390 IEEE COMMUNICATIONS LETTERS, VOL. 14, NO.

5, MAY 2010

Optimal Joint Source and Relay Beamforming for

MIMO Relays with Direct Link
Yue Rong, Member, IEEE

Abstract—In this letter, we investigate the optimal structure In this letter, we derive the optimal structure of the source
of the source precoding matrix and the relay amplifying matrix precoding matrix when the direct link is included. Interest-
for non-regenerative multiple-input multiple-output (MIMO) ingly, we prove that the source precoding matrix has a transmit
relay communication systems with the direct source-destination
link. We show that both the optimal source precoding matrix beamforming structure. By exploiting the optimal structure
and the optimal relay amplifying matrix have a beamforming of the source and relay matrices, we develop an iterative
structure. Based on this structure, an iterative joint source joint source and relay beamforming algorithm to minimize
and relay beamforming algorithm is developed to minimize the the MSE of the signal waveform estimation at the destination
mean-squared error (MSE) of the signal waveform estimation. node. Numerical example demonstrates the effectiveness of
Numerical example demonstrates an improved performance of
the proposed algorithm. our algorithm.
Index Terms—Beamforming, direct link, MIMO relay, linear
non-regenerative relay. II. S YSTEM M ODEL
We consider a three-node MIMO communication system
I. I NTRODUCTION where the source node transmits information to the destination
node with the aid of one relay node. The source, relay, and

R ECENTLY, non-regenerative multiple-input multiple-

output (MIMO) relay communications have attracted
much research interest. For a MIMO relay system, there are
destination nodes are equipped with 𝑁𝑠 , 𝑁𝑟 , and 𝑁𝑑 antennas,
respectively. Due to its merit of simplicity, a non-regenerative
strategy is applied at the relay node to amplify and forward the
two independent links between the source and the destination received signals. The signal vector received at the destination
nodes: the source-relay-destination link and the direct source- node over two consecutive time slots is
destination link. Many works studied the optimal relay ampli- [ ]
fying matrix for the source-relay-destination channel. In [1] y𝑑 (𝑡 + 1)
y(𝑡) ≜
and [2], the optimal relay amplifying matrix which maximizes y𝑑 (𝑡)
[ ] [ ]
the source-destination mutual information (MI) was derived. H𝑟𝑑 FH𝑠𝑟 H𝑟𝑑 Fv𝑟 (𝑡) + v𝑑 (𝑡 + 1)
= Bs(𝑡) + (1)
In [3] and [4], the relay amplifying matrix was designed H𝑠𝑑 v𝑑 (𝑡)
to minimize the mean-squared error (MSE) of the signal where y𝑑 (𝑡 + 1) and y𝑑 (𝑡) are 𝑁𝑑 × 1 signal vectors received
waveform estimation at the destination. A unified framework at the destination through the source-relay-destination link
was developed in [5] to jointly optimize the source precoding and the direct source-destination link, respectively, H𝑠𝑑 , H𝑟𝑑 ,
matrix and the relay amplifying matrix for a broad class of H𝑠𝑟 are the channel matrices for the source-destination, relay-
objective functions. All these works did not consider the direct destination, and source-relay links with dimension 𝑁𝑑 × 𝑁𝑠 ,
source-destination link. 𝑁𝑑 × 𝑁𝑟 , 𝑁𝑟 × 𝑁𝑠 , respectively, s(𝑡) is the 𝑁𝑏 × 1 source
In practice, the direct source-destination link provides valu- signal vector, F is the 𝑁𝑟 × 𝑁𝑟 relay amplifying matrix, B is
able spatial diversity to the MIMO relay system and should not the 𝑁𝑠 ×𝑁𝑏 source precoding matrix, v𝑟 (𝑡) is the 𝑁𝑟 ×1 noise
be ignored. Obviously, the relay amplifying matrix designed vector at the relay, v𝑑 (𝑡 + 1) and v𝑑 (𝑡) are the 𝑁𝑑 × 1 noise
for the source-relay-destination link only [1]-[4] is not optimal vectors at the destination at time 𝑡 + 1 and 𝑡, respectively.
when the direct link is included. Recently, it was shown in [6] We assume that the source signal vector satisfies
that with the direct link, the optimal relay amplifying matrix E[s(𝑡)(s(𝑡))𝐻 ] = I𝑁𝑏 and all noises are independent and
has a general beamforming structure. In [7], an alternating identically distributed (i.i.d.) additive white Gaussian noise
algorithm was proposed to optimize both the source and relay (AWGN) with zero mean and unit variance. Here E[⋅] stands
matrices. However, the algorithm in [7] is strictly suboptimal for the statistical expectation, I𝑛 is an 𝑛 × 𝑛 identity matrix,
and has a high computational complexity, since the optimal and (⋅)𝐻 denotes the Hermitian transpose.
beamforming structure of the relay amplifying matrix was
When a linear receiver is used at the destination node, the
not exploited. Moreover, the structure of the optimal source
estimated signal waveform is given by
precoding matrix with the direct link was not investigated in
[6] and [7]. ŝ(𝑡) = W𝐻 y(𝑡) (2)
Manuscript received December 18, 2009. The associate editor coordinating where W is a 2𝑁𝑑 × 𝑁𝑏 weight matrix. The receiver weight
the review of this letter and approving it for publication was R. Nabar. matrix which minimizes the signal waveform estimation error
Y. Rong is with the Department of Electrical and Computer Engineering,
Curtin University of Technology, Bentley, WA 6102, Australia (e-mail: is the Wiener filter given by [8]
[email protected]). ( )−1
Digital Object Identifier 10.1109/LCOMM.2010.05.092429 W = H̄H̄𝐻 + C̄ H̄ (3)
1089-7798/10$25.00 ⃝
c 2010 IEEE
RONG: OPTIMAL JOINT SOURCE AND RELAY BEAMFORMING FOR MIMO RELAYS WITH DIRECT LINK 391

