Z T E Communications

www.zte.com.cn/magazine/English ISSN 1673-5188CODEN ZCTOAKZTE COMMUNICATIONS VOLUME 9 NUMBER 4 DECEMBER 2011ZTE COMMUNICATIONSDecember 2011, Vol.9 No.4Special Topics:Advances in Digital Front-End and Software RF Processing: Part ⅡAdvances in Mobile Data CommunicationsD:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P1

Frost & Sullivan Recognizes ZTE as 2011LTE Vendor of the YearZTE Corporation announced on November 10, 2011that it has been named LTE Vendor of the Year by leadingconsulting firm Frost & Sullivan.The Frost & Sullivan LTE Vendor of the Year award isgiven to companies that have achieved excellence inLTE. ZTE won the award for its growing number ofsecured contracts and also for its patent applications.The company’s industry leading equipment has enteredhigh-end markets such as Europe, Japan, and theZTE Corporation announced on November 3, 2011 thatit has become the global leader in the CDMA base stationmarket, with a 32.6 percent share in first half 2011.According to a recent IDC analytical report on theglobal CDMA market, ZTE has increased its shipments ofCDMA base stations steadily in recent years. As of theend of first half of 2011, the company’s shipment ofbase-stations had exceeded 320 thousand units,pushing it to the top spot in the global CDMA base stationmarket.ZTE is the first company in the telecom industry tointroduce a CDMA/LTE dual-mode system. Thisinnovative CDMA/LTE product can assist operators tosmoothly upgrade their existing networks to futurenetworks and faster evolve to LTE systems. This solution,along with the company’s EV-DO Rev. B system, isrecognized as industry leading.The global CDMA market has grown in recent years,especially in Asia Pacific, where it has expanded rapidly.Coupled with a loss in market share by severalestablished North American CDMA suppliers, this hasprompted ZTE to seek new opportunities in CDMAmarkets across the globe. Through these efforts, theUnited States. ZTE is one of the world’s most competitiveLTE solutions providers.ZTE’s influence in the global LTE market has steadilygrown. In first half 2011, the number of new LTE contractsexceeded the total number of LTE contracts for the wholeof 2010. ZTE has won 28 contracts for LTE commercialapplication and has deployed test networks inconjunction with more than 90 operators worldwide. Todate, more than 100,000 ZTE LTE terminal units havebeen ordered worldwide. The terminals, which arepurchased by high-end operators, contribute to thecommercialization of LTE technology.“ZTE is honored to be recognized by Frost & Sullivanas a leader in the LTE field,”said Mr. Richard Lihe Ye, thesenior director of wireless product operation for ZTECorporation.“The award illustrates the company’scommitment to the development and commercializationof next-generation technology.”ZTE is a leader in developing LTE technologies. Todate, ZTE has applied to ETSI for 235 essential patentsfor LTE standards, accounting for approximately7 per cent of the total number of EP applications globally.(ZTE Corporation)ZTE Becomes Global Leader in CDMA BaseStation Market with 33% Sharecompany has acquired a leading market share inemerging nations such as, China, Indonesia, and India. Italso has achieved significant breakthroughs in NorthAmerica.IDC Analyst John Byrne believes that“ZTE is the leaderof the global CDMA market and will be in the strongestposition in emerging global markets, as been the case inthe global CDMA market. ZTE will continue to creategreater value for operators.”As of the end of first half 2011, ZTE’s CDMA productshave been put into large-scale commercial use by morethan 120 operators in more than 70 countries worldwide.The company has built more than 80 CDMA20001xEV-DO networks in 60 countries and regions includingthe Czech Republic, India, Indonesia, and the UnitedStates in the same period. In addition, ZTE has won 23contracts for commercial application of LTE solutionsfrom operators including CSL, Telenor, Sonaecom, andH3G and has deployed LTE test networks in conjunctionwith more than 80 operators across the globe. To date,ZTE has received orders for more than 100,000 LTEterminals from several leading global operators.(ZTE Corporation)D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P2

Special TopicPolyphase Filter Banks for Embedded Sample Rate Changes in Digital Radio Front-EndsMehmood Awan, Yannick Le Moullec, Peter Koch, and Fred HarrisSingel ChannelReceiverRFLO-1_BIFBBADCADCDigital BBBaseBandProcDigitalIFLO-1_BDigital BBBaseBandProcSplitter Network…RFLO-1_A…LO-1_CLO-N_BIF…ADCADCBaseBandProc…RFLO-AIFADCLO-1_CLO-N_BSingle ChannelReceiver…BaseBandProc…LO-N_ALO-N_CLO-N_C(a)(b)ADC: analog-to-digital converter BB: baseband IF: internediate frequency LO: local oscillator RF: radio frequency▲Figure 1. (a) First generation of digital radio architecture for an N-channel receiver, and (b) Second generation of digital radio architecturefor an N-channel receiver.advantages: the spectral images arecontrolled so that they are below thequantization noise floor of the ADCinvolved in the conversion process, andthe digital filters before and after themixers are designed to have linearphase characteristics [2]. The secondgeneration of radio, with digitalfront-end, is a realizable version ofSDR. The range of applications forsecond-generation architecture,shown in Fig. 1(b), is restricted to thosewith IF center frequencies of a couple ofhundred megahertz. This is due to thelimited dynamic range of high-speedADCs. The dynamic range is oftenextended by using a hybrid scheme inwhich the initial complexdown-conversion is performed byanalog I/Q mixers, and channelization isperformed digitally after the ADC. DSPtechniques are applied to the digitizedI/Q data to balance the gain and phaseoffsets in the analog ADC [3].A digital front-end with a standarddesign that includes frequencyselection, bandwidth reduction, andsample rate reduction is one of the mostpower- and time-critical functionalitiesof an SDR terminal. This is due to thelarge bandwidth and high dynamicrange of the signal to be processed.Consequently, the digital signals mayhave high sample rates and large wordlengths. High sample rates not onlyincrease power consumption but alsomake the use of time-shared hardwareinfeasible [4]. On the other hand,multirate signal processing specifiesnew ways of performing DSP tasks, andthese ways are not normally available intraditional DSP designs. A multiratepolyphase filter can perform the tasksof a multichannel receiver. In such areceiver, an input signal iscomposed of many equal-bandwidth,equally spacedfrequency-division-multiplexed (FDM)channels. These channels are digitallydown-converted to baseband(bandwidth is constrained by digitalfilters) and subjected to a sample ratereduction commensurate with thebandwidth reduction. This significantlyreduces the number of systemresources required to performmultichannel processing and,consequently, reduces costs [2], [3].The remainder of this paper isorganized as follows: In section 2, webriefly introduce a polyphasechannelizer and describe how it isformulated from a conventionalchannelizer. In section 3, we categorizethe embedded resampling casesdescribed in [5] into five differentresampling cases: maximallydecimated, under-decimated,over-decimated, and combinedup- and down-sampled with singleand multiple commutator stride lengths.In section 4, we perform MatLabsimulations and the simulation resultsdemonstrate the performance ofpolyphase channelizers that deliver thetargeted output sample rates. Section 5concludes the paper.2 IntroductionA multirate polyphase filter canperform the tasks of a multichannelreceiver. These tasks are equivalent todown-conversion, filtering, andresampling of multiple narrowbandsignals [5]. The step-by-stepconversion of a standardsingle-channel demodulator into amultichannel polyphase channelizer isdescribed in [2] and [3]. A briefintroduction is given here. In thestandard single-channel demodulationprocess, shown in Fig. 2(a), thecarrier-centered spectrum is translatedto baseband (where a filter reduces thebandwidth), and a resampler reducesthe sample rate in proportion to thebandwidth reduction. Standardsingle-channel demodulation isdescribed by-jθkny (n,k) =[x (n)e ]×h(n) (1)N-1y (n,k) = Σx (n-r)e h(r) (2)r=0-jθk (n-r)where x (n) is the carrier-centeredinput signal, θ k is carrier angularfrequency for kth channel, h(n) is thebaseband filter, and y (n,k) is the outputbaseband signal for k th channel.According to the EquivalencyTheorem [3], down-conversionfollowed by baseband filtering can bereordered so that filtering at the carrieroccurs first, followed by04ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P8

Polyphase Filter Banks for Embedded Sample Rate Changes in Digital Radio Front-EndsSpecial TopicMehmood Awan, Yannick Le Moullec, Peter Koch, and Fred Harrisdown-conversion. This is the oppositeof the traditional channelizationprocess. Fig. 2(b) shows this reorderedoperation, which is also described by-jrθky (n,k) = e Σ x (n-r)h(r) e . (3)To reduce the work involved indown-converting and then discardingthe samples during resampling, theheterodyne and down-sampler arereordered, and only retained samplesare down-converted, as shown inFig. 2(c). In this case, the frequency ofthe heterodyne at the reduced samplerate is Mθ k rad/sample rather than theoriginal frequency of θ k. If the centerfrequency, θ k, is a multiple of the outputsample rate 2π/M, that is, k(2π/M), thenthe center frequency is aliased to 0 byM-to-1 resampling. Under thiscondition, the down-sampledheterodyne defaults to unity and can bediscarded, as shown in Fig. 2(d).For the computed output for eachinput, M -1 of these computed outputsamples are discarded by thedown-sampler. To reduce thisworkload, the resampling and thefiltering operations are reordered sothat one output is computed for every Minput sample. This is achieved byapplying the Noble identity [3], whichdescribes how a filter processing everyMth input sample followed by an outputM-to-1 down-sampler is equivalent toan input M-to-1 down-samplerfollowed by a filter processing everyinput sample. The originalup-converted filter is partitioned to Msubfilters that operate at the reducedoutput rate rather than the original inputrate. The mathematical expressions in(4) describe the mapping of the filter’sZ-transforms at the input rate to a sumof Z-transforms at the output rate:N-1-jθk n N-1r=0MG(Z ) = Σ h(n)e Z -nn=0N -1M-1 MM-1= Σ Z -r Me Σ h(r+nM)Z -nM .r=02πj kn= Σ Σ h(r+nM) e Z -(r+nM )r=0 n=0Nj2π -1kr Mn=0j2πk (r+nM)MThe phase rotators in each subfilterare constant for that subfilter. Fig. 2(e)shows the block diagram for (4). Theoutput resampler is pulled to the inputside of each filter stage by applying thex(n )x(n )-jθkneDigitalLowpass FilterH(Z)DigitalBandpass Filter-jθkH(Z e )H0(Z)FDM: frequency division multiplexingNoble identity. The input delayelements Z -r and the resampler at eachstage are replaced by a rotary switchcalled a commutator.In the final step of forming thepolyphase filter bank, the sum formedby the phase rotators is one output portof a discrete Fourier transform (DFT).The DFT can be implemented as a fastFourier transform (FFT) to extract timesamples of each narrowband processlocated at multiples of the outputsample rate (that has been aliased tobaseband by the resampler) [5]. This isshown in Fig. 2(f) and given byThe relationship between thesampling frequency, channel spacing,and number of channels for thepolyphase channelizer iswhere f s is the input samplingfrequency, N is number of channels(a)y (n,k)(c)PolyphasePartitionH2(Z)MMj0k(2π/Μ)ej1k(2π/Μ)e-jMθkney(nM,k)FFT: fast Fourier transform▲Figure 2. Transformation of a standard single-channel channelizer into a polyphasechannelizer with M channels: (a) heterodyning, filtering, and down-sampling for a standardchannelizer, (b) reordered channelizer using Equivalency Theorem, (c) reordering of theresampler, (d) down-converter with center frequency at a multiple of output sample ratealiases to baseband by M -to-1 down-sampling, (e) single-channel polyphase channelizer,and (f) polyphase channelizer for M channels.x (n )H1(Z)j2k(2π/Μ)y (nM,k )x(n ) ex (n )…M-1…HM -1(Z)(e)…j (M-1)k (2π/Μ)e(4) y (nM,k) = Σ y r (nM)e M . (5)r=0y (n,k)2πj kry (nM,k )f s= N·Δf (6)(FDM)DigitalBandpass Filterx (n )-jθkH(Z e )DigitalBandpass Filter-j(2π/Μ)kH(Z e )…-jθkney (n,k )(b)(d)PolyphasePartitionH0(Z)H1(Z)H2(Z)…HM -1(Z)(f)(FFT size), which is the same as M here,and Δf is the inter-channel spacing [3].The polyphase filter channelizer usesthe input M-to-1 resampling to aliasthe spectral terms residing at multiplesof the output sample rate to baseband.This means that for a standardpolyphase channelizer processing Minput samples at a time, the outputsample rate is the same as the channelspacing. When operating in this mode,the system is a maximally decimatedfilter bank. We experimented withpolyphase filter banks using embeddedresampling and here presentunder-decimated, over-decimated,and combined up- anddown-sampling (for single and multiplecommutator stride lengths) modes.3 Non-MaximallyDecimated Filter BankWe have briefly presented thepolyphase filter bank channelizer inMMM -PointFFTy (nM,k )y (nM,k )y (nM,0)y (nM,1)y (nM,2)…y (nM,M-1)December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 05D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P9

Special TopicPolyphase Filter Banks for Embedded Sample Rate Changes in Digital Radio Front-EndsMehmood Awan, Yannick Le Moullec, Peter Koch, and Fred Harris-15 -12 -6 0 6 12 15 f (MHz)-f s /2 M-f s /2f s =30 MHz▲Figure 3. A 30 MHz FDM signal with 5 channels separated by 6 MHz center frequencies.Each channel has a 2.5 MHz symbol rate shaped by a square root Nyquist filter with20% excess bandwidth.which the output sample rate is thesame as the channel spacing.However, in practice, an output samplerate that is different from the channelspacing is often required. To uncouplethe output sample rate from the channelspacing, a straightforward approach isto resample each channel with P/Qresamplers [5]. By changing the valuesof P and Q, the required sample ratecan be obtained. An alternative is toembed the resampling process in (i) thepolyphase commutator, that is, in theinteraction between input data registersand the polyphase coefficients, and in(ii) the interaction between thepolyphase outputs and the FFT input.This alternative only requires a statemachine to schedule the interactions,and there is no computational cost.Two schemes [5] for theseinteractions are (i) serpentine shiftingthe input data that interacts with a fixedset of coefficients and circular bufferingthe filtered data prior to FFT, and (ii)sliding the cyclic data-load thatinteracts with cyclic-shifted coefficientmemory. In the serpentine shift andcircular buffering scheme, an input dataset (not equal to M ) is always fed to thesame registers, and the polyphasesubfilter coefficients are fixed. Let usconsider a single-tapped delay linewhere all the data is moved further tothe right before the next input data setis loaded. The data is moved by anaddress equal in length to the nextinput data set. By folding thisone-dimensional tapped delay line intothe two-dimensional memory of thepolyphase filter, the data move is aserpentine shift between the columns.Because this non-equal M input dataf s2.5 MHzset is loaded, the data time-originmoves with respect to the FFTtime-origin. To keep these two originsaligned, the computed output of thepolyphase filter is circular-shifted bythe residue address of the datatime-origin mod M before the FFT isperformed. In the sliding cyclicdata-load and cyclic shift of thecoefficient memory scheme, the dataregisters are fixed instead of beingcycled, and the coefficient sets arerotated. The input data is fed as slidingcyclic load by the input commutator to afixed set of registers, and the subfiltercoefficients are cyclic-shifted by thesame residue address of the datatime-origin mod M before FFT isperformed. Taking individual subfiltersinto account, the first scheme seems torequire more read and write operationsto synthesize the serpentine data shift.However, this shift is, rather, achievedby circular wrapping of block memory(an address control task). In the secondscheme, only the loading subfilter getsa data shift.To demonstrate the embeddedresampling, we here describe theexample shown in Fig. 3. A system has5 channels separated by 6 MHz centerfrequencies. Each channel has a2.5 MHz symbol rate shaped by asquare root Nyquist filter (with 20%excess bandwidth) to form a 30 MHzFDM channel. To satisfy the Nyquistcriteria at the output sample rate, theoutput sample rate must be greaterthan the occupied channel bandwidth.The occupied channel bandwidth of3 MHz (symbol rate plus excessbandwidth) is selected to be smallerthan the channel bandwidth of 6 MHz toallow down-sampling by large factorswithin the channelizer. Thedown-sample channelizer uses a30-tap prototype low-pass filter witharound 60 dB side-lobe attenuationthat is partitioned into a 5-pathpolyphase filter with 6-tap subfilters.Both the data bank and filter coefficientbank are two-dimensional memories of5 rows and 6 columns, and each rowcorresponds to a subfilter. According to(6), the output sample rate for themaximally decimated system becomes6 MHz, which is the same as thechannel spacing. Four other resamplingfactors are also introduced, and two ofthese have an embedded up-samplingfactor of two. These five resamplingfactors are 5, 3, 6, 5/2, and 15/2,delivering output sample rates of 6, 10,5, 12 and 4 MHz, respectively. Thesecorrespond to maximally decimated,under-decimated, over-decimated,and combined up-and down-samplecases. Each of these cases will bepresented for the sliding cyclic dataload and cyclic shift of the coefficientmemory scheme.Case 1: Maximally-Decimated ModeIn a maximally decimated system,data is loaded in stride lengths of 5mod 5, and a computed output samplehas a 5-to-1 down-sampling. Fig. 4(a)shows the data loading process for thetwo outputs. The subfilter’s dataregister and coefficients are denoted Rand C, respectively. In all the dataloads, data loading starts from subfilterR4 up to R0, and the loaded subfilter’stapped delay line is pushed one tap tothe right before a new data element isloaded. For every computed output, allthe subfilters are fed with input data.Because there is no residue(non-loaded subfilter), there is no offsetbetween the data time-origin and theFFT time-origin. The coefficients of thesubfilters are therefore fixed from C0 toC4. There is only one state machinewhere the 5-point data-loading of theregister bank always performs an innerproduct with the fixed set of subfiltercoefficients. Table 1 shows the registerloading sequence and correspondingsubfilter coefficients.Case 2: Under-Decimated ModeIn a non-maximally decimated06ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P10

Polyphase Filter Banks for Embedded Sample Rate Changes in Digital Radio Front-EndsSpecial TopicMehmood Awan, Yannick Le Moullec, Peter Koch, and Fred Harris56▲Figure 4. Data loading processes for a 5-path polyphase filter performing (a) 5-to-1 down-sampling(case 1), (b) 3-to-1 down-sampling (case 2), (c) 6-to-1 down-sampling (case 3), (d) 5/2 down-sampling(case 4), and (e) 15/2 down-sampling (case 5).▼Table 1. Register loading sequence andcorresponding subfilters coefficients for a5-path maximally decimated systemState0Loading SequenceR4 R3 R2 R1 R0Filter CoefficientsC0 C1 C2 C3 C4▼Table 2. Register loading sequence andthe corresponding subfilter coefficients for3-to-1 sample rate change in a 5-pathpolyphase filterState01234R0R1R2R3R4R0R1R2R3R4n+9n+8n+7n+6n+5n+10n+9n+8n+7n+112 nd Data Loadn+4n+3n+2n+1n+02 nd Data Loadn+5n+4n+3n+2n+6Loading SequenceR2 R1 R0R4 R3 R2R1 R0 R4R3 R2 R1R0 R4 R3n+0n+1Filter CoefficientsC0 C1 C2 C3 C4C3 C4 C0 C1 C2C1 C2 C3 C4 C0C4 C0 C1 C2 C3C2 C3 C4 C0 C1(under-decimated) system, data isloaded in stride lengths of 3 mod 5, anda computed output sample has a3-to-1 down-sampling within a5-stage polyphase filter. The leastcommon multiple (LCM) of 3 and 5 is15, which means that the state enginecycles in 15 inputs, and because3-point data is delivered at a time,there must be 5 distinct states in thestate machine. Fig. 4(b) shows thedata-loading processes for the first two56(a)(c)7R0R1R2R3R4R0R1R2R3R4R0R1R2R3R4n+4n+3n+2n+1n+0n+5n+4n+3n+2n+11 st Data Load1 st Data Loadn+12n+14n+11n+13n+103R0R1R2R3R4n+2n+1n+5n+4n+3n+0 R0 n+22 R1R2R3R4n+4n+1n+3n+02 nd Data Loadn+7n+9n+6n+8n+5n+2n+4n+1n+3n+08(e)R0R1R2R3R4n+7n+4n+6n+3n+52 nd Data Loadn+02 nd Data Load1 st Data Loadn+2n+1n+0R0R1R2R3R4states.In the first state, data loading startsfrom subfilter R2 up to R0; and in thesecond state, data loading starts fromR4 to R2 and so on for the five distinctstates. Consequently, there is a residueof 2 for each data-loading operation.To align the data time-origin with theFFT time-origin, the subfiltercoefficients are cyclic-shifted by theresidue address of the data time-originmod 5. The time-origin that is beingcyclically shifted is also periodic in theLCM of 3 and 5. Thus, the cyclic shift ofthe polyphase subfilter coefficients hasthe same period as the data registerload and is controlled by the same statemachine. The data-loading sequenceis always to the next 3 registers thathave indexing of mod 5, which meansthat the next register to accept datawhen moving from state 0 to state 1 is(R0)-1, which is actually R4. Similarly,the filter coefficients assigned toperform the inner products with theregisters are always offset 3 mod 5relative to the previous filter set. Table 2shows the state machine forregister-loading and coefficients ofeach corresponding subfilter forperforming 3-to-1 down-sampling in a5-stage polyphase filter.33(b)n+2n+1n+0R0 n+2R1R2 n+1R3R4 n+0(d)1 st Data Load1 st Data LoadCase 3: Over-DecimatedModeIn a over-decimatedsystem, data is loaded instride lengths of 6 mod 5,and a computed outputsample has a 6-to-1down-sampling within a5-stage polyphase filter.The LCM of 6 and 5 is 30,which means that the stateengine cycles in 30 inputs,and because 6-point datais delivered at a time, theremust be 5 distinct states inthe state machine. Fig. 4(c)shows the data-loadingprocesses for the first twostates.In the first data load,loading starts from subfiltersR0, R4 up to R0, and thesecond load starts from R4up to R0 and R4 again andso on for the 5 distinctstates. Consequently, there is a residueof 4 for each data-loading operation.To align the data time-origin with theFFT time-origin, the subfiltercoefficients are cyclic-shifted by theresidue address of the data time-originmod 5. Table 3 shows the statemachine for register-loading and thecorresponding coefficients forperforming 6-to-1 down-sampling in a5-stage polyphase filter.Case 4: Combined Up- andDown-Sampling Mode (Single Stride)In the previous three cases,down-sampling was performed bydifferent factors. The polyphase filter isalso capable of embedding theup-sampling factor with thedown-sampling so that the sample ratechange is rational. In this case, we▼Table 3. Register loading sequence and thecorresponding subfilter coefficients for3-to-1 sample rate change in a 5-pathpolyphase filterState01234Loading SequenceR0 R4 R3 R2 R1 R0R4 R3 R2 R1 R0 R4R3 R2 R1 R0 R4 R3R2 R1 R0 R4 R3 R2R1 R0 R4 R3 R2 R1Filter CoefficientsC0 C1 C2 C3 C4C1 C2 C3 C4 C0C2 C3 C4 C0 C1C3 C4 C0 C1 C2C4 C0 C1 C2 C3December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 07D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P11

Polyphase Filter Banks for Embedded Sample Rate Changes in Digital Radio Front-EndsSpecial TopicMehmood Awan, Yannick Le Moullec, Peter Koch, and Fred HarrisdB100-10-20-30-40-50-60-70-80-15 -12 -9 -6 -3 0 3 6 9 12 15100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70-80-80▲Figure 5. (a) Input signal spectrum with 5 channels of 3 MHz bandwidth and 6 MHz channelspacing at 30 MHz sample rate. The spectra of the 5 output channels are (b) 6 MHz for 5-to-1down-sampling, (c)10 MHz for 3-to-1 down-sampling, (d) 5 MHz for 6-to-1 down-sampling,(e) 12 MHz for 5/2 resampling, and (f) 4 MHz for 15/2 resampling.Table 5. The spectral locations of thechannels are reordered as a result ofprocessing the up-sampled data in thepolyphase filter [3], [5]. The 5-pointFFT processes the polyphase dataoutput frequencies in the order[0, 2, 4, 1, 3], which is seen to beindexing stride of 2 mod 5. These arereordered back to their natural order.Figs. 5(e) and (f) show the spectra ofthe 5 output channels at 12 MHz and 4MHz, which correspond to 5/2 and 15/2sample rate changes, respectively.The simulations show that embeddedsample rate changes can besuccessfully implemented in apolyphase channelizer. All the outputchannels have 60 dB of spectralside-lobe attenuation, selected by theprototype low-pass filter. Theprocessing engines used in all the 5cases are identical except that eachhas different state machines, registerloading schemes, and subfiltercoefficient sets.5 ConclusionIn this paper, we have shown theversatility of a polyphase engine that100-10-20-30-40-50-60-70-80-80-80-80-80-5 0 5 -5 0 5 -5 0 5 -5 0 5 -5 0 5100-10-20-30-40-50-60-70-80-80-80-80-80-5 0 5 -5 0 5 -5 0 5 -5 0 5 -5 0 5(a)(c)(e)100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70-80100-10-20-30-40-50-60-70-80100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70-80100-10-20-30-40-50-60-70-80-80-80-80-2 2 0 -2 2 0 -2 2 0 -2 200 -2 2100-10-20-30-40-50-60-70-80100-10-20-30-40-50-60-70performs embedded resampling that isuncoupled from frequency selectionand bandwidth control. Five embeddedresampling modes in polyphase filterbanks have been presented, namely,maximally decimated,under-decimated, over-decimated,and combined up- anddown-sampling. These correspond tosingle, short, long, and multiplecommutator stride lengths. For variousapplications, these modes can be usedfor any required rational sampling-ratechange in an SDR front-end using apolyphase channelizer. The suggestedmodes are highly useful for designingflexible and resource-optimalarchitectures for advanced softwareradios. In a subsequent paper“Hardware Architecture Analysis ofPolyphase Filter Banks PerformingEmbedded Resampling for SoftwareDefined Radio Front-Ends”[6], weanalyze FPGA based hardwarearchitecture of these resamplingengines in terms of area, time, andpower tradeoffs.References[1] A. M. Badda and M. Donati,“The software definedradio technique applied to the RF front-end for(b)-80-80-80-80-2 2 0 -2 2 0 -2 2 0 -2 200 -2 2100-10-20-30-40-50-60-70-80100-10-20-30-40-50-60-70(d)-80-80-80-80-800 -2 2 -2 2 0 -2 2 0 -2 2 00 -2 2(f)100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70100-10-20-30-40-50-60-70cellular mobile systems,”in Software RadioTechnologies and Services, E. Del Re, Ed., Berlin,Germany: Springer-Verlag, 2001.[2] f. harris, C. Dick, M. Rice,“Digital receivers andtransmitters using polyphase filter banks forwireless communications,”in IEEE Trans. Microw.Theory Tech., vol. 51, no. 4, pp 1395-1412, 2003.[3] f. harris, Multirate Signal Processing forCommunication Systems. New York: Prentice Hall,2006.[4] T. Hentschel, M. Henker, G. Fettweis,“The digitalfront-end of software radio Terminals,”IEEEPersonal Commun., vol. 6, no.4, pp 40-46, Aug.1999.[5] f. harris, C. Dick,“Performing simultaneous arbitraryspectral translation and sample rate change inpolyphase interpolating or decimating filters intransmitters and receivers,”in Proc. SoftwareDefined Radio Tech. Conf. and Product Expo, SanDiego, CA, Nov 2002.[6] M. Awan, Y. Le Moullec, P. Koch, and f. harris,“Hardware architecture analysis of polyphase filterbanks performing embedded resampling forsoftware-defined radio front-ends,”to appear inSpecial Issue on Digital Front-End and SoftwareRadio Frequency, ZTE Communications, March,2012.BiographiesMehmood Awan (mura@es.aau.dk) received hisMSc degree in electronic engineering withspecialization in applied signal processing andimplementation from Aalborg University in 2007. Hewas a research assistant for one year and startedhis Ph.D. in resource-optimal SDR front-ends in2008. His research interests include multirate signalprocessing, SDR, hardware architectures, andembedded systems.Yannick Le Moullec (ylm@es.aau.dk) receivedhis Ph.D. degree in electrical engineering fromUniversité de Bretagne Sud, Lorient, France, in2003. From 2003 to 2005, he was a post-doctoralfellow at the Center for Embedded SoftwareSystems, Aalborg University, Denmark. From 2005to 2008, he was an assistant professor at theDepartment of Electronic Systems, AalborgUniversity, where he is now an associate professor.His research interests include methods and tools forHW/SW co-design, embedded systems, andreconfigurable computing.Peter Koch (pk@es.aau.dk) received his M.Sc.and Ph.D. degrees in Electrical Engineering fromAalborg University, Denmark, in 1989 and 1996.Since 1997, he has been an associate professor atthe Department of Electronic Systems, AalborgUniversity, working in the interdisciplinary fieldbetween DSP and resource-optimal real-timearchitectures. From 2006 to 2010, he headed theCenter for Software Defined Radio, AalborgUniversity. His research interests includeoptimization between DSP algorithms andarchitectures, and low-energy HW/SW design.Fred Harris (fred.harris@sdsu.edu) holds theSignal Processing Chair of the CommunicationSystems and Signal Processing Institute at SanDiego State University where he teaches DSP andcommunication systems. He holds 20 patents fordigital receivers, and he lectures around the worldon DSP applications. He is an adjunct of thePrinceton IDA-CCR Center for CommunicationsResearch and is the author of“Multirate SignalProcessing for Communication systems.”December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 09D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P13

Special TopicDesign of Software-Defined Down-Conversion and Up-Conversion: An OverviewYue Zhang, Li-Ke Huang, Carsten Maple, and Qing XuanDesign of Software-DefinedDown-Conversion andUp-Conversion: An OverviewYue Zhang 1 , Li-Ke Huang 2 , Carsten Maple 1 , and Qing Xuan 3(1. Department of Computer Science and Technology, University of Bedfordshire, UK;2. Aeroflex International Ltd, Stevenage, UK;3. Sino-European-Link Ltd, UK)Abstract: In recent years, much attention has been paid to software-defined radio (SDR) technologies for multimode wireless systems.SDR can be defined as a radio communication system that uses software to modulate and demodulate radio signals. This articledescribes concepts, theory, and design principles for SDR down-conversion and up-conversion. Design issues in SDRdown-conversion are discussed, and two different architectures, super-heterodyne and direct-conversion, are proposed. Designissues in SDR up-conversion are also discussed, and trade-offs in the design of filters, mixers, NCO, DAC, and signal processing arehighlighted.Keywords: SDR down-conversion; up-conversion; direct-conversion; super-heterodyne conversion1 IntroductionSoftware-defined radio (SDR) willplay a key role in future radioconfigurations because of theemergence of new wirelesstechnologies and their integration intofourth generation standards such asLTE-Advanced. SDR is a fundamentalpart of many radio systems, whichinclude up-conversion of the discretebaseband signal into a high-resolutionradio signal at the transmitter anddown-conversion of thehigh-resolution radio signal back intobaseband at the receiver [1]. Fig. 1shows an SDR receiver insuper-heterodyne architecture — adual-stage conversion architecture inwhich a radio frequency (RF) signal isdown-converted to intermediatefrequency (IF) in the first stage and thenconverted from IF to baseband in thesecond stage. The hardware in the RF,IF, and baseband sections arecontrollable and reconfigurable usingsoftware (Fig. 1).In Fig. 1, the RF front-end (analogfront-end) is the only analog sectionand includes mixer, low-noise amplifier(LNA), power amplifier (PA), RFcombiner, bandpass filter(anti-aliasing), and antenna. The RFfront-end is responsible for transmittingand receiving RF signals andconverting RF signals to IF. Someadvanced RF front-ends allow a certaindegree of control, for example,frequency tuning, through software.The other parts of the SDRarchitecture are digital processingcomponents. In the IF section, samplingand separation of IF carriers, andup/down frequency conversion tobaseband is performed by digital IFprocessing normally implemented infield programmable gate array (FPGA).Take the receiving path for example.The IF signal is digitized by theanalog-to-digital converter (ADC) andthen converted to baseband by digitalIF processing. Digital IF processingincludes digital down-conversion(DDC) operations such as digitalsynthesizing, digital numerical controloscillator (NCO), digital mixing, I/Qdemodulation, and multirate decimationfiltering. In digital IF processing, digitalup-conversion (DUC) is performed inthe reverse order of DDC. Digitalsystems such as FPGAs are normallyused for IF-to-baseband conversionbecause they are able to handle tightreal-time constraints imposed byhigh-speed sampling and digitalconversion. In the back-end of the SDRarchitecture, the baseband processmainly performs digital communicationsfunctions such as symbol timingrecovery, equalization, modulation, andchannel coding. These functions areusually carried out using digital signalprocessors (DSPs) with slightly relaxedreal-time constraints.SDR architecture tends to usegeneral-purpose digital systems.FPGAs and DSPs in SDR allow thetransmitted signal to be generated and10ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P14

Design of Software-Defined Down-Conversion and Up-Conversion: An OverviewSpecial TopicYue Zhang, Li-Ke Huang, Carsten Maple, and Qing XuanADC: analog-to-digital converterBPF: bandpass filterIF: intermediate frequencythe received radio signal (at thereceiver) to be tuned and detecteddigitally by software. The conventionalmethod involves an analog signalpassing through individual hardwarecomponents, each of which perform aspecific function. SDR systems offergreatly extended programmability,reconfigurability, and definability. SDRis intended to be a radio system that isflexible, versatile, and multistandard.This system is therefore able toaccommodate updates of new radiofunctions.This article highlights severalpossible implementations of receiversand transmitters. Several architecturesfor SDR receiver front-ends arereviewed alongside severalarchitectures for transmitter front-ends.In the conclusion, we summarize thiswork and identify solutions with greatestpotential from our point of view.2 Software-Defined RadioReceiver ArchitecturesIn this section, we review severalfront-end architectures for SDRreceivers.2.1 Super-Heterodyne ReceiverFig. 1 shows a super-heterodynereceiver. Super-heterodyne is aLNA: low-noise amplifierLO: local oscillatorLPF: low-pass filter▲Figure 1. Super-heterodyne receiver architecture.▲Figure 2. Sample rate conversion modulein DDC with integral rate conversion.RFBPF LNA LOf1ImageRemovalandAnti-AliasingFilteringIFBPFDown-SamplingIFVGAADC90°QNCONCO: numerical control oscillatorRF: radio frequencyVGA: viable gain amplifierwell-known receiver architecture inwhich an RF signal received from theantenna is translated to basebandusing a down-conversion mixer,bandpass filter, and amplifier [2]. First,the signal is filtered by a bandpass filter(BPF) to obtain the desired channelsignal. After BPF, the signal level isboosted by an LNA. The signal isconverted to IF because of thedown-conversion mixer, local oscillator(LO), second-stage BPF, andvariable-gain amplifier. Then, it issampled by an ADC and digitallyprocessed by the DDC. At the DDC, I/Qcomponents are extracted anddemodulated to baseband. Because ofthe sample rate difference at IF andbaseband, multirate filters are requiredto perform sample-rate conversion.If the conversion rate is an integer,the sample-rate conversion module inthe DDC performs image-removalfiltering, anti-aliasing filtering, anddown-sampling. This module can bejointly designed by using multistage orsingle-stage filtering. Such a designprovides an equivalent impulseresponse for performingimage-removal filtering andanti-aliasing filtering. The single-stageimplementation is shown in Fig. 2.If the conversion rate is a fraction, theDDC module performs image-removalfiltering, up-sampling,interpolation/anti-aliasing filtering, anddown-sampling [3]. The filteringFigure 3. ▶Sample rate conversionmodule in DDC withfractional rateconversion.ILPFLPFRateConversionRateConversionImageRemovalFilteringoperation cannot be jointly designed inthis case. Fig. 3 shows thesingle-stage implementation of eachfiltering operation.Currently, super-heterodynearchitecture is used for higher-RFfrequency designs such as LTE andIEEE 802.11ac. The main issue forsuper-heterodyne receivers is thehigh-quality image and adjacent filterdesign. Fig. 4 (top) shows the processof converting an RF signal into an IFsignal. The received signal contains thedesired signal at f LO + f IF, a secondsignal close to the image frequency atf LO - f IF, and interference in an adjacentchannel [4]. The image signal isremoved by an image-reject BPFbefore conversion. At this stage,adjacent interference and imageinterference cannot be completelyremoved because of the limitation of theimage-reject BPF. The interferencelevel is attenuated by the image-rejectBPF. Fig. 5 shows that the remainingsignal and interference at the imagefrequency falls into the same band asthe desired signal whendown-converted into IF. Thisdown-conversion is performed by themixer and the first LO. However, theimage signal, while strongly attenuatedby the image-reject BPF, is still presentin the signal band at IF. Thechannel-select BPF cannot remove thisinterference. The higher the IF that ischosen, the better the image that canbe rejected by a filter because theseparation between the desired signaland image is greater. A considerabledrawback of using a high IF is thatsignals in adjacent channels are notsufficiently attenuated by theimage-reject filter and a subsequent IFfilter becomes necessary. The lower theIF, the steeper the absolute slope of thefilter. As a compromise, two differentIFs can be used, which creates adual-IF topology. A further drawback ofthe super-heterodyne receiver is itsUp-SamplingInterpolationandAnti-AliasingFilteringDown-SamplingDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 11D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P15

