This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts

for publication in the IEEE GLOBECOM 2005 proceedings.

Joint Channel Estimation and Prediction

for OFDM Systems
Ian C. Wong and Brian L. Evans
Wireless Networking and Communications Group
Dept. of Electrical and Computer Engineering
1 University Station C0803
The University of Texas at Austin, Austin, Texas 78712
Email: {iwong, bevans}@ece.utexas.edu

Abstract Adaptive OFDM improves the system throughput d x

by adjusting transmission parameters based on channel state

information (CSI) estimated at or received by the transmitter. xtx
The improvement, however, is conditioned on the quality of the
CSI, which can be compromised by estimation and quantization
errors, and more significantly by delay. Channel prediction has
been previously proposed to combat feedback delay. In this paper,
a novel OFDM channel prediction algorithm that uses a 2 y
1dimensional frequency estimation to determine the time-delays
and Doppler frequencies of each propagation path is investigated.
The algorithm assumes a general far-field scatterer, frequency-
selective wireless channel model and hence is applicable to a w
wide variety of wireless channel conditions. It is shown that the
method requires less feedback information and has better mean-
squared error performance than previous methods. We provide
simulation results for IEEE 802.16e channels. Fig. 1. Adaptive OFDM System Block Diagram
I. I NTRODUCTION
Adaptive OFDM systems overcome the limitation of con-
We employ a general far-field scatterer, frequency-selective
ventional OFDM by allowing the transmitter to vary the power,
wireless channel model [5], which models each propagation
modulation, and coding on each subcarrier depending on the
path as a complex exponential with unknown Doppler fre-
current channel state information (CSI) [1]. This requires the
quency, time delay, and complex amplitude. A simplified 21-
transmitter to have knowledge of the CSI, which can be ob-
dimensional frequency and amplitude estimation algorithm
tained through feedback from the receivers channel estimates,
that accounts for the specialized structure in the wireless
or through its own estimates in a time division duplex (TDD)
channel is proposed. The algorithm has the advantage of
reciprocal channel. In high mobility environments, where the
producing reliable channel estimates, and thus is a joint OFDM
Doppler frequency is high and the channel changes rapidly,
estimation and prediction solution. The proposed method can
the CSI used by the transmitter would be outdated due to the
be applied to a wide variety of wireless channel conditions, and
processing and feedback delays.
In [1], delayed CSI was shown to negatively impact the requires less feedback information and has better prediction
capacity and bit error rate of the adaptive OFDM system. performance than previously proposed methods.
Furthermore, it was shown that the use of channel prediction
can improve the performance of the system. In [2], channel II. S YSTEM M ODEL
prediction over a longer range was shown to improve the
A. OFDM System
performance of adaptive OFDM in a low-mobility environ-
ment. In [3], decision-directed and adaptive short-term channel The adaptive OFDM system model considered in this paper
prediction on the time-domain channel taps was proposed. is given in Fig. 1. The input bits are initially mapped by a
In [4], an unbiased channel power predictor was applied to bank of adaptive encoders into Ndata complex data symbols
the time-domain channel taps, and a preliminary evaluation Xdata (n, k) which corresponds to the kth subcarrier in the
of frequency domain channel prediction on all the subcarriers nth OFDM block. The constellation density for each encoder
was also presented. would depend on the predicted state of the wireless channel,
In this paper, we predict OFDM channels by using classical in which various bit and power allocation strategies may be
estimation principles. These principles assume that the param- used to either maximize the data rate or to minimize the power
eters being estimated are unknown deterministic quantities. given a bit error rate (BER) constraint [6].

IEEE Globecom 2005 2255 0-7803-9415-1/05/$20.00 2005 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.

