The distributional matrix correspondence problem was first shown to be complete for DistNP by Gurevich [Gur90]. A detailed proof was given in [BG95]. The proof presented here is based on [BG95].
Recall that we use
to denote
the set of unimodular matrices, and a unimodular matrix
is a two-by-two matrix of integer entries such that
.
A matrix is represented
by a list of its entries in binary notation.
We begin with a discussion on positive unimodular matrices.
Positive matrices
A unimodular matrix is called positive if it has no negative
entries. Positive matrices form a monoid
while
forms a multiplicative group called
modular group. In this section, by a column we mean a column of
two relatively prime nonnegative integers. We may view
a positive matrix as the pair of its columns. If
is a column,
we use
to represent the upper and
the lower components of
. Let
and
be two columns. Write
if
and
, and write
if
and either
or
. Define
to be the maximal
entry of a positive matrix
. Let

Proof.(1) is obvious. For part (2), note that if
, then
from (1), so
. For part (3), if
occurs twice in the same row or the
same column, then
divides
and so
.
Otherwise, we first note that the determinant being positive implies
that
cannot occur on
and
,
so the only case left is
. In this case,
and so
.
Proof.that Lemma 6.14(2) implies that
and
generate a free monoid.
Also note that
forms a free monoid on the operation of
concatenation.
So it suffices to show that
every non-unit positive matrix
is a product of
and
. Define
and
. We prove the lemma by induction on
.
Since the entries of the main diagonal cannot be 0,
.
Note that
and
are the only non-unit
matrices of weights
. So the case
is obvious.
Suppose
. Then
.
We assume that
occurs in
, the case that
occurs in
is
similar. If
, then
and so
. Similarly, if
then
.
Thus, the column
has nonnegative entries and so
.
Since
,
.
By the induction hypothesis,
is a product
of matrices
and
. By Lemma 6.14(1),
.
We call the greater column of a non-unit positive matrix is called the major column. In the case of the unit matrix, we call either column major. The other column is call minor.
Proof.loss of generality, assume that
is not the
unit matrix. It follows that both components of the major
column are positive. Suppose that
is the major column (the
case that
is the major column is similar). We will show that
is the least column such that
.
Let
be any column such that
. Then
for some
because
and
are relatively prime. Hence,
and
. If
, then
either
or
is negative. Thus,
and
therefore
,
.
The minor column
can be found in polynomial time by using the extended
Euclid's algorithm [Knu73].
Let
be a natural number. Denote by
the length of
the binary notation of
. Let
be a positive matrix,
and
.
Proof.suffices to prove that the number
of positive matrices of size
is
. Let
denote the number
of positive integers
that are prime to
, and
. Then
. Thus,
.
There are exactly two isomorphisms of
to
.
One takes
to 0
and
to 1, denoted by
, and the other one takes
to 1 and
to 0. Let
denote the isomorphism
from
to
. By an induction on the length of
,
it is easy to check that if
starts with 0 (respectively, 1),
then the lower (respectively, upper) row of
is major.
Note that the size of a matrix
may not be polynomially related
to the length of the corresponding string
.
For example, a matrix
whereas
. This implies that
is not p-time computable. However,
we can show that
is ap-time computable. Let
be a
positive matrix. Denote
by
the probability distribution of
which is
proportional to
by Lemma 6.17.
Let
be proportional to
. .
Proof.
be a positive matrix, then
can be computed by the following recursive algorithm.
If
is the unit matrix then
is the empty string (the identity
of the monoid
).
Suppose that
differs from the unit matrix. If
is the major column, let
and
, then
. If
is the major column, let
and
, then
. This algorithm runs in time
proportional to
. Next, we show that
is polynomial on
-average.
Suppose that
is a nonempty binary string and let
be
the two entries of the major row of
.
First, we note
that
can be
uniquely represented by a finite continued fraction
,
where
is a nonnegative integer,
with
is a positive integer, and
is an integer
.
In other words,
, where
,
,
,
,
,
.
This fact is true for any positive rational number and its proof can be
found numerous number theory books
such as [HW88][Ste69].
Let
.
Then by an induction on
, we
can show that
.
If
, then
and
.
Suppose
. Assume that
. The case that
is similar. Let
be the major row of
.
Then
by Lemma 6.14 and Remark 6.4.
It suffices to show that if
, then
, and
if
, then
. Since
is not the
unit matrix, neither
nor
is zero. In any case,
. If
, write
, then
,
so
. If
, write
,
then
, so
.
Yao and Knuth [YK75]
showed that, for any positive integers
,

Let
, and
form the
major row of
. Therefore,
. Then
and so
. By Lemma 6.17,
, so
.
It follows that
is polynomial on
-average.
The isomorphism
is p-time computable but the distribution
does not dominate the distribution
with respect to
.
Thus,
fails to reduce
to
.
Proof., to the contrary, that
is (weakly) dominated by
with
respect to
.
Since
is one-one, there
is a function
such that
is polynomial
on
-average. Thus, there is an
such that

Hence,
. We will show that
this is impossible.
Let
. By the definition of
and
Lemma 6.17,
.
Let
be the sum of the entries
of the major row of
. Clearly,
. Hence, it
suffices to show that the expectation
is not bounded by any polynomial of
.
We may restrict our attention to
. Let
and
range over strings
of length
.
Define
.
Then there exists an
such that for every
,
. To see this, let
be the two entries of the major row
of
, and
. Then

