International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.

2, December 2011

A FULL PERFORMANCE ANALYSIS OF CHANNEL

ESTIMATION METHODS FOR TIME VARYING
OFDM SYSTEMS

Aida Zaier1 and Ridha Bouallègue2

1
National Engineering School of Tunis, Tunis University, Tunisia,
[email protected]
2
High School of Communications,Tunis Ariana, Tunisia
[email protected]

ABSTRACT
In this paper, we have evaluated various methods of time-frequency-selective fading channels estimation in
OFDM system and some of them improved under time varying conditions. So, these different techniques
will be studied through different algorithms and for different schemes of modulations (16 QAM, BPSK,
QPSK,…).

Channel estimation gathers different schemes and algorithms, some of them are dedicated for slowly time
varying (such as block type arrangement insertion, Bayesian Cramer-Rao Bound, Kalman estimator,
Subspace estimator,…) whereas the others concern highly time varying channels (comb type insertion,…) .
There are others methods that are just suitable for stationary channels like blind or semi blind estimators.
For this aim, diverse algorithms were used for these schemes such as Least Squares estimator LS, Least
Minimum Squares LMS, Minimum Mean-Square-ErrorMMSE, Linear Minimum Mean-Square-Error
LMMSE, Maximum Likelihood ML,…to refine estimators shown previously.

KEYWORDS
OFDM Systems, Channel Estimation, Pilot arrangement, Blind estimation, Subspace LS, LMS, MMSE,
LMMSE, ML

1. INTRODUCTION

The increasing require for high-bit-rate digital mobile communications has incited the appearance
of Orthogonal Frequency-Division Multiplexing (OFDM) for achieving good performance in high
rate data transmission. It is also an effective technique that produces a high spectral efficiency and
a good scheme to combat frequency-selective fading channels in wireless communication systems
without forgetting the major property that is subcarrier orthogonality.

Hence, the symbol duration must be significantly larger than the channel delay spread. In
orthogonal frequency division multiplexing (OFDM) [1][2], the entire channel is divided into
many narrow subchannels. Splitting the high-rate serial data stream into many low-rate parallel
streams, each parallel stream modulates orthogonal subcarriers by means of the inverse fast
Fourier transform (IFFT).[3]

DOI : 10.5121/ijmnct.2011.1201 1
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

If the bandwidth of each subcarrier is much less than the channel coherence bandwidth, a
frequency flat channel model can be assumed for each subcarrier. Moreover, inserting a cyclic
prefix (or guard interval) results in an inter-symbol interference (ISI) free channel assuming that
the length of the guard interval is greater than the delay spread of the channel. Therefore, the
effect of the multipath channel on each subcarrier can be represented by a single complex
multiplier, affecting the amplitude and phase of each subcarrier. Hence, the equalizer at the
receiver can be implemented by a set of complex multipliers, one for each subcarrier. [3]

Under multi path spread situation, a dynamic estimation of channel is necessary before the
demodulation of OFDM signals to ensure a coherent detection and since the radio channel is
frequency selective and time-varying for wideband mobile communication systems [4][5].

In the literature, many channel estimation schemes are found and depends on if the channel is
constant, slowly or fast time varying.

Traditionally, channel estimation is achieved by sending training sequences through the channel.
However, when the channel is varying, even slowly, the training sequence needs to be sent
periodically in order to update the channel estimates. Hence, the transmission efficiency is
reduced. The increasing demand for high-bit-rate digital mobile communications makes
blindchannel identification and equalization very attractive, since they do not require the
transmission of a training sequence. Early methods for blind channel equalization exploit higher-
order statistics (H0S) of the outputs that are sampled at symbol rate. [6]

Heath and Giannakis[8]proposed a subspace method using cyclic correlation of the channel
output to blindly estimate the channel in OFDM systems. But the estimated channel error in that
study is large a subspace approach based on second order statistics is proposed to blindly identify
the channel in OFDM systems. The channel identifiability is due to the cyclostationarity inherent
in the OFDM systems with cyclic prefix. We derive a sufficient condition that guarantees all the
channels to be identifiable no matter what their zero locations are. The difference between the
proposed algorithm in this paper and the method in [7] is as follows. The approach in [7] uses the
cyclic correlation that is defined as the Fourier series expansion of the time-varying correlation of
the received data samples. In [6] , the time-invariant autocorrlation of the vector that consists of N
blocks of the received data was used.

For fast time-varying channels, many existing works resort to estimating the equivalent discrete-
time channel taps, which are modeled by the basis expansion model (BEM) [9] [10]. The BEM
methods [9] are Karhunen-Loeve BEM (KL-BEM), prolate spheroidal BEM (PS-BEM), complex
exponential BEM (CE-BEM) and polynomial BEM (P-BEM). The KL-BEM is optimal in terms
of mean square error (MSE), but is not robust to statistical channel mismatches, whereas the PS-
BEM is a general approximation for all kinds of channel statistics, although its band-limited
orthogonal spheroidal functions have maximal time concentration within the considered interval.
The CE-BEM is independent of channel statistics, but induces a large modeling error. Finally, a
great deal of attention has been paid to the P-BEM [10], although its modeling performance is
rather sensitive to the Doppler spread; nevertheless, it provides a better fit for low, than for high
Doppler spreads. In [3], a piece-wise linear method is used to approximate the channel taps, and
the channel tap slopes are estimated from the cyclic prefix or from both adjacent OFDM symbols.
[11]

