Supelec

Random Matrix Theory

for
Wireless Communications

Mérouane Debbah
http://www.supelec.fr
[email protected]

February, 2008
 Presentation

The role of the Cauchy-Stieltjes Transform in Communications

1
 General Multiple Input Multiple Output Model

Model representing multiple-antennas, CDMA, OFDM, ad-hoc networks with

cooperation,...

y = W s + n
Received signal MIMO matrix emitted signal AWGN
N ×1 N ×K K×1 ∼ N (0, σ 2IN )

Let  
· ¸
s1
W =  u U , s=
x

The goal is to detect s.

2
 Communications Notations Used
Receiving vector

• r as ”Received”.
• y as ”Y do we care?”.

Transmitting vector

• s as ”Signal”.
• x as ”Xciting”.

Noise vector

• n as ”Noise” (english).
• b as ”Bruit” (french).
• z as ”Zzzzzzz...” (disturbance)

MIMO matrix

• W when the matrix is non-isometric.

• Θ when the matrix is isometric
• H when we consider multiple-antennas.

3
 Shannon Capacity

Mutual information M between input and output:

M (s; (y, W)) = M (s; W) + M (s; y | W)

= M (s; y | W)
= H(y | W) − H(y | s, W)
= H(y | W) − H(n)

The differential entropy of a complex Gaussian vector x with covariance Q is given by

log2 det(πeQ).

4
 Shannon Capacity

In the case of Gaussian independent entries, since

H 2 H
E(yy ) = σ IN + WQW
H 2
E(nn ) = σ IN

The mutual information per dimension is:

1
CN = (H(y | W) − H(n))
N
1 ³ 2 H 2
´
= log2det(πe(σ IN + WQW )) − log2det(πeσ IN )
N
µ ¶
1 1 H
= log2det(IN + 2 WQW )
N σ

5
 Shannon Capacity

Consider the random variable

µ ¶ N µ ³ ´¶
1 1 H 1 X 1 H
CN = log det IN + 2 WW = log 1 + 2 λk WW
N σ N k=1 σ

When N → ∞ and K/N → α,

Z µ ¶
1
CN → log 1 + 2 t µ(dt) a.s.
σ
dCN 2 4 2
1
→ σ − σ Gµ (−σ ) a.s.
d σ2

The capacity is strongly related to the Cauchy-Stieltjes transform.

6
 Some numerical facts

³ ´
W i.i.d. zero mean with variance 1
N: CN = 1
N log det IN + 1
σ2
WWH

Mean for variable matrix size at 10dB

2.95
Simulations
Theoretical formula

2.9

2.85
b/s/Hz

2.8

2.75

2.7
0 5 10 15
Number of antennas

• With N= 6, the empirical mean is at 0.02% of the asymptotic value.

• With N= 3, the empirical mean is at 0.6% of the asymptotic value.

7
 MMSE Receiver

Model example :

y = Ws + n
= us1 + Ux + n
0
= us1 + n
0 0H H 2 H
E(n n ) = (UU + σ I) = QΛQ
Whitening filter:

−1 H −1 H 1
−2 H 0
ỹ = Λ 2 Q y = Λ 2 Q us1 + Λ Q n
= gs1 + b

b is a white Gaussian noise.

8
 MMSE Receiver

−1 H
ỹ = Λ 2 Q us1 + b

1
Define g = Λ− 2 QH u

The output SINR is maximized with:

H H H
g ỹ = g gs1 + g b

As a consequence, the receiver is:

³ ´ ³ ´−1
H −1 H H −1 H H H 2
g Λ 2 Q =u QΛ Q =u UU + σ IN

Remark: The usual MMSE receiver is the unbiased one:

³ ´−1 1 ³ ´−1
H H 2 H H 2
u WW + σ IN = ¡ ¢−1 u UU + σ IN
H H 2
1 + u UWU + σ IN u

9
 MMSE Receiver

After MMSE filtering, we obtain:

H H H
g ỹ = g gs1 + g b

1
with g = Λ− 2 QH u

Signal to Interference plus Noise Ratio (SINR):

(gH g)2E(| s1 |2) ³ ´−1

H H H 2
βN = H
=g g=u UU + σ IN u
g g

Depends strongly on the choice of U.

