Sei sulla pagina 1di 9

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 54, NO.

10, OCTOBER 2006 3957

Training Design for Multipath Channel and


Frequency-Offset Estimation in MIMO Systems
Mounir Ghogho, Member, IEEE, and Ananthram Swami, Senior Member, IEEE

Abstract—This paper addresses the problem of training de- [10], and [11] were derived for orthogonal frequency-division
sign for frequency-selective channel and carrier frequency-offset multiplexing (OFDM) systems. The work in [12] focused on
(CFO) estimation in single- and multiple-antenna systems under zero-padded single-carrier and cyclic-prefixed multicarrier
different energy-distribution constraints. The performance metric
block transmissions. Training symbols were inserted in each
used here is the Cramér–Rao bound (CRB). The paper first
addresses the CFO-free case and then generalizes the results block in order to track channel variations. By decoupling
to include CFO-corrupted scenarios. Training sequences are channel estimation from data detection and using linear min-
designed that render the CRB for the CFO independent of the imum mean-square error, optimal training designs were derived.
channel zeros. The proposed training designs also facilitate simple In [9], optimum finite-alphabet constant amplitude training
implementation of the maximum-likelihood CFO and channel sequences were proposed for single-carrier systems. The effect
estimators. Simulation results illustrate the merits of the proposed of imperfect channel estimation on the mutual information or
training designs.
capacity was investigated in [20] and [15]. Training designs
Index Terms—Communication channels, frequency estimation, maximizing the mutual information, capacity, or cutoff rate
identification, synchronization, training. were studied in [12], [13], [15], and [21]. Training designs that
are optimal in the presence of CFO have received relatively
little attention. It was investigated for the case of single antenna
I. INTRODUCTION
systems in [6] using the worst-case asymptotic Cramér–Rao
N communication systems, accurate estimation of the carrier bounds (CRBs), i.e., the large-sample CRB associated with the
I frequency-offset (CFO) and the channel impulse response
(CIR) are usually required in order to achieve good symbol de-
worst channel realization. It was found that a white Gaussian
preamble minimizes such a performance metric. Here, we
tection performance [1]. This is particularly important in sys- use the exact CRB as a metric for training design. The idea
tems employing large alphabets, multicarrier modulations, or of using the CRB for training design was also considered in
multiple transmit antennas. In order to save bandwidth, con- [22] in the context of CFO estimation in OFDM systems, and
siderable effort has been devoted to non-data-aided (or blind) in [14] for frequency-selective channel estimation. First, we
estimation methods [2]. However, in most current practical sys- consider the case where there is no CFO and focus on channel
tems, channel/CFO estimation typically relies on transmitting a estimation. We optimize the preamble design under different
known training sequence, which may be prepended, appended, energy-distribution constraints. Then, we tackle the more chal-
or embedded in the data packet. Although this approach wastes lenging problem of training design for both CFO and channel
bandwidth, reliable and computationally simple estimates can estimation. A power-efficient preamble design that reduces the
be obtained. Therefore, this approach has received renewed at- complexity of CFO and channel estimation will be presented.
tention in the last few years [3]–[16]. Semiblind methods have The proposed preamble design can be accommodated easily in
also been reported in [17]–[19]. both serial and block (either cyclic-prefixed or zero-padded)
If training is to be used, it is important to optimize the transmissions. Our findings are novel in both single and mul-
resources allocated for training: the tradeoff is clearly between tiple antenna scenarios.
improved channel estimation and improved rate; the impact The structure of the paper is as follows. The MIMO signal
on the bit error rate (BER) is more subtle since BER depends model will be presented in Section II. In Section III, optimum
both upon the quality of the channel estimate as well as the training design for channel estimation in the CFO-free case
nominal signal-to-noise ratio (SNR). In the absence of CFO, will be investigated. Section IV proposes training designs
optimal training for multiple-input multiple-output (MIMO) for both channel and CFO estimation. Maximum likelihood
frequency-selective channel estimation has been recently ad- estimation will be addressed in Section V. Simulations results
dressed by many authors, e.g., [8]–[12]. The designs in [8], will be presented in Section VI and conclusions will be drawn
in Section VII.
Notation: Superscripts , and denote Hermitian, trans-
pose, and pseudoinverse operators. The trace, statistical expec-
Manuscript received June 17, 2005; revised October 24, 2005. The associate
editor coordinating the review of this manuscript and approving it for publica- tation, and the Kronecker product are denoted by , ,
tion was Dr. Athina Petropulu. and , respectively. The norm of a vector is denoted by
M. Ghogho is with the School of Electronic and Electrical Engineering, Uni- . The identity matrix is denoted by ; the sub-
versity of Leeds, Leeds LS29JT, U. K. (e-mail: m.ghogho@ieee.org).
A. Swami is with the Army Research Laboratory, Adelphi, MD 20783-1197
script will be dropped when there is no ambiguity about the size
USA (e-mail: a.swami@ieee.org). of the matrix. Finally, is the diag-
Digital Object Identifier 10.1109/TSP.2006.879318 onal matrix whose th diagonal entry is .
1053-587X/$20.00 © 2006 IEEE
3958 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 54, NO. 10, OCTOBER 2006

