Robust Adaptive Modulation for Throughput Maximization in

MIMO Systems Combining OSTBC and Beamforming

Antonio Pascual-Iserte1, Miquel Payaró2, Ana I. Pérez-Neira1, Miguel Ángel Lagunas1,21
1 2
Dpt. Signal Theory and Communications Telecommunications Technological Center of
Polytechnic University of Catalonia (UPC) Catalonia (CTTC)
C/ Jordi Girona 1-3 (campus nord UPC-mòdul D5), 08034 Barcelona (Spain)
emails: tonip@gps.tsc.upc.es, miquel.payaro@cttc.es, anuska@gps.tsc.upc.es, m.a.lagunas@cttc.es

ABSTRACT Note, however, that the most common case corresponds to

a situation in which only a channel estimate, which may be
In this paper it is presented a communication system imperfect or noisy, is available at the transmitter. In this
exploiting multiple antennas at both the transmitter and the case, the best option is to carry out a design in which it is
receiver sides, i.e., configuring a Multi-Input-Multi-Output taken into account explicitly the presence of errors in the
(MIMO) channel, which is assumed to be flat fading. The estimate, leading to the so-called robust solutions [6,7].
transmitter is based on an architecture combining an In this paper we present a robust design in which the
Orthogonal Space Time Block Code (OSTBC) and a set of transmitter architecture is based on the combination of an
beamformers designed according to a channel estimate, Orthogonal Space Time Block Code (OSTBC) and a
which may be noisy. These imperfections in the channel beamformer applied to each output of the OSTBC,
estimate are taken into account explicitly by means of an coupling the signal through the different eigenmodes of the
adequate power allocation strategy, leading to a robust estimated MIMO channel [8,9]. A robust power allocation
design. The ultimate objective is to maximize the policy is designed in order to determine the way in which
throughput applying a robust adaptive modulation scheme the total available power, which is limited, is distributed
while guaranteeing a certain Quality of Service (QoS) in among all the eigenmodes of the channel according to the
terms of a maximum allowed Bit Error Rate (BER). maximin philosophy [10,11], instead of the classical
Bayesian approach, as in [8,9]. In [9] only two estimated
I. INTRODUCTION eigenmodes were used. After calculating the power
Wireless communication systems are currently being allocation, it is proposed an adaptive modulation scheme
required to provide higher bitrates while still maintaining that maximizes the system throughput while guaranteeing a
the same quality and the same number of users to be served minimum Quality of Service (QoS) in terms of a maximum
simultaneously. These extremely demanding requirements Bit Error Rate (BER) for any channel belonging to an
are growing due to the appearance of new wireless services uncertainty region around the available channel estimate.
based on the multimedia capabilities of the user terminals. This paper is organized as follows. In section II the
Very important technical challenges arise due to the system and signal models are described. The robust
appearance of these requirements, especially when taking beamforming and adaptive modulation techniques are
into account that the impairments produced by the wireless presented in sections III and IV. Section V is devoted to
channel are more important as the bitrate increases. One the design of the uncertainty regions, and finally, in section
possible and efficient solution in order to cope with all VI some simulations results and conclusions are obtained.
these issues is the exploitation of time, frequency, and II. SYSTEM AND SIGNAL MODELS
space diversities. Space diversity has received much
attention by many researchers in last years, showing that In this section we present the system and signal models for
extremely high gains can be achieved by using multiple the proposed transmitter architecture. The channel is
antennas. Because of this, several standards have included assumed to be flat fading, uncorrelated, and Rayleigh,
the possibility of using multiple antennas, guaranteeing its which means that all the gains are independent and follow
commercial exploitation in next years. a complex Gaussian distribution with zero mean and
The most general case corresponds to a system in variance σ h2 . It is considered that the transmitter and the
which both the transmitter and the receiver have an array
of antennas, configuring the so-called Multi-Input-Multi- receiver have nT and nR antennas, respectively. The (i,j)th
Output (MIMO) channels. The design strategy for these component of the matrix H ∈ nR ×nT represents the gain
configurations depends on the Channel State Information factor of the channel between the jth transmit and the ith
(CSI) available at the transmitter. In case that no CSI is receive antenna. The transmitter only knows a channel
provided, a space-time coding approach should be taken ˆ = H + E , where E represents the error, and
estimate H
[1,2,3], whereas in case that a perfect channel estimate is
whose components are assumed to be independent,
available, an optimal terminal linear filtering perfectly
matched to the channel realization can be applied [4,5]. Gaussian, and zero-mean with a variance equal to σ e2 .

