In this paper we present two theorems concerning relationships between degrees of unsolvability of recursively enumerable sets and their complexity properties.

The first theorem asserts that there are nonspeedable recursively enumerable sets in every recursively enumerable Turing degree. This theorem disproves the conjecture that all Turing complete sets are speedable, which arose from the fact that a rather inclusive subclass of the Turing complete sets, namely, the subcreative sets, consists solely of effectively speedable sets [2]. Furthermore, the natural construction to produce a nonspeedable set seems to lower the degree of the resulting set.

The second theorem says that every speedable set has jump strictly above the jump of the recursive sets. This theorem is an expected one in view of the fact that all sets which are known to be speedable jump to the double jump of the recursive sets [4].

After this paper was written, R. Soare [8] found a very useful characterization of the speedable sets which greatly facilitated the proofs of the results presented here. In addition his characterization implies that an r.e. degree a contains a speed-able set iff a′ > 0′.

Several canonical partition theorems are obtained, including a simultaneous generalization of Neumer's lemma and the Erdös-Rado theorem. The canonical partition relation for infinite cardinals is completely determined, answering a question of Erdös and Rado. Counterexamples are given showing that in several ways these results cannot be improved.

Much recent work in the theory of computational complexity ([Me], [FR]. [SI]) is concerned with establishing “the complexity” of various recursive functions, as measured by the time or space requirements of Turing machines which compute them. In the above work, we also observe another phenomenon: knowing the values of certain functions makes certain other functions easier to compute than they would be without this knowledge. We could say that the auxiliary functions “help” the computation of the other functions.

For example, we may conjecture that the “polynomial-complete” problems of Cook [C] and Karp [K] and Stockmeyer [S2], such as satisfiability of propositional formulas or 3-colorability of planar graphs, in fact require time proportional to nlog2n to be computed on a deterministic Turing machine. Then since the time required to decide if a planar graph with n nodes is 3-colorable can be lowered to a polynomial in n if we have a precomputed table of the satisfiable formulas in the propositional calculus, it is natural to say that the satisfiability problem “helps” the computation of the answers to the 3-coloring problem. Similar remarks may be made for any pair of polynomial-complete problems.

As a further illustration, Meyer and Stockmeyer [MS] have shown that, for a certain alphabet Σ, recognition of the set of regular expressions with squaring which are equivalent to Σ* requires Turing machine space cn for some constant c, on an infinite set of arguments. We also know that this set of regular expressions, which we call RSQ, may actually be recognized in space dn for some other constant d. Theorem 6.2 in [LMF] implies that there is some problem (not necessarily an interesting one) of complexity approximately equal to that of RSQ, which does not reduce the complexity of RSQ below cn . It does not “help” the computation of RSQ.

Let L be a first order logic and the infinitary logic (as described in [K, p. 6] over L. Suslin logic is obtained from by adjoining new propositional operators and . Let f range over elements of ωω and n range over elements of ω. Seq is the set of all finite sequences of elements of ω. If θ: Seq → is a mapping into formulas of then and are formulas of LA . If is a structure in which we can interpret and h is an -assignment then we extend the notion of satisfaction from to by defining

where f ∣ n is the finite sequence consisting of the first n values of f. We assume that has ω symbols for relations, functions, constants, and ω1 variables. θ is valid if θ ⊧ [h] for every h and is valid if -valid for every . We address ourselves to the problem of finding syntactical rules (or nearly so) which characterize validity .