1
Precoding for MIMO Channels in Cognitive Radio
Networks With CSI Uncertainties and for
Compound Capacity
Mohannad H. Al-Ali, Student Member, IEEE and K. C. Ho, Fellow, IEEE
Abstract—Inaccurate knowledge of channel state information
(CSI) limits the performance offered by MIMO communications.
The design of a MIMO precoder should take the CSI uncertainty
into consideration to mitigate its effects. This paper proposes
a method to design the precoder for a secondary user in an
underlay cognitive radio framework that maximizes its trans-
mission performance, where the direct link to its receiver and
the interference link to a primary user have CSI uncertainties.
We model the CSI uncertainties through the Schatten norm
that encompasses most commonly used norm measures such
as the spectral and Frobenious norms. The proposed method
solves iteratively a minimax problem by deriving the optimal
solution for the worst case interference CSI uncertainty, applying
the alternate-iterate technique for the worst case direct link
CSI deviation, and developing a feasible direction projected
subgradient technique for the precoder. Simpler solutions for the
precoder are also derived under some specific norm measure
of CSI uncertainty and certain requirement of transmission
power. Simulations corroborate the expected performance of the
proposed design.
Index Terms—Cognitive radio, compound capacity, minimax,
MIMO, Schatten norm, alternate-iterate, Lagrange dual, leakage
rate.
I. I NTRODUCTION
T HE exploitation of multiple-input multiple-output
(MIMO) techniques for increasing data throughput
often depends on the quality of channel state information
(CSI) available. A MIMO transmitter can utilize a precoder
devised from the available CSI for sending signals over more
favorable spatial directions to increase the spectral efficiency
and ensure the quality of service (QoS) to the end users
[1]. In practice uncertainty appears in the acquired CSI [2],
rendering the expected performance offered by MIMO. It is
necessary to take the uncertainty of CSI into consideration
when designing the precoder, to mitigate the damage it can
cause [2]–[27].
This paper considers an underlay cognitive radio (CR) point-
to-point MIMO system in which a secondary user (SU) seeks
communication in the presence of a primary user (PU) where
its QoS should be assured. The available CSI of the direct link
between SU-Transmit (SU-Tx) and SU-Receive (SU-Rx) and
the interference link between SU-Tx and PU-Receive (PU-Rx)
is not accurate. The CSI error in the direct link is attributed
The authors are with the Electrical Engineering and Computer Science
Department, University of Missouri-Columbia, Columbia, MO 65211 USA.
e-mail: mha6fd@mail.missouri.edu; hod@missouri.edu
This work was supported in part by the Higher Committee for Education
Development (HCED) in Iraq.
to the short coherence time, limited feedback or significant
demand of spectrum resources by PUs [25]. The uncertainty
in interference CSI results from the non-cooperative nature
of PU with SU since both users often belong to different
networks [9]–[27]. The objective is to design a linear precoder
for SU-Tx to improve the SU transmission performance that
is robust to the CSI uncertainties, while ensuring the QoS for
PU. The underlay CR paradigm is less demanding in managing
interferences than some other approaches, such as interference
alignment [18], [19] which requires the availability of CSI for
all link connections and a precoder at each transmitter and a
combiner at each receiver ends.
The uncertainty of CSI can be modelled in a stochastic
or a deterministic manner. Under the stochastic model the
robust design seeks a precoder that optimizes a stochastic
performance metric with probabilistic constraints [21]–[26].
The stochastic approach is usually not amenable in opti-
mization because of the lack of closed-form expressions for
the CSI probability distributions. The deterministic model
uses a quantitative measure together with a compact set
[2]–[17], whose size and shape are usually evaluated from
field measurements, to encapsulate the CSI uncertainties. The
design looks for a precoder under the worst scenario over the
uncertainty sets, leading to a rather involved maximin (or the
equivalent minimax) optimization problem. This work uses a
deterministic model for the uncertainties and provides a novel
solution to the precoder.
Essentially, we are interested in searching a precoder for
Gaussian modelled data that maximizes the mutual informa-
tion (MI) of SU, under its most conservative performance
caused by the channel uncertainties in the SU intended connec-
tion and in the SU to PU interference link. The optimization is
non-concave over the maximization on the precoder and non-
convex over the minimization of the SU’s MI, although it is
concave over the maximization of the amount of interference
to PU. Such a mixed optimization problem does not lend
itself to apply the existing techniques in [2], [3] or the matrix
inequalities introduced in [5], [6]. This paper formulates the
problem as a continuous maximin optimization [28] and solves
one variable at a time while holding the others fixed. We first
derive an explicit solution of the worst case CSI deviation
that would cause the largest amount of interference from SU
to PU. Next, by introducing a nuisance parameter, we obtain
the CSI uncertainty value that yields the least MI at a given
precoder through the alternate-iterate (AI) technique [29].
Finally, the precoder is then updated according to the feasible
2
direction determined by the projection of a subgradient [30].
The optimization process repeats until a stationary solution is
reached. We support the proposed method by providing the
convergence analysis for reaching a stationary point.
It is critical to maintain the QoS of PU in CR. Such a
constraint, nevertheless, creates a non-convex feasible set for
the precoder that would complicate the optimization. To handle
this difficulty, we express the QoS measure in the interference
constraint through the Taylor-series and maintain up to the
first order. Such an expansion provides more conservative re-
quirement on the QoS. As the optimization cycles through, the
linear approximation would be more accurate and eventually
represents the actual constraint at a stationary solution.
The CR problem we are addressing encompasses another
interesting problem in the literature in which PU is absent
that the QoS constraint is not needed. This is termed as
the compound capacity (CC) problem [2]–[6]. The proposed
solution is applicable to the CC problem as well except the
optimization of CSI deviation on the interference link is not
necessary. The optimization of the precoder is in fact simpler
as its feasible set is convex.
We would like to highlight the contributions of our work.
First, to the best of our knowledge, the proposed algorithms
for solving the CR problem, including the CC problem as
well, are new. Most of the related works from the literature in
robust optimization are for unconstrained maximin problems
[28]. Recently, [31] took the interior-point methods with
semi-infinite programming iterative steps and subgradient de-
scent directions to address the constrained minimax problem.
However, the iterative algorithm is applicable for equality
constrained minimax problems and requires the functions
in the equality constraints twice differentiable. Furthermore,
it assumes there is sufficient number of solutions at each
iteration for the inner non-concave maximization to obtain
a descent direction. The CR problem investigated in [16],
[17] assumes the SU direct link is exactly known, which is a
simpler problem and does not require maximin optimization.
Moreover, the work in [26] addresses the power allocation
problem in CR with a single antenna at SU-Rx and at PU-Rx,
under an outage probability criterion for the QoS of PU.
Second, most studies in the literature uses a specific norm,
such as the spectral norm, to define the CSI uncertainty set for
simplifying the solution finding, and consequently yielding a
limited result to the problem. The work here uses the general
matrix norm called the Schatten norm [7], [16] that comprises
a number of frequently used matrix norms such as the spectral
and Frobenious norms. We also do not restrict the transmit
power constraint to be on the total power. As a result the
proposed solution is more general and applicable to wider
situations, including those from [2]–[6] for the CC problem
that PU is absent and the QoS requirement is not needed.
Third, through the Lagrange dual and the Hölder inequality,
we obtain a suboptimal solution for the worst case direct
link CSI uncertainty in the AI process. This solution is more
attractive than solving the constrained problem directly using
the interior point methods, while maintaining comparable
performance from our observations.
Finally, we derive the optimal structure of the CR precoder
matrix when the uncertainty sets are defined by the spectral
norm and the transmit power requirement is always fulfilled.
Furthermore, we prove for the CC problem that the optimal
precoder has eigen-directions equal to those of the available
CSI matrix for the uncertainty set defined by the Schatten
norm, when the power set is unitarily-invariant. Such a result
was shown for the spectral norm defined uncertainty set only
under unitarily-invariant power set [5], [6].
This paper is organized as follows. The system model and
problem formulation are introduced in Section II. Section III
develops an algorithm to solve the CR problem and proves
the convergence of the proposed solution method. Section IV
briefly summarizes the reduced solution to the CC problem. In
Section V, particular solutions for the CR and CC problems
are established. Section VI provides the numerical results and
Section VII gives the conclusion.
Notation: We denote matrices and column vectors by upper-
case and lower-case letters in bold. We mean by Tr(A),
|A|, [A]ij and λi (A) the trace, determinant, ij-th element
and i-th eigenvalue of A. A  0, A  0 and A  0
indicate A is positive semi-definite (PSD), positive definite
(PD) and negative-semidefinite (NSD). The superscripts (·)T
and (·)† are the regular transpose and the Hermitian transpose.
Rm×n and Cm×n are the real and complex spaces of m × n
matrices. diag(a) is a diagonal matrix formed by the elements
of the vector a and vec(A) = [aT1 · · · aTm ]T is a collection
of the√columns of A = [a1 , · · · , am ]. kak is the
√ Euclidean
norm a† a and kakX is the weighted version a† Xa for a
Hermitian and PSD matrix X of appropriate size. We define
the eigenvalue-decomposition (EVD) of the Hermitian matrix
A ∈ Cn×n as A = UA ΛA U†A , where UA ∈ Cn×n is
a unitary matrix and ΛA = diag([λ1 (A) · · · λn (A)]T ) with
the eigenvalues arranged in non-increasing order. I is an
identity matrix of appropriate size. E[·] denotes the statistical
expectation and | · | is the absolute value of a scalar.
II. S YSTEM M ODEL AND A SSUMPTIONS
Let us consider the scenario as shown in Fig. 1 in which
an SU communicates through the MIMO channel H22 ∈
CN2 ×M2 in the presence of a PU, where M2 and N2 are the
number of SU transmit and receive antennas. The signal from
SU interferes with PU through the channel H21 ∈ CN1 ×M2 ,
where N1 is the number of PU-Rx antennas. The channels
are not known exactly due to limited feedback [4], [25]
and the lack of coordination between PU and SU [9]–[27].
The objective is to obtain a linear precoder for SU that
maximizes the MI for its transmission where the available
channel information has uncertainty, subject to the requirement
of maintaining the QoS of PU.
It is more direct here to represent the CSI in terms of
the covariance matrices R2j , j = 1, 2. It is related to the
†
MIMO channels i R2j = H2j H2j for slow fading and
by
h
R2j = E H†2j H2j for fast fading. The available CSI at SU-
Tx is R̂2j . It is not accurate and is different from the actual
by
R̂2j = R2j + ∆2j , j = 1, 2 . (1)