where where the matrix inversion lemma is applied to obtain the

[ ] [ ] objective function in (11) from (4). For 𝑁𝑏 ≥ 2, the problem
H𝑟𝑑 FH𝑠𝑟 H𝑟𝑑 FF𝐻 H𝐻
𝑟𝑑 + I𝑁𝑑 0𝑁𝑑 ×𝑁𝑑
H̄ ≜ B, C̄ ≜ . (11)-(12) does not have a closed-form solution for general
H𝑠𝑑 0𝑁𝑑 ×𝑁𝑑 I𝑁𝑑
H𝑠𝑑 . We should resort to numerical methods, such as the
Here 0𝑚×𝑛 denotes an 𝑚 × 𝑛 matrix with all zeros entries, projected gradient method [9] to solve (11)-(12).
and (⋅)−1 denotes the matrix inversion. Using (1)-(3), the MSE For a fixed F, the source precoding matrix B is optimized
matrix E of the signal waveform estimation is given by by solving the following problem
[ ] ( )
E = E (ŝ(𝑡) − s(𝑡))(ŝ(𝑡) − s(𝑡))𝐻 min tr [I𝑁𝑏 + B𝐻 Ψ1 B]−1 (13)
[ B
= I𝑁𝑏 + B𝐻 H𝐻 𝐻 𝐻 𝐻 𝐻
𝑠𝑑 H𝑠𝑑 B + B H𝑠𝑟 F H𝑟𝑑
]−1 s.t. tr(B𝐻 Ψ2 B) ≤ 𝑃¯𝑟 (14)
×(H𝑟𝑑 FF𝐻 H𝐻 𝑟𝑑 + I𝑁𝑑 )
−1
H𝑟𝑑 FH𝑠𝑟 B . (4) 𝐻
tr(B B) ≤ 𝑃𝑠 (15)

III. O PTIMAL J OINT S OURCE A ND R ELAY B EAMFORMING where Ψ1 ≜ H𝐻 𝐻 𝐻 𝐻

𝑠𝑑 H𝑠𝑑 + H𝑠𝑟 F H𝑟𝑑 (H𝑟𝑑 FF H𝑟𝑑 +
𝐻 𝐻

With the direct link, the relay amplifying matrix and source I𝑁𝑑 ) H𝑟𝑑 FH𝑠𝑟 , Ψ2 ≜ H𝑠𝑟 F FH𝑠𝑟 , 𝑃¯𝑟 ≜ 𝑃𝑟 − tr(FF𝐻 ).
−1 𝐻 𝐻

