In this talk we consider learning and optimizing a rank-2 convex quadratic function over N variables, each taking K discrete values on the unit circle. This problem arises from optimal design of a passive beamformer for intelligent reflecting surface (IRS) in order to maximize the overall channel strength. When K=2 and the quadratic function (or channel state information) is known, we propose a linear time algorithm that is capable of reaching a globally optimal solution of the problem. When the quadratic function is unknown (i.e. in the absence of channel state information) we propose a random max sampling (RMS) method and a conditional sample mean (CSM) method to maximize the quadratic function. We show that RMS method can provide a SNR boost that is linear in N (the number of reflective elements in IRS)，while the CSM can boost the SNR quadratically (in N), all with polynomial number of samples. Field trial results demonstrate the effectiveness of the proposed CSM method in the commercial 5G communication networks, providing >10dB SNR gain in both typical indoor and outdoor scenarios, and with no need to modify the current communication protocols and design.