Advanced Wireless Communications

Bruno Clerckx

Department of Electrical and Electronic Engineeing, Imperial College London

School of Electrical Engineering, Korea University

August 2014

1 / 494
Course Ojectives

• Advanced course on wireless communication and communication theory

– Provides the fundamentals of wireless communications from a 4G and beyond
perspective
– At the cross-road between information theory, coding theory, signal processing and
antenna/propagation theory
• Major focus of the course is on MIMO (Multiple Input Multiple Output) and
multi-user/multi-cell communications
– Includes as special cases SISO (Single Input Single Output), MISO (Multiple Input
Single Output), SIMO (Single Input Multiple Output)
– Applications: everywhere in wireless communication networks: 3G, 4G(LTE,LTE-A),
(5G?), WiMAX(IEEE 802.16e, IEEE 802.16m), WiFi(IEEE 802.11n), satellite,...+ in
other fields, e.g. radar, medical devices, speech and sound processing, ...
• Valuable for those who want to either pursue a PhD in communication or a career in
a high-tech telecom company (research centres, R&D branches of telecom
manufacturers and operators,...).
• Skills
– Mathematical modelling and analysis of (MIMO-based) wireless communication
systems
– Design (transmitters and receivers) of multi-cell multi-user MIMO wireless
communication systems
– Practical understanding of MIMO applications and performance evaluations
2 / 494
 Content

Central question: How to deal with fading and interference in wireless networks?

• Some fundamentals/revision (matrix analysis, probability, information theory)

• Single link: point to point communications
– Fading and Diversity
– MIMO Channels - Modelling and Propagation
– Capacity of point-to-point MIMO Channels
– Space-Time Coding/Decoding over I.I.D. Rayleigh Flat Fading Channels
– Space-Time Coding in Real-World MIMO Channels
– Partial Channel State Information at the Transmitter (CSIT)
– Frequency-Selective MIMO Channels - MIMO-OFDM

• Multiple links: multiuser communications

– Multi-User MIMO - Capacity of Multiple Access Channels (Uplink)
– Multi-User MIMO - Capacity of Broadcast Channels (Downlink)
– Multi-User MIMO - Scheduling, Linear and Non-Linear Precoding, DPC (Downlink)
– Multi-User MIMO (Downlink) with/for Imperfect CSIT

• Multiple cells: multiuser multicell communications

– Introduction to Multi-Cell MIMO
– Capacity of Interference Channel
– Coordinated Scheduling and Power Control
– Coordinated Beamforming and Interference Alignment
– Network MIMO
3 / 494
Content

• Massive MIMO
• Real-World MIMO Wireless Networks
– MIMO and Interference Management in 4G and beyond (LTE, LTE-Advanced,
WiMAX)
– Homogeneous and Heterogeneous Networks
– System-Level Performance Evaluations

• Additional topics (if time permits)

– Modeling of Wireless Networks: Stochastic Geometry
– Wireless Communication and Power Networks - Energy Harvesting

4 / 494
Important Information

• Course webpage: http://www.ee.ic.ac.uk/bruno.clerckx/Teaching.html

• Prerequisite: Good background on Communication Theory

• Lectures
– Week 1: Wednesday, Thursday, Friday from 09.30 till 12.15
– Week 2: Monday, Tuesday, Wednesday, Thursday from 09.30 till 12.15
– Week 3: Monday, Tuesday, Wednesday, Thursday from 09.30 till 12.15
– Week 4: Monday, Tuesday, Wednesday, Thursday from 09.30 till 12.15

• Problem sheets from Imperial College EE4-65/EE9-SO27 Wireless Communications

course available on course webpage (2 types: 1. paper/pencil, 2. matlab)
• Matlab project also available on course webpage
– Encourage students to work on it if time permits.

5 / 494
Important Information

• Reference book

Bruno Clerckx and Claude Oestges,

“MIMO Wireless Networks: Channels,
Techniques and Standards for Multi-
Antenna, Multi-User and Multi-Cell
Systems,” Academic Press (Elsevier),
Oxford, UK, Jan 2013.

• Another interesting reference on wireless communications (more introductory)

“Fundamentals of Wireless Communication,” by D. Tse and P. Viswanath,
Cambridge University Press, May 2005

6 / 494
Some fundamentals/revisions (matrix analysis,
probability, information theory)

7 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Appendix A, B

• T. Cover and J. Thomas, “Elements of Information Theory,” Second Edition, Wiley,

2006.
8 / 494
Matrix properties

• Vector Orthogonality : aH b = 0 (H stands for Hermitian, i.e. conjugate transpose)

• Hermitian matrix: A = AH
• Unitary matrix: AH A = I
• Singular Value Decomposition (SVD) of a matrix H [nr × nt ]: H = UΣVH
– U [nr × r(H)]: unitary matrix of left singular vectors
– Σ = diag{σ1 , σ2 , . . . , σr(H) }: diagonal matrix containing the singular values of H
– V [nt × r(H)]: unitary matrix of left singular vectors
– r(H): the rank of H
We will often look at Hermitian matrices of the form A = HH H whose Eigenvalue
H 2
Value Decomposition (EVD) writesP as A = VΛV with Λ = Σ .
• Trace of a matrix A: Tr {A} = i A(i, i).P P
• Frobenius norm of a matrix A: kAk2F = i j |A(i, j)|2
 
• kAk2F = Tr AAH = Tr AH A
• Tr {AB} = Tr {BA} Q
• Hadamard’s inequality : det (A) ≤ nk=1 A (k, k) if A > 0 of size n × n

9 / 494
Matrix properties
 
A(1, 1)B . . . A(1, n)B
 .. .. 
• Kronecker product:A ⊗ B =  . ... ., 
A(m, 1)B . . . A(m, n)B
• (A ⊗ B) ⊗ C = A ⊗ (B ⊗ C)
• (A ⊗ B)H = AH ⊗ BH
• (A ⊗ B) (C ⊗ D) = (AC ⊗ BD)
• (A ⊗ B)−1 = A−1 ⊗ B−1 if A, B square and non singular.
• det (Am×m ⊗ Bn×n ) = det (A)n det (B)m
• Tr {A ⊗ B} = Tr {A} Tr {B}
2
• Tr {AB} ≥ Tr {A} σmin (B) with σmin (B) the smallest singular value of B
• vec (A) converts [m × n] matrix into mn × 1 vector by stacking the columns of A
on top of one another.


– vec (ABC) = CT ⊗ A vec (B)
• det (I + ǫA) = 1 + ǫTr {A} if ǫ << 1

10 / 494
Gaussian random variable

• Real Gaussian random variable x with mean µ = E {x} and variance σ 2

 
1 (x − µ)2
p (x) = √ exp − .
2πσ 2 2σ 2

Standard Gaussian random variable: µ = 0 and σ 2 = 1

• Real Gaussian random vector
n x of dimensionon with mean vector µ = E {x} and
covariance matrix R = E (x − µ) (x − µ)T :
 
1 (x − µ)T R−1 (x − µ)
p (x) = √
n p exp − .
2π det (R) 2

Standard Gaussian random vector x of dimension n: entries are independent and

identically distributed (i.i.d.) standard Gaussian random variables x1 , . . . , xn
 
1 kxk2
p (x) = √
n exp − .
2π 2

11 / 494
Gaussian random variable

• Complex Gaussian random variable x = xr + jxi : [xr , xi ]T is a real Gaussian

random vector.
• Important case: x = xr + jxi is such that its real and imaginary parts are i.i.d. zero
mean Gaussian variables of variance σ 2 (circularly symmetric complex Gaussian
random variable).
p
• s = |x| = x2r + x2i is Rayleigh distributed
 
s s2
p(s) = 2 exp − 2 .
σ 2σ
• y = s2 = |x|2 = x2r + x2i is χ22 (i.e. exponentially) distributed (with two degrees of
freedom)
1  y 
py (y) = 2
exp − 2 .
2σ 2σ
Hence, µ = E {y} = 2σ 2 .
• More generally, χ2n is the sum of the square of n i.i.d. zero-mean Gaussian random
variables.
• Assume n i.i.d. zero mean complex GaussianP variables h1 , . . . , hn (real and imaginary
parts with variance σ 2 ). Defining u = n 2
k=1 |hk | , the MGF of u is given by
 n
τu 1
Mu (τ ) = E{e } = ,
1 − 2σ 2 τ
12 / 494
Discrete Memoryless Channel

Definition
A discrete channel is defined as a system consisting of an input alphabet X and
output alphabet Y and a probability transition matrix p(y|x) that expresses the
probability of observing the output symbols y given that the symbold x is sent.

Definition
The channel is memoryless if the probability distribution of the output depends
only on the input at that time and is conditionally independent of previous
channel inputs or outputs, i.e. if x1 ,...,xn are inputs, and y1 ,...,yn denote the
corresponding outputs, for n channel uses, then

p(y1 , ..., yn |x1 , ...xn ) = p(y1 |x1 )...p(yn |xn )

Example
Binary Symmetric Channel (BSC): x and y take values in 0,1 such that

1 − p, y = x,
p(y|x) =
p, y = 1 − x.

13 / 494
Entropy

• Entropy is a measure of the average uncertainty of a random variable

Definition
For a discrete random variable X, the entropy H(X) is defined as
 
1 X
H(X) = E = −E {log2 p(X)} = − p(x) log2 p(x),
log2 p(X) x

where p(x) is the probability mass function of X.

• It is the number of bits on average required to describe the random variable.

Example
Let X be a Bernoulli random variable

1, with probability p,
X=
0, with probability 1 − p.

Then H(X) = −p log2 p − (1 − p) log2 (1 − p). For p = 0, 1, there is no

uncertainty on the value of the RV, so no information gained. For p = 1/2,
H(X) (uncertainty/information) is maximized.
14 / 494
Entropy

Lemma
H(X) ≥ 0
1
Proof: 0 ≤ p(x) ≤ 1 such that log2 p(x)
≥0 

Definition
The joint entropy H(X, Y ) of a pair of discrete random variables X and Y
with a joint pmf p(x, y) is defined as
XX
H(X, Y ) = −E {log2 p(X, Y )} = − p(x, y) log2 p(x, y)
x y

15 / 494
Conditional Entropy

• The conditional entropy of a random variable given another is the expected value of
the entropies of the conditional distributions, averaged over the conditioning random
variable
Definition
The conditional entropy H(Y |X) is defined as
X
H(Y |X) = p(x)H(Y |X = x)
x
X X
=− p(x) p(y|x) log2 p(y|x)
x y
XX
=− p(x, y) log2 p(y|x)
x y

= −E {log2 p(Y |X)}

16 / 494
Joint Entropy

Theorem
Chain rule
H(X, Y ) = H(X) + H(Y |X)

Proof:
XX XX
H(X, Y ) = − p(x, y) log2 p(x, y) = − p(x, y) log2 p(x)p(y|x)
x y x y
XX XX
=− p(x, y) log2 p(x) − p(x, y) log2 p(y|x)
x y x y
X XX
=− p(x) log2 p(x) − p(x, y) log2 p(y|x)
x x y

= H(X) + H(Y |X)

Alternatively,
log2 p(X, Y ) = log2 p(X) + log2 p(Y |X)
E {log2 p(X, Y )} = E {log2 p(X)} + E {log2 p(Y |X)}

17 / 494
Relative Entropy

• The relative entropy is a measure of the distance between two distributions.

Definition
The relative entropy between two pmf p(x) and q(x) is defined as
 
X p(x) p(X)
D(p||q) = p(x) log2 = Ep log2
x
q(x) q(X)

Theorem
The relative entropy is always nonnegative D(p||q) ≥ 0 and is zero if and only
if p = q.

18 / 494
Mutual Information

• The mutual information is a measure of the amount of information that one RV

contains about another RV. It is a measure of the dependence between the two RVs.

Definition
For a pair of discrete random variables X and Y with a joint pmf p(x, y) and
marginal pmf p(x) and p(y), the mutual information I(X; Y ) is the relative
entropy between p(x, y) and p(x)p(y)
 
p(X, Y )
I(X; Y ) = D(p(x, y)||p(x)p(y)) = Ep(x,y) log2
p(X)p(Y )
XX p(x, y)
= p(x, y) log2
x y
p(x)p(y)

19 / 494
Mutual Information

• The mutual information I(X; Y ) is the reduction in the uncertainty of one random
variable due to the knowledge of the other
X p(x, y)
I(X; Y ) = p(x, y) log2
x,y
p(x)p(y)
X p(x|y)
= p(x, y) log2
x,y
p(x)
X X
=− p(x, y) log2 p(x) + p(x, y) log2 p(x|y)
x,y x,y
!
X X
=− p(x) log2 p(x) − − p(x, y) log2 p(x|y)
x x,y

= H(X) − H(X|Y )
= H(Y ) − H(Y |X) = I(Y ; X)

• I(X; Y ) = H(X) + H(Y ) − H(X, Y ).

• I(X; X) = H(X) − H(X|X) = H(X)

20 / 494
Mutual Information

Theorem
Nonnegativity of mutual information: For any two random variables X,Y

I(X; Y ) ≥ 0

with equality if and only if X and Y are independent

Theorem
Conditioning reduces entropy: For any two random variables X,Y

H(X|Y ) ≤ H(X)

with equality if and only if X and Y are independent

Proof: 0 ≤ I(X; Y ) = H(X) − H(X|Y ) 

Knowing another RV Y can only reduce on the average the uncertainty in X.

21 / 494
Channel Coding Theorem

Theorem
(a) For a DMC with channel transition pmf p(y|x), we can use i.i.d. inputs
with pmf p(x) to communicate reliably, as long as the code rate satisfies

R < I(X; Y ).

(b) The achievable rate can be maximized over the input density p(x) to obtain
the channel capacity
C = max I(X; Y ).
p(x)

22 / 494
Differential Entropy

Definition
For a continuous random variable X, the differential entropy h(X) is defined as
  Z
1
h(X) = E = −E {log2 p(x)} = − p(x) log2 p(x)dx,
log2 p(x)

where p(x) is the probability density function of X.

Caution: h(X) can be negative.

Example
2
For X ∼ N (µ, σ 2 ), − log2 p(x) = (x−µ)
2σ 2
log2 (e) + 21 log2 (2πσ 2 ). Thus,
h(X) = −E {log2 p(x)} = 2 log2 (e) + 2 log2 (2πσ 2 ) = 21 log2 (2πeσ 2 ). The
1 1

mean does not affect the differential entropy.

Theorem
Consider a RV with zero mean and variance σ 2 . Then h(X) ≤ 1
2
log2 (2πeσ 2 ),
with equality iff X ∼ N (0, σ 2 ).
23 / 494
AWGN Channel

• Real discrete-time AWGN channel

Y = X + N, N ∼ N (0, σ 2 )

where X is power-constrained input E X 2 ≤ Es

• The channel transition density is given by

 
1 (y − x)2
p(y|x) = √ exp −
2πσ 2 2σ 2

24 / 494
AWGN Channel Capacity

Theorem
The capacity of the real AWGN channel is
1 Es
C= max I(X; Y ) = log2 (1 + 2 ).
p(x):E {X }≤Es
2 2 σ

Proof: Consider Y = X + N , with N ∼ N (0, σ 2 ) and E X 2 ≤ Es . Given X = x,
h(Y |X = x) = h(N ), so that h(Y |X) = h(N ) and
I(X; Y ) = h(Y ) − h(Y |X) = h(Y ) − h(N ).
Maximizing
 I(X;
 Y ) comes
 to maximize h(Y ). Since X and N are independent,
E Y 2 = E X 2 + E N 2 ≤ Es + σ 2 . We now know that
1
h(Y ) ≤ log2 (2πe(Es + σ 2 ))
2
and equality is achieved iff Y ∼ N (0, Es + σ 2 ). Y ∼ N (0, Es + σ 2 ) is achieved if
the input distribution is X ∼ N (0, Es ), independent of the noise. We then get
1 1 1 Es
I(X; Y ) = h(Y ) − h(Z) = log2 (2πe(Es + σ 2 )) − log2 (2πeσ 2 ) = log2 (1 + 2 ).
2 2 2 σ

25 / 494
Jensen’s inequality

Theorem
If f is a convex function and X is a random variable,

E {f (X)} ≥ f (E {X}).

26 / 494
Fading and Diversity

27 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 1
Section: 1.2, 1.3, 1.4, 1.5
Appendix A, B

28 / 494
Space-Time Wireless Channels:Discrete Time
Representation
• channel: the impulse response of the linear time-varying communication system
between one (or more) transmitter(s) and one (or more) receiver(s).
• Assume a SISO transmission where the digital signal is defined in discrete-time by
the complex time series {cl }l∈ and is transmitted at the symbol rate Ts .
• The transmitted signal is then represented by
∞
X √
c(t) = Es cl δ(t − lTs ),
l=−∞

where Es is the transmitted symbol energy, assuming that the average energy
constellation is normalized to unity.
• Define a function hB (t, τ ) as the time-varying (along variable t) impulse response of
the channel (along τ ) over the system bandwidth B = 1/Ts , i.e. hB (t, τ ) is the
response at time t to an impulse at time t − τ .
• The received signal y(t) is given by
y(t) = hB (t, τ ) ⋆ c(t) + n(t)
Z τmax
= hB (t, τ )c(t − τ )dτ + n(t)
0
where ⋆ denotes the convolution product, n(t) is the additive noise of the system and
τmax is the maximal length of the impulse response.
29 / 494
Discrete Time Representation

• hB is a scalar quantity, which can be further decomposed into three main terms,

hB (t, τ ) = fr ⋆ h(t, τ ) ⋆ ft ,
where
– ft is the pulse-shaping filter,
– h(t, τ ) is the electromagnetic propagation channel (including the transmit and receive
antennas) at time t,
– fr is the receive filter.
• Nyquist criterion: the cascade f = fr ⋆ ft does not create inter-symbol interference
when y(t) is sampled at rate Ts .
• In practice,
– difficult to model h(t, τ ) (infinite bandwidth is required).
– hB (t, τ ) is usually the modeled quantity, but is written as h(t, τ ) (abuse of notation).
– Same notational approximation: the channel impulse response writes as h(t, τ ) or ht [τ ].
• The input-output relationship reads thereby as
∞
X √
y(t) = h(t, τ ) ⋆ c(t) + n(t) = Es cl ht [t − lTs ] + n(t).
l=−∞

30 / 494
Discrete Time Representation

• Sampling the received signal at the symbol rate Ts (yk = y(t0 + kTs ), using the
epoch t0 ) yields
∞
X √
yk = Es cl ht0 +kTs [t0 + (k − l)Ts ] + n(t0 + kTs )
l=−∞
∞
X √
= Es cl hk [k − l] + nk
l=−∞

Example
At time k = 0, the channel has two taps: h0 [0], h0 [1]
√
y0 = Es [c0 h0 [0] + c−1 h0 [1]] + n0
• If Ts >> τmax ,
– hB (t, τ ) is modeled by a single dependence on t: write simply as hB (t) (or h(t) using
the same abuse of notation). In the sampled domain, hk = h(t0 + kTs ).
– the channel is then said to be flat fading or narrowband
p
y k = E s hk ck + nk
• Otherwise the channel is said to be frequency selective.
31 / 494
Path-Loss and Shadowing

• Assuming narrowband channels and given specific Tx and Rx locations, hk is

modeled as
1
hk = √ hk ,
Λ0 S
where
– path-loss Λ0 : a real-valued deterministic attenuation term modeled as Λ0 ∝ Rη where
η designates the path-loss exponent and R the distance between Tx and Rx.
– shadowing S: a real-valued random additional attenuation term, which, for a given
range, depends on the specific location of the transmitter and the receiver and modeled
as a lognormal variable, i.e., 10 log10 (S) is a zero-mean normal variable of given
standard deviation σS .
– fading hk: caused
by the combination of non coherent multipaths. By definition of Λ0
and S, E |h|2 = 1.

• Alternatively, hk = Λ−1/2 hk with Λ modeled on a logarithm scale

 
R
Λ|dB = Λ0 |dB + S|dB = L0 |dB + 10η log10 + S|dB ,
R0

where |dB indicates the conversion to dB, and L0 is the deterministic path-loss at a
reference distance R0 , and Λ is generally known as the path-loss.
32 / 494
Path-Loss and Shadowing

• Path loss models are identical for both single- and multi-antenna systems.

• For point to point systems, it is common to discard the path loss and shadowing and
only investigate the effect due to fading, i.e. the classical model for narrowband
channels √
y = Es hc + n,
where the time index isremoved
for better legibility and n is usually taken as white
Gaussian distributed, E nk n∗l = σn2 δ(k − l).

• Es can then be seen as an average received symbol energy. The average SNR is then
defined as ρ , Es /σn2 .

33 / 494
Fading

• Multipaths
receiver
diffusion

line-of-sight diffraction

specular reflection

transmitter

• Assuming that the signal reaches the receiver via a large number of paths of similar
energy,
– h is modeled such that its real and imaginary parts are i.i.d. zero mean Gaussian
variablesof variance
σ 2 (circularly symmetric complex Gaussian variable).
– Recall E |h|2 = 2σ 2 = 1.

34 / 494
Fading

• The channel amplitude s , |h| follows a Rayleigh distribution,

 
s s2
ps (s) = 2 exp − 2 ,
σ 2σ
whose first two moments are
r
π
E{s} = σ
2

E{s2 } = 2σ 2 = E |h|2 = 1.

• The phase of h is uniformly distributed over [0, 2π)

35 / 494
Fading

• Illustration of the typical received signal strength of a Rayleigh fading channel over a
certain time interval
10

0
Received signal [dB]

−5

−10

−15

−20

−25

−30
0 1 2 3 4 5
Time [s]

– The signal level randomly fluctuates, with some sharp declines of power and
instantaneous received SNR known as fades.
– When the channel is in a deep fade, a reliable decoding of the transmitted signal may
not be possible anymore, resulting in an error.
– How to recover the signal? Use of diversity techniques
36 / 494
Maximum likelihood detection

• Decision rule: choose the hypothesis that maximizes the conditional density

arg max p(y|x) = arg max log p(y|x)

x x

• If real AWGN y = x + n with n ∼ N (0, σn2 ),

 
1 (y − x)2
p(y|x) = √ exp −
2πσn2 2σn2
and
arg max p(y|x) = arg max (y − x)2
x x
√
• If y = Es hc + n, the ML decision rule becomes

√ 2

arg max y − Es hc
c

37 / 494
Diversity in Multiple Antennas Wireless Systems

• What is the impact of fading on system performance?

• Consider the simple case of BPSK transmission through an AWGN channel and a
SISO Rayleigh fading channel:
– In the absence of fading (h = 1), the symbol-error rate (SER) in an additive white
Gaussian noise (AWGN) channel is given by
s !
2Es p

P̄ = Q 2
=Q 2ρ ,
σn
where Q (x) is the Gaussian Q-function defined as
Z ∞  
∆ 1 y2
Q (x) = P (y ≥ x) = √ exp − dy.
2π x 2
√
– In the presence of (Rayleigh) fading, the received signal level fluctuates as s Es , and
the SNR varies as ρs2 . As a result, the SER
Z ∞ p

P̄ = Q 2ρs ps (s) ds
0
r
1 ρ 
= 1−
2 1+ρ
(ρր) 1
∼
=
4ρ
R∞
although the average SNR ρ̄ = 0 ρs2 ps (s) ds remains equal to ρ.
38 / 494
Diversity in Multiple Antennas Wireless Systems

• How to combat the impact of fading? Use diversity techniques

• The principle of diversity is to provide the receiver with multiple versions (called
diversity branch) of the same transmitted signal.
– Independent fading conditions across branches needed.
– Diversity stabilizes the link through channel hardening which leads to better error rate.
– Multiple domains: time (coding and interleaving), frequency (equalization and
multi-carrier modulations) and space (multiple antennas/polarizations).
• Array Gain: increase in average output SNR (i.e., at the input of the detector)
relative to the single-branch average SNR ρ
ρ̄out ρ̄out
ga , =
ρ̄ ρ
• Diversity Gain: increase in the error rate slope as a function of the SNR. Defined as
the negative slope of the log-log plot of the average error probability P̄ versus SNR


o log2 P̄
gd (ρ) , − .
log2 (ρ)
The diversity gain is commonly taken as the asymptotic slope, i.e., for ρ → ∞.

39 / 494
Diversity in Multiple Antennas Wireless Systems

• Illustration of diversity and array gains

array gain = SNR shift Careful that the curves have been
plotted against the single-branch
Error probability

diversity gain average SNR ρ̄ = ρ !

= slope increase
If plotted against the output aver-
age SNR ρ̄out , the array gain dis-
appears.
AWGN
Rayleigh fading, no spatial diversity
Rayleigh fading with diversity

SNR ρ [dB]

• Coding Gain: a shift of the error curve (error rate vs. SNR) to the left, similarly to
the array gain.
– If the error rate vs. the average receive SNR ρ̄out , any variation of the array gain is
invisible but any variation of the coding gain is visible: for a given SNR level ρ̄out at
the input of the detector, the error rates will differ.
40 / 494
SIMO Systems

• Receive diversity may be implemented via two rather different combining methods:
– selection combining : the combiner selects the branch with the highest SNR among the
nr receive signals, which is then used for detection,
– gain combining : the signal used for detection is a linear combination of all branches,
z = gy, where g = [g1 , . . . , gnr ] is the combining vector.
1 Equal Gain Combining
2 Maximal Ratio Combining
3 Minimum Mean Square Error Combining

• Space antennas sufficiently far apart from each other so as to experience

independent fading on each branch.

• We assume that the receiver is able to acquire the perfect knowledge of the channel.

41 / 494
Receive Diversity via Selection Combining

• Assume that the nr channels are independant and identically Rayleigh distributed
(i.i.d.) with unit energy and that the noise levels are equal on each antenna.
• Choose the branch with the largest amplitude smax = max{s1 , . . . , snr }.
• The probability that s falls below a certain level S is given by its CDF
2
/2σ 2
P [s < S] = 1 − e−S .
• The probability that smax falls below a certain level S is given by
h i
2 nr
P [smax < S] = P [s1 , . . . , snr ≤ S] = 1 − e−S .

• The PDF of smax is then obtained by derivation of its CDF

2
h i
2 nr −1
psmax (s) = nr 2s e−s 1 − e−s .

• The average SNR at the output of the combiner ρ̄out is eventually given by
Z ∞ nr  
X 1 nr ր 1
ρ̄out = ρs2 psmax (s) ds = ρ ≈ ρ γ + log(nr ) + .
0 n=1
n 2nr

where γ ≈ 0.57721566 is Euler’s constant. We observe that the array gain ga is of

the order of log(nr ).
42 / 494
Receive Diversity via Selection Combining

• For BPSK and a two-branch diversity, the SER as a function of the average SNR per
channel ρ writes as
Z ∞ p

P̄ = Q 2ρs psmax (s) ds
0
r r
1 ρ 1 ρ
= − +
2 1+ρ 2 2+ρ
ρր
∼ 3
= .
8ρ2
The slope of the bit error rate curve is equal to 2.

• In general, the diversity gain gdo of a nr -branch selection diversity scheme is equal to
nr . Selection diversity extracts all the possible diversity out of the channel.

43 / 494
Receive Diversity via Gain Combining

• In gain combining, the signal z used for detection is a linear combination of all
branches,
nr
X √
z = gy = gn yn = Es ghc + gn
n=1

where
– gn ’s are the combining weights and g , [g1 , . . . , gnr ]
– the data symbol c is sent through the channel and received by nr antennas
– h , [h1 , . . . , hnr ]T
• Assume Rayleigh distributed channels hn = |hn | ejφn , n = 1, . . . , nr , with unit
energy, all the channels being independent.
• Equal Gain Combining : fixes the weights as gn = e−jφn .
– Mean value of the output SNR ρ̄out (averaged over the Rayleigh fading):
h i2 
P nr √
E n=1 E s |h n |  
π
ρ̄out = 2
= . . . = ρ 1 + (n r − 1) ,
n r σn 4
where the expectation is taken over the channel statistics. The array gain grows
linearly with nr , and is therefore larger than the array gain of selection combining.
– The diversity gain of equal gain combining is equal to nr analogous to selection.

44 / 494
Receive Diversity via Gain Combining

• Maximal Ratio Combining :the weights are chosen as gn = h∗n .

– It maximizes the average output SNR ρ̄out
( )
Es khk4 n o
ρ̄out = 2 E 2
= ρE khk2 = ρnr .
σn khk
The array gain ga is thus always equal to nr , or equivalently, the output SNR is the
sum of the SNR levels of all branches (holds true irrespective of the correlation
between the branches).
– For BPSK transmission, the symbol error rate reads as
Z ∞ p

P̄ = Q 2ρu pu (u) du
0

where u = khk2 is χ2 distribution with 2nr degrees of freedom when the different
channels are i.i.d. Rayleigh
1
pu (u) = unr −1 e−u .
(nr − 1)!
At high SNR, P̄ becomes
 
2nr − 1
P̄ = (4ρ)−nr .
nr
The diversity gain is again equal to nr .
45 / 494
Receive Diversity via Gain Combining

– For alternative constellations, the error probability is given, assuming ML detection, by

Z ∞ r !
ρu
P̄ ≈ N̄e Q dmin pu (u) du,
0 2
( )  
d2
min ρu x2
≤ N̄e E e− 4 (using Chernoff bound Q (x) ≤ exp − )
2

where N̄e and dmin are respectively the number of nearest neighbors and minimum
distance of separation of the underlying constellation.
Since u is a χ2 variable with 2nr degrees of freedom, the above average upper-bound
is given by
  nr
1
P̄ ≤ N̄e
1 + ρd2min /4
 
ρր ρd2min −nr
≤ N̄e .
4
The diversity gain gdo is equal to the number of receive branches in i.i.d. Rayleigh
channels.

46 / 494
Receive Diversity via Gain Combining

• Minimum Mean Square Error Combining

– So far noise was white Gaussian. When the noise (and interference) is colored, MRC is
not optimal anymore.
– Let us
√ denote the combined noise plus interference signal as ni such that
y = Es hc + ni .
– An optimal gain combining technique is the minimum mean square error (MMSE)
combining, where the weights are chosen in order to minimize the mean square error
between the transmitted symbol c and the combiner output z, i.e.,

g⋆ = arg min E |gy − c|2 .
g

– The optimal weight vector g⋆ is given by

g⋆ = hH R−1
ni ,
 H

where Rni = E ni ni is the correlation matrix of the combined noise plus
interference signal ni .
– Such combiner can be thought of as first whitening the noise plus interference by
−1/2 −1/2
multiplying y by Rni and then match filter the effective channel Rni h using
−H/2
hH Rni .
– The Signal to Interference plus Noise Ratio (SINR) at the output of the MMSE
combiner simply writes as
ρout = Es hH R−1ni h.
– In the absence of interference and the presence of white noise, MMSE combiner
reduces to MRC filter up to a scaling factor.
47 / 494
Receive Diversity via Gain Combining

Example
Question: Assume a transmission of a signal c from a single antenna
transmitter to a multi-antenna receiver through a SIMO channel h. The
transmission is subject to the interference from another transmitter sending
signal x through the interfering SIMO channel hi .
The received signal model writes as

y = hc + hi x + n

where n is the zero mean complex additive white Gaussian noise (AWGN)
vector with E{nnH } = σn2 Inr .
We apply a combiner g at the receiver to obtain the observation z = gy.
Derive the expression of the MMSE combiner and the SINR at the output of
the combiner.

48 / 494
Receive Diversity via Gain Combining

Example
Answer: The MMSE combiner g is given by

g = hH R−1
ni

where Rni = E ni nH i with ni = hi x + n.

Hence Rni = hi Px hi + σn2 Inr with Px = E |x|2 , the power of the
H

interfering signal.
Hence,   −1
g = hH hi Px hH 2
i + σn I nr .

At the receiver, we obtain

z = gy = hH R−1 H −1
ni hc + h Rni ni .

49 / 494
Receive Diversity via Gain Combining

Example
Answer: The output SINR writes
 H −1 2
h R n h  P c
i
ρout = n
H o
E hH R−1 H −1
ni i h Rni ni
n
 H −1 2
 h Rn h  P c
=  H −1 i H −1
E h Rni ni ni Rni h
 H −1 2
h Rn h  P c
i
=
hH R−1ni h

= hH R−1ni hPc
 −1
H
= Pc h hi Px hH 2
i + σn I n r h
 −1
= SNR hH INR hi hH i + I nr h

with Pc = E |c|2 , SNR = Pc /σn2 (the average SNR), INR = Px /σn2 (the
average INR - Interference to Noise Ratio). 
50 / 494
MISO Systems

• MISO systems exploit diversity at the transmitter through the use of nt transmit
antennas in combination with pre-processing or precoding.

• A significant difference with receive diversity is that the transmitter might not have
the knowledge of the MISO channel.
– At the receiver, the channel is easily estimated.
– At the transmit side, feedback from the receiver is required to inform the transmitter.

• There are basically two different ways of achieving direct transmit diversity :
– when Tx has a perfect channel knowledge, beamforming can be performed to achieve
both diversity and array gains,
– when Tx has a partial or no channel knowledge of the channel, space-time coding is
used to achieve a diversity gain (but no array gain in the absence of any channel
knowledge).

• Indirect transmit diversity techniques convert spatial diversity to time or frequency

diversity.

51 / 494
Transmit Diversity via Matched Beamforming

• The actual transmitted signal is a vector c′ that results from the multiplication of
the signal c by a weight vector w.
• At the receiver, the signal reads as
√ √
y = Es hc′ + n = Es hwc + n,
where h , [h1 , . . . , hnt ] represents the MISO channel vector, and w is also known as
the precoder.
• The choice that maximizes the receive SNR is given by
hH
w= .
khk
• Transmit along the direction of the matched channel, hence it is also known as
matched beamforming or transmit MRC.
• The array gain is equal to the number of transmit antennas, i.e. ρ̄out = nt ρ.
• The diversity gain equal to nt as the symbol error rate is upper-bounded at high
SNR by  2 −nt
ρdmin
P̄ ≤ N̄e .
4
• Matched beamforming presents the same performance as receive MRC, but requires
a perfect transmit channel knowledge.
52 / 494
Transmit Diversity via Space-Time Coding

• Alamouti scheme is an ingenious transmit diversity scheme for two transmit

antennas which does not require transmit channel knowledge.
– Assume that the flat fading channel remains constant over the two successive symbol
periods, and is denoted by h = [h1 h2 ].
– Two symbols c1 and c2 are transmitted simultaneously from antennas 1 and 2 during
the first symbol period, followed by symbols −c∗2 and c∗1 , transmitted from antennas 1
and 2 during the next symbol period:
p c1 p c2
y1 = E s h1 √ + E s h2 √ + n 1 , (first symbol period)
2 2
p c∗ p c∗
y2 = − Es h1 √2 + Es h2 √1 + n2 . (second symbol period)
2 2
The two symbols are spread over two antennas and over two symbol periods.
– Equivalently
  p   √   
y1 h1 h2 c1 /√2 n1
y= = E s + .
y2∗ h∗2 −h∗1 c2 / 2 n∗2
| {z }| {z }
Hef f c

– Applying the matched filter HH

ef fto the received vector y effectively decouples the
transmitted symbols as shown below
    p h i  √   
z1 y1 2 2 c1 /√2 n1
= HHef f ∗ = E s |h 1 | + |h 2 | I 2 + H H
ef f ∗
z2 y2 c2 / 2 n2
53 / 494
Transmit Diversity via Space-Time Coding

– The mean output SNR (averaged over the channel statistics) is thus equal to
( 2 )
Es khk2
ρ̄out = 2 E = ρ.
σn 2 khk2
No array gain owing to the lack of transmit channel knowledge.
– The average symbol error rate at high SNR can be upper-bounded according to
 2 −2
ρdmin
P̄ ≤ N̄e .
8
The diversity gain is equal to nt = 2 despite the lack of transmit channel knowledge.
0
10
no spatial diversity
transmit MRC
Transmit MRC vs. Alamouti with 2
Alamouti scheme
transmit antennas in i.i.d. Rayleigh
10
−1 fading channels (BPSK).
Observations:
−2 – At high SNR, any increase in the
SER

10
SNR by 10dB leads to a decrease of
SER by 10−n for diversity order n.
10
−3
Alamouti, transmit MRC: 2
No spatial diversity: 1

−4
– Transmit MRC has 3 dB gain over
10
0 2 4 6 8 10 12 Alamouti
SNR [dB]
54 / 494
Indirect Transmit Diversity

• It is also possible to convert spatial diversity to time or frequency diversity, which are
then exploited using well-known SISO techniques.
• Assume that nt = 2 and that the signal on the second transmit branch is
– either delayed by one symbol period: the spatial diversity is converted into frequency
diversity (delay diversity)
– either phase-rotated: the spatial diversity is converted into time diversity
– The effective SISO channel resulting from the addition of the two branches seen by the
receiver now fades over frequency or time. This selective fading can be exploited by
conventional diversity techniques, e.g. FEC/interleaving.

55 / 494
MIMO Systems - Transmission

56 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 1
Section: 1.2.4, 1.3.2, 1.6

57 / 494
Introduction - Previous Lectures

• Discrete Time Representation

√
– SISO: y = √Es hc + n
– SIMO: y = Es hc + n √
– MISO (with perfect CSIT): y = Es hwc + n
• h is fading
– amplitude Rayleigh distributed
– phase uniformly distributed
• Diversity
log2 (P̄ )
– Diversity gain: gdo (ρ) , − log2 (ρ)
ρ̄out ρ̄out
– Array gain: ga , ρ̄
= ρ
• SIMO
– selection combining
– gain combining
• MISO
– with perfect channel knowledge at Tx: Matched Beamforming
– without channel knowledge at Tx: Space-Time Coding (Alamouti Scheme), indirect
(time, frequency) transmit diversity

58 / 494
MIMO Systems

• In MIMO systems, the fading channel between each transmit-receive antenna pair
can be modeled as a SISO channel.
• For uni-polarized antennas and small inter-element spacings (of the order of the
wavelength), path loss and shadowing of all SISO channels are identical.
• Stacking all inputs and outputs in vectors ck = [c1,k , . . . , cnt ,k ]T and
yk = [y1,k , . . . , ynr ,k ]T , the input-output relationship at any given time instant k
reads as √
yk = Es Hk c′k + nk ,
where
– c′k is a precoded version of ck that depends on the channel knowledge at the Tx.
– Hk is defined as the nr × nt MIMO channel matrix, Hk (n, m) = hnm,k , with hnm
denoting the narrowband channel between transmit antenna m (m = 1, . . . , nt ) and
receive antenna n (n = 1, . . . , nr ),
– nk = [n1,k , . . . , nnr ,k ]T is the sampled noise vector, containing the noise contribution
at each receive antenna, such that the noise is white in both time and spatial
dimensions, E nk nH l = σn2I
nr δ (k − l).

• Using the same channels normalization as for SISO channels, E kHk2F = nt nr .
• when Tx has a perfect channel knowledge: (dominant and multiple) eigenmode
transmission
• when Tx has no knowledge of the channel: space-time coding (with c′k = ck )
59 / 494
Space-Time Coding

• MIMO without Transmit Channel Knowledge

• Array/diversity/coding gains are exploitable in SIMO, MISO and ... MIMO
• Alamouti scheme can easily be applied to 2 × 2 MIMO channels
 
h11 h12
H=
h21 h22
• Received signal vector (make sure the channel remains constant over two symbol
periods!)
 √ 
√ c 1 / √2
y1 = Es H + n1 , (first symbol period)
c2 / 2
 √ 
√ −c∗2 /√ 2
y2 = Es H ∗ + n2 . (second symbol period)
c1 / 2
• Equivalently
 
  h11 h12  √   
y1 √  h21 h 22
 c 1 / √2 n1
y= = Es   + .
y2∗  h∗12 −h∗11  c2 / 2 n∗2
h∗22 −h∗21 | {z }
| {z } c
Hef f

60 / 494
Space-Time Coding

2
• Apply the matched filter HH H
ef f to y (Hef f Hef f = kHkF I2 )
 
z1 √ √
z= = Es HH ef f y = Es kHk2F I2 c + n′
z2

where n′ is such that E{n′ } = 02×1 and E{n′ n′H } = kHk2F σn2 I2 .

• Average output SNR ( 2 )

Es kHk2F
ρ̄out = 2 E = 2ρ,
σn 2 kHk2F
Receive array gain (ga = nr = 2) but no transmit array gain!

• Average symbol error rate

 −4
ρd2min
P̄ ≤ N̄e .
8
Full diversity (gdo = nt nr = 4)

61 / 494
Dominant Eigenmode Transmission

• MIMO with Perfect Transmit Channel Knowledge

• Extension of Matched Beamforming to MIMO
√ √
y = Es Hc′ + n = Es Hwc + n,
√
z = gy = Es gHwc + gn.

• Decompose
H
H = U H ΣH V H ,
ΣH = diag{σ1 , σ2 , . . . , σr(H) }.

• Received SNR is maximized by matched filtering, leading to

w = vmax
g = uH
max

where vmax and umax are respectively the right and left singular vectors
corresponding to the maximum singular value of H, σmax = max{σ1 , σ2 , . . . , σr(H) }.
Note the generalization of√matched beamforming (MISO) and MRC (SIMO)!
• Equivalent channel: z = Es σmax c + ñ where ñ = gn has a variance equal to σn2 .
62 / 494
Dominant Eigenmode Transmission

2
• Array gain: E{σmax } = E{λmax } where λmax is the largest eigenvalue of HHH .
Commonly, max{nt , nr } ≤ ga ≤ nt nr .

Example
Array gain changes depending on the channel properties and distribution
√ of Sight: H = 1nr ×nt . Only one singular value is non-zero and equal to
– Line
nt nr : ga = nt nr .
√ √
2
– In the i.i.d. Rayleigh case: for large nt , nr , ga = nt + nr .

• Diversity gain: the dominant eigenmode transmission extracts a full diversity gain of
nt nr in i.i.d. Rayleigh channels.

63 / 494
Dominant Eigenmode Transmission

Example
Question: Show that the optimum (in the sense of SNR maximization) transmit
precoder and combiner in dominant eigenmode transmission is given by the
dominant right and left singular vector of the channel matrix, respectively.
Answer: Let us write
√ √
y = Es Hc′ + n = Es Hwc + n,
√
z = gy = Es gHwc + gn.

where kwk2 = 1 (power constraint). We decompose

H
H = U H ΣH V H , ΣH = diag{σ1 , σ2 , . . . , σr(H) }.

In order to maximize the SNR, we choose g as a matched filter, i.e.

g = (Hw)H such that
r(H)  2
X
gHw = wH HH Hw = wH VH Σ2H VH
H 2
σi2 viH w ≤ σmax
 
w=
i=1

where vi is the i column of VH and σmax = maxi=1,...,r(H) σi .

64 / 494
Dominant Eigenmode Transmission

Example
Answer: The inequality is replaced by an equality if w = vmax . By choosing
w = vmax ,

g = wH HH = vmax
H
V H ΣH U H
H

= σmax uH
max

where umax is the column of UH corresponding to the dominant singular value

σmax of H. If we normalize g such that kgk2 = 1, we can write g = umax . 

65 / 494
Multiple Eigenmode Transmission

• Assume nr ≥ nt an that r (H) = nt , i.e. nt singular values in H. Hence, what about

spreading symbols over all non-zero eigenmodes of the channel:
– Tx side: multiply the input vector c (nt × 1) using VH , i.e. c′ = VH c.
– Rx side: multiply the received vector y by G = UH H.
– Overall,
p
z = Es GHc′ + Gn
p
= E s UH H HVH c + U n
H
p
= Es ΣH c + ñ.
The channel has been decomposed into nt parallel SISO channels given by
{σ1 , . . . , σnt }.
• The rate achievable in the MIMO channel is the sum of the SISO channel capacities
nt
X
R= log2 (1 + ρsk σk2 ),
k=1

where {s1 , . . . , snt } is the power allocation on each of the channel eigenmodes.
• The capacity scales linearly in nt . By contrast, this transmission does not necessarily
achieve the full diversity gain of nt nr but does at least provide nr -fold array and
diversity gains (still assuming nt ≤ nr ).
• In general, the rate scales linearly with the rank of H.
66 / 494
Multiple Eigenmode Transmission

Example
Question: Is the rate achievable in a MIMO channel with multiple eigenmode
transmission and uniform power allocation across modes always larger than
that achievable with dominant eigenmode transmission?
Answer: No! The achievable rate with multiple eigenmode transmission in the
MIMO channel is the sum of the SISO channel achievable rates
r(H)
X
R= log2 (1 + ρsk σk2 ),
k=1

where {s1 , . . . , sr(H) } is the power allocation on each of the channel

eigenmodes. Pr(H)
Two strategies (for a total power constraint k=1 sk = 1):
Pr(H)
• Uniform power allocation: Ru = k=1 log2 (1 + ρ1/r(H)σk2 )
2
• Dominant eigenmode transmission: Rd = log2 (1 + ρσmax )
Ru could be either greater or smaller than Rd . For instance, if σ1 >> 0 and
σk ≈ ǫ for k > 1, Ru ≈ log2 (1 + ρσ12 /r(H)) ≤ Rd for small values of ρ. At
very high SNR, despite the little contributions of σk ≈ ǫ, Ru will become
higher than Rd . 
67 / 494
Multiplexing gain

• Array/diversity/coding gains are exploitable in SIMO, MISO and MIMO but MIMO
can offer much more than MISO and SIMO.

• MIMO channels offer multiplexing gain: measure of the number of independent

streams that can be transmitted in parallel in the MIMO channel. Defined as


R ρ
gs , lim

ρ→∞ log
2 ρ

where R(ρ) is the transmission rate.

• The multiplexing gain is the pre-log factor of the rate at high SNR, i.e.


R ≈ gs log2 ρ

• Modeling only the individual SISO channels from one Tx antenna to one Rx antenna
not enough:
– MIMO performance depends on the channel matrix properties
– characterize all statistical correlations between all matrix elements necessary!

68 / 494
Interference Management

• In wireless networks, co-channel interference is caused by the necessary frequency

re-use.

