Largemimo

'
Large-MIMO: A Technology Whose Time Has Come

A. Chockalingam and B. Sundar Rajan
(achockal,bsrajan@ece.iisc.ernet.in) April 2010
Department of Electrical Communication Engineering (http://www.ece.iisc.ernet.in/) Indian Institute of Science Bangalore 560 012. INDIA
&
'
A. Chockalingam and B. Sundar Rajan: Large-MIMO: A Technology Whose Time Has Come
Dept. of ECE, IISc, Bangalore, April 2010
MIMO System
Tx1 Rx1
$
1
Tx2
Rx2
Input Data Stream
MIMO Encoder
MIMO Detector
Estimated Data
Tx Nt
RxNr
MIMO Channel
Nt : # Transmit Antennas
Nr : # Receive Antennas
&
Transmit Side (e.g., Base Station, Access Point, Set top box)
Receive Side (e.g., Set top box, Laptop, HDTV)
'
Why Multiple Antennas?
$
2
Nt : No. of transmit antennas, Nr : No. of receive antennas

# Antennas Error Probability (Pe ) Capacity (C), bps/Hz
Nt = Nr = 1 (SISO) Nt = 1, Nr > 1 (SIMO) Nt > 1, Nr > 1 (MIMO)
Pe SN R1
C = log(SN R) C = log(SN R) C = min(Nt , Nr ) log(SN R)

min(Nt , Nr )
Pe SN RNr
Pe SN RNt Nr Nt Nr : Diversity Gain
: Spatial Mux Gain
Large Nt , Nr increased spectral efciency

[1] I. E. Telatar, Capacity of multi-antenna Gaussian channels, European Trans. Telecommun., vol. 10, no. 6, pp. 585-595, November 1999. [2] G. J. Foschini and M. J. Gans, On limits of wireless communications in a fading environment when using multiple antennas, Wireless Pers. Commun., vol. 6, pp. 311-335, March 1998.
&
'
Large-MIMO Approach
$
3
Employ large number (several tens) of antennas at the Tx and Rx Achieve high spectral efciencies (tens to hundreds of bps/Hz)
e.g., 100 bps/Hz Data rate (bps) = Spectral efciency (bps/Hz) Bandwidth (Hz)
= 1 Gbps rate in just 10 MHz bandwidth 10 bps/Hz only
Limitation in current MIMO standards

spectral efciency: 2 to 4 transmit antennas
e.g., 750 Mbps in 80 MHz in 802.11n using 4 Tx antennas

&
do not exploit the potential of large spatial dimensions
'
Technological Challenges in Realizing Large-MIMO
$
4
Placement of large number of antennas in communication terminals

Feasible in moderately sized communication terminals (e.g., Set top boxes, Laptops, BS towers) use high carrier frequencies for small carrier wavelengths (e.g., 5 GHz, 60 GHz)
RF technologies
Multiple IF/RF transmit and receive chains
Large-MIMO detection
Need low-complexity detectors
Channel estimation
&
Estimation of large number of channel coefcients
'
16-Antenna Channel Sounding for IEEE 802.11ac (5 GHz)
$
5
&
[3] Gregory Breit et al, 802.11ac Channel Modeling, doc. IEEE 802.11-09/0088r0, submission to Task Group TGac, 19 January 2009.
'
$
6
64 64 MIMO Indoor Channel Sounding (5 GHz)
(a) 64-Antenna/RF hardware at 5 GHz
(b) LOS setup
[4] Jukka Koivunen, Characterisation of MIMO propagation channel in multi-link scenarios, MS Thesis, Helsinki University of
&
Technology, December 2007.
'
Some Recent Wireless Products
$
7
Can see the trend in packaging increasing number of antennas/RF chains in wireless products
&
Source: Internet
'
Detection Complexity: A Key Challenge in Large-MIMO
$
8
Optimal detection has exponential complexity in # transmit

antennas
Need low-complexity algorithms that are near-optimal A possible approach to low-complexity solutions
Seek algorithms from machine learning (ML) Large-dimension problems are routinely addressed in other areas (e.g., computer vision, web search) using ML algorithms Communications area too has beneted from ML algorithms
&
MUD in CDMA (large # users), Turbo/LDPC decoding (large frame sizes)
'
Linear Vector Channels
$
9
Several communication systems can be characterized by the

following linear vector channel model
yc = Hc xc + nc Examples
MIMO
xc Cdt , Hc Cdr dt , yc Cdr , nc Cdr
Coding
dt = Nt , # Tx antennas; Hc : Channel gain matrix;
dr = Nr , # Rx antennas; xc : Tx symbol vector yc : Rx signal vector; nc : Noise vector
dt = dr = K , # users; xc : Tx. bit vector; y : Rx signal vector; nc : Noise vector & c
CDMA
dt = k, # Information bits; dr = n, # Coded bits; xc : Information bit vector Hc : Generator matrix; yc : Rx signal vector; nc : Noise vector
Hc : Cross correlation matrix, %
'
$
10
Optimum Detection
Problem
Obtain an estimate of xc , given yc and Hc
Maximum likelihood (ML) solution

xM L =
arg min
xc
Adt
yc Hc xc
= (xc )
(1)
A: signaling alphabet; (xc ): ML cost

ML cost:
&
H (xc ) = xH HH Hc xc 2 yc Hc xc c c
'
Optimum Detection
$
11
Let A be M -PAM or M -QAM (Two PAMs in quadrature)

yc = yI + jyQ ,
M -PAM symbols take values from {Am , m = 1, , M }, Am = (2m 1 M )

xc = xI + jxQ , nc = nI + jnQ , Hc = HI + jHQ
Convert (1) into a real-valued system model y = Hx + n

