902
7, JULY 1998
Robust Channel Estimation for OFDM Systems
with Rapid Dispersive Fading Channels
Ye (Geoffrey) Li, Senior Member, IEEE, Leonard J. Cimini, Jr., Senior Member, IEEE,
and Nelson R. Sollenberger, Fellow, IEEE
AbstractOrthogonal frequency-division multiplexing (OFDM)
modulation is a promising technique for achieving the high
bit rates required for a wireless multimedia service. Without
channel estimation and tracking, OFDM systems have to use
differential phase-shift keying (DPSK), which has a 3-dB signal-
to-noise ratio (SNR) loss compared with coherent phase-shift
keying (PSK). To improve the performance of OFDM systems
by using coherent PSK, we investigate robust channel estimation
for OFDM systems. We derive a minimum mean-square-error
(MMSE) channel estimator, which makes full use of the time-
and frequency-domain correlations of the frequency response
of time-varying dispersive fading channels. Since the channel
statistics are usually unknown, we also analyze the mismatch of
the estimator-to-channel statistics and propose a robust channel
estimator that is insensitive to the channel statistics. The robust
channel estimator can significantly improve the performance of
OFDM systems in a rapid dispersive fading channel.
Index Terms Diversity combining, orthogonal frequency-
division multiplexing, robust channel estimation, time-varying
dispersive fading.
I. INTRODUCTION
M ULTIMEDIA wireless services require high-bit-rate
transmission over mobile radio channels. To reduce
the effect of intersymbol interference (ISI) caused by the
dispersive Rayleigh-fading environment [1], the symbol du-
ration must be much larger than the channel delay spread. In
orthogonal frequency-division multiplexing (OFDM) [2][9],
the entire channel is divided into many narrow subchannels,
which are transmitted in parallel, thereby increasing the sym-
bol duration and reducing the ISI. Therefore, OFDM is an
effective technique for combating multipath fading and for
high-bit-rate transmission over mobile wireless channels.
To eliminate the need for channel estimation and tracking,
differential demodulation can be used in OFDM systems, at
the expense of a 34-dB loss in signal-to-noise ratio (SNR)
compared with coherent demodulation. Accurate channel esti-
mation [5], [6], [8] can be used in OFDM systems to improve
their performance by allowing for coherent demodulation.
Furthermore, for systems with receiver diversity, optimum
combining can be obtained by means of channel estimators.
In [5], [6], and [8], a channel estimator for OFDM systems
Paper approved by N. C. Beaulieu, the Editor for Wireless Communication
Theory of the IEEE Communications Society. Manuscript received August
20, 1997; revised March 16, 1998. This paper was presented in part at IEEE
ICC98, Atlanta, GA, June 1998.
The authors are with the Wireless Systems Research Department,
AT&T LabsResearch, Red Bank, NJ 07701-7033 USA (e-mail:
liye@research.att.com; ljc@research.att.com; nelson@research.att.com).
Publisher Item Identifier S 0090-6778(98)05169-1.
00906778/98$10.00 1998 IEEE
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO.
has been proposed based on the singular-value decomposition
or frequency-domain filtering. Time-domain filtering has been
proposed in [8] to further improve the channel estimator
performance. However, the best time- or frequency-domain
filtering shapes for channel estimation has not been studied.
In this paper we investigate minimum mean-square-error
(MMSE) channel estimation for OFDM systems. We first
derive the MMSE estimator, which makes full use of the
correlation of the channel frequency response at different times
and frequencies. In particular, for mobile wireless channels, the
correlation of the channel frequency response at different times
and frequencies can be separated into the multiplication of the
time- and frequency-domain correlation functions. Hence, our
MMSE channel estimator can be a frequency-domain filter
using the fast Fourier transform (FFT), followed by time-
domain filters. Since the channel statistics, which depend on
the particular environment, are usually unknown, we present a
robust estimator, that is, an estimator that is not sensitive to the
channel statistics. Computer simulation demonstrates that the
performance of OFDM systems using coherent demodulation
based on our channel estimator can be significantly improved.
The paper is organized as follows. Section II introduces the
statistics of the mobile wireless channel and briefly describes
the OFDM system with channel estimation. Then, Section III
derives the basic MMSE channel estimator for the OFDM
systems. Next, Section IV presents a robust channel estima-
tor design approach. Finally, Section V presents computer
simulation results to demonstrate the effectiveness of this
channel estimator for OFDM systems in rapid dispersive
fading channels.
II. CHANNEL STATISTICS AND OFDM SYSTEMS
Before investigating channel estimation for OFDM systems
in mobile radio channels, we briefly describe the channel sta-
tistics, emphasizing the separation property of mobile wireless
channels, which is crucial for simplifying our MMSE channel
estimator. In this section we also briefly describe an OFDM
system with receiver diversity.
A. Statistics of Mobile Radio Channels
The complex baseband representation [10] of a mobile
wireless channel impulse response can be described by
(1)
where is the delay of the th path and is the
corresponding complex amplitude. Due to the motion of the
LI et al.: ROBUST CHANNEL ESTIMATION FOR OFDM SYSTEMS 903

