Threshold Saturation for Nonbinary Spatially-Coupled LDPC Codes on the Binary
Erasure Channel
Iryna Andriyanova† and Alexandre Graell i Amat‡
†ETIS laboratory, ENSEA/University of Cergy-Pontoise/CNRS
‡Department of Signals and Systems, Chalmers University of Technology, Gothenburg, Sweden
Journ´ees Codage&Cryptographie Les Sept Eaux, March 25, 2014
Background
I Spatially-Coupled LDPC (SC-LDPC) codes are codes with a low-density parity-check matrix, the non-zero elements in which are located in the diagonal band [FZ99].
I (Iterative) Belief Propagation (BP) algorithm is used for decoding.
I SC-LDPC codes showoutstanding asymptotic performancefor lots of channels and communication problems.
I For the binary erasure channel (BEC), the BP threshold of a binary SC-LDPC ensemble achieves theoptimal MAP threshold of the underlying LDPC ensemble (threshold saturation)[KRU11]. Extended to BMS channels [KMRU10].
[FZ99] J. Felstrom and K. Zigangirov, “Time-varying periodic convolutional codes with low-density parity-check matrix,” IEEE Trans. Inform. Theory, vol.45, no.6, pp.2181,2191, Sep 1999.
[KRU11] S. Kudekar, T. J. Richardson, and R. L. Urbanke, “Threshold saturation via spatial coupling: Why convolutional LDPC ensembles perform so well over the BEC,” IEEE Trans. Inform. Theory, vol. 57, pp. 803–834, Feb. 2011.
[KMRU10] S. Kudekar, C. M´easson, T. J. Richardson, and R. L. Urbanke, “Threshold saturation on BMS channels via spatial coupling,” 6th Int. Symp. Turbo Codes & Iterative Inform. Process., 2010.
Background
I Spatially-Coupled LDPC (SC-LDPC) codes are codes with a low-density parity-check matrix, the non-zero elements in which are located in the diagonal band [FZ99].
I (Iterative) Belief Propagation (BP) algorithm is used for decoding.
I SC-LDPC codes showoutstanding asymptotic performancefor lots of channels and communication problems.
I For the binary erasure channel (BEC), the BP threshold of a binary SC-LDPC ensemble achieves theoptimal MAP threshold of the underlying LDPC ensemble (threshold saturation)[KRU11]. Extended to BMS channels [KMRU10].
[FZ99] J. Felstrom and K. Zigangirov, “Time-varying periodic convolutional codes with low-density parity-check matrix,” IEEE Trans. Inform. Theory, vol.45, no.6, pp.2181,2191, Sep 1999.
[KRU11] S. Kudekar, T. J. Richardson, and R. L. Urbanke, “Threshold saturation via spatial coupling: Why convolutional LDPC ensembles perform so well over the BEC,” IEEE Trans. Inform. Theory, vol. 57, pp. 803–834, Feb. 2011.
[KMRU10] S. Kudekar, C. M´easson, T. J. Richardson, and R. L. Urbanke, “Threshold saturation on BMS channels via spatial coupling,” 6th Int. Symp. Turbo Codes & Iterative Inform. Process., 2010.
Background
I An alternative proof technique based onpotential functionsfor coupled scalar and vector recursions [YJNP12].
I DE (BP decoding, vector recursion, BEC(ε)): x(`+1)=f(g(x(`));ε),
wheref(y)andg(x)are the variable and check node updates.
Converges tox(∞)= (0, . . . ,0)forεbelow the BP thresholdεBP.
Vector admissible system
I The functionsf(y;ε)andg(x)non-decreasinginyandx;
I f(y;ε)isdifferentiableinyandg(x)is twice differentiable inx;
I f(0;ε) =f(y; 0) =g(0) =0;
I f(y;ε)isstrictly increasingwithε.
[YJNP12a] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled scalar recursions,” 7th Int. Symp. Turbo Codes & Iterative Inform. Process., 2012. Extended version:
http://arxiv.org/abs/1309.7910
[YJNP12b] —–, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
I An alternative proof technique based onpotential functionsfor coupled scalar and vector recursions [YJNP12].
