On the Impact of Analog-to-Digital Conversion on

Blind I/Q Imbalance Parameter Estimation in

Low-IF Receivers
Ines Riedel, Marcus Windisch, Gerhard Fettweis
Technische Universit at Dresden, Vodafone Chair Mobile Communications Systems,
01062 Dresden, Germany
Email: {riedeli, windisch, fettweis}@ifn.et.tu-dresden.de
AbstractThe low-IF receiver is very attractive, because the
need for an analog image rejection lter is avoided. Instead, the
image rejection is realized by I/Q signal processing. However,
unavoidable imbalances between the I- and Q-branch in the
analog part of the receiver result in a limited image attenuation.
In previous work a blind I/Q imbalance compensation scheme
for low-IF receivers was presented and its performance was
analyzed. However, the impact of the analog-to-digital conversion
(ADC) was not taken into consideration. In particular, the quan-
tization noise, which is unavoidable in practical implementations,
has been neglected. In this paper this blind I/Q imbalance
compensation scheme is analyzed regarding to its sensitivity to
an imperfect ADC. It can be shown by means of analysis as well
as simulation, that the compensation scheme is robust against
ADC effects.
I. INTRODUCTION
A perfect low-IF receiver enables innite image rejection
due to its I/Q signal processing [1], [2]. However, unavoid-
able impairments of the analog complex-valued I/Q down-
conversion lead to a limited attenuation of the image signal.
Considering that the image signal may originate from a
different carrier frequency, it may be originally more than 50
dB stronger than the desired signal. With todays technologies
an amplitude imbalance of 1-2% and a phase imbalance of 1-
2

are realistic, which lead to an image attenuation of only 30-

40 dB [1]. This insufcient attenuation requires an additional
digital I/Q imbalance compensation.
Several digital compensation schemes have been presented
in the literature. Conventional approaches include the off-line
calibration with analog test signals [3], and the application of
signal separation algorithms, such as interference cancellation
and blind source separation [4]. Although leading to signi-
cant improvements under specic receiving conditions, these
methods come with inherent drawbacks [4], [5].
A novel approach for the digital I/Q imbalance compensa-
tion, which is based on a blind I/Q imbalance parameter esti-
mation, has been proposed in [5]. However, in the derivation of
the algorithm an ideal analog-to digital conversion (ADC) was
assumed. The impact of a non-ideal ADC on the performance
of the digital I/Q imbalance compensation will be analyzed in
this paper.
The paper is outlined as follows: In section II we will de-
scribe the problem of I/Q imbalance in low-IF receivers. Sec-
tion III briey introduces the analyzed compensation scheme.
e
j2f
IF
nT
e
+j2f
IF
nT
e
j2f
LO
t
s(n)

(n)
r(t) z
IQ
(t) z
IQ,q
(n)
Fig. 1. Architecture of a low-IF receiver with I/Q Imbalance Compensation
The analysis of the ADC is done in the following section IV.
The performance analysis is presented in section V. The nal
section VI concludes the paper.
II. I/Q IMBALANCE IN LOW-IF RECEIVERS
The low-IF receiver is well known in the literature [1],
[2]. The architecture of a low-IF receiver with I/Q imbalance
compensation is explained in the following. The received radio
frequency (RF) signal is down-converted to the intermediate
frequency (IF). The resulting complex-valued IF signal is low-
pass-ltered and digitized afterwards. This digital IF signal is
down-converted to the baseband (BB). In order to perform the
digital I/Q imbalance compensation, a second digital down-
conversion branch is introduced, as shown in Figure 1.
The ideal IF signal z(t), having no I/Q imbalance, can be
written as:
z(t) = s(t)e
j2fIF t
+ i(t)e
j2fIF t
, (1)
where f
IF
denotes the IF and s(t) denotes the desired signal.
i(t) denotes the baseband equivalent of the so-called image
signal, which is separated by 2f
IF
from the desired signal.
With this notation, the real-valued RF signal r(t) can be
written as:
r(t) = z(t)e
j2fLOt
+ z

(t)e
j2fLOt
, (2)
where f
LO
denotes the frequency of the complex-valued local
oscillator (LO) and the asterisk denotes complex conjugation.
The I- and Q-branch of an ideal complex-valued LO have
equal amplitudes and a phase difference of exactly 90

