Enhanced Multicarrier Techniques for Professional Ad-Hoc and Cell-Based Communications

Transcription

1 Enhanced Multicarrier Techniques for Professional Ad-Hoc and Cell-Based Communications (EMPhAtiC) Document Number D2.3 Multicarrier and SC-FDMA Waveforms - Comparisons and Conclusions Contractual date of delivery to the CEC: 01/09/2014 Actual date of delivery to the CEC: 26/09/2014 Project Number and Acronym: Editor: Authors: Participants: Work package: Security: Nature: EMPhAtiC CASSIDIAN Martinod Laurent (CASSIDIAN), Markku Renfors (TUT), Xavier Mestre (CTTC), David Gregoratti (CTTC), Leonardo Gomes Baltar (TUM), Israa Slim (TUM) CASSIDIAN, TUT, CTTC, TUM WP2 Version: 1.0 Total Number of Pages: 67 Abstract: Public (PU) / Restricted (RE) Report This document exposes various forms of orthogonal frequency division multiplex formats to be subsequently used for comparative evaluations in terms of performances and implementation complexity. A comparison of the implementation complexity based on (I)FFT-PPN combination and the (distributed) fast-convolution is provided. Optimization of the FBMC referent impulse response has been conducted, while the FMT modulation format is considered mostly as the windowing-modification of the CP-OFDM. Keywords: Filter Banks Structures, Waveform Design, Fast Convolution Filter Bank, PMR ICT-EMPhAtiC Deliverable D2.3 1/67

2 Document Revision History Version Date Author Summary of main changes Laurent Martinod (CASSIDIAN) Xavier Mestre (CTTC), David Gregoratti (CTTC) Initial structure of the document with contribution from Cassidian Contributions from CTTC added Markku Renfors (TUT) Contributions from TUT added Leonardo Gomes Baltar, Israa Slim (TUM) Laurent Martinod (CASSIDIAN) Markku Renfors (TUT) Laurent Martinod (CASSIDIAN) Contributions from TUM added Updated version including remarks and comments from partners Final version ICT-EMPhAtiC Deliverable D2.3 2/67

3 Glossary and Definitions Acronym AFB AWGN BER BLER B-PMR BW CP-OFDM EPA FB-MC FBMC/OQAM FB-SC FMT FC FC-FB FDMA FS-FBMC GFDM ICI (I)DFT (I)FFT I/Q ITU LTE MAI MSE OFDM OFDMA QAM OQAM PHY PMR Meaning Analysis Filter Bank Additive White Gaussian Noise Bit Error Ratio Block Error Ratio Broadband PMR Bandwidth Orthogonal Frequency Division Multiplex with Cyclic Prefix Extended Pedestrian-A [channel] Filter Bank Multi-Carrier Filter Bank Multi Carrier with Offset QAM sub-carrier modulation Filter Bank Single-Carrier Filtered Multi Tone Fast Convolution Fast Convolution Filter Bank Frequency Division Multiple Access Frequency Sampled FBMC Generalized Frequency Division Multiplexing Inter-Carrier Interference (Inverse) Discrete Fourier Transform (Inverse) Fast Fourier Transform In-phase/Quadrature (data signal components) International Telecommunication Union Long Term Evolution Multiple Access Interference Mean-Squared Error Orthogonal Frequency Division Multiplexing Orthogonal Frequency Division Multiple Access Quadrature Amplitude Modulation Offset Quadrature Amplitude Modulation Physical layer Professional Mobile Radio ICT-EMPhAtiC Deliverable D2.3 3/67

4 PSD QPSK RB RC RF RRC SC SC-FDMA SER SINR SNDR SNR TETRA UFMC Power Spectral Density Quadrature Phase-Shift Keying Resource Block Raised Cosine Radio Frequency square Root Raised Cosine Sub-Carrier or Sub-Channel Single-Carrier FDMA [OFDM based single-carrier] Symbol Error Rate Signal to Interference plus Noise Ratio Signal to Noise plus Distortion Ratio Signal to Noise Ratio Terrestrial Trunked Radio Universal Filtered Multi-Carrier ICT-EMPhAtiC Deliverable D2.3 4/67

5 Table of Contents 1 Introduction From 3GPP LTE classical OFDM standard to new Filtered Multi Tone (FMT) radio communication system scheme Context Reminder: 3GPP LTE main parameters, assumptions and hypotheses Air Interface Nomenclature Frequency Band Channel bandwidth FFT Size and cyclic prefix Modulation and Coding schemes To summarize Antenna processing schemes Basic reminder of 3GPP LTE physical DL frame structure Basic reminder of 3GPP LTE physical UL frame structure Main physical parameters and associated available data resources Conclusion Filtered Multi Tone (FMT) radio communication system synoptic overview Considered FMT scheme Synoptic of the FMT radio communication system Tx and Rx chains Comparison aspects Spectral considerations Implementation aspects and added complexity Conclusion Further studies on Fast-Convolution Filter Bank based waveform processing Transmission latency of FC-FB based links Synchronous low-delay data Asynchronous low-delay data Latency in continuous transmission FC-FB based single-carrier waveforms FC-FB based FMT waveforms Performance and complexity evaluation of spectrally well-contained waveforms Frequency-sampling FBMC Advanced multicarrier waveforms with CP Universal filtered multicarrier (UFMC) Waveform comparison for heterogeneous PMR application Some conclusions On the Efficient Realization of Polyphase Components in Filter Bank Multicarrier Transmitters Introduction Efficient Modulated Filter Bank Structure Lattice Realization of Polyphase Components CORDIC Implementation of the Lattice Rotations Numerical Evaluation Conclusions Inter-User Distortion in OQAM/FBMC-based OFDMA System Model Distortion Computation Preliminary Considerations Distortion Expression Noise Effect Numerical Results ICT-EMPhAtiC Deliverable D2.3 5/67

6 6.4 Conclusions Conclusions References ICT-EMPhAtiC Deliverable D2.3 6/67

7 Table of Figures Figure 2-1 3GPP LTE main configuration parameters Figure 2-2 Tx/Rx configurations Figure 2-3 3GPP LTE DL frame overview Figure 2-4 3GPP LTE DL frame detailed configuration: PHY channels Figure 2-5 3GPP LTE DL frame CRS structure Figure 2-6 3GPP LTE UL frame overview Figure 2-7 FMT software PHY link level simulator Figure 2-8 Blanked FMT sub-carrier rejection results Figure 2-9 FMT modulator implementation Figure 3-1 Overlap-save processing for fast convolution Figure 3-2 Notations used for the number of samples in different parts of the overlapsave blocks Figure 3-3 FC block structure and RX-TX block alignment for minimum latency Figure 3-4 Example of an extended channel filter weight mask Figure 3-5 FC-FB spectra with different bandwidths obtained from the demonstrator Figure 3-6 Spectrum of FC-FB based FMT in 1.4 MHz LTE-like case with different overlap factors. Sub-carrier bandwidth is 15 khz and roll-off factor Figure 3-7 In-band interference of FC-FB based FMT in 1.4 MHz LTE-like case with different overlap factors. Sub-carrier bandwidth is 15 khz and roll-off factor Figure QAM constellation plot with L S = 12 and L S = Figure 3-9 Polyphase FMT designs with different prototype filter orders Figure 4-1 Non-contiguous FBMC/OQAM spectrum for FC-FB implementations with different IFFT lengths and overlap factors Figure 4-2 Non-contiguous FB-SC spectrum for FC-FB implementations with different IFFT lengths and overlap factors Figure 4-3 Zoom to the 12 sub-carriers wide gap in non-contiguous FBMC/OQAM spectrum in FC-FB implementations with different IFFT lengths and overlap factors Figure 4-4 Zoom to the 12 sub-carriers wide gap in non-contiguous FB-SC spectrum in FC-FB implementations with different IFFT lengths and overlap factors Figure 4-5 Zoom to the 12 sub-carriers wide gap in non-contiguous FMT spectrum in polyphase implementations with different prototype filter lengths Figure 4-6 Zoom to the 12 sub-carriers wide gap in non-contiguous FMT spectrum in FC- FB implementations with different IFFT lengths and overlap factors Figure 5-1 Polyphase network based efficient structure of synthesis filter bank Figure 5-2 OQAM-Staggering Operation for even Figure 5-3 Efficient SFB implementation with reordered polyphase components Figure 5-4 Lattice realization of polyphase matrix Figure 5-5 Rotor structure with only 2 coefficients Figure 5-6 CORDIC based realization of one lattice rotor Figure 5-7 Sub-carrier filter frequency response for the ELT ( ) with word-length Figure 5-8 Sub-carrier filter frequency response for the ELT ( ) with word-length Figure 5-9 Sub-carrier filter frequency response for the LS-optimized PR prototype with and word-length Figure 5-10 Sub-carrier filter frequency response for the LS-optimized PR prototype with and word-length Figure 5-11 Welch s PSD of the transmitted signal for the ELT ( ) and word-lengths and Figure 5-12 Welch s PSD of the transmitted signal for the LS-optimized PR prototype with and word-lengths and ICT-EMPhAtiC Deliverable D2.3 7/67

8 Figure 6-1: Theoretic (line) and empirical (markers) SNDR for the two-user case. Mean received SNR is 20 db (left) and 40 db (right) Figure 6-2: Normalized distortion power Figure 6-3: Symbol error rate as a function of the instantaneous received signal-to-noise ratio ICT-EMPhAtiC Deliverable D2.3 8/67

