Map decoding method using augmented lattices

Abstract

The invention relates to a MAP decoding method of a signal received through a noisy channel, the signal being composed of symbols in an alphabet having a non-uniform probability distribution, the symbols being represented by points in a lattice (Λ). The probability distribution of symbols is modeled using a Gaussian distribution. An augmented lattice (Λexp) is formed from the lattice (Λ) and the ratio (β) between variance of the noise and variance of the Gaussian distribution of symbols. Therefore, the disclosed MAP decoding method consists essentially of decoding using an ML criterion searching the point in the augmented lattice closest to the point representative of the received signal (yexp).

Description

TECHNICAL DOMAIN

This invention relates to the domain of digital telecommunications and more particularly decoding of an information source of which the symbols do not follow a uniform distribution. It is particularly applicable to decoding in networks using a physical layer network coding technique.

STATE OF PRIOR ART

Digital communications are usually based on a discrete channel model between a transmitter capable of transmitting messages composed of symbols and a receiver capable of reproducing the message from the received signal.

A discrete channel can be considered as being a stochastic system accepting symbols x_i belonging to an alphabet A_X as input, and providing symbols y_i belonging to an alphabet A_Y as output, the channel inputs and outputs being related by a probabilistic model defined by probabilities:P(Y₁=y₁, . . . ,Y_m=y_m|X₁=x₁, . . . ,X_n=x_n)  (1)

In practice the channel is noisy, in other words symbols at the output from the channel are affected by noise that is usually assumed to be additive white Gaussian noise (AWGN). R_i=Y_i+Z_i, i=1, . . . , m denote the samples of the received signal samples in which z, are noise samples.

The receiver searches for the most probable sequence of symbols x₁, . . . , x_n given the received signal, in other words it searches for the sequence of input symbols that maximises the conditional probability:P(X₁=x₁, . . . ,X_n=x_n|R₁=r₁, . . . ,R_m=r_m)  (2)that we will denote more simply as P(X|R) where X is the vector of input symbols that is to be estimated and R the vector of the received signal samples.

In other words, we search for the vector {circumflex over (X)} that maximises the probability:P(X|R)=P(R|X)P(X)/P(R)  (3)

Since the term P(R) is independent of X, all that is necessary is to search for the vector {circumflex over (X)} maximising the probability:P(R|X)P(X)  (4)

At this stage, a distinction is made between two types of decoding methods. The first decoding type uses a criterion called the Maximum Likelihood (ML) criterion, based on the assumption that the symbols in the alphabet A_X are equally probable. All that is then necessary is to maximise P(R|X), in other words when the channel has no memory:

$\begin{array}{cc}{\hat{X}}_{\mathrm{ML}}=\underset{X}{\arg \max }P\left(R|X\right)=\prod _{i=1}^m\underset{x_i}{\arg \max }P\left(R_i|x_i\right)& \left(5\right)\end{array}$

This decoding method is optimal when the input symbols are effectively equally probable, in other words when the alphabet A_X has a uniform probability distribution.

A second type of decoding, called Maximum A Posteriori (MAP) decoding, does not make this simplifying assumption. Consequently, conditional probabilities have to be weighted by the probabilities of input symbols according to expression (5). When the channel has no memory, this expression becomes:

$\begin{array}{cc}{\hat{X}}_{\mathrm{MAP}}=\underset{X}{\arg \max }P\left(R|X\right)=\underset{x_i}{\arg \max }\prod _{i=1}^{m}P\left(R_i|x_i\right)P\left(x_i\right)& \left(6\right)\end{array}$

Although a large number of information sources have a non-uniform probability distribution of the symbols of their alphabet (hereinafter denoted non-uniform information sources), ML decoding is usually preferred due to its greater simplicity, despite its sub-optimal nature.

Non-uniform information sources include most multimedia information sources (text, speech, audio, video) due to the strong correlation between successive symbols (bits) contained in them. Although information compression can strongly reduce this redundancy, it does not eliminate it completely.