The combination of data, pilot, and guard symbols form the get the frequency response of the time-varying channel as
N -subcarrier OFDM symbol X(n, k). This is subsequently p 1
transformed into a time domain sequence using the N -point
M

L1
Hc (t, f ) = ar,p ej2r,p t ej2p f (3)
IFFT. In order to avoid intersymbol interference (ISI), and al- p=0 r=0
low the output sequence to be effectively circularly convolved
with the channel, the last samples of the symbol samples Assuming that the OFDM system with symbol period Tsym
{xi (n)}N and subcarrier spacing f have proper cyclic extension and
i=1 are used as a guard interval (GI) and prepended
to the block to form the transmit sequence xi,tx (n) = xl (n), sample timing, the sampled channel frequency response at the
where l = ((i + N ) N ) for i = 0, 1, , N + 1 kth tone of the nth OFDM block can be expressed as
and is the modulo operator. It is assumed that is longer H(n, k)
Hc (nTsym , kf )
than the delay spread of the channel, and that the effects of
p 1
inter-carrier interference can be ignored; hence, the received
M

L1
(4)
= ar,p ej2(fr,p np k)
signal for the kth subcarrier in the nth OFDM block can
p=0 r=0
be written as Y (n, k) = H(n, k)X(n, k) + W (n, k) where
H(n, k) and W (n, k) are frequency domain channel gain where fr,p = r,p Tsym is the normalized Doppler frequency
and additive white Gaussian noise (AWGN) respectively. The and p = p f is the normalized time delay.
channel estimation block takes Y (n, k) as input and forms the
III. OFDM C HANNEL P REDICTION A LGORITHM
channel estimates H(n, k) to detect the transmitted sequence
as Xdata (n, k) = Y (n, k)/H(n, k). We assume a pilot-assisted OFDM system, in which a block
In [2] [3] [4], channel estimates are fed back to the trans- of Nblock pilot symbols is indexed in time by the set N =
mitter to enable prediction. In our method, only the estimated {ni |i = 1, . . . , Nblock } and spaced t apart to be used for
channel parameters are fed back in order for the channel channel estimation. In each OFDM pilot symbol, let K =
prediction block to generate the predicted channel estimates {kj |j = 1, . . . , Npilot } denote the set of pilot indices spaced
H(n + , k), = 1, . . . , where is the number of steps f apart. We can then perform a least-squares estimate of
ahead to predict. the channel at the pilot subcarriers using the received signal
Y (ni , kj ) and the known pilot symbols Xp (ni , kj ) as
B. Wireless Channel Model H(ni , kj ) = Y (ni , kj )/Xp (ni , kj )
The complex baseband representation of the time-varying = H(ni , kj ) + W (ni , kj )/Xp (ni , kj ) (5)
wireless channel is given as [5] p 1

M

L1
=  (ni , kj )
ar,p ej2(fr,p ni p kj ) + W


L1 p=0 r=0
hc (t, ) = p (t)( p (t)) (1)
For notational convenience, let
p=0
 
1:N 1:N 1:N
HLS = h1 pilot | h2 pilot | ... | hNblock
pilot
(6)
where p (t) is the delay, p (t) is the complex amplitude of
the pth multipath tap, and L is the number of propagation be the Npilot Nblock matrix of the least-squares estimates,
paths. Assuming a far-field discrete scatterer model, p (t) can where
be further decomposed as [5]
hl:m
i = [H(ni , kl ) H(ni , kl+1 ) ... H(ni , km )]T (7)
Mp 1


p (t) = ar,p ej2r,p (t)t (2) is the column vector of the estimates on time index ni and
r=0 pilot indices {kj }mj=l .
The initial task for the channel predictor is to determine the
where Mp is the number of rays contributing to the pth path, unknown parameters L, Mp , ar,p , fr,p , and p from the least
and ar,p and r,p (t) are the complex amplitude and Doppler squares estimates reliably. Since these terms are assumed to
frequency, respectively, for the rth ray in the pth path. Note be stationary throughout our estimation and prediction horizon,
that the random phase from the complex exponentials have the next step would be to simply plug in the future time-index
been incorporated into ar,p . Also note that the time delays and to our model and generate the predicted channel. Estimating
Doppler frequencies are all dependent on time. However, we the parameters in (5) is in the form of a two-dimensional com-
can assume that the time delay p (t) and Doppler frequency plex sum-of-sinusoids in white noise parameter estimation,
r,p (t) parameters vary slowly when compared to the OFDM and is quite well studied in the literature for radar and sonar
symbol time, and can be considered constant within the imaging and other array signal processing applications (see [8]
estimation and prediction time horizons. Tracking algorithms and references therein). Although a straightforward application
can also be employed in order to follow the time variations of of these techniques may be used, they are too computationally
these parameters [7], but these are not pursued in this work. intensive for cost-effective online implementation, and do not
Combining (1) and (2) and taking its Fourier transform, we exploit the special structure of the problem.

