Sei sulla pagina 1di 6

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO.

3, MARCH 2012

649

Multiuser MIMO User Selection Based on Chordal Distance


Kyeongjun Ko and Jungwoo Lee, Senior Member, IEEE

AbstractMultiuser MIMO (MU-MIMO) systems have advantages over single-user MIMO systems in terms of system
performance. In MU-MIMO systems, inter-user interference
needs to be dealt with especially when linear processing is used.
The block diagonalization (BD) method is one of techniques that
are widely used to eliminate the inter-user interference. In a
cellular system where there are many users, the subset of users
which maximizes the system performance should be selected
since the base station cannot support all the users in the cell. In
this paper, we propose a low complexity MU-MIMO scheduling
scheme using BD with chordal distance. For a large number
of users, the optimal scheduling technique needs an exhaustive
search, which is impractical. One of the key ideas of this paper
is to use chordal distance as a measure of orthogonality between
different users. Simulation results show the proposed algorithm
has throughput close to the optimal scheduling scheme with lower
complexity than existing low complexity scheduling algorithms.
Index TermsMU-MIMO systems, scheduling, chordal distance.

I. I NTRODUCTION
ULTIPLE-INPUT multiple-output (MIMO) systems
have been drawing a lot of attention because of spectral
efficiency and diversity gain [1], [2]. It is known that selecting the user who has the best channel is optimal in terms
of capacity in multiuser single-input single-output (SISO)
systems [3]. However, it is optimal to serve multiple users
simultaneously in multiuser MIMO systems [4]. Inter-user
interference needs to be dealt with in multiuser MIMO (MUMIMO) systems since data are transmitted to multiple users in
the same frequency simultaneously. Dirty paper coding (DPC)
is an optimal technique to remove the inter-user interference
[5]. It was shown that the interference which is known a
priori to the transmitter does not affect the channel capacity.
DPC is a nonlinear scheme that achieves the channel capacity
in MIMO broadcast channel. However it is impractical for
implementation since its complexity is prohibitive.
In order to lower the complexity, many precoding algorithms for MU-MIMO systems have been proposed. Zeroforcing (ZF) and block diagonalization (BD) have been used
widely among them [6][8]. The ZF scheme removes interuser interference among the selected users by using a precod-

Paper approved by D. I. Kim, the Editor for Spread Spectrum Transmission


and Access of the IEEE Communications Society. Manuscript received March
10, 2011; revised June 28, 2011 and September 10, 2011.
This research was supported in part by the Basic Science Research Program
(KRF-2008-314-D00287, 2010-0013397) and the Mid-Career Researcher Program (2010-0027155) through the NRF funded by the MEST, Seoul R&BD
Program (JP091007, 0423-20090051), the KETEP grant (2011T100100151),
the INMAC, and BK21. This paper was presented in part at the Conference
on Information Sciences and Systems (CISS), Princeton, NJ, March 2008.
The authors are with the School of Electrical Engineering and Computer Sciences, Seoul National University, Seoul 151-744, Korea (e-mail:
smuff@wmspl.snu.ac.kr; junglee@snu.ac.kr).
Digital Object Identifier 10.1109/TCOMM.2012.020912.110060