9 1 Introduction Multicarrier modulation has most of the key elements needed in the challenging new spectrum use scenarios, like opportunistic dynamic spectrum access, cognitive radio, and heterogeneous wireless system coexistence. Characteristic to these situations is the need to adjust the spectral characteristics of the transmitted signal, notably bandwidth and center frequency, to the available unused slots of radio spectrum. To support high data rates, it is often desirable to combine multiple non-contiguous spectrum slots in the transmission. In multicarrier systems, this can be achieved by activating only those sub-carriers that are within the available frequency slots. Orthogonal frequency-division multiplexing (OFDM) using the time-guard interval known as cyclic-prefix (CP) is the most important multicarrier technique and it is extensively utilized in modern broadband radio access systems. However, OFDM has one major limitation in the mentioned coexistence scenarios: limitations in spectral containment, which leads to spectral leakage of the transmitted signal and high sensitivity to interferences from asynchronous spectral components, e.g., in fragmented spectrum use. The traditional channel filtering approach is commonly used for isolating the used frequency band from adjacent channels. However, in case of dynamic and/or fragmented spectrum use, the needed channel filtering becomes difficult to implement. For these reasons consideration of alternative multi-carrier modulation formats for the emerging wireless applications becomes necessary. An alternative scheme for the considered scenarios is offered by the filter bank based methods of waveform processing and channelization filtering. Actually, it is possible to combine both functions in filter bank based implementations. In a first step, the FMT scheme allows us to keep the 3GPP LTE frame structure as it is, without any modification (except the pulse shaping part): the CRS structure does not change, the symbol length is kept identical, etc. In brief, most of the 3GPP LTE parameters are unchanged. Using the FMT scheme, we do not benefit from the CP removal gain in terms of overall throughput, but it could be seen as an intermediate step between 3GPP LTE and FBMC schemes as it introduces better spectral containment without too much PHYsical frame structure modifications. Then, in a second step, more classical FBMC schemes will be presented and discussed. The focus of this Deliverable is to compare the different multicarrier waveforms for the broadband PMR scenario including enhanced OFDM schemes as well as different filter bank based (uniform and non-uniform) multicarrier (FB-MC) and so-called single carrier (FB-SC) schemes. One of the core ideas is to investigate the feasibility of a special implementation scheme for multi-rate filter banks which is based on fast-convolution (FC) processing. The basic idea of fast convolution is that a high-order filter can be implemented effectively through multiplication in frequency domain, after taking DFT s of the input sequence and the filter impulse response. Eventually, the time domain output is obtained by IDFT. Commonly, efficient implementation techniques, like FFT/IFFT, are used for the transforms, and overlap-save processing is adopted for processing long sequences. Performances and complexity evaluation of those schemes will be presented and discussed in the following document. Chapter 6 of the document will focus on Multiple Access Interference (MAI). In theory, orthogonality among the sub-bands avoids Multiple Access Interference. However, practical ICT-EMPhAtiC Deliverable D2.3 9/67

10 systems must be provided with sophisticated synchronization algorithms, since OFDMA receivers are very sensitive to timing errors and carrier frequency offsets. These issues are especially annoying in the uplink, where the base station has to demodulate messages incoming from different uncoordinated users. In most nontrivial cases, carrier frequency offsets cannot be corrected completely and MAI has to be further cancelled at the demodulator output. Under these circumstances, Filter bank Multicarrier (FBMC) modulations seem to offer a better solution than the classic OFDM-like scheme based on the Inverse Discrete Fourier Transform (IDFT). Indeed, in FBMC, sub-bands are obtained by means of low side-lobe filters that are more robust to synchronization imperfections. It should be remarked, though, that FBMC is not a perfect technology and that MAI cannot be completely avoided. Even in single-user, point-to-point FBMC links, imperfect equalization and suboptimal prototype pulses generate undesired Inter-Carrier Interference (ICI). In a multi-user, multiple access system, this ICI propagates outside the sub-carriers assigned to each user, thus distorting other users. This will be presented with more details in the final chapter of the deliverable. ICT-EMPhAtiC Deliverable D2.3 10/67

11 2 From 3GPP LTE classical OFDM standard to new Filtered Multi Tone (FMT) radio communication system scheme 2.1 Context The goal of this chapter is to propose and describe a complete FMT based PHYsical layer link level scheme (dedicated to layer 1 behaviour analysis). That is to say a FMT based radio communication system scheme which includes a complete transmitter chain (including channel coding and modulation) and of course a complete receiver chain (including demodulation and channel decoding). This proposed FMT based radio communication system will of course be derived from an already existing standard: 3GPP LTE standard release 8/9. As a consequence, main parameters will reuse 3GPP LTE standard specifications (rel. 8 & 9). The same 3GPP parameters and frame structure definition applies to the proposed FMT scheme too. The only difference is the replacement of OFDM modulation and demodulation processing blocks by the FMT ones. The following paragraphs will describe in more details the commonalities and envisaged evolutions. It is clear that both waveforms (FMT and OFDM) are not inter-operable with each other. The main goal was here to stick to the 3GPP LTE standard as much as possible in order to keep compliance with upper layers of the standard as we will see hereafter. 2.2 Reminder: 3GPP LTE main parameters, assumptions and hypotheses Air Interface Nomenclature The proposed scheme is focusing on both downlink and uplink mode concerning data transmission, that is to say the OFDMA and SC-FDMA transmission scheme (respectively) for the LTE standard, and associated FMT schemes in our case Frequency Band The classical PMR UHF frequency band is considered, that is to say: MHz Channel bandwidth We consider three LTE systems with different transmission bandwidths. The main parameters of these systems are summarized below, Transmission bandwidth 1.4 MHz 3 MHz 5 MHz Sub-carrier spacing 15 khz 15 khz 15 khz FFT size Useful sub-carriers (inc. DC) Effective bandwidth 1.08 MHz 2.7 MHz 4.5 MHz By default, the chosen LTE channel bandwidth is 1.4MHz. ICT-EMPhAtiC Deliverable D2.3 11/67

12 Concerning the frame allocation, we consider that the whole bandwidth is used, so the 6 RBs (Resource Blocks) are filled with data. As a consequence, the sampling frequency used in the corresponding software simulation is 1.92 MHz. Of course, wider bandwidths could also be considered; and the evolutions and modifications proposed could also be applied for those wider bandwidths. We have just limited the number of considered possible bandwidths due to the current PMR spectrum limitations (5MHz is usually the amount of spectrum devoted to PMR, at least in Europe), but the generalization of the proposed scheme is straightforward for all other cases FFT Size and cyclic prefix As a consequence, FFT size is of course 128. The cyclic prefix used is the extended one. As we are considering PMR frequency bands, only this extended prefix case is applicable, to cover all propagation channel conditions. So the OFDM/SC-FDMA symbol length is 160 samples, as will also be the FMT symbol length Modulation and Coding schemes The 3GPP LTE standard implicates to use the following modulations: QPSK, 16QAM, 64QAM. We will keep the same constellation orders in the FMT scheme. The coding scheme is chosen by giving the desired coding rate. Then, the effective coding rate is calculated depending on the available TBS size. Of course, the closest coding rate is then applied. The whole set of MCS defined in the standard are supported by the proposed FMT evolution. The TBS characteristics and all the physical channel fields sizes defined by 3GPP LTE will be reused in our FMT scheme. In the same way, the entire channel coding specifications and parameters, as well as all the rate matching algorithms will remain unchanged, compared to 3GPP LTE definition. ICT-EMPhAtiC Deliverable D2.3 12/67

13 2.2.6 To summarize The main parameters derived from the 3GPP LTE standard definition, and the two additional specific parameters for FMT, are given in Figure 2-1. Configuration parameters Basic modulation access mode Carrier Frequency Bandwidth LTE Standard + FMT OFDMA/SC-FDMA 450 MHz 1.4 MHz FFT size (Nfft) 128 Sampling Frequency Sub-carrier spacing ( f) CP Symbol duration Number of used sub-carriers (N used ) Pulse shaping filter size (FMT specific) Roll-off factor (FMT specific) 1.92 MHz 15 khz 512/2048*Nfft= samples 72 (6 RBs) L*(Nfft+CP)) Alpha Figure 2-1 3GPP LTE main configuration parameters Antenna processing schemes Multiple-Input Multiple-Output, or MIMO technology, refers to the use of multiple antennas both at the transmitter and receiver. It offers significant increases in data throughput and link range without additional bandwidth or transmit power. It achieves this by higher spectral efficiency (more bits per second per Hertz of bandwidth) and link reliability or diversity. HH Even with a partial use of MIMO technology (when the receiver has a single antenna and multiple-input singleoutput (MISO) and when the transmitter has a single antenna, single-input multiple-output (SIMO)), a gain is still present. Figure 2-2 Tx/Rx configurations MS The 3GPP LTE standard defines SFC (Space Frequency Coding) methods, similar to Alamouti s technique (DL mode). ICT-EMPhAtiC Deliverable D2.3 13/67

14 2.2.8 Basic reminder of 3GPP LTE physical DL frame structure LTE DL frame structure is composed of the following elements: The overall LTE radio frame lasts 10 ms. It breaks up into 10 sub-frames of 1ms each one. Each sub-frame is composed of ms in itself duration slots. Each slot consists of 6 or 7 symbols OFDM. There are thus 12 or 14 symbols OFDM by sub-frame of 1 ms. This number of OFDM symbols depends on the cyclic prefix mode chosen: normal cyclic prefix mode or extended cyclic prefix mode. Then, the structure of the sub-frame breaks up into PRB, Physical Resource Block, which constitutes the unit of allocation for 3GPP LTE. A PRB corresponds to one sub-frame spread out over 12 sub-carriers. It is composed of 14*12=168 Resource Elements (RE): single subcarrier carrying a symbol OFDM in the normal CP configuration. Figure 2-3 3GPP LTE DL frame overview ICT-EMPhAtiC Deliverable D2.3 14/67

