You are on page 1of 6

IEEE ICC 2014 - Wireless Communications Symposium

Convex Optimization based


Multiuser Detection for Uplink Large-scale MIMO
under Low-Resolution Quantization
Shengchu Wang
Department of Electrical Engineering and
Wireless and Mobile Communications R&D Center
Tsinghua University, Beijing, China, 100084
Email: wsc11@mails.tsinghua.edu.cn
AbstractIn large-scale Multiple-Input-Multiple-Output, the
Base Station (BS) is equipped with a large-size antenna array
containing tens even hundreds Radio Frequency channels. If so
many RF front-ends adopt the classical receivers, their cost and
power consumption during the receiving mode would increase
quickly. Therefore, we exploit low-resolution Analog-to-DigitalConvertor to design low-power and low-cost software dened
radio receivers for the BS. However, the new problem is how the
BS fullls channel estimation and Multiuser Detection (MUD)
under low-resolution quantization. In this paper, rst, least
square method is used for channel estimation, and a robust
Maximum Likelihood (ML) MUD problem is constructed to
take into account the channel estimation errors. Second, an
iterative multiuser detector is constructed by relaxing the ML
MUD problem as a convex optimization problem and then
solving the convex problem through the nonmonotone spectral
projected gradient method. Compared with the Minimum Mean
Square Error (MMSE) detector, the proposed detector has lower
computational complexity, and is more suitable for hardware
implementation. Simulation results show that it also outperforms
MMSE.

I.

I NTRODUCTION

Large-scale Multiple-Input-Multiple-Output (MIMO) is a


new multiuser MIMO systems where the BS is equipped
with a large-size antenna array containing tens even hundreds
Radio Frequency (RF) channels, and multiple single-antenna
users are served simultaneously [1], [2], [3]. Compared to the
conventional MIMO system, both the spectral efciency and
the power efciency of large-scale MIMO have been improved
signicantly [2], [3].
However, if so many RF front-ends adopt the design of the
classical Receivers (RXs), such as Zero or Low Intermediate
Frequency (IF) RX [5], etc., the cost and the power consumption during the receiving mode at the BS would increase at
a disastrous speed. Software Dened Radio RX (SDRX) [6],
[7], [8], [9] is a potential solution for this problem, because
many analog devices, such as RF synchronizer, low-pass lters,
etc, are saved. Unfortunately, high-speed and high-resolution
ADCs in SDRX, e.g., ADS5436 [10], call for high power and
cost which are tens times larger than that of the ADCs in
Zero-IF or Low-IF RXs. In this paper, they are replaced by
0 This work was supported by National 973 project (NO. 2013CB329002),
National 863 project (NO.2012AA011402), National Major Project
(NO.2013ZX03004007-002), Program for New Century Excellent Talents in
University (NCET), and Tsinghua-Qualcomm Joint Research Program.

978-1-4799-2003-7/14/$31.00 2014 IEEE

4789

Yunzhou Li and Jing Wang


Wireless and Mobile Communications R&D Center
Tsinghua University, Beijing, China, 100084
Email: {liyunzhou, wangj}@tsinghua.edu.cn
the low-resolutions ADCs [11], [12], where solution has been
sacriced in exchange for power and cost savings, and also
sampling speed improvement.
Although the power and cost at the BS can be decreased
signicantly by adopting the SDRXs with low-resolution ADC,
two new problems appears. The rst one is whether lowprecision ADC can be applied in large-scale MIMO. [13]
showed that the MIMO channel capacity loss due to coarse
quantization was surprisingly small. This fact indicates that, at
least from information theoretical point, low-precision ADC
can be considered in large-scale MIMO. The second one is
how the BS fullls channel estimation and Multiuser Detection
(MUD) based on the low-resolution data. In this paper, we
adopt the Least Square (LS) method for channel estimation
and construct a new iterative detector for MUD.
Different kinds algorithms, such as Maximum Likelihood (ML)[14], modied Minimum Mean Square Error
(MMSE)[15], Decision Feedback Equalization (DFE)[16], etc.,
have been developed for quantized MIMO. However, they
are not suitable for MUD in large-scale MIMO because of
the following reasons. First, when the number of users is
relatively large (e.g., 15), even the modied MMSE [15]
is problematic in hardware implementation due to the high
overhead from the matrix inversion [29], not to mention
ML and DFE. Second, these detectors fail to consider the
possible channel-estimation errors caused by thermal noise and
quantization noise. In this paper, based on [14] and [17], a
Robust ML MUD problem is constructed for quantized largescale MIMO (robust indicates the channel-estimation errors
have been taken into account). Then, it is relaxed as a convexoptimization problem, which is solved by the nonmonotone
Spectral Projected Gradient (SPG) method [18]. Finally, a
robust SPG detector (robust-SPGd) has been constructed.
Compared with MMSE with matrix inversion, robust-SGPd
with matrix-vector multiplications has lower complexity, and is
more suitable for hardware implementation. Simulation results
show that robust-SPGd outperforms MMSE.
The remainder of this paper is organized as follows. Section
II introduces the SDRX with low-resolution ADC, and the
problem models for both channel estimation and MUD in
uplink large-scale MIMO. Section III introduces the robust
ML MUD problem for large-scale MIMO under low-resolution
quantization, and constructs an iterative detector by relaxing
the ML problem as a convex optimization problem and then

