1 Improving the Spectral Efficiency of Nonlinear Satellite Systems through Time-Frequency Packing and Advanced Processing Amina Piemontese, Member, IEEE, Andrea Modenini, Student Member, IEEE, Giulio Colavolpe, Senior Member, IEEE, and Nader Alagha, Member, IEEE Abstract—We consider realistic satellite communications sys- relaxing the orthogonality condition. Faster-than-Nyquist sig- tems for broadband and broadcasting applications, based on naling(FTN,see[2]–[5])isawellknowntechniqueconsisting frequency-division-multiplexed linear modulations, where spec- ofreducingthetimespacingbetweentwoadjacentpulseswell tral efficiency is one of the main figures of merit. For these below that ensuring the Nyquist condition, thus introducing systems, we investigate their ultimate performance limits by 3 using a framework to compute the spectral efficiency when ISI. If the receiver is able to cope with the interference, the 1 suboptimalreceiversareadoptedandevaluatingtheperformance efficiency of the communication system will be increased. In 0 improvements that can be obtained through the adoption of theoriginalpapersonFTNsignaling[2]–[5],thisoptimaltime 2 the time-frequency packing technique. Our analysis reveals that spacing is obtained as the smallest value giving no reduction n introducingcontrolled interferencecan significantlyincrease the oftheminimumEuclideandistancewithrespecttotheNyquist a efficiency of these systems. Moreover, if a receiver which is able J toaccountfortheinterferenceandthenonlinearimpairmentsis case. This ensures that, asymptotically, the ISI-free bit-error 7 adopted, rather than a classical predistorter at the transmitter rate (BER) performanceis reached, at least when the optimal 1 coupled witha simplerreceiver, thebenefitsin terms of spectral detector is adopted. The i.u.d. capacity or information rate efficiencycanbeevenlarger.Finally,weconsiderpracticalcoded (IR), i.e., the average mutual information when the channel schemesandshow thepotentialadvantages of theoptimizedsig- ] inputs are independent and uniformly distributed (i.u.d.) ran- T nalingformatswhencombinedwithiterativedetection/decoding. dom variables, is then computed, still assuming the adoption I . of the optimal detector [6], [7]. However, the complexity of s Index Terms—Nonlinear satellite channels, interchannel in- c this optimal detector easily becomes unmanageable, and no terference, intersymbol interference, information rate, spectral [ efficiency. hints are provided on how to perform the optimization in the 1 more practical scenario where a reduced-complexity receiver v is employed. 4 I. INTRODUCTION In [8], for the AWGN channel, a different approach for 8 improvingtheSEhasbeenconsidered.Theapproachrelieson 1 In satellite systems, orthogonal signaling is often adopted bothtimepackingofadjacentsymbolsandreducingthecarrier 4 to avoid intersymbolinterference(ISI),at least in the absence 1. of nonlinear distortions. For example, in the 2nd-generation spacingoftheadjacentchannelswhenapplicable(multicarrier transmission), thus introducing also interchannel interference 0 satellite digital video broadcasting (DVB-S2) standard [1], a (ICI). In [8] it is assumed that a symbol-by-symbol detector 3 conventional square-root raised-cosine (RRC) pulse shaping 1 workingonthesamplesatthematchedfilteroutputisadopted filter is specified at the transmitter. For an additive white : atthereceiverside,andthecorrespondingIRiscomputed,by v Gaussian noise (AWGN) channel and in the absence of other i impairments,theuseofaRRCfilteratthereceiverandproper also optimizing time and frequencyspacings to maximize the X achievable SE. Hence, rather than the minimum distance, and sampling ensure that optimal detection can be performed on r thus the BER, the SE is the performance figure of merit. In a a symbol-by-symbol basis. On the other hand, it is known addition, a low-complexitymemorylessreceiveris considered that, when finite-order constellations are considered (e.g., rather than the optimal detector employed in [2]–[7]. More phase-shift keying (PSK)), the spectral efficiency (SE) of the complexdetectionalgorithmshavebeenalsoconsideredin[9] communication system, defined here as the information rate for the case of linear channels. normalized to the spectral bandwidth assigned to the channel In this work, we apply the time-frequency packing (TF and the time spent to transmit a symbol, can be improved by packing) technique to nonlinear satellite channels. In particu- lar, we design highly efficient schemes by choosing the time G.Colavolpe,A.Modenini,andA.PiemontesearewitharewithConsorzio Nazionale Interuniversitario perleTelecomunicazioni (CNIT)andUniversita` andfrequencyspacingswhichgivethemaximumvalueofSE. diParma,Dipartimento diIngegneria