precoding matrix optimization problem is written as The following Theorem establishes the structure of the optimal
B.
min tr(E) (5) T HEOREM 1: The optimal B as the solution to the problem
F,B
( ) (13)-(15) is given by
s.t. tr F(H𝑠𝑟 BB𝐻 H𝐻
𝑠𝑟 + I𝑁𝑟 )F
𝐻
≤ 𝑃𝑟 (6)
( 𝐻
) B = M−𝐻 V1 D (16)
tr BB ≤ 𝑃𝑠 (7)
where tr(⋅) denotes the trace of a matrix, (6) and (7) is where MM𝐻 = 𝜇1 I𝑁𝑠 + 𝜇2 Ψ2 , D is an 𝑁𝑏 × 𝑁𝑏 diagonal
the power constraint at the relay node and the source node, matrix, M−1 Ψ1 M−𝐻 = VΣV𝐻 , and V1 contains 𝑁𝑏
respectively, and 𝑃𝑟 > 0, 𝑃𝑠 > 0 is the power budget available columns of V associated with eigenvalues that are greater than
at the relay and source node, respectively. The problem (5)- one. Here Σ is the diagonal eigenvalue matrix, and 𝜇1 ≥ 0,
(7) is highly nonconvex and a closed-form expression of the 𝜇2 ≥ 0 are the Lagrange multipliers.
optimal F and B is intractable. In this letter, we develop an P ROOF: The Lagrangian function associated with the prob-
iterative algorithm to optimize F and B. lem (13)-(15) is given by
First we derive the optimal structure of F and B. For a given
ℒ = tr([I𝑁𝑏 + B𝐻 Ψ1 B]−1 ) + 𝜇1 (tr(B𝐻 B) − 𝑃𝑠 )
B, the relay matrix F is optimized by solving the following
problem +𝜇2 (tr(B𝐻 Ψ2 B) − 𝑃¯𝑟 )
= tr([I𝑁𝑏 + B𝐻 Ψ1 B]−1 ) + tr(B𝐻 MM𝐻 B)
min tr(E) (8)
F
( ) −𝜇1 𝑃𝑠 − 𝜇2 𝑃¯𝑟 . (17)
s.t. tr F(H𝑠𝑟 BB𝐻 H𝐻
𝑠𝑟 + I𝑁𝑟 )F
𝐻
≤ 𝑃𝑟 . (9)
Making the derivative of ℒ with respect to B be zero, we
Let us introduce the following singular value decomposition obtain
H𝑠𝑟 B = U𝑠 Λ𝑠 V𝑠𝐻 , H𝑟𝑑 = U𝑟 Λ𝑟 V𝑟𝐻 ∂ℒ [ ]−2
= − I𝑁𝑏 + B𝐻 Ψ1 B B𝐻 Ψ1 + B𝐻 MM𝐻 = 0 (18)
∂B
where Λ𝑠 and Λ𝑟 are 𝑅𝑠 × 𝑅𝑠 and 𝑅𝑟 × 𝑅𝑟 square diagonal
matrices (i.e., zero singularvalues are excluded). Here 𝑅𝑠 ≜ where the derivatives of ∂tr(ΘX−1 )/∂X = −(X−1 ΘX−1 )𝑇
rank(H𝑠𝑟 B), 𝑅𝑟 ≜ rank(H𝑟𝑑 ), rank(⋅) denotes the rank of and ∂tr(ΘX)/∂X = Θ𝑇 are used. Since M is non-singular,
a matrix. Based on the theorem in [6], the optimal F is given (18) can be equivalently written as
by [ ]−2
− I𝑁𝑏 + B𝐻 MM−1 Ψ1 M−𝐻 M𝐻 B
F = V𝑟 AU𝐻 𝑠 . (10)
×B𝐻 MM−1 Ψ1 M−𝐻 + B𝐻 M = 0. (19)
It can be seen from (10) that F has a generalized beamforming
structure. The relay first performs receive beamforming using Substituting M−1 Ψ1 M−𝐻 = VΣV𝐻 back into (19) and
the Hermitian transpose of the left singular matrix of the assuming M𝐻 B = V1 T, where T is non-singular, we have
effective source-relay channel H𝑠𝑟 B. Then the relay conducts [ ]−2
− I𝑁𝑏 + T𝐻 Σ1 T T𝐻 Σ1 V1𝐻 + T𝐻 V1𝐻 = 0 (20)
a linear precoding operation using A. Finally, a transmit
beamforming is performed by the relay using the right singular where Σ1 contains the 𝑁𝑏 eigenvalues that are greater than 1.
matrix of the relay-destination channel H𝑟𝑑 . Applying the matrix identity of (I𝑚 + XX𝐻 )−1 X = X(I𝑛 +
Substituting (10) back into (8) and (9), the optimal A can X𝐻 X)−1 for any 𝑚 × 𝑛 matrix X, we have
be obtained by solving the following optimization problem [ ]
[ ]−2 1 1 1 1 −2
([ I𝑁𝑏 +T𝐻 Σ1 T T𝐻 Σ12 = T𝐻 Σ12 I𝑁𝑏 +Σ12 TT𝐻 Σ12 .
min tr I𝑁𝑏 + B𝐻 (H𝐻 𝐻
𝑠𝑑 H𝑠𝑑 + H𝑠𝑟 H𝑠𝑟 )B
A (21)
( 𝐻 2 )−1 ]−1 )
−V𝑠 Λ𝑠 A Λ𝑟 A + I𝑅𝑠 𝐻
Λ𝑠 V𝑠 (11) Using (21), (20) is equivalent to
[ ]
1 −2
( ) 1 1 1
s.t. tr A(Λ2𝑠 + I𝑅𝑠 )A𝐻 ≤ 𝑃𝑟 (12) −T𝐻 Σ12 I𝑁𝑏 + Σ12 TT𝐻 Σ12 Σ 2 + T𝐻 = 0. (22)
392 IEEE COMMUNICATIONS LETTERS, VOL. 14, NO. 5, MAY 2010