.
However, analog components always have impairments. The
mismatch of the amplitudes of the branches is commonly
described as amplitude imbalance and the mismatch of the
phases is commonly described as phase imbalance. Apply-
ing the real-valued parameters amplitude imbalance g and
the phase imbalance , the time function of the imperfect
complex-valued LO can be written as:
x
LO
(t) = cos(2f
LO
t) j g sin(j2f
LO
t + )
= K
1
e
j2fLOt
+ K
2
e
j2fLOt
, (3)
where K
1
and K
2
are complex-valued imbalance parameters:
K
1
=
1 + ge
j
2
, K
2
=
1 ge
+j
2
. (4)
The ideal LO is a special case of equation (3) with g = 1
and = 0

(K
1
= 1 and K
2
= 0). The IF signal with I/Q
imbalance z
IQ
(t) results in:
z
IQ
(t) = LP{r(t)x
LO
(t)} = K
1
z(t) + K
2
z

(t), (5)
where LP{} denotes low-pass ltering. This IF signal has to
be converted from the analog to the digital domain. Therefore
the time is discretized:
z
IQ
(n) = K
1
z(n) + K
2
z

(n), (6)
and the amplitudes are quantized. The resulting digital IF
signal z
IQ,q
is as follows:
z
IQ,q
(n) = z
IQ
(n) + e
IF
(n), (7)
where e
IF
(n) denotes the quantization error at the IF. Assum-
ing the low-pass lter is linear in the pass-band, the resulting
BB signal is:
d(n) = LP
_
z
IQ,q
(n)e
j2fIF nT
_
= K
1
s(n) + K
2
i

(n) + LP
_
e
IF
(n)e
j2fIF nT
_
. .
e
d
(n)
. (8)
The BB signal is a superimposition of the desired signal,
the image signal and a resulting BB quantization error. For
clearness the abbreviation e
d
(n) is applied for the resulting
BB quantization error of d(n) in the following. Exemplary
spectra of the RF signal, IF signals and the resulting BB signal
d(n) are shown in Figure 2.
III. I/Q IMBALANCE COMPENSATION
The analyzed compensation algorithm shall only be briey
introduced. A comprehensive presentation can be found in [5]
and [6]. The algorithm aims on a compensation of the impact
of the image signal on the desired signal. The approach is to
evaluate an additional BB signal (see Figure 2):
v(n) = LP
_
r
IF
(n)e
+j2fIF nT
_
= K
1
i(n) + K
2
s

(n) + LP
_
e
IF
(n)e
+j2fIF nT
_
. .
ev(n)
. (9)
For clearness the abbreviation e
v
(n) is applied for the resulting
BB quantization error of v(n) in the following. The relation
between the BB signals can be written using matrix notation:
_
d(n)
v

(n)
_
=
_
K
1
K
2
K

2
K

1
_
. .
K
_
s(n)
i

(n)
_
+
_
e
d
(n)
e

v
(n)
_
(10)
R(f)
f
LO
f
LO
f
s(t)e
j2(f
LO
+f
IF
)t

i(t)e
j2(f
LO
f
IF
)t

(a)
Z
IQ
(f)
f f
IF
f
IF
(b) Z
IQ,q
(f)
f f
IF
f
IF
e
IF
(n)

(c)
(d) D(f)
f
e
d
(n)

(e) V (f)
f
ev(n)

Fig. 2. Representative spectra of the analog RF signal (a), the analog IF

signal (b), the digital IF signal (c) and the resulting digital BB signals d(n)
(d) and v(n) (e), respectively.
The imbalance matrix K describes the mutual interference
of the desired signal and the image signal. This mutual
interference can be removed by applying the inverse of K.
However, K
1
is unknown in practice. The principle of the
analyzed compensation scheme is to blindly gain an estimate

K
1
. In a rst step, an estimate of the product K
1
K
2
is
calculated by evaluating a block of N consecutive samples of
the BB observations, the product

K
1
K
2
yields:

K
1
K
2
=
1
N

N
n=1
d(n)v(n)
1
N

N
n=1
|d(n) + v

(n)|
2
. (11)
Using the estimated product, the estimation of the real-valued
imbalance parameters is possible:
g =
_
14Re
_

K
1
K
2
_

= arcsin
_

g
2
Im
_

K
1
K
2
_
_
(12)
By adapting denition (4) to the complex-valued estimates

K
1
and

K
2
, an estimate of

K
1
can be determined as follows:

