Maximum-SNR Transmit Antenna Selection with

Unimodular Beamforming and Two Receive

Antennas *
Maria Gkizeli

George N. Karystinos

Department of Electrical Engineering

State University of New York at Buffalo
Buffalo, NY 14260 USA
Email:mgkizeli@buffalo.edu

Department of Electronic and Computer Engineering

Technical University of Crete
Chania, 73100, Greece
Email: karystinos@telecom.tuc.gr

Ahstract-The recent increased interest in massive multiple

input multiple-output systems, combined with the cost of the
analog

RF

chains, necessitates

the

use of

efficient antenna

selection (AS) schemes. Capacity or SNR optimal AS has been

considered to require an exhaustive search among all possible
antenna subsets. In this work, we identify in polynomial time a
polynomial-size collection of antenna subsets that contains the
one that maximizes the post-processing receiver SNR when two
receive and an arbitrary number of selected transmit antennas
are employed.

I.

INTRODUCTION

The cost and complexity of the analog RF chains connected

to the inexpensive antenna elements at both sides of multiple
input multiple-output (MIMO) systems [1], [2] is a limiting
factor on the number of the antennas that may operate in
practice. A low-cost engineering technique that reduces the
number of analog chains required is antenna selection (AS)
where a number of limited transmit/receive RF chains are
multiplexed between a selected set of transmit/receive anten
nas. AS for MIMO has been studied extensively in literature
and research works have dealt with either transmit AS (TAS),
receive AS (RAS), or joint transmit-receive AS [3]. However,
since there has recently been an increased interest in large
scale multiple-antenna wireless systems [4], [5], often called
massive MIMO [6], the need for AS algorithms that efficiently
select the antennas is of major importance.
In the past, AS has been considered to maximize ei
ther channel capacity [7]-[12], mllllmum post-processing
SNR [13], statistical quantities such as average through
put [14], or effective SNR [1], [2], [15]. All these works
present suboptimal algorithms, since the optimal solution is
considered to require the evaluation of the metric of interest
over all () possible combinations of AS sets, where N and
K are the numbers of available and selected, respectively,
antennas. Due to the exponential complexity of the exhaustive
search maximum-predetection-SNR AS, a suboptimal hybrid
*This work was supported by the Seventh Framework Program of the
European Commission under Grant PIOF-GA-2011-302963-TRADENET.

algorithm that performs RAS to maximize the receiver prede

tection SNR with complexity O(N) was considered in [16]
where the authors design the beamforming vector at the
transmitter side as the principal right singular vector of the
full channel matrix.
In this paper, we consider the maximum-SNR joint uni
modular beamforming-TAS problem for two receive antennas
and an arbitrary number K of selected transmit antennas and
develop an algorithm that identifies with complexity 0 (N4 )
a polynomial-size set of antenna selection subsets (i.e., can
didate subsets) that contains the optimal one. This result is
possible after introducing an auxiliary vector that partitions
the antenna selection problem into multiple subproblems, each
with complexity O(N). The maximum number (upper bound)
of subproblems (hence, candidate solutions) that we obtain is
The overall complexity of this algorithm is therefore
6
o N4 ) . These antenna subset candidates are then fed to
a suboptimal quadratic programming algorithm that approx
imates the optimal corresponding unimodular beamforming
vector. Our final selection of the antenna subset-beamforming
vector pair is then determined by a polynomial-time exhaustive
search among those candidate selection pairs.

( ).

II.

SIGNAL MODEL AND PROBLEM STATEMENT

We consider a MIMO system that consists of N transmit

and l'I/I receive antennas. The channel between any trans
mit and receive antennas is assumed flat fading and the
M x N complex baseband channel matrix is denoted by
H [hI h2 ... hM]H where h is a complex row vector
that contains the channel coefficients between the N transmit
antennas and the mth receive antenna, m 1, 2, . . . ,M.
The transmitter selects K (out of the N) antennas and
employs unimodular beamforming to transmit symbol xE <C.
We can equivalently say that it uses a beamforming vector
wE <c N subject to the constraint Il wllo K, Iwnl 1 or 0,
n
1, 2, . . . , N.
The downconverted and pulse-matched filtered received
vector of size M x 1 is y H wx + n where n E <c M
is a zero-mean additive white complex noise vector. Since y
=

978-1-4673-5239-0/13/$31.00 2013 IEEE

represents an unknown vector signal in white vector noise,

the maximum-SNR filter is the matched filter f Hw whose
output is given by fH Y IIHw l1 2 X + wHHH n. Then, the
filter output SNR is
=

III.

