602 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO.

3, MARCH 2003

On Parameter Estimation of MIMO Flat-Fading

Channels With Frequency Offsets
Olivier Besson, Member, IEEE, and Petre Stoica, Fellow, IEEE

Abstract—We address the frequency offsets and channel gains advocated. The first consists of using an encoding strategy
estimation problem for a multi-input multi-output (MIMO) that is immune to frequency offsets. This is the purpose of the
flat-fading channel using a training sequence. The general double-differential space-time coding scheme recently pro-
case where the frequency offsets are possibly different for each
transmit antenna is considered. The Cramér–Rao bound (CRB) posed in [15], which allows for information recovery regardless
for the problem at hand is derived. Additionally, we present a of the frequency offsets. The alternative solution is similar to
simple, closed-form expression for the large-sample CRB and many single antenna communication systems to estimate the
show that it depends in a simple way on the channel parameters. frequency offsets with a view to compensate for them. This is
Next, the parameters estimation issue is investigated. First, the the route we take here.
maximum likelihood estimator (MLE), which entails solving an
-dimensional maximization problem where is the number of In this paper, we consider a multiple-input multiple-output
transmit antennas, is derived. Then, we show that the likelihood (MIMO) flat-fading channel affected by Doppler shifts, and we
function can be written as the product of one-dimensional address the problem of estimating the channel gains and the fre-
(1-D) functions if a suitable choice of the training sequence is quency offsets. Despite the fact that it may reduce the informa-
made. Based on this fact, we suggest two computationally simpler tion rate, the use of training sequences to estimate frequency off-
methods. Numerical examples that illustrate the performance of
the estimators and compare it with the CRB are provided. sets and channel gains remains an effective solution, and thus,
we consider this framework. In contrast to previously published
Index Terms—Channel estimation, Cramér–Rao bounds, flat-
fading channels, frequency offsets, maximum likelihood, multiple papers (see e.g., [15]), we do not assume that the frequency off-
antennas. sets are the same for each Tx antenna. The latter assumption is
valid only when the multipath components impinge on the re-
ceiver from a common angle of arrival. Herein, we relax this
I. INTRODUCTION AND DATA MODEL assumption, and therefore, the model considered is more gen-
eral.
S PACE-TIME diversity, which is achieved through the use
of multiple transmit and receive antennas, has recently
been advocated as an efficient means to mitigate fading effects
Let us therefore consider a flat-fading channel with Tx an-
tennas and Rx antennas that is affected by Doppler shifts and
in wireless channels, see, e.g., [1]–[4] and references therein. possibly carrier frequency mismatches as well. Then, the output
In addition to the inherent diversity provided by multiple of the -th Rx antenna can be written as (see, e.g., [15] for a de-
transmit antennas, efficient space-time coding strategies (such tailed derivation of the equation below)
as space-time block coding) have been proposed [5]–[9].
Space-time block codes enable us to achieve full diversity
(1)
while maintaining a simple linear decoding structure at the
receiver. However, their effectiveness relies on accurate channel
state information [10], [11], which can be obtained for instance where we have the following.
through the use of training sequences [12]. To remedy this • The channel gains { } are assumed to be unknown con-
problem, differential encoding schemes have been proposed; stants that do not change over the interval [ ].
see, e.g., [13] and [14]. However, the performance of these • { } are the (known) sequence symbols transmitted via
schemes may seriously degrade in the presence of frequency the Tx antennas. Unless otherwise stated, we do not
offsets that are unavoidably present due to the relative motion make any specific assumption about the way these se-
between the transmitter and the receiver (as well as possible quence symbols are generated. Some specific assumptions
carrier frequency mismatches). Therefore, it is of primary im- will be introduced when needed, e.g., to derive the asymp-
portance to take into account these offsets prior to performing totic CRB.
detection. Toward this end, two alternative strategies can be • { } are the frequency offsets (that may be different for
each Tx and Rx antenna). It should be pointed out that
Manuscript received January 17, 2002; revised October 10, 2002. The work frequency offsets are to be estimated. This is to be con-
of P. Stoica was supported in part by the Swedish Foundation for Strategic Re- trasted with previous works where does not depend
search. The associate editor coordinating the review of this paper and approving
it for publication was Dr. Alex C. Cot. on .
O. Besson is with the Department of Avionics and Systems, ENSICA, • { } is a sequence of zero-mean, independent and iden-
Toulouse, France (e-mail: [email protected]). tically distributed Gaussian random variables such that
P. Stoica is with the Department of Systems and Control, Uppsala University,
Uppsala, Sweden (e-mail: [email protected]).
Digital Object Identifier 10.1109/TSP.2002.808102 (2)
1053-587X/03$17.00 © 2003 IEEE
BESSON AND STOICA: PARAMETER ESTIMATION OF MIMO FLAT-FADING CHANNELS 603

