Arnaldo Spalvieri, Politecnico di Milano

Introduction to MIMO Systems

Arnaldo Spalvieri
Dipartimento di Elettronica e Informazione
Politecnico di Milano, ITALY
Milano, December 2005 1 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Taxonomy of MIMO Channels
We distinguish between two types of MIMO channels
Single-user channels, which are typical of point-to-point systems
equipped with multiple antennas,
Multi-user channels.
Multi-user channels are further divided between
Multiple-access channels, as, for instance, the uplink of cellular
systems,
Broadcast channels, as, for instance, the downlink of cellular
systems.
Milano, December 2005 2 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Examples of MIMO Systems
Microwave point-to-point radio systems with frequency reuse by
cross-pol transmission. These are single-user, 2 2 systems.
Systems based on synchronous CDMA (downlink of UMTS). In
these systems the number of inputs is the number of active users
in the cell, while the number of outputs is 1.
Wireless LAN of the new generation, where the capacity of the
system is improved by dividing the information between n
antennas at the transmit side and equipping the receiver with
m n antennas. In our taxonomy, this system is to be
considered a single-user system. The real system serves more
users by means other than MIMO.
Milano, December 2005 3 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Time-continuous AWGN channel
Let a
k
denote a vector of data at time k, that is
a
k
= (a
k;1
, a
k;2
, , a
k;n
)
T
. Let H(t) denote a mn matrix of impulse
responses. The element h
i,j
(t) is the impulse response between the
j-th input sequence and the i-th output. The i-th output of the
system is the superimposition of noise plus the eects of all the n
inputs through the impulse responses h
i,j
(t), i = 1, 2, , n:
r
i
(t) =

k
n

j=1
a
k;j
h
i,j
(t kT) +w
i
(t).
The entire MIMO system is described as
r(t) =

k
H(t kT)a
k
+w(t).
Milano, December 2005 4 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
MLD and the matrix matched lter
The joint maximum likelihood sequence detector decides in favor of
the vector-valued sequence that minimizes
J =

||r(t)

k
H(t kT)a
k
||
2
dt,
which can be written as
J =

||

k
H(tkT)a
k
||
2
dt2

a
H
k

H
H
(t kT)r(t)dt

||r(t)||
2
dt.
Everything the receiver needs to know about the received signal to
take the decision is the vector-valued sequence whose k-th element is
y
k
=

H
H
(t kT)r(t)dt,
that is the sampled output of the matrix matched lter H
H
(t).
Milano, December 2005 5 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Time-discrete AWGN channel
Let a(z) denote the z-transform of a vector-valued data sequence,
that is a(z) = (A
1
(z), A
2
(z), , A
n
(z))
T
. Let H(z) denote the
z-transform of a mn matrix-valued sequence. The element H
i,j
(z)
is the transfer function between input A
j
(z) and output X
i
(z). The
i-th output of the system is the superimposition of AWGN plus the
eects of all the n inputs through the transfer functions H
i,j
(z),
i = 1, 2, , n:
X
i
(z) =
n

j=1
H
i,j
(z)A
j
(z) +W
i
(z), i = 1, 2, , m.
The entire AWGN MIMO system is described as
x(z) = H(z)a(z) +w(z).
Milano, December 2005 6 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Time-discrete memoryless AWGN channel
In the memoryless AWGN channel all the impulse responses consist
of only one coecient:
H
i,j
(z) = h
i,j
.
We can therefore forget about time, writing the observation as
x = Ha +w.
This model is often adopted for each subcarrier in MIMO OFDM.
Milano, December 2005 7 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Time-discrete memoryless AWGN channel
We assume that the covariance matrix of a is
E{aa
H
} =
2
a
I
n
,
which means that the total transmitted power over all the n inputs is
n
2
a
. Here and in what follows we denote M
n
a square matrix of
dimension n n. Also, we assume that
E{ww
H
} = N
0
I
m
.
Milano, December 2005 8 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Capacity of the memoryless AWGN channel
Consider the Singular Value Decomposition (SVD) of the mn
matrix H
H = U
m
DV
H
n
,
where U
m
and V
n
are unitary matrices (recall that the property of
the unitary matrix M is MM
H
= M
H
M = I) , and D (dimension:
mn) is non-negative and diagonal. The columns of U
m
are the
eigenvectors of HH
H
, the columns of V
n
are the eigenvectors of
H
H
H, the non-negative entries in the diagonal of D are the square
root of the m eigenvalues of HH
H
, called the singular values of H.
This can be seen as follows:
HH
H
= U
m
DV
H
n
V
n
D
H
U
H
m
= U
m
DD
H
U
H
m
= U
m

m
U
H
m
.
Note that, for m > n, at least mn eigenvalues of HH
H
are zero.
Milano, December 2005 9 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Capacity of the memoryless AWGN channel
Using the SVD one can write
x = U
m
DV
H
n
a +w.
We now impose a rotation of the observation space
U
H
m
x = x

