cosine function is [0, ), a phase ambiguity
arises when M + is out of this range and thus the estimation
range of CFO is limited. To solve these issues, we introduce
another algorithm that generates sine estimate. By combining
the two CFO estimates (cosine and sine), we will obtain a
CFO estimator based on the tangent function, which is more
robust to the small M + . To derive the sine estimator, we
define another matrix based on Y as
1
0
2
0 1 .
BH
s Y with Bs =
1
0
2
Similar to (20), we obtain the 22 matrix
H
H
H
2
H
R s = BH
s YY Bs Bs A()A ()Bs + Q Bs Bs
H
2
= BH
s A()A ()Bs + Q I2 .
(23)
2c1 sin2 (M + )
j 2c2 sin(M + )
Rs
j 2c2 sin(M + ) c1
2
+ Q
I2 ,
1
cos I
e jn[
y'I n
tan I
G n nw
y 'Q n
(20)
u1
}.
cos(M
+ ) = {
2
u0
w1 n
w n
CFO
compensation
Stage 2
Stage 1
(24)
2 A signal s(n) is said to be circularly symmetric if its 0th -order complementary autocorrelation function E{s2 (n)} = 0 [10]. Circularly symmetric
signals include many commonly used constellations such as QAM and M PSK (M > 2).
Fig. 2.
M
M
where c1 = (P 3)M Es ( k=0 |h+ (k)|2 + l=0 |h (l)|2 ),
M
M
c2 = (P 3)M Es ( k=0 |h+ (k)|2 l=0 |h (l)|2 ), and Es
is the average power of s(n). From above, we can compute
| sin(M )| from the diagonal elements of (24). In order to do
2
. From the
this, we need to know the estimated value of Q
derivation of (18)-(21), it can be seen that
2
Q
= the smallest eigenvalue of BH YY H B.
(25)
2
In other words,
Q
can be obtained in prior at the cosine
estimation. Also note that for most practical cases, the image
interference produced by I/Q imbalance is minor compared
to the desired signal, which leads to c2 > 0. Therefore, the
polarity of sine is the same as the polarity of {[Rs ]21 }. Using
this fact, (24), and (25), the CFO sine estimate is obtained as
[R ]
2
Q
]
}
{[R
s
s
21
. (26)
11
sin(M
+ ) =
|{[Rs ]21 }| 2 [R ]
2
Q
s 22
tan
.
(27)
tan =
M
cos(M
+ )
As the range of inverse tangent is [, ), the range of the
CFO estimator becomes independent of , or equivalently
tan <
.
M
M
IV. C OMPENSATION OF I/Q I MBALANCE
A structure for the compensation of frequency-dependent
I/Q imbalance and the correction of CFO is proposed by
[21]. The compensation structure is shown in Figure 2. At
Stage 1, the filter and gain mismatches are compensated by a
length-L FIR filter w1 (n) at the I-branch and a proper delay
nw is added to the Q-branch to match the delay introduced
by w1 (n). At Stage 2, the phase mismatch can be compensated by computing yI (n) = (1/ cos()) (yI (n) w1 (n)) +
tan()yQ (n nw ). To further simplify the structure, the gain
factor 1/ cos() at the I-branch can be merged into w(n).
(n) is fed into the CFO
The resulting y (n) = yI (n) + jyQ
compensator to complete the compensation process.
Below we will propose a new method to estimate w(n) and
tan() by adopting the compensation structure in Figure 2. As
we will demonstrate later, the proposed method has a much
better performance than [21]. In the following, we assume that
the CFO has been estimated as in (27) and thus it is known.
Assuming that there is no channel noise, it was shown in
[16] that if the I/Q imbalance is completely eliminated in
PAN and PHOONG: A TIME-DOMAIN JOINT ESTIMATION ALGORITHM FOR CFO AND I/Q IMBALANCE IN WIDEBAND DIRECT-CONVERSION . . .