Special TopicDesign of Software-Defined Down-Conversion and Up-Conversion: An OverviewYue Zhang, Li-Ke Huang, Carsten Maple, and Qing XuanImage InterferencefLO-fI FImageInterferenceInterference(dB Scale)fLOMixedwithLO f1(dB Scale)fLO+fI FSignalImage Reject BPFFreqency (Hz)Channel Select BPFImage Interferenceresolve this problem, differential circuitscan be used for LNA and mixer, and thedesign of these circuits should beoptimized for linearity. Leakage of theLO signal from the mixer and LNAcreates in-band interference for otherreceivers. To resolve this problem,balanced mixers and LNAs with highreverse isolation are used in the design.In I/Q imbalance, the LO outputs are notexactly 90 degrees out of phasebetween the I and Q branches. Thisphase shift causes I/Q mismatch in thebaseband. However, this distortion isfrequency-independent, and digitalcalibration can remove it.BPF: bandpass filterfIFSignalLO: local oscillatorcomplexity, which leads to increasedchip size, more complex circuit design,and increased power consumption inthe direct-conversion receiver.2.2 Direct-Conversion ReceiverAnother approach for front-endarchitecture is the direct-conversionreceiver as shown in Fig. 5. Adirect-conversion receiver is asimplified version of asuper-heterodyne receiver [5], [6].The received RF signal is selected bya BPF and amplified by an LNA. TheBPF removes out-of-band interferenceand noise. After this, the selected RFsignal is directly down-converted toDC by a mixer and converted to digitalby ADC. Only low-pass filters areneeded in the baseband circuits toperform channel filtering. Comparedwith a super-heterodyne receiver, adirect-conversion receiver has feweranalog components. The absence of ahigh-Q filter is a significant advantageof a direct-conversion receiver, as isthe completely integrated CMOSimplementation of the front-end withoutthe signal quality being impaired.Therefore, the direct-conversionreceiver can make use of the high levelof integration for multiband receivers.Issues for the direct-conversionreceiver include DC offsets,even-order distortion, carrier leakage,Freqency (Hz)◀Figure 4.Spectrum conversion in asuper-heterodyne receiver.and I/Q imbalance [7], and eachcomponent has to be carefullydesigned and tuned. DC offset isgenerated by self-mixing a strongsignal at the mixer stage. The strongsignal is normally from interference orthe LO signal. The down-conversionstage involves down-convertingdirectly to zero frequency. A DC offsetis amplified in baseband together withthe received and down-convertedsignal. If the offset has higher amplitudethan the desired signal, the offsetdominates and saturates the followingLNA [8]. In the even-order distortion,spurious signals are created at lowfrequencies because of thedirect-conversion architecture.Because of RF to IF leakage in themixer, products of even-orderdistortion in the LNA and mixer candamage the baseband signal. ToBPFLNAADC: analog-to-digital converterBPF: bandpass filter▲Figure 5. Direct-conversion receiver architecture.I90°QLO1LPFLPFLNALNA3 Software-Defined RadioTransmitter Architectures3.1 Super-Heterodyne TransmitterA super-heterodyne transmitter usesthe reverse procedure of thesuper-heterodyne receiver shown inFig. 6. The signal is created in thedigital domain and then converted intoanalog using simple digital-to-analogconverters (DACs).The complex input I/Q basebandsignal is sampled at a relatively lowrate, typically the digital modulationsymbol rate. The baseband signal isfiltered and converted to a highersampling rate before being convertedinto analog. Starting at the interpolationfilter, the complex input signals passthrough three stages of filtering, each ofwhich changes the sampling rate andperforms associated low-passinterpolation filtering [9]-[11].The three filtering stages are shownin Fig. 6. First, a compensation finiteimpulse response filter, CFIR G(z),ADCADCLNA: low-noise amplifierLO: local oscillatorLPFLPFRateConversionRateConversionLPF: low-pass filter12ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P16

Design of Software-Defined Down-Conversion and Up-Conversion: An OverviewSpecial TopicYue Zhang, Li-Ke Huang, Carsten Maple, and Qing XuanPolyphaseIntepolateby D1:2:8CFIR G(z)PolyphaseIntepolateby D2:2:8PFIR H(z)CFIR: compensation finiteimpulse response filterCIC: cascaded integrator combDPAP -StageCICUpsamplebyR:4-16383▲Figure 6. Super-heterodyne conversion transmitter architecture.BPFprovides a sampling rate that isincreased (in step 2) from 1 to 8. Thefilter also performs Nyquist pulseshaping for the transmitter. Second, aprogrammable FIR filter, PFIR H(z),increases the sampling rate by 2 andcompensates for the passbanddistortion created by the third-stagecascaded integrator comb (CIC) filter.G(z) and H(z) are implemented withefficient multiplier-accumulator (MAC)blocks. Third, the P-stage CICinterpolation filter increases thesampling rate from 4 to 1448. Thedesign parameters of the CIC allow forthe synthesis of a circuit that has fixedsampling-rate increases or one wheresampling-rate increases areprogrammable in real-time. Fourth, theCIC coarse-gain control block isneeded to control the power distortioncaused by CIC filtering. Finally, thecomplex data stream from the filteringstages is converted into an analog I/Qsignal. In the analog domain (afterDACs), the signal is modulated at an IFand is amplified and filtered to eliminateharmonics generated duringmodulation. The signal is thenup-converted into RF by the secondLO. This RF signal is filtered by anotherBPF to remove the image signal and isboosted by an RF PA [12].A super-heterodyne transmitterarchitecture has two advantages. TheI/Q modulator works at IF band, whichmeans the circuit is much easier todesign than at RF band. Also, theLO2CICGoarseGainControlDAC: digital-to-analog converterLO: local oscillatorLPF: low-pass filterPFIR: programmable finite impulseresponse filteroverall signal level can be controlled atIF band. Therefore, it is much easier todesign a high-quality variable-gainamplifier at IF band. However, such anamplifier has the same high complexityas the super-heterodyne receiver.Super-heterodyne transmitterarchitecture is therefore mostly used forwireless measurement instruments orbackhaul wireless communication.3.2 Direct-Conversion TransmitterA direct-conversion transmitter has asimplified version of thesuper-heterodyne transmitterarchitecture [13] (Fig. 7). The design ofthe digital domain has exactly the samearchitecture as the super-heterodynetransmitter. The two DACs convert thedigital signal into analog. The followingIQBPFPAPolyphaseIntepolateby D1:2:8CFIR G(z)DACDACLPFCFIR: compensation finiteimpulse response filterCIC: cascaded integrator comb90°BPFDAC: digital-to-analog converterLO: local oscillatorLPF: low-pass filter▲Figure 7. Direct-conversion transmitter architecture.LO1PolyphaseIntepolateby D2:2:8PFIR H(z)∑P -StageCICUpsamplebyR:4-16383PALPF removes the Nyquist images andimproves the noise floor. Then, thecomplex analog I/Q signals are directlymodulated at RF band by thehigh-specification RF I/Q modulator[14]. A BPF centered at the desiredoutput frequency then removes theimage signals. A final PA can boost thedesired signal to a certain power level.The most significant advantages ofthe direct-conversion transmitter are itsversatility, flexibility, spectral efficiency,and low complexity. These make thetransmitter simpler than thesuper-heterodyne transmitter. Smallchip and circuit size and low powerconsumption can be achieved with adirect-conversion transmitterarchitecture.Direct-conversion transmitterarchitecture is mostly used inmultimode communication systems. Byduplicating the hardware for eachchannel and each standard, amultimode radio DUC can beimplemented.Although direct-conversionarchitecture reduces the complexity ofthe transmitter design, it has someserious drawbacks, including carrierleakage and phase gain mismatch.Gain may need to be controlled at RF.The architecture also requires a PA withgood linearity [15].4 ConclusionIn this article, we have reviewedreceivers and transmitters used in SDRCICGoarseGainControlDACDACLPFLO190°∑PFIR: programmable finite impulseresponse filterDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 13D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P17

Special TopicDesign of Software-Defined Down-Conversion and Up-Conversion: An OverviewYue Zhang, Li-Ke Huang, Carsten Maple, and Qing Xuanfront-ends. Advantages anddisadvantages of each design havealso been analyzed. A well-designedarchitecture for multiband multimodereceivers and transmitters should allowhardware resources to be optimallyshared. The architecture should makeuse of software-controlled componentsand software-programmable devices.A direct-conversion approach wasused in the transmitter and receiverfront-end for implementation in mostcommercial wireless systems.Direct-conversion is the mostappropriate architecture because it islow cost, consumes little power, andhas a high-data-rate radio interface.However, system designers need topay more attention to I/Q imbalance,carrier leakage, and DC offset in adirect-conversion receiver andtransmitter system.References[1] Fa-Long Luo (Ed.), Digital Front-End in WirelessCommunications and Broadcasting. New York:Cambridge University Press, 2011.[2] V. Giannini, J. Craninckx, and A. Baschirotto,Baseband Analog Circuits for Software DefinedRadio. Dordrecht: The Netherlands:Springer-Verlag, 2008.[3] F. Harris, Multirate Signal Processing forCommunication Systems. New Jersey: Prentice-Hall, 2008.[4] Behzad Razavi. RF microelectronics. New Jersey:Prentice-Hall, 1998.[5] A. Abidi,“The path to the software-defined radioreceiver,”in IEEE J. Solid-State Circuits, vol. 42,no.5, pp. 954-966, May 2007.[6] V.Giannini, P. Nuzzo, C.Soens, K.Vengattaramane,M.Steyaert, J.Ryckaert, M.Goffioul, B.Debaillie, J.Van Driessche, J.Craninckx, and M.Ingels,“A 2 MM 2 0.1-to-5 GHz SDR receiver in 45 nmdigital CMOS, ”in IEEE J. Solid-State Circuits, vol.44, no. 12, Dec. 2009, pp. 3486-3498.[7] B. Razavi.“Design considerations fordirect-conversion receivers,”IEEE Trans. Circ.Syst. II: Analog and Digital Signal Processing, vol.44, no. 6, pp. 428-435.[8] R. Svitek and S. Raman,“DC offsets in directconversion receivers: Characterization andimplications,”IEEE Microw. Mag., vol 6, no.3,pp.76-86.[9] Fa-Long Luo (Ed.), Mobile MultimediaBroadcasting Standards: Technology and Practice.Dordrecht: The Netherlands: Springer, 2008.[10] Yue Zhang, J. Cosmas, K.K Loo and M. Bard,“Design and implementation of digital echocancellation on-channel repeater in DVB-T/Hnetworks,”in 57th Annu. IEEE Broadcast Symp.,Washington DC, 2007.[11] Yue Zhang, J Cosmas, M. Bard, and Y.H Song,“Diversity gain for DVB-H by usingBiographiestransmitter/receiver cyclic delay diversity,”in IEEETrans. Broadcasting, vol. 52, no. 4, Dec. 2006, pp.464-474.[12] F.H. Raab, P. Asbeck, S. Cripps, P.B. Kenington, Z.B. Popovic, N. Pothecary, J.F. Sevic and N.O.Sokal,“RF and microwave power amplifier andtransmitter technologies-Part 3,”High Frequency.Electronics., vol. 2, no.5, pp.34-46, Sept 2003.[13] A. Loke and F. Ali,“Direct conversion radio fordigital mobile phones—Design issues, status, andtrends,”IEEE Trans. Microw. Theory Tech., vol. 50,no.11, pp. 2422-2435, Nov. 2002.[14] G-T Gil,“Non-data-aided I/Q mismatch and DCoffset compensation for direct-conversionreceivers,”IEEE Trans. Signal Processing, vol. 56,no. 7, Jul. 2008, pp. 2662-p2668.[15] F. Ellinger. Radio Frequency Integrated Circuitsand Technologies. New York: Springer, 2007.Yue Zhang (yue_zhang@ieee.org) is a senior lecturer in the Department of Computer Science and Technology,University of Bedfordshire. He received his B.Eng. and M. Eng. degrees in 2001 and 2004 from Beijing Universityof Post and Telecommunications. In 2008, he received his Ph.D. degree from Brunel University — where he alsoworked as a research engineer for the EU IST FP6 project, PLUTO. He was responsible for transmitter and receiverdiversity design and measurement for DVB-T/H systems as well as RF/Digital/DSP design for on-channelrepeaters. After 2008, Dr. Zhang worked as a signal processing design engineer in Anritsu. He was responsible forRF/IF, digital and DSP design for wireless communication systems.Li-Ke Huang (like.huang@gmail.com) is a technical expert and algorithms group leader at Aeroflex UK. Hedevelops test and measurement technologies for wireless system engineering and specializes in transceiveralgorithm and architecture designs for the main wireless standards. He is responsible for new product featuresand new technology developments. He received his B.Sc. degree in electronic engineering from ShenzhenUniversity in 1998 and his Ph.D. degree in communication and signal processing from Imperial College London in2002.Carsten Maple (carsten.maple@beds.ac.uk) is pro-vice-chancellor of research and enterprise at the Universityof Bedfordshire. He received his B.Sc. degree in mathematics and his Ph.D. degree in numerical analysis from theUniversity of Leicester. He is a member of the IEEE, and a fellow of the FBCS and CITP. He heads the Center forResearch in Distributed Technologies (CREDIT) at the University of Bedfordshire, which has 40 staff and Ph.D.researchers. His research interests include information security, trust and authentication in distributed systems,graph theory, and optimization techniques. He has led a number of research projects in these areas with fundingtotaling of more than-million from UK EPSRC, EU, and the Department of Trade and Industry (DTI). He has beeneditor or guest editor for several international journals. He has been chairman and session chairman for a numberof international conferences. He has also been invited to present keynote speeches to various internationalconferences. Dr. Maple has published over 70 papers internationally and has been invited to talk on security,robotics, and applied computing on UK radio and television.Qing Xuan (jasx@talk2-1.com) received her Ph.D. degree in power and energy from the University of Bath in1995. She has more than 20 years’experience in the telecom, energy, finance, and aviation sectors. Since March2009, she has co-founded two companies. From January 2005 to March 2009 she was a business strategydirector for Huawei Technologies. Her responsibilities included working with the minister and regulator to developmarkets and introduce investor partners to clients. She also provided supply chain consultancy services tosecured-business clients, including M&A Technology Company. From 2000 to 2004, Dr. Qing Xuan was acommercial and technical architect at Vodafone Group. From 1998 to 2000, she was as design engineer forPanasonic. From 1995 to 1998, she was post-doctoral researcher in the Power Group at the University of Bath.From 1987 to 1991, she was deputy director of the Telecom Department at the Ministry of Power and Water, China.⬅ From P.02Mikko Valkama is full professor and department head of the Departmentof Communications Engineering at Tampere University of Technology(TUT), Finland. He has been involved in organizing conferences such asthe IEEE SPAWC'07 in Helsinki. He was awarded Best Ph.D. Thesis by theFinnish Academy of Science and Letters. His research interests includecommunications signal processing, estimation and detection techniques,signal processing algorithms for software defined flexible radios, andsignal processing for cognitive radio. He is also interested in digitaltransmission techniques, such as different variants of multicarriermodulation methods and OFDM, and radio resource management forad-hoc and mobile networks.Serioja Ovidiu Tatu received his M.Sc. and Ph.D. degrees in electricalengineering from the École Polytechnique, Montréal, in 2001 and 2004.From 2004 to 2005, he was a post-doctoral researcher at the InstitutNational de la Recherche Scientifique-Énergie Matériaux etTélécommunications, Montréal, and is now associate professor at thatinstitute. His current research interests include millimeter-wave circuitdesign, hardware and software radio receivers, radar, and sensor systems.Tomohisa Wada received his B.S. degree in electronic engineering fromOsaka University in 1983. He received his M.S.E.E. degree from StanfordUniversity in 1992 and his Ph.D. degree in electronic engineering fromOsaka University in 1994. He joined the ULSI Laboratory, MitsubishiElectric Corp. in 1983. Since 2001, he has been a professor at theDepartment of Information Engineering, University of the Ryukyus,Okinawa. In 2001, he was the founding member of Magna Design Net Inc.,an LSI design company for communication-related digital signalprocessing such as OFDM. Currently, he is chief scientist at Magna DesignNet Inc. and researches terrestrial video broadcasting, wireless LAN, andWiMAX.14ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P18

Practical Non-Uniform Channelization for Multistandard Base StationsSpecial TopicÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. FarrellPractical Non-UniformChannelization for MultistandardBase StationsÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrell(Callan Institute, Electronic Engineering Department, National University of Ireland, Maynooth (NUIM), Ireland)Abstract: A Multistandard software-defined radio base station must perform non-uniform channelization of multiplexed frequencybands. Non-uniform channelization accounts for a significant portion of the digital signal processing workload in the base stationreceiver and can be difficult to realize in a physical implementation. In non-uniform channelization methods based on generalized DFTfilter banks, large prototype filter orders are a significant issue for implementation. In this paper, a multistage filter design is applied totwo different non-uniform generalized DFT-based channelizers in order to reduce their filter orders. To evaluate the approach, a TETRAand TEDS base station is used. Experimental results show that the new multistage design reduces both the number of coefficients andoperations and leads to a more feasible design and practical physical implementation.Keywords: SDR; non-uniform; channelization; base station; TETRA1 IntroductionRadio spectrum is typicallyallocated using coarse-grainedfrequency division multiplexing.Different radio standards areallocated independent andnon-overlapping frequency bands thatare reserved exclusively for eachstandard. This approach simplifies theradio design because each standardoperates in isolation; however, availablespectrum is not used optimally becausereserved bands may be under-usedsome of the time. A more efficientalternative is to allow multistandardmultiplexing of the frequency band. Thefrequency band is instead multiplexedamong multiple radio standards [1].Channels within the frequency band aredynamically allocated to the radiostandards. If these channels do nothave uniform bandwidth or centerfrequencies, as is typical forheterogeneous radio standards, thendynamic allocation is challenging.In software-defined radio (SDR),analogue-to-digital anddigital-to-analogue conversion isperformed as close as possible to theantenna [2]. Most of the radiocomponents, now in the digital domain,are implemented on a reconfigurableplatform. Reconfigurability makes SDRparticularly suitable for working withmultistandard systems and forx 0(n )x 1(n )x K-1(n )y 0(n )y 1(n )y K-1(n )……WidebandDownlinkSynthesisWidebandUplinkChannelizerBaseStationUplinkupgrading to future standards.In an SDR base station (Fig. 1), thewideband uplink, which comprises allthe mobile station channels, is digitallyconverted immediately after the RFfront-end. Multirate digital signalprocessing (DSP) is then used to filterand shift to baseband all theindependent information channels. Onthe transmitter side, complementaryDownlink▲Figure 1. Asymmetric system formed by SDR multistandard base station and single-carrierbased mobile stations that use different mobile standards and uplink and downlink frequencybands.MS1MSK 1-1MS0MobileStandard 1MS1MSK J-1MS0MobileStandard JMS1MSK 2-1MS0MobileStandard 2December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 15D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P19

Special TopicPractical Non-Uniform Channelization for Multistandard Base StationsÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrellprocessing is carried out. Extraction ofthe uplink channels (channelization) is abigger challenge than downlinksynthesis because more stringentfiltering is required to avoid adjacentchannel interference.Different non-uniform channelizationtechniques have been described in[3]-[5]. In many cases, thesetechniques are based on combining oraltering efficient uniform channelizationmethods, such as uniform complexdiscrete Fourier transform (DFT)modulated transmultiplexers and treequadrature mirror filter banks (TQMFB)[6]. In these filter bank (FB) basedstructures, a synthesis bank multiplexesthe channel frequency at the transmitterside, and an analysis bank performs forequivalent channelization in thereceiver. Such a structure is symmetric;that is, amplitude and phase distortionin one half of the FB are suppressed inthe complementary half, allowingperfect reconstruction. This approachincludes multicarrier techniques suchas OFDM. In contrast, the system inFig. 1 has an asymmetric design; thatis, the uplink signal channelized by theanalysis bank is not created by acomplementary synthesis bank. Thesystem comprises the transmissions ofindependent mobile stations that areassumed to use single-carriertransceivers. The downlink is similarlyasymmetrical. This structure allows thebase station to be compatible withlegacy systems. A professional mobileradio (PMR) base station using theterrestrial trunked radio (TETRA)standard or its high-speed derivative,TETRA enhanced data service (TEDS)[7], is a good application of thisarchitecture.In an asymmetric system, subcarriersand aliasing overlap because ofdown-sampling in the analysis bank.The overlap must be minimized byhigh-order filters [4], and these highorders are a bottleneck in the designand physical implementation of thechannelizer filters. Multirate DSPtechniques, such as multistage filtering,can significantly reduce the filter ordersfor more practical implementation. Inthis paper, we compare twonon-uniform channelizationtechniques, parallel generalizedDFT-FB (GDFT-FB) and recombinedGDFT-FB, by applying multistagefiltering. We then evaluate thesemodified FBs in TETRA and TEDS basestations and compare their filter lengthsand computational loads with those in[4]. Finally, the effect of multistagedesign on the physical implementationof the channelizer is considered.Section 2 describes the non-uniformchannelization requirements of a basestation, specifically, TETRA and TEDSbase stations and their possibleupdates. Section 3 describes theclassic implementation of the twonon-uniform channelization methodsanalyzed in this paper. Designs formethods to be used in TETRA andTEDS base stations are given. Section 4describes a new multistage design forthe two non-uniform channelizationmethods, and improvements in designand implementation are also described.Section 5 concludes the paper.2 Frequency MultiplexedSpectrum forNext-Generation PMRThere are two main releases ofTETRA: TETRA voice and data (V&D)and TEDS. The former supports 25 kHzchannels that are mainly allocated inthe 380-400 MHz frequency band. Thelatter was approved by the ETSI in 2005and supports wideband services using50, 100, and 150 kHz channels.However, the maximum throughputusing TEDS is not enough for real-timeapplications. Therefore, the addition ofa broadband communication systemsimilar to one used in commercial 4Gmobile communications is beinginvestigated [8].In a legacy base station, the antennasystem can only cover 380-400 MHz,and adding extra antennas can beproblematic because ofelectromagnetic restrictions in differentcountries [9]. As a result, amultistandard multiplexed frequencyband has been proposed in whichTETRA and TEDS channels share the380-400 MHz band instead ofoccupying different bands. Prior toTEDS being released, countriesinvested heavily in TETRA V&Dnetworks, so it is important that anyfuture updates are compatible withthese legacy networks.For all PMR standards, includingTETRA and TEDS, the permittedchannel centre frequencies, f c(n ), aredefined byf c(n ) = band _ edge + (n -1) Δf c (1)2where n is the channel number, bandedge is the lower-edge frequency ofthe multiplexed frequency band, and Δf c is the channel spacing [10]. Theavailable spectrum is divided intofrequency sub-bands that are equal tothe channel spacing, and channels arecentered within each sub-band.In common hardware, themultiplexed frequency band solutionrequires fixed-channel allocation.Because the hardware cannot be easilyreconfigured, different sections of thespectrum are allocated for differentchannel sizes. This solution has poorspectrum efficiency because thechannel allocation is not optimized. Thetwo SDR-based approaches describedhere allow reconfiguration so thatchannels can be dynamically allocated,and channels of different sizes can beadapted as required. This means thatspectrum can be used more efficiently.To test these non-uniformchannelization structures, we created ascenario where the standard380-400 MHz PMR band was sharedby TETRA 25 kHz and TEDS 25, 50, and100 kHz channels. In this band, the5 MHz between 380-385 MHz wasreserved for the uplink signal, and the5 MHz between 390-395 MHz wasused for the downlink signal. TEDS150 kHz channels were not considered,but the procedure can be extended to150 kHz channels.3 Non-UniformChannelization Based onGDFT-FBChannelization methods can beuniform or non-uniform depending ontheir capacity to filter channels thathave equal or different bandwidthswithin the same frequency band.16ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P20

Practical Non-Uniform Channelization for Multistandard Base StationsSpecial TopicÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. FarrellUniform channelization usingpolyphase DFT-FBs has beenproposed for real-world applicationsthat have a large number of (uniform)channels. Various methods have alsobeen proposed for non-uniformchannelization [3], [5], [11]-[12].However, methods based on uniformDFT-DB are particularly attractivebecause of its low implementationcomplexity.Rather than using DFT-FB directly[6], GDFT-FB is used. GDFT-FB givesadditional control over channelfrequency allocation and phaseresponse because it has twoparameters in the subfilter complexmodulation [13]. In the filter bank, kfilters are obtained by complexmodulation of the lowpass prototypefilter, H(z ):-(k+k 0)n0(k+k 0)H k(z ) =W K H(zW K ) (2)where W k= exp(j 2π/K), and K is thenumber of sub-bands of the analysisbank. The parameter k 0 determines theway different sub-bands are stackedon the spectrum, and n 0 determines thephases of the different channels. Ifk 0 = 0 and n 0 = 0, the sub-bandspectrum allocation is even-stacked,and the first sub-band is centred atDC. In this case, the GDFT-FBstructure is reduced to the classicDFT-FB. If k 0 = 1/2 and n 0 = 0, thespectrum allocation is odd-stacked;that is, the sub-bands are shifted halfthe channel spacing, and no channel iscentred at DC. An odd-stackedapproach is used in the proposednon-uniform designs because it meetsthe channel allocation restrictionsdefined by (1). Different values of n 0can be chosen to provide extra phaseshifts to the FB outputs. In theeven-and odd-stacked cases, thesub-band spacing isΔf = f s/K (3)where f s is the sampling frequency ofthe wideband multichannel analysisbank input signal.The prototype filter, H(z ), may bedecomposed into its polyphasecomponents according toK-1H(z) = Σ z -i E i(z K ) (4)i =0where E i are the polyphasecomponents [6]. The rest of thesub-filters, H k(z), expressed in (2) canthen be obtained fromThe general implementation of theanalysis GDFT-FB for even- andodd-stacking cases is shown in Fig. 2.From (5), the different complexexponentials are applied to the differentbranches in order to obtain the desiredsub-band stacking and phase shifts.The complex exponentials, and-k 0iW K , can be directly hard-coded intothe polyphase components of the filter-k ibank, and W K denotes the DFT-k 0nDalgorithm. After the DFT, W K isapplied in order to present the differentoutputs, y k(n), centred at DC. Finally,-(k +k 0)n0W K is applied to the outputs forphase shift purposes. Depending onthe values of the decimation factor, D,the filter bank can be critically sampled(K =D) or oversampled (K = LD), whereL denotes the oversampling factor. Themain benefit of an oversampled FB isthat aliasing due to decimation issignificantly reduced [14]. However, anoversampled FB has additionalcomputational load because it runs at ahigher sample rate (by a factor of L).3.1 Parallel GDFT-FBA parallel GDFT-FB is proposed in[4], [15] as a non-uniformchannelization solution for basestations. The wideband signal isx (n)•z -1z -1z -1K-1-(k+kH k (z )=W K Σz -i W K W K E i(z K·W0 )n 0-ki -k 0 i-k k 0K ). (5)i =0DDDD-kW0 kKE 0(z L -k0WKK )E 1(z L -k0WKK )E 2(z L -k0WKK )E K-1(z L -k 0WKK )DFT: discrete Fourier transform▲Figure 2. Generalized DFT analysis filter bank structure.processed through different criticallydecimated odd-stacked GDFT-FBsoperating in parallel. Each FB uniformlydivides the frequency band accordingto specific channel spacing, and theFBs have overlapping frequency. In[15], the transmitter (synthesis) side isdescribed, and in [4], the receiver(analysis) side is described.In the parallel GDFT-FB channelizer,the digitized wideband signal withsample rate f S, is fed into multiple filterbanks running in parallel. There is onefilter bank for each uniform channelfrequency plan. Narrowband channelsare extracted by selecting anappropriate subset of outputs fromeach filter bank. Any permissiblecombination of channels can bespecified by choosing the appropriatefilter banks and channel numbers.Changing channel allocation inreal-time does not require redesign orre-optimization of the filter bankstructure. Only selection of appropriateoutputs needs to be adapted.Fig. 3 shows the parallel GDFT-FBfor the 5 MHz uplink band of aTETRA/TEDS base station. In this band,there are 200 channels for TETRA V&Dand TEDS 25 kHz. For TEDS 50 kHz,there are 100 channels, and for TEDS100 kHz, there are 50 channels. DFTmodulation is implemented using apower-of-2 FFT for efficiency; thus, thefilter bank sub-bands cover abandwidth larger than the frequencyband containing the informationchannels. The excess sub-bandsDFT-k 0nD-(0+k 0)n 0W K W K-k 0nD-(1+k 0)n 0W K W K-k 0nD-(2+k 0)n 0W K W K-k 0nD-(k-1+k 0)n 0W K W Ky 0(n )y 1(n )y 2(n )y K-1(n )December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 17D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P21

Special TopicPractical Non-Uniform Channelization for Multistandard Base StationsÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. FarrellNulls-fs /2 -2.5 MHZNon-Uniform Input29 30 16 9Nulls2.5 MHz f s /2GDFT-FB: generalized discrete Fourier transform filter bankTEDS: TETRA enhanced data serviceTETRAandTEDS25 kHz256-BandGDTF-FBTEDS50 kHz128-BandGDTF-FBTEDS100 kHz64-BandGDTF-FBTETRA: terrestrial trunked radio58◀Figure 3.…… ……… ……… …12830228229256114151141151281785764NullsSub-Band #29NullsNullsSub-Band #16NullsNullsSub-Band #19NullsCritically DecimatedChannels29-12.5 kHz 0 12.5 kHz16-25 kHz 0 25 kHz9-50 kHz 0 50 kHzParallel GDFT-FBchannelization structure.outside the 5 MHz bands arepermanently null.For the input signal in Fig. 3, startingfrom the lower edge(-2.5 MHz), the first two TETRA/TEDS25 kHz channels are extracted frombranches 29 and 30, respectively, ofthe first (25 kHz) FB. The next channelin the multiplexed spectrum, a 50 kHzTEDS channel, is selected as branch 16of the second (50 kHz) FB. Branch 15refers to the same frequency range asbranches 29 and 30 of the 25 kHz FB.Similarly, the next channel in themultiplexed spectrum, a TEDS 100 kHzchannel, is selected as branch 9 of thethird (100 kHz) FB.In the parallel GDFT-FB, the designcan be critically decimated oroversampled. The critically decimateddesign has more aliasing from adjacentbands, whereas an oversampleddesign has a higher computationalload. The level of aliasing interferencethat can be tolerated depends on theradio standard specifications.Parallel configuration is not the bestsolution when the channel spectrum isnot allocated according to (1), that is,when the possible channel centerfrequencies are not constrained to be amultiple of the channel spacing. In thiscase, additional GDFT-FBs must beadded to the parallel structure so thateach center frequency can be coveredby at least one GDFT-FB. Even withsome constraints, the recombinedGDFT-FB structure is a more flexiblenon-uniform channelization method.3.2 Recombined GDFT-FBA non-uniform recombination FB firstdivides a signal into uniformly spacedsub-bands and then recombinescertain groups of sub-bands to formwider bandwidths that are an integermultiple of the uniform spacing(Fig. 4a). The sub-band bandwidth isthe granularity band, and chosenaccording to the applicationrequirements.Non-uniform, recombined FBs havebeen proposed in literature on audioand speech processing [16]. In theseapplications, the FBs are criticallydecimated, so a perfect reconstructionalgorithm with parameter optimization isneeded to cancel the resulting aliasing.However, because of the asymmetricconfiguration shown in Fig. 1, such analgorithm is not possible.Oversampling ensures that aliasing inthe transition bands of the sub-bandbandpass filters is less than that in thecritically sampled case. Non-uniformchannelization using recombined,oversampled FBs has been proposedin [4], [17], [18]. Recombination iscarried out by the structure shown inFig. 4(b). A recombined signal, denotedY k,R(z ), is formed by R contiguoussub-bands allocated from the k thoutput of the GDFT-FB onwards:R -1Y k,R(z ) = Σ e jφ rY k+r (z M )H M (ze -jβ r). (6)r =0Therefore, channels are recombinedby interpolating each of the Rsub-bands by a factor M, frequencyshifting by β r to the correct centerfrequency, and phase correcting by φ rin order to combine these shiftedin-phase channels. To minimizeamplitude distortion in the recombinedchannels, anamplitude-complementary prototypefilter is required [19].Because the GDFT-FB outputs arealready oversampled by a factor L, theinterpolation factor can be M =R/L.Hence, the frequency and phase shiftsare, respectively,β r= π+π(2r+1) for r= 0, ..., R -1 (7)Rφ r= -(MN+N M) β r for r= 0, ..., R -1 (8)2D 2Dwhere N M is the order of theinterpolation filter, and N is the order ofthe GDFT-FB prototype filter. Toimprove the structure in Fig. 4(b), thephase shift, φ r , can occur before theinterpolation. This means the phaseshift is carried out at a lower samplerate than the anti-image filtering andfrequency shift.The FB is designed to cover an uplinkfrequency band of 5 MHz and deliver18ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P22

Practical Non-Uniform Channelization for Multistandard Base StationsSpecial TopicÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrellx (n)y k (n )y k+1(n )…y k+(R -1)(n )K -BandGDFT-FBMM…MH M(z )H M(z )…H M(z )e j β n(b)GDFT-FB: generalized discrete Fourier transform filter bank▲Figure 4. Recombined GDFT-FB channelizer (a) structure,(b) detail on the recombination structure.…Recomb.y R(n)Recomb.narrowband outputs with bandwidthand channel spacing of 25 kHz.However, the sample rate of thenarrowband channels is twice thechannel spacing, that is, K = 2D,because of oversampling. In thisparticular case, the bandwidthscovered are multiples of each other,and the chosen granularity band is25 kHz. Consequently, the TETRA/TEDS 25 kHz channels may beselected directly, without anyrecombination, from the appropriateoutput sub-bands of the GDFT-FB.For the wider 50 kHz and 100 kHzchannels, adjacent output sub-bandsof the FB must be recombined using thestructure in Fig. 4. TEDS 50 kHzchannels are obtained by recombiningtwo outputs. The channels do notrequire additional interpolation prior tofrequency shifting and combiningbecause of the original oversampling. ATEDS 100 kHz channel is obtained byinterpolating each channel by 2 andthen frequency shifting and combining.3.3 Evaluation of ChannelizersThe parallel GDFT-FB andrecombined GDFT-FB have the same01R -1kk +RK -1…… ……(a)y k +R +1(n )y k +R +2(n )y K -2(n )y K -1(n )e j β 1n0 e j φ 0 Σ…e j φ 1e j β R -1ne j φ R -1y 0,R (n)y k,R(n)y k,R(n)fundamental channelizationcapabilities when applied tochannels whosebandwidths are related byinteger multiples of eachother. In thesecircumstances, parallelGDFT-FB is best forschemes with few possiblechannel bandwidths andalignment patterns; forexample, TETRA/TEDS onlyhas 3 such bandwidth andalignment patterns.However, recombination isbest for schemes with agreater variety of channelbandwidths and alignmentpatterns. By decreasing thegranularity bandwidth of therecombined GDFT-FBsub-bands, a wider rangeof recombined bandwidthsis possible. Also, a widerrange of center frequenciesis possible for applicationsthat do not necessarilyconform to (1).Channelization structuresdynamically filter channels with differentbandwidths; however, these structurescan be differentiated according tocomputational load, filter designcomplexity, and filter implementationcomplexity.For complex I and Q input samples,the number of real multiplications andadditions for each complex inputsample in an analysis or synthesisodd-stacked GDFT-FB are,respectively,Lμ GDFT= [4(N+1)K K+4( (log 2(K )-1))+4K ] (9)2L Kα GDFT= [4(N+1- )K 2 3K+2( (log 2(K )-1))+2K ] (10)2where the first term is the number ofarithmetic operations for the N-ordercomplex-valued prototype filter, thesecond term is the number of arithmeticoperations due to the complex-valuedK-point radix-2 FFT [20], and the lastterm is the number of arithmeticoperations due to multiplication by acomplex exponential signal. If theGDFT-FB is oversampled, L > 1. Thenumber of arithmetic operations due tothe K-point FFT could be reduced byusing more efficient FFT algorithms,such as the split-radix FFT algorithm.However, power-of-2 FFT algorithms,such as the radix-2 and radix-4, arepreferred for practical implementationon FPGAs [21].For the parallel GDFT-FB, thecomputational load is the sum of thecomputational loads of eachcomponent GDFT-FB. Because all theGDFT-FBs run in parallel all the time,the total computational load is constantregardless of the channel allocationpattern.In comparison, the computationalload of the recombined GDFT-FBcomprises a fixed part that correspondsto the oversampled GDFT-FB structureand a variable part, whose complexitydepends on the number of recombinedchannels. The additional number of realmultiplications and additions per inputsample in each recombination structure(Fig. 4) are given by (11) and (12),respectively:L Nμ RECOMB = [M×R (2 l+1 4+4 + )] (11)K2 ML2α RECOMB = [M(R (2N l + 2 + )KM+2(R -1))] (12)where R is the number of sub-bands tobe recombined into a wider channel, Mis the interpolation factor required forrecombination, and N I is the order ofthe anti-alias filter required in theinterpolation. To make the evaluationmore concrete, the TETRA and TEDSspecifications are applied to both theparallel GDFT-FB and recombinedGDFT-FB. The specifications require astop-band rejection of 55 dB forsufficient channel selectivity. Apassband ripple of 0.1 dB is selected tominimize the amplitude distortion. Thelength of the FIR prototype filters usedin both channelization structures iscalculated using [22], and theParks-McClellan equiripple algorithm isused [23]. The asymmetric systemdesign in Fig. 1 creates aliasing in thechannels of the receiver analysis bank,which is minimized by tight filterDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 19D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P23