IEEE ICC 2014 - Wireless Communications Symposium





 S I FW
IF
   S I FW





 S I FcQ
IF
   S I FcQ

  




 
4


E "

V# ^ 1F

[# ^ 1G

K# ^ 1W
Z ^ 1W

V  ^ 1F
X

K ^

1W

4F $

[8 ^ 1G

V8 ^ 1F

L

4$



4790

Fig. 1. a) The structure of the Zero IF Receiver [5]; b) The structure of the
SDR Receiver with low-Resolution ADC



Because of the reasons mentioned in Section II-A, we will


focus on the baseband transceiver shown by Figure 2.
In the system, the number of BS antennas is Nt and the
number of users is U . Each user is endowed with the whole
frequency bandwidth, and Single-Carrier Frequency Domain
Equalization (SC-FDE) [20] is considered in uplink. Based on
[20], every user u [U ] rstly transmit their own Unique Word
(UW) su CNc for channel estimation, and then their M Quadrature Amplitude Modulation (M -QAM) symbols xu



[  ^ 1G

B. Problem Models for Channel Estimation and Signal Detection

 !"

X #

For ease of comparison, Figure 1 shows both the existing


Zero-IF RX from [5] and the new SDRX with low-precision
ADC. The principles of Zero-IF RX are omitted here, and can
be found in [5]. In SDRX, out-of-band frequency components
are rejected by the Band-Select Filter (BSF), then the remaining signal is adjusted by the Low Noise Amplier (LNA) and
Automatic Gain Control (AGC), and sampled by the highspeed but low-resolution ADC. Finally, by the digital down
conversion (DDC), baseband data are recovered. SDRX saves
many circuit devices, such as analog frequency synthesizer,
analog Low-Pass Filters (LPFs), etc. Consequently, its cost and
power are much lower than that of Zero-IF RX. For example,
based on [19, Table I], the total power consumption of a typical
Zero-IF RX is 122.35mW. This value is decreased as 25mW
for the SDRX with 2.5mW 4-bit/1.25Gs/s ADC [11]. The
power is reduced by nearly 80 percent.
Because DDC is a complex operation involving decimation,
frequency down-conversion, and low-pass ltering [9], it is still
missing how the baseband takes into account the inuences
of the low-resolution ADC of the RF front-end. However,
we suppose that the baseband data after DDC also have
low precision. Consequently, it is important to rstly validate
whether the baseband communication under low-resolution
quantization is possible. From then on, ADC quantization is
considered at the baseband. How to combine together the RF
front-end and the baseband is our future research topic.




 

 



 

D

S YSTEM M ODEL

A. Software Dened Ration RX for Large-scale MIMO





II.


 

 



Notation: Aij is the (i, j)th elementof a matrix A; the


M N
2
Frobenius norm of A is ||A||F =
i=1
j=1 |Aij | ;
vec(A) stacks the columns of A as a vector; ai is the ith eleN
1
ment of a vector a; the p norm of a is ||a||p = ( j=1 |aj |p ) p ;
IN is the identity matrix with size N N ; 1N is a N
dimensional vector of all 1s; [K] = {0, 1, ..., K 1} for
an integer K; () and () extract the real and imaginary
parts of a complex matrix (vector), respectively; CN (z; m, )
and N (z ; m ,  ) are the Gaussian distribution function for

complex and real random vector z and z , respectively;

means component-wise less or equal to; sign(x) is 1 if x > 0


and 1 otherwise. ()T and () denote transpose and complex
conjugate transpose, respectively. x is the largest integer not
greater than x.



solving the latter through the nonmonotone spectral gradient


method. The simulation results are presented in Section IV,
and the paper is concluded in Section V.