I DE (BP decoding, vector recursion, BEC(ε)):
x(`+1)=f(g(x(`));ε),
wheref(y)andg(x)are the variable and check node updates.
Converges tox(∞)= (0, . . . ,0)forεbelow the BP thresholdεBP.
Vector admissible system
I The functionsf(y;ε)andg(x)non-decreasinginyandx;
I f(y;ε)isdifferentiableinyandg(x)is twice differentiable inx;
I f(0;ε) =f(y; 0) =g(0) =0;
I f(y;ε)isstrictly increasingwithε.
[YJNP12a] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled scalar recursions,” 7th Int. Symp. Turbo Codes & Iterative Inform. Process., 2012. Extended version:
http://arxiv.org/abs/1309.7910
[YJNP12b] —–, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
I An alternative proof technique based onpotential functionsfor coupled scalar and vector recursions [YJNP12].
I DE (BP decoding, vector recursion, BEC(ε)):
x(`+1)=f(g(x(`));ε),
wheref(y)andg(x)are the variable and check node updates.
Converges tox(∞)= (0, . . . ,0)forεbelow the BP thresholdεBP.
Vector admissible system
I The functionsf(y;ε)andg(x)non-decreasinginyandx;
I f(y;ε)isdifferentiableinyandg(x)is twice differentiable inx;
I f(0;ε) =f(y; 0) =g(0) =0;
I f(y;ε)isstrictly increasingwithε.
[YJNP12a] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled scalar recursions,” 7th Int. Symp. Turbo Codes & Iterative Inform. Process., 2012. Extended version:
http://arxiv.org/abs/1309.7910
[YJNP12b] —–, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
I An alternative proof technique based onpotential functionsfor coupled scalar and vector recursions [YJNP12].
I DE (BP decoding, vector recursion, BEC(ε)):
x(`+1)=f(g(x(`));ε),
wheref(y)andg(x)are the variable and check node updates.
Converges tox(∞)= (0, . . . ,0)forεbelow the BP thresholdεBP.
Vector admissible system
I The functionsf(y;ε)andg(x)non-decreasinginyandx;
I f(y;ε)isdifferentiableinyandg(x)is twice differentiable inx;
I f(0;ε) =f(y; 0) =g(0) =0;
I f(y;ε)isstrictly increasingwithε.
[YJNP12a] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled scalar recursions,” 7th Int. Symp. Turbo Codes & Iterative Inform. Process., 2012. Extended version:
http://arxiv.org/abs/1309.7910
[YJNP12b] —–, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
I Potential function [YJNP12b]:
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε),
whereF andGare scalar functions that satisfyF0(y;ε) =f(y;ε)D, and G0(x) =g(x)D, withDa positive diagonal matrix.
Afixed pointof the DE (x=f(g(x);ε)) corresponds to astationary pointof the corresponding potential function (U(x;ε)).
I TheBP threshold: εBP= supε(ε∈[0,1]|U0(x;ε)>0, ∀x∈ X).
I Thepotential threshold: ε∗= supε ε∈(εBP,1]|∆E(ε)≥0, ∀x∈ X ,
where∆E(ε) = infx∈X \U0(x)U(x;ε)for someε,εBP≤ε≤ε∗, and U0(ε) ={x∈ X |x∞=0}is the basin of attraction forx∞=0.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
I Potential function [YJNP12b]:
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε),
whereF andGare scalar functions that satisfyF0(y;ε) =f(y;ε)D, and G0(x) =g(x)D, withDa positive diagonal matrix.
Afixed pointof the DE (x=f(g(x);ε)) corresponds to astationary pointof the corresponding potential function (U(x;ε)).
I TheBP threshold: εBP= supε(ε∈[0,1]|U0(x;ε)>0, ∀x∈ X).
I Thepotential threshold: ε∗= supε ε∈(εBP,1]|∆E(ε)≥0, ∀x∈ X ,
where∆E(ε) = infx∈X \U0(x)U(x;ε)for someε,εBP≤ε≤ε∗, and U0(ε) ={x∈ X |x∞=0}is the basin of attraction forx∞=0.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
I Potential function [YJNP12b]:
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε),
whereF andGare scalar functions that satisfyF0(y;ε) =f(y;ε)D, and G0(x) =g(x)D, withDa positive diagonal matrix.