II. SIGNAL MODEL The channel vector is identifiable if and only if has full
column rank:
Consider a MIMO system with transmit and receive
antennas. Let the baud-sampled discrete-time channel impulse
response (CIR) between the th transmit and the th receive rank
antennas be denoted by ,
where is an upper bound on the length of the longest CIR. Thus a necessary but not sufficient condition for channel iden-
We assume that all the transmit (respectively, receive) antennas tifiability is .
are driven by the same local oscillator, a reasonable assump- The CRB for is a lower bound on the covariance of any
tion when the receive (transmit) antennas are collocated. This unbiased estimate of , i.e., [23]
implies that all pairs of transmit-receive antennas will experi-
ence the same CFO. In order to estimate the CFO and the CIRs, CRB (5)
the transmit antennas send possibly different -point
preambles,1 , , simulta- Since is an AWGN vector, the CRB for is readily obtained
neously. These preambles or training signals are known to the as
receiver. Ignoring the first samples (which could be con-
taminated by interblock interference), the received preamble at CRB (6)
the th receive antenna is modeled as
Thus, the CRBs for the ’s, the channels to receiver , are iden-
(1) tical and mutually decoupled. Further, for any given receive an-
tenna, the CRBs for the CIRs associated with different transmit
antennas are in general coupled; they can be made decoupled
where , if the training sequences are designed such that if
, is the normalized angular CFO, . Optimal design of the training sequences is investigated
in Section IV-A.

B. CFO-Corrupted Scenario
In the more general case where the received signals are cor-
We assume the ’s to be mutually independent circularly sym- rupted by CFO, the CRB for and are obtained as (see
metric complex white Gaussian noise vectors, with covariance Appendix I for a proof)
. We also assume that no a priori information about the sta-
tistics of the channels is available
CRB (7)
Let , ,
, and . The signal model in CRB
(1) can then be rewritten as
(8)
(2)
(3) where and

where the additive white Gaussian noise (AWGN) vector is (9)


defined similar to . (10)
(11)
III. CRAMÉR–RAO BOUNDS
The motivation behind the use of the CRB to design training where is assumed to be tall. The CRB for can be also ex-
sequences stems from the fact that the CRB is a lower bound pressed as
on the mean square error of any unbiased estimator, and the fact
that the maximum-likelihood (ML) estimator (asymptotically) CRB
achieves the CRB. In this section, we first address the CFO-free CRB
scenario.
which clearly shows the impact of CFO estimation error bounds
A. CFO-Free Scenario on channel estimation error bounds. Notice that the CRB for
is equal to that in the CFO-free case plus an extra term that takes
In this case, , and . The received signal
into account the effect of CFO estimation error. In the case of
becomes
single-antenna systems, the preceding CRB expressions reduce
to the ones given in [6].
(4)
Unlike the CFO-free case, the CRBs for different ’s are
1We consider preambles rather than midambles or postfixes for simplicity. now coupled. Further, the CRBs for the CFO and the CIRs are
GHOGHO AND SWAMI: TRAINING DESIGN FOR MULTIPATH CHANNEL AND FREQUENCY-OFFSET ESTIMATION IN MIMO SYSTEMS 3959