• With multiple antennas, it is possible to exploit the difference between the spatial
signatures of the desired vs. the interfering channels to reduce the intra-cell and
inter-cell interference:
– In a single-cell multi-user context, Multi-user MIMO (MU-MIMO)
– In a multi-cell context, Multi-Cell MIMO (MC-MIMO)

69 / 494
Channel Modelling

70 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 2
Section: 2.1.1, 2.1.2, 2.1.3, 2.1.5, 2.2,
2.3.1
– Chapter 3
Section: 3.2.1, 3.2.2, 3.4.1

71 / 494
Double-Directional Channel Modeling

• Space comes as an additional dimension

– directional: model the angular distribution of the energy at the antennas
– double: there are multiple antennas at transmit and receive sides
• Neglecting path-loss and shadowing, the time-variant double-directional channel
nX
s −1

h (t, pt , pr , τ, Ωt , Ωr ) = hk (t, pt , pr , τ, Ωt , Ωr ) ,
k=0

– pt , pr : location of Tx and Rx,

respectively
– ns contributions receiver
diffusion
– time t: variation with time (with the
motion of the receiver)
line-of-sight diffraction
– delay τ : each contribution arrives with a
delay proportional to its path length
specular reflection
– Ωt , Ωr : direction of departure (DoD),
directions of arrival (DoA). In spherical transmitter
coordinates (i.e., the azimuth Θt and
elevation ψt ) on a sphere of unit radius
Ωt = [cos Θt sin ψt , sin Θt sin ψt , cos ψt ]T
72 / 494
Double-Directional Channel Modeling

• In the case of a plane wave, and considering a fixed transmitter and a mobile receiver,
hk (t, pt , pr , τ, Ωt , Ωr ) , αk ejφk e−j∆ωk t δ(τ − τk ) δ(Ωt − Ωt,k ) δ(Ωr − Ωr,k ),
where
– αk is the amplitude of the kth contribution,
– φk is the phase of the kth contribution,
– ∆ωk is the Doppler shift of the kth contribution,
– τk is the time delay of the kth contribution,
– Ωt,k is the DoD of the kth contribution,
– Ωr,k is the DoA of the kth contribution.
• A more compact notation (all temporal variations are grouped into t)
nX
s −1

h (t, τ, Ωt , Ωr ) = hk (t, τ, Ωt , Ωr )
k=0

• Impulse response of the channel (as in Lecture 1, without path loss/shadowing)

ZZ


h(t, τ ) = h t, τ, Ωt , Ωr dΩt dΩr

• Narrowband transmission (the channel is not frequency selective)

ZZZ


h(t) = h t, τ, Ωt , Ωr dτ dΩt dΩr
73 / 494
Wide-Sense Stationary Uncorrelated Scattering
Homogeneous
• Assumption: Wide-Sense Stationary Uncorrelated Scattering Homogeneous
(WSSUSH) channels

• Wide-Sense Stationary:
– Time correlations only depend on the time difference
– Signals arriving with different Doppler frequencies are uncorrelated

• Uncorrelated Scattering:
– Frequency correlations only depend on the frequency difference
– Signals arriving with different delays are uncorrelated

• Homogeneous:
– Spatial correlation only depends on the spatial difference at both transmit and receive
sides
– Signals departing/arriving with different directions are uncorrelated

74 / 494
Spectra

• Doppler spectrum and coherence time

• Power delay spectrum and delay spread
• Power direction spectrum and angle spread
– the power-delay joint direction spectrum


2
Ph τ, Ωt , Ωr = E h t, τ, Ωt , Ωr  ,
– the joint direction power spectrum
Z


A(Ωt , Ωr ) = Ph τ, Ωt , Ωr dτ,

– the transmit direction power spectrum

Z Z


At (Ωt ) = Ph τ, Ωt , Ωr dτ dΩr ,

– the receive direction power spectrum

Z Z


Ar (Ωr ) = Ph τ, Ωt , Ωr dτ dΩt .

75 / 494
Angular Spread

• The channel angle-spreads are defined similarly to the delay-spread

– delay-spread ⇐⇒ channel frequency selectivity
– angle-spread ⇐⇒ channel spatial selectivity
R
Ω A (Ω ) dΩt
Ωt,M = R t t t
At (Ωt ) dΩt
sR
kΩt − Ωt,M k2 At (Ωt ) dΩt
Ωt,RM S = R
At (Ωt ) dΩt

L=2
L=1

L=0

76 / 494
The MIMO Channel Matrix

• Convert the double-directional channel to a nr × nt MIMO channel

 
h11 (t, τ ) h12 (t, τ ) . . . h1nt (t, τ )
 h21 (t, τ ) h22 (t, τ ) . . . h2nt (t, τ ) 
 
H(t, τ ) =  .. .. .. .. ,
 . . . . 
hnr 1 (t, τ ) hnr 2 (t, τ ) . . . hnr nt (t, τ )

where
ZZ


hnm (t, τ ) , hnm t, τ, Ωt , Ωr dΩt dΩr

• For narrowband (i.e. same delay for all antennas)

balanced (i.e. |hnm | = |h11 |) arrays
and plane wave
incidence, hnm t, τ, Ωt , Ωr is a phase shifted version of
h11 t, τ, Ωt , Ωr
Z Z  (n) (1)   (m) (1) 

T T
hnm (t, τ ) = h11 t, τ, Ωt , Ωr e−jkr (Ωr ) pr −pr e−jkt (Ωt ) pt −pt dΩt dΩr

where kt (Ωt ) and kr (Ωr ) are the transmit and receive wave propagation 3 × 1
vectors.
77 / 494
Steering Vectors

• For a transmit ULA oriented broadside to the link axis,

 (m) (1) 
T
e−jkt (Ωt )· pt −pt = e−j(m−1)ϕt (θt ) ,
(m)
(m−1)
where ϕt (θt ) = 2π(dt /λ) cos θt , and dt = pt − pt denotes the
inter-element spacing of the transmit array.
– θt is defined relatively to the array orientation (so θt = π/2 corresponds to the link axis
for a broadside array).

• Steering vector (expressed here for a ULA)

– At the transmitter in the relative direction θt :
at (θt ) = [ 1 e−jϕt (θt ) ... e−j(nt −1)ϕt (θt ) ]T .
– At the receiver in the relative direction θr :
ar (θr ) = [ 1 e−jϕr (θr ) ... e−j(nr −1)ϕr (θr ) ]T .

• Under the plane wave and balanced narrowband array assumptions, the MIMO
channel matrix can be rewritten as a function of steering vectors as
Z Z
(1)

H(t, τ ) = h t, pt , p(1) T
r , τ, Ωt , Ωr ar (Ωr ) at (Ωt ) dΩt dΩr .

78 / 494
A Finite Scatterer MIMO Channel Representation

• The transmitter and receiver are coupled via a finite number of scattering paths with
ns,t DoDs at the transmitter and ns,r DoAs at the receiver.
−→ Replace the integral by a summation (assume for simplicity 2-D azimuthal
propagation)
ns,t ns,r
XX (l,p) (l)
H(t, τ ) = h11 (t, τ )ar (θr(p) )aTt (θt )
l=1 p=1

= Ar Hs (t, τ )ATt
where
– Ar and At represent the nr × ns,r and nt × ns,t matrices whose columns are the
steering vectors related to the directions of each path observed at Rx and Tx
– Hs (t, τ ) is a ns,r × ns,t matrix whose elements are the complex path gains between all
DoDs and DoAs at time instant t and delay τ
(l)
• Assume the columns of At are written as at (θt ), l = 1, ..., ns,t . Let us write
ns,t ns,t
X (l)
X
H = Ar Hs ATt = H̃s (:, l)aTt (θt ) = H(l) ,
| {z }
l=1 l=1
H̃s

where H(l) can be viewed as the channel matrix corresponding to the lth scatterer
(l)
located in the direction of departure θt .
79 / 494
Statistical Properties of the MIMO Channel Matrix

• Assume narrowband channels, the spatial correlation matrix of the MIMO channel
R = E{vec(HH )vec(HH )H }
This is a nt nr × nt nr positive semi-definite Hermitian matrix.
• It describes the correlation between all pairs of transmit-receive channels:
– E {H(n, m)H∗ (n, m)}: the average energy of the channel between antenna m and
antenna n,
(nq)
– rm = E {H(n, m)H∗ (q, m)}: the receive correlation between channels originating
from transmit antenna m and impinging upon receive antennas n and q,
(mp)
– tn = E {H(n, m)H∗ (n, p)}: the transmit correlation between channels originating
from transmit antennas m and p and arriving at receive antenna n,
– E {H(n, m)H∗ (q, p)}: the cross-channel correlation between channels (m, n) and
(q, p).

Example
2x2 MIMO
 
1 t∗1 r1∗ s∗1
 t1 1 s∗2 r2∗  t1 = E {H(1, 1)H∗ (1, 2)}
R= 
 r1 s2 1 t∗2 
s1 r2 t2 1 r1 = E {H(1, 1)H∗ (2, 1)}

80 / 494
Spatial Correlation

• How are these correlations related to the propagation channel?

• Let us consider the case of ULAs and 2-D azimuthal propagation
Z Z


hnm (t) = h11 t, Ωt , Ωr e−j(m−1)ϕt (θt ) e−j(n−1)ϕr (θt ) dθt dθr

where
– ϕr,t (θr,t ) = 2π(dr,t /λ) cos θr,t ,
– dr and dt are
the Rinter-element spacing

at the receive/transmit arrays
– h11 t, Ωt , Ωr , h11 t, τ, Ωt , Ωr dτ .
• Correlation between channels hnm and hqp
Z 2π Z 2π 
 

E hnm h∗qp = E h11 t, Ωt , Ωr 2 e−j(m−p)ϕt (θt ) e−j(n−q)ϕr (θt ) dθt dθr
0 0
Z 2π Z 2π n
2 o −j(m−p)ϕt (θt ) −j(n−q)ϕr (θt )
= E h11 t, Ωt , Ωr  e e dθt dθr ,
0 0
Z 2π Z 2π
= A (θt , θr ) e−j(m−p)ϕt (θt ) e−j(n−q)ϕr (θt ) dθt dθr ,
0 0
where A (θt , θr ) is the joint direction power spectrum restricted to the azimuth
angles.
• The channel correlation is related to both the antenna spacings and the joint
direction power spectrum! 81 / 494
Spatial Correlation

• When the energy spreading is very large at both sides and dt /dr are sufficiently
large, elements of H become uncorrelated, and R becomes diagonal.

Example
Consider two transmit antennas spaced by dt . The transmit correlation writes
as Z 2π
t= ej2π(dt /λ) cos θt At (θt )dθt ,
0

which only depends on the transmit antenna spacing and the transmit direction
power spectrum.
– isotropic scattering : very rich scattering environment around the transmitter with
a uniform distribution of the energy, i.e. At (θt ) ∼
= 1/2π
Z 2π Z 2π
1 1
t= ejϕt (θt ) dθt = ej2π(dt /λ) cos θt dθt
2π 0 2π 0
!
dt
= J0 2π .
λ
The transmit correlation only depends on the spacing between the two antennas.

82 / 494
Spatial Correlation

Example
– highly directional scattering : scatterers around the transmit array are
concentrated along a narrow direction θt,0 , i.e., At (θt ) → δ(θt − θt,0 )

t → ejϕt (θt,0 ) = ej2π(dt /λ) cos θt,0 .

Very high transmit correlation approaching one. The scattering direction is
directly related to the phase of the transmit correlation.
1 – At (θt ) in real-world channels:
neither uniform nor a delta.
κ = 500
– isotropic scattering (κ = 0): first
κ = 100
κ = 10
minimum for dt = 0.38λ
Transmit correlation

0.5
– directional scattering (κ = ∞):
κ=2 correlation never reaches 0
– in practice, decorrelation in rich
0 scattering is reached for
dt ≈ 0.5λ
κ=0 – The more directional the
azimuthal dispersion (i.e. for κ
−0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 increasing), the larger the
Transmit antenna spacing relative to wavelength
antenna spacing required to
obtain a null correlation. 83 / 494
Analytical Representation of Rayleigh MIMO Channels

• Independent and Identically Distributed (I.I.D.) Rayleigh fading

– R = I nt nr
– H = Hw is a random fading matrix with unit variance and i.i.d. circularly symmetric
complex Gaussian entries.
• Realistic in practice only if both conditions are satisfied:
– the antenna spacings and/or the angle spreads at Tx and Rx are large enough,
– all individual channels characterized by the same average power (i.e., balanced array).
• What about real-world channels? Sometimes significantly deviate from this ideal
channel:

– limited angular spread and/or reduced

array sizes cause the channels to become
correlated (channels are not independent
anymore)
– a coherent contribution may induce the
channel statistics to become Ricean
(channels are not Rayleigh distributed
anymore),
– the use of multiple polarizations creates
gain imbalances between the various
elements of the channel matrix (channels
are not identically distributed anymore).
84 / 494
Correlated Rayleigh Fading Channels

• For identically distributed Gaussian channels, R constitutes a sufficient description of

the stochastic behavior of the MIMO channel.
• Any channel realization is obtained by


vec HH = R1/2 vec(Hw ),
where Hw is one realization of an i.i.d. channel matrix.
• Complicated to use because
– cross-channel correlation not intuitive and not easily tractable
– Too many parameters: dimensions of R rapidly become large as the array sizes increase
– vec operation complicated for performance analysis
• Kronecker model: use a separability assumption
R = Rr ⊗ Rt ,
1/2
H = R1/2
r H w Rt

where Rt and Rr are respectively the transmit and receive correlation matrices.
• Strictly valid only if r1 = r2 = r and t1 = t2 = t and s1 = rt and s2 = rt∗ (for 2 × 2)
   
1 t∗1 r1∗ s∗1 1 t∗ r ∗ r ∗ t∗    
 t1 1 s∗2 r2∗   t 1 r∗ t r∗  1 r∗ 1 t∗
R=  r1 s2 1 t∗2   r rt∗
 =   = ⊗
1 t∗  r 1 t 1
s 1 r 2 t2 1 rt r t 1 | {z } | {z }
Rr Rt
85 / 494
Correlated Rayleigh Fading Channels

Example
Question: Assume a MISO system with two transmit antennas. The channel
gains are identically distributed circularly symmetric complex Gaussian but can
be correlated and are denoted as h1 and h2 . Write the expression of the
transmit correlation matrix Rt and derive the eigenvalues and eigenvectors of
Rt as a fonction of the transmit correlation coefficient t.
Answer: We write
 ∗       
h1   E |h1 |2 E{h∗1 h2 } 1 t∗
Rt = E h1 h2 = =
h∗2 E {h1 h∗2 } E |h2 |2 t 1

where t = E {h1 h∗2 } is the transmit correlation coefficient. The SVD leads to
   H
1 1 1 + |t| 0 1 1
Rt = .
t/|t| −t/|t| 0 1 − |t| t/|t| −t/|t|

The eigenvalues are only function of the magnitude of t while the eigenvectors
are only function of the phase of t.


86 / 494
Correlated Rayleigh Fading Channels

Example
Question: Assume the previous example with |t| → 1. Compute the weights of
the matched beamformer (or maximum ratio transmission/transmit MRC).
Answer: With matched beamforming, w = hH / khk where
1/2
h = h w Rt
  p  H
1 1 1 + |t| p 0 1 1
= hw
t/|t| −t/|t| 0 1 − |t| t/|t| −t/|t|
  H
1 1
= 2hw
t/|t| t/|t|

where the last equality comes from the fact that |t| = 1. This shows
 that for
high correlation, the channel direction (h/ khk) is aligned with 1 t∗ /|t| .
Hence  
H 1
w = h / khk = .
t/|t|
Transmission is performed in the direction where all scatterers are located.

87 / 494
Analytical Representation of Ricean MIMO Channels

• In the presence of a strong coherent component which does not experience any
fading over time
– e.g. a line-of-sight field, one or several specular contributions, coherent addition of
reflected and diffracted contributions (in fixed wireless access only).
• All these situations lead to a Ricean distribution of the received field amplitude.
– The relative strength of the dominant coherent component is characterized by the
K-factor K. As the channel contains a coherent component, its amplitude can be
written as  
 
|h(t)| = h̄ + h̃(t) ,

where h̄ is the coherent component, and h̃(t) is the non coherent part, whose energy is
denoted as 2σs2 .
– The K-factor is defined as  2
h̄
K= .
2σs2
– |h(t)| , s′ is Ricean distributed, and its distribution is given, as a function of K, as
" !# !
′ 2s′ K s′2 2s′ K
ps′ (s ) =  2 exp − K  2 + 1 I0   .
h̄ h̄ h̄

– For K = 0, the Ricean distribution boils down to the Rayleigh distribution while for
K = ∞, the channel becomes deterministic (no fading).

88 / 494
Ricean MIMO Channels

• Common Ricean MIMO channel model

r r
K 1
H= H̄ + H̃
1+K 1+K

• The matrix H̃ relates to the Rayleigh component (non-coherent part). It can be

modeled and characterized using

R = E{vec(H̃H )vec(H̃H )H }

• The matrix H̄ corresponding to the coherent component(s) has fixed phase-shift-only

entries (strongly related to the array configuration and orientation)
 jα11 
e ejα12
H̄ = jα21 jα22
e e

– With only one coherent contribution with given DoD and DoA (Ωt,c and Ωr,c ),
H̄ = ar (Ωr,c ) aT
t (Ωt,c )
– For broadside arrays with a pure line-of-sight component, H̄ = 1nr ×nt .

89 / 494
Capacity of point-to-point MIMO Channels

90 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 5
Section: 5.1, 5.2, 5.3, 5.4.1, 5.4.2
(except “Antenna Selection Schemes”),
5.5.1 - “Kronecker Correlated Rayleigh
Channels”, 5.5.2, 5.7, 5.8.1 (except
Proof of Proposition 5.9 and Example
5.4)

91 / 494
Introduction - Previous Lectures

• Transmission strategies
– Space-Time Coding when no Tx channel knowledge
– Multiple (including dominant) eigenmode transmission when Tx channel knowledge
p
z = Es GHc′ + Gn
p
= E s UHH HVH c + U n
H
p
= Es ΣH c + ñ.
Multiple parallel data pipes → Spatial multiplexing gain!

• Performance highly depends on the channel matrix properties

– Angle spread and inter-element spacing
– Spatial Correlation: spread antennas far apart to decrease spatial correlation
– Rayleigh and Ricean distribution

92 / 494
System Model

• A single-user MIMO system with nt transmit and nr receive antennas over a

frequency flat-fading channel.
• The transmit and received signals in a MIMO channel are related by
√
yk = Es Hk c′k + nk
where
– yk is the nr × 1 received signal vector,
– Hk is the nr × nt channel matrix
– nk is a nr × 1 zero mean complex additive white Gaussian noise (AWGN) vector with
E{nk nH 2
l } = σn Inr δ (k − l).
– ρ = Es /σn2 represents the SNR.

• The input covariance matrix is defined as the covariance matrix of the transmit
H
signal c′ (we drop the time index) and writes as Q = E c′ c′ .
• Short-term power constraint: Tr{Q} ≤ 1.
• Long-term power constraint (over a duration Tp >> T ): E {Tr{Q}} ≤ 1 where the
expectation refers here to the averaging over successive codeword of length T .
• Channel time variation: Tcoh coherence time
– slow fading : Tcoh is so long that coding is performed over a single channel realization.
– fast fading : Tcoh is so short that coding over multiple channel realizations is possible.

93 / 494
Capacity of Deterministic MIMO Channels

Proposition

For a deterministic MIMO channel H, the mutual information I is written as

 
I(H, Q) = log2 det Inr + ρHQHH

where Q is the input covariance matrix whose trace is normalized to unity.

Definition
The capacity of a deterministic nr × nt MIMO channel with perfect channel
state information at the transmitter is
 
C (H) = max log2 det Inr + ρHQHH .
Q≥0:Tr{Q}=1

Note the difference with SISO capacity.

94 / 494
 Capacity of Deterministic MIMO Channels

Proof: Denoting the entropy by H(.), the mutual information between input and output
is given by
I(H, Q) = I(c′ ; y |H ),

= H(y |H ) − H(y c′ , H )

= H(y |H ) − H(n c′ , H ).

When the input vector has a covariance Q = E c′ c′H , we have that the covariance of y
is given by 
E yyH = σn2 Inr + Es HQHH ,
since the noise is an additive white Gaussian noise (AWGN). Following similar steps as in
SISO, H(y |H ) is largest when y is zero-mean circularly symmetric complex Gaussian,
which is achieved when c′ is zero-mean circularly symmetric complex Gaussian. Because
the differential entropy H(c′ ) of a zero-mean circularly symmetric
complex Gaussian
input vector c′ with covariance matrix Q is given by log2 det πeQ , we get
 !
2 H

I(H, Q) = log2 det πe σn Inr + Es HQH − log2 det πeσn2 Inr ,
 
H
= log2 det Inr + ρHQH .


95 / 494
Capacity and Water-Filling Algorithm

• What is the best transmission strategy, i.e. the optimum input covariance matrix Q?
• First, create n = min{nt , nr } parallel data pipes (Multiple Eigenmode Transmission)
– Decouple the channel along the individual channel modes (in the directions of the
singular vectors of the channel matrix H at both the transmitter and the receiver)
H
H = U H ΣH V H ,
UH H H
H HVH = UH UH ΣH VH VH = ΣH
– Optimum input covariance matrix Q⋆ writes as
Q⋆ = VH diag {s⋆1 , . . . , s⋆n } VH
H
,
• Second, allocate power to data pipes
– ΣH = diag {σ1 , . . . , σn }, and σk2 , λk
Pn   Pn  ⋆

– Capacity: C(H) = max{sk }n k=1 log2 1 + ρsk λk = k=1 log2 1 + ρsk λk
k=1

Proposition
⋆ . , s⋆n } that maximizes
Pn power allocation strategy {s1 , . . . , sn } = {s1 , . .P
The
n
k=1 log2 (1 + ρλk sk ) under the power constraint k=1 sk = 1, is given by the
water-filling solution,
 +
1
s⋆k = µ − , k = 1, . . . , n
ρλk
P
where µ is chosen so as to satisfy the power constraint n ⋆
k=1 sk = 1.

96 / 494
Water-Filling Algorithm

• Iterative power allocation

– Order eigenvalues λk in decreasing order
of magnitude
– At iteration i, evaluate the constant µ
from the power constraint
n−i+1
!
1 X 1 s3∗ 1
µ(i) = 1+ s2∗ ρ λn
n−i+1 k=1
ρλk 1
s1∗
ρ λn - 1
– Calculate power µ . ..
1
1 1 ρ λ3
sk (i) = µ(i) − , 1 ρ λ2
ρλk
ρ λ1
k = 1, . . . , n − i + 1.
If sn−i+1 < 0, set to 0
– Iterate till the power allocated on each
mode is non negative.

97 / 494
Water-Filling Algorithm

Example
Question: Consider the transmission y = Hc′ + n with perfect CSIT over a
deterministic point to point MIMO channel whose matrix is given by
 
a 0 a 0
H=
0 b 0 b

where a and b are complex scalars with |a| ≥ |b|. The input covariance matrix
is given by Q = E c′ c′H and is subject to the transmit power constraint
Tr {Q} ≤ P .
1 Compute the capacity with perfect CSIT of that deterministic channel.
Particularize to the case a = b. Explain your reasoning.
2 Explain how to achieve that capacity.
3 In which deployment scenario, could such channel matrix structure be
encountered?

98 / 494
Water-Filling Algorithm

Example
Answer:
H
1 Let us write Q = VPV with the diagonal element of P, denoted as Pk
P nt
(satisfying k=1 Pk = P ), refers to the power allocated to stream k. The
capacity with perfect CSIT over the deterministic channel H is given by
min{2,4}  
X Pk
C (H) = max log2 1 + 2 λk
P1 ,...,Pk σn
k=1

where λk refers the non-zero eigenvalue of HH H, respectively equal to

2 |a|2 and 2 |b|2 . Hence,
    
P1 P2
C (H) = max log2 1 + 2 2 |a|2 + log2 1 + 2 2 |b|2 .
P1 ,P2 σn σn
The optimal power allocation is given by the water-filling solution
 +  +
σn2 σ2
P1⋆ = µ − 2 , P2⋆ = µ − n 2
2 |a| 2 |b|
with µ computed such that P1⋆ + P2⋆ = P . 99 / 494
Water-Filling Algorithm

Example
Answer: 2
  2
σn σn
Assuming P1⋆ and P2⋆ are positive, µ = P
2
+ 4
1
|a|2
+ 1
|b|2
. If µ − 2|b|2
≤ 0,
2 2
σn σn
i.e. P
2
+ 4|a|2
− 4|b|2
≤ 0, P2⋆ = 0 and P1⋆ = P . The capacity writes as
 
P
C (H) = log2 1 + 2 2 |a|2 .
σn
σ2 σ2 σ2 σ2 σ2 σn2
If P2 + 4|a|n2 − 4|b|n2 > 0, P1⋆ = P2 − 4|a|n2 + 4|b|n2 and P2⋆ = P2 + 4|a|n2 − 4|b|2
.
The capacity writes as
   
P⋆ P⋆
C (H) = log2 1 + 12 2 |a|2 + log2 1 + 22 2 |b|2 .
σn σn

In the particular case where a = b, uniform power allocation P1⋆ = P2⋆ = P

2
is
optimal and  
P
C (H) = 2 log2 1 + 2 |a|2 .
σn

100 / 494
Water-Filling Algorithm

Example
Answer:
H
2 Transmit along V, given by the two dominant eigenvector of H H. They
are easily computed given the orthogonality of the channel matrix H as
 
1 0
1  0 1 
V= √  .
2 1 0 
0 1

The power allocated to the two streams is given by P1⋆ and P2⋆ . At the
receiver, the precoded channel is already decoupled and no further
combiner is necessary. Each stream can be decoded using the
corresponding SISO decoder.
3 Dual-polarized antenna deployment (e.g. VHVH-VH) with LoS and good
antenna XPD.


101 / 494
Capacity Bounds and Suboptimal Power Allocations

• Low SNR: power allocated to the dominant eigenmode

ρ→0
C (H) → log2 (1 + ρλmax ) .
• High SNR: power is uniformly allocated among the non-zero modes
n 
ρ→∞
X ρ 
C (H) → log2 1 + λk .
n
k=1

• At any SNR
– lower bound
C (H) ≥ log2 (1 + ρλmax ) ,
n 
X ρ 
C (H) ≥ log2 1 + λk .
k=1
n

– upper bound (use Jensen’s inequality Ex {F (x)} ≤ F (Ex {x}) if F concave)

n
" n #!
X   (a) ρ X ⋆
CCSIT (H) = log2 1 + ρs⋆k λk ≤ n log2 1 + sk λ k ,
k=1
n k=1
h ρ i
≤ n log2 1 + λmax .
n
102 / 494
Ergodic Capacity of Fast Fading Channels

• Fast fading:
– Doppler frequency sufficiently high to allow for coding over many channel
realizations/coherence time periods
– The transmission capability is represented by a single quantity known as the ergodic
capacity
• MIMO Capacity with Perfect Transmit Channel Knowledge
– similar strategy as in deterministic channels: transmit along eigenvectors of channel
matrix and allocate power following water-filling
– short term power constraint: water-filling solution applied over space as in
deterministic channels
(  )
C̄CSIT,ST = E max log2 det Inr + ρHQHH
Q≥0:Tr{Q}=1
n
X  
 
= E log2 1 + ρs⋆k λk .
k=1
– long term power constraint: water-filling solution applied over both time and space
X n  
 
C̄CSIT,LT = E log2 1 + ρs⋆k λk .
k=1
– Impact on coding strategy? Use a variable-rate code (family of codes of different rates)
adapted as a function of the water-filling allocation. No need for the codeword to span
many coherence time periods.
103 / 494
MIMO Capacity with Partial Transmit Channel Knowledge

• H is not known to the transmitter → we cannot adapt Q at all time instants

• Rate of information flow between Tx and Rx at time instant k over channels Hk
h i
log2 det Inr + ρHk QHHk .

Such a rate varies over time according to the channel fluctuations. The average rate
of information flow over a time duration T >> Tcoh is
T −1 h i
1 X
log2 det Inr + ρHk QHH k .
T
k=0

Definition
The ergodic capacity of a nr × nt MIMO channel with channel distribution
information at the transmitter (CDIT) is given by
 h i
C̄CDIT , C̄ = max E log2 det Inr + ρHQHH ,
Q≥0:Tr{Q}=1

where Q is the input covariance matrix optimized as to maximize the ergodic

mutual information.
• T >> Tc to average out the noise and the channel fluctuations
104 / 494
I.I.D. Rayleigh Fast Fading Channels: Perfect Transmit
Channel Knowledge
• Low SNR: allocate all the available power to the strongest or dominant eigenmode.
Use log2 (1 + x) ≈ x log2 (e) for x small and get
 
 
C̄CSIT,ST = E log2 1 + ρλmax

∼
= ρE λmax log2 (e)
∼
= ρn log2 (e), N, n → ∞, N/n >> 0.
 
 
C̄CSIT,LT = E log2 1 + ρs⋆max λmax
 ⋆
∼
= ρE smax λmax log2 (e)

Observations: C̄CSIT grows linearly in the minimum number of antennas n.

• High SNR: uniform power allocation on all non-zeros eigenmodes
 ( n )
Xn h ρ i ρ X
C̄CSIT =∼ E log2 1 + λk ∼ nlog
= +E log2 (λk ) .
2
n n
k=1 k=1

Observations: C̄CSIT also scales linearly with n. The spatial multiplexing gain is
gs = n. MISO fading channels do not offer any multiplexing gain.
105 / 494
I.I.D. Rayleigh Fast Fading Channels: Partial Transmit
Channel Knowledge
• Optimal covariance matrix

Proposition
In i.i.d. Rayleigh fading channels, the ergodic capacity with CDIT is achieved
under an equal power allocation scheme Q = Int /nt , i.e.,
(  ) ( n  )
ρ H
X ρ
C̄CDIT = Īe = E log2 det Inr + Hw Hw =E log2 1 + λk .
nt nt
k=1

Encoding requires a fixed-rate code (whose rate is given by the ergodic capacity)
with encoding spanning many channel realizations.
• Low SNR:
(  )
ρ ρ  o
2
C̄CDIT ≥ E log2 1 + kHw kF ≈ E kHw k2F log2 (e) = nr ρ log2 (e)
nt nt

Observations:
– C̄CDIT is only determined by the energy of the channel.
– A MIMO channel only yields a nr gain over a SISO channel. Increasing the number of
transmit antennas is not useful (contrary to perfect CSIT). SIMO and MIMO channels
reach the same capacity for a given nr .
106 / 494
I.I.D. Rayleigh Fast Fading Channels: Partial Transmit
Channel Knowledge
• High SNR:
( n  ) ( n
)
X ρ ρ X
C̄CDIT ≈ E log2 λk = nlog2 +E log2 (λk )
nt nt
k=1 k=1

Observations:
– C̄CDIT at high SNR scales linearly with n (by contrast to the low SNR regime).
– The multiplexing gain gs is equal to n, similarly to the CSIT case.
– C̄CDIT and C̄CSIT are not equal: constant gap equal to n log2 (nt /n) at high SNR.
• Expressions can be particularized to SISO, SIMO, MISO cases. At high SNR,
– SISO (N = n = 1):
n  o
C̄CDIT ≈ log2 (ρ) + E log2 |h|2 = log2 (ρ) − 0.83 = CAW GN − 0.83

– SIMO (nt = n = 1, nr = N ):
C̄CDIT ≈ log2 (nr ρ)
– MISO (nr = n = 1, nt = N ):
n  o nt →∞
C̄CDIT ≈ log2 (ρ) + E log2 khk2 /nt ≈ log2 (ρ) = CAW GN

107 / 494
I.I.D. Rayleigh Fast Fading Channels

• Ergodic capacity of various nr × nt i.i.d. Rayleigh channels with full (CSIT) and
partial (CDIT) channel knowledge at the transmitter.

20
2 x 2 (CSIT)
18 4 x 2 (CSIT)
2 x 4 (CSIT)
16 4 x 4 (CSIT)
2 x 2 (CDIT)
Ergodic capacity [bps/Hz]

14 4 x 2 (CDIT)
2 x 4 (CDIT)
12 4 x 4 (CDIT)

10

0
−10 −5 0 5 10 15 20
SNR [dB]
108 / 494
Correlated Rayleigh Fast Fading Channels: Uniform Power
Allocation
• Assume the channel covariance matrix is unknown to the transmitter
• Mutual information with identity input covariance matrix
(  )
ρ H
Īe = E log2 det Inr + HH .
nt

• Low SNR (  )
ρ 2
Īe ≥ E log2 1 + kHkF .
nt
1/2 1/2
• High SNR in Kronecker Correlated Rayleigh Channels H = Rr Hw Rt (with full
rank correlation matrices) and nt = nr
(  )
ρ H
Īe ≈ E log2 det Hw Hw + log2 det(Rr ) + log2 det(Rt ).
nt

Observations:
– det(Rr ) ≤ 1 and det(Rt ) ≤ 1: receive and transmit correlations always degrade the
mutual information (with power uniform allocation) with respect to the i.i.d. case.
– Īe still scales linearly with min{nt , nr }
109 / 494
Correlated Rayleigh Fast Fading Channels: Partial
Transmit Channel Knowledge
• Assume the channel covariance matrix is known to the transmitter.
Proposition

In Kronecker correlated Rayleigh fast fading channels, the optimal input

covariance matrix can again be expressed as

Q = U R t ΛQ U H
Rt ,

where URt is a unitary matrix formed by the eigenvectors of Rt (arranged in

such order that they correspond to decreasing eigenvalues of Rt ), and ΛQ is a
diagonal matrix whose elements are also arranged in decreasing order.

Power allocation has to be computed numerically. Approximation using Jensen’s

inequality is possible.
• Spatial correlation: beneficial or detrimental?
– receive correlations degrade both the mutual information Īe and the capacity with
CDIT,
– transmit correlations always decrease Īe but may increase C̄CDIT at low SNR
(irrespective of nt and nr ) or at higher SNR when nt > nr (analogous to the full CSIT
case).

110 / 494
Correlated Rayleigh Fast Fading Channels: Partial
Transmit Channel Knowledge
• Mutual information of various strategies at 0 dB SNR as a function of the transmit
correlation |t| in TIMO. Beamforming refers here to the tranmsission of one stream
along the dominant eigenvector of Rt .

3.6

3.5
Mutual information/capacity [bps/Hz]

3.4

3.3

3.2

3.1

2.9

2.8
Optimal allocation
2.7 Equi−power allocation
Beamforming
2.6
0 0.2 0.4 0.6 0.8 1
Magnitude of transmit correlation
111 / 494
Outage Capacity and Probability in Slow Fading Channels

• In slow fading, the encoding still averages out the randomness of the noise but
cannot fully average out the randomness of the channel.
• For a given channel realization H and a target rate R, reliable transmission if
 
log2 det Inr + ρHQHH > R
If not met with any Q, an outage occurs and the decoding error probability is strictly
non-zero.

• Look at the tail probability of log2 det Inr + ρHQHH , not its average!

Definition

The outage probability Pout (R) of a nr × nt MIMO channel with a target rate
R is given by
   
Pout (R) = min P log2 det Inr + ρHQHH < R ,
Q≥0:Tr{Q}≤1

where Q is the input covariance matrix optimized as to minimize the outage

probability.

• More meaningful in the absence of CSI knowledge at the transmitter: the transmitter
cannot adjust its transmit strategy → hopes the channel is good enough
112 / 494
Diversity-Multiplexing Trade-Off in Slow Fading Channels

• Compound channel coding theorem: there exist “universal” codes with rate R
bits/s/Hz that achieve reliable transmission over any slow fading channel realization
which is not in outage.
– CSIT is actually not necessary in slow fading channels if the aim is transmit reliably
when the channel is not in outage.
• For a given R, how does Pout behave as a function of the SNR ρ?

Definition
A diversity gain gd⋆ (gs , ∞) is achieved at multiplexing gain gs at infinite SNR if


R ρ
lim
= gs
ρ−→∞ log
2 ρ


log2 Pout (R)
lim
= −gd⋆ (gs , ∞)
ρ−→∞ log2 ρ

The curve gd⋆ (gs , ∞) as function of gs is known as the asymptotic

diversity-multiplexing trade-off of the channel.

– The multiplexing gain indicates how fast the transmission rate increases with the SNR.
– The diversity gain represents how fast the outage probability decays with the SNR.
113 / 494
Diversity-Multiplexing Trade-Off in I.I.D. Rayleigh Slow
Fading Channels

• Point (0, nt nr ): for a spatial

multiplexing gain of zero (i.e., R is
fixed), the maximal diversity gain (0,nt nr )
achievable is nt nr .

Diversity gain gd*(gs , )

∞
• Point (min {nt , nr } , 0):
transmitting at diversity gain gd⋆ = 0 (1,(nt −1)(nr −1))
(i.e., Pout is kept fixed) allows the
data rate to increase with SNR as
n = min {nt , nr }. (2,(nt −2)(nr −2))

• Intermediate points: possible to (gs , (nt − gs)(nr − gs))

transmit at non-zero diversity and
multiplexing gains but that any (min(nt ,nr ),0)

increase of one of those quantities Spatial multiplexing gain gs

leads to a decrease of the other
quantity.

114 / 494
Diversity-Multiplexing Trade-Off in I.I.D. Rayleigh Slow
Fading Channels

• For fixed rates R = 2, 4, ..., 40 0

10
bits/s/Hz, gs =1.75
g *d =0.25
– The asymptotic slope of each −1
10
curve is four and matches the

Outage probability Pout

maximum diversity gain gd⋆ (0, ∞). gs =1.5
g*d =0.5
−2
10

– The horizontal separation is 2

gs =1.25
bits/s/Hz per 3 dB, which −3
10 g*d =0.75
corresponds to the maximum
multiplexing gain equal to n(= 2). gs =1.0
−4
10 g*d =1.0
gs =0.75
g s =0.5 g*d =1.75
g*d =2.5
• As the rate increases more rapidly −5
10
with SNR (i.e., as the multiplexing 0 10 20 30 40 50 60
SNR [dB]
gain gs increases), the slope of the
outage probability curve (given by Figure: 2 × 2 MIMO i.i.d. Rayleigh fading channels
the diversity gain gd∗ ) vanishes.

115 / 494
Diversity-Multiplexing Trade-Off of a Scalar Rayleigh
Channel h
• Determine for a transmission rate R scaling with ρ as gs log2 (ρ), the rate at which
the outage probability decreases with ρ as ρ increases.
• Outage probability
 

Pout (R) = P log2 1 + ρ |h|2 < gs log2 (ρ)


= P 1 + ρ |h|2 < ρgs

• At high SNR,  
Pout (R) ≈ P |h|2 ≤ ρ−(1−gs )


• Since |h|2 is exponentially distributed, i.e., P |h|2 ≤ ǫ ≈ ǫ for small ǫ

Pout (R) ≈ ρ−(1−gs )

An outage occurs at high SNR when |h|2 ≤ ρ−(1−gs ) with a probability ρ−(1−gs ) .
• DMT for the scalar Rayleigh fading channel gd⋆ (gs , ∞) = 1 − gs for gs ∈ [0, 1].

116 / 494
Space-Time Coding over I.I.D. Rayleigh Flat
Fading Channels

117 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 6
Section: 6.1, 6.2, 6.3 (except “Antenna
Selection” in 6.3.2), 6.4.1, 6.4.2 (except
the Proofs), 6.5.1, 6.5.2, 6.5.3, 6.5.4,
6.5.8, Figure 7.1

118 / 494
Introduction - Previous Lectures

• Previous lecture
– Capacity of deterministic MIMO channels
 
C (H) = max log2 det Inr + ρHQHH .
Q≥0:Tr{Q}=1

– Ergodic capacity of fast fading channels

– Outage capacity and probability of slow fading channels

• MIMO provides huge gains in terms of reliability and transmission rate

– diversity gain, array gain, coding gain, spatial multiplexing gain, interference
management

• What we further need

– practical methodologies to achieve these gains?
– how to code across space and time?
– Some preliminary answers: multimode eigenmode transmission when channel knowledge
available at the Tx, Alamouti scheme when no channel knowledge available at the Tx

119 / 494
Overview of a Space-Time Encoder

• Space-time encoder: sequence of two black boxes

space-time encoder
T
temporal coding
B bits Q symbols
time interleaving space-time nt codeword C
coding
symbol mapping
nt

• First black box: combat the randomness created by the noise at the receiver.
• Second black box: spatial interleaver which spreads symbols over several antennas in
order to mitigate the spatial selective fading.
• The ratio B/T is the signaling rate of the transmission.
• The ratio Q/T is defined as the spatial multiplexing rate (representative of how
many symbols are packed within a codeword per unit of time).

120 / 494
System Model

• MIMO system with nt transmit and nr receive antennas over a frequency flat-fading
channel
• Transmit a codeword C = [c0 . . . cT −1 ] [nt × T ] contained in the codebook C
• At the kth time instant, the transmitted and received signals are related by
√
yk = Es Hk ck + nk
where
– yk is the nr × 1 received signal vector,
– Hk is the nr × nt channel matrix,
– nk is a nr × 1 zero mean complex AWGN vector with E{nk nH 2
l } = σn Inr δ (k − l),
– 2
The parameter Es is the energy normalization factor. SNR ρ = Es /σn .
• No transmit channel knowledge but we
 know
 it
is i.i.d. Rayleigh fading.
• Codeword average transmit power E Tr CCH = T . Assume

E kHk2F = nt nr .
• Channel time variation:
– slow fading : Tcoh >> T and {Hk = Hw }T −1
k=0 , with Hw denoting an i.i.d. random
fading matrix with unit variance circularly symmetric complex Gaussian entries.
 T −1
– fast fading : T ≥ Tcoh and Hk = Hk,w , where Hk,w k=0 are uncorrelated matrices,

each Hk,w being an i.i.d. random fading matrix with unit variance circularly
symmetric complex Gaussian entries.
121 / 494
Error Probability Motivated Design Methodology

• With instantaneous channel realizations perfectly known at the receive side, the ML
decoder computes an estimate of the transmitted codeword according to
T −1
X √ 2

Ĉ = arg min yk − Es Hk ck
C
k=0

where the minimization is performed over all possible codeword vectors C.

• Pairwise Error Probability (PEP): probability that the ML decoder decodes the
codeword E = [e0 . . . eT −1 ] instead of the transmitted codeword C.
• When the PEP is conditioned on the channel realizations {Hk }Tk=0 −1
, it is defined as
the conditional PEP,
vu T −1 
  uρ X
T −1 2
P C → E| {Hk }k=0 = Q t kHk (ck − ek )kF 
2
k=0

where Q (x) is the Gaussian Q-function.

• The average PEP, P (C → E), obtained by averaging the conditional PEP over the
probability distribution of the channel gains.
• System performance dominated at high SNR by the couples of codewords that lead
to the worst PEP.
122 / 494
• Assume a fixed rate transmission, i.e., spatial multiplexing gain gs = 0.
Definition

The diversity gain gdo (ρ) achieved by a pair of codewords {C, E} ∈ C is

defined as the slope of P (C → E) as a function of the SNR ρ on a log-log
scale, usually evaluated at very high SNR, i.e.,
log2 (P (C → E))
gdo (∞) = lim gdo (ρ) = − lim .
ρ→∞ ρ→∞ log2 ρ

PS: gdo (∞) ↔ P (C → E) , gd⋆ (0, ∞) ↔ Pout .

Definition

The coding gain achieved by a pair of codewords {C, E} ∈ C is defined as the

magnitude of the left shift of the P (C → E) vs. ρ curve evaluated at very
high SNR.

• If P (C → E) is well approximated at high SNR by

o
P (C → E) ≈ c (gc ρ)−gd (∞)
with c being a constant, gc is identified as the coding gain. 123 / 494
Fast Fading MIMO Channels

• In i.i.d. Rayleigh fast fading channels, average PEP reads as

Z π/2 TY
−1
1
−nr
P (C → E) = 1 + η kck − ek k2 dβ
π 0 k=0

where η = ρ/(4 sin2 β).

• Upper bound using the Chernoff bound
TY
−1  −nr
ρ
P (C → E) ≤ 1+ kck − ek k2 .
4
k=0

• In the high SNR regime, the average PEP is further upper-bounded by

 ρ −nr lC,E Y
P (C → E) ≤ kck − ek k−2nr
4
k∈τC,E

with lC,E the effective length of the pair of codewords {C, E}, i.e., lC,E = ♯τC,E
with τC,E = {k | ck − ek 6= 0}. Q
• Diversity gain: nr lC,E , coding gain: k∈τC,E kck − ek k−2nr .

124 / 494
The Distance-Product Criterion

• At high SNR, the error probability is naturally dominated by the worst-case PEP

Design Criterion
( Distance-product criterion) Over i.i.d. Rayleigh fast fading channels,
1 distance criterion: maximize the minimum effective length Lmin of the code
over all pairs of codewords {C, E} with C 6= E
Lmin = min lC,E
C,E
C6=E

2 product criterion: maximize the minimum product distance dp of the code over
all pairs of codewords {C, E} with C 6= E
Y
dp = min kck − ek k2
C,E
C6=E k∈τC,E
lC,E =Lmin

• The presence of multiple antennas at the transmitter does not impact the achievable
diversity gain gdo (∞) = nr Lmin but improves the coding gain gc = dp .
• The diversity gain is maximized first, and the coding gain is maximized only in a
second step.

125 / 494
Slow Fading MIMO Channels

• In i.i.d. Rayleigh fast fading channels, average PEP reads as

Z
1 π/2 h  i−nr
P (C → E) = det Int + η Ẽ dβ (1)
π 0
where Ẽ , (C − E) (C − E)H .
• Upper bound using the Chernoff bound
h  ρ i−nr
P (C → E) ≤ det Int + Ẽ
4
r (Ẽ)
Y  ρ
−nr
= 1 + λi Ẽ
i=1
4




with r Ẽ denotes the rank of the error matrix Ẽ and {λi Ẽ } for i = 1, . . . , r Ẽ
the set of its non-zero
eigenvalues.
• At high SNR, ρ4 λi Ẽ >> 1
(Ẽ)
 ρ −nr r(Ẽ) rY

P (C → E) ≤ λi−nr Ẽ
4 i=1

Qr(Ẽ)

• diversity gain: nr r Ẽ , coding gain: i=1 λi Ẽ .
126 / 494
The Rank-Determinant Criterion

Design Criterion

( Rank-determinant criterion) Over i.i.d. Rayleigh slow fading channels,

1 rank criterion: maximize the minimum rank rmin of Ẽ over all pairs of
codewords {C, E} with C 6= E


rmin = min r Ẽ
C,E
C6=E

2 determinant criterion: over all pairs of codewords {C, E} with C 6= E,

maximize the minimum of the product dλ of the non-zero eigenvalues of
Ẽ,
r (Ẽ)
Y

dλ = min λi Ẽ .
C,E
C6=E i=1

If rmin = nt , the determinant criterion comes to maximize the minimum

determinant of the error matrix over all pairs of codewords {C, E} with C 6= E


dλ = min det Ẽ .
C,E
C6=E

127 / 494
The Rank-Determinant Criterion

Definition
A full-rank (a.k.a. full-diversity) code is characterized by rmin = nt . A
rank-deficient code is characterized by rmin < nt .

Example
Rank-deficient and full-rank codes for nt = 2
• Rank-deficient code
   
1 c1 1 e1
C= √ ,E = √
2 c2 2 e2
• Full-rank code
   
1 c1 −c∗2 1 e1 −e∗2
C= √ ,E = √
2 c2 c∗1 2 e2 e∗1

 
H 1 |c1 − e1 |2 + |c2 − e2 |2 0
(C − E) (C − E) =
2 0 |c1 − e1 |2 + |c2 − e2 |2

128 / 494
The Rank-Determinant Criterion

Example
Question: Relying on the rank-determinant criterion, show that delay diversity
achieves full diversity. Assume for simplicity two transmit antennas.
Answer: The codeword for delay diversity can be written as
 
1 c1 c2 . . . cT −1 0
C= √ .
2 0 c1 c2 ... cT −1

Taking another codeword E, different from C,

 
1 e1 e2 . . . eT −1 0
E= √ .
2 0 e1 e2 ... eT −1

The diversity gain is given by the minimum rank of the error matrix over all
possible pairs of (different) codewords, i.e.



rmin = min r Ẽ = min r C − E .
C,E C,E
C6=E C6=E

129 / 494
The Rank-Determinant Criterion

Example
With delay diversity, we have
 
1 c1 − e1 c2 − e2 ... cT −1 − eT −1 0
C−E= √ .
2 0 c1 − e1 c2 − e2 ... cT −1 − eT −1



Obviously, r C − E ≤ 2. Actually, r C − E = 2 as long as C 6= E. Indeed
even in the case where all ck − ek = 0 except for one index k (in order to keep
C 6= E), e.g. k = 1,
 
1 c1 − e1 0 ... 0 0
C−E= √ ,
2 0 c1 − e1 0 ... 0

the rank is equal to 2. Hence diversity gain of 2nr .

130 / 494
The Rank-Determinant Criterion

Example
Question: Assume that c1 , c2 , c3 and c4 are constellation symbols taken from a
unit average energy QAM constellation. Consider the Linear Space-Time Block
Code, characterized by codewords
 
1 c1 + c3 c2 + c4
C= .
2 c2 − c4 c1 − c3
What is the diversity gain achieved by this code over slow Rayleigh fading
channels?
Answer: Check the rank of its error matrix
 
1 d1 + d3 d2 + d4
C−E=
2 d2 − d4 d1 − d3
where dk = ck − ek for k = 1, ..., 4. This code is rank deficient. It is easily seen
that by taking two codewords C and E such that d3 = d4 = 0 and
d1 = d2 = d (which is encountered for any constellations), r (C − E) = 1.
Hence diversity gain of nr .

131 / 494
Information Theory Motivated Design Methodology: Fast
Fading MIMO Channels - Achieving The Ergodic Capacity
• Recall Lecture 5&6: ergodic capacity
n  o
C̄ = max E log2 det Inr + ρHQHH .
Q:Tr{Q}=1

• Perfect Transmit Channel Knowledge

– transmit independent streams in the directions of the eigenvectors of the channel
matrix H.
– For a total transmission rate R, each stream k can then bePencoded using a
capacity-achieving Gaussian code with rate Rk such that n k=1 Rk = R, ascribed a
power λk (Q) and be decodedindependently
of the other streams.
– The optimal power allocation λ⋆k based on the water-filling allocation strategy.
– Capacity achievable using a variable-rate coding strategy (T = Tcoh is enough as long
as the noise can be averaged out).
• Partial Transmit Channel Knowledge
– When the channel is i.i.d. Rayleigh fading, Q = (1/nt ) Int .
– Transmission of independent information symbols may be performed in parallel over n
virtual spatial channels.
– The transmitter is very similar to the CSIT case except that all eigenmodes now receive
the same amount of power.
– Transmit with uniform power allocation over nt independent streams, each stream
using an AWGN capacity-achieving code and perform joint ML decoding (independent
decoding of all streams is clearly suboptimal due to interference between streams).
132 / 494
Information Theory Motivated Design Methodology: Slow
Fading MIMO Channels Achieving The DMT
• Impossible to code over a large number of independent channel realizations →
separate coding leads to an outage as soon as one of the subchannels is in deep fade.
• Joint coding across all subchannels necessary in the absence of transmit channel
knowledge!
• Rank-determinant criterion focuses on diversity maximization under fixed rate.
• What if we want to design codes achieving the diversity-multiplexing trade-off?
Definition


A scheme C ρ , i.e., a family of codes indexed by the SNR ρ, is said to
achieve a diversity gain gd (gs , ∞) and a multiplexing gain gs at high SNR if


R ρ
lim
= gs
ρ→∞ log
2 ρ



log2 Pe ρ
lim
= −gd (gs , ∞)
ρ→∞ log2 ρ



where R ρ is the data rate and Pe ρ the average error probability averaged
over the additive noise, the i.i.d. channel statistics and the transmitted
codewords. The curve gd (gs , ∞) is the diversity-multiplexing trade-off achieved
by the scheme in the high SNR regime. 133 / 494
Space-Time Block Coding (STBC)

• STBCs can be seen as a mapping of Q symbols (complex or real) onto a codeword

C of size nt × T .

• Codewords are uncoded in the sense that no error correcting code is contained in the
STBC.

• Linear STBCs are by far the most widely used

– Spread information symbols in space and time in order to improve either the diversity
gain, either the spatial multiplexing rate (rs = Q
T
) or both the diversity gain and the
spatial multiplexing rate.
– Pack more symbols into a given codeword, i.e., increase Q, to increase the data rate.

Example
Alamouti code: nt = 2, Q = 2, T = 2, rs = 1
 
1 c1 −c∗2
C= √ ∗ .
2 c2 c1

134 / 494
A General Framework for Linear STBCs

• A linear STBC is expressed in its general form as

Q
X
C= Φq ℜ[cq ] + Φq+Q ℑ[cq ]
q=1

where
– Φq are complex basis matrices of size nt × T ,
– cq stands for the complex information symbol (taken for example from PSK or QAM
constellations),
– Q is the number of complex symbols cq transmitted over a codeword,
– ℜ and ℑ stand for the real and imaginary parts.

Definition
Tall (T ≤ nt ) unitary basis matrices are such that ΦH 1
q Φq = Q I T
∀q = 1, . . . , 2Q. Wide (T ≥ nt ) unitary basis matrices are such that
Φq ΦH T
q = Qnt Int ∀q = 1, . . . , 2Q.

Definition
Q
The spatial multiplexing rate of a space-time block code is defined as rs = T
.
A full rate space-time block code is characterized by rs = nt .
135 / 494
A General Framework for Linear STBCs
√
• Apply the vec operator to yk = Es Hk ck + nk and make use of STBC structure
√
Y = Es HX S + N
where
– Y[2nr T × 1] is the channel output vector
   
ℜ  y0 ... yT −1 
Y = vec ,
ℑ y0 ... yT −1
– H[2nr T × 2nt T ] is the block diagonal channel
 
ℜ [H] −ℑ [H]
H = IT ⊗ H ′ , where H′ = ,
ℑ [H] ℜ [H]
– X [2nt T × 2Q] is the linear code matrix
       
ℜ [Φ1 ] ℜ Φ2Q 
X = vec ··· vec ,
ℑ [Φ1 ] ℑ Φ2Q
– S[2Q × 1] is a block of uncoded input symbols
     T
S = ℜ [c1 ] · · · ℜ cQ ℑ [c1 ] ··· ℑ cQ ,
– N [2nr T × 1] is the noise vector
   
ℜ  n0 ... nT −1 
N = vec ,
ℑ n0 ... nT −1
136 / 494
A General Framework for Linear STBCs

• Average Pairwise Error Probability of STBCs

Proposition

A PSK/QAM based linear STBC consisting of unitary basis matrices minimizes

the worst-case PEP (1) averaged over i.i.d. Rayleigh slow fading channels if
(sufficient condition) the unitary basis matrices {Φq }2Q
q=1 satisfy the conditions

Φq ΦH H
p + Φp Φq = 0nt , q 6= p for wide {Φq }2Q
q=1 ,

ΦH H
q Φp + Φp Φq = 0T , q 6= p for tall {Φq }2Q
q=1 .

Proof: Using Hadamard’s inequality and the unitarity of basis matrices,

 

T
min min det Int + η Ẽ ≤ det Int + η Int d2min .
q=1,...,Q dq Qnt
Equality if unitary basis matrices are skew-hermitian. 

137 / 494
A General Framework for Linear STBCs

• Ergodic Capacity in i.i.d. Rayleigh Fading Channels

1 n  ρ o
C̄ = max EH log det I2nr T + HX X T HT
Tr{X X T }≤2T 2T 2

where it is assumed without loss of generality that E SS H = I2Q .

Proposition

In i.i.d. Rayleigh fading channels, linear STBCs with wide (Q ≥ nt T ) matrices

X that satisfy
1
XXT = I2nt T
nt
are capacity-efficient.

By capacity-efficient, we mean a code that maximizes the average mutual

information in the sense that it preserves the capacity without inducing any loss of
capacity if concatenated with capacity-achieving outer codes.

138 / 494
A General Framework for Linear STBCs

• Decoding

Proposition

Applying the space-time matched filter X H HH to the output vector Y

decouples the transmitted symbols
√
Es T
X H HH Y = kHk2F I2T S + X H HH N
Qnt
if and only if the basis matrices are wide unitary
T
Φq ΦH
q = In , ∀q = 1, . . . , 2Q
Qnt t
and pairwise skew-hermitian

Φq ΦH H
p + Φp Φq = 0nt , ∀q 6= p.

The complexity of ML decoding of linear STBCs grows exponentially with nt and Q.

The decoupling property allows each symbol to be decoded independently of the
presence of the other symbols through a simple space-time matched filter.
139 / 494
A General Framework for Linear STBCs

Example

A code such that T = 1, nt = 2, Q = 2, rs = 2 with the following tall basis

matrices
       
1 1 1 0 1 j 1 0
Φ1 = √ , Φ2 = √ , Φ3 = √ , Φ4 = √ ,
2 0 2 1 2 0 2 j
or equivalently, with the following matrix X
 
1 0 0 0
1  0 1 0 0 
X = √  .
2 0 0 1 0 
0 0 0 1

This code is called Spatial Multiplexing. Optimal for worst-case PEP min.
capacity-efficient, large decoding complexity.

140 / 494
A General Framework for Linear STBCs

Example
A code such that T = 2, nt = 2, Q = 2, rs = 1 with the following wide basis
matrices
   
1 1 0 1 0 −1
Φ1 = √ , Φ2 = √ ,
2 0 1 2 1 0
   
1 j 0 1 0 j
Φ3 = √ , Φ4 = √ ,
2 0 −j 2 j 0
or equivalently, with the following matrix X
 
1 0 0 0
 0 1 0 0 
 
 0 0 1 0 
 
1  0 0 0 1 
X = √  .
2 0 −1 0 0 

 1 0 0 0 
 
 0 0 0 1 
0 0 −1 0

This code is called Alamouti code. Optimal for worst-case PEP min. not
capacity-efficient, low decoding complexity. 141 / 494
Spatial Multiplexing/V-BLAST/D-BLAST

• Spatial Multiplexing (SM), also called V-BLAST, is a full rate code (rs = nt ) that
consists in transmitting independent data streams on each transmit antenna.
• In uncoded transmissions, we assume one symbol duration (T = 1) and codeword C
is a symbol vector of size nt × 1.
• From previous results on error probability and capacity,
Proposition
Spatial Multiplexing with basis matrices characterized by square X such that
1
XTX = I2nt
nt
is capacity-efficient (Proposition 6) and optimal from an error rate
minimization perspective (Proposition 5).

Example

nt
1  T 1 X
C= √ c1 ... c nt = √ In (:, q) ℜ [cq ] + jInt (:, q) ℑ [cq ] .
nt nt q=1 t

Each element cq is a symbol chosen from a given constellation. 142 / 494

ML decoding

• Error probability
Z π/2 h  i−nr
1
P (C → E) = det Int + η Ẽ dβ
π 0
Z nt
!−nr
π/2
1 η X
= 1+ |cq − eq |2 dβ
π 0 nt q=1
 −nr nt
!−nr
ρ X 2
≤ |cq − eq |
4nt q=1

The SNR exponent is equal to nr . Due to the lack of coding across transmit
antennas, no transmit diversity is achieved and only receive diversity is exploited.
• Over fast fading channels, we know that it is not necessary to code across antennas
to achieve the ergodic capacity.

Proposition
Spatial Multiplexing with ML decoding and equal power allocation achieves the
ergodic capacity of i.i.d. Rayleigh fast fading channels.

143 / 494
ML decoding

• Over slow fading channels, what is the multiplexing-diversity trade-off achieved by

SM with ML decoding?

Proposition

For nr ≥ nt , the diversity-multiplexing trade-off at high SNR achieved by

Spatial Multiplexing with ML decoding and QAM constellation over i.i.d.
Rayleigh fading channels is given by
 
gs
gd (gs , ∞) = nr 1 − , gs ∈ [0, nt ] .
nt

144 / 494
Zero-Forcing (ZF) Linear Receiver

• MIMO ZF receiver acts similarly to a ZF equalizer in frequency selective channels.

• ZF filtering effectively decouples the channel into nt parallel channels
– interference from other transmitted symbols is suppressed
– scalar decoding may be performed on each of these channels

• The complexity of ZF decoding similar to SISO ML decoding, but the inversion step
is responsible for the noise enhancement (especially at low SNR).
√  T
• Assuming that a symbol vector C = 1/ nt c1 . . . cnt is transmitted, the
output of the ZF filter GZF is given by
 T
z = GZF y = c1 . . . cnt + GZF n

where GZF inverts the channel,

r
nt †
GZF = H
Es

−1
with H† = HH H HH denoting the Moore-Penrose pseudo inverse.

145 / 494
Zero-Forcing (ZF) Linear Receiver

• Covariance matrix of the noise at the output of the ZF filter

n o nt †  † H nt  H −1
E GZF n (GZF n)H = H H = H H .
ρ ρ

• The output SNR on the q th subchannel is thus given by

ρ 1
ρq = , q = 1, . . . , nt .
nt (HH H)−1 (q, q)

• Inversion leads to noise enhancement. Severe degradation at low SNR.

• Assuming that the channel is i.i.d. Rayleigh distributed, ρq is a χ2 random variable
with 2 (nr − nt + 1) degrees of freedom, denoted as χ22(nr −nt +1) . The average PEP
on the q th subchannel is thus upper-bounded by
 −(nr −nt +1)
ρ
P (cq → eq ) ≤ |cq − eq |−2(nr −nt +1) .
4nt
The lower complexity of the ZF receiver comes at the price of a diversity gain limited
to nr − nt + 1. Clearly, the system is undetermined if nt > nr .

146 / 494
Zero-Forcing (ZF) Linear Receiver

• In fast fading channels, the average maximum achievable rate C̄ZF is equal to the
sum of the maximum rates achievable by all layers
min{nt ,nr }
X
C̄ZF = E {log2 (1 + ρq )}
q=1
 
(ρր) ρ 

≈ min {nt , nr } log2 + min {nt , nr } E log2 χ22(nr −nt +1) .
nt
Note the difference with
ρ X n n o
C̄CDIT ≈ nlog2 + E log2 (χ22(N −n+k) ) .
nt
k=1

Spatial Multiplexing in combination with a ZF decoder allows for transmitting over

n = min {nt , nr } independent data pipes.

147 / 494
Zero-Forcing (ZF) Linear Receiver

• In slow fading channels, what is the diversity-multiplexing trade-off achieved by SM

with ZF?
Proposition

For nr ≥ nt , the diversity-multiplexing trade-off achieved by Spatial

Multiplexing with QAM constellation and ZF filtering in i.i.d. Rayleigh fading
channels is given by
 
gs
gd (gs , ∞) = (nr − nt + 1) 1 − , gs ∈ [0, nt ] .
nt

148 / 494
Zero-Forcing (ZF) Linear Receiver

• ZF receiver maximizes the SNR under the constraint that the interferences from all
other layers are nulled out.
– For a given layer q, the ZF combiner gq is such that this layer is detected through a
projection of the output vector y onto the direction closest to H (:, q) within the
subspace orthogonal to the one spanned by the set of vectors H (:, p), p 6= q.
• Assume the following system model with nr ≥ nt
y = Hc + n,
X
= hq c q + hp c p + n
p6=q

th
where hq is the q column of H.
• Let us build the following nr × (nt − 1) matrix by collecting all hp with p 6= q:
 
H−q = . . . hp . . . p6=q ,
 
= U′ Ũ ΛVH

where Ũ is the matrix containing the left singular vectors corresponding to the null
singular values. Similarly we define
 T
c−q = . . . cp . . . p6=q .

149 / 494
Zero-Forcing (ZF) Linear Receiver

• By multiplying by ŨH , we project the output vector onto the subspace orthogonal to
the one spanned by the columns of H′

ŨH y = ŨH hq cq + ŨH H−q c−q + ŨH n

= ŨH hq cq + ŨH n.

• To maximize the SNR, noting the noise is still white, we match to the effective
channel ŨH hq such that
 H  H
z = ŨH hq ŨH hq cq + ŨH hq ŨH n
 H
and the ZF combiner is gq = ŨH hq ŨH = hH H
q ŨŨ .

150 / 494
Minimum Mean Squared Error (MMSE) Linear Receiver

• Filter maximizing the SINR. Minimize the total resulting noise: find G such that
  T 2
E Gy − c1 . . . cnt is minimum.
• The combined noise plus interference signal ni,q when estimating symbol cq writes as
r
X Es
ni,q = hp cp + n.
nt
p6=q

The covariance matrix of ni,q reads as

n o X Es
Rni,q = E ni,q nH i,q = σn2 Inr + hp hH
p
nt
p6=q

and the MMSE combiner for stream q is given by

r  −1
Es H  2 X Es H
gM M SE,q = h q σn I nr + hp hp .
nt nt
p6=q

• An alternative and popular representation of the MMSE filter can also be written as
r  −1 r  −1
nt nt nt H nt
GM M SE = H H H + I nt HH = H HHH + Inr
Es ρ Es ρ
• Bridge between matched filtering at low SNR and ZF at high SNR.
151 / 494
Minimum Mean Squared Error (MMSE) Linear Receiver

• The output SINR on the q th subchannel (stream) is given by

 −1
Es H  2 X Es H
ρq = h q σn I nr + hp hp hq .
nt nt
p6=q

• At high SNR, the MMSE filter is practically equivalent to ZF and the diversity
achievable is thus limited to nr − nt + 1.

152 / 494
Successive Interference Canceler

• Successively decode one symbol (or more generally one layer/stream) and cancel the
effect of this symbol from the received signal.
• Decoding order based on the SINR of each symbol/layer: the symbol/layer with the
highest SINR is decoded first at each iteration.
• SM with (ordered) SIC is generally known as V-BLAST, and ZF and MMSE
V-BLAST refer to SM with respectively ZF-SIC and MMSE-SIC receivers.

• The diversity order experienced by the decoded layer is increased by one at each
iteration. Therefore, the symbol/layer detected at iteration i will achieve a diversity
of nr − nt + i.
• Major issue: error propagation
– The error performance is mostly dominated by the weakest stream.
– Non-ordered SIC: diversity order approximately nr − nt + 1.
– Ordered SIC: performance improved by reducing the error propagation caused by the
first decoded stream. The diversity order remains lower than nr .

153 / 494
Successive Interference Canceler
(∗) 2
1 Initialization: i ←− 1, y(1) = y,G(1) = GZF (H),q1 = arg minj G(1) (j, :)
where GZF (H) is defined as the ZF filter of the matrix H.
2 Recursion:
1 step 1: extract the qith transmitted symbol from the received signal y(i)

c̃qi = G(i) (qi , :) y(i)

where G(i) (qi , :) is the qith row of G(i) ;
2 step 2: slice c̃qi to obtain the estimated transmitted symbol ĉqi ;
3 step 3: assume that ĉqi = cqi and construct the received signal
s
Es
y(i+1) = y(i) − H (:, qi ) ĉqi
nt


G(i+1) = GZF Hqi
i ←− i + 1
(∗) (i+1) 2
qi+1 = arg min G (j, :)
j ∈{q
/ 1 ,...,qi }

where Hqi
is the matrix obtained by zeroing columns q1 , . . . , qi of H. Here

GZF Hqi denotes the ZF filter applied to Hqi .

154 / 494
Successive Interference Canceler

• In fast fading channels, the maximum rate achievable with ZF-SIC

min{nt ,nr }
X
C̄ZF −SIC = E {log2 (1 + ρq )}
q=1
  min{nt ,nr }
(ρր) ρ X 

≈ min {nt , nr } log2 + E log2 χ22(nr −nt +q) = C̄CDIT
nt q=1

The loss that was observed with ZF filtering is now compensated because the
successive interference cancellation improves the SNR of each decoded layer.

Proposition
Spatial Multiplexing with ZF-SIC (ZF V-BLAST) and equal power allocation
achieves the ergodic capacity of i.i.d. Rayleigh fast fading MIMO channels at
asymptotically high SNR.

This only holds true when error propagation is neglected.

155 / 494
Successive Interference Canceler

• MMSE-SIC does better for any SNR

min{nt ,nr }   
X ρ
C̄M M SE−SIC = E {log2 (1 + ρq )} = E log2 det Inr + HHH = Īe ,
q=1
nt

Proposition

Spatial Multiplexing with MMSE-SIC (MMSE V-BLAST) and equal power

allocation achieves the ergodic capacity for all SNR in i.i.d. Rayleigh fast fading
MIMO channels.

Result also valid for a deterministic channel.

156 / 494
Successive Interference Canceler

• In slow fading channels, what is the diversity-multiplexing trade-off achieved by

unordered ZF-SIC?
Proposition

For nr ≥ nt , the diversity-multiplexing trade-off achieved by Spatial

Multiplexing with QAM constellation and unordered ZF-SIC receiver over i.i.d.
Rayleigh fading channels is given by
 
gs
gd (gs , ∞) = (nr − nt + 1) 1 − , gs ∈ [0, nt ] .
nt

The achieved trade-off is similar to the trade-off achieved by a simple ZF receiver.

This comes from the fact that the first layer dominates the error probability since its
error exponent is the smallest.
• By increasing the number of receive antennas by 1,
– with ZF or unordered ZF-SIC, we can either accommodate one extra stream with the
same diversity order or increase the diversity order of every stream by 1,
– with ML, we can accommodate one extra stream and simultaneously increase the
diversity order of every stream by 1.

157 / 494
Impact of Decoding Strategy on Error Probability

• SM with ML, ordered and non ordered ZF SIC and simple ZF decoding in 2 × 2 i.i.d.
Rayleigh fading channels for 4 bits/s/Hz.
0
10
ML
ordered ZF SIC
ZF SIC
ZF

−1
10
BER

−2
10

−3
10

−4
10
0 5 10 15 20 25
SNR [dB]

The slope of the ML curve approaches 2. ZF only achieves a diversity order of

nr − nt + 1 = 1.
158 / 494
Impact of Decoding Strategy on Error Probability

• SM with ML, ZF and MMSE in i.i.d. Rayleigh slow fading channels with
nt = nr = 4 and QPSK.
0
10
ML
MMSE
ZF

−1
10
BER

−2
10

−3
10

−4
10
0 5 10 15 20 25
SNR [dB]

159 / 494
D-BLAST

• So far, transmission of independent data streams exploits a diversity order of at most

nr out of nt nr .
• Lack of coding across antennas: V-BLAST is in outage each time the SINR of a
layer cannot support the rate allocated to that layer.
• Need for a spatial interleaving so that each layer encounters all antennas.
• D-BLAST: V-BLAST transmitter + a stream rotation following the encoding of all
layers. Receiver is similar to V-BLAST.

Example
Consider two layers a and b and nt = 2. Assume that layer a is made of two
streams a(1) and a(2) and layer b of two streams as well b(1) and b(2) . Each
stream can be seen as a block of symbols. The transmitted codeword C is now
written as  (1) 
a b(1)
C= .
a(2) b(2)

160 / 494
Orthogonal Space-Time Block Codes

• O-STBC vs. SM
– Remarkable properties which make them extremely easy to decode: MIMO ML
decoding decouples into several SIMO ML decoding
– Achieve a full-diversity of nt nr .
– Much smaller spatial multiplexing rate than SM.
• Linear STBC characterized by the two following properties
1 the basis matrices are wide unitary
T
Φq ΦH
q = In ∀q = 1 . . . 2Q
Qnt t
2 the basis matrices are pairwise skew-hermitian
Φq ΦH H
p + Φp Φq = 0, q 6= p

or equivalently by this unique property

" Q #
H T X 2
CC = |cq | Int .
Qnt q=1

• Complex O-STBCs with rs = 1 only exist for nt = 2. For larger nt , codes exist with
rs ≤ 1/2. For some particular values of nt > 2, complex O-STBCs with
1/2 < rs < 1 have been developed. This is the case for nt = 3 and nt = 4 with
rs = 3/4.
161 / 494
Orthogonal Space-Time Block Codes

Example
Alamouti code: complex O-STBC for nt = 2 with a spatial multiplexing rate
rs = 1  
1 c1 −c∗2
C= √ ∗ .
2 c2 c1

• basis matrices are unitary and skew-hermitian (discussed before).

 
• CCH = 21 |c1 |2 + |c2 |2 I2 .
• rs = 1 since two symbols are transmitted over two symbol durations.

Example
For nt = 3 , a complex O-STBC expanding on four symbol durations (T = 4)
and transmitting three symbols on each block (Q = 3)
 
c −c∗2 c∗3 0
2 1 ∗ ∗
C= c2 c1 0 c3  .
3
c3 0 −c∗1 −c∗2

The spatial multiplexing rate rs is equal to 3/4.

162 / 494
Orthogonal Space-Time Block Codes

Proposition
O-STBCs enjoy the decoupling property.

Example
Assume a MISO transmission based on the Alamouti code
r  
  Es   c1 −c∗2  
y1 y2 = h1 h2 ∗ + n1 n2
2 c2 c1

or equivalently
  r     
y1 Es h1 h2 c1 n1
= + .
y2∗ 2 h∗2 −h∗1 c2 n∗2
| {z }
Hef f

163 / 494
Orthogonal Space-Time Block Codes

Example
Applying the space-time matched filter HH ef f to the received vector decouples
the transmitted symbols
    r    
z1 H y1 Es   c1 n1
= Hef f ∗ = |h1 |2 + |h2 |2 I2 + HHef f ∗ .
z2 y2 2 c2 n2

Expanding the original ML metric

 r 2  r 2
y1 − Es (h1 c1 + h2 c2 )  + y2 − Es (−h1 c∗2 + h2 c∗1 ) 
   
 2   2 

and making use of z1 and z2 , the decision metric for c1 is

 r 2  r 2
 Es 2 2

  Es 2 2


choose ci iff z1 −
 |h1 | + |h2 | ci  ≤ z1 −
 |h1 | + |h2 | ck  ∀i 6= k
2 2
and similarly for c2 . Independent decoding of symbols c1 and c2 is so
performed.

164 / 494
Orthogonal Space-Time Block Codes

• Error Probability
Z π/2 h  i−nr
1
P (C → E) = det Int + η Ẽ dβ
π 0
Z Q
!−nr nt
π/2
1 T X
= 1+η |cq − eq |2 dβ
π 0 Qnt q=1
 −nr nt Q
!−nr nt
(ρր) ρ T X 2
≤ |cq − eq | .
4 Qnt q=1

Full diversity gain of nt nr .

165 / 494
Orthogonal Space-Time Block Codes

• O-STBCs are not capacity efficient IO−ST BC (H) ≤ Ie (H)

– mutual information of MIMO channel
 r(H) r(H) !
ρ ρ Y  
Ie (H) = log2 1+ kHk2F + . . . + λk HHH .
nt nt k=1

– mutual information of MIMO channel transformed by the O-STBC

 
Q ρT
IO−ST BC (H) = log2 1 + kHk2F .
T Qnt

Proposition
For a given channel realization H, the mutual information achieved by any
O-STBC is always upper-bounded by the channel mutual information with
equal power allocation Ie . Equality occurs if and only if both the rank of the
channel and the spatial multiplexing rate of the code are equal to one.

Corollary
The Alamouti scheme is optimal with respect to the mutual information when
used with only one receive antenna.
166 / 494
Orthogonal Space-Time Block Codes

• Diversity-Multiplexing Trade-off Achieved by O-STBCs

Proposition
The diversity-multiplexing trade-off at high SNR achieved by O-STBCs using
QAM constellations in i.i.d. Rayleigh fading channels is given by
 
gs
gd (gs , ∞) = nr nt 1 − , gs ∈ [0, rs ] .
rs

Proposition
The Alamouti code with any QAM constellation achieves the optimal
diversity-multiplexing trade-off for two transmit and one receive antennas in
i.i.d. Rayleigh fading channels.

167 / 494
Orthogonal Space-Time Block Codes

• Block error rate for 4 different rates R = 4, 8, 12, 16 bits/s/Hz in 2 × 2 i.i.d. slow
Rayleigh fading channels.
0
10

−1
10
(2x2) Block Error Rate

−2
10

−3
10

−4
10
M=4, R=4 <−−−12dB−−−>
M=16, R=8
−5 M=64, R=12
10 M=256, R=16
0 10 20 30 40 50 60
SNR [dB]

– full diversity exploited: gd (gs = 0, ∞) = gdo (∞) = 4.

– the growth of the multiplexing gain is slow: 12 dB separate the curves, corresponding
to a multiplexing gain gs = 1, i.e., 1 bit/s/Hz increase per 3 dB SNR increase.
168 / 494
Other Code Constructions

• Quasi-Orthogonal Space-Time Block Codes

– increase the spatial multiplexing rate while still partially enjoying the decoupling
properties of O-STBCs
– use O-STBCs of reduced dimensions as the building blocks of a higher dimensional code

• Linear Dispersion Codes

– if a larger receiver complexity is authorized, it is possible to relax the skew-hermitian
conditions and increase the data rates while still providing transmit diversity.

• Algebraic Space-Time Codes

– structured codes using algebraic tools
– many of them are designed to achieve the optimal diversity-multiplexing tradeoff

169 / 494
Global Performance Comparison

• Asymptotic diversity-multiplexing trade-off gd (gs , ∞) achieved by several space-time

codes in a 2 × 2 i.i.d. Rayleigh fading MIMO channel.

4.5

4 Optimal trade−off achieved by

Delay diversity, Dayal code, Golden code,
3.5 Alamouti Tilted QAM, MMSE D−BLAST
Diversity gain gd (gs )

3
ZF D−BLAST
2.5

2 SM with ML

1.5
SM with ZF,
unordered ZF V−BLAST
1

0.5

0
0 0.5 1 1.5 2 2.5
Spatial multiplexing gain gs

170 / 494
Global Performance Comparison

• Bit error rate (BER) of several space-time block codes in i.i.d. slow Rayleigh fading
channels with nt = 2 and nr = 2 in a 4-bit/s/Hz transmission. ML decoding is used.

−1
10
SM
Alamouti
Dayal
TAST
B
2,φ
Tilted−QAM
−2
10
BER

−3
10

−4
10
10 12 14 16 18 20
SNR [dB]

171 / 494
Space-Time Coding in Real-World MIMO
Channels

172 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 8
Section: 8.1, 8.2, 8.3.1 (“Rayleigh Slow
Fading Channels”, “How Realistic is the
High SNR Regime Approximation?”),
8.3.4

173 / 494
Introduction - Previous Lectures

• Space-time coding in i.i.d. Rayleigh fading

– Distance-product criterion in fast fading
– Rank-determinant criterion in slow fading

• Real-world channels span a large variety of propagation conditions.

• Some environments may highly deviate from the i.i.d. Rayleigh fading scenario.

• Objectives:
– how codes developed under the i.i.d. Rayleigh assumption behave in more realistic
propagation conditions, i.e., how these codes are affected by non ideal propagation
conditions,
– how a more adapted design criterion might significantly improve their performance.

174 / 494
System Model

• MIMO system with nt transmit and nr receive antennas communicating through a

frequency flat-fading channel
• A codeword C = [c0 . . . cT −1 ] of size nt × T contained in the codebook C is
transmitted over T symbol durations via nt transmit antennas.
• At the kth time instant, the transmitted and received signals are related by
√
yk = Es Hk ck + nk
where
– yk is the nr × 1 received signal vector,
– Hk is the nr × nt channel matrix,
– nk is a nr × 1 zero mean complex additive white Gaussian noise (AWGN) vector with
E{nk nH 2
l } = σn Inr δ (k − l),
2
– The parameter Es is the energy normalization factor, so that the ratio Es /σn
represents the SNR denoted as ρ.
 
• We normalize the codeword average transmit power such that E Tr CCH =T

and assume for simplicity that E kHk2F = nt nr .
• Hk is not i.i.d. anymore!

175 / 494
Radiation Patterns
PL−1 (l) PL−1 (l) (l)

• Decompose the channel Hk = l=0 Hk = l=0 Hk (:, 1) aTt θt,k , where
(l)
(l)
at θt,k is the transmit array response in the direction of departure θt,k .
• PEP argument writes as
T −1 T

−1 L−1
2
X X
2 X (l) T (l)

kHk (ck − ek )kF = Hk (:, 1)at θt,k (ck − ek )

k=0 k=0 l=0

• The original MIMO transmission can be considered as the SIMO transmission of an

equivalent codeword, given at the kth time instant by

aTt ck
• It may be thought of as an array factor function of the transmitted codewords. At
every symbol period,
– the energy radiated in any direction varies as a function of the transmitted codewords.
– for a given codeword and omnidirectional antennas, the radiated energy is not uniformly
distributed in all directions, but may present maxima and minima in certain directions.

176 / 494
Radiation Patterns

177 / 494
Radiation Patterns

• What if the transmit angle spread decreases?

Definition

A MIMO channel is said to be degenerate in the direction of departure θt if all

scatterers surrounding the transmitter are located along the same direction θt .

In the presence of small angle spread at the transmit side, the MIMO channel
degenerates into a SIMO channel where the 1 × T transmitted codeword is given by
aTt (θt ) C
T −1 T −1 
"n #
X X 2 X r

kHk (ck − ek )k2F = T

|Hk (n, 1)|2 .
 
(ck − ek ) at (θt )
k=0 k=0 n=1

Since a space-time code designed for i.i.d. channels is only concerned with C and E,
its interaction with at (θt ) is not taken into account.

178 / 494
Radiation Patterns

Example
The Spatial Multiplexing example for nt = 2: ck = [ c1 [k] c2 [k] ]T
 
c2 [k] −2πj dλt cos θt
cTk at (θt ) = c1 [k] 1 + e
c1 [k]
| {z }
Gt (θt |ck )

0 phase shift π/2 phase shift

90 2 90 2 0
120 60 120 60 10

150 1 30 150 1 30

180 0 180 0

210 330 210 330

240 300 240 300

270 270 SER
−1
10

π phase shift 3π/2 phase shift

90 2 90 2
120 60 120 60

150 1 30 150 1 30

180 0 180 0
d/λ=0.5
d/λ=0.1
210 330 210 330 −2
10
−2 −1 0 1 2 3 4
240 300 240 300
270 270 Phase shift between the two transmitted symbols [rad]

179 / 494
Derivation of the Average PEP

• Conditional PEP
v
u T −1 
  uρ X
P C → E| {Hk }Tk=0
−1
= Q t kHk (ck − ek )k2F 
2
k=0

where Q (x) is the Gaussian Q-function defined as

Z ∞  2
∆ 1 y
Q (x) = P (y ≥ x) = √ exp − dy.
2π x 2
• Average PEP n  o
P (C → E) = EHk P C → E| {Hk }Tk=0
−1
.
• This integration is sometimes difficult to calculate. Therefore, alternatives forms of
the Gaussian Q-function are used.
– Craig’s formula
Z π/2  
1 x2
Q (x) = exp − dβ.
π 0 2 sin2 (β)
– Chernoff bound  
x2
Q (x) ≤ exp − .
2
180 / 494
Derivation of the Average PEP

• We can derive the average PEP as follows

n  o
P (C → E) = EHk P C → E| {Hk }Tk=0 −1

Z  
1 π/2 1
= MΓ − dβ
π 0 2 sin2 (β)
 
1
≤ MΓ −
2
P −1
with MΓ (γ) moment generating function (MGF) of Γ = ρ2 Tk=0 kHk (ck − ek )k2F
Z ∞
∆
MΓ (γ) = exp (γΓ) pΓ (Γ) dΓ
0

Theorem
The moment generating function of a Hermitian quadratic form in complex
Gaussian random variable y = zFzH , where z is a circularly symmetric complex
Gaussian vector with mean z̄ and a covariance matrix Rz and F a Hermitian
matrix, is given by


exp sz̄F (I − sRz F)−1 z̄H
Z ∞
∆
My (s) = exp (sy) py (y) dy =
0 det (I − sRz F) 181 / 494
Derivation of the Average PEP

• Apply to a joint Space-Time Correlated Ricean Fading Channels. Defining

r r
K
1  
H= 11×T ⊗ H̄ + H̃1 H̃2 · · · H̃T
1+K 1+K
D = diag {c1 − e1 , c2 − e2 , . . . , cT − eT } ,
we may write
T
X n o  H  
kHk (ck − ek )k2F = kHDk2F = Tr HDDH HH = vec HH ∆ vec HH
k=1

where ∆ = Inr ⊗ DDH . This is a hermitian quadratic form of complex Gaussian

random variables where
• Define
 H 
H̃ = vec H̃0 H̃1 · · · H̃T −1 ,

H 
H̄ = vec 11×T ⊗ H̄ ,
and 
Ξ = E H̃H̃H
as the spatio-temporal correlation matrix.
182 / 494
Derivation of the Average PEP

• PEP averaged over the space-time correlated Ricean fading channel

Z π/2  
1
P (C → E) = exp −ηK H̄H ∆ (IT nr nt + ηΞ∆)−1 H̄
π 0

(det (IT nr nt + ηΞ∆))−1 dβ

where the effective SNR η is defined as η = ρ/(4 sin2 (β) (1 + K)).

• PEP averaged over the space-time correlated Rayleigh fading channel

Z
1 π/2
P (C → E) = (det (IT nr nt + ηΞ∆))−1 dβ
π 0

where η = ρ/(4 sin2 (β)).

183 / 494
Average PEP in Rayleigh Slow Fading Channels

Example
Question: Assume a spatially correlated slow Rayleigh fading MIMO channel.
Derive the Average PEP for ML receiver.
Answer: The conditional PEP writes as
r 
ρ
P (C → E | H) = Q kH (C − E)k2F .
2
The average PEP over Rayleigh slow fading channels is
Z  
1 π/2 1
P (C → E) = EH {P (C → E|H)} = MΓ − dβ
π 0 2 sin2 (β)

where MΓ (γ) moment generating function (MGF) of Γ = ρ2 kH (C − E)k2F .

Note that
n o  H    
kH (C − E)k2F = Tr HẼHH = vec HH Inr ⊗ Ẽ vec HH

where Ẽ = (C − E) (C − E)H .

184 / 494
Average PEP in Rayleigh Slow Fading Channels

Example
This is a hermitian quadratic form of complex gaussian random variables of the

H
form zFzH (with z = vec HH and F = Inr ⊗ Ẽ) and we can use Theorem
where the mean
 z̄ = 0
is the zero vector

H and the covariance matrix is
Rz = R = E vec HH vec HH .
We then write (with η = ρ/(4 sin2 (β)))
Z
1 π/2
P (C → E) = (det (Inr nt + ηCR ))−1 dβ.
π 0
where CR = RF. With the Kronecker model, R = Rr ⊗ Rt , and
Z
1 π/2    −1
P (C → E) = det Inr nt + η (Rr ⊗ Rt ) Inr ⊗ Ẽ dβ
π 0
Z
1 π/2   −1
= det Inr nt + ηRr ⊗ Rt Ẽ dβ.
π 0


185 / 494
Average PEP in Rayleigh Slow Fading Channels

• Average PEP writes as

Z π/2
1
P (C → E) = (det (Inr nt + ηCR ))−1 dβ
π 0
Z π/2 r(C
YR )
1
= (1 + ηλi (CR ))−1 dβ.
π 0 i=1
R π/2 Qr(CR )
• What happens at infinite SNR? P (C → E) ≈ 1
π 0
η −r(CR ) i=1 λ−1
i (CR ) dβ


– Full rank code, i.e., r Ẽ = nt
nY
t nr nY
t nr






nr
λi (CR ) = λi R Inr ⊗ Ẽ = det R Inr ⊗ Ẽ = det Ẽ det (R) .
i=1 i=1
No interactions between the
channel and the code at very high SNR!
– Non-full rank code, i.e., r Ẽ < nt
r(CR ) r (Ẽ)
Y
Y

λi (CR ) = det Q′ λn
i
r
Ẽ
i=1 i=1



where Q′ is a nr r Ẽ × nr r Ẽ principal submatrix of VC
H RV
C with
H H
Inr ⊗ (C − E) = UC ΛC VC .
186 / 494
Average PEP in Rayleigh Slow Fading Channels

• Rank deficient codes sensitive to spatial correlation!

Example
Question: In a 2 × 2 spatially correlated Rayleigh fading MIMO channel, derive
the Average PEP for Spatial Multiplexing with ML receiver and discuss the
effect of transmit and receive correlation on the performance.
Answer: For SM, C − E is a nt × 1 vector and

P (C → E)
Z
1 π/2    −1
= det Inr nt + η (Rr ⊗ Rt ) Inr ⊗ Ẽ dβ
π 0
Z
1 π/2   −1
= det Inr nt + η (Rr ⊗ Rt ) (Inr ⊗ (C − E)) (Inr ⊗ (C − E)H ) dβ
π 0
Z
1 π/2   −1
= det Inr + η(Inr ⊗ (C − E)H ) (Rr ⊗ Rt ) (Inr ⊗ (C − E)) dβ
π 0
Z π/2    −1
1
= det Inr + η Rr ⊗ (C − E)H Rt (C − E) dβ.
π 0

187 / 494
Average PEP in Rayleigh Slow Fading Channels

Example
For SM over a 2 × 2 MIMO channel,
     
c0 − e0 1 t∗ 1 r∗
C−E= , Rt = , Rr = ,
c1 − e1 t 1 r 1

a = (C − E)H Rt (C − E) = |c0 − e0 |2 +|c1 − e1 |2 +2ℜ {t (c0 − e0 ) (c1 − e1 )∗ } .

Hence
Z π/2
1
P (C → E) = (det (I2 + ηaRr ))−1 dβ.
π 0
Z π/2
1
= (1 + ηa(1 + |r|))−1 (1 + ηa(1 − |r|))−1 dβ.
π 0

At high SNR, assuming a > 0 and |r| < 1, we get

Z Z
1 π/2 1 π/2
−1 −2 −2
P (C → E) ≈ (det (Rr ))−1 η −2 a−2 dβ = 1 − |r|2 η a dβ.
π 0 π 0
As |r| increases, the PEP increases. For large value of |t|, a can be very small
for some error vectors leading to detrimental performance of SM. 188 / 494
Average PEP in Rayleigh Slow Fading Channels

Example
Observations:
– performance of SM in correlated channels depends on the projection of C − E
onto the space spanned by the eigenvectors of Rt :
– worse performance when C − E is parallel to the eigenvector of Rt
corresponding to the lowest eigenvalue.
Intuition: transmitting all the information contained in the unique non-zero
eigenvalue of the error matrix Ẽ in the direction of the space offering the lowest
scatterer density.
– receive correlation induces a coding gain loss independent of the error matrix.


189 / 494
Average PEP in Rayleigh Slow Fading Channels

• Analysis can be extended to finite SNR and space-time correlated Rician channels
(see more in Chapter 8 if interested).

• Main observations:
– Rank deficient codes are very sensitive to spatial correlation.

– Designing codes using the rank-determinant criterion is not sufficient to guarantee a

good performance in spatially correlated Rayleigh slow fading channels when the code
is rank-deficient.

– Designing codes based on the distance-product criterion is not sufficient to guarantee a

good performance in spatially correlated Rayleigh fast fading channels, irrespective of
the rank of the code.

