Approximation errors and model reduction with an application in optical diffusion tomography