In this paper, we examine the fundamental trade-off between radiated power and achieved throughput in wireless multi-carrier, multiple-input and multiple-output (MIMO) systems that vary with time in an unpredictable fashion (e.g. due to changes in the wireless medium or the users' QoS requirements). Contrary to the static/stationary channel regime, there is no optimal power allocation profile to target (either static or in the mean), so the system's users must adapt to changes in the environment "on the fly", without being able to predict the system's evolution ahead of time. In this dynamic context, we formulate the users' power/throughput trade-off as an online optimization problem and we provide a matrix exponential learning algorithm that leads to no regret - i.e. the proposed transmit policy is asymptotically optimal in hindsight, irrespective of how the system evolves over time. Furthermore, we also examine the robustness of the proposed algorithm under imperfect channel state information (CSI) and we show that it retains its regret minimization properties under very mild conditions on the measurement noise statistics. As a result, users are able to track the evolution of their individually optimum transmit profiles remarkably well, even under rapidly changing network conditions and high uncertainty. Our theoretical analysis is validated by extensive numerical simulations corresponding to a realistic network deployment and providing further insights in the practical implementation aspects of the proposed algorithm.