DETECTION METHOD FOR MIMO SYSTEM RECEIVERS USING POINT NETWORK REDUCTION AND ITERATIVE NOISE REDUCTION
Patent Information
- Application Number
- DE602020040639
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-02-12
- Filing Date
- 2020-01-10
- Publication Date
- 2024-11-06
- Estimated Expiration
- 2040-01-10
AI Technical Summary
Current MIMO systems face challenges in achieving optimal detection performance due to the complexity of optimal detection methods and the sub-optimal performance of linear detection methods, especially in high modulation orders and large antenna configurations, which leads to high error rates and increased computational complexity.
A detection method for MIMO systems that employs LLL reduction of the channel matrix to obtain a reduced channel matrix, followed by linear detection using a matrix derived from the reduced channel matrix, with an iterative noise removal process to improve error rates and provide flexible detection results.
The method achieves improved error rate performance as a function of signal-to-noise ratio without significant computational overhead and provides flexible detection results, including soft outputs suitable for channel decoding.
Description
DOMAINE TECHNIQUE
[0001] The present invention relates to the field of multi-antenna telecommunication systems also called MIMO ( Multiple Input Multiple Output ) and more particularly those using detection aided by network reduction ( Lattice Reduction aided MIMO detection ). ÉTAT DE LA TECHNIQUE ANTÉRIEURE
[0002] Multi-antenna or MIMO telecommunications systems ( Multiple Input Multiple Output ) are well known in the state of the art and have been introduced, about ten years ago, in several wireless telecommunication system standards (for example IEEE 802.11n). These systems have, due to their multiplexing capacity and spatial diversity, channel capacities significantly higher than conventional single-antenna SISO systems ( Single Input Single Output ).
[0003] A MIMO transmission system at N T transmitting antennas and N R receiving antennas is characterized by its channel matrix H , of size N R × N T whose elements are the complex attenuation coefficients of the elementary channels between the N T transmitting antennas and N R receiving antennas.
[0004] We have represented schematically in Fig. 1 such a MIMO system. One element h ik of the matrix H gives the elementary channel coefficient between the transmitting antenna of index k and the index receiving antenna i.
[0005] The transmission of a block of symbols in such a system can be modeled by: y = Hs + n where we note s the vector of symbols respectively transmitted by the N T transmitting antennas, y is a vector of size N R representing the signals respectively received by the N R receiving antennas, and n is a white noise vector of size N R whose components are independent Gaussian centered random variables of the same variance, σ n 2 . The channel matrix H is assumed to be known to the receiver, for example by means of a channel estimation from pilot symbols.
[0006] The emitted symbols belong to a modulation alphabet, noted A, for example an M-QAM alphabet.
[0007] It will be assumed in the following that the number of receiving antennas is greater than or equal to the number of transmitting antennas, i.e. N R ≥ N T .
[0008] Since the elements of the vectors and the matrix involved in relation (1) are complex values, their real and imaginary parts are treated separately. In other words, if y,s And n are the real vectors of size respectively n R , n T And n R with n R = 2 N R And n T = 2 N T defined by: y ‾ = ℜ y ℑ y ; s ‾ = ℜ s ℑ s ; n ‾ = ℜ n ℑ n and the channel matrix H is a real matrix of size n R × n T defined by: H ‾ = ℜ H − ℑ H ℑ H ℜ H relation (1) is still valid in real form: y ‾ = H ‾ s ‾ + n ‾
[0009] Different types of detection can be provided at reception.
[0010] Optimal detection uses a maximum likelihood or ML criterion ( Maximum Likelihood ), consisting of searching on the set of possible emitted symbol vectors the one minimizing the Euclidean distance to the received vector, in other words: s ^ ML = arg min s ∈ A N T ‖ y − Hs ‖
[0011] However, this research presents a complexity in M nT< and is therefore not feasible in practice for high modulation orders and a large number of transmitting antennas.
[0012] In practice, suboptimal detection methods, called linear, are used, which allow the different symbols to be separated by means of linear combinations. More precisely, returning to the complex formulation of relation (1), the vector of emitted symbols is estimated by multiplying the vector of received signals by a detection matrix (also called an equalization matrix). W of size n T × n R : r = Wy = WHs + Wn where the detection matrix can take different forms depending on the type of detection chosen: ZF ( Zero Forcing ), MMSE ( Minimum Square Error ), or MRC ( Maximum Ratio Combining ) also called MF ( Matched Filter ) , either : W ZF = H † = H H H − 1 H H W MMSE = H H H + σ n 2 I N R − 1 H H W MF = H H
[0013] The respective advantages and disadvantages of these different types of detection are known: thus a ZF type detection can considerably amplify the noise components and an MMSE type detection, if it reduces this amplification, requires the estimation of the signal to noise ratio and is therefore more complex.
