You are on page 1of 66

Supelec

Random Matrix Theory


for
Wireless Communications

Merouane
Debbah
http://www.supelec.fr
merouane.debbah@supelec.fr
February, 2008

Presentation

MIMO Channel Modelling and random matrices

Where do we stand on Channel Modelling

Google search: MIMO Wireless Channel Modelling

Over 15 000 publications on channel modelling


At a rate of 10 papers per day, 1 500 days (nearly 4 years)!
The models are different and many validated by measurements!
Three conflicting schools

Geometry based channel models.


Stochastic channel models based on channel statistics
Do not model, use test measurements
Not even within each school, all experts agree on fundamental issues.

MIMO System Model

Rx

Tx

The channel is linear, noise is additive

y(t) =
Hnr nt ( )x(t )d + n(t)
ntr

Hnr nt (f, t)X(f ) + N(f )


Y(f, t) =
nt
3

Why do we need a channel model?

Our Vision
Step 1: Collection of information
The user (or base station) download information on his environment (dense, number of
buildings,...) through a localization service process
Step 2: Model generation
A statistical channel model is automatically created (at the base station or the mobile unit)
integrating only that information and not more!
Step 3: High speed connection
The coding scheme is adapted to the (statistical not realization) model
Example

Additive Gaussian: Euclidean distance coding


Rayleigh i.i.d: rank and determinant criteria
This scenario could be called User customized channel model coding service and is a
viable scenario from a Soft Defined Radio perspective.
4

Why do we need a statistical channel model?

Ergodic Channel Capacity: (The receiver knows the channel and the transmitter knows
the statistics)

H
C = maxQE(C(Q)) with C(Q) = log2det Int + n H QH
t

Q = E(XXH ) = I only with i.i.d zero mean Gaussian MIMO model!


The need to model: Statistical channel models stimulate creativity (patents!):

to optimize the codes


to estimate the channel
in order to achieve the information theoretic limits.

This can not be performed with simulation or measurement based models!

Types of questions channel modelling will answer

Multiple versus Single Antenna


SISO AWGN Channel: C = log2(1 + ) log2() at high SNR.
MIMO Channel:

Suppose that the channel matrix is deterministic with equal entries 1 (Pure Line of Sight
case):
C

=
=

log2 det(Inr +
nr
X
i=1

log2(1 +

H
HH ).
nt

i )
nt

log2(1 + nr )

log2()at high SNR

i.i.d Rayleigh fading: C = min(nr , nt) log2() at high SNR.


Is there a Contradiction?
6

Let us start...

Model Construction

The i.i.d Gaussian model

The modeler would like to attribute a joint probability distribution to:

h11(f ) . . . . . . h1nt (f )

...
...
... ...

H(f ) =

...
...
... ...

hnr 1(f ) . . . . . . hnr nt (f )

(1)

Assumption 1: The modeler has no knowledge where the transmission took place (the
frequency, the bandwidth, the type of room, the nature of the antennas...)
Assumption 2: The only things the modeler knows:
For all {i, j},
P
E( i,j | hij |2) = nr ntE

What distribution P (H) should the modeler assign to the channel based only
on that
specific knowledge?

The i.i.d Gaussian model

Principle of maximum entropy


Maximize the following expression:

Z
dHP (H)logP (H) + [nr ntE

dH

nt
nr X
X

| hij | P (H)]

i=1 j=1

Z
+ 1
dHP (H)

Solution:

r X
t
X
1
| hij |2
exp{
P ( H) =
}
n
n
r
t
(E)
E
i=1 j=1

Contrary to past belief, the i.i.d Gaussian model is not an assumption but the result of
finite energy knowledge.
This method can be extensively used whenever additional information is provided in terms
of expected values.
9

Knowledge of the covariance structure

In the general case, under the constraint that:


