Practical Quantize-and-Forward Schemes for the Frequency Division Relay Channel

We consider relay channels in which the source-destination and relay-destination signals are assumed to be orthogonal and thus have to be recombined at the destination. Assuming memoryless signals at the destination and relay, we propose a low-complexity quantize-and-forward (QF) relaying scheme, which exploits the knowledge of the SNRs of the source-relay and relay-destination channels. Both in static and quasistatic channels, the quantization noise introduced by the relay is shown to be signiﬁcant in certain scenarios. We therefore propose a maximum likelihood (ML) combiner at the destination, which is shown to compensate for these degradations and to provide signiﬁcant performance gains. The proposed association, which comprises the QF protocol and ML detector, can be seen, in particular, as a solution for implementing a simple relaying protocol in a digital relay in contrast with the amplify-and-forward protocol which is an analog solution.


INTRODUCTION
The channels under investigation in this paper are quasistatic orthogonal relay channels for which orthogonality is defined accordingly to [1].Since the source-destination channel is assumed to be orthogonal to the relay-destination channel (i.e., the forward channel), the destination receives two distinct signals.For the channels under consideration, there are at least two important technical issues: the relaying protocol and the recombination scheme at the destination.So far, three main types of relaying protocols have been considered in the literature: amplify-and-forward (AF), decodeand-forward (DF), and estimate-and-forward (EF).From the corresponding works, several observations can be made: (a) from information-theoretic studies like [1,2], it appears that the best choice of the relaying scheme depends on the source-relay channel (i.e., the backward channel) signal-tonoise ratio (SNR) and that of the relay-destination channel; (b) there are not many works dedicated to the design of practical EF schemes although the EF protocol has the potential to perform well for a wide range of relay receive SNRs (in contrast with DF which is generally more suited to relatively high SNRs); (c) the AF protocol is generally chosen for its simplicity but implementation-related issues are often ignored.In particular, while the DF protocol is clearly suited to digital implementations, most of the existing research on the AF protocol makes the questionable assumption that relays are perfect analog devices which forward a scaled copy of the received signal.
One of the motivations for the work presented in the paper is precisely to propose low-complexity relaying schemes (comparable to the AF protocol complexity) that can be implemented in a digital relay transceiver (in contrast with the AF protocol) and use the knowledge of the SNRs of the forward and backward channels in order for the relay to optimally adapt to the forward and backward channel conditions.To achieve these goals, the main solution proposed is a quantize-and-forward (QF) protocol for which forwarding is done on a symbol-by-symbol basis and aims to minimize the mean square error (MSE) between the source signal and its reconstructed version at the output of the dequantizer at the destination.Some researchers have also referred to the classic Wyner-Ziv source coding scheme in [3] as QF [4,5].Our practical approach, which ultimately aims to minimize the raw bit-error rate (BER) at the destination for a fixed transmit spectral efficiency and does not exploit error correcting coding, differs from these information-theoretic works.It also differs from other practical studies on EF protocols, such as [6][7][8] in the sense that the corresponding relaying schemes are not analytically optimized by taking the SNRs of the backward and forward channels into account.Rather, our work is based on the joint source-channel coding approach originally introduced in [9] for the Gaussian pointto-point channel where the authors extended the original iterative Lloyd algorithm by designing a scalar quantizer that takes into account the channel through which the quantized Gaussian source is to be transmitted.The authors of [10] applied this approach in the context of the binary symmetric channel (BSC) and proved that the corresponding distortion is a nonincreasing function of the number of iterations of the optimization algorithm.In this paper, we further extend the iterative algorithm of [9] in the context of quasistatic orthogonal relay channels by taking into account both the forward and backward channels and providing a nonrestrictive sufficient condition for convergence of the derived algorithm, similarly to [10].
This paper is organized as follows: in Section 2, the signal model for the orthogonal relay channel, main assumptions, and notation are given; in Section 3, the proposed QF scheme and a modified AF scheme are provided; in Section 4, we propose an ML detector (MLD) in order to account for the quantization noise introduced by the relay; in Section 5, the proposed schemes are evaluated in terms of raw BER and compared with AF, which serves as a reference strategy; concluding remarks are provided in Section 6.

