Method for the optimization, under resource constraint, of the size of blocks of coded data

[0001] 1. Field of Invention

[0002] 2. Discussion of the Related Art

[0003] The present invention concerns a method of optimising the size of a block of coded data intended to be subjected to an iterative decoding, such as the one implemented in a turbodecoder or a turboequaliser. It applies in particular to the so-called third generation telecommunications systems.

[0004] The concept of turbocode was introduced by C. Berrou, A. Glavieux and P. Thitimajshima in their article “Near Shannon Limit Error-Correcting Coding and Decoding: Turbo-codes”, ICC-1993, Conference Proceedings, pp. 1064-1070, and has, as is known, been subject to many developments.

[0005] It will be recalled that turbocoding results from the concatenation of elementary codes separated by interleaving steps. The elementary codes can be of different types: recursive systematic codes (denoted RSC) for convolutional turbocodes or block codes (RS, BCH) for block turbocodes. The latter were described by R. Pyndiah, P. Combelles and P. Adde in an article entitled “A very low complexity block turbo decoder for product codes”, published in Proc. of IEEE Globecom, pp. 101-105, 1996. Different types of concatenation have been envisaged such as parallel concatenation, in which the same information is coded separately by each coder after having been interleaved and serial concatenation in which the output of each coder is interleaved before being coded by the following coder. For a detailed description, reference can be made, for parallel concatenation, to the aforementioned article by Berrou et al. and, for serial concatenation, to the article by S. Benedetto, G. Montorsi, D. Divsalar and F. Pollara, “Serial concatenation of interleaved codes: Performance analysis, design and iterative decoding”, JPL TDA Progr. Rep., Vol. 42-126, August 1996.

[0006] The decoding of a turbocode is conventionally effected by means of an iterative process also referred to as turbodecoding, consisting of a series of elementary decoding operations, each elementary decoding operation receiving weighted input values and supplying weighted output values, relating respectively to the input and, for a serial turbocode, to the output of the corresponding elementary coding operation. The weighted input and output values represent the probabilities of the binary or M-ary symbols at the input and (for serial turbocodes) at the output of the corresponding elementary coder. Deinterleaving and interleaving operations take place as a function of the interleaving operations performed at coding and enable each elementary decoding operation to take account of the data in the same order as that at the input of the corresponding coding operation.

[0007] FIG. 1a illustrates schematically an example of a turbocoder. The turbocoder depicted is a parallel turbocoder of dimension 2. The information x, in the form of a block of data, is coded by a first elementary coder 110 in order to supply a first output y₁ and by a second elementary coder 120, after having been interleaved by the interleaver 115, in order to supply a second output y₂. The systematic information is multiplexed at 130 with the coded information y₁ and y₂ before being interleaved by a channel interleaver 140.

[0008] On reception, the information is decoded by means of the turbodecoder depicted in FIG. 1b. After deinterleaving in the channel deinterleaver 145, the information x, y₁ and y₂ is demultiplexed by the demultiplexer 150. The elementary decoders 160 and 170, for example of the Log MAP type, correspond respectively to the elementary coders 110 and 120. The decoder 160, of the weighted input and output type, receives a priori information on x and the coded information y₁ in order to derive therefrom a posteriori information on x. The difference between the a posteriori information and the a priori information is called extrinsic information. The a priori information on x and the extrinsic information e₁ supplied by the first decoder are interleaved and then added in order to supply to the decoder 170 new a priori information on x (or more precisely on x′, the interleaved version of x). The decoder 170 estimates, from this a priori information and the coded information y₂, a posteriori information on x′. The intrinsic information e′₂ is derived therefrom by difference in 171 before being deinterleaved in the deinterleaver 180 and is then added at 151 to the systematic information in order to supply new a priori information on x to the decoder 160. The decoding steps are then repeated, for a predetermined number of iterations n_iterations. The flexible values at the output of the decoder 170 are submitted to a decision device 190 in order to supply hard values. An error correction check 195 operating on the output of 190 determines whether the decoded block is error-free and, in the affirmative, interrupts the iterations without having to await the predetermined number n_iterations. Alternatively, in order to decide on the stoppage of the iterations, the turbodecoder can use, instead of a CRC, another stoppage criterion on the weighted values, such as the one disclosed, for example, in the article by J. Hagenauer et al. entitled “Iterative decoding of binary block and convolutional codes”, IEEE Transactions on Information Theory, vol. 42, pp. 429-445, published in March 1996, or in the article by M. Moher entitled “Decoding via cross entropy minimization”, in Proc. IEEE Globecom Conf., Houston, Tex., pp. 809-813, published in December 1993.

