Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
5699
AbstractWe propose a learning technique for MIMO communication systems to perform spatial division multiple access with
minimal cooperation between users. In the proposed technique,
each user (in a two-user receivertransmitter pair) learns the null
space of the interference channel to the other user by transmitting
a learning signal and observing an affine function of the other
users interference plus noise power. The only requirement is that
each system broadcasts, through a low-rate control channel, a
periodic beacon that is a function of its noise plus interference
power, which in practice is typically known by each systems receiver and transmitter. Thus, the learning can be made by the two
users transmitters without affecting the communication protocol
between each users receiver and transmitter. The proposed learning scheme is particularly attractive for underlay cognitive radio,
where only the secondary user (SU), which must not interfere with
the primary user (PU), has to learn the null space. In this case, the
PU can broadcast the schemes beacon without being aware of the
SU. Furthermore, if the PU uses a power control mechanism which
maintains a constant signal to interference plus noise ratio, the SU
can learn the null space even without a beacon, i.e., without any
cooperation with the PU.
Index TermsMIMO, interference mitigation, null space, precoding, spectrum sharing, cognitive radio.
I. I NTRODUCTION
Fig. 1. Two MIMO users with equal priority that share the same channel. The
matrix Hi,j , i = j {1, 2} are unknown to both user 1 and 2. The objective
of user i {1, 2} is to learn the null space of Hij , i = j {1, 2}.
1536-1276 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
5700
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
Fig. 2. The addressed cognitive radio scheme. The matrix H12 is unknown
to the secondary transmitter and v1 (k) is a stationary noise (which may
include stationary interference). The SU transmits a beacon {q(k)}T
k=1 , which
indicates the total interference plus noise power at the PU-Rx. Note that if the
latter is known at the PU-Tx, the beacon can be transmitted by the PU-Tx,
leaving the PU-Rx communication protocol unaffected.
i = j {1, 2}.
(2)
5701
t N,
(1)
(3)
5702
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
Fig. 3. The time indexing used in this paper. At the first TF B time slots of
(n). Then, U2 Tx measures the
each TC, U2 Tx transmits its learning signal x
feedback q(n). The exact time t {(n 1)Ts + TF B , . . . , nTs } in which
q(n) becomes available to U2 Tx depends on the feedback scheme as we
discuss in Section IV.
(n)) = ( + A(n)) H
x(n)2 + c + C(n), (4)
q(n) = b (n, x
where , c, are constants which may be unknown, and A(n)
and C(n) are zero mean random variables that represent the
measurement error. The term s(t) is not directly involved in the
learning process. However, as will be shown in Section IV, s(t)
is important in enabling U2 to extract q(n). U2s Tx objective
is to estimate N (H) from {
x(n), q(n)}N
n=1 . In the ideal case,
U2s Tx observes q(n) with A(n) = C(n) = 0. The problem of
designing the learning signals {
x(n)}N
n=1 for estimating N (H)
N
from {
x(n), q(n)}n=1 , is called the ENSL problem.
Before discussing how these measurements may be available
to U2 (Section IV), we describe the advantages of the proposed
scheme. First, the feedback relies on measurements that already
take place in many communications systems; e.g., the noise
plus interference at the receiver (Section IV-A), since such
information is necessary for power, modulation, and/or rate
adaptation. In such cases, U1s Rx has to forward existing information to U2s Tx without taking additional measurements. In
addition, U1 does not even have to be aware of the U2 system.
The former only broadcasts its total interference plus noise
levels to all the users in its vicinity; i.e., U1s Rx can broadcast
a one-dimensional beacon q(n) which indicates its interference
level and the time slots that it is measured. This beacon not only
indicates U1 Rxs total interference, but also the time instances
in which U2s Tx should send a new learning signal. Note that
such a scheme, which puts most of the learning burden on U2s
Tx, is especially useful in underlay CR, where U1 is the PU
and U2 is the SU. Furthermore, the proposed scheme requires
a much simpler synchronization between U1 and U2 than the
symbol-by-symbol synchronization required in standard channel estimation techniques. Specifically, if a beacon feedback
method is used, U2s Tx must synchronize to the symbol rate
of the beacon, which can be transmitted by a low rate channel
separated from the U1 channel. Actually, it can even be an
analog signal, where the information is encoded in the signals
amplitude. In this case, U2s Tx could detect variations in the
beacon and measure the levels of these variations. Then, the PU
is not required to have any handshake and/or synchronization
with the SU, and the latter does not interrupt the formers
data flow. Note that U1s Rx need not necessarily transmit the
beacon q(n) all the time. Rather, it can start transmitting it only
when it senses that the received interference increases above
some threshold or upon receiving a request from U2s Tx. In
5703
= sT P s,
V(n)
= z sT + V(n),
z = H
x(n) =
PL Hr(n),
(n1)Ts +TF B
qN I (n) =
2
11 pk
y1 (t) P (n)H
.