now channel dependent. In Section IV-B, we design training for all . The above condition implies that
sequences that make the CRB of channel-zeros-independent, , , which is a “cyclic-prefix” type
i.e., for a fixed , CRB is independent of the channel condition. Substituting (15) into (6), we note that the optimal
realizations (or channel zeros). training sequence for the design problem in (14) is an orthog-
onal sequence that maximizes . Equation (15) indicates that
the training sequences associated with the different transmit an-
IV. TRAINING SEQUENCE DESIGN tennas should be orthogonal, i.e.,
Since no a priori information about the channels is available
at the transmitter, the total transmit energy allocated to training (16)
is split equally between the transmit antennas, i.e.,
Under condition (C0), the maximum value for is , and
(12) it is obtained when

where , and for (17)


. If the peak-to-average power ratio (PAPR)
is an issue, then the following condition, which is more strin- Thus, under condition (C0), the first and last samples
gent than Condition (C0), will be imposed: of each -point training sequence should be set to
zero. This is a training sequence that is zero-padded and also
has a cyclic prefix (of zeros). On the other hand, under condition
(C1), can take on only one possible value which is
(13) . Therefore, under condition (C1), all sequences
satisfying the orthogonality condition in (15) are optimal. Under
The constant modulus condition (C1) ensures that the PAPRs condition (C0), the condition in (17) is also required.
of all training sequences are unity. This is a desirable property The minimum CRB for is then given by
when the nonlinear distortion due to power amplifiers is of con-
cern. PAPR is a critical issue in OFDM and other multicarrier
CRB CFO (18)
systems; as such condition (C1) is very important in such sys-
tems. As we will see next, the optimal training sequence de-
signs under condition (C0) will differ from those under condi- where under Condition (C0) and
tion (C1). More specifically, since (C0) is less restrictive than under Condition (C1). The CRBs for all the channel coefficients,
(C1), the former will lead to better estimation performance at , ; ; , are now
the expense of a larger PAPR. Before addressing the general decoupled. Furthermore, as predicted, the CRBs under (C0) are
case of training design for CFO and channel estimation, we first lower than those under (C1) since . Indeed, (C0) enables a
study the CFO-free case and derive optimal training sequences more power efficient training scheme than (C1). However, (C1)
for channel estimation. guarantees unit-PAPR training sequences.
Note that under the optimal training design conditions, the
A. CFO-Free Case SNRs at the receive antennas are maximized and are given by

The design problem here is as follows:


SNR (19)
CRB subject to Cm (14)
This implies that the SNR at each receive antenna is independent
of the corresponding channel zeros and the training sequences.
where and or depending on If the optimality conditions are not satisfied, then this desirable
whether Condition (C0) or Condition (C1) is enforced. We could property will be lost. Notice that for the same transmit power
solve the above optimization problem using the Lagrange multi- and channel norms, the SNRs under Condition (C0) are larger
plier approach. However, an easier way of tackling the problem than those under Condition (C1). For Rayleigh fading channels,
is to first use the following result: the ’s are themselves realization dependent and are de-
scribed by chi-squared random variables.
1) Further Details on the Training Design Under (C0):
Under the condition in (17), the ’s are column-circular
with equality only when (e.g., [7] and references therein) matrices since the ’s exhibit a cyclic-prefix structure. Hence,
they can be diagonalized using the discrete Fourier transform
(15) (DFT) matrix as
, where is the DFT of ,
where i.e.,
3960 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 54, NO. 10, OCTOBER 2006

Fig. 1. Training scheme in (23); M = 2, L = 4, N = 8 (11 training symbols in total); a = P =4.

