# //I/ -3

NASA Technical Memorandum 105431

Least Reliable Bits Coding (LRBC) for High Data Rate Satellite Communications

Mark Vanderaar and Paul Wagner Sverdrup Technology, Inc. Lewis Research Center Group Brook Park, Ohio
and

James Budinger National Aeronautics and Space Administration Lewis Research Center Cleveland, Ohio

Prepared for the 14th International Communications Satellite Systems Conference sponsored by the American Institute of Aeronautics and Astronautics Washington, D.C., March 22-26, 1992

NASA
(\'A~A-TM-1054?I) LEAST ttcLIASLF BITS CODING (Lir-C) "C? HIGH DATA 3ATr! SATELLITE COMMUNICATIONS (NASA) 10 p C5CL 173 G3/32 N92-18921 Unclas 0071432

Pb can be written in a matrix form that illustrates the individual bit contributions as follows W|rt>o W|* i W|rt>2 W|rt>3 P2. + 32P2 + 32P3 + 16P4) 24 _1_(24Pi + 16P2 + 40P3 + 16P4) 24 Symbol Number Phase Value Natural Mapping Gray Coding Decision Parsing \ LV NO \8i\l\ LV NO \8/. 001 000 001 010 011 110 111 100 101 011 010 110 111 101 100 and P-) is simply 371/4 Jl Pi = Po. middle.PS where Ps is the symbol error probability as calculated in Appendix B. PQ is the probability that no symbol error occurred and is written as Po = 1 .J. 4 -P3 Alternatively. mrb (middle reliable bit). respectively.. 0 1 2 3 4 0 71/4 TC/2 000 001 010 011 100 101 110 111 000 . Three symbol mappings defined in Table 1 are evaluated for their probability of bit error as a function of their decision regions. and leftmost bits in the bit to symbol mappings./§l~sin(32L)l) pbGC = J_(16P.1 is P 0 . However. OPo 4P. and Mrb (most reliable bit) represent the rightmost.Q [. Matrices for the three mappings studied are OP0 OPo 8P. 8P2 OP0 OP0 8P. 1 = ( l .was transmitted.Q[. To determine P-). The following three expressions result PbNM = -J24 24P2 + 36P3 + 8P4) 1 24 12P3 . = f oL/I^sin (A \ LV No Is P4 WMrbl Wmrb 1 Wmrt>2 Wm*3 Mrb mrb 24 L Irb = P3. 8P4 8P4 OP 4 . b *T 12P3 8P3 16P3 . ED/NO curves of these 8PSK mappings are shown in Figure 2. Due to the relative minimization of the dominating decision region probability P-). To accomplish this. it is noticed that the probability of symbol error can also be written as the sum of two half-planes minus their overlapping area Ps= Q NO where the W^d's are the number of ways in which a bit error b can be caused by a decision region d multiplied by the probability of being in decision region d.sin =22L . 8P4 OP4 OP 4 ./H~sin(*\l)(i. these equations do not indicate which bits are contributing to the weights on the decision regions. .3= QU/^. 16Pi 1 PbNM = L_ 16P2 OP2 *) A. the first quadrant probability P0. 1 . gray coding exhibits the lowest Eb/No to obtain a given BER. The Irb (least reliable bit). 8P2 and combining with the equation for Ps ( from Appendix B) it is seen that Mrb mrb Irb - From the symmetry of the signalling constellation the following decision region probabilities immediately follow . Ps = Pa Ps = Pz P7 = Pi PbQC " Mrb mrb Irb .Po 5 6 7 5TC/4 3II/2 7n/4 Similar equations are used to determine the remaining decision region probabilities and are given as follows Table 1 : 8PSK bit to symbol mappings The BER vs. 4Pi 8P2 16P2 8P3 12P3 OP0 4P.

