Metrics Calculations Utilising Pre-Stored Values

Abstract

The trellis processing apparatus and method achieve a reduced computational structure when implementing a DDFSE-like branch metrics calculation by pre-storing computed values, which can be re-used.

Description

In mobile communication systems, the transmission link suffers from a number of impairments. Two such effects are thermal noise and multipath fading.

Multipath fading can result in Intersymbol Interference (ISI) at the receiver if the delay spread of the channel is larger than the modulation period (‘Digital Communications’, Proakis, 2^nd Edition, McGraw-Hill). Hence, in a given propagation environment, ISI will become more of a problem as the transmission rate increases. For communication systems aimed at providing medium to high data rate services, the presence of ISI can severely limit the link throughput and degrade the quality of service experienced by the user.

Some digital communication systems also introduce ISI by design. This is the case in the E-GPRS system, where the modulation pulse shape used to improve the spectral efficiency of the transmitted signal generates ISI. For more information, see 3GPP TS 05.04 V8.4.0 (2001-11), Technical Specification 3rd Generation Partnership Project; Technical Specification Group GSM/EDGE Radio Access Network; Digital cellular telecommunications system (Phase 2+); Modulation.

Another source of performance degradation experienced by a user of a Time-Division, Multiple-Access (TDMA) cellular communication system is interference generated by other users in the system using the same carrier or adjacent carriers. These interference effects, referred to as co-channel interference and adjacent channel interference respectively, can greatly reduce the capacity of a cellular system.

All of the impairments described above make it difficult for a receiver to reliably recover information that a transmitted signal intended to convey, leading to the use in receivers of complex algorithms for demodulating received signals. The implementation complexity of such algorithms will have a significant impact on a digital receiver in terms of overall silicon die size, processor speed, power consumption and memory requirements. Hence, the use of an efficient receiver architecture offering good transmission link performance is of considerable importance.

In order to improve the reliability of a communication link, Forward Error Correction (FEC) coding can be embedded in a transmitted signal. An FEC coding operation introduces redundancy to a transmitted signal and this redundancy can then be used at a receiver to improve the accuracy of estimates of the transmitted data that are generated by the receiver. However, for FEC coding in a transmitted signal to be most beneficial, it is important that such a signal is demodulated by a receiver in a format which can be best interpreted by an FEC decoding process within the receiver. A number of such receivers have been proposed in the past. See, for example, ‘Optimal decoding of linear codes for minimizing symbol error rate’, L. Bahl, J. Cocke, F. Jelinek, J. Raviv, IEEE Trans. on Information Theory, Volume: 20, March 1974; ‘A Viterbi algorithm with soft-decision outputs and its applications’, J. Hagenauer, P. Hoeher, GLOBECOM'89, Dallas, November 1989; ‘On the equivalence between SOVA and Max-Log MAP decodings’, M. P. C. Fossorier, F. Burkert, S. Lin and J. Hagenauer, IEEE Communications Letters, vol. 2, no. 5, May 1998; ‘Soft information in concatenated codes’, B. Rislow, T. Maseng, O. Trandem, IEEE Trans. on Communications, Vol. 44, Issue 3, March 1996; and ‘TCM on frequency-selective fading channels: a comparison of soft-output probabilistic equalizers’, P. Hoeher, GLOBECOM '90., December 1990. However, the implementation complexity of such prior receiver architectures is usually high.

According to one aspect, the invention provides trellis processing apparatus comprising means for pre-storing different products of channel coefficient values and signal symbol values and means for calculating a branch metric utilising one or more of said pre-stored products.

The invention also consists in a trellis processing method comprising pre-storing different products of channel coefficient values and signal symbol values and calculating a branch metric utilising one or more of said pre-stored products.

In certain embodiments, a reduced state trellis technique, such as DDFSE, is used, providing groups of hypothesised and hard symbol decisions associated with each trellis states. These groups can contain just a single symbol decision.

The invention is can be realised in hardware or as software on a processor, or a combination thereof.

In certain aspects, the invention relates to a reduced state trellis. Such a trellis is one in which the hypothesised variables making up the state descriptions are converted to definite decisions to reduce the number of states present in the trellis. Such trellises are processed by, for example, RSSE and DFSE demodulator architectures.

By way of example only, certain embodiments of the invention will now be described with reference to the accompanying figures, in which:

FIG. 1 presents a model of a communication link;

FIG. 2 describes one possible implementation of the proposed invention;

FIG. 3 presents how the different branch metrics are computed in the embodiment of the invention shown in FIG. 2;

FIG. 4 describes how, in the E-GPRS system, received information can be equalised in two parts; and

FIG. 5 presents an implementation of the branch metric computation that is an alternative to that shown in FIG. 3.

In order to describe the computations performed by a receiver, it is first necessary to present a model of the transmission link in which the receiver participates. A suitable model is presented in FIG. 1 and will now be described in some detail. This model assumes that the information is transmitted in blocks of bits rather than as a continuous stream. However, it should be noted that the invention described later in this document is applicable to both types of transmission.

On the transmitter side 101, the information signal is processed by an error-protection coding unit 102 and a modulation unit 103. The generated signal then goes through a propagation channel 104 and the resulting signal is acquired by a receiver 108 which then attempts to recover the information signal.

