Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
0Activity
0 of .
Results for:
No results containing your search query
P. 1
10.1.1.65

10.1.1.65

Ratings: (0)|Views: 7|Likes:
Published by Robert Bird

More info:

Published by: Robert Bird on Jun 17, 2012
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

06/17/2012

pdf

text

original

 
Marker codes for channels with insertions and deletions
Edward A. RatzerCavendish Laboratory, University of Cambridge, Madingley Road, Cambridge CB3 0HE, UKPhone: +44 (0) 1223 337238, Fax: +44 (0) 1223 354599E-mail: ear23@mrao.cam.ac.uk
Abstract
: At low noise levels marker codes areshown to outperform watermark codes. Full iterativedecoding enhances performance to close to the capac-ity bounds.
Keywords
: Insertion-deletion channel, markercodes, LDPC codes, turbo synchronization, water-mark codes
1. INTRODUCTION
Examples of channels with synchronization errorsinclude:
Serial line
The clock speed of the transmitter maynot be accurately known (for instance due totemperature variations in the clock) so the timeof arrival of each transmitted bit is not known.
Hard disc
Variations in the rotation speed (for in-stance due to mechanical vibrations or shock)mean the position of the head relative to theplatter may be uncertain.
DAT tape
Tape stretch leads to problems similarto those suffered by a hard disc.In this paper the channel will be modelled by ran-dom uncorrelated insertion or deletion events at un-known positions. A flowchart of the channel model isshown in figure 1. The capacity for channels of thiskind is not known exactly. A capacity lower boundfrom [11] and an upper bound from [9] have beenused in this paper.Marker codes [5] were originally designed to beable to deal with single insertion or deletion errorevents. The bit stream to be transmitted has a reg-ular marker (or header) inserted in it. For examplethe marker ‘001’ may be inserted between every 4data bits:01101100101010
01100011100001101000110The decoder can then look out for the markers anduse any shift in their position to deduce bit loss orgain. Errors in the matched sequence can then becorrected with a conventional code. With advancesin computer power probabilistic sequence matching
 c c          © d  d  d  d  d 
1
 p
flip
p
flip
   t   t
 
E'
p
del
p
ins
 p
t
' T c c
 
 TE
Read
t
Insert random bitEmit
t
Emit 1
t
Figure 1: Flow chart describing the insertion-deletion channel with insertion probability
p
ins
, dele-tion probability
p
del
, transmission probability
p
t
=1
 p
ins
 p
del
, and substitution probability
p
flip
. Forsimplicity
p
flip
= 0 for the simulations in this paper.[7] can be carried out. The coding system is shownin figure 2.Watermark codes [3] are a similar scheme, butrather than having bursts of synchronization infor-mation and bursts of data, the information is dis-tributed uniformly. To encode, the data bits areuniformly sparsified and then combined with a wa-termark sequence. To decode, probabilistic resyn-chronization can be carried out with the watermarksequence.In this paper the results in [7] on the properties of the markers alone are extended by comparing com-plete marker code based systems with watermarkcodes. The benefit of iterative probabilistic resyn-chronization is also studied.
2. COMPARISON WITHWATERMARK CODES
In [7] it was shown that for a code system likefigure 2 the capacity of the effective channel seen byan outer code is higher for marker codes than forwatermark codes at low noise levels. The highestrate results for watermark code published [3] are at
 
Effective channelas seen by theouter LDPC codeProbabilisticresynchronizationInsertion/deletionchannelEncode withLDPC codeLDPC decoderSource dataDataAdd markers
Figure 2: The code system used, with serial decoding illustrated
0.00010.0010.010.110.001 0.01 0.1
  p  r   (   b   l  o  c   k  e  r  r  o  r   )
p
ins
= p
del
Capacity bounds
   S  e  r   i  a   l   "  c  a  p  a  c   i   t  y   "
WatermarkCode ACode B
Figure 3:
R
= 0
.
71,
= 4995 codes over an in-sertion/deletion channel with serial decoding. Thewatermark code result is from [3]. Marker code A iswith a LDPC code with with weight 2 and weight 3columns. Marker code B has weight 10 columns inaddition, table 1.
R
= 0
.
71 with a block size of 4995. To match this amarker code of the same overall rate and block sizewas created, code A. The inner code was chosen tobe good at a noise level
p
ins
=
 p
del
=0
.
005 using the ef-fective capacity technique outlined in [7]. The outercode was a low-density parity-check (LDPC) codewith weight 2 and 3 columns. The code parametersare shown in table 1. Decoding was as figure 2.A comparison of the watermark and marker codesis shown in figure 3. The marker code outperformsthe watermark code, despite the watermark codehaving been constructed with a LDPC code definedover a larger field (GF(16)) than the binary LDPCcodes considered in this paper.The marker code is not close to the channel capac-ity bounds. To try to approach the bounds, furtheroptimization of the outer LDPC code was studied.The threshold of an infinite loop-free LDPC code as afunction of column weight distribution was evaluatedusing a Monte Carlo approach [2]. In this procedureit is necessary to know the distribution of messagessent by the channel. Statistics of the messages re-ceived by the outer code were collected, figure 4. Thedistribution is almost symmetrical if given a data se-
LLR
   D  a   t  a   b   i   t  n  u  m   b  e  r
10 5 0 5 10171319
   l  o  g   (  p   d   f   )
−11−10−9−8−7−6−5−4−3−2
(a) At a noise level of 
p
ins
=
p
del
= 0
.
5%.The vertical axis ranges over each databit between two markers
0.0020.0040.0060.0080.010.012pr(ins/del)-10-8-6-4-202LLR | transmission of 000.050.10.150.20.25pdf
(b) Marginalised over all data bits whosetransmitted bit was ‘0’
Figure 4: Probability density functions of messagesreceived by an outer code where the inner code has amarker sequence of ‘01’ between every 19 data bits.The log-likelihood ratio (LLR) is
Pr(bit=1)Pr(bit=0)
.quence of i.i.d. bits. The simulation was done withan all zero-transmission but with noise statistics as if they were i.i.d. data bits. Despite the non-Gaussianform of the messages, degree sequences could notbe found that significantly outperformed good de-gree sequences obtained from optimizations on theGaussian channel [1]. This optimization was carriedout with the message histograms marginalised overall bits, figure 4(b). Perhaps if a different degreesequence for each bit between markers is allowed afurther gain could be achieved, as figure 4(a) showsbits received near markers are more reliable than bits
 
