IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 53, NO.

11, NOVEMBER 2005 4125

Interpolation Based Transmit Beamforming for

MIMO-OFDM With Limited Feedback
Jihoon Choi and Robert W. Heath, Jr., Member, IEEE

Abstract—Transmit beamforming and receive combining are well as additional array gain; thus, it can significantly improve
simple methods for exploiting spatial diversity in multiple-input system performance. This approach, however, requires knowl-
multiple-output orthogonal frequency division multiplexing edge of the transmit beamforming vector at the transmitter.
(MIMO-OFDM) system. Optimal beamforming requires channel
state information in the form of the beamforming vectors for When the uplink and downlink channels are not reciprocal (as
each OFDM subcarrier. This paper proposes a limited feedback in a frequency division duplexing system), the receiver informs
architecture that combines beamforming vector quantization and the transmitter about the desired transmit beamforming vector
smart vector interpolation. In the proposed system, the receiver through a feedback control channel.
sends a fraction of information about the optimal beamforming The beamforming techniques proposed for narrowband
vectors to the transmitter and the transmitter computes the beam-
forming vectors for all subcarriers through interpolation. A new channels can be easily extended to frequency selective channels
spherical interpolator is developed that exploits parameters for by employing orthogonal frequency division multiplexing
phase rotation to satisfy the phase invariance and unit norm prop- (OFDM). The combination of MIMO and OFDM, known as
erties of the transmitted beamforming vectors. The beamforming MIMO-OFDM, converts a broadband MIMO channel into a
vectors and phase parameters are quantized at the receiver and series of parallel narrowband MIMO channels, one for each
the quantized information is provided to the transmitter. The
proposed quantization system provides only a moderate increase OFDM subcarrier [12]–[14]. Transmit beamforming with
in complexity versus over comparable approaches. Numerical receive combining can be performed independently for each
simulations show that the proposed scheme performs better than subcarrier of MIMO-OFDM. In nonreciprocal channels, this
existing diversity techniques with the same feedback data rate. requires that the MIMO-OFDM receiver computes and sends
Index Terms—Antenna diversity, MIMO-OFDM, mutual infor- back to the transmitter the optimal beamforming vector for
mation, spherical interpolation, transmit beamforming. every active subcarrier. Practically, the feedback rate can be
managed by using limited feedback techniques where the
beamforming vectors are quantized using a beamforming
I. INTRODUCTION
codebook designed for narrowband MIMO channels (see, e.g.,

M ULTIPLE-INPUT multiple-output (MIMO) systems,

which use multiple antennas at both transmitter and
receiver, provide spatial diversity that can be used to miti-
[15]–[18]). Unfortunately, even with quantization, the feedback
requirements generally grow in proportion to the number of
active subcarriers.
gate signal-level fluctuations in fading channels [1]–[3]. In In an attempt to reduce the feedback, in this paper we con-
narrowband channels, MIMO systems can provide a diversity sider a new approach to transmit beamforming that combines
advantage in proportion to the product of the number of transmit quantized beamforming with interpolation of the beamforming
and receive antennas. When the channel is unknown to the vectors. The idea of our approach is to exploit the correlation
transmitter, diversity can be obtained by using space-time codes between the beamforming vectors at neighboring subcarriers.
[1]–[3]. When channel state information (CSI) is available at In OFDM, the length of the cyclic prefix (CP) is designed to be
the transmitter, however, diversity can be obtained using a much less than the number of subcarriers to increase spectral ef-
simple approach known as transmit beamforming and receive ficiency (see, e.g., [19] and [20]). Accordingly, the subchannels
combining [4]–[11]. Compared with space-time codes, beam- obtained from the discrete Fourier transform (DFT) of the sam-
forming and combining achieves the same diversity order as pled time-domain channel are significantly correlated. As a re-
sult, in MIMO-OFDM the beamforming vectors determined by
Manuscript received December 18, 2003; revised November 14, 2004. the subchannels are also substantially correlated. Using this fact,
This work was supported by the Texas Advanced Technology Program under we propose a new beamforming scheme that sends back only a
Grant 003658–0380-2003, the National Instruments Foundation, and the fraction of beamforming vectors to the transmitter and recon-
Post-Doctoral Fellowship Program of Korea Science and Engineering Foun-
dation (KOSEF). This work was completed while the first author was with the structs the beamforming vectors for all subcarriers through in-
University of Texas at Austin. The associate editor coordinating the review of terpolation at the transmitter. We assume that the transmit power
this manuscript and approving it for publication was Dr. Martin Haardt. is equally assigned to all subcarriers. Power allocation can be
J. Choi is with the Media Lab. 1, Telecommunication Network Busi-
ness, Samsung Electronics, Suwon, Korea 442-600 (e-mail: jihoon04.choi@ included in the system model with additional feedback but we
samsung.com). defer this to future work. Since the beamforming vectors have
R. W. Heath, Jr. is with Department of Electrical and Computer Engineering, unit norm by the transmit power constraint, they lie on the unit
Wireless Networking and Communications Group, The University of Texas at
Austin, Austin, TX 78712-0240 USA (e-mail: rheath@ece.utexas.edu). sphere. Consequently, it is natural to look toward spherical inter-
Digital Object Identifier 10.1109/TSP.2005.857019 polators to perform interpolation between beamforming vectors
1053-587X/$20.00 © 2005 IEEE

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
4126 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 53, NO. 11, NOVEMBER 2005

Fig. 1. Block diagram of a MIMO-OFDM system with M transmit antennas, M receive antennas, and N subcarriers. The feedback channel conveys a fraction
f     g
of information about w (k ); 1 i M ; 1 k N , which is the coefficient of the beamforming vector for subcarrier k on the ith antenna.

