Edf 96 T

,OW
COMPLEXITY ALGORITHMS IN DIGITAL RECEIVERS

/VE %DFORS $IVISION OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3 ,5,% 3WEDEN 3EPTEMBER
II
!BSTRACT
4HIS THESIS ADDRESSES LOW COMPLEXITY ALGORITHMS IN DIGITAL RECEIVERS 4HIS INCLUDES ALGO RITHMS FOR ESTIMATION DETECTION AND SOURCE CODING ,OW COMPLEXITY ALGORITHMS FOR ESTIMATION AND DETECTION IN THIS THESIS CONCERNS THE APPROXIMATION OF OPTIMAL ALGORITHMS SO THAT A LOW COMPLEXITY IS OBTAINED WHILE MOST OF THE ORIGINAL PERFORMANCE IS MAINTAINED 4WO DIdERENT PROBLEMS ARE STUDIED #HANNEL ESTIMATION IN ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- SYSTEMS AND SEQUENCE DETECTION IN SYSTEMS WITH INTER SYMBOL INTERFERENCE )3) 4HE WORK ON CHANNEL ESTIMATION IS FOCUSED ON WHAT CAN BE VIEWED AS TRANSFORM BASED ESTIMATORS WHERE THE ESTIMATION IS TRANSFORMED TO A DOMAIN THAT ALLOWS THE STATISTICAL PROPERTIES OF THE CHANNEL TO BE USED EbCIENTLY FOR COMPLEXITY REDUCTION %STIMATORS BASED BOTH ON THE DISCRETE &OURIER TRANS FORM $&4 AND ON LOW RANK APPROXIMATIONS ARE ADDRESSED )T IS SHOWN THAT THESE TYPES OF ESTIMATORS WITH PROPER DESIGN CONSIDERATIONS PROVIDE GOOD LOW COMPLEXITY ESTIMATORS EVEN THOUGH THE $&4 BASED ESTIMATORS HAVE A TENDENCY TO SUdER FROM APPROXIMATION ER RORS AT HIGH SIGNAL TO NOISE RATIOS 4HE WORK ON SEQUENCE DETECTION IN SYSTEMS WITH )3) IS A COMPARATIVE STUDY OF DESIGNS IN A CLASS USUALLY REFERRED TO AS COMBINED LINEAR 6ITERBI EQUALIZERS #,6%S 4HE IDEA BEHIND #,6%S IS TO SHORTEN THE IMPULSE RESPONSE OF THE CHANNEL BY A LINEAR PRELTERING BEFORE THE 6ITERBI DETECTOR IS APPLIED 4HE COMPARISON INCLUDES THREE PREVIOUSLY KNOWN DESIGNS AND A MINIMAX DESIGN ! UNIED DESIGN FRAMEWORK IS ALSO DERIVED WHICH MAKES THE COMPARISON EASIER 4HE LAST TWO PARTS OF THE THESIS ARE CONCERNED WITH THE DESIGN OF ROBUST VARIABLE LENGTH CODES AND AN ANALYSIS OF THE DECODING SPEED FOR LOOK UP TABLE BASED DECODERS FOR VARIABLE LENGTH CODES RESPECTIVELY "Y ROBUST VARIABLE LENGTH CODES ARE MEANT XED VARIABLE LENGTH CODES THAT HAVE A LOW SUSCEPTIBILITY TO CHANGES IN THE PROBABILITY DISTRIBUTION OF THE SOURCE 2OBUSTNESS MEASURES ARE DISCUSSED IN GENERAL AND A MEASURE CALLED GRADIENT ROBUSTNESS IS INTRODUCED )T IS SHOWN THAT ONLY A SMALL LOSS IN THE DEGREE OF DATA COMPRESSION CAN RESULT IN SUBSTAN TIALLY INCREASED ROBUSTNESS !S A SIDE EdECT THE PROPOSED DESIGN METHODS ALSO SHORTEN THE LENGTH OF THE LONGEST CODE WORDS ALLOWING EbCIENT STORAGE OF CODE BOOKS &INALLY THE DESIGN AND THE DECODING SPEED OF LOOK UP TABLE BASED DECODERS FOR VARIABLE LENGTH CODES ARE STUDIED 4HESE DECODERS ARE BASED ON USING THE CODED SEQUENCE FOR INDEXING A TABLE THEREBY PROVIDING AN INSTANTANEOUS DECODING 4WO TYPES OF DECODERS ARE STUDIED /NE ALLOWS A TRADE Od BETWEEN DECODING SPEED AND MEMORY REQUIREMENT
III
IV
#ONTENTS
0REFACE !CKNOWLEDGMENTS 4HESIS SUMMARY %STIMATION AND DETECTION 3OURCE CODING VII IX
%STIMATION AND DETECTION

!N ! INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING )NTRODUCTION 3YSTEM MODELS 3YSTEM ENVIRONMENTS 3YNCHRONIZATION #HANNEL ESTIMATION #HANNEL CODING $ISCUSSION 4IME FREQUENCY LATTICE ESTIMATORS FOR /&$

!NALYSIS OF $&4 BASED CHANNEL )NTRODUCTION 3YSTEM MODEL $&4 BASED ESTIMATORS 0ERFORMANCE ANALYSIS #ONCLUSIONS ! %STIMATOR EXPRESSIONS
/&$- CHANNEL ESTIMATION BY SINGULAR VALUE DECOMPOSITION )NTRODUCTION 3YSTEM DESCRIPTION ,INEAR CHANNEL ESTIMATION ACROSS TONES %STIMATOR PERFORMANCE AND DESIGN 'ENERIC LOW RANK ESTIMATOR #ONCLUSIONS V
VI ! /PTIMAL RANK REDUCTION " #HANNEL CORRELATION MATRICES # %STIMATOR MEAN SQUARED ERROR ! COMPARATIVE STUDY OF /&$ )NTRODUCTION 3YSTEM DESCRIPTION %STIMATORS 0ERFORMANCE EVALUATION #ONCLUSIONS ! #OMPLEXITY OF LOW RANK " #ORRELATION MATRICES PILOT BASED CHANNEL ESTIMATORS FOR WIRELESS ESTIMATORS
#OMBINED LINEAR 6ITERBI EQUALIZERS ! COMPARATIVE STUDY AND A MIN IMAX DESIGN )NTRODUCTION 4HE TRANSMISSION SYSTEM MODEL #,6% DESIGN 3IMULATIONS #ONCLUSIONS AND FUTURE WORK
3OURCE CODING
$ESIGN AND ANALYSIS OF ROBUST SOURCE CODES )NTRODUCTION 3YSTEM DESCRIPTION AND BASIC NOTATIONS 2OBUSTNESS %XPERIMENTAL EVALUATION OF GRADIENT ROBUSTNESS $ISCUSSION AND CONCLUSIONS ! !NALYSIS OF THE GRADIENT ROBUSTNESS " 0ROPERTIES OF GRADIENT ROBUST ENTROPY CLOSE %0, CODES # ! NUMERICAL OPTIMIZATION METHOD

! PERFORMANCE ANALYSIS OF LOOK UP TABLE BASED VARIABLE LENGTH DECODERS )NTRODUCTION !LGORITHMS AND A HARDWARE PROPOSAL $ECODING ALGORITHM PROPERTIES #ONCLUSIONS ! #ODING EXAMPLES " ! 0# IMPLEMENTATION # .OTATIONS
0REFACE
4HIS DOCTORAL THESIS IS A RESULT OF MY TIME AS A 0H$ STUDENT AT THE $IVISION OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY ) HAVE BEEN INVOLVED IN THREE DIdERENT PROJECTS SINCE ) ENROLLED AS A 0H$ STUDENT IN "ECAUSE OF THIS ) HAVE HAD THE OPPORTUNITY TO STUDY MANY DIdERENT ASPECTS OF SIGNAL PROCESSING -Y PERSONAL INTEREST IN SIGNAL PROCESSING IS QUITE BROAD AND ) HAVE ENJOYED WORKING WITH DIdERENT APPLICATIONS ) STARTED WITH SOURCE CODING IN AND CONTINUED UNTIL ) RECEIVED MY LICENTIATE DEGREE IN 3HORTLY AFTER THIS ) HAD THE OPPORTUNITY TO WORK WITH LOW COMPLEXITY EQUALIZERS IN DIGITAL RECEIVERS A DIdERENT PROJECT WITHIN THE $IVISION OF 3IGNAL 0ROCESSING ) CHOSE TO GO ON AN EXCURSION ALONG THIS NEW AT LEAST TO ME AND INTERESTING PATH LYING IN FRONT OF ME $URING THIS PERIOD ) BECAME MORE INVOLVED IN OUR ONGOING COOPERATION WITH 4ELIA 2ESEARCH !" ,ULE WHICH THEN LED ME IN THE DIRECTION OF ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- 4HE COOPERATION WITH 4ELIA 2ESEARCH HAS HELPED ME TO ND INTER ESTING AND RELEVANT RESEARCH PROBLEMS AND HAS ALSO RESULTED IN SEVERAL JOINT PUBLICATIONS )N ADDITION TO OUR COOPERATION WITH INDUSTRY WE HAVE A COOPERATION WITH 0ROFESSOR 3ARAH +ATE 7ILSON 0URDUE 5NIVERSITY WHICH HAS STRENGTHENED THE TELECOMMUNICATIONS ASPECTS OF OUR RESEARCH 4HIS HAS ALSO RESULTED IN JOINT PUBLICATIONS OF WHICH SEVERAL ARE INCLUDED IN THIS THESIS )T IS ARMED WITH EXPERIENCES FROM THE ABOVE JOURNEY THROUGH THE WORLD OF SIGNAL PROCESSING THAT ) TOGETHER WITH MY CO AUTHORS HAVE PRODUCED THE MATERIAL PRESENTED IN THIS THESIS &INALLY ) WOULD LIKE TO ADD THAT ) HAVE CHOSEN TO ORGANIZE THIS THESIS INTO TWO TOPICS MOSTLY BECAUSE ) BELIEVE THEY GENERALLY APPLY TO DIdERENT AUDIENCES 4HE TWO TOPICS ARE %STIMATION AND DETECTION AND 3OURCE CODING
VII
VIII
!CKNOWLEDGMENTS
)N THE LINE OF WORK AS A 0H$ STUDENT ) HAVE HAD THE OPPORTUNITY TO GET TO KNOW A LOT OF PEOPLE WHO HAVE BOTH ENRICHED MY LIFE ON THE GREATER SCALE AND HAVE DIRECTLY ASSISTED ME IN MY WORK 4HERE ARE SO MANY WHO HAVE BEEN SO IMPORTANT THAT ) WOULD LIKE TO ACKNOWLEDGE THEM ALL HERE AND NOW BUT WHAT SCARES ME AND PROBABLY MANY IN THIS SITUATION IS THAT THE MIND MAY SLIP AND SOMEONE NOT BE MENTIONED ) AM INDEBTED TO YOU ALL &IRST OF ALL ) WOULD LIKE TO EXTEND MY GRATITUDE TOWARDS MY ADVISOR 0ROFESSOR 0ER /LA "RJESSON WHO HAS SUPPORTED ME IN MANY WAYS DURING THESE YEARS 7E HAVE HAD NUMEROUS DISCUSSIONS ABOUT ALMOST EVERYTHING AND AT TIMES WHEN DISCUSSING INTERESTING IDEAS ALMOST FORGOT ABOUT THE WORK AT HAND 4HE CURIOSITY WE SHARE HAS BEEN A GOOD TEACHER ) AM ALSO IN GREAT DEBT TO 0ROFESSOR 3ARAH +ATE 7ILSON 0URDUE 5NIVERSITY 53! 7E HAD JUST STARTED WORKING ON ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING WHEN SHE VISITED US FOR THE RST TIME IN "EING AN EXPERT ON THE NAMED SYSTEMS SHE HAS GUIDED ME IN A WAY THAT ) AM ETERNALLY GRATEFUL FOR ) HAVE ALSO HAD THE PRIVILEGE TO SPEND SIX MONTHS AS HER GUEST AT 0URDUE 5NIVERSITY IN ) WOULD ALSO LIKE TO ACKNOWLEDGE THE SUPPORT ) HAVE RECEIVED FROM MY CURRENT AND PAST COLLEAGUES AT THE $IVISION OF 3IGNAL 0ROCESSING ) HAVE HAD A GREAT TIME BOTH SOCIALLY AND ON THE JOB %SPECIALLY ) WOULD LIKE TO THANK *AN *AAP VAN DE "EEK AND -AGNUS 3ANDELL WITH WHOM ) HAVE WORKED VERY CLOSELY THE LAST FEW YEARS ! GREAT SOURCE FOR INSPIRATION HAS BEEN THE !PPLIED 3YSTEMS $EVELOPMENT 'ROUP AT 4ELIA 2ESEARCH !" ,ULE /UR FREQUENT DISCUSSIONS AND MEETINGS HAVE SPAWNED MANY INTERESTING QUESTIONS AND OUR WORK TOGETHER HAS HELPED ME TO A BETTER UNDERSTANDING OF THE TERM TECHNICAL RELEVANCE 0AUL 0ETERSEN WHO ) GOT TO KNOW THROUGH 0ROFESSOR 7ILSON HAS BEEN OF GREAT ASSIS TANCE 7E HAVE HAD MANY IN DEPTH DISCUSSIONS ABOUT EVERYTHING FROM TECHNICAL EDITING PARTS OF THIS THESIS TO !MERICAN HISTORY 3OMEONE WHO HAS NOT DIRECTLY AdECTED THE CONTENTS OF THIS THESIS BUT STILL HAS HAD A SIGNICANT IMPACT ON MY GENERAL KNOWLEDGE IN THE AREA OF SIGNAL PROCESSING IS $R #HRISTOPHER +OH /UR DISCUSSIONS ON DIdERENT APPLICATIONS OF SIGNAL PROCESSING HAVE BEEN BOTH INTERESTING AND EDUCATIONAL ,AST BUT NOT LEAST ) WOULD LIKE TO THANK MY FAMILY AND MY FRIENDS WHO HAVE PUT UP WITH ME DURING THESE YEARS ) KNOW IT MAY HAVE SOMETIMES SEEMED THAT ) CHOSE MY WORK BEFORE THEM BUT ) AM GRATEFUL FOR ALL THEIR PATIENCE AND SUPPORT AND ) ONLY HOPE THEY WILL FORGIVE ME IF ) HAVE DISAPPOINTED THEM "ECAUSE OF YOU THIS WORK HAS NOT ONLY BEEN POSSIBLE IT HAS BEEN A PLEASURE IX
4HESIS SUMMARY
4HIS THESIS CONTAINS MATERIAL FROM THREE DIdERENT TOPICS THAT ) HAVE STUDIED 4HE THREE TOPICS ARE CONCERNED WITH CHANNEL ESTIMATION IN ORTHOGONAL FREQUENCY DIVISION MULTIPLEX ING /&$- SYSTEMS COMBINED LINEAR 6ITERBI EQUALIZERS #,6%S AND VARIABLE LENGTH CODING ) HAVE CHOSEN TO PRESENT THE MATERIAL UNDER TWO MAIN TOPICS CHANNEL ESTIMATION IN /&$- AND #,6%S UNDER %STIMATION AND DETECTION 0ARTS AND VARIABLE LENGTH CODING UNDER 3OURCE CODING 0ARTS 4HE TWO MAIN TOPICS ARE TREATED SEPARATELY EVEN IF MOST OF THE MATERIAL IS DEALING WITH COMPLEXITY VERSUS PERFORMANCE 4HE SEVEN PARTS OF THIS THESIS ARE REPRODUCTIONS OF PUBLICATIONS WHOSE ORIGINAL REFER ENCES ARE DISPLAYED BELOW 4HE CHANGES MADE ARE THE FOLLOWING 4O MAKE CROSS REFERENCING AND CITATIONS UNIQUE SECTION AND EQUATION NUMBERING ARE CHANGED AND ALL REFERENCES ARE COLLECTED IN A SINGLE BIBLIOGRAPHY AT THE END
;= / %DFORS - 3ANDELL * * VAN DE "EEK $ ,ANDSTRM AND & 3JBERG !N INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT ;= / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON !NALYSIS OF $&4 BASED CHANNEL ESTIMATORS FOR /&$- 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT 2ELATED PUBLICATION ;= ;= / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON /&$CHANNEL ESTIMATION BY SINGULAR VALUE DECOMPOSITION 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT 0UBLISHED IN PART AS ; = ;= - 3ANDELL AND / %DFORS ! COMPARATIVE STUDY OF PILOT BASED CHANNEL ESTIMATORS FOR WIRELESS /&$- 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NI VERSITY OF 4ECHNOLOGY 3EPT 2ELATED PUBLICATIONS ; = ;= . 3UNDSTRM / %DFORS 0 DLING ( %RIKSSON 0 / "RJESSON AND 4 +OSKI #OM BINED LINEAR 6ITERBI EQUALIZERS ! COMPARATIVE STUDY AND A MINIMAX DESIGN )N 0ROC )%%% 6EHIC 4ECHNOL #ONF PAGES 3TOCKHOLM 3WEDEN *UNE ;= / %DFORS AND 0 / "RJESSON $ESIGN AND ANALYSIS OF ROBUST SOURCE CODES 2ESEARCH 2EPORT 45,%! ,ULE 5NIVERSITY OF 4ECHNOLOGY ,ULE ;= / %DFORS ! %RENDI AND 0 / "RJESSON ! PERFORMANCE ANALYSIS OF LOOK UP TABLE BASED VARIABLE LENGTH DECODERS 2ESEARCH 2EPORT 45,%! ,ULE 5NIVERSITY OF 4ECHNOLOGY ,ULE 0UBLISHED IN PART AS ;=

4HE MAIN CONCERN OF THE MATERIAL PRESENTED IN 0ARTS IS THE APPROXIMATION OF OPTIMAL ALGORITHMS ALLOWING A LOW COMPLEXITY WHILE PRESERVING MOST OF THE ORIGINAL PERFORMANCE )N THE AREA OF DIGITAL COMMUNICATION LOW COMPLEXITY SIGNAL PROCESSING SOLUTIONS ARE BE COMING MORE AND MORE IMPORTANT &OR INSTANCE INCREASING DEMANDS ON SPECTRAL EbCIENCY IN WIRELESS SYSTEMS AND EbCIENT USE OF EXISTING COPPER NETWORKS CONTINUALLY PROMPT FOR MORE SOPHISTICATED SOLUTIONS -ORE SOPHISTICATED SOLUTIONS USUALLY MEANS INCREASED HARD WARE COMPLEXITY AND INCREASED POWER CONSUMPTION +EEPING THE HARDWARE COMPLEXITY AND THE POWER CONSUMPTION DOWN IS THEREFORE OF INTEREST IN FUTURE COMMUNICATION SYS TEMS 4HIS IS ESPECIALLY TRUE FOR MOBILE COMMUNICATION SYSTEMS WHERE PORTABILITY AND BATTERY LIFE TIME ARE CRITICAL 4HIS MATERIAL CONCERNING LOW COMPLEXITY ALGORITHMS DEALS WITH CHANNEL ESTIMATION IN /&$- SYSTEMS AND SEQUENCE DETECTION IN SYSTEMS WITH INTER SYMBOL INTERFERENCE )3) #HANNEL ESTIMATION IN /&$- SYSTEMS 4HERE ARE SEVERAL WAYS OF IMPLEMENTING /&$- SYSTEMS 3OME OF THESE ARE DESCRIBED IN 0ART WHICH IS AN INTRODUCTION TO /&$- AND AN OVERVIEW OF THE RESEARCH AREAS CURRENTLY PURSUED BY THE $IVISION OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY /NE WAY OF IMPLEMENTING AN /&$- SYSTEM IS SCHEMATICALLY DISPLAYED IN &IGURE WHERE - SIGNAL CONSTELLATION POINTS W ARE DIGITALLY MODULATED ONTO - SUBCARRIERS BY AN INVERSE DISCRETE &OURIER TRANSFORM )$&4 4HE OUTPUT FROM THE )$&4 IS THEN TRANSMITTED OVER THE CHANNEL AND NOISE IS ADDED /N THE RECEIVER SIDE THE DEMODULATION IS PERFORMED BY A DISCRETE &OURIER TRANSFORM $&4 WHICH YIELDS THE - NOISY OUTPUTS X
J J
W W
(#%3
.OISE #HANNEL
#%3
X X
W- `
X- `
&IGURE 3CHEMATIC PICTURE OF A DIGITALLY IMPLEMENTED /&$- SYSTEM 5NDER CERTAIN CONDITIONS SEE 0ART THE SYSTEM IN &IGURE CAN BE MODELLED AS SHOWN IN &IGURE 4HIS MODEL CONSISTS OF - PARALLEL TRANSMISSION CHANNELS WITH AT TENUATIONS G AND ADDITIVE NOISE M )F THE /&$- SYSTEM IS TRULY ORTHOGONAL THERE IS NO CROSSTALK BETWEEN TRANSMISSION CHANNELS AND NO )3) IS PRESENT IN THE SYSTEM IE CON SECUTIVE /&$- SYMBOLS TRANSMITTED OVER THE CHANNEL DO NOT INTERFERE WITH ONE ANOTHER
J J
%XPRESSING THE /&$- SYSTEM IN THE FORM OF A MATRIX EXPRESSION YIELDS X 7G

M
W W
G G
M M
X X
W- `
G- `
M- ` X- `
&IGURE /&$- SYSTEM MODELLED AS - PARALLEL CHANNELS WITH CORRELATED ATTENUATIONS AND ADDITIVE NOISE WHERE 7 IS A DIAGONAL MATRIX WITH THE TRANSMITTED CONSTELLATION POINTS ON ITS DIAGONAL G IS A VECTOR OF CHANNEL ATTENUATIONS X IS A VECTOR OF RECEIVED DATA AND M IS A NOISE VECTOR +NOWING 7 THE STRAIGHT FORWARD LINEAR MINIMUM MEAN SQUARED ERROR ,--3% ES TIMATE OF THE CHANNEL G IS ;= B 1 1 X G
GX XX
WHERE 1 IS THE CROSS CORRELATION MATRIX BETWEEN THE CHANNEL G AND THE RECEIVED DATA X AND 1 IS THE AUTO CORRELATION MATRIX OF THE RECEIVED DATA %VEN IF THIS ESTIMATOR IS DESIGNED FOR XED CORRELATION MATRICES THE COMPUTATIONAL COMPLEXITY BECOMES PRO HIBITIVELY LARGE THE ESTIMATION REQUIRES - MULTIPLICATIONS IN AN - `SUBCARRIER SYSTEM 4WO METHODS FOR LOWERING THE COMPUTATIONAL COMPLEXITY OF THIS /&$- CHANNEL ESTIMA TOR ARE INVESTIGATED IN THIS THESIS 0ARTS "OTH METHODS ARE TRANSFORM BASED AND A SCHEMATIC STRUCTURE OF THE ESTIMATORS IS DISPLAYED IN &IGURE !N INTUITIVE EXPLA NATION OF THE GENERAL APPROACH IS THAT BY CHOOSING A PROPER TRANSFORM DOMAIN WE CAN CONCENTRATE THE POWER OF THE CHANNEL TO A FEW TRANSFORM COEbCIENTS AND PERFORM THE ESTIMATION IN A SPACE OF LOWER DIMENSION 4HIS IS POSSIBLE SINCE THE /&$- SYMBOLS ARE BY DESIGN MUCH LONGER THAN THE IMPULSE RESPONSE OF THE TRANSMISSION CHANNEL 4HE FREQUENCY RESPONSE IS THEREFORE OVERSAMPLED IN X AND CAN BE DESCRIBED ACCURATELY IN A LOWER DIMENSION &OR THIS APPROACH TO YIELD A LOW COMPLEXITY SOLUTION THE COMPLEXITY OF THE TRANSFORMS AND THE ESTIMATION IN THE TRANSFORM DOMAIN HAS TO HAVE A COMBINED COMPLEXITY LOWER THAN THE ORIGINAL - MULTIPLICATIONS )N 0ART WE INVESTIGATE THE POSSIBILITY OF USING THE $&4 AS A POWER CONCENTRATING TRANSFORM 4HE $&4 APPROACH HAS ALSO BEEN SUGGESTED FOR DISCRETE TIME CHANNELS BY #HINI ; = /UR ANALYSIS IS DIdERENT FROM THE ANALYSIS IN ;= IN THAT WE USE CONTINUOUS TIME CHANNEL MODELS &OR LOW 3.2S THIS TYPE OF ESTIMATOR PERFORMS RELATIVELY WELL BUT IN ORDER TO AVOID AN ERROR OOR INHERENT IN THIS TYPE OF ESTIMATOR A LARGE NUMBER OF TRANSFORM COEbCIENTS ARE NEEDED /NE OF THE ANALYZED ESTIMATORS WHICH IS A NEW DESIGN BOTH OBTAINS ITS MAXIMAL PERFORMANCE AND MAINTAINS ITS LOW COMPLEXITY WHEN ALL COEbCIENTS ARE USED )N 0ART WE APPLY THE THEORY OF OPTIMAL RANK REDUCTION ;= TO THE SAME CHANNEL ESTIMATION PROBLEM 4HESE LOW RANK ESTIMATORS CAN ALSO BE INTERPRETED IN THE STRUCTURE
GX XX
&IGURE 3CHEMATIC PICTURE OF LOW COMPLEXITY CHANNEL ESTIMATORS OF &IGURE WHERE THE TRANSFORMS NOW DEPEND ON THE STATISTICAL PROPERTIES OF THE CHAN NEL 3INCE THIS APPROACH YIELDS A BETTER OPTIMAL CONCENTRATION OF CHANNEL POWER IN THE TRANSFORM DOMAIN THE ,--3% ESTIMATOR CAN BE APPROXIMATED WITH THE SAME ACCURACY IN A SMALLER SUBSPACE THAN WITH $&4)$&4 (OWEVER AT THE SAME TIME THE TRANSFORMS ARE NO LONGER FAST AND THIS INCREASES THE COMPLEXITY PER USED TRANSFORM COEbCIENT )N THIS PART IT IS ALSO SHOWN THAT A XED LOW RANK ESTIMATOR DESIGN FOR A WORST CASE CHANNEL AND A HIGH 3.2 GIVES A GOOD ESTIMATOR PERFORMANCE UNDER MISMATCH 4HIS IS SIMILAR TO THE RECOMMENDATIONS FOR THE DESIGN OF SINGLE CARRIER PILOT ASSISTED MODULATION 03!SYSTEMS BY #AVERS IN ;= 4HE INVESTIGATED CHANNEL ESTIMATORS HAVE SO FAR ONLY EXPLOITED THE FREQUENCY CORRELA TION OF THE CHANNEL )N 0ART WE TAKE THE ANALYSIS ONE STEP FURTHER BY INCLUDING TIME CORRELATION AS WELL 4O DO THIS SEVERAL CONSECUTIVE /&$- SYMBOLS HAVE TO BE OBSERVED AND WE MODIFY THE MODEL TO X 7G
M
K K K K

WHERE K IS THE TIME INDEX 4HE ANALYZED SYSTEM IS BASED ON A SCENARIO WHERE THE CHANNEL ESTIMATORS USE SCATTERED PILOTS 4HIS MEANS THAT ONLY A SMALL FRACTION OF THE TRANSMITTED DATA IN THE 7 S ARE KNOWN AT THE RECEIVER 4HE OPTIMAL LINEAR CHANNEL ESTIMATOR IN THIS SITUATION IN TERMS OF MEAN SQUARED ERROR IS A TWO DIMENSIONAL $ 7IENER LTER (OWEVER IN ;= (HER ARGUES THAT THE $ 7IENER LTER IS OF TOO HIGH A COMPLEXITY AND PROPOSES THE USE OF SEPARABLE LTERS 7E DESIGN LOW RANK APPROXIMATIONS OF BOTH THE $ 7IENER LTER AND THE SEPARABLE LTER AND COMPARE THE PERFORMANCES OF ALL LTERS 4HE COMPARISON SHOWS THAT THE USE OF SEPARABLE LTERS CAN INCREASE THE PERFORMANCE SUBSTANTIALLY COMPARED TO $ LTERS WITH THE SAME COMPLEXITY &URTHER IF LOW RANK APPROXIMATIONS OF THE TYPE DISCUSSED IN 0ART ARE ADDED TO THESE SEPARABLE LTERS THEN THE PERFORMANCE IMPROVES EVEN MORE IF ONLY MARGINALLY SO
K
3EQUENCE DETECTION IN SYSTEMS WITH INTER SYMBOL INTERFERENCE #OMBINED LINEAR 6ITERBI EQUALIZERS #,6%S ARE STUDIED IN 0ART ! DISCRETE TIME MODEL OF THE ADDRESSED SYSTEM IS DISPLAYED IN &IGURE WHERE W IS A SEQUENCE OF TRANSMITTED DATA G THE CHANNEL IMPULSE RESPONSE AND M IS WHITE 'AUSSIAN NOISE
J J J
MJ WJ GJ XJ
&IGURE #HANNEL MODEL USED IN THE ANALYSIS OF COMBINED LINEAR 6ITERBI EQUALIZERS 4HE OPTIMAL DETECTOR FOR THE SEQUENCE W IS KNOWN AS THE 6ITERBI DETECTOR ;= BUT ITS DRAWBACK IS THE EXPONENTIAL GROWTH OF COMPLEXITY AS THE LENGTH OF THE CHANNEL IMPULSE RESPONSE G INCREASES /NE METHOD TO REDUCE THE COMPLEXITY OF THIS EQUALIZER IS TO APPLY LINEAR PRELTERING TO THE RECEIVED DATA TO SHORTEN THE EdECTIVE IMPULSE RESPONSE OF THE CHANNEL BEFORE APPLYING THE 6ITERBI DETECTOR 4HE GENERAL STRUCTURE OF THIS DETECTOR SHOWN IN &IGURE IS OFTEN REFERRED TO AS A COMBINED LINEAR 6ITERBI EQUALIZER #,6% ; =
J J
XJ
,INEAR PRELTER
BJ 6ITERBI W DETECTOR
&IGURE 4HE #,6% IS A COMBINATION OF A LINEAR PRELTER AND A 6ITERBI DETECTOR 7HEN DESIGNING #,6%S IT IS OFTEN DESIRABLE TO MINIMIZE THE BIT ERROR RATE OF THE RECEIVER 4HE BIT ERROR PROBABILITY DEPENDS ON THE DESIGN PARAMETERS SUCH AS THE CHANNEL MODEL AND THE IMPULSE RESPONSE OF THE PRELTER IN A COMPLICATED AND NON LINEAR WAY 4HE PRE LTERING OF THE RECEIVED DATA PERTURBS THE SIGNAL SPACE AND COLOURS THE CHANNEL NOISE )GNORING THIS COLOURING OR GIVING THE 6ITERBI DETECTOR AN APPROXIMATE CHANNEL MODEL RESULT IN A DISPLACEMENT OF THE DECISION REGIONS FROM THEIR OPTIMAL LOCATIONS CF THE RESIDUAL )3) IN ;= )NSTEAD OF USING THE BIT ERROR RATE AS A DESIGN CRITERION OTHER MORE FEASIBLE CRITERIA ARE USED IN #,6% DESIGN METHODS 4HE CONTRIBUTION IN THIS PART OF THE THESIS IS A COMPARATIVE STUDY OF #,6% DESIGNS FOUND IN THE LITERATURE ; = AND A MINIMAX DESIGN ! UNIED DESIGN FRAMEWORK IS DERIVED WHICH MAKES THE COMPARISON EASIER
3OURCE CODING
6ARIABLE LENGTH CODING BASED ON 3HANNONS 3OURCE #ODING 4HEOREM ;= IS A COMMON DATA COMPRESSION TECHNIQUE 4HE CODE CONSTRUCTION METHODS EG THE ONE PRESENTED BY (UdMAN ;= USE THE PRINCIPLE THAT SYMBOLS WITH A HIGH PROBABILITY OF APPEARANCE SHOULD REQUIRE LESS STORAGE SPACE THAN LESS FREQUENT SYMBOLS 4HIS IS DONE BY ASSIGNING SHORT CODE WORDS TO THE HIGH PROBABILITY SYMBOLS AND LONG CODE WORDS TO THE LOW PROBABILITY ONES %NTROPY CODES OF THIS TYPE SUCH AS VARIABLE LENGTH CODES ARE FOUND AS AN IMPORTANT INTEGRATED PART IN MANY SOURCE CODING SCHEMES 4HE LAST TWO PARTS OF THIS THESIS ADDRESS THE DESIGN OF ROBUST VARIABLE LENGTH CODES 0ART AND FAST LOOK UP TABLE DECODERS 0ART FOR VARIABLE LENGTH CODES
0ART IS CONCERNED WITH ROBUSTNESS MEASURES TO BE USED WITH VARIABLE LENGTH CODES AND THE DESIGN OF ROBUST CODES 2OBUSTNESS IN THIS CASE IS DENED AS LOW AVERAGE CODE WORD LENGTH SUSCEPTIBILITY TO CHANGES IN THE PROBABILITY DISTRIBUTION OF THE SOURCE )N MANY APPLICATIONS THE PROBABILITY DISTRIBUTION OF THE SOURCE IS NEITHER WELL KNOWN NOR TIME INVARIANT )N SUCH A SITUATION AN ADAPTIVE SOURCE CODER SEE EG ; = CAN BE APPLIED TO ACHIEVE GOOD DATA COMPRESSION (OWEVER MANY APPLICATIONS DO NOT DEPEND ON THE BEST POSSIBLE COMPRESSION A HIGH COMPRESSION SPEED OR PORTABLE LOW COST HARDWARE MAY BE MORE IMPORTANT 4HIS OPENS A NICHE FOR ROBUST STATIC SOURCE CODES 4HERE ARE SEVERAL WAYS TO INTRODUCE WHAT WE CALL ROBUSTNESS &OR INSTANCE BY RE STRICTING THE LENGTH OF THE LONGEST CODE WORDS LIKE IN ;= THE RESULTING CODE WILL HAVE A CERTAIN ROBUSTNESS 2OBUSTNESS IS ALSO OBTAINED THROUGH THE SYMBOL SET PARTITIONING USED IN ; = 4HE SITUATION WHERE CODES ARE DESIGNED USING AN INCORRECT ESTIMATE OF THE PROBABILITY DISTRIBUTION HAS BEEN COVERED BY A NUMBER OF AUTHORS SEE EG ; = 7E INTRODUCE A ROBUSTNESS MEASURE NAMED GRADIENT ROBUSTNESS 4HIS MEASURE IS BASED ON THE GRADIENT OF CHANGE OF THE AVERAGE CODE WORD LENGTH WHEN THE SOURCE PROBABILITY DISTRIBUTION CHANGES $ESIGNING ROBUST CODES ACCORDING TO THIS ROBUSTNESS MEASURE ALSO YIELDS CODES WITH A RELATIVELY SMALL VARIATION IN CODE WORD LENGTHS &OR CERTAIN PROBABIL ITY DISTRIBUTIONS ONLY A SMALL LOSS IN AVERAGE CODE WORD LENGTH RESULTS IN A SUBSTANTIAL INCREASE OF ROBUSTNESS )N 0ART WE ANALYZE TWO DECODING ALGORITHMS FOR BINARY VARIABLE LENGTH CODES BASED ON A LOOK UP TABLE APPROACH ! VARIABLE LENGTH CODED SEQUENCE CAN BE DECODED BY A TREE SEARCH "Y MATCHING THE BITS IN THE SEQUENCE AGAINST THE BRANCHES IN THE CODE TREE ONE WILL UPON ARRIVAL AT A LEAF HAVE DECODED ONE SOURCE SYMBOL (OWEVER THE TREE SEARCH IS NOT A VERY FAST DECODING METHOD )F THE AVERAGE CODE WORD LENGTH IS 6 BITS A TRIVIAL TREE SEARCH WILL REQUIRE AN AVERAGE OF 6 MATCHINGS BEFORE A SOURCE SYMBOL IS DECODED !N EbCIENT DECODER OF THIS TYPE WAS PRESENTED BY ( 4ANAKA IN ;= 4HE APPROACH WE HAVE CHOSEN IS TO COMPLETE THE CODE TREE TO A CERTAIN DEPTH OF EVERY BRANCH THUS ALLOWING A TABULAR DECODING WHERE ONE OR SEVERAL SOURCE SYMBOLS ARE DECODED PER CYCLE OF THE ALGORITHM $ECODING IS DONE BY INDEXING A TABLE WITH A XED LENGTH BLOCK FROM THE VARIABLE LENGTH CODE SEQUENCE ! NICE PROPERTY OF ONE OF THE DECODING ALGORITHMS IS THAT IT ALLOWS A TRADE Od BETWEEN DECODING SPEED AND MEMORY REQUIREMENT 6ARIATIONS ON THE THEME HAVE BEEN KNOWN FOR SOME TIME SEE EG ;= WHERE SHORT BLOCKS OF COMPRESSED DATA ARE DECODED USING A SET OF CONNECTED LOOK UP TABLES
0ART !N INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING

!BSTRACT 4HIS REPORT IS AN INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- 4HE FOCUS IS ON SIGNAL PROCESSING AREAS PURSUED BY OUR RESEARCH GROUP AT ,ULE 5NIVERSITY OF 4ECHNOLOGY 7E PRESENT AN HISTORICAL BACKGROUND AND SOME FREQUENTLY USED SYSTEM MODELS 4YPICAL AREAS OF APPLICATIONS ARE ALSO DESCRIBED BOTH WIRELESS AND WIRED )N ADDITION TO THE GENERAL OVERVIEW THE ADDRESSED AREAS INCLUDE SYNCHRONIZATION CHANNEL ESTIMATION AND CHANNEL CODING "OTH TIME AND FREQUENCY SYNCHRONIZATION ARE DESCRIBED AND THE EdECTS OF SYNCHRONIZATION ERRORS ARE PRESENTED $IdERENT TYPES OF CHANNEL ESTIMA TORS ARE DESCRIBED WHERE THE FOCUS IS ON LOW COMPLEXITY ALGORITHMS AND IN THIS CONTEXT ADVANTAGES AND DISADVANTAGES OF COHERENT AND DIdERENTIAL MODULATION ARE ALSO DISCUSSED #HANNEL CODING IS DESCRIBED BOTH FOR WIRELESS AND WIRED SYSTEMS AND POINTERS ARE INCLUDED TO EVALUATION TOOLS AND BITLOADING ALGORITHMS !N EXTENSIVE BIBLIOGRAPHY IS ALSO INCLUDED
4HIS PART IS REPRODUCED FROM / %DFORS - 3ANDELL * * VAN DE "EEK $ ,ANDSTRM AND & 3JBERG !N INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT

)NTRODUCTION TO /&$-

)NTRODUCTION
4HE AIM OF THIS REPORT IS TWOFOLD 4HE RST AIM IS TO PROVIDE AN INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- SYSTEMS AND SELECTED PARTS OF ITS THEORETICAL BACKGROUND 4HE SECOND AIM IS TO DESCRIBE THE AREAS OF RESEARCH WITHIN /&$- THAT ARE PURSUED AT THE $IVISION OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY 4HIS ALSO INCLUDES A BY NO MEANS COMPLETE DESCRIPTION OF RELATED WORK THAT MAY BE OF INTEREST 4HE PRESENTATION IS IN THE FORM OF A SINGLE BODY WHERE WE DO NOT SEPARATE OUR OWN WORK FROM THAT BY OTHERS 4HE TECHNOLOGY WE CALL /&$- IN THIS REPORT IS USUALLY VIEWED AS A COLLECTION OF TRANSMISSION TECHNIQUES 7HEN APPLIED IN A WIRELESS ENVIRONMENT SUCH AS RADIO BROAD CASTING IT IS USUALLY REFERRED TO AS /&$- (OWEVER IN A WIRED ENVIRONMENT SUCH AS IN ASYMMETRIC DIGITAL SUBSCRIBER LINES !$3, THE TERM DISCRETE MULTITONE $-4 IS MORE APPROPRIATE 4HROUGHOUT THIS REPORT WE ONLY USE THE TERM $-4 WHEN EXPLICITLY AD DRESSING THE WIRED ENVIRONMENT &URTHER THE TWO TERMS SUBCARRIER AND SUBCHANNEL WILL BE USED INTERCHANGEABLY 4HE HISTORY OF /&$- HAS BEEN ADDRESSED SEVERAL TIMES IN THE LITERATURE SEE EG ; = WHICH WE HAVE CONDENSED TO THE BRIEF OVERVIEW BELOW 4HE HISTORY OF /&$- DATES BACK TO THE MID S WHEN #HANG PUBLISHED HIS PAPER ON THE SYNTHESIS OF BANDLIMITED SIGNALS FOR MULTICHANNEL TRANSMISSION ;= (E PRESENTS A PRINCIPLE FOR TRANSMITTING MESSAGES SIMULTANEOUSLY THROUGH A LINEAR BANDLIMITED CHAN NEL WITHOUT INTERCHANNEL )#) AND INTERSYMBOL INTERFERENCE )3) 3HORTLY AFTER #HANG PRESENTED HIS PAPER 3ALTZBERG PERFORMED AN ANALYSIS OF THE PERFORMANCE ;= WHERE HE CONCLUDED THAT THE STRATEGY OF DESIGNING AN EbCIENT PARALLEL SYSTEM SHOULD CONCENTRATE MORE ON REDUCING CROSSTALK BETWEEN ADJACENT CHANNELS THAN ON PERFECTING THE INDIVIDUAL CHANNELS THEMSELVES SINCE THE DISTORTIONS DUE TO CROSSTALK TEND TO DOMINATE 4HIS IS AN IMPORTANT CONCLUSION WHICH HAS PROVEN CORRECT IN THE DIGITAL BASEBAND PROCESSING THAT EMERGED A FEW YEARS LATER ! MAJOR CONTRIBUTION TO /&$- WAS PRESENTED IN BY 7EINSTEIN AND %BERT ;= WHO USED THE DISCRETE &OURIER TRANSFORM $&4 TO PERFORM BASEBAND MODULATION AND DEMODULATION 4HIS WORK DID NOT FOCUS ON PERFECTING THE INDIVIDUAL CHANNELS BUT RATHER ON INTRODUCING EbCIENT PROCESSING ELIMINATING THE BANKS OF SUBCARRIER OSCILLATORS 4O COMBAT )3) AND )#) THEY USED BOTH A GUARD SPACE BETWEEN THE SYMBOLS AND RAISED COSINE WINDOWING IN THE TIME DOMAIN 4HEIR SYSTEM DID NOT OBTAIN PERFECT ORTHOGONALITY BETWEEN SUBCARRIERS OVER A DISPERSIVE CHANNEL BUT IT WAS STILL A MAJOR CONTRIBUTION TO /&$- !NOTHER IMPORTANT CONTRIBUTION WAS DUE TO 0ELED AND 2UIZ IN ;= WHO IN TRODUCED THE CYCLIC PREX #0 OR CYCLIC EXTENSION SOLVING THE ORTHOGONALITY PROBLEM )NSTEAD OF USING AN EMPTY GUARD SPACE THEY LLED THE GUARD SPACE WITH A CYCLIC EXTENSION OF THE /&$- SYMBOL 4HIS EdECTIVELY SIMULATES A CHANNEL PERFORMING CYCLIC CONVOLU TION WHICH IMPLIES ORTHOGONALITY OVER DISPERSIVE CHANNELS WHEN THE #0 IS LONGER THAN THE IMPULSE RESPONSE OF THE CHANNEL 4HIS INTRODUCES AN ENERGY LOSS PROPORTIONAL TO THE LENGTH OF THE #0 BUT THE ZERO )#) GENERALLY MOTIVATES THE LOSS /&$- SYSTEMS ARE USUALLY DESIGNED WITH RECTANGULAR PULSES BUT RECENTLY THERE HAS BEEN AN INCREASED INTEREST IN PULSE SHAPING ; = "Y USING PULSES OTHER THAN RECTANGULAR THE SPECTRUM CAN BE SHAPED TO BE MORE WELL LOCALIZED IN FREQUENCY WHICH IS

BENECIAL FROM AN INTERFERENCE POINT OF VIEW /&$- IS CURRENTLY USED IN THE %UROPEAN DIGITAL AUDIO BROADCASTING $!" STANDARD ;= 3EVERAL $!" SYSTEMS PROPOSED FOR .ORTH !MERICA ARE ALSO BASED ON /&$- ;= AND ITS APPLICABILITY TO DIGITAL 46 BROADCASTING IS CURRENTLY BEING INVESTIGATED ; = /&$- IN COMBINATION WITH MULTIPLE ACCESS TECHNIQUES ARE SUBJECT TO SIGNICANT INVESTIGATION SEE EG ; = /&$- UNDER THE NAME $-4 HAS ALSO ATTRACTED A GREAT DEAL OF ATTENTION AS AN EbCIENT TECHNOLOGY FOR HIGH SPEED TRANSMISSION ON THE EXISTING TELEPHONE NETWORK SEE EG ; = 4HIS REPORT IS ORGANIZED AS FOLLOWS )N 3ECTION WE PRESENT COMMON /&$- MODELS INCLUDING CONTINUOUS TIME AND DISCRETE TIME %NVIRONMENTS IN WHICH /&$- SYSTEMS ARE EXPECTED TO WORK ARE SUMMARIZED IN 3ECTION 3YNCHRONIZATION PROBLEMS AND PROPOSED SOLUTION ARE PRESENTED IN 3ECTION #HANNEL ESTIMATION IS ELABORATED ON IN 3ECTION AND CODING IN BOTH WIRELESS AND WIRED /&$- SYSTEMS IS DISCUSSED IN 3ECTION &INALLY IN 3ECTION WE DISCUSS AND SUMMARIZE THE CONTENTS OF THIS REPORT

3YSTEM MODELS
4HE BASIC IDEA OF /&$- IS TO DIVIDE THE AVAILABLE SPECTRUM INTO SEVERAL SUBCHANNELS SUBCARRIERS "Y MAKING ALL SUBCHANNELS NARROWBAND THEY EXPERIENCE ALMOST AT FADING WHICH MAKES EQUALIZATION VERY SIMPLE 4O OBTAIN A HIGH SPECTRAL EbCIENCY THE FREQUENCY RESPONSE OF THE SUBCHANNELS ARE OVERLAPPING AND ORTHOGONAL HENCE THE NAME /&$- 4HIS ORTHOGONALITY CAN BE COMPLETELY MAINTAINED EVEN THOUGH THE SIGNAL PASSES THROUGH A TIME DISPERSIVE CHANNEL BY INTRODUCING A CYCLIC PREX 4HERE ARE SEVERAL VERSIONS OF /&$- SEE EG ; = BUT WE FOCUS ON SYSTEMS USING SUCH A CYCLIC PREX ;= ! CYCLIC PREX IS A COPY OF THE LAST PART OF THE /&$- SYMBOL WHICH IS PREPENDED TO THE TRANSMITTED SYMBOL SEE &IGURE 4HIS MAKES THE TRANSMITTED SIGNAL PERIODIC WHICH
&IGURE 4HE CYCLIC PREX IS A COPY OF THE LAST PART OF THE /&$- SYMBOL PLAYS A DECISIVE ROLL IN AVOIDING INTERSYMBOL AND INTERCARRIER INTERFERENCE ;= 4HIS IS EXPLAINED LATER IN THIS SECTION !LTHOUGH THE CYCLIC PREX INTRODUCES A LOSS IN SIGNAL TO NOISE RATIO 3.2 IT IS USUALLY A SMALL PRICE TO PAY TO MITIGATE INTERFERENCE ! SCHEMATIC DIAGRAM OF A BASEBAND /&$- SYSTEM IS SHOWN IN &IGURE &OR THIS SYSTEM WE EMPLOY THE FOLLOWING ASSUMPTIONS q ! CYCLIC PREX IS USED q 4HE IMPULSE RESPONSE OF THE CHANNEL IS SHORTER THAN THE CYCLIC PREX q 4RANSMITTER AND RECEIVER ARE PERFECTLY SYNCHRONIZED
)NTRODUCTION TO /&$4RANSMITTER
WK WK

#HANNEL
E S M
2ECEIVER
QS QXJ`
3 % # (
7 4 ,
4G
RXJ`
52
RS
F~ & S
25
4G
7 4 , $ #
XK XK
3 % #
X- ` K
W- ` K
&IGURE ! DIGITAL IMPLEMENTATION OF A BASEBAND /&$- SYSTEM #0 AND #0DENOTE THE INSERTION AND DELETION OF THE CYCLIC PREX RESPECTIVELY q #HANNEL NOISE IS ADDITIVE WHITE AND COMPLEX 'AUSSIAN q 4HE FADING IS SLOW ENOUGH FOR THE CHANNEL TO BE CONSIDERED CONSTANT DURING ONE /&$- SYMBOL INTERVAL 4HE DIbCULTIES IN A COMPLETE ANALYSIS OF THIS SYSTEM MAKE IT RATHER AWKWARD FOR THEORETICAL STUDIES 4HEREFORE IT IS COMMON PRACTICE TO USE SIMPLIED MODELS RESULTING IN A TRACTABLE ANALYSIS 7E CLASSIFY THESE /&$- SYSTEM MODELS INTO TWO DIdERENT CLASSES CONTINUOUS TIME AND DISCRETE TIME

#ONTINUOUS TIME MODEL
4HE RST /&$- SYSTEMS DID NOT EMPLOY DIGITAL MODULATION AND DEMODULATION (ENCE THE CONTINUOUS TIME /&$- MODEL PRESENTED BELOW CAN BE CONSIDERED AS THE IDEAL /&$SYSTEM WHICH IN PRACTICE IS DIGITALLY SYNTHESIZED 3INCE THIS IS THE RST MODEL DESCRIBED WE MOVE THROUGH IT IN A STEP BY STEP FASHION 7E START WITH THE WAVEFORMS USED IN THE TRANSMITTER AND PROCEED ALL THE WAY TO THE RECEIVER 4HE BASEBAND MODEL IS SHOWN IN &IGURE
4RANSMITTER
W W W
K
#HANNEL
E S M
K
S S
` S
RS
F~ & S
QS
2ECEIVER K + 3 S S
` S
X X X
K
K
` K
` K
&IGURE "ASE BAND /&$- SYSTEM MODEL q 4RANSMITTER !SSUMING AN /&$- SYSTEM WITH - SUBCARRIERS A BANDWIDTH OF 6 (Z AND SYMBOL LENGTH OF 3 SECONDS OF WHICH 3 SECONDS IS THE LENGTH OF THE CYCLIC PREX THE TRANSMITTER USES THE FOLLOWING WAVEFORMS 6 P D - BO IF S : 3 < BO S OTHERWISE
BO I { J S 3 J 3 3

BO J J
)NTRODUCTION TO /&$WHERE 3 -6

3 .OTE THAT S S
-6 WHEN S IS WITHIN THE CYCLIC PREX : 3 < 3INCE S IS A RECTANGULAR PULSE MODULATED ON THE CARRIER FREQUENCY J6- THE COMMON INTERPRETATION OF /&$- IS THAT IT USES - SUBCARRIERS EACH CARRYING A LOW BIT RATE 4HE WAVEFORMS S ARE USED IN THE MODULATION AND THE TRANSMITTED BASE BAND SIGNAL FOR /&$- SYMBOL NUMBER K IS
BO J J -
R S
K K K K
8
J (
W S ` K3
JK J
WHERE W W W ARE COMPLEX NUMBERS FROM A SET OF SIGNAL CONSTELLATION POINTS 7HEN AN INNITE SEQUENCE OF /&$- SYMBOLS IS TRANSMITTED THE OUTPUT FROM THE TRANSMITTER IS A JUXTAPOSITION OF INDIVIDUAL /&$- SYMBOLS RS
8
K(
R S
K
8 8
-
K(
W S ` K3
JK J

J (
q 0HYSICAL CHANNEL 7E ASSUME THAT THE SUPPORT OF THE POSSIBLY TIME VARIANT IMPULSE RESPONSE F ~ S OF THE PHYSICAL CHANNEL IS RESTRICTED TO THE INTERVAL ~ : 3 < IE TO THE LENGTH OF THE CYCLIC PREX 4HE RECEIVED SIGNAL BECOMES : BO Q S FcR S F ~ S R S ` ~ C~
M ES
BO 3
WHERE M E S IS ADDITIVE WHITE AND COMPLEX 'AUSSIAN CHANNEL NOISE q 2ECEIVER 4HE /&$- RECEIVER CONSISTS OF A LTER BANK MATCHED TO THE LAST PART :3 3 < OF THE TRANSMITTER WAVEFORMS S IE | 3 ` S IF S : 3 ` 3 < S OTHERWISE
BO J J BO J
%dECTIVELY THIS MEANS THAT THE CYCLIC PREX IS REMOVED IN THE RECEIVER 3INCE THE CYCLIC PREX CONTAINS ALL )3) FROM THE PREVIOUS SYMBOL THE SAMPLED OUTPUT FROM THE RECEIVER LTER BANK CONTAINS NO )3) (ENCE WE CAN IGNORE THE TIME INDEX K WHEN CALCULATING THE SAMPLED OUTPUT AT THE J TH MATCHED LTER "Y USING AND WE GET : X Qc S J( Q S 3 ` S CS : : BO : 8 F ~ S W S ` ~ C~ S CS
M E 3 ` S S CS BO BO (
J J S 3 J 3 3 3 J J J J 3 J 3
7E CONSIDER THE CHANNEL TO BE XED OVER THE /&$- SYMBOL INTERVAL AND DENOTE IT BY F ~ WHICH GIVES u : : t: BO 8 X W F ~ S ` ~ C~ S CS
M E 3 ` S S CS BO BO (
3 3 3 J J J

J
)NTRODUCTION TO /&$4HE INTEGRATION INTERVALS ARE 3 S 3 AND ~ 3 S ` ~ 3 AND THE INNER INTEGRAL CAN BE WRITTEN AS
BO BO
WHICH IMPLIES THAT
: BO
3
F ~
S ` ~ C~
: BO
3
F ~

I {J
I {J
P 3 `3
BO
3
6BO
BO P C~ 3 `3 : BO F ~ D

6-
BO
I {J ~ 6-
C~
3 S3
BO
4HE LATTER PART OF THIS EXPRESSION IS THE SAMPLED FREQUENCY RESPONSE OF THE CHANNEL AT FREQUENCY E J 6- IE AT THE J TH SUBCARRIER FREQUENCY t G &
J

6 J -
u
: BO
3
F ~ D
I {J ~ 6-
C~

WHERE & E IS THE &OURIER TRANSFORM OF F ~ 5SING THIS NOTATION THE OUTPUT FROM THE RECEIVER LTER BANK CAN BE SIMPLIED TO
-

8 8
: W
J 3
I {J
J ( -
BO :
3
P 3 `3
J
BO
3
6BO
J
: S CS
BO
M E 3 ` S S CS
J
W G
J
J
J (
BO
S S CS
M
J J

WHERE M NAL
J
E 3 ` S S CS 3INCE THE TRANSMITTER LTERS J S ARE ORTHOGO J BO M : S CS

3 3
:
3
BO
J
I {J
BO
BO P 3 `3

6-
I {JS
BO
3
6BO
BO
3 `3
CS p :J ` J <
WHERE p :J < IS THE +RONECKER DELTA FUNCTION ;= WE CAN SIMPLIFY AND OBTAIN X G W
M
J J J J

WHERE M IS ADDITIVE WHITE 'AUSSIAN NOISE !7'.

