RADIOENGINEERING, VOL. 32, NO.

4, DECEMBER 2023 479

IQ Imbalance Correction
in Wideband Software Defined Radio Transceivers
Borisav JOVANOVIC 1,2, Srdan MILENKOVIC 1,2
1
Dept. of Electronics, Faculty of Electronic Engineering, University of Niš, Aleksandra Medvedeva 14, Niš, Serbia
2
Lime Microsystems, Surrey Tech Centre, Occam Road, Guildford GU2 7YG, Surrey, UK

borisav.jovanovic@elfak.ni.ac.rs

Submitted July 5, 2023 / Accepted September 21, 2023 / Online first October 17, 2023

Abstract. A method for compensation of frequency- that even with careful RF transceiver design only 30 to
selective (FS) in-phase/quadrature (IQ) imbalance of 40 dBc of image suppression can be achieved [1], [3].
a wideband transceiver is proposed in the paper. It is dedi-
The utilization of a static IQ imbalance (IQI) mitiga-
cated for implementation in software defined radio (SDR)
tion method is sufficient for calibration of narrowband
cellular base stations. Both transmitter (TX) and receiver
transceivers. When wideband waveforms are transmitted
(RX) IQ impairments are corrected by complex valued
the IQI becomes frequency selective (FS) and more com-
finite impulse response (FIR) filters which are designed
putationally complex methods are required for compensa-
based on previously found imbalance correction models.
tion. The IQI mitigation is an important part of fifth gener-
The compensation performance is assessed after the
ation (5G) transceiver design. Imbalance deteriorates the
method was implemented in the SDR platform capable of
error vector magnitude (EVM) for which the 5G standards
transmitting signals at different central frequencies. At
have strict requirements [4].
frequencies higher than 3 GHz measured IQ gain and
phase error functions exhibit asymmetrical characteristic. This paper presents a novel IQI reduction method
In order to reduce the level of asymmetry, adopted IQ gain which is dedicated to wideband transceivers. Both receiver
correction model incorporates odd polynomial elements and transmitter IQI are corrected. The method is imple-
while the phase correction model includes even polynomial mented in a software defined radio (SDR) board. The uti-
parts. Regardless of utilized central frequency IQ impair- lized SDR supports different multi communication stand-
ments are efficiently compensated. The advantage of the ards, frequency carriers and channel bandwidths. The par-
proposed method is low complexity. The method doesn't ticular requirement of the SDR is to transmit modulation
require specialized hardware for calibration, instead, it waveforms having 100 MHz bandwidth while operating at
uses the RF loopback. At central frequency of 3.5 GHz, 3.5 GHz central frequency. Initial SDR board measure-
transmitter image rejection ratio (IRR) is increased from ments revealed significant IQ imbalance and emphasized
20 dBc to 45–50 dBc by applying the proposed method. the necessity for efficient IQI compensation method. Novel
After receiver imbalance is compensated, the improvement IQI reduction method is developed, realized using complex
in IRR of more than 25 dBc is achieved. valued digital filters which are added to the digital base-
band processing blocks. Steps for IQI mitigation are thor-
oughly described, starting from mathematical models to
a complete realization in the SDR.
Keywords
This paper is organized as follows. Related work is
Frequency selective IQ imbalance, transmitter, re- given in the following section. In Sec. 3, the method is
ceiver, software defined radio, IQ calibration described. In Sec. 4, implementation is presented followed
by measured results. Section 5 is dedicated for discussion.
The conclusion is drawn in the last section.
1. Introduction
In ideal case, quadrature mixing in radio frequency
(RF) transceivers completely suppresses image frequency
2. Related Work
components [1], [2]. In-phase/quadrature (IQ) imbalances, The IQI includes the contributions from the analogue-
caused by degraded symmetry between I and Q signal to-digital converter (ADCs), digital-to-analogue converters
paths, reflect in the path's unequal frequency responses. (DACs), the analogue low-pass filters (LPFs), as well as
Uneven amplitudes of quadrature mixer local oscillator the signal paths [3]. Refs. [5–7] investigate the mismatches
(LO) signals, and their phase shift, which is not equal to 90 between LPFs of I and Q signal paths as a source of the FS
degrees, additionally contribute the imbalance. It is proven imbalance. In order to equalize the LPF frequency re-

DOI: 10.13164/re.2023.0479
480 B. JOVANOVIC, S. MILENKOVIC, IQ IMBALANCE CORRECTION IN WIDEBAND SOFTWARE DEFINED RADIO TRANSCEIVERS

sponses, additional feedback path, consisting of ADC, is been equipped with an additional RF signal generator pro-
embedded to the transceiver design. The feedback path ducing single tone test signals. Reference [21] utilizes the
returns the transmitter output signal to the baseband pro- cross-power spectrum between I and Q signals to cancel
cessing unit for IQI identification [6], [7]. The imbalance is the linear phase IQI consisting of the time delay deviation
mitigated using digital filters which are added to the digital and the LO phase offset. The IQ amplitude mismatch and
baseband processing blocks [5], [7]. In [7] the test se- phase mismatch are reduced separately. The previous work
quence is transmitted and fed back, while the operations [22] is based on memory polynomial DPD which jointly
are performed at transceiver start-up. The least squares compensates transmitter IQI and PA nonlinearities. Trans-
(LS) optimization method is employed to minimize the mitted signal bandwidth is limited by DPD operation which
difference between the desired filter response and the requires that the DAC/ADC sampling frequency is at least
measured one. The reference [6] extends the operation of five times greater than the signal bandwidth. The maximum
IQI correction circuits of [7] in order to be adaptive. Also, signal bandwidth of 20 MHz is achieved [22].
the imbalance is neutralized during transceiver operation.
The methods found in literature are mostly validated
In [8], [9] the LS time domain approach is used for trans-
in laboratory using test equipment that relies on high-
mitter impairments identification. The impairments are
performance instruments and state-of-the-art transceiver
reduced in the baseband using complex-valued finite im-
development boards. Also, the captured signals are im-
pulse response (FIR) filters.
ported to MATLAB where algorithms are carried out.
The IQI is detrimental to the digital predistortion
The main contributions of the proposed method are:
(DPD) performance [10]. Many publications combine the
IQI reduction with DPD power amplifier (PA) lineariza- • IQ imbalance calibration method, dedicated for wide-
tion. Such methods extend the parallel Hammerstein struc- band transceivers, utilizes the test tones and has ad-
ture [11], Volterra series model [12] and asymmetrical vantage that both transmitter and receiver are com-
complexity-reduced Volterra series model [13] to reduce pensated in the same process.
both the transmitter IQ imbalance and PA nonlinearities.
Impairments are cancelled using the complex-valued filters • Beside the IQ imbalance compensation, amplitude
and DPD monitoring paths [10–13]. Although efficient, responses are flattened. Impairments are compensated
methods suffer from an increase in computational com- for 100 MHz signal bandwidth.
plexity of complex-valued filters compared to independent • The method was successfully implemented in a SDR
IQI compensation [14]. board capable of transmitting waveforms at different
The IQI calibration algorithms can be divided into LO frequencies. The measured results showed that at
training sequence methods and blind methods according to frequencies higher than 3 GHz the IQ gain and phase
whether the test signals are generated during calibration. error functions exhibit asymmetrical characteristic.
Blind calibration does not require special test signals. In- The imbalance is mitigated using complex-valued fil-
stead, it exploits the inherent characteristics of the received ters [9], [19]. Regardless of selected LO frequency,
signal to calculate the imbalance parameters. Blind calibra- and the level of asymmetry in IQ imbalance func-
tion can be based on blind source separation (BSS) tech- tions, the IQ impairments are efficiently compensated.
niques [15] and on signal statistical characteristics [15], [16]. • The method doesn't require specialized hardware for
The authors of [17] proposed a new adaptive algo- calibration; it uses the RF loopback instead.
rithm for imbalance neutralization of receivers, which is
based on backward blind source separation (BBSS) struc-
ture and the fast Newton transversal filter (FNTF) tech- 3. Method for IQ Imbalance Correc-
nique. In [18] a blind calibration method is applied for
receiver’s imbalance reduction. The frequency-domain
tion
statistical characteristics of the received signal are used for
the construction of the classification rule that estimates the 3.1 The Polynomial Models of Gain, IQ Gain
imbalance parameters. A real-valued digital filter is added
to I component path to cancel the IQI. For method valida-
and Phase Imbalance
tion the gain and phase imbalances are generated using Exposed to the source of IQI, the complex-valued
mismatched data. Reference [19] corrects the IQ impair- signal x(n), composed of quadrature components xI(n) and
ments of an arbitrary waveform generator. The solution is xQ(n), is transformed into y(n), consisting of components
based on a complex-valued filter whose structure is denoted with yI(n) and yQ(n):
adopted from [9]. Method performance is assessed by la-
boratory measurements where the RF signal is acquired by xI (n) ⋅ GI (ω ) cos(ω n + ϕI (ω )),
yI ( n ) = (1)
high-performance oscilloscope and uploaded to a PC for xQ (n) ⋅ GQ (ω ) sin(ω n + ϕQ (ω )).
yQ (n) =
IQI estimation. Reference [20] presents the IQ calibration
method for ultra-wideband SDR zero-IF receivers, based We denote the gain of the I and Q channels as GI(ω)
on utilization of complex valued filters. The receiver has and GQ(ω), respectively. φI(ω) and φQ(ω) are correspond-
RADIOENGINEERING, VOL. 32, NO. 4, DECEMBER 2023 481

