This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMAG.2017.2700819, IEEE
Transactions on Magnetics

JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX 2017 1

A Detector-aware LDPC Code Optimization for Ultra-High Density

Magnetic Recording Channels
Lingjun Kong1,4 , Kui Cai2 , Pingping Chen3,2 , and Bing Fan4
1 College of Telecommunication & Inform. Engineering, Nanjing University of Posts & Telecommunications, Nanjing, China
2 Department of Science, Singapore University of Technology and Design, Singapore 487372
3 College of Physics and Information Engineering, Fuzhou University, Fuzhou, China
4 Center for Memory and Recording Research, University of California, San Diego, La Jolla, CA 92093, USA

In this paper, a detector-aware low-density parity-check (LDPC) code optimization algorithm is proposed for the ultra-high
density magnetic recording channel, such as bit-patterned magnetic recording (BPMR) and two-dimensional magnetic recording
(TDMR), where a 2D detector (2D-DET) is employed to combat 2D-ISI effects, by acquiring the variance of the 2D-ISI channel
log-likelihood ratios (LLRs) corresponding to the specific 2D-DET. The new parameter builds a bridge of LDPC code optimization
between the ISI-channel and AWGN channel. Numerical results show that our proposed algorithms are more efficient in running
time than another recently proposed optimization algorithms. Moreover, the protograph based quasi-cyclic (QC) codes using the
proposed optimization strategy, which have a low-complexity QC encoder structure with an effectual parallelizable protograph
decoder composition, enjoy up to 1 order of magnitude performance gain in bit error rate (BER) over the other random codes that
suffer from high error floors.

Index Terms—low-density parity-check (LDPC) codes, protograph LDPC codes, quasi-cyclic (QC) LDPC codes, two dimensional
(2D) intersymbol interference (ISI) channels

I. I NTRODUCTION indispensable for dense magnetic recording. Existing LDPC

codes optimized for symmetric, AWGN-like channels are
IMING at further increasing the data-storage density be-
A yond 1Tb/in2 [1], bit-patterned media recording (BPMR),
shingled writing (SW)/two-dimensional magnetic recording
not optimal for magnetic recording channels due to channel
asymmetry. In [9], the authors designed LDPC codes opti-
mized for partial response channels (PRCs) by means of a
(TDMR) have been emerging as new promising technologies.
density evolution (DE) analysis. Recent works [10] provide
However, with the reduction of the track pitch, inter-track
a comprehensive analysis of the error floor along with code
interference (ITI) becomes a major impairment for these mag-
optimization guidelines for structured and regular non-binary
netic recording systems. ITI combined with the conventional
(NB) LDPC codes in magnetic recording channels.
(down-track) inter-symbol interference (ISI) forms a two-
Not much work has been done on LDPC codes optimized
dimensional (2D) ISI that severely degrades the performance
for 2D-ISI channels. In [11], new LDPC codes optimized for
of data storage systems [2]-[3].
2D-ISI channels detected using the symbol-based BCJR 2D-
A number of approaches have been proposed to com- DET have been designed, and are shown to achieve better
bat 2D-ISI effects, including the use of 2D detectors (2D- performance than the LDPC codes optimized for AWGN
DETs) [4]-[8] that mitigate the interference in 2D directions channels. However, the LDPC codes in [11] were designed
and advanced error-correcting coding (ECC) schemes. The for use with the full 2D-DET, which is highly complex.
symbol-based Bahl-Cocke-Jelinek-Raviv (BCJR) [4] 2D-DET LDPC codes optimized for 2D-ISI channels detected using
provides optimal detection over such channels, but it has the IRCSDFA-GA detector in [12], exhibit excellent error-
prohibitive complexity for the typical 2D data size [5]. In [8], correction capabilities, especially for high codeword lengths.
the well known suboptimal 2D-DET, iterative row-column soft However, the ISI-optimized LDPC codes are designed for a
decision feedback algorithm (IRCSDFA) [6], has been further 2D-ISI channel detected by the specific 2D-DET, and code
simplified using Gaussian approximation. optimization procedure should be operated again for a new
By applying the turbo principle to 2D-DET and decoder, detector. Moreover, the 2D-DET Extrinsic Information Trans-
ECC also provides a significant means to combat 2D-ISI [9]- fer (EXIT) curve cannot be expressed in the closed form, it
[14]. Low-density parity-check (LDPC) code that is well- is required to statistically average over a large number of 2D
known for its capacity-approaching ability in the additive detection by Monte Carlo simulation, and the computational
white Gaussian noise (AWGN) channel is believed to be burden remains a bottleneck of the optimization approaches.
In this paper, we present a detector-aware LDPC code
Corresponding author: Lingjun Kong (email: [email protected]). This
work is supported by China NSF for Young Scientists (Grant No. 61501250, optimization algorithm for a 2D-ISI channel, acquiring the
61401099), Singapore A*Star Public Sector Research Funding (PSF) and variance of the channel log-likelihood ratios (LLRs) corre-
SUTD-MIT IDC research grant, the Natural Science Foundation for Young sponding to the specific 2D-DET. The new parameter which
Scientists of Jiangsu Province (Grant No. BK20140879), SRF for ROCS,
SEM., Natural Science Foundation of the Jiangsu Higher Education Institu- captures the error characteristics of magnetic recording builds
tions (Grant No. 14KJB510021), NUPTSF (Grant No. NY214015). a bridge of LDPC code optimization between the ISI-channel