– The maximization of the coding gain in i.i.d. Rayleigh channels is not a sufficient
condition to guarantee the good performance of a code in correlated channels at finite
SNR, even for full-rank codes.
– Robust code design exist (see Chapter 9 if interested)

190 / 494
Average PEP in Rayleigh Slow Fading Channels

• Performance of full-rank and rank-deficient STBCs in i.i.d. and spatially correlated

channels with nt = 2 and nr = 2.
0 0
10 10
LDC 1 − i.i.d.
LDC 2 − i.i.d.
LDC 1 − correlated
LDC 2 − correlated

−1 −1
10 10
BER

BER
−2 −2
10 10

−3 −3
10 10
SM and new SM − i.i.d.
Hassibi et al. − i.i.d.
SM − correlated
Hassibi et al. − correlated
−4 −4
new SM − correlated
10 10
4 6 8 10 12 14 16 18 20 22 4 6 8 10 12 14 16 18 20 22
SNR [dB] SNR [dB]

191 / 494
MIMO with Partial Channel State Information at
the Transmitter

192 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 10
Section: 10.1, 10.2.1, 10.5, 10.6.1,
10.6.2, 10.6.3, 10.6.4, 10.9

193 / 494
Introduction

• full CSIT
– array and diversity gain
– lower system complexity (parallel virtual transmissions)
– hardly achievable (especially when the channel varies rapidly), costly in terms of
feedback
• Exploiting Channel Statistics at the Transmitter
– low rate feedback link
– statistical properties of the channel (correlations, K-factor) vary at a much slower rate
than the fading channel itself
– The receiver estimates the channel stochastic properties and sends them back to the
transmitter “once in a while” (if channel reciprocity cannot be exploited)
– stationary channel: statistics do not change over time
• Exploiting a Limited Amount of Feedback at the Transmitter
– codebook of precoding matrices, i.e., a finite set of precoders, designed off-line and
known to both the transmitter and receiver.
– The receiver estimates the best precoder as a function of the current channel and feeds
back only the index of this best precoder in the codebook.

194 / 494
System Model

• MIMO system with nt transmit and nr receive antennas communicating through a

frequency flat slow fading channel.
• The encoder outputs a codeword C = [c0 . . . cT −1 ] of size ne × T contained in the
codebook C over T symbol durations.
• Precoder P [nt × ne ] processes the codeword C and the codeword C′ = PC
[nt × T ] is transmitted over nt antennas.
P

1/2
C S C’’ W C’
constellation shaper beamformer

• Linear precoder P = WS1/2

– multi-mode beamformer W whose columns have a unit-norm
– constellation shaper S1/2 (if S real-valued and diagonal, it can be thought of as the
power allocation scheme across the modes)
    
• nomalization: E Tr C′ C′H = T , E Tr CCH = T and Tr PPH = ne ,

E kHk2F = nt nr .
195 / 494
Channel Statistics based Precoding

• Information Theory motivated strategy.

Proposition

In Kronecker correlated Rayleigh fast fading channels, the optimal input

covariance matrix can again be expressed as

Q = U R t ΛQ U H
Rt ,

where URt is a unitary matrix formed by the eigenvectors of Rt (arranged in

such order that they correspond to decreasing eigenvalues of Rt ), and ΛQ is a
diagonal matrix whose elements are also arranged in decreasing order.

Transmit a single stream along the dominant eigenvector of Rt if very large transmit
correlation. Transmit multiple streams with uniform power allocation if very low
transmit correltion.
• Error Probability motivated strategy
P⋆ = arg min max P (C → E)
P Ẽ6=0

– challenging problem for arbitrary codes

– focus on O-STBC
196 / 494
Channel Statistics based Precoding

• O-STBCs in Kronecker Rayleigh fading channels

  
P⋆ = arg max max det Inr nt + ζR Inr ⊗ PPH
P Ẽ6=0
  
= arg max max det Inr nt + ζ Rr ⊗ Rt PPH
P Ẽ6=0

2
where ζ = ηT δ /(Qne ) and δ = dmin .

Proposition
In Kronecker Rayleigh fading channels, the optimal precoder minimizing the
average PEP/SER is given by P = WS1/2 where
– W = U′Rt with U′Rt the nt × ne submatrix of URt containing the ne
dominant eigenvector of Rt , i.e., Rt = URt ΛRt UH
Rt ,

– S1/2 = D, D being a real-valued diagonal matrix accounting for the power

allocation.

• Power allocation strategy follows the water-filling solution.

197 / 494
Channel Statistics based Precoding

Example
Let us consider the Alamouti O-STBC with two transmit antennas
(ne = nt = 2). Denoting S = diag {s1 , s2 }, the transmitted codewords are
proportional, at the first time instant, to
 
1 c1 1 √ 1 √
√ URt S1/2 = √ URt (:, 1) s1 c1 + √ URt (:, 2) s2 c2
2 c2 2 2
and, at the second time instant, to
 
1 −c∗2 1 √ 1 √
√ URt S1/2 = − √ URt (:, 1) s1 c∗2 + √ URt (:, 2) s2 c∗1 .
2 c∗1 2 2
Extreme cases:
• s1 = s2 = 1: Alamouti scheme
• s1 = 2, s2 = 0: beamforming in the dominant eigenbeam
The precoder allocates more power to angular directions corresponding to the
peaks of the transmit direction power spectrum.

198 / 494
Channel Statistics based Precoding

• Performance of a transmit correlation based precoded Alamouti scheme in 2 × 2

transmit correlated (t = 0.7 and t = 0.95) Rayleigh channels
−1 −1
10 10
t = 0 .7, Alamouti t = 0 .95, Al amouti
t = 0 .7, Alamouti with precoder t = 0 .95, Alamouti with precoder
t = 0 .7, single mode eigenbeamforming t = 0 .95, dominant eigenmode transmission
t = 0 .7, dominant eigenmode transmission i.i.d., Al amouti
i.i.d., Al amouti
BER

BER
−2 −2
10 10

−3 −3
10 10
0 2 4 6 8 10 12 0 2 4 6 8 10 12 14
SNR [dB] SNR [dB]

199 / 494
Quantized Precoding: dominant eigenmode transmission

• Assume dominant eigenmode transmission (i.e. beamforming)

√
y = Es Hwc + n,
z = gH y,
√
= Es gH Hwc + gH n

where g and w are respectively nr × 1 and nt × 1 vectors.

• Assuming MRC, the optimal beamforming vector w that maximizes the SNR is given
by
w⋆ = arg max kHwk2
w∈Cw

with Cw set of unit-norm vectors. The best precoder is the dominant right singular
vector of H.
• Reduce the number of feedback bits: limit the space Cw over which w can be chosen
to a codebook called W. The receiver evaluates the best precoder w⋆ among all
unit-norm precoders wi ∈ W (with i = 1, . . . , np ) such that

w⋆ = arg max kHwi k2 .

1≤i≤np
wi ∈W

200 / 494
Quantized Precoding: distortion function

• How to design the codebook? Need for a distortion function, i.e. measure of the
average (over all channel realizations) array gain loss induced by the quantization
process n o
2
df = EH λmax − kHw⋆ k

• Upper-bound
  2 
 H
w⋆  ,

df ≤ EH λmax − λmax vmax
  2 
(a)  H
w⋆ 

= EH {λmax } EH 1 − vmax
| {z }
quality of the channel | {z }
quality of the codebook

where vmax is the dominant right singular vector of H. Equality (a) is only valid for
i.i.d. Rayleigh fading channels.

201 / 494
Quantized Precoding: Lloyd

• Generalized Lloyd Algorithm:

Algorithm
For the given codebook, find the optimal quantization cells using the nearest
neighbor rule. For the so-obtained quantization cells, determine that optimal
quantized precoders using the centroid condition. Iterate till convergence.

• Essential conditions:
– Assume MISO channel.
– centroid condition: the optimal quantized precoder w k of any quantization
cell Rk is
to be chosen as the dominant eigenvector of Rk = E hH h | h ∈ Rk .
– nearest neighbor rule: all channel vectors that are closer to the quantized precoder wk
are assigned to quantization cell Rk , i.e. h ∈ Rk if khk2 − |hwk |2 ≤ khk2 − |hwj |2 ,
• Optimal codebook design for arbitrary fading channels

202 / 494
Quantized Precoding: Grassmannian

• Grassmannian Line Packing

Design Criterion

Choose a codebook W made of np unit-norm vectors wi (i = 1, . . . , np ) such

that the minimum distance
q
2
δline (W) = min 1 − |wkH wl | ,
1≤k<l≤np

is maximized.

– Problem of packing np lines in nt in

such a way that the minimum distance
between any pair of lines is maximized.
– Close to optimal only for i.i.d. Rayleigh
Fading Channels.

203 / 494
Quantized Precoding: How many bits?

• How many feedback bits B = log2 (np ) are required? In i.i.d. channels
n   o
− B
C̄quant ≈ Eh log2 1 + ρ khk2 1 − 2 nt −1 ,
 
− B
leading to an SNR degradation of 10 log10 1 − 2 nt −1 dB relative to perfect
CSIT.
Proposition

In order to maintain a constant SNR or capacity gap between perfect CSIT and
quantized feedback, it is not necessary to scale the number of feedback bits as
a function of the SNR. The multiplexing gain gs is not affected by the quality
of CSIT.
• Achievable diversity gain?
– Antenna selection (AS) is a particular case of a quantized precoding whose codebook is
chosen as the columns of the identity matrix Int .
– AS achieves a diversity gain of nt .
– Sufficient to take a full rank codebook matrix with np ≥ nt to extract the full diversity

204 / 494
Quantized Precoding: Evaluations

• SER of a 3 × 3 MIMO system using 2-bit and 6-bit quantized BPSK-based DET.
−2
10
2-bit D-DET
6-bit Q-DET
ant. selection Example
P-DET

Codebook for quantized DET for

−3
nt = 3 and np = 4.
10
 1   j 
 √3
 √
  −13 
SER

W =  √13  ,  √

3 
,

 √1 −j
−4
10
√
3 3
 −1   −j 
√ √ 
 √13   √ 3 
−1 
 3 , 3 
−1 
√ √j 
0 1 2 3 4 5 6 3 3
SNR [dB]

205 / 494
Quantized Precoding: spatially correlated channels

• Grassmiannian only appropriate for i.i.d. Rayleigh fading.

• Spatial correlation decreases the quantization space compared to i.i.d. channels.
– e.g. Lloyd, adaptive/CDIT-based codebook, DFT (for uniform linear arrays)
• Normalized average distortion (SNR loss) df,n = df /EH {λmax } as a function of the
codebook size np = 2B and the transmit correlation coefficient t with nt = 4.

0.4

0.35

0.3
f,n
Average distortion d

0.25

0.2

0.15

CDIT−based CB, B=3

0.1 CDIT−based CB, B=4 ( 1/2
)
CDIT−based CB, B=5 1/2
|t|−based CB, B=4 R t w1 R t wn p
0.05
DFT CB, B=4 Wc = 1/2 , . . . , 1/2
GLP CB, B=4 R t w1 R t wn p
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Magnitude of the transmit correlation coefficient t

206 / 494
Quantized Precoding: some extensions

• Extension to other kinds of channel models (e.g. spatial/time correlation,

polarization), transmission strategies (e.g. O-STBCs, SM), reception strategies (e.g.
MRC, ZF, MMSE, ML), criteria (e.g. error rate or transmission rate)
• Quantized precoding for SM with rank adaptation and rate maximization

W⋆ = arg max max R.

ne (n )
Wi e ∈Wne

The codebooks Wne are defined for ranks ne = 1, . . . , min {nt , nr }. Rate is
(n )
computed on the equivalent precoded channel HWi e .
– Uniform power allocation and joint ML decoding
 
ρ  (ne ) H H (n )
R = log2 det Ine + Wi H HWi e .
ne
– With other types of receivers/combiner
ne
X   
(n )
R= log2 1 + ρq HWi e .
q=1

where ρq is the SINR of stream q on at the output of the combiner for the equivalent
(n )
channel HWi e .

207 / 494
Frequency-Selective MIMO Channels -
MIMO-OFDM

208 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 11
Section: 11.1, 11.4.1, 11.5.2

209 / 494
Introduction

• Two major different approaches to transmit information over frequency selective

MIMO channels:
– modulating a single carrier over the full bandwidth B.
– converts the frequency selective channel into a set of multiple parallel flat fading
channels in the frequency domain by means of Orthogonal Frequency Division
Multiplexing (OFDM) modulation.

210 / 494
Single-Carrier Transmissions

• Transmission of a codeword C = [c0 . . . cT −1 ] (of size nt × T )

• Presence of L resolvable taps, which are responsible for inter-symbol interference
(ISI)
L−1
X
H[τ ] = H[l]δ (τ − τl )
l=0

– L replicas of the same codeword → Lth order diversity!

• How to design codewords such that a ML decoder is able to efficiently exploit the
frequency and spatial diversity without suffering from the ISI?
• System model
√ L−1X
yk = Es H [l] ck−l + nk
l=0

where yk , nk , Es are defined analogous to frequency flat fading channels.

211 / 494
Virtual Transmit Antenna Array

• Equivalent system model

√  T
yk = Es H cTk ... cTk−L+1 + nk
 
H, H [0] ... H [L − 1]

– L taps may be thought of as virtual transmit antennas

– Virtual transmit array of nt L antennas where the nr × nt L virtual channel matrix is H
and equivalent codeword
 
c0 c1 . . . cT −1
 c−1 c0 . . . cT −2 
 
C= . . .. . 
. . . .
 . . . 
c1−L c2−L . . . cT −L
– Maximum diversity gain of nr nt L

212 / 494
Multi-Carrier Transmissions: MIMO-OFDM

• Frequency domain approach to exploit the frequency diversity

• Basic idea of OFDM: Turn the channel matrix into a circulant matrix via the
addition of a cyclic prefix to the transmitted sequence
– A circulant matrix has the property that its left and right singular vector matrices are
respectively DFT and IDFT matrices.
– The multiplication by an IDFT matrix at the transmitter and by a DFT matrix at the
receiver transforms the frequency selective channel into a diagonal matrix, whose
elements are the singular values of the circulant matrix.
– The original frequency-selective channel in the time domain becomes a set of parallel
flat fading channels in the frequency domain

c0 y0
c1 y1

• •
IDFT CP CP DFT
• •
• •
cT-1 yT-1

• This construction allows for a considerable reduction of complexity in terms of

equalization and demodulation.
213 / 494
Orthogonal Frequency Division Multiplexing (OFDM)

• Fundamental Steps 1 to 6
th
1 Apply an IDFT to the codeword C. We obtain as output at the n time interval
(n = 0, . . . , T − 1),
T −1
1 X 2π
xn = √ ck ej T kn ,
T k=0
or equivalently in a matrix form,
 T  T
x0 . . . xT −1 = DH c0 . . . cT −1 ,
 T T   T
T H
x0 . . . xT −1 = D ⊗ I nt c0 . . . cTT −1
T
.

The T × T matrix DH realizes the IDFT operation. Hence, D is a DFT matrix

reading as
 
1 1 1 ... 1
2π 2π 2π
 1
 e−j T e−j T 2 ... e−j T (T −1) 

1  .
.. .. .
.. .
.. .
..

D= √   . .

T  2π 2π 2π 
 1 e−j T (T −2) e−j T (T −2)2 . . . e−j T (T −2)(T −1) 
2π 2π 2π
1 e−j T (T −1) e−j T (T −1)2 . . . e−j T (T −1)(T −1)
214 / 494
Orthogonal Frequency Division Multiplexing (OFDM)
 
2 Add the guard interval vector Xg = x−(L−1) . . . x−1 of length L − 1 in front of
the codeword X = [ x0 . . . xT −1 ] to avoid inter-symbol interference.
Choose the guard interval vector Xg in such way that x−n = xT −n , for
n = 1,. . . , L − 1. Hence, the guard interval vector becomes
Xg = xT −(L−1) . . . xT −1 and is commonly known as the cyclic prefix.
3 Transmit the OFDM symbol X′ = Xg X of size nt × (T + L − 1).
4 Remove the guard interval (CP) at the receiver and gather T output samples as
 T  T  T
rT0 ... rTT −1 = Hg xT−(L−1) ... xTT −1 + nT0 ... nTT −1

where
 
H [L − 1] ... H [1] H [0] 0nr ×nt ... 0nr ×nt
 .. 
 0nr ×nt H [L − 1] . H [1] H [0] ... 0nr ×nt 
Hg =  
 .. .. .. .. .. .. .. 
 . . . . . . . 
0nr ×nt ... 0nr ×nt H [L − 1] H [L − 2] ... H [0]

is a nr T × nt (T + L − 1) matrix representing the channel seen by the OFDM symbol.

215 / 494
Orthogonal Frequency Division Multiplexing (OFDM)

5 Observe that the choice of the CP creates a blockwise circulant matrix Hcp of size
nr T × nt T
 T T  T  T T
r0 . . . rTT −1 = Hcp xT0 . . . xTT −1 + n0 . . . nTT −1
with
 
H [0] 0nr ×nt ... 0nr ×nt H [L − 1] ... H [1]
 H [1] H [0] 0nr ×nt ... 0nr ×nt ... H [2] 
 
 . .. .. .. .. .. . 
. . . . . . .

 . . 

 H [L − 2] ... H [0] 0nr ×nt ... 0nr ×nt H [L − 1] 
Hcp = .

 H [L − 1] ... H [1] H [0] 0nr ×nt ... 0nr ×nt 


 .. .. .. .. .. .. .. 

 . . . . . . . 
 0nr ×nt ... H [L − 1] H [L − 2] ... H [0] 0nr ×nt 
0nr ×nt ... 0nr ×nt H [L − 1] ... H [1] H [0]


SVD decomposition Hcp = DH ⊗ Inr Λcp (D ⊗ Int ) is such that Λcp is a block
diagonal
 matrix whose blocks are obtained by a blockwise DFT of
H [0] H [1] . . . H [L − 1] , i.e., for the (k, k)th block
L−1
X
(kk) 2π kl
Λcp = H [l] e−j T , k = 0, . . . , T − 1,
l=0

irrespective of the channel matrix. 216 / 494

Orthogonal Frequency Division Multiplexing (OFDM)

6 The use of IDFT matrices at the transmitter is now clear as

 T T    T  T T
r0 . . . rTT −1 = DH ⊗ Inr Λcp cT0 . . . cTT −1 + n0 ... nTT −1

Applying a DFT operation to the received vector, we finally obtain

 T T  T
y0 . . . yTT −1 = (D ⊗ Inr ) rT0 . . . rTT −1
 T  T
= Λcp cT0 . . . cTT −1 + (D ⊗ Inr ) nT0 ... nTT −1 .

The original frequency selective channel has been converted into a set of T parallel
flat fading channels in the frequency domain, the channel gains being given by the
diagonal blocks of Λcp .

217 / 494
Orthogonal Frequency Division Multiplexing (OFDM)

• MIMO-OFDM System Model: The input-output relationship on each parallel

channel k = 0, . . . , T − 1 may be expressed without loss of generality as
√
yk = Es H(k) ck + nk

with
L−1
X
(kk) 2π kl
H(k) = Λcp = H [l] e−j T .
l=0

The nr × 1 vector yk is the received signal to be decoded, and nk is a nr × 1 zero

mean complex AWGN vector with E{nk nH 2 ′
k′ } = σn Inr δ [k − k ].
• If ML decoding is applied, the decoder computes an estimate of the transmitted
codeword according to
T −1
X √ 2

Ĉ = arg min yk − Es H(k) ck .
C
k=0

218 / 494
Orthogonal Frequency Division Multiplexing (OFDM)

• Block diagram of a MIMO-OFDM system

OFDM
interleaver S/P modulator

encoder

interleaver S/P OFDM

modulator

de- OFDM
decoder interleaver demodulator

• Strong analogy with the input-output relationship over a flat fading MIMO channel:
– Temporal dimension replaced by frequency dimension
– Hk ↔ H(k)
– Commonly known as Space-Frequency Coded MIMO-OFDM.
– If the coherence bandwidth of the channel is small, the channel gains H(k) vary
significantly from tone to tone. The channel in the frequency domain can then be
considered as a fast fading channel in the frequency domain.
219 / 494
Virtual Transmit Antenna Array

• Virtual transmit array

√ h
2π kl 2π k(L−1)
iT
yk = Es H cTk ... e−j T cTk ... e−j T cTk + nk

where H is the nr × nt L virtual channel matrix and the equivalent codeword is

 
c0 . . . ck ... cT −1
 .. .. .. .. .. 
 .
 . . . . 

 2π
−j T kl 
C= e ck 
 . .. .. .. 
 . .. 
 . . . . . 
2π
−j T k(L−1) 2π
−j T (T −1)(L−1)
c0 . . . e ck . . . e cT −1

• Equivalent codewords differ from the single-carrier case.

• Maximum diversity gain of nr nt L.

220 / 494
Unified Representation for Single and Multi-Carrier
Transmissions
• Unique virtual nr × Lnt MIMO channel
  √  
Y = y0 · · · yT −1 = Es H C + n0 ··· nT −1
• Equivalent transmitted codewords in the virtual nr × Lnt MIMO representation
 
C(0)
C=
 .. 
. 
C(L−1)

– Space-frequency coded MIMO-OFDM

     
C(0) CD(0) D(0)
 .  
= . 
= [I ⊗
 . 
 .
.   .
.  L C]  .. 
C(L−1) CD(L−1) D(L−1)
| {z }
D
n 2π 2π
o
with D(l) = diag 1, . . . , e−j T kl , . . . , e−j T (T −1)l .
– Single carrier transmissions,
C(l) (m, k) = ck−l (m, 1) , k = 0, . . . , T − 1, m = 1, . . . , nt .
√ 2
• Maximum-likelihood (ML) decoding: Ĉ = arg minC Y − Es H C
221 / 494
Capacity of Frequency Selective MIMO Channels

• MIMO frequency selective channels

– Mutual information obtained by an integration over the frequency band of interest B
  Z h i
1
IF S {H(f )}f , {Q(f )}f = log2 det Inr + ρ(f )H(f )Q(f )H(f )H df
B B
R
subject to B Tr {Q(f )} = PB .
– Capacity  
CCSIT,F S = R max IF S {H(f )}f , {Q(f )}f .
B Tr{Q(f )}=PB
• MIMO-OFDM (neglecting the loss in spectral efficiency due to the cyclic prefix):
– Mutual information
 T −1 T −1
  1 X 1 X h i
IF S H(k) k , Q(k) k = I(k) = log2 det Inr + ρH(k) Q(k) HH
(k)
T k=0 T k=0
PT −1 
subject to k=0 Tr Q(k) = T .
– Capacity
T −1 h i
1 X
CCSIT,F S = max log2 det Inr + ρH(k) Q(k) HH(k)
T T −1 Tr{Q(k) }=T k=0
P
k=0

T −1 X
n
1 X  
= max log2 1 + ρs(k),l λ(k),l .
T PT −1 Pn
s =T k=0 l=1
k=0 l=1 (k),l

Solved using a space-frequency water-filling.

222 / 494
Average Pairwise Error Probability

• Conditional pairwise error probability (PEP)

r 
Es 2
P (C → E|H) = Q kH (C − E)k F
2σn2
• Average PEP in Rayleigh slow fading channels
Z
1 π/2   −1
P (C → E) = det IT nr + ηCR dβ
π 0
where


CR = Inr ⊗ (C − E)H R (Inr ⊗ (C − E))
n

H o
R = E vec HH vec HH


Assuming full rank R, full diversity at high SNR if r CR = nr nt L.
• If each tap is spatially i.i.d. Rayleigh distributed with an average power βl and if
there is no correlation between taps R = Inr ⊗ diag {β0 , . . . , βL−1 } ⊗ Int ,
Z
1 π/2 h
i−nr
P (C → E) = det ILnt + η [diag {β0 , . . . , βL−1 } ⊗ Int ] Ẽ dβ
π 0


where Ẽ , (C − E) (C − E)H . Full diversity at high SNR if r Ẽ = nt L.
223 / 494
Code Design for Single-Carrier Transmissions

Example
Send c0 , . . . , cT −1 over two antennas in a 2-tap Rayleigh fading channel.
• classical delay-diversity scheme
 
1 c0 c1 ... cT −1 0 0
C= √ ,
2 0 c0 c1 ... cT −1 0
  Diversity of
c0 c1 ... cT −1 0 0
3nr only!
1  0 c0 c1 ... cT −1 0 
C= √   .
2 0 c0 c1 ... cT −1 0 
0 0 c0 c1 ... cT −1
• Generalized delay-diversity scheme
 
1 c0 c1 ... cT −1 0 0 0
C= √ ,
2 0 0 c0 c1 ... cT −1 0
  Diversity of
c0 c1 ... cT −1 0 0 0
4nr !
1  0 0 c0 c1 ... cT −1 0 
C= √  .
2 0 c0 c1 ... cT −1 0 0 
0 0 0 c0 c1 ... cT −1
224 / 494
Code Design for Space-Frequency Coded MIMO-OFDM

• Diversity gain
– Define lC,E (l) = ♯τC,E (l) (l = 1, . . . , nt ) with τC,E (l) = {k | ck (l) − ek (l) 6= 0} .
– Define lC,E = ♯τC,E with τC,E = {k | ck − ek 6= 0}.

Proposition
For full rank space-tap correlation matrix R, a pair of space-frequency
codewords {C, E} with an effective length lC,E , effective lengths {lC,E (l)}n t
l=1
and a rank r(Ẽ) achieves the full diversity nt nr L if

r(Ẽ) = nt ,
lC,E (l) ≥ L, ∀l = 1, . . . , nt ,
lC,E ≥ nt L.

225 / 494
Code Design for Space-Frequency Coded MIMO-OFDM

• Coding gain
– Assume each tap l is i.i.d. Rayleigh distributed with an average power βl and no
correlation between taps
h  i
CR = Rf ⊙ 1nr ×nr ⊗ (C − E)H (C − E) ,
R f = I nr ⊗ R F , (space-frequency correlation matrix)
L−1
X  
RF = βl d H
(l) d(l) with d(l) = 1 ... e−j2πkl/T ... e−j2π(T −1)l/T
l=0

Example
PL−1 2π
SISO channel h(k) = l=0 h [l] e−j T kl . Frequency correlation between channel on
subcarrier k and k + K
n o L−1
X n o 2π
E h(k) h∗(k+K) = E |h[l]|2 ej T lK
l=0 | {z }
βl

226 / 494
Code Design for Space-Frequency Coded MIMO-OFDM

– Impact of Frequency Correlation on the Coding Gain

   h  i
det IT nr + ηCR = det IT nr + ηRf ⊙ 1nr ×nr ⊗ (C − E)H (C − E)
  h  inr
= det IT + η RF ⊙ (C − E)H (C − E)
TY−1
"L−1 # !nr
X
≤ 1+η βl kck − ek k2
k=0 l=0

– MIMO-OFDM ≈ narrowband MIMO transmissions over i.i.d. fast fading Rayleigh

channels if Rf is diagonal (L >> 0).
– Frequency correlation reduces the achievable coding gain. Reduce the frequency
correlation between adjacent tones by means of an interleaver.

227 / 494
Code Design for Space-Frequency Coded MIMO-OFDM

• With interleaver, use codes with large effective length and product distance for
OFDM transmissions as in fast fading channels.
• FER of the 16-state space-time trellis code for L = 2, 3 and 4 in uniformly
distributed i.i.d. Rayleigh channels with and without interleaver.
0
10

−1
10
FER

−2
10 L=2, with interleaver
L=3, with interleaver
L=4, with interleaver
fast fading
L=2, no interleaver
L=3, no interleaver
L=4, no interleaver
−3
10
7.5 10 12.5 15 17.5
SNR [dB]

228 / 494
Code Design for Space-Frequency Coded MIMO-OFDM

• Space-Frequency Linear Block Coding

– Orthogonal codes: O-STBC → O-SFBC (time replaced by frequency)
– Make sure that O-SFBC is operated on adjacent subcarriers. Recall that the channel
has to be constant within a O-STBC/O-SFBC block!
– In practice, O-SFBC often preferred over O-STBC.
• Cyclic Delay Diversity
– Adaptation of the generalized delay-diversity (GDD) scheme to OFDM systems.
– Send on each antenna a circularly shifted version of the same OFDM symbol in the
time domain. Hence, the temporal delay introduced on each antenna in the GDD
scheme is transformed into a cyclic delay in the CDD scheme.

∆1 CP

c OFDM x
∆2 CP
modulator

∆ nt CP

1 A sequence c of symbols ck with k = 0, . . . , T − 1 is OFDM modulated.

2 The output sequence x is transmitted on each antenna with a cyclic delay ∆m , m = 1, . . . , nt
so that the output symbol on antenna m (m = 1, . . . , nt ) at time n (n = 0, . . . , T − 1) is given
by x(n−∆m )mod T .
3 Finally, CP is added on each antenna, analogous to conventional OFDM transmissions.
4 At the receiver, the cyclic prefix is removed, and OFDM demodulation and decoding are
performed.
229 / 494
Code Design for Space-Frequency Coded MIMO-OFDM

– Analogous to GDD, CC converts a MIMO channel into a SIMO channel with enhanced
frequency selectivity. The subsequent frequency diversity is extracted by appropriate
outer codes.
– A cyclic shift in the time domain corresponds to the multiplication by a phase shift in
the frequency domain. Therefore, the received signal in the frequency domain reads as
s
Es
yk = heq,(k) ck + nk
nt

where the equivalent SIMO channel matrix on the kth tone, denoted as heq,(k) , is
given by
nt
X 2π
heq,(k) = H(k) (:, m) e−j T k∆m
m=1
and H(k) is the DFT of the impulse response evaluated on the kth subcarrier
(kk) PL−1 2π
H(k) = Λcp = l=0 H [l] e−j T kl .

230 / 494
Code Design for Space-Frequency Coded MIMO-OFDM

• CDD vs. GDD vs. O-SFBC

– over GDD: reduced guard interval.
– over O-SFBC/O-STBC: increased flexibility and scalability to any nt , no rate loss if
nt > 2, no requirement on constant channel over several tones, unlike in O-SFBC and
O-STBC.
– CDD receiver is essentially the same as a classical SIMO receiver.
– the number of states of the outer code necessary to exploit the full diversity is much
larger with CDD than with O-SFBC.
• Precoder cycling
– Codevector on subcarrier k writes as ck = wk ck where wk is the precoding vector and
ck is a complex symbol.
– The precoder changes every M contiguous physical subcarriers.
– Appropriate design (using Grassmanian Line Packing) of the precoders converts the
MIMO channel into a frequency selective SIMO channel.

231 / 494
Multi-User MIMO - Multiple Access Channels
(Uplink) & Broadcast Channels (Downlink)

232 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 12
Section: 12.1, 12.2, 12.3, 12.4

233 / 494
Introduction

• So far, we looked at a single link/user. Most systems are multi-user!

• How to deal with multiple users? What is the benefit of MIMO in a multi-user
setting?
• MIMO Broadcast Channel (BC) and Multiple Access Channel(MAC)

Differences between BC and MAC:

– there are multiple independent receivers (and therefore multiple independent additive
noises) in BC while there is a single receiver (and therefore a single noise term) in MAC.
– there is a single transmitter (and therefore a single transmit power constraint) in BC
while there are multiple transmitters (and therefore multiple transmit power
constraints) in MAC.
– the desired signal and the interference (originating from the co-scheduled signals)
propagate through the same channel in the BC while they propagate through different
channels in the MAC.
234 / 494
MIMO MAC System Model

• Uplink multi-user MIMO (MU-MIMO) transmission

– total number of K users (K = {1, . . . , K}) distributed in a cell,
– nt,q transmit antennas at mobile terminal q (we simply drop the index q and write nt if
nt,q = nt ∀q)
– nr receive antenna at the base station
• Received signal (we drop the time dimension)
K
X
yul = Λq−1/2 Hul,q c′ul,q + nul
q=1

where
– yul ∈ nr
– Hul,q ∈ nr ×nt,q models the small scale time-varying fading process and Λ−1
q refers
to the large-scale fading accounting for path loss

and shadowing
– nul is a complex Gaussian noise CN 0, σn 2I
nr .
• User q’s input covariance matrix
 is defined as the covariance matrix of the transmit
H
signal of user q as Qul,q = E c′ul,q c′ul,q .
• Power constraint: Tr{Qul,q } ≤ Es,q .

235 / 494
MIMO MAC System Model

• By stacking up the transmit signal vectors and the channel matrices of all K users,
h ′ ′
iT
c′ul = cul,1
T T
, . . . , cul,K ,
h i
−1/2 −1/2
Hul = Λ1 Hul,1 , . . . , ΛK Hul,K ,

the system model also writes as

yul = Hul c′ul + nul .

Hul is assumed to be full-rank as it would be the case in a typical user deployment.

• Long term SNR of user q defined as ηq = Es,q Λ−1 2
q /σn .

• Note on the notations: the dependence on the path loss and shadowing is made
explicit in order to stress that the co-scheduled users experience different path losses
and shadowings and therefore receive power.
• We assume that the receiver (i.e. the BS in a UL scenario) has always perfect
knowledge of the CSI, but we will consider strategies where the transmitters have
perfect or partial knowledge of the CSI.

236 / 494
Capacity Region of Deterministic Channels

• In a multi-user setup, given that all users share the same spectrum, the rate
achievable by a given user q, denoted as Rq , will depend on the rate of the other
users Rp , p 6= q → Trade-off between rates achievable by different users!
• The capacity region C formulates this trade-off by expressing the set of all user rates
(R1 , . . . , RK ) that are simultaneously achievable.
Definition

The capacity region C of a channel Hul is the set of all rate vectors
(R1 , . . . , RK ) such that simultaneously user 1 to user K can reliably
communicate at rate R1 to rate RK , respectively.

Any rate vector not in the capacity region is not achievable (i.e. transmission at
those rates will lead to errors).
Definition

The sum-rate capacity C of a capacity region C is the maximum achievable

sum of rates
K
X
C= max Rq .
(R1 ,...,RK )∈C
q=1
237 / 494
Rate Region of MIMO MAC

• For given input covariance matrices Qul,1 , . . . , Qul,K , the achievable rate region is
defined by
1 The rate achievable by a given user q with a given transmit strategy Qul,q cannot be
larger than its achievable rate in a single-user setup, i.e.
 
Λ−1
q
Rq ≤ log2 det Inr + 2 Hul,q Qul,q HH ul,q , q = 1, . . . , K
σn

where Qul,q = E c′q c′q H is subject to the power constraint Tr{Qul,q } ≤ Es,q .
2 The sum of the rates achievable by a subset S of the users should be smaller than the
total rate achievable
P when those users “cooperate” with each other to form a giant
array with nt,S = q∈S nt,q transmit antennas subject to their respective power
constraints, i.e.
 
X 1
Rq ≤ log2 det Inr + 2 Hul,S Qul,S HH ul,S
q∈S
σn
 
1 X −1
= log2 det Inr + 2 Λq Hul,q Qul,q HH ul,q ,
σn q∈S
h i
−1/2 −1/2
with Hul,S = Λi Hul,i , . . . , Λj Hul,j ,
 i,j∈S
Qul,S = diag Qul,i , . . . , Qul,j i,j∈S , subject to the constraints Tr{Qul,q } ≤ Es,q .
• The rate region looks like a K-dimensional polyhedron with K! corner points on the
boundary.
238 / 494
Rate Region of a Two-User MIMO MAC

• This rate region is a pentagon with two corner points A and B.

• Remarkably, at point A, user 1 can transmit at a rate equal to its single-link MIMO
rate and user 2 can simultaneously transmit at a rate R2′ > 0 equal to
 
Λ−1 Λ−1
R2′ = log2 det Inr + 12 Hul,1 Qul,1 HHul,1 +
2
2
Hul,2 Qul,2 HH
ul,2
σn σn
 
Λ−1
− log2 det Inr + 12 Hul,1 Qul,1 HH ul,1
σn
 !−1 
−1
Λ Λ−1
= log2 det Inr + 22 Hul,2 Qul,2 HHul,2 I nr + 1
2
H Q H H
ul,1 ul,1 ul,1 .
σn σn
239 / 494
Capacity Region of MIMO MAC

• We have assumed so far specific input covariance matrices.

– A different choice of the beamforming matrix and the power allocation leads to a
different transmit strategy Qul,q and generally a different shape of the pentagon (or
more generally the K-dimensional polyhedron).
– The trade-off between user rates is therefore affected by the choice of the input
covariance matrices.
– The optimal set of input covariance matrices that maximizes the sum-rate can be found
using a generalization of the single-link water-filling solution (Detail in the book).
• The capacity region is equal to the union (over all transmit strategies satisfying the
power constraints) of all the K-dimensional polyhedrons.

Proposition

The capacity region CM AC of the Gaussian MIMO MAC for a determinsitic

channel Hul is the union of all achievable rate vectors (R1 , . . . , RK ) given by
 P 
[  (R1 , . . . ,RK ) : q∈S Rq ≤  
P Λ−1
q H .
 log2 det Inr + q∈S σ2 Hul,q Qul,q Hul,q , ∀S ⊆ K 
Tr{Qul,q }≤Es,q n
Qul,q ≥0,∀q

240 / 494
Capacity Region of a Two-User MIMO MAC

• Due to the union of pentagons, the capacity region of the two-user MIMO MAC
does not look like a pentagon in general.

• However, with a single antenna (nt,q = 1), the capacity region remains a pentagon
because a single data is transmitted per user at the full power, i.e. Es,q .

241 / 494
Capacity Region of SISO MAC

Corollary

 X X

CM AC = (R1 , . . . , RK ) : Rq ≤ log2 1 + ηq |hul,q |2 , ∀S ⊆ K
q∈S q∈S

where ηq = Λ−1 2
q Es,q /σn .

Example
Two-user SISO: CM AC is the set of all rates pair (R1 , R2 ) satisfying to


Rq ≤ log2 1 + ηq |hul,q |2 , q = 1, 2


R1 + R2 ≤ log2 1 + η1 |hul,1 |2 + η2 |hul,2 |2 .

R2′ = log2 1 + η1 |hul,1 |2 + η2 |hul,2 |2 − log2 1 + η1 |hul,1 |2
   2 
η2 |hul,2 |2 Λ−1
2 |hul,2 | Es,2
= log2 1 + = log2 1 + .
1 + η1 |hul,1 |2 σn2 + Λ−1 2
1 |hul,1 | Es,1

242 / 494
Capacity Region of SIMO MAC

Corollary
( P )
(R1 , . . . ,hRK ) : q∈S Rq ≤ i
CM AC = P
log2 det Inr + q∈S ηq hul,q hH
ul,q , ∀S ⊆ K

where ηq = Λ−1 2
q Es,q /σn .

Example
Two-user SIMO: CM AC is the set of all rates pair (R1 , R2 ) satisfying to

 
Rq ≤ log2 1 + ηq khul,q k2 = log2 det Inr + ηq hul,q hH ul,q , q = 1, 2
 
R1 + R2 ≤ log2 det Inr + η1 hul,1 hH H
ul,1 + η2 hul,2 hul,2 .

   
R2′ = log2 det Inr + η1 hul,1 hH H H
ul,1 + η2 hul,2 hul,2 − log2 det Inr + η1 hul,1 hul,1
  −1 
= log2 det Inr + η2 hul,2 hH H
ul,2 Inr + η1 hul,1 hul,1
  −1 
= log2 1 + η2 hH H
ul,2 Inr + η1 hul,1 hul,1 hul,2
243 / 494
Achievability of the Capacity Region

• For nt = 1, the SIMO MAC architecture is reminiscent of the Spatial Multiplexing

architecture discussed for a single-link MIMO channel.
• We can therefore fully reuse the various receiver architectures derived for single-link
MIMO.
• Recall the optimality of the MMSE V-BLAST (also called Spatial Multiplexing with
MMSE-SIC receiver)

Proposition
MMSE-SIC is optimal for achieving the corner points of the MIMO MAC rate
region.

• The exact corner point that is achieved on the rate region depends on the stream
cancellation ordering:
– Point A, user 2 is canceled first (i.e. all streams from user 2) such that user 1 is left
with the Gaussian noise and can achieve a rate equal to the single-link bound.
– Assuming nt = 1, R2′ = log2 (1 + ρq ) where ρq is the SINR of the MMSE receiver for
user 2’s stream treating user 1’s stream as colored Gaussian interference.

244 / 494
Comparisons with TDMA

• TDMA allocates the time resources in an orthogonal manner such that users are
never transmitting at the same time (line D-C in the rate region).

• SISO: both TDMA and SIC exploit a single degree of freedom but TDMA rate
region is strictly smaller than the one achievable with SIC.

• SIMO: TDMA incurs a big loss compared to SIMO MAC (with MMSE-SIC) as it
only exploits a single degree of freedom despite the presence of min {nr , K} degrees
of freedom achievable with SIMO MAC at high SNR.

• MIMO: As nt increases, the gap between the TDMA and MIMO MAC rate regions
decreases.

245 / 494
Ergodic Capacity Region of Fast Fading Channels: Perfect
CSIT
• The ergodic capacity region is the set of achievable long-term average rates
R1 , . . . , RK where the averaging is taken w.r.t. all channel realizations.
• The rate region can therefore be extended to fast fading channels as
(  )
X 1 X −1 H
Rq ≤ E log2 det Inr + 2 Λq Hul,q Qul,q Hul,q , ∀S ⊆ K
q∈S
σn q∈S

where the input covariance matrices are subject to power constraints.

• short-term power constraint: Tr{Qul,q } = Es,q
– similar to deterministic channels

• long-term power constraint: E Tr{Qul,q } = Es,q where the average power is
computed over a duration Tp >> T
– complicated scenario
– Qul,q and its trace change according to the channel gain subject to the constraint that
the average Tr{Qul,q } over a duration Tp should equal Es,q .
– change defined by a power control policy that maps a channel realization to Tr{Qul,q }
∀q
– However, there may be multiple power control policies that meet the long-term power
constraint. The ergodic capacity region is then given by the union of the capacity
regions, each region corresponding to a given power control policy.

246 / 494
Fast Fading - Perfect CSIT

• SISO MAC with long-term power constraint

– Sum-rate maximization strategy: allow a single user to transmit at a time! That user is
the one with the largest weighted channel gain
 
Λ−1
q
hul,q 2
q ⋆ = arg max
q νq
where νq is a Lagrangian multiplier chosen to satisfy the power constraint. The other
users remain quiet until their own weighted channel gain becomes the largest.
– Reminiscent of the water-filling power allocation. A user is allocated more power when
its channel is good and less power when its channel is bad.
– Dynamic TDMA based on channel measurement and dynamic user selection and power
control is optimal to maximize the sum-rate! → multi-user diversity!
• SIMO MAC with long-term power constraint
– The power is allocated to more than one user at a time.
– As nr increases, irrespectively of the number of users K, the optimal power allocation
relying on CSIT provides a marginal gain over the constant power allocation strategy
that utilizes only the path loss and shadowing information (but no small scale fading
information). Hence, perfect CSIT is of decreasing value as nr increases.
– The multi-user diversity gain indeed decreases as nr increases due to channel hardening
effect.

247 / 494
Fast Fading - Partial Transmit Channel Knowledge

• Hul,q is not known to the transmitter q ∀Q → we cannot adapt Qul,q at all time
instants
• Rate of information flow between a subset of users S and Rx at time instant k over
channels Hul,k,q ∀q ∈ S
 
1 X −1
log2 det Inr + 2 Λq Hul,k,q Qul,q HH
ul,k,q .
σn q∈S

Such a rate varies over time according to the channel fluctuations. The average rate
of information flow over a time duration T >> Tcoh is
T −1  
1 X 1 X −1
log2 det Inr + 2 Λq Hul,k,q Qul,q HH
ul,q .
T σn q∈S
k=0

• The rate region is a K-dimensional polyhedron in general and a pentagon in the

two-user case. The corner points are still achieved by MMSE-SIC.

248 / 494
Fast Fading - Partial Transmit Channel Knowledge

• The ergodic capacity region is obtained as the union of all the K-dimensional
polyhedrons whose corresponding input covariance matrices satisfy the power
constraints.
Proposition
The ergodic capacity region C¯M AC of the Gaussian fast fading MIMO MAC is
the set of all achievable rate vectors (R1 , . . . , RK ) given by
 P 
[  (R1 , . . . , RK) : q∈S Rq ≤  
P Λ−1
q H .
 E log2 det Inr + q∈S σ2 Hul,q Qul,q Hul,q , ∀S ⊆ K 
Tr{Qul,q }≤Es,q n
Qul,q ≥0,∀q

• T >> Tc to average out the noise and the channel fluctuations.

• Assuming i.i.d. Rayleigh fading for all the users, equal power allocation, i.e.
Es,q
Qul,q = nt,q Int,q , is optimal to achieve the entire ergodic capacity region of the
P
MIMO MAC and the sum-rate capacity scales linearly with min(nr , K q=1 nt,q ).
• TDMA incurs a loss compared to MMSE SIC for SISO, SIMO and MIMO MAC.

249 / 494
Outage Capacity and Probability in Slow Fading Channels

• The transmitters have only partial transmit channel knowledge in the form of the
channel distribution information.
S
• MAC outage event O = S OS where
(   X )
1 H
OS = Hul : log2 det Inr + 2 Hul,S Qul,S Hul,S < Rq .
σn q∈S

• Outage probability of the MAC is defined as the probability that the target rate
vector (R1 , . . . , RK ) lies outside the achievable rate region.

Definition

The outage probability Pout (R1 , . . . , RK ) of a MIMO MAC with target rate
vector (R1 , . . . , RK ) is given by
!
[
Pout (R1 , . . . , RK ) = min P OS .
{Qul,q ≥0,Tr(Qul,q )≤Es,q }∀q S

250 / 494
Diversity-Multiplexing trade-off of i.i.d. Rayleigh Slow
Fading Channels
• Assume that all users have the same transmit power constraint Es,q = Es ∀q and
experience independent and identically distributed channels with Λq = Λ (so that
ηq = η ∀q) and Hq being i.i.d. Rayleigh fading.
• Asymptotic (i.e. large η) diversity-multiplexing trade-off of the K-user MIMO MAC
Definition
⋆
A diversity gain gd,M AC (gs,1 , . . . , gs,K , ∞) is achieved for the set of K-tuple
multiplexing gains (gs,1 , . . . , gs,K ) if


Rq η
lim
= gs,q , ∀q
η−→∞ log
2 η


log2 Pout (R1 , . . . , RK ) ⋆
lim
= −gd,M AC (gs,1 , . . . , gs,K , ∞)
η−→∞ log2 η
⋆
The curve gd,M AC (gs,1 , . . . , gs,K , ∞) as function of (gs,1 , . . . , gs,K ) is known
as the asymptotic diversity-multiplexing trade-off of the MIMO MAC.