where is the Kronecker delta, i.e., if The CRB for the estimation of is obtained as the inverse of
and if . and can be written as
The problem is to estimate the { } and { }. It can be Re Im
viewed as a generalization of the problem of frequency esti-
Im Re
mation in the presence of time-varying amplitude to multiple
signals and amplitudes that are partly known [ ] and partly
unknown [ ]. Below, we derive the CRB for the problem at Im
hand. Additionally, the problem of estimating the channel gains Re Re
and frequency offsets is addressed.
Im Re (6)
II. CRAMÉR–RAO BOUND
with .
This section is devoted to the derivation of the CRB. First, Proof: See Appendix A
we consider the exact (i.e., finite-sample) CRB and then derive The following remarks are in order.
the asymptotic (i.e., large-sample) CRB. While the former is a
• The first fact to be noted is that the FIM and, hence, the
rather complicated function of the channel gains and the fre-
CRB are block-diagonal. Hence, there is a decoupling be-
quency offsets, the latter exhibits a simple dependence on .
tween the estimation errors corresponding to the param-
Only the main results will be given while their proofs are de-
eters of two different Rx antennas. This implies that the
ferred to the appendices.
estimation of the parameters of interest can be carried out
A. Exact CRB independently for each Rx antenna.
• The CRB for the estimation of can be readily derived
Let us define the vector of the parameters of interest as
from (6) and is given by (see Appendix A)

.. (3) Re
.
(7)
• The CRB for the estimation of is given by
with
Re (8)
Re Im
• The CRB for the estimation of and depends on all
through the matrices .
• The CRB for the estimation of and also depends on
Proposition 1: The Fisher Information Matrix (FIM) for the all ; more precisely, it depends on the differences
estimation of is block-diagonal, i.e., ; see (37). When the frequency offsets at the th
Rx antenna happen to be all equal (see, e.g., [15]), i.e.,
when , the CRB for estimation of the th
.. frequency offset no longer depends on . Indeed, in the
. (4) latter case, one has
..
.
diag

where each is the FIM corresponding to the estimation of with , and the matrices and only
and is given by depend on . Furthermore, if it is known a priori that
, it can be shown (see Remark 5 in
Re Im Im Appendix A) that
Im Re Re (5)
Im Re Re

where , and
with (9)

with ,
.. .. .. .. , and .
. . . .
.. .. .. ..
. . . . B. Asymptotic CRB
The exact CRB, although obtained in closed form, exhibits
diag a rather complicated dependence with respect to (w.r.t.) the
diag channel gains and the frequency offsets. Thus, it does not
604 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 3, MARCH 2003

provide immediate insights into the influence of the channel on In this case, which is most commonly encountered, the asymp-
the estimation accuracy. To remedy this problem, we derive the totic CRB no longer depends on the frequency offsets but only
asymptotic (large-sample) CRB. The latter will be a much sim- on the training sequences powers. Furthermore, the errors asso-
pler function of and . Furthermore, as will be illustrated in ciated with the estimation of and are not correlated for
Section V, the asymptotic CRB is close to the exact CRB even . A similar result holds for and . However, there is
for small . Therefore, it is an interesting tool to evaluate the a coupling between the estimation errors corresponding to
influence of the channel parameters and the training sequences and .
onto the estimation performance. The asymptotic CRB can also Remark 1: When the frequency offsets are a priori known to
serve as a measure to design the ; see the next section. be equal, then the model should be reparameterized as a function
In order to obtain the asymptotic CRB, we assume that the of ; see Remark 5 in Appendix I and the dis-
are realizations of zero-mean stationary random processes, cussion at the end of the previous section. Using the arguments
and we examine the limiting behavior of the FIM. The main of Appendix B, it is straightforward to show that the asymptotic
result is given in the following proposition whose proof can be CRB is given in this case by
found in Appendix B.
Proposition 2: Under mild conditions on (see
Appendix B), the asymptotic CRB for the estimation of
can be written as
(16)

(10) with and where

is the correlation matrix of the .
with
III. SELECTING THE TRAINING SEQUENCES
(11)
In this section, we briefly discuss how the sequences
could be designed. Similarly to [16], a sensible approach would
and where and are defined through
be to minimize the CRB for joint estimation of the channel gains
and the frequency offsets. The rationale behind this approach is
that the CRB is asymptotically achieved by the maximum like-
Re (12) lihood estimator (MLE). However, the CRB turned out to be a
rather complicated function of the channel gains and the fre-
The previous equation provides a simple expression of the quency offsets; see (7) and (8). Hence, we examine the asymp-
asymptotic CRB as a function of the channel gains and the totic CRB that i) no longer depends on the frequency offsets and
correlation properties of the pilot symbols. It is applicable to ii) is a simple function of the channel gains. As previously said,
a fairly large class of training sequences, including the case the asymptotic CRB is, in most situations, given by (13) rather
where corresponds to the output of a space-time block than (10). Indeed, as soon as for or
code. It should be observed that the asymptotic CRB no longer whenever the frequency offsets at each Rx antenna are different,
depends on the values of ; it only depends on whether they (13) holds true. Since this is likely to be the most usual situation,
are equal or not. Furthermore, when the random processes we focus on (13). As (13) only depends on through their
and are uncorrelated [more precisely, it suffices that powers , we propose to select the powers transmitted via each
for ] or when , Tx antenna to minimize the trace of . Since the total trans-
the matrix becomes real-valued and diagonal, and the mitted power must be constrained, the previously stated goal
asymptotic CRB takes the very simple form amounts to solving the following problem:

Tr subject to (17)

(13) However (to within an irrelevant scaling factor)

where diag with . In Tr

particular, (13) implies that

with
asCRB (14)

asCRB (15)
BESSON AND STOICA: PARAMETER ESTIMATION OF MIMO FLAT-FADING CHANNELS 605

The problem in (17) is readily solved by noting that for details)

(19)

where . Since the term in the sum

of (19) depends on and only, each of these terms can be
minimized independently, and for each Rx antenna, one is left
with the problem of minimizing

(20)
with equality if . Imposing the constraint, the solution
of (17) is thus with respect to and . It is well known that for a given ,
the minimizer of (20) is
(18)
(21)

Inserting (21) in (20), the frequency offsets estimate is obtained

This solution depends on the channel gains, which are unknown. as
Hence, only a two-step procedure could be used where the most
recent estimates of the channel gains may be used to compute (22)
the optimal powers for the next block of training symbols. How-
ever, this method can work well only if the channel gains do not Note that (22) is an -dimensional minimization problem. In
change significantly from one burst to the other. If this is not addition, the matrix depends in a compli-
the case, then we lack the information needed for an optimal cated way on . Finally, evaluation of the function in (22) re-
design and choosing all equal may be a reasonable possible quires a matrix inversion for each value of . However, the
choice. To summarize, an “optimal” power selection depends criterion in (22) can be evaluated using the fast Fourier trans-
on the channel; without information about the channel, we may form (FFT). We illustrate this property for the case of ,
choose . which will be studied in the numerical examples. Let
Remark 2: When the frequency offsets are common to all denote the symbols sent via the th Tx
Tx antennas [and when this fact is known and hence taken antenna, and let . First, note that
into account in the model], the asymptotic CRB is given by
(16). Using arguments similar to those in [16], it can be shown (23)
that choosing is minmax optimal. In other words,
minimizes the trace of the asymptotic CRB (16)
corresponding to the worst-case channel, i.e., it is the solution where diag , and stands for the Schur–
Hadamard (i.e., element-wise) product. Then, it is straightfor-
to const. Tr .
ward to show that the criterion in (22) can be written as (24),
shown at the bottom of the page, where . The
IV. CHANNEL GAINS AND FREQUENCY OFFSETS ESTIMATION three terms , , and can be evaluated
In this section, we consider the problem of estimating the on a regular grid using an FFT. Additionally, only
channel gains and the frequency offsets. We begin with exam- depends on the training sequences [which are known] and
ining the MLE for the problem at hand. Since the MLE en- could be stored. Hence, the computational complexity of the
tails solving an -dimensional minimization problem, a simpler MLE can be reduced by using the above properties.
method that enables us to decouple the -dimensional problem
in one-dimensional (1-D) problems will be proposed. This is B. Computationally Simpler Approach
achieved through a specific choice of the training sequence, as Although the MLE can be implemented using the FFT, it is
described below. still an -dimensional problem. Hence, it is of practical interest
to search for an estimator with reduced complexity. A possible
A. Maximum Likelihood Estimation simplification of the MLE consists of using the following fact
Under the assumptions made, the negative log-likelihood proved in Appendix B. When for , the
function is given, up to an additive constant, by (see Appendix A matrix converges to a real-valued diagonal matrix

Re
(24)
606 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 3, MARCH 2003

whose entries are the powers of the training sequences. Hence, MLE only for a specific choice of the training sequence. Note
the function in (22) converges to that although the number of data samples has been halved the
“aperture” remains the same as 2 for is still .
Let and .
Observe that in (29) is the maximum likelihood estimator of
(25) if one considers as a constant-amplitude exponential
signal. Indeed, for , we have from (28)

