All the reductions we have studied so far are variants of many-one reductions. Other types of reductions such as truth-table reductions and Turing reductions can be similarly defined for distributional problems. The reader should find no difficulty in defining them with the desired properties. It is a long-standing open problem whether Turing reductions (or truth-table reductions) are provably more powerful than many-one reductions on NP problems. Not much has been done for distributional NP problems. However, it is not known whether truth-table or Turing reductions can help identify additional average-case NP-complete problems encountered in practice. For one thing, we note that all the known natural NP-complete problems are indeed many-one complete. Nevertheless, NP search problems can be reduced to NP decision problems by Turing reductions. For distributional problems, randomized truth-table reductions (defined below) are sufficient [BCGL92].
By Theorem 4.3, an ap-time randomized algorithm that
produces a correct output with probability at least
can be
iterated to get a correct output with probability 1.
We also note that Theorem 4.1 remains true if one uses
deterministic Turing reductions [Gur91a].
If in a p-time (ap-time) randomized Turing reduction
all queries can be generated at the beginning
independent of other queries,
then the reduction is called a p-time (ap-time)
randomized truth-table
reduction. Randomized Turing reductions are closed for randomized ap-time
solvable problems and are transitive in the special case where on
input
only queries of length at least
are asked, for
a constant
[BCGL92].
As pointed out in [BCGL92], the standard
Turing reduction
of a search problem to the corresponding decision problem,
which asks queries to the set of prefixes of
a solution on input
, fails on domination when the query string
is too long. This is because
has distribution
with input distribution
, which will
be too small to dominate
when
is large.
However, if the search problem has
a unique solution, then instead of asking for
prefixes of a solution, one can ask non-adaptively and
deterministically for the
th bit of
the solution with probability
. The domination
condition is therefore satisfied. Based on this idea,
Ben-David et al. [BCGL92] obtained the following result.

Proof Sketch.simplicity, assume that
implies
. Let
, and let
be a set of standard
universal hash functions from
strings of length
to strings of length
(e.g.,
matrices over the finite field
).
Define
as follows.
An instance will be a binary string
and an integer
, and
an instance will be in
if there exists
a
such that
,
,
,
, and the
th bit of
is 1.
The distribution
is defined by
if
and
;
and 0 otherwise.
Construct a truth-table reduction
as follows.
On input
, for every
value of
,
first
selects uniformly at random
an
in
and a
of length
.
Let
.
then makes queries
, and
to
and
outputs
.
Clearly,
runs in time polynomial in
.
To check the validity,
let
and assume
(the case
that
is straightforward). Let
if
and
otherwise.
Then
.
It follows that for a randomly selected
and
of length
, the
expected number of
satisfying
is between
and
. By Markov's inequality
and the property of universal
hashing that pairs of hashing points are pairwise independent,
the probability that there is a unique
such that
is great than or equal
to
for some constant
. In such
a case
returns a correct answer
.
To check the domination condition, it suffices to note that
appears as a query with probability
.