[0014] The emitted symbols are then estimated by making a decision ( dec ) hard or soft on the components of the equalized vector, r : s ^ ZF = dec W ZF y s ^ MMSE = dec W MMSE y s ^ MF = dec W MF y
[0015] For example, in the case of a hard decision, for each component of the equalized vector, we search for the element of the alphabet, that is, the point of the modulation constellation closest to this component.
[0016] If a linear or LD detector ( Linear Detector ) is relatively simple, its performance depends on the conditioning of the matrix Hand more precisely its deviation from orthogonality. Indeed, if the eigenvalues of the matrix H H< H the noise components are low ( Wn in expression (3)) are amplified and the error rate (BER) is high.
[0017] To overcome this difficulty, it has been proposed to perform linear detection aided by point lattice reduction as described in the article by D. Wübben entitled "Near-maximum likelihood detection of MIMO systems using MMSE-based lattice reduction" published in Proc. IEEE International Conference on Communications, vol. 2, June 2004, pp. 798-802.
[0018] We first recall that a network of points Λ in ℝ n is defined as the set of linear combinations with integer coefficients of a plurality of vectors v 1,..., v n forming a base V of ℝ n , in other words: Λ = ∑ i = 1 n a i v i | a i ∈ ℤ n
[0019] The base V generator of the network is represented by the matrix V = ( v 1 | v 2 |...| v n ) in the canonical basis of ℝ n , in other words the matrix V is the matrix whose column vectors are the basis vectors V .
[0020] The same network of points, Λ , can be generated by different bases. Thus, if V is a generating basis of the network, represented by the matrix V and if T is a unimodular matrix (in other words a matrix with integer coefficients and whose absolute value of the determinant is equal to 1), the columns of the matrix V' defined by: V ′ = VT generate the same network of points, Λ .
[0021] The reduction of a network of points Λ consists of finding a generating basis of this network formed by vectors of shortest length and closest to orthogonality.
[0022] The deviation from the orthogonality of the base V is given by the indicator: od V = ∏ i = 1 n ‖ v i ‖ vol Λ Or vol (Λ) = |det( V )|called the volume of the network, depends only on A where det( V ) is the determinant of the matrix V . Note that the orthogonality indicator is always greater than or equal to 1, equality corresponds to the case of an orthogonal basis.
[0023] The reduction of the point lattice can be approximated by means of the LLL (Lenstra-Lenstra-Lovâsz) algorithm. A reduced LLL lattice basis, Ṽ of parameter δ of the network is obtained from a decomposition QR of the matrix V . This basis is given by the column vectors of the matrix Ṽ = Q̃R̃ checking: r ˜ l , k ≤ 1 2 r ˜ l , l pour 1 ≤ l ≤ k ≤ n δ r ˜ k − 1 , k − 1 2 ≤ r ˜ k , k 2 + r ˜ k − 1 , k 2 pour 1 < k ≤ n
[0024] The parameter δ gives the quality of the base Ṽ , the approximation is all the better as the value δ is close to 1.
[0025] The points of a modulation constellation M -QAM, can be considered, up to a common translation and multiplicative factor, as elements of ℤ 2 . So the real vector s ‾ ∈ ℤ n T and the real vector Hs is an element of the point network Λ H generated by the column vectors of H in space ℝ n T . The vector y is therefore a point of ℝ n T resulting from the sum of a network vector Λ H and a noise vector.
[0026] The reduction of the network and in particular its LLL reduction makes it possible to obtain a basis for this network composed of short vectors close to the orthogonality condition. If H̃ is the matrix of the channel represented in this new basis, we have: H ˜ = H ‾ T Or T is a unimodular matrix of size n T × n T . The set of column vectors of T is called reduced basis.
[0027] Relation (4) can then be rewritten in the form: y ‾ = H ‾ T T − 1 s ‾ + n ‾ = H ˜ z ‾ + n ‾ Or z = T -1< s represents the received signal in the reduced basis and the matrix H̃ is better conditioned than the matrix H . It is important to understand that the point H̃z is identical to the point Hs of the Λ H network.
