algebraic structures for distributed source codingdineshk/thesis/ita_09_thesis_overview.pdf ·...
TRANSCRIPT
Algebraic Structures for Distributed Source Coding
Dinesh Krithivasan and S. Sandeep Pradhan
University of Michigan
ITA 2009
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 1 / 21
Introduction
Random coding - common in information theory
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Introduction
Random coding - common in information theory
Structured codes - explored mainly for complexity reasons
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Introduction
Random coding - common in information theory
Structured codes - explored mainly for complexity reasons
Structured codes have performance unattainable by random codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Introduction
Random coding - common in information theory
Structured codes - explored mainly for complexity reasons
Structured codes have performance unattainable by random codes
Can achieve expurgated bound on the BSC
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Introduction
Random coding - common in information theory
Structured codes - explored mainly for complexity reasons
Structured codes have performance unattainable by random codes
Can achieve expurgated bound on the BSC
Some works on first order performance gains offered by linear codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Introduction
Random coding - common in information theory
Structured codes - explored mainly for complexity reasons
Structured codes have performance unattainable by random codes
Can achieve expurgated bound on the BSC
Some works on first order performance gains offered by linear codes
This thesis: Unified framework for using structured codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Introduction
Random coding - common in information theory
Structured codes - explored mainly for complexity reasons
Structured codes have performance unattainable by random codes
Can achieve expurgated bound on the BSC
Some works on first order performance gains offered by linear codes
This thesis: Unified framework for using structured codes
Attains/Exceeds known performance limits for many problems
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Introduction
Random coding - common in information theory
Structured codes - explored mainly for complexity reasons
Structured codes have performance unattainable by random codes
Can achieve expurgated bound on the BSC
Some works on first order performance gains offered by linear codes
This thesis: Unified framework for using structured codes
Attains/Exceeds known performance limits for many problems
Suggests practical code constructions
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 2 / 21
Thesis Overview
Very general framework - includes many famous problems
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 3 / 21
Thesis Overview
Very general framework - includes many famous problems
Describe avenues for rate gains
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 3 / 21
Thesis Overview
Very general framework - includes many famous problems
Describe avenues for rate gains
Role of structured codes in obtaining these rate gains
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 3 / 21
Thesis Overview
Very general framework - includes many famous problems
Describe avenues for rate gains
Role of structured codes in obtaining these rate gains
Gaussian sources - Lattice codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 3 / 21
Thesis Overview
Very general framework - includes many famous problems
Describe avenues for rate gains
Role of structured codes in obtaining these rate gains
Gaussian sources - Lattice codes
Discrete sources - Group codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 3 / 21
Thesis Overview
Very general framework - includes many famous problems
Describe avenues for rate gains
Role of structured codes in obtaining these rate gains
Gaussian sources - Lattice codes
Discrete sources - Group codes
Applications to sensor networks, video coding etc.