Let
. Then
. Consider the function
.
Note that
, so
.
Thus,
is concave.
Since
, the chord
between the points
and
lies strictly above the
chord
between the points
and
. Clearly,
the center of the interval
coincides with the
center 1.5 of the interval [1,2], and so the center of
lies directly above the center of
.
Note that the arithmetical mean
of numbers 1 and
exceeds their geometrical mean
.
Thus,
. The desired
is
.
Thus,

It follows that
(
) and therefore
is not bounded by any polynomial of
.
Positive Matrix Correspondence
Let
be a nondeterministic Turing machine. Let
,
and
be proportional to
.
The size of
is
.
From the proof of Theorem 4.1, it is easy to see that
there is a Turing machine
such that
is complete for DistNP. Moreover, we can assume that
accepts
iff
halts on
.
Let
be a distribution proportional to
.
Proof.will reduce
to an appropriate
.
One might be tempted to simply take
and to use the
identity function as a reduction. By Proposition 6.19,
this approach fails to meet the domination requirement.
For every binary string
, let
be the positive integer
with binary representation
. If
,
let
. We construct
such that, on any input
,
first computes
, then simulates
on
.
Define a good-input domain
such that for all
,

where
represents the greatest common divisor of
and
.
Clearly,
is certifiable.
Let
be the number of positive integers that are relatively
prime to and less than
.
Then
[HW88].
Note that if
, then
so is
provided
that
. Hence, half of the integers counted by
,
and so
. Thus,

Therefore, the rarity function
,
which implies that
is nonrare.
By Lemma 6.16, for each
, there is a unique
positive matrix
with
being its first and major
column. Define the reduction
by

where
and
is the time that
needs to
convert
into
.
By the definition of
,
it is easy to verify that
halts on
iff
halts on
in
steps iff
halts within
steps on
iff
halts on
. So for all
,
iff
.
Next, we check that
is ap-time computable with respect to
. Since
,
can be computed in
time polynomial on
-average. Now we consider
, which is the time that
needs to compute
from . Clearly
is bounded by a polynomial
of
since
,
, and
are all p-time computabl
e.
It follows that
is polynomial on
-average.
We now check that
satisfies the domination property.
Note that
is one-one and so we can use Lemma 2.1.
We have
.
Note that
,
,
and
, so
.
Similar to the proof of Theorem 5.2, it is straightforward to
reduce
to the distributional positive matrix correspondence
problem by taking
in that proof.
The reduction is given by
, where
represents the instance of
of
.
The isomorphism
carries out the entire proof into the matrix
setting. This provides a proof to the following theorem.
Matrix Correspondence
We now turn our attention to unimodular matrices that may or may not
be positive. The technical terms we used to describe positive matrices such
as major columns, major rows, and the size of the matrix will apply
to unimodular matrices with respect to absolute values.
In this part, we use
to represent absolute value
unless otherwise stated.
Let
be a column. Denote by
the column of
and ,
and
. Let
be a unimodular
matrix. Write
.
Proof.. If one of the numbers
and
is positive and
the other is negative, then
, which is impossible. If the two numbers
are both positive, then
;
otherwise,
.
2. Suppose that
is not one of the four matrices listed, and
suppose that neither
nor
. Without loss of
generality, assume that
and
. (The other
case is similar.) Clearly, either
or
.
Hence, by assumption,
, a contradiction.
3. Suppose, to the contrary,
that
has two or more entries with value
.
If
occurs in both entries in the same row or in the same column,
then
divides
, which is impossible.
Thus, suppose that
occurs exactly twice in different columns and
different row. Without loss of generality, assume that
it occupies the second diagonal, i.e.,
.
(The other case is similar.) If
, then the
determinant is negative, which is a contradiction. Hence,
.
By part 1 above,
and
, a contradiction.
Proof.first show that for every two unimodular matrices
and
, there is an integer
such that
.
Suppose
(the case
that
is similar).
Then
, and so
.
Let
to obtain the claim.
Suppose
, then
, and do
for some
.
The claim is therefore obvious.
To prove the lemma, it suffices to consider the case when
is
the major column. (If instead
is the major column,
then the unimodular matrix
will have the major column on
the left.) Moreover, it suffices to consider that
is positive, for
otherwise, the unimodular matrix
will have the major
column on the left and the major column is positive.
Let
be the major entry of
, i.e.,
.
Let
be another matrix with major column
.
Then as seen above, there is a
such that
.
Since
,
. If
, then
; if
, then
. Note that if
(respectively,
) then
is the major column of
(respectively,
). Since by assumption that
is positive,
and
. By Lemma 6.22(1),
is
either positive or negative. If
is positive (respectively, negative),
then
is negative (respectively, positive).
The size of a unimodular matrix
, denoted by
,
is defined to be the length of
in binary notation. We assume that all unimodular
matrices of the same
size have an equal chance to be selected.
Proof.
be the collection of unimodular matrices
with
.
Let
be the collection of matrices
such that
the major row of
is positive. For every
,
exactly one of the two matrices
and
belongs to
. It remains to show that the probability of a random matrix
from
being positive is
.
Since the major component of an
matrix is greater than 1,
the minor component of the major column cannot be zero.
Let
be the collection of
matrices such that the
minor component of the major column is positive.
For every
matrix
, let
be the result of multiplying
by
the diagonal of
that contains the minor component of
the major column. Exactly one of the two matrices
,
belongs to
. It follows that
contains exactly
one half of the elements of
. By Lemma 6.23,
the probability of a random
matrix being positive
is
.
Clearly, the identity function reduces the distributional positive matrix correspondence problem to the distributional matrix correspondence problem.