Afixed pointof the DE (x=f(g(x);ε)) corresponds to astationary pointof the corresponding potential function (U(x;ε)).
I TheBP threshold: εBP= supε(ε∈[0,1]|U0(x;ε)>0, ∀x∈ X).
I Thepotential threshold: ε∗= supε ε∈(εBP,1]|∆E(ε)≥0, ∀x∈ X ,
where∆E(ε) = infx∈X \U0(x)U(x;ε)for someε,εBP≤ε≤ε∗, and U0(ε) ={x∈ X |x∞=0}is the basin of attraction forx∞=0.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
I Potential function [YJNP12b]:
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε),
whereF andGare scalar functions that satisfyF0(y;ε) =f(y;ε)D, and G0(x) =g(x)D, withDa positive diagonal matrix.
Afixed pointof the DE (x=f(g(x);ε)) corresponds to astationary pointof the corresponding potential function (U(x;ε)).
I TheBP threshold: εBP= supε(ε∈[0,1]|U0(x;ε)>0, ∀x∈ X).
I Thepotential threshold: ε∗= supε ε∈(εBP,1]|∆E(ε)≥0, ∀x∈ X ,
where∆E(ε) = infx∈X \U0(x)U(x;ε)for someε,εBP≤ε≤ε∗, and U0(ε) ={x∈ X |x∞=0}is the basin of attraction forx∞=0.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
Proof of threshold saturation [YJNP12b]:
1. DefineU(x;ε)for the (uncoupled) vector recursion.
2. Derive the potential function for the coupled vector recursion.
3. Show that, below potential threshold1, theonly fixed point of the DE is x∞=0.
1For several systems the MAP threshold and the potential thresholdε∗are identical.
I Proved threshold saturation of irregular LDPC codes for aSlepian-Wolf problem with erasures, joint decoding of irregular LDPC codes on an erasure multiple-access channel, andprotograph codes on the BEC.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
Proof of threshold saturation [YJNP12b]:
1. DefineU(x;ε)for the (uncoupled) vector recursion.
2. Derive the potential function for the coupled vector recursion.
3. Show that, below potential threshold1, theonly fixed point of the DE is x∞=0.
1For several systems the MAP threshold and the potential thresholdε∗are identical.
I Proved threshold saturation of irregular LDPC codes for aSlepian-Wolf problem with erasures, joint decoding of irregular LDPC codes on an erasure multiple-access channel, andprotograph codes on the BEC.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
Proof of threshold saturation [YJNP12b]:
1. DefineU(x;ε)for the (uncoupled) vector recursion.
2. Derive the potential function for the coupled vector recursion.
3. Show that, below potential threshold1, theonly fixed point of the DE is x∞=0.
1For several systems the MAP threshold and the potential thresholdε∗are identical.
I Proved threshold saturation of irregular LDPC codes for aSlepian-Wolf problem with erasures, joint decoding of irregular LDPC codes on an erasure multiple-access channel, andprotograph codes on the BEC.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
Proof of threshold saturation [YJNP12b]:
1. DefineU(x;ε)for the (uncoupled) vector recursion.
2. Derive the potential function for the coupled vector recursion.
3. Show that, below potential threshold1, theonly fixed point of the DE is x∞=0.
1For several systems the MAP threshold and the potential thresholdε∗are identical.
I Proved threshold saturation of irregular LDPC codes for aSlepian-Wolf problem with erasures, joint decoding of irregular LDPC codes on an erasure multiple-access channel, andprotograph codes on the BEC.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
Proof of threshold saturation [YJNP12b]:
1. DefineU(x;ε)for the (uncoupled) vector recursion.
2. Derive the potential function for the coupled vector recursion.
3. Show that, below potential threshold1, theonly fixed point of the DE is x∞=0.
1For several systems the MAP threshold and the potential thresholdε∗are identical.
I Proved threshold saturation of irregular LDPC codes for aSlepian-Wolf problem with erasures, joint decoding of irregular LDPC codes on an erasure multiple-access channel, andprotograph codes on the BEC.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Background
Proof of threshold saturation [YJNP12b]:
1. DefineU(x;ε)for the (uncoupled) vector recursion.
2. Derive the potential function for the coupled vector recursion.
3. Show that, below potential threshold1, theonly fixed point of the DE is x∞=0.
1For several systems the MAP threshold and the potential thresholdε∗are identical.
