You are on page 1of 6


3, MARCH 2009 1133

Smooth Beamforming for OFDM

Magnus Sandell, Senior Member, IEEE, and Vishakan Ponnampalam, Senior Member, IEEE

Abstract—This paper investigates the combination of beam- considered across the subcarriers, there is no guarantee that the
forming and channel state information (CSI) smoothing for beamformer will be smooth since the optimum beamforming
OFDM. While both can offer significant performance gains, care solution is not unique. In this paper we will investigate the
must be taken when using both of them at the same time. Two
different sets of beamforming weights can produce the same use of orthogonal iterations [4] to compute the beamforming
beamforming gains but be unsuitable for CSI smoothing due to weights, which not only results in a smooth beamformer but
sudden changes in the frequency domain. By using orthogonal also low complexity due to its incremental approach. This
iterations to compute the beamforming weights, it is shown technique is an improvement over the smooth beamformer in
that a smooth beamformer is produced which is suitable for [3], which, in addition to its inferior performance, is in practice
CSI smoothing at the receiver. An additional benefit of this
technique is that the beamforming weights are easy to compute limited to a two-antenna system. In short, this paper offers two
by exploiting the frequency correlation of the channel. The insights to beamforming with receiver CSI smoothing:
performance is evaluated for an OFDM system based on the • It is not necessary to compute the exact eigenmodes to
IEEE 802.11n draft standard for which the gains are quantified. get the most out of beamforming. It is a flat optimum, so
We also present results obtained from a hardware test bed, which
verify the performance of the proposed smooth beamforming if the beamformer is near enough to the optimal solution,
technique. essentially all the benefits from the beamforming can be
Index Terms—Beamforming, channel estimation, OFDM,
• This observation can be used to efficiently compute the
MIMO, SVD, eigenvectors.
eigenmodes using orthogonal iterations, since it will be
shown that only one iteration is sufficient. This also
I. I NTRODUCTION ensures a smooth beamformer, facilitating CSI smoothing

B EAMFORMING can achieve significant gains in

multiple-input, multiple-output (MIMO) systems. By us-
ing the channel state information (CSI) at the transmitter, the
at the receiver.
This paper is organised as follows: the system model,
beamforming and CSI smoothing are described in Section II.
channel can be shaped according to its eigenmodes, which The novel beamforming weight computation is explained in
can improve the link performance. Eigenmode beamforming Section III and analysed in Section IV. Simulation results
uses the singular vectors of the channel matrix as beam- illustrating the performance is presented in Section V and
forming weights [1], resulting in orthogonal columns of the conclusions are drawn in Section VI.
beamformed channel matrix. This means that transmission
is done in the dominant subspace of the MIMO channel. II. BACKGROUND
Eigenmode beamforming assumes full channel knowledge
A. System model
at the transmitter, which can be facilitated by feedback or
channel reciprocity. A MIMO OFDM system with M transmit, N receive
Another feature that may significantly improve performance antennas and K subcarriers can be modelled as
is CSI smoothing at the receiver. In an OFDM system, rk = Hk xk + vk , k = 1, · · · , K (1)
subcarrier channel coefficients are correlated across frequency
which may be exploited to improve the initial per-tone channel where rk ∈ CN is the received signal, Hk ∈ CN ×M is
estimate obtained from a preamble, e.g., through a Wiener- the channel, xk ∈ CM is the transmitted signal, vk ∈ CN
filtering approach in the frequency domain [2]. As the beam- is the noise and the subscript k denotes the subcarrier in-
forming and CSI smoothing techniques are independent, it is dex. The elementwise variances
the variables
 in (1) are
2 2
desirable to combine them in order to get the full benefits σh2 = E |Hk (n, m)| , σx2 = E |xk (m)| and σv2 =
of both. However, in order for this to work, the frequency 2
E |vk (n)| , with the SNR defined as
response of the beamformed channel (which the receiver sees)
must have no "discontinuities" or sudden changes. We call M σx2 σh2
such a beamformer smooth [3]. If beamforming is not jointly SN R  . (2)
Manuscript received March 5, 2008; revised July 17, 2008; accepted August
8, 2008. The associate editor coordinating the review of this paper and B. Beamforming
approving it for publication was D. Dardari.
This paper was presented in part at the GLOBECOM Conference in In eigenmode beamforming [1], the beamforming weights
November 2007. are the right singular vectors Vk of the channel Hk , such that
The authors are with Toshiba Research Europe Ltd, Telecommunications the transmitted signal is xk = Vk sk , where sk ∈ SJ are the
Research Laboratory, 32 Queen Square, Bristol BS1 4ND, UK (e-mail: data symbols, S is the constellation alphabet and J ≤ M is the
Digital Object Identifier 10.1109/TWC.2008.080312 number of spatial streams. We let Vk ∈ CM×J denote the J
c 2009 IEEE