J
4HE BENET OF A CYCLIC PREX IS TWOFOLD IT AVOIDS BOTH )3) SINCE IT ACTS AS A GUARD SPACE AND )#) SINCE IT MAINTAINS THE ORTHOGONALITY OF THE SUBCARRIERS "Y RE INTRODUCING THE TIME INDEX K WE MAY NOW VIEW THE /&$- SYSTEM AS A SET OF PARALLEL 'AUSSIAN CHANNELS ACCORDING TO &IGURE !N EdECT TO CONSIDER AT THIS STAGE IS THAT THE TRANSMITTED ENERGY INCREASES WITH THE LENGTH OF THE CYCLIC PREX WHILE THE EXPRESSIONS FOR THE RECEIVED AND SAMPLED SIGNALS 2 STAY THE SAME 4HE TRANSMITTED ENERGY PER SUBCARRIER IS J S J CS 3 3 ` 3 AND THE 3.2 LOSS BECAUSE OF THE DISCARDED CYCLIC PREX IN THE RECEIVER BECOMES
J BO
2-1
KNRR
` KNF ` o

G W
K
K
M
K
X G ` M `
K K
K
W `
-
K
X `
-
K
&IGURE 4HE CONTINUOUS TIME /&$- SYSTEM INTERPRETED AS PARALLEL 'AUSSIAN CHANNELS
- RTAB@QQHDQR 2O@BHMF aE 6-
&IGURE ! SYMBOLIC PICTURE OF THE INDIVIDUAL SUBCHANNELS FOR AN /&$- SYSTEM WITH - TONES OVER A BANDWIDTH 6 WHERE o 3 3 IS THE RELATIVE LENGTH OF THE CYCLIC PREX 4HE LONGER THE CYCLIC PREX THE LARGER THE 3.2 LOSS 4YPICALLY THE RELATIVE LENGTH OF THE CYCLIC PREX IS SMALL AND THE )#) AND )3) FREE TRANSMISSION MOTIVATES THE 3.2 LOSS LESS THAN D" FOR o &IGURE DISPLAYS A SCHEMATIC PICTURE OF THE FREQUENCY RESPONSE OF THE INDIVIDUAL SUBCHANNELS IN AN /&$- SYMBOL )N THIS GURE THE INDIVIDUAL SUBCHANNELS OF THE SYSTEM ARE SEPARATED 4HE RECTANGULAR WINDOWING OF THE TRANSMITTED PULSES RESULTS IN A SINC SHAPED FREQUENCY RESPONSE FOR EACH CHANNEL 4HUS THE POWER SPECTRUM OF THE /&$SYSTEM DECAYS AS E )N SOME CASES THIS IS NOT SUbCIENT AND METHODS HAVE BEEN PROPOSED TO SHAPE THE SPECTRUM )N ;= A RAISED COSINE PULSE IS USED WHERE THE ROLL Od REGION ALSO ACTS AS A GUARD SPACE SEE &IGURE )F THE AT PART IS THE /&$- SYMBOL
BO
&IGURE 0ULSE SHAPING USING THE RAISED COSINE FUNCTION 4HE GRAY PARTS OF THE SIGNAL INDICATE THE EXTENSIONS

INCLUDING THE CYCLIC PREX BOTH )#) AND )3) ARE AVOIDED 4HE SPECTRUM WITH THIS KIND OF PULSE SHAPING IS SHOWN IN &IGURE WHERE IT IS COMPARED WITH A RECTANGULAR PULSE 4HE OVERHEAD INTRODUCED BY AN EXTRA GUARD SPACE WITH A GRACEFUL ROLL Od CAN BE A GOOD
&IGURE 3PECTRUM WITH RECTANGULAR PULSE SOLID AND RAISED COSINE PULSE DASHED INVESTMENT SINCE THE SPECTRUM FALLS MUCH MORE QUICKLY AND REDUCES THE INTERFERENCE TO ADJACENT FREQUENCY BANDS /THER TYPES OF PULSE SHAPING SUCH AS OVERLAPPING ;= AND WELL LOCALIZED PULSES ; = HAVE ALSO BEEN INVESTIGATED

$ISCRETE TIME MODEL
!N ENTIRELY DISCRETE TIME MODEL OF AN /&$- SYSTEM IS DISPLAYED IN &IGURE #OM PARED TO THE CONTINUOUS TIME MODEL THE MODULATION AND DEMODULATION ARE REPLACED BY AN INVERSE $&4 )$&4 AND A $&4 RESPECTIVELY AND THE CHANNEL IS A DISCRETE TIME CONVOLUTION 4HE CYCLIC PREX OPERATES IN THE SAME FASHION IN THIS SYSTEM AND THE CAL CULATIONS CAN BE PERFORMED IN ESSENTIALLY THE SAME WAY 4HE MAIN DIdERENCE IS THAT ALL INTEGRALS ARE REPLACED BY SUMS
4RANSMITTER
WK WK
#HANNEL
E :J< M
2ECEIVER
Q:J<
3 % # (
7 4 ,
"/
R:J<
F:LJ<
"/
7 4 , $ #
XK XK
3 % #
X- ` K
W- ` K
&IGURE $ISCRETE TIME /&$- SYSTEM &ROM THE RECEIVERS POINT OF VIEW THE USE OF A CYCLIC PREX LONGER THAN THE CHANNEL WILL TRANSFORM THE LINEAR CONVOLUTION IN THE CHANNEL TO A CYCLIC CONVOLUTION $ENOTING CYCLIC CONVOLUTION BY ] WE CAN WRITE THE WHOLE /&$- SYSTEM AS X
K
E #%3 (#%3 W ] F
M #%3 (#%3 W ] F
M
K K K K K K

K K
WHERE X CONTAINS THE - RECEIVED DATA POINTS W THE - TRANSMITTED CONSTELLATION POINTS F THE CHANNEL IMPULSE RESPONSE OF THE CHANNEL PADDED WITH ZEROS TO OBTAIN A LENGTH OF E THE CHANNEL NOISE 3INCE THE CHANNEL NOISE IS ASSUMED WHITE AND 'AUSSIAN - AND M E REPRESENTS UNCORRELATED 'AUSSIAN NOISE &URTHER WE USE THAT THE THE TERM M #%3 M $&4 OF TWO CYCLICALLY CONVOLVED SIGNALS IS EQUIVALENT TO THE PRODUCT OF THEIR INDIVIDUAL $&4S $ENOTING ELEMENT BY ELEMENT MULTIPLICATION BY a THE ABOVE EXPRESSION CAN BE WRITTEN X W a #%3 F
M W a G
M
K K K K K K K K K K K
WHERE G #%3 F IS THE FREQUENCY RESPONSE OF THE CHANNEL 4HUS WE HAVE OBTAINED THE SAME TYPE OF PARALLEL 'AUSSIAN CHANNELS AS FOR THE CONTINUOUS TIME MODEL 4HE ONLY DIdERENCE IS THAT THE CHANNEL ATTENUATIONS G ARE GIVEN BY THE - POINT $&4 OF THE DISCRETE TIME CHANNEL INSTEAD OF THE SAMPLED FREQUENCY RESPONSE AS IN
K K K

! TIME FREQUENCY INTERPRETATION
4HE MODELS DESCRIBED ABOVE ARE TWO CLASSICAL MODELS OF /&$- WITH A CYCLIC PREX ! MORE GENERAL MODEL SUITABLE FOR EG PULSE SHAPING IS TO VIEW /&$- AS TRANSMISSION OF DATA IN A LATTICE IN THE TIME FREQUENCY PLANE #ONSIDER RST A TRANSMITTED /&$- SIGNAL RS 8 RS W S
JK JK JK
WHERE THE FUNCTIONS S ARE TRANSLATIONS IN TIME BY ~ AND IN FREQUENCY BY y OF THE PROTOTYPE FUNCTION O S IE
JK
JK
S O S ` K~ D
I {Jy S
4HIS CREATES A TWO DIMENSIONAL $ LATTICE IN THE TIME FREQUENCY PLANE ; = SEE &IGURE 5SUALLY THE PROTOTYPE FUNCTION IS CHOSEN AS THE RECTANGULAR WINDOW O S y v S v ~ 4HE SPACING IN THE FREQUENCY DIRECTION IS THEN y ~ ` 3 WHERE 3 IS THE LENGTH OF THE CYCLIC PREX &OR A DISCUSSION ON THE IMPACT OF PROTOTYPE FUNCTIONS SEE !PPENDIX ! %ACH TRANSMITTED DATA SYMBOL IN THE LATTICE EXPERIENCES AT FADING SEE WHICH SIMPLIES EQUALIZATION AND CHANNEL ESTIMATION 4HE CHANNEL ATTENUATIONS AT THE LATTICE POINTS ARE CORRELATED AND BY TRANSMITTING KNOWN SYMBOLS AT SOME POSTIONS THE CHANNEL ATTENUATIONS CAN BE ESTIMATED WITH AN INTERPOLATION LTER ; = 4HIS IS A $ VERSION OF PILOT SYMBOL ASSISTED MODULATION WHICH HAS BEEN PROPOSED FOR SEVERAL WIRELESS /&$- SYSTEMS SEE EG ; = ! MORE DETAILED DESCRIPTION OF /&$- CHANNEL ESTIMATION IS GIVEN IN 3ECTION
~ BO BO

)MPERFECTIONS
$EPENDING ON THE ANALYZED SITUATION IMPERFECTIONS IN A REAL /&$- SYSTEM MAY BE IGNORED OR EXPLICITLY INCLUDED IN THE MODEL "ELOW WE MENTION SOME OF THE IMPERFECTIONS AND THEIR CORRESPONDING EdECTS

&IGURE ,ATTICE IN THE TIME FREQUENCY PLANE 4HE DATA SYMBOLS W AT THE LATTICE POINTS q $ISPERSION
JK
ARE TRANSMITTED
"OTH TIME AND FREQUENCY DISPERSION OF THE CHANNEL CAN DESTROY THE ORTHOGONALITY OF THE SYSTEM IE INTRODUCE BOTH )3) AND )#) ;= )F THESE EdECTS ARE NOT SUbCIENTLY MITIGATED BY EG A CYCLIC PREX AND A LARGE INTER CARRIER SPACING THEY HAVE TO BE INCLUDED IN THE MODEL /NE WAY OF MODELLING THESE EdECTS IS AN INCREASE OF THE ADDITIVE NOISE ;= q .ONLINEARITIES AND CLIPPING DISTORTION /&$- SYSTEMS HAVE HIGH PEAK TO AVERAGE POWER RATIOS AND HIGH DEMANDS ON LINEAR AMPLIERS ;= .ONLINEARITIES IN AMPLIERS MAY CAUSE BOTH )3) AND )#) IN THE SYSTEM %SPECIALLY IF THE AMPLIERS ARE NOT DESIGNED WITH PROPER OUTPUT BACK Od /"/ THE CLIPPING DISTORTION MAY CAUSE SEVERE DEGRADATION 4HESE EdECTS HAVE BEEN ADDRESSED IN EG ; = 3PECIAL CODING STRATEGIES WITH THE AIM TO MINIMIZE PEAK TO AVERAGE POWER RATIOS HAVE ALSO BEEN SUGGESTED SEE EG ; = q %XTERNAL INTERFERENCE "OTH WIRELESS AND WIRED /&$- SYSTEMS SUdER FROM EXTERNAL INTERFERENCE )N WIRE LESS SYSTEMS THIS INTERFERENCE USUALLY STEMS FROM RADIO TRANSMITTERS AND OTHER TYPES ELECTRONIC EQUIPMENT IN THE VINCINITY OF THE RECEIVER )N WIRED SYSTEMS THE LIMIT ING FACTOR IS USUALLY CROSSTALK WHICH IS DISCUSSED IN MORE DETAIL IN 3ECTION )NTERFERENCE CAN BE INCLUDED IN THE MODEL AS EG COLOURED NOISE

3YSTEM ENVIRONMENTS
4WO MAJOR GROUPS OF COMMUNICATION SYSTEMS ARE THOSE WHO OPERATE IN WIRELESS AND WIRED ENVIRONMENTS &OR INSTANCE WHEN DESIGNING A WIRELESS /&$- SYSTEM THE FADING CHANNEL

IS USUALLY A MAJOR OBSTACLE WHILE FOR A WIRED /&$- AKA $-4 SYSTEM CROSSTALK AND IMPULSIVE NOISE ARE MORE DIbCULT TO HANDLE )N THE FOLLOWING SECTIONS WE BRIEY DISCUSS THE WIRELESS AND WIRED ENVIRONMENTS

7IRELESS SYSTEMS
)N WIRELESS SYSTEMS RADIO SYSTEMS CHANGES IN THE PHYSICAL ENVIRONMENT CAUSE THE CHAN NEL TO FADE 4HESE CHANGES INCLUDE BOTH RELATIVE MOVEMENT BETWEEN TRANSMITTER AND RECEIVER AND MOVING SCATTERERSREECTORS IN THE SURROUNDING SPACE 7HEN DEVELOPING NEW STANDARDS FOR WIRELESS SYSTEMS CHANNEL MODELS ARE USUALLY CLAS SIED ACCORDING TO THE ENVIRONMENT IN WHICH THE RECEIVER OPERATES 4HESE ENVIRONMENTS ARE OFTEN DESCRIBED IN TERMS LIKE 2URAL AREA "USINESS INDOOR ETC -ODELS OF THIS TYPE ARE SPECIED BY EG THE %UROPEAN TELECOMMUNICATIONS STANDARDS INSTITUTE %43) ; = )N THEORETICAL STUDIES OF WIRELESS SYSTEMS THE CHANNEL MODELS ARE USUALLY CHOSEN SO THAT THEY RESULT IN A TRACTABLE ANALYSIS 4HE TWO MAJOR CLASSES OF FADING CHARACTERISTICS ARE KNOWN AS 2AYLEIGH AND 2ICIAN ;= ! 2AYLEIGH FADING ENVIRONMENT ASSUMES NO LINE OF SIGHT AND NO XED REECTORSSCATTERERS 4HE EXPECTED VALUE OF THE FADING IS ZERO )F THERE IS A LINE OF SIGHT THIS CAN BE MODELLED BY 2ICIAN FADING WHICH HAS THE SAME CHARACTERISTICS AS THE 2AYLEIGH FADING EXCEPT FOR A NON ZERO EXPECTED VALUE /FTEN PROPERTIES OF A THEORETICAL MODEL ARE CHARACTERIZED BY ONLY A FEW PARAMETERS SUCH AS POWER DELAY PROLE AND MAXIMAL $OPPLER FREQUENCY 4HE POWER DELAY PROLE | a DEPENDS ON THE ENVIRONMENT AND A COMMON CHOICE IS THE EXPONENTIALLY DECAYING PROLE | ~ D
~ ~
QLR
IS THE ROOT MEAN SQUARED 2-3 VALUE OF THE POWER WHERE ~ IS THE TIME DELAY AND ~ DELAY PROLE 3EVERAL OTHER CHOICES ARE POSSIBLE SEE EG ;= 4HE MAXIMAL $OPPLER FREQUENCY E vd CAN BE DETERMINED BY
QLR C
E
B
Cvd
E
U B
WHERE THE CARRIER FREQUENCY IS E (Z THE SPEED OF THE RECEIVER IS U MS AND THE SPEED OF LIGHT IS B { b # MS )SOTROPIC SCATTERING IS COMMONLY ASSUMED IE THE RECEIVED SIGNAL POWER IS SPREAD UNIFORMLY OVER ALL ANGLES OF ARRIVAL WHICH RESULTS IN A 5 SHAPED $OPPLER SPECTRUM 4HIS IS USUALLY REFERRED TO AS A *AKES SPECTRUM ;= AND IS DETERMINED BY THE MAXIMAL $OPPLER FREQUENCY "EFORE WE START DISCUSSING THE DIdERENT SCENARIOS ENCOUNTERED IN WIRELESS SYSTEMS THERE ARE A FEW THINGS THAT MAY BE SAID ABOUT /&$- ON FADING CHANNELS IN GENERAL 4HE INTER CARRIER SPACING OF THE SYSTEM HAS TO BE CHOSEN LARGE COMPARED TO THE MAXIMAL $OPPLER FREQUENCY OF THE FADING CHANNEL TO KEEP THE )#) SMALL ; = 4HIS IS FURTHER DISCUSSED IN !PPENDIX ! )F THE ORTHOGONALITY OF THE SYSTEM IS MAINTAINED THE BASIC /&$- STRUCTURE DOES NOT NECESSITATE TRADITIONAL EQUALIZING (OWEVER TO EXPLOIT THE DIVERSITY OF THE CHANNEL PROPER CODING AND INTERLEAVING IS REQUIRED ;=

7E HAVE CHOSEN TO DISCUSS THE WIRELESS ENVIRONMENT IN TWO CONTEXTS THE TRANSMISSION FROM A BASE STATION TO MOBILE TERMINALS DOWNLINK AND THE TRANSMISSION FROM MOBILE TERMINALS TO A BASE STATION UPLINK 4HE REASON FOR THE CHOSEN CONTEXTS IS THAT ONE OR BOTH USUALLY ARE REPRESENTED IN EVERY WIRELESS SYSTEM AND THEY REQUIRE QUITE DIdERENT DESIGN STRATEGIES 4HE MOST FREQUENTLY DISCUSSED WIRELESS /&$- SYSTEMS ARE FOR BROADCASTING EG DIGITAL AUDIO AND DIGITAL VIDEO AND ONLY CONTAIN A DOWNLINK SINCE THERE IS NO RETURN CHANNEL #ELLULAR SYSTEMS ON THE OTHER HAND HAVE BOTH A DOWNLINK AND AN UPLINK $OWNLINK
! SCHEMATIC PICTURE OF THE DOWNLINK ENVIRONMENT IS SHOWN IN &IGURE )N THIS CASE MOBILE TERMINAL NUMBER M RECEIVES THE SIGNAL R S TRANSMITTED FROM THE BASE STATION THROUGH ITS OWN CHANNEL F S AND THE RECEIVED SIGNAL Q S IS GIVEN BY
M M
S RcF
S
#H AN NE L
NEL #HAN
#HAN NEL
4ERMINAL
"ASE STATION
4ERMINAL +
4ERMINAL
&IGURE 4HE WIRELESS DOWNLINK ENVIRONMENT 4HIS ENVIRONMENT IMPLIES THAT EACH RECEIVER TERMINAL ONLY HAS TO SYNCHRONIZE TO THE BASE STATION AND FROM ITS POINT OF VIEW THE OTHER TERMINALS DO NOT EXIST 4HIS MAKES SYNCHRONIZATION RELATIVELY EASY AND ALL PILOT INFORMATION TRANSMITTED FROM THE BASE STATION CAN BE USED FOR CHANNEL ESTIMATION AND SYNCHRONIZATION 4HE DOWNLINK ENVIRONMENT HAS BEEN THOROUGHLY INVESTIGATED SEE EG ; = ,ARGE PORTIONS OF WORK PRESENTED ON SYSTEMS OF THIS KIND HAVE BEEN CONCERNED WITH DIGITAL AUDIO SEE EG ; = AND DIGITAL VIDEO SEE EG ; = BROADCASTING 5PLINK
M
! SCHEMATIC PICTURE OF THE UPLINK ENVIRONMENT IS SHOWN IN &IGURE )N THIS CASE THE BASE STATION RECEIVES THE TRANSMITTED SIGNAL R S FROM MOBILE TERMINAL M THROUGH CHANNEL F S AND THE TOTAL RECEIVED SIGNAL Q S AT THE BASE STATION IS A SUPERPOSITION
M
Q S
8
* M(
R cF
M

NEL #HAN
)NTRODUCTION TO /&$#HAN NEL
#H AN NE L
4ERMINAL
"ASE STATION
4ERMINAL +
4ERMINAL
&IGURE 4HE WIRELESS UPLINK ENVIRONMENT OF SIGNALS FROM ALL MOBILE TERMINALS 4HE MAJOR PROBLEM HERE IS THE SUPERPOSITION OF SIGNALS ARRIVING THROUGH DIdERENT CHANNELS &OR THE BASE STATION TO BE ABLE TO SEPARATE THE SIGNALS FROM EACH RECEIVER A SUbCIENT ORTHOGONALITY BETWEEN RECEIVED SIGNALS FROM DIdERENT TERMINALS HAS TO BE ACHIEVED 3EVERAL METHODS FOR OBTAINING THIS HAVE BEEN PROPOSED 4HESE INCLUDE COMBI NATIONS OF /&$- AND CODE DIVISION TIME DIVISION AND FREQUENCY DIVISION MULTIPLE ACCESS #$-! 4$-! AND &$-! RESPECTIVELY !LL THREE HAVE BEEN PROPOSED IN ;= AND &$-!/&$- IS CURRENTLY UNDER INVESTIGATION IN EG ; = )NDEPENDENT OF THE METHOD CHOSEN TO SEPARATE SIGNALS FROM DIdERENT TERMINALS THE SYSTEM SYNCHRONIZATION IS ONE OF THE MAJOR DESIGN ISSUES 4O AVOID INTERFERENCE ALL MO BILE TERMINALS HAVE TO BE JOINTLY SYNCHRONIZED TO THE BASE STATION &URTHER IF COHERENT MODULATION IS USED AS IN ; = THE DIdERENT CHANNELS FROM THE USERS HAVE TO BE ESTIMATED SEPARATELY

7IRED SYSTEMS
7HEN STUDYING WIRED COMMUNICATION SYSTEMS AND TRANSMISSION CHARACTERISTICS OF CABLES A DISTINCTION IS OFTEN MADE BETWEEN SHIELDED CABLES LIKE COAXIAL CABLES AND UNSHIELDED CABLES LIKE TWISTED WIRE PAIRS #OAXIAL CABLES HAVE MUCH BETTER TRANSMISSION PROPERTIES FOR BROADBAND SIGNALS THAN DO WIRE PAIRS %XCEPT FOR COMPUTER NETWORKS COAXIAL CABLES AND WIRE PAIRS CURRENTLY EXIST IN TWO BASICALLY DIdERENT NETWORK TOPOLOGIES 7IRE PAIRS ARE THE DOMINATING CABLE TYPE IN TELEPHONE ACCESS NETWORKS THAT ARE BUILT FOR POINT TO POINT AND TWO WAY COMMUNICATION #OAXIAL CABLES ARE USUALLY FOUND IN CABLE 46 SYSTEMS A NETWORK TOPOLOGY THAT IS PRIMARILY INTENDED FOR BROADCASTING AND NOT FOR POINT TO POINT COMMUNICATION 4HE CABLE 46 SYSTEMS SOMETIMES CONTAIN AMPLIERS THAT MAKE BIDIRECTIONAL COMMUNICATION ALMOST IMPOSSIBLE (OWEVER CABLE 46 NETWORKS ARE CURRENTLY BEING UPGRADED TO SUPPORT BIDIRECTIONAL COMMUNICATION 7E FOCUS ON WIRE PAIRS AND WILL NOT DISCUSS COAXIAL CABLES FURTHER 4HE COPPER WIRE PAIR DOES NOT CHANGE ITS PHYSICAL BEHAVIOR SIGNICANTLY WITH TIME AND IS THEREFORE CONSIDERED A STATIONARY CHANNEL ;= 4HIS MAKES IT POSSIBLE TO USE A TECHNIQUE CALLED BIT LOADING ;= SEE 3ECTION WHICH MAKES GOOD USE OF THE SPECTRALLY SHAPED CHANNEL 7HEN BIT LOADING IS USED IN A WIRED /&$- SYSTEM IT IS OFTEN REFERRED TO AS $-4 3INCE /&$- IN COMBINATION WITH BIT LOADING MAKES EbCIENT USE OF AVAILABLE BAND

WIDTH IT HAS BECOME A GOOD CANDIDATE FOR DIGITAL SUBSCRIBER LINE $3, SYSTEMS $3, IS ANOTHER NAME FOR DIGITAL HIGH SPEED COMMUNICATION IN THE TELEPHONE ACCESS NETWORK 7HEN THE BIT RATE OdERED IN DOWNSTREAM DIRECTION TO THE SUBSCRIBER IS LARGER THAN THE BIT RATE IN UPSTREAM DIRECTION TO THE BASE IT IS CALLED AN ASYMMETRIC DIGITAL SUBSCRIBER LINE !$3, !$3, IS SUITABLE FOR APPLICATIONS LIKE VIDEO ON DEMAND GAMES VIRTUAL SHOPPING INTERNET SURNG ETC WHERE MOST OF THE DATA GOES FROM THE BASE TO THE SUB SCRIBER )N THE 53! THERE EXISTS AN !$3, STANDARD THAT SUPPORTS DOWNSTREAM BIT RATES FROM TO -BITSS ;= 4HE BIT RATES OF THE UPSTREAM RETURN PATH USUALLY RANGES BETWEEN AND KBITSS ;= 3TANDARDS FOR SYMMETRICAL $3,S HAVE ALSO EMERGED TO SUPPORT VIDEO CONFERENCING AND OTHER SERVICES WITH HIGH DATA RATE IN THE UPSTREAM DIRECTION 4HE RST SYMMETRIC $3, SYSTEM WAS CALLED HIGH BIT RATE DIGITAL SUBSCRIBER LINE ($3, ;= WHICH CURRENTLY SUPPORTS BIT RATES BETWEEN AND -BITSSEC IN BOTH DIRECTIONS ; = &OR DIGITAL SUBSCRIBER LINES WITH HIGHER BIT RATES THAN ($3, AND !$3, THE TERM VERY HIGH BIT RATE DIGITAL SUBSCRIBER LINE 6$3, IS USED 3UBSCRIBER LINE TRANSFER FUNCTION
4HE CHARACTERISTICS OF THE WIRE PAIR CHANNEL HAVE BEEN STUDIED IN A NUMBER OF PAPERS ; = ! THOROUGH DESCRIPTION OF THE TRANSFER FUNCTION OF COPPER WIRES AND NOISE SOURCES IS GIVEN BY 7ERNER IN ;= &OR $3,S USING A LARGE FREQUENCY RANGE SEVERAL -(Z OR HIGHER THE ATTENUATION FUNCTION CAN BE APPROXIMATED AS J' E C J D
CJ
WHERE C IS THE LENGTH OF THE CABLE AND J IS A CABLE CONSTANT 4HIS MODEL IS OFTEN USED WHEN 6$3, AND ($3, SYSTEMS ARE ANALYZED ; = .OISE AND CROSSTALK
4HE MOST IMPORTANT NOISE SOURCES IN THE SUBSCRIBER LINE ENVIRONMENT ARE CROSSTALK FROM OTHER WIRE PAIRS IN THE SAME CABLE RADIO FREQUENCY 2& NOISE FROM NEARBY RADIO TRANS MITTERS AND IMPULSE NOISE FROM RELAYS SWITCHES ELECTRICAL MACHINES ETC !7'. IS GENERALLY NOT A LIMITING FACTOR IN DIGITAL SUBSCRIBER LINES FOR SHORT CABLES BUT BECOMES MORE IMPORTANT WITH INCREASING CABLE LENGTH )N EG ;= !'7. P IS INCLUDED IN THE CHANNEL MODEL WITH A SPECTRAL DENSITY OF ` D"M(Z x6 (Z )MPULSE NOISE IS DIbCULT TO CHARACTERIZE COMPLETELY BUT SOME EdORTS HAS BEEN MADE TO MODEL THIS KIND OF DISTURBANCES ; = 4HE NORMAL WAY TO MITIGATE THE EdECTS OF IMPULSE NOISE ON A $-4 SYSTEM IS TO ADD D" TO THE SYSTEM MARGIN ;= AND TO USE SPECIALLY DESIGNED CODES ;= )T SHOULD BE NOTED THAT $-4 IS MORE RESISTANT TO IMPULSE NOISE THAN SINGLE CARRIER SYSTEMS SUCH AS CARRIERLESS AMPLITUDEPHASE #!0 MODULATION ;= 4HE IMPACT OF 2& NOISE ON A $3, SYSTEM CAN BE REDUCED SIGNICANTLY WITH /&$AND BIT LOADING ;= 2& NOISE CAN BE MODELLED AS NARROWBAND DISTURBANCE WITH KNOWN SPECTRAL DENSITY AND THE BIT ERROR RATE "%2 CAN BE PRESERVED BY TRANSMITTING FEWER SOMETIMES ZERO BITS ON THE DISTURBED SUBCHANNELS

4HERE ARE BASICALLY TWO DIdERENT FORMS OF CROSSTALK NEAR END CROSSTALK .%84 AND FAR END CROSSTALK &%84 .%84 OCCURS AT THE CENTRAL ObCE BASE STATION WHEN THE WEAK UPSTREAM SIGNAL Q S IS DISTURBED BY STRONG DOWNSTREAM SIGNALS R S SEE &IGURE &%84 IS CROSSTALK FROM ONE TRANSMITTED SIGNAL R S TO ANOTHER Q S IN THE SAME DIRECTION SEE &IGURE AND APPEARS AT BOTH ENDS OF THE WIRE LOOP
&IGURE .EAR END CROSSTALK .%84
&IGURE &AR END CROSSTALK &%84 4HE SPECTRAL DENSITY OF .%84 IS MODELLED IN ;= AS /
-
E / E J E
R -

AND THE SPECTRAL DENSITY OF &%84 AS /

R %
E C / E J E J' E C J C / E J E D
R % R % -
CJ
C
%

WHERE / E IS THE SPECTRAL DENSITY OF THE TRANSMITTED SIGNALS J AND J ARE CONSTANTS DEPENDING ON THE TYPE OF CABLE HOW WELL BALANCED THE CABLES ARE AND THE NUMBER OF DISTURBING COPPER PAIRS ;= .OTE THAT .%84 DOES NOT DEPEND ON THE LENGTH OF THE WIRE PAIR )N &IGURE WE DISPLAY AN EXAMPLE OF THE SPECTRAL DENSITY OF A RECEIVED SIGNAL .%84 AND &%84

3YNCHRONIZATION
/NE OF THE ARGUMENTS AGAINST /&$- IS THAT IT IS HIGHLY SENSITIVE TO SYNCHRONIZATION ERRORS IN PARTICULAR TO FREQUENCY ERRORS ;= (ERE WE GIVE AN OVERVIEW OF THREE SYNCHRO NIZATION PROBLEMS SYMBOL CARRIER FREQUENCY AND SAMPLING FREQUENCY SYNCHRONIZATION !LSO THE EdECTS OF PHASE OdSETS AND PHASE NOISE ARE DISCUSSED
)NTRODUCTION TO /&$
/

2 J ' E CJ - E

NEXT
= " D N I Z ( 7 ; Y T I S N E D L A R T C E P S R E W O 0

% E C FEXT

&REQUENCY ;-(Z=

&IGURE 0OWER SPECTRAL DENSITY OF ATTENUATED SIGNAL .%84 AND &%84

3YMBOL SYNCHRONIZATION
4IMING ERRORS
! GREAT DEAL OF ATTENTION IS GIVEN TO SYMBOL SYNCHRONIZATION IN /&$- SYSTEMS (OWEVER BY USING A CYCLIC PREX THE TIMING REQUIREMENTS ARE RELAXED SOMEWHAT 4HE OBJECTIVE IS TO KNOW WHEN THE SYMBOL STARTS 4HE IMPACT OF TIMING ERRORS HAS BEEN ANALYZED IN ; = ! TIMING OdSET GIVES RISE TO A PHASE ROTATION OF THE SUBCARRIERS 4HIS PHASE ROTATION IS LARGEST ON THE EDGES OF THE FREQUENCY BAND )F A TIMING ERROR IS SMALL ENOUGH TO KEEP THE CHANNEL IMPULSE RESPONSE WITHIN THE CYCLIC PREX THE ORTHOGONALITY IS MAINTAINED )N THIS CASE A SYMBOL TIMING DELAY CAN BE VIEWED AS A PHASE SHIFT INTRODUCED BY THE CHANNEL AND THE PHASE ROTATIONS CAN BE ESTIMATED BY A CHANNEL ESTIMATOR )F A TIME SHIFT IS LARGER THAN THE CYCLIC PREX )3) WILL OCCUR 4HERE ARE TWO MAIN METHODS FOR TIMING SYNCHRONIZATION BASED ON PILOTS OR ON THE CYCLIC PREX !N ALGORITHM OF THE FORMER KIND WAS SUGGESTED BY 7ARNER AND ,EUNG IN ;= 4HEY USE A SCHEME WHERE THE /&$- SIGNAL IS TRANSMITTED BY FREQUENCY MODULATION &- 4HE TRANSMITTER ENCODES A NUMBER OF RESERVED SUBCHANNELS WITH KNOWN PHASES AND AMPLITUDES 4HE SYNCHRONIZATION TECHNIQUE WITH MODICATIONS IS APPLICABLE TO /&$SIGNALS TRANSMITTED BY AMPLITUDE MODULATION 4HEIR ALGORITHM CONSISTS OF PHASES POWER DETECTION COARSE SYNCHRONIZATION AND NE SYNCHRONIZATION 4HE RST PHASE POWER DETECTION DETECTS WHETHER OR NOT AN /&$- SIGNAL IS PRESENT BY MEASURING THE RECEIVED POWER AND COMPARE IT TO A THRESHOLD 4HE SECOND PHASE COARSE SYNCHRONIZATION IS USED TO ACQUIRE SYNCHRONIZATION ALIGNMENT TO WITHIN f SAMPLES 4HIS PERFORMANCE IS NOT ACCEPTABLE BUT THIS PHASE SERVES TO SIMPLIFY THE TRACKING ALGO RITHM WHICH CAN ASSUME THAT THE TIMING ERROR IS SMALL 4HE COARSE SYNCHRONIZATION IS DONE BY CORRELATING THE RECEIVED SIGNAL TO A COPY OF THE TRANSMITTED SYNCHRONIZATION

SIGNAL 4O ND THE PEAK OF THIS CORRELATION WITH ENOUGH ACCURACY A DIGITAL LTER IS USED TO PROVIDE INTERPOLATED DATA VALUES AT FOUR TIMES THE ORIGINAL DATA RATE )N THE LAST PHASE OF THE SYNCHRONIZATION NE SYNCHRONIZATION THE SUBCHANNELS WITH PILOTS ARE EQUALIZED WITH THE ESTIMATED CHANNEL OBTAINED FROM PILOTS 3INCE THE COARSE SYNCHRONIZATION GUAR ANTEES THAT THE TIMING ERROR IS LESS THAN f THE CHANNEL IMPULSE RESPONSE IS WITHIN THE CYCLIC PREX 4HE REMAINING PHASE ERRORS ON THE PILOT SUBCHANNELS ARE DUE TO TIMING ERROR AND CAN BE ESTIMATED BY LINEAR REGRESSION 4HERE ARE ALSO SYNCHRONIZATION ALGORITHMS BASED ON THE CYCLIC PREX )N ;= THE DIdERENCE BETWEEN RECEIVED SAMPLES SPACED - SAMPLES APART IS FORMED Q J ` Q J
- 7HEN ONE OF THE SAMPLES BELONGS TO THE CYCLIC PREX AND THE OTHER ONE TO THE /&$SYMBOL FROM WHICH IT IS COPIED THE DIdERENCE SHOULD BE SMALL /THERWISE THE DIdERENCE BETWEEN TWO UNCORRELATED RANDOM VARIABLES WILL HAVE TWICE THE POWER AND HENCE ON AVERAGE WILL BE LARGER "Y WINDOWING THIS DIdERENCE WITH A RECTANGULAR WINDOW OF THE SAME LENGTH AS THE CYCLIC PREX THE OUTPUT SIGNAL HAS A MINIMUM WHEN A NEW /&$SYMBOL STARTS 4HIS IDEA IS MORE FORMALLY ELABORATED IN ; = 4HE LIKELIHOOD FUNCTION GIVEN THE OBSERVED SIGNAL Q J WITH A TIMING AND FREQUENCY ERROR IS DERIVED IN ; = 4HIS FUNCTION IS MAXIMIZED TO SIMULTANEOUSLY OBTAIN ESTIMATES OF BOTH TIMING AND FREQUENCY OdSETS 7ITH NO FREQUENCY OdSET THE LIKELIHOOD FUNCTION WITH RESPECT TO A TIMING OdSET t IS
t +
c t
8
J (t
2-1 1D FQ J Q J
- G ` JQ J ` Q J
- J 2-1
2-1

&OR MEDIUM AND HIGH 3.2S 2-1 A MAXIMUM LIKELIHOOD -, ESTIMATOR BASED ON c t ESSENTIALLY APPLIES A MOVING AVERAGE TO THE TERM JQ J ` Q J
- J IE THE SAME AS THE ESTIMATOR IN ;= (OWEVER FOR SMALL 3.2 VALUES THE CROSSCORRELATION Q J Q J
ALSO HAS TO BE TAKEN INTO ACCOUNT ! SIMILAR PROCEDURE IS USED IN ;= WITH THE DIdERENCE THAT THE INPHASE AND QUADRATURE PARTS OF THE OBSERVED SIGNAL Q J ARE QUANTIZED TO BIT BEFORE t IS ESTIMATED 4HIS YIELDS A SYMBOL SYNCHRONIZER WITH A LOW COMPLEXITY THAT CAN BE USED IN AN ACQUISITION MODE 3YNCHRONIZATION IN THE UPLINK IS MORE DIbCULT THAN IN THE DOWNLINK OR IN BROADCASTING 4HIS IS DUE TO THE FACT THAT THERE WILL BE A SEPARATE OdSET FOR EACH USER 4HIS PROBLEM HAS NOT YET BEEN GIVEN MUCH ATTENTION IN THE LITERATURE (OWEVER A RANDOM ACCESS SEQUENCE IS USED TO SYNCHRONIZE THE MOBILE AND THE BASE STATION IN ;= )NTERFERENCE DUE TO NON SYNCHRONIZED TRANSMISSION HAS BEEN INVESTIGATED IN ;= #ARRIER PHASE NOISE
#ARRIER PHASE NOISE IS CAUSED BY IMPERFECTIONS IN THE TRANSMITTER AND RECEIVER OSCILLATORS &OR A FREQUENCY SELECTIVE CHANNEL NO DISTINCTION CAN BE MADE BETWEEN THE PHASE ROTATION INTRODUCED BY A TIMING ERROR AND A CARRIER PHASE OdSET ;= !N ANALYSIS OF THE IMPACT OF CARRIER PHASE NOISEh IS DONE IN ;= 4HERE i IT IS MODELLED AS A 7IENER PROCESS t S WITH $ Ft S G AND $ t S
S ` t S {n JSJ WHERE n IN (Z DENOTES THE ONE SIDED D" LINEWIDTH OF THE ,ORENTZIAN POWER DENSITY SPECTRUM OF THE FREE RUNNING CARRIER

GENERATOR 4HE DEGRADATION IN 3.2 IE THE INCREASE IN 3.2 NEEDED TO COMPENSATE FOR THE ERROR CAN BE APPROXIMATED BY t u n $ # D" { { KM 6 -
R
WHERE 6 IS THE BANDWIDTH AND $ - IS THE PER SYMBOL 3.2 .OTE THAT THE DEGRADATION INCREASES WITH THE NUMBER OF SUBCARRIERS $UE TO THE RAPID VARIATIONS OF THE PHASE NOISE IT MAY CAUSE LARGE PROBLEMS !NALYSIS OF THE IMPACT OF PHASE NOISE IN CODED SYSTEMS HAS BEEN DONE IN ;=
R

3AMPLING FREQUENCY SYNCHRONIZATION
4HE RECEIVED CONTINUOUS TIME SIGNAL IS SAMPLED AT INSTANTS DETERMINED BY THE RECEIVER CLOCK 4HERE ARE TWO TYPES OF METHODS OF DEALING WITH THE MISMATCH IN SAMPLING FREQUENCY )N SYNCHRONIZED SAMPLING SYSTEMS A TIMING ALGORITHM CONTROLS A VOLTAGE CONTROLLED CRYSTAL OSCILLATOR IN ORDER TO ALIGN THE RECEIVER CLOCK WITH THE TRANSMITTER CLOCK 4HE OTHER METHOD IS NON SYNCHRONIZED SAMPLING WHERE THE SAMPLING RATE REMAINS XED WHICH REQUIRES POST PROCESSING IN THE DIGITAL DOMAIN 4HE EdECT OF A CLOCK FREQUENCY OdSET IS TWOFOLD THE USEFUL SIGNAL COMPONENT IS ROTATED AND ATTENUATED AND IN ADDITION )#) IS INTRODUCED )N ;= THE BIT ERROR RATE PERFORMANCE OF A NON SYNCHRONIZED SAMPLED /&$SYSTEM HAS BEEN INVESTIGATED )T IS SHOWN THAT NON SYNCHRONIZED SAMPLING SYSTEMS ARE MUCH MORE SENSITIVE TO A FREQUENCY OdSET COMPARED WITH A SYNCHRONIZED SAMPLING SYS TEM &OR NON SYNCHRONIZED SAMPLING SYSTEMS IT WAS SHOWN THAT THE DEGRADATION IN D" DUE TO A FREQUENCY SAMPLING OdSET DEPENDS ON THE SQUARE OF THE CARRIER INDEX AND ON THE SQUARE OF THE RELATIVE FREQUENCY OdSET %RRORS IN THE SAMPLING FREQUENCY FOR $-4 SYSTEMS HAVE BEEN ANALYZED IN ;=

#ARRIER FREQUENCY SYNCHRONIZATION

&REQUENCY ERRORS
&REQUENCY OdSETS ARE CREATED BY DIdERENCES IN OSCILLATORS IN TRANSMITTER AND RECEIVER $OPPLER SHIFTS OR PHASE NOISE INTRODUCED BY NON LINEAR CHANNELS 4HERE ARE TWO DESTRUC TIVE EdECTS CAUSED BY A CARRIER FREQUENCY OdSET IN /&$- SYSTEMS /NE IS THE REDUCTION OF SIGNAL AMPLITUDE THE SINC FUNCTIONS ARE SHIFTED AND NO LONGER SAMPLED AT THE PEAK AND THE OTHER IS THE INTRODUCTION OF )#) FROM THE OTHER CARRIERS SEE &IGURE 4HE LATTER IS CAUSED BY THE LOSS OF ORTHOGONALITY BETWEEN THE SUBCHANNELS )N ;= 0OLLET ET AL ANALYTICALLY EVALUATE THE DEGRADATION OF THE "%2 CAUSED BY THE PRESENCE OF CARRIER FREQUENCY OdSET AND CARRIER PHASE NOISE FOR AN !7'. CHANNEL )T IS FOUND THAT A MUL TICARRIER SYSTEM IS MUCH MORE SENSITIVE THAN A SINGLE CARRIER SYSTEM $ENOTE THE RELATIVE FREQUENCY OdSET NORMALIZED BY THE SUBCARRIER SPACING BY aE WHERE a% IS THE FREQUENCY OdSET AND - THE NUMBER OF SUBCARRIERS 4HE DEGRADATION # IN 3.2 IN D" CAN THEN BE APPROXIMATED BY t u - a a% $ $ # D" { { aE { KM - KM 6 -
% 6R R

&IGURE %dECTS OF A FREQUENCY OdSET a% REDUCTION IN SIGNAL AMPLITUDE p AND INTERCARRIER INTERFERENCE q .OTE THAT THE DEGRADATION IN D" INCREASES WITH THE SQUARE OF THE NUMBER OF SUBCARRIERS IF a% AND 6 ARE XED )N ;= -OOSE DERIVES THE SIGNAL TO INTERFERENCE RATIO 2(1 ON A FADING AND DISPER SIVE CHANNEL 4HE 2(1 IS DENED AS THE RATIO OF THE POWER OF THE USEFUL SIGNAL TO THE POWER OF THE INTERFERENCE SIGNAL )#) AND h ADDITIVE i NOISE (E ASSUMED THAT ALL CHANNEL ATTENUATIONS G HAVE THE SAME POWER $ JG J !N UPPER BOUND ON THE DEGRADATION IS
J J
# D" v KNF

R RHM {aE
$ -
RHMB aE

WHERE RHMB W RHM {W {W 4HE FACTOR IS FOUND FROM A LOWER BOUND OF THE SUMMATION OF ALL INTERFERING SUBCARRIERS )N &IGURE THE DEGRADATION IS PLOTTED AS A FUNCTION OF THE NORMALIZED FREQUENCY OdSET aE IE RELATIVE TO THE SUBCARRIER SPAC ING 4HE SYNCHRONIZATION REQUIREMENTS FOR AN /&$- SYSTEM HAVE BEEN INVESTIGATED IN ;= 4HE CONCLUSION THEREIN IS THAT IN ORDER TO AVOID SEVERE DEGRADATION THE FREQUENCY SYNCHRONIZATION ACCURACY SHOULD BE BETTER THAN &REQUENCY ESTIMATORS
3EVERAL CARRIER SYNCHRONIZATION SCHEMES HAVE BEEN SUGGESTED IN THE LITERATURE !S WITH SYMBOL SYNCHRONIZATION THEY CAN BE DIVIDED INTO TWO CATEGORIES BASED ON PILOTS OR ON THE CYCLIC PREX "ELOW FOLLOWS A SHORT OVERVIEW OF SOME OF THEM 0ILOT AIDED ALGORITHMS HAVE BEEN ADDRESSED IN ;= )N THAT WORK SOME SUBCARRIERS ARE USED FOR THE TRANSMISSION OF PILOTS USUALLY A PSEUDO NOISE 0. SEQUENCE 5SING THESE KNOWN SYMBOLS THE PHASE ROTATIONS CAUSED BY THE FREQUENCY OdSET CAN BE ESTIMATED 5NDER THE ASSUMPTION THAT THE FREQUENCY OdSET IS LESS THAN HALF THE SUBCARRIER SPACING THERE IS A ONE TO ONE CORRESPONDENCE BETWEEN THE PHASE ROTATIONS AND THE FREQUENCY OdSET

&IGURE $EGRADATION IN 3.2 DUE TO A FREQUENCY OdSET NORMALIZED TO THE SUBCARRIER SPACING !NALYTICAL EXPRESSION FOR !7'. DASHED AND FADING CHANNELS SOLID
4O ASSURE THIS AN ACQUISITION ALGORITHM MUST BE APPLIED )N ;= SUCH AN ALGORITHM IS B CONSTRUCTED BY FORMING A FUNCTION WHICH IS SINC SHAPED AND HAS A PEAK FOR E ` E )T WAS FOUND THAT BY EVALUATING THIS FUNCTION IN POINTS 3 APART AN ACQUISITION COULD BE OBTAINED BY MAXIMIZING THAT FUNCTION 4HIS ACQUISITION SCHEME WAS CONRMED BY COMPUTER SIMULATIONS TO WORK WELL BOTH FOR AN !7'. CHANNEL AND A FADING CHANNEL ! RELATED TECHNIQUE IS TO USE THE CYCLIC PREX WHICH TO SOME EXTENT CAN BE VIEWED AS PILOTS 4HE REDUNDANCY OF THE CYCLIC PREX CAN BE USED IN SEVERAL WAYS EG BY CREATING A FUNCTION THAT PEAKS AT ZERO OdSET AND NDING ITS MAXIMIZING VALUE ; = OR BY DOING MAXIMUM LIKELIHOOD ESTIMATION ; = )N ;= IT IS ASSUMED THAT THE CYCLIC PREX HAS THE SAME SIZE AS THE /&$- SYMBOL IE THE USEFUL SYMBOL IS TRANSMITTED TWICE IN ;= AVERAGING IS PERFORMED TO REMOVE THE DATA DEPENDENCE AND IN ;= DECISION DIRECTION IS USED )N ;= THE LIKELIHOOD FUNCTION FOR BOTH TIMING AND FREQUENCY OdSETS IS DERIVED BY ASSUMING A NON DISPERSIVE CHANNEL AND BY CONSIDERING THE TRANSMITTED DATA SYMBOLS W UNCORRELATED "Y MAXIMIZING THIS FUNCTION A SIMULTANEOUS ESTIMATION OF THE TIMING AND FREQUENCY OdSETS CAN BE OBTAINED )F THE FREQUENCY ERROR IS SLOWLY VARYING COMPARED THE /&$- SYMBOL RATE A PHASE LOCKED LOOP 0,, ;= CAN BE USED TO REDUCE THE ERROR FURTHER
J
)T IS INTERESTING TO NOTE THE RELATIONSHIP BETWEEN TIME AND FREQUENCY SYNCHRONIZATION )F THE FREQUENCY SYNCHRONIZATION IS A PROBLEM IT CAN BE REDUCED BY LOWERING THE NUMBER OF SUBCARRIERS WHICH WILL INCREASE THE SUBCARRIER SPACING 4HIS WILL HOWEVER INCREASE THE DEMANDS ON THE TIME SYNCHRONIZATION SINCE THE SYMBOL LENGTH GETS SHORTER IE A LARGER RELATIVE TIMING ERROR WILL OCCUR 4HUS THE SYNCHRONIZATIONS IN TIME AND FREQUENCY ARE CLOSELY RELATED TO EACH OTHER


#HANNEL ESTIMATION
-ODULATION CAN BE CLASSIED AS DIdERENTIAL OR COHERENT 7HEN USING DIdERENTIAL MODU LATION THERE IS NO NEED FOR A CHANNEL ESTIMATE SINCE THE INFORMATION IS ENCODED IN THE DIdERENCE BETWEEN TWO CONSECUTIVE SYMBOLS 4HIS IS A COMMON TECHNIQUE IN WIRELESS SYS TEMS WHICH SINCE NO CHANNEL ESTIMATOR IS NEEDED REDUCES THE COMPLEXITY OF THE RECEIVER $IdERENTIAL MODULATION IS USED IN THE %UROPEAN $!" STANDARD ;= 4HE DRAWBACKS ARE ABOUT A D" NOISE ENHANCEMENT ;= AND AN INABILITY TO USE EbCIENT MULTIAMPLITUDE CONSTELLATIONS (OWEVER DIdERENTIAL SCHEMES CAN BENET FROM ASSISTANCE BY A CHANNEL ES TIMATOR ;= !N INTERESTING ALTERNATIVE TO COHERENT MODULATION IS DIdERENTIAL AMPLITUDE AND PHASE SHIFT KEYING $!03+ ; = WHERE A SPECTRAL EbCIENCY GREATER THAN THAT OF $03+ IS ACHIEVED BY USING A DIdERENTIAL CODING OF AMPLITUDE AS WELL 4HIS REQUIRES A NONUNIFORM AMPLITUDE DISTRIBUTION #OHERENT MODULATION HOWEVER ALLOWS ARBITRARY SIGNAL CONSTELLATIONS AND IS AN OBVIOUS CHOICE IN WIRED SYSTEMS WHERE THE CHAN NEL HARDLY CHANGES WITH TIME )N WIRELESS SYSTEMS THE EbCIENCY OF COHERENT MODULATION MAKES IT AN INTERESTING CHOICE WHEN THE BIT RATE IS HIGH AS IN DIGITAL VIDEO BROADCAST $6" ; = 4HE CHANNEL ESTIMATION IN WIRED SYSTEMS IS FAIRLY STRAIGHTFORWARD AND IS NOT DISCUSSED IN DETAIL BELOW 7E CONCENTRATE ON CHANNEL ESTIMATION IN WIRELESS SYSTEMS WHERE THE COMPLEXITY OF THE ESTIMATOR IS AN IMPORTANT DESIGN CRITERION 4HERE ARE MAINLY TWO PROBLEMS IN THE DESIGN OF CHANNEL ESTIMATORS FOR WIRELESS /&$SYSTEMS 4HE RST PROBLEM CONCERNS THE CHOICE OF HOW PILOT INFORMATION DATASIGNALS KNOWN AT THE RECEIVER SHOULD BE TRANSMITTED 4HIS PILOT INFORMATION IS NEEDED AS A REFERENCE FOR CHANNEL ESTIMATION 4HE SECOND PROBLEM IS THE DESIGN OF AN ESTIMATOR WITH BOTH LOW COMPLEXITY AND GOOD CHANNEL TRACKING ABILITY 4HESE TWO PROBLEMS ARE INTERCONNECTED SINCE THE PERFORMANCE OF THE ESTIMATOR DEPENDS ON HOW PILOT INFORMATION IS TRANSMITTED

0ILOT INFORMATION
#HANNEL ESTIMATORS USUALLY NEED SOME KIND OF PILOT INFORMATION AS A POINT OF REFERENCE ! FADING CHANNEL REQUIRES CONSTANT TRACKING SO PILOT INFORMATION HAS TO BE TRANSMITTED MORE OR LESS CONTINUOUSLY $ECISION DIRECTED CHANNEL ESTIMATION CAN ALSO BE USED ;= BUT EVEN IN THESE TYPES OF SCHEMES PILOT INFORMATION HAS TO BE TRANSMITTED REGULARLY TO MITIGATE ERROR PROPAGATION 4O THE AUTHORS KNOWLEDGE THERE IS VERY LITTLE PUBLISHED ON HOW TO TRANSMIT PILOT INFORMATION IN WIRELESS /&$- (OWEVER AN EbCIENT WAY OF ALLOWING A CONTINUOUSLY UPDATED CHANNEL ESTIMATE IS TO TRANSMIT PILOT SYMBOLS INSTEAD OF DATA AT CERTAIN LOCATIONS OF THE /&$- TIME FREQUENCY LATTICE 4HIS CAN BE VIEWED AS A GENERALIZATION OF PILOT SYMBOL ASSISTED MODULATION 03!- IN THE SINGLE CARRIER CASE 03!- IN THE SINGLE CARRIER CASE WAS INTRODUCED IN ;= AND THOROUGHLY ANALYZED IN ;= !N EXAMPLE OF THIS IS SHOWN IN &IGURE WHERE BOTH SCATTERED AND CONTINUAL PILOT SYMBOLS ARE SHOWN )N A PRELIMINARY DRAFT OF THE %UROPEAN $6" STANDARD ;= PILOT INFORMATION IS SPECIED TO BE TRANSMITTED ON BOOSTED SUBCARRIERS BOTH SCATTERED AND AS CONTINUAL PILOT CARRIERS "OOSTED SUBCARRIERS MEANS THAT PILOT INFORMATION IS TRANSMITTED AT HIGHER POWER THAN
)NTRODUCTION TO /&$THE DATA

&IGURE !N EXAMPLE OF PILOT INFORMATION TRANSMITTED BOTH SCATTERED AND CONTINUAL ON CERTAIN SUBCARRIERS )N GENERAL THE FADING CHANNEL CAN BE VIEWED AS A $ SIGNAL TIME AND FREQUENCY WHICH IS SAMPLED AT PILOT POSITIONS AND THE CHANNEL ATTENUATIONS BETWEEN PILOTS ARE ESTIMATED BY INTERPOLATION 4HIS ENABLES US TO USE THE $ SAMPLING THEOREM TO PUT LIMITS THE DENSITY OF THE PILOT PATTERN ;= (OWEVER AS IN THE SINGLE CARRIER CASE ;= THE PILOT PATTERN SHOULD BE DESIGNED SO THAT THE CHANNEL IS OVERSAMPLED AT THE RECEIVER

%STIMATOR DESIGN
!SSUMING THAT THE PILOT PATTERN IS CHOSEN THE OPTIMAL LINEAR CHANNEL ESTIMATOR IN TERMS OF MEAN SQUARED ERROR -3% IS A $ 7IENER LTER +NOWING THE STATISTICAL PROPERTIES OF THE CHANNEL SUCH AN ESTIMATOR CAN BE DESIGNED USING STANDARD TECHNIQUES ;= 4HE COMBINATION OF HIGH DATA RATES AND LOW BIT ERROR RATES NECESSITATES THE USE OF ESTIMATORS THAT HAVE BOTH LOW COMPLEXITY AND HIGH ACCURACY 4HESE TWO CONSTRAINTS ON THE ESTI MATORS WORK AGAINST EACH OTHER -OST ESTIMATORS WITH HIGH ACCURACY SUCH AS THE $ 7IENER LTER HAVE A LARGE COMPUTATIONAL COMPLEXITY WHILE ESTIMATORS OF LOWER COMPLEX ITY USUALLY PRODUCE A LESS ACCURATE ESTIMATE 4HE ART IN DESIGNING CHANNEL ESTIMATORS IS NDING A GOOD TRADE Od BETWEEN COMPLEXITY AND PERFORMANCE 4HE ISSUE OF REDUCING THE COMPUTATIONAL COMPLEXITY WHILE MAINTAINING MOST OF THE PERFORMANCE HAS BEEN ADDRESSED IN SEVERAL PUBLICATIONS )N ;= SEPARABLE LTERS ARE APPLIED INSTEAD OF A $ NITE IMPULSE RESPONSE &)2 LTER 4HE USE OF SEPARABLE LTERS

INSTEAD OF FULL $ LTERS IS A STANDARD TECHNIQUE USED TO REDUCE COMPUTATIONAL COMPLEXITY IN MULTIDIMENSIONAL SIGNAL PROCESSING ;= 5SING THIS TECHNIQUE THE ESTIMATION IS RST PERFORMED IN THE FREQUENCY DIRECTION USING A $ &)2 LTER AND THEN IN THE TIME DIRECTION USING A SECOND $ &)2 LTER 4HIS RESTRICTS THE OBTAINABLE $ IMPULSE RESPONSES TO THOSE THAT ARE THE OUTER PRODUCT OF TWO $ LTERS 4HIS RESULTS IN A SMALL PERFORMANCE LOSS BUT THE GREATLY REDUCED COMPLEXITY USUALLY MOTIVATES THE USE OF SEPARABLE LTERS ; = ! SECOND APPROACH IN THE REDUCTION OF COMPUTATIONAL COMPLEXITY IS BASED ON USING TRANSFORMS THAT CONCENTRATE THE CHANNEL POWER TO A FEW TRANSFORM COEbCIENTS THUS AL LOWING EbCIENT CHANNEL ESTIMATION TO BE PERFORMED WITH LITTLE EdORT IN THE TRANSFORM DOMAIN ,OW COMPLEXITY ESTIMATORS OF THIS TYPE BASED ON BOTH THE $&4 ; = AND ON OPTIMAL RANK REDUCTION ; = HAVE BEEN PROPOSED 4HIS TECHNIQUE USUALLY YIELD ESTI MATORS OF HIGH PERFORMANCE AND LOW COMPLEXITY BUT MAY RESULT IN AN IRREDUCIBLE ERROR OOR UNLESS SPECIAL CARE IS TAKEN 4HESE EdECTS ARE ANALYZED IN DETAIL IN ; = ! COMPARATIVE ANALYSIS OF PILOT BASED ESTIMATORS PRESENTED IN ;= SHOWS THAT THE COM BINATION OF SEPARABLE LTERS AND LOW RANK APPROXIMATIONS CAN GIVE HIGH PERFORMANCE ES TIMATORS OF LOW COMPLEXITY WHERE THE GREATEST PORTION OF THE REDUCED COMPLEXITY STEMS FROM THE USE OF SEPARABLE LTERS

0ERFORMANCE EXAMPLE
&IGURE SHOWS AN EXAMPLE ON THE DIdERENCE IN CODED BIT ERROR RATE BETWEEN COHERENT AND DIdERENTIAL MODULATION 4HE SIMULATIONS ARE PERFORMED FOR A WIRELESS SUBCARRIER /&$- SYSTEM WITH A SAMPLE CYCLIC PREX 4HE CHANNEL IS 2AYLEIGH FADING WITH RELATIVE $OPPLER FREQUENCY AND THE SYSTEM USES TRELLIS CODED MODULATION WITH PHASE SHIFT KEYING 03+ ACCORDING TO ;= 4HE DIdERENTIAL MODULATION IS PERFORMED IN THE FREQUENCY DIRECTION SINCE THE FREQUENCY CORRELATION IS GREATER THAN THE TIME CORRELATION ;= 4HE PERFORMANCE OF COHERENT MODULATION IS PRESENTED FOR BOTH KNOWN CHANNEL AND WITH A LOW COMPLEXITY ESTIMATOR ;= 4HE LOW COMPLEXITY CHANNEL ESTIMATOR USES A PILOT SYMBOL DENSITY AND REQUIRES ONLY MULTIPLICATIONS PER ESTIMATED CHANNEL ATTENUATION 4HIS GURE ILLUSTRATES THAT COHERENT MODULATION WITH LOW COMPLEXITY CHANNEL ESTIMA TION CAN OUTPERFORM DIdERENTIAL MODULATION %VEN THOUGH THE CHANNEL CORRELATION IN THE FREQUENCY DIRECTION IS LARGE IN THIS CASE THE BEGINNING OF AN ERROR OOR FOR THE DIdERENTIAL MODULATION IS CLEARLY VISIBLE FOR $ - GREATER THAN ABOUT D" 4HIS ERROR OOR IS OF THE SAME TYPE AS THE ONE EXPERIENCED IN THE SINGLE CARRIER CASE WHEN THE CHANNEL IS FADING ;=
A

#HANNEL CODING
4HIS CHAPTER DESCRIBES CODING IN /&$- SYSTEMS 4HE CODING PROBLEM IS QUITE DIdERENT FOR THE WIRELESS AND THE WIRED CASE )N THE LATTER CASE THE CHANNEL IS STATIC AND TECHNIQUES LIKE BIT LOADING AND MULTIDIMENSIONAL CODING ARE APPROPRIATE /N A FADING CHANNEL THE MAIN DIdERENCE BETWEEN AN /&$- SYSTEM AND A SINGLE CARRIER SYSTEM IS THE INTERLEAVING 7E WILL GIVE A SHORT OVERVIEW OF CODING ON BOTH WIRELESS AND THE WIRED CHANNELS

&IGURE !N EXAMPLE ON THE DIdERENCE BETWEEN COHERENT AND DIdERENTIAL 03+ IN A 2AYLEIGH FADING ENVIRONMENT

7IRELESS SYSTEMS
5SING A TIME DOMAIN EQUALIZER IT IS POSSIBLE TO OBTAIN AN , BRANCH DIVERSITY IF THE CHANNEL CONSISTS OF , RESOLVABLE PATHS ;= )N /&$- THE EQUALIZER DOES NOT GIVE YOU ANY DIVERSITY SINCE ALL SUBCHANNELS ARE NARROWBAND AND EXPERIENCE AT FADING ;= (OWEVER THE STRUCTURE OF /&$- OdERS THE OPPORTUNITY TO CODE ACROSS THE SUBCARRIERS )N ;= IT IS SHOWN THAT WITH AN , PATH CHANNEL IT IS POSSIBLE TO OBTAIN AN , BRANCH DIVERSITY THROUGH CODING (ENCE IN THIS DIVERSITY CONTEXT A MULTI CARRIER SYSTEM IS COMPARABLE TO A SINGLE CARRIER SYSTEM "ESIDES COMBATING FADING CODING HAS ALSO BEEN PROPOSED TO DEAL WITH LONG ECHOES THAT CAUSES )3) BETWEEN SUBSEQUENT /&$- SYMBOLS ;= 4HE DESIGN OF CODES FOR /&$- SYSTEMS ON FADING CHANNELS FOLLOWS MANY OF THE STAN DARD TECHNIQUES 7HAT IS SPECIAL ABOUT /&$- IS THE TIME FREQUENCY LATTICE AND THE POSSIBILITY TO USE TWO DIMENSIONS FOR INTERLEAVING AND CODING 4O ILLUSTRATE HOW TO USE THIS STRUCTURE WE GIVE AN OVERVIEW OF TWO SYSTEMS THE %UROPEAN $!" STANDARD AND A TRELLIS CODED SYSTEM BY (HER 4HE $!" SYSTEM USES DIdERENTIAL MODULATION WHICH AVOIDS CHANNEL ESTIMATION WHILE THE OTHER SYSTEM USES A MULTIAMPLITUDE SIGNAL CONSTEL LATION WHICH REQUIRES CHANNEL ESTIMATION $IGITAL !UDIO "ROADCASTING
$IGITAL BROADCASTING TO MOBILE RECEIVERS WAS UNDER CONSIDERABLE INVESTIGATION IN THE LATE S ; = 4HE STANDARD FOR $!" WAS SET IN %UROPE TO USE /&$- ;= WHILE IT IS STILL UNDER INVESTIGATION IN THE 53! ;= 4HE %UROPEAN $!" SYSTEM USES DIdERENTIAL QUADRATURE PHASE SHIFT KEYING $103+ TO AVOID CHANNEL ESTIMATION !N ARGUMENT FOR THIS IS BASED ON SIMPLE AND INEXPENSIVE RECEIVERS FOR CONSUMERS 4HE CHANNEL ENCODING PROCESS IS BASED ON PUNCTURED CONVOLUTIONAL CODING WHICH ALLOWS BOTH EQUAL AND UNEQUAL ERROR PROTECTION ;= !S A MOTHER CODE A RATE CONVOLUTIONAL CODE WITH CONSTRAINT LENGTH AND OCTAL POLYNOMIALS IS USED 4HE PUNCTURING PROCEDURE

ALLOWS THE EdECTIVE CODE RATE TO VARY BETWEEN AND )NTERLEAVING IS PERFORMED IN BOTH TIME AND FREQUENCY 4HE FORMER IS A KIND OF BLOCK IN TERLEAVING AFTER WHICH THE BITS ARE MAPPED TO 103+ SYMBOLS 4HE FREQUENCY INTERLEAVER WORKING WITH THE 103+ SYMBOLS FOLLOWS A PERMUTATION RULE OF THE SUBCARRIERS )N ONE OF THE THREE TRANSMISSIONS MODES THIS PERMUTATION RULE IS e e M e M `
LNC M
4HIS PERMUTATION DENES THE SET Fe e e e G F G ACCORDING TO WHICH THE INTERLEAVING PATTERN IS CHOSEN !FTER THE FREQUENCY INTERLEAVING THE 103+ SYMBOLS ARE DIdERENTIALLY MODULATED ON EACH SUBCARRIER 4RELLIS CODED /&$-
!S AN EXAMPLE OF A MULTIAMPLITUDE COHERENT AND CODED /&$- SYSTEM WE GIVE A SHORT OVERVIEW OF A SYSTEM INVESTIGATED BY (HER 4HE ORIGINAL INVESTIGATION ;= ADDRESSES A DIGITAL AUDIO BROADCASTING SCENARIO BUT THE CONCEPT IS MORE GENERAL (HERS INVESTI GATION IS THEREFORE ONE OF THE MOST REFERENCED PAPERS ON /&$- (IS SYSTEM IS A POWER AND BANDWIDTH EbCIENT CONCATENATED CODING SYSTEM FOR DATA TRANSMISSION ON TIME AND FREQUENCY SELECTIVE MOBILE FADING CHANNELS ! CONCATENATED CODING IS USED IN CONJUNCTION WITH DOUBLE INTERLEAVING AND SLOW FREQUENCY HOPPING TO PROVIDE DIVERSITY !N OVERVIEW OF THE SYSTEM IS DEPICTED IN &IGURE 4HE OUTER CODES ARE RATE COMPATIBLE PUNC
&IGURE /VERVIEW OF THE SYSTEM INVESTIGATED BY (HER ;= TURED CODES 2#0# DERIVED FROM THE RATE CODE WITH CONSTRAINT LENGTH 4HE OUTER INTERLEAVING SCHEME IS APPLIED TO BREAK UP ERROR BURSTS FROM THE INNER CODING SYSTEM 3INCE THESE BURSTS ARE TYPICALLY MUCH SHORTER THAN THE FADING BURSTS THE OUTER INTERLEAVING CAN BE MUCH SIMPLER THAN THE INNER INTERLEAVING 4HE INNER CODE IS BINARY TRELLIS CODED MODULATION 4#- WITH ONE DIMENSIONAL SIGNAL CONSTELLATION 4HE REASON FOR THIS CHOICE IS THAT THEY WERE FOUND TO PROVIDE A GOOD DIVER SITY FACTOR AT A VERY LOW DECODER COMPLEXITY ;= 4HE CODE USED IS A ONE DIMENSIONAL

STATE CODE WITH LEVEL UNIFORM PULSE AMPLITUDE MODULATION 0!- 4HE 0!OUTPUT SYMBOLS ARE THEN COMBINED TO A POINT QUADRATURE AMPLITUDE MODULATION 1!- CONSTELLATION AND INTERLEAVED TO BREAK UP CHANNEL MEMORY )N THE RECEIVER THE 6ITERBI ALGORITHM 6! IS USED FOR DECODING 4HIS ALGORITHM IS CAPABLE OF USING THE CHAN NEL STATE INFORMATION OBTAINED FROM A PILOT SEQUENCE SEE 3ECTION 4HE DECODING IS PERFORMED BY MINIMIZING THE METRIC n 8n nB n ) nG n JX ` W J
M M M M
WHERE B G IS THE CHANNEL ESTIMATE X IS THE DEINTERLEAVED OBSERVATION AFTER EQUALIZATION THE REAL OR IMAGINARY PART AND W IS A POTENTIAL CODEWORD "ECAUSE OF THE OUTER CODE THE 6! SHOULD BE MODIED TO PROVIDE RELIABILITY ESTIMATES TOGETHER WITH THE DECODED SEQUENCE 4HIS ENABLES SOFT DECODING OF THE OUTER CODE AS WELL "Y APPLYING A SOFT OUTPUT 6ITERBI ALGORITHM 3/6! AN IMPROVEMENT OF ABOUT D" IS OBTAINED ;= 4OGETHER WITH INNER CODING MULTICARRIER SIGNALLING AND SLOW FREQUENCY HOPPING THE INTERLEAVER PROVIDES DUAL TIMEFREQUENCY DIVERSITY !N EXAMPLE OF SLOW FREQUENCY HOPPING IS DEPICTED IN &IGURE %ACH PROGRAM CONSISTS OF A NUMBER OF SUBCARRIERS AND /&$- SYMBOLS THE
M M M
&IGURE 3LOW FREQUENCY HOPPING %ACH PROGRAM / USES A BANDWIDTH ! AND CHANGES FREQUENCY BAND AFTER 3
H GNO
PARAMETERS ! AND 3 ARE CHOSEN TO MAXIMIZE THE DIVERSITY OF THE SYSTEM ! SIMILAR BUT SOMEWHAT MORE GENERALIZED HOPPING SCHEME IS PRESENTED IN ;=
GNO

/THER SYSTEMS
4HERE HAVE BEEN OTHER CODED /&$- SYSTEMS PROPOSED AND ANALYZED IN THE LITERATURE )N ; = AN /&$-&- SYSTEM IS INVESTIGATED AND SIMULATED /&$- IS PROPOSED IN %UROPE AS THE TRANSMISSION TECHNIQUE FOR THE NEW $6" SYSTEM ;= WHERE A MULTIRESO LUTION SCHEME IS USED TOGETHER WITH JOINT SOURCECHANNEL CODING ; = 4HIS ALLOWS SEVERAL BIT RATES AND THEREBY A GRACEFUL DEGRADATION OF IMAGE QUALITY IN THE FRINGES OF THE BROADCAST AREA
#ODING ON FADING CHANNELS
#ODES FOR FADING CHANNELS HAVE BEEN INVESTIGATED FOR A LONG TIME AND THE SEARCH FOR GOOD CODES IS STILL GOING ON !N OVERVIEW OF THIS SUBJECT IS FOUND IN ;= $ESIGN OF CODES HAS BEEN ANALYZED IN ; = AND FOR TRELLIS CODED MULTIPLE 03+ IN ; = !SYMMETRIC 03+ CONSTELLATIONS ARE CONSIDERED IN ;= AND CODE DESIGN FOR 2ICIAN FADING CHANNELS IS INVESTIGATED IN ;= 0ERFORMANCE BOUNDS HAVE BEEN DERIVED FOR 2AYLEIGH FADING CHANNELS IN ; = AND FOR 2ICIAN FADING CHANNELS IN ; = 5SUALLY PERFORMANCE ANALYSIS OF CODES ASSUMES PERFECT KNOWLEDGE OF THE CHANNEL (OWEVER IN ; = AN ANALYTICAL METHOD WAS INTRODUCED THAT ALLOWS NON IDEAL CHANNEL INFORMATION 4HIS WAS LATER GENERALIZED TO INCLUDE NON IDEAL INTERLEAVING ; = 4HIS METHOD HAS BEEN USED TO ANALYZE A CODED /&$- SYSTEM WITH PILOT BASED CHANNEL ESTI MATION ON 2AYLEIGH FADING CHANNELS ;= ! MAJOR BENET OF USING ANALYTICAL METHODS FOR EVALUATION OF CODED SYSTEMS IS THAT A CODED BIT ERROR RATE CAN BE OBTAINED QUICKLY AFTER SYSTEM MODICATIONS WITHOUT TIME CONSUMING SIMULATIONS