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 3 / 21
A Distributed Source Coding Problem
Xn
Y n
f1(·)
f2(·)
Z
Encoder 1
Encoder 2
Decoder
R1
R2
Ed(X, Y, Z) ≤ D
Set of encoders observe different components of a vector source
Central decoder receives quantized observations from the encoders
Decoder interested in minimizing a joint distortion criterion
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 4 / 21
Joint Distortion Criterion
Distortion criterion depends on all the sources and decoder
reconstructions
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 5 / 21
Joint Distortion Criterion
Distortion criterion depends on all the sources and decoder
reconstructions
Special cases of this framework
Slepian-Wolf problem
Wyner-Ziv problem
Ahlswede-Korner-Wyner problem
Berger-Yeung problem
Berger-Tung problem
Korner-Marton binary XOR problem
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 5 / 21
Joint Distortion Criterion
Distortion criterion depends on all the sources and decoder
reconstructions
Special cases of this framework
Slepian-Wolf problem
Wyner-Ziv problem
Ahlswede-Korner-Wyner problem
Berger-Yeung problem
Berger-Tung problem
Korner-Marton binary XOR problem
Best known rate region - Berger-Tung based
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 5 / 21
Motivation for our coding scheme
Suppose decoder receives auxiliary random variables U ,V
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 6 / 21
Motivation for our coding scheme
Suppose decoder receives auxiliary random variables U ,V
Encoders don’t communicate with each other: V −Y −X −U
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 6 / 21
Motivation for our coding scheme
Suppose decoder receives auxiliary random variables U ,V
Encoders don’t communicate with each other: V −Y −X −U
Decoder reconstructs G(U ,V )
G(U ,V ) , arg minG :Z=G(U ,V )d (X ,Y , Z )
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 6 / 21
Motivation for our coding scheme
Suppose decoder receives auxiliary random variables U ,V
Encoders don’t communicate with each other: V −Y −X −U
Decoder reconstructs G(U ,V )
G(U ,V ) , arg minG :Z=G(U ,V )d (X ,Y , Z )
Decoder not interested in (U ,V ). Only in G(U ,V )
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 6 / 21
Motivation for our coding scheme
Suppose decoder receives auxiliary random variables U ,V
Encoders don’t communicate with each other: V −Y −X −U
Decoder reconstructs G(U ,V )
G(U ,V ) , arg minG :Z=G(U ,V )d (X ,Y , Z )
Decoder not interested in (U ,V ). Only in G(U ,V )
Encode such that decoder can only reconstruct what it needs
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 6 / 21
An Illustrative Example
X ,Y - 3 bit correlated binary sources, dH (X ,Y )≤ 1
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 7 / 21
An Illustrative Example
X ,Y - 3 bit correlated binary sources, dH (X ,Y )≤ 1
Lossless reconstruction of Z = X ⊕2 Y ∈ {000,001,010,100}
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 7 / 21
An Illustrative Example
X ,Y - 3 bit correlated binary sources, dH (X ,Y )≤ 1
Lossless reconstruction of Z = X ⊕2 Y ∈ {000,001,010,100}
Berger-Tung based coding scheme:
Reconstruct sources X ,Y . Compute Z = X ⊕2 Y
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 7 / 21
An Illustrative Example
X ,Y - 3 bit correlated binary sources, dH (X ,Y )≤ 1
Lossless reconstruction of Z = X ⊕2 Y ∈ {000,001,010,100}
Berger-Tung based coding scheme:
Reconstruct sources X ,Y . Compute Z = X ⊕2 Y
Sum rate: H(X ,Y ) = 5 bits
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 7 / 21
An Illustrative Example
X ,Y - 3 bit correlated binary sources, dH (X ,Y )≤ 1
Lossless reconstruction of Z = X ⊕2 Y ∈ {000,001,010,100}
Berger-Tung based coding scheme:
Reconstruct sources X ,Y . Compute Z = X ⊕2 Y
Sum rate: H(X ,Y ) = 5 bits
Can we do better?
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 7 / 21
An Illustrative Example contd.
A linear coding scheme:
Z1 ⊕ Z2
Z1 ⊕ Z3
X1X2X3
1 1 0
1 0 1
1 1 0
1 0 1
Y1Y2Y3
X1 ⊕ X2
X1 ⊕ X3
Y1 ⊕ Y2
Y1 ⊕ Y3
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 8 / 21
An Illustrative Example contd.
A linear coding scheme:
Z1 ⊕ Z2
Z1 ⊕ Z3
X1X2X3
1 1 0
1 0 1
1 1 0
1 0 1
Y1Y2Y3
X1 ⊕ X2
X1 ⊕ X3
Y1 ⊕ Y2
Y1 ⊕ Y3
Sum rate: 2+2 = 4 bits = 2H (Z )
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 8 / 21
An Illustrative Example contd.
A linear coding scheme:
Z1 ⊕ Z2
Z1 ⊕ Z3
X1X2X3
1 1 0
1 0 1
1 1 0
1 0 1
Y1Y2Y3
X1 ⊕ X2
X1 ⊕ X3
Y1 ⊕ Y2
Y1 ⊕ Y3
Sum rate: 2+2 = 4 bits = 2H (Z )
Significant features:
Identical binning at both encoders.