[0028] If we perform a linear detection at reception, we have in the real domain: r ‾ = W ‾ y ‾ = W ‾ H ˜ z ‾ + W ‾ n ‾ with r ‾ = ℜ r J r And W ‾ = ℜ W ˜ − J W ˜ J W ˜ ℜ W ˜ where the linear detection matrix is obtained this time from H̃ : W ˜ ZF = H ˜ † = H ˜ H H ˜ − 1 H ˜ H W ˜ MMSE = H ˜ H H ˜ + σ n 2 I N R − 1 H ˜ H W ˜ MF = H ˜ H
[0029] For reasons of notational simplification, we will abandon in the following the above-mentioned lines for real vectors and matrices. Thus, relation (15) is written in simplified form: r = W ˜ y = W ˜ H ˜ z + W ˜ n
[0030] Those skilled in the art will, however, understand that the vectors and matrices appearing in equation (16) are real-valued and that their dimensions are twice those appearing in equation (6).
[0031] A first method of linear detection aided by point network reduction or LR-LD ( Lattice Reduction aided Linear Detection ) was described in the aforementioned article. It consists of quantifying each component z i , i = 1, ..., n T of z on the i th vector of the reduced basis. The quantization of the component z i is performed by searching for the point belonging to the projection of the product constellation A NT< on the i-th vector. In other words, in each of the directions carried by a vector of the reduced basis, we seek the point of the projection of the constellation A NT< on this direction which is closest to the component of z on this vector.
[0032] A second linear detection method aided by network reduction, also described in the aforementioned article, hereinafter referred to as LR-LD-SIC ( Lattice Reduced Linear Detection with Successive Interference Concettation ) consists of performing a QR decomposition of the channel matrix H̃ (Or Q , R are respectively an orthogonal matrix and an upper triangular matrix) then to detect the components of z , starting with the component z nT not interfered by other components (due to the upper triangular structure of R ). The columns of H̃ can be swapped (which is equivalent to swapping the transmitting antennas) so that the signal-to-noise ratio of the component z nT is as high as possible. In other words, we carry out for each permutation of the columns of H̃ a QR decomposition and we select the permutation leading to the coefficient r nTnT the highest.
[0033] The LR-LD method has the disadvantage of performing a suboptimal truncation since depending on the order in which the components of z are truncated, the performances in terms of BER can be very different, for the same value of signal to noise ratio (SNR).
[0034] The LR-LD-SIC method is efficient when the linear detection is of the MMSE type but it requires a significant amount of calculations, in particular for MIMO configurations with a large number of antennas, which is hardly compatible with the computing capacity of a mobile terminal.
[0035] Furthermore, neither the LR-LD nor the LR-LD-SD method is capable of providing soft outputs, such as those required for soft-input channel decoding when a turbo code, a convolutional code, or an LDPC code is used at transmission.
[0036] An object of the present invention is therefore to provide a detection method for a MIMO system receiver aided by dot array reduction, which has good performance in terms of error rate as a function of signal-to-noise ratio, without requiring significant calculations. A subsidiary object of the present invention is to provide such a MIMO detection method aided by dot array reduction which is further capable of providing a detection result in the form of soft values. EXPOSÉ DE L'INVENTION
[0037] The present invention is defined by a detection method for a MIMO system receiver at N T transmitting antennas and N R receiving antennas, characterized by a channel matrix H , each transmitting antenna transmitting a symbol belonging to a modulation constellation, the vector of the transmitted symbols belonging to a constellation produced in space ℝ 2 N T , the channel matrix H undergoing LLL reduction to provide a reduced channel matrix, H̃ , such as H̃ = HT Or T is a unimodular matrix, the vector of signals received by the receiving antennas being equalized by means of a linear detection matrix W̃ , obtained from the reduced channel matrix, H̃ , to provide an equalized vector r, in a reduced basis formed from the column vectors of the matrix T , the equalized vector being represented by c = Tr in a reduced basis formed from the column vectors of the matrix T , in which we define, in the representation according to the reduced basis, a non-noisy reception zone Z ε surrounding the product constellation with a predetermined tolerance margin ε , the equalized vector then being subjected to an iterative noise suppression process in which, at each iteration, the component (ℓ) of the equalized vector is sought in the reduced base, furthest from the noise-free reception zone, and a noise vector is subtracted from the equalized vector c whose modulus is a fraction of the tolerance margin and the direction is that of the component thus determined, the process being iterated until the representative point of the equalized vector is located in said noise-free reception zone or a predetermined number of iterations is reached, the result of the detection being obtained by making a hard or soft decision on the equalized vector at the end of the iterative process.