M2 N2
ˆ= R + ∆
R
SU-Transmit
22 22 22
SU-Receive
H 22
N1
PU-Receive
Fig. 1. An underlay MIMO CR network with single SU and single PU. CSI
for the SU direct link and for the interference link from SU to PU is not
accurate.
∆2j ∈ CM2 ×M2 represents the error in acquiring CSI.
We model the channel acquisition error as deterministic
using the non-empty compact and convex set Uj , j = 1, 2,
whose shape is defined by the Schatten norm1 [7], [16],
Uj = {∆2j : k∆2j kSpj ≤ j , R̂2j − ∆2j  0} , j = 1, 2 .
(2)
The parameter j > 0 controls the size of the uncertainty set
and k · kSpj denotes the Schatten norm of order pj ∈ [1, ∞].
The Schatten norm of a matrix A ∈ Cn×m is given by [33,
Proposition 9.2.3]

 Pmin(n,m) p 1/p

σi , 1≤p<∞
kAkSp = i=1 (3)
σ ,
1 p=∞
where σi , i = 1, 2, · · · , min(n, m), are the singular values of
A arranged in non-increasing order. In (2), the requirement
R̂2j − ∆2j  0 ensures the actual CSI R2j is PSD and it
implicitly restricts ∆2j to be a Hermitian matrix.
The linear precoder we would like to obtain is represented
by the matrix Q ∈ CM2 ×M2 . It processes the SU signal
before transmission to improve the performance. The metric
for maximization is
∆
φ(Q, ∆22 ) = log |I + R22 Q| (4a)
= log |I + (R̂22 − ∆22 )Q| (4b)
where log(·) is the natural logarithm. (4b) resembles the MI
of SU. It is indeed the MI for slow fading, and is the upper
bound for the MI under fast fading h wherei R22 is defined
with expectation, i.e., R22 = E H†22 H22 . Please refer to
Appendix A for the details. Precoding optimization based on a
tractable bound of a performance metric is common in CR with
single [23], [24] or multiple antennas [34], [35]. The metric
(4b) defines the information rate over the SU transmission
link and henceforth we shall call φ(·, ·) as the SU rate. The
SU rate in (4) assumes the SU signal follows a Gaussian
distribution [1] so that it is accurate and has a simple form.
In practice the distribution of SU signal may be different. For
finite-alphabet signals such as those created from quadrature
amplitude modulation (QAM), the rate expression should be
defined using the constellation-constraint MI instead [32]. In
such a case, it will lead to a different solution for the precoder.
1 The
developed techniques and solutions in this work are not limited to
the Schatten norm based uncertainty set. They can be easily extended to any
non-empty compact and convex uncertainty set.
3
The precoder matrix Q is not a free variable. It should
be Hermitian and PSD, and should satisfy the power budget
defined by its elements or eigenvalues. We shall use a convex
set Q to represent these requirements together.
The maximization of the SU rate is under the constraint that
the interference caused by SU-Tx to PU-Rx is under a certain
limit. The interference level is quantified by
∆
g(Q, ∆21 ) = log |I + R21 Q| (5a)
= log |I + (R̂21 − ∆21 )Q| . (5b)
The measure (5b) is often referred as the leakage rate [20] and
was recently investigated in [16], [17] for the CR design.
It would be more convenient to use −φ(·, ·) instead of φ(·, ·)
for the development in the sequel. In mathematical form, the
problem we would like to address can be posed as
(P-I): min. max. − φ(Q, ∆22 ) (6a)
Q∈Q ∆22 ∈U2
s.t. max. g(Q, ∆21 ) ≤ rL . (6b)
∆21 ∈U1
Note that (6a) is equivalent to max. min. φ(Q, ∆22 ). We
Q∈Q ∆22 ∈U2
seek a conservative precoder that guarantees performance
improvement as much as possible under the worst transmission
scenario for SU. (6b) ensures the interference tolerance limit
rL > 0 is not exceeded for all possible CSI uncertainties
expected in the interference link.
P-I is a constrained continuous minimax problem and it is
rather difficult to solve for several reasons. First, the objective
function −φ(Q, ∆22 ) is not concave in ∆22 . Second, the
interference constraint (6b) is not convex in Q. Third, the
optimization on the left side of (6b) does not have a closed
form expression in terms of Q, unlike the studies in [14]–[16].
In addition to the CR problem, P-I becomes the CC problem
[2] if the constraint (6b) is absent. That is, we only consider
the transmission of SU without worrying the QoS of PU. The
CC problem has found interests in the research community as
well [2]–[6].
III. C OGNITIVE R ADIO P ROBLEM
We would like to solve P-I over the optimization variables
∆21 , ∆22 and Q. They are coupled together and it is
extremely difficult for joint optimization. Rather, we resort
to the approach similar to coordinate descent by iteratively
optimizing over the variables one by one in sequence; while
optimizing in one variable the others are held fixed at their
latest values. Let us assume that after the previous iteration
we have obtained the precoder solution Q(l) , where l denotes
the iteration count.
A. Solution for ∆21
Utilizing the interference constraint (6b) requires the opti-
mal value of the subproblem
(P-I.1): max. g(∆21 , Q) . (7)
∆21 ∈U1
P-I.1 is concave in ∆21 for a given precoder.
To proceed, we first introduce a proposition for the set of
possible solutions for ∆21 .
4

Proposition 1: For a given precoder matrix Q the optimal
uncertainty matrix ∆21 to the problem P-I.1 must be NSD.
Proof: The key idea for the proof is to show that the function
g(·, ·) in P-I.1 would have a larger value when ∆21 is NSD
than PSD. Please refer to Appendix B for the details.
As a result, R̂21 − ∆21  0 is automatically satisfied and
does not need to be included in the set U1 .
We shall derive an analytic solution of ∆21 for the problem
(7) at Q = Q(l) . Note that ∆21 does not depend on
∆22 explicitly. Let us begin by defining the matrix T(l) =
(l) (l) −1
Q (I + R̂21 Q ) , which is PSD according to Lemma 2
in Appendix C. It is direct to verify the subproblem P-I.1 in
(7) at Q = Q(l) is equivalent to
(P-I.2): max. log |I − ∆21 T(l) |
−∆21 0
s.t. k∆21 kpS1p ≤ p11 .
1
B. Solution for ∆22
Using a similar procedure as for Proposition 1 in Appendix
B, we can show that ∆22 should be PSD for the solution to
P-I.3. Incorporating this result, the uncertainty set for ∆22 in
 (2) becomes
Ũ2 = {∆22 : k∆22 kSp2 ≤ 2 , R̂22 − ∆22  0, ∆22  0} .
(13)
Given the intermediate solution for the precoder Q(l) , the
∆ objective is to find ∆22 ∈ Ũ2 that maximizes −φ(Q(l) , ∆22 ).
We shall make use of the following lemma to convert −φ(·, ·)
to a different and yet equivalent form to simplify the solution
finding.
Lemma 1: Let A ∈ Cn×n be any PD matrix (not neces-
(8a) sarily Hermitian). Consider the function z(S) = − Tr(SA) +
(8b) log |S| + n, where S ∈ Cn×n is any PD matrix. Then

The solution to P-I.2 is obtained through the EVD rep- max. z(S) = − log |A| . (14)
S0
resentations of T(l) and ∆21 . Let T(l) = UT(l) ΛT(l) U†T(l)
with eigenvalues λ1 (T(l) ) ≥ · · · ≥ λM2 (T(l) ) and ∆21 = The optimization on the left is achieved when S = A−1 .
U∆21 Λ∆21 U†∆21 with eigenvalues −λ1 (∆21 ) ≥ · · · ≥
Proof: It is direct to show (14) as z(S) is concave and
−λM2 (∆21 ). From the Q Hadamard’s inequality [40, Ch. 9,
n n×n differentiable in S. 
B.5] that states |A| ≤ i=1 [A]ii with A ∈ C being
Hermitian and PSD, the objective function (8a) is maximized When using the fact that ∆22 is PSD and applying (14) to
when U∆21 = UT(l) . the function −φ(·, ·) defined in (4b), (11) becomes
Introducing the Lagrange multiplier µ ≥ 0 for the constraint  
max. − Tr Q(l) W2 (R̂22 − ∆22 ) + f (W2 )(15)
(8b), the Lagrange dual for P-I.2 under U∆21 = UT(l) is ∆22 ∈Ũ2 ,W2 0
PM2 (l)


max. i=1 log 1 + |λi (∆21 )|λi (T ) where W2 is a nuisance variable that is PD and
[Λ∆21 ]ii ≥0
f (W2 ) = − Tr(W2 ) + log |W2 | + M2 . (16)
P 
M2 p1 p1
−µ i=1 |λ i (∆21 )| −  1 . (9)
The optimization problem (15) is biconcave, i.e., it is
Taking derivative with respect to λi (∆21 ) and setting it to
concave in ∆22 or W2 but not in both [29]. We shall
zero give
solve it based on the AI approach using the alternate convex

1 1
+ search (ACS) method [29]. In particular, each iteration of ACS
λi (∆21 ) = − − (10) consists of the following two steps:
µp1 |λi (∆21 )|p1 −1 λi (T(l) )
• Fix ∆22 and maximize (15) over W2 . The resulting
for p1 ∈ [1, ∞), where the value x+ = max(x, 0). For a solution for W2 is in closed-form
given µ, we can solve (10) for λi (∆21 ) through a simple line  −1
search.
PMApplying another level of line search yields µ such W2 = I + (R̂22 − ∆22 )Q(l) . (17)
that i=1 2
|λpi 1 (∆21 )| = p11 .
For p1 = ∞, it is straightforward to validate that ∆21 = • Fix W2 at the value obtained in (17) and maximize
−1 I is optimal since the constraint (8b) is equivalent to (15) over ∆22 , which is uniquely determined from the
−∆21  1 I. following subproblem
Putting together the eigenvectors and the eigenvalues, we
(l)
have the solution to P-I.2 and it is denoted as ∆21 . (P-I.5): max. Tr (Q(l) W2 ∆22 ) (18a)
∆22 0
We next obtain the solutions for ∆22 and Q. For ease of
s.t. k∆22 kSp2 ≤ 2 (18b)
illustration, (6a) in the problem P-I is expressed as two sub-
problems R̂22 − ∆22  0 . (18c)
∆ The subproblem (18) has a linear objective function and
(P-I.3): Ψ(Q) = max. − φ(Q, ∆22 ) , (11)
∆22 ∈U2 convex constraints in ∆22 . It can be solved using the
(P-I.4): min. Ψ(Q) . (12) interior-point methods [36] or some software optimization
Q∈Q
packages such as CVX [37]. Alternatively, we shall
P-I.3 is not concave in ∆22 at a given Q and P-I.4 is convex propose an attractive suboptimal solution for (18) and
in Q over the set Q. provide the corresponding structure of ∆22 .