vehicle, s are wide-sense stationary (WSS) narrow-

band complex Gaussian processes, which are independent for
different paths. (a)
We assume that has the same normalized correlation
function for all and, therefore, the same normalized
power spectrum Hence

(2)
where is the average power of the th path.
Using (1), the frequency response of the time-varying radio
channel at time is (b)
Fig. 1. OFDM system with channel estimator. (a) Transmitter. (b) Diversity
receiver.

(3)
kind, and its Fourier transform (FT) is
Hence, the correlation function of the frequency response for
different times and frequencies is if
(10)
otherwise
In the above expression , and is the Doppler
frequency, which is related to the vehicle speed and the
carrier frequency by

(11)

(4) where is the speed of light. For example, for a system with
carrier frequency GHz, Hz when the user
where is the total average power of the channel impulse is moving at 60 mi/h.
response defined as
B. OFDM Systems with Channel Estimator
(5) The OFDM system with channel estimation considered in
this paper is shown in Fig. 1. The ReedSolomon (RS) code
(6) across tones is utilized in the system to correct the burst errors
resulting from frequency-selective fading. Since the phase of
each subchannel can be obtained by the channel estimator,
It is obvious that Without loss of
coherent phase-shift keying (PSK) modulation is used here to
generality, we also assume that , which, therefore, enhance the system performance.
can be omitted from (4). For a diversity receiver, the signal from the th antenna at
From (4), the correlation function of can be sep- the th tone and the th block can be expressed as
arated into the multiplication of a time-domain correlation (12)
and a frequency-domain correlation is
In the above expression is additive Gaussian noise
dependent on the vehicle speed or, equivalently, the Doppler
from the th antenna at the th tone and the th block, with
frequency, while depends on the multipath delay zero-mean and variance We also assume is inde-
spread. With the separation property, we are able to simplify pendent for different s, s, or s. , the frequency
our MMSE channel estimator described in the next section. response at the th tone and the th block corresponding to
For an OFDM system with block length and tone spacing the th antenna, is assumed independent for different s,
(subchannel spacing) , the correlation function for different but with the same statistics. is the signal modulating
blocks and tones can be written as the th tone during the th block and is assumed to have unit
variance and be independent for different s and s.
(7) With knowledge of the channel parameters, can be
where estimated as by an MMSE combiner

(8)
From Jakes model [11] (13)