t=(n1)Ts +1
(5)
c = v2 nr (TF B nt ),
qT RP (n) =
1
T s PL
nTs
y1 (t)22 .
(6)
t=(n1)Ts +1
2
P H11 2F
+ nr v ,
PL
PL
= 1,
c=
nr v2
P 2 H11 2F
+
,
PL
PL
5704
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
Fig. 4. Upper bound on the interference reduction for all feedback schemes.
(a) Ts = 1 ms, TF B = (nt + 1)/Tsym ms, T /v2 = 5 dB. (b) Ts =
5 ms, TF B = (nt + 1)/Tsym ms, T /v2 = 5 dB. (c) Ts = 5 ms, TF B =
(nt + 7)/Tsym ms, T /v2 = 5 dB. (d) Ts = 5 ms, TF B = (nt +
7)/Tsym ms, T /v2 = 10 dB.
the beginning
of the n-TC, U1s Tx transmits the pilot signals
P
= [p1 , . . . , pT ] during TF B slots,
Pn = P (n 1)P,
FB
and U1s Rx estimates its SINR as
(n1)T
s +TF B
(n) =
(n1)T
s +TF B
2
11 pk
y1 (t) H
where
=
2 ,
11 Pn
Y(n) H
(8)
T
,
(n)
(9)
(11)
nTs
(16)
2
with a = 2H
x(n)2 1T
TF B P 1TF B /v .
Proof: See Appendix C.
For large T , E{P (n)} can be approximated as
T 1T
2 T nr (TF B nt )
TF B P 1TF B
H
x(n)2 + v
,
2
2
H11 P
H11 P
F
t=(n1)Ts +TF B +1
2
E {P (n)}
E {qN B (n)} =
H21 2F + nr2 v2 ,
PL
PL
(13)
where
2
PL T 1T
TF B P 1TF B Hr(n)
2
H11 P
F
2 2 + 2a (na + 2) + na (na + 2)
E P (n)2 = T a
n2b 6nb + 8
(nb /2 1)(nb /2 2)v4
5/2
,
+
O
T
4
H11 P
v2 T nr (TF B nt )
3/2
+ O(T ).
2
H11 P
(14)
3 In
1
2v22 E {P (n)} H21 2F
PL2 (Ts TF B )
+ E P (n)2 + H21 H21 2F + v42 nr2 , (15)
y1 (t)22
11 Pn 2
H
F
E {P (n)} =
t=(n1)Ts +1
t=(n1)Ts +1
qN B (n) =
5705
.
(n) to learn G =
In the ENSL algorithm, U2s Tx chooses x
H H Cnt nt . From G it is sufficient to obtain N (H) since
x N (H) x N (G).
G is a Hermitian semi-positive definite matrix with an eigenvalue decomposition G = WW where W is a unitary matrix, i.e., W W = I and is a diagonal matrix whose elements
are non-negative. The columns of W are the orthonormal
eigenvectors of G. We assume in this stage that q(n) has no
noise, i.e.,
q(n) = Hr(n)22 + c = r(n) H Hr(n) + c
= r(n)Gr(n)
+ c,
= G. Note that in each TC, U2s Tx obtains a linear
where, G
and that n2 + 1 TCs are sufficient
equation of the elements of G
t
since there are nt real diagonal, nt (nt 1)/2
to learn G,
complex upper diagonal elements and the unknown c. It is not
= N (G).
necessary to learn since N (G)
5706
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
entries of G. Then,
gl,l = S (rl,m (0, 0)) S(0),
(
gl,m ) = cl,m (/4, 0),
(
gl,m ) = cl,m (/4, /2),
1 l = m nt ,
1 l < m nt ,
1 l < m nt ,
(17)
where
cl,m (, ) =S(rl,m (, ))S(0)
gl,l cos2 ()
gm,m sin2 ().