0018-9464 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMAG.2017.2700819, IEEE
Transactions on Magnetics

JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX 2017 2

and AWGN channel. According to the proposed approach procedure [11], the effective mutual information of variable
that has lower computational complexity, the existing code node decoder (VND) combined with the 2D-DET IE,V N D is
optimization techniques previously developed for the AWGN DL DL
channel which are not effective can be re-used in 2D-ISI div
X X
IE,V N D = λi ×IE,V ND = λi ×
channels. i=1 i=1
The rest of this paper is organized as follows. Section II r ! (4)
2 div 2
presents the 2D-ISI channel model. Section III introduces the J (div − 1) × [J −1 (IA,V N D )] + [J −1 (IE,DET )]
detector-aware LDPC code optimization for 2D-ISI channels
and then the computational complexity is analyzed. The pro- di
where λi = (1−Rcv)×dc ×ςi , ςi is the fraction of nodes incident
posed schemes are validated in Section IV. Finally, conclu-
sions are drawn in Section V . to variable nodes of degree div , dc denotes the check node
div
degree, J(·) and J −1 (·) are given in [11], and IE,V N D de-
notes the combined 2D-DET and VND Extrinsic Information
II. 2D-ISI CHANNEL MODEL Transfer (EXIT) function of a degree−div variable node.
The 2D channel response before detection can be repre- Fig.2 shows the conditional probability density functions
sented by an LM × LN channel response matrix H2D : (PDFs) of the output LLRs of symbol-based BCJR 2D-DET
in an information bit density of 4Tb/in2 channel using Monte
  Carlo simulations, where a LDPC code with rate of 8/9 is
h(0,0) h(0,1) ··· h(0,LN −1)
 h(1,0) h(1,1) ··· h(1,LN −1)  employed. It is easy to observe that the re-modeled symmetric
H2D =  .. .. .. ..  Gaussian distribution with mean k and variance $ = 2k and
 . . . .

the LLRs of 2D-DET are well matched for transmitted coded
h(LM −1,0) h(LM −1,1) ··· h(LM −1,LN −1) bit 0s and 1s, respectively. Thus, we conclude that the LLRs
(1)
of a 2D-DET can be approximated by a symmetric Gaussian
where LM and LN denote the interference lengths in the
distribution giving a specific 2D-DET and the EXIT chart can
vertical and horizontal directions, respectively.
be applied to optimize LDPC codes for such 2D-ISI channels.
The binary data distributed in an array with M rows and
Then the (4) can be written as (5),
N columns is denoted by x(i, j) ∈ {−1, +1}, with i =
DL
1, 2, · · · , M , and j = 1, 2, · · · , N . Then, the received signal X q 
i −1 2
at position(i, j) can be expressed by IE,V N D = λi ×J (dv − 1) × [J (IA,V N D )] + $
i=1
M −1 LX
LX N −1 (5)
∗ ∗ where $ is the estimated variance of output LLRs of 2D-DET.
r(i, j) = h(m, n)x(i , j ) + w(i, j) (2)
m=0 n=0 The LDPC code optimization problem can be accomplished
by means of linear programming [11]. Following the pro-
where i∗ = i − m + dM and j ∗ = j − n + dN . Here, dM and
posed procedure, protograph-based QC-LDPC codes can be
dN are non-causality offsets, and w(i, j) is the AWGN with
2 constructed according to the optimized degree sequences for
zero mean and variance σw . In the coded 2D-ISI channel, the
2D-ISI channels, where the QC structure is maintained while
signal-to-noise ratio (SNR) per information bit is defined as
maximizing the girth of the Tanner graph in the progressive
M −1 LXN −1 edge-growth (PEG) algorithm [11]. The proposed strategy of
LX
!
2 2
Eb /N0 = 10×log10 |h(m, n)| /2σw /Rc (3) detector-aware codes design is summarized in Algorithm 1.
m=0 n=0

