Sei sulla pagina 1di 64

Precoding and Massive MIMO

Precoding and Massive MIMO

Jinho Choi
School of Information and Communications
GIST

October 2013

1 / 64
Precoding and Massive MIMO

1. Introduction

2. Overview of Beamforming Techniques

3. Cooperative (Network) MIMO


3.1 Multicell with a common group of users
3.2 Multicell with different groups of users

4. Massive MIMO
4.1 Partial cooperation
4.2 Pilot contamination in massive MIMO
4.3 Pilot contamination precoding

5. Conclusions

2 / 64
Precoding and Massive MIMO
1. Introduction

1. Introduction
I Beamforming with antenna arrays has been studied for
wireless communications since early 90s.
I Transmit and receive beamforming have been considered for
cellular systems to improve the signal-to-noise ratio (SNR) or
extend the coverage.
I A better beamforming gain can be achieved if the number of
antennas in an array is large.
I In cellular systems, as base stations (BSs) can have a number
of antennas, beamforming can be easily employed at BSs.
I In this case, transmit beamforming becomes downlink
beamforming and receive beamforming becomes uplink
beamforming.

3 / 64
Precoding
Beamformingandand
Massive MIMO
Massive MIMO
1.
1. Introduction
Introduction

Downlink beamforming = transmit beamforming at BS

Beamformer

w1

Data symbols

w2

Dynamic or static
Beampattern

wL

44 // 64
60
Precoding and Massive MIMO
1. Introduction

I Unfortunately, due to various problems, the number of


antenna elements in an array for downlink beamforming is
limited.
I However, in 5G, transmit (or downlink) beamforming has been
considered seriously to extend the coverage.
I In this tutorial, we review existing downlink beamforming
approaches and focus on cooperative and noncooperative
approaches in multi-cell MIMO systems.
I In the end, we wil attempt to highlight the differences between
network MIMO and massive MIMO in multi-cell systems.

5 / 64
Precoding and Massive MIMO
1. Introduction

How transmit beamforming works:


h1
TX 1

h2
TX 2

Let h1 and h2 denote the channel coefficients from TX antennas 1


and 2, respectively. If TX antenna k transmits wk s, where wk is
the weight and s is the signal to be transmitted, the received signal
becomes

r = h1 w1 s + h2 w2 s + n = (h1 w1 + h2 w2 )s + n,

where n is the background noise.


6 / 64
Precoding and Massive MIMO
1. Introduction

The SNR at the receiver becomes


|h1 w1 + h2 w2 |2 Es ||h||2 ||w||2 Es
SNR = ,
N0 N0
where the equality holds if wk hk .
I A full diversity order is equal to the number of antennas.
I The more transmit antennas, the better performance.
I However, the channel state information (CSI) is required for
transmit beamforming to achieve the maximum SNR.
I If the transmitter does not know the CSI, the receiver has to
feed back it to the transmitter.
I With limited feedback, the number of antennas cannot be
large, which has been one of the major drawbacks of transmit
beamforming, which may not be a drawback in TDD mode
(using the channel reciprocity).
7 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

2. Overview of Beamforming Techniques


I There are three different systems where beamforming can be
employed:
A. Point-to-point MIMO (single-user beamforming)
B. Multiuser (point-to-multipoint) MIMO (multiuser
beamforming)
C. Multipoint-to-multipoint (or network) MIMO (cooperative
beamforming)
I We can also consider a different system in which beamforming
plays a crucial role:
D. Massive MIMO: this is the case where single-user beamforming
is employed in a multi-cell system
I For cellular systems,
I single-cell: point-to-point or multiuser MIMO
I multi-cell: network MIMO (there is inter-cell interference
problem), massive MIMO
8 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

A. Point-to-point MIMO
I To achieve diversity gain:
I beamforming with CSI at transmitter
I space-time coding (STC) without CSI at transmitter
I To achieve multiplexing gain:
I SVD with CSI at transmitter
I BLAST techniques without CSI at transmitter
I There are also other techniques that enjoy the trade-off
between diversity and multiplexing gains
Ref. Zheng and Tse, Diversity and Multiplexing: A Fundamental
Tradeoff in Multiple Antenna Channels, IEEE TIT, May 2003

