Per-Antenna Power Constrained MIMO

Transceivers Optimized for BER

Ching-Chih Weng and P. P. Vaidyanathan
Dept. of Electrical Engineering, MC 136-93
California Institute of Technology, Pasadena, CA 91125, USA
E-mail: [email protected], [email protected]

Abstract— This paper considers the linear transceiver op- along with precoder, under the total power constraint, was also
timization problem for multi-carrier multiple-input multiple- found [13], [6].
output (MIMO) channels with per-antenna power constraints. Instead of total power constraint, in this paper we con-
Because in practical implementations each antenna is limited
individually by its equipped power amplifier, this paper adopts sider the more realistic per-antenna power constraints on the
the more realistic per-antenna power constraints, in contrast transmitter [7], [11], since in practice each antenna is limited
to the conventional sum-power constraint on the transmitter individually by its equipped power amplifier. The optimization
antennas. Assuming perfect channel knowledge both at the process consists of two steps. We first formulate the MMSE
transmitter and the receiver, the optimization problem can be (Minimum Mean Square Error) linear transceiver design with
transformed into a semi-definite program (SDP), which can
be solved by convex optimization tools. Furthermore, several per-antenna power constraints to be a semi-definite program
objective functions of the MIMO system, including average bit (SDP), which can be efficiently solved by convex optimization
error rate, can also be optimized by the introduction of the tools. Then, among the family of MMSE transceivers, we
majorization theory. 1 develop a method to find the one that minimizes the average
Index Terms — Per-antenna Constraints, BER Optimiza- bit error rate as well as many other Schur-convex objective
tion, Schur Convexity, Semi-definite Programming, MIMO
Transceivers. functions. The proof of the optimality is also given in the
paper.
I. I NTRODUCTION This paper is structured as follows. In Section II, after a
brief introduction, the optimal receiver matrix as well as the
In this paper we consider the optimization of multi-carrier
resulting error covariance matrix will be given. The formula-
multiple-input multiple-output (MIMO) communication sys-
tion of MMSE transceiver optimization problem in the form of
tems with per-antenna power constraints. In the MIMO system,
a SDP will also be described. In Section III, the optimization
the transmitter has M antennas sending independent infor-
of average BER problem will be discussed. Section IV presents
mation to the receiver equipped with J antennas. The signal
the numerical simulation results related to the topics discussed
vector consisting of M sub-streams is assumed to be linearly
in the paper. The final conclusions of the paper are summarized
transformed by the channel matrix H, and corrupted by the
in section V.
additive Gaussian noise.
Based on the assumption of having perfect channel state II. MMSE T RANSCEIVER D ESIGN WITH P ER -A NTENNA
information, the transmitter can use appropriate precoding, and P OWER C ONSTRAINTS
jointly with the equalization scheme at the receiver side, better
performance can be achieved. It is known that the optimal In this section we describe the MMSE transceiver optimiza-
equalization technique is the maximum likelihood receiver. tion with per-antenna power constraints. This serves as the
However, due to the heavy computation load, usually the first step toward minimizing the average BER. We consider
linear precoding and equalization scheme, or other suboptimal a communication system with M transmit and J receive
techniques, such as linear precoder with decision feedback antennas. This gives rise to a MIMO channel that can be
equalizers, are utilized. Under those schemes, several authors represented by a channel matrix. Consider the system model
considered the optimization of the MIMO communication in Fig. 1, where s is the M ×1 transmitted symbol vector, H is
systems. In [12], the authors considered the linear transceiver the channel matrix with dimension J × M , which is assumed
optimization problem under the total power constraint. By in- to be fixed during the transmission of a block, and known at
troducing the majorization theory, several objective functions, both sides of the communication links. We assume zero-mean
which can be categorized as Schur-convex or Schur-concave unit-energy uncorrelated (white) symbols, i.e., E[ss † ] = IM .
[10] functions of the mean squared errors, were optimized. The received signal prior to equalization is
Recently, the optimal solution for decision feedback equalizer
y = HFs + n. (1)
1 Thiswork is supported in parts by the ONR grant N00014-06-1-0011, and
the TMS scholarship 94-2-A-018 of the National Science Council of Republic where F is the precoding matrix, and n is the additive Gaus-
of China, Taiwan. sian noise with covariance matrix R n , i.e., n ∼ CN(0, Rn ).
The per-antenna power constraints can be formulated as Since M and J are constants, R n is known, and the MMSE
† † † depends only on W, which is a function of F as in (7).
(E[Fss F ])ii = (FF )ii ≤ Pi , ∀i = 1, 2, · · · , M (2)
Furthermore, if we define
n U ≡ FF† (10)
y ^s then according to (7) we can write
s
M
F M
H J
G
W = (HUH† + Rn )−1
M
(11)
By the above formulation, we can re-cast the problem (9)
Fig. 1. Communication model. as follows:

Usually the performance of the communication system min T r(WRn ) (12)

U
depends on the mean square error (MSE). To design the s.t. (a) (U)ii ≤ Pi , ∀i = 1, 2, · · · , M
system, we first derive the optimum receiver matrix G which
(b) U  0
minimizes the MSE, when H, F, and R n are given. The error
vector in Fig. 1 of the transmitted symbols before the hard (c) W = (HUH† + Rn )−1
decision device is where U  0 means that the matrix U is positive semi-
e = Gy − s = (GHF − I)s + Gn. (3) definite. The problem formulation in (12) is not convex though.
Similar to [9], by introducing an auxiliary matrix variable W 0 ,
Under the assumption that the transmitted symbols have no we can cast the problem in (12) as the following equivalent
correlation with the additive noise, the covariance matrix of problem:
the error vector is
min T r(W0 Rn ) (13)
E[ee† ] = (GHF − I)(GHF − I)† + GRn G† (4) U,W0

It is well known that if F is given [12], [9], the optimum s.t. (a) (U)ii ≤ Pi , ∀i = 1, 2, · · · , M
receiver matrix would be the linear minimum MSE (LMMSE) (b) U  0
receiver or so called Wiener filter: (c) W0  (HUH† + Rn )−1
† † † † −1
Gopt = F H (HFF H + Rn ) , (5) The equivalence of problem formulation (12) and (13) can
and the resulting error covariance matrix (MSE matrix) can be easily explained as follows: In (13) we are minimizing the
be written as trace of matrix W 0 Rn , where Rn is positive semi-definite.
Therefore we have [9]
E ≡ E[ee† ] = I − F† H† WHF (6)
T r(W0 Rn ) ≥ T r((HU† H† + Rn )−1 Rn )
where
since W0  (HUH† + Rn )−1 . The optimal solution of
W = (HFF† H† + Rn )−1 (7)
problem (13) ensures the equality
This can be rewritten using matrix inversion lemma [4]
W0 = (HU† H† + Rn )−1
† † †
E ≡ E[ee ] = (I + F H R−1
n HF)
−1
(8)
holds, where U is also the optimal solution for (12). Note that
In the following we will describe how to solve the MMSE the constraint
problem with the per-antenna power constraints. This is done
by formulating the problem as a semi-definite program (SDP) W0  (HUH† + Rn )−1
as we shall see.
holds if only if the following linear matrix inequality holds
The MMSE problem with per-antenna power constraints can
(p. 472 in [4])
be formulated as  
HU† H† + Rn I
min T r(E) (9) 0 (14)
F I W0
s.t. (FF† )ii ≤ Pi , ∀i = 1, 2, · · · , M Therefore, our final form of problem formulation can be
where T r(E) denotes the trace of the matrix E. From (6) written as
T r(E) = T r(I − F† H† WHF) min T r(W0 Rn ) (15)
U,W0
= M − T r(F† H† WHF) s.t. (a) (U)ii ≤ Pi , ∀i = 1, 2, · · · , M
= M − T r(WHFF† H† ) (b) U  0
= M − T r(W(W−1 − Rn ))  
HUH† + Rn I
(c) 0
= M − J + T r(WRn ) I W0
 In (15) the objective function is linear, and the constraints Note that for any unitary matrix Ψ, the product F mmse Ψ will
are either linear or positive semi-definite, which implies the still be an MMSE solution. This is the freedom we have, to
problem (15) is a SDP problem [15]. This ensures that the optimize the average bit error rate. Suppose matrix V is the
global minimum of (15) can be found in polynomial time. The unitary matrix which diagonalizes F †mmse H† R−1n HFmmse :
algorithmic complexity of solving SDP using interior points †
F†mmse H† R−1
n HFmmse = VΣV . (17)
methods is O(n6.5 log(1/ε)), where ε is the solution accuracy,
and n is the problem size [15], [1]. Once we solve the problem Then the MSE matrix in (8) is
(15), the precoder matrix F is the Cholesky factorization of
the solution U in (15). E[ee† ] = (I + F†mmse H† R−1
n HFmmse )
−1

