ALGORITHMS FOR DIOPHANTINE EQUATIONS

PROEFSCHRIFT

ter verkrijging van de graad van Doctor aan de Rijksuniversiteit te Leiden, op gezag van de Rector Magnificus Dr. J.J.M. Beenakker, Hoogleraar in de Faculteit der Wiskunde en Natuurwetenschappen, volgens besluit van het College van Dekanen te verdedigen op woensdag 6 januari 1988 te klokke 14.15 uur
1

BENJAMIN

MARPIUS MARNIX DE WEGER

.-

geboren te Delft in 1958

1987

Centrum voor Wi.?ktinde en Informatica, Amsterdam

Samenstelling van de promotiecommissie:

promotor referent overige leden

prof. dr . R. Tijdeman dr. F. Beukers prof .dr. P.L. Cijsow prof-dr. G. van Dijk prof-dr. J.P. Hurre prof .dr. H.N. Spijker

Chapter

1.

Introduction. Algorithms for diophantine equations. The Celfond-Baker method.

g 1.1.

5 1.2.

g 1.3. Theoretical diophantine approximation. 5 1.4. Computational diophantine approximation.

5 1.5. The procedure for reducing upper bounds.

Chapter

2.

Preliminaries. Algebraic number theory. Some auxiliary results. p-adic numbers and functions. Lower bounds for linear forms in logarithms. Numerical methods. Algorithms for diophantine approximation. Introduction. Homogeneous one-dimensional approximation in the real case: continued fractions. Inhomogeneous one-dimensional approximation in the real case: the Davenport lemma. The L3-lattice basis reduction algorithm, theory. 3 The-L -d&Xicekasis reduction algorithm, p~actice algurithm. Homogeneous multi-dimensional approximation in the real case: real approximation lattices. Inhomogeneous multi-dimensional approximation in the real case: an alternative for the generalized Davenport lemna.

5 2.1.

5 2.2.
5 2.3.
5 2.4. 5 2.5.
Chapter 3.

g 3.1.

8 3.2.

Finding all short lattice points: the Fincke and Pohst

Het onderzoek dat geleid heeft tot dit proefschrift stond in het kader van het onderzoeksproject Diophantische Approximaties van de Nederlandse Stichting voor de Wiskunde (SCIC), dat gesubsidieerd werd door de Nederlandse

Inhomogeneous zero-dimensional approximation in the p-adic case. Homogeneous one-d&nsional approximation in the p-adic case: p-adic continued fractions and approximation lattices of p-adic numbers.

Organisatie voor Zuiver Uetenschappelijk Onderzoek (ZWO).

Hoqogeneous multi-dimensional approximation in the p-adic case: p-adic approximation lattices. Inhomogeneous one- and multi-dimensional approximation in the p-adic case. Useful sublattices of p-adic approximation lattices. Chapter Chapter 8.

The standard example. Tables. The Thue equation. Introduction. From the Thue equation to a linear form in logarithms. Upper bounds. Reducing the upper bound.

0 8.1.

4. S-integral elements of binary recurrence sequences.

Introduction. Binary recurrence sequences. The growth of the recurrence sequence. Upper bounds. Symmetric recurrences: an elementary method.

8 8.2. 0 8.3.
8.4.

5 4.1.

3 4.2.

5 8.5.

An application: integral points on the elliptic curve

y2-x3-4~+l. The Thue-Mahler equation, an outline.

4.3. 4.4. 4.5. 4.7. 4.9.

References. Samenvatting. Curriculum vitae.

8 4.6.

A basic lemna. and some trivial cases.

The reduction algorithm in the hyperbolic case. The generalized Ramanujan-Nagell equation.

Q 4.8. The reduction algorithm in the elliptic case.

84.10. Chapter

A mixed quadratic-exponential equation.

The inequality
0

5.

<x

y < y6

in S-integers.

5 5.1.

Introduction. Upper bounds for the solutions. Reducing the upper bounds in the multi-dimensional case. Tables. The equation Introduction. Upper bounds. The p-adic aperoximation lattices.
-

5.2. 5.4. 5.5.

8 5.3. Reducing the upper bounds in the one-dimensional case.

Chapter

6.

in S-Integers

Q 6.1.
0 6.2.

9 6.3. 3 6.4.

Reducing the upper bounds in the one-dimensional case. Reducing the-upper bounds in the multi-dimensional case. Examples related to the abc-conjecture. Tables. The sum of tvo S-units being a square. Introduction. The case

6.5.

8 6.6. 8 6.7.
Chapter

7.

5 7.1.
0 7.2.

0 7.3. 8 7.4. 5 7.5.

Tovards generalized recurrences. Towards linear forms in logarithms. Upper bounds for the solutions: outline. Upper bounds for the solutions: details. The reduction technique.

5 7.6.

7.7.

1.1.

Algorithms for diophantine equations.

This thesis deals with certain types of diophantine equations. An equation is a mathematical formula. expressing equality of two expressions that involve one or more unknowns (variables). Solving an equation means finding all solutions. i.e. the values that can be substituted for the unknowns such that the equation becomes a true statement. An equation is called a diophantine equation if the solutions are restricted to be integers in some sense. usually the ordinary rational integers (elements of that. Examples of diophantine equarions that will be studied in this thesis are x2 + 7 (the (5c.n)
Z ) or some subset of

2 " equation. having only the solutions given by

Ramanujan-Nagell

(1.3).

(3.41, (5.5). (11.71, (181.15) , see Chapter 4);

(a purely exponential equation, having only the solutions (2.1.01, (3.1.1). (S.3.1). (7.1.31

(x,y,z)

(1.0.0).

see Chapter 6);

(an elliptic equation. having only 22 solutions. of which the largest are (x.y)

(1274.f45673)

see Chapter 8). The three examples mentioned here are

only exaatples of classes of equations which we study. We also study (in Chapter 5) a diophancine inequality (a formula expressing that one expression is larger than another, vhere solutions are again restricted to integers). In the following discussion the statements about diophantine equations also hold for this inequalicy.

What the equations treated in this book have in c o m n is that they can all be solved by the same method. This method consists essentially of three

parts: a transformation step, a application of the Gelfond-Baker theory, and a diophantine approximation step. We explain these steps briefly. First, one transforms the equation to a purely exponential equation or inequality, i.e. a diophantine equation or inequality where the unknowns are all in the exponents, such as in the second example given above. Each type of diophantine equation needs a particular kind of transformation. so that it is difficult to be more specific at this point. In some instances, such as in the second example above, this transformation is easy, if not trivial. In other instances, as in the first example above, it uses some arguments from algebraic number theory, or, as in the third example above, a lot of them. In general, such a purely exponential equation has the form

real, or complex sense, or in the p-adic provides effectively computable lower

sense). bounds

The Gelfond-Baker for the absolute

theory values

(respectively p-adic values) of such linear forms in logarithms of algebraic numbers. In many cases these bounds have been explicitly computed. Comparing the so-found upper and lower bounds it is possible to obtain explicit upper bounds for the solutions of the exponential diophantine equation or inequality, leading to upper bounds rather general nature. We remark that many authors have given effectively computable upper bounds for the solutions of a wide variety of diophantine equations, by applying the method sketched above. For a survey, see Shorey and Tijdeman [1986]. Often these authors were satisfied with the knowledge of the existence of such bounds, and they did not actually compute them. If they computed bounds. they did not always determine all the solutions. In this thesis, solving an and a corresponding pure y exponential inequality looks like equation will always mean: explicitly finding all the solutions. After the second step, the problem of solving the diophantine equation is reduced to a finite problem, which is the treated in the third part of the where
0 < 6 < 1 and t. si. c i s aij. 6 are constants with t. si E N belong to some algebraic extension of 0 and where the ci, aij "ij are the unknowns in X .,_We-now suppose that the number of terms: t on the left

for the solutions of

the original

equation. This second step, unlike the first (transformation) step, is of a

method. Namely, since we have found explicit upper bounds for the absolute values of the (integral) unknowns, we have to check only finitely many possibilities for the unknowns. However, the word f i n i t e does not meKn the same as small or t r i v i a l . In fact, the constants appearing in the bounds that the Celfond-Baker theory provides for linear forms in logarithms are rather large. Therefore, in practice the upper bounds that can be obtained in this as way for the solutions of purely exponential equations can be for instance --

hand side is equal to 2. This restriction is essential for the second step. in which we use results from the so-called theory of linear forms in logarithms, also known as the Celfond-Baker theory. (Some special exponential equations uith more terns can also be treate&-by the C e 1 f o n d ~ B d e e - m by ~ reducing them to exponential inequalities with two terms. cf. Stroeker and b Tijdeman [1982], Alex [1985a], [I985 1 , Tijdeman and Wang [I9871.)

large a s - p . This is far too large to admit simple entameranon of a11 the possibilities, s v e n with the fastest of computers today. Notwithstanding, it is generally assumed that the upper bounds found in this

An exponential equation or inequality such as (1.1) or (1.2) gives rise to a

l i n e a r form in l o g a r i t h

way are far from the actual largest solution. Therefore, it is worthwile to search for methods to reduce these upper bounds to a size that can be more
-

easily handled. Often it is possible to devise such a method using directly certain properties of the original diophantine equation, for example that where the large solutions must satisfy certain congruences modulo large numbers, or are algebraic constants, and the ni are integral unknovns. some reciprocity condition (fqr some examples, see Grinstead [1978], Brown (19851, Pinch [1987], and Path6 (19831). The disadvantage of such rathods is that they work only for that particular type of diophantine equation. so that

Pi

Here, the logarithms can be real or complex, or can be p-adic. This relation between equation and linear form in logarithms is such that for a large solution of the equation the linear form is extremely close to zero (in the

in general for each type of equation a new reduction method must be devised. It would therefore be interesting to have methods for reducing upper bounds for the solutions of inequalities for linear forms in logarithms. They would be useful for solving any type of diophantine problem that leads to such inequalities. Such methods are provided by that part of the theory of diophantine

A rheoretfcal method is capable of proving results that hold for some large
set of values of the parameters. The Gelfond-Baker theory is of a theoretical nature, since it yields upper bounds for the solutions of many equations in terms of their parameters. Other examples are reciprocity. that shows that odd, at least 5 . solutions ID1 < 296 (x.n) x2 the theory of quadratic D is

D
-

--

2"

has no solutions at all if

and not congruent to of x2 + D

7 (mod 8) , and the theory of

[1981] used n < 435 to

hypergeometric functions, which then moreover n < 18

approximation that is concerned with studying how close to zero a linear form can be for given values of the variables. Recently important progress has been made in this field, leading to practically efficient algorithms. which can be employed for many diophantine equations to show that in a certain interval [X1.XO] X log , no solutions exist. Usually Xl is of the order of

- 2"

Beukers

satisfy 2 -2loglDl

+ 10-21oglDI

show that the and if

. Theoretical

methods are often

too general to be able to produce all the solutions of a given equation.

An algorithmic method is a method that is guaranteed to work for any set of

values of the parameters, but has to be applied separately to each particular set of parameter values, in order to produce all the solutions. The methods used in this thesis are mainly of such an algorithmic nature. For the equation x2 + D a 2 " (and actually for a more general equation) we will aboveall the give an algorithmic method in Chapter 4. In fact, since Beukers' made algorithmic by providing a simple method of enumerating

magnitude of

the theoretical upper bound for the solutions is substituted. a new upper bound X1 is found. For many equations the upper bound Xo is well within reach of practical application of these algorithms, within only a few minutes of computer time. This thus leads in for finding all the solutions of many types of diophantine equations. for which alternative methods have not yet been found or employed with success. It is mainly in this third part of the method that new developments can be reported. The arguments we use in the first and second parts are mainly clas$ical. They are applied to types of equations that have been studied before. and also to new types of equations. The method outlined above. and used in this thesis to solve many examples of sense lies between ' a d explain below. parameter hocn methods and "theoretical" methods. Let a given. As This we shall practice to methods

. When for

Xo

mentioned result provides a small upper bound for the solutions, it can be solutions below the upper bound. In order to make the Gelfond-Baker theory algorithmic, enumeration of all possibilities is impractical. Therefore more ingenious ways of determining all the solutions below a large upper bound

have to be fougd. We remark that Beukers' method for the more general - n . - p equation x2 + D also has an ad hoc aspect. since it does not work for any value of p . Our method of Chapter 4 does not have this disadvantage. An ideal towards - which one might strive in solyLng dioph+riine equations is
-

var ieus-iEophmthe

to devise a computer algorithm, which only has to be fed with the parameters of the equation, and after a short tile gives a list of all the solutions as output. The user should have a guarantee (in the strictest, mathematical sense of proof), that no solutions have been missed. At first sight the lethod outlined above, and described in this dissertation, seeas to be a good candidate to be developed into such a general applicable algorithm. Namely, the second step is of a quite general nature. providing

set of diophantine equatrons with an unspecified an example of such a set, consider the x2 equation

in it be x, n

generalized Ramanujan-Nagell parameter. and

+ D

2 "

where

is a

are the unknowns.

--

An ad hoc method is a method for solving the equation for specific values of
the parameters only. However. it map not work at all for other than these particular values. The first example of solving an equation of the type 7 x2 + D 2" occurring in the literature is that by Nagell [I9481 of D

upper bounds for exponential diophantine equations that are explicit in the
parameters of the equation. Also +e third step, the algorithmic diophantine approximation part, works in principle for any set of values substituted for

The method he used is of an ad hoc nature, since it depends heavily on the special choice of

7 for the parameter D

the parameters. However, the computations have to be performed separately for

methods, because the application of the theory of linear forms in logarithms each particular set of values. The main difficulties in devising such a 'diophantine machine' are in the first part of the method outlined above, especially if some algebraic number theory is used. Developments taking place in the- theory of algorithic algebraic nrmber theory on computing fundamental units and on finding factorizations of prime numbers in algebraic extensions, are of importance here. We believe that when suitable algorithms of this kind are available, it will be possible in principle to make such a 'diophantine machine'.
he
1 5 .

is more complicated for these equations, and moreover the upper bounds are essentially x too large. An example of this kind is the Catalan equation
a
A.

in integers a. b, x, y , a11 2 2

Catalan conjectured in 1844

that this equation has only the solution (a,b.x,y) (3.2.2.3) Tijdeman [1976] proved that the solutions of the Catalan equation are bounded by a computable number. This number can be taken to be exp(exp(exp(exp(730)))) , according to Langevin [1976]. However, we fail to see how the wthods that we describe in the forthcoming chapters can be applied for completely solving the Catalan equation.

ie
,

generality of such an algorithm is restricted by the generality of the first step. the transformation to the linear form in logarithms. In this thesis we use computer algorithms only if the magnitude of the computational tasks makes this necessary, and keep to *manualm work otherwise. In this way we also try to keep the presentation of the methods lucid. The reader should be aware of the fact that the computer programs and their results are part of the proofs of many of our theorems on specific diophantine equations. It is however impossible to publish all details of these progrand computations. The interested reader may obtain the details from the author by request, and is invited to check the computations himself. The book by Shorty and Tijdeman [I9861 gives a good survey of the diophantine equations for which computable4per bounds for the solutions can be found using the Celfond-Baker method ( ; e e and Schintel
-

hother dio~hantineequation, that for centuries has attracted the attention n of many mathematicians, is the Fermat equation x + yn zn in integers It is conjectured to have no x, y, z, n , with n r 3 and x - y - z + 0 solutions. Faltings [I9831 proved that for fixed n the number of solutions is finite. His proof is ineffective, and not based on the Gelfond-Baker theory. g he Gelfond-~aker theory seems not to be strong enough to dela with the Fermat equation in its full generality, not even if n is fixed. For partial results on the Fermat equation that have been obtained using this theory. see Tijdeman [1985] and Chapter 11 of Shorey and Tijdeman I1986). Ue remark that for many diophantine equations recently important progress

has been - - made in determining upper bounds for ,the number of solutions. See Evertse [1983] and Evertse, Gy6ry. Stewart and ~ijdeman[1987] for a survey.

also Shorey, van der Poorten, Tijdenan (19821). Some of these

(19771, and Stroeker and Tijdeman

equations can be completely salved by the wthods described in chis thesis, arPong wfiich khere are purely- exponential e q u a t i ~ m , e q e i o n s involvi binary recurrence sequences. and Thue equations and hue-Hahler equations. Especially the latter t w o are of importance in'various other parts 5 number theory. various For example. they are the key to solving equations arising in algebraic number theory. The

These results are often remarkably sharp. but ineffective, so that they cannot be used for actually finding the solutions.
. , - --- - TO conclude this section we give an overview of the reraining chapters of

ord dell equations and el fond-Baker

this thesis. It is divided into three parts: Chapter 1 is introductory, Chapters 2 and 3 give the necessary preliminaries, and Chapters 4 to 8 deal
with various types of diophantine equations.

m t h o d was used to actually solve a diophantine equation for the first time in tke work of Baker and Davenport (19691, who solved the system of diophantine equations 3.2-2-y 2 8-?-7-z

Sections 1.2 to 1.5 give a short introduction to the Gelfond-Baker theory. diophantine approximation theory, the algorithmic aspects of diophantine approxiPlation, and the procedure for reducing upper bounds, respectively, for

the no-specialist. Chapter 2 contains the preliminary results that we need . f , f 0 1 algebraic number theory a8d from the theory of p-adic nrmbers and
f;mctions. and quotes in full. detail the theorfrom the Gelfond-Wer
0 ~ which

Other equations occurring in the literature for which upper bounds for the solutioru can be computed, cannot be treated as easily by our a l g o r i ~ i c

we need. It concludes with s o w remarks on numerical retho&.

Chapter 3 gives in detail the algorithms in the field of diophantine approximation theory that we apply in the subsequent chapters. The remaining Chapters 4 to 8 are each devoted to a certain type of diophantine equation. Let Let only. Chapter 4 deals with elements of binary recurrence sequences ('generalized Fibonacci sequences") that are in equation x2 + k E S
(

b 1. published in Tzanakis and & Weger (19871, and in de Weger [I987

1.2.