According to the model, a transmission block {u_k}_{k∈(1, . . . ,K)} is made of K information bits, u_k ∈{0,1}. Error protection is added to these bits in order to improve the reliability of the transmission. The error protection coding unit 102 generates from the block {u_k}_{k∈(1, . . . ,K)} a transmission block {d_k}_{k∈(1, . . . ,D)} made of D (where D>K) information bits, d_k Å{0,1}.

A number of different techniques can be used by unit 102 for the error protection coding process. For example, in the E-GPRS system, convolutional codes with varying coding rates are used by the transmitter. Interleaving and bit reordering techniques can also be used in order to improve the resilience of the receiver to bursts of errors.

In the modulation unit 103, the information bits d_k are grouped into C sets of M bits (it can be assumed, without loss of generality, that D=M×C). Each of these sets can be denoted Δ_k, where Δ_k{=d_M×k, . . . , d_{(M×k)+(M−1)}.

Each set of M information bits Δ_k is modulated onto the complex plane using a modulation scheme M that maps sets of M bits onto the complex plane. This mapping is performed by a modulation unit 103. For example, in the case of an 8 PSK modulation, the modulation M can be expressed as:

$\begin{array}{c}M\left({\Delta }_k\right)=M\left(\left\{d_{3\times k},d_{3\times k+1},d_{3\times k+2}\right\}\right)\\ =\exp \left(2·j·\pi \times \left(\frac{\begin{array}{c}\left(4·d_{3\times k}\right)+\left(2·d_{3\times k+1}\right)+\\ \left(1·d_{3\times k+2}\right)\end{array}}{8}\right)\right)\end{array}$

A slightly modified version of the 8 PSK modulation described by the above equation is used in the E-GPRS system.

A set of M information bits {d_M×k, . . . , d_{(M×k)+(M−1)}} can be uniquely identified with a single decimal number i {0≦i≦2^M−1} calculated using the one-to-one function D described in the following equation:

$D\left(\left\{d_0,\dots ,d_{M-1}\right\}\right)=\sum _{i=0}^{M-1}d_i\times 2^i$

This equation maps the M binary values in a set Δ_k to a unique value in the set {0, . . . , 2^M−1}. The set of information bits d_b (b∈0, . . . ,M−1) that verify the above equation for a given value of i can be denoted D_b⁻¹(i) (b∈{0, . . . ,M−1}).

The C modulated symbols representing a transmission block are transmitted over the air and in the process are distorted by the propagation channel 103. Assuming a general model with memory for the propagation channel, the samples {S_k}_{k∈(1, . . . ,C)} at the input of a receiver can be expressed as:

$\{\begin{array}{c}{s}_k=F\left(c_k,{\xi }_{k-1}\right)\\ {\xi }_k=S\left(c_k,{\xi }_{k-1}\right)\end{array}\hspace{1em}$

Here, c_k=M(Δ_k) and ξ_k represents the state (memory) of the propagation channel when the k^th modulated symbol is transmitted. Note that any filtering performed by either the transmitter and/or the receiver can be incorporated in the propagation channel model. The mappings F and S used to model the propagation channel can be time varying. However, to simplify the notations, it is assumed in this document that these mappings do not depend on time. Note, however, that the approach described here is also applicable to time-varying channels.

In most cases, the propagation channel mappings F and S can be modelled as linear filtering operations:

$\{\begin{array}{c}{s}_k=\sum _{i=0}^{L-1}h_ic_{k-i}\\ {\xi }_k=\left\{c_k,\dots ,c_{k-L+1}\right\}\end{array}\hspace{1em}$

In the above example, it is assumed that the memory of the channel is limited to L modulated symbols. In reality, the memory of the channel may be infinite. However, for any desired power threshold T, it is usually possible to find a value L such that:

$\left(\sum _{k\ge i}{h_k}^2\right)\le T\mathrm{for}i\ge L-1$

Hence, by selecting the threshold T such that the residual power is low enough, it is possible to assume that the channel memory is limited. When this is done, the channel mapping can be described with just a set of filter coefficients {h_i}_{i∈{0, . . . ,L−1}}.

In the model described here, it has been assumed, without loss of generality, that the filter representing the channel propagation is causal.

In order to recover the transmitted symbols c_k, the receiver will need to know the propagation channel mapping. However, the receiver will usually not have prior knowledge of the propagation channel conditions. It may nevertheless be possible for the receiver to generate estimates of the channel coefficients {h_i}_{i∈{0, . . . ,L−1}} which can be used in place of the true values. For example, in the EGPRS system, a transmitted signal will incorporate a pattern, referred to as training sequence, which is known to the receiver and the receiver can use this training sequence to generate an estimate of the propagation channel conditions.

At the receiver 108, the signal {S_k}_{k∈(1, . . . ,C)} is first processed by the signal conditioning unit 105, a receiver front-end, to generate a new sequence of received symbols {r_k}_{k∈(1, . . . ,C)} that is input to a demodulation unit 106.