1) Suboptimal Solution to P-I.5: We solve (18) by consid-
ering an upper bound of its Lagrange dual function and denote
the resulting solution as ∆∗22 . The following theorem defines
the structure of the solution.
Theorem 1: Given that the matrix Q(l) W2 in (18a) is
Hermitian and PSD. Let F∗ ∈ CM2 ×M2 be a PSD matrix
obtained through (21) and it is Hermitian. Define the matrix
∆
E2 = Q(l) W2 −F∗ whose EVD is E2 = UE2 ΛE2 U†E2 where
λ1 (E2 ) ≥ · · · ≥ λM2 (E2 ). A suboptimal solution to (18) that
optimizes an upper bound of its Lagrange dual function is
∆∗22 = U∆∗22 Λ∆∗22 U†∆∗ , where
22
1. The eigenvectors of ∆∗22 are the same as those of UE2 ,
i.e., U∆∗22 = UE2 .
2. The eigenvalues of ∆∗22 , λ1 (∆∗22 ) ≥ · · · ≥ λM2 (∆∗22 ) ≥
0, are:
i. If p2 ∈ (1, ∞)
q2
λi (∆∗22 ) = (|λi (E2 )|/|λ1 (E2 )|) p2 λ1 (∆∗22 ) , i = 2, · · · , M2
(19)
where p2 is the Schatten norm order in the set Ũ2 and q2 =
(1 − 1/p2 )−1 . The maximum eigenvalue λ1 (∆∗22 ) is
"M # −1
p2
X 2
q
λ1 (∆∗22 ) = 2 (|λk (E2 )|/|λ1 (E2 )|) 2 . (20)
k=1
ii. If p2 = 1, λ1 (∆∗22 ) = 2 and λi (∆∗22 ) = 0 for i =
2, · · · , M2 .
iii. If p2 = ∞, ∆∗22 = 2 I.
Proof: Please refer to Appendix C. 
Corollary 1: The matrix ∆∗22 from Theorem 1 is the optimal
solution for (18) if the matrix F∗ = 0.
Proof: Please see the end of Appendix C. 
Corollary 2: The matrix F∗  0 presented in Theorem 1
2
d(Q) − d̄(l) (Q(l) , ∆22 )
−1(l) , (24)
represents the optimal solution of the following optimization
subproblem
∆
min. h(F) = 2 kQ(l) W2 − FkSq2 + Tr(R̂22 F) . (21)
F0
Proof: Please refer to Appendix D.  d̄(l) =−N(l) vec(Ψ(l) (l)
The problem (21) is convex in F. For q2 ∈ [1, ∞), the
function h(F) is differentiable and hence (21) can be solved
efficiently using the non-monotone spectral projected gradient
(NMSPG) method with superlinear convergence [38]. We
have included the first derivative of the function h(F) in
Appendix D that will be useful in obtaining the numerical
results presented in Section VI. For q2 = ∞, h(F) is not
differentiable and we reformulate the problem (21) as linear
semi-definite programming (SDP) in Appendix D, which can
be solved in polynomial time using an interior-point based
method [36].
The proposed suboptimal solution for P-I.5 through Theo-
rem 1 and (21) is more computationally efficient than solving
(18) directly using the interior-point methods. The NMSPG
algorithm for (21) has a linear time complexity and involves
h(F) and its gradient only. The interior-point methods, on the
other hand, have polynomial time complexity. Furthermore,
Theorem 1 yields the optimal solution for P-I.5 when the
constraint (18c) is always satisfied.
5
The computation of F∗ from (21) does not depend on the
solution ∆∗22 . Hence, using F∗ in E2 yields ∆∗22 analytically.
The AI process starts with ∆22 = γ∆22 I where γ∆22 ≈ 0,
and will reach an accumulation or a stationary point [29,
Theorem 4.7]. The solution of W2 is discarded and the
(l)
solution of ∆22 is the intermediate solution ∆22 for use to
update Q.
C. Solution for Q
(l) (l)
We now update Q, using Q(l) , ∆21 and ∆22 . The corre-
sponding optimization problem P-I.4 is solved using a feasible
descent direction method. The method uses scaled gradient
projection to generate the feasible direction for updating the
precoder matrix [30, Sec. 2.3]. The gradient of Ψ with respect
to Q may not be unique and the descent direction is obtained
by a subgradient instead [39]. For simplicity, let us use
(l)
Ψ(l) (Q) to denote −φ(Q, ∆22 ).
(l)
A direction vector at Q is
d(Q) = vec(Q) − vec(Q(l) ) . (22)
The update equation for Q in the next iteration is
vec(Q(l+1) ) = vec(Q(l) ) + αl d(l) (23)
where the parameter αl ∈ (0, 1] is the step-size at the l-th
iteration that will be discussed later. The feasible direction
vector is d(l) = d(Q̆(l) ). Q̆(l) ∈ Q is obtained through scaled
subgradient projection by using a quadratic approximation
of Ψ(l) (Q) around Q(l) , which is equivalent to solving the
following Euclidean projection problem [30, Sec. 2.3]
1
(l)
Q̆(l) = arg min.
Q∈Q 2 N
where d̄(l) is a descent direction that may not be feasible and
is given by
g (Q ) , (25)

(l)
−1
(l)
Ψ(l) (l)
g (Q )=− I + (R̂22 − ∆22 )Q
(l)
(R̂22 − ∆22 ) . (26)
2 2
The matrix N(l) ∈ CM2 ×M2 is PD symmetric that approxi-
mates the Hessian inverse with respect to Q, and the choice
(l)
of it will be discussed at the end of this Section. Ψg (Q(l) )
represents a subgradient of Ψ with respect to Q evaluated
at Q(l) , please see Lemma 5 in Appendix E. The projection
problem (24) can be solved using the interior-point methods
[36] or an optimization package [37].
The precoder Q in P-I must be solved under the interference
constraint (6b) to guarantee the QoS of PU. The constraint
makes the feasible set of Q non-convex. The non-convexity
in the constraint is handled through the Taylor-series expan-
sion up to the first order. The linearized constraint is more
conservative than the original [16] and the two become closer
and eventually equivalent as we approach the solution.
To ensure d(l) is a feasible direction for the problem P-I,
(l)
Q̆ is determined by the Euclidean projection (24) together
6

with the interference constraint, gradient method [38]

1 (l)
Q̆(l) ∈ arg min. kd(Q) − d̄(l) (Q(l) , ∆22 )k2N−1(l) (27a) N(l) =s(l) I ,
2
  
Q∈Q s(l) =min smax , max smin , ρ(l−1) x(l−1)† x(l−1) (31)
(l)
s.t. log |I + R21 Q| ≤ rL (27b)
where s(l) is known as the spectral step-length, 0 ≤ smin ≤
(l) ∆ (l) smax ≤ ∞ are two pre-determined constants and
where R21 = R̂21 − ∆21 .
The solution of the projection
problem (27) exists as the feasible set is closed and it may
x(l−1) = vec(Q(l) ) − vec(Q(l−1) ) , (32a)
not be unique due to the non-convex constraint (27b). We (l−1)
shall obtain one possible solution out of many by applying y = vec(Ψ(l) (l)
g (Q )) − vec(Ψ(l−1)
g (Q(l−1) )) , (32b)
(l−1) (l−1)† (l−1)
a local approximation through linearization to the constraint ρ = 1/(y x ). (32c)
(27b) to obtain a tractable formulation. Expanding the left side
of (27b) around a nominal point Q̃(k) , (27) becomes Algorithm 1 illustrates the major steps for solving P-I,
through the AI and feasible direction projected subgradient
1 (l) (FDPS) techniques. In Algorithm 1, we use a diminishing step-
Q̃(k+1) = arg min. kd(Q) − d̄(l) (Q(l) , ∆22 )k2N−1(l) (28a)
Q∈Q 2 size rule for αl in (30). The algorithm would converge to a
s.t. Tr (M(k) Q) ≤ r(k) (28b) critical point and reach a solution for Q. This is shown in
Appendix E with the aid of the Danskin’s min-max Theorem
where the variables M(k) and r(k) are defined as [36, Proposition 4.5.1].
 −1
(l) (l) Algorithm 1 Solving P-I Using the AI and FDPS Techniques
M(k) = I + R21 Q̃(k) R21 , (29a)
(l) Input: Provide the initial matrices Q(1) , N(1) ; the Schatten
r(k) = rL − log |I + R21 Q̃(k) | + Tr (M(k) Q̃(k) ) .(29b)
norm order pj and the error size j for j = 1, 2; the LR limit
The matrix M(k) represents the partial derivative of the left rL . Define the step-size rule for αl ; set the iteration counter
side of (27b) with respect to Q evaluated at Q̃(k) . Note that l = 0.
k = 0, 1, · · · here represents the iteration index for the first Step 1. l = l + 1.
order approximation and (28) is iterated over k to reach a Step 2. If Q(l) satisfies a certain stopping criterion: STOP.
(l)
solution. Step 3. Find ∆21 by the explicit solution of (8).
(l)
The proposed formulation (28) would guarantee a feasible Step 4. Obtain the solution ∆22 by iterating (17) and (18).
(l) (l) (l)
solution to the subproblem (27) as the feasible set of (28) Step 5. Determine d̄ (Q , ∆22 ) using (25), solve (27) for
(l)
is a subset of that in (27). This means that (28b) provides Q̆ through (28).
a conservative approximation of (27b). Hence, as k increases Step 6. Evaluate αl from a diminishing step-size rule.
the precoder matrix Q̃(k) converges to a solution for (27). Step 7. Obtain the precoder Q(l+1) by solving (30) and N(l+1)
The iteration begins with the initial value Q̃(0) = γCR I where using (31). Go back to Step 1.
γCR ≈ 0 and the converged value is Q̆(l) .
As the feasible region for Q in P-I is not convex, direct A typical choice for initialization is Q(1) = γQ I and
(1)
update using (23) is no longer valid. The solution for Q N = I, where γQ ≈ 0. We would like to clarify that the
requires the optimization optimization problem P-I.3 shown in (11) is non-concave in
∆22 and the feasible set of Q in P-I is non-convex. There
Q(l+1) ∈ arg min. kd(Q) − αl d(l) (Q̆(l) )k2 (30a) could be multiple stationary points and there is no guarantee
Q∈Q
the proposed algorithm will give the global optimum solution.
(l)
s.t. log |I + R21 Q| ≤ rL . (30b) For better result, one can use several different initializations
for Q(1) and select the solution that gives the best objective
There could be multiple solutions to (30) due to the non- value.
convex optimization problem, and applying the approach as in It is straightforward to extend the proposed CR solution
(28)-(29) with linearization of (30b) would lead to a solution. for the presence of multiple PUs. In such a case, a separate
To complete the solution finding, the details for selecting solution for the amount of interference CSI uncertainty for
the step-size αl and the PD matrix N(l) are as follows. each PU is evaluated according to Section III-A. Furthermore,
Step-Size the number of constraints is expanded for the PUs in (30b)
We shall use an exogenous step-size that can be fixed or di- when obtaining the precoder.
minishing [30], [39]. Such a step-size ensures the convergence
of (23) to a stationary point, as shown in Appendix E. IV. C OMPOUND C APACITY P ROBLEM
Hessian Inverse
When dropping (6b) in P-I, we have the CC problem
The Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm
can yield a good estimate of the Hessian inverse as the iteration (P-II): min. max. − φ(Q, ∆22 ) . (33)
Q∈Q ∆22 ∈U2
proceeds [28, Ch. 4]. To simplify the algorithm and ensure
stable behavior, we simply choose N to be a scalar multiple of The CC problem is considered to be challenging to solve in
an identity matrix by following the approach from the spectral the literature due to the inner maximization over ∆22 before
 7