9 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

Beamforming for point-to-point MIMO


I It is often known as single-user beamforming
I To maximize the SNR, the principle of matched filtering (MF)
can be employed when CSI is available, which is known as
maximal ratio transmission (MRT) scheme in the context of
MIMO.
I There are also other approaches without CSI or partial CSI:
IBlind beamforming (or long-term transmit beamforming) with
statistical properties of channels
I Semi-blind beamforming (with partial CSI)
I Diversity beamforming (with channel coding): compared to

blind beamforming, it can achieve a diversity gain


Ref. J. Choi, Diversity eigenbeamforming for coded signals, IEEE
TCOM, June 2008.

10 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

Coded
1
10
Deigen
Eigen 1
Optimized

2
10
BER

3
10

4
10

4 6 8 10 12 14 16
SNR (dB)

Performance of blind beamforming (Eigen-1), diversity


beamforming (D-eigen and Optimized)

11 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

B. Multiuser MIMO
I There exists interference due to the presence of multiple
signals to be transmitted to multiple users.
I Dirty paper coding (DPC) can achieve the channel capacity
by suppresing known interference. However, its
implementation is not easy.
I Multiuser beamforming can provide a reasonable performance
with low-complexity.
I A better performance can be achieved with multiuser diversity
& user selection
I CSI at BS is required to mitigate the (intra-cell) interference:
I No feedback in TDD: channel reciprocity can be used
I Feedback in FDD: excessive overhead
I Resource allocation (including power control) becomes a
crucial.
12 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

Capacity of multiuser MIMO


I Dirty paper coding: it can achieve the capacity, but difficult
to implement
Csum = E max log det(I + HPHH )
Pk
Ptotal
E log det(I + HHH )(equal power)
K
I Multiuser diversity: it is simple, but overall throughput is low
Cmd = E log(1 + Ptotal max ||hk ||2 )
k
I Multiuser beamforming: if user selection is combined, it its
upper bound is proportional to the number of selected users
(scaling law is applied):
X
Cmb E log(1 + Pk ||hk ||2 )
kUorthogonal

13 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

30
C (equal power)
sum
C
md
Cmbeam (approx.)
25

20

15

10

0
0 2 4 6 8 10 12 14 16 18 20
Achievable Rate

60 users, 4 users are selected in multiuser downlink beamforming

14 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

Conventional Multiuser Beamforming Problem (without user


selection)
Base station

Mobile terminal

Mobile terminal
Mobile terminal

I wk : beamforming vector to user k; hk : channel vector to user


k, Pk = E[|sk |2 ]: TX power to user k multiuser beamforming,
where sk is the signal to user k
PK
I TX signal: k=1 wk sk
15 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

Multiuser joint beamforming and power control:

{Pk , w
P
k } = arg min k Pk
Pq |hH
q wq |
2
(
H 2 2 q
P
subject to k6=q Pk |hq wk | +q
2
||wq || = 1

I This is an optimization problem to minimize the total


transmission power subject to SINR constraints.
I Based on the uplink-downlink duality, this problem has been
solved.

16 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

C. Network (or Cooperative) MIMO


I Highly sophisticated systems to deal with intra-cell as well as
inter-cell interference (ICI)
I Known techniques
I Cooperative transmissions (e.g., CoMP)
I Interference alignment
I Distributed nature
I Backhaul transmissions
I Interference channel models
I There should be the trade-off between performance and
complexity (or overhead)

17 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

TX RX

TX RX

TX RX

18 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

Remarks on beamforming in network MIMO:


I Advantages
I Easy to implement
I Optimal solutions to most beamforming optimization problems
are known (QoS can be guaranteed)
I Spatial multiplexing gain with reasonable performances
I Disadvantages
I Not capacity-achieving schemes
I Most (downlink) beamforming methods require CSI at BS.
I Furthermore, in network MIMO, BSs in cooperation need to
share their CSI.
I Dilemma in FDD mode
I Due to limited CSI feedback, it is prohibitive to use large
antenna arrays.
I However, with arrays of a few antenna elements, a significant
performance improvement may not be achieved.
19 / 64
Precoding and Massive MIMO
2. Overview of Beamforming Techniques