Non-uniformity of the probability distribution of the symbols to be decoded is not necessarily the result of the information source but it may be due to the communication strategy used. Thus, relay terminals in cooperative networks using a physical layer network coding technique, receive combinations of signals originating from different source terminals. An introduction to physical layer network coding is given in the article by S. Zhang et al. entitled «Physical Layer Network Coding», published in Proc. Of ACM Mobicom, 2006, pp. 358-365. Although the probability distribution of symbols transmitted by the source terminals is uniform, the combination of these symbols at the relay terminal is not. Thus, in the simple example shown in FIG. 1 in which two source terminals S₁,S₂ transmit symbols x₁ and x₂ to two relay terminals R₁,R₂ that decode them before recoding them if necessary and retransmitting them to a destination terminal D, if gains of unit channels are assumed between sources and relays, the relays will both receive the sum symbol x₁+x₂. If the source terminals use a BPSK alphabet with uniform probability, the probabilities of sum symbols −2, 0, 2 will be ¼, ½ and ¼ respectively. More generally, if a relay terminal receives BPSK symbols from a plurality of K source terminals, the sum symbol received from the relay will follow a binomial probability distribution B(K,2). The same conclusion is valid for QPSK symbols, the probability distribution then being binomial on each axis in phase and in quadrature.

In the two cases mentioned above, original non-uniformity of the information source, non-uniformity induced by the network, the receiver needs to decode information symbols belonging to an alphabet with a non-uniform probability distribution.

As mentioned above, the optimal decoding method is MAP decoding but it is difficult to achieve in practice due to its high complexity. In particular, in the context of network coding, an exhaustive search has to be made on all possible combinations of symbols from the different source terminals. Furthermore, ML decoding is sub-optimal in terms of the error rate. Furthermore, existing ML decoding methods in a telecommunication system using network coding are also very complex.

The article by Nevat et al. entitled «Detection of Gaussian constellations in MIMO systems under imperfect CSI» published in IEEE Trans. on Com., vol. 58, No. 4, April 2010, pp. 1151-1160 discloses a MAP detector in a MIMO system, the received symbols having a quasi-Gaussian amplitude distribution following formatting in transmission.

Consequently, the purpose of this invention is to disclose a method of decoding information symbols belonging to an alphabet with a non-uniform probability distribution, that is significantly simpler than MAP decoding by an exhaustive search and that has a lower error rate than that obtained by conventional decoding by maximum likelihood (ML).

PRESENTATION OF THE INVENTION

This invention is defined by a MAP decoding method of a signal received through a noisy channel, said signal being composed of symbols belonging to a predetermined alphabet, affected by additive white Gaussian noise, the probability distribution of the symbols within this alphabet being non-uniform, the symbols being represented by points in a lattice Λ={x|x=Ma, a∈Z^N} generated by a matrix M with dimension N×N, in which:

• • a vector y with dimension N representative of the received signal is determined;
  • an augmented vector

$y_{\exp }=\left[\begin{array}{c}y\\ 0_N\end{array}\right]$ is created in which 0_N is a null vector with dimension N;

• • an augmented lattice Λ_exp={x|x=M_expa,aϵZ^N} with generator matrix

$M_{\exp }=\left[\begin{array}{c}M\\ \beta M\end{array}\right]$ is formed in which

$\beta =\frac{{\sigma }^2}{{\sigma }_V^2}$ is the ratio between variance of the noise and variance of a Gaussian probability distribution modelling the non-uniform probability distribution of the symbols in the alphabet;

• • a search is made among the points in the augmented lattice to find the point closest to the vector representative of the received signal, â;
  • the received symbol is estimated from the generating matrix M and the closest neighbour, {circumflex over (v)}_MAP*=Mâ.

Advantageously, the search for the closest neighbour is limited to a finite sub-set Ω_a of the lattice representing a constraint on the signal energy.

According to a first embodiment of the invention, the noisy channel is a channel composed of a plurality K of elementary channels between the source terminals and a relay terminal of a network and each of the symbols is composed of a combination of elementary symbols transmitted by the corresponding source terminals to said relay terminal, each said elementary symbol belonging to an elementary alphabet of a source terminal.

