上:Advances in Wireless Communication 课堂笔记(上)
中:Advances in Wireless Communication 课堂笔记(中)
Advanced modulation techniques, Iterative detection and decoding
Modulation in 5G
Mainly uses QAM schemes
Going back to structure of PDSCH for an M-ary modulation (M-QAM), the encoded interleaved rate-matched scrambled bits are splited into chunck of m bits and the mapped to the complex modulated symbols
Modulation coding scheme MCS
Modulation scheme | Modulation order |
---|---|
QPSK | 2 |
16-QAM | 4 |
64-QAM | 6 |
256-QAM | 8 |
release 17:1024-QAM | 10 |
MCS table between 0 and 31 (5 bits representation)
MCS index | Modulation order | code rate (*1024) | spectral efficiency |
---|---|---|---|
0 | 2 | 120(lowest rate=120/1024) | 0.234 |
… | … | … | … |
27 | 8 | 948 | 7.4 |
Bounds on symbol error probability
M-ary modulation with signal set {
S
1
(
t
)
,
S
2
(
t
)
,
.
.
.
,
S
n
(
t
)
S_1(t),S_2(t),...,S_n(t)
S1(t),S2(t),...,Sn(t)} Let
E
i
j
E_{ij}
Eij be the event that
P
(
y
(
t
)
∣
S
j
(
t
)
)
>
P
(
y
(
t
)
∣
S
i
(
t
)
)
P
(
e
r
r
o
r
∣
S
i
(
t
)
)
=
P
(
⋃
j
=
1
,
j
≠
i
M
E
i
j
)
P
(
e
r
r
o
r
∣
S
i
(
t
)
)
<
∑
j
=
1
,
j
≠
i
M
P
(
E
i
j
)
⇒
∑
j
=
1
,
j
≠
i
M
Q
(
D
i
j
/
2
N
o
)
≤
(
M
−
1
)
Q
(
D
m
i
n
(
i
)
2
N
o
)
P(y(t)|S_j(t))>P(y(t)|S_i(t))\\ P(error|S_i(t))=P(\bigcup ^M_{j=1,j\ne i}E_{ij})\\ P(error|S_i(t))<\sum ^M_{j=1,j\ne i}P(E_{ij})\\ \Rightarrow \sum^M_{j=1,j\ne i}Q(D_{ij}/\sqrt{2N_o})\le(M-1)Q(\frac {D_{min}(i)}{\sqrt{2N_o}})
P(y(t)∣Sj(t))>P(y(t)∣Si(t))P(error∣Si(t))=P(j=1,j=i⋃MEij)P(error∣Si(t))<j=1,j=i∑MP(Eij)⇒j=1,j=i∑MQ(Dij/2No)≤(M−1)Q(2NoDmin(i))
Union bound
P
(
⋃
i
=
1
n
A
i
≤
∑
i
=
1
n
P
(
A
i
)
)
P(\bigcup_{i=1}^nA_i\le\sum^n_{i=1}P(A_i))
P(⋃i=1nAi≤∑i=1nP(Ai))
D
m
i
n
(
i
)
D_{min}(i)
Dmin(i)
D m i n ( i ) = m i n D i j 1 ≤ j ≤ m , j ≠ i D_{min}(i)=min D_{ij}\qquad 1\le j\le m,j\ne i Dmin(i)=minDij1≤j≤m,j=i
AWGN:
P
(
E
i
j
)
=
P
(
n
≥
D
i
j
2
)
=
Q
(
D
i
j
2
N
o
)
P(E_{ij})=P(n\ge\frac {D_{ij}}2)=Q(\frac {D_{ij}}{\sqrt{2N_o}})
P(Eij)=P(n≥2Dij)=Q(2NoDij)
n: one component of Gaussian noise~
N
(
0
,
N
o
2
)
\mathcal N(0,\frac {N_o}2)
N(0,2No)
Also for linear bound
P
(
e
r
r
o
r
∣
S
i
(
t
)
)
≥
m
a
x
P
(
E
i
j
)
(
i
≠
j
)
=
Q
(
D
m
i
n
(
i
)
/
2
N
o
)
P(error|S_i(t))\ge maxP(E_{ij})\;(i\ne j)=Q(D_{min}(i)/\sqrt{2N_o})
P(error∣Si(t))≥maxP(Eij)(i=j)=Q(Dmin(i)/2No)
For the average probability of error
P
a
v
(
e
r
r
o
r
)
=
1
M
∑
i
=
1
m
P
(
e
r
r
o
r
∣
S
i
(
t
)
)
≤
(
M
−
1
)
Q
(
D
m
i
n
2
N
o
)
P_{av}(error)=\frac1M \sum^m_{i=1}P(error|S_i(t))\le (M-1)Q(\frac {D_{min}}{\sqrt{2N_o}})
Pav(error)=M1∑i=1mP(error∣Si(t))≤(M−1)Q(2NoDmin)
Bit-level demodulation
Example:
QPSK symbol(
