Jul 14, 1992 - Time, Numerically. Stable. Integer. Relation. Algorithm. Hclam-xn. R. P. Ferguson and David. H. Bailey. RNR. Tcctmical. Report. RNR-91-032.
A Polynomial
Time,
Numerically
Hclam-xn
Stable
R. P. Ferguson
RNR
Tcctmical
and
Report
July
Integer
Relation
Algorithm
David
H. Bailey
RNR-91-032
14, 1992
Abstract Let
x = (xl,x2,'"
relation
if there
Beginning
,::,_) be a vector
exist
in 1977
integers
sever.d
al not
of real all zero
algorithms
(with
numbers, such
proofs)
ai given x. The most e:l]cient of these existing run time and the precision required of the input) numerically. recover
It often
relations
requires
a numeric
in mo&_st-sized
test
a relation
in a nurrber
this
algorithm
employs
the
numerical
difficulties
stability
admits
algorithms obtained
Ferguson Mail Stop
is with Internet: T045-1,
that stable
plague
other
runs
he].amanf©super, Field,
with
discovered
to recover
O. the
(in terms of very unstable
of digits
reduction relation
lower
to reliably
CA 94035.
run
Bailey Internet:
Center, is with
in n.
procedure, algorithms. times
are numerically
Research org.
by a polynomial
17100 NASA
dbailey©nas
Because
it is free Furthermore,
on average
can be used to prove
this algorithm
th,, Supercomputing
Moff,_tt
matrix
this stability
using
an integer
relations, which we have named the PSLQ algorithm terminates
is bounded
integer
implementation
in use. Finally,
from computer
MD 20715.
a numerically
an effic;ent
currently
been
relation algorithms drawback of being
level in the thousands
for finding integer in this paper that
of iterations
that
to possess
+ a2x2 + "" + a,_:r,_ =
problems.
We present here a new algorithm the "PSLQ" algorithm. It is proved with
alXl
have
integer has the
precision
x is said
that
that
than
relation
from its other
bounds
accurate.
Science Alnes
Drive,
Research
.nasa.gov.
Bowie, Center,
1. Introduction Let x = (Xl,X2,-.. relation
if there
By an integer
,x_) be a vector of real numbers, integers ai not all zero such that
exist
relation
algorithm,
computer implementation to produce bounds within One
application
we mean
x is said to possess an integer alxl + a2x2 + ... + anx,_ = O.
an algorithm
that
is guaranteed
has sufficient numeric precision) to recover which no integer relations can exist.
of an integer
wherein one determines what words, a subset sum problem
relation
algorithm
is to solve
"subset
application is discussed in [11] and [5]. of an integer relation algorithm is to determine
a certain
constant,
mathematical
of a polynomial
x = (1, _, _2,..., relation resulting
of degree
a,_-l)
precision).
and
whose
value
can
be computed
This
can
be done
n or less.
applying
an integer
Even
if no relation
for example,
that
relation
is found,
a calculation
the constaslt
cannot
n or less whose coefficients are smaller application is discussed in [4]. The problem Euclid,
of finding
who gave
terminates,
yielding
Poincare,
Minkowski,
iterative algorithms have been found. first
discovered
integer
to high
been
relation
of other
a bound
among
or not
precision, If any
Bernstein, to work with
(Ferguson) relation
newer
require much algorithm. 2. The
HJLS
Tile
HJLS respects.
for HJLS
algorithms less precision
are
among
and
in the input
faster
of approximate
x vector
been
of their
couuterexamples
mentioned
above
discovered,
was
including
a
algorithm [12], the "H,ILS" the "PSOS" [4] algorithm.
implemented
to recover
none
Jacobi,
in 1977 [7]. In the intervening
have
when
by
either
by Euler,
However,
properties
R. Forcade
This
was first studied
sequence
others.
desired
can
of degree
to two real numbers,
for n > 3, and numerous
the
algorithm
by the algorithm.
an infinite
algorithms
significantly
relation
integer then the machine
any polynomial
a set of real numbers applied
is a
a by setting
to x.
for n > 2 has been attempted
Brun,
algorithm
satisfy
established
when
or produces
proven
integer
with an integer
non-recursive variant of the original algorithm [8], the "LLL" algorithm [10] (which is based on the LLL algorithm), and
several
whether
for a constant
algorithm
possibly
which,
relation,
by one of the authors
a number
relations
of this problem
Perron, have
than
algorithm
all exact
The generalization
The
integer
an iterative
relations.
These
problems,"
is found to hold (within the limits of the available machine precision), ai are precisely the coefficients of a polynomial satisfied by a (to within
establish,
years
sum
the ai and
subset of a certain list of integers has a given sum. In other is an integer relation problem where the relation coefficients
ai are zero or one. This Another application root
(provided
the integers
the
oil a computer
relation
than
and
the original
Algorithm algorithm
is superior
For one thing,
to recover
a relation
among
these
it has been
proven
existing
is bounded
by a polynomial
that
integer
the number
relation
algorithms
of iterations
in n [10], whereas
this property are lacking for the other algorithms. Further, the HJLS algorithm to be the most efficient of these algorithms in terms of its ability to recover the
in
required proofs appears relation
of
satisfiedby an input vectorknownonly to limited precision.Finally, basedon the authors' experience,HJLS appear,lto requirethe lowestaveragecomputertime to recovera relation amongpreviouslyexisting algorithms. Unfortunately, the H.JLSalgorithm has one seriousdrawback: it is extremely unstable numerically. Even for modest n and small relations, enormous numeric precision
is often
17-long
required
vector
(1,a,
(1, 0, 0, 0, -3860, required
for the
must
outward
where
relation
sympto_n
a
with
integer
= 31/4-
This
the authors is that
in a matrix,
one which
working
run
with
precision.