H R2dr 2dt , y R2dr , x R2dt , n R2dr
H=@ 0 HI HQ HQ HI 1
T T A , y = [yI yQ ]T , x = [xT xT ]T , n = [nT nT ]T . I Q I Q
(2)
ML solution
&
xM L
arg min
S: 2dt -dimensional signal space (Cartesian product of A1 to A2dt ;

from which xi takes values, i
xS
y Hx
xT HT Hx 2yT Hx,
(3)
Ai : M -PAM signal set
= 1, , 2dt ).
ML Complexity:Exponential in dt
'
Optimum Detection
$
12
Maximum a posteriori (MAP) solution

Consider square M -QAM
M -PAM constellation (q1) (0) (1) denote the q = log2 ( M ) constituent bits of xi Let bi , bi , , bi
Each entry of x belongs to a
xi can be written as
q1
xi =
j=0
2j bi ,
(j)
i = 0, 1, , 2dt 1
Let the bit vector b
b =
Dening c
(0) b0
{1}2qdt be written as
(q1) (0) b0 b1 (q1) b1
(0) b2dt 1
(q1) b2dt 1
&
= [20 21 2q1 ], x can be written as x = (I2dt c)b
'
$
13
Optimum Detection
Rx signal model can be written as

y = H(I2dt c)b + n
= H R2dr 2qdt
MAP estimate of bi ,
(j) bi
(j)
i = 0, , 2dt 1,
j = 0, , q 1 is
arg max
a {1}
p bi = a | y, H
(j)
&
Complexity: Exponential in qdt
'
Sub-optimum Solutions
$
14
Matched lter (MF)
x M F = HT y
Zero-forcing (ZF) solution
xZF = H1 y
Minimum mean square error (MMSE) solution
xM M SE = (H + 2 I)1 y
These suboptimum solution vectors can be used as initial vectors

&
in search algorithms to improve performance further
'
Near-Optimal Algorithms for Large dt
$
15
Near-ML algorithms
Local neighborhood search based Likelihood ascent search (LAS) Reactive tabu search (RTS)
Near-MAP algorithms
Message passing based Belief propagation (BP)
&
Probabilistic association (PDA)
'
LAS Algorithm
$
16
Search for good solution vectors in the local neighborhood Neighborhood denition
Neighbors that differ in one coordinate
e.g., Consider A = {1}; x = [1, 1, 1, 1] 1-bit away neighbors of x: N1 (x) = [1, 1, 1,1], [1, 1,1, 1], [1,1, 1, 1], [1, 1, 1, 1]
Neighbors that differ in two coordinates 2-bit away neighbors of x:
N2 (x) = [1, 1,1, 1], [1,1, 1, 1], [1, 1, 1, 1], [1, 1, 1, 1], [1, 1,1, 1], [1,1, 1,1]
&
Choose best neighbor based on ML cost: ( ) = xT HT H 2yT H x x x
'
LAS Algorithm
START Compute initial solution vector Find the neighborhood of the solution vector Find the best vector in the neighborhood
$
17
Does this vector have a better cost function than that of the current solution vector? Yes
No
Make this neighbor as the current solution vector
END
&
'
An Illustration of LAS Search Path
$
18
&
'
$
19
&
'
$
20
&
'
$
21
&
'
$
22
Local minima trap
&
'
Large-Dimension Behavior of LAS in V-BLAST [5],[6]
$
23
* 1-LAS: 1-symbol away neighborhood * BER improves with increasing Nt (large-dimension effect)
26
10
24
ZFSIC ZFLAS SISO AWGN performance
10
22
Average received SNR required (dB)
20
10
Bit Error Rate
18
16
10
Increasing # antennas improves BER performance 10

4
14 Uncoded BER target = 103 Nt = Nr
10
ZFLAS (1 X 1) ZFLAS (10 X 10) ZFLAS (50 X 50) ZFLAS (100 X 100) ZFLAS (200 X 200) ZFLAS (400 X 400)
12
10
Almost SISO AWGN performance
10
For large # antennas ZFLAS, MFLAS, MMSELAS perform almost same 8 9 10
6 0 10
10
10
10
5 6 7 Average Received SNR (dB)
Number of Antennas, N t = Nr
[5] K. V. Vardhan, S. K. Mohammed, A. Chockalingam, B. S. Rajan, A low-complexity detector for large MIMO systems and multicarrier CDMA systems, IEEE Jl. Sel. Areas in Commun. (JSAC), vol. 26, no.3, pp. 473-485, April 2008. efciencies, & IEEE ICC2008, Beijing, May 2008. [6] S. K. Mohammed, K. V. Vardhan, A. Chockalingam, B. Sundar Rajan, Large MIMO systems: A low-complexity detector at high spectral
'
$
24
Complexity of 1-LAS in V-BLAST
Consider Nt = Nr Total complexity comprises of 3 main parts

1. Computing initial vector (e.g., ZF, MMSE): 2. Computing HT H: 3. Search operation:
O(Nt2) per symbol
O(Nt2 ) per symbol O(Nt ) per symbol (through simulations)
So, overall average per-symbol complexity: O(Nt2)

hundreds of spatial dimensions
&
This low-complexity allows detection of V-BLAST signals in

%
'
Large # Dimensions: The Key
$
25
Observation
In V-BLAST, LAS algorithm achieves near-ML performance, but only when the # antennas is in hundreds hundreds of antennas may not be practical
Note
LAS requires large # dimensions to perform well but, all dimensions need not be in space alone
Q1: Can large # dimensions be created with less # Tx antennas? A1: Yes. Use time dimension as well. Approach: Non-orthogonal STBCs Q2: Can LAS modied to work well for smaller (tens) dimensions?
&
A2: Yes. Approach: Escape strategies from local minima
'
An Escape Strategy from Local Minima [7]
$
26
Multistage LAS (M-LAS)
Start the algorithm as 1-LAS On reaching the local minima,
nd 2-symbol away neighbors of the local minima choose the best 2-symbol away neighbor if it has lesser cost than local
minima
run 1-LAS from this best neighbor till a local minima is reached
Expect better performance. Complexity is increased a little, but not by an order Escape strategy with 3-symbol away neighborhood on reaching local minima
Another promising strategy is reactive tabu search