[21]–[25]. The nonuniqueness of optimal beamforming vectors for the proposed scheme are analyzed in Section IV. Through
[17], however, compromises the performance of standard spher- numerical simulations, the proposed beamforming method is
ical interpolators. compared with existing diversity techniques in terms of bit error
In this paper, we propose a new spherical interpolator based rate (BER) and mutual information in Section V. Finally, we pro-
on a modification of Watson’s interpolator [23]. To resolve the vide some conclusions in Section VI.
ambiguities in the optimal beamforming vector, our interpolator
employs a parameter for phase rotation, denoted as , which is II. SYSTEM OVERVIEW
chosen by optimizing a cost function based on the channel gain
A MIMO-OFDM system with transmit beamforming and
or the mutual information. Because the optimization is difficult
receive combining, using transmit antennas and receive
to solve in closed form, we solve for the phase rotation by per-
antennas, and subcarriers is illustrated in Fig. 1. At the trans-
forming either a numerical grid search or by solving a closed
mitter, the subcarrier modulates the symbol using
form solution with an approximate cost function.
the beamforming vector
To enable practical implementation with limited feedback,
and launches the resulting signal into the propagation envi-
we combine our new spherical interpolator with beamforming
ronment. Assuming that the sampled impulse response of the
vector quantization as proposed in [15]–[18]. With our ap-
channel is shorter than the CP, the channel for the th subcarrier
proach, the receiver quantizes the beamforming vectors for a
after the DFT can be described by a -by- channel matrix
fraction of the subcarriers by selecting the best element from
whose entries represent the complex channel gains expe-
the codebook. Using the quantized vectors, the receiver then
rienced by subcarrier . After processing with the combining
selects a near-optimal quantized phase for each quantized
vector , the combined
vector based on the phase cost function. After being informed
signal at subcarrier can be expressed as
about the quantized beamforming vectors and phases, the trans-
mitter uses the proposed spherical interpolator to reconstruct
(1)
the beamforming vectors for all the subcarriers. We analyze the
computational complexity and feedback requirements for the where is the -dimensional noise vector whose entries
proposed method and compare with several related diversity have the independent and identically distributed (i.i.d.) complex
techniques. Computer simulations show that the proposed Gaussian distribution with zero mean and variance . We as-
scheme outperforms existing diversity techniques with compa- sume that the power is allocated equally across all subcarriers;
rable complexity and feedback bits and performs close to the thus, is a constant, and , where
ideal beamforming with full feedback information. Compared means the 2-norm of to maintain the overall power
with prior work that deals with unquantized partial channel constraint.
information [26]–[29], our approach focuses on methods for Since the signal model in (1) is identical to that of a narrow-
reducing feedback requirements by subsampling in frequency band MIMO system, and can be chosen to maximize
and quantizing separately the optimal beamforming vectors the signal to noise ratio (SNR) using the results on beamforming
computed from the instantaneous channel state (versus the first and combining for narrowband MIMO systems [4], [6]. Without
or second order statistics of the channel) at each subcarrier. loss of generality we can fix . Then, the SNR for
This paper is organized as follows. Section II reviews MIMO- subcarrier can be written as
OFDM communication with beamforming and combining. The
proposed interpolation based beamforming scheme is presented
in Section III, while the complexity and feedback requirements (2)

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
CHOI AND HEATH: INTERPOLATION-BASED TRANSMIT BEAMFORMING FOR MIMO-OFDM 4127

where is the effective channel Section III that the beamforming vectors are correlated; thus,
gain. The SNR maximizing solution chooses the transmit the amount of feedback data is proportional to only
beamforming vector and receive combining vector (where is normally much less than ) by sending back only
based on to maximize the SNR of the subcarrier in (2). a fraction of vectors combined with intelligent interpolation
For any , under the unit norm constraint on , it is at the transmitter. We combine this result with beamforming
possible to show that the SNR maximizing solution uses max- vector quantization [15]–[17] to reduce the feedback required
imum ratio combining (MRC) with and provide a specific recipe that can be followed for practical
implementation in MIMO-OFDM systems.
(3)
III. INTERPOLATION BASED BEAMFORMING
Since the combining is performed digitally, it is reasonable FOR MIMO-OFDM
to use optimum MRC processing instead of a suboptimum In this section, we present the motivation and detailed de-
technique such as antenna selection. Choosing , however, scription of our approach for interpolation based beamforming
depends in part on how much CSI is available at the transmitter. in MIMO-OFDM systems. First, we show that the correlation
For example, selection diversity transmission chooses between the beamforming vectors is similar to the subchannel
from the set of the columns of the -by- identity matrix correlation with one receive antenna. Next, we use these results
thus requiring only bits of information of the CSI to motivate subsampling the beamforming vectors in frequency
per subcarrier at the transmitter. Maximum ration transmission followed by beamformer interpolation across subcarriers. We
(MRT), on the other hand, chooses as the right singular introduce a beamforming vector interpolator based on spherical
vector of corresponding to the largest singular vector linear interpolator with additional parameters for phase rotation,
of [6], [8] for every . In Rayleigh fading channels, which takes the invariance of the optimal beamforming vectors
both approaches achieve full diversity but MRT also obtains into account. We optimize the phase rotation parameters in the
the maximum amount of array gain. Unfortunately, MRTs sense of maximizing the minimum effective channel gain or mu-
improved performance comes at the expense of more extensive tual information. Finally, the proposed beamforming method is
CSI at the transmitter in the form of the complete channel combined with quantization of beamforming vectors for prac-
impulse responses or equivalently the optimum beamforming tical implementation.
vectors .
In this paper, we consider a burst-mode communication A. Correlation Between Beamforming Vectors
system where data is transmitted in frames that consist of
For notational convenience, we define the -dimen-
multiple MIMO-OFDM symbols from the system described
sional vector representing the channel gains for subcarrier as
in Fig. 1. We assume that the channel is fixed in a frame
but varies randomly between frames (commonly known as vec (4)
the block fading model). Further, we assume that full CSI is
not available to the transmitter, but there exists a low-rate, where vec is the vector obtained by stacking the columns of
error-free, zero-delay feedback link from the receiver back to one on top of another, and . Generally, the corre-
the transmitter. To accommodate the low-rate requirement, we lation between subchannel and subchannel (assuming
assume that the number of feedback bits per frame is limited modulo addition) is defined as
to a constant. The receiver uses CSI (assumed perfect in this
paper) obtained from a training phase at the beginning of the (5)
frame to determine the feedback bits corresponding to the set of
beamforming vectors for all OFDM
symbols in the frame. Similarly, the correlation between and can be
To respect the feedback constraints, in this paper we de- expressed as
velop a quantization strategy for transmit beamforming in
(6)
MIMO-OFDM communication systems. A conventional vector
quantization approach would be to quantize the impulse re- The definition in (6), however, is not a good correlation indi-
sponse of the channel in the time domain. The amount of cator for beamforming vectors due to the nonuniqueness of the
feedback, though, can be substantial since for a MIMO impulse optimal beamformers, which we discuss more in Section III-B.
response with length , the number of coefficients to be quan- Briefly, when is the optimal beamforming vector for the
tized is . Additionally, a high-resolution quantization subcarrier , is also optimal from (2). In practice, the
is required because we have to take the DFT and then the arbitrary phase is determined by the algorithm that computes
eigendecomposition of the quantized channel. In our case, we the beamforming vector. The phase ambiguity means that
are primarily interested in the dominant singular vector of each can take a different value for the same channel because the ex-
; thus, we do not need to inform the transmitter about pectation of (6) is dependent on the arbitrary phase . For this
the complete channel. Consequently, we develop a method that reason, we measure correlation using a different definition of
allows us to inform the transmitter about the desired transmit the beamformer correlation which is independent of :
beamforming vectors , within the
constraints of the limited feedback channel. We argue in (7)

