ITERATIVE DECODER FOR DECODING A CODE COMPOSED OF AT LEAST TWO CONSTRAINT NODES

Abstract
An iterative decoder, comprises:
Description
TECHNICAL FIELD

Various example embodiments relate to iterative decoding of a code composed of at least two constraint nodes, such as turbocodes or LDPC (Low Density Parity Check) codes.


BACKGROUND

During the past two decades, Turbocodes and Low Density Parity Check (LDPC) codes have emerged as a crucial component of many communication standards. These codes can offer performance close to the Shannon limit for error rates above the decoder's error floor.


Particularly, LDPC codes are widely used in communications standards like DVB-S2, DVB-S2X, IEEE 802.3, etc. LDPC codes can be efficiently decoded by Message-Passing (MP) algorithms that use a Tanner graph representation of the LDPC codes. The Belief-Propagation (BP) decoder has excellent decoding performance in the waterfall region but at a cost of a high computational complexity. The Min-Sum (MS) and Offset Min-Sum (OMS) decoders are simplified version of the BP that are much less complex but at a cost of a slight decoding performance degradation in the waterfall region. Reducing the bit-size representation of message is a technique that further reduces the complexity of the decoder, again, at a cost of performance loss.


In the last fifteen years, the quantization problem has been extensively studied over the Binary-Input Additive White Gaussian Noise (BI-AWGN) channel. From these works, it can be concluded that 6 bits of quantization gives almost optimal performance. It can also be observed that all paper on finite precision use the “no-decision value”, i.e., a Log-Likelihood Ratio (LLR) equal to 0 (without defined sign, thus) in their messages representation. This is also true for the recent work on Non-surjective Finite Alphabet Iterative Decoders (NS-FAIDs) which provides a unified framework for several MS-based decoders like Normalized MS (NMS), OMS, Partially OMS.


In “Improved low-complexity low-density parity-check decoding”, Cuiz et al., a weighted bit-flipping (WBF) algorithm is proposed called improved modified WBF (IMWBF). The IMWBF uses the bit of sign plus an extra bit to give a weight to the message, thus, exchange messages are on two bits. However, like all Bit Flipping algorithm, the same message it broadcasted by a given variable node to its associated check nodes.


In “Non-surjective finite alphabet iterative decoders”, Nguyen-Ly Thien Truong et al.


and “Finite alphabet Iterative Decoders-Part I: Decoding Beyond Belief Propagation on the Binary Symmetric Channel”, Shiva Kumar Planjery et al., the authors propose an algorithm named Non-Surjective Finite Alphabet Iterative Decoder (NS-FAID). A characteristic of this algorithm is that the 0 value is encoded in all the messages it exchanges.


SUMMARY

However, from all these work, it appears there is still a need for a decoder having good performance with a relative low complexity, i.e. a decoder which works well with few quantization bits.


In a first example embodiment, an iterative decoder is configured for decoding a code composed of at least two constraint nodes and having a codeword length N, said decoder comprising:

    • N variable nodes (VNs) vn, n=1 . . . N, each variable node vn being configured to receive a log-likelihood ratio LLR In of the channel decoded bit n of the codeword to be decoded, said LLR In being defined on a alphabet As of q quantization bits, q being an integer equal or greater than 2;
    • M constraint nodes (CNs) cm, m=1 . . . M, 2≤M<N;
    • the variable nodes and the constraint nodes being the nodes of a Tanner graph; variable nodes and constraint nodes being configured to exchange messages along edges of the Tanner graph;
    • each variable node vn being configured for estimating the value yn of the nth bit of the codeword to be decoded and for sending messages mvn→cm, messages belonging to As, to the connected constraint nodes cm, the set of connected constraint nodes being noted V(vn), defining the degree dv of the variable node vn as the size of the set V(vn), and V(vn)\{cm} being the set V(vn) except the constraint node cm, and then,
    • each constraint nodes cm being configured for testing the received message contents with predetermined constraints and sending messages mcm→vn, messages belonging to an alphabet Ac, to the connected variable nodes vn;
    • the message emissions being repeated until the decoding of the codeword is achieved successfully or a predetermined number of iteration is reached; and
    • the LLR In and the messages mvn→cm and mcm→vn are coded according to a sign-and-magnitude code in the alphabets As and Ac symmetric around zero, As={−Nq, . . . , −1, −0, +0, +1, . . . ,+Nq}, Ac={−Nq′, . . . , −1, −0, +0, +1, . . . ,+Nq′} in which the sign indicates the estimated bit value and the magnitude represents its reliability; and
    • each variable node vn, for each iteration l, is configured to compute:
    • sign-preserving factors, sum of the ponderated sign of LLR In and the sum of the signs of all messages, except for the message coming from cm, received from the connected constraint nodes at iteration /:
      • custom-character=ξ×sign (In)+Σc∈V(vn)\{cm} sign (custom-character) where ξis a positive or a null integer;
    • messages to connected constraint nodes cm for iteration l+1 computed in two steps
      • custom-character=ln+½×custom-characterc∈V(vn)\{cm} (custom-character) and
      • custom-character=(sign(custom-character), custom-character(floor(abs(custom-character)))) where S is a function from the set of value that can take floor (abs(custom-character)) to the set As.


Advantageously, the decoder uses an implementation that always preserve the sign of the messages. Consequently, it can achieve the same convergence threshold as a classical decoder using a representation with more bits.


This embodiment may comprises other features, alone or in combination, such as:

    • ξ equal to 0 if dv=2, equal to 1 if dv>2 and dv is odd, and equal to 2 if dv>2 and dv is even;
    • the constraints applied by the constraint nodes are parity check, convolution code or any block code;
    • the codeword is coded by LDPC and the constraint nodes are check nodes;
    • the function S is defined as S(x)=min(max(x−λ(x), 0),+Nq), where λ(x) is an integer offset value depending on the value of x;
    • λ(x) is a random variable that can take it value in a predefined set of values according to a predefined law;
    • alphabets As and Ac are identical; and/or
    • the means comprises
    • at least one processor; and
    • at least one memory including computer program code, the at least one memory and computer program code configured to, with the at least one processor, cause the performance of the decoder.


In a second example embodiment, an iterative decoding method for decoding a code composed of at least two constraint nodes and having a codeword length N, by an iterative decoder as disclosed here above, comprises:

    • reception and storage of the LLR In by the variable node vn, n=1 . . . N;
    • emission of messages mvn→cm;
    • verification of the constraints by the constraint nodes cm;
    • emission of messages mcm→vn containing a possible candidate value for the variable n solving the constraints;
    • computation by the variable nodes of custom-character=ξ×sign (In)+Σc∈V(vn)\{cm} sign(custom-character) where ξ is a positive or a null integer, and of messages to connected constraint nodes cm for iteration l+1 computed in two steps
    • custom-character=In+½×custom-characterc∈V(vn)\{cm} (custom-character) and
    • custom-character=(sign(custom-character), custom-character(floor(abs(custom-character)))) where S is a function from the set of value that can take floor (abs(custom-character)) to the set As; and.
    • iteration from emission of messages mvn→cm to their computation for the next iteration until the codeword is decoded or the number of iteration reaches a predetermined count.


In a third example embodiment, a digital data storage medium is encoding a machine-executable program of instructions to perform the method disclosed here above.





BRIEF DESCRIPTION OF THE FIGURES

Some embodiments are now described, by way of example only, and with reference to the accompagnying drawings, in which:


The FIG. 1 schematically illustrates an iterative decoder;


The FIG. 2 illustrates a flow chart of an iterative decoding method;


The FIG. 3 schematically illustrate a mapping used for the noise model γ;


The FIG. 4 illustrates the FER performance of MS and SP-MS decoders for (3, 6)-regular LDPC code;


The FIG. 5 illustrates the FER performance of MS and SP-MS decoders for (4, 8)-regular LDPC code;


The FIG. 6 illustrates the FER performance of MS and SP-MS decoders for (5, 10)-regular LDPC code;


The FIG. 7 illustrates the FER performance of MS and SP-MS decoders for the IEEE 802.3 ETHERNET code;


The FIG. 8 illustrates the FER performance of MS and SP-MS decoders for the WIMAX rate 1/2 LDPC code;


The FIG. 9 illustrates the FER performance of MS and SP-MS decoders for the WIMAX rate 3/4 B LDPC code;


The FIG. 10 illustrates the FER convergence comparison on (5, 10)-regular LDPC code at Eb/N0=3.75 dB;


The FIG. 11 illustrates the FER convergence comparison on the IEEE 802.3 ETHERNET code at Eb/N0=4.75 dB;


The FIG. 12 illustrates the FER convergence comparison on the WIMAX rate 1/2 LDPC code at Eb/N0=1.75 dB; and


The FIG. 13 illustrates the FER performance of MS and SP-MS decoders for the IEEE 802.3 ETHERNET code with various alphabet lengths.


The FIG. 14 illustrates the performances of the SP-MS decoder in comparison with other decoders.





DESCRIPTION OF EMBODIMENTS

As a preliminary remark, the following described embodiments are focused on LDPC codes. However, the man skilled in the art may transpose without particular difficulties, the teaching to a code having at least two constraint nodes, particularly turbocodes. In the latter case, the constraint nodes use convolution as predetermined rule.


In reference to FIG. 1, an iterative decoder 1 is configured for decoding LDPC codes having a codeword length N. It means that the message was coded on N bits.


In the illustrated example, N=6. However, in practice, the codeword length N is much larger and can be typically around 64 000.


The decoder 1 comprises N variable nodes (VNs) vn, n=1 . . . N. It means that the variable node vn receives the result In of the signal quantization for the bit n of the received message. The signal quantization will not be described further as it is a step well known from the man skilled in the art.


The result In is a combination of the estimated value, either 0 or 1, and a likelihood that this estimated value is the correct one expressed as a log-likelihood ratio LLR. In is defined on a alphabet AL of qch quantization bits, qch being an integer equal or greater than 3.


In is coded on a particular alphabet AL, or representation, called Sign-Magnitude. The alphabet is symmetric around zero. AL={−Nqch, . . . , −1, −0, +0, +1, . . . , +Nqch} in which the sign indicates the estimated bit value and the magnitude represents its likelihood.


The following table gives the represention on 3 bits for a classical decoder using a 2-Complement representation and for the described decoder 1 using a Sign-Magnitude representation


















Estimated


Sign-



value
Likelihood
2-Complement
Magnitude









1
3
101
111



1
2
110
110



1
1
111
101



1
0
000
100



0
0

000



0
1
001
001



0
2
010
010



0
3
011
011










The variable nodes VNs use for all their computation an alphabet As constructed similarly than the alphabet AL but with q quantization bits, q≤qch.


The decoder 1 comprises also M constraint nodes CNs cm, m=1 . . . M, 2≤M<N.


The variable nodes VNs and the constraint nodes CNs are the nodes of a Tanner graph. The edges of the Tanner graph defines the relationship between variable nodes and constraint nodes along which they exchange messages.


The message sends by variable node vn to constraint node cm at iteration custom-character is named custom-character and the message sends by constraint node cm to variable node vn at iteration custom-character is named custom-character


The set of constraint nodes connected in the Tanner graph to the variable node vn is named V(vn). The degree dv of the variable node vn is defined as the size of the set V(vn), i.e. the number of constraint nodes connected to the variable node. And V(vn)\{cm} is defined as the set V(vn) except the constraint node cm.


Each variable node vn estimates the value yn of the nth bit of the codeword to be decoded and send messages custom-character,messages belonging to As, to the connected constraint nodes cm.


At the first iteration, each variable node vn sends to its connected constraint nodes the value In.


At each iteration custom-character, custom-character>1, each variable node vn computes sign-preserving factors custom-character which are the sum of the ponderated sign of LLR In and the sum of the signs of all messages, except for the message coming from cm, received from the connected constraint nodes. custom-character=ξ×sign (In)+Σc∈V(vn)\{cm} sign (custom-character) (1) where ξ is a positive or a null integer.


ξ is equal to 0 if dv=2, equal to 1 if dv>2 and dv is odd, and equal to 2 if dv>2 and dv is even.


Then the variable node vn compute the messages to send to the constraint nodes cm for iteration custom-character+1 in two steps. custom-character=In+½×custom-characterc∈V(vn)\{cm}(custom-character) (2) and custom-character=(sign(custom-character), custom-character(floor(abs(custom-character)))) (3) where S is a function from the set of value that can take floor (abs(custom-character)) to the set As.


Each constraint node cm tests the received message contents with predetermined constraints. The constraints may be parity check, convolution code or any block code. Typically for LDPC code, the constraint is parity check and the constraint nodes are thus named check nodes.


After analyzing the constraints, each constraint nodes cm sends messages custom-character to the connected variable nodes vn.


The messages custom-characterbelongs to an alphabet Ac. Alphabet Ac uses also a Sign-Magnitude representation but with a set of value similar or different than As. The alphabet As and Ac are thus symmetric around zero as AL.


Alphabet Ac may use a Sign-Magnitude representation with a set of value similar or different than As in terms or cardinality, i.e. Ac could be equal to {−3,−2,−1,−0,+0,+1,+2,+3} while As can be equal to {−1,−0,+0,+1} for example. By “similar”, one should understand equal or approximatively equal.


In a particular embodiment, the function S may be defined as S(x)=min(max(x−λ(x), 0),+Nq), where λ(x) is an integer offset value depending on the value of x. More particularly, λ(x) is a random variable that can take its value in a predefined set of values according to a predefined law. A careful choice of the offset value will allow a smooth convergence of the iterative process by decreasing the impact of high likelihood during propagation.


The message emissions are repeated until the decoding of the codeword is achieved successfully or a predetermined number of iteration is reached.


Consequently, the method for decoding is the following, FIG. 2:

    • Reception and storage of the LLR In by the variable node vn, n=1 . . . N, step 21;
    • Emission of messages custom-character step 23;
    • Verification of the constraints by the constraint nodes cm, step 25;
    • Emission of messages mcm→vn(1) containing a possible candidate value for the variable n solving the constraints, step 27;
    • Computation by the variable nodes of equations (1), (2) and (3), step 29;
    • Iteration of steps 23 to 29 until the codeword is decoded or the number of iteration reaches a predetermined count.


In the following sections, a comparison of classical quantized decoders with the disclosed decoder, which will be called Sign-Preserving Min-Sum (SP-MS) decoder, will be exposed as well as some theorical analysis and modelisation showing the improvements of the disclosed decoder. Finally, some experimental results will be disclosed.


Basic Notions of Classical Quantized Decoders and LDPC Codes

An LDPC code is a linear block code defined by a sparse parity-check matrix H=[hmn] of M rows by N columns, with M<N. The usual graphical representation of an LDPC code is made by a Tanner graph which is a bipartite graph G composed of two types of nodes, the variable nodes (VNs) vn, n=1 . . . N and the check nodes (CNs) cm, m=1 . . . M. A VN in the Tanner graph corresponds to a column of H and a CN corresponds to a row of H, with an edge connecting CN cm to VN vn exists if and only if hmn≠0.


Let us assume that v is any VN and c is any CN. Let us also denote V(v) the set of neighbors of a VN v, and denote V(c) the set of neighbors of a CN c. The degree of a node is the number of its neighbors in G. A code is said to have a regular column-weight dv=|V(v)| if all VNs v have the same degree dv. Similarly, if all CNs c have the same degree dc=|V(c)|, a code is said to have a regular row-weight dc. In case of irregular LDPC codes, the nodes can have different connexion degrees, defining an irregularity distribution, which is usually characterized by the two polynomials λ(x)=Σi=2dv,max λixi−1, and ρ(x)=Σj=2dc,max ρjxj−1. The parameters λi (respectively ρj) indicate the fraction of edges connected to degree i VNs (respectively degree j CNs). For regular codes, the polynomials reduce to monomials, λ(x)=xd−1 and ρ(x)=xdc−1.


Let x=(x1, . . . ,xN)∈{0,1}N denote a codeword which satisfies HxT=0. In the following examples, x is mapped by the Binary Phase-Shift Keying (BPSK) modulation and transmitted over the BI-AWGN channel with noise variance σ2. The channel output y=(y1, . . . , yN) is modeled by yn=(1−2xn)+zn for n=1, . . . , N, where zn is a sequence of independent and identically distributed (i.i.d.) Gaussian random variables with zero mean and variance σ2. The decoder produces the vector {circumflex over (x)}=({circumflex over (x)}1, . . . ,{circumflex over (x)}N)∈{0,1}N which is an estimation of x. To check if {circumflex over (x)} is a valid codeword, we verify that the syndrome vector is all-zero, i.e. H{circumflex over (x)}T=0.


For classical quantized decoders the finite message alphabet custom-characterc is defined as custom-characterc={−Nq, . . . , −1,0,+1, . . . ,+Nq} and consists of Ns=2Nq+1 states, with Nq=2(q−1)−1 where q is the number of quantization bits. Let us denote AL the decoder input alphabet, and denote custom-charactercustom-character the iteration number. Let us also denote custom-charactercustom-characterc the message sent from VN v to CN c in the custom-characterth iteration, and denote custom-charactercustom-characterc the message sent from CN c to VN v in the custom-characterth iteration.


The LLR that can be computed at the channel output is equal to:










LLR


(

y
n

)


=


log


(


Pr


(



y
n



x
n


=
0

)



Pr


(



y
n



x
n


=
1

)



)


=



2


y
n



σ
2


.






(
11
)







We assume that custom-characterc=AL, hence, LLR(yn) has to be quantized and saturated. For classical decoders, let us denote the quantizer by custom-character: custom-character→AL, defined as






custom-character(a)=custom-character(└α×a+0.5┘, Nq),   (12)


where └ ┘depicts the floor function and custom-character(b, Nq) is the saturation function clipping the value of b in the interval [−Nq, Nq], i.e. custom-character(b, Nq)=min (max (b, Nq),+Nq). The parameter α is called channel gain factor and is used to enlarge or decrease the standard deviation of quantized values at the decoder input. The value of α can be seen as an extra degree of freedom in the quantized decoder definition that can be analyzed and optimized for quantized decoders on the BI-AWGN channel. Note that if α is too large most of quantized values will be saturated to Nq. With those notations, we define the quantized version of the intrinsic LLR that initialize the classical decoder by the vector I=(I1, . . . ,IN)∈custom-characterCN, with





In=custom-character(LLR(yn)) ∀n=1, . . . ,N.   (13)


A MP decoder exchanges messages between VNs and CNs along edges using a Tanner graph. During each iteration, the VN update (VNU) and CN update (CNU) compute outgoing messages from all incoming messages.


Let us briefly recall the VNU and CNU equations for the Min-Sum based decoders, before introducing the Sign-Preserving Min-Sum (SP-MS) decoders. For this purpose, we define the discrete update functions for quantized Min-Sum based decoders. Let Ψv:AL×custom-characterc(dv−1)custom-characterc, denote the discrete function used for the update at a VN v of degree dv. Let Ψc:custom-characterc(dc−1)custom-characterc, denote the discrete function used for the update at a CN c of degree dc.


Thus the update rule at a CNU is given by









=



Ψ
c



(


{
}


v



V


(

c
m

)



\


{

v
n

}




)


=


(




v



V


(

c
m

)



\


{

v
n

}






sign
(
)


)

.






(
14
)







And the update rule at a VNU is expressed as






custom-characterv(In, {custom-character}c∈V(vn)\{cm})=Λ(custom-character),   (15)


where the function Λ(.) and the unsaturated v-to-c message custom-character are defined by





Λ(a)=sign (acustom-character(max(|a|−λv, 0),Nq).






custom-character=In+∈c∈V(vn)\{cm}custom-character


The alphabet of custom-character, denoted custom-characterU, is defined as custom-characterU={−Nq×dv, . . . ,−1,0,+1, . . . ,+Nq×dv}. We define the classical OMS decoder with offset value λv∈{+1, . . . ,+(Nq−2)}, where the special case of λv=0 corresponds to the classical MS decoder. It must be noted that the discrete functions Ψv and Ψc satisfy the symmetry conditions. Now, let us further custom-character=(custom-character, . . . ,custom-character) denote the a posteriori probability (APP) in the custom-characterth iteration. Let us also denote custom-characterapp the alphabet of APPs with custom-characterapp={−Nq×(dv+1), . . . ,−1,0,+1, . . . ,+Nq×(dv+1)}. The APP custom-charactercustom-characterapp is associated to a VN vn, n=1,2, . . . ,N. The APP update at a VN vn of classical MS-based decoders is given by






custom-characterv(In{custom-character}c∈V(vn))=Inc∈V(vn)custom-character  (16)


From the APP, {circumflex over (x)}n can be computed as {circumflex over (x)}n=(1−sign (custom-character))/2 if custom-character≠0, otherwise, {circumflex over (x)}n=0 if In>0 and {circumflex over (x)}n=1 if In≤0, for n=1, . . . ,N.


We must take into account that at the initialization of MP decoders, variable-to-check messages custom-character are initialized by In at custom-character=0, i.e. mvn␣cm(0)=In where vn ∈V(cm).


Sign-Preserving Min-Sum (SP-MS) Decoders

In the classical MS-based decoders, the value of the v-to-c message can be zero, see (5). In that case, the erased message, i.e. custom-character=0, does not carry any information and does not participate in the convergence of the decoder. In this paper, we propose a new type of decoder, with a modified VNU using a sign preserving factor, which never propagates erased messages.


Quantization Used for SP-MS Decoders

Using the sign-and-magnitude representation one can obtain a message alphabet which is symmetric around zero and which is composed of Ns=2q states. Hence the message alphabet for SP-MS decoders denoted by custom-characters is defined as custom-characters={−Nq, . . . ,−1,−0,+0,+1, . . . ,+Nq}. The sign of a message m ∈ custom-characters indicates the estimated bit value associated with the VN to or from which m is being passed while the magnitude |m| of m represents its reliability. In this paper, it is assumed that custom-characterL=As. An example of the binary representation of custom-characterc and custom-characters for q=3 is shown in Table 1, one can see that −0 is represented by 1002, +0 is represented by 0002, etc.









TABLE 1







Binary representation of the quantized values.











Classical Decoder
Sign Preserving Decoder












m ϵ AC
q = 3 bits
m ϵ AS
q = 3 bits
(sign (m), |m|)














−3
101
−3
111
(−1, 3)


−2
110
−2
110
(−1, 2)


−1
111
−1
101
(−1, 1)



100
−0
100
(−1, 0)


0
000
+0
000
(+1, 0)


+1
001
+1
001
(+1, 1)


+2
010
+2
010
(+1, 2)


+3
011
+3
011
(+1, 3)









The quantization process defined in (12) is replaced by






custom-character*(a)=(sign(a),custom-character(┌a×|a|┐−1,Nq)),   (17)


where ┌ ┐ depicts the ceiling function. The quantized LLR is thus defined as In=custom-character*(LLR(yn))∈custom-characters for n=1, . . . ,N. Let us define the update rules for Sign-Preserving decoders.


Sign-Preserving Min-Sum Decoders

One can note from (14) that the CNU by construction determines the sign of each outgoing message, thus the CNU generates outgoing messages that always belong to custom-characters, therefore, the CNU remains identical. In the case of the VNU, (15) should be modified to ensure that the outgoing message will always belong to custom-characters. To preserve always the sign of the messages, let us denote by custom-character the sign-preserving factor of the message custom-character, defined as






custom-character=ξ×sign(In)+Σc∈V(vn)\{cm}sign(custom-character),   (18)


where the values of ξ depends on the value of the column-weight dv of a VN vn, thus we have









ξ
=

{




0
,



if





d
v

=
2

,






1
,



if






d
v

>

2





and






(


d
v






mod





2

)



=
1

,






2
,



if





d
v

>

2





and






(


d
v






mod





2

)



=
0.









(
19
)







Note that the other values of ξ give worse decoding performance.


From (18), one can note that, by construction, custom-character is the sum of dv (resp. dv+1) values in {−1, +1} if dv is odd (resp. if dv is even and greater than 2), and custom-character∈{−1, +1} for the special case of dv =2. Thus, custom-character is always an odd number.


The update rule of the Sign-Preserving Offset Min-Sum (SP-OMS) VNU is changed from (5) to






custom-characterv(In,{custom-character}c∈V(vn)\{cm})=(sign(custom-character), custom-character(max(|custom-character|−λv, 0),Nq))   (20)


where custom-character is the unsaturated v-to-c message of Sign-Preserving decoders, given by







=


Λ
*



(


2

+

I
n

+




c



V


(

v
n

)



\


{

c
m

}







)



,




the function Λ*(.) is defined by Λ*(a)=(sign(a),└|a|┘). Note that Λ*(.) is applied on a non-null value since by construction, the fractional part of (custom-character)/2 is 0.5.


We define a Sign-Preserving Min-Sum (SP-MS) decoder by setting λv=0. For Sign-Preserving decoders, we have custom-characterU={−Nq×dv−└dv/2┘), . . . , −1,−0,+0,+1, . . . , +Nq×dv+└dv/2┘}.