ing matrix, which is the pseudo-inverse of the selected users


channels. It is typically used in multiple-input single-output
(MISO) systems. On the other hand, the BD scheme uses
null space which is computed by singular value decomposition
(SVD). For a given user, a matrix is constructed by stacking
all the users channels except its own channel, and finds the
null space by SVD. However, the number of users who can be
simultaneously supported with BD is limited by the number
of transmit antennas and the number of receive antennas since
each users precoding matrix must lie in the null space of all
the other users channels. It can be considered as an extension
of ZF, and can be used for MU-MIMO systems. In MU-MIMO
systems where there are many users in a cell, the optimal
scheduling (user selection) is computationally prohibitive. The
optimal strategy is a brute-force approach to find the best
subset of users exhaustively. Many suboptimal user selection
schemes have been proposed to reduce the computational
complexity of the optimal (exhaustive) user selection scheme
[9],[10],[12].
A suboptimal user selection scheme by using Gram-Schmidt
orthogonalization (GSO) in MU-MISO systems was proposed
in [9]. However, it may be difficult to use GSO in MUMIMO systems because the channel is a matrix while GSO
can only be used for vectors. For MU-MIMO systems,
two low complexity user scheduling algorithms, the capacity
based algorithm and the Frobenius norm based algorithm,
were proposed in [10]. These algorithms are based on a
greedy method, and achieve performance close to the optimal
scheduling algorithm. But their computational complexity is
relatively high. The capacity based algorithm needs frequent
computation of SVD, and the Frobenius norm based algorithm
also needs heavy GSO computation. Therefore, our motivation
is to further reduce the complexity of user selection in MUMIMO systems.
In this paper, we propose a new low complexity scheduling algorithm with the BD scheme to maximize the total
throughput. We introduce chordal distance as a new distance
metric. The scheduling algorithm select a new user who
has the maximum chordal distance with previously selected
users channels. It has lower computational complexity since it
requires less GSO computation than the Frobenius norm based
algorithm. The proposed scheduling algorithm is extended
to the case where the SVD-based receiver (Rx) processing
is used, which will be shown to improve the sum-capacity
significantly. The rest of this paper is organized as follows.
Section II and Section III introduce the system model and the
background, respectively. The proposed algorithm is described
in Section IV. Section V presents the computational complexity analysis, and Section VI presents the simulation results
for the proposed scheduling algorithm as well as the receiver

c 2012 IEEE
0090-6778/12$31.00 

650

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 3, MARCH 2012

processing (SVD) algorithm. Finally, conclusions are given in


Section VII.

to eliminate the interference term. The precoding matrix Vk


can be computed by SVD

II. S YSTEM M ODEL


We consider an MU-MIMO downlink system with a single
base station (BS) which has M transmit antennas, and KT
users with N receive antennas. We assume that the receivers
estimate their channels perfectly, and the BS knows the exact
channel state information (CSI) of all the users. The broadcast
channel of an MU-MIMO system is then given by
yk = Hk x + nk ,

k = 1, . . . , KT

(1)

where Hk (k = 1,. . . ,KT ) is the N M channel matrix at


the kth user whose channel entries are independent identically
distributed (i.i.d.) complex Gaussian with zero mean and unit
variance, nk is the complex white Gaussian noise vector for
the kth user whose elements are also i.i.d. complex Gaussian
with zero mean and unit variance, and yk is the received signal
vector at the kth user. The transmitted signal, x, is given by
x=

K 

Pi Vi si

(2)

i=1

where K is the number of the simultaneously selected users, si


is the ith symbol vector with E[si 2 ] = 1, Pi is the allocated
power for the ith selected user, and Vi is the M nk precoding
matrix for the ith selected user.

H(1)
..
.

(k1)
H

H(k+1)

..

.
H(K)

= U
(k)


U(k)

0
0

 (k)H 

(5)
H
R(k)

The columns of R(k) in (5) spans the null space of all H(j) s
(1 j K, j = k), and thus it becomes Vk .
From (3), we have
yk =

P k Hk Vk sk + nk .

(6)

An MU-MIMO system is decomposed into K independent


SU-MIMO systems
by BD. The dimension of Vk is M nk
K
where nk = M i=1,i=k N . The number of the transmit
antennas should be larger than the sum of the number of
receive antennas of any K 1 users for the existence of the
null space of all H(j) s (1 j K, j = k). The condition to
be satisfied is then given by
M > (K 1)N.

(7)

From (7), the maximum of K is  M


N  where a is the
minimum integer number not smaller than a.

III. BACKGROUND
A. Block Diagonalization
The received signal for the kth user denoted by (1) can be
divided into the intended signal term, the interference term,
and the noise by using (2), which is given by
yk =

P k Hk Vk sk + Hk

K


Pj Vj sj + nk .