Proof: See Appendix D.
in which
Proposition 4 implies an algorithm to estimate G
(n)), given in (4).
S(x) is replaced by its noisy version b(n, x
(n))
In the case where the number of the samples of b(n, x
is equal to the number of equations, this is the least squares
The estimate of G,
denoted as G,
is obtained
estimate of G.
(n) represents a difvia |R| = n2t + 1 TCs, where for each n, x
ferent element of R. A pseudo-code description of the proposed
TG
= vi1 , . . . , vint nr ,
(18)
i ,
viv Gv
v
q v.
(19)
= ENSL(b, nt )
Algorithm 1 G
Let b(n, x) be defined in (4)
c b(0, 0); n 1;
for l = 1, . . . , nt do
gl,l b(n, rl,1 (0, 0)) c; n + +;
end for
for l = 1, . . . , nt 1 do
for m = l + 1, . . . , nt do
gl,l +
gm,m )/2)b(n, rl,m (/4, 0))+
c;
l,m ((
n + +;
gl,l +
gm,m )/2)b(n, rl,m (/4, /2))+
c;
l,m ((
n + +;
gl,m l,m + il,m ;
;
gm,l gl,m
end for
end for
E=G
Denote as n0 , nl,l , nl,m , and nl,m + 1 the indices of the TCs in
gl,m ), and (
gl,m ) are estimated, respectively,
which c, gll , (
and as
0 = Var {q(n0 )} , l,l = Var {q(nl,l )} ,
l,
and as
l,m = rl,m + il,m ,
with
(20)
l<m
where TG
is defined in (18).
Proof: See Appendix E.
In what follows, we calculate this bound for each of the
feedback schemes of Section IV.
5707
(TF B nt )nr
c2 =
2
1T
TF B P 1TF B
2
Then, assuming [H]i,j CN (0, H
) and
Proposition 4 in Proposition 5 it follows that
v2
PL
2
.
substituting
quadratic function of H
x(n)2 . Assuming for simplicity that
2
2
v = v2 , it holds from (15) that
2 2
1
T v
Var {qN B (n)} =
c1 O T2 + c2 O
Ts TF B
PL
2 4
T v
+ c3 O
+ O(T ) , (23)
PL2
where
c1 =
2
E HTG
|H
c2 =
2E {H2F }
+ (6n2t 4nt ) c2
2 c1 (2nt + 1)
2
.
2 n n (2n + 1) + (6n2 4n ) c ,
= BN I = 2 c1 H
t r
t
t 2
t
(21)
where (a) follows from Jensens inequality. In the case where
2
is greater than the noise floor, the first term under the
PL H
square root is dominant and therefore
2
8nt nr (2nt + 1)v2 H
BN I
.
(22)
PL 1T
TF B P 1TF B
T
Note that maxP 1T
TF B P 1TF B = 1TF B 1TF B = TF B . Thus,
1/2
the bound in (22) decreases as TF B
. Furthermore, U2s Tx
learning-signals power reduces the bound as PL 1/2 .
2) Total Received Power Feedback (TRP): We now analyze
the behavior of the upper bound in (20) for Var{qT RP (n)}
(7). To simplify the exposition, we use the fact that
P 2 H11 H11 2F /PL2 is the dominant term in Var{qT RP (n)}
since U1 Txs power P must be large enough such that
P H11 2 is much greater than PL Hr(n) and v2 , as discussed in Section III. Thus, since
2
E HTG
|H
2
.
4n2t nt P 2 H11 H11 2F
BT RP =
T s PL
+ O P v2 H11 2F + O v2 PL Hr(n)2
1/2
+ O P H11 Hr(n)2 PL + O v2
it follows that
(4n2t nt ) H11 H11 F /( Ts PL ).
Hr(n)
.