1
This work has been partially financed by the Spanish Government under projects GIRAFA TIC2002-04594-C02-01 (jointly financed
by FEDER), MARQUIS FIT-070000-2003-257 (MEDEA+ A111), and the integrated action HF2001-0055.
According to these assumptions, the statistics of the real H H
B(i ) = PT UPT(i ) , T( r ) T( r ) = I, T(i ) T(i ) = I
H

channel H conditioned to the estimate Ĥ are expressed as:

P = diag ( p1 ,…, pnT )
h = vec ( H ) , hˆ = vec H
ˆ ( )
As shown in this signal model, nT channel uses or
( ) (
f h | hˆ ∼ CG hˆ b , C , ) hˆ b =
SNR est ˆ
1+SNR est
h (1)
periods of time are needed to transmit each complex
symbol since the length of the spreading sequences is equal
to nT. Obviously, this implies a reduction of the useful data
σ h2 σ h2
C= I, SNR est = rate by a factor nT. If a higher rate transmission is desired,
1+SNR est σ e2 then the following modulation scheme should be applied:

( )
R
where vec(A) represents the vectorization consisting in
XT ( n) = PT UP ∑ Tk( r ) sk( r ) (n) + jTk( i ) sk( i ) (n)
H H
(2)
stacking all the columns of the matrix A, and hˆ b is the k =1

Bayesian MMSE estimate of the channel.

We assume that the transmitted symbols are complex where R complex symbols are transmitted simultaneously,
s(n) = s ( r ) (n) + js (i ) (n) with a normalized power: and therefore, the final symbol rate is R .
nT
2
s {
σ = E s ( n)
2
} = 1 . Let XT (n) be the matrix whose rows In this work it is considered that the receiver knows
{ }
R
are the signal sequences transmitted through the antennas perfectly which are the spreading matrices Tk( r ) , Tk(i ) ,
k =1
obtained by a linear mapping of the symbol s(n):
and also the current realization of the MIMO channel H.
XT (n) = B( r ) s ( r ) (n) + jB(i ) s (i ) (n) . The matrices B( r ) and Based on this fact, an optimum Maximum Likelihood
B(i ) are assumed to be real and square with dimensions (ML) detection could be applied. Depending on the way in
nT × nT . The transmit power limitation is formulated as: which the spreading sequences are designed, the
complexity of the ML detector can be extremely simplified
PT ≤ PTmax , where PTmax is the maximum allowed power to the case of applying only a linear processing to the
{
and PT = E XT (n) 2
F } = 12 ( B (r ) 2
F
+ B( i )
2

