1: \begin{abstract}

2: \textcolor{black}{Considering backhaul consumption in practical systems, it may not be the best choice to engage all the time in full cooperative MIMO  for interference mitigation.} 

3: In this paper, we  propose a novel downlink partial cooperative MIMO

4: (Pco-MIMO) physical layer   (PHY) scheme, which allows flexible tradeoff between

5: the {\em partial data cooperation} level and the {\em backhaul

6: consumption}. Based on this Pco-MIMO scheme,  we consider  \textcolor{black}{dynamic} transmit power and  rate  allocation according to the \textcolor{black}{imperfect} channel state information at transmitters (CSIT) and  the queue state information (QSI) to minimize the average delay \textcolor{black}{cost} 

7: subject to \textcolor{black}{average} backhaul consumption constraints and average power

8: constraints. The delay-optimal control problem is formulated as an

9: infinite horizon average cost constrained partially observed Markov

10: decision process (CPOMDP). \textcolor{black}{By} exploiting the special structure in our

11: problem, we  derive an {\em equivalent Bellman Equation} to solve

12: the CPOMDP. \textcolor{black}{To reduce computational complexity} and facilitate distributed implementation, we propose

13: a distributed online learning algorithm to estimate the per-flow

14: potential functions and Lagrange multipliers (LMs) and

15: a distributed online stochastic partial  gradient algorithm to

16: obtain the power and rate control policy.  The proposed \textcolor{black}{low-complexity distributed} solution is 

17: based on local observations of the system states at the BSs and is very robust against model variations.  We also  prove the convergence and the asymptotic optimality of the proposed solution.

18: \end{abstract}

19: