1644
4, APRIL 2009
Semianalytic BER for PSK
Mark T. Core, Ray Campbell, Member, IEEE, Patricia Quan, and Joan Wada, Member, IEEE
Abstract—Semianalytic bit error rate (BER) estimation is a
well-known method for evaluating the BER of a digital commu-
nication system. The main utility of the method is the significant
time savings in computation relative to Monte Carlo simulation.
Despite this advantage, no known reference defines the procedure
for computing exact BER for M-ary phase shift keying (PSK)
with ISI and AWGN using the semianalytic method. This letter
defines an efficient procedure for computing exact semianalytic
BER for modulation formats with circular constellations when
the noise component of the decision variable has a circularly
symmetric Gaussian distribution. The technique is demonstrated
for 8PSK over the Digital Video Broadcasting-Satellite-Second
Generation (DVB-S2) channel.
Index Terms—Computer aided analysis, digital communica-
tion, error analysis, phase shift keying, simulation.
I. I NTRODUCTION
C OMPUTER aided methods for determining the probabil-
ity of bit error Pb or BER of a digital communication
system are widely used when an accurate analytic expression
for BER is not available. The most general technique for
BER estimation is the Monte Carlo method. A simulation is
performed by passing a known data sequence through a system
model and counting the bit errors at the receiver decision
device. A simple and unbiased estimator of the BER is the
sample mean of the simulated BER. If N is the number of
bits evaluated, then as N → ∞ the estimate will converge to
the true BER. In practice, the rule of thumb that N should
be on the order of 10/Pb produces a 90% confidence interval
of about two on the BER estimate. Simulations of systems
operating at very low error probabilities therefore require
prohibitively large run times to accurately estimate BER using
the Monte Carlo method.
There are several techniques for accelerating the estimation
of BER [1]-[4]. The method considered here is called the
semianalytic or quasianalytic method. With this method the
problem is broken into two parts, the first dealing with the
signal component, and the second dealing with the noise
component of the decision variable at the receiver. For a
known data sequence and a static channel, the demodulated
and sampled signal component at the decision device is
deterministic and can be evaluated by simulation. The data
sequence used in the simulation should have specific properties
in order to emulate random data with equiprobable symbols.
Using this noise free sequence of demodulated and detected
Manuscript received December 17, 2007; revised April 21, 2008; accepted
June 6, 2008. The associate editor coordinating the review of this letter and
approving it for publication was C.-X. Wang.
M. Core is with Broadcom Corp., Irvine, CA 92617 (e-mail:
mcore@jps.net).
R. Campbell, P. Quan, and J. Wada are with The Boeing Company,
Anaheim, CA 92803.
Digital Object Identifier 10.1109/TWC.2009.071412
1536-1276/09$25.00 
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO.
decision variables, then if the probability density of the noise
component is known, the BER can be computed. This is
done by integrating the noise probability density function
(pdf) centered about each of the simulated noise free decision
variable points over each error event region. Each integral is
then weighted by the corresponding number of bit errors and
the sample mean over all the weighted integrals produces the
BER. Because of the complexity of this procedure for M-ary
PSK, previous references have presented only approximations
[4]. For ideal M-ary PSK with known symbol probabilities
and arbitrary symbol bit-mapping it is possible to compute
the exact BER for the additive white Gaussian noise (AWGN)
channel in an efficient way [5]. By extending these methods
to include the effects of intersymbol interference (ISI), an
efficient procedure for computing semianalytic BER will be
developed.
II. PSK C ONSTELLATION WITH ISI AND AWGN
The M signal waveforms for ideal M-ary PSK are repre-
sented as
sm (t) = g(t)cos(ωc t + θm ) m = 0, 1, ...,M − 1 (1)
where g(t) is a pulse waveform used to shape the spectrum
of the transmitted signal and θm is the information-bearing
phase angle which takes M possible values
(2m + 1) π
θm = m = 0, 1, ..., M − 1. (2)
M
If an ideal PSK signal is optimally demodulated, then using
complex phasor notation, each of the complex decision vari-
ables takes one of the following M values
Sm = εejθm m = 0, 1, ..., M − 1 (3)
where ε is the energy of the spectrum shaping pulse and is
given in terms of bit energy Eb by ε = log2 (M ) Eb . Fig.
1 shows the ideal symbol locations Sm and decision regions
Rm for 8PSK.
When distortions due to channel effects or modem imper-
fections are present, the received decision variables will differ
from the M ideal points, and their locations will be data
dependent due to ISI. In this context, ISI will refer to the
effects of both linear and non-linear time invariant distortions
with memory. Assuming equiprobable symbols, then in order
to completely characterize the ISI of a channel with L symbol
periods of memory, it is sufficient to consider all possible
sequences of L symbols. A maximal length pseudorandom
M L symbol sequence will satisfy this property. For M=2,
linear feedback shift registers can be used to generate maxi-
mal length pseudorandom bit sequences. For M>2, efficient
c 2009 IEEE
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 4, APRIL 2009 1645