Z

hihj PH|Q(H)dH = qi,j


CN

for (i, j) [1, . . . , N ]2 (N = nr nt). Then using Lagrangian multipliers,


Z
log(PH|Q(H))PH|Q(H)dH
L(PH|Q) =
CN

+
+

1
X

CN

PH|Q(H)dH

Z
i,j

CN

hihj PH|Q(H)dH qi,j .

we obtain:

1
H 1
PH|Q(H) =
exp (vec(H) Q vec(H)) .
det( Q)

10

Existence of Correlation

Question
What to do if we know the existence of correlation but not its exact value?
Answer

Z
P ( H) =

Z
P (H, Q)dQ =

P (H | Q)P (Q)dQ

1- Determine the a priori distribution of the covariance matrix based on limited information
at hand
2- Marginalize with respect to the a priori distribution

11

Construction of the a priori

Let us determine the a priori distribution of the covariance


Suppose that we only know that E(Trace(Q)) = nr ntE (The covariance is not fixed but
varies due to mobility for example)
Result. The maximum entropy distribution for a covariance matrix Q under the constraint
E(Trace(Q)) = nr ntE is such as:
Q = UU

where:

U is haar unitary distributed matrix.

= diag 1, ..., nr nt is diagonal matrix with independent Laplacian distributions.

P (Q)dQ =

1
E nr nt

n n 1

r r
n=0

(n!(n + 1)!)e

Trace(Q)
E
i>j (i

j ) dUd

Note that Q is nothing else than a Wishart matrix with nr nt degrees of freedom.
12

MIMO Channel distribution with correlation

What do we need to do?


Z
P ( H) =
P (H | Q)P (Q)dQ

Z Z
=

Qn r n t
i=1

Trace(vec(H)vec(H)H UH 1 U)

nr nr 1

(n!(n
n=0
E nr nt

+ 1)!)e

Trace(Q)
E
i>j (i

j ) dUd

We need to integrate over U and !


Difficult problem...but well known in statistical physics!

13

MIMO Channel distribution with correlation


Harish-Chandra, Differential Operator on a Semi-Simple Lie Algebra, Amer. J. Math. 79
87-120 (1957)

Harish-Chandra, 1923-1983
Harish-Chandra integral

mTrace(1 UU1 )

e
UU (m)

dU =

1 k
det(e j
)

(1)()
14

MIMO Channel distribution with correlation

Maximum Entropy Analytical MIMO Channel Models, M. Guillaud, M. Debbah and A.


Moustakas, submitted to IEEE transactions on Information Theory, 2006.
Solution. P (H) is given by
nr nt

P (H)

X
n=1

Trace(HH ) n+nr nt2


2
)
Kn+nr nt2(2
2(
E

s
Trace(HHH )
).
E

(1)nnr nt
[(n 1)!]2(nr nt n)!
Kn(x) are bessel functions of order n.
We have therefore an explicit form that can be used for design when correlation exists in
the MIMO model but we are not aware of the explicit value of the correlation!

15

MIMO Channel distribution with correlation

What happens when we know of the existence of a channel covariance matrix with rank L?
Using the same methodology (and integration on a lower subspace), we obtain:

(L)

(H)

2
Trace(HHH )

L
X

i=1

1
2

[(i 1)!] (L i)!

L+i
Trace(HHH )

E0

Ki+L2 2L

Trace(HHH )

E0

Kn(x) are bessel functions of order n.

16

Some distributions

iid Gaussian (2)


(16)

MaxEnt Px (x)

0.1

MaxEnt P(12)(x)
x

0.08

(8)

PDF of x=||H||

2
F

MaxEnt Px (x)
MaxEnt P(4)(x)

0.06

(2)

MaxEnt Px (x)
0.04

0.02

0.02
0

10

20

30
Energy x

40

50

60

Examples of the limited-rank covariance distribution for a 4 4 MIMO matrix and


L = 2, 4, 8, 12 and 16, and 2 with 16 degrees of freedom, for E0 = 16.

