We propose a theoretical framework for adding assignments and dynamic data to functional languages without violating their semantic properties. This differs from semi-functional languages like Scheme and ML in that values of expressions remain static and side-effect-free. A new form of abstraction called observer is designed to encapsulate state-oriented computation from the remaining purely applicative computation. The type system ensures that observers are combined linearly, allowing an implementation in terms of a global store. The utility of this extension is in manipulating shared dynamic data embedded in data structures. Evaluation of well-typed programs is Church-Rosser. Thus, programs produce the same results whether an eager or lazy evaluation order is used (assuming termination). A simple, sound logic permits reasoning about well-typed programs. The benefits of this work include greater expressive power and efficiency (compared to applicative languages), while retaining simplicity of reasoning.

Functional programming languages have traditionally lacked complete, flexible, and yet referentially transparent I/O mechanisms. Previous proposals for I/O have used either the notion of lazy streams or continuations to model interaction with the external world. We discuss and generalize these models and introduce a third, which we call the systems model, to perform I/O. The expressiveness of the styles are compared by means of an example. We then give a series of surprisingly simple translations between the three models, demonstrating that they are not as different as their programming styles suggest, and implying that the styles could be mixed within a single program. The need to express non-deterministic behavior in a functional language is well recognized. So is the problem of doing so without destroying referential transparency. We survey past approaches to this problem, and suggest a solution in the context of the I/O models described. The I/O system of the purely func...