F ) is the real received signal samples. In order to obtain this low

complexity optimum receiver, and also to decouple the
transmit power. According to this, the received signal detection of the real and imaginary parts of the transmitted
samples can be expressed as: XR (n) = HXT (n) + W(n) , symbols, the spreading sequences defined by the complex
OSTBC should be used. Note, however, that if this kind of
where W(n) models the contribution of the white Gaussian
sequences are exploited, the maximum symbol rate could
noise (AWGN) of variance σ w2 . only be obtained for 2 transmit antennas. The Signal to
In this paper we constrain the matrices B ( r ) and B ( i ) to Noise Ratio (SNR) at the detection stage for this proposed
have a predefined structure that forces the symbols to be architecture at the transmitter can be expressed as [3]:
transmitted through all the spatial eigenmodes of the
1 PT
estimated channel. This is achieved by using a set of
beamformers. In order to decouple the transmission
SNR =
σ 2
tr ( B H H H HB ) =
σ w2
(
tr U H H H HUP 2 ) (3)
w
through these eigenmodes, a set of orthonormal spreading
sequences is used. This is represented in the following where it has been considered only the transmission of one
equations by the matrices T( r ) and T( i ) , whose columns symbol since, thanks to the use of OSTBC, the detection of
are the orthonormal spreading sequences for s ( r ) (n) and different symbols is independent, and the SNR is equal for
all of them. Note that the SNR does not depend on the
s (i ) (n) , respectively. The set of power parameters { pk }kT=1
n
spreading sequences, provided they are orthonormal.
are responsible for giving a relative importance to the In the following figure it is shown the proposed
different estimated eigenmodes. It can be shown that the architecture for the transmitter:
power constraint at the transmitter side can be
∑
nT p1
reformulated in terms of these parameters as: pk = 1 . k =1
#1

u1
ˆ HH
H ˆ = UΛU H , U = u u nT  , U U = I H
b b  1 nT
{sk(n)}k=1
OSTBC pnT

(
Λ = diag λ1 ,…, λnT , λ1 ≥ ) ≥ λnT
# nT
nT u nT
B( r ) = PT ∑ pk uk t k (r ) H
= PT UPT (r ) H

k =1

Fig. 1. General architecture for the transmitter.

 To summarize, the architecture proposed in this paper ε . Note that M ({ pk } , H ) is proportional to the SNR,
is based on the concatenation of an OSTBC and a set of nT
beamfomers (similarly to [8] and [9]), each one and therefore, the ultimate goal is its maximization.
corresponding to an eigenmode of the estimated MIMO According to this, the maximin optimization problem for a
channel, and applied to an output of the OSTBC. The set of concrete channel estimate can be formulated as [10]:
{ pk }k =1
nT
max min M ({ pk } , H )
powers are responsible for scaling the outputs of
(5)
the OSTBC (see Fig. 1). In the following sections it is { pk } H∈Η( Hˆ b )
shown how to calculate the set of powers according to
different criteria, such as a maximin robust technique. The expression M ({ p } , H )
k
has a quadratic term on

III. A MAXIMIN APPROACH TO ROBUST the error δ . For usual values of the estimation SNR, this
BEAMFORMING term can be neglected. Taking this into account, the
expression of M(·) can be simplified using the equality
tr ( ABCD) = vec ( DH ) (C ⊗ A ) vec ( B ) , as follows:
In [4] and [5], the SNR expression in (3) was maximized H T
for the case where the available channel estimate was
assumed to be perfect: H = H ˆ =H ˆ . The SNR optimal
b nT
solution was to transmit only through the best estimated M = ∑ λk pk + pT AH δ + δH Ap M (6)
spatial mode of the channel, u1 . In [4] and [5] no OSTBC k =1

was considered, which is equivalent to having active only

the first output of the OSTBC by forcing where the matrix A H is constructed by choosing the rows
p1 = 1, pk = 0, k > 1 in our scheme. 1 + k (1 + nT ) , k = 0,…, nT − 1 of UT ⊗ UH Hˆ H , and
b ( )
In practical systems the channel estimate is not perfect, T
and therefore, in this section a robust design is proposed, p =  p1 pnT  . The first step to calculate the
so that it is taken into account explicitly that there may be
errors or noise in the available channel estimate, obtaining
ˆ
optimum p consists in finding the channel in Η H b ( ) or,
a solution that is less sensitive to these errors. There exist equivalently, the vector δ such that
2
δ ≤ ε , that
different ways of modeling the error in the channel
estimate and including robustness capabilities in the minimizes the modified cost function M ({ p } , H ) .
k
As
design. In this work, we have taken a maximin approach.
In this strategy, the objective is to look, given a power M (⋅) depends linearly on δ , the minimum is achieved
distribution, for the worst channel within an uncertainty when δ = ε , and the vector δ minimizing M ( ⋅) is:
region Η H ( )
ˆ
b
around the channel estimate, i.e., to look
δ* = − ε
Ap . According to this, the function to be
for the real channel that minimizes the SNR for a concrete Ap
power distribution. Obviously, the lower the estimation
SNR, the higher the size of the uncertainty region, since maximized with respect to the power parameters { pk } is:
less information about the real channel is provided by the
nT
estimate. Once this worst channel is found, the second step
consists in maximizing the SNR for this worst channel by max ∑ λk pk − 2 ε Ap (7)
{ pk } k =1
designing the power parameters adequately. Let us now
nT
present some useful definitions:
s.t. ∑p k = 1, pk ≥ 0
( )
k =1
M ({ pk } , H ) = tr U H HUPH H 2