The Celfond-Baker method.

pl,

. . .,

ps

be a fixed set of distinct primes. pl.

be the set of positive integers composed of primes

....

ps

In Section 1.1 we have explained that before applying the Celfond-Baker method to some diophantine equation, the equation should be transformed into a purely exponential diophantine equation or inequality with not too many terms (cf. (1.1). (1.2) ) . In this section we sketch the arguments from the

. and gives their application to mixed

Gelfond-Baker theory that lead to upper bounds for the variables of this exponential equation/inequality. Let us first treat the case of the inequality (1.2). have the form It may be assumed to

quadratic-exponential equations, such as the generalized Ramanujan-Nagell k fixed) . The diophantine approximation part of this chapter is interesting for two reasons: the p-adic approximation is very simple. and in the case of the recurrence having negative discriminant. a nice interplay of p-adic and real/complex approximation arguments occurs. The research for Chapter 4 was done partly in cooperation with A. Petha from Debrecen. The results have been published in Petha and de Weger [I9861 and de b Ueger [I986 1 . Chapter 5 deals with the diophantine inequaliq are in
x,
y,

where the

ai

are fixed algebraic numbers. In the examples we mi

- maxinil . and
tail

Cop 6 i

are

positive constants.

study, we encounter one of the

S z

. and

0 < x

y < y6

. where
z

6 E (0.1)

is fixed. Chapter 6 deals with

x + y

- .

X.

following two cases: either all the real case, if N

are real, or

for all

In

where

is large enough, the linear form in logarithms

S , which can be considered as the padic analogue of the

A

. - - are the simplest examples of inequality of Chapter 5. These two equations

log laO

diophantine equations that can be treated by our method. Since they are already purely exponential equations. the first (transformation) step is trivial. -So the linear forms in logarithms are directly related to the equations-thewe lves, Therefore they serve as good examples_-to ge U clear understanding of the diophantine approximation part of our method. results of &se chapters have been published in de Weger [1987~]. x The
- - .
-

m s t satisfy

---

- -

--

Ci linear form

for some

. In the complex case,-the same inequality (1.3)

. -

follows for the

Chapter 7 studies the equation equation. studied in Chapter 6 .

- .
r2

where

x, y E S

. and

z E Z

-..'

Log . a

1ni-Log ai + k-Log(-1)
i-1

This equation is a further generalization of the generalized ~amanujarr-Nagell

.
-licit

i-( Arg no +

1 ni-Arg ai + k i i
i-1

In Chapter 8 a procedure is given to solve Thue equations. that works in principle for Thue equations of any degree. It is applied to find all x3 4.x + 1 We also oention integral points on the elliptic curve y2 briefly how Thue-Hahler equations can be dealt with. This chapter has been written jointly with N. Tzanakis from Iraklion. The results have been

where the Log and Arg functions take their principal values. Now we can *ply one of the many results ,from the Gelfond-Baker theory. giving an lower bound for IAl . , i n terms of

N , e.g. the follwing theorem.

THEOREM 1 .I. (Baker [1972l),

Let

A : be a s above. There e x i s t computable

Applying (1.4) and Theorem 1.2 for a11 possible

Ci

, ! C we obtain constants .

cons t a n t s

C1, C2 , depending on the

ai o n l y , such t h a t i f

then

with

We know that N <

A
C

+
1

.
6

Combining (1.3) and Theorem 1.1 we obtain

C .H 5 P bound for H
If H
S

for some constant- C5 , then this immediately yields an upper

If

H > C -H

+
N

conjugate of the ai. log Ci)

Bj

then it can be shown that there exists a 5 P ' denoted with a prime sign, for which

+ C2 6

log N

.
for a constant
C6

It follows that

is bounded from above. (cf. the proof of Theorem 1.4, pp. 45-49, of Shorey and Tijdeman [1986]). Now we can apply Theorem 1.1. This yields

Next. consider the exponential equation (1.1). We can write it as

where the the resp. i, j

ai.

Bj

are fixed algebraic numbers. Let where i. j

Ini1 , Imj 1

be the maximum of HP run through the set of indices for which ai

It follows that H

is bounded from above.

Bj

are non-units . Let

be the maximum of the

I n1 , m p

where

run through the set of all indices. Suppose that

is a rational such that

If it happens that none of the ai, application of Theorem 1.2 suffices.

Bj

are units. then of course the

prime lying above

Bj

for some j n aii-1) s

. There

are constants c1, c2

ord ( a p o ' i , l
i

We remark that, in order to be able to completely solve a diophantine

C ,

+
i

equation, it is clrucial that all constants can be computed explicitly. Therefore K c a n only use' the bounds from the Gelfond-Bakcr theory that are
may write down a p-adic linear
-

Assuming that

ord (a ) P i form in logarithms

for all

. we

completely explicit. We give details of such theorems in Section 2.4.

. .

--

4 2 _lo5,aO

+ -i-1 x.ma -lo$ai,

cl
+

----

1 . 3 .

Theoretical diophantiire CppS6ximaf;ion. this section we briefly mention some

for which, if m

is large enough. it follows-that

5

In
(1.4)

--

results

f r o ~ diophantine

ord (A) P

C2-m~

approximation theory, thus giving a background to the next section. We refer to Koksna [1937], Cassels [I9571 (Chapters I and 111) and to Hardy and Wright (19791 (Chapters XI and XXIII), for further details. The simplest form of diophantine approximation in the real case is that of approxhation of a real number 6 by rational numbers p/q It is vell

We are now in a position to apply the following result from the p-adic Celfond-Baker theory. Here. N v

naxlnil

a Let
h

, p

be as above. ai

There e x i s t computable constants

p

C3, C4 , depending o n l y on t h e

and on
C

known that if
4p.q) E Z * I . _

such t h a t if

A z 0

then

with

is irrational, $hen there exist infinitely many solutions (p.q) 1 ofithe diophantine inequality

ord (A) < C3 P

C4-log N

6 are such ~ l l convergents from the continued fraction expansion of solutions. The convergents are simple to compute f6r any particular 8 E R .

L~ - p i J < Q ' " ~

for

1.

.... n .

As we shall show in Section 1.4, the L3-algorithm may be applied to this

One way o f generalizing this is to study simultaneous approximations to a set of real numbers ... en i.e. rationaL--approximationsto 6L all

general form. We actually compute solutions of systems of inequalities that are slightly weaker in the sense that the right hand side is multiplied by a constant larger than 1. We now consider inhowgeneous approximation. This means that for all there is an inhomogeneous tern pi in the linear form i Li , viz. i

el.

having the same denominator. It is well known that the system of inequalities

has infinitely many solutions in the case n

(pl. ...,pn.q )

if at least one of the

Bi

is Li

irrational. But it is much harder to find solutions of such inequalities than

-Pi

Some multi-diaensional continued fraction algorithms

m I q j - b i j for j-1

- 1.
i

....

n .

have been devised (cf. Brentjes (19811 for a survey), but they seem not to have the desired simplicity and generality. We shall see later how we can apply the so-called ~ ~ - a l ~ o r i tto h athis ~ problem. Another way of generalizing the simplest case of diophantine approximation is to study linear forms. such as

Again, there exists a constant c

such that the system

Li

- pi

I <

C-Q

-m/n

for

1. pi m

..., n ,

under some independence condition on the is Kronecker's theorem. The simplest case

-n-1

and

dij

has a solution. This

comes down to

Lwhere 8 unknowns in

m lqj-ej j-1 To conclude this section, we remark that there is a p-adic analogue of this
-

are given real numbers. and ql, .... % are the Z . Put _4 maxlqil -. A classical theorem guarantees the of the inequality existence of a solution .(p,ql,...,%)
6

theory of diophantine approximation, founded by Uahler and Lutz. replace -in the above considerations R 'by CI the absolute value value

the p-adic

1 - Ip

--

Note that the case r q1 . ~ 1 in ~ this 0 e be low.

(pl. .pn.ql. .q .qJ then the by any c o w e x noW p l . . ..pn.ql. m p-adic analogues of the theorems of Uinkowski and Kronecker are essentially
-)

--.

. ..

I-1

If we by

and

P * the measure Q

for an approximation

--

1 becomes our first-i~quaITty On divtding bythe 13-algorithm is very useful. as we shall See

analogous to the above mentioned results in the real case. See Koksna (19371 for references to Uahler's work, and Lutz (19511.
-

-----

--

--

We can incorporate the NO generalizati~nsabove in a further generalization. that of simultaneous approxination of linear forms. Let real numbers eij, be given for i

1 . 4 .

Computational diophantine approximation.

Li

1, m

..., n

, j
for

1. i

....
1.

Put

In this section we give some idea of practically solving the diophantine

approxiration problems that we encounter in solving diophantine equations. In we give no rigorous treatment. We neglect worst cases, and to vork in For a nore rigorous theoretical treatment we refer to a forthcoming subsequent
A

1 qj-bij
j-1

.... n .

A celebrated theorem of Uinkovski states that there exists a solution (pl.. .pnsql.. .. q , J of the system of inequalities

..

chapters of t h i s t h e s i s many examples a r e given, showing t h a t our methods a r e indeed useful i n practice. Applying the method i n practice may be the b e s t way of acquiring the necessary ~ i n ~ e r s ~ i t z e ;hl n ~ ef fo r the method. W e s h a l l deal with problem. Let constant
Q0
8

( c f . L e v t r a , Lenstra and b v 4 s z [1982], Proposition (1.11) 3.4). ( i i ) , f o r any given point nearest l a t t i c e point, v i z .

and our L e m m a y t o the

y E in , the distance from

the following computational

E R Q

diophantine

i j' i n t e g r a l unknowns with

- max[qj 1

be given, and l e t

Let

Li

pl. . . pn, ql, . ., 9 , be be a s above. Let a positive say, be given. Find

..

approximation

(cf. our Lemmas 3.5 and 3.6). This algorithm enjoys the property t h a t these lower bounds a r e usually near t o the a c t u a l minimal s o l u t i o n s . I n a t y p i c a l s i t u a t i o n , where the l a t t i c e is not too d i s t o r t e d , the vectors reduced b a s i s a l l have about the same length, which is of magnitude of

assumed t o be a r a t h e r large number. 10''

a lower bound f o r the value of

ri of the t h e order of

where

(pl,.

.. ,pn,ql,. . .,qJ

runs through the s e t of values with

Q s

QO .
The value of J!(r.y) t(r) large a s t h a t . I f a s well a s the lower bounds computed f o r i t , a r e about a s y is not too close t o a l a t t i c e point. the same holds f o r (it
is

From t h e theory outlined i n Section 1 . 3 it follows t h a t one w i l l be s a t i s f i e d i f t h i s lower bound is of the s i z e problem may be formulated. Related problems i n diophantine approximation theory a r e those of a c t u a l l y f i n d i n g a good o r the b e s t solution of maxlLi-p i ( < c f o r a fixed c > 0 . As w e s h a l l see. the L3-algorithm is a very useful tool f o r finding good solutions. The problem of finding the best W e note solution however seems t o be our applications of
3

Q7'n . For the p-adic

case an analogous

Moreover, the running time of the algorithm is good, both i n the sense polynomial-time in the length of the input-

theoretical

parameters), and i n the p r a c t i c a l sense. To solve the problem of finding a lower bounds f o r above. w e take the l a t t i c e large a s
Q * " : /

as follovs. Let
I.

as formulated maxlLi-pi! be an integer. a t l e a s t a s n

e s s e n t i a l l y more d i f f i c u l t .

t h a t i n most of

The l a t t i c e

of dimension

+ m

is defined by

s o l v i n g diophantine equations it s u f f i c e s t o have a s u i t a b l e lower bound f o r =ILi-pil

specifying a b a s i s , namely the column vectors

k,. ..., &-

of t h e matrix-

The computational t o o l t h a t w e use t o solve the afore-mentioned problems is 3 - k n-stra. the so-ca-1led-L - l a t t i c e b a s i s reduction algorithm. described i n - Lenstra and Lavlsz [1982]. W e s h a l l give d e t a i l s of t h i s a l g o r i t d . i n Chapter

3.

Below w e briefly

indicate how

it

can be used

to

solve diophantine

approximation problems.

Let
basis

I.

be a l a t t i c e i n P1,

in . The L3-algorithm accepts a s input an a r b i t r a r y

(The symbol

means t h a t a l l not e x p l i c i t l y given e n t r i e s a r e zero).

.... P ,

of

I .

. hs

output it gives another basis

cl.

.... %

of

Applying the L3-algorithm t o t h i s l a t t i c e w e f i n d a reduced b a s i s . of which the b a s i s vectors w i l l hlVe1Tengths of about s i z e of QO c~'(*~)
C

t h e same l a t t i c e

r ,

t h a t is a so-called reduced basis. The concept reduced

means something l i k e nearly orthogonal. From a reduced b a s i s it is possible t o cormpute lower bounds f o r the following two quantities: ( i ) . t h e length of t h e non-zero l a t t i c e point t h a t is nearest t o the o r i g i n , viz.

which is roughly the

Generally speaking, the l a r g e r

is. the l a r g e r the lengths

of t h e b a s i s vectors of a reduced b a s i s w i l l be (and t h e l a r g e r t h e lover bounds f o r C(I.) and C(r,y) w i l l be).

Let us f i r s t t r e a t t h e homogeneous case, i . e . the l a t t i c e point

11

8 . [ql..

..,%,pl,. ..pJT . I t

Bi

0 for a l l is equal t o

. Consider

The

above

process

describes

how

to

find

lover

bounds

for

systems with

of QO

diophantine i n e q u a l i t i e s . where and max(Li-pi[

It w i l l be c l e a r from the above t h a t it is not (ql,

d i f f i c u l t t o f i n d good s o l u t i o n s , i . e .

. . . ,q,,pl. .. . ,pn)

near t o the b e s t possible value. I n p a r t i c u l a r ,

. *

the b a s i s

Li
size QO

- j-1 1 q -[C-8 J ij
-

for

1,

...,

.
C(r)

vectors of a reduced b a s i s a r e adequate f o r the homogeneous case. and f o r the inhomogeneous case the l a t t i c e points near t o

w i l l be such s o l u t i o n s . The

From the application of the L3-algorithm w e f i n d a lower bound f o r

.W e

of

l a t t i c e points near t o respect to a

t

a r e not d i f f i c u l t t o f i n d once a reduced b a s i s is sl, s E R a r e the coordinates of y with n then one may take the l a t t i c e points with si
( i

assume it t o be large enough ( i f t h i s is not the case, w e try a

C
C

available. S p e c i f i c a l l y , i f coordinates

....

somewhat l a r g e r value f o r l a t t i c e defined f o r t h i s constant cl such t h a t

and perform the L3-algorithm So w e may assume t h a t

again f o r the
is a small

reduced b a s i s ,

).

there

i E .Z

t h a t a r e near t o

1.

.... n

).

I n the d e f i n i t i o n of the matrix above the expressions

r these expressions w e have constructed a l a t t i c e r c zWn The L3-algorithm can be adapted t o work exact f o r i n t e g r a l . i .e.
W e have ILi-C-LiI

[C-dij ] occur. Using t h a t is completely

m-Qo ,

SO

w e may assume t h a t f o r small constants

those l a t t i c e s . s o t h a t rounding-of f e r r o r s a r e avoided (cf . Section 3.5). C2* C3 The "errors" occur i n the difference between the

Li

and the cl.

C-Li 6

and a r e

thus kept under control by choosing the proper constants B y t h e choice of

C

c 2 , ,c .

Of

course one should take c a r e t o have the numerical values of the t h i s l a s t bound has the required s i z e . a r e zero. W e take
A possible v a r i a t i o n of

pi

correct

t o a s u f f i c i e n t precision.

W e

s h a l l discuss

such numerical

iJ

and the

problems b r i e f l y i n Section 2.5. Next, w e study the inhomogeneous case, where not a l l t h e same l a t t i c e

Bi

I"

a s i n the homogeneous case 8

ij

(note t h a t the l a t t i c e

the above diophantine approximation problem is t o Li

d e f i n i t i o n depends only on the

and the

C ) . Consider the point

give weights t o t h e . l i n e a r forms

i.e . t o look f o r a lower bound f o r

From t h e reduced b a s i s found by the L3-algorithm w e have a lower bound f o r m , y - L A s s u w ,th,at l a t t i c e point

it is l a r g e enaugh,anbof-sFzeQo

.. W e

take the s *

h e r e the

3[ql..
[ql

...%.pl,. ..
qm,y-pl.

P ~ T a s i n the homogeneous case. Then

. -

a r e fixed p o s i t i v e numbers. This s i t u a t i o n can be d e a l t with wi - . e a s i l y .by replacing t h e C 's i n the (n+i) t h - r o w of t h e G i r i x - b y proper

_
wi

_-_

x
vhere

-Y-

.-.-. -

--

.-.. in-c.pJ

.
.... n
.

constants depending on

Another v a r i a t i o n is the problem where not a l l the v a r i a b l e s same upper bound

QO'. To i l l u s t r a t e t h i s , assume t h a t

qj 1 , and t h a t

the

Li-

[C-Bi]

lqj-[C-dij] J -1

for

i-1.

The same reasoning a s i n the homogeneous case now y i e l d s the desired r e s u l t . Note t h a t i f w e have performed the ~ ~ - a l ~ o r i c once h m f o r given with d i f f e r e n t bij

N o w suppose t h a t f o r s o w

Q1

>

Q2 ( i t w i l l be handy t o have

Q2

I Q 1

) w e

w e may

a r e i n t e r e s t e d i n the s o l u t i o n s with

use t h e r e s u l t t o t r e a t the homogeneous case, and many inhomogeneous cases

pi

's

a s well, a s long a s the

iJ

's

a r e the same.

) q j l 5 Q2

for' j

ml+l.

..., q

, Then it is obvious that

Next let

be of the size of

ml m-m +1 Q1 -Q2 , and take the matrix Hence the lattice

can be described as the set

1 -1
1

-aij

pi o

for

1,

.... n

The ~ ~ - a l ~ o r i tprovides hm a lower bound for the length of the nonzero vectors For a lattice point sow c ' , c cn (ql..

... k . i - ~ - p ) ~ we

. It follows that lGpl > c'.QI/C . This variant is useful when

-(m-m +1) for some 'Q2 a combination of real and p-adic

expect that -(mI-1) > c"-Q1

ILC-PI > c0Q1

for

t i
t

in this set, which is of the same sire as This yields the desired result, if p For the inhomogeneous case, put
y

p p . n ' ( n + m )

, and that of

90 .

is taken large enough.

i r
I

techniques is used, such as for the Thue-Mahler equation (see Section 8.6). We conclude this section by giving the analogous method of p-adic diophantine approximation. We assume that the are p-adic integers. Let No
p E No we denote by 7 " )

(0..