4
H11 PF
(a)
BT RP 2P
c3 =
(24)
+2
1/2
Hr(nl,m )
O(T )+O
l<m
T v
1/2
PL
+
1/2
T
(25)
Observe that this bound does not depend on PL as long as
PL v2 , which indicates the existence of an upper floor on the
achieved interference. The numerical results in Fig. 4 corroborate this observation. In particular, Fig. 4 plots the average value
of the normalized upper bounds presented above as a function
of the interference to noise ratio
.
IN R = PL /v2 ;
i.e., it uses Monte Carlo simulations to estimate
E 2 E|H
PU B = E
,
G
(26)
5708
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
(nt nr ) (nr + 1) + 1
2
(2)
1
v(2) v(1) = u
(1)
(1)
(1)
1
= u(2)
U(0) u
(1) U(0)
U
0
(1)
1 = 0,
= u(2) (1) u
(27)
U
= ENSL(Hm
U
eq , nr + 1);
(m)
1 ;
vm = U(m1) u
(m)
(m)
m = [
nr +1 ];
u2 , . . . , u
U
(m)
(m) ]
U(m) = [wnt nr m , U(m1) U
end for
PIR = E
,
(28)
G
where T is the resulting precoding matrix after one run of the
.
ENSL algorithm, as a function of IN R = PL /L12 v2 , which
represents the interference to noise ratio of U1s Rx averaged over the fast fading, for each feedback scheme for static
channels; i.e., Fd = 0 Hz. The figures also plot the respective
normalized analytical upper bound (26). For comparison, we
plot the worst case interference reduction when T is a random
unitary matrix. We used nt = 3, nr = 2, and t = 10 dB. In
Fig. 5(c), we use a larger interval, Ts = 10 msec, than Fig. 5(a)
and (b), because it takes more time for U1s Rx to estimate the
TRP beacon (see Section V-B2). The results show that the TRP
scheme has a similar performance as the NPI scheme with a
sensing interval Ts larger by an order of magnitude. Also, the
NB scheme has similar performance as the NPI scheme when
IN R is small, even though U1 and U2 do not cooperate in
5709
Fig. 5. Simulation result of (a) NPI, (b) NB, (c) TRP schemes, for TF B =
1 ms. (d) Presents a comparison between the ENSL and R-ENSL algorithms
for nt = 8, nr = 1 for the NPI beacon. The results were averaged over 103
Monte Carlo trials.
5710
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
Fig. 7. Simulation results of NPI, NB, and TPR schemes as a function of the
number of transmit antennas nt for Rayleigh fading (K = 0) and Fd = 3 Hz.
(a) NPI performance as a function of nt . Ts = 1 ms, INR = 10 dB, Fd =
3 Hz, K = 0. (b) NB (Ts = 1 ms) and TPR (Ts = 10 ms) performance as a
function of nt . INR = 10 dB, Fd = 3 Hz, K = 0.
users transmitters without affecting the communication protocol between each users receiver and transmitter. This property
is very appealing in cases were the receivers are end user equipment, which is difficult to upgrade, or if one wishes to avoid
performing computations related to interference mitigation at
the receivers to save energy consumption.
An important future research direction is to extend this
framework to more than two interfering systems. Specifically,
note that the proposed algorithms work without modifications
in cases where only one of the interfering systems performs
null space learning, while the remaining interferers keep the
statistics of their inflicted interference unchanged; i.e., keep
their precoding matrices and power levels constant. However,
if two or more systems try to learn their null space towards
a specific user concurrently, the feedback they are going to
receive will not be an affine function of the inflicted interference
that originated from both of them, and thus the proposed
schemes will not work.
A PPENDIX A
2
11 Pn , and that
Note that qN I (n) = Y(n) H
F
11 Pn = H11 Pn + V(n)
H11 Pn + V(n)
Pn Pn
Y(n) H
= V(n)P
(29)
5711
then,
To simplify notation, we omit index n from Y, V,
V
11 Pn 2 = Tr (P ) V
VP
= Tr{VP
}
Y H
F
= Tr{z sT P z s} + v,
(30)
where z = H
x(n) and v = 2(Tr{z sT P V }) +
Tr{VP V } is a non zero-mean noise. Because P is a
projection matrix it can be decomposed as
U
,
P = U
(31)
E{V V}U
E Tr{VP V } = Tr U
U}
= nr 2 (TF B nt ),
= nr v2 Tr{U
v
v }2 = 2E{a
a} + E{bb} E{
v }2
Var{
v } = E{
v v } E{
= 2v2 (sT P s)H
x(n)2
+ v4 (TF B nt )nr ((TF B nt )nr + 1)
2
nr v2 (TF B nt )
v } the result follows.