7IRED SYSTEMS
!N IMPORTANT DIdERENCE BETWEEN A WIRED AND A WIRELESS SYSTEM IS THE CHARACTERISTICS OF THE CHANNEL )N THE WIRED CASE THE CHANNEL IS OFTEN CONSIDERED STATIONARY WHICH FACILI TATES A NUMBER OF TECHNIQUES TO IMPROVE THE COMMUNICATION SYSTEM #HANNEL CODING IN COMBINATION WITH A TECHNIQUE CALLED BIT LOADING IS OFTEN EMPLOYED FOR THIS PURPOSE -UL TIDIMENSIONAL TRELLIS CODES ARE WELL SUITED FOR THE CHANNEL CODING 7HEN USING BIT LOADING THE SUBCHANNELS ARE ASSIGNED INDIVIDUAL NUMBERS OF BITS ACCORDING TO THEIR RESPECTIVE 3.2S !N /&$- BASED COMMUNICATION SYSTEM USING BIT LOADING IS OFTEN REFERRED TO AS A $-4 SYSTEM "IT LOADING
"IT LOADING IS A TECHNIQUE THAT IS USED FOR MULTICARRIER SYSTEMS OPERATING ON STATIONARY CHANNELS ;= ! STATIONARY CHANNEL MAKES IT POSSIBLE TO MEASURE THE 3.2 ON EACH SUB CHANNEL AND ASSIGN INDIVIDUAL NUMBERS OF TRANSMITTED BITS ! SUBCHANNEL WITH HIGH 3.2 THUS TRANSMITS MORE BITS THAN A SUBCHANNEL WITH LOW 3.2 &IGURE SHOWS A SCHEMATIC PICTURE OF 3.2 AND HOW THE NUMBER OF BITS ON EACH SUBCHANNEL VARY ACCORDINGLY 7HEN PERFORMING BIT LOADING ONE USUALLY OPTIMIZES FOR EITHER HIGH DATA RATE LOW AVERAGE TRANSMITTING ENERGY OR LOW ERROR PROBABILITY 4YPICALLY TWO OF THESE ARE KEPT CONSTANT AND THE THIRD IS THE GOAL FOR THE OPTIMIZATION 7HICH PARAMETER SHOULD BE OPTIMIZED DEPENDS ON THE SYSTEM ITS ENVIRONMENT AND ITS APPLICATION 7HEN THERE IS ONLY ONE SYSTEM OPERATING ON A CABLE THIS SYSTEM NEITHER INTERFERES WITH NOR IS INTERFERED BY OTHER SYSTEMS 4HIS MEANS THAT CONTROLLING THE TRANSMITTING POWER TO REDUCE CROSSTALK IS NOT NECESSARY 'IVEN A DATA RATE AND A BIT ERROR PROBABILITY WHATEVER TRANSMISSION ENERGY NEEDED TO ACHIEVE THESE GOALS CAN BE USED WITHIN REASONABLE LIMITS )N A MULTI SYSTEM ENVIRONMENT WHERE THERE ARE SEVERAL SYSTEMS TRANSMITTING IN THE SAME CABLE THE PROBLEM IS MORE COMPLICATED SINCE THE SYSTEMS EXPERIENCE CROSSTALK 4HE LEVEL
)NTRODUCTION TO /&$3.2 "ITS

3UBCARRIER
3UBCARRIER
&IGURE #HANNEL 3.2 LEFT AND CORRESPONDING NUMBER OF BITS ON EACH SUBCARRIER RIGHT OF CROSSTALK IS PROPORTIONAL TO THE TRANSMITTING POWER IN THE SYSTEMS SEE AND )T IS THEREFORE DESIRABLE TO HAVE AN EQUAL TRANSMISSION POWER IN ALL SYSTEMS TO OBTAIN EQUAL DISTURBANCE SITUATIONS )N A MULTI SYSTEM ENVIRONMENT THE AVERAGE TRANSMITTING POWER IS USUALLY XED AND THE OPTIMIZATION IS FOR EITHER HIGH DATA RATE OR LOW BIT ERROR RATE "IT LOADING ALGORITHMS
4HERE ARE SEVERAL TECHNIQUES FOR BIT LOADING IN $-4 SYSTEMS AND SOME OF THESE ARE DESCRIBED IN ; = !S MENTIONED EARLIER THERE ARE SEVERAL PARAMETERS THAT ONE CAN OPTIMIZE FOR -OST ALGORITHMS OPTIMIZE FOR HIGH DATA RATE OR LOW BIT ERROR RATE 'IVEN A CERTAIN DATA RATE AND AN ENERGY CONSTRAINT THE (UGHES (ARTOGS ALGORITHM PROVIDES THE BIT LOADING FACTORS THAT YIELD MINIMAL BIT ERROR RATE SEE EG ;= 4HE IDEA BEHIND THE (UGHES (ARTOGS ALGORITHM IS TO ASSIGN ONE BIT AT A TIME TO THE SUBCHANNELS 4HE ALGORITHM CALCULATES THE ENERGY COST TO SEND ONE BIT MORE ON EACH SUBCHANNEL 4HE SUBCHANNEL WITH THE SMALLEST ENERGY COST IS THEN ASSIGNED THE BIT 4HIS PROCEDURE IS REPEATED UNTIL A DESIRED BIT RATE IS OBTAINED #HOW HAS SHOWED THAT COMPLEXITY OF THE (UGHES (ARTOGS ALGORITHM IS PROPORTIONAL TO THE NUMBER OF SUBCHANNELS AND THE NUMBER OF BITS TRANSMITTED IN A $-4 FRAME ;= (E ALSO SUGGESTS A SUBOPTIMAL ALGORITHM OF LOWER COMPLEXITY IN ;= !N ALGORITHM THAT MAINTAINS AN EQUAL BIT ERROR PROBABILITY OVER ALL SUBCHANNELS GIVEN A DATA RATE AND AN ENERGY CONSTRAINT IS PRESENTED BY &ISCHER IN ;= ! SUBOPTIMAL WAY OF PERFORMING BIT LOADING TO ACHIEVE A HIGH DATA RATE WHILE MAIN TAINING A CONSTANT SYMBOL ERROR PROBABILITY ACROSS ALL SUBCHANNELS IS PRESENTED BY 4U ;= )N HIS ALGORITHM THE BIT LOADING FACTORS ARE CALCULATED ACCORDING TO t A KNF
J J
u $ F o
` KNF " *}
J J C J J

WHERE A IS THE NUMBER OF BITS CARRIED BIT LOADING FACTOR ON SUBCARRIER J $ THE AVERAGE SYMBOL TRANSMISSION ENERGY F THE CHANNEL ATTENUATION AND } THE NOISE VARIANCE 4HE CODING GAIN IS DENOTED o AND THE CONSTELLATION EXPANSION FACTOR DUE TO CODING IS DENOTED " &URTHER TO OBTAIN A DESIRED SYMBOL ERROR RATE OF / THE DESIGN CONSTANT * IS CHOSEN
J J C D
TO *
D
)NTRODUCTION TO /&$v 0 t

/ -
uw
D
WHERE - IS THE NUMBER OF NEAREST NEIGHBORS %XPRESSION CAN BE VIEWED AS THE UNION BOUND FOR A 1!- CONSTELLATION WITH SOME MODICATIONS FOR CODING WHERE * IS THE 3.2 REQUIRED TO OBTAIN AN ERROR PROBABILITY / 4HE CHANNEL 3.2 INCLUDING CODING $ F o } IS DIVIDED BY THE 3.2 REQUIRED TO TRANSMIT ONE BIT &INALLY THE NUMBER OF BITS NEEDED IN THE CODING KNF " IS SUBTRACTED TO GET THE NUMBER OF BITS CARRIED BY SUBCHANNEL J )F THE NUMBER OF SYSTEMS TRANSMITTING IN A CABLE VARY THE AMOUNT OF CROSSTALK WILL VARY ACCORDINGLY 4O HANDLE THE SITUATION WHERE THE NUMBER OF TRANSMITTING SYSTEMS VARY ONE CAN EITHER DO THE BIT LOADING FOR A WORST CASE OR EMPLOY ADAPTIVE BIT LOADING #HOW ;= PRESENTS SUCH AN ADAPTIVE ALGORITHM CALLED THE BIT SWAP ALGORITHM WHICH IS DESIGNED FOR THE CASE WHEN A XED DATA RATE IS SPECIED 7HEN TRYING TO MAXIMIZE THE DATA RATE WITH A CONSTANT TRANSMITTING POWER IT IS OPTIMAL TO ALLOW BIT LOADING FACTORS TO SPAN A CONTINUOUS RANGE OF VALUES 4U PRESENTS SOME RESULTS IN ;= ON HOW THE GRANULARITY OF BIT LOADING FACTORS AdECTS THE OBTAINED DATA RATE IN SUCH A SYSTEM /NE WAY TO GET NON INTEGER BIT LOADING FACTORS IS TO USE MULTIDIMENSIONAL CODES -ULTIDIMENSIONAL CODES ALLOW A FRACTIONAL NUMBER OF BITS PER $ SYMBOL TO BE TRANSMITTED ON EACH SUBCHANNEL &OR $ $ AND $ CODES THE GRANULARITIES BECOME BIT BITS AND BITS PER $ SYMBOL RESPECTIVELY !NOTHER TECHNIQUE REFERRED TO AS ENERGY LOADING IS TO ALLOW SOME SORT OF NE TUNING OF THE TRANSMITTED ENERGY ON THE SUBCHANNELS IE ADJUSTING THE ENERGY $ IN SO THAT IT CORRESPONDS TO ONE OF THE SUPPORTED BIT LOADING FACTORS (OWEVER ENERGY LOADING ONLY WORKS IF THE TUNING IS SMALL WHICH REQUIRES MANY BIT LOADING FACTORS AND A SIDE EdECT IS THAT A MORE COMPLEX SIGNAL CONSTELLATION MAPPERDEMAPPER IS REQUIRED
D J J C J J

#HANNEL CODING
#ODING IN $-4 HAS BEEN ANALYZED IN EG ; = ! TYPICAL CODING SCHEME FOR $-4 CONSISTS OF AN OUTER CODE AN INTERLEAVER AND AN INNER CODE $UE TO THE TYPE OF APPLI CATIONS THAT $-4 IS DESIGNED TO CARRY SEE 3ECTION IT IS APPEALING TO KEEP A LOW DELAY BETWEEN TRANSMITTER AND RECEIVER 4HIS LIMITS THE INTERLEAVING DEPTH AND AdECTS THE CHOICE OF ERROR CORRECTING CODES 4HE CODING SCHEME ANALYZED IN ;= USES AN OUTER 2EED 3OLOMON CODE AND AN INNER TRELLIS CODE 4HE 2EED 3OLOMON CODE AND INTERLEAVING ARE DESIGNED TO REDUCE ERRORS DUE TO IMPULSE NOISE "Y USING ONLY ONE CODER THAT CODES ACROSS SUBCARRIERS THE DELAY IS SMALL COMPARED TO THE CASE WHERE ONE TRELLIS CODE IS USED FOR EACH SUBCARRIER 4HIS IS DUE TO THE 6ITERBI DECODERS NEED FOR A CERTAIN DECISION DEPTH TO MAKE A GOOD DECISION )N THE INVESTIGATED SYSTEM THE AMOUNT OF DATA SENT IN ONE $-4 FRAME IS ENOUGH FOR THE 6ITERBI ALGORITHM TO MAKE A DECISION -ULTIDIMENSIONAL CODES ARE WELL SUITED FOR $-4 SYSTEMS "Y USING SEVERAL $ CON STELLATIONS ON DIdERENT SUBCHANNELS IT IS EASY TO CREATE MULTIDIMENSIONAL CONSTELLATIONS !S DESCRIBED EARLIER THE MULTIDIMENSIONAL CODES ALLOW FRACTIONAL BITS TO BE TRANSMITTED WHICH REDUCES THE GRANULARITY OF THE BIT LOADING FACTORS

$ISCUSSION
4HIS SECTION IS BOTH A DISCUSSION AND A SUMMARY OF THE MATERIAL PRESENTED EARLIER IN THIS REPORT /NE OF THE MAJOR ADVANTAGES OF /&$- IS ITS ROBUSTNESS AGAINST MULTIPATH PROP AGATION (ENCE ITS TYPICAL APPLICATIONS ARE IN TOUGH RADIO ENVIRONMENTS /&$- IS ALSO SUITABLE IN SINGLE FREQUENCY NETWORKS SINCE THE SIGNALS FROM OTHER TRANSMITTERS CAN BE VIEWED AS ECHOES IE MULTIPATH PROPAGATION 4HIS MEANS THAT IT IS FAVORABLE TO USE /&$- IN BROADCASTING APPLICATIONS SUCH AS $!" AND $6" 4HE USE OF /&$- IN MULTIUSER SYSTEMS HAS GAINED AN INCREASING INTEREST THE LAST FEW YEARS 4HE DOWNLINK IN THOSE SYSTEMS IS SIMILAR TO BROADCASTING WHILE THE UPLINK PUTS HIGH DEMANDS ON EG SYNCHRONIZATION 4HE FUTURE OF /&$- AS A TRANSMISSION TECHNIQUE FOR MULTIUSER SYSTEMS DEPENDS ON HOW WELL THESE PROBLEMS CAN BE SOLVED )N WIRED SYSTEMS THE STRUCTURE OF /&$- OdERS THE POSSIBILITY OF EbCIENT BITLOADING "Y ALLOCATING A DIdERENT NUMBER OF BITS TO DIdERENT SUBCHANNELS DEPENDING ON THEIR INDIVIDUAL 3.2S EbCIENT TRANSMISSION CAN BE ACHIEVED !LTHOUGH OTHER SYSTEMS HAVE BEEN PROPOSED /&$- IS THE DOMINATING TECHNIQUE ON EG DIGITAL SUBSCRIBER LINES .OTE THAT /&$- OFTEN GOES UNDER THE NAME $-4 WHEN USED IN WIRED SYSTEMS WITH BITLOADING 4HERE ARE ALSO PROBLEMS ASSOCIATED WITH /&$- SYSTEM DESIGN 4HE TWO MAIN OBSTA CLES WHEN USING /&$- ARE THE HIGH PEAK TO AVERAGE POWER RATIO AND SYNCHRONIZATION 4HE FORMER PUTS HIGH DEMANDS ON LINEARITY IN AMPLIERS 3YNCHRONIZATION ERRORS IN BOTH TIME AND FREQUENCY DESTROY THE ORTHOGONALITY AND CAUSE INTERFERENCE "Y USING A CYCLIC PREX THE TIMING REQUIREMENTS ARE SOMEWHAT RELAXED SO THE BIGGEST PROBLEMS ARE DUE TO HIGH FREQUENCY SYNCHRONIZATION DEMANDS $EGRADATION DUE TO FREQUENCY ERRORS CAN BE CAUSED BOTH BY DIdERENCES IN LOCAL OSCILLATORS AND BY $OPPLER SHIFTS ! GREAT DEAL OF EdORT IS THEREFORE SPENT ON DESIGNING ACCURATE FREQUENCY SYNCHRONIZERS FOR /&$- !S IN ANY DIGITAL COMMUNICATION SYSTEM THERE ARE TWO ALTERNATIVES FOR MODULATION COHERENT OR DIdERENTIAL 4HE %UROPEAN $!" SYSTEM USES DIdERENTIAL 103+ WHILE THE PROPOSED SCHEME FOR $6" IS COHERENT 1!- $IdERENTIAL 03+ IS SUITABLE FOR LOW DATA RATES AND GIVES SIMPLE AND INEXPENSIVE RECEIVERS WHICH IS IMPORTANT FOR PORTABLE CON SUMER PRODUCTS LIKE $!" RECEIVERS (OWEVER IN $6" THE DATA RATE IS MUCH HIGHER AND LOW BIT ERROR RATES ARE DIbCULT TO OBTAIN WITH DIdERENTIAL 03+ ! NATURAL CHOICE FOR $6" IS THEREFORE MULTIAMPLITUDE SCHEMES $UE TO THE STRUCTURE IN /&$- IT IS EASY TO DESIGN EbCIENT CHANNEL ESTIMATORS AND EQUALIZERS 4HIS IS ONE OF THE APPEALING PROPERTIES OF /&$- WHICH SHOULD BE EXPLOITED TO ACHIEVE HIGH SPECTRAL EbCIENCY #ODING IN WIRELESS /&$- SYSTEMS DOES NOT DIdER MUCH FROM CODING IN WIRELESS SINGLE CARRIER SYSTEMS 4HE MAIN DIdERENCE IS THAT INTERLEAVING IN /&$- ALLOWS SYMBOLS TO BE SPREAD IN BOTH TIME AND FREQUENCY 4HE POSSIBILITY TO INTERLEAVE IN FREQUENCY OVERCOMES THE DRAWBACK OF NOT OBTAINING DIVERSITY FROM THE EQUALIZER 3INCE EACH SUBCHANNEL EX PERIENCES AT FADING CODE DESIGN AND PERFORMANCE ANALYSIS DEVELOPED FOR AT FADING CHANNELS CAN BE USED $ECODING CAN BE PERFORMED WITH A 6ITERBI DECODER WHERE THE METRIC DEPENDS ON THE ESTIMATED CHANNEL ATTENUATIONS 4HIS MEANS THAT SYMBOLS ARE WEIGHED WITH THEIR RESPECTIVE CHANNEL STRENGTH 4HIS REDUCES THE EdECT OF ERRORS CAUSED BY SYMBOLS TRANSMITTED DURING A FADE

!CKNOWLEDGEMENTS
4HE AUTHORS WOULD LIKE TO EXTEND THEIR GRATITUDE TO THE STAd AT 4ELIA 2ESEARCH !" ,ULE AND THE COLLEAGUES AT THE $IVISION OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY FOR PROVIDING VALUABLE COMMENTS AND CORRECTIONS
!
4IME FREQUENCY LATTICE
)N THIS APPENDIX WE GIVE A BRIEF OVERVIEW OF THE EdECTS OF PULSE SHAPING 7E ONLY CONSIDER TIME AND FREQUENCY DISPERSION AND EXCLUDE CHANNEL NOISE IN THIS ANALYSIS 7E DESCRIBE THE /&$- SIGNAL AS 8 RS W S
JK JK JK
WHERE THE FUNCTIONS
JK
S ARE TRANSLATIONS OF A PROTOTYPE FUNCTION O S

S
JK
S O S ` K~ D
S
I {Jy S

4HIS ALLOWS US TO INTERPRET THE PULSE SHAPING PROBLEM 4HE RECEIVER USES THE FUNCTIONS S THAT ARE TRANSLATIONS OF A POSSIBLY DIdERENT PROTOTYPE FUNCTION O S
JK Q
JK
S O S ` K~ D
Q
I {Jy S
)N /&$- THESE FUNCTIONS FULL l S

JK
J K
p :J ` J K ` K <
WHERE Ha aI DENOTES THE %UCLIDEAN INNER PRODUCT ;= (ENCE THE TRANSMITTER AND RECEIVER FUNCTIONS ARE BI ORTHOGONAL ;= 4HIS SIMPLIES THE RECEIVER SINCE : l m R S S R S S CS W
JK
JK
JK
(OWEVER A TIME OR FREQUENCY DISPERSIVE CHANNEL DESTROYS THIS ORTHOGONALITY "Y CAREFULLY CHOOSING O S AND O S THE EdECTS OF THE LOSS OF ORTHOGONALITY CAN BE KEPT LOW !N /&$- SYSTEM MUST BE SUbCIENTLY RESISTANT TO BOTH TIME AND FREQUENCY DISPERSION 4HE FORMER CAN DEALT WITH BY INTRODUCING A GUARD SPACE USUALLY IN THE FORM OF A CYCLIC PREX WHILE THE LATTER IS OFTEN APPROACHED BY PULSE SHAPING )N SYSTEMS WITH A CYCLIC PREX AND NO PULSE SHAPING THE FUNCTIONS O S AND O S ARE CHOSEN AS THE RECTANGULAR PULSE ALTHOUGH OF DIdERENT LENGTHS 4HE RECEIVER PROTOTYPE FUNCTION O S IN THIS CASE IS SHORTER THAN THE TRANSMITTER PROTOTYPE FUNCTION O S WHICH CORRESPONDS TO THE REMOVAL OF THE CYCLIC PREX ! COMMON PROPAGATION MODEL IS OBTAINED BY ASSUMING THAT THE CHANNEL CONSISTS OF A NUMBER OF ELEMENTARY PATHS ;= WHERE EACH PATH IS DESCRIBED BY A DELAY A FREQUENCY OdSET AND A COMPLEX ATTENUATION 4HUS BY INVESTIGATING THE EdECTS OF A STATIC DELAY AND FREQUENCY OdSET THE SENSITIVITY TO A FADING MULTIPATH CHANNEL CAN BE EVALUATED ;= 4HIS
S Q S Q Q S

ANALYSIS CAN BE MADE WITH THE CROSS AMBIGUITY FUNCTION ;= OF THE PROTOTYPE FUNCTIONS O S AND O S : ~ E O S O S ` ~ D CS
S Q
WHICH CAN BE VIEWED AS A CROSSCORRELATION FUNCTION IN THE TIME FREQUENCY PLANE 4HE BI ORTHOGONALITY OF AND S REQUIRES THAT : l m S S D O S O S ` M~ D CS
JK JK I {LK~ JK J LK M
I {E S
I {LK~
I {Ly S
M~ Ly p :M L<
4HIS IS A CONDITION ON THE SAMPLES OF THE CROSS AMBIGUITY FUNCTION AT POSITIONS M~ Ly 4HE CROSS AMBIGUITY FUNCTION SHOULD BE ZERO FOR ALL M L BUT A DELAY OR FREQUENCY OdSET WILL DESTROY THE ORTOGONALITY SINCE ~ E IS NOT SAMPLED AT ITS ZEROS 7ITH A DELAY a~ AND A FREQUENCY OdSET ay THE SIGNAL POWER IS J a~ ay J AND THE INTERFERENCE POWER CAN BE UPPER BOUNDED BY ` J a~ ay J ;= 4HIS BOUNDS THE SIGNAL TO INTERFERENCE RATIO 3)2 FROM BELOW AS 3)2 w J a~ ay J ` J a~ ay J
4HUS THE CROSS AMBIGUITY FUNCTION IS A MEASURE OF THE INTERFERENCE IN THE SYSTEM CAUSED BY A DELAY OR A FREQUENCY OdSET )N ;= A PROTOTYPE FUNCTION IS CREATED WHICH IS CLAIMED TO HAVE A NEAR OPTIMUM CROSS AMBIGUITY FUNCTION 4HE CROSS AMBIGUITY FUNCTION FOR A RECTANGULAR PULSE IN A SYSTEM WITH CYCLIC PREX IS SHOWN IN &IGURE .OTE THAT
&IGURE !MBIGUITY FUNCTION FOR A RECTANGULAR PULSE AND CYCLIC PREX WITH LENGTHS ~ AND 3 RESPECTIVELY
BO
THE SYSTEM IS INSENSITIVE TO A TIME DELAY LESS THAN 3 FUNCTION IS AT AT THE TOP
BO
SINCE THE CROSS AMBIGUITY

4HE MAIN PROBLEM WITH USING A RECTANGULAR PULSE O S IS THAT IT IS NOT WELL LOCALIZED IN FREQUENCY $ENOTE THE TIME AND FREQUENCY WIDTHS RESPECTIVELY OF THE UNIT ENERGY SIGNAL O S BY : aS S JO S J CS : aE E J/ E J CE P WHERE / E IS THE &OURIER TRANSFORM OF O S 4HEN aS ~ AND aE FOR THE RECTANGULAR PULSE 4HIS FREQUENCY SPREAD OF ENERGY IS THE REASON FOR )#) IN THE CASE OF TRANSMISSION OVER FREQUENCY DISPERSIVE CHANNELS ;= 4HUS OTHER PULSES HAVE BEEN SOUGHT HAS A MINIMAL TIME TO OVERCOME THIS PROBLEM 3INCE THE 'AUSSIAN PULSE O S D BANDWIDTH PRODUCT ;= IT HAS BEEN USED TO FORM SUITABLE FUNCTIONS ; = 0ROLATE SPHEROIDAL WAVE FUNCTIONS ;= HAVE ALSO BEEN USED TO MINIMIZE OUT OF BAND ENERGY IN PULSES UNDER CERTAIN CONDITIONS ; =
{S
0ART !NALYSIS OF $&4 BASED CHANNEL ESTIMATORS FOR /&$!BSTRACT )N THIS PAPER WE ANALYZE THE PERFORMANCE OF THREE LOW COMPLEXITY CHANNEL ESTI MATORS BASED ON THE DISCRETE &OURIER TRANSFORM $&4 FOR ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- SYSTEMS %STIMATORS OF THIS TYPE HAVE BEEN ANALYZED FOR DISCRETE TIME CHANNELS AND WE EXTEND THIS ANALYSIS TO CONTINUOUS TIME CHANNELS 7E PRESENT ANALYT ICAL EXPRESSIONS FOR THEIR MEAN SQUARED ERROR -3% AND EVALUATE THEIR COMPLEXITY VERSUS SYMBOL ERROR RATE 3%2 FOR 1!- 4HE ANALYSIS SHOWS THAT THIS TYPE OF ESTIMATORS MAY EXPERIENCE AN IRREDUCIBLE ERROR OOR AT HIGH 3.2S (OWEVER IN ONE OF THE THREE ESTIMATORS THE ERROR OOR CAN BE ELIMINATED WHILE THE COMPLEXITY STAYS LOW AND THE PERFORMANCE IS MAXIMIZED
4HIS PART IS REPRODUCED FROM / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON !NALYSIS OF $&4 BASED CHANNEL ESTIMATORS FOR /&$- 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NI VERSITY OF 4ECHNOLOGY 3EPT

$&4 BASED CHANNEL ESTIMATORS

)NTRODUCTION
7IRELESS DIGITAL COMMUNICATION SYSTEMS USING COHERENT SIGNALING SCHEMES SUCH AS A QUADRATURE AMPLITUDE MODULATION 1!- REQUIRE ESTIMATION AND TRACKING OF THE FADING CHANNEL )N GENERAL THIS MEANS A MORE COMPLEX RECEIVER THAN FOR DIdERENTIAL MODULA TION SCHEMES SUCH AS DIdERENTIAL PHASE SHIFT KEYING $03+ WHERE THE RECEIVERS OPERATE WITHOUT A CHANNEL ESTIMATE ;= )N ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$SYSTEMS ; = $03+ IS APPROPRIATE FOR RELATIVELY LOW DATA RATES SUCH AS IN THE %URO PEAN DIGITAL AUDIO BROADCAST $!" SYSTEM ;= (OWEVER FOR MORE SPECTRALLY EbCIENT /&$- SYSTEMS COHERENT MODULATION IS MORE APPROPRIATE 7E ADDRESS LINEAR ESTIMATORS FOR /&$- WHERE ALL CHANNEL ATTENUATIONS IN A RECEIVED SYMBOL ARE ESTIMATED SIMULTANEOUSLY 5SING THE LINEAR MINIMUM MEAN SQUARED ERROR ,--3% ESTIMATOR WHICH TAKES ADVANTAGE OF THE CORRELATION BETWEEN ALL - SUBCARRI ERS REQUIRES AN - b - MATRIX MULTIPLICATION 4HIS COMPLEXITY CAN BE LARGE DEPENDING ON THE NUMBER OF SUBCARRIERS IN THE SYSTEM 4HIS PAPER PRESENTS AND ANALYZES THREE LOW COMPLEXITY SUBOPTIMAL APPROXIMATIONS OF THE ,--3% CHANNEL ESTIMATOR 4HESE ESTIMATORS ALL SHARE THE PROPERTY THAT THEY USE THE DISCRETE &OURIER TRANSFORM $&4 TO ESTIMATE THE CHANNEL IN THE TIME DOMAIN %STIMATORS OF THIS TYPE HAVE BEEN PROPOSED ; = BUT ONLY ANALYZED FOR DISCRETE TIME CHANNELS ;= 4HE ADDRESSED ESTIMATORS TAKE THE - NOISY FREQUENCY DOMAIN OBSERVATIONS AND TRANS FORM THEM TO THE TIME DOMAIN BY AN INVERSE $&4 )$&4 4HE LINEAR ESTIMATION IS THEN PERFORMED IN THE TIME DOMAIN AND THE RESULT TRANSFORMED BACK TO THE FREQUENCY DOMAIN BY A $&4 4HE TRANSFORMS CAN BE IMPLEMENTED WITH FAST ALGORITHMS REQUIRING ONLY A FEW MULTIPLICATIONS PER ESTIMATED ATTENUATION BUT THERE ARE STILL - COEbCIENTS TO ESTI MATE SIMULTANEOUSLY IN THE TIME DOMAIN (OWEVER AN /&$- SYMBOL TIME IS BY DESIGN MUCH LARGER THAN THE LENGTH OF THE CHANNEL 4HE TIME DOMAIN ESTIMATION TAKES ADVAN TAGE OF THE FACT THAT THIS CONCENTRATES THE CHANNEL POWER TO A RELATIVELY SMALL NUMBER OF TIME DOMAIN SAMPLES 4HREE OF THE STRATEGIES FOR DOING TIME DOMAIN APPROXIMATIONS ARE APPROXIMATING TIME DOMAIN SAMPLES WITH LOW CHANNEL POWER AS ZERO IGNORING CROSS CORRELATIONS AND IGNORING DIdERENCES IN VARIANCE 4HE THREE ESTIMATORS ANALYZED HERE USE THESE THREE STRATEGIES CUMULATIVELY !FTER PRESENTING THE /&$- SYSTEM MODEL IN 3ECTION WE INTRODUCE THE THREE $&4 BASED ESTIMATORS IN 3ECTION )N 3ECTION WE PRESENT AN ANALYSIS OF THE AVERAGE MEAN SQUARED ERROR -3% AND SHOW THAT THERE IS AN IRREDUCIBLE -3% OOR INHERENT IN $&4 BASED LOW COMPLEXITY ESTIMATORS 7E ALSO ILLUSTRATE THEIR PERFORMANCE BY PRESENTING THE UNCODED 1!- SYMBOL ERROR RATE FOR A TONE /&$- SYSTEM ! COMPLEXITY VERSUS PERFORMANCE COMPARISON IS DONE WHICH SINGLES OUT THE SECOND OF THE THREE ESTIMATORS AS A GOOD TRADE Od ! SUMMARY AND CONCLUDING REMARKS APPEAR IN 3ECTION

3YSTEM MODEL
&IGURE DISPLAYS THE /&$- BASE BAND SYSTEM USED IN THIS PAPER 7E ASSUME THAT THE USE OF A CYCLIC PREX #0 BOTH PRESERVES THE ORTHOGONALITY OF THE SUBCARRIERS AND ELIMI NATES INTER SYMBOL INTERFERENCE )3) BETWEEN CONSECUTIVE /&$- SYMBOLS ;= &URTHER

THE CHANNEL F ~ S IS ASSUMED TO BE SLOWLY 2AYLEIGH FADING AND CONSIDERED CONSTANT DUR ING ONE /&$- SYMBOL 4HE NUMBER OF TONES IN THE SYSTEM IS - AND THE LENGTH OF THE CYCLIC PREX IS + SAMPLES
W W
(#%3
.OISE # F ~ S #
#%3
X X
W- `
X- `
&IGURE "ASE BAND /&$- SYSTEM 4HE CYCLIC PREX IS NOT SHOWN IN THIS GURE
&IGURE 4HE /&$- SYSTEM DESCRIBED AS A SET OF PARALLEL 'AUSSIAN CHANNELS WITH CORRELATED ATTENUATIONS 5NDER THESE ASSUMPTIONS WE CAN DESCRIBE THE SYSTEM AS A SET OF PARALLEL 'AUSSIAN CHANNELS ;= SHOWN IN &IGURE WITH CORRELATED ATTENUATIONS t G &
J
J -3
u J - `
R

WHERE & a IS THE FREQUENCY RESPONSE OF THE CHANNEL F ~ S DURING THE /&$- SYMBOL AND 3 IS THE SAMPLING PERIOD OF THE SYSTEM )N MATRIX NOTATION WE DESCRIBE THE /&$SYSTEM AS X 7G
M
R
WHERE X IS THE RECEIVED VECTOR 7 IS A DIAGONAL MATRIX CONTAINING THE TRANSMITTED SIGNAL POINTS G IS A CHANNEL ATTENUATION VECTOR AND M IS A VECTOR OF INDEPENDENT AND IDENTICALLY DISTRIBUTED COMPLEX ZERO MEAN h 'AUSSIAN i NOISE VARIABLES WITH VARIANCE } 7ITHOUT LOSS OF GENERALITY WE ASSUME THAT $ JG J
M J

$&4 BASED ESTIMATORS
4HE TASK OF THE CHANNEL ESTIMATOR IS TO ESTIMATE THE CHANNEL ATTENUATIONS G FROM THE OBSERVATIONS X GIVEN THE TRANSMITTED SYMBOLS 7 &OR THE SAKE OF A TRACTABLE ANALYSIS WE ASSUME THE W S TO BE KNOWN AT THE RECEIVER 3INCE /&$- SYSTEMS ARE DESIGNED SUCH THAT THE SYMBOL TIME IS SIGNICANTLY LONGER THAN THE DURATION OF THE CHANNEL IMPULSE RESPONSE THE INVERSE $&4 OF THE CHANNEL ATTENUATION VECTOR G HAS MOST OF ITS POWER CONCENTRATED TO RELATIVELY FEW SAMPLES !S AN ILLUSTRATION OF THIS POWER CONCENTRATION &IGURE SHOWS THE CHANNEL POWER IN THE TIME DOMAIN FOR TWO CHANNEL TYPES SAMPLE SPACED AND NON SAMPLE SPACED 3AMPLE SPACED CHANNELS ARE CHANNELS THAT HAVE ALL FADING IMPULSES AT INTEGER MULTIPLES OF THE SYSTEM SAMPLING RATE AND FOR WHICH THE $&4 GIVES OPTIMAL POWER CONCENTRATION ;=
J
3AMPLE SPACED CHANNEL

0OWER
.ONSAMPLE SPACED CHANNEL

0OWER
4IME
"XBKHB OQDEHW .%#, RXLANK "XBKHB OQDEHW .%#, RXLANK
4IME
&IGURE 3CHEMATIC PICTURES OF THE TIME DOMAIN POWER DISTRIBUTION OF THE CHANNEL IE OF (#%3 G $&4 BASED ESTIMATION OF SAMPLE SPACED CHANNELS IS ADDRESSED IN ;= AND THE THREE ESTIMATORS WE PRESENT AND ANALYZE ARE GENERALIZATIONS THEREOF FOR NON SAMPLE SPACED CHAN NELS &OR CLARITY WE RST CALCULATE THE LINEAR MINIMUM MEAN SQUARED ERROR ,--3% ESTIMATE OF G 7E BASE OUR ESTIMATES ON THE ,3 ESTIMATE THE BACKROTATED OBSERVATIONS E 7 M IS A VECTOR OF INDEPENDENT 'AUSSIAN NOISE VARIABLES WITH COVARIANCE WHERE M ` a B CONSTITUTES A SUbCIENT STATISTIC SINCE MATRIX 1E E } 77 4HE ,3 ESTIMATE G 7 IS NON SINGULAR 4HE ,3 ESTIMATE IS A NOISY OBSERVATION OF THE CHANNEL ATTENUATIONS AND CAN BE SMOOTHED USING CORRELATION PROPERTIES OF THE CHANNEL 4HE OPTIMAL LINEAR ESTIMATOR IN TERMS OF MEAN SQUARED ERROR -3% IS ; =
' MM M KR
B 7 X G
M E G
KR

B 67 G B G
KR

'
WHERE 67 1
GG
GG
` a s
} 77
M

' GG
h i AND 1 $ GG IS THE AUTO COVARIANCE MATRIX OF THE CHANNEL VECTOR G !T THIS POINT WE RECOGNIZE THAT THE WEIGHTING MATRIX 67 IS OF SIZE - b - AND DEPENDS ON THE TRANSMITTED DATA 7 !S A RST STEP TOWARDS LOW COMPLEXITY ESTIMATORS WE WANT TO ND A WEIGHTING MATRIX THAT IS INDEPENDENT OF THE TRANSMITTED DATA 4HIS CAN BE B TO BE OUR OBSERVATION AND DERIVE AN ,--3% ESTIMATOR THAT OBTAINED BY CONSIDERING G CONSIDERS 7 TO BE STOCHASTIC WITH INDEPENDENT AND UNIFORMLY DISTRIBUTED CONSTELLATION POINTS )N THAT CASE THE AUTO COVARIANCE MATRIX OF THE NOISE BECOMES 1E E PEI ( i h i h WHERE n $ JW J $ JW J IS A CONSTELLATION FACTOR n FOR 1!- AND h i 2-1 $ JW J } IS THE PER SYMBOL SIGNAL TO NOISE RATIO B FROM G B NOW BECOMES 4HE ,--3% ESTIMATE OF G
KR n MM J J J M KR KR
B G
KLLRD
B 6G
KR
u
WHERE THE XED WEIGHTING MATRIX IS GIVEN BY t 61 1
GG GG
n ( 2-1

4HIS ,--3% ESTIMATOR STILL REQUIRES - MULTIPLICATIONS PER ESTIMATED ATTENUATION AND WE USE IT BOTH AS A REFERENCE AND AS A STARTING POINT IN THE DERIVATION OF THE $&4 BASED LOW COMPLEXITY ESTIMATORS 7E NOW USE THE PROPERTY OF /&$- SYSTEMS IDENTIED ABOVE AND IN ; = G IS THE SAMPLED FREQUENCY RESPONSE OF A CHANNEL WITH SHORT TIME DURATION COMPARED TO THE /&$- SYMBOL LENGTH AND HENCE ITS ASSOCIATED CYCLIC IMPULSE RESPONSE F (#%3 G HAS ONLY A FEW TAPS WITH SIGNICANT POWER )F WE PERFORM THE ESTIMATION IN THE TIME DOMAIN WE CAN REDUCE THE COMPLEXITY OF THE ESTIMATION BY USING THIS POWER CONCENTRATION 4HIS PROMPTS THE ESTIMATOR STRUCTURE IN &IGURE THE ,3 ESTIMATE IS TRANS r WHERE s B 4HE SMOOTHING IS THEN PER B (#%3 G FORMED INTO ITS TIME DOMAIN EQUIVALENT F FORMED BY A LINEAR TRANSFORMATION B 0F B F
KR KR KR
B #%3 F B 4HE IMPOR AND THE RESULT IS TRANSFORMED BACK TO THE FREQUENCY DOMAIN G TANT BENET OF THIS ESTIMATOR STRUCTURE IN TERMS OF COMPLEXITY IS THE LOW COMPLEXITY OF THE $&4)$&4 IMPLEMENTED AS FAST TRANSFORMS AND THE TIME DOMAIN POWER CONCENTRATION 4HIS OdERS A SIMPLICATION OF WITHOUT SACRICING TOO MUCH IN PERFORMANCE /UR APPROACH IS TO ND SPARSE APPROXIMATIONS OF THE ,--3% ESTIMATORS EQUIVALENT TIME DOMAIN SMOOTHING MATRIX 0 % 6%
KR '
WHERE % IS THE - b - UNITARY $&4 MATRIX AND 6 IS DENED IN 4HIS WILL REDUCE THE NUMBER OF REQUIRED MULTIPLICATIONS AND THUS THE ESTIMATOR COMPLEXITY ! STRAIGHTFORWARD B THAT CONTAIN MORE NOISE THAN CHANNEL POWER WAY IS TO SIMPLY IGNORE THE COEbCIENTS IN F AND ONLY TRANSFORM THE REMAINING ELEMENTS BACK TO THE FREQUENCY DOMAIN &OR SAMPLE SPACED CHANNELS THIS IS A FRUITFUL APPROACH ; = SINCE THE MAJOR PART OF THE COEbCIENTS ONLY CONTAIN NOISE AND NO CHANNEL POWER )F THE CHANNEL IS NOT SAMPLE SPACED HOWEVER THE CHANNEL POWER IS STILL CONCENTRATED BUT DISTRIBUTED OVER ALL COEbCIENTS $UE TO THE
KR

,INEAR TRANSFORMATION
X
&REQUENCY ` W B G
KR
4IME
B F
KR
&REQUENCY
B F B G
)$&4
X `
-
W` ` B G `
KR-
$&4
B ` F
-
B F
KR-
`
B G `
-
&IGURE 3TRUCTURE OF $&4 BASED CHANNEL ESTIMATORS WHERE THE LINEAR TRANSFORMATION B 0F B CONSISTS OF A MATRIX MULTIPLICATION F
KR
LOST CHANNEL POWER IN THE IGNORED COEbCIENTS THE SIMPLICATION CAUSES AN IRREDUCIBLE ERROR OOR ;= 7E NOW MOVE THROUGH THREE SIMPLICATION STEPS AND OBTAIN THREE DIdERENT LOW COMPLEXITY ESTIMATORS OF WHICH THE LAST IS THE STRAIGHTFORWARD APPROACH DESCRIBED ABOVE 4HE GENERAL CONCEPT IS BASED ON REDUCING THE NUMBER OF NON ZERO ELEMENTS IN THE TIME DOMAIN MATRIX MULTIPLICATION WITH THE AIM OF REDUCING THE COMPUTATIONAL COM PLEXITY AND PRESERVING THE PERFORMANCE 4HE THREE ESTIMATORS ARE SELECTED AS FOLLOWS 3EE !PPENDIX ! FOR A DETAILED DERIVATION q %STIMATOR ! B THAT HAVE THE HIGHEST CHANNEL POWER WE "Y CHOOSING THE , COEbCIENTS IN F RESTRICT THE LINEAR TRANSFORM IN THE TIME DOMAIN TO A XED MATRIX OF SIZE , b , )F , IS CHOSEN MUCH SMALLER THAN - THE COMPLEXITY REDUCTION COMPARED TO THE ,--3% IS CONSIDERABLE 4HE COMPLEXITY OF THE TIME DOMAIN PROCESSING IN THIS CASE IS , - MULTIPLICATIONS PER ESTIMATED ATTENUATION 4HIS ESTIMATOR CONVERGES TO THE ,--3% WHEN , - 7E HAVE PRESENTED A RELATED ESTIMATOR PREVIOUSLY IN ;=
KR
q %STIMATOR " &URTHER REDUCTIONS IN COMPLEXITY CAN BE DONE BY IGNORING CROSS CORRELATION BETWEEN B AND ONLY WEIGHTING THEM INDIVIDUALLY 4HIS ESSENTIALLY THE , CHOSEN TAPS IN F MEANS THAT WE RESTRICT THE TIME DOMAIN PROCESSING TO BE A DIAGONAL , b , MATRIX MULTIPLICATION 3INCE THIS MATRIX ONLY HAS , NON ZERO ELEMENTS THE COMPLEXITY OF THIS PROCESSING IS ,- MULTIPLICATIONS PER ESTIMATED ATTENUATION 4O THE AUTHORS KNOWLEDGE THIS ESTIMATOR HAS NOT BEEN PRESENTED BEFORE
KR
q %STIMATOR # )N THIS LAST ESTIMATOR WE FURTHER RESTRICT THE TIME DOMAIN PROCESSING TO ONLY USE THE , CHOSEN COEbCIENTS DIRECTLY AS INPUT TO THE $&4 4HIS MEANS RESTRICTING THE MATRIX TO AN , b , IDENTITY MATRIX WHICH DOES NOT REQUIRE ANY MULTIPLICATIONS

$&4 BASED CHANNEL ESTIMATORS AT ALL 7HEN , - THIS ESTIMATOR CONVERGES TO THE ,3 ESTIMATOR 4HIS ESTIMATOR IS SIMILAR TO THE ESTIMATOR DESIGNED FOR SAMPLE SPACED CHANNELS IN ; =
%STI MATOR ,--3% ! " # ,3
2ESTRICTION CUMULATIVE .! 5SE - COEd $IAGONAL MATRIX )DENTITY MATRIX .!
,INEAR TRANSFORMATION .! v w 1, , 1 w v $, , 1! w v ), , 1" .!
%dECTIVE MATRIX SIZE . b. - b- b - DIAG - b - IDENT .!
2EQUIRED MULTATTEN . LOG . LOG .

,
, -
LOG .
4ABLE !NALYSED $&4 BASED ESTIMATORS 3EE !PPENDIX ! FOR DETAILS 4HE OUTLINED ESTIMATORS ! # ARE SUMMARIZED IN 4ABLE WHERE WE HAVE ALSO INCLUDED THE ,--3% AND ,3 ESTIMATORS AS REFERENCES 4HE TABLE SHOWS THE TOTAL COM PUTATIONAL COMPLEXITY INCLUDING THE )$&4 AND $&4 FOR ESTIMATORS !# )N GENERAL UNLESS , IS CLOSE TO - THE COMPLEXITY DECREASES IN 4ABLE FROM THE ,--3% TO THE ,3

0ERFORMANCE ANALYSIS
4HE ESTIMATORS PRESENTED IN THE PREVIOUS SECTION ALL HAVE DIdERENT COMPUTATIONAL COM PLEXITIES AND THE DESIGN VARIATIONS GIVE THEM DIdERENT PERFORMANCES AS WELL 4O ILLUSTRATE THE PERFORMANCE OF THESE ESTIMATORS WE CALCULATE THE -3% AND USE FORMULAE FROM ;= TO OBTAIN THE UNCODED 1!- SYMBOL ERROR RATE 3%2 4HE PARAMETERS WE HAVE CHOSEN FOR THE /&$- SYSTEM ARE - SUBCARRIERS AND A CYCLIC PREX OF LENGTH + SAMPLES 4HE IMPULSE RESPONSE OF THE PHYSICAL CHANNEL F ~ S 8
L
S p ~ `~
CONSISTS OF INDEPENDENT 2AYLEIGH FADING m IMPULSES UNIFORMLY DISTRIBUTED ~ OVER THE LENGTH OF THE CYCLIC PREX AND WITH A CONSTANT POWER DELAY PROLE &ROM THIS CHANNEL MODEL WE CALCULATE THE AUTO COVARIANCE MATRIX 1 OF THE CHANNEL ATTENUATIONS G ;= 4HE CYCLIC IMPULSE RESPONSE F % G HAS A CORRESPONDING h i AUTO COVARIANCE MATRIX 1 % 1 % 7E ASSUME THAT THE VARIANCES o $ JF J ARE IN DECREASING ORDER IE o w o w o 4HE AUTO COVARIANCE MATRIX OF THE , USED TRANSFORM COEbCIENTS 1 IS THUS THE UPPER LEFT , b , CORNER OF 1
L L GG ' FF ' GG J J FF, FF
OBTAIN A COMPLEXITY MEASURE WE HAVE ASSUMED THAT . IS A POWER OF TWO AND THAT THE $&4 AND THE )$&4 REQUIRES . LOG . COMPLEX MULTIPLICATIONS EACH ;=
4O

4HROUGH DIRECT CALCULATION OF THE AUTO COVARIANCE MATRIX OF THE ESTIMATION ERROR |r sr s } B B 1 $ G`G G`G
' D0 D0
FOR ALL ESTIMATORS WE OBTAIN THEIR RESPECTIVE AVERAGE -3%S ,2$ ` 3Q@BD 1 D0 D0
a
D0 D0

.OTE THAT THE DIAGONAL ELEMENTS OF 1 ARE THE INDIVIDUAL ERROR VARIANCES FOR EACH CHANNEL ATTENUATION 4HE CALCULATIONS ARE DERIVED IN !PPENDIX ! 4HE NAL -3% EXPRESSIONS FOR THE ,--3% ESTIMATOR THE LOW COMPLEXITY ESTIMATORS !# AND THE ,3 ESTIMATOR ARE DISPLAYED IN 4ABLE !PPENDIX ! ALSO CONTAINS A NOTE ON SAMPLE SPACED CHANNELS THAT IS OF INTEREST WHEN COMPARING THIS ANALYSIS WITH THE ANALYSIS IN ;= !S DISPLAYED IN THE TABLE ESTIMATORS ! " AND # EXPERIENCE AN ERROR OOR ,2$ DUE TO THE CHANNEL POWER IN THE - ` , EXCLUDED CHANNEL TAPS 4HIS ERROR OOR IS THE SAME FOR ALL $&4 BASED LOW COMPLEXITY ESTIMATORS 4HE INDIVIDUAL RANKING OF THE LOW COMPLEXITY ESTIMATORS IN TERMS OF -3% FOR A XED , IS ,2$ v ,2$ v ,2$
! "
%STIMATOR ,--3% ! " # ,3
!VERAGE -3% wJ0 n

-
PEI
n
J( ,

-
0 0
n wJ- 2-1
PEI
n
J( ,
wJ, n wJ, 2-1 oJ oJ
-3%

-
PEI
n
J(
2-1
-3%
, -
PEI
-3%
PEI
wJ, EIGENVALUES OF 2FF, o J DIAGONAL ELEMENTS OF 2FF IN DECREASING ORDER -3%

-
0
J (,
oJ
-3% OOR
FF,
4ABLE !VERAGE -3% FOR THE INVESTIGATED ESTIMATORS 1 DENOTES THE COVARIANCE MATRIX OF THE , DOMINATING TAPS IN F .OTE THAT THE w S ARE EIGENVALUES OF 1
JFF
5SING THE FORMULAE FROM ;= WE DISPLAY 1!- 3%2 CURVES FOR TWO DIdERENT NUMBERS OF INCLUDED TRANSFORM COEbCIENTS , AND , IN &IGURE 4HE 1!- 3%2 OF THE FULL ,--3% ESTIMATOR AND THE ,3 ESTIMATOR ARE ALSO INCLUDED IN THE GURE AS REFERENCES .OTE THAT FOR , THERE IS NO VISIBLE DIdERENCE IN 3%2 BETWEEN THE THREE LOW COMPLEXITY ESTIMATORS AND THAT THE 3%2 LEVELS Od AT HIGH 3.2S DUE TO

THE ERROR OOR )T IS ONLY FOR HIGHER VALUES OF , THAT A DIdERENCE IN 3%2 IS NOTICEABLE BETWEEN THE LOW COMPLEXITY ESTIMATORS )T IS ALSO NOTEWORTHY THAT EVEN AT , OUT OF A POSSIBLE - THE ERROR OOR IS VISIBLE AND THE LOW COMPLEXITY ESTIMATORS PERFORM WORSE THAN THE ,3 ESTIMATOR ABOVE A CERTAIN 3.2 &OR , THE GURE ALSO SHOWS THAT THE PERFORMANCE OF THE ESTIMATORS IS DECREASING FROM ! TO #

1!- 3%2

1!- 3%2

-
,--3% ! " # ,3
-
,--3% ! " # ,3

3.2 ;D"=
3.2 ;D"=

&IGURE 5NCODED 1!- 3%2 FOR THREE SYSTEMS USING THE LOW COMPLEXITY ESTIMATORS ! " AND # RESPECTIVELY .OTE %STIMATORS ! " AND # HAVE THE SAME 3%2 FOR , 4HE PERFORMANCE OF THE LOW COMPLEXITY ESTIMATORS ESPECIALLY FOR HIGH 3.2S DEPENDS STRONGLY ON THE NUMBER OF INCLUDED TAPS !N ANALYSIS OF THIS BEHAVIOR SHOWS THAT THE THREE LOW COMPLEXITY ESTIMATORS ARE QUITE DIdERENT IN TERMS OF COMPLEXITY VERSUS PERFORMANCE 4HIS IS ILLUSTRATED IN &IGURE WHICH SHOWS THAT THE COMPLEXITY NEEDS TO BE HIGH TO MAKE ESTIMATOR ! THE BEST IN TERMS OF 3%2 )N GENERAL ESTIMATORS " AND # ARE MORE EbCIENT PER REQUIRED MULTIPLICATION &OR A SUBCARRIER SYSTEM THE COMPLEXITY OF ESTIMATOR # IS ALWAYS MULTIPLICATIONS PER ATTENUATION WHILE ESTIMATOR " HAS A COMPLEXITY OF BETWEEN AND MULTIPLICATIONS PER ATTENUATION DEPENDING ON THE NUMBER OF TAPS USED )F WE WANT TO ELIMINATE THE ERROR OOR ENTIRELY WE HAVE TO USE ALL TAPS , IN ESTIMATORS !# %dECTIVELY THIS TURNS ESTIMATOR ! INTO THE HIGH COMPLEXITY ,--3% ESTIMATOR AND ESTIMATOR # INTO THE LOW PERFORMANCE ,3 ESTIMATOR (OWEVER ESTIMATOR " WHICH HAS LOWER COMPLEXITY THAN ESTIMATOR ! AND BETTER PERFORMANCE THAN ESTIMATOR # IS A GOOD COMPROMISE BETWEEN THE TWO %STIMATOR ! DOES NOT HAVE A COMPLEXITY LOW ENOUGH TO COMPETE WITH THE OTHER TWO AND THE APPROXIMATIONS IN ESTIMATOR # SEEM TO BE TOO CRUDE TO PROVIDE A GOOD ESTIMATE 4HIS ENABLES US TO SINGLE OUT ESTIMATOR " AS THE

1!- 3%2
,3 #
3.2 D"

" ! ,--3%

-ULTIPLICATIONS PER ESTIMATED ATTENUATION

&IGURE 3%2 VERSUS ESTIMATOR COMPLEXITY AT 3.2 D" .OTE THE DISCONTINUITIES OF THE ABSCISSA BEST TRADE Od BETWEEN COMPLEXITY AND PERFORMANCE AMONG THE THREE ESTIMATORS ANALYZED !S A NAL COMPARISON WE PRESENT THE 1!- 3%2 PERFORMANCE OF ESTIMATOR " WITH ALL TAPS INCLUDED THE ,--3% ESTIMATOR AND THE ,3 ESTIMATOR IN &IGURE 4HE 3%2 PERFORMANCE OF ESTIMATOR " IS GOOD FOR LOW 3.2S WHERE IT RESEMBLES THAT OF THE ,--3% ESTIMATOR AND THIS AT LESS THAN OF THE COMPLEXITY NUMBER OF MULTIPLICATIONS OF THE ,--3% !T HIGH 3.2S IT CONVERGES TO THE 3%2 OF THE ,3 ESTIMATOR BUT THERE IS NO ERROR OOR AND THE PERFORMANCE NEVER BECOMES WORSE THAN THAT OF THE ,3 ESTIMATOR %STIMATOR " HAS MORE THAN A D" GAIN OVER THE ,3 ESTIMATE FOR 3.2S LESS THAN D" AND A D" GAIN FOR 3.2S BELOW D"

#ONCLUSIONS
)N THIS PAPER WE HAVE PRESENTED THE -3% AND 3%2 PERFORMANCES OF THREE LOW COMPLEXITY $&4 BASED ESTIMATORS !# 5SING , OF THE - TIME DOMAIN COEbCIENTS YIELDS ESTIMA TORS WITH COMPLEXITY EXCLUDING THE $&4)$&4 OF , - ,- AND MULTIPLICATIONS PER ESTIMATED CHANNEL ATTENUATION RESPECTIVELY /NLY THE RST OF THESE ESTIMATORS IS PO TENTIALLY OF HIGH COMPLEXITY SINCE , v - 7E HAVE PROVIDED ANALYTICAL EXPRESSIONS FOR THE MEAN SQUARED ERROR OF ALL THREE ESTIMATORS AND SHOWN THAT IF THE NUMBER OF INCLUDED TAPS IS LESS THAN - THEY SUdER FROM AN IRREDUCIBLE ERROR OOR 4HE ERROR OOR CAN BE REMOVED IF ALL TAPS ARE USED IN THE LINEAR TRANSFORM BUT ONLY ES TIMATOR " MAINTAINS BOTH ITS PERFORMANCE AND ITS LOW COMPLEXITY IN THIS CASE 4HE OTHER TWO DESIGNS EITHER EXPERIENCE A DRASTICALLY INCREASED COMPLEXITY ESTIMATOR ! OR CON VERGE TO THE POOR PERFORMANCE OF THE ,3 ESTIMATOR ESTIMATOR # %STIMATOR " MAINTAINS GOOD PERFORMANCE WITH LOW COMPLEXITY BY IGNORING THE RELATIVELY SMALL CROSS CORRELATION

1!- 3%2

%STIMATOR
,--3% " - ,3

MULT

3.2 ;D"=
&IGURE 5NCODED 1!- 3%2 FOR ESTIMATOR " WHERE ALL TAPS ARE USED , - #URVES FOR THE ,--3% AND ,3 ESTIMATORS ARE INCLUDED AS REFERENCES BETWEEN THE TIME DOMAIN CHANNEL COEbCIENTS 3O WE CONSIDER ESTIMATOR " USING ALL COEbCIENTS TO BE THE MOST SUITABLE OF THE $&4 BASED CHANNEL ESTIMATORS )N THE INVESTI GATED SYSTEM IT HAS ALMOST THE SAME PERFORMANCE AS THE ,--3% ESTIMATOR FOR LOW 3.2S AND THIS AT LESS THAN OF THE COMPLEXITY &URTHER IN TERMS OF SYMBOL ERROR RATE THIS ESTIMATOR HAS MORE THAN A D" GAIN OVER THE ,3 ESTIMATOR FOR 3.2S LESS THAN D"
!
%STIMATOR EXPRESSIONS
)N THIS APPENDIX WE PROVIDE EXPRESSIONS FOR THE LINEAR TRANSFORMATIONS OF THE THREE LOW COMPLEXITY ESTIMATORS AND THE -3% FOR EACH ESTIMATOR !T THE END WE HAVE ALSO INCLUDED A NOTE ON SAMPLE SPACED CHANNELS WHICH IS OF INTEREST WHEN COMPARING OUR ANALYSIS TO THE ONE IN ;= 4O SIMPLIFY THE MATRIX NOTATION WE ASSUME THAT THE COEbCIENTS IN THE CYCLIC IM B ARE ORDERED ACCORDING TO DECREASING CHANNEL POWER 4HIS IS JUSTIED PULSE RESPONSE F SINCE PERMUTATIONS OF THE $&4)$&4 COEbCIENTS DO NOT CHANGE THE ESTIMATORS IT ONLY CHANGES THE ORDER IN WHICH h THE i COEbCIENTS ARE INDEXED 4HE CHANNEL POWER IN COEbCIENT F IS DENOTED o $ JF J WHICH ARE THE DIAGONAL ELEMENTS OF 1 B "EFORE WE START THE DERIVATION WE WRITE THE ESTIMATOR STRUCTURE &IGURE IN MATRIX NOTATION B %0% 7 X %0% G B G
KR KRJ J J FF ' ' 0 KR
WHERE 0 IS THE MATRIX REPRESENTING THE LINEAR TRANSFORMATION IN THE TIME DOMAIN AND % IS THE - b - UNITARY $&4 MATRIX .OTE % % &URTHER WE NEED THE AUTO COVARIANCE MATRIX OF F WHICH IS ` a 1 $ FF % 1 %
' ' ' FF GG

7E ARE MINIMIZING THE ,2$ AND THUS NEED AN EXPLICIT EXPRESSION FOR THE AUTO COVARIANCE MATRIX OF THE ESTIMATION ERROR |r sr s } B0 ` G G B0 ` G 1 $ G t u n %0% 1
( %0 % ` 2-1 1 %0 % ` %0% 1
1
' D0 D0 ' ' ' GG ' ' ' ' GG GG GG
WHICH GIVES THE ,2$ ACCORDING TO 4O SAVE SPACE THE CALCULATIONS BELOW ARE NOT PRESENTED IN DETAIL (OWEVER THE RESPEC TIVE 0 MATRICES OF THE ESTIMATORS ARE SUBSTITUTED IN AND THE FOLLOWING EQUALITIES ARE APPLIED ` a q 3Q@BD 414 3Q@BD 1 WHEN 4 IS A UNITARY MATRIX
'
q 3Q@BD # 1# 3Q@BD # # 1 WHEN # AND # ARE DIAGONAL MATRICES c b 5 AND 4 CONTAINS THE RST , q 1 4 1 4 WHERE % 4 COLUMNS OF %
FF, ' , GG , , , ,
q 4 5
' ,
WHERE 4
AND 5
ARE DENED ABOVE
4O SIMPLIFY THE ,2$ EXPRESSIONS WE ALSO USE THAT THE TRACE OF A MATRIX IS EQUIVALENT TO THE SUM OF ITS EIGENVALUES ;= &URTHER RELATION IMPLIES THAT 1 AND 1 SHARE THE SAME EIGENVALUES ;= WHICH IS USED TO AVOID SEPARATE NOTATIONS
GG FF
q ,3 ESTIMATOR 4HE ,3 ESTIMATE IS USED AS INPUT TO THE )$&4 IN &IGURE AND ITS -3% IS ,2$ n 2-1
4HIS ESTIMATE ONLY REQUIRES ONE MULTIPLICATION PER ESTIMATED ATTENUATION 3INCE IT IS THE INPUT TO THE REST OF THE ESTIMATORS THIS ONE MULTIPLICATION WILL SHOW UP IN THE FOLLOWING COMPLEXITY EXPRESSIONS q ,--3% ESTIMATOR 4HE ,--3% ESTIMATOR IS GIVEN IN AND AND ITS ,2$ BECOMES t u n ,2$ 3Q@BD 1 (` 1
( 1 2-1
GG GG GG
w n 8 - 2-1 ( w
J JGG
Jn

FF
PEI
WHERE THE w
J-
S ARE EIGENVALUES OF 1
AND 1
)MPLEMENTING THIS ESTIMATOR AS A MATRIX MULTIPLICATION AS IN REQUIRES -

MULTIPLICATIONS PER ESTIMATED ATTENUATION
q %STIMATOR !
)MPOSING THE RST RESTRICTION ON THE LINEAR TRANSFORMATION v w 0 0

, ,
THE MINIMAL -3% IS OBTAINED IF 0 WHERE 1

FF, ,
t
FF,
1
FF,
n
( 2-1
FF
u
IS THE UPPER LEFT , b , CORNER OF 1 t (` 1 u n

( 1 2-1
4HE ,2$ OF THIS ESTIMATOR IS ,2$ 3Q@BD 1 ,
FF,
FF,
FF,
,2$
w n 8 - 2-1 ( w
J J,
J, n
,2$ AND
GG -
PEI
FF,
WHERE THE w
J,
S ARE EIGENVALUES OF 1 ,2$ 3Q@BD 5 `

' -
1 5
a
,
8 o - (
J J ,

WHERE 5 CONTAINS THE LAST - ` , COLUMNS OF % 7E CALL ,2$ THE -3% OOR SINCE IT ONLY DEPENDS ON THE NUMBER OF EXCLUDED TAPS - ` , AND LOWER BOUNDS THE -3%
,
)MPLEMENTING THIS ESTIMATOR ACCORDING TO &IGURE REQUIRES KNF -

TIPLICATIONS PER ESTIMATED ATTENUATION q %STIMATOR " !PPLYING THE SECOND RESTRICTION ON THE LINEAR TRANSFORMATION v w # 0
, , "
MUL
WHERE #
CH@F p p p
WE OBTAIN A MINIMAL -3% IF o p o

PEI
J J n J ,
4HE -3% OF THIS ESTIMATOR IS
o n 8 ,2$ - 2-1 ( o
J J
J n
,2$
PEI
WHERE ,2$ IS GIVEN BY )MPLEMENTING THIS ESTIMATOR ACCORDING TO &IGURE REQUIRES KNF -
TIPLICATIONS PER ESTIMATED ATTENUATION
, -
MUL
$&4 BASED CHANNEL ESTIMATORS q %STIMATOR # !PPLYING THE LAST RESTRICTION ON THE LINEAR TRANSFORMATION v w ( 0
, , #