optimization in Q can take place. Indeed, analytical solutions Let us define the PSD matrix R222 = UR̂22 (ΛR̂22 −
for P-II are known under very few cases: when H22 belongs to 2 I)+ U†R̂ . The inequality k∆22 ks∞ ≤ 2 from the uncer-
22
an isotropic set [2], the available channel Ĥ22 is rank one [3], tainty set Ũ2 is equivalent to −∆22  −2 I where ∆22 is
and the uncertainty set is defined for the spectral norm only PSD. Adding both sides by R̂22 gives
(p2 = ∞) [5], [6]. In this work, the considered uncertainty set
U2 is governed by the Schatten norm and the feasible set Q R̂22 − ∆22  R222 (34a)
is more general than the previous studies [2]–[6]. 2
⇒ log |I + (R̂22 − ∆22 )Q| ≥ log |I + R22 Q| . (34b)
The proposed algorithm for CR is applicable to the CC
problem as well. There are two main changes. The first one As a result, the solution of ∆22 that yields the minimum MI
is that we do not need to update ∆21 in each iteration. The is ∆22 = 2 I.
second is that the feasible set of Q is convex because of the Similarly, let R211 = R̂21 + 1 I. The inequality k∆21 ks∞ ≤
absence of the constraint (6b), making it easier to solve. 1 from the uncertainty set U1 is equivalent to −∆21  1 I.
Algorithm 2 summarizes the main framework to find a Adding both sides by R̂21 yields
solution for the CC problem. It would converge to a critical R̂21 − ∆21  R211 (35a)
point for ∆22 and Q. The proof of convergence follows a ⇒ log |I + (R̂21 − ∆21 )Q| ≤ log |I + 1
R21 Q| . (35b)
similar analysis as in Appendix E.
The largest amount of interference to PU occurs when ∆21 =
Algorithm 2 Solving P-II Using the AI and FDPS Approach −1 I.
Input: Provide the initial matrices Q(1) and N(1) ; the Schatten Using the uncertainty matrix solutions ∆22 = 2 I and
norm order p2 and the uncertainty size 2 . Set counter l = 0. ∆21 = −1 I, P-I in (6) becomes
Step 1. l = l + 1. (P-III): max. log |I + R222 Q| (36a)
Step 2. If Q(l) satisfies a suitable stopping criterion: STOP. Q0
(l)
Step 3. Obtain the solution ∆22 by iterating (17) and (18). s.t. log |I + R211 Q| ≤ rL (36b)
(l)
Step 4. Calculate Ψg (Q(l) ) using (26) and d̄(l) using (25); where the power limitation in the feasible set Q is assumed
solve the subproblem (24) for Q̆(l) and obtain d(l) from (22) satisfied. The following proposition shows that the optimal
at Q = Q̆(l) . precoder should diagonalize the matrix R222 .
Step 5. Determine αl using a fixed or diminishing step-size
rule. Proposition 2: Let U ∈ CM2 ×r with r ≤ M2 be (semi-
Step 6. Obtain the precoder Q(l+1) from (23) and N(l+1) )unitary and Σ ∈ Rr×r a diagonal matrix with non-negative
using (31). Go back to Step 1. non-increasing elements. The optimal precoder solution for the
problem P-III would diagonalize the covariance matrix R222
and has the structure
Similar to Algorithm 1, the initialization is Q(1) = γQ I
‡ ‡
and N(1) = I with γQ ≈ 0. The optimization problem P- Q = (R222 )− 2 UΣU† (R222 )− 2 (37)
I.3 in (11) is non-concave in ∆22 . The possibility remains 1
where (R222 ) 2is the matrix square root of R222
and (·)‡
of having multiple stationary points that some could be local
denotes the pseudo-inverse of (·).
optima. Using several different initializations for Q(1) could ‡ ‡

improve the chance of obtaining the global optimum solution. Define the PSD matrix R̄ = (R222 )− 2 R211 (R222 )− 2 whose
†
EVD is R̄ = UR̄ ΛR̄ UR̄ where UR̄ ∈ CM2 ×r and ΛR̄ ∈
Rr×r are (semi-)unitary and diagonal matrices. Then, the
V. PARTICULAR S OLUTIONS selection U = UR̄ Ia would optimize (36) where Ia is the anti-
Under some specific conditions the proposed solutions for identity matrix of appropriate size. The matrix Σ is obtained
the CR and CC problems can be reduced to simpler forms for by solving the problem
evaluation.
(P-IV): max. log |I + Σ| (38a)
Σ0

A. Particular Solution for P-I s.t. log |I + Ia ΛR̄ Ia Σ| ≤ rL . (38b)

We shall consider the solution of the CR problem P-I under
the following two conditions.
• The uncertainty sets U1 and U2 for R̂21 and R̂22 are
defined under pj = ∞, j = 1, 2 in the Schatten norm
(3), which is equivalent to the spectral norm. Using the
spectral norm will lead to the most conservative solution
among the class of unitarily-invariant norms [4];
• The power limitation constraints for the precoder in
the feasible set Q are inactive in the presence of the
interference constraint (6b). Such situation occurs often
in a CR network where the performance is interference
limited rather than power-budget restricted.
Proof: Please refer to Appendix F. 
P-IV is not convex in Σ due to the constraint (38b). (38b)
can become tractable by using the Taylor-series expansion and
P-IV can be solved by the iterative linearization technique as
presented in Section III-C.
B. Particular Solution for P-II
When the power of the precoder Q is restricted to be in a
unitarily-invariant convex set Qu , for example, the total power
restriction
Qu = {Q : Q  0, Tr(Q) ≤ PT } (39)
8
or the maximum power limitation
Qu = {Q : Q  0, λ1 (Q) ≤ Pmax } , (40)
the following theorem states the precoder for the CC problem
has the same eigenvectors as the available channel covariance.
Theorem 2: Consider the EVD representation Q =
UQ ΛQ U†Q and R̂22 = UR̂22 ΛR̂22 U†R̂ where the elements
22
along the diagonal of ΛQ and ΛR̂22 are arranged in non-
increasing order. For any ∆22 ∈ U2 , define the matrix
∆
∆o22 = JU†R̂ ∆22 UR̂22 J, where J ∈ Rn×n is a diagonal
22
matrix with diagonal elements either +1 or −1. If Q ∈ Qu
where Qu is a unitarily-invariant compact and convex set, the
optimal precoder for the CC problem P-II has eigenvectors
equal to those of the channel R̂22 available, i.e., UQ = UR̂22 .
Additionally, P-II becomes
(P-V): max. min. log |I + (ΛR̂22 − ∆o22 )ΛQ | (41)
ΛQ ∈Qu ∆o o
22 ∈U2
where the uncertainty set U2o is
U2o = {k∆o22 kSp2 ≤ 2 , ΛR̂22 − ∆o22  0, ∆o22  0}. (42)
Proof: Please see Appendix G.  the Schatten norm orders pj and normalized uncertainty error
The conclusion that the precoder has the same eigenmodes
as the available channel has previously been demonstrated for
the CSI uncertainty set defined by the spectral norm only under
unitarily-invariant convex set for the transmission power [5],
[6]. We generalize the results here and prove that the same
conclusion carries for the CSI uncertainty set defined by the
Schatten norm.
As apparent from (41), the outer maximization for the CC
problem simplifies considerably since the unknowns are M2
eigenvalues rather than M2 (M2 + 1)/2 independent elements
of Q.
The modification to the FDPS method developed in Section
III-C for P-V is straightforward. In particular, the matrices Q
and R̂22 become diagonal counterparts ΛQ and ΛR̂22 with
diagonal elements arranged in non-increasing order. Further-
more, ∆22 is now ∆o22 , and Q and U2 turn into Qu and U2o .
We can simplify P-V further by restricting the matrix ∆o22
to be diagonal. Consequently, both Q and ∆22 share the same
eigenvectors as R̂22 , which leads to a channel-diagonalizing
solution. In this case, we only need to obtain the eigenvalues
of Q and ∆22 . The restriction that ∆o22 has diagonal structure
will likely lead to a suboptimal solution, unless the channel
uncertainty set U2 is defined by the spectral norm [5], [6].
VI. N UMERICAL R ESULTS
We shall provide numerical examples to illustrate the per-
formance of proposed precoder solutions. The available CSI
is set as R̂2j = Ĥ†2j Ĥ2j for j = 1, 2, where the elements
of Ĥ2j are IID and follow a Gaussian distribution with zero
mean and unit variance. The SU rate reported is the worst
case rate by the design methodology and is the average over
the randomly generated matrices R̂2j . The uncertainty error
size is j = wj kR̂2j kS∞ where the normalized parameter
wj ∈ (0, 1) is used to adjust the amount of uncertainty. The
noise variance is fixed to unity and hence the total SU transmit
power PT is equal to the signal-to-noise ratio (SNR), i.e.,
PT = SNR. In addition to the power sets in (39) and (40), we
consider the per-transmit dimension power set defined as
Q = Q : Q  0, Q  diag([P1 , · · · , PM2 ]T )