The APP update at a VN vn of Sign-Preserving decoders is given by






custom-characterv(In,{custom-character}c∈V(vn))=In+½×ξ×sign(In)+531c∈V(vn)(custom-character+½×sign(custom-character)).   (21)


The alphabet of APPs for Sign-Preserving decoders is given by custom-characterapp={−(Nq×(dv+1)+(dv+ξ)/2), . . . , −1,0,+1, . . . ,+(Nq×(dv+1)+(dv+ξ)/2)}. From the APP, {circumflex over (x)}n can be computed as {circumflex over (x)}n =sign(In) if custom-character=0, otherwise, {circumflex over (x)}n=sign(custom-character) for n=1, . . . ,N.


Sign-Preserving Noise-Aided Min-Sum Decoders

In order to define the noisy version of the SP-MS decoder, named Sign-Preserving Noise-Aided Min-Sum (SP-NA-MS) decoder, we first introduce the constraints on the noise models, and then we present a noise model that we use to perturb noiseless decoders.


1. Probabilistic Error Model for SP-NA-MS Decoders

We assume that the noisy message alphabet is denoted by custom-characters. The noisy message custom-character is obtained after corrupting the noiseless message custom-character with noise. To simplify the notations is this section, we use mu to denote any custom-character and {tilde over (m)} to denote any custom-character


DE analysis of SP-NA-MS decoders can be performed only using memoryless noise models which must satisfy the following condition of symmetry






Pr({tilde over (m)}=ψ2|mu1)=Pr({tilde over (m)}=−ψ2|mu=−ψ1), ∀ψ1custom-characteru and ψ2custom-characters.


This noise model injects some randomness at the VNU of SP-NA-MS decoders. Thus the noisy-VNU is symmetric, allowing to use the all-zero codeword assumption necessary in DE. Since the addition of noise in VNUs is independent of the sign of the messages, we will suppose in the sequel without loss of generality that the messages mu and {tilde over (m)} are always positive.


Now, let us denote γ:custom-characterucustom-characters the function which transforms mucustom-characteru into {tilde over (m)}=γ(mu)∈custom-characters with the random process defined by the conditional probability density function (CPDF) Pr({tilde over (m)}|mu). In this document, the CPDF Pr({tilde over (m)}|mu) for the noise model γ is given by










Pr


(


m
~

|

m
u


)


=

{




1
,






if






m
~


=


m
u

=



+
0






or





if






m
~


=

+

N
q





,


m
u

>

+

N
q



,







φ


(

m
u

)


,






if






m
~


=


m
u

-
1


,




m
u



{


+
1

,

+
2

,





,

+

N
q



}



,







1
-

φ


(

m
u

)



,






if






m
~


=

m
u


,




m
u



{


+
1

,

+
2

,





,

+

N
q



}



,






0
,




otherwise
.









(
22
)







where φ(mu) is defined as










φ


(

m
u

)


=

{





φ
0

,



if





m
u

=

+

N
q



,







φ
a

,



if





m
u



{


+
2

,





,

+

(

Nq
-
1

)



}


,







φ
s

,



if





m
u

=

+
1


,









(
23
)







The noise model analyzed is parametrized by three different transition probabilities φ=(φsa0). The choice of these three transition probabilities is a compromise between complexity and the process of the border effects in custom-characters.


The reasoning behind γ is to implement a probabilistic offset with the purpose of always keeping the sign of the messages. With φ=(φaaa), we can implement the SP-MS decoder setting φa=0 and the SP-OMS with λv=1 setting φa=1. The effect of the noise on the extreme values of the message alphabet custom-characters is studied with φs and φ0. Thanks to γ, we can implement a SP-NA-MS decoder whose behaviour is a probabilistic weighted combination of a SP-MS decoder and a SP-OMS decoder. As an example, γ is depicted in FIG. 3 for (q=3, Nq=3).


2. Sign-Preserving Noise-Aided Min-Sum Decoders

A SP-NA-MS decoder is defined by injecting some randomness during the VNU processing. The SP-NA-MS decoder is implemented perturbing unsaturated v-to-c messages custom-character with noise. Hence, the update rule for a noisy-VNU is given by






custom-character=γ(custom-character),   (24)


We can note that γ is a symmetric function that performs the saturation function.


Density Evolution for Sign-Preserving Decoders

The goal of DE is to recursively compute the probability mass function (PMF) of the exchanged messages in the Tanner graph along the iterations. DE allows us to predict if an ensemble of LDPC codes, parametrized by its degree distribution, decoded with a given MP decoder, converges to zero error probability in the limit of infinite block length.


In order to derive the DE equations for Sing-Preserving decoders, custom-character, k ∈ custom-characters, denote the PMF of noiseless c-to-v messages in the custom-characterth iteration. Similarly, let custom-character(k), k ∈ custom-characters, denote the PMF of noiseless v-to-c messages in the custom-characterth iteration. Also, let Ω(0)(k), k ∈ AL, be the initial PMF of messages sent at custom-character=0. To deduce the noisy DE equations, let custom-character(k), k ∈ custom-characters, denote the PMF of noisy v-to-c messages in the custom-characterth iteration. We consider that the all-zero codeword is sent over the BI-AWGN channel.


Initialization

DE is initialized with the PMF of the BI-AWGN channel with noise variance σ2 as follows











p
vtoc

(
0
)




(
k
)


=

{




F


(
k
)




if



k
=

-

N
q









F


(
k
)


-

F


(

k
-
1

)





if




-

N
q


<
k


-
1








F


(
0
)


-

F


(

-
1

)





if



k
=

-
0








F


(
1
)


-

F


(
0
)





if



k
=

+
0








F


(

k
+
1

)


-

F


(
k
)





if




+
1


k
<

+

N
q








1
-

F


(
k
)





if



k
=

+

N
q











(
25
)







where






F


(
k
)



=


1



2

π




σ
n








-


k




e



-


(

t
-

μ
n


)

2


/
2



σ
n
2




fsimi














dt
,




(
26
)







with






σ
n


=



(

2
/
σ

)

×
α





and






μ
n


=


(

2
/

σ
2


)

×

α
.















DE update for CNU


The input of a CNU is the PMF of the noisy messages going out of a noisy VNU, i.e. custom-character. For a CN of degree dc, custom-character is given by






custom-character=Σ(1, . . . , idc−1):Ψc(i1, . . . ,idc−1)=kcustom-character(i1) . . . custom-character(idc−1), 0.1 cm∀k ∈ custom-characters.   (27)


Considering the diffe-rent connection degrees of CNs of irregular LDPC codes, we have






custom-characterdc=2dc,maxρdc×custom-character(k), 0.3 cm ∀k custom-characters   (28)


DE update for VNU


We know that γ perturbs unsaturated values. For this reason, we first compute the PMF of unsaturated v-to-c messages of a VN of degree dv, i.e. custom-character, with the following equation






custom-character(k)Σ(t,i1, . . . ,idv−1):Ψv(t,i1, . . . ,idv31 1)=(0)(t)custom-character(i1) . . . custom-character(idv−1), 0.1 cm ∀k ∈custom-characterU.   (29)


And second, the noise effect is added to the PMF of unsaturated v-to-c messages to obtain the corrupted PMF






custom-character(k)=custom-character(ipγ (i,k), 0.1 cm ∀k ∈ custom-characters,   (30)


where pγ is the transition probability of the VN noise, pγ also performs the saturation effect.


In this document, we use only the transition probabilities of the noise model γ defined here above. Although of course other noise models can be used.


Then the effect of the different connection degrees of VNs is considered using the following relation






custom-character(k)=∈dv=2dv,maxλdv×custom-character(k), 0.3 cm ∀k ∈ custom-characters   (31)


For SP-NA-MS decoders, the DE update for VNU is implemented with (29), (30), and (31) where the effect of noise injection is added at VNUs. We can deduce that the DE update for VNU of SP-MS decoders setting pγ(i,k)=1 if i=k, otherwise pγ(i,k)=0, i.e. φ32 (0,0,0).


Asymptotic Bit Error Probability

The asymptotic bit error probability can be deduced from the PMF of the APPs, which is obtained from the DE equations. Let custom-character denote the bit error probability at iteration custom-character, which is computed from the PMF of all incoming messages to a VN in the custom-characterth iteration, and defined by (f) I .0






custom-charactercustom-character(0)+Σi=−(Nq×(dv,max+1)+(dv,max+ξ)/2)−1 custom-character(i)   (32)


where custom-character(k), k ∈ custom-characterapp, denotes the PMF of the APP at the end of the custom-characterth iteration for Sign-Preserving decoders. We can compute custom-character (k) as follows






custom-characterλdv×custom-character(k), 0.3 cm ∀k ∈ custom-characterapp


where custom-character (k) is computed as







=






(

t
,

i
1

,





,

i

d
v



)

:


Ψ
v



(

t
,

i
1

,





,

i

d
v



)



=
k






Ω

(
0
)




(
t
)





(

i
1

)






(

i

d
v


)




,



k


app







The evolution of custom-character with the iterations characterizes whether the Sign Preserving decoder converges or diverges in the asymptotic limit of the codeword length. When the number of iterations custom-character goes to infinity, we obtain the asymptotic error probability pe(+∞). For SP-MS decoders which is a noiseless decoder, the decoder converges to zero error probability and successful decoding is declared, i.e. pe(+∞)=0.


In the case of SP-NA-MS decoders, contrary to the noiseless case, pe(+∞) is not necessarily equal to zero when the noisy DE converges and corrects the channel noise. It depends mainly on the chosen error model and the computing units to which it is applied. For noisy decoders, the lower bound of the asymptotic bit error probability, denoted pe(lb), has a mathematical expression that we can compute, but for other noise models is very difficult to find it. Hence, the bit error probability is lower-bounded as custom-character≥pe(lb)>0.


Density Evolution threshold

The DE threshold δ is expressed as a crossover probability (δ=ε*) for the BSC or as a standard deviation (δ=σ*) for the BI-AWGN channel, with the objective of separating two regions of channel noise parameters. The first region composed of values smaller than δ corresponds to the region where the DE converges to the zero error probability fixed point in less than Lmax iterations of the DE recursion. The second region composed of values greater than δ corresponds to when the DE does not converge. In this later case, the DE converges to a fixed point which does not represent the zero error probability. Then the DE threshold can be considered as a point of discontinuity between these two regions.


We compute the DE threshold performing a dichotomic search and stopping when the bisection search interval size is lower than some precision prec. The DE estimation procedure is performed choosing a target residual error probability η>pe(lb), and declaring convergence of the noisy DE recursion when pe(+∞) is less than or equal to η.


The noisy-DE threshold is a function of the code family, parametrized by its degree distribution (λ(x),ρ(x)), of the number of precision bits q, of the value of the channel gain factor α, and of the values of the transition probabilities of the noise model (φsa0). The algorithm 1 describes the procedure to compute the noisy-DE threshold a for the SP-NA-MS decoders for a fixed precision q, a fixed degree distribution (λ(x),ρ(x)), a fixed channel gain factor a, and a fixed noise model parameters (φsa0).


For the BI-AWGN channel, δ is the maximum value of σ or the minimum SNR at which the DE converges to a zero error probability can be expressed as







δ
db

=

10







log
10



(

1

2

R






σ

*
2




)







where σ*=δ, and R is the rate of the code.


In this paper, we use δ to jointly optimize the noise model parameters (φsa0) and the channel gain factor a for a fixed precision q and a fixed degree distribution (λ(x),ρ(x)) as follows











(


φ
s
*

,

φ
a
*

,

φ
0
*

,

α
*


)

=

arg







max

(


φ
s

,

φ
a

,

φ
0

,
α

)





{


δ
~



(


λ


(
x
)


,

ρ


(
x
)


,
q
,
α
,

(


φ
s

,

φ
a

,

φ
0


)


)


}

.












(
33
)







The optimization of the transition probabilities of the noise model γ, and the channel gain factor a is made using a greedy algorithm which computes a local maximum DE threshold. For noiseless decoders, the optimization (33) is reduced to the optimum channel gain factor a* which is computed performing a grid-search.


Asymptotic Analysis of Sign-Preserving Min-Sum Decoders
Asymptotic Analysis of SP-MS Decoders for Regular LDPC Codes

In this section, we consider the ensemble of (dv,dc)-regular LDPC codes with code rate R ∈ {1/2,3/4} for dv ∈ {3,4,5}, R=0.8413 for the IEEE 802.3 ETHERNET code, and quantized decoders with q ∈ {3,4}.


The DE thresholds of the noiseless classical MS and OMS decoders are given in Table 2. It can be seen that the OMS is almost always superior to the MS for the considered cases, except for the regular dv=3 LDPC codes with low precision q=3.









TABLE 2







DE thresholds of classical MS and OMS decoders


(dv, dc,)-regular LDPC code, BI-AWGN channel














(dv, = 3, dc, =6)
(dv = 4, dc = 8)
(dv = 5, dc = 10)
(dv = 6, dc = 32)
















q
λv
α*
δdb
α*
δdb
α*
δdb
α*
δdb



















3 bits
0
0.9375

1.7888

0.8125
2.7360
0.6300
3.4117
0.455
4.0812



1
1.0625
2.2039
1.25

2.3219

1.1500

2.7079

0.84

3.5928



4 bits
0
2.0
1.6437
1.625
2.5389
1.2500
3.1772
1.035
3.8154



1
1.875

1.3481

1.75

1.7509

1.5900

2.2306

1.28

3.1685



5 bits
0
4.0
1.6132
3.25
2.4948
2.3000
3.1126
1.985
3.7506



1
2.625

1.2154

2.0

1.7061

1.6900

2.2089

1.45

3.1400



















(dv = 3, dc = 12)
(dv = 4, dc = 16)
(dv = 5, dc = 20)
















q
λv
α*
δdb
α*
δdb
α*
δdb






3 bits
0
0.625

2.7316

0.6875
3.1550
0.5600
3.6449




1
0.9375
3.1343
0.9375

3.0632

0.9200

3.2312




4 bits
0
1.25
2.5646
1.375
2.9441
1.4000
3.3917




1
1.5

2.4484

1.5

2.5292

1.3900

2.7620




5 bits
0
2.5
2.5268
2.75
2.8991
2.4700
3.3373




1
2.25

2.3040

1.875

2.4606

1.6100

2.7238










In Table 3, we indicate the noisy and noiseless DE thresholds obtained with (33), we also show the DE gains obtained comparing the best thresholds indicated in bold in Table 2 and the noisy (resp. noiseless) thresholds of SP-NA-MS decoders (resp. SP-MS decoders). Moreover, we list the best noisy DE thresholds of NAN-MS decoders.









TABLE 3







Noisy DE thresholds of SP-NA-MS decoders and DE


thresholds of SP-MS decoders













SP-NA-MS decoders
NAN-MS
SP-MS decoders



















(dv, dc)
q
α*
φs*
φa*
φo*
{tilde over (δ)}db
DE gain
{tilde over (δ)}db - NIV
α*
s*, φa*, φo*)
{tilde over (δ)}db
DE gain






















(3, 6)
3 bits
0.96
0.987
0.712
0.000

1.4994


0.2894

1.5711
0.95
(1, 1, 0)

1.5096


0.2792




4 bits
1.79
1.000
1.000
0.000

1.2688


0.0793

1.2877
1.79
(1, 1, 0)

1.2688


0.0793



(3, 12)
3 bits
0.72
1.000
0.725
0.000

2.5421


0.1895

2.5989
0.71
(1, 1, 0)

2.5468


0.1848




4 bits
1.34
1.000
0.938
0.000

2.3596


0.0888

2.3777
1.36
(1, 1, 0)

2.3600


0.0884



(4, 8)
3 bits
1.01
0.900
1.000
1.000

1.9820


0.3399

2.1056
1.01
(1, 1, 1)

1.9824


0.3395




4 bits
1.54
1.000
1.000
1.000

1.7306


0.0203

1.7411
1.54
(1, 1, 1)

1.7306


0.0203



(4, 16)
3 bits
0.75
1.000
0.962
0.712

2.7448


0.3184

2.8121
0.78
(1, 1, 1)

2.7459


0.3173




4 bits
1.30
1.000
1.000
1.000

2.4941


0.0351

2.5077
1.30
(1, 1, 1)

2.4941


0.0351



(5, 10)
3 bits
1.12
1.000
1.000
0.987

2.4908


0.2171

2.6417
1.12
(1, 1, 1)

2.4908


0.2171




4 bits
1.57
1.000
1.000
1.000

2.2196


0.0110

2.2306
1.57
(1, 1, 1)

2.2196


0.0110



(5, 20)
3 bits
0.87
1.000
1.000
0.838

3.0106


0.2206

3.1400
0.89
(1, 1, 1)

3.0137


0.2175




4 bits
1.39
1.000
1.000
1.000

2.7412


0.0208

2.7596
1.39
(1, 1, 1)

2.7412


0.0208



(6, 32)
3 bits
0.74
1.000
1.000
1.000

3.3963


0.1965

3.5766
0.74
(1, 1, 1)

3.3963


0.1965




4 bits
1.18
1.000
1.000
1.000

3.1787


−0.0102

3.1685
1.18
(1, 1, 1)

3.1787


−0.0102










Several conclusions can be derived from this analysis. First, the DE thresholds of the SP-NA-MS decoders are almost always better than the DE thresholds of the noiseless classical decoders. The DE gains for the SP-NA-MS decoders are quite important for q=3, the largest gain obtained is around 0.3399 dB for (dv=4,dc=8). While the DE gains are smaller for the largest precision q=4. We can observe a loss of around 0.0102 dB for (dv=6,dc=32) and q=4. From this analysis, we can conclude that the preservation of the sign of messages and the noise injection are more and more beneficial as the decoders are implemented in low precision. Second, when comparing the noisy thresholds of SP-NA-MS and NAN-MS decoders, one can observe that the SP-NA-MS decoders achieve better DE thresholds for almost all (dv,dc)-regular LDPC codes tested, the only exception appears for the regular (dv=6,dc=32) LDPC code and q=4. The largest gain obtained, when comparing the SP-NA-MS thresholds and NAN-MS thresholds, is around 0.1803 dB for the regular (dv=6,dc=32) LDPC code and q=3. A third remark comes from the interpretation of the optimum φ* obtained through the DE analysis. We have φ*0=0 for regular dv=3 LDPC codes, this makes sense because dv=3 is small enough to transform {circumflex over (m)}=±1 into {tilde over (m)}=±0, which gives to custom-character a reliability of zero and which could not help to extrinsic messages become more and more reliable at each new decoding iteration. For regular dv>3 LDPC codes, we have almost always 6100 *0=1, hence, one can conclude that for regular dv>3 LDPC codes, the transformation from {circumflex over (m)}=±1 to {tilde over (m)}=±0, does not affect the decoding process. Note that in [?], the transformation from {circumflex over (m)}=±1 to {tilde over (m)}=0 should not be allowed since {tilde over (m)}=0 ∈ Ac erase the bit value. Finally, all SP-NA-MS decoders can be implemented as deterministic decoders since the values of the transition probabilities are close or equal to 0 or 1. The optimum noise parameters φ* are close to (φ*s,φ*a,φ*0)=(1,1,0) for the regular dv=3 LDPC codes. While in the case of the regular dv>3 LDPC codes, φ* are close to (φ*s,φ*a,φ*0)=(1,1,1) which correspond to a deterministic SP-OMS decoder.


Asymptotic Analysis of SP-MS Decoders for Irregular LDPC codes

In the previous section we have seen that the optimum noise parameters (φ*s,φ*a,φ*0) and the respective gains of SP-NA-MS decoders depend on the VN degree. For LDPC codes with irregular VN distribution, we propose therefore to extend our approach by considering a noise injection model γ with different values of the transition probabilities for the different connection degrees.


We denote by γ(2)(2)=(φs(2)a(2)0(2)) the model which injects noise at VNs of degree dv=2. Similarly, let γ(3)(3)=(φs(3)a(3)0(3)) denote the noise model for the VNs of degree dv=3. Finally, we decide to use the same model for all other VNs with degrees dv≥4, denoted γ(≥4)(≥4)=(φs(≥4)a(≥4)0(≥4)).


The optimization of the transition probabilities for an irregular LDPC code with distribution (λ(x),ρ(x)) is still performed by the maximization of the noisy DE thresholds:










(


φ


(
2
)

*


,

φ


(
3
)

*


,

φ


(


4

)

*


,

α
*


)

=

arg







max

(


φ

(
2
)


,

φ

(
3
)


,

φ

(


4

)


,
α

)





{


δ
~



(


λ


(
x
)


,

ρ


(
x
)


,
q
,
α
,

φ

(
2
)


,

φ

(
3
)


,

φ

(


4

)



)


}

.







(
34
)







For our analysis, we consider the ensemble of irregular LDPC codes which follow the distribution of the rate R ∈ {1/2,3/4}, length N=2304 code described in the WIMAX 726 x +2746 x2 +3706 xs standard. The degree distribution for the rate 1/2 code is λ(x)=22/76x+24/76x2+30/76x5 and ρ(x)=48/76x5+28/76x6, while for the rate 3/4 B code is λ(x)=10/88x+36/88x2+42/88x5 and ρ(x)=28/88x13+60/88x14. For these distributions, we indicate in Table 4 the DE thresholds of the noiseless MS decoder and the noiseless OMS decoder.









TABLE 4







DE thresholds of noiseless MS decoders and noiseless OMS decoders with


offset value λν = 1 for the WIMAX degree distribution


Irregular LDPC code, BI-AWGN channel












R = 1/2

R = 3/4















q
λν
α*
δdb
α*
δdb 1*


















3 bits
0
0.44
1.8310
0.66
2.8236




1
0.40
5.2283
0.50
3.4406



4 bits
0
1.07
1.3941
1.29
2.6150




1
0.80
2.8140
1.42
2.2416



5 bits
0
2.30
1.3013
2.50
2.5654




1
1.55
1.1828
1.97
2.1637










Noisy DE thresholds are summarized in Table 5, where we indicate the optimum values of a and of the noise parameters for the different degrees. Those results confirm the conclusions of the regular LDPC codes analysis: (i) the DE thresholds of SP-NA-MS decoders are better than the DE thresholds of NAN-MS decoders, (ii) the optimum value for φ*0 is 0 or it is close to 0 for dv=3 VNs and for q=3, and (iii) some of the optimized models are not probabilistic since the optimized values of the transition probabilities are very close to 0 or 1.









TABLE 5







DE thresholds of SP-NA-MS and SP-MS decoders


for the WIMAX degree distribution













SP-NA-MS decoders
NAN-MS
SP-MS decoders




















R
q
α*
dv
φs*
φa*
φo*
{tilde over (δ)}db
DE gain
{tilde over (δ)}db-NIV
α*
s*, φa*, φo*)
{tilde over (δ)}db
DE gain























½
3 bits
0.65
2
0.000
0.000
0.000

1.3997


0.4313

1.6236
0.65
(0,0,0)

1.4003


0.4307






3
0.000
0.162
0.000




(0,0,0)







≥ 4
1.000
1.000
1.000




(1,1,1)





4 bits
1.24
2
0.000
0.000
0.000

0.9547


0.4394

0.9995
1.24
(0,0,0)

0.9582


0.4359






3
0.250
1.000
0.375




(0,1,0)







≥ 4
1.000
1.000
1.000




(1,1,1)




¾
3 bits
0.81
2
0.000
0.000
0.000

2.4433


0.3803

2.5323
0.81
(0,0,0)

2.4451


0.3785






3
1.000
0.687
0.000




(1,1,0)







≥ 4
1.000
1.000
1.000




(1,1,1)





4 bits
1.48
2
1.000
0.788
0.000

2.2110


0.0306

2.2138
1.49
(1,1,0)

2.2111


0.0305






3
1.000
1.000
1.000




(1,1,1)







≥ 4
1.000
1.000
1.000




(1,1,1)











Another conclusion can be driven from these tables. From the DE analysis we can conclude that the noise should not be injected on degree dv=2 VNs for the case of low precision q=3, since we obtain always (φs(2)a(2)0(2))≃(0,0,0). While for the largest precision q=4, the noise should be injected on degree dv=2 VNs for some cases. These observations, combined with the fact that the optimum values for φ(≥4)* are always 1, lead to the conclusion that injecting noise in SP-NA-MS decoders for irregular LDPC codes is especially important for the degree dv=3 VNs (inject the noise on degree dv=2 VNs will be depend on the degree distribution and the precision used)


Finally, the gains for SP-NA-MS decoder for irregular codes are larger than for the regular codes. The gain of the rate 1/2 code is 0.4313 dB for the lower precision q=3, and 0.4394 dB for the largest precision q=4. In the case of the rate 3/4 B code, the gains for the two considered precision q=3 and q=4 are smaller than the rate 1/2 code, a gain of 0.3803 dB for q=3, and a gain of 0.0306 dB for q=4.


Finite Length Performance

In this section we present the frame error rate (FER) performance for noiseless classical MS, noiseless classical OMS, and SP-MS decoders. We analyze the quantized decoder performance over the BI-AWGN channel.


The considered decoders are the ones with the best DE thresholds, indicated in bold in Table 2 and Table 3. The noiseless classical OMS decoder performance for quantization q=5 are also show as benchmark.


A maximum of 100 iterations has been set for dv=3, dv=4, and dv=5 LDPC decoders. FIG. 4 shows the FER performance comparisons between the classical MS, classical OMS, and SP-MS decoders, for the two considered precisions q=3 and q=4, and for the regular (dv=3, dc=6) QC-LDPC code. FIG. 5 and FIG. 6 draw the same curves for the regular (dv=4, dc=8) QC-LDPC code and the regular (dv=5, dc=10) QC-LDPC code, respectively.


A first conclusion is that the finite length FER performance are in accordance with the gains predicted by the DE analysis. We observe in the waterfall (i.e. at FER=10−2) an SNR gain for the SP-MS decoders which corresponds to the threshold differences (Table 3 and Table 5): around 0.27 dB (q=3,dv=3,R=1/2), 0.06 dB for (q=4,dv=3,R=1/2), 0.32 dB for (q=3,dv=4,R=1/2), the same performance for (q=4,dv=4,R=1/2), 0.20 dB for (q=3,dv=5,R=1/2), the same performance for (q=4,dv=5,R=1/2). We have made the same analysis for LDPC codes with rate R=3/4, i.e. (dv=3,dc=12), (dv=4,dc=16), and (dv=5,dc=20), and obtained the same conclusions.


Simulation results for the IEEE 802.3 ETHERNET code are provided on FIG. 7 with a maximum of 30 iterations. Again, the SNR gains in the waterfall correspond to what was predicted with the DE analysis, with a 0.19 dB gain for q=3 and the same performance for q=4.


Similarly, FIG. 8 and FIG. 9 present simulation results for the WIMAX rate 1/2 LDPC code and the WIMAX rate 3/4 B LDPC code, respectively, for a maximum of 100 iterations. We observe again that the SNR gains in the waterfall region are in agreement with the gains predicted by the DE analysis, with a 0.40 dB gain for (q=3, R=1/2), a 0.40 dB gain for (q=4, R=1/2), a 0.37 dB gain for (q=3, R=3/4), and a 0.03 dB gain for (q=3, R=3/4).


Additionally, for the WIMAX rate 1/2 LDPC code, the 3-bit SP-MS decoder has the same FER performance as the 4-bit MS decoder. In the waterfall region, the 4-bit SP-MS decoder has the same FER performance as the 5-bit OMS decoder, while in the error floor region, the 5-bit OMS decoder has better FER performance than the 4-bit SP-MS decoder. In the case of the WIMAX rate 3/4 B LDPC code, The SP-MS decoders have better FER performance than the MS and the OMS decoders in the error floor region.


As a remark, we can also see that the preservation of the sign of messages does not seem to have an influence in the error floor of the decoders, since all the curves have similar slopes in the low FER region. This means that the preservation of the sign of messages does not correct the dominant error events due to trapping sets.


The study presented in details here above is a particular case where qch=q AL=AS=AC.


Another particular case of study occurs when qch=q+1, and AL≠AS=AC. For these conditions some results obtained in the study of the IEEE 802.3 ETHERNET code are presented.


In the initialization stage of the SP-MS decoder, i.e. first iteration, the variable-to-check messages are computed as mvn→cm(1)=min(max(In,−Nq), +Nq) where cm ∈ V(vn).


First, we present the DE thresholds δ obtained for different precision qch and q. The following table shows the DE thresholds for optimized SP-MS decoders.


















(dv, dc)
(qch, q)
α
δ





















(6, 32)
(3, 2)
0.74
3.3979




(3, 3)
0.74
3.3963




(4, 3)
1.22
3.1740




(4, 4)
1.18
3.1787










From the table we can conclude that the SP-MS decoder implemented with precision qch=3 and q=2 (resp. qch=4 and q=3) has the same performance as the SP-MS decoder implemented with precision qch=q=3 (resp. qch=q=4). In the implementation part this has a great impact, since it goes from the precision q=qch of the messages to the precision q=qch−1, this reduces the number of wires between VNs and CNs in an ASIC implementation.


Simulation results for the IEEE 802.3 ETHERNET code are provided on FIG. 13 with a maximum of 30 iterations. From the finite length FER performance, we can conclude that the performances of the SP-MS are in accordance with the performances predicted by the DE analysis.


In general, the choice of the alphabets AL, AS, and AC (i.e. the precision of LLRs, v-to-c messages, and c-to-v messages) will depend on the code with which the SP-MS decoder works. FIG. 14 illustrates the performances of the SP-MS decoder in comparison with other decoders. Considering the case of qch=q+1 and AL≠AS=AC for the regular (dv=4,dc=8) QC-LDPC code, we can note in the FIG. 14 that the SP-MS decoder exhibits poor performance having an early error floor. Note that for q=2, we have AS=AC={−1 ,−0, +0,+1} and AL={−3,−2,−1,−0,+0,+1,+2,+3}. In FIG. 14 one can also see that the SP-MS decoder considering qch=q and AL=AS=AC has the best performance. We have AL=AS=AC ={−3,−2,−1,−0,+0,+1 ,+2,+3} for q=3. Another particular case of study occurs when considering AL=AS≠AC, this means that the precision of the v-to-c messages mvn→cm(I) and the c-to-v m(cm→vn(I) messages are different. Let us consider the case of qch=3 bits of precision for the LLR and the v-to-c messages, hence AL=AS={−3,−2,−1,−0,+0,+1,+2,+3}. Let us also consider the case of q=2 bits of precision for the c-to-v messages, hence AC={−1,−0,+0,+1}. The CNs have 3-bit inputs and must generate 2-bit outputs. Using the update rule at CN, see equation (14), the c-to-v mcm→vn(I) messages belongs to the alphabet AS, i.e. qch=3 bits of precision. In order to use q=2 bits of precision for the c-to-v messages we consider the following mapping.









=

{





(


sign
(
)

,
0

)

,





if






abs


(
)











{

0
,
1

}






and





if






<
11








(


sign


(
)


,
1

)

,





if






abs


(
)











{

2
,
3

}






and





if






<
11








(


sign


(
)


,
0

)

,





if






abs


(
)











{
0
}






and





if







11








(


sign


(
)


,
1

)

,





if






abs


(
)











{

1
,
2
,
3

}






and





if







11


















Then the c-to-v messages belong to the alphabet AC. In the VNs, the c-to-v messages with amplitude abs(mcm→vn(I))=1 are interpreted as messages of amplitude 2. With this interpretation the VNs perform the update rules. Simulations results using a maximum of 100 iterations are provided on FIG. 14. From the results we can see that the SP-MS decoder with AL=AS≠AC has better performance than the MS and the OMS decoders. Comparing the SP-MS decoders, we clearly observe that the best SP-MS decoder is obtained using the same precision for messages and LLRs. We can also see that the early appearance of the error floor is eliminated by using 3 bits for the v-to-c messages and 2 bits for the c-to-v messages. In the implementation part this has a great impact because the number of wires in an ASIC implementation is reduced. It can be observed that for example AL≠AS=AC can be a good choice for the IEEE 802.3 ETHERNET code, AL=AS=AC can be a good choice for the WIMAX rate 1/2 LDPC code, and AL=AS≠AC can be a good choice for the regular (dv32 4,dc=8) QC-LDPC code.


A person of skill in the art would readily recognize that steps of various above-described methods can be performed by programmed computers. Herein, some embodiments are also intended to cover program storage devices, e.g., digital data storage media, which are machine or computer readable and encode machine-executable or computer-executable programs of instructions, wherein said instructions perform some or all of the steps of said above-described methods. The program storage devices may be, e.g., digital memories, magnetic storage media such as a magnetic disks and magnetic tapes, hard drives, or optically readable digital data storage media. The embodiments are also intended to cover computers programmed to perform said steps of the above-described methods.


The functions of the various elements shown in the figures, including any functional blocks labeled as “processors”, may be provided through the use of dedicated hardware as well as hardware capable of executing software in association with appropriate software. When provided by a processor, the functions may be provided by a single dedicated processor, by a single shared processor, or by a plurality of individual processors, some of which may be shared. Moreover, explicit use of the term “processor” or “controller” should not be construed to refer exclusively to hardware capable of executing software, and may implicitly include, without limitation, digital signal processor (DSP) hardware, network processor, application specific integrated circuit (ASIC), field programmable gate array (FPGA), read only memory (ROM) for storing software, random access memory (RAM), and non volatile storage. Other hardware, conventional and/or custom, may also be included. Similarly, any routers shown in the figures are conceptual only. Their function may be carried out through the operation of program logic, through dedicated logic, through the interaction of program control and dedicated logic, or even manually, the particular technique being selectable by the implementer as more specifically understood from the context.


As used in this application, the term “circuitry” may refer to one or more or all of the following:

    • hardware-only circuit implementations (such as implementations in only analog and/or digital circuitry) and
    • combinations of hardware circuits and software, such as (as applicable):
    • a combination of analog and/or digital hardware circuit(s) with software/firmware and
    • any portions of hardware processor(s) with software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions) and
    • hardware circuit(s) and or processor(s), such as a microprocessor(s) or a portion of a microprocessor(s), that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation.


This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and/or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device.


It should be appreciated by those skilled in the art that any block diagrams herein represent conceptual views of illustrative circuitry embodying the principles of the invention. Similarly, it will be appreciated that any flow charts, flow diagrams, state transition diagrams, pseudo code, and the like represent various processes which may be substantially represented in computer readable medium and so executed by a computer or processor, whether or not such computer or processor is explicitly shown.


The description and drawings merely illustrate the principles of the invention. It will thus be appreciated that those skilled in the art will be able to devise various arrangements that, although not explicitly described or shown herein, embody the principles of the invention and are included within its spirit and scope. Furthermore, all examples recited herein are principally intended expressly to be only for pedagogical purposes to aid the reader in understanding the principles of the invention and the concepts contributed by the inventor(s) to furthering the art, and are to be construed as being without limitation to such specifically recited examples and conditions. Moreover, all statements herein reciting principles, aspects, and embodiments of the invention, as well as specific examples thereof, are intended to encompass equivalents thereof.












Algorithm 1 - Computation of the noisy DE threshold















1) [Initialization]


 Initialize interval limits [{tilde over (δ)}1, {tilde over (δ)}2] with {tilde over (δ)}1 < {tilde over (δ)}2, such that DE succeeds


 for {tilde over (δ)} = {tilde over (δ)}1 and fails for {tilde over (δ)} = {tilde over (δ)}2. Further define {tilde over (δ)}m = ({tilde over (δ)}1 + {tilde over (δ)}2)/2


2) [While |{tilde over (δ)}2 − {tilde over (δ)}1| > prec]


 a) [Perform noisy DE]


  i) [Initialize noisy DE]


   Noisy DE is initialized with the equation (15) and σ = {tilde over (δ)}m


  ii) [Iteration Loop]


   A) [Compute PMF]


    Apply recursively the sequence of five equations (27), (28), (29),


    (30), and (31) for Lmax iterations.


   B) [Break Iteration]


    The iteration loop breaks when either the pe(custom-character) ≤ η or Lmax is


    reached.


 b) [Noisy DE succeeds]


  If pe(custom-character) ≤ η, the noisy DE has converged and we update {tilde over (δ)}1 ={tilde over (δ)}m, {tilde over (δ)}2 =


  {tilde over (δ)}2 and am = ({tilde over (δ)}1 + {tilde over (δ)}2)/2.


 c) [Noisy DE fails]


  If pe(Lmax) > η, the noisy DE has not converged and we update {tilde over (δ)}1 = {tilde over (δ)}1,


  {tilde over (δ)}2 = {tilde over (δ)}m and {tilde over (δ)}m = ({tilde over (δ)}1 + {tilde over (δ)}2)/2.


 d) [Tolerance]


  Compute the size of the interval |{tilde over (δ)}2 − {tilde over (δ)}1| and stops the procedure if


  it is smaller than the threshold tolerance ( e.g. 10−10).


3) [Threshold]


 {tilde over (δ)} = {tilde over (δ)}m is the noisy-DE threshold.








Claims
  • 1. Iterative decoder configured for decoding a code composed of at least two constraint nodes and having a codeword length N, said decoder comprising: N variable nodes (VNs) vn, n=1 . . . N, each variable node vn being configured to receive a log-likelihood ratio LLR In of the channel decoded bit n of the codeword to be decoded, said LLR In being defined on a alphabet AL of qch quantization bits, qch being an integer equal or greater than 2;M constraint nodes (CNs) cm, m=1 . . . M, 2≤M>N;the variable nodes and the constraint nodes being the nodes of a Tanner graph; variable nodes and constraint nodes being configured to exchange messages along edges of the Tanner graph;each variable node vn being configured for estimating the value yn of the nth bit of the codeword to be decoded and for sending messages mvn→cm, messages belonging to an alphabet As of q quantization bits, q ≤qch, to the connected constraint nodes cm, the set of connected constraint nodes being noted V(vn), defining the degree dv of the variable node vn, as the size of the set V(vn), and V(vn)\{cm} being the set V(vn) except the constraint node cm, and then,each constraint nodes cm being configured for testing the received message contents with predetermined constraints and sending messages mcm→vn, messages belonging to an alphabet Ac to the connected variable nodes vn;the message emissions being repeated until the decoding of the codeword is achieved successfully or a predetermined number of iteration is reached; andthe LLR In and the messages mvn→cm and mcm→vn are coded according to a sign-and-magnitude code in the alphabets AL, As and Ac symmetric around zero, AL={−Nqch, . . . ,−1,−0,+0,+1, . . . ,+Nqch}, As={−Nq, . . . ,−1,−0,+0,+1, . . . ,+Nq}, Ac={−Nq′, . . . ,−1,−0,+0,+1, . . . ,+Nq′}in which the sign indicates the estimated bit value and the magnitude represents its reliability; andeach variable node vn, for each iteration l, is configured to compute:sign-preserving factors, sum of the ponderated sign of LLR In and the sum of the signs of all messages, except for the message coming from cm, received from the connected constraint nodes at iteration I:=ξ×sign(In)+Σc∈V(vn)\{cm} sign () where ξ is a positive or a null integer;messages to connected constraint nodes cm for iteration I+1 computed in two steps=In+1/2×+Σc∈V(vn)\{cm}() and=(sign(),(floor(abs()))) where S is a function from the set of value that can take floor(abs())to the set As.
  • 2. The iterative decoder of claim 1 wherein ξ equal to 0 if dv=2, equal to 1 if dv>2 and dv is odd, and equal to 2 if dv>2 and dv is even.
  • 3. The iterative decoder of claim 1, wherein the constraints applied by the constraint nodes are parity check, convolution code or any block code.
  • 4. The iterative decoder of claim 3, wherein the codeword is coded by LDPC and the constraint nodes are check nodes.
  • 5. The iterative decoder of claim 1, wherein the function S is defined as S(x)=min(max(x−λ(x), 0),+Nq), where λ(x) is an integer offset value depending on the value of x.
  • 6. The iterative decoder of claim 5, wherein λ(x) is a random variable that can take it value in a predefined set of values according to a predefined law.
  • 7. The iterative decoder of claim 1, wherein alphabets As and Ac are identical.
  • 8. The iterative decoder of claim 1 wherein the means comprises at least one processor; andat least one memory including computer program code, the at least one memory and computer program code configured to, with the at least one processor, cause the performance of the decoder.
  • 9. An iterative decoding method for decoding a code composed of at least two constraint nodes and having a codeword length N, by an iterative decoder according to claim 1, said method comprising: reception and storage (21) of the LLR In by the variable node vn, n=1 . . . N;emission (23) of messages mvn→cm;verification (25) of the constraints by the constraint nodes cm;emission (27) of messages mcm−vn containing a possible candidate value for the variable n solving the constraints;computation (29) by the variable nodes of−ξ×sign (In)+Σc∈V(vn)\{cm} sign () where ξ is a positive or a null integer, and of messages to connected constraint nodes cm for iteration l+1 computed in two steps=In+1/2×+Σc∈V(vn)\{cm}() and=(sign(),(floor(abs()))) where S is a function from the set of value that can take floor (abs()) to the set As; anditeration from emission of messages mvn→cm to their computation for the next iteration until the codeword is decoded or the number of iteration reaches a predetermined count.
  • 10. A computer readable medium encoding a machine-executable program of instructions to perform a method according to claim 9.
Priority Claims (1)
Number Date Country Kind
18306599.4 Dec 2018 EP regional
PCT Information
Filing Document Filing Date Country Kind
PCT/EP2019/083544 12/3/2019 WO 00