D. Massive MIMO
I TDD mode to exploit the channel reciprocity for the CSI at
BS and make use of large antenna arrays
I Multi-cell, but noncooperative systems
I Due to noncooperative transmissions, no backhaul
communications between BSs are required.
I But, systems can suffer from ICI.
I Using massive antenna arrays, ICI can be mitigated.
I Key issues: pilot contamination, array calibration, etc.
I There are other advantages. One of them is that the
short-term fading disappears by the law of large numbers.
I This makes resource allocation easy and reduces the burden of
channel coding.
20 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO

3. Cooperative (Network) MIMO


I Network MIMO is to fully exploit the spatial gain in a
multi-cell system
I In network MIMO, intercell interference can be mitigated by
cooperation between BSs through backhaul links:
Ref. 1 Karakayali, et al., Network coordination for spectrally efficient
communications in cellular systems, IEEE Comm. Mag.,
2006.
Ref. 2 Gesbert, et al., Multi-cell MIMO cooperative networks: a new
look at interference, IEEE JSAC, 2010.
I Provided that each BS in cooperation is equipped with an
antenna array, cooperative beamforming can mitigate intercell
as well as intracell interference in downlink transmissions.
I Regardless of its implementability, it may provide performance
bounds.
21 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO

BS

BS BS

BSs in cooperation need to share users data and/or CSI to their


users through a backhaul network.
22 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO

Levels of cooperation between BSs through backhaul links:


I full cooperation: BSs share all CSI and signals (this is
equivalent to the case of a big BS with distributed arrays or a
single big cell)
I partial cooperation 1: BSs share all CSI, but not signals
I partial cooperation 2: BSs share all signals, but not CSI
I partial cooperation 3: each BS has its local CSI and signals
(but limited information can be exchanged between BSs for
such as user allocation)
The cooperation would be limited by backhaul overhead. Thus,
effective partial cooperation, not full cooperation, is desirable in
practice.

23 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.1 Multicell with a common group of users

3.1 Multicell with a common group of users


1. All BSs in cooperation support a common group of users
(these users would be cell-edge users).
2. The cooperation between BSs is limited (if there is no
limitation, this scenario is equivalent to a single-cell).
3. Each BS has local CSI, which is the channel vectors from the
BS to all users in a common group, while other BSs CSI is
unknown.
4. There could be some cooperation between BSs such as user
allocation.

24 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.1 Multicell with a common group of users

ZF-DPC precoding in a multicell system


Ref. Ho, et al., Decentralized precoding for multicell MIMO
downlink, IEEE TWC, 2011.
Assumptions:
I There are Q BSs in cooperation and N users per cell (there
are K = QN users in total).
I Each BS is equipped with an antenna array of L elements.
I The CSI from each BS to all K users (not just N in the cell)
is known by each BS, while the CSI from the other BSs to
users is not known.
I That is, at BS q, the CSI to K users are known, but not the
CSI from BS q 0 6= q to K users.
I This partial CSI can allow block diagonalization to suppress
intercell interference
I Within a cell, then DPC is used to suppress intracell
interference
25 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.1 Multicell with a common group of users

Key variables:
Hq : L K downlink channel matrix from BS q to all K users
Gq : L N submatrix of Hq (channel vectors to users in cell q from B
Wq : L N beamforming matrix at BS q
sq : N 1 signal vector at BS q
rq : N 1 signal vector received by users in cell q
nq : N 1 noise vector received by users in cell q
Let
Hq = [Gq Hq ].
The received signal vector at users in cell q is given by
Q
X
rq = GH
q Wp sp + nq
p=1

Block diagonalization to avoid intercell interference:


HH
q Wq = 0.
26 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.1 Multicell with a common group of users

For block diagonalization, orthogonal projection matrix can be


used: 1 H
P H
q = I Hq Hq Hq Hq .
The beamforming matrix can be

Wq = q P
q Gq ,

where q is the normalization factor that is given by



N
q =
||Pq Gq ||F

The average power per beamforming vector is unity.

