The traffic matrix (TM) is a chief input in many network design and planning applications. In this paper, we propose a model, called the spherically additive noise model (SANM). In conjunction with iterative proportional fitting (IPF), it enables fast generation of synthetic TMs around a predicted TM. We analyze SANM and IPF's action on the model to show theoretical guarantees on asymptotic convergence, in particular, convergence to the well-known gravity model.