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 8 / 21
An Illustrative Example contd.
A linear coding scheme:
Z1 ⊕ Z2
Z1 ⊕ Z3
X1X2X3
1 1 0
1 0 1
1 1 0
1 0 1
Y1Y2Y3
X1 ⊕ X2
X1 ⊕ X3
Y1 ⊕ Y2
Y1 ⊕ Y3
Sum rate: 2+2 = 4 bits = 2H (Z )
Significant features:
Identical binning at both encoders. Binning performed by Linear codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 8 / 21
An Illustrative Example contd.
A linear coding scheme:
Z1 ⊕ Z2
Z1 ⊕ Z3
X1X2X3
1 1 0
1 0 1
1 1 0
1 0 1
Y1Y2Y3
X1 ⊕ X2
X1 ⊕ X3
Y1 ⊕ Y2
Y1 ⊕ Y3
Sum rate: 2+2 = 4 bits = 2H (Z )
Significant features:
Identical binning at both encoders. Binning performed by Linear codes
Lossless reconstruction. Entire space is binned
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 8 / 21
Jointly Gaussian sources and linear function reconstruction
Source (X1, X2) is bivariate Gaussian
X1, X2 ∼N (0,1), E(X1 X2)= ρ > 0
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 9 / 21
Jointly Gaussian sources and linear function reconstruction
Source (X1, X2) is bivariate Gaussian
X1, X2 ∼N (0,1), E(X1 X2)= ρ > 0
Decoder interested in Z = X1 −c X2 with mean square distortion, c > 0
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 9 / 21
Jointly Gaussian sources and linear function reconstruction
Source (X1, X2) is bivariate Gaussian
X1, X2 ∼N (0,1), E(X1 X2)= ρ > 0
Decoder interested in Z = X1 −c X2 with mean square distortion, c > 0
Berger-Tung based coding scheme
Encoders: Quantize X1 to W1, X2 to W2. Transmit W1,W2
Decoder: Reconstruct Z = E(Z |W1,W2)
Optimal for c < 0 with Gaussian test channels
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 9 / 21
Nested Lattice codes
Lattice codes: Equivalent to linear codes for continuous sources
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 10 / 21
Nested Lattice codes
Lattice codes: Equivalent to linear codes for continuous sources
Lattice code: points in Rn closed under addition
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 10 / 21
Nested Lattice codes
Lattice codes: Equivalent to linear codes for continuous sources
Lattice code: points in Rn closed under addition
Need to do quantization first - nested lattice codes
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 10 / 21
Nested Lattice codes
Lattice codes: Equivalent to linear codes for continuous sources
Lattice code: points in Rn closed under addition
Need to do quantization first - nested lattice codes
Quantizer - nearest neighbour
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 10 / 21
Nested Lattice codes
Lattice codes: Equivalent to linear codes for continuous sources
Lattice code: points in Rn closed under addition
Need to do quantization first - nested lattice codes
Quantizer - nearest neighbour
Binning - coset leader of the
quantized fine lattice point in
the coarse lattice
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 10 / 21
Nested Lattice codes
Lattice codes: Equivalent to linear codes for continuous sources
Lattice code: points in Rn closed under addition
Need to do quantization first - nested lattice codes
Quantizer - nearest neighbour
Binning - coset leader of the
quantized fine lattice point in
the coarse lattice
Fine code - Quantizes the source
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 10 / 21
Nested Lattice codes
Lattice codes: Equivalent to linear codes for continuous sources
Lattice code: points in Rn closed under addition
Need to do quantization first - nested lattice codes
Quantizer - nearest neighbour
Binning - coset leader of the
quantized fine lattice point in
the coarse lattice
Fine code - Quantizes the source
Coarse code - bins fine code−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 10 / 21
Rate region using nested lattice codes
Nested lattice codes Λ2 ⊂Λ11,Λ2 ⊂Λ12
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 11 / 21
Rate region using nested lattice codes
Nested lattice codes Λ2 ⊂Λ11,Λ2 ⊂Λ12
