1188 IEEE COMMUNICATIONS LETTERS, VOL. 17, NO.

6, JUNE 2013

Optimal Relay Selection and Beamforming in

MIMO Cognitive Multi-Relay Networks
Quanzhong Li, Qi Zhang, Renhai Feng, Liping Luo, and Jiayin Qin

Abstract—For the non-regenerative multiple-input multiple- We assume that each of the SUs and PUs is equipped with
output (MIMO) cognitive multi-relay network, we propose an single antenna and each relay has multiple antennas. The
optimal relay selection and beamforming scheme which maxi- scenario is typical for device-to-device communications [8]
mizes the capacity of the secondary user by selecting the best
cognitive MIMO relay subject to the transmit power constraints where two mobile phones in a multi-cell environment commu-
at the relays and the interference power constraints at the nicate directly with the help of multiple mobile devices, such
primary users. In our proposed scheme, the relay selection and as laptops. In conventional cooperative transmission without
beamforming optimization problem is solved separately by em- spectrum sharing, the MIMO relay network where each of
ploying convex semidefinite programming (SDP) through rank- the SUs is equipped with single antenna and each relay is
one relaxation and Charnes-Cooper transformation. It is shown
that our proposed scheme achieves higher average capacity than equipped with multiple antennas was also studied in [9]-[10].
the convectional relay beamforming schemes and relay selection The joint relay selection and beamforming scheme for the
schemes. relay network where the source is equipped with multiple
Index Terms—Cognitive radio, spectrum sharing, multiple- antennas and each relay has single antenna was also studied
input multiple-output (MIMO), relay selection, beamforming. in [11].
For non-regenerative MIMO cognitive multi-relay networks,
I. I NTRODUCTION we propose the optimal relay selection and beamforming
scheme which maximizes the capacity of the SU by selecting

T HE spectrum sharing multiple-input multiple-output

(MIMO) cognitive relay networks, where the secondary
user (SU) communications are assisted by the MIMO relay,
the best cognitive MIMO relay subject to the transmit power
constraints at the relays and the interference power constraints
at the PUs. This relay selection and beamforming optimization
have raised great interest due to its capability to improve is solved in two steps. In the first step, the optimal relay beam-
the performance of the SU significantly [1]-[3]. In [1]-[2], forming matrices are carefully designed and in the second
the relay beamforming scheme for non-regenerative MIMO step, the relay with the largest capacity employing the opti-
cognitive relay networks was studied to maximize the system mal beamforming is selected. The beamforming optimization
capacity. In [3], the relay beamforming scheme was studied problem for each relay is transformed into a convex semidef-
to minimize the sum of mean square error (MSE) of SUs. inite programming (SDP) by using the rank-one relaxation
The above-mentioned works consider the cognitive single and the Charnes-Cooper transformation. To recover the rank-
relay networks. It is well known that the multiple relays one beamforming solution for the selected relay, the relay
provide more diversity gain than the single relay for the transmit power minimization problem is considered, whose
SU in the cognitive relay networks. For cognitive multi-relay rank-one solution is also optimal to the original beamforming
networks, relay selection was widely studied in [4]-[7] for optimization problem. Simulation results are provided to show
one-way and two-way cognitive communications. The works that our proposed scheme achieves higher average capacity
in [4]-[7] consider the situation that each of the multiple relays than the convectional relay beamforming schemes and relay
is deployed with a single antenna. The research on the relay selection schemes.
selection and beamforming for a MIMO cognitive multi-relay Notations: The † and ∗ denote the conjugate transpose and
network is missing. conjugate operation, respectively. The ⊗ denotes Kronecker
In this letter, we consider a non-regenerative MIMO cogni- product. The vec(·) denotes to stack the columns of a matrix
tive multi-relay network, where two SUs and multiple relays into a vector. The || · || denotes Frobenius norm. The tr(·) and
share the same spectrum with multiple primary users (PUs). rank(·) denote the trace and rank of a matrix, respectively.
Manuscript received March 5, 2013. The associate editor approving it for
publication was N.-D. Dao.
This work was supported by the National Natural Science Foundation of P.
II. S YSTEM M ODEL
R. China (61173148 and 61202498), the Industry-University-Research Project We consider a non-regenerative MIMO cognitive multi-
of Guangdong Province and the Ministry of Education (2011B090400581),
the Scientific and Technological Project of Guangzhou City (12C42051578 relay network, where a pair of SU transceivers and K cog-
and 11A11060133). nitive relays are allowed to share the same spectrum with M
Q. Li, Q. Zhang, R. Feng, and J. Qin are with the School of primary users (PUs). Each of the SUs and PUs is equipped
Information Science and Technology, Sun Yat-Sen University, Guangzhou
510006, Guangdong, China (e-mail: [email protected], with a single antenna and each relay is equipped with N
[email protected], [email protected], [email protected]). antennas. We assume that there is no direct link between the
L. Luo is with the College of Information Science and Engineering, SU transmitter (SU-Tx) and the SU receiver (SU-Rx). We also
Guangxi University for Nationalities, Nanning 530006, Guangxi, China (e-
mail: [email protected]). assume that the system operates in the time division duplex
Digital Object Identifier 10.1109/LCOMM.2013.051313.130496 (TDD) mode. During the first time slot, the SU-Tx transmits
1089-7798/13$31.00 
c 2013 IEEE
LI et al.: OPTIMAL RELAY SELECTION AND BEAMFORMING IN MIMO COGNITIVE MULTI-RELAY NETWORKS 1189