17

Some distributions

1
0.9
0.8
0.7

CDF

0.6
0.5

MaxEnt rank 1
0.4
0.3
0.2

MaxEnt rank 2
MaxEnt rank 3
MaxEnt rank 7
MaxEnt rank 16
iid Gaussian

0.1
0
0

5
10
15
20
mutual information (nats) at 15dB SNR, n =n =4
r

25

CDF of the instantaneous mutual information of a 4 4 flat-fading channel for the MaxEnt
model with various covariance ranks, at 15dB SNR.

18

Presentation

Asymptotic Analysis of MIMO systems

19

Representation of a multiple-antenna system: example

Rx

Tx

nr sr

sr st

st nt

stnt : matrix of direction of departure of size st nt.


nr sr : matrix of direction of arrival of size nr sr .
sr st i.i.d Gaussian matrix of size sr st.
This model is known as the Maxent model (the Kronecker model, Sayeeds virtual
representation and the key-hole model can be shown to be particular cases).

20

CLT for MIMO systems

M. Debbah and R. Muller,


MIMO Channel Modelling and the Principle of Maximum

Entropy, IEEE Transactions on Information Theory, Vol. 51 , pp. 1667 - 1690, May, No.
5-2005

r
y

=
r
=

Hs + n
nt
1
nr sr sr st stnt s + n

nt sr st

For many types of random matrices:

H
2
limn
log2det Inr + HH
nt N (0, ).
nt
nt , nr =
t

21

General result for H =

Result: Let et be the eigenvalues of matrices


an sr st i.i.d zero mean Gaussian matrix):

nt
X

log2(1 + ir) +

i=1

nr
X

1
sr st

H
1
sr

and

H
1
st

respectively ( is

log2(1 + iq) nr qr

i=1
2

= 2 log(1 g(r, q))


"
#"
#
nt
nr
X
X
1
i
1
i
2
2
g(r, q) =
(
)
(
)
nt i=1 1 + iq
nt i=1 1 + ir
n

t
1 X
i
r=
nt i=1 1 + iq
n

r
1 X
i
q=
nt i=1 1 + ir

22

Elements of Proof

H
+ HH
nt

1
log det
nt

r
1 X
nr
log(1 + i)
nt i=1
nt

I nr

nr
nr 1 X
nr
log(1 + i)
nt nr i=1
nt
Z
nr
nr
log(1 + )dF 1 HHH ()
nr
nt
nt

23

Elements of Proof

More specifically, let

1 H
H
= V V
sr
1
H
H
= V V
st
V and V are unitary matrices while and are diagonal matrices representing

1
1
1
r1
H
r2
H t2
1
1 t2
2
respectively the eigenvalues of matrices sr P P and s P P . The
t
1
1
1
1
non-zero eigenvalues of matrix n1r HHH = nr s1r s Pr 2 Pt 2 H Pt 2 H Pr 2 H are the
t
1
1
1
1
H
H H
H
1
2
2
2
2
same as 11 = nr [(V V)][(V V)]. Without loss of generality,
we will suppose that sr nr . Therefore, the spectra of n1r HHH and 1H
1 are related

by:

sr
sr
)(x 0) +
f
H (x)
nr
nr 11
and their Stieltjes transforms are related as:
fHHH (x) = (1

mHHH (z) = (1

sr
sr 1
) +
m
H (z)
nr z
n r 1 1

24

Elements of Proof

Matrix VV is an i.i.d zero mean Gaussian matrix with unit variance (only unitary
transforms are applied). Therefore, matrix 1 =

1
1
1
2
2
[ (V V ) ]

nr

is a sr st

random matrix composed of independent entries with zero mean and variances

1
1 i j
. The weak convergence of the empirical eigenvalue distribution of
nr i j = s r

1H
1 to a limiting distribution holds under certain assumptions and is an application of

the theorem of Girko.

25

Remarks on capacity of the double directional model

Proposition: In the high SNR regime, the mean mutual information of the double
directional model converges to:

Z
Z
min nt, nr , sr
dSdoa(), st
dSdod() log2()
>0

>0

R
R

The factor min nt, nr , sr >0 dSdoa(), st >0 dSdod() denotes the multiplexing gain.
R

R
dS
()
and
dSdod()) express the correlation factor of the sr and st
doa
>0
>0
scatterers respectively.
The MIMO mutual information is limited by the scattering environment!

