ALTERNATIVE INTERLEAVING SCHEMES FOR INTERLEAVED ORTHOGONAL FREQUENCY DIVISION MULTIPLEXING Karthik Subramanian and K.V.S Hari Department of Electrical Communication Engineenng Indian Institute of Science Bangalore 560012, INDIA 080-293-2745 karthik@dsplah.ece.iisc.emet.in,hari@ece.iisc.emet.in Absrrucr-We propose a matrix model for the Interleaved Orthogonal Frequency Division Multiplexing System OOFDM) [31, [41, [SI. Using this model, we show that the interleaving scheme originally proposed for IOFDM is not unique in terms of preserving system complexity. We further show that a class of interleaving matrices yields IOFDM systems with the same complexity as the original IOFDM system. Furthermore, we show that the interleaving scheme does not affect the BER performance of the system as long as the interleaving mamx is unitary. : W ~ ~ ' ]E, 1 by the D' blocks*, D' = d i U ~ [ l : W p K..., { O ; 1,.. . P - 1). Each component K x 1block is then subjected to a K-point DlT, as represented by the blocks FE and the composite PK x 1 block is subjected to interleaving, and finally a zero pad of length L equal to the maximum channel length is appended to the IOFDM blockJhe size of the block with zero-pad added is M = P K L. This block sequence !c(rk) is now converted to serial and then fed to a single antenna for transmission across the channel. + 1. INTRODUCTION Orthogonal frequency division multiplexing (OFDM)[l], [2] is an effective technique for data transmission over frequency selective channels. However, due to the redundancy introduced in the form of the zero-pad or the cyclic prefix, OFDM has a loss in code rate. Interleaved OFDM [31,[41,[51 is an OFDM system which offers an improvement in code rate without an increase in the FFT size used, by interleaving several data blocks using a specific scheme before the addition of the cyclic prefixhero pad. IOFDM was proposed with a specific interleaving scheme. We propose a simplified matrix model for IOFDM and use this model to show that the interleaving scheme is not unique in terms of yielding an efficient receiver structure. We further state a result giving a class of interleaving schemes that yield IOFDM systems with the same complexity as the original IOFDM system. We also show that as long as the interleaving matrix is unitary, it does not affect the BER of the system. (&Ix PXI _ I 2. A BRIEFREVIEWOF IOFDM Figure (1) shows an IOFDM Transmitter. The information sequence s(n) consists of complex numbers obtained by mapping a binary sequence to a constellation, and is blocked into a vector sequence s(n) of size K x 1'. Next, P consecutive blocks of size K x 1 are blocked to get a composite block of size P K x 1. This block is pre-processed by multiplying the block elements by scalars, which is represented 0-7803-7651-x103/$17.00 ~ m r i u ~ K is the OFDM block size i FFT size, and P is the interleaving facta Heme the IOFDM block size i s P K x I. ' -/Yid. ,PX I, Figure 2. IOFDM Receiver To illustrate the interleaving scheme, consider a sequence3 [U,,, U,,. . . a K - - l , bo, bl .,.. bx- ,IT. ( We have consid~ 2WK = P $ 3Tsmds far manspose. H is the Hermitian operator, an underline denotes a vector. FK is a K x K DfT mtnx. Communications-/ /\J 47, ered P = 2 hcrc ). 'The interleaved sequence is now [(LO; b u , a ~ , b , ; . . ; O K - I , b1<-1]"'. If we choose the interleaving factor P to he I , IOFDM reduces to the familiar OFDM system. At the receiving end, thc serial sequence is blocked into blocks of size Pfi' x 1, and the zero pad is strippcd from each block. The blocks are now subjected to a PI< point DFT. This is represented.hy the V matrix, where V = FA,!,.R z J ~RKI, , representing the zero-pad removal. Equalization is done by simple scalar division as represented by the Rt blocks. given by '.",,<+,, R~ = d i , l , ~ [ H ( p j ~ ~ ) , H ( c J ~ ) VI t (0; A' l ) ,where H(eJ?