, and is the leading matrix of . constellations were also investigated. In this subsection, we as-
Therefore, (15) and (17) are equivalent to sume ( being an integer) and impose a cyclic
prefix structure on the training sequences, i.e.,
, . In this case, the orthogonality condition
in (15) becomes

(20)

and

(24)

Equation (24) is similar to (20). Note that the cyclic-prefix struc-


(21) ture does not affect the optimality of the training design. It pro-
vides a subclass of the general class of optimal designs satis-
The second condition forces the last samples of the training fying the condition in (15) and (C1).
sequences to be zero, as it is required under condition (C0); see An example of sequences satisfying both (C1) and (24) is the
(17). chirp
Now, we give a design example when where
is an integer. In Appendix II, we show that sequences
with the following DFTs satisfy the design conditions in (20)
and (21) (25)

where if is odd and if is even, ,


and the ’s are arbitrary phases. Similar designs were proposed
in [16] for cyclic-prefixed single carrier systems.
(22)
B. Training Design in the Presence of CFO
where , , are arbitrary phases in . In the It is obvious from (7) and (8) that the training sequences
frequency-domain, each training sequence is an L-tooth comb, that minimize CRB and CRB are channel depen-
with a spacing of bins; the combs for adjacent transmitters dent. To circumvent this problem, the asymptotic CRB and
are offset by frequency bins. Note that the above design is the worst-channel scenario were used in [6]. Here, we take a
similar to that proposed for MIMO OFDM systems (see, e.g., different direction. We investigate training designs that make
[10]). It is, however, worth pointing out that here we achieve this CRB channel independent, i.e., independent of the channel
design without assuming any modulation or time guard (e.g., zeros. Since we cannot make both CRB and CRB
cyclic prefix) for the transmitted preamble. In the time domain, channel independent, we derive a channel-independent upper
the above sequences are given by bound on CRB . Although the channel-independent
CRBs are not the minimum CRBs, which are channel de-
pendent, they have the nice property of being constant for all
channel realizations provided the norm of is constant. In the
case of random channels, the above property also guarantees
(23) maximum space and multipath diversity CFO estimation since
the CRB for will be inversely proportional to .
for . Each training sequence consists of an impulse In order for CRB in (7) to be channel independent, the
train, spaced apart, amplitude modulating a different tone. training sequences have to satisfy the following condition:
If , then the above sequences are the only sequences
satisfying the optimal design criteria under Condition (C0) (see (26)
Appendix II). The training design in (23) has both leading and
trailing zeros. Hence, it can be accommodated easily in both where is a positive real scalar. Such sequences will be re-
zero-padded and cyclic-prefixed block transmission systems. A ferred to as channel-independent performance training (CIPT)
block diagram of this training design is given in Fig. 1, for the sequences. Finding all CIPT sequences seems to be a difficult
case where , , and . task. Here, we give examples of such sequences.
2) Further Details on the Training Design Under (C1): Ex- Consider sequences of size where with
amples of orthogonal training designs satisfying condition (C1) being an integer, which have a cyclic prefix structure,
were given in [9], where suboptimal training designs using PSK i.e., for , and such that
GHOGHO AND SWAMI: TRAINING DESIGN FOR MULTIPATH CHANNEL AND FREQUENCY-OFFSET ESTIMATION IN MIMO SYSTEMS 3961

their -point DFTs, i.e., DFT of , satisfy Proposition 1: If and the training sequences are
(c.f., (22)) designed as in (27), then the CRB for is channel-independent
and is given by

CRB (32)

(27) where . For the sequences in (23)


and (25), which are special cases of those in (27), and
where the ’s are arbitrary phase values and respectively. Further, the trace of CRB
. For such sequences, we obtain is upper bounded by the channel-independent quantity in (31).
; hence We remark that when , the above channel-indepen-
dent CRB for is identical to the asymptotic CRB devel-
(28) oped in [6]. Further, in the case of single antenna systems, pe-
riodic training was used for CFO estimation in [3] and a simple
where is a matrix of ones. After some algebra, we CRB expression was developed. However, the CRB in [3] was
can show that the above sequences satisfy the condition in (26) channel dependent because the SNR was channel dependent.
with given by Indeed, the periodicity in of the training sequence is a nec-
essary but not sufficient condition for channel-independent per-
formance of CFO estimators.

