1.

INTRODUCTION
In high-speed digital communication systems, it is of interest to have a full-duplex, i.e., simultaneous bidirectional, transmission between the transmitter and the receiver, as it generally leads to lower latency and greater data throughput. However, full-duplex transmission results in an unwanted interference of the transmitted signal on the collocated received signal, which is known as echo. The echo can be prevented by using different methods such as, electrically shielding the transmitter and the receiver, frequency division duplexing or deploying digital echo cancellers (a combination of these methods can also be used).In digital subscriber line (DSL) systems, which are widely used for broadband communication over the existing telephone lines, a single pair of wires is used for full-duplex transmission. Thus, a hybrid circuit (i.e., a four-wire to two-wire interface) is used to separate the transmitter and receiver circuits from the line. However, because of the impedance mismatch in the hybrid, the transmitted signal is not completely blocked and leaks into the collocated receiver, causing an undesirable echo. Therefore an echo canceller can be used which removes the echo by subtracting an emulated echo from the received signal at the receiver. In a basic echo canceller, an estimation of the echo is first generated by performing the linear convolution between the transmitted signal (causing the echo) and the estimated echo channel; the estimated echo is then subtracted from the received signal. In this method, the echo channel is modeled by a finite impulse response filter which is adaptively updated. In DSL systems based on the discrete multitone (DMT) modulation, the channel is divided into orthogonal sub channels (or tones), and data is modulated and demodulated by the use of the inverse discrete Fourier transform (IDFT) and DFT operations. The direct time domain implementation of the basic echo canceller for DMT-based systems is characterized with both high computational cost and slow
Department of ECE,GEC,Thrissur 1

convergence; however, this can be improved by exploiting the structure in the transmitted signal. Mixed time and frequency domains adaptive echo canceller for DMT-based system can be viewed as a constrained optimization problem, where a cost function is minimized over an extended linear space. This extended space contains the weights of the finite-impulse- response (FIR) filter modeling the echo channel both in the time and frequency domains. In most of the echo cancellation methods for DMT-based systems, the frequency-domain adaptive echo weights are mapped into the time domain filtering weights used for the echo emulation by means of the inverse Fourier transform. In constrained echo canceller this mapping can be regarded as a linear constraint on an extended set of weights. Based on this interpretation, we introduce a constrained adaptive echo canceller structure for DMT-based systems.

Department of ECE,GEC,Thrissur 2

2. MULTICARRIER MODULATION
Two main modulation schemes are currently being used to implement ADSL: carrier less amplitude/phase (CAP), a single carrier modulation scheme based on quadrature amplitude modulation (OAM); and discrete multi-tone (DMT), a multichannel modulation scheme. The choice between them naturally depends on how well they perform in the presence of impairments on the existing copper twisted-pair access cabling , because these can limit the transmission capacity. In addition, high-bit rate services carried by ADSL must not interfere with other services, particularly plain old telephone service (POTS), that are being transported simultaneously over the same lines. In essence, multicarrier modulation superimposes a number of carriermodulated waveforms to represent the input bit stream. The transmitted signal is the sum of these subchannels (or tones), which have the same bandwidth and equally spaced center frequencies.. Early problems with maintaining an equal spacing between tones have been resolved with the introduction of digital signal processors, which can accurately synthesize the sum of modulated waveforms, and the FFT (fast fourier transform), which can efficiently compute this sum. 2.1 DISCRETE MULTI-TONE MODULATION DMT is a form of multicarrier modulation in which each tone is QAM modulated on a separate carrier. The lowest carriers are not modulated, thereby avoiding interference with POTS. DMT modulation is optimal for band-limited communication channels (such as twisted pair telephone cables), which exhibit large differences in gain and phase with frequency. When the modem is initialized, the number of bits assigned to a tone can be set to compensate for differences in these transmission characteristics. Subsequently, if conditions on the line alter slowly, this
Department of ECE,GEC,Thrissur 3

