4.7 Robust Group -Blind Multiuser DetectionConsider the received signal of (4.97). As noted in Chapter 3, in group-blind multiuser detection, only a subset of the K users' signals need to be demodulated. Specifically, suppose that the first ( K ) users are the users of interest. Denote and as matrices containing, respectively, the first and the last ( K “ ) columns of S . Similarly define the quantities , , and . Then (4.97) can be rewritten as Equation 4.123
Equation 4.124
Let the autocorrelation matrix of the received signal and its eigendecomposition be Equation 4.125
We next consider the problem of nonlinear group-blind multiuser detection in non-Gaussian noise. Denote and . Then (4.124) can be written as Equation 4.126
Equation 4.127
for some . The basic idea here is to get an estimate of the sum of the undesired users signals, , and to subtract it from r . This effectively reduces the problem to the form treated in Section 4.6. To that end, denote Equation 4.128
Equation 4.129
Next, we outline the method for estimating the signal in (4.127). Denote Equation 4.130
In what follows we assume that the Huber penalty function is used. We first obtain a robust estimate of f by the following iterative procedure: Equation 4.131
Equation 4.132
The robust estimate of f translates into a robust estimate of z , which by Proposition 4.2, in turn translates into a robust estimate of , as Equation 4.133
Using the above-estimated , the desired users' signals are then subtracted from the received signal to obtain Equation 4.134
Next, the subspace components of the undesired users' signals are identified as follows. Let Equation 4.135
where the dimension of the signal subspace in (4.135) is K “ . We then have Equation 4.136
Equation 4.137
for some or in its real-valued form, Equation 4.138
Equation 4.139
A robust estimate of is then obtained from (4.139) using an iterative procedure similar to (4.131)-(4.132). Finally, the estimated undesired users' signals are subtracted from the received signal to obtain Equation 4.140
Equation 4.141
Note that in order to form , the complex amplitudes of the desired users, , must be estimated, which can be done based on , as discussed in Section 3.4. Note also that such an estimate has a phase ambiguity of p , which necessitates differential encoding and decoding of data. The signal model (4.141) is the same as the one treated in Section 4.6. Accordingly, define the following cost function based on the Huber penalty function: Equation 4.142
where denotes the j th row of the matrix . Let be the stationary point of which can also be found using an iterative method similar to (4.105)-(4.106). The Hessian of at the stationary point is given by Equation 4.143
with Equation 4.144
The estimate of the desired users' bits based on the slowest-descent search is now given by [let sign ( )] Equation 4.145
with Equation 4.146
The robust group-blind multiuser detection algorithm for synchronous CDMA with non-Gaussian noise is summarized below. Algorithm 4.5: [Robust group-blind multiuser detector ”synchronous CDMA]
Simulation ExamplesWe consider a synchronous CDMA system with a processing gain N = 15, the number of users K = 8, and random phase offset and equal amplitudes of user signals. The number of desired users is = 4. Only the spreading waveforms of the desired users are assumed to be known to the receiver. The noise parameters are = 0.01, k = 100. The BER curves of the robust blind detector of Section 4.5 (Algorithm 4.2) and the slowest-descent-search robust group-blind detector with Q = 1 and Q = 2 are shown in Fig. 4.13. It is seen that significant performance improvement is offered by the robust group-blind local-search-based multiuser detector in non-Gaussian noise channels over the (nonlinear) blind robust detector discussed in Section 4.5. Figure 4.13. BER performance of a slowest-descent-based group-blind multiuser detector in non-Gaussian noise: synchronous case. N = 15, K = 8, = 4, = 0.01, k = 100.
|