(3)

j=1,j=k

To simplify the notation, we assume that the users are appropriately re-indexed after the user selection.
The first term of (3) is the desired signal for the kth user,
and the second term is the interference term. BD eliminates
the interference term by finding the null space of the channel
matrices of the other users. To be more precise, the precoding
matrix Vk is found by removing the interference term as in

H(1) Vk = 0
H(1)
..
..

.
(k1).

(k1)
H

Vk = 0

H
(4)

H(k+1) Vk = 0
H(k+1) Vk = 0

..
..

.
.
H(K) Vk = 0

H(K)

where H(k) is the channel matrix of the kth selected user. Vk


should be in the null space of each H(j) (1 j K, j = k)

B. Chordal Distance
Let us introduce a distance metric between subspaces.
Grassmannian space G(m, n) is the set of all the ndimensional subspaces of Euclidean m-dimensional space
[11]. We need to define the distance between two elements
of G(m, n). We associate principal angles i [0, /2], and
principal vectors ui P and vi Q for i = 1, . . . , n with two
n-planes (n-dimensional hyperplanes) P and Q as follows.
Choose u1 P and v1 Q having length 1, and such that
u1 v1 is maximal. Inductively, define ui P and vi Q
having length 1 and such that ui vi is maximal, subject to
the conditions ui uj = 0 and vi vj = 0 for all 1 j < i
[13]. Then set i = arccos(ui vi ). The chordal distance is
given by
dc (P, Q) =


sin2 1 + + sin2 n

(8)

A generator matrix for an n-plane P G(m, n) is an


m n matrix whose columns span P . The orthogonal group
O(m) acts on G(m, n) by right multiplication of a generator
matrix. Applying a suitable element of O(m) and choosing
appropriate basis vectors for the planes, we can reduce the
generator matrices of any given pair of n-planes P, Q with

KO and LEE: MULTIUSER MIMO USER SELECTION BASED ON CHORDAL DISTANCE

n m/2 [14] to the forms of

cos 1
0
0
cos
2


1 0 0
..
..
.
.
0 1 0


0
0
.. .. . .
..

. .
. .
0

sin 1
0 0 1
0
sin
2

,
0 0 0
..
..


.
.
0 0 0


0
0
. . .

.
. . ..
.. ..
0
0

0 0 0 .
..
..
.
0
0

651

B. Chordal Distance Based MU-MIMO Scheduling Algorithm

..
.

..
.

..
.

0
0
..
.

cos n

0
.
..
.

sin n

..
.

The sum capacity of an MU-MIMO system with the optimal


scheduling scheme can then be written as
Ropt (S) =

(9)

If A is a generator matrix for P whose columns are orthogonal


unit vectors, the projection is represented by the matrix
Pp = AAH . From (9), we can see that trace(Pp ) is n. For
P, Q G(m, n), with orthonormal generator matrices A and
B, and principal angles 1 , . . . , n , a simple calculation using
(9) shows that


d2c (P, Q) = n cos2 1 + + cos2 n
= n trace(AAH BB H )
1
= Pp Qp 2F
2

(10)

where Pp and Qp are the corresponding projection matrices.


Note that the chordal distance definition can be extended to a
case where P and Q have different dimensions. For example,
P G(m, n) can be an n-plane, and Q G(m, n) can be an
n -plane (n n).
IV. L OW C OMPLEXITY S CHEDULING A LGORITHM
A. Power Allocation
An MU-MIMO system can be divided into K independent
parallel SU-MIMO systems by BD. We can use water-filling
scheme since BS knows perfect channel information about
all receivers. When we assume each receiver gets N streams,
there are total N K streams, and we use water-filling with
these streams. The capacity with water-filling can be written
by [12]


n

P i
i
log2 1 +
(11)
C(H(S), P ) =
M
i=1
where P is the total transmit power, n is the number of total
streams, S is set of the selected users, H(S) is the combined
channel of S, and i (i = 1, . . . , n) is the positive singular
value
of the
of H(S). Note that i satisfies i =
+ith channel

n
M
and i=1 i = M . The sum rate of the MU P i
MIMO system with BD and water-filling among the selected
users is then given by

