1: \begin{abstract}

2: The problem of modulation classification for a multiple-antenna (MIMO)

3: system employing orthogonal frequency division multiplexing (OFDM)

4: is investigated under the assumption of unknown frequency-selective

5: fading channels and signal-to-noise ratio (SNR). The classification

6: problem is formulated as a Bayesian inference task, and solutions

7: are proposed based on Gibbs sampling and mean field variational inference.

8: The proposed methods rely on a selection of the prior distributions

9: that adopts a latent Dirichlet model for the modulation type and on

10: the Bayesian network formalism. The Gibbs sampling method converges

11: to the optimal Bayesian solution and, using numerical results, its

12: accuracy is seen to improve for small sample sizes when switching

13: to the mean field variational inference technique after a number of

14: iterations. The speed of convergence is shown to improve via annealing

15: and random restarts. While most of the literature on modulation classification

16: assume that the channels are flat fading, that the number of receive

17: antennas is no less than that of transmit antennas, and that a large

18: number of observed data symbols are available, the proposed methods

19: perform well under more general conditions. Finally, the proposed

20: Bayesian methods are demonstrated to improve over existing non-Bayesian

21: approaches based on independent component analysis and on prior Bayesian

22: methods based on the `superconstellation' method. \end{abstract}