[0038] The linear detection matrix is notably chosen from ZF, MMSE or MF detection matrices defined respectively by: W ˜ ZF = H ˜ † = H ˜ H H ˜ − 1 H ˜ H W ˜ MMSE = H ˜ H H ˜ + σ n 2 I N R − 1 H ˜ H W ˜ MF = H ˜ H Or σ n 2 is the variance of the noise affecting the signal received by a receiving antenna and I NR is the size identity matrix N R .
[0039] The modulation constellations relating to the different transmitting antennas are typically QAM constellations. The modulation constellations relating to the different antennas are advantageously chosen to be identical.
[0040] Advantageously, to determine the component of the equalized vector in the reduced base, furthest from the noise-free zone, we seek the index ℓ such that l = arg max i c i − s max − ε / ν i Or c i is the i th component of the equalized vector c , s max is the maximum amplitude of the QAM constellation, and ν i = max j ≠ i 〈 t i | t j 〉 where is the i th < column vector of T .
[0041] The noise vector subtracted from the equalized vector c can be determined by a l sgn c l . μ . ε ν l Or aℓ is the ℓ-th column of the matrix A , Or A is the Gram matrix of the reduced basis, A = TT T< , sgn(c ℓ ) is the sign of the ℓ-th component of the equalized vector c And µ is the value of said fraction. In particular, we can choose μ = 1 2 .
[0042] According to a first variant, a hard decision is made on the equalized vector at the end of the iterative process, by searching for the point of the product constellation closest to this vector.
[0043] According to a second variant, prior to the iterative process the variances are initialized, σ n , m 2 , of the noise affecting each of the components of the equalized vector c by σ n , m 2 = σ n 2 ∑ i = 1 N T W ˜ m , i 2 + σ a 2 ∑ j = 1 j ≠ m N T β ˜ m , j 2 / β ˜ m , m 2 Or σ n 2 is the variance of the noise affecting the signal received by a receiving antenna, σ a 2 is the average power of the modulation constellation, W̃ m , iis the element of the matrix W̃ to the line m and the column i, β̃ m , j is the element of the matrix β = W̃H to the line m and the column j and that at each iteration we then update these variances by σ n , m 2 = σ n , m 2 + μ 2 ε ν l 2 A m , l 2 Or A m ,ℓ is the element at line m and the column ℓ of the Gram matrix, A .
[0044] In the latter case, we can calculate for each symbol emitted by a transmitting antenna m the soft values of the different bits b m,q , b m,q being the q ème< bit of the word used to generate the symbol emitted s m , the soft values being given by L b m , q = 1 2 σ n , m 2 min u ∈ S m , q 1 ‖ c − u ‖ 2 − min u ∈ S m , q 0 ‖ c − u ‖ 2 Or σ n , m 2 is the value of the variance affecting the m -th component of the equalized vector c at the end of the iterative process, S m , q 0 is the set of points of the product constellation admittingb m , q =0 and S m , q 1 is the set of points of the product constellation admitting b m,q =1. BRÈVE DESCRIPTION DES DESSINS
[0045] Other characteristics and advantages of the invention will appear on reading a preferred embodiment of the invention, described with reference to the attached figures among which: There Fig. 1 schematically represents a MIMO system known from the state of the art; The Fig. 2 schematically represents a flowchart of a linear detection method aided by a point network and successive noise reduction according to a first embodiment of the invention; The Fig. 3 schematically illustrates an example of noise reduction in one step of the method of the Fig. 2 ; There Fig. 4 schematically represents a flowchart of a linear detection method aided by point network and successive noise reduction according to a second embodiment of the invention; The Fig. 5 represents the throughput of the MIMO system as a function of the signal-to-noise ratio for a detection method according to the second embodiment of the invention and according to a detection method according to the state of the art. EXPOSÉ DÉTAILLÉ DE MODES DE RÉALISATION PARTICULIERS
[0046] In the following, we will consider a MIMO system as described in relation to the Fig. 1 . The information symbols emitted by the N T transmitting antennas belong to QAM constellations. Without loss of generality, it will be assumed that the modulation alphabet (or constellation) used for the different transmitting antennas is identical, in other words a block of transmitted symbols belongs to a product constellation A NT< . We will also assume that the transmission power of each antenna is normalized to 1. In the absence of noise, the received signal can be represented by a point in the network of points Λ H generated by the column vectors of H in space ℝ n T with n T = 2 N T .In the presence of noise, the received signal can be considered as a point of ℝ n T obtained by means of a translation by a noise vector of a point of the network Λ H .
