1 / after dispreading. Thus the signal-to-interference-plus-noise ratio becomes /( )
,
0 +
∑
=1 ≠= K
J k
k k
J J
J SF P N P
SINR . If the
spreading factor (SF) is moderate and the interfering users is large (>10), the MAI can be modeled as AWGN according to central limit theorem [33]. In the cellular environment, if the power of each user within a cell is not controlled appropriately, SINR of the user with small received power can be dramatically decreased by user with large received power, i.e., the near-far problem occurs. However, since the MAI are not noises, SINR can be increased if the MAI are removed. It has been shown in 1980s that the system capacity can be increased by an optimal detector with high complexity [80], and thus led to a new field known as multiuser detection (MUD).
1.2 Multiuser Detection
1.2.1 Optimum Multiuser Detection
The optimal multiuser detector derived by Verdu [80] can use either maximum a posteriori (MAP) detection or maximum likelihood (ML) sequence detection. For a K-user system, the complexity of the optimal detector is O(|A|K) where |A| is the alphabet size (two
for binary). In addition to high computational complexity, the detector required the knowledge of the noise variance through the channel, as well as the amplitudes at receiver end, the spreading codes, and timing of all K users. Although the detector is not feasible for practical use, it provides large capacity gain over conventional matched filter (MF).
Afterwards, a large number of researches have been triggered to find a sub-optimal receiver with lower complexity and less required information with slight sacrifice in performance [22], [54], [78], [79], [81].
1.2.2 Sub-Optimal Detection
There are many kinds of sub-optimum MUDs announced in the past decade, and these existing detectors can be categorized in many ways. Based on the implementation-oriented categorization, the detectors can be classified to centralized or non-centralized ones. Another way is to classify the detectors to linear or non-linear ones.
Centralized vs. Decentralized
The centralized detectors jointly detect each user’s data, while the decentralized detectors detect data of the user or users of interest according to the received signal composed of multiple users’ data. The decentralized detectors (single-user reception) require no spreading codes or received signal information of other users. The orthogonal filter is well known, which executes updating using the Minimum Mean-Squared Error (MMSE) algorithm so that the spreading code replica used for despreading would be orthogonal to the spreading code of signals of other users (including multipath signals). Although the orthogonal filter has an easier configuration than centralized detectors, it cannot be applied to scrambling codes that have much longer iteration period than the symbol length (long codes). In contrast, centralized detectors uses the reception signals and decoding data sequence of users to reduce the interference of other users in a mutually dependent manner. Generally speaking, the
centralized detectors are used in base stations, and the decentralized detectors can be used either in base stations or mobile receivers.
Linear vs. Nonlinear
The linear detectors perform linear filtering according to a specific criterion to suppress the MAI. There are two main kinds of linear MUD, named decorrelator and minimum mean square error (MMSE) detector [48], [49]. These suboptimum MUDs are similar to the zero-forcing and the MMSE equalizers used to combat inter-symbol interference in a single-user channel [1].
Although the linear detectors are easily to analyze, they restrict the system performance.
The non-linear detectors are known to achieve better performance in an iterative manner. The first kind of generalized non-linear detectors adopt linear detector as the pre-processing unit to suppress interference, and they suffer the same problems as linear detectors, i.e., matrix inversion must be performed. The second kind of generalized non-linear detectors, often referred to as interference cancellation (IC) [18], perform data decision using the output of matched filter without matrix inversion. The IC scheme tends to partially or fully remove the MAI terms at MF or RAKE outputs. A number of interference cancellation detectors have been proposed [22], [53], [54], [78], [90]. These detectors use soft or hard decisions to reconstruct interfering signal from part or all of the interfering users and subtracted the interfering signal from the received signal. The user of interest can then expect detection without or with only part of MAI if all decisions of interfering users are correct. Otherwise, error decisions of these interferers contribute double interfering signal. IC can be classified into three categories: parallel IC (PIC), successive IC (SIC), and hybrid IC (HIC). These ICs have tradeoff on computational complexity, processing delay and error performance.
The SIC detects user data and cancels multiple access interference in a serial manner. The performance of SIC is influenced by the cancellation order. The PIC simultaneously
processes all K users, canceling their interference after they have all been decoded independently [22], [78], [90], [92]. To alleviate performance saturation due to poor estimated interference in early stages, the partial PIC [22] is often used to partially cancel the estimated MAI from early stages. SIC also tends to remove partial interference from the decision statistics. The difference between SIC and PPIC is that PPIC removes partial interference from all users while SIC removes interference from users that are more reliable than the desired one. The PIC takes the advantage of having lower latency than the SIC at the sacrifice of adding more computational complexity. If there are K users in the system, the computational complexity of PIC are proportional to SK where S is the stage of PIC and that of SIC is proportional to K. As for latency, it is proportional to S for PIC and K for SIC. In addition to simplicity, the SIC performs better and is more robust than PIC [23], [53].
However, there are several drawbacks that prevent SIC becoming a widely used technique [6]: (1) the total decoding time of SIC increases linearly with the number of users;
(2) it is thought that the optimum power control for SIC are far more complicated than the equal power control for conventional receiver or PIC: although SIC with controlled user power distribution [6], [16], [42] can reduce the other-cell interference [5], [31] and increase system capacity [19], [87]; (3) the SIC is sensitive to estimation error due to error propagation.
Several techniques have been proposed to solve these drawbacks [2], [7], [17], [35], [68]. Pipelined [35], [68] and hybrid [40], [89] techniques are utilized to combat the problem of long latency of SIC when K is large. The HIC attempts to compromise the characteristics of SIC and PIC, i.e., all users in the system are separated into several parts, a part of users are detected in parallel, removed from the received signal, and then another part of users is detected in parallel. In [89], the so-called groupwise serial interference cancellation (GSIC) is the hybrid version of PIC and SIC which perform interference
cancellation in systems with variable SF. Recently, a frame-error-rate based outer-loop power control is shown to be applicable to SIC [17]. In [7], a simple iterative algorithm to achieve the optimal power control distribution is given. In [2], the authors show that power control can be done in commercial CDMA without modification. Recent work shows that SIC is a practical IC where its simplified version is employed as part of a commercial device to increase the cdma2000 EV-DO Rev A reverse link voice over IP (VoIP) capacity by about 15 percent [37].