Transcript Undecidability

Undecidability
Church-Turing Thesis
A Turing machine that halts on all inputs is
the precise formal notion corresponding to
the intuitive notion of an algorithm
 An “algorithm” means a precisely defined
set of instructions
 This thesis cannot be formally proven
 See this link.

Consequence of Church-Turing thesis: If a
problem cannot be solved by a Turing
machine then it cannot be solved by a
human using a precisely defined sequence
of instructions, either
 Knowing a problem is undecidable helps
avoid wasted work
 Can extend Church-Turing thesis to
semidecidability (e.g. theorem proving)

Universal Turing machine
A UTM is a general purpose computer
 Can simulate any other Turing machine
 How simple can it be?
 U: Given description “T” of machine T and
“x” of input to T, simulates T on x and halts
if and only if T halts on x.
 Need to represent arbitrary TM in fixed
alphabet; see this link.

Details of U are uninteresting. But try to
understand how one or more UTM work.
 Examples:

–
–
–
–
–
Hopcroft and Ullman 40 state UTM
A 4 state 6 symbol UTM
Minsky 7 state 4 symbol UTM
Wolfram’s Universal Turing Machines
Details for Hopcroft and Ullman machine
Unsolvability of Halting Problem
A Turing machine can be represented by a
sequence of 4-tuples (state, symbol, new
state, direction or symbol) or 5-tuples of the
form (state, symbol, new state, direction,
new symbol)
 Thus a TM can be represented by a string of
symbols
 This string can be encoded in binary

Thus a TM can be represented as a binary
string
 Not all binary strings represent TM; if a
binary string does not represent a TM we
can say it represents a TM that always halts
right away
 A binary string can be interpreted as a
binary integer j
 Thus each binary integer j corresponds to a
TM (call it Tj )

The integer j can also be represented in
unary
 This permits j to be read by a TM having
only one symbol other than blank in its
alphabet

Every TM can be represented as Tj for some
j
 Let  be the problem whose base set is the
positive integers, and such that the right
answer for an integer j is “yes” if Tj does
not halt on input j, “no” otherwise
 Claim:  is not partially decidable.
 Proof: By contradiction.

–
Suppose  is partially decidable. Then there is
a Turing machine M that always gives the right
answer and halts on input j if Tj does not halt
on input j, and M loops otherwise.
–
–
–
–
–
–
–
–
But then M must be Ti for some i.
Then for any j, M halts on input j if and only if
Tj does not halt on input j.
In particular, for j = i, this is true.
Thus M halts on input i if and only if Ti does
not halt on input i.
But M is Ti .
Thus Ti halts on input i if and only if Ti does not
halt on input i.
This is a contradiction.
Thus M does not exist and  is not partially
decidable.
Note that self-reference is essential for this
proof.
 The problem occurs when Tj reads input j.
 Thus Tj is reading a description of itself.
 Self awareness is part of consciousness.
 Does this mean that consciousness is in
some way releted to undecidability?

Corollary:  is not decidable either.
 Thus the halting problem is not decidable.
 The halting problem H: The base set is the
set of pairs (T,x) where T is a Turing
machine and x is the input to T.
 The right answer is “yes” if T halts on input
x and “no” otherwise.
 If H were decidable then  would be too.
So H is undecidable.
 H is partially decidable (use a universal
Turing machine).

H is not partially decidable (if it were then
H would be decidable).
 H is not decidable on any computing
device.
 Don’t waste time trying to solve it.
 Illustration: In a certain village the barber
shaves everyone who does not shave
himself, and only those people who do not
shave themselves.
 Who shaves the barber?


Unsolvability of H results from TM running
on their own descriptions. A kind of selfawareness. Is this significant?
Other Unsolvable Problems
Show other problems unsolvable by
showing that if you could solve them, you
could solve the halting problem.
 Many problems are known to be unsolvable.
 Examples:

–
–
–
Blank tape halting problem
Hilbert’s tenth problem
Theorem proving in first-order logic
But many of these unsolvable problems are
partially decidable.
 A proof that the blank tape halting problem
is undecidable:

Unsolvability of BTHP




Let M have start state s, blank B, and these 5tuples:
– (s,0,t,0,R) (s,1,t,1,R) (s,B,y,B,R)
– (t,0,s,0,R) (t,1,s,1,R) (t,B,n,B,R)
M tests if a binary string has even length.
Let M’ have start state a, blank symbol B, all 5
tuples of M, and these 5-tuples:
– (a,B,b,0,R) (b,B,c,1,R) (c,B,d,1,L) (d,1,s,1,L)
Then M’ started on blank tape writes 011, moves
left to the 0, and transfers control to M.
Thus M halts on input 011 if and only if M’
halts on blank tape.
 If we could solve the blank tape halting
problem then we could use this to tell
whether M halts on input 011.
 The same idea works in general: If we
could solve BTHP we could solve the
general halting problem.
 Because the halting problem is unsolvable,
BTHP is unsolvable.

