Adaptive Coded Modulation for Transmission over Fading Channels - Adaptive Signal Processing in Wireless Communications - page 84

at bit error rates of 10 -5 for average received SNRs under 20 dB. Thus, there is a significant
gain in system performance when transmitter CSI is available in uncoded systems.
However, as pointed out in [24, 27], the assumption of perfect CSI is dangerous when
channel estimates are outdated or noisy, as would be the case with realistic delay in the
feedback path from the receiver to the transmitter. In particular, the conditional den-
sity function given in equation (3.2) becomes Rician (rather than a delta function), and
hence the conditional channel acts like a fading channel. For example, for the example
described in section 3.3.2.1, adaptive signaling assuming perfect channel estimation can
miss its target bit error rate by two orders of magnitude—even for the relatively high
correlation coefficients of ρ = 0.96. In this case, bad predictions, which are relatively
uncommon, lead to instantaneous error rates that are orders of magnitude above the
target and thus dominate system performance. Using the design method of equation
(3.2) reveals that there are still significant gains in adaptive signaling versus nonadap-
tive signaling when transmitter CSI is not perfect—even when the correlation coeffi-
cient drops as low as ρ = 0.96. We conclude from this section that adaptive signaling is
particularly effective for simple, low-latency systems such as adaptive uncoded QAM
systems [24, 27].
3.3.3 Coded Modulation Structures
As discussed in section 3.2.3, the conditional channel given an outdated measurement
can vary from almost Rayleigh to almost AWGN—depending on both the channel
estimate and the mean square prediction error σ 2 . For small σ 2 , the conditional chan-
nel is nearly always Rician with a large noncentrality component [53, p. 811] (hence
approaching AWGN), whereas for large σ 2 , the channel often approaches Rayleigh. It is
well known that coded modulation structures optimized for AWGN channels [e.g., 63]
are not well matched to Rayleigh channels [58]. Thus, the characterization of the mean
square prediction error in a given system determines the types of coded modulation
structures to be employed.
For systems where the mean square prediction error is anticipated to be nearly zero,
coded modulation structures designed for the AWGN channel can be employed without
interleaving [31]. For systems with a moderate amount of channel prediction error, struc-
tures designed for a Rayleigh fading channel can have aspects of structures designed for
an AWGN channel embedded in them [27]. Finally, for adaptive systems where the mean
square prediction error is expected to be large, adaptive bit-interleaved coded modula-
tion (BICM) [50] is preferable. These three types of structures are presented below.
3.3.3.1 Coding Structures with (Nearly) Perfect Prediction
If the current channel fading X ( t ) is known accurately at the transmitter (i.e., σ 2 , the
prediction error of an MMSE predictor, is small), the effective channel given the out-
dated estimate is roughly AWGN. Thus, coding structures designed for AWGN chan-
nels [63] should be employed [32]. In particular, a base trellis-coded modulation scheme
[63] tuned to the average received SNR can be selected, and then uncoded bits can be
added or deleted based on the channel estimate. This structure is shown in Figure 3.5 . As