gwyn@BRL.ARPA (VLD/VMB) (10/08/85)
Several years ago, there was a very nice article on variable- precision floating point representation. I think it was by Leslie Lamport and published in CACM, but my memory is fuzzy. The idea was, with a given number of bits, as the exponent became larger (in absolute value), bits were stolen from the mantissa. This gave maximum significance around 1.0 and permitted a much wider dynamic range than usual. One might think that arithmetic on such critters would be horribly complicated, but in fact it was not much worse than usual and perhaps even somewhat better for addition/subtraction.