resulting in a reduced probability of bit error. For this high data rate application. This structure of sixteen decision regions is shown in Figure 8. 111.Prr\s Pm7 = P4/2 Pm8 = Pm7 Pain = Pm4 Pm14 = Pmi Pm9 = Pm6 Pmi2 = Pm3 Pm15 = Pm0 Prtiio = Prris Pmi3 = Pm2 i=U1 • ' VI and the resulting average decoded bit error probability is y^Vcd These expressions are used to determine the performance of the three 8PSK modulation schemes that are RS coded in a bit by bit fashion. Figure 7 shows the best scheme from each mapping. 6 show the performance of the various coding schemes for each bit mapping.PbDP OPo 4Pi 8P2 24 12P3 . Reed-Solomon (RS) block codes are studied over convolutions! codes because of the commercial availability of RS codec VLSI chips at data rates in the hundreds of megabits per second. called Pb1 mrb ci a"d Pb1 Mrb ci is calculated as M-1 2M-1 I (bix®bjL2x)(b| M Casel 171 Case 2 Case3 199 241 241 207 237 237 Case 4 CaseS 213 233 233 227 227 227 i =0 o \ L j even 2M-1 2 M _i |rt|pmn+ 2 / NONE NONE )jx ® bj±l x ||bj|rb® b_i|ri) Pmn iodd i =o Table 2 : k Values for RS(n.Pi where Pb is the channel bit rate and m= Iog2(n+1) is the number of bits in an RS character. Bit Irb mrb Mrb Three types of error events are considered to determine the effect of soft-decision information when only the Irb is coded.Pmi t = (n . Knowing this. with special notice of the PI terms that occupy the second row verifies that some bits are more likely to be in error than others. half and quarter-plane probabilities are again used. denoted Pm. the best mapping uses gray coding. further parsing of the decision regions is needed. 18 l\l I Pm2 = Pi . Again. This compares to Viterbi algorithm convolutional decoders available at tens of megabits per second.2Pi . The average channel character error rate Pc .Pa . The less reliable bits result in the bulk of the errors. The use of soft-decision information [5] can be used to further improve the performance of LRBC. block codes are well suited to time-shared decoding and independent burst decoding .Po .P o l 2\\ LV N.Po . the uncoded bits can be modified as necessary. To find the probabilities associated with each minor decision region. Irb. First. The following equations result Reading the columns. 5. For example. the received decision region is necessarilly in error. Figure 3 shows the individual bit error rates of gray coded 8PSK where the Irb contributes more to the overall BER than the mrb and Mrb. In each case the ensemble code rate is 8/9. is determined from the channel bit rate by Pm M- . the probability that an uncoded mrb or Mrb is in error given a correction has been made on the coded Irb. Also. An approximate expression for the decoded character error probability of a t error correcting RS code is [4] Pm6 = Pa . Figures 4. The likely symbol transmitted is the nearest one (in terms of Euclidean distance) with the corrected coded bit(s). When a coded bit correction is made. n=255 and k is shown in Table 2.2 : LRBC Based on Reed-Solomon Codes The remainder of this paper focuses on coding that takes advantage of the unequal bit by bit error probabilities.k) in the absence error location information is Pmo = Po 12 Pm.k) . as individual codewords are independent. 8P 4 OPo OP0 ' Mrb" 4P.k)/ 2 where n is the total number of characters per codeword and k is the number of information characters per codeword. because of the minimization of the dominate PI term. 16Pi mrb 8P2 OP2 12P 3 16P3 .Pi . 8P 4 OP 4 .-l([l-aU^sin(2t)lf . The character error correcting capability of the RS(n. To quantify the performance of LRBC with soft-decision information.