IEEE Globecom 2005 2256 0-7803-9415-1/05/$20.00 2005 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.

We propose using a 2 1-dimensional sinusoidal parameter B. Estimate the number of paths L

estimation approach. Notice that we can rewrite (5) as Estimation of the number of paths L is essentially a model-


L1 selection problem, wherein the Minimum Description Length
H(ni , kj ) =  (ni , kj )
cp (ni )ej2p kj + W (8) (MDL) is the method most often used due to its consistency
p=0 [8]. We employ the MDL appropriate for the modified covari-
ance averaging technique [10] given as
where
Mp 1



K  1 Nblock (K)
f K
cp (ni ) = ar,p ej2fr,p ni (9) k=+1 k
r=0 L = arg min log 1 K f
1K1
K k=+1 k
is the complex gain for the pth propagation path1 , and is the
sampled version of (2). 1
+ (2K + 1) log Nblock
Notice that (8) is now a much simpler one-dimensional 4
(12)
complex sum-of-sinusoids parameter estimation, where vari-
ous standard iterative and non-iterative algorithms have been where {fk }K
k=1 are the non-increasingly ordered estimated
proposed in the literature (see [9] and the references therein). eigenvalues of the estimated autocorrelation matrix Rf .
In this type of estimation problem, the difficulty lies in the
estimation of the frequencies of the complex exponentials C. Estimate the time-delays {p }L
p=1
(p in our case), since these frequencies enter the model in Given the estimated number of paths L, we can now find
a non-linear fashion. Once these frequencies are estimated, the time delays {p }L
p=1 . In [9], the ESPRIT method was
the complex amplitudes cp (ni ) are easily computed as a shown to perform better than other eigen-analysis techniques
linear regression. Although the maximum-likelihood estimates for frequency estimation. Thus, this is the method we pursue
for the frequencies are desirable, techniques to find these in our algorithm. Let
exactly or even approximately are highly complex iterative
procedures that are not guaranteed to converge [9]. Hence, we V1 = [IK1 0K1 ]V (13)
opted to base our algorithm on the non-iterative eigen-analysis
and
technique Estimation of Signal Parameters via Rotational In-
V2 = [0K1 IK1 ]V (14)
variance Techniques (ESPRIT) as described in [9]. We discuss
the steps in estimating the parameters below. where IK1 is the (K 1) (K 1) identity matrix, 0 is a
column vector of K 1 zeros, and V = [v1 | . . . | vL ] is
A. Estimate the autocorrelation matrix a matrix whose columns are the eigenvectors associated with
In order to use eigen-analysis based techniques, the autocor- the L largest eigenvalues of Rf . Then
relation function of the underlying process has to be estimated.
p = arg(
p )/f , p = 1, . . . , L (15)
Since there are Nblock OFDM symbols, we can use an average
of the frequency autocorrelation estimates for each symbol where arg (x) is the radian phase angle of the complex number
generated using the modified covariance method [9], x, and {
p }L
p=1 are the eigenvalues of the L L matrix

1 
N

1
block
L = (V1H V1 )1 V1H V2 (16)
Rf = (ni ) + J(ni )H J (10)
Nblock 2
i=1 D. Estimate the complex amplitudes {cp (ni )}L
p=1
where Given the time delay estimates {p }L
p=1 , the maximum-