WE DO NOT HAVE ANY CHOICE IN THE DESIGN EXCEPT FOR , AND THE -3% BECOMES ,2$ , n
,2$ - 2-1
WHERE AGAIN ,2$ IS GIVEN BY )MPLEMENTING THIS ESTIMATOR ACCORDING TO &IGURE REQUIRES KNF -
MULTIPLI CATIONS PER ESTIMATED ATTENUATION ! NOTE ON SAMPLE SPACED CHANNELS 4HE ABOVE EXPRESSIONS ARE DERIVED FOR A GENERAL CASE BUT THEY HAVE SOME INTERESTING PROPERTIES FOR SAMPLE SPACED CHANNELS THAT ARE WORTH NOTING #ONSIDER A CHANNEL IMPULSE RESPONSE 8 m S p ~ ` L3 F ~ S
+ L R L(
WHERE THE FADING AMPLITUDES m S OF THE SAMPLE SPACED IMPULSES ARE INDEPENDENT 4HEN % DIAGONALIZES 1 AND 1 BECOMES DIAGONAL WITH ONLY + NON ZERO ELEMENTS B IS THEREFORE EQUIVALENT TO THE J TH LARGEST 4HE CHANNEL POWER IN THE J TH COEbCIENT OF F EIGENVALUE w OF 1 AND 1 AND THE EIGENVALUES OF 1 BECOMES w w o OF WHICH ONLY THE RST + ARE NON ZERO (ENCE THE ,2$S FOR THE ESTIMATORS BECOME
L GG FF KR JGG FF FF, J, JJ
%STIMATOR !VERAGE -3% 0 ,--3% PEI

n oJ J (
! " # ,3

-
0 0
oJ
2-1
n
PEI
n
oJ oJ
J( ,
2-1
n
,2$
,2$

-
PEI
n
oJ oJ
J(
2-1
, -
PEI
,2$
PEI
&IRST OF ALL ESTIMATOR ! IS NOW EQUIVALENT TO ESTIMATOR " AND SINCE o FOR J w + BOTH ESTIMATOR ! AND " ARE EQUIVALENT TO THE ,--3% ESTIMATOR FOR , w + &URTHER SINCE ,2$ FOR , w + WE CAN CHOOSE , + IN ESTIMATOR # WHICH REDUCES THE NOISE COMPARED TO THE ,3 ESTIMATOR TO A FRACTION +- 4HIS LAST OBSERVATION WAS ALSO DONE IN ;=
J

0ART /&$- CHANNEL ESTIMATION BY SINGULAR VALUE DECOMPOSITION

!BSTRACT )N THIS PAPER WE PRESENT AND ANALYSE LOW RANK CHANNEL ESTIMATORS FOR ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- USING THE FREQUENCY CORRELATION OF THE CHANNEL ,OW RANK APPROXIMATIONS BASED ON THE DISCRETE &OURIER TRANSFORM $&4 HAVE BEEN PRO POSED BUT THEY SUdER FROM POOR PERFORMANCE WHEN THE CHANNEL IS NOT SAMPLE SPACED 7E APPLY THE THEORY OF OPTIMAL RANK REDUCTION TO LINEAR MINIMUM MEAN SQUARED ERROR ,--3% ESTIMATORS AND SHOW THAT THESE ESTIMATORS WHEN USING A XED DESIGN ARE RO BUST TO CHANGES IN CHANNEL CORRELATION AND SIGNAL TO NOISE RATIO 3.2 4HE PERFORMANCE IS PRESENTED IN TERMS OF UNCODED SYMBOL ERROR RATE 3%2 FOR A SYSTEM USING 1!-
4HIS PART IS REPRODUCED FROM / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON /&$- CHANNEL ESTIMATION BY SINGULAR VALUE DECOMPOSITION 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT

36$ BASED CHANNEL ESTIMATORS

)NTRODUCTION
7IRELESS DIGITAL COMMUNICATION SYSTEMS USING MULTI AMPLITUDE MODULATION SCHEMES SUCH AS QUADRATURE AMPLITUDE MODULATION 1!- REQUIRE ESTIMATION AND TRACKING OF THE FADING CHANNEL )N GENERAL THIS MEANS A MORE COMPLEX RECEIVER THAN FOR DIdERENTIAL MODULATION SCHEMES SUCH AS DIdERENTIAL PHASE SHIFT KEYING $03+ WHERE THE RECEIVERS OPERATE WITHOUT A CHANNEL ESTIMATE ;= )N ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- SYSTEMS $03+ IS APPROPRIATE FOR RELATIVELY LOW DATA RATES SUCH AS IN THE %UROPEAN DIGITAL AUDIO BROADCAST $!" SYSTEM ;= (OWEVER FOR MORE SPECTRALLY EbCIENT /&$- SYSTEMS COHERENT MODULATION IS MORE APPROPRIATE 4HE STRUCTURE OF /&$- SIGNALLING ALLOWS A CHANNEL ESTIMATOR TO USE BOTH TIME AND FREQUENCY CORRELATION 3UCH A TWO DIMENSIONAL ESTIMATOR STRUCTURE IS GENERALLY TOO COM PLEX FOR A PRACTICAL IMPLEMENTATION 4O REDUCE THE COMPLEXITY SEPARATING THE USE OF TIME AND FREQUENCY CORRELATION HAS BEEN PROPOSED ;= 4HIS COMBINED SCHEME USES TWO SEPARATE &)2 7IENER LTERS ONE IN THE FREQUENCY DIRECTION AND THE OTHER IN THE TIME DIRECTION )N THIS PAPER WE PRESENT AND ANALYSE A CLASS OF BLOCK ORIENTED CHANNEL ESTIMATORS FOR /&$- WHERE ONLY THE FREQUENCY CORRELATION OF THE CHANNEL IS USED IN THE ESTIMATION 7HATEVER THEIR LEVEL OF PERFORMANCE IT MAY BE IMPROVED WITH THE ADDITION OF A SECOND LTER USING THE TIME CORRELATION ; = 4HOUGH A LINEAR MINIMUM MEAN SQUARED ERROR ,--3% ESTIMATOR USING ONLY FRE QUENCY CORRELATION HAS LOWER COMPLEXITY THAN ONE USING BOTH TIME AND FREQUENCY COR RELATION IT STILL REQUIRES A LARGE NUMBER OF OPERATIONS 7E INTRODUCE A LOW COMPLEXITY APPROXIMATION TO A FREQUENCY BASED ,--3% ESTIMATOR THAT USES THE THEORY OF OPTIMAL RANK REDUCTION /THER TYPES OF LOW RANK APPROXIMATIONS BASED ON THE DISCRETE TIME &OURIER TRANSFORM $&4 HAVE BEEN PROPOSED FOR /&$- SYSTEMS BEFORE ; = 4HE WORK PRESENTED IN THIS PAPER WAS INSPIRED BY THE OBSERVATIONS IN ;= WHERE IT IS SHOWN THAT $&4 BASED LOW RANK CHANNEL ESTIMATORS HAVE LIMITED PERFORMANCE FOR NON SAMPLE SPACED CHANNELS AND HIGH 3.2S !FTER PRESENTING THE /&$- SYSTEM MODEL AND OUR SCENARIO IN 3ECTION WE IN TRODUCE THE ESTIMATORS AND DERIVE THEIR COMPLEXITIES IN 3ECTION 7E ANALYSE THE SYMBOL ERROR RATE 3%2 PERFORMANCE IN 3ECTION WHERE WE ALSO DISCUSS DESIGN CONSID ERATIONS 4HE PROPOSED LOW RANK ESTIMATOR IS COMPARED TO OTHER ESTIMATORS IN 3ECTION AND A SUMMARY AND CONCLUDING REMARKS APPEAR IN 3ECTION

3YSTEM DESCRIPTION
3YSTEM MODEL
&IGURE DISPLAYS THE /&$- BASE BAND MODEL USED IN THIS PAPER 7E ASSUME THAT THE USE OF A CYCLIC PREX #0 ;= BOTH PRESERVES THE ORTHOGONALITY OF THE TONES AND ELIMINATES INTER SYMBOL INTERFERENCE )3) BETWEEN CONSECUTIVE /&$- SYMBOLS &URTHER THE CHANNEL F S ~ IS ASSUMED TO BE SLOWLY FADING SO IT IS CONSIDERED TO BE CONSTANT

DURING ONE /&$- SYMBOL 4HE NUMBER OF TONES IN THE SYSTEM IS - AND THE LENGTH OF THE CYCLIC PREX IS + SAMPLES
&IGURE "ASE BAND MODEL OF AN /&$- SYSTEM #0 DENOTES THE CYCLIC PREX 5NDER THESE ASSUMPTIONS WE CAN DESCRIBE THE SYSTEM AS A SET OF PARALLEL 'AUSSIAN CHANNELS SHOWN IN &IGURE WITH CORRELATED ATTENUATIONS G 4HE ATTENUATIONS ON
J
&IGURE 4HE /&$- SYSTEM DESCRIBED AS A SET OF PARALLEL 'AUSSIAN CHANNELS WITH CORRELATED ATTENUATIONS EACH TONE ARE GIVEN BY G &
J
J -3
u J - `
R
WHERE & a IS THE FREQUENCY RESPONSE OF THE CHANNEL F S ~ DURING THE /&$- SYMBOL AND 3 IS THE SAMPLING PERIOD OF THE SYSTEM )N MATRIX NOTATION WE DESCRIBE THE /&$SYSTEM AS
R
X 7G
M

WHERE X IS THE RECEIVED VECTOR 7 IS A MATRIX CONTAINING THE TRANSMITTED SIGNALLING POINTS ON ITS DIAGONAL G IS A CHANNEL ATTENUATION VECTOR AND M IS A VECTOR OF IID COMPLEX ZERO MEAN 'AUSSIAN NOISE WITH VARIANCE }
M

#HANNEL MODEL
,
7E ARE USING A FADING MULTI PATH CHANNEL MODEL ;= CONSISTING OF , IMPULSES F ~
J
8
J (
m p ~ `~ 3
J J R

WHERE m ARE ZERO MEAN COMPLEX 'AUSSIAN RANDOM VARIABLES WITH A POWER DELAY PROLE t ~ )N THIS PAPER WE HAVE USED , IMPULSES AND TWO VERSIONS OF THIS CHANNEL MODEL
J
q 3YNCHRONIZED CHANNEL 4HIS IS A MODEL OF A PERFECTLY TIME SYNCHRONIZED /&$SYSTEM WHERE THE RST FADING IMPULSE ALWAYS HAS A ZERO DELAY ~ AND OTHER FADING IMPULSES HAVE DELAYS THAT ARE UNIFORMLY AND INDEPENDENTLY DISTRIBUTED OVER THE LENGTH OF THE CYCLIC PREX 4HE IMPULSE POWER DELAY PROLE t ~ "D DECAYS EXPONENTIALLY ;=
J ~ J ~ QLR
q 5NIFORM CHANNEL !LL IMPULSES HAVE THE SAME AVERAGE POWER AND THEIR DELAYS ARE UNIFORMLY AND INDEPENDENTLY DISTRIBUTED OVER THE LENGTH OF THE CYCLIC PREX

3CENARIO
/UR SCENARIO CONSISTS OF A WIRELESS 1!- /&$- SYSTEM DESIGNED FOR AN OUTDOOR ENVIRONMENT THAT IS CAPABLE OF CARRYING DIGITAL VIDEO 4HE SYSTEM OPERATES AT K(Z BANDWIDTH AND IS DIVIDED INTO TONES WITH A TOTAL SYMBOL PERIOD OF xS OF WHICH xS IS THE CYCLIC PREX /NE /&$- SYMBOL THUS CONSISTS OF SAMPLES -
+ FOUR OF WHICH ARE CONTAINED IN THE CYCLIC PREX + 4HE UNCODED DATA RATE OF THE SYSTEM IS -"ITSEC 7E ASSUME THAT ~ SAMPLE FOR THE SYNCHRONIZED CHANNEL
QLR

,INEAR CHANNEL ESTIMATION ACROSS TONES
)N THE FOLLOWING WE PRESENT THE ,--3% ESTIMATE OF THE CHANNEL ATTENUATIONS G FROM THE RECEIVED VECTOR X AND THE TRANSMITTED DATA 7 7E ASSUME THAT THE RECEIVED /&$SYMBOL CONTAINS DATA KNOWN TO THE ESTIMATOR EITHER TRAINING DATA OR RECEIVER DECISIONS 4HE COMPLEXITY REDUCTION OF THE ,--3% ESTIMATOR CONSISTS OF TWO SEPARATE STEPS )N THE RST STEP WE MODIFY THE ,--3% BY AVERAGING OVER THE TRANSMITTED DATA OBTAINING A SIMPLIED ESTIMATOR )N THE SECOND STEP WE REDUCE THE NUMBER OF MULTIPLICATIONS REQUIRED BY APPLYING THE THEORY OF OPTIMAL RANK REDUCTION ;=

,--3% ESTIMATION
4HE ,--3% ESTIMATE OF THE CHANNEL ATTENUATIONS G IN GIVEN THE RECEIVED DATA X AND THE TRANSMITTED SYMBOLS 7 IS ;= B G
KLLRD
1
GG
GG
` a s B
} 77 G
' M
KR

WHERE B G
KR
36$ BASED CHANNEL ESTIMATORS 7 X

M
X X X aaa W W W

IS THE LEAST SQUARES h i ,3 ESTIMATE OF G } IS THE VARIANCE OF THE ADDITIVE CHANNEL NOISE AND 1 $ GG IS THE CHANNEL AUTOCORRELATION 4HE SUPERSCRIPT a DENOTES (ERMITIAN TRANSPOSE )N THE FOLLOWING WE ASSUME WITHOUT LOSS OF GENERALITY THAT THE VARIANCES OF i h THE CHANNEL ATTENUATIONS IN G ARE NORMALIZED TO UNITY IE $ JG J 4HE ,--3% ESTIMATOR IS OF CONSIDERABLE COMPLEXITY SINCE A MATRIX INVERSION IS NEEDED EVERY TIME THE TRAINING DATA IN 7 CHANGES 7E REDUCE THE COMPLEXITY OF THIS ESTIMATOR BY AVERAGING OVER THE h TRANSMITTED i DATA ;= IE WE REPLACE THE TERM 77 IN WITH ITS EXPECTATION $ 77 !SSUMING THE SAME SIGNAL CONSTELLATION h i ON ALL TONES AND EQUAL PROBABILITY ON ALL CONSTELLATION POINTS WE HAVE $ 77 $ FJW J G ( WHERE ( IS THE IDENTITY MATRIX $ENING THE AVERAGE SIGNAL TO NOISE RATIO AS 2-1 $ FJW J G } WE OBTAIN A SIMPLIED ESTIMATOR
' ' GG J ' ' ' J J M
B1 G WHERE
t 1
GG
GG
n (
2-1
u
B G
KR

h i h i n $ JW J $ JW J
J J
IS A CONSTANT DEPENDING ON THE SIGNAL CONSTELLATION )N THE CASE OF 1!- TRANSMISSION n "ECAUSE 7 IS NO LONGER A FACTOR IN THE MATRIX CALCULATION THE INVERSION OF 1
PEI ( DOES NOT NEED TO BE CALCULATED EACH TIME THE TRANSMITTED DATA IN 7 CHANGES &URTHERMORE IF 1 AND 2-1 ARE KNOWN BEFOREHAND OR ARE SET TO XED NOMINAL VALUES THE MATRIX 1 1
PEI ( NEEDS TO BE CALCULATED ONLY ONCE 5NDER THESE CONDITIONS THE ESTIMATION REQUIRES - MULTIPLICATIONS PER TONE 4O FURTHER REDUCE THE COMPLEXITY OF THE ESTIMATOR WE PROCEED WITH LOW RANK APPROXIMATIONS BELOW
n GG GG n GG GG

/PTIMAL LOW RANK APPROXIMATIONS
/PTIMAL RANK REDUCTION IS ACHIEVED BY USING THE SINGULAR VALUE DECOMPOSITION 36$ ;= 4HE 36$ OF THE CHANNEL AUTOCOVARIANCE MATRIX IS 1
GG
4c4
'

WHERE 4 IS A UNITARY MATRIX CONTAINING THE EIGENVECTORS AND c IS A DIAGONAL MATRIX CONTAINING THE SINGULAR VALUES w w w w w w ON ITS DIAGONAL )N !PPENDIX ! IT IS SHOWN THAT THE OPTIMAL RANK O ESTIMATOR IS
-
B 4aO 4 G B G
' O
KR

WE ARE DEALING WITH (ERMITIAN MATRICES THE wK S ARE ALSO EIGENVALUES (OWEVER WE USE THE TERMINOLOGY OF THE 36$ SINCE IT IS MORE GENERAL AND CAN BE USED IN OPTIMAL RANK REDUCTION OF NON (ERMITIAN MATRICES
3INCE
36$ BASED CHANNEL ESTIMATORS WHERE a IS A DIAGONAL MATRIX WITH THE VALUES
O

p
J
wJ wJ
2-1
J O ` J O - `

6IEWING THE ORTHONORMAL MATRIX 4 AS A TRANSFORM THE SINGULAR VALUE w OF 1 IS THE CHANNEL POWER VARIANCE CONTAINED IN THE J TRANSFORM COEbCIENT AFTER TRANSFORMING B 3INCE 4 IS UNITARY THIS TRANSFORMATION CAN BE VIEWED AS ROTATING THE THE ,3 ESTIMATE G B VECTOR G SO THAT ALL ITS COMPONENTS ARE UNCORRELATED ;= 4HE DIMENSION OF THE SPACE OF ESSENTIALLY TIME AND BAND LIMITED SIGNALS LEADS US TO THE RANK NEEDED IN THE LOW RANK ESTIMATOR )N ;= IT IS SHOWN THAT THIS DIMENSION IS ABOUT !3
WHERE ! IS THE ONE SIDED BANDWIDTH AND 3 IS THE TIME INTERVAL OF THE SIGNAL !CCORDINGLY THE MAGNITUDE OF THE SINGULAR VALUES OF 1 SHOULD DROP RAPIDLY AFTER ABOUT +
LARGE VALUES WHERE + IS THE LENGTH OF THE CYCLIC PREX ! 3 3 +3 AND !3
+
7E PRESENT THE CHANNEL POWER CONTAINED IN THE RST COEbCIENTS IN &IGURE 4HE CALCULATIONS ARE BASED ON OUR SCENARIO AND THE TWO CHANNEL MODELS THE SYNCHRONIZED AND THE UNIFORM 4HE MAGNITUDE OF THE CHANNEL POWER DROPS RAPIDLY AFTER ABOUT J IE COEbCIENTS WHICH IS CONSISTENT WITH THE OBSERVATION THAT THE DIMENSION OF THE SPACE SPANNED BY 1 IS APPROXIMATELY +
THAT IS
IN THIS CASE
' J GG SG KR KR GG R R GG
&IGURE 2ELATIVE CHANNEL POWER w %JG J IN THE TRANSFORM COEbCIENTS FOR THE TWO EXAMPLE CHANNELS
J J
! BLOCK DIAGRAM OF THE RANKÒ ESTIMATOR IN IS SHOWN IN &IGURE WHERE THE ,3 ESTIMATE IS CALCULATED FROM X BY MULTIPLYING BY 7
TRANSFORM IN THIS SPECIAL CASE OF LOW RANK APPROXIMATION IS THE +ARHUNEN ,OEVE AKA (OTELLING TRANSFORM OF H
4HE

&IGURE "LOCK DIAGRAM OF THE RANKÒ CHANNEL ESTIMATOR

%STIMATOR COMPLEXITY
4HE LIMITING FACTOR OF THE RANKÒ ESTIMATORS IS AN ERROR OOR SEE 3ECTION 4O ELIMINATE THIS ERROR OOR UP TO A GIVEN 3.2 WE NEED TO MAKE SURE OUR ESTIMATOR RANK IS LARGE ENOUGH 4HIS PROMPTS AN ANALYSIS OF THE COMPUTATIONAL COMPLEXITY OF THE RANKÒ ESTIMATOR 4HE IMPLEMENTATION WE HAVE CHOSEN IS BASED ON WRITING AS A SUM OF RANK MATRICES WHICH GIVES US THE EXPRESSION $ % 8 8 B B B G G p T T P T G
O O ' J O J J KR J J KR
$ % B B IS THE %UCLIDIAN INNER PRODUCT 4HE LINEAR COM WHERE P p T AND T G T G BINATION OF O VECTORS OF LENGTH - ALSO REQUIRES O- MULTIPLICATIONS 4HE ESTIMATION THUS REQUIRES O- MULTIPLICATIONS AND THE TOTAL NUMBER OF MULTIPLICATIONS PER TONE BECOMES O )N COMPARISON WITH THE FULL ESTIMATOR WE HAVE MANAGED TO REDUCE THE NUMBER OF MULTIPLICATIONS FROM - TO O PER TONE 4HE SMALLER O IS THE LOWER THE COMPUTATIONAL COMPLEXITY BUT THE LARGER THE APPROXIMATION ERROR BECOMES &OLLOWING THE ANALYSIS IN 3ECTION WE CAN EXPECT A GOOD APPROXIMATION WHEN O IS IN THE RANGE OF SAMPLES IN THE CYCLIC PREX WHICH IS USUALLY MUCH SMALLER THAN THE NUMBER OF TONES - ! LEGITIMATE QUESTION AT THIS POINT IS WHAT HAPPENS FOR A SYSTEM WITH MANY TONES AND MANY SAMPLES IN THE CYCLIC PREX 4HE NUMBER OF CALCULATIONS PER TONE CAN BE CONSIDERABLE IF A RANKÒ ESTIMATOR IS USED DIRECTLY ON ALL TONES IN THE SYSTEM /NE SOLUTION TO THIS PROBLEM IS A PARTITIONING OF THE TONES INTO REASONABLE SIZED BLOCKS AND AT A CERTAIN PERFORMANCE LOSS PERFORM THE ESTIMATION INDEPENDENTLY IN THESE BLOCKS "Y DIVIDING THE CHANNEL ATTENUATIONS INTO * EQUALLY SIZED BLOCKS THE BANDWIDTH IN EACH BLOCK IS REDUCED BY A FACTOR * 2EFERRING AGAIN TO THE DIMENSION OF THE SPACE OF ESSENTIALLY TIME AND BANDLIMITED SIGNALS ;= THE EXPECTED NUMBER OF ESSENTIAL BASE VECTORS IS REDUCED FROM +
TO +*
(ENCE THE COMPLEXITY OF THE ESTIMATOR DECREASES ACCORDINGLY 4O ILLUSTRATE THE IDEA LET US ASSUME WE HAVE A SYSTEM WITH - TONES AND A + SAMPLE CYCLIC PREX 4HE UNIFORM CHANNEL CORRELATION BETWEEN THE ATTENUATIONS G AND G IN THIS SYSTEM IS SEE !PPENDIX "
' J J J J J KR KR L M
J (
J (
LM

Ì {+
L`M -
IF L M IF L M
I {+
L`M -

4HIS ONLY DEPENDS ON THE DISTANCE BETWEEN THE TONES L ` M AND THE RATIO BETWEEN THE LENGTH OF THE CYCLIC PREX AND THE NUMBER OF TONES +- 4HE TONE SYSTEM CAN BE DESCRIBED BY X 7 G M
! ! ! ! X 7 G M THAT IS AS PARALLEL TONE SYSTEMS X 7 G
M J
J J J J
7E HAVE THE SAME CHANNEL CORRELATION IN EACH SUBSYSTEM AS WE HAVE IN THE TONE SCENARIO IN THIS PAPER +- "Y ESTIMATING THE CHANNEL ATTENUATIONS G IN EACH SUB SYSTEM INDEPENDENTLY WE NEGLECT THE CORRELATION BETWEEN TONES IN DIdERENT SUB SYSTEMS BUT OBTAIN THE SAME -3% PERFORMANCE AS IN OUR TONE SCENARIO
J

%STIMATOR PERFORMANCE AND DESIGN
7E PROPOSE A GENERIC LOW RANK FREQUENCY BASED CHANNEL ESTIMATOR IE THE ESTIMATOR IS DESIGNED FOR XED NOMINAL VALUES OF 3.2 AND CHANNEL CORRELATION (ENCE WE NEED TO ANALYSE HOW THE RANK CHANNEL CORRELATION AND 3.2 SHOULD BE CHOSEN FOR THIS ESTIMATOR SO THAT IT IS ROBUST TO VARIATIONS IN THE CHANNEL STATISTICS IE MISMATCH !S A PERFORMANCE MEASURE WE USE UNCODED SYMBOL ERROR RATE 3%2 FOR 1!- SIGNALLING 4HE 3%2 IN THIS CASE CAN BE CALCULATED FROM THE MEAN SQUARED ERROR -3% WITH THE FORMULAE IN ;=

h i 4HE MEAN SQUARED ERROR RELATIVE TO THE CHANNEL POWER $ JG J OF THE RANKÒ ESTIMATOR IS MAINLY DETERMINED BY THE CHANNEL POWER CONTAINED IN THE TRANSFORM COEbCIENTS AND CAN BE EXPRESSED SEE !PPENDIX #
J
2ANK REDUCTION
t 8 LRD O w `p - (
O J J J J

J
n
p 2-1
8
w - (
J O

WHERE w AND p ARE GIVEN BY AND RESPECTIVELY 4HE -3% IS A MONOTON ICALLY DECREASING FUNCTION OF 2-1 AND CAN BE BOUNDED FROM BELOW BY THE LAST TERM LRD O
8 w v LRD O - (
J J O

WHICH IS THE SUM OF THE CHANNEL POWER IN THE TRANSFORM COEbCIENTS NOT USED IN THE ESTIMATE 4HIS -3% OOR LRD O WILL GIVE RISE TO A ERROR OOR IN THE SYMBOL ERROR RATES

4HE ERROR OOR IS THE MAIN LIMITATION ON THE COMPLEXITY REDUCTION ACHIEVED BY OPTIMAL RANK REDUCTION !S AN ILLUSTRATION &IGURE DISPLAYS THE 3%2 RELATIVE TO THE CHANNEL VARIANCE FOR THREE DIdERENT RANKS AS A FUNCTION OF THE 3.2 4HE RANKS CHOSEN ARE O AND AND THE CHANNEL USED IN THE EXAMPLE IS THE SYNCHRONIZED CHANNEL 4HE CORRESPONDING 3%2 OORS ARE SHOWN AS HORIZONTAL LINES &OR O THE 3%2 OOR IS RELATIVELY SMALL AND THE 3%2 OF THE RANK` ESTIMATOR IS COMPARABLE TO THE ORIGINAL FULL RANK ESTIMATOR IN THE RANGE TO D" IN 3.2 "Y CHOOSING THE APPROPRIATE RANK ON THE ESTIMATOR WE CAN ESSENTIALLY AVOID THE IMPACT FROM THE 3%2 OOR UP TO A GIVEN 3.2 7HEN WE HAVE FULL RANK O - NO 3%2 OOR EXISTS
&IGURE ,OW RANK ESTIMATOR SYMBOL ERROR RATE AS A FUNCTION OF 3.2 WITH RANKS O AND #ORRESPONDING 3%2 OORS SHOWN AS HORIZONTAL LINES 3YNCHRONIZED CHANNEL "ASED ON THE CHANNEL POWERS PRESENTED IN &IGURE WE SHOW THE CORRESPONDING 3%2 OORS RELATIVE TO THE CHANNEL VARIANCE IN &IGURE !FTER ABOUT RANK` THE 3%2 OOR DECREASES RAPIDLY 7E ARE THEREFORE ABLE TO OBTAIN A GOOD ESTIMATOR APPROXIMATION WITH A RELATIVELY LOW RANK

3%2 PERFORMANCE UNDER MISMATCH
)N PRACTICE THE TRUE CHANNEL CORRELATION AND 3.2 ARE NOT KNOWN 4O GET A GENERAL EXPRESSION FOR THE ESTIMATOR 3%2 WE DERIVE IT UNDER THE ASSUMPTION THAT THE ESTIMATOR IS DESIGNED FOR CORRELATION 1 AND SIGNAL TO NOISE RATIO 2-1 BUT THE TRUE VALUES ARE E DENOTES A CHANNEL WITH DIdERENT STATISTICS THAN = RESPECTIVELY WHERE G 1E E AND 2-1 G 4HIS ALLOWS US TO ANALYSE THIS ESTIMATORS SENSITIVITY TO DESIGN ERRORS 5NDER THESE ASSUMPTIONS THE RELATIVE -3% OF THE RANKÒ ESTIMATE BECOMES SEE !PPENDIX # w v 8 n 8 LRD O x `p
p
x - ( - ( = 2-1
GG GG O J J J J J J O

&IGURE %STIMATOR 3%2 OOR AS A FUNCTION OF ESTIMATOR RANK #IRCLES SHOW THE 3%2 OORS APPEARING IN &IGURE WHERE x IS THE J DIAGONAL ELEMENT OF 4 E VARIANCE OF THE TRANSFORMED CHANNEL 4 G |r sr E $ 4 G 4
SG J ' ' '
'
1E E 4 CF )T CAN BE INTERPRETED AS THE UNDER CORRELATION MISMATCH SINCE s } E G 4 1E E 4

GG ' ' GG
E ARE NO LONGER UNCORRELATED (OWEVER DUE TO )T SHOULD BE NOTED THAT THE ELEMENTS OF 4 G THE FACT THAT THE POWER DELAY PROLE IS SHORT COMPARED TO THE /&$- SYMBOL THE RST O ELEMENTS CAN BE EXPECTED TO CONTAIN MOST OF THE POWER 4HIS PROPERTY WILL ENSURE ONLY A SMALL PERFORMANCE LOSS WHEN THE ESTIMATOR IS DESIGNED FOR WRONG CHANNEL STATISTICS )F RANKÒ ESTIMATORS ARE USED IN A REAL SYSTEM THE SENSITIVITY TO MISMATCH IN BOTH CHANNEL CORRELATION AND 3.2 ARE IMPORTANT 7E WILL SHOW THAT A RANKÒ ESTIMATOR BASED ON THE UNIFORM CHANNEL MODEL AND A NOMINAL 3.2 CAN BE USED AS XED GENERIC ESTIMATOR WITH ONLY A SMALL LOSS IN AVERAGE PERFORMANCE 7E DIVIDE THE MISMATCH ANALYSIS INTO TWO PARTS RST WE ANALYSE THE 3%2 WHEN WE HAVE A MISMATCH IN CHANNEL CORRELATION AND LATER WE ANALYSE THE 3%2 WHEN WE HAVE A MISMATCH IN 3.2
'

)NCORRECT CHANNEL CORRELATION
= BUT INCORRECT CHANNEL CORRELATION &ROM WITH NO 3.2 MISMATCH 2-1 2-1 1 1E E WE OBTAIN THE PERFORMANCE FOR THE CORRELATION MISMATCH CASES 7E COM PARED THE PERFORMANCE OF OUR CHANNEL ESTIMATOR IN TWO MISMATCH SITUATIONS I USING THE A UNIFORM CHANNEL WHEN THE TRUE CHANNEL MODEL WAS THE SYNCHRONIZED CHANNEL AND II USING THE SYNCHRONIZED CHANNEL WHEN THE TRUE CHANNEL MODEL WAS THE UNIFORM CHANNEL 4HE RESULTING CHANNEL ESTIMATES THAT WERE USED IN THE DETECTION OF THE DATA PRODUCED NO NOTICABLE DIdERENCE IN SYMBOL ERROR RATES LESS THAN D" CHANGE IN EdECTIVE 3.2
GG GG

FOR AN AVERAGE 3.2 UP TO D" 4O SHOW THE DIdERENCES MORE CLEARLY THE -3%S ARE DISPLAYED IN &IGURE &OR THE MISMATCHED CASES MARKED WITH q THE UNIFORM DESIGN IS MORE ROBUST IE THE ERROR IN CASE OF MISMATCH IS LOWER 7ITH THE RESTRICTION THAT THE TRUE CHANNEL HAS A POWER DELAY PROLE SHORTER THAN THE CYCLIC PREX DESIGNING FOR A UNIFORM POWER DELAY PROLE CAN BE SEEN AS A MINIMAX DESIGN
&IGURE -3% FOR CORRECT AND MISMATCHED DESIGN 4HE LATTER IS MARKED WITH DOTS q

)NCORRECT 3.2
&INALLY WE EVALUATE THE SENSITIVITY TO MISMATCH IN DESIGN 3.2 FOR A RANK ESTIMATOR 7HEN THERE IS NO MISMATCH IN CHANNEL CORRELATION AND NOMINAL 3.2S OF AND D" ARE USED IN THE DESIGN THE SENSITIVITY TO 3.2 MISMATCH IS NOT THAT LARGE (OWEVER IN &IGURE WE PRESENT THE 3%2 FOR THE SAME RANK ESTIMATORS BUT WITH THE DIdERENCE THAT THE TRUE CHANNEL CORRELATION IS MISMATCHED WITH THE DESIGN CORRELATION )N THIS SECOND CASE THERE IS A CLEAR DIdERENCE BETWEEN THE TWO DESIGNS THE HIGHER THE NOMINAL DESIGN 3.2 THE BETTER THE OVERALL PERFORMANCE OF THE ESTIMATOR IN THE RANGE TO D" IN 3.2 )T SHOULD BE NOTED THAT A ,--3% ESTIMATOR DESIGNED FOR A LARGE 3.2 APPROACHES THE ,3 ESTIMATOR

'ENERIC LOW RANK ESTIMATOR
)F WE WANT A ROBUST GENERIC CHANNEL ESTIMATOR DESIGN FOR /&$- SYSTEMS OF THE LOW RANK TYPE THE ANALYSIS IN THE PREVIOUS SECTION SUGGESTS THE USE OF THE UNIFORM CHANNEL CORRELATION AND A RELATIVELY HIGH 3.2 AS NOMINAL DESIGN PARAMETERS 4HE DESIGN OF SUCH AN ESTIMATOR ONLY REQUIRES KNOWLEDGE ABOUT THE LENGTH OF THE CYCLIC PREX THE NUMBER OF TONES IN THE SYSTEM AND THE TARGET RANGE OF 3.2S FOR THE APPLICATION )F THE RECEIVER

&IGURE 2ANK ESTIMATOR 3%2 WHEN 3.2S OF AND D" ARE USED IN THE DESIGN 4HE ESTIMATORS ARE DESIGNED FOR INCORRECT CHANNEL CORRELATION CANNOT AdORD AN ESTIMATOR THAT INCLUDES TRACKING OF CHANNEL CORRELATION AND 3.2 THIS CHANNEL ESTIMATOR WORKS REASONABLY WELL FOR XED 3.2 AND CHANNEL CORRELATION

0ERFORMANCE GAIN
&OR THE SCENARIO USED IN THIS PAPER 3EC WE CHOOSE A RANK ESTIMATOR WITH UNIFORM DESIGN AND 2-1 D" 4HE PERFORMANCE OF THIS ESTIMATOR IS PRESENTED IN &IG WHERE THE 3%2 FOR THE ,3 ESTIMATE AND KNOWN CHANNEL ARE ALSO SHOWN !S CAN BE SEEN THE LOW RANK ESTIMATOR IS D" BETTER THAN THE ,3 ESTIMATOR AND LESS THAN D" FROM THE KNOWN CHANNEL

#OMPARISON TO &)2 LTERS
!N ALTERNATIVE TO USING LOW RANK ESTIMATORS TO SMOOTH THE CHANNEL ESTIMATES IS TO USE A &)2 LTER INSTEAD (ENCE WE WILL COMPARE OUR PROPOSED LOW RANK ESTIMATORS TO &)2 LTERS OF THE SAME COMPLEXITY 4HE &)2 LTERS ARE O TAPS 7IENER LTERS ;= IE O MULTIPLICATIONS PER TONE THAT ARE DESIGNED FOR THE SAME CHANNEL CORRELATION AND 3.2 AS THE LOW RANK ESTIMATORS &IGURE SHOWS THE 3%2 FOR RANKÒ ESTIMATORS IN COMPARISON WITH &)2 LTERS OF THE SAME COMPUTATIONAL COMPLEXITY 7HEN THE COMPLEXITY IS MULTIPLICATIONS PER TONE ! THE RANKÒ ESTIMATOR HAS ABOUT D" ADVANTAGE IN 3.2 OVER THE &)2 LTER IN THE RANGE OF 3.2S SHOWN 7HEN THE NUMBER OF CALCULATIONS GOES DOWN TO MULTIPLICATIONS PER TONE " THE 3%2 OOR OF THE RANKÒ ESTIMATOR BECOMES VISIBLE AND THE &)2 LTER PERFORMS BETTER AT 3.2S ABOVE D" (OWEVER IT SHOULD BE NOTED THAT THE PERFORMANCE OF THE LOW RANK ESTIMATORS DEPEND HEAVILY OF THE SIZE OF THE CYCLIC PREX )F THE CYCLIC PREX WERE TO BE DECREASED RELATIVE TO

&IGURE 3%2 FOR 1!- TRAINING DATA AND A SYNCHRONIZED CHANNEL 4HE GENERIC RANK` ESTIMATOR DESIGNED FOR A UNIFORM CHANNEL AND D" IN 3.2 IS COMPARED TO THE ,3 ESTIMATOR AND KNOWN CHANNEL AT THE RECEIVER THE /&$- SYMBOL THE LOW RANK ESTIMATOR WOULD INCREASE ITS PERFORMANCE 4HIS IS DUE TO THE FACT THAT THE DIMENSION OF THE CHANNEL WHOSE DURATION IS ASSUMED TO BE SHORTER THAN THE CYCLIC PREX DECREASES AND CAN THUS BE REPRESENTED WITH FEWER COEbCIENTS /N THE OTHER HAND IF THE CYCLIC PREX INCREASES IN SIZE MORE COEbCIENTS ARE NEEDED TO AVOID LARGE APPROXIMATION ERRORS (ENCE WHETHER OR NOT THE LOW RANK ESTIMATOR IS BETTER THAN THE &)2 LTER DEPENDS ON THE RELATIVE SIZE OF THE CYCLIC PREX AND THE ALLOWED COMPLEXITY

4HE USE OF TIME CORRELATION
4HE LOW RANK ESTIMATOR PRESENTED IN THIS PAPER IS BASED ON FREQUENCY CORRELATION ONLY BUT THE TIME CORRELATION OF THE CHANNEL CAN ALSO BE USED 4HE TWO DIMENSIONAL ,--3% ESTIMATOR CAN BE SIMPLIED USING THE SAME TECHNIQUE WITH RANK REDUCTION AS DESCRIBED HERE (OWEVER IN ;= IT IS SHOWN THAT SUCH AN ESTIMATOR GIVES AN INFERIOR PERFORMANCE FOR A XED COMPLEXITY (ENCE IT SEEMS THAT SEPARATING THE USE OF FREQUENCY AND TIME CORRELATION IS THE MOST EbCIENT WAY OF ESTIMATING THE CHANNEL /THER APPROACHES TO USE THE TIME CORRELATION IS EG TO USE A DECISION DIRECTED SCHEME ;= OR &)2 LTERS ; = 4HE FORMER CAN BE USED IN A SLOW FADING ENVIRONMENT WHERE IT OdERS GOOD PERFORMANCE FOR A MINIMAL COMPLEXITY AND THE LATTER IS PREFERRED IN CASE OF FAST FADING )T IS POSSIBLE TO USE A BANK OF &)2 LTERS AND CHOOSE THE MOST APPROPRIATE ACCORDING THE ESTIMATED $OPPLER FREQUENCY ;=

&IGURE 3%2 COMPARISON BETWEEN THE RANKÒ ESTIMATORS AND &)2 7IENER LTERS OF THE SAME COMPLEXITY "OTH ESTIMATORS ARE DESIGNED FOR THE UNIFORM CHANNEL AND D" 3.2 ! MULTIPLICATIONS PER TONE AND " MULTIPLICATIONS PER TONE

#ONCLUSIONS
7E HAVE INVESTIGATED LOW COMPLEXITY LOW RANK APPROXIMATIONS OF THE ,--3% CHANNEL ESTIMATOR FOR NON SAMPLE SPACED CHANNELS 4HE INVESTIGATION SHOWS THAT AN ESTIMATOR ERROR OOR INHERENT IN THE LOW RANK APPROXIMATION IS THE SIGNICANT LIMITATION TO THE ACHIEVED COMPLEXITY REDUCTION 7E SHOWED THAT A GENERIC LOW RANK ESTIMATOR DESIGN BASED ON THE UNIFORM CHANNEL CORRELATION AND A NOMINAL 3.2 CAN BE USED IN OUR TONE SCENARIO #OMPARED WITH THE FULL ,--3% THERE IS ONLY A SMALL LOSS IN PERFORMANCE UP TO A 3.2 OF D" BUT A REDUCTION IN COMPLEXITY WITH A FACTOR -O &OR SYSTEMS WITH MORE SUBCHANNELS THIS GAIN IS EVEN LARGER 4HE GENERIC ESTIMATOR DESIGN ONLY REQUIRES KNOWLEDGE ABOUT THE LENGTH OF THE CYCLIC PREX THE NUMBER OF TONES IN THE SYSTEM AND THE TARGET RANGE OF 3.2S FOR THE APPLICATION 7E ALSO COMPARED LOW RANK ESTIMATORS TO &)2 LTERS ACROSS THE TONES 4HE COMPARISON SHOWED THAT AT LOW COMPLEXITIES AND HIGH 3.2S THE &)2 LTERS IS THE PREFERABLE CHOICE DUE TO THE ERROR OOR IN THE LOW RANK APPROXIMATION (OWEVER IF WE CAN ALLOW UP TO MULTIPLICATIONS PER TONE IN OUR SCENARIO THE LOW RANK ESTIMATOR IS MORE ADVANTAGEOUS !LSO THE LOW RANK ESTIMATORS IMPROVE THEIR PERFORMANCE AS THE CYCLIC PREX DECREASES IN SIZE
!
/PTIMAL RANK REDUCTION

' KR
4HE OPTIMAL RANK REDUCTION IS FOUND FROM THE CORRELATION MATRICES N O B 1 B $ GG 1

GGKR GG
1B AND THE 36$ 1

GGKR

B
GKR GKR
N O B B $ G G 1
' KR KR
GG
n ( 2-1
1B

GKR GKR
0 #0
'
WHERE 0 AND 0 ARE UNITARY MATRICES AND # IS A DIAGONAL MATRIX WITH THE SINGULAR VALUES C w C w a a a w C ON ITS DIAGONAL 4HE BEST LOW RANK ESTIMATOR ;= IS THEN
-
B 0 G
O O
#

w 0 1
'
GLS GLS

B G
KR

WHERE # IS THE O b O UPPER LEFT CORNER OF # IE WE EXCLUDE ALL BUT THE O LARGEST SINGULAR VECTORS )N THIS PAPER WE HAVE 1 B 1 AND 1B B 1
PEI ( AND WE NOTE THAT THEY SHARE THE SAME SINGULAR VECTORS IE THE ONES OF 1 4c4 4HUS WE MAY EXPRESS AS
n GGKR GG GKR GKR GG ' GG
t t 4c4 4 c
'
n 4c c
( 2-1
u
u u n ( 4 2-1
' '

'
4 0 #0
t 0 0 4 AND # c c
4HE RANK O ESTIMATOR NOW BECOMES B G v
O
n ( 2-1
u

4 v 4
# #
u u n B 4 4 c
( 4 G 2-1 wt u v w n a B B c
( 4 G 4 4 G 2-1 w t t
' ' KR ' O ' KR
KR
WHERE a IS THE O b O UPPER LEFT CORNER OF

O
n ac c
( 2-1
u
CH@F
w w aaa w
PEI w
PEI
n n -
.OTE THAT 0 0 SINCE WE ARE ESTIMATING THE SAME TONES AS WE ARE OBSERVING IE SMOOTHING AND AN EIGENVALUE DECOMPOSITION COULD BE USED TO ACHIEVE OPTIMAL RANK RE DUCTION )N THE GENERAL CASE WHEN EG PILOT SYMBOL ASSISTED MODULATION ;= IS USED AND THERE ARE KNOWN SYMBOLS PILOTS ON ONLY A PART OF THE SUBCHANNELS WE HAVE 0 0 SINCE 1 B AND 1B B DONT SHARE THE SAME SINGULAR VECTORS THE MATRICES ARE NOT EVEN OF THE SAME SIZE (ENCE THE MORE GENERAL 36$ MUST BE USED WHICH MOTIVATES THE NOMENCLATURE IN THIS ARTICLE
GGKR GKR GKR

"
#HANNEL CORRELATION MATRICES

,
5SING THE CHANNEL MODEL IN THE ATTENUATION ON TONE J BECOMES G

J
8
H(
m D
H
I {
J ~ - H
AND THE CORRELATION MATRIX FOR THE ATTENUATION VECTOR G h i 1GG $ GG :Q <
' LM
CAN BE EXPRESSED AS ~ S INDEPENDENT : : 9 aaa E ~ Q

J , LM ~J ,
, J
8
H(
t ~ D
H
L`M -
I {~ H
L`M -
C~ C~
: 8
H(
J(
~H
~ t ~ D
H H
I {~ H
C~
H ~J J

WHERE t ~ IS THE MULTI PATH INTENSITY PROLE AND E ~ IS THE PROBABILITY DENSITY FUNC TION OF ~ 4HE CORRELATION MATRICES OF THE THREE CHANNELS USED IN THIS PAPER ARE CALCULATED BELOW
J
q 3YNCHRONIZED CHANNEL 4HE PROBABILITY DISTRIBUTIONS FOR THE DELAYS ARE E

~
~ ~
H
~H
p ~ | + IF ~ : +< H , OTHERWISE

H
AND THE POWER DELAY PROLE IS t ~ NORMALIZING Q TO UNITY GIVES US

JJ
" D r
~ ~ QLR
3UBSTITUTING IN AND

{I
L`M -
LM
L`M I { ~ QLR
` D r +
, ` ~ ` D

,
s
~ QLR
QLR
+ ~ QLR
q 5NIFORM CHANNEL 4HE PROBABILITY DISTRIBUTIONS FOR THE DELAYS ARE | + IF ~ : +< E ~ H , OTHERWISE
H ~H H
AND THE POWER DELAY PROLE IS CONSTANT t ~ NORMALIZING Q TO UNITY GIVES US

JJ
" 3UBSTITUTED IN AND IF L M IF L M
LM

Ì {+
L`M -
I {+
L`M -

#
%STIMATOR MEAN SQUARED ERROR
)N THIS APPENDIX WE DERIVE THE -3% OF THE RANKÒ ESTIMATOR IN 7E ALSO PRESENT THE -3% OOR WHICH BOUNDS THE ACHIEVABLE -3% FROM BELOW IN LOW RANK APPROXIMATIONS OF THE ,--3% ESTIMATOR 4O GET A GENERAL EXPRESSION FOR THE MEAN SQUARED ERROR FOR THE RANKÒ APPROXIMATION OF THE ,--3% ESTIMATOR WE ASSUME THAT THE ESTIMATOR HAS BEEN DESIGNED FOR CHANNEL CORRELATION 1 AND SIGNAL TO NOISE RATIO 2-1 BUT THE REAL CHANNEL E HAS THE CORRELATION 1 E AND THE REAL SIGNAL TO NOISE RATIO IS 2-1 = &ROM AND G E
M B G E WHERE THE NOISE TERM M E 7 M HAS THE AUTOCOVARIANCE MATRIX WE HAVE G E B 1E E PEI ( 4HE ESTIMATION ERROR D G ` G OF THE RANKÒ ESTIMATOR IS wu w t v v a a ' E`4 E D 4 (` 4 G 4' M
GG GG KR n MM O O O O O
AND THE MEAN SQUARED ERROR IS LRD O 4O SIMPLIFY THE EXPRESSION WE USE THAT E AND M E ARE UNCORRELATED HENCE THE CROSS TERMS ARE CANCELLED IN THE EXPECTATION q G ` a q 3Q@BD 4 4 3Q@BD IF 4 IS A UNITARY MATRIX AND 3Q@BD
! 3Q@BD
3Q@BD ! ;= 0 q 3Q@BD # # @ C WHEN # IS A DIAGONAL MATRIX WITH THE ELEMENTS C ON ITS DIAGONAL AND NOT NECESSARILY A DIAGONAL MATRIX HAS DIAGONAL ELEMENTS @
' J JJ J J JJ
h i 3Q@BD $ D D ' O O

5SING IN THE MEAN SQUARED ERROR BECOMES LRD O t v wu t v a a ' 3Q@BD 4 ( ` 4 1G EG E4 ( ` v w v w a a ' 4 4 1M 4' EM E4 8 8 n 8 x `p
x
p = 2-1 ( ( ( u t 8 8 n x `p
p
x - ( - ( = 2-1
O ' O O O O J J J J J J O J O J J J J J J O ' O

wu
'
'

WHERE x IS THE CHANNEL POWER IN THE J TH TRANSFORM COEbCIENT IE THE JTH DIAGONAL E 4 4HE -3% CAN BE LOWER BOUNDED LRD O w LRD O BY ELEMENT OF THE MATRIX 4 1 WHAT WE CALL THE -3% OOR 8 LRD O x - (
J GG J J O

' J GG

` a )F THERE IS NO MISMATCH IN 2-1 OR CHANNEL CORRELATION WE HAVE x CH@F 4 1 4 = 2-1 AND THE -3% BECOMES w AND 2-1
J
t 8 w `p LRD O - (
O J J

J
n
p 2-1
8 w - (
J J O

0ART ! COMPARATIVE STUDY OF PILOT BASED CHANNEL ESTIMATORS FOR WIRELESS /&$!BSTRACT 4HIS REPORT DEALS WITH PILOT BASED CHANNEL ESTIMATION IN WIRELESS /&$- SYS TEMS 7E ASSUME THAT THE RECEIVER IS ABLE TO USE ALL TRANSMITTED PILOTS WHICH IS THE CASE IN BROADCASTING AND IN THE DOWNLINK OF A MULTIUSER SYSTEM &OUR ESTIMATORS OF WHICH TWO HAVE BEEN PROPOSED IN THE LITERATURE ARE COMPARED BOTH IN TERMS OF MEAN SQUARED ERROR AND BIT ERROR RATES )N THE LATTER CASE WE SIMULATE A MULTIUSER SYSTEM WHICH IN CORPORATES CHANNEL CODING 4HE CHANNEL ESTIMATION IS A TWO DIMENSIONAL PROBLEM TIME AND FREQUENCY AND BOTH SEPARABLE AND NON SEPARABLE ESTIMATORS ARE INVESTIGATED 7E DE SIGN LOW RANK APPROXIMATIONS OF THESE ESTIMATORS AND COMPARE THE PERFORMANCE AT GIVEN COMPLEXITIES 4HE COMPARISON SHOWS THAT THE USE OF SEPARABLE ESTIMATORS INCREASES THE PERFORMANCE SUBSTANTIALLY COMPARED TO NON SEPARABLE ESTIMATORS WITH THE SAME COMPLEX ITY &OR THE SCENARIO INVESTIGATED IN THIS REPORT THE PERFORMANCE IS FURTHER IMPROVED BY APPLYING LOW RANK APPROXIMATIONS TO SEPARABLE ESTIMATORS
4HIS PART IS REPRODUCED FROM - 3ANDELL AND / %DFORS ! COMPARATIVE STUDY OF PILOT BASED CHANNEL ESTIMATORS FOR WIRELESS /&$- 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT

#OMPARATIVE STUDY OF CHANNEL ESTIMATORS

)NTRODUCTION
7IRELESS ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING /&$- IS CURRENTLY USED AND PRO POSED FOR SEVERAL BROADCASTING APPLICATIONS 4HE %UROPEAN STANDARD FOR DIGITAL AUDIO BROADCAST $!" ;= USES /&$- WITH DIdERENTIAL PHASE SHIFT KEYING $03+ 4HIS IS SUITABLE FOR LOW BIT RATE SYSTEMS BUT WHEN HIGHER BIT RATES ARE REQUIRED MULTIAMPLI TUDE MODULATION IS MORE APPROPRIATE 0ROPOSALS FOR DIGITAL VIDEO BROADCASTING ; = HAVE INCLUDED MULTIAMPLITUDE MODULATION /&$- 4HESE SCHEMES CAN BE MADE DIdEREN TIAL WHICH OdERS THE ADVANTAGE OF AVOIDING CHANNEL ESTIMATION $IdERENTIAL AMPLITUDE AND PHASE SHIFT KEYING $!03+ ;= IS AN EXAMPLE OF THIS APPROACH )N $!03+ HOW EVER THE CONSTELLATION POINTS ARE NON UNIFORMLY DISTRIBUTED IN THE SIGNAL SPACE WHICH REDUCES PERFORMANCE 4HERE MAY ALSO BE METRIC DIbCULTIES CONCERNING DECODING #O HERENT MODULATION ON THE OTHER HAND GIVES BETTER PERFORMANCE BUT BECAUSE OF THE NECESSARY CHANNEL ESTIMATION IT REQUIRES MORE COMPLEXITY AT THE RECEIVER )T IS OF INTER EST THEREFORE TO INVESTIGATE THE PERFORMANCE OF COHERENT /&$- SYSTEMS USING CHANNEL ESTIMATORS WITH DIdERENT LEVELS OF COMPLEXITY )N THIS PAPER WE ANALYZE LOW COMPLEXITY COHERENT DEMODULATION RECEIVER SCHEMES SUITABLE FOR HIGH BIT RATE /&$- /NE WAY OF ESTIMATING THE CHANNEL IN A AT FADING ENVIRONMENT IS TO MULTIPLEX PILOTS KNOWN SYMBOLS INTO THE TRANSMITTED SIGNAL &ROM THESE SYMBOLS ALL CHANNEL ATTENU ATIONS ARE ESTIMATED WITH AN INTERPOLATION LTER 4HIS TECHNIQUE IS CALLED PILOT SYMBOL ASSISTED MODULATION 03!- AND WAS INTRODUCED FOR SINGLE CARRIER SYSTEMS BY -OHER AND ,ODGE ;= AND ANALYZED BY #AVERS ;= 3INCE EACH SUBCHANNEL IN /&$- IS AT FADING 03!- CAN BE GENERALIZED TO TWO DIMENSIONS WHERE PILOTS ARE TRANSMITTED IN CER TAIN POSITIONS IN THE TIME AND FREQUENCY GRID OF /&$- 4HE CHANNEL ESTIMATION IS THEN PERFORMED BY A TWO DIMENSIONAL INTERPOLATION (HER ;= PROPOSES TO USE NITE IMPULSE RESPONSE &)2 LTERS FOR THIS AND TO SEPARATE THE USE OF TIME AND FREQUENCY CORRELATION (E ARGUES THAT THIS IS A GOOD TRADE Od BETWEEN COMPLEXITY AND PERFORMANCE 4HE SPACING OF PILOT SYMBOLS IN 03!- FOR SINGLE CARRIER SYSTEMS WAS INVESTIGATED IN ;= )T WAS FOUND THAT THE OPTIMUM SPACING WAS SOMEWHAT CLOSER THAN THE .YQUIST RATE IE THE INVERSE OF THE BANDWIDTH OF THE CHANNEL COVARIANCE FUNCTION 7E GENERALIZE THIS RESULT TO TWO DIMENSIONS FOR THE /&$- TIME FREQUENCY GRID 5SING A DENSE PILOT PATTERN MEANS THAT THE CHANNEL IS OVERSAMPLED IMPLYING THAT LOW RANK ESTIMATION METHODS ;= CAN WORK WELL 4HIS TYPE OF LOW COMPLEXITY ESTIMATION PROJECTS THE OBSERVATIONS ONTO A SUBSPACE OF SMALLER DIMENSION AND PERFORMS THE ESTIMATION IN THAT SUBSPACE "Y OVER SAMPLING THE CHANNEL IE PLACING THE PILOT SYMBOLS CLOSE TO EACH OTHER THE OBSERVATIONS ESSENTIALLY LIE IN A SUBSPACE AND LOW RANK ESTIMATION IS VERY EdECTIVE )N THIS REPORT WE PRESENT AND ANALYZE PILOT BASED /&$- CHANNEL ESTIMATORS THAT RELY ON BOTH TIME AND FREQUENCY CORRELATION OF THE FADING CHANNEL 4HE ESTIMATORS ARE LINEAR AND FEEDFORWARD IE NO DECISION DIRECTION OR FEEDBACK IS USED 7E DIVIDE THEM INTO TWO CLASSES DIMENSIONAL $ AND SEPARABLE ESTIMATORS 4HE LATTER USES DIMENSIONAL $ INTERPOLATION LTERS IN THE TIME AND FREQUENCY DIRECTIONS SEPARATELY )N EACH CLASS WE COMPARE A &)2 7IENER LTER ;= WITH A LOW RANK APPROXIMATION OF THE LINEAR MIN IMUM MEAN SQUARED ERROR ,--3% ESTIMATOR ;= 4HE ESTIMATORS ARE COMPARED BOTH IN TERMS OF MEAN SQUARED ERROR -3% AND CODED BIT ERROR RATE "%2 4HE SYSTEM AND THE SCENARIO ARE INTRODUCED IN 3ECTION 4HE ESTIMATORS ARE DESCRIBED IN 3ECTION

AND THEIR PERFORMANCE IS PRESENTED IN 3ECTION BOTH IN TERMS OF MEAN SQUARED ERROR AND CODED BIT ERROR RATE &INALLY IN 3ECTION WE PRESENT CONCLUSIONS

3YSTEM DESCRIPTION
/&$- SYSTEM
)N THIS REPORT WE CONSIDER AN /&$- SYSTEM OPERATING IN A 2AYLEIGH FADING CHANNEL ENVIRONMENT 4HIS SYSTEM USES A CYCLIC PREX ;= WHICH IS A COPY OF THE LAST PART OF THE /&$- SYMBOL AND ACTS AS A GUARDSPACE BETWEEN CONSECUTIVE /&$- SYMBOLS (ENCE IF THE IMPULSE RESPONSE OF THE CHANNEL IS SHORTER THAN THE CYCLIC PREX INTER SYMBOL INTERFERENCE )3) IS AVOIDED &URTHERMORE IF THE CHANNEL IS ASSUMED CONSTANT DURING ONE /&$- SYMBOL INTER CARRIER INTERFERENCE )#) IS ALSO AVOIDED ;= )N &IGURE A A SCHEMATIC VIEW OF THE BASE BAND /&$- SYSTEM IS SHOWN 4HE MOD
&IGURE /&$- SYSTEM A "ASE BAND MODEL B PARALLEL SUBCHANNELS MODEL #0 AND #0 DENOTE THE INSERTION AND DELETION OF THE CYCLIC PREX RESPECTIVELY ULATION ON - SUBCARRIERS IS PERFORMED BY AN INVERSE DISCRETE &OURIER TRANSFORM )$&4 IN THE TRANSMITTER ;= 3IMILARLY DEMODULATION IS DONE WITH A DISCRETE &OURIER TRANSFORM $&4 IN THE RECEIVER 4HE EdECTIVE SYMBOL LENGTH IS 3 -3 WHERE 3 IS THE SAMPLING PERIOD OF THE SYSTEM !DDING A CYCLIC PREX #0 WITH A LENGTH OF 3 +3 MAKES THE TOTAL SYMBOL LENGTH 3
3 )F )3) AND )#) ARE ELIMINATED WE CAN DESCRIBE THE SYSTEM AS A SET OF PARALLEL 'AUSSIAN CHANNELS ;= SHOWN IN &IGURE B WITH CORRELATED CHANNEL
R R & R &
#OMPARATIVE STUDY OF CHANNEL ESTIMATORS ATTENUATIONS J G & -3

J J

u J - `
R
WHERE & a IS THE FREQUENCY RESPONSE OF THE CHANNEL F ~ DURING THE /&$- SYMBOL 4HE RECEIVED SIGNAL X ON SUBCHANNEL J CAN THUS BE DESCRIBED AS X G W
M
J J J J J J

WHERE W IS THE TRANSMITTED DATA SYMBOL AND M THE CHANNEL NOISE AT SUBCARRIER J 4HE RELATION HOLDS FOR EVERY /&$- SYMBOL THUS CREATING A TWO DIMENSIONAL GRID WITH FREQUENCY SUBCARRIERS ON ONE AXIS AND TIME /&$- SYMBOLS ON THE OTHER

3CENARIO
4HE CHANNEL ESTIMATION IS BASED ON PILOTS TRANSMITTED AT CERTAIN POSITIONS IN THE TIME FREQUENCY GRID OF THE /&$- SYSTEM 4HE CHANNEL ATTENUATIONS ARE ESTIMATED BY MEANS OF INTERPOLATION BETWEEN THESE PILOTS WHERE WE ASSUME THAT THE CHANNEL ESTIMATORS CAN USE ALL TRANSMITTED PILOTS 4HIS IS THE CASE IN EG BROADCASTING OR IN THE DOWNLINK OF A MULTIUSER SYSTEM )N BOTH THESE CASES THERE IS ONLY ONE PHYSICAL CHANNEL BETWEEN THE TRANSMITTER AND THE RECEIVER 4HUS CHANNEL ATTENUATIONS IN NEIGHBORING TIME FREQUENCY GRIDPOINTS ARE HIGHLY CORRELATED A FEATURE THAT CAN BE USED FOR CHANNEL ESTIMATION )N THE UPLINK OF A MULTIUSER SYSTEM ON THE OTHER HAND EACH USER HAS THEIR OWN PHYSICAL CHANNEL SO CHANNEL ATTENUATIONS STEMMING FROM DIdERENT MOBILE TRANSMITTERS MUST BE ASSUMED TO BE UNCORRELATED 4O ESTIMATE THE ATTENUATIONS FOR ONE USER ONLY PILOTS TRANSMITTED BY THAT USER CAN BE USED 4HUS THE UPLINK IS QUITE DIdERENT FROM OUR SCENARIO AND WILL NOT BE CONSIDERED HERE 4HE INVESTIGATED /&$- SYSTEM HAS A BANDWIDTH OF -(Z AND IS OPERATING IN THE '(Z FREQUENCY BAND 4HE NUMBER OF SUBCARRIERS IS - WHICH MAKES THE EdECTIVE SYMBOL LENGTH xS 4HE ENVIRONMENT IS A MACROCELL WHICH IS ASSUMED TO HAVE A MAXIMUM DELAY SPREAD OF xS AND A MAXIMUM $OPPLER FREQUENCY OF (Z 4HUS THE MAXIMUM $OPPLER FREQUENCY RELATIVE TO THE INTER CARRIER SPACING IS E WHICH CORRESPONDS TO A VEHICLE SPEED OF KMH 4HE POWER DELAY PROLE IS EXPONENTIALLY DECAYING WITH ROOT MEAN SQUARE 2-3 WIDTH ~ xS 4O ELIMINATE )3) WE USE A GUARD SPACE OF xS WHICH CORRESPONDS TO + SAMPLES 4HE LENGTH OF THE /&$SYMBOL IS
xS WHICH MAKES THE RELATIVE SIZE OF THE GUARD SPACE 4HE CORRESPONDING 2-1 LOSS IS D" /UR SYSTEM MODEL ASSUMES THAT THE CHANNEL IS CONSTANT DURING AN /&$- SYMBOL )N REALITY )#) OCCURS DUE TO CHANNEL FADING DURING THE TRANSMISSION OF AN /&$- SYMBOL ; = (OWEVER WITH A MAXIMUM RELATIVE $OPPLER FREQUENCY OF THE SIGNAL TO )#) RATIO IS D" ;= 4HIS IS NEGLIGIBLE IN THE 2-1 RANGES WE ARE LOOKING AT AND CONSEQUENTLY WE IGNORE THE )#) AND USE THE MODEL IN )NTERLEAVING IS PERFORMED SEPARATELY IN FREQUENCY AND TIME OVER A FRAME CONSISTING OF /&$- SYMBOLS 4HIS CORRESPONDS TO A MAXIMUM DELAY OF MS &IRST INTER LEAVING IN FREQUENCY IS DONE BY PLACING CONSECUTIVE DATA SYMBOLS SUBCARRIERS APART )NTERLEAVING IS THEN PERFORMED IN TIME WITH A UNIQUE PATTERN FOR EACH SUBCARRIER 4HESE
#MAX
RMS