Λ11,Λ12 - good lattice source codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 11 / 21
Rate region using nested lattice codes
Nested lattice codes Λ2 ⊂Λ11,Λ2 ⊂Λ12
Λ11,Λ12 - good lattice source codes
Λ2 - good lattice channel code
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 11 / 21
Rate region using nested lattice codes
Achievable tuples satisfy 2−2R1 +2−2R2 ≤
(
σ2z
D
)−1
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 11 / 21
Rate region using nested lattice codes
Achievable tuples satisfy 2−2R1 +2−2R2 ≤
(
σ2z
D
)−1
For certain sources and distortions: better than Berger-Tung based
scheme
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 11 / 21
Rate region using nested lattice codes
Achievable tuples satisfy 2−2R1 +2−2R2 ≤
(
σ2z
D
)−1
For certain sources and distortions: better than Berger-Tung based
scheme
Gains based on alignment of [1, −c] with eigenvectors of source
covariance matrixD. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 11 / 21
Lattice coding vs Berger-Tung based coding
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.10
2
4
6
8
10
12
14
16
18
20
Distortion D
Sum
rat
e R
1 + R
2
Comparison between the two coding schemes
Berger−Tung Sum RateLattice Sum Rate
ρ = 0.95c = 1
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 12 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
Decoder interested only in G(U ,V )
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
Decoder interested only in G(U ,V )
Function G(U ,V ) equivalent to addition in some abelian group G
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
Decoder interested only in G(U ,V )
Function G(U ,V ) equivalent to addition in some abelian group G
Abelian groups decomposable into primary cyclic groups Zpr
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
Decoder interested only in G(U ,V )
Function G(U ,V ) equivalent to addition in some abelian group G
Abelian groups decomposable into primary cyclic groups Zpr
g ∈G ⇔ g = (g1, . . . , gk ), gi ∈Zpiei
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
Decoder interested only in G(U ,V )
Function G(U ,V ) equivalent to addition in some abelian group G
Abelian groups decomposable into primary cyclic groups Zpr
g ∈G ⇔ g = (g1, . . . , gk ), gi ∈Zpiei
Suffices to build codes over Zpr
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
Decoder interested only in G(U ,V )
Function G(U ,V ) equivalent to addition in some abelian group G
Abelian groups decomposable into primary cyclic groups Zpr
g ∈G ⇔ g = (g1, . . . , gk ), gi ∈Zpiei
Suffices to build codes over Zpr - nested group codes
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Discrete sources and arbitrary distortions
Arbitrary discrete sources and arbitrary distortions
Fix test channels: PX Y UV = PX Y PU |X PV |Y
Decoder interested only in G(U ,V )
Function G(U ,V ) equivalent to addition in some abelian group G
Abelian groups decomposable into primary cyclic groups Zpr
g ∈G ⇔ g = (g1, . . . , gk ), gi ∈Zpiei
Suffices to build codes over Zpr - nested group codes
Extension to arbitrary groups through sequential coding
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 13 / 21
Nested Group codes
Group code over Znpr : C <Z
npr
C = ker(φ) for some homomorphism φ : Znpr →Z
kpr
(C1,C2) nested if C2 ⊂C1
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 14 / 21
Nested Group codes
Group code over Znpr : C <Z
npr
C = ker(φ) for some homomorphism φ : Znpr →Z
kpr
(C1,C2) nested if C2 ⊂C1
We need:
C1 <Znpr : “good” source code
C2 <Znpr : “good” channel code
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 14 / 21
Nested Group codes
Group code over Znpr : C <Z
npr
C = ker(φ) for some homomorphism φ : Znpr →Z
kpr
(C1,C2) nested if C2 ⊂C1
We need:
C1 <Znpr : “good” source code
Can find un ∈C1 jointly typical with source xn
C2 <Znpr : “good” channel code
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 14 / 21
Nested Group codes
Group code over Znpr : C <Z
npr