27 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.1 Multicell with a common group of users

G1
BS1 user 1
W1 H 1

H 2
BS2 user 2
W2 G2

Then, the received signal vector becomes


Q
X
rq = GH
q q P
p Gp sp + nq
p=1

= q GH
q Pq Gq sq + nq .
28 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.1 Multicell with a common group of users

Within rq , there exists intracell interference. Thus, sq has to be


DPCed signals.

The sum rate becomes


Q
X
E log2 det I + |q |2 q H
 
C= q
q=1

where E[nq nH H H
q ] = I, E[sq sq ] = I, and q = Gq Pq Gq .

29 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

3.2 Multicell with different groups of users


I Each BS has its own group of users.
I There are Q BSs and each BS has N users.
I Each BS is equipped with L antennas for beamforming.
I In this case, local CSI at BS is not sufficient to suppress the
intercell interference.
I In order to migitate both intercell and intracell interference by
beamforming, SINR could be considered.

30 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Key variables:
Hq,k : the L N channel matrix from BS k to the users in cell q
Wq : the L N beamforming matrix
sq : the N 1 signal vector from BS q

I Local CSI at BS k is {H1,k , H2,k , . . . , HQ,k }.


I Wq = [w1;q . . . wN ;q ] decides the transmission power if
E[sq sH
q ] = I, i.e.,
Pi;q = ||wi;q ||2 .

31 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

The received signal at cell q is given by


Q
X
rq = HH
q,k Wk sk + nq .
k=1
At user i in cell q, the received signal becomes
Q
X N
X
ri;q = hH
i;q,k wu;k su;k + ni;q
k=1 u=1
X
= hH
i;q,q wi;q si;q + hH
i;q,q wu;q su;q
u6=i
| {z }
=intracell
N
XX
+ hH
i;q,k wu;k su;k +ni;q
k6=q u=1
| {z }
=intercell
32 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

I The SINR becomes


|hH
i;q,q wi;q |
2
SINRi;q =P PN
H 2 H 2 2
P
u6=i |hi;q,q wu;q | + k6=q u=1 |hi;q,k wu;k | + i;q

In general, any beamforming algorithm that minimizes the


SINR requires full CSI, which requires significant backhaul
transmissions.
I Define the intercell interference at cell q as
N
XX
Zi;q = |hH 2
i;q,k wu;k | .
k6=q u=1

Note that this interference is not a function of Wq .

33 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Well-known multiuser beamforming problems:


I Maximization of SINR with power constraints

maxWq SINRi;q
subject to ||Wq ||2F Pq

I Minimization of power with SINR constraints

min q ||Wq ||2F


P
subject to SINRi;q i;q

I See SINR-based Joint Power Control and Beamforming


to solve the above problem

34 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Interference leakage based approaches are also popular:


I Define the interference leakage from BS q to cell k as

Ik (Wq ) = ||HH 2
k,q Wq ||F .

I To minimize the leakage to the other cells, we can have the


following constraint:

X X
Ik (Wq ) = tr WqH Hk,q HHk,q
Wq Iq,leak .
k6=q k6=q

35 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Distributed optimization approaches:


I Forming a global optimization problem
I Decompose it into local optimization problems
I Each local optimization problem is solved at a BS and update
variables through backhaul links
I Do iterations until a satisfactory performance is achieved
I Key references:
I Palomar and Chiang, A tutorial on decomposition methods
for network utility maximization, IEEE JSAC, 2006.
I Chiang, et al., Layering as optimization decomposition: a
mathematical theory of network architectures, Proc. IEEE,
2007.

36 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Example: Dual decomposition


I The minimization of power with SINR constraints:

min Q
P
q=1 Cq (Wq )
subject to SINRi;q (Wq ) i;q

where Cq (Wq ) = ||Wq ||2F .