Several others methods was investigated to estimate channel by using the Cramer-Rao Bounds
(CRBs), which give fundamental lower limits of the Mean Square Error (MSE) achievable by any
unbiased estimator. A Modified CRB (MCRB), easier to evaluate than the Standard CRB
(SCRB), has been introduced in [12] [13]. The MCRB proves useful when, in addition to the

2
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

parameter to be estimated, the observed data also depend on other unwanted parameters. More
recently, the problem of deriving CRBs, suited to time-varying parameters, has been addressed
throughout the Bayesian context. In [9], the authors propose a general framework for deriving
analytical expression of on-line CRBs. [14]

2. SYSTEM MODEL
The baseband OFDM system is practically the same for all the schemes of channel estimations
and differs just from the bloc of the channel but some schemes can add another bloc used
especially for interpolation or for equalization,…
Then we will present in this section the major systems used by the majority of channel estimation
schemes.

2.1.OFDM System for Channel Estimation based on Pilot Arrangement

The OFDM system model used for training sequence (pilot signal) consists of mainly of a mapper
bloc forwarded by a S/P conversion, then there is an insertion pilot bloc follow up of an IDFT
calculation of the information data. After that we find the guard insertion bloc and a P/S
conversion before reaching the channel which is affected by an AWGN noise. The data stream
will be converted on a parallel stream, then the guard interval is removed and we will sail towards
the frequency domain. Channel estimation is afterward performed before carrying out a P/S
conversion and attainment of the demapper bloc to restore back the data stream.

We will give here an overview about the routing circuit of the transmission and estimation
process according to pilot arrangement channel estimation.
Assuming first that a 16 QAM modulation will be used here, the input binary data is mapped
according to this modulation scheme. After crossing the S/P bloc, the pilot used here will be
inserted in all sub carriers of one OFDM symbol with a specific period or uniformly between data
sequence for a block pilot type estimator or in some specific sub carriers for the comb pilot type.
Then, the data sequence will pass up the IDFT block for the transformation to time domain and
the expression of x(n) (N being the DFT length) is given as follow:

     0,1,2, … . ,
 1

 (1)
   # $%& '/

!"

After that, the cyclic prefix will be inserted to preserve orthogonality of the sub carriers on the

hand. The guard time which contains this cyclic prefix having a length
 is chosen to be grater
one hand and to evade inter symbol interference between adjacent OFDM symbols on the other

than the delay spread. Then the resulting symbol is:



,   
 , 

1, . . . , 1
  
,   0,1, … ,
 1
(2)

After a P/S conversion, the OFDM symbol will cross the channel expected to be frequency
selective and time varying fading channel with an additive noise and will be given by:

3
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

)     ⊗ hn

w
(3)
h(n) is the channel impulse response represented by [16]
7

-   -. # $ 1 23 45 6 λ  τ8 , 0 9n 9 N1
/0

.!"
(4)
Where r is the total number of propagation paths, hi is the complex impulse response of the ith
path, fDi is the ith path Doppler frequency shift, λ is the delay spread index, T is the sample
period and τ8 is the ith path delay normalized by the sampling time. [17]
The guard interval is removed after the return to continuous domain from the S/P block and the

)  ;<= 
 9  9
 1
resulting signal is:

)  ) >

 ?   0,1,2, … ,
 1
(5)

The frequency form of this resulting signal will be expressed as follow:

@  )   0,1,2, … ,
 1

 ∑ )# $ 1 
  /0B5 (6)
 '!"

By supposing a transmission without an inter symbol interference ISI, the relation between the
frequency components is [18]:

@  C

D (7)

Where C  -,  is the inter-carrier interference because of the Doppler
frequency and D  E
The pilot signals are then extracted and cross the channel estimation block, after that the
estimated channel He(k) for the data sub-channel is obtained and the transmitted data is estimated.
At last, the binary data sequence is recovered by the signal demapper block.

2.2.OFDM System for Blind Estimation

In OFDM systems, the serial data are converted into M parallel streams. Each parallel data stream
modulates a different carrier. The frequency separation between the adjacent carriers is 1/T,
where T is the symbol duration for the parallel data that is M times of the symbol duration for the
serial data. Let us consider an OFDM signal in the interval (nT,(n+1)T) as [6]:

M

FG   HI # $JKL

I!"