far away from markers.Simulations using a good degree sequence fromthe Gaussian channel, code B, are also shown in fig-ure 3. The waterfall region is not much closer to theinsertion-deletion capacity, but it is close to the se-rial “capacity” (defined from the effective capacity of the channel seen by the outer code [7]).
3. A COMPLETE ITERATIVESYSTEM
It has been shown [6], [10] that extending loopybelief propagation as used to decode LDPC codes toinclude estimating channel state can be beneficial.For the insertion-deletion resynchronization phase itis expected that if we give the resynchronizationalgo-rithm more information about the likely transmissionthe accuracy of the resynchronization will increase.Watermark and regular marker codes were em-pirically discovered to have similar performance near
R
= 0
.
5 [7]. Therefore
R
= 0
.
5 codes with a block-size of 4000 (to match [3]) were studied.For benchmarks, simulations were done of markercodes with the serial decoding algorithm from sec-tion 2. The first simulation (code C) was with iden-tical markers of length 3 and a LDPC outer codewith weight 2 and weight 3 columns. Figure 5 showsthat the watermark code has better error floor be-haviour. With identical markers catastrophic errorpropagation is possible as the resynchronization canbe shifted by a multiple of a marker interval. CodeD was similar to code C, however the markers werepseudo-randomly chosen from a set of two differentmarkers. This improved the error floor and led tobetter performance than the watermark code. Anirregular LDPC outer code (chosen to be good onthe Gaussian channel) was also tested and a smallfurther improvement found, code E.Simulations were carried out with codes D and Eusing iterative resynchronization. The algorithm issimilar to the serial resynchronization algorithm butwith extrinsic information from the LDPC decoderfed back into the resynchronization stage (alteringPr(
r
|
path) in [7, eq 2]). The resynchronization wascarried out every five iterations of the LDPC code toincrease the decoding speed as a LDPC code itera-tion is faster than probabilistic resynchronization.The simulation results are also shown on figure 5.The figure shows that the iterative approach signifi-cantly outperforms the serial approach and that thewaterfall regions can be close to the channel capac-ity. It is worth noting that the ranking of the codesis reversed. The choice of a code to be used shouldbe made when the decoding algorithm is known.To look into this swap in performance, extrinsicinformation transfer (EXIT) charts [8] for the sys-tems were studied at a noise level of 
p
ins
=
 p
del
=0
.
04.
0.00010.0010.010.110.01 0.1
  p  r   (   b   l  o  c   k  e  r  r  o  r   )
p
ins
= p
del
Capacity bounds
   S  e  r   i  a   l   "  c  a  p  a  c   i   t  y   "
CDEEDWatermarkSerialIterative
Figure 5: The benefit of full iterative decoding with
R
= 0
.
5,
= 4000 codes. Identical marker codesare shown simulated with iterative and non-iterativeresynchronization. Also the reduction in error floorfrom identical markers (C) to non-identical markers(D) is shown. The outer LDPC code in codes C andD have only weight 2 and 3 columns, code E hasweight 10 columns in addition. The watermark coderesult is from [3] and the details of the marker codesare in table 1.The EXIT chart of the resynchronization was ob-tained by a Monte Carlo approach and a quadraticfunction fitted, figure 6(a). With no informationpassed from the code to the resynchronization stagethe maximum rate of the LDPC outer code with se-rial decoding is shown. As the input informationis increased the curve increases but does not reach1. This is due to remaining uncertainty in the exactsynchronization path and whether the bit in questionmay have been deleted (and possibly reinserted).The behaviour of the decoding algorithm can beseen in terms of messages being passed between thevariable nodes and check nodes. The variable nodetransfer function includes the EXIT chart of theresynchronization calculated above. The code thatperforms well on the Gaussian channel leads to anEXIT chart with an intersection, figure 6(b), ie de-coding is not expected to converge. With the codewith only weight 2 and weight 3 columns no intersec-tion is observed, figure 6(c), and therefore decodingis expected to converge.The width of the swath between the curves can beused as a metric to choose a better degree sequence.Better degree sequences were searched for using aglobal optimization package [4]. Better thresholdscould only be found by increasing the number of weight 2 columns above the number of rows in theparity-check matrix. This is not expected to producea good code as short cycles in weight 2 columns leadto low weight codewords.

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->