Npilot  H likelihood estimate for the L Nblock matrix C of complex
1 (jK+1):j (jK+1):j
(ni ) = hi hi (11) amplitudes is given by
Npilot
j=K  1 H
C = EH E E HLS (17)
and J is the exchange matrix with ones on the anti-diagonal
and zeros elsewhere, and K is the size of the autocorrelation where
matrix chosen to be greater than the maximum possible C = [c1:N
1
block
| ... | c1:N
L
]
block T
(18)
number of paths2 Lmax and less than Npilot .
cq:r
p = [cp (nq ) cp (nq+1 ) ... cp (nr )]T (19)
1 Note that this is the common model used for frequency selective fading and
channels, where {cp (ni )}L p=1 are typical assumed to be uncorrelated wide- ej21 k1 ej2L k1
sense stationary narrow-band complex Gaussian processes with Jakess power
.. .. ..
spectrum [5]. We do not make this assumption here, and is thus a more general E= . . . (20)
model. j2 1 kNpilot j2 L kNpilot
2 We make the reasonable assumption that L
max is known a priori. This
e e
value is typically determined by the propagation environment, and the desired
accuracy of the channel characterization. is the Npilot L Fourier transform matrix.

IEEE Globecom 2005 2257 0-7803-9415-1/05/$20.00 2005 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.

TABLE I
Note that each of the Nblock columns of C in (17) is actu- C OMPUTATIONAL COMPLEXITY FOR THE PROPOSED ALGORITHM
ally the time-domain channel tap estimate for the frequency-
selective channel. Hence, we can write Step Computation

HM L = WC (21) Autocorrelation Estimation LNblock I 2

Estimation of No. of Paths (MDL) L( 34 I 3 + I 2 + Mp I Mp2 )
where the Ndata Nblock matrix HM L is the maximum- Frequency Estimation (ESPRIT) L(2I Mp2 + 32 Mp3 + 2Mp2 )
likelihood estimate for the frequency response of the channel Amplitude Estimation L(2Nblock Mp2 23 Mp3 )
at the Ndata data carrying subcarriers, and the Ndata L Channel Prediction LMp
Fourier transform matrix is given as
j2 ( N d ) N d TABLE II
e 1 2 ej2L ( 2 )
C OMPUTATIONAL COMPLEXITY FOR B URG P REDICTION
.. .. ..
W= . . . (22)
Step Computation
ej21 ( 2 +1) ej2L ( 2 +1)
Nd Nd

MDL and coefficients comp. L3Nblock p2
Using (21) as the channel estimates, demodulation of the Prediction Filtering Lp
OFDM symbol can then be performed. Note that minimum Interpolation filter design L 34 B 3
mean squared error estimates (MMSE) may also be generated Interpolation L(2Bt + 1)
when we assume that the channel is wide-sense-stationary with
uncorrelated scatterers (WSSUS), and this is similar to the
parametric channel estimation algorithm proposed in [7]. IV. P ERFORMANCE C OMPARISON
In comparing the proposed algorithm with previous meth-
E. Estimate Mp , fr,p , and ar,p
ods, it has been shown in [11] that prediction on the time-
Notice that we can now replace the left-hand side of (9) domain channel taps requires less complexity and still per-
cp (ni ) with its corresponding estimate cp (ni ), and proceed forms better than prediction on the subcarriers directly. Thus,
similarly as the previous steps. we need only to compare our methods with those that predict
For each path p = 1, . . . , L, on the time-domain channel taps [3] [4]. Furthermore, it was
Estimate the autocorrelation function Rp across time as in
t
shown in [7] that parametric channel estimation performs
III-A, but removing the averaging in (10), and replacing better than its non-parametric counterpart. Thus, comparing
(ni ) with with [3] and [4] directly would give our approach an unfair
N
  H advantage since their methods use a non-parametric estimator.
1
block

(p) = c(iI+1):i c(iI+1):i We compare our algorithm to using standard Wiener linear pre-
Nblock p p
diction on the downsampled time-domain channel taps in (18).
i=I

Estimate Mp with the MDL using the non-increasing Furthermore, since the Burg algorithm performs better than
eigenvalues of Rtp , which is (12) with the L, K, and the standard Yule-Walker method in deriving the appropriate
linear prediction coefficients [2], we used the Burg method in
the constant in the penalty term of 1/4, replaced by Mp ,
our simulations. We also used the MDL as described in [8]
I, and 1/2 respectively.
Mp to find the appropriate model orders for the prediction filters.
Estimate {fr,p }r=1 as in Section III-C, by replacing L We call this method the Burg Prediction method. We compare
by Mp and K by I in (13)-(16). the two methods in terms of complexity, required feedback
Estimate ap = [a1,p . . . aMp ,p ]T as in III-D, given as information, and mean-squared error performance.
 1 H
ap = EH t Et Et cp A. Complexity
where We consider only the computations after the channel esti-
mation step, i.e. Sections III-E and III-F, since this is where
j2 fMp n1
ej2f1,p n1 e the two methods differ. Due to space constraints, we present a
.. .. ..
Et =
. . .