15 RBs = 600 subcarriers ICT EMPhAtiC Date: 26/09/2014 Overview of a typical LTE DL frame structure (physical channels) One radio frame = 10ms One subframe = 1 ms slot slot 1 slot 2 slot 10 slot RB DC Subcarrier Reference signals for antenna port 0 Reference signals for antenna port 1 PDCCH Secondary sync signals Primary sync signals PBCH Figure 2-4 3GPP LTE DL frame detailed configuration: PHY channels ICT-EMPhAtiC Deliverable D2.3 15/67

16 The scheme presented above is a typical configuration (typical release 8 configuration with Common Reference Signal configuration). Please take care that, depending on antenna configuration scheme, RS structure will differ. In the case of multi-antennas, besides the symbols of useful RS for the considered antenna port, other REs are dedicated to channel estimation purpose. These REs correspond to the positions of the RS of the other antennas ports to be considered and taken into account, and thus are not used for the transmission. Figure 2-5 3GPP LTE DL frame CRS structure As a consequence, we could easily see that the overall number of radio resources elements (RE) dedicated to useful data transmission in the frame greatly depends on the system configuration. Then, the usage of multi-antenna configuration within the transmission will of course help to increase data throughput thanks to transmission diversity schemes, but in the same time, we shall consider the loss of available REs for user data. As already explained, the frame structure remains identical in the FMT scheme: that is to say that Radio Resources Elements (REs) and their respective allocation do not change compared to 3GPP LTE definition of Physical Channels. The only modification is the pulse shaping function during modulation/demodulation process. ICT-EMPhAtiC Deliverable D2.3 16/67

17 2.2.9 Basic reminder of 3GPP LTE physical UL frame structure LTE UL frame structure is composed of the following elements: The duration of the sub-frame is still 1 ms, and it is composed of two 0.5 ms slots as previously. There are 6 or 7 SC-FDMA symbols per slot (depending on cyclic prefix mode): thus, 5 resp. 6 SC-FDMA symbols for the data transmission and the central symbol used as the reference signal for the demodulation of data (DM-RS). Overview of a typical LTE UL frame structure (physical channels) Figure 2-6 3GPP LTE UL frame overview ICT-EMPhAtiC Deliverable D2.3 17/67

18 Main physical parameters and associated available data resources In order to be able to calculate the available data flow (to compute maximum available throughput), we have listed main parameters: BW size (MHz) : 1.4, 3, 5, Modulation order : QPSK; 16QAM; 64QAM Channel coding ratio : see MCS defined in the standard HARQ (max. retransmission) : 1; 2; 3; 4; 5; 6; 7; 8 Nb of CCE per DCI : 2; 4; 8 Number of DL antennas : 1; 2 PDCCH size (nb of symbols) : 1; 2; 3; 4 The presented general parameters are those which are most important for the dimensioning of available dataflow. They will influence the calculations leading to the final estimate amongst available resources for traffic. This number of available radio resources for traffic is obtained by withdrawing among all the resources of the system those which are used for signalling and control and thus non available for the transfer of voice or data. Bandwidth size: The size of BW will change the calculation of the total resources and the useful resources of the frame, since it conditions the number of RB in the frequency domain (RB = allocation unit). Number of TX antennas: The antenna configuration will influence the number of resources standardized for the RS (sounding reference) thus non-usable for the traffic. MCS: This parameter determines the number of bits contained in a symbol and thus a RE (Resource Element). The number of CCE per DCI will impact consequently on the number of possible DCI, that is to say the maximum number of connections for a given PDCCH size. HARQ max. retransmission Some of those general parameters will be useful both for DL and UL calculations, whereas some of them will be dedicated to DL or UL only Conclusion As described just before, the whole 3GPP LTE different physical channel specifications in terms of number of resources (REs) is kept unchanged in the case of our FMT scheme. The length of the OFDM symbols and FMT symbols remains identical (160 samples at 1.92 Msamples/s for the BW=1.4 MHz case for example). The different slots, sub-frames and radio frame lengths and configurations are identical too. In brief, the overall number of REs and their respective allocation to physical channel fields do not change at all. ICT-EMPhAtiC Deliverable D2.3 18/67

19 In that way, we could assure that we maintain compliance and inter-operability with 3GPP LTE upper layers. This is main assumption of this FMT scheme proposition: we could easily replace the classical OFDM structure by our FMT structure, keeping the rest of the radio communication system unchanged. In that way, we benefit from the better spectral containment brought by the FMT characteristics compared to classical OFDM. Of course, due to the better spectral behaviour of the proposed FMT evolution, we could reduce frequency guard bands compared to 3GPP LTE, or reuse additional sub-carriers if we want to fill the 3GPP LTE initial bandwidths, but this is not the subject here. 2.3 Filtered Multi Tone (FMT) radio communication system synoptic overview We will introduce the proposed FMT scheme and implement it in an existing 3GPP LTE Physical Link Layer chain synoptic. We replaced the OFDM (IFFT + CP insertion) modulation block by the FMT modulation one at emitter side and, in the same way, replaced the OFDM demodulation block by the FMT demodulation processing at receiver side Considered FMT scheme For implementation matters, we consider the FMT scheme and parameters described in the dedicated paragraph in document D2.1, related to FMT approach. To remind, we derived parameters from 3GPP LTE standard, and we added 2 additional parameters (temporal pulse shaping filter support and roll-off factor). Configuration parameters Basic modulation access mode (LTE) Carrier Frequency (PMR case) Bandwidth (LTE) FMT derived from LTE OFDMA (downlink mode) 450 MHz 1.4 MHz FFT size (LTE) 128 Sampling Frequency (LTE) Sub-carrier spacing (LTE) CP (LTE) Symbol duration (LTE) Number of used sub-carriers (LTE) Pulse shaping filter size in samples (FMT specific) 1.92 MHz 15 khz 512/2048*NFFT=1/4 160 samples: Nfft+CP 72 (6 RBs) L*(Nfft+CP)) with L = 12 Roll-off factor (FMT specific) Alpha Tx = 0.22 Alpha Rx = 0.18 As described in details in document D2.1, the pulse shaping filter used in this FMT scheme is a classical square root raised cosine filter applied in the frequency domain. In addition, the frame structure and frame formatting used in order to perform the analysis corresponds exactly to the 3GPP LTE standard release 8/9 features. The various physical ICT-EMPhAtiC Deliverable D2.3 19/67

20 channels and pilots (Reference Signals) configurations implemented in the FMT scheme are compliant with the standard. In particular, for channel estimation purpose, we used the CRS (Common Reference Signals) scheme: the positions and RS sequence and values are then obtained following the corresponding specifications Synoptic of the FMT radio communication system Tx and Rx chains The following synoptic describes the FMT radio communication system chain model (DL case). OFDM modulation and demodulation processing blocks have been replaced by FMT modulation and demodulation processing algorithms. As it is highlighted in the following synoptic, the only changes correspond to the replacement of CP insertion and suppression of the OFDM scheme by the FMT pulse shaping feature. Both OFDM and FMT scheme are presented, and the FMT evolution is superimposed on the OFDM one. As we could see in the synoptic, the modifications due to the FMT scheme are quite limited and localized. As we will analyse in the deliverable D9.2, the overall performances in terms of BER/BLER will remain very similar whatever the considered OFDM / FMT scheme. Moreover, as far as channel estimation processing algorithm is considered, we will see in D9.2 that the same algorithm is applicable in both cases with similar results. As a consequence, the receiving chain shall not be strongly modified to take into account the FMT modifications. ICT-EMPhAtiC Deliverable D2.3 20/67

21 Figure 2-7 FMT software PHY link level simulator BURST MAPPING PN seq. generator PHY-channel mapping Pilots insertion Data bits CRC encoder Randomizer FEC encoder Interleaver Rate matching QAM mapper MIMO encoder Frame structure INIT CHANNEL CODING SYMBOL MIMO IFFT FMT OFDM CP remove FFT AWGN Multiple paths summation Uncorrelated fadings PROPAGATION CHANNEL Delays Multiple paths CP insert. OFDM FMT MIMO decoder & Channel estimation Channel Equalization Pilots removal BER/FER calculation + EVM PHY-channel demapping Received Data bits QAM Demod. CRC decoder Rate dematching Derandomizer Deinterleaver FEC decoder ICT-EMPhAtiC Deliverable D2.3 21/67

22 alpharx Magnitude (db) ICT EMPhAtiC Date: 26/09/ Comparison aspects Spectral considerations The main metric we used to qualify the benefit brought by the usage of the FMT scheme is the rejection measured in case of one sub-carrier removal. As described in the dedicated paragraph in document D2.1, related to FMT approach, we have the following results: by applying a pulse-shape filtering as described previously, one notices that this rejection level increased (FMT spectrum is shown in red, compared to the previous OFDM one, in blue). 0 Filtered LTE signal with blank sub-carrier Frequency x 10 5 Please remind the following results from D2.1 deliverable, which resume the achievable EVM and rejection theoretical performances depending on chosen roll-off factors for the FMT parameters. EVM with filter length:12, 64QAM, 1 subcarrier removed Rejection with filter length:12, 64QAM, 1 subcarrier removed alphatx alphatx Figure 2-8 Blanked FMT sub-carrier rejection results Rejection max: 49.7dB (AlphaTx: 0.24) EVM min: 1.0% (AlphaTx: 0.21, AlphaRx: 0.18) EVM max: 4.7% EVM at maximum Rejection: 49.7dB, 1.1% (AlphaTx: 0.24, AlphaRx: 0.17) ICT-EMPhAtiC Deliverable D2.3 22/67