(43)
where Pm , m = 1, · · · , M2 , is the power limit for the m-
th antenna. Unless otherwise stated, we simply set the power
values as Pmax = 0.7PT and Pm = 0.75m PT for illustration
purpose.
For all the proposed precoder designs, we compute αl in
(23) by a diminishing step-size rule such that the rule αl+1 =
1/(l+1) is used for the CR problem while αl+1 = αl (1−βαl )
with β = 10−2 ∈ (0, 1) and α0 = 1 for the CC problem.
The proposed solutions shown in Algorithm 1 and Algorithm
2 are computed with one initialization of Q(1) only. The
initialization constants γ∆22 , γQ and γCR are set to 10−4 and
the constants smin and smax are 10−7 and 1020 . We shall denote
the solution for ∆22 as AI(O) when solving P-I.5 optimally
using CVX [37] and AI(SO) as the suboptimal solution.
We first examine Algorithm 1 and the proposed solutions
for the CR problem in Section III under different values of
wj , j = 1, 2. Fig. 2 illustrates the SU rate in (nats/s/Hz) as
the leakage rate limit rL (nats/s/Hz) increases. It includes the
robust solution from [16] for comparison. The antenna setting
is M2 = N2 = N1 = 4 and SNR = 5dB. We use the AI(O)
method to obtain the uncertainty element ∆22 . It is obvious
from Fig. 2 that the proposed solutions for Q and ∆21 provide
higher rate than that in [16]. It should be noted that the work
[16] does not consider CSI uncertainty in the direct link of
SU; therefore, we limit the comparison in Fig. 2 for w2 = 0
only. As expected, increasing the uncertainty error w2 for fixed
values of w1 will deteriorate the SU rate. Also, the proposed
solution when p1 = p2 = ∞ produces the most conservative
solution among the Schatten norm orders.
Fig. 3 depicts the SU rate versus the uncertainty error w1 of
the interference link for antenna settings of M2 = N2 = N1 =
2 or 4, where the leakage rate limit and the uncertainty error
of the direct link are kept at rL = 3 (nats/s/Hz) and w2 = 0.5.
The AI(SO) method is applied to obtain ∆22 . The results show
that the SU rate decreases as the uncertainty error w1 increases
for a fixed value of w2 , and the decrease is more significant for
the four antenna than the two antenna setting. The uncertainty
sets defined with p1 = p2 = 1 have the smallest size among all
other Schatten norm orders and SU has the highest rate. The
effect of the Schatten norm order on the SU rate is apparent.
We next consider Algorithm 2 to obtain the designs for
the CC problem presented in Section IV when the normalized
uncertainty error w2 = 0.3 or 0.5 and the Schatten norm order
p2 = 2. Both precoder solutions AI(O)-FDPS or AI(SO)-FDPS
are provided. To illustrate the effectiveness of the proposed
FDPS technique, we compare the proposed precoder solutions
with the beamforming precoder that transmits signals over
the principal eigenvalue of the worst channel and with the
equal power allocation precoder that distributes power equally
among the antennas. Both the beamforming and equal power
precoders require ∆22 and it is taken from the proposed
 9

5 12
CC, AI(O)-FDPS, w2 = 0.3
CC, AI(SO)-FDPS, w2 = 0.3
4.5
10 CC, Equal power, w2 = 0.3
4 CC, Beamforming, w2 = 0.3
CC, AI(O)-FDPS, w2 = 0.5
3.5 CC, AI(SO)-FDPS, w2 = 0.5

SU Rate (nats/s/Hz)
8
SU Rate (nats/s/Hz)

CC, Equal power, w2 = 0.5

3
CC, Beamforming, w2 = 0.5

2.5 6

2
CR, Proposed solution, p1 = p2 = 1, w1 = 0.4, w2 = 0 4
1.5 CR, Proposed solution, p1 = p2 = 2, w1 = 0.4, w2 = 0
CR, Solution from [16], w1 = 0.4, w2 = 0
1
CR, Proposed solution, p1 = p2 = 1, w1 = 0.4, w2 = 0.25 2

0.5 CR, Proposed solution, p1 = p2 = 2, w1 = 0.4, w2 = 0.25

CR, Proposed solution, p1 = p2 = ∞, w1 = 0.4, w2 = 0.25
0 0
0 1 2 3 4 5 6 -10 -5 0 5 10 15 20
rL (nats/s/Hz) SNR (dB)

Fig. 4. Rate for SU from the proposed CC solution, the beamforming and the
Fig. 2. Rate for SU from the proposed CR solution and from [16] for different equal power solutions versus SNR (dB), M2 = N2 = 4. The results shown
values of rL (nats/s/Hz), SNR = 5dB, M2 = N2 = N1 = 4. The results are the average over 100 realizations of R̂22 .
shown are the average over 100 realizations of R̂22 and R̂21 .

4.5
CR, Proposed solution, p1 = p2 = 1 CC, AI(O)-FDPS, p2 = 2
3
CR, Proposed solution, p1 = p2 = 2 CC, AI(SO)-FDPS, p2 = 2
4
2.8 CR, Proposed solution, p1 = p2 = ∞ CC, Equal power, p2 = 2

M2 = N2 = 4, N1 = 4 3.5 CC, Beamforming, p2 = 2

2.6 M2 = N2 = 2, N1 = 2 CC, AI(O)-FDPS, p2 = ∞
SU Rate (nats/s/Hz)

CC, AI(SO)-FDPS, p2 = ∞

SU Rate (nats/s/Hz)
3
2.4 CC, Equal power, p2 = ∞
2.5 CC, Beamforming, p2 = ∞
2.2
2
2
1.5
1.8
1
1.6
0.5
1.4
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Uncertainty error w1 Uncertainty w2

Fig. 3. Rate for SU from the proposed CR solution versus the uncertainty Fig. 5. Rate for SU from the proposed CC solution, the beamforming and
error w1 , rL = 3 (nats/s/Hz), w2 = 0.5, SNR = 5dB, M2 = N2 = N1 = 2 the equal power solutions at different values of w2 , M2 = N2 = 4. The
or 4. The results shown are the average over 100 realizations of R̂22 and R̂21 . results shown are the average over 100 realizations of R̂22 .

AI(O)-FDPS solution. Fig. 4 illustrates the average SU rate Q is relaxed. The SU rate increases with rL as expected. The
as SNR increases. The antenna setting for SU is M2 = N2 = increase in the SU rate is more sensitive at smaller normalized
4. Fig. 4 indicates that the proposed precoder significantly uncertainty radii w1 and w2 .
outperforms the beamforming and equal power ones. The Fig. 7 shows the SU rates for the CC problem as SNR
performance of the AI(O)-FDPS and AI(SO)-FDPS is almost varies. The curves denoted by “Q diagonalization” and “Q,
identical at different SNR and w2 values. ∆22 diagonalization” correspond to the optimal precoder
Fig. 5 shows the average SU rates when the error w2 that diagonalizes the available channel estimate R̂22 and the
increases with SNR = 4dB and p2 = 2 or ∞. The power and suboptimal solution in which both Q and ∆22 diagonalize R̂22
antenna settings of Fig. 5 are the same as those in Fig. 4. The for any value of p2 . The SU rates for both precoder solutions
proposed precoders show better SU rate than the beamforming are very comparable. The proposed precoders are applicable to
and equal power designs. The uncertainty set defined with a wide class of situations and covers some precoder solutions
p2 = ∞ has the largest shape among the Schatten norm from the literature that are obtained under more restrictive
uncertainty sets; hence the proposed solution at p2 = ∞ would scenarios. This is illustrated in the figure where the proposed
produce the most conservative SU rate. Similar to Fig. 4, Q diagonalization precoder yields identical performance with
the precoders from AI(O)-FDPS and AI(SO)-FDPS are quite the precoder from [5], [6] that is applicable to p2 = ∞ only.
comparable in performance at different values of p2 and w2 .
Finally, we investigate the proposed particular solutions for
VII. C ONCLUSION
the CR and CC problems in Section V. Fig. 6 evaluates the
performance of the particular solution for the CR problem We have developed in this paper the solution of a robust
from Section V-A at different values of w1 = w2 as the leakage precoder for an SU in an underlay cognitive radio framework,
rate limit rL increases, where p1 = p2 = ∞ and the power set where the available CSI of the direct link of SU and the
10

14
CR, Particular solution, w1 = w2 = 0 where the expectation is taken over the channel H22 . Using
12
CR, Particular solution, w1 = w2 = 0.15
first the fact that [33, Corollory 2.8.5]
CR, Particular solution, w1 = w2 = 0.35
CR, Particular solution, w1 = w2 = 0.55
log |I + AB| = log |I + BA| (45)
10
SU Rate (nats/s/Hz)

for any A ∈ Cn×m and B ∈ Cm×n , and next the concavity

8
of the log | · | function over the set of PSD matrices, we have
6
E[log |I + H22 QH†22 |] = E[log |I + H†22 H22 Q|] (46a)
4 ≤ log |I + E[H†22 H22 ]Q| . (46b)