I Proved threshold saturation of irregular LDPC codes for aSlepian-Wolf problem with erasures, joint decoding of irregular LDPC codes on an erasure multiple-access channel, andprotograph codeson the BEC.
[YJNP12b] A. Yedla, Y.-Y. Jian, P. Nguyen, and H. Pfister, “A simple proof of threshold saturation for coupled vector recursions,” 2012 IEEE Inform. Theory Workshop. Extended version: http://arxiv.org/abs/1208.4080
Motivation
I Threshold saturation for binary SC-LDPC codes so far.
I Nonbinary SC-LDPC codes:
I Construction method for nonbinary SC-LDPC codes [UKS11].
I Thethreshold saturationalso occurs for nonbinary SC-LDPC codes. Contrary to uncoupled ensembles, the BP threshold of nonbinary SC-LDPC codesimproves with field size and tends to the Shannon limit[PGiA13].
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4 0.42 0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6
hBP
ε m
Figure: BP EXIT functions for nonbinary(3,6)SC-LDPC code ensembles over GFm2.
[UKS11] H. Uchikawa, K. Kasai, and K. Sakaniwa, “Design and performance of rate-compatible non-binary LDPC convolutional codes,” 2011. [Online]. Available: http://arxiv.org/abs/1010.0060
[PGiA13] A. Piemontese, A. Graell i Amat, and G. Colavolpe, “Nonbinary spatially-coupled LDPC codes on the binary erasure channel,”IEEE Int. Conf. Commun., ICC’2013.
Motivation
I Threshold saturation for binary SC-LDPC codes so far.
I Nonbinary SC-LDPC codes:
I Construction method for nonbinary SC-LDPC codes [UKS11].
I Thethreshold saturationalso occurs for nonbinary SC-LDPC codes. Contrary to uncoupled ensembles, the BP threshold of nonbinary SC-LDPC codesimproves with field size and tends to the Shannon limit[PGiA13].
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4 0.42 0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6
hBP
ε m
Figure: BP EXIT functions for nonbinary(3,6)SC-LDPC code ensembles over GFm2.
[UKS11] H. Uchikawa, K. Kasai, and K. Sakaniwa, “Design and performance of rate-compatible non-binary LDPC convolutional codes,” 2011. [Online]. Available: http://arxiv.org/abs/1010.0060
[PGiA13] A. Piemontese, A. Graell i Amat, and G. Colavolpe, “Nonbinary spatially-coupled LDPC codes on the binary erasure channel,”IEEE Int. Conf. Commun., ICC’2013.
Motivation
I Threshold saturation for binary SC-LDPC codes so far.
I Nonbinary SC-LDPC codes:
I Construction method for nonbinary SC-LDPC codes [UKS11].
I Thethreshold saturationalso occurs for nonbinary SC-LDPC codes.
Contrary to uncoupled ensembles, the BP threshold of nonbinary SC-LDPC codesimproves with field size and tends to the Shannon limit[PGiA13].
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4 0.42 0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6
hBP
ε m
Figure:BP EXIT functions for nonbinary(3,6)SC-LDPC code ensembles over GFm2.
[UKS11] H. Uchikawa, K. Kasai, and K. Sakaniwa, “Design and performance of rate-compatible non-binary LDPC convolutional codes,” 2011. [Online]. Available: http://arxiv.org/abs/1010.0060
[PGiA13] A. Piemontese, A. Graell i Amat, and G. Colavolpe, “Nonbinary spatially-coupled LDPC codes on the binary erasure channel,”IEEE Int. Conf. Commun., ICC’2013.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andG exist.
I Proof of threshold saturation.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andG exist.
I Proof of threshold saturation.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andG exist.
I Proof of threshold saturation.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andG exist.
I Proof of threshold saturation.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD
=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andG exist.
I Proof of threshold saturation.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andG exist.
I Proof of threshold saturation.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andGexist.
I Proof of threshold saturation.