SYSTEM MODEL
The source is assumed to be represented by a discrete-time signal x, which takes its value in the finite set of equiprobable symbols X = {x 1 , . . ., x Ms } and is subject to a unit average power constraint: E[|x 2 |] = 1.For sake of simplicity, square M s -QAM symbols with independent real and imaginary parts are assumed.More importantly, the samples of the source, denoted by x(n) where n is the time index, are assumed to be independent and identically distributed (i.i.d.) as in [9,10].In the context of digital communications, this assumption is generally valid because of interleaving, dithering, or equivalent operations.In order to limit the relay and receiver complexity, we will not exploit the interactions between the quantizer and the error correcting coders, possibly present at the source and relay.Therefore the assumption made on the source samples and channel model (described just below) implies that there is loss of optimality by assuming scalar quantizers, that is, symbol-by-symbol forwarding at the relay, instead of vector quantizers [11].At each time instant n the source broadcasts the signal x(n), which is received by the destination and relay nodes.The received baseband signals can be written: where w sd and w sr are zero-mean circularly symmetric complex Gaussian noises with variances σ 2 sd and σ 2 sr , respectively.The complex coefficients h sd and h sr are, respectively, the gains of the source-destination and source-relay channels.In this paper, for simplicity of presentation, most of the derivations are conducted for static channels, so h sd and h sr are constant over the whole transmission.Therefore, the presence of these gains makes only sense in the quasi-static case whereas in the context of static channels they could be removed.In this case, these quantities are assumed to be constant over a block duration and vary from block to block.In the simulation part both cases will be analyzed and Rayleigh block-fading will be assumed for modeling the channel gains in the case of quasistatic channels.In this case, for each block, h sd and h sr are the realizations of two independent Gaussian complex random variables.Note that, thanks to the independency assumption between all the fading gains, the presence of the relay will provide more degrees of freedom in the channel, which will be exploited at the receiver through signal combiners that will provide a diversity order of two instead of one (this assertion can be proven for the two combiners provided in this paper, for more information see [12]).Therefore, one has to keep in mind that in quasistatic channels the performance gain due to the presence of the relay can also come from the qualities of the source-relay and relaydestination channels but is, in general, essentially due to the higher diversity order.In static channels (namely, Gaussian channels or fading channels with a strong Rician component) only a gain in terms of SNR can be expected.
The relay forwards the cooperation signal x r (n) to the destination.We assume memoryless and zero-delay relaying.The memoryless assumption is a consequence of the previously mentioned independence assumptions while the zerodelay assumption can be satisfied by resynchronizing the direct and cooperation signals at the destination.Under these assumptions x r (n) = f (x sr (n)) for some memoryless function which is chosen to satisfy a unit average power constraint E[|x r | 2 ] = 1.Since the relay function and channels are memoryless, in the sequel we will at times omit the time index n from the signals.For the QF protocol the relaying function comprises a zero-memory quantization operation (denoted by Q) followed by an M r -QAM modulation (denoted by M).In the case of the clipped AF protocol, there is no modulation since the relay is assumed to generate a continuous signal.The cooperation signal received at the destination is written: where the notation is defined above.Orthogonality between the received cooperation signal y rd and direct signal y sd can be implemented by frequency division (FD).The optimal bandwidth allocation issue is beyond the scope of this paper, thus we assume that y sd and y rd have the same bandwidth.At the destination, two types of combiners can be assumed.We will use either a conventional maximum ratio combiner (MRC) or a more sophisticated detector, namely the MLD, which will be derived in Section 4. The reason for introducing the latter combiner will be clearly explained in Section 4. Figure 1 summarizes the system model when QF is assumed.The notation D stands for decoder, which jointly incorporates the demodulation and de-quantization operations.On the other hand, when the relay amplifiesand-forward, D is the identity operator and Q and M are replaced with a linear function in the AF case and a nonlinear function in the clipped AF case (Section 3.2).