K
1
=
1
|

K
1
|
2
|

K
2
|
2
_

K

K
2

2

K
1
_
. (13)
The multiplication of the disturbed BB signals with

K
1
results in the estimated BB signals s(n) and

i(n):
_
s(n)

(n)
_
=

K
1

_
d(n)
v

(n)
_
. (14)
IV. MODEL OF THE ANALOG-TO-DIGITAL CONVERSION
The I/Q imbalance compensation takes place in the BB.
In order to determine the performance of the compensation
scheme, the impact of the ADC has to be analyzed. However,
the ADC takes place at the IF. The relation between the ADC
at the IF and the resulting impact on the BB is analyzed in
this section.
An important issue for the effective resolution of an ADC
is its saturation level. While too high signal amplitudes lead
to clipping, too low amplitudes result in a decrease of the
effective resolution. These effects can be modeled by dening
an input power backoff (IBO) as well as a peak-to-average
power ratio (PAPR).
The IBO describes the relation between the maximum
possible output power
2
out,max
of the ADC and the variance

2
x
of the input signal:
IBO =

2
out,max

2
x
(15)
The statistics of the input signal are considered using the
PAPR:
PAPR =

2
x,max

2
x
(16)
where
2
x,max
denotes the maximum signal power. Under the
condition IBOPAPR clipping is avoided. The quantization
error of the ADC is considered as independent, uniformly
distributed noise, known as the pseudo quantization noise
model [7]. Under this assumption the signal-to-noise ratio
(SNR) for a quantization with b bits results in:
SNR = 3 4
b
/ IBO, (17)
which can be equivalently written in dB-scale as:
SNR
[dB]
= 4.77dB + b 6.02dB IBO
[dB]
. (18)
In order to describe a real ADC, more errors than the quan-
tization errors have to be considered. The signal is additionally
distorted by aperture jitter, ambiguity of the comparator,
nonlinearities of the characteristic and several noise processes.
An excellent analysis can be found in [8]. These effects can
be modeled using an effective number of bits (ENOB) instead
of the nominal number of bits b. Hence the SNR is as follows:
SNR = 3 4
ENOB
/ IBO. (19)
This formula models an ADC considering signal statistics as
well as the impact of distorting effects of a real ADC. In a
low-IF receiver, the ADC takes place at the IF. Therefore the
SNR of the quantized IF signal z
IQ
(n) is dened as:
SNR =
P
zIQ
P
eIF
, (20)
where P
zIQ
denotes the power of z
IQ
(n) and P
eIF
denotes
the power of the IF quantization error e
IF
(n), as dened by
(7). Applying (6) and assuming that the desired signal and the
image signal are uncorrelated, P
zIQ
results in:
P
zIQ
=
_
|K
1
|
2
+|K
2
|
2
_
(P
s
+ P
i
), (21)
where P
s
and P
i
denote the power of the desired signal and
the image signal, respectively.
For the analysis of the BB signals, the power of the IF
quantization error P
eIF
is not a useful measure. Instead, only
those spectral components, which superimpose the IF signals
of interest are critical, as depicted in Figure 2. In general, the
power of these spectral components can be adjusted by noise-
shaping techniques. However, it is reasonable to consider the
special case of a uniformly distributed quantization noise. In
this case, the portion of the quantization noise corresponding
to the bandwidth of the BB signals is:
P
eBB
=
P
eIF
OSR
, (22)
where OSR denotes an oversampling ratio, which is deter-
mined by the ratio of the BB sampling frequency to the IF
sampling frequency. By merging (19) - (22), the power of the
equivalent baseband quantization noise can be calculated as
follows:
P
eBB
=
_
|K
1
|
2
+|K
2
|
2
_
(P
s
+ P
i
)
SNR
eff
. (23)
SNR
eff
is the effective SNR, which models all effects related
to the resampling and quantization process:
SNR
eff
=
P
zIQ
P
eBB
= SNR OSR
= 3 OSR 4
ENOB
/ IBO (24)
As expected, the SNR
eff
depends on the properties of the ADC.
The higher ENOB or OSR are, the higher is the SNR
eff
. A
large IBO leads to a degradation.
V. PERFORMANCE ANALYSIS
A. Theoretical Analysis
1) Performance without Digital Compensation: In order to
evaluate the performance of the compensation scheme, we
need an appropriate reference. As reference we consider the
BB signal without digital compensation. Equation (10) leads
to:
d(n) = K
1
s(n) + K
2
i