I Using Lagrangian multipliers,
Q
X Q X
X N
min Cq (Wq ) i;q (SINRi;q (Wq ) i;q )
q=1 q=1 i=1
Q N
!
X X
= min Cq (Wq ) i;q (SINRi;q (Wq ) i;q )
Wq
q=1 i=1
| {z }
=local optimization
37 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

I At each BS, the local optimiztion is carried out as follows:


N
q = arg min Cq (Wq )
X
W i;q SINRi;q (Wq )
Wq
i=1

I At a higher level (i.e., at a central unit), the master dual


problem is to update s:
Q
X N
X
min gq () i;q i;q , 0,

q=1 i=1

where
N
q ) Cq ( W
q ).
X
gq () = i;q SINRi;q (W
i=1

38 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

I To update s at a higher level, the following gradient method


can be used:
h  i+
i;q (t + 1) = i;q (t) SINRi;q (W q ) i;q ,

where [x]+ = max{0, x} and > 0 is the step-size.


I Thus, BSs need to send their SINRs to a central unit and this
central unit sends s to BSs.
I This approach would have a slow convergence rate, but it
does not require excessive backhaul transmissions between
BSs and a central unit.

39 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Example Minimization interference leakage problem with SINR

min Q
P PQ
q=1 k=1 Ik (Wq )
subject to SINRi;q (Wq ) i;q

A direct decomposition can be considered. At each BS,

min i wiH Xwi


P
|hH
i wi |
2
subject to P
|hH w |2 +Z i
q6=i i q i

X k6=q Hk,q HH
P
I
k,q
I wi wi;q
I hi hi;q
I 2
Zi Zi;q + i;q
I i i;q
40 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Since the ICI Zi;q (Wq ) in the SINR depends on the other
beamforming matrices, an iterative algorithm is required to solve
the NML beamforming problem, which consists of the following
two key steps.
I Solving the local minimization problem (at BS q):
(`)
Wq = arg min Cq (Wq )
|hH
i;q,q wi;q |
2
subject to P H 2 (`1) 2
i;q
u6=i |hi;q,q wu;q | +Zi;q (Wq )+i;q

I updating the ICI:


N
(`) (`)
XX
Zi;q (Wq ) = |hH 2
i;q,k wu;k | ,
k6=q u=1

which is fed back by the users (not by other BSs) to BS q.


41 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Low-Dimensional Beamforming
I Consider the following eigendecomposition:
X
X= Hk,q HH H
k,q + I = EE ,
k6=q

where E = [e1 . . . eL ] and = diag(1 , . . . , L ).


I The local cost function becomes
X
Ik (W) = tr WH XW

Cq (W) =
k6=q

I The complexity can be reduced if a subspace of E used as



i = E
wi w 1/2 vi ,

= [e1 . . . eM ] and M < L.


where E
42 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

I With vi of size M 1, we have


N
X
Cq (W) ||vi ||2
i=1
I The local optimization at BS q is carried out to minimize
Cq (W) through vi of size M 1.
I The subspace beamforming with optimized vi (not wi ) can
L 2

reduce the complexity by a factor of M .
I For highly correlated channels, this approximation may not
result in a significant loss.
I The complexity of the eigendecomposition of X may not be
significant if second order statistics are used:
X
X E[X] = E[Hk,q HH H
k,q ] + I = EE .
k6=q

43 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Simulation Results
I Three different approaches are used:
I Approach A: The cooperative min-power beamforming with
target SINR
I Approach B: The NML beamforming with target SINR
I Approach C: The noncooprative minimum total power (NMP)
beamforming with the following cost function:
Cq (Wq ) = ||Wq ||2F ,
which is to reduce the transmission power rather than
interference leakage (an egoistic approach).
I Approach A is cooperative, while Approaches B and C are
noncooperative (i.e., no backhaul communications between
BSs)
I The elements of channel matrices are independent Gaussian
(no spatial correlation is considered)
44 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Feasibility results: The NMP beamforming (Approach C) cannot


achieve target SINR, but both cooperative min-power beamforming
and NML beamforming (Approaches A and B) can achieve target
SINR
20
Cooperative
Noncoop. (minpower)
18 Noncoop. (minleakage)

16

14
Achieved SINR (dB)

12

10

2
4 6 8 10 12 14 16 18 20
Target SINR (dB)

L = 30, Q = 3, N = 5, and M = L = 30
45 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

Approach A performs better than Approaches B and C, but


requires backhaul communications between BSs
80
Cooperative
Noncoop. (minpower)
Noncoop. (minleakage)
70