POLYNOMIAL-COMPLEXITY OPTIMAL TRANSMIT

ANTENNA SELECTION

A. M

1 receive antenna

We consider the trivial case of transmit antenna selection

with one receive antenna (M
1). We do so to identify
some
interesting
properties
that
will
be useful in the design of
(1)
the algorithm for optimal transmit antenna selection with two
receive antennas (1\1 2) in the next subsection.
Since M 1, the channel matrix becomes H hH and,
Eq. (1) shows how the predetection SNR is related to the
for
a fixed selection set I, we have H:,I h. Then, the
beamforming vector.
inner
maximization term in (2) becomes
Our objective is to jointly select K antennas and optimize
the unimodular beamforming vector w to maximize the prede
max
max I hw l
(3)
hI(k)W(k)
tection SNR in (1). That is, we seek w that solves the problem
=

wEflK

max

WE(flU{O})N
Il wll = K

IIHw l1

max

Ie[N]
I II =K

max

wEflK

IIH:,IWII ,

(2)

However, 12:[=1 hI(k)w(k)1

Il h Illl, with equality if

Itk=1

<

2:[=l lhI(k)w(k)1

w ejarg(hr).

where n { lsi E C lsi I} and [N] {1,2, ..., N} .

In (2), the optimization problem has been rewritten as two
nested problems; the outer one is the antenna selection prob
lem where set I contains the indices of the K selected
antennas and the inner one is the beamforming problem for the
particular antenna selection where w E nK is the "pruned"
beamforming vector that consists of the K nonzero loadings
of w E (n U {O} ) N. For a fixed antenna selection I, the
problem of identifying the optimal unimodular beamforming
vector -which alone is a unimodular quadratic programming
(UQP) problem- has not been efficiently solved so far. Subop
timal approaches to tackle this problem include semidefinite
relaxation (SDR) [17].
If a solver existed, then a straightforward approach to
solve (2) and identify the optimal set I would be an exhaustive
search among all () cardinality-K subsets of [N]. However,
if the number of selected transmit antennas K is a linear
function of the total number of transmit antennas N, then
such a solver would be impractical even for moderate values
of N, since its complexity grows exponentially with N. Even
in the case where K is not a function of N, if N is large
enough (e.g. N 100 as in massive MIMO [5]) and K is,
for example, equal to or greater than 5, then the complexity
of the exhaustive search would still be too large to consider
for practical implementation. In the next section, we present
an efficient algorithm that, for any fixed number of receive
antennas M
1 or M 2 and any unrestricted number
of selected transmit antennas K (that is, even if K grows
linearly with N), identifies in time that is polynomial in N
a polynomial-size collection of antenna selection subsets I
that includes the optimum subset in (2). For each antenna
subset candidate, we calculate the corresponding beamforming
vector via a UQP approximation algorithm. The selected pair
of candidate antenna subset and corresponding beamforming
vector is then determined by a polynomial-time exhaustive
search among those antenna candidate-beamformer pairs.

wEflK

(4)

Using the optimal beamforming vector in (4), eq. (2) becomes

max
(5)
Ihil
Ih
Ie[N] l Illl Ie[N] L
.
'EI
I
I
I I =K
I I =K
The optimal set I in (5) is the one that consists of the indices
of the K largest elements of Ihl.
To describe the latter step, we introduce function select
which selects the k largest (in magnitude) elements of a vector,
as follows.
max

select(u; k)

(6)
Il u Illl
Ie[N]
I I I =k
That is, select(u; k) computes Ihll, Ih 21, ..., Ih NI and returns
the indices of the largest k values. It turns out that the outcome
of select(u; k) is a set I such that IhiI 2: IhjI for any i E I
and j E [N] - I. The computational cost of select(u; k) is
=

argmax

'

O(N) [18].

We conclude that the maximum-SNR transmit antenna

selection with one receive antenna is provided by I
select(h; K) whose complexity is linear in the number of
available transmit antennas N. Optimal beamforming is then
calculated by (4). In the developments that follow, select is
critical in identifying in polynomial time a polynomial-size
collection of antenna subsets that includes the optimal one
when M 2.
=

B.