the mrb and Mrb look like gray coded QPSK. i . i . At this point the . Soft-decision quantization error is a factor. but not both. It is expected that the LRBC Pb3 Mrb ci = Pb Mrb Pdb Irb implementaion using commercially available VLSI technology will exhibit about a factor of 100 reduction in combined mass where x = mrb and Mrb. The LRBC without multi-stage . The reason decision parsing is slightly better than a natural mapping has to do with the mapping of the mrb and Mrb. ieven decoding requires either one time-shared RS codec operating at the full data rate. the M number of quantization bits (and thus the size of the lookup table) is being studied. M-1 M-1 Pbx=JM _ _^_^ (bj x ffibrx)(bjirt. Figure 11 illustrates a comparison between the best LRBC code with channel information.Pdb Irb) Pb1 Mrb ci = Pb Mrb (i s .ffib||rtJP|. the best LRBC code without channel information.2j > 0 Figure 11 is an existing implementation at 225 Mbps by COMSAT Laboratories [6].2j < 0 ECL gate arrays to perform the add-compare-select function of the 16 state Vrterbi decoder. in error is then The third type of error event is the probability of bit error given a Irb decoding errror or erasure has occurred.2j < 0 and thus Pb1 mrb ci = Pb mrb (1 .corrections.Pdb Irb) where Pdb Irb ' the decoded bit error probability of the Irb. Optimization for power efficient applications such as high rate return links This probability is plotted for the three mapping schemes using a RS(255. In the future. In decision parsing. so the most likely error event Pm4 will result in an error on either the mrb or Mrb. It requires a set of 8 semi-custom 2j-i -1 . Comparisons were presented on the different approaches to LRBC in terms of performance and complexity. At higher . The performance of a convolutional coding scheme plotted in i -2j. but can result in both a mrb and Mrb error. An existing equivalent convolutional coding scheme was also used for comparison. CONCLUSION AND FUTURE WORK Pb mrb ci =Pb1 mrb ci + Pb2 mrb ci + Pb3 mrb ci Pb Mrb ci =Pb1 Mrb ci + Pb2 Mrb ci + Pb3 Mrb ci and overall probability of bit error is Pb — (Pb Irb + Pb mrb ci +Pb Mrb ci) 3 An analysis of the performance of LRBC has been presented. mrb and Mrb corrections.most likely error event when the mapping is gray coded is dependent on Pm^. This is written as Eb/No (greater than approximately 7 dB in the gray coded case and 8 dB in the other cases) the RS decoded Irb begins to exhibit very low probability of error. a Pm^ decision region results in an Irb correction and thus if necessary. 2j-i -1. spectrally efficient 2 information bps/Hz LRBC can improve power efficiency over BPSK. or three parallel RS decoders 2M-1 £ (b j x ©bj±i x ) bj| r b ©bj. This maintains the ensemble code rate of 8/9. 171) code on the Irb and channel correction information on the mrb and Mrb in Figure 9. The use of concatenated coding schemes will also be investigated. and for comparison.j| 0 i= 0 Pb2 mrb ci = Pb mrb Pb2 Mrb ci = Pb Mrb The other issue is the hardware and computational complexity.where i-2j cannot equal 0 or 1 (not an error event) and the minor decision region index n is n= i -2j. In the other mappings. It is shown that a rate 8/9. An LRBC \ =0\ 2 l\ 21 scheme coding only the Irb and not using side information iodd would require only one decoder at 1/3 the data rate and some buffering. In this case gray coding is no longer the best mapping. Depending on the bit error rate a service class requires would dictate the code to be chosen. r j Pmn each operating at 1/3 the data rate. i . With a natural mapping. Pm4 is also the most likely error event.2i > 0 i . and Pb3 mrb ci = Pb mrb Pdb lib hard decision logic. The rate difference in the codes will also require a degree of buffering. Figure 10 illustrates the effect of using shorter RS codes on the decision parsed mapping. the best uncoded 8PSK scheme. in fact it is the worst. For completeness. uncoded BPSK scheme. The second type of error event is the probability of bit error when no Irb correction or error has ocurred. The probability that a mrb or Mrb are and power consumption. If the amount of coding gain was acceptable for again the minor decision region index n is the application. The LRBC scheme with multi-stage decoding needs one RS codec operating at 1/3 of the information data rate. a board of state memory. other types of block codes will be compared to the performance of the RS codes. point from a time domain simulation (with a union bound based extrapolated data point) of a rate 8/9 convolutional coding system [6]. it would be the simplest to implement.. This is evaluated as IV. Multi-stage decoding requires a buffer to wait for M-1 2M-1 codeword decisions and a lookup table to make the = _! i jx © b±2 JI b j jrb© b_i |rJ Pmn.