V. MAXIMUM-LIKELIHOOD ESTIMATION
Note that the above class of sequences encompasses the se-
quences proposed in (23) and (25). Since is an AWGN vector, the ML estimate of and are
As to the CRB for , we first use (8) to obtain found to be

CRB (33)

(29)
(34)

where is given in (11). For the CIPT design in (27), we have where is defined in (9). When and the training
that and (using (28)) sequences satisfy the CIPT design condition in (27), we obtain

CRB

(30) where is the matrix of ones. Therefore, the above


ML estimates are greatly simplified and are given by

where ; so . A channel-inde-
pendent upper bound for the above quantity is given by (35)

CRB (36)
(31)
where indicates the real part of and is the correlation
where is the maximum eigenvalue of
. Note that the upper bound is equal to CRB
in the CFO-free case plus an extra term which depends on how
the ’s in (27) are chosen. Maximum energy efficiency (re-
spectively, unit PAPR) imposes constraints on the ’s as il- We observe that in the case of single antenna systems, the ML
lustrated in (22) (respectively, (25)). Furthermore, note that the estimate of is identical to the repetitive-slot-based nonlinear
above upper bound is strictly channel independent in the sense least squares estimate in [3], [5], [24]. This holds true only
that it does not even depend on the norm of . Numerical ex- for sequences satisfying (27), i.e., if the number of repetitive
amples in the simulation section show that this upper bound is slots in the training sequence after removing the cyclic prefix
very tight, i.e., it is very close to the exact CRB. We could also is not equal to , then the repetitive-slot-based nonlinear
minimize the upper bound with respect to the possible CIPT se- least-squares estimates are not ML. Note that the acquisition
quences using a numerical optimization routine. We have found range of the above ML estimator of is ,
that the special CIPT sequences in (23) are solutions to this op- which decreases with and . Thus, for large or/and ,
timization problem. the proposed training design is viable for fine CFO estimation
We summarize these results in the following proposition. only, i.e., for small values of .
3962 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 54, NO. 10, OCTOBER 2006

If , the ML estimator in (35) is given in closed-


form as

(37)

In the case of single-antenna systems, this estimator reduces to


the one proposed in [4]. If , no closed-form solution
is available for the optimization problem in (35). In this case,
the ML estimator can be initialized or even replaced by simpler
estimators along the same lines as those developed for single-
antenna systems in [3], [24].
The mean-square errors (MSEs) of the ML estimates asymp-
totically (i.e., large ) achieve the CRB. In the CFO-free
scenario, the ML estimator of is obtained from (34) after
replacing by zero. In this case, the MSE of the ML estimate
achieves the CRB for any number of samples .
Finally, if we use the CIPT training design in (23) with
, , which is a special case of the design in (27),
Fig. 2. CRBs on CFO estimation for different channel realizations.
further simplifications of the ML channel estimates can be ob-
tained as

(38)
Again, in the CFO-free case, in the above channel estimate
should be set to zero.

VI. SIMULATION RESULTS


We focus on the interesting case of channel estimation in the
presence of CFO. In this case, since the estimation problem
is nonlinear, the CRB is achieved by the ML estimates only
asymptotically. The following simulations results shed some
light on the finite-sample analysis of these estimates.
Consider a system with one receive and two transmit an-
tennas. The channels are assumed independent and have
taps each. The ’s, are assumed uncorrelated zero-mean
Gaussian random variable with exponential power delay pro-
file , where is a scaling factor Fig. 3. MSEs and (averaged) CRBs for CFO estimation versus SNR.
that ensures that . The length of the
training sequence is set to 19, i.e., . In each Monte
Carlo simulation, a new CFO is randomly drawn from the in- since for the above channel model, the received signal power
terval and a new channel realization is gener- is on average equal to the transmitted power. These MSEs are
ated. Two types of training sequences are used to estimate the obtained using 1000 Monte Carlo runs. For the CIPT design, the
CFO and CIR: 1) independent pseudonoise (PN) sequences, as closed-form CFO estimator in (37) was used since ,
recommended in [6] in the case of single-input systems, and and the channel coefficients were estimated using the simple
2) our CIPT sequence design in (23). First, in order to illustrate expression in (38). For the PN design, a gradient-type numerical
the advantage in using CIPT over PN sequences, we display in optimization was used to estimate .
Fig. 2 the CRBs for for different realizations of the normal- Figs. 3 and 4 suggest that the proposed CIPT design not
ized channel, i.e., with kept constant. It can be seen that only leads to simple estimation algorithms but it also provides
for the CIPT design CRB is the same for all realizations of a better averaged (over the channel realizations) CFO and
the normalized channel, but it can significantly vary in the case channel estimation performance than the PN training design.
of PN-based training design. In what follows, we do not nor- Furthermore, Fig. 4 shows that the channel-independent upper
malize the channels. Figs. 3 and 4 display the MSE of the ML bound on CRB given in (31) is quite tight.
CFO and channel estimates versus the average SNR, which is Figs. 5 and 6 display performance versus for a fixed SNR.
defined here as It is seen that the CIPT design outperforms the PN-based design
for all values of . The gap in performance decreases when
SNR increases.
GHOGHO AND SWAMI: TRAINING DESIGN FOR MULTIPATH CHANNEL AND FREQUENCY-OFFSET ESTIMATION IN MIMO SYSTEMS 3963