This is a convex optimization problem that can be

(= tr U H H b) (
ˆ HHˆ UP 2 + tr U H H
b )
ˆ H ∆UP 2
b solved numerically using efficient algorithms such as the
interior point method [12]. Once the optimum power
+ tr ( U ∆ Hˆ UP ) + tr ( U ∆ ∆UP )
{ }
H H 2 H H 2
b parameters pk* have been obtained, the designer can
ˆ ) = {H : H = H
Η (H ˆ + ∆, ∆ ≤ ε } calculate which is the worst channel for this power
b b F
distribution. If a concrete QoS is attained for this worst
ˆ
where Η H ( ) is the uncertainty region, that can also be
b
channel H* ∈Η H ˆ
b( )and this collection of power
parameters, then it can be guaranteed that this QoS will be
expressed in terms of the vectorized form of the real and
also attained for any channel in the uncertainty region. For
( ) {
estimated channels: Η hˆ b = h : h = hˆ b + δ, δH δ ≤ ε , and } example, the SNR for all the channels in the uncertainty
that can be represented geometrically as a sphere centered
at the Bayesian channel estimate with a radius equal to
region will be higher than or equal to PT M pk* , H* .
2
σw
({ } )
 IV. ROBUST ADAPTIVE MODULATION uncertainty region. In a real deployment, it is also possible
that a channel realization is outside the uncertainty region,
In this section we show an application of the robust design and therefore, this QoS cannot be guaranteed for any
for a power-limited system in which the instantaneous channel. Note, however, that it is possible to calculate
transmit power is upper-bounded by PTmax ( PT ≤ PTmax ). explicitly which is the probability that a channel is inside
The goal is to maximize the throughput while guaranteeing the uncertainty region, i.e., it is possible to find a
a minimum QoS in terms of a maximum BER for any relationship between the size of the uncertainty region and
channel in the uncertainty region. This can be achieved by the probability that the real channel is inside this region.
applying adaptive modulation techniques (see [9,13]). As it has been stated in section II, the error δ = h − hˆ b
Let BER L (SNR) be the function that establishes the
in the Bayesian channel estimate is a white properly
relationship between the SNR and the BER for a L-QAM Gaussian distributed vector with zero mean and variance
modulation. Obviously, given a certain SNR, the BER
σ h2 (1 + SNR est ) . Consequently,
−1 2
increases as the number of levels in the modulation L also δ follows a chi-
increases. Taking this into account, the proposed robust square distribution with 2nR nT degrees of freedom and
adaptive modulation is based on the following steps:
normalized variance σ h2 ( 2 + 2SNR est ) . Let Pout be the
−1