(8)

4
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

Where HI  are symbols resulting from a modulation constellation like 16 QAM.EI is the
frequency of mthcarrier that is N . The M samples that are sampled at G  
O , O 
%& 
4 4
0,1, … , P  1 are as follow:
M

FP
O   HI # $ Q I.
/0

I!"
(9)

From this equation, the M samples can be seen as the inverse discrete Fourier transform (IDFT)
of a block for M input symbols.
Theoretically speaking, when the number of carriers is large enough, symbol duration T is much
larger than the duration of FIR channel; IS1 is negligible. However, for the high-bit-rate
communications, it is impractical to choose very large M to make
ISI negligible. Therefore, a cyclic prefix of length P is added into each block of IDFT output at
the transmitter. The length of the prefix is chosen to be longer than the length of the channel
impulse response in order to avoid inter-block interference (IBI). That results with total
cancellation of IS1 and inter carrier interference (ICI). The input data will be as follow:
M

FP
R
O   HI # Q
/0
$ .S

I!"
O  0,1, … , P
R  1

Where FP
R
O, O  0,1, … , R  1 denotes the cyclic prefix.
(10)

Assuming that the channel is frequency selective affected by an additive white Gaussian noise
(AWGN), The received signal r(n) will be degraded by theses two conditions. We will suppose
also that the length L of the channel impulse response is known. Assuming that blocks are
synchronized and carrier frequency offset is corrected [19], the receiver removes the first P

samples of received signal to obtain ). , O  0,1, … , P  1. If the cyclic prefix duration is equal
symbols corresponding to the cyclic prefix and performs an M-point DFT on the remaining

or more than the channel duration, i.e.R T U, it is shown that [20]

2W
(11)
).   H. C V OX
Y. 
P

Where H(.) is the frequency response of the channel. It is evident from the equation above that the
ISI is completely cancelled and the effect of the channel at the receiver is simply a complex gain
and AWGN

2.3.OFDM System for Channel Estimation based on Kalman Filter

For this scheme of estimation, the OFDM system model used is a standard one in which the
information symbols are grouped into blocks and inverse discrete Fourier transform (IDFT) is
performed on each block and cyclic prefix (CP) added before they are fed into the modulator and
transmitted. At the receiver, DFT is performed on each received OFDM symbol after the CP is
removed. With proper CP extensions, carrier synchronization and sample timing of tolerable
leakage, the sample from the kth subcarrier for the nth OFDM symbol is [21]:

)   C s[ 

w[ ,   1, … ,
, ∞ ]
] ∞
(12)

5
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

Where s[  is the kth information symbol of the nth OFDM symbol, C  is the gain of the kth
sub-channel during the nth OFDM symbol, w[  is the noise, and N is the total number of the
subcarriers. In this study, we assume that s[  is drawn from a BPSK constellation {−1, +1}
independently for different k and n, and w[  is a circular Gaussian random variable with zero
mean and variance ^'% and i.i.d for different k and n. We also assume that the channel is Rayleigh

fading process. Specifically, we have the correlation of the channel gain C  as
fading and we use Jakes’ model [22] for the power spectral density and Doppler spectrum of the

= ,_ N  `C Hbc n  m

(13)
1  g2Wh  ^L /
 e" 2W;f N
1
4W²h  ²^L² /²

where;f is the maximum Doppler frequency, ^L the maximum delay spread of the channel and T
the OFDM symbol duration, and e" .  is the zeroth-order Bessel function of the first kind.[21]
In general, the dynamics of the gain C  can be well modeled by an autoregressive model AR.
Defining h[n] as -kl m kC , … , C l4 , the kth order AR model for h[n] can be written as:

-kl    nkOl-k  Ol
opkl
.! (14)

Where nk1l, … , nkl and o are N*N matrices and pkl is a N*1 vector white Gaussian process.
nk1l, … , nklando are the model parameters which are obtained by solving a Yule Walker
equation using the expression of = ,_ N.

By defining kl m k-4 kl, … , -4 k  

1ll4 , we write:
Based on the AR model of the channel, a state-space model for the OFDM system can be built.

kl  qk  1l
rskl

(15)
The equation above can be seen as the state equation with skl being an AWGN noise. C and G

nk1l nk2l u nkl

are given by:

q  t  0 u 0 w
0 … v  v 0

r:  ko, 0y , … , 0y l4

 and0 are respectively N*N identity matrix and all zeros matrix. The observation equation of
the state space model is a vector version of y[n]

ykl  klkl
Ekl
(16)

Where )kl m k) kl, … , ) kl l4 , D m k{kl, 0 , … , 0 l4 , Ekl m kE kl, … , E kl l4 and
{kl is an N*N diagonal matrix with F klbeing its kth diagonal entry.

6
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

2.4.OFDM System Model based Bayesian Cramer-Rao Bounds (BCRBS)

sub-carriers and a cyclic prefix of length
 . Supposing that | denotes the sampling time and
The model adopted in this case is very near to the classical system model that is consists of N

s 


 , the duration of an OFDM symbol will be   s| .

symbol ' given as '  k' } % ~ , ' } %

1~ , … , ' } %  1~l4 ' kl are
  
In the case of a transmission over a multipath Rayleigh channel and for a nth transmitted OFDM

modulated 4-QAM or 16QAM ), the nth received OFDM symbol )'  k)' } % ~ , )' } %

 

1~ , … , )' } %  1~l4 is as follow: [14]



)'  C' '

E' (17)

whereE' is a N × 1 zero-mean complex Gaussian noise vector with covariance matrix ^ ²  , and
C' is a N × N diagonal matrix with diagonal elements given by:

†
1
kC' l  k€_  # $%& 1 /„… l
B‚ƒ ƒ
'
,


_! ((18)

L is the total number of propagation paths, €_ is the lth complex gain of variance ^‡_
²