Special TopicPractical Non-Uniform Channelization for Multistandard Base StationsÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrellspecifications. This aliasing isespecially significant in the criticallysampled case [14].To approximate the order of thedifferent filters, the following equation isused [24]:-20log 10( δ pδ s)-13N ≈ . (13)14.6[(ω stop-ω pass)/2π]The filter order is a function of thepassband ripple, δ p, stop-bandattenuation, δ s , and normalizedtransition bandwidth (ω stop, - ω pass)where ω = 2πf /f s .Consequently, for thesame values of δ p, δ s, f pass, and f stop , thesample frequency, f s , determines thenormalized transition band and filterorder. For channelizers with a largenumber of channels, the samplefrequency of the wideband signal ismuch larger than the normalizedtransition band, and this leads to veryhigh orders.When applied to the TETRA andTEDS standards, the 25 kHz prototypefilter for both channelizer structures hasan order of 8085 taps. For the parallelGDFT-FB, the required prototype filterorders for the 50 kHz and 100 kHz FBsare 3584 and 1444 taps, respectively.However, these filter orders are onlytheoretical; the actual filter orders maybe higher [4] because interchannelinterference and aliasing produced bydecimation in the analysis bank causesthe filters’frequency response todeteriorate [14]. As the number ofchannels increases, aliasing alsoincreases, and additional filteroverdesign is required.Each GDFT-FB requires only oneprototype filter design; however, in thecase of TETRA/TEDS, the large filterorders make designing the filter andimplementing the correspondingchannelizer structures impractical.Large filter orders can be expected forother radio standards with similarspecifications.This impracticality arises because theefficient implementation of a filter in areconfigurable hardware platformtypically requires fixed-pointrepresentation of filter coefficients.Coefficients must be quantized to theword length of the device, and theresulting quantization error may changethe filter transfer function [25]. In an FIRfilter, these changes can lead todeviations in the magnitude response(particularly the stop-bandattenuation), and this renders the filterunsuitable for certain applications. Theproblem is greater with high-orderfilters because quantization errorsaffect the position of filter zeros in thez-domain, and the distance betweenzeros is reduced. Therefore, thefrequency response of high-orderfilters is more sensitive to smallchanges in the zero positions than thefrequency response of low-order filters.4 Modified GDFT-FB forFilter Coefficient ReductionAlthough the parallel GDFT-FB andrecombined GDFT-FB are flexible andmore efficient than other structures,high filter orders are still required forTETRA/TEDS. For high-orderchannelization structures and, moregenerally, in communicationapplications, FIR filters are chosenbecause of their linear phase response.To achieve a linear phase response inan FIR filter of length N+1, where N isthe order, (N+1)/2 coefficientscontribute to the magnitude response ofthe filter, and the other half of thecoefficients provide the linear phaseproperty [22]. As a consequence, anFIR filter design generally requires alarger number of coefficients comparedwith other designs. For applicationswhere perfect linear phase response isnot required, a minimum-phase FIR orinfinite impulse response (IIR) filter canprovide a functionally equivalentmagnitude response with a reducednumber of filter coefficients.When linear phase response isrequired, multistage filtering [26] is auseful technique that can be applied toFIR filter design to reduce the totalnumber of filter coefficients. Multistagefiltering is most commonly applied tointerpolators and decimators that havelarge sample rate conversion factors. Ina multistage filter design, an originalfilter is factorized into multiplecomponent filters which, whencascaded, produce the original filtermagnitude and phase response.Component filters have more relaxedspecifications than the original filter.Therefore, the number of coefficients ineach component filter is smaller (oftenmuch smaller) than the original filter.The design of each component filter isalso simplified, and the totalcomputational load is often less thanthe original filter.Here, we show that multistagefiltering can be extended to modulatedFBs and, in particular, to the GDFT-FB.In FB literature, the term multistagetypically refers to a structure in whichmultiple FB stages are cascaded toform a complete FB [5], [27]. In theapproach presented here, themultistage technique is applied to theprototype filter of only one FB.There are various ways in which themultistage technique could be appliedto the prototype filter. The specificationof the prototype filter, H(z ), and itspolyphase components, is relaxed, anda half-band filter,H B (z ), is applied toevery sub-band output to obtain theoriginal filtering specifications. Thisdesign is shown in Fig. 5.In the single-stage design, theprototype filter transition band is narrowand centered at π/K. In the multistagedesign, the prototype filter transitionband is shifted so that it starts at π/Kand its width is increased so that itextends to 2π/K and includesfrequency components from adjacentchannels. The half-band filter on eachsub-band output provides the originalsharp transition band and alsoeliminates the undesired frequencycomponents from the adjacentchannels passed by the relaxedprototype filter. This filtering process isshown in Fig. 6.For a multistage GDFT-FB, the FBcannot be critically decimated, but ithas to be oversampled by 2, that is,K = 2D. Oversampling prevents theundesired adjacent-channelfrequencies from aliasing with thechannel of interest. Also, oversamplingby 2 allows the use of half-band filters.This is attractive because half-bandfilters have desirable impulse andfrequency response properties inFig. 7(a). Half-band filters have asymmetric impulse response where20ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P24

Practical Non-Uniform Channelization for Multistandard Base StationsSpecial TopicÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrell-k 0nD-(0+k 0)n 0W K W Kx (n)y 0(n )E 0(z L -k• W 0 KDK )H B(z )z -1z -1z -1DDDE 1(z L -k0WKK )E 2(z L -k0WKK )E K-1(z L -k 0WKK )▲Figure 5. Multistage analysis GDFT-FB. A half-band filter is added after every sub-band output.every second coefficient is zero;therefore, in implementation, onlyapproximately 1/4 of the half-band filtercoefficients have to be computed. Interms of frequency response,half-band filters allocate the -6 dBpoint at exactly π/2 radians. This makesthem useful for recombined GDFT-FBsbecause the magnitudecomplementary property is easilyachieved. Further advantages can begained from the two-band polyphaseimplementation of the half-band filterusing (4). This implementation allowsdown-sampling to be performedbefore the filtering. Also, because of thezero-value coefficients, one of thepolyphase branches is formed by apure delay and is followed by themiddle filter coefficient [23]. Combined,these two advantages lead to fewerDFTDFT: discrete Fourier transformH (e jω )-k 0nD-(1+k 0)n 0W K W Ky 1(n )-k 0nD-(2+k 0)n 0W K W Ky 2(n )y K-1(n )-k 0nD-(k-1+k 0)n 0W K W K-π -4π/K -3π/K -2π/K -π/K DC π/K 2π/K 3π/K 4π/K π ω(a)HB (e jω )-π -π/2 DC π/2(b)π: Filters MagnitudeResponse: Input ChannelsωH B(z )H B(z )H B(z )z 0(n )z 1(n )z 2(n )z K -1(n )◀Figure 6.multistage GDFT-FBfiltering operationsapplied to one of theinformation channels.(a) GDFT-FB filteringthe channel centered atDC with relaxedtransition band. (b)Half-band filterperforming the sharpfiltering of the channeland eliminating thefrequency componentsfrom adjacent channels.operations per second than in thenon-polyphase implementation.By adding extra filter operations to (9)• ••• • •• • 0-N /20-6DCHB (e jω ) (dB)hB [n ]•0.5π/2(a)•• • •• n•N /2πand (10), the number of realmultiplications and additions percomplex input for the channelizer isnow determined bywhere N B is the order of the half-bandfilters.This multistage GDFT-FB canreplace the single-stage GDFT-FB inthe parallel and recombined GDFT-FBchannelization methods without anysignificant changes being made tothese methods. For the recombinedGDFT-FB, the phase shift introduced inthe recombination structure shown inFig. 4 and described by (7) is nowchanged towhich compensates for the changedphase response slope introduced bythe half-band filters on each sub-bandoutput.4.1 Channelizers Evaluation Usingmultistage GDFT-FBTo demonstrate the advantages ofreducing the coefficients in themultistage design, the sameTETRA/TEDS examples used for the▲Figure 7. (a) Half-band filter impulse and frequency response (b) Efficient polyphaseimplementation of half-band filter plus decimator.ωLKμ M-GDFT= [4(N+1)+4( (log 2(K)-1))K2(N+4K+2K B+1)] (14)8L K 3Kα M-GDFT= [4(N+1- )+2( (log 2(K )K 2 2x (n )z -1 2-1))+2K+2K ] (15)42N Bφ r= -[M (N+N B)+N M]β r2D 2 2for r = 0, ..., R-1 (16)HB 0(z )y (n )0.5z N B 1/2(b)December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 21D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P25

Special TopicPractical Non-Uniform Channelization for Multistandard Base StationsÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrell▼Table 1. Filter orders calculated for TETRA/TEDS non-uniform channelizers using multistageGDFT-FBParallel GDFT-FBRecombined GDFT-FBGDFT-FB OutputSub-Band Bandwidth (kHz)two non-uniform channelizationmethods are used here. The theoreticalfilter orders are calculated again using(13). Even though the multistage designhas two filters instead of one, it has twomain advantages: its prototype filter isdesigned with a much larger transitionband, and the half-band filter isdesigned for a much lower sample rate.Both factors contribute to reducing thefilter orders shown in Table 1.A reduction in the number ofcoefficients also implies a reduction inthe number of operations per complexinput sample (Table 2). For themultistage recombined GDFT-FB inTable 2, the figures represent only theGDFT-FB part of the channelizer. Thestructure used to recombine the TEDS50 kHz and 100 kHz channels is thesame for both single and multistagecases and is not included for thatreason. The saving on multistagefiltering computational load is greatestin the recombined GDFT-FB.The performance of the twonon-uniform channelizers can becompared in terms of their total numberof real multiplications per sample fordifferent sub-band allocation patterns[4]. Both architectures were tested inscenarios where the frequency band iscovered with a single type of channel,for example, 25, 50, or 100 kHz, orwhere it is covered by various types ofchannels, for example, 50 and 100 kHz.255010025Single-Stage Numberof CoefficientsN = 8085N = 3584N = 1444N = 8085GDFT-FB: generalized discrete Fourier transform filter bankmultistage Numberof CoefficientsN = 1294, N B = 64N = 595, N B = 58N = 291, N B = 46N = 1294, N B = 64▼Table 2. Number of coefficients and computational load of the TETRA/TEDS non-uniformmultistage channelizers compared with single-stage designPercentage of Coefficients Compared with Single-Stage DesignPercentage of Real Multiplications Compared with Single-Stage DesignPercentage of Real Additions Compared with Single-Stage DesignmultistageParallel GDFT-FB16.7%70.6%92.3%GDFT-FB: generalized discrete Fourier transform filter bankmultistageRecombined GDFT-FB16.0%35.3%45.0%The test results are shown in Fig. 8.The parallel GDFT-FB has a constantcomputational load that is independentof the sub-band allocation pattern. Onthe other hand, the computational loadof the recombined GDFT-FB variesaccording to the number and type ofrecombined channels. The load is at aminimum when the whole frequencyband is covered by 25 kHz channelsbecause no recombining is required.The load is at a maximum when thefrequency band is fully occupied byTEDS 100 kHz channels; that is,recombination is applied to everyGDFT-FB output sub-band. In contrastto the results obtained in [4], therecombined GDFT-FB using multistagefiltering is more efficient than theFigure 8. ▶Performance of parallelGDFT-FB andrecombined GDFT-FBwhen multistage filteringis applied comparedwith [20].Real Multiplications per Input Sample280260240220200180160140120100010equivalent parallel GDFT-FB in everysituation. The reason for this is thatmultistage filtering has a greater effecton the recombined GDFT-FB.For the multistage channelizers, thefilter orders were again obtained usingtheoretical calculations (section 3.3).However, real implementations requirehigher filter orders because ofundesired effects, such as aliasing, inthe FBs (Fig. 9). In Fig. 9, a channelizerfor eight 25 kHz channels wasdesigned using the single-stage andmultistage GDFT-FBs. The theoreticalorder for the single-stage prototypefilter was 253, whereas that for themultistage prototype filter was 42, andthe half-band filter order was 64. Allfilters were designed as optimalequiripple filters.Because of aliasing, the magnituderesponse does not exhibit 55 dB ofattenuation in the stopband for thesingle-stage design. For the multistagedesign, the cascade of the two filtersproduces decay in the stopband thatprovides the desired attenuation inmost of the stopband. In the passband,both filters provide a ripple within the0.1 dB limits. To conform to the requiredspecifications, an overdesign of theideal prototype filter is needed tocompensate for the aliasing. Thisusually leads to filter orders that arelarger than the theoretical orders,especially in the single-stage design.Optimization techniques are outside thescope of this paper, butKaiser-Window filters could possiblybe used instead of optimal equiripple20 30 40 50 60 70 80Number of Wideband Channels (>25 KHz): Parallel FB: Recombined FB 25 kHz only: Recombined FB 50 kHz only: Recombined FB 50:50 split: Recombined 100 kHz only9010022ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P26

Practical Non-Uniform Channelization for Multistandard Base StationsSpecial TopicÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. FarrellMagnitude Response (dB)Magnitude Response (dB)100-10-20-30-40-50-60-70-80-90-10000.050.040.030.020.010.00-0.01-0.02-0.03-0.04-0.0500.10.050.20.100.3 0.4 0.5 0.6Frequency (rad/π)(a) Full magnitude response0.150.20 0.25 0.30Frequency (rad/π)(b) Passband ripple detailones. In normal circumstances, filtersdesigned with the Kaiser-Windowmethod need more coefficients thanoptimized filters. On the other hand,they have decay in their stopband thatminimizes aliasing when applied toFBs [28].4.2 Fixed-Point FPGA ChannelizerImplementationPhysical implementation of parallelGDFT-FB and recombined GDFT-FBchannelizers is aided by efficient FIRdesign and FFT processing tool boxessuch as the Xilinx LogiCORE FIRCompiler [29] and Xilinx LogiCORE FFT[21] (which can be used with Virtex andSpartans FPGAs). These toolboxessupport most of the component parts ofthe parallel GDFT-FB and recombinedGDFT-FB structures. Nevertheless,such toolboxes often have limitations;for example, the FIR compiler limitsfilters to 1024 coefficients. With minorfilter optimization, the multistagechannelizers could be implementedwith these toolboxes. However, thecoefficient limitation prevents thesetoolboxes being used to implementsingle-stage channelizers because: Classic GDFT-FB: Multistage GDFT-FB0.70.80.9 1.0◀Figure 9.Classic and multistageGDFT-FB outputmagnitude response foran 8-channels TETRAchannelizer.they require much higher filter orders.Single-stage implementations mightstill be possible using a customizeddesign, but this implementationrequires longer development time.Mobile communication systemsgenerally use modulation schemesbased on real-value in-phase andquadrature (I/Q) signals. For thesemodulation schemes, complex signalprocessing is useful for simplifying thesignal operations and notation. Insteadof considering the real-value I/Q inputsignals separately, in complex signalprocessing the I/Q signals are mappedon a complex input value as the realand imaginary parts respectively [30] .However, the system must bephysically implemented using realsignal operations. Consequently, thereal-value I and Q components areprocessed in two different physicalpaths. Filtering or frequency mixing hasto be applied to both componentsseparately. For the filterimplementations, depending on thedesired odd or even sub-bandstacking configuration of the GDFT-FB,the coefficients of the FB prototype areeither real- or complex-valued, asshown in (5). For real filters, thereal-value coefficients of the prototypefilter H(z ) are applied independently tothe I and Q components (Fig. 10),where R (z ) = H (z ) and Q (z) = 0. Incontrast, complex digital filterimplementations require moreresources because their complexcoefficients need to be reduced to realvalues [30]. For these implementations,the two real-value filter components, R(z ) and Q (z ), are obtained from thecomplex filter H (z ):: Classic GDFT-FBH (z )+H *(z): Multistage GDFT-FB R (z ) =2(17)H (z )-H *(z)jQ (z ) =2(18)0.350.400.450.50Like real FIR filters, complex FIRfilters have symmetric properties thatcan be exploited in the R (z) and Q (z )implementation [31]. Therefore, acomplex filter requires twice thenumber of multipliers as a real digitalfilter of the same length. Theeven-stacked GDFT-FB withreal-value coefficients enables efficientphysical implementation. Fixed-pointrepresentation is also important inphysical implementation. For multiratefilter banks, using fixed-pointarithmetic to represent coefficients andsignals produces errors such asanalogue-to-digital input quantizationnoise, coefficient quantization errors,rounding errors, overflow errors, andsub-band quantization errors[32]-[33]. In FIR prototype filters,coefficient quantization errors do notaffect the linear phase responsecharacteristics but can affect themagnitude response. By using themultistage design with lower-orderfilters, the zeros in the z -plane arefurther apart; therefore, the filtermagnitude response is lesssensitive to quantization error than infilters with higher-order single-stagedesigns [25].5 ConclusionIn this paper, we have described thepractical implementation of twonon-uniform channelization techniquesbased on GDFT-FBs. These FBs areefficient and flexible, but the use of FIRDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 23D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P27

Special TopicPractical Non-Uniform Channelization for Multistandard Base StationsÁlvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrellx I (n )x Q (n )••Q (z )filters generally leads to impracticallyhigh filter orders. To overcome thisproblem, a multistage filtering structurewas applied to the GDFT-FB to reducethe number of coefficients. Taking thisapproach, the number of filtercoefficients and operations per inputsample are reduced compared with thesingle-stage design. These reductionsmake the physical implementation morepractical for fixed-point implementationusing existing FPGA tools.AcknowledgmentThe authors wish to thankJean-Christophe Schiel and FrançoisMontaigne for their assistance andsupport. Also the authors extend thanksto the sponsors EADS and IRCSET forthe Ph.D. program.References[1] P. Leaves, et al., "Dynamic spectrum allocation incomposite reconfigurable wireless networks,"Communications Magazine, IEEE, vol. 42, pp.72-81, 2004.[2] T. Hentschel, "Channelization for Software DefinedBase-Stations," Annales de Telecommunications,May/June 2002.[3] R. Mahesh, et al., "Filter Bank Channelizers forMultistandard Software Defined Radio Receivers,"Journal of Signal Processing Systems, SpringerNew York, 2008.[4] A. Palomo Navarro, et al., "Non-UniformChannelization Methods for Next Generation SDRPMR Base Stations," in IEEE Symposium inComputers and Communications (ISCC 2011), 2011.[5] C. Zhao, et al., "Reconfigurable Architectures forLow Complexity Software Radio Channelizers usingHybrid Filter Banks," in Communication systems,2006. ICCS 2006. 10th IEEE Singapore InternationalConference on, 2006, pp. 1-5.[6] P. P. Vaidyanathan, Multirate Systems and FilterBanks: Prentice Hall PTR, 1993.[7] W. A. Abu-Al-Saud and G. L. Stuber, "Efficientwideband channelizer for software radio systemsusing modulated PR filterbanks," Signal Processing,IEEE Transactions on, vol. 52, pp. 2807-2820, 2004.[8] ECC, "Public protection and disaster relief spectrumrequirements (ECC report 102)," HelsinkiJanuary2007.[9] Electromagnetic compatibility and Radio spectrumR (z ) yI (n )R (z )Q (z )-yQ(n )◀Figure 10.Physical implementationof a real digital filter(shaded) and complexdigital filter applied to I/Qsignals.Matters (ERM); TETRA Enhanced Data Service(TEDS); System reference document, ETSI TR 102491 V1.2.1, 2006-05.[10] Planning criteria and coordination of frequencies inthe land mobile service in the range 29.7-921MHz, T/R 25-08, 2008.[11] A. Eghbali, et al., "A Farrow-structure-basedmulti-mode transmultiplexer," in Circuits andSystems, 2008. ISCAS 2008. IEEE InternationalSymposium on, 2008, pp. 3114-3117.[12] A. Eghbali, et al., "Dynamic Frequency-BandReallocation and Allocation: from Satellite-BasedCommunication Systems to Cognitive Radios,"Journal of Signal Processing Systems, pp. 1-17,2009.[13] R. E. Crochiere and L. R. Rabiner, Multirate DigitalSignal Processing: Englewood Cliffs (NJ): PrenticeHall, 1983.[14] Q.-G. Liu, et al., "Simple design of oversampleduniform DFT filter banks with applications tosubband acoustic echo cancellation," SignalProcessing, vol. 80, pp. 831-847, 2000.[15] A. Palomo Navarro, et al., "Overlapped polyphaseDFT modulated filter banks applied to TETRA/TEDS SDR base station channelization," presentedat the Royal Irish Academy Communication andRadio Science Colloquium, 2010.[16] X. M. Xie, et al., "Design of linear-phaserecombination nonuniform filter banks," SignalProcessing, IEEE Transactions on, vol. 54, pp.2809-2814, 2006.[17] F. J. M. G. Harris, R., "A receiver structure thatperforms simultaneous spectral analysis and timeseries channelization," in SDR 09 TechnicalConference and Product Exposition, 2009.[18] H. Johansson and P. Lowenborg, "Flexiblefrequency-band reallocation networks usingvariable oversampled complex-modulated filterbanks," EURASIP J. Appl. Signal Process., vol.2007, pp. 143-143, 2007.[19] S. Radhakrishnan Pillai and G. H. Allen,"Generalized magnitude and powercomplementary filters," in Acoustics, Speech, andSignal Processing, 1994. ICASSP-94., 1994 IEEEInternational Conference on, 1994, pp. III/585-III/588 vol.3.[20] P. Duhamel and M. Vetterli, "Fast fouriertransforms: a tutorial review and a state of the art,"vol. 19, ed: Elsevier North-Holland, Inc., 1990, pp.259-299.[21] Xilinx, "LogiCORE IP Fast Fourier Transform v7.1Product Specification," ed, 2010.[22] T. Saramäki, "Finite impulse response filter design,"in Handbook for Digital Signal Processing, S. K.Mitra and J. F. Kaiser, Eds., ed New York , NY,USA: John Wiley & Sons, 1993, pp. 155-227.[23] F. J. Harris, Multirate Signal Processing forCommunication Systems: Prentice Hall PTR, 2004.[24] J. Kaiser, "Nonrecursive Digital Filter Design Usingthe IO-Sinh Window Function," in IEEEInternational Symposium on Circuits and Systems,1974.[25] J. G. Proakis and D. G. Manolakis, Digital signalprocessing: Principles, Algorithms andApplications: Pearson Prentice Hall, 2007.[26] L. Milic, Multirate Filtering for Digital SignalProcessing: Information Science Reference, 2009.[27] C. Y. Fung and S. C. Chan, "A multistagefilterbank-based channelizer and itsmultiplier-less realization," in Circuits andSystems, 2002. ISCAS 2002. IEEE InternationalSymposium on, 2002, pp. III-429-III-432 vol.3.[28] K. F. C. Yiu, et al., "Multicriteria design ofoversampled uniform DFT filter banks," SignalProcessing Letters, IEEE, vol. 11, pp. 541-544,2004.[29] Xilinx, "IP LogiCORE FIR Compiler v5.0 ProductSpecification," ed, 2011.[30] K. W. Martin, "Complex signal processing is notcomplex," Circuits and Systems I: Regular Papers,IEEE Transactions on, vol. 51, pp. 1823-1836,2004.[31] F. Bruekers, "Symmetry and Efficiency in ComplexFIR Filters," Philips Research Laboratories,Eindhoven, 2009.[32] M. Abo-Zahhad, "Current state and futuredirections of multirate filter banks and theirapplications," Digital Signal Processing, vol. 13,pp. 495-518, 2003.[33] T. Karp and A. Mertins, "Implementation ofbiorthogonal cosine-modulated filter banks withfixed-point arithmetic," in Circuits and Systems,2001. ISCAS 2001. The 2001 IEEE InternationalSymposium on, 2001, pp. 469-472 vol. 2.BiographiesÁlvaro Palomo Navarro (apalomo@eeng.nuim.ie)received his B.Eng. degree in telecommunicationsengineering from the Polytechnic University ofMadrid (UPM) in 2006. He carried out his final yearproject in the Signal Processing Department at BTH,Sweden. Between 2006 and 2007 Álvaro worked asa test engineer of GSM intelligent networks. Since2007 he has been a Ph.D. candidate at the CallanInstitute at NUI Maynooth. His main researchinterests include multirate digital signal processing,adaptive equalization, software-defined radio andMultistandard wireless communications systems.He also collaborates with the Electronic EngineeringDepartment as occasional lecturer and tutor.Rudi Villing (rudi.villing@eeng.nuim.ie), MIEEE,received his B.Eng. degree in electronicengineering from Dublin City University in 1992. Hespent the next 10 years working in thetelecommunications software industry and becamea specialist in telecommunications managementnetworks (TMN). He worked as a product architectwith Euristix Ltd. and Marconi Plc, creating thearchitecture for strategic network managementtechnologies before joining the ElectronicEngineering Department at NUI Maynooth in 2002.He received his Ph.D. degree from NUI Maynooth.His research interests include wirelesscommunications and applications and perceptualsignal processing.Ronan J. Farrell (ronan.farrell@eeng.nuim.ie)received his B.E. and Ph.D. degrees in electronicengineering from University College Dublin in 1993and 1998. He is currently a senior lecturer at theNational University of Ireland, Maynooth, anddirector of the Callan Institute for applied ICT. Hisresearch interests include physical layercommunication technologies, in particular, adaptivereceivers, Pas, and active antenna arrays. He iscurrently the strand leader responsible for radiotechnologies in the SFI-funded Centre forTelecommunications Research.24ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P28

Crest Factor Reduction for OFDM Using Selective Subcarrier DegradationSpecial TopicR. Neil BraithwaiteCrest Factor Reduction for OFDMUsing Selective SubcarrierDegradationR. Neil Braithwaite(Powerwave Technologies, Santa Ana, CA 92705, USA)Abstract: This paper describes a crest factor reduction (CFR) method that reduces peaks in the time domain by modifying selected datasubcarriers within an OFDM signal. The data subcarriers selected for modification vary with each symbol interval and are limited to thosesubcarriers whose data elements are mapped onto the outer boundary of the constellation. In the proposed method, a set of peaks areidentified within an OFDM symbol interval. Data subcarriers whose data element has a positive or negative correlation with the set peakare selected. For a subcarrier with an outer element and a significant positive correlation, a bit error (reversal) is intentionally introduced.This moves the data element to the opposite side of the constellation. Outer elements on negatively-correlated subcarriers are increasedin magnitude along the real or imaginary axis. Experimental results show that selecting the correct subcarriers for bit reversals andoutward enhancements reduces the peak-to-average power ratio (PAPR) of the OFDM signal to a target value and limits in-banddegradation measured by bit error rate (BER) and error vector magnitude (EVM).Keywords: crest factor reduction; OFDM; PAPR; wireless communication system; digital transmitterO1 Introductionrthogonal frequency-divisionmultiplexing (OFDM) waveformscan have largepeak-to-average power ratios(PAPR). Crest factor reduction (CFR)reduces peaks at the expense of signalquality. Usually, degradation isdistributed throughout the signal’sfrequency bandwidth in the form ofspectral leakage and in-band errors.However, because the OFDM signal iscreated in the Fourier domain,degradation can be concentrated tospecific subcarriers (frequency bins ofthe FFT).In the proposed method, CFR is usedsparingly to bound the PAPR whentransmitting at high power levels. Thisdiffers from the usual approach ofmaximizing PAPR reduction. The boundis met, and bit error rate (BER) is limitedwithout deviating from the OFDMstandard — for example, WorldwideInteroperability for Microwave Access(WiMAX) 802.16 [1]. This peak powerbound allows the power amplifier (PA)in the transmitter to be designed forgreater power-added efficiency (PAE)[2]. It may be necessary to enforce asecond PAPR bound if the dynamicrange of the digital-to-analogconverter (DAC) cannot preventclipping for all possible signals.The remainder of this introductiondescribes OFDM signal generation aswell as the subcarrier phase alignmentsthat create peaks. Section 2 contains areview of CFR methods. Section 3introduces new CFR methods forintentionally degrading selectedsubcarriers whose data elements arecorrelated (phase aligned) with thepeaks of the OFDM signal. Section 4contains experimental results that showa WiMAX signal can be crest-factorreduced to a target PAPR byintroducing an acceptable amount ofin-band degradation. This degradationis measured by the BER and errorvector magnitude (EVM).1.1 OFDM Transmit OverviewCreation of an OFDM symbol forradio frequency (RF) transmission isshown in Fig. 1. The symbol is part of adata stream that has been 1) encoded,2) modulated, 3) converted from serialto parallel as a 256 sample block, 4)converted to the time domain using aninverse fast Fourier transform (IFFT), 5)extended using a cyclic prefix (CP) to256 + CP samples, and 6) convertedback into a serial data stream. The datastream is then converted from DAC,lowpass filtered, up-converted to RF,then amplified by a PA. The signals Y(k)and y (n) are important for the crestfactor method presented in this paper.These symbols correspond to theFourier domain and time domain dataDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 25D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P29

Special TopicR. Neil BraithwaiteCrest Factor Reduction for OFDM Using Selective Subcarrier DegradationDataInputEncoderY (k)1 Modulator:BPSK, QPSK, 1 Serial 25616-QAM, orto64-QAMParallelDigitalTransmitterInputDACLowpassFilterIFFTy (n )256 AppendCyclicPrefixPA256+CP ParalleltoSerialRFTransmissionTo DigitalTransmitter1QPSK, it varies between constellationelements for 16-QAM and 64-QAM.Higher magnitudes are found at theouter elements of the 16-QAM and64-QAM constellations. As a result, itcan be assumed that many of theseouter constellation elements arepresent in the data subcarriers when alarge peak appears in the time domain.BPSK: binary phase-shift keyingCP: cyclic prefixDAC: digital-to-analog converterIFFT: inverse fast Fourier transformLO: Local oscillatorPA: power amplifier▲Figure 1. RF transmission of one OFDM symbol [3].blocks, respectively, of the OFDMsymbol.The encoding includes coders andinterleavers to allow for error correctionat the receiver. The modulator allowsrate changes, and the rate is selectedaccording to the receivedsignal-to-noise ratio (SNR) and BER atthe receiver. The rates from lowest tohighest are binary phase-shift keying(BPSK), quadrature phase-shift keying(QPSK), 16-quadrature amplitudemodulation (16-QAM), and 64-QAM.In the serial-to-parallel conversion, afrequency domain representation iscreated, and the data elements areassigned to different subcarriers of theOFDM signal. The IFFT transforms thedata from the frequency domain to thetime domain. The cyclic prefix is a copyof the tail of the time domain block andis appended to the beginning of thetime domain block. The cyclic prefixprotects against intersymbolinterference (ISI) caused by multipathRF propagation.Within an OFDM symbol, thesubcarriers may be data subcarriers,pilot subcarriers, or null subcarriers. Nodata elements are mapped onto the nullsubcarriers, which include the outerguard-band frequencies and the DCsubcarrier. The pilots areBPSK-modulated and assigned tospecific subcarriers. The remainingsubcarriers are used for datatransmission, which may be modulatedusing BPSK, QPSK, 16-QAM, or64-QAM.The OFDM-transmitted signal is asequence of symbols sent as aLOQAM: quadrature amplitude modulationQPSK: quadrature phase-shift keyingdown-link (DL) subframe. Thissubframe comprises a preamble, framecontrol header (FCH), and DL bursts.QPSK is used for the preamble andBPSK for the FCH. BPSK, QPSK,16-QAM, or 64-QAM (except for theBPSK pilots) is used for the DL bursts.The preamble and FCH are sent first,and the DL bursts are sent in order ofthe modulation rates (lower rates aresent first).The OFDM symbol at the output is atime-domain data sequence. Althoughthe individual data subcarriers aretransmitted using simple modulationmappings, the magnitude in the timedomain varies significantly. This is dueto the IFFT operation that forms eachtime-sample from a sum of 200random-phase variables (56 of the 256subcarriers are null subcarriers). Phasealignment of subcarriers in thefrequency domain results in largepeaks in the time domain.Peak-forming phase alignment in thefrequency domain varies according tothe position of the peak after the IFFTwithin the time block. A peak at timet peak within the interval t = [0, 255] ismaximized by the following subcarrierphases:θ align(t_peak)(k ) = -k·Δω·t peak +θ (t peak) (1)where Δω = 2π (N = 256), k is theNsubcarrier frequency index (DC = 0),and θ (t peak) is the phase of the complextime sample at t peak.The magnitude of the subcarriersalso has an effect on peaking. Althoughthe magnitude is constant for BPSK and2 Crest Factor ReductionLarge peaks cause problemsbecause PAs become less efficient asthe PAPR of the RF signal increases.Limiting the PAPR is necessary for amore efficient transmitter design. Thisprocess is called CFR. Past CFRmethods for OFDM signals include clipand filter, partial transmit sequence(PTS), selective mapping (SLM), tonereservation, and constellationextension. The following is a summaryof these CFR methods and theirsuitability for OFDM signals.The direct CFR method involvesclipping peaks of the time signal y (n )when they exceed a specified level, L.The clipped signal y clip(n ) isL· y (n ) for y (n ) >Ly clip(n ) = y (n )(2)y (n )otherwise.The excess peak waveformy peaks(n ) becomesy peaks(n ) = y (n ) - y clip(n ). (3)Clipping moves the constellationelements from their assigned positions.The difference between the actual andassigned positions in the IQ space iscalled the constellation error or EVM.The allowable relative constellationerror for WiMAX, averaged oversubcarriers, frames, and packets,depends on the rate modulation wherethe most difficult specification is -31.0dB for the 3/4 rate 64-QAM. Clippingtends to distribute excess peak energyover the 256 subcarriers, including thenull and pilot subcarriers. As well asEVM limits on the data subcarriers,there are also limits on spuriousemissions and adjacent channelleakage ratio (ACLR). These limits arespecifically for the null subcarriers26ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P30

Crest Factor Reduction for OFDM Using Selective Subcarrier DegradationSpecial TopicR. Neil Braithwaitewithin the guard-band frequencies.The energy from the clipped peaksmust be constrained to reside primarilyon the in-band subcarriers.Filtering attenuates the excess peakenergy present in the null subcarriers(except for the DC subcarrier). That is,Δy (n ) = Σy peaks(n -τ)·h (τ) (4)τwhere h (τ ) is a filter kernel. Thecrest-factor-reduced signal beingtransmitted, denoted y CFR(n ), becomesy CFR(n ) = y (n ) - α·Δy (n ) (5)where α is a scaling term used tocontrol the EVM introduced by the CFR[4], [5]. In some clip and filterimplementations [6], the excess peakwaveform y peaks(n ) is replaced by asequence of delta functions. Each deltafunction is located at a peak andassigned a magnitude and phase thatmatch the excess value of the peak.This sequence is filtered using (4).Using α in (5) for controlling EVM ofOFDM signals is not effective becausethe EVM limit varies according to thefunction of the subcarrier and the ratemodulation of the data (BPSK, QPSK,16-QAM, or 64-QAM). It is beneficialfor the receiver to have accurate pilotinformation, which means theCFR-induced EVM should be zero forpilot subcarriers. Multiaccess OFDM,such as OFDMA and LTE, may havedifferent rate modulations on differentsubcarrier blocks. Because theallowable EVM is larger for BPSK andQPSK than for 16-QAM and 64-QAM,some researchers [7] convert theexcess peak signal back into thefrequency domain and apply thefiltering and rate-dependent EVMcontrol by weighting the subcarriers.The result is then converted back intothe time domain to produce the desiredΔy (n ) and y CFR(n ).The EVM control for a data subcarrierdepends on the modulation type. CFRattempts to introduce EVM on datasubcarriers to reduce peaks withoutincreasing the BER at the receiver.Thus, QPSK allows more EVM than64-QAM because the distancebetween neighboring points in theconstellation is greater. However, someresearchers [8] have exploited the factthat points on the outer boundary of theconstellation have no neighbors in theoutward-expanding direction. As aresult, EVM need not be limited in theoutward direction because it does notcause bit errors at the receiver.Outward expansion is also used in theproposed CFR and is discussed insection 3.In a different class of CFR methods,CFR disrupts the subcarrier phasealignments that create peaks. Thesemethods include partial transmitsequence (PTS) [9]-[11] and selectivemapping (SLM) [12]. The subcarriersare multiplied by a set of differentphase-shift vectors that produce a setof potential time sequences. The timesequence with the lowest PAPR istransmitted. Information about thephase-shift vector used must be sentto the receiver to allow demodulation.This is considered a disadvantage ofthe method. A PTS method specificallyfor LTE and not requiring thephase-shift vector to be sent to thereceiver is described in [13]. NeitherPTS nor SLM is used in the proposedCFR method.A third CFR method involvesreserving some of the subcarriers aspeak-reducers; that is, the reservedsubcarriers do not carry any data. Thismethod is called tone reservation [14],[15]. Once a peak is detected in thetime domain, the magnitudes andphases of the reserved subcarriers areselected to reduce the peak.Unfortunately, data throughput isreduced because fewer datasubcarriers are available. It is possibleto use tone reservation as a form of clipand filter with EVM control. The EVM isset to zero for all data subcarriers thatare not reserved for peak reduction.The scale factor, α , may have to beincreased above unity for such animplementation in order to compensatefor the sparseness of the reservedsubcarriers in the frequency domain.Reserved tones are not used in theproposed CFR method.A fourth CFR method involves alteringthe constellation mapping so thatelements are not unique. This isreferred to as constellation extension[16]. A redundant mapping of input bitsis used so that opposing points on theconstellation, d IQ and -d IQ, represent thesame data at the receiver. Theadvantage of constellation extension isthat π radian phase shifts can beintroduced onto some of the subcarrierswith the goal of disruptingpeak-forming phase alignments. Thedownside of this method is that one bitis lost in the constellation mapping,which reduces throughput for QPSK,16-QAM, and 64-QAM to 1/2, 3/4, and5/6 of the original value, respectively.Also, the Gray code mapping specifiedin the standard must be abandoned.The proposed CFR method similarlyintroduces large phase shifts ontoselected subcarriers; however, thephase shifts are applied to the standardGray code mapping instead of usingthe redundant constellation mappingalready described.3 Proposed CFR MethodThe proposed CFR method does notalter the WiMAX (or other OFDM)standard nor does it require additionalinformation to be sent to the receiver.The accuracy of the null and pilotsubcarriers is preserved by restrictingconstellation errors to selected datasubcarriers. CFR modifications areapplied only to subcarriers whose dataelements are mapped onto the outerpositions of the constellation. For BPSKand QPSK, the outer elements includethe entire constellation. For 16-QAMand 64-QAM, there are 12 (of 16) and30 (of 64) outer elements, respectively.The outer elements have specialproperties in terms of constellationerrors and BER. These properties areexploited in the proposed CFR method.In a 16-QAM constellation, the IQmapping and decision boundariesused by the receiver define therelationship between constellationerrors and BER. The 16-QAM mappingis a Gray code (Fig. 2). Also shown inFig. 2 are the decision boundaries ofthe receiver. It is of interest to determinea) the largest constellation error thatcan be tolerated without causing a biterror, and b) the largest constellationerror caused by a single bit reversal.The number of bit reversals betweenDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 27D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P31