26

Mutual information compliance

For a given frequency f , a model will be called capacity complying if it minimizes:

| F (C, f ) Fempirical(C, f ) | dC

(2)

Here Fempirical(C, f ) is the empirical cdf given by measurements.

How to derive the theoretical cdf F (C) of the capacity?

F (C) = 1 Q(
with

since C = log2det Int

C nt
)

Z
2
1
t2
Q(x) =
dte
2 x

H
+ n H H is asymptotically Gaussian.
t

27

Measurements of channels

Measurements have been performed:

at 2.1 GHz
at 5.2 GHz
In the following scenarios:

Indoor
Atrium
Urban Open Place
Urban Regular High Antenna
Urban Regular Low Antenna

28

Sounder characteristics at 2.1 GHz

Measurement frequency
Measurement bandwidth
Delay resolution
Sounding signal
Transmitter antenna
Element spacing
Receiver antenna
Element spacing

2.1 GHz
100 MHz
10ns
linear frequency chirp
8 element ULA
71.4 mm (0.5 )
32 (8 4) element
73.0 mm (0.51 )

29

Sounder characteristics at 5.2 GHz

Measurement frequency
Measurement bandwidth
Sounding signal
Transmitter antenna
Element spacing
Receiver antenna

5.255 GHz
100 MHz
linear frequency chirp
8 element ULA
28.54 mm (0.5 )
8 element ULA

30

Antennas

The receiver is an 8 4 antenna whereas the transmitter is an 8 elements ULA.

31

Example: Urban Open Place at 2.1 GHz

32

Is mutual information Gaussian at 2.1 GHz

Are the measured mutual information Gaussian?


Measured
Gaussian approximation

0.9

0.8
Urban RegularHIgh Antenna

0.7

Indoor
0.6
CDF

Urban Open Place

0.5

0.4
Atrium
0.3
Urban regular low Antenna
0.2

0.1

0
15

16

17

18

19

20
b/s/Hz

21

22

23

24

25

Mutual information has a Gaussian behavior for an 8 8 system at 2.1Ghz!

33

Is mutual information Gaussian at 5.2 GHz

Are the measured mutual information Gaussian?


1

Measured
Gaussian approximation

0.9
Atrium

0.8

0.7
Urban Open Place

0.6
CDF

Urban Low Antenna


0.5
Indoor
0.4

0.3

0.2

0.1

0
15

16

17

18

19

20

21

22

23

24

b/s/Hz

Mutual information has a Gaussian behavior for 8 8 system at 5.2Ghz!

34

Rice Model Case

Single Antenna/Multiple Antennas


SISO Channel: Capacity of the AWGN channel is higher than the ergodic capacity of a
Rayleigh fading channel (Jensens inequality).
Rice MIMO Channel: Suppose that the matrix representing the channel has deterministic
entries equal to 1:

=
=

log2det(Inr +
nr
X
i=1

log2(1 +

H
HH ).
nt

i )
nt

log2(1 + nr )

log2()at high SNR

MIMO i.i.d zero mean Gaussian : C = min(nr , nt)log2() at high SNR.


Is there a contradiction?
35

Rice Model Case

Diversity versus path loss trade-off


Free Space Propagation:

max
2
r

Propagation with reflections:

max
r 2n+2
(n is the number of reflections)

There is a trade-off between the gain in SNR due to the line of sight component and the
multiplexing gain due to multiple reflections.
The analysis of MIMO Rice channels is important to determine the interval of interest.