2 receive antennas

When two receive antennas are utilized (M

2), the
channel matrix is H [hI h2]H, and the problem of selecting
the optimal subset of K transmit antennas in (2) becomes more
challenging. In this section, we show that a polynomial-size
set of antenna selection subsets that contains the optimal one
can be identified with complexity 0 (N4 ) for any number of
selected transmit antennas K.
=

IH ::nc(,B) I
First, we redefine the problem space by introducing the Notice that, for any n E [N], Iun(,B)I
auxiliary angles E [0,] and B E ( -7r,7r] and defining IHi,n sin + H2',nejO cos I, i.e., every element of lu(,B)I
the unit-norm 2 x 1 vector
is a continuous function, or a surface, of (,B). When, due
to (14), we select the K largest elements of Iu(,B)I at a
) .
(7) given point (,B) as function select requires, we actually
c(,)B eJosin(
cos()
compare the surfaces IU1 (,B)I, IU2(,B)I, ..., IUN(,B)I at
In the following, we will see that and B help us identify point (,B). The optimal selection I c [N] in (11), is met
a polynomial number of locally optimal candidate selection if we scan the entire space 1> and collect the locally optimal
sets I. The optimal solution of (2) will be among the locally selection I(,B) for any point (,B) E 1>.
A natural question that arises is the following. How many
optimal ones.
Due to the unity of the norm of c(), from Cauchy-Schwarz selection subproblems are induced if we scan all values of
(, B) E 1>? An answer to the previous question identifies
Inequality, we obtain, for any vector a E ([;2,
exactly the number of locally optimal solutions for (2).
(8)
laHc(,)B 1 Ilallllc(,B)11 Ilall
The auxiliary angles ,B now become relevant in answering
the
above question. Due to the continuity of the surfaces
The equality above is achieved if and only if c() is collinear
(,
B)I, we expect that in an area around (,B) the selection
IUn
with a within a phase rotation such that its first element is
subset
I(
,B) will be retained because either the sorting of the
real positive, i.e., if and only if
surfaces does not change (although the surfaces vary) or the
(9) sorting of the surfaces changes but the group of the K surfaces
c(,B) : e-jarg(al).
11 ll
with the higher value is retained (that is, the change of the
sorting
occurs either within the higher K surfaces or within
We note that, for any a E ([;2, there always exists a pair
the lower N - K surfaces). Hence, we expect the formation of
of angles (,B) E 1>, where 1> [O,J X ( -7r,7r] , such regions in 1> within which the locally optimal selection subset
that (9) is satisfied. Therefore, from (8), we obtain Iiall
I is unique. In the sequel, we determine all these regions, show
max(,)O EiP laHc(,B)1 for any a E ([;2.
that their number is less than or equal to 6 (), and present a
If we substitute a with H:,IW in (2), then our optimization polynomial-time algorithm that identifies the selection subsets
problem becomes
I that are associated with these intervals.
We begin by noting that, as we scan the space 1>, the
max max IIH:,IWII
(10)
IC[N] wEOK
selection subset I does not change unless two surfaces inter
III=K
sect (which implies that the sorting of the surfaces changes).
C(,B) I .
max max max IwHHH
I
Therefore,
to identify all regions that retain their selection
,
.
IC[N] wEOK (,O)EiP
subset I, it suffices to examine when two surfaces intersect.
III=K
We note that this is a necessary, but not sufficient, condition for
6
We define u(,B) HHc(,B) and change the order of the a change of I, since the intersecting surfaces may correspond
maximizations in (10) to obtain
to elements of Iu(,B)I that, before they intersect, both belong
to I or neither belongs to I. In the latter case, although the
(11) magnitude sorting of the surfaces changes, the selection subset
max max max IWHUI(,B) I
(,O)EiP IC[N] wEOK
I does not.
III=K
Two surfaces, say Iun(,B)I and Ium(,B)I, intersect when
For any given pair of angles (,B) E 1> and any given Iun(,B)I Ium(,B)I. We note that, for any n, m E [N], the
selection subset I c [N], the inner maximization in (11) is intersection of IUn(,B)I and IUm(,B)I always exists, since
achieved, as in (3) and (4), by the pruned beamforming vector all surfaces meet 0 for some ,B.I As a result, the intersection
a curve on the (,B)-plane which
(12) of two surfaces determines
w(,B; I) ejarg(ur(,e)),
6
we define as Ln,m {(,B) E 1>: Iun(,B)I Ium(,B)I}.
resulting in the value maxwEoK IwHUI(,B)1
To illustrate this, in Fig. 1, we set N to 4, consider an arbitrary
Ilu I(,B)11 1 Then, the optimization problem in (11) 2 x 4 channel matrix H, and plot curves Ln,m, for any n, m E
becomes
[N] with n i= m.
We observe the regions that are formed; within each region,
max max Ilu I(,B)11 1
(13)
(,O)EiP IC[N]
the
selection subset I remains the same. We also observe that,
III=K
in most of the cases, each region "touches" an intersection of
where, for any given (,B) E 1>, the inner maximization is two or three curves.2 That is, we can identify these regions
achieved by the subset that consists of the indices of the K by examining intersections between curves Ln,m. In addition,
largest elements of lu(,B)I, i.e.,
i
=