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
4128 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 53, NO. 11, NOVEMBER 2005

To make our comparison we also consider a similar modification

to the subchannel correlation:

(8)

Now, we compare the subchannel correlation with the beam-

former correlation through computer simulations. For illustra-
tion purposes we generated the time domain channel impulse
response following the ETSI/BRAN Channel Models A, B, and
C [30]. Note that the root mean square (RMS) delay spread of
Channel Models A, B, and C is 50, 100, and 150 ns, respec-
tively. It was assumed that channel impulse responses between
different transmit and receive antenna pairs are independent.
Fig. 2 shows when , , and .
For comparison, given by (8) was plotted for ,
, and . The amount of correlation for Fig. 2. Comparison between  (d) with one receive antenna and  (d) with
and was the greatest for Model A because the coherence two receive antennas when M = 4, and N = 64. This figure shows the
bandwidth of the channel is inversely proportional to the RMS similarity between  (d) and  (d).
delay spread.
Of particular interest is that the correlation between the beam- the missing subcarriers. Conventional interpolators, however,
forming vectors with two receive antennas is very similar to the are not suitable for beamformer interpolation because they do
correlation between the channel vectors with one receive an- not preserve the unit norm constraint on . An alternative
tenna. This is a trend we also observed for other combinations approach is to use spherical interpolation algorithms (see, e.g.,
of receive antennas and transmit antennas. One explanation is [21]–[23]) that naturally enforce the unit norm constraint by
given in [17], where it was shown that the distribution of the construction. One such algorithm, which is introduced in [23],
optimal beamforming vector in a narrowband independent iden- is to weight average of points on a sphere in Euclidean space
tically distributed complex Gaussian matrix channel is indepen- and then renormalize the result back to the unit sphere. For ex-
dent of the number of receive antennas. Based on these observa- ample, a th-order interpolator computes
tions, we will use the coherence bandwidth predicted by
with one receive antenna to estimate the coherence of the beam-
forming vectors. (9)

B. Proposed Interpolation Based Beamforming Method where and are weights values such that each
A simple method for exploiting the correlation of neighboring and . When , the interpolated vector
beamforming vectors is to combine adjacent subcarriers into a is a point on the spherical line between points and .
cluster and use the beamforming vector corresponding to the Unfortunately, the algorithm in (9) (and, generally, the
center subcarrier in the cluster. This method will be referred to class of spherical interpolation algorithms) cannot be applied
as clustering.1 If we combine subcarriers into one cluster, to our approach due to the lack of uniqueness of the op-
the amount of feedback is reduced by . As the cluster size timal beamforming vector as we discussed in Section III-A.
increases, however, the performance of beamforming de- To emphasize this point, recall the effective channel gain
grades because of the distortion experienced by subcarriers near in (2). When is the optimal
the cluster boundary. This limits the cluster size and the beamforming vector maximizing the effective channel gain,
consequent reduction of feedback information is restricted. As also maximizes the effective channel gain. In other
an alternative, we consider an interpolation-based beamforming words, the optimal beamforming vector is not a unique point
method where we send back a subsampling of the beamforming but represents a complex line. As a consequence, algorithms
vectors and use a smart interpolation at the transmitter to fill in that are used to compute the optimal beamforming vector, e.g.,
the gaps. the singular value decomposition (SVD), typically choose the
Suppose that is divided by the subsampling rate . The phase to force the first coefficient of each to be real. The
receiver evaluates the optimal beamforming vectors for the se- phase, however, has a dramatic impact on the resulting inter-
lected subcarriers and sends the subsampled beamforming vec- polated vectors as calculated through (9) as well as by other
tors to the transmitter. The spherical interpolators in [21]–[23]. This makes it difficult to
transmitter reconstructs the beamforming vectors for all subcar- directly apply spherical interpolation to perform beamformer
riers by interpolating for interpolation.
To solve this problem, we propose a new algorithm through
1We derive this terminology based on work in [31] where clustering was used modification of the interpolater in (9). For notational conve-
to reduce the required feedback for an OFDM system with adaptive modulation. nience, let us define . Given