= V(I + Σ)−1 V† (18)

In view of Theorem 1, which will be proved later, the
optimal F has the form
F = Fmmse VΦ (19)
where Φ denotes the unitary matrix such that the MSE matrix
as in (8) has the identical diagonal elements. Examples of such
Φ are the Hadamard matrix and the discrete Fourier transform
(DFT) matrix [8], [3].
The performance of a digital communication system ulti-
mately is given by the bit error rate (BER). Assuming a Gray
X=5 encoding is used to map the bits into the constellation points,
X+Y=8
X+Y 8
Y=5 the BER can be approximately obtained from the symbol error
rate (SER), i.e.,
BER ∼
= SER/k,
where k is the number of bits per-symbol. Under the Gaussian
Fig. 2. Visualization of the feasible region of the solution of (15) for the assumption the SER can be approximately expressed as a
two antenna case. function of SINR:

P e(SIN R) = αQ( βSIN R), (20)
Fig. 2 shows the visualization of the feasible region of the where Q is Q-function defined as in [12] and α and β
solution U in problem (15) for the two antenna case. For are constants depending on the constellation. The following
simplicity, we assume lemma is useful in developing the BER optimality of F in
 
X Z (19).
U= Lemma 1. In the high SINR region, the bit error rate is a
Z Y
convex increasing function of the MSE. That is, BER(E ii ) is
where Z is a real number. Therefore the constraint U  0 a convex increasing function of E ii , where Eii is defined as
forms a positive semi-definite cone. The constraints U kk ≤ 5 the ith diagonal term of the MSE matrix in (8).
define the halfspaces on one side of the hyperplanes U kk = 5. Proof: See Appendix H in [12].
The total power constraint Tr(U) ≤ 8, is also a halfspace The high SINR regions represent the low BER regions,
formed by the hyperplane cutting through the semi-definite which are often the focus of the analysis in communication
cone. The convex problem is solved by running the interior systems. The following theorem shows that the system with F
point method [2] in the feasible region, which is formed by in (19), and with the corresponding G in (5), has the minimum
the intersection of the semi-definite cone and the halfspaces. average BER.
III. BER O PTIMIZATION OF T RANSCEIVER WITH Theorem 1. Among all systems in Fig. 1 which satisfy (2),
P ER -A NTENNA P OWER C ONSTRAINTS the F in (19) and the corresponding G in (5) are the optimal
solutions that minimize the average BER, which is defined as
In this section we derive the optimal transceiver for the av-
M
erage bit error rate (BER), with per-antenna power constraints. 1 
This method is based on distributing the MSE identically in BER = BER(Eii ) (21)
M i=1
each substream.
.
Suppose we already found the solution U mmse to the
Proof: For the system {F mmse , Gmmse }, sup-
problem (15). The optimal MMSE solution F mmse can be
pose the diagonal terms of the MSE matrix are E =
taken as any Cholesky factor of U mmse , so that
{E11 , E22 , · · · , EMM }, where Fmmse is a Cholesky factor
Fmmse F†mmse = Ummse . (16) of Ummse , and Gmmse is the corresponding LMMSE filter.
  
Suppose there is another system, {F , G }, which has the di- −1
    10
agonal terms of the MSE matrix E = {E11 , E22 , · · · , EMM }.
Trivially we have
−2
 10
E  E
where E  is defined as the vector with all elements equal to the −3
10
arithmetic mean of E  , and ’’ denotes the majorization as in
[10]. According to Lemma 1, BER(·) is a convex function

BER
−4
10
of Eii . Therefore the average BER is a Schur-convex function
of E (p. 70 in [10]). Then we have OPT−MMSE−nodist
−5
10 OPT−MMSE−dist

1
M  Naive−nodist
BER = M i=1 BER(Eii ) (22) Naive−dist
1
M 
−6
10
≥ BER( M i=1 Eii ) (23)
1
M
≥ BER( M i=1 Eii ) (24) −7
10
= BER(19) (25) 0 10 20
α (dB)
30 40 50

where BER(19) is defined as the BER obtained by using

F in (19). The inequality in (23) follows from the fact that Fig. 3. Comparing four different transceivers for channel H1 , with per-
antenna power constraints α[5, 4, 3, 2].
Schur-convex function preserves majorization. The inequality
in (24) is due to the fact that E is the corresponding MSE of
The plots in Fig. 3 and Fig. 4 are with respect to the
the solution to problem (15), and the fact that BER(·) is an
variable α, where α controls the total power. “OPT-MMSE-
increasing function. The equality in (25) is because F in (19)
nodist” denotes the optimal MMSE design but without dis-
is the one that makes all diagonal terms of the MSE matrix
tributing the MSE in each sub-stream, i.e., the precoder is
identical.
taken to be a Cholesky factor F mmse as in (16). “OPT-
Therefore, any other system satisfying (2) will have larger
MMSE-dist” denotes the optimal MMSE design with the
average BER than the system with F as in (19) and the
MSE in each sub-stream identical, which corresponds to using
corresponding LMMSE receiver G as in (5). This concludes
F as in (19). “Naive-nodist” denotes the case where the
our proof.
power constraints are satisfied by using the simple choice
Note that the proof given above relies only on the fact that
Fnaive = diag(α[P1 , P2 , · · · , PM ]). “Naive-dist” corresponds
the average BER is a Schur-convex and increasing function
to the case where the precoder matrix is F = F naive VΦ as
of MSE’s. Therefore the same concept can also be applied to
in (19), so that (8) has identical diagonal elements.
other objective functions that have these two properties. Many
examples of such objective functions are provided in [12]. For
all such objective functions, the system with F as in (19) and −1
10
the corresponding LMMSE receiver G as in (5) is always the
optimal one! −2
10
IV. N UMERICAL R ESULTS
In this section we present our simulation results relevant to −3
10
the discussions carried out in previous sections. After formu-
lating the problem as in (15), we use the convex optimization
BER

−4
10
tool “SeDuMi” [14] to obtain the optimal solution.
OPT−MMSE−nodist
In both Fig. 3 and Fig. 4 we choose M = 4, J = 5, and per- OPT−MMSE−dist
antenna power constraints to be [P 1 , P2 , P3 , P4 ] = α[5, 4, 3, 2], −5
10 Naive−nodist
where α controls the total power. The constellation are all Naive−dist

QPSK. The noise is additive white Gaussian for each case, −6

10
with covariance matrix R n = 4I for Fig. 3, and R n = 6.25I
for Fig. 4. The channels used in the simulations are H 1 and −7
H2 , for Fig. 3 and Fig. 4 respectively, given as follows: 10
0 10 20 30 40 50
 −1.59 − 0.62i −0.94 − 0.85i −0.44 − 0.43i −0.70 − 0.50i  α (dB)
1.27 + 0.88i 0.44 − 1.66i 0.75 + 0.34i −0.58 + 1.17i
H1 = −0.40 − 0.05i
0.22 − 1.16i
0.17 − 0.70i
−0.00 − 0.80i
−2.48 + 0.51i
1.07 − 1.23i
0.31 − 0.18i
0.80 + 0.79i
−0.80 − 0.99i −2.09 + 0.13i 0.88 − 0.22i −1.86 + 0.18i Fig. 4. Comparing four different transceivers for channel H2 , with per-
antenna power constraints α[5, 4, 3, 2].
and
 0.01 + 0.47i −0.19 + 0.43i 0.84 + 0.86i 2.27 − 1.37i