I(,B)

argmaxllu I(,B)11 1
IC[N]
III=K

select(u(,B); K) . (14)

I,

For any surface IUn (<,1>,8)1

H!;n c(<,I>,8) we can always find a vector
is orthogonal to H:,n.
2The case where a region does not touch any intersection is examined in
the end of this subsection.

c(<,I>,8) which

n/2 ,-------,-----_,
'"

- ----- ...
, ... - -

nl4

"
"
,

'

,
,
,

""'... ....

"

'

/'0

' ,,/

,
,

,. ,,/
"", ...
....
,C ,
- ......
- ...
..........
. .
-: : : = .. ,
O L----------
....... ,'!..
-nl2
o
nl2
-n
e

::

Fig. 1. An illustration of intersection curves n,m, for n, mE {l, 2, 3, 4}

with n i= m, and formed regions, resulting from N
4 surfaces.
=

we can concentrate only on intersections of three curves and

ignore intersections of two curves, since the latter change
the sorting of surfaces but do not "generate" a new subset
I that has not been generated by a neighboring three-curve
intersection.
To examine all three-curve intersections, we observe that
each such intersection corresponds to a three-surface in
tersection (otherwise, more curves would pass through the
intersection point, something that happens w.p.O). Hence, it
suffices to find when three surfaces, say Iun(,e) l, Ium(,e) l,
and IUl(,e) l, where n, m,l E [N] with n i- m, n i- l,
and m i- l, intersect. That is, we have to find (,e) that
satisfies IHn c(,e)1
IHmc(,e)1
HC(,e) or,
equivalently,
=

(15)

A,

for some /-l E JR. A solution to (15), i.e., an intersection

between the three surfaces, exists if and only if there exist
ej)'HH /-l E JR that make the matrix A
sin-

A,

HH ]

ejwFi.:Ji
.,n Fi.:H'
.,l
gular. We define the 2 x 1 vector d [H:,m H:,d-1 H:,n and
=

For any combination of three surfaces, the above procedure,

with complexity 0(1), examines if they intersect and, if
so, also computes the two intersection points (,e). Each
intersection point is a vertex of six regions. Then, function
I
select(u(, e); K) computes, with complexity O(N),
the indices of the K largest surfaces. We recall that, at the
intersection point, all three surfaces that intersect have the
same value. If I contains the indices n, m,l of the three
surfaces that intersect or it does not contain either of them,
then I is the common optimal selection subset for all six
neighboring regions. Otherwise, we have to consider three
different cases for the index among n, m, and l that belongs or
does not belong to I. Overall, for each combination of three
surfaces, we obtain at most 6 candidate sets I with complexity
O(N). Since the total number of surfaces is N, the overall
number of candidates is upper bounded by 6 G).
Finally, regarding the case where a region does not touch
any intersection, we can see that such a region touches a
curve n,m which does not intersect with any other curve. For
example, in Fig. 1, such a curve is ,2. Therefore, when we
1
examine the three-surface intersections, as described before,
we should mark the corresponding curves and, in the end, if
a curve, say n,m, is unmarked, then we should examine it
separately as follows. We just need to pick a point on n,m
by solving IHn c(,e)1
IHm c(,e)1 or, equivalently,
(ej.\Hn - Hm) c ( ,e) 0 for some E JR. Actually, we
can set to zero, obtain c ( ,e) as the unit-norm vector in the
null space of H H
" n - H Hm , and, from (7), uniquely determine
the intersection point (,, e). We note that the existence of
such unmarked curves does not increase the complexity of our
algorithm since, at the same time, we avoid the computation of
many candidates that would have been generated if the curve
had participated in any three-curve intersection(s).
In the end, we keep at most 6()
0 (N3 ) candidate
selection subsets I. This way, with complexity 0 (N4 ) , inde
pendently of K, we provide a smaller collection of possible
antenna subsets (in which resides the optimal one) to any UQP
algorithm for optimal beamforming approximation. Then, we
compare all AS-beamforming pairs against the metric of in
terest in (2). In the next section, we describe the beamforming
approximation algorithm.
=