36

Rice Model Case

Rice versus Rayleigh


Rayleigh fading:

Rayleigh channels have been studied extensively in the case of non-line of sight.
In the case of i.i.d entries, ergodic capacity increase is min(nr , nt) bits per second per
hertz for every 3dB increase at high SNR.
Highly scattered environnements increase ergodic capacity.
Rice fading:

Many propagation scenarios have line of sight components (Inter-base station


transmissions,..).
Capacity studies limited to rank 1 Rice fading or based on very loose bounds.
The impact of line sight with scattering is still an issue!

37

Rice Model Case

General Model

s
H=

s
K
A+
K+1

1
B
K+1

A is the deterministic line of sight component part of the matrix such as kAk2F = ntnr
B is a gaussian unit variance zero mean i.i.d matrix.
The average energy of the channel is normalized according to:

Assumption: The matrix size

H
1
AA
nt

the empirical eigenvalue distribution of


limit function F A .

grows large with =


H
1
nt AA

nr
nt

E(tr(HHH ))
nr nt

= 1.

remaining fixed such as

converges in distribution to a deterministic

nt

Remark: The rank of A increases at the same rate as the number of antennas.

38

Rice Model Case

Remember the useful result


Let mH(z) be the Stieljes transform of the limiting eigenvalues FH() of

Z
mH(z) =

HHH
nt :

dFH()
z

As nr = nt ,

Z
mH(z) =

K
mH (z)+K+1

d FA()

mH (z)
z
K+1 + 1 +

1
K+1

The asymptotic eigenvalue distribution of HH


nt is completely determined knowing only
FA(), and K and not the particular fluctuations of the fading!

39

Rice Model Case

Perfect Channel Knowledge at the transmitter


The channel capacity per receiving antenna converges almost surely, as nr = nt ,
to:

Z +
1

C1 =
ln d FH()
ln 2 1

Z +
1


d FH() =
1

Remark:

The capacity is achieved through waterfilling power allocation.


The channel should change slowly enough in order to have complete channel
knowledge at the transmitter.

40

Rice Model Case


Channel knowledge of the limiting singular value distribution of A
L. Cottatellucci, M. Debbah, The Effect of Line of Sight Components on the Asymptotic
Capacity of MIMO Systems, 2004 IEEE International Symposium on Information
Theory,Chicago, June 27 - July 2, 2004, USA
Since the transmitter has no knowledge of the eigenvector structure, the transmitted
power is equally distributed among the antennas.
The channel mutual information per receiving antenna converges almost surely, as
nr = nt , to:

1
C=
ln 2

Z
0

1
x

1
1 mH
x

dx

Remark:
1
z(mH (z)+1)+(1) (Marchenko Pastur
R d FA()
=
z (distribution of the mean)

If K 0, mH(z) =
If K , mH(z)

Distribution)

41

Rice Model Case

Perfect Knowledge of line of sight matrix A


and A. Lapidoth, The Capacity of a MIMO Ricean Channel is Monotonic in the
D. Hosli
Singular Values of the Mean, 5th International ITG Conference on Source and Channel
Coding (SCC), Erlangen, Nuremberg, jan, 2004
The optimal covariance matrix Q is shown to have the same eigenvectors as AH A.

42

Rice Model Case


Perfect Knowledge of line of sight matrix A
The capacity is then given by:

1
b
log2 det
J(H, UQU
)=
nr

1
=
log2 det
nr
1
=
log2 det
nr

e = V HU =
H
H

Inr

Inr

Inr

K b
K+1 A

b H HH
+ HUQU
nt
e b eH H
+ VH
QH V
nt

e b eH
+ HQ H
nt

q
+

H
1
V
BU
K+1

VHBU is a random matrix with i.i.d zero mean Gaussian entries

q
The line of sight matrix

K b
K+1 A

is diagonal.
43

Rice Model Case


Perfect Knowledge of line of sight matrix A
The asymptotic channel capacity per receive antenna converges almost surely to:

C2(H) =

b (,F (),K,)
Q
A

1
b
J(H, UQU
)=
ln 2
H

max
Z

1
x

b H)
J(H, UQU

1
1 m2
x

dx

R
(m1(z) z(K + 1))q(K + 1) d FA()
(1 )q0(K + 1)

m
(z)
=
+

1
(K + 1 + m2(z)q)(m1(z) z(K + 1)) + (K + 1)Kq
(K + 1 + m2(z)q0)

m
(z)
=
2

(K + 1 + qm2(z))(K + 1) d FA()
(K + 1 + m2(z)q)(m1(z) z(K + 1)) + q(K + 1)K

b and q0 = q(0, F (), K, ).


q = q(, F (), K, ) denotes the diagonal entries of Q

Remark: The solution is not waterfilling on the mean matrix!