4$

K8 ^ 1W

X 8

Fig. 2. Baseband transceiver between the Nt BS antennas and the U uplink


single-antenna users

XcNd . {su }U
u=1 contain unit-power elements and are orthogonal
with each other. So the inner product < si , sj > is equal to
is theM -QAM
Nc if i = j and 0 otherwise. X c = X + i X
alphabet where X = {(2k + 1)Am , k [ M 2]} with

power normalization factor Am (e.g. 1 10 for 16-QAM).
During the channel training phase, the Nt Nc received
pilot matrix at the BS is given by
Y = Qc (HS + W)
 

(1)

where H = [ h1 h2 ... hU ] , hu CNt CN (0, H )


with H = INt is the at-fading channel between the uth
T
user and the BS, S = [ s1 s2 ... sU ] , W CNt Nc is
the Gaussian noise matrix containing independent identically
distributed (i.i.d.) entries CN (0, 2 ) with variance 2 , and
Qc () is the complex quantizer which is introduced at the end
of this subsection.
Because there exists no Inter-Symbol-Interference (ISI)
under at-fading channel, Nd (the length of the data sequence)
is assumed to be one without loss of generality. In the uplink,
the received Nt 1 vector at the BS is
y = Qc (Hx + w)
 

(2)

where x = [ x1 x2 ... xU ] are the QAM symbols from


different users, w is i.i.d. Gaussian noise with variance 2 , and

IEEE ICC 2014 - Wireless Communications Symposium

Qc () is the complex quantizer.


Finally, we will introduce the design of the complex
quantizer Qc (). Qc (B) quantizes the (i, j)th element of a
R
I
R
) + i Q(Bij
) where Bij
and
complex matrix B as Q(Bij
I
Bij are the real and imaginary parts of Bij respectively. Q(x)
is the midrise b-bit quantizer [21] dened as
r = Q(x)


sign(x) ( |x| + ), |x| < G +

2
2,
=

sign(x) G,
otherwise.

(3)

where is the quantization step and G = (2b1 1/2) is


the saturation level. For the quantized measurement r in (3),
the upper and lower quantization boundaries are dened as

+
if r = G
up
,
(4)
V (r) =
r + /2 otherwise


and
V low (r) =

if r = G
.
r /2 otherwise

(5)

Z in (1) and z in (2) are the measurements before quantization. Their elements are distributed as CN (0, U + 2 )
because H is an i.i.d. Gaussian matrix with zero mean and
unit variance, and the entries of S and x have unit power.
[22] provided the optimal quantization step (denoted as )
for a b-bit uniform ADC (e.g. is 0.3352 for 4-bit ADC) to
minimize the mean square error (MSE) ||x Q(x)||2 where
x is a standard normal variable. Consequently, in order to
minimize||Y Z||F and ||y z||2 , in (3) should be set

as = (U + 2 ) 2 . Until now, the complex quantizer
Qc () appeared in (1) and (2) has been xed.
III.

ROBUST LARGE-SCALE MIMO MUD UNDER


THE CHANNEL-ESTIMATION ERRORS

A. Least-Square Channel Estimation


Additive noise model is adopted to analyze the quantization
effects [23], and (1) is transformed as
Y = HS + W + Wq ,

(6)

where Wq CNt Nc is the quantization noise.


In the ADC (3), saturation happens if the input sigprobability is
nal exceeds G +
2 , and the saturation

 +
2
erfc(N (2 2)) where erfc(x) = 2/ x et dt and
N = 2b . This probability is usually small (e.g. 0.0073 for
the 4-bit ADC). Consequently, most elements of the real and
imaginary parts of Wq are approximately uniformly distributed among [/2, /2]. Similar as [24], they are relaxed
as

Gaussian variables with zero mean and variance 2 12 [24].
By combining W and Wq together, (6) is simplied as

Y = HS + W,

(7)

is the approximate Gaussian noise with variance


where W


2 = 2 + 2 6.
Based on (7), the LS estimation of H is
LS = YS (SS )1 .
H

(8)

4791

Because UWs are orthogonal with each other, there exists


SS = Nc INt . Consequently, (8) is transformed as
= H + E,
LS = H + 1 WS
(9)
H
Nc
is the channel-estimation error mawhere E = 1/Nc WS
based
trix. The vector vec(E) is 1/Nc (S INt ) vec(W)
T