[0047] We will further assume, as before, that the elements of the matrix H are known to the receiver, for example by means of a channel estimate made from pilot symbols.
[0048] The idea behind the invention is to carry out noise suppression by iterative noise reduction on the different components of the equalized signal, in the reduced base of the point network.
[0049] There Fig. 2 schematically represents a flowchart of a linear detection method aided by point network and successive noise reduction according to a first embodiment of the invention.
[0050] In step 210, an LLL reduction of the point network Λ is performed H generated by the column vectors of the channel matrix, expressed in real form, i.e. of size n R × n T . This LLL reduction is known in itself, it uses a QR decomposition of the channel matrix ( H = QR Or Q is an orthogonal matrix and R an upper triangular matrix) and in return provides matrices Q̃ , R̃ And T with H̃ = Q̃R̃ Or H̃ is the base ( δ - )LLL reduced by Λ H (or channel matrix in the reduced basis), Q̃ is an orthogonal matrix and R̃ an upper triangular matrix and T is a unimodular transformation matrix of size n T × n T such as H̃ = HT , the column vectors of T forming the reduced base.
[0051] In step 220, a linear detection of the received signal is carried out by means of an equalization matrix. W̃ .
[0052] This linear detection could be, for example, a ZF type detection ( W̃ = W̃ ZF), MMSE type ( W̃ = W̃ MMSE ) or MF type ( W̃ = W̃ MF ), in a manner known per se. If we note y the vector received by the receiving antennas and r = W̃y the equalized vector. The latter can be considered as a real vector at n T = 2 N T elements.
[0053] In step 230, a representation of the equalized signal in the reduced base is obtained by calculating the vector c = Tr of size n T .
[0054] In step 240, the Gram matrix is calculated A = TT T< of the reduced basis and if the Gram matrix is non-diagonal, we go to step 250. The Gram matrix gives the scalar products of the column vectors of T .
[0055] Note that if the matrix A is diagonal, the column vectors of T are orthogonal to each other. In this case, the detection is completed at step 290.
[0056] In step 250, we determine for each vector of the reduced basis, that is to say each column vector t i i = 1, ..., n T , of the matrix T the column vector t k such as : k = arg max j ≠ i 〈 t i | t j 〉 Or < t i | t j 〉 represents the scalar product of the vectors t i And t j of the reduced basis. In other words, for each vector of the reduced basis, we determine the column vector which least well verifies the orthogonality condition. We thus construct a vector v of size n T whose elements are the elements are ν i = max j ≠ i 〈 t i | t j 〉 . A simple way to construct this vector is to search on each row of A , the non-diagonal element with the highest absolute value.
[0057] We then enter an iterative noise reduction loop in 260.
[0058] In step 260, it is determined whether the following condition is true: ∃ i ∈ 1 , … , n T tel que c i > s max + ε / ν i Or s max is the maximum amplitude of a point in the constellation and ε is a positive value representing a predetermined margin of error. The choice of the value of the margin of error is the result of a compromise between the speed of convergence and the tolerance of error on the detection result; the smaller this margin, the slower the convergence and the less erroneous the result will be.
[0059] If condition (18) is not verified, in other words if all the components of the equalized vector verify c i ≤ s max + ε / v i we finish the detection in 290.
[0060] Failing this, we determine in 270 the component ℓ of the equalized vector which deviates the most from the product constellation, more precisely which deviates the most from a safety zone (also called noise-free reception) Z ε in space ℝ n T surrounding the constellation produced with a noise margin of width ε : l = arg max i c i − s max − ε / ν i
[0061] In other words, we are looking for the direction of maximum error with respect to the product constellation. It is essential to note that this search is carried out in the reduced basis, in other words, in a non-orthogonal basis.
[0062] We then perform a noise reduction in 280 by subtracting a fraction from the equalized vector µ of the margin of error according to the direction of the component determined in the previous step: c = c − a l sgn c l . μ ε ν l where a, is the column vector of index ℓ of the Gram matrix, A . We will choose for example μ = 1 2 .
[0063] In other words, we bring the equalized vector closer to the noise-free zone Z ε by reducing the distance to this area in the direction of maximum error. Preferably, a fraction of 1 2 but it is clear that other fractions can be used.
