Theory of Computation

Reducibility

Bow-Yaw Wang

Academia Sinica

Spring 2012

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 1 / 32

Reducibility

In mathematics, many problems are solved by “reduction.”

Recall the reduction from Eulerian path to Eulerian cycle.
I Suppose EC(G) returns true iff G has a Eulerian cycle.
I Let s, t be nodes of a graph G.
I To check if there is a Eulerian path from s to t in G.
I Construct a graph G0 that is identical to G except an additional edge
between s and t.
I If EC(G0 ) returns true, there is a Eulerian path from s to t.
I If EC(G0 ) returns false, there is no Eulerian path from s to t.
Instead of inventing a new algorithm for finding Eulerian paths,
we use EC(G) as a subroutine.
We say the Eulerian path problem is reduced to the Eulerian cycle
problem.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 2 / 32

Reducibility

Let us say A and B are two problems and A is reduced to B.

If we solve B, we solve A as well.
I If we solve the Eulerian cycle problem, we solve the Eulerian path
problem.
If we can’t solve A, we can’t solve B.
To show a problem P is not decidable, it suffices to reduce ATM to
P.
We will give examples in this chapter.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 3 / 32

The Halting Problem for Turing Machines
The halting problem is to test whether a TM M halts on a string w.
As usual, we first give a language-theoretic formulation.
HALTTM = {hM, wi : M is a TM and M halts on the input w}.

Theorem 1
HALTTM is undecidable.

Proof.
We would like to reduce the acceptance problem to the halting
problem. Suppose a TM R decides HALTTM . Consider
S = “On input hM, wi where M is a TM and w is a string:
1 Run TM R on the input hM, wi.
2 If R rejects, reject.
3 If R accepts, simulate M on w until it halts.
4 If M accepts, accept; if M rejects, reject.”
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 4 / 32
Emptiness Problem for Turing Machines
Consider ETM = {hMi : M is a TM and L(M) = ∅}.

Theorem 2
ETM is undecidable.

Proof.
We reduce the acceptance problem to the emptiness problem. Let the
TM R decides ETM . Consider
S = “On input hM, wi where M is a TM and w a string:
1 Use hMi to construct
M1 = “On input x:
1 If x 6= w, reject.
2 If x = w, run M on the input x. If M accepts x, accept.”
2 Run R on the input hM1 i.
3 If R accepts, reject; otherwise, accept.”
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 5 / 32
Regularity Problem for Turing Machines
Consider
REGULARTM = {hMi : M is a TM and L(M) is regular}.

Theorem 3
REGULARTM is undecidable.

Proof.
Let R be a TM deciding REGULARTM . Consider
S = “On input hM, xi where M is a TM and x a string:
1 Use hMi to construct
M2 = “On input x:
1 If x is of the form 0n 1n , accept.
2 Otherwise, run M on the input w. If M accepts w, accepts.”
2 Run R on the input M2 .
3 If R accepts, accept; otherwise, reject.”
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 6 / 32
Rice’s Theorem
Theorem 4
Let P be a language consisting of TM descriptions such that
1 P is not trivial (P 6= ∅ and there is a TM M with hMi 6∈ P);
2 If L(M1 ) = L(M2 ), hM1 i ∈ P iff hM2 i ∈ P.
Then P is undecidable.

Proof.
Let R be a TM deciding P. Let T∅ be a TM with L(T∅ ) = ∅. WLOG, assume hT∅ i 6∈ P.
Moreover, pick a TM T with hTi ∈ P. Consider
S = “On input hM, wi where M is a TM and w a string:
1 Use hMi to construct
Mw = “On input x:
1 Run M on w. If M halts and rejects, reject.
2 If M accepts w, run T on x.”
2 Run R on hMw i.
3 If R accepts, accept; otherwise, reject.”
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 7 / 32
Language Equivalence Problem for Turing Machines

Consider

EQTM = {hM1 , M2 i : M1 and M2 are TM’s with L(M1 ) = L(M2 )}.

Theorem 5
EQTM is undecidable.