Fig. 4. MSEs and (averaged) CRBs for channel estimation versus SNR.
Fig. 7. MSEs and (averaged) CRBs for CFO estimation versus M.

Fig. 5. MSEs and (averaged) CRBs for CFO estimation versus N. M


Fig. 8. Normalized MSEs and (averaged) CRBs for channel estimation versus
.

Finally, we investigate performance versus , the number


of transmit antennas. In order to better highlight the effect of
on MSE and CRB, we increase the number of samples
when increases; more specifically, we set . This
also allows for closed-form CFO estimation. The results when
SNR 10 dB are shown in Figs. 7 and 8. Since increases
with , the CRB for decreases with as illustrated in
Fig. 7. Indeed, for , (32) becomes inversely propor-
tional to , i.e.,

CRB

In Fig. 8, we plot in order to assess the average


estimation performance for each channel. It is seen that the nor-
malized CRBs do not significantly vary with . This suggests
that as long as is proportional to , the CRBs for the indi-

N.
vidual channels are not affected by . Fig. 8 also shows that
Fig. 6. MSEs and (averaged) CRBs for channel estimation versus the gain of the CIPT design over the PN design is uniform over
3964 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 54, NO. 10, OCTOBER 2006

. Finally, the gap between the upper-bound on CRB and The FIM is readily obtained as the equation shown at the top of
the average CRB for seems to increase with . the page, where .
The CRB for is given by the inverse of . Using the same
VII. CONCLUSION mathematical manipulations as in Appendix I of [6] along with
Optimal training designs were derived for MIMO frequency- (39), we obtain (7) and (8).
selective channel estimation under two energy constraints. A
subclass of these designs were shown to render the Cramér–Rao APPENDIX II
bound for the CFO channel independent. Using these designs, a
Here, we show that sequences whose DFTs are given in (22),
channel-independent upper bound on the aggregate CRB of the
where with being an integer, satisfy the design
CIRs was also derived. Simulations results show that the pro-
conditions in (20) and (21). First, it is obvious that the sequences
posed training designs outperform PN-sequence-based training
in (27) satisfy the first condition. Next, we show that within this
in terms of both CFO and channel MSE/CRB. Further, when the
class of sequences, the sequences in (22) are the only sequences
proposed training designs are used, a significant saving in com-
satisfying the condition in (21). Indeed, under (27), condition
putational complexity of the ML estimators can be achieved.
(21) is equivalent to
An interesting direction for future work is the extension of the
proposed designs to the case where different MIMO multipath
branches experience different CFOs or Doppler shifts, which
was investigated under the assumption of flat fading in [25].