Im{s} Im{s}
q2
S2 S1 q1

R2 R1
S3 S0 h2 rk
R3 R0 h3 h1 θk
Re{s} Re{s}

R4 R7
S4 S7 c1 h4
R5 R6

S5 S6

Fig. 1. Ideal 8PSK symbols Sm and decision regions Rm . Fig. 2. 8PSK decision regions showing half-planes, quadrant planes, and
correction planes. A single noise free decision variable sk with ISI is also
shown.

methods for generating maximal length pseudorandom symbol

sequences have also been proposed [6][7]. With the addi- decision variable points will determine the BER of the system.
tion of cyclic prefixes and postfixes of L symbols each, a The bit error probability is therefore given by
simulation using one cycle of a M L length pseudorandom
symbol sequence is sufficient to emulate equiprobable data M −1 L

symbols. Therefore, performing a simulation using M L + 2L 1 

Pb = L Pb (sk , m) (7)
symbols from a maximal length M L symbol sequence, and M
k=0
discarding the first and last L demodulated and detected
decision variable points, the resulting M L decision variable where Pb (sk , m) is the bit error probability for decision
points will completely characterize the effect of the system variable sk with correct decision region Rm . An expression
ISI on the signal. This set of decision variables can be defined for Pb (sk , m) is
in terms of their respective magnitudes and phases or in-phase M−1
1 
and quadrature components Pb (sk , m) = wmn Pn/m (sk ). (8)
log2 (M ) n=0

sk = rk ejθk = ik + jqk k = 0, 1, ..., M L − 1. (4) The term Pn/m (sk ) is the probability that the received deci-
sion variable sk from region Rm would fall in decision region
When AWGN is present at the receiver input, the decision Rn under the influence of noise, and wmn is the number of bit
variables are errors that occurs in this event. For an arbitrary bit mapping,
the bit error weight matrix W with elements wmn can be
y k = sk + n k k = 0, 1, ..., M L − 1. (5) generated from the weight matrix for the binary bit mapping
For a receiver having an arbitrary discrete time detection filter by swapping rows and then columns according to the desired
with impulse response h(n), the noise component nk at the symbol assignment.
filter output is a sequence of complex Gaussian distributed Equation (8) can be simplified by noting that each received
random variables each with mean zero and variance given by decision variable can be rotated to decision region R0 provided
that the bit error weight matrix W is appropriately transformed
∞  as well. The rotated decision variable from decision region Rm
 N0 π   is
σ 2 = σr2 + σi2 = N0
2
|h(n)| = H(ejω )2 dω.
n=−∞
2π −π
2πm
(6) sk = e−j M sk . (9)
where σr2 and σi2 are the variances of the real and imaginary
components respectively, and N0 is the one-sided power The required transformation for the W matrix consists of a
spectral density (psd) of the AWGN. cyclic left shift of row m by m positions. This operation can
be expressed as
III. B IT E RROR R ATE C ALCULATION 
wmn = wm(n+m)modM . (10)
Since a pseudorandom symbol sequence emulates equiprob-
able symbols, averaging the BER over all of the resulting In terms of the transformed variables, (8) becomes
1646 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 4, APRIL 2009

