Episode 3 Overview of the theory of computation • Turing machines • The traditional concepts of computability, decidability and recursive enumerability • The limitations of.

Download Report

Transcript Episode 3 Overview of the theory of computation • Turing machines • The traditional concepts of computability, decidability and recursive enumerability • The limitations of.

Episode 3
Overview of the theory of
computation
• Turing machines
• The traditional concepts of computability, decidability and recursive enumerability
• The limitations of the power of Turing machines
• The Church-Turing thesis
• Mapping reducibilty
• Turing reducibility
• Kolmogorov complexity
0
3.1
Components of a Turing machine (TM)
Input tape
2
0
-
-
-
-
Work tape
4
0
0
-
1
#
Output tape
4
0
0
-
-
-
- -
-
Read-only
2
$
- -
Read-write
-
-
-
No direct access
Control
(transition function)
- blank symbol
Has a finite number of states, two of which, Start and Halt, are special.
3.2
How a TM works
Input tape
2
0
-
-
-
-
Work tape
-
- -
-
-
-
Output tape
-
- -
-
-
-
- -
-
-
-
-
-
- -
• At the beginning, the machine is in its Start state, the input tape has an input on it,
the other tapes are blank, and the two scanning heads are in their leftmost positions.
Such a situation is shown above, with input being 20.
• After the computation starts, the machine goes from one state to another, writes on the
work tape, and moves its scanning heads left or right according to its transition function,
which precisely prescribes what to do in each particular case, depending on the current
state and the contents of the two scanned cells (on the input and work tapes).
• If and when the machine enters the Halt state, it halts, and whatever string w is to
the left of the work-tape scanning head, will be automatically copied onto the output
tape. Such a string w is considered to be the output of the machine. Slide 3.1 shows
such a situation, with output 400.
3.3
Variants of Turing machines
There are numerous reasonable variations of the Turing machine
model, and they all turn out to be equivalent. For example:
• More often than not, there is just a single tape for everything (input,
output, work).
• Some models allow multiple work tapes.
• The tape can be infinite in both directions (have no beginning).
• And so on.
One of the advantages of the particular model we have chosen here is
that it can be easily adapted to cases where infinitely many inputs and
outputs are allowed. In such cases, entering the Halt state will only have
the effect of sending a (next) output to the output tape, but otherwise the
machine would not halt.
Input
Output
2
0
- 3
4
0
0
- 1
- 9
-
0
- 2
-
...
1
0
-
...
0
3.4
Computability
Definition 3.1. Let f: AB be a function.
(a) We say that a TM M computes f iff, for every string wA,
whenever M receives w as input, M (halts and) outputs string
v such that v = f(w).
(b) When such a machine M exists, we say that f is computable.
There is no need to separately consider functions with more than one
(but finitely many) arguments, as tuples of strings can be encoded
through (just) strings.
For example, the function f(x,y)=x+y can be thought of as a
single-argument function that returns string “3” on input string “(2,1)”,
returns string “20” on input string “(8,12)”, etc.
3.5
Decidability
Definition 3.2. Let S be a set of strings over a given alphabet A.
(a) We say that a TM M decides S iff, for every string wA*,
whenever M receives w as input, M (halts and) outputs:
1 (or “yes”, or “accept”) if wS;
0 (or “no”, or “reject”) if wS.
(b) When such a machine M exists, we say that S is decidable.
As every set S can be thought of as and identified with its characteristic
function, decidability is just a special case of computability.
The concept of decidability extends to predicates and relations,
as the latter can be thought of as sets.
3.6
Recursive enumerability
Definition 3.3. Let S be a set of strings over a given alphabet A.
(a) We say that a TM M recognizes S iff, for every string wA*,
whenever M receives w as input:
if wS, then M (halts and) outputs “yes”;
if wS, then M either (halts and) outputs something else
(such as “no”), or never halts.
(b) When such a machine M exists, we say that S is recursively
enumerable.
Again, this definition extends from sets to predicates and relations.
3.7
Computational problems in the traditional sense
Computational problems in the traditional sense, as established in
the traditional theory of computation, are nothing but functions,
sets, predicates or relations (to be computed or decided or recognized).
We will follow that tradition, and often say “problem” instead of
“function”, “relation”, etc.
