Joint channel estimation and decoding using belief propagation
(BP) on factor graphs requires probability densities as messages
since continuous parameters are involved.
For coded OFDM on frequency-selective channels, we propose
to replace these densities by simpler messages where the channel
estimate is accurately modeled as a Gaussian mixture.
Upward messages in the graph include extrinsic information and
downward messages carry the parameters of the Gaussian mixture
for the channel estimate. The proposed BP performs as well as
expectation maximization (EM) for high pilot overhead, and it
is more stable than EM for low pilot overhead.