*) is the channel frequency response evaluated at the I"' subcartier frequency. The block sequence is now subjected to a P point IDFT, de-interleaved and fed to the detector. This yields an estimate of the information sequence a(.). A complete exposition of lOFDM can be found in [4]. We define LpX as Where each &,j is a P x I< matrix of zeros, with a 1 at the (i,j)"' entry ~ 3. MATRIXMODELFOR IOFDM Figute ( 3 ) shows thc matrix models for the IOFDM transmitter and receiver. s ( ? ~is) the complex modulated information sequence. which is blocked into a PI( x 1 vector sequence, ~ ( n )G-' . is the pre-processor, C ~ isKa PIC x P K block i the interIDFT comprising fi' x K IDFT hlocks 4 . L p ~ is 7'. I S the zero-pad leaving inatrix, and TZP= [ I f K Ozxl,li] addition matrix. Thus we have at the output of the zero-pad block: L p x thus defined is the interleaving matrix corresponding to the original IOFDM interleaving scheme. The signal after passing through the linearly convolutive channel with impulse-response h ( n ) and being corrupted with additive white Gaussian noise, is represented as -dn.) = H : ( T I ) + y(n) (5) H is the channel-convolution matrix [I], ?(TI.) is the noise component and ~ ( nis)the signal at the receiver's front-end. The operations performed on the received signal can he written as follows: Where Rzp = [ I ~ KO p K x L ] is the zero-pad removal matrix, and FPK is a PK-point DFT matrix. We know that the product H = RZPH TZP.is n circulant matrix [I],and can - If be factorized as H = F,, 'fJ(71) D F P K [ 6 ] ,and hence we have = D FPKLPK C P Ka ( l l ) -td ( 7 L ) (9) "(PK-1 D = d i q ] [ H ( l ) H ( c j % ) , .._,H ( r J w ) ]is a diagonal matrix and is easily inverted by scalar division. The issue now is to efficiently invert the product A = FPK LPX C I J X . which has the following structure: A= IDo K . Do ... D1 DI .. . DP-I WK-"KD p - I (K-IJXI 1 Where D, := diag(1, I.Vj,f<, . . .,LVpK ). We make use of the following fact to factorize A intvan easily invertible product: The matrix Z = TENCON 2003 / 148 We know that can be FPK LPK C P KG-I = L' CEp L factorized as Z = L'r C g p L, where L is the original interleave matrix defined in (4). and C g P is a P K x PI< block DFT, comprising K blocks of P x P DFT's. That is, the matrix Z is merely a re-indexedversion ofthe block DFT matrix C;.p. (14) This equation is now re-written as Defininx G = dioy[Do.D1,. . . , D P - ~ ]we . see that A = Z G = LT C a p L G. In other words, FpK Lph. C P K = L'" cpp L G # , L G ~ ( n+)$(n) v(n) = D LT C (10) (11) ~(11)= G-',q(n),the vector infomiation sequence ~ ( 1 % )in the following way, that yields an efficient receiver structure: Since e(,,.)can he recovered from - cKp (DL~)-' =L ~ ' We want to find the class of matrices Lpx such that G in the equation above is diagonal. Now, consider the product F"9 K L$ti C g p . It can he seen that this product can be written as (12) It can he seen that the matrix equation above corresponds to the IOFDM receiver structure shown in the previous section. Where, J is a diagonal matrix given by 4. INTERLEAVING SCHEMES PRESERVING SYSTEM COMPLEXITY From the previous section, we see that at the receiver: ).( U = FPKLPK C P Ks(n)+D-' P O J = [ :0 FPKRzr $n) (13) The signal ~ ( ncan ) be recovered from this equation as long as the matrix LPK is invertible. In other words, the interleaving scheme is not unique. Permutation matrices are attractive because of their low complexity of implementation. For any choice of LPK that is invertible, we can always find a pre-processing matrix G - ' that will ensure the block-DFT structure at the receiver. For the IOFDM system, G-' is a diagonal matrix, resulting in P K complex multiplications. In the worst case, a full preprocessing matrix G-' can have a complexity of (PI<)' multiplications, which is a considerable increase in complexity. Hence this motivates us to look for interleaving matrices L ~ I that ( result in a pre-processing matrix G-' with low complexity of multiplication. A diagonal or an antidiagonal G-' would be a good low-complexity solution, since this would result in only P K multiplications for the preprocessing operation. However, an antidiagonal G-' cannot be used in the current IOFDM framework, since the precoding matrix C is a (forward) block-diagonal. Hence we concentrate on interleaving matrices that result in a diagonal pre-processing matrix. We immediately notice that the use of the original interleaving matrix L P K as defined in equation.(3.3) results in a pre-processing matrix G - ' that is diagonal. We state our question as What is the class qfmatrices LPK that re.salts iii U ~iagonalpre-pmcessingmatrix G-'for. the IOFDM system Y . J': . . 0. 0 ... .'., ._ ... 