• The DMT in the MAC differs from that of the single-link MIMO channel by the fact
that coding can only P be performed across antennas belonging to the same user and
not jointly across all nt,q antennas.
251 / 494
Diversity-Multiplexing trade-off of i.i.d. Rayleigh Slow
Fading Channels
⋆
• Asymptotic symmetric (nt,q = nt and gs,q = gs ∀q) DMT gd,M AC (gs , ∞) of MIMO
nr
MAC for nt > K+1

– Lightly loaded: multiple access

is provided without
compromising individual users’
performance and admitting
more users in the system does
not degrade the users’
performance.
– Heavily loaded: tradeoff of a
giant MIMO system made of
Knt transmit antennas
transmitting at a multiplexing
rate Kgs . User performance is
affected by the presence of
other users.

⋆ nr
• Asymptotic symmetric DMT gd,M AC (gs , ∞) of MIMO MAC for nt ≤ K+1
is the
same as single-link MIMO.
252 / 494
MIMO BC System Model

• Downlink multi-user MIMO (MU-MIMO) transmission

– total number of K users (K = {1, . . . , K}) distributed in a cell,
– nr,q receive antennas at mobile terminal q (we simply drop the index q and write nr if
nr,q = nr ∀q)
– nt transmit antenna at the base station
• Received signal (we drop the time dimension)

yq = Λq−1/2 Hq c′ + nq

where
– yq ∈ nr,q
– Hq ∈ nr,q ×nt models the small scale time-varying fading process and Λ−1
q refers to
the large-scale fading accounting for path loss and

shadowing
– nq is a complex Gaussian noise CN 0, σn,q2 I
nr,q .
• The input covariance
 matrix
is defined as the covariance matrix of the transmit
signal as Q = E c′ c′H .
• Power constraint: Tr{Q} ≤ Es .

253 / 494
MIMO BC System Model

• By stacking up the received signal vectors, the noise vectors and the channel
matrices of all K users,
h iT
y = y1T , . . . , yK
T
,
h iT
n = nT1 , . . . , nTK ,
h iT
−1/2 T −1/2
H = Λ1 H1 , . . . , ΛK HTK ,
the system model also writes as
y = Hc′ + n.
H is assumed to be full-rank as it would be the case in a typical user deployment.
• SNR of user q defined as ηq = Es Λ−1 2
q /σn,q .
• Perfect instantaneous channel state information (CSI) at the Tx and all Rx.
• Generally speaking, c′ is written as the superposition of statistically independent
signals c′q
XK
c′ = c′q .
q=1

The input covariance matrix of user q is defined as Qq = E c′q c′H
q .
254 / 494
Capacity Region of two-user SISO Deterministic BC

• In two-user SISO MAC, point A was obtained by canceling user 2’s signal first such
that user 1 is left with Gaussian noise.
• Let us apply the same philosophy to the SISO BC:
– transmit c′ = c′1 + c′2 , with power of c′q denoted as sq
– user 1 cancels user 2’s signal c′2 so as to be left with its own Gaussian noise
– user 2 decodes its signal by treating user 1’s signal c′1 as Gaussian noise.
• Achievable rates of such strategy (with sum-power constraint s1 + s2 = Es )
 
Λ−1
1 s1 2
R1 = log2 1 + 2 |h1 |
σn,1
!
2
Λ−12 |h2 | s2
R2 = log2 1 + 2 2 .
σn,2 + Λ−1
2 |h2 | s1

• Careful! For user 1 to be able to correctly cancel user 2’s signal, user 1’s channel has
to be good enough to support R2 , i.e.
!
2
Λ−1
1 |h1 | s2
R2 ≤ log2 1 + 2 2 .
σn,1 + Λ−11 |h1 | s1

• The channel gains normalized w.r.t. their respective noise power should be ordered
2
Λ−1
2 |h2 | Λ−1 |h1 |2
2
≤ 1 2 .
σn,2 σn,1 255 / 494
Capacity Region of two-user SISO Deterministic BC

• If the ordering condition is satisfied, the above strategy achieves the boundary of the
capacity region of the two-user SISO BC for any power allocation s1 and s2
satisfying s1 + s2 = Es .
• The capacity region is given by the union of all rate pairs (R1 , R2 ) over all power
allocations s1 and s2 satisfying s1 + s2 = Es .

256 / 494
Capacity Region of K-user SISO Deterministic BC
−1/2
• Define hq = Λq hq /σn,q . Assume |h1 |2 ≥ |h2 |2 ≥ . . . ≥ |hK |2 .
Proposition

With the ordering |h1 |2 ≥ |h2 |2 ≥ . . . ≥ |hK |2 , the capacity region CBC of the
Gaussian SISO BC is the set of all achievable rate vectors (R1 , . . . , RK ) given
by
   
 2 
[ |h q | s q
(R1 , . . . , RK ) : Rq ≤ log2 1 + hP i  , ∀q .
PK
sq :
 1 + |h |2 q
q−1
s p

q=1 sq =Es p=1

Proposition
The sum-rate capacity of the SISO BC is achieved by allocating the transmit
power to the strongest user
   
CBC = log2 1 + Es max |hq |2 = log2 1 + max ηq |hq |2 .
q=1,...,K q=1,...,K

Recall that the MAC sum-rate capacity is obtained with all users simultaneously
transmitting at their respective full power.
257 / 494
Achievability of the SISO BC Capacity Region

• Receiver cancellation - Superposition coding with SIC and appropriate ordering :

– User ordering: decode and cancel out weaker users signals before decoding their own
signal.
– The weakest user decodes only the coarsest constellation. The strongest user decodes
and subtracts all constellation points in order to decode the finest constellation.
• Transmitter cancellation - Dirty-Paper Coding (DPC)
– Assume a system model y = hc′ + i + n with i, n Gaussian interference and noise.
Simply subtracting i for transmit signal is not a good idea!

Proposition
If Tx has full (non-causal) knowledge of the interference, the capacity of the dirty
paper channel is equal to the capacity of the channel with the interference completely
absent.

– By encoding users in the increasing order of their normalized channel gains, DPC
achieves the capacity region of the SISO BC.

Example
Assume |h1 |2 ≥ |h2 |2 . By treating user 2’s signal c′2 as known Gaussian interference
at Tx and encoding user 1’s signal c′1 using DPC, user 1 can achieve a rate as high as
if user 2’s signal was absent. User 2 treats user 1’s signal as Gaussian noise.

258 / 494
Achievability of the SISO BC Capacity Region

Proposition
With the appropriate cancellation/encoding ordering, Superposition Coding
with SIC and Dirty-Paper Coding are both optimal for achieving the SISO BC
capacity region.

Proposition

The SISO BC sum-rate capacity is achievable with dynamic TDMA (to the
strongest user), Superposition Coding with SIC (with the appropriate
cancellation ordering) and Dirty-Paper Coding (with the appropriate encoding
ordering).

259 / 494
Capacity Region of MIMO BC and its Achievability

• MAC with multiple Rx antennas provides a tremendous capacity increase compared

to suboptimal TDMA. So does BC with multiple Tx antennas!
• MIMO BC difficult problem: users’ channels cannot be ranked anymore.
• Assume an increasing encoding order from user 1 to K:
1 Encode user 1’s signal into c′1 .
2 With full knowledge of c′1 , encode user 2’s signal into c′2 using DPC: c′1 appears
invisible to user 2 but c′2 appears like a Gaussian interference to user 1.
3 With full knowledge of user 1 and user 2’s signals, encode user 3’s signal into c′3 using
DPC.
4 ... till K users are encoded.
• A given user q sees signals from users p > q as a Gaussian interference but does not
see any interference signals from users p < q:
2 I −1 P  H
– Covariance of Noise plus Interference at user q: σn,q nr,q + Λq Hq p>q Qp Hq .
– With a MMSE receiver that whitens the colored Gaussian interference (same as in
MAC)
    −1 
X
−1 H 2 −1 H
Rq = log2 det Inr,q + Λq Hq Qq Hq
 σn,q Inr,q + Λq Hq  Qp H q
 
p>q

• Capacity region: Repeat

P for all covariance matrices Q1 , . . . , QK satisfying the
sum-power constraint q Tr {Qq } = Es and all user ordering.
• Only DPC can achieve the MISO/MIMO BC sum-rate capacity.
260 / 494
BC-MAC Duality

• There are similarities between BC and MAC:

– Both MAC and BC deal with received signal(s) expressed as a sum of K (Gaussian)
codewords scaled by the wireless channel and perform SIC at the receiver(s).
– The receiver in SISO MAC receives the sum of K (Gaussian) codewords (after
propagating through the wireless channels) and decodes each of those signals using SIC.
– In the degraded SISO BC, the transmitter sends a sum of K (Gaussian) codewords
using superposition coding and each receiver also decodes its own codeword using SIC.

• Can those similarities be formally characterized? Yes, by the MAC-BC or UL-DL

duality.
• Interestingly, the BC capacity region can be characterized in terms of the capacity
region of a dual MAC and vice-versa.
– By dual MAC, we here refer to the channel obtained by converting the transmitter in
the BC into a receiver and by converting the receivers in the BC into the transmitters.
– The BC and dual-MAC have the same channel gains and the noise variances at their
respective receivers are equal.
– The power constraint for the BC equals the sum of the individual power constraints of
the dual MAC.

261 / 494
SISO BC-MAC Duality
 −1/2 −1/2 T
• Assuming a SISO BC over a deterministic channel h = Λ1 h1 , . . . , Λ K hK
2 2
with receiver noise powers σn,1 , . . . , σn,K , we express the SISO BC in the equivalent
system model with unit variance receiver noises and normalized channel gains
−1/2
h q = Λq hq /σn,q such that h = [h1 , . . . , hK ]T .
• The system model y = hc′ + n for SISO then writes equivalently as

ydl = hc′dl + ñdl

where ñ is a complex Gaussian noise CN 0, IK .
• We can also define the MAC where user q’s uplink channel hul,q is given by hq and
the receiver noise power is equal to one as

yul = hT c′ul + ñul

where ñ is a complex Gaussian noise CN 0, 1 .
• The MAC is the dual of the BC and vice-versa.

262 / 494
SISO BC-MAC Duality

• Usefulness: Given the difficulty to characterize the BC capacity region, the

MAC-to-BC duality is very helpful to express the BC capacity region as a function of
the capacity region of its dual MAC.
Proposition
The capacity region of a Gaussian SISO BC with power constraint Es over a
deterministic channel h = [h1 , . . . , hK ]T with unit variance receiver noise,
denoted explicitly as CBC (Es , h), is equal to the union of the capacity regions
of thePdual MAC with individual power constraints Es,q (q = 1, . . . , K) such
that K q=1 Es,q = Es
[
CBC (Es , h) = CM AC (Es,1 , . . . , Es,K , h)
{Es,q }
PK
: q=1 Es,q =Es
∀q
( P )
[ (R1 , . . . , RK ) : q∈S Rq ≤

= P
log2 1 + q∈S |hq |2 Es,q , ∀S ⊆ K
{Es,q }∀q :
PK
q=1 Es,q =Es

where CM AC (Es,1 , . . . , Es,K , h) is capacity region of the SISO MAC with the
−1/2
channel gains Λq hul,q replaced by the normalized channel gain hq and the
2
noise power σn = 1. 263 / 494
SISO BC-MAC Duality

• Two-user SISO BC capacity region characterized in the terms of the capacity region
of its dual MAC

264 / 494
MIMO BC-MAC Duality

• The system model y = Hc′ + n can be written equivalently as

ydl = Hc′dl + ñdl

 T −1/2
Λq Hq
where H = HT1 , . . . , HTK with Hq = σn,q
and ñdl is a complex Gaussian noise


CN 0, I q nr,q .
P

• The dual uplink channel has K users and nt receive antennas

yul = HH c′ul + ñul

where c′ul is the vector of transmitted signals from the K users, yul is the received
signal vector

at the nt receive antennas and ñul is a complex Gaussian noise
CN 0, Int .

265 / 494
MIMO BC-MAC Duality

Proposition
The capacity region of the MIMO BC with power constraint Es over a
deterministic channel H is equal to the union of the capacity region P of the dual
MIMO MAC with individual power constraints Es,q such that K q=1 Es,q = Es
[  
CBC (Es , H) = CM AC Es,1 , . . . , Es,K , HH
{Es,q }∀q : K
P
q=1 Es,q =Es
 P 
[  (R1 , . . . ,RK ) : q∈S Rq ≤  
= P H
 log2 det Int + q∈S Hq Qul,q Hq , ∀S ⊆ K 
{Qul,q ≥0}∀q ,
q=1 Tr{Qul,q }≤Es
PK

where CM AC Es,1 , . . . , Es,K , HH is teh MIMO MAC capacity region with the
channel matrix Hul replaced by HH and the noise power σn2 = 1.

266 / 494
MIMO BC-MAC Duality

Proposition
The sum-rate capacity of the MIMO BC is equal to the sum-rate capacity of
the sum power dual MIMO MAC
 
CBC (H, Es ) = CM AC HH , Es
 K
X 
= max log2 det Int + HH
q Qul,q Hq .
{ }
Qul,q ≥0
∀q
,
q=1
{ }
PK
q=1 Tr Q ul,q ≤E s

267 / 494
Bounds on Sum-Rate Capacity of MIMO BC

• Define
– ñ = min {nt , Knr } (nr,q = nr ∀q is assumed),
– λmax,q as the dominant eigenvalue of HH q Hq ,
• Achievable multiplexing gain of ñ
Proposition

The sum-rate capacity of MIMO BC for a deterministic channel H, achievable

with DPC, CBC (H), is lower-bounded as
ñ 
X ηq 
CBC (H) ≥ CBF (H) = log2 1 + αq2
q=1
ñ

for some non-zero channel gains α12 , . . . , αñ

2
and is upper-bounded as
 
1
CBC (H) ≤ nt log2 1 + max ηq λmax,q ,
nt q=1,...,K
CBC (H) ≤ CCSIT (H) ,
K
X
CBC (H) ≤ CCSIT (Hq ) .
q=1
268 / 494
Comparisons with TDMA

• SISO
– Similarly to MAC, TDMA rate region is contained in the BC capacity region.
– The gap between the BC capacity region and the TDMA rate region increases
proportionally with the asymmetry between users normalized channel gains.
– TDMA achieves the sum-rate capacity of SISO BC.
• MIMO
– The maximum sum-rate CT DM A (H) is the largest single-link capacity among K users
CT DM A (H) = max CCSIT (Hq ) .
q=1,...,K


– Define λmax,q and λmax as the largest eigenvalue of HH H
q Hq and H H, λk Hq Hq
H
H
as the non-zero eigenvalues of Hq Hq and n = min {nt , nr }. Assume nr,q = nr ∀q.

Proposition
The maximum TDMA sum-rate, CT DM A (H), is lower bounded as
 
CT DM A (H) ≥ log2 1 + max ηq λmax,q ,
q=1,...,K
n   
X ηq
CT DM A (H) ≥ CCSIT (Hq ) ≥ log2 1 + λk H q H H
q ,
k=1
n

for q = 1, . . . , K, and is upper bounded as

 
1
CT DM A (H) ≤ n log2 1 + max ηq λmax,q .
n q=1,...,K 269 / 494
Comparisons with TDMA

Proposition

For channels H1 , . . . , HK , SNR ηq , number of receive antennas nr , the gain of

DPC over TDMA is upper-bounded by the minimum between the number of
transmit antennas nt and the number of users K
CBC (H)
≤ min {nt , K} .
CT DM A (H)

Intuition:
TDMA exploits at least one spatial dimension with the largest effective SNR among all
users.
– DPC exploits up to nt dimensions. Since the quality of each of those nt dimensions
cannot be larger than the single dimension used in the TDMA lower bound, DPC
cannot achieve a rate larger than nt times the TDMA capacity.

Proposition
For any nt , nr and K, at high SNR (Es → ∞, i.e. ηq → ∞ ∀q),

CBC (H) min {nt , Knr }

≈ .
CT DM A (H) min {nt , nr }
270 / 494
Ergodic Capacity Region of SISO Fast Fading Channels

Perfect CSIT
P
• short-term power constraint q sq = Es : similar to deterministic channels
P
• long-term power constraint EH q sq (H) = Es :
– Power control policy that maps a channel realization to a set of transmit power.
– Ergodic capacity region is given by the union of all achievable rate regions over all
power control policies that satisfy the long-term power constraint.
– Sum-rate maximization: the sum-rate capacity can also be achieved by transmitting to
the strongest user in each fading state
Λ−1
q |hq |
2
q ⋆ = arg max |hq |2 = arg max 2
q=1,...,K q=1,...,K σn,q
and the power in each fading state can be optimized following the time domain
water-filling solution.
– Observe the similarity with SISO MAC. The user to be selected is slightly different.
• In MISO and MIMO BC, not sufficient to transmit to a single user at a time to
achieve the sum-rate capacity (similarly to SIMO/MIMO MAC).

Partial CSIT
• No channel ordering and no way know the interference to other users.
• TDMA is an appropriate strategy. Huge loss compared to perfect CSIT!
271 / 494
Outage Capacity and Probability in Slow Fading Channels

• The notion of diversity-multiplexing trade-off is more meaningful in the absence of

CSIT when the transmitter cannot adapt its transmit strategy as a function of the
channel realization.
• However, SISO/MIMO BC both critically depend on CSIT. With only partial channel
knowledge at the transmitter, the performance drops significantly.
• Assume perfect CSIT

Proposition

For a MISO BC with nt transmit antennas and K ≤ nt single-antenna users

(whose concatenated channel matrix entries are Rayleigh i.i.d.) and given the
fixed rates R1 , . . . , RK ,
⋆
gd,BC (0, . . . , 0, ∞) ≤ nt .

The diversity gain of MU-MIMO precoding in MISO BC is not larger than the
diversity gain of transmit beamforming in a single link.

272 / 494
Multi-User MIMO - Scheduling and Precoding
(Downlink)

273 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 12
Section: 12.1,12.5,12.6,12.8

274 / 494
Introduction

• BC: K >> nt , MAC: K >> nr → All users cannot be scheduled at the same time.
– Which users to schedule?
– How to account for fairness?

• DPC is optimal in MIMO BC but is very complex to implement.

– Can we derive suboptimal strategies? Yes, there are various linear and non-linear
precoding techniques
– How to design suboptimal linear precoders?
– What is the performance of those precoders combined with scheduling?

• What if we do not have perfect channel knowledge at the transmitter to design the
precoders in MIMO BC?

275 / 494
System Model

• Downlink multi-user MIMO (MU-MIMO) transmission

– total number of K users (K = {1, . . . , K}) distributed in a cell,
– nr,q receive antennas at mobile terminal q (we simply drop the index q and write nr if
nr,q = nr ∀q)
– nt transmit antenna at the base station
• Received signal (we drop the time dimension)

yq = Λq−1/2 Hq c′ + nq
where
– yq ∈ nr,q
– Hq ∈ nr,q ×nt models the small scale time-varying fading process and Λ−1
q refers to
the large-scale fading accounting for path loss and

shadowing
– nq is a complex Gaussian noise CN 0, σn,q2 I
nr,q .
• Long term SNR of user q defined as ηq = Es Λ−1 2
q /σn,q .
• Generally speaking, c′ is written as the superposition of statistically independent
signals c′q
XK
c′ = c′q .
q=1

• Power constraint: Tr{Q} ≤ Es with Q = E c′ c′H .
276 / 494
System Model - Linear Precoding

• scheduled user set, denoted as K ⊂ K, is the set of users who are actually scheduled
(with a non-zero transmit power) by the transmitter at the time instant of interest.
• The transmitter serves users belonging to K with ne dataP streams and user q ∈ K is
served with nu,q data streams (nu,q ≤ ne ). Hence, ne = q∈K nu,q .
• Linear Precoding
X X
c′ = Pc = WS1/2 c = Pq c q = Wq S1/2
q cq
q∈K q∈K

where
– c is the symbol vector made of ne unit-energy independent symbols.
– P ∈ nt ×ne is the precoder subject to Tr{PPH } ≤ Es , made of two matrices: a
power control diagonal matrix denoted as S ∈ ne ×ne and a transmit beamforming
matrix W ∈ nt ×ne .
– Pq ∈ nt ×nu,q , Wq ∈ nt ×nu,q , Sq ∈ nu,q ×nu,q , and cq ∈ nu,q are user q’s
sub-matrices and sub-vector of P, W, S, and c, respectively.
• The received signal yq ∈ nr,q is shaped by Gq ∈ nu,q ×nr,q
and the filtered
received signal zq ∈ nu,q at user q writes as
z q = G q yq ,
X
= Λq−1/2 Gq Hq Wq S1/2
q cq + Λq−1/2 Gq Hq Wp S1/2
p c p + Gq nq .
p∈K, p6=q

277 / 494
Multi-User Diversity

• In single-link systems, channel fading is viewed as a source of unreliability mitigated

through diversity techniques (e.g. space-time coding).

• In multi-user communications, fading is viewed as a source of randomization that

can be exploited!

• Multi-User (MU) diversity is a form of selection diversity among users provided by

independent time-varying channels across the different users.

• Provided that the BS is able to track the user channel fluctuations (based on
feedback), it can schedule transmissions to the users with favorable channel fading
conditions, i.e. near their peaks, to improve the total cell throughput.

• Recall that MU diversity was already identified as part of the sum-rate maximization
in SISO BC.

278 / 494
Multi-User Diversity Gain in SISO

• Assume that the fading distribution of the K users are independent and identically
(Λ−1
q = Λ−1 and channel gains hq are drawn from the same) Rayleigh distributed
2
and that users experience the same average SNR ηq = η (σn,q = σn2 ) ∀q:
yq = Λ−1/2 hq c′ + nq .
• Assume MU-SISO where one user is scheduled at a time in a TDMA manner: select
the user with the largest channel gain.
• Mathematically same as antenna selection diversity.
• Average SNR gain
– Average SNR after user selection ρ̄out
  K
X 1
ρ̄out = E η max |hq |2 =η .
q=1,...,K
q=1
q
– SNR gain provided by MU diversity gm
K
ρ̄out X 1 K→∞
∼
gm = = = log(K).
η q=1
q
gm is of the order of log(K) and hence the gain of the strongest user grows as log(K)!
• Heavily relies on CSIT (partial or imperfect feedback impacts the performance) and
independent user fading distributions (correlated fading or LOS are not good for MU
diversity)
279 / 494
Multi-User Diversity Gain in SISO

• Sum-rate capacity
  
2
C̄T DM A = E {CT DM A } = E log2 1 + η max |hq | .
q=1,...,K

– low SNR  
C̄T DM A ≈ E max |hq |2 η log2 (e) ≈ gm Cawgn .
q=1,...,K
Observations: capacity of the fading channel log(K) times larger than the AWGN
capacity.
– high SNR (Use Jensen’s inequality: Ex {F (x)} ≤ F (Ex {x}) if F concave)
  
C̄T DM A ≈ log2 (η) + E log2 max |hq |2 ,
q=1,...,K
  
≈ Cawgn + E log2 max |hq |2 ,
q=1,...,K
(a)
  
≤ Cawgn + log2 E max |hq |2 ,
q=1,...,K

= Cawgn + log2 (gm ) .

Observations: capacity of a fading channel is larger than the AWGN capacity by a
factor roughly equal to log2 (gm ) ≈ log log (K).
• Fading channels are significantly more useful in a multi-user setting than in a
single-user setting 280 / 494
Multi-User Diversity

• In MU-MIMO, the performance is function of the channel magnitude but also of the
spatial directions and properties of the channel matrices.
• MU diversity offers abundant spatial channel directions and allows to appropriately
choose users with good channel matrix properties or spatial separations.
• Opportunistic Beamforming: precode multiple streams along the unitary precoding
matrix W (orthogonal beams). For a large number of users, thanks to MU diversity,
each beam matches one user channel with a high probability and orthogonality of
beams prevents users from experiencing multi-user interference
K→∞
yq = Λq−1/2 hq WS1/2 c + nq = Λq−1/2 khq k s1/2
q cq + n q .

– The terminal only measures the effective channel, i.e. the channel precoded by each
beam, and reports the SNR (or CQI) for one or multiple beam(s).
– Works well only for very large K.

281 / 494
Multi-User Diversity

• Few fundamental differences with classical spatial/time/frequency diversity:

– Diversity techniques, like space-time coding, mainly focus on improving reliability by
decreasing the outage probability in slow fading channels. MU diversity on the other
hand increases the data rate over time-varying channels.

– Classical diversity techniques mitigate fading while MU diversity exploits fading.

– MU diversity takes a system-level view while classical diversity approaches focus on a

single-link. This system-level view becomes increasingly important as we shift from
single-cell to multi-cell scenarios.

282 / 494
Resource Allocation, Fairness and Scheduling Criteria

• An appropriate scheduler should allocate resources (time, frequency, spatial, power)

to the users in a fair manner while exploiting the MU diversity gain.
• Goal of the resource allocation strategy at the scheduler: maximize the utility metric
U.  ′⋆ ⋆ ⋆
c , G , K = arg ′ max U
c ,G,K⊂K
′⋆ ⋆
where c is the optimum transmit vector, G denotes the optimum set of receive
beamformers, and K⋆ ⊂ K refers to the optimum subset of users to be scheduled.
• Two major kinds of resource allocation strategies:
– rate-maximization policy : maximizes the sum-rate - no fairness among users
– fairness oriented policy, commonly relying on a proportional fair (PF) metric:
maximizes a weighted sum-rate and guarantees fairness among users.
• Those two strategies can be addressed by using two different utility metrics:
 ′⋆ ⋆ ⋆ X
c , G , K = arg ′ max wq Rq
c ,G,K⊂K
q∈K

where
– rate-maximization approach: wq = 1
γ
– proportional fair approach: wq = R̄q (R̄q is the long-term average rate of user q and
q
γq is the Quality of Service (QoS) of each user).
283 / 494
Practical Proportional Fair Scheduling

• The long-term average rate R̄q of user q is updated using an exponentially weighted
low-pass filter such that the estimate of R̄q at time k + 1, denoted as R̄q (k + 1), is
function of the long-term average rate R̄q (k) and of the current rate Rq (k) at
current time instant k as outlined by
(
(1 − 1/tc ) R̄q (k) + 1/tc Rq (k), q ∈ K⋆
R̄q (k + 1) =
(1 − 1/tc ) R̄q (k) , q∈/ K⋆

where tc is the scheduling time scale and K⋆ refers to the scheduled user set at time
k. The resources should thus be allocated at time instant k as
 ′⋆ ⋆ ⋆ X Rq (k)
c , G , K = arg ′ max γq .
c ,G,K⊂K
q∈K
R̄q (k)

• The scheduling time scale tc is a design parameter of the system that highly
influences the user fairness and the performance
– Very large tc : assuming all users experience identical fading statistics and have the
same QoS, the PF scheduler is equivalent to the rate-maximization scheduler, i.e. users
contributing to the highest sum-rate are selected.
– Small tc : assuming all users have the same QoS, the scheduler divides the available
resources equally among users (Round-Robin scheduling). No MU diversity is exploited.
284 / 494
Proportional Fair Scheduling

• Sum-rate of SISO TDMA with PF scheduling at SNR=0 dB as a function of the

number of users K, the scheduling time scale tc and the channel model
p
hk = ǫhk−1 + 1 − ǫ2 nk
with ǫ the channel time correlation coefficient.
2.6
ε=0.98 t =1e4
2.4 ε=0.72
c

ε=0
TDMA sum−rate [bps/Hz]

2.2

1.8
tc=50
1.6

1.4

1.2
t =1.1
c
1

0.8
0 10 20 30 40 50
Number of Users K

285 / 494
User Grouping

• Given the presence of K users in the cell, the scheduler for MU-MIMO aims at
finding the best scheduled user set among all possible candidates within K.

• The exhaustive search is computationally intensive. Assuming a single stream

P per user and ne ≤ min {nt , K}, a search like (with
transmission
R (K) = q∈K wq Rq )

K⋆ = arg max R (K)

K⊂K
ne ≤min{nt ,K}

requires to consider a large number of different sets and has a complexity that
quickly becomes cumbersome as K increases.

286 / 494
User Grouping

• Lower complexity user grouping algorithms: Greedy User Selection, Semi-orthogonal

User Selection.
• Greedy User Selection consists in successively adding a user to the tentative
scheduled user set only if the weighted sum-rate is increased.


– Initialization step: We fix n = 1, K(0) = ∅, R K(0) = 0 and Done = 0.
– Iteration-n: While (n ≤ min {K, nt }) and (not Done), select the user
 
q (n) = arg max R K(n−1) ∪ q (n) .
q∈K\K



If R K(n−1) ∪ q (n) <R K(n−1) , K(n) = K(n−1) and Done = 1, otherwise
K (n) =K (n−1) ∪q (n) and we move the next iteration n + 1. The final scheduled user
set K = K(n) .

287 / 494
User Correlation

• Scheduling orthogonal users provides a larger sum-rate by removing naturally the

multi-user interference.

• The probability of finding orthogonal users is an important indicator of the

performance of MU-MIMO.
• Among K active single-antenna users, the probability of finding nt (semi-)orthogonal
users is investigated.
– Denote as Knt all sets of nt users among the K active users, and as K one of those
sets,
– user correlation u ( )
 
 H
u = E{hw,q } min max h̄k h̄l  ,
∀q K∈Kn k,l∈K
t

288 / 494
User Correlation

• Average value of u as a function of |t| and K, where the average value here refers to
the averaging over different sets of transmit correlation matrices Rt,q (|tq | = |t| ∀q,
phases of the correlation coefficients tq are randomly generated).
n =4
t
1
t=0
0.9 |t|=0.5, 360°
|t|=0.95, 360°
Average User Correlation u

0.8
|t|=0.95, 70°
0.7

0.6

0.5

0.4

0.3

0.2
4 6 8 10 12 14 16 18 20
Number of Users K

– User correlation decreases with K at a higher rate in spatially correlated channels.

– Transmit correlation decreases the probability of finding semi-orthogonal users for
moderate K while it increases such probability for large K.
– Spatially correlated fading at TX detrimental (resp. beneficial) to MU-MIMO for small
(resp. large) K.
– Cell sectorization increases user correlation. 289 / 494
Precoding with Perfect Transmit Channel Knowledge

• Single-link Spatial Multiplexing: Multiple Eigenmode Transmission relies on CSI

knowledge at both the transmitter and the receiver and splits the spatial channel
equalization between the transmitter and the receiver. As a result, the channel is
decoupled into multiple parallel data pipes.
• Unfortunately, this approach cannot be applied to MU-MIMO as the receivers do not
cooperate.
• MIMO BC, DPC optimal but extremely complex. Any suboptimal strategies?
• In MU-MIMO where CSI is available at the transmitter, precoding techniques
reminiscent of the receiver architectures for SM
precoding transmitter side receiver side
Linear Matched Beamforming (MBF) MRC
Linear Zero-Forcing Beamforming ZF
(ZFBF)
Linear Regularized Zero-Forcing Beamform- MMSE
ing (R-ZFBF)
Non-linear Tomlinson-Harashima Precoding SIC
(THP)
Non-linear Vector Perturbation (VP) sphere decoder

290 / 494
Achievable rate

• Maximum rate achievable by user q with linear precoding is

nu,q
X
Rq = log2 (1 + ρq,l )
l=1

where ρq,l denotes the SINR experienced by stream l of user q

2 2
Λ−1
q |gq,l Hq pq,l | Λ−1
q |gq,l Hq wq,l | sq,l
ρq,l = 2 =
2
Il + Ic + kgq,l k σn,q Il + Ic + kgq,l k2 σn,q
2

with pq,l = wq,l sq,l (resp. gq,l ) the precoder (resp. combiner) attached to stream l
of user q, Il the inter-stream interference and Ic the intra-cell interference (also
called multi-user interference)
X −1 X −1
Il = Λq |gq,l Hq pq,m |2 = Λq |gq,l Hq wq,m |2 sq,m ,
m6=l m6=l
nu,p
X X X nX
u,p

Ic = Λ−1 2
q |gq,l Hq pp,m | = Λ−1 2
q |gq,l Hq wp,m | sp,m .
p∈K m=1 p∈K m=1
p6=q p6=q

2
Λ−1
q |hq wq | sq
• If nr = 1, the SINR of user q simply reads as ρq = −1 2 .
| | 2
P
p∈K Λq hq wp sp +σn,q
p6=q
291 / 494
Zero-Forcing Beamforming (ZFBF)

• Most popular MU-MIMO precoder. Assume single receive antenna per user.
• Channel Direction Information (CDI) of user q: h̄q = hq / khq k.
• Idea is to force the intra-cell interference Ic to zero: the precoder of a user q, wq , is
chosen such that hp wq = 0 ∀p ∈ K \ q. Only possible if ne ≤ nt !
• Define h iT
−1/2 T −1/2 T
H = Λi hi , . . . , Λ j hj = DH̄
i,j∈K
with
n o
−1/2 −1/2
D = diag Λi khi k , . . . , Λj khj k ,
i,j∈K
h iT
H̄ = h̄Ti , . . . , h̄Tj .
i,j∈K

The ZFBF aims at designing W = [wi , . . . , wj ]i,j∈K such that HW is diagonal.

• Assuming ne ≤ nt and H̄ is full rank, the precoders can be chosen as the normalized
columns of the right pseudo inverse of H
 −1  −1
F = HH HHH = F̄D−1 = H̄H H̄H̄H D−1 .

Transmit precoder wq for user q ∈ K: wq = F(:, q)/kF(:, q)k = F̄(:, q)/kF̄(:, q)k
where F(:, q) is to be viewed as the column of F corresponding to user q.
292 / 494
Zero-Forcing Beamforming (ZFBF)
 T
• Assuming that c = ci , . . . , cj i,j∈K
, the received signal of user q ∈ K is

yq = Λq−1/2 hq wq s1/2
q cq + n q = d q cq + n q ,
−1/2 1/2 −1/2 khq k 1/2
with dq = Λq h q wq s q = Λ q s .
kF̄(:,q)k q
Observations: MU-MIMO channel with ZFBF is split into ne parallel
(non-interfering) channels.
• The rate achievable by user q is given by
2

Rq = log2 1 + d2q /σn,q .
d2q is low if H is badly conditioned but would get larger if users’ CDI are orthogonal
or quasi-orthogonal.
– reminiscent of the loss caused by noise enhancement incurred by the linear ZF
• For large K, better conditioning of matrix H through the use of user grouping.
• By uniformly allocating the power across user streams sq = Es /ne and by choosing
ne = min {nt , K}, d2q /σn,q
2
= αq2 ηq /ne with αq2 = |hq wq |2 = khq k2 /kF̄(:, q)k2
min{nt ,K} 
X ηq 
CBF (H) = log2 1 + αq2 .
q=1
ñ

At high SNR with ηq = η, CBF (H) ≈ min {nt , K} log2 (ηq ). The multiplexing gain
min {nt , K} is achieved (same as with DPC). 293 / 494
Zero-Forcing Beamforming (ZFBF)

• Illustration of ZFBF precoding for a two-user scenario: (a) non-orthogonal user set,
(b) quasi-orthogonal user set.

294 / 494
Block Diagonalization (BD)

• Extension of ZFBF to multiple receive antennas and multiple streams per user.
• Constraints on the transmit filters targeting user q ∈ K
Λp−1/2 Hp Wq = 0, ∀p 6= q, p ∈ K
nr K̃q ×nt
• Denoting K̃q = K \ q of size K̃q = ♯K̃q , the interference space H̃q ∈ is
h iT
H̃q = . . . Λp−1/2 HTp . . . .
p∈K̃q

• BD filter design forces Wq to
lie in the null space of H̃q : null space of H̃q to be
strictly larger than 0 → r H̃q < nt .
• An orthogonal basis of the null space of H̃q is obtained by taking its SVD
 H
H̃q = Ũq Λ̃q Ṽq Ṽq′

where Ṽq′ refers to the eigenvectors of H̃q associated with the null singular values.
• Assuming
the zero-interference constraint is possible for all users in K̃q and that
r Hq Ṽq′ = nu,q , Wq writes as a linear combination of columns of Ṽq′ as
Wq = Ṽq′ Aq
with some nu,q × nu,q unitary matrix Aq .
295 / 494
Block Diagonalization (BD)

• Multi-user interference is eliminated and each user experiences an equivalent

single-user MIMO channel H̃eq,q = Hq Ṽq′ , for which the optimal solution is obtained
by transmitting along the nu,q dominant right singular vectors of H̃eq,q
 
  Λ̃eq,q 0  H
H̃eq,q = Ũeq,q Ũ′eq,q ′
Ṽeq,q Ṽeq,q
0 0
where Ṽeq,q refers to the nu,q dominant right singular vectors.
• The final beamformer for user q writes as
Wq = Ṽq′ Ṽeq,q .
• Applying Gq = ŨH
eq,q , the equivalent channel of each user is

zq = Gq yq = Λq−1/2 Λ̃eq,q S1/2

q c q + Gq nq .

• Achievable sum-rate (with λ̃eq,q,m diagonal entries of Λ̃2eq,q )

!
X nX u,q
Λ−1
q λ̃eq,q,m
log2 1 + sq,m 2
q∈K m=1
σn,q
P Pnu,q
For a sum-power constraint q∈K m=1 sq,m = Es , the optimal power allocation is
obtained by water-filling.
296 / 494
Tomlinson-Harashima Precoding (THP)

• Tomlinson-Harashima Precoding (THP) originally designed to cope with ISI in SISO

point-to-point channel when the channel impulse response is known to Tx
– an alternative to a receiver-based decision-feedback equalizer (DFE).
– DFE for ISI channels is the analog to the SIC receivers used for MIMO channels

• For a SISO channel, the intended signal at time instant k, h [0] ck , is affected by the
PL−1
ISI ik = l=1 h [l] ck−l (ignoring the noise)

• If h [l] ∀l known to Tx and given that the previous transmitted symbols ck−l are
known to Tx, ISI ik is known and the transmitter can make use of that knowledge.
– DFE or SIC prone to error propagation, but THP not affected as the previously
transmitted symbols ck−l are perfectly known to Tx.

297 / 494
THP

• Assume the received signal

y = c̃ + i + n
with
– c̃ a precoded version of the desired symbol c belonging to a symbol constellation B
with average transmit power Es ,
– i the interference (e.g. ISI), Gaussian distributed with variance σi2
– n is the Gaussian noise with variance σn 2

• If the Rx treats i as noise, achievable rate given by log2 1 + Es /(σi2 + σn2 ) .

• If i known to Tx and not known at Rx: how to design c̃ ?

2

– Subtracting i to c, i.e. c̃ = c − i? No! Rate is log2 1 + Es − σi2 /σn assuming
Es ≥ σi2 . Significant power penalty especially for large i.
– IDEA: THP infinitely replicates the constellation B and transmits c̃ = Qc (i) − i where
Qc (i) is the replica of c closest to i.

298 / 494
THP

• Replication of the QPSK constellation made of symbols ◦, , ♦, N: the set of N, i.e.

{N}, corresponds to the equivalence class of N.

299 / 494
THP

• Set of replicas of a constellation symbol c denoted as the equivalence class of c.

• Qc (i) is therefore the point in the equivalence class of c closest to i.
• Qc (i) viewed as a quantizer of i and c̃ = Qc (i) − i is the quantization error.
• Received signal
y = c̃ + i + n = Qc (i) + n.
The receiver finds the point in the replicated constellation closest to y and decodes
to the equivalence class containing that point.
• The error probability of THP is roughly the same as if c ∈ B is transmitted in the
absence of interference!
– The quantization error Qc (i) − i is always bounded even when i is very large.
– Constellation points c located at the border of the constellation B experience a slightly
higher decoding error probability due to the presence of the replicas and the probability
of confusing c with points belonging to the replicated constellations.
– In the limit of high SNR, the performance gap is negligible.
– The power consumption of THP  is slightly higher than in the absence of interference:
For random interference i, E |c̃|2 is slightly higher than the average power of
constellation B.

300 / 494
THP

• At low SNR, enhancement possible by scaling i with a coefficient α

c̃ = Qc (αi) − αi

i.e. the transmitter finds the point in the equivalence class of c closest to αi and
transmits the quantization error between that point and αi.

• The receiver scales the received signal by α

αy = α (c̃ + n) + αi

and finds the constellation point nearest to αy.

• A suitable value for α is equal to the MMSE scaling factor Es /(Es + σn2 ). By doing
so, αy is a linear MMSE estimate of c̃ but shifted by αi.

301 / 494
n-dimensional THP and DPC

• System model
y = c̃ + i + n
• Choose
c̃ = Qc (αi) − αi
• With high dimensional coding, such
precoding technique achieves the same capacity
as AWGN channel log2 1 + Es /σn2 , i.e. as in the absence of interference.

• This scheme is commonly known as Costa precoding or Dirty-Paper Coding (DPC).

• Alternative representation: view the replicated constellation as a lattice L
– use a modulo operation such that c̃ = [c − αi] mod L where [a] mod L = a − QL (a)
with QL (a) the lattice vector quantizer based on lattice L.
– At the receiver, the received signal is passed through the modulo operation
z = αy mod L before taking the decision on c.

302 / 494
THP for MU-MIMO

• THP used as a MU-MIMO precoder by precoding sequentially the users data.

– Assume that the user data are encoded in an increasing order.
– Any signal generated at step p can be exploited to encode data at step q, for q > p.

• Transmit vector as
c′ = Pc̃ = WS1/2 c̃.

• Assuming a predefined K and a predefined user ordering in increasing order of the

user index, the signal at user q writes as
X −1/2
yq = Λq−1/2 hq wq s1/2
q c̃q + iq + Λq hq wp s1/2
p c̃p + nq
p>q

where
P −1/2 1/2
– Known interference at Tx: iq = p<q Λq hq wp sp c̃p (recall that c̃p , p < q, have
been previously computed).
– The interference from p > q is treated as an additional noise.

303 / 494
THP for MU-MIMO

• Relying on THP for SISO channels, we design c̃q

" #
α q iq
c̃q = cq − −1/2 1/2
mod L
Λq h q wq s q
where αq is the scaling factor of user q.
– Operation repeated sequentially for the ne data streams to be encoded.
– THP for MU-MIMO precoding has a very reasonable complexity as it involves only
computation of a sequence of ne complex scalar quantizations (modulo operation).
• At receiver q, the received signal yq is passed through the modulo operation
αq yq
zq = −1/2 1/2
mod L
Λq h q wq s q
before taking the decision on cq .
• Assuming THP can subtract the interference perfectly, i.e. as DPC, the SINR of user
q reads as
2
Λ−1
q |hq wq | sq
ρq = P −1 2 2
p>q Λq |hq wp | sp + σn,q
P
and Rq = log2 (1 + ρq ) and R (K) = q∈K wq Rq .
– Note the difference with conventional SINR expression!
– Note the similarity with capacity of MISO BC
– Sum-rate further maximized over all unitary beamforming vectors and power
allocations satisfying the sum-power constraint. 304 / 494
QR-THP (or ZF-THP)

• Choose the transmit beamformer W such that the interference due to p > q is
eliminated, i.e. zero-forced.
• Let H = RQ be the QR decomposition of H
– R is a ne × ne lower triangular
– Q is a ne × nt matrix with orthonormal rows.

• By selecting W = QH , the system model is simplified as

yq = Λq−1/2 hq wq s1/2
q c̃q + iq + nq .

– Since HW = R is lower diagonal, the interference caused by users p > q is forced to

zero for user q.

• Encoding of c̃q can be obtained by successive THP modulo operation.

−1/2 1/2 1/2
• The SINR simply boils down to ρq = d′2 2 ′
q /σn,q with dq = Λq h q wq s q = Rqq sq
th
with Rq,q the (q, q) entry of R.
• DPC can be implemented using THP.

305 / 494
Sum-Rate Scaling Laws

• For a large number of users in i.i.d. Rayleigh Fading

Proposition
Assuming ηq = η, ∀q, for nt , nr and η fixed, the average maximum sum-rates
of TDMA, DPC and BF in i.i.d. Rayleigh fading channels (across antennas and
users) scales as
Kր
 η 
C̄T DM A ∼ n log2 1 + log (K)
n  
Kր Kր η
C̄BC ∼ C̄BF ∼ nt log2 1 + log (nr K)
nt

where n = min {nt , nr }.

Observations:
– log (nr K): K users with nr receive antennas effectively act as a set of nr K
independent users.
Kր
– C̄BC ∼ nt log2 log K: full spatial multiplexing gain and MU diversity gain are
exploited with DPC.
Kր
– C̄BC ∼ C̄BF : as K gets large, there is a large probability to find a set of orthogonal
user channels to transmit over, each of those channels having a channel gain growing
roughly as log(K).
306 / 494
Global Performance Comparison

• Sum-rate of linear (left) and non-linear (right) MU-MIMO precoders vs SNR in

nt = 4, K = 20 i.i.d. Rayleigh fading channels
60 60
ZFWF DPC
greedy−ZFWF MSR−DE
50 MSR 50 greedy−THP
greedy−MSR greedy−ZFWFVP
MSR−DE
Sum−rate [bps/Hz]

Sum−rate [bps/Hz]
40 40

30 30

20 20

10 10

0 0
−5 0 5 10 15 20 25 30 35 40 −5 0 5 10 15 20 25 30 35 40
SNR [dB] SNR [dB]

Observations: ZFBF without user selection (ZFWF) performs poorly. ZFBF with
user selection (greedy-ZFWF) is a competitive strategy for MU-MIMO broadcast
channels, in terms of both performance and complexity.
Keep in mind the assumptions: perfect CSIT, the same average SNR for all users
and a max-rate scheduler (i.e. there is no fairness issue involved here).
307 / 494
Multi-User MIMO (Downlink) with Imperfect
CSIT

308 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 12
Section: 12.9

309 / 494
Additional References

• M. Maddah-Ali and D. Tse, Completely state transmitter channel state information

is still very useful, IEEE Trans. Int. Theory, vol. 58, no. 7, 2012.
• C. Hao and B. Clerckx, Imperfect and Unmatched CSIT is Still Useful for the
Frequency Correlated MISO Broadcast Channel, IEEE ICC 2013, June 2013.
• C. Hao and B. Clerckx, MISO Broadcast Channel with Imperfect and (Un)matched
CSIT in the Frequency Domain: DoF Region and Transmission Strategies, IEEE
PIMRC 2013, September 2013.
• R. Tandon, S. Jafar, S. Shamai Shitz, and H. Poor, “On the synergistic benefits of
alternating CSIT for the MISO broadcast channel,” IEEE Trans. Inf. Theory., vol.
59, no. 7, pp. 4106-4128, 2013.
• J. Chen and P. Elia, Optimal DoF Region of the two-user MISO BC with general
alternating CSIT, ArXiv 2013.