[0009] The principle of iterative decoding was transposed to equalisation by C. Douillard et al. in an article entitled “Iterative Correction of Intersymbol Interference: Turbo-equalization”, published in European Trans. Telecomm., Vol. 6, N^o 5, September/October, pages 507-511. This iterative equalisation method, also referred to as turboequalisation, starts from the principle that a channel affected by intersymbol interference (ISI) can be considered to be equivalent to a convolutional coder and consequently the concatenation of a channel coder, an interleaver and a transmission channel can itself be considered to be the equivalent of a turbocoder.

[0010] FIG. 2b illustrates schematically the structure of a turboequaliser. It will be assumed that the data were, on the same side as the transmitter depicted in FIG. 2a, the subject of channel coding in a coder 201 before being interleaved in an interleaver 202 and subjected to M-ary to symbol mapping in the modulator 203. The data are transmitted in the form of a block of symbols interleaved by the channel interleaver 204. The turboequaliser comprises first of all a channel deinterleaver 210 followed by a weighted output equaliser of the Log-MAP type supplying flexible values of coded data. These data are deinterleaved in the deinterleaver 230 before being decoded by a decoder 240 of the Log-MAP type with weighted output. The flexible values issuing from the decoder 240 are submitted to a decision device 250 supplying the corresponding hard values. The weighted values at the input of the decoder 240 are subtracted from the output values in order to supply extrinsic information e. After interleaving, the extrinsic information is on the one hand subtracted at the output of the equaliser 220 and on the other hand remodulated before being transmitted to the equaliser. From the symbols received and the remodulated extrinsic information, the equaliser 220 proceeds with a new a priori estimation. The turboequaliser thus carries out a predetermined number of iterations on a block of symbols. An error correction check 260 at the output of 250 diagnoses the presence or absence of errors, the iteration process is interrupted without the iterations without having to wait for the predetermined number n_iterations. The stoppage criterion can alternatively relate to the weighted values, as seen above.

[0011] It should be noted in relation to the design that both for turbocoding and for turboequalisation the iterations can be carried out one after the other within the same processor or in dedicated circuits arranged in a pipeline, each being responsible for an iteration of given rank.

[0012] In the cases of turbocoding and turboequalisation, the performance in terms of bit error rate (BER) or packet error rate (PER) is all the better, the greater the length of the size N of the internal interleaver (IL). This is because the increase in the size of the internal interleaver, that is to say the size of the block on which the interleaving is carried out, improves both the Hamming weight distribution of the turbocode and the efficacy of the iterative decoding. The interleaving gain varies according to the type of turbocode. Hereinafter the term “performance” will be used indifferently for the BER or the PER on the decoded data.

[0013] FIG. 3 shows, by way of example, the interleaving gain for the turbocoder of FIG. 1a, whose elementary coders are RSC (Recursive Systematic Convolutional) coders of polynomials (13,15)_oct, when the size N of the block changes from 200 to 1000.

[0014] When the signal to noise ratio (SNR or C/N standing for Carrier/Noise) is low, it is known from the state of the art that high values of N can be used for improving the performance of the system. However, the increase in N is to the detriment of the latency of the receiver, that is to say the time at the end of which a block of data or symbols received is actually available in decoded form. This is because the latency time at the receiver can be written: $\begin{array}{cc}{T}_{\mathrm{latency}}=T_{\mathrm{rec}}+\frac{N}{D_d}·n_{\mathrm{iterations}}& \left(1\right)\end{array}$

[0015] where D_d is the decoding rate per iteration and T_rec is the time at the end of which the block of size N is available in deinterleaved form, at the receiver. If a channel interleaving of size M, with M≧N, is applied at the transmitter, this will give T_rec=M/D_u where D_u is the useful information rate. If no channel interleaving is used, the reception time will simply be T_rec=N/D_u.