For example, we may refer to the function f(x)=x2 as “the problem of
finding the square of a given number”, and refer to the set (predicate,
relation) {x | x is an even number} as “the problem of telling whether
a given number is even”, or “the evennees problem”, etc.
3.8
The limitations of the power of TMs
Alas, not all reasonable problems can be handled by TMs.
Examples of undecidable or incomputable problems:
1. The acceptance problem for Turing machines, i.e. the problem of telling whether a
given Turing machine accepts (meaning outputting “yes”) a given input.
2. The halting problem for Turing machines, i.e. the problem of telling whether a
given Turing machine ever halts on a given input.
3. The problem of finding the Kolmogorov complexity of a given number.
4. The problem of telling whether a given polynomial equation has integral roots.
5. The problem of first-order logical validity.
Examples of problems that are not recursively enumerable:
6. The complements of the above problems.
7. The problem of second-order logical validity.
8. The problem of telling whether a given formula of arithmetic is true.
3.9
The Church-Turing thesis
Algorithms = Turing machines = Computers
intuitive concept
mathematical concept
intuitive concept
Of course, here “=” should be understood not in the strict sense, but
in the sense of equivalence of power.
That is, according to the Church-Turing thesis, the problems that have
algorithmic solutions also have Turing machine solutions (deciders,
recognizers, etc.), and vice versa.
Thus, the existence of undecidable or incomputable problems means
the existence of problems that have no algorithmic solutions, problems
that no computers can ever handle.
3.10
Definition of mapping reducibility
Definition 3.4. Let A and B be sets of strings over an alphabet .
(a) We say that a function f: ** is a mapping reduction
from A to B iff f is computable and, for every w*,
wA iff f(w)B.
(b) When such a function f exists, we say that A is mapping
reducible to B, and write AmB. In the literature, mapping
reducibility is more often called “many-one reducibility”.
*
*
A
B
f
f
3.10
Definition of mapping reducibility
Definition 3.4. Let A and B be sets of strings over an alphabet .
(a) We say that a function f: ** is a mapping reduction
from A to B iff f is computable and, for every w*,
wA iff f(w)B.
(b) When such a function f exists, we say that A is mapping
reducible to B, and write AmB. In the literature, mapping
reducibility is more often called “many-one reducibility”.
*
*
A
B
f
f
3.10
Definition of mapping reducibility
Definition 3.4. Let A and B be sets of strings over an alphabet .
(a) We say that a function f: ** is a mapping reduction
from A to B iff f is computable and, for every w*,
wA iff f(w)B.
(b) When such a function f exists, we say that A is mapping
reducible to B, and write AmB. In the literature, mapping
reducibility is more often called “many-one reducibility”.
*
*
A
B
f
f
3.11
Using mapping reducibility for proving decidability/undecidability
Theorem 3.5. If AmB and B is decidable, then A is decidable.
Proof: Let DB be a decider for B and f be a mapping reduction from
A to B. We describe a decider DA for A as follows.
DA= “On input w:
1. Compute f(w).
2. Run DB on input f(w) and output whatever DB outputs.” 
( means “end of proof ”).
Thus, the decidability of a problem A can be proven by finding a mapping reduction
from A to some problem B which is already known to be decidable.
Or, the undecidability of a problem B can be proven by finding a mapping reduction
from A to B, where A is a problem already known to be undecidable.
Theorem 3.5 remains valid with “recursively enumerable” instead
of “decidable”.
3.12
A mapping reduction of the acceptance problem to the halting problem
For every TM M, let M’ be the following TM:
M’ = “On input x:
1. Run M on x.
2. If M outputs “yes”, accept.
3. If M outputs anything else, enter an infinite loop.”
Thus, • If M outputs “yes” on input x, then M’
accepts x
• If M outputs anything else on input x, then M’ never halts on x
• If M never halts on input x, then M’
never halts on x
To summarize, M accepts x iff M’
halts on x
Let then f be the function defined by f(M,w)=(M’,w).
Is f computable? Of course.
And, since we have
(M,w)ACCEPTANCE_PROBLEM iff f(M,w)HALTING_PROBLEM,
f is a mapping reduction of the acceptance problem to the halting
problem.