&
A. Chockalingam, B. S. Rajan, A low-complexity near-ML performance achieving algorithm for large MIMO detection, [7] S. K. Mohammed,
IEEE ISIT2008, Toronto, June 2008.
'
$
27
Performance of M-LAS
* 3-LAS performs better than 1-LAS
10
0
10
10
MMSELAS, Nt=Nr=32 MMSEMLAS, Nt=Nr=32 MMSELAS, Nt=Nr=64 MMSEMLAS, Nt=Nr=64 AWGN SISO
10
Spatial Multiplexing, 4QAM

Bit Error Rate
Bit Error Rate
10
10
MMSE initial filter
BER improves with increasing Nt
10
Spatial Multiplexing 4QAM, MMSE initial filter
10
10
10
Nt = Nr = 16, MMSEMLAS Nt = Nr = 32, MMSEMLAS Nt = Nr = 64, MMSEMLAS Nt = Nr = 64, MMSE only Nt = Nr = 128, MMSEMLAS Nt = Nr = 256, MMSEMLAS AWGNonly SISO
10
10
6 8 Average Received SNR (dB)
10
12
6 Average Received SNR (dB)
10
12
(e) 3-LAS versus 1-LAS
(f) 3-LAS
&
'
Space-Time Block Codes
$
28
Provide redundancy across space and time Goal of space-time coding

Achieve the maximum Tx-diversity of Nt (i.e., full-diversity), high rate, decoding at low-complexity
An STBC is usually represented by a p nt matrix

rows: time slots;
p: # time slots
columns: Tx. antennas;
nt : # Tx. antennas
s12 s22 s42 s1nt s2nt
&
sij denotes the complex number transmitted in the ith time slot on the j th Tx antenna
s11 s 21 X = sp1
spnt
'
Space-Time Block Codes
$
29
Rate of an STBC,
k p k : number of information symbols sent in one STBC
r=
p: number of time slots in one STBC Higher rate means more information carried by the code XH X = |x1 |2 + |x2 |2 + + |xk |2 Int , xk and their conjugates
A matrix X is said to be a Orthogonal STBC if
Elements of X are linear combinations of x1 ,
2-Tx Antennas Codes (2 2 Alamouti Code)

X = & x1 x 2 x2 x 1 , k = 2, p = 2, r = 1, orthogonal STBC %
x1 , x2 , , xk are information symbols
'
Linear-Complexity Decoding of OSTBCs
$
30
Consider Alamouti code with nt = 2, nr = 1 Received signal in ith slot, yi , i = 1, 2, is

y1 y2 = h1 x1 + h2 x2 + n1 = h1 x + h2 x + n2 2 1
ML decoding amounts to
computing
x1 x2
= y1 h + y2 h2 1 = y1 h y2 h1 2
decoding x1 by nding the symbol in the constellation that is closest to x1 and decoding x2 by nding the symbol that is closest to x2
&
This decoding feature is called Single-Symbol Decodability (SSD)
'
Orthogonal vs Non-Orthogonal STBCs
$
31
Orthogonal STBCs are more widely known

advantages
e.g., 2 2 Alamouti code (Rate-1; 2 symbols in 2 channel uses)
linear complexity ML decoding, full transmit diversity

major drawback
rate falls linearly with increasing number of transmit antennas Non-orthogonal STBCs: less widely known
e.g., 2 2 Golden code (Rate-2; 4 symbols in 2 chl uses; same as V-BLAST)
&
advantages High-rate (same as V-BLAST, i.e., Nt symbols/channel use) Full Transmit diversity best of both worlds (in terms of data rate and transmit diversity) What is the catch decoding complexity
'
Non-Orthogonal STBCs
$
32
Golden code [8] (2 2 non-orthogonal STBC)

X =
where
x1 + x2 i(x3 + x4 )
and
1 5 2
x3 + x4 x1 + x2
k = 4, p = 2, r = 2
1+ 5 2
Features
Information Losslessness (ILL) Full Diversity (FD) Coding Gain (CG)
Golden code is a perfect code -
Perfect codes [9] achieve all the above three features
[8] J.-C. Belore, G. Rekaya, and E. Viterbo, The golden code: A 2 2 full-rate space-time code with non-vanishing determinants, IEEE Trans. on Information Theory, vol. 51, no. 4, pp. 1432-1436, April 2005.
&
[9] F. E. Oggier, G. Rekaya, J.-C. Belore, and E. Viterbo, Perfect space-time block codes, IEEE Trans. on Information Theory, vol. 52, no. 9, pp. 3885-3902, September 2006.
'
High-Rate Non-Orthogonal STBCs from CDA for any Nt
$
33
High-rate non-orthogonal STBCs from Cyclic Division Algebras (CDA) for arbitrary # transmit antennas, n, is given by the n n matrix [10]
6 Pn1 i 6 i=0 x1,i t 6 6 Pn1 i 6 i=0 x2,i t 6 X=6 . 6 . 6 . 6P 6 n1 x i 4 i=0 n2,i t Pn1 i i=0 xn1,i t 2 Pn1
i=0
x0,i ti
Pn1
i=0
i xn1,i n ti
i i i=0 x0,i n t Pn1 i i i=0 x1,i n t
Pn1
Pn1
Pn1
i=0
2i xn2,i n ti
. . .
2i i i=0 xn1,i n t Pn1 2i i i=0 x0,i n t