C (H(S)V(S), P )
(12)
R(S) =
S

where V(S) is the precoding matrix of S.

max

S{1, ,KT },1|S|K

R(S)

(13)

From (11) to (13), the optimal scheduling algorithm uses


brute-force exhaustive search over all possible user sets. The
number
 user
 subsets with the optimal scheduling algorithm
K of
is i=1 KiT . However, when KT is large, the search complexity of the optimal scheduling algorithm is too high.
We propose a new low complexity MU-MIMO scheduling
with chordal distance. Orthogonality among channel matrices
of the selected users in BD is a critical issue because of the
characteristics of BD. For example, the Frobenius norm of Hk
can be large but its orthogonality with the other selected users
can be small, which is not desirable. In (6), we should consider
not Hk , but Hk Vk since the effective channel is Hk Vk . Even
if the Frobenius norm of Hk is large, the Frobenius norm of
Hk Vk can be small since Vk eliminates not only the interuser interference, but also the common subspace of Hk and
the channel matrices of the other selected users. If users are
close to orthogonal, the common subspace removed by Vk
can be made small. We use the chordal distance to select the
users to maximize the channel matrix distance between the
previously selected users and the new user.
The chordal distance between plane A and B is given by
1
H
dc (A, B) = Ao AH
o Bo Bo  F
2

(14)

where Ao and Bo are orthonormal bases for subspaces A


and B, respectively. One of the characteristics of chordal
distance is that it can be measured between two different
dimensional objects. For example, it is possible to estimate the
distance between a plane and a vector in a 3-dimensional space
with chordal distance. The chordal distance between channel
matrices of the previously selected users and the channel
matrix of a candidate user can be estimated accordingly. The
summary of the proposed algorithm is described in Table I.
First, the user who has the maximum Frobenius norm of
Hk (k T ) is selected among all the users. Since the effective
channel is Hk Vk (k T ), we must consider not only the
chordal distance between Po and Gk,o , but also the Frobenius
norm of Hk . The Frobenius norm of the effective channel
is roughly proportional to the chordal distance between Po
and Gk,o , and the Frobenius norm of Hk . Therefore the next
selected user is the user who has the maximum value of
Ck dc (Po , Gk,o ), where is the weight factor between the
chordal distance and the Frobenius norm. When is large, the
Frobenius norm is weighted more than the chordal distance,
and vice versa. depends on N and K, as will be shown in
Section VI.
V. C OMPUTATIONAL C OMPLEXITY A NALYSIS
The complexity is counted as the number of flops, denoted
as . A real addition, a multiplication, and a division are
counted as one flop. Therefore a complex addition needs two
flops, and a complex multiplication needs six flops [10]. We
assume KT
K, N M, and K =  M
N . For an N M

652

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 3, MARCH 2012

complex valued matrix H, the number of typical matrix


operation [10] is summarized as follows.

K  
BS performs exhaustive search over all i=1 KiT possible
user sets in the optimal scheduling algorithm. Thus the order
of the optimal scheduling algorithm is given by [10]
 


KT
OP
K 48(K 1)2 + 8 N 2 M +
K
24(K 1)N M 2 + (54(K 1)3 + 2K 2 + 126)N 3 +

8KN

 
KT
K M3
O
K


1
K
O KT K K+ 2 M 3
(15)

where the Stirlings approximation to a factorial is used in the


last line.
B. Suboptimal Scheduling Algorithm
The computational complexity of the capacity based algorithm and the Frobenius norm based algorithm in [10] is
summarized here. The complexity order of the capacity based
algorithm is expressed by [10]
CA <

i=2

[48i(i 1)2 + 48i]N 2 M

+ [24i(i 1) + 32i]N M 2
+ (54i(i 1)3 + 54i)N 3

+ 2i2 N 2 + 8iN (KT i + 1)

+ KT (48N 2 M + 24N M 2 + 54N 3 + 2N 2 + 8N )


O(KT K 2 M 3 ).
(16)
The complexity order of the Frobenius norm based algorithm is given by [10]
F

K 

[8(i 1)3 + 18(i 1)2 + 18(i 1)]N 2 M
i=2

+ [2(i 1)2 + 4(i 1)]N M


(KT i + 1) + 4KT N M
O(KT K 2 M 3 ).

