370 likes | 946 Views
Decidability. A decision problem is a problem for which the answer is either "yes" or "no". A decidable problem is a decision problem that has a solution algorithm. A semidecidable problem is a decision problem for which there is a decision procedure
E N D
Decidability • A decision problem is a problem for which the answer is either "yes" or "no". • A decidable problem is a decision problem that has a solution algorithm. • A semidecidable problem is a decision problem for which there is a decision procedure • that correctly answers "yes" if the answer is "yes", • and never answers "yes" if the answer is "no".
Decidability and semidecidability • Decidable problems correspond to recursive languages; semidecidable problems correspond to r.e. languages. • since we’re assuming algorithms halt on all input • All decidable problems are semidecidable • so all recursive languages are r.e. • Semidecidable problems needn't be decidable • there may be no decision procedure that halts on all inputs
The universal problem • The universal problem asks whether M accepts w, given a TM M and string w • here w is a string over {0,1}, which is also M's input alphabet • We've seen this problem to be semidecidable • so Lu = {<M,w> | M accepts w} is an r.e.language, if encoded as described above • The universal problem is undecidable • so undecidable problems exist!!!!!
Undecidable problems • To show that undecidable problems exist, it’s easiest to look at a special case of the universal problem • Just as we can number TMs, we can also number strings over {0,1} • One mapping between i and wi converts i to binary and ignores the leading digit • every bit string arises in this way from some i • Now we can ask whether Mi accepts wi
A non-r.e. language • If it were decidable whether M accepts w, it would be decidable whether Mi accepts wi • And it would be decidable, given i, whether Mi fails to accept wi • given i, we could construct Mi and wi • and then decide whether Mi accepts wi • So let Ld = { wi | Mi doesn't accept wi} • Ld (the diagonal language) can’t even be r.e. • the intuition is that of Linz's Figure 11.2, p. 280
Showing that Ld is not r.e. • To be r.e., Ld must be accepted by some TM Mk. • We may ask whether Mk accepts wk • If so, then wk is in not in Ld, by its definition • but by definition of Mk, wk is in Ld • If not, then wk is in Ld, by its definition. • but by definition of Mk, wk is not in Ld • These contradictions show that no such Mk exists, and thus that Ld is not r.e.
Recursive and r.e. languages • If L is recursive, then so is L' • since the algorithm for deciding L will also decide L' if we complement its output • If L and L' are both r.e., then both are recursive, since • both L and L' have TMs M and M’ • given an input x, we may run both TMs on x • one will halt and return "yes" • if M halts, we answer "yes" for L and "no" for L' • if M’ halts, we answer "yes" for L' and "no" for L
Examples of recursive and r.e. languages • Ld is not r.e., so it's not recursive. • Ld' is r.e. • Mu may be modified to accept it • Ld' is not recursive • if it were, then Ld would be • Lu is r.e. but not recursive. • if it were recursive, then Ld' would be
Recursive and r.e. languages, conclusions • Not all languages are r.e. • Ld is not r.e. • The recursive languages are a proper subset of the r.e. languages • Lu is r.e. but not recursive • The CFLs are a proper subset of the recursive language • recall that {anbncn} is recursive but not a CFL
Properties of r.e. languages • Our final and rather astonishing result deals with properties of the r.e. languages. • For each a property P (e.g., being finite), we consider the question of whether a given r.e., language L has property P • We assume that L is represented by a TM M • Rice's theorem states that there is a decision algorithm for this question only when the answer is always "yes" or always "no"
Rice's theorem, restated • That is, for any property P we know that one of three cases holds: • P is true for any r.e. language • P is true for no r.e. language • P is true for some but not all r.e. languages • We know that in cases 1 and 2 there are trivial solution algorithms • We call such properties trivial properties. • Rice's theorem states that there are no solution algorithms for case 3
The halting problem • The proof of Rice's theorem is not difficult • But it may be instructive to consider a similar argument first • For a given TM M and input x, we may ask whether it halts on all input • This is the halting problem
Undecidability of the halting problem • To show undecidability, we may use the language of algorithms instead of TMs. • We'll suppose there’s an algorithm A for deciding the halting problem. • We'll then use A to construct an algorithm for deciding membership in Lu. • Since no such algorithm can exist, this contradiction shows that A cannot exist either
The algorithm for Lu • The algorithm U for Lu would take its input <M,x> and feed M and x to A. • If A returned “no”, then so would U • correctly, since if M can’t accept x unless it halts • If A returned “yes”, then U would return the result of simulating M on x • correctly, since the simulation halts. • Since U always halts and always determines whether M accepts w, it’s an algorithm for membership in Lu.
Proof of Rice's Theorem • Consider a nontrivial property P. • We first assume that the empty language {} does not have property P. • Since P is nontrivial, there's a (nonempty) r.e. language L that does have P • And since L is r.e., L is accepted by a TM ML
Rice's Theorem: a claim • The proof of Rice's theorem is by contradiction. • We’ll show that if there's a TM MP for deciding P, then Lu is recursive. • The proof depends on the following claim: • Claim: there's an algorithm for constructing TMs from pairs (M,w). The resulting TM M‘ • accepts L when M accepts w, and • accepts {} otherwise.
Proof of the claim • M' can be constructed to work as follows: • note: M' has its own input x, rather than M or w • Given M and w, simulate M on w. • If M doesn’t accept w, then M' doesn’t accept its input x. • If M accepts w, then M' simulates ML on x. • This M' behaves properly • if M does not accept w, M' accepts {} • otherwise M' accepts L.
Conclusion • So we may get a halting TM for Lu by taking (M,w), building M', and feeding M' to MP. • MP answers "yes" if and only if M accepts w. • Finally, the construction above works if {} does have property P • if an extra step is added to complement the output of MP (which does halt on all inputs) • Since no such halting TM exists, the proof is complete, and no nontrivial property of the r.e. languages is decidable
The Chomsky hierarchy • Each of the following classes of languages is properly contained in the succeeding class: • regular languages • deterministic context-free languages • context-free languages • context-sensitive languages • recursive languages • recursively enumerable languages
Grammars and the Chomsky hierarchy • Every regular language can be generated by a grammar of a more restrictive form than a CFG. • Every context-sensitive language can be generated by a grammar of a more general form than a CFG. • Every r.e. language can be generated by a grammar of a more general form than a CFG.
Decidability of CFG problems • Given CFGs G1 and G2, it’s undecidable whether • G1 is ambiguous • L(G1) is S* for any S • L(G1) = L(G2) • L(G1) is a subset of L(G2) • L(G1) ∩ L(G2) = {}