b
1
,
b
2
b_1,b_2
b1,b2), received
y
(
y
1
,
y
2
)
y(y_1, y_2)
y(y1,y2)
y
=
s
i
+
(
n
1
,
n
2
)
y=s_i+(n_1,n_2)
y=si+(n1,n2)
gray-labling
ML decision for
b
1
b_1
b1
b
1
=
0
b_1=0
b1=0 if
P
(
y
∣
b
1
=
0
)
>
P
(
y
∣
b
1
=
1
)
P(y|b_1=0)>P(y|b_1=1)
P(y∣b1=0)>P(y∣b1=1)
P
(
y
∣
b
1
=
0
,
b
2
=
0
)
⏟
S
1
P
(
b
2
=
0
)
⏟
1
2
+
P
(
y
∣
b
1
=
0
,
b
2
=
1
)
⏟
S
2
P
(
b
2
=
1
)
⏟
1
2
\underbrace{P(y|b_1=0,b_2=0)}_{S_1}\underbrace{P(b_2=0)}_\frac 12+\underbrace{P(y|b_1=0,b_2=1)}_{S_2}\underbrace{P(b_2=1)}_{\frac 12}
S1
P(y∣b1=0,b2=0)21
P(b2=0)+S2
P(y∣b1=0,b2=1)21
P(b2=1)
Similarly
P
(
y
∣
b
1
=
1
)
=
1
2
P
(
y
∣
S
3
)
+
1
2
P
(
y
∣
S
4
)
P(y|b_1=1)=\frac12P(y|S_3)+\frac12P(y|S_4)
P(y∣b1=1)=21P(y∣S3)+21P(y∣S4)
Decide
b
1
=
0
b_1=0
b1=0 if
y
2
≥
0
y_2\ge0
y2≥0
Decide
b
2
=
0
b_2=0
b2=0 if
y
1
≥
0
y_1\ge0
y1≥0
P(error in
b
1
b_1
b1)=
P
(
n
2
≤
−
1
)
=
Q
(
1
σ
)
P(n_2\le-1)=Q(\frac1{\sigma})
P(n2≤−1)=Q(σ1)
this is the optimal labeling rule
An alternative which is worse:
Suppose
S
1
S_1
S1 is transmitted
P
(
c
o
r
r
e
c
t
b
2
∣
S
1
)
=
(
1
−
Q
(
1
σ
)
)
2
+
Q
(
1
σ
)
2
=
1
−
2
Q
(
1
σ
)
+
2
Q
(
1
σ
)
2
P(correct \;b_2|S_1)=(1-Q(\frac1{\sigma}))^2+Q(\frac 1{\sigma})^2=1-2Q(\frac1{\sigma})+2Q(\frac1{\sigma})^2
P(correctb2∣S1)=(1−Q(σ1))2+Q(σ1)2=1−2Q(σ1)+2Q(σ1)2
P(error in
b
2
)
b_2)
b2)=
2
Q
(
1
σ
)
−
2
Q
(
1
σ
)
2
>
Q
(
1
σ
)
⇒
Q
(
1
σ
)
>
2
Q
(
1
σ
)
2
⇔
1
2
>
Q
(
1
σ
)
2Q(\frac1{\sigma})-2Q(\frac1{\sigma})^2>Q(\frac1{\sigma})\\\Rightarrow Q(\frac1{\sigma})>2Q(\frac1{\sigma})^2\\\Leftrightarrow\frac12>Q(\frac1{\sigma})
2Q(σ1)−2Q(σ1)2>Q(σ1)⇒Q(σ1)>2Q(σ1)2⇔21>Q(σ1)
In general for M_PSK gray labeling is an ording if binary strings on the circle, such that two successive values differ in only on bit
8PSK gray coding and 16PSK gray coding
8PSK:
16PSK
liklihood ratios for bit
b
i
b_i
bi:
l
o
g
P
(
y
∣
b
i
=
0
)
P
(
y
∣
b
i
=
1
)
log\frac {P(y|b_i=0)}{P(y|b_i=1)}
logP(y∣bi=1)P(y∣bi=0)
this is LLR for bit
b
i
b_i
bi, which will be passed to the decoder
Bit-Interleaved coded Modulation (BICM)
16 QAM with gray mapping:
consider a symbol
x
=
(
x
1
,
x
2
,
x
3
,
x
4
)
,
x
1
x=(x_1,x_2,x_3,x_4),x_1
x=(x1,x2,x3,x4),x1and
x
3
x_3
x3 observe identical and independence channels
The bits
x
2
x_2
x2 and
x
4
x_4
x4 also observe indentical an independence channels
In general BICM with M-ary modulation
M
=
2
m
M=2^m
M=2m
can be decomposed into m parallel channels
For 16-QAM, two types of channel, two from each type
x
1
x_1
x1 &
x
2
x_2
x2,
x
3
x_3
x3 &
x
4
x_4
x4 correlation
Coded modulation capacity and BICM capacity
channel capacity
c
=
c=
c=
s
u
p
P
x
{sup} \atop {P_x}
Pxsup
I
(
x
;
y
)
I(x;y)
I(x;y)
P
x
P_x
