You are on page 1of 6

Semi Definite Relaxation (SDR) in Multiple Input

Multiple Output (MIMO) Detection


Paban Sarma

Ripan Deuri

Roll No. 14104104

Roll No. 14104135

AbstractThe paper consists of studies on application of semi


definite programming in MIMO detection. Several ways to tackle
the NP-Hard ML MIMO detection via Convex SDR relaxation
is studied and implemented. At first the focus is on MIMO
systems with 16-QAM constellations followed generalization of
the methods to higher order QAM of the form 4q . Performance of
these techniques are verified by simulation and the performance
metric observed is symbol error rate w.r.t SNR per bit. Also the
equivalence among the techniques are studied and they verified
by simulation. In this case metric observed is optimal objective
value produced by the methods with system size and SNR.

I. I NTRODUCTION
Multiple Input Multiple Output (MIMO) systems are very
common in modern day communication system. For instance
multiple user communication and multiple antenna channels
can be named. The uses of multiple transmit and receive antenna enables higher system performance provided transmitted
symbols are efficiently detected. This fact has made the MIMO
detection problem an important topic among the scholars of
communication systems.
The Maximum Likelihood (ML) detector is the most efficient detector for a MIMO system however it is much costly
in terms of computational effort. ML problem is known to be
a NP-Hard due to its exponential complexity with system size.
In the series of low complexity sub-optimal detector the Zero
Forcing (ZF) detector that relaxes the discrete nature of constellation point is computationally efficient but performance is
compromised. ZF with decision feedback enhances the performance, but it is still lower than the sphere decoders. Sphere
decoders are the best performance sub-optimal detectors, but
it is limited to smaller system size due to its complexity and
compromised performance in low SNR condition.[1]
In recent times semidefinite relaxation (SDR) technique has
been widely used for sub-optimal MIMO detection. And the
motivation that leads to think about SDR is that quadratic
optimization problem can be converted to SDP by some
suitable transformation or introduction of dummy variable[2].
At the very beginning the SDR is successfully applied to
QPSK constellation which motivates the researchers to apply
the same for higher order QAM constellations. In this paper we
have discussed and demonstrated different MIMO detection
methods for QAM constellations based on SDR.

II. MIMO S YSTEM M ODEL


A multiple input multiple output system with Nt transmit
antenna and Nr receive antenna can be expressed as
yc = Hc sc + wc

(1)

Where sc is an Nt 1 transmitted symbol vector and each


element is drawn from a set Sc that contains the set of all
possible constellation points, Hc is a complex channel matrix
of size Mr Nt , wc is additive white Gaussian noise vector
of dimension Nr 1, and yc is the complex received vector
of dimension Nr 1. An equivalent model with real entities
can be described as
  
   
yR
HR HI sR
wR
=
+
yI
HI
HR sI
wI
(2)
y = Hs + w
where (.)R and (.)I represents real and imaginary part of
the complex entity (.). Also, the dimensions of y, s, w and
H are 2Nr 1, 2Nt 1, 2Nr 1 and 2Nr 2Nt respectively.
In model (2) the elements of s belong to a real set S =
{1, 3 . . . , (2q 1)} for 4q -QAM constellations.
Therefore, the ML detection problem is formulated as
follows
2
min ||y Hs||2
(3)
s.t
si S i = 1, 2, . . . 2Nt
III. SDR DETECTORS FOR 16-QAM MIMO
SDR techniques for MIMO system using 16QAM constellations are based on convex relaxation of the constraint
si {1, 3} in MIMO detection problem (3).
A. Polynomially inspired (PI) SDR
For 16-QAM constellations Si {1, 3} is equivalent to
writing
(si 1)(si + 1)(si 3)(si + 3) = 0
= s4i 10s2i + 9 = 0
= s2i ui = 0 ; u2i 9ui + 10 = 0;

(4)

i = 1, 2, . . . , 2Nt
With this constraint the MIMO detection problem in (3) is
equivalent to solving the minimization problem
min ky Hsk22
s,u

s.t.

s2i ui = 0;

i = 1, 2, . . . , 2Nt

u2i

i = 1, 2, . . . , 2Nt

9ui + 10 = 0;

(5)

