Similarity measuring is a key problem for 3D model retrieval. In this paper, we propose a novel shape descriptor "Thickness Histogram" (TH) by uniformly estimating thickness of a model using statistical methods. It is translation and rotation-invariant, discriminative to different shapes, and very efficient to compute with the Shape Distribution (SD) proposed by Osada etc. For high performance of the retrieval, we propose a robust method for translating the directional form of the statistical distribution to the harmonic representation. By summing up energies at different frequencies, a matrix shape signature is formed to provide an exhaustive characterization of 3D geometry. Experiments show that the performance of the statistical harmonic representation is among the top ones of existing shape descriptors.