44

Rice Model Case


Simulations: general results
FA = 14 ( 3) + 34 ( 13 ), = 1, = 5dB.
Rice channel capacity for FA()=(3)/4+3 (1/3)/4, =1, =5 dB
1.85

C1: Asymptotic evaluation


C1: Simulation with 8 8 MIMO system
C2: Asymptotic evaluation
C2: Simulation with 8 8 MIMO system
C3: Asymptotic evaluation
C3: Simulation with 8 8 MIMO system

Capacity (bit/s/Hz)

1.8

Rayleigh channel

1.75

1.7

1.65

1.6 2
10

10

10

10

10

Rice factor K

10

10

Close match with ergodic capacities for a system with 8 8 antennas.


Mean feedback is sufficient for values of K > 10.

45

Rice Model Case


Simulations: Low and high SNR regime
FA = 14 ( 3) + 34 ( 31 ), = 1.
Rice channel with F ()=1/4 (3)+3/4 (1/3), =1.

Rice channel with FA()=1/4 (3)+3/4 (1/3), =1.

0.18

=0 dB

=0 dB
0.25

0.14

2 dB

0.12

1 dB

(C C )/C

4 dB

0.15

1 dB

0.1

2 dB

0.08

(C1C3)/C3

0.2

0.16

6 dB

0.06

0.1

8 dB

0.04

10 dB

0.02

4 dB
6 dB
8 dB
10 dB

0.05
0

0 2
10

10

10

10

Rice factor K

10

10

10

0.02 2
10

10

10

10

10

10

10

Rice factor K

At high SNR, the effect of feedback on the capacity diminishes.


For low SNR, the effect of feedback is quite important and depends mainly on the Ricean
factor K .

46

Rice Model Case


Simulations: reduced rank, mean matrix known at the transmitter.
n
8n
8
FA = 8t () + n t ( 8n
), = 1, = 10.
t

Rice channel capacity C for F ()=p ()+p (1/p ), =1, =10.


2

3.5

p = 0
0
1/8
2/8

4/8
2.5

5/8

Capacity C , (bits/s/Hz)

3/8

6/8
1.5

7/8
1 2
10

10

10

10

Rice factor K

10

10

10

Rank reduction yields a dramatric effect on the capacity.


Transition phase with respect to K around [1,10].

47

Rice Model Case


Simulations: rank 1, perfect channel knowledge at the transmitter.
n 1
FA = tn () + n1 ( nt), = 1, = 10.
t

Rice channel capacity C1 for FA()=(nt1)/nt ()+(nt)/nt, =1, =10.


3

n= 2
t

3
2

Capacity C , (bits/s/Hz)

2.5

5
1.5

6
7
8
10

16
0.5 2
10

23
1

10

10

10

10

10

10

Rice factor K

Asymptotic analysis able to capture the rank 1 case.


Close match with a small number of antenna elements.

48

Rice Model Case


Rice or Rayleigh?
Asymptotic analysis of Ricean MIMO channels show that Rice fading is not always better
than Rayleigh fading.
The result depends in fact on:

the limiting behavior of the mean matrix.


the ratio =
the SNR .

nr
nt .

the Ricean factor K .


through various fixed point equations depending on the channel state of knowledge
available at the transmitter.