ing channel phases. ω is the normalized angular frequency come this, the operation of complex-valued filters is re-
given by: quired [9].
fC The amplitude response g(ω) of the circuit with real
ω = 2π (2) valued components is modeled by even function:
fS
g (ω ) =
a0 + a2ω 2 + a4ω 4 . (6)
where fC is the signal carrier frequency and fS is the sam-
pling frequency. Transceiver’s amplitude response should Symmetric form for IQI gain is given by:
be a constant value, at least in the pass-band of interest.
γ even (ω ) =
b0 + b2ω 2 + b4ω 4 . (7)
The amplitude function GI(ω) is taken for a gain function,
denoted with g(ω). With the increase of signal bandwidth, The asymmetric gain IQI function γ(ω) is:
and also, with the increase of the transceiver's LO central
frequency, the g(ω) becomes frequency dependent and γ (ω ) =b0 + b1ω + b2ω 2 + b3ω 3 + b4ω 4 . (8)
improvement of g(ω) flatness is required. The symmetric phase imbalance function is modeled
The IQI is described via imbalance gain and phase with odd polynomial:
functions. The IQI gain function γ(ω) is defined by [1]: ϕodd (ω ) =c0 + c1ω + c3ω 3 . (9)
G (ω )
γ (ω ) = I . (3) The phase IQI model is modified to have asymmetric
GQ (ω ) form:
ϕ (ω ) =c0 + c1ω + c2ω 2 + c3ω 3 + c4ω 4 . (10)
The phase imbalance φ(ω) is defined as the difference
between I and Q component phases [1] The coefficients with index zero in the polynomials
ϕ (ω ) = ϕI (ω ) − ϕQ (ω ). (4) given by (6–10) represent the static gain and phase values.
While static gain IQI is taken into account by desired am-
The imbalance has a significant impact on the trans- plitude responses, static phase imbalance is omitted from
ceiver performance, which leads to the incomplete image desired phase responses and is cancelled by a static phase
signal rejection. For quantification of IQI effects the image correction block.
rejection ratio (IRR) can be used. The IRR is defined as the In order to find the g(ω), γ(ω) and φ(ω) coefficients
ratio between the intermediate-frequency (IF) signal ampli- sets of measurements are performed at angular frequencies
tude, produced by the desired input frequency and signal ωi, i = 0,1,...,N, uniformly distributed over baseband band-
amplitude generated by the image frequency. For a given width range. The frequency values are selected on both
γ(ω) and φ(ω), the IRR is equal to [1]: sides around the DC. For each ωi, the values for gi, γ i and
φi are determined. The method for gi, γi and φi calculation
1 + γ (ω ) + 2γ (ω ) cos ϕ (ω )
2

IRR (ω ) = 10 log . (5) will be explained later in detail in the implementation sec-
1 + γ (ω ) − 2γ (ω ) cos ϕ (ω )
2
tion. Based on measured values gi, γi and φi, the polynomial
coefficients are calculated after mean square error (MSE) is
With increase of signal bandwidth, γ(ω) and φ(ω) be- minimized between measured data and polynomial models.
come FS and IQI must be neutralized in the baseband using The g(ω) coefficients are found after the following system
digital filters. This is the approach we follow. However, of equations is solved:
the filter design can be simplified when polynomial models
for g(ω), γ(ω) and φ(ω) are designed first. Namely, many  s0 s2 s4   a0  T
s N N N

points of the desired filter amplitude and group delay re-  2 s4 s6   a2  =  ∑ gi ∑ giωi2 ∑ giωi4  . (11)
sponses will not be measured. Instead, they are calculated  s4
=  i 0=i 0 =i 0 
s6 s8   a4 
based on previously found models g(ω), γ(ω) and φ(ω).
The advantage of this approach resides in reduced number The elements sj are defined as:
of measurement points which speeds up the filter design N
and whole calibration process. s j = ∑ ωi j . (12)
i =0
The circuits with real valued components have posi-
tive symmetrical amplitude response around the DC, and Similarly, the γ(ω) and φ(ω) coefficients are found
also, negative symmetrical phase response [2]. In this case, after systems of equations given by (13), (14) are solved:
γ(ω) is constrained to be an even polynomial of ω, while T
T  N

φ(ω) (after removal of the DC phase offset) an odd poly- [ sk +l −2 ]k ,l =1,...,5 b j −1  j =1,...,5 = ∑ γ iωi j −1  , (13)
nomial. The correction circuits can be constructed using  i =0  j =1,...,5
two real valued digital filters which are positioned in I and T
Q signal paths. However, in the RF, the symmetry may be T N 
degraded. It was seen by measurements that utilized SDR [ sk +l −2 ]k ,l =1,...,5 c j −1  j =1,...,5 =  ∑ ϕiωi j −1  . (14)
 i =0  j =1,...,5
transceiver reveals asymmetric γ(ω) and φ(ω). To over-
482 B. JOVANOVIC, S. MILENKOVIC, IQ IMBALANCE CORRECTION IN WIDEBAND SOFTWARE DEFINED RADIO TRANSCEIVERS

3.2 The FIR Filter Specification

The block diagram of the circuit dedicated for IQ cor-
rection is depicted in Fig. 1. The correction scheme is
based on a complex FIR filter adopted from [9]. The circuit
is composed of four real valued FIR filters. The FIR_II and
FIR_QQ are positioned in I and Q signal paths. The other
two, the FIR_IQ and FIR_QI, are located in the cross
paths. The digital filters have length M. Besides, separate
correctors are dedicated for receiver and transmitter IQI
mitigation.
For the construction of FIR_IQ and FIR_QI the digi-
Fig. 1. The IQ compensator architecture.
tal differentiator FIR filters are used whose length M is
an odd number. The frequency response of a digital differ-
entiator is given by [23]:

H diff (ω =
) K ⋅ jω (15)

where the parameter K represents its gain. The impulse

response is given by [23]:

 cos (π ( n − α ) )
K ,n ≠α
hdiff ( n ) =  n −α (16)
 0, n = α

where α = (M − 1)/2. In an effort to improve amplitude
response linearity the filter coefficients from (16) are modi-
fied by the Hamming window function [23].
In IQI analysis we assume that the signals at the in-
puts XI and XQ (given in Fig. 1) are equal to cos(ω0n) and
Fig. 2. The frequency components of a) the FIR_IQ input,
sin(ω0n) and the K is a positive value. Also, we assume b) the FIR_IQ output, c) FIR_QI input and d) FIR_QI
that the filters FIR_II and FIR_QQ are bypassed (their output.
outputs are delayed for the delay of FIR filters and not
K(ω) is calculated by:
changed in gain and phase).
Using Euler’s complex expansion of cosine and sine b1 + b3ω 2
K (ω ) = . (19)
functions, the signals XI and XQ can be represented as 2 + b1ω + b3ω 3
a sum of two components positioned at ω0 and −ω0:
The asymmetrical φeven(ω) is corrected by a phase im-
1 1 balance produced by a time delay between FIR_IQ and
X=I cos (ω0 n=
) exp ( jω0 n ) + exp ( − jω0 n ) ,
2 2 (17) FIR_QI impulse responses. Namely, when impulse re-
j j sponse of FIR_QI is delayed relative to the response of
X Q = sin (ω0 n ) = exp ( − jω0 n ) − exp ( jω0 n ) . FIR_IQ, the signal at I path output is given by:
2 2
YI = cos (ω n ) + ω K (ω ) cos(ω n − 2πω ⋅ delay ) =
The FIR_IQ modifies XI into XIQ, while the FIR_QI
produces at its output XQI. The graphical representation of cos (ω n ) ⋅ (1 + ω K (ω ) cos(2πω ⋅ delay )) + (20)
XI, XQ, XIQ and XQI decomposition is shown in Fig. 2. The
sin (ω n ) ⋅ ω K (ω ) sin(2πω ⋅ delay ).
XIQ components amplitudes are 0.5Kω0 and −0.5Kω0, at ω0
and −ω0. In the case of XQI they are both equal to 0.5Kω0. The introduced phase shift neutralizes the φeven(ω):
The FIR_QI increases the gain in I path by 0.5Kω0 while
the Q path gain is reduced by the same amount. The gain ϕeven (ω ) = c2ω 2 + c4ω 4 =
imbalance is (1 + Kω0) / (1 − Kω0).
 ω K (ω ) ⋅ sin ( 2πω ⋅ delay )  (21)
The gain imbalance, induced by FIR_IQ and FIR_QI, arctan   .
is equalized with the odd gain imbalance function γodd(ω):  1 + ω K (ω ) ⋅ cos ( 2πω ⋅ delay ) 

1 + ω K (ω ) The delay value, expressed in 1/fS units, is calculated

1+ γ odd (ω ) = 1+ b1ω + b3ω 3 = . (18) from (21) and it is used for the construction of a fractional
1 − ω K (ω ) delay (FD) FIR filter [21] whose frequency response is
RADIOENGINEERING, VOL. 32, NO. 4, DECEMBER 2023 483

denoted with Hsinc(ω, delay). The FD filter and digital 3.3 Amplitude and Phase Response, the
differentiator share the same filter length. Group Delay of FIR Filter
The impulse response hsinc(n, delay) is given by [23]: The transfer function of M tap FIR filters is given by:
hsinc ( n, delay ) = M −1
H ( z ) = ∑ hk z − k . (29)
 sin (π ( n − α − delay ) ) (22) k =0
 , n − α − delay ≠ 0,
 n − α − delay The frequency response of the filter has the form:
 1, n − α − delay = 0.

( ω ) R ( ω ) − jI ( ω ) ,
H= (30)
The amplitude response K(ω) from (21) is approxi-
mated by an FIR filter named FIR_K. The desired ampli- while the real and imaginary parts are calculated as:
tude and phase responses of FIR_K are specified by: M −1
R (ω ) = ∑ hk cos ( kω ),
AFIR _ K (ω ) = K (ω ), k =0 (31)
ϕFIR _ K (ω ) = 0, (23) M −1
I (ω ) = ∑ hk sin ( kω ).
τ FIR _ K (ω ) = 0. k =0

The FIR_IQ and FIR_QI are produced by convolution The amplitude and phase of the complex function are:
of three filters: the digital differentiator, FIR_K and FD:
A (ω ) R (ω ) + I (ω ) ,
2 2
=
H FIR_IQ (ω , delay ) =jω ∗ K (ω ) ∗ H sinc (ω , 0), (32)
(24) I (ω )
H FIR_QI (ω ) =jω ∗ K (ω ) ∗ H sinc (ω , delay ). ϕ (ω ) = arg H (ω ) = − arctan .
R (ω )
The odd function φodd(ω) is compensated by FIR_II
whose amplitude response contains the inverse of g(ω). Finally, the normalized group delay is given as:
The FIR_II filter is described by: ∂ϕ (ω ) R (ω ) Rk (ω ) + I (ω ) I k (ω )
1
τ (ω ) =
− =
− , (33)
R (ω ) + I (ω )
2 2
g FIR_II (ω )
= = , ϕ FIR_II (ω ) ϕodd (ω ) . (25) ∂ω
g (ω )
M −1

The γeven(ω) is corrected by FIR_QQ. The amplitude Rk (ω ) = ∑ hk k cos ( kω ),

response of FIR_QQ contains the inverse of g(ω). The
k =0 (34)
M −1
desired amplitude and phase responses are given by: I k (ω ) = ∑ hk k sin ( kω ).
γ even (ω )
k =0

g FIR_QQ (ω ) = ,
g (ω ) (26)
3.4 Iterative Procedure for Coefficients
ϕFIR_QQ (ω ) = 0.
Calculation
Splitting the operations for correction of γeven(ω) and Identical numerical optimization procedures are used
φodd(ω) between FIR_II and FIR_QQ relaxes filter specifi- for construction of FIR_K, FIR_II and FIR_QQ. The coef-
cation, and consequently, reduces the number of filtering ficients are determined under two constraints. In the first,
taps. By substituting γeven(ω) and φodd(ω), the FIR_II ampli- the amplitude response A(ω) from (32) should approximate
tude and group delay become: AFIR_K(ω), AFIR_II(ω) and AFIR_QQ(ω) for filters FIR_K,
1 FIR_II and FIR_QQ, respectively. The second constraint
AFIR_II (ω ) = , considers the group delay τ(ω) from (33) which is made to
a0 + a2ω 2 + a4ω 4 (27) be as close as possible to τFIR_K(ω), τFIR_II(ω) and
1 dϕdI (ω ) c1 + 3c3ω 2 τFIR_QQ(ω).
τ FIR_II (ω ) =
− = .
In the signal pass band, P frequency points are se-
Ts dω 2π
lected. Ps − 1 out of band frequency points are chosen as
Similarly, the FIR_QQ amplitude and group delay are well and the desired amplitude response is constrained for
specified by: A(ω). Out-of-band amplitude constraint prevents the result-
ing amplitude response to have large out-of-band gain
b0 + b2ω 2 + b4ω 4
AFIR_QQ (ω ) = , which may produce data overflow when the filter is im-
a0 + a2ω 2 + a4ω 4 (28) plemented. An arbitrary weighting function Aw(ω) is used
τ FIR_QQ (ω ) = 0. to control the approximation accuracy in certain frequency
bands. For example, out-of-band Aw(ω) can be set to a very
 484 B. JOVANOVIC, S. MILENKOVIC, IQ IMBALANCE CORRECTION IN WIDEBAND SOFTWARE DEFINED RADIO TRANSCEIVERS

small value which in turn improves in-band amplitude The matrix I(i) is updated with [24]:
response approximation.
∆h∆h T I ( i ) ∇g∇g I ( i )
T

Group delay constraint is defined in a similar way. I ( i + 1)= I ( i ) + − . (41)