[0016] For a given maximum latency time or set latency time and a predetermined number of iterations, formula (1) makes it possible to obtain the block size giving the maximum performance.

[0017] Conversely, a reduction in the latency time can be obtained by reducing the size of the blocks, at the cost however of a correlative reduction in the performance of the system. In order to regain acceptable performance, it is known that the transmission power can be increased, which naturally burdens the resources of the system.

[0018] The problem at the basis of the invention is to determine, for a given maximum latency time and in general terms for given resources, a block size N affording better performance than in the state of the art.

[0019] The solution afforded by the invention is defined by a method of optimising the size of coded data blocks, intended to be subjected to an iterative decoding, comprising a first step evaluating a resource available for the decoding of a block of normal size and a second step seeking, amongst a plurality of block sizes which are submultiples of the normal size by an integer factor greater than or equal to 1 and requiring on average a number of iterations compatible with the said available resource, the one which makes it possible to obtain the lowest error rate at the output of the iterative decoding.

[0020] The available resource is for example an available time or energy for receiving all the data of a block and for effecting the iterative decoding on the said block. If the data have been interleaved by a channel interleaver, the said available resource can be a time or energy available for obtaining the data of a block in deinterleaved form and for effecting the iterative decoding on a block of deinterleaved data.

[0021] Advantageously, the first step also determines, for the normal block size, the maximum number of iterations compatible with the said available resource. There are then estimated, according to a required maximum error rate, the minimum number of iterations which must be effected by the iterative decoding, for the current block size and the signal to noise ratio.

[0022] The second step may not be effected if the said maximum number of iterations exceeds the said minimum number of iterations.

[0023] According to one embodiment, the second step selects, from amongst the sizes of blocks of which are submultiples of the normal size and amongst different maximum numbers of iterations, the block size and the largest maximum number of iterations associated with this size, giving a mean number of iterations compatible with the said available resource. The optimum size which makes it possible to obtain the lowest error rate at the output of the iterative decoding is sought amongst the selected submultiple sizes and for the associated maximum numbers of iterations.

[0024] For a size which is a submultiple by a given factor k and a given maximum number of iterations, the mean number of iterations can be determined according to the signal to noise ratio as the mean value of the number of iterations which would be effected by the iterative decoding for each block in a succession of blocks of submultiple size, the iterations being stopped on a block of the said submultiple size if a predetermined reliability criterion is satisfied or if the number of iterations for this block reaches the said given maximum number of iterations.

[0025] Advantageously, the mean numbers of iterations for different submultiple sizes, different maximum numbers of iterations and different signal to noise ratios are stored in a table. The table can be updated as the iterative decoding continues. Where applicable, the mean numbers of iterations will be obtained by interpolation from values available in the table.

[0026] According to a first variant, after the data in a block of initial size has been coded as a sequence of sub-blocks of optimum size, the sub-blocks are decoded, one by one, by means of a succession of iterations of the iterative decoding, the iterations being stopped for a sub-block if a predetermined reliability criterion is satisfied or if the number of iterations reaches the maximum number of iterations associated with the said optimum size.

[0027] According to a second variant, after the data in a block of initial size have been coded as a sequence of sub-blocks of optimum size, the sub-blocks are decoded by effecting successively on each sub-block an iteration of the iterative decoding, an iteration not being performed on a sub-block if a predetermined stop criterion is satisfied or if the number of iterations reaches the maximum number of iterations associated with the said optimum size. Alternatively, an iteration will not be performed for a sub-block if a predetermined stop criterion is satisfied or if the said available resource is exhausted.

[0028] The invention is also defined by a device for the iteration decoding of blocks of data coded by a turbocoder having means for implementing the optimisation method disclosed above, the said means supplying an optimum block size and the device also comprising means for transmitting optimum block size information to the turbocoder.

[0029] The invention also concerns a coding/decoding system comprising a turbocoder adapted to code blocks of data and an iterative decoding device as defined above, adapted to decode the blocks of data coded by the turbocoder, the latter comprising means for receiving the said optimum block size information and for modifying the size of at least one internal interleaver according to the said received information.

