Title: Multiple Access Channels with Correlated Sources and Feedback Sufficient Conditions for SourceChanne
1Multiple Access Channels with Correlated Sources
and Feedback(Sufficient Conditions for
Source-Channel Separation)
- Sriram Vishwanath
- in collaboration with
- W. Wu, B. Smith and S. Sridharan
- University of Texas at Austin
2Overview
- Single user
- MAC with correlated sources tough problem
- MAC with feedback tough problem
- Why combine the two? easier problem (perhaps)
- Source channel separation for MAC channels
- Source channel separation in general networks
- Conclusions
3Single User DMC
Rx
Source and Tx
- Source-channel separation holds
- Feedback leads to no increase in capacity
4Source-Channel Separation
- Source-channel separation inspires many of our
current architectures - Pros (Big)
- Allows for coding at separate layers
- Complexity reduction
- Caution an asymptotic result not true for
systems with stringent block length constraints
5Role of Feedback
- Apparently useless
- It can, in many cases (e.g. Gaussian) greatly
reduce - Coding complexity
- Code latency
6MAC Channel
Tx1
Rx
Tx2
- In general
- NO source channel separation
- Feedback increases capacity region
7Capacity of MAC with Feedback
- In general an open problem
- Two user Gaussian case Ozarow
- Result Capacity region is the cutset outer bound
- R1 lt I(X1YX2)
- R2 lt I(X2YX1)
- R1 R2 lt I(X1,X2Y)
- over all cdfs F(x1,x2)
8Gaussian MAC with feedback
- Gaussian inputs extremize the outer bound
- R1 lt log(1 (1-?2)P1)
- R2 lt log(1 (1-?2)P2)
- R1 R2 lt log(1 P1 P2 2?vP1P2)
- Achieved by modified Schalkwijk-Kailath Coding
Scheme - Feedback
- Increases region
- Simplifies coding
- Reduces latency
9One more gain from feedback..
- Feedback induces source-channel separation
10MAC channel with correlated sources
U
Tx1
Rx
Tx2
V
- U and V correlated and no feedback
- Well known that there is no source channel
separation Cover, El Gamal Salehi , Gastpar,
Dueck
11MAC channel with correlated sources (and no
feedback)
- Finding actual capacity region nearly impossible
- Only extreme cases, like UV
- UV results in a MISO capacity gain
- Even if U and V are 99.9 correlated, capacity
region a very hard problem (!) - Isolating even 1 bit of common information very
difficult Gacs Korner, Ahlswede Csiszar
12The meat of the talk
- For two-user Gaussian MAC with feedback source
channel separation holds - For a general MAC with feedback if the
correlation between U and V is greater than a
threshold value then source channel separation
holds - Proof in all cases Fanos inequality
13Two user Gaussian MAC with Feedback Correlation
- Achievability Separate Coding
Tx1 SK
binning SW
U
Rx SK Decoder
SW decoder
Tx2 SK
binning SW
V
14Two user Gaussian MAC with Feedback
- By separate source channel coding (binning with
Ozarow coding) we achieve - H(UV) lt I(X1YX2)
- H(VU) lt I(X2YX1)
- H(U,V) lt I(X1,X2Y)
- for any joint c.d.f F(x1,x2)
- Converse By Fanos
- Intuition Arbitrary correlation between X1 and
X2 already possible. Correlation between U and V
is thus useless.
15Converse
- nH(UV) lt ?H(YiYi-1,Vn) (Usual Fanos)
- - ?H(YiYi-1,Vn,Un)
- lt ?H(YiYi-1,Vn,X2i)
- - ?H(YiYi-1,Vn,Un,X1i,X2i)
- lt?H(YiX2i)
- - ?H(YiX1i,X2i)
- and thats it!
16DMC MAC with Correlated Sources and Feedback
- Achievable with separate source-channel coding
- H(UV) lt I(X1YX2,T)
- H(VU) lt I(X2YX1,T)
- H(U,V) lt I(X1,X2Y)
- for any p(t)p(x1t)p(x2t)
- T is the cooperation auxiliary random variable
- Note Ability to cooperate limited by first two
inequalities.
17Achievability proof
H(VU) lt I(X2YX1,T)
Tx1 Decodes V
U
Rx Decodes U V
Tx2 Decodes U
V
H(U,V) lt I(X1,X2Y)
H(UV) lt I(X1YX2,T)
- Extension of Cover, El Gamal Salehi argument
18Achievable Region vs. Outer Bound
- Achievable Region
- H(UV) lt I(X1YX2,T)
- H(VU) lt I(X2YX1,T)
- H(U,V) lt I(X1,X2Y)
- Cutset outer Bound
- H(UV) lt I(X1YX2)
- H(VU) lt I(X2YX1)
- H(U,V) lt I(X1,X2Y)
- Missing term I(TYXi) in the first two
inequalities of Achievable region -
for p(t)p(x1t)p(x2t)
19When does separation hold?