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:43 from IEEE Xplore. Restrictions apply.

singular vectors corresponding to the J largest singular values B EAMFORMING WEIGHT COMPUTATION USING ORTHOGONAL ITERATIONS
of Hk and they span the dominant subspace. Hence the task
of the beamformer is to compute the right singular vectors  0 = IM ×J
1) Initialise V
of the channel. In this paper, we propose to compute them 2) for k = 1 : K
3) Ak = HH k Hk
incrementally, from the first subcarrier to the last. Due to the 4) Vk = V  k−1
frequency correlation, the singular vectors change gradually 5) for i = 1 : Nit
across frequency so we are essentially faced with a subspace 6) B = Ak V k
7) "Skinny" QR decomposition B = V  k Rk with V
 k ∈ CM ×J
tracking problem [5], which appears in applications such as (orthonormal columns) and Rk ∈ CJ ×J (upper triangular)
adaptive antenna array processing [6]. Note that we do not 8) end
consider power or bit loading as that does not affect the 9) If required, compute squared singular values as
2 = diag (R ) , 1 ≤ j ≤ J
calculation of beamforming weights or CSI smoothing. 10) end

C. CSI smoothing
We assume that the data packet contains a time-multiplexed [4]. The orthogonal iteration algorithm is similar to the method
preamble from which the MIMO channel is estimated, al- of QR-iterations but it only computes the dominant subspace
though it is also possible to use scattered pilots embedded [4].
in the data packet. The least squares (LS) estimate is obtained We can now devise two variants of the SVD calculation:
on a per-tone basis by simply dividing the received value (i) compute it from "scratch" by starting from the identity
by the pilot value; however, it is possible to improve this matrix, Q(0) = IM×J ; (ii) start with the matrix computed on
initial estimate by exploiting the frequency correlation of the previous subcarrier. We call the former non-incremental
the channel. Hence we apply an FIR filter to smooth the and the latter incremental computation. Since the channels on
channel and use the 2P + 1 nearest subcarriers, which means two consecutive subcarriers are likely to be highly correlated,
P subcarriers on each side. At the band edges and around it can be expected that the incremental approach would
the DC subcarrier, the nearest active subcarriers are used. A be advantageous. The complete algorithm for incrementally
Wiener filter is designed to minimise
 the mean-squared error computing beamforming weights is summarised in Table I. For
  the non-incremental approach, line four would be changed to
(MSE) of the channel estimate, E Hk − Hk  , and can  k = IM×J . It should be noted that we constrain the diagonal
be computed from the frequency correlation of the channel elements of Rk in the QR-decomposition to be real and
and the noise variance. The filter only acts in the frequency positive in order to maintain smoothness between subcarriers.
domain, hence it is applied on a per-antenna basis. Since the
channel frequency correlation is not known, we use the generic IV. C ONVERGENCE A NALYSIS
design described in [7]. It should be pointed out that different
filters could be used for the different streams since they exhibit As long as two sets of beamforming weights span the same
different frequency correlation as noted in [3]; however, we subspace, they will have the same effect on the beamforming,
use the same filters in this paper and leave the filter design i.e., the same performance. Hence the convergence of the
for future studies. orthogonal iterations can be analysed by studying the distance
The beamforming will make the frequency correlation of between the actual and estimated dominant subspaces. If X
the beamformed channel different from the original channel. and Y are matrices with orthonormal columns, the distance
There is also a risk that beamforming will introduce "dis- between the subspaces they span can be defined as [4]
continuities" in the channel; the singular value decomposition  
d (X, Y)  XXH − YY H  (5)
(SVD) is not unique since it may produce two sets of singular 2
vectors that span the same space but are different. Hence where 0 ≤ d (X, Y) ≤ 1. From the algorithm in Table I and
it is crucial to maintain some kind of smoothness of the using the analysis of orthogonal iterations in [4, Sec. 8.2.4],
beamformers, which means coupling the beamforming weight we then have
computation on different subcarriers. 
 d Vk , V k−1
k ≤
d Vk , V  λ (6)
1 − d Vk , Vk−1
The right singular vectors of H are the same as the
eigenvectors of A = HH H and eigenvectors of Hermitian where λj is the jth largest eigenvalue of Ak , i.e., λJ is the
matrices can be found by using orthogonal iterations [4] from smallest eigenvalue of the dominant subspace and λJ+1 is the
an initial matrix Q(0) ∈ CM×J with orthonormal columns and largest eigenvalue of the complementary subspace. If Ak =
performing the iterations Ak−1 + Ek , the perturbation analysis of eigenvectors [8] says
B(i) = AQ(i−1) , i = 1, 2, · · · (3) that
Ek Vk−1 F
B(i) = Q(i) R(i) : QR decomposition (4) d (Vk , Vk−1 ) ≤ . (7)
λJ − λJ+1
It can be shown that R(i) converges to a diagonal matrix Combined with the triangle equality
containing the eigenvalues of A and that Q(i) converges to an  
 k−1 ≤ d (Vk, Vk−1 ) + d Vk−1, V
d Vk, V  k−1 (8)
orthonormal basis for the dominant subspace of dimension J