C = ker(φ) for some homomorphism φ : Znpr →Z
kpr
(C1,C2) nested if C2 ⊂C1
We need:
C1 <Znpr : “good” source code
Can find un ∈C1 jointly typical with source xn
C2 <Znpr : “good” channel code
Can distinguish between typical channel noise sequences
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 14 / 21
Group source codes: Example
Consider Z4 = {0,1,2,3}
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 15 / 21
Group source codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 15 / 21
Group source codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group source code for (X ,U ,PXU )
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 15 / 21
Group source codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group source code for (X ,U ,PXU )
Assume U =Z4
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 15 / 21
Group source codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group source code for (X ,U ,PXU )
Assume U =Z4
Exists for large n if 1n
log |C1| ≥ log 4−min{H (U |X ),2|H (U |X )−1|+}
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 15 / 21
Group source codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group source code for (X ,U ,PXU )
Assume U =Z4
Exists for large n if 1n
log |C1| ≥ log 4−min{H (U |X ),2|H (U |X )−1|+}
Optimal random code’s size: H (U )−H (U |X ) = I (X ;U )
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 15 / 21
Group source codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group source code for (X ,U ,PXU )
Assume U =Z4
Exists for large n if 1n
log |C1| ≥ log 4−min{H (U |X ),2|H (U |X )−1|+}
Optimal random code’s size: H (U )−H (U |X ) = I (X ;U )
Penalty for imposing group structure
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 15 / 21
Group Channel codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 16 / 21
Group Channel codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group channel code for (Z ,S ,PZ S )
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 16 / 21
Group Channel codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group channel code for (Z ,S ,PZ S )
Assume Z =Z4
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 16 / 21
Group Channel codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group channel code for (Z ,S ,PZ S )
Assume Z =Z4
Exists for large n if1n log |C2| ≤ log 4−max{H (Z |S),2(H (Z |S)−H ([Z ]1|S))}
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 16 / 21
Group Channel codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group channel code for (Z ,S ,PZ S )
Assume Z =Z4
Exists for large n if1n log |C2| ≤ log 4−max{H (Z |S),2(H (Z |S)−H ([Z ]1|S))}
Optimal random code’s size: log 4−H (Z |S)
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 16 / 21
Group Channel codes: Example
Consider Z4 = {0,1,2,3}
One non-trivial subgroup 2Z4 = {0,2}
Good group channel code for (Z ,S ,PZ S )
Assume Z =Z4
Exists for large n if1n log |C2| ≤ log 4−max{H (Z |S),2(H (Z |S)−H ([Z ]1|S))}
Optimal random code’s size: log 4−H (Z |S)
Penalty for presence of non-trivial subgroups
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 16 / 21
Coding Strategy
Nested group codes C2 <C11,C12
yn
C11
C12
vn
un
xn
H2un
H2vn
H2zn
1n log |C11| ≥ log 4−
min{H (U |X ),2|H (U |X )−1|+}
1n
log |C12| ≥ log 4−
min{H (V |Y ),2|H (V |Y )−1|+}
1n log |C2| ≤ log 4−
max{H (Z ),2(H (Z )−H ([Z ]1))}
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 17 / 21
Achievable Rates
The set of tuples (R1,R2,D) that satisfy
R1 ≥ max{H (Z ),2(H (Z )−H ([Z ]1))}−min{H (U |X ),2|H (U |X )−1|+}
R2 ≥ max{H (Z ),2(H (Z )−H ([Z ]1))}−min{H (V |Y ),2|H (V |Y )−1|+}
D ≥ Ed (X ,Y ,G(U ,V ))
are achievable.
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 18 / 21
Achievable Rates
The set of tuples (R1,R2,D) that satisfy
R1 ≥ max{H (Z ),2(H (Z )−H ([Z ]1))}−min{H (U |X ),2|H (U |X )−1|+}
R2 ≥ max{H (Z ),2(H (Z )−H ([Z ]1))}−min{H (V |Y ),2|H (V |Y )−1|+}
D ≥ Ed (X ,Y ,G(U ,V ))
are achievable.
