On the Performance of Rate-Compatible LDPC Codes under All-Integer Quantization Decoding

. From a practical point of view, quantization is necessary for the design of any low-complexity LDPC decoder. However, the quantization scheme with the minimum computational complexity called all-integer quantization has not been well investigated since, intuitively, it seems to signifcantly degrade the decoding performance. It is found in this work that the utilization of 4-bit all-integer quantization for MSA-based LDPC decoder is adequate to achieve the decoding performance of the nonquantized case which can be considered as the lower bound on the decoding performance. Moreover, this simple quantization scheme can also be applied to the hardware and cost-efcient rate-compatible LDPC codes.


Introduction
Nowadays, it is well known that low-density parity-check (LDPC) codes can provide great error correction performance over many types of communication channels [1][2][3]. Terefore, LDPC codes are chosen for various communication standards, such as DVB-S2 [4], optical fber [5], and 3GPP [6]. Teir decoding algorithm, known as the sumproduct algorithm (SPA), is the important factor that yields the LDPC codes with high decoding performance [7]. However, one might consider that the check node processing of SPA is a bottleneck of LDPC decoding [8]. In order to solve this problem, a simplifed version of SPA, called the min-sum algorithm (MSA), has been proposed to greatly reduce the computational complexity [9].
Unlike theoretical studies that assume infnite precision decoding, the implementation of any LDPC decoder has to rely on fnite precision, i.e., quantization [10]. It is known that the number of quantization bits relates to both the decoding performance and the computational complexity of decoding [10]. Specifcally, LDPC decoding with a sufcient number of quantization bits can provide good error-correcting performance [11]. In contrast, a smaller number of quantization bits can incur a loss in decoding performance but a lower computational complexity can be expected [11]. Terefore, LDPC decoder based on MSA must be practically designed to trade-of between performance and computational complexity [12].
Note that, for SPA-based algorithms, three major processes involving LDPC decoding are initialization, check node processing, and variable node processing [13]. By using the same quantization bits for all decoding processes, the fnite precision efects on LDPC decoding are investigated in [14]. Te simulation results in this work show that the SPA is very sensitive to the efect of quantization. Furthermore, it is also demonstrated that, compared with SPA, the MSA is less sensitive to the quantization efect. Te decoding performance by using the diferent quantization bits at variable and check node processing is provided in [15]. With the assumption of fnite precision, the performance of the enhanced versions of MSA, including normalized MSA and ofset MSA, is studied in [16]. For all works mentioned before, distances for all quantized level are the same, i.e., uniform quantization. It is reported in [16] that more sophisticated quantization for MSA can outperform the uniform quantization. An improved MSA specifcally designed for 5G LDPC codes is proposed [17,18], and their simulation results are carried out over fnite precision efect.
Contrary to the conventional quantization scheme that relies on both integer and fractional parts, our work focuses attention on the all-integer quantization decoding, i.e., quantization without decimals, to further minimize the computational complexity of the LDPC decoder. We frst demonstrate that, for fxed rate LDPC codes, the decoding performance of MSA with only 4-bit integer quantization is almost identical to that of MSA with infnite precision. After getting this result, our hypothesis is that rate-compatible LDPC codes could have higher sensitivity to the efect of integer quantization than the fxed rate LDPC codes. To the best of our knowledge, the performance of rate-compatible LDPC codes under integer quantization has not yet been reported. So, the performance of this particular LDPC coding scheme is deeply investigated in this work. Surprisingly, our results reveal that rate-compatible LDPC codes, constructed from both puncturing and shortening, can exhibit excellent decoding performance under 4-bit integer quantization.
Te remainder of this paper is organized as follows. Section 2 presents min-sum decoding for rate-compatible LDPC codes. Te quantization decoding is described in Section 3. We show simulation results and discussions in Section 4. Finally, Section 5 presents the conclusions.