Busy Beaver Problem (step
counting version)
BB(n) is the largest number of steps taken
by any n state TM with a binary alphabet,
starting on blank tape, that eventually halts
 Suppose there were 5 TM having 3 states.
Suppose M1 took 3 steps and halted, M2
took 50 steps and halted, M3 looped, M4
took 7 steps and halted, and M5 looped.
Then BB(3) = 50.

Now BB(n) is not computable and no
computable function is always larger.
 Proof: If we could compute BB(n) we
could solve the halting problem. (How?)
 But with a few extra bits for each n we
could compute BB(n).

–
Just say which TM runs the longest and halts.
Thus with a few extra bits we could solve
the halting problem. These bits will never
be known unless by revelation.
 BB(n) grows very fast, faster than any
computable function.

The problem is classifying Turing
computations that do not obviously halt or
loop.
 This will forever be impossible to do unless
some computing device is more powerful
than a Turing machine. The behavior of
Turing machines is too subtle for us to fully
understand.
 See this link
 There is no easy way to distinguish the
beginning of a halting computation from the
beginning of a looping computation.

If there were an easy way to look at the
beginning of a computation and tell that it
was going to halt, then one would quickly
know that the other computations were
looping.
 If there were an easy way to look at the
beginning of a computation and tell that it
was going to loop, then one would quickly
know that the other computations would
halt.
 Neither is possible. Halting and looping
computations are very hard to distinguish.

Known bounds on BB(n)

The function Sigma(n) denotes the maximal
number of tape marks which a Turing
Machine (TM) with n internal states and a
two-way infinite tape can produce onto an
initially empty tape and then halt. The
function S(n) denotes the maximal number
of steps (shifts) which such a TM can do (it
needs not produce many tape marks). The
following table gives some known values:

n Sigma(n)
S(n)

1
1
1

2 4
6

3
6
21

4 13
107

5
>= 4098
>= 47,176,870

6
> 1.29*10865 > 3*101730

Some researchers believe that Sigma(5) will
never be known exactly. There are too
many Turing machines to consider and their
behavior is too subtle.
Following are transition tables and values of
S and Sigma for a few Turing machines.
 Try simulating them for a few steps to
understand how they work!
 The difficulty of understanding such
machines helps us to see “why” the halting
problem is unsolvable.
 No matter how advanced we become, there
will be infinitely many machines that are
beyond our understanding.
 We will not know whether they eventually
halt. Our math is not strong enough to say.

Record 5-state Turing Machines

The following table lists for each input state and symbol the output
state, symbol and tape movement. The initial state is A, the halting
state is H. The blank symbol is 0 and the tape is 2-way infinite.

No A0 A1 B0 B1 C0 C1 D0 D1 E0 E1 ones steps
1 B1L C1R C1L B1L D1L E0R A1R D1R H1L A0R 4098 47176870
2 B1L A1L C1R B1R A1L D1R A1L E1R H1R C0R 4098 11798826
3 B1L D0R C1R D1L A1R C1R H1L E1L A1L B0L 4097 23554764
4 B1L A1L C1R D1L A1R C1R H1L E0L C1R B1L 4097 11798796
5 B1L A1L C1R D0L A1R C1R H1L E1L E0R B1L 4096 11804910
6 B1L A1L C1R D0L A1R C1R H1L E1L C1R B1L 4096 11804896
7 B1L D1L C1R E0R A0L B0R E1L H1L C1R C1L 1471 2358064







Record 6-state Turing machines

This is a list of 6-state TMs producing many
ones, found by J.Buntrock and H.Marxen by
scanning around 10% of all 6-state TMs
with a program from June 2000 to October
2000. The blank symbol is 0 and the tape is
2-way infinite.