(PI 1T ...O,-By).. ..,-Bn

.
-

aij. Bi

are in Q and, moreover, that P N U (0) For any p-adic integer 7 and

and consider the set

the unique rational integer such that

7 = 7 " )

" ) (mod pp) , 0 s 7

pp

<

pp

.
<*In
_ C._

Bi
, and
in the

1 qj -aij
j -1

pi (mod pp)

for

1,

..., n

Let

p E N

be such that

is roughly the same size as

-assume.-that y is large e n o u g ~ t A ~ d o g of u e the constant real case above). Take for c o l b vectors of the matrix

'

Then x E r if and only if + y E r so r* is a translated lattice. A lower bound for4ri-y) n a ~ y i e l d sthe desired result. _Again variations are possible. a s i n the real case. e.g. by replacing on the (n+i) th row the p by different p pi

' I

the lattice of which a basis is given by the

--

. It

is even possible in this way to

treat more than one prime

at the same t h e .

We conclude this section with three remarks. Firstly, in the case that the dimension of the lattice under consideration is only 2, the ~ ~ - a l ~ o o r i t h is m essentially the continued fraction algorithm. and so yields nothing new. For the p-adic continued fraction algorithm, see de Weger [1986a]. Secondly, the inhomogeneous case of diophantine approximation of one linear form of real Consider the lattice point numbers can also be treated

by

what is known as Davenport's

lemma. cf. Baker

and Davenport [I9691 (and its multi-dimensional generalization. cf. Ellison

25

[1971a]). W e w i l l r e t u r n t o t h i s i n Chapter 3, and explain there why w e p r e f e r our method. Finally, one of the nice features of the above method
C

I t w i l l be c l e a r from Section 1.4 t h a t the methods outlined there a r e of use f o r solving t h i s problem. For case w e expect, by choosing
C
QO

w e take

No

.W e

a t l e a s t of s i z e

c1
have

I n the r e a l

, that

of

practical
)

diophantine approximation i s t h a t i f an extreme solution e x i s t s . then i n the homogeneous case the l a t t i c e (with proper constant or
p

w i l l be

f o r a small constant
lAl

c'

I t follows by combining the two i n e q u a l i t i e s f o r

d i s t o r t e d . This means t h a t the reduced basis w i l l not be as nice as expected, f o r example there might be a b a s i s vector i n it t h a t is s u b s t a n t i a l l y shorter than the other ones. In the inhomogeneous case the existence of an extreme s o l u t i o n means t h a t there is a l a t t i c e point extremely near t o the extremal solution is presented e x p l i c i t l y (e.g. a s one of the vectors of the reduced b a s i s ) . by replacing i n the above reasoning (r) or

that

The

algorithm detects such an extraordinary s i t u a t i o n a t once, and i n most cases i n the homogeneous case One can check whether t h i s (r,y) by lower bounds f o r

So the upper bound of log No

No

for

i s reduced t o an upper bound No

Nl

of the s i z e

which is a considerable improvement indeed. W e now may apply the

N1

procedure with

instead of

and repeat u n t i l no f u r t h e r improvement

extremal solution a c t u a l l y s a t i s f i e s the original equation, and then proceed a l l vectors i n the l a t t i c e except the extremal one. These new lower bounds e solved diophantine w i l l i n general be of the expected s i z e . However. when w equations i n practice. w e have never met such an extraordinary s i t u a t i o n .

is obtained. In practice i t appears almost always t o be the case t h a t i n t h a t

s i t u a t i o n the reduced upper bound is near t o the a c t u a l l a r g e s t solution. anyway so small t h a t simple methods of finding a l l the s o l u t i o n s below t h a t bound s u f f i c e . I n the p-adic case an analogous reduction of upper bounds can be reached.
A

following a similar argument. W e have f o r the l i n e a r form 1.5. The procedure f o r reducing upper bounds. seen i n Section 1.2 how u p p bounds ~ inequalities f o r - the solutions of can be found. the In where c ord (A) P

( c f . (1.4)).

sc +c
1

2"'"" *

W e have

exponential

and equations -occurring there

c a r e small constants, and mj is pne of t h e variables. 1'. 2- _ - -. Moreover, the variables a r e bounded by a large constant that is

No

Section 1.4 w e have studied some diophantine approximation theory from a p r a c t i c a l point of view. N o w these two things come together.
-

e x p l i c i t l y known. W e take

JL

such t h a t

pp is a t l e a s t of s i z e :+$t

SO

--

t h a t t h e lower bound f o r the s h o r t e s t nonzero vector i n l * r g z than the

I '

(or

r
I?

) is
-

--

--

_ ;

- . Then ~ it ~ follpws

From

the

application of

the Gelfond-Baker

theory w e are

left

with

of the t r a n s l a t e d l a t t i c e

r*

t h a t the elements , o r t h e l a t t t c e

) cannot be solutions of (1.2).

Therefore.

following problem. W e have a l i n e a r form'

so that w e f i n d a new upper bound f o r

mj

where the

B _and d
A

h e r s is now of no importance anymore), and the W e how that

a r e constants ( t h a t they a r e logarithms of algebraic n

order t o obtain a reduced upper bound f o r

is extremely close t o 0 , namely

a r e i n t e g r a l unknowns.

mj in H I f t h i s is not y e t s u f f i c i e n t P t o derive a t once a reduced upper bound f o r H then ve can do s o by

is about

log No / log p

.W e

t h a t is of t h e s i z e of

JL

which

repeat t h i s procedure f o r a l l t h e

applying a reduction s t e p f o r r e a l l i n e a r forms, where w e may take advantage of t h e f a c t t h a t f o r some of the variables a much b e t t e r upper bound has j u s t been found ( c f . t h e second v a r i a t i o n i n Section 1.4).

where

a r e (small) constants. and No for

N
No

e x p l i c i t upper bound

. This

v x 1n,

. Finally,
los0

w e have an

whole procedure as f a r a s possible.

-.

Again we repeat t h e

is very l a r g e ,

say.

element

a E K

be defined by

For algebraic integers. of norm 2.1. Algebraic number theory. unit

c

NKB(a) E Z
a,

The u n i t s a r e p r e c i s e l y the elements denote the i d e a l generated by a

21

T w o elements
a

such t h a t

B
Let

of

K a r e c a l l e d associates i f there is a
(a)

c.B

Associated elements generate the same i d e a l , and d i s t i n c t generators of an I n t h i s s e c t i o n we quote r e s u l t s from algebraic number theory t h a t w e use throughout the remaining chapters. W e refer t o Borevich and Shafarevich [I9661 o r other text-books Let are let
K D

i d e a l a r e associated. There e x i s t only f i n i t e l y many non-associated algebraic integers i n

OK

with given norm. The r i n g of algebraic i n t e g e r s is denoted by a D be elements of

OK

on algebraic number theory f o r f u l l d e t a i l s . Q , of degree

Let

al,

Then be a f i n i t e algebraic extension of embeddings

o : K

Z-al x

+C .

..., ... x
OK

that are

Q-linearly independent. i f i t i s a subring of the

2-aD i s c a l l e d an order of

[K:Q]

. There
d , and

'maximal order' In
K

irreducible

Let

a E K

be an element of degree

a .

> 0

be the leading c o e f f i c i e n t of i t s minimal polynomial over h(a) by

any algebraic integer can be w r i t t e n as a product of

W e define the (logarithmic) height

elements. Here an irreducible element (prime element) is an element t h a t has no i n t e g r a l divisors but its own associates. However, t h i s decomposition i n t o primes need not be unique. Ideals can a l s o be decomposed i n t o prime i d e a l s . and t h i s decomposition is unique. A principal ideal is an i d e a l generated by

where t h e product is taken over a l l embeddings does n o t depend on the f i e l d

a

. Note
K

that t h i s definition

a s i n g l e element

T w o f r a c t i o n a l i d e a l s a r e c a l l e d equivalent i f t h e i r
I t is well known t h a t there a r e only f i n i t e l y many

al.

... ad
h(a)

Hence, i f

the conjugates of

then the above d e f i n i t i o n applied f o r

- Q(a)

are

quotient is principal.

yields

equivalence classes. Their number is c a l l e d the class number ideal a

it is always t r u e t h a t

% .

For an

d _. -1z-log(aO( I . i-1

I~~-I))

" K a

i s a p r i n c i p a l i d e a l . T h e o m of the

(integral) ideal For a prime i d e a l

is defined by

NK/P (a)

#(oK/Q)

such t h a t

there is always a r a t i o n a l prime number ._p

p -is-a-dfvfsur uf (p)
Let t h e r e be s r e a l and cl.

.W e

saythat

p --liesabove -p p
divides (p)

. The-ramffi cation
. The
residue class

2 -t

non-real

embeddings (with

D -- s

index
2 - t ).
. -

degree

_.9

is the l a r g e s t power t o which

is the integer such t h a t

Then D i r i c h l e t ' s independent u n i t s of u n i t s of

K

Unit Theorem s t a t e s t h a t there e x i s t s

.... c r
1 C

, where

a system of

- 1,
E

such t h a t t h e group W e denote by ord (a)

-

is given by a a a root of unity, ai

< - l1 - ...-

for

i-1..

...r

the i d e a l

the exact power t o which t h e p r i w i d e a l a a w e write ord ( a ) for ord ( ( a ) )

divides

For f r a c t i o n a l i d e a l s

t h i s number can of course be

There a r e only f i n i t e l y many roots of unity i n units t h a t generate the torsion-free

. Any

negative. For numbers s e t of independent ord ( a ) P

p a r t of the u n i t group Is c a l l e d a

. Note

that

ord (a)/e

PO.

system of fundamental unf ts.

can be defined f o r a l l a
E

The nraber

is c a l l e d an algebraic integer i f

a .

.W e vill

r e t u r n t o t h i s i n Section 2.3,

which

. Let

t h e norm of an

deals with p-adic number theory.

2 .$.

Some auxiliary lemmas.

In this section we give a few simple auxiliary lemmas. The first one enables
us to find an upper bound in closed form for.some real number

x > 1

that is If
-

bounded by a polynomial in
LEMMA 2.1.

log x . See Petha and de Weger [1986]. Lemma 2.3. b > 0 , and l e t
x E R, x > 1 satisfy

- e2

c s e2 , then note that

x I a

2 h (e /h) -(log xlh

So we may assume
0

in this case. The result follows. x and log(l+x) are near if 1x1

Let
5

a 2 0 ,

h 2 1 ,

The next lemmas make explicit that

is

h b-(log x) .
then

small in the real and complex case, respectively.

If

b > (e2/h)h x <

LEHHA 2.2,

Let

If a < 1 and
ga

1x1 < a

then

zh-(alfi+blfi-log(hh-b)lh
then

.
and

Ilog(l+x) l <

-10 (1-a) -

1x1

and i f

2 (e /hlh
5

2h-(a1fi+2.e2)h

Proof,

We may assume that x-a+b-(logx) h

is the largest solution of

Proof. at x

Note that

log l+x)/x

is a strictly positive and strictly decreasing 1x1 < a always less than its value

.
we infer

function for 1x1 < 1

- -a

Hence it is for

. The same s true for the function x/(ex-1)

Let

BY

(z1+z2)lfi I z i f i + zlfi 2 X I " '

,a y '
?-

+ c -log(x1&)
y by

,
xlfi

JBWi 2.3,

0 < a I

If

1x1 c a

then

where

- h-blfi . Define
-

(l+y) -c.log c

From

log c

< log(c.log c)

If a < 2 . Iei-x-ll< a and

1x1 <

then

-.

it follows that

J ' r o o f L which implies x > ch-(log clh . Hence y > 0

Note that x

le

i-x -11 a

2 -! s i n ( : x )

and that 0
Ix

2-sin(:-x)/x

is a

Now.

positive and even function, that decreases on c-log c

(l+y).c-10%~- x l f i a alfi + c-log(l+y)

minimal value at c-loglog c

< a

. Hence

it takes its

. The

first inequality now follows. The second one

0

can be proved in a similar way.

<a l " '

+ c.y + c-log c + c-loglog c

.
2.3.
p-adic numbers and functions.

Hence

In this section we mention the facts about p-adic numbers and functions that

If c r e2

it follows that

we use. For details we refer to Bachman [I9641 and Koblitz [1977]. [1980].

I /

W e assume t h a t the reader i s f a m i l i a r with the f i e l d of p-adic and t h e p-adic valuation

numbers 0P a s defined i n ord . Note t h a t the ordinary ord P P 0 coincides with the d e f i n i t i o n given i n Section 2.1. W e denote by n the P P completion of the algebraic closure of 0 i . e . the f i e l d t o which a l l P * -- p-adic theory is applied. Every nonzero number a
i
0

A n equivalent d e f i n i t i o n is u n i t y such t h a t by ordp(<-r)

logp({)
0

- logp((/r)

, where

i s a root of

>

I n t h i s way the p-adic ord (() > l/(p-1) P then

logarithm is a well

defined function. Note t h a t

logp(() if

l i e s i n the s u b f i e l d of

D
P

generated

< . Finally

w e note t h a t

has a p-adic expansion ordp(<)

a where with then k

1 ui-P

i-k

.
ui

- ordP( l o5(I+())

- ordP ( a )
.

and the p-adic d i g i t s 0 0 , and

are in

0, 1,

...,

p-1 1 , k

uk z 0
a

The number

can be represented i n t h i s way by taking

and a l l d i g i t s equal t o

is c a l l e d a p-adic

ord (0) .o by d e f i n i t i o n . I f ord ( a ) 1 0 P P integer. The s e t of p-adic integers i s denoted by

2.4.

Lover bounds for l i n e a r forms i n logarithms. theory full

0 I n t h i s section w e quote i n d e t a i l the r e s u l t s from the Gelfond-Baker that w e use. algebraic They y i e l d lower bounds f o r l i n e a r forms W e do not always give the numbers. theorems in their

A p-adic unit i s an a E 0 with ord ( a ) 0 . For any p-adic integer P P P P-1 a and any p E No there e x i s t s a unique r a t i o n a l integer a(') ui-pi i-0 satisfying

i n logarithms of

g e n e r a l i t y . since i n t h i s t h e s i s only l i n e a r forms with r a t i o n a l unknowns occur, whereas most Gelfond-Baker theorems a r e formulated f o r l i n e a r forms with algebraic unknowns. W e selected r e s u l t s t h a t give completely e x p l i c i t

For by

ord ( a ) 2 k P

w e also write

constants. The f i r s t r e s u l t i n t h i s f i e l d f o r a l i n e a r form i n logarithms k 0 (mod p )

with a t l e a s t three terms is due t o Baker (19661, and i n the p-adic The p-adic norm is defined der Poorten [1977]. W e will use more recent, sharper results,

case t o due to

Coates [1969], [1970]. For a survey of t h i s theory, see Baker [I9771 and van Wiildschmidt [I9801 and Y u ;[1987~]. Further improvements of the constants have been reached, but too recently t o be taken i n t o account i n t h i s t h e s i s .

14,
-

-ord (a) P P .
-

I n Pection-2 -1 we have seen h o w1 0 define __ord P

extensions of

and

. For

any

the p-adic logarithm

lo

(-l+a)

a E 0 with ordp(a) > P by the Taylor s e r i e s

ord on algebraic P l/(p-1) w e can define

First w e deal with real/complex

--

r e s u l t of Waldschnidt [198'0]

l i n e a r forms i n logarithms.
. --

W e quote the

LEHUA 2 . 4 t W a l d ~ & n i d t ) ~ Let

al,

be a number field witb

( n z 2 )
I

[K:O]
V1,

-.-,an E K ,

a d

bl,

..., b n Z
1/D s V1

. Let
and

. Let .... Vn be

-D

--- -

posicive real numbers satisfying

This
t h logarithmic

lo^((^-(^)
e

function

has

the

well

known

properties,

such

as

... s Vn

1ogP(e1)

logp(()

f o r a11
(

el,

Z2

f o r which it i s defined. p I n 0 P the is odd). Using

i f and only i f

is a root of unity.

only r o o t s of u n i t y a r e the (p-1)

t h roots of unity ( i f

where

log a

l p t h e s e p r o p e r t i e s , t h i s logarithmic function can be extended t o a l l ( E C k 0 , a s follows. Let k E N such t h a t ord (( -1) > l/(p-1) with ord (e) P P Then

the logarfthm of

for
a

. Let

1,

.... n
V+ j

- -( V

is an arbitrary but fixed determination of

,1)

for

n, n-1 , and put

Put

max lbil BLln

. If

then

l A l > exp ( -2 e ( n ) - n 2 - n .Dn+2 -V1-.

..- V n - l o g ( e - D n-1 . ~ +) Let

where

e(n)

- min

known t h a t

[Q(hl..

replace t h e factor

8 - n + 51, 1 0 - n + 33, 9 - n + 39 ) . I f , moreover, i t i s 9 - n + 26 and - ., A n ) :0] 2n , then we can take e ( n ) 2-n n+4 n i n the above bound f o r Ihl by n .

V. 2 max ( h ( a . 1 , f - ( l o g p ) / d )

for

1.

.... n
.vn-l)

.
,

such t h a t

V1 I

...

Vnl

v;-~

Waldschmidt's main theorem does not give the constant of

e(n)

as detailed as

w e do, but he does s o i n h i s proof, c f . p. 283. W e remark t h a t improvements t h e above bounds have r e c e n t l y been found by Blass. Glass, Heronk and d S t e i n e r [1987C], (1987 1 , Loxton. Hignotte, van d e r Poorten and Waldschmidt [1987]. and Philippon and Waldschmidt (19871. For t h e case bound has been given by Hignotte and Waldschmidt [1978]. n B 2 max
2 mix

lbll.

.-., l b n l ,

) ,
f p - ( l o g p)/d

-2 .

a sharp

3 l o g ( l c - B . l o g Bo. 4-n

) .