A)
vec(B)
[30].
Its
covariance
on vec(ABC)
=
(C

matrix is
2 Nc IU Nt based on the orthogonality of S and the
So the
Gaussian probability distribution function (pdf) of W.
columns of E are independent and distributed
as CN (0, E )

2
2
INt with E
=
2 Nc . Consequently, the
where E = E
LS is
conditional pdf of H over H
LSLS ) = CN ( H
LS , IU E ),
p(H|H
1

where
= H (E + H )
= INt with

2
). The proof for (10) can be found in [17].
1 (1 + E

(10)
=

B. Robust Maximum-Likelihood Detection under the ChannelEstimation Errors


Based on (2), the ML estimation of x under H is
maximum p(y|x, H) ,
xXcU

(11)

where the conditional probability p(y|x, H) is



p(y|x, H) = p(z D(y)|x, H) =
p(z|x, H)dz, (12)
D(y)

where

p(z|x, H) = CN (z; Hx, 2 INt )

(13)

and
D(y) = {z CNt |V low (yic )  zic  V up (yic );
i {1, 2, ..., Nt }, c {R, I}}
is the quantized hyper-rectangular cube in C2Nt [14].
Based on (12) and (13), (11) becomes to be

CN (z; Hx, 2 INt )dz .
maximum
xXcU

D(y)

(14)

(15)

However, accurate H is difcult to be acquired. In this paper,


LS (8). One naive method is
we only know its LS estimation H
treating HLS as true H and then (15) becomes a mismatchedML-MUD problem as

LS x, 2 INt )dz.
CN (z; H
(16)
maximum
xXcU

D(y)

But (16) is ineffective because it does not consider the


LS and H. In order to
possible mismatch E in (9) between H
take into account the channel-estimation errors, the objective
function of (11) is modied as

EH|H
LS [p(y|x, H)|HLS ]
 
LS )dH
=
[
p(z|x, H)dz]p(H|H
,
H
D(y)


LS )dH]dz
[ p(z|x, H)]p(H|H
=
D(y)

(17)

which averages the objective function of (11) over all realization of the unknown channel H conditioned on its available

IEEE ICC 2014 - Wireless Communications Symposium

LS by using the pdf (10).


estimation H
Based on [17], there exists

LS )dH = CN (z; Hx,

p(z|x, H)]p(H|H
2 INt ),

(18)

2
= H
LS and
2 = 2 + E
||x||22 .
where H
By substituting (18) into (17), we have



CN (z; Hx,
2 INt )dz.
EH|H
[p(y|x, H)|H] =

(19)

D(y)

We replace the objective function of (15) as (19), and obtain


the robust-ML-MUD problem


CN (z; Hx,
2 INt )dz.
(20)
maximum
xXcU

D(y)

C. Spectral Projected Gradient detector for Quantized largescale MIMO


According to (20), the real-valued robust-ML-MUD problem is

N (z ; Ax ,
2 I2Nt )dz ,
(21)
maximum
x X 2U

D  (y )


T
2, y =
, x =
where
2 =
(yT ) (yT)



T
(H) (H)
,A=
, and quanti(xT ) (xT )

(H)
(H)
zation hyper-rectangular D (y ) is dened as {z R2Nt |li 
zi  ui } where li = V low (yi ) and ui = V up (yi ).
By taking negative logarithm and some other algebraic
manipulations, (21) becomes to be

2

maximum fml (x ),


x X 2U

where fml (x ) =

2N
t
i=1

log((


ui aT
l aT x
i x
)( i i )).

(22)
Sim-

ilar as (20), (22) still has exponential complexity because it


involves an exhaustive search over X 2U . Consequently, it is
almost impossible to be implemented in large-scale MIMO
where U is usually large.
Subsequently, (22) is approximately solved by two steps.
Firstly, it is relaxed as a convex optimization problem. Because
the elements of x are the In-phase and Quadrature parts
of the
x, their values are constraint into

M -QAM vector
[( M 1)Am , ( M 1)Am ]. Based on this fact, (22) is
relaxed as
fml (x )
maximum
x
,
(23)
subject to x

2U
where is the feasible set dened as
= {x R |

Tm 12U
x
Tm 12U } with Tm = ( M 1)Am .