[0030] The invention also concerns a device for coding blocks of data, the device having means for implementing the optimisation method as defined above, the said means supplying an optimum block size, the device also comprising means for adaptively modifying the size of the blocks of coded data according to the said optimum block size.

[0031] Finally, the invention concerns a device for the turboequalisation of blocks of data coded by a coder and modulated, the device having means for implementing the optimisation method defined above, the said means supplying an optimum block size, the said device also comprising means for transmitting optimum block size information to the coder.

[0032] The characteristics of the invention mentioned above, as well as others, will emerge more clearly from a reading of the following description of an example embodiment, the said description being given in relation to the accompanying drawings, amongst which:

[0044] The general idea at the basis of the invention is to take advantage of the fact that, in an iterative decoder, for a predetermined maximum number of iterations n_iterations which will also be referred to as the set number, the block can be entirely decoded before the last iteration provided for. For example, if the stop criterion for the iterations is a perfect ideal criterion (zero BET, that is to say with a complete absence of errors in the block), the mean number of iterations, {overscore (n)}_iterations, is often less than the set number, as can be discerned in FIG. 4a. This figure shows the change in {overscore (n)}_iterations as a function of the signal to noise ratio for different values of n_iterations and for a block size of N=1000. FIG. 4b shows this change for a block size N=10,000. In the two examples, it can be seen that, for a given maximum number of iterations, n_iterations, the mean number of iterations does not reach the maximum, and this all the more so when the signal to noise ratio is high. Naturally, in practice, there is no perfect criterion: the absence of errors is tested, for example, by means of a CRC and the iterations are stopped as soon as the CRC no longer detects any errors. The conclusions concerning the change in {overscore (n)}_iterations with respect to n_iterations do however remain valid.

[0045] FIG. 5 shows schematically a histogram of the number of iterations necessary for obtaining a zero BET (and therefore PET). As, in practice, it is not possible to effect an infinite number of iterations, there is fixed, as seen above, a set number of iterations n_iterations compatible with the latency time and the size of the interleaver according to (1) is derived therefrom. The mean {overscore (n)}_∞ of the histogram corresponds to an infinite set number. For a given set number n_iterations, the mean {overscore (n)}_iterations is taken on the hatched area of the curve.

[0046] Let it be assumed that there is given a set latency time T and that there has been determined a pair of values (n_iterations, N) satisfying T_latency≦T where T_latency is given by (1). n_iterations and T_latency will hereinafter be denoted respectively n_iterations⁽¹⁾ and T_latency⁽¹⁾.

[0047] Let it be assumed now that the size N of the block, and therefore the size of the internal interleaver of the turbocoder or of the turboequaliser, is divided by a non-zero integer k. The latency time T_latency^(k) at the end of which the k blocks of reduced size N/k are decoded is then expressed as: $\begin{array}{cc}{T}_{\mathrm{latency}}^{\left(k\right)}=T_{\mathrm{rec}}^{\left(k\right)}+\sum _{i=1}^k\frac{N/k}{D_d}·n_{\mathrm{iterations}}\left(i\right)=T_{\mathrm{rec}}^{\left(k\right)}+\frac{N}{D_d}·{\stackrel{\_}{n}}_{\mathrm{iterations}}^{\left(k\right)}& \left(2\right)\end{array}$
where T_rec^(k) is a reception time at the end of which the k blocks of size N/k are available, where applicable in interleaved form, at the receiver. In the absence of channel interleaving, this reception time is equal to $\frac{N}{k·D_u}+\sum _{i=2}^k\delta t_{\mathrm{rec}}^{\left(i\right)}\mathrm{where}\frac{N}{k·D_u}$
is the reception time of the 1^st block of size N/k and δt_rec⁽ⁱ⁾ is the time which the receiver must wait between the end of the decoding of the (i-1)^th block and the end of the reception of the i^th block. Consequently this gives: $\frac{N}{k·D_u}\le T_{\mathrm{rec}}^{\left(k\right)}<\frac{N}{D_u}.$
On the other hand, when a channel interleaving of size M_k with M_k≧N/k has been applied at the transmitter, this gives: $T_{\mathrm{rec}}^{\left(k\right)}=M_k/D_u+\sum _{i=2}^{N/M_k}{\delta }_{\mathrm{rec}}^{\left(i\right)}$
where δ_rec⁽ⁱ⁾ is the waiting time between the end of the deinterleaving of the (i-1)^th group of k.M_k/N blocks and their decoding on the one hand and the end of reception of the i^th group on the other hand. In practice, M_k=N will be taken and therefore T_rec^(k)=M_k/D_u.