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
CHOI AND HEATH: INTERPOLATION-BASED TRANSMIT BEAMFORMING FOR MIMO-OFDM 4129

, the proposed interpolator computes for subcarrier minimum channel gain can happen at any subcarrier in the in-
terpolation interval. Yet we assume that the subcarrier
has the worst effective channel gain to simplify the
(10) cost function in (11). From this assumption (11) is approximated
to
where is the linear weight value, ,
, and is a parameter for phase rotation. Note
that are obtained
by and (equivalently ). (13)
While the spherical interpolator in (9) with only utilizes
and , the proposed interpolator evaluates the beam-
where . Note that this
forming vector from and . Essentially, the role of
cost function does not guarantee that we maximize the minimum
is to remove the distortion caused by the arbitrary phase rotation
channel gain over all subcarriers and, thus, is suboptimal relative
of the optimal beamforming vectors. Notice that the parameters
to (11). Our simulations, though, show that the cost function in
are conveyed to the transmitter along
(13) performs comparably to the grid search in (12).
with the selected beamforming vectors .
By differentiating the cost function in (13) with respect to ,
C. Phase Optimization for the Proposed Interpolator the optimal solution satisfies

To maximize the performance of the proposed interpolator, Imag

the receiver evaluates the optimal phases (14)
based on a communication-theoretic performance metric. where
Ideally the phase parameter would be determined to minimize ,
the average BER performance of the proposed interpolator. This , and Imag is the imaginary part of . By Theorem
criterion, however, makes it complicated to define the cost func- 1 in the Appendix, Imag ; thus, there always exists
tion with respect to especially when modulation, coding, and that satisfies (14). Suppose that and
interleaving are taken into account. Instead, we propose to find let denote . Then the solutions for (14) are given
by maximizing the minimum of the effective channel gains by
for all subcarriers. We find through simulations that this crite- (15)
rion tends to decrease the average BER. For capacity compar-
if
isons, we find the optimal in the sense of maximizing the (16)
if
mutual information. In both cases, we find the optimal solution
through a numerical grid search and/or mathematical analysis where . Now, the derivative of
after some approximations of the cost function. (14) is denoted as
In (10), is only used for computing
, and thus, the optimal maximizing the min- Real
imum effective channel gain can be found as (17)
where Real is the real part of . Since (17) decides the sign
(11)
of the second derivative of the cost function, the solution max-
imizing the cost function becomes one of the solutions and
where . Due to the normalization factor
such that Real is negative. This solution will be
in , it is difficult to solve for (11) analytically.
referred to as the closed-form solution.
Instead, we use a numerical grid search by uniformly quantizing
Alternatively, can be determined to maximize the sum rate
the phase and modifying (11) as
of all the OFDM subcarriers. This is useful for information theo-
(12) retic comparisons. When waterfilling is considered, the optimal
phase maximizing the sum rate is found by
where , and
is the number of quantized levels which determines the perfor-
mance and complexity of the search.
As an alternative to the grid search, we simplify (11) through (18)
an approximation and finally derive a closed-form solution. The where is the transmit power allocated to the subcarrier
advantage of a closed form solution is that we can use it to re- and by the power constraint. Since the
duce search time or search complexity, especially for large . optimal determined by waterfilling across subcar-
On average, suffers from the largest dis- riers is a function of , it is complicated to find the optimal .
tortion induced by interpolation, because the subcarrier For example, the numerical grid search computes the effective
is the furthest from the subcarriers and channel gain and evaluates through waterfilling for each
, whose beamforming vectors are available through candidate of , thereby requiring an extensive computational
feedback. Since the instantaneous channel gain is a function of cost. Also, waterfilling requires additional feedback information
the channel matrix and the beamforming vector, however, the about power allocation among subcarriers. To overcome these

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
4130 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 53, NO. 11, NOVEMBER 2005

difficulties, we assume that the transmit power is equally as- IV. COMPUTATIONAL COMPLEXITY AND
signed to all subcarriers. When the grid search is used, (18) is FEEDBACK REQUIREMENTS
simplified to This section compares the complexities and feedback require-
ments for implementing selection diversity for each subcarrier,
clustering, the proposed grid search with beamformer quantiza-
tion, and ideal beamforming with full feedback of the quantized
(19) beamforming vectors.
Following the approach in (13), can be approximately found In the complexity analysis, we roughly calculate the number
by solving (20), shown at the bottom of the page. From the con- of complex multiplications and the number of complex addi-
cavity of the log function, this equation is equivalent to the so- tions. The computation at the transmitter includes the inverse
lution to (13). DFT, transmit beamforming, and beamformer interpolation
for the proposed algorithm. The computation at the receiver
D. Interpolation Combined With Beamformer Quantization includes the DFT, receive combining, quantization of beam-
For practical implementation, we propose to quantize the forming vectors, and determination of the phase rotation
beamforming vectors using the codebooks designed for nar- parameters for the proposed algorithm. For simplicity, the
rowband MIMO systems [16]–[18]. These codebooks are computational load for synchronization and channel estimation
specially designed to take into account the phase invariance of is not considered as it is likely to be comparable for all ap-
the optimal beamforming vector. Given , the quantized proaches. The DFT or inverse DFT can be implemented using
beamforming vectors are determined by selecting the element fast Fourier transform (FFT) algorithms such as the power
of the codebook that maximizes the effective channel gain as of two decimation-in-time and decimation-in-frequency algo-
follows: rithms, which require complex multiplications
and complex additions [34]. The computational
(21) load for MIMO-OFDM is multiplied by and at the
transmitter and the receiver, respectively. When quadrature
where is the codebook including all possible beamforming amplitude modulation (QAM) is used, the complexity for
vectors, and . transmit beamforming is almost negligible. Receive com-
The quantized beamforming vectors are used to find the op- bining needs complex multiplications and additions,
timal and is uniformly quantized on . As an aside, respectively. The beamforming vectors are quantized at the
instead of separately quantizing and , the receiver by (21) performing multiplica-
whole vector can be quantized by spherical tions and additions per subcarrier, where
quantization based on the Euclidean distance [32], [33]. This ap- is the cardinality of the chosen beamforming codebook. In
proach, however, requires almost the same amount of feedback the proposed algorithm, beamformer interpolation requires
as with separate quantization, yet makes it more complicated to complex multiplications and complex additions.
determine the phase . Therefore, we only consider the sepa- The phase rotation parameter is determined by (22), which
rate quantization of and . requires complex
When quantized beamforming vectors are used, the interpo- multiplications and additions.
lated vector is evaluated by replacing The complexity calculations at the transmitter and receiver
with in (10). Since we are quantizing , it is nat- for various diversity techniques are summarized in Table I.
ural to use the grid search to find the best . By modifying (12) Selection diversity necessitates the least operations because it
and (19), we obtain the following equations that optimize for does not require the quantization of beamforming vectors. Ideal
the quantized beamforming vectors in the sense of maximizing beamforming quantizes the beamforming vectors for all subcar-
the minimum effective channel gain riers; thus, it needs extensive computational load. Compared to
clustering, the proposed method requires additional operations
(22) to interpolate beamforming vectors at the transmitter and find
the phase rotation parameters at the receiver: In Table II, the
or maximizing the mutual information number of complex multiplications and the number of additions
are increased by 35.3% and 29.6%, respectively.
Table III compares the required number of feedback bits
per frame for various diversity techniques. The information
(23) required for antenna selection is bits per subcarrier.