PATTERNS ARE RANDOM PERMUTATIONS AND CHANGED EVERY FRAME "Y HAVING DIdERENT INTER LEAVING PATTERNS ON ALL SUBCARRIERS CHANNEL ATTENUATIONS ARE INTERLEAVED IN BOTH TIME AND FREQUENCY 4HIS PRODUCES AN ALMOST PERFECT INTERLEAVING WITH NO SIGNICANT PERFORMANCE LOSS &OR ERROR CORRECTION A RATE CONVOLUTIONAL CODE WITH THE OCTAL POLYNOMIALS IS USED IE THE CODE POLYNOMIALS ARE ;=
F #

#
#
#
#!
#
#
#
#!
$
F
4HE CONSTRAINT LENGTH IS M AND A TAIL OF M

ZEROS IS APPENDED TO CLEAR THE ENCODERS MEMORY 4HE RECEIVER USES A SOFT DECISION 6ITERBI DECODER WITH A TRUNCATED MEMORY LENGTH OF M BITS 4HE BITS ARE MODULATED USING "03+ IN EACH DIMENSION AND THE INPHASE AND QUADRATURE PARTS ARE CONCATENATED TO FORM 103+ SYMBOLS 4HIS MAKES THE DATA RATE OF THE SYSTEM -BITS
$ $

#HANNEL MODEL
)N OUR ANALYSIS WE USE THE WIDE SENSE STATIONARY UNCORRELATED SCATTERING 73353 CHANNEL MODEL INTRODUCED IN ;= "Y CONSIDERING THE CHANNEL TO BE CONSTANT OVER ONE /&$SYMBOL THE INSTANTANEOUS FREQUENCY RESPONSE OF THE , `PATH CHANNEL AT TIME S IS 8 D & E S P , (
, M M #M I t M {%#M S
{E ~ M

SG

WHERE t IS THE PHASE % THE $OPPLER FREQUENCY AND ~ THE DELAY OF THE M PATH !LL THESE PARAMETERS ARE INDEPENDENT RANDOM VARIABLES 4O OBTAIN 2AYLEIGH FADING WITH THE *AKES SPECTRUM ;= AND AN EXPONENTIALLY DECAYING POWER DELAY PROLE WITH 2-3 VALUE ~ WE CHOOSE THE PROBABILITY DENSITY FUNCTIONS AS ;=
M
RMS
%#
O t { %
t # D ~ RM S ~
{%#M A X

D
O ~
#
`~ ~ RM S `3BO ~ RM S

%# %#M AX
v t { J % J%
#
#MAX
v~ v3
BO
4HE RANDOM VARIABLES % AND ~ CAN EASILY BE OBTAINED FROM A UNIFORMLY DISTRIBUTED RANDOM GENERATOR WITH OUTPUTS : < BY USING THE INVERSES OF THE DESIRED CUMULATIVE DISTRIBUTION FUNCTIONS ;=

0ILOT PATTERN
"Y USING A TWO DIMENSIONAL GENERALIZATION ;= OF PILOT SYMBOL ASSISTED MODULATION ;= KNOWN SYMBOLS PILOTS ARE TRANSMITTED ON CERTAIN POSITIONS IN THE TIME FREQUENCY GRID 4HE NUMBER OF PILOTS TO USE IS A TRADE Od BETWEEN DATA RATE AND CHANNEL ESTIMATION PERFORMANCE (OWEVER BY VIEWING THE CHANNEL ESTIMATION IN THE TIME FREQUENCY GRID AS

A TWO DIMENSIONAL INTERPOLATION FUNDAMENTAL LIMITS ON THE DENSITY OF PILOTS CAN BE DE RIVED 4HE SCATTERED PILOT SYMBOLS CAN BE SEEN AS NOISY SAMPLES OF THE TWO DIMENSIONAL STOCHASTIC SIGNAL & E S 4HESE SAMPLES HAVE TO BE PLACED CLOSE ENOUGH TO FULL THE SAM PLING THEOREM AND AVOID ALIASING .OTE THAT THE EdECTIVE 3.2 IS LOWERED BY USING MANY PILOTS SINCE A SMALLER PART OF THE TRANSMITTED POWER IS USED FOR DATA SYMBOLS 3INCE & E S IN IS THE &OURIER TRANSFORM OF THE CHANNEL IMPULSE RESPONSE AT TIME S WHICH IS ASSUMED TO BE CONSTANT FOR ONE /&$- SYMBOL THE AUTO COVARIANCE FUNCTION OF & E S IS THE SPACED FREQUENCY SPACED TIME CORRELATION FUNCTION OF THE CHANNEL ;=
"
&&
aE aS $ F& E S & E ` aE S ` aS G

C
"
aE aS
MAX
4HE BANDWIDTH OF THIS FUNCTION IS ! THE $OPPLER SPREAD IN THE aE DIRECTION AND ~ THE MULTIPATH SPREAD IN THE aS DIRECTION ;= &OR THE ANALYZED /&$- SYSTEM WE HAVE ! ~
#MAX C
%
#MAX
MAX
+3
R
E -3
#MAX R
WHERE E IS THE MAXIMUM $OPPLER FREQUENCY RELATIVE TO THE INTER CARRIER SPACING )F WE ASSUME THAT PILOTS ARE PLACED - SUBCARRIERS APART IN EVERY - /&$- SYMBOLS WE HAVE t u G & Ja K a- -
+ 3 -3
E S E JK S R R
SINCE THE INTER CARRIER SPACING IS - 3 AND THE DURATION OF AN /&$- SYMBOL IS -
+ 3 4O FULLL THE SAMPLING THEOREM ;= WE NEED
R R

+
`
a

#MAX #MAX
)N THE ANALYZED SYSTEM WE HAVE - + AND E E
WHICH GIVES
` a

)N ;= WHERE 03!- FOR SINGLE CARRIER SYSTEMS IS ANALYZED IT IS SHOWN THAT THE "%2 CAN BE LOWERED BY PLACING THE PILOT SYMBOLS CLOSER THAN THAT SPECIED BY THE SAMPLING THEOREM .OTE THAT THERE EXISTS A PILOT SPACING WHICH OPTIMIZES THE TRADE Od BETWEEN IMPROVED CHANNEL ESTIMATION AND REDUCED 3.2 ON THE DATA SYMBOLS "Y VARYING THE PILOT SPACINGS - AND - IT WAS FOUND THAT - AND - WAS CLOSE TO OPTIMAL IN TERMS OF "%2 4HE USED PILOT PATTERN IS SHOWN IN &IGURE 4HIS MEANS THAT { OF THE BANDWIDTH AND THE TRANSMITTED POWER IS USED FOR PILOTS .OTE THAT THE CHANNEL IS OVERSAMPLED WHICH MEANS THAT LOW RANK ESTIMATORS CAN BE VERY EdECTIVE ;=
E S E S

&IGURE 4HE PILOT PATTERN USED IN THE SYSTEM 0ILOT SYMBOLS ARE MARKED WITH GREY SQUARES )N OUR STUDY THE PILOT SYMBOLS HAVE THE SAME AVERAGE POWER AS THE DATA SYMBOLS (OWEVER A TECHNIQUE CALLED BOOSTED PILOTS CAN ALSO BE USED WHICH IS PROPOSED FOR $6" ;= 4HIS MEANS THAT THE PILOT SYMBOLS ARE TRANSMITTED WITH A HIGHER AVERAGE POWER THAN THE DATA SYMBOLS 4HE AVERAGE 3.2 ON THE DATA SYMBOLS IS REDUCED BUT THE CHANNEL ESTIMATES ARE BETTER SINCE THE 3.2 AT THE PILOT SYMBOLS IS INCREASED 4HUS BY CHOOSING A SUITABLE POWER LEVEL FOR THE PILOT SYMBOLS THE BIT ERROR RATE CAN BE DECREASED

%STIMATORS
)N /&$- SYSTEMS THE OPTIMAL LINEAR ESTIMATOR IN THE MEAN SQUARED ERROR SENSE IS A $ BOTH TIME AND FREQUENCY LTER (OWEVER THE COMPLEXITY OF THIS ESTIMATOR IS USUALLY TOO LARGE FOR IT TO BE OF PRACTICAL USE ! NUMBER OF SUBOPTIMAL LOW COMPLEXITY CHANNEL ESTIMATORS HAVE BEEN SUGGESTED IN THE LITERATURE SEE EG ; = 7E WILL INVESTIGATE TWO CLASSES OF ESTIMATORS DIMENSIONAL AND SEPARABLE 4HE USE OF SEPARABLE LTERS IS A COMMON METHOD TO REDUCE COMPLEXITY IN MULTIDIMENSIONAL SIGNAL PROCESSING ;= &OR BOTH SEPARABLE AND NON SEPARABLE ESTIMATORS WE LOOK AT &)2 7IENER LTERS AND LOW RANK APPROXIMATIONS OF ,--3% ESTIMATORS 7E COMPARE ALL ESTIMATORS PERFORMANCES FOR TWO LEVELS OF COMPLEXITY 3INCE THEY ARE ALL LINEAR ESTIMATORS A REASONABLE MEASURE OF COMPLEXITY IS THE AVERAGE NUMBER OF MULTIPLICATIONS PER ESTIMATED ATTENUATION )N THE SEQUEL WE USE THE FOLLOWING NOTATION 4HE BACKROTATED OR LEASTSQUARES ESTIMATED CHANNEL ATTENUATIONS AT PILOT POSITIONS ARE DENOTED BY O
JK JK
X W
JK JK JK
WHERE X IS THE RECEIVED SIGNAL AT SUBCARRIER J IN /&$- SYMBOL K AND W IS THE CORRE SPONDING TRANSMITTED PILOT SYMBOL 4HE NAL ESTIMATE OF THE CHANNEL ATTENUATIONS G ARE LINEAR COMBINATIONS OF THE O S WHERE THE COEbCIENTS ARE CHOSEN ACCORDING TO EACH
JK JK

&IGURE 4WO DIMENSIONAL &)2 7IENER LTER 4HE ESTIMATED TONE b IS A LINEAR COMBINATION OF THE PILOT TONES d ESTIMATORS STRUCTURE "Y ARRANGING THE AVAILABLE ,3 ESTIMATES AT PILOT POSITIONS IN A VECTOR O AND THE CHANNEL ATTENUATIONS TO BE ESTIMATED IN A VECTOR G THE LINEAR MINIMUM MEAN SQUARED ERROR ESTIMATOR OF G IS ;= B 1GO 1 O G OO
WHERE 1GO IS THE CROSS COVARIANCE MATRIX BETWEEN G AND O AND 1OO IS THE AUTO COVARIANCE MATRIX OF O $EPENDING ON THE NUMBER OF PILOTS USED AND THEIR RELATIVE LOCATIONS THE SIZE OF O AND THE CORRESPONDING AUTO COVARIANCE MATRIX 1OO WILL CHANGE !LSO DEPENDING ON THE NUMBER OF ESTIMATED ATTENUATIONS THE SIZE OF G WILL CHANGE &URTHERMORE 1GO DEPENDS ON THE RELATIVE POSITIONS BETWEEN ESTIMATED ATTENUATIONS AND THE USED PILOT POSITIONS "ELOW WE ADDRESS SEVERAL CHOICES ON USED PILOTS AND ESTIMATED ATTENUATIONS

$ LTERS
4HE $ 7IENER LTER IS OPTIMAL IN TERMS OF -3% IF COMPLEXITY IS NOT CONSIDERED (OW EVER FOR A XED COMPLEXITY THE NUMBER OF LTER TAPS THAT CAN BE USED IS QUITE SMALL 7E USE THIS ESTIMATOR AS A REFERENCE AND INVESTIGATE A REDUCED COMPLEXITY ESTIMATOR WHICH IS DERIVED USING THE THEORY OF OPTIMAL RANK REDUCTION ;= $ ESTIMATOR
)F THE ALLOWED COMPLEXITY IS * MULTIPLICATIONS PER ATTENUATION THE TWO DIMENSIONAL LTER USES THE * PILOTS CLOSEST TO THE ESTIMATED ATTENUATION )N &IGURE WE DISPLAY AN EXAMPLE OF THE SEVEN PILOT POSITIONS USED * TO ESTIMATE ONE CHANNEL ATTENUATION &OR EVERY ESTIMATED CHANNEL ATTENUATION THERE IS A SET OF * ASSOCIATED PILOTS /PTIMAL WEIGHTS ARE CALCULATED ACCORDING TO &OR THE ESTIMATOR WITH THE LOWER COMPLEXITY WE WILL USE THE CLOSEST PILOTS AND FOR THE HIGHER COMPLEXITY THE CLOSEST

&IGURE 3TRUCTURE OF THE LOWRANK DIMENSIONAL ESTIMATOR 4HE TONES TO ESTIMATE ARE MARKED WITH b AND THE PILOTS USED ARE MARKED WITH d ,OWRANK $ ESTIMATOR
4HE LOWRANK $ ESTIMATOR IS IN A SENSE AN APPROXIMATION OF THE OPTIMAL $ ESTIMATOR IN THE PREVIOUS SECTION 4HE LOW COMPLEXITY IS ACHIEVED BY A GENERALIZATION OF THE IDEAS IN ;= 4O ALLOW A LOW RANK APPROXIMATION * ATTENUATIONS G ARE ESTIMATED SIMULTANEOUSLY USING THE * CLOSEST PILOTS O )F THE ATTENUATIONS TO BE ESTIMATED AND THE PILOTS USED ARE CHOSEN PROPERLY THE ESTIMATOR CAN BE WELL APPROXIMATED BY A LOW RANK ESTIMATOR THEREBY REDUCING THE COMPLEXITY CONSIDERABLY WHILE MAINTAINING MOST OF THE PERFORMANCE .OTE THAT THE ESTIMATED ATTENUATIONS CAN BE CHOSEN ARBITRARILY IN THE TIME FREQUENCY GRID )N &IGURE AN EXAMPLE IS GIVEN FOR THE LOCATION OF ESTIMATED ATTENUATIONS * AND THE USED PILOT SYMBOLS * 4HE ESTIMATOR BECOMES B & O G
G O G O Q
WHERE & IS A LOW RANK 7IENER LTER ;= &ROM THE SINGULAR VALUE DECOMPOSITION 36$
Q
1GO 1OO

4f5
'
WHERE 4 AND 5 ARE UNITARY MATRICES AND f IS A DIAGONAL MATRIX ;= THE LOW RANK 7IENER LTER IS DETERMINED BY ;= & 4f 5 1OO
' Q Q
WHERE f IS A * b * DIAGONAL MATRIX CONTAINING THE Q LARGEST SINGULAR VALUES 4HE COMPLEXITY OF THIS ESTIMATOR IS FOUND IN !PPENDIX ! TO BE u t * Q
*
Q G O O G
MULTIPLICATIONS PER ESTIMATED ATTENUATION WHERE Q IS THE RANK OF THE ESTIMATOR NUMBER OF SINGULAR VALUES USED * THE NUMBER OF PILOTS USED AND * THE NUMBER OF ATTENUATIONS TO BE ESTIMATED
O G

O

&OR THE LOWER COMPLEXITY ESTIMATOR WE CHOSE * PILOTS IN THE TIME DIRECTION AND IN THE FREQUENCY DIRECTION AND * ATTENUATIONS TO ESTIMATE IN THE TIME DIRECTION AND IN THE FREQUENCY DIRECTION 4HE LATTER WERE PLACED IN THE MIDDLE OF THE FORMER IN ORDER TO EXPLOIT AS MUCH CORRELATION AS POSSIBLE 7ITH A RANK OF Q THE NUMBER OF MULTIPLICATIONS PER ATTENUATION IS ACCORDING TO %Q u t "

G
LOW
&OR THE HIGHER COMPLEXITY WE CHOSE PILOTS IN THE TIME DIRECTION AND IN THE FREQUENCY DIRECTION AND ATTENUATIONS TO ESTIMATE IN THE TIME DIRECTION AND IN THE FREQUENCY DIRECTION 4HE RANK USED WAS GIVING u t "

HIGH
MULTIPLICATIONS PER ATTENUATION

3EPARABLE LTERS
3INCE $ LTERS IN GENERAL TEND TO HAVE A LARGE COMPUTATIONAL COMPLEXITY THE OUTER PRODUCT OF TWO $ LTERS CAN GIVE A GOOD TRADE Od BETWEEN PERFORMANCE AND COMPLEX ITY 4HIS IS A STANDARD TECHNIQUE IN MULTIDIMENSIONAL SIGNAL PROCESSING ;= AND IT HAS ALSO BEEN PROPOSED FOR PILOT BASED CHANNEL ESTIMATION IN /&$- SYSTEMS ;= )N PILOT BASED ESTIMATION SCHEMES THE MAJOR ADVANTAGE IS THE LOW NUMBER OF MULTIPLICATIONS PER USED PILOT 4HIS ALLOWS THE ESTIMATOR TO BE BASED ON MORE PILOTS THUS IMPROVING THE PERFORMANCE "ASED ON THE PILOT PATTERN CHOSEN THE GENERAL CONCEPT USED IN THIS REPORT IS SHOWN IN &IGURE WHERE A $ LTER IS APPLIED IN THE FREQUENCY DIRECTION 4HEREAFTER A $ LTER IS APPLIED IN THE TIME DIRECTION TO COMPLETE THE INTERPOLATION TO ALL POINTS IN THE GRID 7E INVESTIGATE BOTH AN ESTIMATOR BASED ON THE PROPOSAL IN ;= AND A VARIANT THEREOF WHICH ALLOWS THE USE OF MORE PILOTS BY A LOW RANK APPROXIMATION IN THE FREQUENCY DIREC TION 3EPARABLE &)2 LTERS
4HE USE OF SEPARATE $ &)2 LTERS IN THE TIME AND FREQUENCY DIRECTIONS HAS BEEN PRO POSED BY (HER IN ;= &IRST ALL ATTENUATIONS IN /&$- SYMBOLS CONTAINING PILOTS ARE ESTIMATED 4HIS IS DONE WITH A &)2 7IENER LTER OF LENGTH * 4WO ATTENUATIONS ON DIdERENT POSITIONS RELATIVE TO THE PILOTS WILL NEED DIdERENT LTERS SO THERE WILL BE DIdERENT LTERS FOR THIS ESTIMATION .OTE THAT THESE LTERS ARE NON CAUSAL IN THE SENSE THAT THEY WILL USE PILOTS ON BOTH SIDES OF THE ESTIMATED ATTENUATION IN ORDER TO EXPLOIT THE CLOSEST PILOTS !FTER THIS PROCEDURE THERE WILL BE ESTIMATES OF ALL ATTENUATIONS IN EVERY - /&$- SYMBOL &)2 7IENER LTERS OF LENGTH * ARE NOW USED IN THE TIME DIRECTION TO OBTAIN ESTIMATES OF ALL ATTENUATIONS (ERE THERE WILL BE - ` DIdERENT LTERS DEPENDING
E E SG S S S

&IGURE 3EPARABLE LTER BASED ON ONE DIMENSIONAL LTERS IN FREQUENCY AND TIME DIRECTIONS &ILTERING IN THE FREQUENCY DIRECTION IS PERFORMED RST ON WHICH ATTENUATION IS ESTIMATED 4HESE LTERS CAN BE NONCAUSAL WHICH WILL INTRODUCE A DELAY IN THE SYSTEM )F THIS DELAY CANNOT BE ACCEPTED CAUSAL LTERS MUST BE USED 4HE TOTAL NUMBER OF MULTIPLICATIONS PER ESTIMATED ATTENUATION IS *
* E S S
SINCE THE FREQUENCY DIRECTION LTER HAS TO BE APPLIED IN ONLY ONE OUT OF EVERY - /&$SYMBOLS 4HIS FACT CAN BE USED IN THE DESIGN OF THE LTERS SINCE THE FREQUENCY DIRECTION LTER TAPS ARE CHEAPER IN TERMS OF COMPLEXITY "Y REDUCING THE NUMBER OF TAPS IN THE TIME DIRECTION LTER BY ONE - TAPS CAN BE ADDED TO THE FREQUENCY DIRECTION LTER WITHOUT CHANGING THE TOTAL COMPLEXITY &OR THE LOWER COMPLEXITY WE USED TAPS IN THE FREQUENCY LTER AND TAPS IN THE TIME LTER 4HIS RESULTS IN "
MULTIPLICATIONS PER ATTENUATION &OR THE HIGHER COMPLEXITY AND TAPS WERE USED FOR THE FREQUENCY AND TIME LTERS RESPECTIVELY 4HIS MEANS A COMPLEXITY OF
S E
LOW
" MULTIPLICATIONS PER ATTENUATION
HIGH

,OW RANK SEPARABLE LTER
"Y USING OBSERVATIONS FROM ;= WHERE LOW RANK APPROXIMATIONS OF CHANNEL ESTIMATORS ARE PRESENTED WE REPLACE ONE OF THE &)2 LTERS IN THE ESTIMATOR PROPOSED BY (HER )NSTEAD OF &)2 7IENER LTERS IN BOTH DIRECTIONS A LOW RANK APPROXIMATION OF THE FREQUENCY DIRECTION

,--3% ESTIMATOR IS USED IN COMBINATION WITH THE TIME DIRECTION &)2 LTER (ENCE A FREQUENCY DIRECTION LTERING IS PERFORMED FOR EACH /&$- SYMBOL CONTAINING PILOTS !N OBVIOUS WAY OF DOING THIS LTERING IS TO ESTIMATE ALL ATTENUATIONS IN AN /&$- SYMBOL USING ALL PILOTS (OWEVER WHEN USING ALL PILOTS THE COMPLEXITY REDUCTION IS NOT SO LARGE THAT IT CAN COMPETE WITH A SHORT &)2 7IENER LTER 4HE NUMBER OF MULTIPLICATIONS PER ATTENUATION IS SEE !PPENDIX ! t u * Q
*
O G
WHERE Q IS RANK USED * AND * ARE THE NUMBER OF ATTENUATIONS TO ESTIMATE AND NUMBER OF PILOTS USED RESPECTIVELY 3INCE PILOTS FAR AWAY FROM THE ESTIMATED ATTENUATIONS ARE WEAKLY CORRELATED THEY DO NOT CONTRIBUTE MUCH TO THE ESTIMATE "Y EXCLUDING THEM THE COMPLEXITY GOES DOWN WHILE THE PERFORMANCE IS ALMOST THE SAME (ENCE THE /&$SYMBOL IS PARTITIONED INTO A NUMBER OF SUB SYMBOLS WHERE THE ATTENUATIONS ARE ESTIMATED USING ONLY THE * PILOTS CLOSEST TO THE SUB SYMBOL CONSISTING OF * SUBCARRIERS ;= )N &IGURE AN EXAMPLE IS SHOWN FOR * CHANNEL ATTENUATIONS AND * PILOTS ,OW RANK APPROXIMATIONS CAN BE DONE FOR THE TIME DIRECTION LTERING AS WELL BUT IN THIS REPORT WE HAVE CHOSEN TO USE AN &)2 LTER INSTEAD
G O O G G O
&IGURE 3TRUCTURE OF THE LOWRANK ESTIMATOR IN THE FREQUENCY DIRECTION 4HE * ATTENUATIONS TO ESTIMATE ARE MARKED WITH b AND THE * PILOT SYMBOLS USED ARE MARKED WITH d
G O
7E CHOSE FOR THE LOWER COMPLEXITY TO ESTIMATE ATTENUATIONS IN THE FREQUENCY DIRECTION USING PILOTS AND A RANK OF 4HIS RESULTS IN
MULTIPLICATIONS PER ATTENUATION IN THE FREQUENCY DIRECTION #OMBINED WITH A TAP TIME LTER THIS GIVES A TOTAL OF "
MULTIPLICATIONS PER ATTENUATION &OR THE HIGHER COMPLEXITY WE USED PILOTS TO ESTIMATE ATTENUATIONS WITH A RANK OF 4OGETHER WITH A TIME LTER WITH TAPS THIS GIVES a `
! # "
MULTIPLICATIONS PER ATTENUATION 4HE FOUR INVESTIGATED ESTIMATORS ARE SUMMARIZED IN 4ABLE
LOW HIGH

#OMPLEXITY AND USED PILOTS
3INCE ALL ESTIMATORS ARE LINEAR THE CHANNEL ESTIMATES ARE LINEAR COMBINATIONS OF A NUMBER OF PILOTS 4HE NUMBER OF PILOTS USED DEPENDS ON THE COMPLEXITY AND THE TYPE OF ESTIMATOR
%STIMATOR $
#OMPARATIVE STUDY OF CHANNEL ESTIMATORS 3TRUCTURE 5SES THE * CLOSEST PILOTS

O G O
,OW RANK $ %STIMATES * ATTENUATIONS USING * PILOTS AND A RANK OF Q 3EPARABLE 3EPARABLE &)2 LTER WITH * FREQUENCY AND * TIME TAPS
E S
Q
*E -S
MULTATT s r *
O *O *G
PILOTS *
O
*
E
*
*O *G
* *
S
,OW RANK SEPARABLE
%STIMATES * ATTENUATIONS USING * PILOTS AND A RANK OF Q FREQUENCY AND * TAPS &)2 LTER TIME
G O S
Q -S
* *
O
4ABLE 4HE FOUR CHANNEL ESTIMATORS INVESTIGATED IN THIS REPORT 4HE AVERAGE NUMBER OF MULTIPLICATIONS PER ATTENUATION AND THE NUMBER OF PILOTS THE ESTIMATORS ARE BASED ON ARE SHOWN IN 4ABLE 4HE ESTIMATORS USE THE PARAMETERS DESCRIBED IN THE PREVIOUS SECTION !S CAN BE SEEN IN THE TABLE FOR A XED COMPLEXITY THE NUMBER OF PILOTS USED IN THE ESTIMATES CAN BE INCREASED BY POSING RESTRICTIONS ON THE ESTIMATORS SUCH AS SEPARABILITY AND LOW RANK (OWEVER BECAUSE A LARGE NUMBER OF PILOTS ARE USED MANY WILL BE ONLY WEAKLY CORRELATED WITH THE ESTIMATED CHANNEL ATTENUATIONS AND WILL NOT THEREFORE CONTRIBUTE MUCH 4HUS CONSIDERING CORRELATION MISMATCH IN THE DESIGN THE ESTIMATION MAY ACTUALLY BE DEGRADED BY INCREASING THE NUMBER OF USED PILOTS 4HE MAIN ADVANTAGE OF USING MANY PILOTS IS THAT THE IMPACT OF THE CHANNEL NOISE IS REDUCED BY A LARGE AVERAGING

$ESIGN ASPECTS
)N ORDER TO MAKE THE CHANNEL ESTIMATORS ATTRACTIVE TO IMPLEMENT WE ASSUME THAT THEY ARE XED IE DESIGNED FOR BOTH A XED CHANNEL CORRELATION AND A XED 3.2 4HE FREQUENCY CORRELATION IS DETERMINED BY THE POWER DELAY PROLE OF THE CHANNEL ;= AND THE TIME CORRELATION IS DETERMINED BY THE $OPPLER FREQUENCY ;= .EITHER THE POWER DELAY PROLE NOR THE $OPPLER FREQUENCY ARE KNOWN BY THE RECEIVER 5SING OUR CHANNEL MODEL WE DESIGN THE ESTIMATORS FOR A MAXIMUM RELATIVE $OPPLER FREQUENCY OF AND A UNIFORM POWER DELAY PROLE OVER THE LENGTH OF THE CYCLIC PREX 4HE USE OF THESE WORST CASE PARAMETERS FOLLOWS THE RECOMMENDATIONS IN ;= WHERE PILOT SYMBOL ASSISTED %STIMATOR $ ,OW RANK $ 3EPARABLE &)2 ,OW RANK SEPARABLE ,OW COMPLEXITY MULTATT PILOTS (IGH COMPLEXITY MULTATT PILOTS
4ABLE !VERAGE NUMBER OF MULTIPLICATIONS PER TONE AND THE NUMBER OF PILOTS THE ESTIMATORS ARE BASED ON
#OMPARATIVE STUDY OF CHANNEL ESTIMATORS 0ARAMETER .O OF SUBCARRIERS 4IME DISPERSION 0OWER DELAY PROLE 4RUE | xS "D $ESIGN | xS 3 ~ 3 OTHERWISE D"
BO

~ xR
-AX REL $OPPLER FREQUENCY 3.2 6ARYING
~ 3 OTHERWISE
BO
BO
4ABLE $ESIGN AND TRUE VALUES OF SYSTEM PARAMETERS 4HE CONSTANT " IS A NORMALIZA TION FACTOR MODULATION IS ANALYZED &OR THE DETERMINATION OF THE CORRELATION MATRICES SEE !PPENDIX " #ONTRARY TO THE WORST CASE RECOMMENDATIONS FOR XED DESIGN CORRELATIONS THE XED DESIGN 3.2 SHOULD BE CHOSEN TO A BEST CASE 4HIS IMPLIES A CLOSE TO OPTIMAL PERFORMANCE FOR 3.2S BELOW THE DESIGN 3.2 WHERE THE EdECTS OF 3.2 MISMATCH ARE SMALL COMPARED TO THE OVERALL NOISE LEVEL 7E HAVE CHOSEN THE DESIGN 3.2 TO D" 7E EVALUATE THE ESTIMATORS UNDER MISMATCH IE THEY ARE DESIGNED FOR THE WRONG CHANNEL CORRELATION AND 3.2 0ARAMETERS FOR THE DESIGN AND THE TRUE VALUES OF THE CHANNEL STATISTICS ARE SHOWN BELOW IN 4ABLE .OTE THAT THERE IS NO MISMATCH IN $OPPLER FREQUENCY 'IVEN THE DESIGN FOR RELATIVE $OPPLER FREQUENCY THE PERFORMANCE IS APPROXIMATELY THE SAME FOR E v ;= 4HE 3.2 IS DENED AS THE TRANSMITTED ENERGY PER DATA BIT OVER THE CHANNEL NOISE VARIANCE h i h i $ JG J $ JW J h i 2- 1 a A $ JM J
#MAX J J J
WHERE A DENOTES THE NUMBER OF INFORMATION BITSSYMBOL )N OUR CASE WE HAVE A

0ERFORMANCE EVALUATION
7E EVALUATE THE PERFORMANCE OF THE FOUR INVESTIGATED ESTIMATORS BOTH IN TERMS OF MEAN SQUARED ERROR -3% AND CODED BIT ERROR RATE "%2 4HE -3% IS THEORETICALLY CALCULATED WHILE THE "%2 IS SIMULATED 4O SIMPLIFY THE -3% CALCULATIONS WE IGNORE EDGE EdECTS IN THE TIME FREQUENCY GRID AND ASSUME THAT IT IS OF INNITE EXTENT &OR A SYSTEM WHERE THE NUMBER OF SUBCARRIERS IS MUCH LARGER THAN THE LENGTH OF THE ESTIMATOR THESE EdECTS CAN BE IGNORED

-EAN SQUARED ERROR

B &O G
3INCE ALL ESTIMATORS ARE LINEAR THEY CAN BE EXPRESSED AS BY COLLECTING ALL USED PILOTS IN A VECTOR O AND CALCULATING THE CORRESPONDING ESTIMATOR B CAN BE EXPRESSED AS ;= MATRIX & 4HE COVARIANCE MATRIX OF THE ERROR D G ` G 1DD 1GG ` 1GO & ` &1GO
&1OO &
' ' '

&OR ALL ESTIMATORS THERE WILL BE DIdERENT MEAN SQUARED ERRORS DEPENDING ON THE ESTIMATED ATTENUATION )N ORDER TO COMPARE THE ESTIMATORS WE ONLY LOOK AT THE AVERAGE ERROR OVER ALL ATTENUATIONS IE 8 1DD M M TR 1DD -3% * ( *
*G G M G
WHERE * IS THE NUMBER OF ESTIMATED ATTENUATIONS AND TR DENOTES THE TRACE OF A MATRIX ;= )N &IGURE THE -3% FOR THE ESTIMATORS WITH THE LOWER COMPLEXITY LEVEL ARE SHOWN AS A FUNCTION OF 3.2 )T SHOULD BE NOTED THAT THE NON SEPARABLE ESTIMATORS HAVE AN ERROR OOR THAT IS ALREADY VISIBLE AT LOW 3.2 &OR THE SEPARABLE ESTIMATORS THE ERROR CURVES LEVEL OUT FOR VERY HIGH 3.2S BUT THIS ERROR OOR IS HARDLY NOTICEABLE IN THE GURE "ECAUSE THERE WILL ALWAYS BE AN INTERPOLATION ERROR EVEN IN THE NOISELESS CASE ALL THE ESTIMATORS HAVE AN ERROR OOR 4HIS IS DUE TO NITE LTER LENGTHS &OR 2-1 D" THE LOW RANK SEPARABLE ESTIMATOR IS D" BETTER THAN SEPARABLE &)2 LTERS 4HE -3% CURVES FOR THE HIGHER COMPLEXITY ARE SHOWN IN &IGURE 4HE ERROR OORS HAVE NOW BEEN LOWERED AND ARE ONLY NOTICEABLE FOR THE $ ESTIMATOR .OTE ALSO THAT THE MUTUAL ORDERING IS THE SAME AS FOR THE LOWER COMPLEXITY IE THE LOW RANK SEPARABLE ESTIMATOR IS THE BEST AND THE $ ESTIMATOR IS THE WORST 4HE DIdERENCE BETWEEN THE LOW RANK SEPARABLE ESTIMATOR AND SEPARABLE &)2 LTERS IS NOW ABOUT D" )N BOTH &IGURES AND IT CAN BE SEEN THAT SEPARABLE ESTIMATORS PERFORM BETTER THAN NON SEPARABLE 4HIS WAS NOTED IN ;= WHERE IT WAS ARGUED THAT SEPARABLE ESTIMATORS PROVIDE A GOOD TRADE Od BETWEEN COMPLEXITY AND PERFORMANCE
G

"IT ERROR RATE
4HE ESTIMATORS HAVE BEEN SIMULATED IN THE CODED SYSTEM )N &IGURE THE CODED "%2 IS SHOWN FOR THE ESTIMATORS WITH THE LOWER COMPLEXITY { MULTIPLICATIONS PER ATTENUATION !S A REFERENCE A SYSTEM WITH PERFECT KNOWLEDGE OF THE CHANNEL AT THE RECEIVER IS INCLUDED !S OBSERVED ABOVE ALL ESTIMATORS SUdER FROM AN ERROR OOR WHICH APPEARS DUE TO THE INTERPOLATION )N &IGURE THIS IS ONLY NOTICEABLE FOR THE $ ESTIMATOR 4HE OTHER ESTIMATORS DO NOT HAVE THIS DRAWBACK FOR 2-1 D" AND THEY PERFORM BETTER )T IS ALSO NOTICEABLE THAT ESTIMATORS THAT PERFORM WELL IN TERMS OF -3% ALSO HAVE A LOW "%2 WHICH IS EXPECTED 4HE BEST ESTIMATOR IS THE LOW RANK SEPARABLE ESTIMATOR WHICH IS ONLY ABOUT D" WORSE THAN KNOWN CHANNEL AND D" BETTER THAN SEPARABLE &)2 LTERS .OTE THAT FOR -3% THIS LATTER DIdERENCE WAS D" )N &IGURE THE "%2 CURVES FOR THE ESTIMATORS WITH THE HIGHER COMPLEXITY ARE SHOWN !GAIN THE "%2 WITH KNOWN CHANNEL IS INCLUDED AS A REFERENCE 4HE LOW RANK SEPARABLE ESTIMATOR IS NOW ONLY D" AWAY FROM KNOWN CHANNEL AND STILL ABOUT D" BETTER THAN SEPARABLE &)2 LTERS )N THIS GURE WE NOTE THAT THE PER FORMANCE OF ALL ESTIMATORS HAS INCREASED BUT THE ORDERING BETWEEN THEM IS NOT CHANGED IE THE LOW RANK SEPARABLE ESTIMATOR IS THE BEST AND THE $ IS THE WORST 4HE INVESTIGA TION HERE SUGGESTS THAT THIS HOLDS FOR MOST COMPLEXITY LEVELS (OWEVER FOR OTHER TYPES OF

&IGURE -EAN SQUARED ERROR RELATIVE TO CHANNEL POWER FOR THE FOUR ANALYSED ESTIMA TORS WITH THE LOWER COMPLEXITY MULTIPLICATIONS PER TONE
&IGURE -EAN SQUARED ERROR RELATIVE TO THE CHANNEL POWER FOR THE FOUR ANALYSED ESTIMATORS WITH THE HIGHER COMPLEXITY MULTIPLICATIONS PER TONE

&IGURE #ODED BIT ERROR RATE WITH THE LOW COMPLEXITY ESTIMATORS
&IGURE #ODED BIT ERROR RATE WITH THE HIGH COMPLEXITY ESTIMATORS

CHANNELS AND SCENARIOS SUCH AS THE UPLINK ANOTHER ESTIMATOR MIGHT BE BETTER 3EPARATE STUDIES ARE REQUIRED FOR THESE CIRCUMSTANCES &INALLY THE TWO LEVELS OF COMPLEXITY FOR THE LOW RANK SEPARABLE ESTIMATOR WHICH WAS SHOWN TO BE THE BEST ARE COMPARED TO KNOWN CHANNEL )N &IGURE IT CAN BE SEEN THAT THE LOW AND HIGH COMPLEXITY ESTIMATORS ARE ABOUT D" AND D" AWAY FROM KNOWN CHANNEL RESPECTIVELY 4HE "%2 WILL DECREASE WITH INCREASING COMPLEXITY BUT TO
&IGURE "%2 OF KNOWN CHANNEL AND THE SEPARABLE LOW RANK ESTIMATOR OF LOW COM PLEXITY MULTATT AND HIGH COMPLEXITY MULTATT GET REALLY CLOSE TO THE PERFORMANCE OF KNOWN CHANNEL A VERY HIGH COMPLEXITY IS NEEDED 4HIS PROMPTS AN ANALYSIS OF THE TRADE Od BETWEEN COMPLEXITY AND PERFORMANCE BUT THIS IS BEYOND THE SCOPE OF THIS REPORT

#ONCLUSIONS
)N THIS REPORT WE HAVE INVESTIGATED FOUR /&$- CHANNEL ESTIMATORS SUITABLE FOR BROAD CASTING OR FOR THE DOWNLINK IN A MULTIUSER SYSTEM 4HE ESTIMATORS USE PILOTS IE KNOWN SYMBOLS TRANSMITTED IN CERTAIN POSITIONS IN THE TIME FREQUENCY GRID OF /&$- 4WO CLASSES OF ESTIMATORS DIMENSIONAL AND SEPARABLE WERE INVESTIGATED 7ITHIN EACH CLASS WE COM PARED AN &)2 7IENER LTER WITH A LOW RANK ,--3% ESTIMATOR 4HROUGH ANALYTICAL CAL CULATIONS OF THE -3% AND SIMULATION OF THE CODED "%2 IT WAS FOUND THAT THE SEPARABLE ESTIMATORS WERE THE BEST FOR A XED COMPLEXITY 7ITHIN THE CLASS OF SEPARABLE ESTIMATORS THE LOW RANK ESTIMATOR WAS SHOWN TO BE ABOUT D" BETTER THAN THE &)2 ESTIMATOR FOR THE CODED "%2 4WO LEVELS OF COMPLEXITIES WERE INVESTIGATED AND MULTIPLICATIONS PER ESTIMATED ATTENUATION AND IT WAS FOUND THAT FOR THE CODED "%2 THE FORMER IS D" FROM KNOWN CHANNEL AND THE LATTER D" ! NATURAL CONTINUATION OF THIS INVESTIGATION

OF CHANNEL ESTIMATION IN /&$- SYSTEMS IS A MORE COMPREHENSIVE STUDY OF THE TRADE Od BETWEEN COMPLEXITY AND PERFORMANCE 7E HAVE USED A PILOT PATTERN WHERE OF THE TRANSMITTED SYMBOLS ARE KNOWN 4HIS PILOT PATTERN IS SUbCIENT TO OBTAIN GOOD ESTIMATIONS OF THE CHANNEL ATTENUATIONS WHILE INTRODUCING ONLY A SMALL OVERHEAD 7E HAVE ASSUMED THAT THE RECEIVER CAN USE ALL PILOT SYMBOLS THAT ARE TRANSMITTED 'ENERALLY THIS IS NOT THE CASE IN THE UPLINK IN A MULTIUSER SYSTEM WHERE THE CHANNEL ESTIMATION CAN BE BASED ONLY ON PILOTS TRANSMITTED BY A SINGLE USER (ENCE FOR THE UPLINK A SEPARATE STUDY MUST BE MADE TO INVESTIGATE THE PERFORMANCE OF THE CHANNEL ESTIMATORS
!
#OMPLEXITY OF LOW RANK ESTIMATORS

B& O G
Q Q
4HE LOW RANK $ ESTIMATOR CAN BE FORMULATED AS 8

Q J(
F F O
' J J
8
Q J(
HF OI F
J J O
SINCE & IS A RANK Q MATRIX 4HE INNER PRODUCTS HF OI REQUIRE * MULTIPLICATIONS EACH IE A TOTAL OF Q* MULTIPLICATIONS 4HE LINEAR COMBINATION IS OVER Q VECTORS OF LENGTH * IE REQUIRES Q* MULTIPLICATIONS 3INCE * ATTENUATIONS ARE SIMULTANEOUSLY ESTIMATED THE NUMBER OF MULTIPLICATIONS PER ATTENUATION BECOMES t u * Q*
Q* Q
* *
J O G G G O G O G G
"
1
&&
#ORRELATION MATRICES
aE aS $ F& E S & E ` aE S ` aS G 8 $ D D , (
, I tM t M I { %#M S MM
4HE AUTO CORRELATION OF THE CHANNEL MODEL IS
%#M S
I { E ~ M
E ~ M
3INCE ALL RANDOM VARIABLES ARE INDEPENDENT WE HAVE 1

&&
aE aS
i h i 8 h $ D $ D , ( h i h i $ D $ D 1 aE 1 aS
, I {%#M S I { E~M M I {%#M S I { E~M E S
IE THE CHANNEL CORRELATION IS SEPARABLE 4HE EXPECTATIONS CAN BE FOUND FROM STANDARD &OURIER TRANSFORMS ;= h i 1 aS $ D ) {% aS ` a RM S h i `D 1 aE $ D RM S ` D
I {aE~
S I {%#M S #MAX 3BO ~ I { E I { E E~M 3BO ~
RMS

WHERE ) a IS THE ZEROTH ORDER "ESSEL FUNCTION OF THE RST KIND .OTE THAT THE CORRELATION FUNCTION FOR THE UNIFORM POWER DELAY PROLE CAN BE OBTAINED BY LETTING ~
RMS
UNIFORM
` D aE I { aE 3
I {
E 3BO
BO
4HE CORRELATION BETWEEN CHANNEL ATTENUATIONS SEPARATED BY J SUBCARRIERS AND K /&$SYMBOLS IS i h $ G G Q J Q K

J K J JK K E K
WHERE Q
E
J
S
Q K
RMS RMS
a ` D ERM S 1 ` D ERM S
I {J E ~ t t u u + 1 K -
+ 3 ) {E
K t
E
J -3
~
I {J-
+~
RMS
#MAX
AND E ~ ~ 3 IS THE 2-3 SPREAD RELATIVE TO THE SAMPLING INTERVAL 3INCE THE ,3 ESTIMATES AT PILOT POSITIONS ARE
R
JK
X W
JK JK
G
JK
M W
JK JK
THE CROSS CORRELATION AND THE AUTO CORRELATION ARE i h Q J ` J Q K ` K $ G O | } h i $ O O Q J`J Q K`K
} $ p J ` J K ` K JW J
JK J K E S JK J K E S M J
5SING THESE FUNCTIONS THE AUTO CORRELATION 1 CALCULATED
OO
AND THE CROSS CORRELATION 1
GO
CAN BE

0ART #OMBINED LINEAR 6ITERBI EQUALIZERS ! COMPARATIVE STUDY AND A MINIMAX DESIGN
!BSTRACT #OMBINED LINEAR 6ITERBI EQUALIZER #,6% IS A TERM OFTEN USED FOR A CLASS OF DIGITAL RECEIVERS REDUCING THE COMPLEXITY OF THE 6ITERBI DETECTOR BY ASSUMING AN APPROXIMATE CHANNEL MODEL TOGETHER WITH LINEAR PRE EQUALIZATION OF THE RECEIVED DATA 7E RECONSIDER A WEIGHTED LEAST SQUARES DESIGN TECHNIQUE FOR #,6%S BY INTRODUCING A MINIMAX CRITERION FOR SUPPRESSING THE STRONGEST COMPONENT OF THE RESIDUAL INTERSYMBOL INTERFERENCE 0REVIOUSLY IN ;= WE HAVE STUDIED THE PERFORMANCE OF SOME PROPOSED #,6% DESIGN METHODS AND EVALUATED THEM BY SIMULATED BIT ERROR RATES (ERE WE INVESTIGATE THE PERFORMANCE OF THE MINIMAX DESIGN AND OF THE #,6% DESIGNS FOUND IN LITERATURE ; = FOR TWO '3- TEST CHANNELS 7E ALSO PRESENT A COMPARISON OF THE #,6% DESIGNS BASED ON A COMMON QUADRATIC OPTIMIZATION CRITERION FOR THE SELECTION OF THE CHANNEL PRELTER AND THE DESIRED IMPULSE RESPONSE
4HIS PART IS REPRODUCED FROM . 3UNDSTRM / %DFORS 0 DLING ( %RIKSSON 0 / "RJESSON AND 4 +OSKI #OMBINED LINEAR 6ITERBI EQUALIZERS ! COMPARATIVE STUDY AND A MINIMAX DESIGN )N 0ROC )%%% 6EHIC 4ECHNOL #ONF PAGES 3TOCKHOLM 3WEDEN *UNE

#OMBINED LINEAR 6ITERBI EQUALIZERS

)NTRODUCTION
4HE -AXIMUM ,IKELIHOOD 3EQUENCE $ETECTOR -,3$ IS A PROCEDURE FOR ESTIMATING A SEQUENCE OF BITS FROM A SEQUENCE OF CHANNEL OUTPUT OBSERVABLES GIVEN A MODEL OF THE COMMUNICATION SYSTEM )N THE PRESENCE OF INTERSYMBOL INTERFERENCE )3) THE 6ITERBI ALGORITHM 6! PROVIDES AN EbCIENT WAY OF COMPUTING THE -,3$ ; = (OWEVER THE 6! STILL BECOMES IMPRACTICAL WHEN THE TIME SPREAD OF THE )3) IS LARGE BECAUSE OF THE EXPONENTIAL RELATION BETWEEN )3) TIME SPREAD AND 6! COMPLEXITY 4HE COMPLEXITY OF A 6ITERBI DETECTOR CAN BE REDUCED BY GIVING THE 6! AN APPROXIMATE CHANNEL MODEL WITH A SHORTER TIME SPREAD THAN THAT OF THE ORIGINAL CHANNEL 4HE CLASS OF RECEIVERS EMPLOYING THIS TECHNIQUE TOGETHER WITH LINEAR PRE EQUALIZATION OF THE RECEIVED DATA ARE OFTEN REFERRED TO AS COMBINED LINEAR 6ITERBI EQUALIZERS #,6%S SEE ; = /THER CLASSES OF RECEIVERS ADDRESSING THE SAME COMPLEXITY PROBLEM CAN BE FOUND IN EG ; = )N THIS PAPER WE FOCUS ON #,6%S 7HEN DESIGNING #,6%S IT IS OFTEN DESIRABLE TO MINIMIZE THE BIT ERROR RATE OF THE RECEIVER 4HE BIT ERROR PROBABILITY DEPENDS ON THE DESIGN PARAMETERS SUCH AS THE CHANNEL MODEL AND THE IMPULSE RESPONSE OF THE PRELTER IN A COMPLICATED AND NON LINEAR WAY 4HE PRE LTERING OF THE RECEIVED DATA PERTURBS THE SIGNAL SPACE AND COLOURS THE CHANNEL NOISE )GNORING THIS COLOURING OR GIVING THE 6! AN APPROXIMATE CHANNEL MODEL RESULT IN A DISPLACEMENT OF THE DECISION REGIONS FROM THEIR OPTIMAL LOCATIONS CF THE RESIDUAL )3) IN ;= )NSTEAD OF USING THE BIT ERROR RATE AS A DESIGN CRITERION OTHER MORE FEASIBLE CRITERIA ARE USED IN #,6% DESIGN METHODS ; = )N THIS PAPER WE INVESTIGATE THE PERFORMANCE OF A MINIMAX #,6% DESIGN 4HREE OTHER PRINCIPAL TECHNIQUES FOR DESIGNING #,6%S ; = ARE ALSO OVERVIEWED AND COMPARED

4HE TRANSMISSION SYSTEM MODEL
! CONTINUOUS TIME MODEL OF A TRANSMISSION SYSTEM AS DESCRIBED IN ;= MAY WITHOUT INFORMATION LOSS BE REPRESENTED BY A DISCRETE TIME MODEL AS IN &IG
ni bi h
y i
Receiver
&IGURE $ISCRETE TRANSMISSION SYSTEM MODEL )N BLOCK TRANSMISSION SYSTEMS BLOCKS OF INFORMATION ARE TRANSMITTED SO THAT THE DE CODING CAN BE PERFORMED INDEPENDENTLY ON EACH BLOCK SEE EG ;= &OR SUCH SYSTEMS IT IS CONVENIENT TO USE THE MATRIX FORMALISM ;= X 'A
M


-
WHERE THE TRANSMITTED MESSAGE IS CODED ANTIPODALLY IN A F`

G AND THE CHANNEL OBSERVABLES ARE X 2 4HE TIME INVARIANT CHANNEL IMPULSE RESPONSE IE THE )3) COEbCIENTS IS REPRESENTED BY A 4OEPLITZ BAND MATRIX ' )T IS A KNOWN -
+ ` b MATRIX WITH THE IMPULSE RESPONSE G OF LENGTH + IN ITS COLUMNS ARRANGED SO THAT MATRIX MULTIPLICATION CORRESPONDS TO CONVOLUTION 4HE NOISE VECTOR M 2 IS A JOINTLY 'AUSSIAN ZERO MEAN RANDOM VECTOR WITH A - 1 DISTRIBUTION WHERE 1 DENOTES THE NOISE COVARIANCE MATRIX
+ + M M

#,6% DESIGN
4HE PERFORMANCE EVALUATION OF THE #,6%S IN THIS PAPER IS PERFORMED ON BLOCK TRANSMIS SION SYSTEMS BUT HISTORICALLY THE DESIGN METHODS HAVE FOCUSED ON CONTINUOUS TRANSMISSION SYSTEMS AND 6!S WITH INNITE HORIZON IE FA G ( 4O RELATE TO EXISTING METHODS WE HAVE CHOSEN TO CONNE THE DEVELOPMENT OF THE MINIMAX DESIGN TO A CRITERION BASED ON SE QUENCES OF INNITE LENGTH 3OME REMARKS ON #,6% DESIGN FOR BLOCK TRANSMISSION SYSTEMS CAN BE FOUND IN SECTION ! DESIGN MODEL IS PRESENTED IN &IG CF ;= WHERE P IS THE DESIRED IMPULSE RESPONSE $)2 IE THE CHANNEL MODEL GIVEN TO THE 6! )F THE TIME DELAY MODELLED BY THE LTER C IS ZERO THE SYSTEM IS EQUIVALENT TO THE ONE FOUND IN ;= 3INCE THIS PAPER ONLY CONSIDERS CAUSAL CHANNEL IMPULSE RESPONSES THE INTRODUCTION OF THE TIME DELAY OdERS THE POSSIBILITY OF PLACING THE ENERGY OF THE $)2 IN AN ARBITRARY POSITION 4HIS POSSIBILITY IS ACCOUNTED FOR IN ; = BY ALLOWING THE CHANNEL IMPULSE RESPONSE ANTI CAUSAL COMPONENTS
H H
ni h bi d q
yi
To VA
&IGURE $ISCRETE DESIGN MODEL FOR INNITE SEQUENCES 4HE ERROR
H
IN &IG CAN BE EXPRESSED AS

H
:A c G c O ` C c P
M c O<
H 3 H
IS GIVEN BY
WHERE c DENOTES CONVOLUTION "Y USING KWK , W ,W THE VARIANCE OF h i $ K'O ` #PK 1
KOK1
H
A M
WHERE O AND P ARE VECTORS CONTAINING THE IMPULSE RESPONSE OF THE PRELTER O AND THE $)2 P RESPECTIVELY AND 1 AND 1 ARE COVARIANCE MATRICES FOR THE TRANSMITTED SEQUENCE A AND THE NOISE M ' AND # ARE 4OEPLITZ BAND MATRICES SUCH THAT THE MULTIPLICATIONS 'O
A M

AND #P DESCRIBE CONVOLUTION 3INCE THE LENGTH OF G c O IS GREATER THAN THE LENGTH OF C c P THE SUBTRACTION 'O ` #P IN NECESSITATES THAT THE SIZE OF # IS CHOSEN SUCH THAT THE DIMENSIONS AGREE 4HE RESIDUAL )3) IS DENED BY THIS DIdERENCE AS RESIDUAL )3) 'O ` #P IE AS THE )3) THAT IS NOT ACCOUNTED FOR IN THE 6! CF ;=

! SELECTION OF PRELTER DESIGNS
)N THIS SECTION WE COMPARE DESIGN METHODS FOR O AND P 4HE METHODS ARE DUE TO &ALCONER AND -AGEE ;= &REDRICSSON ;= AND DLING ET AL ;= WHERE THE LAST ONE IS A WEIGHTED LEAST SQUARES METHOD INSPIRED BY ;= "Y INTRODUCING A WEIGHTING MATRIX 6 IN SELECTIVE WEIGHTING OF THE RESIDUAL )3) BECOMES POSSIBLE 4HIS IS USED TO FORMULATE THE DESIGN CRITERION
) O P K6 'O ` #P K 1
KOK1
A M

)N THE SEQUEL BOTH THE CHANNEL INPUT AND THE CHANNEL NOISE ARE ASSUMED WHITE AND STATIONARY HENCE 1 ( AND 1 } ( #OMPLETING THE SQUARE IN GIVES
A M M
) O P K O ` O K
KPK !
N

WHERE ' 6 6'

} ( ` a ! # 6 ( ` 6' ' 6 6#
3 3 M 3 3 3 3
O
'

6 6#P
3
)N THE PRESENCE OF NOISE THE MATRICES WITH RESPECT TO O IS OBTAINED IF O O
AND ! ARE POSITIVE DENITE SO THE MINIMUM OF
'

6 6#P
3

WITH A RESIDUAL ERROR OF KPK ! 4HE MOST STRAIGHTFORWARD DESIGN APPROACH IS TO ASSUME 6 ( IE A UNIFORM WEIGHT ING OF THE RESIDUAL )3) AND TO ND THE GLOBAL MINIMUM OF BY USING IN COMBI NATION WITH NDING A P THAT MINIMIZES KPK ! UNDER SOME CONSTRAINT EG KPK 4HIS IS DONE BY &ALCONER AND -AGEE IN ;= WHERE THE PRELTER OBTAINED IS O AND THE $)2 P
%, %,
'

#P
%,

THE NORMALIZED EIGENVECTOR CORRESPONDING TO THE SMALLEST EIGENVALUE OF !