Optimum quantize-and-forward
The most natural way to quantize and forward the signal received by the relay is to quantize x sr in order to minimize the distortion , map the quantizer output onto a QAM modulation and send it to the destination.As σ 2 rd →0 and the number or quantization bits increases, this quantization strategy becomes optimum since it achieves the performance of a 1 × 2 single-input multiple-output (SIMO) system.On the other hand, if x sr is quantized with a high number of bits and sent through a bad cooperation channel, minimizing D 00 is no longer optimal.This is why minimizing can be more efficient as shown by [9,10,13,14] in the context of the point-to-point Gaussian channel.In the context of the relay channel we know that the source-relay channel quality also plays a role in the receiver performance.Therefore, we propose minimizing the MSE between the reconstructed signal x rd and the original source signal x, that is, , by assuming the SNRs of the forward and backward channels are known to the relay.The disadvantage of minimizing D 11 is that in the high cooperation regime the SIMO performance is not reached.We will comment on this point further (Section 5.2).
Let us turn our attention to the quantizer itself.Since the signal to be quantized is complex, the quantizer is composed of two "subquantizers," one for the real part of x sr and one for its imaginary part.Quantizing consists in mapping the signal x sr into a pair of rational numbers belonging to where L= 2 b/2 and b is the total number of quantization bits.As the real and imaginary parts of the signal received by the relay are assumed to be independent, the two subquantizers can be designed independently and in the same manner.This is why, from now on, we restrict our attention to the subquantizer of the real part of x sr .The subquantizer maps x R sr Re(x sr ) onto the finite set of be the set of the transition levels of the subquantizer.The aforementioned mapping is done as follows: ) then its representative is v R j , where j ∈ {1, 2, . . ., L}.The quantizer output is then mapped onto the constellation following the idea of [15].The mapping is done in such a manner that close representatives in the signal space are assigned to close symbols in the modulation space.
Therefore, the most likely decision errors which appear in the neighborhood of the symbol associated with the input representative will result in a slight increase in distortion.We now describe the quantizer optimization procedure.To find the optimal pair of subquantizers at the relay, we minimize the MSE D 11 as follows.The distortion can be written as ( As D R 11 and D I 11 can be optimized independently and identically, we focus, hence forth, on minimizing D R 11 .The latter quantity can be shown to expand as where , the forward channel statistics) and sr ) is the probability density function (pdf) of the real noise component Re(w sr ) of the signal received by the relay (i.e., the backward channel statistics).Given a number of quantization bits, we now optimize the subquantizer Q R by minimizing D R 11 with respect to the transition levels {u R } ∈{1,...,L} and the representatives {v R } ∈{1,...,L} .For fixed transition levels, the optimum representatives are the centroids of the corresponding quantization cells which are obtained by setting the partial derivatives of D R 11 to zero: When the representatives are fixed, it is not trivial, in general, to determine the transition levels explicitly as is the case for conventional channel optimized quantizers such as [10] for which the backward channel is not present.The difficulty is due to the presence of the function φ(•) in the MSE expression (for more information see Appendix A).Determining the transition levels then requires the use of an exhaustive search algorithm.However, note that there are simple cases such as a 4-QAM source, which is used in the simulations (Section 5), where both the optimum representatives for fixed transition levels and optimum transition levels for fixed representatives can be found.For a 4-QAM constellation, we have (x R , x I ) ∈ {−A, +A} 2 .For fixed transition levels, we have for all ∈ {1, . . .
and for fixed representatives we have Note that in (7) the strict positiveness of the argument of the logarithm ensures the existence of the optimum transition levels.We are now in position to provide the complete iterative optimization procedure.Let i and be the iteration index and the current value of the estimation error criterion of the iterative algorithm.The algorithm is said to have converged when reaches max .
with the sets (say V R (0) and U R (0) ) obtained from [10], which correspond to a local optimum since the backward channel is not taken into account.(ii) Step 2. Set i→i+1.For the fixed partition U R (i−1) use ( 6) to find the optimal codebook V R (i) .For the fixed codebook V R (i) use (7) to obtain the optimal partition U R (i) .

If the realizability condition
L is not met, stop the procedure and keep the transition levels provided by the previous iteration.(iii) Step 3. Update as follows: If ≥ max , then go to Step 2, stop otherwise.
As with other iterative algorithms (e.g., the EM algorithm), one cannot easily prove or ensure, in general, convergence to the global optimum.When the backward channel was not present, the authors of [10] proved that the distortion obtained by applying the generalized Lloyd algorithm is a nonincreasing function of the number of iterations.The authors provided a sufficient condition under which the procedure is guaranteed to converge towards a local optimum.The corresponding condition is not restrictive since it can be imposed through the realizability constraint of the transition levels [10] to the iterative procedure without loss of optimality.Recall that this constraint consists in imposing u R to be an increasing function of .It turns out a similar result can be derived in our context (see Appendix A) if one assumes a zero-mean channel input (i.e., E[X R ] = 0) and the backward channel noise to be Gaussian.The obtained condition is as follows: at each iteration step, ∀ ∈ {1, . . .
If this condition is met, the MSE will be a nonincreasing function of the iteration index.
To conclude this section we will make a few comments on the complexity of the proposed protocol.Compared to vector quantizers [16], the proposed solution is much simpler since the creation, storage and computation complexities (for more information see, e.g., [17]) both grow exponentially with the cell dimension (which is 1 for scalar quantizers).If one wants to further decrease the complexity of the quantizer, it is possible to simplify the proposed algorithm by imposing the quantizer to be uniform (equispaced transition levels and representatives).Since the uniform quantizer is entirely specified by its quantization step there is only one parameter to be determined.We will not conduct a complexity analysis here but it can be checked (Appendix C) that the ratio of the optimum QF protocol complexity to that of the uniform version is of the order of the number of iterations of the proposed algorithm, which is typically between 5 and 10 in simulations.The uniform QF protocol can be obtained by using [9] and by specializing the results presented here.The performance of the corresponding scheme will be presented in the simulation part.