Basic and Background
In order to comprehend our work, the overall design diagram for the LDPC code is described in this section. Te system shown in Figure 1 can be explained as follows. Initially, the information vector m is generated and then sent to the rate-compatible low-density parity-check (RC LDPC) code. After that, the codeword c is produced. Te details of RC LDPC codes will be explained below. To straightforwardly deal with the time-varying channel, many pairs of encoders and decoders must be employed to support various code rates. However, the hardware cost is very high. Rate-compatible (RC) codes are the solution to this problem since they can adapt the code rate according to the channel conditions by using only a single pair of encoder and decoder, i.e., mother code. From a fxed rate mother code, shortening techniques can be used to provide a series of lower rate codes whereas puncturing can be applied to generate a series of higher rate codes.
Te idea of shortening is to insert known symbols instead of actual message symbols. Figure 2 shows the distinction between the code rates after encoding and removing. Firstly, the known symbols with length k s bits are inserted into the actual information with length k m bits. Te information vector m with a length k � k m + k s bits is sent to the encoding process. Note that the positions and values of the known symbols are known to both the transmitter and receiver. Ten, the encoder forms the codeword c of length n bits. Te code rate obtained after the encoding is R � k/n. After that, the known symbols are removed from the codeword and the code rate change to R s � k m /n s , i.e., shortened code rate. So, the code rate after shortening depends on the number of known symbols, and the length of the actual codeword is shortened to n s � n − k s bits.
By using the (4, 2) shortened code with the (6, 4) mother code as an example, the process of shortening is illustrated in Figure 3. Te information vector with a length k � 4 bits is the input of the encoder. Te known symbols of length k s � 2 bits are inserted instead of the actual message bits. After the encoding process, all the known symbols are removed. Ten, the actual codeword with the length n s � 4 bits is produced.
Te shortening process at the receiver side is described via Figure 4. Te known symbols are inserted back into the predefned positions before decoding. Tose known symbols, which can be considered perfect knowledge for a decoder, are fxed through the decoding process. Ten, by removing known symbols, the decoder fnally produces the estimation.
It is essential to note that the position of the inserted known symbols infuences the decoding performance [19,20]. Terefore, the shortening position should be carefully chosen. Te method used to select the position of the known symbols is called the shortening algorithm. Tere are many shortening algorithms proposed in the literature. Te best known shortening algorithm is the uniform shortening algorithm [21]. So, such an algorithm as per employed to per our work.
Note again that puncturing is a technique to obtain a series of higher rate codes from a single mother code. Te certain parity symbols of the codeword vector will be punctured before transmission. Figure 5 shows a block diagram of the puncturing LDPC code for the encoding process. Te information vector, length k bits is encoded to acquire a codeword c. After that, some parity symbols will be removed from the codeword. So, the code length is shorter and the code rate is increased.
To readily understand, the example of puncturing from code rate 1/2 into 3/4 is illustrated in Figure 6. Te information vector of length k � 3 bits is encoded. Ten, the encoder produces a codeword of length n � 6 in which the  length of parity symbols is 3 bits. After that, two parity symbols are punctured, so the length of the codeword is 4 bits. Finally, the punctured code rate is 3/4.
Like shortening, the selection of puncturing symbols infuences the decoding performance. Te best known puncturing algorithm is called grouping and sorting and is proposed in [22]. However, the simple puncturing of consecutive parity bits is used for simplicity. For punctured position, the decoder must be initiated with an ambiguous state, e.g., a zero loglikelihood ratio. Ten, the punctured position can be recovered by their neighbors that have nonzero values.