Suppose t h a t
I n t h e p-adic case we quote two r e s u l t s : one due t o Schinzel 119671 (Theorem 1 ) f o r the case of a l i n e a r form i n logarithms with two terms. and another b f o r t h e general case. due t o Y u ( 1 9 8 7 ~ 1(Theorem 1 , s e e a l s o Y u (1987 I ) . W e note t h a t Yu's bounds improve much upon t h e r e s u l t s of van der Poorten's van der Poorten [1977]. Moreover.
-

o r d (a ) P j

for

1,

..., n

. that
(2.1)
abl

[~(a;'?

..., a y q ) :L]
ord (b ) P j

n q ,

that

ord (b ) P n

for

1.

...,

n , and

bn -...a n

z 1

. Then

proofs seem t o contain some e r r o r s . W e

give Schinzel's r e s u l t f o r quadratic f i e l d s only.

-

,
Let
p
be prime. Let

LEEMA 2.5 (Schinzel).

let
D

be t h e discriminant o f

be a squarefree i n t e g e r , and
(

(pAp-1) (2+)

(Pa( l o g p ) / d ) - ( n + 2 T ~ l

- ..- - vn -

Q(/A)
,

Let

- .fn/('

and
- -..

- xn/x'

be elements o f
L

K , vhere
-

4 ' . t'. x ' , X" are algebraic integers. Put

----A-

l o g max

I~-DI'''.

I t *- x * l . I t * . x n l . I t W . x * l .I t " . x w l
K

1
7 E K

where Let
p

IrD
p

denotes t h e maximal absolute value o f t h e conjugates o f with norm

N p

be a prime ideal o f

pP

.
4

Put

-- ordP(p)

.If

.f

or

i s a p-adic u r i c and

tn

+-

.
and
Cl(p,n)

2/p.log

p ,

i s given by t h e following t a b l e , w i t h f o r

p r 5

then

Put

a ( n L 2 ) be nonzero algebraic numbers. n [L:Q] . Let . , bn be rational i n t e g e r s . bl. Let p be a prime ideal o f L , l y i n g above t h e rational prime p . Let e P be t h e ramification index, and f t h e residue c l a s s degree o f p . Write P Lp f o r t h e completion o f L w i t h respect t o ord . (Note t h a t f o r a l l

2.6 tYulL Let

ale

...,

Q(al,.

..,an)

, d

..

Remark,

Yu [ 1 9 8 7 ~ ] Anounces that the 'indepedence condition* (2.1) can be

log(l+x)

removed. This may be at the cost of somewhat larger constants.

- x - x2/2 +. x 3/3 - ... ,

2.1 x

or by the more rapidly converging series 2.5. Numerical methods. For In solving diophantine equations using computational methods from diophantine approximation theory, as we will do in Chapters 4 to 8, it is necessary to have logarithms (real, complex or p-adic) of algebraic numbers available to a large enough precision (maybe several hundreds of digits). We will not go deeply into the problems of computing such approximations, but make only a few remarks on it in this section. which is a matter of a few divisions of a multi-precision number with a To start with, the precision with which most computers (mainframes as well as personal computers) work, is insufficient for our purposes. Usually at most double precision (52 bits. equivalent to 15 decimal digits), or at best X1 and x3
E

l+x log*-x 1x1

x 3/3 + x 5/5 +

... ) 1x1 the

very small this method works fast, whereas for larger log 1.0001, log 1.00000001 and k 1
E No

following idea works well. Compute approximations to the desired precision of log 1.1,

, say, and store them. Now compute

x1 E [l,1.1)

such that

rational number with small numerator and denominator (11 and 10) only, that can be done fast. Next. compute x2 E [1,1.0001) and k2 E No such that

quadruple precision (112 bits, equivalent to 33 decimal digits) is standard available. This is not sufficient for our purposes. not only because we may require larger precision. but also because we want to have the rounding off errors under control, to be sure that no solution of a diophantine equation is missed by unexpected consequences of rounding off errors.

x2. 1.0001 k2 ' k3 E No such that

[1,1.00000001) and

Then compute Packages for computations with arbitrary precision are available and very useful. e.g. difficult
.-

compute

log x

log x3 by

by the Taylor series. which converges very fast. and

- - -

the HP package of R.P. write one's own

Brent (cf. Brent [1978]). It is not for simple manipulations on log x

to

package

multi-precision numbers. such as addition, multiplication and division (cf .

- .-

log x3 + kg-log 1.00000001 + kq-log 1.0001

-

kl-log 1.1 .
-

k t h --(I 9811 for efficienr a l g o r t t h m ) . ~

- -.
.

such package- are -available for

When computing 2 1 1 thls, -one should take care of_ having the rounding off errors at each addition/multiplication under control. This can e g. be done by doing-all computations twice, rounding off in different directions at each step, such that finally a small interval is found in which the exact number lies (with mathematical certainty). Computation of arctan x is done by the Taylor series

manipulations on p-adic numbers, but the programs are similar to those for real numbers. Computing roots of polynomials with integral coefficients can be done by Newton's (only) method, both in the real and the p-adic case. One should make sure desired precision, preferably not of the root into the by substituting the found approximation that the result obtained is correct --the
I

arctan x The number

r

polynomial and checking that the result is 0 within the desired precision, but (also) by theoretical error estimates for the Newton method. Computing logarithms can be done by the Newton method too. However. we did it by using the Taylor series
r

- x3/3

+ x5/5

- ... .

3.l4lS9. . .

c p be computed rapidly by this series for the

arctan function, by the identity

16-arctan

Doing p-adic arithmetic has the advantage above real arithmetic that rounding off errors do not tend to become larger, as long as one is not dividing by a number with large p-adic order. If computed by the Taylor series log (l+x) P ord (x) > 0 P then log (l+x) P can be

The above considerations are sufficient for doing exact computations with the L3-algorithm. as we present it in Section 3.5. We also use the simple continued fraction algorithm in some instances. This we do as follows. Suppose we want to compute the continued fraction expansion of a real number
6 , that we have approximated by rational numbers

- x - x2/2 + x3/3 +
2.(X

el, e2

such that

and also it may be useful to compute logp If x

. l

k E N

for some small

0 (mod p)

and

3 x /3

x /5

and then logp x can be computed. since

6

e2

*1 exactly. As far as they coincide, they coincide also with the

6

We can compute the continued fraction expansions of

continued fraction expansion of is needed so far that the xk = 1 (mod p) , and then

. If the continued fraction expansion of

convergent with denominator

1 (mod p)

k th

there exists a

such that

be known exactly, for a given (large) constant least as small as

Xo

. then

qk > 0' should be at

xi2 .

and the above given Taylor series can be used to compute the denominators of the terms. Hence. to find the first log (l+x) p

logp x

Note that

Almost all computer calculations done for the research of this thesis were performed on an IBM 3083 computer at the Centraal Rekeninstituut of the University of Leiden, using the Fortran-77 language. Also some computations were done at a VAX 11/750 computer at the Rekencentrum of the University of Twente.

in computing the above mentioned Taylor series there will be factors

. it is not enough to compute only the first

k

p in p-adic digits of

the Taylor series, but the first k

terms of p/ordp(x) terms must be taken into account. where

is the smallest integer satisfying ksord (x) P

log k/log p r p

For rapid convergence of Taylor series it^?^ desirable to apply them only for numbers x with large p-adic order. For example,

--

1 - j-(

converges not as fast as log 4 3

1 --log3 6 4 3

7-32

3 - 62 4 -' 7 - 3 /2 + 7 3 /3 - ... )

: itself, whereas the numbers P subfield of 0 . P

Let c, 6

Bi, @

are allowed to be in some larger

be positive constants. Put

maxlxil

Let

Xo

be a (large)

positive constant. In the real case we shall always assume that

3.1.

Introduction.

In this section we give details of the computational methods we use to reduce upper bounds for the solutions of diophantine equations. Our starting point will always be a linear form involving the unknowns), A

. that is close to 0 (in the real or p-adic

Let

be real constants, with c2 > 0 . In the p-adic case we shall cl, c2 and assume that x > 0 for some index j E (1.....n) j

sense, with the word "close" defined explicitly in terms of an inequality together with a large but explicitly known upper ord (A) 1 cl + c2-XJ P bound for the absolute values of the unknowns. Our aim is to reduce the upper bound by showing that there are no solutions between the new and the old upper bound. Let p al,

(3.3) (3.4)

XSX,.
Our aim is to find a constant X1 real case (3.2) can be replaced by

of the size of

X s X1

log Xo

such that in the

and in the p-adic case the bound xj

$

. Let

..., en, xl, . . .

. xn

be given numbers, in R , or in n for a fixed prime P ' be unknowns in Z . Put

xj s Xo

(a consequence of (3.4))

can be improved to

5.

In the forthcoming sections we treat all cases. according to the 3 classification given above. We insert Sections 3.4, 3.5 on the L -algorithm. which will be our main computational tool, Section 3.6 We classify such linearfor~saccord_ing to three criteria: on finding short vectors in lattices, and Section 3.13 on certain sublattices that are3eful for our applications. n

+ + +

homogeneous if fi

- 0 , inhomogeneous if + 0 one-dimensional if n - 2 , multi-dimensional if

@

;
;

r 3

real if all the numbers are in R , p-adic if all the numbers are in Q n

.
3.2.
fractions.
-

The reason that the case

H o m o g e n e ~ e - d i m e n s i o n a lapproximation..in-the real _ case :rpatinued

---

is called one-dimensional is that in the

homogeneous case the linear form

We first study the case

leads to studying the simple, one-dimensional continued fraction expansion of

a1/e2 . The
A

inhomogeneous case with

X-6

1 , viz. 5
!

Put

- a1/e2

. -We-assume

that

i s -irrational. Let the continued

-6+

fraction expansion of 6 be given by

is not of any interest in the real case, but it is of interest in the p-adic case. We call this the zero-dimnsional case. In the p-adic case we require that the quotients and let the convergents pn/% for n

- 0.

1. 2,

...

be defined by

i I i / 6 ,

and

@/*j

are in

4 d
r

We may apply Lemma 3.1(i)

JEMU 3.2,

directly, or as follows.

Let
A

It is well known that the convergents satisfy the inequalities

-(ak+l)

,
k
s a t i s f y i n g (3.7)

where t h e maximum is taken o v e r a1 1 i n d i c e s and (3.2) hold f o r

I f (3.1)

xl. x2

with

X r X1 , then

and that if

p/q

satisfies the inequality

1 1 X X < x-l~g(~.(~+2)/(821) + z - l ~ g

.
X b follows. We can apply Lemma only.

Remark, then p/q must be one of the convergents (cf. Hardy and Wright [1979], Proof,
1 8 1 1

From Lemma 3.2 an upper bound for

2.1 here. but Lemma 2.1 is sharp for large Theorems 163, 171 and 184). We may assume without loss of generality that and that such that (x1,x2! X (3.1) and (3.5) yield

< 1e21 , that x1 >

0 ,

-1 -c -exp(C-X)
The result follows by applying Lemma 3.1(i). In practice it does not often occur that

.
0

1 . From (3.1) it follows that there exists a number implies X

xl

and (3.6) for

(p,q)

X*
now

(-x2 ,xl)

. We

have the following criteria.

is large. Therefore this lemma

LElMA 3.1.
(-x2,x1)

(i) . (p . q ) k k

I f (3.1) and (3.2) hold f o r f o r an index

xl. x2

with

X r X* , then

is useful indeed. Summarizing, this case comes dovn to computing the continued fraction of a

that s a t i s f i e s

s -1 + l o @

.x0+1] /log(+4))
a,+l

(3.7)

real number to a certain precision, and establishing that it has no extremely large partial quotients. This idea has been applied in practice by Ellison b [I971 1. by Cijsouw, - Uorlaar and Tijdcman (appendix to - Stroeker and Tijdeman
) [1_982]

Moreover, the p a r t i a l q u o t i e n t

satisfies

and 1 ' )

by__Hunt

and

van .der_._ Paorten--(unpPublished)-

for

solving-

diophantine equations, by Steiner [I9771 in connection with the Syracuse ('3-N

problem, and by Cherubini and Walliser fL987) (using a small

home computer only) for determining all imaginary quadratic number fields with class number 1. We shall use it in Chapters 4 and 5.
then (3.1) h o l d s f o r

(-x2,x1)

(pk,qk)

.
(-x2,x1)

B o o L an index

(i). By
k qk

)(t

X*
qk

and (3.6) it follows that is at least the Xo

. Since

- . -

(pk,qk)

for

3.3.

Inhomogeneous o n e - d i m e n s i ~ l approximation i n the real case : the

follows from

- -Xs
xl

(k+l) th Fibonacci number, (3.7)

D8venport lemma. The next case is when

. To

prove (3.8), apply (3.1) and the first

inequality of (3.5). (if). Combine (3.9) with the second inequality of (3.5).
0

A
1

has t h e ' f o r n
b

A-B+x.t3

+ x - 6

2'

where

p . r 0

We

then use the so-called Davenport lemma, which was [1969]. It is. like the homogeneous

of

for which the values of

Iq-$l
q-$

are all extremely small. In practicq is always far enough from the nearest

introduced by algorithm. Put again 8

Baker and Davenport

it appears to be the case that integer (the values of interval

one-dimensional case, based on the simple one-dimensional continued fraction

Is-$[

seem to be distributed randomly over the

[0,0.5] ). This method has been used in practice by Baker and

a1/a2

. and put - 8/e2 . Then we have

$

Davenport [I9691 as we already mentioned. by Ellison. Ellison, Pesek. Stahl and Stall [1972]. and by Steiner [1986]. We shall use it in Chapter 4.

3.4. Let p/q be a convergent of

6

3 The L -lattice basis reduction algorithm, theory.

. with

q > Xo . We now have the following To deal with linear forms with the case n

result.

-2

a straightforward generalization of

would be to study multi-dimensional continued fractions. For to have the &sired efficiency and

, Suppose t h a t , in t h e above n o t a t i o n , 3.3. (Daven~ort)

a good survey of this field, see Brentjes [1981]. However, the available algorithms in this field seem not generality. Fortunately, since 1981 there is a useful alternative, which in a

( b y 1.1 o f (3.1).

we d e n o t e t h e d i s t a n c e t o t h e n e a r e s t i n t e g e r ) . Then t h e s o l u t i o n s

sense is also a generalization of the onedimensional continued fraction algorithm. In 1981, L. bv&z invented an algorithm, that has since then become known as 3 the L -algorithm. It has been published in Lenstra. Lenstra and LovAsz [1982], Fig. 1, p. 521. Throughout this and the next section we refer to this

(3.2) s a t i s f y

Proof.

From (3.5) and (3.10) we infer

paper as " W " . The algorithm computes from an arbitrary basis of a lattice in
-

Rn

another basis of this lattice,

+
7

so-called r e d u c e d basis, which has

certain nice properties (its vectors are nearly orthogonal). The algorithm has many important applications in a variety of mathematical fields, this leads to (3.11). such as _ the factorization te Riele has ,of- polynomials Of

{ F ) _,public-key
are its

-_

cryptography (Lagarias and Odlyzko [1985]), and the disproof of the Mertens

> Xo

Conjecture

(Odlyzko and 525. The

[1985]). a very

interest to us theoretical

If (3.10) is not true for the first convergent with denominator one should try some further convergents. If than (3.10) as desired. If no
q

. then

applications to diophantine approximation. which already had been noticed in

W ,

is not essentially larger of size log Xo ,

p.

algorithm

good

complexity

Xo , then (3.11) yields a reduced upper bound for X of the size of Xo

(polynomial-time in the length of the input parameters), very well in practical computations. Let

and p e r f o m also

can be found that also satisfies

(a situation which is very unlikely to occur, as experiments shov),

then not all is lost. since then only very few exceptional possible solutions have to be checked. See Baker and Davenport [I9691 for details. Summarizing, we see that in this case the essential idea is that an extremely large solution of (3.1) and (3.2) leads to a large range of convergents p/q

r c

Rn

be a lattice. that is givem by the basis

introduce the concept of a reduced basis of vectors

* hi

... . We pl. , according to ttt, p.516. The

p

.%

(1-1,

..., n

ahdtherealnumbers

are inductively defined by

"

i 93

( l s j < i S n )

Then PI,

h;.

.... P,

.. . , n b*
of

is an orthogonal basis of

R"

W e c a l l the l a t t i c e basis

with

Z ,
" 1 ;

r :

E R

Let

i0 be the l a r g e s t index such t h a t

r.

reduced
for

if 1r j

Then, since

Ipinjl a :

< i a n ,
for 1< i

all

and

""

span the same l i n e a r . space a s hl. for -i is the projection of si+l on the orthogonal complement 0

---.

1 ,

+ 0

"* hi

* 2 3 * L lhi-l12 + i , i l i l l

of t h i s l i n e a r space, i t follows that r i

. ;

a n

* . io
0

Hence, by ( 3 . 1 2 ) .

Hence a reduced b a s i s i s nearly orthogonal. For a reduced basis w e have, by Ilt (1.7).

P1. ...,

( r )

{r:2-lhi{ i-1

C2-l< l 2
io

. and

mil

2 2- n 1 2 - l l l

for

1.

.... n .
el. .... n c be a reduced 1 si-si f o r s l , . . . , sn E R , w i t h
3.5, n Let
s
basis o f t h e l a t t i c e

W e remark t h a t a l a t t i c e may have more than one reduced b a s i s , and t h a t the ordering of the b a s i s vectors is not a r b i t r a r y . The L3-algorithm accepts as input any b a s i s

let

. .

of

, and i t computes a reduced basis

y E R" be a given point, t h a t is

.. . , E ~ ,

of t h a t l a t t i c e . The properties of reduced bases t h a t a r e of t(r) the length of the s h o r t e s t non-zero

most i n t e r e s t t o us a r e the following. Let not a l a t t i c e point. W e denote by vector i n the l a t t i c e , viz.

i-1 t h e l a r g e s t index such t h a t t ( r , y ) r 2-

not a l l

si

in

Z . Let

i0 be

io

B Z . Then

Proof, and we denotelby

I

The proof of t h i s lemna resembles t h a t of Lermna 3.4. y

t(r,y)

the distance from

t o the l a t t i c e point nearest

the l a t t i c e point n e a r e s ~t o

So

Iz-yI

- t(r.y)

Let

be

Write

t o it. v i z .

--

. .
ri E Z
s il

with
I

From a reduced b a s i s

lower bounds f o r both

t(r)

alp__ - t ( r , y )

can be ril

ri.