∑_! ^‡_  1 and ˆ_  | is the lth delay (ˆ_ ]
 )

with
† ²

The L individual elements of €_ are uncorrellated with respect to each other. They are wide-
'

spectrum [23] with Doppler frequency fd. It means that €_ are correlated complex gaussian
'
sense stationary narrowband complex Gaussian processes, with the so-called Jakes’ power

variables with zero -means and correlation coefficients given by:

‰‡_  ` }€_ €_ ~  ^‡_ e" 2W;f 

' 'Š‹ %

(19)

)'  ŒOH' €'

E'
Then, the observation model for the nth OFDM symbol can be re-written as:
(20)

Where €'  k€ , … , €† l4 is a U  1vector and Fis the
 UFourier matrix defined by:[14]
' '

3. CHANNEL ESTIMATION ALGORITHMS

All the channel estimation schemes presented in the previous section are achieved by using the
standard estimators such as Least Squares algorithm (LS), Minimum Mean Square Error
(MMSE), Linear Minimum Mean Square Error (LMMSE), Low rank LMMSE, Maximum
Likelihood (ML),…

7
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

Thus we will present here the majority if these algorithms and scow in the next section their
behavior and their impact on channel estimations.

3.1 Least Square Estimator:

The LS estimator is shown to be the basic algorithm and gives regular results. Used with
practically all the schemes of channel estimation, the LS estimator will be expressed as a ratio
between the input data sequence and the output.
Denote X as the input data sequence and Y as the output one, then the LS estimator of the channel

)" ) ) 4
attenuation vector can be simply expressed as:
-Ž_|    )   … 
"  
(22)

3.2.Least Mean Square Estimator:

The LMS estimator uses one tap LMS adaptive filter at each pilot frequency. The first value is
found directly through LS and the following values are calculated based on the previous
estimation and the current channel output.

The LMS estimator is used mainly for the tracking of the channel and is usually clustered with an
equalizer or a decision feedback equalizer.

3.3.Minimum Mean Square Error Estimator::

Inter symbol interference ISI is supposed usually dropped by the guard interval, thus the equation
of Y [17] will be written as:

@  -
D

(23)

Where
  ŒOH0, 1, … , 
 1

@  ŒOH@0, @1, … , @
 14

D  ŒOHD0, D1, … , D
 14

C  ŒOHC0, C1, … , C
 14   -

D"" u D
"

‘ ’ v ’ “
D u D
" 

1 $%&” 5 •
D'  #


1
(24)

Then, the MMSE estimate of h is given by [17]:

8
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

CMM–—  ‰˜™ ‰šš


@
(25)

‰˜™  `hY  ‰˜˜  ‹  ‹

Where
(26)

‰™™  `@@  ‰˜˜  ‹  ‹

^ %  (27)

of Y. ‰™™ is the auto-covariance matrix of h and ^ % is the noise variance `|D|².

design respectively the cross covariance matrix between h and Y and the auto- covariance matrix

3.4.LMMSE Estimator:

The LMMSE estimate of the channel attenuations h from the received data y and the transmitted

-Ž_II|  ‰žž…Ÿ ‰ž -Ž

symbols Xis [17][24]
…Ÿ ž…Ÿ _|

 ‰žž ‰žž
^'%  ž   -Ž_|
(28)

)" ) ) 4
Where
-Ž_|    )   … 
"   (29)

is the least-squares (LS) estimate of h, ^'% is the variance of the additive channel noise, and the
covariance matrices are

‰žž  `--‹

‰žž…Ÿ  `--Ž_| ‹
(30)

 `-Ž_| -Ž_| 
‹
‰ž…Ÿ ž…Ÿ
(31)
(32)

The LMMSE estimator (24) is of considerable complexity, since a matrix inversion is needed

averaging over the transmitted data [15], i.e. we replace the term  ž  in (27) with its
every time the training data in Xchanges. We reduce the complexity of this estimator by

expectation

` ž  . Assuming the same signal constellation on all tones and equal probability on all
constellation points, then we get ` ž   `|1/ |², here Iis the identity matrix. Defining

{
‰  `  %  /^'%

the average signal-to-noise ratio as

(33)

, we obtain a simplified estimator

9
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

¡ 
-Ž  ‰žž V‰žž
X -Ž_|
{
‰ (34)

Where ¡  `  %  ` ¢ / ¢ ²is a constant depending on the signal constellation. In the case of 16-

£B
QAM transmission, ¡  17/9. Because Xis no longer a factor in the matrix calculation, no
inversion is needed when the transmitted data in Xchanges. Furthermore, if ‰žž and SNR are

known before hand or are set to fixed nominal values, the matrix
‰žž ”‰žž
–§ • needs to be calculated only once. Under these conditions the estimation
¦

requires N multiplications per tone.