APPENDIX I
The CRB derivations given here are a straightforward exten- Let and
sion of those in [6] to the case of MIMO systems. Thus, only an . The above equation becomes
outline of the derivations is provided.
For any unbiased estimates, and of and , we have
that
CRB
CRB The only solution to the above equation is
, , which concludes the proof.
The unknown parameters in the signal model (3) are , and If , then the sequences in (27) are the only sequences
. Since the additive noise is Gaussian and is modeled as satisfying condition (20). Therefore, the sequences in (22) are
a deterministic parameter vector, the CRBs for and the only sequences satisfying both (20) and (21).
are decoupled. To derive CRB , we first derive the CRB for
, where and denote the REFERENCES
real and imaginary parts of . Then, using [1] J. Proakis, Digital Communications, 3rd ed. New York: McGraw-
Hill, 1995.
[2] G. Vazquez and J. Riba, “Non-data aided digital synchronization,” in
Signal Processing Advances in Wireless and Mobile Communications:
we get Trends in Channel Estimation and Equalization, G. Giannakis, Y. Hua,
P. Stoica, and L. Tong, Eds. Upper Saddle River, NJ: Prentice-Hall,
2001, vol. 2, ch. 9.
CRB CRB [3] M. Morelli and U. Mengali, “Carrier-frequency estimation for trans-
(39) mission over selective channels,” IEEE Trans. Commun., vol. 48, pp.
where CRB is the CRB for . 1580–1589, Sep. 2000.
[4] P. H. Moose, “A technique for orthogonal frequency division multi-
Since the additive noise is white and circular, the received plexing frequency offset correction,” IEEE Trans. Commun., vol. 42,
signal is complex circular with mean pp. 2908–2914, Oct. 1994.
and covariance matrix . In this case, the Fisher information [5] J. Li, G. Liu, and G. B. Giannakis, “Carrier frequency offset estimation
for OFDM-based WLAN’s,” IEEE Signal Process. Lett., vol. 8, pp.
matrix (FIM) for is given by [23] 80–82, Mar. 2001.
[6] P. Stoica and O. Besson, “Training design for frequency offset and fre-
quency-selective channel estimation,” IEEE Trans. Commun., vol. 51,
no. 11, pp. 1910–1917, Nov. 2003.
GHOGHO AND SWAMI: TRAINING DESIGN FOR MULTIPATH CHANNEL AND FREQUENCY-OFFSET ESTIMATION IN MIMO SYSTEMS 3965