Clipped amplify-and-forward
In this section, we propose a modified version of the AF protocol.Our motivation for proposing this new version of AF is threefold.First, it optimizes the same performance criterion as for the QF schemes, that is, the end-to-end distortion.Second, it allows us to fairly compare the scalar QF schemes with the scalar AF scheme given the fact that the conventional AF does not exploit the knowledge of the SNRs of the sourcerelay and relay-destination channels.Third, the clipped AF bridges the gap between the QF and AF protocols since it allows us to isolate the clipping effect naturally introduced by the QF schemes.So, we now replace the quantizer with a piecewise linear saturation function, which simply clips samples with magnitude above a chosen threshold β > 0. The linear threshold function which operates independently on the real and imaginary parts of the signal is defined as where we considered the case of the real part.We see that the relay acts like a perfect AF relay in the region [−β, β] and limits values outside this region.Our motivation for using this function is to assess the benefits from clipping x sr but in some context better relaying functions can be used.For example the authors of [18] derived the best relaying function in the sense of the raw BER when no direct link is assumed and a BPSK modulation is used both at the source and relay.In our context, the goal is different and the extension of [18] to the case of QAM modulations does not seem to be trivial.In the same spirit, [19] proposed an optimized relaying function in the sense of the mutual information when no direct link is assumed.The rationale for the proposed function ( 9) is that it preserves the important soft information but does not needlessly expend power relaying large noise samples.Furthermore, it only requires the optimization of a single parameter, that is, the clipping level β.In spite of the seeming simplicity of the relaying function, however, calculating the p.d.f. of saturated Gaussian signals is known to be intractable [20].After passing the received signal through the saturation function, the signal is scaled by some real parameter α which is chosen to satisfy the average unit power constraint.Of course, in order to ensure coherent reception at the relay node, the incoming signal also has to be equalized.
Here, our choice is the MMSE (minimum MSE) equalizer: where α is such that E[|x r | 2 ] = 1, and f R β (•) and f I β (•) are defined identically since the real and imaginary parts are assumed to be i.i.d.Due to the fact that the data and noise are independent and by calculating the first and the second-order moments (Appendix B) for the random clipped gaussian variable, we find that which can be set equal to 1 to find the scaling factor α β that satisfies the power constraint.Note that Q is the classical error function: ∞ x e −t 2 dt.To find the clipping level β, we minimize the MSE between the source signal andthe signal received by the destination on the cooperation channel: where we note that α β is effectively a function of β since any change in the clipping level affects the scaling required to satisfy the power constraint.The β that minimizes this func-tion cannot be written in closed form.However, it is purely a function of the source-relay and relay-destination SNRs, so it can be computed numerically offline using (13) and the calculation of the first and second-order moments for the clipped Gaussian (Appendix B) and stored in a lookup table.Note that this implies that the relay needs to know the SNRs on its channel both to the transmitter as well as to the receiver, which was also the case in the QF protocol.

