IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO.

2, MARCH 2004 443

Gain, Phase Imbalance, and Phase Noise Effects

on Error Vector Magnitude
Apostolos Georgiadis, Member, IEEE

Abstract—The error vector magnitude (EVM) is extensively

applied as a measure of communication systems’ performance. In
this paper, the effects of gain, phase imbalance, and phase noise on
EVM are examined. The work is focused on single-carrier, linear,
and memoryless modulated signals, such as phase-shift keying and
quadrature amplitude modulation (QAM). The EVM is calculated
under the assumption that the transmitted signal consists of
zero-mean uncorrelated inphase and quadrature components that
are corrupted by additive white Gaussian noise. The contributions Fig. 1. System model.
of this paper are as follows. First, an expression for the EVM is
derived using a simple model that accounts for linear transmitter
and receiver imperfections, inspired by the works of Cavers and
dence on sideband suppression (SSB) and phase noise. Finally,
Liao, 1993. Second, a union bound on the symbol error rate (SER) differences between the EVM and SER dependence on ampli-
is derived. The root mean square EVM is shown to be independent tude and phase imbalance are emphasized.
of the constellation shape. The SER, however, is sensitive to the
individual transmitted symbols and, therefore, the constellation II. SYSTEM MODEL
shape. The resulting equations are used to examine the relation
between EVM, sideband suppression, and phase noise. The system is comprised of a quadrature modulator, channel,
Index Terms—Error vector magnitude (EVM), gain imbalance, and quadrature demodulator [1], as shown in the block diagram
phase imbalance, phase noise, phase-shift keying (PSK), quadra- of Fig. 1. The signal at each stage is represented by a 2 1
ture amplitude modulation (QAM), sideband suppression (SSB). column vector that contains the inphase and quadrature compo-
nents of its low-pass representation.
A transmitted symbol is mapped to a constellation point rep-
I. INTRODUCTION
resented by a 2 1 vector . The modulator imperfections are

R ECENTLY, the importance of error vector magnitude

(EVM) in transceiver characterization has been empha-
sized [2], [3]. Its popularity relies on the fact that it contains
included in the model with the use of a 2 2 matrix
1 vector . The matrix
and a 2
represents a linear transform of ,
resulting from the gain and phase imbalance of the transmitter.
information on the various transmitter and receiver imper- The vector is used to include any direct current (dc) offsets
fections, making it very useful for the radio designer. While present at the modulator input. The channel is represented by the
there have been a number of simulations and measurements matrix . Additive Gaussian noise corrupts the signal before
of EVM [4]–[6], a general analytical expression relating EVM it enters the receiver. Correspondingly, the demodulator imper-
to various system imperfections has not been listed, to the fections are included with the combination of a matrix and a
author’s knowledge. vector containing imbalances and dc offsets, respectively.
The transmitted signal is assumed to consist of zero-mean un- Assuming a gain imbalance and a phase imbalance , the
correlated inphase and quadrature components. Single-carrier, modulator or demodulator matrix becomes [1]
linear, and memoryless modulated signals, such as phase-shift
keying (PSK) and quadrature amplitude modulation (QAM),
are considered. Moreover, transmission through an ideal addi- (1)
tive white Gaussian noise (AWGN) channel is assumed. The
modulator and demodulator imbalances change the correlation The phase imbalance has been divided equally among the and
properties of the signal and white noise that are applied to their channels. In the following, the transmitter imbalances will be
inputs. denoted with an index and the receiver with an index .
Initially, a system model is constructed based on the formu- The channel is assumed to be ideal unit gain represented by
lation by [1]. The model is then used to derive expressions that the matrix . The phase difference between the transmitter
provide EVM and a union bound on symbol error rate (SER). and receiver oscillators is included in the channel response as
The EVM expression is then explored to investigate its depen- a phase rotation

Manuscript received July 29, 2002; revised July 28, 2003 and November 17, (2)
2003.
The author was with Global Communication Devices, Systems Laboratory,
North Andover, MA 01845 USA (e-mail: [email protected]). The transmitted symbol vectors are normalized to the average
Digital Object Identifier 10.1109/TVT.2004.823477 symbol energy . As a result, the power of the signal present at
0018-9545/04$20.00 © 2004 IEEE

Authorized licensed use limited to: University of Cyprus. Downloaded on January 10,2022 at 08:59:27 UTC from IEEE Xplore. Restrictions apply.
444 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 2, MARCH 2004

the receiver input can be accounted for with a gain factor ,

multiplying the channel matrix .
Using the above formulation, the following equations provide
the signal at any stage of the system presented in Fig. 1.

