Error analysis of linear recursions in floating point

A stochastic error model for floating point arithmetic is developed and used to study the effects of finite wordlength on linear recursive formulas. Optimal realizations exist, but they are highly sensitive.