- Define which is the desired QoS in the system in

probability that the real channel, given the channel
terms of the maximum BER, also called target BERt.
estimate, is inside the uncertainty region, i.e., the
- Given a certain channel and uncertainty region,
probability that the desired QoS is achieved. Taking this
calculate H* ∈Η H b ( )
ˆ and p* .
k { } into account, ε = f −1 ( Pout ) , where f ( ⋅) is the cumulative
- Calculate the maximum achievable SNR for the worst density function of the chi-square distribution.
({ } )
max
channel as SNR max = PT M pk* , H* . The main problem is that, for high Pout the errors in the
2
σw channel estimate may be too high so as to consider the first
- Calculate the maximum number of levels L of the 2
order approximation in (6). For example, if ε ≥ hˆ b , the
modulation format that fulfills:
 P max  worst channel would be h* = 0 and δ* = −hˆ b , and not the
 w
({ } )
BER L  T 2 M pk* , H*  ≤ BER t .
σ  Ap . The
one resulting from the approximation δ* = − ε
- If the previous equation cannot be fulfilled for any Ap
value of L, set L=0. In this case, no signal is
following proposal avoids this by defining an uncertainty
transmitted since the QoS requirement cannot be
region that does not contain the channel h = 0 .
satisfied for all the channels in the uncertainty region
while fulfilling the transmission power constraint.
B. Adhoc Approach
- For the selected value of L, calculate the necessary
instantaneous transmission power as: It is also possible to calculate the parameter ε without
σ w2
PT = BER −1
( BER )
t having any relationship with Pout . In this case, we propose
L
M ({ p } , H )
*
k
*
2
the following adhoc rule: ε = (1 + SNR est )−1 hˆ b . In this
Summarizing, this algorithm proposes a robust adaptive design, the channel h = 0 does not belong to the
modulation technique in which the throughput is
uncertainty region. Besides, as SNR est increases, the size
maximized while a certain QoS can be guaranteed to the
user given a channel estimate and a transmission power of the uncertainty region decreases, as expected. It can be
constraint. In the simulations section, it will be shown how shown that, usually, for this kind of uncertainty regions,
the robust design already presented in this paper can the first order approximation in (6) is quite accurate.
provide higher bitrates than other non-robust solutions.
VI. SIMULATION RESULTS AND CONCLUSIONS
V. DESIGN OF THE UNCERTAINTY REGIONS
In this section we provide some simulation results obtained
In the previous section it has been assumed that the from the application of our robust adaptive modulation
uncertainty regions are established and that the parameter technique when compared with the non-robust approach.
ε , related to the size of the uncertainty region, has been In Fig. 2 it is shown the mean value of the throughput
appropriately defined. In this section we show two with respect to the maximum allowed transmit power PTmax
different ways of calculating this parameter. for two different cases of number of antennas: 4x4 and
8x8. Besides, the curves obtained by applying the robust
A. Outage Probability power allocation for fixed modulation schemes, such as
QPSK and 16-QAM, are also plotted as illustrative
As previously stated, the objective of the robust design is examples. For the 4x4 MIMO configuration, the estimation
to guarantee a certain QoS for any channel in the
SNR is fixed to 9 dB, and for the 8x8 it has been fixed to REFERENCES
13 dB. It can be seen that the throughput of the non-robust
approach gets closer to that of the robust technique as the [1] V. Tarokh, N. Seshadri, and A. R. Calderbank,
estimation SNR increases: 2-3 dB gap for SNRest = 9 dB, “Space-Time Codes for High Data Rate Wireless
and 1-2 dB gap for SNRest = 13 dB. In addition to that, it Communication: Performance Criterion and Code
must be also noticed that the 8x8 configuration shows Construction”, IEEE Trans. Inform. Theory, vol. 44,
always higher throughput than the 4x4 system, as pp. 744-765, March 1998.
expected. [2] V. Tarokh, H. Jafharkani, and A. R. Calderbank,
“Space-Time Block Codes from Orthogonal
System throughput vs. maximum available transmission power

Robust with AM
Designs”, IEEE Trans. Inform. Theory, vol. 45, no. 5,
Non−Robust with AM
Fixed Modulation (Robust) pp. 1456-1467, July 1999.
7
[3] G. Ganesan and P. Stoica, “Space-Time Block Codes:
A Maximum SNR Approach”, IEEE Trans. Inform.
Theory, vol. 47, no. 4, pp. 1650-1656, May 2001.
Throughput (bits s−1 Hz−1)