signals to the relays and during the second time slot, the The problem (6) is a joint optimization of the relay selection
selected kth relay, k ∈ {1, 2, · · · , K}, multiplies the received and the beamforming matrices of all the relays. Similar to
signals by a beamforming matrix and forwards them to the [4]-[5], the joint optimization problem (6) can be solved in
SU-Rx. The received signals at the SU-Rx is expressed as1 two steps. The first step is to design the optimal beamforming
matrices of all the relays, and the second step is to find the
y = h†2k Fk h1k x + h†2k Fk nr + z (1) optimal relay. Accordingly, we first solve the beamforming
where x ∈ C1×1 denotes the transmit signal at the SU- optimization problem at each relay, which is equivalently
Tx with E[|x|2 ] = Ps ; h1k ∈ CN ×1 and h2k ∈ CN ×1 expressed as
 
denote the channel response from the SU-Tx to the kth f † Ps hh† f
relay and that from the kth relay to the SU-Rx, respectively; max  
f
f † σr2 H2 H†2 f + σd2
nr ∼ CN (0, σr2 I) denotes the additive white Gaussian noise  
vector received at the kth relay; z ∼ CN (0, σd2 ) denotes the s.t. f † Ps H1 H†1 + σr2 I f ≤ PR (7)
AWGN received at the SU-Rx; Fk ∈ CN ×N denotes the  
linear beamforming matrix at the kth relay. For simplicity, f † Ps G1m G†1m + σr2 G2m G†2m f ≤ Im ,
the subscript k in the matrix Fk is omitted in the following
m ∈ {1, 2, · · · , M }
letter. The received signal-to-noise ratio (SNR) at the SU-Rx
is expressed as where f = vec(F), h = h∗1k ⊗ h2k , H1 = h∗1k ⊗ I, H2 =
I ⊗ h2k , G1m = h∗1k ⊗ g2m , and G2m = I ⊗ g2m .
Ps |h†2k Fh1k |2 The problem (7) is a fractional quadratically constrained
SNR = . (2)
σr2 ||h†2k F||2 + σd2 quadratic problem (QCQP) with multiple quadratically con-
straints, which is difficult to solve. To solve the problem
The transmit power of the kth relay is expressed as
(7), we transform it to the following fractional semidefinite
PRk = Ps ||Fh1k ||2 + σr2 ||F||2 (3) programming (SDP),
tr(A1 W)
which is limited by the maximum allowable transmit power max
W0 tr(A2 W) + σd2
of the relay, denoted as PR . (8)
To protect the communication of the PU, we consider that s.t. tr(A3 W) ≤ PR
the interference power at the mth PU from the SU-Tx and the tr(Bm W) ≤ Im , m ∈ {1, 2, · · · , M }
kth relay should not exceed a given maximum tolerable level
where A1 = Ps hh† , A2 = σr2 H2 H†2 , A3 = Ps H1 H†1 + σr2 I,
Im . Thus, in the first time slot, the interference power at the
mth PU raised by the SU-Tx is limited by Bm = Ps G1m G†1m + σr2 G2m G†2m , W = ff † and the rank-