4HERE ARE OTHER METHODS THAT DO NOT MINIMIZE BUT STILL OFTEN RENDER A LOWER PROBABILITY OF BIT ERROR THAN THE METHOD ABOVE /NE SUCH METHOD WAS PRESENTED BY

&REDRICSSON IN ;= &OR THE PURPOSE OF PRESERVING THE SIMILARITY TO THE EXPRESSIONS IN THE ORIGINAL REFERENCE WE PRESENT HIS RESULT IN THE &OURIER DOMAIN 0 E ' E / E J' E J
}
% %

AND J0 E J
%
J' E J
} }
M M

%QUATION CORRESPONDS TO OF &ALCONER AND -AGEE 4HE $)2 HOWEVER IS DETERMINED BY SPECTRUM MATCHING THE LEFT HAND SIDE TO THE RIGHT HAND SIDE OF EQUATION GIVEN A LENGTH CONSTRAINT ON P 4HIS MATCHING WAS ADDRESSED BY "EARE IN ;= WHERE AN ADAPTIVE ALGORITHM WAS PRESENTED ! WEIGHTED LEAST SQUARES 7,3 DESIGN APPROACH FOR O AND P WAS PRESENTED BY DLING ET AL IN ;= 4HE $)2 IN THIS DESIGN METHOD IS ASSIGNED AN EXACT COPY OF THE CORRE SPONDING POSITIONS OF 'O THUS GIVING THE 6! A CORRECT CHANNEL MODEL FOR THOSE POSITIONS IE P # 'O
3 6 +2 6 +2
4HE PRELTER O
6 +2
IS CHOSEN BY MODIFYING TO O

6 +2
'

P 6 6# P
3 K 6 +2

WHERE THE POSITION K IS CHOSEN AS A SUITABLE INDEX IN P "Y CHOOSING THE WEIGHTING MATRIX IN AS 6 CH@F a a a a a a a a a a a a POS
J
WHERE THE ZEROES COINCIDE WITH THE $)2 POSITION AND J IS CHOSEN SO THAT IT COINCIDES WITH IS AVOIDED THE KTH POSITION OF P 4HIS ENSURES THAT THE TRIVIAL ALL ZERO SOLUTION FOR O AND THAT THE ENERGY IN 'O IS CONCENTRATED TO THE $)2 INTERVAL %QUATION CAN NOW BE SIMPLIED TO THE FORM IN WHICH IT WAS PRESENTED BY DLING ET AL
6 +2 6 +2
O
J
6 +2
'

WHERE p IS A VECTOR WITH IN POSITION J AND ZEROES ELSEWHERE

! MINIMAX DESIGN
4HE CRITERION OF HAS THE TOTAL ENERGY OF THE WEIGHTED RESIDUAL )3) AS ONE PART )T CAN BE DISCUSSED IF ALSO THE DISTRIBUTION OF THE RESIDUAL )3) IS OF ANY IMPORTANCE FOR THE PROBABILITY OF BIT ERROR OF THE RESULTING RECEIVER (ERE WE INVESTIGATE AN APPROACH THAT TAKES INTO ACCOUNT THE MAXIMUM ABSOLUTE VALUE OF THE RESIDUAL )3) THEREBY SHAPING THE DISTRIBUTION OF THE SAME ,ET US USE THIS MINIMAX APPROACH AND SOLVE N O O @QF LHM L@W ) O
,,
#OMBINED LINEAR 6ITERBI EQUALIZERS WHERE ) O IS THE HTH ELEMENT IN THE VECTOR
H

) O J6 'O ` p J
} KOK
J M J 3

4HE EXPRESSION J6 'O ` p J DENOTES THE VECTOR OF THE ABSOLUTE VALUES OF EACH ELEMENT IN 6 'O ` p 4HE $)2 P # 'O THE WEIGHTING MATRIX 6 AND p ARE CHOSEN AS IN THE 7,3 DESIGN METHOD
J ,, ,, J

3IMULATIONS
4O EVALUATE THE PERFORMANCES OF THE #,6%S FOR THE DIdERENT PRELTER AND DESIRED IMPULSE RESPONSE DESIGN METHODS DESCRIBED IN SECTION WE HAVE SIMULATED THE BLOCK TRANSMIS SION SYSTEM OF SECTION 7E HAVE USED TWO '3- TEST CHANNELS ;= THE 4YPICAL 5RBAN AREA CHANNEL 45 AND THE 2URAL !REA CHANNEL 2!
0.4 0.2 0 0 4
Typical Urban Impulse resp.
10
20
0.8 0.6 0.4 0.2 0 0 2
Rural area Impulse resp.
10
20
Typical Urban Freq. resp.
Rural area Freq. resp.
0 0
0.2
0.4
0 0
0.2
0.4
&IGURE )MPULSE AND FREQUENCY RESPONSES OF THE '3- 4YPICAL 5RBAN CHANNEL LEFT AND 2URAL !REA CHANNEL RIGHT 7HEN IMPLEMENTING THE MINIMAX DESIGN WE HAVE USED A SEQUENTIAL QUADRATIC PRO GRAMMING METHOD PROVIDED BY THE MINIMAX ROUTINE OF THE -ATLAB3, /PTIMIZATION 4OOLBOX3, ;= TO SOLVE STARTING WITH AN INITIAL PRELTER O GIVEN BY 4O DETERMINE THE PRELTER O AND THE $)2 P OF &REDRICSSON WE HAVE SOLVED EQUATION BY MEANS OF A LEAST SQUARES T OF THE SPECTRUM COEbCIENTS OF P TO THE RIGHT HAND SIDE OF 4HE $)2 P WERE THEN CALCULATED ANALYTICALLY FROM THESE COEbCIENTS 4HE 2! CHANNEL HAS A SMOOTH FREQUENCY RESPONSE AND IS COMPARATIVELY EASY TO MATCH 4HE MATCHING OF P TO THE 45 CHANNEL IS MORE DIbCULT DUE TO THE SPECTRAL DIPS OF THIS CHANNEL SEE &IG !S A CONSEQUENCE AT HIGH 3.2S THE NUMBER OF POINTS USED IN THE LEAST SQUARES T MIGHT HAVE TO BE ADJUSTED IN ORDER TO OBTAIN REAL COEbCIENTS


Typical Urban Area Channel
10
-1
Bit error rate (BER)
10
-2
x Minimax + Fredricsson
* WLS o FaMa
10
-3
-5
5 10 Q:3 taps, P:53 taps
15
&IGURE "IT ERROR RATE FOR THE '3- 4YPICAL 5RBAN CHANNEL 3IMULATION RESULTS ARE SHOWN IN &IG AND FOR #,6% DESIGNS BY THE &ALCONER AND -AGEE METHOD ;= O THE 7,3 METHOD ;= c THE -INIMAX METHOD X AND THE METHOD OF &REDRICSSON ;= 4HE SIMULATIONS ON THE 2! CHANNEL ARE ALSO COMPARED WITH THE FULL COMPLEXITY 6ITERBI DECODER ;= DOTTED LINE IN &IG 3UCH A COMPARISON WAS NOT FEASIBLE IN SIMULATIONS ON THE 45 CHANNEL BECAUSE OF THIS CHANNELS LENGTH )N ALL CASES WE HAVE CHOSEN THE LENGTH OF THE DESIRED IMPULSE RESPONSE TO COEbCIENTS AND THE BLOCK LENGTH TO BITS 4HE PRELTER LENGTH IS SET TO ONE COEbCIENT MORE THAN TWICE THE )3) IE COEbCIENTS FOR THE 2! CHANNEL AND FOR THE 45 CHANNEL

#ONCLUSIONS AND FUTURE WORK
4HE BASIC IDEA IN MOST DESIGN METHODS FOR #,6%S IS REECTED BY THE CRITERION ) O P 4HE MAIN DIdERENCE BETWEEN THE METHODS PRESENTED IN THE LITERATURE ; = LIES IN THE WAY THE $)2 P IS CHOSEN &ALCONER AND -AGEE ND THE GLOBAL MINIMUM OF THE CRITERION UNDER THE CONSTRAINT KPK 4HE 7,3 METHOD AND THE METHOD PROPOSED BY &REDRICSSON HAS SLIGHTLY DIdERENT APPROACHES WHERE THE SUBSEQUENT PROCESSING OF DATA BY THE 6! IS TAKEN INTO CONSIDERATION WHEN APPLYING THE CRITERION &REDRICSSON USES A PROJECTION ON THE MINIMUM DISTANCE ERROR SEQUENCE GIVING AN EdECTIVE SIGNAL TO NOISE RATIO ;= TAILORED FOR THE 6! 4HE 7,3 METHOD ENSURES THAT THE $)2 GIVEN TO THE 6! IS A TRUE REPLICA OF THE TOTAL SYSTEM IMPULSE RESPONSE IN THE CORRESPONDING TIME INTERVAL )N OUR SIMULATIONS THE 7,3 RECEIVER SHOWS A SUPERIOR PERFORMANCE AND IS INDEED CLOSE TO THE FULL COMPLEXITY 6ITERBI DECODER 4HE RECEIVER PROPOSED BY &REDRICSSON PERFORMS

Rural Area Channel

10
-1
Bit error rate (BER)
10
-2
x Minimax + Fredricsson
* WLS
o FaMa .. Full Viterbi
10
-3
-10
-5
0 Q:3 taps, P:13 taps
10
&IGURE "IT ERROR RATE FOR THE '3- 2URAL AREA CHANNEL WELL ON THE 45 CHANNEL WHICH IS SLIGHTLY SURPRISING CONSIDERING THE SPECTRAL SHAPE OF THIS CHANNEL 4HE RESULTS IN THE PRESENTED SIMULATIONS AGREE WITH APPLICABLE OBSERVATIONS IN ; = AND IN OUR EARLIER INVESTIGATIONS ; = 4HE NEW METHOD BASED ON THE MINIMAX DESIGN WAS INTRODUCED AS AN ATTEMPT TO SHAPE THE DISTRIBUTION OF THE RESIDUAL )3) IN A FASHION FAVOURABLE TO THE 6! "EING SIMILAR TO THE 7,3 METHOD IT GIVES A TRUNCATED VERSION OF THE TOTAL SYSTEM IMPULSE RESPONSE AS A $)2 TO THE 6! BUT INSTEAD OF MINIMIZING THE CRITERION ) O P IT SUPPRESSES THE LARGEST RESIDUAL )3) COEbCIENT 4HE PERFORMANCE IS ALMOST UP TO PAR WITH THE 7,3 RECEIVER WHICH INDICATES A POTENTIAL FOR THE CONCEPT OF SHAPING THE RESIDUAL )3) 4HE HITHERTO DISCUSSED METHODS FOR #,6% DESIGN ARE DERIVED FOR CONTINUOUS TRANS MISSION SYSTEMS (OWEVER MANY CONTEMPORARY AND FUTURE COMMUNICATION SYSTEMS ARE OF BLOCK TRANSMISSION TYPE EG THE CELLULAR TELEPHONE SYSTEMS OF %UROPE '3- *APAN *$# AND THE 53! !$# 4O OUR KNOWLEDGE THERE ARE TODAY NO #,6% DESIGN METHODS THAT TAKE ADVANTAGE OF THE STRUCTURE OF SUCH SYSTEMS !N INCREASED UNDERSTANDING OF THE PROPERTIES OF BLOCK TRANSMISSION SYSTEMS COULD RESULT IN IMPROVED RECEIVERS WITH RESPECT TO BIT ERROR PROBABILITY AS WELL AS REDUCED IMPLEMENTATION COMPLEXITY AND COST )N #,6%S DEVELOPED FOR BLOCK TRANSMISSION SYSTEMS THIS COULD BE REECTED BY THE TIME INVARIANT LINEAR PRELTERING BEING REPLACED BY EG A GENERAL MATRIX MULTIPLICATION IN ORDER TO UTI LIZE THE EDGE EdECTS AT THE BLOCK BOUNDARIES FOR PERFORMANCE IMPROVEMENT 4HE NOISE CORRELATION DUE TO THE PRELTER O IS ANOTHER IMPORTANT ISSUE IN CONNECTION WITH #,6%S CF ;= 4HIS IS RECOGNIZED BY &REDRICSSON AND "EARE BUT NOT CONSIDERED IN THE OTHER DESCRIBED METHODS 7E REGARD THE ABOVE ISSUES AS KEY COMPONENTS IN THE DEVELOPMENT OF NEW #,6%S

3OURCE CODING

0ART $ESIGN AND ANALYSIS OF ROBUST SOURCE CODES

!BSTRACT )N THIS REPORT WE ARE PROPOSING A FRAMEWORK FOR ROBUSTNESS ANALYSIS OF SOURCE CODES IN COMBINATION WITH A DESIGN PROCEDURE FOR NDING AN OPTIMAL ROBUST CODE 4HE ROBUSTNESS IS DENED AS LOW SUSCEPTIBILITY TO CHANGES IN THE PROBABILITY DISTRIBUTION OF THE SOURCE )N MANY APPLICATIONS THE PROBABILITY DISTRIBUTION IS NOT WELL KNOWN NOR TIME INVARIANT (ENCE BY INTRODUCING A ROBUSTNESS MEASURE WE CAN DESIGN A CODE BASED ON A RELATIVELY POOR ESTIMATE OF THE PROBABILITY DISTRIBUTION WITHOUT RISKING A CATASTROPHIC INCREASE OF THE AVERAGE CODE WORD LENGTH IF THE ESTIMATE WAS INCORRECT OR THE SOURCE CHAR ACTERISTICS CHANGE 4HE BASIC CONCEPTS ARE ROBUSTNESS IN GENERAL AND GRADIENT ROBUSTNESS IN PARTICULAR 4HE GRADIENT ROBUSTNESS 2R # OF A VARIABLE LENGTH CODE # WITH CODE WORD LENGTHS LI s` r0 I I . ` IS DENED AS 2R # L ` L WHERE L IS THE ARITHMETIC MEAN OF THE I I CODE WORD LENGTHS )N THE EXPERIMENTAL PART IT IS SHOWN THAT EVEN A SMALL LOSS IN THE DEGREE OF COMPRESSION CAN GIVE A SUBSTANTIAL INCREASE OF THE ROBUSTNESS OF THE CODE 4HE PROCEDURE FOR NDING THE OPTIMAL ROBUST CODE IS COMPUTATIONALLY INTENSE BUT AN EASY TO COMPUTE APPROXIMATION FORMULA IS GIVEN 4HE EXPERIMENTS PERFORMED USING THE APPROXIMATION FORMULA ONLY LEAD TO MINOR DIdERENCES IN THE OBTAINED ROBUSTNESS COMPARED TO THE OPTIMAL ROBUST CODE 3OME OF THE OBSERVATIONS DONE IN THE EXPERIMENTS ARE q 4HE LENGTHS OF THE LONGEST CODE WORDS ARE REDUCED WHEN INTRODUCING ROBUSTNESS THUS REDUCING THE REQUIRED MEMORY SPACE FOR STORING CODE BOOKS q 7HEN THE EXPERIMENTS ON ENTROPY CLOSE CODES ARE APPLIED TO (UdMAN CODES WE GET ROUGHLY THE SAME RESULTS 4HE DIdERENCES IS DUE TO THE QUANTIZATION OF CODE WORD LENGTHS IN THE (UdMAN CODE q 4HREE DIdERENT METHODS FOR MEASURING THE BEHAVIOUR OF THE GRADIENT ROBUST CODES COMPARED TO THE OPTIMAL DATA COMPRESSION CODES ARE SHOWING PROMISING RESULTS 4HE ANALYSIS AND EXPERIMENTS IN THIS REPORT ONLY COVER A SMALL PART OF THE POSSIBILITIES BUT THE RESULTS OBTAINED ARE PROVIDING INCENTIVE FOR FURTHER STUDIES ! ROBUSTNESS MEASURE WE ARE PLANNING TO INVESTIGATE IN THE FUTURE IS BASED ON THE ARITHMETIC MEAN OF THE CODE WORD COSTS 4HIS PART IS REPRODUCED FROM / %DFORS AND 0 / "RJESSON $ESIGN AND ANALYSIS OF ROBUST SOURCE CODES 2ESEARCH 2EPORT 45,%! ,ULE 5NIVERSITY OF 4ECHNOLOGY ,ULE

2OBUST SOURCE CODES
2OBUST SOURCE CODES

)NTRODUCTION
4HIS REPORT IS CONCERNED WITH TWO MAJOR POINTS ONE IS TO PROPOSE ROBUSTNESS MEASURES TO BE USED WITH STATIC CODE BOOK SOURCE CODERS AND THE OTHER IS TO PROPOSE DESIGN PROCEDURES FOR SUCH CODES 2OBUSTNESS IN THIS CASE IS DENED AS LOW AVERAGE CODE WORD LENGTH SUSCEPTIBILITY TO CHANGES IN THE PROBABILITY DISTRIBUTION OF THE SOURCE )N MANY APPLICATIONS THE PROBABILITY DISTRIBUTION OF THE SOURCE IS NEITHER WELL KNOWN NOR TIME INVARIANT )N A SITUATION WHERE SOME KIND OF DATA COMPRESSION IS NEEDED AN EdECTIVE BUT COMPUTATIONALLY INTENSE METHOD TO SOLVE THE PROBLEM WITH A TIME VARYING PROBABILITY DISTRIBUTION IS TO USE AN ADAPTIVE SOURCE CODER SEE EG ;= AND ;= (OW EVER MANY APPLICATIONS ARE NOT DEPENDENT ON THE BEST POSSIBLE COMPRESSION RATE A HIGH COMPRESSION SPEED OR PORTABLE LOW COST HARDWARE MIGHT BE MORE IMPORTANT "ECAUSE OF THE NATURE OF THESE LIMITATIONS A STATIC CODER IS OFTEN A SUITABLE SOLUTION ,ONG TIME REGISTRATION OF %#'S ; = WHERE DATA COMPRESSION CODING IS USED IS A TYPICAL AP PLICATION WHERE A ROBUST SOURCE CODER IS SUITABLE SINCE ONE OF THE REASONABLE SYSTEM SPECICATIONS IS SMALL AND PORTABLE HARDWARE )F WE DECIDE TO USE A STATIC SOURCE CODE THERE ARE SEVERAL DIdERENT METHODS AVAILABLE FOR DESIGNING THE CODE /NE EXTREME IS TO ASSIGN CODE WORDS WITH EQUAL LENGTH AND THE OTHER EXTREME IS TO DO SOME KIND OF ESTI MATION OF THE PROBABILITY DISTRIBUTION OF THE SOURCE AND DESIGN AN OPTIMAL CODE FOR THAT PROBABILITY DISTRIBUTION 4HE INTENTION OF THIS REPORT IS TO PROPOSE TOOLS FOR THE ANALYSIS AND CONSTRUCTION OF A SPECTRUM OF STATIC CODES IN BETWEEN THESE TWO EXTREMES 4HERE ARE SEVERAL METHODS TO INTRODUCE WHAT WE CALL ROBUSTNESS /NE APPROACH IS TO PUT RESTRICTIONS ON THE LENGTH OF THE LONGEST CODE WORDS SEE EG ;= AND "RJESSON ET AL USED A PARTITIONING OF THE SYMBOL SET IN ; = 4HE SITUATION WHERE CODES ARE DESIGNED USING AN INCORRECT ESTIMATE OF THE PROBABILITY DISTRIBUTION HAS BEEN COVERED BY A NUMBER OF AUTHORS SEE EG ;= AND ;= %VEN THOUGH PARTS OF THE PROBLEM HAVE BEEN ADDRESSED BEFORE WE HAVE NOT FOUND A UNIED APPROACH TO THE CONCEPT WE CALL ROBUSTNESS )N #HAPTER WE OUTLINE THE CONDITIONS UNDER WHICH THE INVESTIGATION IS PERFORMED AND SOME BASIC NOTATIONS ARE INTRODUCED )N #HAPTER THE CONCEPT OF ROBUSTNESS IN GENERAL AND GRADIENT ROBUST CODES IN PARTICULAR ARE DISCUSSED TOGETHER WITH THE ANALYSIS OF SOME BASIC PROPERTIES OF GRADIENT ROBUST CODES 4HE MAIN RESULTS OF THE BEHAVIOUR OF GRADIENT ROBUST CODES ARE EVALUATED BY EXPERIMENTS IN #HAPTER ! DISCUSSION IN COMBINATION WITH SOME CONCLUSIONS IS PRESENTED IN #HAPTER

3YSTEM DESCRIPTION AND BASIC NOTATIONS

3
4HE SOURCE CODING SYSTEM ANALYSED IN THIS REPORT IS SHOWN IN &IGURE WHERE THE SOURCE IS DENED BY A PAIR H3 OI OF SOURCE SYMBOLS 3 FR R R G AND A VECTOR OF b c PROBABILITIES O O O a a a O &URTHER THE SYMBOLS PRODUCED BY THE SOURCE ARE ASSUMED TO BE INDEPENDENT 4HE SOURCE CODE IS DENED BY A SET OF CODE WORDS # FB B B G WHERE EVERY CODE WORD B IS ASSOCIATED WITH A COST K IN SOME SUITABLE UNIT FOR INSTANCE NATS OR BITS ! SET FREQUENTLY USED IN THE SEQUEL IS THE - ` DIMENSIONAL PROBABILITY SIMPLEX a TO WHICH THE PROBABILITY DISTRIBUTION O BELONGS
J J -

2OBUST SOURCE CODES
&IGURE 3OURCE CODING SYSTEM $ENITION 4HE - ` DIMENSIONAL PROBABILITY SIMPLEX IS DENED AS THE FOLLOWING SUBSET OF THE %UCLIDEAN SPACE 1
-
P1 J
-
8
H( -
P > P w
H H

WHICH IS A CLOSED SUBSET OF THE - CELL : < /NE OF THE QUANTITIES WE ARE TAKING INTO CONSIDERATION IS THE AVERAGE COST PER CODED SYMBOL 4HIS QUANTITY FOR THE ABOVE DESCRIBED SYSTEM IS GIVEN BY $ENITION $ENITION &OR A SOURCE CODING SYSTEM CONSISTING OF A SOURCE H3 OI AND A CODE # THE AVERAGE CODING COST FUNCTION 6 # O WHEN O a IS DENED BY
-
6 # O
8
H(
OK
H H

WHERE K IS THE COST ASSIGNED TO THE CODE WORD B

J J
)N A CASE WHERE WE CONSIDER VARIABLE LENGTH CODERS AND DECIDE TO USE THE LENGTH OF EACH CODE WORD AS THE CODING COST K EXPRESSION REPRESENTS THE AVERAGE CODE WORD LENGTH )N THIS SITUATION 3HANNONS 3OURCE #ODING 4HEOREM ;= STATES THAT THE COST 0 O KNF O OF THE 6 " O IS ALWAYS GREATER THAN OR EQUAL TO THE ENTROPY ' O ` ( SOURCE
J H H H

2OBUSTNESS
4HE REASON FOR INTRODUCING ROBUSTNESS IS THAT WE WOULD LIKE TO BE ABLE TO DESIGN STATIC SOURCE CODES THAT ARE CHARACTERIZED BY A HIGH ROBUSTNESS OF THE CODING COST AGAINST CHANGES IN THE PROBABILITY DISTRIBUTION OF THE SOURCE WITHOUT INCREASING THE CODING COST TOO MUCH 4HE EXPRESSION HIGH ROBUSTNESS AGAINST CAN BE INTERPRETED AS LOW SUSCEPTIBILITY TO $IdERENT APPLICATIONS OR SITUATIONS MIGHT CHANGE THE MEANING OF ROBUSTNESS AND CODING COST )N THIS CHAPTER WE HAVE CHOSEN TO FOCUS ON TWO LEVELS OF RENEMENT OF THE CONCEPT ROBUSTNESS 4HE RST LEVEL IS A GENERAL DESCRIPTION OF WHAT WE MEAN BY SOURCE CODING ROBUSTNESS AND THE SECOND LEVEL IS A CONCRETE EXAMPLE OF A ROBUSTNESS MEASURE GRADIENT ROBUSTNESS
2OBUST SOURCE CODES

0ROPERTIES OF A ROBUSTNESS MEASURE
7ITHOUT CONCENTRATING ON A PARTICULAR ROBUSTNESS MEASURE THERE ARE A FEW REASONABLE PROPERTIES THAT CAN BE ATTACHED TO ROBUSTNESS MEASURES IN GENERAL 7E ASSUME A RO BUSTNESS MEASURE TO BE A REAL VALUED FUNCTION INTRODUCING A PARTIAL ORDERING OF DIdERENT CODES AND ASSIGNING THE HIGHEST ROBUSTNESS TO A CODE WHERE THE CODE WORD COSTS ASSOCI ATED WITH EACH CODE WORD ARE ALL EQUAL 4HE DENITION BELOW IS THE ONE USED THROUGHOUT THIS REPORT $ENITION ,ET 5 DENOTE A SET OF POSSIBLE CODES FOR AN - `SYMBOL SOURCE &URTHER TO EVERY CODE " 5 ASSUME THAT THERE IS A VECTOR K :K K a a a K < OF POSITIVE REAL VALUED CODE WORD COSTS 7E DENE A ROBUSTNESS MEASURE 1h # BASED ON THE CRITERION h AS A FUNCTION HAVING THE FOLLOWING TWO PROPERTIES
3 m m m
4HE ROBUSTNESS 1h # IS A REAL VALUED NON NEGATIVE FUNCTION ON THE SET OF ALL SOURCE CODES ACCORDING TO SOME CRITERION h IE 1h # 5 `
-
1
3 -
AND IF THE ELEMENTS OF THE CODE WORD COST VECTOR K :K K a a a K < ASSOCIATED WITH THE CODE # 5 ARE ALL EQUAL IE K K K THEN 1h # w 1h # # 5
m m -
2EMARK !S A CONSEQUENCE OF $ENITION WE SAY THAT IF 1h # 1h # THEN THE CODE # IS A MORE ROBUST CODE THAN # 7HICH ROBUSTNESS CRITERION h TO BE USED IS A CHOICE DEPENDENT ON THE SITUATION IN WHICH THE ANALYSIS IS DONE )N THE NEXT SECTION WE INTRODUCE A ROBUSTNESS MEASURE WHICH HAS THE PROPERTIES OF $ENITION

'RADIENT ROBUST CODES
4O DEMONSTRATE THE ROBUSTNESS CONCEPT WE HAVE CHOSEN TO INTRODUCE A ROBUSTNESS MEA SURE BASED ON A DIRECTIONAL DERIVATIVE OF THE AVERAGE CODING COST 0ICTURE THE FOLLOWING SITUATION ! PROBABILITY DISTRIBUTION OF THE SOURCE IS ESTIMATED AND IT IS SUSPECTED TO CHANGE SLIGHTLY WITH TIME (OWEVER IT IS NOT KNOWN IN WHICH DIRECTION THE CHANGE WILL TAKE PLACE 4HE GRADIENT ROBUSTNESS MEASURE IS BASED ON A DIRECTIONAL DERIVATIVE OF THE AVERAGE CODING COST SEE EXPRESSION IN THE DIRECTION D WITH RESPECT TO THE PROBABILITY DISTRIBUTION O 4HIS CAN BE EXPRESSED AS
8 6 # O 6 # O
GD ` 6 # O KHL DK D G (
G H H H H -

WHERE D IS A UNIT VECTOR IN A VALID DIRECTION IN THE PROBABILITY SIMPLEX a 0 ! VALID D DIRECTION D IN THE PROBABILITY SIMPLEX a WILL HAVE TO SATISFY THE CONDITION (
H
,ET % BE THE SET OF UNIT VECTORS IN THE VALID DIRECTIONS THEN
-
2OBUST SOURCE CODES
%
D1 J
-
8
H(
D >
H
8
H(
D
H
3INCE THE DIRECTION OF CHANGE IS UNKNOWN WE CHOSE D TO BE THE DIRECTION IN WHICH THE DIRECTIONAL DERIVATIVE IS THE LARGEST AND DENE THE ROBUSTNESS 1 # TO BE THE RECIPROCAL 1 # L@W
D6
0
H(
DK

H H
4HE MAXIMIZATION OF THE DENOMINATOR OF EXPRESSION CAN BE DONE BY APPLYING STAN DARD LINEAR ALGEBRA OPERATIONS IN THE %UCLIDEAN SPACE 1 4HE MAXIMUM EQUALS THE LENGTH OF THE PROJECTION OF THE GRADIENT R6 # O ONTO THE HYPER PLANE COINCIDING WITH THE PROBABILITY SIMPLEX a )T CAN BE SHOWN THAT V U U8 8 8 T L@W DK K ` K D6 ( ( (
H H H I H H I
4HE INTERESTED READER CAN ND A MORE DETAILED ANALYSIS OF EXPRESSION IN !PPENDIX ! 5SING THE ABOVE CALCULATIONS WE ARE READY TO PROPOSE A DENITION OF GRADIENT ROBUSTNESS $ENITION 4HE GRADIENT ROBUSTNESS 1 # OF A SOURCE CODE # WITH CODE WORD COSTS K :K K a a a K < IS DENED AS
3 -
1 # Q0
H(
`
K `K
H
a 0
-

WHERE K IS THE ARITHMETIC MEAN OF THE CODE WORD COSTS IE K -
H(
K
H
/NE CAN EASILY VERIFY THAT THE GRADIENT ROBUSTNESS SATISES THE PROPERTIES OF $ENITION

4HE DESIGN OF ROBUST SOURCE CODES
!PART FROM THE ANALYSIS OF ROBUSTNESS IT IS DESIRABLE TO HAVE A DESIGN PROCEDURE THAT MAKES IT POSSIBLE TO USE A ROBUSTNESS MEASURE FOR THE DESIGN OF ROBUST SOURCE CODES 4HE RST DESIGN STEP IS TO ESTIMATE THE PROBABILITY DISTRIBUTION FOR THE SOURCE IN QUESTION )F THE SOURCE IS TIME INVARIANT AND THE PROBABILITY DISTRIBUTION IS ESTIMATED WITH A SUb CIENT ACCURACY THE MOST STRAIGHTFORWARD METHOD IS TO DESIGN AN OPTIMAL STATIC SOURCE CODE EG A (UdMAN CODE FOR THAT PARTICULAR DISTRIBUTION (OWEVER IF THE PROBABILITY DISTRIBUTION OF THE SOURCE IS TIME VARYING THE DESIGN PROCEDURE FOR A STATIC SOURCE CODE IS NOT STRAIGHTFORWARD
2OBUST SOURCE CODES

"Y USING A SUITABLE ROBUSTNESS MEASURE WE HAVE A TOOL FOR NDING A ROBUST STATIC SOURCE CODE TO BE USED ON A SOURCE WITH A PROBABILITY DISTRIBUTION THAT CHANGES SLIGHTLY WITH TIME $ENITION STATES THAT A CODE WITH EQUAL CODE WORD COSTS IS THE MOST ROBUST CODE (ENCE WE CAN ND THE MOST ROBUST CODE BY CHOOSING A CODE WITH EQUAL CODE WORD COSTS %VEN THOUGH THIS SEEMS LIKE A NICE PROPERTY THE UNIVERSALLY MOST ROBUST CODE IS PROBABLY NOT A VALID OPTION IN A TYPICAL APPLICATION )N A CASE WHERE THE CODE WORD COSTS EQUALS THE CODE WORD LENGTHS THE UNIVERSALLY MOST ROBUST CODE IS A XED LENGTH CODE IE NO DATA COMPRESSION IS ACHIEVED 4O BE ABLE TO USE THE ROBUSTNESS MEASURE FOR DESIGN OF ROBUST SOURCE CODES WITH SOME DEGREE OF DATA COMPRESSION WE WILL HAVE TO EXCLUDE SOME OF THE POSSIBLE CODES 4HIS CAN BE DONE BY DENING A SET OF ADMISSIBLE CODES 5 DUE TO SOME CRITERION 4HE CRITERION CAN FOR INSTANCE BE A LIMIT ON THE AVERAGE CODE COST 6 # O AT THE ESTIMATED PROBABILITY DISTRIBUTION O 7E THEREFORE SUGGEST THE FOLLOWING PROCEDURE FOR THE DESIGN OF ROBUST SOURCE CODES $ENE A SUITABLE ROBUSTNESS CRITERION h AND ND A CORRESPONDING ROBUSTNESS MEA SURE 1h # 3ELECT A SET OF ADMISSIBLE CODES 5 | 5 WHICH FULLS SOME CRITERION
NOS
&IND THE MOST ROBUST CODE #
WITHIN THE SET 5 IE

NOS
# )N THIS REPORT THE CODE #

NOS
@QFL@W 1h #
4R

IS CALLED THE OPTIMAL ROBUST CODE IN THE SET 5

4HE CODING COST FOR GRADIENT ROBUST CODES
4HE DENITION OF GRADIENT ROBUST CODES WAS DONE WITHOUT TAKING THE ACTUAL CODING COST UNDER CONSIDERATION (OWEVER WHEN WE HAVE FOUND OUR ROBUST CODE THERE ARE A NUMBER OF ALGEBRAIC AND GEOMETRIC INTERPRETATIONS THAT CAN BE DONE 3UPPOSE THAT WE HAVE AN ESTIMATED PROBABILITY DISTRIBUTION O AN OPTIMAL MINIMAL COST CODE # SUCH THAT 6 # O IS MINIMIZED AND AN OPTIMAL ROBUST CODE # SUCH THAT 1h # IS MAXIMIZED &URTHER ASSUME THAT THE CODE WORD COSTS FOR # AND # ARE GIVEN BY THE VECTORS K AND K RESPECTIVELY )F THE PROBABILITY DISTRIBUTION OF THE SOURCE CHANGES FROM THE ESTIMATED O TO P EXPRESSION GIVE US THE CODING COSTS
. . 1 1 . 1 . 1 -
6 # P
.
8
H(
PK
H .H

AND 6 # P
1
8
H(
PK
H 1H

1

3INCE THE AIM OF DESIGNING A ROBUST CODE IS TO OBTAIN A CODE # WITH A DIdERENT BEHAVIOUR THAN THE OPTIMAL MINIMAL COST CODE # WE ASSUME THAT K K 4HE TWO
. . 1

-
2OBUST SOURCE CODES
EXPRESSIONS AND ARE LINEAR FUNCTIONS ON 1 BUT WE ARE ONLY INTERESTED IN THE PARTS WHERE P a ! SCHEMATIC PICTURE OF THE TWO HYPER PLANES IS SHOWN IN &IGURE
-
&IGURE 3CEMATIC PICTURE OF THE AVERAGE CODING COST HYPER PLANES 4HE INDEPENDENT VARIABLE IS P 4HE INTERSECTION OF THE TWO PLANES IE FOR THE POINTS P a WHERE 6 # P 6 # P GIVE US A BOUNDARY BETWEEN TWO DISJOINTED PARTS OF THE SIMPLEX /N ONE SIDE IT IS MORE BENECIAL TO USE # AND ON THE OTHER SIDE IT IS MORE BENECIAL TO USE # /N THE BOUNDARY THE CHOICE OF CODE DOES NOT AdECT THE AVERAGE CODING COST )F WE CALCULATE THE DIdERENCE IN AVERAGE CODING COST IE
. 1 . 1
a6 P 6 # P ` 6 # P
. 1 1

IT IS OBVIOUS THAT # IS MORE BENECIAL THAN # IF a6 P AND VICE VERSA $EPENDING ON THE CONDITIONS WE ARE WORKING UNDER THE INTERPRETATION OF THE GOODNESS OF OUR ROBUST CODE MIGHT VARY (ERE ARE TWO POSSIBLE GOODNESS MEASURES q )F WE WANT TO LIMIT THE WORST CASE AVERAGE CODING COST WE MIGHT BE INTERESTED IN THE LARGEST VALUE ON 6 # P ON SOME PRE DENED SET OF POSSIBLE PROBABILITY DISTRIBUTIONS FOR THE SOURCE )N THIS CASE THE VALUE OF 6 # P IS IRRELEVANT
1 .
q )F WE HAVE A CONDITIONAL PROBABILITY ON THE SOURCE STATISTICS O P J O BASED ON THE ESTIMATED O WE MIGHT WANT TO LOOK AT THE EXPECTED GLOBAL REDUCTION OF CODING COST : 6 O O P J O a6 P CP
- `
2OBUST SOURCE CODES

)F THE GOODNESS MEASURE IS VERY IMPORTANT IE IF THE APPLICATION IS DEPENDING HEAVILY ON THE GOODNESS MEASURE A PARTICULAR ROBUSTNESS MEASURE 1h # MIGHT BEE NEEDED )F THIS IS NOT THE SITUATION SOME KIND OF GENERAL ROBUSTNESS MEASURE CAN BE USED INSTEAD

'RADIENT ROBUST ENTROPY CLOSE VARIABLE LENGTH CODES
)F WE USE THE DENITION OF GRADIENT ROBUST SOURCE CODES IT IS FAIRLY STRAIGHTFORWARD TO APPLY THE CONCEPT TO ENTROPY CLOSE VARIABLE LENGTH CODES /UR DENITION OF ENTROPY CLOSE CODES IS GIVEN BELOW $ENITION !N ENTROPY CLOSE VARIABLE LENGTH CODE # P IS A CODE WHERE THE CODE WORD COST EQUALS THE SELF INFORMATION OF THE CODED SOURCE SYMBOL )F # P IS DESIGNED FOR A PROBABILITY DISTRIBUTION P :P P a a a P < THEN THE COST VECTOR K :K K a a a K < ASSOCIATED WITH # P HAS THE FOLLOWING ENTRIES
3 3 -
K ` KNF P
J J
4HE ABOVE DENITION IMPLIES THAT AN ENTROPY CLOSE CODE # P DESIGNED FOR A SOURCE WITH PROBABILITY DISTRIBUTION P :P P a a a P < WILL HAVE AN AVERAGE CODING COST 6 # P THAT EQUALS THE ENTROPY ' P 4HE MOTIVATION FOR FOCUSING ON ENTROPY CLOSE CODES ARE THE ANALYTICAL EXPRESSIONS IN P FOR THE CODE WORD COSTS )F WE WORK WITH OTHER CODES FOR INSTANCE (UdMAN CODES THERE MIGHT NOT EXIST SUCH ANALYTICAL EXPRESSIONS )N #HAPTER WE WILL SEE THAT THE RESULTS OBTAINED FOR ENTROPY CLOSE CODES CAN TO SOME EXTENT BE APPLIED TO (UdMAN CODES &OR ENTROPY CLOSE CODES THE AVERAGE CODING COST EQUALS THE AVERAGE CODE WORD LENGTH OF THE CODE DUE TO THE FACT THAT WE INTERPRET THE COSTS THE ` KNF P S AS CODE WORD LENGTHS !S A CONSEQUENCE OF $ENITION THE ENTROPY CLOSE CODE ITSELF AND THE AVERAGE CODE COST AT THE ESTIMATED PROBABILITY DISTRIBUTION ARE COMPLETELY DENED BY THE CODE CONSTRUCTION PROBABILITY DISTRIBUTION P )N THE SEQUEL WE WILL CHANGE THE NOTATIONS FOR ENTROPY CLOSE CODES TO OBTAIN LESS COMPLICATED EXPRESSIONS 4HE CHANGES ARE
3 H
# P 6 # P O 1 # P
P 6 P O 1 P

)F THE GRADIENT ROBUSTNESS IS TO BE USED FOR THE DESIGN OF AN OPTIMAL GRADIENT ROBUST CODE WE NEED A SET OF ADMISSIBLE CODES 6 4HE RESTRICTION CHOSEN IS TO LIMIT THE EXPECTED CODING COST AT THE ESTIMATED PROBABILITY DISTRIBUTION O )F 6 P O IS RESTRICTED TO BE LESS THAN OR EQUAL TO 6 WE DENE THE SET OF ESTIMATION POINT LIMITED CODES ACCORDING TO THE DENITION BELOW $ENITION 4HE SET OF ESTIMATION POINT LIMITED %0, CODES IS DENED AS WHERE O IS THE ESTIMATED SOURCE PROBABILITY DISTRIBUTION 6 IS THE UPPER LIMIT ON THE EX PECTED CODING COST AT O AND a IS THE SET OF ALL POSSIBLE CODES PROBABILITY DISTRIBUTIONS FOR THE SOURCE
-
6 O FP a
J 6 P O v 6 G

2OBUST SOURCE CODES
!SSUME THAT WE HAVE AN ESTIMATE O OF A SOURCE PROBABILITY DISTRIBUTION &URTHER ASSUME IT IS DESIRED TO DESIGN AN OPTIMAL GRADIENT ROBUST ENTROPY CLOSE CODE WITHOUT CHANGING THE AVERAGE CODING COST AT O MORE THAN A FACTOR
p IE A GRADIENT ROBUST ENTROPY CLOSE CODE WHICH SATISES 6 P O v
p 6 O O
p ' O
WHERE p IS A POSITIVE REAL VALUED NUMBER SUCH THAT

p ' O KNF - 4HIS RESTRICTION ON p DOES NOT AdECT THE ANALYSIS IN A NEGATIVE SENSE SINCE THE ONLY RESTRICTION IS THAT ONLY DATA COMPRESSION CODES ARE CONSIDERED )N THE SEQUEL WE WILL REFER TO p AS THE DESIGN PARAMETER FOR GRADIENT ROBUST ENTROPY CLOSE CODES ! CONSEQUENCE OF THE ABOVE IS THAT THE XED LENGTH CODE IS NOT AN ADMISSIBLE CODE 4HE SET OF ADMISSIBLE CODES IS 6 O FP a
p p -
WHICH IS THE SET OF %0, CODES SEE $ENITION WHERE 6

p ' O 4HE OPTIMAL GRADIENT ROBUST CODE IN 6 O IS GIVEN BY P
NOS
J 6 P O v
p ' O G

@QFL@W 1 P
PS O
p

WHICH IS A VARIATION OF %XPRESSION 4HE FOLLOWING PROPERTIES OF THE OPTIMAL ROBUST CODE CAN BE OBTAINED FROM THE EXPRES SIONS AND 4HE INTERESTED READER CAN ND PROOFS IN !PPENDIX " 0ROPERTY !N OPTIMAL GRADIENT ROBUST ENTROPY CLOSE CODE EXISTS FOR EVERY ESTIMATED PROBABILITY DISTRIBUTION O WHICH IS AN INTERIOR POINT OF THE SIMPLEX a AND EVERY SET OF ENTROPY CLOSE %0, CODES 6 O
p
0ROPERTY 4HE SET OF ENTROPY CLOSE %0, CODES 6 O IS CONVEX

p
0ROPERTY 4HE OPTIMAL ROBUST ENTROPY CLOSE CODE IS FOUND BY MAXIMIZING THE GRADIENT ROBUSTNESS 1 P UNDER THE CONSTRAINT 6 P O
p ' O 4HE RST PROPERTY ENSURES THAT THERE IS AN OPTIMAL ROBUST CODE AND BOTH THE SECOND AND THE THIRD PROPERTY WILL TO SOME EXTENT SIMPLIFY THE OPTIMIZATION PROCEDURE FOR ND ING THE OPTIMAL ROBUST CODE 0ROPERTY IMPLIES THAT WHEREVER WE ARE IN THE SET 6 O WE HAVE FREE SIGHT TO THE OPTIMAL SOLUTION AND 0ROPERTY ALLOWS US TO CONCENTRATE ON THE BORDER OF THE SET 6 O
p p

%XPERIMENTAL EVALUATION OF GRADIENT ROBUSTNESS
4HE AIM OF THIS EXPERIMENTAL PART IS BOTH TO ELUCIDATE THE ROBUSTNESS CONCEPT AND TO EVALUATE SOME OF THE CONSEQUENCES IT HAS ON THE CODE CONSTRUCTION !S MENTIONED IN #HAPTER THE MAIN PROBLEM IS TO MAXIMIZE THE ROBUSTNESS 1 P ON A SET OF ADMISSIBLE
2OBUST SOURCE CODES

p

CODES 6 O 4HIS MAXIMIZATION CAN BE EXPRESSED AS A SYSTEM OF -

NON LINEAR EQUATIONS OBTAINED BY APPLICATION OF ,AGRANGIAN MULTIPLIERS TO ENSURE THE TWO CONSTRAINTS Pa AND
-

6 P O
p ' O
(OWEVER THE OBTAINED SYSTEM OF EQUATIONS IS NOT EASILY SOLVED FOR LARGE VALUES ON - 4HEREFORE THE EXPERIMENTS IN THIS CHAPTER ARE PERFORMED BY USING A MODIED GRADIENT SEARCH OVER THE SET 6 O &OR THE INTERESTED READER WE HAVE INCLUDED A DESCRIPTION OF THE NUMERICAL OPTIMIZATION METHOD IN !PPENDIX # ,ATER IN THIS CHAPTER WE ALSO PROPOSE A COMPUTATIONALLY CHEAP APPROXIMATION FORMULA FOR THE OPTIMAL SOLUTION
p

0ROPERTIES EXAMINED
)N THIS REPORT WE HAVE CONCENTRATED ON A FEW PROPERTIES TO EXAMINE 4HE BASIC QUESTIONS TREATED IN THE EXPERIMENTS ARE THE FOLLOWING q !RE THERE ANY GENERAL CHANGES IN THE CODE WORD LENGTHS THAT CAN BE SEEN IN OUR EXPERIMENTS WHEN GRADIENT ROBUSTNESS IS INTRODUCED q 7HAT IS THE DEPENDENCE BETWEEN ACHIEVED ROBUSTNESS AND THE DESIGN PARAMETER p q )S IT POSSIBLE TO ND ROBUST CODES THAT ARE CLOSE TO THE OPTIMAL ROBUST CODE WITHOUT USING A COMPLEX NUMERICAL ALGORITHM q )S IT POSSIBLE TO EXTEND THE RESULTS OBTAINED FOR ENTROPY CLOSE CODES TO (UdMAN CODES q !RE THERE OTHER CODE DESIGN METHODS THAT HAVE SIMILAR ROBUSTNESS PROPERTIES

0ROBABILITY DISTRIBUTIONS USED IN THE EXPERIMENTS
!S A DEMONSTRATION OF THE PROPERTIES OF GRADIENT ROBUST SOURCE CODES A NUMBER OF EXPER IMENTS ARE PERFORMED WHERE DIdERENT SITUATIONS ARE ANALYZED ! SET OF EXPERIMENTS CAN NOT BE EXHAUSTIVE BUT THE EXPERIMENTAL SITUATIONS WE HAVE CHOSEN WILL ELUCIDATE SOME MORE OR LESS GENERAL PROPERTIES 4HE EXPERIMENTS ARE BASED ON A FAMILY OF MIXED ,APLACIAN AND 'AUSSIAN LIKE PROB ABILITY DISTRIBUTIONS 4HE TWO BASIC STRUCTURES ARE THE FOLLOWING b c E m E m E m a a a E - ` m
3 + + + +
AND WHERE
E n
&
E n
&
E n
&
a a a E - ` n
&
E H m D
+
H -m
8
I (
I -m
H - `
AND
2OBUST SOURCE CODES
E H m D
&
- m
H
8
I (
- m
I
H - `
4HE PROBABILITY DISTRIBUTIONS E m AND E n CAN BE INTERPRETED AS PARAMETERIZED SAMPLED AND TRUNCATED VERSIONS OF THE CONTINUOUS ,APLACIAN AND 'AUSSIAN DISTRIBUTIONS 4HE TRUNCATION IS DONE IN A WAY THAT ONLY THE PART ABOVE THE MEAN IS USED 4HE PARAME TERS ARE IN BOTH CASES NORMALIZED WITH RESPECT TO THE NUMBER OF SOURCE SYMBOLS - SUCH THAT THE SAME SHAPE OF THE SAMPLED CONTINUOUS DISTRIBUTION WILL OCCUR FOR DIdERENT - S )N &IGURE E m AND E n ARE PRESENTED FOR A SYMBOL SOURCE WITH PARAMETERS AND RESPECTIVELY
+ & + &
&IGURE 0ROBABILITY DISTRIBUTIONS FOR VARIOUS PARAMETER VALUES 4HE PROBABILITY DISTRIBUTIONS CHOSEN FOR THE EXPERIMENTS ARE O O O O
$ $ $ $
E E E E
& + & +

E
E
E
E
+ +

' ' ' '
O O O O

:BITSYMBOL< :BITSYMBOL< :BITSYMBOL< :BITSYMBOL<
4HE PROBABILITY DISTRIBUTIONS O a a a O ARE MIXTURES OF LOW ENTROPY PARAMETERS AND AND MEDIUM ENTROPY DISTRIBUTIONS PARAMETERS AND 7E HAVE CHOSEN NOT TO CONSIDER PROBABILITY DISTRIBUTIONS WHERE THE ENTROPY IS CLOSE TO KNF SINCE SOURCES WITH SUCH DISTRIBUTIONS DO NOT GAIN MUCH FROM DATA COMPRESSION CODING

%XPERIMENTS ON OPTIMAL GRADIENT ROBUST CODES
)N THE SEQUEL WE WILL PRESENT SOME EXPERIMENTAL RESULTS TRYING TO ANSWER THE QUESTIONS RAISED IN 3ECTION 4HE RST QUESTION CONCERNED THE CHANGE IN CODE WORD LENGTH WHEN INTRODUCING A ROBUST CODE INSTEAD OF AN OPTIMAL COMPRESSION CODE !LL OUR EXPERIMENTS SHOW THE SAME PATTERN 4HE SHORT CODE WORDS ARE ALMOST UNCHANGED WHILE THE LONG CODE WORDS ARE SHORTENED ! VISUALIZATION OF THE CHANGES IN CODE WORD LENGTH IS PRESENTED IN &IGURE WHERE O IS USED AS THE PROBABILITY DISTRIBUTION OF THE SOURCE 4HE TWO
2OBUST SOURCE CODES

GRADIENT ROBUST CODES WERE DESIGNED USING DESIGN PARAMETERS p AND p 4HE LINE FOR WHICH p IS THE OPTIMAL COMPRESSION CODE
&IGURE #ODE WORD LENGTHS USING O 4HE RESULT PRESENTED IN &IGURE CAN BE INTERPRETED BY CONSIDERING THE FOLLOWING ARGUMENTS q THE OPTIMIZATION PROCEDURE TRIES TO MINIMIZE THE STANDARD DEVIATION OF THE CODE WORD LENGTHS THUS SHRINKING THE SPAN BETWEEN THE LONGEST AND SHORTEST CODE WORDS q AND THE CODING COST IS TO BE KEPT BELOW A CERTAIN LEVEL THUS FORCING MOST OF THE CHANGES TO TAKE PLACE WHERE THE PROBABILITIES ARE THE SMALLEST 4HE SECOND QUESTION WAS CONCERNED WITH THE DEPENDENCE BETWEEN ACHIEVED ROBUSTNESS AND THE DESIGN PARAMETER p 7E HAVE PERFORMED NUMERICAL EXPERIMENTS ON THE FOUR PROB ABILITY DISTRIBUTIONS O O WHERE p WAS SELECTED IN THE INTERVAL : < IE BETWEEN AND INCREASED AVERAGE CODING COST AT THE ESTIMATED PROBABILITY DISTRIBUTION )N &IGURE THE ACHIEVED ROBUSTNESS RELATIVE TO THE ROBUSTNESS OF THE OPTIMAL COMPRESSION CODE IS PRESENTED FOR DIdERENT VALUES ON p !S ILLUSTRATED IN &IGURE EVEN SMALL VALUES ON THE DESIGN PARAMETER p CAN GIVE QUITE A SUBSTANTIAL INCREASE OF THE ROBUSTNESS 4HIS EdECT IS DUE TO THE FACT THAT THERE ARE A LARGE NUMBER OF LONG CODE WORDS THAT CAN BE SHORTENED WITHOUT LOSING MUCH IN COMPRESSION

!N APPROXIMATION FORMULA FOR GRADIENT ROBUST CODES
4HE THIRD QUESTION IN 3ECTION IS CONCERNED WITH THE POSSIBILITY OF NDING A SIMPLIED PROCEDURE FOR THE DESIGN OF GRADIENT ROBUST CODES 4HE NUMERICAL ALGORITHM OUTLINED IN !PPENDIX # IS COMPUTATIONALLY INTENSE WHICH IS THE CASE OF ANY OPTIMIZATION ALGORITHM OPERATING ON A NON LINEAR PROBLEM WHERE MANY VARIABLES ARE INVOLVED AND WE ARE THEREFORE

3.5 3.0 2.5 2.0 1.5 1.0 0.5
2OBUST SOURCE CODES
Relative robustness
p1
0 0 0.02 0.04 0.06 Design parameter d
p2
p3
0.08
p4
0.10
&IGURE !CHIEVED ROBUSTNESS FOR DIdERENT VALUES ON p TRYING TO ND A SIMPLIED DESIGN METHOD "ASED ON THE FACT THAT THE OPTIMIZATION PROCEDURE IS MINIMIZING THE STANDARD DEVIATION OF THE CODE WORD LENGTHS UNDER THE CONSTRAINT 6 P O
p ' O WE HAVE CHOSEN TO APPLY A SIMILAR OPERATION ON THE PROBABILITY DISTRIBUTION /NE OF THE MOTIVATIONS FOR THIS IS THAT THE CODE WORD LENGTHS ARE STRICTLY DECREASING CONTINUOUS FUNCTIONS OF THEIR RESPECTIVE DESIGN PROBABILITIES 3INCE BOTH THE UNIFORM DISTRIBUTION O :- - a a a - < AND THE CORRESPOND ING ENTROPY CLOSE CODE WORD LENGTHS HAVE ZERO STANDARD DEVIATION WE TRY TO ND A GRADIENT ROBUST CODE ON THE STRAIGHT LINE FROM O TO O 4HE PROBABILITY DISTRIBUTIONS ON THE STRAIGHT LINE ARE GIVEN BY
3 TMH TMH
P O w ` w O
wO
TMH
w : <
p

p
4HE SET OF ADMISSIBLE CODES IS CONVEX 0ROPERTY HENCE THERE IS EXACTLY ONE w w O IN THE INTERVAL : < SUCH THAT P O w O IS A BOUNDARY POINT OF 6 O IE 6 `w O O
w O O
p p p TMH
O
p ' O

p
,ET EXPRESSION BE THE DENITION OF A FUNCTION HENCEFORTH REFERRED TO AS w O 4HE VALUE ON w O CAN EASILY BE CALCULATED BY APPLYING SOME METHOD FOR NDING SOLUTIONS TO ONE VARIABLE EQUATIONS EG THE BISECTION METHOD WITH STARTING POINTS w AND w )NSERTING w O INTO WE GET AN EXPRESSION FOR A DISTRIBUTION P CHARACTERIZING AN APPROXIMATION OF THE OPTIMAL SOLUTION P
+ 4 p w NOS
P `w O O
w O O
p p w
TMH

7E HAVE PERFORMED THE SAME EXPERIMENTS AS IN THE PREVIOUS SECTION ON THE FOUR DIS TRIBUTIONS O O WHERE THE APPROXIMATION FORMULA WAS USED INSTEAD OF THE OPTIMIZATION ALGORITHM !S A COMPARISON BETWEEN CODE WORD LENGTHS OF THE OPTIMAL
2OBUST SOURCE CODES

SOLUTION AND THE APPROXIMATIVE SOLUTION WE PRESENT THE CODE WORD LENGTHS FOR THE AP PROXIMATIVE SOLUTION IN &IGURE )F WE COMPARE THE CODE WORD LENGTHS WITH THE ONES PRESENTED IN &IGURE WE CAN OBSERVE THAT THE MEDIUM LENGTH CODE WORDS ARE THE ONES THAT DIdER THE MOST
&IGURE !PPROX SOLUTION CODE WORD LENGTHS USING O "Y EXPERIMENTAL EVALUATION WE OBSERVED THAT THE APPROXIMATION FORMULA HAVE ROUGHLY THE SAME EdECT AS THE OPTIMIZATION ALGORITHM 5SING THE SAME VALUES ON THE DE SIGN PARAMETER p AS IN &IGURE WE PERFORMED THE SAME EXPERIMENTS FOR THE APPROXIMA TIVE SOLUTION AS FOR THE OPTIMAL SOLUTION AND FOUND THAT THERE ARE ONLY MINOR DIdERENCES 4HE RELATIVE DIdERENCE IN ROBUSTNESS BETWEEN THE OPTIMAL GRADIENT ROBUST CODE AND THE APPROXIMATION GIVEN BY IS PRESENTED IN &IGURE WHERE THE RELATIVE DIdERENCE IN ROBUSTNESS IS GIVEN BY 1 P ` 1 P 1 P .OTE THAT 1 P v 1 P SINCE P IS THE OPTIMAL ROBUST SOLUTION &ROM &IGURE IT CAN BE CONCLUDED THAT FOR THE EXPERIMENTS PERFORMED THERE IS ONLY A SMALL LOSS IN OPTIMALITY IF WE USE THE APPROXIMATION INSTEAD OF THE MORE COMPLICATED OPTIMIZATION ALGORITHM 4HE ROBUSTNESS OF THE CODE GIVEN BY CAN BE EXPRESSED EXPLICITLY IN TERMS OF O AND p 4HE GRADIENT ROBUSTNESS IS
NOS w NOS w NOS NOS
1 P
R 0
r` KNF P (

wH

-
I (
-
KNF P
s
wI

WHERE P
wH
w O `w O O
p p H

2OBUST SOURCE CODES
&IGURE $IdERENCE BETWEEN OPTIMAL AND APPROXIMATIVE SOLUTION

!PPLICATION TO (UdMAN CODES
%NTROPY CLOSE CODES ARE SELDOM USED IN REAL SITUATIONS $IdERENT IMPLEMENTATIONS OF ARITHMETIC CODING WILL TO SOME EXTENT GIVE CODES WHICH ARE CLOSE TO THE ACTUAL ENTROPY OF A SOURCE AND IF THE PROBABILITY DISTRIBUTION HAS SOME SPECIAL PROPERTIES OTHER TYPES OF SOURCE CODERS WILL GIVE ENTROPY CLOSE CODING 3UCH A SITUATION WILL OCCUR FOR (UdMAN CODES IF THE PROBABILITY DISTRIBUTION IS ON THE FORM O WHERE M ARE INTEGERS (OWEVER THE OBTAINED RESULTS FROM THE PREVIOUS SECTIONS SHOULD BE APPLICABLE TO SOME EXTENT ON (UdMAN CODES 4HE APPLICATION TO (UdMAN CODES IS PERFORMED IN THE FOLLOWING MANNER
H MH H
q %STIMATE THE PROBABILITY DISTRIBUTION OF THE SOURCE AND CALL THE ESTIMATE O q 0ICK A SUITABLE VALUE FOR THE DESIGN PARAMETER p q &IND THE OPTIMAL GRADIENT ROBUST ENTROPY CLOSE CODE CHARACTERIZED BY P
NOS
q !PPLY (UdMANS ALGORITHM ON P
NOS
TO GET A (UdMAN CODE #
'
NOS
4HE EXPERIMENTS ARE PERFORMED ON THE SAME PROBABILITY DISTRIBUTIONS AS IN THE PRE VIOUS SECTIONS THUS ALLOWING US TO COMPARE THE RESULTS IN A STRAIGHTFORWARD MANNER &OR THE PROBABILITY DISTRIBUTIONS O O THE EXPERIMENTS SHOW THE SAME TYPE OF BEHAVIOUR FOR THE CODE WORD LENGTHS AS THE ENTROPY CLOSE CODES 4HE LONG CODE WORDS ARE SHORTENED AND THE SHORTER ONES ARE ALMOST UNCHANGED !S A DEMONSTRATION OF THE CHANGES WE HAVE CHOSEN TO PLOT THE CODE WORD LENGTHS SEE &IGURE FOR PROBABILITY DISTRIBUTION O WHERE THE DESIGN PARAMETER WAS CHOSEN TO p OPTIMAL COMPRESSION CODE p AND p )N &IGURE WE HAVE PLOTTED THE ACHIEVED ROBUSTNESS RELATIVE TO THE ROBUSTNESS OF THE (UdMAN CODE FOR THE ORIGINAL PROBABILITY DISTRIBUTION IE 1 # P 1 # O FOR DIdERENT VALUES ON THE DESIGN PARAMETER p "Y COMPARING WITH &IGURE WE CAN
' NOS '
2OBUST SOURCE CODES

&IGURE (UdMAN CODE WORD LENGTHS USING O SEE THAT HIGHER VALUES ON THE RELATIVE ROBUSTNESS SEEM TO BE ACHIEVED FOR (UdMAN CODES 3OMETHING ELSE NOTEWORTHY IS THAT THE CURVES OF RELATIVE ROBUSTNESS ACHIEVED FOR THE DIF FERENT PROBABILITY DISTRIBUTIONS CROSS EACH OTHER WHICH WAS NOT THE SITUATION FOR ENTROPY CLOSE CODES SEE &IGURE
&IGURE !CHIVED ROBUSTNESS FOR (UdMAN CODES 4HERE ARE SEVERAL DIdERENCES BETWEEN ENTROPY CLOSE CODES AND (UdMAN CODES THAT ARE AdECTING THE ACHIEVED ROBUSTNESS 7HILE THERE ARE AN UNCOUNTABLE NUMBER OF DIdERENT ENTROPY CLOSE CODES FOR EVERY - `SYMBOL SOURCE THERE ARE ONLY A NITE NUMBER OF DIdERENT (UdMAN CODES 4HE CODE WORD LENGTHS FOR (UdMAN CODES CAN ONLY TAKE INTEGER VALUES THUS GIVING LARGE RELATIVE DIdERENCES FOR HIGH PROBABILITY CODEWORDS %VEN THOUGH THERE

2OBUST SOURCE CODES
ARE DIdERENCES THE METHOD SEEMS TO WORK FAIRLY WELL FOR THE PROBABILITY DISTRIBUTIONS USED IN THE EXPERIMENTS %VEN THOUGH THE DESIGN PARAMETER p LIMITS THE CODING COST FOR ENTROPY CLOSE CODES AT THE PROBABILITY DISTRIBUTION O WE CAN NOT EXPECT THE SAME THING TO APPLY FULLY FOR (UdMAN CODES DUE TO THE DIdERENCES MENTIONED ABOVE )N &IGURE WE HAVE PRESENTED THE ACTUAL OBTAINED RELATIVE DIdERENCE 6 # P O ` 6 # O O 6 # O O FOR THE DIdERENT VALUES ON THE DESIGN PARAMETER p 4HE DOTTED LINE SYMBOLIZES THE RELATIVE DIdERENCE WE TRIED TO OBTAIN IE p
' NOS ' '
&IGURE !CTUAL DIdERENCES IN CODING COST FOR THE (UdMAN CODES 4AKING THE CURVES IN &IGURE UNDER CONSIDERATION WE GET ONE PART OF THE EXPLA NATIONS FOR THE FACT THAT THE CURVES IN &IGURE CROSS EACH OTHER "Y PRESENTING THE ACHIEVED ROBUSTNESS AGAINST THE ACTUAL INCREASE OF CODING COST INSTEAD OF THE DESIGN PA RAMETER p SEE &IGURE WE GET A MORE VALID PICTURE OF THE SITUATION (OWEVER THERE ARE STILL DIdERENCES BETWEEN THE ENTROPY CLOSE CODES AND (UdMAN CODES THAT CANNOT BE EXPLAINED FULLY BY THE DIdERENCE BETWEEN p AND THE ACTUAL INCREASE OF CODING COST %VEN THOUGH THERE ARE A NUMBER OF QUESTIONS FOR WHICH WE HAVE NO EXPLICIT ANSWERS EXCEPT THAT IT IS RELATED TO THE QUANTIZATION OF THE CODE WORD LENGTHS THE BASIC PROPERTIES FOR GRADIENT ROBUST ENTROPY CLOSE CODES SEEM TO SOME EXTENT BE VALID FOR (UdMAN CODES TOO 4HE BASIC PROPERTIES SUCH AS THE SHORTENING OF THE LONG CODE WORDS ARE THE SAME FOR BOTH THE (UdMAN CODES AND THE ENTROPY CLOSE CODES

! PERFORMANCE EVALUATION
!S MENTIONED IN 3ECTION THERE ARE SEVERAL POSSIBLE METHODS FOR EVALUATING THE GOODNESS OF A ROBUST SOURCE CODE )N THE CASE OF GRADIENT ROBUST SOURCE CODES THE MOST NATURAL WAY IS TO TRY TO EXTRACT INFORMATION FROM THE ROBUSTNESS MEASURE 1 # (OWEVER THE THERE ARE OTHER METHODS WHICH PROBABLY ARE EASIER TO INTERPRET )N THIS SECTION WE ARE FOCUSING ON THE FOLLOWING APPROACHES
2OBUST SOURCE CODES

&IGURE !CHIEVED ROBUSTNESS VS ACTUAL INCREASE OF CODING COST q 4HE RST APPROACH WHICH IS BASED ON THE ASSUMPTION THAT WE DO NOT KNOW THE DIRECTION OF CHANGE WE CALCULATE THE CODING COST FOR BOTH THE OPTIMAL COMPRESSION CODE 6 # P AND THE ROBUST CODE 6 # P IN THE MOST UNFAVOURABLE DIRECTION RESPECTIVELY AND SEE WHICH BEHAVES THE WORST 4HIS TYPE OF INTERPRETATION IS THE ONE CLOSEST TO THE DENITION OF GRADIENT ROBUSTNESS !N INTERPRETATION OF THIS MEASURE IS THAT WE LOOK AT TWO CODERS SEPARATELY AND CONSIDER THE WORST CASE DIRECTIONS RESPECTIVELY
. 1
q !S A SECOND APPROACH WE LOOK AT THE DIdERENCE IN CODING COST a6 P 6 # P ` 6 # P AND CHOSE THE WORST DIRECTION OF CHANGE !N INTERPRETATION OF THIS MEA SURE IS THAT WE LOOK AT TWO CODERS WORKING IN PARALLEL ON THE SAME SOURCE AND MEASURE THE DIdERENCE IN CODING COST IN A WORST CASE DIRECTION
. 1
q 4HE THIRD APPROACH IS BASED ON THE CODING COST AT THE UNIFORM DISTRIBUTION O :- - - < WHERE WE COMPARE THE CODING COSTS 6 # O AND 6 # O 4HIS MEASURE CAN BE USED IN SITUATIONS WHERE THERE ARE BURSTS OF NOISE IN THE SYS TEM SUCH THAT THE PROBABILITY DISTRIBUTION OF THE SOURCE BECOMES ALMOST AT !N APPLICATION WHERE THIS IS A CONCEIVABLE SITUATION IS DATA ACQUISITION IN COMBINATION WITH POORLY ATTACHED WIRING OR OTHER ARTIFACT SOURCES
TMH 1 3 . TMH
TMH
)N THE RST APPROACH WE LOOK AT THE EXPRESSIONS FOR THE CODING COSTS IN THE MOST UNFAVOURABLE DIRECTIONS D AND D RESPECTIVELY 4HE CODING COST CAN BE EXPRESSED AS
. 1
6 C 6 # O
CD
. .
6 # O
C 1 # C 1 #

AND 6 C 6 # O
CD
1 1 1
6 # O

2OBUST SOURCE CODES
&IGURE 0ERFORMANCE MEASURED IN THE WORST DIRECTIONS RESPECTIVELY WHERE C IS THE %UCLIDEAN DISTANCE FROM O SEE &IGURE "Y DENITION OF %0, CODES WE KNOW THAT
p 6 # O 6 # O HENCE CAN BE EXPRESSED AS C 6 C
p 6 # O
1 #
. 1 1 . 1
%VEN THOUGH THE MOST UNFAVOURABLE DIRECTION MOST CERTAINLY IS DIdERENT FOR THE TWO CODES WE CAN BY SETTING 6 C 6 C SAY THAT THE POINT OF BREAK EVEN LIES AT A DISTANCE 6 # O 1 # 1 # C p 1 # ` 1 #
. 1 . . 1 6 1 .
7E SHOULD BE SOMEWHAT CAREFUL WITH THE INTERPRETATION OF C SINCE IT IS BASED ON TWO DIdERENT DIRECTIONS OF CHANGE OF THE SOURCE PROBABILITY DISTRIBUTION (OWEVER WHEN THE %UCLIDEAN DISTANCE BETWEEN THE ACTUAL SOURCE PROBABILITY DISTRIBUTION P AND THE ESTIMATED PROBABILITY DISTRIBUTION O EXCEED THE VALUE ON C IN DIdERENT DIRECTIONS THE WORST CASE PERFORMANCE MIGHT BE BETTER IF WE USE THE ROBUST CODE 4O ILLUSTRATE THIS CONCEPT WE PRESENT 6 C p AND 6 C FOR p AND p 4HE PROBABILITY DISTRIBUTION WE CHOSE TO WORK WITH ARE O &IGURE AND O &IGURE
6 6 . 1
4HE SECOND APPROACH CHOSEN FOR THE PERFORMANCE EVALUATION IS BASED ON THE DIdERENCE IN CODING COST BETWEEN THE TWO CODES WHICH CAN BE EXPRESSED AS
-
a6 P 6 # P ` 6 # P
. 1
8
H(
P aK
H H

WHERE K AND K ARE THE COST VECTORS ASSOCIATED WITH THE CODES " AND " RESPECTIVELY AND aK K ` K 4HE DIRECTIONAL DERIVATIVE IN THE WORST CASE DIRECTION D OF THIS FUNCTION CAN BE CALCULATED BY USING THE SAME LINEAR ALGEBRA METHODS AS IN 3ECTION
. 1 . 1 H .H 1H
2OBUST SOURCE CODES

&IGURE 6 C AND 6 C IN THE WORST CASE DIRECTION RESPECTIVELY USING O "REAK EVEN POINTS WHERE THE DASHED LINES CROSS THE SOLID LINE
. 1
AND !PPENDIX ! (ENCE
V U U8 ` a 6 P 6 O
GD ` 6 O L@W L@W KHL T aK ` aK 6 D 6 G (
D D G H H

WHERE aK EXPRESSED AS
I (
aK 4HE DIdERENCE IN CODING COST IN THE WORST DIRECTION CAN BE

I
V U U8 ` a a6 C 6 # O ` 6 # O
CT aK ` aK
. 1 H H(

WHERE C IS THE %UCLIDEAN DISTANCE FROM THE ESTIMATED PROBABILITY DISTRIBUTION O SEE &IGURE %XPRESSION CAN BE REDUCED TO V U U8 ` a a6 C `p6 # O
CT aK ` aK
. H H(
BY TAKING THE DENITION OF %0, CODES INTO CONSIDERATION 4HE POINT OF BREAK EVEN CAN BE OBTAINED BY SETTING a6 C AND WE GET C p Q0 6 # O àK ` aKa (
. H

&ROM EXPRESSION WE GET THE DISTANCE C ABOVE WHICH IT IS MORE BENECIAL TO USE THE ROBUST CODE IF THE DIRECTION OF CHANGE IS THE WORST POSSIBLE 4HE DIdERENCE BETWEEN

2OBUST SOURCE CODES
&IGURE 6 C AND 6 C IN THE WORST CASE DIRECTION RESPECTIVELY USING O "REAK EVEN POINTS WHERE THE DASHED LINES CROSS THE SOLID LINE
. 1
THE DISTANCE AND THE DISTANCE IS THAT WE MAKE THE CALCULATIONS IN THE SAME DIRECTION FOR BOTH CODES IN THE LATTER ONE 4O ILLUSTRATE THIS SECOND METHOD WE PRESENT a6 C FOR p AND p 4HE PROBABILITY DISTRIBUTIONS ARE THE SAME AS IN THE RST METHOD OF PERFORMANCE EVALUATION IE O &IGURE AND O &IGURE 4HE LAST AND PROBABLY THE MOST EASILY INTERPRETED APPROACH ON THE GOODNESS OF THE ROBUST CODE IS WHERE WE ONLY LOOK AT THE CODING COST AT THE UNIFORM PROBABILITY DISTRIBUTION 7E HAVE CHOSEN TO MEASURE THE GOODNESS AS THE DIdERENCE IN CODING COSTS IE a6 6 # O `6 # O
TMH . TMH 1 TMH
5SING EXPRESSION WE HAVE ILLUSTRATED THE DIdERENCE IN AVERAGE CODING COST AT THE AT PROBABILITY DISTRIBUTION SEE &IGURE WHERE WE USE THE SAME PROBABILITY DISTRIBUTIONS AS BEFORE 4HE CURVES ARE DETERMINED BY USING DIdERENT VALUES ON THE DESIGN PARAMETER p 4HE DASHED LINES SYMBOLIZE THE DESIGN PARAMETERS USED IN THE PREVIOUS TWO METHODS FOR ANA LYZING THE PERFORMANCE