bit assignment can be changed "on the fly."Over long distances, a DMT-based ADSL transmission system approaches the fundamental capacity limit of 13.6 Mbps. However, over the distances typically found in the access network (a few kilometers), the maximum capacity drops to about 6 Mbps. DMT technique employs frequency division multiplexing by dividing the bandwidth of a twisted pair, extending up to 1.1 MHz, into many smaller bands of 4 KHz each. A "sub-carrier" is defined for each of these frequency bands, with their frequencies at the centre of their respective bins. In the case of ADSL, the carriers are spaced at 4.3125 KHz intervals, with 249 of those used for the downstream data in the range from 26 KHz to 1.1 MHz and 25 of those used for the upstream data in the range from 26 KHz to 133 KHz. Echo cancellation is used to enable full duplex operation over overlapping frequency ranges. The alternative is to use FDM ADSL where the upstream channel is 26 - 134 KHz and the downstream channel above 138 KHz but, this has the disadvantage of a reduced bandwidth. The stream of data is first divided into several parallel streams, and these are used to modulate a sub-carrier each, independently of the others. The bandwidth efficiency attainable in each of these sub bands extends from 0 to around 15 b/s/ Hz. At low frequencies, where copper wire attenuation is low and signal-to-noise (SNR) ratio is high, it is common to use a very dense constellation supporting 10 b/s/ Hz or more. In unfavourable line conditions, the modulation can be relaxed to accommodate a lower SNR, typically 4 b/s/ Hz or less In principle, tones are independent of one another. However, in practice some intersymbol interference (ISI) occurs because the tail of one symbol corrupts the start of the following symbol. Fortunately, this can be virtually eliminated by adding a small number of samples (the cyclic prefix) to each DMT symbol. Any ISI is then limited to the prefix that is removed after demodulation by the FFT.

Department of ECE,GEC,Thrissur 4

2.1.1 BENEFITS OF DMT Both DMT and CAP are similarly complex during implementation in real applications. However, DMT has a number of important advantages: * DMT achieves near optimum use of the line capacity because it can adaptively allocate information and transmit power across the available bandwidth. * Bits rates can be adjusted in small increments of a few tens of kilobits per second, enabling DMT hardware to be programmed to support a wide range of data rates in both directions with the possibility of changing these rates "on the fly." * Power spectral density can be adjusted very flexibly. Forbidden bands (e.g. amateur radio band) can thus be avoided simply by imposing a suitable frequency mask. * DMT is good at coping with interference from multiple radio frequency sources. * DMT offers greater immunity to transient noise bursts than CAP; this performance can be enhanced by using interleaving and forward error correction. * Only minimal equalization is required; in contrast to CAP, DMT does not require a feed back equalizer, so there is no error propagation. * DMT is the more cost-effective option on complex lines with numerous bridged taps. 2.1.2 DISADVANTAGES OF DMT FFT processing introduces some delay, and a high peak to average ratio in the DMT transmit signal can lead to clipping noise. Complex initialization procedures result in a relatively long activation time. However, correct dimensioning can minimize the first two problems and ensure that any delay is negligible, even for telephony. The complex initialization procedure also can be considered as an advantage as it increases the flexibility of an ADSL system based on DMT.

Department of ECE,GEC,Thrissur 5

3. BASIC ECHO CANCELLER

Echo cancellation method for full duplex transmission is illustrated in fig.(3.1) There is a transmitter (TR) and receiver on each end of the connection and a hybrid is used to provide a virtual four-wire connection between the transmitter on each end and the receiver on opposite end. The echo canceller is an adaptive transversal filter that adaptively learns the response of the hybrid, and generates a replica of that response which is subtracted from the hybrid output signal. to yield an echo free received

Fig(3.1)Block diagram of basic echo canceller

The echo canceller notation is shown in fig(3.2). The local transmitter signal u(t) at port A generates the undesired echo signal y(t). This signal is superimposed at the output of the hybrid (port D) with the far transmitted signal x(t). The canceller takes the advantage of its knowledge of the local transmitter to generate a replica of the echo ye(t).This replica is subtracted from the echo to generate error signal e(t).This
Department of ECE,GEC,Thrissur 6

error signal is used to update the weights of adaptive filter which models the echo channel.

Fig(3.2) Operation of a basic echo canceller

Department of ECE,GEC,Thrissur 7

4. BACKGROUND
In this section, we examine the echo canceller structure using CES, for DMTbased DSL systems. In this method, echo cancellation is performed by subtracting the emulated echo from the received signal using both time and frequency representations. If the echo is periodic (i.e., generated by a circular convolution in the time domain), it can be regenerated with reduced complexity in the frequency domain. Therefore, in the echo canceller with CES, the only process performed in the time domain is to make the echo periodic by reconstructing and canceling the echo generated by the tail and the head, respectively, of the transmitted signal; the remaining echo is then removed in the frequency domain with one complex multiplication per tone. A mathematical formulation of this method is presented in this paper based on the frame synchronous DSL systems. Extension to the frame asynchronous and multirate systems are also briefly discussed . 4.1 SYNCHRONOUS ECHO CANCELLATION A multicarrier structure is assumed where DMT modulation and demodulation are performed by the use of IDFT/DFT of equal length both at the transmitter and receiver. The transmitted time domain symbol at symbol period is represented by u(k)=[u0k,uN-1k]T , which is obtained as the point IDFT of the vector U(k) which contains the QAM modulated data of the transmitter in the frequency domain. Later, a cyclic prefix with the length v (which is larger than the assumed length of the far end channel) is added to each time domain symbol. This setup ensures that the channel is equivalent to a bank of parallel independent sub channels. The true echo channel, assumed to be of length M N samples is represented by the zero-padded vector of size Nx1. The latter models the effect of the hybrid circuit, the digital and analog front end filters and the time domain equalizer (TEQ).