10
 Example: the i.i.d. model case

³ ´−1
H H 2
βN = u UU + σ IN u
h i
2
Suppose the matrix W = [Wij ] has i.i.d. elements, E [Wij ] = 0, E Wij = 1/N .
Example : IS95.

You remember that important lemma?

u vector N × 1 with i.i.d elements. Each element : zero mean and variance 1/N .
A matrix N × N independent of u. Then, under some assumptions,

H 1
u Au − trace (A) → 0 a.s.
N
when N → ∞.

Application : u et U independent, so
µ³ ´−1¶
1 H 2
βN − trace UU + σ IN → 0 a.s.
N

11
 Example: the i.i.d. model case

µ³ ´−1¶ ³ ³ ´´
1 H 2 1 H
trace UU + σ IN = trace f UU
N N
where f (t) = 1/(t + σ 2).

Since UUH has a limiting Marchenko-Pastur distribution law µ, we have

Z
1 2
βN → β = µ(dt) = G µ (−σ ) a.s.
t + σ2

Solution : s
1−α 1 (1 − α)2 (1 + α) 1
β= − + + +
2σ 2 2 4σ 4 2σ 2 4
The SINR at the output of the MMSE receiver is exactly the Cauchy-Stieltjes transform.

12
 MMSE and Capacity

2
SINRMMSE = Gµ(−σ )
dC 2 4 2
1
= σ − σ G µ (−σ )
d σ2

dC 1
= SINRMMSE −
σ2 σ2

The derivative of the capacity is strongly related to the performance of the MMSE receiver!

13
 When Wiener meets Shannon

D. Guo, S. Shamai, and S. Verdú, “Mutual Information and Minimum Mean-Square Error in
Gaussian Channels,” IEEE Trans. Information Theory, vol. 51, no. 4, pp. 1261-1283, Apr.
2005.

D. P. Palomar and S. Verdú, “Gradient of Mutual Information in Linear Vector Gaussian

Channels,” IEEE Trans. Information Theory , vol. 52, no. 1, pp. 141-154, Jan. 2006.

Norbert Wiener, 1894-1964

Claude Shannon, 1916-2001

14
 The Universality of Wiener’s solution

Still, no explanation for why an MMSE estimator, which is seemingly an artifact from the
world of analog signals, plays such a key role in achieving channel capacity.

Perhaps, best stated by Shannon himself in 1948 paper acknowledging Wiener:

”Credit should also be given to Prof. N. Wiener whose elegant solution of the problem of
filtering and prediction of stationary ensembles has considerably influenced the writer’s
thinking in this field.”

15
 Outage performance

You remember the CLT result?

2
lim N (C − µ) → N (0, σ )
K→∞, K =α
N
1 1 H
C= N log det(I + σ2
WW ) is the mutual information per dimension and R = N C .

• To optimize the network (CDMA for example), only the mean and the variance is
needed!
• Application: How to derive the outage mutual information? Let q be the outage
probability and Rq be the :

Z Tq
q = P (R ≤ Rq ) = p(R)dR
−∞
−1
Rq = N µ + σQ (1 − q)
R∞ 2
− t2
with Q(x) = √1 e dt
2π x

16
 Presentation

Asymptotic Analysis of (MC)-CDMA systems

17
 Uplink CDMA

• K users in the cell.

• Code Division Multiple Access (CDMA): Simultaneous communication of all the users
to the base station using different codes.

18
 Uplink CDMA: General fading model

The N × 1 received signal vector y at the base station has the form:
√ √ √
y = H1w1 P 1s1 + H2w2 P 2s2 + ... + HK wK P k sK + n

• s = (s1, . . . , sK ) is the emitted symbol vector.