Px: input distribution
For AWGN with input power P, noise power N, bandwidth W C = 1 2 l o g ( 1 + P / W ) = 1 2 l o g ( 1 + S N R ) C=\frac12 log(1+P/W)=\frac12 log(1+SNR) C=21log(1+P/W)=21log(1+SNR)
Coded modulation capacity
I
(
X
,
Y
)
I(X,Y)
I(X,Y)
X
X
X belongs to the constellation set of modulation with uniform distribution
x
=
(
x
1
,
x
2
,
x
3
,
x
4
)
x=(x_1,x_2,x_3,x_4)
x=(x1,x2,x3,x4)
I
(
x
1
;
y
)
+
I
(
x
2
;
y
)
+
I
(
x
3
;
y
)
+
I
(
x
4
;
y
)
≤
I
(
x
1
,
x
2
,
x
3
,
x
4
;
y
)
I(x_1;y)+I(x_2;y)+I(x_3;y)+I(x_4;y)\le I(x_1,x_2,x_3,x_4;y)
I(x1;y)+I(x2;y)+I(x3;y)+I(x4;y)≤I(x1,x2,x3,x4;y)
For BICM, with M = 2 m M=2^m M=2m (size of the constellation) in parallel channels with inputs x 1 , x 2 , x 3 , . . . , x m x_1,x_2,x_3,...,x_m x1,x2,x3,...,xm, C B I C M = ∑ i = 1 m I ( x i ; y ) C_{BICM}=\sum^m_{i=1}I(x_i;y) CBICM=∑i=1mI(xi;y)
Coded modulation capacity and BICM capacity
BICM capacity (dash-dotted lines)
with Gray labeling, the difference between BICM capacity and coded modulation capacity is very small
Basics of classical multiple access techniques
Multiuser Communication
Two major scenarios
Broadcast Channel (downlink):
Multiple access channel (uplink):
Two early method of multiple access techniques
- FDMA (frequency division multiple access)
- TDMA (time division multiple access)
cross-talk or interference
3. implementation of filters at all possible frequency bands
4. crosstalk
guard band to reduce crosstalk
for TDMA, time synchronization is very challenging as the number of user grow
In both TDMA and FDMA, we allocate orthogonal resources either in time or frequency to users
S
1
(
t
)
S_1(t)
S1(t) and
S
2
(
t
)
S_2(t)
S2(t) are orthogonal:
∫
S
1
(
t
)
S
2
(
t
)
d
t
=
0
\int{S_1(t)S_2(t)}dt=0
∫S1(t)S2(t)dt=0
CDMA code division multiple access (CDMA)
In CDMA, each users is assigned a distinct signature sequence (or waveform) which the user employs to spread the signal
CDMA is a spread spectrum technique, where symbol rate is much smaller than the bandwidth
symbol rate is proportion to
1
T
\frac 1T
T1
These techniques provide frequency diversity against “selective fading”
selective fading:
one of the most challenging parts of CDMA systems is the design of equalizes
h
(
t
)
h(t)
h(t):h(0),h(1),…,h(d) — all channel gain
y
(
t
)
=
x
(
t
)
∗
h
(
t
)
,
y
[
n
]
=
x
[
n
]
∗
h
[
n
]
y(t)=x(t)*h(t),y[n]=x[n]*h[n]
y(t)=x(t)∗h(t),y[n]=x[n]∗h[n]
orthogonal frequency division multiplexing (OFDM)
X
0
,
X
1
,
.
.
.
,
X
N
−
1
X_0,X_1,...,X_{N-1}
X0,X1,...,XN−1 denote N complex symbols (in frequency domain)
Taking the IFFT
x
n
=
1
N
∑
k
=
0
N
−
1
X
k
e
j
2
π
n
k
N
f
o
r
n
=
0
,
1
,
.