The demodulation unit 106 derives estimates of the Log-Likelihood Ratios (LLRs) of the coded bits. The LLR of the coded bit d_k is expressed as:

${\lambda }_k=\log \left(\frac{P\left(d_k=1R\right)}{P\left(d_k=0R\right)}\right)$

where k belongs to the set {1, . . . ,C} and R denotes the set {r_k}_{k∈{1, . . . ,C}}.

The LLRs are then input to an error-decoding unit 107 which generates estimates of the transmitted information sequence, {u_k}_{k∈(1, . . . K)}.

FIG. 2 illustrates a demodulation architecture that can be used to implement the demodulation unit 106 of FIG. 1. The demodulator architecture shown in FIG. 2 is derived from the Delayed Decision Feedback Sequence Estimation (DDFSE) approach presented in ‘Delayed decision-feedback sequence estimation’, A. Duel-Hallen, C. Heegard, IEEE Trans. on Communications, Vol. 37, Issue 5, May 1989. It will, however, be apparent to one skilled in the art that the demodulator architecture that will shortly be described in conjunction with FIG. 2 can also be extended to different equalisation approaches, of which an example is Reduced State Sequence Estimation (RSSE) as described in ‘Reduced-state sequence estimation for coded modulation of intersymbol interference channels’, M. V. Eyuboglu, S. U. H. Qureshi, IEEE Journals on Selected Areas in Communications, Volume: 7, Issue: 6, August 1989.

In order to facilitate the description of the demodulator architecture of FIG. 2, it is first useful to describe the computations involved in a demodulator architecture using a “full trellis” approach and then the computations involved in a DDFSE demodulator architecture. The different computations involved in “full trellis” and DDFSE Viterbi processing will not be detailed and only those computations that facilitate the description of the invention will be presented. For a more detailed description of the computations involved in a “full trellis” demodulator architecture, see, for example, ‘TCM on frequency-selective fading channels: a comparison of soft-output probabilistic equalizers’, P. Hoeher, GLOBECOM '90, December 1990. For a more detailed description of the computations involved in a DDFSE demodulator architecture, see, for example, ‘Delayed decision-feedback sequence estimation’, A. Duel-Hallen, C. Heegard, IEEE Trans. on Communications, Vol. 37, Issue 5, May 1989. The computations involved in a demodulator architecture using a “full trellis” approach will now be described.

In a “full trellis” Viterbi demodulator, the number of states in the trellis used to perform the Viterbi decoding, denoted S, is equal to 2^M×L. Moreover, any given state in the trellis has 2^M “merging states” leading into it.

When the symbol r_k is received, the metrics associated with the different states in the trellis are updated according to the equation:

${\gamma }_m\left(k+1\right)=\underset{i=0,\dots ,2^M-1}{\max }\left({\chi }_m^{k+1}\left(i\right)\right\}$ $0\le m\le S-1$

The candidate metrics X_m^k+1(i) are calculated using the state metrics from the previous iteration with the recursive equation:

X _m ^k+1(i)=γ_p(m,i)(k)−B(r_k, p(m,i),m)

Here, p(m,i) {0≦m≦S−1;0≦i≦2^M−1} is the index of the state leading into the state with index m in the path that corresponds to a hypothesis transmitted complex symbol c equal to M(D⁻¹(i)).

The branch metrics are calculated using the following equation:

$B\left(r_k,p\left(m,i\right),m\right)={r_k-\sum _{u=0}^{L-1}h_u{\tilde{c}}_{k-u}}^2$

As a “full trellis” approach is being used, the hypothesis samples {tilde over (c)}_k−u used in the above branch metric calculation are taken from the states p(m,i) and m that the branch metric interconnects. The candidate metrics X_m^k+1(i) are then used to calculate LLRs in a known manner. The computations involved in a demodulator architecture using a DDFSE approach will now be described.

In a DDFSE demodulator architecture, a reduction in the number of trellis states is achieved by splitting the channel into two sections. The first section of the channel (with length L_ƒ) is processed in a way similar to the full-state approach. However, for the remainder of the channel (with length L_r such that L=L_ƒ+L_r), the modulated symbols derived from previous decisions are used rather than testing all the possible hypotheses. Using this method, the number of states for which trellis processing is performed is reduced from 2^M×(L−1) to 2^M×(Lƒ−1). To reflect this change in the number of states, the branch metric computations need to be modified as follows:

$B\left(r_k,p\left(m,i\right),m\right)={r_k-\sum _{u=0}^{L_f-1}h_u{\tilde{c}}_{k-u}-\sum _{u=L_f}^{L_f+L_r-1}h_u{\hat{c}}_{k-u}\left(p\left(m,i\right)\right)}^2$

It can be seen from the above equation that the computations involving the first L_ƒ taps are identical to those performed in the full trellis approach. However, for the last L_r taps, the hypothesis symbols {tilde over (c)}_k−u are replaced by hard decision symbols ĉ_k−u(p(m,i)) . These hard decision symbols are generated using the decisions that are taken during the selection of candidate metrics to become new state metrics in accordance with the equation:

${\gamma }_m\left(k+1\right)=\underset{i=0,\dots ,2^M-1}{\max }\left\{{\chi }_m^{k+1}\left(i\right)\right\}$ $0\le m\le S-1$

This equation is also used in the “full trellis” approach. The hard decision symbols are associated with each state in the trellis and are updated as the state metrics are updated such that the selection involved in determining γ_m(k+1) determines ĉ_k(m).