(3)
(4)
(5)

with
Fig. 2. EVM definition.
(6)
(7) The result of the receiver phase imbalance is that the noise pro-
(8) cesses corrupting the and channels are now correlated (14).
A property that will be useful in the following calculations is
(9) now presented. Let a zero-mean vector process with variance
(10) . Also assume that the process components are uncorrelated
and have equal variances. Given the process and any matrix
The combined effects of the transmitter, receiver imbalances, , the following property is proved in Appendix I:
and phase noise are included in the channel matrix (6). Note
that the transmit modulator dc offsets are referenced to its (15)
input (3), whereas the receive demodulator offsets are refer-
enced to the receiver output (5). where is the trace of matrix .
The received signal (5) is finally expressed as the sum of One process that has the same properties as is the white-
the transmitted symbol scaled by the average symbol energy noise vector process with . Moreover, the random
and an error vector (10). This is also pictured in Fig. 2. signal vector has the same properties with . There-
One can see that there are two distinct contributions to the error fore, the following holds:
vector, one from the thermal noise and one from the system
imperfections . It should also be clear that the instantaneous (16)
error vector depends on the transmitted symbol (through
and, therefore, the corresponding constellation point).
Another important observation is that the noise present at
(17)
the demodulator output is a linear transform of the thermal noise
(8). The noise vector is represented by its mean vector
(18)
and its covariance matrix , which are equal to
(19)
(11)
The results of (16) and (18) are immediately obvious.
(12)
Moreover, (17) is easily proven by taking the inner product
using (1). The linear transforms caused by the trans-
where the superscript H denotes the Hermitian (conjugate trans- mitter and receiver imperfections and , and also the
pose) operation. In the following, it is assumed that all vector
phase noise included in , preserve the variance (energy)
and matrix components are real. As a result, inner and outer of the input vector signal or noise. This can be expressed as
products are formed using only the transpose T operator. . As a result, it
Due to the transform caused by the imbalances is stated without proof that the cascade of these transforms,
at the receiver, the output noise mean and covariance matrix represented by the channel matrix , also leaves the input
become random process variance unchanged or

(13) (20)

III. EVM CALCULATION

The EVM is the envelope of the error vector (10). In
matrix form, this is equal to

(21)
(14) (22)

Authorized licensed use limited to: University of Cyprus. Downloaded on January 10,2022 at 08:59:27 UTC from IEEE Xplore. Restrictions apply.
GEORGIADIS: GAIN, PHASE IMBALANCE, AND PHASE NOISE EFFECTS ON EVM 445

with . It is customary to refer to the rms value of EVM, in the absence of any imperfections, is set by the SNR
the error vector. In this case, using the results of the previous .
section, one gets
IV. SER CALCULATION
(23)
Using the formulation of Section II, a union bound on the
where the fact that was used. It remains, symbol error probability can be derived. A maximum likelihood
therefore, to calculate the mean square value of the instanta- (ML) receiver correlates the received signal with every possible
neous error vector over the transmitted symbols set transmitted symbol and selects the one that corresponds to the
maximum correlation as the received signal. A union bound on
the probability of symbol error [7] is calculated by taking the
sum of the individual symbol error probabilities between the
(24) transmitted symbol and every other symbol in the transmit al-
phabet. The procedure is described in Appendix II. The union
The last two terms are equal to zero because . upper bound on the symbol error probability is equal to (29) and
Using (9), (10), and (20), one gets (30), shown at the bottom of the page, where the indices and
run over all the possible transmitted symbols.
(25) Each transmitted symbol may have a different symbol error
probability. This corresponds to the fact that different symbols
may have different EVM and average energy values. Although
Also, since , the mean square of the error in the calculation, these differences were averaged
vector becomes into a value that was independent of the modulation type, which
is not possible in the error-probability calculation because of the
(26)
nonlinearity of the function.
The error vector is usually expressed as a percentage of the
signal energy. Therefore, when constellations that contain sym- V. TRANSMITTER GAIN AND PHASE IMBALANCE
bols of different energy (such as QAM) are employed, one has As a first example, one can apply the above results to the
the option to normalize EVM to the peak symbol energy or the testing procedures of a transmit modulator. It is customary
average symbol energy, or even the minimum symbol energy. to measure the quality of the modulator by applying a complex
From (26), it is obvious that if one expresses EVM relative to the tone to its input. This is equivalent to applying the vector
peak symbol energy, its value depends on the modulation type to the and modulator in-
through the peak-to-average energy ratio (PAV). Therefore, it is puts. Since is a deterministic signal, one uses time averages
better to normalize using the average energy. In any case, the instead of ensemble averages to calculate its moments. It is easy
final result becomes to verify that is zero mean with components that are uncorre-
lated and have powers equal to 1/2. As a result, (27) is also valid.
(27) The output of the modulator ideally consists of one tone at
a distance of from the carrier [1]. However, if the modu-
lator exhibits gain and phase imbalance, a second tone
at is also present. Moreover, if there are dc offsets as well,
a carrier component is also present at the modulator output. The
modulator performance regarding gain and phase imbalance is
(28)
usually expressed by SSB, which is the power ratio of the spu-
rious versus the desired tone when a complex tone is present at
where and . its input. It is equal to [1]
Equation (27) proves that EVM is a good measure of the
transceiver quality and, in the case of linearly memoryless mod-
ulated signals, is independent of the modulation type that is (31)
being transmitted. Moreover, one can see that the minimum