Department of ECE,GEC,Thrissur 8

In the synchronous case where the received and echo frames are aligned, the echo symbol affecting the received symbol y(k) at symbol period can be described completely in terms of the two consecutive transmitted symbols u(k-1) and u(k) (two symbols are needed because the length of the echo channel is usually longer than the cyclic prefix). For clarity, we refer to these transmitted symbols as the echo reference symbols. The echo symbol is generated by the linear convolution of the echo reference symbols and the echo channel and can be expressed in the form y(k)=(k)h (1)

where (k) is an NXN Toeplitz matrix consisting of the elements from symbols u(k1) and u(k).The true echo channel is unknown in practice, so the emulated echo is estimated by replacing by which is the weight vector of the adaptive filter modeling the echo channel at time. Using (1), the emulated echo is then given by ye(k)=(k)w(k) (2)

The echo weights can be obtained adaptively using various methods, e.g., the least mean square (LMS) algorithm, in which an error signal is used to update the weights iteratively .The error signal is the difference between the received signal and the emulated echo, i.e., e(k)=y(k)-ye(k) (3)

As proposed in (2)in order to avoid the matrix multiplication in (2), the echo emulation is performed partially in the time and frequency domains. This can be achieved by rewriting the matrix as a sum of a circulant matrix and a correction matrix (k) as follows: (k)=(k)+(k) where , (k) defined as (k)=C{u0k,uN-1k} (5) (4)

Department of ECE,GEC,Thrissur 9

is the periodic part of (k) , and contains only the elements of symbol u(k) , and (k)= (k)-(k) expanded as ye(k)=(k)w(k)+(k)w(k) (6) is an upper triangular matrix. Using (4) the emulated echo can be

The circular term (k)w(k) can be implemented in the frequency domain with less complexity, by diagonalizing the circulant matrix (k) . This can be achieved by using DFT and IDFT matrices, resulting into the decomposition (k)=FN-1diag{U(k)}FN (7)

Where U(k)= FN u(k).Therefore, the error signal in the frequency domain (i.e., E(k)= FN e(k).) can be written as E(K)=FN(y(k)-(k)w(k))-diag{U(K)}W(k) (8)

where W(k)=FNw(k) contains the echo channel weights in the frequency domain. As shown in Fig. 1, partial echo cancellation is performed in the time domain, and the residual signal is transferred into the frequency domain, where the remaining echo is emulated and cancelled on a per tone basis. Finally, the error signal is used to update the echo weights. An approximate LMS update in the frequency domain is used, as given by W(k+1)=W(k)+diag{U*(K)}E(k) (9)

where is the step size. The use of the approximate diagonal update (9) in place of the exact LMS update deteriorates the convergence but decreases the complexity. The inverse Fourier transform is then applied to map the updated weights into the time domain where the last elements are zeroed. The complete block diagram for this echo canceller is shown in

Department of ECE,GEC,Thrissur 10

Fig.4.1

1.

4.2 FRAME ASYNCHRONOUS ECHO CANCELLATION In the previous section, the frame synchronous echo cancellation has been discussed. In practical systems, there is a frame misalignment between the echo reference symbols and the received echo symbols. If this misalignment is set to a predetermined value at the transceiver on one side, then the misalignment at the other side is imposed by the propagation delay of the channel. Therefore, only one transceiver can perform echo cancellation synchronously, and at the other side it should be done asynchronously. By adding an extra pair of DFT/IDFT to the echo canceller it becomes possible to adjust the alignment between echo frames and received