Special TopicR. Neil BraithwaiteCrest Factor Reduction for OFDM Using Selective Subcarrier DegradationDecisionBoundaryΔ1101 1001 1000 11000101 0001 0000 01000111 0011 0010 01101111 1011 1010 1110Hamming Distance =1Constellation Error=Δ▲Figure 2. 16-QAM constellation with Gray code mappingand decision boundaries. Outer elements can havedifferent constellation errors for Hamming distance = 1.the actual and received elements isreferred to as the Hamming distance.The largest constellation error thatcan be tolerated without causing a biterror depends on the position of thedata element. A single bit error occurswhen the constellation error causes thereceived element to cross one of thehorizontal or vertical boundaries. Thedistance between an interior elementand the closest boundary is Δ/2;however, this allowable error, whichincludes additive noise, is sharedbetween the transmitter, receiver, andpropagation channel. For the outerelements, there is at least one directionwhere no decision boundary exists: theoutward direction away from either theI axis or Q axis (depending on theposition within the constellation). At thecorner elements, there are twodirections unconstrained by decisionboundaries. Intentionally increasing theconstellation errors for outer elements inthese unconstrained directions doesnot increase the BER [6].To determine the largest constellationerror caused by a single bit reversal, weneed to look at the Gray code mapping.Neighboring elements in the horizontaland vertical directions have a Hammingdistance of one. If we assume theactual and received constellationelements differ by a Hamming distanceof one, the constellation error for aninterior point is Δ. For outer elements, aHamming distance of unity canproduce a constellation error of 3Δ(Fig. 2). Thus, large constellation errorscan be created from a single bitreversal on an outer element.Using the CFR method,constellation errors arecreated. The proposedmethod concentrates theconstellation errors on thedata subcarriers that producethe least amount of BER. Thatis, the goal is to generate asmuch constellation error asnecessary for the CFR whilecreating the minimumHamming distance betweenthe actual and receivedelements. Data subcarrierswith outer constellationelements are ideal for CFR.Two CFR methods are possible:outward enhancement and bit reversal.The former increases the I- orQ-component magnitude for all datasubcarriers that have an outer elementand negative correlation to the peak.The latter reverses the sign of the I- orQ-component for subcarriers that havean outer element and large positivecorrelation to the peak. Because signreversal causes a bit error, it is usedmore sparingly than outwardenhancement.As well as reducing the peak value, itis also important not to significantlyincrease the secondary peaks in thetime block. To avoid enhancingsecondary peaks, certain subcarrierscannot be used for peak reduction.Only subcarriers with a negativecorrelation to the primary peak and allsecondary peaks are used for theoutward enhancements. Onlysubcarriers with a positive correlation tothe primary and secondary peaks areHammingDistance =1ConstellationError=3ΔModulator:BPSK, QPSK,16-QAM, or64-QAMSerialtoParallelBPSK: binary phase-shift keyingCFR: crest factor reductionY (k)▲Figure 3. CFR Schematic [3].IFFTIFFTBitReversalFindpeaksCFRy (n )OutwardEnhanceIFFT: inverse fast Fourier transformPAPR: peak-to-average power ratioconsidered for bit reversal. The numberof secondary peaks specified must belimited to avoid eliminating too manysubcarriers from the CFR process.The CFR method is shown in Fig. 3.The OFDM system in Fig. 1 is modifiedso that the PAPR is measured in thetime domain, after the IFFT and beforethe addition of the cyclic prefix. If thePAPR is small enough (less than 8.5 dBfor example), the original OFDM datablock is used for transmission. If thePAPR is too large, the CFR OFDM datablock is used. Before CFR is applied,the primary and secondary peaks areidentified within the time block. TheCFR module reduces the primary peakand does not increase the secondarypeaks. CFR is applied in the Fourierdomain. The CFR signal is thenconverted to a time block using an IFFT.The CFR OFDM data block is notcomputed when the PAPR of theoriginal signal y (n ) is below 8.5 dB.The CFR module is shown in greaterdetail in Fig. 4. CFR uses the phasealignment profile described by (1) foreach of the primary and secondarypeaks. The phase alignment profile iscross-correlated with the real andimaginary components of thesubcarriers containing outer elements.The cross correlations for tpeak areC Re (k ;t peak) = Re{Y (k )}·cos{θ align(t_peak)(k )} (6)C Im (k ;t peak) = Im{Y (k )}·sin{θ align(t_peak)(k )}. (7)The peak is formed by the sum ofmany subcarriers. Because of thephase term θ (t peak) in (1), subcarrierswith positive cross-correlationscontribute to the peak whereas thosewith negative cross-correlationsMeasurePAPR of256 block (a)Use (a) ifPAPR

Crest Factor Reduction for OFDM Using Selective Subcarrier DegradationSpecial TopicR. Neil BraithwaiteMod.StoPFind datasubcarriers withouter elementsSeparate realand imaginarycomponentsY (k) y (n )Find PositiveCSCIFFTFindpeaksComputephase profilefor peak tp (1)tp (1) CrossCorrelationFindNegativeCSCIntersection:CSC Set 1FindPositiveCSCFind N largesttp (1) corr. fromCSC Set 1CFR: crest factor reduction CSC: correlated subcarrier components IFFT: inverse fast Fourier transform▲Figure 4. CFR for the case of N bit reversals per OFDM symbol plus outward enhancement [3].attenuate the peak. Applying a bitreversal, that is, changing the sign ofeither Re{Y(k)} or Im{Y(k )}) to asubcarrier, reverses thecross-correlation. A positive tonegative change reduces the peak.Increasing the magnitude of eitherRe{Y(k )} or Im{Y(k )} for a subcarrierpossessing a negativecross-correlation also reduces thepeak. Thus, a coordinated effort tocreate or enhance the negativecross-correlation on many subcarriersresults in effective CFR.There is a risk that a secondary peakwill increase in response to the CFR ofthe primary peak. It would be poor useof bit reversals and outwardenhancements if the CFR transformed asecondary peak into a primary peak. Toprevent this, the cross-correlations arecomputed relative to the secondarypeaks as well. The intersection of thesets of positively-correlatedcomponents for each peak is used as apool of available subcarriers for a bitreversal. The available subcarrierpossessing the largest positivecorrelation to the peak is selected. Theintersection of theComputephase profilefor secondarypeak tp (2)tp (2) CrossCorrelationFindNegativeCSCApply NbitreversalsComputephase profilefor secondarypeak tp (3)tp (3) CrossCorrelationFindPositiveCSCIntersection:CSC Set 2FindNegativeCSCApply outwardenhancement toall of Set 2negatively-correlated components foreach peak is also computed. Theoutward enhancement is applied to allthe available subcarrier componentsfrom the negatively-correlated set. Theenhancement is a scalar multiple of theoriginal value, for example, 1.05Re{Y(k )} or 1.05 Im{Y(k )}.In this approach, a secondary peakhas a magnitude that is above athreshold defined as a fraction of theprimary peak. The fraction isdetermined by the amount of peakreduction sought from the CFR. Thepotential increase in the secondarypeak is directly related to the decreasein the primary peak. Currently, thetarget PAPR is set to 8.5 dB. When theoriginal PAPR is greater than 8.5, 9.2, or9.7 dB, the fractional thresholds forsecondary peaks are 0.85, 0.8, and0.75 of the primary peak, respectively.These thresholds were obtained byexperimentation.Selecting too many secondary peakscan be problematic because theintersection of the correlated subcarriercomponent sets (Fig. 4) may become anull set, preventing any CFR. To avoidthis problem, the fractional threshold for…CFRN·bitReversals,OutwardEnhancementIFFTsecondary peaks is raised, ifnecessary, until the number of selectedpeaks is three or less. For these rareoccurrences, the number of bitreversals is reduced to avoidexcessively enhancing the secondarypeaks, but this comes at the expense ofincreased PAPR. Thus, to limit thenumber of secondary peaksconsidered, it is sometimes necessaryto increase the target PAPR for a givenblock, even if this results in clipping atthe PA.Because a bit reversal typicallyreduces the peak by about 0.4 dB, it isnecessary to specify additional bitreversals for large peaks. The numberof bit reversals for an OFDM symbol is1, 2, 3, or 4 when the original PAPRexceeds 8.7, 9.2, 9.7, and 10.1 dB,respectively. There are two approachesto implementing N bit reversals whenN > 1. Either all N bit reversals areapplied at once (Fig. 4), or single-bitreversals are applied recursively Ntimes (Fig. 5). For N = 0(PAPR < 8.7 dB), only outwardenhancement is used.Single bit reversal applied recursivelyN times is shown in Fig. 5. Thisapproach requires additional IFFTs tobe computed. Because the primary andsecondary peaks are re-computedafter each bit reversal, the fractionalthreshold for the secondary peaks is setto 0.85. Outward expansion is appliedafter the last bit reversal has beencompleted.In addition to bit reversals andoutward expansions, a third CFRapproach can be used. In thisapproach, constellation error isdistributed over all elements, not justthe outer elements. The phase profilesfor the primary and secondary peaksare multiplied by a scalar term and thenadded to the Fourier coefficients. Thisintroduces constellation errors similar toclipping, except that the affectedsubcarriers are selected, which allowsthe pilot and null subcarriers to betransmitted without error. Because thephase profiles of the primary andsecondary peaks have already beencomputed (Fig. 4), the additionalcomputational cost is minimal. The sizeof the scalar term controls theDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 29D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P33

Special TopicR. Neil BraithwaiteCrest Factor Reduction for OFDM Using Selective Subcarrier DegradationModulator:BPSK, QPSK,16-QAM, or64-QAMN bit reversalsStoPBPSK: binary phase-shift keyingCFR: crest factor reductionY (k) y (n )IFFTCFR1-bitReversalFindpeaksCFR1-bitReversal▲Figure 5. CFR for N bit reversals per OFDM symbol plus outward enhancement usingthe recursive implementation [3].constellation error in this third part ofthe CFR, and the scalar term is typicallyselected to be small in value. That is,this is not the primary source of peakreduction. In the proposed CFRmethod, the third approach is used toassist with reducing the largest peaks.The scale term is made to be anincreasing function of the PAPR of theOFDM symbol.In the CFR approach shown in Fig. 4,the cross-correlation is computed forall subcarrier components associatedwith the primary and secondary peaks.If only bit reversals are used for CFR,the cross-correlation for the secondarypeak need only be computed for thesubcarrier components associated withthe primary peak (Fig. 6). This can bethought of as a serial implementation ofa correlated subcarrier componentsearch (Fig. 4), which requires fewercomputations on average because theset of available subcarrier componentsbecomes smaller as subsequentsecondary peaks are tested.4 ResultsThe CFR approach in Fig. 4 isapplied to a WiMAX DL subframecomprising two QSPK symbols for theIFFTFindpeaksBitReversalMeasurePAPR…IFFTFindPeaksCFRMeasurePAPROutwardEnhanceIFFT: inverse fast Fourier transformQAM: quadrature amplitude modulation(0)(1)MeasurePAPR…Use up to Nbit reversalsto achievePAPR

Crest Factor Reduction for OFDM Using Selective Subcarrier DegradationSpecial TopicR. Neil Braithwaite-0.5-1.0-1.5-2.0-2.5β-3.0-3.5-4.0-4.5-5.0234CFRCCDF5 6 7PAPR (dB)CCDF: complementary cumulative distribution functionCFR: crest factor reductionPAPR: peak-to-average power ratio▲Figure 7. CCDF for the original and crestfactor reduced WiMAX time waveforms.The vertical axis represents the 10 βprobability that the signal power |y (n)| 2exceeds the level P o = PAPR*P ave, whereP ave is the average signal power.after filtering. In practice, however, thePAPR is defined using the digital timedomain signal y (n) and is based on thesignal power statistics rather than theabsolute peak. The practical peak is thelevel P o , for which the signal powery (n) 2 has a 10 β probability ofexceeding this peak. Thecomplementary cumulative distributionfunction (CCDF) of y (n) 2, in which β isplotted as a function of P o, is a usefuldescription of the signal. In this paper,two probability thresholds are selectedfor defining the signal peak: β= -4and -5.The CCDFs of the original andcrest-factor-reduced OFDM timesequences are shown in Fig. 7. Usingthe 10 -4 probability threshold for theCFR OFDM time sequence, PAPR is8.42 dB, which is a reduction of 1.2 dBfrom the 9.62 dB PAPR of the originalsignal. Using the 10 -5 probabilitythreshold, PAPR of the CFR and originalOFDM time sequences are 8.56 dB and10.54 dB, respectively. That is, CFRreduces PAPR by 2 dB when the 10 -5threshold is used. Thecrest-factor-reduced waveform meetsthe target PAPR of 8.5 dB for the 10 -4probability threshold, which is usedmore commonly in PA design than the10 -5 threshold.The BER introduced by the CFR is0.00019 (327 bit reversals from89OriginalCCDF10111730460 bits sent). The EVM, includingthe bit reversals, is 0.0725, which ishigh for 64-QAM. The limit for 3/4 rate64-QAM is 0.0282. However, whenusing the easier measure that excludesthe contribution from bit reversals, EVMis 0.0119. This EVM level is consideredacceptable. Most of the EVM for thissecond measure is due to the outwardexpansion of the constellation, whichdoes not increase BER at the receiver.This approach does not involve tryingto achieve the lowest CFR. Instead, abounded CFR is created to ease thedesign of the PA and digital circuitryand to generate a low BER. With thistype of CFR, 90% of the CFR symbolsare transmitted without modification.For such pass-through cases, only thePAPR measurement is required, and noadditional IFFTs are computed.5 ConclusionA CFR method that is suitable forWiMAX and other OFDM signals hasbeen presented. Degradationassociated with CFR is restricted toselected data subcarriers whose dataelements are mapped onto the outerboundary of the IQ constellation.Subcarrier components correlated tothe phase profiles of the primary andsecondary peaks are identified formodification by outward expansionsand bit reversals. Thepeak-to-average power for the WiMAXsignal is reduced to a target level, andthis allows the PA in the transmitter tobe designed for high PAE. CFR isachieved with an acceptable amount ofin-band BER and EVM degradation.References[1] IEEE 802.16-2009, IEEE Standard for Local andMetropolitan Area Networks Part 16: Air Interface forBroadband Wireless Access Systems,P802.16Rev2/D9, Jan. 2009.[2] S. C. Cripps, RF Power Amplifiers for WirelessCommunications, Norwood, MA: Artech House,1999.[3] R. N. Braithwaite,“Crest factor reduction systemand method for OFDM transmission systems usingselective sub-carrier degradation,”US patentapplication 20070140367, June 21, 2007.[4] M. J. Hunton,“System and method for peak powerreduction in spread spectrum communicationssystems,”US patent 6449362, Sept. 10, 2002.[5] M. J. Hunton,“System and method for post filteringpeak power reduction in multi-carriercommunications systems,”US patent 7,095,798,Aug. 22, 2006.[6] G. A. Awater, R. D. de Wild, A. Hendrik,“Transmission system and method employing peakcancellation to reduce the peak-to-average powerratio,”US patent 6,175,551, Jan. 16, 2001.[7] M. J. Hunton,“OFDM communications systememploying crest factor reduction with ISI control,”US patent application 20070058743, March 15,2007.[8] B. S. Krongold and D. L. Jones,“PAR reduction inOFDM via active constellation extension,”IEEETrans. Broadcasting, vol. 49, no. 3, pp. 258-268,Sept. 2003.[9] S. H. Han and J. H. Lee,“PAPR reduction of OFDMsignals using a reduced complexity PTStechnique,”IEEE Signal Proc. Lett., vol. 10, no. 11,pp. 887-890, Nov. 2004.[10] L. J. Cimini, Jr. and N. R. Sollenberger,“OFDMcommunication system and method having areduced peak-to-average power ratio,”USpatent 6928084, Aug. 9, 2005.[11] A. Alavi and C. Tellambura,“PAPR reduction ofOFDM signals using partial transmit sequence: anoptimal approach using sphere decoding,”IEEECommun. Lett., vol. 9, no. 11, pp. 982-984, Nov.2005.[12] C.-L. Wang and Y. Ouyang,“Low-complexityselected mapping schemes for peak-to-averagepower ratio reduction in OFDM systems,”IEEETrans. Signal Processing, vol. 53, no. 12, pp.4652-4660, Dec. 2004.[13] R. N. Braithwaite,“Crest factor reduction fordown-link LTE by transmitting phase shiftedresource blocks without side information,”European Wireless Conf., EuWIT 2009, Rome, Italy,pp. 13-16.[14] S. Hosokawa, S. Ohno, K. D. Teo, and T.Hinamoto,“Pilot tone design for peak-to-averagepower ratio reduction in OFDM,”IEEE Int. Sym.Circ. Syst. (ISCAS 2005), Kobe, Japan, pp.6014-6017.[15] L. Guan and A. Zhu,“Gaussian pulse-basedtwo-threshold parallel scaling tone reservation forPAPR reduction of OFDM signals,’Int. J. DigitalMultimedia Broadcasting, vol. 2011, Article ID470310, 9 pages, 2011. doi:10.1155/2011/470310.[16] Y. J. Kou, W.-S. Lu, and A. Antoniou,“Peak-to-average power ratio reductionalgorithms for OFDM systems via constellationextension,”IEEE Int. Sym. Circ. Syst. (ISCAS2005), Kobe, Japan, pp. 2615-2618.[17] R. N. Braithwaite,“General principles and designoverview of digital predistortion,”chapter in DigitalProcessing for Front End in WirelessCommunication and Broadcasting, F. Luo (Ed.),Cambridge Univ. Press, 2011.BiographyR. Neil Braithwaite (nbraithwaite@pwav.com)received his B.Sc. degree in electrical engineeringfrom the University of Calgary in 1985. He receivedhis M.Sc. and Ph.D. degrees from the University ofBritish Columbia in 1989 and 1992. From 1992 to1995, he conducted postdoctoral research at theUniversity of California, Riverside. From 1985 to1987 and 1995 to 2002, he worked for ComputingDevices Company (Canada), Nortel (Canada), andAgilent Laboratories (USA). Since 2002, he hasbeen working for Powerwave Technologies (USA).He is the author of several papers and patents, aswell as a recent book chapter on digitalpredistortion in RF power amplifiers [17].December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 31D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P35

Special TopicAn Antenna Diversity Scheme for Digital Front-End with OFDM TechnologyFa-Long Luo, Ward Williams, and Bruce GladstoneAn Antenna Diversity Scheme forDigital Front-End with OFDMTechnologyFa-Long Luo, Ward Williams, and Bruce Gladstone(Element CXI, San Jose, CA 95131, USA)Abstract: In this paper, we propose a new antenna diversity scheme for OFDM-based wireless communication and digital broadcastingapplications. Compared with existing schemes, such as post-fast Fourier transform (FFT), pre-FFT, and polyphase-based filter-bank,the proposed scheme performs optimally and has very low computational complexity. It offers a better compromise betweenperformance, power consumption, and complexity in real-time implementation of the receivers of broadband communication and digitalbroadcasting systems.Keywords: OFDM; digital front-end; MIMO; cross-layer processing; diversity; antenna1 IntroductionIn modern wireless broadbandcommunication and digitalbroadcasting, orthogonalfrequency-division multiplexing(OFDM)-based modulation schemesare usually used. However, an OFDMsystem has very poor reception whenthere is noise, interference, or movingobjects. To solve this problem, manytechnologies have been proposed andused in real applications [1], [2].Among these technologies,beamforming-based diversitytechnology is the most promising. Ituses multiple antennas at the receiverside and spatial filtering to optimizereception in noisy and mobileenvironments. Figs. 1, 2, and 3 showthree representative solutions: pre-fastFourier transform (pre-FFT), polyphasefilter-bank, and post-FFT [1], [4], [5],[7], [8].In Fig. 1, X 1(n), X 2(n), ... , X M(n) arethe received signals in one of Mantennas, and W 1(n), W 2(n), ... , W M(n)are the weights applied to theseantenna signals. All the weightedsignals are summed to one channel,such as one antenna’s output and thenforwarded for further FFT processing,which is required in any OFDM-basedreceiver. These weights are designedto meet an optimization criterion, andthe most common criterion is themaximum ratio combination (MRC).With MRC, the optimized weight vector,W (n) = [W 1(n), W 2(n), ..., W M(n)], canbe obtained byW opt(n) = E [X (n)×Y (n)] (1)Figure 2.▶Polyphase bankscheme.x 1(n )AnalysisAnalysisx 2(n )…AnalysisxM(n )where X (n) = [X 1(n), X 2(n), ..., X M(n)]and y (n) =W H (n)X (n). Pre-FFT cangive 5-10% gain in bit-error reductionx 1(n )x 2(n )xM(n )w 1(n )w 2(n )…wM(n )FFT: fast Fourier transform▲Figure 1. Pre-FFT scheme [1].FilterBank 1FilterBank 2…FilterBank L+y (n )Synthesis y (n )FFTFFTZ(1)Z(2)Z(N)Z(1)Z(2)Z(N)FFT: fast Fourier transform32ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P36

An Antenna Diversity Scheme for Digital Front-End with OFDM TechnologySpecial TopicFa-Long Luo, Ward Williams, and Bruce Gladstonex 1(n )x 2(n )xM(n )Multiplications with Weights forDifferent Frequency BinsFFTFFT…FFT▲Figure 3. Post-FFT scheme [1].and is not complex in implementation.However, this processing is done onlyin the time domain, and the sameweights are used over the entirefrequency band (all-carriers). Becauseof the shortcomings of simple pre-FFT,a polyphase filter bank scheme isproposed (Fig. 2).In a polyphase scheme, eachantenna signal is divided into a numberof frequency band signals, and differentweights apply to different bands. Thisscheme can further improveperformance at the cost of increasedcomputational complexity that arises asa result of multichannel antenna signaldecomposition (analysis) andmultiplication of weights in eachchannel. Different weights in differentfrequency bins and bands can be usedin the frequency domain, and this iscalled post-FFT (Fig. 3).A polyphase scheme is a specialcase of post-FFT. Post-FFT performsthe best but has the greatestcomputational complexity. Pre-FFT hasthe least computational complexity butimproves performance the least.Another approach is to divide Mantennas into L groups. Each groupuses pre-FFT and obtains L outputs.With these L outputs, post-FFT is thenused to obtain the desired outputs.Compared with the full post-FFTscheme, this alternative avoids M-LFFT computations at the cost ofreduced performance. A diversity+++Z(1)Z(2)Z(N)FFT: fast Fourier transformscheme that provides abetter compromisebetween performance,complexity, and powerconsumption is highlydesired.2 The ProposedSchemeHere, we propose anew scheme in whichthe number of variablescomputed is reducedfrom N ×M (in post-FFT)to N+M. The schemeshown in Fig. 4 performsjust as well as thepost-FFT scheme.M antenna signals are weighted bythe corresponding weight as theprocessing made in the pre-FFTalgorithm in Fig. 1. These M weightedsignals are added together to passthrough an N-tap finite impulseresponse (FIR) eigenfilter, after whichthey are input for FFT processing. Thenumber of unknown variables in theproposed scheme is N +M, which issignificantly less than the N ×M neededin post-FFT. Furthermore, only one FFToperation is needed in the proposedscheme instead of M FFT operations inthe post-FFT scheme. Here we willprove that the post-FFT scheme in Fig.3 and the proposed scheme in Fig. 4give the same frequency-domainoutputs, Z(1), Z(2),...,Z(N). Using theoptimization criterion in post-FFT, thevariable matrix, W N ×M, in Fig. 3 can berewritten asW N ×M = [W N,1, W N,2, ..., W N,M]== W C [C 1, C 2, ..., C M] (2)where W N,i is the i th column of theN ×M matrix and denotes theweights after FFT processing of thei th antenna in Fig. 3. W C is anN-dimensional vector, and C i is ascalar (corresponding to eachantenna). This suggests thatFigs. 3 and 5 give the same outputs.Because each block in Fig. 5 islinear processing, changing theorder of the blocks results in thesame performance. Hence, thescheme in Fig. 4 is obtained.Now we determine the N +M weightsin proposed scheme.W E, X h, and X t denote the weightvector of the eigenfiltering, thesequence of the head-guided interval,and the tail of the symbols,respectively. If the norm is a constant(unity, for instance), the weight vector isobtained in such a way that the error isminimized. That is,minW E E‖W H X h - W H X t‖ 2St.‖W H W‖2 = 1. (3)According to beamforming matrixtheory [3], the solution to theoptimization problem is the eigenvector(minor component) corresponding tothe smallest eigenvalue of thecorrelation matrix R, which isE [(X h -X t) (X h -X t) H ]. (4)This is also the reason that thefiltering in Fig. 4 is called eigenfiltering.In practical implementation, the minorcomponent in the following adaptivealgorithm can be updated:W (t+1)=W(t)+γs (t )(X h-X t-s(t )W (t))s (t )=W H (t )(X h-X t) (5)where γ is a constant. This algorithmis as complex as LMS algorithm, andthe required multiplications are only onthe order of N. An algorithm, such asMRC used in pre-FFT, can be used todetermine the weights applied to eachantenna before eigenfiltering.3 Comparison andDiscussionCompared with existing solutions, thex 1(n )x 2(n )xM(n )w 1(n )w 2(n )…wM(n )FIREigenfiltering▲Figure 4. Proposed scheme.FFTZ(1)Z(2)Z(N)FFT: fast Fourier transformFIR: finite impulse responseDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 33D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P37

Special TopicAn Antenna Diversity Scheme for Digital Front-End with OFDM TechnologyFa-Long Luo, Ward Williams, and Bruce Gladstonex 1(n )x 2(n )xM(n )FFTFFT…FFT▲Figure 5. Alternative scheme for post-processing.proposed scheme has the followingfeatures:•The parameters to be computed ina real-time implementation arereduced from M ×N to N +M, andperformance remains the same as thatof post-FFT.•Only one FFT operation is neededinstead of multiple FFT computations orpolyphase filtering analyses. Table 1shows computational complexity forvarious standards.•The related parameters can beadaptively updated from the receivedsamples of the each antenna. Theadaptive algorithm used is as simple asthe LMS algorithm, which has thecomputational complexity on the orderof the number of unknown parameters.With these features, the proposedscheme is a new tool for improvingreception quality in broadband wirelesscommunications and digitalbroadcasting. It will have very manypractical uses in devices.4 Conclusions and FutureWorkIn this paper, we have proposed anew and efficient diversity scheme forOFDM-based receivers. We have alsoshown the accuracy and effectivenessof the proposed scheme. Whenembedding this algorithm into realsilicon, it is often desirable that oneC 1C 2CMWcVectorMultiplicationSummationFFT: fast Fourier transformplatform support allexisting standards (andtheir various modes) withhigh power efficiency, lowcost, and short time tomarket [9]. Traditionalintegrated circuittechnologies, such asapplication-specificintegrated circuits(ASICs) and digital signalprocessors (DSPs), arenot highly flexible andpower efficient. An ASICsolution gives highperformance with lowpower consumption andprice, but it cannotsupport multiplestandards nor is itsufficiently flexible. A DSPis highly flexible but performs poorlyand consumes much power. It mayinclude high-speed arithmeticoperations, such asmultiply-accumulate, but thealgorithms require extensiveprogramming and more parallelismthan can be offered by a general DSP.One alternative is to use a generalDSP or reduced instruction setcomputing (RISC) processor plushardware accelerators that aredesigned and optimized to implementFFT algorithm as well as eigenfilteringand its adaptive algorithm. In otherwords, an eigenfiltering unit performs abasic filter computation, an FFT unitperforms FFT computations, and anadaptive unit implements theoperations defined by (5) in section 2.These optimized accelerators maymeet performance goals, but theaccelerators are very narrow in theirapplicability, and this significantlyreduces the flexibility of theprocessor-based solution.An alternative that offers flexibilityand parallelism is devices with FPGAs.These devices may combineprocessors with a programmable arrayof low-level logic devices, but they areexpensive, and performance is limitedat high temperatures.In a future paper, we will introduce➡To P.41▼Table 1. Parameters related to implementation of FFT in different standardsStandardFFT SizeDuration(μs)Single FFTCycles (k)Z(1)Z(2)Z(N)SymbolNumber (K)MIPSISDB-T256512102420484096819225250410082.565.7612.828.261.4133.1242110.2411.5212.8123.0425.6028.20112.80132.80133.12DVB-H20484096819222444889628.261.4133.14.52.21.1127135146DVB-T2048819222489628.2133.14.5CMMB: China Mobile multimedia broadcastingDTMB: digital terrestrial multimedia broadcastDVB-H: digital video broadcasting-handheldDVB-T: digital video broadcasting-terrestrialFFT: fast Fourier transform1.1127146FLO4096738.0261.41.483.2CMMB10244096409.612.861.42.430.72147.00DTMB1024409650012.861.4225.6132.8T-DMB256512102420482502.565.7612.828.2FLO: forward link onlyISDB-T: integrated services digital broadcasting-terrestrialMIPS: Million Instructions Per SecondT-DMB: digital multimedia broadcasting via terrestrial410.2423.0451.20112.2WiMAX(Mobile)1285121024204891.41.125.7612.828.21112.321263.36140.80310.20WiMAX(Fixed)256642.5615.6340.0134ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P38

A Histogram-Based Static-Error Correction Technique for Flash ADCsSpecial TopicArmin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul DehghaniA Histogram-Based Static-ErrorCorrection Technique for Flash ADCsArmin Jalili 1 , J Jacob Wikner 2 , Sayed Masoud Sayedi 1 , and Rasoul Dehghani 1(1. Department of Electrical and Computer Engineering, Isfahan University of Technology, Isfahan 84156-83111, Iran;2. Department of Electrical Engineering, Linköping University, SE-581 83 Linköping, Sweden)Abstract: High-speed, high-accuracy data converters are attractive for use in most RF applications. Such converters allow directconversion to occur between the digital baseband and the antenna. However, high speed and high accuracy make the analogcomponents in a converter more complex, and this complexity causes more power to be dissipated than if a traditional approach weretaken. A static calibration technique for flash analog-to-digital converters (ADCs) is discussed in this paper. The calibration is based onhistogram test methods, and equivalent errors in the flash ADC comparators are estimated in the digital domain without any significantchanges being made to the ADC comparators. In the trimming process, reference voltages are adjusted to compensate for static errors.Behavioral-level simulations of a moderate-resolution 8-bit flash ADC show that, for typical errors, ADC performance is considerablyimproved by the proposed technique. As a result of calibration, the differential nonlinearities (DNLs) are reduced on average from 4 LSBto 0.5 LSB, and the integral nonlinearities (INLs) are reduced on average from 4.2 LSB to 0.35 LSB. Implementation issues for thisproposed technique are discussed in our subsequent paper,“A Histogram-Based Static-Error Correction Technique for Flash ADCs:Implementation Aspects.”Keywords: Calibration; flash ADC; offset; trimming; uniform distribution1 IntroductionAn analog-to-digital converter(ADC) is an essential part of anRF receiver, but in manysystems, an ADC limitsperformance.To relax the dynamic rangerequirements of an ADC, a voltage gainamplifier (VGA) and filters (foradjacent-channel blocking) are usedbefore the ADC. Typically, linear analogfilters with sharp transition bands andgood VGA designs are used. Suchfilters are difficult to implement,especially in an integrated system andin newer CMOS technologies. Thedesign of these blocks should besimplified, and more should berequired of the ADCs. In today’scellular applications, the mainchallenge is to reduce analogcomplexity as much as possible andshift it to the digital part of the systemthat relies on digital signal processing(DSP). Analog complexity is reduced byminimizing the number of (mostly)discrete elements, such as high-Qfilters, and lowering the number ofanalog cascaded stages indown-conversion. Reduction in analogcomplexity allows the system to bemore integrated and makes it morecompatible with a system-on-chip(SOC) design.An important factor in ADC design isthe power-accuracy trade-off. This ismore critical in cellular applications,where portability significantly affectspower consumption, especially that ofthe converter. A highly dynamic rangein the converter is sometimesachievable at the cost of high powerconsumption. Calibration optimizes thepower-accuracy trade-off, and thedesired accuracy can be obtained withreduced power consumption. Byreducing analog complexity andincreasing digital complexity,calibration makes the converter moreconducive to process-scaling andmore compatible with newer CMOStechnologies. However, this comes atthe cost of increased digital complexitybecause a larger part of the ADC isdigital rather than analog.In this paper, we focus on flash ADCswith low to moderate resolutions, that is,3 to 8 bits. These resolutions are usedextensively as the sub-ADC part ofsigma-delta ADCs, in pipelined ADCs,or in applications ultrawideband wherelow resolution and very high-speedconverters are needed. In a standardDCS-1800 for a low-IF receiver, ahigh-resolution, low-bandwidthconverter is required. Such a convertercan be realized using a sigma-deltaADC.The accuracy of a flash ADC ismainly affected by the offsets of itscomparators and references. Toimprove accuracy, components that arewell-designed and comparatively largein terms of area and power areDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 35D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P39

Special TopicA Histogram-Based Static-Error Correction Technique for Flash ADCsArmin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul Dehghanicommonly used. This results in highercost and, perhaps more importantly,longer design time and limitedprocess-scaling if the ADC isembedded in a system on chip (SOC).In addition to the material presentedin this paper, we give a more detailedanalysis of the histogram-basedstatic-error correction technique in [1],taking implementation issues intoaccount. We study thetransistor-implementation of a 5-bit,1 GHz ADC in a 1.2 V, 65 nm CMOSprocess, and we evaluate theimprovement in performance broughtabout by the calibration technique. Wealso suggest further improvements tothe analog design so that analogcomplexity is further reduced in favor ofdigital complexity.1.1 CalibrationCalibration techniques for flash ADCshave been reported in [2]-[6]. In mostof these techniques [2]-[4], [6], theanalog part — especially thehigh-speed analog signal path — ofthe flash converter is affected by thecalibration circuitry. In some of thesetechniques [3], [4], offset compensation(trimming) is performed inside thecomparator structures. In [3], trimmingis performed by adjusting the internalresistive loads of the comparatorpreamplifiers. However, these loads areoften difficult to trim accurately becauseof their high gain and nonlinearity, andtrimming might be inefficient. In [2] and[6], linear trimming is performedefficiently by adjusting the voltagesgenerated by the reference ladder. In[2], an accurate digital-to-analogconverter (DAC) generates a calibrationsignal, but the design of such a DACstill has a high degree of analogcomplexity. In [6], each comparator ischosen sequentially and calibrated.This affects the main ADC structure;specifically, extra calibration circuitryaffects the high-speed input signalpath. In [5], trimming is performed byadjusting the bulk voltages of thecomparator input transistors to keep thehigh-speed signal path intact.However, the trimming is nonlinearbecause of the nonlinear relationshipbetween the bulk and thresholdvoltages of the MOS transistor.1.2 Algorithm ExampleA histogram-based calibrationtechnique for flash ADCs is an exampleof an error-correction technique forfront-end ADCs. The theory ofhistogram test methods is well-known[7]-[13], and in this paper, a histogrammethod is used to extract the equivalentinput-referred offset of eachcomparator in the flash ADC. This is anestimation process. Trimming is thenperformed by adjusting the referencevoltage levels in order to compensatefor the static error sources. Theestimation process is fully digital, andno changes are made to thecomparator structure. Trimming isperformed by adding arrays of analogswitches that are connected to thereference taps of the comparators sothat the high-speed signal path of theconverter is not affected.To illustrate the calibration technique,behavioral-level simulations can beperformed. Individual error sources canthen be isolated, and their effects canbe analyzed separately. In [1], atransistor-level ADC that uses thecalibration technique is described, andpractical implementation issues arediscussed.In section 2, main sources of staticerror are modeled. In section 3, thecalibration technique is described, andin section 4, the estimation process isdescribed in mathematical detail. Insection 5, the trimming process andrelated practical issues are discussed.In section 6, simulation results are-VOS,i++V r,i•-given. Section 7 concludes the paper.2 Examples of Static ErrorsIn an n-bit flash ADC, 2 n -1 referencelevels are commonly generated by aresistor ladder. For each referencelevel, a comparator is used. Thecomparators compare the input signalwith the reference levels, andcombined, the comparators generate athermometer code at their outputs. Adigital decoder that filters out sparkleerrors is usually used. Static errorsaffect the accuracy of the decisionlevels in the converter. Errors in thereference levels are mainly caused byresistor mismatch and comparatoroffsets. These two errors can bemodeled (and combined) as a voltagesource, V os , at the input of eachcomparator, as shown in Fig. 1(a). For alow-resolution ADC, the matching ofthe resistors is usually much better thanthe nominal resolution of the flashconverter, so the reference voltagesfrom the resistor ladder can be treatedas ideal values [6]. However, this is notthe case for higher-resolution flashADCs.In Fig. 1(a), the comparator isassumed to be ideal, and the errorvoltage, V os, in practice would bedescribed by statistical variation overmany individuals. This is shown in Fig. 1(b), where V r,i is the ideal referencelevel of the ith comparator, and V os, i isthe corresponding offset voltage sourceof the ideal reference. The ith intervalbetween two adjacent reference levelsis denoted I i. In an ideal converter, the▲Figure 1. (a) Static errors with voltage sources, V os,i, and their effects for (b) smalland (c) large offsets.I iV r,iV r,i - VOS,iV r,i -1 - V OS, i -1V r,i -1(a) (b) (c)V r,iV r,i -1 - V OS, i -1V r,i - VOS,iV r,i -136ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P40