one constraint rank(W) = 1 has been relaxed.
Ps |g1m |2 ≤ Im , m ∈ {1, 2, · · · , M } (4) Using the Charnes-Cooper transformation [12], the frac-
tional SDP (8) can be further converted into a convex SDP.
where g1m ∈ C1×1 denotes the channel response from the Let W = S/ν and tr(A2 W) + σd2 = 1/ν. The problem (8)
SU-Tx to the mth PU. From (4), it is determined
 that the  is equivalently transformed into
transmit power of the SU-Tx is Ps = min Ps , min |gI1m m
m | 2
max tr(A1 S)
where Ps is the maximum allowable transmit power of the S0,ν≥0
SU-Tx. In the second time slot, the interference power at the s.t. tr(A2 S) + σd2 ν = 1 (9)
mth PU raised by the kth relay is limited by tr(A3 S) ≤ νPR
† † tr(Bm S) ≤ νIm , m ∈ {1, 2, · · · , M }
Ps |g2m Fh1k |2 + σr2 ||g2m F||2 ≤ Im (5)
The problem (9) is a convex SDP, which can be solved
where g2m ∈ CN ×1 denotes the channel response from the
effectively using the interior point method [13]. The last
kth relay to the mth PU.
and most important issue is whether the rank-one relaxation
problem (9) is tight to the original optimization problem (7),
III. O PTIMAL R ELAY S ELECTION AND B EAMFORMING i.e., whether the SDP (9) has an optimal rank-one solution.
The optimization problem of relay selection and beam- Unfortunately, the existence of the rank-one solution to the
forming for a non-regenerative MIMO cognitive multi-relay SDP (9) is still unknown. In the following, we will propose a
network is formulated as, method to find this rank-one solution. We consider the relay
  transmit power minimization problem, from (8), which is
1 Ps |h†2k Fh1k |2
arg max max log2 1 + min tr(A3 W)
k F 2 σr2 ||h†2k F||2 + σd2 W0

s.t. Ps ||Fh1k ||2 + σr2 ||F||2 ≤ PR (6) tr(A1 W)

s.t. ≥ γ − Δγ (10)
† † tr(A2 W) + σd2
Ps |g2m Fh1k |2 + σr2 ||g2m F||2 ≤ Im ,
tr(Bm W) ≤ Im , m ∈ {1, 2, · · · , M }
k ∈ {1, 2, · · · , K}, m ∈ {1, 2, · · · , M }
where γ is the optimal objective function value of the problem
1 Here we adopt the same assumption in [1-2, 4, 7] that the interference (9) and Δγ is a non-negative constant with 0 ≤ Δγ < γ. Thus,
from the PU transmitter is neglected. we have the following result:
1190 IEEE COMMUNICATIONS LETTERS, VOL. 17, NO. 6, JUNE 2013

Proposition 1: If Wo is the optimal solution to the problem

3
(10), then Wo can achieve at least (1 − Δγ γ )-optimal to the ORSB, K=3, N=3
problem (8). Especially, when Δγ = 0, Wo is also optimal MRR−OPMRT, K=1, N=9
MRR−OPMRT, K=3, N=3
to the problem (8). 2.5
MRR−MRT, K=1, N=9
Proof: Suppose that W1 and W2 are the optimal solutions MRR−MRT, K=3, N=3