310 / 494
Introduction

• In practice, perfect CSIT is hard to obtain in both FDD and TDD systems.
– TDD could make use of reciprocity
• Only partial CSIT is available at the transmitter.
• There are two major impairments that prevent from obtaining perfect CSIT:
• inaccurate CSI measurement and feedback (due to channel estimation errors and
limited feedback overhead)
• feedback delay (due to processing delay at the mobile and the BS and the frame
structure).

311 / 494
Quantized Feedback-Based Precoding

• ZFBF with uniform power allocation, single receive antenna per mobile and
quantized feedback ĥq
wq = F̂(:, q)/kF̂(:, q)k
where  −1
F̂ = ĤH ĤĤH D−1

with h iT
Ĥ = ĥTi , . . . , ĥTj
i,j∈K

• Every user q quantizes its channel using a Bq -bits codebook Wq of codevectors vq,i ,
i = 1, . . . , np,q = 2Bq . the best codevector vq⋆ for user q is selected as

vq⋆ = arg max |hq vq,i |2

1≤i≤np,q

The quantized version of the channel direction writes as the nt × 1 row vector
ĥq = (vq⋆ )H .
• With imperfect CSIT, multi-user interference cannot be canceled out perfectly by the
ZFBF filter.

312 / 494
Sum-Rate Analysis

• Expected rate achieved by user q with perfect CSIT and uniform power allocation
(assume ne co-scheduled users)
  
ηq
R̄CSIT,q = EH log2 1 + |hq wZF,q |2
ne
• Expected rate achieved by user q with quantized CSIT and uniform power allocation

R̄LF,q = EH,Wq {log2 (1 + ρq )}

where ηq
n
|hq wq |2
ρq = ηq Pe
1+ ne p∈K,p6=q |hq wp |2
• The rate loss for user q incurred by the quantized feedback

∆R̄q = R̄CSIT,q − R̄LF,q .

• A multi-user MIMO scheme with imperfect CSIT/quantized feedback is affected by
the quantization error at two levels:
– Residual interference term in the SINR ρq that does not vanish with the SNR ηq ,
therefore inducing a ceiling effect as the SNR increases
– Accurate CQI evaluation becomes challenging
313 / 494
Precoding with Partial Transmit Channel Knowledge

• Assume a predefined set of ne > 1 co-scheduled users

Proposition

The limited feedback-based ZFBF using a codebook with Bq bits of feedback

incurs a rate loss for user q (relative to perfect CSIT-based ZFBF) that is
upper bounded as
 
ηq
∆R̄q . log2 1 + (ne − 1) df,q
ne
where df,q is the average distortion function of user q
n o n  2 o
2
df = Eh λmax − khw⋆ k = Eh khk2 − khk2 h̄w⋆  .

314 / 494
Precoding with Partial Transmit Channel Knowledge

• Performance of channel statistics-based codebook (CDIT-CB) and DFT codebook

with Greedy user selection for B = 2, 3, 4, nt = 4, |t| = 0.95 and K = 10.
|t|=0.95, K=10
40
Perfect CSIT
35 CDIT CB, B=4
CDIT CB, B=3
30 CDIT CB, B=2
Sum−rate [bps/Hz]

DFT, B=4
25 DFT, B=3
DFT, B=2
20

15

10

0
0 5 10 15 20 25 30
SNR [dB]

315 / 494
Scalable Feedback

• Number of feedback bits necessary to maintain a rate loss of ∆R̄q ≤ log2 (b) bps/Hz
for user q

Proposition
In order to maintain a rate loss ∆R̄q between limited feedback ZFBF and
perfect CSIT-based ZFBF smaller than log2 (b) bps/Hz for user q, the number
of feedback bits Bq should scale according to
– i.i.d. Rayleigh fading channels
Bq ≈ (nt − 1) log2 (ηq ) − (nt − 1) log2 (b − 1).
– spatially correlated Rayleigh fading channels (with CDIT-based codebook)

Bq ≈ (rq − 1) log2 (ηq ) − (rq − 1) log2 (b − 1)

2
!  
σ2,q (ne − 1) nt
+ (rq − 1) log2 2
+ (rq − 1) log2
σ1,q ne
2 and σ 2 the rank and the two dominant eigenvalues values of
with rq , σ1,q 2,q
user-q transmit correlation matrix.

316 / 494
Scalable Feedback

• Assuming ne = nt , ηq = η and rq = r ∀q
Deployment DL throughput UL overhead
i.i.d. nt log2 (η) nt (nt − 1) log2 (η)
Spatially Correlated nt log2 (η) nt (r − 1) log2 (η)
• Performance of ZFBF with perfect CSIT and ZFBF with channel statistics-based
codebook and scalable feedback without user selection (nt = 4, ne = 4)
Λ =diag{2,1,1,0}
t
20
Sum−rate [bps/Hz]

Perfect CSIT B=15

scalable feedback B=11
10
B=8
1.6 dB
B=5
B=1
0
0 5 10 15 20
Λt=diag{8/3,4/3,0,0}
30
Sum−rate [bps/Hz]

B=11
Perfect CSIT B=9
20 scalable feedback
B=7
B=6
10
B=4
B=1 B=2
0
0 5 10 15 20 25 30
SNR [dB]

317 / 494
Outdated Feedback-Based Precoding

• Is outdated CSIT useless?

– No! In a two-user MISO BC with outdated CSIT, a sum DoF of 4/3 can be achieved, a
33% DoF enhancement compared to conventional TDMA approach.
• Assume a two-user two transmit antenna MISO BC with delayed CSIT.
– The transmission occurs over three coherence times.
– Each coherence time is made of T time slots over which the channel is constant.
– Channel vector of user 1 on coherence time k as hk = hk,1 hk,2 ,
 
gk = gk,1 gk,2 for user 2.
– Channel coefficients are assumed constant within a coherence time and change
independently from one coherence time to the next one.
– The CSI is assumed to be available at Tx only at the next coherence time.
• Denoting the transmit signal on time slot t of coherence time k as xk,t , the received
signals at user 1 and 2, respectively denoted as yk,t and zk,t , write as

yk,t = hk xk,t + nk,t ,

zk,t = gk xk,t + wk,t ,

where nk,t ∼ CN (0,

 1) and w k,t ∼ CN (0, 1) are AWGN. We consider a long-term
power constraint E xHk,t xk,t ≤ ρlt .

318 / 494
Outdated Feedback-Based Precoding

• Consider two independent 2 × T codewords, C = [c1 , . . . , cT ] and C′ = [c′1 , . . . , c′T ]

respectively intended for user 1 and user 2.
   
• Normalized such that E Tr CCH = E Tr C′ C′H = T.
• MAT strategy transmits codeword C to user 1 and codeword C′ to user 2 over 3T
time slots using the MAT strategy. On time slot t, MAT consists in transmitting
√
1 x1,t = ρct in coherence time 1,
√
2 x2,t = ρc′t in coherence time 2,
√  T
3 the overheard interference x3,t = ρ g1 ct + h2 c′t 0 in coherence time 3.

• A long-term average transmit power (where averaging is also taken over the channel
realizations) of ρlt = 4/3ρ is consumed and twice as much power is spent on
coherence time 3 as in coherence time 1 and 2. We will refer to ρ as the SNR.

319 / 494
Outdated Feedback-Based Precoding

• The equivalent system model for user 1 at time instant t = 1, . . . , T can then be
written as
       
y1,t h1,1 h1,2 0 0 n1,t
√
 y2,t  = ρ  √  c′t +
0 0  ct + ρ  h2,1 h2,2 n2,t  .
y3,t h3,1 g1,1 h3,1 g1,2 h3,1 h2,1 h3,1 h2,2 n3,t
| {z } | {z }
rank two rank one

• Interference elimination
     
y1,t √ h1,1 h1,2 n1,t
ỹt = = ρ ct + .
y3,t − h3,1 y2,t h3,1 g1,1 h3,1 g1,2 n3,t − h3,1 n2,t

This is an equivalent 2 × 2 MIMO channel. User 1 decodes 2 symbols in 3 time slots.

• Similar for user 2. Hence, per user DoF of 2/3 and sum DoF of 4/3.
K K
• Extendable to K-user MISO BC: sum DoF of 1 +...+ 1
1+ 2
≈ ln K
.
K

320 / 494
Imperfect CSIT-Based Precoding

• Design MU-MIMO for imperfect CSIT

• Given imperfect feedback in the frequency domain,
– What is the maximum achievable rate region?
– What are the optimal/suboptimal transmission and reception strategies?
– How to optimally make use of feedback resources?

• Consider a two-user MISO-OFDMA Broadcast Channel, with arbitrary values of the

CSIT qualities across L subbands.
Qualities of the Imperfect CSIT
User User

Rx1 Subband

…
Tx
Subband
Rx2

…
…
Subband

321 / 494
System Model

• Transmit signal vector

 in subband
 j denoted as xj subject ot a per-subband based
power constraint E kxj k2 ∼P (P is the SNR).
• The observations at user 1 and 2, yj and zj respectively, are given by
yj =hH
j xj +ǫj1 ,

zj =gjH xj +ǫj2 ,
where ǫj1 and ǫj2 are unit power AWGN noise.
• hj and gj are the CSI in subband i of user 1 and user 2, respectively. The CSI are
i.i.d. across users and subbands.
• Imperfect CSIT: ĥj of user 1 and ĝj of user 2. Error vectors h̃j = hj − ĥj and
g̃j = gj − ĝj with the covariance matrix E[h̃j h̃H 2 H 2
j ]=σj1 I2 and E[g̃j g̃j ]=σj2 I2 .
2 ∼P −aj and σ 2 ∼P −bj . a and b are respectively interpreted as the quality of the
– σj1 j2 j j
CSIT of user 1 and user 2 in subband j, given as follows
2
logσj1 2
logσj2
aj = lim − , bj = lim − .
P →∞ logP P →∞ logP
– aj and bj vary within the range of [0,1]. aj =1 (resp. bj =1) is equivalent to perfect
CSIT and aj =0 (resp. bj =0) is equivalent to no CSIT.
• DoF per user and per channel use (assuming S channel uses)
Rk
dk , lim , k = 1, 2,
P →∞ S log P
322 / 494
System Model

• Note E[|hH ⊥ 2 H ⊥ 2 H ⊥ 2 H ⊥ ⊥H
j ĥj | ]=E[|(ĥj +h̃j ) ĥj | ]=E[|h̃j ĥj | ]=E[h̃j ĥj ĥj h̃j ]∼P −aj . and
E[|gjH ĝj⊥ |2 ]∼P −bj .
• The average
P CSIT quality ofPLuser 1 and user 2 are respectively expressed as
ae = L1 L j=1 aj and be = L
1
j=1 bj .

Definition
PL Problem: Find transmission strategies that maximize the DoF region in a
scenario such that ae =be .
– L = 1: aj = bj ∀j.
– L = 2: a1 +a2 =b1 +b2
a1 =b1 and a2 =b2 : two P1 on each subband
a1 6=b1 and a2 6=b2 : P2 where the transmitted signal in each subband is correlated to each other

$%&'"(" $%&'")" $%&'"(" $%&'")"

*+,,-./"0" !" #" *+,,-./"0" !" !"
*+,,-./" #" !" *+,,-./" #" #"
(a) Unmatched (b) Matched

Figure: Two-subband based frequency correlated BC.

323 / 494
ZFBF

• ZFBF: designed for perfect CSIT (aj = bj = 1)

• Transmit signal on subband j with uj for user 1 and vj for user 2
⊥ ⊥
xj = ĝj uj + ĥj vj

• Received signals on subband j

⊥ ⊥
yj = hH H
j ĝj uj + hj ĥj vj + ǫj1
⊥ ⊥
zj = gjH ĝj uj + gjH ĥj vj + ǫj2

leading to a sum DoF of 2.

• If imperfect CSIT: aj = β for user 1 and bj = α for user 2
⊥ ⊥
yj = hH H
j ĝj uj + hj ĥj vj +ǫj1
| {z } | {z }
P P 1−β
⊥ ⊥
zj = gjH ĝj uj + gjH ĥj vj +ǫj2
| {z } | {z }
P 1−α P

leading to a sum DoF of β + α (rate of uj + rate of vj )

324 / 494
 3/2
S3
3/2
• S3 : designed for alternating CSIT in a two subband case, i.e. the transmitter has
perfect CSIT of only one user at a time (a1 = 1, b1 = 0, a2 = 0, b2 = 1)
• Transmit signal on subband 1 and 2 with u0 to be decoded by both users but
intended to user 1 or 2 and u2 intended for user 1 and v1 intended for user 2.

x1 = [u0 ,0]T +ĥ⊥

1 v1 ,

x2 = [u0 ,0]T +ĝ2⊥ u2 .

• Received signals at each user

y1 = h∗11 u0 + hH ĥ⊥ v1 +ǫ11 , ∗

z1 = g11 u0 + gH ĥ⊥ v1 +ǫ12 ,
| {z } | 1 {z1 } | {z } | 1 {z1 }
P P0 P P

y2 = h∗21 u0 + hH ⊥
2 ĝ2 u2 +ǫ21 , ∗
z2 = g21 u0 + g2H ĝ2⊥ u2 +ǫ22 ,
| {z } | {z } | {z } | {z }
P P P P0

leading to a sum DoF of 3/2.

325 / 494
 3/2
S3

• If imperfect CSIT: a1 = b2 = β, b1 = a2 = α

y1 = h∗11 u0 + hH ĥ⊥ v1 +ǫ11 , ∗

z1 = g11 u0 + gH ĥ⊥ v1 +ǫ12 ,
| {z } | 1 {z1 } | {z } | 1 {z1 }
P P 1−β P P

y2 = h∗21 u0 + hH ⊥
2 ĝ2 u2 +ǫ21 , ∗
z2 = g21 u0 + g2H ĝ2⊥ u2 +ǫ22 ,
| {z } | {z } | {z } | {z }
P P P P 1−β

leading to a sum DoF of 1/2(β[rate ofu0 ] + 1[rate ofv1 ] + 1[rate ofu2 ]) = 1 + β/2.

326 / 494
P1
3/2
• Note ZFBF performs a space-only precoding while S3 performs a
space-frequency/time precoding
• P1 : Can we do better than ZFBF when aj =bj ∀j?
• Transmit the signal in each subband by superposing a common message with
ZFBF-precoded private messages. Focus on subband 1 for simplicity

x1 =[ c1 ,0]T + ĝ1⊥ u1 + ĥ⊥ v1 ,

|{z} | {z } | 1{z }
P −P a1 P a1 /2 P a1 /2

where c1 is the common message broadcast to both users and u1 and v1 are symbols
intended for user 1 and user 2 respectively.
• Integration of broadcasting (ZFBF) and multicating/FDMA.
• Received signal at each user

y1 = h∗11 c1 + hH ĝ⊥ u1 + hH ĥ⊥ v1 + ǫ11 , ∗

z1 = g11 c1 + gH ĝ⊥ u1 + gH ĥ⊥ v1 + ǫ12 ,
| {z } | 1 {z1 } | 1 {z1 } |{z} | {z } | 1 {z1 } | 1 {z1 } |{z}
P P a1 /2 P0 P0 P P0 P a1 /2 P0

where the private symbols u1 and v1 are drowned by the noise respectively at user 2
and user 1 due to partial ZFBF.

327 / 494
P1

• Decodability:
– Both users decode the common message first with rate (1−a1 )logP by treating the
private message as noise,
– Using SIC, each user can decode their private message with rate a1 logP only subject
to noise, after removing the common message,

• Assume aj =bj = β, P∞ leads to a sum DoF of 1 − β + β + β = 1 + β > 2β

achieved by ZFBF only.
– The DoF pairs (1,β) and (β,1) are achieved if we consider the common message is
intended for user 1 and user 2 respectively.

• Assume a1 =b1 = β and a2 =b2 = α, sum DoF of

1/2(1 − β + β + β + 1 − α + α + α) = 1 + β+α
2
> β + α achieved by ZFBF only.

328 / 494
P2

• P2 problem with L = 2 such that ae = b2 with a1 ≥ b1 and a2 ≤ b2 : a1 = b2 = β

and a2 = b1 = α and β ≥ α
• The transmission blocks in subband 1 and 2 are expressed as

x1 = [c1 , 0]T +ĝ1⊥ u1 +[u0 ,0]T +ĥ⊥

1 v1 ,

x2 = [c2 , 0]T +ĥ⊥ T ⊥

2 v2 +[u0 ,0] +ĝ2 u2 .

where
– Common messages u0 , c1 and c2 to be decoded by both users (intended for user 1 and
user 2 respectively or exclusively for user 1 or user 2 or for both users).
– Note that we do not precode common messages as it does not impact the DoF.
– u1 and u2 are symbols intended for user 1, while v1 and v2 are symbols intended for
user 2.
3/2
• Integrate ZFBF, S3 and FDMA/multicasting.
• Power and rate allocation
subband 1 Power Rate subband 2 Power Rate
c1 P −P β 1−β c2 P −P β 1−β
u1 P α /2 α u2 P β /2 β
u0 (P β − P α )/2 β−α u0 (P − P α )/2
β
β−α
v1 P β /2 β v2 P α /2 α
329 / 494
P2

• Received signals at each user

y1 = h∗11 c1 + hH ĝ⊥ u1 + h∗ u0 + hH ĥ⊥ v1 +ǫ11 ,
| {z } | 1 {z1 } | 11 {z } | 1 {z1 }
P Pα Pβ P0
∗
z1 = g11 c1 + g1H ĝ1⊥ u1 ∗
+ g11 u0 + g1H ĥ⊥
1 v1 +ǫ12 ,
| {z } | {z } | {z } | {z }
P P0 Pβ Pβ

y2 = h∗21 c2 + hH ĝ⊥ u2 + h∗ u0 + hH ĥ⊥ v2 +ǫ21 ,

| {z } | 2 {z2 } | 21 {z } | 2 {z2 }
P Pβ Pβ P0
∗
z2 = g21 c2 + g2H ĝ2⊥ u2 ∗
+ g21 u0 + g2H ĥ⊥
2 v2 +ǫ22 .
| {z } | {z } | {z } | {z }
P P0 Pβ Pα

• Decodability:
– c1 and c2 are respectively decoded first by treating all the other terms as noise.
– Afterwards, user 1 decodes u0 and u1 from y1 using SIC. With the knowledge of u0 ,
u2 can be recovered from y2 .
– Similarly, user 2 decodes u0 and v2 from z2 via SIC. v1 can be decoded from z1 by
eliminating u0 .
α+β
• Sum DoF of 1 + 2
– The DoF pair (1, α+β
2
and ( α+β
) 2
,1) are achieved if we consider the common messages
are intended for user 1 and user 2 respectively.
• When β=α, P2 degrades to 2 parallel P1 and no common message u0 is sent. 330 / 494
P2

• Received signal and decoding procedure of the optimal scheme for the P2
Received Signal
User 1 User 2

Subband 1

Subband 2

• The key point: the transmitter broadcasts u0 twice, i.e. subband 1 and 2.
– User 1 (resp. user 2) observes u0 with higher power than u1 (resp. v2 ) in subband 1
(resp. 2) and receives u0 with the same power level as u2 (resp. v1 ) in subband 2
(resp. 1).
– The common message u0 can be decoded by both users but in different subbands.
– Can be generalized to solve PL ,L≥3 problem by generating multiple streams of u0 and
sending each of them twice.
331 / 494
Weighted-Sum DoF Interpretation of P2 region

• Decompose the two subbands P2 with a1 = b2 = β and a2 = b1 = α

$%&'"(" $%&'")"
*+,,-./"0" !" #"
*+,,-./" #" !"
into subchannels
User 1 User 2 User 1 User 2 User 1 User 2

α
A 1 1
β-α
 1 0
1-β
A% 0 0
B 1 1 B̂ 0 1 A% 0 0

where
– Ã, B̃: no CSIT, each with channel use 1−β;
– Â (B̂): perfect CSIT of user 1 (2), with channel use β−α;
– Ā, B̄: perfect CSIT of both users, with channel use α.

332 / 494
Weighted-Sum DoF Interpretation of P2 region

• DoF region of P2 can be interpreted as a weighted-sum representation of the DoF

region of each subchannel

Du = (1 − β)D̃ + (β − α)D̂ + αD̄

– Subchannel à and B̃ are the BC with no

CSIT Unmatched CSIT, β =0.8, α=0.5
1.2
( β +α d 1 +d 2 =1+ β +α
2 =1.65
à B̃ 2 ,1)
D =D = D̃ : d1 + d2 ≤ 1.
1
– Subchannel Ā and B̄ are the BC with ( 1−β ) D̃ ( 1−β ) D̃
0.8
perfect CSIT of both users ( 1−β ) D̃

d2
D Ā = D B̄ = D̄ : d1 ≤ 1, d2 ≤ 1. 0.6 (β −α) D̂ (β −α) D̂

0.4 (1, β +α
2 )
– Subchannel  and B̂ have an alternating
CSIT setting with two states: I1 I2 =P N 0.2 α D̄
3/2
and I1 I2 =N P (as in S3 ) (β −α) D̂ ( 1−β ) D̃
0
 B̂
(D +D )/2 = D̂ :d1 + d2 ≤ 1.5, 0 0.2 0.4 0.6 0.8 1
d1
d1 ≤ 1, d2 ≤ 1.

333 / 494
Weighted-Sum DoF Interpretation of P1 region

• Decompose the subband into subchannels

– Ã, B̃: no CSIT, each with channel use 1−β and 1−α;
– Ā, B̄: perfect CSIT of both users, with channel use β and α respectively.
• DoF region of P1 can be interpreted as a weighted-sum representation of the DoF
region of each subchannel
β+α β+α
Dm = (1 − )D̃ + D̄.
2 2

Matched CSIT, β =0.8,α=0.5

1.2
( β +α
2 ,1)
d 1 +d 2 =1+ β +α
2 =1.65
1

0.8
(1− α+β
2 ) D̃ (1− α+β
2 ) D̃
d2

0.6

0.4 (1, β +α
2 )

α+β
2 D̄
0.2
(1− α+β
2 ) D̃
0
0 0.2 0.4 0.6 0.8 1
d1
334 / 494
Mode switching among sub-optimal strategies
3/2
• P2 integrates FDMA/multicast, ZFBF, S3 . What about a simple switching
strategy?
– FDMA only: sum DoF dFΣ =1
– ZFBF only: sum DoF dZ
Σ =β+α
3/2 β
– S3 only: sum DoF dS
Σ =1+ 2
• The best strategy among the 3 sub-optimal strategies can achieve at least 80% of
the optimal sum DoF performance as
max(dF Z S opt
Σ , dΣ , dΣ ) ≥ 0.8 × dΣ , ∀β, α ∈ [0, 1].

1 1
ZFBF
0.8 0.8 ZFBF
0.6 3 Optimal 0.6
S32
α

α
3
0.4 0.4
S 32
0.2 3 0.2 FDMA 3
FDMA & S 3
3/2 S3
2
& S 32
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
β β

(a) ρ=90% (b) ρ=80%

3/2
Figure: Unmatched case, switching among FDMA, ZFBF and S3 .

335 / 494
Mode switching among sub-optimal strategies

• P1 integrates FDMA/multicast, ZFBF.

• The best strategy among the 2 sub-optimal strategies can achieve at least 66.7% of
the optimal sum DoF performance as

max(dF Z opt
Σ , dΣ ) ≥ 2/3 × dΣ , ∀β, α ∈ [0, 1].

1 1

0.8 ZFBF 0.8

ZFBF
0.6 0.6

α
α

0.4 Optimal 0.4

0.2 0.2 FDMA

FDMA
0 0
0 0.5 1 0 0.5 1
β β

(a) ρ=75% (b) ρ=66.7%

Figure: Matched case, switching among FDMA and ZFBF.

336 / 494
Introduction to Multi-Cell MIMO

337 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 13
Section: 13.1, 13.2, 13.3

338 / 494
Introduction

• Current wireless networks primarily operate using a frequency reuse 1 (or close to 1),
i.e. all cells share the same frequency band
• Interference is not only made of intra-cell (i.e. multi-user interference), but also of
inter-cell (i.e. multi-cell) interference.
• Cell edge performance is primarily affected by the inter-cell interference.

339 / 494
Wideband/long-term SINR

• For user q in cell i, the wideband/long-term SINR is commonly evaluated by ignoring

the effect of fading but only account for path loss and shadowing
Λ−1
q,i Es,i
SIN Rw,q = P .
2
σn,q + j6=i Λ−1 q,j Es,j

• Provides a rough estimate of the network performance. Function of major

propagation mechanisms (path loss, shadowing, antenna radiation patterns,...), base
stations deployment and user distribution.
• CDF of SIN Rw,q in a frequency reuse 1 network (cells share the same frequency
band) with 2D and 3D antenna patterns in urban macro deployment.
100

90

80

70

60
CDF [%]

50

40

30

20

10 3D
2D
0
−10 −5 0 5 10 15 20 25
Downlink wideband SINR [dB] 340 / 494
Classical Inter-Cell Interference Mitigation

• Divide-and-conquer approach:
– fragmenting the network area into small zones independently controlled from each other
– making progressively use of advanced error correction coding, link adaptation, frequency
selective scheduling and lately single-user and multi-user MIMO in each of those zones.

Figure: Frequency Reuse Partitioning. Figure: Static Fractional Frequency Reuse.

341 / 494
Towards Multi-Cell MIMO: Coordination and Cooperation

• Jointly allocate resources across the whole network (and not for each cell
independently) and use the antennas of multiple cells to improve the received signal
quality at the mobile terminal and to reduce the co-channel interferences.

• Two categories:
– Coordination: No data sharing (user data is available at a single transmsitter) - CSI
sharing. Modelled by an Interference Channel and Interfering Broadcast/Multiple
Access Channel
– Cooperation: Data sharing (user data is available at multiple transmsitters) - CSI
sharing. Modelled by a Broadcast Channel (for Downlink) and Multiple Access Channel
(for Uplink)

342 / 494
Towards Multi-Cell MIMO: Coordination and Cooperation

343 / 494
Network Deployments

344 / 494
System Model - Interference Channel

• Interfering broadcast/multiple access channel

– For each transmitter i (one per cell), the intended receivers (i.e. users) are in cell i.
– Each receiver (i.e. user) is only interested in what is being sent by the corresponding
transmitter.
– Transmitters and receivers do not cooperate but only coordinate their transmissions by
sharing CSI information. In the downlink, one transmitter does not have access to the
codewords sent by other transmitters and cannot perform DPC. In the uplink, one
receiver never has access to other received signals and cannot perform SIC.
• General downlink multi-cell multi-user MIMO network with a total number of KT
users distributed in nc cells.
• Ki users in every cell i, nt,i transmit antennas at BS i, nr,q receive antennas at
mobile terminal q.
• The received signal of a given user q in cell i is
−1/2
X −1/2
yq = Λq,i Hq,i c′i + Λq,j Hq,j c′j +nq
j6=i
| {z }
inter-cell interference

where
– yq ∈ nr,q ,

– 2 I
nq is a complex Gaussian noise CN 0, σn,q nr,q ,
−1
– Λq,i refers to the path-loss and shadowing between transmitter i and user q,
– Hq,i ∈ nr,q ×nt,i models the MIMO fading channel between transmitter i and user q.
345 / 494
Linear Precoding

• scheduled user set of cell i, denoted as Ki , as the set of users who are actually
scheduled by BS i at the time instant of interest

• Transmit ne,i streams in each cell i using MU-MIMO linear precoding

1/2
X X 1/2
c′i = Pi ci = Wi Si ci = Pq,i cq,i = Wq,i Sq,i cq,i
q∈Ki q∈Ki

where
– ci is the symbol vector made of ne,i unit-energy independent symbols
– Pi ∈ nt,i ×ne,i is the precoder made of two matrices, namely a power control
diagonal matrix denoted as Si ∈ ne,i ×ne,i and a transmit beamforming matrix
Wi ∈ nt,i ×ne,i .
– Pq,i ∈ nt,i ×nu,q , Wq,i ∈ nt,i ×nu,q , Sq,i ∈ nu,q ×nu,q , and cq,i ∈ nu,q are
user q’s sub-matrices and sub-vector of Pi , Wi, Si , and
ci , respectively.
– The input covariance matrix at cell i is Qi = E c′i c′H
i subject to the transmit power
constraint Tr{Qi } ≤ Es,i .

346 / 494
Linear Precoding

nr,q
• The received signal yq ∈ of user q ∈ Ki
−1/2
X −1/2
yq = Λq,i Hq,i Pq,i cq,i + Λq,i Hq,i Pp,i cp,i
p∈Ki , p6=q
| {z }
intra-cell (multi-user) interference
X X −1/2
+ Λq,j Hq,j Pl,j cl,j +nq .
j6=i l∈Kj
| {z }
inter-cell interference

• Apply a receive combiner to stream l of user q in cell i

−1/2
X −1/2
zq,l = gq,l yq = Λq,i gq,l Hq,i pq,i,l cq,i,l + Λq,i gq,l Hq,i pq,i,m cq,i,m
m6=l
| {z }
inter-stream interference
X −1/2
X X −1/2
+ Λq,i gq,l Hq,i Pp,i cp,i + Λq,j gq,l Hq,j Pl,j cl,j +gq,l nq .
p∈Ki , p6=q j6=i l∈Kj
| {z } | {z }
intra-cell (multi-user) interference inter-cell interference

347 / 494
Achievable Rate

• By treating all interference as noise, the maximum rate achievable by user q in cell i
with linear precoding is
nu,q
X
Rq,i = log2 (1 + ρq,l ) .
l=1
• The quantity ρq,l denotes the SINR experienced by stream l of user-q and writes as
S
ρq,l = .
Il + Ic + Io + kgq,l k2 σn,q
2

where S refers to the received signal power of the intended stream, Il the
inter-stream interference, Ic the intra-cell interference (i.e. interference from
co-scheduled users) and Io the inter-cell interference and they write as
S = Λ−1 2
q,i |gq,l Hq,i pq,i,l | ,
X −1
Il = Λq,i |gq,l Hq,i pq,i,m |2 ,
m6=l
nu,p
X X
Ic = Λ−1 2
q,i |gq,l Hq,i pp,i,m | ,
p∈Ki ,p6=q m=1
X
Io = Λ−1 2
q,j kgq,l Hq,j Pj k .
j6=i
348 / 494
Achievable Rate

Example
Given the precoders in all cells, what is the SINR of stream l of user-q in cell i?
• Noise plus interference: Il + Ic + Io + kgq,l k2 σn,q
2 H
= gq,l Rni gq,l where
X
Rni = Λ−1
q,i Hq,i pq,i,m (Hq,i pq,i,m )
H

m6=l
nu,p
X X
+ Λ−1
q,i Hq,i pp,i,m (Hq,i pp,i,m )
H

p∈Ki ,p6=q m=1

X
+ Λ−1
q,j Hq,j Pj (Hq,j Pj )
H 2
+ σn,q Inr,q
j6=i

is the covariance matrix of the noise plus interference.

−1/2
• MMSE combiner for stream l: gq,l = Λq,i (Hq,i pq,i,l )H Rni −1
• SINR ρq,l experienced by stream l of user-q
2
Λ−1
q,i |gq,l Hq,i pq,i,l |
ρq,l = H
= Λ−1 H
q,i (Hq,i pq,i,l ) Rni
−1
Hq,i pq,i,l .
gq,l Rni gq,l
349 / 494
Extension to OFDMA Networks

• Scheduled user set of cell i on subcarrier k, denoted as Kk,i ⊂ Ki , is the subset of

users ∈ Ki who are actually scheduled on subcarrier k.
• The received signal after receive filtering of a user q ∈ Kk,i scheduled in cell i on
subcarrier k writes as
−1/2 1/2
zk,q = Λq,i Gk,q H(k),q,i Wk,q,i Sk,q,i ck,q,i
X −1/2 1/2
+ Λq,i Gk,q H(k),q,i Wk,p,i Sk,p,i ck,p,i
p∈Kk,i , p6=q
X X −1/2 1/2
+ Λq,j Gk,q H(k),q,j Wk,l,j Sk,l,j ck,l,j + Gk,q nk,q .
j6=i l∈Kk,j

• The power constraint in an OFDMA network writes as

T −1 nu,k,q
X  TX
−1 X X
Tr Sk,j = sk,q,j,m ≤ Es,j , ∀j.
k=0 k=0 q∈Kk,j m=1

350 / 494
System Model - Broadcast and Multiple Access Channel

• If transmitters (resp. receivers) in different cells are allowed to cooperate and can
share any information through an ideal backhaul, the MIMO IC effectively becomes
the MIMO BC (resp. MIMO MAC).
– Giant MIMO BC in the downlink and MIMO MAC in the uplink.
• Focus on DL. By stacking up the transmit signal vectors, the received signal at user
q is
yq = H q c ′ + n
where
h ′ ′
iT h iT h iT
c′ = c1T , . . . , cnTc , y = y1T , . . . , yKT
T
, n = nT1 , . . . , nTKT ,
h i
−1/2 −1/2
Hq = Λq,1 Hq,1 . . . Λq,nc Hq,nc ,
• The transmit signal vector in the DL multi-cell cooperation writes as
KT
X
c′ = c′q
q=1
h ′ ′
iT
where c′q = cq,1
T T
, . . . , cq,n c
.
′
T = 0 ∀j 6= i where cell i is the serving cell of user q (i.e. q ∈ K ).
– In the MIMO IC, cq,j i

• Careful: power constraint per base  Tr QjPn≤c Es,j ∀j and not sum-power
Pncstation
constraint across base stations j=1 Tr Qj ≤ j=1 Es,j ! 351 / 494
Network Architecture: Multi-Cell Measurement, Clustering
and Transmission
• All interfering links do not affect equally user q’s performance.
– Dominant interfering links with small path losses/shadowing contribute to a high
interference while other interfering links are almost invisible to user q.
– Only the CSI of the dominant interfering link should actually be measured and made
available to the transmitters (the CSI of other links may be ignored).

• The MC measurement set of user q ∈ Ki whose serving cell is i is defined as the set
of cells about which channel state/statistical information related to their link to the
MT is reported and is expressed based on long-term channel properties as
( )
 Λ−1 Es,i
 q,i
Mq = j  −1 ≤δ
Λq,j Es,j

for some threshold δ and assuming maximum power transmission.

• MC user : a user whose MC measurement set is strictly larger than one (i.e. includes
at least the MT’s serving cell). The MC users set of cell i is defined as
Pi = {q ∈ Ki | ♯Mq > 1}.

352 / 494
Multi-Cell Measurement, Clustering and Transmission

• The MC-requested user set of cell i is defined as the set of MC users that have cell i
in their MC measurement set, i.e.

Ri = { l| i ∈ Ml , ♯Ml > 1} .

Note that the MC-requested user set can also be viewed as the victim user set of cell
i as it is the set of users who could be impacted by cell i in the absence of multi-cell
cooperation/coordination.
• The MC clustering set of user q ∈ Ki on subcarrier/time instant k is defined as the
set of cells (BS) participating in the multi-cell coordination/cooperation.
• The MC transmission set Tk,q is a subset of the MC clustering set and is defined by
the BS or set of BSs actively transmitting data to MT q.

353 / 494
Distributed and Centralized Architecture

354 / 494
User-Centric and Network-Predefined Clustering

• user-centric clustering: each UE/MT has its owm clustering set. Clustering sets
dynamically selected and may overlap.
• network predefined clustering: cells are statically clustered and MTs are only served
by one cluster. Clusters do not overlap.

355 / 494
Capacity of the Interference Channel

356 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 13
Section: 13.4

357 / 494
SISO Interference Channel

• What is the capacity region of the two-user SISO IC?

• ηq,i = Λ−1 2
q,i Es,i /σn,q
– long-term SNR when user q is served by cell i
– long-term INR (interference to noise ratio) when q is a victim user of cell i

• η̃q,i = ηq,i |hq,i |2 can be thought of as an instantaneous SNR or INR

• two-user SISO IC: transmitter 1 (i.e. cell 1) communicates with user 1 and
transmitter 2 (i.e. cell 2) with user 2
– achievable rate region function of η̃1,1 , η̃2,2 , η̃1,2 , η̃2,1
– symmetric SISO IC characterized by η̃1,1 = η̃2,2 = η̃d and η̃1,2 = η̃2,1 = η̃c
– symmetric rate: Rsym = max(R1 ,R2 )∈CIC min {R1 , R2 } where R1 and R2 are the
rates achievable by user 1 and 2 respectively in the two-user SISO IC and CIC is the
capacity region of the SISO IC
358 / 494
Very Weak Interference Regime

• Conditions: η̃2,1 << η̃1,1 and η̃1,2 << η̃2,2 (or simply by η̃c << η̃d in the symmetric
case)
• The interfering signal is treated as noise and encoding/decoding as in the absence of
interference is sufficient
– divide-and-conquer approach mentioned earlier (e.g., with frequency-reuse in cellular
systems)

359 / 494
Weak Interference Regime

• Conditions: η̃2,1 < η̃1,1 and η̃1,2 < η̃2,2 (or simply by η̃c < η̃d in the symmetric case)
• Capacity unknown in general but the best known achievable region has been
proposed by Han-Kobayashi.
• Capacity outer-bound lies within 1 bit of the capacity inner-bound achieved by the
Han-Kobayashi (HK) scheme
• Main idea behind Han-Kobayashi scheme:
– split each transmitter information into two parts, i.e., a common and a private message.
– A codebook shared between both transmitters is used to construct independently the
common messages at each transmitter.
– The private messages are constructed from independent codebooks.
– Each receiver jointly decodes the common messages (and therefore partially cancel off
part of the interference) by treating the private messages as interference, cancels the
common messages from the received signal and then decodes the intended private
message.

• Assume for simplicity a symmetric interference channel

360 / 494
Weak Interference Regime

• Common and private messages of user i as ci,c and ci,p , i = 1, 2.

• A fraction x of the transmit power is allocated to the common message while the
remaining fraction 1 − x is allocated to the private message.
• View SISO IC as formed by two SISO MACs:
– M AC1 : 3 virtual transmitters respectively sending c1,p , c1,c and c2,c to receiver 1,
with c2,p treated as noise.
– M AC2 : 3 virtual transmitters respectively sending c2,p , c1,c and c2,c to receiver 2,
with c1,p treated as noise.
– Achievable rate region: intersection of the capacity regions of those two SISO MACs
• Assume for simplicity that R1,c = R2,c = Rc and R1,p = R2,p = Rp .
• Private message  
η̃d (1 − x)
Rp = log2 1 +
1 + η̃c (1 − x)
• Common message
   
η̃d x η̃c x
R1,c = Rc ≤ log2 1 + , R2,c = Rc ≤ log2 1 +
1 + ηI 1 + ηI
 
η̃d x + η̃c x
R1,c + R2,c = 2Rc ≤ log2 1 +
1 + ηI
where ηI = η̃d (1 − x) + η̃c (1 − x) is the interference from private messages.
• In the weak interference regime, η̃c < η̃d , so that the first inequality can be discarded.
361 / 494
Weak Interference Regime

• Hence
 
η̃d (1 − x)
Rsym = Rp + Rc = log2 1 +
1 + η̃c (1 − x)
    
η̃c x 1 η̃d x + η̃c x
+ min log2 1 + , log2 1 +
1 + ηI 2 1 + ηI
• x chosen so that the interference level caused by the private message has the same
level as the other user’s noise level.
– interference caused by the private message has little impact on the other user’s
performance.
– does not prevent each user from experiencing a relatively large private message rate as
long as η̃d > η̃c .

– η̃c (1 − x) ≈ 1, i.e. 1 − x ≈ 1/η̃c and x ≈ η̃c − 1 /η̃c
• Assuming η̃d >> 1 and η̃c >> 1 and η̃d > η̃c

1 1
Rsym ≈ min log2 (η̃d ) + [log2 (η̃d ) − log2 (η̃c )] ,
2 2

max {log2 (η̃c ) , log2 (η̃d ) − log2 (η̃c )} .

362 / 494
Mixed Interference Regime

• Conditions: η̃2,1 ≥ η̃1,1 and η̃1,2 < η̃2,2 or η̃2,1 < η̃1,1 and η̃1,2 ≥ η̃2,2 .
• Not meaningful for the symmetric case
• The capacity is also unknown but the best known achievable region relies on the
Han-Kobayashi scheme.

363 / 494
Strong Interference Regime

• Conditions: η̃2,1 ≥ η̃1,1 and η̃1,2 ≥ η̃2,2 (or simply, η̃c ≥ η̃d in the symmetric case).
• The capacity region has been identified:
– The interfering signal can be decoded along with the desired signal, i.e. each user is
able to decode both messages.
– By decoding first the interfering signal, the rate of the desired signal is improved.
Unfortunately, the decodability of the interfering signal puts a constraint on the other
users’ rates, therefore resulting in a tradeoff between the interfering signal rate and the
desired signal rate.
– The two-user SISO IC capacity region is expressed as the intersection of the capacity
regions of the two SISO MAC formed by the two transmitters and each receiver q = 1, 2
Ri ≤ log2 (1 + η̃q,i ) , i = 1, 2
R1 + R2 ≤ log2 (1 + η̃q,1 + η̃q,2 ) .
• Given the strong interference regime η̃2,1 > η̃1,1 and η̃1,2 > η̃2,2 , the intersection
simply writes
Proposition

The capacity region CIC of the Gaussian two-user SISO IC with strong
interference is the set of all achievable rate pair (R1 , R2 ) such that

Ri ≤ log2 (1 + η̃i,i ) , i = 1, 2
R1 + R2 ≤ min {log2 (1 + η̃1,1 + η̃1,2 ) , log2 (1 + η̃2,2 + η̃2,1 )} .
364 / 494
Strong Interference Regime

• Symmetric

Corollary

The capacity region CIC of the symmetric Gaussian two-user SISO IC with
strong interference is the set of all achievable rate pair (R1 , R2 ) such that

Ri ≤ log2 (1 + η̃d ) , i = 1, 2
R1 + R2 ≤ log2 (1 + η̃d + η̃c ) .

• In the strong interference regime, the capacity region of the two-user SISO IC is a
pentagon (as in two-user SISO MAC).
• The symmetric rate (it is actually the symmetric capacity) simply writes as
1
Rsym = log2 (1 + η̃d + η̃c ) ,
2
1
≈ max {log2 (η̃d ) , log2 (η̃c )} ,
2
1
≈ log2 (η̃c ) .
2

365 / 494
Very Strong Interference Regime

• Can the capacity region, under some interference conditions, become a square only
determined by the inequalities Ri ≤ log2 (1 + η̃i,i ) , i = 1, 2?
– i.e. each transmitter can communicate with its receiver at a rate equal to the one
achievable without any interference
• Possible whenever

log2 (1 + η̃1,1 ) + log2 (1 + η̃2,2 )

≤ min {log2 (1 + η̃1,1 + η̃1,2 ) , log2 (1 + η̃2,2 + η̃2,1 )} .

• Very strong interference regime conditions:

– If η̃1,1 + η̃1,2 ≤ η̃2,2 + η̃2,1 , η̃1,2 ≥ η̃2,2 + η̃1,1 η̃2,2
– If η̃2,2 + η̃2,1 ≤ η̃1,1 + η̃1,2 , η̃2,1 ≥ η̃1,1 + η̃1,1 η̃2,2
• The interference is so strong that each user performs SIC by decoding the interfering
message first and subtracting it from the received signal before decoding its own
message.

• Each transmitter can communicate with its receiver at a rate Ri = log2 1 + η̃i,i for
i = 1, 2, as in the absence of any interference.
• The symmetric rate (symmetric capacity) simply writes as

Rsym = log2 (1 + η̃d ) .

366 / 494
Very Strong Interference Regime

• The very strong interference conditions can be viewed from another angle that is
reminiscent of the SIC behavior in SISO BC.
• When user 1 decodes user 2’s signal in the very strong interference regime, it treats
its own signal as noise. Hence, for user 1 to be able to cancel correctly user 2’s
signal, the interfering channel between transmitter 2 and user 1 has to be strong
enough to support R2 , i.e.
2
!  
Λ−1
1,2 |h1,2 | Es,2 η̃1,2
R2 ≤ log2 1 + 2 2 = log 2 1 + .
σn,1 + Λ−1 1,1 |h1,1 | Es,1
1 + η̃1,1


Given that user 2 wants to receive its message at a rate R2 = log2 1 + η̃2,2 , this
puts the constraints
 

η̃1,2
log2 1 + η̃2,2 ≤ log2 1 + ,
1 + η̃1,1
which equivalently writes as η̃1,2 ≥ η̃2,2 + η̃1,1 η̃2,2 . The other condition is obtained
similarly by looking at user 2’s requirement to decode user 1’s message correctly.