Here, τ(ω) and τw(ω) represent target group delay and ∆h T ∇g ∇g T I ( i ) ∇g
group delay weighting function, respectively. The group
delay constraint is defined only for the signal pass band.
Based on the above definitions, the cost function is 4. Implementation
constructed to minimize the difference between the desired The IQI correction method is validated through prac-
response and that measured from I and Q channels respec- tical hardware implementation in the LimeSDR QPCIe
tively. SDR board [25]. The transceiver, whose IQ imbalance is
λ P −1 calibrated, is LMS7002M IC [25]. The transceiver covers
∑τ (ω ) (τ (ω ) − τ (ω ) − β )
2
=E w l l d l the central frequency range of several hundreds of MHz to
2 l =1 (35) 3.8 GHz [25]. The PC is equipped with the SDR board
(1 − λ ) P + P −1 A ω A ω − A ω 2 . inserted in the peripheral component interconnect express
∑ w ( l ) ( ( l ) d ( l ))
s

+ (PCIe) slot. Beside the SDR, the measurement setup in-

2 l =1
cludes the spectrum analyzer and RF signal generator
There are two more coefficients introduced in (35). which are used in IRR measurements. In the setup the SDR
Namely, λ is the penalty factor which defines relative im- TX output is connected via RF cable to the input of
portance of the amplitude and group delay constraints Keysight E4440 spectrum analyzer. The Keysight E8267D
while β is an additional parameter which represents laten- signal generator output is connected to RX input of the
cy. It is fixed here to (M − 1)/2 and is not changed during SDR board. The signal generator and spectrum analyzer
the optimization. The optimization problem defined by (35) are not used during the calibration process; they are uti-
is nonlinear and Davidon-Fletcher-Powell (DPF) method is lized during IRR measurements.
used to solve it numerically [24].
The software consists of the LimeSuite [25], the soft-
The FIR coefficients are stored in the vector h, which ware implementing digital modulator and the application
elements are changed in an iterative DPF process. dedicated to IQ calibration. The SDR is configured using
LimeSuite. The configuration files are loaded into
h(i ) = [ h0 (i ) h1 (i )... hM −1 (i )] . (36) LMS7002M ICs [25], the on-board DAC and ADC sample
The required DPF partial derivatives are calculated as rates are set to 245.76 MHz and 122.88 MHz, respectively.
follows: Upon loading the configuration files, the SDR transmitter
and receiver are in uncalibrated state. The software imple-
∂ε P −1
∂τ (ωl ) menting the digital modulator functions generates
= λ ∑τ w (ωl ) (τ (ωl ) − τ d (ωl ) − β ) 100 MHz wideband waveforms at a rate of 122.88 MS/s
∂hk l =1 ∂hk (37) and it is explained in [26] in detail.
P + Ps −1
∂A (ωl )
+ (1 − λ ) ∑ Aw (ωl ) ( A (ωl ) − Ad (ωl ) ) . The SDR board, whose block diagram is depicted in
l =1 ∂hk Fig. 3, includes two transceiver ICs, an Altera Cyclone V
FPGA chip, 14-bit external ADC and DACs [25]. To in-
The starting vector h(0) has all zero value elements.
crease the capacity of a radio link, two-by-two multiple-
The DPF utilizes the identity matrix I which has M × M input and multiple-output (MIMO) transceiver is imple-
dimension. In every iteration I(i) elements are changed; in mented on SDR board. The SDR board has many other
the starting point I(0) has all zero elements, except the options than shown in Fig. 3. For clarity, only minimum
elements on the main matrix diagonal which are equal to hardware blocks are presented in the figure, which are
one. The gradient function is [24]: relevant for a method description.
T
 ∂f ∂f ∂f  The signals, generated by digital modem [26], are fed
∇f ( h ( i ) ) =
 ...  . (38)
into the SDR board via PCIe at a rate of 122.88 MS/s and
 ∂h0 ∂h1 ∂hM −1  h =h( i )
they are processed by transmit path (TX) signal processing
In iteration the vector h elements are changed for [24]: blocks, as shown in Fig. 3.
An oversampling of factor one is used before the
∆h(i ) = h(i + 1) − h(i ) = −η I ∇f (h(i )) . (39)
DACs, yielding a data rate of 245.76 MS/s. The data rate is
The parameter η is calculated from the constraint that constrained by DAC rate maximum of 250 MS/s. Transmit-
the f(h(i + 1)) is minimized [24]. The change of gradient ted signal bandwidth is 100 MHz and the interpolation
function value is [24]: block is required to eliminate unwanted DAC-related sig-
nal images. At the transmit side, the IQ samples pass
∆g (i ) = ∇f (h(i + 1)) − ∇f (h(i )) . (40) through the signal interpolation block, the TX static IQI and
RADIOENGINEERING, VOL. 32, NO. 4, DECEMBER 2023 485

coefficients via the Serial Peripheral Interface (SPI). The

coefficients are sent to the FPGA over the PCIe/SPI. In
order to automate the IRR measurement operations, the
functionality of software application is extended to control
spectrum analyzer and RF signal generator via LAN con-
nection. The application performs the RX and TX IRR
measurements. After calibration is finished, the filter coef-
ficients can be stored in a file. When the transceiver is
powered up, the calibration process can be skipped. The
coefficients are read from the recorded file and loaded into
the complex filter registers.

4.1 The Calibration Routines

The receiver and transmitter imbalances are separate-
ly extracted. A calibration setup is used that was presented
Fig. 3. The implementation platform. in [16]. This setup uses the single-frequency tones to cali-
DC offset correctors, followed by IQI compensators. The brate the receiver and then proceeds to use the calibrated
IQI correction filters operate at a sample rate of receiver as a measuring device. The transmitter supplies
245.76 MS/s. The 18-bit arithmetic precision does not the test tones into the receiver over RF loopback. The test
impact the algorithm performance. The TX IQI compensa- tones are generated either by transmitter LO or by numeri-
tor is composed of four 15-tap FIR filters whose structure cally controlled oscillator (NCO) which is embedded in the
is given in Fig. 1. For realization of one FIR filter the fol- TX static calibration block. The amplitudes of the signal
lowing FPGA resources are spent: 330 adaptive logic and corresponding image tone are determined using spec-
modules (ALM), 350 combinatorial adaptive look-up ta- tral analysis. Besides, the static TX and RX IQ correctors
bles (ALUT), 700 dedicated registers and 9 DSP blocks. are configured to minimize the amplitudes of unwanted
Frequency up conversion from baseband (BB) to RF and images.
down conversion from RF to BB, are performed by trans- The block diagrams of RX and TX static IQ correc-
mitter and receiver chains. tors are presented in Fig. 4. The transmitter's static IQ
At the receiving side, the signals are sampled by 14- corrector parameters include gain correction codes
bit ADCs at a rate of 122.88 MS/s. The data rate is con- (txGain_I and txGain_Q) and phase correction code (the
strained by on-board ADC maximum data rate of txAlpha). Similarly, the RX static IQ corrector parameters
160 MS/s. Signals are further processed by digital blocks include rxGain_I, rxGain_Q and rxAlpha.
implemented in FPGA. The receive (RX) chain blocks
include the static IQ and DC correctors followed by RX
IQI correction circuit which share identical structure as the
circuit in TX path.
The IQI calibration procedures are executed at trans-
ceiver start-up. The board incorporates only a few addi-
tional circuits supporting the calibration process. These
circuits consist of the RF switches that form the RF loop-
back path from transmitter output to receiver input. There-
fore, the RF loopback introduces only minimal modifica-
tions in SDR hardware (i.e. no additional mixers, ADCs or
other expensive monitoring equipment). In the calibration
process the transmitter plays the role of a test signal gener-
ator while the receiver is used as a monitoring device. In
each measurement point the IQI parameters are extracted
by analyzing the received signal spectrum.
The complex filter coefficients are calculated by
C/C++ software application that runs on a CPU core. Inter-
nally, the algorithm operations can be divided into calibra-
tion routines, the coefficient’s calculation operations and
FIR filter programming operations. After filter coefficients
are calculated they are transferred to the SDR board via
PCIe. The correction filters have provision to change their Fig. 4. Static IC IQI and DC offset correctors.
486 B. JOVANOVIC, S. MILENKOVIC, IQ IMBALANCE CORRECTION IN WIDEBAND SOFTWARE DEFINED RADIO TRANSCEIVERS