Redefining x of dimension 4Nt +1 as xT = [xT uT 1]T and


X = xT x, the decomposition of X is as follows

xxT xuT x
X1,1 X1,2 X1,3

(6)
X = X2,1 X2,2 X2,3 = uxT uuT u

xT
uT 1
X3,1 X3,2 X3,3
X is a PSD matrix with rank 1; since the rank constraint is
non convex, the convex SDR version of the ML problem in
(5) is formulated eliminating the rank constant as

X0
(7)

X4Nt +1,4Nt +1 = 1
diag(X1,1 ) X2,3 = 0

diag(X2,2 ) 10X2,3 + 9X2,3 = 0


T

H H 0 HT y
0
0
where, Q = 0
yT H 0 yT y
It can be proved that the Lagrangian bi-dual formulation of
problem (5) leads to the SDP (7) above [3].
B. Bound constrained (BC) SDR
BC-SDR [4] exploits the fact of QAM constellations being
symmetric around the origin and existing of bounds l and u
such that l s2i u for i = 1, 2, . . . , 2Nt . In 16-QAM
constellation l and u are 1 and 9 respectively.
Introducing a new variable t {1},the objective function
of the ML detection problem in (3) can be simplified as
ky Hsk22 = sT HT Hs 2yT Hs + yT y
= sT HT Hs 2yT Hts + t2 yT y

(8)

= x Qx
Where,
 
 T
s
H H
; Q=
t
yT H

HT y
yT y


(9)

Now, xT Qx = tr(xT QT x) = tr(QxxT ) = tr(QX) with


X = xxT . X is a PSD matrix with rank 1. Again relaxing the
non-convex rank constraint the BC-SDR can be formulated as
min tr(QX)
X

s.t.

X0
1 Xi,i 9;

s.t.

bi {1};

(12)
i = 1, 2, . . . , 4Nt

= HW and introducing x = [bT 1]T the VA


Now, using H
SDR can be formulated in a similar approach like (10) as,
min tr(QX)
X

Xi,i = 1, i = 1, 2, . . . , 4Nt + 1

(13)

X0

x=

min ky HWbk22

s.t.

min tr(QX)
s.t.

Therefore, defining b = [bT1 bT2 ]T and W = [I 2I], where I


is and identity matrix of dimension 2Nt 2Nt . therefore the
problem (3) is equivalent to

(10)
i = 1, 2, . . . , 2Nt

X2Nt +1,2Nt +1 = 1;
The detected signal constellations points belonging to S can
be obtained from optimal X by suitable technique.
C. Virtually antipodal (VA) SDR
An element s {S} can always be represented as a
weighted sum of antipodal element b {1}. For 16-QAM
the VA representation of s is given by
s = b1 + 2b2

(11)

Q here has a similar definition like (7) with H replaced by

H.Also
X = xxT is a rank one PSD matrix and its rank
constraint is relaxed to form the convex VA-SDR problem
above.
D. Recovering constellation points from SDR solution
is the solution in (7),
1) Eigen Value Decomposition: Let X
(10) or (13) of dimension n n.In (7) and (10) n is 2Nt + 1
and in (13) it is 4Nt + 1. The EVD based methods takes the
If v is the eigen
solution as the principle eigen vector of X.
vector corresponding to the largest eigenvalue in eigenvalue
= VVT then in case of PI and BC SDR
decomposition X
optimum s is found as
si = quant(vi /vn ); i = 1, 2, . . . , 2Nt

(14)

Where, quant(.) denotes the closest approximation of the


argument to a point belonging to S = {1, 3}.
In case of VA-SDR with sign(.) denoting the sign of the
argument, the optimal b and s predicted as
bi = sign(vi /vn ); i = 1, 2, . . . , 4Nt
(15)

s = [I 2I]b
The solution from this method is the best solution if optimal
is a rank one matrix i.e., it has only one distinct nonzero
X
eigenvalue.
2) Randomization: In this approach a correlated random
for i =
Gaussian vector vi generated such that vi N (0, X)
1, 2, . . . r. The feasible vi producing the least objective value
in corresponding problem is considered as solution v. larger
the number of randomization steps r, more is the accuracy in
this approach.
After v is chosen with randomization procedure, the next
step leading to selection of optimal
s is same as (14) and (15).
3) Simple rounding: It can be noted that the minimization
variable X in the problems is of the form
 
 T


x 
x x x
x 1 =
X=
1
xT
1
Where, x is s in case of BC/PI-SDR and b in case of VA nn we can retrieve
SDR. Therefore from optimal solution X
vn1 comprising of the first n1 elements of the nth column
as
of X
v = X(1 : n 1, n)
(16)
With v in hand the next step is same as (14) and (15).