More general rate region possible
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 18 / 21
Group codes vs Berger-Tung based coding
0.02 0.04 0.06 0.08 0.1 0.12 0.14
1.2
1.4
1.6
1.8
Distortion D
Comparison of the two lower convex envelopes
Sum
Rat
e R
1 + R
2 Berger−Tung based coding schemeGroup code based coding scheme
Reconstruction of XOR with Hamming distortion
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 19 / 21
Group codes vs Berger-Tung based coding
0.02 0.04 0.06 0.08 0.1 0.12 0.14
1.2
1.4
1.6
1.8
Distortion D
Comparison of the two lower convex envelopes
Sum
Rat
e R
1 + R
2 Berger−Tung based coding schemeGroup code based coding scheme
Reconstruction of XOR with Hamming distortion
Implies Berger-Tung bound not tight for more than 2 users
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 19 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
Lossless and lossy compression using abelian group codes - achievable
rates
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
Lossless and lossy compression using abelian group codes - achievable
rates
Nested linear codes achieve Shannon rate-distortion bound for
arbitrary discrete sources and additive distortions
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
Lossless and lossy compression using abelian group codes - achievable
rates
Nested linear codes achieve Shannon rate-distortion bound for
arbitrary discrete sources and additive distortions
Nested linear codes recover known rate regions of
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
Lossless and lossy compression using abelian group codes - achievable
rates
Nested linear codes achieve Shannon rate-distortion bound for
arbitrary discrete sources and additive distortions
Nested linear codes recover known rate regions of
Berger-Tung problem
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
Lossless and lossy compression using abelian group codes - achievable
rates
Nested linear codes achieve Shannon rate-distortion bound for
arbitrary discrete sources and additive distortions
Nested linear codes recover known rate regions of
Berger-Tung problem
Wyner-Ziv problem, Wyner-Ahlswede-Korner problem
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
Lossless and lossy compression using abelian group codes - achievable
rates
Nested linear codes achieve Shannon rate-distortion bound for
arbitrary discrete sources and additive distortions
Nested linear codes recover known rate regions of
Berger-Tung problem
Wyner-Ziv problem, Wyner-Ahlswede-Korner problem
Yeung-Berger problem
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Other results
Existence of “good” nested lattices - arbitrary finite levels of nesting
Lossless and lossy compression using abelian group codes - achievable
rates
Nested linear codes achieve Shannon rate-distortion bound for
arbitrary discrete sources and additive distortions
Nested linear codes recover known rate regions of
Berger-Tung problem
Wyner-Ziv problem, Wyner-Ahlswede-Korner problem
Yeung-Berger problem
Slepian-Wolf problem, Korner-Marton problem
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 20 / 21
Future research directions
Extension of lattice coding schemes to arbitrary continuous sources
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 21 / 21
Future research directions
Extension of lattice coding schemes to arbitrary continuous sources
Construction of practical group codes using LDPC/LDGM codes with
iterative decoding
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 21 / 21
Future research directions
Extension of lattice coding schemes to arbitrary continuous sources
Construction of practical group codes using LDPC/LDGM codes with
iterative decoding
Use of structured codes in multi-terminal channel coding problems
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 21 / 21
Future research directions
Extension of lattice coding schemes to arbitrary continuous sources
Construction of practical group codes using LDPC/LDGM codes with
iterative decoding
Use of structured codes in multi-terminal channel coding problems
Connections between information theory (typicality) and abstract
algebra (subgroups)
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 21 / 21
Future research directions
Extension of lattice coding schemes to arbitrary continuous sources
Construction of practical group codes using LDPC/LDGM codes with
iterative decoding
Use of structured codes in multi-terminal channel coding problems
Connections between information theory (typicality) and abstract
algebra (subgroups)
Codes over non-abelian groups
D. Krithivasan & S.S. Pradhan (U of M) Algebraic Structures for DSC ITA 2009 21 / 21