. . .
. . .
Pn1
i=0
Pn1
i=0
Pn1
i=0
x1,i n x2,i n x3,i n

. . .
(n1)i i t
(n1)i i t
3 7 7 7 7 7 7 7 7 7 7 7 5
(n1)i i t
Pn1
n = e
j2 n
,j=
n2 complex data symbols in one STBC matrix (i.e., n complex data symbols per channel use) = t = 1: Information-lossless (ILL); =e
5j
1, and xu,v , 0 u, v n 1 are the data symbols from a QAM alphabet

and t
i=0 Pn1 i=0
i xn3,i n ti i xn2,i n ti
Pn1
i=0 Pn1 i=0
2i xn4,i n ti 2i xn3,i n ti
Pn1
(n1)i i t i=0 xn1,i n Pn1 (n1)i i t i=0 x0,i n
= ej : Full diversity and ILL
Ques: Can large (e.g., 32 32) STBCs from CDA decoded?
Ans: LAS algorithm can.
&
[10] B. A. Sethuraman, B. Sundar Rajan, V. Shashidhar, Full-diversity high-rate space-time block codes from division algebras, IEEE Trans. on Information Theory, vol. 49, no. 10, pp. 2596-2616, October 2003.
'
Linear Vector Channel Model for NO-STBC
$
34
(n, p, k) STBC is a matrix Xc Cnp , n: # time slots, p: # tx antennas, k : # data symbols in one STBC; (n = p and k = n2 for NO-STBC from CDA) Received space-time signal matrix
Yc = Hc Xc + Nc ,
Consider linear dispersion STBCs where Xc can be written in the form

k
Xc
(i)
=
i=1
x(i) A(i) c c
(i)
Applying vec(.) operation

vec (Yc )
where Ac CNt p is the weight matrix corresponding to data symbol xc

k X i=1 k X i=1
x(i) vec (Hc A(i) ) + vec (Nc ) c c x(i) (Ipp Hc ) vec (A(i) ) + vec (Nc ) c c
&
'
$
35
Linear Vector Channel Model for NO-STBC

Dene yc = vec (Yc ) CNr p , Hc = (I Hc ) CNr pNt p ,
ac = vec (Ac ) CNt p ,
(i) (i)
nc = vec (Nc ) CNr p

k
System model can then be written in vector form as

yc =
i=1
x(i) (Hc a(i) ) + nc c c

(4)
= Hc xc + nc Hc CNr pk , whose ith column is Hc ac , i = 1, , k

(i)
(i)
xc Ck , whose ith entry is the data symbol xc
Convert the complex system model in (4) into real system model as before
&
Apply LAS algorithm on the resulting real system model
'
LAS Performance in Decoding NO-STBCs [11]

10
0
$
36
10
ILLonly STBCs, 4QAM Nr = Nt, 2Nt bps/Hz 10

1
MMSEonly (No LAS) (1, 2, 3, 4) (1) : 4x4 STBC, MMSEonly
10
Rate1/3 turbo Rate1/2 turbo Rate3/4 turbo Min SNR for capacity = 42.6 bps/Hz Min SNR for capacity = 64 bps/Hz Min SNR for capacity = 96 bps/Hz
10
Bit Error Rate
(2) : 8x8 STBC, MMSEonly (3) : 16x16 STBC, MMSEonly (4) : 32x32 STBC, MMSEonly 4x4 STBC, 1LAS
Bit Error Rate
10
32 x 32 ILLonly STBC Nt = Nr = 32, 16QAM Soft LAS outputs
10
8x8 STBC, 1LAS 16x16 STBC, 1LAS (5) : 32x32 STBC, 1LAS (5, 6)
10
Min SNR = 3.32 dB
8x8 STBC, 2LAS 16x16 STBC, 2LAS 10

5
BER improves with increasing Nr =Nt.

10
4
(6) : 32x32 STBC, 2LAS 4x4 STBC, 3LAS 8x8 STBC, 3LAS SISO AWGN 2 4 6 8 10 Average Received SNR (dB) 12 14
10
10
Min SNR = 6.83 dB
10
Min SNR = 11.12 dB

15 20 Average Received SNR (dB)
10
4x4 STBC, 2LAS
25
30
(g) Uncoded 8 8, 16 16, 32 32 NO-STBC, 4-QAM
(h) Turbo Coded 32 32 NO-STBC, 16-QAM
[11] S. K. Mohammed, A. Zaki, A. Chockalingam, B. Sundar Rajan, High-Rate Space-Time Coded Large-MIMO Systems: Low-Complexity Detection and Channel Estimation, IEEE Journal on Sel. Topics in Signal Processing (IEEE JSTSP):
&
Special Issue on Managing Complexity in Multiuser MIMO Systems, vol. 3, no. 6, pp. 958-974, December 2009.
'
Comparison with Other Architectures/Detectors [11]

Complexity No. MIMO Architecture/Detector Combinations (xed Nt (in # real operations per bit) at 5 102 uncoded BER SNR required to achieve 5 102 uncoded BER
$
37
= Nr = 16 and 32 bps/Hz
for all combinations)
i) ii) iii) iv) v) vi)
16 16 ILL-only CDA STBC (rate-16),

4-QAM and 1-LAS detection ( Proposed scheme [11] )
3.473 103
6.8 dB
16 16 ILL-only CDA STBC (rate-16),

4-QAM and ISIC algorithm [Choi, Ciof] Four 4 4 stacked rate-1 QOSTBCs, 256-QAM and IC algorithm [Jafarkhani] Eight 2 2 stacked rate-1 Alamouti codes, 16-QAM and IC algorithm [Jafarkhani]
1.187 105 5.54 106 8.719 103 4.66 104 1.75 104
11.3 dB
24 dB
17 dB
16 16 V-BLAST (rate-16) scheme,

4-QAM and sphere decoding
7 dB
&
16 16 V-BLAST (rate-16) scheme,

4-QAM and V-BLAST detector (ZF-SIC)
13 dB
'
Effect of Spatial Correlation?
$
38
Spatially correlated MIMO fading channel model by Gesbert et al [12]

Dr
s dr dt t
r Nr RX s Dt
Nt T X s
Figure: Propagation scenario for the MIMO fading channel model 1
correlated channel matrix: -
1 1/2 1/2 1/2 H = Rr ,dr Gr RS ,2Dr /S Gt Rt ,dt S

%
&
[12] D. Gesbert, H. Bolcskei, D. A. Gore, A. J. Paulraj, Outdoor MIMO wireless channels: Models and performance prediction, IEEE Trans. on Commun., vol. 50, pp. 1926-1934, December 2002.
'
Spatial Correlation Degrades Performance [11]

