This paper deals with the estimation of time-varying Multi-Carrier Direct-Sequence Code DivisionMultiple Access (MC-DS-CDMA) fading channels using a training-aided scheme. Our approach consists in using an optimal filtering based on a linear state-space model of the fading channel system. In that case, two issues have to be investigated: 1) what kind of optimal filtering can be used? 2) how to estimate the state-space matrices? Thus, Kalman filtering can be considered. It is optimal in the H2 sense providing the underlying state-space model is Gaussian and accurate. However, as these assumptions may no longer be satisfied in real cases, we propose to study the relevance of H∞ filtering. More particularly, when an explicit AR model is used for the channel, our first solution consists in estimating the fading channel and its AR parameters by means of two-cross-coupled H∞ filters. Instead of AR model based-estimators, our second contribution is to view the channel estimation as a realization issue where the state-space matrices are estimated by using subspace methods for system identification without any a priori explicit model for the channel.