Secondly, (23) is a convex optimization problem about


minimizing a differential convex function fml (x ) [24] on a
closed convex set . Projected Gradient (PG) method [25] are
especially suitable for such kind of problems. However, PG
is considered to be slow. In order to improve the convergence
speed of PG, [18] developed nonmonotone Spectral PG (SPG)
method by associating the nonmonotone line search schemes
[26] with the spectral step-length [27] into PG. Finally, we
exploit the nonmonotone SPG to solve the robust-ML-MUD
problem (23). Finally, an iterative multiuser detector named as

4792

robust-SPGd is construct and shown by Table I. It contains


three main parts. First, Step 1.1 calculates the descent direction
dk , and calls for P () to ensure x+ in the Step 1.2 and 1.3
always stay in the feasible set . Second, the step-length is
determined by iterating between Step 1.2 and Step 1.3. Finally,
Step 2 calculates the safeguarded inverse Rayleigh quotient
k+1 for Step 1.1 of the next iteration.
By repeating the above derivation on the mismatched-MLMUD problem (15), a new detector named as mismatchedSPGd can also be constructed. It is similar as robust-SPGd

in A is replaced with HLS and
2 is set as 2 2.
except H
TABLE I.

P ROCEDURES OF ROBUST-SPG D

Initialization. x0 is the starting point. The orthogonal


projection P (z) for z R2U is dened as {sign(zi )
min(|zi |, Tm ), f or i [2U ]}. Initialize the nonmonotone
parameter J  1, a small parameter min > 0, a large parameter
max > min , a damping parameter (0, 1), safeguarding
parameters 0 < 1 < 2 < 1, a small positive number tol for
termination judgement. Set the iteration index k = 0 and the
safeguarded inverse Rayleigh quotient 0 [min , max ].
Step 1 (Backtracking)
Step 1.1. Set = 1. Compute dk = P (xk + k g(xk )) xk
and the negative gradient direction g(x ) as
g(x ) = AT fml (Ax ),
where, based on [24], fml (z)i =

2
exp(
u2
i /2)exp(li /2)
u

i exp(t2 /2)dt

(24)
with

li

u
i = (ui zi )/
and
li = (li zi )/
for i [2Nt ].
Step 1.2. Set x + = x k + dk .
Step 1.3. If
fml (x+ ) 

max

0jmin{k,J1}

f (xkj )+ < dk , g(xk ) >, (25)

then set xk+1 = x+ , sk = xk+1 xk , tk = g(xk+1 ) g(xk ),


and go to Step 2. If (25) does not exist, randomly choose value
for from [1 , 2 ], and return back to Step 1.2.
Step 2.(Update Spectral step-length)
If < sk , tk > 0, set k+1 = max ; otherwise, update k+1 as
k ,sk >
}}.
min{max , max{min , <s
<sk ,tk >
Step 3. (Termination)
If ||xk+1 xk ||2 /||x k ||2  tol, output x k+1 as the nal
estimation of x and stop. Otherwise, set k := k + 1, and go
back to Step 1 for the next iteration.

D. Complexity Analysis of the robust-SPGd


In the robust-SGPd, out-iteration starts from Step 1 and
ends at Step 3, and contains an inner iteration iterating between
Step 1.2 and Step 1.3 to x . In the inner iteration, only
and x+ are changed. Consequently, the computation complexity comes from the matrix-vector multiplication Ax+ =
Axk + Adk in the left term fml (x+ ) of (25). Axk and
Adk can be pre-calculated and pre-stored. From then on,
although keeps changing, Ax+ can be updated by simple
vector addition. Consequently, the complexity associated with
the inner iteration is o(Nt U ). Except the inner iteration, outiteration also contains two matrix-vector multiplications in
(24), so its complexity is o(Nt U ).
Consequently, the total complexity of the robust-SPGd
is o(KNt U ) where K is the total number of iterations for

IEEE ICC 2014 - Wireless Communications Symposium

TABLE II.

robust-SPGd to terminate. Simulation results show that K is


small especially when the ratio = Nt /U is large; K is
almost irrelevant to the problem size (U or Nt ) once is
xed. Then, under any xed 1, the robust-SPGd has
complexity o(Nt U ). Therefore, its complexity is one order of
magnitude lower than the o(Nt U 2 ) of Minimum Mean Square
Error (MMSE). Moreover, robust-SGPd with matrix-vector
multiplications is more suitable for hardware implementation
than MMSE with matrix inversion. Actually, when U is large
(e.g. 15), MMSE is problematic in hardware implementation
due to high overhead from the matrix inversion [29]. In the
contrary, the robust-SPDd does not have a similar problem.