Department of ECE,GEC,Thrissur 11

Fig(4.2) frames at the receiver independently of their original alignment. The extra IDFT maps the frequency domain emulated echo to the time domain where it is then subtracted serially; DFT maps the error signal back to the frequency domain where it is used in the adaptive update. If the error signal does not contain the far end signal, the convergence of the echo canceller improves. This method, known as double talk cancellation, can be easily implemented if the echo canceller and demodulator share the DFT at the receiver side. However, adding the extra DFT/IDFT pair at the receiver side, used for asynchronous echo canceller inhibits the low complexity implementation of the double talk cancellation. Therefore, an asynchronous echo canceller that can be implemented with just one additional DFT at the transmitter is presented, which enables the implementation of the double talk cancellation. In a synchronous echo canceller, the echo is completely determined from two consecutive echo reference symbols. However, in the asynchronous case, where there is a misalignment or delay of between the echo and received frames, three consecutive echo reference symbols u(k-1),u(k) , and u(k+1) , are needed to

completely determine the echo signal. Therefore, the echo is given by y(k)=a(k)h a(k)=a(k)+a(k) (10) (11)

All the operations are similar to that of synchronous case and we get

Department of ECE,GEC,Thrissur 12

E(K)=FN(y(k)-a(k)w(k))-diag{Ua(K)}W(k) 4.3 MULTIRATE ECHO CANCELLATION

(12 )

In DSL systems, different data rates are used for the downstream and the upstream. Therefore, if the symbol rate is equal in both directions, more samples are received at the remote terminal (RT) where higher data rates are required than at the central office (CO). Different multirate echo cancellation schemes are available 4.3.1 INTERPOLATED ECHO CANCELLER We assume that in the RT transceiver the transmitted signal bandwidth is k times smaller than that of the received signal. Because of the non-ideal reconstruction filter at the DAC, the higher frequencies of the transmitted signal leak into the received signal. This aliasing is modeled by interpolation of the transmitted signal before the convolution, which in the time domain corresponds to padding (k-1) zeros between adjacent samples. Thus, the error signal is given by E(K)=FN(y(k)-i(k)w(k))-diag{MU(K)}W(k) (13 )

where i(k) is defined as in the previous section with the difference that here, it is constructed from the interpolated (i.e., zero-padded) data; in the frequency domain, the interpolation is performed by the use of a matrix M which is the vertical concatenation of identity matrices M=[ IN.IN] T of order NN , the DFT matrix is of size and the weight vectors and are of length . The echo channel weight is subsequently updated by W(k+1)=W(k)+diag{MU*(K)}E(k) 4.3.2 DECIMATED ECHO CANCELLER At the CO transceiver the transmitted signal bandwidth is k times larger than that of the received signal and aliasing occurs because of the non-ideal anti-aliasing filter at the ADC. In order to consider the effect of the aliasing in the echo canceller the transmitted signal is decimated; in the time-domain, this
Department of ECE,GEC,Thrissur 13

(14 )

corresponds to downsampling by and in the frequency-domain, it corresponds to a block and add operation, i.e., pre-multiplication by MT . Thus, the error signal is given by E(k)=FN(y(k)-d(k)w(k))-diag{MTU(K)}W(k) Where d(k) contains the decimated data in the time domain. The weight vectors and are of length kN and are updated by W(k+1)=W(k)+diag{U*(K)}ME(k) where matrix M is used to replicate the error vector E(k) (16) (15 )

Department of ECE,GEC,Thrissur 14

5. LINEARLY CONSTRAINED ADAPTIVE ECHO CANCELLATION

As discussed in the previous section, the echo cancellation in DMT-based DSL systems relies on the dual filtering in the time and frequency domains, where the filtering weights are mapped from the frequency domain into the time domain. Here echo cancellation problem is considered as a constrained optimization problem on an extended linear space containing both the time and frequency domain weights. The constraints originally are used to ensure the proper mapping of the weights from the frequency domain into the time domain. Later, additional constraints can be added to improve the performance of the system. 5.1 LCA ECHO CANCELLER BASED ON CONSTRAINED LMS ALGORITHM Let us consider a frame synchronous system with equal rates at the transmitter and the receiver. The echo cancellation for DMT-based systems can be represented as a constrained optimization problem, where a cost function is minimized under certain constraints. The optimization is performed over an extended linear space, which is spanned by the extended weight vector, containing the weights both in the time and frequency domains, as given by (k) = w(k) W(k) 2N1 Where (k) is a vector of size 2N1 . As seen in the previous section, the error signal is used to update the adaptive weights. This signal can also be rewritten in terms of the extended weight vector, as follows: E(k)=FNy(k)-[FN(k)|diag{U(K)}](k) (17)