where the integrand term is

M−1

1
Pb (sk , m) = w Pn/0 (sk ). (11)
log2 (M ) n=1 mn
f (x;
i, M, qk , σ) = 

In this way, only the error event probabilities for decision tan( 2(1−i)π )x+qk tan( 2iπ
M )x+qk (18)
erfc M
σ − erfc σ .
variables from decision region R0 need to be computed. This
situation is depicted in Fig. 2. Expressing (11) in vector
notation produces The values for the correction plane probabilities ci (k) speci-
fied by (17) and (18) can be evaluated by numerical integra-
1 tion.
Pb (sk , m) = wT (k)p(k) (12)
log2 (M ) Expressions for the half-plane, quarter plane, and correction
where p(k) is a column vector with elements Pn/0 (sk ) for plane probabilities can also be written in terms of the desired
n=1,2,. . . ,M-1 and wT (k) is a row vector with elements taken probabilities Pn/0 (sk )
from row m and columns 1 through M-1 of the weight matrix
W  . Note that column 0 need not be included because this M
+(i−1)
corresponds to a correct decision, and the weight is therefore
2  M
hi (k) = Pn/0 (sk ) i = 1, 2, ..., (19)
always 0. Substituting (12) into (7) produces 2
n=i
L
M−1
1
Pb = L wT (k)p(k). (13) M
+(i−1)
M log2 (M )
k=0
4  M
qi (k) = Pn/0 (sk ) i = 1, 2, ..., (20)
The main difficulty in evaluating (13) is the computa- n=i
4
tion of the individual error event probabilities Pn/0 (sk ) for
n=1,2,. . . ,M-1. The probabilities could be evaluated for each
M
decision variable point by numerically computing M-1 double +(i−1)
2  M
integrals over the error event regions. By generalizing the ci (k) = Pn/0 (sk ) i = 1, 2, ..., − 1 . (21)
4
method presented in [5] to include ISI, a significant improve- n= M
2 −(i−1)
ment in computational speed and accuracy can be achieved
relative to direct numerical integration. This is done by ex- The system of linear equations specified by (19)-(21) can be
pressing a set of half-plane, quarter-plane, and correction- expressed compactly in matrix-vector notation as
plane probabilities in terms of the desired Pn/0 (sk ) and
solving this system of equations for Pn/0 (sk ). For an M-ary z(k) = Ap(k) (22)
PSK decision variable specified by sk as shown in Fig. 2
and circularly symmetric additive Gaussian distributed noise, where z(k) is a known column vector of h(k), q(k), and c(k)
M/2 independent half-plane probabilities and M/4 independent values, A is a matrix of known coefficients, and p(k) is an
quarter-plane probabilities can be written unknown column vector of Pn/0 (sk ) values. The matrix A
 and the form of the vector z(k) are determined by the PSK
 2πi 
1 rk sin − θk M modulation order M. For 8PSK they are given by
M
hi (k) = erfc i = 1, 2, ..., (14)
2 σ 2
⎡ ⎤ ⎡ ⎤
1 1 1 1 0 0 0 h1 (k)
  M ⎢ 0 1 1 1 1 0 0 ⎥ ⎢ h2 (k) ⎥
qi (k) = hi (k) 1−hi+M/4 (k) . (15)
i = 1, 2, ..., ⎢ ⎥ ⎢ ⎥
4 ⎢ 0 0 1 1 1 1 0 ⎥ ⎢ h3 (k) ⎥
⎢ ⎥ ⎢ ⎥
In addition to the 3M/4 probabilities specified by (14) and A=⎢
⎢ 0 0 0 1 1 1 1 ⎥ z(k) = ⎢
⎥ ⎢ h4 (k) ⎥.
⎥
(15), the remaining required M/4-1 independent probabilities ⎢ 1 1 0 0 0 0 0 ⎥ ⎢ q1 (k) ⎥
⎢ ⎥ ⎢ ⎥
can be found using the following correction plane probabilities ⎣ 0 1 1 0 0 0 0 ⎦ ⎣ q2 (k) ⎦
0 0 0 1 0 0 0 c1 (k)
(23)
−((x+ik )2 +(y+qk )2 )
∞ tan( 2iπ
M )x
ci (k) = 1
e σ2 dydx The system of equations specified by (22) can be solved for
πσ2 0 2(1−i)π
tan( M )x the unknown probabilities
i = 1, 2, ..., M
4 −1
(16)
where ik = rk cos(θk ) and qk = rk sin(θk ). The expressions p(k) = A−1 z(k). (24)
for the probabilities (14)-(16) are given by integrating the
circularly symmetric Gaussian pdf over the specified regions. Substituting (24) into (13) produces the desired expression for
Evaluating the inner integral of (16) produces the bit error probability