(20)

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
CHOI AND HEATH: INTERPOLATION-BASED TRANSMIT BEAMFORMING FOR MIMO-OFDM 4131

TABLE I TABLE III

jWj
(a) REQUIRED NUMBER OF COMPLEX MULTIPLICATIONS FOR BEAMFORMING IN FEEDBACK REQUIREMENTS OF VARIOUS DIVERSITY TECHNIQUES FOR
MIMO-OFDM. (b) REQUIRED NUMBER OF COMPLEX ADDITIONS MIMO-OFDM WHEN M N K
= 4, = 64, = 8, = 4, AND P = 16
FOR BEAMFORMING IN MIMO-OFDM

V. SIMULATION RESULTS

To illustrate the performance of the proposed approach, we

present Monte Carlo simulations for a system with parameters:
, , , , and CP length of 16. We
assumed that the discrete-time channel impulse response was
generated according to the ETSI/BRAN Channel Model B in
[30]; the channels between different transmit and receiver an-
tenna pairs were independent; the channel was fixed for a frame
and randomly varied between frames in Figs. 3–7; was
i.i.d. complex Gaussian with zero mean; the transmit power was
equally allocated to all subcarriers; QPSK was used for the BER
simulations; and the receiver used MRC with perfect channel
knowledge. In all simulations except Fig. 6, it was assumed that
the feedback channel had no delay and no transmission error.
Every point of the simulation results was obtained by averaging
over more than 3000 independent realizations of the channel and
the noise.
Simulation results are presented in the following two subsec-
tions. In the first part, we simulate the transmit beamforming
schemes without quantization to provide an upper bound on the
TABLE II
= 4, M performance with quantization. Next, we consider the quantiza-
jWj
REQUIRED NUMBER OF COMPLEX OPERATIONS WHEN
M N K
= 2, = 64, = 8, = 4, AND P = 16 tion of beamforming vectors and , shown in Section III-D.

A. Unquantized Beamforming
We compare the proposed beamforming scheme with clus-
tering and ideal beamforming with feedback of all beamforming
vectors. The dominant right singular vector was used for the
beamforming vector for MRT/MRC and quantization was not
considered.
Fig. 3 compares the uncoded BER performance for various
beamforming methods. In the figure, “grid search” and “closed-
When the transmit antenna with the maximum channel gain is form solution” describe the proposed interpolators using the
selected independently for each subcarrier, the total feedback grid search in (12) and the closed-form solution in (13), respec-
information becomes bits. The beamforming vec- tively. The grid searches with and both per-
tors are quantized by using the codebook , thereby requiring formed comparably to the closed-form solution. The proposed
feedback bits per beamforming vector. The proposed methods outperformed clustering, yet exhibited slight diversity
algorithm needs additional bits of feedback to loss compared to the ideal beamforming.
transfer the information about the phase rotation parameters. Fig. 4 presents the mutual information for the whole OFDM
Note that the proposed grid search requires much less feedback symbol obtained using various beamforming techniques. Be-
than selection diversity and ideal beamforming. For example, in cause we only consider beamforming, the gains in sum rate are
Table III, the proposed method needs only 48 feedback bits per much smaller than if used multiple eigenvectors of the channel.
frame, which is about one third of feedback for selection diver- For comparison, we plotted the ideal beamforming and the pro-
sity and about one fifth of feedback for the ideal beamforming. posed grid search with waterfilling across the subcarriers while

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
4132 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 53, NO. 11, NOVEMBER 2005