In one exemplary application, the elementary alphabets of the source terminals are identical and the variance σ_V² of the Gaussian probability distribution modelling the non-uniform distribution of symbols in the alphabet is obtained by σ_V²=Kσ_X², where σ_X² is the variance of the probability distribution of elementary symbols within an elementary alphabet.

In this case, the variance σ_X² is obtained by σ_X²=NP_max where P_max is the maximum transmission power of a source terminal.

Advantageously, the search for the nearest neighbour is made using a sphere decoder and sphere decoding uses an enumeration of points according to Pohst's algorithm.

Alternately, the search for the closest neighbour could be made by decoding using a stack decoder with a spherical bound.

BRIEF DESCRIPTION OF THE DRAWINGS

Other characteristics and advantages of the invention will become clear after reading preferred embodiments of the invention, with reference to the appended figures among which:

FIG. 1 shows a simplified example of a cooperative network using physical layer network coding;

FIG. 2 shows an error rate curve as a function of the signal to noise ratio for an exemplary application of the MAP augmented lattice decoding method according to one embodiment of this invention;

FIG. 3 shows a flowchart of the MAP augmented lattice decoding method according to one embodiment of this invention.

DETAILED PRESENTATION OF PARTICULAR EMBODIMENTS

In the following, we will consider an alphabet of information symbols with a non-uniform probability distribution. In other words, the symbols in this alphabet are not equally probable.

This alphabet may for example be the alphabet used by a non-uniform information source or it may be the result of a combination at a relay in a cooperative network making use of physical layer network coding.

For illustration purposes and without reducing generalisation, we will disclose the decoding method according to the invention as part of a cooperative network. To achieve this, we will assume that K source terminals S₁, . . . , S_K transmit K symbols x₁, . . . , x_K respectively, and that a relay terminal receives a signal corresponding to a combination of these symbols, namely:

$\begin{array}{cc}\frac{1}{N}E\left\{{x}^2\right\}\le P_{\mathrm{ma}x}& \left(8\right)\end{array}$ where g_k, k=1, . . . , K are the gains of each of the channels between the source terminals and the relay terminal. Without reducing generality, we will assume that these gains are equal to 1.

Symbols x_k are elements of a finite alphabet A_X=Λ∩Ω, called the elementary alphabet, in which Λ is a lattice of dimension N, and Ω is a convex of ^N, containing the origin and defining the transmission power constraint of a source terminal (assumed to be identical for all terminals).

For example, if the symbols x_k are elements in a QAM alphabet, the lattice dimension will be 2. Remember that a lattice with dimension N is generally defined by a matrix M with dimension N×N, in which the column vectors of the matrix are called the lattice generator vectors, a point x in the lattice being defined in a vector form by x=Ma where aϵZ^N is a vector with dimension N in which the elements are integers. However, it is understood that the origin of the lattice is not considered for a QAM alphabet (a≠0_N). Nevertheless it will be understood that we can always consider gains g_k equal to 1, provided that the matrix M can be modified.

The convex Ω is typically a sphere of ^N centred at the origin, the radius of the sphere being given by the maximum transmission power of the source terminal. Thus, if P_max is the maximum transmission power of a source terminal, the convex Ω can be defined by points x, such that:

$\begin{array}{cc}y=\sum _{k=1}^K{g}_k{x}_k+z& \left(7\right)\end{array}$ where E{.} is the mean value.

The vector z in the expression (7) is a noise vector with dimension N the components of which are random Gaussian variables with zero mean and variance σ².

The relay terminal is provided with a decoder to decode the sum

$v=\sum _{k=1}^K{x}_k.$ After decoding if applicable, the relay terminal transmits it to the destination terminal.