where Rc is the code rate. B. Computational complexity analysis

Normally, the offsets are set to be dM = LM2−1 and dN = The symbol-based BCJR 2D-DET trellis has 2M (LN −1)
LN −1
2 . It is also assumed that a boundary of −1 around the
states and 2M input and output branches for each state,
data block for initializing the 2D-DET, i.e., x(i, j) = −1 for therefore its complexity is O(N · 2M LN ). The complexity of
(i, j) ∈
/ {(i, j) |1 ≤ i ≤ M, 1 ≤ j ≤ N }. the IRCSDFA-GA, whose state space dimension of the ISI
trellis of either component detector is reduced enormously by
employing the GA, is O(ICD · M N · ((LM − 1)LN 2LN +
III. A D ETECTOR - AWARE LDPC C ODE O PTIMIZATION
LM (LN − 1)2LM ), where ICD denotes the iteration number
In this section, we present a detector-aware LDPC code between two component detectors in the IRCSDFA-GA de-
optimization approach for 2D-ISI channels and compare the tector [8]. In [11]-[12], the 2D-DET EXIT curve cannot be
computational complexity with the methods in [11]-[12]. expressed in the closed form, so it is essential to statistically
average over a huge number of 2D detection simulations.
Specifically, once the a-prior mutual information IA,DET and
A. Detector-aware LDPC Code Optimization 2D-ISI channel output sequence r are generated, the 2D-
In the 2D-ISI channel, the soft information messages gener- DET is run. The 2D-DET output LLRs are used to calculate
ated by the 2D-DET are exchanged within the LDPC decoder the IE,DET (see [11]), which is measured by estimating the
iteratively to perform joint detection and decoding, where p(ζ|x =1 ) and p(ζ|x =−1 ) from a histogram of the extrinsic
the block diagram is given in Fig.1. In the code design information values, where ζ denotes the element of LLRs.

0018-9464 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMAG.2017.2700819, IEEE
Transactions on Magnetics

JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX 2017 3

Algorithm 1 Detector-aware LDPC code design algorithm TABLE II

1: Initialize: code rate Rc , code length Nc , initial chan- T HE OPTIMIZED DEGREE SEQUENCES OF LDPC CODES

nel Eb /N0 , the step ∆s , target cost Tcost . λ =

λ2 λ3 λ7 λ8 λ9 λ10 λ15
(λ2 , λ3 , · · · , λDL ) and ρ = (ρ2 , ρ3 , · · · , ρDR ).
The proposed 0.0644 0.5953 0.2267 0.1136
2: Run the 2D-DET once and calculate the a-posteriori
Scheme 1 0.1445 0.8146 0.0378 0.0031
mutual information of the 2D-DET.
Scheme 2 0.0622 0.4016 0.5362
3: Model the distribution of LLRs as an Gaussian with
variance $
4: Based on (5), select the best λ sequences to fit the check
IV. S IMULATION R ESULTS
node decoder (CND) EXIT curve as closely as possible
to the EXIT curve of the combined 2D-DET and VND. In this section, we present the simulation results with the
5: Design LDPC parity check matrix according the optimized LDPC schemes proposed in Section III. The symbol-based
degree sequences. Go to next step if protograph-based QC- BCJR-aware code optimization is proposed and verified for a
LDPC code is to be presented, or END. 4Tb/in2 2D-ISI channel, however, it can easily be extended
6: Adjust the degree sequence based on the code length for other 2D detector-aware schemes. In the simulations, we
Nc , code rate Rc and replication fact J, and obtain the set the maximum number of iterations between the symbol-
protograph base matrix Hp (Lpm × Lpn ). based BCJR 2D-DET and the LDPC belief-propagation (BP)
7: Permute the edges of the nodes in the J replicas of the decoder to 10, and the maximum number of BP iterations
base protograph matrix Hp : in LDPC decoder to 30, all codes have rate of Rc = 8/9.
8: for i = 1 to Lpn do The 2D channel response matrix (CRM) is assumed to be (6),
9: for j = 1 to J do corresponding to a magnetic recording density of 4Tb/in2 [11].
10: if (j == 1) then  
0.0368 0.1435 0.0368
11: Select the check nodes with larger local girth based
H4T =  0.2299 0.8966 0.2299  (6)
on the modified PEG algorithm
0.0368 0.1435 0.0368
12: else
13: //Cyclically shift others to get the new position Llj . In Fig.3, we compare the BER performance of our proposed
14: for l = 1 to Nid do LDPC codes with the 2D-ISI-optimized (Scheme 1) [11] and
15: //Nid is the number of connections of the check AWGN-optimized (Scheme 2) for two different code lengths
nodes connected with nodes v i of Hp . in a 2D-ISI channel with recording density of 4Tb/in2 , where
Ll