COMBINING SCHEMES
When the AF protocol is assumed at the relay, the optimum combiner in terms of raw BER is the MRC.When using the clipped version of the AF protocol this is no longer true since the equivalent additive noise in the relay-destination channel is not Gaussian.As already mentioned, calculating the pdf of saturated Gaussian signals is known to be intractable.Therefore, we will still use the MRC at the destination when the clipped AF is used.We will see through the simulation analysis that this issue does not seem to be critical but deriving a better combiner might be seen as an extension of this work.On the other hand, when QF is assumed, using the MRC at the destination can lead to a significant performance loss.In this respect the authors have shown in [21] that using the DF protocol with a conventional MRC when the relay is in bad reception conditions can severely degrade the BER performance at the destination with respect to the case without cooperation.This is in part because the relay generates non-Gaussian residual decoding noise that is correlated with the useful signal.For the QF protocol the combiner choice might look less critical since the relay does not make a decision on the transmitted symbols.However, for a low number of quantization bits and relay receive SNR, the answer is not clear.This is why we not only consider the MRC but also propose a more sophisticated detector (namely, the MLD) adapted to the QF protocol, which is derived as follows.
Assume the symbol transmitted by the source is x and the quantizer output Q(x sr ) = v i .The likelihood p ML = p(y sd , x rd | x) can be factorized as where p(y sd | x) = (1/πσ 2 sd ) exp (−|y sd − h sd x| 2 /σ 2 sd ).To expand the second term p( x rd | x), we recall that X rd ∈ V R × V I = {v 1 , v 2 , . . ., v Mr }, and we make use of the channel transitions probabilities P k, between complex representatives 6 EURASIP Journal on Wireless Communications and Networking (see Section 3.1) where we have defined P R k, for the real part of complex representatives.We have where the index j corresponds the symbol of the relay alphabet (i.e., {1, . . ., M r }) associated with the pair of representatives (v R , v I m ).Now, by denoting s = (s 1 , . . ., s N ), the vector of bits associated with the source symbol x allows us to express the log-likelihood ratio for the nth bit: where the sets S (i) 1 and S (i) 0 are defined by then s n = 1 and s n = 0 otherwise.

SIMULATION ANALYSIS
We assume a 4-QAM source and consider different simulation scenarios with the following parameters: (i) the channels can be either static (Gaussian or purely Rician) or quasistatic (Rayleigh block-fading model); in the latter case the channels are constant over a block duration; each block comprises 100 symbols; we note that the case of static channels can correspond to real situations in wireless communications, for example, fixed users using laptops connected to a hot-spot; (ii) the relative quality of the relay: SNR sr [dB] = SNR sd [dB] + ρ, where ρ ∈ {−5 dB, 0 dB, +10 dB}; (iii) the number of quantization bits used by the QF protocol: b ∈ {2, 6} (i.e., b/2 bits per subquantizer); (iv) the relay-destination channel quality: SNR rd [dB] ∈ {0 dB, 10 dB, 30 dB} with SNR rd = 1/σ 2 rd ; (v) the relaying scheme: AF, optimally clipped AF, uniform QF, and optimum QF; for reference, we will consider the case where no relay is available (a BPSK is then used at the transmitter in order to make a fair comparison in terms of spectral efficiency) and also the full cooperation case; the latter is defined as follows: σ rd →0 and the AF protocol is used; we will refer to this case as the SIMO bound; (vi) the combining scheme at the receiver: MRC or MLD.

Optimum QF versus uniform QF
All the simulations we performed showed one significant drawback of the uniform QF relaying protocol.Both in static and quasistatic channels, the receiver performance, when using the uniform QF protocol with MRC or MLD, is sensitive to the choice of the number of quantization bits.This tendency is clearly more marked for static channels.For example, see Figures 2 and 3. Figure 2 shows that using the uniform QF with b = 6 bits can lead to a significant performance loss.This appears when the source-relay SNR is sufficiently large and the cooperation channel has medium quality.In this situation it is better to decode and forward than quantize and forward a signal that is not robust to cooperation channel noise.When b/2 = 1 the uniform QF roughly behaves like DF while it behaves more like AF for b = 6, which explains why the performance is better for b = 6 in Figure 3.Our interpretation is that the uniform QF has only one degree of freedom (namely, its quantization step) to adapt to SNR sr and SNR rd .For a fixed number of bits, there will always be scenarios where the performance of the uniform QF can be much less than the optimum relaying scheme (AF, DF, or optimum QF) used in the considered setup.On the other hand, the number of quantization bits has much less influence on the performance of the optimum QF when the MLD is employed at the receiver.By analyzing many simulations, which are not provided here due to lack of space, we have observed  that it is generally better to choose a sufficiently high number of bits (typically 3 bits per dimension) regardless of the SNRs of the different channels.Our explanation is that the optimum quantizer produces a grid of centroids that looks like the source constellation.The constellation in the output of the quantizer looks like a constellation with 2 resolution levels: there are 4 clouds (for a 4-QAM) of centroids, with each cloud comprising 2 b−2 centroids that are typically concentrated around the cloud center.Depending on SNR sd and SNR sr , the optimum QF can adapt both the location of the cloud centers and the points around each center.