- If (T,X1,X2) be a triplet such that
- (X1,X2) extremizes the outer bound
- (X1,X2) conditionally independent given T
- And for this triplet
- H(UV) lt I(X1YX2,T)
- H(VU) lt I(X2YX1,T)
- Then
- That extremum point of the outer bound is
achievable - Source-channel separation holds for that boundary
point - Intuition If U and V correlated beyond a
threshold value - (highly correlated), then source channel
separation holds.
20A more general network
U1
Tx1
U1,Um
Rx1
U2
Tx2
Network
Ui
Txi
U1,Um
Rxn
Um
Txm
21Towards a general statement
- Statement 1 For a system with m transmitters
(messages), n receivers where - Condition 1 Every receiver wishes to receive all
n messages - Condition 2 The cutset outer bound can be
achieved for n independent messages - Then source-channel separation holds
- Proof By Fanos again
22An even more general (but loose statement)
- Statement 2 For a system with m transmitters
(messages), n receivers where - Condition 1 Every receiver wishes to receive all
n messages - Condition 2 U1,Um are highly correlated
(beyond a threshold value) - Then source-channel separation holds
- Proof By Fanos again
23Conclusions
- The ability to cooperate at the channel level
induces source-channel separation - For a 2 user MAC with feedback, we always have
source channel separation - For an arbitrary m Tx, n Rx network, it may
hold if correlations are large enough
24Sensor Networks
- Many distributed nodes can make measurements and
cooperate to propagate information through the
system, perhaps to a single endpoint - Key Observation Nodes located near each other
may have correlated information. What effect
does this have on information flow?
25The Relay Channel
- Introduced by van der Meulen
- Discrete Memoryless Relay Channel consists of
- An Input X.
- A Relay Output Y1.
- A Relay Sender X1, which can depend upon
previously received Y1. - A Channel Output Y.
- A conditional probability distribution
p(y,y1x,x1).
Relay, Receives Y1 Inputs X1
Transmitter, Input X
Receiver, Receives Y
26The Relay Channel in Sensor Networks
- A Relay Channel in a sensor network differs in
that both sender and relay have access to
correlated sources of information. - Can exploit the relays source as side
information - If a MAC channel, i.e. no channel between
transmitter and relay - Then correlation less than 100 is nearly useless
(Gacs, Korner) - If relay channel, correlation is very useful!
27The Relay Channel
- Capacity of General Relay Channel is Unknown
- Upper Bound by Cut-Set Argument
- Achievable Rate by Cover, El Gamal Thm 1
- This Rate is Achieved by Block-Markov Coding
- Introduce a correlation between the transmitter
input and the relay - Decode and Forward Policy
28Block-Markov Coding
- Overview of Block-Markov Coding for Classic Relay
Channel - Relay Terminal completely decodes a message index
w from the set 1..2nR sent by the transmitter - Relay sends the bin index of the message that it
received to aid the receiver in decoding - This is helpful, because the transmitters
codeword is dependent on the bin index that the
relay is transmitting in the same block - Codebook generation
- Fix any p(x,x1) as
- Generate 2nR0 codewords x1n as
- Index them as x1n(s)
- For each of these x1n codewords, generate 2nR xn
codewords as - Index these as xn(ws)
- Independently bin the messages w into the 2nR0
bins s
29Block-Markov Coding
- Encoding
- Messages sent in a total of B blocks
- In the first block
- Relay sends codeword x1n corresponding to a
pre-determined null message, say x1n(f) - Transmitter sends codeword xn, dependant on the
first message w1 and the null message, say
xn(w1f) - In block b
- Assuming relay correctly decoded message wb-1
sent in the previous block, relay sends codeword
x1n(sb-1) - Transmitter sends xn(wbsb-1)
- Same bin index s that relay is simultaneously
sending - Shifted by one block to allow relay to decode
current message
30Block-Markov Transmission
1
2
B
B-1
Block
Message
w1
w2
wB-1
wB
Transmitter
xn(wBwB-1)
xn(w2w1)
Relay
xrn(f)
xrn(w1)
xrn(wB-1)
At the end of each block
w1
wB-1
Relay Decodes
w2
wB
Receiver Decodes
w1
wB-2
wB-1
31Block-Markov Decoding
Y1X1
Source U
X
Y
- Decoding
- At the Relay
- Can determine message if R lt I(XY1X1)
- At the Receiver
- Can determine message if R lt I(X, X1 Y)
- Thus we get R lt min(I(XY1X1), I(X, X1 Y))
32Relay Channel with Correlated Side Information
- Key addition to the coding strategy List
Decoding - With no side information, block-Markov requires
that all of H(U) be sent to relay - If relay has V correlated with U, only need to
push H(UV) information across the channel - Will show that this relay channel has achievable
rate -
- Implies that if correlation between sources, i.e.
