Constellation Shaping using LDPC Codes

It is well-known that a Gaussian source distribution is required for maximum information transfer across a Gaussian channel. In a coded modulation system an equiprobable symbol constellation loses at most 1.53 dB when compared to a Gaussian source. To bridge this shaping gap, a code can be used to make the source distribution more Gaussian over an expanded constellation that results in lower average transmitted energy. Trellis shaping uses convolutional codes and the Viterbi algorithm for minimizing the transmitted energy. In this work, we propose trellis shaping using low-density parity-check codes as the shaping codes. We show that the 2-state min-sum algorithm over the Tanner graph can be used to efficiently implement the energy minimization. This is a more than 4-fold decrease in complexity over 4-state convolutional code-based trellis shaping. Using one of our simple shaping codes, we have observed a shaping gain of up to 0.65 dB (with CER = 1.26; PAPR = 3.86) (as compared with CER=1.41 and PAPR=3.3 for convolutional-code based trellis shaping with similar shaping gain). This encouraging result indicates that more complex LDPC-based approaches will do even better. We also present simulation results to show that constellation shaping provides similar gains over wireless channels under slow fading conditions.