the optimized degree sequences of the LDPC codes are given
16: Llj = j−1J × J + Llj−1 + 1 mod J
17: end for in Table II. Fig.3 shows that the proposed LDPC codes (red
18: end if lines) has a performance gain of about 0.2 dB over the codes
19: end for of Scheme 2 (blue dashed lines) at BER of 10−5 for all
20: end for two different code lengths (N c = 4608, 9216) in a 2D-ISI
channel with recording density of 4Tb/in2 . Furthermore, there
is only small loss in performance of the proposed codes when
compared to the codes which are designed by the significantly
TABLE I
C OMPUTATIONAL C OMPLEXITY OF D IFFERENT O PTIMIZATION S CHEMES
more complex optimization approach (Scheme 1) at BER of
10−5 for all two different code lengths, and the proposed codes
Optimization schemes Computational complexity have a lower error floor at high SNR region.
2D BCJR-based Opt. [13] O(Lλ · L~ · N · 2M LN ) To further evaluate the validity of the proposed detector-
Proposed BCJR-aware Opt. O(N · 2M LN )
aware LDPC code optimization algorithm, we design the
protograph based QC-LDPC code according to Algorithm 1
O(Lλ · L~ · I · M N · ((LM − 1)LN 2LN
2D GA-based Opt. [14] as described in Section III. Fig.4 shows the BER comparison
+LM (LN − 1)2LM ))
between the protograph-based QC-LDPC code and random-
O(I · M N · ((LM − 1)LN 2LN like protograph LDPC codes [15]. As shown, the proposed
Proposed GA-aware Opt.
+LM (LN − 1)2LM )) protograph based QC-LDPC code is able to provide up to
1 order of magnitude performance gain relative to the other
random codes that suffer from high error floor, and no error
floor is observed down to a BER of about 10−8 . Note
Instead of running the 2D-DET Lλ · L~ times, we ap- that the protograph based QC structure only causes a slight
ply the 2D-DET once in the proposed optimization scheme, performance degradation compared with others.
where Lλ is the interval of IA,DET and L~ is the 2D-
DET running time for the histogram to estimate the p(ζ|x =1 )
V. C ONCLUSION
and p(ζ|x =−1 ). The computational complexities of the 2D
BCJR-based optimization, 2D GA-based optimization, and A low-complexity detector-aware LDPC code optimization
corresponding detector-aware ones are listed in Table I which algorithm has been proposed for dense 2D-ISI magnetic
are for off-line optimization of the designed codes. recording channels. The new variance of the channel LLRs

0018-9464 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TMAG.2017.2700819, IEEE
Transactions on Magnetics

JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. XX, XX 2017 4

−2
10

−3
10

−4
10
Proposed Protograph QC−LDPC code
Random LDPC code [13]
Random Protograph code: 2D−P1 [15]

BER
−5
10 Random Protograph code: 2D−P2 [15]

−6
10

Fig. 1. Block diagram of joint 2D-DET and decoding for a 2D-ISI channel. −7
10

−8
10
0.12 4.2 4.4 4.6 4.8 5 5.2 5.4
Eb/N0(dB)

0.1
Fig. 4. BER performance for the protograph-based QC-LDPC code and
random-like LDPC codes in a 4Tb/in2 channel.
0.08

original extrinsic LLR for bit 0

f(LE,DET−DEC)