Department of ECE,GEC,Thrissur 15

The error signal can be rewritten as E(K)=FNy(k)-H(k)(k) Where (k) is echo reference matrix. Using this extended notation, the echo cancellation can be expressed as a constrained optimization, where the error power is minimized subject to a linear constraint on . This can be formulated as min E[||E(k)||2] s.t CH (k)=g where C is constraint matrix of size 2NMC In our derivation, we assume a general set of Mc linear constraints described by the constraint matrix C of size 2NMC and vector of size Mc 1, where Mc N and matrix is full column rank. The constraint is primarily used to ensure the Fourier transform relation between the weights in the time and frequency domains. However, additional constraints can be added to improve the system performance, as discussed later.Using the method of Lagrange multipliers the constrained optimization can be transformed into an unconstrained one. Therefore, the constraint can be appended to the cost function, as given by J=minE[||E(k)||2]+(H(k)C-gH)+H(CH(k)-g) (19) (18)

Where is the Lagrange multipliers vector of size Mc 1. Complying with the method in constrained gradient-descent optimization is used where the weight vector is adaptively updated by the gradient of the constrained cost function, as follows (k+1)=(k)-wHJ where wHJ=R(k)-P+C R-correlation matrix (21) (20)

Department of ECE,GEC,Thrissur 16

P-cross correlation matrix Based on constrained LMS algorithm the iterative update of the extended weight is given by (k+1)=wq+PC((k)+(k)E(k)) (22)

where wq =C(CHC)-1g is the quiescent term and depends only on the constraint. C(CHC)-1CH represents the projection onto the subspace spanned by the constraint matrix PC = IN-PC is the projection onto its orthogonal complement. Also0<<max Where is the largest eigenvalue of the matrix PCR PC (23)

Department of ECE,GEC,Thrissur 17

6. CONCLUSION
Multicarrier transmission metlhods have long been known to optimize the performance of data transceivers on bandlimited communication channels. One form of multicarrier transmission, known as discrete multitone modulation (DMT), is particularly attractive for its ability to be implemented using efficient digital signal processing techniques.. Given a basic DMT system, it is possible to increase the aggregale data rate with fullduplex transmission using echo cancelation. However, DMT echo cancelation at first appears difficult because of the computational complexity required in a straightforward implementation to cancel the cross-echoes produced by each carrier into every carrier. Here we introduced a constrained adaptive echo canceller for DMTbased DSL systems where echo is cancelled partially in the frequency and time domain. As a basic to his topic we discussed about discrete multi tone modulation(DMT), working principle of a basic echo canceller and also synchronous, asynchronous, and multi rate echo canceller.

Department of ECE,GEC,Thrissur 18

7. REFERENCES
[1.] Neda Ehtiati, Member, IEEE, and Benot Champagne, Senior Member IEEE, IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 57, NO. 1, JANUARY 2009 [2.] T. Starr, J. M. Cioffi, and P. J. Silverman, Understanding Digital Subscriber Line Technology. Englewood Cliffs, NJ: Prentice-Hall, 1999. [3.] M. Ho, J. M. Cioffi, and J. A. C. Bingham, Discrete multitone echo cancelation, IEEE Trans. Commun., vol. 44, pp. 817825, Jul. 1996. [4.] D. C. Jones, Frequency domain echo cancellation for discrete multitone asymmetric digital subscriber line transceivers, IEEE Trans. Commun., vol. 43, pp. 16631672, Feb./Mar./Apr. 1995. [5.] G. Ysebaert, K. Vanbleu, G. Cuypers, M. Moonen, and K. V. Acker, Double talk cancellation in echo cancelled DMT-systems, in Proc. Eur. Signal Processing Conf., Toulouse, France, Sep. 2002, vol. II, pp. 381384. [6.] G. Ysebaert, K. Vanbleu, G. Cuypers, M. Moonen, and J. Verlinden, Echo cancellation for discrete multitone frame-asynchronous ADSL transceivers, in Proc. IEEE Int. Conf.Communications, Seattle, WA, May 2003, vol. 4, pp. 24212425. [7.] G. Ysebaert, F. Pisoni, M. Bonaventura, R. Hug, and M. Moonen, Echo cancellation in DMT-receivers: Circulant decomposition canceler, IEEE Trans. Signal Process., vol. 52, no. 9, pp. 26122624, Sep. 2004. [8.] L. Griffiths and C. Jim, An alternative approach to linearly constrained adaptive beamforming, IEEE Trans. Antennas Propag., vol. 30, no. 1, pp. 2734, Jan. 1982.

Department of ECE,GEC,Thrissur 19

[9.] DIGITAL COMMUNICAION 2nd edition by Edward A Lee and David G Messerschmitt [10.] WIRELESS COMMUNICATION By Andrea Goldsmith

Department of ECE,GEC,Thrissur 20