The demodulator architecture of FIG. 2, which will now be described, is based on the DDFSE approach. The architectures shown in FIGS. 2 and 3 and FIGS. 2 and 5 can be implemented as software using a suitable data processor in conjunction with suitable memory resources.

The demodulator architecture of FIG. 2 achieves a significant reduction in complexity by separating the sequence of L channel coefficients {h_i}_{i∈{0, . . . ,L-31 1}} into two distinct sections in the same manner as the DDFSE technique. For the first L_ƒ channel coefficients, all possible hypotheses for the modulated symbols are tested. For the remaining L-L_ƒ=L_r channel coefficients, however, the different hypotheses are replaced for each state by a single value derived from previous decisions. By so doing, the number of states in the trellis, denoted as S, is reduced from 2^L−1 down to 2^Lƒ−1. It is to be noted, however, that this reduction in the number of states to be processed in the demodulation architecture of FIG. 2 can lead to a significant performance reduction. It is therefore important that prior to the trellis processing, the signal is conditioned properly. Such a signal conditioning technique is described in ‘Equalization Concepts for EDGE’, W. H. Gerstacker, R. Schober, IEEE Trans. on Wireless Communications, January 2002.

When the symbol r_k is received, the state metrics associated with the different states in the trellis are updated in the demodulation architecture of FIG. 2 according to the following equation:

${\gamma }_m\left(k+1\right)=\underset{i=0,\dots ,2^M-1}{\max }\left\{{\chi }_m^{k+1}\left(i\right)\right\}$ $0\le m\le S-1$

Note that, for a given modulation scheme M, there exists a one-to-one relationship, denoted H(m), between each state index m and the set of hypotheses {{tilde over (c)}_k, . . . , {tilde over (c)}_k−Lƒ+2} for the transmitted information symbols. H is a function that maps a value from the set {0, . . . , S−1} to a value in a set of L_r−1 complex numbers. The total number of candidate metrics in the trellis is equal to S×2^M.

It should also be noted that it is possible to add or subtract any value, constant over the different states, to this update without changing the soft decisions generated by the algorithm. Such an approach may be used when the proposed receiver architecture is implemented using fixed-point number representation in order to reduce the computational complexity.

The candidate metrics X_m^k+1(i) are calculated in the demodulation architecture of FIG. 2 using the state metrics from the previous iteration with the recursive equation:

X _m ^k+1=γ_p(m,i)−B(r_k, p(m,i),m)

Here, p(m,i) {0≦m≦S−1;0≦i≦2^M−1} is the index of the state leading into the state with index m along the path that corresponds to a hypothesis transmitted complex symbol {tilde over (c)} equal to M (D⁻¹(i)). The calculation of the branch metrics B will be described shortly. Each branch metric, and each candidate metric, in the trellis is associated with a unique set of hypotheses for the modulation symbols {{tilde over (c)}_k, . . . , {tilde over (c)}_k−Lƒ+1}. For a given received symbol index k, the L_ƒ symbols which are uniquely associated with the ending state with index m and the modulation symbol index i are denoted as ζ(m,i,δ)={tilde over (c)}_k−δ0≦δ≦L_ƒ−1

The branch metrics are calculated in the demodulation architecture of FIG. 2 using the equation:

$B\left(r_k,p\left(m,i\right),m\right)={r_k-\sum _{u=0}^{L_f-1}h_u{\tilde{c}}_{k-u}-\sum _{u=L_f}^{L_f+L_r-1}h_u{\hat{c}}_{k-u}\left(p\left(m,i\right)\right)}^2$

The modulation symbols ĉ_k−u(p(m,i)) are derived from previous decisions from the trellis processing. Their actual computation will be described in a latter part of this document.

Note that the branch metric described in the above equation can be calculated in other ways. It is possible to modify the computation of the branch metric, based on, for example, knowledge of, or an estimation of, the statistical distribution of noise affecting the transmitted block. It is also possible to change the subtraction in the equation for X_m^k+1(i) into an addition. This may be useful if the processor on which the demodulator architecture of FIG. 2 is being implemented contains dedicated ‘minimum finding’ instructions.

In the demodulation architecture of FIG. 2, estimates of the A-Posteriori probabilities, computed in the logarithmic domain, of the 2^M different modulation symbols are derived from the candidate metrics using the following equation:

${\xi }_i=\log \left(P\left(c_k=M\left(D^{-1}\left(i\right)\right)R\right)\right)\approx \underset{\underset{ϛ\left(m,j,L_f-1\right)=c}{\underset{0\le j\le 2^M-1}{0\le m\le S-1}}}{\max }\left({\chi }_m^{k+1}\left(j\right)\right)$

The set of computations specified by the above equation for deriving the symbol A-Posteriori probabilities is not optimum in terms of demodulation performance. However, it is significantly less complex than the optimum approach (as described in ‘Optimal decoding of linear codes for minimizing symbol error rate’, L. Bahl, J. Cocke, F. Jelinek, J. Raviv, IEEE Trans. on Information Theory, Volume: 20, March 1974) and the performance loss is kept to an acceptable level in most conditions.