10
0
$
39
Uncoded (i.i.d. fading) Uncoded (correlated fading) Uncoded SISO AWGN Rate3/4 Turbo Coded (i.i.d. fading) Rate3/4 Turbo Coded (correlated fading) Min. SNR for capacity = 48 bps/Hz (i.i.d. fading) Min. SNR for capacity = 48 bps/Hz (correlated fading)
10
Bit Error Rate
10
16x16 ILLonly STBC, 16QAM Nt = Nr, LAS detection
10
Min. SNR = 11.1 dB (i.i.d.) Min. SNR = 12.6 dB (correlated)
10
Correlated MIMO channel parameters: fc = 5 GHz, R = 500 m, S = 30 dt = dr = 4 cm. Dr = Dt = 20 m. t = r = 90 deg.
10
10
15
20 25 30 35 Average Received SNR (dB)
40
45
50
Figure : Uncoded/coded BER performance of 1-LAS detector i) in i.i.d. fading, and ii) in correlated 2
MIMO fading in [3] with fc
= 5 GHz, R = 500 m, S = 30,Dt = Dr = 20 m, t = r = 90 , and
&
dt = dr = 2/3 = 4 cm. 16 16 STBC, Nt = Nr = 16, 16-QAM, rate-3/4 turbo code, 48 bps/Hz.

Spatial correlation degrades performance.
'
Increasing # Receive Dimensions Helps!

10
0
[11]
$
40
Nt = Nr = 12, uncoded Nt = 12, Nr = 18, uncoded Uncoded SISO AWGN Nt = Nr = 12, rate3/4 turbo coded Nt = 12, Nr = 18, rate3/4 turbo coded Min. SNR for Capacity = 36 bps/Hz (Nt = Nr = 12) Min. SNR for capacity = 36 bps/Hz (Nt = 12, Nr = 18)
10
Bit Error Rate
10
12x12 ILLonly STBC, 16QAM Nt = 12, Nr = 12,18, 1LAS detection
10
Min. SNR = 12.6 dB (Nr = 12)
Min. SNR = 9.4 dB (Nr = 18)
10
Correlated MIMO chl parameters: fc = 5 GHz, R = 500 m, S = 30 Nrdr = 72 cm, dt = dr D = D = 20 m = = 90 deg.

t r r t
10
10
15
20 25 30 35 Average Received SNR (dB)
40
45
50
Figure : Effect of Nr > Nt in correlated MIMO fading in [3] keeping Nr dr constant and dt = dr . 3
Nr dr = 72 cm, fc = 5 GHz, R = 500 m, S = 30, Dt = Dr = 20 m, t = r = 90 , 12 12
&
ILL-only STBC,
Nt = 12, Nr = 12, 18, 16-QAM, rate-3/4 turbo code, 36 bps/Hz. Increasing # receive
dimensions alleviates the loss due to spatial correlation.
'
Channel Estimation in Large-MIMO?
$
41
Training based channel estimation [11]
Send 1 Pilot matrix followed by Nd data STBC matrices
1 0 11111 00000 111111 000000 1 0 11111 00000 111111 000000 1 111111 0 000000 1 111111 0 000000 1 111111 0 000000 1 111100 000011 11 00 11111 00000 0 1 0
1 Pilot Matrix Space

Data STBCs

1 Frame
11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 00000 11111 1111 00000 0000
Data STBCs
Pilot Matrix
time
1 frame length (in # of channel uses), T = (Nd + 1)Nt 1 pilot matrix length (in # of channel uses), = Nt
[coherence time]
Obtain an MMSE estimate of the channel matrix during pilot phase Use estimated channel matrix to detect data matrices using LAS detection Iterate between detection and channel estimation
&
'
MIMO Capacity with Estimated CSIR

T 4 E logdet INt T
70 60 Perfect CSIR 1P + 8D (HH bound) 1P + 1D (HH bound) 16 x 16 MIMO System
$
42
Hassibi-Hochwald (H-H) bound [13] on capacity with estimated CSIR:

C 2 3 c HH H c 5 + Nt (1 + d ) + p Nt 2 2 d p
Hc
Ergodic Capacity (bps/Hz)
50 40 30 24 bps/Hz 21.3 bps/Hz 20 12 bps/Hz 10 0 4
4 6 8 Average SNR (dB)
4.3 dB
7.7 dB
10
12
14
16
Figure: H-H capacity bound [13] for 1P+8D (T = 144, = 16, p = d = 1) and 1P+1D (T = 32, = 4 -
16, p = d = 1) training for a 16 16 MIMO channel.

[13] B. Hassibi and B. M. Hochwald, How much training is needed in multiple-antenna wireless links?, IEEE Trans. on
Information Theory, vol. 49, no. 4, pp. 951-963, April 2003. &
'
How Much Training is Required?
$
43
Figure: Capacity as a function of Nt with SNR = 18 dB and Nr = 12. For a given Nr , SNR (), and 5
&
coherence time (T ), there is an optimum Nt [13].
'
BER Performance with Estimated Channel Matrix [11]