The following BB frequencies are selected for meas- to the target RX_fLO + fi. The image signal is minimized by
urement points fi [MHz] = {−50, −45, −40, −30, −20, −10, employing the RX static IQ corrector. For each fi, the fol-
10, 20, 30, 40, 45, 50}. The angular frequency values ωi, lowing corrector codes are acquired: rxGain_3_Ii,
i = 0,1, …,11, are derived after fi values are divided with rxGain_3_Qi and rxAlpha_3i. The rx_γi and rx_φi are cal-
sample rate values. For the transmitter, the sample rate is culated by (44), (45). The rx_gi is derived from the ampli-
245.76 MS/s while the receiver sample rate is equal to tude of the signal positioned at fi..
122.88 MS/s. The calibration operations can be divided in
phases designated as 1, 2 and 3. The calibration results are
txGain _ 3 _ Qi (44)
rx _ γ i = ,
the values for gain - gi and IQI parameters - γi and φi, txGain _ 3 _ I i
which are obtained for each ωi. The phases are explained
 txAlpha _ 3i  (45)
as follows. rx _ ϕi = 2 ⋅ arctan  .
 2048 
The calibration phase 1 is the preparation step for
phase 2. In phase 1 the receiver is calibrated. The test tones The gi, γi and φi values, gained for RX and TX at fi,
are generated by transmitter LO. At the beginning of are used for calculation of the g(ω), γ(ω) and φ(ω) (using
phase 1, the RX LO frequency is fixed to target TX LO (11–14)). The mathematical background for conversion of
frequency (the target TX fLO) and is not changed further the g(ω), γ(ω) and φ(ω) into the coefficients of FIR filters
during phase 1. The RX DC offset is corrected. (the RX and TX FIR_II, FIR_QQ, FIR_K) is described in
the previous section.
The DC correction codes rxDC_1_I and rxDC_1_Q
are calculated and programmed in the RX static IQ correc-
tor (shown in Fig. 4). In each point, the test tone is gener- 4.2 Transmitter Measurement
ated by tuning the TX LO to the frequency equal to target
TX_fLO+fi, where target TX_fLO is a constant. After the RF The method performance is estimated in cases when
loopback is established and data is read, a spectral analysis transmitter is tuned to different LO frequencies: TX_fLO =
is performed over the received data. A desired signal com- {2.0 GHz, 2.3 GHz, 2.6 GHz, 2.9 GHz, 3.2 GHz and
ponent, observed at f = fi, is accompanied by an undesired 3.5 GHz}. The calibration is performed for all fLO from the
image, positioned at f = −fi. The image signal amplitude is set. The routines, described in Sec. 4.1, are executed for
un-calibrated and calibrated transmitter in order to calcu-
minimized by changing the parameters of the RX static IQ
late the TX IQ gain and phase errors. The TX IQ gain and
corrector. The phase 1 outputs, which correspond to fi, are
phase imbalance values are determined by (42) and (43).
the RX gain (rxGain_1_Ii, rxGain_1_Qi) and phase
(rxAlpha_1i) correction codes. The described operations Normalized amplitude response, as a function fLO and
are repeated for all fi. fBB, is given in Fig. 5. The curve labeled as “Meas.” is
obtained by measured data before the proposed method is
In phase 2 the transmitter is calibrated. The test tones
applied. The results, which are labeled as “Corr.” curves,
are generated by TX NCO located in the static TX IQ cor-
represent corrected results, derived after the IQI is mini-
rector block (Fig. 4). At phase 2 beginning, the TX and RX
mized. When amplitude response is considered, the results
LO are set to target TX_fLO. Then, the TX DC offset is
presented in the figure prove that the corrected amplitude
calibrated. In each measurement point, the NCO frequency
response is flatter than the measured one. However, the
is set to fi and RX IQ correction codes rxGain_1_Ii,
corrected results are not ideally flat because only even
rxGain_1_Qi, rxAlpha_1i are loaded into the RX IQ correc-
polynomials are compensated.
tor. The RF loopback is formed and the amplitudes of sig-
nals positioned at fi, and −fi are measured. The image sig- Measured and corrected IQI phase values are shown
nals are reduced by adjusting the TX static IQ corrector. in Fig. 6. In the most challenging case, when TX_fLO =
The phase 2 results are correction codes for TX IQ gain 3.5 GHz, the phase imbalance ranges from 8 degrees, ob-
(txGain_2_Ii, txGain_2_Qi) and phase (txAlpha_2i). The tained for fBB = −50 MHz, up to 16 degrees, gained for
IQI parameters tx_γi and tx_φi are calculated by (42), (43). fBB = 50 MHz. As it can be seen from the figure, the cor-
The tx_gi is found as the amplitude of the signal positioned rected phase IQI is efficiently minimized, regardless of
at fi, normalized to the amplitude maximum. tuned TX_fLO.
txGain _ 2 _ Qi (42) The gain IQI results, measured at different TX_fLO,
tx _ γ i = ,
txGain _ 2 _ I i are given in Fig. 7. At lower TX_fLO frequencies (up to
2.6 GHz) the measured IQI curves are symmetrical. When
 txAlpha _ 2i  (43) TX_fLO is increased, the symmetry is degraded. After the
tx _ ϕi = 2 ⋅ arctan  .
 2048  method is applied, the gain IQI is efficiently reduced; the
corrected values are very close to the ideal case.
In phase 3, the RX IQI calibration procedure is re-
peated. The operations are similar to those executed in The IRR results are acquired by measurements in
phase 1. The transmitter LO is again used as a tone genera- which single tone test signals are generated by SDR
tor. The RX LO is now tuned to the target RX fLO. The RX transmitter NCO. The baseband signal is in the range
DC offset is calibrated. In each point, the TX LO is tuned fBB = [−50 MHz, 50 MHz]. The transmitter output is observed
RADIOENGINEERING, VOL. 32, NO. 4, DECEMBER 2023 487

1.2 1.2
Meas. 2.0GHz Corr. 2.3GHz Meas. 2.0GHz Corr. 2.3GHz
Corr. 2.0GHz Meas. 2.6GHz 1.15 Corr. 2.0GHz Meas. 2.6GHz
1.1
Normalized ampl. []

Meas. 2.3GHz Corr. 2.6GHz Meas. 2.3GHz Corr. 2.6GHz

1.1

IQ gain error []
1
1.05

0.9 1
0.95
0.8
0.9
0.7 0.85
-40 -20 0 20 40 -40 -20 0 20 40
Baseband frequency [MHz] Baseband frequency [MHz]
a) a)

1.2
Meas. 2.9GHz Corr. 3.2GHz 1.2
Meas. 2.9GHz Corr. 3.2GHz
Corr. 2.9GHz Meas. 3.5GHz
1.1 1.15 Corr. 2.9GHz Meas. 3.5GHz
Normalized ampl. []

Meas. 3.2GHz Corr. 3.5GHz

1.1 Meas. 3.2GHz Corr. 3.5GHz

IQ gain error []
1
1.05
0.9 1

0.8 0.95
0.9
0.7
-40 -20 0 20 40 0.85
Baseband frequency [MHz] -40 -20 0 20 40
b) Baseband frequency [MHz]
b)
Fig. 5. Transmitter normalized measured and corrected
amplitude response obtained at a) 2.0 GHz, 2.3 GHz, Fig. 7. Transmitter IQ gain imbalance at: a) 2.0 GHz,
2.6 GHz, b) 2.9 GHz, 3.2 GHz and 3.5 GHz. 2.3 GHz, 2.6 GHz, b) 2.9 GHz, 3.2 GHz, 3.5 GHz.