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:43 from IEEE Xplore. Restrictions apply.

d(V ,V ) = 0.05
k−1 k−1
0.9 d(Vk−1,Vk−1) = 0.1
d(V ,V ) = 0.2 0.18
k−1 k−1
M aximum λJ +1 /λJ

Subspace distance
0.7 0.14


0.4 0.06


0.2 0
0 0.02 0.04 0.06 0.08 0.1 0 5 10 15 20
Ek Vk−1 F /λJ Subcarrier index

Fig. 1. The maximum eigenvalue ratio that will ensure a reduction in Fig. 2. Convergence of the orthogonal iterations (N = M = 3, J = 2,
subspace distance as a function of the perturbation of the channel matrix. AR(1)-process with correlation coefficient 0.99).

we then have correlation coefficient 0.99. The convergence rate is high and
Ek Vk−1 F 
+ d Vk−1, V we will show in the next section that the error floor does not
λJ −λJ+1 λJ +1
k ≤
d Vk , V . degrade the BER performance noticeably. Although not shown
 2 λJ
Ek Vk−1 F  k−1 here, the convergence of the singular values follow a similar
1− λJ −λJ+1
+ d Vk−1, V
pattern and are virtually indistinguishable from the actual ones
(9) after only a few subcarriers.
To ensure that the distance
the actual
 and estimated

subspaces is reduced, d Vk, V  k < d Vk−1, V  k−1 , we

Ek Vk−1 F  k−1
+ d Vk−1, V Our simulations are based on the IEEE 802.11n system
λJ 1− λ λJ+1
 J draft [9] and we consider the 20 MHz bandwidth mode with
  2 λJ K = 64 subcarriers, a cyclic prefix of 16 symbols, 100
1 − Ek Vk−1 F + d V 
λ k−1, Vk−1 J+1
byte packets and a preamble for channel estimation. All the
λJ 1− λJ
impairments such as frequency offset, phase noise and non-

<  k−1 .
d Vk−1, V (10) linear PAs, as defined in [10], are applied and an MMSE
decoder is used at the receiver. Cyclic delay diversity (CDD)
From (6), it is clear that the subspacedistance at subcarrier is used but it is important to note that it must be removed
k depends on the starting point, i.e., d Vk, V  k−1 , and the before CSI smoothing is applied, otherwise it will not work.
Feedback CSI is provided to the transmitter with the same
convergence rate λJ+1  /λJ < 1. Hence given the previous SNR as the forward link, the channel coefficients are quantised
subspace distance d Vk−1, V  k−1 and the magnitude of the
with 6 bits per real and imaginary part and the feedback
perturbation Ek Vk−1 F /λJ , from (10) we can compute the delay is 5 ms. Simulations not shown here indicated that a
maximum eigenvalue ratio that will ensure that the subspace 2P + 1 = 7-tap CSI smoothing filter was a good trade-off
distance is reduced. In Figure 1, this is plotted against the between performance and complexity, and was hence used
perturbation for some initial distance values. As can be seen, for the simulations. The filter was designed as described in
the larger the perturbation is, the smaller the eigenvalue ratio Section II-C, assuming a uniform power-delay profile equal
must be to ensure a distance reduction. This is natural since to the cyclic prefix (16 symbols) and an SNR of 30 dB; this
the eigenvalue ratio determines how quickly the orthogonal resulted in the coefficients
iterations converge and hence a smaller ratio can allow for a
larger perturbation.
the larger the previous subspace wk = {0.03 + 0.05j, 0.01 − 0.01j, 0.23 − 0.21j, 0.46,