SNR 20 dB

20

SNR 10 dB

15

Optimal
Capacity based
Frobenius norm based
Chordal distance

10
SNR 0 dB
5

0
5

A. Optimal Scheduling

K 


25
Sum Capacity (bits/s/Hz)

 Frobenius norm H2F : 2M N real multiplication +


2M N real addition = 4M N flops.
 Gram-Schmidt orthogonalization (GSO) : 4M N 2 2M N
real multiplication + 4M N 2 2M N real addition +
2M N real divisions = 8M N 2 2M N flops.
 Water-filling over n eigenmodes : (1/2)(n2 + 3n) real
multiplication + (n2 +3n) real additions + (1/2)(n2 +3n)
real divisions = 2n2 + 6n flops.
 Singular value decomposition (SVD) 24M N 2 +
48M 2 N + 54M 3 flops.

30

10

15
20
25
30
The total number of users (Kt)

35

40

Fig. 1. Sum-capacity comparison of various scheduling algorithms when


M = 4, N = 2, = 0.4.

C. Chordal Distance based Scheduling Algorithm


In the chordal distance scheduling algorithm, the complexity
order is given by
CO

K 

[8(i 1)2 N 2 M 2(i 1)N M + 7(i 1)N M 2 ]
i=2

+ [8N 2 M 2N M + 7N M 2 + 3M 2 ]

(KT i + 1) + 4KT N M

O(KT M 3 ).
(18)
The following is the detailed explanation of (18).
i = 1 : the Frobenius norm of KT users needs
4KT M N flops.
H
i 2 : Po Po is calculated once for each i. Po needs
2 2
8(i 1) N M 2(i 1)N M flops by GSO. Therefore
Po PoH needs 8(i1)2 N 2 M 2(i1)N M +7(i1)N M 2
flops. Since the size of matrix Gk,o is independent of i,
2
2
Gk,o GH
k,o needs 8N M 2N M + 7N M flops. Finally,
H
the Frobenius norm of (A Gk,o Gk,o ) needs 6M 2 . The
proposed scheduling algorithm has lower complexity than
the optimal algorithm and the two suboptimal algorithms
in [10].
As shown in the above computational complexity analysis,
all low complexity scheduling algorithms have complexity
proportional to KT except for the optimal scheduling algorithm, which depends on KTK . The chordal distance based
scheduling algorithm has the lowest computational complexity
among other scheduling algorithms. algorithms. Note that it
does not depend on K any more in (18).
VI. S IMULATION R ESULTS

(17)

In this section, we compare the performance of the proposed scheduling algorithm with the optimal scheduling algorithm and two suboptimal algorithms in [10]. Fig. 1 and
Fig. 2 compare the sum capacities of various scheduling
algorithms with respect to the total number of users (KT )

KO and LEE: MULTIUSER MIMO USER SELECTION BASED ON CHORDAL DISTANCE

40

14

35

12

SNR 20 dB
25
SNR 10 dB

Optimal
Capacity based
Frobenius norm based
Chordal distance

Capacity based
Frobenius norm based
Chordal distance based

10
Run Time (ms)

30
Sum Capacity (bits/s/Hz)

653

20
15

8
6
4

10

SNR 0 dB
2

5
0
5

0
5

10

15
20
25
30
The total number of users (Kt)

35

40

Fig. 2. Sum-capacity comparison of various scheduling algorithms when


M = 6, N = 2, = 0.22.

30

Sum Capacity (bits/s/Hz)

25
20

Frobenius w/ Rx proc.
Chordal w/ Rx proc.