The sum symbol

$v=\sum _{k=1}^K{x}_k$ belongs to an alphabet A_V resulting from superposition of elementary alphabets A_X. Given that the lattice Λ is stable by addition (Λ with the addition has a group structure), the sum symbol v still belongs to the lattice Λ. The result is that the lattice Λ_V defining the alphabet A_V is a sub-set of Λ(Λ_V⊂Λ). The alphabet A_V of sum symbols is such that A_V=Λ_V∩Ω_V where Ω_V is a convex of ^N: reflecting the transmission power constraint of K source terminals.

MAP decoding at the relay would result in a search for the sum vector {circumflex over (v)}_MAP such that:

$\begin{array}{cc}{\hat{v}}_{\mathrm{MAP}}=\underset{v\in {\Lambda }_V}{\mathrm{argmax}}\left(P\left(v|y\right)\right)=\underset{v\in {\Lambda }_V}{\mathrm{argmax}}\left(P\left(v\right)P\left(y|v\right)\right)& \left(9\right)\end{array}$ also such that:

$\begin{array}{cc}{\hat{v}}_{\mathrm{MAP}}=\underset{v\in {\Lambda }_V}{\mathrm{argmin}}\left(-\ln \left(P\left(v\right)\right)+\frac{{y-v}^2}{2{\sigma }^2}\right)& \left(10\right)\end{array}$ taking account of the fact that

$P\left(y|v\right)=\exp \left(-\frac{{y-v}^2}{2{\sigma }^2}\right).$

The basic idea of the invention is to model the sum symbol v using a random Gaussian variable with zero mean and variance σ_V²=Kσ_X²=KNP_max. This approximation is justified by application of the central limit theorem to the random variables x_k.

Based on this approximation, the probability distribution law of the sum symbol is given as follows:

$\begin{array}{cc}P\left(v\right)=\frac{1}{{\left({\sigma }_V\sqrt{2\pi }\right)}^N}\exp \left(-\frac{{v}^2}{2{\sigma }_V^2}\right)& \left(11\right)\end{array}$

In substituting this estimation of P(v) in the expression (10), we obtain a new decoding criterion:

$\begin{array}{cc}{\hat{v}}_{{\mathrm{MAP}}^{*}}=\underset{v\in {\Lambda }_V}{\mathrm{argmin}}\left(N\ln \left({\sigma }_V\sqrt{2\pi }\right)+\frac{1}{2{\sigma }_V^2}{v}^2+\frac{1}{2{\sigma }^2}{y-v}^2\right)& \left(12\right)\end{array}$

Considering that the first term does not depend on v, decoding consists of determining:

$\begin{array}{cc}{\hat{v}}_{{\mathrm{MAP}}^{*}}=\underset{v\in {\Lambda }_V}{\mathrm{argmin}}\left({y-v}^2+\frac{{\sigma }^2}{{\sigma }_V^2}{v}^2\right)& \left(13\right)\end{array}$

A vector, y_exp called an augmented vector, with dimension 2N, resulting from the vertical concatenation of vector y and a null vector with dimension N, is added, namely

$y_{\mathrm{ex}p}=\left[\begin{array}{c}y\\ 0_N\end{array}\right]$ and a full rank matrix

$B=\left[\begin{array}{c}{I}_N\\ \beta I_N\end{array}\right]$ where I_N is the unit matrix with dimension N×N and β is a coefficient reflecting the ratio between the noise variance and the variance of the sum symbol to be decoded, namely

$\beta =\frac{{\sigma }^2}{{\sigma }_V^2}.$

The decoding criterion (13) can then be expressed as follows:

$\begin{array}{cc}{\hat{v}}_{{\mathrm{MAP}}^{*}}=\underset{\underset{a\in {\Omega }_a}{v=\mathrm{Ma}}}{\mathrm{argmin}}\left({y_{e\mathrm{xp}}-\mathrm{BMa}}^2\right)& \left(14\right)\end{array}$ in which the condition on the transmission power limit is shown by constraining the vector a to belong to a subset of Z^N such that the transmission power constraint is satisfied, in other words

${\Omega }_a=\left\{a\in Z^N|\frac{1}{N}E\left\{{\mathrm{Ma}}^2\le P_{\mathrm{ma}x}\right\}\right\}.$