Name = a

ones = 17485734

steps = 95547257425490

ones > 1.7*10^7 steps > 9.5*10^13

A: B1R C0L

B: A1L A0R

C: D0L Z1R

D: E1R D1L

E: F0L E0L

F: F1R B0L

Name = b

ones = 36109969

steps = 1137975066814256

ones > 3.6*10^7

steps > 1.1*10^15

A: B1R C0L

B: A1L E0R

C: D1L F0L

D: B0L C1L

E: B1R A1R

F: B1R Z1R

Name = c

ones = 36109970

steps = 758650111948072

ones > 3.6*10^7

steps > 7.5*10^14

A: B1R C0L

B: A1L E0R

C: D1L F1L

D: B0L C1L

E: B1R A1R

F: C0L Z1R

Name = q

ones > 6.4*10^462

steps > 6.1*10^925

A: B1R B0L

B: C0R B1L

C: D1R A0L

D: E1L F1L

E: A1L D0L

F: Z1R E1L

Name = r

ones > 1.2*10^865

steps > 3.0*10^1730

A: B1R F0L

B: C0R D0R

C: D1L E1R

D: E0L D0L

E: A0R C1R

F: A1L Z1R
Simulation Results
Check these links for the results of
simulations of machines 1,2,3,a,r:
 Busy Beaver Simulation 1
 Busy Beaver Simulation 2
 Busy Beaver Simulation 3
 Busy Beaver Simulation a
 Busy Beaver Simulation r (the record
holder)

Chaitin’s Omega
A real number between 0 and 1
 If we could compute it we could solve the
halting problem.
 Omega is “random.” Writing n bits of
omega requires an n bit program. Omega
has no simple patterns.
 See this link and this link.

 = pP
-|p|
2
where P is the set of all halting Turing
machines and |p| is the length in bits of the
encoding of p.
 Let Pn be the set of TM that halt in n steps
or less.

n = pPn
-|p|
2
Then n approaches  as n approaches .
Omega
0.3
0.2
Omega
0.15
Omegan
OmegaP
0.1
0.05
0
Steps
Probability
0.25
Steps
n can be computed to any accuracy.
 To solve the halting problem using :

–
–
–
–

If M halts in n steps then answer “yes.”
If  - n < 2-|M| then answer “no.”
Do this for n = 1, 2, 3, … until an answer is
given.
Eventually either a “yes” or “no” answer will
be given.
Thus  will never be known. It is
unknowable unless it is given by revelation.
Partial solutions to the halting
problem
A loop tester is a Turing machine M that on
input i, halts if Ti loops on input i, and M
loops otherwise.
 A partial loop tester is a Turing machine M
such that if M halts on input i, then Ti loops
on input i.
 No loop tester exists. But partial loop
testers exist.

Improving partial loop testers





Suppose M is a partial loop tester. Let M be Tj.
What does M do on input j?
If M halts on input j then Tj loops on input j. But
Tj is M. Contradiction.
Therefore M loops on input j. But M does not
“know” that it loops on input j.
M cannot solve this instance of the loop testing
problem. But we can solve it.
Thus we excel every partial loop tester.
Let b(M) (“better” M) be M modified so
that it halts on input j.
 Thus b(M) knows that M halts on input j.
 It is easy to construct Turing machine b(M)
from M.
 b(M) is also a partial loop tester.
 But b(M) does not “know” that b(M) halts
on input k where b(M) is Tk.
 Consider b(b(M)). This machine knows
that b(M) halts on input k. Et cetera.

In this way we construct b(M), b2(M),
b3(M), et cetera. Each halts more often than
the previous one.
 Let b(M) be a Turing machine that halts on
input i if any of M, b(M), b2(M), et cetera
halt on input i. This improves all bn(M).
 Such a machine can be constructed by
simulating M, b(M), b2(M), … in parallel.
 Then one can improve b(M) to get
b(b(M)), b(b(b(M))) et cetera.
 Notation: b+1(M), b+2(M) … .

One obtains in the same way b 2 (M),
b3(M), b4(M), …, and eventually b(M).
2

 This can be written b (M).
3
4


 Similarly one obtains b (M) b (M)
5

b (M) … .
 This goes on for a long time!
 Related to the ordinals.

A paradox of consciousness
Suppose H is a Turing machine encoding
human thought about looping of Turing
machines.
 Suppose H halts on input i if a human with
lots of time can eventually deduce
mathematically that Ti loops on input i. H
loops otherwise.
 Then H is a partial loop tester. Therefore H
loops on input j where H is Tj.





Because H loops on input j, this means that
humans cannot deduce that Tj loops on input j.
Tj is H, so humans cannot deduce that H loops on
input j.
But we just showed that H loops on input j. We
are humans. Contradiction.
Possibilities:
– H is not a partial loop tester; human logic is
faulty.
– Humans can never know H.
– Human thought cannot be represented by a
Turing machine.

All possibilities are disturbing.
–
–
–
The first means that we can’t think correctly
even when reasoning mathematically.
The second means that we can never know how
we think.
The third means that our thought in some way
exceeds the capacity of a Turing machine.
Note that not much human logic is needed
to reason that H loops on input j.
 Some people think that our consciousness,
or self-awareness, permits us to excel
Turing machines.

Another possibility: H is a partial loop
tester but we can never know that H is a
partial loop tester.
 This implies that human mathematical
thought employs reasoning that human
thought cannot justify.
 What would be an example of this?