• wk is the N × 1 k-th user code.
• Hk is the N × N channel matrix of user k.
• N is the spreading length.
√
• P k is the amplitude of user k.
• n is a N × 1 white complex gaussian noise vector of variance σ 2.

19
 Uplink CDMA: Flat fading model

D. N Tse and S. Hanly, ”Linear Multiuser Receivers: Effective Interference, Effective

Bandwidth and User Capacity”, IEEE Transactions on Information Theory, Vol.45, No. 2,
Mar. 1999

S. Verdu and S. Shamai, ”The Impact of Frequency-Flat Fading on the Spectral Efficiency
of CDMA”, IEEE Transactions on Information Theory, Vol. 45, No. 2, Mar. 1999

The N × 1 received signal vector y at the base station has the form:
√ √ √
y = h1w1 P 1s1 + h2w2 P 2s2 + ... + hK wK P K sK + n
1
= WHP 2 s +n

• s = (s1, . . . , sK ) is the emitted symbol vector.

• wk is the N × 1 k-th user code.
• H = diag[h1, ..., hK ] an K × K matrix with hk flat fading of user k.
• P = diag[P1, ..., PK ] an K × K matrix with Pk power of user k.
• N is the spreading length.
• n is a N × 1 white complex gaussian noise vector of variance σ 2.

20
 MMSE receiver for Uplink flat fading CDMA

SINR of user k:
k 2 k
SINR (K, N ) = Pk | hk | π (K, N )

where  −1
K
X
k H 2 H 2
π (K, N ) = wk  Pi | hi | wiwi + σ I wk .
i=1,i6=k

In the case of i.i.d codes,

1 ³ ´−1
k H H 2
π (K, N ) → trace WHPH W + σ I
N
µ ¶
K
→ β
N
Z
f (λ)dλ
→
λ + σ2
2
→ GWHPHH WH (−σ )

21
 MMSE receiver for Uplink flat fading CDMA

You remember that famous result...

2
β = GWHPHH WH (−σ )
is such as:

à Z !
tν(dt)
GWHPHH WH (z) = G0.I z−α
1 + tGWHPHH WH (z)

ν(dt) is the limiting distribution of the fading matrix HPHH .

In the case of i.i.d Rayleigh fading and equal power P = 1, ν(dt) = e−tdt

µ Z ¶
tν(dt) 1
β = G0I −α = R tν(dt)
1 + tβ 2
σ +α 1+tβ

22
 MMSE receiver for Uplink flat fading CDMA

As a consequence, the SINR for user k is given by:

k 2
SINR (K, N ) =| hk | Pk β

where β is solution of:

1
β= PK
1 |hi |2 Pi
σ2 + N i=1,i6=k 1+|h |2 P β
i i

23
 MMSE receiver for Uplink flat fading CDMA

Performance measure with MMSE Receiver:

The spectral efficiency of the system is the number of bits/s/Hz that can be reliably
transmitted.

K
1 X
γ = Ci
N i=1
K
1 X 2 i
= log2(1+ | hi | Piπ (K, N ))
N i=1
³ ´
2
= αEh log2(1+ | h | P β(α))

Eb 2 −1
= (Cσ )
N0

24
 How can one extend the result to other receivers

Optimal Filter: The total spectral efficiency of the system is given by:

µ ¶
1 1 H H
γ= log det I + 2 WHPH W
N σ

and
µ ¶
dγ 2 σ2 1 H H
= σ − Trace I + 2 WHPH W
d σ12 N σ
σ4 ³ ´
2 2 H H
= σ − Trace σ I + WHPH W
N
2 4
→ σ −σ β

25
 How can one extend the result to other receivers

Matched filter: SINR of user k:

k Pk | hk |2 wkH wk
SINR (K, N ) = H
wk (UHPHH UH + σ 2I) wk

In the case of i.i.d codes,

H
wk wk → 1
³ ´ 1 ³ ´
H H H 2 H H 2
wk UHPH U + σ I wk → Trace UHPH U + σ I
N
Z
2
→ (λ + σ )f (λ)dλ
2 2
→ αEh,P (P | h | ) + σ

SINR of user k:
k Pk | hk |2
SINR (K, N ) →
αEh,P (P | h |2) + σ 2

26
 How can one extend the result to other receivers

MMSE SIC Detector: Principle of the algorithm

2 3
s1

= = HN [ ] 64 75 + n
.
.
Step 1 r1 y w1 ;


; w K .
.

