Low Complexity Estimation of Fast Fading Radio

Channels for Higher Order Modulation

Alireza Movahedian and Michael McGuire
Department of Electrical and Computer Engineering
University of Victoria, Victoria, BC, Canada
E-mail: {amovahed,mmcguire}@ece.uvic.ca

AbstractThis paper proposes a technique for near optimal

estimation of doubly selective fast-fading radio channels with
low computational cost. The channel is characterized by an
autoregressive (AR) model. A Kalman filter combined with a
zero phase filter smoother is used to estimate the channel gains
using the detected symbols in a turbo equalization and decoding
scheme. It is demonstrated that, compared to other methods, this
technique provides a superior bit error rate performance with a
low computational cost for fast fading channels. The proposed
technique is applied to higher order modulation schemes and
the error-rate performance is evaluated. In contrast to other
techniques for fast fading channels, the error floor for this
technique is at a much higher SNR level.
Index TermsChannel estimation, fast fading channels,
Kalman filters, zero-phase filters, autoregressive models, turbo
equalization

I. I NTRODUCTION
Accurate channel estimation for fast-fading radio channels
using the previously proposed methods generally calls for
high cost estimators, which are not feasible for many mobile
computing applications. The problem of fast-fading channel
estimation for extremely high fading rates up to 1% of
the symbol rate has been tackled by several authors, most
notably in [1] and [2], where extended Kalman filters(EKF)
are proposed, and in [3] and [4] where a zero phase filter
(ZPF) is used to de-noise the channel estimates, and thereby
combat the computational complexity. Using sufficiently large
state vectors, these methods may be applied to higher order
modulation schemes, but the computational cost becomes
prohibitive, particularly when the radio channel entails a large
number of taps.
For lower fading rates of up to 0.1% of the sample rate,
low complexity and near optimal techniques are available
for orthogonal frequency division multiplexing (OFDM) and
frequency-domain equalization schemes [5][7]. However,
these methods rely on the fundamental assumption that the
channel is nearly static over the duration of long blocks of
symbols, so they fail to work at the fading rate addressed in
this paper.
It has been demonstrated that with a ZPF, the computational
cost of the channel estimator can be significantly reduced compared to previously proposed methods [3]. In [3], a Kalman
filter (KF) is used to estimate channel based on a complexexponential basis expansion model (CE-BEM). However, the
complexity of their approach can be prohibitive for higher

order modulations where a larger number of basis functions are

needed to provide the required channel estimation accuracy.
This paper employs the approach of [3] for estimating the
channel. However, instead of CE-BEM, an autoregressive (AR)
model is fitted on channel variations. It is shown that the
computational complexity is considerably reduced, yet giving
a superior estimation accuracy. The improvement in the biterror rate (BER) is significant for a 256-QAM scheme, where
the BER does not show any error floor for an Eb /N0 of as
high as 17dB.
The paper is organized as follows. Section II describes the
system model. Section III elaborates the proposed method of
channel estimation including a technique for block processing at the receiver. In Section IV, a complexity analysis is
presented and comparison are made with the EKF approach.
Simulation results in terms of BER performance are demonstrated in Section V. An extrinsic information transfer(EXIT)
chart analysis is given in Section VI. Section VII concludes
the paper.
Notations: Vectors and matrices are denoted with lowercase
and uppercase bold characters, respectively. The matrix Kronecker product is denoted with . IM is the square M M
matrix. The transpose and complex conjugate transpose of
matrix are denoted with superscripts T and H, respectively.
The expectation operator is denoted E(). We use (x)ij to
refer to elements i to j of vector x, and (A)1i for the square
submatrix of A formed by the first i rows and columns.
Notation (A)i,j represents the element in row i and column
j of matrix A. Similar to MATLAB syntax, given some
integer M , (x)iM n refers to a vector, whose j-th element is
(x)(j1)M +i , when (j 1)M +i n. The diagonal of matrix A
is denoted diag(A). For vector x, diag(x) denotes a diagonal
matrix with the elements of x on its diagonal.
II. S YSTEM MODEL
The transmitter is a bit-interleaved coded modulation
(BICM) system, consisting of a convolutional encoder followed by an interleaver, and an M-ary modulation scheme, as
shown in Fig 1. We use the notation of [2] for single-carrier
signaling and samples. The sampling period of the signal Ts is
identical to the symbol period. A block of lp pilots are inserted
before each block of ls data symbols of the modulated signal to
generate the symbol sequence {s(n), n = 1, 2, . . . , N } of zero
mean and unit variance. Pilot segments consist of an impulse