60
Total transmission power (dB)

50

40

30

20

10
4 6 8 10 12 14 16 18 20
Target SINR (dB)

L = 30, Q = 3, N = 5, and M = L = 30

46 / 64
Precoding and Massive MIMO
3. Cooperative (Network) MIMO
3.2 Multicell with different groups of users

I Low-complexity versions of Approach B (NML): tradeoff


between complexity and performance
I If M 20, no performance degradation is observed (note that
uncorrelated channels are used in simulations)
36

34
Total transmission power (dB)

32

30

28

26

24
5 10 15 20 25 30
M

L = 30, Q = 3, N = 5, and = 10 dB
47 / 64
Precoding and Massive MIMO
4. Massive MIMO

4. Massive MIMO
I Proposed by Marzetta in 2010 to effectively mitigate ICI
without any cooperation between BSs
I A massive MIMO system consists of BSs with large antenna
arrays.
I The number of antenna elements is about 100.
I Frequency reuse factor is 1 and orthogonal channels within a
cell (no intra-cell interference)
I TDD to exploit the channel reciprocity.
I The channel vector, h, from a BS to a user can be factorized
as p
h = u,
where is a parameter for large-scale fading and
u CN 0, L1 I is a random vector for small-scale fading,


where L  1.
48 / 64
Precoding and Massive MIMO
4. Massive MIMO

Base Station

Large Array

narrow beam

user user

In each cell, each user has an orthogonal channel (i.e., no intra-cell


interference) and hk,q denotes the channel vector from BS q to the
user in cell k.
49 / 64
Precoding and Massive MIMO
4. Massive MIMO

For a large L, the matched-filter (MF) beamforming can be used.


The beamforming vector is given by
hk,k
wk = .
||hk,k ||

Since the received signal at the user in cell k is


X
rk = hHk,k wk sk + hHk,q wq sq + nk ,
q6=k

the SINR becomes


Pk ||hk,k ||2
SINRk = P H 2 2
.
q6=k Pq |hk,q wq | + k

50 / 64
Precoding and Massive MIMO
4. Massive MIMO

ICI can be mitigated as L as the inner product of two


random vectors, x = hk,k and y = hk,q , k 6= q, approaches 0.
0.25

0.2

0.15
|<x,y>|2

0.1

0.05

0
0 10 20 30 40 50 60 70 80 90 100
L: number of antennas

This is the key idea of massive MIMO. That is, without any
cooperation between BSs, it is possible to mitigate ICI by
increasing L (similar to CDMA). 51 / 64
Precoding and Massive MIMO
4. Massive MIMO

Approximation for a large L:


I ||hk,q ||2 k,q
I uH
k,q wq CN (0, 1/L)

X X 1X
Pq |hH 2 2
k,q wq | +k = Pq k |uH 2 2
k,q wq | +k Pq k,q +k2 .
L
q6=k q6=k q6=k

I The approximate SINR is


LPk k,k
SINRk = P 2
q6=k Pq k,q + Lk

I In SINR, small-scale fading terms disappear the SINR is


much less fluctuated over time.

52 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.1 Partial cooperation

4.1 Partial cooperation


I Since the frequency reuse factor is 1, there exists ICI although
it may not be significant for a large L (as shown in the SINR
expression).
I ICI can be further mitigated if ZF or MMSE beamforming can
be used in cooperation with adjacent BSs as in network
MIMO .
I However, it requires massive CSI exchange, which would
provide a marginal gain at the expense of excessive backhaul
communications.
I There might be a tradeoff between ICI mitigation (high ICI for
a small L) performance and the backhaul overhead (a large
overhead for a large L) in the case of cooperation.

53 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.1 Partial cooperation

I In order to improve the performance further, partial


cooperation can be considered.
I Not exchange the CSI of fast-fading, but the CSI of
slow-fading, which is k for large-scale fading.
I Then, the joint power control is carried out to minimize the
total power of BSs in cooperation with SINR constraints.
I This partial cooperation can guarantee target SINR in massive
MIMO.