C ONVERGENCE SPEED OF ROBUST-SPG D UNDER


DIFFERENT MIMO CONFIGURATIONS

MIMO Cong.

Cond.

K Zero

K MF

K MMSE

64x8
64x16
64x32
128x16
128x32
128x64
256x32
256x64
256x128

3.2
6.7
24.6
3.6
7.5
27.9
3.9
8.0
30.1

10.6
15.3
54.8
10.8
15.5
55.7
10.9
15.5
55.9

10.0
14.6
52.1
10.5
14.8
53.0
10.7
14.9
53.1

8.0
12.7
46.3
8.4
13.0
46.8
8.5
13.1
46.9

10

S IMULATION

In the robust-SPGd and mismatched-SPGd, the free parameters are set as follows: max = 1, min = 0.01, = 0.1,
1 = 0.2, 2 = 0.8, J = 5, tol = 103 . Spectral step-length
0 is initialized as 1. We adopt 4-bit ADC at the BS and
16QAM modulation among all the uplink users. The MIMO
size is denoted as Nt U . Signal to Noise Ratio (SNR) is
dened as 10 lg U2 .

10

10

SER

IV.

10

10

A. Convergence properties of the robust-SPGd


This experiment reveals how the convergence speed of the
robust-SPGd is inuenced by the condition number of the
Hessian matrix of fml (x ) and the accuracy of initial point
x0 . x0 is initialized as 0 or hard decisions from MF or hard
decisions from MMSE. SNR is xed as 10dB.
In Table II, Cond indicates the averaged condition number of the Hessian matrices of {fml (xk )}k where xk is the
signal estimation at the kth iteration of the robust-SPGd.
K Zero , K MF and K MMSE indicate the averaged number
of iterations for robust-SPGd to terminate under three different
initializing strategies. We repeat the Monte-carlo simulations
1000 times, and record Cond and the number of iterations.
The recorded results are averaged and shown in Table II.
Based on Table II, there exist K Zero > K MF > K MMSE .
This means that the more accurate the initial point x0 is, the
quicker the robust-SGPd converges. The number of iterations
is strongly inuenced by the Hessian condition number. In
general, the larger the ratio is, the smaller the condition
number is, and the faster robust-SPGd converges. For example,
K MF is 52.1 in 64x32 MIMO with = 2, and decreased
as 10.0 in 64x8 MIMO with = 8. Once is xed,
although the MIMO size is scaled up (shown by the shaded
rows), the condition number is almost unchanged, so is the
number of iterations. For example, K MF is about 10 in 64x8,
128x16, 256x32 MIMOs whose s are all equal to 8. So
the convergence speed of the robust-SGPd is irrelevant to Nt
and U once their ratio is xed. This further proves that the
complexity of robust-SPGd is one order lower than that of
MMSE under xed (please refer to Section III-D for more
details).
B. robust-SPGd vs mismatched-SPGd
In this experiment, robust-SPGd and mismatched-SPGd
are compared under different SNRs and U s. The number of
BS antennas is xed as 64. Figure 3 shows that, when U
is 8, robust-SPGd outperforms mismatched-SGPd by nearly
1dB; the performance gain is diminished as 0.3dB when U is

4793

64x8 robust
64x8 mismatch
64x16 robust
64x16 mismatch
64x32 robust
64x32 mismatch

10

10

10

15

20

25

30

35

SNR dB

Fig. 3. The SER performances of robust-SPGd (robust) and mismatchedSPGd (mismatched) in the MIMO systems where the number of BS antenna
is xed as 64, and the number of users is varied as 8, 16 and 32.