Min-Sum Algorithm.
Despite the excellent performance of the traditional LDPC decoder based on the sum-product algorithm (SPA), it requires high decoding complexity. Te min-sum algorithm (MSA) is an alternative to SPA with substantially lower complexity. Te MSA is the approximate algorithm of SPA that provides small decoding performance degradation. Te procedures of min-sum decoding are described as follows: At the receiver side, the log-likelihood ratios (LLR) of the received vector are expressed by where log means log e . c i denotes the i-th codeword, and r i is the i-th received. For the initialization step, log-likelihood ratios value for the additive white Gaussian noise (AWGN) channel is calculated by where σ 2 is the noise variance, and i � 1, 2, 3, . . . , n. Te variable-to-check message M is given by where j � 1, 2, 3, . . . , k. Te check-to-variable message E can be calculated by where l � 1, 2, 3, . . . , l max and l max denotes the maximum number of decoding iterations. Let N(j) be the set of all variable nodes connected with j-th check node. Te updated LLR value llr * is represented by Te value llr * i is made the hard decision to get the decoding result as follows: where z ⊤ � [z 1 , z 2 , z 3 , . . . , z n ]. Estimation at each iteration is denoted by z. Te decoding terminate when the estimation satisfes Hz ⊤ � 0 or the maximum number of iterations is reached. Otherwise, the variable-to-check message is updated by where M(i) is the set of all check nodes connected with i-th variable node. Ten, the iteration began with equation (4).

Quantization Decoding
It is common to use infnite precision operations in the theoretical study of LDPC decoding. However, it is difcult to execute in hardware implementation. Hence, there is a great need to use a fnite precision scheme, i.e., quantization, which constitutes a low-complexity and efcient decoder. For practical LDPC decoding, the reduction of data and complexity in the LDPC decoder leads to a decrease in area, power consumption, and latency. Particularity, the quantization without a decimal part results in a signifcantly greater reduction of computational complexity.
Te notation (q: f) are used to represent uniform quantization scheme. Let q denote the total number of quantization bits and f be the bit length of fractional part of the value. To quantize the infnite precision into fnite precision, quantization limit is defned as Te interval between every two adjacent quantized values is 1/2 f . Te set of quantized values is expressed as follows: where Q J � J/2 f and J � 2 q− 1 − 1.
As mentioned earlier, only all-integer quantization is utilized in this work. Terefore, the notation of uniform quantization given above can be represented by (q: 0). For example, the (4: 0) quantization scheme will provide the set of quantized values as follows: For the sake of completeness, the formula to perform quantization with constraint (q: f) is given as follows: where x is the input with infnite precision and y ∈ Q denotes the quantized output value. It is worth noting again that, f is always zero since all-integer quantization is only our focus.

Simulation Results
Te bit error rate (BER) performances of LDPC codes under fnite precision decoding are presented in this section. A BPSK-AWGN channel is employed for all the simulations. Te sum-product and min-sum algorithms are used as the LDPC decoder. Following [15], the maximum iteration of the LDPC decoder is fxed at 20. Various block lengths and code rates involved in this section are chosen to the 5G New Radio (NR) standard [23].

Integer Quantization Efects on SPA Decoding.
Teoretically, infnite precision, i.e., nonquantization is assumed for all LDPC decoding algorithms. However, the practical implementation of any LDPC decoder relies on quantization. As mentioned before, integer quantization implies the lower decoding complexity and the relationship between the number of quantization bits and decoding performance should be addressed. So, this motivates us to frst investigate the traditional and powerful SPA-based LDPC decoder. Figures 7 and 8 illustrate the BER performance of integer decoding for diferent numbers of quantization bits. It is found that integer quantization causes severe performance degradation in all cases. Since the SPA decoding requires high resolution to calculate, even 6 bits for integer quantization are not sufcient to provide good performance. Terefore, with quite poor performance, it is not exaggerated to state that SPA decoding is very sensitive to integer quantization and integer arithmetic as well.

Integer Quantization Efects on MSA Decoding.
Before showing the decoding performance of MSA under integer quantization, MSA decoding performance with a diferent maximum number of iterations is depicted in Figures 9 and  10. It is obviously seen that the BER performance gradually reduces when the maximum number of iterations is increased. However, it is observed that the BER performance tends to saturate at 100 decoding iterations. Importantly, the use of the maximum number of iterations of 20 causes a small performance loss, i.e., 0.3-0.4 dB compared with employing 100 iterations. Terefore, MSA with 20 decoding iterations, which is quite a good compromise between time consuming and BER performance, is selected to exhibit the efect of integer quantization in this work.
To show the robustness of MSA decoding to the integer quantization efect, Figures 11 and 12 present the performance of MSA under this circumstance. It is obviously seen that the decoding with 4 bits integer quantization can provide the performance very close to the infnite precision decoding. Although the performance of MSA decoding is slightly poorer than that of SPA, the MSA is more powerful than SPA for low precision decoding, i.e., integer quantization with 2 or 4 bits. Te MSA decoding is an approximation algorithm. In the check-to-variable messages procedure, for quantization decoding, the error probability for the MSA is lower than that of SPA. Hence, it is reasonable to think that the MSA decoder is less sensitive to fnite precision.
To justify the robustness of MSA under integer quantization decoding, a simple analysis is carried out by observing the quantized values and nonquantized values for both SPA and MSA. Te check-to-variable message or check node processing is focused on mean squared error (MSE) calculating because SPA and MSA are diferent only at this step. Te check node processing for SPA is given by