A Histogram-Based Static-Error Correction Technique for Flash ADCsSpecial TopicArmin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul DehghaniV inPDFGeneratori th interval is-VRf in (x )where i = 1, 2,..., 2 n , and the boundariesare given by V r,0=-V R andV r, = +V R so that the ADC range is 2V R.In practice, these intervals are changedby offset errors, as inFig. 1(b). In the case of large offsetvalues, I i may disappear, and anotherinterval may be generated. In practice,many of the reference levels aredisplaced from their ideal locations,and combined with the comparatoroffset errors, these reference levels mayoverlap.3 Calibration TechniqueTo calibrate the ADC, the errorvoltage sources, V os,i (Fig. 1), have to beestimated and compensated for. Duringestimation, the offset voltage or, as inmany cases, just its polarity, isdetermined [3], [5], [6]. In trimming, thederived offset value can becompensated for by using analogcircuit-level techniques.In the proposed technique,estimation and trimming are performedwithout the comparator structure beingmodified. This ensures that thehigh-speed input signal is notadversely affected. Estimation can beperformed without modifying theconverter structure by relying on theinput signal characteristics, as iscommonly done in histogram-basedtest methods.The basis of the proposed techniqueis shown in Fig. 2. The probabilitydensity functions (PDFs) of the inputand output signals of the converter areADC: analog-to-digital converterPDF: probability density functionV r,i -1≤ I i≤ V r,i (1)XVRXFlash ADC-VRf out(x )VRXf in(x ) and f out(x ), respectively. Thereference voltage of the ADC is V R, andconversion is only valid for inputs in therange -V R < V in < V R because the ADCwould otherwise saturate at either end.An ADC system operates in normal orcalibration mode. In normal mode, theinput signal, V in, is connected to theflash ADC input (node x in Fig. 2), andthe ADC converts normally. Incalibration mode, the output of the PDFgenerator is connected to the ADCinput. The PDF generator produces ananalog signal with a known probabilitydensity function, f in(x). The output PDF,f out(x ), is a quantized version of f in(x).The output PDF is a discrete function interms of I i. The input PDF, however, is acontinuous function. Because of staticerrors, the two PDFs are not equal. Inhistogram-basedcalibration, thedifference betweenthese two PDFs is usedto extract the ADCerrors.Fig. 3 shows an n-bitflash ADC that useshistogram-basedcalibration. Thecalibration units areindicated by the dashedline. An array ofmultiplexers (MUX) isused to select a set ofcomparators forcalibration. The outputsof the array areconnected to theinterval detector block.This block determineswhich intervals, I i(i = 1,2,...,16), the input◀Figure 2.A histogram-basedcalibration technique.V inVR-VRPDFGeneratorsamples belong to. The counter recordsthe number of samples for eachinterval. The ideal number of samplesfor each Ii is calculated because theinput PDF must be known withreasonable accuracy.The estimation block estimates theerror voltage source, V os,i, by using thedifference between N i and N i , where N iis the desired number of samplesbelonging to I i, and N i ( ~ denotes thenon-ideal case) is the measurednumber of samples belonging to I i. Theestimations are then used by thetrimming block to compensate for staticerrors. As described in section 5, thistrimming block compensates for ADCerror by adjusting the voltage taps ofthe resistor ladder. This adjustmentchanges the ADC reference levels,which are generated by the referenceselection block.4 Offset Error EstimationFor simplicity, the proposedcalibration technique is used in a flashADC with 16 decision levels, whichcorresponds to a resolution of n = 4bits. The calibration technique is thenextended to higher resolutions thathave greater hardware complexity. Inan IC implementation, extending theresolution limits the number of blocks,Reference SelectionT/H+-+-+-+-inMUX: multiplexers PDF: probability density function T/H: track and hold circuit▲Figure 3. An n -bit flash converter. Additional calibration unitsare indicated by the dashed line.…TrimmingCalibration UnitMUXArrayIntervalDetectorCounterEstimationDecoderDigitalOutput2 n 41December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 37D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P

Special TopicA Histogram-Based Static-Error Correction Technique for Flash ADCsArmin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul DehghaniI 1V bot =-V RV r, 1 V r, 1V r, 1 - VOS, 1(a)▲Figure 4. The effect of static error values oninterval I i for (a) small and (b) large errors.(which grows exponentially with ADCresolution) so that the critical area doesnot increase too much.4.1 Estimation for a Low-ResolutionADCIf the PDF circuit generates an analogsignal with a uniform distribution(constant PDF) in the range [-V R, V R],then for an ideal 4-bit converter, andfor every sufficiently large batch of Ninput samples, on averageN i = N/16 (2)samples in each interval i = 1, 2,..., 16.In a non-ideal 4-bit converter withinterval I i, as in Fig. 1(b), the expectednumber of samples is~ 2VR-Vos,i+ Vos,i -1V bot =-V RV r, 1 - VOS, 1N i =16×N. (3)2V RBy rearranging (3), a recursiveexpression is obtained:~N i-N iV os,i = V os,i -1 -2V R· . (4)NEquation (4) can now be used toestimate the offset values, assumingV os,0 = V os,16= 0. According to (4),comparators 1 and 15 give two initialestimated values:~V os,1 = -2VN R· 1-1(5)N 16and~V os,15 = -2VN R· 16-1N 16. (6)Either one of (5) or (6) is a sufficientstarting condition for the calibrationalgorithm. However, to minimizecomputational error accumulationduring error estimation, the offset errorsrelated to comparator 2 to 7 are(b)estimated using the initial condition (5),and offset errors related to comparators8 to 14 are estimated using initialcondition (6).For large offset values, as shown inFig. 1(c), I i may disappear, creating anegative interval. In the case of anoverlap, (4) must be modified so thatthe estimation can handle errors givenby~N i+N iV os,i = V os,i -1 +2V R· . (7)NThis modified version of (4) is usedinstead of (4). The case of small offsetvalues (V os,i) in comparator 1 is shown inFig. 4(a), where (5) is valid. For largeerror values, the reference level may bepushed below the bottom voltage -V R,as in Fig. 4(b). In this case, V os,1 isdecreased during trimming until thesituation shown in Fig. 4(a) is reached.Again, (5) is valid and can be used inthe estimation process. The samestrategy is used for the upper-endusing a modified version of (6).4.2 Extending to HigherResolutionsTo extend the low-resolutiontechnique to a general n-bit converter,a segmented structure can be used(Fig. 5). This structuresaves hardware forhigher-resolution ADCsbecause a substantial partof the hardware growslinearly rather thanexponentially. Thesebenefits are described inmore detail in [1].In the segmentedapproach, m = 2 n - 1comparators are dividedinto 16 = 2 4 subgroups(starting with the original4-bits). Each groupconsists of i = 2 n - 4comparators, except for thelast group, which containsi - 1 comparators. Sixteenmultiplexers are used tochoose the output of onecomparator from eachgroup and connect these tothe 16 -decision-levelestimation block. At anyVR-VRtime during the calibration, 16comparators are chosen and theiroffsets are estimated. The results arestored in a RAM. Then, another set of16 comparators is selected for theestimation. This procedure continuesuntil all offsets of all comparators havebeen estimated and stored in the RAM.Only one state, corresponding to astandard 4-bit flash ADC, containsi - 1 comparators. This is the statewhen the first input of each multiplexeris chosen.5 ADC Trimming ProcessWe take the trimming approachsuggested by Chen et al. in [6].Besides having single LSB steps for theideal reference voltages, the resistorladder also generates smaller steps,that is, fractions of an LSB to allow thereference voltages to be more finelyadjusted. This implies that trimming isdone without any changes being madeto the comparator.The approach is shown in Fig. 6. Anappropriate voltage tap is selected bythe trimming circuit in order tocompensate for the offset value. In theresistor ladder, the coarse tapsgenerate the ideal reference voltages,+-+-+-+-+-+-+-+-+-+-+-+-… …… ………'1'lMUX2 #11•lMUX2 #i1RAM•……lMUX2 #151lMUX2 #161……••Select timesMUX: multiplexers4-bit Estimation Block▲Figure 5. Extending 4-bit ADC calibration to an n -bitconverter with l = 2 n - 4 comparators in each group.Trimming38ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P42

……A Histogram-Based Static-Error Correction Technique for Flash ADCsSpecial TopicArmin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul Dehghani…Coarse Tap V r, i + 1 •+FineTapsV r, i + (M -1)δV r •V r, i + (M -2)δV r •Coarse Tap…V r, i + 2δV r•V r, i +δV r•V r, i •…▲Figure 6. Trimming a comparator by selecting taps from the resistor ladder.V i + 1V i •V i + 1……Reference ResistorLadderLatch•▲Figure 7. Kickback noise in a latchedcomparator.…V inV r,I. In addition, there are M resistorsbetween two consecutive coarse tapsthat generate M - 1 fine taps. Each finetap generates a voltage of V r,i +δV r ,whereV r,i +1-V r,iδV r = (8)Mis the fine-level step — a fraction of theoriginal coarse LSB step.Our aim is to create a calibrationprocess that does not adversely affectthe performance of the ADC operatingin normal mode. However, in practice,this is not possible. Any calibrationchanges or influences the ADCstructure, and this can affect the overallperformance.Because more switches are added tothe ADC’s design (Fig. 6), kickbacknoise, a well-known problem, emerges.In a clocked comparator, the largevariations on the regeneration nodes inthe latch circuit are fed back through a………parasitic capacitorto the sensitiveinputs of thecomparator. ThisCR idisturbs the inputand reference-voltages. A modelof kickback noiseis shown in Fig. 7for an ordinarydifferential-paircomparator inputstage.Trimming To simulate theeffects of kickbacknoise, a dynamiccomparator withrelatively largekickback is chosen. This comparator isused in a 6-bit flash ADC architecturewith a 0.5 V signal range, that is, V R=0.25 V. The number of sub-LSB stepsis M = 4, see (8). The amplitudedistribution for simulated kickback indifferent comparators is shown in Fig. 8.The effect of the kickback noisedepends on the reference level and theimpedance of the resistor ladder at aparticular node. If resistance betweenthe taps of the ladder is very low, thekickback is very small. The noiseamplitude is proportional to theimpedance of each tap of the resistorladder:R i = i -i 2·R (9)2 nwhere i = 1, 2,...,2 n - 1 is the tapnumber, and R is the unit resistance ofthe reference ladder. According to (9),a symmetrical shape can be expected.However, the amount of kickback noisealso depends on the input signal levelas well as the reference level createdby the voltage and capacitancedifferences between the gate and drainFigure 8. ▶Kickback noise fordifferent tap numbers.The solid line representsthe actual curve and thedotted line is thepredicted curve whenimpedance variationsare taken into account.Normalized Kickback Noise Amplitude1.00.80.60.40.20010of input transistors. In particular, thevalues of the parasitic capacitors, C gsand C gd , change. Therefore, simulatednoise is asymmetrical. In Fig. 8, wecompare the normalized kickback noiselevels for ideal and voltage-dependentcases.When extra switches are inserted forselecting finer reference levels, moredecoupling capacitance is, byimplication, added to different taps inthe resistor ladder. Adding trimmingcircuitry does affect kickback noise butin a positive way.An increased time constant is animportant problem when the resistorladder charges the comparator input.Using analog switches that are toosmall in order to save area makes theproblem worse. However, the timeconstant appears as a configurationtime, T conf, which is the time the systemtakes to settle after the trimmingcommand has been applied (Fig. 9).The perturbation is the effect of thekickback noise. The amplitude of thekickback noise is V n, and the residualkickback error, caused by the resistorladder’s long settling time in responseto kickback disturbance, is ΔV n .The additional analog switches arethe configuration switches, and their on/off state is changed at the end of eachcalibration cycle. During theconverter’s normal operation, theirstate is fixed. Depending on theresistance of the resistor ladder and thesize of the switches, the configurationtime differs from a few clock cycles toseveral. The design of the circuits mustinclude a time constant that isnegligible compared with the overallcalibration time. A more detailed pictureof the effect of kickback noise on thecalibration technique is given by thetransistor-level simulation in [1]. The2030Tap Number4050: Ideal Curve: Actual Curve60December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 39D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P43

Special TopicA Histogram-Based Static-Error Correction Technique for Flash ADCsArmin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul DehghaniV r, jV r, j~T conf▲Figure 9. Settling behavior of the resistorladder in response to changing fromreference level V r,i to V r, j (as a result ofinserting a trimming command) also tokickback noise.simulation uses a 65 nm CMOS with1.2 V power supply.6 Simulation ResultsTo determine the efficiency of theproposed calibration technique, abehavioral-level model of an 8-bitflash ADC using the segmentedcalibration technique was designedand simulated. A triangular signalgenerator (TSG) was used to realize auniform distribution. The effects of anunwanted, non-uniform distributioncaused by, for example, nonlinearityare discussed in [1]. The TSG, wasconfigured so that a triangularwaveform with a frequency of about100 kHz was obtained, and thesampling frequency of the ADC was amodest 400 MHz.The system design parameters andapplied-error sources are shown inTable 1. Each calibration cycle containsthe estimation of equivalent offsetvoltages for a set of 16 comparatorschosen by the multiplexer array. In thecase of interval overlap, one morecalibration cycle is required. Further on,two calibration cycles are required tofully compensate for the TSGcomparator offsets. Because an 8-bitconverter is being studied, there are 16sets of comparators to trim, and 64cycles are needed to fully estimate andtrim all the error sources in the ADCstructure.The estimation was implemented inregister-transfer level (RTL) andsynthesized on a Xilinx Spartan 3FGPA. The post-synthesis code wasused in the simulations, and 693 slices,…532 slice flip flops, and 1280 four-inputlookup tables were also used.One hundred Monte Carlo analyseswere run, and the performance of theADC before and after calibration wasmeasured. The static performance interms of DNL and INL is shown in Figs.10(a) and (b), respectively. For about98% of the Monte Carlo runs, theconverter’s performanceconsiderably improved. Aftercalibration, the DNL was reducedon average from 4 to 0.5 LSB, andthe INL was reduced on averagefrom 4.2 to 0.35 LSB. The failing2% were cases where the offsetwas large enough to end upoutside the correctable range.and trimming is performed by adjustingthe comparator reference voltagesthrough arrays of analog switches.A behavioral-level 8-bit flash ADCwas designed, and typical errors wereadded to the comparators, resistorladder, and PDF generator. The resultsof 100 Monte Carlo analysis show thatconverter performance was improvedΔV n100▼Table 1. Design parametersParameterADC ResolutionOffset Distribution (σos)Offset Distribution of TSG Comparators (σos,tsg)TSG NonlinearityEffective Resolution of Resistor LadderReference Voltage, VR7 ConclusionsNumber of Samples Required in Each Calibration Cycle, NThis paper describes acalibration technique for flashADCs that is based on a histogramtest method. The ADC iscalibrated without its structurebeing adversely affected.Estimation is performed digitally,Number of Calibration Cycles for a Complete CorrectionTotal Number of Samples per Calibration InstantNumber of Fine Voltage Steps, MWord Length in Digital PartTSG: triangular signal generator7: Before Calibration6: After Calibration5432100 10 20 30 40 50 60 70 80 90V nDNL (LSB)DNL (LSB)765432100102030Simulation Index40 50 60Simulation Index▲Figure 10. (a) DNL and (b) INL of the 8-bit flash ADC before and after calibration.(a)(b)7080: Before Calibration: After Calibration90100Value825 mV5 mV5 %7 bits1 V2 13642 19412 bits40ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P44

A Histogram-Based Static-Error Correction Technique for Flash ADCsSpecial TopicArmin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul Dehghaniconsiderably in 98% of the runs.This work can be used as agrounding for the study andimplementation of ADC calibrationtechniques. Prospects and limitations ofthe technique are also highlighted inthis work. An IC implementation of anADC using the proposed calibrationtechnique is detailed in [1].References[1] J. Wikner, A. Jalili, S. M. Sayedi, and R. Dehghani,“A histogram-based static error correctiontechnique for flash ADCs: Implementation aspects,”accepted for publication in ZTE Communications,Mar. 2012.[2] H. Yu and M.-C. F. Chang,“A 1 V 1.25 GS/s 8-bitself-calibrated flash ADC in 90 nm digital CMOS,”IEEE Trans. Circ.. Syst.. II, vol. 55, no. 7, pp.668-672, Jul. 2008.[3] C.-W. Lin Y.-Z. Lin and S.-J. Chang,“A 5-bit 3.2GS/s flash ADC with a digital offset calibrationscheme,”IEEE Trans. Very Large Scale Integration(VLSI) Syst., vol. 18, no.3, pp. 509-513, Mar. 2010.[4] V. Srinivas, S. Pavan, A. Lachhwani, and N.Sasidhar,“A distortion compensating flash analogto-digital conversion technique,”IEEE J. SolidState Circ., vol. 41, no. 9, pp. 1959-1969, Sep.2006.[5] Jin Liu Junjie Yao and Hoi Lee,“Bulk voltagetrimming offset calibration for high-speed flashADCs,”IEEE Trans. Circ. Syst. II, vol. 57, no. 2, pp.110-114, Feb. 2010.[6] Chun-Ying Chen; M. Q. Le and Kwang Young Kim,“A low power 6-bit flash ADC with referencevoltage and common-mode calibration,”IEEE J.Solid-State Circ., vol. 44, no. 4, pp. 1041-1046,Apr. 2009.[7] J. Larrabee, F.H. Irons and D.M. Hummels,“Usingsine wave histograms to estimate analog-to digitalconverter dynamic error functions,”IEEE Trans.Instrum. Meas., vol. 47, no. 6, pp. 1448-1456, 1998.[8] J. Doernberg, H.-S. Lee, and D.A. Hodges,“Full-speed testing of A/D converters,”IEEE J.Solid-State Circ., vol. 19, no. 6, pp. 820- 827, Dec.1984.[9] Y. Betrand, F. Azais, S. Bernard, and M. Renovell,“Towards an ADC BIST scheme using thehistogram test technique,”in IEEE Proc. EuropeanTest Workshop, Cascais, Portugal, May 2000, pp.53-58.[10] U. Eduri and F. Maloberti,“Online calibration of aNyquist-rate analog-to-digital converter usingoutput code-density histograms,”IEEE Trans.Circ. Syst. I, vol. 51, no. 1, pp. 15-24, Jan. 2004.Biographies[11] Degang Chen, Xin Dai and R. Geiger,“Acost-effective histogram test-based algorithm fordigital calibration of high precision pipelinedADCs,”in Proc. IEEE Int. Symp.Circ.Syst., Kobe,Japan, May 2005, pp. 4831-4834.[12] Degang Chen Le Jin and R. Geiger,“A digital selfcalibration algorithm for ADCs based on histogramtest using low-linearity input signals,”in IEEE Int.Symp. Circ. Syst., Kobe, Japan, May 2005, pp.1378-1381.[13] J. Elbornsson and J.-E. Eklund,“Histogram basedcorrection of matching errors in subranged ADC,”in Proc. 27th European Solid-State Circ. Conf.,Villach, Austria, 2001, pp. 555-558.Armin Jalili (arminj@ec.iut.ac.ir) received his B.Sc. and M.Sc. degrees in electrical engineering from IsfahanUniversity of Technology (IUT) in 2004 and 2006. He is currently working towards his Ph.D. degree in electricalengineering at IUT. His research interests include ADC design.J. Jacob Wikner (jacob.wikner@liu.se) received his Ph.D. from the Department of Electrical Engineering,Linköping University, Sweden, in 2001. He has worked as a research engineer at Ericsson Microelectronics, senioranalog design engineer at Infineon Technologies, and senior design engineer and chip architect at SiconSemiconductor. Dr. Wikner has been an associate professor at Linköping University since 2009. His researchinterests include biologically inspired architectures, high-speed ADC and DAC, and general analog andmixed-signal design. He holds six patents, has published 40 scientific papers, and has co-authored“CMOS DataConverters for Telecommunications.”He is the co-founder of CogniCatus and AnaCatum Design.Sayed Masoud Sayedi (m_sayedi@cc.iut.ac.ir) received his B.Sc. and M.Sc. degrees in electrical engineeringfrom Isfahan University of Technology (IUT), Iran, in 1986 and 1988. He received his Ph.D. degree in electronicsfrom Concordia University in 1996. From 1988 to 1992, and since 1997, he has worked at IUT, where he is currentlyan associate professor in the Department of Electrical and Computer Engineering. His research interests includeVLSI fabrication processes, low power VLSI circuits, vision sensors, and data converters.Rasoul Dehghani (dehghani@cc.iut.ac.ir ) received his B.S.E.E., M.Sc., and PhD degrees in electricalengineering from Sharif University of Technology (SUT), Iran, in 1988, 1991 and 2004. From 1987 to 1991, heworked on design and implementation of different electronic circuits and systems at SUT. From 1991 to 1998, hewas involved in implementing various electronic circuits focused on industrial applications. From 1998 to 2004, heworked with Emad Co. in Tehran and Jaalaa Company in Kuala Lumpur as a senior design engineer. Since 2006,he has been an assistant professor at IUT. His research interests include RF IC design for wireless communication,frequency synthesis, and low-voltage low-power circuits.⬅ From P.34elemental computing array (ECA),which is designed to achieve a bettercompromise between performance,flexibility, cost, and power efficiency.With ECA, the FGPA scheme can haveperformance, power, and cost similar toASICs. ECA is far more flexible than aDSP with accelerators. It is also muchcheaper and consumes far less powerthan FPGA devices.References[1] D. H. Pham, J. Gao, T. Tabata, H. Asato, S. Hori, T.Wada,“Implementation of joint pre-FFT adaptivearray antenna and post-FFT space diversitycombining for mobile ISDB-T receiver,”IEICETrans., vol. E91-B, no.1, pp.127-137, 2008.[2] Fa-Long Luo, Multimedia Mobile BroadcastingStandards: Technology and Practice Springer, 2008.[3] Fa-Long Luo, R. Unbehauen, Applied NeuralNetworks for Signal Processing. Cambridge:Cambridge University Press, 1999.[4] I. Nobuo, T. Kenichi,“HDTV Mobile Reception inAutomobiles,”Proc. IEEE, vol.94, no.1,pp.274-280, 2006.[5] T. Tabata, H. Asato, Dang Hai Pham, M. Fujimoto,N. Kikuma, S. Hori, T. Wada.“Experimental study ofadaptive array antenna system for ISDB-Thigh-speed mobile reception,”Proc. IEEE Int.Symp. Antenna and Propag., pp.1697-1700,Hawaii, 2007.[6] Fa-Long Luo and R. Unbehauen,“A minorsubspace analysis algorithm,”in IEEE Trans. NeuralNetw., vol. 8, no.5, 1997, pp1149-1155.[7] M. Kornfeld, G. May,“DVB-H and IPdatacast-broadcast to handheld devices,”IEEEBiographiesTrans. Broadcasting, vol. 53, no.1, pp. 161-170,2007.[8] M. Chari, F. Ling, A. Mantravadi, R. Krishnamoorthi,R. Vijayan, G. Walker, R. Chandhok,“FLO physicallayer: An overview,”IEEE Trans. Broadcasting, vol.53, no. 1 pp. 145-160, Mar. 2007.[9] Fa-Long Luo, Digital Front-End in WirelessCommunications and Broadcasting, Circuits andSignal Processing. Cambridge: CambridgeUniversity Press, 2011.Fa-Long Luo (falong.luo@elementcxi.com) is chief scientist at Element CXI. He has been the editor-in-chief ofthe International Journal of Digital Multimedia Broadcasting since 2007. Fa-Long Luo is now the chairman of theIEEE Industry DSP Standing Committee and technical board member of the IEEE Signal Processing Society. Hehas 28 years of research and industry experience in multimedia, communications, and broadcasting withreal-time applications and standards. Fa-Log Luo has received worldwide recognition. He has authored andedited 4 books, more than 100 technical papers and 18 patents.Ward Williams (ward.williams@elementcxi.com) is vice president of marketing for Element CXI. Ward has broadtechnology marketing and business development experience in senior management positions. He has worked forcompanies such as Xilinx, Chips and Systems, S3/SonciBLUE, and Hewlett-Packard. He frequently speaks atinternational technology shows or forums such as IEEE ICASSP and Global Mobile Congress. His researchinterests include digital front-end, software-defined radio, single-chip solutions, SoC-based hardware, andsoftware IPs.Bruce Gladstone (bruce.gladstone@elementcxi.com) has worked in the semiconductor and EDA industries forover 20 years. After graduating from the University of Illinois, he went to Silicon Valley to work in IC design,applications, and technical marketing positions. Bruce has worked for companies such as AMD, Western Digital,Synopsys, and other companies in California and Japan. As a senior director at Element CXI, he is focused onputting new-generation programmable semiconductors into the next-generation communications products.December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 41D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P45

Guest EditorialAdvances in Mobile Data CommunicationsSean CaiThis special issue of ZTE Communications focuses onrecent advances in mobile data communications for theICT and telecommunications industries. Theever-increasing amount of mobile data traffic has beenthe subject of many studies. This research area is widelyapplicable to contemporary technology and networkoptimization techniques.The priority of most operators is to further expand theirexisting 3G networks and focus on new 3G technologiesbeing developed to meet ever-increasing data traffic anddata speed. Optimization techniques of 3G and 4G networksare being explored in order to improve network performance.Today, operators are concerned about adopting the rightbackhaul technologies, leveraging existing Wi-Fi networks tooffload data traffic, building and evolving the core network,and introducing mobile cloud computing in order to supportexpanded business in the future.In this special issue of ZTE Communications, we called forpapers that provide greater insight into improved mobile datanetworks. Inter cell interference coordination (ICIC),coordinated multipoint transmission (CoMP), multiple inputmultiple output (MIMO), adaptive beamforming, and networkload balancing (NLB) are introduced. These techniques mayrequire operators to review the way they engineer and deploytheir networks. Techniques to further improve networkperformance are discussed in this issue. High-level backhaultechnologies and core network evolution paths are alsodiscussed in this issue.In“Enhanced Cell Edge Performance with Transmit PowerShaping and Multipoint/Multiflow Techniques,”by PhilipPietraski et al., a new concept called Fuzzy Cells is introducedto improve cell coverage. Fuzzy Cells enable user equipment(UE) to access the full system bandwidth in a sector by usinga component carrier (CC) transmitted at higher power than theother CC frequencies. If at least three CC frequencies areused, three sets of cell boundary locations can be created sothat no UE is simultaneously in the cell-edge region for everyCC. Cell coverage can be improved using the techniquesproposed in the article.In“Spatial Load Balancing in Wide-Area WirelessNetworks,”by Kambiz Azarian et al., NLB and single-carriermultilink (SCML) are compared. These techniques are used toincrease network capacity and improve user experience. InNLB, each sector-carrier periodically provides its loadingLi Moinformation in order to improve overall network performance.In SCML, the combined burst rate of a UE can be significantlyimproved between neighboring cells.In“Uplink Power Control for MIMO-OFDMA CellularSystems,”by Rongzhen Yang, uplink power controlalgorithms, including FPC used in 3GPP LTE andLTE-Advanced, are compared. A novel power controlalgorithm is proposed for MIMO-OFDMA systems. Asimplified maximum sector throughput (SMST) algorithm isalso proposed to maximize the sector throughput by adjustinguplink transmit power.In“Mobile Backhaul Solutions,”by Li Mo et al., mobilebackhaul requirements are reviewed, and backhaul solutionsare proposed with timing synchronization, multiprotocol labelswitching (MPLS); operation administration, and maintenance(OAM), and protection taken into consideration.We thank all the authors for their excellent papers and thereviewers that had helped us with their careful and detailedcritical reviews. We are very grateful to the editorial board fortheir helpful suggestions in improving the quality of the papersin this issue.BiographiesSean Cai (SeanCai@zte.com.cn) is the vice president of wireless at ZTE.He is responsible for developing 4G technologies as well as strategicmarketing and technical support for business development. He leads atechnical team in USA, working closely with leading operators. He isactively involved in the working group activities of various standardbodies. He has more than 20 years’professional experience intelecommunications, and he specializes in advanced wireless productsresearch and development. Prior to joining ZTE, he held various seniorpositions in infrastructure ASIC architecture design, CDMA, and opticaldevelopment. He has been awarded multiple international patents intelecommunications. He was recognized as one of the“Mobile Movers &Shakers”in 2008 by RCR Wireless News for his leadership in thedevelopment of new wireless technology.Li Mo (mo.li@zte.com.cn) received his B.Eng. degree from theDepartment of Electrical Engineering, Queen’s University, in 1989. Aftergraduation, he worked at IBM, Nortel, and Fujitsu before joining ZTE in2001. Currently, Dr. Mo is the chief architect for ZTE USA and works as amarketing specialist for ZTE headquarters in Shenzhen. Dr. Mo hasworked in the networking industry for more than 20 years, and hasnumerous publications and U.S. patents. His research interests includefixed and mobile core networks, session control, QoS, and routing. Dr. Mois also an active member of IEEE, ITU, ETSI, and IETF.42ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P46

Enhanced Cell-Edge Performance with Transmit Power-Shaping and Multipoint, Multiflow TechniquesSpecial TopicPhilip Pietraski, Gregg Charlton, Rui Yang and Carl WangEnhanced Cell-Edge Performancewith Transmit Power-Shaping andMultipoint, Multiflow TechniquesPhilip Pietraski, Gregg Charlton, Rui Yang, and Carl Wang(InterDigital Communications, LLC., King Of Prussia,PA19406, U. S. A.)Abstract: In this paper, we present a technique called“fuzzy cells”that builds on the multicarrier features of Long TermEvolution-Advanced (LTE-A) and high-speed packet access (HSPA). Multiple carriers are aggregated to create a larger systembandwidth, and these carriers are transmitted at different powers by each sector antenna. This creates a set of cell-edge locations thatdiffer from one frequency to the next. System-level simulations are performed to estimate individual user and average throughput for ahexagonal deployment of 3-sector base stations. For moderately high loads, a fuzzy cell deployment can improve tenth percentile(cell-edge) user throughput by 100% and can improve average throughput by about 30% compared with a reuse 1 scheme. Fuzzy cellsreduce inter-cell interference in the same way as higher-order reuse schemes and allow users to access the full system bandwidth.Keywords: LTE; fractional frequency reuse; multicarrier; flow aggregation; system simulation1 IntroductionIn current and evolving cellularsystems (LTE is the example in thispaper) [1], it is generally verydifficult to create a uniform userexperience because throughput at thecell edge is limited by interference fromother cells. In standard frequencyreuse 1, the cell-edge downlink (DL)signal to interference plus noise ratio(SINR) can be less than 0 dB, whichlimits user throughput even if there is alarge number of resource blocks (RBs)for UEs. The cell-edge problem isshown in Fig.1.For UEs close to an E-UTRANNode B (eNB), the DL SINR is quitehigh. However, as the UE moves awayfrom the eNB, the DL SINR decreasesuntil the UE is midway between theeNBs (at the cell edge). A handover(HO) then occurs; the UE passes to thenext eNB, and SINR starts to increaseagain. SINR is lowest near the celledge. In a typical two-dimensionalhexagonal deployment, this cell-edgecondition affects much of thedeployment area. In an OFDM system,such as LTE Release 8, the cell-edgeproblem can be mitigated to someextent by organizing UEs into groupsaccording to different parts of thespectrum and using different transmitpowers for these groups.Cell-edge UEs are groupedlogically. They are assigned a specificpart of the spectrum and are allocateda potentially higher power, as in softfrequency reuse (SFR), or higherfrequency reuse transmission, as infractional frequency reuse (FFR).Although the data signal received bycell-edge UEs may be improved, thisapproach has limitations. Take, forexample, the cell coverage and controlchannel. The control channel spans thesystem bandwidth and is sharedamong UEs. Thus, the same extendedrange that a frequency selective,UE-specific data channel might attainis not possible for UEs at the cell edge.More importantly, some UEs will nothave access to the full systembandwidth. In the FFR case shown inFig.2, the spectrum is split into twoparts: reuse 3 and reuse 1. Some of theUEs are classified as cell edge, andothers are classified as cell center. Theclassification determines which parts ofthe spectrum are available to the UEs.The details depend on the particularFFR scheme, but the cell-edge UEsare expected to have improved SINR inthe red, blue, or green frequency areas.Improved SINR comes at the price ofreduced accessible bandwidth for UEs.SINRUser LocationSINR: signal to interference plus noise ratio▲Figure 1. Cell edge in a one-dimensionaldeployment.December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 43D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P47

Special TopicEnhanced Cell-Edge Performance with Transmit Power-Shaping and Multipoint, Multiflow TechniquesPhilip Pietraski, Gregg Charlton, Rui Yang and Carl Wang231In LTE Release 10, carrieraggregation (CA) of componentcarriers (CC) allows spectrum to beaggregated. Each CC has a bandwidththat is compatible with Release 8, butUEs can attach to and be scheduled onmultiple CCs. Each of these CCs canbe considered a cell of its own with adifferent carrier frequency and,potentially, different coverage areas.Differences in coverage may occur as aresult of differences in propagationconditions; however, system transmitpower, HO thresholds (for example, cellbiasing), and even sector antennadirection can be changed to modify theDL coverage. This solves the controlchannel problem; the power of thecontrol channel can be adjusted foreach CC, thus enabling SFR. However,the problem remains that UEs cannotaccess the full system bandwidth.The cell-edge problem in Fig.1 hasalso been approached from theperspective of intercell interference(ICI) and ICI mitigation usingcoordinated base stations [2]. Adjacentcooperative base station clusters cangenerate different interference levels indifferent frequency subchannels byoptimizing power allocation to thosesubchannels. In this way, SINR near thecluster edge can be improved withoutthe need for fast coordination betweenthe clusters. In [3], this approach wasanalyzed in a one-dimensional space,and power allocation was optimizedbased on an SINR balancing conditionat fixed locations in the cluster.Therefore, power allocation dependedonly on the system’s geometry, and theeffects of adjusting the transmit poweron cell loading were not taken intoaccount. The fuzzy cells techniquerelies on system capabilities that enabletransmitting multiple powers ondifferent carriers and transmitting dataFrequencySector 1Sector 2Sector 3to UEs from different base stations ondifferent CCs.2 Fuzzy Cells◀Figure 2.FFR spectrum distribution.2.1 DefinitionsSector antenna: an antenna thatprovides coverage to an area in one ormore of the CC frequencies that makeup the system bandwidth. In theexamples discussed herein, sectorantennas support either three or six CCfrequencies.Site: a collection of co-located sectorantennas. Each site is composed ofthree sector antennas, and all sectorantennas at a site are part of the sameeNB.Hexagon: a geographical areadefined by site locations only. Ahexagon is defined by geometry and isnot affected by transmit antenna powerand shadowing.CC frequency: a frequency bandcarrying the data and control signals ofa particular CC. The sum of all the CCfrequencies is the total systembandwidth.CC : is a component carrier in anaggregation of carriers. This comprisesthe CC frequency and sector antenna.Figure 3. ▶Cell-edge improvementin one-dimensionalfuzzy cell.SINR of Each Component CarrierFreq 1 = BlueFreq 2 = RedIts coverage area is determined byUE-CC selection rules and parameters,for example, transmit power and cellbias. It is close to the usual definition ofa cell or sector.Cell edge UEs: the set of UEs inwhich throughput levels are at mostthose represented by the tenthpercentile of UE throughputs.UE perceived throughput: thenumber of bits downloaded by a UEdivided by the time that the UE has datain its buffer to be transmitted.2.2 Fuzzy Cell ConceptThe key to fuzzy cells is enabling aUE to connect to the best CC in eachCC frequency regardless of the site inwhich the CC originates. Enabling theUE to connect in this way allows a UE toaccess the full system bandwidth whenSFR-type power allocation is used.Then, a good SINR can be achieved ineach CC frequency within the systembandwidth. Multisite connection isshown in Fig.3. As the UE crosses thecell boundary in a given CC frequency,it drops that CC and picks up a new CCfrom a different site in the same CCfrequency. In this way, the UE alwayshas access to full system bandwidtheven midway between sites. The celledge is blurred, creating the concept ofa“fuzzy”cell. The resulting cell-edgeperformance is enhanced, and thisimprovement can be measured in termsof equivalent SINR (eSINR).Equivalent SINR is the SINR neededto maintain the same throughput overthe entire system bandwidth (assumingthe same SINR is observed over theentire system bandwidth). To calculateEffective SINRover Both CarriersCell-EdgeRegion RemovedUser LocationSINR: signal to interference plus noise ratio44ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P48

Enhanced Cell-Edge Performance with Transmit Power-Shaping and Multipoint, Multiflow TechniquesSpecial TopicPhilip Pietraski, Gregg Charlton, Rui Yang and Carl WangSINR (dB)453525155-5BS1: SINR for Equal Power Carriers: SINR in Carrier 1: SINR in Carrier 2: Effective SINR over Full BWA C BSINR: signal to interference plus noise ratio▲Figure 5. Two-dimensional color map ofSINR for reuse 1 baseline.eSINR, first, the SINR for the UE in eachCC with which it is associated isdetermined. The SINR for each of theseCCs is then mapped into an achievablethroughput using the Shannon Capacityformula. The achievable throughput ineach CC is summed, and the result ismapped back into the eSINR, given bywhere n is the index of the CC, N is thenumber of CC frequencies, and SINRnis the SINR of the UE in CC frequency n.Fig. 4 shows SINR and eSINR in thecase of two isolated base stations, eachwith two CCs transmitted with a powerdifference of 12 dB.The red and green curves representSINR for CC frequency 1 and CCfrequency 2, respectively. The blackcurve is the eSINR when the transmitpowers of both frequencies and basestations are identical. The blue curve isthe improved eSINR resulting fromdifferent transmit powers at the basestations. Cell-edge eSINR is improvedby multisite connections and differentD1614121086420-2eSINR =2 ∑n log2(1+SINRn)/N -1 (1)-4BS2◀Figure 4.eSINR for two isolated basestations.coverage areas for the two CCfrequencies.2.3 Fuzzy Cell SINR in aTwo-Dimensional DeploymentFig. 5 shows a two-dimensional plotof SINR for a typical hexagonal site withthree sectors per site. With reuse 1,each cell has a coverage area thatapproximates the shape of a hexagon.Assuming the eNB transmit powers aresufficiently large to limit systeminterference, the worst possible SINR isless than -4 dB without shadowing(Fig.5). With shadowing, the SINR may beeven lower. This scenario is used as abaseline case.2.4 eSINR in a Fuzzy Cell DeploymentUnlike the cell boundaries in thereuse 1 baseline (Fig. 5), cellboundaries in a fuzzy cell deploymentare no longer hexagonal (Fig. 6). In thefirst CC frequency, applying highertransmit powers from some sectorscreates larger cells. In another CCfrequency, the same phenomenonoccurs, but the pattern is rotated. If atleast three CC frequencies are used,three sets of cell boundary locations are150010005000-500-1000SINR CC1-1500-1500 -1000 -500 0 500 1000 15002520151050-1500-1500 -1000 -500 0 500 1000 1500▲Figure 6. Two-dimensional SINR for a fuzzy cell deployment.150010005000-500-1000SINR CC2created so that no UE is simultaneouslyin the cell-edge region for every CC.The cell boundaries are moved bychanging the transmit power and bymoving from one CC frequency to thenext. In Fig. 6, ΔP between high andlow power transmitters is 12 dB. Themagenta contours delineate whereSINR = 0 dB and are an estimate of thecell boundaries.In Fig. 7, the eSINR of the fuzzy celldeployment is shown alongside that ofthe baseline. Although the maximumeSINR is lower in the fuzzy cells than inthe baseline, there is no place in thefuzzy cells where eSINR drops below 0.This shows that fuzzy cell configurationbetter services cell-edge users. Thecumulative distribution functions (CDFs)for eSINR in fuzzy cells and baseline(Fig. 8) show that eSINR at the 10thpercentile is improved byapproximately 4 dB when using a fuzzycell configuration.2.5 UE-CC Association StrategyEvery UE can be associated with thesector antenna in each frequency thatprovides the best biased SINR. Biasingartificially extends the coverage area ofsmall cells or is used for loadbalancing. With a small bias, ICI islimited because the cell boundary isonly slightly extended. However, moreusers are associated with CCstransmitted at high power, and thisdecreases the resources available toeach user in those CCs. Conversely, alarge bias better balances theresources available to each user at theexpense of smaller SINR improvement.A tradeoff is necessary to obtain thebest combination of cell-edge andoverall average throughput. Ideally,cell-edge throughput is significantlyimproved at the cost of a modest2520151050150010005000-500-1000SINR CC3-1500-1500 -1000 -500 0 500 1000 15002520151050December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 45D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P49