(9)
where is the zeroth-order Bessel function of the first However, the multipath channel parameters are time varying
904 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 7, JULY 1998

and are usually unknown. Hence, a channel estimation algo-

rithm must be derived to obtain accurate estimation of the
channel parameters.
Since the channel corresponding to each antenna has the
same statistics, the channel estimator for each antenna has
the same coefficients. Furthermore, the estimator for each
antenna works independently since the signal from the other Fig. 2. Channel estimator for OFDM systems.
antennas carries no information about the channel parameters
corresponding to this antenna. Therefore, the subscript is
which is uniquely determined by
eliminated from in the next two sections.
(23)
III. MMSE CHANNEL ESTIMATION
If the reference generator in Fig. 1 can generate an ideal The direct current (dc) component in can be found
reference , then a temporal estimation of can by
be obtained as
(24)
(14)
The s and s are the corresponding eigenvalues and
where the superscript denotes the complex conjugate. As eigenvectors of the frequency-domain correlation matrix ,
indicated in the previous section, s for different s defined as
and s are correlated; therefore, an MMSE channel estimator
can be constructed as follows:
.. .. .. ..
. . . .
(15)
(25)
where s are selected to minimize is obviously a unitary matrix.
(16) The MMSE channel estimator derived from (20) is shown
in Fig. 2. The unitary linear inverse transform and trans-
in the above expression is the number of tones in each form in the figure perform the eigendecomposition of the
OFDM block. frequency-domain correlation. The estimator turns off the zero
Denote or small s to reduce the estimation noise. For those large
s, linear filters are used to take advantage of the time-
.. domain correlation.
. It should be noted that a channel estimator described in
(17) [5] uses only the frequency-domain correlation functions.
.. However, the MMSE estimator described here exploits the
.
channel correlations in both the time- and frequency-domains,
resulting in better performance than the estimator in [5].
As shown in the appendix, the average MMSE of the
(18)
channel estimator is

(19)

Then, using the separation property (7), it can be shown (see

the Appendix) that the estimator coefficients are given by

(20)
(26)
In the above expression is a diagonal matrix with the
th diagonal element For Jakes model, Then by direct calcu-
lation
(21)
(27)
where is a stable one-sided FT

(22)
(28)
LI et al.: ROBUST CHANNEL ESTIMATION FOR OFDM SYSTEMS 905

where band-limited time-domain correlation gives the worst mean-

square-error (MSE) performance among all channels.
(29)

if IV. ROBUST CHANNEL ESTIMATOR DESIGN

if Once the channel statistics, such as the time-domain corre-

lation and frequency-domain correlation, are known, the opti-
(30) mum channel estimator can be designed. However, in mobile
wireless links the channel statistics depend on the particular
If the time-domain correlation is ideal -band-limited, i.e., environments, for example, indoor or outdoor, urban or sub-
urban, and change with time. Hence, it is not robust to design
if a channel estimator that tightly matches the channel statistics.
(31)
otherwise In this section we first analyze the performance degradation
then due to a mismatch of the estimator to the channel statistics and
develop a robust estimator design approach. Then, we briefly
(32) discuss the design of finite-tap robust estimators.

A. Mismatch Analysis
If an MMSE channel estimator is designed to match a
(33) channel with time- and frequency-domain correlations
and , respectively, then its coefficients are
determined from (20) by
For any -band-limited function satisfying (37)

(34) where the definitions of and are similar to those of

and except that and there are replaced by
and , respectively.
we have
For a channel with time- and frequency-domain correlations
and , rather than and , from (16), the
MSE for the designed channel estimator is

(35)

with equality if and only if almost everywhere.

In the above derivation the inequality for all
has been used. Hence

(38)

where and
if
(36) (39)
otherwise.
The above inequality suggests that a channel with an ideal Substituting (37) into (38), we obtain a general formula for
906 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 7, JULY 1998

the MSE of the mismatched channel estimator If , that is, the time-domain correlation of
the designed estimator is ideal -band-limited, then, for any
channels with zero outside and

(45)

we have

(46)