23 Rejection at minimum EVM: 48.1dB, 1.0% (AlphaTx: 0.21, AlphaRx: 0.18) Optimal point: (AlphaTx: 0.22, AlphaRx: 0.18) => Rejection: 48.7dB, EVM: 1.0% In conclusion, in the case of one sub-carrier removal, we increase the rejection level by nearly 40 db compared to OFDM case. Moreover, if we consider removing additional adjacent sub-carriers, this value will increase at bit more for the bi and tri-adjacent removed sub-carriers. Trying to achieve better rejection is not necessarily a good option. Of course, from a digital point of view, it is possible, but we also have to consider the behaviour of the whole transmission chain in the reality: the High Power Amplifier (HPA) will also add some adjacent channel spectral regrowth due to its intrinsic non-linearities. Thanks to digital pre-distortion (DPD) algorithms, the overall adjacent channel spectral regrowth at the system output will be reduced, but not suppressed. Please refer to the D3.3 deliverable to have deeper details and results on this topic. As a final result, we could say that the achieved results on increasing the rejection by more than 40 db is in the right range of order. This will allow to reduce needed frequency guard bands in order to improve coexistence and cohabitation from both broadband and narrowband systems. Please refer also to D2.2 deliverable about coexistence issues due to interference limitations brought by high level of broadband noise close to the carrier Implementation aspects and added complexity As far as implementation and complexity aspects are concerned, we have taken the following hypotheses and assumptions: Thanks to the convolution theorem which states that the Fourier transform of a convolution is the point wise product of Fourier transforms, FMT implementation of the pulse shaping function is performed after IFFT (TX side), or before FFT (RX side), in the temporal domain. As a consequence, the classical OFDM chain remains identical and we only replace the CP insertion/suppression processing blocks by the FMT pulse shaping processing ones. Using this implementation, the filtering convolution process is replaced by a simple overlap-add feature to process the point wise product and accumulation all along the filter temporal support length, and then the FMT frame (multiplications and additions on the temporal support of the pulse shaping filter). Please refer to the following FMT modulator synoptic focus for dedicated description. Real channel estimation (MAP approach algorithm) is based on single tap equalization. So, it applies on both schemes equally. No changes are required for this algorithm regarding FMT implementation scheme, and then, no additional complexity. Please refer to D3.1 deliverable to have the detailed description of the MAP algorithm used for this FMT implementation case, and also to D9.2 deliverable in order to have the corresponding performances and comparisons in terms of BER/BLER results for both OFDM and FMT schemes. As a result, we do not change at all the rest of the receiver (only the pulse shaping feature is added at receiver side). The rest of the demodulation processing is kept identical. The algorithms used are thus the same whatever the case: OFDM of FMT schemes. ICT-EMPhAtiC Deliverable D2.3 23/67

24 Figure 2-9 FMT modulator implementation To conclude, regarding Tx or Rx side, the only additional complexity is composed of L*(Nfft+CP)) complex multiplications and additions for each equivalent OFDM symbol, which is acceptable in terms of complexity. The main constraint is the additional delay due to this filtering process (on both Tx and Rx side). In our hypotheses, we have limited the temporal support of the pulse shaping filter to 1 ms. As a result, we have an additional delay of 1 ms in the overall processing. This is a relatively small delay, but regarding 3GPP LTE H-ARQ process timings constraints, it is not so negligible. This is the main drawback of this FMT scheme. ICT-EMPhAtiC Deliverable D2.3 24/67

25 2.5 Conclusion As already explained, this FMT implementation has several benefits in terms of implementation: the classical OFDM digital processing chain is conserved, only adding the FMT filtering block after IFFT computation (Tx side), or before FFT computation (Rx side), the 3GPP LTE upper layer (MAC) could be kept identical: same number of resources (RE) and same pilot ratio (RS), and moreover, it provides very close performances results in simulation (see D9.2 deliverable). Thus, we are able to take benefit of improved spectral rejection close to the carrier (gain of more than 40 db) thanks to the FMT additional processing, with a relatively low added complexity, being able to fit the 3GPP LTE upper layer structures in terms of resources allocation. The proposed scheme allows us to replace OFDM processing block by FMT ones without any additional evolutions or modifications. The radio resources available are not changed, the general frame structure is not modified. This really constitutes a smooth way to introduce new filter bank waveform implementation within the 3GPP LTE actual standard. ICT-EMPhAtiC Deliverable D2.3 25/67

26 3 Further studies on Fast-Convolution Filter Bank based waveform processing The FC-FB structure was introduced in the EMPhAtiC Deliverable D2.1 for flexible multicarrier waveform processing. As a reminder about its principle, the analysis filter bank (AFB) processing flow is illustrated in Figure 3-1 while Figure 3-2 highlights the used notations. The idea of FC-FB is to use multi-rate fast-convolution processing for implementing the sub-channel filters of analysis and synthesis filter banks. Overlap-save processing is utilized to accurately approximate linear convolution by the structure that implements a cyclic convolution by nature. In this structure, the sub-band widths, center frequencies, and transition band shapes of different sub-carrier multiplexes can be tuned independently of each other. This gives also the possibility to process different waveforms in different sub-bands simultaneously. In the basic scheme, the weight coefficients are designed to optimize the pass-band and stop-band frequency response characteristics, primarily targeting at minimizing the in-band interference of the transmission link and maximizing stop-band attenuation to minimize out-of-band interference coupling. The same weight multipliers can be used also for implementing the sub-carrier equalizers, as explained in Deliverable D3.1. Figure 3-1 Overlap-save processing for fast convolution. ICT-EMPhAtiC Deliverable D2.3 26/67

27 Figure 3-2 Notations used for the number of samples in different parts of the overlapsave blocks. Below we refer repeatedly to the FC-FB design following the EMPhAtiC demonstrator parameterization. This is a 1.4 MHz LTE-like case with FBMC/OQAM waveform 1.92-MHz sampling rate, 72 active sub-carriers out of M=128 total, 15-kHz sub-carrier spacing, FFT length N = 8M=1024, IFFT length L k = 16, overlap factor of 2L O,k /L k = 6/16, that is, an overlap of 6 samples is used in the short transforms. In this chapter, we will first analyse the physical layer transmission latency of FC-FB based transmission links. Then we study the use of FC-FB for generating and processing FMT and filter bank based single carrier (FB-SC) waveforms. 3.1 Transmission latency of FC-FB based links Here we consider different elements introducing latency in an FC-FB based transmission link, with the general aim of trying to minimize the overall latency introduced by the waveform processing layer. We consider three different setups: (i) synchronous low-delay data, (ii) asynchronous low-delay data, and (iii) transmission of continuous data stream Synchronous low-delay data We assume that a certain amount of low-delay data is to be transmitted, along with data that is not critically time-constrained. Furthermore, it is assumed that the transmission frame structure is synchronized to the availability of the critical data. This model could be applied, e.g., for the transmission of control data, like channel state information. The latency is minimized by using the block structure of Figure 3-3, i.e., placing the low-delay data within FC data blocks as early as possible, but in such a way that they are not used in the overlap processing. It is not necessary to assume that the FC blocks in the TX and RX processing are aligned, and by choosing the block offset as shown in the figure, the latency is minimized. Now we can analyse the elements introducing transmission latency as follows: At time t b, all the data to be transmitted in the bth transmission block becomes available. Also the data to be carried in the overlap part of block b 1 need to be available. The transmission block of length M is created through FC processing from L S OQAM half-symbols of the current block, L O last OQAM half-symbols of the previous block, and L O first OQAM half-symbols of the following block. Further, we S ICT-EMPhAtiC Deliverable D2.3 27/67

28 assume that L LD half-symbols carry low-delay data, and these appear in the block as shown in Figure 3-3. t b t 1 =t b +t TX low-delay data of block TX processing: b b + 1 Received data: t 2 =t b +t TX t P b + 2 b b + 1 b + 2 t 3 =t b +t TX t P t D RX processing: b b + 1 t b : data for block b available, TX FC processing starts t 1 : waveform ready, transmission starts t 3 : data for RX block b available, RX FC processing starts t 4 : low-delay data of block b available b + 2 t 4 =t b +t TX +t RX t P t D t D =(2L O +L LD )M/(Lf s ) Figure 3-3 FC block structure and RX-TX block alignment for minimum latency. At time t b t TX the transmission of the non-overlapping part block b starts. Here t TX is the transmitter processing latency, including the FC processing. At time t t t (2 M M ) / f the samples needed for the FC processing of b TX P O LD s bth RX block have arrived at the receiver. Here t P is the channel propagation delay, M LD M LLD / L, and f s is the sampling rate. At time t t t t (2 M M ) / f the data symbols from the bth received b P TX RX O LD s block become available for further processing. Here t RX is the receiver processing latency. The overall transmission latency of the link is thus tsync t P ttx t RX (2 MO M LD) / fs. As an example, with the parameters selected for the EMPhAtiC demonstrator ( M 1024, L 16, L 10, M 192, f 1.92 ), and assuming that the low-delay data is mapped into S O s one OFDM-OQAM symbol, the latency becomes t t P ttx t RX μs. Evidently, coding and interleaving functions introduce additional latency in practice. Depending on the used pilot pattern, also channel estimation may also increase the latency. However, these additional elements are not included in the following discussions. In this scheme, the lower limit of the latency, t ynch,min t (2 M M ) / f, depends on s P O LD s the length of the overlapping part of the transmission block, but not on the overall block length. For given frequency response characteristics, the required length of the overlapping ICT-EMPhAtiC Deliverable D2.3 28/67

