on compute-and-forward with feedback€¦ · gaussianmultiple-accesschannel w 1 e 1 x 1 y d wˆ...
TRANSCRIPT
![Page 1: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/1.jpg)
On Compute-and-Forward with Feedback
Or OrdentlichJoint work with Uri Erez and Bobak Nazer
Information Theory WorkshopJerusalem, IsraelApril 27, 2015
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 2: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/2.jpg)
Gaussian Multiple-Access Channel
W1 E1 X1
Y D W1, W2
W2 E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 3: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/3.jpg)
Gaussian Multiple-Access Channel
W1 E1 X1
Y D W1, W2
W2 E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
Capacity Region
R1 <1
2log(1 + h21SNR)
R2 <1
2log(1 + h22SNR)
R1 + R2 <1
2log(1 + ‖h‖2SNR)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 4: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/4.jpg)
Gaussian Multiple-Access Channel
W1 E1 X1
Y D W1, W2
W2 E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 5: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/5.jpg)
Gaussian Multiple-Access Channel
W1 E1 X1
Y D W1, W2
W2 E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
Feedback Capacity Region (Ozarow 84)
R1 <1
2log(1 + (1− ρ
2)h21SNR)
R2 <1
2log(1 + (1− ρ
2)h22SNR)
R1 + R2 <1
2log(1 + (‖h‖2 + 2ρ|h1h2|)SNR)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 6: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/6.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
R1 = R2 = R = knlog p
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 7: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/7.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
R1 = R2 = R = knlog p
We assume full CSI everywhere
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 8: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/8.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
R1 = R2 = R = knlog p
We assume full CSI everywhere
Only lower and upper bounds are known (Nazer & Gastpar 11)
1
2log
(
1
2+ min{h21, h22}SNR
)
≤ Ccomp ≤ 1
2log
(
1 + min{h21, h22}SNR)
At high SNR the bounds coincide. At low SNR separation is optimal
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 9: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/9.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 10: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/10.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
How Much Does Feedback Help?
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 11: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/11.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
w2
w2
How Much Does Feedback Help?
Upper bound remains the same Ccomp ≤ 12 log
(
1 + min{h21, h22}SNR)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 12: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/12.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
How Much Does Feedback Help?
Upper bound remains the same Ccomp ≤ 12 log
(
1 + min{h21, h22}SNR)
No non-trivial lower bounds are known
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 13: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/13.jpg)
Gaussian MAC - Compute-and-Forward
w1 ∈ Fkp E1 X1
Y D w1 ⊕w2
w2 ∈ Fkp E2 X2
Z ∼ N (0, 1)
h1
h2
Y = h1X1 + h2X2 + Z
How Much Does Feedback Help?
Upper bound remains the same Ccomp ≤ 12 log
(
1 + min{h21, h22}SNR)
No non-trivial lower bounds are known
In this work we derive a novel lower bound
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 14: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/14.jpg)
Gaussian MAC - Compute-and-Forward
Main Result
For any 0 < ρ ≤ 1 let
ρ1 = 1− (1− ρ)
(
h2
h1
)2
,
Rc =1
2log+
(
1
2+ (1− ρ)h22SNR
)
,
R ′ =1
2log
(
1 +(h1
√ρ1 + h2
√ρ)2SNR
1 + 2(1− ρ)h22SNR
)
.
Any computation rate satisfying
R < max0<ρ≤1
min
(
R ′ + Rc ,1
2log
(
1 + (1− ρ)h22SNR)
)
is achievable with feedback.