3.5.Optimal Low - rank Estimator

This estimator is generally used to overcome the high computational complexity of the LMMSE
algorithm

The optimal rank reduction of the estimator in (30), using the singular value decomposition
(SVD), is obtained by exclusion of base vectors corresponding to the smallest singular values
[25]. We denote the SVD of the channel correlation matrix

‰žž  ¨ΛU ª (35)

p" , p , … , p and Λis a diagonal matrix, containing the singular values
whereUis a matrix with orthonormal columns

«" T « T u T « T 0 on its diagonal. This allows the estimator in (18) to be written

-Ž  ¨∆U ª -Ž_|
(36)

6  ¯   0,1, … ,
 1on its diagonal. The best rank -p approximation of the estimator in
­B
where∆is a diagonal matrix containing the values

­B ®
°1±
(18) then becomes

0 ªŽ
-ŽŠ  ¨ V Š X U -_|
0 0
∆

(37)

Where ∆Š is the upper left p× p corner of ∆.

Viewing the unitary matrix U ª as a transformy, the singular value « of ‰žž is the channel energy
contained in the kth transform coefficient after transforming the LS estimate-Ž_| . [17].

3.6.Maximum Likelihood Estimator

This estimator is well adapted for slowly time varying channel.

expressed as:   k" … . Ml where I  k",I , … , ² ,I l4 is an OFDM symbol. For a
Considering an OFDM system model with Nc carriers and M OFDM symbols. The input data X is

single OFDM symbol, the received signal is:

10
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

)  ³C

(38)

Where ³ denotes the Hadamard product of the columns of  with H.

I  k² ´ ,I … ² ,I ",I … ² ,I l4 where
Š is the size of the prefix and  
After IFFT and CP insertion, the transmitted signal is

k" … . Ml. [25]

After prefix removal and DFT computing, the received signal is given as follow:

)I  I ³C
 (39)

Where h designs the channel -  kh" … . hµ l with
ž 9
Š and H is the FFT of h; C 
kH" … . H² l4 .

As (34) shows, the OFDM system can be described as a set of parallel Gaussian channels.
Because the time- domain channel has a finite length (smaller than the prefix length in a

without loss of generality   k1 1 … .1l4 , y will be written as [25]:

welldesigned OFDM system), these parallel channels feature correlated attenuations. Considering,

-
) } ~
  C

0
(40)

Where  is a
¶ 
¶ FFT matrix. The vector ) is a Gaussian random variable with mean
k- 0l4 and covariance q'' . However, the signal part of y is contained only in the space
spanned by its mean. Separating the “signal subspace” from the “noise-only subspace,” the
received signal can be rewritten as

-
)  kž ' l } ~

0 (41)

=  ž# )  -
ž#   -
s
The reduced signal space is then drawn by:

(42)
Where s is a zero mean Gaussian noise of covariance q¸¸  ž‹ q'' ž and ž#  ž‹ ž  ž‹ . If
q''  ^'% ² , s is a white Gaussian noise of covariance matrix ^'% µ . The reduced space
(Gaussian) signal has a log-likelihood function expressed by [25]

log ;=   logW detq¸¸   ž# )  -‹ q¸¸


ž# )  -
(43)
Maximizing this log-likelihood with respect to y leads to the ML estimator given by

¿  ž ž# )  RÀ )
C µ
(44)
whereRÀµ denotes the orthogonal projection on the column space of ž (RÀµ  ž ž‹ ž ž‹
3.7.Kalman Channel Estimator

Assuming that the OFDM system model used here is whose described in section I, then the state-

C . The algorithm is standard and is given below [21]:

space model of (14) and (15) allows us to use Kalman filter to adaptively track the channel gain

11
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

1. Initialize the Kalman Filter with k0l  0Š and Σ"  Σ where Σis the stationary
covariance of kl and can be computed analytically from the expression of = ,_ N given in
(12) .

2. For each n, do the Kalman Filter update according to:

P'  qΣ' q ‹
rr ‹ ,

ΓL  klP' ‹ kl
^Ã%  ,

Ä'  P' ‹ klΓ',

kl  qk  1l
Ä' )kl  klqk  1l,

Σ'  >Š  Ä' ' ?P' ,

3. Channel estimate at instance n is

-Žkl  k , Å , … , Å lkl

(45)
Noted that the algorithm needs the information symbol F kl , so is working in the training or
decision- feedback mode.
The vector Kalman-filter algorithm gives the optimal linear estimate of the channel. Its drawback
is the high complexity. Considering that the dimension of the state vector is pN which can be
significantly high when there is large number of subcarriers.

a per-sub-channel fashion. Consider the kth sub-channel gain C . It can be modeled as a one-
One solution to reduce the complexity of the Kalman-filter channel estimator is to implement it at

dimensional AR process
Š

C     H. C   O
^p kl
.!
where the model parameter H , … , HŠ and ^ can be computed from the correlation = , 0 of (12)
(46)

according to a Yule- Walker equation. Note that these parameters do not have index k. This is
because all the subcarriers have the same statistics and fit in the same AR model. This may
greatly simplify the channel estimator for many components are shared by the estimator for each
sub-channel. [21]
Tracking the similar process as that of the previous section, the state-space model for the kth sub-

 kl  q k  1l
s kl
channel is obtained and expressed as:
(47)

) kl  Πklkl
E kl (48)

Where  kl  kC , … , C   Æ

1l4 ,   k^, 0, … ,0l, Œ  F k0l, 0, … ,0l

H H% u HŠ
qt 1 0 u 0 w
0 u 1 0

12
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

The Kalman filter for the model of  kl is a p-dimensional one which is much simpler. We
present it as follows.