= Da

+w

obtaining an observation model statistically equivalent to x. The

capacity of the i-th sub-channel is
C
i
= log
2

1 +

2
a

i
N
0

, b/2D,
where

i
is the i-th singular value.
Milano, December 2005 10 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Capacity of the memoryless AWGN channel
The capacity of the MIMO memoryless channel is
C =
m

i=1
C
i
= log
2

m
i=1
(1 +

2
a

i
N
0
)

,
which can be written as
C = log
2

det(I
m
+

2
a
N
0

m
)

.
Henceforth we write
C = log
2

det(U
m
(I
m
+

2
a
N
0

m
)U
H
m
)

,
arriving to the celebrated log-det formula
C = log
2

det(I
m
+

2
a
N
0
HH
H
)

, b/2D.
Milano, December 2005 11 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Capacity of the AWGN channel
The capacity of the memoryless channel is extended in a
straightforward way to the general case. Let

m
(z) = H(z)H
H
(z

)
be the spectral matrix that characterizes the system. We have
C =
T
2

/T
/T
log
2

det(I
m
+

2
a
N
0

m
(e
jT
))

d, b/2D.
Milano, December 2005 12 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
MLD for the time-discrete memoryless AWGN channel
The MLD rule is to minimize
||x Ha||
2
= x
H
x 2(a
H
H
H
x) +a
H
H
H
Ha.
Note that all we need to know about the received vector is H
H
x, that
is the output of the time-discrete Matrix Matched Filter (MMF). As
in the SISO channel, the size of the search space becomes huge when
the number of points of the constellations and/or n growth.
Milano, December 2005 13 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
MMF bound
At the MMF output we have
y =
n
a +n,
where
n
= H
H
H, and the correlation matrix of n is N
0

n
.
Suppose that only the j-th user is active. We can take the maximum
likelihood decision about a
j
by slicing
y
j
= a
j

m

i=1
|h
i,j
|
2
+n
j
.
The MMF lower bound on the SNR for the j-th user is therefore
SNR
j
=

2
a

m
i=1
|h
i,j
|
2
N
0
.
Milano, December 2005 14 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Linear detection
The MMF followed by the slicer is the conventional receiver in
CDMA, where the columns of H are the signatures of the users. It
may be strongly suboptimal when the columns of H are not
orthogonal and many users are active. Moreover, the conventional
receiver is particularly susceptible to the near-far problem.
Milano, December 2005 15 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Linear detection
The received 1 m vector x is multiplied by the n m matrix C
y = Cx = CHa +Cw.
The zero forcing matrix is
C = (H
H
H)
1
H
H
, y = a +n,
where the correlation matrix of the noise is
E{nn
H
} = E{Cww
H
C
H
} = N
0
CC
H
= N
0

1
n
.
The MSE for the i-th user is
E{|y
i
a
i
|
2
} = E{|n
i
|
2
} = N
0
(
1
n
)
i,i
.
Milano, December 2005 16 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Linear detection
The MMSE matrix is
C =
2
a
(
2
a
H
H
H+N
0
I
n
)
1
H
H
.
The covariance matrix of the sum of ISI and noise at the slicers is

2
a
N
0
(
2
a
H
H
H+N
0
I
n
)
1
.
Milano, December 2005 17 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Space-time causality
H(z) is space-time causal i H
k
= 0 for k < 0 and H
0
is lower
triangular. In the memoryless case we forget about time, and space
causality means that
y
i
=
i

j=1
h
i,j
x
j
.
In other words, we have H = H
0
and therefore the output y
i
depends
only on the inputs x
1
, x
2
, , x
i
.
Milano, December 2005 18 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Matrix spectral factorization
Let R be a spectral matrix. We concern ourselves with the
memoryless case, hence the entries of R are scalars. The matrix
spectral factorization reads
R = (I B)(I B)
H
,
where B is strictly lower triangular, and is positive denite and
diagonal. We have det() = det(R). The spectral factor can be
worked out by Cholesky factorization. Specically, the Matlab
instruction chol works as follows:
chol(R) =
1/2
(I B)
H
.
Milano, December 2005 19 Arnaldo Spalvieri
Arnaldo Spalvieri, Politecnico di Milano
Decision feedback detection
Let R be the inverse of the covariance matrix of the sum of
interference and noise after some linear interface. For the zero forcing
interface we have
R
zf
= N
1
0
H
H
H,
for the MMSE linear interface we have
R
mmse
= (
2
a
N
0
)
1
(
2
a
H
H
H+N
0
I
n
).
The matrix spectral factorization leads to a form of decision feedback
detection. Assuming correct decisions, the covariance matrix of the
sum of interference and noise at the slicers is
1
. Note that, unlike
in the case of the ISI channel, here the order of detection plays an
important role. Specically, the best thing to do is to detect rst the
signal that has better SDR at the output of the linear interface.
Milano, December 2005 20 Arnaldo Spalvieri