20 dB

15
10

10 dB

5
0 dB
0
15

20
25
30
35
The total number of users (Kt)

40

Fig. 3.
Sum-capacity comparison between the chordal distance based
algorithm with Rx processing and the Frobenius norm based algorithm with
Rx processing when M = 4, N = 2, = 0.1.

and SNR with fixed M, N , and K. Fig. 1 is for the case


of M = 4, N = 2, = 0.4, and Fig. 2 is for the case of
M = 6, N = 2, = 0.22. It is shown that the proposed
algorithm achieves the sum capacity similar to the optimal
scheduling algorithm. When M = 4, N = 2, the proposed
algorithm has slightly lower sum capacity compared to the
capacity based algorithm, but outperforms the Frobenius norm
based algorithm. When M = 6, N = 2, it has the sum capacity
close to the Frobenius norm based algorithm. The capacity
based algorithm performs slightly better than the proposed
algorithm and the Frobenius norm based algorithm.
Fig. 3 shows the sum capacity comparison between the
chordal distance based algorithm with receiver processing and
the Frobenius norm based algorithm with receiver processing
when M = 4, N = 2, = 0.1. In the receiver processing
scheme, SVD is used to decompose the channel of each user
into multiple eigenmodes. It is assumed that the Hermitian of
the left singular matrix of SVD is multiplied to the received
signal of each user. Its scheduling process is similar to

10

15
20
25
30
The total number of users (Kt)

35

40

Fig. 4. Average run time comparison for various scheduling algorithms when
M = 6, N = 2, = 0.22.

Table I except that a basic unit for scheduling is not a user


anymore, but an eigenmode (stream) decomposed by SVD.
In other words, Table I is used as an eigenmode (stream)
selection algorithm instead of a user selection algorithm. It
is observed that the chordal distance based algorithm with
receiver processing performs better than the Frobenius norm
based algorithm with receiver processing especially for large
number of users. When M = 4, N = 2, the maximum number
of streams in a scheduling scheme with Rx processing is 4,
whereas the maximum number of users in a scheduling scheme
without RX processing is M/N = 2. Since the Rx processing
scheme can serve up to 4 users, and the non-Rx processing
scheme can only serve up to 2 users, the former performs
better than the latter due to multiuser diversity, which can be
observed from Fig. 1 and Fig. 3. It can be easily shown that
the complexity order of the chordal distance based algorithm
with receiver processing (SVD) is also the same as (18).
Fig. 4 shows the average CPU run time of various scheduling algorithms with respect to the total number of users
with M = 6, N = 2, = 0.22. The three low complexity
scheduling algorithms have linear behavior with respect to
the total number of users as indicated by the complexity
equations. It is observed that the proposed algorithm has lower
complexity than the other two algorithms because the capacity
based algorithm computes SVD heavily, and the Frobenius
norm based algorithm needs more GSO computation than the
proposed algorithm.
Table II shows empirical values that maximizes the sum
capacity for non-receiver processing schemes. Note that
is dependent on the number of transmit antennas and the
number of receive antennas. As M gets larger, gets smaller
when N is 2. That is, when N is fixed, is inversely
proportional to K since K is M/N . This may be because
more common subspace is eliminated as K gets larger, and
orthogonality becomes more important. When K is 2, is
inversely proportional to N , which is also because more
common subspace is eliminated as N becomes larger. For
the simulations of receiver processing schemes in Fig. 3, we
use different (0.1) from Table II because the empirically

654

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 60, NO. 3, MARCH 2012

TABLE I
C HORDAL D ISTANCE BASED M ULTIUSER MIMO S CHEDULING A LGORITHM
Step I: Initialization
T = {1, 2, . . . , KT };
Gk =

HT
k;

Ck = Hk F ;

S = {s1 };

T = T s1 ;

s1 = max Ck ;
k

R = 0;

Po = ;

Step II: Loop


FOR i = 2 to K


Po = Po Gsi1 o ; A = Po PoH ;


si = max Ck A Gk,o GH
k,o F ,
kT

where Gk,o is orthonormal bases of Gk ;