10
0
$
44
10
1P+1D;T=32; 12 bps/Hz 1P+8D;T=144; 21.3 bps/Hz 1P+24D;T=400; 23.1 bps/Hz 1P+48D;T=784; 23.5 bps/Hz Perfect CSIR; 24 bps/Hz
Bit Error Rate
10
10
16x16 ILLonly STBC Nt=Nr=16, 4QAM Rate 3/4 turbo code 1LAS detection Iterative Det/Est (4 iterns.)
10
10
10
12
14
16
18
20
Average Received SNR (dB) Figure: Turbo coded BER performance of LAS detection and channel estimation as a function of co6
herence time,
T = 32, 144, 400, 784 (Nd = 1, 8, 24, 48), for a given Nt = Nr = 16. 16 16
CSIR approaches to those with perfect CSIR in slow fading (i.e., large T ). &
ILL-only STBC, 4-QAM, rate-3/4 turbo code. Spectral efciency and BER performance with estimated
'
Other Promising Large-MIMO Detection Algorithms
$
45
Reactive Tabu Search [14] Probabilistic Data Association [15] Belief Propagation [16],[17] These algorithms exhibit large-dimension behavior; i.e., their bit error performance improves with increasing Nt .
[14] N. Srinidhi, S. K. Mohammed, A. Chockalingam, and B. S. Rajan, Low-Complexity Near-ML Decoding of Large Non-Orthogonal STBCs using Reactive Tabu Search, IEEE ISIT2009, Seoul, June 2009. [15] S. K. Mohammed, A. Chockalingam, B. S. Rajan, Low-complexity near-MAP decoding of large non-orthogonal STBCs using PDA, IEEE ISIT2009, Seoul, June 2009. [16] S. Madhekar, P. Som, A. Chockalingam, B. S. Rajan, Belief Propagation Based Decoding of Large Non-Orthogonal STBCs, IEEE ISIT2009, Seoul, June 2009.
&
[17] P. Som, T. Datta, A. Chockalingam, B. S. Rajan, Improved Large-MIMO Detection using Damped Belief Propagation, IEEE ITW2010, Cairo, January 2010.
'
$
46
Reactive Tabu Search
Another iterative local search algorithm

A metaheuristics algorithm cannot guarantee optimal solution, but generally gives near optimal solution
Uses tabu mechanism to escape from local minima or cycles

This is meant to ensure efcient exploration of the search space
Certain vectors are prohibited (made tabu) from becoming solution vectors for certain number of iterations (called tabu period) depending on the search path
The reactive part adapts the tabu period

& %
'
RTS Algorithm [14]

E
$
47
START Compute initial solution vector
Yes
Is any move nontabu?
Find the neighborhood of the solution vector

No
Find the best vector in the neighborhood
Make the oldest move performed as nontabu
Does this vector have the best cost function found so far?
No
Yes
Make this neighbor as the current solution vector Update tabu matrix to reflect current and past P moves Check for repetition of the solution vector
Is the move to this vector tabu?

Yes
No
Update tabu period P based on repetition
stopping
No
criterion satisfied?
Exclude the vector from the neighborhood

A B E C D
Yes
&
END
'
An Illustration of RTS Search Path
$
48
&
'
$
49
&
'
$
50
&
'
$
51
&
'
$
52
&
'
$
53
&
'
$
54
&
'
$
55
Global Minima
&
'
Performance of RTS in V-BLAST

10
0
$
56
10
VBLAST, 4QAM SNR = 10 dB, MMSE initial vector 10

Bit Error Rate
Bit Error Rate
8x8 VBLAST 16x16 VBLAST 32x32 VBLAST 64x64 VBLAST SISO AWGN
10
LAS
10
VBLAST, 4QAM MMSE initial vector
10
10
10
10
10
50
100
150 200 250 300 350 400 Maximum number of iterations, max_iter
(a) Convergence of RTS
450
500
10
16x16 VBLAST, LAS 32x32 VBLAST, LAS 64x64 VBLAST, LAS 16x16 VBLAST, RTS 32x32 VBLAST, RTS 64x64 VBLAST, RTS SISO AWGN 2 4 6 8
RTS
10
12
Average received SNR (dB)

(b) RTS versus LAS
[18] N. Srinidhi, S. K. Mohammed, A. Chockalingam, B. S. Rajan, Near-ML Signal Detection in Large-Dimension Linear Vector
&
Channels Using Reactive Tabu Search, Online arXiv:0911.4640v1 [cs.IT] 24 Nov 2009.
'
$
57
Performance/Complexity of RTS in V-BLAST

* RTS performs better than LAS at the same order of LAS complexity
10
0
24
log (Average no. of real operations)
LAS, 32x32 VBLAST, 64QAM RTS, 32x32 VBLAST, 64QAM SISO AWGN, 64QAM LAS, 32x32 VBLAST, 16QAM RTS, 32x32 VBLAST, 16QAM SISO AWGN, 16QAM
22 20 18 16 14 12 10
10
Bit Error Rate
RTS (overall) RTS (search part) LAS (overall) LAS (search part) 3 Nt
Nt
2
VBLAST, 4QAM BER = 0.01
10
10
10
10
15
20
25
30
35
40
45
50
4.2
4.4
4.6
4.8
5.2
5.4
5.6
5.8
log 2 (N t )
(d) Complexity
&
(c) Performance in 32 32 V-BLAST, M -QAM
'
$
58
Performance of RTS in NO-STBC [14]

* RTS performs better than LAS
10
0
STBC, 4QAM MMSE initial vector
10
STBC, 16QAM MMSE initial vector 10

Bit Error Rate
1
10
Bit Error Rate
LAS, 8x8 STBC RTS, 8x8 STBC LAS, 16X16 STBC RTS, 16x16 STBC SISO AWGN
10
4x4 STBC (LAS) 4x4 STBC (RTS) 8x8 STBC (LAS) 8x8 STBC (RTS) 12x12 STBC (LAS)
4x4 STBC
10
8x8 STBC
10
8x8 STBC
10
10
12x12 STBC
12x12 STBC (RTS) SISO AWGN
10
16x16 STBC
10
10
12
10
15

(e) 4-QAM
20 25 30 35 Average received SNR (dB)
40
45
50
&
(f) 16-QAM
'
Probabilistic Data Association
$
59
Originally developed for target tracking Used in digital communications recently PDA
A reduced complexity alternative to a posteriori probability (APP) detector/decoder/equalizer. Has been applied in
Multiuser detection in CDMA (Luo et al 2001, Huang and

Zhang 2004, Tan and Rasmussen 2006)
&
Turbo equalization (Yin et al 2004)
'
PDA Based Large-MIMO Detection [15]
$
60
Iterative algorithm
In each iteration, 2qk statistic updates (one for each bit) are performed
(j) Likelihood ratio of bit bi in an iteration is