.
.
,
N
−
1
x_n=\frac1{\sqrt N}\sum^{N-1}_{k=0}X_ke^{\frac{j2\pi nk}N}\qquad for \;n=0,1,...,N-1
xn=N1∑k=0N−1XkeNj2πnkforn=0,1,...,N−1
x
i
x_i
xi s are in the time domain
Then
x
0
,
x
1
,
.
.
.
,
x
N
−
1
x_0,x_1,...,x_{N-1}
x0,x1,...,xN−1 are passed through a D/A (digital to analog) converter resulting in the baseband OFDM signal
x
(
t
)
x(t)
x(t),
x
(
t
)
=
1
N
∑
k
=
0
N
−
1
X
k
e
j
2
π
k
t
T
N
0
≤
t
≤
T
N
x(t)=\frac1{\sqrt N}\sum^{N-1}_{k=0}X_ke^{\frac{j2\pi k_t}{T_N}}\quad0\le t\le T_N
x(t)=N1∑k=0N−1XkeTNj2πkt0≤t≤TN . Where
T
N
T_N
TN is the duration of OFDM signal.
The sub-carrier frequencies are
f
i
=
i
/
T
N
,
i
=
0
,
1
,
.
.
.
,
N
−
1
f_i=i/T_N,\;i=0,1,...,N-1
fi=i/TN,i=0,1,...,N−1 and the samples
x
0
,
x
1
,
.
.
.
,
x
N
−
1
x_0,x_1,...,x_{N-1}
x0,x1,...,xN−1 represent samples of
x
(
t
)
x(t)
x(t) every
T
N
/
N
T_N/N
TN/N second
The block diagram of an OFDM system
A major advantage of OFDM comparing with CDMA sis its resilience to ISI(inter-symbol interference) because the data rate on each sub-channel is very low ( B / N < < B c B/N<<B_c B/N<<Bc, coherence bandwidth), B c B_c Bc proportional to 1 / D 1/D 1/D, where D is delay spread. Channel equalization is much simpler since channel gain can be consider flat at each sub-channel
Non-orthogonal multiple access techniques
Cyclic prefix for OFDM:
x 0 , x 1 , . . . x N − 1 ⏟ data symbols in time domain I F F T ← X 0 , X 1 , . . . , X N − 1 ⏟ data symbols in frequency domain y [ n ] = x [ n ] ∗ c [ n ] = ∑ k = n − μ + 1 n x [ k ] c [ n − k ] \underbrace {x_0,x_1,...x_{N-1}}_{\text{data symbols in time domain}} {IFFT \atop \leftarrow} \underbrace{X_0,X_1,...,X_{N-1}}_{\text{data symbols in frequency domain}} \\ y[n]=x[n]*c[n]=\sum^n_{k=n-\mu+1}x[k]c[n-k] data symbols in time domain x0,x1,...xN−1←IFFTdata symbols in frequency domain X0,X1,...,XN−1y[n]=x[n]∗c[n]=k=n−μ+1∑nx[k]c[n−k]
channel delay spread is at most
μ
T
N
/
N
\mu T_N/N
μTN/N
In the frequency domain, see
X
⋅
C
X\cdot C
X⋅C
If we consider the convolution of
x
N
−
μ
,
.
.
.
,
x
N
−
1
,
x
0
,
x
1
,
.
.
.
.
x
N
−
1
{x_{N-\mu},..., x_{N-1},x_0,x_1,....x_{N-1}}
xN−μ,...,xN−1,x0,x1,....xN−1 with
c
0
,
c
1
,
.
.
.
,
c
μ
{c_0,c_1,...,c_{\mu}}
c0,c1,...,cμ and remove the last
μ
\mu
μ samples, we get a linear convolution which is equivalent to circular convolution
Therefore in the frequency domain, we have
x
^
k
=
c
k
×
x
k
\hat{x}_k=c_k\times x_k
x^k=ck×xk
c
k
c_k
ck: FFT of the sequence (
c
0
,
c
1
.
.
.
c
μ
⏟
μ
+
1
,
0
,
0
,
.
.
.
⏟
N
−
μ
+
1
\underbrace{c_0,c_1...c_\mu}_{\mu+1},\underbrace{0,0,...}_{N-\mu+1}
μ+1
c0,c1...cμ,N−μ+1
0,0,...)