1.81 − 0.23i −1.14 − 0.62i −1.43 + 2.48i 0.63 − 0.83i
H2 = 2.47 + 1.34i 3.66 − 2.19i 1.78 + 0.27i 1.71 − 1.41i From the simulations, we can see that the proposed op-
−0.82 − 0.45i −1.19 − 1.51i 0.96 − 0.74i −0.08 − 1.67i
1.38 − 0.15i 0.82 − 0.17i 0.26 − 0.97i −1.23 + 1.64i timization method is superior to naive designs. Also, both
the MSE-minimization step and MSE-distribution step are
beneficial for minimizing the average BER.
V. C ONCLUDING R EMARKS
We have presented a convex formulation for optimal
transceiver design with per-antenna power constraints. The
problem is re-formulated as a SDP, which can be solved
efficiently by convex optimization tools. The optimization for
average BER as well as other Schur-convex functions are
also discussed. Several simulation results were presented to
demonstrate the advantages of the optimized designs.
R EFERENCES
[1] A. Antoniou, and W. S. Lu, “Practical optimization: algorithms and
engineering applications,” Springer. 2007.
[2] S. Boyd, and L. Vandenberghe, Convex optimization, Cambridge. 2004.
[3] Y. Ding, T. N. Davidson, Z. Luo, and K. M. Wong, “Minimum BER
precoders for zero-forcing equalization,” IEEE Trans. Sig. Proc. pp. 2410
- 2423, Sep. 2003.
[4] R. A. Horn, C. R. Johnson, “Matrix analysis,” Cambridge Univ. Press
1985.
[5] Y. Jiang, J. Li, and W. W. Hager “Joint transceiver design for MIMO
communications using geometric mean decomposition, ” IEEE Trans. Sig.
Proc., pp.3791-3803, Oct. 2005.
[6] Y. Jiang, D. P. Palomar, and M. K. Varanasi, “Precoder Optimization for
Nonlinear MIMO Transceiver Based on Arbitrary Cost Function,” Proc.
Conference on Information Sciences and Systems. March, 2007.
[7] K. H. Lee, and D. P. Petersen, “Optimal linear coding for vector channels,”
IEEE Trans. Comm. pp. 1283 - 1290, Dec. 1976.
[8] Y. P. Lin, and S. M. Phoong, “BER minimized OFDM systems with
channel independent precoders,” IEEE Trans. Sig. Proc. pp. 2369 - 2380,
Sep. 2003.
[9] Z. Luo, T. N. Davidson, G. B. Giannakis, and K. M. Wong, “Transceiver
optimization for block-based multiple access through ISI channels,” IEEE
Trans. Sig. Proc. pp. 1037 - 1052, Apr. 2004.
[10] A. W. Marshall, I. Olkin, “Inequalities: Theory of majorization and its
applications,”Academic Press. 1979.
[11] H. S. Malvar, and D. H. Staelin, “Optimal pre- and postfilters for
multichannel signal processing,” IEEE Trans. Acoustics, Speech, and
Signal Processing pp. 287 - 289, Feb. 1988.
[12] D. P. Palomar, J. M. Cioffi, and M. A. Lagunas, “Joint Tx-Rx beam-
forming design for multicarrier MIMO channels: a unified framework for
convex optimization,” IEEE Trans. Sig. Proc. pp. 2381 - 2401, Sept.
2003.
[13] M. B. Shenouda, T. N. Davidson, “A framework for designing MIMO
systems with decision feedback equalization or Tomlinson-Harashima
precoding,” Proc. of the ICASSP, pp. 209-212, April. 2007.
[14] J. F. Sturm, “Using SeDuMi 1.02, a Matlab toolbox for optimization
over symmetric cones” Optim. Meth. Software. vol. 11-12 pp. 625 - 653,
1999.
[15] L. Vandenberghe, and S. Boyd, “Semidefinite programming,” SIAM Rev.
pp. 49 - 95, 2001.
[16] F. Xu, T. N. Davidson, J. K. Zhang, S. S. Chan, and K. M. Wong,
“Design of block transcivers with MMSE decision feedback detection,”
Proc. of the ICASSP, pp. 1109 - 1112, March. 2004.