Proof.
We reduce the emptiness problem to the language equivalence
problem this time. Let the TM R decide EQTM and TM M1 with
L(M1 ) = ∅. Consider
S = “On input hMi where M is a TM:
1 Run R on hM, M1 i.
2 If R accepts, accept; otherwise, reject.”
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 8 / 32
Computation History

Definition 6
Let M be a TM and w an input string. An accepting computation
history for M on w is a sequence of configurations C1 , C2 , . . . , Cl where
C1 is the start configuration of M on w;
Cl is an accepting configuration of M; and
Ci yields Ci+1 in M for 1 ≤ i < l.
A rejecting computation history for M on w is similar, except Cl is a
rejecting configuration.

Note that a computation history is a finite sequence.

A deterministic Turing machine has at most one computation
history on any given input.
A nondeterminsitic Turing machine may have several
computation histories on an input.
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 9 / 32
Linear Bounded Automaton

control

0 0 1 0 1 1 1 0
Figure: Schematic of Linear Bounded Automata

Definition 7
A linear bounded automaton is a Turing machine whose tape head is
not allowed to move off the portion of its input. If an LBA tries to
move its head off the input, the head stays.

With a larger tape alphabet than its input alphabet, an LBA is able
to increase its memory up to a constant factor.
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 10 / 32
Acceptance Problem for Linear Bounded Automata

Consider

ALBA = {hM, wi : M is an LBA and M accepts w}.

Lemma 8
Let M be an LBA with q states and g tape symbols. There are exactly qngn
different configurations of M for a tape of length n.

An LBA has only a finite number of different configurations on an

input.
Many langauges can be decided by LBA’s.
I For instance, ADFA , ACFG , EDFA , and ECFG .
Every context-free langauges can be decided by LBA’s.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 11 / 32

Acceptance Problem for Linear Bounded Automata
Theorem 9
ALBA is decidable.

Proof.
Consider
L = “On input hM, wi where M is an LBA and w a string:
1 Simulate M on w for qngn steps or until it halts. (q, n, and g are
obtained from hMi and w.)
2 If M does not halt in qngn steps, reject.
3 If M accepts w, accept; if M rejects w, reject.”

The acceptance problem for LBA’s is decidable. What about the

emptiness problem for LBA’s?
ELBA = {hMi : M is an LBA with L(M) = ∅}.
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 12 / 32
Emptiness Problem for Linear Bounded Automata

Theorem 10
ELBA is undecidable.

Proof.
We reduce the acceptance problem for TM’s to the emptiness problem
for LBA. Let R be a TM deciding ELBA . Consider
S = “On input hM, wi where M is a TM and w a string:
1 Use hMi to construct the following LBA:
B = “On input hC1 , C2 , . . . , Cl i where Ci ’s are configurations of M:
1 If C1 is not the start configuration of M on w, reject.
2 If Cl is not an accepting configuration, reject.
3 For each 1 ≤ i < l, if Ci does not yield Ci+1 , reject.
4 Otherwise, accept.”
2 Run R on hBi.
3 If R rejects, accept; otherwise, reject.”
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 13 / 32
Universality of Context-Free Grammars

Consider a problem related to the emptiness problem for CFL’s

ALLCFG = {hGi : G is a CFG and L(G) = Σ∗ }.

Let x be a string. Write xR for the string x in reverse order.

I For example, 100R = 001, levelR = level.
I Another example,
乾隆： 客上天然居 居然天上客
紀曉嵐： 人過大鐘寺 寺鐘大過人
Let C1 , C2 , . . . , Cl be the accepting configuration of M on input w.
Consider the following string in the next theorem:

#hC1 i#hC2 iR # · · · #hC2k−1 i#hC2k iR # · · · #hCl i#

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 14 / 32

Universality of Context-Free Grammars
Theorem 11
ALLCFG is undecidable.