Therefore, the “asymptotic” concentrated likelihood function (30)

is a sum of terms, each of them depending solely on a
single . Hence, maximizing the right-hand side of (25) with , , , and
enables us to replace the original -dimensional problem by . If is a phase-modulated signal, then is truly
1-D problems. Additionally, the criterion to be minimized constant-amplitude exponential signal in white noise. This ob-
is recognized as the periodogram of and can be servation paves the way to the use of a myriad of techniques
implemented efficiently using the FFT. However, it is our already proposed in the literature; see e.g., [17]. For instance, a
experience that maximizing the right-hand side of (25) provides very simple way to estimate was proposed in [18] and con-
estimates with poor accuracy. The main reason of this poor sists of estimating the frequencies as
performance lies in the fact that (25) holds only asymptotically.
For example, for , neglecting in small data samples
the third term in the numerator of (24) and the denominator angle (31)
results in too crude an approximation.
However, the previous discussion suggests that a diagonal where { } is a window designed to minimize the variance of
would enable us to split the original -dimensional the estimator; see [18] for details. The above estimator is simple
problem into 1-D problems. In the sequel, we therefore and does not necessitate phase unwrapping. A high SNR ap-
look for a way to make diagonal for any and any proximation was also proposed in [18], which further simpli-
. Toward this end, we make use of the degrees of freedom fies the estimator. It consists of interchanging the operations of
provided by the choice of . For the sake of clarity, we taking the angle and summation; more precisely, the estimate is
consider the case of in the sequel (extension to an computed as angle . However,
arbitrary is straightforward). From the inspection of (23), it
there is a loss of performance, especially for low signal-to-noise
is clear that
ratios. Further refinements that consist of performing linear re-
is a diagonal matrix (26) gression over the phase of correlation lags [with ]
could be investigated; however, this would increase the com-
The latter condition can be satisfied, for instance, if we choose putational burden and unavoidably require phase unwrapping,
which is a delicate task. Hence, we focus on (31), which we
refer to as the correlation-based estimator.
(27) Remark 3: We reiterate the fact that, although (29) is a
true ML estimate, it is valid only when satisfies (27). In
for (we here assume that is even). In the contrast, (22) is applicable to any training sequence; however,
previous equation, and are arbitrary sequences. With it requires solving an -dimensional problem. Regarding
the choice (27), is diagonal for any and any . The the performances of these two methods, the following ob-
signal at the th Rx antenna takes the following form: servations should be made. The asymptotic CRB [see (14)
and (15)] depends on the training sequences only through
their powers . Usually, the total power transmitted is
(28)
constrained, i.e., is (on average) a constant. Hence,
Using (25), the frequency offsets estimates are now given by for a given power sent via the th TX antenna, the sequence
{ } results in the same asymptotic CRB
as sending { }
or { }: is even
exactly the same if is a phase-modulated signal. This
suggests that selecting as in (27) will not penalize the
accuracy, provided that the total Tx power is constant. This is
(29)
an additional argument to use the SMLE (29).
Remark 4: For the “standard” frequency estimation
The FFT can be used to compute efficiently the above functions. problem, for which , one usually encounters problems
Once and have been calculated, the channel gains are in high-resolution scenarios. Indeed, when frequency offsets
computed from (21). We stress the fact that and above are closely spaced, the matrix is nearly rank-deficient,
are the true MLEs of and for the choice (27). However, which gives rise to numerical problems. Herein, we can
we refer to it as the specific MLE (SMLE) since (29) is the use the flexibility of choosing to our advantage. First,
BESSON AND STOICA: PARAMETER ESTIMATION OF MIMO FLAT-FADING CHANNELS 607

Fig. 1. +
CRB (solid lines), asymptotic CRB (“3”) and MSE of the SMLE (“ ”), and the correlation-based estimator (“o”) versus the training sequence length.
SNR = 15 dB.
note that converges to a scaled identity matrix, performance of the SMLE (22) and the correlation-based esti-
regardless of the frequency offsets separation. Moreover, mator (31) proposed in the previous section is illustrated and
we can choose such that ; this is the case for compared with the CRB. In all simulations, the number of Tx
instance if a space-time encoder is built from the designs and Rx antennas is set to and , respectively. For
proposed in [7], [8]. Consequently, the sake of clarity, we only display the results for the first Rx
is likely to be small for antenna (similar results were obtained for the second Rx an-
, regardless of the values { }. In fact, even for tenna). The training sequences are phase-modulated (QPSK)
, can be signals generated according to (27) with
arbitrarily small [and even equal to zero] for . Hence, the and . With this choice, the covariance matrix
matrix is usually well conditioned, which in turn im- of is the identity matrix. The additive noise
plies that (22) should not give rise to numerical problems with is zero-mean, complex-valued, and Gaussian distributed with
closely spaced frequencies. Furthermore, when the training variance SNR, where SNR denotes the signal-to-noise
sequence is selected as in (27), is diagonal, irrespec- ratio.
tive of the frequency offsets separation. This is a significant In the simulations, only the “random” part of the model,
advantage over the conventional frequency estimation problem. i.e., the additive noise , is varied from trial to trial. A
Finally, observe that the number of frequencies to be estimated specific channel is used in all simulations; more precisely,
is known, which is an additional appealing feature. is drawn from an i.i.d. zero-mean complex Gaussian
distribution with variance equal to 1, and this channel
is used in all runs. For our particular channel, we have
V. NUMERICAL EXAMPLES . The fre-
quency offsets are selected as . Hence,
The aim of this section is twofold. First, the validity of the they are rather close to each other. This example is chosen
asymptotic formula (13) is checked; moreover, we study the to show that the additional degrees of freedom provided by
number of samples , which is required for the asymptotic CRB enable us to deal with closely spaced frequency offsets.
to be an accurate approximation of the exact CRB. Second, the Observe that estimating with would be a hard
608 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 3, MARCH 2003

