Topology estimation is a critical problem in distribution grids that affect real-time control and optimization of grid operations. It is however hindered by the limited presence of real-time meters at the grid nodes and lines. This paper discusses a framework to learn the radial topology in distribution grids using limited nodal meters and no line meters. The most significant feature of the learning framework is that it is able to provably learn the topology as long as two hidden nodes (without meters) are not adjacency to one another in the operational grid. Our learning problem allows for greater number of hidden nodes than known prior work in this area. Further the algorithm does not require historical data for the hidden nodes and estimates their injection covariances. The learning algorithm uses ordered trends as well as equality constraints in nodal voltage fluctuations that arise from the radial topology. The efficiency of the designed algorithm is discussed by presenting simulation results for topology recovery in test radial grids.