Journal of Electrical and Computer Engineering
While this step for MSA is defned by (4). Specifcally, it is known that check node processing for MSA is an approximation of that of SPA in order to achieve lower decoding complexity. Terefore, it is worth to note that the values involved in MSA decoding are already based on approximated values or truncated values. So, it is intuitive to think that the MSA under integer decoding, i.e., further approximation of decoding value, is less sensitive to SPA with integer decoding.
To clearly show this, let y v and x v be the v-th quantized value and v-th nonquantized value at check node processing, respectively, where v � 1, 2, . . . , N and N is the total number of observation. Te MSE between quantized value and nonquantized value can be expressed by Tis parameter is defned to quantify the quantization error between quantized values and nonquantized values for both the MSA and SPA decoding algorithms. Te MSE analysis is shown in Table 1. Te E b /N 0 used for block length  of 1008 bits and 264 bits is 3.5 and 5 dB, respectively, to achieve a BER of 10 − 6 . Additionally, the observed values of about 10 6 samples are used. It is clear that, in all cases, the MSE values of SPA are much greater than those incurred from MSA. So, the SPA decoding is very sensitive to the integer quantization efect, but this quantization efect seems to be negligible for MSA decoding, as shown in Figures 11 and 12    puncturing techniques From the previous section, our results reveal that the utilization of 4-bit integer quantization is sufcient for MSA decoding to achieve good decoding performance. Terefore, the rate-compatible LDPC codes with 3 and 4 bit integers are explored in this subsection. Figures 13 and 14 show the BER performance of shortening LDPC code with integer quantization. It can be seen that the decoding with 4-bit integer quantization for shortened LDPC codes still provides performance very close to that of infnite precision decoding. Regrettably, the shortened LDPC code with 3-bit integer quantization decoding leads to slight performance degradation. Next, the performance of puncturing LDPC codes under integer quantization decoding is presented. Te integer quantization efect on the puncturing LDPC is shown in Figures 15 and 16. It is observed that the simulation results are similar to those of shortening. Although the MSA decoder with puncturing is infuenced by the reduction of the parity bits, it can provide good decoding performance. For 4bit integer quantization, the BER performance of puncturing approaches the ideal case. Unlike the shortening case, the decoding performance is poorer when the quantization scheme is based on 3-bit integer quantization. So, the efect of integer quantization on the puncturing scheme is more severe when the quantization level is reduced.

Conclusions
Te LDPC decoding with low-resolution leads to a reduction of computational complexity, power consumption, and increase of speed. Te BER performance of rate-compatible LDPC codes under integer quantization decoding are presented in this paper. It is found that the rate-compatible LDPC codes with 4-bit integer quantized MSA decoding algorithms are able to attain a 0.1 dB BER performance loss compared to MSA decoding with infnite precision. As a result, under integer quantization, an efcient implementation of a high-throughput LDPC decoder can be realized. Te operation of the MSA decoder with integer quantization can be implemented via look-up tables (LUT), facilitating simple hardware design. Our work can be applied to the research in 5G system. Motivated in part of the performance gap caused by using 3-bit integer quantization. We think that the performance improvement would be an interesting topic for future research.

Conflicts of Interest
Te authors declare that they have no conficts of interest.   Journal of Electrical and Computer Engineering 7