Fig. 5. Uncoded BER performance with quantization of beamforming vectors

when M = 4, M N K
= 2, = 64, and = 8. Channel was fixed in a frame
Fig. 3. Uncoded BER performance without quantization of beamforming
M M N K
and the grid search in (22) was used for the proposed interpolator.
vectors when = 4, = 2, = 64, and = 8. Grid search in (12)
and closed-form solution in (13) were used.
B. Quantized Beamforming
The beamforming vectors were quantized by using the code-
book designed according to (22) in [17]. We used the code-
book with that requires 4 bits of feedback per beam-
forming vector. was uniformly quantized by the grid search
with . For comparison, we simulated antenna selection
diversity [7], clustering, and ideal beamforming. Also, the the-
oretical bound of orthogonal space-time block codes (OSTBC)2
[2] was considered. In selection diversity, the transmit antenna
was selected independently for each subcarrier. The feedback
requirements for the diversity techniques under consideration
are provided in Table III.
Fig. 5 shows the uncoded BER performance when the
beamforming vectors were quantized. The proposed method
outperformed clustering and the theoretical lower bound of
OSTBC, yet exhibited slight loss in diversity order compared
to the ideal beamforming. Selection diversity, as expected,
achieves full diversity order but exhibits some array gain loss.
Thus, the proposed method performs comparably to selection
diversity. Note that the proposed scheme requires only 48
Fig. 4. Mutual information without quantization of beamforming vectors feedback bits per frame, which is 3/8 of the feedback required
when M = 4, M = 2, N = 64, and K
= 8. Grid search in (19) and for selection diversity.
closed-form solution in (20) were used. To study the effects of channel estimation error and/or feed-
back delay, we employ the first order autoregressive model given
all other curves were obtained without waterfilling. The pro- by
posed techniques had higher mutual information than the clus-
tering and performed within 0.3 dB of the ideal beamforming (24)
without waterfilling. Since the diversity provided by multiple
antennas reduces channel fluctuation between subcarriers, the where is the predicted channel at the transmitter,
mutual information gain by waterfilling was small. Notice that is the th element of , is i.i.d. com-
waterfilling requires additional feedback information for power plex Gaussian noise with zero mean and unit variance, and
assignment among subcarriers.
Figs. 3 and 4 show that the grid search with pro-
vides a good tradeoff between the performance and complexity.
Thus, the grid search with will be used in the following 2It was shown that there are no complex OSTBC with full rate when M  3,
simulations. but the theoretical performance bound can be obtained for any M .

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
CHOI AND HEATH: INTERPOLATION-BASED TRANSMIT BEAMFORMING FOR MIMO-OFDM 4133

Fig. 6. Uncoded BER performance with quantization of beamforming vectors Fig. 7. Mutual information with quantization of beamforming vectors when
when E =N = 7 dB, M = 4, M = 2, N = 64, and K = 8. Channel M = 4, M = 2, N = 64, and K = 8. The grid search in (23) was used for
prediction error was considered and the proposed interpolator is the same in the proposed interpolator.
Fig. 5.

. Then, the mean square error (MSE) between

and is denoted as

(25)

Fig. 6 shows the BER degradation caused by the channel pre-

diction error at dB. Similar to Fig. 5, the proposed
method performed better than clustering and selection diversity.
Note that the proposed beamforming method is better than the
theoretical bound of OSTBC as long as the MSE dB.
This gives an indication when estimation error and/or delay are
significant enough to mitigate the advantages of closed-loop di-
versity techniques over open-loop techniques.
Fig. 7 compares the mutual information when quantized
beamforming vectors are used. As before, the proposed method
outperformed clustering, selection diversity, and the theoretical Fig. 8. Uncoded BER performance with quantization of beamforming vectors
when E =N = 8 dB, M = 4, M = 2, N = 64, and K = 8. Channel
upper bound of OSTBC. In addition, the proposed performed was time-varying within a frame, and beamforming vectors were periodically
very close to the ideal beamforming: the SNR loss was less updated through feedback. The proposed interpolator is the same in Fig. 5.
than 0.2 dB.
The uncoded BER performance in time-varying channels enough, there is no good reason to use closed-loop diversity
is shown for various Doppler frequencies in Fig. 8, where modes.
denotes the Doppler frequency normalized by the OFDM The coded BER values in time-varying channels with
symbol duration . In the figure, it was assumed that all the are presented in Fig. 9. For channel coding, we used a
beamforming methods have the same feedback ratio over data convolutional code with generator polynomials and
bits. Taking into account the feedback requirements shown with coding rate 1/2, along with the interleaver and
in Table II, the frame length was set as 40, 60, 160, and 320 deinterleaver defined in [20],3 and soft Viterbi decoding. The
symbols for clustering, the proposed, selection diversity, and frame length was the same as that in Fig. 8. Each OFDM symbol
the ideal beamforming, respectively. The proposed method had transmitted 64 bits, which is a half of the uncoded case, and
the lowest BER when . The the feedback bits are just 1.25% of data bits for all cases. The
BER performance of ideal beamforming increases rapidly with proposed scheme outperformed other diversity techniques and
because it has the longest framelength and, thus, largest exhibited a 2.5-dB gain over selection diversity and a 2.8-dB
beamformer update period. OSTBC performed best as long 3In [20], only 48 subcarriers among 64 are used for data transmission. To use
as since its performance is irrelative to all the subcarriers for our simulation, the index of the interleaver and deinter-
. This confirms the intuition that when the Doppler is large leaver was extended.

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
4134 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 53, NO. 11, NOVEMBER 2005

combine spatial multiplexing with special per-tone unitary

precoding matrices along the lines of [35]. An extension of
our proposed approach is to evaluate the precoding matrices
for the subcarriers through interpolation. While the proposed
interpolator for transmit beamforming could be used, it does
not retain the orthogonality between the column vectors of
the interpolated matrix. Developing an interpolator for spatial
multiplexing remains a future research topic.