Rtemp = R(Stemp ), where Stemp = S + {si };
IF (Rtemp < R)
break;
ELSE
S = Stemp ; R = Rtemp ; T = T {si };
END-IF
END-FOR

TABLE II
E MPIRICALLY C HOSEN A CCORDING TO M AND N

PP
PP N
M
P
P
4
6
12

0.4
0.22
0.07

0.28
0.1

0.15

optimized of receiver processing schemes is different from


that of non-receiver processing schemes.
VII. C ONCLUSION
In this paper, we proposed a new low complexity MUMIMO scheduling algorithms. One of the key ideas of this
paper is to use chordal distance as an orthogonality measure
in the proposed user selection algorithm. Simulation results
show that the proposed algorithm achieves the sum rate close
to the optimal (exhaustive) search algorithm. It also has the
sum rate similar to the existing low complexity scheduling
algorithms while its complexity is lower. In particular, it has an
advantage for a system where the number of selected users is
large. The proposed chordal distance based algorithm for user
selection appears to be promising for practical MU-MIMO
systems of next-generation wireless standards such as 3GPP
LTE or WiMAX (IEEE 802.16m).
R EFERENCES
[1] I. E. Telatar, Capacity of multi-antenna Gaussian channels, European
Trans. Telecommun., vol. 10, no. 6, pp. 585595, 1999.
[2] G. J. Foschini and M. J. Gans, On limits of wireless communication
a fading environment when using multiple antennas, Wireless Pers.
Commun., vol. 6, no. 3, pp. 311335, 1998.
[3] R. Knopp and P. Humblet, Information capacity and power control in
single cell multiuser communications, in Proc. 1995 IEEE ICC, pp.
331335.

[4] G. Caire and S. Shamai, On the achievable throughput of a multiantenna Gaussian broadcast channel, IEEE Trans. Inf. Theory, vol. 49,
pp. 16911706, July 2003.
[5] M. Costa, Writing on dirty paper, IEEE Trans. Inf. Theory, pp. 439
441, May 1983.
[6] J. Lee and N. Jindal, Dirty paper coding vs. linear precoding for MIMO
broadcast channels, in Proc. 2006 Asilomar Conf. Signal, Syst. Comp.
[7] L. U. Choi and R. D. Murch, A transmit preprocessing technique for
multiuser MIMO systems using a decomposition approach, IEEE Trans.
Wireless Commun., vol. 3, no. 1, pp. 2024, 2004.
[8] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, Zero-forcing methods for downlink spatial multiplexing in multiuser MIMO channels,
IEEE Trans. Signal Process., vol. 52, no. 2, pp. 461471, 2004.
[9] T. Yoo and A. J. Goldsmith, On the optimality of multiantenna
broadcast scheduling using zero-forcing beamforming, IEEE J. Sel.
Areas Commun., vol. 24, pp. 528541, Mar. 2006.
[10] Z. Shen, R. Chen, J. G. Andrews, J. R. W. Heath, and B. L. Evans, Low
complexity user selection algorithms for multiuser MIMO systems with
block diagonalization, IEEE Trans. Signal Process., vol. 54, no. 9, pp.
36583663, 2006.
[11] J. H. Conway, R. H. Hardin, and N. J. A. Sloane, Packing lines, plane,
etc.: packings in Grassmannian spaces, Exper. Math, vol. 5, no. 2, pp.
139159, 1996.
[12] X. Zhang and J. Lee, Low complexity multiuser MIMO scheduling
with channel decomposition using capacity upperbound, IEEE Trans.
Commun., vol. 56, no. 6, pp. 871876, June 2008.
[13] G. H. Golub and C. F. Van Loan, Matrix Compution, 2nd edition. Johns
Hopkins Univ. Press, 1989.
[14] Y. C. Wong, Differential geometry of Grassmannian manifolds, Proc.
Nat. Acad. Sci. pp. 589594, 1967.

Potrebbero piacerti anche