15-399, 80-317/617 Constructive Logic
Overview

The material in this class can be roughly divided into three parts.

  1. Constructive Logic
  2. Proofs and Programs
  3. Decidable Fragments

We review the contents of each part in turn.

1. Constructive Logic

The underlying principle of our development of logic (and later type theory) is the separation of judgments from propositions. A judgment is an object of knowledge, while an evident judgment is something we know. Examples of judgments are ``A is a proposition'' and ``A is true''. The meaning of a proposition is determined by what counts as a verification of it. These consideration lead us to introduction and elimination rules for logical connectives, plus the notion of local reduction which verifies that the elimination rules are sound with respect to the introduction rules.

For the purpose of guiding proof search we further introduce the notion of normal deduction. Roughly, a proof is normal if it can be found by using only introduction rules from below (from the goal) and only elimination rules from above (the assumptions). Normal proofs are complete for the propositional fragment (every true proposition has a normal proof). As an application, we can show the independence of certain schematic propositions such as the so-called law of excluded middle.

A common alternative meaning explanation is via notational definitions. We use this, for example, to define negation and equivalence of propositions.

As a final, purely logical concept we consider universal and existential quantification over elements of an unspecified domain. This yields first-order intuitionistic logic. If we also assume the law of excluded middle (which cannot be justified using our machinery of judgments), then we obtain classical logic in which every proposition is assumed to be true or false. This destroys the constructive nature of proofs which is subject of the second part of the course.

2. Proofs and Programs

We introduce proof terms to directly express the evidence for the truth of a proposition. Proofs terms constitute a small functional programming languages, in which propositions play the role of types. This correspondence is known as the Curry-Howard isomorphism between proofs and programs. The basis for computation are local reductions, expressed as reduction rules on proof terms.

Under this interpretation, implication corresponds to function types, conjunction to product types, and disjunction to disjoint sum types. Thus we have strong logical guidance in the design of programming languages and type systems, which is the first major application of logic in computer science we consider.

To write interesting programs we need specific data types, such as Booleans, natural numbers, and lists. These can be introduced via introduction and elimination rules, following precisely the same methodology as in our development of logic. Elimination rules takes the form of schemas for definition by cases or primitive recursion (elimination on types) and proof by cases or induction (elimination on propositions).

We briefly investigate arithmetic (the theory of natural numbers), but then generalize to other structure data types and discuss various programming concepts. This includes reasoning about data representation (exemplified by an implementation of numbers as bit strings), general recursion (related to complete induction), and reasoning about data structure invariants (captured by dependent types).

With dependent types we have essentially replicated the full structure of propositions at the level of types. In order to avoid excessive redundancy, those concepts are therefore often collapsed, which can be justified by the correspondence between proofs and programs. In general, we see that proofs concerning data types contain computationally uninteresting part, and we show how to consistently eliminating those parts, thereby contracting proofs to programs.

3. Decidable Fragments

There is no terminating algorithm which can decide whether a given proposition of first-order logic or arithmetic is true, even if we restrict ourselves to a certain formal system. This did not directly prohibit our applications in programming languages, since the problem of proof checking (or the related type-checking) is decidable by design. However, it complicates other applications of logic where problems are translated into logic to be solved by theorem proving.

We therefore conclude the course by considering two interesting decidable fragments that have found numerous practical applications.

The first is the theory of Booleans. The elimination rule here is not induction, but simply proof by cases. The theory of Booleans with equality, where AND, OR, NOT, FORALL, EXISTS are given directly by notational definitions, turns out to be a decidable fragment of type theory. This includes in particular the problem SAT (Boolean satisfiability) which is NP-complete. Nonetheless, many practical problems can be mapped to SAT and solved efficiently, such as k-coloring of graphs, state-space planning, or Hamiltonian circuits. The straightforward implementation in type theory, however, is too inefficient so we introduce ordered binary decision diagrams (OBDDs) that are frequently used in practice.

The second is the theory of temporal properties of finite state systems. Part of this theory can be developed easily in type theory, but for the full computation tree logic (CTL) it is an open question how to accomplish this effectively from first principles. Again, under the assumption of excluded middle we develop a model-checking algorithm that decides, for a given formula and finite state system (= model), if the formula is true on the model. We considered this for the temporal operators X (next), G (globally), F (eventually), and U (until) each considered along all (A) or some (E) computation paths. The model-checking algorithm can be implemented by using OBDDs, representing both the states and transition relation by Boolean functions.


[ Home | Schedule | Assignments | Handouts | Software | Overview ]

awodey@cmu.edu
Steve Awodey