APPENDIX
Theorem 1: For any -dimensional complex vectors , ,
, , let us define and as

(26)
(27)

Then, and satisfy . In addition, let us define

Fig. 9. BER performance with quantization of beamforming vectors when , , ,
M = 4, M = 2, N = 64, and K = 8. Channel was time-varying within a
frame with f T = 0:0005. A convolutional code with rate 1/2 was used along and . From (14), we get and Imag .
with interleaving and the proposed interpolator is the same in Fig. 7. Therefore, Theorem 2 implies that Imag .
Proof: Let us consider
gain over the ideal beamforming at BER . Note that the
proposed scheme had huge performance gain over the OSTBC.
Overall, these results demonstrate that the proposed beam-
forming vector quantization and interpolation scheme is advan-
tageous in terms of BER and mutual information compared with
existing closed-loop diversity techniques. (28)

VI. CONCLUSIONS For any non-negative real values and , we have .

Thus
In an attempt to reduce the feedback information required
for closed-loop MIMO-OFDM, a new transmit beamforming
scheme with limited feedback and beamformer interpolation
was proposed. In the proposed scheme, the receiver sends back
only a fraction of information about the beamforming vectors
along with additional parameters for phase rotation used in the
interpolation. The transmitter calculates the beamforming vec-
tors for all the subcarriers using a new spherical interpolator.
It was demonstrated that the proposed interpolator can be opti-
mized by adjusting the parameters for phase rotation depending
on different error rate or sum rate inspired criteria. Through
computer simulations, it was shown that the proposed method
has considerable performance benefits with modest feedback
and complexity requirements compared with other closed-loop
diversity techniques. (29)
For the beamformer interpolation, we only considered the
From (29), we get .
spherical linear interpolator. In fact, it is apparent that higher
order interpolation can decrease the distortion of the beam-
REFERENCES
forming vectors; however, the determination of becomes
more complicated because the higher order interpolation re- [1] V. Tarokh, H. Jafarkhani, and A. R. Calderbank, “Space-time codes for
high data rate wireless communication: performance criterion and code
quires joint optimization of . Therefore, construction,” IEEE Trans. Inf. Theory, vol. 44, no. 2, pp. 744–765, Mar.
the use of a higher order interpolation requires a practical 1998.
method for finding the optimal phases with reasonable [2] , “Space-time block codes from orthogonal designs,” IEEE Trans.
Inf. Theory, vol. 45, no. 5, pp. 1456–1467, Jul. 1999.
complexity. [3] S. M. Alamouti, “A simple transmit diversity technique for wireless
Our work proposed herein is limited to a MIMO system communications,” IEEE J. Sel. Areas Commun., vol. 16, no. 10, pp.
that transmits a single data stream per subcarrier. The capacity 1451–1458, Oct. 1998.
[4] J. B. Andersen, “Antenna arrays in mobile communications: gain, diver-
of MIMO-OFDM, however, can be increased by supporting sity, and channel capacity,” IEEE Antennas Propag. Mag., vol. 42, no.
the transmission of multiple data streams. One solution is to 2, pp. 12–16, Apr. 2000.

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.
CHOI AND HEATH: INTERPOLATION-BASED TRANSMIT BEAMFORMING FOR MIMO-OFDM 4135