Average capacity (bps/Hz)

to the problem (8) and (10), respectively. It is easy to see that
W1 is feasible to the problem (10). Then we have 2

tr(A3 W2 ) ≤ tr(A3 W1 ) ≤ PR (11)

1.5

where the second inequality comes from the fact that W1 is

the optimal solution to the problem (8). From (11), we note
that W2 is feasible to the problem (8), which implies that 1

tr(A1 W2 )
≤ γ. (12)
tr(A2 W2 ) + σd2 0.5
0 2 4 6 8 10 12 14 16 18 20
P /σ2 (dB)
Since W2 is the optimal solution to (10), we also have R

tr(A1 W2 )
≥ γ − Δγ (13) Fig. 1. Average capacity versus the ratio of maximum allowable transmit
tr(A2 W2 ) + σd2 power of the relay to the noise power, PR /σ2 ; average capacity comparison
of our proposed ORSB scheme, the MRR-OPMRT scheme in [2], the MRR-
Combining (12) and (13), we obtain MRT scheme in [2] where M = 2, I1 /σ2 = I2 /σ2 = 0 dB.

tr(A1 W2 )
γ − Δγ ≤ ≤γ (14)
tr(A2 W2 ) + σd2
which indicates that W2 can achieve at least (1− Δγ γ )-optimal scheme, the maximal ratio reception-orthogonally projected
tr(A1 W2 )
to the problem (8). When Δγ = 0, we have tr(A2 W2 )+σ2 = γ, maximum ratio transmission (MRR-OPMRT) scheme in [2],
d
that is, W2 is also the optimal solution to the problem (8).  the maximal ratio reception-maximum ratio transmission
From Proposition 1, we know that if the SDP (10) has an (MRR-MRT) scheme in [2]. For the MRR-OPMRT scheme,
optimal rank-one solution, the SDP (9) combining with (10) the beamforming matrix is given by F = α(I − VV † )h2k h†1k
is tight to the original optimization problem (7). where α is employed to satisfy the constraints in (6)
Proposition 2: For any small positive value of Δγ, the and V is from the singular value decomposition (SVD),
† † †
optimal solution to the SDP (10) is rank-one. [g21 ; g22 ; · · · ; g2M ] = UΣV† . For the MRR-MRT scheme,
Proof: See Appendix A.  the beamforming matrix is given by F = αh2k h†1k . For the
From Proposition 2, we know that the optimal rank-one MRR-OPMRT and MRR-MRT schemes, we select the relay
solution to the SDP (10) with any small positive Δγ can which achieves the largest average capacity when K > 1.
be obtained. Thus, according to Proposition 1, the optimal From Fig. 1, it is found that when K = 3 and N = 3,
rank-one solution to the original optimization problem (7) is our proposed ORSB scheme has substantial capacity improve-
achieved when Δγ → 0. ments over the MRR-OPMRT and MRR-MRT schemes. Since
It is noted that the complexity to solve the SDP (9) is the MRR-OPMRT and MRR-MRT schemes are proposed for
O(N 6.5 ) [14]. For the K relays, the complexity of our MIMO cognitive single-relay networks, when the number of
proposed optimal relay selection and beamforming scheme is available relaying antennas is the same, i.e., KN is the same,
O(KN 6.5 ). Therefore, our proposed scheme is suitable for we compare our proposed ORSB scheme where K = 3 and
the situation when the number of antennas in each relay, N , N = 3 with the MRR-OPMRT and MRR-MRT schemes that
is not large. The complexity of our proposed scheme increases all the available antenna are equipped at the single relay. It
linearly with the increase of the number of relays. is found that at high PR /σ 2 , our proposed ORSB scheme
outperforms the MRR-MRT scheme. When PR /σ 2 = 20 dB,
IV. S IMULATION R ESULTS the performance gap between our proposed ORSB scheme and
the MRR-OPMRT scheme is only about 0.4 bps/Hz.
In this section, we present the computer simulation results
of our proposed optimal relay selection and beamforming In Fig. 2, we present the average capacity comparison of our
scheme. We assume that in the MIMO cognitive multi-relay proposed optimal relay selection and beamforming (ORSB)
network, all the channel responses are independent and iden- scheme, the joint relay selection and power allocation (JRS-
tically distributed (i.i.d) complex Gaussian random variables PA) scheme in [4] and the relay selection (RS) scheme in [6]
with zero-mean and unit variance. We also assume that the when the number of available relaying antennas is the same.
noise variances σr2 = σd2 = σ 2 . The ratio of the maximum For our proposed ORSB scheme, K = 5 and N = 3. For the
allowable transmit power at the SU-Tx to the noise power, JRS-PA and RS schemes, K = 15 and N = 1. From Fig. 2, it
Ps /σ 2 , is 10 dB. In all simulations, the average capacity is is observed that our proposed ORSB scheme has substantial
plotted with 1000 randomly generated channel realizations. capacity improvements over the JRS-PA and RS schemes. This
In Fig. 1, we present the average capacity comparison of our is because that by exploiting the multiple antennas at the relay,
proposed optimal relay selection and beamforming (ORSB) the average capacity can be improved significantly.
LI et al.: OPTIMAL RELAY SELECTION AND BEAMFORMING IN MIMO COGNITIVE MULTI-RELAY NETWORKS 1191