distance, d Vk−1, Vk−1 , the larger the eigenvalue ratio can 0.23 + 0.21j, 0.01 + 0.01j, 0.03 − 0.05j} . (11)
be since it is easier to reduce the distance in this case.
It should be noted that the above analysis only shows As the most challenging case is a high data rate in a frequency-
when a subspace distance reduction is guaranteed. Viewing selective channel (which can make CSI smoothing difficult),
the channel as a stochastic process across the subcarriers, we use 64QAM modulation and a rate 2/3 convolutional code
the subspace distance might actually increase occasionally in channel model D unless otherwise noted. The IEEE 802.11n
due to a large change between subcarriers but on average it channel models cover six fundamental cases, ranging from
will decrease until it reaches an equilibrium. In Figure 2 the model A (frequency-flat) to model F (150 ns RMS delay
average subspace distance is shown for an AR(1)-process with spread) with different spatial correlations [11].

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:43 from IEEE Xplore. Restrictions apply.

10 3
Singular value 1
Singular value 2

Incremental SVD, per−tone estimation

Non−incremental, per−tone estimation 2
Incremental SVD, CSI smoothing
10 Non−incremental SVD, CSI smoothing



−2 0.5

1 2 3 4 5 6 7 8 0
10 20 30 40 50 60
Number of iterations Subcarrier index

Fig. 3. PER as a function of the iterations in the orthogonal iterations Fig. 4. Singular values of a channel realisation as a function of frequency.
algorithm (SNR = 20 dB).

A. Computer simulations seeking, and that the singular values are plotted merely as a
First we will investigate the impact of the incremental representation of these subspaces. Requiring the beamformer
computation and the number of orthogonal iterations, Nit , per to have maximum smoothness means that tracking of a par-
subcarrier. In Figure 3, the packet-error rate (PER) is shown ticular subspace, which initially was the dominant one, may
as a function of Nit when using M = N = 3 antennas become a weaker one, i.e., corresponding to a smaller singular
and J = 2 data streams. With per-tone channel estimation, value. This corresponds to the solid line in Figure 4. This is
more iterations do not improve performance except for the not a problem when J = M , i.e., the number of streams equals
non-incremental approach (initialise the eigenvectors as the the number of transmit antennas, since all singular vectors are
identity matrix instead as the eigenvectors on the previous used for beamforming. On the other hand, if J < M the J
subcarrier), where one iteration is not enough. When CSI dominant singular vectors should be chosen, which means that
smoothing is used, the non-incremental orthogonal iterations the tracking method in [3] might pick weaker singular vectors
produce a beamformer that is not smooth and its performance and not achieve the full beamforming gains. This problem
is actually worse than using per-tone channel estimation. With can be solved by sorting the calculated singular values and
incremental computation, however, there is a gain, although it always choosing the singular vectors corresponding to the
decreases with the number of iterations until it also performs largest ones. This corresponds to switching from the solid line
worse than per-tone estimation. This might seem counterintu- to the dashed line in Figure 4 when they cross over. However
itive but from the results for the per-tone channel estimation this might mean that some smoothness is lost since there is
it is obvious that one iteration is enough to produce a good an abrupt change from one subspace to another. Hence there
beamformer. With more iterations, the beamforming weights is a trade-off between beamforming and CSI smoothing.
will differ more between two neighbouring subcarriers and Our proposed algorithm solves the smooth beamforming
hence CSI smoothing will suffer. Therefore the best choice is problem by being a dominant subspace tracker. By using
to use only one iteration, which also keeps the complexity to a only one iteration per subcarrier, smoothness is ensured (no
minimum. For the rest of the paper we will use one iteration sudden changes can appear). Although it doesn’t compute
per subcarrier, Nit = 1, when computing the beamforming the actual dominant subspace on each subcarrier due to the
weights. That makes the complexity of this algorithm very frequency-selective channel, it is close enough to get all the
low since it only involves two matrix multiplications (HH k Hk beamforming gains. Hence we can achieve the full benefits
and Ak V  k−1 ) and a QR decomposition per subcarrier. of beamforming and CSI smoothing while also offering a
The orthogonal iteration method is now compared to the simple implementation. In Figure 5, all the above schemes are
technique in [3]. That algorithm is based on finding the unitary compared for a system with M = N = 2 antennas and J = 1
rotation between the two SVDs on consecutive subcarriers. stream: (i) orthogonal iterations, as proposed in this paper;
By parameterising a 2 × 2 unitary matrix with one variable (ii) the method in [3] with smoothness requirements; (iii) the
(other degrees of freedom can be ignored due to the nature of method in [3] with sorting of the singular values. When per-
the problem), the rotations are found by solving a non-linear tone channel estimation is used, schemes (i) and (iii) perform
equation. As there are two solutions, the one maintaining the identically. The performance of scheme (ii) is degraded since
most smoothness is chosen. However, as the singular values of the smallest singular vector is sometimes used. With CSI
the channel matrix vary across the frequency band, they might smoothing at the receiver, the performance of schemes (i) and
change order. This is illustrated in Figure 4 where two singular (iii) improve although the orthogonal iterations outperform
values are plotted as a function of frequency. Note that they the method in [3] with sorting. This is due to producing a
each correspond to a singular vector, which is what we are smoother beamformed channel. It should also be pointed out