[5] T. K. Y. Lo, “Maximum ratio transmission,” IEEE Trans. Commun., vol. [28] P. Xia, S. Zhou, and G. B. Giannakis, “Adaptive MIMO-OFDM based
47, no. 10, pp. 1458–1461, Oct. 1999. on partial channel state information,” IEEE Trans. Signal Process., vol.
[6] P. A. Dighe, R. K. Mallik, and S. S. Jamuar, “Analysis of transmit-re- 52, no. 1, pp. 202–213, Jan. 2004.
ceive diversity in Rayleigh fading,” IEEE Trans. Commun., vol. 51, no. [29] S. Zhou and G. B. Giannakis, “Adaptive modulation for multi-antenna
4, pp. 694–703, Apr. 2003. transmissions with channel mean feedback,” IEEE Trans. Wireless
[7] S. Thoen, L. V. der Perre, B. Gyselinckx, and M. Engels, “Perfor- Commun., to be published.
mance analysis of combined transmit-SC/receive-MRC,” IEEE Trans. [30] J. Medbo and P. Schramm, Channel models for HIPERLAN/2 in dif-
Commun., vol. 49, no. 1, pp. 5–8, Jan. 2001. ferent indoor scenarios, in ETSI/BRAN 3ERI085B, Mar. 1998.
[8] C.-H. Tse, K.-W. Yip, and T.-S. Ng, “Performance tradeoffs between [31] R. Grünheid, E. Bolinth, and H. Rohling, “A blockwise loading algo-
maximum ratio transmission and switched-transmit diversity,” in Proc. rithm for the adaptive modulation technique in OFDM systems,” in Proc.
IEEE PIMRC, vol. 2, Sept. 2000, pp. 1485–1489. IEEE Veh. Technol. Conf. Fall, vol. 2, Oct. 2001, pp. 948–951.
[9] R. W. Heath Jr. and A. Paulraj, “A simple scheme for transmit diversity [32] J. Hamkins and K. Zeger, “Asymptotically dense spherical codes—Part
using partial channel feedback,” in Proc. IEEE Asilomar Conf. Signals, I: Wrapped spherical codes,” IEEE Trans. Inf. Theory, vol. 43, no. 6, pp.
Syst., Comput., vol. 2, Nov. 1998, pp. 1073–1078. 1774–1785, Nov. 1997.
[10] M. Kang and M.-S. Alouini, “Largest eigenvalue of complex wishart [33] , “Asymptotically dense spherical codes—Part II: Laminated spher-
matrices and performance analysis of MIMO MRC systems,” IEEE J. ical codes,” IEEE Trans. Inf. Theory, vol. 43, no. 6, pp. 1786–1798, Nov.
Sel. Areas Commun., vol. 21, no. 4, pp. 418–426, Apr. 2003. 1997.
[11] H. Shi, M. Katayama, T. Yamazato, H. Okada, and A. Ogawa, “An adap- [34] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal
tive antenna selection scheme for transmit diversity in OFDM systems,” Processing, Second ed. Englewood Cliffs, NJ: Prentice-Hall, 1999.
in Proc. IEEE Veh. Technol. Conf. Fall, vol. 4, Oct. 2001, pp. 2168–2172. [35] D. J. Love and R. W. Heath Jr., “Limited feedback unitary precoding
[12] G. G. Raleigh and J. M. Cioffi, “Spatio-temporal coding for wireless for spatial multiplexing,” IEEE Trans. Inf. Theory, vol. 51, no. 8, pp.
communication,” IEEE J. Sel. Areas Commun., vol. 46, no. 3, pp. 2967–2976, Aug. 2005.
357–366, Mar. 1998.
[13] H. Bölcskei, D. Gesbert, and A. J. Paulraj, “On the capacity of OFDM-
based spatial multplexing systems,” IEEE Trans. Commun., vol. 50, no.
2, pp. 225–234, Feb. 2002.
[14] H. Bölcskei, M. Borgmann, and A. J. Paulraj, “Impact of the propaga- Jihoon Choi received the B.S., M.S., and Ph.D. de-
tion environment on the performance of space-frequency coded MIMO- grees from the Korea Advanced Institute of Science
OFDM,” IEEE J Sel. Areas Commun., vol. 21, no. 4, pp. 427–439, Apr. and Technology (KAIST), Daejon, Korea, in 1997,
2003. 1999, and 2003, respectively.
[15] A. Narula, M. J. Lopez, M. D. Trott, and G. W. Wornell, “Efficient use From 2003 to 2004, he was with the Department of
of side information in multiple-antenna data transmission over fading Electrical and Computer Engineering, the University
channels,” IEEE J. Sel. Areas Commun., vol. 16, no. 8, pp. 1423–1436, of Texas at Austin, where he performed research on
Oct. 1998. MIMO-OFDM systems with limited feedback. Since
[16] D. J. Love and R. W. Heath Jr., “Equal gain transmission in multi-input 2004, he has been with the Samsung Electronics,
multi-output wireless systems,” IEEE Trans. Commun., vol. 51, no. 7, Suwon, Korea. With Samsung, he has participated
pp. 1102–1110, Jul. 2003. in Portable Internet (PI) and Enhanced Broadcast
[17] D. J. Love, R. W. Heath Jr., and T. Strohmer, “Grassmannian beam- Multicast Service (BCMCS) projects. His research interests include MIMO,
forming for multiple-input multiple-output wireless systems,” IEEE multicarrier modulation, equalization, and space-time coding for wireless
Trans. Inf. Theory, vol. 49, no. 10, pp. 2735–2747, Oct. 2003. communications.
[18] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, “On beam-
forming with finite rate feedback in multiple-antenna systems,” IEEE
Trans. Inf. Theory, vol. 49, no. 10, pp. 2562–2579, Oct. 2003.
[19] Digital Broadcasting Systems for Television; Framing Structure,
Channel Coding, and Modulation for Digital Terrestrial Services, ETSI Robert W. Heath, Jr. (S’96–M’01) received the B.S.
DVB Stand. ETS 300 7XX, Geneva, Switzerland, 1996. and M.S. degrees from the University of Virginia,
[20] Part 11: Wireless LAN Medium Access Control (MAC) and Physical Charlottesville, in 1996 and 1997, respectively, and
Layer (PHY) Specifications-High-Speed Physical Layer in the 5 GHz the Ph.D. degree from Stanford University, Stanford,
Band, IEEE Stand. 802.11a, LAN/MAN Standards Committee, 1999. CA, in 2002, all in electrical engineering.
[21] S. R. Buss and J. P. Fillmore, “Spherical averages and applications to From 1998 to 1999, he was a Senior Member of the
spherical splines and interpolation,” ACM Trans. Graphics, vol. 20, no. Technical Staff at Iospan Wireless Inc., San Jose, CA,
2, pp. 95–126, Apr. 2001. where he played a key role in the design and imple-
[22] K. Shoemake, “Animating rotation with quaternion curves,” in Proc. mentation of the physical and link layers of the first
ACM SIGGRAPH, vol. 19, San Franscisco, CA, Jul. 1985, pp. 245–254. commercial MIMO-OFDM communication system.
[23] G. S. Watson, Statistics on Spheres. New York: Wiley, 1983. From 1999 to 2001, he served as a Senior Consultant
[24] G. Wagner, “On means of distances on the surface of a sphere (lower for Iospan Wireless Inc. In 2003, he founded MIMO Wireless Inc., a consulting
bounds),” Pacific J. Math., vol. 144, no. 2, pp. 389–398, 1990. company dedicated to the advancement of MIMO technology. Since January
[25] , “On means of distances on the surface of a sphere. II (upper 2002, he has been with the Department of Electrical and Computer Engineering,
bounds),” Pacific J. Math., vol. 154, no. 2, pp. 381–396, 1992. The University of Texas at Austin, where he serves as an Assistant Professor as
[26] S. Zhou and G. B. Giannakis, “Optimal transmitter eigen-beamforming part of the Wireless Networking and Communications Group. His research in-
and space-time block coding based on channel mean feedback,” IEEE terests include interference management in wireless networks, sequence design,
Trans. Signal Process., vol. 50, no. 10, pp. 2599–2613, Oct. 2002. and all aspects of MIMO communication including antenna design, practical re-
[27] , “Optimal transmitter eigen-beamforming and space-time block ceiver architectures, limited feedback techniques, and scheduling algorithms.
coding based on channel correlations,” IEEE Trans. Inf. Theory, vol. 49, Dr. Heath serves as an Associate Editor for the IEEE TRANSACTIONS ON
no. 7, pp. 1673–1690, Jul. 2003. COMMUNICATIONS and the IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY.

Authorized licensed use limited to: Ural Federal University. Downloaded on January 29,2021 at 17:16:46 UTC from IEEE Xplore. Restrictions apply.