(40) Therefore

1) Time-Domain Correlation Mismatch: Assume that the

frequency-domain correlation of the channel estimator matches (47)
that of the channel, i.e.,
Hence, if an OFDM channel estimator is designed using
(41) as the time-domain correlation, then the time-domain
for , and the time-domain correlation is correlation mismatch of the estimator will not degrade its
mismatched. Then performance. This suggests that a robust channel estimator
should use as the time-domain correlation.
(42) 2) Frequency-Domain Correlation Mismatch: To analyze
the frequency-domain correlation mismatch, we assume that
and the time-domain correlation of the designed estimator is the
same as that of the channel, that is, , and the
frequency-correlation matrix of the designed estimator has the
same eigenvectors as that of the channel. That is, can be
eigendecomposed into

or (48)

where and . and for

are generally different.
Although the second assumption seems strange, it is, in fact,
reasonable. As indicated in [5], with tolerable leakage, both
matrices and can be approximated by the discrete FT
(DFT) matrix that is defined as

(43)
The first term in the above expression represents the MSE
variation due to the mismatch. By means of (24) and (21),
can be further simplified into
(44)
.. ..
. .
(49)
The leakage of the above approximation depends on the guard
interval and the channel multipath delay profile. If the delay
of a path is an integer multiple of the sampling interval ,
that is, , then all of the energy from the path will be
mapped to Otherwise, if the delay is a noninteger multiple
of the sampling interval, that is, , then most
of its energy will be contained in and , although the
energy will leak to all s. Hence, if the maximum delay
spread is , then for all ,
where is the symbol duration of OFDM.
LI et al.: ROBUST CHANNEL ESTIMATION FOR OFDM SYSTEMS 907

Applying the above two assumptions to (40), we get , and Note

that is the DFT matrix defined in (49). In this case the
average MSE of the robust estimator is

(57)

For any channel with and , the average

MSE should be
Fig. 3 shows the MSE of the robust estimator that matches
different Doppler frequencies and delay spreads. As shown, the
MSE is almost a constant if is fixed. In particular, let

(58)

(50) where is the baud of the OFDM system. For

OFDM systems satisfying and ,
the average MSE of the robust channel estimator is
where

(51)

In the above derivation we have used (23) and (21). From its
definition, and for any
If the channel estimator is designed such that

for (59)
(52)
for
If the channel estimator is designed to match the Doppler
then, for any channel with for and
spectrum given by (10) and the s given by (52), then
, we have

(53) for (60)

for
However, if
(61)

(54) and therefore, from (28), we have

then
(55)
since for any
From the above discussion, a robust estimator, which is
insensitive to the channel statistics, should have an ideal
band-limited time-domain correlation (62)
and a frequency-domain correlation matrix
Hence, compared with the estimator that tightly matches the
.. .. .. Doppler spectrum, the performance degradation of the robust
. . . (56) channel estimator is negligible.

B. FIR Channel Estimator

where
Here, we briefly introduce the design of finite-tap robust
estimators. Note that from the previous section, for a robust
908 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 7, JULY 1998

with length , we have

(66)
for where
if
(67)
otherwise.
From (66), it is obvious that for For

(68)

where
(a)
(69)

..
.
.. .. .. ..
. . . .
..
.
(70)
(71)
Using derivations similar to those in Section III, the FT of
the coefficient matrix of the designed FIR channel estimator
is determined by
(72)
(b)
where is a diagonal matrix with
Fig. 3. Normalized MSE of channel estimator versus Doppler frequency
(a) when SNR = 10 dB and td = 0, 5, 10, 20, and 40s and (b) when (73)
SNR = 10 dB and fd td = 0:000 08, 0:0008, and 0:008, respectively.