2
Hence the metric φ(·, ·) in (4a) is an upper bound of (44).
We can follow similar procedure to show that the function
0
0 1 2 3 4 5 6 7
g(Q, R21 ) (5a) is no less than E[log |I + H21 QH†21 |], where
rL (nats/s/Hz)
the expectation is applied to the channel H21 .
Fig. 6. Rate for SU from the particular CR solution versus rL , M2 = N2 =
N1 = 5. The results shown are the average over 100 realizations of R̂22 and A PPENDIX B
R̂21 . P ROOF OF P ROPOSITION 1
CC, Q Diagonalization proposed, p2 = 2, w2 = 0.2
Consider NSD ∆21 ∈ U1 and PSD ∆ˇ 21 ∈ U1 . It is obvious
14 CC, Q, ∆ 22 Diagonalization proposed, p2 = 2, w2 = 0.2
CC, Q Diagonalization proposed, p2 = ∞, w2 = 0.2 ˇ
that −∆21  −∆21 . Hence
CC, Q, ∆ 22 Solurion from [5, 6], p2 = ∞, w2 = 0.2
12
CC, Q Diagonalization proposed, p2 = 2, w2 = 0.4 ˇ 21  R̂21 − ∆21
R̂21 − ∆ (47a)
CC, Q, ∆ 22 Diagonalization proposed, p2 = 2, w2 = 0.4
10
CC, Q Diagonalization proposed, p2 = ∞, w2 = 0.4 1/2
Q (R̂21 − ∆ ˇ 21 )Q1/2  Q1/2 (R̂21 − ∆21 )Q1/2 (47b)
SU Rate (nats/s/Hz)

CC, Q, ∆ 22 Solution from [5, 6], p2 = ∞, w2 = 0.4

8 ˇ 21 )Q1/2 | ≤
log |I + Q1/2 (R̂21 − ∆
6 log |I + Q1/2 (R̂21 − ∆21 )Q1/2 | (47c)
ˇ 21 )Q| ≤ log |I + (R̂21 − ∆21 )Q| (47d)
log |I + (R̂21 − ∆
4

2
0
-10 -5
Fig. 7. Rate for SU from the particular CC solutions versus SNR (dB),
M2 = 6, N2 = 3. The results shown are the average over 100 realizations
of R̂22 .
interference link to PU have uncertainties and the QoS of
PU must be maintained. In the absence of PU, the proposed
solution addresses the CC problem in the literature. The CSI
uncertainties are modelled by convex sets with the Schatten
norm measure. We formulate the design as a maximin opti-
mization problem that seeks a precoder to maximize the MI
of SU under the most conservative anticipated performance
for SU while ensuring the QoS of PU under all possible CSI
uncertainties expected in the interference link. The maximin
problem is solved iteratively, using the explicit solution for
the worst case CSI deviation of the interference link, the AI
technique for the worst case CSI error of the direct link,
and the FDPS method for the precoder. Convergence analysis
of the proposed solution is established. Simpler solutions for
some specific choices of the norm measure and transmission
power restrictions are also derived. Simulations validate the
performance improvement of the proposed precoders for the
CR and CC problems.
A PPENDIX A
The MI for the transmission link from SU-Tx to SU-Rx is
defined as [1]
E[log |I + H22 QH†22 |] (44)
where Q1/2 is the matrix square root of Q. (47b) follows
since for Hermitian matrices A, B ∈ Cn×n such that A  B,
0 5 10 15 20
SAS†  SBS† where S ∈ Cm×n [33, Proposition 8.1.2].
SNR (dB)
(47c) comes from the fact that the function log | · | is strictly
increasing on the set of PD Hermitian matrices [40, Ch. 16,
F.2.b], and (47d) is by applying (45). It is obvious from (47d)
that the function g(·, ·) will take a larger value if ∆21 is NSD
rather than PSD.
A PPENDIX C
P ROOF OF T HEOREM 1
The proof requires a number of lemmas. We first summarize
them below,
Lemma 2: Given the Hermitian and P(S)D matrices A, B ∈
Cn×n . The matrix A(I+BA)−1 is then Hermitian and P(S)D
matrix.
Proof: Let A1/2 be the matrix square root of A. From direct
algebraic manipulation,
A(I + BA)−1 = A1/2 A1/2 (I + BA1/2 A1/2 )−1 (48a)
= A1/2 (I + A1/2 BA1/2 )−1 A1/2 .(48b)
(48b) is due to the fact (I + AB)−1 A = A(I + BA)−1 [33,
Fact 2.16.16]. The Hermitian property can be easily seen from
the right side of (48b). Note that A1/2 BA1/2 is P(S)D. By
the definition of the P(S)D property, the right side is P(S)D
and hence the left side. 
Lemma 3: Let A, B ∈ Cn×n be Hermitian with eigenvalues
λ1 (A) ≥ · · P
· ≥ λn (A) and λ1 (B) ≥ · · · ≥ λn (B). Then,
n
Tr(AB) ≤ i=1 λi (A)λi (B) and the equality holds if A
and B are diagonal [33, Fact 5.12.4].
 11

Lemma 4: (Hölder inequality [33, Proposition 9.1.6]) Let From the Schatten norm definition in (3) and (19), when
p, q ∈ [1, ∞] such that 1/p + 1/q = 1, and let a, b ∈ Cn with the constraint (18b) is active for reaching a solution of P-II.3,
elements a1 , · · · , an and b1 , · · · , bn . Then, we obtain λ1 (∆∗22 ) as
M2  q
|aH b| ≤ kakp kbkq , (49) p2 ∗ p2
X |λk (E2 )| 2
2 =λ1 (∆22 ) + λ1 (∆∗22 )p2 . (54)
|λ1 (E2 )|
where k · kp is the Hölder norm [33, Proposition 9.1.4] k=2
P (54) leads to (20) directly.
( n |ai |p )1/p , 1 ≤ p < ∞ For p2 = 1, from (51a) a possible choice for λi (∆∗22 ) that
i=1
kakp = (50)
 max |ai |, p = ∞. achieves the equality in (52f) and fulfills (18b) is λ1 (∆∗22 ) =
i∈1,··· ,n
2 and λi (∆∗22 ) = 0 for i 6= 1. For p2 = ∞, using (51c)
Equality in (49) holds if and only if we can select λi (∆∗22 ) = 2 to reach the equality in (52f)
and satisfy (18b). The above two choices for λi (∆∗22 ) are
valid regardless of the value of the matrix F∗ in E2 . If the

|ai ||bi | = kbk∞ |ai |, p = 1, (51a)
eigenvalues of E2 as well as ∆∗22 are considered variables to


|a1 |1/q |an |1/q

= · · · = , 1 < p < ∞, (51b) be evaluated, the conditions (51a) or (51c) can lead to other
|b1 |1/p |bn |1/p
choices for λi (∆∗22 ).



|ai ||bi | = kak∞ |bi |, p = ∞. (51c)

Proof of Corollary 1: If F∗ is the zero matrix, then the
matrix E2 would be PSD and the inequality in (52e) becomes
equality, which makes ∆∗22 the optimal solution for the
From Lemma 2 and (17) the matrix (Q(l) W2 ) is Hermitian
problem P-II.3 when p2 ∈ (1, ∞). For p2 = 1 or p2 = ∞
and PSD.
the choice (λi (∆∗22 ) = 2 for i = 1 and zero otherwise) or
We shall begin the proof of Theorem 1 as follows. Let the
(λi (∆∗22 ) = 2 for every i) is optimal as E2 becomes a fixed
Lagrange multipliers associated with (18b) and (18c) be γ ≥ 0
matrix and the equality condition (51a) or (51c) is determined
and F  0. We first form the Lagrangian L for (18) and its
by ∆∗22 only.
corresponding dual function D(·, ·) is

D(F, γ)= max. L(∆22 , F, γ) (52a) A PPENDIX D

∆22 0
E VALUATION OF THE M ATRIX F∗
= max. Tr(Q(l) W2 ∆22 ) − γ(k∆22 kSp2 − 2 )
∆22 0 Proof of Corollary 2: From Appendix C, an upper bound
the dual function D(·, ·) of the problem (18) is the right
 
+ Tr (R̂22 − ∆22 )F (52b) for
  side of (52f). D(·, ·) would take the following values
(l)
= max. Tr (Q W2 − F)∆22 − γk∆22 kSp2 + γ2 (
∆22 0 ∞, γ < kΛE2 kSq2 ,
D(F, γ) ≤ (55)
+ Tr (R̂22 F) (52c) γ2 + Tr (R̂22 F), γ ≥ kΛE2 kSq2 .
≤ max. Tr (ΛE2 Λ∆22 ) − γkΛ∆22 kSp2 + γ2 We can find the variables γ and F by minimizing the upper
Λ∆22 0
bound of D(·, ·) when γ ≥ kΛE kSq2 . That is,
+ Tr (R̂22 F) (52d)
≤ max. kΛE2 Λ∆22 kS1 − γkΛ∆22 kSp2 + γ2 min. γ2 + Tr (R̂22 F) (56a)
Λ∆22 0 γ≥0, F0

+ Tr (R̂22 F) (52e) s.t. γ ≥ kΛE2 kSq2 . (56b)

≤ max. kΛ∆22 kSp2 kΛE2 kSq2 − γ + γ2 It is evident that γ should be chosen as its smallest possible
Λ∆22 0
value, γ = kΛE2 kSq2 . Substituting the optimal value of γ back
+ Tr (R̂22 F) . (52f) into (56a) produces the problem (21).
Derivative of h(F): For q2 ∈ [1, ∞), the derivative of the
where E2 = (Q(l) W2 − F). The inequality in (52d) is due
function h(F) in (21) with respect to F is
to Lemma 3 and equality happens when U∆22 = UE2 .
Hence, we conclude that U∆∗22 = UE2 in order to maximize  q q2 1/q2
†
L(∆22 , F, γ). d Tr E2 E2
The inequality Tr (ΛE2 Λ∆22 ) ≤ k(ΛE2 Λ∆22 )kS1 in (52e) Dh (F) = 22 + R̂22 (57a)
dF
is from [33, Fact 9.14.3]. The inequality in (52f) is due to q q2 −2
−q /p
Hölder inequality in Lemma 4. Equality for (52f) can be = − 2 kE2 kSq 2 2 E†2 E2 E2 + R̂22 . (57b)
2
achieved provided that the eigenvalues of the suboptimal so-
lution ∆∗22 and that of E2 satisfy (51). Thus, for p2 ∈ (1, ∞), In (57b), the matrix (E† E ) should be non-singular for q < 2,
2 2 2

|λ1 (E2 )|1/p2 |λi (E2 )|1/p2 otherwise we can consider its pseudo-inverse to obtain a well-
1/q2
= ··· = 1/q2
(53) defined derivative Dh (F) [33, Sec. 6.1].
λ1 (∆∗22 ) λi (∆∗22 ) We shall provide next a reformulation for (21) q when
and (19) in Theorem 1 follows. q2 = ∞. Since kE2 kS∞ is the largest eigenvalue of E†2 E2
12

according to (3), the optimization problem (21) can be casted and Q∗ . Using (23) and (25),
as
kvec(Q(l+1) ) − vec(Q∗ )k2N−1(l)
≤ kvec(Q(l) ) + αl d(l) − vec(Q∗ )k2N−1(l) (61a)
min. 2 t + Tr (R̂22 F) (58a)
(l) (l) ∗
t≥0, F0
≤ kvec(Q ) + αl d̄ − vec(Q )k2N−1(l) (61b)
tI E†2
 