Moreover, it should be noted that, in the above equation for ξ_i, only the maximum candidate metric is used to derive the symbol A-Posteriori probabilities. It is, however, possible for someone skilled in the art to modify the equation to include all the candidate metrics, rather than just the maximum one.

The LLRs for the set of transmitted bits {d_M×(k−L+1), . . . , d_{M×(k−L+1)+M−1}} can then be estimated in the demodulation architecture of FIG. 2 from the symbol A-Posteriori probabilities using the equation:

${\lambda }_{M\left(k-L+1\right)+b}=\underset{i\in \left\{0,\dots ,2^M-1D_b^{-1}\left(i\right)=1\right\}}{\max }\left({\xi }_i\right)-\underset{i\in \left\{0,\dots ,2^M-1D_b^{-1}\left(i\right)=0\right\}}{\max }\left({\xi }_i\right)$ $b\in \left\{0,\dots ,M-1\right\}$

It should be noted that the LLRs could be improved by grouping first all the metrics corresponding to a decision 1, then all the metrics corresponding to a decision 0 and calculating the difference. Such an approach, although introducing a further computational load, provides better estimates of the LLRs since all the candidate metrics are used rather than just the maximum values. Non-linear processing techniques can be used for the process of combining the LLRs into the two groups.

From the foregoing description, it can be seen that the different computations required for the derivation of the LLRs are very complex. The demodulation architecture of FIG. 2 utilises techniques that reduce the computational complexity of the demodulation process. This reduction in computational complexity is achieved mainly by pre-computing and storing some of the quantities required in the process of LLR calculation. By so doing, it is possible to avoid the duplicating some computations and thus achieve a more efficient receiver architecture.

The specific processes shown in FIG. 2 for implementing the above-described calculations for demodulating a received signal will now be described. The following description assumes that L_ƒ has been set to two. However, the proposed demodulator architecture can easily be extended to the cases where L_ƒ is greater than two.

As soon as the channel mapping {h_i}_{i∈{0, . . . ,L−1}} is known (or is estimated), a product memory 201 can be populated. This table contains all the possible values for the complex products:

h_k×c

where k assumes all possible values in the range 0≦k≦L−1 and c takes all possible values that are permitted for a modulation symbol in the modulation scheme M. For retrieval purposes, the product values h_k×c are indexed by the values k and i, where i is the modulation symbol index i which verifies M (D⁻¹(i))=c for the value of c participating in the product.

These products are calculated for each channel coefficient h_k (of which there are L) and each possible modulated symbol value c (of which there are 2^M). Hence, the modulated channel table contains 2^M×L complex entries.

The product memory 201 is calculated only once and it does not need to be re-computed for the update of the state metrics γ_m(k+1) upon the presentation of each successive received symbol r_k.

The processing of the received symbol r_k starts with the computation of symbol differences in unit 206. This involves the computation of all the complex symbol differences r_k−h₀×c for all the possible values that a modulated symbol c can assume. These complex values are then stored in a symbol difference memory 202. The symbol difference memory 202 is dimensioned to contain the 2^M complex symbol differences r_k−h₀×c. These computations and the population of the symbol difference memory 202 are performed once for each received symbol in the burst (it will be recalled that a burst structure is being assumed for the transmitted signal for the purposes of this embodiment).

A state metrics memory 204 is provided and contains the values of all S state metrics awaiting adaptation upon receipt of the next symbol of the transmitted burst. The state metrics memory 204 therefore contains S entries. The size of each entry in the state metrics memory 204 depends on the selected numerical representation that is used for, and the desired accuracy required for, the state metrics. The size of these entries can be derived as a trade-off between demodulation performance and memory requirements.

A decision memory 203 is provided and this contains the set of previous decisions ĉ_k−u(m) associated with each trellis state m. As there are S states and L_r previous decisions ĉ_k−u(m) associated with each trellis state, the decision memory 203 has to be dimensioned to hold S×L_r previous decisions. Rather than storing the actual sequence of previous modulation symbol decisions, it is more efficient to store the corresponding symbol indices i. For each modulated symbol c, the modulation symbol index i which verifies M(D⁻¹(i))=c is stored in the decision memory. If this approach is selected, only M bits are required to store each i value, so the decision memory 203 is only S values deep and L_r×M bits wide.

The values held in the product memory 201, the symbol difference memory 202, the decision memory 203 and the state metrics memory 204 are combined in unit 207 to calculate all possible S×2^M new candidate metrics. The computation of a given candidate metric first requires the calculation of the associated branch metric.

The branch metric computations are performed sequentially for each branch starting state, indexed by m, in sequence. For the branch starting state m, the contribution of the past decisions to the branch metrics is first calculated and stored:

$\mu =\sum _{u=L_f}^{L_f+L_r-1}h_u{\hat{c}}_{k-u}\left(m\right)$

It should be noted that even though there are 2^M branch metrics associated with each branch starting state m, the computation of the above equation only needs to be performed once for each branch starting state. Moreover, the value of μ is not required for the processing of the following states. Hence, for μ, memory provision only has to be made to store a single value as this value can be overwritten when progressing to the candidate metric calculations for the next value of m.