Fig. 1. Transmitter

of magnitude lp = 2L + 1, guarded by L zeros on each

side, where L + 1 denotes the number of the channel taps.
This pilot arrangement is shown to be optimal in [8] in terms
of bounds on capacity.
In the receiver, a turbo-equalizer and decoder structure similar to that of [1], [2] is employed, as illustrated in Fig 2. However, the channel estimation and equalization are performed
separately by using two Kalman filters, a channel estimator
KF and an equalizer KF. At each iteration, a smoothing zero
phase filter is applied to the initial estimated channel gains,
g(n; l). This zero phase filter allows highly accurate channel
estimates at very low computational cost [3]. The smoothed
estimated gains, g(n; l), are then input to the equalizer fixedlag KF, producing data symbol estimates s(n), along with
their variances 2 (n). A (soft-in-soft-out) SISO demapper is
then used to generate the extrinsic log-likelihood ratio (LLR)
LM
e {c(k)} on the coded bits. The extrinsic LLRs are deinterleaved and then fed to a SISO decoder. The extrinsic
information is obtained by subtracting the a priori LLR from
the output LLR. The decoder generates LLR information,

LD
a {c(k )}, which is used by the LLR-to-symbol block to
update symbol estimates s(n) and their variance (n) for
the next iteration, as well as providing the SISO de-mapper
with new extrinsic information. The equalizer is realized by
a special form of fixed-lag KF as proposed in [1]. The SISO
blocks are implemented based on [2].
The channel is assumed to be a doubly-selective Rayleigh
fading channel with a Doppler frequency of fd , subject to
additive white Guassian noise (AWGN). Denoting the channel
gain of path lth at time n by g(n; l), the channel output is
described as
L

y(n) = g(n; l)s(n l) + v(n),

(1)

l=0

where v(n) denotes AWGN with variance v2 .

III. C HANNEL E STIMATION
A Kalman filter is used to estimate the channel gains. The
KFs output is de-noised by a ZPF with a component IIR filter
designed using an elliptical approximation. In a ZPF, the input
is processed by the component IIR filter, in the forward and
then backward direction. A ZPF is a linear-phase filter, which
provides the high-selectivity of IIR filters, without introducing
phase distortion. The same performance may be provided by a
linear-phase finite impulse response (FIR) filter, but it requires
a large number of taps to give the same selectivity, which is
not acceptable in terms of cost [9].
The ZPF has an unwanted transient response at the two
ends of the output blocks. So, a block processing method is

Fig. 2. Receiver structure (as used in [4])

Fig. 3. Block processing (as used in [4])

employed to alleviate the transient effects of the ZPF. The

ZPF is applied to extended blocks, formed by appending extra
samples from the preceding and following blocks to the current
sample block. As such, the unwanted transient effects will
only affect the extra samples. The extra samples are discarded
before the de-noised channel gains being sent to the equalizer
KF. The block processing technique works as follows. First,
the received signal is divided into blocks of N samples. Each
block contains symbols for Ni bits, where Ni is the interleaver
length. The samples processed by the ZPF are shown in
Fig 3. Each block is processed in two stages. At time step
i, the channel gains and data symbols for blocks i and i + 1
are estimated by the receiver using the iterative algorithm.
However, only the symbol decisions for block i are forwarded
to the receiver output, as the results for block i+1 are likely
to be contaminated by the transient effects of the ZPF. At the
end of each step, the final M samples of the estimated channel
gains for each propagation path of block i are stored to provide
the initial samples for the ZPF in the next time step.
The parameters of the IIR filter within the ZPF are selected
so that the contribution of the the error due to the passband
ripple, Er , and stop-band leakage, Es are less than the Wiener
bound of the optimal mean square error (MSE) for the channel
estimates. The optimal MSE is determined by the power