summary of the computational complexity analysis in Tables
j2 fMp nNblock I and II. The details of the Burg Prediction can be found
ej2f1,p nNblock e in [2] [11]. The parameter p is the prediction filter order,
F. Predict the channel 2Bt + 1 is the interpolation filter order, and is the number
Now that we have estimated all the parameters needed in of steps ahead to predict. Note that in the proposed algorithm,
our model, we just plug in these parameters into our model in it is the autocorrelation estimation that dominates the required
(4) to find our predicted channel as computations, and has complexity O(LNblock I 2 ). For the
Burg Prediction, it is the MDL step that dominates, and thus
p 1

M

L1

has complexity O(LNblock 3p2 ). In our numerical simulations,
H(n + , k) = ar,p ej2(fr,p (n+)p k) (23) I = 60, and the average filter order was p = 35, thus the
p=0 r=0 complexity of the two methods are comparable.

IEEE Globecom 2005 2258 0-7803-9415-1/05/$20.00 2005 IEEE

This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.

Predicted Channel for path =1, SNR =10 dB Predicted Channel 1 coherence time ahead, SNR =10 dB
B. Required Feedback Information 5

In previous methods, the actual channel estimates need

0

to be known by the transmitter. Thus 8LNblock bytes must 5 0

Channel Gain (dB)

Channel Gain (dB)

be fed back, assuming each complex channel estimate is
10
represented by two 32-bit single-precision floating point num-
bers. In the proposed method, only the time-delays {p }Lp=1 ,
15 5

Mp ,Lp
doppler frequencies {fr,p }r=1,p=1 , and complex amplitudes 20
Actual Channel
Proposed Method, ML Estimates
Burg Prediction
Mp ,Lp
{ar,p }r=1,p=1need to be fed back. This amounts to 4L(1 + 25
Prediction Start Point

2200 2300 2400 2500 2600 2700 2800 2900

10
50
0 100 150

3Mp ) bytes. Using the simulation parameters below, our OFDM Symbol Index OFDM Data Subcarrier Index

1+3M
proposed method requires only 2Nblockp 25% of feedback Fig. 2. Snapshot of predicted channel with SNR= 10 dB. Channel trace for
required compared to previous methods. the 1st path is shown on the left, and the frequency domain predicted channel
for 1 coherence time ahead is shown on the right.

C. Mean-squared Error Prediction NMSE vs. SNR, 10 coherence times ahead Prediction NMSE vs. Length of Prediction, SNR=10 dB
5 8

The OFDM system considered is based on the IEEE 802.16e 10 9

mobile broadband wireless system [12] in the ETSI Vehicular 15

10

A channel environment (6-tap frequency-selective Rayleigh 20

11

NMSE in dB

NMSE in dB
fading channel model). We simulate the Rayleigh fading chan- 25
11.5004 12

nel using the modified Jakes simulator with 32 propagation

30
13

paths. The OFDM system has bandwidth BW = 5 MHz,

35
12.6799
7.484 8.4044 14

carrier frequency fc = 2.6 GHz, and a Doppler frequency of

40
Proposed Method, ML Estimates
15
144 45
fd 180Hz. A sampling frequency of fs = 125 BW = 5.76
Proposed Method, MMSE Estimates
Burg Prediction
50 16
MHz, and a guard interval of Ngi = 64 samples is used, giving 10 15 20 25
SNR in dB
30 35 1 2 3 4
Prediction length in Coherence Times
5

an OFDM symbol period of tsym = (N +Ngi )/fs = 55.56s.