29 part is relatively independent of the overall block length. In other words, the length of the useful part of the data block can be increased without increasing the lower limit for the latency. Naturally, the TX and RX processing latencies with practical hardware would increase with increasing block length. In the case where the TX and RX hardware is just fast enough to support continuous transmission, t t M / f and the latency becomes TX RX S s ts ynch,max tp (2 M M LD) / fs. In the EMPhAtiC demonstrator case this would lead to the latency of tmax t P μs. The above analysis applies to all FC-FB based waveforms, including FBMC/OQAM, FMT, and FB-SC. It should also be noted that the maximum number of low-delay half-symbols in the presented scheme is 2L S L. For overlapping factors 0.5 or higher, the scheme does not work anymore in the presented form. However, if low-delay date is transmitted only in every other FC block, the idea is still workable Asynchronous low-delay data Here also, we consider low-delay data transmission, but without assuming that the generation of this data is synchronized to the transmission frame structure. In the worst case, there is an additional delay approaching the length of the (non-overlapping) transmission block, i.e., the link latency becomes tasync tsync MS / fs t P ttx t RX ( MO M LD) / fs. Now the latency depends on the total FC block length and the number of multicarrier symbols used for low-delay data. With the demonstrator parameters, tasync,min t P 600 μs Latency in continuous transmission When considering the transmission of a continuous data stream fully utilizing the transmission resources, the latency is minimized by synchronizing the FC block structure in the receiver to that of the received signal. Then the timing offset between transmitter and receiver is tp trx. The resulting latency is tcont t P ttx t RX ( M MS ) / fs. With the demonstrator parameters, tcont,min t P 867 μs. We can also check the latency for low-delay data when the transmitter and receiver FC processing is synchronized in the mentioned way. Clearly, the worst-case latency of asynchronous data is the same as the latency of continuous data transmission. Synchronous low-delay data could be placed at the end of the transmission frame, in which case the latency becomes t P ttx t RX ( M M LD) / fs. With the demonstrator parameters and one MC symbol of low-delay data, the corresponding value is 600 μs. 3.2 FC-FB based single-carrier waveforms The basic sub-channel filter design introduced in deliverable D2.1, with 100 % roll-off, can be extended to wider bandwidths by using the same weight coefficients: The pass-band is extended by adding bins with value 1 and the stop-bands are extended by adding the same number of 0-valued weights (see Figure 3-4 for an example). With this modification, the ICT-EMPhAtiC Deliverable D2.3 29/67

30 pulse-shaping filtering maintains the Nyquist characteristics and the root-raised cosine nature of frequency response. Figure 3-4 Example of an extended channel filter weight mask. With this approach, the extended sub-channels use 2x oversampling in the transmitter and receiver. While the transition bandwidth is constant, the roll-off factor is reduced when the bandwidth is extended. In the basic design, the roll-off is 1. In general, when the bandwidth is increased by the factor of K, K QAM-modulated single-carrier symbols are transmitted within one FB-MC symbol, and the roll-off becomes 1/K. In case of multiuser FDMA operation (e.g., cellular uplink or ad-hoc operation), a single unused sub-carrier is enough as a guard band. The feasible parameterization of FB-SC depends on the parameters of the base design with ( B) 100 % roll-off. Assume that the short transform length of the base design is L and the ( B) ( B) ( B) length of the non-overlapping part is L and that the same overlapping factor 1 L / L is used for FB-SC with different bandwidths. Then the essential criterion is that S L S ( B) LS L ( B) L takes an integer value. The resulting bandwidth can be expressed as ( B) L L W f, 2 where f is the FFT bin spacing. The resulting roll-off is Examples: L 16, L 10, f 15 / 8 khz ( B) ( B) S L L L 16, 24, 32, 40, 48, W 30, 37.5, 45, 52.5, 60, khz 1, 0.67, 0.5, 0.4, 0.33, L 8, L 4, f 15 / 20 khz ( B) ( B) S ( B) L 8, 10, 12,, 32,, 64, W 6, 6.75, 7.5,, 15,, 30, khz 1, 0.8, 0.667,, 0.25,, 0.125,. S ICT-EMPhAtiC Deliverable D2.3 30/67

31 PSD [db] ICT EMPhAtiC Date: 26/09/2014 We can see different levels of flexibility, depending on the parameters of the base design. In Figure 3-5 we can see FB-SC spectra with bandwidths of 15 khz, 90 khz, 180 khz, 360 khz, and 1080 khz, using the parameters of Example 1. The short transform lengths are 16, 6x16, 12x16, 24x16, and 72x16, respectively. The real weight mask of the demonstrator is used as the base design. Re-optimizing the weight mask for each bandwidth improves the stop-band attenuation up to about 3 db. The peak-to-average power ratio (PAPR) characteristics of FB- SC designs are evaluated in Deliverable D3.3, and they are found to be similar or slight better (with relatively low bandwidth) than those of the OFDM-based SC-FDMA designs [1]. In order to evaluate the computational complexity of FC-FB based FB-SC, we recall that the structure includes a long FFT/IFFT transform of length M, short transform of length L, and ( B) L 2 nontrivial real weight coefficients, when real weight mask is utilized. Implementing the folding (due to 2x oversampling) in FFT domain, the short transform length can be easily reduced to L/2 on the transmitter side. The same can be done on the receiver side after channel equalization, which can be done in FFT domain by adjusting the weights based on the estimated channel (so-called embedded equalizer, see deliverable D3.1). In this case, L complex weights are needed. Table 3-1 shows the complexity of RX processing of FB-SC and SC-FDMA in terms of real multiplications per symbol. The FB-SC designs are based on the EMPhAtiC demonstrator parameters with real (TX) or complex (RX) weights. For channel equalization, one complex multiplication is used for each active FFT bin, both in the SC-FB and SC-FDMA cases. We can see that the multiplication and addition rates of SC-FB are about times the corresponding rates in the SC-FDMA cases. Concerning both FB-SC and SC-FDMA, the multiplication rates for TX processing are found to be 5 10 % lower in narrowband cases and up to 30 % lower in wideband cases, compared to those of the RX processing. In the addition rates, the differences between TX and RX processing are small SC-FB 24 SC-FB 12 SC-FB 6 SC-FB 1 SC-FB Subcarrier index] Figure 3-5 FC-FB spectra with different bandwidths obtained from the demonstrator FC-FB design with real weights ( ( B ) ( B ) L 16, L 10, f 15 / 8 khz ). S ICT-EMPhAtiC Deliverable D2.3 31/67

32 Table 3-1. Computational complexity of SC-FB vs. SC-FDMA for different bandwidths in 1.4 MHz LTE like case. The results are for receiver processing with FFT-domain equalization. SC-FB design uses the EMPhAtiC demonstrator parameters. Bandwidth in Real additions/symbol Real multiplications/symbol sub-carriers SC-FDMA SC-FB SC-FDMA SC-FB FC-FB based FMT waveforms We can construct an FMT system by utilizing the FB-SC designs of Section 3.2 and placing the sub-channels at minimum distance from each other without overlap. For spectrally efficient FMT, the sub-channel roll-off should be relatively small, like 0.25 or smaller. From the previous section, we see that with the FC-FB parameterization used for FBMC/OQAM (e.g., the demonstrator case), this leads to relatively high sub-channel bandwidth, like 75 khz. In the FC-FB design, short transform length of the base design should be at least ( B) L 8 (i.e., each transition band should consist of at least 3 FFT bins), in order to reach reasonable in-band interference and stop-band attenuation. Example 2 of Section 3.2 gives a practical parameterization for such a design, with highest FFT bin spacing and lowest transform lengths. Focusing on the 15 khz bandwidth case, the main parameters are: M=2560, L=32, L S ={12, 16, 20}. Complex weight coefficients are used for improved spectral characteristics. Figure 3-6 shows the resulting transmitted power spectra with these parameters, for a 1.4 MHz LTE-like case with two 1 RB-wide spectral gaps. In the center of the gaps, the power level is at about -73 / -63 / -58 db with respect to the active sub-carrier for L S = 12 / 16 / 20, respectively. Figure 3-7 shows the in-band interference with the same overlap factors. The interference is significantly increased at the edge symbols when the overlap factor is reduced. The worst-case MSE values are -43 / -34 / -28 db for L S = 12 / 16 / 20, respectively. Still L S = 20 might be considered feasible, except for high-order constellations. Figure 3-8 shows also the constellation plots for L S = 12 and L S = 20. Also polyphase implementations for FMT [2] were designed with the same parameters for two prototype filter orders 1600 and 1920 (corresponding to polyphase overlap factors L={10, 12}) and inband MSE level of -35 db. The resulting PSDs are shown in Figure 3-9. We can see that the lower filter order is sufficient for reaching -60 db PSD level in the spectral gap. The computational complexity of FC-FB implementations of FMT can be evaluated based on the ideas of Section 3.2. For example, with L S =20 and 60 active FMT sub-carriers, the computation rates for transmitter processing are evaluated as 37 real multiplication per symbol and 215 real additions per symbol. For receiver processing, with FFT-domain equalization, the corresponding numbers are 44 real multiplications per symbol and 220 real additions per symbol. One notable aspect regarding the implementation is that length 2560=5x512 FFT/IFFT is needed. ICT-EMPhAtiC Deliverable D2.3 32/67