Contributions
In this talk
I Proof of threshold saturation for nonbinary SC-LDPC codes to the potential thresholdfor transmission over the BEC, using the technique based on potential functions.
I (dv,dc,m)and(dv,dc,m,L,w)ensembles over the general linear group.
Contributions
I We prove themonotonicity of the variable node and check node updates for nonbinary LDPC codes, and theexistence of a fixed point in the DE.
I Existence and computation ofU(x;ε): Can be obtained by finding the functionsF andGas the solution of asystem of linear equations.
I Anecessary conditionon the existence ofF andG(thusU(x;ε)), assuming diagonalD=⇒Not verifiedfor nonbinary codes!
I AsymmetricDis sufficientso thatF andGexist.
I Proof of threshold saturation.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Code symbolsα0, . . . , α2m−1 ∈ S=GFm2
=⇒The messages exchanged in the BP decoding areprobability vectors of length2m,(p0, . . . ,p2m−1), wherepi is theprobability that the symbol isαi.
I A message(p0, . . . ,p2m−1)has dimensionk if it has2k non-zero entries, i.e., the symbol is known to beone out of2k possible symbols.
I For theBECthe non-zero entries of a message are all equal, and the message is equivalent to asubspace of GFm2 of dimensionk.
I DE simplifies to exchange ofprobability vectors of lengthm+ 1,
(˜p0, . . . ,˜pm), where˜pi is theprobability that the message has dimensioni.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Code symbolsα0, . . . , α2m−1 ∈ S=GFm2 =⇒The messages exchanged in the BP decoding areprobability vectors of length2m,(p0, . . . ,p2m−1), wherepi is theprobability that the symbol isαi.
I A message(p0, . . . ,p2m−1)has dimensionk if it has2k non-zero entries, i.e., the symbol is known to beone out of2k possible symbols.
I For theBECthe non-zero entries of a message are all equal, and the message is equivalent to asubspace of GFm2 of dimensionk.
I DE simplifies to exchange ofprobability vectors of lengthm+ 1,
(˜p0, . . . ,˜pm), where˜pi is theprobability that the message has dimensioni.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Code symbolsα0, . . . , α2m−1 ∈ S=GFm2 =⇒The messages exchanged in the BP decoding areprobability vectors of length2m,(p0, . . . ,p2m−1), wherepi is theprobability that the symbol isαi.
I A message(p0, . . . ,p2m−1)has dimensionk if it has2k non-zero entries, i.e., the symbol is known to beone out of2k possible symbols.
I For theBECthe non-zero entries of a message are all equal, and the message is equivalent to asubspace of GFm2 of dimensionk.
I DE simplifies to exchange ofprobability vectors of lengthm+ 1,
(˜p0, . . . ,˜pm), where˜pi is theprobability that the message has dimensioni.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Code symbolsα0, . . . , α2m−1 ∈ S=GFm2 =⇒The messages exchanged in the BP decoding areprobability vectors of length2m,(p0, . . . ,p2m−1), wherepi is theprobability that the symbol isαi.
I A message(p0, . . . ,p2m−1)has dimensionk if it has2k non-zero entries, i.e., the symbol is known to beone out of2k possible symbols.
I For theBECthe non-zero entries of a message are all equal, and the message is equivalent to asubspace of GFm2 of dimensionk.
I DE simplifies to exchange ofprobability vectors of lengthm+ 1,
(˜p0, . . . ,˜pm), where˜pi is theprobability that the message has dimensioni.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Code symbolsα0, . . . , α2m−1 ∈ S=GFm2 =⇒The messages exchanged in the BP decoding areprobability vectors of length2m,(p0, . . . ,p2m−1), wherepi is theprobability that the symbol isαi.
I A message(p0, . . . ,p2m−1)has dimensionk if it has2k non-zero entries, i.e., the symbol is known to beone out of2k possible symbols.
I For theBECthe non-zero entries of a message are all equal, and the message is equivalent to asubspace of GFm2 of dimensionk.
I DE simplifies to exchange ofprobability vectors of lengthm+ 1,
(˜p0, . . . ,˜pm), where˜pi is theprobability that the message has dimensioni.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I x`◦= (x◦0` , . . . ,x◦m` );x◦i` is the probability that a messagefrom variable nodesat iteration`hasdimensioni.