2K
s
3
s1

Step 2 r2 = = HN [
y w1 ;


; w K 1 ] 64 .
.
.
75 + n
s K 1

StepK rK = y = HN w1 s1 + n .

• At each step, the symbols are assumed to be perfectly retrieved.

• At each step, an MMSE equalization taking into account the retrieved symbols is
performed.

27
 Uplink CDMA: flat fading model

Performance Measure with MMSE SIC Detector

K
1 X
γ = Ci
N i=1
K
1 X 2 i
= log2(1+ | hi | Piπ (i, N ))
N i=1
Z α ³ ´
2
= Eh log2(1+ | h | P β(x)) dx
0

Performance Measure with MMSE Detector:

³ ´
2
γ = αEh log2(1+ | h | P β(α))

Important Result: It has been shown that the spectral efficiency of the MMSE SIC
scheme is equal to the spectral efficiency of the optimum receiver (Varanasi & Guess).

28
 Uplink CDMA: Rayleigh flat fading model at 10dB

Single User Gaussian Channel

Single User Rayleigh fading Channel

4
Orth MF/MMSE/Optimum Random Optimum
Spectral Efficiency

2
Random MMSE

Random MF
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Ratio Number of users/Spreading length

Spectral efficiency is only a function of the load of the system, the statistics of the channel
and noise variance.

29
 Uplink CDMA: Frequency Selective fading model

The N × 1 received signal vector y at the base station has the form:
√ √ √
y = H1w1 P 1s1 + H2w2 P 2s2 + ... + HK wK P k sK + n

Toeplitz structure
H
Hi ∼ F DiF

1
wi are supposed to be i.i.d zero mean Gaussian codes with variance N

Equivalent model:

√ √ √
ỹ = D1w̃1 P 1s1 + D2w̃2 P 2s2 + ... + D̃K w̃K P k sK + ñ
³ ´ 1
= C ¯ W̃ P 2 s + ñ

30
 Uplink CDMA: Frequency Selective fading model

Optimal filter: Let h(f, x) (0 ≤ τ ≤ 1) and P (x) (0 ≤ x ≤ α with α = K N ) be

respectively the channel frequency power profile and the attenuations of user [N x].

The total spectral efficiency is given by:

dγ σ4 ³ ´
2 2 H
=σ − Trace σ I + (C ¯ W̃)P(C ¯ W̃)
d σ12 N

31
 Uplink CDMA: Frequency Selective fading model

Optimal filter: You remember Girko’s result?

³ ´ Z 1
1 2 H
lim Trace σ I + (C ¯ W̃)P(C ¯ W̃) = v(f )df
N →∞ N 0

with

" Z #−1
α 2
2 P (x) | h(f, x) | dx
v(f ) = σ + R1
0 1 + P (x) 0 v(u) | h(u, x) |2 du

In the case where P (x) = 1,

" Z #−1
α 2
2 | h(f, x) | dx
v(f ) = σ + R1
0 1+ 0
v(u) | h(u, x) |2 du

32
 Uplink CDMA: Frequency Selective fading model

A. Tulino, L. Li and S. Verdu, ”Spectral Efficiency of Multicarrier CDMA”, IEEE

Transactions on Information Theory, vol. 51, No. 2, February 2005

MMSE receiver: The SINR at the output of the MMSE receiver is given by:
k
SINR = β(k/N )

where
Z 1
| h(f, x) |2 df
β(x) = P (x) R α |h(f,u)|2P (u)du
0 2
σ + 0 1+β(u)