If we define the augmented matrix M_exp with dimension 2N×N by

$M_{\exp }=\mathrm{BM}=\left[\begin{array}{c}M\\ \beta M\end{array}\right],$ the decoding criterion is finally expressed in the following form:

$\begin{array}{cc}{\hat{v}}_{{\mathrm{MAP}}^{*}}=\underset{\underset{a\in {\Omega }_a}{v=\mathrm{Ma}}}{\mathrm{argmin}}\left({y_{\exp }-M_{\exp }a}^2\right)& \left(15\right)\end{array}$ in other words, this is equivalent to searching for the point in an augmented lattice Λ_exp, generated by the matrix M_exp, that is closest to the point y_exp, representing the received signal. Consequently, it will be understood that it is equivalent to a conventional ML decoding in an augmented space, with dimension 2N, the lattice Λ_exp and the point representative of the received signal y_exp both being dimension 2N.

The search for the closest neighbour of y_exp in the augmented lattice Λ_exp may be made classically using sphere decoding. It should be remembered that sphere decoding can limit the search for the closest neighbour to lattices belonging to a noise ball centred at the point representing the received signal. A description of sphere decoding is given in the article by E. Viterbo et al. entitled «A universal lattice code decoder for fading channels» published in IEEE Transactions on Information Theory, vol. 45, pages 1639-1642, July 1999. A spherical bound stack decoder (SB) can also be used as described in article by G. Rekaya Ben-Othman entitled «The spherical bound stack decoder» published in Proc. of IEEE International Conference on Wireless & Mobile Computing, Networking & Communication, pp. 322-327.

The search for the closest neighbour using the sphere decoder could in particular use a Pohst enumeration, known in itself.

The sphere decoder starts from a noise sphere with radius C centred at the point y_exp representative of the received signal. Points a in the lattice Λ_exp belonging to this sphere satisfy the following relation:∥y_exp−M_expa∥²≤C²  (16)

The search for the closest neighbour must also take account of the fact that the point a must satisfy the power constraint, in other words aϵΩ_a.

In a preliminary step, the augmented matrix M_exp is subject to a QR decomposition, in other words M_exp=QR where Q is a matrix with dimension 2N×N in which the column vectors are orthogonal and R is an upper triangular matrix with dimension N×N. If we denote ρ=M_exp⁻¹y_exp (Zero Forcing solution) and ξ=ρ−a, relation (16) can be written again as follows:

$\begin{array}{cc}R\xi \le C^2\mathrm{or}\text{:}& \left(17\right)\\ \sum _{i=1}^N{p_{\mathrm{ii}}\left({\xi }_i+\sum _{j=i+1}^N{p}_{\mathrm{ij}}{\xi }_j\right)}^2\le C^2& \left(18\right)\end{array}$ where p_ii=r_ii², i=1, . . . , N and

$p_{\mathrm{ij}}=\frac{r_{\mathrm{ij}}}{r_{\mathrm{ii}}},$ j=i+1, . . . , N.

The search begins with the N^th component of a:

$\begin{array}{cc}-\frac{C}{\sqrt{p_{\mathrm{NN}}}}\le {\xi }_n\le \frac{C}{\sqrt{p_{\mathrm{NN}}}}& \left(19\right)\end{array}$ namely, considering the fact that a_nϵZ:

$\begin{array}{cc}\lceil {\rho }_N-\frac{C}{\sqrt{p_{\mathrm{NN}}}}\rceil \le a_n\le \lfloor {\rho }_N+\frac{C}{\sqrt{p_{\mathrm{NN}}}}\rfloor & \left(20\right)\end{array}$

Similarly, the search for other components is limited to intervals:

$\begin{array}{cc}\lceil -\sqrt{\frac{T_i}{p_{\mathrm{ii}}}}+S_i\rceil \le a_i\le \lfloor \sqrt{\frac{T_i}{p_{\mathrm{ii}}}}+S_i\rfloor & \left(21\right)\end{array}$ where

$S_i={\rho }_i+\sum _{j=1}^N{p}_{\mathrm{ij}}{\xi }_j$ and T_i=T_i-1−p_ii(S_ii−a_i) in which T_N=C². The search is made from the last component to the first, the choice of a component with a given index of a reducing the search interval for the component with a lower index.