(n) + e
d
(n). (25)
Based on equation (25) we dene an image-and-noise-to-
signal ratio (INSR):
INSR
0
=
|K
2
|
2
P
i
+ P
e
d
|K
1
|
2
P
s
=

K
2
K
1

2
P
i
P
s
. .
ISR0
+
P
e
d
|K
1
|
2
P
s
. .
NSR0
, (26)
where P
e
d
denotes the power of the BB quantization error
e
d
(n). The reference INSR
0
is composed of two terms: an
image-to-signal ratio (ISR
0
) representing the impact of the I/Q
imbalance and a noise-to-signal ratio (NSR
0
). The NSR
0
is the
performance of a reception without any I/Q imbalance. This
limit holds also, if a digital compensation of I/Q imbalance is
applied.
Now, we consider the special case of a uniformly distributed
quantization noise (P
e
d
= P
eBB
). The NSR
0
as a function of
the power ratio of the desired signal and the image signal
P
s
/P
i
is depicted in Figure 3. For a dominating image signal
(P
i
P
s
) the ADC is saturated by the image signal, hence
the NSR
0
is increased. For (P
s
P
i
) the NSR
0
is determined
by the SNR
eff
. Figure 4 shows the composition of the INSR
0
,
80 60 40 20 0 20 40 60 80
100
80
60
40
20
0
20
P
s
/ P
i
in dB
N
S
R
0

i
n

d
B
SNR
eff
= 40 dB
SNR
eff
= 60 dB
SNR
eff
= 80 dB
SNR
eff
= 100 dB
NSR
0
(SNR
eff
= 40 .. 100 dB)
Fig. 3. Noise-to-signal ratio due to quantization without I/Q imbalance
based on NSR
0
and ISR
0
. The plots are parameterized by the
image rejection ratio (IRR):
IRR =

K
1
K
2

2
, (27)
which describes the nite image suppression due to the I/Q
imbalance. Figure 4 shows, that the INSR
0
is signicantly
degraded by the nite IRR.
2) Performance with Digital Compensation: By merging
(10) and (14), the estimated BB signals s(n) and

i(n) after
the digital compensation can be expressed as:
_
s(n)

(n)
_
=

K
1
K
_
s(n)
i

(n)
_
+

K
1

_
e
d
(n)
e

v
(n)
_
. (28)
Applying the abbreviations a
11
= K
1

K

1
K

2

K
2
and
a
12
=K

1

K
2
+K
2

K

1
, the product

K
1
K can be written as:

K
1
K =
1
|

K
1
|
2
|

K
2
|
2
_
a
11
a
12
a

12
a

11
_
. (29)
Hence, the estimated desired BB signal s(n) can be written
as:
s(n) =
a
11
s(n) + a
12
i

(n) +

K

1
e
d
(n)

K
2
e

v
(n)
|

K
1
|
2
|

K
2
|
2
. (30)
In the case of a perfect estimation (

K = K), the term a
12
will
be zero. However, due to the nite accuracy of the estimation,
a residual interference of the image signal will persist. Hence,
we dene the INSR after the digital compensation:
INSR
c
=
|a
12
|
2
P
i
+|

K
1
|
2
P
e
d
+|

K
2
|
2
P
ev
|a
11
|
2
P
s
=

a
12
a
11

2
P
i
P
s
. .
ISRc
+
|

K
1
|
2
P
e
d
+|

K
2
|
2
P
ev
|a
11
|
2
P
s
. .
NSRc
, (31)
where P
ev
denotes the power of the BB quantization error
e
v
(n). In analogy to the INSR
0
(no compensation), the INSR
c
is determined by the imperfect compensation (ISR
c
) and the
impact of the ADC (NSR
c
).
80 60 40 20 0 20 40 60 80
80
60
40
20
0
20
40
60
P
s
/ P
i
in dB
I
N
S
R
0

i
n

d
B
IRR = 20 dB
IRR = 30 dB
IRR = 40 dB
NSR
0
( SNR
eff
= 60 dB )
ISR
0
( IRR = 20 .. 40 dB )
INSR
0
( SNR
eff
= 60 dB, IRR = 30 dB )
Fig. 4. Image-and-noise-to-signal ratio (no compensation of I/Q Imbalance)
In the following we will derive analytic expressions of ISR
c
and NSR
c
. In order to simplify our analysis we incorporate
the following useful properties: For realistic I/Q imbalances
denition (4) yields K
1
1, K
2
0, hence |K
1
| |K
2
|.
Furthermore, we assume the estimates to be close to, but not
necessarily identical to the desired values, i.e.