Stages 1 and 2, then the signals yI (n) and yQ
(n) in Figure 2
3
satisfy
yI (n + 2M ) yI (n) = 2 sin(M + )yQ
(n + M ),
2 sin(M + )yI (n + M ) = yQ
(n + 2M ) yQ
(n).
(28)
C(n) =
d(n) =
,
2357
d(n)
J (w, tan()) =
tan()
n=M+L1
(31)
that minimize (31) can be
and tan()
The solutions w
obtained by solving a linear least square (LLS) equation. The
solution is given in closed-form:
d(M + L 1)
C(M + L 1)
C(M + L)
d(M + L)
.
.
..
..
tan()
d ((P 2)M 1)
C ((P 2)M 1)
(32)
where the superscript denotes the pseudoinverse.
The case of M + = 0: As noted in [19], the time-domain
estimation algorithms for I/Q imbalance suffer from a poor
performance when M + = 0. This can be seen from (5)
and (7). When M + = 0, the received samples y(n) in
(5) becomes periodic (in the absence of noise) for any I/Q
imbalance. Thus any estimation algorithm that exploits the
periodicity of training sequence for I/Q estimation will fail.
In our method, similar result is occurred when M + is a
3 Note that [16] considers only frequency-independent I/Q imbalance.
Furthermore, though the expressions in (28) were derived in [16], they were
used for CFO estimation (not I/Q imbalance). In this paper, we show how
the same expressions can be used for I/Q estimation.
V. C OMPLEXITY I SSUE
In this section, we will analyze the complexity of the
proposed algorithm and make comparisons to some of the
existing methods. For the CFO estimators derived in Section
3, the main computation is the calculation of the 33 matrix
YY H . Recall that the 3(P 3)M matrix Y has the
structure shown in (10). The second and third rows of Y
are the shifted versions of the first row of Y. We can use
a sliding window approach to compute the elements of YY H
efficiently. If we do this, by careful calculation, it can be
verified that the computation of YYH needs (3P 6)M
complex multiplications and (3P 6)M 7 complex additions.
For the estimator in Section 4, the main complexity is the LLS
solution in (32).
The complexity of the proposed algorithms is summarized
in Table I in terms of real additions (ADD) and real multiplications (MUL), along with existing methods in [21], [22].
For simplicity, we only show the most dominant term of
each operation. The CFO estimation in [21] requires a sample
covariance matrix and a one-dimensional search algorithm,
whereas the proposed method and [22] do not. The computations of the P P sample covariance matrix and each cost
function require 2P 2 M and 12P 2 ADD/MUL respectively.
Since the search algorithm depends on the search resolution
as well as the search range, we denote the number of search
points as and thus the complexity of one-dimensional search
algorithm is 12P 2 . In Table I, we assume that all the LLS
problems are solved using Householder solution [27], which
requires n2 (m n/3) ADD and MUL for an m n LLS
problem. In comparison, the complexity of [21] is much higher
than the other two if P is large. On the other hand, we can see
that when L > 2 (the length of w(n) is larger than 2), which
2358
TABLE I
C OMPLEXITY C OMPARISON
[21]
[22]
Proposed
ADD/MUL
12P 2 + 2P M P + (L + 1)2
2P M (L + 2)2
2P M (L + 1)2 + 7
Proposed
[21]
No compensation
Example in Section 6
35
7.47 104
1.57 104
1.38 104
Method
40
25
20
15
10
Fig. 3.
10
15
20
25
SNR (dB)
30
35
40
50
45
40
Image Rejection Ratio (dB)
30
35
30
25
20
Proposed
[21]
No compensation
15
10
5
20
10
0
10
Subcarrier Index
20
30
Fig. 4. IRR versus subcarrier index with SNR= 20dB and perfect knowledge
of CFO (M = 0.078).