33
 Uplink CDMA: Frequency Selective fading model

In the case where P (x) = 1,

Z 1
| h(f, x) |2 df
β(x) = R α |h(f,u)|2du
0 2
σ + 0 1+β(u)

34
 Uplink CDMA: Frequency Selective fading model

MMSE receiver: For the high number of users finite case, let hk (i) be the frequency
response of user k on frequency i, then the SINR β k at the output of the MMSE receiver
is given by:

N
k Pk X | hk (i) |2
β = PK
N i=1 σ 2 + 1 Pl |hl (i)|2
N l=1,l6=k 1+β l

35
 Uplink CDMA: Flat fading versus frequency selective

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Flat fading represented by diamond curve

Frequency selective with 5 paths represented by star curve
Frequency selective with 128 paths represented by circle curve.

36
 Uplink CDMA: should we synchronize users?

N. Bonneau, M. Debbah, E. Altman and G. Caire, ”When to Synchronize in Uplink CDMA”,

2005 IEEE International Symposium on Information Theory, 4-9 September 2005,
Adelaide, Australia.

Question: When is it useful to use orthogonal uplink signaling?

Answer: Intuitively,

• In the case of flat fading, orthogonality is preserved.

• In the case of frequency-selective fading, orthogonality is destroyed.

37
 Orthogonal Uplink CDMA: Estimation of the SINR for
linear receivers

For the matched filter, g = H1w1.

³P ´2
N 2 2
i=1 | h1(i) | | wi1 |
SINR1 = PN PK PN
| h1(i) |2| wi1 |2 σ 2 + k=2 | ∗ ∗ 2
i=1 i=1 h1 (i) hk (i)wi1 wik |

³P ´−1
K
For the MMSE receiver, g = w1H HH
1 i=1 HiwiwiH HH
i +σ I .2

η1
SINR1 =
1 − η1
³P ´−1
K
with η1 = w1H HH
1 i=1 HiwiwiH HH
i +σ I 2
H1 w 1 .

For given K and N , it is extremely difficult to get an insight on the SINR expressions.
K
Therefore, they will be analyzed in the asymptotic case when N → ∞ and N → α upon
certain assumptions.

38
 Orthogonal Uplink CDMA: multipath channel

For user k, the model of the channel is given by

L−1
X
ck (τ ) = cpk φ(τ − τpk )
p=0

where φ is the transmit pulse filter.

The Fourier transform of c(τ ) after pulse matched filtering at the receiver is
PL−1 −j2πf τkp 2 1 if − W 2 ≤ f ≤
W
2
hk (f ) = c
p=0 pk e | Φ(f ) | where Φ(f ) =
0 otherwise
N −1
Sampling at frequencies f1 = − W2 , f2 = − W
2 +
1
NW, . . . , fN = − W
2 + N W, we
obtain the coefficients hk (i):

L−1
X −j2π i W τpk jπW τpk
hk (i) = hk (fi) = cpk e N e
p=0

39
 Orthogonal Uplink CDMA: Code Structure Model

Orthogonal Case:

• models the case of synchronized users.

• User codes wk are columns extracted from a Haar unitary (random) matrix V.
• The Haar distribution: For each deterministic unitary matrix Q, V and QV have the
same distribution.
1
• Generation: V = X(XH X) 2 where X has i.i.d complex Gaussian unit variance entries

i.i.d Case:

• models the asynchronous case.

• Entries of W are Gaussian i.i.d., with zero mean and variance 1/N .

Remark: Asymptotically (as N → ∞), MC-CDMA is equivalent to Direct-sequence

CDMA for a multipath channel of length L.

40
 Orthogonal Uplink CDMA: SINR with Matched Filter

K
Proposition: When N → ∞ and N → α, the SINR with Matched filter is:
W
1 R 2 |h (f )|2 df
W −W 1
Orthogonalcodes :SINRorth = 2
σ 2 +α(%−ξ1 )
W
1 R 2 |h (f )|2 df
W −W 1
i.i.d. codes :SINRiid = 2
σ 2 +α%

where
· ¸
RW
h i Ehk | 1
W
2
W h1(f )h∗k (f )df |2
2 − 2
% = Ehk | hk (f ) | and ξ1 =
1
RW
2 2
W − W | h1 (f ) | df
2

The asymptotic SINR depends only on a few meaningful parameters: α, σ 2, and the
distribution of the elements of H!