367 / 494
Degrees of Freedom - Multiplexing Gain

Definition

The number of generalized degrees of freedom (or multiplexing gain) is defined

as
Rsym (η̃d , η̃c )
gs (α) = lim
.
η̃d ,η̃c →∞:
log2 (η̃c )
=α log2 η̃d
log2 (η̃d )

Proposition

The achievable number of generalized degrees of freedom (i.e. per-user

multiplexing gain) of the two-user Gaussian SISO IC is given by

1
1 − α 0 ≤ α < 2


 1
α

 2
≤ α < 32
gs (α) = 1 − α2 2
3
≤α<1

 α


2 1 ≤ α<2


1 2 ≤ α.
log2 (η̃c )
where log2 (η̃d )
= α.
368 / 494
 Degrees of Freedom - Multiplexing Gain

Figure: Achievable multiplexing gain per user of the two-user Gaussian SISO IC (α = INR/SNR).
369 / 494
More than Two-User SISO Interference Channels

• The extension to more than two users is far from being clear.
– In the very weak interference regime, the optimality of treating the interference as
noise has been established for scenarios with more than two users.
– In the strong interference regime, the extension of the two-user strategy to more than
two users is not straightforward in general.
– In the weak and mixed interference regimes, the situation is even less clear.

• In the nc -user case (or nc -cell case), is the degree of freedom be of the order of
1/nc ?
• Fortunately not! In a nc -user interference channel where the intended and interfering
signals are of comparable strength (i.e. medium interference regime), it is possible
with Interference Alignment to achieve a multiplexing gain per user of 1/2 despite
the presence of nc interfering users!
– As the transmit power of each base station increases, every user will be able to
simultaneously achieve half of the capacity he could achieve in the absence of the
interference from other users.

370 / 494
More than Two-User SISO Interference Channels

• In a general nc -user SISO IC, it is challenging to characterize the achievable

multiplexing gain as a function of the SNR and INR of all links.
• The multiplexing gain is therefore commonly evaluated by taking the transmit
powers to inifinity, leading to infinite SNR and INR, but without constraining the
ratio between SNR and INR.
Definition
The achievable multiplexing gain of user i is defined as
Ri
lim
= gs,i
Es −→∞ log2 ηi

where ηi = Es Λ−1 2
i,i /σn,i . The total achievable multiplexing gain at the network
level is defined as
nc
X Rj
lim
= gs,sum .
Es −→∞
j=1
log 2 ηj

• Meaningful at asymptotically high SNR and INR, not necessarily at finite SNR and
INR!
371 / 494
More than Two-User SISO Interference Channels

• Time/frequency-varying interference networks are not fundamentally

interference-limited
Proposition
In the nc -user time/frequency-varying SISO IC with an infinite number of
symbol extensions, the total multiplexing gain gs,sum (or number of degrees of
freedom) is nc /2.

– time-varying: channel coefficients vary from one channel use to the another
– beamforming over multiple symbol extensions of the time-varying channel.
– every user must be able to partition its observed signal space into two subspaces of
equal size: 1) one for the desired signals, 2) one for the waste basket for all the
interference terms, under the constraint that the vector spaces corresponding to the
interference must exactly align at every user receiver within the waste basket.

372 / 494
MIMO Interference Channels

• Use multiple antennas over static channels rather than time-varying channels
• Two-user MIMO IC
Proposition

The two-user MIMO IC with nt,1 , nt,2 antennas at the two transmitters and
nr,1 , nr,2 antennas at their respective receivers has a maximum multiplexing
gain

gs,sum = min {nt,1 + nt,2 , nr,1 + nr,2 , max {nt,1 , nr,2 } , max {nt,2 , nr,1 }} .

– If nt,1 = nt,2 = nr,1 = nr,2 = n, gs,sum = n.

– The way antennas are distributed at both ends significantly impacts the multiplexing
gain of the MIMO interference channel. A (nt,1 , nt,2 , nr,1 , nr,2 ) = (1, n − 1, n − 1, 1)
MIMO IC with a total of n transmit antennas and n receive antennas would only
achieve a maximum multiplexing gain of 1. Distributed processing at both ends
severely limits the multiplexing gain!
• In a three-user (nc = 3) MIMO IC with n > 1 antennas at each transmitter and each
receiver and static (constant) channels, gs,sum = 3n/2.
• Extension to more general settings but but achievable multiplexing gains commonly
known only for specific antenna configurations.
373 / 494
Capacity of Multiple Access and Broadcast Channels

• Two-user MIMO BC and MAC

Proposition
The two-user MIMO BC with nt transmit antennas and nr,1 , nr,2 receive
antennas has a total multiplexing gain

gs,sum = min {nr,1 + nr,2 , nt } .

The two-user MIMO MAC with nr receive antennas and nt,1 , nt,2 transmit
antennas at the two transmitters has a total multiplexing gain

gs,sum = min {nt,1 + nt,2 , nr } .

– If nt,1 = nt,2 = nr,1 = nr,2 = n, gs,sum = 2n with BC and MAC. Twice as much as
MIMO IC!
– The way antennas are distributed at both ends does not affect the multiplexing gain in
MIMO BC and MIMO MAC, contrary to the MIMO IC!
• Note: In multi-cell cooperation, no sum power constraint anymore but per BS power
constraint

374 / 494
Coordinated Scheduling and Power Control

375 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 13
Section: 13.5, 13.6

376 / 494
Multi-Cell Multi-User Diversity

• Benefit of a large number of users per cell in a multi-cell network?

– Reminiscent of the single-cell multi-user diversity.
• Assume SISO with a single user scheduled on any given spectral resource slot.
• The SINR of user q in cell i in a SISO IC simply writes as
2
Λ−1
q,i |hq,i | si
ρq = P 2
j6=i Λ−1 2
q,j |hq,j | sj + σn,q

with si and sj the transmit powers.

(a) (b)
• Upper- and lower-bounds ρq,lb ≤ ρq ≤ ρq,ub
2 2
Λ−1
q,i |hq,i | Es Λ−1
q,i |hq,i | si
ρq,lb = P 2 , ρq,ub = max = ηq |hq,i |2 .
j6=i Λ−1 2
q,j |hq,j | Es + σn,q
si 2
σn,q
• With a rate maximization policy, the scheduler in cell i picks up the user whose SINR
is the largest.
Xnc
Cn,lb ≤ Cn = max Rq,i ≤ Cn,ub
S,K
i=1
where
nc
X   nc
X  
Cn,lb = log2 1 + max ρq,lb , Cn,ub = log2 1 + max ρq,ub .
q∈Ki q∈Ki
i=1 i=1 377 / 494
Multi-Cell Multi-User Diversity

• In the symmetric configuration, the users within a cell i are assumed to be located at
the same distance from the base station i, i.e. Λq∈Ki ,i = Λ ∀i, such that they
experience the same average SNR η.

Proposition

In a symmetric network configuration with the fading hq,i independent and

identically Rayleigh distributed across users, for a fixed number of cells nc and
an asymptotically large number of users per cell Ki = K, ∀i, the upper and
lower bounds on the average SINR in cell i and on the network capacity scale as
 
Kր Kր
ρ̄ub = E max ρq,ub ∼ η log K, C̄n,ub ∼ nc log log K,
q∈Ki
 
Kր Kր
ρ̄lb = E max ρq,lb ∼ η log K, C̄n,lb ∼ nc log log K.
q∈Ki

• Same scaling law for ub and lb: degradation created by inter-cell interference
becomes negligible when the number of users is large!
• With a rate maximization scheduler, the network is not interference limited as long
as K is large enough.
378 / 494
Multi-Cell Multi-User Diversity

• In the asymmetric configuration, the users are uniformly distributed in each cell such
that the path-loss is determined by the distance between the user and its serving cell.
Λ−1
q,i is a random variable i.i.d. across users and cells.

Proposition

In an asymmetric network configuration with the fading hq,i independent and

identically Rayleigh distributed across users, for a fixed number of cells nc and
an asymptotically large number of users per cell Ki = K, ∀i, the upper and
lower bounds on the SINR in cell i and on the network capacity scale as
 
Kր ǫ Kր ǫ
ρ̄ub = E max ρq,ub ∼ κub K 2 , C̄n,ub ∼ nc log K,
q∈Ki 2
 
Kր ǫ Kր ǫ
ρ̄lb = E max ρq,lb ∼ κlb K 2 , C̄n,lb ∼ nc log K
q∈Ki 2
where κub and κlb are scaling factors.

• Same scaling law for ub and lb but the presence of unequal path-losses among users
enhances the multi-user diversity, therefore resulting in a larger growth rate in the
asymmetric case (log K) compared to the symmetric case (log log K).
379 / 494
Multi-Cell Multi-User Diversity

• Similar behavior was already observed in single-cell MU-MIMO where the intra-cell
interference is shown to have a negligible impact on the capacity when random
beamforming and opportunistic scheduling are performed under the assumption that
the number of users in the cell is large enough.

• Possible to achieve the optimal network capacity under a totally distributed network
architecture that does not require any exchange of CSI and coordination among cells.
Indeed, each cell can perform single-cell scheduling simply relying on the report of a
CQI ρq,lb .

380 / 494
Multi-Cell Resource Allocation

• Maximize a network utility metric rather than a cell utility metric

• Narrowband transmission
nc X
X
{S⋆ , W⋆ , G⋆ , K⋆ } = arg max wq Rq,i
S,W,G,K⊂K
i=1 q∈Ki

– Weights wq account for rate maximization or network-wide proportional fairness

• Multi-carrier (MIMO-OFDMA) transmission

nc T −1
1 XX X
{S⋆ , W⋆ , G⋆ , K⋆ } = arg max wq R(k),q,i
S,W,G,K⊂K T i=1 q∈K
k=0 k,i

381 / 494
Coordinated Power Control

• Assume SISO narrowband transmissions, i.e. nt = nr = 1

• What are the benefits of performing power control (as well as joint power control
and user scheduling) to mitigate inter-cell interference ?
nc
X
{S⋆ , K⋆ } = arg max wq Rq,i .
S,K⊂K
i=1,q∈Ki

• Same scaling laws for lower- and upper-bounds on the network capacity
(a)symmetric configurations for large K and with a rate maximization approach.
– → The use of power control, even though optimal, does not further increase the
network capacity, and transmitting at full power is optimal in the asymptotic case of a
large number of users.

Proposition
Assuming a fading channel hq,i to be independent and identically Rayleigh
distributed across users, for a fixed number of cells nc and an asymptotically
large number of users per cell Ki = K, ∀i, the network capacity with optimal
power control and rate-maximization based scheduling scales as
Kր
C̄n ∼ nc log log K in a symmetric network configuration, and as
Kր
C̄n ∼ nc 2ǫ log K in an asymmetric network configuration.
382 / 494
High and Low SINR Regimes

• Imagine that the users to be scheduled K have been selected, and focus on optimal
power allocation.
nc
X
{S⋆ } = arg max wq Rq,i
S∈S
i=1,q∈Ki

where S = {S | es,i ≤ si ≤ Es,i , i = 1, . . . , nc } is the feasible set of power allocation

strategies.
Proposition
In the high and low SINR regimes, the optimal power control S ⋆ , is binary, i.e.
S⋆ ∈ Sbnc , where Sbnc is the set of 2nc − 1 corner points of S, excluding the
all-es,i point (i = 1, . . . , nc ).

• The original problem can be converted into the following user scheduling and power
control exhaustive search problem
nc
X
{S⋆ , K⋆ } = arg max
n
wq Rq,i .
S⊂Sb c ,K⊂K
i=1,q∈Ki

An exhaustive search is conducted over the sets Sbnc and K to find the optimal S⋆
and K⋆ .
383 / 494
Two-Cell Clusters

• Interestingly, in the specific two-cell network (nc = 2), binary power control is not
only optimal at low SINR but in the whole SINR range for a network relying on a
rate maximization policy

Proposition


In the two-cell case, the network sum-rate maximizing power allocation s⋆1 , s⋆2
is binary and always takes one of the following three power allocation
candidates: (Es,1 , es,2 ), (es,1 , Es,2 ) and (Es,1 , Es,2 ).

• Implications:
– The transmit power can be quantized to two values without loss of capacity.
– This makes the power allocation strategy particularly simple and of very low overhead.
– The decision cannot be taken based on local CSI only as the optimal decision requires
simultaneous CSI from both cells, therefore requiring some form of centralized
scheduler.
– Popular in LTE-A in the name of on-off power control or coordinated silencing. Robust
to CSI measurement and CSI feedback impairments once implemented at the subframe
level in OFDMA networks
• The binary allocation is not optimal anymore for a network-wide proportional
criterion for which the weights w1 and w2 are different, and for a more general
set-up containing more than two cells.
384 / 494
OFDMA Networks

• Multi-cell coordinated OFDMA networks consists in deriving a joint scheduling and

power allocation scheme that decides, on each subcarrier, upon the transmit power
level and the user to be scheduled in each cell.
– one more dimension since: multiple users can be allocated different frequency resources.
– The rate of a given user would typically be improved by increasing its bandwidth
allocation or transmit power:
the former leading to a bandwidth allocation loss for other users in the cell,
the latter leading to an increase of the inter-cell interference (analogous to the narrowband
system).
– Any improvement of the rate of one user affects the rate of the other users in the
network

• Assume SISO
• User assignments for all sub-carriers and all cells: K = {Ki }n c
i=1 where
Ki = {Kk,i }∀k .
• Power allocation: S = {Si }n c
i=1 where Si = {sk,i }∀k

385 / 494
Objective function

• Coordinated power control and scheduling in OFDMA Networks

nc T −1
1 X X
{S⋆ , K⋆ } = arg max wq R(k),q,i
S,K⊂K T i=1
k=0,q∈Kk,i

where
R(k),q,i = log2 (1 + ρk,q )
and  2
Λ−1
q,i h(k),q,i
  sk,i
ρk,q = P 
−1 
2 ,
j6=i Λq,j h(k),q,j sk,j + σ 2

n,k,q

under the constraint

T
X −1
sk,i ≤ Es,i , ∀i.
k=0

• Non-convex problem. The globally optimal solution may not be found but
near-optimal solutions can be obtained using iterative algorithms. The key idea relies
on an iterative optimization of scheduling and power allocation as discussed below.

386 / 494
Optimality Conditions

• For a predefined set of scheduled users, the optimal power allocation problem must
satisfy the Karush-Kuhn-Tucker (KKT) conditions
• Lagrangian of the optimization problem
nc T −1 nc T −1
!
X X X X
L (S, K, ν) = wq R(k),q,i + νi Es,i − sk,i
i=1 k=0,q∈Kk,i i=1 k=0

where ν = {νi }n c
i=1 is the set of Lagrange multipliers associated with the power
constraint in each cell.
• The solution should satisfy
∂L
=0
∂sk,i
and !
T
X −1
νi Es,i − sk,i = 0,
k=0
PT −1
under the constraints νi ≥ 0, sk,i ≥ 0 and k=0 sk,i ≤ Es,i , for i = 1, . . . , nc and
k = 0, . . . , T − 1.
387 / 494
Optimality Conditions
∂L
• ∂sk,i
= 0 leads to
∂R(k),q,i X ∂R(k),q′ ,m
wq + wq′ ,m = νi ,
∂sk,i ∂sk,i
m6=i
′
with q ∈ Kk,i and q ∈ Kk,m . Equivalently we can write
∂R(k),q,i
wq − Πk,i = νi
∂sk,i
P
where we define Πk,i = m6=i Πk,i,m with
∂R(k),q′ ,m ∂R(k),q′ ,m ∂Ik,q′ ,m
Πk,i,m = −wq′ ,m = −wq′ ,m
∂sk,i ∂Ik,q′ ,m ∂sk,i
| {z }
−πk,q ′ ,m

(a)  
= wq′ ,m πk,q′ ,m Λ−1 h(k),q′ ,i 2
q ′ ,i
P 
−1 
2
where Ik,q′ ,m = l6=m Λq′ ,l h(k),q′ ,l sk,l is the total interference received by user

q ′ in cell m and πk,q′ ,m is defined as the non-negative quantity that represents the
marginal increase in rate of user q ′ in cell m per unit decrease in total interference
on subcarrier k.
388 / 494
Optimality Conditions

• Power allocation
P  2
1 wq j6=i Λ−1
q,j h(k),q,j
 2
 sk,j + σn,k,q
= sk,i +   2
log 2 (νi + Πk,i ) Λ−1 h(k),q,i 
q,i
P
for i = 1, . . . , nc and k = 0, . . . , T − 1, where Πk,i = m6=i Πk,i,m and
!
1 1 ρ2k,q′
πk,q′ ,m =   .
log 2 Λ−1
′
h(k),q′ ,m 2 sk,m 1 + ρk,q′
q ,m

389 / 494
Interference Pricing

• View πk,q′ ,m as a price charged to other cells for generating interference to user q ′
•
∂R(k),q,i
wq − Πk,i = νi
∂sk,i
is a necessary and sufficient optimality condition for the problem in which each cell i
specifies a power level sk,i on subcarrier k to maximize the following surplus function

Υ(k),i = wq R(k),q,i − sk,i Πk,i ,

assuming fixed sk,j with j 6= i and πk,q′ ,m with m 6= i.

• Rather than maximizing selfishly its own utility metric (i.e. weighted sum-rate), cell i
maximizes the difference between its utility and its payment owed to the interference
created to the victim users in the neighboring cells:
– The payment is given by the transmit power sk,i times Πk,i . Πk,i writes as a weighted
sum of victim users’ prices, with the weights equal to QoS weights times the channel
gains between cell i and the victim users.
– Representative of the effect of allocating additional transmit power at cell i on the
weighted rate of all victim users in neighboring cells.
– A high value of Πk,i suggests that cell i must pay a high price for assigning power on
subcarrier k.

390 / 494
Iterative Scheduler

• In the downlink, the inter-cell interference is only function of the power levels and is
independent of the user scheduling decisions.
• This suggests that the user scheduling and the power allocation can be carried out
separately.
• An iterative scheduler can be derived so that the best user to schedule are first found
assuming a fixed power allocation, then the best power allocation are computed for
the fixed scheduled users.
• Assuming a fixed power allocation, given the independence of the inter-cell
interference on the scheduled users,

q ⋆ ∈ K⋆k,i = arg max Υ(k),i = arg max wq R(k),q,i , ∀i, k.

Ki Ki

391 / 494
Iterative Scheduler

• Initialization step: We first fix the maximum number of iteration Nmax and fix
n = 0. We initialize S(0) using e.g. a binary power control strategy, and compute
K(0) and Π(0) .
(n)
• Iteration-n: For each cell i = 1, . . . , nc , we update the power allocation Si based
(n−1) (n−1) (n−1)
on K ,Π and Sj for ∀j 6= i (i.e. assuming the transmit powers in the
other cells remain fixed) as follows
 +
2 2
(n) 1 w q σI,k,q + σn,k,q (n−1)
sk,i =   −  2  , q ∈ Kk,i
log 2 ν + Π(n−1)
i
−1 
Λ h(k),q,i
q,i

k,i

where X  2 (n−1)
2
σI,k,q = Λ−1
q,j h(k),q,j sk,j
  .
j6=i
P −1 (n)
The parameters νi are obtained from the power constraint Tk=0 sk,i ≤ Es,i . After
obtaining the power allocation S(n) , the user selection K(n) and finally Π(n) are
obtained. The procedure is repeated till convergence or till the number of iterations
reaches Nmax .

392 / 494
Modified Iterative Water-Filling

• Modified iterative water-filling

 +
2 2
(n) 1 w q σI,k,q + σn,k,q (n−1)
sk,i =   −  2  , q ∈ Kk,i
log 2 νi + Π(n−1) Λ−1
q,i h(k),q,i
 
k,i

• Similar to the point-to-point water-filling algorithm

– The main difference lies in the fact that the power is allocated accounting for the
combined noise and interference and that the water-filling level νi is modified by the
additional pricing term Πk,i .
– As the inter-cell interference increases, the water-filling level decreases, which results in
a lower power allocation.
– The water-filling level is also affected by the proportional fairness weights wq in such a
way that the water-filling level gets higher as the weight increases.
– As a result, BSs allocate more power on subcarriers that serve users with either high
priorities or better channel qualities but transmit power is decreased on subcarriers
where transmission causes excessive interference to victim users in adjacent cells.

393 / 494
Feedback and Message Passing Requirements

• Centralized implementation:
– Each user q in cell i to report nc T channel measurements as the channel from user q
to any base station j over all subcarriers must be known.
– Each cell i forwards the CSI to a centralized controller.
– The central controller performs the modified iterative water-filling and informs each cell
about the scheduled user and the transmit power on each subcarrier.

• Distributed implementation:
– Cells exchange with each other messages and rely on the feedback information from
the users
– Each cell is assumed to be aware of local CSI, i.e. CSI that can be measured by its user
and reported, plus the messages exchanged between cells.
– From cell i perspective:
(n−1) 2
2 2
and Λ−1

the report to cell i from users q ∈ Kk,i of σI,k,q + σn,k,q q,i h(k),q,i
 
the reception by cell i from each cell m 6= i in the MC clustering set of user q of the tax
(n−1)
information Πk,i,m at iteration n − 1
(n−1)
the transfer of the tax information Πk,j,i at iteration n − 1 to cooperating cell j.

394 / 494
Coordinated Beamforming and Interference
Alignment

395 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 13
Section: 13.7

396 / 494
Coordinated Beamforming

• Rate achievable by user q in cell i with linear precoding is

nu,q
X
Rq,i = log2 (1 + ρq,l ) .
l=1

• The quantity ρq,l denotes the SINR experienced by stream l of user-q

S
ρq,l = .
Il + Ic + Io + kgq,l k2 σn,q
2

where S refers to the received signal power of the intended stream, Il the
inter-stream interference, Ic the intra-cell interference (i.e. interference from
co-scheduled users) and Io the inter-cell interference
S = Λ−1 2
q,i |gq,l Hq,i wq,i,l | sq,i,l ,
X −1
Il = Λq,i |gq,l Hq,i wq,i,m |2 sq,i,m ,
m6=l
nu,p
X X
Ic = Λ−1 2
q,i |gq,l Hq,i wp,i,m | sp,i,m ,
p∈Ki ,p6=q m=1
X 2
1/2
Λ−1

Io = q,j gq,l Hq,j Wj Sj .
F
j6=i
397 / 494
Zero-Forcing Beamforming and Block Diagonalization

• Zero-Forcing Beamforming (ZFBF) based coordination is a natural extension of

MU-MIMO precoding based on ZFBF or BD.
• Forcing the interference to zero at either the input or the output of the receiver.
• If we want to zero-force at the input of the receiver, the constraints on the ZFBF
transmit filters, targeting user q ∈ Ki , follow from (??) as
−1/2
Λs,i Hs,i Wq,i = 0, ∀s 6= q, s ∈ Ki ,
−1/2
Λl,i Hl,i Wq,i = 0, ∀l ∈ Kj ∩ Ri .
• Denoting the set of user indices
K̃q,i = {Ki , Kj ∩ Ri }∀j6=i \ q
nr K̃q,i ×nt
whose size is K̃q,i = ♯K̃q,i , we define the interference space H̃q,i ∈ as
h iT
−1/2 T
H̃q,i = . . . Λs,i Hs,i . . . .
s∈K̃q,i

The zero forcing constraint forces Wq,i to lie in the null space of H̃q,i .
• Serving cell channel (between the user and the serving BS) but also the interfering
cells channels (between the user and the interfering BSs in the MC measurement
set) need to be known at the transmitter.
398 / 494
Interference Alignment

• Extends the coordinated zero-forcing beamforming (or block diagonalization) to

jointly design the transmit precoders and the receive combiners in every cell and for
every user.
• IA restricts the interference at every receiver input to a subset of the received signal
space (i.e. the interference is aligned in that subset) and arranges the desired signal
in the complementary subset such that it can be perceived as interference-free at the
receiver output.
• Such alignment is receiver specific in the sense that some signals may appear aligned
in a given space at receivers where they constitute interference while they remain
distinguishable at other receivers where they are desired.
• Focus on maximizing the degrees of freedom in the network.
• An underlying assumption is that the SNR/INR are high enough.
• Assume a predefined set of scheduled users and a single user transmission, where
every cell schedules only a single user at a given time instant.
– The scheduled user index q is chosen as i and cq,i and Wq,i write as ci and Wi

399 / 494
Conditions for Interference Alignment

• Assumptions:
– nc cells fully connected with nc users with nt transmit antennas at each BS (nt,i = nt
∀i) and nr ≤ nt (nr,i = nr ∀i) receive antennas at each MT.
– Each MT receives ne < nr data streams from its serving BS.

• Divide the nr -dimensional observation space at the receiver into a ne -dimensional

signal space and a nr − ne -dimensional interference space and design jointly the
transmit and receive filters such that every interference is aligned into the
nr − ne -dimensional interference space.
• Interference alignment possible if

C (H1,2 W2 ) = C (H1,3 W3 ) = · · · = C (H1,nc Wnc ) ,

C (H2,1 W1 ) = C (H2,3 W3 ) = · · · = C (H2,nc Wnc ) ,
..
.
C (Hnc ,1 W1 ) = C (Hnc ,2 W2 ) = · · · = C (Hnc ,nc −1 Wnc −1 ) ,

where C (A) is the column space of a matrix, i.e., the vector space spanned by the
column vectors of matrix A.

400 / 494
Conditions for Interference Alignment
 
 

 ne < nr , C  a1 a2  = C
• Given b1 b2 is satisfied if ∃ G such that
G a1 a2 = G b 1 b 2 = 0
• Interference is aligned before receive shaping in such a way that after receive
shaping, it is completely canceled out and the received signal yi of the scheduled
user in cell i lies in the d-dimensional signal space.
• The transmit filter Wi and the receive shaping Gl are obtained as solutions of the
set of nc (nc − 1) equations
−1/2
Λl,i Gl Hl,i Wi = 0, l 6= i, ∀i, l = 1, ..., nc .

• Knowledge of global CSI required. Strictly speaking, Hi,i ∀i not needed.

401 / 494
Closed Form Solutions

• Assume
– nc = 3, n = nt = nr with n even for simplicity.
– The precoder Wi is of dimension n × n/2.
– The channel matrices are full rank.
• Show that there exist n/2 non-interfering paths between transmitter i and receiver i
for each i = 1, 2, 3, i.e. a total multiplexing gain of 3n/2.
• The interference can be zero-forced if the dimension of the interference space is
≤ n/2  

– e.g. at receiver 1, r H1,3 W3 H1,2 W2 = n/2. Recall that
r (H1,2 W2 ) = n/2 and r (H1,3 W3 ) = n/2!
• The conditions for IA
C (H1,2 W2 ) = C (H1,3 W3 ) ,
C (H2,1 W1 ) = C (H2,3 W3 ) ,
C (H3,1 W1 ) = C (H3,2 W2 ) ,
• Given that channel matrices are invertible,


C (W2 ) = C H−1
3,2 H3,1 W1 ,


C (W3 ) = C H−1
2,3 H2,1 W1 .

C (W1 ) = C (TW1 ) ,
where T = H−1 −1 −1
3,1 H3,2 H1,2 H1,3 H2,3 H2,1 . 402 / 494
Closed Form Solutions

• Set W1 = eig (T) whereeig (T) refers to n/2 dominant eigenvectors of T. Hence
W1 = t1 . . . tn/2 .
• Stricter conditions

H2,1 W1 = H2,3 W3 ,
H3,1 W1 = H3,2 W2 ,

leding to W2 = H−1 −1
3,2 H3,1 W1 and W3 = H2,3 H2,1 W1 .

• To zero-force interference, the desired signal must be linearly independent of the

interference at the receivers.
 
– e.g. at receiver 1, we need a full rank matrix H1,1 W1 H1,2 W2 .
 
– Multiplying by H−11,1 , columns of
t1 . . . tn/2 At1 . . . Atn/2 should be
linearly independent, with A = H−1 −1
1,1 H1,2 H3,2 H3,1 ,
– Satisfied given that A is a random full rank linear transformation.
– Similar observations hold true for receivers 2 and 3,

• All 3 receivers can decode n/2 streams using zero-forcing.

403 / 494
Closed Form Solutions

• At receiver 1, the eigenvalue decomposition of the interference matrix is given by

   
H1,2 W2 H1,3 W3 = U(1) U(0) ΛVH .

• Given the alignment of H1,2 W2 and H1,3 W3 , U(0) refers to the n/2 singular
 H
vectors corresponding to zero singular values. Hence by selecting G1 = U(0) ,
we obtain
−1/2 1/2
X −1/2 1/2
y1 = Λ1,1 G1 H1,1 W1 S1 x1 + Λ1,j G1 H1,j Wj Sj xj + G1 n1 ,
j=2,3
−1/2 1/2
= Λ1,1 G1 H1,1 W1 S1 x1 + G1 n1 .
−1/2
User 1 perceives an equivalent channel given by Heq,1,1 = Λ1,1 G1 H1,1 W1 and is
not affected by multi-cell interference.
• The interference alignment presented so far creates an interference-free subspace but
does not attempt to maximize the desired signal strength within the desired signal
subspace.
– IA solution is not a function of the direct channels Hi,i
– sub-optimal at low and medium SNR
– Enhancement: precoding in a second stage (IA being the first stage) along the
eigenvectors of Heq,1,1 and applying water-filling based on its singular values.
404 / 494
IA Illustration

• Illustration of interference alignment on a three-user interference channel

405 / 494
Iterative Solution

• Closed form solutions for IA have been found for specific settings only.
• In general, with nc > 3, nt 6= nr and ne,i streams for transmitter i, analytical
solutions to the IA problem are difficult to obtain.
• Iterative solution:
(n−1) (n)
– Assuming the transmit precoders Wi at iteration n − 1, the receiver shaping Gi
are first computed at iteration n.
(n) (n)
– The updated transmit precoders Wi can then be computed based on all Gi .
– The process iterates until convergence.

• The iterative algorithm alternates between the original and reciprocal networks.
– The reciprocal network consists in switching the roles of transmitter and receiver.
– We denote a variable in the reciprocal network with a bar on top.
– In the reciprocal network, the channel matrix writes as H̄j,i = HH i,j .
– IA conditions in the reciprocal network write as Ḡj H̄j,i W̄i = 0, ∀j 6= i.
– W̄i = GH i and Ḡi = Wi .
H

406 / 494
Iterative Solution

• Within each network only the receive filters are updated to minimize the total
inter-cell leakage interference.
– In the original network, the total inter-cell interference leakage at receiver i due to all
interfering transmitters is given by
n o
Io,i = Tr Gi Qi GH i

with X
Qi = Λ−1 H H
i,j Hi,j Wj Sj Wj Hi,j .
j6=i

• Design the receive shaping Gi such that it lies in the space spanned by the ne,i
eigenvectors corresponding to the ne,i smallest eigenvalues of Qi . Writing
Q i = U Q i ΛQ i U H
Qi with the entries of ΛQi ranked by increasing order of
magnitude, Gi = UH Qi (:, 1 : ne,i ).

• At every iteration, the computation of the receive filter Gi is performed in the

original network while the computation of the transmit filter Wi (or equivalently the
receive shaping Ḡi ) is computed in the reciprocal network.

407 / 494
Iterative Solution
(0)
• Initialization step: Start with arbitrary precoding matrices Wi with
(0)
H (0)
Wi Wi = Ine,i .
• Iteration-n: Alternate between the original and the reciprocal networks:
1 In the original network, compute the interference covariance matrix at each receiver i
(n)
X (n−1) (n−1)
H H
Qi = Λ−1
i,j Hi,j Wj S j Wj Hi,j ,
j6=i

and fix the receive shaping in the original network and the transmit beamformer in the
reciprocal network respectively as
(n) (n) (n)
H
Gi = UH(n) (:, 1 : ne,i ) , W̄i = Gi .
Qi

2 In the reciprocal network, compute the interference covariance matrix at each receiver
j, X
(n) (n) (n)
H H
Q̄j = Λ̄−1
j,i H̄j,i W̄i S̄i W̄i H̄j,i ,
i6=j
and fix the receive shaping in the reciprocal network and the transmit beamformer in
the original network respectively as
(n) (n) (n)
H
Ḡj = UH(n) (:, 1 : ne,i ) , Wj = Ḡj .
Q̄j

For simplicity, uniform power allocation (Sj = S̄j = Es,j /ne,j Ine,j ) is often
assumed.
408 / 494
Iterative Solution

• nr,i = nt,i = n ∀i, ne,i = ne ∀i, uniform power allocation

60
nc=1, 2x2, ne=2
55
n =1, 4x4, n =4
50 c e
n =2, 2x2, n =1
45 c e
nc=2, 2x2, ne=2
Sum−rate [bps/Hz]

40
n =2, 4x4, n =2
c e
35
n =3, 2x2, n =1
c e
30 nc=3, 4x4, ne=2
25
20
15
10
5
0
0 5 10 15 20 25 30
SNR [dB]

Figure: Sum-rate of IA vs SNR in various configurations (nc , nr × nt , ne ) in i.i.d. Rayleigh

fading channels. 409 / 494
MIMO Interfering Broadcast/Multiple Access Channels
w1
  H1,1 v1

 H1,2 v 2

w1 ~
y1 H1,3 v3 = H1, 4 v 4


H1,1 v1
~
y2 w2
w2
H1, 2 v 2
H1,3v3 = H1,4 v4




w3 ~
y3

 
~
y4
w4

Figure: The system model of the two-cell interfering MIMO-MAC when nt = 2, nr = 3, and
two cells with two users in each cell.

410 / 494
CSI Feedback and Message Exchange

• IA relies heavily on accurate CSI knowledge!

• Recall the sensitivity of MU-MIMO to inaccurate CSIT.
• MIMO Interfering MAC with quantized feedforward

40
Perfect feedback
Scaled B as in Theorem 3
35
Perfect feedback − 4log (2)
2
Number of feedback bits: 10
Achievable sum rate [bit/sec/Hz]

30
Number of feedback bits: 8
Number of feedback bits: 6
25

20

15

10

0
0 5 10 15 20 25 30
SNR [dB]

411 / 494
Other Beamformers

• Matched Beamforming
– important when it comes to Massive MIMO

• Joint Leakage Suppression

• Maximum Network Sum-Rate Beamforming
• Beamforming with Assigned Target SINR
• Balancing Competition and Coordination
• Opportunistic Beamforming

412 / 494
Coordinated Scheduling, Beamforming and Power
Control

413 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 13
Section: 13.8

414 / 494
Coordinated Scheduling, Beamforming and Power Control

• Coordinated scheduler, beamformer and power control in MIMO-OFDMA

nc T −1
1 XX X
{S⋆ , W⋆ , K⋆ } = arg max wq R(k),q,i
S,W,K⊂K T i=1 q∈K
k=0 k,i

where
R(k),q,i = log2 (1 + ρk,q )
with  2
Λ−1
q,i h(k),q,i wk,q,i
  sk,q,i
ρk,q = Pn P  2 ,
c
j=1 u∈Kk,j Λ−1
q,j h(k),q,j wk,u,j
  sk,u,j + σ 2
n,k,q
(u,j)6=(q,i)

under the constraint

T
X −1 X
sk,q,i ≤ Es,i , ∀i.
k=0 q∈Kk,i

415 / 494
Optimality Conditions

• Assuming a fixed user schedule and transmit beamformers

 
Xnc T
X −1 X nc
X T
X −1 X
L (S, W, K, ν) = wq R(k),q,i + νi Es,i − sk,q,i 
i=1 k=0 q∈Kk,i i=1 k=0 q∈Kk,i

where ν = {νi }n c
i=1 is the set of Lagrange multipliers associated with the power
constraint in each cell.
• The solution should satisfy
∂L
=0
∂sk,q,i
and  
T
X −1 X
νi Es,i − sk,q,i  = 0,
k=0 q∈Kk,i
PT −1 P
under the constraints νi ≥ 0, sk,q,i ≥ 0 and k=0 q∈Kk,i sk,q,i ≤ Es,i , for
i = 1, . . . , nc and k = 0, . . . , T − 1.

416 / 494
Iterative Scheduler

• Power allocation follows the modified iterative water-filling (accounting for

beamforming)
!+
2 2
1 wq σI,k,q + σn,k,q
sk,q,i = − −1  2
ln 2 νi + Πk,q,i Λ h(k),q,i wk,q,i 
q,i

where
nc
X X
2
 2
σI,k,q = Λ−1
q,j h(k),q,j wk,u,j
  sk,u,j .
j=1 u∈Kk,j
(u,j)6=(q,i)

• Interference pricing interpretation: each cell i attempts to maximize on subcarrier k

the following surplus function for every single user q which this cell i aims to schedule

Υ(k),q,i = wq R(k),q,i − sk,q,i Πk,q,i ,

assuming fixed sk,u,j and Πk,u,j ∀ (u, j) 6= (q, i).

– pricing mechanism that accounts for the impact of beamforming and power allocation
on the interference created to co-scheduled users and adjacent cells.
– any variation of the transmit beamformer of a given user in a cell alters the interference
created to other users in the network.
417 / 494
Iterative Scheduler

• Scheduler:
– For a fixed user schedule and transmit power per beam, optimize the beamforming
vectors.
– For fixed beamforming vectors and power allocation, the user scheduling is done per
beam by finding the user that maximizes wq R(k),q,i .
– For a fixed beamformers and user schedule, the power levels are updated.

• Recall that SINR of each user needs to be accurately computed by the BS at every
iteration!
– Inaccurate SINR prediction hampers the appropriate selection of the users, the
transmission ranks and the beamformers at every iteration of scheduler and ultimately
the whole link adaptation and the convergence of the scheduler

418 / 494
A General Framework of Coordination

• Previous iterative scheduler motivates the design of a general framework of

coordination as used in CoMP
– CSI exchange between cells
– Dynamically and iteratively take decisions on the users to schedule, on the appropriate
subcarriers, on their corresponding beamformers and on the power levels to maximize a
network utility metric

• Initialization step: Each cell decides upon which users to schedule in SU or

MU-MIMO mode and the corresponding transmit precoders and power levels
assuming no coordination between cells. In cell i,
n o
(0) (0) (0) (0) (0)
Si , Wi , Gi , Ki = arg max Ui

with
T −1  
(0) 1 X X (0)
Ui = wq R(k),q,i Pk,q,i
T
k=0 q∈K
(0)
k,i

419 / 494
A General Framework of Coordination

• Iteration-n: Each cell revisits its decision regarding the users to be scheduled and
their transmit precoders, based on the decisions taken by other cells in iteration
n − 1.
– The scheduling decisions in a given cell i are not only function of the utility metric of
users scheduled by that cell but also of the utility metric of victim users that have been
tentatively scheduled by other cells in iteration n − 1.
– Cell i allocates resources such that
n o
(n) (n) (n) (n) (n) (n−1)

S i , W i , Gi , K i = arg max Ui Ki , R i .
Ki ∈ Ki
(n−1)

Pk,i Ki , Ri
(n−1)

Ki , Ri : function of its served user set and its victim user set at iteration n − 1!
– Utility metric of cell i at iteration n
T −1  
(n) 1 X X (n) (n−1)
Ui = wq R(k),q,i Pk,q,i , P (n−1)
T k=0 (n)
k,j∈Mq
q∈Kk,i
| {z }
Single-cell weighted sum-rate
 
(n−1)
− Π i Ri .
| {z }
Tax to be paid due to the interference created to vicitm users in adjacent cells

420 / 494
Massive MIMO

421 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 5,12,13
Section: 5.4, 12.2, 12.6, 12.8, 13.7

422 / 494
Introduction

423 / 494
Point-to-Point i.i.d. Channels

Proposition
In i.i.d. Rayleigh fading channels, the ergodic capacity with CDIT is achieved
under an equal power allocation scheme Q = Int /nt , i.e.,
(  )
ρ
C̄CDIT = Īe = E log2 det Inr + Hw HH w ,
nt

or equivalently,
( n  )
X ρ
C̄CDIT = Īe = E log2 1 + λk ,
nt
k=1

where n = r(Hw ) is the rank of Hw and {λ1 , λ2 , . . . , λn } are the non-zero

eigenvalues of Hw HH
w.

424 / 494
Point-to-Point i.i.d. Channels

Proposition
The ergodic capacity of i.i.d. Rayleigh fast fading channels with CDIT is given
by Z ∞
C̄CDIT = Īe = n log2 (1 + ρλ/nt )pλ (λ)dλ,
0

where pλ (λ) is the distribution of a randomly selected (non-ordered) eigenvalue

of Tw (Tw = Hw HH H
w for nt > nr and Tw = Hw Hw for nr > nt ).

425 / 494
Point-to-Point i.i.d. Channels

• Three particular cases of MIMO systems:

1 nt = N and nr = 1 (MISO)
N  
X N
C̄CDIT = eN/ρ log2 (e) Ep ,
p=1
ρ
2 nt = 1 and nr = N (SIMO)
N  
X 1
C̄CDIT = e1/ρ log2 (e) Ep ,
p=1
ρ
3 nt = nr = n = N
N −1 X
k 2l
(  
N/ρ
X X (−1)m 2l
C̄CDIT = e log2 (e)
22k−m l
k=0 l=0 m=0
   m+1  )
2k − 2l 2l X N
× Ep
k−l m ρ
p=1
 
1
≈ e1/ρ log2 (e)E1
ρ
( )
p
log2 (e) p
2
+(N − 1) 2log2 1 + 4ρ + 1 − 4ρ + 1 − 1 − 2 .
4ρ
426 / 494
Point-to-Point i.i.d. Channels

Ergodic capacity/N [bps/Hz]

MISO
4
SIMO
MIMO N x N

0
0 5 10 15 20 25 30
Number of antennas N

• Observations: Asymptotically,
– C̄CDIT scales linearly with N for squared MIMO systems at any SNR,
– For MISO and SIMO systems, C̄CDIT /N decreases toward zero with increasing N
(and the decrease is faster for MISO than for SIMO).
– If the number of antennas is increased only at one side, the asymptotic rate of the
capacity growth is equal to zero. 427 / 494
Large Antenna Array Regime in Point-to-Point i.i.d.
Channels
• For nt = nr = N , we may take the limit for N → ∞ of C̄CDIT /N , and obtain

C̄CDIT p
log2 (e) p
2

lim = 2log2 1 + 4ρ + 1 − 4ρ + 1 − 1 − 2
N →∞ N 4ρ

• Observations:
– limN →∞ C̄CDIT /N is a constant, which only depends on the SNR.
– The capacity scales with N = n at any SNR in the large antenna array regime!

• More general scenario of N > n? Consider

– N = nr → ∞, with n = nt fixed,
– N = nt → ∞, with n = nr fixed,
– N = nt → ∞, n = nr → ∞, in a constant ratio N/n > 1.

428 / 494
Large Antenna Array Regime in Point-to-Point i.i.d.
Channels
• N = nr → ∞, with n = nt fixed:
– W/N = HH H/N converges to In as N → ∞.
– This implies that, for a fixed value of n, the n eigenvalues of W/N approach one, i.e.
the empirical distribution pλ′ (λ′ ) (where λ′ , λ/N ) approaches δ(λ′ − 1).
– Hence,  
C̄CDIT N
lim = log2 1 + ρ .
N =nr →∞ n n
• N = nt → ∞, with n = nr fixed:
– The empirical distribution of the eigenvalues of W/N = HHH /N also converges
almost surely to δ(λ′ − 1),
– Hence
C̄CDIT
lim = log2 (1 + ρ),
N =nt →∞ n
which is equal to the capacity of a SISO AWGN channel!
• N = nt → ∞, n = nr → ∞, in a constant ratio N/n > 1:
– pλ′ (λ′ ) (where λ′ , λ/N ) can be computed
– Hence
   
C̄CDIT n N N
lim = log2 1 + ρ + ρ − ρβ + 1 − log2 (1 − β) − log2 (e) β
n→∞ n N n n
N
with β a function of n
and ρ.

429 / 494
Large Antenna Array Regime in Point-to-Point i.i.d.
Channels
• The capacity in the large antenna regime scales linearly with n at any SNR
– Recall that in the non-asymptotic case, the linear increase in n is observed only at high
SNR!
– The growth rate is only function of the SNR and the ratio N/n.

• The convergence is very fast, so that the large antenna array regime is reached
already for values of n as small as 3.
• CCDIT is Gaussian distributed, with a mean value given by in previous slide, and a
variance decreasing as 1/N in the first two scenarios. In the third scenario, the
variance dependence towards ρ and N/n is more complex.
• Under the large antenna array regime, the channel becomes much more deterministic
and the channel matrices better conditioned (as opposed to random),
– see e.g. the distribution of the eigenvalues of W/N in scenarios 1 and 2.

430 / 494
Large Antenna Array Regime in Multi-User Channels

• Recall
Proposition
For channels H1 , . . . , HK , SNR ηq , number of receive antennas nr , the gain of
DPC over TDMA is upper-bounded by the minimum between the number of
transmit antennas nt and the number of users K
CBC (H)
≤ min {nt , K} .
CT DM A (H)

• In the limit of large nt and with a fixed K and nr , the DPC gain over TDMA
equals K:
Proposition
For any fixed nr , K and effective SNR ηq ∀q, the gain of DPC over TDMA for
a large number of transmit antennas (nt → ∞) in fading channels that are
independent and identically Rayleigh distributed across antennas and
independent across users writes as

C̄BC nt ր
∼ K.
C̄T DM A
431 / 494
Intuition

• Fading channels are independent and identically Rayleigh distributed across antennas
and independent across users.
• Hence, by the law of large numbers, the nr K rows of H become mutually
orthogonal as nt becomes large,
1
lim Hl HH
p = Inr δlp , ∀l, p = 1, . . . , K,
nt →∞ nt

• By transmitting to the best user, TDMA exploits only nr orthogonal dimensions

while DPC can use up to nr K dimensions, i.e. K times as many signaling
dimensions as TDMA.
• For all SNR! Yes, this factor K is translated into a factor K increase in rate in the
limit of a large nt because the received SNR linearly increases with nt and effectively
reaches the high SNR regime for large nt .