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 15: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/15.jpg)
Gaussian MAC - Compute-and-Forward
h2
0 0.5 1 1.5 2
Rat
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8h
1=1, SNR=2
Upper bound
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 16: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/16.jpg)
Gaussian MAC - Compute-and-Forward
h2
0 0.5 1 1.5 2
Rat
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8h
1=1, SNR=2
Upper boundComputation rate - full CSI
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 17: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/17.jpg)
Gaussian MAC - Compute-and-Forward
h2
0 0.5 1 1.5 2
Rat
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8h
1=1, SNR=2
Upper boundComputation rate - full CSISeparartion with feedback
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 18: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/18.jpg)
Gaussian MAC - Compute-and-Forward
h2
0 0.5 1 1.5 2
Rat
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8h
1=1, SNR=2
Upper boundComputation rate - full CSISeparartion with feedbackFeedback computation rate
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 19: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/19.jpg)
Lattice List Decoder (Song & Devroye 13)
Fine lattice Λc
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 20: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/20.jpg)
Lattice List Decoder (Song & Devroye 13)
Fine lattice Λc , coarse lattice Λ⊆ Λc
C = Λc ∩ V
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 21: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/21.jpg)
Lattice List Decoder (Song & Devroye 13)
Fine lattice Λc , coarse lattice Λ, intermediate lattice Λs , Λ ⊆ Λs ⊆ Λc
C = Λc ∩ V
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 22: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/22.jpg)
Lattice List Decoder (Song & Devroye 13)
AWGN channel y = x+ z, R > C
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 23: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/23.jpg)
Lattice List Decoder (Song & Devroye 13)
AWGN channel y = x+ z, R > C
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 24: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/24.jpg)
Lattice List Decoder (Song & Devroye 13)
AWGN channel y = x+ z, R > C
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 25: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/25.jpg)
Lattice List Decoder (Song & Devroye 13)
AWGN channel y = x+ z, R > C
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 26: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/26.jpg)
Lattice List Decoder (Song & Devroye 13)
AWGN channel y = x+ z, R > C
Decode a list of codewords: L = {c ∈ C : c ∈ [y + Vs ] mod Λ}
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 27: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/27.jpg)
Lattice List Decoder (Song & Devroye 13)
AWGN channel y = x+ z, R > C
Decode a list of codewords: L = {c ∈ C : c ∈ [y + Vs ] mod Λ}|L| = log
(
Vol(Vs)Vol(Vc )
)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 28: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/28.jpg)
Lattice List Decoder (Song & Devroye 13)
Theorem (Song & Devroye 13)
It is possible to decode a list with size 2n(R−C) that contains the truecodeword w.h.p. using a lattice list decoder
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 29: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/29.jpg)
Compute-and-Forward with Feedback - High-Level
High-level overview of our coding scheme
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 30: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/30.jpg)
Compute-and-Forward with Feedback - High-Level
High-level overview of our coding scheme
Block Markov coding
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 31: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/31.jpg)
Compute-and-Forward with Feedback - High-Level
High-level overview of our coding scheme
Block Markov coding
In the end of each block user i can decode wi using the feedback link
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 32: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/32.jpg)
Compute-and-Forward with Feedback - High-Level
High-level overview of our coding scheme
Block Markov coding
In the end of each block user i can decode wi using the feedback link
In each block, each user superimposes encoding of a new messageand encoding of the sum of messages from the last block
The encoding of the sum is transmitted coherently
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 33: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/33.jpg)
Compute-and-Forward with Feedback - High-Level
High-level overview of our coding scheme
Block Markov coding
In the end of each block user i can decode wi using the feedback link
In each block, each user superimposes encoding of a new messageand encoding of the sum of messages from the last block
The encoding of the sum is transmitted coherently
The receiver decodes the coherent part first, and then a list ofcandidates for the new sum
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 34: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/34.jpg)
Compute-and-Forward with Feedback - High-Level
High-level overview of our coding scheme
Block Markov coding
In the end of each block user i can decode wi using the feedback link
In each block, each user superimposes encoding of a new messageand encoding of the sum of messages from the last block
The encoding of the sum is transmitted coherently
The receiver decodes the coherent part first, and then a list ofcandidates for the new sum
A compute-and-forward variant of Cover-Leung 81
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 35: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/35.jpg)
Compute-and-Forward with Feedback
For simplicity assume h1 = h2 = 1
Decoding w(k)1 ⊕ w
(k)2 , k = 1, . . . ,N over N + 1 blocks
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 36: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/36.jpg)
Compute-and-Forward with Feedback
For simplicity assume h1 = h2 = 1
Decoding w(k)1 ⊕ w
(k)2 , k = 1, . . . ,N over N + 1 blocks
Both users encode their messages using the same lattice code C, suchthat x
(k)i = f
(
w(k)i
)
∈ C
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 37: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/37.jpg)
Compute-and-Forward with Feedback
For simplicity assume h1 = h2 = 1
Decoding w(k)1 ⊕ w
(k)2 , k = 1, . . . ,N over N + 1 blocks
Both users encode their messages using the same lattice code C, suchthat x
(k)i = f
(
w(k)i
)
∈ C
First block: x(1)i =
√1− ρx
(1)i
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 38: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/38.jpg)
Compute-and-Forward with Feedback
For simplicity assume h1 = h2 = 1
Decoding w(k)1 ⊕ w
(k)2 , k = 1, . . . ,N over N + 1 blocks
Both users encode their messages using the same lattice code C, suchthat x
(k)i = f
(
w(k)i
)
∈ C
First block: x(1)i =
√1− ρx
(1)i
Receiver sees
y(1) =√
1− ρ
(
x(1)1 + x
(1)2
)
+ z(1).