[7] S. Ohno and G. B. Giannakis, “Optimal training and redundant pre- [23] S. Kay, Fundamentals of Statistical Signal Processing: Estimation
coding for block transmissions with application to wirelesss OFDM,” Theory. Englewood Cliffs, NJ: Prentice-Hall, 1993.
IEEE Trans. Wireless Commun., vol. 50, no. 12, pp. 2113–2123, Dec. [24] M. Ghogho and A. Swami, “Unified framework for a class of fre-
2002. quency-offset estimation techniques for OFDM,” presented at the Int.
[8] Y. Li, “Simplified channel estimation for OFDM systems with multiple Conf. Acoustics, Speech, Signal Processing (ICASSP), Montreal, QC,
transmit antennas,” IEEE Trans. Wireless Commun., vol. 1, no. 1, pp. Canada, 2004.
67–75, Jan. 2002. [25] O. Besson and P. Stoica, “On parameter estimation of MIMO flat-
[9] C. Fragouli, N. Al-Dhahir, and W. Turin, “Reduced-complexity fading channels with frequency offsets,” IEEE Trans. Signal Process.,
training schemes for multiple-antenna broadband transmissions,” in vol. 51, no. 3, pp. 602–613, Mar. 2003.
Proc. IEEE Wireless Communications Networking Conf. (WCNC),
2002, pp. 78–83.
[10] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for Mounir Ghogho, (M’98) received the Ph.D. degree
MIMO OFDM systems in mobile wireless channels,” IEEE Trans. from the National Polytechnic Institute of Toulouse
Signal Process., vol. 51, no. 6, pp. 1615–1624, Jun. 2003. in 1997.
[11] H. Minn and N. Al-Dhahir, “Optimal training signals for MIMO From 1997 to 2001, he was a Research Fellow at
OFDM channel estimation,” in Proc. IEEE GLOBECOM 2004, Nov. the University of Strathclyde. In 2002, he joined the
2004, pp. 219–224. University of Leeds, Leeds, U.K., where he is cur-
[12] X. Ma, L. Yang, and G. B. Giannakis, “Optimal training for MIMO rently an Associate Professor (Reader). His research
frequency-selective fading channels,” IEEE Trans. Wireless Commun., interests are in signal processing for communications
vol. 4, no. 2, pp. 453–466, Mar. 2005. and networking.
[13] S. Adireddy, L. Tong, and H. Viswanathan, “Optimal placement of Dr. Ghogho served as an Associate Editor of the
training for unknown channels,” IEEE Trans. Inf. Theory, vol. 48, no. IEEE SIGNAL PROCESSING LETTERS from 2001 to
8, pp. 2338–2353, Aug. 2002. 2004. He is currently an Associate Editor of the IEEE TRANSACTIONS ON
[14] M. Dong and L. Tong, “Optimal design and placement of pilot symbols SIGNAL PROCESSING and a member of the IEEE SPCOM and SPTM Technical
for channel estimation,” IEEE Trans. Signal Process., vol. 50, no. 12, Committees.
pp. 3055–3069, Dec. 2002.
[15] B. Hassibi and B. M. Hochwald, “How much training is needed in mul-
tiple-antenna wireless links,” IEEE Trans. Inf. Theory, vol. 49, no. 4,
pp. 951–963, Apr. 2003. Ananthram Swami (SM’96) received the B.S.
[16] J. Coon, M. Beach, and J. McGeehan, “Optimal training sequences degree from the Indian Institute of Technology,
for channel estimation in cyclic-prefix based single-carrier systems Bombay, India; the M.S. degree from Rice Univer-
with transmit diversity,” IEEE Signal Process. Lett., vol. 11, no. 9, pp. sity, Houston, TX; and the Ph.D. degree from the
729–732, Sep. 2004. University of Southern California, Los Angeles, all
[17] M. Ghogho and A. Swami, “Semi-blind frequency offset synchroniza- in electrical engineering.
tion for OFDM,” presented at the IEEE Int. Conf. Acoustics, Speech, He has held positions with Unocal Corporation,
Signal Processing (ICASSP), Orlando, FL, May 2002. the University of Southern California, CS-3, and
[18] C. Budianu and L. Tong, “Channel estimation for space-time orthog- Malgudi Systems. He is currently a Senior Research
onal block codes,” IEEE Trans. Signal Process., vol. 50, no. 10, pp. Scientist with the U.S. Army Research Laboratory,
2515–2528, Oct. 2002. Adelphi, MD, where his work is in the broad area of
[19] S. Zhou and G. B. Giannakis, “Single-carrier space–time block coded signal processing for communications. He was a Statistical Consultant to the
transmissions over frequency-selective channels,” IEEE Trans. Inf. California Lottery, developed a Matlab-based toolbox for non-Gaussian signal
Theory, vol. 49, no. 1, pp. 164–179, Jan. 2003. processing, and has held visiting faculty positions at INP, Toulouse, France.
[20] M. Medard, “The effect upon channel capacity in wireless communica- Dr. Swami is a member and Vice-Chair of the IEEE Signal Processing
tions of perfect and imperfect knowledge of the channel,” IEEE Trans. Society’s Technical Committee on Signal Processing for Communications, an
Inf. Theory, vol. 46, no. 3, pp. 933–946, May 2000. Associate Editor of the IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS,
[21] S. Misra, A. Swami, and L. Tong, “Optimal training over the Gauss- and of the IEEE TRANSACTIONS ON SIGNAL PROCESSING, and a member of
Markov fading channel: a cutoff rate analysis,” presented at the Proc. the Editorial Board of the IEEE Signal Processing Magazine. He has served
Int. Conf. Acoustics, Speech, Signal Processing (ICASSP), Montreal, as an associate editor for IEEE SIGNAL PROCESSING LETTERS and IEEE
QC, May 2004. TRANSACTIONS ON CIRCUITS AND SYSTEMS II. He was Co-Organizer and
[22] M. Ghogho, A. Swami, and G. B. Giannakis, “Optimized null-subcar- Co-Chair of the 1993 IEEE-SPS HOS Workshop, the 1996 IEEE-SPS SSAP
rier selection for CFO estimation in OFDM over frequency-selective Workshop, and the 1999 ASA-IMA Workshop on Heavy-Tailed Phenomena.
fading channels,” presented at the IEEE GLOBECOM, San Antonio, He was Co-Guest Editor of the 2004 IEEE Signal Processing Magazine Special
TX, Nov. 2001. Issue on Signal Processing for Networking.

Potrebbero piacerti anche