*-

si. si E R

--

--

Let

il

be the l a r g e s t index such t h a t

computed, according t o the following r e s u l t s .

. Then,

reasoning a s i n the 'proof of L e m m a 3.4, w X find

LEJU'U 3.4.

(Lenstra. Lenstra and Lovasz 119821),

Let

el,

..., %

be a
Using (3.12) it follows t h a t

reduced b a s i s o f t h e l a t t i c e

r . Then

Proof.

This is Proposition (1.11)

E

from ttt. W e r e c a l l the proof here. Let

be the l a t t i c e point with minimal length

1 ~ 1

Obviously, then si1

E

il 2 i0

If

C(r)

. Write

Z. sil + ril

il

iO--' the r e s u l t follows a t once. I f

hence

lril-sill

2 1

il > i0 0 and the r e s u l t follows.

The above lemma is rather weak in the extraordinary situation that extremely close to an integer. If one of the other integer, we can apply the following variant. si

is

io

Obviously,
then til follows.

il z i0

If il

il

i0

the result follows at once. If r - t i 2 1 > 6

il > i0

is not close to an

EZ, t
il

+ r

, hence

, and the result

0

3.6. Let pl, % be a reduced b a s i s o f &he.-lattice I' and l e t n y s i - s i f o r sl. sn E R with not a11 si i n Z . Suppose that i-1 such that 0 < 62 s f there i s an index i0 and constants 61

....

....
i

Remark,

3 Babai [I9861 showed that the L -algorithm can be used to find a

lattice point 3.5 or 3.6.

LI with

(x-y[ I c-t(r.y) for a constant

depending on the

dimension of the lattice only. This result can also be used instead of L e m a

Isi!

61

for

iO+l,

.... n

3.5.
Then

The L3 -lattice basis reduction algorithm, practice.

Below we describe the variant of the L3-algorithm that we use in this thesis to solve diophantine equations. This variant has been &signed to work with integers only, so that rounding-off errors are avoided completely. In the algorithm as stated in ttL, Fig. 1, p. 521, non-integral rational numbers may occur, even if the input parameters are all integers. Let di

Proof.

With notation as in the proof o$ Lemma 3.5. let si , for i 2 i0 + 1

nearest to

. and

ti

- si

ti i0

be the integer

for

. Put

c 2"

be a lattice with basis vectors (1.3). (1.24).

hl. .

* .-.P, . Define hi, "ij

di

as in ttL (1.2),

respectively. The i. j

can be used as

denominators for all numbers that appear in the original algorithm ( W , p. with ti E R

. Let

il be the larmst index such that

523). Thus. put for all relevant indices

They are integral, by m 1 . 2 8 ) :

(1.29).

Notice that, with

Bi

* - lkil
--

-3 -

Now, We can now rewrite the algorithm in terms of Bi* " i . j

in stead of hi. i .l , thus eliminating all non-integral rationals. We give this variant

ri. dim

of the ~'-al~orithm in Fig. 1. All the lines in this variant are evident from and. using (3.12). applying (3.13) and (3.14) to the corresponding lines in the original algorithm, except the lines (A), (B) and (C), which will be explained below. We added a few lines to the algorithm, in order to compute the matrix of the transformation from the initial to the reduced basis. Let with column vectors

hl. ...,

t be the matrix , the initial basis of the lattice r

r e

3 1 % Variant of the L -algorithm.

which is the input for the algorithm. We say: the basis basis

is the matrix associated to

bl. . . . ,

Ti :X i Sj

hi

... , which the algorithm delivers as output. Then we define cl, this transformation matrix V by

.c

Let

be the matrix associated to the reduced

:- (b OG 1 ;

-i

ri

:- (d

J -s-1
i

for j-1. ....i-1 ;

for i-1, ...,n ;

S

-r )/d j-1 i.j j

di :- (si.si)/di-l k:-2; (1) perform (*) for if t

k-1 ;

2 2 4~4c-~-4c < 3-4ce1 - 4-Xk,k-1 go to (2) ; k-2.

perform (*) for t if k

.... 1

- S0 -1 . Denote the column vectors of by ..., , and the * T row vectors of v e T . Ye feed the algorithm with 1 and by yl . ... . n I f ' also. All manipulations in the algorithm done on the hi are also done on the u . and the v ! ~are adjusted accordingly. This does not affect the -i computation time seriously. The algorithm now gives as output matrices e . 1 ' and lfw-I , such that e is associated to a reduced basis, e - S - V . and U' - U - V . Note that is not computed explicitly, unless 1 - J (the
, so
S
U

More generally, let

be the matrix of a transformation from some

yl.

SO

to

-1

terninate ;

unit matrix). in which case 'Y*

It follows that

k :- k+l ; go to (1) ; so Sil U* is the matrix of the transformation from So to

C

Note that if as well.

is known. then it is not much extra effort to compute e-l

We now explain why lines (A). (B) and (C) are correct. (A): From U t (1.2) it follows that

. -

for

k+l,

.... n

Define for

- 0, 1, ..., i-1

--

Then
(*)

si(0) Pi , and

si(i-1)

- si .

The

si(j)

is exactly the vector

computed in (A) at the j th if 2-IXk,tl > dt then 'k,ddt

;

step, since

[r

:- integer nearest to

Xk,& :-

r-X 483 r-dt .

for j

1.

.... G l

This explains the recursive formula in line (A). It remains to show that the

occurring vectors

ti(j)

are integral. This follows from

- 'k'k-1-'i,k-1
Finally, from (3.19) we see

% 2 ' i . k

which is integral by ttt p. 523, 4. 11. (B), (C): Notice that the third and fourth line. starting from label (2). in the original algorithm, are independent of the first, second and fifth line. Thus a permutation of these lines is allowed. We rewrite the first, second and fifth line as follows, where we indicate variables that have been changed with a prime sign. and (8) follows.

In our applications we often have a lattice

, of vhich a basis is given

such-that the associated matrix, 1 say, has the special form

where the '5.k-1

:- J5.k-1-"i,k-1
+

ei

are large integers, that may have several hundreds of decimal

~l-"k,k-l-%,k-l~-"i,k ;

(3.18) (3.19)

digits. We can compute a reduced basis of this lattice directly, using the matrix space
d

itself as input for the ~ ~ - a l ~ o r i t h m But . it may save time and split up the computation into several steps vith increasing

pivk :- %.k-l

- pk,k-l*piok ;

to

where (3.18) and (3.19) hold for for i

0. 1.

.. . ,

k-2

k+l.

.... n . The
i

accuracy, as follows.

and by (3.16) also for

di k

remain unchanged

Now, (3.15) is

Let

be a natural number (the number of steps). and let the j -1.

be a natural

equivalent to
n

number' such that i-1,

ei

have

about

k-C

(decimal---digLts. -For

..., n

and

...,

k put

which explains (C). From (3.17) we find

Thus, the j
hence \,k-l remains unchanged. From (3.18) we obtain 'k-k-1 1 *i.k-I+ %-1 relevant j

)
the

are blocks of n x n

C consecutive digits of

Bi

. Define

for the

matrices

A 1 ---.%-1

whence, by multiplying by

%-l-dic_l

and using (3.20)'

If
l
C

~n( j ) w e have

e J
i

L(C.) n 4-j/n
J .

Put

rl

( c ) :

qj
,

(uieh) (3 1

d -TI

1,

...,

and the special shape of n-1 and h

- 1,

...,

and

w e have

c (3 i.h

.
(

then by )
fir

i .h

Then it follows a t once t h a t

r
Notice &at

J +l
tj
*

- e.r. + *, J dk . since
. (

j r 1
11

let

and
U

-I

such t h a t

1 1 j -1
Y

. For some s1 beknown~atrice~.Thenweapplythe~~~-algorith.

e(k) i 0% . Put VO
U-'

-I .

Instead of applying the L3-algorithm once with times, with factor S1,

.I

a s inpur. v e apply it

J -1

, and

W e thus find matrices

J '

11

and

... ,

Sk

a s input. Thus w e reduce the computation time by a

N o w put

j+l

- e.eJ + oJ -r1 a r e defined f o r

j

For

between

2.5-11 and

3 - n t h i s expression is maximal, about

is required is a l s o reduced.

0 . 4 - n2

By induction

e and ll J ' 1 J s .pr-f e.1 -1-1 J+1 1 j j-1

S

- 1,

So the reduction i n computation time is considerable ( a f a c t o r 10 already f o r

.. . ,

?' &

Note t h a t

5 ). The storage space t h a t

since the C-k

l a r g e s t numbers t h a t appear i n the input have

+

C-(l+(k-l)/n)

i n s t e a d of

Dj

digits. d

s o the
S

-%-I dl , w e have
0

S SU-'

J J-1

s a t i s f y the same recursive r e l a t i o n a s the

&-! - dJ 1
-

f i r a11

. Hence
-

Since

3.6.

Finding a l l short l a t t i c e points: the Fincke and P o b t algorithm. C(r) or C(r,y)

and. it follows t h a t l a t t i c e , which is

I.

-&

Sometimes i t is not s u f f i c i e n t t o have a lower bound f o r the same only. I t may be useful t o know exactly +alIfyectors- x E or (x-yl

and

dk since

a r e a s s ~ c i a-t e d t o bases of

. Moreover,

ek
z

is output of the ~ ~ - a l ~ o r i t h i m t.

is a s s o c i a t e d t o a reduced b a s i s of

--

f o r a given constant

. There

r -- s u c h c h a t 1x1 s C e x i s t s an e f f i c i e n t algorithm

f o r finding a l l solutions t o these problems.

This algorithm vas devised by

Fincke and Pohst (19851, c f . t h e i r (2.8) and (2.12). W e g i v e a d e s c r i p t i o n of

L m t us now analyse the cooputation time. For a matrix

M w e denote by L(A) 3 t h e maxinal number of (decimal) d i g i t s bf .its e n t r i e s . I f the L -algorithm is v i t h a s output a matrix e , then according t o t h e a p p l i e d t o a matrix S

experiences of Lenstra, Odlyzko ( c f . Lenstra [1984]. p. 7) and ourselves. the computation time is proportional to L(d13 i n practice. Since
C

is

t h i s algorithm below. The input of the algorithm is a matrix lattice points d

whose column v e c t o r s span the

r , x
E I.

and a constant vith

1x1 S

a s s o c i a t e d t o a reduced b a s i s , w e assume t h a t L(e) u 1log(aet r ) / n

4

> 0

.
(X

The output is a list o f a l l l a t t i c e

a p a r t from

Figure 2. W e use the notation

.
,
L(DJ)

and

xi

.X

.W e

g i v e t h e algorithm i n
2

f o r t h e column vectors-of

iJ )

f o r matrices

4, S, I, 9 , 1 ,

I n our s i t u a t i o n .

3)

and since

det C

det d

The algorithm can a l s o be used f o r finding a l l vectors

of which t h e

55

Firmre 2 L The Fincke and Pohst Algorithm.

distaqce t o a given non-lattice point Namely, l e t

i s a t most a given constant

qil

:- aij

for

1S i I j

~j /qii qji :- Q i j * q i j :- q. qkt :- qkt - qkieqiC f o r i + l I k I C

n ; for

1 5 i

<j

I n ;

In

for

1I i I

and l e t

ri

be the integer nearest t o

for a l l

. Put

rii :- /qii rij : r

compute

for

1s i s n ;
0

i , r
~1-I;

for

1s j < i 5 n ;
of 2-'

compute a row-reduced version 1-I -11-I ; t h a t 1-I compute 1 let

- R-a

1-I

and

11, U'

such

Then
1E

ly-11

<

C'

f o r some constant also g

C'

( C*

n - p-zlbil
y

w i l l do).

Since

it s u f f i c e s t o search f o r a l l l a t t i c e points

; r

and compute f o r each such such t h a t

determine a permutation

1
s

---

- z+

with Ig-yl

lgl s

C*

since

<

implies

1 ' be the matrix with columns

for i

r-'(i) qiJ
:- ail

qji :- qij qkC :- qkC

for

1 5 i

j I n ;

qij :- q i p i i for qkieqiC f o r i + l s k

1I i < j n ; 5 C s n for 1 s i

3.7.
I

Homogeneous multidimensional a p p r o x h t i o n

i n t h e real case:

real

n ;

approximation l a t t i c e s . Let the l i n e a r form

A

have the form

W e assume t h a t

r 2

The case

has already been discussed i n n

Section 3.2, but the method of t h i s s e c t i o n works a l s o f o r

it is i n t h i s case e s s e n t i a l l y the same method.

In fact.
-

Let

be a large enough integer, t h a t is of t h e order of magnitude of

-- -

X :

y E N approximation lattice

Let

be a constant (we w i l l explain its uke l a t e r ) . W e &fin= t h e

I ' by giving the matrix

(5)

if

xi

for

1s i

compute and p r i n t go t o (2).

terminate ;

U-(x u-l(l) '

(n)

'x r

of which the column vectors is a s u b l a t t i c e of

A l a t t i c e point

hi..,

...

a r e a b a s i s of t h e l a t t i c e . Then yn-l- [ y - C ~ 6 ~ ] which is of s i r e

zn

of determinant

I .

has the form

and (3.22) follows a t once. where t h e xi a r e i n t e g e r s , and

n A - 1 ~ ~ - [ 7 - C -. 6~]
i-1

Condition (3.21) can be checked by computing a reduced b a s i s of t h e l a t t i c e - - --3 r by t h e L -algorithm, and applying Lemma 3.4. The parameter y is used t o keep the "rounding-off The length of the v e c t o r
g

error"

Clearly, both

is c l o s e t o
1A1

Xo

and

7-C-A

now measures . r e l a t i v e l y small. This is of importance only i f

C C

which a r e e x a c t l y t h e two numbers we want t o balance

is n o t very

e express t h i s i n t h e following lemma. with each o t h e r . W

&EMMA 3.7.

large.

usually only i f one wants t o make a f u r t h e r reduction s t e p a f t e r t h e f i r s t

Let X1 be a p o s i t i v e number such that

s t e p has already been made. For l a r g e

I t may be necessary, i f
C

, simply take

- 1.

is not very l a r g e . t o use a more r e f i n e d method

Then (3.1) has no s o l u t i o n s w i t h

of reducing t h e upper bound. To do s o , w e use the following lemma, which i s a s l i g h t refinement of Lemma 3.7, together with the a l g o r i t h o of Fincke and lxil for different
i

Pohst ( c f . Section 3.6). I t i s p a r t i c u l a r l y u s e f u l i n the s i t u a t i o n t h a t one has d i f f e r e n t upper bounds f o r the Remark. size Proof,
X :

W e apply t h i s lemma f o r

X1

- Xo

I f condition (3.21) then f a i l s . C of t h e l o g Xo

we must take a l a r g e r constant

. then

338p. Suppose t h a t f o r a s o l u t i o n o f (3.1)

I f it holds f o r a constant
X

(3.22) y i e l d s a reduced lower bound f o r

of s i z e
XI

. -

Let

xl.

....

be a s o l u t i o n of (3.1) with

0 < X

Consider

the l a t t i c e point

---

-- -

i-1 '
as above. Then
-

with

Proof.

Define t h e l a t t T c e point

a s i n t h e proof of L e &

3.7.

By (3.23)

and (3.24)

and

1
The r e s u l t follows a t once by (3.1). W e proceed a s follows. Choose a constant
0

which i s a t lost have

n-X1

B y (3.1).

(3.21) and t h e d e f i n i t i o n of

C(r)

we

t h e upper bounds f o r bound f o r

1xi 1

imply (3.24).

--

Co

such t h a t i f

1x1

>

Co

then

I n t h a t case v e have a ncv upper we have a n upper bound f o r t h e

from ( 3 -25). 1 1 ; case

length of t h e v e c t o r

.W e

1x1

d Co

compute a l l l a t t i c e p o i n t s s a t i s f y i n g t h i s bound

by the algorithm of Fincke and Pohst, and check them for (3.1). Summarizing, the reduction method presented above is based on the fact that a large solution of (3.1) corresponds to an extremely short vector in an appropriate approximation lattice. Since we can actually prove by - ---computations that such short vectors do not exist. it follows that such large solutions do not exist. We will apply the above described techniques in Chapter 5.

Then the solutions of (3.1).

(3.2) satisfy

x
Proof,

1 < z.logl$ l+l/(n-l)-c/lenl-c*-(n-l)-~J.

The result follows at once from

3.8.

Inhomogeneous multi-dimensional approximation in the real case: an To apply this generalized Davenport method in practice, it is necessary to (pl,. . .pn-l ,q) . We indicated in Section 1.4 how this can be done with the L3-algorithm. As lattice we take compute the simultaneous approximations the one associated to the following matrix:

alternative for the generalized Davenport lemma. Let

be the most general linear form that we study, viz.

where

(the case

-2

1 has been dealt with in Section 3.3, but can of where C

-C

be incorporated here also). To deal with this inhomogeneous case, two methods are available. The first method is a generalization of the method the homogeneous case of the previous section. Davenport that we discussed in Section 3.3. The second method is closer to n is a constant of size Xo

. Then

c1 ,

the first basis vector of a

a ';X

First we explain briefly [1971a] (where only the case the generalized Davenport method. n 3 is treated). Put

reduced basis. will have length of the size of can be written as See Ellison

But

sl

for some and

pl,

..., pn-l.
--

. It can be
,

expected that
-__/-

q
----

is of size
-

' : x

,
-- --

Let q

(pls...,pnl,q)

be a simultaneous approximation to

of the size of

xghl

such that. for l+l/(n-1)

di.

1.

..., n-1

...,

6k1 with

are of the size

9 ' .

c1

X n

q l l n

Ide-p /ql i i l , as desired.

SO

that

are of the size

x~-~/c-x~-~

' / q 1q-pi/q/ < c for a small constant c '

The above method has been applied in practice to solve Thue and Thue-Hahler

equations by Agrawal, Coates. Hunt and van der Poorten [1980] (using multidimensional continued fractions instead of the L3-algorithm). Peth6 and b 1. So Schulenberg [I9871 and Blass, CLass, Heronk and Steiner [1987~],(1987

it has proved to be useful. However, we prefer another method, for several reasons. Firstly, it is close to the homogeneous case as described in the previous section, whereas the generalized Davenport method has no obvious

counterpart

for

the

homogeneous

case.