Therefore the effect of channel, c ( t ) c(t) c(t) can be completely removed by a simple equalization in the frequency domain
Uplink AWGN channel
consider the discrete time model with two users
y
=
x
1
+
x
2
+
n
,
n
∼
N
(
0
,
N
0
)
y=x_1+x_2+n,\;n\sim \mathcal{N}(0,N_0)
y=x1+x2+n,n∼N(0,N0) users have powewr constraint
user K power is
P
K
P_K
PK
R
1
<
1
2
l
o
g
(
1
+
P
1
/
N
o
)
R_1<\frac 12 log(1+P_1/N_o)
R1<21log(1+P1/No)— maximum capacity user 1 can achieve,
R
1
R_1
R1 : rate for user 1 (
R
1
+
R
2
R_1+R_2
R1+R2) sum rate
user 1 power
P
1
P_1
P1 user 2 power
P
2
P_2
P2
Capacity region pairs of achievable
(
R
1
,
R
2
)
(R_1,R_2)
(R1,R2)
{user1-
P
1
⨁
P_1\bigoplus
P1⨁ user2-
P
2
P_2
P2} — superuser of power
P
1
+
P
2
P_1+P_2
P1+P2 and rate
R
1
+
R
2
R_1+R_2
R1+R2
R
1
+
R
2
<
1
2
l
o
g
(
1
+
P
1
+
P
2
N
0
)
R_1+R_2<\frac12log(1+\frac{P_1+P_2}{N_0})
R1+R2<21log(1+N0P1+P2)
How to achieve corner points A B
SIC: successive interference cancellation to achieve A. Decode
x
1
x_1
x1 assuming
x
2
x_2
x2 is noise. Cancel out
x
1
x_1
x1 and decode
x
2
x_2
x2, assuming
x
1
x_1
x1 is known and cancelled.
what rate
R
1
R_1
R1 we can get?
R
1
<
1
2
l
o
g
(
1
+
P
1
P
2
+
N
o
)
R_1<\frac 12log(1+\frac{P_1}{P_2+N_o})
R1<21log(1+P2+NoP1)
l
o
g
(
1
+
P
1
P
2
+
N
o
)
+
l
o
g
(
1
+
P
2
N
o
)
=
l
o
g
(
1
+
P
1
+
P
2
N
o
)
log(1+\frac{P_1}{P_2+N_o})+log(1+\frac{P_2}{N_o})=log(1+\frac{P_1+P_2}{N_o})
log(1+P2+NoP1)+log(1+NoP2)=log(1+NoP1+P2), so we can achieve A, B
How can we achieve power between A B ?
time sharing.
λ
\lambda
λ fraction of time operate at A. (
1
−
λ
1-\lambda
1−λ) fraction of time operate at A
achieve
λ
A
+
(
1
−
λ
)
B
\lambda A+(1-\lambda)B
λA+(1−λ)B
Any convex combination of A and B is achievable
The difference with an orthogonal method (FDMA or OFDM) for instance,
α
\alpha
α fraction of Bandwidth is given to user one and (
1
−
α
1-\alpha
1−α) fraction is given to user 2
C
=
W
l
o
g
P
N
o
W
C=Wlog\frac{P}{N_oW}
C=WlogNoWP (bit/second)
C
1
=
α
W
l
o
g
P
1
α
N
o
W
C_1=\alpha Wlog\frac{P_1}{\alpha N_o W}
C1=αWlogαNoWP1
C
2
=
(
1
−
α
)
W
l
o
g
P
1
(
1
−
α
)
N
o
W
C_2=(1-\alpha) Wlog\frac{P_1}{(1-\alpha) N_o W}
C2=(1−α)Wlog(1−α)NoWP1
in the discrete domain
R
1
=
α
2
l
o
g
P
1
α
N
o
R_1=\frac{\alpha}2 log\frac{P_1}{\alpha N_o }
R1=2αlogαNoP1
R
2
=
(
1
−
α
)
2
l
o
g
P
1
(
1
−
α
)
N
o
R_2=\frac{(1-\alpha)}2 log\frac{P_1}{(1-\alpha) N_o }
R2=2(1−α)log(1−α)NoP1
Extension to K-user uplink
y
=
x
1
+
x
2
+
.
.
.
+
x
k
+
n
y=x_1+x_2+...+x_k+n
y=x1+x2+...+xk+n
power of user i is
P
i
P_i
Pi
we can form a superuser for any subset
S
⊂
{
1
,
.
.
.
,
k
}
S\subset \{1,...,k\}
S⊂{1,...,k} of user. For any
S
⊂
{
1
,
.
.
.