(29)

(30)

Authorized licensed use limited to: University of Cyprus. Downloaded on January 10,2022 at 08:59:27 UTC from IEEE Xplore. Restrictions apply.
446 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 2, MARCH 2004

!
1
Fig. 3. EVM variation with transmitter gain and phase imbalance (SNR
+ ).

Fig. 4. EVM measurement and calculation comparison for various

It is easy to see from (31) that different combinations of gain modulation formats (SNR = 45 dB).
and phase imbalance correspond to the same SSB. These com-
binations are defined by

Const (32)

It is interesting to examine how SSB is related to the

and, more specifically, whether they have a similar
dependence on the transmitter imperfections. In this case,
the channel-transfer function consists of only as a perfect
receiver and perfect-phase reference
are assumed. Moreover, the dc offsets of the transmitter are
ignored as they are not considered in the SSB measurement.
Using (27)

Fig. 5. EVM variation with SSB for different SNR values.

(33)
the VSA provides EVM values relative to the peak symbol en-
The variation of with transmitter gain and phase im- ergy, they had to be scaled according to for the
balance is plotted in Fig. 3. In this plot, the SNR is assumed to be 16QAM case (28).
infinite so that the only contributions to are due to the Using (32) and taking advantage of the symmetry of (33), one
imperfections. It is clear from (33) that the constant can express as a function of SSB and or SSB and
curves are given by
(35)
Const (34)
where or . If one assumes that
which means that the and SSB dependence on and the modulator will usually exhibit relatively small imbalances,
is different. then one can approximate and derive
A measurement setup was constructed using an Agilent ESG
E4438C signal generator as a transmitter and an 89 600 Vector (36)
Signal Analyzer as a receiver. The ESG allows for the introduc-
tion of gain and phase imbalance to the transmitted signal and
the VSA provides a direct measurement of the Eq. (36) provides the desired relation between the measured
for various modulation formats. The measurement results for an SSB and , under the assumption that one of the mod-
are plotted in Fig. 4, showing a very good agree- ulator imbalances is relatively small. The variation of
ment with the calculated values. It should be noted that because with SSB for different SNR values is shown in Fig. 5.

Authorized licensed use limited to: University of Cyprus. Downloaded on January 10,2022 at 08:59:27 UTC from IEEE Xplore. Restrictions apply.
GEORGIADIS: GAIN, PHASE IMBALANCE, AND PHASE NOISE EFFECTS ON EVM 447

VI. EVM AND PHASE NOISE

The modulator and demodulator translate in frequency the
signal that is applied to their inputs by multiplication with a
radio-frequency (RF) LO carrier. The LO signal is provided
by a narrow-band phase-locked loop (PLL). Phase noise results
in spreading of the LO signal in the frequency domain, which
would otherwise be an ideal tone.
The probability density function (pdf) of the local oscillator
(LO) phase can be derived by solving a partial differential
equation known as the Fokker–Planck equation [8], [9]. This
problem is beyond the scope of this work. When the SNR
in the PLL bandwidth is high, the pdf of the LO phase error
approximates a Gaussian distribution [9]. Therefore, the phase
noise may be represented by a zero-mean Gaussian variable
whose variance defines the LO rms phase error.
In Section II, the channel matrix was used to include in
the model a fixed-phase difference between the receiver and Fig. 6. EVM variation with LO phase error without gain and phase imbalance.
transmitter local oscillators (2). In general, one may calculate
a conditional using (27) and integrate using a bols (constellation shape) and the correlation properties of the
Gaussian pdf for to get the mean square . noise vector (14), (30).
Assuming a perfect modulator and demodulator, the channel The dependence of SER on the correlation properties of the
becomes , so that using (2) and (27) one gets noise vector suggests that SER is sensitive to where AWGN
is included in the system model (Fig. 1). Therefore, based on
(37) Fig. 1, the cases of ideal demodulator and ideal mod-
ulator result in the same EVM, but in different SER.
Given an rms LO phase error , the mean square is This is displayed with the following example using quaternary
then equal to phase-shift keying (QPSK) modulation.
The is calculated using (27). Since a modulator is
compared with a demodulator with the same imbalances
, it is obvious that their EVM performances are the same.
The SER is calculated using (30). When the imbalances are at-
(38) tributed to the transmitter, one gets
The integral is calculated using [10]