Then, since Var{qN I (n)} = Var{
(35)
(32)
A PPENDIX B
and
Tr{z sT P P z s}
= Tr (z sT )(1 P )(1 P )(z s)
= Tr z (sT P )z (P s)
= Tr (zz ) (sT P P s)
qT RP (n) =
(33)
Thus, E{
v }=nr v2(TF Bnt) since E{2(Tr{zsT P V })}=
0. Therefore, the E{qN I (n)} follows. Also, from (33) we see
that A(n) = 0, n.
It remains to derive the variance of qN I (n). To this
end, we calculate the variance of v = 2(a) + b, where a =
Tr{z sT P V } and b = Tr{VP V }. Because the entries of V are i.i.d. circular Gaussian random variables,
the random variables a, b are also circular Gaussian and
+ b)(a + a
+ b)} = 2E{a
a} +
therefore E{
v v } = E{(a + a
T F B PL
y1 (t)22
t=(n1)Ts +1
= Hr(n) + c + C(n),
= Tr{zz }Tr{sT P s}
= (sT P s)H
x2 .
nTs
(34)
v2 P H11 2F
+
.
PL
PL
(36)
5712
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
A PPENDIX C
At each TC, U1s Tx chooses the next P (n) based on P (n
1), T and the energy measurements U1s RX performs at the
first TF B slots. To simplify the notation, we omit the index n
from Y(n) and V(n). We begin by analyzing the following
quantity:
P (n)
1
=
(n) P (n 1)T
2
H
x(n) 1T
a
TF B P +VP
F
=
2 = ,
b
x(n) 1T
H11 Pn +H
TF B Pn Pn +VPn Pn
F
(37)
where
2
11 Pn 2 = H
x(n) 1T
a = Y H
TF B P + VP
F
F
and
2
11 Pn
b = H
F
2
= H11 Pn + H
x(n) 1T
TF B Pn Pn + VPn Pn F .
We now derive the conditional distribution of 1/
(n) given
P (n 1). To this end, we first derive the distribution of a
and b. Since P is a projection matrix, there exists U
TF B (TF B nt )
j u
l v
k ]} = Tr [E {
i} u
l ]
j u
vi u
vk v
E {Qij Qkl } = E {Tr [
j u
l = v2 ki Tr {
l }
= Tr v2 ki ITF B u
uj u
j = v2 ki lj .
lu
= v2 ki Tr u
(39)
2
+
In the same manner, b = Pn Pn + VPn Pn F = U
T
is a ma 2 , where = H11 Pn + H
x
(n)
1
and
U
VU
F
TF B
to an orthonortrix whose columns complete the columns of U
TF B
mal basis of C
; i.e., [U, U] [U, U] = ITF B . Note that the
are circular Gaussian i.i.d. and that VU
and
entries of VU
x(n) 1T
2 H11 Pn + H
TF B Pn Pn F
b =
2
v
respectively, where the latter equality follows from (33).
The degrees of freedom are given by na = 2nr (TF B nt )
and nb = 2nr nt , respectively. Therefore, given P (n 1),
(1/
(n))(nb /na ) is a doubly non-central F-distributed random
variable [35], with expectation
1 nb
|P (n1) = E F(na ,nb ,la ) 1 F1 (1, nb /2, b /2)
E
(n) na
(41)
funcwhere 1 F1 (a, b, c) is the confluent hypergeometric
j
tion, defined as 1 F1 (a, b; z) = ((b)/(a))
j=0 (z (a +
j)/j!(b + j)) and F(na ,nb ,a ) is a non-central (single)
F random variable with na degrees of freedom in the numerator, non-centrality parameter a and nb degrees of freedom in the denominator. Because, E{F(na ,nb ,a ) } = nb (na +
a )/na (nb 2) for nb > 2, (41) it follows that
E{P (n)|P (n1)}= P (n1)T
na +a
1 F1 (1, nb /2,b /2).
nb 2
(42)
The series expansion of the confluent hypergeometric function [36, p. 466] is given by
1 F1 (a, b; z)
(38)
Let = H
x(n) 1T
TF B + V, then a = U + QF , where
nr (TF B nt )
Q = VU C
is a circular Gaussian random matrix with i.i.d. entries. To see that, consider E{Qij Qkl } =
= 1 F1 1, nb /2; (y + x)2
= (nr nt 1)(y + ax)2 + O (x y)3 ,
(43)
P
P
/
.