33 Inband interference PSD [db] ICT EMPhAtiC Date: 26/09/ L S /L=20/32 L S /L=16/32 L S /L=12/ Subchannel Figure 3-6 Spectrum of FC-FB based FMT in 1.4 MHz LTE-like case with different overlap factors. Sub-carrier bandwidth is 15 khz and roll-off factor L S =20 L S =16 L S = Symbol phase within IFFT block Figure 3-7 In-band interference of FC-FB based FMT in 1.4 MHz LTE-like case with different overlap factors. Sub-carrier bandwidth is 15 khz and roll-off factor ICT-EMPhAtiC Deliverable D2.3 33/67

34 PSD [db] Im Im ICT EMPhAtiC Date: 26/09/2014 FMT, 64 QAM, L S /L=12/32 FMT, 64 QAM, L S /L=20/ Re Re Figure QAM constellation plot with L S = 12 and L S = Order: 10x160 Order: 12x Subchannel Figure 3-9 Polyphase FMT designs with different prototype filter orders. Considering the polyphase FMT structure (including also the implementation structure of Section 2.3), the transmitter processing takes one IFFT of length 128 and the pulse shaping L* N CP L*160 to be implemented for each transmitted QAM filter of order FFT symbol block, consisting of 60 symbols in our example case. The filter has real coefficients and it operates on complex data. Using split-radix algorithm for IFFT, the computation rates are evaluated for the lower order case as 62 real multiplication and 92 real additions per symbol. Complexity evaluation for different designs is included in Section 4.4. Based on the model of Section 3.1, the lower limit for the transmission latency in the FC- FMT implementation can be calculated as t, in t 583 μs with L S =20, and tsync,min t P sync m 667 μs for L S = 18. The upper limit is tsync,max t P 2.75 ms in both cases. P ICT-EMPhAtiC Deliverable D2.3 34/67

35 4 Performance and complexity evaluation of spectrally wellcontained waveforms In this chapter we evaluate different multicarrier and single-carrier waveforms regarding their feasibility to the heterogeneous PMR application. The focus is on techniques which are able to reach about -60 db level in the spectral gaps corresponding to the width of an LTE resource block (about 180 khz). This is a crucial requirement to support the coexistence of legacy narrowband signals of the TETRA family in the same frequency band, while reaching high spectral efficiency for the broadband PMR (B-PMR) data communication system. Beyond 60 db suppression level is a natural target, since about -60 db level of PSD in spectral gaps has already been demonstrated with a practical base-station type high power amplifier (HPA) in EMPhAtiC Deliverable D3.3. In this chapter we evaluate the feasibility of different FB-MC and FB-SC waveforms to meet these requirements, considering also other performance and implementation related aspects. Before presenting the comparison, we will briefly introduce certain additional waveforms and implementation structures, which are relevant in this context and not addressed in earlier EMPhAtiC deliverables. We include in the discussions various other advanced waveforms which have been proposed for future wireless system development, like generalized frequency division multiplexing (GFDM) and universal filtered multicarrier (UFMC). 4.1 Frequency-sampling FBMC The idea of FS-FBMC was presented by Bellanger et al. in [3] and [4] and it has been widely studied in the 5GNOW project [5], [6]. It can be seen as a special case of FC-FB based FBMC/OQAM implementation, where only one sample is used from each FC processing block. In other words, the overlap factor of FS-FBMC is ( L 1) L. In this case the system becomes linear time invariant and no cyclic distortion takes place. The use of high overlap factor results in relatively high computational complexity. Sliding FFT techniques [7] help to reduce the complexity somewhat. On the other hand, a large overlap factor allows the use of low short transform length. L 8 is the common choice in FS-FBMC designs based on the PHYDYAS prototype filter with K 4. In this case the long transform length is 4 times the number of sub-carriers. Motivated by the FS-FBMC idea, one possible design approach for FC-FB is to use a small short transform length, like L 8, while slightly reducing the overlap factor, i.e., using two (or more) samples from each processing block instead of just one. This has very significant effect on the complexity, while the performance can still be expected to be sufficient. Figure 4-1 shows the PSDs 1 of such designs, in comparison to the demonstrator design parameters ( L 16, L 10 ) and a case with L 16 and L 8. Complex FFT-domain weight S coefficients are used in all cases. The FFT-domain weights are optimized separately for each case, using the ideas presented in D2.1. Interestingly, in the FS-FBMC case, the optimized S 1 In this chapter, PSDs are obtained by using an FC-FB-based analysis filter bank with the same subband spacing as in the transmitter, but with long IFFT length ( L 24 ) and high overlap factor (23/24). ICT-EMPhAtiC Deliverable D2.3 35/67

36 weights are practically the same as in the PHYDYAS design, and the FS-FBMC implementation results in the same PSD as the traditional polyphase implementation. Considering the in-band interference, the MSE value depends on the sample position within the processing block. In the following, the highest value within the processing block is indicated. For the designs with L 8, the resulting MSE in the detected symbols (with ideal noise-free channel) is -50 db, -45 db, -31 db and -28 db for LS {1,2,3,4}, respectively. Clearly, with L 8 the case where two samples from each processing block are utilized is still quite feasible, regarding both the PSD and in-band interference, and it gives a 50 % reduction in the computation rate of long transforms compared to the FS-FBMC case with L 1. With smaller overlap factors, the in-band interference becomes critical, especially if S high-order constellations are used, while the out-of-band spectral leakage might still be considered acceptable. For comparison, the in-band MSE levels in the designs with L 16 are -40 db and -58 db for LS 10 and LS 8, respectively. One important issue in this context is that small overlap makes it difficult to utilize the FFTdomain timing offset compensation method presented in Deliverable D3.1. It was verified that in the design cases with LS / L {1/ 8, 2 / 8, 8/16,10 /16} the timing offset compensation range of ±half of the sub-carrier symbol interval can be reached without significantly degrading the PSD or in-band interference characteristics. However, with L / L {3/ 8, 4 / 8} the compensation range is rather limited. S In conclusion, while considering FBMC/OQAM waveform targeting at short FC-FB block length, 60 db out-of-band PSD suppression, about -40 db in-band interference level, and wide timing offset tuning range, the indicated design cases with L / L {1/ 8, 2 / 8, 8/16,10 /16} are found to be feasible. Also designs with L {10,12,14} S would be possible, but they would lead to long transform lengths which are not powers of 2. Figure 4-2 shows the spectra for a non-contiguous FB-SC scenario 2 with L / L {1/ 8, 2 / 8, 8/16,10 /16} using the base designs without any further optimization. S We can observe that the benefit of using smaller overlap factors is significantly smaller than in the corresponding FBMC/OQAM cases. 2 This could be a scenario of two uplink users with different data rates. ICT-EMPhAtiC Deliverable D2.3 36/67

37 PSD [db] PSD [db] ICT EMPhAtiC Date: 26/09/ L S /L=4/8 L S /L=3/8 L S /L=2/8 L S /L=1/8 L S /L=10/16 L S /L=8/ Subchannel Figure 4-1 Non-contiguous FBMC/OQAM spectrum for FC-FB implementations with different IFFT lengths and overlap factors. 0 L S /L=2/ L S /L=1/8 L S /L=10/16 L S /L=8/ Subchannel Figure 4-2 Non-contiguous FB-SC spectrum for FC-FB implementations with different IFFT lengths and overlap factors. ICT-EMPhAtiC Deliverable D2.3 37/67

38 4.2 Advanced multicarrier waveforms with CP Generalized frequency division multiplexing (GFDM) is a recent multicarrier scheme, which has been widely studied in the 5GNOW project [6], [8]. The main elements of GFDM include the following: Pulse-shaping filtering is applied to the sub-carriers. This is implemented in a cyclic manner for a block of multicarrier symbols. Cyclic prefix (CP) is inserted at the symbol block level, not for each sub-carrier symbol. QAM sub-carrier modulation is used with overlapping sub-carriers, instead of OQAM. This introduces inter-carrier interference (ICI), and interference cancellation schemes needs to be used for detection. Also other than root-raised cosine type pulse shapes have been considered, and roll-off values smaller than 1 are often considered. Since CPs are used at block level, the CP-overhead is significantly reduced compared to CP- OFDM. This also allows using considerably wider sub-channels without excessive CP overhead, thus making it possible to reach reasonable block length and latency. Due to cyclic processing and the use of CP, rectangular windowing is applied to the multicarrier symbol blocks. This introduces side lobes to the spectrum in the same way as in OFDM. However, the side lobes are narrower and decay faster than in an OFDM system with the same sub-carrier spacing. Various methods have also been considered for reducing the side lobes of GFDM waveforms. One possibility is to use zero-symbols in the beginning and end of the symbol block [8]. However, this leads to significant loss in the spectrum efficiency. The use of cancellation carriers with GFDM has been shown to effectively suppress the side lobes [9]. Also time-domain windowing has been found to be effective for side lobe suppression in GFDM [8]. An effective implementation structure for GFDM was proposed in [8]. It can also be seen as an application of fast-convolution processing, now with zero-overlap of the processing blocks. The block-level CP idea has also been proposed in the FBMC/OQAM context in [10]. This scheme, called windowed CP circular OQAM (WCP-COQAM), has a lot of similarity with GFDM. However, here the real-domain orthogonality of sub-carriers is maintained due to OQAM modulation. Also time-domain windowing has been proposed to be included, in order to suppress the spectral side lobes. Similar cyclic convolution and block-level CP based variant of FMT, cyclic block FMT (CB-FMT), has also been proposed in [11]. 4.3 Universal filtered multicarrier (UFMC) UFMC is based on the OFDM model, i.e., generating first sub-carriers using the IFFT, and then filtering groups of sub-carriers to improve the spectral containment. In case of LTE-like system parameterization, a natural approach is to do the filtering at the resource block level. The used filters have relatively short impulse response, such that the impulse response tails are accommodated in the guard intervals between the multicarrier symbols. This also means that the filter transition bands are wider than what we have considered in the FLO type FB-MC and FB-SC designs. This is based on the fact that for an FIR filter, the filter order is primarily determined by the width of the transition band. ICT-EMPhAtiC Deliverable D2.3 38/67