Fig. 2. +
CRB (solid lines), asymptotic CRB (“3”) and MSE of the SMLE (“ ”), and the correlation-based estimator (“o”) versus SNR. N = 16.

problem, requiring a large number of samples ( ). For lation-based estimator to accurately estimate the two fre-
each simulation, 500 Monte Carlo trials are run. quencies, due to the knowledge of and the way it has
In a first simulation, the influence of is studied, whereas been chosen.
the SNR is fixed at SNR dB. In Fig. 1, we plot the CRB, In a second simulation, the influence of SNR is studied for
asymptotic CRB, and the empirical mean-square error (MSE) . The results are displayed in Fig. 2. Again, it can
of the estimators. The following facts can be observed. be seen that the asymptotic CRB provides an accurate approx-
• The asymptotic CRB is very close to the exact CRB, even imation of the exact CRB. A threshold effect is observed for
for short data samples, which are typically . This both the SMLE and the correlation-based estimator. For SNR
is an important result as it implies that the very simple above the threshold, the two methods provide efficient estimates
expression of the asymptotic CRB can be used instead while a breakdown is observed below the threshold. Note that
of the complicated expression of the CRB for practical the threshold is lower for the SMLE than for the correlation-
training sequences lengths. based estimator; in other words, the SMLE performs better in
• Since the asymptotic CRB is very close to the exact CRB low SNR environments. It should also be pointed out that the
and since the former depends only on the training se- threshold does not occur at the same SNR for and . This
quence powers, it implies that the exact CRB is not very fact can be explained as follows. The estimators are based on
sensitive to the particular realization of . Indeed, in , which is given by (30). In our definition of the SNR
other simulations not reported here, we verified that the [remember that SNR], the factors are not taken
variations of the CRB are negligible when is changed into account. Hence, the “effective” SNR, considering as
[provided that its power is kept constant]. a constant-amplitude exponential signal, is SNR SNR
• The SMLE has aperformance very close to the CRB forboth : For our choice of , this results in an SNR loss
the frequency offsets and the channel gains estimation. The of 4.52 dB and 0.57 dB for and , respectively.
correlation-based estimator is as accurate as the SMLE. This is why a higher SNR is required to estimate . Once
• Even though is small and the number of sam- above the threshold, the correlation-based estimator performs
ples is small, there is no problem for the SMLE and corre- as well as the SMLE and achieves the CRB.
BESSON AND STOICA: PARAMETER ESTIMATION OF MIMO FLAT-FADING CHANNELS 609

VI. CONCLUSIONS Now, observe that depends only on and . It follows

In this paper, we addressed the channel and frequency off- that the FIM is block-diagonal, i.e.,
sets estimation problem for a MIMO flat-fading channel when
the frequency offsets may differ from one Tx antenna to an-
other. We considered the data-aided framework, where known ..
. (35)
training sequences are used to probe the channel. A closed-from ..
expression for the Cramér-Rao bound was derived. A simpler, .
large-sample CRB was also presented; we illustrated the fact
that the asymptotic CRB is very close to the exact CRB even where each is the FIM corresponding to estimation of
for a small number of pilot symbols. It was shown that as the Re Im . We now derive all entries of
number of samples grows, the CRB no longer depends on the in a step-by-step manner. From (32), we have
frequency offsets but only on the channel gains. This holds true,
irrespective of the frequency offsets separation, provided that (36a)
the symbols sent at the same instant on the Tx antennas are Re
uncorrelated. Starting from the maximum likelihood estimator,
we proposed a specific training sequence choice that enables us (36b)
Im
to simplify the MLE and leads to a simple estimator. Further-
more, a correlation-based method was also presented. The two (36c)
methods were shown to provide accurate estimates.
which, along with (34), implies that, for
APPENDIX A
DERIVATION OF THE CRB Re
In this Appendix, we provide an expression of the CRB
for the estimation of . First, note that the vector Re (37a)
of received signals can be written as

Im
with ,
, and
Im (37b)
(32)