1. Initialize the Kalman Filter with  kl  0Š and Σ ,"  Σ where Σis the stationary
covariance of  kl and can be computed analytically from the expression of = ,_ N given in
(12) .

2. For each n, do the Kalman Filter update according to:

P'  qΣ' q ‹
‹ ,

Ç ,'  Œ klP' Œ ‹ kl

^Ã% ,

Ä ,'  P' Œ‹ kl/Ç ,' ,

 kl  q k  1l
Ä ,' ) kl  Œ klq k  1l,

Σ,'  >Š  Ä ,' Œ kl?P' ,

3. Channel estimate at instance n is

¿ kl  k1,0, … ,0l kl

C
. (49)

The Kalman filter estimator proposes an extended version but its drawback is when non linearity
is high, the solution consists of the extension of the approximation by taking additional terms in
Taylor’s series

3.8.Subspace Approach for Blind Estimation

the vector H  kH" , H , … , HM l4 as the nth block of data and F 
We consider a data model for the OFDM system model described in the section I. Let us indicate

kFy , kFy% , … , kF" l 4

as a sequence of asa sequence of the nthblock of the IDFT
output and embedded cyclic prefix, where K = M + P and P is the length of cyclic prefix.
Denoting D È # $ Q , O  0,1, … , Ä  1 , g  0,1, … , P  1, the nth transmitted data sequence will
/0

be:

F  DH

By considering N blocks of data,  kH4 , H  14 , … , H 

14 l4 ; s kF4 , F 
(50)

É   Ê D where  is an

identity matrix and Ê is the

14 , … , F 

14 l4 and D
Kronecker product; the received signal is expressed as:

ÉH

=  CF
  CD

With F  D É H and ris an (NK - L) 1 vector. Let the nthblock of received signal be denoted as
(51)

=  k=y, k=y%, … , k=" l4 then =  k=4 , =  14 , = 

24 , = 

14 1: Ä  U4 l4 where = 

14 1: Ä  U4 is a Matlab notation standing for the first K -
L elements of = 

1. b is a noise vector that is assumed to be zero mean white

13
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

Gaussian noise with variance matrix ^²y† and be mutually independent with the input symbol
sequence. His an (NK-L) x NK matrix defined as:[6]

h" … h† 0 … 0
0 h" u h† 0 u u 0
CË Ì
’ u v u v ’
h" … h†
(52)
0 …u 0

É n then (51) becomes

Denote n È CD
=  nH


Since the matrix nshould be full column rank for the channel to be identified, we give a sufficient
(53)

condition for full rank requirement; this will be verified under the assumption U 9 R
[6].
In fact it is the inserted prefix that makes matrix na “tall” matrix and be possible to be full
column rank. The full column rank condition can always be satisfied aslong asN is chosen to be

The identification problem is based on the 
Ä  U  
Ä  U autocorrelation matrix ‰7 of

large enough.

the measurement where ‰7  `== ‹ . Considering the expression given in (53), this
autocorrelation matrix will be:

‰7  n‰Í n‹
^²
(54)
If matrix A is fullcolumn rank and the autocorrelation of input ‰Í , is alsofull rank, then
=H#n  =H#n‰Í n‹ . Let us define the noise subspace for ‰7 , to be the subspace
generated by PN - L eigenvectors corresponding to the smallest eigenvalue, and let r 
kr , … , rS† lbe the matrix containing those eigenvectors. Then, r spans the null space of
n‰Í n‹ and is orthogonal to its range space:

r.‹ n  0 O  1, … , R
 U
(55)

square sense. Let us define Î as:

For determining the identifiability of the channel, the equation above will be solved in the least

G ," … G ,Ð 0 … … 0
0 G ," u G 0 u u 0
Î Ë ,Ð
Ì
(56)
’ u v u v ’
0 …u 0 G ," … G ,Ð

Where e  KN  L  1. Next, r 4 C  -4 Î . This leads to the minimization problem:

S†

-Ž  arg min  rŽ ‹ C D
ÉDÉ ‹ C ‹ rŽ
!
(57)

S†

 arg min  -‹ ÎÖ D
ÉDÉ ‹ ÎÖ ‹ -
!

S†

 arg min  -‹ Ψ -
!

14
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

Where Ψ  ∑S† Ö É É ‹ ֋ Ž
! Î D D Î . Minimization is subject to the constraint Ø-Ø  1 Therefore, -is
given by the eigenvector corresponding to the smallest eigen value of Ψ. [6]

4. SIMULATIONS RESULTS
In this section, we will evaluate the performance of each method of channel estimation presented
in the previous sections in term of Bit Error Rate BER and Mean Square Error MSE, Root Mean
Square Error RMSE and some figures will display the estimated samples behind the transmitted
samples.