estimated N(k0,2k0) for bit 0

0.06
original extrinsic LLR for bit 1
R EFERENCES
estimated N(k ,2k ) for bit 1
1 1
[1] R. Wood, M. Williams, A. Kavcic, and J. Miles, “The feasibility of
0.04 magnetic recording at 10 Tb/in2 on conventional media,” IEEE Trans.
Magn., vol. 45, no. 2, pp. 917-923, Feb. 2009.
0.02
[2] Y. Wang, B. V. K. Vijaya Kumar, “Multi-track joint detection for
shingled magnetic recording on bit patterned media with 2-D sectors,”
IEEE Trans. Magn., vol. 52, no. 7, Article#: 3001507, Jul. 2016.
0
−30 −20 −10 0 10 20 30
[3] J. Yao, E. Hwang, B. V. K. Vijaya Kumar, et al., “Two-track joint
LE,DET−DEC
detection for two-dimensional magnetic recording (TDMR),”in Proc.
IEEE ICC, London, UK, Jun. 8-10, 2015, pp.418-424.
[4] L. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of
Fig. 2. PDFs of the 2D-ISI channel L-values detected by symbol-based BCJR linear codes for minimizing symbol error rate,” IEEE Trans. Inform.
2D-DET in a 4Tb/in2 magnetic recording channel with Eb /N0 = 4dB. Theory, vol. 20, no. 2, pp. 284-287, Mar. 1974.
[5] B. M. Kurkoski, “Towards efficient detection of two-dimensional
intersymbol interference channels,” IEICE Trans. Fundamentals, vol.
corresponding the specific 2D-DET captures the error char- E91-A, no. 10, Oct. 2008.
[6] T. Cheng, B. J. Belzer, and K. Sivakumar, “Row-column soft-decision
acteristics of 2D-ISI magnetic recording channels. Based on feedback algorithm for two-dimensional intersymbol interference,”
the detector-aware code optimization approach, new LDPC IEEE Signal Processing Lett., vol. 14, no. 7, pp. 433-436, Jul. 2007.
codes have been designed for 4Tb/in2 channels. Simulation [7] J. Yao, K.C. Teh, and K. H. Li, “Joint iterative detection/decoding
scheme for discrete two-dimensional interference channels,” IEEE
results have verified that our proposed protograph based QC- Trans. Commun., vol. 60, no. 12, pp. 675-681,Dec. 2012.
LDPC codes can perform as well as the random LDPC codes, [8] J. Zheng, X. Ma, Y. L. Guan, K. Cai, and K. S. Chan, “Low-
obtain up to 1 order of magnitude performance gain over other complexity iterative row-column soft decision feedback algorithm
for 2D inter-symbol interference channel detection with Gaussian
random codes, and no error floor is observed down to a BER approximation,”IEEE Trans. Magn., vol. 49, no. 8, pp. 4768-4773,
of about 10−8 , while retaining the implementational benefits. Jun. 2013.
[9] A. Kavcic, X. Ma, and M. Mitzenmacher, “Binary intersymbol
interference channels: Gallager codes, density evolution and code
performance bounds,” IEEE Trans. Inform. Theory, vol. 49, no. 7,
10
−2
pp. 1636–1652, Jul. 2003.
[10] A. Hareedy, B. Amiri, R. Galbraith, and L. Dolecek, “Non-Binary
LDPC Codes for Magnetic Recording Channels: Error Floor Analysis
and Optimized Code Design,” IEEE Trans. Commun., vol. 64, no. 8,
10
−3
pp. 3194-3207, Aug. 2016.
[11] L. Kong, Y. L. Guan, J. Zheng, G. Han, K. Cai, and K. S. Chan,
“EXIT-chart-based LDPC code design for 2D ISI channels,” IEEE
10
−4 Trans. Magn., vol. 49, no. 6, pp. 2823-2826, Jun. 2013.
[12] Y. L. Guan, G. J. Han, L. J. Kong, K. S. Chan, K. Cai and J.
BER

Zheng, “Coding and signal processing for ultra-high density magnetic

Proposed:N =4608 recording channels,” in Proc. Int. Conf. on Comp., Networking and
10
−5 c
Scheme 1:Nc=4608 Commun. (ICNC), Hawaii, USA, Feb. 3-6, 2014, pp. 194-199.
Scheme 2:N =4608
c
[13] E. Hwang, R. Negi, and B. Kumar, “Additive encoding low-density
Proposed:Nc=9216 parity-check (AE-LDPC) codes for two-dimensional magnetic record-
10
−6 Scheme 1:N =9216
c ing (TDMR),” in Proc. Int. Conf. on Comp., Networking and Commun.
Scheme 2:Nc=9216
(ICNC), Hawaii, USA, Jan. 30-Feb. 2, 2012, pp. 481-485.
4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 [14] M. Mehrnoush, B. J. Belzer, K. Sivakumar and R. Wood, “EXIT
E /N (dB)
b 0
chart-based IRA code design for TDMR turbo-equalization system,”
IEEE Trans. Commun., vol. 65, no. 4, Apr. 2017.
[15] P. P. Chen, L. J. Kong, Y. Fang, and L. Wang, “The design of
Fig. 3. BER performance comparison with different schemes in a 4Tb/in2 protograph LDPC codes for 2-D magnetic recording channels,” IEEE
channel employed the same symbol-based BCJR 2D-DET. Trans. Magn., vol. 51, no. 11, Article#: 3101704, Nov. 2015.

0018-9464 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.