,
k
}
,
∑
i
∈
S
R
i
<
1
2
l
o
g
(
1
+
∑
i
∈
s
P
i
N
o
)
S\subset \{1,...,k\},\;\sum_{i\in S}R_i<\frac12log(1+\sum_{i\in s}\frac{P_i}{N_o})
S⊂{1,...,k},∑i∈SRi<21log(1+∑i∈sNoPi)(
R
i
R_i
Ri: rate for user i).
2
k
−
1
2^{k}-1
2k−1 constrains must be satisfied
R
1
<
1
2
l
o
g
(
1
+
P
1
N
o
)
R
2
<
1
2
l
o
g
(
1
+
P
2
N
o
)
R
3
<
1
2
l
o
g
(
1
+
P
3
N
o
)
R
1
+
R
2
R
1
+
R
3
R
2
+
R
3
R
1
+
R
2
+
R
3
R_1<\frac12log(1+\frac{P_1}{N_o})\\ R_2<\frac12log(1+\frac{P_2}{N_o})\\ R_3<\frac12log(1+\frac{P_3}{N_o})\\ R_1+R_2\\ R_1+R_3\\ R_2+R_3\\ R_1+R_2+R_3
R1<21log(1+NoP1)R2<21log(1+NoP2)R3<21log(1+NoP3)R1+R2R1+R3R2+R3R1+R2+R3
6(=3!) corner point correspond
to one of the 6 possible ordering of user 1 2 3
SIC decode user 1 assuming
x
2
,
x
3
x_2, x_3
x2,x3 are noise, decode
x
2
x_2
x2 give
x
1
x_1
x1(cancel out) and assuming
x
3
x_3
x3 as noise and finally decode
x
3
x_3
x3 given
x
1
x_1
x1 and
x
2
x_2
x2(cancel out)
R
1
<
1
2
l
o
g
(
1
+
P
1
P
2
+
P
3
+
N
o
)
R
2
<
1
2
l
o
g
(
1
+
P
2
P
3
+
N
o
)
R
2
<
1
2
l
o
g
(
1
+
P
3
N
o
)
R_1<\frac 12log(1+\frac{P_1}{P_2+P_3+N_o})\\ R_2<\frac 12log(1+\frac{P_2}{P_3+N_o})\\ R_2<\frac 12log(1+\frac{P_3}{N_o})
R1<21log(1+P2+P3+NoP1)R2<21log(1+P3+NoP2)R2<21log(1+NoP3)
The sum-capacity(maximum of Sum-rate) is given by
C
s
u
m
=
1
2
l
o
g
(
1
+
∑
k
=
1
k
P
k
N
o
)
C_{sum}=\frac12log(1+\frac{\sum_{k=1}^kP_k}{N_o})
Csum=21log(1+No∑k=1kPk)
If the users powers are equal then sum-capacity is the same for orghogonal and non-orthogonal cases
Downlink AWGN channel
y 1 = h 1 x + n 1 y 2 = h 2 x + n 2 n 1 & n 2 ∼ N ( 0 , N o ) ∣ h 1 ∣ ≤ ∣ h 2 ∣ y_1=h_1x+n_1\\ y_2=h_2x+n_2\\ n_1 \& n_2\sim \mathcal N(0,N_o)\\ |h_1|\le|h_2| y1=h1x+n1y2=h2x+n2n1&n2∼N(0,No)∣h1∣≤∣h2∣
Superposition coding, interference cancellation methods
strategy super position coding
x
=
x
1
(
signal intended for user 1
)
+
x
2
(
signal intended for user 2
)
x=x_1(\text{signal intended for user 1})+x_2(\text{signal intended for user 2})
x=x1(signal intended for user 1)+x2(signal intended for user 2)
user 1 treats
x
2
x_2
x2 as noise
R
1
<
1
2
l
o
g
(
1
+
P
1
∣
h
1
∣
2
P
2
∣
h
1
∣
2
+
N
o
)
R_1<\frac12log(1+\frac{P_1|h_1|^2}{P_2|h_1|^2+N_o})
R1<21log(1+P2∣h1∣2+NoP1∣h1∣2)
user 2 decodes
x
1
x_1
x1 and cancel it out then decode
x
2
x_2
x2,
R
2
<
1
2
l
o
g
(
1
+
P
2
∣
h
2
∣
2
N
o
R2<\frac12log(1+\frac{P_2|h_2|^2}{N_o}
R2<21log(1+NoP2∣h2∣2
visual super position coding(user 1 and user 2 both use QPSK):
for the orthogonal case, suppose a fraction of bandwidth is user 1 and
1
−
α
1-\alpha
1−α fraction is allocated for user2
rate of user 1:
R
1
=
α
2
l
o
g
(
1
+
P
1
∣
h
1
∣
2
α
N
o
)
R_1=\frac{\alpha}2log(1+\frac{P_1|h_1|^2}{\alpha N_o})
R1=2αlog(1+αNoP1∣h1∣2)
rate of user 2:
1
−
α
2
l
o
g
(
1
+
P
2
∣
h
2
∣
2
(
1
−
α
)
N
o
)
\frac{1-\alpha}2log(1+\frac{P_2|h_2|^2}{(1-\alpha) N_o})
21−αlog(1+(1−α)NoP2∣h2∣2)
In general, with the ordering
∣
h
1
∣
≤
∣
h
2
∣
≤
.