SINRiid is always inferior to SINRorth.

41
 Orthogonal Uplink CDMA: Simplifying assumptions

Channel model: We will suppose that for the channel model

1. The fading coefficients are i.i.d. Gaussian with

h i %
2
E [cpk ] = 0 and E | cpk | =
L

2. The delays are uniformly distributed according to the bandwidth

p
τpk =
W

42
 Orthogonal Uplink CDMA: Simplification of the
asymptotic expressions of the SINR

The SINR for the Matched filter in this case becomes

PL−1 2
orth p=0 |cp1 |
SINR = 2
σ +α%(1− 1 )
L
PL−1 2
iid p=0 |cp1 |
SINR = σ2+α%

As a consequence:

SINRorth σ 2 + α%
= 2 ¡ ¢
SINRiid σ + α% 1 − L1

When σ 2 → 0,

SINRorth L
→
SINRiid L−1

In a two-path channel, gain of 3 dB; in a 5-path channel, gain of less than 1 dB.

43
 Orthogonal Uplink CDMA: SINR with MMSE receiver

The proof is still under study.....

K
Proposition: When N → ∞ and N → α, the SINR of user 1 with MMSE receiver is given
by:
Z W
orth 1 2 1
SINR = ³ ´ df
W −W α σ2 α% SINRorth
2 + |h (f )|2 1 −
1+SINRorth 1 L SINRorth +1

Z W
i.i.d 1 2 1
SINR = df
W −W α σ2
2 1+SINRi.id + |h1 (f )|2

When L → ∞, both expressions are equal.

44
 Orthogonal Uplink CDMA: Simulations, ρ = 1,
SNR = 10dB, L = 1

Single User Gaussian Channel

Single User Rayleigh fading Channel

4
Orth MF/MMSE/Optimum Random Optimum
Spectral Efficiency

2
Random MMSE

Random MF
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Ratio Number of users/Spreading length

In a one-path channel, orthogonality gain is maximal.

45
 Orthogonal Uplink CDMA: Simulations, ρ = 1,
SNR = 10dB, L = 5

Single User Gaussian Channel

Single User Rayleigh fading Channel

4
Orth Optimum Random Optimum
Spectral Efficiency

Orth MMSE
3

Random MMSE
Orth MF

Random MF
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Ratio Number of users/Spreading length

As L increases, orthogonality gain decreases for any receiver.

46
 Is it useful to synchronize in uplink CDMA?

Spectral efficiency always increases with the use of orthogonal codes.

The orthogonality gain depends mainly on the number of paths and the load of the system.

As a consequence, adaptive synchronization protocols for future multiple access CDMA

schemes could be used to increase the rate.

To fully assess the gain, studies need to be conducted to determine the amount of
overhead signaling for a given number of users and bandwidth.

47
 Downlink CDMA: Frequency Selective fading model

M. Debbah, W. Hachem, P. Loubaton and M. de Courville, ”MMSE Analysis of Certain

Large Isometric Precoded Systems”, IEEE Transactions on Information Theory, Vol. 49,
May, No.5- 2003

y(n) = HΘs(n) + n(n)

• H : Toeplitz matrix N × N associated with the channel transfer function

XM
−k
h(z) = hk z .
k=0
• Θ = [w1, V], N × K matrix of codes for the different users.
• Without loss of generality, we focus our analysis on user 1:

³ ´−1
H H H H 2
SINR = w1 H HVV H + σ I Hw1

48
 Downlink CDMA: i.i.d case

MMSE receiver

µ ³ ´−1¶−1
H H 2 H
SINR = w1 VV + σ H H w1
µ ³ ´−1¶−1
1 H 2 H
→ Trace VV +σ H H
N

You remember the result of Bai and Silverstein?