PAN and PHOONG: A TIME-DOMAIN JOINT ESTIMATION ALGORITHM FOR CFO AND I/Q IMBALANCE IN WIDEBAND DIRECT-CONVERSION . . .
2359
10
10
[22]
Proposed (cosine)
Proposed (tangent)
[21]
10
[22]
Proposed (cosine)
Proposed (tangent)
10
10
10
10
10
10
10
10
10
10
10
10
15
20
SNR (dB)
25
30
Fig. 5.
10
35
Fig. 7.
1.5
0.5
0.5
1.5
10
10
SNR=10dB, [22]
SNR=10dB, Proposed (tangent)
SNR=25dB, [22]
SNR=25dB, Proposed (tangent)
10
10
2
10
10
10
10
10
10
10
10
Fig. 6.
10
[22]
Proposed (cosine)
Proposed (tangent)
[21]
10
15
20
SNR (dB)
25
30
35
10
Fig. 8.
10
15
P
20
25
2360
10
10
A PPENDIX
D ERIVATION OF (24)
10
as
10
No compensation
[22]
Proposed
[21]
Perfect compensation
10
10
15
20
SNR (dB)
with
11 =
25
30
35
L+ L
(P 2)M1
12 =
22 =
L+ L
n=M
h (k1 )h (k2 )
(P 2)M1
s (n k1 )s(n k2 ),
n=M
10
10
(P 2)M1
No compensation
[22]
Proposed
[21]
Perfect compensation
10
15
20
SNR (dB)
s(n k1 )s (n k2 )
n=M
(P 2)M1
10
Fig. 10.
k1 )s(n k2 ),
n=M
k1 =0 k2 =0
10
s(n k1 )s (n k2 ),
(P 2)M1
h+ (k1 )h (k2 )
ej2n s(n
k1 =0 k2 =0
10
k1 =0 k2 =0
L+
Fig. 9.
12
22
(P 3)M Es ,
0,
k1 = k2
otherwise
ej2n s (n k1 )s(n k2 ) 0, k1 , k2 ,
n=M
25
30
35
PAN and PHOONG: A TIME-DOMAIN JOINT ESTIMATION ALGORITHM FOR CFO AND I/Q IMBALANCE IN WIDEBAND DIRECT-CONVERSION . . .
[23]
[24]
[25]
[26]
[27]
2361
pensation for carrier frequency offset and I/Q imbalance, IEEE Trans.
Commun., vol. 58, no. 2, Feb. 2010.
L. Anttila, M. Valkama, and M. Renfors, Frequency-selective I/Q
mismatch calibration of wideband direct-conversion transmitters, IEEE
Trans. Circuits and Systems II, vol. 55, no. 4, pp. 359363, Apr. 2008.
Y. Zou, M. Valkama, and M. Renfors, Pilot-based compensation of
frequency-selective I/Q imbalances in direct-conversion OFDM transmitters, 2008 IEEE Conf. on Vehicular Technology Fall.
Y.-C. Pan and S.-M. Phoong, A new algorithm for carrier frequency
offset estimation in the presence of I/Q imbalance, 2010 IEEE Conf.
on Vehicular Technology Spring Spring.
Y.-C. Pan and S.-M. Phoong, A new estimation algorithm for CFO
and I/Q imbalance in wideband direct-conversion receivers, 2011 IEEE
Global Telecommunications Conf.
G. H. Golub, and C. F. Van Loan, Matrix Computations, 3rd edition.
The Johns Hopkins University Press, 1996.
Yen-Chang Pan (M09) was born in Taipei, Taiwan,
in 1983. He received the B.S. and M.S. degree in
communication engineering from National ChiaoTung University (NCTU), Hsinchu, Taiwan, in 2006
and 2008 respectively. He is now a Ph.D. student at
Graduate Institute of Communication Engineering,
National Taiwan University (NTU), Taipei, Taiwan.
His current research interests include discrete-time
signal processing for the estimation of front-end
non-idealities in communication systems.