[0064] Note that the equalized vector, updated using relation (20), does not only see its component c ℓ be modified since A is not a diagonal matrix.
[0065] Thus the detection algorithm will be able to pass several times over the same direction, correcting an error introduced in another direction.
[0066] The algorithm continues by returning to test step 260.
[0067] The iterations end when the equalized vector is in the noise-free reception area Z ε (in which case the noise will have been removed) or, failing that, when a maximum number of iterations is reached (absence of convergence). In the event of convergence, the equalized vector is at worst at a distance (1 - µ ) ε of the constellation and in the case where μ = 1 2 , in the middle of the margin of error.
[0068] In step 290, a quantization of the equalized vector that has been the subject of the iterative noise reduction is performed. This quantization is performed as in the prior art, by searching for the point in the constellation closest to the point representing the equalized vector.
[0069] There Fig. 3 illustrates the noise reduction step 280 in an example. In this example, it is assumed that there is a single transmitting antenna ( n T = 2) and that the modulation constellation used was a 4-QAM constellation. The product constellation is therefore reduced to this 4-QAM constellation.
[0070] The produced constellation has been represented in the reduced base, in other words the figure represents the image of the modulation alphabet A by the transformation T -1< .
[0071] The reduced basis consists of the vectors t 1 = 0 1 And t 2 = 1 1 and so the transformation matrix is T = 0 1 1 1 . The Gram matrix is then A = 1 1 1 2 .
[0072] The noise-free modulation symbols, i.e. the points of the modulation constellation are represented at A, B, C, D. The received signal, after equalization, i.e. the vector r, is represented by the point M. The vector c is assumed here to be − 2 − 1 .
[0073] The noise-free reception area, in other words the area Z ε surrounds the modulation constellation with a margin of error ε .
[0074] According to the direction of the vector t 2, point M belongs to the zone Z ε and therefore we do not reduce the noise in this direction.
[0075] On the other hand, depending on the direction of the vector t 1, point M does not belong to the zone Z ε and therefore it is brought back to this area in this direction by means of the expression (16) where μ = 1 2 . c = c − a l − 1 . 1 2 ε 1 = c + ε 2 a 1 = − 2 + ε 2 − 1 + ε 2
[0076] We therefore see that the new equalized vector is closer to the area Z ε of ε 2 in the direction t 1 but that on the other hand its component according to t 2 has been modified. This explains why in certain cases, one can leave the zone Z ε according to management t 2 and that an iteration in the direction t 2 is therefore necessary again.
[0077] In a second embodiment of the invention, a soft detection of the transmitted symbols is carried out. This will be the case, for example, when the successive symbols transmitted are the result of a coding of a binary sequence, for example the turbocoding of a packet of bits. In such a case, the convolutional decoder (using, for example, a SOVA algorithm or a conventional Viterbi algorithm) needs to receive as input the soft values of the bits of the sequence taking into account the symbols received.
[0078] We first consider the case of a SISO system, i.e. with a single transmitting antenna and a single receiving antenna.
[0079] If we note b m , q THE q ème< bit of the word used to generate the symbol emitted s m belonging to the modulation alphabet A, the receiver must evaluate the soft value or LLR ( Log Likelihood Ratio ), L ( b m , q ) defined by: L b m , q = ln P b m , q = 1 | y P b m , q = 0 | y = ln P y | b m , q = 1 P y | b m , q = 0 where we recall that vector y represents the received signal (and where it was assumed that the probabilities a priori of the two binary values were identical).
[0080] Conditional probabilities P ( y | b m , q = X ) (with X = 0 or 1) are determined from: P y | b m , q = X = ∑ u ∈ S m , q X P y | s = u Or P ( y | s = u ) is the conditional probability of receiving the vector y knowing that the symbol emitted is u ∈ S m , q X , S m , q X representing the set of symbols of the alphabet A admitting b m , q = X .
[0081] In a classical way, we can approximate the sum of the terms appearing on the right of the relation (23) by the predominant term (so-called Max-Log approximation): P y | b m , q = X ≈ max u ∈ S m , q X P y | s = u
[0082] Given the Gaussian density of the conditional probability, the soft value of the bit b m,q is expressed in the following form: L b m , q = 1 2 σ n 2 min u ∈ S m , q 1 ‖ y − Hu ‖ 2 − min u ∈ S m , q 0 ‖ y − Hu ‖ 2 Or H is the channel matrix (here reduced to a single element since the system is SISO).