K
1
K
1
and

K
2
K
2
.
First we analyze the term NSR
c
. The additional assumption
of similar noise powers P
e
d
P
ev
leads to:
NSR
c

|K
1
|
2
P
e
d
|K
1
K
2
|
2
P
s

P
e
d
P
s
. (32)
In contrast, the ISR
c
is more complicated. The results of the
estimation depend on the certain realization and are therefore
not deterministic. However, the performance can be evaluated
in a statistical sense. It has been shown in [6], that the
expectation of ISR
c
can be approximated by:
E{ISR
c
}

K
2
K
1

2
E
_

K1K2
K
1
K
2

2
_
P
i
P
s
, (33)
where
K1K2
=

K
1

K
2
K
1
K
2
denotes the absolute estimation
error of the product K
1
K
2
and E{} denotes expectation.
Our next goal is to analyze the expectation of the squared
magnitude of the relative estimation error on the right hand
side of (33). Therefore, the numerator and denominator of
(11) have to be analyzed using (8) and (9). For clearness, we
introduce the following abbreviations with the representative
complex-valued signals x(n) and y(n):

P
x
=
1
N
N

n=1
|x(n)|
2

R
xy
=
1
N
N

n=1
|x(n)y

(n)|
2
(34)
With this notation the numerator of (11) can be written as:
1
N
N

n=1
d(n)v(n) =K
2
1

R
si
+K
2
2

R

si
+K
1
(

R
se

v
+

R
ie

d
)
+K
1
K
2
(

P
s
+

P
i
)+K
2
(

iev
+

R

se
d
)+

R
e
d
e

v
.
Assuming |K
2
(

R

iev
+

R

se
d
)| |K
1
(

R
se

v
+

R
ie

d
)| yields:
1
N
N

n=1
d(n)v(n) K
2
1

R
si
+ K
1
(

R
se

v
+

R
ie

d
)
+ K
1
K
2
(

P
s
+

P
i
)+

R
e
d
e

v
. (35)
The denominator of equation (11) can be written as:
1
N
N

n=1
|d(n)+v

(n)|
2
=

P
s
+

P
i
+

P
e
d
+

P
ev
+2Re
_

R
si
+

R
se
d
+

R
sev
+

R
ie

d
+

R
iev
+

R
e
d
e

v
_
.
This term can be simplied with the following assumptions:
We assume that the sum of the cross-correlations of a BB sig-
nal with the quantization errors is smaller than the correspond-
ing signal power: |

R
se
d
|+|

R
sev
|

P
s
and |

R
ie

d
|+|

R
iev
|

P
i
.
The additional assumption, that the cross-correlation of two
signals is smaller than the sum of the corresponding signal
powers: |

R
si
|

P
s
+

P
i
and |

R
e
d
e

v
|

P
e
d
+

P
ev
yields for
the denominator of (11):
1
N
N

n=1
|d(n) + v

(n)|
2

P
s
+

P
i
+

P
e
d
+

P
ev
. (36)
Applying (35) and (36), the expectation of the squared
magnitude of the relative estimation error can be written as:
E
_

K1K2
K
1
K
2

2
_
= E
_
_
_

K
1

K
2
K
1
K
2
K
1
K
2

2
_
_
_
(37)
=E
_
_
_

K
2
1

R
si
+K
1
(

R
se

v
+

R
ie

d
)+

R
e
d
e

v
K
1
K
2
(

P
e
d
+

P
ev
)
K
1
K
2
(

P
s
+

P
i
+

P
e
d
+

P
ev
)

2
_
_
_
When splitting the expectation term in equation (37), several
products of cross-correlations arise. In order to simplify the
terms, we assume a sufciently large block size N. The esti-
mates of the cross-correlations converge under this condition
to their expectations. Furthermore, we assume independent
correlation functions and uncorrelated signals. Under these
conditions the following equation with the representative sig-
nals x, y, a and b holds:
E{

R
xy

R
ab
} = E{

R
xy
}E{

R
ab
} = R
xy
R
ab
= 0. (38)
The signal powers converge for a sufciently large block size
N as well to their expectations