L
M−1
∞ −(x+ik )2
1
ci (k) = 2√1πσ2 0 e σ2 f (x; i, M, qk , σ) dx (17) Pb = L wT (k)A−1 z(k). (25)
M log2 (M )
i = 1, 2, ..., M
4 −1
k=0
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 4, APRIL 2009
MOD IMUX TWTA OMUX DEMOD
n(t)
Fig. 3. Simulation model.
1.5
1 10
0.5
Im{s}
0 Fig. 5. Semianalytic and Monte Carlo BER as a function of SNR for uncoded
-0.5
-1
-1.5
-1.5 -1 -0.5 0 0.5 1 1.5
Re{s}
Fig. 4. Noise free decision variable points sk used for the semianalytic BER
computation of uncoded 8PSK and the DVB-S2 channel.
IV. A PPLICATION E XAMPLE
To serve as an example of the proposed method, the
BER for uncoded 8PSK with the DVB-S2 satellite channel
was evaluated and compared to the results of Monte Carlo
simulation. A block diagram of the model used for the
simulations is shown in Fig. 3. The characteristics of the
channel model are specified in the DVB-S2 standard [8].
The complex baseband simulations were performed using
an 8PSK pseudorandom symbol sequence sampled at 16
samples/symbol. A Gray coded symbol bit mapping was used
and the symbol pulse shaping filter had a square-root raised-
cosine magnitude response with a roll-off factor of 0.35. A
10th order rational transfer function was used to approximate
the specified frequency response of the input multiplexer
(IMUX) filter and a 6th order rational transfer function was
used for the output multiplexer (OMUX) filter. The signal
at the traveling wave tube amplifier (TWTA) model input
was scaled to a level corresponding to 1dB of input backoff
(IBO). The TWTA produces a passband nonlinearity that was
modeled using the non-linearized amplitude modulation-to-
amplitude modulation (AM/AM) and amplitude modulation-
to-phase modulation (AM/PM) conversions specified in the
standard. The output from the OMUX filter was applied to
a square-root raised-cosine detection filter with roll-off factor
0.35 and then interpolated by 4. The optimal sample phase,
carrier phase, and gain were then jointly determined to produce
decision variables with the minimum mean squared error
relative to the ideal symbol locations.
1647
0
10
Semianalytic
Monte Carlo
−2
10
−4
10
BER
−6
10
−8
10
−10
−12
10
0 5 10 15 20
E /N (dB)
b 0
8PSK and the DVB-S2 channel.
From [8] it can be seen that the passband channel memory
spans approximately 5 symbol periods due to the IMUX
and OMUX filters, so a value of L=5 was chosen. For the
semianalytic BER computation, only one cycle of a length
32768 pseudorandom symbol sequence with a small cyclic
prefix and postfix was simulated. After discarding the decision
variable points corresponding to the prefix and postfix, the
resulting set of points shown in Fig. 4 was used to compute
the BER according to (25).
For the Monte Carlo BER estimate, 10173 cycles of the
32768 length pseudorandom symbol sequence were simulated
for a total of approximately 109 bits. AWGN was applied to
the receiver detection filter and the resulting decision variable
noise components were scaled and added to the deterministic
components to emulate the desired effective SNR at the
receiver input. Bit errors were then counted to produce BER
estimates. The results of the semianalytic and Monte Carlo
simulations are plotted in Fig. 5. The differences between the
results are consistent with the expected estimation error for
a Monte Carlo simulation of 109 bits in length. While the