Special TopicEnhanced Cell-Edge Performance with Transmit Power-Shaping and Multipoint, Multiflow TechniquesPhilip Pietraski, Gregg Charlton, Rui Yang and Carl Wang150010005000-500-1000-1500-1500 -1000 -500 0 500 1000CDF (%)100503020105321-5eSINR0 5 10Effective SINR (dB)CDF: cumulative distribution functionSINR: signal to interference plus noise ratio▲Figure 8. CDF of eSINR of fuzzy cells and baseline.reduction in overall throughput.However, for cells that are realisticallyloaded, it is possible to improve bothcell-edge and average throughput.The CCs providing the largest-biasedSINRs in each frequency are assignedto each UE. The bias applied to theSINR involves optimizing one of thesystem parameters so that the tenthpercentile user throughput ismaximized for each averagethroughput value.2.6 Impact of Backhaul Data FlowSplit on PerformanceFig. 9 shows the splitting of thedownlink data flow in a fuzzy celldeployment. The split occurs in theeNB. The application data packet isdownloaded from the core network tothe serving eNB (eNB1) via S1interface. The eNB1 decides how tosplit received data for each UE, and thedata payload is split in two. Part 1 issent to the UE directly from the eNB1,and part 2 is forwarded to thecooperating eNB (eNB2) via X2interface before being sent to the UE bythe eNB2. Status signals are frequentlyeSINR1500 -1000 500 0 500 1000 1500: Baseline: Fuzzy, 12 dB20151050-5◀Figure 7.Two-dimensional SINR forbaseline (left) and eSINRfor fuzzy cells (right).sent over the backhaul to controldata-flow splitting.X2 latency modifies the start andend times of the period wherefuzzy cell operation producesgains. When the applicationpacket arrives at eNB1, data isimmediately transmitted to the UEbut only from eNB1. Fuzzy celloperation begins only after theinitial X2 data forwarding delay toeNB2; that is fuzzy cell gains areonly possible when both eNBshave data to send. Similarly, if oneof the eNBs empties its buffer forthe UE before the other eNB, thenthe fuzzy cell gains are also lost. IfeNB1 finishes transmitting part 1 beforeeNB finishes transmitting part 2, then atthat point, only eNB 2 can transmit tothe eUE.The packet size for the traffic modelin [2] is 2 MB. For a 2 MB packet and anassumed X2 delay of 10-20 ms, thedelay is negligible and not modeled inthe simulations.152.7 Estimating PerformanceBecause of the complexity ofcellular systems, it is difficult topredict gains without extensivesimulation; however, someestimation is possible. For lightlyloaded cells, UEs do not contendmuch for resources in any giventransmission time interval (TTI).Inthis case, the throughputdistribution can be largelypredicted by the eSINRdistribution, and the eSINR canbe mapped to an achievablethroughput. If we consider onlythe long-term SINR in each CCfrequency in the system, we cancompute the achievableCDF (%)1005030201053210throughput for each UE in the systemusing all its connected CCs. Thelong-term SINR can be used as arough estimate of throughput in afading environment. Fig.10 showsachievable throughput distribution for apower difference of 10 dB. From thisgraph, we can expect cell-edgethroughput to be nearly 85% better thanin the baseline and still have similaraverage UE throughput. Furthermore, iftwo of the three CCs are switched off ateach sector antenna (indicated by apower difference of 100 dB in Fig. 10),eSINR distribution is further improved.The tenth percentile throughput isnearly 160% better than in the baseline,and there is nearly 30% increase inaverage UE throughput. This estimate isonly accurate if there is little or nocompetition among UEs for radioresources. As the loads increasetowards full buffer-like operation, theload imbalance between large- andsmall-coverage area CCs has agreater impact on performance, andCore Network(MME/S-GW)S11 2eNB1MME: mobility management entityS-GW: serving gatewayBackhaul▲Figure 9. Data- flow splitting at eNB.1020 30 40Throughput (Mbit/s)CDF: cumulative distribution functioneNB2▲Figure 10. eSINR based achievable throughput CDFfor a 10 dB power difference.12X2502: Baseline: Fuzzy, 10 dB: Fuzzy, 100 dB6046ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P50

Enhanced Cell-Edge Performance with Transmit Power-Shaping and Multipoint, Multiflow TechniquesSpecial TopicPhilip Pietraski, Gregg Charlton, Rui Yang and Carl Wang10th Percentile User Throuput in Active Reception (Mbit/s)121086425.00◆◆◆◆◆ ◆◆◆◆ ◆◆◆◆◆◆◆ ◆◆◆ ◆◆10.00▲Figure 11. Comparison of algorithmic schemes for high traffic with shadowing.eSINR alone cannot be used forestimation. In other words, improvingthe SINR in some CCs by increasingtransmit power must be weighedagainst the increased load these CCswill have because of their largercoverage areas. It is not clear what theoptimum tradeoff is, and we must relyon simulations to provide performanceestimates. Smaller power differencesare also preferable as the traffic loadincreases.Even in a heavily loaded system, thatis, 50% resource use, there are manyTTIs where a UE has little competitionfor resources in a given CC.Consequently, there are some TTIs inwhich the eSINR is a predictor ofperformance.3 Simulation ResultsThroughput based simulations werecarried out to compare the performanceof different types of fuzzy celldeployments to the performance of thebaseline case and FFR schemes. Weconsider two main metrics ofimportance: average UE perceivedthroughput and the cell edgeUE-perceived throughput. Becausethere are two metrics of interest, it isinstructive to present data as thepossible tradeoff between them. Thereis tradeoff between averageUE-perceived throughput and thecell-edge UE-perceived throughput.For example, even in the simple◆ ◆◆ ◆◆15.00◆◆◆◆20.00Average User Throughput Active Reception (Mbit/s): Fuzzy, Multi-Site: Fuzzy, Single-Site: FFR: Baseline25.00baseline case with PF scheduler, atradeoff between cell-edge andaverage throughput can be achievedby varying the fairness exponent (β ) inthe scheduler. By sweeping β, wegenerate the parametric curve ofcell-edge user throughput versusaverage throughput. There are severalparameters that affcet the performanceof FFR and fuzzy cell cases. To obtainthe tradeoff curves, optimizations wereperformed to find system parametersthat maximized the tenth percentilethroughput subject to different averagethroughput constraints. The otherparameters optimized were the cellbiasing and the difference in powerbetween the high tx power CC and thetwo lower-power CCs on each sectorantenna. Similar optimizations are donefor the parameters that describe theFFR and SFR techniques.3.1 High-Volume TrafficHere, a comparison of fuzzy, FFR,and baseline schemes forhigh-volume, bursty traffic isdiscussed. High-volume traffic meansthe traffic load corresponds to RU ofabout 50% in the baseline scenario.Two types of fuzzy results are shown inFig.11: multisite fuzzy and single-sitefuzzy. The former are results fromsimulation of multiflow data deliveryscheduling on the best CC regardlessof the site at which the CC originated.This is the same fuzzy cell scenariopreviously described. The latter areresults from simulation of multiflow datadelivery scheduling restricted to CCsoriginating from the same site. This ismore compatible with LTE-A Release10.Multisite fuzzy cells doubleedge-user throughput and improveaverage user throughput by roughly30% compared with FFR and baselineschemes.3.2 Low-Volume TrafficHere, a comparison of fuzzy, FFR,and baseline schemes for low-volume,bursty traffic is discussed. Low-volumetraffic implies an RU of about 10% in thebaseline case. Similar to the highvolume traffic case, low volume trafficfuzzy results are shown in Fig. 12.Multisite fuzzy doubles improvement inedge-user throughput and improvesaverage user throughput by roughly60% compared with FFR and baselineschemes.4 ConclusionsFuzzy cell technology combinesmultipoint, multifrequency transmissiontechniques, and intelligent parameterselection (e.g. transmission powerdifferences between carriers), and it isbuilt on top of carrier aggregation. Itcan be applied to any multicarriersystem, including HSPA+ and LTE.Fuzzy cells is a simple,high-performance, robust scheme thatmitigates the cell-edge problem andimproves system performance.We haveshown the cell edge versus averageperformances that can be achievedwith standard frequency reuse 1, FFR,and fuzzy cells. Fuzzy cells roughlydoubles cell-edge performancecompared with frequency reuse 1 andFFR. Fuzzy cells also significantlyimprove average performance.In fuzzy cells, the data flow to eachUE in a CC is independent in terms ofradio resource scheduling and hybridautomatic repeat request (HARQ).Therefore, allocations among differentsites do not need to be tightlycoordinated. The need for stringentlatency is relaxed compared with CoMPinter-site coordination via the X2interface. To support techniques suchDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 47D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P51

Special TopicEnhanced Cell-Edge Performance with Transmit Power-Shaping and Multipoint, Multiflow TechniquesPhilip Pietraski, Gregg Charlton, Rui Yang and Carl Wang10th Percentile User Throuput in Active Reception (Mbit/s)121086425.00▲Figure 12. Comparison of algorithmic schemes for low traffic with shadowing.as CoMP, investment needs to be madein low-latency, high-capacitybackhaul, for example, fiber optics.However, fuzzy cells can be realizedwith existing backhaul. Fuzzy cells canbe realized before other advancedtechniques and without investment innew low-latency backhaul.AppendixA system-level simulator usinghexagonal cellular layout is used tovalidate our concept. The simulationparameters are based on LTE urbanBiographies◆◆◆◆: Fuzzy, Multi-Site: Fuzzy, Single-Site: FFR: Baseline10.00◆◆◆◆macro cell no line-of- sight. UEs aredropped randomly with uniformdistribution. The simulation parametersare summarized in Table 1. Two siterings are populated with users, andstatistics are collected only from UEsnear the center site.References[1] Further advancements for E-UTRA physical layeraspects,3GPP TR 36.814, 2010.[2] R. Chang, Z. Tao, J. Zhang, C-C Jay Kuo,“A graphapproach to dynamic fractional frequency reuse(FFR) in multi-cell OFDMA networks,”Proc. IEEEInternational Conf. Commun., Dresden, Jun. 2009,pp. 1-6.Phil Pietraski (philip.pietraski@interdigital.com) received his BSEET degree from De Vry University in 1987. Hereceived his BSEE, MSEE, and Ph.D. degrees from Polytechnic University, Brooklyn, NY in 1994, 1995, and 2000.He also has a graduate certificate in wireless communications from Polytechnic University and studied acceleratorphysics at the U.S. Particle Accelerator School (hosted by universities including Harvard and Duke). He joinedInterDigital Communications in 2001 and is currently a principal engineer who leads research in wirelesscommunications. He holds more than 30 patents in the field. Prior to his transition to communications in 2000, hewas a research engineer at Brookhaven National Laboratory, National Synchrotron Light Source, responsible fordeveloping spectroscopic and imaging, solid-state and gaseous, X-ray detectors. He has also conductedresearch at the Polytechnic University for the Office of Naval Research (ONR) in underwater source localizationusing passive SONAR.Gregg Charlton obtained his BSEE from Carnegie-Mellon University in 1982 and his MSEE in EE systems fromthe University of Michigan, Ann Arbor in 1986. While at Michigan, he received a research fellowship and worked asa research assistant in the EECS department. Mr. Charlton has been with InterDigital since 2000 and is currently amember of technical staff in the advanced air interfaces organization. Prior to working at InterDigital, he worked forLockheed Martin Global Telecommunications, General Electric Aerospace, AT&T Bell Laboratories, and TRW. Hisresearch interests include analysis of mobile network deployments, wireless communication theory, and satellitecommunications.Rui Yang received his M.S. and Ph.D. degrees in electrical engineering from the University of Maryland in 1987and 1992. He joined InterDigital Communications in 2000 and has 11 years experience researching anddeveloping wireless communication systems. His interests include digital signal processing and physical layerdesign for wireless devices. He has more than 10 patents has publiushed several papers. He is currently anengineering manager at InterDigital Communications.Carl Wang obtained his BSEE from Polytechnic University, Brooklyn, NY, in 1994. He received his MSE incomputer engineering from the Carnegie-Mellon University in 2006. Mr. Wang has been with InterDigital since1996 and is currently a member of technical staff in the advanced air interfaces organization. Prior to working atInterDigital, he worked for Raytheon Company in digital signal data processing. His research interests includeanalysis of mobile network deployments, wireless communication theory, and communication protocol andarchitecture.◆◆15.00◆◆◆◆◆◆ ◆◆◆20.00Average User Throughput Active Reception (Mbit/s)◆ ◆◆◆25.00[3] G. Caire, S. Ramprashad, H. Papadopolous, C.Pepin and C. Sundeberg,“Multiuser MIMO downlinkwith limited inter-cell cooperation: Approximateinterference alignment in time, frequency and space,”46th Annu. Allerton Conf. Commun., Control, andComput., Illinois, Sept. 2008, pp 730-737.▼Table 1. System simulation parametersCarrier FrequencyBandwidth per CCNumber of CC frequenciesNumber of Antennas per LinkAverage CC Transmit PowerNumber of RBs per CCNumber of SectorAntennas per SiteInter-Site Distance (ISD)Total Number of SitesNumber of UEs per SiteUE DistributionChannel ModelPath LossStreet Width (W)Avg. Building Height (h)hBShUTShadow FadingShadowing StandardStandard Deviation (σ)Correlation Dist (dcor)Multipath ChannelUE thermal Noise DensityNoise FigureTTI LengthSector Antenna ParametersAntenna PatternPeak Antenna Gain3 dB Beamwidth (θ3dB)Front/Back Ratio (Am)SchedulerBetaLink Performance ModelInterference ModelX2 LatencyTraffic ModelPacket Size (S)User Arrival Rate (λ)Simulation Duration2 GHz3 MHz6 or 3 ( 6 CCs are used forcomparison with FFR techniquesand to allow the systembandwidth to be split into reuse1 and reuse 3)1 Tx, 1 Rx41 dBm153500 m1930 (10 per Sector Antenna)Uniform DistributionRandom Drop-PL = 161.04 - 7.1 log10 (W ) +7.5 log10 (h ) - [24.37- 3.7(h/hBS ) 2 ]log10 (hBS ) + [43.42 -3.1 log10 (hBS )] [log10 (d )-3] +20 log10(fc ) -{3.2 [(log10 (11.75 hUT ) ] 2 - 4.97}20 m20251.5 m-6 dB50 mEVA 30 kph-174 dBm / Hz7 dB1 ms-A(θ)=-min[ 12(θ/θ3dB) 2 ,A m]14 dBi3 dB @ ± 70 Degree20 dBProportional Fair. Common PoolBuffer. Perfect CSI Information.1 RB Granularity.Optimized in SimulationEESMAWGN with Power Based onPL and per CC Tx PowerSetting0 ms3GPP FTP Model 11 MB0.00063 (high load) or0.00013 (low load) Packets/TTI/UE Corresponds toApproximately 50% and 10%Resource Utilization50 s/drop48ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P52

Spatial Load Balancing in Wide-Area Wireless NetworksSpecial TopicKambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid AttarSpatial Load Balancing inWide-Area Wireless NetworksKambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid Attar(Qualcomm Inc., San Diego, CA 92121, U.S.A)Abstract: Load balancing is typically used in the frequency domain of cellular wireless networks to balance paging, access, and trafficload across the available bandwidth. In this paper, we extend load balancing into the spatial domain, and we develop twoapproaches—network load balancing and single-carrier multilink—for spatial load balancing. Although these techniques are mostlyapplied to cellular wireless networks and Wi-Fi networks, we show how they can be applied to EV-DO, a 3G cellular data network. When adevice has more than one candidate server, these techniques can be used to determine the quality of the channel between a server andthe device and to determine the load on each server. The proposed techniques leverage the advantages of existing EV-DO networkarchitecture and are fully backward compatible. Network operators can substantially increase network capacity and improve userexperience by using these techniques. Combining load balancing in the frequency and spatial domains improves connectivity within anetwork and allows resources to be optimally allocated according to the p-fair criterion. Combined load balancing further improvesperformance.Keywords: CDMA 2000; EV-DO; DO-RevC; high rate packet data (HRPD); network load balancing (NLB); single-carrier multilink (SCML)1 IntroductionDemand for data in wirelessnetworks is spatiallynon-uniform, which leads tochokepoint sectors, and it is alsotime-varying, which means thatchokepoint sector locations change.Chokepoint sectors operate at or closeto the maximum load; therefore, theydetermine network performance, userperception of network performanceand, ultimately, user experience. In anetwork with fixed assets andunbalanced loading, spatial loadbalancing uses network resources fromneighbors of chokepoint sectors toserve users in the chokepoint sector.This increases the capacity of thenetwork.Field data collected fromcommercially deployed networks showsthat only a small fraction of sectors arechokepoints at any given time and thatchokepoint sectors typically haveseveral lightly loaded neighbor sectors.Fig. 1 shows the distribution of sectorloading (per-sector forward link (FL)slot use) during peak hour in ametropolitan commercial EV-DOnetwork where sectors with low slot useFrequency0.400.350.300.250.200.150.100.050-200 5 102030are lightly loaded. Operators addhardware by splitting cells near thechokepoint sectors or increasebandwidth by adding carriers to the▲Figure 1. Per-sector FL slot use during peak hour (mean FL use is 26%).405060FL Slot Utilization (%)70809010010.90.80.70.60.50.40.30.20.10120Probability < XDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 49D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P53

Special TopicSpatial Load Balancing in Wide-Area Wireless NetworksKambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid Attarchokepoint sectors. Such actions leadto networks that are designed forworst-case scenarios and are thereforeunderused most of the time.The spatial load-balancingtechniques discussed in this paper arenetwork load balancing (NLB) andsingle-carrier multilink (SCML).In NLB, a device is opportunisticallyreassigned from a heavily loadedsector to a lightly loaded sector, even ifthe channel quality of the heavilyloaded sector is better. The end resultis increased capacity for the chokepointsector from offloading the mostexpensive users (poor channel qualitymeans that the network has to allocatemore resources for a given amount oftraffic) to the neighbor sector.With SCML, any device that cansimultaneously process two or moreindependent data streams can achievethe benefits of a multicarrier network,even though the device may be in thehand-off region of a single-carrierdeployment.With multicarrier EV-DO, the EV-DOcarriers that serve a mobile device usedifferent carrier frequencies and mayoriginate from the same cell or differentcells. With SCML, the same carrierfrequency serves the mobile devicewith independent data streams andfrom different sectors. NLB improvesperformance when the chokepointsector is loaded, and SCML improvesperformance at all load levels. In thispaper, we describe how thesetechniques augment load-balancing inthe frequency domain.In Section 2, we provide necessarybackground information. In sections 3and 4, we discuss the concepts,algorithms, and implementations ofNLB and SCML. In section 5, weperform simulations to determine theperformance of NLB and SCML. Insection 6, we make concluding remarksand discuss future work.2 Background2.1 Server SelectionWhen the access terminal (AT) in awireless network has a choice ofserving sectors, it typically selects theserver with the best channel quality.(Here, hysteresis based on channelquality and time is not discussed for thesake of simplicity.) In EV-DO, the ATselects the best forward layer (FL)server from the candidate servers thathave a corresponding reverse link (RL)of acceptable quality. If the DRCLockbit on, the RL quality is acceptable; ifthe DRCLock bit is off, the RL quality isunacceptable. The DRCLock bit itself isset according to the quality of the RLoverhead channels received at thebase station.2.2 Multicarrier OperationWhen an AT in EV-DO is assignedmultiple carriers, those carriers mayoriginate from the same sector ordifferent sectors (one forward-linkserving sector for each carrier) [1], [2].Commercial EV-DO uses independentschedulers on each sector carrier, andmulticarrier EV-DO uses multilink radiolink protocol (RLP) on each sectorcarrier. Multilink RLP enables thenetwork to split a data stream at thetransmitter across multiple independentschedulers and to combine the datastream at the receiver while ensuringearly detection of packet loss andin-order delivery of RLP packets [3].When multiple links (or carriers) areassigned to an AT, the networkdistributes forward-link data acrossavailable links. Data throughput oneach link can be different, and thethroughput varies over time. Enhancedflow control (EFC) ensures FL data isdistributed across links according tolink throughput and that the datadistribution adapts to changes in linkthroughputs.Each forward link in EV-DO requiresreverse-link feedback to indicate FLchannel quality and acknowledgements(ACK/NACK) for physical-layer HARQ.EV-DO supports multiple feedbackmultiplexing modes. RL overhead ofmultiple FL carriers can be transmittedover a single RL carrier to minimize theRL pilot overhead.2.3 FL Load MetricsWe introduce a new metric for theloading at each sector. This metric isthe average number of non-emptyqueues at the sector carrier. It iscalculated for each slot (which takes1.66 ms) and is fed to a filter with atime-constant of a few seconds togenerate the Neff metric.Networks are at timesbackhaul-limited, and this leads to theair link at the BTS being underused. Abackhaul-limited sector-carrier maynot appear to be loaded because ofqueue under-run. Therefore, backhaullimit must be accounted for whencalculating loading level at eachsector-carrier. A metric for backhaullimit is the number of unfilled flowcontrol requests for a sector-carrierthat has non-empty AT queues at theBSC.2.4 Operator DeploymentsOperators deploy wireless datanetworks according to data demand.Hence, operator deployments are notuniform and have spatial variations inthe number of carriers deployed.Typically, a large area in an operatornetwork has three or less deployedcarriers. In regions of the network withjust one carrier, multicarrier EV-DOcannot improve performance of ATswith multicarrier capability.2.5 Smart Carrier ManagementEV-DO access networks (ANs)assign the best-available carrier to anAT by transporting the load levelsacross carriers (on the forward andreverse links) and the channel qualityfrom the ATs to each of the candidatecarriers.3 Network Load BalancingNLB is shown in Fig. 2. PN (b) is thebest FL sector for the user; however,PN (b) is heavily loaded and theneighboring sector, PN (a), is lightlyloaded. If PN (b) is selected as theserving sector, it has bettersignal-to-interference plus noise ratio(SINR) than PN (a) but is scheduledmuch less time than in PN (a) becauseof heavy loading. Thus, even at the costof FL SINR degradation, the AT’sthroughput can increase when the FLloading differential is sufficient to offsetthe SINR degradation.50ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P54

Spatial Load Balancing in Wide-Area Wireless NetworksSpecial TopicKambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid Attarf1Load=20%Single-CarrierDevice153 k 307 kLoad=80%Chokepoint SectorAn important consideration in NLB isits interaction with the partial loading ofneighboring sectors. The SINR of thetraffic channel in the chokepoint sectoris higher than the SINR of the pilotchannel (only applicable to EV-DO)because of lesser interference from thelightly loaded neighboring sector. If theAT is served by the lightly loadedneighbor, then the chokepoint sectorwith high neighbor loading gives stronginterference. If a network has a largenumber of ATs with the ability tospatially null the strong interferer, thenetwork can gain more from NLB. Inaddition, offloading users to lightlyloaded neighboring sectors drivesthose sectors towards higher slot use,and these sectors then cause moreinterference to the chokepoint sector.This effect limits the loading differentialacross sectors where NLB is beneficial.The NLB algorithms also specifyminimum FL channel quality to ensurethe AT does not choose an extremelyweak FL server, regardless of the loaddifference between neighboringsectors. This allows the AT to continueto receive the FL control channels withthe desired performance. There are twotypes of NLB algorithm: one for existingATs deployed by operators and anotherfor new ATs.3.1 Network Load Balancing for LegacyATsImplementation of the NLB algorithmfor legacy ATs does not require the ATto have any additional capability. Eachsector-carrier periodically provides itsload information (Neff) to the BSC. TheBSC also receives FL pilot strengthinformation from ATs via aRouteUpdateRequest in EV-DO every2 to 4 seconds. With this information,the BSC can select suitable legacy ATsand induce them to change servingSwitch Betwwen(PN(a), f1) and (PN(b), f1)PN(a) PN(b) PN(c)◀Figure 2.Network load balancing.sectors by setting the DRCLock bit (RLquality indicator) that corresponds tothe chokepoint sector to off (indicatingpoor RL quality).3.2 Network Load Balancing forNew ATsThe AN assists NLB for new ATs.Each sector-carrier periodicallyprovides its Neff to the BSC. The BSCpropagates this to all members in theneighbor list of each sector-carrier.Each sector-carrier periodicallybroadcasts its load level — as well asthe load level for all sector-carriers inits neighbor list — on the synchronouscontrol channel. The server selectionalgorithm at the access terminal thenuses not only FL channel quality and RLchannel quality (DRCLock bit) but alsothe FL load level from the candidateserving sectors on each carrier. Theserver selection metric is modified touse the FL channel quality and the FLload level, and the constraint wherebythe AT selects a serving sector fromsectors with acceptable RL quality ismaintained.4 Single-Carrier MultilinkEV-DO ATs are multicarrier enabled,which means they can receive data onmultiple links regardless of whetherthese links are from the same sectorson the same carrier. This capability canbe harnessed with SCML to allowmulticarrier-enabled ATs to haveimproved performance even in regionsof the network with less than threecarriers. Because SCML is beneficialonly in handoff regions (typicallycharacterized by lower SINR andless-than-ideal performance), itdelivers gains where they are neededmost. With SCML, the network canserve the AT using multiple servers on acarrier. Each FL server creates a linkbetween the AN and AT. SCML allowsmultiple links on a single carrier, andthat is why it is called single-carriermultilink. An AT’s active set are thosesectors that control the power of the ATon the RL and can serve the AT on theFL. SCML enhances FL data rate whenall sectors in the AT’s active set arelightly loaded, and it balances the loadwhen some chokepoint sectors are inthe AT’s active set. For an AT that canreceive FL data from n linkssimultaneously, SCML allows the AN toassign additional links and enhance FLthroughput when the number of carriersassigned to AT is < n and the AT hasmultiple sectors in the active set. SCMLleverages the advantages of EV-DO,and in the following, we explain somechanges to existing EV-DO.4.1 Reverse Link Feedback MultiplexingFor each additional link assignedusing SCML, an additional set of RLfeedback channels is required. This issent over a single RL carrier usingreverse link feedback multiplexing. Themultiplexing is supported by methodssuch as basic feedback multiplexing(BFM) and enhanced feedbackmultiplexing (EFM).4.2 Distributed Network SchedulingThe network can more flexiblyschedule FL data when an AT is servedon multiple links. Distributed networkscheduling (DNS) improves networkefficiency by biasing FL data deliverytowards a better link [4]. When FL slotuse is 100%, DNS fairly schedulesacross all ATs regardless of the numberof links assigned. When FL slot use isless than 100%, DNS allows SCML ATsto achieve some trunking gain by usingfree slots on each link, which results ina higher FL data rate for SCML ATs.4.3 Receiving Diversity and InterferenceNullingAn SCML AT may be simultaneouslyserved from multiple links on samecarrier, resulting in self interference.The AT is simultaneously served fromtwo sectors that interfere with eachother. Even if no other AT is present inthe system, each link assigned to theDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 51D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P55

Special TopicSpatial Load Balancing in Wide-Area Wireless NetworksKambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid AttarNLB Gain (%)504030201001●●●●2: 1 Rx: 2 Rx3 4 5Demand RatioNLB: network load balancing▲Figure 3. NLB gain as a function of demand ratio.SCML AT receives interference fromanother link(s). To achieve gains frommultiple links, the AT must overcomethe interference. Some EV-DOreceiving diversity implementations canspatially null a strong interferer, and thisis the key to SCML. The network shouldlimit SCML assignment to receivingdiversity ATs with spatial nullingcapabilities. ATs with more Rx antennasare typically capable of superior spatialnulling of strong interferers, andtherefore, SCML performance improveswith an increase in number of AT Rxantennas.4.4 Smart Carrier ManagementThe smart carrier managementalgorithm can be modified to take theSCML capability of a device intoaccount. The network allocates andremoves SCML based on signalconditions, and this ensures thatadditional links can provide reasonablethroughput. Links across carriers arepreferred because they are orthogonaland provide trunking (multiplexing)gain. If the AT can support more linksthan assigned carriers, AN uses SCMLto assign additional links on a carrieraccording to available FL controlresources and RL load on that carrier.The additional links provide trunkinggain in the spatial dimension. Once theAT server selection algorithm has beenassigned multiple links by thesmart-carrier management algorithm,the AT server-selection algorithmchooses the best link(s) by using the FL●●6●●78channel quality and FL load.This is similar to the NLBalgorithm for new ATs. TheAN has multiple paths that itcan use to serve FL data,and DNS allows networkthroughput to be optimized.5 PerformanceSimulations reveal theperformance gained usingNLB and SCML. Theperformance results are validfor a seven-cell layout with acenter-cell demand model;that is, the center cell hasmultiple times the number ofusers of neighboring cells. The ratio ofthe number of users in the center cell tousers in the neighboring cell is thedemand ratio. Except for Fig. 3, whichshows load-balancing gain as afunction of demand ratio, all otherfigures are valid for a demand ratio ofthree. Users have an http traffic model;that is, they repeat downloading asample http webpage according to arandom arrival process with a fixedrate. To determine performance in aparticular scenario, for example,dual-antenna or dual-carrier deviceswith load balancing enabled,simulations are performed for a rangeof download rates. Each simulation isthen characterized in terms ofdownload rate (the average number ofwebpages downloadedper minute perhigh-demandsector-carrier) and the tailof the page delay. The tailof the page delay isdefined as follows: Thedelay for eachdownloaded http page isthe time between when thefirst byte of the page isqueued in the BSC to thetime the last byte isdownloaded by the user.Then, the page delay CDF(across all users and allwebpages) is formed, andthe ninety percentile delayis chosen as the tail pagedelay. The tail page delaySeconds●●versus rate curve for the scenario isthen formed from all such rate-delaypairs. Both single and dual-antennadevices are simulated. Single andmulticarrier devices are also simulated.5.1 Load-Balancing Gain forSingle-Carrier DevicesFig. 4 shows the tail page delayversus rate curves for single-carrierdevices with one or two antennas. Loadbalancing provides significant gain onlywhen the load level is sufficiently large,that is, when the solid red and bluecurves (which represent the baselineand new NLB cases for two-antennadevices, respectively) fall on top ofeach other at low rates. Because of thelow load differential at low load levels,no user is offloaded to another sectorand there is no load balancing gain.Also in Fig. 4, the load-balancing gainsfor single-antenna devices (new orlegacy) are not as large as those fordual-antenna devices. Once a user isoffloaded from a heavily loaded sectorto a lightly loaded one, the heavilyloaded sector appears as a stronginterferer. This occurs because the userreceives a stronger signal from theheavily loaded sector than from thelightly loaded one. Dual-antennadevices can spatially null some of theinterference. Single-antenna deviceslack such capability and thus achievelower load balancing gains. Multiplereceiving antennas and special null of a●●Tail of Page Delay●10 1 ●● : 2 Rx,Baseline●● : 2 Rx,NewNLB●: 2 Rx,Legacy NLB●●● : 1 Rx,Baseline●●● : 1Rx,NewNLB● : 1 Rx,Legacy NLB510152025Web-Page (minute/sector-carrier)NLB: network load balancing▲Figure 4. NLB gain for single-carrier devices with one andtwo antennas (600 kB per web page).●●●●●●52ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P56

Spatial Load Balancing in Wide-Area Wireless NetworksSpecial TopicKambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid AttarSecondsTail of Page Delay●●10 10●●5●●NLB: network load balancing▲Figure 5. Load-balancing gain devices with two antennas(600 kB per web-page).●●strong interferer becomes even moreimportant for SCML because usersusually suffer self-interference.Consequently, NLB and SCML arerecommended only for dual-antennadevices.Fig. 4 also shows that although theload-balancing gain for legacy devicesis not as large as the gain for newdevices, they are comparable. The gainis reduced because the BSC, whichinduces load balancing for legacyterminals, does not have accurateinformation about user SINR.Fig. 3 shows that the load-balancinggain increases with the demand ratio,that is, the ratio of users in the centercell to users in the neighbor cell. This isa result of more users being offloadedfrom the heavily loaded center cell tothe lightly loaded neighboring cells asthe demand differential increases. Loadbalancing provides substantial gainsover a wide range of demand ratios.5.2 Load-Balancing Gain for MulticarrierDevicesFig. 5 shows the tail page delayversus rate curves for single andmulticarrier devices (the maximumnumber of assigned carriers is two).The single and multicarrier deviceshave two antennas. Load-balancinggain improves when a second carrier isadded. As carriers are added, usersexperience trunking (multiplexing) gain.10●●●15●●●●: 2Rx,2Carr,Baseline: 2Rx,2Carr,NLB: 2Rx,1Carr,Baseline: 2Rx,1Carr,NLB20●●Web-Pages/Minute/Sector-Carrier25●30The horizontal distancebetween the magenta andred curves in Fig. 5represents the trunkinggain of dual-antennadevices when a secondcarrier is added. Trunkinggain is also a function ofload level. As the loadlevel increases, thetrunking gain decreases.The magenta and redcurves in Fig. 5 merge asrate is sufficientlyincreased. Both baselineand NLB cases achievetrunking gain when asecond carrier is added;the trunking gainachieved in the latter caseis more substantialbecause some users hand off to lightlyloaded neighbors.In these cases, there is a dual benefitfrom NLB—spatial load-balancingcombined with load-balancing in thefrequency domain—and this magnifiesthe gains from spatial load balancing.NLB combined with multicarrierdeployments increases networkcapacity when the network is loadedand also substantially improves UE.This means faster web-pagedownloads when the network is not asloaded.Fig. 6 and Fig. 7 show therelationship between load balancingand trunking gains for two-antennadevices at low (7 second tail pagedelay) and high (13 second tail pagedelay) load levels. Specifically, X in Fig.6 denotes the network capacity for asingle-carrier deployment withdual-antenna devices were the tailpage delay does not exceed 7seconds. The coefficients in Fig. 6represent the per-carrier networkcapacity gain when load balancing isenabled or additional carriers areadded. At low load levels, per-carriertrunking gain (63%) created by addinga second carrier is much larger than thegain from load balancing (5%). NLBprovides significant gain when sectorsare sufficiently loaded; trunking gain isgreater when load level is smaller. Incontrast, Fig. 7 shows that at high loadlevels, load balancing provides thesame gain (26%) as adding the secondcarrier. Thus, combining multicarrierdeployments with per-carrier loadbalancing provides substantial gainover a wide range of load levels.5.3 SCML PerformanceFig. 8 shows FL burst rate as asingle-carrier user moves between twoneighboring cells. With no SCML, auser’s burst rate reduces as the ATmoves away from the center of the celland approaches the cell edge. Theburst rate starts increasing again as theAT approaches the center of theneighboring cell. With SCML, a user’sburst rate is significantly improved nearthe cell edge because the user isserved by two sectors at the cell edge.Thus SCML can provide ATs with amore uniform UE as the user moveswithin the network.Fig. 9 shows the tail page delayversus rate curves for dual-antennasingle-carrier devices. The clusterconsists of seven sectors, and thecenter sector has three times thenumber of users of its neighboringsectors. Load balancing providessubstantial gains only when load level is1.05X 2.02X 2.54XNLB NLB NLBX 1.64X 1.84X1-Carr 2-Carr 3-CarrNLB: network load balancing▲Figure 6. NLB gains for dual-antennamulticarrier devices at a low load level(7 second tail page delay).1.26Y 1.65Y 1.90YNLB NLB NLBY 1.26Y 1.33Y1-Carr 2-Carr 3-CarrNLB: network load balancing▲Figure 7. NLB gains for dual-antennamulticarrier devices at a high load-level(13 seconds tail page delay).December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 53D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P57

Special TopicSpatial Load Balancing in Wide-Area Wireless NetworksKambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid AttarSecondsThroughput Mbit/s10 12.52.01.51.00.50.1●●20●●●●0.240●●●0.3NLB: network load balancingSCML: single-carrier multilink: Baseline: NLB: SCML●●high, whereas SCML providessubstantial gains at all load levels. Atlow load levels, SCML exploits spatialtrunking across sectors, whereas athigh load levels, SCML asymptotes toNLB where the AT is effectively servedby one sector only. Single-carriermultilink is therefore a form ofsoft-network load balancing.SCML: single-carrier multilink▲Figure 9. SCML gain for dual-antenna single-carrierdevices (300 kB per webpage).0.4Throughput● ●●● ●0.50.6Distance from Center Cell (km)6 ConclusionsThis paper gives an overview of loadbalancing in the spatial dimension. NLBand SCML are introduced, and relatedalgorithms, implementation, andperformance are presented. We also●: Baseline: SCML60●Tail of Page Delay80●●●100●Web-Pages/Minute/Sector-Carrier●●●●0.7120●●●●0.81400.9◀Figure 8.Burst rate for asingle-carrier user movingbetween two neighboringcells.show that load balancing inthe spatial dimension andfrequency dimension canbe combined to amplifygains in networks wherecapacity is constrained.The proposed methods arefully backward compatibleand leverage theadvantages of EV-DOnetworks thus increasingthe return on investment forthese networks. NLB, inparticular, is applicable tolegacy devices. Thesetechniques improve userexperience when thesector is not operating atload.Further improvements toNLB and SCML can bemade with higher-order receivingdiversity at the device. This enablesfurther improvements in spatial nulling,which allows the device to besimultaneously served by even moresectors on the same carrier.Similar techniques, such assingle-frequency dual cell (SFDC), arebeing developed for UMTS networksand are applicable to LTE networks aswell. NLB and SCML are bothapplicable to local-area wirelessnetworks.References[1] CDMA2000 High Rate Packet Data Air InterfaceSpecification, 3GPP2 C.S0024-C version 1.0, April2010.[2] R. Attar, D. Ghosh, C. Lott, Mingxi Fan, P. Black, R.Rezalifar and P. Agashe,“Evolution of CDMA2000cellular networks: multicarrier EV-DO,”IEEECommun. Mag., vol. 44, Issue 3, pp. 46-53, Mar.2006.[3] CDMA 2000 High Rate Packet Data SupplementalServices, 3GPP2 C.S0063-B version 1.0, May 2010.[4] R. Gowaikar, C Lott, A. Jafarian, D. Ghosh, K.Azarian and R. Attar,“Distributed Scheduling forWireless Networks,”in Proc. IEEE ISIT, 2011.BiographiesKambiz Azarian (kambiza@qualcomm.com)received his Ph.D. in electrical engineering fromOhio State University. He joined Qualcomm in 2007and has since been working on 3G and 4G cellularsystems.Ravindra Patwardhan joined Qualcomm in 1998.He has worked on ISDN, Globalstar, Telematics,1xRTT, 1xEV-DO, and UMB.Christopher Lott received his Ph.D. in controltheory from the University of Michigan. He hasworked on 1xEV-DO systems design for more than10 years and is currently the DO systems lead atQualcomm. He has designed aspects ofcommunications systems, including LTE, GPS, andInmarsat.Donna Ghosh received her M.S. and Ph.D.degrees from Pennsylvania State University. Shejoined Qualcomm in 2003 and has been working on1xEV-DO system design and standardization withan emphasis on MAC design. Her researchinterests include resource allocation, distributedalgorithms, and wireless systems.Radhika Gowaikar received her M.S. and Ph.D.degrees in electrical engineering from CaliforniaInstitute of Technology. She joined Qualcomm in2006 and is interested in distributed algorithms forresource allocation in wireless networks. She hasworked on these for the 1xEV-DO system.Rashid Attar received his B.E. in electronics fromBombay University in 1994 and his MSEE fromSyracuse University in 1996. He is a senior directorof engineering in Corporate Research andDevelopment (CR&D), Qualcomm. He joinedQualcomm in 1996 and was first engaged in theintegration of IS-95 systems. Since 1998, he hasworked on the system design, prototype,development, standardization, andcommercialization of 1xEV-DO (Rev0, RevA, RevB,and DO-Adv) and 1x-Adv. He is currently co-leadof the CR&D DO-Advanced effort and acting leadof the Qualcomm CDMA Technologies (QCT)CDMA2000 modem systems team.AD IndexA1 and Back Cover:ZTE Corporation54ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P58