In
(43)
we
used
the
triangle
inequality
1T
v
TF B n n F
H
x(n) 1T
x(n)1T
TF B Pn Pn + H11 Pn F H
TF B Pn Pn
F + H11 Pn F . We now derive expressions that enable us to
use the assumption x y. We begin with
1 F1 (1, nb /2; b /2)
5713
where
E
x
= (nr nt 1) y 2 1 + O
+ O (x + y)3
y
x
= (nr nt 1)y 2 + O
+ O (x + y)3
3
y
= (nr nt 1)y 2 + O (x + 1)y 3 + xy 4 ,
where in the last equality we used the fact that
1
1
3
=O
O (x + y)
= O(y 3 + xy 4 ).
y 3 (1 + x/y)3
+ a )
3
+ xy
2 + O (x + 1)y
2H11 PF
4
T 1T
2 T nr (TF B nt )
TF B P 1TF B
H
x(n)2 + v
2
2
H11 P
H11 P
F
F
3
.
+ O (x + 1)y
nb
na
2
(nb /2 2)(na /2 + 2)
(nb /2)(na /2)
,
(45)
and, with the same reasoning as in (43),
1 F1 (2, nb /2; b /2)
= (nb /21)(nb /22) (y(1+x/y))4 +O (x+y)5
= (nb /2 1)(nb /2 2)y 4 + O (x + 1)y 5 + xy 6 .
(46)
Using (44) and (9) it holds that E{P (n)2 |P (n 1)} = (na /nb )2
P (n1)2 E{(F(na ,nb ,la ) )2 }1 F1(2,nb /2,b /2). Thus, after substituting (45), (46), and using the inequality P (n 1)
t v2 , (16) follows.
E {P (n)|P (n 1)}
= T
= (nr nt 1) (y(1 + x/y))2 + O (x y)3
v2 (na
2
A PPENDIX D
(n) = rl,m (, ) the idealUsing the specific input signals x
istic feedback signal can be written
.
(n) G
x(n) + c
q(n) = x
= gl,l cos2 (n ) + gm,m sin2 (n )
+ |gl,m | sin(2n ) cos(gl,m n ) + c.
(47)
= 0 then
Obviously, if U2s Tx does not send any signal, i.e., x
(n) =
the feedback will only observe c, i.e., c = H02 . If x
rl,m (0, 0), for any m = l, then
q(n) = gl,l + c gl,l = b (nl,l , rl,1 (0, 0)) c.
(n) = rl,m (/4, 0) then
If x
q(n) = gl,l /2 + gm,m /2 + |gl,m | cos(gl,m ) + c
(gl,m ) = cl,m (/4, 0).
(48)
(49)
A PPENDIX E
We have that
2
GT = T (G E + E)T
HTG
= TG
G
G
G
(50)
TG
+ TG
.
GTG
ETG
5714
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 13, NO. 10, OCTOBER 2014
and
= n +1 (G),
By (19), TG
r
GTG
T ET = max xT ET x
G
G
G
G
R EFERENCES
x=1
max
y=1,yC(TG
)
(51)
Thus,
2
HTG
E + nr +1 (G).
(52)
(b) 2
E {|
glm glm |2 }
lm
Var{
gl,l } + 2
Var{
gl,m }.
E HTG
l<m
(53)
The estimation error in G occurs due to measurements errors in the feedback (4). As it is presented in Algorithm 1,
c is estimated at n = 0 as c = q(0) = b(0, 0) = C(0).
Then, from Proposition 4 we see that gll = q(nl,l )
gl,l + gm,m ) c and (
gl,m ) =
c, (
gl,m ) = q(nl,m ) 0.5(
c} = 0 ,
0.5(
gl,l + gm,m ) q(nl,m + 1) + c. Therefore, Var{
Var{
gll } = l,l + 0 , and
1
Var {(
gl,m )} = (l,l + m,m ) + rl,m +
4
1
Var {(
gl,m )} = (l,l + m,m ) + il,m +
4
3
0 ,
2
3
0 .