P
x
E{

P
x
}=P
x
. With these
assumptions and the additional reasonable assumption
P
e
d
+P
ev
P
s
+P
i
, the expectation of the squared magnitude
of the relative estimation error results in:
E
_

K1K2
K
1
K
2

2
_

1
(P
s
+P
i
)
2
_

K
1
K
2

2
E
_
|

R
si
|
2
_
+
1
|K
2
|
2
_
E
_
|

R
se

v
|
2
_
+E
_
|

R
ie

d
|
2
_
_
+
E
_
|

R
e
d
e

v
|
2
_
|K
1
K
2
|
2
+(P
e
d
+P
ev
)
2
_
80 60 40 20 0 20 40 60 80
80
70
60
50
40
30
20
10
0
P
s
/ P
i
in dB
I
N
S
R
c

i
n

d
B
N = 100
N = 1,000
N = 10,000
N = 100,000
NSR
0
( SNR
eff
= 60 dB )
E{ISR
c
}|
SNR
eff

( N = 100 .. 100,000 )
E{INSR
c
} ( SNR
eff
= 60 dB, N = 10,000 )
Fig. 5. Composition of the image-and-noise-to-signal ratio with I/Q imbal-
ance compensation
As derived in [6], the variance of correlation functions can be
written as: E
_
|

R
xy
|
2
_
=
1
N
P
x
P
y
. Hence the E{ISR
c
} yields:
E{ISR
c
}
1
N
1
_
1+
Ps
Pi
_
2
+

K
2
K
1

2
P
i
P
s
(P
e
d
+P
ev
)
2
(P
s
+P
i
)
2
(39)
+
1
N|K
1
|
2
P
i
P
s
1
(P
s
+P
i
)
2
_
P
s
P
ev
+P
i
P
e
d
+
P
e
d
P
ev
|K
1
|
2
_
.
This general result simplies for the special case of a uni-
formly distributed quantization noise P
e
d
=P
ev
=P
eBB
. By
merging (23) into (39) we get:
E{ISR
c
} E{ISR
c
}

SNR
eff

+ , (40)
where
E{ISR
c
}

SNR
eff

=
1
N
1
_
1+
Ps
Pi
_
2
(41a)
=
1
SNR
eff
P
i
P
s
_
1
N
+
1
NSNR
eff
+
4
SNR
eff

K
2
K
1

2
_
. (41b)
The E{ISR
c
} is composed of two terms: The term
E{ISR
c
}

SNR
eff

represents the impact of the imperfect digital

image suppression due to nite estimator block length N in the
absence of any quantization noise. This result corresponds to
the results of [6]. The second term represents the degradation
of the estimation of the I/Q imbalance parameters due to the
impact of the ADC.
Similarly, by merging (23) into (32), the NSR
c
yields:
NSR
c
NSR
0

1
SNR
eff
_
1+
P
i
P
s
_
(42)
The application of (40) and (42) to the expectation of the
INSR
c
of equation (31) results in:
E{INSR
c
} E{ISR
c
}

SNReff
+ NSR
0
+ (43)
The performance E{INSR
c
} after the digital compensation
is composed of three terms. The rst term represents the
80 60 40 20 0 20 40 60 80
80
70
60
50
40
30
20
10
0
10
20
P
s
/ P
i
in dB
IRR = 34dB, SNR
eff
=72.73dB
I
N
S
R
c

i
n

d
B
N = 100
N = 1,000
N = 10,000
Theory: E{INSR
c
}
Theory: INSR
0
Theory: NSR
0
Simulation: mean INSR
c
Simulation: mean INSR
0
Simulation: mean NSR
0
Fig. 6. Comparison of NSR
0
, INSR
0
and E{INSRc} applying theoretical
analysis and simulation results
+f
IF
f
IF
fsym fs = 4fsym
f
IF
+f
IF
fsym
Fig. 7. Overview of the simulation
performance of the digital compensation scheme with a nite
block size N but no limitations due to the ADC. The second
term NSR
0
represents the impact of the ADC on the desired
signal. The third term represents the impact of the ADC on
the estimation of the I/Q imbalance parameters.
A comparison of (41b) and (42) yields NSR
0
for
realistic values of the SNR
eff
and the block size N. Hence
can be neglected in (43):
E{INSR
c
} E{ISR
c
}