Uplink Power Control for MIMO-OFDMA Cellular SystemsSpecial TopicRongzhen Yang and Hujun YinUplink Power Control forMIMO-OFDMA Cellular SystemsRongzhen Yang 1 and Hujun Yin 2(1. Intel China Corp., Shanghai 200241, P.R. China;2. Intel Corp., 2200 Mission College Blvd, CA 95054-1549, U.S.A.)Abstract: In this paper, we propose a novel uplink power control algorithm, SMST, for multiple-input multiple-output orthogonalfrequency-division multiple access (MIMO-OFDMA).We perform an extensive system-level simulation to compare different uplinkpower control algorithms, including the FPC adopted in 3GPP LTE and LTE-Advanced. Simulations show that SMST adopted in IEEE802.16m outperforms other algorithms in terms of spectral efficiency, cell-edge performance, interference control, and trade-off controlbetween sector-accumulated throughput and cell-edge user throughput. The SMST performance gain over FPC can be more than 40%.Keywords: uplink power control; inter-cell interference; OFDMA; MIMO1 IntroductionUplink power control is a criticalfeature of CDMA cellularsystems. It is used to alleviatethe near-far problem caused byintracell interference. The newgeneration of broadband wirelesstechnologies, including WiMAX andLTE, are designed to provide higherbandwidth, higher peak throughput,and higher spectral efficiency [1],[2].These wireless technologies are basedon orthogonal frequency-divisionmultiple access (OFDMA), in which allthe uplink transmissions are orthogonalwithin one cell [3]. Therefore, intracellinterference is minimal compared withtechnologies based on 3G CDMA.Fourth-generation WiMAX and LTEwireless standards are designed tosupport up to 20 MHz channelbandwidth, high-order 4 × 2 and 4 × 4multiple-input multiple-output (MIMO),and aggressive frequency reuse(reuse 1) to improve spectral efficiencyand user throughput. Especially forcell-edge users, intercell interferenceis a significant problem caused byaggressive frequency reuse. Uplinkpower control is an importantmechanism for controlling intercellinterference and improving cell-edgeuser experience (UE), even inbroadband wireless systems based onOFDMA[4]-[7]. As opposed to 3G,where closed-loop power control istypically used to control both intracelland intercell interference, the mainobjective of uplink power control inOFDMA-based systems is to controlintercell interference [8],[9].In this paper, we investigate controlof uplink transmit power in order tomaximize sector and cell-edgespectral efficiency. These are essentialparameters for next-generationbroadband wireless systems [10]-[13].A simplified maximum-sectorthroughput (SMST) algorithm isproposed that maximizes sectorthroughput by adjusting uplink transmitpower.2 Uplink Power ControlMechanisms2.1 BackgroundPower control has been studiedextensively since the introduction ofcellular systems. A power-controlmechanism can be closed-loop powercontrol (CLPC), open-loop powercontrol (OLPC), or CLPC-OLPCcombined power control. In CLPC,power control is centralized at the basestation (BS). The mobile station (MS)provides feedback on link quality, andthe BS calculates the uplink transmitpower level for the MS and instructs theMS to transmit at that level. In OLPC,the MS measures link quality andcalculates the uplink transmit powerlevel (based on predeterminedequations) in a distributed manner. TheBS may influence the MS by adjustingcertain parameters in the equation, butthe BS does not directly control the MStransmit power.2.1.1 CLPCIn a CDMA system, CLPC allowscommands to be sent from the BS sothat power can be quickly adjusted. Inan OFDMA system, intracellinterference is not significant, soadaptive modulation and coding (AMC)and hybrid automatic repeat request(HARQ) are used to provide fast linkadaptation for the data channel. CLPCis mainly used for fixed-rate controlchannel when channel fading exceedsDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 55D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P59

Special TopicUplink Power Control for MIMO-OFDMA Cellular SystemsRongzhen Yang and Hujun YinNeighbor Cell 1Home CellChannel Loss: CL0Channel Loss: CL1Channel Loss: CL2Neighbor Cell 2power (linear) of one data subcarrier ata given time;•i: cell index. i = 0 is the home BS,and i = 1-N is the neighboring BSs;•CL i: The instantaneous channelloss (linear) from MS transmitter to onereceiving antenna of a BS;•NI i: The noise-plus-interferencelevel (linear) per subcarrier at a BSreceiving antenna.the power margin. In CLPC in the BS,the received signal quality is monitoredand power-adjustment commands aresent.2.1.2 OLPCOLPC is mainly implemented in theMS. The MS measures downlink signalstatus, compensates the uplink pathloss, and controls interference toneighboring BSs. With OLPC, the MSneeds some static/semi-staticconfiguration parameters to besignaled by the BS and does notrequire short-term inputs. This savesoverall signaling overhead.2.1.3 CLPC-OLPC Combined PowerControlIn current 4G wireless standards,both CLPC and OLPC are combined tobalance flexibility with signalingoverhead. OLPC is mostly used to savesignaling overhead.2.2 Uplink Power Control AlgorithmsDepending on the goal of the powercontrol mechanism, the uplinkpower-control algorithm can be basedon signal-to-noise ratio (SNR),signal-to-interference plus noise ratio(SINR), or Internet of things (IoT).2.2.1 SNR-Based AlgorithmThe goal of an SNR-based algorithmis to maintain the desired receivedsignal strength in the BS. AnSNR-based algorithm does not takeinto consideration the uplink-receivedinterference power level. This type ofalgorithm is simple to implement andalways convergent. Fractional powercontrol (FPC) in LTE is an SNR-basedalgorithm.2.2.2 SINR-Based AlgorithmThe goal of an SINR-based algorithmis to maintain the desired receivedSINR level in the BS while taking intoaccount the uplink interference powerlevel. SMST used in IEEE 802.16m is anSINR-based algorithm that defines thedesired received SINR goal for eachMS and takes into account themeasured downlinksignal-to-interference ratio (SIR),uplink MIMO mode, and BS antennaconfiguration.2.2.3 IoT-Based algorithmThe goal of an IoT-based algorithmis to maintain the desired uplinkinterference level in the BS. In general,an IoT-based algorithm can stabilizethe interference level in the uplink tohelp control interference andmodulation-coding-scheme (MCS)level estimation.3 Maximizing Throughputwith Power Control3.1 System ModelingFig. 1 shows an uplink interferencemodel for one MS in a typicalMIMO-OFDMA cellular system.At any time, the MS uses the uplink totransmit to its home BS and causesinterference to neighboring BSs. Tosimplify the analysis, simple inputsimple output (SISO) is assumed, andthe parameters for the model are•P s: The MS uplink transmission3.2 Maximum Throughput CriteriaThe aggregated cell throughput is▲Figure 1. Uplink interference for one MS in a typical MIMO-OFDMA cellular system. R Aggregated = ∑R i (1)iIf we assume that all BSs occupy thesame bandwidth, (1) can be modified:R Aggregated=∑BW ×SE i =BW ×∑SE iii→SE Aggregated=∑SE i (2)For uplink power control of each MS,aggregated spectrum efficiency SE isprioritized as the target of the algorithm,which can be expressed asTo solve (3) for each MS, thefollowing process is used:1) P S is the transmission power of onesubcarrier that is initialized to 02) A power increase, ΔP S , is assumedin order to calculatei) SE gain : as the power increases, thisis the spectrum efficiency gain that theMS can achieve in the home BS on thesubcarrier.ii) SE loss : as the power increases, thisis the spectrum efficiency loss the MSinflicts on the neighboring BSs on thesubcarrier.3) SE gain and SE loss are comparedi) If SE gain > SE loss, P S = P S + ΔP S , goback to 2ii) If SE gain ≤ SE loss, then the PS isoptimum on the subcarrier for the MS.This process is used to evaluate theSE gain and SE loss for each powerincrease of ΔP S .At any time, when a power increaseof ΔP S is assumed, the SE gain of the MSon the subcarrier can be obtained bywhereiP tx=argmax∑SE i (3)iP txSE gain=log(1+SINR (Home) New)- log(1+ SINR (Home) Original)) (4)56ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P60

Uplink Power Control for MIMO-OFDMA Cellular SystemsSpecial TopicRongzhen Yang and Hujun YinP SCLSINR (Home)0Original= (5)NI 0(P S+ΔP S)Home CellHome ChannelLoss: CLHVirtual Cell ChannelLoss: CLiCLSINR (Home)0New= (6)NI 0Combining (4), (5), and (6), SE gain canbe expressed asVirtual Neighbor CellΔP S( )CL 0SE gain = log 1+ . (7)NI 0 +P SCL 0The SE loss on the subcarrier in oneneighbor BS i (i = 1-N) can beexpressed asSE loss(i) =log(1+ SINR (i ) Original)-log(1+SINR (i ) New) . (8)where SINR (i ) Original= ,NI iSSINR(i ) New =i,NI i+ΔI iS i is the received signal power on theΔP Ssubcarrier of BS i, and ΔI i= isCL ithe increased inference power causedby ΔP S.From (8), the total SEloss on thesubcarrier in all neighbor BSs can beexpressed asNSE loss = ∑SE loss(i ) . (9)i = 1Therefore, the optimum power on thesubcarrier can be calculated byincreasing P S of the subcarrier by ΔP Ssteps from 0 until the SE gain and SE loss onthis subcarrier no longer satisfiesSE gain ≥ SE loss . (10)If the MS is assigned resources foruplink transmission by the BS, optimaltransmission power for all assigneduplink subcarriers can be calculated inorder to achieve the overall optimumcell throughput. This algorithm is calledthe maximum sector throughput (MST)algorithm.4 Simplified MaximumSector ThroughputAlgorithm4.1 MST Algorithm Simplified FormThe MST algorithm is only useful inS i▲Figure 2. Uplink interference from one MS.theory and is impossible to beimplemented in practice. The algorithmrequires the home BS or MS toaccurately know the received signalpower and interference power on allneighbor BSs on each subcarrier at thetime of uplink transmission. Makingsome simple assumptions, wedeveloped a practical SMST algorithm.First, we modeled one virtual neighborBS (or sector) that accounts for allinterference impact on SEloss (Fig. 2).We assume that all BSs have thesame downlink transmission powerlevel. The total downlink referenceinterference power from the virtual BSto the MS isP DL_I = = ∑ (11)CL I i = 1 CL iwhere P DL_Preamble is the downlinkpreamble power used as the referencesignal to measure downlink path loss,and CL I is the virtual downlink pathloss. The downlink reference signalpower of the MS isDownlink SIR can be measured asP DL_ S, so thatP DL_ ISIR DL=P DL_PreambleNP DL_PreambleP DL_PreambleP DL_s = . (12)CL HCLSIRIDL= . (13)CL HThe SE gain can be expressed as1+SINR(New,MRC)SE gain= log . (14)1+SINR Original, MRCOne virtual neighbor model is used,and one MS keeps the sametransmission power spectral density(PSD) for all data tones (PSD Data). Onlyone transmission stream is used, andthere are Nr receiving antennas at theBS (an MRC receiver is assumed).It is further assumed that eachreceiving antenna on the BS sufferssimilar noise and interference levelNI H,Ant , and the path loss fromtransmission antenna to each receivingantenna is similar to CL H , then (14) canbe re-written asSE gain= logN r×(PSD Data+ΔPSD Data)1+CL H×NI H,AntN r×PSD Data1+CL H×NI H,AntΔPSD DataCL H= log 1+. (15)N r×PSD DataNI H,Ant+CL HSimilarly, the new SE loss on the virtualneighbor BS can be expressed asSE loss= logI1+SINR(Original,MRCI )1+SINR New,MRCN r×SNR I,Ant×P Noice,Ant1+NI I,Ant= log (16)N r×SNR I,Ant×P Noice,Ant1+ΔPSDNIDataI,Ant+CL Iwhere NI I,Ant, is the noise plusinterference for each antenna on thevirtual BS, P Noice,Ant, is the white noise foreach antenna on the virtual BS, SNR I,Antis the average SNR level for eachantenna on the virtual BS, andSNR I,Ant × P Noise,Ant, is used to estimatereceived signal PSD for each antennaon the virtual BS.The optimum overall SE can beobtained when the following condition ismet:maxΔPSDData→0(SE H+SE I) ⇒ ΔSE= SE gain-SE loss= 0 (ΔPSD Data→0 ) . (17)From (13), (15), (16) and (17), we canDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 57D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P61

Special TopicUplink Power Control for MIMO-OFDMA Cellular SystemsRongzhen Yang and Hujun YinderiveNISINRI,Ant1H,Ant= ×(1+ ) ×NI H,Ant N r×SINR I,AntSIR1DL- (18)N rwhere, SINR H,Ant is the received averageSINR for each antenna on the home BS,andSIN I,Ant×P Noise,AntSINR I,Ant = .NI I,Antand is converted into a linear value toalign with the linear result of (19).Equation (21) is a core part of the IEEE802.16m [3] uplink power-controlalgorithm.Equation (21) is the SMST algorithmcombined with minimum target SINRthreshold. It expresses the suitableuplink target SINR in each receivingantenna on a home BS. If OLPC isapplied, the PSD of each datasubcarrier used by the MS isEquation (18) can be furthersimplified:PSD Tx=L+SINR Target(dB)+NL+Offset (22)where L is the average downlink pathSINR H,Ant =γ ×SIR1DL- (19) loss measured by MS base on BSN rtransmission power level and receivedwhere γ is a derived parameter to signal power level, NI is the averagecontrol the interference to other cells: noise and interference level for eachsubcarrier at the BS antenna, theNIγ =I,Ant 1× 1+ . (20) information is broadcasted from BS toNI H,Ant( N r×SINR I,Ant)MS, and Offset is the MS-specificThe parameter,γ , is linearly related power offset decided by the BS.to the ratio of virtual-cell average NI If (21) is used for conventional CLPClevel to home-cell average NI level. design, MS needs to report theHigh γ results in high interference over measured downlink SIR value SIR DL tothermal (IoT), and low γ results in low BS periodically, and BS then uses (21)IoT.to calculate the target SINR. Comparedwith the measured SINR, the difference4.2 Limitation of Minimum Transmission is compensated by CLPC commands.RateEquation (21) is the solution for uplinkEquation (19) provides the optimal single stream. In IEEE 802.16m [11],solution for the SMST algorithm. The two options for uplink multistreamresulting target SINR on each antenna power control were discussed. Theis expressed as a linear value. From power level of each stream in uplink(19), some MSs could have negative multistreams can be kept the same astarget SINR because of the very low the single stream or the power level ofmeasured downlink SIR. For these MSs, each stream in uplink multistreams canthe results of (19) indicate that any be reduced to keep the sum ofpower assigned to these MSs reduces power/interference similar to the singleoverall throughput. Because all active stream. Both options improveMSs must support a minimumperformance in different cell sizes andtransmission rate to keep them online, in different scenarios, so there is onethe minimum target SINR (to support additional control parameter added intothe minimum transmission rate) should (18) to support both options:be set as the threshold. Then, (19) isSINR MINmodified as follows:SINR Target =10log10(max(10 10 ,SINR MIN1SINR Target =10log10(max(10 10 ,γ ×SIR DL- ))N rγ ×SIR1-β ×10log10(TNS ) (23)DL- )) . (21)N rwhere β is the newly added parameterIn (21), the resulting target SINR is and can be set by the BS as 0converted to decibels so that the(disabled) or 1 (enabled) fortransmission power can beenvironment performance tuning. TNSconveniently calculated. The minimum is the total number of uplinkSINR threshold is expressed in decibels multistreams.Equation (23) is the final derivation ofSMST adopted into IEEE 802.16m [3]for data-channel power control. 16 mcontrol-channel power-control designwas discussed in [14].5 Algorithm Evaluation andComparison5.1 Evaluation ConsiderationsBefore evaluating uplink powercontrol algorithms, the following generalpoints for system-level simulation needto be considered:•sector-cumulated throughput andcell-edge throughput. These are themetrics of overall performancedetermined by uplink power-controlalgorithms.•fairness control curve. In general,the uplink power control algorithmprovides a trade-off betweensector-cumulated throughput andcell-edge throughput. One new-formcurve is used to show the sector SE andcell-edge SE (Figs. 3, 4, 8 and 9).•IoT control. IoT is the key metric ofuplink interference in a systemevaluation. Effective IoT control needsto be proven for an uplink power controlalgorithm.As (22) and (23) show, there aresome key parameters that need to beevaluated and discussed for realimplementation:•γ plays a key role in IoT andfairness. In the evaluation, the results ofdifferent γ values are discussed•SINR MIN (dB) is the minimumthreshold for cell-edge user SINR.•L is the average downlinkpath-loss measured by the MS. This iskey to deciding the transmission powerlevel. For different productimplementations, the long-termaverage or short-term average havecause different effects.•The open-loop power adjustmentis decided by the MS. The rate of powercontrol may be used in productimplementation. Fast (per-frame) orslow (per 50 frames) control rates areevaluated.5.2 Results and DiscussionThe evaluation scenario in [15]58ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P62

Uplink Power Control for MIMO-OFDMA Cellular SystemsSpecial TopicRongzhen Yang and Hujun Yin▼Table 1. Key parameters for IEEE 802.16 mULPC evaluation scenarioSite LayoutParameterCarrier Frequency(GHz)System Bandwidth(MHz)Reuse FactorFrame Duration(Preamble+DL+UL)( ms)Number of OFDMSymbols in UL FrameFFT Size(tone)Useful ToneNumber of LRULRU TypeNumber of Usersper SectorCMIMO SupportUser LocationSite to Site Distance(m)ChannelMax Power in MS(dBm)Antenna ConfigurationHARQTarget PERLink to System MappingSchedule TypeResourceAssignment BlockPenetration Loss(dB)Control OverheadValueHexagonal Grid,19 3-SectorSites, Wrap-Around2.5101518102486448DRU10NoUniform Distribution500eITU-Ped B,3 km/h231×2 SIMOOn (Max Retrains:4/Sync)0.2RBIRPF8 LRU200 for SE Calculation(not Defined yet)CMIMO: Collaborative MIMODRU: Distributed Resource UnitHARQ: hybrid automatic repeat requestLRU: least recently usedRBIR: Resource Block Information RateSIMO: Single-Input-Multi-Outputfollows the IEEE 802.16m EvaluationMethodology Document [16], but somesimplifications are made for uplinkpower control. The key parameters ofthe evaluation scenario are listed inTable 1.In the OLPC algorithm, path loss andpower control rate are related toimplementation. The path loss L isapplied to (22) to calculate thetransmission power for each subcarrier.L is estimated by the MS throughdownlink signaling measurement. Forthe extreme evaluation setting, the pathloss L is assumed with two conditions:•instantaneous path loss. This is theideal accurate instantaneous (current▼Table 2. Evaluation results of different gamma valuesGamma00.20.40.60.81.01.2Full Power(No PC)SINR MIN(dB)0000000N/A2.70833.23873.77384.11494.54824.79014.91553.3585frame) downlink value used in thesimulation•average path loss. This is theestimated path loss is the verylong-term (~50 frames, smoothingfactor = 0.02) average downlink pathloss. Ideally, short-term fast fading issmoothed.In product implementation, the pathloss for a specific product should besomewhere between the two extremes,depending on the averaging factor. TheSINR MIN (dB) is set as 0 dB, and theresults are shown in Table 2.With different γ values, theperformance trade-off between sectorSE and cell-edge SE is shown in Fig. 3Celledge SE: Throught (bit/s/Hz)0.080.070.060.050.040.030.020.7Sector Throughput(Mbit/s) Cell-Edge Throughput(kbit/s) Sector SE Cell-Edge SEγ=0.0 γ=0.2 γ=0.40.8256.5689262.6901252.3136222.1170180.5995150.1525102.126932.5632SMST: simplified maximum sector throughput0.72220.86361.00641.09731.21291.27731.31080.89560.06840.07010.06730.05920.04820.04000.02720.0087based on the results of Table 2.For OLPC, the power adjustment ratebased on (22) can be per frame(rate = 1) or infrequent (rate > 1). Ingeneral, the power control rate isassumed to be 1; that is, in each frame,OLPC is applied for uplink transmission.However, there are some parametersrelated to product implementation, suchas path loss, L , and downlink signalversus interference SIR DL measurementand estimation. NI updating, broadcastby the base station, is not performer perframe. The change of power dictatedby OLPC may be longer than oneframe. Therefore, in one extreme case,a control rate of 50 frames is also: SMST, Instantaneous Pathloss, Power Control Rate=1: SMST, Average Pathloss, Power Control Rate=1γ=0.6γ=0.80.9 1.0 1.1 1.2Sector SE: Average Throughput (bits/s/Hz)γ=1.0γ=1.2▲Figure 3. Performance tradeoff curves: instantaneous path loss vs. averaged path loss. Thecalculation of sector spectrum efficiency (SE) and cell-edge SE is defined in [16].1.31.4December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 59D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P63

Special TopicUplink Power Control for MIMO-OFDMA Cellular SystemsRongzhen Yang and Hujun YinCelledge SE: Throught (bit/s/Hz)0.080.070.060.050.040.030.020.8γ=0.20.9γ=0.4▲Figure 4. Performance tradeoff curves: control rate of 1 vs. control rate of 50.simulated to verify the robustness of theSMST algorithm.In Fig. 3 and Fig. 4, the robustness ofSMST compared with path-lossestimation and power updating rate isvery clear. The average path lossshows minor gain in sector averagethroughput but minor loss in cell-edgeSE. The reason is that the estimatedaverage path loss can help the MSmaintain stable transmission power thatis determined by OLPC. Then, thesignal/interference estimation on the BSis more accurate, and this is veryimportant for the adaptive andmodulation coding (AMC) process toassign the desired MCS level to the MS.The average path loss shows minor lossin a cell-edge SE becauseinstantaneous path-loss estimationallows the MS to perform fast-linkadaption by OLPC power change.Similarly, Fig. 4 shows that a slowpower-updating rate of 50 has somegain in maximum sector SE and someloss in cell-edge SE.Other aspects of uplink powercontrol, such as user throughputdistribution, IoT control, and powerdistribution, are shown in one of thefollowing four cases, because it is: SMST, Instantaneous Pathloss, Power Control Rate=1: SMST, Instantaneous Pathloss, Power Control Rate=50γ=0.6γ=0.81.0 1.1 1.2 1.3Sector SE: Average Throughput (bits/s/Hz)SMST: simplified maximum sector throughputclosely related to real productimplementation. The evaluation resultsare presented as an empiricalcumulative distribution function (CDF).Fig. 5 shows one example of how thecontrol factor,γ, controls the trade-offCDF1.00.90.80.70.60.50.40.30.20.100γ=1.0γ=1.2CDF: cumulative distribution function▲Figure 5. User throughput CDF by control factor γ.1.41.55001000User Throughput (kbit/s)between sectors to accumulatethroughput and fairness in cell-edgeperformance.The IoT distribution shows theeffectiveness of interferencemanagement.Fig. 6 shows the relationship betweenIoT distribution and control factor γ.When γ = 0, the SMST algorithmdegenerates to a fixed SINR targetmethod:SINR Target=SINR MIN(dB) . (24)The IoT distribution of γ = 0 is theminimum IoT value in the case where allMSs try to maintain the minimumtransmission rate expected by the BS.The IoT value, an importantmeasurement for intercell interferencecontrol, is jointly decided by SINR MIN(dB) and γ. The SINR (dB) determinesthe base value of IoT distribution, and γdetermines the IoT differential based onthe base value. Fig. 7 shows IoTdistribution for different SINR MIN valuesfor γ = 0.2.5.3. Comparison of Different UplinkPower Control AlgorithmsIn this section, we compare theperformance of SMST with fractionalpower control (FPC) that was1500:γ=0.0:γ=0.2:γ=0.4:γ=0.6:γ=0.8:γ=1.0:γ=1.2:γ=1.460ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P64

Uplink Power Control for MIMO-OFDMA Cellular SystemsSpecial TopicRongzhen Yang and Hujun YinCDF1.00.90.80.70.60.50.40.30.20.10.00▲Figure 6. IoT CDF by control factor γ.incorporated into 3GPP E-UTRA afterRelease8.During the development of 3GPP LTEspecifications, FPC was proposed[18]-[20] and used [17] as an OLPCmethod for uplink data channel physicaluplink shared channel (PUSCH).The key to FPC is to compensate thefraction of the path loss by control factorα, which can be expressed as PSD ofeach subcarrier, the same as in (22):510IoT (dB)CDF: cumulative distribution function15:γ=0.0:γ=0.2:γ=0.4:γ=0.6:γ=0.8:γ=1.0:γ=1.2:γ=1.4large range.For the scenario defined in Table 1,the results of SMST and FPC are shownon Fig. 8.The performance trade-off of FPC islocated in the range of P 0 largerthan -79 dBm. If the P 0 is not wellselected, for example, when P 0 < -791.020dBm, the sector SE and cell-edge SEdegrades at the same time. Similarevaluation results can be found in [23]and [24], where the performance ofsector throughput and cell-edgethroughput are shown separately by IoTlevel. Here, the IoT level is directlycontrolled by P 0.When the selected optimumperformance point of FPC is P 0 = 79dBm with fixed α = 0.8, the sector SE is0.9375, and cell edge SE is 0.0447.Compared with the SMST algorithmresults without optimum searching forSINR MIN( the value is just set as 0 dB),there is still big performance gap. IfSMST maintains the same sector SE asFPC optimum point of 0.9375, the SMSTcan provide cell-edge SE of 0.0687,which is a 53.69% gain over 0.0447 ofFPC optimum cell-edge SE. If SMSTmaintains the same cell-edge SE asFPC optimum point of 0.0447, the SMSTcan provide sector SE of 1.2404, whichis a 32.31% gain over the 0.9375 of FPCoptimum-sector SE. The averageperformance gap between SMST andFPC is 43%. In the second method of P 0selection, P 0 is transformed fromα valueand cell-edge target SNR, assuggested in [23] and given byP 0=α×(SNR Target+PSD n)+(1-α)×PSD Max (26)γ=0.2PSD Tx=P 0+α ×PL . (25)When α = 0 , the FPC algorithmdegenerates to a fixed transmissionPSD algorithm without affecting pathloss. When α = 1, the FPC algorithmprovides full path loss compensationand degenerates to the fixed targetreceived signal strength (RSS)algorithm. The 3GPP specification [19]defines α from 0 to 1 asα∈{0,0.4,0.5,0.6,0.7,0.8,0.9,1}.Theparameter, P 0 , is also important in FPCalgorithm. The P 0 is mainly defined intwo ways. In the first method of P 0selection,α is fixed as one selectedoptimum value, such as 0.8, in mostpublished results, and P 0 is searchedfor the optimum value in eachsimulation scenario. P 0 is explicitlysignaled by eNodeB [21],[22] with aCDF0.90.80.70.60.50.40.30.20.10.0051015IoT (dB)CDF: cumulative distribution function▲Figure 7. IoT CDF by control factor SINR MIN.: SINR MIN = -5 dB: SINR MIN = -4 dB: SINR MIN = -3 dB: SINR MIN = -2 dB: SINR MIN = -1 dB: SINR MIN = 0 dB: SINR MIN = 1 dB: SINR MIN = 2 dB: SINR MIN = 3 dB2025December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 61D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P65

Special TopicUplink Power Control for MIMO-OFDMA Cellular SystemsRongzhen Yang and Hujun YinCelledge SE: Throughput (bit/s/Hz)0.08: SMST, Average Pathloss, Power Control Rate=1: FPC,Fixed α=0.5, Different P0 value0.070.060.05P0=-83 dBM -81 dBM-79 dBM0.04-80 dBM0.03P0=-84 dBM -82 dBM -75 dBM-74 dBM0.02-73 dBM-71 dBM0.01-65 dBM0.7 0.8 0.9 1.0 1.11.2 1.3Sector SE: Average Throughput (bit/s/Hz)FPC: Fractional power control SMST: simplified maximum sector throughput▲Figure 8. Comparison of SMST and FPC performance curves (with fixed α = 0.8).1.4point of FPC algorithm is difficult toidentify. One selected performancepoint for FPC (SNR Target = 6 dB ,α = 0.6)is used for performance comparison, itssector SE is 1.0799, and its cell-edgeSE is 0.0350. If SMST maintains thesame sector SE as FPC selected pointof 1.0799, the SMST can providecell-edge SE of 0.0607, which is a73.43% gain over 0.0350 of FPCselected cell-edge SE. If SMSTmaintains the same cell-edge SE asFPC optimum point of 0.0350, the SMSTcan provide sector SE of 1.2904, whichis 19.49% gain over 1.0799 of FPCselected sector SE. Also, the averagegain of SMST compared with FPC canbe as much as 46.46%. From Fig. 8 andFig. 9, SMST performs 40% better thanFPC.Celledge SE: Throughput (bit/s/Hz)0.080.070.060.050.040.030.020.010.000.7α=1.00.8FPC: Fractional power controlα=0.8α=0.4▲Figure 9. Comparison of SMST and FPC performance (with fixed-target SNR ).where SNR Target is the cell edge targetSNR, PSD n is the thermal noise PSD,and PSD Max is the maximumtransmission power PSD for assignedresource size.Equation (26) provides theconnection betweenαand P 0. As aresult, if α = 0 , then (25) becomesα=0.0PSD Tx=PSD Max . (27)α=0.60.9 1.0 1.1Sector SE: Average Throughput (bit/s/Hz): SMST, Average Pathloss, Power Control Rate=1: FPC,Target SNR=6 dB: FPC,Target SNR=8 dB: FPC,Target SNR=10 dB: FPC,Target SNR=12 dBα=0.21.2SMST: simplified maximum sector throughputFPC degenerates to a maximumtransmission power scheme. If α = 1,then (25) becomesPSD Tx=SNR Target. (28)1.3FPC degenerates to a fixedreceiving-target SNR method. Acomparison of the performances offixed-target SNR is shown in Fig. 9.In Fig. 9, the optimal performance1.46 Conclusion andDiscussionIn this paper, we have proposed anovel uplink power control algorithm,SMST, for MIMO-OFDMA systems. Wealso performed extensive system-levelsimulations to compare different uplinkpower control algorithms, including theFPC adopted in 3GPP LTE andLTE-Advanced. Our results show thatSMST adopted in IEEE 802.16moutperforms other algorithms in terms ofspectral efficiency, cell-edgeperformance, interference control, andtradeoff control betweensector-accumulated throughput andcell-edge user throughput. The SMSTperformance gain over FPC can bemore than 40%.References[1]“IEEE 802.16m System Requirements Document(SRD).”[Online]. Available:http://www.wirelessman.org/tgm/docs/80216m-07_002r9.pdf[2]“IEEE 802.16m System Description Document(SDD) .”[Online]. Available:http://www.wirelessman.org/tgm/docs/80216m-09_0034r2.zip[3]“IEEE Std 802.16m, Amendment to IEEE Standardfor Local and Metropolitan Area Networks - Part 16:Air Interface for Broadband Wireless AccessSystems - Advanced Air Interface.”[Online].Available: http://www.wirelessman.org/pubs/80216m.html[4] Dongcheol Kim,Wookbong Lee, Jin Sam Kwak,Yeong-Hyeon Kwon, Sungho Moon, Jong YoungHan, and HanGyu Cho,“Proposed Text on PowerControl Section for the IEEE 802.16m Amendment.”➡To P.6762ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P66

Mobile Backhaul SolutionsSpecial TopicLi Mo, Fei Yuan, and Jian YangMobile Backhaul SolutionsLi Mo 1 , Fei Yuan 2 , and Jian Yang 2(1. ZTE(USA) Inc., Richardson, TX 75080, U.S.A;2. ZTE Corporation, Shenzhen 518004, P.R.China)Abstract: In this paper, we give an overview of mobile backhaul solutions and propose an MPLS-centered solution that takes intoaccount timing synchronization, OAM, and protection. We also propose an evolved protection bandwidth allocation mechanism thatmakes the transport network as efficient as possible.Keywords: mobile backhaul; multiprotocol label switching (MPLS); optical networking; protection1 IntroductionRecent advances in cellulartechnology have necessitatedsignificant improvement in themobile backhaul network. Withincreased use of smartphones andlaptops, mobile communication hasbeen moving from voice-centeredinfrastructure to data-centeredinfrastructure.New mobile technologies, such ashigh-speed packet access (HSPA+)and LTE, have significantly increaseddownlink and uplink speeds over theradio link. Improvements over the radioaccess link demand similarimprovement in bandwidth on thebackhaul network.However, current backhaul networksare based on legacy technologies thatare optimized for voice transport. Tobenefit from advances in radio accesslink technologies, the backhaul networkneeds to be transformed from a slow,TDM-centered network into ahigh-speed, data-centered network.Driven by bandwidth demand of theInternet, engineers have been buildinghigh-speed, data-centered networksfor many years. However, building amobile backhaul network is verydifferent from building a commercialdata network. A mobile backhaulnetwork requires•QoS-based traffic with strictrequirements on delay and jitter•high availability•timing synchronization•frequency and phasesynchronization for optimal handover•coordinated multipointtransmission (CoMP) and interferencecancellation in LTE-Advanced•enhanced OAM•multiple transport modes thatsupport data traffic and TDM traffic forlegacy base stations•symmetric bandwidth requirementsfor the up-stream and down-streambetween the central office and basestations.In traditional wired broadbandaccess, bandwidth is usuallyasymmetrical. As the use of cloudservices increases, symmetricalbandwidth in both directions becomesmore important when designing thenetwork.In this paper, anEthernet/MPLS-based solution isproposed to satisfy mobile backhaulrequirements. The solution addressesthe following issues:•Depending on the situation, themigration path from the backhaulnetwork (based on legacy technology)to the new infrastructure (capable ofsupporting the new cellulartechnologies such as LTE) may bedifferent.•High availability necessitatesprotection infrastructure. A means ofefficiently reserving bandwidth on theprotection path for an MPLS-basednetwork needs to be articulated.•OAM techniques of data-centerednetworks need to be incorporated intotraditional transport networks.•Traffic aggregation needs to beperformed closer to the base stations.In this paper, a new trend in mobileradio access network architecture isbriefly discussed. In this architecture,called cloud radio access network(C-RAN), baseband processing of theradio signal is centralized. Althoughthere are many advantages to acentralized approach in terms of powerconsumption and project engineering,among other things, C-RAN alsorequires a very different transportnetwork between the radio amplifiersand antenna, and the central locationwhere baseband signals are processed.2 Basic RequirementsDuring mobile communication,connectivity from the base station to thecentral office occurs over the mobilebackhaul network.In this backhaul network, traffic fordifferent generations of the radioaccess network is transported. Differentgenerations of cellular technologieshave different architectural names toindicate cell-site equipment and theDecember 2011 Vol.9 No.4 ZTE COMMUNICATIONS 63D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P67