I
I

'.24
,
,
,

\
\
,
- ... ... ,
....
..
.
,
... ,,-

,
,
-B-

..

Id.

the scalar D
Then, after algebraic computations, we
can prove that IDI ::; 1 is a necessary and sufficient condition
to make A singular. If the condition is not satisfied, then
no intersection exists between the three surfaces (and three
corresponding curves). If the condition is satisfied, then we set
'IjJ -angle(d1d2) cos-1 D and w angle ([1 e-N] d)
and can prove, after algebraic computations, that A becomes
singular when
w and /-l
'IjJ + w. Then, c ( ,e) is the
unit-norm vector in the null space of A. Finally, from (7),
we can uniquely determine the intersection point (,e). Since
there are two values of 'IjJ that make A singular, we obtain
two intersection points.

IV.

BE AMFORMING ApPROXIMATION ALGORITHM

We seek W that solves (2) for a fixed antenna selection I.

Because no efficient solver for (2) is available, we consider
its relaxed version
max

Wl,W2Efl K

wfHIH:,IW2

(16)

and observe that, for a fixed beamformer W2, the vector WI

ejarg(HIHHIW2),
that maximizes (16) is given by WI
while, for fixed WI, the optimal W2 becomes W2
ejarg(HIHHIW1). This way, we perform a cyclic maximization
that is similar to [19]. Starting from an initial beamforming
vector wO
principal right singular vector of (HIH:,I)'
=

10-2"",------.------,----,-------r--,------,--r-----.------,

O
10' ,------,---,--r--.--,

, , \ , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
,

'.
'.

"","
'.

'.

"",

,."

"
,

10

",

"

'.

.,,'

.,
"
,.
--------

cr:
w
CD

- - - --- - -

--

-' -

10-

'

" " " , Random (M=2)

,-,-

, , , , , " Exhaustive search

Optimal (M=1)

- - - Proposed (bound)

-- Proposed (M=2)
10

20

30

40
50
60
70
Total Number of Antennas N

80

100

90

Fig. 2. Bit error rate versus total number of transmit antennas N for M
receive antennas and selection of K
6 transmit antennas,

-- Proposed (actual)
_-L___l__-L_========

10

20

30
40
50
60
70
Total Number of Antennas N

80

90

Fig. 3. Complexity versus total number of transmit antennas N for M

receive antennas and selection of K
6 transmit antennas,

100

we perform cyclic iterations of the form

REFERENCES

(17)
and continue the iterative algorithm for a predefined number
of steps or until Ilw (t+ 1) - w (t)II :s; E. In the simulations that
follow, we set E 0.01 and the maximum number of possible
iterations to 100.
=

V.

10 0

SIMULATION RESULTS

To illustrate our developments we consider a MIMO system

with N transmit and 2 receive antennas and perform K
6 TAS. The presented results are averages over 1000 i.i.d.
Rayleigh channel realizations.
In Fig. 2, we set the total transmit SNR to -3dB and plot
the bit error rate (BER) as a function of the available transmit
antennas N when K
6 transmit antennas are selected
according to the proposed algorithm with complexity 0 (N4).
We compare against random antenna selection and we also
include the BER of the optimal TAS when 1 transmit antenna
is occupied, as a reference; the latter has complexity 0 (N).
In Fig. 3, we plot the corresponding complexity (in terms
of number of candidate AS sets that are examined) of our
proposed algorithm, together with the corresponding upper
bound 6 (), and the exhaustive-search complexity. Our al
gorithm offers significant computation gains in comparison
with the exhaustive-search solution, without losing optimality.
These gains are offered in conjunction with performance
improvement, as indicated in Fig. 2.
=

ACKNOWLEDGMENT

The authors would like to acknowledge Prof. Aggelos

Bletsas for suggesting the idea of developing a polynomial
time algorithm for the MIMO antenna selection problem.