spectral density (PSD) of signal and noise [10]. At high SNR

where the noise power is much smaller than the signal power,
the optimal MSE can be approximated by 2v2 fD , where
fD = fd Ts is the normalized Doppler frequency. The Doppler
frequency may be estimated using the methods described
in [11]. To simplify the problem, the estimation error of the
KF is assumed to be white with variance Pe , and the PSD of
the channel gains is assumed to be constant over [fD , fD ].
Then, Er Rp2 Pl and Es Rs4 Pe , where Pl = 1/(L + 1) is the
power of each channel path, Rp is the passband ripple, and
Rs is the stopband attenuation of the ZPF. The filter order Nf
is selected so that the transition band is an order of magnitude
less than the ZPF passband width fD .
The evolution of each path over time is characterized by the
AR(p) process with
p

g(n; l) = aj g(n j; l) + w(n; l),

[
g (n; 0) g(n; L)]T as presented in Algorithm 1, where
g(n; l) = E[g(n; l)y(1), , y(n + p)] = (h(n + p))lp+p (8)
for l = 0, 1, , L.
The KF makes use of the detected data symbols, s(n),
fed back by the LLR-to-symbol block. These symbols are
contaminated with detection error, u(n), of variance (n);
that is s = s(n) + u(n). To take this error into account in
the KF calculations, the noise variance is augmented with an
additional term to give the effective noise v (n) given by
T

v (n) = [
g (n; 0)...
g (n; L)] [u(n)...u(n L)] + v(n) (9)
Assuming that the detection errors are uncorrelated with each
other and independent from the channel gains,
L

v2 = v2 + E [g(n; l)2 ] (n l)

(10)

l=0

(2)

j=1

for l = 0, . . . , L, where w(n; l) represents the noise process

of path l. Define the vector of the AR model for path l as
a = [a1 ap ]T . By solving the Yule-Walker equations [12],
one obtains
a = R1
(3)
l rl
where Rl represents the p p correlation matrix for tap l,
defined as
(Rl )i,j = Pl J0 (2fD i j),
(4)
with J0 () denoting the first-kind, zeroth order Bessel function.
Also, (rl )i = Pl J0 (2fD i) for i = 2, ..., p + 1. The variance
2
of the channel noise process for path l, w
(l), is calculated
as
2
w
(l) = Pl aH Rl a
(5)
for l = 0, 1, ..., L. The system state is characterized by an (L +
1)p 1 vector, defined as

In Algorithm 1, E g(n; l)2 ] is approximated with (h(nn

2
2
1))1pLp+1 + (diag(P))1pLp+1 .
At each iteration, the KF processes a block of length NE =
2N + ME consisting of the current and next blocks, as well as
ME symbols from the previous block used for training the KF.
Then, an extended block of length 2N + M is passed through
the ZPF, where M symbols are taken from the previous block
as shown in Fig. 3. The first M symbols of the ZPFs output
are corrupted by the transient response of the ZPF and are
discarded before being applied to the equalizer.
The Soft-Input/Soft Output (SISO) detector and decoder
uses a fixed-lag Kalman Filter equalizer as described in [1].
The -lag equalizer will estimate the data symbol for time
period n using measurements up until time n + . The computational cost of the equalizer is O (2 3 ) multiplications per
sample.

IV. C OMPLEXITY A NALYSIS

T
h(n) = [g(n; 0)g(n p + 1; 0) g(n; L)g(n p + 1; L)] .
The computational complexity of the EKF method of [2] is
(6)
2
approximated as O ( [ + Q(L + 1)] ) floating point multiThe system equations are given by
plications per symbol, per iteration, where and Q denote the
h(n) = Fh(n 1) + w(n),
(6a) equalizer delay and the number of basis functions, respectively.
y(n) = E(n)h(n) + v(n),
(6b) However, this method has a significantly high estimation error
floor which is too high to allow low bit error rate operation for
where F = IL+1 K with
any modulation scheme with an order higher than 16-QAM.
a1 a2 a2 . . . ap
Since higher-order modulations are being proposed to provide

1
0
0
...
0

the higher spectral efficiency needed for the latest generation

1
0
...
0 ,
K = 0
(7) of multimedia wireless applications. The computational cost

of the CE-BEM based method of [3] (called KF/ZPF-CE
0
2
0
...
1
0

BEM herein) is on the order of O (2 3 + 2 [Q(L + 1)] ).