dell’Informazione, VialeG.P.Usberti, We assume a realistic satellite channel where nonlinear dis- 181A, I-43124 Parma, Italy, e-mail: [email protected], [email protected], tortionsoriginatefromthe presenceof a high-poweramplifier [email protected]. N. Alagha is with the European Space Agency (ESA-ESTEC),Noordwijk, TheNetherlands (HPA). The considered system is also affected by ISI, due ThisworkisfundedbytheEuropeanSpaceAgency,ESA-ESTEC,Noord- to the presence of input and output multiplexing (IMUX and wijk,TheNetherlands, undercontract no.4000102300. OMUX)filtersplacedbeforeandaftertheHPAandintention- The paper was presented in part at the IEEE Intern. Conf. Commun. (ICC’12), Ottawa, Canada, June2012. ally introducedby the adoptionof the time packingtechnique 2 aswell.AlthoughICIisalsopresentduetofrequencypacking, Satellitetransponder we limit our investigation to systems in single-carrier-per transponderoperation(i.e., each transponderis devoted to the x(t) h(ℓ)(t) HPA h(ℓ)(t) i o amplification of the signal coming from only one user) where a single-userreceiveris employed,and considertwo different Fig.1. Satellite transponder foruser ℓ. approaches to detection for nonlinear channels, namely the use of a detector taking into account the nonlinear effects and a more traditional scheme based on predistortion and where x(ℓ) is the symbol transmitted by user ℓ during the memorylessdetection.Inthecaseofpredistortion,weconsider k kth symbolinterval,and F is the frequencyspacing between u the dynamic data predistortion technique described in [10], adjacentchannels.1Thetransmittedsymbolsbelongtoagiven [11], whereas in the case of advanced detection we consider zero-meanM-arycomplexconstellation.InDVB-S2 standard a receiver based on an approximate signal model described the base pulse p(t) is an RRC-shaped pulse with roll-off in [12]. In this latter case, we also employ the channel short- factor α (equal to 0.2, 0.3, or 0.35 depending on the service ening technique (CS) [13], recently proposed for nonlinear requirements).Noticethat,inordertoleaveoutbordereffects, satellite channels, to limit the complexity of the detection the summations in (1) extend from −∞ to +∞, namely an algorithm. It should be noted that we apply the TF packing infinite number of time epochs and carriers are considered. technique to nonlinear satellite channels for which, usually, As commonly assumed for broadband and broadcasting even by using RRC pulses there is still ISI at the receiver. systems, on the feeder uplink (between the gateway and the The proposed TF packing technique promises to provide satellite) the impact of thermal noise can be neglected due increased SEs at least for low-order modulation formats. In to a high transmit signal strength. Hence, in our analysis, we fact, when dense constellations with shaping are employed, have considered a noiseless feeder uplink. Although the TF we fall in a scenario similar to that of the Gaussian channel packing can be applied to other and more general scenarios, with Gaussian inputs for which orthogonal signaling with we consider here a single-carrier per transponder scenario, no excess bandwidth (rectangular shaping pulses) is optimal where differentcarriersundergoindependentamplification by (although this is mainly true for the linear channel and not different transponders on board of satellite, each of which in the presence of a nonlinear HPA, since shaping increases works with a single carrier occupying its entire bandwidth. thepeak-to-averagepowerratio).ImprovingtheachievableSE In this case, the on-board power amplifier can operate closer withoutincreasingthe constellationordercanbe considerably tosaturationandhenceimproveitsefficiency.Thetransponder convenientsinceitiswellknownthatlow-orderconstellations model for user ℓ, shown in Fig. 1, is composed of an IMUX are more robust to channel impairments such as time-varying filter h(ℓ)(t) which selects the ℓth carrier, an HPA, and an phase noise and non-linearities. It is expected that the use of i OMUXfilterh(ℓ)(t)whichreducestheout-of-bandpowerdue low-ordermodulationinconjunctionwithTFpackingprovides o to the spectral regrowth after nonlinearamplification [1]. The similar advantages in terms of robustness against channel HPA is a nonlinear memoryless device defined through its impairments. AM/AMandAM/PMcharacteristics,describingtheamplitude The proposed approach to improve the SE is very general and phase distortions caused on the signal at its input. andthe caseofsatellite systemsforbroadbandandbroadcast- The outputsof differenttranspondersare multiplexedagain ing applications must be thus considered just as an example in the downlink to form the signal s(t), and we assume to illustrate the benefits that can be