2
1
(l,l + m,m ) + l,m + 30 .
2
l<m
G.
The
latter
bound is
[1] S. Jafar and M. Fakhereddin, Degrees of freedom for the MIMO interference channel, IEEE Trans. Inf. Theory, vol. 53, no. 7, pp. 26372642,
Jul. 2007.
[2] Q. Spencer, A. Swindlehurst, and M. Haardt, Zero-forcing methods for
downlink spatial multiplexing in multiuser MIMO channels, IEEE Trans.
Signal Proces., vol. 52, no. 2, pp. 461471, Feb. 2004.
[3] L. Ruan and V. Lau, Dynamic interference mitigation for generalized partially connected quasi-static MIMO interference channel, IEEE Trans.
Signal Proces., vol. 59, no. 8, pp. 37883798, Aug. 2011.
[4] A. Goldsmith, S. Jafar, I. Maric, and S. Srinivasa, Breaking spectrum
gridlock with cognitive radios: An information theoretic perspective,
Proc. IEEE, vol. 97, no. 5, pp. 894914, May 2009.
[5] R. Zhang and Y.-C. Liang, Exploiting multi-antennas for opportunistic
spectrum sharing in cognitive radio networks, IEEE J. Sel. Topics Signal
Process., vol. 2, no. 1, pp. 88102, Feb. 2008.
[6] G. Scutari, D. Palomar, and S. Barbarossa, Cognitive MIMO radio,
IEEE Signal Process. Mag., vol. 25, no. 6, pp. 4659, Nov. 2008.
[7] G. Scutari and D. Palomar, MIMO cognitive radio: A game theoretical
approach, IEEE Trans. Signal Process., vol. 58, no. 2, pp. 761780,
Feb. 2010.
[8] K. Hamdi, W. Zhang, and K. Letaief, Opportunistic spectrum sharing
in cognitive MIMO wireless networks, IEEE Trans. Wireless Commun.,
vol. 8, no. 8, pp. 40984109, Aug. 2009.
[9] S. Zhou, B. Muquet, and G. Giannakis, Subspace-based (semi-) blind
channel estimation for block precoded space-time OFDM, IEEE Trans.
Signal Process., vol. 50, no. 5, pp. 12151228, May 2002.
[10] Z. Ding and L. Qiu, Blind MIMO channel identification from second
order statistics using rank deficient channel convolution matrix, IEEE
Trans. Signal Process., vol. 51, no. 2, pp. 535544, Feb. 2003.
[11] C. Shin, R. Heath, and E. Powers, Blind channel estimation for MIMOOFDM systems, IEEE Trans. Veh. Technol., vol. 56, no. 2, pp. 670685,
Mar. 2007.
[12] G. Scutari, D. Palomar, and S. Barbarossa, Competitive design of multiuser MIMO systems based on game theory: A unified view, IEEE J. Sel.
Areas Commun., vol. 26, no. 7, pp. 10891103, Sep. 2008.
[13] G. Scutari, D. Palomar, and S. Barbarossa, The MIMO iterative
waterfilling algorithm, IEEE Trans. Signal Process., vol. 57, no. 5,
pp. 19171935, May 2009.
[14] F. Chen, W. Su, S. Batalama, and J. Matyjas, On transceiver beamformer
design for multi-source multi-destination wireless networks, in Proc.
IEEE GLOBECOM, Dec. 2011, pp. 16.
[15] R. Zhang, Y.-C. Liang, and S. Cui, Dynamic resource allocation in
cognitive radio networks, IEEE Signal Process. Mag., vol. 27, no. 5,
pp. 102114, May 2010.
[16] E. Larsson and E. Jorswieck, Competition versus cooperation on the
MISO interference channel, IEEE J. Sel. Areas Commun., vol. 26, no. 7,
pp. 10591069, Sep. 2008.
[17] L. Zhang, Y.-C. Liang, Y. Xin, and H. V. Poor, Robust cognitive beamforming with partial channel state information, IEEE Trans. Wireless
Commun., vol. 8, no. 8, pp. 41434153, Aug. 2009.