E(n) = [s(n) s(n1) s(nL)] eT0 , e0 = [1 01(p1) ]T ,
w(n) = [w(n; 0) 01p1 ... w(n; L) 01p1 ]T .
The autocorrelation matrix for w(n) is Qw
=
2
2
diag ([w
(0) 01p1 ... w
(L) 01p1 ]).
A KF is used to obtain gain estimates, g
(n)

The proposed method has a cost of O (2 3 + 2 [p(L + 1)] ),

where p represents the order of the AR model. The cost of
the ZPF is comparatively negligible for small Nf , and is not
considered. In the simulations, we used L = 2 (3-tap channel),
Nf = 5, p = 4 and = 5. Comparing with the EKF, where
typically = 5 and Q = 9, the complexity of the proposed

Algorithm 1 Channel estimation Kalman filter

1: h(00)
E [h(0)]
= 0(L+1)p1
2:

P(00)

1
I
(L+1) (L+1)p

for n = 1, 2, . . . , NE do

1n 1)
4:
h(nn
1) Fh(n
3:

5:

P(nn 1) FP(n 1n 1)FH + Qw

6:
7:

(n) [(n) (n 1) . . . (n L)]

2
2

rgg (nn1) (h(nn1))

+(diag(P))

8:

v2 (n) v2 (n) + (n)

rgg (nn 1)

9:

K(n) P(nn 1)EH (n)

10:

1pLp+1

[E(n)P(nn 1)EH (n) + v2 ]

h(nn)
h(nn
1)

1pLp+1

+K(n) [y(n) E(n)h(nn

1)]
11:

P(nn) [Ip(L+1) K(n)E(n)] P(nn 1)

12:

if n > p 1 then

13:
14:
15:

(n p + 1) = (h(nn))
g
ppLp+p
end if
end for

for k = 1, 2, . . . , p 1 do

(NE k + 1) = (h(nn))
17:
g
pjpLp+pj
16:

18:

end for

method is about one order of magnitude less. For KF/ZPFCE-BEM with 256-QAM, one needs to select Q = 9, at least.
Therefore, the cost of KF/ZPF-CE-BEM is at least three
times that of the proposed method. Even with the extra cost,
KF/ZPF-CE-BEM exhibits a severe error floor for the case
of 256-QAM, as illustrated in the next section.
V. S IMULATIONS
An equipower fast-fading Rayleigh channel with L + 1 = 3
taps and Pl = 1/(L + 1) = 1/3 was considered. The channel was simulated based on [13]. A sampling interval of
Ts = 25 106 , and a Doppler frequency of fd = 400Hz were
used, giving a normalized Doppler frequency of fD = 0.01.
This Doppler frequency corresponds to a vehicle moving at
the speed of 216 km/h, when the carrier frequency is 2GHz. A
convolutional code of rate 1/2 and octal generator [133, 171]
were employed. Every symbol block comprised N = 104
symbols, in which lp = 5 pilots were inserted per ls = 20 data
symbols.
Pilot segments consisted of an impulse of magnitude

lp , guarded by two zeros on each side. The pilot cost is

10 log10 (lp + ls )/ls 0.97dB, and is not taken into account on
the x-axis of the BER plots. The ZPFs normalized passband
edge frequency was fp = fD = 0.01. To pick other parameters,
we note that the Wiener bound on the estimation error at the
highest SNR of interest, that is at Eb /N0 = 17dB for 256-QAM

with a rate 1/2 code, is calculated as W 2fD v2 104 .