Proof.
We reduce the acceptance problem for TM’s to the universalty problem. We construct
a nondeterministic PDA D that accepts all strings if and only if M does not accept w.
The input and stack alphabets of D contain symbols to encode M’s configurations.
D = “On input #x1 #x2 # · · · #xl #:
1 Do one of the following branches nondeterministically:
I If x1 6= hC1 i where C1 is the start configuration of M on w, accept.
I If xl 6= hCl i where Cl is a rejecting configuration of M, accept.
I Choose odd i nondeterministically. If xi 6= hCi, xRi+1 6= hC0 i, or C
does not yield C0 (C, C0 are configurations of M), then accept.”
I Choose even i nondeterministically. If xRi 6= hCi, xi+1 6= hC0 i, or C
does not yield C0 (C, C0 are configurations of M), then accept.”
M accepts w iff the accepting computation history of M on w is not in L(D) iff
CFG(D) 6∈ ALLCFG .
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 15 / 32
Post Correspondence Problem (PCP)
 
t
A domino is a pair of strings:
b
    
t1 t2 tk
A match is a sequence of dominos ··· such
b1 b2 bk
that t1 t2 · · · tk = b1 b2 · · · bk .
The Post correspondence problem is to test whether there is a
match for a given set of dominos.
PCP = {hPi : P is an instance of the PCP with a match}
Consider
       
b a ca abc
P= , , ,
ca ab a c
A match in P:
     
a b ca a abc
ab ca a ab c
Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 16 / 32
The Modified Post Correspondence Problem

The modified Post correspondence problem is a PCP where a

match starts with the first domino. That is,

MPCP = {hPi : P is an instance of the PCP with a match

starting with the first domino}

Theorem 12
PCP is undecidable.

Proof idea.
We reduce the acceptance problem for TM’s to PCP. Given a TM M and
a string w, we first construct an MPCP P0 such that hP0 i ∈ MPCP if and
only if M accepts w. The MPCP P0 encodes an accepting computation
history of M on w. Finally, we reduce MPCP P0 to PCP P.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 17 / 32

The Post Correspondence Problem

Proof.
Let the TM R decide MPCP. Let M = (Q, Σ, Γ, δ, q0 , qaccept , qreject ) be the given TM and
w = w1 w2 · · · wn the input. The set P0 of dominos has
 
#
as the first domino. Begin with the start configuration
#q0 w1 w2 · · · wn #
(bottom).

# q0 w1 w2 ··· wn #

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 18 / 32

The Post Correspondence Problem

Proof (cont’d).
 
qa
if δ(q, a) = (r, b, R) with q 6= qreject . Reads a at state q (top); writes b and
br
moves right (bottom).
 
cqa
if δ(q, a) = (r, b, L) with q 6= qreject . Reads a at state q (top); writes b and
rcb
moves left (bottom).
 
a
if a ∈ Γ. Keeps other symbols intact.
a

# q0 0 1 0 0

# q0 0 1 0 0 # 2 q7 1 0 0

δ(q0 , 0) = (q7 , 2, R)

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 19 / 32

The Post Correspondence Problem

Proof (cont’d).
   
# #
and Matches previous # (top) with a new # (bottom). Adds xy
# xy#
when M moves out of the right end.

# q0 0 1 0 0 #

# q0 0 1 0 0 # 2 q7 1 0 0 #

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 20 / 32

The Post Correspondence Problem

Proof (cont’d).
   
aqaccept qaccept a
and if a ∈ Γ. Eats up tape symbols around qaccept .
qaccept qaccept
 
qaccept ##
. Completes the match.
#

# 2 1 qaccept 0 0 # # qaccept # #

··· ···

# 2 1 qaccept 0 0 # 2 1 qaccept 0 # # qaccept # #

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 21 / 32

The Post Correspondence Problem

Proof (cont’d).
So far, we have reduced the acceptance problem of TM’s to MPCP. To complete the
proof, we need to reduce MPCP to PCP.
Let u = u1 u2 · · · un . Define
?u = ∗ u1 ∗ u2 ∗ ··· ∗ un
u? = u1 ∗ u2 ∗ ··· ∗ un ∗
?u? = ∗ u1 ∗ u2 ∗ ··· ∗ un ∗

Given a MPCP P0 :      
t1 t2 tk
, ,...,
b1 b2 bk
Construct a PCP P:
       
?t1
?t2 ?tk ∗♦
, ,..., ,
b2 ?
?b1 ? bk ? ♦
 
?t1
Any match in P must start with the domino .
?b1 ?