I y`◦= (y◦0` , . . . ,y◦m` );y◦i` is the probability that a message from check nodesat iteration`has dimensiondimensioni.
I Atcheck nodes, the BP decoder computes thecheck node updates (sum of the subspaces corresponding to the incoming messages),
y`◦=g◦(x`◦−1).
I Atvariable nodes, the decoder computes thevariable node updates (intersection of the subspaces of the incoming messages), corresponding to the incoming messages,
x`◦=f◦(y`◦;ε).
Background DE for nonbinary LDPC ensembles Threshold saturation Conclusion
DE for (d
v, d
c, m ) LDPC code ensembles on the BEC
• xℓ◦= (x◦0ℓ , . . . ,x◦mℓ );x◦iℓ is the probability that a messagefrom variable nodesat iterationℓhasdimensioni.
• yℓ◦= (y◦0ℓ , . . . ,yℓ◦m);yℓ◦i is the probability that a message from check nodesat iterationℓhas dimensiondimensioni.
Threshold Saturation for Nonbinary SC-LDPC Codes on the BEC | A. Graell i Amat and I. Andriyanova 8 / 16
(x◦0, . . . ,x◦m)
(y◦0, . . . ,y◦m)
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I x`◦= (x◦0` , . . . ,x◦m` );x◦i` is the probability that a messagefrom variable nodesat iteration`hasdimensioni.
I y`◦= (y◦0` , . . . ,y◦m` );y◦i` is the probability that a message from check nodesat iteration`has dimensiondimensioni.
I Atcheck nodes, the BP decoder computes thecheck node updates (sum of the subspaces corresponding to the incoming messages),
y`◦=g◦(x`◦−1).
I Atvariable nodes, the decoder computes thevariable node updates (intersection of the subspaces of the incoming messages), corresponding to the incoming messages,
x`◦=f◦(y`◦;ε).
Background DE for nonbinary LDPC ensembles Threshold saturation Conclusion
DE for (d
v, d
c, m ) LDPC code ensembles on the BEC
• xℓ◦= (x◦0ℓ , . . . ,x◦mℓ );x◦iℓ is the probability that a messagefrom variable nodesat iterationℓhasdimensioni.
• yℓ◦= (y◦0ℓ , . . . ,yℓ◦m);yℓ◦i is the probability that a message from check nodesat iterationℓhas dimensiondimensioni.
Threshold Saturation for Nonbinary SC-LDPC Codes on the BEC | A. Graell i Amat and I. Andriyanova 8 / 16
(x◦0, . . . ,x◦m)
(y◦0, . . . ,y◦m)
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I x`◦= (x◦0` , . . . ,x◦m` );x◦i` is the probability that a messagefrom variable nodesat iteration`hasdimensioni.
I y`◦= (y◦0` , . . . ,y◦m` );y◦i` is the probability that a message from check nodesat iteration`has dimensiondimensioni.
I Atcheck nodes, the BP decoder computes thecheck node updates (sum of the subspaces corresponding to the incoming messages),
y`◦=g◦(x`◦−1).
I Atvariable nodes, the decoder computes thevariable node updates (intersection of the subspaces of the incoming messages), corresponding to the incoming messages,
x`◦=f◦(y`◦;ε).
Background DE for nonbinary LDPC ensembles Threshold saturation Conclusion
DE for (d
v, d
c, m ) LDPC code ensembles on the BEC
• xℓ◦= (x◦0ℓ , . . . ,x◦mℓ );x◦iℓ is the probability that a messagefrom variable nodesat iterationℓhasdimensioni.
• yℓ◦= (y◦0ℓ , . . . ,yℓ◦m);yℓ◦i is the probability that a message from check nodesat iterationℓhas dimensiondimensioni.
Threshold Saturation for Nonbinary SC-LDPC Codes on the BEC | A. Graell i Amat and I. Andriyanova 8 / 16
(x◦0, . . . ,x◦m)
(y◦0, . . . ,y◦m)
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Thefixed-point DE equationforx◦=x∞◦ is x◦=f◦(g◦(x◦);ε).
Decoding issuccessfulwhen itconverges tox∞◦ = (1,0, . . . ,0).