obtained through the that the adjacent users have a frequency separation of F , applicationoftheTFpackingparadigmcoupledwithadvanced d usually equal to that in the uplink. The useful signal at the receiver processing. user terminal is still the sum of independent contributions, The remainder of this paper is organized as follows. In one for each transponder (although these contributions are Section II, we introduce the system model. The framework no more, rigorously, linearly modulated due to the nonlinear that we use to evaluate the SE of satellite systems is detailed transformation of the on-board HPA). The received signal inSectionIII,whereasdifferentapproachestothedetectionfor is also corrupted by the downlink AWGN, whose low-pass theconsideredchannelaredescribedinSectionIV.Numerical equivalent w(t) has power spectral density (PSD) 2N . The results are reported in Section V, where we show how the 0 low-passequivalentof the receivedsignalhas thusexpression proposed technique can improve the SE of DVB-S2 systems. Finally, conclusions are drawn in Section VI. r(t)=s(t)+w(t). II. SYSTEMMODEL We evaluate the ultimate performance limits of this com- We consider the forward link of a transparent satellite munication system when single-user detection is employed at system, where synchronous users employ the same linear thereceiverside.Theproposedtechniqueconsistsofallowing modulation format, shaping pulse p(t), and symbol interval interference in time and/or frequency by reducing the values (or time spacing) T, and access the channel according to of T, Fu, and Fd, (partially) coping with it at the receiver, a frequency division multiplexing scheme. The transmitted in order to increase the SE. In other words, T, Fu, and Fd signal in the uplink can be expressed as are chosen as the values that give the maximum value of the x(t)= x(ℓ)p(t−kT)ej2πℓFut, (1) 1Inthisscenario,wewillusetheterms,“channels”,“users”,and“carriers” k interchangeably. ℓ k XX 3 SE.Thesevaluesdependontheemployeddetector—thelarger The aim of the proposed technique is to find the values of the interferencethat the receivercan cope with, the larger the F and T providing,foreach value of the signal-to-noiseratio SE and the lower the values of time and frequency spacings. (SNR),themaximumvalueofSEachievablebythatparticular Notice that, since we are considering single-user detection, receiver,optimalfortheconsideredspecificauxiliarychannel. thereceiverisnotabletodealwiththeinterferenceduetothe Namely, we compute overlap of different channels. In this case, the optimization η = max η(F,T). (4) of the frequency spacings is actually an optimization of M F,T>0 the frequency guard bands generally introduced in satellite Typically, the dependence on the SNR value is not critical, systems to avoid the nonlinear cross-talk, since a single-user in the sense that we can identify two or at most three SNR receiver can tolerate only a very small amount of ICI. regions for which the optimal spacings practically have the The considered nonlinear satellite channel reduces, as a same value. particularcase,tothelinearchannel,providedthattheHPA is For fair comparisons in terms of SE, we need a proper drivenfarfromsaturation.Hence,alltheconsiderationsinthis definition of the SNR. We define the SNR as the ratio P/N paper can be straightforwardly extended to the linear channel between the transmit power when the HPAs are driven at case. A few results can be found in [9]. saturation and the noise power(in the consideredbandwidth). Denoting by U the number of users and by W the signal III. OPTIMIZATION OF THESPECTRALEFFICIENCY bandwidth, P/N can as written as We describe the framework used to evaluate the ultimate P UP P = lim sat = sat , (5) performance limits of the considered satellite system and to N U→∞N0((U −1)F +W) N0F perform the optimization of the time and frequency spacings. where P is the peak power at the output of an HPA in To simplify the analysis, we will assume F = F = F. We sat u d response to a continuous wave input, denoted as the ampli- perform this investigation by constraining the complexity of fier saturation power. P is independent of the bandwidth theemployedreceiver.Inparticular,asmentioned,weassume sat W. The SNR definition as given in (5) is independent of that a single-user detector is used. For this reason, without the transmit waveform and its parameters. This provides a loss of generality we only consider the detection of symbols x(0) = {x(0)} of user with ℓ = 0.2 In addition, we also common measure to compare the performance of different k solutions in a fair manner. The output back-off for each consider low-complexity receivers taking into account only a waveform (or modulation scheme) is defined as the power