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 39: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/39.jpg)
Compute-and-Forward with Feedback
For simplicity assume h1 = h2 = 1
Decoding w(k)1 ⊕ w
(k)2 , k = 1, . . . ,N over N + 1 blocks
Both users encode their messages using the same lattice code C, suchthat x
(k)i = f
(
w(k)i
)
∈ C
First block: x(1)i =
√1− ρx
(1)i
Receiver sees
y(1) =√
1− ρ
(
x(1)1 + x
(1)2
)
+ z(1).
R < Rcomp , 12 log
(
12 + (1− ρ)SNR
)
is needed for decoding
w(1)1 ⊕ w
(1)2
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 40: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/40.jpg)
Compute-and-Forward with Feedback
For simplicity assume h1 = h2 = 1
Decoding w(k)1 ⊕ w
(k)2 , k = 1, . . . ,N over N + 1 blocks
Both users encode their messages using the same lattice code C, suchthat x
(k)i = f
(
w(k)i
)
∈ C
First block: x(1)i =
√1− ρx
(1)i
Receiver sees
y(1) =√
1− ρ
(
x(1)1 + x
(1)2
)
+ z(1).
R < Rcomp , 12 log
(
12 + (1− ρ)SNR
)
is needed for decoding
w(1)1 ⊕ w
(1)2
In our case R > Rcomp and the receiver can decode a list L(1) of
candidates for w(1)1 ⊕ w
(1)2 with size |L(1)| = 2n(R−Rcomp)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 41: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/41.jpg)
Compute-and-Forward with Feedback
Using the feedback link, user i can decode w(1)
iif
R <1
2log(1 + (1− ρ)SNR)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 42: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/42.jpg)
Compute-and-Forward with Feedback
Using the feedback link, user i can decode w(1)
iif
R <1
2log(1 + (1− ρ)SNR)
Both users can compute v(1) = w(1)1 ⊕ w
(1)2
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 43: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/43.jpg)
Compute-and-Forward with Feedback
Using the feedback link, user i can decode w(1)
iif
R <1
2log(1 + (1− ρ)SNR)
Both users can compute v(1) = w(1)1 ⊕ w
(1)2
Both users apply the same binning function B : [2nR ] 7→ [2nR′
],R ′ < R , to obtain B
(
v(1))
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 44: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/44.jpg)
Compute-and-Forward with Feedback
Using the feedback link, user i can decode w(1)
iif
R <1
2log(1 + (1− ρ)SNR)
Both users can compute v(1) = w(1)1 ⊕ w
(1)2
Both users apply the same binning function B : [2nR ] 7→ [2nR′
],R ′ < R , to obtain B
(
v(1))
Each user encodes B(
v(1))
to x(1)cohr using the same codebook C′ with
rate R ′ and average power SNR
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 45: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/45.jpg)
Compute-and-Forward with Feedback
Using the feedback link, user i can decode w(1)
iif
R <1
2log(1 + (1− ρ)SNR)
Both users can compute v(1) = w(1)1 ⊕ w
(1)2
Both users apply the same binning function B : [2nR ] 7→ [2nR′
],R ′ < R , to obtain B
(
v(1))
Each user encodes B(
v(1))
to x(1)cohr using the same codebook C′ with
rate R ′ and average power SNR
In addition, each user encodes a new message w(2)i to the codeword
x(2)i and transmits
x(2)i =
√ρx
(1)cohr +
√
1− ρx(2)i
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 46: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/46.jpg)
Compute-and-Forward with Feedback
Channel output is
y(2) = 2√ρx
(1)cohr +
√
1− ρ
(
x(2)1 + x
(2)2
)
+ z(2)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 47: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/47.jpg)
Compute-and-Forward with Feedback
Channel output is
y(2) = 2√ρx
(1)cohr +
√
1− ρ
(
x(2)1 + x
(2)2
)
+ z(2)
Can decode x(1)cohr if
R ′ ≤ 1
2log
(
1 +4ρSNR
1 + 2(1− ρ)SNR
)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 48: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/48.jpg)