(l) ∗
s.t 0 (58b) = kvec(Q ) − vec(Q )k2N−1(l)
E2 tI (l) † ∗
−2αl vec(Ψ(l) (l)
g (Q )) (vec(Q ) − vec(Q ))

where we have introduced t such that E†2 E2 ≤ t2 I and used
the Schur complement property of a partitioned PSD matrix
[33, Definition 6.1.8]. Thus, (21) reduces to (58) when q2 =
∞, which is an SDP problem of an affine objective function
with a linear matrix inequality constraint.
A PPENDIX E
We shall establish the convergence of Algorithm 1 to a
stationary point with exogenous step-size. For a given Q, ∆21
has an optimal solution and the AI method is ensured to yield
a solution for ∆22 [29]. We therefore only need to show the
convergence of the FDPS iteration for Q. It is achieved based
on the Danskin’s min-max Theorem [36, Proposition 4.5.1]
which is repeated here for completeness.
Lemma 5: (Danskin’s min-max Theorem [36, Proposition
4.5.1]) Let Y be a compact set and let the function f : Cn×m ×
Y → R be continuous and differentiable such that f (·, Y) :
Cn×m → R is convex for each Y ∈ Y. Then,
∆
1. The maximum function p(X) = max. f (X, Y) is convex
Y∈Y
and has a subgradient OX f (X, Ȳ) with respect to X, where
∆ |I + AB| ≥ |I + Ia ΛA Ia ΛB | .
Ȳ ∈ Y(X) and Y(X) = {Ȳ : f (X, Ȳ) = max. f (X, Y)} is
Y∈Y
the set of maximizing points for p(·).
2. If f (·, Y) is differentiable for any Y ∈ Y and ∇X f (X, ·)
is continuous on Y for each X, then
∂p(X) = conv{∇X f (X, Ȳ) : Ȳ ∈ Y(X)} (59)
where ∂p(X) is the subdifferential of the convex function
p(X) at X, and conv{·} denotes the convex hull.
To make use of the min-max theorem, we associate the
variables to our problem as follows: f (·, ·) is −φ(·, ·), p(·) is
(l)
Ψ(·), X is Q, Y is ∆22 , and Ȳ is ∆22 .
The following assumption is needed. Assumption 1: The
Hessian inverse approximation matrix N(l) , l = 0, 1, 2 · · · ,
satisfies the following condition
ckzk2 ≤ kzk2N(l) ≤ c̄kzk2 (60)
where 0 < c ≤ c̄. Assumption 1 indicates that the eigenvalues
of N(l) are bounded from below and above, which can be
easily checked to ensure it is valid.
Let Q∗ be an optimal point for P-I in (6). We begin with
+αl2 kvec(Ψ(l) (l) 2
g (Q ))kN(l) (61c)
≤ kvec(Q(l) ) − vec(Q∗ )k2N−1(l) − 2αl (Ψ(Q(l) ) − Ψ(Q∗ ))
+αl2 c̄kvec(Ψ(l) (l) 2
g (Q ))k . (61d)
The inequalities in (61a) and (61b) are due to the fact that
Q(l+1) from (30) is a projection of that from (23) and the
descent direction d(l) from (27) is a projection of d̄(l) . Note
(l)
that vec(Ψg (Q(l) )) is a subgradient for Ψ according to part
2 of Lemma 5, and kd̄(l) k2N−1(l) = kvec(Ψg (Q(l) ))k2N(l) . The
last two terms in (61d) follows from the subgradient definition
[36, Sec. 4.2] and Assumption 1, respectively. The quantity
(l) (l)
kvec(Ψg (Q(l) ))k2 is bounded from above for any ∆22 as
the feasible set Q is compact.
From (61d), we can proceed to prove the convergence of
the proposed FDPS method for exogenous step-size using the
procedure presented in [39].
A PPENDIX F
P ROOF OF P ROPOSITION 2
Lemma 6: For Hermitian PSD matrices A ∈ Cn×n
and B ∈ Cn×n whose EVDs are A = UA ΛA U†A and
B = UB ΛB U†B , where the eigenvalues are arranged in non-
increasing order. Then, we have
(62)
The equality is achieved when UA = UB Ia [2, Lemma 1].
We can prove the precoder structure (37) is optimal using
a similar approach to that in [41, Theorem 1].
We can show that U = UR̄ Ia optimizes (36) as fol-
lows. Substituting (37), the left side of the constraint (36b)
becomes log |I + R̄UΣU† |. According to (62), we have
log |I+ R̄UΣU† | ≥ log |I+Ia ΛR̄ Ia Σ|. Clearly, when the left
side of (36b) is replaced by the expression log |I+Ia ΛR̄ Ia Σ|,
the problem (36) would have a larger feasible set and hence
U = UR̄ Ia maximizes P-V.
A PPENDIX G
P ROOF OF T HEOREM 2
Lemma 7: Let J be the set of all L = 2n diagonal matrices
J that have either +1 or −1 as the diagonal elements. Let
A ∈ Cn×n be a Hermitian matrix and DA bePa diagonal
L
matrix such that [DA ]ii = [A]ii . Then, DA = L1 J∈J JAJ
[40, Ch 9].
∆ ∆
Define the matrices Q̄ = U† QU and ∆¯ 22 =
R̂22 R̂22
U†R̂ ∆22 UR̂22 .Under the unitarily-invariant set Qu , the CC
22
problem P-II can be expressed as
  
min. Ψ(Q̄) = max. − log I + (ΛR̂22 − ∆
 ¯ 22 )Q̄ (63)
the weighted Euclidean distance between Q(l+1) from (30) Q̄∈Qu ¯ 22 ∈U¯2
∆

∆ ¯ 22 kS ≤ 2 , Λ
where the uncertainty set Ū2 = {k∆ p2
¯ ¯
∆22  0, ∆22  0}. It is equivalent to U2 since the Schatten
norm is invariant under the transformation and so does the
PSD property [33, Proposition 8.1.2]. Let ∆o22 = J∆
From (63) and using (45), we have
13
R̂22 − [15] I. Wajid, M. Pesavento, Y. C. Eldar, and D. Ciochina, “Robust downlink
beamforming with partial channel state information for conventional and
cognitive radio networks,” IEEE Trans. Signal Process., vol. 61, no. 14,
pp. 3656-3670, Jul. 2013.
¯ 22 J. [16] M. H. Al-Ali and K. C. Ho, “Transmit precoding in underlay MIMO
cognitive radio with unavailable or imperfect knowledge of primary
interference channel,” IEEE Trans. Wireless Commun., vol. 15, no. 8,
  pp. 5143-5155, Aug. 2016.
¯ = max. − log I + (JΛ J − J∆
Ψ(JQJ) ¯ 22 J)Q̄ (64a) [17] M. H. Al-Ali and K. C. Ho, “Robust transmit precoding for underlay
¯ 22 ∈Ū2 R̂ 22
∆ MIMO cognitive radio with interference leakage rate limit,” in Proc. IEEE
 
o ICASSP, Shanghai, China, Mar. 2016, pp. 3001-3005.
= max. − log + (Λ − ∆ )Q̄ (64b)
 