432 / 494
Sum-Rate Capacity of Massive MISO BC

• Using the BC-MAC duality, sum-rate capacity of the MISO BC

 XK 
CBC (H, Es ) = max log2 det Int + hH
q sul,q hq ,
{sul,q } q=1
h i
= max log2 det Int + HH Sd H ,
{sul,q }
where SdP= diag {sul,1 , . . . , sul,K } and the maximization is performed over sul,q ≥ 0
∀q with K q=1 sul,q ≤ Es .
• Assuming a large number of transmit  antennas nt , decorrelation
leads to
1/nt HHH ≈ Λd where Λd = diag Λ−1 2 −1 2
1 /σn,1 , . . . , ΛK /σn,K .
• CBC with a large number of transmit antennas approximates as
C̄BC ≈ CBC (H, Es ) ≈ max log2 det [IK + nt Λd Sd ] ,
{sul,q }
XK


= max log2 1 + nt Λ−1 2
q /σq sul,q .
{ ul,q } q=1
s

• Assuming Λ−1 2
q /σq is the same for all users, the optimal power allocation boils down
to the uniform power allocation sul,q = Es /K ∀q.
433 / 494
Sum-Rate of Linear-Precoded Massive MISO BC

• At high and low SNR, linear beamforming (BF) techniques (based on ZFBF) can
achieve the same scaling rate as DPC.

Proposition
At high SNR and low SNR, DPC and BF have the same scaling rate
Es ր
CBC (H) ∼ CBF (H)
Es ց
CBC (H) ∼ CBF (H)

434 / 494
Sum-Rate of Linear-Precoded Massive MISO BC

• In the limit large nt , thanks to the decorrelation, matched beamforming achieves,

similarly to DPC or ZFBF, a factor K increase in rate compared to TDMA.
– matched beamformer ws = h̄H H
s = hs / khs k ∀s = 1, . . . , K
– the SINR ρq of user q
Λ−1 2
q |hq wq | sq nt ր Λ−1
q n t sq
ρq = P −1 2
≈ 2
p∈K Λq |hq wp | sp + 2
σn,q σn,q
p6=q

– achievable sum-rate
K
X

C̄BF ≈ CBF ≈ log2 1 + nt Λ−1 2
q /σn,q sq .
q=1

– Assuming Λ−1 2
q /σq is the same for all users, the uniform power allocation sq = Es /K
∀q maximizes C̄BF .
– Strong similarity with the sum-rate capacity: Matched beamforming achieves the
sum-rate capacity in the very large transmit antenna regime if sq are chosen equal to
sul,q .
– Rate approximations are only valid in the large antenna regime for K/nt → 0. If
K, nt → ∞ with the ratio nt /K = α, CBF with matched beamforming exhibits an
error floor in the limit of large SNR.

435 / 494
Massive SIMO MAC

• Similar observations also hold in the multiple access channels (i.e. in the uplink) with
a large number of receive antennas. Assuming a SIMO MAC (with single antenna
transmitters) with large nr ,
K
X
C̄M AC ≈ CM AC ≈ log2 (1 + nr ηq ) .
q=1

• Under the large receive antenna regime, this sum-rate is achievable with a simple
receive matched filter.

436 / 494
Large Antenna Array Regime in Multi-User Channels

• The transmit (and also receive) beamforming gain approximates as nt for large nt .
As nt increases, the value of khq k2 , being a χ22nt distributed random variable,
concentrates indeed more and more around its mean.
• The SINR and the sum-rates become exclusively a function of Λq and not of the
fading (that is so useful to benefit from MU diversity)!
• For large nt , C̄BF /C̄T DM A = K!
Proposition
For any fixed nr , K and effective SNR ηq ∀q, the gain of DPC/BF over TDMA
for a large number of transmit antennas (nt → ∞) in fading channels that are
independent and identically Rayleigh distributed across antennas and
independent across users writes as

C̄BC/BF nt ր
∼ K.
C̄T DM A


– For large nt , C̄T DM A ≈ log2 1 + nt maxq=1,...,K {ηq } and
PK nt η q

C̄BF ≈ q=1 log2 1 + K (assuming uniform power allocation).
• Transmit/Receive beamforming and MU diversity are somehow not complementary.
A large nt benefits the array gain and multiplexing gain but restricts the MU
diversity gain. 437 / 494
Channel Hardening and Scheduling

• This behavior is called channel hardening and originates from the fact that, with
2
transmit and/or receive beamforming, the variance of rate σR decreases with nt and
nr as nt and nr tend to infinity, respectively.
• In general, any transmission scheme that exploits spatial diversity reduces the
multi-user diversity gain because of the channel hardening effect.
• Scheduling in Massive MIMO becomes much simpler!

438 / 494
Linear Precoding: Matched Beamforming

• As nt increases, with perfect CSIT, the matched beamformer of user q becomes

orthogonal to co-scheduled users (s 6= q) channels. Hence, the multi-user
interference is naturally eliminated.
• Massive MIMO is spectrally and energy efficient:
– Assume a single receive antenna for simplicity, and transmit with a precoder ws = h̄H
s
and a transmit power ss = Es /nt ∀s = 1, . . . , K
– For large nt , the SINR ρq of user q simplifies as
Λ−1 2
q |hq wq | Es /nt nt ր Λ−1 2
q khq k Es /nt Λ−1
q Es
ρq = P ≈ ≈ = ηq
p∈K Λ−1 2 2
q |hq wp | Es /nt + σn,q
2
σn,q 2
σn,q
p6=q

and the sum-rate is equal to

K
X
CBF (H) ≈ C̄BF ≈ C̄BF ≈ log2 (1 + ηq )
q=1

for a total transmit power KEs /nt .

– By matched beamforming with a power Es /nt per user in a large MISO system (i.e.
the transmit power is scaled down proportionally to 1/nt ), each of the K users gets
the same rate as if it were scheduled on a SISO AWGN channel with a transmit power
Es (and received SNR ηq ) without any intra-cell interference and without any fading!
– Assuming ηq = η ∀q, the total achievable sum-rate writes as K times the SISO AWGN
rate.
– The transmit power is scaled down proportionally to 1/nt and the multiplexing gain
increased proportional to K! 439 / 494
Linear Precoding: Zero-Forcing Beamforming

• ZFBF precoding: wq for user q ∈ K writes as

wq = F(:, q)/kF(:, q)k = F̄(:, q)/kF̄(:, q)k

with
 −1
F = HH HHH ,
 −1
= H̄H H̄H̄H D−1 .
| {z }
F̄

• Massive MIMO effect also benefits ZFBF:

– As nt grows, HHH and H̄H̄H become better conditioned, thereby simplifying the
computation of the matrix inverse.
– In the limit where user channels are orthogonal, HHH and H̄H̄H are diagonal and
ZFBF boils down to MBF.

440 / 494
Sum-Rate Evaluations

• Performance of MBF, ZFBF, DPC, IF in nt = 4, 16, 64 and K = 4 i.i.d. Rayleigh

fading channels.
– IF stands for interference free and is the upper bound on the performance obtained
assuming perfect matched beamforming, no intra-cell interference and an uniform
power allocation across the four users, leading to a sum-rate of
PK 2

q=1 log2 1 + ηq /K khq k .

70
4 − MBF
4 − ZFBF
60
4 − IF
16 − MBF
50 16 − ZFBF
Sum−rate [bps/Hz]

16 − IF
64 − MBF
40 64 − ZFBF
64 − IF
30 4 − DPC

20

10

0
0 5 10 15 20 25 30 35 40
SNR [dB]
441 / 494
Sum-Rate Evaluations

• As nt grows, the gap between IF and ZFBF shrinks significantly:

– Severe gap exists in the four transmit antenna case between IF, DPC and ZFBF,
– The gap completely vanishes with 64 transmit antennas with ZFBF performing as well
as an IF system.
– Hence the performance gain of advanced precoding techniques does not justify the
complexity increase.

• MBF on the other hand performs relatively poorly (except at low SNR)
– sum-rate performance fundamentally limited by intra-cell interference and his SINR is
limited at high SNR by the ratio α = nt /K.
– MBF requires a much larger number of antennas to reach the same performance as
ZFBF.

• In summary, for K >> nt and nt >> K, simple linear precoding schemes provide
very competitive alternatives to more complex (non-linear) strategies.

442 / 494
Inter-Cell Interference

• As the number of transmit antennas nt increases, assuming perfect CSIT, the

matched beamformer of user q in cell i becomes orthogonal to co-scheduled users’
channels (s 6= q) in cell i but also to victim users’ channels in adjacent cells. Hence,
the intra-cell and inter-cell interference is naturally eliminated as evidenced by
(assuming for simplicity nt,i = nt ∀i and nr,q = nr ∀q)
1
lim Hl,i HH
p,i = Inr δlp , ∀l, p ∈ Ki ,
nt →∞ nt
1
lim Hl,i HH
p,j = Inr δlp , ∀l ∈ Ki , p ∈ Kj .
nt →∞ nt

• Inter-cell interference is naturally mitigated and the need for multi-cell coordination
or cooperation therefore vanishes as the number of transmit antennas increases!
• Same for Uplink

443 / 494
Practical issues

• Cost of a BS may increase.

• Cost of the network infrastructure (backhaul and coordination) may decrease.
• Transmit power can be decreased proportionally to nt .
– RF electronics behind every antenna is therefore able to operate at a significantly lower
power operating point.

• More antennas packed in a limited volume, increased spatial correlation and antenna
coupling.
– But recall that spatial correlation/LoS can be beneficial to MU-MIMO.

• Accurate CSIT
– FDD?
– Reciprocity in TDD (if perfect calibration is performed), but the presence of pilot
contamination originating from the reuse of the same pilots by users in different cells
degrades uplink channel estimates and therefore limits the performance of Massive
MIMO

444 / 494
Real-World MIMO Wireless Networks

445 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 14

446 / 494
System Requirements

• peak rate
– highest theoretical throughput achievable with SU-MIMO spatial multiplexing but are
typically not achieved in practical deployments.
– e.g. 8x8 Spatial multiplexing with 8 streams transmission

• cell average spectral efficiency

– average spectral efficiency of a cell (with K users).
– much more representative of throughput encountered in practice

• cell edge user spectral efficiency

– spectral efficiency achieved by at least 95% of the users in the network.
– much more representative of throughput encountered in practice

447 / 494
Frame Structure

• Multiplexing/Access
– DL: OFDM
– UL: DFT-Spread OFDM (SC-FDM)
• Frame structure
– OFDMA/SC-FDMA create a
time-frequency grid composed of
time-frequency resources
– A resource block (RB) is formed by 12
consecutive REs in the frequency
domain for a duration of 7 OFDM
symbols in the time domain.
– A subframe is formed of 14
consecutive OFDM/SC-FDM symbols.
– Scheduling and data transmission is
performed at the RB-level with the
minimum scheduling unit consisting of
two RBs within one subframe.
– First 3 symbols used to carry control
information.

448 / 494
Key Downlink Technologies

• Antenna configurations: 2, 4 or 8 transmit antennas and a minimum of 2 receive

antennas
• LTE Rel. 8 (finalized in Dec 2008):
– Up to 4x4 (up to 4 stream transmission)
– Transmit diversity (to protect against fading) using Orthogonal Space-Frequency Block
Coding (O-SFBC) for 2Tx, non-orthogonal SFBC for 4TX
– Open-loop (for high speed) Spatial Multiplexing with rank adaptation based on
predefined precoders
– Closed-loop (for low speed) Spatial Multiplexing based on codebook precoding
– Stone-age MU-MIMO based on common reference signals (CRS)
• LTE Rel. 9 (finalized in Dec 2009):
– Up to 4x4 (up to 4 stream transmission)
– Introduction of demodulation reference signals (DM-RS)
– Enhancement of MU-MIMO to support ZFBF-like precoding
• LTE-A Rel. 10 (finalized mid 2011):
– Up to 8x8 (up to 8 stream transmission)
– New channel measurement reference signals (CSI-RS)
– New feedback mechanisms for 8Tx (dual codebook W1 W2 structure)
– HetNet - eICIC
• LTE-A Rel. 11 (finalized in Dec 2012):
– Coordinated Multi-Point Transmission/Reception (CoMP) for Homogeneous (Macro)
and heterogeneous (pico, DAS) networks
Dynamic cell/point selection combined with dynamic ON/OFF blanking
449 / 494
Key Uplink Technologies

• Antenna configurations: 1, 2 or 4 transmit antennas in the uplink with a minimum of

2 receive antennas
• LTE Rel. 8 (finalized in Dec 2008):
– single antenna transmission and transmit antenna selection
– MU-MIMO
• LTE-A Rel. 10 (finalized mid 2011):
– Spatial Multiplexing with codebook
– Transmit diversity (for control channels)
• LTE-A Rel. 11 (finalized in Dec 2012):
– Coordinated Multi-Point Transmission/Reception (CoMP)

450 / 494
Antenna Deployments

451 / 494
Reference Signals

Dedicated RS (DRS) Common RS (CRS)

For demodulation For demodulation and measurement
Targets a specific terminal Shared among a group of terminals
Terminal specifically precoded Commonly non-precoded
Overhead proportional to the number Overhead proportional to the number
of transmitted streams of transmit antennas
Sent in RBs where data is present Sent in all RBs
Channel estimation less flexible Channel estimation more flexible
452 / 494
Reference Signals

453 / 494
Channel State Information (CSI) feedback

• Three main feedback information:

– Rank Indicator (RI): the preferred number of streams (denoted as layers in LTE) a user
would like to receive
– Precoding Matrix Indicator (PMI): the preferred precoder in the codebook
– Channel Quality Indicator (CQI): the rate achievable with each stream (used to
perform link adaptation)

• Open-Loop relies only on RI and CQI

– High mobility or limited CSI feedback prevent the use of PMI

• Closed-Loop (Spatial Multiplexing and MU-MIMO) rely on RI, CQI and PMI
– If Spatial Multiplexing, the actual precoder is the same as the one selected by the user
(PMI)
– If MU-MIMO based on CRS, the actual precoder is the same as the one selected by the
user (PMI)
– If MU-MIMO based on DM-RS, the actual precoder (e.g. ZFBF) is computed based on
the one selected by the user (PMI).

454 / 494
Network Deployments

455 / 494
Network Deployment Scenarios

0 uncoordinated macrocell deployment: each macrocell is controlled independently of

its neighbors,
1 macro intra-site homogeneous deployment: coordination between the cells (sectors),
e.g. cells 1,2,3, controlled by the same macro base station (where no standardized
backhaul interface is needed),
2 macro inter-site homogeneous deployment: coordination between cells, e.g. cells
1,5,9, belonging to different radio sites from a macro network,
3 macro-pico heterogeneous deployment: macrocell overlaid with low power open
access points with possible coordination between the macrocell and low power
transmission/reception points within its coverage, each point controlling its own cell
(with its own cell identity),
4 distributed antennas: the same deployment as (3), except that the low-power
transmit/receive points constitute distributed antennas of the macrocell, and are
thus all associated with the macrocell identity,
5 macro-femto heterogeneous deployment: macrocell overlaid with low power closed
access points with no standardized interface between the macro and femtocells.

456 / 494
Macro-Pico Heterogeneous Deployment

• Cell association
– Connect a terminal (for both DL and UL links) to the cell with the strongest received
DL power.
– In HetNet, owing to the transmit power difference between the high power and low
power node, it is preferable to connect to the cell with the lowest path-loss in the UL
and connect to the cell with the strongest received DL power in the DL.
– The UL coverage area therefore becomes larger than the DL coverage area, leading to
different cell associations in the DL and UL and particularly complexifying the system.

• Cell loading
– Balance the load among macro- and picocells in highly loaded cells in order to
maximize the resource reuse between cells.

457 / 494
Macro-Pico Heterogeneous Deployment

• To allow flexible load balancing and keep the same cell association for DL and UL,
LTE-A supports cell range expansion (CRE)
– The range of the low-power node is controlled by a cell association bias.
– Expansion so that a UE may be associated with a cell which does not provide the
strongest received signal power.
– Unfortunately, range expansion brings inter-cell interference whose severity increases as
the cell association bias increases.

• Enhanced ICIC: Almost Blank Subframes (ABSF) introduced in LTE-A

– Semi-static form of the time-domain ON-OFF (binary) power control
• CoMP: Dynamic cell/point selection combined with dynamic ON/OFF blanking

458 / 494
Macro-Femto Heterogeneous Deployment

• In heterogeneous femto networks, femtocells are randomly deployed by consumers

without coordination.
• Since they cannot rely on the X2 interface in LTE-A, femtocells cause severe
downlink and uplink interference to adjacent cells.
• Downlink dead-zone problem: When a non-CSG UE is located in the vicinity of a
Femto BS (HeNB), the harsh interference from HeNB will create an outage area.
Alt 2. ICIC via over-the-air broadcasting

Aggressors
Alt 1. ICIC via UE relaying

Home eNB

Home UE
Macro eNB Macro UE Victim UE
Dead-zone

DSL gateway

ICIC via X2 interface

• When the probability of being located in the vicinity of HeNB and the deployment
ratio of HeNB are both low, static ICIC is sufficient.
– silencing resources by e.g. ON-OFF, blank subframes
• When the deployment ratio of HeNB is high and the outage probability becomes
noticeable, dynamic ICIC is useful.
459 / 494
Beyond LTE-A: Massive Multi-Cell and Massive
Multi-Antenna Networks

460 / 494
System-Level Performance Evaluations

461 / 494
Reference Book

• Bruno Clerckx and Claude Oestges, “MIMO Wireless Networks: Channels,

Techniques and Standards for Multi-Antenna, Multi-User and Multi-Cell Systems,”
Academic Press (Elsevier), Oxford, UK, Jan 2013.

– Chapter 15

462 / 494
System Level Assumptions

• System Level Simulations compliant with LTE-A evaluation methodology

• Assumptions:
– DL synchronized LTE-Advanced network based on FDD and 10 MHz bandwidth made
of 50 resource blocks (RB).
– Homogeneous network
– 57 (=19x3) cell sites and 10 users dropped per cell.
– Full-buffer traffic
– Link adaptation: Adaptive coding, modulation and transmission rank combined with
HARQ based on Chase Combining (target BLER 10%)
– Proportional Fair scheduling
• Careful about notations: nt × nr as in LTE terminology!
• Performance metric
– cell average throughput [bits/s/Hz/cell]: sum of the average throughput of each user
in a cell.
– cell edge troughput [bits/s/Hz/user]: 5th percentile of the CDF of the user average
throughput.

463 / 494
 Single-User MIMO
Antenna deployment (with 6RB) Antenna configuration (for ULA)

cell average throughput

cell average throughput

3 4
ULA 2x2 (4,15)
[bits/s/Hz/cell]

2.75 DP 3.5
37% 4x2 (4,15)

[bits/s/Hz/cell]
40% 48%
3 4x4 (4,15)
2.5 43% 50% 65%
58% 2x2 (0.5,8)
2.5
77% 4x2 (0.5,8)
2.25 2 4x4 (0.5,8)
2 1.5
8x2 (0.5,15) 4x2 (0.5,15) 4x2 (4,15) 1 RB 3 RB 6 RB 50 RB

0.15 0.2
cell edge throughput

cell edge throughput

ULA 2x2 (4,15)
[bits/s/Hz/user]

DP

[bits/s/Hz/user]
0.15 4x2 (4,15)
0.1
4x4 (4,15)
0.1
2x2 (0.5,8)
0.05 4x2 (0.5,8)
0.05
4x4 (0.5,8)
0 0
8x2 (0.5,15) 4x2 (0.5,15) 4x2 (4,15) 1 RB 3 RB 6 RB 50 RB

Observations: Observations:
• 4 × 2 ≥ 2 × 2, 4 × 4 >> 4 × 2 (symmetric
• 8 × 2 > 4 × 2 (larger transmit array gain conf. better).
and SM gain).
• gain of large nr higher for less correlated
• DP > ULA for cell average if nt large (SM scenarios and large subband size.
gain & large array gain on each pol).
• ULA>DP for cell edge (large array gain
due to spatial correlation).
• Large AS < small AS (large array gain >
large SM gain). 464 / 494
 Single-User MIMO
Statistics of the transmission rank
°
Channel estimation errors (ULA,3RB)
3 RB, ULA, 4λ, 15 AS
1

cell average throughput

4
probability

2x2
4x2 7%
0.5 3.5
(4,15) (0.5,8)

[bits/s/Hz/cell]
4x4 6%
3
0
1 2 3 4 4% 4%
2.5
3 RB, ULA, 0.5λ, 8°AS 7% 6%
1
probability

2x2 2
2x2 4x2 4x4 2x2 4x2 4x4
0.5 4x2
4x4
0.2

cell edge throughput

0 ideal CSI−RS, ideal DM−RS
1 2 3 4
non−ideal CSI−RS, ideal DM−RS

[bits/s/Hz/user]
6 RB, 0.5λ, 15°AS 0.15 non−ideal CSI−RS, non−ideal DM−RS
1
probability

4x2 ULA
0.5 8x2 ULA 0.1
8x2 DP
0
1 2 3 4 0.05
2x2 4x2 4x4 2x2 4x2 4x4
number of transmitted streams

Observations: Observations:
• loss of 6 to 7% in the cell average and cell
• higher ranks as nt , nr , spacing and AS edge throughputs.
increase.
• higher ranks for DP than ULA.
• rank-1 transmission most encountered.
• symmetric antenna set-up: full rank
transmission negligible.

465 / 494
 Single-User MIMO

Feedback subband size Quantized and unquantized PMI

cell average throughput

cell average throughput

3.5
3.5 quantized PMI
(4,15) (0.5,8)

[bits/s/Hz/cell]
[bits/s/Hz/cell]

3 unquantized PMI
3 11.5%
14% 13%
2.5 16% 11% 6%
31% 6.5%
42% 39% 30%
2 2.5
1.5
2x2 4x2 4x4 2x2 4x2 4x4 2
4x2 ULA (0.5,15) 4x2 DP (0.5,15) 8x2 DP (0.5,15)
0.2
cell edge throughput

0.12

cell edge throughput

4% 7%
[bits/s/Hz/user]

0.15

[bits/s/Hz/user]
1 RB
3 RB 0.1
0.1
6 RB
0.05 50 RB
0.08 13.5%
0
2x2 4x2 4x4 2x2 4x2 4x4 0.06
4x2 ULA (0.5,15) 4x2 DP (0.5,15) 8x2 DP (0.5,15)

Observations: Observations:
• as the subband size increases, the • loss incurred by codebook quantization
performance decreases (due to channel larger in DP compared to ULA deployments
frequency selectivity) and in 8 × 2 compared to 4 × 2.
• less pronounced in spatially correlated
environment (0.5, 8).

466 / 494
 Multi-User MIMO
Antenna deployment Antenna configuration
cell average throughput

cell average throughput

4 6
ULA 4x2 (0.5,15)
[bits/s/Hz/cell]

[bits/s/Hz/cell]
3.5 DP 5 8x2 (0.5,15)
8x4 (0.5,15)
3 4

2.5 3

2 2
8x2 (0.5,15) 4x2 (0.5,15) 4x2 (4,15) DP ULA

0.15 0.2
cell edge throughput

cell edge throughput

ULA
[bits/s/Hz/user]

[bits/s/Hz/user]
DP
0.1 0.15

0.05 0.1

0 0.05
8x2 (0.5,15) 4x2 (0.5,15) 4x2 (4,15) DP ULA

Observations: Observations:
• 8x4 provides significant gain over 8x2
• ULA > DP for cell average for any nt
– 8Tx ZFBF is far from nulling out MU
• ULA > DP at the cell edge interference
• 0.5λ > 4λ – more pronounced in DP

467 / 494
 Multi-User MIMO
MU-MIMO dimensioning without overhead CSI measurement and feedback

cell average throughput

3.5

non−id. CSI/DM−RS

non−id. CSI/DM−RS

non−id. CSI/DM−RS
cell average throughput

4.5
max−2 streams 3 30%

unq. CDI/q. CQI

id. CSI/DM−RS
4
[bits/s/Hz/cell]

max−4 streams 14%

[bits/s/Hz/cell]

unq. CDI/CQI

unq. CDI/CQI

unq. CDI/CQI
2.5

non−id. CSI
3.5

id. DM−RS

q. CDI/CQI
3 2

2.5 14% 1.5

2 1
4x2, q. CDI 4x2, unq. CDI 8x2, q. CDI 8x2, unq. CDI

0.125

cell edge throughput

0.15
cell edge throughput

max−2 streams
[bits/s/Hz/user]

[bits/s/Hz/user]
0.125 max−4 streams
16% 23% 42%
0.1 0.1
9%
0.075

0.05 0.075
4x2, q. CDI 4x2, unq. CDI 8x2, q. CDI 8x2, unq. CDI

Observations: Observations:
• ranking of losses in increasing order of
• 4 × 2: 4 streams > 2 streams with accurate
severity: CSI-RS < q. CQI < DM-RS <<
feedback, 2 streams > 4 streams with
q. CDI (assuming 6RB subband size and
quantized feedback
feedback delay)
• 8 × 2: 4 streams > 2 streams if accurate
and quantized feedback

468 / 494
 Multi-User MIMO
Dynamic switching based on quantized Dynamic switching based on unquantized
feedback feedback
cell average throughput

cell average throughput

4 5
SU − SU report 9% SU − SU report
38%

[bits/s/Hz/cell]
3.5 MU − rank−1 report MU − rank−1 report
[bits/s/Hz/cell]

MU − SU report 19% 6% 4 SU/MU − SU report

3 33%
SU/MU − SU report 5%
36%
17% 3 20%
5%
2.5

2 2
4x2 DP (4) 4x2 DP (0.5) 4x2 ULA 8x2 DP 8x2 ULA 4x2 DP (0.5,15) 4x2 ULA (0.5,15) 8x2 DP (0.5,15)

0.15 0.15
cell edge throughput

cell edge throughput

[bits/s/Hz/user]
0.125 0.125 14%
[bits/s/Hz/user]

0.1 0.1
11%
0.075 8% 0.075

0.05 0.05
4x2 DP (4) 4x2 DP (0.5) 4x2 ULA 8x2 DP 8x2 ULA 4x2 DP (0.5,15) 4x2 ULA (0.5,15) 8x2 DP (0.5,15)

Observations: Observations:
• MU-MIMO and SU/MU-MIMO dynamic • dynamic SU/MU-MIMO ≥ MU-MIMO
switching bring negligible gain over SU in based on rank-1 report
4 × 2 DP (4,15)
• MU-MIMO and SU/MU-MIMO dynamic
switching bring only 5% gain over SU in
4 × 2 DP (0.5,15)
• MU with rank-1 report > MU w/
469 / 494
Multi-User MIMO
Multi-User Diversity Multi-User Diversity
user average throughput cell average throughput

2.8 50

w.r.t. 2 users [%]

SU − SU report
[bits/s/Hz/cell]

2.6 40
MU − rank−1 report

relative gain
2.4 30 SU/MU − SU report
2.2 SU − SU report 20
MU − rank−1 report
2 10
SU/MU − SU report
1.8 0
2 4 6 8 10 12 14 2 4 6 8 10 12 14

1.25 5

relative gain w.r.t.

SU −SU report [%]
[bits/s/Hz/user]

SU − SU report
1 0
MU − rank−1 report
0.75 SU/MU − SU report
−5
0.5
−10 MU − rank−1 report vs. SU − SU report
0.25 SU/MU − SU report vs. SU − SU report
0 −15
2 4 6 8 10 12 14 2 4 6 8 10 12 14
Average number of users per cell Average number of users per cell

Observations:

• SU-MIMO > MU-MIMO for small K (less than 4) while MU-MIMO > SU-MIMO for large
K.
• MU-MIMO relies more heavily on multi-user diversity than SU-MIMO
• SU/MU-MIMO dynamic switching outperforms both SU-MIMO and MU-MIMO for any
number of users.

470 / 494
Multi-Cell MIMO

Which users benefit from intra-site cooperation?

600 600
8 7 8 7

9 9
400 400

11 10 5 4 11 10 5 4
200 12 6 200 12 6
y [m]

y [m]
0 2 1 0 2
2 1

3 3

−20014 13 20 19 −20014 13 20 19

15 21 15 21
−400 −400
17 16 17 16

18 18
−600 −600
−500 −400 −300 −200 −100 0 100 200 300 400 500 −500 −400 −300 −200 −100 0 100 200 300 400 500
x [m] x [m]

600
8 7

9
400

11 10 5 4
200 12 6
y [m]

0 2
2 1
3

−20014 13 20 19

15 21
−400
17 16

18
−600
−500 −400 −300 −200 −100 0 100 200 300 400 500
x [m] 471 / 494
Multi-Cell MIMO

Which users benefit from inter-site cooperation?

600 600
8 7 8 7

400 400 9
9

11 10 5
5 4 11 10 5
5 4
200 12 6 200 12 6
y [m]

y [m]
0
2 1 0
2 1
3 3

−20014 13 20 19 −20014 13 20 19

15 21 15 21
−400 −400
17 16 17 16

18
−600 −600
−500 −400 −300 −200 −100 0 100 200 300 400 500 −500 −400 −300 −200 −100 0 100 200 300 400 500
x [m] x [m]

600
8 7

400 9

11 10 5
5 4
200 12 6
y [m]

0
2 1
3

−20014 13 20 19

15 21
−400
17 16

18
−600
−500 −400 −300 −200 −100 0 100 200 300 400 500
x [m] 472 / 494
Multi-Cell MIMO: Intra-site vs. Inter-site Clustering

• Percentage of users whose CoMP (MC) measurement set size is 1 to 6 for inter-site
and intra-site (10dB triggering threshold).
CoMP measurement set size 1 2 3 4 5 6
Inter-site CoMP 53% 23% 18% 3% 2% 1%
Intra-site CoMP 75% 19% 6% 0% 0% 0%

Observations:

• By constraining the cooperation to intra-site deployments, the percentage of CoMP users is

significantly decreased.
• No benefit to consider CoMP measurement set sizes larger than 3 for such triggering
threshold.

473 / 494
Multi-Cell MIMO: User-centric vs. Network Predefined
Clustering
• Percentage of users whose CoMP (MC) measurement set size is 1 to 6 for
user-centric and network predefined cooperating (or clustering) set (10dB triggering
threshold).
CoMP measurement set size 1 2 3 4 5 6
user-centric cooperating set 53% 23% 18% 3% 2% 1%
inter-site network predef. cooperating set 80% 14% 6% 0% 0% 0%
• Percentage of users whose CoMP (MC) measurement set size is 1 to 6 for two
different network predefined cooperating (or clustering) sets (10dB triggering
threshold).
CoMP measurement set size 1 2 3
intra-site network predefined cooperating set 75% 19% 6%
inter-site network predefined cooperating set 80% 14% 6%

Observations:

• Network predefined clustering reduces the occurence of CoMP (MC) users

• CoMP measurement set size depends on the network predefined cooperating set.
– intra-site network predefined cooperating set provides larger CoMP measurement sets
than an inter-site network predefined cooperating set
474 / 494
Multi-Cell MIMO: Feedback Overhead

• Percentage of users whose CoMP measurement set size is 1,2 or 3 as a function of

the triggering threshold with a network predefined clustering.
Inter-site Intra-site
100 100
size−1 CoMP measurement set size−1 CoMP measurement set
90 size−2 CoMP measurement set 90 size−2 CoMP measurement set
size−3 CoMP measurement set size−3 CoMP measurement set
80 80

70 70
percentage of users

percentage of users
60 60

50 50

40 40

30 30

20 20

10 10

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
triggering threshold [dB] triggering threshold

Observations:
• With a 10dB triggering threshold
Deployment Absolute overhead Overhead increase
intra-site K ∗ B ∗ (0.75 ∗ B + 0.19 ∗ 2 + 0.06 ∗ 3) = 1.31KB 31%
inter-site K ∗ B ∗ (0.53 ∗ B + 0.23 ∗ 2 + 0.24 ∗ 3) = 1.71KB 71%
• User centric clustering requires higher overhead than network predefined clustering.
• With network predefined clustering, inter-site requires less overhead than intra-site. 475 / 494
Coordinated Scheduling and Beamforming

• Assumptions:
– Coordinated SU-MIMO in homogeneous network: one user scheduled at a time in each
cell on a given time/frequency resource
– Network level iterative coordinated scheduling and beamforming based on interference
pricing, Signal-to-Leakage-and-Noise-Ratio (SLNR) filter design and user-centric
clustering

– Unquantized but average (at the subband level) CSIT

476 / 494
Coordinated Scheduling and Beamforming

Iterative CSCB SU-MIMO vs. SU-MIMO Iterative CSCB SU-MIMO vs. SU-MIMO
nt × nr = 4 × 2 ULA (4,15) and (0.5,8) nt × nr = 4 × 4 ULA (4,15)
cell average throughput

cell average throughput

2.8 4.2
2% SU SU
3.6%
[bits/s/Hz/cell]

[bits/s/Hz/cell]
2.6 CSCB SU 16 iterations CSCB SU

3.8
2.9%
2.4 9.5% 2%
3.6 10.5%
2.2 3.4
4.5%
2 3.2
(4,15) 1RB (0.5,8) 1RB (4,15) 3RB (0.5,8) 3RB 1 3 6

0.11
cell edge throughput

cell edge throughput

SU
[bits/s/Hz/user]

0.14

[bits/s/Hz/user]
CSCB SU
0.09 31% 0.13
26.5% 0.12
42% 31.2% 18.3%
0.07 19% 0.11
6.6%
0.1
0.05 0.09
(4,15) 1RB (0.5,8) 1RB (4,15) 3RB (0.5,8) 3RB 1 3 6
subband size [RB]

Observations:

• Gain of about 30% coordination at the cell edge for a CSI overhead increase of 71%
• Big loss as the CSIT accuracy decreases
– ... and this assumed unquantized CSI, user receiver implementation assumed known at
the Tx, perfect CSI measurement, no delay

477 / 494
 Coordinated Scheduling and Beamforming

Statistics of the transmission rank

°
Number of iterations nt × nr = 4 × 2 ULA
1 RB, 4x2 ULA, 0.5λ, 8 AS

cell average throughput

0.8 2.8
SU 4 iterations

[bits/s/Hz/cell]
0.6 CSCB SU 2.6 8 iterations
probability

16 iterations
0.4 2.4

0.2 2.2

0 2
1 2 (4,15) 1RB (0.5,8) 1RB (4,15) 3RB (0.5,8) 3RB
°
3 RB, 4x2 ULA, 0.5λ, 8 AS
0.11

cell edge throughput

SU

[bits/s/Hz/user]
6%
0.6
probability

CSCB SU 0.09
11%
0.4
0.07 13%
0.2

0 0.05
1 2 (4,15) 1RB (0.5,8) 1RB (4,15) 3RB (0.5,8) 3RB
number of transmitted streams

Observations: Observations:

• Coordination allows cell edge users to • Convergence rate depends on deployment

benefit from spatial multiplexing gains

478 / 494
Coordinated Scheduling and Beamforming

Ideal and non-ideal link adaptation

cell average throughput

2.6
SU
2.5
[bits/s/Hz/cell]
CS/CB SU 4 iter.
2.4 4% 3.5% CS/CB SU 4 iter. − id. LA
2.3
2.2 7%
9%
2.1
2
(4,15) 3RB (0.5,8) 3RB

0.1
cell edge throughput
[bits/s/Hz/user]

0.09
0.08
32% 24% 43%
0.07
0.06
0.05
(4,15) 3RB (0.5,8) 3RB

Observations:
• Most of the potential gain lost due to inaccurate LA.
• Inaccurate CQI prediction hampers the appropriate selection of the users, the transmission
ranks and the beamformers at every iteration of scheduler and ultimately the whole link
adaptation and the convergence of the scheduler

479 / 494
Coordinated Scheduling and Beamforming

CS vs. CS/CB

cell average throughput

2.8
CS

[bits/s/Hz/cell]
2.6 CSCB

2.4

2.2

2
(4,15) 1RB (0.5,8) 1RB (4,15) 3RB (0.5,8) 3RB

0.1
cell edge throughput
[bits/s/Hz/user]

0.09

0.08

0.07

0.06
(4,15) 1RB (0.5,8) 1RB (4,15) 3RB (0.5,8) 3RB

Observations:

• CS only brings all the gains

480 / 494
 Robust Joint Scheduling and Rank Coordination

B. Clerckx et al., A Practical Cooperative Multicell MIMO-OFDMA Network based on Rank

Coordination, IEEE Trans. on Wireless Comm. vol. 12, no. 4, pp. 1481-1491, April 2013.

Improve cell edge user experience by Rank coordination: Each cell edge UE
enabling robust multi-streams transmission recommends the interfering cells to use a
to cell edge users transmission rank that is the most beneficial
to its performance

481 / 494
Robust Joint Scheduling and Rank Coordination

Simulation assumptions SU-MIMO with and without coordination

cell average throughput

3.8
SU
dyn. RR SU (A) 1% 1%

[bits/s/Hz/cell]
3.7 dyn. RR SU (B)
4.5%
stat. RR SU
3.6

3.5

0.14

cell edge throughput

SU
dyn. RR SU (A)

[bits/s/Hz/user]
0.13
dyn. RR SU (B) 4%
6.5%
0.12 stat. RR SU

0.11 25.5%

0.1

Observations:

• Better performance gain with a significantly lower feedback overhead and scheduler
complexity
• About 20% gain at the cell edge with only 2-bit additional feedback compared to
uncoordinated SU-MIMO

482 / 494
Heterogeneous Networks

Two-tier Macro - Femto

483 / 494
Heterogeneous Networks: Femto

Two-tier Macro - Femto

Macro MT (×) inside and outside the apartment complex, femto MT (o), femto BS (∗).

484 / 494
Heterogeneous Networks: Femto

Downlink Dead-Zone Problem: Wideband SINR distribution

1
0.5 FL=1, DR=2.5%
FL=1, DR=10% FL=1, DR=5%
FL=1, DR=25% 0.8 FL=1, DR=10%

Cumulative density function

0.4
Cumulative density function

FL=1, DR=50% FL=1, DR=20%

FL=1, DR=75% FL=1, DR=30%
FL=1, DR=100% FL=1, DR=40%
0.3 0.6
FL=1, DR=0% (baseline) FL=1, DR=60%
FL=1, DR=80%
0.2 FL=1, DR=100%
0.4
DR increases
0.1
DR increases
0.2
0

−60 −50 −40 −30 −20 −10 0 10 0

Macro user wideband SINR [dB] −40 −20 0 20 40 60 80
Femto user wideband SINR [dB]

Observations:

• Significant degradation as the deployment ratio (DR) increases

• Interference to Macro MT originates from a small number of dominant Femto BS
• Femto user subject to the increase of the interference level from Femto BS as DR increases

485 / 494
Heterogeneous Networks: Femto

Downlink Dead-Zone Problem: Throughput Analysis

1 1
FL=6, DR=5%
FL=6, DR=10%
0.8 0.8 FL=6, DR=20%
Cumulative density function

Cumulative density function

FL=6, DR=30%

0.6 0.6

DR increases
0.4 0.4

FL=6, DR=0% (baseline)

FL=6, DR=5%
0.2 FL=6, DR=10% 0.2
DR increases
FL=6, DR=20%
FL=6, DR=30%
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0 0.5 1 1.5 2 2.5 3 3.5
Macro user throughput [bits/s/Hz] Femto user throughput [bits/s/Hz]

Observations:

• A non-negligible cell average throughput loss for macro

• The macro outage probability (about 13.5%) matches the ratio between the HeNB cluster
area and the cell sector area.
• The femto cell throughput is affected by the increase of inter-femto cells interference.

486 / 494
Heterogeneous Networks: Femto

Downlink Dead-Zone Problem: Throughput Analysis

cell/area average throughput [bits/s/Hz/(cell or area)]

180
Macro cell Macro cell +
160 Femto cell 72 Femto cells
Area
140 Macro cell +
48 Femto cells
120

100
Macro cell +
80 24 Femto cells

60 Macro cell +
12 Femto cells
40

20
Macro cell only
0
0% 5% 10% 20% 30%
deployment ratio (DR) [%]

Observations:

• The cell area throughput is boosted and benefits from the increase in the cell-splitting gain
of femto cells.
• As the DR (i.e., the number of femtocells) is doubled, the area throughput is however not
doubled due to the inter-cell interference increase.
487 / 494
Heterogeneous Networks: Femto

• PF-TDMA (all frequency resources allocated to a single user on a subframe basis)

was assumed.
• PF-FDMA (resources allocated to users on RB-level and multiple users supported in
different allocations in the same subframe)?
outage cell average throughput

3
[bits/s/Hz/cell]

2.5
2
1.5
0% 10% 20% 30%
15
[bits/s/Hz/user] probability [%]

PF−TDMA
10 PF−FDMA
5
0
0% 10% 20% 30%
cell edge throughput

0.12
0.08
0.04
0
0% 10% 20% 30%
deployment ratio (DR) [%]

Observations:
• PF-FDMA helps the macro user performance (Outage still large but decreased).
• Exploit frequency selective scheduling in interference-limited environments.
488 / 494
Static Binary ON/OFF Power Control in Heterogeneous
Networks
Common Resource Silencing for non reliable backhaul with PF-FDMA

• Femto resource
utilization factor α
• Common silencing
resource is allocated
exclusively for macro
MT protection.
• No femto MT can be
allocated on that
common resource.

489 / 494
Static Binary ON/OFF Power Control in Heterogeneous
Networks
Femtocell throughput Macro and femto throughput
1 3
DR=10%, α=0.5

Femto user average throughput [bits/s/Hz]

DR=10%, α=0.7 α=1
2.5
Cumulative density function

0.8 DR=10%, α=0.9 α=1

α=0.9
DR=10%, α=1.0
DR=30%, α=0.5 2
0.6
DR=30%, α=0.7 α=0.7
DR=30%, α=0.9
1.5 α=0.9
DR=30%, α=1.0
0.4
α decreases α=0.5
α=0.7
1
DR increases
0.2 α=0.5
0.5
α=0 α=0
FL=6, DR=30%
0 FL=6, DR=10%
0 0.5 1 1.5 2 2.5 3 3.5
Femto user throughput [bits/s/Hz] 0
0 0.02 0.04 0.06 0.08 0.1 0.12
macro cell edge (5% user) throughput [bits/s/Hz]

Observations:

• As α decreases, a decrease of the femto MT throughput is observed because of the inability

of femto cells to use some of the resources.
• The frequency-domain common resource silencing boosts the macro cell edge (macro MT
5%) throughput but decreases the femto MT throughput.
• Any enhancement of the victim macroMT throughput goes with a femto MT throughput
degradation.
• the inter-femto cell interference is not negligible and significantly affects the performance:
need to protect victim macro MT and victim femto MT 490 / 494
 Dynamic Binary ON/OFF Power Control in Heterogeneous
Networks
Macro performance - Dynamic PF-FDMA Dynamic vs. static PF-FDMA coordination
outage cell average throughput

2.6 2.6
[bits/s/Hz/cell]

[bits/s/Hz/cell]
macro cell av.
w/ dynamic PC
2.4 6.1% 2.5
14.9% w/ static PC, α=0.9
2.2 2.4 w/ static PC, α=0.5

2 2.3
0% 10% 20% 10% 20%

femto cell av. macro cell edge

10 0.1
[bits/s/Hz/user] probability [%]

[bits/s/Hz/user] [bits/s/Hz/user]
w/o dynamic PC w/ dynamic PC
w/ dynamic PC 0.09 w/ static PC, α=0.9
5
0.08 w/ static PC, α=0.5

0 0.07
0% 10% 20% 10% 20%
cell edge throughput

0.12 3
w dynamic PC
0.08 69% w/ static PC, α=0.9
1281% 2
0.04 w/ static PC, α=0.5

0 1
0% 10% 20% 10% 20%
deployment ratio (DR) [%] deployment ratio (DR) [%]

Observations:

• PF-TDMA/FDMA eliminate outage

• PF-FDMA > PF-TDMA but gain of coordination smaller in PF-FDMA than PF-TDMA
• PF-FDMA increases the overhead of inter-cell message passing
• Dynamic coordination enables a better tradeoff between macro throughput and femto
throughput compared to static coordination

491 / 494
Heterogeneous Networks - DAS

• Distributed Antenna System (DAS)

– Numerous nodes
create more cell
boundaries
– Overlay of macro and
small cells enlarges
the interference zone.
– More UEs become
eligible to benefit
from CoMP in HetNet
• Dynamic point selection with dynamic blanking (ON/OFF power control) in
clustered heterogeneous deployments
N Av. area thrpt cell edge thrpt
Rel. 10 (0dB RE,no ABSF) 4 16.41 0.0574
Rel. 10 (20dB RE,60% ABSF) 4 16.50 (1%) 0.0668 (16%)
DAS with DS 4 15.55 (-5.2%) 0.0698 (21.6%)
DAS with DS/DB 4 16.68 (1.6%) 0.0840 (46.3%)
Rel. 10 (0dB RE,no ABSF) 10 22.33 0.0708
Rel. 10 (20dB RE,60% ABSF) 10 23.76 (6%) 0.0937 (32%)
DAS with DS 10 22.66 (1.5%) 0.0820 (15.8%)
DAS with DS/DB 10 23.27 (4.2%) 0.1067 (50.7%)
• cooperation/coordination gain larger in heterogeneous than homogeneous
492 / 494
Some Conclusions

• Potential gain of single-cell and multi-cell MIMO in theory but benefits may vanish
in practical scenarios
– Sensitivity to CSI measurement: channel estimation errors particularly large for cell
edge users
– CSI feedback inaccuracy: Limited feedback, Subband feedback with strong frequency
selectivity within subband, Particularly problematic in dual-polarized antenna
deployments
– Latency of the feedback and the backhaul
– Feedback and message exchange overhead: Target cell edge users
– Inaccurate link adaptation: due to feedback inaccuracy, BS does not know the receiver
at the mobile terminal, Traffic model, Fast variation of the inter-cell interference
– Scheduler convergence and complexity
– Many other issues left: time/frequency synchronization, antenna calibration, ...
• Sensitivity different depending on SU-MIMO, MU-MIMO, Muti-Cell MIMO

493 / 494
Some Conclusions

• Current wireless system design is at the network level

– Lots of aspects interact with each other
• Network designs become more and more sensitive to impairments

• Gap between theory and practice gets much bigger as we move from single-cell to
(cooperative/coordinated) multi-cell designs
• Account for impairments

494 / 494