[0083] Relation (25) extends easily to a MIMO system using linear detection: L b m , q = 1 2 σ n , m 2 min u ∈ S m , q 1 ‖ r − u ‖ 2 − min u ∈ S m , q 0 ‖ r − u ‖ 2 the vector u then representing the vector of the symbols emitted (and therefore an element of the constellation produced), Hbeing the complex matrix of the size channel N T × N T , r = Wy being the equalized vector where y represents the size vector N R signals received by the different receiving antennas, and where σ n , m 2 represents the noise variance affecting the component r m of the equalized vector. Due to this equalization, the noise variance is not identical for the different components: σ n , m 2 = σ n 2 ∑ i = 1 N T W ˜ m , i 2 + σ a 2 ∑ j = 1 j ≠ m N T β ˜ m , j 2 / β ˜ m , m 2 Or W̃ m , i is the element of the matrix W̃ to the line m and the column i , in other words corresponding to the receiving antenna m and to the broadcast antenna i , β m,j is the element of the matrix β = W̃H to the line m and the column j , σ a 2 is the average power of the modulation constellation (assumed to be identical for all transmitting antennas) and σ n 2 is the variance of the noise affecting the signal received by a receiving antenna (the variance is assumed to be identical for all receiving antennas).
[0084] The calculation of soft values according to (26) and (27) must be adapted in the case where an iterative noise reduction is carried out in addition to the equalization, as illustrated in relation to the Fig. 4 .
[0085] There Fig. 4 schematically represents a flowchart of a linear detection method aided by point network and successive noise reduction according to a second embodiment of the invention.
[0086] Steps 410 to 480 are identical to steps 210 to 280 already described in relation to the Fig. 2 and their description will therefore not be repeated here.
[0087] On the other hand, an initialization step was added in 455 as well as a noise power update step in 475.
[0088] More precisely, we initialize in 455 the noise variances (or noise powers) affecting the different components of the equalized vector by means of relation (27).
[0089] Then, at each iteration we update the noise variances in 475 σ n , m 2 by means of: σ n , m 2 = σ n , m 2 + μ 2 ε ν l 2 A m , l 2 Or A m ,ℓ is the element at line m and the column ℓ of the Gram matrix, A . It will be understood from expression (28) that only the variances relating to the components impacted by the noise reduction are updated.
[0090] Finally, in step 490, instead of making a hard decision as in step 290, the distances of the equalized vector to the different points of the product constellation are calculated, then soft values of the different bits are calculated using: L b m , q = 1 2 σ n , m 2 min u ∈ S m , q 1 ‖ c − u ‖ 2 − min u ∈ S m , q 0 ‖ c − u ‖ 2 noise variances σ n , m 2 affecting the different components of the equalized vector having been updated during successive iterations according to (28).
[0091] If turbo coding has been carried out on transmission to generate symbols on each of the antennas, BCJR type decoding can be carried out using the soft values provided by step 490 as input.
[0092] There Fig. 5 represents the throughput of the MIMO system as a function of the signal-to-noise ratio for a detection method according to the second embodiment of the invention and according to a detection method according to the state of the art.
[0093] The MIMO system used was a 2x2 system with 16-QAM modulation. A high-performance turbo coding R = 1 2 on 6144-bit packets was performed on transmission before 16-QAM modulation for each antenna.
[0094] The system flow rate is given by r (1- or ) Q< Or ris the transmission speed in bit / s, Q is the packet size and or is the bit error rate.
[0095] Curve 510 represents the system throughput when performing a ZF-type linear detection and a soft value calculation according to (22) and (23) at the reception, i.e. without the Max-Log approximation, before performing a BCJR decoding.
[0096] Curve 520 represents the system throughput when performing ZF type linear detection at reception, iterative noise suppression according to the second embodiment of the invention, and calculation of the soft values according to (22) and (23) before performing BCJR decoding.
[0097] Curve 530 represents the system flow rate under the same conditions as 510 but using the Max-Log approximation, in other words using the soft value expressions according to (26) and (27).
[0098] Curve 540 represents the throughput of the system according to the second embodiment of the invention, but using the Max-Log approximation, in other words using the soft value expressions according to (26) and (27) and the iterative updating of the variances according to (28), before performing channel decoding.
[0099] A significant improvement in throughput is observed when using the second embodiment of the invention, cf. curves (520) and (540), instead of a state-of-the-art detection method, cf. curves (510) and (530).