estimator and is the FT of The estimation error of the FIR

estimator can be found by
(63)
(74)
Let
Hence, for the robust FIR channel estimator, the in Fig. 2
.. .. is the DFT matrix and the s for are
. .
(64) In Fig. 4 the average MSE of a robust FIR channel estimator
.. ..
. . is shown as a function of its length. From the figure, for the
estimator matching a 40-Hz maximum Doppler frequency, a
50-tap FIR estimator is needed to exploit the time-domain
correlation of the channel parameters, while for one matching a
.. .. 200-Hz maximum Doppler frequency, only a five-tap channel
. .
estimator is sufficient.
(65)
.. ..
. . V. REFERENCE GENERATION AND COMPUTER SIMULATION
In this section we demonstrate the performance improve-
where is the coefficient of the designed channel ment of an OFDM system with our robust channel estimator.
estimator. Then, from (A.15), for the FIR channel estimator First, we briefly describe the simulated OFDM system.
LI et al.: ROBUST CHANNEL ESTIMATION FOR OFDM SYSTEMS
(a)
(c)
Fig. 4. Normalized MSE versus the length of FIR channel estimator under different channel conditions.
A. System Parameters
In our simulation, we use a two-path Rayleigh-fading chan-
nel model [12] with delay from 0 to 40 s and Doppler
frequency from 10 to 200 Hz. The channels corresponding
to different receivers have the same statistics. Two antennas
are used for receiver diversity.
To construct an OFDM signal, assume that the entire
channel bandwidth, 800 kHz, is divided into 128 subchannels.
The four subchannels on each end are used as guard tones and
the rest (120 tones) are used to transmit data. To make the tones
orthogonal to each other, the symbol duration is 160 s. An
additional 40- s guard interval is used to provide protection
from ISI due to channel multipath delay spread. This results
in a total block length s and a subchannel symbol
rate kBd.
To compare the performance of the OFDM system with and
without the channel estimation, PSK modulation with coherent
demodulation and differential PSK (DPSK) modulation with
differential demodulation are used, respectively. As in [4], a
(40,20) RS code, with each code symbol consisting of three
quadrature PSK/differential quadrature PSK (QPSK/DQPSK)
symbols grouped in frequency, is used in the system. Hence,
each OFDM block forms an RS codeword. The RS decoder
909
(b)
(d)
erases ten symbols, based on signal strength, and corrects five
additional random errors. Hence, the simulated system can
transmit data at 1.2 Mb/s before decoding or 600 kb/s after
decoding, over an 800-kHz channel.
B. Reference Generation
An ideal reference is assumed in the derivation of the chan-
nel estimator in Section III. In practical systems a reference can
be generated during a training block. In subsequent blocks a
reference is generated using the received signals. We consider
four possible reference generating schemes.
1) Undecoded/Decoded Dual Mode Reference: If the RS
decoder can successfully correct all errors in an OFDM
block, the reference for the block can be generated by
the decoded data; hence Otherwise,
2) Undecoded Reference: , no matter
whether the RS decoder can successfully correct all
errors in a block or not.
3) Decoded/CMA Dual Mode Reference: The constant
modulus algorithm (CMA) is one of the most effective
adaptive blind equalization algorithms [13], [14]. It
can be also used to generate a reference for the
910
(a)
(b)
Fig. 5. (a) WER and (b) NMSE of robust estimator with different references
versus SNR for the system without training blocks when the 50-tap channel
estimator matches the channel with fd = 40 Hz and td = 20 s.
OFDM channel estimator. If the RS decoder can
successfully correct all errors in a block, the reference
for the block can be generated from the decoded data;
hence Otherwise, the reference can
use the projection of on the unit circle, i.e.,
4) Error Removal Reference: If the RS decoder can suc-
cessfully correct all errors in a block, the reference
for the block can be generated by the decoded data.
Otherwise, the s are used instead of s about 3 dB better than using differential detection. However,
for , respectively.
C. Simulation Results
Figs. 58 demonstrate the performance of the channel es-
timator using different references under different channel
conditions. To get insight into the average behavior of the
channel estimator, we have averaged the performance over
10 000 OFDM blocks.
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 7, JULY 1998
(a)
(b)
Fig. 6. (a) WER and (b) NMSE of robust estimator with different references
versus SNR for the system with 1% training blocks when the 50-tap channel
estimator matches the channel with fd = 40 Hz and td = 20 s.
Fig. 5 shows the word-error rate (WER) and normalized
MSE (NMSE) for the OFDM system with only the first block
as a training or synchronizing block and for a channel with
Hz and s. From the figure, when
dB, the channel estimator using the decoded/undecoded dual
mode reference or the undecoded reference can estimate the
channel parameters with NMSE as small as 22 dB. Hence,
the WER of the system is almost the same as the performance
assuming ideal knowledge of the channel parameters, which is
due to error propagation through the references, there is a noise
threshold for the channel estimator. For the estimators with the
decoded/undecoded dual mode reference and the undecoded
reference, the noise thresholds are 7 and 8 dB, respectively.
For the estimators using the other two reference schemes, the
noise thresholds are 12 and 19 dB, respectively. Once the SNR
is larger than the noise threshold, the system performance will
be significantly better than that of the OFDM system using
DQPSK.
LI et al.: ROBUST CHANNEL ESTIMATION FOR OFDM SYSTEMS 911