[0048] n_iterations(i) is the number of iterations effected by the decoder on the i^th block of size N/k using a predetermined stoppage criterion and for a given set number of iterations n_iterations^(k);

[0049] {overscore (n)}_iterations^(k) is the mean number of iterations, taken on the set of k blocks. It should be noted that {overscore (n)}_iterations^(k) in general depends on the set number n_iterations^(k) chosen for the block size N/k and on the signal to noise ratio. Hereinafter the term (N/k, n_iterations^(k)) will be used for a turbocode operating on blocks of size N/k and decoded by means of at most n_iterations^(k) per block.

[0050] It should first of all be noted that, in general, the size M_k of the channel interleaving decreases (in the broad sense) with k, and therefore T_rec^(k) is a decreasing function of k: T_rec^(k+1)≦T_rec^(k) whether or not a channel interleaving is applied.

[0051] It should also be noted that the stoppage criterion can tolerate a predetermined residual error rate. Use can be made, as mentioned above, of a stoppage criterion operating on a weighted output or outputs or extrinsic information and the stoppage of the iterations can be triggered when a predetermined reliability threshold is reached. In all cases, the stoppage criterion relates to the reliability of the blocks of decoded data and for this reason will also be referred to as the reliability criterion. The stoppage criterion used is chosen so as to be identical for all values of k.

[0052] The invention is based on the fact that it is generally possible to find an integer k and a set number n_iterations^(k) such that the constraint T_latency^(k)≦T is satisfied while giving better performance than for k=1. The relationship T_latency^(k)≦T is generally less constraining than T_latency⁽¹⁾≦T since it involves the mean number of iterations ({overscore (n)}_iterations^(k)) rather than the maximum number of iterations (n_iterations⁽¹⁾).

[0053] The comparison between the two situations is illustrated in FIG. 6. An iteration is represented by a horizontal line extending over the length of the block. The line is dotted if errors remain after the iteration and solid in the contrary case. The top part of the figure illustrates the process of iterative decoding on a block of size N and for a set number n_iterations⁽¹⁾=3 satisfying the latency constraint 3N/D_d≦T (it is assumed in order to simplify that there is no channel interleaving and that the reception time is zero). The iterative process has not converged at the end of the 3^rd iteration and the decoded block has residual errors. The bottom part of the figure illustrates the situation when the size of the block is divided by 10. The set number n_iterations⁽¹⁰⁾ has been chosen so as to be equal to 8. It can be seen that, for the great majority of the blocks, the iterative process converges well before n_iterations⁽¹⁰⁾ and even before n_iterations⁽¹⁾ iterations: in this example, only block N^o 8 still has errors at the end of the 8^th iteration. The latency constraint is well complied with since {overscore (n)}_iterations⁽¹⁰⁾=3. Division of the block made it possible, for constant latency and calculation resources, to “concentrate” the iterations on the critical blocks. Thus the performance of the receiver is, before division of the block, that of a turbocode of length N with 3 decoding iterations whilst, after division, it is that of a turbocode of length N/10 after 8 decoding iterations.