increased as 16, and disappears when U is 32. Consequently,


robust technique is more effective when Nt is much larger
than U . One possible reason is that, when = NUt is large,
thermal noise is averaged out [1] and the channel estimation
errors becomes the dominant source of errors. Consequently,
the robust-SPGd outperforms its mismathed counterpart by
considering the channel-estimation errors specially.
C. Symbol Error Rate Performances
In this experiment, we compare the SER performances
of different multiuser detectors for quantized large-scale MIMO. These detectors are quant-MMSE where quantization
noise is treated as Gaussian, modied-MMSE from [15],
and the proposed robust-SPGd. For ease of comparison,
we also present the SER performance of the un-quantized
MIMO where MMSE detector (denoted as unquant-MMSE)
in conjunction with LS channel estimator are adopted.
Figure 4 shows that, in the 64x32 MIMO, modiedMMSE has minor performance gains compared to quantMMSE. When SER is 102 (101 ),robust-SPGd outperforms the two quantized MMSE detectors (modied-MMSE
and quant-MMSE) by 6dB (1.5 dB), but has SNR loss
4dB(0.8dB) compared to unquant-MMSE. When SNR is
larger than 35dB, error oor appears in all the three detectors
for quantized MIMO. But the error oor of robust-SPGd
(102.7 ) is lower than that of quant-MMSE (102.13 ) or
modied-MMSE (102.16 ). In 64x8 MIMO, robust-SPGd

IEEE ICC 2014 - Wireless Communications Symposium

10

10

SER

10

10

10

10

10

64x32 unquantMMSE
64x32 quantMMSE
64x32 modifiedMMSE
64x32 robustSPGd
64x8 unquantMMSE
64x8 quantMMSE
64x8 modifiedMMSE
64x8 robustSPGd

10

15

20

25
SNR dB

30

35

40

45

50

Fig. 4. SER performances. unquant-MMSE indicates MMSE in MIMO


without quantization; quant-MMSE treats quantization noise as Gaussian;
modied-MMSE comes from [15]; robust-SPGd indicates the proposed
detector; solid lines for 64x32 MIMO and dash lines for 64x8 MIMO.

performs similarly to unquant-MMSE in general, and can


even outperform the latter when SNR is lower than 12dB. One
reason for this is that, in (23), the robust-SPGd exploits the fact
that the maximum amplitudes of the QAM symbols are nite.
The other two quantized MMSE detectors are always inferior
to unquant-MMSE, although their performance losses are
small in the low SNR region.
In summary, the robust-SPGd always outperforms MMSE
in quantized large-scale MIMO, and the performance gain is
more apparent when the number of BS antennas are not much
more than the number of users. It performs similarly or even
outperforms the unquantized MMSE when Nt is much larger
than U .
V.

C ONCLUSION

In this paper, we propose software dened ratio receivers


with low-resolution ADC for the base station antennas in
large-scale MIMO. Then we construct an iterative multiuser
detector for uplink large-scale MIMO under low-resolution
quantization. Compared to MMSE, the new detector has lower
complexity, more superior performances and is more suitable
for hardware implementation. It is also robust to the channel
estimation errors.
R EFERENCES
[1] F. Rusek, D. Persson, B. K. Lau, E. G. Larsson, T. L. Marzetta, O. Edfors,
and F. Tufvesson, Scaling Up MIMO: Opportunities and challenges with
very large arrays, IEEE Signal Process. Magazine, vol. 30, pp. 40-60,
Jan. 2013.
[2] E. G. Larsson, F. Tufvesson, O. Edfors, and T. L. Marzetta, Massive
MIMO for Next Generation Wireless Systems, arXiv:1304.6690 [cs.IT],
2013.
[3] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, Energy and spectral efciency of very large multiuser MIMO systems, IEEE Trans. Commun.,
vol. 61, no. 4, pp. 1436-1449, Apr. 2013.
[4] M. Jiang and L. Hanzo, Multiuser MIMO-OFDM for next generation
wireless systems, Proceed. of the IEEE, vol. 95, no. 7, pp. 1430-1469,
July 2007.
[5] P.-I. Mak, S.-P U, and R. P. Martins, Transceiver architecture selection:
Review, state-of-the-art survey and case study, IEEE Circuits Syst. Mag.,
vol. 7, pp. 6-25, 2007.