Therefore, under the stated assumptions, Re

is distributed as
Re (37c)
(33)

with . For notational conve-

nience, let us also introduce the following vectors and matrices Im
(along with their dimensions) that will be of use in subsequent
derivations:

Im (37d)
diag
diag
Re
For distributed as in (33), it is known [17], [19] that the FIM for
estimation of is block-diagonal, i.e., the estimation Re (37e)
of is decoupled from that of . Therefore, in the sequel, we
only consider the FIM for , which we denote by . The th
element of the latter is given by [19], [17] Re

Re

Re (34)
Re (37f)
610 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 3, MARCH 2003

Rewriting (37) in a more compact form yields the expression for and therefore, the CRB for the estimation of is given by
shown in (38) at the bottom of the page. Let us now proceed
Re Im
to the derivation of the CRB. Since the FIM is block-diagonal,
the CRB is also block-diagonal; hence, in the sequel, we focus Im Re
on . For the sake of clarity, let us partition
as Im
Re Re

Im Re
where corresponds to the lower right corner of (39)
, and the other matrices definitions follow obviously from
inspection of (38). Accordingly, for notational convenience, let It should be pointed out that there is coupling in the CRB be-
us define , , and tween the estimation errors of Re , Im , and . In par-
. The inverse of is given by [20] ticular, it implies that the channel gain estimation errors may not
be circular. The CRB associated with the estimation of is ob-
tained as the lower right corner of the previous matrix, yielding

Re

Re

Observing that
where . The CRB for the
estimation of is obtained as the th diagonal element of
Re Im . In order to derive the CRB associated with the es-
Im Re timation of (and not only its real and imaginary part), let us
define and note that
it follows that

Im Re
Re Im Im The CRB for the estimation of is thus given by (40), shown
Im Re Re at the bottom of the page.
Remark 5: The previous derivations can be modified in a
Re
straightforward manner when it is known a priori that the fre-
quency offsets are common to all transmit antennas, i.e., when
and hence for . In this case, one should define the
vector of interest as Re Im . The
Re Re derivatives in (36) should then be modified to

Similarly (41a)
Re

Re Im Im (41b)
Im
Im Re Re
Im (41c)
Re

Re Im Im
Im Re Re (38)
Im Re Re

Re Re
(40)
Re Re
BESSON AND STOICA: PARAMETER ESTIMATION OF MIMO FLAT-FADING CHANNELS 611

Mimicking the derivations that led to (38), one can readily for some integer . Indeed, one has
show that the FIM corresponding to the estimation of
can be written as (42), shown at the bottom of the page,
where . The lemma for the inverse of partitioned (47a)
matrices can again be invoked to obtain the CRB associated
with the estimation of , and subsequently, the CRB for
, resulting in
(47b)

with , , (47c)
Therefore, in the sequel, we concentrate on above and derive
and .
an asymptotic expression for it. The statistical average of is
given by
APPENDIX B
DERIVATION OF THE ASYMPTOTIC CRB
In this appendix, we derive an expression for the asymptotic
(i.e., large-sample) CRB that hopefully results in a much simpler
(48)
dependence w.r.t. the channel and the pilot symbols. Toward this
end, we consider that the are realizations of (possibly cor- where and where we used the fact that
related) zero-mean random stationary processes that satisfy the
so-called mixing conditions [21] (see below). Next, we evaluate
the limit of the statistical average of the quantities appearing in .
the FIM. Finally, we show that as tends to infinity, these quan- We now show that under the mild conditions stated in (43)
tities converge in mean-square sense to their statistical average, (49)
which enables us to derive the asymptotic CRB.
In the sequel, we therefore assume that are where the limit should be understood in the mean-square sense.
zero-mean random sequences such that all their joint cumulants Toward this end, let us examine the variance of
of all orders satisfy

cum

(43)
where . This condition
means that samples of that are well separated in time
are approximately independent. Although it may appear some-
what technical, we stress the fact that it is applicable to a fairly
large class of signals; see our Remark 6. Let us introduce the
normalized FIM

(44)

and corresponding normalized CRB with

(45)

From the inspection of (38), it is readily verified that each el-

ement of the normalized FIM can be written in the following
form: cum

(46) (50)

Re Im Im
Im Re Re (42)
Im Re Re
612 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 51, NO. 3, MARCH 2003

In order to prove (49), it remains to show that the three terms in is obtained as the inverse of (54). After some straightforward
the last equality of (50) converge to zero. As the proofs are sim- calculations, similar to those used for inverting the finite-sample
ilar, we only prove the result for the first term, which is referred FIM, the asymptotic (normalized) CRB is readily seen to be
to as . This is achieved by writing that
Re Im
Im Re

Im
Re

Im Re
(55)

with Re . Moreover, similarly to

what was done in Appendix A, the asymptotic CRB for the
estimation of can be readily evaluated, and one ends up with
the following expression:

(56)