Secondly,

it

actually

produces

3 . 1 0 .

+t

X1

be a positive constant such that

solutions for which the linear form A under the condition the

is almost as near to zero as possible Finally, if a linear relation between the method detects these

Xo

Thirdly, an analogous method for the p-adic

2 2 W , y ) 2 (C(n+2) +(n-l)y ) .X1

Then (3.1) has no s o l u t i o n s w i t h

case can be given (see Section 3.11).

ai

exists, but had not been noticed before (a situation that sometimes Thue equations),

occurs when one solves e.g.

relations. by finding explicitly an extremely short lattice vector giving the coefficients of the relation. Concerning computation time we think that the tvo methods are about equally fast. The method works as follovs. We take the approximation lattice chosen properly, i .e. a reduced basis its inverse C

Peolark, We apply this lemma for

we musc take a larger constant C

Xl

Xo

If condition (3.26) then fails, C

If it holds for a constant of size

5

of the
log Xo

X size :
exactly as
7 . C

, then (3.27) yields a reduced lower bound for X

xl,

in the homogeneous case, cf. the previous section, with constants

sl. ...

.c

X is of the size :
of

.
e

proofL Let

Compute vith the L3-algorithm the lattice point be the matrix associated to

...,

x n

be a solution of (3.1) vith

0<X

XI

Consider

r .

Let

this basis, and compute also the transformation matrix

Y '

Note that

. and

t vith

hence also

e '

1 - Y , and

are easy to with

compute, namely by

Put
--

7i

[y-c-fi] +

X0 . Then

and the L3-algorithm. Let

zn

be defined by and

--

where the coefficients

si E R

can be computed by

To be more precise, if g/[y-C-6n] as

f1

has

as

n th column, then By (3.1). -(3.26) and the definition of

n th column, so

C(r,y)

the result follovs. since

Again we may prove refinements of the above l e m a , similar to Lemma 3.8 the homogeneous case. We explained in Section 3.5. how to apply the Fincke

63

and Pohst algorithm in the inhomogeneous case. ye do not work that out here. Summarizing, the method described above is based on the fact that a large solution of (3.1) in the inhomogeneous case leads to a lattice point extremely near to a fixed point in

LEUHA 3 -11
such that

Let

X1

be a positive constant. Let

be the minimal index

zn .

We can actually prove by some

Then (3.28) has no solutions with

computations that such lattice points do not exist, so that such extreme solutions do not exist. The method outlined in this section is used in Chapter 8. Note that in the case as the Davenport lemma. n

the method is essentially the same Remark& We apply the lemma with X1

Xo

The assumption behind the lemma ui r are distributed satisfying (3.29)

is that in the p-adic expansion.of 6 ' 3.9. Inhomogeneout zero-dimensional approximation in the p-adic case. randomly over
(

no long sequences of zeroes appear.

In fact, it seems that in our applications the numbers 0, 1,

. ..,

p-l 1

Then the minimal

In the p-adic case we start with a very simple linear form a very simple reduction method applies. Let
A

A , to which also

will not be much larger than upper bound of size Proof- Let log Xo

be

log Xo/log p , and then (3.30) yields a reduced as desired. ord (6'-x) 2 r P

X1

satisfy (3.28). Suppose that

. Then

such that @ / ~ E O ~and X E Z , x > 0 . It is obvious P that in the real case with such a simple linear form A inequality (3.1) has only finitely many solutions (we even don' t need (3.2)) , and that they are for 4,

ED

x
By

-1

i r+l ui-p (mod p 1 i-0

x r 0

it follows from (3.29) that

easy to compute. In the p-adic method. Put

6 '

case however. inequality (3.3) and a red

infinitely many solutions. so we do need a bound like (3.4).

- -@/a .

Then 8 ' E 0

. Inequality

(3.3) now becomes

-. In

which contradicts the assumption x follows from (3.28). the

. Hence

ord (6'-x) P

and (3.30)
0

proof it is esseiiEiZi1thae- x - r 0 ui

It --isrhowever--not-

difficult to formulate a similar result that holds for all p z 2 'for p-adic digits

x E Z
z 0

by

-looking, if
z p-1

, and if p

that -are notonly ui with

but also

for p-adic digits

ui z u ~ + ~

Then (3.28) has no solutions if

ord (6') < 0 . Hence we may assume that P is a p-adic integer. Let the p-adic expansion of 6' be

8 '

A method very similar to the one described above was used by Wagstaff [1979].
(19811 for solving congruences such as in Chapter 4 . 5 " 2 (mod 3n)

. We

apply the method

where ui

ui E

( 0.

1.

.... p-1

for all

No

Compute the p-adic digits

far enough to be able to apply the following reduction lema.

3 1

Homogeneous one-dimensional approximation in the p-adic case : p-adic

Figure 3 ,

p-adic approximation algorithm.

continued fractions and approximation lattices of p-adic numbers.

x
Let
A

:- [l,a(qT

have the form (1)

-

if

1x1 > 1x1

; y :- (O,~P)= ;

interchange a such that

and

y ;

compute K E Z
y :- y

I y-K-XI

is minimal ;

K-a ;

where assume that

e2

S A P

such that

- al/e2

E Q

ord (6) 1 0 P

. Now

, and

xl, x2 E Z

if We may

1x1

>

1x1

, interchange 3 and y , and go to (1) ;

print

.
explicitly. Then we can

With this algorithm it is possible to compute (I" ) Ir apply the following lemma. So (3.3) now means that the rational number the p-adic number
6

x2/x1

is p-adically close to 3.12, Let X1

be a constant such that

In analogy of the real case it seems reasonable to study p-adic continued fraction algorithms. However, a p-adic continued fraction algorithm that provides all best approximations to a p-adic number seems not to exist. Therefore we introduce the concept of p-adic approximation lattices, as was From . this paper we adopt the best approximation done in de Weger [ 1 9 ~ 6 ~ ] algorithm. which is a generalization of the algorithm of Hahler [1961]. Chapter IV. This algorithm goes back also on the euclidean algorithm. and thus is close to a continued fraction algorithm. But it is not a p-adic continued fraction algorithm in the sense that a p-adic number is expanded into a continued fraction. and that the approximations are then fiXihd by truncating the continued fraction. Recall that for ordn(84(Ir))
r

0-) > /2.x1

Ir

Then (3.3) has no solutions with

Bepark,

We take

lemna for X1

such that

is of the sire of

: X

a d apply the Xo

Xo

Then we expect that

that (3.31) is a reasonable condition.

(rP )
-

is of the size of
.. .-

so
b

Proof.
6")

Apply the proof of Lemma 3.14 (in the next section) for

E No
--

the rational integer

is defined by
.

The above -method has been applied by Agrawal. -PootCeK (19-801. We use-it in Chapteys 6-and 7.

Coates, Hunt and van der

-----

.I tI
i

and

0s d ( * ) I '
P

<

.=define-&

any

& NO

thebadis

'*

approximation -lattice

by a matrix to which a basis of

I '

' 4 '

is

3.11.
Homogeneous rmlti-dimensional approximation in the p-adic case: p-adic .pproximation lattices. We now study the case

.!

associated. namely the matrix

Then it is easy to seethat

(cf. L e m a 3.13 in the next section. where we prove a more general result). The following algorithm computes the point of minimal length in

. )

r . P

where

ei

E Q

L 2

. We

P may assume that

such that

E Q xi E 2 for all i d P ' ord ( 8 ) is minim1 for i n P i

8 /d

. Put

i. j

and with

6 ;

- ailen
for all

for i

Remark, 1,

We take X1

..., n-1 .

lemma for

such that

pP

is of :the size of : X C ( r ) P

Xo

. Then we

expect that

is of the size of

, and apply the Xo , so

Then 6 ' E Z i P

that (3.33) is a reasonable condition. Put proofL Let xl,

..., xn
(xl..
5

be a solution of (3.3) with

Xl

Then (3.33) I . P

prohibits the point by Lemma 3.13, approximation lattices can be generalized directly from the one-dimensional case. Namely, for any p E No we define The definition of the p-adic

. ..XJ=
p-1

ord ( A ' ) P

from being a lattice point in and (3.34) follows from (3.3).

. Hence.
O

r
P

We will apply the results of this section in Chapters 6 and 7.

as the lattice associated to the matrix . 3.12. case. Finally we study an inhomogeneous p-adic form Inhomogeneous one- and multi-dimensional approxinurtion in the p-adic

Then we have the following result. 3 -13, The l a t t i c e

rP

, associated t o the above defined matrix

P ' where and

i s equal t o the s e t

8 . ei
n 2 2

n such that j/6 6 /6 E Q and xi E Z for all i. j P j ' i j P We assume that ordp(bi) is minimal for i n , and that

ord (B) r ord (6 ) P P n

. Put
for i p '

.-

F a L

For any such that S p .

xn
Hence

- (xl.. . . . x d T
. Then
-

E F

xi

zi

P for n-1

there exists a i

1.

.. .
.

n-1

(zl.. and

...zJT

zn
3

6;

-t9i/6n

1. n-1

...,

n-1 ,

B'

8 / 6 ,

n-1

1 ~~-6~(') + m-pP
--

i-1

-1

A '
~ ~ - (mod 6 ' $ ) ~
-

Alen

- 1 xi-*f + Xn .
i-1 1
-- .
'

.
Then

i-1

ord-(A*) r p

ord ( A ' ) - t P

Conversely, for any

E

8'. 6 ; E Z

( x
-

.. ., x T
0

such that

lattices
p E

there obviously~sxists a

zn

such that

SP-z

.
3.12.

No

P the lattice

-rp

for all i As p-adic approximation lattices we take the -P ~ti3Ztiwerk -&fined for the homogeneous case, i .e. for any

P .

that is associated to the matrix

. -

--

(see Section

3.11). C(rP)

Put further

Using the ~ ~ - a l ~ o r i twe h . can compute a lower bound for

. Then we

can

e apply the following lemma, which is a direct generalization of L

3.14, Let

be a constant such that

where

cl.

..., E,

is a reduced basis of

3 . 6 we can compute a lower bound for

C(r,y)

P '

and

si E R FBy-Lema 3.5 or is useful in view of the

WP) > /n-% .

Then ( 3 . 3 ) has no solutions w i t h

. This

following lema.
t

.b

J B l M3.15%

The s e t

rP (y)

-5+

i s equal t o t h e s e t

3.13.

Useful sublattices of p-adic approximation lattices.

rp(y)

(xl ..... X J E~ 2" 1 ordp(Aa) 2 p I

.
Note that

In our p-adic applications of solving diophantine equations via linear forms, we always have linear forms in logarithms of algebraic numbers, i.e. in

ProofL Let

&

(xl .....x,)~

satisfy I

y E

rp .

the By Lelma 3.13 we have

and

a i 8 s are p-adic logarithms of algebraic numbers, say

In Section 2.3 we have seen that for a ordp(logp(0) The left hand side is just ord (A') P , which proves the lemma.
i
D

( E Op

if

ordp(E()

> l/(p-1)

then

ord ( 1 5 0 P

In our applications we apply this to

Obviously, the length of the shortest vector in (unless in the case lattice) is equal to l(rP.y) following useful l e m a . Let X1 be a constant such that

y E T , , ) . We have the

C(y)

(a translated

> >
5

for which

ord ( ( 1 ) is large. This implies that ordp(lo$(()) is large P too, on which we based the definition of our approximation lattices. However, ord (lo

the converse is not necessarily true:

L -

C(r .y) > Jn-X1

C ,

Then (3.3) has no s o l u t i o n s w i t h

imply that ord ( ( 1 ) is large. This is due the fact that the p-adic P logarithm is a multi-branched function. To be more precise. for any root of 0 (cf. Section 2.3). In 0 there exist unity c E Op we have logp(<) P only the (p-1) th roots of unity if p is odd, and only 21 as roots of

$ to

(4))

being large does not

unity if odd, and

-2

- -1

Let if

<
p

be a primitive,fp-1) 2
(

th

root of unity if

is

It follows 'that

implies that for some k E Remark. l e u for We take

--

0, 1.

..., p-2

ord (lo (()) being large $ P (or k E ( 0, 1 ) if p 2 )

KO

-5 .

is of the size of : X and apply the P .y)--ls -of the size of Xo so-; Then we e x p e c t X a t such that
pp
.

that (3.35) is a reasonable condition.

Let

--

Boof 3.15.

xl.

.... xn

be a solution of (3.3) with

X s X1

. Then

(3.35)
0

prohibits the point ord (A') < p-1 P

( x l . .

.- , X J ~

from being in

rP(y)

Hence. by L o -

x , x such that ord ( ( 1 ) (or . n if one wishes) is large, is a sublattice r* (or rP ) of r P P P C In the following lemma we shall prove this fact, and indicate how a basis of this sublattice can be found. Then we can work vith this sublattice instead ord ( ( f l ) of by all
I'

It turns out that the set of

....

.-

and (3.36) follows from (3.3).

itself. Of course. in Lemmas 3.12. 3 - 1 6 and 3.16 we can replace I ' P P ' : For shplicity we assume that a; E Qp for these sublatxices r* I i

Ve shall not apply the above l e m in this thesis, so we have included it here only for the sake of completeness. However, when solving Thue-Hahler equations (see Section 8 . 6 ) . it will be of use.

. We

take

aO

translated lattices I ' (y), 'I (y) P P

-* .
P'

and leave it to the reader to define appropriate

#

LPIHA
a11
i

3.17&

Let

al,

md

ord (lo (ai)) P $

..., an,$ 0P

for the case

aO r. 1

.
xl. ord (ai) 0 P x E Z n

be given numbers w i t h i

minimal for

.t

...,

for

. Put

the only roots of unity in (mod (p-1))

Q for p 2, 3 . Note that P r . The points x of a linear function on

k(x)

is are are

characterized by
Put f o r any
C, E

(p-1)/2

I k(x)

N0

characterized by

P P

(xl,.. . .xn) E

zn

I ordp(logp(0)

2 P +

.
-3
.
P

that for

- 1, ..., n-1

(p-1) I k(x)

. It

r*
P

C1

and the points

of

r :

follows from the definitions in the lemma

r*
1

(xl....,xn) E 2" 1 ord ((21) r P + Po I P (xl.. . . .xn) E 2" 1 ordP((-1)

P

.
.
p Note that Write

rX P
Then then
#

pO

rPC rPC r

are l a t t i c e s . If
p

r*
P

I .

'4

. Let further
f o r any

Define

k(r)

.-

they are a l l equal. I f

hl. . . .
r
as
P

a 3
.XJ'

Let

hl. ...
by

. 4 be a

.
#

and

, : P

. . .. c

l ,

are both bases of

I ' P

basis o f

(xl..

..

rP

for integers yi. yi

Let

. Then

it follows that

h i . .. . b <

be a b a s i s o f

rP such that

Put f o r

- 1.

.... n-1

and

p 2 5

So

if and only if

yn I yn

. and

E I .

if and only if

# 7 ,

# 1 yn

.
0

This proves the result.

and f o r

p r 3

also

Further put f o r

p 2 5

and f o r

p 2 3

also

Them

PI. ... h ,

i s a b a s i s of

rP
p

and

hl.

. . ..

is a basis o f

I .

Proof.

It is trivial that

$E

equalities of the lattices for

C rP , and that they are lattices. The

2. 3

follow from the fact that

fl are

72

hyperbolic case this suffices to be able co find all solutions of (4.1). This is based on a trivial observation of the exponential growth of )Gn( in this case. In the elliptic case the situation is essentially more complicated. lGnl can be obtained from the complex Then information on the growth of Gelfond-Baker theory. Therefore in this case we have to combine the p-adic arguments Acknovledgements. The research for this chapter has been done partly in with the one-dimensional homogeneous or inhomogeneous real diophantine approximation method, cf. Sections 3.2 and 3.3. We shall give explicit upper bounds for the solutions of (4.1) which are small enough to admit the practical application of the reduction algorithms, if the parameters of the equation are not too large. Petha (19851 pointed out 4.1. Introduction. that essentially better upper bounds hold for all but possibly one solutions. The generalized Ramanujan-Nagell equation

cooperation with A. Peth6 from Debrecen. The results have been published in b Pethi5 and de Weger [I9861 and de Weger [I986 1 .

In this chapter we present a reduction algo~ithm for the following problem.

OD

Let

A, B. Go. G1

be integers, and let the recurrence sequence

(Gn'n-~ be

defined by Gn+l- A-Gn - B-Gn-l for Assume that let pl, equation
A

- 1. 2.

... .
w

where

k E Z s and

is fixed, and

x. zl,

...,

E No

are the unknowns, can be

A > 0

- A2 - 4 - B
ps

reduced to a finite number of equations of type (4.1) with is not a square. Let be a nonzero integer. and (4.2) with survey),

.. .,

be distinct prime numbers. We study the diophantine

-1

. Equation
Bremner,

has a long history (cf. Hasse [1966], Beukers [I9811 for a applications in coding theory (cf. [1983], HacWilliams and Sloane

interesting

Calderbank, Hanlon, Morton and Wolfskill been 'solved

[1977]. and Tzanakis and Wolfskill [1986]. [1987]). Examples of (4.2) have using the Gelfond-Baker theory -by Hunt and van der Poorten They used real or complex, not p-adic linear forms in

in nonnegative integers cases)

n, ml,

.. .