E. Equivalence in the Methods


The three SDR techniques discussed above are equivalent
to each other and the optimal objective value produced is
same for all as the objective function of each of them can
be expressed as a single unified expression as follows,
f (S, s) = tr(HT HS) 2yT Hs + yT y
S = ssT ; si {1, 3} i

(17)

Also the feasible sets of each of the three methods are


identical[4], i.e
FP I = FBC = FV A
(18)
Where, FP I ,FBC , and FV A represents feasible sets for PISDR, BC-SDR and VA-SDR respectively.
IV. E XTENSION TO HIGHER ORDER CONSTELLATION
A. PI SDR for 4q QAM
Considering q = 3 i.e. for 64-QAM the constraints on
MIMO ML problem can be simplified as s2i R; i =
1, 2, . . . 2Nt where, R = {r1 , r2 , r3 , r4 } = {1, 9, 25, 49}.
Now
5
4
X
Y
pl ul1 = 0
(19)
(ui ri ) =
uR
i=1

l=1

Therefore, the PI-SDR for 64 QAM can be reformulated as,


min

U,u,S,s

s.t.

f (S, s)

U11
U=
U21
T

U12
U2,2

V. N EAR M AXIMUM L IKELIHOOD D ECODING A LGORITHM


In this method, the transmitted vector is expanded as a
linear combination of all the possible constellation points in
each dimension. Using this formulation, the minimization in
Euclidean distance space is expressed as a binary quadratic
minimization problem.
A. Problem Formulation
Considering the MIMO system model as in (2), the transmitted symbol vector s in this case is x of size N 1 with
N = 2Nt . Clearly elements of x is selected from a real symbol
set S = {s1 , s2 , . . . , sM }.
xi = ui (1)s1 + ui (2)s2 + + ui (M )sM i = 1, . . . , N (23)
Where,
ui (j) {0, 1}and

N
X

ui (j) = 1, i = 1, 2, . . . , N

(24)

j=1

Let u = [u1 (1) . . . u1 (M ) . . . . . . uN (1) . . . uN (M )]T is an


N N M binary vector such that Au = 1N , where A =
IN 1TM . From the equations in (23) and (24), x can be
expressed as
x = Su
(25)
Where, S = IN [s1 , . . . , sM ] is an N N M coefficient
matrix. Thus the system model can be written as

 
u
,u = 1
u2

y = HSu + n

(20)

S  ss , U  uu

(26)

At the receiver, the ML detection rule is given by

p1 1 + p2 u1 + p3 diag(U11 ) + p4 diag(U12 )

= arg min ||
x
y Hx||2
xi S

+ p5 diag(U22 ) = 0

(27)

(20) is a valid formulation because the constraints restricts ui = s2i , diag(U11 )i = u2i , diag(U12 )i =
u3i , diag(U22 )i = u4i . Also it has eliminated the rank one
constraints i.e. S = ssT , U = uuT .

is the ML estimated input vector and y


is the
Where, x
received vector. The equivalent form of the problem can be
formulated as

B. BC-SDR for 4q QAM


It is observed that the elements of s in our system model
() si {1, 3 (2q 1)}, or alternatively s2i
{1, 9, . . . , (2q 1)2 }. keeping this is mind the BC SDR is
formulated as
min
f (S, s)

Au=1N

min

||
y HSu||2

(28)
Let P = ST HT HS and c = ST HT y
. So the optimization
problem can be written as
min uT Pu + 2cT u
s.t.

(21)

S  ssT
q

1 diag(S) (2 1) 1
C. VA-SDR for 4q QAM
The VA-SDR for higher order QAM is the most generalized
version of 16 order QAM.It has the same form as in (13),
with the optimization variable X of dimension 2qNt ant the
matrices W and b redefined as
h
i
W = I 2I 4I . . . 2q1 I
(22)
h
iT
T
T
q1 T
b = bT
2b
4b
.
.
.
2
b
q
1
2
3

Au = 1N

(29)

u {0, 1}N M

U,u,S,s

s.t.

min uT ST HT HSu 2
yT HSu

Au=1N

(29) is a quadratic minimization problem with binary variable


u. In the following section, we discuss several SDP relaxation
models for this problem.
B. SDP Solution
The objective function in (29) is scalar and hence can be
modified as

uT Pu + 2cT u = tr uT Pu + 2cT u

1 uT
(30)