portionoftheactualchannelmemory.Undertheseconstraints, ratio (in dBs) between the unmodulated carrier at saturation we compute the IR, i.e., the average mutual information and the modulated carrier after the OMUX. We point out whenthechannelinputsarei.u.d.randomvariablesbelonging that equations (3)-(5), although derived for nonlinear satellite to a given constellation. Provided that a proper auxiliary channels, can be used for linear AWGN channels also, by channelcan be definedforwhichthe adoptedlow-complexity replacing P in (5) with the user’s transmitted power. receiver is optimal, the computed IR represents an achievable sat Without loss of generality, T and F in (3)-(5) can be lower bound of the IR of the actual channel, according to normalized to some reference values T and F . We will mismatched detection [14]. B B denoteν =F/F andτ =T/T .Inthenumericalresults,we Denoting by y(0) a set of sufficient statistics for the detec- B B will chooseT and F as the symboltime andthe frequency tion of x(0), the achievable IR, measured in bit per channel B B spacingadoptedin the DVB-S2standard,whichis considered use, can be obtained as as a benchmark scenario. 1 p(y(0)|x(0)) I(x(0);y(0))= lim E log , (2) K→∞K ( p(y(0)) ) IV. AUXILIARYCHANNELMODELS The system model described in Section II is representative where K is the numberof transmittedsymbols.Theprobabil- of the considered scenario and has been employed in the ity density functions p(y(0)|x(0)) and p(y(0)) are computed information-theoretic analysis and in the simulations results. by using the optimal maximum-a-posteriori (MAP) symbol In this section, we describe the employed auxiliary channel detectorfortheauxiliarychannel,while theexpectationin (2) modelsandthecorrespondingoptimalMAPsymboldetectors. is with respect to the input and output sequences generated As explained in Section III, they are used to compute two accordingtotheactualchannelmodel[15].Inthenextsection, lowerboundsontheSEfortheconsideredchannel[15].Since we will discuss two different low-complexity detectors for these lower boundsare achievable by those receivers,we will nonlinearsatellitechannelsandwewilldefinethecorrespond- say that the computed lower bounds are the SE values of the ing auxiliary channels. considered channel when those receivers are employed. We can define the user’s bandwidth as the frequency sepa- ration F between two adjacent carriers. The achievable SE is A. Memoryless model and predistortion at the transmitter thus η = 1 I(x(0);y(0)) [b/s/Hz]. (3) When the HPA AM/AM and AM/PM characteristics are FT properly estimated and feeded back to the transmitter, the sequence of symbols {x(ℓ)} can be properly predistorted 2Assuming a system with an infinite number of users, the results do not k dependonaspecificuser. to form the sequence {x′(ℓ)} that is transmitted instead, k 4 in order to compensate for the effect of the non-linearity from the classical one by neglectingsome selected terms. For and possibly to reduce the ISI. Here we consider the dy- further details, the reader can refer to [12]. We are looking namic datapredistortiontechniquedescribedin [10], [11] and for an auxiliary channel and the corresponding optimal MAP also suggested for the application in DVB-S2 systems [1], symbol detector. As mentioned, only single-user receivers where the symbol x′(ℓ) transmitted by user ℓ at time k is are considered here. Hence, we will assume that, apart from k a function of a sequence of 2L + 1 input symbols, i.e., AWGN, only user ℓ = 0 is present and that the simplified p x′ = f(x ,...,x ,...,x ). The mapping f at the model (7) holds. Optimal single-user MAP symbol detection k k−Lp k k+Lp transmitter is implemented through a look-up table (LUT), for this case can be performed through a bank of filters which is computed through an iterative procedure performed followed by a conventional BCJR detector [18] with proper off-line and described in [10], [11]. This procedure searches branchmetricsandworkingonatrelliswhosenumberofstates the best trade-off between the interference reduction and exponentially depends on the channel memory [12]. the increase of the OBO. The complexity at the transmitter Inthe case of PSK modulations,being|x |2 =1,the signal i dependson the numberof symbolsaccounted for throughthe model(7)simplifiestoalinearmodulationwithshapingpulse parameter L . The transmitted signal is thus h¯(t) = NV−1h(2i+1)(t) [12]. An approximate (being the p i=0 modelin (7) an approximation)set of sufficient statistics y(0) x(t)= x′k(ℓ)p(t−kT)ej2πℓFt canthusPbeobtainedbysamplingtheoutputofafiltermatched Xℓ Xk to h¯(t). Under the assumption that only user with ℓ = 0 is whereas, at the receiver, a simple single-user memoryless present, the kth element of y(0) is