(j) i
P y|bi
(j) (j)
= +1 = 1
(j)
P bi P bi
(j) (j)
= +1 = 1
(j)
P y|bi
= i
= i
Received signal vector y can be written as

2k1 q1
(j) hqi+j bi
+
l=0
m=0
hql+m bl
m=q(il)+j
(m)
+n
e = n (interf erence+noise vector)
&t
h : tth column of H
'
$
61
(j) (j) Dene pj+ = P (bi = +1) and pj = P (bi = 1) i i
To compute i , approximate the distribution of n to be Gaussian Mean of y

2k1 q1
(j)
j+ = E(y|bi = +1) = hqi+j + i

l=0 (j)
m=0
(j)
hql+m (2pm+ 1) l
m=q(il)+j
Covariance of y
&
j = E(y|bi = 1) = j+ 2hqi+j i i
2k1 q1 m+ hql+m hT (1 pm+ ) ql+m 4pl l
Cj = 2 I2Nr p + i
l=0
m=0
m=q(il)+j
'
$
62
(j) Using j and Cj , P (y|bi = 1) can be written as i i
P (y|bi = 1) =
(j)
e(yi
j T ) (Cj )1 (y i 1
j ) i
(2)Nr p |Cj | 2 i
Using (5), ij can be written as

ij ij = e
((yj+ )T (Cj )1 (yj+ )(yj )T (Cj )1 (yj )) i i i i i i
(j) Compute i using i(j) and i(j) (j) Update the statistics of bi as (j) P (bi (j)
= +1|y) =
1+
(j) i
(j) P (bi
= 1|y) =
1 1 + i
(j)
&
This completes one iteration of the algorithm
'
$
63

(j) (j) Updated values of P (bi = +1|y) and P (bi = 1|y) for all i, j are
fed back as a priori probabilities to the next iteration
Algorithm terminates after a certain number of iterations At the end of the last iteration,
decide bi as +1 if i 1, and 1 otherwise
(j) (j)
In coded systems
feed i s as soft inputs to the decoder
& %
(j)
'
Performance of PDA in V-BLAST

10
0
$
64
* PDA algorithm also exhibits large-dimension behavior

Nt = Nr = 8 Nt = Nr = 16 Nt = Nr = 32 Nt = Nr = 64 Nt = Nr = 96 SISO AWGN
10
10
Bit Error Rate
10
VBLAST MIMO, Nt = Nr 4QAM, m = 5
10
Performance improves with increasing Nt = Nr.
10
10
12
14
16
Average Received SNR (dB)
Figure: BER performance of PDA based detection of V-BLAST MIMO for Nt = Nr = 8, 16, 32, 64, 96, 7
&
4-QAM, and m
= 5 iterations.
'
$
65
Performance of PDA in NO-STBC [15]
10
LAS, 4x4 STBC PDA, 4x4 STBC Nt x Nt nonorthogonal ILL STBCs Nt = Nr, 4QAM LAS, 8x8 STBC PDA, 8x8 STBC LAS, 16x16 STBC PDA, 16x16 STBC SISO AWGN
10
Nt x Nt nonorthogonal ILL STBCs Nt = Nr, 16QAM Number of iterations m=10 for PDA MMSE init. vec. for LAS 10
1
10
LAS, 4x4 STBC PDA, 4x4 STBC LAS, 8x8 STBC PDA, 8x8 STBC LAS, 16x16 STBC PDA, 16x16 STBC SISO AWGN
10
Bit Error Rate
10
Bit Error Rate
Number of iterations m = 10 for PDA MMSE initial vector for LAS 10

3
BER improves with increasing Nt = Nr
10
10
10
BER improves with increasing Nt=Nr
. 10
5
12
14
16
10
10
15
35
40
45
(a) 4-QAM
(b) 16-QAM
&
'
Large-MIMO Detection Based on Graphical Models
$
66
Belief propagation (BP) is proven to work in cycle-free graphs BP is often successful in graphs with cycles as well MIMO graphical models are fully/densely connected Graphical models with certain simplications/assumptions work
successfully in large-MIMO detection 1. Use of Pairwise Markov Random Field (MRF) based graphical model in conjunction with message/belief damping [16],[17] 2. Use of Factor Graph (FG) based graphical model with Gaussian
&
Approximation of Interference (GAI) [17]
'
Performance of Damped BP on Pairwise MRF [17]

10
0
$
67
10
Nr=Nt=16 Nr=Nt=24 SISO AWGN
10
VBLAST, BPSK, message damped BP Message damping factor m = 0.2 # BP iterations = 5
10
Nt=Nr=4 Nt=Nr=8 Nt=Nr=16 Nt=Nr=24 Nt=Nr=32 SISO AWGN
Bit Error Rate
Bit Error Rate
10
10
VBLAST, BPSK Average received SNR = 8 dB Message damped BP # BP iterations = 5
10
BER improves with increasing Nt=Nr
10
10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
10
10
Message Damping Factor ( m)
Average Received SNR (dB)
(c) Performance as a function of damping factor
(d) BP on pairwise MRF exhibits large-dimension behavior
Damping signicantly improves performance

&
Order of per-symbol complexity: O(Nt2 )
'
Large-MIMO Detection using BP on FGs [17]
$
68
Each entry of the vector y is treated as a function node

(observation node)
Key ingredient: Gaussian approximation of the interference

interf erence 2Nt
Each symbol, xi {1}, is treated as a variable node
yi = hik xk +
j=1,j=k
hij xj + ni ,
= zik
2 is modeled as CN (zik , zik ) with zik
PNt
j=1,j=k
hij E(xj ),
and
%
&
2 zik =
P2Nt
2 j=1,j=k |hij | Var(xj ) +
2 2
, where hij is the (i, j)th element in H
'
Large-MIMO Detection using BP on FGs [17]
$
69
With xi s {1}, the log-likelihood ratio (LLR) of xk at observation node i, denoted by k , is i k i 2 p(yi |H, xk = 1) = 2 (h (yi zik )) = log ik p(yi |H, xk = 1) zik
LLR values computed at observation nodes are passed to variable nodes. Using these LLRs, variable nodes compute the probabilities exp( l=i k ) l pk+ = pi (xk = +1|y) = i 1 + exp( l=i k ) l
and pass them back to the observation nodes.
At the end, xk is detected as

&
This message passing is carried out for a certain number of iterations.