= tr LP
u uuT


Where, LP :=

0
c

cT
P

Define W = [bv IN VM (M 1) ]. Then, x can be written


as
x = Wp

Let
M N = {X MM N : X1M = 1N , xij {0, 1}, i, j}
(31)
Since, the constraints Au = 1N , ui {0, 1}N M are equivalent to u = vec(U), U M N , the minimization problem
in (29) can be written as

1 uT

min tr LP
(32)
u uuT
s.t.

u = vec(U), U M N

The feasible points of (32) are given by





1
1
1 uT =
Yu =
u
u

and

Y=

xT

xxT

pT WT

Wp

WppT WT

Let

=
Q

0TN (M 1)

bv

IN VM (M 1)

M(N M +1)((M 1)N +1)

Then
Q
T
Y = QR

uT

uuT

Where,

Yu is a rank one and positive semi-definite matrix. Also,


diag(uuT ) = u Yu and consequently, diag(Yu ) = Yu:,0 =
Yu T0,: and diag(uuT ) = u, where Yu:,0 and Yu0,: denotes
the first row and the first column of the matrix Yu respectively.
The SDP relaxation of the problem can be achieved by
removing the rank one restriction from the feasible set. A
larger convex set F is defined by taking the convex hull of
the original set M N i.e.,
F := conv {Yu : u = vec(U), U M N }

(34)

thereby, removing the rank one restriction from the problem.


So the relaxation model for the original minimization problem
given in (29) can be formulated as
min tr (LP Y)
s.t. Y F

(33)

(35)

It is clear that Yu are the feasible points of F. Since Yu


are rank one matrices, they lies in the set of extreme points
of F. So all the rank one matrices of the F in (35) for some
u is the optimal solution of (29).
The set F has no interior points. In the next relaxation
model F is replaced by another larger set containing F.
Let X MM N such that 1TM X = 1TN . If x = vec(X)
MM N 1 , using lemma 1 [5], x can be expressed as
i
1 h
x=
1M N 1 (IN VM (M 1) )1(M 1)N 1
M
+(IN VM (M 1) )
x
where x
= vec(X(1 : (M 1), 1 : N )) and

IM 1
MM (M 1)
VM (M 1) =
1TM 1
Let pT = [1 x
T ]
and
i
1 h
bv =
1M N 1 (IN VM (M 1) )1(M 1)N 1
M

R = ppT =

(36)

x
T

x
x
T

(37)

Clearly, R  0 and r00 = 1, rii = r0i , i. If Y is an extreme


point of F, then rij {0, 1}, otherwise rij [0, 1] i, j
{0, 1, . . . , N (M 1)}.
Using the above result a set Fr containing F is defined as
follows

Fr = Y SN M +1 : R S(M 1)N +1 , R  0,
(38)
Q
T , diag(Y) = Y0,:
r00 = 1, Y = QR
Thus, second relaxation model can be formulated as
T LQ Q)R

min tr(Q
s.t.

T RQ)
= (1, (Q
T RQ)
0,1:n )T
diag(Q

(39)

R0
(39) is the quasi-ML decoding rule for the MIMO model
in (2). From its optimal solution R, and using the relation
(36) Y can be obtained and a randomization procedure can
be employed to recover u.
VI. S IMULATION R ESULTS AND D ISCUSSION
To simulate the SDR detectors for MIMO we have considered that the elements of the channel matrix as Rayleigh
in magnitude i.e both real and imaginary parts are i.i.d.
Gaussian with variance 21 . Also the noise elements are circular
symmetric Gaussian. To check Symbol error rate performance
Pav
we have considered SNR per bit i.e. SN R = 2k
2 , Pav is
2
the average constellation power and is noise variance per
dimension in a 2k -ary constellation.For each step about 10000
Monte-Carlo simulations are performed; 200 for one channel
matrix averaged over 50 times.
Fig.1 and Fig.2 shows the performance for 16-QAM detectors discussed in section III in 2x2 and 4x4 MIMO system
respectively. The performance are superior to ZF detector in
both cases. We have also verified ML detector performance
in 2x2 system as it requires search over 22k = 256 possible
symbol vectors.Approximation techniqes used in both these
simulations are eigenvalue decomposition.Fig.3 demonstrates