[18] Y. J. Zhang and A. M.-C. So, Optimal spectrum sharing in MIMO cognitive radio networks via semidefnite programming, IEEE J. Sel. Areas
Commun., vol. 29, no. 2, pp. 362373, Feb. 2011.
[19] Z. Chen et al., Interference mitigation for cognitive radio MIMO systems
based on practical precoding, Physical Communication, 2012.
[20] H. Yi, Nullspace-based secondary joint transceiver scheme for cognitive
radio MIMO networks using second-order statistics, in Proc. IEEE Int.
Conf. Commun., May 2010, pp. 15.
[21] Y. Noam and A. Goldsmith, Blind null-space learning for MIMO underlay cognitive radio with primary user interference adaptation, IEEE
Trans. Wireless Commun., vol. 12, no. 4, pp. 17221734, Apr. 2013.
[22] Y. Noam and A. J. Goldsmith, The one-bit null space learning algorithm and its convergence, IEEE Trans. Signal Process., vol. 21, no. 24,
pp. 61356149, Dec. 2013.
[23] R. Zhang, F. Gao, and Y.-C. Liang, Cognitive beamforming made practical: Effective interference channel and learning-throughput tradeoff,
IEEE Trans. Commun., vol. 58, no. 2, pp. 706718, Feb. 2010.
[24] K. Huang and R. Zhang, Cooperative feedback for multiantenna
cognitive radio networks, IEEE Trans. Signal Process., vol. 59, no. 2,
pp. 747758, Feb. 2011.
[25] H. Wang, Y. Huang, K.-K. Wong, and L. Yang, Asymptotic performance
of two-user interference channels using coordinated zero-forcing, IEEE
Commun. Lett., vol. 16, no. 5, pp. 608611, May 2012.
[26] R. Zhang, On active learning and supervised transmission of spectrum sharing based cognitive radios by exploiting hidden primary
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
radio feedback, IEEE Trans. Commun., vol. 58, no. 10, pp. 29602970,
Oct. 2010.
S. Huang, X. Liu, and Z. Ding, Decentralized cognitive radio control
based on inference from primary link control information, IEEE J. Sel.
Areas Commun., vol. 29, no. 2, pp. 394406, Feb. 2011.
J. Wang, G. Scutari, and D. P. Palomar, Robust MIMO cognitive radio via
game theory, IEEE Trans. Signal Process., vol. 59, no. 3, pp. 11831201,
Mar. 2011.
M. Biguesh and A. Gershman, Training-based MIMO channel estimation: A study of estimator tradeoffs and optimal training signals, IEEE
Trans. Signal Proces., vol. 54, no. 3, pp. 884893, Mar. 2006.
M. Wax and T. Kailath, Detection of signals by information theoretic
criteria, IEEE Trans. Acoust., Speech, Signal Process., vol. 33, no. 2,
pp. 387392, Apr. 1985.
A. P. Liavas, P. A. Regalia, and J.-P. Delmas, Blind channel approximation: Effective channel order determination, IEEE Trans. Signal Process.,
vol. 47, no. 12, pp. 33363344, Dec. 1999.
N. S. Zuckerman, Y. Noam, A. J. Goldsmith, and P. P. Lee, A self-directed
method for cell-type identification and separation of gene expression
microarrays, PLoS Comput. Biol., vol. 9, no. 8, p. e1003189, 2013.
A. Tulino and S. Verd, Random Matrix Theory and Wireless Communications. Delft, The Netherlands: Now Publishers Inc., 2004.
K. S. Miller, Complex Stochastic Processes: An Introduction to
Theory and Application, Advanced Book Program. Reading, MA,
USA: Addison-Wesley, 1974.
M. Bock, G. G. Judge, and T. Yancey, A simple form for the inverse
moments of non-central 2 and f random variables and certain confluent
hypergeometric functions, J. Econometrics, vol. 25, no. 1, pp. 217234,
1984.
W. J. Thompson, Atlas for Computing Mathematical Functions: An Illustrated Guide for Practitioners, With Programs in C and Mathematica.
Hoboken, NJ, USA: Wiley, 1997.
R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge,
U.K.: Cambridge Univ. Press, 1987.
5715