I f◦andg◦arenot monotone! =⇒Not an admissible system.
I Does DE for nonbinary LDPC codes converge to afixed point?
Idea
Rewrite the DE equation using complementary cumulative distribution function (CCDF) vectors.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Thefixed-point DE equationforx◦=x∞◦ is x◦=f◦(g◦(x◦);ε).
Decoding issuccessfulwhen itconverges tox∞◦ = (1,0, . . . ,0).
I f◦andg◦arenot monotone! =⇒Not an admissible system.
I Does DE for nonbinary LDPC codes converge to afixed point?
Idea
Rewrite the DE equation using complementary cumulative distribution function (CCDF) vectors.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Thefixed-point DE equationforx◦=x∞◦ is x◦=f◦(g◦(x◦);ε).
Decoding issuccessfulwhen itconverges tox∞◦ = (1,0, . . . ,0).
I f◦andg◦arenot monotone!
=⇒Not an admissible system.
I Does DE for nonbinary LDPC codes converge to afixed point?
Idea
Rewrite the DE equation using complementary cumulative distribution function (CCDF) vectors.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Thefixed-point DE equationforx◦=x∞◦ is x◦=f◦(g◦(x◦);ε).
Decoding issuccessfulwhen itconverges tox∞◦ = (1,0, . . . ,0).
I f◦andg◦arenot monotone! =⇒Not an admissible system.
I Does DE for nonbinary LDPC codes converge to afixed point?
Idea
Rewrite the DE equation using complementary cumulative distribution function (CCDF) vectors.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Thefixed-point DE equationforx◦=x∞◦ is x◦=f◦(g◦(x◦);ε).
Decoding issuccessfulwhen itconverges tox∞◦ = (1,0, . . . ,0).
I f◦andg◦arenot monotone! =⇒Not an admissible system.
I Does DE for nonbinary LDPC codes converge to afixed point?
Idea
Rewrite the DE equation using complementary cumulative distribution function (CCDF) vectors.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
I Thefixed-point DE equationforx◦=x∞◦ is x◦=f◦(g◦(x◦);ε).
Decoding issuccessfulwhen itconverges tox∞◦ = (1,0, . . . ,0).
I f◦andg◦arenot monotone! =⇒Not an admissible system.
I Does DE for nonbinary LDPC codes converge to afixed point?
Idea
Rewrite the DE equation using complementary cumulative distribution function (CCDF) vectors.
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
Definition
Given a probability vectorx◦, define the CCDF vectorx= (x1, . . . ,xm), where xi=Pm
k=ix◦k. We also definexm+1= 0. Then, it follows that x◦i=xi−xi+1. Note also thatx0= 1.
I We definenew vector functionsf = (f0, . . . ,fm)andg= (g0, . . . ,gm), fi=
m
X
k=i
f◦k(y◦;ε) =
m
X
k=i
f◦k(y−1−y;ε), gi=
m
X
k=i
g◦k(x◦) =
m
X
k=i
g◦k(x−1−x), wherex−1= (1,x1, . . . ,xm−1).
DE for (d
v, d
c, m) LDPC code ensembles on the BEC
Definition
Given a probability vectorx◦, define the CCDF vectorx= (x1, . . . ,xm), where xi=Pm
k=ix◦k. We also definexm+1= 0. Then, it follows that x◦i=xi−xi+1. Note also thatx0= 1.
I We definenew vector functionsf = (f0, . . . ,fm)andg= (g0, . . . ,gm), fi=
m
X
k=i
f◦k(y◦;ε) =
m
X
k=i
f◦k(y−1−y;ε), gi=
m
X
k=i
g◦k(x◦) =
m
X
k=i
g◦k(x−1−x), wherex−1= (1,x1, . . . ,xm−1).
Convergence of the DE for nonbinary LDPC codes
I Using CCDF vectors, the DE equation can be written in anequivalent formas
x=f(g(x);ε). (1)
Theorem 1
The functionsf(y;ε)andg(x)are increasing inyandx. Corollary
The DE for regular nonbinary LDPC codes given by (1)converges to a fixed point.
Successful decoding
Successful decoding corresponds to convergence of the DE equation to the fixed pointx∞=0= (0,0, . . . ,0).