As indicated earlier, for memory efficiency reasons, it is preferred to store modulation symbol indices i rather than the complex modulated symbols ĉ_k−u(m) in the decision memory 203. Each product value h_k×c required for μ is retrieved from the product memory 201 by reading the location within that memory that is indexed by the values k and i of the product concerned. Hence, the computation of the complex value μ only requires L_r complex additions and L_r memory accesses (one memory access representing the retrieval of one complex value h_k×c).

Next, the contribution of the hypothesised symbols to the branch metrics is added to the partial sum μ such that μ becomes:

$\mu =\sum _{u=1}^{L_f-1}h_u{\tilde{c}}_{k-u}+\sum _{u=L_f}^{L_f+L_r-1}h_u{\hat{c}}_{k-u}\left(m\right)$

The values h_u×{tilde over (c)}_k−u for u=1 to L_ƒ−1 are retrieved from the product memory 201 (again by indexing the memory 201 using u and i) and are accumulated with μ on a recursive basis to achieve the result set out in the preceding equation. This update to μ involves L_ƒ−1 memory accesses and L_ƒ−1 complex addition operations.

This partial sum μ is then stored and used to calculate the different branch metrics associated with the branch starting state indexed by m.

In order to calculate the 2^M branch metrics associated with the current branch starting state, the partial sum μ is combined with each entry from the symbol difference memory 202 in turn. For each branch metric, the partial sum μ is subtracted from an entry from the symbol difference memory 202 to produce:

$r_k-\sum _{u=0}^{L_f-1}h_u{\tilde{c}}_{k-u}-\sum _{u=L_f}^{L_f+L_r-1}h_u{\hat{c}}_{k-u}\left(p\left(m,i\right)\right)$

The squared magnitude of this result is then computed and the result, a branch metric, is stored in the candidate metrics memory 205. To calculate the next branch metric for the branch starting state, the next value in the symbol difference memory is read out and μ is subtracted from it. The calculation of the branch metrics is particularly efficient since the entries in the symbol difference memory are simply read out from beginning to end for combination with μ.

The candidate metrics memory 205 contains all 2^M×S candidate metrics in the trellis although at this point, the entries in that memory merely contain branch metrics. The 2^M metrics associated with the branch starting state k are written into the locations k×2^M to (k+1)×2^M−1. The squared magnitudes associated with the branch starting state k are written sequentially.

The different computations performed for the derivation of the branch metrics are summarised in FIG. 3. The contribution to a given branch metric of the past modulation symbol decisions ĉk−u (u=L_ƒ to Lƒ+L_r−1) is first calculated in unit 301 using the past decisions from the memory 203 and values from the product memory 201. These computations are performed for each of the S trellis states.

The contribution to the branch metric of the hypothesised symbols {tilde over (c)}_k−u for u=1 to L_ƒ−1 associated with the branch starting state is then added in unit 302. The resulting quantity is then combined in unit 303 with the symbol difference from the memory 202 corresponding to the symbol decision implied by the branch in question. The result is then squared in unit 304 to calculate the branch metric. The computations described in this paragraph need to be carried out for each of the 2^M×S branch metrics.

As was indicated earlier, at this point, the values in the candidate metrics memory 205 do not correspond to candidate metrics since the contribution of the previous state metrics has not been included yet. Hence, each branch metric value is retrieved in turn from the memory and is then added to the state metric of the starting state of the trellis branch in question. It should be noted that each state metric is common to 2^M candidate metrics. After this addition process, each candidate metric value is stored in its original location in the candidate metrics memory 205. In other words, a Read-Modify-Write (RMW) process is performed for the 2^M candidate metrics corresponding to the branch starting state k.

The candidate metrics computations described in the preceding paragraph are repeated, for each branch starting state in turn, until the candidate metrics memory 205 has been populated in its entirety. Once this is done, the state metrics and the past decision symbols are updated by unit 208. The decision memory 203 and the state metrics memory 204 are updated as an output of this process, as will be described later.

In the candidate metrics calculation unit 207, candidate metrics were generated for each branch starting state in turn. This permits the re-use of the pre-computed values of memories 201 and 202, thus reducing the number of computations required for calculating the state metrics. However, in unit 208 that updates the state metrics, the processing is sequenced according to the branch ending, rather than the branch starting, states. Hence, the candidate metrics are not retrieved from the candidate metrics memory 205 for use by unit 208 with a fixed incremental offset of 1. Instead, a constant incremental offset equal to 2^M is used. Assume, for example, the branch ending state with index m is processed. In such a case, the values in the candidate metrics memory 205 at the locations m, m+2^M, . . . , m+S−2^M are retrieved by unit 208. As each state metric is retrieved, the best value (i.e. the minimum value) is kept. Once all 2^M metrics have been processed, the minimum value corresponds to the updated state metric γ_m(k+1).