.
.
≤
∣
h
k
∣
|h_1|\le|h_2|\le...\le|h_k|
∣h1∣≤∣h2∣≤...≤∣hk∣
x
=
x
1
+
x
2
+
.
.
.
+
x
k
x=x_1+x_2+...+x_k
x=x1+x2+...+xk
Then the boundary of the achievable capacity region of the k-user downlink AWGN is given by the parameterized rate tuple
R
k
=
1
2
l
o
g
(
1
+
P
k
∣
h
k
∣
2
N
o
+
(
∑
j
=
1
k
P
j
)
∣
h
k
∣
2
)
k
=
1
,
2
,
.
.
.
,
K
R_k=\frac12log(1+\frac{P_k|h_k|^2}{N_o+(\sum_{j=1}^kP_j)|h_k|^2})\qquad k=1,2,...,K
Rk=21log(1+No+(∑j=1kPj)∣hk∣2Pk∣hk∣2)k=1,2,...,K
where
P
=
∑
j
=
1
k
P
j
P=\sum_{j=1}^kP_j
P=∑j=1kPj is the power splits among the users
Multiuser fading channels
The single user case:
Two types of fading
{
1) slow fading
2) fast fading
\begin{cases} \text{1) slow fading} \\ \text{2) fast fading} \end{cases}
{1) slow fading2) fast fading
generic model for fading
y
=
h
x
+
n
,
h
:
channel gain
y=hx+n,\;h:\text{channel gain}
y=hx+n,h:channel gain
typical model for the distribution of h h h is the Rayleigh distribution
The received SNR:
∣
h
∣
2
S
N
R
−
(
t
r
a
n
s
m
i
t
S
N
R
)
|h|^2SNR-(transmit\; SNR)
∣h∣2SNR−(transmitSNR)
R: transmission Rate
Communication is successful is
R
<
1
2
l
o
g
(
1
+
∣
h
∣
2
S
N
R
)
R<\frac 12log(1+|h|^2SNR)
R<21log(1+∣h∣2SNR)
probability of outage: P o u t ( R ) = P { 1 2 l o g ( 1 + ∣ h ∣ 2 S N R ) ≤ R } P_{out}(R)=P\{\frac 12log(1+|h|^2SNR)\le R\} Pout(R)=P{21log(1+∣h∣2SNR)≤R} — depends on the distribution of h
For example: for Rayleigh fading channel(pdf of
h
=
h
2
e
−
h
2
2
δ
2
,
δ
h=\frac h2e^{-\frac {h^2}{2\delta^2}},\delta
h=2he−2δ2h2,δ is the parameter of fading),
P
o
u
t
(
R
)
=
1
−
e
x
p
(
−
2
R
−
1
S
N
R
)
P_{out}(R)=1-exp(-\frac{2^R-1}{SNR})
Pout(R)=1−exp(−SNR2R−1)
at high SNR,
P
o
u
t
(
R
)
∼
2
R
−
1
S
N
R
P_{out}(R)\sim \frac{2^R-1}{SNR}
Pout(R)∼SNR2R−1(decays out
1
S
N
R
\frac1{SNR}
SNR1)
Notions of capacity:
ε
\varepsilon
ε — outage capacity
P
o
u
t
(
R
)
≤
ε
P_{out}(R)\le \varepsilon
Pout(R)≤ε
Let F be the cdf of
∣
h
∣
2
|h|^2
∣h∣2, then outage =
1
2
l
o
g
(
1
+
F
−
1
(
1
−
ε
)
S
N
R
)
\frac12log(1+F^{-1}(1-\varepsilon)SNR)
21log(1+F−1(1−ε)SNR)
Another notion of capacity ergodic capacity
∫
n
1
2
l
o
g
(
1
+
∣
h
∣
2
S
N
R
)
f
H
(
h
)
d
h
\int_n\frac12log(1+|h|^2SNR)f_H(h)dh
∫n21log(1+∣h∣2SNR)fH(h)dh
f
H
(
h
)
f_H(h)
fH(h): pdf of h
distribution of h across time is f H ( h ) f_H(h) fH(h) ergodic capacity: average of capacity across time as the # of time-frames grows large
Slow fading: the communication delay (time frame) is relatively short compared with channel coherence time
In fast fading:
y
=
h
x
+
n
y=hx+n
y=hx+n
h
:
h:
h: i.i.d for each symbol transmission
for transmitting L symbols, outage probability P o u t ( R ) = P { 1 L ∑ l = 1 L 1 2 l o g ( 1 + ∣ h l ∣ 2 S N R ) < R } P_{out}(R)=P\{\frac1L\sum_{l=1}^L\frac12log(1+|h_l|^2SNR)<R\} Pout(R)=P{L1∑l=1L21log(1+∣hl∣2SNR)<R}
uplink fading channel:
slow fading, suppose we have k users
P
o
u
t
u
p
l
i
n
k
(
R
)
=
P
{
1
2
l
o
g
(
1
+
S
N
R
∑
k
∈
S
∣
h
k
∣
2
)
<
∣
s
∣
R
f
o
r
s
o
m
e
S
C
{
1
,
2
,
.