channelisassumedcorrespondingtotheauxiliarychannel(for user with ℓ=0) yk(0) = x(k0−)igi+ηk, yk(0) =x(k0)+nk (6) Xi where where n is azero-meancircularlysymmetricwhiteGaussian k g = h¯(t)h¯∗(t−iT)dt process with PSD 2(N +N ), N being a design parameter i 0 I I which can be optimized through computer simulations—an Z andη isaGaussianprocesswithE{η η∗}=2N g .Vector increase of the assumed noise variance can improve the com- k k+i k 0 i y(0) can be written as puted achievable lower bound on the spectral efficiency [8]. y(0) =Gx(0)+η (8) B. Model with memory and advanced detection whereGisaToeplitzmatrixobtainedfromthesequence{g } i A valid alternativeto nonlinearcompensationtechniquesat whereasη is a vectorcollectingthe noisesamples. According the transmitter relies upon the adoptionof advanceddetectors to the CS approach,the consideredauxiliary channelis based which can manage the nonlinear distortions and the ISI. In on the following channel law [17] this work, we consider detection based on an approximate signalmodeldescribedin[12],whichcomesfromasimplified p(y(0)|x(0))∝exp 2R(y(0)†Hrx(0))−x(0)†Grx(0) , Volterra series expansion of the nonlinear channel. In the (9) (cid:16) (cid:17) following, we will consider PSK and amplitude/phase shift where Hr and Gr are Toeplitzmatricesobtainedfromproper keying (APSK) modulations typically employed in satellite sequences{hr}and{gr},andareknownaschannelshortener i i systems [1]. To limit the receiver complexity with a limited andtargetresponse,respectively[17].Matrix Hr representsa performance degradation, we also apply a CS technique [16]. linearfilteringofthesufficientstatistics(8),and Gr istheISI In fact, when the memory of the channel is too large to be to be set at the detector (different from the actual ISI) [17]. taken into account by a full complexity detector, an excellent In (9), the noise variance has been absorbed into the two performancecanbeachievedbyproperlyfilteringthereceived matrices. In order to reduce the complexity, we constrain the signal before adopting a reduced-state detector [16]. A very target response used at the receiver to effectiveCStechniqueforgenerallinearchannelsisdescribed in [17], while its extension to nonlinear satellite systems is gr =0 |i|>L , (10) i r reported in [13]. which implies that the memory of the detector is L instead The approximatemodel (employedfor receiverdesign pur- r poses only) of the contribution s(0)(t) of user ℓ = 0 to the of the true memory of the channel. The CS technique finds a closed form of the optimal {hr} and {gr} which maximize receivedusefulsignals(t),canbebasedonthefollowingvth- i i the achievable IR (2). We point out that if the memory L is order(with v beingany oddinteger) simplified Volterra-series r larger than or equal to the actual channel memory the trivial expansion [12] solution is Hr = I/2N and Gr = G/2N , where I is the 0 0 NV−1 identity matrix. s(0)(t)≃ x(0) |x(0)|2ih(2i+1)(t−nT) , (7) n n For APSK modulations, an approximate set of sufficient Xn Xi=0 h i statistics can be obtained through a bank of filters matched whereN =(v+1)/2,andh(2i+1)(t)arecomplexwaveforms to the pulses h(2i+1)(t), i=0,...,N −1. The CS technique V V givenbyalinearcombinationsofthetheoriginalN Volterra has been recently extended to this case in [13], and leads to V kernels. This simplified Volterra-series expansion is obtained an auxiliary channel whose law has the same form of (9), 5 r(t) h¯∗(−t) Hr 5 a) detGe trion QPSK 8PSK 16APSK r(t) h(1)∗(−t) soft 4 32APSK Hr inform. 64APSK b) ... NV ×NV ... detGe rtion Hz] 3 h(v)∗(−t) b/s/ η [ 2 Fig. 2. Block diagram of the considered suboptimal receivers for a) PSK andb)APSKmodulations. 1 0 where the channelshortenerand the targetresponseare block -12 -8 -4 0 4 8 12 16 20 24 matrices (see [13] for details). Psat/N0F [dB] The resulting receivers for PSK and APSK modulations are shown in Fig. 2. They are optimal for the considered Fig. 3. Spectral efficiency of DVB-S2 modulations with roll-off 0.2, data auxiliary channel models. Interestingly, when L = 0 the r predistortion, andmemoryless detection. Comparison with aconstellation of optimalchannelshortenerbecomestheminimummeansquare increased cardinality (64APSK). error (MMSE) feedforward equalizer of [19], applied to the signal model (7). cardinality,andinFig.3wealsoshowtheSEforthe64APSK V. NUMERICAL RESULTS modulation[20]. It can be seen that the 64APSK modulation, due to the higher impact of the nonlinearities, allows to A. Spectral efficiency increase the SE only at high SNR values and it seems there Considering the DVB-S2 system as a benchmark scenario, is no hope to improve the SE in the low and medium SNR we now show the improvement, in terms of SE, that can be regions. obtained by adopting the TF packing technique joint with an An