$ISCUSSION AND CONCLUSIONS
)N THIS REPORT A LARGE NUMBER OF DIdERENT ASPECTS HAVE BEEN INVESTIGATED REGARDING ROBUST CODES IN GENERAL AND GRADIENT ROBUST CODES IN PARTICULAR 4HE INVESTIGATIONS HAVE BEEN PERFORMED BOTH THROUGH ANALYTICAL EXPRESSIONS AND THROUGH EXPERIMENTAL EVALUATIONS 3OME ASPECTS OF THE INVESTIGATION THAT SHOULD BE NOTED ARE THE FOLLOWING q 4HE GRADIENT ROBUST CODES EXAMINED ARE ONLY ONE CLASS OF ROBUST CODES AND MANY APPLICATIONS WOULD PROBABLY GAIN MORE FROM OTHER TYPES OF ROBUSTNESS MEASURES
2OBUST SOURCE CODES

&IGURE 0ERFORMANCE MEASURED IN THE WORST DIRECTION OF THE DIdERENCE 4HE QUESTION OF HOW TO DECIDE WHICH ROBUSTNESS MEASURE IS THE BEST FOR A PARTICULAR SITUATION IS AN OPEN QUESTION AT THIS POINT q 4HE EXPERIMENTS ONLY CONSIDER SOURCES WHERE THE NUMBER OF SYMBOLS IS - 4HE REASON FOR THIS IS THAT WE WANTED TO USE AS MANY SYMBOLS AS POSSIBLE WITHOUT INCREASING THE EXECUTION TIME FOR THE NUMERICAL ALGORITHMS TOO MUCH q 4HE EXPERIMENTS ARE PERFORMED ON A LIMITED NUMBER OF DIdERENT PROBABILITY DIS TRIBUTIONS WHICH IS A FACT THAT WILL HAVE TO BE TAKEN INTO CONSIDERATION WHEN THE RESULTS ARE INTERPRETED (OWEVER THE EXPRESSIONS USED ARE CONTINUOUS AND BEHAVES FAIRLY WELL EXCEPT FOR THE ONES IN THE APPLICATION TO (UdMAN CODES WHERE THE CODE WORD LENGTHS ARE NON ANALYTICAL INTEGER APPROXIMATIONS OF THE ONES OBTAINED FOR ENTROPY CLOSE CODES q 7E HAVE NOT SHOWN THAT THE OPTIMAL GRADIENT ROBUST %,0 CODE IS UNIQUE 4HERE MIGHT IN FACT BE SEVERAL LOCAL OPTIMA IN THE EXPRESSIONS USED FOR ROBUST CODES OF WHICH WE ONLY ND ONE WHEN USING THE OPTIMIZATION PROCEDURE DESCRIBED IN !PPENDIX # %VEN THOUGH WE HAVE NOT BEEN ABLE TO SHOW THAT THE OPTIMAL ROBUST CODE IS UNIQUE OUR HYPOTHESIS IS THAT IN FACT IT IS UNIQUE 4HE PROBLEM IN SHOWING THE UNIQUENESS GOES BACK TO SHOWING THAT THE SET OF CODES WITH GRADIENT ROBUSTNESS GREATER THAN OR EQUAL TO SOME XED VALUE IS CONVEX 7E WILL CONTINUE WORKING ON THIS PROBLEM )N THE EXPERIMENTS WE HAVE OBSERVED THAT THE GRADIENT ROBUST CODES FOR THE PROB ABILITY DISTRIBUTIONS EXAMINED SEE #HAPTER HAVE CERTAIN USEFUL PROPERTIES 4HE PROPERTIES WE HAVE TRIED TO ELUCIDATE ARE MORE OR LESS CONNECTED TO CERTAIN SITUATIONS WHICH MIGHT ARISE IN REAL APPLICATIONS (OPEFULLY WITHOUT OVERRATING THE CONTENTS OF THE ANALYSIS AND THE EXPERIMENTS THE FOLLOWING CONCLUSIONS ARE MADE q /UR DENITION OF ROBUSTNESS MEASURE FOR SOURCE CODES AND THE PROPOSED DESIGN PROCEDURE FOR OPTIMAL ROBUST SOURCE CODES PROVIDE A FRAMEWORK FOR A LARGE NUMBER

2OBUST SOURCE CODES
&IGURE a6 C FOR p AND p USING O "REAK EVEN POINTS WHERE THE DASHED LINES CROSSES THE SOLID LINE OF DIdERENT CLASSES OF ROBUST CODES ! CERTAIN CLASS OF ROBUST SOURCE CODES CAN BE CREATED BY FOLLOWING THE DIRECTIONS OF #HAPTER q "Y INTRODUCING GRADIENT ROBUST CODES WE HAVE OBSERVED THAT THE MAXIMAL CODE WORD LENGTH OF THE CODE IS REDUCED &OR LOW ENTROPY SOURCES THE DECREASE IN MAXIMAL CODE WORD LENGTH IS SUBSTANTIAL EVEN FOR SMALL LOSSES IN ACHIEVED COMPRESSION )N SOME APPLICATIONS IT IS BENECIAL TO REDUCE THE LENGTH OF THE LONGEST CODE WORDS ! GREAT ADVANTAGE OF SHORT MAXIMAL CODE WORD LENGTH IS THAT THE MEMORY SIZE REQUIRED TO STORE CODE BOOKS CAN BE REDUCED SUBSTANTIALLY SEE EG ;=;= AND ;= ! NUMBER OF PAPERS HAVE BEEN PUBLISHED WHERE THE CONSTRUCTION OF LENGTH LIMITED (UdMAN CODES AND PATH LENGTH LIMITED BINARY SEARCH TREES ARE ADDRESSED SEE EG ;=;= AND ;= q )N OUR EXPERIMENTS THE GRADIENT ROBUSTNESS OF CODES INCREASES SUBSTANTIALLY WITH MINOR LOSS OF OPTIMALITY OF THE COMPRESSION RATE THUS MAKING GRADIENT ROBUST CODES LESS SENSITIVE TO CHANGES IN THE PROBABILITY DISTRIBUTION OF THE SOURCE "Y TAKING THIS PROPERTY INTO CONSIDERATION WE CAN CONSTRUCT VARIABLE LENGTH CODES BASED ON FAIRLY POOR ESTIMATES OF THE SOURCE PROBABILITY DISTRIBUTION WITH AN ACCEPTABLE BEHAVIOUR 4HE ESTIMATION OF THE SOURCE PROBABILITY DISTRIBUTION FROM THE POINT OF VIEW OF MINIMUM REDUNDANCY SOURCE CODING HAS BEEN ADDRESSED IN EG ;= q 7E PRESENT AN APPROXIMATION FORMULA FOR NDING AN APPROXIMATION OF THE OPTIMAL GRADIENT ROBUST CODE 4HE COMPUTATION OF THE APPROXIMATIVE SOLUTION IS VERY FAST COMPARED TO THE COMPUTATION OF THE OPTIMAL SOLUTION USING THE MODIED GRADIENT SEARCH 4HE APPROXIMATION FORMULA ONLY REQUIRES THE SOLUTION OF A ONE VARIABLE EQUATION WHEREAS THE OPTIMIZATION PROCEDURE IS WORKING WITH THE SAME NUMBER OF VARIABLES AS THE NUMBER OF SOURCE SYMBOLS )N THE EXPERIMENTS THE DIdERENCE IN ROBUSTNESS BETWEEN THE OPTIMAL SOLUTION AND THE APPROXIMATION WAS SMALL
2OBUST SOURCE CODES

&IGURE a6 C FOR p AND p USING O "REAK EVEN POINTS WHERE THE DASHED LINES CROSSES THE SOLID LINE q 7HEN APPLYING THE EXPERIMENTS ON GRADIENT ROBUST ENTROPY CLOSE CODES TO (UdMAN CODES WE HAVE OBSERVED THAT THE PROPERTIES ARE ROUGHLY THE SAME 4HE OBTAINED RO BUSTNESS EVEN SEEMS TO BE SLIGHTLY HIGHER FOR (UdMAN CODES (OWEVER THE METHOD WE USED FOR INTRODUCING GRADIENT ROBUSTNESS TO (UdMAN CODES IS SOMEWHAT ELUSIVE WHEN IT COMES TO DEPENDENCE BETWEEN THE DESIGN PARAMETER p AND THE ACTUAL VALUE ON THE INCREASE OF THE CODING COST 4HIS PROBLEM IS DUE TO THE INTEGER QUANTIZATION OF THE CODE WORD LENGTHS OF (UdMAN CODES q 4HE PERFORMANCE EVALUATIONS DONE ARE BASED ON THREE DIdERENT APPROACHES FOR MEASURING THE GOODNESS OF A GRADIENT ROBUST CODE /NE OF THE APPROACHES IS BASED ON THE ROBUSTNESS MEASURE ITSELF AND THE OTHER TWO ARE INCLUDED FOR COMPARISON AND BECAUSE THEY PROBABLY ARE EASIER TO INTERPRET )N THE EXPERIMENTS IT WAS EVIDENT THAT SMALL CHANGES IN THE AVERAGE CODING COST CAN GIVE SUBSTANTIAL ADVANTAGES IN DIdERENT WORST CASE SITUATIONS 4HE LAST APPROACH WE USED IN THE PERFORMANCE ANALYSIS WAS BASED ON THE DIdERENCE IN AVERAGE CODING COST AT THE AT PROBABILITY DISTRIBUTION ! NOTEWORTHY PROPERTY OF THIS LAST METHOD IS THAT ONE OF THE POSSIBLE ROBUSTNESS MEASURES BASED ON IT IMPLIES A MINIMIZATION OF THE ARITHMETIC MEAN OF THE CODE WORD LENGTHS WHEREAS THE GRADIENT ROBUSTNESS MEASURE IMPLIES A MINIMIZATION OF THE UNWEIGHTED STANDARD DEVIATION OF THE CODE WORD LENGTHS IE WE HAVE A CLOSE CONNECTION TO THE WELL KNOWN QUANTITIES ARITHMETIC MEAN AND STANDARD DEVIATION
! MORE THOROUGH ANALYSIS AND THE INTRODUCTION OF OTHER ROBUSTNESS MEASURES ARE UNDER CONSIDERATION SINCE THE RESULTS OBTAINED IN THIS REPORT SHOW THAT THE INTRODUCTION OF ROBUSTNESS CAN GIVE CERTAIN BENECIAL PROPERTIES TO SOURCE CODES AT A LOW COST

2OBUST SOURCE CODES
&IGURE a6 FOR DIdERENT VALUES ON THE DESIGN PARAMETER p USING O AND O $ASHED LINES SYMBOLIZE THE PARAMETER VALUES USED IN THE PREVIOUS EVALUATIONS
TMH
!
!NALYSIS OF THE GRADIENT ROBUSTNESS 0 )N 3ECTION WE STATE THAT THE MAXIMIZATION OF D K OVER THE SET % IS GIVEN BY
-
V U U8 8 8 T L@W DK K ` K D6 ( - ( (
H H H H H I
H(
H H

I

WHERE THE SET % IS GIVEN BY % D1 J

-
8
H(
D >
H
8
H(
D
H

4O PROVE WE USE STANDARD LINEAR ALGEBRA OPERATIONS WHICH BE FOUND IN 0 CAN MOST TEXTBOOKS ON THE SUBJECT EG ;= &IRST WE OBSERVE THAT D K IS AN INNER ( PRODUCT HD KI WHERE D IS CONSTRAINED TO BELONG TO A SUBSPACE OF 1 &ROM LINEAR ALGEBRA WE KNOW THAT N THE MAXIMIZATION O CAN BE DONE BY ORTHOGONAL PROJECTION OF K ONTO 0 THE SUBSPACE D 1 J D FOLLOWED BY A CALCULATION OF THE LENGTH OF THE ( PROJECTION 4HE ORTHOGONAL PROJECTION KD OF K IS GIVEN BY
H H H H H
WHERE M y : a a a < IS THE ORTHOGONAL DIRECTION OF THE SUBSPACE &URTHER THE LENGTH OF KD IS V U U8 8 JKD J JK ` HM KI MJ T K ` K - ( (
3 H I H I
KD K ` HM KI M
WHICH COMPLETES THE PROOF OF
2OBUST SOURCE CODES

"
0ROPERTIES OF GRADIENT ROBUST ENTROPY CLOSE %0, CODES
)N THIS APPENDIX WE ARE PROVIDING PROOFS FOR THE PROPERTIES PRESENTED IN 3ECTION 8 &OR CONVENIENCE WE CONSIDER KNF W TO BE THE NATURAL LOGARITHM /THER BASES WILL ONLY CHANGE THE EXPRESSIONS BY A CONSTANT $ETAILED INFORMATION ON THE MATHEMATICAL TOOLS USED IN THE PROOFS CAN BE FOUND IN ;= 0ROPERTY !N OPTIMAL GRADIENT ROBUST ENTROPY CLOSE CODE EXISTS FOR EVERY ESTIMATED PROBABILITY DISTRIBUTION O WHICH IS AN INTERIOR POINT OF THE SIMPLEX a AND EVERY SET OF ENTROPY CLOSE %0, CODES 6 O 0ROOF 7E WANT TO SHOW THAT THERE IS A P 6 O MAXIMIZING 1 P ON 6 O 3INCE
p p p
1 P R 0
r ` KNF P
(
H

-
I (
-
KNF P
s

IT IS SUbCIENT TO SHOW THAT THE EXPRESSION

-
) P
p
8
H(
8 ` KNF P
KNF P - (
H I

I

p
ATTAINS ITS MINIMUM ON 6 O 3INCE 6 P O v

p ' O ON 6 O AND O IT IS CLEAR THAT NO P CAN BE ZERO DUE TO THE EXPRESSIONS AND ) O IS A CONTINUOUS FUNCTION ON THE INTERIOR OF THE SIMPLEX a HENCE ESPECIALLY ON 6 O &OR P O THE AVERAGE CODING COST ATTAINS ITS MINIMUM 6 O O ' O 4HE SET 6 O IS COMPACT SINCE IT IS THE INVERSE IMAGE ) & OF A COMPACT SET & :' O
p ' O < AND 6 O | a ! CONTINUOUS FUNCTION ON A COMPACT SET ATTAINS ITS EXTREMES ON THE SET IE ) O ATTAINS ITS MINIMUM ON 6 O (ENCE AN OPTIMAL ROBUST %0, CODE EXISTS IN 6 O
H H p p p p p
0ROPERTY 4HE SET OF ENTROPY CLOSE %0, CODES 6 O IS CONVEX 0ROOF 7E KNOW THAT P 6 O 6 P O v
p ' O 4HE SET 6 O IS CONVEX IF AND ONLY IF P ` w Q
wR 6 O FOR ALL w Q 6 O AND R 6 O 4HE AVERAGE CODING COST AT O IS
p p p p p p -
6 P O `
8
H(
O KNF ` w Q
wR v ` w 6 Q O
w6 R O v
p ' O
H H H p
DUE TO THE CONVEXITY OF ` KNF W (ENCE THE SET 6 O IS CONVEX
0ROPERTY 4HE OPTIMAL ROBUST ENTROPY CLOSE CODE IS FOUND BY MAXIMIZING THE GRADIENT ROBUSTNESS 1 P UNDER THE CONSTRAINT 6 P O
p ' O

2OBUST SOURCE CODES
0ROOF &ROM 0ROPERTY WE KNOW THAT THE DESIGN PROBABILITY DISTRIBUTION FOR THE OPTIMAL ROBUST CODE CAN BE FOUND IN 6 O )F THERE IS AN INTERIOR MINIMUM POINT P THE GRADIENT R) P MUST BE ZERO OR ORTHOGONAL TO THE PROBABILITY SIMPLEX a SEE EG ;= 4HE GRADIENT IS P
p )
R) P WHERE
P
)

P- `

) P KNF P -P
J J
P
-
H(

H -

)F THE GRADIENT IS TO BE ZERO OR ORTHOGONAL TO THE SIMPLEX a IT IS NECESSARY THAT ALL COMPONENTS OF R) O ARE EQUAL SINCE THE ORTHOGONAL DIRECTION OF THE SIMPLEX IS b c a a a )N THE CASE WHERE ALL P ARE EQUAL IT IS TRIVIAL THAT THE GRADIENT IS ZERO BUT THIS POINT DOES NOT BELONG TO 6 O )N THE CASE WHERE ALL P ARE NOT EQUAL WE EXPRESS THE ORTHOGONALITY AS P P J K KNF 1 KNF 1 -P -P P P ( (
3 H p H J K J H H K H H
3INCE ALL P ARE NOT EQUAL THERE MUST BE A LARGEST AND A SMALLEST ONE $ENOTE THESE BY Pvd AND Pvrw )F THE EQUALITY HOLDS FOR ALL J AND K THEN IT MUST HOLD FOR J AND K CORRESPONDING TO Pvrw AND Pvd IE P P KNF 1 vrw KNF 1 vd P - Pvrw - Pvd P ( (
H H H H H
#ONSIDERING THAT
Pvrw 1
H(
P AND 1 vd
H(

H
WE HAVE OBTAINED A CONTRADICTION (ENCE THERE ARE NO INTERIOR MINIMUM POINTS AND THE DESIGN PROBABILITY P FOR THE OPTIMAL GRADIENT ROBUST CODE MUST BE LOCATED ON THE BOUNDARY OF 6 O IE WHERE 6 P O
p ' O
#
! NUMERICAL OPTIMIZATION METHOD

p
4HE METHOD USED FOR THE OPTIMIZATION IS DESIGNED FOR THE CONDITIONS CONSIDERED IN THIS REPORT 3INCE WE KNOW FROM 0ROPERTY AND 0ROPERTY THAT THE SET 6 O IS CONVEX AND THAT THE OPTIMIZING POINT RESIDES ON THE BOUNDARY WE DEVELOPED A MODIED VERSION OF A GRADIENT SEARCH ALGORITHM FOR NDING THE OPTIMIZING POINT 4HE CONNECTION TO GRADIENT SEARCH IS OBVIOUS BUT THERE ARE SOME DIdERENCES ONE SHOULD BE AWARE OF )N &IGURE A VISUALIZATION IS DONE OF THE OPTIMIZATION SITUATION 4HE DASHED TRI ANGULAR SHAPED CURVES ARE CONTOURS OF THE ROBUSTNESS MEASURE 1 P SEE THE EXPRESSIONS
2OBUST SOURCE CODES

&IGURE !N ILLUSTRATION OF THE OPTIMIZATION SITUATION 4HE INDEPENDENT VARIABLE IS O AND 4HE DASHED ELLIPTICAL CURVES ARE CONTOURS OF THE CODING COST FUNCTION 6 P O SEE THE EXPRESSIONS AND 4HE OPTIMUM POINT P IS GIVEN BY EXPRESSION REVISITED HERE
NOS
NOS
@QFL@W 1 P
PS O
p

p
7E KNOW FROM 0ROPERTY THAT THE P IS TO BE FOUND ON THE BOUNDARY OF 6 O REPRESENTED BY THE SOLID ELLIPTICAL CURVE IN &IGURE )NSTEAD OF MOVING ALONG THE BOUNDARY OF 6 O WE HAVE CHOSEN A VARIATION OF THE GRADIENT SEARCH METHOD 4HE STEPS OF THE ITERATIVE SEARCH FOR P ARE DESCRIBED BELOW )T SHOULD BE NOTED THAT THE NOTATIONS Q0 0 W FOR THE INNER PRODUCT AND THE NORM ARE HW XI W X AND KWK ( ( RESPECTIVELY
NOS p NOS H H H H H
#HOOSE A STARTING POINT P 6 O AN INCREMENTAL STEP SIZE w A DIRECTION CHANGE INTEGER * AND INITIALIZE THE COUNTER M
p
)F M IS NOT AN INTEGER MULTIPLE OF * THEN SELECT THE DIRECTION OF CHANGE TO BE C R1 P

M M

OTHERWISE SELECT THE DIRECTION OF CHANGE TO BE C R6 P O `

M
HR6 P O R1 P I a R1 P KR1 P K
M M M M

4HE NEXT POINT P

M
2OBUST SOURCE CODES IS GIVEN BY P

M M
P
waq aC
M M M M M M p

WHERE q IS A POSITIVE SCALAR FOR WHICH P

q a C IS A BOUNDARY POINT OF 6 O )NCREMENT M AND GO TO STEP IF P IS NOT CLOSE ENOUGH TO THE OPTIMIZING POINT P
M NOS
4HE PRINCIPAL DIdERENCE BETWEEN A GRADIENT SEARCH BASED ON THE DIRECTION ALONE AND A GRADIENT SEARCH BASED ON ALTERNATION BETWEEN DIRECTION AND IS SHOWN IN &IGURE WHERE PATH ) AND )) SYMBOLIZE THE TWO METHODS RESPECTIVELY !S ILLUSTRATED IN THE GURE A GRADIENT SEARCH BASED ONLY ON R1 P WILL IN MOST CASES NOT CONVERGE TO P 4HIS OBSERVATION HAS BEEN VERIED THROUGH EXPERIMENTS DURING THE DEVELOPMENT OF THE ALGORITHM
M NOS
&IGURE )LLUSTRATION OF THE OPTIMIZATION PATHS FOR TWO DIdERENT GRADIENT METHODS 4HE STARTING POINT P IS CHOSEN AS A ATTENED VERSION OF THE ESTIMATED PROBABILITY DIS TRIBUTION O BY MOVING TOWARDS THE RECTANGULAR PROBABILITY DISTRIBUTION Q : a a a < ALONG A STRAIGHT LINE IE P ` | O
|Q
3
WHERE | IS CHOSEN SUCH THAT 6 P O v

p ' O !N ANALYTICAL EXPRESSION FOR SUCH A | CAN BE OBTAINED BY OBSERVING THAT 6 ` | O
|Q O v ` | ' O
| KNF - (ENCE IF WE CHOSE | TO BE |
p

p' O KNF - ` ' O
IT IS ENSURED THAT P 6 O 4HE ABOVE ALGORITHM IS NOT SELF ADJUSTING AND THEREFORE THE PARAMETERS w AND * MIGHT HAVE TO BE CHANGED TO OBTAIN A PROPER CONVERGENCE OF P TO P 4HE CONVERGENCE CAN BE MONITORED BY PLOTTING THE VALUES OF
M NOS
T
M
HR6 P O R1 P I KR1 P K a KR6 P O K

M M M M

2OBUST SOURCE CODES AND

V 6 P O
M M NOS NOS

NOS
DURING THE ITERATIONS OF THE ALGORITHM !T THE OPTIMUM P R6 P O AND R1 P ARE POINTING IN THE SAME DIRECTION SEE EG;= &URTHER WE KNOW THAT THE OPTIMIZING POINT RESIDES ON THE BOUNDARY OF 6 O IE WHERE 6 P O
p ' O (ENCE IF T AND V
p ' O WE HAVE AN INDICATION ON THE CONVERGENCE P P )N THE EXPERIMENTS PERFORMED THE ALGORITHM SEEMS TO CONVERGE WITHIN A FEW HUNDRED ITERATIONS WHEN THE NUMBER OF SOURCE SYMBOLS IS - w AND * )N SOME EXTREME CASES * WAS CHOSEN TO A HIGHER VALUE TO OBTAIN CONVERGENCE ! MORE THOROUGH ANALYSIS OF THE OPTIMIZATION ALGORITHM IS NOT PERFORMED IN THIS REPORT SINCE IT IS OUT OF THE SCOPE OF THIS REPORT 4HE ONLY REASON FOR DESIGNING A SPECIAL ALGORITHM IN THIS CASE IS A MATTER OF SPEEDING UP THE CALCULATIONS DURING THE EXPERIMENTS
p M M M NOS

2OBUST SOURCE CODES
0ART ! PERFORMANCE ANALYSIS OF LOOK UP TABLE BASED VARIABLE LENGTH DECODERS

!BSTRACT )N THIS REPORT WE ARE PRESENTING TWO DECODING ALGORITHMS FOR BINARY VARIABLE LENGTH CODES BASED ON A LOOK UP TABLE APPROACH 4HIS TYPE OF ALGORITHMS ARE FAMILIAR TO SOFTWARE DESIGNERS BUT THE AUTHORS HAVE NOT BEEN ABLE TO ND ANY SCIENTIC MATERIAL WHERE THE DECODING SPEED IS ANALYSED 4HE ALGORITHMS WILL DECODE ONE OR SEVERAL SOURCE SYMBOLS FOR EACH CYCLE OF THE ALGORITHM 4HE LOOK UP TABLE APPROACH IS BASED ON A COMPLETION OF THE TREE REPRESENTATION OF THE VARIABLE LENGTH CODE $ECODING IS DONE BY INDEXING A TABLE WITH A XED LENGTH BLOCK FROM THE CONCATENATED VARIABLE LENGTH CODE SEQUENCE /NE RESTRICTION IS THAT THE XED LENGTH BLOCK MUST BE AT LEAST AS LONG AS THE MAXIMAL CODE WORD LENGTH IN THE USED VARIABLE LENGTH CODE "OTH THE SPEED AND MEMORY REQUIREMENT PROPERTIES OF THE ALGORITHMS ARE STUDIED )T IS SHOWN THAT THE LOOK UP TABLE ALGORITHMS ARE CONSIDERABLY FASTER THAN THE TREE SEARCH ALGORITHM SEEN IN MOST TUTORIAL BOOKS AT A COST OF MEMORY REQUIREMENTS EXPONENTIALLY GROWING WITH THE XED LENGTH BLOCK SIZE /N THE ONE HAND THE MEMORY REQUIREMENTS OF THE LOOK UP TABLE DECODERS ARE LARGE BUT ON THE OTHER HAND THE COST PER BIT OF MEMORY HAS SHOWN A RAPIDLY DECREASING TREND DURING THE LAST FEW DECADES $EDICATED HARDWARE PROPOSALS ARE PRESENTED FOR THE LOOK UP TABLE ALGORITHMS AS WELL AS A SOFTWARE EVALUATION OF THE DECOMPRESSION SPEEDS ON AN I 0# USING A BIT # COMPILER
4HIS PART IS REPRODUCED FROM / %DFORS ! %RENDI AND 0 / "RJESSON ! PERFORMANCE ANALYSIS OF LOOK UP TABLE BASED VARIABLE LENGTH DECODERS 2ESEARCH 2EPORT 45,%! ,ULE 5NIVERSITY OF 4ECHNOLOGY ,ULE

6ARIABLE LENGTH DECODERS

)NTRODUCTION
6ARIABLE LENGTH CODING BASED ON 3HANNONS 3OURCE #ODING 4HEOREM ;= IS A COMMON DATA COMPRESSION TECHNIQUE 4HE CODE CONSTRUCTION METHODS EG THE ONE PRESENTED BY (UdMAN ;= USES THE PRINCIPLE THAT SYMBOLS WITH A HIGH PROBABILITY OF APPEARANCE SHOULD REQUIRE LESS STORAGE SPACE THAN LESS FREQUENT SYMBOLS 4HIS IS DONE BY ASSIGN ING SHORT CODE WORDS TO THE HIGH PROBABILITY SYMBOLS AND LONG CODE WORDS TO THE LOW PROBABILITY ONES ! VARIABLE LENGTH CODED SEQUENCE CAN BE DECODED BY A TREE SEARCH "Y MATCHING THE BITS IN THE SEQUENCE AGAINST THE BRANCHES IN THE CODE TREE ONE WILL UPON ARRIVAL AT A LEAF HAVE DECODED ONE SOURCE SYMBOL (OWEVER THE TREE SEARCH IS NOT A VERY FAST DECODING METHOD )F THE AVERAGE CODE WORD LENGTH IS 6 BITS A TRIVIAL TREE SEARCH WILL REQUIRE AN AVERAGE OF 6 MATCHINGS BEFORE A SOURCE SYMBOL IS DECODED 4HE APPROACH WE HAVE CHOSEN IS TO COMPLETE THE CODE TREE TO A CERTAIN DEPTH OF EVERY BRANCH ALLOWING A TABULAR DECODING WHERE ONE OR SEVERAL SOURCE SYMBOLS ARE DECODED PER CYCLE OF THE ALGORITHM $IdERENT TYPES OF SIMILAR ALGORITHMS HAVE BEEN USED IN COMPRESSION SOFTWARE BUT AS FAR AS THE AUTHORS KNOW NO SCIENTIC MATERIAL HAS BEEN PUBLISHED WHERE THE SPEED PERFORMANCE IS ANALYSED 4O MAKE SURE THAT THERE IS NO AMBIGUITY IN THE ANALYSIS WE CONSTRUCT A SET OF DECODERS WHICH WE ARE USING IN THE ANALYSIS )N #HAPTER THE BASIC STRUCTURES AND THE CODING ALGORITHMS ARE INTRODUCED 4HE TABULAR DECODING CONSISTS OF TWO GROUPS OF ALGORITHMS EACH GROUP CONSISTING OF A LOOK UP TABLE CONSTRUCTION ALGORITHM AND A VARIABLE LENGTH DECODER 4HE RST TWO ALGORITHMS ARE DESCRIBING A LOOK UP TABLE DECODER AND THE TWO LATTER ONES AN EXTENDED LOOK UP TABLE DECODER #HAPTER ALSO CONTAINS A PROPOSAL ON DEDICATED HARDWARE FOR HIGH SPEED DECODING )N #HAPTER THE DECODING SPEED AND THE REQUIRED MEMORY SPACE OF THE LOOK UP TABLE DECODING ALGORITHMS ARE ANALYSED 4HE POSSIBLE PROBLEMS WITH LARGE MEMORY REQUIREMENTS CAN BE DEALT WITH BY INTRODUCING SMALLER CHANGES IN THE DECODING AND ENCODING ALGORITHMS ! GROUPING STRATEGY CAN BE APPLIED ON THE VARIABLE LENGTH CODE TO REDUCE THE MEMORY REQUIREMENTS SEE FOR INSTANCE ; = AND THERE ARE ALSO ALGORITHMS AVAILABLE FOR DESIGNING LENGTH LIMITED VARIABLE LENGTH CODES SEE FOR INSTANCE ; = 4HESE METHODS FOR SOLVING MEMORY PROBLEMS ARE NOT INCLUDED IN THE REPORT )F A SHORT INTRODUCTION TO THE ALGORITHMS IS DESIRED IT IS RECOMMENDED TO HAVE A BRIEF LOOK AT THE CHAPTERS AND IN COMBINATION WITH THE TWO APPENDICES ! AND " WHICH CONTAIN A FEW EXAMPLES ON DECODING AND A 0# IMPLEMENTATION OF THE ALGORITHMS

!LGORITHMS AND A HARDWARE PROPOSAL
)N THIS CHAPTER THE CODING AND DECODING ALGORITHMS ARE DESCRIBED 4HE CODING ALGORITHM IS A STRAIGHTFORWARD IMPLEMENTATION BUT IT IS INCLUDED FOR THE PURPOSE OF COMPLETENESS 4HE ALGORITHM IMPROVEMENTS IN THIS REPORT CONCERN THE DECODING PART WHERE TWO DIdERENT TYPES OF LOOK UP TABLES ARE USED TO ACHIEVE A FAST DECODING PROCEDURE 4HE ENCODERDECODER MODEL USED IN THIS CHAPTER IS PRESENTED IN &IGURE WHERE

IS A DATA SEQUENCE OF SOURCE SYMBOLS AND ! IS A SEQUENCE OF BITS REPRESENTING THE CORRESPONDING VARIABLE LENGTH CODE SEQUENCE
A B
Encoder
Storage/ Channel
Decoder
&IGURE %NCODER$ECODER MODEL 4O ENSURE THAT THERE IS NO AMBIGUOUSNESS IN THE MODEL DESCRIPTION USED A FEW BASIC DENITIONS ARE STATED -OST OF THE NOTATIONS USED ARE LISTED IN !PPENDIX # $ENITION - DENOTES 3 THE SET OF NON EMPTY SEQUENCES CONSISTING OF ELEMENTS FROM THE SET - IE - - (
M M
$ENITION 4HE SET OF SOURCE SYMBOLS 3 FR J H - ` G IS A SET OF SYMBOLS FROM WHICH THE SOURCE PRODUCES A SEQUENCE 3 h i $ENITION ! BINARY VARIABLE LENGTH PREX CODE # " F G J H - ` IS A SET OF CODE WORDS WHERE EVERY CODE WORD " HAS A CERTAIN LENGTH K BITS AND NO " IS THE PREX OF ANY OTHER CODE WORD " I H SEE FOR INSTANCE ;=
H H H H H I
$ENITION "Y A TRIMMED BINARY VARIABLE LENGTH CODE WE MEAN A CODE # THAT HAS NO EMPTY LEAVES IN THE TREE REPRESENTATION CAN BE EXPRESSED AS EQUALITY 0 OF THE CODE 0 4HIS IN THE +RAFT INEQUALITY ;= v IE WHERE K IS THE CODE ( ( WORD LENGTH ACCORDING TO $ENITION
H KH H KH H
2EMARK !S LONG AS WE ARE STUDYING BINARY VARIABLE LENGTH CODES THE CODE CAN ALWAYS BE TRIMMED ACCORDING TO $ENITION WITHOUT ANY LOSS IN COMPRESSION RATE )N FACT THE COMPRESSION RATE IS ENHANCED BY TRIMMING A CODE 4HIS IS DUE TO THE FACT THAT THE NUMBER OF LEAVES CAN BE INCREASED OR DECREASED BY ONE BY ADDING OR REMOVING ONE NODE $ENITION 6ARIABLE LENGTH CODING OF A SYMBOL FROM THE SOURCE SYMBOL SET 3 ACCORD ING TO $ENITION WITH A VARIABLE LENGTH CODE # ACCORDING TO $ENITION IS DENED AS A MAPPING 5+" 3 # SUCH THAT R `
H
SC4
"
H - `
"EFORE DESCRIBING THE ENCODING AND DECODING ALGORITHMS A FEW FUNCTIONS AND A CON CATENATION OPERATOR ARE INTRODUCED WHICH ARE TO BE USED ON SEQUENCES LATER ON $ENITION ,ET ! BE BINARY SEQUENCE ! n n n VALUATION OF THAT SEQUENCE IS DENOTED BY U@K !
L L
WHERE n F G 4HE

H
U@K !
8
H(
n
H

L

$ENITION ! BINARY SEQUENCE ! n n n OF LENGTH L WITH A CERTAIN VALUA TION M FH J H ` G IS DENOTED BY AHMRDP L M
L
AHMRDP L M !
RTBG SG@S U@K ! M
2EMARK 4HE FUNCTIONS U@K ! IN $ENITION AND AHMRDP L M IN $ENITION DO NOT NECESSITATE CALCULATION ON AN ORDINARY COMPUTER SINCE THE BITS n n n IS THE INTERNAL REPRESENTATION OF THE VALUATION
L
$ENITION ,ET AND ! BE TWO SEQUENCES m m m AND ! n n n BELONGING TO THE SAME SET - THE CONCATENATION OPERATOR h IS DENED BY
L
h ! m m m
n

n n
4HE ALGORITHM DESCRIPTION IN THE SEQUEL IS DONE IN AN ABSTRACT WAY BUT IN !PPENDIX ! AN EdORT IS MADE TO ELUCIDATE THE THEORY BY A FEW EXAMPLES

%NCODING ALGORITHM
4HE ALGORITHM DESCRIBED IN THIS SECTION IS A STANDARD VARIABLE LENGTH CODING PROCEDURE 4HE BASIC STRUCTURE OF THIS TYPE OF CODING WAS RST PROPOSED BY 3HANNON ;= IN !LGORITHM %NCODING (AVING A SEQUENCE OF SOURCE SYMBOLS m m m THE FOLLOWING ALGORITHM CAN BE USED FOR ENCODING INTO A BINARY SEQUENCE OF CONCATENATED VARIABLE LENGTH CODE WORDS ! 4HE ALGORITHM IS BASED ON $ENITION
L

H ! ! ! h 5 +" m HH
HE H L FNSN
H
!FTER COMPLETING THE ALGORITHM ! CONTAINS THE VARIABLE LENGTH CODED VERSION OF THE SE QUENCE

! LOOK UP TABLE DECODER
4HE LOOK UP TABLE DECODER PROPOSED IN THIS SECTION WILL DECODE ONE SOURCE SYMBOL FOR EACH MATCHING THUS SPEEDING UP THE DECODING PROCEDURE COMPARED TO A TREE SEARCH ALGORITHM 4HE DECODING ALGORITHM IS BASED ON A TABULAR METHOD WHERE THE LEAVES OF A COMPLETED CODE TREE ARE USED AS THE TABULAR ENTRIES #OMPLETING THE CODE TREE IS TO BE INTERPRETED AS AN EXTENSION OF EVERY BRANCH TO A CERTAIN DEPTH Kvd WHERE Kvd IS THE MAXIMAL CODE WORD LENGTH IN THE CODE USED &IRST A DECODE TABLE IS INTRODUCED IN THE FORM OF A DECODE ARRAY WHICH IS CONSTRUCTED ACCORDING TO !LGORITHM

!LGORITHM ,OOK UP TABLE ,ET # BE PRE DENED TRIMMED VARIABLE LENGTH PREX CODE ACCORDING TO $ENITION AND DENOTE THE MAXIMAL CODE WORD LENGTH IN # BY Kvd (AVING AN ARRAY IMPLEMENTATION OF OUR LOOK UP TABLE CDBNCD :H< 3 FOR H L@W ` THE CONTENTS ARE DETERMINED BY THE FOLLOWING ALGORITHM
K

H I " " h AHMRDP Kvd ` K I CDBNCD :U@K " < R I I

HE I L@W FNSN HH
HE H - FNSN
H H H K KH
!FTER COMPLETING THE ABOVE ALGORITHM THE CDBNCD :a< ARRAY CAN BE USED IN THE DECODING ALGORITHM TO PERFORM A DECODING 4HE RST CODE WORD OF A VARIABLE LENGTH CODED SEQUENCE ! CAN NOW BE DECODED BY INDEXING CDBNCD :a< WITH THE VALUATION OF THE RST Kvd BITS OF ! 4HE BITS OF THE NEXT CODE WORD IN THE SEQUENCE WILL NOT EdECT THE DECODED SYMBOL RETURNED BY CDBNCD :a< 4HIS IS DUE TO THE STEPS OF !LGORITHM !FTER THE BITS OF THE RST CODE WORD ANOTHER SET OF Kvd BITS ARE TAKEN 2EPEATING THIS THE WHOLE SEQUENCE CAN BE DECODED !LGORITHM ,OOK UP TABLE DECODER ,ET ! BE THE REPRESENTATION OF A BINARY CONCATENATED VARIABLE LENGTH CODE SEQUENCE ! n n n FOR INSTANCE CREATED BY !LGORITHM AND LET Kvd BE THE MAXIMAL CODE WORD LENGTH IN THE CODE # !SSUMING A DECODE ARRAY CDBNCD :a< IE A LOOK UP TABLE CREATED BY !LGORITHM THE DECODING OF ! INTO IS EXECUTED IN THE FOLLOWING MANNER
M

I b C CDBNCD U@K n n hC I I
KDMFSG C HE I M FNSN
I
I
n
I KL@W
WHERE KDMFSG C IS A FUNCTION RETURNING THE LENGTH OF THE CODE WORD ASSOCIATED WITH THE SOURCE SYMBOL C 7HEN COMPLETING THE ALGORITHM CONTAINS THE DECODED SOURCE SYMBOL SEQUENCE 2EMARK )N STEP OF !LGORITHM THE INDEX OF THE BINARY SEQUENCE ! MIGHT EXCEED THE WELL DENED BITS OF ! IE EXCEED M ` )F THIS SITUATION OCCURS THE UNDENED BITS MAY BE LLED WITH ANY CONTENTS EG ZEROS OR ONES 2EMARK 4HE FUNCTION KDMFSG R CAN b BE IMPLEMENTED AS c AN ARRAY CONTAINING THE CODE WORD LENGTHS K EG KDMFSG :H< K K K DENING KDMFSG 3
H H -
FH J H Kvd G BY R `
H
uhwpq
KDMFSG :H<

!N EXTENDED LOOK UP TABLE DECODER
! MORE EbCIENT DECODING ALGORITHM THAN IN THE PREVIOUS SECTION CAN BE DERIVED IN A SOMEWHAT DIdERENT MANNER )NSTEAD OF COMPLETING THE CODE TREE TO THE FULL DEPTH OF EVERY BRANCH DECODING THE RST SYMBOL BY A TABULAR INDEXATION IT IS POSSIBLE TO COMPLETE THE TREE TO A PRE DENED DEPTH K w Kvd BY A RECURSIVE EXTENSION OF ITSELF 3UCH AN APPROACH WILL ALLOW MORE THAN ONE SOURCE SYMBOL TO BE DECODED AT EACH INDEXATION !NY DECODING ALGORITHM CAN BE USED TO PERFORM THE RECURSIVE EXTENSION OF THE TREE BUT TO MAKE THE DESCRIPTION OF !LGORITHM SHORTER THE DECODE ARRAY FROM !LGORITHM IS BIT BLOCKS AND DECODING AS MANY USED 4HE DECODING PROCEDURE IS BASED ON TAKING K SYMBOLS AS POSSIBLE IN THAT CERTAIN BLOCK 4HE CONTENTS OF A DECODE ARRAY BASED ON A RECURSIVELY EXTENDED TREE IS DETERMINED BY !LGORITHM
CDOSG CDOSG
!LGORITHM %XTENDED LOOK UP TABLE !SSUME A PRE DENED TRIMMED PREX FREE VARIABLE LENGTH CODE # ACCORDING TO $ENITION WHERE THE MAXIMAL CODE WORD LENGTH IN # IS Kvd (AVING A LOOK UP TABLE CDBNCD :H< 3 FOR H L@W ` CREATED BY !LGORITHM THE CONTENTS OF THE EXTENDED LOOK UP TABLE CDBNCD :I < 3 FOR I ` IS DETERMINED BY THE FOLLOWING ALGORITHM
K $ KCDOSG

I # I VGDQD " " AHMRDP K J b c C CDBNCD U@K L@W J J

KDMFSG C HE J v K SGDM # # h C FNSN CDBNCD :I < # I I
HE I FNSN
CDOSG J J J K CDOSG $ KCDOSG $
KCDOSG
4HE DECODE ARRAY CDBNCD :a< CAN NOW BE USED IN THE EXTENDED LOOK UP TABLE DECODER 2EMARK )N STEP OF !LGORITHM THE INDEX OF THE BINARY SEQUENCE " MIGHT EXCEED THE WELL DENED BITS OF " IE EXCEED K ` )F THIS SITUATION OCCURS THE UNDENED BITS MAY BE LLED WITH ANY CONTENTS EG ZEROS OR ONES
CDOSG
!LL CODE WORDS FULLY CONTAINED IN THE RST K BITS OF A BINARY SEQUENCE ! CAN NOW BE DECODED BY INDEXING CDBNCD :a< WITH THE VALUATION OF THOSE BITS !T THE END OF THE LAST CODE WORD ANOTHER SET OF K BITS IS TAKEN 2EPEATING THIS THE WHOLE SEQUENCE CAN BE DECODED
CDOSG $ CDOSG
!LGORITHM %XTENDED LOOK UP TABLE DECODER ,ET ! REPRESENT A BINARY CON CATENATED VARIABLE LENGTH CODE SEQUENCE ! n n n FOR INSTANCE CREATED BY !LGO RITHM !SSUMING A LOOK UP TABLE CDBNCD :a< CREATED BY !LGORITHM WITH A CHOSEN
M $
DEPTH K
CDOSG
6ARIABLE LENGTH DECODERS THE DECODING OF ! INTO I IS EXECUTED IN THE FOLLOWING MANNER H
$
# CDBNCD U@K n n h# I I
KDMFSG # HE I M FNSN
I
n
I KCDOSG
I

WHERE KDMFSG # IS A FUNCTION RETURNING THE SUM OF THE LENGTHS OF THE CODE WORDS ASSOCIATED WITH THE SOURCE SYMBOL SEQUENCE # 7HEN COMPLETING THE ALGORITHM CONTAINS THE DECODED SOURCE SYMBOL SEQUENCE 2EMARK )N STEP OF !LGORITHM THE INDEX OF THE BINARY SEQUENCE ! MIGHT EXCEED THE WELL DENED BITS OF ! IE EXCEED M ` )F THAT SITUATION OCCURS THE UNDENED BITS MAY BE LLED WITH ANY CONTENTS EG ONES OR ZEROS &URTHER TO MAKE SURE THAT NO EXTRA SYMBOLS ARE ADDED TO # WILL HAVE TO BE TRUNCATED UNTIL KDMFSG # HAS THE SAME SIZE AS THE WELL DENED BITS OF n n n
I I I KCDOSG
2EMARK !S IN !LGORITHM THE FUNCTION KDMFSG # CAN BE IMPLEMENTED AS AN ARRAY

! DEDICATED HARDWARE PROPOSAL
)N THIS 3ECTION A PROPOSAL IS MADE ON A DEDICATED HARDWARE FOR THE PRESENTED DECOD ING ALGORITHMS 4HE HARDWARE WILL BE SOMEWHAT DIdERENT FOR THE TWO ALGORITHMS 4HE EXTENDED LOOK UP TABLE ALGORITHM IS IMPLEMENTED IN A WAY THAT WILL ALLOW HIGH SPEED DECODING WITHOUT CARING ABOUT THE REQUIRED MEMORY SPACE )N THE LOOK UP TABLE IMPLE MENTATION MORE CARE IS TAKEN OVER THE REQUIRED MEMORY SPACE AT THE COST OF A LOWER DECOMPRESSION SPEED !CCORDING TO 3ECTION THE LOOK UP TABLE DECODER WILL NEED A CDBNCD :a< ARRAY WITH L@W ENTRIES CONTAINING ONE SOURCE SYMBOL EACH &URTHER A KDMFSG :a< ARRAY SEE 2EMARK WITH - ENTRIES IS NEEDED FOR THE CODE WORD LENGTHS )F WE PUT THE CDBNCD :a< ARRAY IN ONE 2!- RANDOM ACCESS MEMORY AND THE KDMFSG :a< ARRAY IN A SECOND 2!- THE DECODER CAN BE IMPLEMENTED AS IN &IGURE &OR THE EXTENDED LOOK UP TABLE ALGORITHM THERE IS A DIdERENT SITUATION 4HERE ARE TWO MAJOR DIdERENCES THE KDMFSG a FUNCTION HAS TO RETURN THE TOTAL CODE WORD LENGTH OF A SEQUENCE OF SOURCE SYMBOLS AND THE ENTRIES IN THE LOOK UP TABLE WILL HAVE TO CONTAIN SEQUENCES OF SOURCE SYMBOLS )N ORDER TO PREVENT A CALCULATION OF KDMFSG # IT IS STORED FOR EVERY SEQUENCE IN THE CDBNCD :a< ARRAY 4HIS WILL IMPLY THAT THE KDMFSG :a< ARRAY WILL CONTAIN ENTRIES IE THE SAME NUMBER OF ENTRIES AS THE CDBNCD :a< ARRAY 5SING THE FACT THAT THE NUMBER OF ENTRIES ARE THE SAME THE TWO ARRAYS CAN BE PLACED IN THE SAME 2!- INDEXING THEM WITH THE SAME ADDRESS IE THE SAME K BITS OF THE BINARY CODE SEQUENCE 4O BE ABLE TO STORE A SOURCE SYMBOL SEQUENCE # CONTAINING M SYMBOLS IT HAS BEEN B WHERE # OCCUPIES THE RST M POSITIONS CHOSEN TO REPRESENT # AS A XED LENGTH SEQUENCE #
K $ KCDOSG $ CDOSG

Shift register (lmax bits)
Buffer Storage/ Channel

Look-up table memory (RAM)
length(D)
Code word length memory (RAM)
&IGURE $EDICATED LOOK UP TABLE HARDWARE % B 4HE MEMORY REPRESENTATION OF SUCH A SEQUENCE CAN BE INTERPRETED AS A TUPLE M # 4HE B WILL HAVE TO BE LARGE ENOUGH TO STORE ALL SEQUENCES IN NUMBER OF AVAILABLE POSITIONS IN # THE LOOK UP TABLE 4HE PROPOSED HARDWARE IMPLEMENTATION IS SHOWN IN &IGURE
Shift register (ldepth bits)
Buffer Storage/ Channel
Look-up table and code sequence length memory (RAM)
length ( D ) D n
&IGURE $EDICATED EXTENDED LOOK UP TABLE HARDWARE 2EMARK 4HE BUdER AND SHIFT REGISTER IMPLEMENTATION CAN BE DONE IN A WAY THAT WILL ALLOW ANY SHIFT TO BE DONE IN CONSTANT TIME /NE APPROACH IS TO IMPLEMENT THE SHIFT REGISTER AS A NET OF BINARY LOGIC IE A BARREL TYPE SHIFT REGISTER

$ECODING ALGORITHM PROPERTIES
4HE PROPERTIES WE ARE INVESTIGATING IN THIS CHAPTER ARE THE DECODING SPEED AND THE MEMORY REQUIREMENTS OF THE DECODING ALGORITHMS 4HE SPEED OF AN ALGORITHM IS MEASURED

IN THE ASYMPTOTIC NUMBER OF DECODED SOURCE SYMBOLS PER ALGORITHM CYCLE 4HIS MAKES THE MEASURE IMPLEMENTATION INDEPENDENT 4HERE ARE SEVERAL METHODS FOR STORING THE DECODING TABLES AND EACH METHOD WILL REQUIRE A DIdERENT AMOUNT OF MEMORY 4HE MEMORY SIZES CONSIDERED IN THE ANALYSIS ARE BASED ON THE HARDWARE IMPLEMENTATIONS PROPOSED IN 3ECTION $ENITION !LGORITHM CYCLE 4HE ALGORITHM CYCLE USED FOR MEASURING THE ALGO RITHM SPEED CONSISTS OF THE TWO BASIC STEPS SHOWN BELOW /NE EXECUTION OF THE STEPS H AND HH IS DENED AS ONE ALGORITHM CYCLE H HH 'ET BIT S FROM THE STORAGECHANNEL $ECODE BIT S DIdERENT APPROACH FOR THE ALGORITHMS
)N THE TREE SEARCH ALGORITHM THE ALGORITHM CYCLE CONSISTS OF GET ONE BIT AND MATCH WITH THE BRANCHES IN THE CODE TREE AND IN THE LOOK UP TABLE AND EXTENDED LOOK UP TABLE ALGORITHMS !LGORITHM AND IT CONSISTS OF THE STEPS $ENITION !LGORITHM SPEED ,ET b b 7 3 0 # BE A MARKOV CHAIN DE SCRIBING THE NUMBER OF DECODED SOURCE SYMBOLS AT ALGORITHM CYCLE M OF DECODING ALGORITHM 7 USING THE CODE # ON THE SOURCE 3 WITH PROBABILITY DISTRIBUTION 0 4HE SPEED OF ALGO RITHM 7 UNDER THESE CONDITIONS IS DENED AS THE FOLLOWING ASYMPTOTIC NUMBER OF DECODED SOURCE SYMBOLS PER ALGORITHM CYCLE
M M
8 3 0 # KHL L ( b
L

%XPRESSION OF $ENITION WILL BE USED IMPLICITLY IN THE SEQUEL WITHOUT REF ERENCES

!LGORITHM SPEED
)N THE SPEED ANALYSIS OF THE ALGORITHMS THE AVERAGE CODE WORD LENGTH IS USED AS THE INDEPENDENT VARIABLE 4HE FOLLOWING DENITION WILL GIVE THE OBSERVED AVERAGE CODE WORD LENGTH FOR A CERTAIN CODING SITUATION $ENITION 4HE THE AVERAGE CODE WORD LENGTH WHEN USING THE CODE # ON A SOURCE WITH PROBABILITY DISTRIBUTION 0 IS GIVEN BY 6 # 0 :BITSYMBOL<
-
6 # 0
8
H(
OK
H H

)NFORMATION ON AVERAGE CODE WORD LENGTH CAN BE FOUND IN SEVERAL TEXTS CONCERNING VARIABLE LENGTH CODING SEE FOR INSTANCE ; = 4HE RST TWO PROPERTIES ARE GIVEN WITHOUT FORMAL PROOF AND ALL THREE OF THEM ARE BASED ON THE ASYMPTOTIC NUMBER OF DECODED SOURCE SYMBOLS PER ALGORITHM CYCLE ACCORDING TO $ENITION 4HE AVERAGE CODE WORD LENGTH 6 # 0 OF $ENITION IS USED AS THE INDEPENDENT VARIABLE

0ROPERTY 4REE SEARCH SPEED 4HE ORIGINAL TREE SEARCH ALGORITHM WILL ON AVER AGE DECODE ONE SOURCE SYMBOL PER 6 # 0 BIT MATCHINGS AGAINST THE BRANCHES IN THE CODE TREE (ENCE THE DECODING SPEED IS h
32
3 0 #
6 # 0

0ROPERTY ,OOK UP TABLE SPEED !S MENTIONED IN 3ECTION THE LOOK UP TA BLE DECODER IN !LGORITHM WILL DECODE ONE SOURCE SYMBOL PER ITERATION (ENCE THE DECODING SPEED IS h 3 0 #
+3
0ROPERTY %XTENDED LOOK UP TABLE SPEED 4HE DECODING SPEED OF !LGORITHM IS BOUNDED BY t u K
K vh 3 0 # v L@W 6 # 0 6 # 0
CDOSG CDOSG $+3
0ROOF ,ET US START WITH THE UPPER BOUND &IRST ASSUME THAT L IS THE NUMBER OF SYMBOLS IN THE SEQUENCE AND THAT M IS THE NUMBER OF BITS IN THE CODE SEQUENCE ! WHEN USING THE CODE # ON 4HE MOST FAVOURABLE SITUATION WOULD BE IF THE DECODER ACTUALLY DECODED K BITS EVERY STEP OF THE ALGORITHM )F THE DECODER HAD TO USE J STEPS TO DECODE ! INTO THE TOTAL NUMBER OF DECODED BITS WOULD BE JK 4HE OBSERVED AVERAGE CODE WORD LENGTH OF ML BITSSYMBOL IN COMBINATION WITH JK M THEN PRODUCES THE UPPER BOUND L L K h 3 0 # v L J MK 6 # 0
CDOSG CDOSG CDOSG CDOSG $ +3 CDOSG
)F WE CONSIDER THE LOWER BOUND IT IS QUITE OBVIOUS THAT THE EXTENDED LOOK UP TABLE DECODER WILL DECODE AT LEAST ONE SYMBOL PER ITERATION (OWEVER IT IS NOT A VERY APPEALING LOWER BOUND 4HEREFORE CONSIDER A SITUATION WHERE TWO CONSECUTIVE ALGORITHM STEPS ARE STUDIED )F THE DECODER MANAGES TO DECODE J BITS DURING THE RST ALGORITHM STEP THE NEXT K ` J BITS IN THE SEQUENCE MUST BELONG TO ONE AND ONLY ONE CODE WORD OF LENGTH GREATER OR EQUAL TO K ` J
HENCE IN THE NEXT STEP OF THE ALGORITHM AT LEAST K `J
BITS WILL BE DECODED &ROM THIS SITUATION WE CAN CONCLUDE THAT EVERY TWO STEPS OF THE ALGORITHM AT LEAST K
BITS WILL BE DECODED 5SING THE SAME TYPE OF ARGUMENTATION AS IN EXPRESSION WE ARRIVE IN
CDOSG CDOSG CDOSG CDOSG
$ +3
3 0 # w
K
L 6 # 0
CDOSG

%XPRESSION IN COMBINATION WITH THE AT LEAST ONE SYMBOL PER ITERATION STATEMENT PRODUCES t u K
h 3 0 # w L@W 6 # 0
CDOSG $+3
"Y COMBINING THE TWO EXPRESSIONS AND THE PROOF IS COMPLETED

4HEOREM 5SING THE THREE DECODING ALGORITHMS TREE SEARCH LOOK UP TABLE AND EX TENDED LOOK UP TABLE ON THE SAME SYMBOL SOURCE WITH THE SAME VARIABLE LENGTH CODE THE FOLLOWING RELATION WILL HOLD FOR THE SPEED OF THE ALGORITHMS h
32 32
3 0 # v h
+3
3 0 # v h
$ +3
$+3
3 0 #
WHERE h 3 0 # h 3 0 # AND h 3 0 # ARE THE INDIVIDUAL SPEEDS OF THE ALGORITHMS ACCORDING TO THE EXPRESSIONS AND
+3
0ROOF &OR EVERY VARIABLE LENGTH CODE # THE AVERAGE CODE WORD LENGTH WILL BE GREATER OR EQUAL TO ONE BIT PER SYMBOL IE 6 # 0 w 4HIS IS DUE TO THE FACT THAT EVERY CODE WORD MUST HAVE A LENGTH OF AT LEAST ONE BIT 5SING THIS WE GET THE FOLLOWING BOUNDS ON h 3 0 #
32
h
32 +3
32
3 0 # v
4HIS IMPLIES h 3 0 # v h 3 0 # 5SING THE EXPRESSIONS AND THE CONCLUSION IS THAT THE RELATION IN 4HEOREM HOLDS
3INCE ONLY UPPER AND LOWER BOUNDS ARE DERIVED ON h A SIMPLE EXPERIMENTAL EVALU ATION OF THE SPEED IS PRESENTED IN &IGURE ALONG WITH THE THEORETICAL SPEED PROPERTIES 4HE EXPERIMENTAL DATA WAS CREATED USING A UNIFORMLY DISTRIBUTED STOCHASTIC VARIABLE t < AND A MAPPING x y E t t @t

$ +3
WHERE OBSERVATIONS OF E t :< WERE USED AS THE INPUT SEQUENCE TO THE ALGORITHM 4HE PARAMETER @ IN EXPRESSION WAS SET TO DIdERENT VALUES TO CHANGE THE SHAPE OF THE DISTRIBUTION OF E t IE TO CHANGE THE ENTROPY )T CAN BE SHOWN THAT t /Q u ` W W ` @t
@W v W v @

&ROM THE EXPRESSIONS AND IT FOLLOWS THAT r s ` M /Q E t vM ` @ M
x y v M v @

E WHERE M IS AN INTEGER 4HIS WILL IMPLY THAT THE PROBABILITY OF THE SMALLEST r VALUE ON s t WILL INCREASE WITH @ AND THE REST OF THE PROBABILITIES ARE SCALED TO MAKE /Q E t v 4HE REASON FOR CHOOSING THE ABOVE DESCRIBED METHOD FOR CREATING TEST DATA IS A MATTER OF A SIMPLE REALIZATION IN THE # PROGRAMMING LANGUAGE )N &IGURE WE HAVE PLOTTED THE THEORETICAL VALUES AND THE RESULT FROM THE EXPERI MANTAL EVALUATION !T LEAST FOR THIS EXPERIMENT THE EXPERIMENTAL VALUES ARE NOT TOO FAR FROM THE UPPER BOUND ON h
$+3

16

Tree-search
Decoded symbols per algorithm cycle
14 12
Look-up table Extended look-up table (upper and lower bound)

+
Extended look-up table (experimental)
10 8 6
+ + +
4
2 0
Average code word length, W [bit/symbol]
&IGURE $ECODING SPEED USING A SOURCE WITH - SYMBOLS AND K
CDOSG

-EMORY REQUIREMENTS
&ROM THE LOOK UP TABLE CONSTRUCTION ALGORITHMS !LGORITHM AND IT IS OBVIOUS THAT THE MEMORY REQUIREMENTS OF THE TABLES ARE LARGE FOR VARIABLE LENGTH CODES WITH A LARGE MAXIMUM CODE WORD LENGTH 4HERE ARE SEVERAL DIdERENT WAYS OF STORING THE LOOK UP TABLES FOR THE DECODING ALGO RITHMS HOWEVER THE STRUCTURES CHOSEN IN THIS SECTION ARE BASED ON THE PROPOSED HARDWARE IMPLEMENTATION OF THE DECODING ALGORITHMS SEE 3ECTION &OR INSTANCE BY USING DY NAMIC MEMORY ALLOCATION THE REQUIRED MEMORY SPACE FOR THE EXTENDED LOOK UP TABLE ALGORITHM CAN BE REDUCED BUT IN THAT CASE A DEDICATED HARDWARE IMPLEMENTATION WILL BE LESS EbCIENT AND MORE COMPLEX "EFORE PROCEEDING WITH THE ANALYSIS THE FOLLOWING LEMMA IS STATED WHICH WILL BE OF CONSIDERABLE USE IN THE FORTHCOMING ANALYSIS ,EMMA 4HE SMALLEST MEMORY SPACE REQUIRED TO STORE ONE SYMBOL FROM A SET - FL J H * ` G OF SIZE * SYMBOLS IS
H
LDL - DKNF * E
:AHS<
M M
0ROOF (AVING A MEMORY OF SIZE M BITS ONE CAN REPRESENT ONE OUT OF DIdERENT SYMBOLS 4O BE ABLE TO STORE ONE OUT OF * SYMBOLS ONE HAS TO CHOOSE w * 3INCE M IS AN INTEGER WE WILL HAVE TO PICK THE SMALLEST INTEGER M SATISFYING THE INEQUALITY IE M DKNF * E THUS PROVING ,EMMA
&OR BOTH DECODING ALGORITHMS THE LOOK UP TABLE DECODER AND THE EXTENDED LOOK UP TABLE DECODER THE MEMORY SPACE REQUIRED FOR THE DECODE ARRAY IE THE LOOK UP TABLE WILL BE EXPONENTIALLY GROWING WITH THE DEPTH OF THE COMPLETED TREE )N THE LOOK UP TABLE ALGORITHM IE !LGORITHM A TABLE WITH L@W ENTRIES IS USED EACH ENTRY CONTAINING ONE SYMBOL FROM THE SET OF SOURCE SYMBOLS !PART FROM THE LOOK UP TABLE A TABLE WITH CODE WORD LENGTHS IS NEEDED SEE 2EMARK WITH A NUMBER OF ENTRIES CORRESPONDING TO THE NUMBER OF CODE WORDS IE THE NUMBER OF SOURCE SYMBOLS
K

4HE RESULT OF THE ABOVE DISCUSSION WILL IMPLY THE FOLLOWING PROPERTIES BASED ON ,EMMA 0ROPERTY ,OOK UP TABLE MEMORY !SSUME A SET 3 OF - SOURCE SYMBOLS AND A MAXIMAL CODE WORD LENGTH OF Kvd BITS 4HE TOTAL MEMORY SPACE REQUIRED FOR THE LOOK UP TABLE AND THE TABLE OF CODE WORD LENGTHS IS :BIT< i
+3
L@W LDL 3
- LDL FH J H Kvd G \ [Z ] \ [Z ] LOOK UP TABLE CODE WORD LENGTH TABLE
K

L@W DKNF - E
- DKNF Kvd E
K
)N THE EXTENDED LOOK UP TABLE ALGORITHM TWO ARRAYS ARE USED CDBNCD :a< AND KDMFSG :a< WITH ENTRIES EACH 4HE DECODED SEQUENCE # CAN CONTAIN AT MOST K SYMBOLS WILL HAPPEN IF ONE OF THE B MUST CODE WORDS HAS A LENGTH OF ONE BIT THEREFORE THE XED LENGTH SYMBOL SEQUENCE # BE OF LENGTH K TO BE ABLE TO STORE AN ARBITRARY CODE
$ KCDOSG CDOSG CDOSG
0ROPERTY %XTENDED LOOK UP TABLE MEMORY !SSUME A SET OF SOURCE SYMBOLS 3 FR J H - ` G AND A CHOSEN DEPTH K OF THE RECURSIVELY EXTENDED TREE 4HE TOTAL MEMORY SPACE REQUIRED BY THE EXTENDED LOOK UP TABLE IS :BIT<
H CDOSG
$+3