are expressed as
3
ORSB, K=5, N=3 ∂L/∂W = 0 (17a)
JRS−PA, K=15, N=1
2.5 RS, K=15, N=1 QW = 0 (17b)
W 0, Q 0, μ ≥ 0, λm ≥ 0, ∀m. (17c)
Average Capacity (bps/Hz)

2
From (16) and (17a), we have
M
1.5 Q = A3 + μγ̂A2 + λm Bm −μA1 .
m=1 (18)
Θ
1
It is noted that Θ is a positive-definite matrix which has full
0.5
rank, i.e., rank(Θ) = N 2 . Since A1 = Ps hh† , we have
   
N 2 = rank Q + μPs hh† ≤ rank(Q) + rank μPs hh†
0
0 5 10 15 20 ⇒ rank(Q) ≥ N 2 − 1. (19)
P /σ2 (dB)
R
2 2
From (19), the rank of Q is either N or N −1. If rank(Q) =
Fig. 2. Average capacity versus the ratio of maximum allowable transmit N 2 , W = 0 from (17b). However, if W = 0, γ̂σd2 ≤ 0
power of the relay to the noise power, PR /σ2 ; average capacity comparison from (15) which violates the assumption that the SDP (15) is
of our proposed ORSB scheme, the JRS-PA scheme in [4] and the RS scheme
in [6] where M = 2, I1 /σ2 = I2 /σ2 = 0 dB. feasible. Thus, rank(Q) = N 2 − 1. The (17b) is satisfied only
when W lies in the nullspace of Q whose dimension is one.
Thus, the optimal W to the problem (10) is rank-one.
V. C ONCLUSIONS
In this letter, we have proposed the optimal relay selection R EFERENCES
and beamforming scheme for the non-regenerative MIMO [1] Q. Li, L. Luo, and J. Qin, “Optimal relay precoder for non-regenerative
MIMO cognitive relay systems with underlay spectrum sharing,” Elec-
cognitive multi-relay network which maximizes the capacity tron. Lett., vol. 48, no. 5, pp. 295–297, Mar. 2012.
of the SUs by selecting the best cognitive MIMO relay [2] K. Jitvanichphaibool, Y. C. Liang and R. Zhang, “Beamforming and
subject to the transmit power constraints at the relays and the power control for multi-antenna cognitive two-way relaying,” in Proc.
2009 IEEE WCNC, pp. 1–6.
interference power constraints at the PUs. Simulation results [3] H. Mu and J. K. Tugnait, “MSE-based source and relay precoder design
have shown that our proposed scheme achieves higher average for cognitive radio multiuser two-way relay systems,” in Proc. 2012 IEEE
capacity than the convectional relay beamforming schemes WCNC, pp. 742–747.
[4] L. Li, X. Zhao, H. Xu, G. Y. Li, D. Wang, and A. Soong, “Simplified relay
proposed for the cognitive single relay network. It is also selection and power allocation in cooperative cognitive radio systems,”
shown that when the number of available relaying antennas is IEEE Trans. Wireless Commun., vol. 10, no. 1, pp. 33–36, Jan. 2011.