Monte Carlo simulation produced no errors when the SNR was
greater than 22dB, the semianalytic method produced accurate
results at all signal-to-noise ratios.
V. C ONCLUSION
A fast and accurate method for computing semianalytic
BER for PSK with ISI and AWGN has been presented. Two
simplifications were applied relative to directly evaluating the
semianalytic BER using numerical integration. First, the deci-
sion variable points were each rotated to decision region R0 ,
and the bit error weight matrix was appropriately transformed.
Second, the process of numerically computing M-1 double
integrals for each decision variable point was simplified to
require only the evaluation of complementary error functions
and M/4-1 single integrals. The technique is applicable to
ideally demodulated and matched filter detected PSK with a
static channel distortion, and produces an exact BER result
when the received noise has a circularly symmetric Gaussian
1648 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 4, APRIL 2009
distribution. The basic approach can also be applied for any
circularly symmetric noise distribution by simply re-deriving
the half-plane and correction-plane probabilities of (14) and
(16) for the desired pdf.
An extension of the technique to include synchronization
errors in the demodulator is described in [3] and is referred
to as the statistical averaging method. In a similar way, the
technique can also be extended to frequency non-selective
slowly fading channels [9]. The BER specified by (25) that
accounts for time invariant distortions such as high power
amplifier nonlinearity and non-ideal filtering is first expressed
as a function of αrk , αik , and αqk , where α accounts for
the effect of fading. The resulting expression is then averaged
over the pdf of α. Depending on the pdf considered, some
of the integrals may be evaluated analytically to produce
a computationally efficient procedure for determining exact
BER in a flat fading environment.
R EFERENCES
[1] M. Jeruchim, “Techniques for estimating the bit error rate in the simula-
tion of digital communication systems," IEEE J. Select. Areas Commun.,
vol. SAC-2, pp. 153-170, Jan. 1984.
[2] W. Tranter and K. Kosbar, “Simulation of communication systems," IEEE
Commun. Mag., vol. 32, no. 7, pp. 26-35, July 1994.
[3] M. Jeruchim, P. Balaban, and K. Shanmugan, Simulation of Communi-
cations Systems. New York: Plenum, 1992.
[4] F. Gardner and J. Baker, Simulation Techniques: Models of Communica-
tions Signals and Processes. New York: Wiley, 1997.
[5] M. Irshid and I. Salous, “Bit error probability for coherent M-ary PSK
systems," IEEE Trans. Commun., vol. COM-39, pp. 349-352, Mar. 1991.
[6] E. Pinto and J. Brandao, “On the efficient use of computer simulated
digital signals to evaluate performance parameters," IEEE J. Select. Areas
Commun., vol. SAC-6, pp. 52–57, Jan. 1988.
[7] M. Kavehrad and P. Balaban, “Digital radio transmission modeling and
simulation," in Proc. IEEE GLOBECOM ‘89, Dallas, TX, pp. 1289-1294,
Nov. 1989.
[8] ETSI EN 302 307 V1.1.1 (2005-03); “Digital video broadcasting (DVB);
second generation framing structure, channel coding and modulation
systems for broadcasting interactive services, news gathering and other
broadband satellite applications."
[9] J. Proakis, Digital Communications, 3rd ed. New York: McGraw Hill,
1995.