49

Rice Model Case


More Reading
J. Dumont, P. Loubaton, S. Lasaulce and M. Debbah, On the Asymptotic Performance of
MIMO Correlated Ricean Channels, in Proc. IEEE Int. Conf. on Acoustics, Speech and
Sig. Proc. (ICASSP), Montreal, Canada, Mar. 2005, pp. 813-816.
J. Dumont, W. Hachem, P. Loubaton and J. Najim, On the asymptotic Analysis of mutual
information of MIMO Rician correlated channels, IEEE-EURASIP Int. Sym. on Control,
Commun. and Sig. Proc. (ISCCSP 2006), Marrakech, Morocco, Mar. 2006
A. L. Moustakas and S.H. Simon, Random matrix theory of multi-antenna
communications: the ricean case, J. Phys. A: Math. Gen. 38 (2005), 10859-10872.
J. Dumont, P. Loubaton and S. Lasaulce, On the capacity Achieving Transmit Covariance
Matrices of MIMO Rician Channels: A Large System Approach, GlobeCom 2006,
San-Francisco, USA, FIRST PATENT DUE TO FRANCE TELECOM based on Random
matrices.

50

Presentation

Asymptotic design of receivers

51

Model

y
Received signal
N 1

W
code matrix
N K

P2
Amplitude Gains
KK

s
emitted signal
K1

n
AWGN
N (0, 2IN )

The goal is to detect s.


Linear MMSE Filter:
LMMSE =

1H
H
2
P W

1
R+ I

R = WPW

52

Matrix Inversion

Let R be non-singular.
Let i be the eigenvalues of R.
Then,

QK

k=1 (R k I) = 0
Cayley Hamilton Theoreom

PK
I + k=1 k Rk1 = 0
with appropriate k0s.

Solution to matrix inversion problem given the eigenvalues

K
X

k R

k=1

53

Linear Multi-Stage Detection

M. Honig and W. Xiao, Performance of Reduced Rank Linear Interference Suppression,


IEEE Transactions on Information Theory, vol. 47, No.5, July 2001.

2 1
Linear MMSE Filter: LMMSE = R + I
Approximation by Power Series:
Cayley-Hamilton Theorem yields:

R+ I

K1
X

i
iR

i=0

'

D1
X

i R

for D < K

i=0

For a N K matrix W with i.i.d random spreading and R = WWH , the optimum

weights converge almost surely as K, N with = K


N , and can be given in closed

form.
54

Semi-Universal Weights

R. Muller
and S. Verdu, Design and Analysis of Low-Complexity Interference Mitigation

on Vector Channels, IEEE Journal on Selected Areas in Communications, vol. 19, no. 8,
pp. 1429-1441, August 2001.
Filter shall be independent from the realization of the random matrix W, but may use its
statistics.
For most large random matrices, as K = N , many finite dimensional functions of
the eigenvalues, e.g, the filter coefficients free.
The asymptotic limits depend only on parts of the statistics of the random matrix.
The weights can be calculated off-line with the help of random matrix theory and free
probability theory.

55

Weight Design

is given by the Yule-Walker equations:

m1
m2
...
mD+1

m2 + m1
m3 + 2m2
...
mD+2 + 2mD+1

with the total moments:

m3 + m2
m4 + 2 m3
...
mD+3 + 2mD+2
n

...
...
...
...

mD+2 + mD+1
mD+3 + 2mD+2
...
m2D+2 + 2m2D+1

w0
w1
...
wD+1

H n

mn = E( ) = Trace(WW )
.

56

Example for Weight Design

Random with i.i.d entries


D=2
D=3

D=4

w0 = 2m1 + 2 + 2
w1 = 1
w0 = 2m1 + 3 + 4 + 3 2
w1 = 2m2 3 3
w2 = 1
w0 = 2m1 + 4 + 6 + 6 2 + 4 3
w1 = 2m2 6 9 6 2
w2 = 2m3 + 4 + 4
w3 = 1
n

1X n n i
mn =
C C
n i=1 i i1

57

Rate of Convergence

Ph. Loubaton and W. Hachem, Asymptotic Analysis of Reduced Rank Wiener Filters,
Information Theory Workshop 2003, Paris, France
Theorem: Suppose that the elements of the matrix are i.i.d zero-mean random variable
with finite fourth moment. Then the multi-user efficiencies of all users converge almost

surely as N, K but K
N fixed to:

D+1 =

1
1+

2
+

with 0 = 0 for optimally chosen weights.