the same, our proposed scheme outperforms the convectional [5] P. Ubaidulla and S. Aı̈ssa, “Optimal relay selection and power allocation
for cognitive two-way relaying networks,” IEEE Wireless Commun. Lett.,
relay selection schemes. vol. 1, no. 3, pp. 225–228, Jun. 2012.
[6] Y. Zou, J. Zhu, B. Zheng, and Y.-D. Yao, “An adaptive cooperation
A PPENDIX A: P ROOF OF P ROPOSITION 2 diversity scheme with best-relay selection in cognitive radio networks,”
IEEE Trans. Signal Process., vol. 58, no. 10, pp. 5438–5445, Oct. 2010.
We rewrite the SDP (10) as [7] Y. Li and A. Nosratinia, “Spectrum sharing with distributed relay selec-
tion and clustering,” IEEE Trans. Commun., vol. 61, no. 1, pp. 53–62,
min tr(A3 W) Jan. 2013.
W0 [8] P. Janis, C.-H. YU, K. Doppler, C. Ribeiro, C. Wijting, K. Hugl, O.
s.t. γ̂tr(A2 W) − tr(A1 W) + γ̂σd2 ≤ 0 (15) Tirkkonen, and V. Koivunen, “Device-to-device communication under-
laying cellular communications systems,” Int. J. Commun., Network and
tr(Bm W) − Im ≤ 0, m ∈ {1, 2, · · · , M } System Sciences, doi: 10.4236/ijcns.2009.23019, pp. 169–247, 2009.
[9] C. Li, L. Yang, and W.-P. Zhu, “Two-way MIMO relay precoder design
where γ̂ = γ − Δγ. The problem (15) is a convex SDP and it with channel state information,” IEEE Trans. Commun., vol. 58, no. 12,
is verified that the problem (15) satisfies Slater’s constraint pp. 3358–3363, Dec. 2010.
[10] B. K. Chalise and L. Vandendorpe, “Optimization of MIMO relays for
qualification condition [13]. Thus, the strong duality holds multipoint-to-multipoint communications: nonrobust and robust designs,”
and the Karush-Kuhn-Tucker (KKT) conditions can be used IEEE Trans. Signal Process., vol. 58, no. 12, pp. 6355–6368, Dec. 2010.
to prove Proposition 2. The Lagrangian function of the SDP [11] A. Yang, C. Xing, Z. Fei, and J. Kuang, “On the performance of
joint relay selection and beamforming with limited feedback for AF
(15) is cooperative networks,” in Proc. 2011 IEEE VTC – Fall, pp. 1–5.
M [12] A. Charnes and W. W. Cooper, “Programming with linear fractional
L =tr(A3 W) + λm (tr(Bm W) − Im ) functionals,” Naval Res. Logist. Quarter., vol. 9, pp. 181–186, 1962.
m=1 (16) [13] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge Uni-
+ μ(γ̂tr(A2 W) − tr(A1 W) + γ̂σd2 ) − tr(WQ) versity Press, 2004.
[14] A. Ben-Tal and A. Nemirovski, Lectures on Modern Convex Optimiza-
where μ ≥ 0, λm ≥ 0, and Q 0 are the Lagrangian tion: Analysis, Algorithms, and Engineering Applications. Ser. MPS-
dual variables. The KKT conditions related with the proof SIAM Series on Optimization, 2001.