(a) (a)

(b) (b)
Fig. 7. (a) WER and (b) NMSE of robust estimator with different references Fig. 8. (a) WER and (b) NMSE of robust estimator with different references
versus SNR for the system with 1% training blocks when the 50-tap channel versus SNR for the system with 10% training blocks when the five-tap channel
estimator matches the channel with fd = 200 Hz and td = 5 s. estimator matches the channel with fd = 200 Hz and td = 5 s.

To suppress the error propagation, training blocks are pe- spread, then for all channels with Hz and s,
riodically inserted in the data stream. Fig. 6 illustrates the the system performance is not worse than the channel with
WER and NMSE for the systems with 1% training blocks. Hz and s, as indicated by Fig. 9. However,
In this situation the noise threshold disappears and, hence, the for channels with Hz or s, such as
OFDM system with channel estimation has better performance Hz and s or Hz and s, the
than the one without channel estimation when the SNR ranges system performance degrades dramatically. On the other hand,
from 0 to 20 dB. as indicated in Fig. 10, if the estimator is designed to match a
For a channel with a Doppler frequency as large as 200 Hz, Doppler frequency or delay spread larger than the actual ones,
as indicated in Fig. 7, if 1% training blocks are used, then the system performance degrades only slightly compared with
the OFDM system with channel estimation has about a 1.5- estimation that exactly matches the channel Doppler frequency
dB SNR improvement compared to the one without channel and delay spread.
estimation. If the training blocks are increased from 1% to
10% of the data, then the required SNR for
is reduced by 1 dB compared with the one with 1% training VI. CONCLUSIONS
blocks, but the required SNR for is almost the We have presented the design of robust channel estimators
same as before, as shown in Fig. 8. for OFDM systems that make full use of the time- and
Figs. 9 and 10 illustrate the robustness of our channel frequency-correlations of the rapid dispersive fading wireless
estimator. As indicated in the analysis in Section IV, if a channel and are insensitive to the channel statistics. Com-
channel estimator is designed to match the channel with 40- puter simulation demonstrates that channel estimation gives
Hz maximum Doppler frequency and 20- s maximum delay about 2.5-dB improvement when the Doppler frequency is
912 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 7, JULY 1998

(a) (a)

(b) (b)
Fig. 9. (a) WER and (b) NMSE of the 50-tap channel estimator matching Fig. 10. (a) WER and (b) NMSE of the 50-tap channel estimator matching
fd = 40 Hz and td = 20 s for channels with different fd s and td s. different fd s and td s for the channel with fd = 10 Hz and td = 5 s.

40 Hz, and about 1.5-dB improvement when the Doppler

or
frequency is as large as 200 Hz. This channel estimation
used together with antenna arrays has the potential to provide
significant suppression of cochannel interference in OFDM
systems.
(A.2)
APPENDIX
DERIVATION OF MMSE CHANNEL ESTIMATOR
for and
By the orthogonality principle [15], the s are
where
determined by

(A.1) (A.3)
Direct calculation shows that (A.1) is equivalent to
With the separation property (7) of the correlation of the
channel frequency response, (A.2) can be simplified to