The approximation converges to the exact MMSE performance as a continued fraction.
For optimally weights wi, the approximation error decreases exponentially with the
number of stages D .
D

< Const(1 + SN R)

58

Individual Weight Design

Allow for different weight for different users:

R+ I

K1
X

w
iR

i=1

D1
X

WiR

for D < K and all Wi diagonal

i=0

Weight design by the same Yule-Walker equations, but with the k-partial moments:
h
i
k
H n
mn = (SS )
kk

For users with different powers, individual weight design is better.


Do the k-partial moments converge asymptotically?
59

Convergence of partial moments

Let the random matrix S fulfill the same conditions as before. Let A be an K K
diagonal matrix such as its singalur value distribution converges almost surely as
K to a non-random limit distribution. Let R = ASSHAH
Then Rlkk , the k-th diagonal element of Rl converges, conditioned on akk , the k-th
diagonal element of A, almost surely, as K = N to
Rlkk =| akk |2

Pl
q=1

R
R
q
Rq1
kk mlq with mq = Trace(R )

The total moments of R are conveniently given by the recursion:


R
ml

l
X
q=1

R
mlq

K
1 X
2
q1
| akk | Rkk
lim
K K
k=1

60

Model

The system is described by a virtual N R K spreading matrix

=
S

h11s1
h21s1
...
hR1s1

h12s2
h22s2
...
hR2s2

...
...
...
...

h1K sK
h2K sK
...
hRK sK

Note that with the Kronecker product :

sk = hk sk

are not jointly independent even if those ones of S and H


Note also that the entries of S
are.

61

Ressource Pooling Result

S. Hanly and D. Tse, Resource Pooling and Effective Bandwidths in CDMA Systems with
Multiuser Receivers and Spatial Diversity, IEEE Transactions on Information Theory, vol.
47(4), May 2001, pp. 1328-1351
Theorem Let the chips of any user be i.i.d zero-mean random variables with finite 6th
moment and the antenna array channel hrk follow the i.i.d Complex Gaussian Model.

Then the multiuser


efficiency MMSE of the linear MMSE detector converges for all users

almost surely as N, K but = K


N and R fixed to the deterministic unique positive
solution of the fixed point equation:
Z

x
1
=1+
dPA2 (x)
2
MMSE
R
+ MMSEx
if the power of the users converge weakly to the limit distribution PA2 with:

| Ak | =| Ak |
2

R
X

| hrk |

r=1

62

Ressource Pooling result for correlated MIMO Channels

L. Cottatellucci, R. Muller,
A Generalized Resource Pooling Result for Correlated

Antennas with Applications to Asynchronous CDMA International Symposium on


Information Theory and its Applications, Parma, Italy, October 2004
Theorem Let the chips of any user be i.i.d zero-mean random variables with finite 6th
moment and the empirical distribution of the channel gains hrk across the users
converge, jointly for all receive antennas r to an R-dimensional joint limit distribution PH.

Then, with linear MMSE detection, the SINR


converges as N, K but = K
N and R
fixed, conditioned on the channel gains of user k to:

hH
k Ahk
2
where A is the deterministic unique positive definite solution of the matrix valued fixed
point equation

Z
1

=I+

xxH
dPH(x)
2 + xH Ax

Asymptotic Performance is characterized by an R R matrix.


63

Multi-stage Detection for Correlated Ressource Pooling

As the dimension of S grows, the following result hold:

the k partial moments conditioned on hk converge


Explicit Recursive expressions for them are now available

The proof follow along the same lines as before.

The system is sensitive to the correlation at the receiving side.

64

Last Slide

THANK YOU!

65

You might also like