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:43 from IEEE Xplore. Restrictions apply.

0 0
10 10
Orthogonal iterations, per−tone est.
[3] with smoothness, per−tone est.
[3] with sorting, per−tone est.
Orthogonal iterations, CSI smoothing
[3] with smoothness, CSI smoothing
[3] with sorting, CSI smoothing
Full SVD, known channel

Packet−error rate
Packet−error rate

10 −1

No BF, per−tone
BF, per−tone
No BF, CSI smoothing
−2 BF, CSI smoothing
8 10 12 14 16 18 20 22 24 −2
SNR [dB] 10
10 12 14 16 18 20
SNR [dB]

Fig. 5. PER for three different beamforming computation methods and two
channel estimation techniques (M = N = 2 antennas and J = 1 data Fig. 7. PER on the hardware platform with beamforming and CSI smoothing
stream. turned on and off.

ensuring that received signals were subject to all practical
OFDM impairments. Within the receiver, frame detection,
synchronisation and automatic gain control (AGC) were im-
plemented on the FPGA with all further processing being
conducted in software on a host PC. This allows for the rapid
Packet−error rate

implementation of new algorithms, as well as the ability to

directly compare algorithm performance by processing each
received packet multiple times; once for each algorithm being
Orthogonal iterations, per−tone est.
[3] with smoothness, per−tone est.
[3] with sorting, per−tone est. We used a system with M = N = 2 antennas and J = 1
Orthogonal iterations, CSI smoothing
[3] with smoothness, CSI smoothing data stream. The feedback had 6 bits quantisation per real and
[3] with sorting, CSI smoothing
Full SVD, known channel imaginary part and no delay. The beamforming weights were
computed using the incremental technique with one orthogonal
18 20 22 24 26 28 30
SNR [dB]
iteration. In Figure 7, the PER is shown for all combinations
of beamforming and CSI smoothing turned on and off. When
Fig. 6. PER for three different beamforming computation methods and two beamforming is turned off, spatial spreading is used [9]. As
channel estimation techniques (M = N = 2 antennas and J = 2 data
streams). can be seen, the benefits of using beamforming and CSI
smoothing are additive and are not mutually exclusive. The
evaluation was carried out in an office environment with
that the method in [3] only works for two antennas in practice people moving around. Although this is only a snapshot
(extensions are very difficult), while the orthogonal iteration of a typical operating environment, it shows that smooth
technique works for any number of antennas and data streams. beamforming does work in a practical system.
To illustrate the versatility of the proposed method, the
results for J = 2 data streams are shown in Figure 6. Again VI. C ONCLUSIONS
there is a gain when using orthogonal iterations and CSI
In this paper we have presented a smooth beamforming
smoothing. The reason why the method in [3] doesn’t appear
technique which can improve the performance of an OFDM
to offer any improvements with CSI smoothing might be the
system. By incrementally computing the beamforming weights
impairments, e.g., the results in [3] were obtained by assuming
across subcarriers, smoothness of the beamformed channel can
perfect knowledge of the channel correlation, which is not a
be ensured which allows for CSI smoothing at the receiver. We
realistic assumption.
have used Wiener filtering in the frequency domain, which has