Therefore, . Similar calculations can be con- Remark 6: The above derivations were carried out under the
ducted with the other terms in (50), and we arrive at the main hypothesis that (43) holds true. Despite the fact that (43) may
result appear somehow technical, it applies to many signals encoun-
tered in practice. More precisely, the asymptotic formula (54)
(51) is valid whenever the are stationary processes such that
is of finite duration or exponentially de-
caying (see [20, App. B] for related derivations). We also stress
Applying (51) to (47), we get the fact that (54) is valid when corresponds to the output of
a space-time block code (STBC) designed with the procedures
(52a) described e.g., in [7] and [9]. A sketch of the proof is now given.
Any STBC can be represented as a transmission matrix
(52b) where is the number of time slots needed to transmit sym-
bols. Since at every samples new and independent symbols
(52c) enter the space-time encoder, it follows that

where the ( )th element of is given by for (57)

(53) where or . Observe that and

are correlated for , even though we have
Observe that the “correlation matrix” is diagonal whenever for [symbols sent at the same time on
for or when , even two different Tx antennas are independent]. In fact, the finite-
though the processes and may be correlated. Using length property in (57) is all that is required to prove (54). Since
(38) and (52), one can obtain the asymptotic FIM as (54), shown the main step in deriving (54) is to show (51) or, equivalently,
at the bottom of the page. The asymptotic CRB for estimation of to show that the three terms in (50) converge to zero, we focus

Re Im Im
Im Re Re (54)
Im Re
BESSON AND STOICA: PARAMETER ESTIMATION OF MIMO FLAT-FADING CHANNELS 613

on the latter property. More precisely, one has [15] Z. Liu, G. Giannakis, and B. Hughes, “Double-differential space-time
block coding for time-selective fading channels,” IEEE Trans.
Commun., vol. 49, pp. 1529–1539, Sept. 2001.
[16] P. Stoica and O. Besson, Training Sequence Design Freq. Offset Channel
Estimation Freq. Selective Channels, Nov. 2001, submitted for publica-
tion.
[17] P. Stoica and R. Moses, Introduction to Spectral Analysis. Upper
Saddle River, NJ: Prentice-Hall, 1997.
[18] S. Kay, “A fast and accurate single frequency estimator,” IEEE Trans.
Acoust., Speech, Signal Processing, vol. 37, no. 12, pp. 1987–1990, Dec.
1989.
[19] , Fundamentals of Statistical Signal Processing: Estimation
Theory. Englewood Cliffs, NJ: Prentice-Hall, 1993.
[20] T. Söderström and P. Stoica, System Identification. London, U.K.:
Prentice-Hall Int., 1989.
[21] A. Dandawaté and G. Giannakis, “Asympotic theory of kth-order cyclic
moment and cumulant statistics,” IEEE Trans. Inform. Theory, vol. 41,
pp. 216–232, Jan. 1995.