For all the algorithms considered, the channel adopted is a Rayleigh fading one with either 16-
QAM or BPSK modulations

MSE for LS,MMSE and LMMSE algorithms

0
10
MMSE
LS
LMMSE

-1
10
mean squared error

-2
10

-3
10
2 4 6 8 10 12 14 16 18
SNR in DB

Figure 1 : MSE vs SNR with a 16 QAM modulation

BER vs SNR dB for a block pilot arrangement withFFTsize=1024

LS
-0.32 LMMSE
10 Lr-LMMSE

-0.33
10
BER

-0.34
10

0 5 10 15 20 25 30 35 40
SNR dB

Figure 2 : SNR vs BER with a 16 QAM modulation

15
International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

SNR vs MSE for MMSE and LS Estimators with BPSK modulation

-1
10

-2
10

mean squared error

-3
10

-4
10
5 10 15 20 25
SNR in DB

Figure 3 : SNR vs MSE with BPSK modulation

Original and Estimated signal with Kalman Estimator

1.6
transmitted data
estimated data
1.4

1.2

0.8

0.6

0.4

0.2
0 5 10 15 20 25 30 35

Figure 4: Kalman Filter Estimator

Combined plot
700
original
600 estimated
MSE

500

400

300

200

100

-100
0 10 20 30 40 50 60 70 80 90 100

Figure 5: MSE evaluation for an Extended Kalman Filter Estimator

16
International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

System output with LMS Estimator

2

1.5

True and estimated output

0.5

-0.5

-1

-1.5

-2
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Samples

Figure 6: LMS Estimation display of transmitted and received symbols

Error curve between transmitted and estimated signal with LMS estimator
0
10

-1
10
Error value

-2
10

-3
10

-4
10
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Samples

Figure 7: LMS Error Curve between transmitted and estimated signals

Real part of Channel Transmitted Symbols
2 4
Estimated
Accurate 2
1
0
0
-2

-1 -4
0 5 10 15 20 -4 -2 0 2 4

Imag Part of Channel Estimated Symbols

1 0.4
Estimated
0.5 Accurate 0.2

0 0

-0.5 -0.2

-1 -0.4
0 5 10 15 20 -0.4 -0.2 0 0.2 0.4
hb/h=0.044933+0.32646i

Figure 8: Subspace method for blind estimation with a 16 QAM modulation

17
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

Phase
10

RMSE (radians)
-5

-10

-15

-20

-25
-30 -20 -10 0 10 20 30
SNR(dB)

Figure 9: RMSE vs SNR for a ML Estimator of the phase

Amplitude
-5

-6

-7

-8
RMSE

-9

-10

-11

-12
-30 -20 -10 0 10 20 30
SNR(dB)

Figure 10: RMSE vs SNR for a ML Estimator of the amplitude

Frequency
35

30

25

20

15
RMSE (Hz)

10

-5

-10

-15
-30 -20 -10 0 10 20 30
SNR(dB)

Figure 11: RMSE vs SNR for a ML Estimator of the frequency

4. CONCLUSIONS

In this paper, the main methods and algorithms of channel estimation for a time varying channel
was presented and some of them performed. Theses methods can be structured in two major
categories: one is dedicated for slowly time varying channels and the one for the fast time varying
channels.

The performance of the channel was evaluated in term of bit error rate and mean square error
versus SNR and for some of them the root mean squared error. Some of results show the
difference between the transmitted and received data

18
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

The results illustrate that the Least Square is the basis of the major channel estimation algorithms
and gives usually a high BER comparatively with MMSE and LMMSE algorithms. The ML and
blind estimators are also robust estimators even if they remain more complex.
As a prospective of our work is that we can extend all these methods to the case of MIMO OFDM
channel estimation.

REFERENCES
[1] R. W. Chang, “Synthesis of band-limited orthogonal signals for multichannel data transmission,” Bell Syst.
Tech. J., vol. 45, pp. 1775– 1796, Dec. 1966.

[2] L. J. Cimini, “Analysis and simulation of a digital mobile channel using orthogonal frequency division
multiplexing”, IEEE Trans. Commun., vol. COM-33, pp. 665–675, July 1985.

[3] Y. Seok Choi, P. J. Voltz, and F. A. Cassara, “On Channel Estimation and Detection for Multicarrier Signals
in Fast and Selective Rayleigh Fading Channels”, IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 49, NO.
8, AUGUST 2001

[4] N. Chotikakamthorn and H. Suzuki, “On Identifiability of OFDM Blind Channel Estimation”, 0-7803-5435-
4/99/ 1999 IEEE

[5] A. R. S. Bahai and B. R. Saltzberg, “Multi-Carrier Digital Communications: Theory and Applications of
OFDM”, Kluwer Academic/Plenum, 1999.
[6] X. Cai and Ali N. Akansu, “A Subspace Method for Blind Channel Identification in OFDM systems”, 0-
7803-6283-7/00/ 0 2000 IEEE.

[7] G.B. Giannakis, “Filterbanks for Blind Channel Identification and Equalization,” IEEE Signal Processing
Letters,vol. 4, pp. 184-187, June 1997.

[8] R. W. Heath and G.B. Giannakis, “Exploiting input cyclostationary for blind channel identification in OFDM
systems,” IEEE l”s. Signal Processing, vol. 47, no. 3, pp. 848-856, March 1999.

[9] Z. Tang, R. C. Cannizzaro, G. Leus, and P. Banelli, “Pilot-assisted time varying channel estimation for
OFDM systems,” IEEE Trans. Signal Process., vol. 55, pp. 2226-2238, May 2007.

