Variable-Radix Coding of the Reals

Recently proposed real number systems like Posits and Elias codes make use of tapered accuracy resulting from variable-length coding of exponents and significands. Several quite different interpretations of these number systems have been provided, though most often these rely on some combination of fixed- and variable-length codes for exponent and significand. We provide a new perspective on these number systems that unifies known representations while suggesting new ones. Our framework is based on multibit radix representations that encode the exponent in unary, the leading nonzero digit in a variablelength code, and the remaining digits in fixed-length binary code. We show how Posits, the various Elias codes, and ieee 754 like representations can be expressed in this framework. Moreover, we show that Posits and the Elias γ and δ codes represent the leading digit using the canonical Huffman code for a probability distribution given by Benford's law, which governs the probability of leading digits. We further show that Posits correspond to the use of a fixed radix while Elias δ and ω codes are based on simple sequences of increasing radix. Our approach provides for an intuitive and uniform framework for representing numbers that reveals a visual mapping between codewords and the binary representation of real numbers obscured by prior frameworks. This new interpretation suggests a generalization of Posits and other number systems and provides simple rules for designing information-theoretically optimal codes.