In simulating our channel prediction algorithm, we used
Npilot = 26 pilot subcarriers spaced f = 8 subcarriers apart
across the middle of the OFDM symbol, and Nblock = 100
OFDM pilot symbols spaced t = 25 symbols apart. The
autocorrelation matrix size for time-delay estimation in (11)
is chosen to be K =  53 Npilot  = 16, and similarly for the
doppler frequency estimation I =  53 Nblock  = 60.
We ran our proposed algorithm using ML estimates as in
(17) and MMSE estimates as in [7], and compared it with the
Burg Prediction method. We generated 100 different channel
realizations and 100 iterations for each realization. Figure 2
shows a snapshot of the channel prediction algorithm for the
proposed algorithm using ML estimates and Burg Prediction
methods for SNR= 10 dB. It is shown qualitatively that the
proposed method predicts the channel better than Burg Predic-
tion. Figure 3 shows the prediction normalized mean square
error (NMSE) comparisons for the three different methods
predicting 10 tcoh ahead for varying SNR (left figure), and
predicting in varying lengths for SNR=10 dB (right figure).
The proposed methods outperform the Burg Prediction method
for all SNR values and prediction lengths, and the advantage
is more pronounced as the SNR and the prediction length
increase. This is because using the parametric model allows
us to focus our estimation on the principal components of the
underlying process, and this greatly enhances the prediction
performance. This is analogous to the performance improve-
ment brought about by using parametric channel estimation
versus non-parametric estimation as observed in [7]. Addition-
ally, the improvement in using the MMSE estimates versus the
ML estimates is minimal, and is only seen for lower SNRs.
Fig. 3. Normalized mean-squared error (NMSE) performance for the three
different methods. The figure on the left shows NMSE vs. SNR, and the figure
on the right shows NMSE vs. prediction length.
R EFERENCES
[1] M. R. Souryal and R. L. Pickholtz, Adaptive Modulation with Imperfect
Channel Information in OFDM, in IEEE Proc. Int. Conf. Comm., 2001,
pp. 18611865.
[2] A. Forenza, Exploitation of the Fixed Wireless Channel through the
Link-Adaptation Algorithm and the Channel Prediction, Masters thesis,
Institut Eurocom, 2001.
[3] D. Schafhuber and G. Matz, MMSE and Adaptive Prediction of Time-
Varying Channels for OFDM Systems, IEEE Trans. Wireless Commun.,
vol. 4, no. 2, pp. 593602, Mar 2005.
[4] M. Sternad and D. Aronsson, Channel estimation and prediction for
adaptive OFDM downlinks [vehicular applications], in Proc. IEEE
Vehicular Technology Conference, vol. 2, 2003, pp. 12831287.
[5] W. C. Jakes, Microwave Mobile Communications. Wiley, 1974.
[6] T. Keller and L. Hanzo, Adaptive modulation techniques for duplex
OFDM transmission, IEEE Trans. Veh. Technol., vol. 49, no. 5, pp.
18931906, Sep 2000.
[7] B. Yang, K. Letaief, R. Cheng, and Z. Cao, Channel estimation for
OFDM transmission in multipath fading channels based on parametric
channel modeling, IEEE Trans. Commun., vol. 49, no. 3, pp. 467479,
Mar 2001.
[8] S. M. Kay, Modern Spectral Estimation : Theory and Application.
Prentice Hall, 1988.
[9] P. Stoica and R. L. Moses, Introduction to Spectral Analysis. Prentice
Hall, 1997.
[10] G. Xu, I. R.H. Roy, and T. Kailath, Detection of number of sources
via exploitation of centro-symmetry property, IEEE Trans. Signal
Processing, vol. 42, no. 1, pp. 102112, Jan 1994.
[11] I. C. Wong, A. Forenza, R. Heath, and B. Evans, Long range channel
prediction for adaptive OFDM systems, in Proc. IEEE Asilomar Conf.
on Sig., Sys., and Comp., vol. 1, 2004, pp. 732736.
[12] Air Interface for Fixed and Mobile Broadband Wireless Access Systems,
Std. 802.16e/D5, 2004.

IEEE Globecom 2005 2259 0-7803-9415-1/05/$20.00 2005 IEEE