[10] S. Tomasin, A. Gorokhov, H. Yang, and J.-P.Linnartz, “Iterative interference cancellation and channel
estimation for mobile OFDM,” IEEE Trans. Wireless Commun., vol. 4, no. 1, pp. 238-245, Jan. 2005.

[11] H. Hijazy and L. Ros, “Joint Data QR-Detection and Kalman Estimation for OFDM Time-Varying Rayleigh
Channel Complex Gains”, 170 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 58, NO. 1, JANUARY
2010.
[12] B. Z. Bobrovosky, E. Mayer-Wolf and M. Zakai, “Some classes of global Cramer-Rao bounds” in Ann.
Statistics, vol. 15, pp. 1421-1438, 1987.

[13] A. D’Andrea, U. Mengali and R. Reggiannini, “The modified Cramer-Rao bound and its application to
synchronization problems” in IEEE Trans. Commun., vol. 42, pp. 1391-1399, Apr. 1994.

[14] H. Hijazy and L. Ros, “On-line Bayesian Cramer-Rao Bounds for OFDM Slowly Varying Rayleigh Multi-path
Channel Estimation ”, 978-1-4244-2489-4/08/2008 IEEE, IEEE ISWCS 2008.

[15] SinemColeri, Mustafa Ergen, AnujPuri, and Ahmad Bahai, A study of Channel Estimation in OFDM
Systems.IEEE TRANSACTIONS ON BROADCASTING, VOL. 48, NO. 3, SEPTEMBER 2002.

[16] R. Steele, Mobile Radio Communications. London, England: Pentech Press Limited, 1992.

[17] Aida Zaier, RidhaBouallegue, “LMMSE Channel Estimation for Block Pilot Insertion In OFDM Systems
Under Time Varying Conditions”, in MMS Conference 2011.

[18] M. Hsieh and C.Wei, “Channel estimation for OFDM systems based on comb-type pilot arrangement in
frequency selective fading channels,” IEEE Trans. Consumer Electron., vol. 44, no. 1, Feb. 1998.

19
 International Journal of Mobile Network Communications & Telematics (IJMNCT) Vol.1, No.2, December 2011

[19] J.-J van de Beek, M. Sandell, and P. 0. Borjesson, “ML estimation of time and frequency offset in OFDM
systems,” IEEE Transactions on Signal Processing, vol. 45, pp. 1800-1805, July 1997.

[20] J.M. Cioffi, “A multicarrier primer”, Amatio Communications Corporation and Stanford University.

[21] Wei Chen and Ruifeng Zhang, “Kalman-Filter Channel Estimator for OFDM Systems In Time and Frequency
-Selective Fading Environment”,0-7803-8484-9/04 IEEE 2004 ICASSP 2004.

[22] W. C. Jakes (ed), Microwave Mobile Communications,IEEE Press, New York, 1994.

[23] W. C. Jakes, Microwave Mobile Communications. IEEE Press, 1983.

[24] Jan-Jaap van de Beek, OveEdfors, Magnus Sandell, Sarah Kate Wilson, and Per Ola BÄorjesson. “On
channel esimation in OFDM systems”, In Proc. VTC'95, volume 2, pages 715{719, Rosemont, IL, July 1995. IEEE.

[25] Louis L. Scharf. Statistical Signal Processing: Detection, Estimation, and Time Series Analysis. Addison-
Wesley, 1991.

[26] L. Deneire, P. Vandenameele, L. Perre, B. Gyselinckx, and M. Engels, “A Low-Complexity ML Channel

Estimator for OFDM”, IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 51, NO. 2, FEBRUARY 2003.

[27] H. L. Van Trees, Detection, estimation, and modulation theory: Part I, Wiley, New York, 1968.

[28] S. Bay, C. Herzet, J. M. Brossier, J. P. Barbot and B. Geller, “Analytic and asymptotic analysis of Bayesian
Cramer-Rao bound for dynamical phase offset estimation” in IEEE Trans. Signal Processing, vol. 56, pp. 61-70, Jan.
2008.

Authors

PR. RIDHA BOUALLEGUE

Received the Ph.D degrees in electronic engineering from the National Engineering School of Tunis. In
Mars 2003, he received the Hd.R degrees in multiuser detection in wireless communications. From
September 1990.He was a graduate Professor in the higher school of communications of Tunis
(SUP’COM), he has taught courses in communications and electronics. From 2005 to 2008, he was the
Director of the National engineering school of Sousse. In 2006, he was a member of the national
committee of science technology. Since 2005, he was the laboratory research in telecommunication
Director’s at SUP’COM.From 2005, he served as a member of the scientific committee of validation of
thesis and Hd.R in the higher engineering school of Tunis. His recent research interests focus on mobile
and wireless communications, OFDM, OFDMA, Long Term Evolution (LTE) Systems. He’s interested
also in space-time processing for wireless systemsand CDMA systems.

ZAIER AIDA

Received the B.S. degree in 2005 from National Engineering School of Gabes, Tunisia, and M.S. degree
in 2006 from Polytechnic School of Sophia Antipolis of Nice Frrance. Her Research interests focus on
channel estimation and synchronization of OFDM and MIMO-OFDM channels under very high mobility
conditions.

20