Convergence of the DE for nonbinary LDPC codes
I Using CCDF vectors, the DE equation can be written in anequivalent formas
x=f(g(x);ε). (1)
Theorem 1
The functionsf(y;ε)andg(x)are increasing inyandx.
Corollary
The DE for regular nonbinary LDPC codes given by (1)converges to a fixed point.
Successful decoding
Successful decoding corresponds to convergence of the DE equation to the fixed pointx∞=0= (0,0, . . . ,0).
Convergence of the DE for nonbinary LDPC codes
I Using CCDF vectors, the DE equation can be written in anequivalent formas
x=f(g(x);ε). (1)
Theorem 1
The functionsf(y;ε)andg(x)are increasing inyandx.
Corollary
The DE for regular nonbinary LDPC codes given by (1)converges to a fixed point.
Successful decoding
Successful decoding corresponds to convergence of the DE equation to the fixed pointx∞=0= (0,0, . . . ,0).
Convergence of the DE for nonbinary LDPC codes
I Using CCDF vectors, the DE equation can be written in anequivalent formas
x=f(g(x);ε). (1)
Theorem 1
The functionsf(y;ε)andg(x)are increasing inyandx.
Corollary
The DE for regular nonbinary LDPC codes given by (1)converges to a fixed point.
Successful decoding
Successful decoding corresponds to convergence of the DE equation to the fixed pointx∞=0= (0,0, . . . ,0).
Potential Function
I x=f(g(x);ε)is avector admissible system.
Can define a potential function
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε). However...
U(x;ε)with diagonalDdoes not exist! =⇒Conditions on the existence of U(x;ε).
Potential Function
I x=f(g(x);ε)is avector admissible system. Can define a potential function
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε).
However...
U(x;ε)with diagonalDdoes not exist! =⇒Conditions on the existence of U(x;ε).
Potential Function
I x=f(g(x);ε)is avector admissible system. Can define a potential function
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε).
However...
U(x;ε)with diagonalDdoes not exist! =⇒Conditions on the existence of U(x;ε).
Properties of D and calculation of U ( x ; ε)
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε)
Theorem 4 [U(x;ε)for the (dv,dc,m)ensemble]
F(y;ε)andG(x)exist (henceU(x;ε)exists)if there exist sets of values{djs}, {ϕ(i1,...,im)}and{µ(k1,...km)}that satisfy the following equations,
(isϕ(i1,...,is,...,im)=Pm j=1djsϕ(j)(i
1,...is−1,...,im)(ε) ktµ(k1,...,kt,...,km)=Pm
j=1djtγ(j)(k
1,...,kt−1,...km)
, (2)
for all possiblem-tuples(i1, . . . ,im)and(k1, . . . ,km)and allisandkt varying from1tom. The coefficientsϕ’s andγ’s in (2) are given by
ϕ(j)(i
1,...,im)(ε) = coeff(fj(x;ε),x1i1· · ·xmim), γ(k(j)
1,,...km)= coeff(gj(x),x1i1· · ·xmim).
Similar result for the coupled case. General resultfor coupled vector systems.
Properties of D and calculation of U ( x ; ε)
U(x;ε) =g(x)DxT−G(x)−F(g(x);ε)
Theorem 4 [U(x;ε)for the (dv,dc,m)ensemble]
F(y;ε)andG(x)exist (henceU(x;ε)exists)if there exist sets of values{djs}, {ϕ(i1,...,im)}and{µ(k1,...km)}that satisfy the following equations,
(isϕ(i1,...,is,...,im)=Pm j=1djsϕ(j)(i
1,...is−1,...,im)(ε) ktµ(k1,...,kt,...,km)=Pm
j=1djtγ(j)(k
1,...,kt−1,...km)
, (2)
for all possiblem-tuples(i1, . . . ,im)and(k1, . . . ,km)and allisandkt varying from1tom. The coefficientsϕ’s andγ’s in (2) are given by
ϕ(j)(i
1,...,im)(ε) = coeff(fj(x;ε),x1i1· · ·xmim), γ(k(j)
1,,...km)= coeff(gj(x),x1i1· · ·xmim).
Similar result for the coupled case. General resultfor coupled vector systems.