I(UV)/H(U) is large enough, we have found
capacity!
33Block Markov Coding
- Encoding
- Identical to Block-Markov
- Decoding at Relay
- Form two lists of message indices
- One of Un jointly typical to Y1n when X1n is
known - Second of Un jointly typical with Vn
- Choose Un that appears in both lists
- Rlt I(XY1X1) I(UV)
34Example MIMO Channel
- Example where Capacity Found
- Description of example MIMO System
- Transmitter to Relay is Point-to-Point with gain
h - Relay and Transmitter each have single antenna,
power constraints P and P1 - Receiver has two antennae and matrix gain H
- Independent Gaussian noise Model with unit noise
power - Desire to transmit the single source U
Y1hX ?1
X
YHx,x1T ?Y1, ?Y2T
35Example Channel
YHx,x1T ?Y1, ?Y2T
- Assumption The transmitter to relay link is
better than the transmit to destination link - Result With 23 correlation full cooperative
capacity can be obtained - With no cooperation data rate is 1 bit/sec
- With block-Markov but no correlated relay - 1.34
bit/sec - With correlated relay, block-Markov and joint
source-channel coding 1.55 bit/sec! - On the flip side, for given data rate, lower
power consumption
36Summary
- Started with the observation that nodes in a
sensor network may have correlated data - Showed that can use this side information to
increase rate (or decrease power usage) - This is a joint source-channel coding strategy
- For some relay channels with correlated sources,
capacity can be found
37Part II Network Coding in Interference Networks
38Overview
- Introduction
- Interference vs. Non-Interference Networks
- Network Coding in Non-Interference Networks
- Gaussian Interference Networks
- Classic Network Coding Example on a Gaussian
Network - Network Coding on Bow-Tie Network
- Network Coding with Node Cooperation
- Summary
39Introduction
- Show by series of examples
- Network-coding in interference network can act
same as network coding in equivalent
non-interference network - Can provide a benefit in interference network
when it does not in the equivalent
non-interference network - Node-cooperation and network-coding used together
can additionally increase throughput
40Non-Interference vs. Interference Networks
- Arbitrary network configuration with conditional
probability distributions across links
Source Node 2
Source Node 1
Terminal Node 1
Terminal Node 3
Terminal Node 2
41Non-Interference vs. Interference Networks
42Classic Network Coding
- Ahlswede, Cai, Li, Yeung, 2000
- Model
- Links transmit single symbol with no errors
- Nodes can perform linear operations on received
symbols - Result
- Multicasting from single source to multiple
receivers can be performed at min-cut max-flow
rate ACLY 2000 - Example
- Ubiquitous routing vs. network coding example
43The Network Coding Example
Network Coding
Routing
1.5 bits to each receiver
2 bits to each receiver
44Gaussian Interference Networks
- Use additive complex white Gaussian noise
channels in the illustrative examples - Each node has transmit power constraint
- Received signal Y??Xi ? ? N(0,N)
- Point-to-point channel
- Capacity Rlg(1P/N)
- Multiple access channel with independent messages
- Sum Rate Capacity
- R1R2lg(1 P1/N P2/N)
P
P1
X1
Y
P2
X2
45Gaussian Interference Networks
- Broadcast Channel with independent messages
- Sum Rate Capacity
- R1R2 lg(1P/N)
- Multicasting over Broadcast Channel
- Same message can be sent to both receivers
- at point-to-point rate
- Broadcast channel with common message
- First example Classic network-coding network
configuration with Gaussian channels - Use P3/2 and N1 at all nodes
- To compare with bit-pipe model
46Gaussian Network Coding Example
A
B
P
P
P
X
C2
Clg(13/2)
Y
C1
P
Clg(13/23/2)2
D1
D2
47Interference Network Example Routing
- Strategies for the classic configuration in
interference model - Ignore center nodes X,Y completely
- Receive lg(5/2) bits/transmission each
- Routing strategy
- Constraints if multicasting at broadcast nodes
- MACs 2 bits/transmission
- Direct links lg(5/2) bits/transmission
- RARBRXYRT1 is viable
- Route so that X-Y and Y-D links carry As
information - 1 bit to D1, 2 bits to D2, 1.5 bits on average
- Can show with linear programming that no better
pure routing strategy exists
RA
RB
RA
RB
RXY
RT
RT
48Interference Network Example Network Coding
- Instead, use the same strategy as in the
