1. Presented by Rajatha Raghavendra
Large System Performance of Linear Multiuser Receivers in Multipath Fading Channels Authors – Jamie Evans & David Tse
Presented by Rajatha Raghavendra

2. Outline Multi user receivers Performance measures
Data estimator performance
Impact of channel estimation
Simulation results
Conclusions

3. Conventional receiver for CDMA
Matched filter - Correlation of received signal with all PN sequences.
Detection - Highest peak for autocorrelation.
But PN sequences are not fully orthogonal in practice.
Results in Multiple Access Interference(MAI).
Matched filter receiver – Has a bank of correlators which correlate the received signal with the signature PN sequence of all users. Highest SNR obtained (autocorrelation) is considered as actual received signal. Correlation of received signal with other sequences gives MAI.

4. Multi-User Detection receiver
Knowledge of other user’s channel and signature code helps in mitigating MAI at output of matched filter.
Types of linear receivers:
Decorrelator – requires signature sequence. Applies inverse of correlation to output of matched filter.
LMMSE - requires channel knowledge. Minimizes the error between estimated data and actual data with the help of training sequences.
Focus on Linear receivers - linear transformation on the output of matched receivers.

5. Block diagram of M.U.D.
Data estimator is the heart of the MUD.
Data estimator estimates the data of each user by observing the received data over one symbol period.
Needs channel estimates which are time-varying due to multipath fading.

6. Performance measure of MUD
SIR is a measure of performance.
SIR for random signature sequence is random.
David Tse – asymptotically, for large number of users, SIR converges to a deterministic quantity.
Extension – Channel has multipath fading components.
Only channel estimates(mean & covariance) are known.
In this paper, perfect knowledge of channel cannot be obtained due to time varying channel. So SIR is a function of mean & covariance of the channel.

7. Concept of Effective Interference
System with K users, N spreading gain, ak received power
where
where
- Effective interference of k users on user1
For estimated channel
where
- The estimated channel gain of user k
- The error variance

8. Data Estimator performance
For a multipath fading channel with L resolvable paths
where
Interference looks like (L-1) users with power and one user with power

9. Data Estimator performance
Overall interference caused by user k
When channel is known perfectly, then the interferer looks like a single interferer with power
When no channel knowledge is available, the interferer looks like L interferers with power

10. Data Estimator performance
One high power interferer is weaker than several low powered interferers with same total power.
Therefore channel estimation is an important factor in improving the performance.
Uncertainty results in single interferer becoming L dimensional.

11. Channel Estimation Performed during training sequences.
Estimation window size is less than coherence time.
Mean Square Error
where
As estimation window length increases, is approximated to which is the same as absence of other users.

12. Simulation Results
Eq 12
Asymptotically, normalized SIR converges to the theoretical value of 0.38
K/N = 0.5
N= 32, 64, 128, 256

13. Simulation Results
Ideal – channel is known perfectly
Worst case – channel is not known
Ideal LMMSE (o), worst case LMMSE (+), Decorrelator (x), and matched filter (*)
Ideal LMMSE & worst case LMMSE performance is almost the same in frequency flat fading channel.

14. Simulation Results Results are shown for Frequency Selective fading .
The matched filter (*), the Decorrelator (X), and the LMMSE receiver (o). Curves are shown for estimation window lengths of (from the top) infinity (perfectly known channel), 10, 2, and finally for the case when nothing is known about the channels.

15. Simulation Results
Plots of performance loss for the LMMSE receiver for Flat fading channel(L=1). Results are shown for channel estimator window lengths of (from the top) = 1, 2, and 5.

16. Conclusions
Asymptotic performance with random sequences is equal to the performance when the sequences are independent.
In multipath fading, the receivers making accurate channel estimates performs better than those without channel knowledge.
LMMSE performs better than decorrelator and matched filter.

17. THANK YOU!!