alternative way of improving the SE is based on a advanced processing at the receiver. Let us consider a typical reduction of the roll-off factor. In Fig. 4, we consider QPSK DVB-S2 scenario where, at the transmitter, the shaping pulse and16APSK modulationsin a scenariowherepredistortionat p(t) has a RRC spectrum, whereas the IMUX and OMUX thetransmitterandsymbol-by-symboldetectionatthereceiver filters and the nonlinear characteristics of the HPA are those are still employed. We show the SE improvementthat can be reported in [1, Figs. H.12 and H.13]. The standard considers obtainedbyreducingtheroll-offtoα=0.05.4Wecanobserve the following modulation formats: QPSK, 8PSK, 16APSK, that the roll-off reduction improves the SE with respect to and 32APSK. To combat ISI and nonlinear distortions, a DVB-S2 for all SNR values. On the other hand, as shown data predistorter is employed at the transmitter whereas at in Fig. 4, better results can be obtained by allowing TF the receiver a symbol-by-symbol detector is assumed. Here, packing.ThevaluesofT andF arechosenasthoseproviding we consider the predistorter described in Section IV-A, with the largest SE. This search is carried out by evaluating (4) L =2 for QPSK, 8PSK 16APSK, andL =1 for 32APSK. p p on grid of values of T and F (coarse search), followed by TheSEresultshavebeenobtainedbycomputingtheIR in(2) interpolation of the obtained values (fine search). We point by means of the Monte Carlo method described in [15]. For out that these curves have been obtained without reducing each case, we also performed a coarse optimization of the the roll-off factor, which is still α = 0.2, and employing the noise variance to be set at receiver [8], and of the amplifier same predistorter and symbol-by-symbol receiver adopted in operation pointthroughthe OBO. Unless otherwise specified, the DVB-S2 system. the roll-off factor of the RRC pulses is α=0.2, which is the With the aim of further improving the performance, we lowest value considered in the standard. now consider TF packing and a system without predistortion We first consider the achievable SE of our benchmark at the transmitter but using the advanced detection algorithm scenario,andin Fig.3we report η asafunctionofP /N F sat 0 described in Section IV-B, joint with CS (L = 1). The r forthefourmodulationformatsofthestandard(QPSK,8PSK, Volterra order of the model (7) is v = 5. The results for 16APSK, and 32APSK). We verified that comparable SE val- QPSK and16APSKmodulationsarereportedinFig.5,where ues can be also obtained by using, instead of the predistorter, we also showthe DVB-S2benchmarkcurvesdiscussedabove the advanced detection scheme of Section IV-B with L = 0 r andthecurvesrelatedtoTFpackingwhenpredistortionatthe (MMSE detection). We also consider two alternative ways transmitterandmemorylessdetectionat the receiverare used. that, at least in the case of a linear channel, can be used Theseresultsshowtheimpressiveimprovementachievableby to improve the SE without resorting to TF packing.3 The simplest approach relies on the increase of the modulation 4We properly modify the transmitted signal such that it occupies the same bandwidth as that of the signal with roll-off 0.2. We verified that no 3On a nonlinear satellite channel, due to the increased peak-to-average improvement can be obtained by resorting to a more sophisticated receiver power ratio, their application must be carefully considered since not nec- based on linear ornonlinear equalization in addition to orin substitution of essarilyproduces theexpected benefits. thepredistorter. 6 5 4 QPSK, DVB-S2 QPSK, T pack. QPSK, α=0.05 3.5 QPSK, W opt. QPSK, TF pack. QPSK, T pack., W opt. 4 16APSK, DVB-S2 3 QPSK, TF pack. 16APSK, α=0.05 QPSK, TF pack., W opt. z] 3 16APSK, TF pack. z] 2.5 H H b/s/ b/s/ 2 [M 2 [M η η 1.5 1 1 0.5 0 0 -12 -8 -4 0 4 8 12 16 20 24 -12 -8 -4 0 4 8 12 16 20 24 Psat/N0F [dB] Psat/N0F [dB] Fig. 4. Improvements, interms ofspectral efficiency, that can be obtained Fig. 6. Spectral efficiency of QPSK modulation with TF packing and forQPSK and 16APSK modulations byreducing ofthe roll-off (α=0.05) bandwidthoptimizationbyadoptingtheadvancedreceiverwithCS(Lr =1). or by adopting the TF packing technique. In all cases, predistortion at the transmitter andmemoryless detection atthereceiver areemployed. decreasing F, this is no more true for our nonlinear channel 5 since IMUX and OMUX bandwidths are kept fixed. Hence, QPSK, DVB-S2 an increased value of W also increases the ISI. The benefit QPSK, TF pack., pred. QPSK, TF pack., adv. det. of the bandwidth optimization is twofold: it can be used as 4 16APSK, DVB-S2 an alternative to frequency packing (e.g., in cases where the 16APSK, TF pack., pred. z] 3 16APSK, TF