[0054] FIG. 7 illustrates the performance of a turbocode coded by the turbocoder of FIG. 1a consisting of two elementary coders of polynomials (13,15)_oct, when the length of the block (and therefore of the interleaver 115) changes from N=10,000 to N=1000 and when the set number of iterations changes correlatively from n_iterations⁽¹⁾=3 to n_iterations⁽¹⁰⁾=8. This comparison is valid only for {overscore (n)}_iterations⁽¹⁰⁾≦3, which corresponds to C/N≦−0.75 dB, as can be seen in FIG. 4a. It can be seen in FIG. 4b that, for a signal to noise ratio of −0.75 dB and n_iterations⁽¹⁾=3, the mean number of {overscore (n)}_iterations⁽¹⁾ is equal to the maximum value n_iterations⁽¹⁾. On the other hand, under the same conditions, with the same signal to noise ratio and n_iterations⁽¹⁰⁾=3, it can be seen in FIG. 4a that the mean number of iterations {overscore (n)}_iterations⁽¹⁰⁾ has departed from its maximum value. n_iterations⁽¹⁰⁾ can be increased as long as {overscore (n)}_iterations⁽¹⁰⁾ remains below 3, which is the case up to n_iterations⁽¹⁰⁾=8. It can be seen in FIG. 7, for the range C/N≧−0.75 dB, that both in terms of BET and PET the division of the block size by a factor 10 affords a substantial improvement in the performance because of the increase in the maximum number of iterations for constant latency.

[0055] In general terms, the division of a block by a factor k makes it possible to change, for constant latency, from a turbocode of length N associated with a maximum number of iterations n_iterations⁽¹⁾ to a turbocode (N/k, n_iterations^(k)) of length N/k associated with a maximum number of iterations n_iterations^(k). If, in order to comply with the latency constraint: {overscore (n)}_iterations^(k)=n_iterations⁽¹⁾, then n_iterations^(k)>n_iterations⁽¹⁾, which allows a concentration of the iterations on the deficient blocks. The choice of the factor k is made according to the performance of the turbocodes (N/k, n_iterations^(k)) satisfying T_latency^(k)≦T. The choice of k is in fact a compromise between the interleaving gain, which is all the lower, the smaller N/k, and the gain due to the iterative decoding (n_iterations^(k)>n_iterations⁽¹⁾), which is all the higher, the greater k is.

[0056] FIG. 8 gives schematically a flow diagram for the method of optimising the block size according to a first embodiment of the invention. The transmission depicted at 810 uses an interleaving length N. On reception, there is determined at 820, from N and from an estimation at 821 of the calculation resources and/or the maximum latency time T, the set number of iterations n_iterations⁽¹⁾ by means of equation (1). Likewise there is determined at 830, from N and the signal to noise ratio C/N, by means of a table 831, the number of iterations n_req required to reach a required level of BET or PET. Next, at 840, n_iterations⁽¹⁾ and n_req are compared. If n_iterations⁽¹⁾≧n_req, the size of the block cannot be divided (since this would, for equal performances, result in an increase in the number of iterations) and the optimisation process is terminated. On the other hand, if n_iterations⁽¹⁾<n_req, there are sought at 850, in a table 851, having regard to the signal to noise ratio, all the pairs (N/k, n_iterations^(k)), where k≧1 and divisor of N, such that T_latency^(k)≦T. The value of k giving the best performance (BET, PET) is then selected, as well as the (largest) associated maximum number of iterations n_iterations^(k) making it possible to satisfy T_latency^(k)≦T. At 860 it is tested whether the value of k selected is greater than a threshold value k_min guaranteeing the averaging of the actual numbers of iterations. In the negative, the division of the size of the block (k=1) is not carried out. On the other hand, in the affirmative, the factor k is transmitted to the turbocoder on a return channel. The turbocoder adapts the size of the blocks to the value N/k and the size of its internal interleaver or interleavers accordingly.

[0057] According to a variant of the invention, the test 840 is not carried out and there are sought directly at 850 the value of k and the associated number n_iterations^(k) giving the best performance under the latency constraint. It is thus ensured that the best possible performance will be obtained subject to compliance with the said constraint.

[0058] The resources can be estimated at 821 by means of a dedicated resource management algorithm, or, a posteriori, by examining the maximum number of iterations effected for a block size N/k.

[0059] The tables 831 and 851 can be produced in advance or subsequently from the performance obtained and the actual numbers of iterations. They can also be produced in advance or completed as use of the system continues. The values unavailable in the tables can be calculated by interpolation from the existing values.