Search intervals (20) and (21) can be expressed using the simplified relation:

$\begin{array}{cc}{b}_i^{-}\le a_i\le b_i^{+},i=1,\dots ,N{b}_i^{-}=\lceil -\sqrt{\frac{T_i}{p_{\mathrm{ii}}}}+S_i\rceil \mathrm{and}{b}_i^{+}=\lfloor \sqrt{\frac{T_i}{p_{\mathrm{ii}}}}+S_i\rfloor & \left(22\right)\end{array}$

Search intervals also have to be restricted to take account of the maximum transmission power constraint, aϵΩ_a. This constraint imposes that each component a_i in a should remain between two bounds:ω_i⁻≤a_i≤ω_i⁺, i=1, . . . ,N  (23)

Therefore, finally, the search interval I_i for each component a_i is reduced to:sup(ω_i⁻,b_i⁻)≤a_i≤inf(ω_i⁺,b_i⁺)  (24)

Decoding then continues as follows; the first step is to choose a component a_N within the interval I_N=[sup(ω_N⁻,b_N⁻),inf (ω_N⁺,b_N⁺)], and the next step is to search for a candidate for component a_N-1 within the interval I_N-1=[sup(ω_N-1⁻,b_N-1⁻),inf(ω_N-1⁺,b_N-1⁺)] for which the bounds were calculated from a_N. If there is no value a_N-1 within this interval, then we return to level N to select another candidate for component a_N. The process continues from step to step until level 1 is reached. Once a vector â has been found for which the components satisfy all conditions (24), the radius of the sphere is updated and the search process is iterated by scanning all intervals I_i, until the vector â closest to y_exp is found. The symbol {circumflex over (v)}_MAP* is given by {circumflex over (v)}_MAP*=Mâ.

FIG. 2 shows the symbol decoding error rate at the relay, as a function of the signal to noise ratio.

In the case shown, it has been assumed that N=4, K=2 and a maximum transmission power P_max=1.

210 relates to the curve corresponding to the classical maximum likelihood (ML) decoding method and 220 relates to the curve corresponding to the MAP augmented lattice decoding method according to this invention. This figure confirms that the augmented lattice decoding method enables to achieve a lower symbol error rate for a given signal to noise ratio or to achieve a better signal to noise ratio for a given target symbol error rate. It should have been noted that the gain is higher when the dimension N of the lattice is high.

Those skilled in the art will understand that other search algorithms for the closest neighbour within a lattice and particularly Schnorr-Euchner's decoding could be used instead of the sphere decoder method, without going outside the scope of this invention.

We will disclose the MAP augmented lattice decoding method below, in a general framework.

The decoding method aims at decoding a signal received through a noisy channel, said signal being composed of information symbols belonging to a predetermined alphabet. The probability distribution of information symbols within this alphabet is non-uniform, in other words some information symbols are more frequent than others. Noise affecting the channel is assumed to be white and Gaussian.

It is assumed that symbols in the alphabet can be represented by points in a lattice Λ with dimension N defined by the relation {x|x=Ma, aϵZ^N} in which M is a lattice generator matrix. This condition is satisfied particularly when symbols belong to a QAM type modulation constellation. The alphabet is a finite sub-part of Λ, defined as being the intersection between said lattice and a convex Ω of ^N. However in the case of a QAM alphabet, it should be noted that the origin of the lattice is excluded. The different points in the lattice are weighted by different probability distributions due to the non-uniformity of the probability distribution of symbols in the alphabet.

According to the invention, the probability distribution of symbols in the alphabet is modeled by a Gaussian distribution with variance σ_V².

The signal to be decoded is expressed in the form y=x+z, where x is a point in Λ∩Ω and z is a noise vector with dimension N of which the components are random variables with variance σ².