39 4.4 Waveform comparison for heterogeneous PMR application Summarizing the previous discussions of this chapter and earlier studies under EMPhAtiC WP2, it would be quite difficult to reach the expected level of spectral containment for the heterogeneous PMR application using CP-OFDM with side lobe suppression. The methods considered in Section 2.1 of the Deliverable D2.1 approach the arithmetic complexity of FB- MC schemes but are still far from the expected side lobe suppressions level. Also the sufficiently effective methods reported in [12] exhibit quite excessive computational complexity. Side-lobe suppression methods are needed also in the block-wise CP schemes, like GFDM, but it seems that they can be used in more effective manner than in basic CP- OFDM. In any case, designing GFDM, WCP-COQAM, or CB-FMT system to reach the targeted spectral containment is expected to involve critical trade-offs between error-rate performance, PSD characteristics, overheads in spectrum efficiency, and algorithmic complexity. Also the spectral characteristics presented for UFMC in the literature do not meet the demands. Consequently, we do not consider these waveforms as interesting alternatives for the heterogeneous PMR application. The waveforms included in the evaluation are FBMC/OQAM, FMT, and SC. For the multicarrier waveforms, the traditional polyphase structure and FC based implementation structure are considered. Frequency spread FBMC (FS-FBMC) is also included in the discussion, as a special case of FC. Regarding FBMC/OQAM, only the frequency-limiter orthogonal (FLO) type is able to meet the required level of spectral containment. In the polyphase implementation, the PHYDYAS prototype filter design with overlapping factor K=4 is used due to its good spectral characteristics. Regarding the SC waveform, FC based implementation and basic time-domain implementation are considered. The latter one is based on the impulse responses corresponding to FS-FBMC designs and the length of the pulse-shaping filter is equal to the FFT length, i.e Time-domain implementation is useful as such on the transmitter side. On the receiver side, the channel equalizer would be needed, but time domain equalizer is not practical. Therefore the FB-SC approach, implementing channel equalization together with the pulse shaping, is preferred. Figure 4-3, Figure 4-4, Figure 4-5, and Figure 4-6 show the PSDs of different designs in a 12 sub-carrier (1 RB) gap for different waveforms. The characteristics of different designs are included in Table 4-1. These include the numbers of real multiplications and real additions per processed symbol as a coarse implementation complexity metric [14]. Also the in-band interference level is indicated, together basic metrics for the out-of-band PSD in a spectrum gap. The design cases have been selected to meet or exceed in a reasonable way the targeted PSD level of -60 db in the spectrum gap and -35 db in-band interference levels. ICT-EMPhAtiC Deliverable D2.3 39/67

40 PSD [db] PSD [db] ICT EMPhAtiC Date: 26/09/ L S /L=2/8 L S /L=1/8 L S /L=10/16 L S /L=8/ Subchannel Figure 4-3 Zoom to the 12 sub-carriers wide gap in non-contiguous FBMC/OQAM spectrum in FC-FB implementations with different IFFT lengths and overlap factors L S /L=2/8 L S /L=1/8 L S /L=10/16 L S /L=8/ Subchannel Figure 4-4 Zoom to the 12 sub-carriers wide gap in non-contiguous FB-SC spectrum in FC-FB implementations with different IFFT lengths and overlap factors. ICT-EMPhAtiC Deliverable D2.3 40/67

41 PSD [db] PSD [db] ICT EMPhAtiC Date: 26/09/ Order: 10x 160 Order:12x Subchannel Figure 4-5 Zoom to the 12 sub-carriers wide gap in non-contiguous FMT spectrum in polyphase implementations with different prototype filter lengths L S /L=16/32 L S /L=12/ Subchannel Figure 4-6 Zoom to the 12 sub-carriers wide gap in non-contiguous FMT spectrum in FC- FB implementations with different IFFT lengths and overlap factors. ICT-EMPhAtiC Deliverable D2.3 41/67

42 Scheme Complexity 12 sub-carriers TX/RX Complexity 72 sub-carriers TX/RX Maximal transform length Inband interf. [db] Out-of-band interference in 1 RB gap [db] Adds Mults Adds Mults Edge Centr. Polyphase FBMC/OQAM, K=4 513/ /269 86/96 43/ FC-FB FBMC/OQAM L=8, L S =1 (FS-FBMC) L=8, L S = L=16, L S = L=16, L S = Polyphase FMT, Order 12x / /376 86/96 61/ < FC-FB FMT Order 10x / /322 77/87 52/ < L=32, L S = L=32, L S = Traditional single carrier Order 1024, TX only FC-FB single-carrier L=8, L S = L=8, L S = L=16, L S = L=16, L S = Table 4-1. Characteristics of different spectrally well-contained waveforms and implementation structures. LTE-like parameterization with 128 sub-carriers and 15 khz subcarrier spacing. Notes: Complexity metrics given for 12 and 72 active sub-carriers. All FC-FB transmitters and receivers use complex weights and receivers use embedded equalizers (utilizing FFT-domain weights). 3-tap equalizer assumed for polyphase FBMC/OQAM and polyphase FMT. In FCbased FMT and FB-SC implementations, the short transform length is L /2, in FBMC/OQAM it is L. 4.5 Some conclusions - FBMC/OQAM and FB-SC reach similar spectral characteristics and spectrum efficiency. The considered FMT designs have about 20 % reduced spectrum efficiency due to the used roll-off of FC-FB implementation of FBMC/OQAM and SC waveforms is competitive against traditional designs in terms of arithmetic complexity; the downside is higher transform lengths. On the other hand, increased transform sizes is a common feature of many recent waveform proposals, like GFDM. But the main benefits of FC- FB is high flexibility, multimode capability, and effective support for asynchronous ICT-EMPhAtiC Deliverable D2.3 42/67

43 operation, as documented in the EMPhAtiC WP3 deliverables and the results of FS- FBMC studies of the 5GNOW project. - FC-FB is not very effective for implementation of FMT with LTE-like parameterization and medium/small roll-off ( 0.25). - FS-FBMC reaches very good spectral characteristics, at the cost of high complexity. Other FC-FB parameterizations reach the target with much lower complexity. Having small transform lengths is a favoured property, but minimum lengths in FC-FB lead to high overlap factors, and high complexity. LS / L 2 / 8 is a possible choice if the transform length is critical but, e.g., LS / L 8 /16 provides good performance with clearly lower complexity. ICT-EMPhAtiC Deliverable D2.3 43/67

44 5 On the Efficient Realization of Polyphase Components in Filter Bank Multicarrier Transmitters In this chapter, we perform an analysis of different realizations of the polyphase networks usually employed in OQAM-FBMC. The main objective is to evaluate the complexity and robustness of the structures with a very low complexity implementation of the multiplication by the prototype filter coefficients. We evaluate a direct form, a lattice and a Coordinate Rotation by Digital Computer (CORDIC)-based lattice structure of the polyphase components. For the first two structures a Canonical Signed Digit (CSD) representation of the coefficients is utilized and for the later CORDIC is used. We evaluate the complexity obtained after the quantization in terms of additions and shifts. By considering the application of multicarrier systems we show the effect of the coefficient quantization on the spectrum of the sub-carrier filters and of the multicarrier transmit signal Power Spectral Density (PSD). 5.1 Introduction Multiplications by the prototype filter coefficients are the most computationally intensive operations in the realization of the filter bank for multicarrier modulation. During the design of the prototype filter, the coefficients calculated are represented with the accuracy of the computer employed. However, in practical implementations of the filter bank structure the coefficients will be implemented with finite-precision arithmetic using digital signal processors (DSPs) or VLSI chips. In this case the coefficients will have to be quantized, usually using rounding, and the time and frequency responses will be deviated from the original response. If the sub-carrier frequency response or the out-of-band radiation should meet some prescribed specification, the quantized prototype may even fail to do that. Moreover, the sensitivity to coefficient quantization will depend on the structure employed and this is the main motivation of this study. We will perform a comparison of different structures for the implementation of the prototype filter. By using a fixed-point binary representation of the coefficients, e.g. two s complement, the multiplication can be substituted by shift and addition operations. However, using those binary representations will necessitate a large number of additions and shifts. An alternative representation is the Canonical Signed Digit (CSD) number representation [23], [24] that includes also subtractions. In addition to that, CSD is a signed digit representation that minimizes the number of non-zero digits by avoiding consecutive non-zero digits and thus reducing the number of partial product additions/subtractions. The shift operations can be implemented by only connecting the wires to the corresponding bit positions or by using shift registers. The hardware complexity can be then evaluated in terms of the total number of additions/subtractions and shift operations. It is worth noting that the signal word-length will determine the maximum size of the shift registers. The critically sampled modulated filter banks (MFBs) have some very efficient structures that allow the implementation of pairs of polyphase components using lattice structures. Those lattice structures are known to preserve the perfect reconstruction [25], [31]. In addition to that, the lattice structures can be implemented using coordinate rotation by ICT-EMPhAtiC Deliverable D2.3 44/67