Claims
1. Detection method for a MIMO system receiver with NT transmit antennas and NR receive antennas, characterised by a channel matrix H, each transmit antenna transmitting a symbol belonging to a modulation constellation, the vector of the transmitted symbols belonging to a product constellation in the space ℝ 2 N T , the channel matrix H undergoing a LLL reduction in order to provide a reduced channel matrix H̃, such that H̃ = HT where T is a unimodular matrix, the vector of the signals received by the receive antennas being equalised by means of a linear detection matrix W̃, obtained from the reduced channel matrix H̃, in order to provide an equalised vector r, the equalised vector being represented by c = Tr in a reduced basis formed by the column vectors, tj, j = 1, ..., 2NT of the matrix T, characterised in that a receive area unperturbed by noise Zε is defined, in the representation according to the reduced basis, which area surrounds the product constellation with a predetermined tolerance margin ε, the equalised vector in the reduced basis then undergoing an iterative noise cancellation process wherein, upon each iteration, a search is carried out for the component, ℓ, of the equalised vector in the reduced basis, that is located the furthest from the receive area unperturbed by noise, and a noise vector is subtracted from the equalised vector in the reduced basis, c, the noise vector being determined by a l sgn c l . μ . ε ν l where ν l = max j ≠ l 〈 t l | t j 〉 and where aℓ is the ℓ-th column of the matrix A, where A is the Gram matrix of the reduced basis A = TTT, sgn(cℓ) is the sign of the f -th component of the equalised vector c and µ is the value of a fraction, the process being iterated until the point representative of the equalised vector is located within said receive area unperturbed by noise or until a predetermined number of iterations is reached, the result of the detection being obtained by making a hard or soft decision on the equalised vector at the end of the iterative process.
2. Detection method according to claim 1, characterised in that the linear detection matrix is selected from the ZF, MMSE or MF detection matrices respectively defined by: W ˜ ZF = H ˜ † = H ˜ H H ˜ − 1 H ˜ H W ˜ MMSE = H ˜ H H ˜ + σ n 2 I N R − 1 H ˜ H W ˜ MF = H ˜ H where σ n 2 is the variance of the noise affecting the signal received by a receive antenna and INR is the identity matrix of size NR.
3. Detection method according to claim 1 or 2, characterised in that the modulation constellations relative to the different transmit antennas are QAM constellations.
4. Detection method according to claim 3, characterised in that the modulation constellations relative to the different antennas are identical.
5. Detection method according to claim 4, characterised in that in order to determine the component of the equalised vector in the reduced basis, that is located the furthest from the area unperturbed by noise, the index ℓ is sought such that l = arg max i c i − s max − ε / ν i where ci is the ith component of the equalised vector C, smax is the maximum amplitude of the QAM constellation.
6. Detection method according to any of the claims 1 to 5, characterised in that μ = 1 2 .
7. Detection method according to one of the preceding claims, characterised in that a hard decision is made on the equalised vector at the end of the iterative process, by searching for the point of the product constellation that is located the closest to this vector.
8. Detection method according to any of the claims 1 to 5, characterised in that prior to the iterative process, the variances σ n , m 2 of the noise affecting each of the components of the equalised vector c are initialised by σ n , m 2 = σ n 2 ∑ i = 1 N T W ˜ m , i 2 + σ a 2 ∑ j = 1 j ≠ m N T β ˜ m , j 2 / β ˜ m , m 2 where σ n 2 is the variance of the noise affecting the signal received by a receive antenna, σ a 2 is the average power of the modulation constellation, W̃m,i is the element of the matrix W̃ in the row m and in the column i, β̃m,j is the element of the matrix β = W̃H in the row m and in the column j and that at each iteration, these variances are then updated by σ n , m 2 = σ n , m 2 + μ 2 ε ν l 2 A m , l 2 where Am,ℓ is the element in the row m and in the column ℓ of the Gram matrix A.
9. Detection method according to claim 8, characterised in that for each symbol transmitted by a transmit antenna m , the soft values of the different bits bm,q are computed, where bm,q is the qth bit of the word having been used to generate the transmitted symbol sm, the soft values being given by L b m , q = 1 2 σ n , m 2 min u ∈ S m , q 1 ‖ c − u ‖ 2 − min u ∈ S m , q 0 ‖ c − u ‖ 2 where σ n , m 2 is the value of the variance affecting the m -th component of the equalised vector c at the end of the iterative process, S m , q 0 is the set of points of the product constellation, assuming bm,q = 0 and S m , q 1 is the set of points of the product constellation, assuming bm,q = 1.