Based on the discussion made in Section III, we selected the
passband ripple Rp = 0.1dB, stopband attenuation Rs = 15dB,
and the filter order Nf = 5. In block processing, M = 4000
and ME = 1000 symbols from the previous block were used to
deal with the unwanted ZPFs transient response and to train
the KF, respectively (see Section III).
An AR process of order p = 4 represented the gains variations. A lower p gives inferior BER, while higher order AR
models become computationally unstable. The variance of the
2
noise process as given by (5) is computed as w
(l) 61013
for l = 0, 1, ..., L. A better tracking behaviour is obtained
with higher values (also reported in [2]), so we selected
2
w
(l) = 107 . The equalizer KF was a fixed-lag KF as
described in [1], with = 5 lags.
The bit-error rate performances of Gray-mapped 16-QAM,
64-QAM and 256-QAM receivers are demonstrated in Fig. 4.
In this figure, legend KF/ZPF; CE-BEM(9) refers to a
CE-BEM based channel modeling using KF and ZPF with
Q = 9 bases; legend EKF; CE-BEM(Q=9) refers to the EKF
method of [2]. The proposed method is labeled with KF/ZPF;
AR(4). The detector/decoder performance with perfect channel state information (CSI) performance is indicated by the
line marked Perfect channel. For the competitor methods,
Q = 9 was selected despite the fact that the computational
cost is at least twice as high as that of the proposed method.
The length of the KFs state vector in the ZP/KF, CE-BEM
method was Q(L + 1) = 27, and that of the channel and
symbol estimator EKF was + Q(L + 1) = 32, while the
length of the state vector for the proposed method is only
p(L+1) = 12. It can be seen that, in all cases, the performance
of the proposed method over the SNRs of interest is within
0.3dB from the perfect channel performance and outperforms
the other methods. Specifically, no BER floor is perceptible.
The CE-BEM based KF/ZPF method performs well for the 16QAM case, but suffers serious error floor under higher order
modulations. The EKF method did not converge to low BERs
over the indicated range for the 256-QAM case.
VI. EXIT C HART A NALYSIS
The extrinsic information transfer (EXIT) chart is a means
by which the exchange of extrinsic information between the
SISO modules in an iterative receiver is depicted [14]. It can
be used to examine and predict the convergence properties of a
turbo receiver with different modulations and error correction
codes, without the need to perform comprehensive simulations.
The approximate BER performances of different techniques
can also be compared.
In the case of a turbo-equalizer and decoder, the EXIT
chart illustrates the interaction between the transfer function
of the SISO decoder and that of the equalizer. A transfer
function characterizes the gain of a SISO module in terms
of the input/output mutual information. For the SISO decoder,
the output extrinsic mutual information denoted with IA , is
a function of the input mutual information, denoted with IE .
The transfer function of the equalizer is defined likewise. Since

10

0.9

10

0.8
2

10

0.7
0.6

IE

BER

10

0.5

10

10

10

0.4

16QAM
64QAM
256QAM
EKF; CEBEM(9)
KF/ZPF; CEBEM(9)
KF/ZPF; AR(4)
Perfect channel

Equalizer curve (EKF)

0.2

Sample Trajectory

0.1

10

Equalizer curve (AR4)

0.3

10
11
Eb/N0(dB)

12

13

14

15

16

17

Fig. 4. BER vs. Eb /N0 under different modulation schemes

the output of the decoder is the input to the equalizer and

vice versa, the two functions can be depicted in a single plot.
Furthermore, the exchange of the extrinsic information can be
shown by a zigzag trajectory, going back and forth between
the two curves. The trajectory starts from the origin and is
comprised of horizontal and vertical lines, representing the
operation of the decoder and equalizer, respectively. By examining the trajectory, one can figure out whether the receiver
would eventually converge to a high mutual information state,
corresponding to low BERs. This is the case if a tunnel exists
from the starting point toward the high mutual information
area.
The convergence properties of the proposed method is
analyzed with an EXIT chart. Fig 5 shows the EXIT curves for
a 256-QAM receiver at Eb /N0 = 17dB. A sample trajectory
illustrates how the extrinsic information is exchanged between
the SISO components of the turbo receiver. A comparison
with the EKF-based method of [2] is also made. It can be
seen that the proposed method will reliably converge to high
mutual information state, thus giving low bit error rates. On the
contrary, the EKF cannot make the above-mentioned tunnel for
the receiver to converge to a low BER. In addition, the equalizer and decoder curves intersect at higher values of ID , IE
with the proposed method, thereby predicting a superior BER
performance. These conclusions are in compliance with the
BER results presented previously.
VII. C ONCLUSION
An accurate channel estimation approach with a computational cost that is significantly less than the previously
proposed methods was presented. An AR model was fitted to
the channel variations. The lag of the AR model was exploited
by the channel estimator KF to obtain more accurate estimates
by reducing the in-band estimation error which cannot be
removed by the ZPF. It is shown that the proposed technique
gives excellent error performance compared to the previous
art.