Once the best candidate metric has been found for a given branch ending state α, the associated past decision symbol information ĉ_k−u(α) (u=L_ƒ to L_ƒ+L_r−1) is updated. This is achieved by taking the past decision symbols held in the memory 203 for the state β corresponding to the branch-starting state of the branch producing the best candidate metric and combining them with ĉ_k-Lƒ+1(β), i.e. where u=L_ƒ−1.

When the past decisions associated with a given state are stored as a single word of L_r×M bits, the update of the information in the decision memory can be implemented in very efficient manner. The new decision word can be calculated by first shifting by M bits the decision word of the branch starting state corresponding to the best candidate metric. The index i of the earliest (position u=L_ƒ−1) hypothesis symbol corresponding to this branch starting state is then combined with the decision word by using bit masking and bitwise OR operations.

Once the contents of the decision memory 203 and the state metrics memory 204 have been updated, the symbol A-Posteriori probabilities are calculated in unit 209. The 2^M symbol probabilities are then stored in a symbol probabilities memory 210. The symbol probability associated with the symbol c=M(D⁻¹(i)) is calculated by reading in turn the candidate metrics at the locations 2^M×i to 2^M×(i−1)+1 within memory 205 and finding the best value. This process is repeated for all 2^M possible modulation symbols. Because the candidate metrics are read from the start of the memory 205 till the end with an increment in memory location equal to one, it is possible to implement these memory accesses in a very efficient manner.

Finally, in the last stage 211 of the demodulator architecture of FIG. 2, the LLRs for the M bits associated with the current received symbol r_k are calculated. These values are calculated using the symbol probabilities stored in the symbol probabilities memory 210. The LLRs for the M bits are calculated one after the other, in no particular order. For the LLR of a given bit, the maximum (or minimum depending on the convention used for the branch metrics) of the 2^M−1 symbol probabilities corresponding to the modulation symbols with the bit set to 1 is calculated. The same process is repeated for the 2^M−1 probabilities corresponding to symbols with the bit set to 0. The difference between these two values is equal to the LLR of the given bit.

It should be noted that it is possible to combine the computations performed in unit 209 with those required for the derivation of the candidate metrics. Doing so would remove the need for the symbol probabilities memory 210. However, this would also mean that the best candidate metrics would have to be tracked as they were being computed.

In summary, a substantial reduction in computational complexity can be achieved by pre-computed and storing the following quantities

• • 2^M×L complex products of individual channel coefficients and individual modulation symbols
  • 2^M×S candidate metrics
  • 2^M complex differences between the current received symbol and all modulation symbols modulated by the first propagation channel tap ho
  • 2^M A-Posteriori symbol probabilities

The proposed demodulator architecture also requires storage for the S different state metrics and the S different sequences of previous modulation symbols. However, this memory requirement is present for any receiver using the DDFSE architecture. Moreover, it has been shown that the amount of memory required for the past decision symbols can be reduced if the modulation symbol indices are stored rather than the actual complex modulation symbols. Moreover, this storage approach makes the update of the ĉ_k−u (u=L_ƒ to Lƒ+L_r−1) information very easy when a new symbol is received. Finally, it should be noted that, depending on the processor used to implement the proposed demodulation approach, the size of the state metrics and past decision symbols memories may need to be doubled in order to avoid over-writing information before it has been completely used.

The initialisation of the contents of the decision memory 203 and of the state metrics memory 204 will depend on the actual communication systems in which the proposed demodulator architecture is being used. If a sequence of known training symbols is inserted as part of the transmitted information, it is possible to use this information to set the initial past decision words ĉ_k−u (u=L_ƒ to Lƒ+L_r−1). The state metrics can also be initialised to values which will bias following decisions toward these known symbols. For example in the E-GPRS system, a training sequence is inserted in the middle of the burst of transmitted information. It is then recommended to perform the equalisation of the received information in two sequences which start from the training sequence, as indicated in FIG. 4. The training bits can then be used to initialise the decision memory 203 and the state metrics memory 204.

It should also be noted that that the implementation of the derivation of the LLRs from the A-Posteriori probabilities may vary depending on the overall architecture. For example, if the receiver architecture is based on the RSSE approach, the processing of the A-Posteriori probabilities needs to be modified to take into account that the bit decisions can be made with different delays.

Another technique that can be used to further reduce the complexity of the demodulation computations set out above will now be described. This complexity reduction technique applies to the branch metric computations and is presented in FIG. 5. The reduction in complexity is achieved by splitting the L_r channel coefficients into two sets. For the first L_r^St coefficients {h_u}_{Lƒ≦u≦Lƒ+LrSt−1}, the contribution to the branch metrics of the associated modulation symbols are calculated for each branch starting state in unit 502. However, for the last L_r^Sy (where L_r^Sy+L_r^Sy=L_r) coefficients, {h_u}_{Lƒ+LrSt≦u≦Lƒ+LrSy−1}, the contribution to the branch metric of the modulation symbols is only calculated once in unit 501 with the result being used for all states. Hence, a complexity reduction of a factor of S is achieved for the computations associated with the last L_r^Sy channel taps.

The technique that was presented for the update of the past decision symbols ĉ_k−u (u=L_ƒ to Lƒ+L_r−1) is very efficient since it only requires bit shifting and bit setting operations to be performed. This technique, however, may not be optimum in terms of demodulation performance. Hence, an alternative approach will now be described.