Fig. 1. Performance of the three 16-QAM SDR techniques in 2x2 MIMO


system, compared with ZF and ML detector

Fig. 4. Performance of SDR detectors in NxN MIMO with 64-QAM


constellation

Fig. 2. Performance of the three 16-QAM SDR techniques in 4x4 MIMO


system, compared with ZF detector

Fig. 5. Average optimal objective value produced by SDR methods in 4x4


MIMO with SNR(16/64-QAM constellation)

Fig. 3. 16 QAM BC-SDR performance in 4x4 MIMO with different


approximation technique

Fig. 6. average optimal value produced by SDR methods with system size
(16-QAM constellation)

Fig. 7. Complexity performance (Average run time) of the 16-QAM SDR


detectors with system size

The first two models of [5] for Near ML decoding algorithm


are verified in Fig.8 for 16QAM constellation in a 2x2 MIMO
system and results are compared with conventional ZF and
ML detector.
From the simulation results BC-SDR technique is observed
to be the most efficient one in terms of complexity and
SER performance as well. It is also successfully inherited
to 64-QAM constellations whereas PI-SDR demands more
computational effort in higher order constellation and VA-SDR
lacks performance though it is computationally easy.
From a theoretical view it can be noted the the first
three techniques discussed relaxes the constraints on symbols
belonging to a discrete set. In PI-SDR the discrete constraints
are relaxed as quadratic relations, where BC SDR relaxes their
square within a bound using symmetry of QAM constellation.
The VA-SDR represents the symbols vectors as weighted sum
of antipodal elements and solves for the antipodal vectors. In
the near-ML decoding the SDP is formed to find a binary
vector that actually selects exactly one symbol from the set
containing all possible symbols. Performance of this method
is enhanced by imposing more constraints and increasing
complexity of implementation.
VII. C ONCLUSION
Different SDR methods discussed in the paper gives a
better solution for sub-optimal MIMO detection problem. The
first three methods are mainly for 4q -QAM constellations
whereas the near ML approach can be extended over any Mary constellation. The three 16-QAM SDRs and also its higher
order extensions are equivalent to each other and SDP solution
leads to same objective value. Also the SER performance of
these detectors are far superior to ZF detector and observed
to be near to optimal performance.
R EFERENCES

Fig. 8. SER performance of the Near-ML decoding approaches, in 2x2 MIMO


system using 16-QAM constellation

the performance of different Approximation techniques for BC


SDR in 16 QAM and they are found to be approximately same.
64-QAM equivalent models for BC-SDR and VA-SDR are
demonstrated in Fig.4. Performance of VA SDR is poor in this
case and needs special approximation method to improve it.
BC-SDR performs better than ZF as expected.
To check the equivalence among the BC/VA/PI SDR the
average optimal value is considered and it is found to be
same in each cases. The simulated results are for 4x4 MIMO
using 16-QAM and 64-QAM respectively shown in Fig.5. It
can be noted that VA SDR is equivalent to BCSDR even in
higher order QAM in terms of optimal value but detection
performance is poor. Also the equivalence in optimal value
with system size is simulated in Fig.6. Complexities of the
SDR methods are compared in terms of average run time to
solve the SDP with different system size and simulated results
are shown in Fig.7. The time considered here doesnt include
the approximation time as it is same in all the SDRs.

[1] E. G. Larsson, MIMO Detection Methods: How They Work, IEEE Signal
Processing Magazine, May 2009, pp. 91-95.
[2] Z. Q. Luo et al., Semidefinite Relaxation of Quadratic Optimization
Problems, IEEE Signal Processing Magazine, May 2010, pp. 20-34
[3] A. Wiesel et al., Semidefinite Relaxation for Detection of 16-QAM
Signaling in MIMO Channels, IEEE Signal Processing Letters, Vol. 12,
No. 9, September 2005, pp. 653-656.
[4] W. K. Ma et al., The Equivalence of Semidefinite Relaxation MIMO
Detectors for Higher-Order QAM, IEEE Journal of Selected Topics in
Signal Processing, Vol. 3, No. 6, December 2009, pp. 1038-1052.
[5] A. Mobasher et al., A Near-Maximum-Likelihood Decoding Algorithm for
MIMO Systems Based on Semi-Definite Programming, IEEE transactions
on Information Theory, Vol. 53, No. 11, November 2007, pp. 3869-3886.

You might also like