[1] A. F. Molish and M. Z. Win, "MIMO systems with antenna selection,"

IEEE Microwave Magaz., pp, 46-56, Mar. 2004,
[2] S, Sanayei and A, Nosratinia, "Antenna selection in MIMO systems,"
IEEE Commun. Magaz., pp, 68-73, Oct. 2004.
[3] D. A. Gore and A, J, Paulraj, "MIMO antenna subset selection with
space-time coding," IEEE Trans. Signal. Proc., vol. 50, pp. 2580-2588,
Oct. 2002.
[4] F. Rusek, D. Persson, B. K. Lau, E. G, Larsson, T. L. Marzetla,
O. Edfors, and F. Tufvesson, "Scaling up MIMO: Opportunities and
challenges with very large arrays," to appear in IEEE Signal Process.
Mag. Available online: http://arxiv.org/abs/1201.3210
[5] 1. Hoydis, C. Hoek, T. Wild, and S. ten Brink, "Channel measurements
for large antenna arrays," in Proc. ISWCS 2012, pp, 811-815, 2012,
[6] 1. Hoydis, S, ten Brink, and M. Debbah, "Massive MIMO: How
many antennas do we need?," in Proc. IEEE Allerton Conference on
Communication, Control, and Computing, Monticello, IL, Sept. 2011,
pp.545-550,
[7] A. Gorokhov, D. A, Gore, and A, 1. PauJraj, "Receive antenna selection
for MIMO spatial multiplexing: Theory and algorithms," IEEE Trans.
Signal. Proc., vol. 51, pp. 2796-2807, Nov, 2003,
[8] A. Gorokhov, D. A, Gore, and A, J. Paulraj, "Receive antenna selection
for MIMO flat-fading channels: Theory and algorithms," TEEE Trans.
In! Theory, vol. 49, pp. 2687-2696, Oct. 2003.
[9] YO-S, Choi, A, F. Molish, M. Z. Win, and 1. H. Winters, "Fast algorithms
for antenna selection in MIMO systems," in Proc. VTC 2003 Fall, vol.
3, pp, 1733-1737, Oct. 2003,
[10] M. Ghavari-Alkansari and A. B. Gershman, "Fast antenna subset selec
tion in MIMO systems," IEEE Trans. Signal. Proc., vol. 52, pp. 339-347,
Feb. 2004,
[II] A. Dua, K. Medepalli, and A. J, Paulraj, "Receive antenna selection
in MIMO systems using convex optimization," IEEE Trans. Wireless
Commun., vol. 5, pp, 2353-2357, Sept. 2006,
[12] L. Zhou and M. Shimizu, "Fast recursive algorithm for efficient transmit
antenna selection in spatial multiplexing systems," in Proc. VTC 2009Fall, pp. 1-5, Sept. 2009.
[13] R. W. Heath, S, Sandhu, and A. J, Paulraj, "Antenna selection for spatial
multiplexing systems with linear receivers," TEEE Commun. Letters, vol.
5, pp, 142-144, Apr. 2001.
[14] D. G, Gore, R. W. Heath, and A, J. Paulraj, "Transmit selection in spatial
multiplexing systems," TEEE Commun. Letters, vol. 6, pp, 491-493, Nov,
2002.
[15] A. F. Molish, M. Z, Win, and J, H, Winters, "Reduced-complexity
transmit/receive-diversity systems," IEEE Trans. Signal. Proc., vol. 51,
pp. 2729-2738, Nov, 2003,

[16] P. Sudarshan, N. B. Mehta, A. F. Molish, and 1. Zhang, "Channel

statistics-based RF pre-processing with antenna selection," IEEE Trans.
Wireless Commun., vol. 5, pp. 3501-3511, Dec. 2006.
[17] Z.-Q. Luo, W.-K. Ma, A. M.-C. So, Y. Ye and S. Zhang, "Semidefinite
relaxation of quadratic optimization problems," IEEE Sign. Proc. Mag.,
vol. 27, pp. 20-34, May 2010.
[18] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction
to Algorithms, 2nd. ed. MIT Press 2001, Part II, Chapter 9.
[19] P. Stoica, H. He, and 1. Li, "On designing sequences with impulse-like
periodic correlation, IEEE Sign. Proc. Letters, vol. 16, pp. 703-706, Aug.
2009.