pack., adv. det. frequenciesoftheon-boardoscillatorscannotbemodifiedand, H hence, frequency packing is not an option), or it can be used s/ b/ to improve the results of TF packing. In Fig. 6, we consider [M 2 QPSK modulationandtheadvancedreceiverwith L =1.As η r expected, the combination of TF packing with the bandwidth 1 optimizationgivesthebestresults.Wealsoshowtheresultsin caseonlytimepackingoronlythebandwidthoptimizationare adopted.Interestingly,theSE oftimepackingwith bandwidth 0 -12 -8 -4 0 4 8 12 16 20 24 optimizationisquitesimilartothatachievablebyTFpacking. Psat/N0F [dB] Finally,tosummarizetheresults,Fig.7showstheSEforall DVB-S2 modulations (QPSK, 8PSK, 16APSK and 32APSK) with TF packing, bandwidth optimization, and the advanced Fig. 5. Spectral efficiency for QPSK and 16APSK modulations with TF receiver. For clarity, we show only one curve which, for each packing and advanced detector (TF pack., adv. det.). Comparison with the abscissa, reports only the largest value of the four curves DVB-S2 scenario and the case of TF packing when a predistorter and a symbol-by-symboldetector areadopted(TFpack.,pred.). (the “envelope”). In the same figure, we also plot three other SE curves obtained by using predistortion and a memoryless receiver. The lowest one is that corresponding to the DVB- TF packing combined with the considered advanced receiver, S2 scenario (one curve which is the “envelope” of all four which, with a memory of only one symbol, can cope with curves in Fig. 3), the SE curve for the 64APSK modulation, much more interference than the schemes employing the andtheSEcurveincaseofroll-offα=0.05reduction.Inthis predistorter and a memoryless detector. latter case, we considered all modulationswith cardinality up Havingremovedthe constraintoforthogonalsignaling,one to 64, and hence this curve represents the effect of both roll- more degreeof freedomin the SE optimizationis represented off reduction and cardinality increase with respect to DVB- by the bandwidth W of the shaping pulse p(t) (in case S2. The figure shows that TF packing and advanced receiver of orthogonality, it is W = (1 + α)/T instead).5 Hence, processing allows a SE improvement of around 40% w.r.t. guided by the same idea behind the TF packing technique, DVB-S2 at high SNR. we also optimized W, further increasing both ICI and ISI duetotheadjacentusersandtotheIMUXandOMUXfilters, B. Modulation and coding formats respectively.WhereasontheAWGNchannelthisoptimization What information theory promises can be approached by is implicit in TF packing, in the sense that we can obtain the using proper coding schemes. All the considered modulation same ICI by fixing F and increasing W or by fixing W and and coding formats (MODCODs) use the low-density parity- check (LDPC) codes with length 64800 bits of the DVB-S2 5Thesymbolrate,orequivalently thebandwidth,ofconventional DVB-S2 canalsobefurtheroptimized foraperformance improvement. standard.WeadopttheoptimizedvaluesforT,F,andW and 7 5 5 DVB-S2 SE, DVB-S2 4.5 64APSK 4.5 SE, TF pack, W opt. α=0.05 MODCODs, DVB-S2 4 4 TF pack., W opt. MODCODs, TF pack, W opt. 3.5 3.5 Hz] 3 z] 3 H [b/s/M 2 .25 η [b/s/ 2 .25 η 1.5 1.5 1 1 0.5 0.5 0 0 -12 -8 -4 0 4 8 12 16 20 24 -12 -8 -4 0 4 8 12 16 20 24 Psat/N0F [dB] Psat/N0F [dB] Fig.7. Spectral efficiency ofTFpacking withbandwidth optimization (TF Fig. 8. Modulation and coding formats of the DVB-S2 standard and pack.,W opt.).ComparisonwithDVB-S2,64APSK androll-offreduction. comparisonwiththosedesignedfortheproposedTFpackingtechniquewith optimized bandwidth. TABLEI DETAILSOFTHEMODCODSBASEDONTFPACKING. TABLEII DETAILSOFTHEDVB-S2MODCODS.INTHISCASE,τ =1ANDν=1. rate τ ν Wopt P/N0F η [dB] [b/s/Hz] rate P/N0F [dB] η [b/s/Hz] 1/3 0.833 1.00 +20% -1.7 0.53 1/2 0.1 0.66 QPSK 1/2 0.750 0.90 +20% 2.2 0.98 QPSK 3/5 1.4 0.79 3/5 0.750 0.90 +20% 3.6 1.18 3/4 3.2 0.99 1/2 0.731 0.95 +30% 5.3 1.43 3/5 4.6 1.19 8PSK 3/5 0.731 0.95 +30% 7.4 1.72 8PSK 3/4 7.3 1.49 2/3 0.731 0.95 +30% 8.5 1.91 8/9 10.0 1.77 2/3 0.792 0.90 +20% 11.1 2.48 3/4 10.9 1.99 16APSK 16APSK 3/4 0.750 0.90 +20% 14.1 2.94 4/5 12.0 2.12 2/3 0.731 0.95 +30% 15.3 3.18 5/6 12.7 2.21 3/4 0.731 0.95 +30% 17.5 3.58 8/9 14.6 2.35 32APSK 5/6 0.731 0.95 +30% 19.5 3.98 3/4 14.3 2.48 8/9 0.731 0.95 +30% 21.2 4.24 4/5 15.6 2.65 32APSK 5/6 16.5 2.76 8/9 19.2 2.94 9/10 19.9 2.98 the advanced detector described in Section IV-B. We assume that1/T =27.5MbaudandF =41.5MHz, and use these B B values to normalize the time and frequency spacings. Due to efficiency of a nonlinear satellite system employing linear the soft-input soft-output nature of the considered detection modulationswith finite constellations. As a first step, through algorithm, we can adopt iterative detection and decoding. We an information-theoretic analysis, we computed the