5
[4] A. Pascual-Iserte, A. I. Pérez-Neira, and M. A.
Lagunas, “Joint Beamforming Strategies in OFDM-
16QAM MIMO Systems”, in Proc. IEEE Int. Conf. Acoust.,
3 Speech, Signal Proc., ICASSP’02, May 2002, vol. 3,
MIMO 8x8
pp. 2845-2848.
[5] A. Pascual-Iserte, A. I. Pérez-Neira, and M. A.
1
QPSK
Lagunas, “On Power Allocation Strategies for
MIMO 4x4 Maximum Signal to Noise and Interference Ratio in
0
−5 0 5 10 15 20 25 30 35 an OFDM-MIMO System”, accepted at IEEE Trans.
Maximum Available Power (dB)

Fig. 2. Mean throughput vs maximum transmit power. Wireless Comm., April 2003.
[6] F. Rey, M. Lamarca, and G. Vázquez, “Optimal
In Fig. 3 it is plotted the mean value of the efficiency, Power Allocation with Partial Channel Knowledge
defined as η = R , for different values of SNRest, for the for MIMO Multicarrier Systems”, in Proc. IEEE
PT Vehic. Techn. Conf., VTC’02, Sept. 2002, vol. 4, pp.
4x4 MIMO configuration with a maximum transmission 2121-2125.
power of 20 dB. As it is concluded from the figure, the [7] A. Pascual-Iserte, A. I. Pérez-Neira, and M. A.
robust adaptive modulation presents always a higher Lagunas, “Exploiting Trnamission Spatial Diversity
efficiency than the non-robust system. in Frequency Selective Systems with Feedback
Channel”, in Proc. IEEE Int. Conf. Acoust., Speech,
0.4
System Efficiency for various levels of Channel Estimation SNR
Signal Proc., ICASSP’03, April 2003.
Robust System
Non−Robust System [8] G. Jöngren, M. Skoglund, and B. Ottersten,
0.35
“Combining Beamforming and Orthogonal Space-
0.3
Time Block Coding”, IEEE Trans. Inform. Theory,
Efficiency − Rate / Power (bits s−1 dB−1)

vol. 48, no. 3, pp. 611-627, March 2002.

0.25 [9] S. Zhou and G. B. Giannakis, “Adaptive Modulation
for Multi-Antenna Transmission with Channel Mean
0.2
Feedback”, in Proc. IEEE Int. Conf. Comm., ICC’03,
0.15 May 2003.
[10] S. A. Kassam and H. V. Poor, “Robust Techniques
0.1
for Signal Processing: A Survey”, Proceedings of the
0.05
IEEE, vol. 73, no. 3, pp. 433-481, March 1985.
[11] S. A. Vorobyov, A. B. Gershman, and Z-Q. Luo,
0
4 6 8 10 12 14 16 “Robust Adaptive Beamforming Using Worst-Case
Channel Estimation SNR (dB)
Performance Optimization: A Solution to the Signal
Fig. 3. Mean efficiency vs estimation SNR (SNRest)
Mismatch Problem”, IEEE Trans. Signal Proc., vol.
In this paper we have presented a robust adaptive 51, pp. 313-324, Febr. 2003.
modulation scheme which is seen to make a more efficient [12] S. Boyd and L. Vandenberghe, Introduction to
use of the power obtaining higher throughputs than Convex Optimization with Engineering Perspectives,
traditional non-robust approaches, while guaranteeing a Course Notes. Stanford University, 2000.
predefined QoS. Future work is to be done on the design of [13] S.T. Chung and A.J. Goldsmith, “Degrees of freedom
a system in which the MIMO channel is diagonalized and a in adaptive modulation: A unified view,” IEEE Trans.
robust adaptive modulation scheme is applied over the Comm., vol. 49, pp. 1561–1571, Sep. 2001.
different generated spatial subchannels.