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 22 / 32

Computable Functions

Definition 13
f : Σ∗ → Σ∗ is computable if some Turing machine M, on input w, halts
with f (w) on its tape.

Usual arithmetic operations on integers are computable functions.

For instance, the addition operation is a computable function
mapping hm, ni to hm + ni where m, n are integers.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 23 / 32

Mapping Reducibility

Definition 14
A language A is mapping reducible (or many-one reducible) to a
languate B (written A ≤m B) if there is a computable function
f : Σ∗ → Σ∗ such that

w ∈ A if and only if f (w) ∈ B, for every w ∈ Σ∗ .

f is called the reduction of A to B.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 24 / 32

Properties of Reducibility

Theorem 15
If A ≤m B and B is decidable, A is decidable.

Proof.
Let the TM M decide B and f the reduction of A to B. Consider
N = “On input w:
1 Construct f (w).
2 Run M on f (w).
3 If M accepts, accept; otherwise reject.

Corollary 16
If A ≤m B and A is undecidable, then B is undecidable.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 25 / 32

Examples

Example 17
Give a mapping reduction of ATM to HALTTM .

Proof.
We need to show a computable function f such that hM, wi ∈ ATM if
and only if hM0 , w0 i ∈ HALTTM whenever hM0 , w0 i = f (hM, wi).
Consider
F = “On input hM, wi:
1 Use hMi and w to construct
M0 = “On input x:
1 Run M on x.
2 If M accepts, accept.
3 If M rejects, loop.”
2 Output hM0 , wi.”

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 26 / 32

Examples

Example 18
Give a mapping reduction from ETM to EQTM .

Proof.
The proof of Theorem 5 gives such a reduction. The reduction maps
the input hMi to hM, M1 i where M1 is a TM with L(M1 ) = ∅.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 27 / 32

Transitivity of Mapping Reductions

Lemma 19
If A ≤m B and B ≤m C, A ≤m C.

Proof.
Let f and g be the reductions of A to B and B to C respectively. g ◦ f is a
reduction of A to C.

Example 20
Give a mapping reduction from ATM to PCP.

Proof.
The proof of Theorem 12 gives such a reduction. We first show
ATM ≤m MPCP. Then we show MPCP ≤m PCP.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 28 / 32

More Properties about Mapping Reductions

Theorem 21
If A ≤m B and B is Turing-recognizable, then A is Turing-recognizable.

Proof.
Similar to the proof of Theorem 15 except that M and N are TM’s, not
deciders.

Corollary 22
If A ≤m B and A is not Turing-recognizable, then B is not
Turing-recognizable.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 29 / 32

More Properties about Mapping Reductions

Observe that A ≤m B if and only if A ≤m B.

I The same reduction applies to A and B as well.
Recall that ATM is not Turing-recognizable.
In order to show B is not Turing-recognizable, it suffices to show
ATM ≤m B.
I ATM ≤m B implies ATM ≤m B. That is, ATM ≤m B.

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 30 / 32

Equivalence Problem for TM’s (revisited)

Theorem 23
EQTM is neither Turing-recognizable nor co-Turing-Recognizable.

Proof.
We first show ATM ≤m EQTM . Consider
F = “On input hM, wi where M is a TM and w a string:
1 Construct
M1 = “On input x:
1 Reject.”
M2 = “On input x:
1 Run M on w. If M accepts, accept.”
2 Output hM1 , M2 i.”

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 31 / 32

Equivalence Problem for TM’s (revisited)

Proof (cont’d).
Next we show ATM ≤m EQTM . Consider
G = “On input hM, wi where M is a TM and w a string:
1 Construct
M1 = “On input x:
1 Accept.”
M2 = “On input x:
1 Run M on w.
2 If M accepts w, accept.”
2 Output hM1 , M2 i.”

Bow-Yaw Wang (Academia Sinica) Decidability Spring 2012 32 / 32