FIG. 3 diagrammatically shows the flowchart for the augmented lattice decoding method according to a general embodiment of the invention.

In step 310, a vector y is formed, representative of the received signal in a space with dimension N where N is the dimension of the lattice Λ. This vector may for example be obtained by demodulation of the received signal.

In step 320, the augmented vector

$y_{\exp }=\left[\begin{array}{c}y\\ 0_N\end{array}\right]$ is constructed.

In step 330, an augmented lattice Λ_exp with generator matrix

$M_{\exp }=\left[\begin{array}{c}M\\ \beta M\end{array}\right]$ is formed, in which M is the generator matrix of the lattice Λ and

$\beta =\frac{{\sigma }^2}{{\sigma }_v^2}$ is the ratio between the variance of the noise and the variance of the Gaussian probability distribution modelling the non-uniform probability distribution of symbols in the alphabet.

In step 340, a search is made for the closest neighbour to the augmented vector y_exp in the augmented lattice Λ_exp, in other words the vector â of Ω_a minimising the distance ∥y_expM_expa∥², where Ω_a is a part of Z^N reflecting a maximum energy constraint of the received signal.

In step 350, the received symbol in the sense of the MAP criterion is estimated by {circumflex over (v)}_MAP*=Mâ.

It will be understood that the search for the closest neighbour in the lattice can be made using a sphere decoder or a stack decoder with spherical bound, as mentioned above. The different known sphere decoding variants are applicable in this case, particularly those implementing enumeration of points making use of Pohst's algorithm.

Those skilled in the art will understand that decoding of sum symbols for a channel of a cooperative network is a special case of decoding in FIG. 3 since these sum symbols form part of an alphabet obtained by superposition of the elementary alphabets of the different source terminals and that the distribution of sum symbols within this alphabet is non-uniform.

Finally as mentioned above, other closest neighbour search algorithms in a lattice, and particularly Schnorr-Euchner's decoding, could be used without going outside the scope of this invention.

Claims

1. A MAP decoding method implemented on a decoder for decoding a signal received through a noisy channel, wherein the method comprises: receiving, by a relay terminal of a network having a decoder, the signal transmitted by at least one source terminal through a noisy channel, said noisy channel comprising an elementary channel between each source terminal and the relay terminal, the signal being composed of symbols belonging to a predetermined alphabet, affected by an additional Gaussian white noise, a probability distribution of the symbols within the predetermined alphabet being non-uniform, the symbols being represented by points in a lattice Λ={x|x=Ma, aϵZN} generated by a matrix M with dimension N×N;determining, by the decoder, a vector y with dimension N representative of the received signal;determining, by the decoder, an augmented vector

2. The MAP decoding method according to claim 1, wherein said searching for the closest neighbour â is limited to a finite sub-set Ωa of the lattice Λ representing a constraint on a signal energy.

3. The MAP decoding method according to claim 1, wherein said noisy channel comprises a plurality of K elementary channels between the source terminals and the relay terminal, each symbol comprising a combination of elementary symbols transmitted by the corresponding source terminals to said relay terminal, each elementary symbol belonging to an elementary alphabet of a source terminal.

4. The MAP decoding method according to claim 3, wherein the elementary alphabets of the source terminals are identical and the variance σV2 of the Gaussian probability distribution modelling the non-uniform probability distribution of symbols in the alphabet is obtained by σV2=KσX2 where σX2 is the variance of the probability distribution of elementary symbols within an elementary alphabet.

5. The MAP decoding method according to claim 4, wherein the variance σX2 is obtained by σX2=NPmax where Pmax is the maximum transmission power of a source terminal.

6. The MAP decoding method according to claim 1, wherein the search for the closest neighbour â is made using a sphere decoder.

7. The MAP decoding method according to claim 6, wherein the sphere decoding uses an enumeration of points according to Pohst's algorithm.

8. The MAP decoding method according to claim 1, wherein the searching for the closest neighbour â comprises decoding, using a stack decoder with a spherical bound.