30
Meas. 2.0GHz Corr. 2.3GHz 10
Meas. 2.0GHz Corr. 2.3GHz
25 Corr. 2.0GHz Meas. 2.6GHz 0 Corr. 2.0GHz Meas. 2.6GHz
Image suppression [dBc]
IQ phase error [deg]

20 Meas. 2.3GHz Corr. 2.6GHz -10 Meas. 2.3GHz Corr. 2.6GHz

15 -20
-30
10
-40
5
-50
0
-60
-5 -70
-40 -20 0 20 40 -40 -20 0 20 40
Baseband frequency [MHz] Baseband frequency [MHz]
a) a)
30
Meas. 2.9GHz Corr. 3.2GHz 10
25 Corr. 2.9GHz Meas. 3.5GHz Meas. 2.9GHz Corr. 3.2GHz
0
IQ phase error [deg]

Meas. 3.2GHz Corr. 3.5GHz Corr. 2.9GHz Meas. 3.5GHz

Image suppression [dBc]

20
-10 Meas. 3.2GHz Corr. 3.5GHz
15 -20
10 -30
5 -40

0 -50
-60
-5
-40 -20 0 20 40 -70
Baseband frequency [MHz] -40 -20 0 20 40
b) Baseband frequency [MHz]
b)
Fig. 6. Measured and corrected transmitter phase IQI
functions acquired at: a) 2.0 GHz, 2.3 GHz, 2.6 GHz, Fig. 8. TX IQI image level suppression at: a) 2.0 GHz,
b) 2.9 GHz, 3.2 GHz and 3.5 GHz. 2.3 GHz, 2.6 GHz, b) 2.9 GHz, 3.2 GHz and 3.5 GHz.
488 B. JOVANOVIC, S. MILENKOVIC, IQ IMBALANCE CORRECTION IN WIDEBAND SOFTWARE DEFINED RADIO TRANSCEIVERS

using a spectrum analyzer whose central frequency is set to

Meas. 2.0GHz Corr. 2.3GHz
TX_fLO. The image suppression level (the inverse of IRR)
5

IQ phase imbalance [deg]

Corr. 2.0GHz Meas. 2.6GHz
as a function of fLO and fBB is shown in Fig. 8. The ampli- Meas. 2.3GHz Corr. 2.6GHz
tude of the main test signal, positioned at TX_fLO + fBB, and
the image component, positioned at TX_fLO − fBB, are 0
measured. The IRR is determined by the difference be-
tween the amplitudes of the signal and its image. The IRR -5
results clearly demonstrate an improvement in image rejec-
tion. When fLO = 3.5 GHz and the method is not applied,
the IRR is only 20 dBc. When IQI is reduced and the trans- -10
ceiver operates at the same fLO, the IRR values ranges from -40 -20 0 20 40
45 to 55 dBc. The enhancement in IRR is more than 25 dBc. Baseband frequency [MHz]
a)

4.3 Receiver Measurement

Meas. 2.9GHz Corr. 3.2GHz
The SDR receiver is calibrated and IQ imbalance is 5

IQ phase imbalance [deg]

Corr. 2.9GHz Meas. 3.5GHz
measured at LO frequencies RX_fLO={2.0 GHz, 2.3 GHz, Meas. 3.2GHz Corr. 3.5GHz
2.6 GHz, 2.9 GHz, 3.2 GHz and 3.5 GHz}. The measure- 0
ment routines, described in Sec. 4.1, are separately per-
formed for un-calibrated and calibrated receiver in order to
get measured and corrected RX IQ gain and phase imbal- -5
ance values. The RX gain and phase imbalances are calcu-
lated using (44) and (45), respectively. -10
Normalized receiver amplitude response, as a func- -40 -20 0 20 40
tion of fLO and fBB, is given in Fig. 9. The phase imbalance Baseband frequency [MHz]
b)
values are depicted in Fig. 10. As it can be seen from
Fig. 10, at fLO = 3.5 GHz the measured phase imbalance Fig. 10. RX phase IQI measured at a) 2.0 GHz, 2.3 GHz,
function shows asymmetry. The corrected phase imbalance 2.6 GHz; b) 2.9 GHz, 3.2 GHz and 3.5 GHz.
values prove method efficiency in IQI reduction. The RX
gain IQI is depicted in Fig. 11. 1.08
Meas. 2.0GHz Corr. 2.3GHz
1.06
1.3 Corr. 2.0GHz Meas. 2.6GHz
1.04
IQ gain imbalance []

Meas. 2.0GHz Corr. 2.3GHz Meas. 2.3GHz Corr. 2.6GHz

1.2 Corr. 2.0GHz Meas. 2.6GHz 1.02
Normalized ampl. []

1.1 Meas. 2.3GHz Corr. 2.6GHz 1

0.98
1
0.96
0.9
0.94
0.8 0.92
0.7 0.9
-40 -20 0 20 40
0.6 Baseband frequency [MHz]
-40 -20 0 20 40 a)
Baseband frequency [MHz]
a)
1.08
1.3 Meas. 2.9GHz Corr. 3.2GHz
Meas. 2.9GHz Corr. 3.2GHz 1.06
Corr. 2.9GHz Meas. 3.5GHz
1.2 Corr. 2.9GHz Meas. 3.5GHz 1.04
IQ gain imbalance []

Meas. 3.2GHz Corr. 3.5GHz

Normalized ampl. []

1.1 Meas. 3.2GHz Corr. 3.5GHz

1.02
1 1
0.98
0.9
0.96
0.8
0.94
0.7 0.92
0.6 0.9
-40 -20 0 20 40 -40 -20 0 20 40
Baseband frequency [MHz] Baseband frequency [MHz]
b) b)
Fig. 9. RX normalized amplitude response at: a) 2.0 GHz, Fig. 11. RX gain IQI as a function of fBB at a) 2.0 GHz,
2.3 GHz, 2.6 GHz; b) 2.9 GHz, 3.2 GHz, 3.5 GHz. 2.3 GHz, 2.6 GHz; b) 2.9 GHz, 3.2 GHz, 3.5 GHz.
RADIOENGINEERING, VOL. 32, NO. 4, DECEMBER 2023 489

10 repeated. Special factory or laboratory calibration is not

Meas. 2.0GHz Corr. 2.3GHz required.
0 Corr. 2.0GHz Meas. 2.6GHz
Image suppression [dBc]

-10 Meas. 2.3GHz Corr. 2.6GHz Reduction of FIR filter length enables savings of
-20 FPGA resources, which makes the method suitable for
-30
realization in FPGA. It is worth mentioning that beside the
proposed FS IQ corrector, the other digital blocks are also
-40
required in transmitter paths, such as crest factor reduction
-50 (CFR) and post-CFR FIR filters. For implementation of
-60 these digital blocks, significant amount of FPGA resources
-70 is spent [27]. Moreover, the hardware is optimized to
-40 -20 0 20 40 occupy minimum resources on FPGA for transceiver 2×2
Baseband frequency [MHz]
a) MIMO operation.
When gain and phase imbalance functions are sym-
10 metrical a real-valued digital filter can be utilized for FS IQ
Meas. 2.9GHz Corr. 3.2GHz
0 Corr. 2.9GHz Meas. 3.5GHz mitigation. However, when these criteria are not fulfilled,
Image suppression [dBc]

-10 Meas. 3.2GHz Corr. 3.5GHz a complex filter is required. As measured results proved, at
-20 central frequencies greater than 3 GHz, the TX and RX
gain imbalance functions are asymmetrical. In order to
-30
reduce the asymmetry, compared to the symmetrical form
-40 given by (7), the adopted γ(ω) is designed to incorporate
-50 odd polynomial elements with exponents equal to 1 and 3
-60 (8). The results showed that the receiver φ(ω) becomes
-70
asymmetrical at fLO greater than 3 GHz. At same frequen-
-40 -20 0 20 40 cies the TX φ(ω) doesn't possess this property. Compared
Baseband frequency [MHz]
b) to the symmetric form given in (9), the implemented φ(ω)
(10), includes even polynomial parts whose exponents are
Fig. 12. The RX IQ image suppression at a) 2.0 GHz, 2.3 GHz, equal to 2 and 4.
2.6 GHz; b) 2.9 GHz, 3.2 GHz and 3.5 GHz.
The complex filter is designed based on previously
The RX image suppression level is shown in Fig. 12.
determined IQ imbalance models. It consists of four real
Single tone signals are produced by RF signal generator
valued FIR filters. Two FIR filters are positioned in I and
starting from RX_fLO −50 MHz to RX_fLO +50 MHz with
Q paths. The others are located in cross paths (see Fig. 1).
the increment step of 5 MHz. The amplitude of test signals
Digital differentiator and fractional delay FIR filters are
is equal to −30 dBm. The signals are then fed to the input
used for realization of cross path filters. They cancel the
of SDR receiver which LO is tuned to RX_fLO. At each fBB,
asymmetric portion of IQI. Specifically, the gain of cross
the amplitudes of the main test tone, positioned at
path filters compensates the odd part of γ(ω). The even part
RX_fLO+ fBB, and the image component, positioned at
of φ(ω) function is neutralized by the phase difference
RX_fLO − fBB are determined.
which is introduced by the delay between cross path filter
impulse responses.