Comparison between the different relaying protocols
Many simulations showed us the following trend: in quasistatic channels, the receiver performs quite similarly no matter which relaying protocol (AF, clipped AF, or optimum QF) is used, provided that the preferred combining scheme is employed (i.e., the MRC is used for AF and clipped AF, and MLD is used for optimum QF).This is essentially due to the averaging effect of the channel conditions.Figure 4 compares the receiver performance of AF + MRC with optimum QF + MLD. Figure 5 shows that the conventional and clipped AF protocols perform similarly.However, the relaying strategy is more influential in static channels.Figure 6 shows a typical example.Other simulations with different numbers of quantization bits and SNR values can be roughly summarized as follows: for low and medium transmit or cooperation powers, the optimum QF provides the best performance whereas the performance loss in the high cooperation regime is always small, which means that the SIMO bound is almost achieved by opti-    mum QF in the latter regime.Also the AF tends to perform better than the QF protocol in situations where the source-relay channel is bad.Now let us comment on the effect of clipping the signal received by a relay using the AF protocol in static channels.The obtained performance gain obtained by clipping depends on SNR sr and SNR rd .For low and medium cooperation channel qualities, this  gain typically ranges from 0.5 dB to 1.5 dB, depending on SNR sr .In the high cooperation regime, it is small and can even be slightly negative since the clipped AF minimizes the distortion while the AF reaches the SIMO bound when SNR rd →∞.

Importance of the combining scheme for the QF protocol
As already mentioned, when optimum QF is assumed, the facts that the receiver performance is not very sensitive to the number of quantization bits and is close to that obtained by the AF protocol is in part due to the use of the MLD instead of MRC.This can be shown both in static and quasistatic channels.In this subsection, we want to illustrate this point by an explicit comparison.Figures 7 and 8, respectively, represent the receiver performance over quasistatic channels in two markedly different scenarios: (a) a good relay, a good cooperation channel, and b = 6; (b) a bad relay, a medium quality cooperation channel, and b = 2.In both cases the MLD brings a significant performance gain, which shows the importance of using a receiver structure adapted to the assumed relaying scheme.

CONCLUSION
We have proposed a low-complexity quantize-and-forward scheme, which exploits the knowledge of the SNRs of the source-relay and relay-destination channels.In static channels it generally performs close to or better than the conventional or clipped AF protocols.Also, based on knowledge of the SNRs, clipping can provide a nonnegligible (and almost free in terms of complexity) gain with respect to the conventional AF, whose value depends on the different SNRs.Over Rayleigh block-fading channels, we have seen that the optimum QF protocol, provided it is associated with an ML detector, has generally similar performance to the conventional or clipped AF protocols, whatever the simulation scenario.Although the clipped AF and QF protocols can be shown not to be strictly equivalent for a high number of quantization bits (because of the presence of the dequantizer at the end of relay-destination channel), the following comment can be made: since the optimum QF protocol is both scalar and simple and generally performs closely to the AF protocol, this shows that the proposed solution can be seen as a way of implementing a channel-optimized AFtype protocol in a digital relay transceiver.Now, if the relay and receiver complexity can be relaxed, the proposed approach can be improved by exploiting the structure inherent to channel coding, which can be seen as an extension of this work.

APPENDICES A. A SUFFICIENT CONDITION FOR CONVERGENCE OF THE MSE IN THE OPTIMUM QUANTIZER DESIGN
First, we derive the MSE expression in our context: Assume the transition levels to be fixed.Then the MSE is a strictly convex function of v R over R. Indeed, the second partial derivative of the MSE with respect to v is given by the following expression: x R j )dt.For all ∈ {1, . . ., L}, the strict positiveness of this second derivative implies that updating the representatives v according to (5) cannot increase the overall MSE.Now, assume the representatives are fixed.The second partial derivative of the MSE with respect to u can be expanded as follows: Today such sequential design flow is reaching its limits due to:

EURASIP Journal on Embedded Systems
• The complexity of today's systems designed with the emerging submicron technologies for integrated circuit manufacturing • The intense pressure on the design cycle time in order to reach shorter time-to-market, reduce development and production costs • The strict performance constraints that have to be reached in the end, typically low and/or guaranteed application execution time, integrated circuit area, overall system power dissipation An alternative approach to a traditional design flow, called algorithmarchitecture matching, aims to leverage the design flow by a simultaneous study of both algorithmic and architectural issues, taking into account multiple design constraints, as well as algorithm and architecture optimizations, not only in the beginning but all the way throughout the design process.
Introducing such design methodology is also necessary when facing the new emerging applications such as high-performance, low-power, low-cost mobile communication systems and/or smart sensors-based systems.
This design methodology will have to face also future architectures based on multiple processor cores and dedicated coprocessors to achieve the required efficiency.NoC-based communications will become also mandatory for many applications to enable parallel interconnections and communication throughputs.Adaptive and reconfigurable architectures represent a new computation paradigm whose trend is clearly increasing.
This forms a driving force for the future evolution of embedded system designs methodologies.
This special issue of the EURASIP Journal of Embedded Systems is intended to present innovative methods, tools, design methodologies, and frameworks for algorithm-architecture matching approach in the design flow including system level design and hardware/software codesign, RTOS, system modeling and rapid prototyping, system synthesis, design verification and performance analysis and estimation.Because in such design methodology the system is seen as a whole, this special issue will also cover the following topics: • New and emerging architectures: SoC, MPSoC, configurable computing (ASIPs), (dynamically) reconfigurable systems using FPGAs • Smart sensors: audio and image sensors for high performance and energy efficiency • Applications: Automotive, medical, multimedia, telecommunications, ambient intelligence, object recognition, cryptography, wearable computing This special issue is open to all contributions.Authors are invited to submit their papers addressing the domain of design and architectures for signal and image processing.We also strongly encourage authors who presented a paper to the DASIP 2007 workshop to submit an extended version of their original workshop contributions.
Authors should follow the EURASIP Journal on Embedded Systems manuscript format described at the journal site http://www.hindawi.com/journals/es/.Prospective authors should submit an electronic copy of their complete manuscript through the journal Manuscript Tracking System at http://mts.hindawi.com/according to the following timetable:

Figure 1 :
Figure 1: System model for the quantize-and-forward protocol.

Figure 2 :
Figure 2: Influence of b on the performance when the uniform QF protocol is used over static channels: raw BER versus SNR sr at the output of the MRC for SNR rd = 10 dB with SNR sr = SNR sd + 10 dB.
xsr p x sr , x rd = v i | x dx sr = xsr p x sr | x p x rd = v i | x sr dx sr = Mr j=1 xsr∈Sj p x sr | x p x rd = v i | x sr dx sr = coop Uniform QF minimizing D 11 with b = 2 Uniform QF minimizing D 11 with b = 6 SIMO bound

Figure 3 :
Figure 3: Influence of b on the performance when the uniform QF protocol is used over static channels: raw BER versus SNR sr at the output of the MRC for SNR rd = 10 dB with SNR sr = SNR sd − 5 dB.
coop Relay Optimal QF with MLD b = 6 Amplify-and-forward SIMO bound

Figure 4 :
Figure 4: Comparison between the optimum QF (b = 6) and the AF schemes in quasistatic channels for SNR rd = 40 dB, SNR sr = SNR sd + 10 dB.

Figure 5 :
Figure 5: Comparison between the AF and clipped AF protocols in quasistatic channels for SNR sr = SNR sd + 10 dB.

Figure 6 :
Figure 6: Comparison of the different relaying schemes (AF, clipped AF, optimum QF with b = 6) in static channels for SNR sr = SNR sd + 10 dB with SNR rd = 10 dB.

Figure 7 :
Figure 7: Influence of of the combining scheme for the optimal QF scheme (b = 2) in quasistatic channels with {SNR rd = 40 dB, SNR sr = SNR sd + 10 dB}.

Figure 8 :
Figure 8: Influence of of the combining scheme for the optimal QF scheme (b = 2) in quasistatic channels with {SNR rd = 10 dB, SNR sr = SNR sd − 10 dB}.
h t t p : / / w w w .h i n d a w i .c o m / j o u r n a l s / e s / Special Issue onDesign and Architectures for Signal and Image ProcessingCall for PapersThe development of complex applications involving signal, image, and control processing is classically divided into three consecutive steps: a theoretical study of the algorithms, a study of the target architecture, and finally the implementation.