0
10 SNRrd ≜ 𝜎𝑟2 𝑃𝑟 /𝑁𝑟 , SNRsd ≜ 𝜎𝑑2 𝑃𝑠 /𝑁𝑠 for the source-
−1
relay, relay-destination, and source-destination links, respec-
10 tively. We simulate a scenario where the distance between the
−2
relay and destination nodes is fixed, while the source-relay
10 distance (and thus also the source-destination distance) are
varying. In particular, we set SNRrd = 20dB and SNRsd =
BER

−3
10 SNRsr − 10dB. All simulation results are averaged over 1000
−4
independent channel realizations.
10 We compare the performance of the proposed iterative joint
beamforming (JointBF) algorithm with the suboptimal (Sub-
−5 SubOpt Algorithm
10
RelayBF Algorithm Opt) algorithm in [5] which optimizes only the source-relay-
−6
JointBF Algorithm destination link, and the relay-only beamforming (RelayBF)
10
0 5 10 15 20 algorithm in [6]. For both the JointBF and RelayBF algo-
SNR (dB)
sr
rithms, the projected gradient method is applied to optimize
A in the relay amplifying matrix.
Fig. 1. BER versus SNRsr . 𝑁𝑠 = 𝑁𝑑 = 2, 𝑁𝑟 = 6, SNRrd = 20dB, Fig. 1 shows the performance of three algorithms in terms
SNRsd = SNRsr − 10dB. of bit-error-rate (BER) versus SNRsr using the QPSK con-
stellation. It can be seen that the SubOpt algorithm has the
−1 worst performance, since it does not consider the direct link.
Solving (22) for T, we obtain TT𝐻 = Σ1 2 − Σ−1 1 , which
is valid since the diagonal elements of Σ1 are greater than The JointBF algorithm outperforms the other two algorithms
−1 1 in the whole SNRsr range, since it jointly optimizes the source
1. Consequently, we have T = (Σ1 2 − Σ−1 1 ) , and B =
2