Compute-and-Forward with Feedback
Channel output is
y(2) = 2√ρx
(1)cohr +
√
1− ρ
(
x(2)1 + x
(2)2
)
+ z(2)
Can decode x(1)cohr if
R ′ ≤ 1
2log
(
1 +4ρSNR
1 + 2(1− ρ)SNR
)
The decoder looks for a unique w ∈ Fkp in L(1) ∩ B−1
(
v(1))
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 49: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/49.jpg)
Compute-and-Forward with Feedback
Channel output is
y(2) = 2√ρx
(1)cohr +
√
1− ρ
(
x(2)1 + x
(2)2
)
+ z(2)
Can decode x(1)cohr if
R ′ ≤ 1
2log
(
1 +4ρSNR
1 + 2(1− ρ)SNR
)
The decoder looks for a unique w ∈ Fkp in L(1) ∩ B−1
(
v(1))
If R ′ > R − Rcomp such a w ∈ Fkp will be found with probability 1
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 50: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/50.jpg)
Compute-and-Forward with Feedback
Channel output is
y(2) = 2√ρx
(1)cohr +
√
1− ρ
(
x(2)1 + x
(2)2
)
+ z(2)
Can decode x(1)cohr if
R ′ ≤ 1
2log
(
1 +4ρSNR
1 + 2(1− ρ)SNR
)
The decoder looks for a unique w ∈ Fkp in L(1) ∩ B−1
(
v(1))
If R ′ > R − Rcomp such a w ∈ Fkp will be found with probability 1
Next, the decoder subtracts x(1)cohr from y(2) and decodes a list L(2) of
candidates for v(2) = w(2)1 ⊕w
(2)2
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 51: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/51.jpg)
Compute-and-Forward with Feedback - Rate Constraints
Correct decoding through feedback link
R <1
2log(1 + (1− ρ)SNR)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 52: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/52.jpg)
Compute-and-Forward with Feedback - Rate Constraints
Correct decoding through feedback link
R <1
2log(1 + (1− ρ)SNR)
Correct decoding of xcohr
R ′ ≤ 1
2log
(
1 +4ρSNR
1 + 2(1− ρ)SNR
)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 53: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/53.jpg)
Compute-and-Forward with Feedback - Rate Constraints
Correct decoding through feedback link
R <1
2log(1 + (1− ρ)SNR)
Correct decoding of xcohr
R ′ ≤ 1
2log
(
1 +4ρSNR
1 + 2(1− ρ)SNR
)
Unique element in intersection of list and bin
R ′> R − 1
2log
(
1
2+ (1− ρ)SNR
)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 54: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/54.jpg)
Compute-and-Forward with Feedback - Rate Constraints
Achievable Rate
R < min
{
1
2log(1 + (1− ρ)SNR),
1
2log
(
1 +4ρSNR
1 + 2(1 − ρ)SNR
)
+1
2log
(
1
2+ (1− ρ)SNR
)}
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 55: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/55.jpg)
Compute-and-Forward with Feedback - Rate Constraints
Achievable Rate
R < min
{
1
2log(1 + (1− ρ)SNR),
1
2log
(
1
2+ (1 + ρ)SNR
)}
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 56: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/56.jpg)
Compute-and-Forward with Feedback - Rate Constraints
Achievable Rate
R < min
{
1
2log(1 + (1− ρ)SNR),
1
2log
(
1
2+ (1 + ρ)SNR
)}
Setting ρ = 14SNR we get
R <12 log
(
34 + SNR
)
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback
![Page 57: On compute-and-Forward with Feedback€¦ · GaussianMultiple-AccessChannel W 1 E 1 X 1 Y D Wˆ 1,Wˆ 2 W 2 E 2 X 2 Z ∼ N(0,1) h 1 h 2 Y = h 1X 1 +h 2X 2 +Z CapacityRegion R 1](https://reader034.vdocument.in/reader034/viewer/2022050407/5f84b4a61310e002952f6c8f/html5/thumbnails/57.jpg)
Summary and Conclusions
We studied the problem of computing a linear function from theoutput of a Gaussian MAC with feedback
We derived a new coding scheme for this scenario
For a symmetric setting our scheme achieves R = 12 log
(
34 + SNR
)
The scheme can be extended to noisy feedback and more than 2 users
Our scheme works in blocks. Can we find a scalar, a laSchalkwijk-Kailath 66 scheme?
Ordentlich, Erez, Nazer On compute-and-Forward with Feedback