3.13
Definition of Turing reducibility
An oracle for a set (relation) B is an external device that is capable of
reporting whether any given string w is a member of B.
An oracle Turing machine (OTM) is a modified Turing machine that
has the additional capability of querying an oracle.
Example: Construct an OTM O with an oracle for the acceptance
problem, such that O decides the nonacceptance problem (the
complement of the acceptance problem).
O = “On input (M,w), where M is a TM and w is a string:
1. Query the oracle to determine whether M accepts w.
2. If the oracle answers NO, accept; if YES, reject.”
Definition 3.6. We say that a problem A is Turing reducible to a
problem B, written ATB, iff there is an OTM M with an oracle for
B, such that M decides A (or M computes A, if A is a function
rather than a relation).
3.14
Using Turing reducibility for proving decidability/undecidability
Theorem 3.7. If ATB and B is decidable, then A is decidable.
Proof. If B is decidable, then we may replace the oracle for B by an
actual procedure that decides B. Thus we may replace the OTM that
(using an oracle for B) decides A by an ordinary TM that decides A. 
Does this proof go through for “recursively enumerable” instead of
“decidable”? (see the end of Slide 3.11)
No! In fact, one can prove the opposite. For example, as we saw on the previous
slide, the nonacceptance problem is Turing reducible to the acceptance problem
(generally, any problem is Turing reducible to its complement). And the acceptance
problem is obviously recursively enumerable (why?). But the nonacceptance
problem is not recursively enumerable (why?).
It also follows from the comment at the end of Slide 3.11 that the nonacceptance
problem is not mapping reducible to the acceptance problem.
Does mapping reducibility always imply Turing reducibility?
Turing reducibility is the weakest form of reducibility.
Yes!
3.15
Turing reductions of the acceptance problem to the halting problem
Example: Show that the acceptance problem is Turing reducible to the halting problem.
Solution 1. When receiving the question “Does M accept w?”, replace M by M’ (as
described on Slide 3.12), and ask the oracle whether M’ halts on w. Repeat whatever
answer you get from the oracle. 
Solution 2. When receiving the question “Does M accept w?”, ask the oracle if M halts
on input w.
If the oracle says “No”, you also say “No”.
If the oracle says “Yes”, start running (simulating) machine M on input w until
you see it has halted. If its output is “Yes”, you also say “Yes”. Otherwise say “No”. 
Why could not we just do simulation as described in the last paragraph of Solution 2,
without any preliminary use of the oracle? Because there would be no guarantee that
the simulation would ever end.
3.16
Kolmogorov complexity
All Turing machines can be listed in the lexicographic order of their
descriptions:
M0, M1, M2, M3, ..., Mi, ....
Number i can thus be considered the code of the machine Mi.
Definition 3.8. Let m be a natural number. The Kolmogorov
complexity of m is the smallest number i such that machine Mi
outputs m on input 0.
Note: More often, Kolmogorov complexity is defined not as the
above number i itself, but as the size |i| of that number, i.e. the logarithm
of i. For our purposes, however, this makes no difference, and we will
stick to Definition 3.8 as it is given.
Importance: Kolmogorov complexity can be seen as a mathematical
counterpart of the intuitive concepts of “randomness” or “amount of
information”. The greater the Kolmogorov complexity of a given object,
the more random it is and the more information it contains.
3.17
Turing reduction of the Kolmogorov complexity problem to the halting
problem
Example: Show that the Kolmogorov complexity problem (i.e. the problem of finding
the Kolmogorov complexity of a given number) is Turing reducible to the halting
problem.
Solution. After receiving the question “What is the Kolmogorov complexity of m?”,
initialize variable i to 0, and do the following:
1. Ask the oracle if the machine Mi halts on input 0.
2. If the oracle says “No”, increment i to i+1, and go back to Step 1.
3. If the oracle says “Yes”, simulate Mi on input 0 until it halts. If you see that the
output of Mi is m, return i as your output. Otherwise, increment i to i+1 and go back to
Step 1. 
How many times does the above algorithm use the oracle? Unlike the example on
Slide 3.15, here the oracle is queried more than once. The number of queries, however,
can be shown to be bounded by a certain linear function of m.
Theorem 3.9. The Kolmogorov complexity problem is not computable.