SNReff
+ NSR
0
. (44)
The composition of the E{INSR
c
} is depicted in Figure 5.
B. Simulation
The theoretical results were veried by the results of
MATLAB simulations. For an overview see Figure 7. As
source signals, we used randomly generated QPSK signals
with a symbol rate of f
sym
= 3.84 MHz. These signals
were upsampled using a root-raised cosine (RRC) lter to a
sample rate of f
s
=4f
sym
, hence the oversampling ratio was
OSR=4. Afterwards the BB signals were upconverted to the IF
f
IF
=f
sym
, respectively. The I/Q imbalance was modeled
at the IF according to equation (6). The IF signal with I/Q
imbalance was quantized with an uniform mid-rise quantizer.
Based on the quantized IF signal the observations d(n) and
v(n) were generated by an appropriate down-conversion, RRC
ltering and decimation. The proposed compensation scheme
was applied in order to reconstruct the desired signal s(n).
Figure 6 shows the results of the theoretical analysis in
comparison to results of the simulation for the exemplary I/Q
imbalance parameters g =1.02 and =2

(IRR=34 dB) and

the exemplary ADC parameters ENOB=12 Bit and IBO=10 dB
(SNR
eff
=72.73 dB). It can be easily seen, that the simulation
results conrm the results of the theoretical analysis.
VI. CONCLUSION
The impact of a non-ideal analog-to-digital conversion
(ADC) on the performance of a novel blind I/Q imbalance
compensation has been analyzed in this paper. We developed
a set of equations for the theoretical performance evaluation
with and without digital I/Q imbalance compensation. The
validity of these equations has been conrmed by computer
simulations.
The quantization noise caused by the nite number of
discrete levels of the ADC has been identied as one of the
key problems in the practical implementation of the low-IF
receiver. Even in the presence of no I/Q imbalance, a powerful
image signal can saturate the ADC, such that the signal quality
of the desired signal is impaired. However, the situation is
signicantly aggravated in the presence of I/Q imbalance. In
this case, interference by the insufciently attenuated image
signal is the dominant source of signal degradation.
Fortunately, the interference by the image signal can be
removed by a digital post-correction of the I/Q imbalance.
Our analysis shows, that the analyzed blind I/Q imbalance
parameter estimation algorithm is insensitive to the presence
of quantization errors. Depending on the requirements of the
communication standard, the compensation algorithm can be
dimensioned, such that the performance with digital compen-
sation is arbitrarily close to the reference case of a low-IF
receiver with ADC errors, but without any I/Q imbalance.
Hence, the demands to the image rejection capabilities of the
analog front-end are signicantly reduced due to the digital
I/Q imbalance compensation.
REFERENCES
[1] Jan Crols and Michiel S. J. Steyaert. Low-IF Topologies for High-
Performance Analog Front Ends of Fully Integrated Receivers IEEE
Trans. on Circuits and Systems II, volume 45, pages 269282, Mar. 1998.
[2] Shahriar Mirabbasi and Ken Martin. Classical and modern receiver
architectures IEEE Communications Magazine, volume 38, pages 132
139, Nov. 2000.
[3] Jack P. F. Glas. Digital I/Q imbalance compensation in a Low-IF receiver.
In IEEE Global Communications Conference, volume 3, pages 1461
1466, Nov. 1998.
[4] Mikko Valkama, Markku Renfors and Visa Koivunen. Advanced methods
for I/Q imbalance compensation in communication receivers. IEEE Trans.
on Signal Processing, volume 49, Oct. 2001.
[5] Marcus Windisch and Gerhard Fettweis. Blind I/Q Imbalance Parameter
Estimation and Compensation in Low-IF Receivers. In Proc. 1st Intl.
Symposium on Control, Communications and Signal Processing (ISCCSP
2004), Hammamet, Tunisia, Mar. 2004.
[6] Marcus Windisch and Gerhard Fettweis. Performance Analysis for Blind
I/Q Imbalance Compensation in Low-IF Receivers. In Proc. 1st Intl.
Symposium on Control, Communications and Signal Processing (ISCCSP
2004), Hammamet, Tunisia, Mar. 2004.
[7] Bernard Widrow, Istvan Kollar and Ming-Chang Liu. Statistical theory of
quantization IEEE Trans. on Instrumentation and Measurement, volume
45, pages 353361
[8] Robert H. Walden. Analog-to-Digital Converter Survey and Analysis
IEEE Journal on Selected Areas in Communications, pages 539550,
Apr. 1999.