45 digital computer (CORDIC) [26], [27] in order to improve the quality of the coefficient quantization with a reduced complexity. 5.2 Efficient Modulated Filter Bank Structure We consider here uniform exponentially modulated filter banks. In uniform filter banks, all the sub-channels have the same sampling rate, all the analysis and synthesis filters in the sub-channels, and, respectively, have the same bandwidth and are derived from a single prototype filter denoted as. In exponentially modulated filter bank, both and are obtained by exponentially modulating Where and. We also assume here, without loss of generality, that the length of the prototype, where is a time overlapping factor. determines not only the complexity of the filter banks, but also the length of its memory. An efficient structure of the synthesis filter bank is shown in Figure 5-1, where represents the -th polyphase component of type-1 of and provide the following relation (2) Since we have assumed that the prototype has a length of component will have a length of., each polyphase Figure 5-1 Polyphase network based efficient structure of synthesis filter bank. ICT-EMPhAtiC Deliverable D2.3 45/67

46 The block performs a OQAM staggering of the real and imaginary parts of the low-rate signals. Figure 5-2 depicts the internal structure of. Figure 5-2 OQAM-Staggering Operation for even. We assume here that the prototype filter was designed such that perfect reconstruction (PR) [25], [28], [24] is fulfilled. In that case, and in ideal channel conditions, the sub-carriers are completely orthogonal. Furthermore, the prototype has to have a symmetrical impulse response. In general, the methods to design the prototype filter to guarantee PR sets constraints on the polyphase filters [29]. For the case that the prototype filter has a length of ( ), there exists a closed form expression of the filter s coefficients derived by Malvar named extended lapped transform (ELT) [30]: For the cases where there is no closed form solution and the prototype has to be obtained by numerical optimization methods [31], [32], [24]. It is worth noting that although the prototype has a symmetrical impulse response, necessary to achieve PR, one multiplier per prototype filter coefficient has to be realized in the polyphase network. This is a particularity of this polyphase components based filter bank structure. 5.3 Lattice Realization of Polyphase Components By considering Figure 5-1 again, it is possible to show that there are pairs of power complementary polyphase components [25], [28]. The polyphase components and have the same but shifted outputs. So these polyphase components can be grouped as shown in Figure 5-3 where there is a need to have a permutation matrix denoted as. ICT-EMPhAtiC Deliverable D2.3 46/67

47 Figure 5-3 Efficient SFB implementation with reordered polyphase components. The polyphase components pairs can be jointly implemented using a non-recursive lattice structure [27], [25]. A non-recursive lattice structure consists of rotations and delay sections: The lattice structure of each pair of the polyphase components is shown in Figure 5-4. The rotation angles can be found successively from the coefficients of the prototype filter by polynomial degree reduction [25]. Later, the rotor coefficients will be quantized. Figure 5-4 Lattice realization of polyphase matrix. The lattice structure based on rotors with 4 coefficients can have the coefficients easily quantized also using the CSD representation. And again the multiplication with the coefficients will be substituted by shifts and additions. Since the coefficients are sines and cosines they already have a normalized dynamic range. The main advantage of this lattice structure is that the PR is preserved independent of the quantization of the coefficients. The problem is that the structure in Figure 5-4 has a higher number of coefficients/multiplications than the direct form. For this reason the rotors can be modified ICT-EMPhAtiC Deliverable D2.3 47/67

48 in order to have only 2 multipliers each plus two multipliers for each polyphase component as shown in Figure 5-5. Figure 5-5 Rotor structure with only 2 coefficients. Now the lattice rotors with only 2 multipliers reduce the complexity in almost a half, but the coefficients in each of them are tangent functions and their dynamic range are theoretically infinity, depending on the resulting angles. It becomes hard to find an optimum quantizer and the quantization completely changes the frequency behavior of the filters. It is worth noting that still PR can be preserved. As an alternative the quantized 2 coefficients rotors can be realized using CORDIC based rotors [33], [27]. 5.4 CORDIC Implementation of the Lattice Rotations The principle of CORDIC is applied to realize the rotation by an angle [26]. As a result, the multiplication by will need only a small number of shift and addition operations. To avoid the direct computation of the trigonometric function the rotation by the angle is successively approximated by a sequence of elementary rotations by the angles,,. These are chosen in such a way that, which for binary data only requires a simple shift by bit. Thus, we have where the sign factor is chosen according to the different CORDIC algorithms as Elementary-Angle-Set Angle Recoding (EAS-AR) [34], [35], [26] and the extended CORDIC structures that promise higher precision for the same complexity like, for example, Extended EAS-AR [26]. The total number of approximation steps depends on the required accuracy. If the word-length of the data is bit, the least significant bit has weight, which should be equal to the smallest angle. So, we stop at. One algorithm that can be used for the calculation of the s is the Elementary-Angle-Set Angle Recoding (EAS-AR) [34], [35], [26]. There are extended CORDIC structures, with corresponding algorithms to calculate the parameters, that promise higher precision for the same complexity like, for example, Extended EAS-AR [25]. Each rotor can be implemented as in the structure in Figure 5-6, where the multipliers are now only shifts and depending on the, the micro-rotation will not exist. (5) ICT-EMPhAtiC Deliverable D2.3 48/67

49 Figure 5-6 CORDIC based realization of one lattice rotor. 5.5 Numerical Evaluation To evaluate the effect of prototype coefficient quantization in an FBMC system we consider here a system with a total of sub-carriers. Two prototypes with different lengths will be considered: for we took the ELT from (4) and for we took a PR optimized prototype that minimizes the energy in the stop-band (Least Squares) using the method from [32], [36]. We consider three structures: direct form, lattice and CORDIC-based lattice. For the direct form and the lattice structure the coefficients are directly quantized using CSD representation. Furthermore, the dynamic range of the quantizer is adjusted for the direct form in order to minimize the effects of saturation in the description of the coefficients. For the lattice structure no such an adjustment is needed, because the coefficients are limited to the unity. For the CORDIC-based lattice structure we have employed the EAS-AR algorithm to obtain the micro-rotations. To evaluate the different structures after coefficient quantization, we have considered first the total number of additions/subtractions involved in the realization of the polyphase network for the synthesis filter bank. We assume that the complexity of the DFT will be the same for all structures and we leave it away for the complexity evaluation. In Table 1 and 2 the total number of additions is listed for word-lengths varying from to 12 for each prototype. For the ELT, one can see that the lattice structure provides a much higher number of additions/subtractions compared to the direct-form and the CORDICbased lattice structures, they need a similar number of additions/subtractions. Of course that both lattice structures preserve the orthogonality between the sub-carriers in opposition to the direct form that does not. Word-length Direct Form Lattice Lattice-CORDIC Table 1: Number of additions/subtractions for the ELT ( ) for different word-lengths In the case of the longer prototype, the LS-optimized, the number of additions of the lattice structure increases dramatically compared to the direct form, but the lattice-cordic only ICT-EMPhAtiC Deliverable D2.3 49/67

50 results in increases on the order of 22-80%. Word-length Direct Form Lattice Lattice-CORDIC Table 2: Number of additions for the LS-optimized PR prototype with word-lengths for different In Tables 3 and 4 we have calculated the total number of shifts also for different wordlengths for the two prototypes. For the ELT we can see that the lattice-cordic structure has a slightly lower number of shifts and the lattice has more than 100% increase compared to the direct form. Word-length Direct Form Lattice Lattice-CORDIC Table 3: Number of shifters for the ELT ( ) for different word-lengths In the case of the longer prototype the roles change and the lattice-cordic structure has now an increase between 17-22% in the total number of shifts compared to the direct form. Word-length Direct Form Lattice Lattice-CORDIC for different word- Table 4: Number of shifters for the LS-optimized PR prototype with lengths As a second figure to illustrate the effects of coefficient quantization we have calculated frequency response of the sub-carrier pulse, i.e. the prototype frequency response, for unquantized direct form and all the three structures with quantized coefficients. In Figure 5-7 we have depicted the frequency response for and in Figure 5-8 for, both for the ELT. We can see that for the deviation of the frequency response, especially in the stop-band, is very small for all the structures when compared to the unquantized version. For we can see a larger deviation for both lattice structures compared to the direct form. Again, one should note that only the lattice structures preserve the orthogonality at the end. ICT-EMPhAtiC Deliverable D2.3 50/67

51 Figure 5-7 Sub-carrier filter frequency response for the ELT ( ) with word-length Figure 5-8 Sub-carrier filter frequency response for the ELT ( ) with word-length In Figure 5-9 and Figure 5-10 we have plotted the frequency responses for the LS-optimized PR prototype for and. In the second case we can see that there are minor deviations for all structures but it is not possible to say which is higher. In the first case all ICT-EMPhAtiC Deliverable D2.3 51/67

52 structures show a significant deviation but a decrease in the stop-band attenuation is minimal for all of them, with the lattice-cordic showing a slightly worse attenuation. Figure 5-9 Sub-carrier filter frequency response for the LS-optimized PR prototype with and word-length Figure 5-10 Sub-carrier filter frequency response for the LS-optimized PR prototype with and word-length ICT-EMPhAtiC Deliverable D2.3 52/67