0 0 ] Jp-' (17) and J is J =1[ ,l.i;;i<, ,:;W . . . , W)$'']'. Inserting this into equation (3.15), we get LPK = L Cpti J L" L G CFK = L C ~ ~ J G C $ ~ = L C ~ ~ D ' C ~ ~ . ( I The matrices J and G are both diagonal, and hence their product D' = J G is diagonal as well. L is the original interleaving matrix used in the IOFDM system [4]. Result: -To get a diagonal pre-processirlg matrix, the interleaving matrix LIJKmust be of the form r~',o ... o 1 Where the matrices D'I,1 E {O, 1,.. . , ( P - 1))are diagonal matrices. In other words, the interleaving matrix should be - . Communications4 /149 a permuted P K x P K block diagonal matrix, with each of the IC x K diagonal blocks being a circulant matrix. Such an interleaving matrix will yield an IOFDM system with the same complexity as the original IOFDM system. by each symbol. The noise vecLor at the output of the deinterleaver is given by q(n) = L'Clf, L D-' Fpti RCP~ ( n ) where the cyclic-prefix removal matrix RCP is given by RCP = [ O F K x f , I $ K x P K ] ' . The total energy in the vector RCP~ ( n E, ) . is given by Ret?iarkr 1. The factorization F j j L ~ ~ CPK K G-' = L' C g p L is not unique. The factorization is valid for P-circular shifted versions of the matrix L. However this does not affect the . validity of the result in equation (3.18) in any way. This is because, LT C z P L1 = LT C g p Lz. for all L1 and Lr which are P-circular shifted versions of L. The equality is due to the block-diagonal symmetric structure o f C 2 , 2. The choice of interleaving scheme does not affect the performance of the system in any way. In the next section, we see that the BER of the lnterleavkd OFDM system is independent of the interleavingscheme used. This is bome out by simulations as well. 3. We see that for LPK to be a permutation matrix, each of the constituent blocks of the block diagonal matrix CPK D' Cpfcmust be a (circulant) pennutation matrix. Le, theelementsofeachdiagonalhlockD1,l E { O , l , . . . , K - l } should be the K t h roots of unity, in some order. (This is because the eigenvalues of a circular-shift permutation matrix of order IC are the h't.hroots of unity in some order.) 5 . PROBABILITY OF ERRORFOR IOFDM Etu = I ((RGPU(7L))w (RCPw ( ~ L ) ) ) Since the matrix Rep is unitary, the expression on the RHS above becomes 8 ( ( ~ ( n ()~)( ~ n ) ) )= p (PIC L ) Nu WS,which is the same as the energy in the vector before the removal of the cyclic prefix. However, the average noise PSD has now changed, since the number of symbols in the vector has become P K from P K + L. The noise PSD in each subcarrier is given by No All matrices in the matrix product in the noise vector under consideration are unitary, except for the diagonal matrix D-' . This vector represents equalization by scalar division for each subcarrier. Hence, the noise PSD after equalization in each subcarrier is given by Rqq(L)= No The average number ofhad bits per received vector of PIC symbols (after processing at the receiver) is given by the sum of the bad hits in each subcarrier. Hence the total number ofbad bits per block of P K symbols is given by Pbrr.,c Thus we obtain an expression for the probability of error (averaged over several blocks of PI( symbols each) as: + w. &!$&, Xft-') (m). We define Pbcr,c($)to be the Probability of Error expresEa being the sion for the signal constellation c at the SNR energy per bit, and h'o being the noise power spectral density in WattslHz. We know that the Probability of Error of a conventional OFDM system using IC subcarriers can he derived %, (w) [7], with paw,, as pep,lm = CL;' pbe?,c as defined above. For the sake of convenience, we assume in this section that in the IOFDM system, the received vector is processed and ML detected only after the de-interleaving at the receiver. In other words, the positions of the ML detector, and the de-interleaver are interchanged. This does not affect the performance of the system, as we shall see. The received vector after processing now becomes ~ ( I I= ) + ~ ( n ) L'r C z p L D-' FPKRzp ~ ( n ) It can be seen from the equation above that the encrgyper bit I is unchanged at the receiver. However, the noise energy at the receiver is no longer No. The noise vector under consideration is L7' C z , L D-' FpK RZP~ ( T L ) .We analyse the noise vcctor for both the cyclic prefix and the zero-padding cases. IOFDM with Zero Padding We now examine the probability of error