4794

[6] J. Mitola, The Software Radio Architecture, IEEE Commun. Mag., vol.
33, no. 5, pp. 26-38, May 1995,.
[7] A. A. Abidi, The path to the software-dened radio receiver, IEEE
Jour. Solid-state Circuits, vol. 42, no. 5, pp. 954-966, May 2007.
[8] D. M. Akos, M. Stockmaster, J. B. Y. Tsui, and J. Caschera, Direct
bandpass sampling of multiple distinct RF signals, IEEE Trans. on
Commun., vol. 47, no. 7, pp. 983-988, July 1999.
[9] H. J. Kim, J. U. Kim, J. H. Kim, H. Mei, and I. S. Lee, The
design method and performance analysis of RF subsampling frontend
fro SDR/CR receivers, IEEE Trans. on Indus. Electronics, vol. 57, no.
5, pp. 1518-1525, May 2010.
[10] http://www.ti.com/lsds/ti/data-converters/high-speed-adc-greater10msps-products.page.
[11] G. Van der Plas, S. Decoutere, and S. Donnay, A 0.16 pJ/conversionstep
2.5 mW 1.25 GS/S 4b ADC in a 90 nm digital CMOS process, in IEEE
ISSCC Dig. Tech. Papers, Feb. 2006, pp. 566-567.
[12] S. W. M. Chen, and R. W. Brodersen, A 6-bit 600-MS/s 5.3-mW
asynchronous ADC in 0.13- m CMOS, IEEE Jour. Solid-state Circuits,
vol. 41, no. 12, pp. 2669-2680, Dec. 2006.
[13] J. A. Nossek and M. T. Ivrlac, Capacity and coding for quantized
MIMO systems, in Intern. Wireless Commun. and Mobile Computing
Conf. (IWCMC), pp. 1387-1392, July 2006.
[14] A. Mezghani, M. S. Khou, and J. A. Nossek, Maximum likelihood
detection for quantized MIMO systems, in Proc. International ITG
Workshop Smart Antennas, pp. 278-284, 2008.
[15] A. Mezghani, M. S. Khou, and J. A. Nossek, A Modied MMSE
Receiver for Quantized MIMO Systems, in Proc. International ITG
Workshop Smart Antennas, Feb. 2007.
[16] A. Mezghani, M. S. Khou, and J. A. Nossek, Spatial MIMO Decision
Feedback Equalizer Operating on Quantized Data, In Proc. IEEE Int.
Conf. Acoust., Speech, Signal Processing (ICASSP), Apr. 2008.
[17] S. M. S. Sadough, M. A. Khalighi, and P. Duhamel, Improved iterative
MIMO signal detection accounting for channel-estimation errors, IEEE
Trans. on Vehicular Tech., vol. 58, no. 7, pp. 3154-3167, 2009.
[18] E. G. Birgin, J. M. Martinez, and M. Raydan, Nonmonotone spectral
projected gradient methods on convex sets, SIAM J. Optim., vol. 10, pp.
1196-1211, 2000.
[19] Y. Li, B. Bakkaloglu, and C. Chakrabarti, A system level energy model
and energy-quality evaluation for integrated transceiver front-ends, IEEE
Trans. VLSI Sys., vol. 15, no. 1, pp. 90-103, Jan. 2007.
[20] F. Pancaldi, G. M. Vitetta, R. Kalbasi, N. Al-Dhahir, M. Uysal, and H.
Mheidat, Single-carrier frequency domain equalization, IEEE Signal
Process. Mag., vol. 25, no. 5, pp. 37-56, Sep. 2008.
[21] J. Proakis, Digital Communication, 5th edition, Mc Graw Hill, 2008.
[22] J. Max, Quantizing for Minimum Distortion, IEEE Trans. on Inform.
Theory, vol. 6, pp. 7-12, March 1960.
[23] R. M. Gray, D. L. Neuhoff, Quantization, IEEE Trans. on Inform.
Theory, vol. 44, no. 6, pp. 2325-2383, Oct. 1998.
[24] A. Zymnis, S. Boyd, and E. Candes, Compressed sensing with quantized measurements, IEEE Signal Process. Letters, vol. 17, no. 2, pp.
149-152, Feb. 2009.
[25] P. Calamai and J. More, Projected gradient methods for linearly
constrained problems, Mathematical Programming, 39(1):93-116, 1987.
[26] L. Grippo, F. Lampariello, and S. Lucidi, A nonmonotone line search
technique for Newtons method, SIAM J. Numer. Anal., pp. 707-716,
1986.
[27] J. Barzilai and H. M. Borwein, Two point step size gradient methods,
IMA J. Number. Anal., pp. 141-148, 1988.
[28] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Cambridge Univ. Press, 2004.
[29] C. Shepard, H. Yu, N. Anand, L. E. Li, T. L. Marzetta, R. Yang, and
L. Zhong, Argos: Practical many-antenna base stations, in ACM Int.
Conf. Mobile Computing and Networking (MobiCom), Aug. 2012
[30] K. B. Petersen and M. S. Pedersen, The Matrix Cookbook, 2008
Available: http://orion.uwaterloo.ca/ hwolkowi/matrixcookbook.pdf.

You might also like