For any given branch ending state, a new past decision word needs to be created from the 2^M old decision words of the 2^M starting branch starting states that have branches leading into the branch ending state. For each bit position in the new decision word, a bit is derived by gathering a set of bits consisting of the bits appearing at the same position in the group of 2^M old decision words. The state metrics of states whose past decision words contributed an even parity bit to the set are combined in a weighted sum and the state metrics of states whose past decision words contributed an odd parity bit to the set are likewise combined in a weighted sum. If the sum associated with the odd parity bit or bits is larger than the sum associated with the even parity bit or bits, then the bit at the present position in the decision word for the branch ending state is set to 1; otherwise, the bit at the present position in the decision word for the branch ending state is set to 0. Various techniques can be used for the computation of the weighted sums. In one possible implementation, the weights are derived from the probabilities associated with the different branch starting states. These probabilities can be derived from the state metrics.

Thus far, certain embodiments of the invention have been described in terms of architectures. As indicated earlier, these architectures can be implemented as software with suitable underlying hardware, such as that shown in FIG. 6.

FIG. 6 illustrates a generic structure that can represent a mobile telephone, base station or similar receiver in which the invention can be implemented. The receiver 601 comprises an antenna 602 for acquiring wireless signals, an RF section 603, an analogue-to-digital conversion (ADC) section 604, a data processor 605 and a memory 606. In practice, the receiver will contain many other elements but only those necessary for an explanation of the implementation of the invention have been shown.

Signals received at the antenna 602 are down-converted in frequency and amplified at RF section 603. The signals are then converted into digital signals by ADC section 604 and passed to processor 605. The processor 605 performs the operations necessary to extract and utilise the information payload of the acquired signals, relying on the memory 606 to provide storage of signal values and other data as required. The processor 605 undertakes the procedures of the architectures shown in FIGS. 2 and 3 and 2 and 5, utilising memory resource 606 as necessary to provide storage of data required by those procedures.

Claims

1. Trellis processing apparatus for estimating a transmitted signal from a received signal acquired through a channel, wherein the received signal comprises a sequence of received symbols, the transmitted signal comprises a sequence of transmitted symbols, the channel is characterised by a channel response estimate comprising a series of channel coefficients each corresponding to a different time delay and the apparatus comprises: a memory containing differences between a received symbol and different products, each product being a product of the channel coefficient corresponding to the least time delay and a different one of the possible transmitted symbols; a processor for calculating a summation of products, each product being a product of a possible transmitted signal symbol and a channel coefficient other than the one corresponding to the least time delay and the summation being specific to a branch-start state in a trellis; and wherein the processor is arranged to combine the result of the summation with different differences from the memorv means in the calculation of branch metrics leading to different branch-end states in the trellis.

2. Apparatus according to claim 1, wherein the memory contains products, each product being a product of a channel coefficient and a possible transmitted signal symbol, and the processor is arranged to employ at least one product from the memory in the calculation of the summation and/or one or more of the differences.

3. Trellis processing apparatus for calculating branch metrics in a reduced state trellis comprising a processor adapted to pre-calculate for a trellis state a sum of products of channel coefficient values and signal symbol values corresponding to at least the oldest part of the section of the channel that is not hypothesised by the state such that the sum can be re-used in calculating branch metrics for different trellis branches leading out of different states and further comprising a memorv for storing the products.

4. Apparatus according to claim 1 or 3, wherein the processor is arranged to compute candidate metrics for states in the trellis from the branch metrics that it has calculated.

5. A trellis processing method for estimating a transmitted signal from a received signal acquired through a channel, wherein the received signal comprises a sequence of received symbols, the transmitted signal comprises a sequence of transmitted symbols, the channel is characterised by a channel response estimate comprising a series of channel coefficients each corresponding to a different time delay and the method comprises: storing a group of differences between a received symbol and different products, each product being a product of the channel coefficient corresponding to the least time delay and a different one of the possible transmitted symbols; a calculating step comprising calculating a summation of products, each product being a product of a possible transmitted signal symbol and a channel coefficient other than the one corresponding to the least time delay and the summation being specific to a branch-start state in a trellis; and wherein the calculating step further comprises combining the result of the summation with different differences from said stored group in the calculation of branch metrics leading to different branch-end states in the trellis.

6. A method according to claim 5, further comprising storing a set of products, each product being a product of a channel coefficient and a possible transmitted signal symbol, and wherein the calculating step is arranged to employ at least one product from the set in the calculation of the summation and/or one or more of the differences.

7. A trellis processing method for calculating branch metrics in a reduced state trellis comprising precalculating for a trellis state a sum of products of channel coefficient values and signal symbol values corresponding to at least the oldest part of the section of the channel that is not hypothesised by the state such that the sum can be re-used in calculating branch metrics for different trellis branches leading out of different states and storing the products in a memory.

8. A method according to claim 5 or 7, wherein the calculating step further comprises computing candidate metrics for states in the trellis from the branch metrics that it has calculated.

9. A program for causing data processing apparatus to perform a method according to claim 5 or 7.