H.ILS
has
level of over
in a H.ILS run
entry
in the current
vector
consider the
only 100 digits
H.ILS,
precision
failure large
a computer
For example,
21/4 .
using
a numeric
of numerical
is that
properly.
0, 0, 0, 1). Although
to be recovered
of _n extremely
failure
to work
0, 0, 0, -20,
be performed
to the nearest
of numerical
algorithm
a2, ...,at6),
due to the appearance exactly
the
0, 0, 0, -666,
as input
computations One
fcr
relation
or so of a are
have
found
10,000
digits.
the
program
cannot
Another
produces
the
aborts
be rounded
outward
a bound
that
symptom
on the
norm
of
possible relations that e::cludes a relation known to exist. In this second type of failure, comparison with runs us,ng higher precision reveals that some matrix entries had become so corrupted
by numerical
error
that
all significance
In some trials, computer runs using recovering a relation usii,g only moderate would
not be recovered
precision. relation
But
at that goc.d
will be missed
If one
asks
computer many
such
run
cases
here means
what using
appears that
point
fortune
in the cannot
a_ the point lewd
where
of precision
computer
of digits,
are the same
on,
it should
lost.
if it were performed and
it is just
with higher
as likely
that
the
before
the
be recovered. for a lnodest-sized
to one using
to te thousands
run
be relied
is required
H.]LS is identical
all decisions
had been
the H.ILS algorithm succeed "by accident" in levels of numeric precision, whereas the relation
"infinite"
based
the
precision,
authors'
and all relation
problem then
the
experience.
bounds
are the same
answer
in
"Identical" (to 8 digits
or so) up to and including the point of recovery. Cases that require very high precision are infrequent for n < 1(* but are quite common for larger n. There does not appear to be any
a priori
means
of d,_termining
the
required
H.ILS working
precision
level for a given
problem. As one would memory
and
than other disadvantage that
expect,
processing
these
very high levels of numeric
time,
although
amazingly
previously kaown integer relation of the H.]I_S numerical instability
a relation
of a certain
size has
been
precision
enough,
excluded,
because
that
the sequeuce
previous
run,
but one can never
The root cause from the fact that which
is known
of norm
bounds
be certain
is the same of the
large amounts
H,]LS is still faster
algorithms. However, is that one can never
after running H,ILS for a while might be completely corrupted confidence in a bound Jesult can be enhanced by increasing verifying
require
the
of
on average
the most significant know with certainty
bound
that
is obtained
with numerical error. One's the numeric precision and
up to the
point
determined
in the
result.
of this numerical instability is not known, but it is believed to derive H.JL'_ is based upon the Gram-Schmidt orthogonalization algorithm,
to be ut merically
unstable
[9].
3
3. The PSLQ Algorithm Recentlyoneof the authors (Ferguson)discovereda new polynomial lation
algorithm.
This
sum of squares
algorithm
scheme
has been
like the PSOS
a LQ (lower trapezoidal--orthogonal) The PSLQ algorithm exhibits its ability
to recover
avoiding does
not
appear
numerical double
precision
a working
in x vectors
and multiprecision range
precision
obtained
Let x E R '_ be a nonzero
on
arithmetic sizes.
is only
computer
runs n-tuple
to only limited average, that
than
0
hi,j
--
Si+l si
hi,.i
=
-
xix'i
while completely of PSLQ n,
the
a combination
of
to be faster
than
HJLS
that
by using
that
of the
input
data,
on
bound
are reliable. x = (Xl,X2,'
.. ,Xn_l,X,_).
Define
2
=
including
one can show
squares, s.i , for x by %. = _j 2/x/_
= 1.1547....
Suppose
we are given
three
matrices,
H, A, B,
where H is a n × (n - 1) lower trapezoidal matrix and A and B are n x n integral matrices, with B = A -1. An iteration of the algorithm PSLQ is defined by the following three steps. 1. Replace 2. Select
H by DH. an integer
3. Replace Theorem. bination
H by RjHG.i,
norm with
of an)
1 such that
A by RjDA
such relation
the following
1. A relation
and
m.
set of three
for x will appear
iterations 2. The
1 < j < n -
-yJlh.;,jl >_ "/Ih_,,[ for all i, 1 < i < n - 1.
B by BERj.
Fix 7 > 2/,/3 and set 62 = 3/4 - 1/72. Suppose some integral of the entries of x E R" is zero, so that x has an integer relation.
be the least beginning
j.
Normalize matrices:
as a column
x so that
Ix] = 1 and
linear comLet M > 1 iterate
PSLQ
H = H=, A = I,_, B = I_. Then
of B after
fewer
than
_n2(n
+ 1)log(Mn
2)
of PSLQ.
norm
of such
a relation
for x appearing
as a column
of B is no greater
than
v/-;ilHIIBPIM. 3. If after B, then
a number there
(,f iterations
are Jlo relations
of PSLQ of norm
no relation less than
the
has yet bound
appeared 1/IHI.
in a column
of
6.
Proof
of the
Suppose invertible
PSLQ
Algorithm
m E Z '_ is a relation C E GL(n,
Z) and
1 _< Icm'l = [cpmq Note Gj.
that
after
Perform
m, including 1