KCDOSG
"
# LDL 3
LDL FH J H K G ! K \ [Z ] \ [Z ] XED LENGTH SEQUENCE NUMBER OF VALID SYMBOLS uhwpq
CDOSG CDOSG # CDOSG

KCDOSG
DKNF - E
DKNF K
CDOSG
)N &IGURE THE EXPRESSIONS AND ARE PLOTTED AS A FUNCTIONS OF THE DEPTH OF THE COMPLETED TREE .OTE THAT THE MEMORY SIZE IS GIVEN IN BYTES IE IN BLOCKS OF BITS

#ONCLUSIONS
)N THIS REPORT IT IS SHOWN THAT IT IS POSSIBLE TO ACHIEVE A CONSIDERABLE SPEED UP IN THE VARIABLE LENGTH DECODING PROCEDURE COMPARED TO THE TREE SEARCH ALGORITHM BY LOOKING AT BLOCKS OF RECEIVED BITS AND UTILIZING A LOOK UP TABLE FOR DECODING 4HE LOOK UP TABLE DECODER !LGORITHM WILL DECODE ONE SOURCE SYMBOL PER ALGO RITHM CYCLE AND THE EXTENDED LOOK UP TABLE DECODER !LGORITHM WILL DECODE ONE OR MORE SOURCE SYMBOLS PER ALGORITHM CYCLE (OWEVER THE MEMORY REQUIREMENTS ARE LARGE IN BOTH CASES &OR ANY GIVEN VARIABLE LENGTH CODE THE SIZE OF THE LOOK UP TABLE !LGORITHM WILL BE DETERMINED BY THE MAXIMAL CODE WORD LENGTH ! LARGER TABLE WILL NOT INCREASE THE DECODING SPEED AND THEREFORE WE HAVE A XED MEMORY REQUIREMENT IN THIS CASE 4HE SIZE OF THE EXTENDED LOOK UP TABLE !LGORITHM CAN BE CHOSEN BY INCREASING THE DEPTH OF THE COMPLETED CODE TREE ! LARGER EXTENDED LOOK UP TABLE WILL IMPLY A HIGHER

109

108
N= 6 409
Memory size [byte]
107
e abl p t k-u o o l ed end Ext

able p t k-u Loo
N=
256
096 N= 4
106
56 N= 2
105
104 14 16 18 20 22 24
Completed tree depth (l max resp. l depth)
&IGURE 2EQUIRED MEMORY SIZE DECODING SPEED 4HEREFORE IT IS BENECIAL THAT THE DEPTH OF THE COMPLETED TREE BE AS LARGE AS POSSIBLE ON CONDITION THAT THERE IS ENOUGH MEMORY AVAILABLE TO STORE THE TABLE /N ONE HAND THE MEMORY REQUIREMENTS OF THE LOOK UP TABLE DECODERS ARE LARGE BUT ON THE OTHER HAND THE COST PER BIT OF MEMORY HAS SHOWN A RAPIDLY DECREASING TREND DURING THE LAST FEW DECADES )F THE MAXIMAL CODE WORD LENGTH OF AN OPTIMAL VARIABLE LENGTH CODE MAKE THE LOOK UP TABLE SIZE EXCEED THE FEASIBLE ADDRESS SPACE DUE TO HARDWARE LIMITATIONS OR MEMORY COSTS THERE ARE SEVERAL METHODS FOR SHORTENING THE MAXIMAL CODE WORD LENGTH IE DESIGNING OPTIMAL VARIABLE LENGTH CODES WITH A RESTRICTED MAXIMAL CODE WORD LENGTH SEE FOR INSTANCE ; =
!
#ODING EXAMPLES
4HE THREE EXAMPLES CONTAINED IN THIS APPENDIX ARE BASED ON THE THEORY IN #HAPTER DESCRIBING THE ENCODING AND DECODING PROCEDURES %XAMPLE %NCODING #ONSIDER THE VARIABLE LENGTH CODE REPRESENTED AS A BINARY TREE IN GURE &ROM THIS TREE IT IS RELATIVELY STRAIGHTFORWARD TO CREATE THE ENCODING TABLE
s0
0
s1
0 1
s2 s3
&IGURE #ODE TREE

6ARIABLE LENGTH DECODERS )NDEX 3YMBOL #ODE 7ORD R " R " R " R " 4ABLE 6ARIABLE LENGTH CODE
5SING TABLE A SEQUENCE R R R R R R R R R R WILL BE CODED INTO THE BINARY SEQUENCE ! " h " h " h " h " h " h " h " h " h h h h h h h h 2EMARK 4HE BINARY SEQUENCE WILL IN THE CONSIDERED IMPLEMENTATION BE STORED IN A DATA STRUCTURE IN WHICH THE COMPUTER ALLOWS BIT OPERATIONS AND USUALLY HOLDS BITS 4HIS WILL IMPLY THAT THE LAST BIT SEGMENT OF ! MAY NOT BE ENTIRELY LLED WITH RELEVANT DATA )N ORDER TO KNOW WHEN TO STOP DECODING THE NUMBER OF SYMBOLS WILL BE STORED TOO %XAMPLE ,OOK UP TABLE DECODING &IRST APPLY !LGORITHM ON THE VARIABLE LENGTH CODE IN ORDER TO CONSTRUCT THE LOOK UP TABLE 4HE LOOK UP TABLE CONSTRUCTION IS EQUIVALENT TO COMPLETING THE BINARY TREE IN &IGURE THUS OBTAINING THE TREE IN &IGURE 4HIS
0 0 1
s0 s0 s0 s0 s1 s1 s2 s3
0 1 0
0 1
Depth
&IGURE #OMPLETED TREE WILL PROVIDE THE DECODE ARRAY IN 4ABLE WHICH IS USED TO DECODE THE BINARY SEQUENCE ! GIVEN IN %XAMPLE 4HE DECODING IS DESCRIBED BY THE FOLLOWING NINE STEP EXECUTION WHERE A SLIDING WINDOW OF SIZE Kvd BITS IS USED
6ARIABLE LENGTH DECODERS )NDEX 3YMBOL R R R R R R R R 4ABLE ,OOK UP TABLE

2SDO
6HMCNV 2XLANK +DMFSG R R R R R R R R R
%XAMPLE %XTENDED LOOK UP TABLE DECODING .OW APPLY !LGORITHM ON THE LOOK UP TABLE FROM %XAMPLE IN ORDER TO CONSTRUCT THE EXTENDED LOOK UP TABLE )N THIS EXAMPLE IT HAS BEEN CHOSEN TO SET K Kvd BUT IT CAN BE SET TO ANY VALUE K w Kvd 4HE DECODE ARRAY CONSTRUCTION IS EQUIVALENT TO COMPLETING THE BINARY TREE IN &IGURE BY IT SELF IN A RECURSIVE MANNER THUS OBTAINING THE TREE IN &IGURE 4HIS
CDOSG CDOSG
0 0 1
s0 s0 s0 s0 s0 s0 s1 s0 s1 s0 s1 s2 s3
0 1 0
0 1
Depth
&IGURE #OMPLETED TREE BY RECURSIVE EXTENSION PROCEDURE PROVIDES THE EXTENDED DECODE ARRAY IN 4ABLE WHICH IS USED TO DECODE THE
)NDEX 3YMBOL R R R R R R R R R R R R R
4ABLE %XTENDED LOOK UP TABLE BINARY SEQUENCE ! GIVEN IN %XAMPLE 4HE DECODING IS DESCRIBED BY THE FOLLOWING SIX STEP EXECUTION WHERE A SLIDING WINDOW OF SIZE K BITS IS USED
CDOSG
2SDO
6HMCNV 2XLANKR +DMFSG R R R R R R R R R
"
! 0# IMPLEMENTATION
4HIS APPENDIX DESCRIBES A 0# IMPLEMENTATION OF THE LOOK UP TABLE DECODER AND THE EXTENDED LOOK UP TABLE DECODER 4HE FOLLOWING EQUIPMENT WAS USED q )"-0# CLONE I-(Z -" 2!- q 7ATCOM # COMPILER 2ATIONAL 3YSTEMS $/3 EXTENDER 4HE EXECUTION SPEED OF THIS IMPLEMENTATION WILL NOT BE PROPORTIONAL TO THE EXECU TION TIME EXPECTED BY A DEDICATED HARDWARE SINCE DYNAMIC MEMORY ALLOCATION HAS BEEN USED AND THE STANDARD SHIFT OPERATIONS OF THE I PROCESSOR DOES NOT PROVIDE THE SAME SPEED AS A DEDICATED BARREL SHIFTER (OWEVER THIS INFORMATION IS INCLUDED TO GIVE AN EX EMPLICATION OF THE DECODING SPEED OBTAINED ON A 0# 4HE DECODING TABLE CONSTRUCTION ALGORITHMS ARE NOT INCLUDED IN THIS DESCRIPTION )N THE PROGRAM THE BINARY SEQUENCE TO DECODE IS LOCATED IN THE 2!- MEMORY WHERE THE BUdER SIZE CAN BE CHOSEN ARBITRARY AS LONG AS THE COMPUTER HAVE SUbCIENT AMOUNT OF 2!- )F THE SEQUENCE DOES NOT T IN THE 2!- DISK OPERATIONS ARE INCLUDED THAT WILL AUTOMATICALLY LL THE BUdER WITH NEW DATA WHEN IT IS EMPTY "OTH IMPLEMENTATIONS ARE BASED ON THE FOLLOWING TWO FUNCTIONS TMRHFMDC KNNJ M AHSR HMS M UNHC CQNO M AHSR HMS M

WHERE THE LOOK?N?BITS FUNCTION RETURNS THE VALUE OF THE RST N BITS IN THE SEQUENCE AND THE DROP?N?BITS FUNCTION DROPS THE RST N BITS OF THE SEQUENCE
"
4HE LOOK UP TABLE DECODER
4HE BASIC DATA STRUCTURES USED IN THE LOOK UP TABLE DECODER TO DESCRIBE THE CODE ARE THE FOLLOWING SEE !LGORITHM HMS KDMFSG :-< HMS CDBNCD :,< "ASICALLY THE MAIN DECODING LOOP IN THE PROGRAM WILL HAVE THE FOLLOWING STRUCTURE WHERE IT IS ASSUMED THAT THE NUMBER OF CODED SOURCE SYMBOLS NO?OF?SYMBOLS IS KNOWN FROM HEADER INFORMATION ON THE CODE SEQUENCE ENQ H H MN NE RXLANKR H

F CDBNCDC RXLANK CDBNCD :KNNJ M AHSR K L@W < CQNO M AHSR KDMFSG :CDBNCDC RXLANK< c 2SNQD CDBNCDC RXLANK RNLDVGDQD c G
"
4HE EXTENDED LOOK UP TABLE DECODER
4HE BASIC DATA STRUCTURE USED IN THE EXTENDED LOOK UP TABLE DECODER TO DESCRIBE THE CODE IS THE FOLLOWING SXODCDE RSQTBS S@AKD DMSQX F HMS RXLA MN HMS c RXLA RSQ HMS KDMFSG G CDBNCD$ :,< WHERE SYMB?NO AND
SYMB?STR IS THE REPRESENTATION OF THE SEQUENCE # OF !LGORITHM 4HE MAIN LOOP IN THE PROGRAM WILL HAVE THE FOLLOWING STRUCTURE WHERE IT IS ASSUMED THAT THE NUMBER OF CODED SOURCE SYMBOLS NO?OF?SYMBOLS IS KNOWN FROM HEADER INFOR MATION ON THE CODED SEQUENCE LIKE IN THE LOOK UP TABLE DECODER ENQ H H MN NE RXLANKR H

F AHSR KNNJ M AHSR KCDOSG CQNO M AHSR CDBNCD$ :AHSR< KDMFSG c 2SNQD SGD CDBNCD$ :AHSR< RXLA MN RXLANKR ONHMSDC @S AX CDBNCD :AHSR< RXLA RSQ RNLDVGDQD c G
"
$ECOMPRESSION SPEED
5SING THE 0# IMPLEMENTATION OF THE DECODING ALGORITHMS AN EXPERIMENTAL EVALUATION OF THE DECOMPRESSION SPEED HAS BEEN DONE )N &IGURE THE DECOMPRESSION SPEED USING SEVEN DIdERENT BYTE SOURCE LES THE SAME DATA AS IN 3ECTION IS PRESENTED 4HE LE SIZE IS SMALL ENOUGH TO ALLOW THE DECOMPRESSION TO BE DONE WITHOUT BUdER LLING DISK ACCESS

3 x10
5
2.5
Decoded symbols per second
+
2
Look-up table Extended look-up table

+ +
1.5
+ +
+ + + + + + + +
0.5
0 0 1 2 3 4 5 6 7 8
Average code word length, W [bit/symbol]
&IGURE $ECODING SPEED USING A I 0#
#
.OTATIONS
! " B # # 3 M
/RDERED SEQUENCES OF SOME TYPE &IXED LENGTH ORDERED SEQUENCE 3ETS 4HE CARTESIAN PRODUCT b b \ b - [Z ] 3ET DENITION .UMBERED MEMBERS OF AN ORDERED SEQUENCE 0ROBABILITY VARIABLE .UMBERED SET MEMBER " IS A SEQUENCE #ONSTANT NUMERIC VALUES )NTEGER VARIABLES 3TOCHASTIC VARIABLES -EASURED QUANTITIES &UNCTIONMAPPING !RRAY !SSIGNMENT -APPING -APPING DENITION 5PWARD ROUNDING TO THE NEAREST INTEGER CEILING $OWNWARD ROUNDING TO THE NEAREST INTEGER OOR
H M TIMES
FLDLADQ J BNMCHSHNMG mn O RKL" * - L M Kvd K H I b t E t h i M@LD a M@LD :a<

H H H H H H H H
CDOSG
` DaE BaC
"IBLIOGRAPHY
;= / %DFORS - 3ANDELL * * VAN DE "EEK $ ,ANDSTRM AND & 3JBERG !N INTRO DUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT ;= / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON !NALYSIS OF $&4 BASED CHANNEL ESTIMATORS FOR /&$- 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT ;= / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON /&$CHANNEL ESTIMATION BY SINGULAR VALUE DECOMPOSITION 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT ;= - 3ANDELL AND / %DFORS ! COMPARATIVE STUDY OF PILOT BASED CHANNEL ESTIMATORS FOR WIRELESS /&$- 2ESEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY 3EPT ;= . 3UNDSTRM / %DFORS 0 DLING ( %RIKSSON 0 / "RJESSON AND 4 +OSKI #OMBINED LINEAR 6ITERBI EQUALIZERS ! COMPARATIVE STUDY AND A MINIMAX DESIGN )N 0ROC )%%% 6EHIC 4ECHNOL #ONF PAGES 3TOCKHOLM 3WEDEN *UNE ;= / %DFORS AND 0 / "RJESSON $ESIGN AND ANALYSIS OF ROBUST SOURCE CODES 2ESEARCH 2EPORT 45,%! ,ULE 5NIVERSITY OF 4ECHNOLOGY ,ULE ;= / %DFORS ! %RENDI AND 0 / "RJESSON ! PERFORMANCE ANALYSIS OF LOOK UP TABLE BASED VARIABLE LENGTH DECODERS 2ESEARCH 2EPORT 45,%! ,ULE 5NIVERSITY OF 4ECHNOLOGY ,ULE ;= * * VAN DE "EEK / %DFORS - 3ANDELL 3 + 7ILSON AND 0 / "RJESSON /N CHANNEL ESTIMATION IN /&$- SYSTEMS )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES #HICAGO ), *ULY ;= / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON /&$CHANNEL ESTIMATION BY SINGULAR VALUE DECOMPOSITION )N 0ROC )%%% 6EHIC 4ECHNOL #ONF PAGES !TLANTA '! !PR ;= / %DFORS - 3ANDELL * * VAN DE "EEK 3 + 7ILSON AND 0 / "RJESSON !N APPLICATION OF THE SINGULAR VALUE DECOMPOSITION TO /&$- CHANNEL ESTIMATION )N 0ROC 2ADIOVETEN +ONF PAGES ,ULE *UNE
;= - 7AHLQVIST # STBERG * * VAN DE "EEK / %DFORS AND 0 / "RJESSON ! CONCEPTUAL STUDY OF /&$- BASED MULTIPLE ACCESS SCHEMES 0ART !IR INTERFACE REQUIREMENTS 4ECHNICAL 2EPORT 4DOC %43) 34# 3-' MEETING NO (ELSINKI &INLAND -AY ;= * * VAN DE "EEK / %DFORS 0 / "RJESSON - 7AHLQVIST AND # STBERG ! CONCEPTUAL STUDY OF /&$- BASED MULTIPLE ACCESS SCHEMES 0ART #HANNEL ESTIMATION IN THE UPLINK 4ECHNICAL 2EPORT 4DOC %43) 34# 3-' MEETING NO (ELSINKI &INLAND -AY ;= * * VAN DE "EEK / %DFORS 0 / "RJESSON - 7AHLQVIST AND # STBERG ! CONCEPTUAL STUDY OF /&$- BASED MULTIPLE ACCESS SCHEMES 0ART 0ERFORMANCE EVALUATION OF A CODED SYSTEM 4ECHNICAL 2EPORT 4DOC %43) 34# 3-' MEETING NO $SSELDORF 'ERMANY 3EPT ;= * * VAN DE "EEK / %DFORS 0 / "RJESSON - 7AHLQVIST AND # STBERG #HANNEL ESTIMATION IN THE UPLINK OF AN /&$- SYSTEM )N 0ROC .23 3YMP PAGE !UG ;= / %DFORS 0 / "RJESSON AND ! %RENDI !NALYSIS OF A FAST ALGORITHM FOR LOOK UP TABLE DECODING )N 0ROC 2ADIOVETEN +ONF PAGES ,UND 3WEDEN !PR ;= 2ADIO BROADCASTING SYSTEMS $IGITAL !UDIO "ROADCASTING $!" TO MOBILE PORTABLE AND XED RECEIVERS %43 %43) %UROPEAN 4ELECOMMUNICATIONS 3TANDARDS )NSTITUTE 6ALBONNE &RANCE &EB ;= $IGITAL BROADCASTING SYSTEMS FOR TELEVISION SOUND AND DATA SERVICES %UROPEAN 4ELECOMMUNICATIONS 3TANDARD PR%43 $RAFT VERSION !PR ;= 4RANSMISSION AND RECEPTION 4ECHNICAL 2EPORT '3- 2ECOMMENDATION VER SION %43) 6ALBONNE &RANCE -AR ;= 7ORKING DOCUMENT TOWARDS %423-' SELECTION PROCEDURES FOR THE CHOICE OF RADIO TRANSMISSION TECHNOLOGIES OF THE UNIVERSAL MOBILE TELECOMMUNICATIONS SYSTEM 5-43 4ECHNICAL 2EPORT $423-' %43) 6ALBONNE &RANCE ;= - !BRAMOVITZ AND ) 3TEGUN (ANDBOOK OF MATHEMATICAL FUNCTIONS WITH FORMULAS GRAPHS AND MATHEMATICAL TABLES .UMBER IN !PPLIED -ATH 3ERIES .AT "UREAU OF 3TAND 7ASHINGTON $# 53! ;= - !LARD AND 2 ,ASSALLE 0RINCIPLES OF MODULATION AND CHANNEL CODING FOR DIGITAL BROADCASTING FOR MOBILE RECEIVERS %"5 2EVIEW 4ECHNICAL !UG ;= * " !NDERSEN AND " , !NDERSEN &IRST ORDER FREQUENCY SELECTIVE EdECTS ON PHASE MODULATIONS IN A FADING CHANNEL 4ECHNICAL 2EPORT #/34 4$ %52/ #/34 &IRENZE *AN
;= # "ARBOSA -AXIMUM LIKELIHOOD SEQUENCE ESTIMATORS ! GEOMETRIC VIEW )%%% 4RANSACTIONS ON )NFORMATION 4HEORY )4 ;= # 4 "EARE 4HE CHOICE OF THE DESIRED IMPULSE RESPONSE IN COMBINED LINEAR 6ITERBI ALGORITHM EQUALIZERS )%%% 4RANSACTIONS ON #OMMUNICATION ;= * * VAN DE "EEK - 3ANDELL AND 0 / "RJESSON -, ESTIMATION OF TIMING AND FREQUENCY OdSET IN MULTICARRIER SYSTEMS 2ESEARCH 2EPORT 45,%! $IVISION OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECHNOLOGY ;= * * VAN DE "EEK - 3ANDELL - )SAKSSON AND 0 / "RJESSON ,OW COMPLEX FRAME SYNCHRONIZATION IN /&$- SYSTEMS )N 0ROC )NT #ONF 5NIVERSAL 0ERSONAL #OMMUN PAGES 4OKYO *APAN .OV ;= 3 "ENEDETTO % "IGLIERI AND 6 #ASTELLANI $IGITAL 4RANSMISSION 4HEORY 0RENTICE (ALL )NC %NGLEWOOD #LIdS .EW *ERSEY ;= $ "ENGTSSON AND $ ,ANDSTRM #ODING IN A DISCRETE MULTITONE MODULATION SYSTEM -ASTERS THESIS ,ULE 5NIVERSITY OF 4ECHNOLOGY !PR ;= * ! # "INGHAM -ULTICARRIER MODULATION FOR DATA TRANSMISSION !N IDEA WHOSE TIME HAS COME )%%% #OMMUN -AG -AY ;= 0 / "RJESSON $ATA #OMPRESSION AND 3IGNAL %STIMATION WITH !PPLICATION TO $IGITIZED %LECTROCARDIOGRAMS $OCTORAL THESIS 5NIVERSITY OF ,UND ;= 0 / "RJESSON AND / 0AHLM 0REDICTIVE CODING FOR %#' DATA COMPRESSION 4ECHNICAL 2EPORT 42 4ELECOMMUNICATION 4HEORY 5NIVERSITY OF ,UND -AY ;= % & #ASAS AND # ,EUNG /&$- FOR DATA COMMUNICATION OVER MOBILE RADIO &- CHANNELS 0ART ) !NALYSIS AND EXPERIMANTAL RESULTS )%%% 4RANS #OMMUN -AY ;= % & #ASAS AND # ,EUNG /&$- FOR DATA COMMUNICATION OVER MOBILE RADIO &CHANNELS 0ART )) 0ERFORMANCE IMPROVEMENT )%%% 4RANS #OMMUN !PR ;= * #AVERS AND 0 (O !NALYSIS OF THE ERROR PERFORMANCE OF TRELLIS CODED MODULATION IN 2AYLEIGH FADING CHANNELS )%%% 4RANS #OMMUN *AN ;= * + #AVERS !N ANALYSIS OF PILOT SYMBOL ASSISTED MODULATION FOR 2AYLEIGH FADING CHANNELS )%%% 4RANS 6EHIC 4ECHNOL .OV ;= 2 7 #HANG 3YNTHESIS OF BAND LIMITED ORTHOGONAL SIGNALS FOR MULTICHANNEL DATA TRANSMISSION "ELL 3YSTEM 4ECH * $EC ;= ! #HINI -ULTICARRIER MODULATION IN FREQUENCY SELECTIVE FADING CHANNELS 0H$ THESIS #ARLETON 5NIVERSITY /TTAWA #ANADA
;= ! #HINI - 3 %L 4ANANY AND 3 ! -AHMOUD 4RANSMISSION OF HIGH RATE !4PACKETS OVER INDOOR RADIO CHANNELS )%%% * 3ELECT !REAS #OMMUN !PR ;= 0 3 #HOW "ANDWIDTH OPTIMIZED DIGITAL TRANSMISSION TECHNIQUES FOR SPECTRALLY SHAPED CHANNELS WITH IMPULSE NOISE 0H$ THESIS 3TANFORD 5NIVERSITY #! -AY ;= 0 3 #HOW . !L $HAHIR AND * - #IOb ! MULTICARRIER % ($3, TRANSCEIVER SYSTEM WITH CODED MODULATION %UR 4RANS 4ELECOMMUN 2EL 4ECHNOL -AY ;= 0 3 #HOW * - #IOb AND * ! # "INGHAM ! PRACTICAL DISCRETE MULTITONE TRANSCEIVER LOADING ALGORITHM FOR DATA TRANSMISSION OVER SPECTRALLY SHAPED CHANNELS )%%% 4RANS #OMMUN &EB ;= , * #IMINI !NALYSIS AND SIMULATION OF A DIGITAL MOBILE CHANNEL USING ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING )%%% 4RANS #OMMUN #/- *ULY ;= * - #IOb 0ERSONAL COMMUNICATION ;= & #LASSEN AND ( -EYR &REQUENCY SYNCHRONIZATION ALGORITHMS FOR /&$- SYSTEMS SUITABLE FOR COMMUNICATION OVER FREQUENCY SELECTIVE FADING CHANNELS )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES 3TOCKHOLM 3WEDEN *UNE ;= , #OHEN 4IME FREQUENCY ANALYSIS 0RENTICE (ALL .EW *ERSEY ;= * 7 #OOK 7IDEBAND IMPULSIVE NOISE SURVEY OF THE ACCESS NETWORK "4 4ECHNOL *OURN *ULY ;= 4 DE #OUASNON 2 -ONNIER AND * " 2AULT /&$- FOR DIGITAL 46 BROADCASTING 3IGNAL 0ROC 3EPT ;= 4 - #OVER AND * ! 4HOMAS %LEMENTS OF )NFORMATION 4HEORY *OHN 7ILEY 3ONS ;= * 2 #OX AND + , 2IPLEY #OMPACT DIGITAL CODING OF ELECTROCARDIOGRAPHIC DATA )N 0ROC (AWAII #ONF 3YSTEM 3CI
SG
;= 3 . #ROZIER AND $ $ &ALCONER 2EDUCED COMPLEXITY SHORT BLOCK DATA DETECTION TECHNIQUES FOR FADING TIME DISPERSIVE CHANNELS )%%% 4RANSACTIONS ON 6EHICULAR 4ECHNOLOGY !UG ;= & $AdARA AND / !DAMI ! NEW FREQUENCY DETECTOR FOR ORTHOGONAL MULTICARRIER TRANSMISSION TECHNIQUES )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES #HICAGO ), *ULY
;= & $AdARA AND ! #HOULY -AXIMUM LIKELIHOOD FREQUENCY DETECTORS FOR ORTHOGONAL MULTICARRIER SYSTEMS )N 0ROC )NTERN #ONF #OMMUN PAGES 'ENEVA 3WITZERLAND -AY ;= 2 $INIS 0 -ONTEZUMA AND ! 'USMO 0ERFORMANCE TRADE OdS WITH QUASI LINEARLY AMPLIED /&$- THROUGH A TWO BRANCH COMBINING TECHNIQUE )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES !TLANTA '! !PR ;= $ $IVSALAR AND - + 3IMON 4RELLIS CODED MODULATION FOR BITSS TRANS MISSION OVER A FADING MOBILE SATELLITE CHANNEL )%%% * 3ELECT !REAS #OMMUN 3!# &EB ;= $ $IVSALAR AND - + 3IMON 4HE DESIGN OF TRELLIS CODED -03+ FOR FADING CHANNELS 0ERFORMANCE CRITERIA )%%% 4RANS #OMMUN 3EPT ;= $ $IVSALAR AND - + 3IMON 4HE DESIGN OF TRELLIS CODED -03+ FOR FADING CHANNELS 3ET PARTITIONING FOR OPTIMUM CODE DESIGN )%%% 4RANS #OMMUN 3EPT ;= $ $IVSALAR AND - + 3IMON -ULTIPLE SYMBOL DIdERENTIAL DETECTION OF -03+ )%%% 4RANS #OMMUN -AR ;= $ $IVSALAR AND - + 3IMON -AXIMUM LIKELIHOOD DIdERENTIAL DETECTION OF UNCODED AND TRELLIS CODED AMPLITUDE PHASE MODULATION OVER !7'. AND FADING CHANNELS -ETRICS AND PERFORMANCE )%%% 4RANS #OMMUN *AN ;= * $U AND " 6UCETIC 4RELLIS CODED 1!- FOR FADING CHANNELS %UR 4RANS 4ELECOMMUN 2EL 4ECHNOL -AY ;= $ % $UDGEON AND 2 - -ERSEREAU -ULTIDIMENSIONAL DIGITAL SIGNAL PROCESSING 0RENTICE(ALL %NGLEWOOD #LIdS .* ;= 6 %NGELS AND ( 2OHLING $IdERENTIAL MODULATION TECHNIQUES FOR A -"ITS RADIO CHANNEL USING ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING 7IRELESS 0ERS #OMMUN ;= 6 %NGELS AND ( 2OHLING -ULTILEVEL DIdERENTIAL MODULATION TECHNIQUES $!03+ FOR MULTICARRIER TRANSMISSION SYSTEMS %UR 4RANS 4ELECOMMUN 2EL 4ECHNOL .OV ;= $ $ &ALCONER AND & 2 -AGEE !DAPTIVE CHANNEL MEMORY TRUNCATION FOR MAXIMUM LIKELIHOOD SEQUENCE ESTIMATION 4HE "ELL 3YSTEM 4ECHNICAL *OURNAL .OV ;= . &ALLER !N ADAPTIVE SYSTEM FOR DATA COMPRESSION )N 2ECORD OF THE TH !SILOMAR #ONFERENCE ON #IRCUITS 3YSTEMS AND #OMPUTERS PAGES .AVAL 0OSTGRAD UATE 3CHOOL -ONTEREY #ALIFORNIA
;= + &AZEL 0ERFORMANCE OF CONVOLUTIONALLY CODED #$-!/&$- IN A FREQUENCY TIME SELECTIVE FADING CHANNEL AND ITS NEAR FAR RESISTANCE )N 0ROC )NTERN #ONF #OMMUN VOLUME PAGES .EW /RLEANS ,! .OV ;= 2 & &ISCHER AND * " (UBER ! NEW LOADING ALGORITHM FOR DISCRETE MULTITONE TRANSMISSION )N 0ROC 'LOBECOM ,ONDON 5+ ;= $ ' &ORNEY *R -AXIMUM LIKELIHOOD SEQUENCE ESTIMATION OF DIGITAL SEQUENCES IN THE PRESENCE OF INTERSYMBOL INTERFERENCE )%%% 4RANSACTIONS ON )NFORMATION 4HEORY )4 -AY ;= $ ' &ORNEY *R 4HE 6ITERBI ALGORITHM 0ROCEEDINGS OF )%%% -AR ;= 3 ! &REDRICSSON /PTIMUM TRANSMITTING LTER IN DIGITAL 0!- SYSTEMS WITH A 6ITERBI DETECTOR )%%% 4RANSACTIONS ON )NFORMATION 4HEORY )4 *ULY ;= 3 ! &REDRICSSON *OINT OPTIMIZATION OF TRANSMITTER AND RECEIVER LTER IN DIGITAL 0!- SYSTEMS WITH A 6ITERBI DETECTOR )%%% 4RANSACTIONS ON )NFORMATION 4HEORY )4 -AR ;= 0 &RENGER AND ! 3VENSSON ! DECISION DIRECTED COHERENT DETECTOR FOR /&$- )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES !TLANTA '! !PR ;= & 'AGNON AND $ (ACCOUN "OUNDS ON THE ERROR PERFORMANCE OF CODING FOR NON INDEPENDENT 2ICIAN FADING CHANNELS )%%% 4RANS #OMMUN &EB ;= 2 # 'ALLAGER )NFORMATION 4HEORY AND 2ELIABLE #OMMUNICATION *OHN 7ILEY 3ONS .EW 9ORK ;= - 2 'AREY /PTIMAL BINARY SEARCH WITH RESTRICTED MAXIMAL DEPTH 3)!- * #OMPUT *UNE ;= - 'HOSH !NALYSIS OF THE EdECT OF IMPULSE NOISE ON MULTICARRIER AND SINGLE CARRIER 1!- SYSTEMS )%%% 4RANS #OMMUN &EB ;= % . 'ILBERT #ODES BASED ON INACCURATE SOURCE PROBABILITIES )%%% 4RANS )NFORM 4HEORY )4 -AY ;= ' 'OLUB AND # VAN ,OAN -ATRIX #OMPUTATIONS .ORTH /XFORD !CADEMIC *OHNS (OPKINS !CADEMIC 0RESS ND EDITION ;= ! 'RACE /PTIMIZATION 4OOLBOX FOR USE WITH -AT,AB 4HE -ATH7ORKS )NC .ATICK -ASS *UNE ;= 2 'ROSS AND $ 6EENEMAN #LIPPING DISTORTION IN $-4 !$3, SYSTEMS %LECTRON ,ETT .OV
;= - 'UDMUNDSON AND 0 / !NDERSSON !DJACENT CHANNEL INTERFERENCE IN AN /&$SYSTEM )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES !TLANTA '! !PR ;= 2 (AAS !PPLICATION DES TRANSMISSIONS PORTEUSES MULTIPLES AUX COMMUNICATIONS RADIO MOBILES 0HD THESIS %COLE .ATIONAL 3UPRIEURE DES 4LCOMMUNICATIONS 0ARIS &RANCE *AN )N %NGLISH ;= 0 ( (ALPERN /PTIMUM NITE DURATION .YQUIST SIGNALS )%%% 4RANS #OMMUN #/- *UNE ;= - (ANKAMER ! MODIED PROCEDURE WITH REDUCED MEMORY REQUIREMENT )%%% 4RANS #OMMUN #/- *UNE ;= 4 (ASHIMOTO ! LIST TYPE REDUCED CONSTRAINT GENERALIZATION OF THE VITERBI ALGO RITHM )%%% 4RANSACTIONS ON )NFORMATION 4HEORY )4 .OV ;= 0 (O * #AVERS AND * 6ARALDI 4HE EdECTS OF CONSTELLATION DENSITY ON TRELLIS CODED MODULATION IN FADING CHANNELS )%%% 4RANS 6EHIC 4ECHNOL !UG ;= 0 (O AND $ &UNG %RROR PERFORMANCE OF MULTIPLE SYMBOL DIdERENTIAL DETECTION OF 03+ SIGNALS TRANSMITTED OVER CORRELATED 2AYLEIGH FADING CHANNELS )%%% 4RANS #OMMUN /CT ;= 4 # (U AND + # 4AN /PTIMAL COMPUTER SEARCH TREES AND VARIABLE LENGTH ALPHABETIC CODES 3)!- * !PPL -ATH ;= 4 # (U AND + # 4AN 0ATH LENGTH OF BINARY SEARCH TREES 3)!- * !PPL -ATH -AR ;= * (UANG AND , , #AMPBELL 4RELLIS CODED -$03+ IN CORRELATED AND SHADOWED 2ICIAN FADING CHANNELS )%%% 4RANS #OMMUN .OV ;= $ ! (UdMAN ! METHOD FOR THE CONSTRUCTION OF MINIMUM REDUNDANCY CODES )N 0ROCEEDINGS OF THE )2% NUMBER PAGES 3EPT ;= 0 (HER 4#- ON FREQUENCY SELECTIVE LAND MOBILE FADING CHANNELS )N 0ROC 4IR RENIA )NT 7ORKSHOP $IGITAL #OMMUN 4IRRENIA )TALY 3EPT ;= 0 (HER ! STATISTICAL DISCRETE TIME MODEL FOR THE 73353 MULTIPATH CHANNEL )%%% 4RANS #OMMUN .OV ;= 0 (HER * (AGENAUER % /dER # 2APP AND ( 3CHULZE 0ERFORMANCE OF AN 2#0# CODED /&$- BASED DIGITAL AUDIO BRAODCASTING $!" SYSTEM )N 0ROC 'LOBECOM VOLUME PAGES ;= 7 # *AKES -ICROWAVE -OBILE #OMMUNICATIONS *OHN 7ILEY AND 3ONS .EW 9ORK
;= 3 ( *AMALI AND 4 ,E .GOC ! NEW STATE 03+ 4#- SCHEME FOR FAST FADING SHADOWED MOBILE RADIO CHANNELS )%%% 4RANS 6EHIC 4ECHNOL &EB ;= 2 *OHANNESSON )NFORMATIONSTEORI GRUNDVALEN FR TELE KOMMUNIKATION 3TU DENTLITTERATUR ,UND 3WEDEN ;= ! *ONES AND 4 7ILKINSON #OMBINED CODING FOR ERROR CONTROL AND INCREASED RO BUSTNESS TO SYSTEM NONLINEARITIES IN /&$- )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES !TLANTA '! !PR ;= ! *ONES 4 7ILKINSON AND 3 "ARTON "LOCK CODING SCHEME FOR REDUCTION OF PEAK TO MEAN ENVELOPE POWER RATIO OF MULTICARRIER TRANSMISSION SCHEMES %LECTRON ,ETT $EC ;= ' + +ALEH #HANNEL EQUALIZATION FOR BLOCK TRANSMISSION SYSTEMS )%%% *OURNAL 3EL !REAS #OM *AN ;= ) +ALET 4HE MULTITONE CHANNEL )%%% 4RANS #OMMUN &EB ;= ) +ALET AND 3 3HAMAI /N THE CAPACITY OF A TWISTED WIRE PAIR 'AUSSIAN MODEL )%%% 4RANS #OMMUN -AR ;= 4 +ELLER ET AL 2EPORT ON DIGITAL AUDIO RADIO LABORATORY TESTS 4ECHNICAL REPORT %LECTRONIC )NDUSTRIES !SSOCIATION -AY ;= , ' +RAFT ! DEVICE FOR QUANTIZING GROUPING AND CODING AMPLITUDE MODULATED PULSES -ASTERS THESIS $EPT OF %% -)4 #AMBRIDGE -ASS ;= ( * ,ANDAU AND ( / 0OLLAK 0ROLATE SPHERIODAL WAVE FUNCTIONS &OURIER ANALYSIS AND UNCERTAINTY ))) 4HE DIMENSION OF THE SPACE OF ESSENTIALLY TIME AND BAND LIMITED SIGNALS "ELL 3YSTEM 4ECH * ;= , , ,ARMORE AND $ 3 (IRSCHBERG ! FAST ALGORITHM FOR OPTIMAL LENGTH LIMITED (UdMAN CODES *OURNAL OF THE !#- *ULY ;= 4 ,ARSSON ! 3TATE 3PACE 0ARTITIONING !PPROACH TO 4RELLIS $ECODING 0H$ THESIS #HALMERS 5NIVERSITY OF 4ECHNOLOGY 'TEBORG ;= " ,E &LOCH - !LARD AND # "ERROU #ODED ORTHOGONAL FREQUENCY DIVISION MUL TIPLEXING 0ROC )%%% *UNE ;= " ,E &LOCH 2 (ALBERT ,ASSALLE AND $ #ASTELAIN $IGITAL SOUND BROADCASTING TO MOBILE RECEIVERS )%%% 4RANS #ONSUMER %LECTRONICS !UG ;= $ ! ,ELEWER AND $ 3 (IRSCHBERG $ATA COMPRESSION !#- #OMPUTING 3URVEYS 3EPT ;= ( ,I AND * + #AVERS !N ADAPTIVE LTERING TECHNIQUE FOR PILOT AIDED TRANSMISSION SYSTEMS )%%% 4RANS 6EHIC 4ECHNOL !UG
;= # -ANAGEMENT #OMMITTEE $IGITAL LAND MOBILE RADIO COMMUNICATION &INAL RE PORT 4ECHNICAL REPORT #OMMISSION OF THE %UROPEAN #OMMUNICATION ,UXEMBOURG ;= " -ANDERSSON 0 / "RJESSON . ' (OLMER + ,INDSTRM AND ' 3ALOMONSSON $IGITAL LTERING OF ULTRASONIC ECHO SIGNALS FOR INCREASED AXIAL RESOLUTION )N 0ROC .ORDIC -EETING -ED "IOL %NG ,INKPING 3WEDEN *UNE ;= " -ARTI %UROPEAN ACTIVITIES ON DIGITAL TELEVISION BROADCASTING FROM COMPANY TO COOPERATIVE PROJECTS %"5 2EVIEW 4ECHNICAL ;= 2 ' -C+AY 0 * -C,ANE AND % "IGLIERI %RROR BOUNDS FOR TRELLIS CODED -03+ ON A FADING MOBILE SATELLITE CHANNEL )%%% 4RANS #OMMUN $EC ;= 0 * -C,ANE ! RESIDUAL INTERSYMBOL INTERFERENCE ERROR BOUND FOR TRUNCATED STATE 6ITERBI DETECTORS )%%% 4RANSACTIONS ON )NFORMATION 4HEORY )4 ;= - -OENECLAEY AND - VAN "LADEL $IGITAL ($46 BROADCASTING OVER THE #!46 DISTRIBUTION SYSTEM 3IGNAL 0ROC )MAGE #OMMUN $EC ;= - , -OHER AND * ( ,ODGE 4#-0 ! MODULATION AND CODING STRATEGY FOR 2ICIAN FADING CHANNELS )%%% * 3ELECT !REAS #OMMUN $EC ;= 0 -ONSEN &EEDBACK EQUALIZATION FOR FADING DISPERSIVE CHANNELS )%%% 4RANSAC TIONS ON )NFORMATION 4HEORY )4 ;= 0 -OOSE ! TECHNIQUE FOR ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING FREQUENCY OdSET CORRECTION )%%% 4RANS #OMMUN /CT ;= ! -LLER /&$- TRANSMISSION OVER TIME VARIANT CHANNELS )N 0ROC )NT "ROADC #ONV NUMBER PAGES !MSTERDAM .ETHERLANDS 3EPT ;= * .ILSSON #ODING TO CONTROL THE SIGNAL WAVEFORM IN - ARY 03+ MULTICARRIER COMMUNICATIONS )N 0ROC 2ADIOVETEN +ONF PAGES ,ULE 3WEDEN *UNE ;= 2 VAN .OBELEN AND $ 0 4AYLOR !NALYSIS OF THE PAIRWISE ERROR PROBABILITY OF NONIN TERLEAVED CODES ON THE 2AYLEIGH FADING CHANNEL )%%% 4RANS #OMMUN !PR ;= ! /PPENHEIM AND 2 3CHAFER $ISCRETE TIME SIGNAL PROCESSING 0RENTICE (ALL ;= / 0AHLM 0 / "RJESSON AND / 7ERNER #OMPACT DIGITAL STORAGE OF %#'S #OMPUTER 0ROGRAMS IN "IOMEDICINE -AY ;= ! 0ELED AND ! 2UIZ &REQUENCY DOMAIN DATA TRANSMISSION USING REDUCED COMPU TATIONAL COMPLEXITY ALGORITHMS )N 0ROC )%%% )NT #ONF !COUST 3PEECH 3IGNAL 0ROCESSING PAGES $ENVER #/
;= 4 0OLLET AND - -OENECLAEY 3YNCHRONIZABILITY OF /&$- SIGNALS )N 0ROC 'LOBE COM VOLUME PAGES 3INGAPORE .OV ;= 4 0OLLET 0 3PRUYT AND - -OENECLAEY 4HE "%2 PERFORMANCE OF /&$- SYSTEMS USING NON SYNCHRONIZED SAMPLING )N 0ROC 'LOBECOM VOLUME PAGES 3AN &RANCISCO #! .OV ;= 4 0OLLET - VAN "LADEL AND - -OENECLAEY "%2 SENSITIVITY OF /&$- SYS TEMS TO CARRIER FREQUENCY OdSET AND 7IENER PHASE NOISE )%%% 4RANS #OMMUN &EB-AR!PR ;= * 0ROAKIS $IGITAL COMMUNICATIONS 0RENTICE (ALL RD EDITION ;= 3 1URESHI AND % .EWHALL !N ADAPTIVE RECEIVER FOR DATA TRANSMISSION OVER TIME DISPERSIVE CHANNELS )%%% 4RANSACTIONS ON )NFORMATION 4HEORY )4 *ULY ;= + 2AMCHANDRAN ! /RTEGA + - 5Z AND - 6ETTERLI -ULTIRESOLUTION BROAD CAST FOR DIGITAL ($46 USING JOINT SOURCECHANNEL CODING )%%% * 3ELECT !REAS #OMMUN *AN ;= * . 2EDDY !PPLIED &UNCTIONAL !NALYSIS AND 6ARIATIONAL -ETHODS IN %NGINEERING %NGINEERING -ECHANICS -C'RAW (ILL 3INGAPORE ;= # 2EINERS AND ( 2OHLING -ULTICARRIER TRANSMISSION TECHNIQUE IN CELLULAR MOBILE COMMUNICATION SYSTEMS )N 0ROC )%%% 6EHIC 4ECHNOL #ONF PAGES 3TOCKHOLM 3WEDEN *UNE ;= * 2INNE AND - 2ENFORS 4HE BEHAVIOUR OF ORTHOGONAL FREQUENCY DIVISION MULTI PLEXING IN AN AMPLITUDE LIMITING CHANNEL )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES .EW /RLEANS ,! -AY ;= ( 2OHLING + "RNINGHAUS AND 4 -LLER 0ERFORMANCE OF COHERENT /&$- #$-! FOR BROADBAND MOBILE COMMUNICATION )N 0ROC 2!#% -OBILE #OMMUN 3UMMIT PAGES #ASCAIS .OV ;= ( 2OHLING AND 2 'RNHEID -ULTICARRIER TRANSMISSION TECHNIQUE IN MOBILE COM MUNICATION SYSTEMS )N 0ROC 2!#% -OBILE #OMMUN 3UMMIT PAGES #ASCAIS .OV ;= 7 2UDIN 0RINCIPLES OF -ATHEMATICAL !NALYSIS -ATHEMATICS -C'RAW (ILL 3IN GAPORE EDITION ;= - 2USSELL AND ' 3TBER )NTERCHANNEL INTERFERENCE ANALYSIS OF /&$- IN A MOBILE ENVIRONMENT )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES #HICAGO ), *ULY ;= - 3ABLATASH 4RANSMISSION OF ALL DIGITAL ADVANCED TELEVISION 3TATE OF THE ART AND FUTURE DIRECTIONS )%%% 4RANS "ROADC *UNE
;= " 2 3ALTZBERG 0ERFORMANCE OF AN EbCIENT PARALLEL DATA TRANSMISSION SYSTEM )%%% 4RANS #OMMUN #/- $EC ;= - 3ANDELL * * VAN DE "EEK AND 0 / "RJESSON 4IMING AND FREQUENCY SYNCHRO NIZATION IN /&$- SYSTEMS USING THE CYCLIC PREX )N )NTERN 3YMP 3YNCH PAGES %SSEN 'ERMANY $EC ;= - 3ANDELL 3 + 7ILSON AND 0 / "RJESSON 0ERFORMANCE ANALYSIS OF CODED /&$- ON FADING CHANNELS WITH NON IDEAL INTERLEAVING AND CHANNEL KNOWLEDGE 2E SEARCH 2EPORT 45,%! $IV OF 3IGNAL 0ROCESSING ,ULE 5NIVERSITY OF 4ECH NOLOGY 3EPT ;= ' 3ANTELLA /&$- WITH GUARD INTERVAL AND SUBCHANNEL EQUALIZATION IN A RESOLUTION TRANSMISSION SCHEME FOR DIGITAL TELEVISION BROADCASTING )N 0ROC )NTERN #ONF #OMMUN VOLUME PAGES .EW /RLEANS ,! -AY ;= ( 3ARI ' +ARAM AND ) *EANCLAUDE 4RANSMISSION TECHNIQUES FOR DIGITAL TERRISTRIAL 46 BROADCASTING )%%% #OMMUN -AG &EB ;= , , 3CHARF 3TATISTICAL SIGNAL PROCESSING $ETECTION ESTIMATION AND TIME SERIES ANALYSIS !DDISON 7ESLEY ;= # 3CHLEGEL 4RELLIS CODED MODULATION ON TIME SELECTIVE FADING CHANNELS )%%% 4RANS #OMMUN ;= # 3CHLEGEL AND $ * #OSTELLO "ANDWIDTH EbCIENT CODING FOR FADING CHANNELS #ODE CONSTRUCTION AND PERFORMANCE ANALYSIS )%%% * 3ELECT !REAS #OMMUN $EC ;= 7 & 3CHREIBER !DVANCED TELEVISION SYSTEMS FOR TERRISTRIAL BROADCASTING 3OME PROBLEMS AND SOME PROPOSED SOLUTIONS 0ROC )%%% *UNE ;= # % 3HANNON ! MATHEMATICAL THEORY OF COMMUNICATION PART ) "ELL 3YST 4ECH * *ULY ;= ! 3IEMINSKI &AST DECODING OF THE (UdMAN CODES )NFORMATION 0ROCESSING ,ETTERS *AN ;= - + 3IMON AND $ $IVSALAR 4HE PERFORMANCE OF TRELLIS CODED MULTILEVEL $03+ ON A FADING MOBILE SATELLITE CHANNEL )%%% 4RANS 6EHIC 4ECHNOL -AY ;= 3 " 3LIMANE AND 4 ,E .GOC 4IGHT BOUNDS ON THE ERROR PROBABILITY OF CODED MODULATION SCHEMES IN 2AYLEIGH FADING CHANNELS )%%% 4RANS 6EHIC 4ECHNOL &EB ;= ' 3TRANG ,INEAR !LGEBRA AND )TS !PPLICATIONS !CADEMIC 0RESS /RLANDO &LORIDA EDITION
;= ' , 3TBER AND - 2USSELL 4ERRESTRIAL DIGITAL VIDEO BROADCASTING FOR MOBILE RECEPTION USING /&$- )N 0ROC 'LOBECOM VOLUME PAGES 3INGAPORE .OV ;= # % 7 3UNDBERG AND . 3ESHADRI #ODED MODULATION FOR FADING CHANNELS !N OVERVIEW %UR 4RANS 4ELECOMMUN 2EL 4ECHNOL -AY ;= ( 4ANAKA $ATA STRUCTURE OF HUdMAN CODES AND ITS APPLICATION TO EbCIENT ENCODING AND DECODING )%%% 4RANS )NFORM 4HEORY )4 *AN ;= # 4ELLAMBURA 1 7ANG AND 6 + "HARGAVA ! PERFORMANCE ANALYSIS OF TRELLIS CODED MODULATION SCHEMES OVER 2ICIAN FADING CHANNELS )%%% 4RANS 6EHIC 4ECH NOL .OV ;= # 4ELLAMBURA 1 7ANG AND 6 + "HARGAVA 0ERFORMANCE OF TRELLIS CODED MODU LATION SCHEMES ON SHADOWED MOBILE SATELLITE COMMUNICATION CHANNELS )%%% 4RANS 6EHIC 4ECHNOL &EB ;= 0 * 4OURTIER 2 -ONNIER AND 0 ,OPEZ -ULTICARRIER MODEM FOR DIGITAL ($46 TERRESTRIAL BROADCASTING 3IGNAL 0ROC )MAGE #OMMUN $EC ;= * # 4U 4HEORY DESIGN AND APPLICATION OF MULTI CHANNEL MODULATION FOR DIGITAL COMMUNICATIONS 0H$ THESIS 3TANFORD 5NIVERSITY #! *UNE ;= ! 6AHLIN AND . (OLTE /PTIMAL NITE DURATION PULSES FOR /&$- )%%% 4RANS #OMMUN *AN ;= ) 6AJDA %STIMATION AND CODING OF INFORMATION SOURCES 0ROBLEMS OF #ONTROL AND )NFORMATION 4HEORY ;= - 6ETTERLI AND * +OVACEVIC 7AVELETS AND SUBBAND CODING 0RENTICE (ALL %NGLE WOOD #LIdS .* ;= % 6ITERBO AND + &AZEL (OW TO COMBAT LONG ECHOES IN /&$- TRANSMISSION SCHEMES 3UB CHANNEL EQUALIZATION OR MORE POWERFUL CHANNEL CODING )N 0ROC 'LOBECOM VOLUME PAGES 3INGAPORE .OV ;= $ # VAN 6OORHIS #ONSTRUCTING CODES WITH BOUNDED CODE WORD LENGTHS )%%% 4RANS )NFORM 4HEORY )4 -AR ;= " 6UCETIC AND * $U 4HE EdECTS OF PHASE NOISE ON TRELLIS CODED MODULATION OVER 'AUSSIAN AND FADING CHANNELS )%%% 4RANS #OMMUN ;= - 7AHLQVIST 2 ,ARSSON AND # STBERG 4IME SYNCHRONIZATION IN THE UPLINK OF AN /&$- SYSTEM )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES !TLANTA '! !PR ;= 7 $ 7ARNER AND # ,EUNG /&$-&- FRAME SYNCHRONIZATION FOR MOBILE RADIO DATA COMMUNICATION )%%% 4RANS 6EHIC 4ECHNOL !UG
;= , 7EI AND # 3CHLEGEL 3YNCHRONIZATION REQUIREMENTS FOR MULTI USER /&$- ON SATELLITE MOBILE AND TWO PATH 2AYLEIGH FADING CHANNELS )%%% 4RANS #OMMUN &EB-AR!PR ;= 3 " 7EINSTEIN AND 0 - %BERT $ATA TRANSMISSION BY FREQUENCY DIVISION MUL TIPLEXING USING THE DISCRETE &OURIER TRANSFORM )%%% 4RANS #OMMUN #/- /CT ;= * * 7ERNER 4HE ($3, ENVIRONMENT )%%% * 3ELECT !REAS #OMMUN 3!# !UG ;= 3 ' 7ILSON $IGITAL MODULATION AND CODING 0RENTICE (ALL .EW *ERSEY 53! ;= 3 + 7ILSON $IGITAL AUDIO BROADCASTING IN A FADING AND DISPERSIVE CHANNEL 0H$ THESIS 3TANFORD 5NIVERSITY #! !UG ;= 3 + 7ILSON 2 % +HAYATA AND * - #IOb 1!- MODULATION WITH ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING IN A 2AYLEIGH FADING ENVIRONMENT )N 0ROC )%%% 6EHIC 4ECHNOL #ONF VOLUME PAGES 3TOCKHOLM 3WEDEN *UNE ;= 9 7U AND " #ARON $IGITAL TELEVISION TERRESTRIAL BROADCASTING )%%% #OMMUN -AG -AY ;= & 8IONG ! :ERIK AND % 3HWEDYK 3EQUENTIAL SEQUENCE ESTIMATION FOR CHANNELS WITH INTERSYMBOL INTERFERENCE OF NITE OR INNITE LENGTH )%%% 4RANSACTIONS ON #OMMUNICATIONS ;= ' 9OUNG + 4 &OSTER AND * 7 #OOK "ROADBAND MULTIMEDIA DELIVERY OVER COPPER "4 4ECHNOL *OURN /CT ;= % :EHAVI AND * + 7OLF /N THE PERFORMANCE EVALUATION OF TRELLIS CODES )%%% 4RANS )NFORM 4HEORY )4 -AR ;= 4 . :OGAKIS * 4 * !SLANIS AND * - #IOb !NALYSIS OF A CONCATENATED CODING SCHEME FOR A DISCRETE MULTITONE MODULATION SYSTEM )N 0ROC )%%% -ILITARY #OMMUN #ONF VOLUME PAGES ,ONG "RANCH .* /CT ;= 4 . :OGAKIS 0 3 #HOW * 4 !SLANIS AND * - #IOb )MPULSE NOISE MITIGATION STRATEGIES FOR MULTICARRIER MODULATION )N 0ROC )NTERN #ONF #OMMUN PAGES 'ENEVA -AY ;= 4 . :OGAKIS AND * - #IOb 4HE EdECTS OF TIMING JITTER ON THE PERFORMANCE OF A DISCRETE MULTITONE SYSTEM )%%% 4RANS #OMMUN *ULY ;= 7 9 :OU AND 9 7U #/&$- !N OVERVIEW )%%% 4RANS "ROADC -AR ;= 0 DLING ,OW #OMPLEXITY $IGITAL 2ECEIVERS ,ICENTIATE THESIS ,ULE 5NIVERSITY OF 4ECHNOLOGY
;= 0 DLING 4 +OSKI AND 0 / "RJESSON ! REDUCED COMPLEXITY 6ITERBI EQUALIZER USED IN CONJUNCTION WITH A PULSE SHAPING METHOD )N 0ROC )NTERN 3YMP 3IGNAL 0ROC !PPL PAGES 1UEENSLAND !USTRALIA !UG

Edf 96 T

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Edf 96 T

Uploaded by

Copyright:

Available Formats

,OW

COMPLEXITY ALGORITHMS IN DIGITAL RECEIVERS

%STIMATION AND DETECTION

%STIMATION AND DETECTION

%XPRESSING THE /&$- SYSTEM IN THE FORM OF A MATRIX EXPRESSION YIELDS X  7G

%STIMATION AND DETECTION

0ART  !N INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING

#ONTINUOUS TIME MODEL

)NTRODUCTION TO /&$WHERE 3  -6

 WHICH IMPLIES THAT

E 3 ` S S CS 3INCE THE TRANSMITTER LTERS J S ARE ORTHOGO J BO M : S CS 

WHERE M IS ADDITIVE WHITE 'AUSSIAN NOISE !7'. 

$ISCRETE TIME MODEL

! TIME FREQUENCY INTERPRETATION

)NTRODUCTION TO /&$#HAN NEL

&IGURE  .EAR END CROSSTALK .%84 

AND THE SPECTRAL DENSITY OF &%84 AS /

3AMPLING FREQUENCY SYNCHRONIZATION

#ARRIER FREQUENCY SYNCHRONIZATION

)NTRODUCTION TO /&$THE DATA

  #ODING ON FADING CHANNELS

)NTRODUCTION TO /&$3.2 "ITS

4IME FREQUENCY LATTICE

WHERE THE FUNCTIONS

S ARE TRANSLATIONS OF A PROTOTYPE FUNCTION O S 

)N /&$- THESE FUNCTIONS FULL l S 

  SINCE THE CROSS AMBIGUITY

$&4 BASED CHANNEL ESTIMATORS

$&4 BASED CHANNEL ESTIMATORS

$&4 BASED CHANNEL ESTIMATORS

$&4 BASED CHANNEL ESTIMATORS

$&4 BASED ESTIMATORS

3AMPLE SPACED CHANNEL

.ONSAMPLE SPACED CHANNEL

$&4 BASED CHANNEL ESTIMATORS

WHERE THE XED WEIGHTING MATRIX IS GIVEN BY t 6 1 1

$&4 BASED CHANNEL ESTIMATORS

%STI MATOR ,--3% ! " # ,3

,INEAR TRANSFORMATION .! v w 1, ,  1    w v $, ,  1!   w v ), ,  1"    .!

%dECTIVE MATRIX SIZE . b. - b- b - DIAG - b - IDENT .!

2EQUIRED MULTATTEN .  LOG . LOG .

$&4 BASED CHANNEL ESTIMATORS

%STIMATOR ,--3% ! " # ,3

!VERAGE -3%  wJ0 n 

wJ, n wJ,  2-1 oJ oJ 

wJ, EIGENVALUES OF 2FF,  o J DIAGONAL ELEMENTS OF 2FF IN DECREASING ORDER -3% 

$&4 BASED CHANNEL ESTIMATORS

$&4 BASED CHANNEL ESTIMATORS

-ULTIPLICATIONS PER ESTIMATED ATTENUATION

$&4 BASED CHANNEL ESTIMATORS

$&4 BASED CHANNEL ESTIMATORS

ARE DENED ABOVE

)MPLEMENTING THIS ESTIMATOR AS A MATRIX MULTIPLICATION AS IN  REQUIRES -

$&4 BASED CHANNEL ESTIMATORS

)MPOSING THE RST RESTRICTION ON THE LINEAR TRANSFORMATION v w 0  0    

THE MINIMAL -3% IS OBTAINED IF 0 WHERE 1

IS THE UPPER LEFT , b , CORNER OF 1  t (` 1 u n

4HE ,2$ OF THIS ESTIMATOR IS  ,2$  3Q@BD 1 ,

S ARE EIGENVALUES OF 1 ,2$  3Q@BD 5 `

)MPLEMENTING THIS ESTIMATOR ACCORDING TO &IGURE  REQUIRES KNF -

WE OBTAIN A MINIMAL -3% IF o p   o

4HE -3% OF THIS ESTIMATOR IS

%STIMATOR !VERAGE -3%  0  ,--3% PEI 

%XPRESSING THE /&$- SYSTEM IN THE FORM OF A MATRIX EXPRESSION YIELDS X 7G

0ART !N INTRODUCTION TO ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING

)NTRODUCTION TO /&$WHERE 3 -6

WHICH IMPLIES THAT

E 3 ` S S CS 3INCE THE TRANSMITTER LTERS J S ARE ORTHOGO J BO M : S CS

WHERE M IS ADDITIVE WHITE 'AUSSIAN NOISE !7'.

&IGURE .EAR END CROSSTALK .%84

)NTRODUCTION TO /&$THE DATA

#ODING ON FADING CHANNELS

S ARE TRANSLATIONS OF A PROTOTYPE FUNCTION O S

)N /&$- THESE FUNCTIONS FULL l S

SINCE THE CROSS AMBIGUITY

WHERE THE XED WEIGHTING MATRIX IS GIVEN BY t 61 1

,INEAR TRANSFORMATION .! v w 1, , 1 w v $, , 1! w v ), , 1" .!

%dECTIVE MATRIX SIZE . b. - b- b - DIAG - b - IDENT .!

2EQUIRED MULTATTEN . LOG . LOG .

!VERAGE -3% wJ0 n

wJ, n wJ, 2-1 oJ oJ

wJ, EIGENVALUES OF 2FF, o J DIAGONAL ELEMENTS OF 2FF IN DECREASING ORDER -3%

ARE DENED ABOVE

)MPLEMENTING THIS ESTIMATOR AS A MATRIX MULTIPLICATION AS IN REQUIRES -

)MPOSING THE RST RESTRICTION ON THE LINEAR TRANSFORMATION v w 0 0

IS THE UPPER LEFT , b , CORNER OF 1 t (` 1 u n

4HE ,2$ OF THIS ESTIMATOR IS ,2$ 3Q@BD 1 ,

S ARE EIGENVALUES OF 1 ,2$ 3Q@BD 5 `

)MPLEMENTING THIS ESTIMATOR ACCORDING TO &IGURE REQUIRES KNF -

WE OBTAIN A MINIMAL -3% IF o p o

%STIMATOR !VERAGE -3% 0 ,--3% PEI

0ART /&$- CHANNEL ESTIMATION BY SINGULAR VALUE DECOMPOSITION

36$ BASED CHANNEL ESTIMATORS 7 X

&IGURE "LOCK DIAGRAM OF THE RANK`O CHANNEL ESTIMATOR

1E E 4 CF )T CAN BE INTERPRETED AS THE UNDER CORRELATION MISMATCH SINCE s } E G 4 1E E 4

4HE OPTIMAL RANK REDUCTION IS FOUND FROM THE CORRELATION MATRICES N O B 1 B $ GG 1

1B AND THE 36$ 1

5SING THE CHANNEL MODEL IN THE ATTENUATION ON TONE J BECOMES G

CAN BE EXPRESSED AS ~ S INDEPENDENT : : 9 aaa E ~ Q

q 3YNCHRONIZED CHANNEL 4HE PROBABILITY DISTRIBUTIONS FOR THE DELAYS ARE E

p ~ | + IF ~ : +< H , OTHERWISE

3UBSTITUTING IN AND

" 3UBSTITUTED IN AND IF L M IF L M

#OMPARATIVE STUDY OF CHANNEL ESTIMATORS ATTENUATIONS J G & -3

4HE CONSTRAINT LENGTH IS M AND A TAIL OF M

aE aS $ F& E S & E ` aE S ` aS G

)N THE ANALYZED SYSTEM WE HAVE - + AND E E

WHICH GIVES

MULTIPLICATIONS PER ATTENUATION

" MULTIPLICATIONS PER ATTENUATION

-AX REL $OPPLER FREQUENCY 3.2 6ARYING

WHERE A DENOTES THE NUMBER OF INFORMATION BITSSYMBOL )N OUR CASE WE HAVE A