−𝐻 − 1
−1 21 and relay matrices. In fact, it achieves a higher diversity order
M V1 (Σ1 − Σ1 ) . Thus, we prove Theorem 1 with
2
than the other algorithms.
−1 1
D = (Σ1 2 − Σ−1 1 ) .
2 □
From (16) we see that B has a transmit beamforming struc-
ture, where the directions of the beams are given by M−𝐻 V1 , V. C ONCLUSION
while D represents the power allocation at each beam. The un- In this letter, we have proved the optimal beamforming
known Lagrange multipliers 𝜇1 and 𝜇2 in (16) can be obtained structure of the source precoding matrix for non-regenerative
by solving the dual optimization problem associated with the MIMO relay systems with the direct source-destination link.
problem (13)-(15) as explained in the following. Substituting An iterative joint source and relay beamforming algorithm
−1 ¯
(16) into (17) we have ℒ = tr(2Σ1 2 − Σ−1 1 ) − 𝜇1 𝑃𝑠 − 𝜇2 𝑃𝑟 . is developed to minimize the MSE of the signal waveform
The dual optimization problem of the original problem (13)- estimation.
(15) is given by
− 12 ¯
max tr(2Σ1 − Σ−1
1 ) − 𝜇1 𝑃𝑠 − 𝜇2 𝑃𝑟 . (23) R EFERENCES
𝜇1 ≥0,𝜇2 ≥0

Since the dual problem is convex, the problem (23) can be [1] X. Tang and Y. Hua, “Optimal design of non-regenerative MIMO
efficiently solved by the interior-point method [10]. wireless relays,” IEEE Trans. Wireless Commun., vol. 6, pp. 1398-1407,
Apr. 2007.
We have proved that in order to jointly minimize tr(E), [2] O. Muñoz-Medina, J. Vidal, and A. Agustı́n, “Linear transceiver design
both the optimal B and F have a general beamforming in nonregenerative relays with channel state information,” IEEE Trans.
structure as given in (10) and (16), respectively. Exploiting Signal Process., vol. 55, pp. 2593-2604, June 2007.
[3] W. Guan and H. Luo, “Joint MMSE transceiver design in non-
this optimal structure, the problem (5)-(7) can be solved by regenerative MIMO relay systems,” IEEE Commun. Lett., vol. 12,
an iterative
√ algorithm. This algorithm is first initialized at pp. 517-519, July 2008.
B = 𝑃𝑠 /𝑁𝑠 I𝑁𝑠 . Then F is updated by solving the problem [4] A. S. Behbahani, R. Merched, and A. M. Eltawil, “Optimizations of a
MIMO relay network,” IEEE Trans. Signal Process., vol. 56, pp. 5062-
(11)-(12) with a fixed B, and B is updated by solving the 5073, Oct. 2008.
problem (13)-(15) with a given F. The updating of F and B [5] Y. Rong, X. Tang, and Y. Hua, “A unified framework for optimizing lin-
is operated in an alternating fashion. Note that the conditional ear non-regenerative multicarrier MIMO relay communication systems,”
IEEE Trans. Signal Process., vol. 57, pp. 4837-4851, Dec. 2009.
updates of F and B may either decrease or maintain but cannot [6] Y. Rong and F. Gao, “Optimal beamforming for non-regenerative MIMO
increase the objective function tr(E). Monotonic convergence relays with direct link,” IEEE Commun. Lett., vol. 13, pp. 926-928,
of F and B follows directly from this observation. Dec. 2009.
[7] Y. Rong, “MMSE-based non-regenerative multicarrier MIMO wireless
relay communications with direct source-destination link,” in Proc. IEEE
IV. N UMERICAL E XAMPLE Int. Conf. Commun., Dresden, Germany, June 2009.
We simulate a MIMO relay system with 𝑁𝑠 = 𝑁𝑑 = 2 [8] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Englewood Cilffs, NJ: Prentice Hall, 1993.
and 𝑁𝑟 = 6. All channel matrices have Gaussian entries [9] D. P. Bertsekas, Nonlinear Programming, 2nd Edition. Belmont, MA:
with zero-mean and variances 𝜎𝑠2 /𝑁𝑠 , 𝜎𝑟2 /𝑁𝑟 , 𝜎𝑑2 /𝑁𝑠 for Athena Scientific, 1995.
H𝑠𝑟 , H𝑟𝑑 , and H𝑠𝑑 , respectively. Consequently, the signal- [10] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Cambridge University Press, 2004.
to-noise ratios (SNRs) are defined as SNRsr ≜ 𝜎𝑠2 𝑃𝑠 /𝑁𝑠 ,