[0060] Naturally, after division of the size of the block, the system can subsequently return to the largest size block or even to the initial block size if the resource constraint (for example the latency constraint) is relaxed or if the quality of service (QoS) no longer requires optimum performance.

[0061] According to a second embodiment of the invention, the process of optimising the size of the block is effected on transmission and the system does not have any return channel between the receiver and the transmitter. The algorithm does not include the estimation of the resources of the receiver but decides, for a given type of application, to improve the error rate (BET, PET) without increasing the latency time of the receiver.

[0062] When the size N of the block has been divided by a factor k, the iterative process can be effected in two possible ways.

[0063] According to a first variant, the 1^st decoding iteration is effected for each of the k blocks, and then the 2^nd decoding iteration is effected for the blocks with residual errors after the ₁^st iteration and so on until the maximum number of iterations is reached or better still until the said available resource or the time allotted for the decoding of the k blocks is exhausted. At each pass the blocks having residual errors are identified by means of the stoppage criterion. In this way, it is certain that the available resource or the allotted latency will be used in an optimum fashion and “lingering” too long on the decoding of a block, which could be prejudicial to the following blocks if k is not sufficiently large to ensure sufficient averaging of the number of iterations, is avoided.

[0064] In the case of the example in FIG. 6, the implementation of this variant would be as follows:

[0065]

1	Iteration #1 for blocks 1 to 10
2	Iteration #2 for blocks 1 to 10	−> blocks 1, 2, 5, 6, 10
		correct
3	Iteration #3 for blocks 3, 4, 7, 8, 9	−> blocks 3, 4, 7, 9 correct
4	Iteration #4 for block 8
5	Iteration #5 for block 8
6	Iteration #6 for block 8
7	Iteration #7 for block 8
8	Iteration #8 for block 8

[0066] This first decoding variant is advantageous when the channel interleaving time T_ilc is fixed or when k is small. This is because, in the first case, it is necessary to wait for the length of time T_ilc before being able to decode a block and, in the second case, it makes it possible to remedy any averaging defect.

[0067] According to a second variant, each of the k=10 blocks is successively decoded. At a maximum n_iterations⁽¹⁰⁾=8 decoding iterations are effected successively on each block using the stop criterion for ideally reaching {overscore (n)}_iterations⁽¹⁰⁾=3 iterations per block on average over all the k blocks.

[0068] In the case of the example in FIG. 6, the implementation of this variant would be as follows:

[0069]

1.	iterations 1 to 2 for block 1
2.	iterations 1 to 2 for block 2
3.	iterations 1 to 3 for block 3
4.	iterations 1 to 3 for block 4
5.	iterations 1 to 2 for block 5
6.	iterations 1 to 2 for block 6
7.	iterations 1 to 3 for block 7
8.	iterations 1 to n(10)iterations = 4 for block 8
9.	iterations 1 to 3 for block 9
10.	iterations 1 to 2 for block 10

[0070] This second variant is advantageous when the channel interleaving time T_ilc is not fixed or when k is large. This is because, if the time T_ilc is not fixed, it can be reduced, possibly at the cost of a loss of performance, to a value corresponding to an interleaving on k′<k consecutive blocks of size k/N. As a result the latency time T_latency^(k) given by (2) can also be reduced. This gain in latency time can be converted into a performance gain by increasing the maximum number of iterations n_iterations^(k) for each of the k blocks.

[0071] It is essential to note that the latency constraint which has essentially been dealt with up to now is only one example of a resource constraint to be complied with for effecting the iterative decoding. Resource can also mean the energy expended by a processor. This will notably be the case if the iterative decoding is dealt with by a fast processor: the constraint relates in this case more to the number of operations to be performed and therefore to the energy expended than to the calculation time itself. In more general terms, resource will mean an additive physical quantity, or a combination of such physical quantities, liable to be expended by the calculation means for effecting the iterative decoding.

[0072] In addition, although the description of the invention has essentially been illustrated as a process of the optimisation of the size of turbocoded blocks and resource constraint, it also applies, in a similar manner, to a process of optimising the size of turboequalised blocks under resource constraint, whether this optimisation is carried out at the turboequaliser (the first embodiment) or at the transmitter (the second embodiment).