Olivier Besson (M’92) is currently an Associate Professor with the Department

of Avionics and Systems, ENSICA, Toulouse, France. His research interests are
in the general area of statistical signal and array processing with applications in
communications and radar.
Dr. Besson is the co-recipient of the IEE Clarke-Griffiths premium award and
the Eurasip Best Paper Award in 1998. He is a member of the SAM Technical
Committee of the IEEE Signal Processing Society and serves as an Associate
Editor for IEEE TRANSACTIONS ON SIGNAL PROCESSING.
which proves that goes to zero as tends to infinity. Conse-
quently, the asymptotic expression of the FIM holds for the case
where an STBC is used to generate the training sequences. Petre Stoica (F’94) received the D.Sc. degree in
authmatic control from the Polytechnic Institute of
Bucharest (BPI), Bucharest, Romania, in 1979 and an
REFERENCES honorary doctorate degree in science from Uppsala
University (UU), Uppsala, Sweden, in 1993.
[1] A. Paulraj and C. Papadias, “Space-time processing for wireless com-
He is Professor of system modeling with the
munications,” IEEE Signal Processing Mag., vol. 14, pp. 49–83, Nov.
Department of Systems and Control at UU. Previ-
1997.
ously, he was a Professor of system identification
[2] A. Naguib, N. Seshadri, and A. Calderbank, “Increasing data rate over
and signal processing with the Faculty of Automatic
wireless channels,” IEEE Signal Processing Mag., vol. 17, pp. 76–92,
May 2000. Control and Computers at BPI. He held longer
[3] G. Ganesan and P. Stoica, “Space-time diversity,” in Signal Processing visiting positions with Eindhoven University of
Advances in Wireless & Mobile Communications, G. Giannakis, Y. Hua, technology, Eindhoven, The Netherlands; Chalmers Univerity of Technology,
P. Stoica, and L. Tong, Eds. Upper Saddle River, NJ: Prentice-Hall, Gothenburg, Sweden (where he held a Jubilee Visitinbg Professorship); UU;
2001, vol. 2, Trends in Single and Multi-User Systems, ch. 2, pp. 59–87. The Universitry of Florida, Gainesville; and Stanford Univerity, Stanford, CA.
[4] Z. Liu, G. Giannakis, B. Muquet, and S. Zhou, “Space-time coding His main scientific interests are in the areas of system identification, time series
for broadband wireless communications,” Wireless Commun. Mobile analysis and prediction, statistical signal and array processing, spectral analysis,
Comput., vol. 1, no. 1, pp. 35–53, Jan.-Mar. 2001. wireless communications, and radar signal processing. He has published
[5] S. Alamouti, “A simple transmit diversity technique for wireless com- seven books, ten book chapters, and some 450 papers in archival journals and
munications,” IEEE J. Select. Areas Commun., vol. 16, pp. 1451–1458, conference records on these topics. The most recent book he co-authored, with
Oct. 1998. R. Moses, is entitled Introduction to Spectral Analysis (Englewood Cliffs,
[6] V. Tarokh, N. Seshadri, and A. Calderbank, “Space-time codes for high NJ: Prentice-Hall, 1997). Recently, he edited two books on signal processing
data rate wireless communications: Performance criterion and code con- advances in wireless communicaitons and mobile communications (Englewood
struction,” IEEE Trans. Inform. Theory, vol. 44, pp. 744–765, Mar. 1998. Cliffs, NJ: Prentice-Hall, 2001). He is on the editorial boards of five journals
[7] V. Tarokh, H. Jafarkhani, and A. Calderbank, “Space-time block codes in the field: Journal of Forecasting; Signal Processing; Circuits, Signals,
from orthogonal designs,” IEEE Trans. Inform. Theory, vol. 45, pp. and Signal Processing; Digital Signal Processing—A Review Journal; and
1456–1467, July 1999. Multidimensional Ssytems and Signal Processing. He was a Co-Guest Editor
[8] G. Ganesan and P. Stoica, “Space-time block codes: A maximum snr for several special issues on system identification signal processing, spectral
approach,” IEEE Trans. Inform. Theory, vol. 47, no. 4, pp. 1650–1656, analysis, and radar for some of the aforementioned journals, as well as for
May 2001. PROCEEDINGS OF THE IEEE.
[9] , “Space-time diversity using orthogonal and amicable orthogonal Dr. Stoica was co-recipient of the IEEE ASSP Senior Award for a paper on
designs,” Wireless Personal Commun., vol. 18, no. 2, pp. 165–178, Aug. statistical aspects of array signal processing. He was also recipient of the Tech-
2001. nical Achievement Award of the IEEE Signal Processing Society for fundamental
[10] V. Tarokh, A. Naguib, N. Seshadri, and A. Calderbank, “Space-time contributions to statistical signal processing with applications in time-series anal-
codes for high data rate wireless communications: Performance criteria ysis, system identification, and array signal processing. In 1998, he was the recip-
in the presence of channel estimation errors, mobility and multiple ient of a Senior Individual Grant Award of the Swedish Foundation for Strategic
paths,” IEEE Trans. Commun., vol. 47, pp. 199–207, Feb. 1999. Research. He was also co-recipient of the 1998 EURASIP Best Paper Award
[11] G. Ganesan and P. Stoica, “Utilizing space-time diversity for wireless
for Signal Processing for a work on parameter estimation of exponential signals
communications,” Wireless Personal Commun., vol. 18, no. 2, pp.
with time-varying amplitude, a 1999 IEEE Signal Processing Society Best Paper
149–163, Aug. 2001.
[12] T. Marzetta, “BLAST training: Estimating channel characteristics for Award for a paper on parameter and rank estimation of reduced-rank regression,
high capacity space-time wireless,” in Proc. 37th Annual Allerton Conf. a 2000 IEEE Third Millenium Medal, and the 2000 W. R. G. Baker Prize Paper
Commun., Contr., Comput., Monticello, IL, Sept. 1999, pp. 958–966. Award for a paper on maximum likelihood methods for radar. He was a member
[13] B. Hughes, “Differential space-time modulation,” IEEE Trans. Inform. of the international program committees of many topical conferences. From 1981
Theory, vol. 46, pp. 2567–2578, Nov. 2000. to 1986, he was the Director of the International Time-Series Analysis and Fore-
[14] H. Jafarkhani and V. Tarokh, “Multiple transmit antenna differential casting Society, and he has been a member of the IFAC Technical Committee on
detection from generalized orthogonal designs,” IEEE Trans. Inform. Modeling, Identification, and Signal Processing since 1994. He is also a member
Theory, vol. 47, pp. 2626–2631, Sept. 2001. of the Romanian Academy and a Fellow of the Royal Statistical Society.