for an IOFDM system using zero-padding. The noise vector under considerationnow becomesg(n) = L" C KL~D-'FPKRZPg ( r i ) . where Rzp, the overlap-add matrix is given by Rzp = [IpI< Izp] where Izp is a matrix comprising thb first L columps of the identity matrix I P K . The overlap-add matrix RZP is not unitary. The average energy in the ~ e c t o r Rap ~ ( nis )then given by 8 ( ( R ; r p ~ ) ~ ( R ; r p g Defin)). ing 9 = [wow' . . . wpx+f,-1IT, we see that the average energy in the vector Rzp &I is given by IOFDM with c.vclicprefrr Let 11he the number of bits per symbol in the constellation used. Then the total energy in the noise vector w(n)that is added to the signal at the receiver front-end is given by p (PK L ) Nu where W8 is the bandwidth occupied + The expectation of the cross terms is zero because the w ; s are uncorrelated. As in the earlier section, the noise PSD in each subcarrierafter equalization becomes Rq,l(l)= No m. TENCON 2003 / 150 The probability of error for the IOFDM system with zero padding is then given by This expression is the same as that for the Pe of an IOFDM system using the cyclic prefix. Effect of the Inlevleaving Scheme on P,: We see that the Probability of error of the IOFDM system is not affected by the choice of the “interleaving” matrix, as long us rhe nrutrix is unitary. Since any pemiutatiqn matrix is unitary, the stacked data vectors at the transmitter can be interleaved in anyfashion without qfecting the bit-error YOtio of the IOFDMsystem. This is borne out by the simulation results as well. If the “interleaving” matrix chosen is not unitary, it will affect the effective signal-to-noise ratio ($), and hence the Probability of error as well. Compari.son of P, between IOFDM and OFDM: The probability of error for OFDM and IOFDM are given below (assuming that we know the exact P, of the constellation used.) * ,0-1 I 1 0 ,I Io z 1D U J10 E b W dB Figure 4 BER comparison, IOFOM with direrent interleave schemes that the choice of the interleaving scheme does not affect the system performance as long as the interleaving matrix LS unitary Simdation results have been presented to hear out the same. REFERENCES Zhengdeo Wang, Georgios E. Giannakis, “Wircless Multicarrier Communications - where Fourier meets -Shannon”, IEEE Signal Processing Magazine. 2 9 4 5 , May 2000. S. B Weinstein, Paul M. Ehert. “Data Transmission by Frequency-Division Multiplexing Using the Discrete Fourier Transfonn”, IEEE Trurisucrion,s on G~mmnriication Technoloo, COM-19, 628-634, October I971 We see that the probability of error of an IOFDM system with Zi’ subcarriers and an interleaving factor of P is exactly the same as that of an OFDM system with P K subcamers. Furthermore, for K s , the Pis are approximately the same for both OFDM and IOFDM systems with A’ subcarriers. So for large P K (with K ), the probability of error of an IOFDM system is very close to that of an OFDM system. V.G.S Prasad. K.VS Hari, “Interleaved Orthogonal Frequency Division Multiplexing System” in Proc. lC4SSP 2002, Orlando, FL. May 2002,111.2745-2748,2002. V.G.S Prasad, Miiltiple-lirput Mri/tiple-Oiitput k h niques for interleaved Orthogonal Frequency Division Mdfiplexing, Master’s Thesis, Depanment of Electrical Communication Engineering, Indian lnstiture ofScience, January 2002. V.G.S Prasad, K.V.S Hari, “Interleaved Orthogonal Fre- 6 . S I M U L A T I ORESULTS N In figure (4) we see the results of a comparison of the BER‘s ofOFDM, IOFDM with the original interleaving scheme, and IOFDM with two different interleaving schemes. In the first scheme, no interleaving is done and the blocks are inerely stacked. In the second one, a randomly generated permutation matrix was used for the interleaving scheme. The simulations were performed for the GSM HT channel. I. CONCLUSIONS It has been shown that a class of interleaving schemes exist for IOFDM which yield systems with the same complexity as the original IOFDM system. Furthermore; it has been shown quency Division Multiplexing (IOFDM) System” IEEE Transactions on Signul Processing, To Appear. Gene H. Golub, Charles van Loan, Matrix Conrpiitulions, 3rd ed. Baltimore. M D Johns Hopkins University Press, 1996. Xiang-Gen Xia, “Precoded and Vector OFDM robust to Channel Spectral Nulls”, IEEE Transactions on Communications, vol. 49, no. 8, pp. 1363-1374. August 2001.