Decoder
0.1

0.2

0.3

0.4

0.5
IA

0.6

0.7

0.8

0.9

Fig. 5. EXIT chart for a 256-QAM receiver at Eb /N0 = 17dB.

R EFERENCES
[1] X. Li and T. F. Wong, Turbo equalization with nonlinear kalman
filtering for time-varying frequency-selective fading channels, Wireless
Communications, IEEE Transactions on, vol. 6, no. 2, pp. 691700,
2007.
[2] H. Kim and J. K. Tugnait, Turbo equalization for doubly-selective
fading channels using nonlinear kalman filtering and basis expansion
models, Wireless Communications, IEEE Transactions on, vol. 9, no. 6,
pp. 20762087, 2010.
[3] M. McGuire, A. Movahedian, and M. Sima, Zero phase smoothing of
radio channel estimates, in 20th European Signal Processing Conference (EUSIPCO 2012), 2012.
[4] A. Movahedian and M. McGuire, Estimation of fast-fading channels
for turbo receivers with high-order modulation, Vehicular Technology,
IEEE Transactions on, vol. 62, no. 2, pp. 667 678, Feb. 2013.
[5] P. Wan, M. McGuire, and X. Dong, Near-optimal channel estimation for
ofdm in fast-fading channels, Vehicular Technology, IEEE Transactions
on, vol. 60, no. 8, pp. 37803791, 2011.
[6] D. Liu and M. Fitz, Joint turbo channel estimation and data recovery
in fast fading mobile coded ofdm, in Personal, Indoor and Mobile
Radio Communications, 2008. PIMRC 2008. IEEE 19th International
Symposium on, Sept. 2008, pp. 1 6.
[7] L. He, S. Ma, Y.-C. Wu, and T.-S. Ng, Iq imbalance compensation: A
semi-blind method for ofdm systems in fast fading channels, in Circuits
and Systems (APCCAS), 2010 IEEE Asia Pacific Conference on, Dec.
2010, pp. 362 365.
[8] X. Ma, G. B. Giannakis, and S. Ohno, Optimal training for block
transmissions over doubly selective wireless fading channels, Signal
Processing, IEEE Transactions on, vol. 51, no. 5, pp. 13511366, 2003.
[9] A. Antoniou, Digital Signal Processing: Signals, Systems, and Filters.
Toronto, Ontario: McGraw-Hill, 2006.
[10] P. Z. Peebles, Probability, random variables, and random signal principles. McGraw-Hill, 1993.
[11] J. Hua, L. Meng, X. Xu, D. Wang, and X. You, Novel scheme for
joint estimation of snr, doppler, and carrier frequency offset in doubleselective wireless channels, Vehicular Technology, IEEE Transactions
on, vol. 58, no. 3, pp. 1204 1217, march 2009.
[12] J. Candy, Model-Based Signal Processing, ser. Adaptive and
Learning Systems for Signal Processing, Communications and
Control Series. John Wiley & Sons, 2005. [Online]. Available:
http://books.google.ca/books?id=PHxMKTsVELUC
[13] D. J. Young and N. C. Beaulieu, The generation of correlated rayleigh
random variates by inverse discrete fourier transform, Communications,
IEEE Transactions on, vol. 48, no. 7, pp. 11141127, 2000.
[14] S. ten Brink, Convergence behavior of iteratively decoded parallel
concatenated codes, Communications, IEEE Transactions on, vol. 49,
no. 10, pp. 17271737, 2001.