Special TopicMobile Backhaul SolutionsLi Mo, Fei Yuan, and Jian YangBase Station万 万 万 万 万万 万 万 万 万万 万 万 万 万 万 万 万万 万 万 万 万 万 万 万万 万 万 万 万万 万 万 万 万CO:central officeBackhaul NetworkMSC-S: message switch center-server CO Site ◀Figure 1.Mobile backhaul network.corresponding equipment in the centraloffice (CO). In this paper, the cell-siteequipment is called base station unlessotherwise noted (e.g. in LTE, the basestation is eNodeB). The central officeequipment is called message switchcenter-server (MSC-S) for voice, andgateway for data (e.g. SAE-GW forEPC, SGSN for UTRAN). The backhaulnetwork is shown in Fig. 1.Connectivity between the basestation and the CO site is determinedby the technology used in the RAN, thelocation of the cell site, the bandwidthrequirements, and local regulations.In a GSM (voice and limited datatraffic) and UMTS/CDMA 2000(voice-centered with increasing datatraffic) mobile network, many cell sitesuse microwave as backhaultechnology. However, in the LTE era ofwide spectrum and high data rate, themicrowave-based backhaul networkdoes not have adequate bandwidth. Inthis paper, we discuss a fiber-basedmobile backhaul network.The current practice for buildingmobile backhaul networks is a flatIP-centered architecture. To supportlegacy mobile technologies, one of thekey requirements of a mobile backhaulnetwork is support for a native TDMinterface. Support for such an interfaceusually comes in the form of circuitemulation, for example, PWE3emulation, in the flat IP network.Different generations of mobilenetwork also have different delay andjitter requirements. For LTE, NextGeneration Mobile Network Alliance(NGMN) specifies a maximum delay of5 ms (one way), and there are nospecific jitter requirements.The common requirement for LTEbandwidth, which also depends onspectrum width and the number ofMSC-SGatewaycarrier-sections, is around300-400 Mbit/s. In LTE, the physicallayer can be gigabit Ethernet, 10GEthernet (in sharing mode), GPON, or10G EPON.Normal EPON is not included in thephysical layer because the sharingcapability of the link is a maximum of1G. 10G GPON is also omitted becausecurrent 10G GPON has asymmetricbandwidth.Fig. 2 shows the basic requirementsfor a mobile backhaul for advancedmobile technologies such as LTE andLTE-Advanced.3 MPLS-Based SolutionsIn this section, the specifics of usingmultiprotocol label switching (MPLS)technology for mobile backhaulapplication are articulated.Apart from the final access link(which can be xPON-based), insidethe mobile backhaul network, thelayer-2 technology is Ethernet. WithEthernet, technologies such asLayer-2/Layer-3 virtual private network(VPN), pseudo-wire for circuitemulation in supporting TDM traffic,and MPLS can be easily supported.For mobile backhaul applications, apure IP-based network with no MPLScan also provide adequate service withincreases in delay and jitter. TheseBase Station万 万 万 万 万万 万 万 万 万万 万 万 万 万 万 万 万万 万 万 万 万 万 万 万万 万 万 万 万万 万 万 万 万TDM and IPAccess TechnologiesGigabit Ethernet10 G EthernetGPON, 10 G EOPNincreases are due to an increase inprocessing at each hop. With an IPtransport network supporting multipleservices, and considering IP router portcost and processing complexity, mobilebackhaul can best be implemented viaan MPLS-based network.Another contender for mobilebackhaul could be optical transportnetwork (OTN) with ODU-0 to transportgigabit Ethernet signals. But an OTNnetwork is TDM in nature and cannot beas flexibly deployed as an MPLSnetwork.3.1 QoS and CoSTo make the network scalable so thatit can provide many different types ofservices, CoS is usually used withqueuing and discard techniques. Amobile backhaul network needs toqueue and discard priorities from theradio access network.In almost all radio accesstechnologies, the relative queuing and/or discard priorities are marked. Afterbaseband processing of the radiosignals at the entrance of the mobilebackhaul network, the markings on theradio signals need to be remappedonto the technologies used in thebackhaul network.The core technology used in themobile backhaul network is MPLS overEthernet. For layer-2 (Ethernet), the802.3p bits can indicate the priority ofthe packet. At the starting point of theMPLS label switched path (LSP), theEXP bits can also indicate the priority ofthe packet.3.2 Timing SynchronizationIn a mobile backhaul network withEthernet and MPLS, timing needs to besynchronized for better handoffsupport, better voice/video quality,Backhaul NetworkFigure 2. ▶Mobile backhaul networkaccess technologies. CO:central office MSC-S: message switch center-serverMSC-SGatewayCO Site64ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P68

Mobile Backhaul SolutionsSpecial TopicLi Mo, Fei Yuan, and Jian YangBase Station万 万 万 万 万万 万 万 万 万万 万 万 万 万 万 万 万万 万 万 万 万 万 万 万万 万 万 万 万万 万 万 万 万Working path with Bandwidth ReservationBackhaul NetworkProtection path with Bandwidth ReservationCO:central officeMSC-S: message switch center-serverlower interference, and betterbandwidth (e.g. CoMP).There are a number of ways ofsupporting timing synchronization,each of which has different advantagesand disadvantages:•GPS based timing. This is accuratefor frequency and can also be used forphase synchronization. The drawbackis that GPS is often unavailablebecause of a poor environment foracquiring GPS signals or because theU.S. government has encrypted thesignals.•packet-based timingsynchronization (IEEE1588v2). Thissynchronization mechanism has beenwidely evaluated over the past a fewyears. It is adequate for synchronizationin universal mobile telecommunicationssystem (UMTS) networks. A CDMAnetwork uses GPS timing. Theperformance of this mechanism insynchronizing timing for moretime-stringent requirements, such asCoMP, still requires further study.•synchronous Ethernet. Thisscheme only works if Ethernet isprovided end-to-end. In this case, thelast access link also needs to beEthernet, and an xPON-based solutionis excluded. Synchronous Ethernet canbe deployed if the carrier is not usingxPON-based technologies for theaccess link between the backhaulnetwork and the base station.3.3 OAM for MPLS-Based NetworkOAM is an important andfundamental in transport networks[1]. Itcontributes to•reducing operational complexityand costs because it allows for efficientand automatic detection, localization,handling, and diagnosis of defects. ItMSC-SGatewayCO Site◀Figure 3.Protection path andbandwidth reservation inthe MPLS network.also minimizes service interruptionsand operational repair time.•enhancing network availability byensuring that defects, such as thoseresulting in misdirected customertraffic, are detected, diagnosed, anddealt with before a customer reports theproblem.•meeting service and performanceobjectives. OAM allows service-levelagreements (SLAs) to be verified in amultimaintenance environment andallows service degradation caused bypacket delay or packet loss to bedetermined.A backhaul solution should supportend-to-end OAM in a multivendorenvironment and simplify networkoperations with OAM tools.Currently, the telecommunicationstandardization sector (ITU-T) SG15and Internet engineering task force(IETF) are cooperating on MPLStransport profile (MPLS-TP). Comparedwith MPLS, OAM is a very importantcharacteristic of MPLS-TP. Faultmanagement (FM) OAM functions, suchas continuity check (CC), continuityverification (CV), andremote defect indication(RDI), can automaticallydetect and localizedefects that occur innetwork. Performancemanagement (PM) OAM6functions, such as packetloss measurement (LM),packet delaymeasurement (DM), andthroughput measurement,can diagnose servicedegradation. OAMfunctionality is also thekey for networksurvivability andtriggering network protection.4 Protection BandwidthReservationMPLS-based mobile backhaul is amesh network with bandwidthassurance. When high availability isrequired, bandwidth also needs to beassured when the network experiencesa limited fault, such as a single fiber cutor a single MPLS transport nodal failure.The current method of supportinghigh-availability is to provide a workingpath and a protection path. Therequired bandwidth is allocated on boththe working path and protection path. Ifthe working path experiences a failure,the protection path is used. Becausethe bandwidth is also allocated on theprotection path, bandwidth is assuredfor the backhaul network. Thisarrangement is shown in Fig. 3.The protection bandwidth reservationscheme assures bandwidth when thereis limited network failure, but sharedprotection between many different cellsites is not taken into account. Thiscritical flaw is shown by the simple ringtopology in Fig. 4.In the arrangement in Fig. 4, theprotection bandwidth reserved on thelink between nodes 5 and 6 (same asother links) is the sum of the bandwidthfor the working paths A and B. If weassume a single failure on any nodeand take into consideration the ringstructure, the maximum bandwidthrequired for protection is the maximumbandwidth for working paths A and B.Working Path A1Protection Path B5 Protection Path A 34WorkingPath B▲Figure 4. Current working and protection-reservation method.2December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 65D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P69

Special TopicMobile Backhaul SolutionsLi Mo, Fei Yuan, and Jian YangBandwidth reservation for any link isperformed in the egress direction. Theadjacent node is concerned with theingress direction, which is its egressdirection on the link concerned.For bandwidth reservation in apacket network, equivalent bandwidthis usually used. For any packet stream,the equivalent bandwidth is usuallyderived from the peak bandwidth,average bandwidth, maximum burstsize, buffer size, and packet lossprobability. The equivalent bandwidth issomewhere between the averagebandwidth and the peak bandwidth.The theory of equivalent bandwidth,sometimes also called effectivebandwidth, is discussed in [2] and [3].In the following, for simplicity, thebandwidth required for reservationimplies the equivalent bandwidth.A link k on node Y is given by Y(k),k = 1,..., K y . There are K y links for Y.The egress bandwidth on a particularlink has four main features:1) Normal working bandwidth. This isthe bandwidth required if there is nofailure inside the network. For a nodallink Y(k), 1 ≤ k ≤ K y, the bandwidth,B N(Y(k)) is the sum of the equivalentbandwidths of all the working pathsover Y(k).2) Bandwidth may be absent on Y(k)because of the failure of link i of node J(i.e. J (i ) failure). For any working pathover Y(k) and J (i ), when there is J(i )failure, the working-path traffic isabsent if the working path goes overJ (i ) before Y(k). There is a differencebetween 1+1 protection [4] and MPLSfast reroute protection [5]. The workingpath traffic is absent for 1+1 protectionand may be absent for MPLS fastreroute, depending on topology and thedetour path. Depending on theprotection scheme, the working-pathtraffic may be absent if the workingpath goes over before Y(k). In thispaper, this bandwidth is B a(Y(k)),J (i )),where 1≤ k ≤K y and 1≤ i ≤K J.3) Bandwidth may be absent onY(k) due to failure of J. For a completefailure of J, the bandwidth absent onlink Y(k) is B a(Y(k)),J (0)). In this case,the notation is the same as that forbandwidth absent because of linkfailure, and J (0) denotes the nodalRRUTo CoreBBU+RRUBBUTo CoreRRU+BBUTraditional RAN BBU CRANTo CoreBBU: baseband unit CRAN: cloud radio access network RAN: radio access network RRU: radio remote unit▲Figure 5. From traditional RAN to C-RAN.failure.4) Protection bandwidth on Y(k),1≤ k ≤K y because of failure on J (i ),0≤ i ≤K J.. This failure includes bothlink and nodal failures. In such failures,the equivalent bandwidth for theprotection path is B P(Y(k),J (i )), where1≤ k ≤K y and 0≤ i ≤K J.All the required qualities, B N(Y(k)),B a(Y(k),J (i )), and B P(Y(k),J (i )), where0≤ i ≤K J for any 1≤k≤K y, areavailable once the working path and theprotection scheme is determined.With this information, we are ready todefine a quantityP (Y(k),J (i )),Y≠J (1)where P (Y(k),J (i )) = B P (Y(k),J (i ))-B a(Y(k),J (i )),and 0 ≤ i ≤ K J.If there are N nodes in the network,for all values of J, where 1≤J≤N,J ≠Y,and for all values of i, where 0≤ i ≤K J,a list in descending order can beconstructed based on P(Y(k),*) for aparticular Y(k), 1≤k≤K y. This list isgiven byP (Y(k),J 1(i 1)),P (Y(k),J 2(i 2)),…P (Y(k),J M(i M)) (2)Nwhere M =∑l =1, l≠YK l+N-1. The notationJ x is defined as J x∈(1, ...,N ),J x≠Y and 1≤ x ≤M.In the ordered list, we haveP (Y(k),J 1(i 1))≥P (Y(k),J 2(i 2))≥…≥…P (Y(k),J M(i M)) (3)For a single link or node failure in theTo CoreRRUTransportNetworkRRURRUnetwork, the protection bandwidth to bereserved for Y(k) is P (Y(k),J 1(i 1)).If the topology constraint is ignored,the upper bound for bandwidthreservation on Y(k) that is necessary tohandle m faults is the sum of the first mmembers of the ordered list. That is,m∑P (Y(x),J k(i k)) (4)k=1By using this method for reservingprotection bandwidth, the mobilebackhaul network is efficient andcompetitive. A tighter upper boundbased on topology and traffic flow isalso possible for multiple failures, andthis will be left for future discussion.5 New RAN Architecture:C-RANTo this point, there has been nochange to the traditional mobilenetwork architecture, where basestations consist of the radio unit (RRU)and baseband processing unit (BBU).The RRU and BBU are connectedwith fiber optics. The RRU is locatedwith the antenna on top of the celltower, and the BBU sits on the ground.The BBUs for many different cell sitescan be co-located to providecentralized baseband processing. Theresulting architecture is called cloudradio access network (C-RAN) and isshown in Fig. 5.Because the modulated radio signalsare transported between RRU and66ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P70

Mobile Backhaul SolutionsSpecial TopicLi Mo, Fei Yuan, and Jian YangBBU, the requirements in terms ofbandwidth, delay, and jitter are muchmore stringent than those in mobilebackhaul networks. The preferredmethod of transport between the BBUand RRU are direct fiber connection orwavelength division multiplexing (WDM)if conservation of fiber links is required.Because of fiber connectivity inadvanced mobile technology,variations, such as C-RAN, ontraditional RAN architecture becomefeasible.C-RAN architecture is still in itsinfancy. Further division of the workloadbetween BBU and RRU according totransport requirements warrants furtherstudy.6 ConclusionIn this paper, we have reviewedmobile backhaul requirement forcurrent and future mobile accesstechnologies. An MPLS-based mobilebackhaul solution is proposed, andtiming synchronization, MPLS OAM,and protection have been discussed. Anew mechanism for reservingprotection bandwidth is described. Thismechanism ensures that protectionbandwidth is efficiently reserved for asingle fault and that an upper-boundprotection bandwidth estimationmechanism is provided for multiplefaults.References[1] Requirements for Operations, Administration, andMaintenance (OAM) in MPLS Transport Networks,RFC5860, 2010.[2]ChengShang Chang, Performance guarantees incommunication networks, Springer, 2000.[3] Jean-Yves Le Boudec and Patrick Thiran, NetworkCalculus, Springer, 2001.[4] Linear Protection Switching for Transport MPLSNetworks, ITU-T G.8131, 2006.[5] Fast Reroute Extensions to RSVP-TE for LSPTunnels, RFC 4090, 2005.BiographiesLi Mo (mo.li@zte.com.cn)received his B.Eng.degree from the Department of ElectricalEngineering, Queen’s University, in 1989. Aftergraduation, he worked at IBM, Nortel, and Fujitsubefore joining ZTE in 2001. Currently, Dr. Mo is thechief architect for ZTE USA and works as marketingspecialist for ZTE headquarters in Shenzhen. Dr. Mohas worked in the networking industry for more than20 years, and has numerous publications and U.S.patents. His research interests include fixed andmobile core networks, session control, QoS, androuting. Dr. Mo is also an active member of IEEE,ITU, ETSI, and IETF.Fei Yuan (yuan.fei@zte.com.cn) is the acting vicegeneral manager of the Department of StandardDevelopment and Industry Relations at ZTECorporation. He has been working intelecommunications for eighteen years and has astrong technical background in a range of fields. FeiYuan guides the standardization and advancedresearch teams with innovative strategies and ideas.Jian Yang (yang.jian90@zte.com.cn) joined ZTEafter receiving her bachelor's degree from XiDianUniversity. She is now the vice general standardengineer of bearer networks and has made manycontibutions to MPLS-TP standard development,both in the ITU-T and IETF.⬅ From P.62[Online]. Available:http://www.ieee802.org/16/tgm/contrib/C80216m-09_0634.doc[5] Rongzhen Yang, Ali T. Koc, PapathanassiouApostolos, W. G, Hujun Yin, Nageen Himayat,Yang-seok Choi and Shilpa Talwar,“Additionalmaterial related to the Recommended AWD TextProposal section 11.1 in C802.16m-09/0546 (PowerControl).”[Online]. Available: http://www.ieee802.org/16/tgm/contrib/C80216m-09_0703.ppt[6] Jeongho Park, Jaehee Cho, Heewon Kang, HokyuChoi, Jihyung Kim, Wooram Shin and Dong SeungKwon,“Proposed Text of Power Control Section forthe IEEE 802.16m Amendment Working Document(revision1).”[Online]. Available:http://www.ieee802.org/16/tgm/contrib/C80216m-09_0612r1.pdf[7] P. Wang, A. Boariu, Joon Chun, Xiaoyi Wang,Zexian Li,“Proposed Text on Power Control Sectionin Amendment Text.”[Online]. Available: http://www.ieee802.org/16/tgm/contrib/C80216m-09_0545.doc[8] Rongzhen Yang, Apostolos Papathanassiou, W.Guan, Ali T. Koc, Hujun Yin and Yang-seok Choi,“Supporting material for Uplink OLPC ProposalC80216m-09/0844.”[Online]. Available: http://www.ieee802.org/16/tgm/contrib/C80216m-09_0845.ppt[9] Jeongho Park, Suryong Jeong, Jaehee Cho,Heewon Kang and Hokyu Choi,“Performancecomparison between various Open Loop PowerControl schemes.”[Online]. Available: http://www.ieee802.org/16/tgm/contrib/C80216m-09_1042.pdf[10] Rongzhen Yang, Apostolos Papathanassiou, W.Guan, Ali T. Koc, Hujun Yin, Yang-seok Choi,Jeongho Park, Suryong Jeong, Jaehee Cho,Heewon Kang, Hokyu Choi, Dong-cheol Kim,Wookbong Lee, HanGyu Cho, Jihyung Kim, DongSeung Kwon, Xiaoyi Wang and Zexian Li,“IEEE802.16m Amendment Text Proposal for UplinkOpen-Loop Power Control.”[Online]. Available:http://www.ieee802.org/16/tgm/contrib/C80216m-09_0844r4.doc[11] Rongzhen Yang, Apostolos Papathanassiou, W.Guan, Hujun Yin, Yang-seok Choi, Dong-cheolKim, Wookbong Lee, HanGuy Cho and Jin SamKwak,“Uplink Power Control Rule for UplinkMulti-Streams Transmission.”[Online]. Available:http://www.ieee802.org/16/tgm/contrib/C80216m-09_1596r1.ppt[12] Rongzhen Yang, Hujun Yin, Yang-seok Choi andApostolos Papathanassiou,“Discussion of CLPCby PC-A-MAP IE for TDD and FDD.”[Online].Available: http://www.ieee802.org/16/tgm/contrib/C80216m-09_2661.ppt[13] Rongzhen Yang, Hujun Yin, Yang-seok Choi andApostolos Papathanassiou,“Uplink Power ControlOffset Design Discussion.”[Online]. Available:http://www.ieee802.org/16/tgm/contrib/C80216m-09_2660.ppt[14] Rongzhen Yang, Xinrong Wang, Hongmei Sun, Y.Zhu, Yi Hsuan, Alexei Davydov, Hujun Yin,Yang-seok Choi, Jeongho Park, Wookbong Leeand Wilson Tim,“Discussion of Control ChannelPower Control Design.”[Online]. Available:http://www.ieee802.org/16/tgm/contrib/S80216m-09_2846.ppt[15] Wookbong Lee, jeongho. park, Yang Rongzhen,Koc, Ali T; Zeira Eldad, W. Fan, robson domingos,minoh, HanGyu Cho; Dong-cheol Kim;JongYoung Han; Fred Vook, Mark Cudak, Andrebarreto, and Xiaoyi.Wang,“EvaluationMethodology for Uplink Power Control Algorithm.”[Online]. Available:http://www.ieee802.org/16/tgm/contrib/C80216m-09_1167.doc[16]“IEEE 802.16m Evaluation MethodologyDocument.”[Online].Available: http://wirelessman.org/tgm/core.html#08_004[17] Evolved Universal Terrestrial Radio Access(E-UTRA); Physical layer procedures, 3GPP TS36.213 v9.1.0.[18] Evaluation of Slow Power Control Techniques onthe System Performance of Uplink SC-FDMA,3GPP R1-061754.[19] Uplink Power Control for E-UTRA, 3GPPR1-062612.[20] Uplink Power Control for E-UTRA, 3GPPR1-062861.[21] Uplink Power Control for E-UTRA - Range andRepresentation of P0, 3GPP R1-074850.[22] Evolved Universal Terrestrial Radio Access(E-UTRA); Radio Resource Control (RRC);Protocol specification, 3GPP TS 36.331 v9.2.0.[23]“Performance of Uplink Factional Power Control inUTRAN LTE”, VTC Spring 2008, IEEE.[24] Anil M.Rao,“Reverse Link Power Control forManaging Inter-cell Interference in OrthogonalMultiple Access System”, VTC-2007 Fall, IEEE.BiographiesRongzhen Yang (rongzhen.yang@intel.com)received his B.S. and M.S.degrees in electricalengineering from Shanghai Jiaotong University in1997 and 2000. He was a software engineer atIntel's China Software Lab from 2000 to 2002, aresearcher at Intel Research from 2003 to 2004, anda wireless security architect at Intel APAC WirelessSecurity from 2005 to 2006. He is currently a systemengineer of advanced wireless technology at IntelCorporation. Rangzhen Yang’s research interestsinclude technology development for advancedwireless standards in personal, local, and wide areawireless networks. He has written more than 10journal and conference articles and has more than40 patents.Hujun Yin (hujin.yin@intel.com) received his B.S.and an M.S. degrees from Shanghai JiaotongUniversity in 1995 and 1998. He received his Ph.D.in electrical engineering from the University ofWashington in 2001. From 2001 to 2002, he was asenior researcher at AT&T Research, where hefocused on wireless network management andoptimization research. From 2002 to 2004, he was asenior wireless architect at ViVATO, where he wasresponsible for smart antenna enhanced 802.11WLAN network solutions. Dr. Yin is currently aprincipal engineer and director of advancedwireless technology at Intel Corporation. He leadstechnology development for advanced wirelessstandards in personal, local, and wide area wirelessnetworks. Dr.Yin has more than 20 journal andconference articles and has over 50 patents.December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 67D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P71

Lecture Series The Internet of Things and Ubiquitous Intelligence (4)Dongliang Xie and Yu WangThe Internet of Things andUbiquitous Intelligence (4)Dongliang XieYu Wang(State Key Laboratory of Networking and Switching Technology, Beijing University of Posts andTelecommunications, Beijing 100876, P. R. China)Editor's Desk:The traditional Internet is oriented towards person-to-person connection, whereas theInternet of things (IoT) is oriented towards connections between inanimate objects. IoTcovers a larger range of connections and involves more semantics than traditional Internet.Traditional Internet and telecom networks focus on information transfer, but IoT focuses oninformation services. By combining sensor networks, Internet, telecom networks, and cloudcomputing platform, IoT can sense, recognize, affect, and control the physical world. Thephysical world can be unified with the virtual world and human perception. In this part, wediscuss cloud computing and the cyber-physical system (CPS).services applications. It expandshardware capacity, makes softwarereconfiguration reduces easier,overheads of software virtual machinesis reduced, and supports multipleoperating systems. With virtualization,use of the server is greatly enhanced.8.2.2 Data Storage and ManagementIn cloud computing, data is highlyreliable and highly available because ofstorage redundancy. Data storagetechnologies involve cluster computing,data redundancy, and distributedstorage.The frequency of data access ismuch higher than that of data updatebecause of one characteristic of cloudcomputing—analyzing the stored andaccessing massive data [1]. Therefore,cloud data management focuses onoptimizing the data access operation.8 Cloud ComputingAdistributed CPU is the corecomponent in cloud computingand plays a pivotal role in thedevelopment of the Internet ofthings. Combining cloud computingand the IoT can boost the entireindustry and value chain if capabilityand resource sharing, quick servicedeployment, expansion of newhuman-thing interactive services, anddeep data mining are optimized. Whenthe Internet of things reaches a certainlevel, its dependence on cloudcomputing will be greater.8.1 Concept and Current StatusWith the development of electronics,communications, computers, andnetwork technologies, cloud computingis an inevitable stage in the evolution ofturing computing to network computing.Based on the Internet, networkcomputing provides applications withhardware, infrastructure, platform,software, and storage services. Theterm“cloud”vividly describes thecharacteristics of cloud computing:virtualization, transparency, scalability,and elasticity. It is an apt term fordescribing the abstraction of underlyinginfrastructure.Cloud computing integratesdistributed computing, parallelcomputing and grid computing. At itscore, it virtualizes computing, resourcesof large data centers and providesthese to users as a service.8.2 Key TechnologiesCloud computing is a form ofdata-intensive super computing. It hasunique data storage and managementtechnologies and a uniqueprogramming model.8.2.1 VirtualizationVirtualization is at the core of cloudcomputing, and is the basis for all cloud8.2.3 Parallel ProgrammingThe cloud computing programmingmodel must be very simple, andcomplicated parallel execution and taskscheduling must be hidden in thebackground. The programming modelmust be transparent, enabling users todevelop programs for specificpurposes.Most cloud computing systems adoptGoogle’s MapReduce [2]. This is adistributed parallel programming modeland also an efficient task-schedulingmodel.8.3 Combining with the Internet of ThingsCloud computing and IoT arecomplementary. Development of IoTrequires the powerful processing andstorage capabilities of cloudcomputing. If cloud computinginfrastructure is used to mine, process,and analyze mass data collected on theubiquitous sensing layer, IoT canquickly, accurately, and intelligently68ZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P72

The Internet of Things and Ubiquitous Intelligence (4)Lecture SeriesDongliang Xie and Yu Wangmanage and control the physical worldand provide technical support forubiquitous services. The IoT will also bethe largest user of cloud computingservices, which will, in turn, lay thefoundation for more successful cloudcomputing.9 Cyber-Physical SystemCyber-physical system (CPS) wasfirst proposed by the US NationalScience Foundation (NSF) in 2006. CPSis expected to become the third wave ofinformation technology following thecomputer and the Internet. CPSubiquitous sensing, control, andcomputing and the close integration ofhuman, machines, and things are theultimate goals of IoT.9.1 CPS OverviewCPS is complicated amultidimensional, embedded systemthat is integrated with computing,network and physical environments.With the integration of 3C (computation,communication, and control)technologies, CPS delivers real-timesensing, dynamic control, andinformation services for largeengineering systems. Throughinteractive, cyclic feedback incomputing and physical processes,CPS closely integrates these processesand adds or expands new functions inreal time. In this way, a physical entitycan be monitored and controlled in asecure, reliable, efficient, and real-timemanner [5].9.2 CharacteristicsCPS has the following characteristics:•It integrates computing,communication, and control because itembeds computing capability deeplyinto each physical subsystem. It aimsfor precise control over the physicalprocesses in networks.•It requires the integration ofcomputing and control technologies. Toconnect the cyber world to the physicalworld, CPS should integrate computingtechnologies (which are discreteevent-relevant and indifferent to timeand space) with the controltechnologies (which are continuousprocess-relevant and timespace-focused), enabling the discretecomputing process to interact andtightly couple with the continuousphysical process.•It is adaptive to dynamic changesin the physical world; it has powerfulreorganization and reconstructioncapabilities; it can esaily; and it is easyto connect with other CPS subsystems.•Because it has multidimensionaltime-space complexity, CPS should bean open, trustable, and predictableembedded system. The embeddedcomputing system of CPSinterconnects and interoperates withother cyber systems via the Internet.Furthermore, CPS has stepped intofields that are closely related to nationalinfrastructure and people’s daily lives.It is sensitive to security, and itstechnologies and products must beprecise and highly reliable.9.3 ChallengesTo realize the goals of CPS, there aresome technical challenges to beaddressed. A good architecture iscritical in maintaining scalability,durability, diversity and continuoustechnical innovation in CPS [3]. It is alsothe key to customer investment.To ensure sustainable developmentof CPS, the architecture should bestable, a clear roadmap should bedevised that can lead to a targetedstudy based on core CPS architecture,and a development schedule should bein place so that technical achievementsare planned and predicted. The naturaldevelopment cycle should be balanced.There are several challenges in thedevelopment of CPS. The theoreticalbasis differs from the applicationmodel. The key to the success of atechnology is its deployment in actualenvironments. It is also necessary tomaintain the integrity, reliability andsecurity of the architecture. Because ofuncertainty in the external environmentand potential changes, CPS shouldrespond to a system failure or maliciousattack automatically, autonomously,and quickly. According to thecharacterisics described in 9.2 the OSand architecture should be able tomanage redundant resources on theinterconnected device layers, monitorerrors in user applications and physicalcomponents, and quickly recover CPSfrom a system failure. Electricity CPS isa critical infrastructure, andguaranteeing its security is an importantissue.10 SummaryThe IoT, cloud computing, and CPSinteract and couple with each other. IoTis the intuitive application of CPS, andcloud computing provides technicalsupport for IoT information service. Thedemands of economy and society onthe Internet of things and CPS go farbeyond existing applications.Therefore, from the perspective ofeconomic development and technicalinnovation, IoT and CPS are a trend inthe development of global informationtechnologies and industry. They willhave a profound effect on existingindustrial structure, become the corecompetence in the new industrialstructure, and cultivate new economicgrowth.References[1] Q.Chen and Qianni Deng,“Cloud computing andits key technologies,”in Journal of ComputerApplications, vol.29,no.9,pp.2562-2567,2009.[2] J.Dean and S. Ghemawat,“MapReduce:SimpliedData Processing on Large Clusters,”in Proc. 6thUSENIX Symp. on Operation Syst. Design andImplementation (OSDI'04), New York,2004,pp.137150.[3] T.Aldridge, G.Allee and A.Gorius,“Syst.Architecture and Industry Structure as InterrelatedFoundations for Progress in Cyber-physical EnergySyst.,”in Proc.National Workshop on ResearchDirections for Future Cyber-physical Energy Syst.,Baltimore,2009.BiographiesDongliang Xie (xiedl@bupt.edu.cn) is a directorand associate professor at the Broadband NetworkCenter of State Key Laboratory of Networking andSwitching Technology, Beijing University of Postsand Telecommunications. He researches wirelessand mobile network technologies, wireless sensornetworks, mobile Internet QoS, network cooperation,and ubiquitous intelligence. He has published morethan 40 papers.Yu Wang (wang0yu@gmail.com) is a master’sstudent at the State Key Laboratory of Networkingand Switching Technology, Beijing University ofPosts and Telecommunications. She researchesconvergence of wireless sensor network andubiquitous network.December 2011 Vol.9 No.4 ZTE COMMUNICATIONS 69D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P73

GUEST PAPERZTE CommunicationsTable of Contents Volume 9, Numbers 1-4, 2011Volume-Number-PageChinese Tri-Network Convergence: Characteristics and Challenges………………………………………………Hequan Wu 9-1-01SPECIAL TOPICSMobile Cloud Computing and ApplicationsGuest Editorial ……………………………………………………………………………………………………… Chengzhong Xu 9-1-03A Survey of Mobile Cloud Computing ………………………………………… Xiaopeng Fan, Jiannong Cao, and Haixia Mao 9-1-04Mirroring Smartphones for Good: A Feasibility Study ……… Bo Zhao, Zhi Xu, Caixia Chi, Sencun Zhu, and Guohong Cao 9-1-09A Cloud-Based Virtualized Execution Environment for Mobile Applications…………………………………Shih-Hao Hung,Tei-Wei Kuo, Chi-Sheng Shih, Jeng-Peng Shieh, Chen-Pang Lee, Che-Wei Chang, and Jie-Wen Wei 9-1-15Building a Platform to Bridge Low End Mobile Phonesand Cloud Computing Services……………………………………………Fung Po Tso, Lin Cui, Lizhuo Zhang, and Weijia Jia 9-1-22WiFace: A Secure Geosocial Networking SystemUsing Wi-Fi Based Multihop MANET …………………… Lan Zhang, Xuan Ding, Zhiguo Wan, Ming Gu, and Xiangyang Li 9-1-27A Case for Cloud-Based Mobile Search …………………… Yan Gao, Li Fu, Zhenwei Zhang, Shengmei Luo, and Ping Lu 9-1-33An On-Demand Security Mechanism for Cloud-BasedTelecommunications Services ………………………… Zhaoji Lin, Ping Lu, Shengmei Luo, Feng Gao, and Jianyong Chen 9-1-37Microwave/RF Technologies for Future Wireless CommunicationsGuest Editorial ………………………………………………………………………………………… Ke-li Wu and Keqiang Zhu 9-2-01RF Technologies and Challenges for Future MBR Systemsin Cellular Base Stations …………………… Hongyin Liao, Baiqing Zong, Jianli Wang, Keqiang Zhu, and Changjiang Cao 9-2-02Broadband Power Amplifiers for Unified Base Stations ………………………………………………………… Pengcheng Jia 9-2-08Single Mode DR Filters for WirelessBase Stations ……………………………………… Ji-Fuh Liang, Guo-Chun Liang, Marco Song, George He, and Tony An 9-2-12Design of a Magneto-Electric Dipole Element for Mobile CommunicationBase Station Antennas ………………………………………………………………………… …Hang Wong and Kwai Man Luk 9-2-20Advanced Synthesis Techniques for Microwave Filters …………………………………………………… Richard J Cameron 9-2-27Advances in Digital Front-End and Software RF Processing: Part IGuest Editorial………………………… Jun Fang, Fa-Long Luo, Mikko Valkama, Serioja Ovidiu Tatu, and Tomohisa Wada 9-3-01Adaptation of a Digitally Predistorted RF Amplifier Using Selective Sampling …………………………… R. Neil Braithwaite 9-3-03A New Two-Branch Amplification Architecture and its Applicationwith Various Modulated Signals………………………………………………………W. Hamdane, A. B. Kouki, and F. Gagnon 9-3-13FPGA Implementation of a Power Amplifier Linearizerfor an ETSI-SDR OFDM Transmitter …………………………………………………………… Suranjana Julius and Anh Dinh 9-3-22Design Technologies for Silicon-Based High-EfficiencyRF Power Amplifiers: A Brief Overview……………………………………Ruili Wu, Jerry Lopez, Yan Li, and Donald Y. C. Lie 9-3-28Multi-Gbit/s 60 GHz Transceiver Analysis Using FDM Architectureand Six-Port Circuit …………………………………… Nazih Khaddaj Mallat, Emilia Moldovan, Serioja O. Tatu, and Ke Wu 9-3-36Millimeter-Wave Heterodyne Six-Port Receiver: New Implementationand Demodulation Results…………………………………………………………… D. Hammou, E. Moldovan, and S. O. Tatu 9-3-42Advances in Digital Front-End and Software RF Processing: Part IIGuest Editorial………………………… Jun Fang, Fa-Long Luo, Mikko Valkama, Serioja Ovidiu Tatu, and Tomohisa Wada 9-4-01Polyphase Filter Banks for Embedded Sample Rate Changesin Digital Radio Front-Ends …………………………… Mehmood Awan, Yannick Le Moullec, Peter Koch, and Fred Harris 9-4-03December 2011 Vol.9 No.4 ZTE COMMUNICATIONS ⅠD:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P74

ZTE CommunicationsTable of Contents Volume 9, Numbers 1-4, 2011Volume-Number-PageDesign of Software-Defined Down-Conversion and Up-Conversion:An Overview………………………………………………………… Yue Zhang, Li-Ke Huang, Carsten Maple, and Qing Xuan 9-4-10Practical Non-Uniform Channelization for Multi-StandardBase Stations…………………………………………………………Álvaro Palomo Navarro, Rudi Villing, and Ronan J. Farrell 9-4-15Crest Factor Reduction for OFDM Using Selective Subcarrier Degradation ……………………………… R. Neil Braithwaite 9-4-25An Antenna Diversity Scheme for Digital Front-Endwith OFDM Technology…………………………………………………… Fa-Long Luo, Ward Williams, and Bruce Gladstone 9-4-32A Histogram-Based Static-Error Correction Techniquefor Flash ADCs ………………………………… Armin Jalili, J Jacob Wikner, Sayed Masoud Sayedi, and Rasoul Dehghani 9-4-35Advances in Mobile Data CommunicationsGuest Editorial ……………………………………………………………………………………………………Sean Cai and Li Mo 9-4-42Enhanced Cell-Edge Performance with Transmit Power-Shapingand Multipoint, Multiflow Techniques …………………………… Philip Pietraski, Gregg Charlton, Rui Yang and Carl Wang 9-4-43Spatial Load Balancing in Wide-Area WirelessNetworks…………Kambiz Azarian, Ravindra Patwardhan, Chris Lott, Donna Ghosh, Radhika Gowaikar, and Rashid Attar 9-4-49Uplink Power Control for MIMO-OFDMA Cellular Systems…………………………………… Rongzhen Yang and Hujun Yin 9-4-55Mobile Backhaul Solutions…………………………………………………………………………Li Mo, Fei Yuan, and Jian Yang 9-4-63RESEARCH PAPERSMultifrequency Networking Solution for TD-SCDMA ……………………………… Min Jin, Wenbo Wang, and Mugen Peng 9-1-41ZP-CI/OFDM: A Power Efficient Wireless Transmission Technology …………………… Pei Gao, Xiaohu Chen, Jun Wang 9-1-45Research on the Next Generation Naming System …………………………………………………Fuhong Lin, Changjia Chen 9-1-49Privacy-Preserving Protocol for Data Stored in the Cloud …………………………… Hongyi Su, Geng Yang, and Dawei Li 9-2-36A Mobility Management Solution Based on ID/Locator Separation ………… Yuhong Li, Yunjing Hou, and Shiduan Cheng 9-2-39Self-Adaptive QoS Control in Cognitive Networks That Is Basedon Service Awareness………………………………………………………………Chengjie Gu, Shunyi Zhang, and Yanfei Sun 9-2-44Security Service Technology for Mobile Networks ………………………………………… Aiqun Hu, Tao Li, and Mingfu Xue 9-3-49DEVELOPMENT FIELDSAdaptive Multiantenna Technology …………………………………………………Huahua Xiao, Dengkui Zhu, and Liujun Hu 9-1-54A P2PSIP System with Intelligent Routing Functionon the Media Plane ……………………………………………… Yongsheng Hu, Zhenwu Hao, Jun Wang, and Naibao Zhou 9-2-49Research on LTE Network Coverage Planning ………………………………………………………… Jun Gu and Ren Sheng 9-3-55OPERATIONAL APPLICATIONSGreen Base Station Solutions and Technology ………………………………………………… Zhiping Chen and Licun Wang 9-1-58Architecture and Key Technology of Distributed IntelligentOpen Systems ………………………………………………………………… Xiaoyu Tong, Yunyong Zhang, and Bingyi Fang 9-2-53Cloud Computing in Mobile Communication Networks……………………………………………………………Xinzhi Ouyang 9-3-59LECTURE SERIESThe Internet of Things and Ubiquitous Intelligence (1) ……………………………………………… Dongliang Xie, Yu Wang 9-1-62The Internet of Things and Ubiquitous Intelligence (2) ……………………………………………… Dongliang Xie, Yu Wang 9-2-58The Internet of Things and Ubiquitous Intelligence (3) ……………………………………………… Dongliang Xie, Yu Wang 9-3-63The Internet of Things and Ubiquitous Intelligence (4) ……………………………………………… Dongliang Xie, Yu Wang 9-4-68ⅡZTE COMMUNICATIONSDecember 2011 Vol.9 No.4D:\EMAG\2010-05-26/VOL8\COVER.VFT——5PPS/P75