.
.
,
k
}
}
P_{out}^{uplink}(R)=P\{\frac12log(1+SNR\sum_{k\in S}|h_k|^2)<|s|R \;for\; some \;SC \;\{1,2,...,k\}\}
Poutuplink(R)=P{21log(1+SNRk∈S∑∣hk∣2)<∣s∣RforsomeSC{1,2,...,k}}
suppose all users are transmitting at rate
R
R
R
At low SNR, this can be approximates
P
o
u
t
u
p
l
i
n
k
=
P
{
1
2
∣
h
k
∣
2
S
N
R
<
R
f
o
r
s
o
m
e
k
∈
{
1
,
2
,
.
.
.
,
k
}
}
≈
k
P
o
u
t
P_{out}^{uplink}=P\{\frac12|h_k|^2SNR<R \;for\; some \;k\in\{1,2,...,k\}\}\approx kP_{out}
Poutuplink=P{21∣hk∣2SNR<Rforsomek∈{1,2,...,k}}≈kPout
Ergodic capacity (for both slow and fast fading)
C
s
u
m
=
E
[
1
2
l
o
g
(
1
+
∑
k
=
1
k
∣
h
k
∣
2
P
)
]
∑
k
=
1
k
∣
h
k
∣
2
P
:
S
N
R
,
N
o
R
1
+
R
2
+
.
.
.
+
R
k
<
1
2
l
o
g
(
1
+
∑
k
=
1
k
∣
h
k
∣
P
o
)
∑
k
=
1
k
∣
h
k
∣
P
o
:
e
f
f
e
c
t
i
v
e
W
R
o
f
t
h
e
s
u
p
e
r
u
s
e
r
C_{sum}=E[\frac12log(1+\sum_{k=1}^k|h_k|^2P)]\\ \sum_{k=1}^k|h_k|^2P:SNR,N_o\\ R_1+R_2+...+R_k<\frac12log(1+\sum_{k=1}^k|h_k|P_o)\\ \sum_{k=1}^k|h_k|P_o:effective WR of the superuser
Csum=E[21log(1+k=1∑k∣hk∣2P)]k=1∑k∣hk∣2P:SNR,NoR1+R2+...+Rk<21log(1+k=1∑k∣hk∣Po)k=1∑k∣hk∣Po:effectiveWRofthesuperuser
How does this compose with the sum-capacity of uplink without fading?
(Assuming
E
[
∣
h
k
∣
2
]
=
1
E[|h_k|^2]=1
E[∣hk∣2]=1)
By Jensen’s inequality:
E
[
l
o
g
(
1
+
∑
k
=
1
k
∣
h
k
∣
2
P
N
o
)
]
≤
l
o
g
(
1
+
∑
k
=
1
k
∣
h
k
∣
2
P
N
o
)
=
l
o
g
(
1
+
k
P
N
o
)
E[log(1+\frac{\sum_{k=1}^k|h_k|^2P}{N_o})]\le log(1+\frac{\sum_{k=1}^k|h_k|^2P}{N_o})=log(1+\frac{kP}{N_o})
E[log(1+No∑k=1k∣hk∣2P)]≤log(1+No∑k=1k∣hk∣2P)=log(1+NokP)
l
o
g
(
1
+
k
P
N
o
)
log(1+\frac{kP}{N_o})
log(1+NokP): sum-capacity without fading