(A.4)
LI et al.: ROBUST CHANNEL ESTIMATION FOR OFDM SYSTEMS 913

Let then

..
.
(A.14)
..
.
or

..
.
(A.5)
.. (A.15)
.
for , where and are the th
(A.6) elements of and , respectively.
For the with , from [16, Appendix A]
Then, (A.4) can be written in vector form as

(A.16)
(A.7)

Let the eigendecomposition of be where is a stable one-sided FT

(A.8) (A.17)

where is a unitary matrix and is a diagonal matrix with

which is uniquely determined by
diagonal elements It is clear that
(A.18)
(A.9)
The dc component in can be found by
Substituting (A.8) into (A.7) gives
(A.19)

For the with , we have for

(A.10) , which can be also written as (A.16).
Therefore, from (A.13)
Consequently

(A.11)
(A.20)
The in the above expression is the pseudoinverse of ,
which is also a diagonal matrix with elements where is a diagonal matrix with the th diagonal element

if
(A.12) (A.21)
if

Denote which is zero for the with If we define

(A.13) (A.22)
914 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 46, NO. 7, JULY 1998

then ACKNOWLEDGMENT
The authors would like to thank J. Chuang and V. Tarokh for
their insightful comments, and J. Winters and L. Ariyavisitakul
(A.23) for helpful discussions.

REFERENCES
The MMSE channel estimator for OFDM systems derived
(A.23) is shown in Fig. 2. [1] J. C.-I. Chuang, The effect of time delay spread on portable radio
communications channels with digital modulation, IEEE J. Select.
Substituting (A.2) into (16), the MMSE for the estimation Areas Commun., vol. SAC-5, pp. 879889, June 1987.
of is [2] L. J. Cimini, Jr., Analysis and simulation of a digital mobile channel us-
ing orthogonal frequency division multiplexing, IEEE Trans. Commun.,
vol. COM-33, pp. 665675, July 1985.
(A.24) [3] L. J. Cimini, Jr., B. Daneshrad, and N. R. Sollenberger, Clustered
OFDM with transmitter diversity and coding, in Proc. 1996 IEEE
Global Telecommunications Conf., London, U.K., Nov. 1996, pp.
703707.
Let [4] L. J. Cimini, Jr. and N. R. Sollenberger, OFDM with diversity and
coding for advanced cellular internet services, in Proc. 1997 IEEE
Global Communications Conf., Phoenix, AZ, Nov. 1997, pp. 305309.
[5] J.-J. van de Beek, O. Edfors, M. Sandell, S. K. Wilson, and P. O.
Borjesson, On channel estimation in OFDM systems, in Proc. 45th
.. (A.25) IEEE Vehicular Technology Conf., Chicago, IL, July 1995, pp. 815819.
. [6] , OFDM channel estimation by singular value decomposition, in
Proc. 46th IEEE Vehicular Technology Conf., Atlanta, GA, Apr. 1996,
pp. 923927.
[7] I. Kalet, The multitone channel, IEEE Trans. Commun., vol. 37, pp.
Then, from (A.23) 119124, Feb. 1989.
[8] V. Mignone and A. Morello, CD3-OFDM: A novel demodulation
scheme for fixed and mobile receivers, IEEE Trans. Commun., vol.
44, pp. 11441151, Sept. 1996.
[9] S. B. Weinstein and P. M. Ebert, Data transmission by frequency-
division multiplexing using the discrete Fourier transform, IEEE Trans.
Commun., vol. COM-19, pp. 628634, Oct. 1971.
[10] R. Steele, Mobile Radio Communications. New York: IEEE Press,
1992.
[11] W. C. Jakes, Jr., Ed., Microwave Mobile Communications. New York:
(A.26) Wiley, 1974.
[12] Y. (G.) Li, J. H. Winters, and N. R. Sollenberger, Spatial-temporal
equalization for IS-136 TDMA systems with rapid dispersive fading and
where is a diagonal matrix with the th diagonal element co-channel interference, IEEE Trans. Veh. Technol., to be published.
[13] D. N. Godard, Self-recovering equalization and carrier tracking in two-
dimensional data communication systems, IEEE Trans. Commun., vol.
COM-28, pp. 18671875, Nov. 1980.
[14] J. R. Treichler and B. G. Agee, A new approach to multipath correction
of constant modulus signals, IEEE Trans. Acoust., Speech, Signal
Processing, vol. ASSP-31, pp. 349372, Apr. 1983.
[15] A. Papoulis, Probability, Random Variables, and Stochastic Processes,
3rd ed. New York: McGraw-Hill, 1991.
[16] J. Salz, Optimum mean-square decision feedback equalization, Bell
Syst. Tech. J., vol. 52, pp. 13411371, Oct. 1973.