resulted in significant gains over simple per-tone estimation.
B. Hardware verification Our beamforming weight calculation solution is based on
In order to verify that the proposed algorithm will work using orthogonal iterations to track the dominant subspace
in practical systems it was trialled on a custom hardware test and we have showed that only one iteration per subcarrier
bed. This FPGA-based test bed was configured to generate is needed. This captures practically all the beamforming gains
IEEE 802.11n [9] transmissions with 2 transmit antennas, 2 and ensures that the beamformer is smooth across frequency
receive antennas, and 20MHz bandwidth at a carrier frequency as changes made between consecutive subcarriers are small.
of 5.24GHz. No oscillators or other synchronisation signals The smooth beamformer was also verified on a hardware test
were shared between the transmitter and receiver, therefore bed, showing its performance gains in practice.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:43 from IEEE Xplore. Restrictions apply.

VII. ACKNOWLEDGEMENTS [5] P. Strobach, “Low-rank adaptive filters," IEEE Trans. Signal Processing,
vol. 44, no. 12, pp. 2932-2947, Dec. 1996.
The authors would like to acknowledge the fruitful discus- [6] P. Comon and G. Golub, “Tracking a few extreme singular values and
sions with their colleagues at Toshiba Research Europe and vectors in signal processing," Proc. IEEE, vol. 78, no. 8, pp. 1327-1343,
the support of its directors. They would like to in particular Aug. 1990.
[7] O. Edfors, M. Sandell, J.-J. van de Beek, S. K. Wilson, and P. O. Bör-
thank Dr D. McNamara and Dr A. Lillie for their help with jesson, “OFDM channel estimation by singular value decomposition,"
the hardware verification. IEEE Trans. Commun., vol. 46, no. 7, pp. 931-939, July 1998.
[8] C. Davis and W. Kahan, “The rotation of eigenvectors by a perturbation
R EFERENCES III," SIAM J. Num. Anal., vol. 7, no. 1, pp. 1-46, Mar. 1970.
[9] IEEE 802.11n/D2.0, “Draft Amendment to Standard for Information
[1] I. Medvedev, R. Walton, J. Ketchum, S. Nanda, B. Bjerke, M. Wallace, Technology - Telecommunications and information exchange between
and S. Howard, “Transmission strategies for high throughput MIMO systems - Local and Metropolitan networks - Specific requirements - Part
OFDM communications," in Proc. Intern. Conf. Commun., vol. 4, Seoul, 11: Wireless LAN Medium Access Control (MAC) and Physical Layer
South Korea, pp. 2621-2625, May 2005. (PHY) specifications: Enhancements for Higher Throughput," IEEE Std
[2] T. Onizawa, M. Mizoguchi, T. Sakata, and M. Morikura, “A simple 802.11, Feb. 2007.
adaptive channel estimation scheme for OFDM systems," in Proc. IEEE [10] A. Stephens et al., “IEEE 802.11 TGn comparison criteria."
Veh. Technol. Conf. Fall, vol. 1, Amsterdam, The Netherlands, pp. 279- IEEE 802.11 - 03/814r30, May 2004. [Online] Available:
283, Sept. 1999.
[3] C. Shen and M. Fitz, “MIMO-OFDM beamforming for improved [11] V. Erceg et al., “TGn channel models.” IEEE 802.11 - 03/940r4, May
channel estimation," in Proc. Globecom, vol. 1, San Francisco, CA, 2004. [Online] Available:
pp. 1-5, Nov. 2006.
[4] G. Golub and C. van Loan, Matrix Computations. Johns Hopkins
University Press, 3rd ed., 1996.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:43 from IEEE Xplore. Restrictions apply.