non-interference network - Set all links to rate 1 bit/transmission
- Send A B on links X-Y, Y-D1, and Y-D2
- Each terminal receives 2 bits/transmission
- Rate of 2 bits/transmission is maximum with
- independent codebooks
- cut C1 across multiple access channel
- Exceeding the independent codebooks rate is
- the subject of the node-cooperation section
- Essentially, taking the same actions as the
non-interference network
A
B
P
P
P
X
Y
C1
P
D1
D2
49Aside Multicasting
- In the previous example, all broadcast nodes
operated by multicasting - The same amount of information crosses cut C1 in
either mode - But, greater amount crosses each of C2 and C3 in
the multicasting mode
P
(1-?)P
C1
C1
?P
C2
C3
C2
C3
R1
R2
R1
R2
Multicasting
Broadcasting
R1R2lg(1P)
R1R2lg(1P)
50Example Multicasting Outperforms Broadcasting
- In the following example, multicasting is
superior to broadcasting - Three independent sources
- Achievable rate across cut C1 when center node
multicasts - Final term removes rate of common message
- When center node broadcasts
S0
S2
S1
P
P
P
R1
R2
C1
51Effect of Multicasting
- Intuition Broadcast requirement of the
interference network acts like a T
configuration in the non-interference network - Adding a bottleneck link of equal capacity onto
the node in the non-interference case constrains
the two downlinks to send the same signal - Otherwise, not operating at full capacity
- Analogy useful for next example
P
Bottleneck link
Multicasting in Interference Network
T Configuration in Non-Interference Network
52Bow-Tie Network
- Non-interference network routing performs
optimally - Interference Network Strategies
- Ignore center node X
- Allows lg(13/2) bits/transmission to each
terminal node - Routing Center node X alternately chooses A or
B - to forward
- Rate of 1.5 bits/transmission to each terminal
node - Network Coding
- Node X multicasts A B to both destination nodes
- Rate of 2 bits/transmission to each terminal node
is optimal - Network Coding is useful in some interference
networks when it provides no benefit in the
non-interference network with the same
configuration
B
A
X
D1
D2
53Node Cooperation
- Network coding is one method of inducing
correlations on a network - Other methods
- Using correlated codebooks at separate nodes
- Block-Markov coding for relay channel
- For non-interference networks, independent coding
shown optimal - Demonstrate interference network counter-example
- Example Network coding and node-cooperation in
tandem increasing performance
U,V
V
U
P0
P
P
D1
D2
54Node Cooperation
U,V
V
U
- Nodes T1 and T2 have access to data sources U and
V, respectively, while node T0 has access to both
U and V - Nodes T1 and T2 have power constraint P, while
node T0 has constraint P0 - When PP03/2
- Routing 1.5 bits/transmission
- Network coding with independent codebooks 2
bits/transmission
T1
T0
T2
P0
P
P
D1
D2
55Node Cooperation
U,V
V
U
T1
T0
T2
P0
P
P
- Network coding with node cooperation scheme
- Three codebooks for U message, V message, and U
V message - Equal timesharing over two modes of operation
- In first mode,
- Node T1 transmits U codeword scaled to power P?
- Node T2 transmits V codeword scaled to power P- ?
- Node T0 splits its power to transmit both U and U
V - Second mode Reverse of first mode
D1
D2
56Node Cooperation
U,V
V
U
T1
T0
T2
P ?
?0
P- ?
P0- ?0
- Network coding with node cooperation scheme
- T1 and T0 cooperate to send the codeword for U to
D1 - T0 sends codeword for U V to both D1 and D2
- T2 sends codeword for V to D2
- Destinations receive their own sources directly
- Destinations can decode the opposite source with
U V message - With appropriate choice of parameters
- Average rate of 2.06 is achievable at each node
D1
D2
57Summary
- Bit-pipe non-interference model is not complete
- Neglects interactions between channels
- In wireless scenario, imposes orthogonality
constraint on channels - Any gains due to node-cooperation necessarily
lost - Noisy channel model captures these effects
- More difficult to handle
- Results for non-interference networks can be
duplicated - Configurations exists where bit-pipe model has no
network coding gain, while interference model
benefits from network coding - Node cooperation strategies can increase the
throughput beyond that rate achieved by using
independent codebooks