spectral distinguishbetweenlocaliterations,withintheLDPCdecoder, efficiency achievable through this technique showing, with and global iterations, between the detector and the decoder. reference to the DVB-S2 scenario, that without an advanced Here,weallowamaximumof5globaliterationsand20local processingatthereceiver,the potentialgainsareverylimited. iterations. On the other hand, a detector which takes into account a BER resultshavebeencomputedbymeansofMonteCarlo memory of only one symbol, and thus with a very limited simulationsandarereportedintheSEplaneinFig.8using,as complexity increase, it is possible to obtain a gain up to 40% reference,aBERof10−6.Inthesamefigure,theperformance intermsofspectralefficiencywithrespecttotheconventional oftheDVB-S2MODCODsisalsoshownforcomparison.We use of the current standard. All these considerations can be recallthatforthempredistortionatthetransmitterandsymbol- extended to other channels and scenarios. by-symboldetectionatthereceiverareadopted.Moreover,for them we have τ =1 and ν =1. The details of the considered REFERENCES MODCODs are reported in Tables I and II. These results are in perfectagreementwith the theoreticalanalysisand confirm [1] ETSIEN301307DigitalVideoBroadcasting(DVB);V1.1.2(2006-06), Second generation framing structure, channel coding and modulation that the TF packing technique can provide an impressive systems for Broadcasting, Interactive Services, News Gathering and performance improvementw.r.t. the DVB-S2 standard. other Broadband satellite applications, 2006. Available on ETSI web site(http://www.etsi.org). [2] J.E.Mazo,“Faster-than-Nyquistsignaling,”BellSystemTech.J.,vol.54, VI. CONCLUSIONS pp.1450–1462, Oct.1975. [3] J. E.Mazo and H. J. Landau, “On the minimum distance problem for We have investigated the TF packing technique, joint with faster-than-Nyquist signaling,” IEEETrans. Inform. Theory, pp. 1420– anadvancedprocessingatthereceiver,toimprovethespectral 1427,Nov.1988. 8 [4] A. Liveris and C. N. Georghiades, “Exploiting faster-than-Nyquist signaling,” IEEETrans.Commun.,vol.47,pp.1502–1511,Sept.2003. [5] F. Rusek and J. B. Anderson, “The two dimensional Mazo limit,” in Proc.IEEEInternationalSymposiumonInformationTheory,(Adelaide, Australia), pp.970–974,Nov.2005. [6] F. Rusek and J. B. Anderson, “On information rates of faster than Nyquist signaling,” in Proc. IEEE Global Telecommun. Conf., (San Francisco, CA,U.S.A.),Nov.2006. [7] F. Rusek and J. B. Anderson, “Maximal capacity partial response signaling,” in Proc. IEEE Intern. Conf. Commun., pp. 821–826, June 2007. [8] A. Barbieri, D. Fertonani, and G. Colavolpe, “Time-frequency pack- ing for linear modulations: spectral efficiency and practical detection schemes,”IEEETrans.Commun.,vol.57,pp.2951–2959,Oct.2009. [9] A. Modenini, G. Colavolpe, and N. Alagha, “How to significantly improve the spectral efficiency of linear modulations through time- frequency packing and advanced processing,” in Proc. IEEE Intern. Conf.Commun.,pp.3299–3304, June2012. [10] G.Karam andH.Sari, “AData Predistortion Technique with Memory forQAMRadioSystems,”IEEETrans.Commun.,vol.39,pp.336–344, Feb.1991. [11] E.Casini,R.DeGaudenzi,andA.Ginesi,“DVB-S2modemalgorithms design and performance over typical satellite channels,” Intern. J. of Satellite Communications and Networking, vol. 22, pp. 281–318, May/June 2004. [12] G.ColavolpeandA.Piemontese,“NovelSISODetectionAlgorithmsfor Nonlinear Satellite Channels,” IEEEWireless Commun.Letters, vol. 1, pp.22–25,Feb.2012. [13] G. Colavolpe, A. Modenini, and F. Rusek, “Channel shortening for nonlinearsatellitechannels,”IEEECommun.Letters,vol.16,pp.1929– 1932,Dec.2012. [14] N. Merhav, G. Kaplan, A. Lapidoth, and S. Shamai, “On information rates for mismatched decoders,” IEEE Trans. Inform. Theory, vol. 40, pp.1953–1967,Nov.1994. [15] D.M.Arnold,H.-A.Loeliger,P.O.Vontobel,A.Kavcˇic´,andW.Zeng, “Simulation-based computation of information rates for channels with memory,” IEEE Trans. Inform. Theory, vol. 52, pp. 3498–3508, Aug. 2006. [16] D.D.FalconerandF.Magee,“Adaptivechannelmemorytruncationfor maximumlikelihoodsequenceestimation,”BellSystemTech.J.,vol.52, pp.1541–1562,Nov.1973. [17] F.RusekandA.Prlja,“OptimalchannelshorteningforMIMOandISI channels,” IEEETrans. Wireless Commun., vol. 11, pp. 810–818, Feb. 2012. [18] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, “Optimal decoding of linearcodesforminimizingsymbolerrorrate,”IEEETrans.Inform.The- ory,vol.20,pp.284–287,Mar.1974. [19] S.BenedettoandE.Biglieri,“Nonlinearequalization ofdigitalsatellite channels,”IEEEJ.Select.AreasCommun.,vol.1,pp.57–62,Jan.1983. [20] K. P. Liolis and N. S. Alagha, “On 64-APSK constellation design optimization,” Proc. Intern. Work. on Signal Processing for Space Commun.,pp.1–7,Oct.2008.