M. the terms of Q(x) are found by the ratio of the k and k-1 terms. Presently.e The error term e that results in an upper bound for Ps is defined as .ik-k! where 7 is Euler's constant.-2Q . March 1985. NASA Conference Publication 3132. IEEE. J. IEEE Trans.-| / k-k! yk-1 \ _ y (k-1) REFERENCES [1] Imai. [3] M. J. no. 371-377.: OnBoard Switching and Processing. 3. pp. Hatzipapafotiou. vol.. The ratio of the k to k-1 terms is yk The erfc(x) can be written in series notation as terrrik termk. Prentice Hall. "Mathematical Handbook of Formulas and Tables". [9] M. May 1977. March 1991. Comm. "Bounds and Approximations for Rapid Evaluation of Coherent MPSK Error Probabilities". Comm. Irshid and I. H. to General Electric Company Military and Data Systems Operations. 2*(2k + 1) -k! term k . July 1990. 3. Proceedings of the Space Communications Technology Conference : Onboard Processing and Switching. "Digital Communications Fundamentals and Applications". COM-33. L. 7.V . vol. No. S. [6] COMSAT Laboratories. x2 (2k . "Final Report: Advanced Modulation Technology Development for Earth Station Demodulator Applications". 39. NAS3-24680..from the space exploration initiative program will be addressed. the direct evaluation of this alternating series becomes more inaccurate. Proc. 306. J. This ratio is written as [4] B. IEEE Trans. 271-273. and Stevens.. [8] C.ln(y) + £ -*— k. Sklar.1206-1212. Comm. Spiegel. McGraw-Hill Book Company. 3. [5] D. It will be integrated into a programmable modem developed by COMSAT Laboratories for NASA Lewis [7]. 78. No. submitted to NASA Lewis Research Center under NASA contract no. Salous. No. 349-352. vol. pp. pp.Ej(y) Ei(y) 271 'frh where y is y = 3Eb No and Ej (y) is the exponential integral defined in [9] as the series 00 k Ei(y) = -y . pp.erfc For the 8PSK case x is defined as ^ No -sirte) \8I A exp.. "Codulation Study Final Report". 183. Hirakawa. Thus. APPENDIX Q(x) is defined as Q(x) = J. January 1989. M. vol.: "A New Multilevel Coding Method Using Error-Correcting Codes". "Bit Error Probability for Coherent M-ary PSK Systems". H. a LRBC hardware prototype is being developed at NASA Lewis Research Center. Chie. 1968. S. [7] Poklemba. November 1991.1) 2 (2k + 1 )• k APPENDIX B A tight bound on the probability of symbol error for 8PSK is written as [8] P.. Budinger. As k becomes large. 1988.. pp. The LRBC hardware will be compared to COMSAPs Coded Trellis Modulation (CTM) hardware in a laboratory environment. 263-271. 23. pp. pp. I. IEEE Trans.G.: "Programmable Digital Modem". 5-2! in series notation this reduces to 7-3! [2] Harrold.

Mrb .0 T Decision Parsing A2 / A3 \ / A1 AO A5 I i \ A6 A7 Figure 1: 8PSK Decision Regions 6 8 10 14 Eb/No(dB) Figure 2: Uncoded 8PSK Comparison 0 T 0 -r -2 -2 -4 Irb :k=171 mrb: None Mrb: None mrb.Q Q. -6 -6 0) o §> Irb :k=213 mrb: k=233 Mrb: k=233 lib: k=227 mrb: k=227 Mrb: k=227 Irb :k=199 mrb: k=241 Mrb: k=241 -8 -8 -1 0 -10 -1 H—I-H—I—HH—h-H—I—I—I—I—I—I—I—I 202 4 6 8 10 12 14 Eb / No (dB) Figure 3: Gray Coding Bit by Bit Error Probabilities -12 -I—I—I—h 2 4 -I—I—r6 8 10 Eb / No (dB) Figure 4 : LRBC with Natural Mapping .

0 T o T Irb :k=213 mrb: k=233 Mrb: k=233 Irb: k=227 mrb: k=227 Mrb: k=227 Irb : k=207 mrb: k=237 Mrb: k=237 lib :k=213 mrb: k=233 Mrb: k=233 Irb: k=227 mrb: k=227 Mrb: k=227 -12 6 8 1 0 14 2 4 6 8 10 Eb/No(dB) Figure 5: LRBC with Gray Coding Eb / No (dB) Figure 6: LRBC with Decision Parsing 0 T \B5 B6 B7 Natural Mapping Irb :k=213 mrb: k=233 Mrb: k=233 B4/ / B3^W / / B2 B8 iV'''' B1 B9 \ \ N ' BO B15 \B14 B13v \ v -12 12 14 Figure 8: Minor Decision Regions Eb / No (dB) Figure 7: LRBC Comparison .

85)RS(63. 171)RS(127.0 T 0 T -2 -2 Gray Coded -4 RS(31.21) RS(255. 43) b> o -10 -12 H—I—I 6 8 6 8 1 0 12 14 JO 12 Eb / No (dB) Figure 9: Multl-Stage LRBC Eb / No (dB) Figure 10: Multi-Stage Decision Mapped LRBC 0 T Gray Coded 8PSK BPSK -10 -12 2 4 6 8 10 14 Eb/No(dB) Figure 11: Overall Comparison .