2Nr
xk = sgn
i=1
k i
'
$
70
Message Passing on Factor Graphs [17]
Figure: Message passing between variable nodes and observation nodes. 8
&
'
Performance of BP on FGs with GAI [17]

10
0
$
71
10
Nt = Nt; 4QAM # BP iterations = 20 Message damping factor = 0.4 10

1
10
Nt=Nr; 4QAM; STBCs from CDA # BP iterations = 20 Message damping factor = 0.4
Bit Error Rate
10
Bit Error Rate
10
8x8 VBLAST 10
3
16x16 VBLAST 24x24 VBLAST
10
4x4 STBC 8x8 STBC
10
32x32 VBLAST 64x64 VBLAST SISO AWGN

10
4
16x16 STBC SISO AWGN
10
10
12
10
10
12

(a) V-BLAST
(b) Non-Orthogonal STBC
BP with GAI achieves near-optimal performance for increasing Nt = Nr with O(Nt ) per-symbol complexity
& %
'
Large-MIMO Applications/Standardization
$
72
Potential Applications
Fixed Wireless IPTV/HDTV distribution (e.g., in 5 GHz band)
potentially big markets in India

High-speed back haul connectivity between BSs/BSCs using high data rate large-MIMO links (e.g., in 5 GHz band) Wireless mesh networks
Large-MIMO in Wireless Standards?

LTE-Advanced, WiMax (IEEE 802.16m)
Multi-Gigabit Rate LAN/PAN (e.g., in 5GHz / 60 GHz band)
Evolution of WiFi standards (IEEE 802.11ac and 802.11ad)

&
Can consider 12 12, 16 16, 24 24, 32 32 MIMO systems
'
Concluding Remarks
$
73
Low-complexity detection
critical enabling technology for large-MIMO no more a bottleneck
Large-MIMO systems can be implemented Large-MIMO approach scores high on spectral efciency and
QAM size) operating SNR compared to other approaches (e.g., increasing
Standardization efforts can consider reaping the benets of

&
large-MIMO in their evolution
'
$
74
Thank You
&

Largemimo

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Largemimo

Uploaded by

Copyright:

Available Formats

'

Large-MIMO: A Technology Whose Time Has Come

Dept. of ECE, IISc, Bangalore, April 2010

Input Data Stream

Receive Side (e.g., Set top box, Laptop, HDTV)

Dept. of ECE, IISc, Bangalore, April 2010

Why Multiple Antennas?

Nt : No. of transmit antennas, Nr : No. of receive antennas

Nt = Nr = 1 (SISO) Nt = 1, Nr > 1 (SIMO) Nt > 1, Nr > 1 (MIMO)

C = log(SN R) C = log(SN R) C = min(Nt , Nr ) log(SN R)

Pe SN RNt Nr Nt Nr : Diversity Gain

: Spatial Mux Gain

Large Nt , Nr increased spectral efciency

Dept. of ECE, IISc, Bangalore, April 2010

= 1 Gbps rate in just 10 MHz bandwidth 10 bps/Hz only

Limitation in current MIMO standards

e.g., 750 Mbps in 80 MHz in 802.11n using 4 Tx antennas

do not exploit the potential of large spatial dimensions

Dept. of ECE, IISc, Bangalore, April 2010

Technological Challenges in Realizing Large-MIMO

Placement of large number of antennas in communication terminals

Dept. of ECE, IISc, Bangalore, April 2010

16-Antenna Channel Sounding for IEEE 802.11ac (5 GHz)

Dept. of ECE, IISc, Bangalore, April 2010

64 64 MIMO Indoor Channel Sounding (5 GHz)

(a) 64-Antenna/RF hardware at 5 GHz

(b) LOS setup

Technology, December 2007.

Dept. of ECE, IISc, Bangalore, April 2010

Some Recent Wireless Products

Dept. of ECE, IISc, Bangalore, April 2010

Detection Complexity: A Key Challenge in Large-MIMO

Optimal detection has exponential complexity in # transmit

MUD in CDMA (large # users), Turbo/LDPC decoding (large frame sizes)

Dept. of ECE, IISc, Bangalore, April 2010

Linear Vector Channels

Several communication systems can be characterized by the

xc Cdt , Hc Cdr dt , yc Cdr , nc Cdr

dt = Nt , # Tx antennas; Hc : Channel gain matrix;

dr = Nr , # Rx antennas; xc : Tx symbol vector yc : Rx signal vector; nc : Noise vector

dt = dr = K , # users; xc : Tx. bit vector; y : Rx signal vector; nc : Noise vector & c

Dept. of ECE, IISc, Bangalore, April 2010

Maximum likelihood (ML) solution

A: signaling alphabet; (xc ): ML cost

Dept. of ECE, IISc, Bangalore, April 2010

Let A be M -PAM or M -QAM (Two PAMs in quadrature)

M -PAM symbols take values from {Am , m = 1, , M }, Am = (2m 1 M )

Convert (1) into a real-valued system model y = Hx + n

S: 2dt -dimensional signal space (Cartesian product of A1 to A2dt ;

Ai : M -PAM signal set

Dept. of ECE, IISc, Bangalore, April 2010

Maximum a posteriori (MAP) solution

Let the bit vector b

= [20 21 2q1 ], x can be written as x = (I2dt c)b

Dept. of ECE, IISc, Bangalore, April 2010

Rx signal model can be written as

Complexity: Exponential in qdt

Dept. of ECE, IISc, Bangalore, April 2010

Matched lter (MF)

Zero-forcing (ZF) solution

Minimum mean square error (MMSE) solution

These suboptimum solution vectors can be used as initial vectors

in search algorithms to improve performance further