Lossless coding for sources of floating-point and fixed-precision numbers

A procedure for constructing binary models for sources of floating-point and fixed-precision numbers is given and used together with the minimum description length principle to find optimal models. These models are used to design noiseless source codes which take into consideration the practical tradeoff between minimizing memory cost and expected code length. As an example, the authors show the impossibility of compressing the Gaussian source of floating-point numbers beyond a certain point regardless of the variance