5. Discussion The method performance is assessed at different fLO,

starting from 2 GHz to 3.5 GHz. The IRR of the un-
A novel method for correction of the FS IQI is pre- calibrated transmitter ranges between 20 dBc and 35 dBc,
sented. The method relies on utilization of single tone test depending on fLO. Different configurations of
signals and spectral analysis for IQI extraction. Both re- compensation circuits are investigated. First, the
ceiver and transmitter IQ imbalances are calibrated for the performance is estimated of the circuit consisting only of
baseband signal bandwidth of 100 MHz. two real valued FIR filters, positioned in I and Q signal
The IQI is reduced and the amplitude response is flat- paths. The measured results proved that when transceiver
tened. The advantage of the proposed method compared to LO is tuned at higher LO frequencies the utilization of real
the methods found in literature is that the hardware over- valued FIR_II and FIR_QQ cannot remove the IQ images.
head is minimized. This is achieved by employing the For example, at fLO = 3.5 GHz, after calibration is done, the
transmitter as a test signal generator. Also, the receiver resulting TX IRR is only 38 dBc. The utilization of
plays the role of a measuring device. complex filters gives better results. When the proposed
method is applied, the corrected amplitude response
The RF signal loopback, implemented by on-board becomes flatter than in the case of uncalibrated transmitter.
RF switches, represents the only additional hardware Also, the corrected phase and gain IQI become closer to
which supports the calibration process. The calibration is the ideal values over the entire pass band of 100 MHz.
executed at the transceiver start-up and can be periodically When TX IQI is corrected (given in Fig. 1) the IRR is
490 B. JOVANOVIC, S. MILENKOVIC, IQ IMBALANCE CORRECTION IN WIDEBAND SOFTWARE DEFINED RADIO TRANSCEIVERS

improved by more than 25 dBc. The resulting IRR reaches 6. Conclusion

45–50 dBc.
In the paper the method for compensation of wide-
The RX measured results confirm improvement in IQ band transceiver IQ impairments is described. The method
image rejection. The measured IRR of uncalibrated corrects the imbalances of both the receiver and transmit-
receiver, operating at LO frequencies less than 3.5 GHz, is ter. The measured results that are presented in the paper are
approximately equal to 30 dBc. At fLO = 3.5 GHz, the IRR obtained after the method has been implemented in a SDR-
is only 20 dBc. After the proposed method is applied, the based RF transceiver. Particularly, the transceiver is used
corrected phase IQI values get close to zero degrees. Also, for transmission of the wideband modulation waveforms; it
the corrected γ(ω) gets near to the ideal case. By applying operates at a central frequency of 3.5 GHz where the
the RX IQ complex filter, the IRR becomes greater than transmitted signal bandwidth is 100 MHz. The advantage
50 dBc. The improvement at fLO = 3.5 GHz is more than of the method is that it does not require special hardware
30 dBc. for calibration operations. The RF loopback is added to
The methods from literature are focused on either re- support the calibration process. In calibration setup the
ceiver or transmitter calibration. In our case the imbalances transmitter is utilized as a test signal generator, while the
of both receiver and transmitter are corrected. The TX IQI receiver is a measuring device. The method performance is
mitigation techniques which are found in literature require assessed by IRR measurements. The IRR is measured in
implementation of additional hardware for IQI detection. cases when the transceiver is tuned to different LO fre-
The monitoring path is realized by external instruments or quencies, starting from 2.0 GHz to the target frequency of
dedicated ADCs [28]. The measured results are produced 3.5 GHz. The results prove efficient IQ imbalance suppres-
using laboratory equipment that relies on high-performance sion increasing the quality of wideband signals. The design
signal generators. Moreover, the calibration procedures are methodology is generic and the proposed solution is suita-
realized in MATLAB. In [29] the TX mitigation method ble for implementation in other field-programmable RF
improves the IRR over 45 dBc. When a ten-tone waveform base stations. For the future work the development of
is applied, the method yields increase in IRR by 10–15 dBc a method that mitigates the IQ impairments which appear
[29]. In [13] 10 dBc IRR enhancement is achieved for in massive MIMO transceivers is envisaged.
a baseband signal bandwidth of 80 MHz. The algorithm
gives more than 50 dBc IRR on both sidebands which indi-
cates that the FS IQI is mitigated well. In [18] the experi-
mental results give the receiver IRR improvement of Acknowledgments
30 dBc compared to uncompensated case. The disad-
vantage of the method from [18] is that it does not utilize This research was funded by The Ministry of Educa-
complex filters, and therefore, cannot neutralize asymmet- tion, Science and Technological Development of the Re-
ric IQI. The RX IQI reduction procedure from [20] pro- public of Serbia.
vides the IRR better than 65 dBc over 600 MHz bandwidth
for a transceiver operating in 1 GHz to 6 GHz frequency
range. The comparison of the proposed method with the
results found in literature is given in Tab. 1. References
[1] CAVERS, J. K., LIAO, M. W. Adaptive compensation for
imbalance and offset losses in direct conversion transceivers. IEEE
Transactions on Vehicular Technology, 1993, vol. 42, no. 4,
Ref. Ref. Ref. Ref. Pro- p. 581–588. DOI: 10.1109/25.260752
[13] [18] [20] [29] posed [2] RAZAVI, B. Design considerations for direct-conversion
Signal bandwidth receivers. IEEE Transactions on Circuits and Systems II: Analog
150 200 600 400 100
[MHz] and Digital Signal Processing, 1997, vol. 44, no. 6, p. 428–435.
Central frequency DOI: 10.1109/82.592569
1 3 1–6 11 2.0–3.5
[GHz]
[3] VALKAMA, M. Advanced I/Q Signal Processing for Wideband
T Sample rate
10000 / / 800 245.76 Receivers: Models and Algorithms. Ph.D. dissertation. Tampere
X [MS/s]
(Finland): Tampere University of Technology, 2001.
Before comp. ISBN: 9521507152
40–45 / / 20 20–30
IRR [dBc]
After comp. [4] 3GPP. Technical Specification Group Radio Access Network; NR;
50–55 / / 45 45–55 Base Station (BS) Radio Transmission and Reception. 2019,
IRR [dBc]
R Sample rate Valbonne, France, Tech. Rep.
/ 491.7 1250 122.88
X [MS/s] [5] KISS, P., PRODANOV, V. One-tap wideband I/Q compensation
Before comp. for zero-IF filters. IEEE Transactions on Circuits and Systems I:
/ 52.2 25–40 20–30
IRR [dBc] Regular Papers, 2004, vol. 51, no. 6, p. 1062–1074.
After comp. DOI: 10.1109/TCSI.2004.829233
/ 81.6 65 40–60
IRR [dBc]
[6] LIM, A. G. K. C., SREERAM, V., WANG, G. Q. Digital
Tab. 1. Performance comparison of different methods. compensation in IQ modulators using adaptive FIR filters. IEEE
RADIOENGINEERING, VOL. 32, NO. 4, DECEMBER 2023 491

Transactions on Vehicular Technology, 2004, vol. 53, no. 6, (MIKON). Warsaw (Poland), 2020, p. 209–213. DOI:
p. 1809–1817. DOI: 10.1109/TVT.2004.836934 10.23919/MIKON48703.2020.9253894
[7] TUTHILL, J. CANTONI, A. Efficient compensation for [21] MENG, J., WANG, H., YE, P., et al. I/Q linear phase imbalance
frequency-dependent errors in analog reconstruction filters used in estimation technique of the wideband zero-IF receiver. Electronics.
IQ modulators. IEEE Transactions on Communications, 2005, 2020, vol. 9, no. 11, p. 1–14. DOI: 10.3390/electronics9111787
vol. 53, no. 3, p. 489–496. DOI: 10.1109/TCOMM.2005.843455
[22] JOVANOVIĆ, B., MILENKOVIĆ, S. Transmitter IQ imbalance
[8] DING, L., MA, Z., MORGAN, D. R., et al. Compensation of mitigation and PA linearization in software defined radios.
frequency-dependent gain/phase imbalance in predistortion Radioengineering, 2022, vol. 31, no. 1, p. 144–154. DOI:
linearization systems. IEEE Transactions on Circuits and Systems 10.13164/re.2022.0144
I: Regular Papers, 2008, vol. 55, no. 1, p. 390–397. DOI:
10.1109/TCSI.2007.910545 [23] INGLE, V. K., PROAKIS, J. G. Digital Signal Processing Using
MATLAB: A Problem Solving Companion. 4th ed. Boston (USA):
[9] ANTTILA, L., VALKAMA, M., RENFORS, M. Frequency- Cengage Learning, 2015. ISBN: 978-1305635128
selective IQ mismatch calibration of wideband direct-conversion
transmitters. IEEE Transactions on Circuits and Systems II: [24] FLETCHER, R. Practical Methods of Optimization. 2nd ed. New
Express Briefs, 2008, vol. 55, no. 4, p. 359–363. DOI: York (USA): John Wiley & Sons, 1987. ISBN: 978-0-471-91547-8
10.1109/TCSII.2008.919500 [25] LIMEMICROSYSTEMS, UK. LimeSDR qPCIe (datasheet).
[10] CAVERS, J. K. The effect of quadrature modulator and [Online] Cited 2023-06-28. Available at:
demodulator errors on adaptive digital predistorters for amplifier https://wiki.myriadrf.org/LimeSDR-QPCIe
linearization. IEEE Transactions on Vehicular Technology, 1997, [26] KIELA, K., JURGO, M., MACAITIS, V., et al. 5G standalone and
vol. 46, no. 2, p. 456–466. DOI: 10.1109/25.580784 4G multi-carrier network-in-a-box using a software defined radio
[11] ANTTILA, L., HANDEL, P., VALKAMA, M. Joint mitigation of framework. Sensors, 2021, no. 21, p. 1–18.
power amplifier and I/Q modulator impairments in broadband DOI: 10.3390/s21165653
direct-conversion transmitters. IEEE Transactions on Microwave [27] JOVANOVIĆ, B., MILENKOVIĆ, S. PA linearization by digital
Theory and Techniques, 2010, vol. 58, no. 4, p. 730–739. predistortion and peak to average power ratio reduction in software
DOI: 10.1109/TMTT.2010.2041579 defined radios. Journal of Circuits, Systems and Computers, 2020,
[12] CAO, H., SOLTANI TEHRANI, A., FAGER, C., et al. I/Q vol. 29, no. 9, p. 1–19. DOI: 10.1142/S0218126620501479
imbalance compensation using a nonlinear modeling approach. [28] FAN, L., LI, Y. ZHAO, M. Joint IQ imbalance and PA
IEEE Transactions on Microwave Theory and Techniques, 2009, nonlinearity pre-distortion for highly integrated millimetre-wave
vol. 57, no. 3, p. 513–518. DOI: 10.1109/TMTT.2008.2012305 transmitters. In Proceedings of 2014 IEEE Globecom Workshops
[13] LI, W., ZHANG, Y., HUANG, L. K., et al. Self-IQ-demodulation (GC Wkshps). Austin (USA), 2014, p. 399–404. DOI:
based compensation scheme of frequency-dependent IQ imbalance 10.1109/GLOCOMW.2014.7063464
for wideband direct-conversion transmitters. IEEE Transactions on [29] KIM, M., MARUICHI, Y., TAKADA, J. I. Parametric method of
Broadcasting, 2015, vol. 61, no. 4, p. 666–673. DOI: frequency dependent I/Q imbalance compensation for wideband
10.1109/TBC.2015.2465138 quadrature modulator. IEEE Transactions on Microwave Theory
[14] MKADEM, F., FARRES, M. C., BOUMAIZA, S., et al. Com- and Techniques, 2013, vol. 61, no. 1, p. 270–280. DOI:
plexity reduced Volterra series model for power amplifier digital 10.1109/TMTT.2012.2228215
predistortion. Analog Integrated Circuits and Signal Processing,
2014, vol. 79, p. 331–343. DOI: 10.1007/s10470-014-0266-4
[15] VALKAMA, M., RENFORS, M., KOIVUNEN, V. Blind source
separation based I/Q imbalance compensation. In Proceedings of
the IEEE 2000 Adaptive Systems for Signal Processing,
Communications, and Control Symposium. Lake Louise (Canada),
2000, p. 310–314. DOI: 10.1109/ASSPCC.2000.882491
About the Authors ...
[16] CETIN, E., KALE, I., MORLING, R. Adaptive compensation of
analog front-end I/Q mismatches in digital receivers. In Borisav JOVANOVIĆ was born in 1979 in Niš, Republic
Proceedings of 2001 IEEE International Symposium on Circuits of Serbia. He received his Ph.D. degree from the Dept. of
and Systems (ISCAS 2001). Sydney (Australia), 2001, vol. 4,
p. 370–373. DOI: 10.1109/ISCAS.2001.922250
Electronics, Faculty of Electrical Engineering, University
of Niš, Serbia, in 2016, which is entitled with “Advanced
[17] ZEKKARI, C., DJENDI, M., GUESSOUM, A. Efficient adaptive Methods of Designing Digital Integrated Circuits in Na-
filtering algorithm for IQ imbalance compensation TX/RX
systems. IET Signal Processing, 2018, vol. 12, no. 5, p. 566–573.
nometer Technologies with Special Emphasis on Speed,
DOI: 10.1049/iet-spr.2017.0448 Static and Dynamic Consumption”. His current research is
focused on digital techniques for predistortion and correc-
[18] PENG, X., WANG, Z., MO, J., et al. A blind calibration model for
I/Q imbalances of wideband zero-IF receivers. Electronics, 2020, tion of static and dynamic IQ imbalance of 4G and 5G RF
vol. 9, no. 11, p. 1–16. DOI: 10.3390/electronics9111868 transmitters.
[19] LI, Y., CHEANG, C. F., MAK, P. I., et al. Joint-digital- Srđan MILENKOVIĆ was born in 1962 in Leskovac,
predistortion for wireless transmitter’s I/Q imbalance and PA Republic of Serbia. He received his Ph.D. degree from the
nonlinearities using an asymmetrical complexity-reduced Volterra
series model. Analog Integrated Circuits and Signal Processing, Dept. of Electronics, Faculty of Electrical Engineering,
2016, vol. 87, no. 1, p. 35–47. DOI: 10.1007/s10470-016-0724-2 University of Niš in 1996 and his Ph.D. thesis deals with
[20] ROSOLOWSKI, D. W., KORPAS, P. IQ-imbalance and DC-offset
artificial neural networks. His research is focused on the
compensation in ultra wideband zero-IF receiver. In Proceedings design of software radio defined telecommunication
of the 23rd International Microwave and Radar Conference systems.