54 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.1 Partial cooperation

16

14

12
average/minimum SINR (dB)

10

Average SINR (equal power)


2
Average SINR (opt. power)
Minimum SINR (equal power)
Minimum SINR (opt. power)
0
0 2 4 6 8 10 12 14 16
target SINR (dB)

Arrays of L = 100 elements, 3-cell, 4 users per cell

55 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.2 Pilot contamination in massive MIMO

4.2 Pilot contamination in massive MIMO


I A main issue of massive MIMO is pilot contamination.
I As the coherence time is limited and the number of users per
cell can be large, the same set of orthogonal pilot sequences
can be used in all cells, which results in interference from
adjacent cell during the uplink training.
I The estimated uplink channel can be contaminated by the
interference and the performance of beamforming can be
degraded.
I This problem is called the pilot contamination. To mitigate
this problem, we can
I take into account interfering pilot signals from adjacent cells in
estimating uplink channels;
I adjust precoding vectors to reduce the interference.

56 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.2 Pilot contamination in massive MIMO

Two-cell example (TDD: channel reciprocity)

I Assume that MRT precoding vectors are used from estimated


channels.
57 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.3 Pilot contamination precoding

4.3 Pilot contamination precoding


I For a large L, statistical precoding can be used to mitigate
the intercell interference resulting from pilot contamination.
I This scheme is called the pilot contamination precoding
(PCP).
I Assume that there are K cells. The estimated channel vector
at the BS in cell k is
K
k =
X
h hq,k + nk .
q=1

I The MF beamforming vector is used as:


k
h
wk = .
k ||
||h
58 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.3 Pilot contamination precoding

I Received signal at the user in cell k:


K
X
xk = hH
k,q wq sq + vk
q=1
K K
X sq X H
= hk,q ht,q + vk ,
q

q=1 t=1

where k ||.
k = ||p
h
I Since hk,q = k,q uk,q and uk,q are iid, we have
uH
k,q ut,q k,q as L w.p. 1
I Thus, as L , we have
K
X sq
xk = k,q + vk .
q

q=1

59 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.3 Pilot contamination precoding

I Staking xk s:
x = [x1 . . . xK ]T
 
k,q
= [s1 . . . sK ]T + [v1 . . . vK ]T
q
= As + v,

where [A]k,q = k,q
q
.
I To avoid the interference, s can be replaced with
s Bs,
where B = A1 . This means that the BSs need to exchange
the channel information (elements in A) for precoding.
I The resulting received signal is
x = s + v.
60 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.3 Pilot contamination precoding

I Since each BS needs to estimate k,q and q ||, which


q = ||h
are related to slow fading coefficients only, their estimation
can be done precisely, and their exchange through a backhaul
network may not result in a heavy overhead.
I Note that the transmission power depends on B = A1 .
I DPC or VP technique can help reduce the total transmission
power when precoding is used.

61 / 64
Precoding and Massive MIMO
4. Massive MIMO
4.3 Pilot contamination precoding

Improvement by partial cooperation against pilot contamination


1.5 0.8

0.6
1

0.4

0.5
0.2

0 0

-0.2
-0.5

-0.4

-1
-0.6

-1.5 -0.8
-1.5 -1 -0.5 0 0.5 1 1.5 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8

without PCP with PCP

62 / 64
Precoding and Massive MIMO
5. Conclusions

5. Conclusions
I Beamforming is a simple and effective means to increase SNR
or coverage, which can be employed in more complicated
systems as follows:
I Network MIMO
I Cooperation between BSs with not too big arrays
I More controls (for better performance)
I Backhaul overhead for cooperation might be a critical issue
I Distributed optimization will play a key role in reducing
backhaul overhead
I Massive MIMO
I Noncooperation, each BS with a big array might be robust
enough against ICI
I Less controls (for implementation)
I There exists ICI, and due to it, there is difficulty to guarantee
certain performance in terms of SINR.
I As a remedy, partial cooperation could be considered.
63 / 64
Precoding and Massive MIMO
5. Conclusions

I Network MIMO versus Massive MIMO:


I Network MIMO is to reduce cooperation (full to partial
cooperation)
I Massive MIMO is to introduce cooperation (no cooperation to
partial cooperation)
I In the end, network MIMO meets massive MIMO.

64 / 64

Potrebbero piacerti anche