(39)

Finally, the is

(40)
(41)
The variation with the LO phase error is plotted in
Fig. 6. When the receiver, however, is distorting the signal, the symbol-
error probability becomes
VII. EVM AND SER CONSIDERATIONS
The modulator and demodulator imbalances result in linear
transforms ( , ) of the signal and white noise. As a result,
they change the correlation properties of the processes that are
applied at their inputs. It was shown (27) that the is
independent of the individual transmitted symbols (i.e., the con-
stellation shape). Moreover, it depends on the variance of the
noise vector process, which is preserved by the linear transforms
, (17). In contrast, the SER depends on the variance values
of each correlation metric (48) due to the nonlinearity of the (42)
function. As a result, it depends on the transmitted sym-

Authorized licensed use limited to: University of Cyprus. Downloaded on January 10,2022 at 08:59:27 UTC from IEEE Xplore. Restrictions apply.
448 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 53, NO. 2, MARCH 2004

The results are plotted in Fig. 7. In order to verify the accuracy

of the calculations, the calculated results are complemented by
simulation results using Matlab. It is seen that a system that has a
transmitter with 10 phase imbalance and 3-dB gain imbalance
requires 1.4 dB more SNR to achieve a symbol error rate of
than a system with the same imbalances at its receiver. Simula-
tion results for 8-PSK modulated signals are also included in
Fig. 7. In this case, the degradation is 0.6 dB. Therefore, due to
the linear transformation that the noise vector undergoes when
passing through the quadrature demodulator, the system SER
performance appears to be less sensitive to receiver gain and
phase imbalance than to transmitter.
A second observation from (42) is that, for small receiver
phase imbalance values, the SER is independent of the gain im-
balance. Setting and , one gets

(43)
Fig. 7. Theoretical (TH) and simulated (SIM) performance under transmitter
and receiver imperfections (10 phase imbalance and 3-dB gain imbalance).
which is independent of . This is not possible in (41), where
the imperfections are due to the transmitter. This is a result of the properties and, therefore, requires careful modeling of the var-
constellation shape. It can be shown following a similar proce- ious noise contributors in the system model. Future work con-
dure that if the QPSK constellation is rotated by 45 , the SER in sists of including in the system model transmitter or receiver
the case of small receiver gain imbalance is independent of the nonlinearities through power or Volterra series.
receiver phase imbalance. Therefore, it appears that the shape of
the constellation determines whether the receiver gain or phase APPENDIX I
imbalance has a stronger effect on the error rate. USEFUL QUADRATIC
In summary, it is important for SER calculations to carefully Consider a column vector process and a 2 2
include the AWGN in the system model. Each noise contributor
should be identified and introduced separately before or after the matrix . The quadratic is equal to
modulator and demodulator matrices. Regarding the receiver,
for example, noise generated by components (filters, amplifiers) (44)
that are placed after the quadrature demodulator is not affected
by the transformation and should not be referred at Moreover, let the process be zero mean , with com-
the input of the receiver. Moreover, within the demodulator it- ponents that are uncorrelated , and have equal
self, there are noise contributors that may be input referred and variances .
others that should only be introduced at its output depending In this case, the expected value of the quadratic becomes
on the specific circuit implementation. However, noise gener-
ated by the low-noise amplifier stage (LNA), which precedes (45)
the quadrature demodulator, may be input referred. In practical
receiver implementations, it is desired to place enough gain in where is the trace of the matrix .
the LNA such that its noise figure determines the overall re-
ceiver noise figure. As a result, the LNA noise is the main noise APPENDIX II
contributor and, since it can be input referred, the system model UNION BOUND ON SER
of Fig. 1 may be applicable. Similar considerations apply for the Let the transmitted symbol be . Also let the correlation
transmitter. metric of the received signal and a symbol be .A
symbol error occurs if is less than with or,
VIII. SUMMARY equivalently, if their difference is less than zero.
The effects of linear transmitter and receiver imperfections (46)
such as gain and phase imbalance, as well as phase noise on
(47)
the , have been investigated. An analytical expression
was derived for and SER, when the transmitted sig-
nals consist of zero mean uncorrelated inphase and quadrature (48)
components. It was found that is independent of the
constellation shape. Moreover, is independent of the A union bound on the probability of a symbol error [7] can
noise vector correlation properties. The SER, on the other hand, be derived by taking the sum of the individual symbol-error
is sensitive to both the constellation shape and noise correlation probabilities between the transmitted symbol and every other