(A.27)

The average MMSE is Ye (Geoffrey) Li (S92M95SM97) was born in

Jiangsu, China. He received the B.S.E. and M.S.E.
degrees from the Department of Wireless Engi-
neering, Nanjing Institute of Technology, Nanjing,
China, in 1983 and 1986, respectively, and the
Ph.D. degree from the Department of Electrical
Engineering, Auburn University, Auburn, AL, in
1994.
From 1986 to 1991 he was a Teaching Assistant
and then a Lecturer with Southeast University, Na-
jing, China. From 1991 to 1994 he was a Research
and Teaching Assistant with Auburn University, Auburn, AL. From 1994
to 1996 he was a Postdoctoral Research Associate with the University
of Maryland, College Park, MD. Since 1996 he has been with AT&T
LaboratoriesResearch, Red Bank, NJ. His general research interests include
(A.28) statistical signal processing and wireless mobile systems with emphasis on
signal processing in communications.
Dr. Li is currently serving as a guest editor for a special issue on signal
processing for wireless communications for the IEEE JOURNAL ON SELECTED
where denotes the trace of matrix, defined as the sum AREAS IN COMMUNICATIONS and as an Editor for Wireless Communication
of the diagonal elements of the matrix. Theory for the IEEE TRANSACTIONS ON COMMUNICATIONS.
LI et al.: ROBUST CHANNEL ESTIMATION FOR OFDM SYSTEMS
Leonard J. Cimini, Jr. (S77M72SM89) was
born in Philadelphia, PA, on April 19, 1956. He
received the B.S.E. (summa cum laude), M.S.E.,
and Ph.D. degrees in electrical engineering from the
University of Pennsylvania, Philadelphia, in 1978,
1979, and 1982, respectively.
Since 1982 he has been with AT&T Laborato-
riesResearch, Red Bank, NJ, where his current
research interests are in wireless communications
systems. He is also an Adjunct Professor at the
University of Pennsylvania.
Dr. Cimini is a member of Tau Beta Pi and Eta Kappa Nu. He has been
very active in the IEEE Communications Society and is currently serving
as Area Editor for Wireless Communications for the IEEE TRANSACTIONS
ON COMMUNICATIONS and Editor-in-Chief of the IEEE JOURNAL ON SELECTED
AREAS IN COMMUNICATIONS Wireless Communications Series. He received a
National Science Foundation Fellowship during his graduate work.
915
Nelson R. Sollenberger (S78M81SM90F96)
received the B.S.E. degree from Messiah College,
Grantham, PA, in 1979, and the M.S.E. degree from
Cornell University, Ithaca, NY, in 1981, both in
electrical engineering.
He currently heads the Wireless Systems Re-
search Department, AT&T LaboratoriesResearch,
Red Bank, NJ. His department performs research on
next-generation wireless systems concepts and tech-
nologies including high-speed transmission meth-
ods, smart antennas and adaptive signal processing,
system architectures and radio link techniques to support wireless multimedia,
and advanced voice services. From 1979 through 1986 he was a member of
the Cellular Radio Development Organization, Bell Laboratories, where he
investigated spectrally efficient analog and digital technologies for second-
generation cellular radio systems. In 1987 he joined the Radio Research
Department, Bellcore, where he was Head of the department from 1993
to 1995. At Bellcore, he investigated concepts for the Personal Access
Communications System (PACS).