o o
I R̂ 22 22 
∆22 ∈U2
= Ψ(Q̄) (64c)
where (64b) follows as Ū2 is equivalent to U2o [33, Proposition
8.1.2]. The function Ψ(Q̄) is convex in Q̄, and hence
" L
# L
1 X ¯ 1 X ¯
Ψ(Q̄) = Ψ(JQJ) ≥ Ψ( JQJ) (65a)
L L
J∈J J∈J
= Ψ(DQ̄ ) (65b)
where (65a) is due to the Jensen’s inequality and (65b) are
obtained from Lemma 7. The inequality in (65a) becomes
equality when Q̄ is a diagonal matrix, which can be achieved
when UQ = UR̂22 .
R EFERENCES
[1] A. Goldsmith, Wireless Communications. Cambridge Univ. Press, 2005.
[2] D. P. Palomar, J. M. Cioffi, and M. A. Lagunas, “Uniform power
allocation in MIMO channels: A game-theoretic approach,” IEEE Trans.
Inf. Theory, vol. 49, no. 7, pp. 1707-1727, Jul. 2003.
[3] A. Wiesel, Y. C. Elder, and S. Shamai, “Optimization of the MIMO
compound capacity,” IEEE Trans. Wireless Commun., vol. 6, no. 3, pp.
1094-1101, Mar. 2007.
[4] S. Loyka and C. D. Charalambous, “On the compound capacity of a
class of MIMO channels subject to normed uncertainty,” IEEE Trans.
Inf. Theory, vol. 58, no. 4, pp. 2048-2063, Apr. 2012.
[5] Y. Sun, C. E. Koksal, and N. B. Shroff, “Capacity of compound of MIMO
Gaussian channels with additive uncertainty,” IEEE Trans. Info. Theory,
vol. 58, no. 4, pp. 2048-2063, Apr. 2012.
[6] S. Loyka and C. D. Charalambous, “Novel matrix singular value inequal-
ities and their applications to uncertain MIMO channels,” IEEE Trans.
Info. Theory, vol. 61, no. 12, pp. 6623-6634, Dec. 2015.
[7] J. Wang, M. Bengtsson, B. Ottersten, and D. P. Palomar, “Robust MIMO
precoding for several classes of channel uncertainty,” IEEE Trans. Signal
Process., vol. 61, no. 12, pp. 3056-3070, Jun. 2013.
[8] N. Shariati, J. Wang, M. Bengtsson, “Robust training sequence design for
correlated MIMO channel estimation,” IEEE Trans. Signal Process., vol.
62, no. 1, pp.107-120, Jan. 2014.
[9] K. T. Phan, S. A. Vorobyov, N. D. Sidiropoulos, and C. Tellambura,
“Spectrum sharing in wireless networks via QoS-aware secondary mul-
ticast beamforming,” IEEE Trans. Signal Process., vol. 57, no. 6, pp.
2323-2335, Jun. 2009.
[10] H. Du, T. Ratnarajah, M. Pesavento, and C. B. Papadias, “Joint
transceiver beamforming in MIMO cognitive radio network via second-
order cone programming,” IEEE Trans. Signal Process., vol. 60, no. 2,
pp. 781-792, Feb. 2012.
[11] M. Beko, “Efficient beamforming in cognitive radio multicast transmis-
sion,” IEEE Trans. Wireless Commun., vol. 11, no. 11, pp. 4108-4117,
Nov. 2012.
[12] Y. W. Huang, Q. Li, W.-K. Ma, and S. Z. Zhang, “Robust multicast
beamforming for spectrum sharing-based cognitive radios,” IEEE Trans.
Signal Process., vol. 60, no. 1, pp. 527-533, Jan. 2012.
[13] Y. Zhang, E. Dall’Anese, and G. B. Giannakis, “Distributed optimal
beamformers for cognitive radios robust to channel uncertainties,” IEEE
Trans. Signal Process., vol. 60, no. 12, pp. 6495-6508, Dec. 2012.
[14] G. Zheng, K.-K. Wong, and B. Ottersten, “Robust cognitive beamform-
ing with bounded channel uncertainties,” IEEE Trans. Signal Process.,
vol. 57, no. 12, pp. 4871-4881, Jul. 2009.
[18] B. Xie, Y. Li, H. Minn, and A. Nosratinia, “Adaptive interference
alignment with CSI uncertainty,” IEEE Trans. Commun., vol. 61, no. 2,
pp. 792-801, Feb. 2013.
[19] H. Du and T. Ratnarajah, “Robust joint signal and interference alignment
for MIMO cognitive radio network,” in Proc. IEEE WCNC, Paris, France,
Apr. 2012, pp. 448-452.
[20] M. Pei, A. Mukherjee, A. L. Swindlehurst, and J. Wei, “Rank mini-
mization designs for underlay MIMO cognitive radio networks with com-
pletely unknown primary CSI,” in Proc. IEEE GLOBECOM, Anaheim,
CA, Dec. 2012, pp. 1150-1155.
[21] K.-Y. Wang, A. M.-C. So, T.-H. Chang, W.-K. Ma, and C.-Y. Chi,
“Outage constrained robust transmit optimization for multiuser MISO
downlinks: Tractable approximations by conic optimization,” IEEE Trans.
Signal Process., vol. 62, no. 21, pp. 5690-5707, Nov. 2014.
[22] Y. J. Zhang and A. M. So, “Optimal spectrum sharing in MIMO
cognitive radio networks via semidefinite programming,” IEEE J. Sel.
Areas Commun., vol. 29, no. 2, pp. 362-373, Feb. 2011.
[23] B. Maham, P. Popovski, X. Zhou, and A. Hjørungnes, “Cognitive
multiple access network with outage margin in the primary system,” IEEE
Trans. Wireless Commun., vol. 10, no. 10, pp. 3343-3353, Oct. 2011.
[24] L. Sboui, Z. Rezki, and M.-S. Alouini, “Achievable rate of spectrum
sharing cognitive radio systems over fading channels at low-power
regime,” IEEE Trans. Wireless Commun., vol. 13, no. 11, pp. 6461-6473,
Nov. 2014.
[25] E. Dall’Anese, S.-J. Kim, G. B. Giannakis, and S. Pupolin, “Power
control for cognitive radio networks under channel uncertainty,” IEEE
Trans. Wireless Commun., vol. 10, no. 10, pp. 3541-3551, Oct. 2011.
[26] B. Maham and P. Popovski, “Cognitive multiple-antenna network with
outage and rate margins at the primary system,” IEEE Trans. Veh.
Technol., vol. 64, no. 6, pp. 2409-2423, Jun. 2015.
[27] Y. Xu, X. Zhao, and Y.-C. Liang, “Robust power control and beam-
forming in cognitive radio networks: A survey,” IEEE Commun. Survey
Tutorials, vol. 17, no. 4, pp. 1834-1857, 2015.
[28] B. Rustem and M. Howe, Algorithms for Worst-case Design and
Applications to Risk Management. Princeton, NJ, USA: Princeton Univ.
Press, 2002.
[29] J. Gorski, F. Pfeuffer, and K. Klamroth, “Biconvex sets and optimization
with biconvex functions - A survey and extensions,” Math. Methods of
Operation Research, vol. 66, no. 3, pp. 373-407, May 2007.
[30] D. P. Bertsekas, Nonlinear Programming. Belmont, MA, USA: Athena
Scientific, 2nd ed., 1999.
[31] B. Rustem, S. Žaković, and P. Parpas, “An interior point algorithm
for continuous minimax: Implementation and computation,” Optimization
Methods Software, vol. 23, no. 6, pp. 911-928, Dec. 2008.
[32] C. Xiao and Y. R. Zheng, “On the mutual information and power
allocation for vector Gaussian channels with finite discrete inputs,” in
Proc. IEEE GLOBECOM, New Orleans, LA, 2008, pp. 1-5.
[33] D. S. Bernstein, Matrix Mathematics: Theory, Facts, Formulas, 2nd ed.
Princeton, NJ, USA: Princeton Univ. Press, 2009.
[34] F. Gao, R. Zhang, Y.-C. Liang, and X. Wang, “Design of learning-based
MIMO cognitive radio systems,” IEEE Trans. Veh. Technol., vol. 59, no.
4, pp. 1707-1720, May 2010.
[35] J. Mao, J. Gao, Y. Liu, G. Xie, and X. Li, “Power allocation over fading
cognitive MIMO channels: An ergodic capacity perspective,” IEEE Trans.
Veh. Technol., vol. 61, no. 3, pp. 1162-1173, Mar. 2012.
[36] D. P. Bertsekas, A. Nedić, and A. E. Ozdaglar, Convex Analysis and
Optimization. Belmont, MA, USA: Athena Scientific, 2003.
[37] M. Grant and S. Boyd, “CVX: Matlab software for disciplined convex,”
Mar. 2014 [Online]. Available: http://cvxr.com/cvx
[38] E. G. Birgin, J. M. Martı́nez, and M. Raydan, “Nonmonotone spectral
projected gradient methods of convex sets,” SIAM J. Optimization, vol.
10, no.4, pp. 1196-1211, Jun. 2000.
[39] A. Nedić and D. P. Bertsekas, “Incremental subgradient methods for
nondifferentiable optimization,” SIAM J. Optimization, vol. 12, no. 1, pp.
109-138, Jul. 2001.
14

[40] A. W. Marshall, I. Olkin, and B. C. Arnold, Inequalities: Theory of

Majorization and Its Applications, Springer, 2nd ed., 2009.
[41] H. Yu and V. K. N. Lau, “Rank-constrained schur-convex optimization
with multiple trace/log-det constraints,” IEEE Trans. Signal Process., vol.
59, no. 1, pp. 304-314, Jan. 2011.

Mohannad H. Al-Ali (S’11) was born in Iraq. He

received the B.Sc. and M.Sc. degrees in Electrical
Engineering from the University of Basrah, Basrah
province, Iraq in 2003 and 2007, respectively. He
expects to receive his Ph.D. degree in Electrical
and Computer Engineering from the University of
Missouri, Columbia, USA on May 2017.
Mr. Al-Ali joined the Asiacell Telecommunication
Company, Basrah province, Iraq as a member of the
Project Management Department from August 2005
to December 2011. Since April 2009, he has been a
faculty in the Computer Engineering Department at the University of Basrah.
His current research interests are cognitive radio, nonlinear optimization, game
theory for wireless communications, and statistical signal processing.








 K. C. Ho (S ’89 – M ’91 – SM ’00 – F ’09) was

 born in Hong Kong. He received the B.Sc. degree

 with First Class Honors in Electronics in 1988 and

 the Ph.D. degree in Electronic Engineering in 1991.
 He was a research associate in the Royal Military
College of Canada from 1991 to 1994. He joined the


Bell-Northern Research, Montreal, Canada in 1995



 as a member of scientific staff. He was a faculty in

 the Department of Electrical Engineering at the Uni-

 versity of Saskatchewan, Saskatoon, Canada from

 September 1996 to August 1997. Since September
1997,

 he has been with the University of Missouri and is currently a Professor
 the Electrical Engineering and Computer Science Department. His research
in
interests are in sensor array processing, source localization, subsurface object
detection, wireless communications and adaptive processing.
He was active in the development of the ITU-T Standard Recommendations
from 1995 to 2012. He was the Rapporteur of ITU-T Q15/SG16: Voice
Gateway Signal Processing Functions and Circuit Multiplication Equip-
ment/Systems from 2009 to 2012, and the Associate Rapporteur of ITU-T
Q16/SG16: Speech Enhancement Functions in Signal Processing Network
Equipment in 2012. He was the Editor of the ITU-T Recommendations G.160:
Voice Enhancement Devices from 2006 to 2012, G.168: Digital Network
Echo Cancellers from 2000 to 2012 and G.799.2: Mechanism for Dynamic
Coordination of Signal Processing Network Equipment from 2004 to 2009.
Dr. Ho served as a Technical Co-Chair of the IEEE International Conference
on Acoustics, Speech and Signal Processing 2016 (ICASSP2016). He was
the Past Chair (2015), Chair (2013-2014) and Vice-Chair (2011-2012) of the
Sensor Array and Multichannel (SAM) Technical Committee of the IEEE
Signal Processing Society. He participated in the organizing committees of
the IEEE SAM 2008 Workshop and the IEEE CAMSAP 2011. He was an
Associate Editor of the IEEE Transactions on Signal Processing (2003-2006,
2009-2013) and the IEEE Signal Processing Letters (2004-2008). He received
the Senior Faculty Research Award in 2009 and 2014, the Junior Faculty
Research Award in 2003 and the Teaching Award in 2006 from the College
of Engineering at the University of Missouri. He is an inventor of 22 patents
in the United States, Canada, Europe and Asia on geolocation and signal
processing for wireless communications.