.
A

We will study both the cases of (the 'hyperbolic' and 'elliptic'

(unpublished).

positive and negative dAscriminant

logarithms. As far as we know, none of the proposed methods to treat (4.2) gives rise to an algorithm which works for arbitrary values o f pi's

.- I-was

s h a m

by^ Hahler [1934] that ) 1 . 4 (

A > 0

has only finitely many-

solutions. For the case

Schinzel [I9671 has given an effectively

.
t

k and the --.. .=-.-- whereas Tzanakis' elementary method (cf. Tzanakis (1983)) seems to be
s .a 1Our method has both

computable upper bound for the solutions. b Mignotte [1984a], [I984 ] indicated how in s o w instances (4.1) with method will work for any equation (4.1) with seems not to be generalizable for
A > O ,

the only one that can be generalized to

properties. 1 This chapter is (GnI organized as follows. In Section 4.2

-

can be solved by congruence techniques. It is however not clear that his s -- 1

we

give

some The n

Moreover. his method

preliminaries on binary recurrence sequences. In Section 4.3 we study the grovth of solving l G , (

s > 1

Pethi3 [I9851 has given a reduction

both

in the hyperbolic and the elliptic case. v E N

algorithm, based on the Gelfond-Baker method, to treat (4.1) in the case w - s o l .

hyperbolic case is trivial, and in the elliptic case we give a method for

< v

for a fixed

. by

proving an upper bound for

that has particularly good depedence on Our reduction algorithms are based on a simple case of p-adic diophantine approximation, namely the zero-dimensional case. cf. Section 3.9. In the (4.1).

v , and by showing how to reduce recurrences.

such an upper bound. Section.4.4 gives upper bounds for the solutions of Section 4.5 treats a special case: that of 'symmetric'

For this special type of recurrence sequences our *duction

algorithms fail.

but elementary arguments will always work for solving (4.1) in these cases. Section 4.6 gives a lemma on which the p-adic part of the reduction procedure is based, and some trivial cases are excluded. In Section 4.7 we give the algorithm for reducing upper bounds for the solutions of (4.1) in the case A > 0 , with some elaborated examples. The same is done for the case A < 0 in Section 4.8. Section 4.9 shows how to treat the generalized Ramanujan-Nagell equation (4.2).
as an application of the hyperbolic case of (4.1).

Namely, if

(G2sG3)

I
p

( G I . ) )

then

p 1 (G1.G2)

and if

(A,B)

then

and if P 1 (G

common factor

.Gn +1) then p 1 Gn for all " 0 0 can be divided out in equation (4.1).

n r no , so the

LEMMA 4.1.

Let

n, ml.

....
i

As an

assumptions. we have f o r

- 1. .... s

be a solutionof (4.1) either

example we btermine a11 integers x x2

such that

x2

mi

7 has no prime factors

Then, or

the above

or

larger than 20, thus extending the result of Nagell (19481 on the equation + 7 2n (the original Ramanujan-Nagell equation). Finally in Section

4.10 we give an application of the elliptic case of (4.1) to a certain type of mixed quadratic-exponential diophantine equation, analogous to the application of the hyperbolic case to solving (4.2). As an example. we determine the solutions Suppose Pi 1

PrePf.

X. ml, m2. n

of

Pi 1 Gn for 111 n Then, by 0.8 B ,

Then pi 1 A hence. from (4.3) and (BOG1) 1 , Thus. mi 0 or n 0 . Next suppose pi 1 B .

Now,

and

are algebraic integers. so their

4.2.
Let by

Binary recurrence sequences.

aD

nornegative. It follows that they are zero. Put

E
be defined

and for all

adic orders are i-A-p-A . Note that

132

A. B. G ~ : G1

be given integers. Let the sequence

(GnIn4

Gwl
--

, a.

A-Gn
-

- BSG*~
x2 a/fl

for n
%-

1, 2,

... .
7-

(4.3)

suppose that (Go ,C1)

__

.-

Let

fl

be the roots of

- A-x + B

. We

assume that

- 4-B

.-Hence-la
+

pi 1 6

then we infer that 0

. Next

suppose

--

is not a square. and that not degenerate). Put

is not a root of unity (i.e.-the sequence is Since

ord (A-/A) pi

ord (p-/A) pi

--

- pi 1 pi 1 E,

G n z for all
-

n -

since

then

ord (E) pi

* . / A

and

*-/A

are algebraic.integers, the result follows.

Then n r O

and

are conjugates in K

From Lemma 2 . 1 it follows that we may assume without loss of generaliq that

@(/A)

. It

is well known that for all

(4.6) holds for i 0 for i ord (w) pi is trivial if A > 0

-.

1. 1.

.... s .... s . The

t8

Of course. we may also assume that special case s 0 in equation (4.1)

and will,be treated in the next section for all

(cf. Shorey and Tijdeman [1986]. Theorem C.1). (4.1).

Since our aim is to solve

we see from (4.3) that we may assume without loss of generality that

4.3.

The grovth of the recurrence sequence.

E
First we treat the hyperbolic case A > 0 . Note that la1 rr 1/31 , since the

-A-p-A ,
(

sequence is not degenerate. So we may assume sequence (Gn)n-O

m

lo1 >

u2
U ;
C1 C3

- A-max

r, log B ) ,

u3

We have the
--

A-max ( r, log E

following, almost trivial, result on the exponentiality of the growth of the

Let

---

-(

min (

u2. u3

. U; -

max (

u2. u3

.
+
B) ) -4/log B

3.362~10~~-~~.~~-10~ C2 (2log(4-e-U3) e-~~) , log(r/2- lp( ) + C1-C2 + Cl log(4.Cl/log

Note that

> 0
Let

.
A > 0

THEOREM 4.4,

Let

v E R , v 2 1 max

. All

solutions

n r 0 of (4.7) satisfy

4 n < C3 + --log

( v. 2. IGO.p-/~() .

LEKMA 4.2,

If

n r no and

then

n lGnl 1 y-lo1 it follows for lpl-1

.
n r no that Remark, Note that C3 does not depend on
v

ProofL By (4.5).

la1 >

no > 0

I G ~ ~ . ~ Ul I~ -+ ~ p-Q-~l a IAI

jl-nr 7 .

The following corollary of Theorem 4.4 is immediate COROLLARY 4.5,

Let

A < 0

. Any

solution

n, ml, ..., m

of (4.1) satisfies

We apply this to (4.1) as follows. COROLIARY 4.3, n 2 no

Let

A > 0 . Any solution

n. ml.

. . .,

of (4.1) vith

satisfies

Proof (of Theorem 4T4) .- Note that

lo1

- 1/31 - /B

r ( 2

We hive from (4.7)

We may assume n r 2
- -

. Let

-X/p

2ri-*-

Next we study the elliptic case

A < 0
v

. Since

and a/@ is not a root of unity, la1

- i < c s i .

Let

ko..klsZ

r 2

. Since
IAI

and

(a,@)

and

(X.p)

are pairs of complex conjugates,

- I@I

Then

Ik 1 5 1 + A - n s n for j j .

2ri-0 ' a/@ e ,with - A be such thar 1 j - $ + n - p + k

- -

0. 1

. Put

<+$' I s f .

Ipl

Let

v E R ,

r 1

be given, We study the diophantine for

inequality

0, 1

. By

Lenra 2.3 and (4.8) we have an upper bound for

IAIJ

We apply a result of Waldschmidt (see Section 2.3) from the complex theory of linear forms in logarithms, which gives an upper bound for particularly good in v n that is It may happen that A1

. See also

Kiss (19791. Let

In that case.

+ n - 9 E Z , hence

-(A/p)

- ( a / ~ )-~1

, and it follows that Gn

showed that this implies

IRnI i this. Kiss derived an upper bound for

- A-an + p-fin - 0 . Kiss [I9791 2 -\Go! , where Rn - (an-gn)/(a-8) . From

n .,We shall follow his argument, but

the diophantine inequality

we apply another, sharper result from the Gelfond-Baker theory than Kiss did.
...

Note that, by

- /B

which follows from (4.9).

$ j

where dj

vo 0

-+ . We

- max

( v, 2 - I G O - p - / ~ 1 ]/4- [pi , and

have to distinguish between the homogeneous case

the inhomogeneous case

. We apply the methods that have been

$3

and

described in Sections 3.2 and 3.3 respectively. Unlike in other chapters, here we give the results in the form of precisely defined algorithms. Now and . A j
z 0

, since by

n 1 2

the contrary would imply

p E Q

impossible, since a/fi

is not a root of unity. Thus, take A z 0 , and j

-1

, which is
if A1 z 0 First we study the homogeneous case algorithm. Let N the bound found in Theorem 4.3.

1)J

We have

the following for example

otherwise. Then IAj I s

be an upper bound for the solutions of (4.11).

x m-MX ( v. 2- l ~ we

~ - ~ - / l) ~ l
lower bound

.
for choose IAjl V1 -

(4- 9)

From
z

Lemma

2.4

uu(j,n.21kj1)

s 2-n

can

derive W

so that

- log(2.n)

. We

ALGORITHM H . (reduces given upper bound for (4.11) in the case

+j

. The

Note

that

number

a//3

satisfies

JIluS2 cp, B, [PI, vo. N Qut~ut:new, reduced bound

.
N* for n

0 ).

.
such t h a t

(i) (initialization) Choose

no 1 2/1og B

n0/2 B /no 2 2-v0 ;

No :- N ; compute t h e continued f r a c t i o n hence

A - log B . And h(a//3) :

- -A/p
V2

satisfies IpI , w i t h

hence

h(-A/p)

: . l o g

. Thus

- u;-+
+

Y3

and t h e denominators

ql.

""

qt0+1

o f t h e convergents o f

U ;

satisfy the requirements

Lo

SO

large that

qto 5 NO < -qto+l ; 1 :- 0 i

for Lemma 2.4. Ve find

-

(ii) (compute new bound) Ai : max(al, log(2.e-u;)

)
)

...,

; c q t e the l m g e s t

log(2.n) IAjl > exp (XI-(

)
-

integer
-

Ni+l

such t h a t

- . -

- exp (-a1-( log n + C2

a b

(4.10)
1

) .

Combinfig (4.9) and (4.10) we find n

< a + b-log n

where

- m2

and

'uch t h a t

(log aax

( v.

2. I G ~ - P - / ) A~ + log-

qc i+l

Ni+l < qli+l+l :

(iii) (terminate loop)

2-C1/log B

if

no INi+l < Ni

i :- i

1 , g s ! g ~(ii) ;

idas N*
4.6, 1 . 6 8 1 ~ 1 0 log ~ ~ Bg- B , ~ -YX(I, ( ~ * ~ log ~ E) log(2-e-u;)

:- mu(no.Ni+l)

, LEM

The result now follows from Lemma 2.1, since b

2-Cl/log B

Algorithm H t e n u f n a t e s . I n e q u a l i t y (4.11) w i t h

solutfons with
0

*<n

<N

.
Ni
x

has

vhich is certainly larger than e2

proof.

Termination is obvious, since all

are integers. Note that

We now want to reduce the bound found in Theorem 4.3. We do this by studying

~ ~ ' ~ / xis an increasing function for

r 2/log B . Hence, if n r no

> 4 - N i and

iqCi-$j 1

> 2-Ni/qti (if such ti does not

ti such that q&i > 4 - N i ) i

I
Ikjl/n

1 9 1

Ikjl/n

2 V ~ - B - ~ / <~ 1/2n / ~ that n

exist. choose the minimal

It follows (cf.

(3.6))

pd%

. Then % 5

Ikj l/n

is a convergent of

eoto

(ii) ;

and (cf. (3.5)).

2

N* :-Ni ; S m P
-

$.

I Icl
Suppose

-P

J ~ 1 > 1/(ae1+2)-%
for some

.
m
5

LEUHA *

4.7,

Algorithm I terminates. Inequality (4.11) with

n d Ni

i 2 0 . Then

Li

N < n < N only the finicely many solutions found by the algorithm. Proof. some It is clear that the algorithm terminates. Suppose that ia 0 n
5

has for

Hence.

then if

.Iqi'

-$jj( >

2-Ni/qt i

. we have
- -

Ni

for

It follows that if Ni+l 2 no

then n

N i+l

0
'

Next we study the inhomogeneous case bound for n satisfying (4.11)

$j + 0

Again, let

be an upper

a qci-wj+n-Ptkjl+ n/qc i
It follows that n a Ni+l

qci vo B-"I2

+ N i/qCi .

If

lqi' -$ji

c 2-Ni/qCi

. then

LOPVT;
Output;

9 . $j. B. vo. N

N*
for a11 but a finite number of

new. reduced upper bound

explicitly given

n . (i) (initialization) No :- [N] ; compuce che continued fraction then K + napC kj'q&i 0 since it is an integer. qC .v0.~-n/2 5 i .By ( p ) 1 it f~llowsthat n no (mod q& ) . Since qti > Ni * thei i only possibility is n no If qL -v0a-"i2 > , then n Ni+l 0 1 1 0 ~ ~ i 0 imediat~ly. ~f

1. . . . to , with to SO large and the convergents pi/qi for i > 2 - N / q . (If such Lo cannot be that qc > 4 - N O and 0 O &o so large that qc > 4-NO ) ; vithin reasonable take , fo-und -- -- - - . cine, 0 i:-0;

1q40-$jl

t.

- .

'

--

-. -

(ii) (compute new bound)

- .- -

We remark. that in practice one almost always finds an

ti

such that

if

Ni

is large enough.

compute K E Z 0

if

-Sj 1 " : ; compute n 0 Z qci no < qC , i t K no-pCi= 0 (mod q ) ; i i ' n no is a solution of (4.11). & b print ~ an
with

1 K-

.
4 . 4 .

Upper bounds.

In this section we will derive explicit upper bounds for the solutions of ( 4 1 ) both in the hyperbolic and elliptic cases. O u r first step is the application of the p-adic Cheory of linear forms in logarit-. the s . . e log n way in both cases. We use it to find a bound for vhich works in terms of

approprfate message ; Ri+l :- 12-log(4-qC -v0)/log B] ; i (iii) (terminate loop)

if

Ni+l < Ni i :- i

. Then we

mi combine this with the results of Section 4.3 on the growth of

1 ; compute the minimal

ti <

such that

the recurrence sequence, which for the solutions of (4.1) yield a bound for n in terms of the mi no 1 2 (Corollaries 4.3 and 4.5). D be the discriminant of Q(/A)

In the case C5

A > 0

let

no > max

( 2, loglA/pl/logla/81 ) )

and put

Assume that

Let

Put

L
Let d

c6 C ,

log P /

( loglrl

+ min(O.log(~/lwl))

lax ( 8-C4-(log 2 7 - ~ ~ - 841-C4 ~ ~ ) ) ~ . .

log max

( le-~l~'~. I
A

A
1,

. 18-~44-. 18.r./~l )

.
In the case

be the squarefree part of pi

pi

For

A < 0 , put 4 - max { C3 + m-log(2IG~-P-/AI).

...,

put

2 2
pi

if

pi

d , pi

-1

otherwise, or if pi > 2,

if pi

2, d - 5 (mod 8)

e)i

-1 ,

-1

otherwise, PI P ~ - L -+P2fi ~ ~ log n ]3 0

cgSi

log

c ~ for ) ~

- 1.

...,

Then we have the following result, giving explicit upper bounds for the

solutions of (4.1).

THEOREM 4.9.
4.8,
The solutions of (4.1) with

Let

n, ml,
then

..., m
then

be a solution of (4.1).

n 2 no 1,

satisfy

(i). If (ii). If

A > 0

and

n r no

n < C5-C6 and mi < C g e i for

m < C6 i

mi <

log

n13

for

. . .. s .

A < 0

n < C7

and

1,

Proof-

(i). Corollary 4.3 yields n < C5-m . By Lemma 4.8 we now have 3 m < ~ ~ . ( l o g n ) ~ < ~ ~ - ( l o. g ~ ~ - m )

-P

Rewrite (4.1). using (4.5). as If

C4-C5 > te2/3)3

, we apply Lemma 2.1 with

ve find rn < 8.C4.(log Then, by (4.6). n

-

h a 4 0. b C4 - C 5' 2 2 7 . ~ ~ - If ~ ~C4.C5 ) ~ S (e /313 , then

3 , and

< C5-m < C4-C5(log n ) ) a (e /3)

-

- (log n) 3

from which ve deduce

n < 12564

. Now.

m < C4- (ldg h7<-841-cG

. --

(ii). From Lemma 4.8 and Corollary 4.5 we see that

x'

.e

n we i d , i

n ord ( - ) Pi

pi-ord ( - 1 , pi

from which the result follows, since n 2 n Put

The result now follows from L e u 2.1. since

4-Cqlog P/log B

> (e /3)3

I J

4.5.

Symmetric recurrences: an elementary method.

L -

If

has only finitely many nonzero p-adic digits, then there 6 . 5 , or i f r

exIst an Before we give our reduction method for the upper bounds following from Theorem 4.9, we treat in this section separately the cases of 'symmetric' recurrences. for which the reduction methods fail. The reduction methods make use of the zero-dimensional p-adic diophantine approximation, as explained in Section 3.9, applied to the p-adic linear form loge@ for p

(if d -1 ) G where o p or

r E NO

and a

a E Q

such that

. Further,

G a-R , or Gn a -Sn-r , or n n- r -3 ) Gn a - U ( w ) or 6-Vn(w) n 0 if A < 0 . d

Proof. By ord (6) L 0 we have P definition of 6 we infer

for some

No . From the

+ n-log P
ps

9
means that we must study the p-adic number hence
rl

p , .

...,

. This

- (b/aIr-(ec/A)
Gn

is a root of unity. It follows that we can write

an-r + q.an"
+1

It nay however happen that this number 6 is zero, or that all digits in the p-adic expansion of 6 are zero from a certain point on. Then obviously the reduction process of Section 3.9 breaks down, since it is based on the assumption that the p-adic non-zero digits. Define the following special 'symmetric recurrences'. For in Section 6.2, let Rn for d a, expansion of
6

First let

Then

A > 0

and

contains sufficiently many

as defined

Note that

- an -an a - p
( d

--

'

sn

- an + an .
A ) also

-1

is the squarefree part of

'BY

(GO,G1)

-1

it f o l l ~ ~that s

+*-or

- 1,

o r l/(a-p)

. respectively.

and the assertion follows.

.

and for

6 --3-rls&

(with

Or

f * r

- $' (14(-~))
--

Next suppose
-

JBJ 2

. Then

Since

(a.8)
for all n E Z

Note that 2.sZn ,

Go A - (l+q) E Z possibilities are

(B.G1) ' ' 1

.
51

we have

a.fi

c---r q-a f B

By

(A.B)
6

(1) , and from

1 pr

it then follows that

- 1 r
q

we have
0

So

. The
.

T , T ;

un(~). u n ( ; ) - R ~ 3 .R~,

vnb) .vn(~).sn .

- sjn .
i I

fp, f p

if

- -3

result now follows easily, since for and moreover f/(-1) if for all d

the only -1 , and

0

We have the following lemma. We assume that

ordP (8) 2 0

In the cases of Lemma 4.10 we can treat (4.1) as follows. The smallest index n

g(m-p ) > 0

such that grows d & l y

yap

I Gn

grows exponentially with

Also.