Authorized licensed use limited to: University of Cyprus. Downloaded on January 10,2022 at 08:59:27 UTC from IEEE Xplore. Restrictions apply.
GEORGIADIS: GAIN, PHASE IMBALANCE, AND PHASE NOISE EFFECTS ON EVM 449

symbol in the transmit alphabet. Finally, one has to average over [2] K. Voelker, “Apply error vector measurements in communications de-
all possible transmitted symbols. Assuming equal transmission sign,” Microwaves RF, pp. 143–152, 1995.
[3] R. Hassun, M. Flaherty, R. Matreci, and M. Taylor, “Effective evaluation
probabilities for all symbols, one gets of link quality using error vector magnitude techniques,” in Proc. IEEE
Wireless Commun. Conf., 1997, pp. 89–94.
[4] T. Nakagawa and K. Araki, “Effect of phase noise on rf communication
signals,” in Proc. IEEE VTC’00, vol. 2, 2000, pp. 588–591.
[5] H. Ku and J. S. Kenney, “Estimation of error vector magnitude using
(49) two-tone intermodulation distortion measurements,” in Proc. IEEE/MTT
One can calculate the mean and variance of Microwave Symp. Dig., vol. 1, 2001, pp. 17–20.
using (48) as [6] J. L. Pinto and I. Darwazeh, “Phase distortion and error vector magnitude
for 8-psk systems,” Electron. Lett., vol. 37, pp. 437–438, 2001.
[7] J. G. Proakis, Digital Communications, 3rd ed. New York: McGraw-
(50) Hill, 1995.
[8] A. J. Viterbi, Principles of Coherent Communication. New York: Mc-
(51) Graw-Hill, 1969.
[9] , “Phase-locked loop dynamics in the presence of noise by
(52) Fokker–Planck techniques,” Proc. IEEE, vol. 51, pp. 1737–1753, Dec.
1963.
[10] M. Abramowitz and I. Stegan, Handbook of Mathematical Func-
In this case, it is easily proven that [11] tions. New York: Dover, 1970.
[11] W. C.. Lindsey and M. K. Simon, Telecommunication Systems Engi-
(53) neering. New York: Dover, 1973.

Apostolos Georgiadis (S’94–M’03) was born

(54) in Thessaloniki, Greece, in 1971. He received
the B.S. degree in physics and the M.S. degree
in electronics physics, radioelectrology from the
In this way, the union upper bound on the symbol-error prob- Aristotle University of Thessaloniki, Greece, in
ability is equal to 1993 and 1996, respectively. He received the Ph.D.
degree in electrical engineering from the University
of Massachusetts, Amherst, in 2002.
In 1995, he was with Radio Antenna Communi-
cations, Milan, Italy, working on wire antennas used
in UHF and VHF communications. In the summer of
2000, he was with Telaxis Communications, South Deerfield, MA, where he
worked on antennas and circuits for millimeter wave (LMDS) applications. In
January 2002, he joined Global Communications Devices (GCD), North An-
dover, MA, as a Systems Engineer and worked on CMOS transceivers for wire-
(55) less network applications. In June 2003, he joined Bermai, Inc., as an RF/Analog
Systems Architect. In August 2003, he returned to Greece, where he is currently
fulfilling a military service requirement in the Greek army.
REFERENCES Dr. Georgiadis is a Fulbright Scholar. He received the Outstanding Teaching
Assistant Award from the University of Massachusetts, Amherst, in 1997 and
[1] J. K. Cavers and M. W. Liao, “Adaptive compensation for imbalance 1998. He also was the recipient of the Eugene M. Isenberg Award from the Isen-
and offset losses in direct conversion transceivers,” IEEE Trans. Veh. berg School of Management, University of Massachusetts, Amherst, in 1999 and
Technol., vol. 42, pp. 581–588, Nov. 1993. 2000.

Authorized licensed use limited to: University of Cyprus. Downloaded on January 10,2022 at 08:59:27 UTC from IEEE Xplore. Restrictions apply.