 
Z 1
df
SINR =  
α σ2
0 +
SINR+1 |h(f )|2

49
 Downlink CDMA: orthogonal case

³ ´−1
H H H H 2
SINR = w1 H HVV H + σ I Hw1

Main difference between i.i.d. and isometric case :

H
¡ H H
¢
2 −1
w1 and H HVV H + σ I H are not independent.

You remember the result in the orthogonal case?

H
¡ H H
¢
2 −1
Let A = H HVV H + σ I H

H 1 H
w1 Aw1 → tr((I − VV )A)
N −K

¡ ¢−1
The idea (not detailed here) is to show that (I − VVH )HH HVVH HH + σ 2I H can
be approximated by a non-commutative polynomial of VVH and HH H and therefore one
can show that the SINR converges almost surely to a quantity β .

50
 Downlink CDMA: orthogonal case

β1
Other expression of β : One can rewrite η1 = 1+β1 such as

³ ´−1
H H H H 2
η1 = w1 H HΘΘ H + σ I Hw1

K
1 X k
η = lim η
K→∞ K
k=1

1 ³ H H 2 −1 H H
´
η = lim tr (HΘΘ H + σ I) HΘΘ H
K→∞ K

N Z
1 1 X λi 1 λ
= lim = lim dρN (λ)
α N →∞ N i=1 λi + σ 2 α N →∞ λ + σ2

ρN is the empirical eigenvalue distribution of HΘΘH HH

51
 Downlink CDMA: orthogonal case

Previous results in the course show that ρN converges almost surely to a compactly
supported measure, which can be calculated explicitly.

β 1
R t
Therefore, 1+β
=η= α t+σ 2
PΘTΘH X(t)dt = − α1 ΦΘTΘH X(− σ12 )

52
 Application

Context: X = HHH , T = Diag(IK , 0) and ΘTΘH , Θ Haar distributed N × N random

unitary matrix, with respective empirical eigenvalue distribution µN and νN .

dµN (t) → dµ(t) = p(t)dt

νN (t) → ν(t) = αδ(t − 1) + (1 − α)δ(t − 0)

αz
ΦΘTΘH =
1−z
z
ΞΘTΘH =
α+z

1+z
ΞΘTΘH X = ΞX(z)
α+z

53
 Application

Now, we have:

¡ ¢ 1 + ΦΘTΘH X(z)
z = ΞΘTΘH X ΦΘTΘH X (z) = ΞX(ΦΘTΘH X(z))
α + ΦΘTΘH X(z)

à !
α + ΦΘTΘH X(z)
ΦΘTΘH X(z) = ΦX α
1 + ΦΘTΘH X(z)

Since
1 αβ
ΦΘTΘH X(− ) = −
σ2 β+1
one gets immediately the following result:

 
Z 1
df
β=  
α σ2 β
0
β+1 + |h(f )|2
(1 − α (1+β) )

54
 Downlink CDMA: comparison

• i.i.d codes:  
Z 1
df
SINR =  
α σ2
0 +
SINR+1 |h(f )|2
• orthogonal codes:
 
Z 1
df
SINR =  
0 α
+ σ2
(1 − α SINR )
SINR+1 |h(f )|2 (1+SINR)

• Conclusion: For a target SINR, performance of isometric precoded system with noise
variance σ 2 equivalent to i.i.d. one with noise variance (1 − α SINR )σ 2
1+SINR

55
 Downlink CDMA

Spectral efficiency of different systems with Eb/N0=10dB

6

MMSE Haar−precoded
SIC Haar−precoded
4
spectral efficiency bit/s/Hz

MMSE i.i.d−precoded
SIC i.i.d−precoded
Gaussian channel
single user bound
limit MMSE i.i.d
3

0
0 1 2 3 4 5 6 7 8
alpha

• For low loaded systems, non-linear (very complex) receivers do not make sense.
• Orthogonal codes impact the performance only at high loads.

56
 Last Slide

THANK YOU!

57