Gn Hence

grows exponentially with

Gg b - p1

n , as follows from Lemma 4.2 and Theorem 4.4. exponentially with C It follows that

as the mi tend to infinity. w-p1 -...-pss cannot keep up with G g(a) m is large enough, there exists a prime q ~t follows that ii (Rn), (Sn) have

Gn-Hn-Rn

- R3n/3

Since

Rn

has nice divisibility properties. we have Gn and Hn

useful information on the prime divisors of

We find:

such that a . Now the sequences I Gg(a) but q t special divisibility properties, such as Rn Sn

I Rm if and only if n I m ,

Skn

for odd k
Iord2(S3)

Now, n 2 1

Gn

ord (S ) 2 n

for all

.
q ( G

Making use of this kind of properties it can be proved that

a I Gn . This gives an upper bound for the solutions of (4.1). those solutions a I Gn but q l Gn . We give two examples. Exam~le, Let a
6

whenever since for

0 (mod 16) if and only if 4 n 0 (mod 12) . Note that G4-H4-R4 R12/3 -2 - 5 - 7 - 1 1 - 2 3 . Considering 4 the sequences modulo 5, 7, 11 and 23 we find that 2 7 11 -23 1 Gn-Hn for
and only if n and a11 n

0 (mod 16)

if and only if

n = 8 (mod 12)

4 (mod 12) ,

Rn

--

Hn = 0 (mod 16)

if

0 (mod 4) m s 3

implies

. It

- -

and in fact

11 1 Gn

whenever

16 1 Gn lGnl

. Thus
I8

Gn

2 2 .

follows from Section 3 how to solve

8 0

- 16, B - 1, Go - 1, G1 - 8, w - 1, pl - 2, p2 - 11 . Then 8 - 3-(7, A -: . so A is a root of unity. Hence , for both p - 2 and p - 11 . Note that we have a sequence of type
A

3 ( 7 .

fi

We note that a process as described above can always be applied when dealing with a situation as in Lemma 4.10. There is an alternative way, that we will mention in the next section. It provides immediately a very sharp upper bound for the mi

Sn

here. We have

4.6.
G (mod 16) n G (mod 11) n 2 Gn (mod 11 ) It follows that ordli(Cn) > 0 odd

A basic lemma. and some trivial cases.

8
0
3

-1 6 6

8 8

1 1 1

8
8

-1 6

8
0

We introduce some notation, and then give an almost trivial lemma that is at the heart of our reduction methods for both the hyperbolic and the elliptic cases. Let for i

88
ord (G ) 2 n if and o G3

8
0 or

6
n ~

88 is even or odd, and

Cn

3 , according as

1,

.... s
-

~ 3 (mod-&) i Now,

. Note

that

has exactly 3

it is larger than and and 11 1

2 -11
q

G3 factors 2 , and

F 3 G 1 factor
-

11

. But
2
By Lemna 4.1 the pi-adic logarithms of 105~(a/fi) z 0 conjkated l o g ( ( / ( ' ) /A-Q pi 3.p.

88

. Hence

- -

there is a prime 11

q , distinct from Gn

such that m
2

whenever

Gn

. Thus

; ' I F 2 n q

solutions with

z 0, so that there remain only three solutions:

. Note

has no -1, 0
--

that it is not necessary to know the value of n

A

explicitly. 23 1 Gn

In this case it is if and only if E z r a m ~ l e ~Let A

23 , and indeed it is easy to show directly that 3 (mod 6)

a/#

and

exist. Note that

since the sequence

(G ) n a d

is not degenerate. Note that for 1 0 are conjugates, hence Bi are in

'

also

lo$<

.
G1 1

5. B

(1+p)/3

. We

solve Gn Hn

to the sequence

--

- -d .
13, Go

. Then
Rn

The sequence Gn to

A-an

+ A-2 and

-27, a

/A-QP bi E Q

.
pi

Hence both numerator and denominator of

so

. Hence.

if Qi

fi

we can vrite

+
)

A-an +

1.2

is related

( an

- zn )/(

-OL

by

where

ord (ei) and u E ( 0 , 1, i.4 pi following lemma l o c a l i z e s the elements of ( C ) n terms of the pi-adic expansion of ki

...,

pi-1

for a l l

The

Section 4.3 s u f f i c e t o &a1 with t h i s inequality instead of (4.7). Another t r i v i a l case is t h a t of satisfy
mi 5 l/(pi-l)

with many f a c t o r s

Pi

in

- ei
+

JEHM 4.11,

Let

NO

. If

SO.

ord (ai) < 0 . Then the solutions of (4.1) pi e m m a 4.11, by L

ord (Gn) pi

ei > l/(pi-1)

then mi
Since n
E

ord (Gn) Pi

- gi

fi

ei

+ ord ( n a i ) Pi

.
Z
and max
mi

ord ( n a i ) Pi w e have ord (t1-6~) pi

ord (ai) < 0 Pi

ord (ai) pi

. Hence

Proof,

B y L e m m a 4.1 w e have ord (Gn) pi

+ ord
pi

(ai).

ei

ord pi

[b]n-[~]] w e have

~ ~ d ~ , [ [ $ ] - b ] ~ - l ].

Thus w e may assume without l o s s of generality t h a t

With

ord (0 ordPi(logpi(l+O). pi ord (Gn) + ei pi

( A ) (

ord (0 pi and i t f o l l a t h a t

> l/(pi-1).

Hence

- 1. ....

s , and t h a t i n f i n i t e l y many

pi-adic

ord (ai) r 0 Pi d i g i t s uiSC of

for a l l

ei

are

nonzero.

ord pi

[ n.logpi b]
+
fi

+ logpi

[$] ]

4.7.

The reduction algorithm i n the hyperbolic case.

d i g i t s of that

ord ( n a i ) pi

First w e give the reduction algorithm f o r the case L e m m a 4.11 and Corollary 4.3 only. Let solutions n , ml ,
N

A > 0
N

I t is based on n f o r the a s i n Theorem

be an upper bound f o r

.. ., m
.

of (4.1)
-

For example,

W e have t o exclude some t r i v i a l c a s e s - f i r s t .

ai

The case where a l l pi-adic

C5 C6

4.9.

,
A

from a c e r t a i n point on are a l l zero has been d e a l t with i n holds f o r a11

i

the previous section. But t h i s case can a l s o be d e a l t with a s follows. Note

ei

1,

...
-

AII;ORITHn (reduces given upper bounds f o r (4.1) i f

a. for

>0

).

with the same

-

r , which i s the
-. __

from l p m 4-10, Thus, by L e a a 4.11,

__

B,
n

A. P. w, pl,

..., pS,

.
mi
0

--

Qufaut: new, reduced upper bounds -Hi -for

m i s max ( gi
Then we have, i f
A

ord (n-r), pi

- ei

) s

gi

1-+ ord
pi

(n-r)

. (4.12) -

f o r bi--&

-.... s

. d - N

*
-

( i) ( i n i t i a l i z a t i o n ) Choose'a n
Y :-

no

such that

> 0 , by Corollary 4.3,

111

- lPl-l~/Bl
+

-"o

.
if if if pi pi pi

ng >

log(p/A 1/log1 a/B I ;

gi :- ord (A) pi from which a good upper bound f o r proof of L e m m a 4.10 y i e l d s bi n can be derived. And i f

ord ( l o ~ ~ ( a / p ) ) pi 3/2 1 112

< 0 , the

0 , whence, by (4.12).

hi :- ord (A) pi

2
3 5

1
}

for

1.

.... s

f o r some constant

vO

. Only

minor changes i n the r e s u l t s and algorithms of

(ii) (computation of the

ei's) Compute for

1.

.... s

the first

ri

pi-adic digits uiSc of hence. by

-

u i*si.j

2 0 ,

i + O ; pi r N o a n d u i. ri :- ri + 1 (iii) (further initialization, start outer loop) s i ,o i 1. ..., s ; j :- 1 ; where ri is so large that

for which contradicts our assumption. Thus, mi Then from Corollary 4.3 it follows that
d

gi + s

i . J

for

1,

.... s

(iv) (start inner loop) i :- 1 ; K, :- ,false, ; J (v) (computation of the new bounds for mi , terminate inner loop)
( s N 0 I p ; ~ N : . i n isj 1 -1 Ff si,j < si,j-l then K ,true, ;

and

i.s

+ O 1 ; hence n < N
0

Ff

i < s

---

-1p : i ' ~

Remark 1. (vi) computation of the new bound for n , terminate outer loop) N

. 1
J

In general, one expects that

will not be much larger than will be zero. Then

i.e. not too many consecutive pi-adic digits of is about as large as

terminate in three or

log N . In practice. the algorithm will often j-1 four steps, near to the largest solution. The s , the bottleneck of the algorithm is the

if

N r n O and K j J j :-j+l;w(iv);

computation time is polynomial in

N* :- ux

ni
D E O R E n 4.12,

:-

computation of the pi-adic logarithms. for i

(
(

N ~ . no
hi. gi
+

:
siej )

1,

...,

s ; &gp. terminates.

Remark 2 , For a prime

Petha I19851 pi

gives for such that

s Cn

-1
+
0

a different reduction algorithm. and

he computes the function g(n) , defined for

With all

the above assumptions, Algorithm P has no solutions with

Equation (4.1) wit91 A > 0

*5

smallest index pi-adic limit

A

n 2 0
lim g(u)
UiQ
-- -

py

as the

I Gn

Note that if the

< N , mi > Mi for

exists, then by Lemma 4.11 it is equal to

.
(Gn)&
- d

---k 3.

Proof. exist

Since the pi-adic expansion of

is assumed to be infinite, there 2.1 holds for all j

s
$

If B

- +I

(hence

A > 0 ), we can extend the sequence

ri with the required properties. It is clear that

t 0

ri < s

to negative indices by the recursionformula

and that Si.j 1. i

...,

r 1 s N . So si#j si#j-l j j-1 s for all there is a j such that N S no or si, 1.j-1 j s . In the latter case, K remains ,falsei ; in both cases the

1.0 ' Since

Gnel (cf. (4.3)). with for n E Z ( c n ) &

A-B-G n

- BSG,+~

for n

0. -1, -2,

...

-I

Then (4.5) is true for

n < 0 also. We can solve equation (4.1) (G;I)L

algorithm terminates. We prove by induction on n < N j that For j hold for all j

n < N for some j r 1 . Suppose there exists an 1-1 . From Lemma 4.11 we have " i > gi "i,j
+

-i

mi

for i 1, .... s , and i . J 0 , it is clear that n < No Suppose gi

not necessarily nonnegative, by applying Algorithm P twice: once

, and once for the sequence

GA

Note that

~~-(p.a~+l-B") and

, defined by

C ; I

G-n

such that

Now. instead of applying Algorithm P twice, we can mpdify it, so that it works for all replace n by replaced by n E Z , as follows. Lemmas 4.8 and 4.11 remain correct if we In1

e2
so

0.A9359 05530 7330A LA223 96230 3A006 A3366 83368

In Theorem 4.9

the lower bound for

no

must be

r1 r2

- 90

8270 .... , (since u 1.89 (since u , ,

29

Ul,go 6. 1 1 2 '

> No

289 > No 1.
).

no > max and

( 2. Iloglr/All/logla/BL

llog

(iii) (v)-(vi)

~ ~ , ~ - 9 1 . ~ ~ , ~ - 3 0 ; ~ slS1 s slS3 sle4

S

7 has to be replaced by 7

ri

( IAI - Irl -la/BI

0 " '

, lrl

v 111

(v)-(vi) Similar modifications should be made in step (i) of Algorithm P. Further, in step (ii), should be chosen so large that (v)-(vi) Hence In1 -5

90, s2,1 s 2.2 6, s2,3 6, s2,4

- 29, K1 -

2. K2 1. K3 1, K4

&trueL. N1 < 76.9 ; *true.. ,true" ,false., N2 < N3 < N4 < 8.7 ; 5.8 ; 5.8

5, ml 5 6 . m2 I 2

. We
1 2 n

have 0 1

if

pi+2

then

i pi z N 0 r -1

and

u. + O , l.ri u i.ri
z u

u. + p - 1 ; l,ri i,ri-1
*

n Gn

4
373

-3 64

-2 11

1
4

2 23

3
134

5 4552

Pi

2 No

and

2174

781

and similar modifications have to be made in step (v) for changes. Theorem 4.12 remains true with n replaced by

S L ~

. With

these

So there are 5 solutions: with

-3. -2. -1. 0. 1 .

In1

We conclude this section with an example. Examle, a

4.8.

The reduction algorithm in the elliptic case.

Let 2-/2. 32

-;3

6. B

and

no

1, Go

2-/2. A

. With
of

e60

1. G1

4. w

1. pl

1 + 2-/2 )/4./2, p
we have
7

2. p2

11

Then

We now present an algorithm to reduce upper bounds for the solutions of (4.1) in the case Theorem 4.9.
-

( -1 + 2-/2 )/4-/2

A < 0
N

The idea is to apply alternatingly Algorithms P and one be an upper bound for n , for example n

1.142~10~ we ~ find above

C4 < 2 . 4 9 ~ 1 0.~With ~ the

of H and I. Let

- C7

as in

modifications

Reaark

> 0.323, C , 2
-

2 : 1 . 1 ; C6 < 2 . 6 2 ~ 1 0 ~ C ~ .- C < 4 . 6 2 ~ 1 0 ~ ~ Hence a11 solutions of Gn - - 5 . - - ---~a .x ( y . m 2 ) < 2 . 6 2 ~ 1 0 . ~ We ~ perform the reduction satisfy In1 < 4 . 6 2 ~ 1 0 ~ m Algorithm 0.u u u

P step by

step.

(We .write the p-adic

2..

. and

number--^^^-^^ to

< 1.76,

bLCORITIl C. (reduces upper.bounds

for (4. L) in-the case-A for n , and

Xi

C-R ) .
for
P-

-J

as

a, B, A , C D P~. -.., ps, N * Put~ut; new, reduced upper bounds N i

if
).

p > 10

we denote the digits larger than 9 by the

1 .

..., s .
+
3 if pi

for

symbols A. B, C,

...

(i) (initialization) No :- [N] ; j :- 1 ; gi :- ord (A) pi

- )
2 of

for

1.

.... s

(ii) (computation of the 6 ; ' s .

first

9 , Ip )

Compute for

1. ...

the

ri pi-adic d i g i t s u

i . C

qC

> 4-N j and

(qc

does not e x i s t , choose the minimel where

$

-$I > 2.
+1
for

~ ~ / (if q such ~ ~ tj

such that

ri

i s so large that

Log(-X/p)/2si

u d0;compute i,ri , and the continued fraction

' pi 2No

and

qc, J

) ; j :- j

; g g & ~ (iii) ;

(v) (termination) N* :- Nj-l ; Hi :- Hi,

1.

.... s

; STPP.

The following theorem now follows at once from the proofs of Lemmas 4 -6. 4.7 and Theorem 4.12.

with the convergents largethat

'0

pi/qi
0

for

I,

.. ., Co

, where Co

i s so

O 4.13, R ~

Algorithm

terminates. ~ q ~ t i o (4.1) n with and

mi

q < N <qC to-1 - 0

if $ . r 0

if $ - 0 ;

solutions with
can be found i n a reasonable

N* < n < N

> Mi

for

1.

.. . . s

has no

, apart from

O amount o f time,

iq 1 > 2-N /q

and such

Co

those spotted by the algori t m .

We conclude this section with an example.

qco

> 4-NO otherwise;

For
i

(iii) (one step of Algorithm P)

- 1. .... s

put

H : - m a x ( h i , g i + m i n t a ~ N o ~ P ~ ~a N ndu .rOI); i . J j-1 i,s (iv) (one step of Algorithm H or I)

,-F
and
X

if

$-0 then A :- =(al

let
( 2

+ 4-7

1, B ) / / 7

2, GO

. Let .

2. G1 3 fl. Pl

- -.

then A 3. P2 7

-.

-7. a ( 1 + d-7 )/2 2 We have with no

.....at
j

; v :- Iwl-

s H piisJ , i-1

the following results:

16 29 30 C4 < 6.40~10 , C3 < 9.14~10 , C7 < 7.42~10 , Further, gl

-(C~,~.C~,~) < 2.30x10~~ Theorem 4.9 we may choose

;

choose no r 2/log B

such that

n0/2
B

No

/no

r v/2- lpl

- 1.
. We

g2

0 , hl

1. h2

-0

BY

7.42~10~'

have

compute the largest integer N such chat -1 - Nj/2 B /N,S(A+2)-~/4.lpl ;

--

--. -

Li; NJ < N1-1 -a compute t ificb q4j-l Nj < qe ; J . j :- j + 1- ; (iii) ;6 J.

-

compute

n O ~ z , ~

1 ; K+nOSpC - 0 (wdqC j-1 J -1 if n no i s a solution of (4.1)

~ i .w qith ~ 3-1

"

N ,

< Nj-l
compute the minimal
C j

Now we choose to

- .
61

since

<C

J-1

such that

and qg q6 M

Iq6,*$I

0.24487 ... > 2.N0/q61 N1

37, and we find

.
74

0.104 ...

We have C1

637

- 1291 > 4x74 . and

n
S

2.3 Hence

10102 > 4x637 7, M2,2

and

Iqg.e5
N2

4 , and we find

Iq6-$i
N3
5

3 , and we find
60, ml
5

- 60 . In the next step we

.

6, m2

1.3 * find no improvement. It is a matter of straightforward computation G

- 0.49398 > 2x74/1291

Next we choose

0.38745 ... > 2~637/10102

67, 1.1 9 , since

Then. from (4.13).

Next we choose

C2

-6
M

We have , since

We have

Put

u i-r+l

pi . Then (4.14) and (4.15) lead to

to check that there are only the following 6 solutions of

- fT1-T2 :
with

4.9.

The generalized Ramanujan-Nagell equation.

- y.1-r+l 1

pi , and also to

The most interesting application of the reduction algorithms of the preceding section seems to be the solution of the generalized Ramanujan-Nagell equation (4.2). Let k be a nonzero integer, and let pl, "., ps be distinct prime numbers. Then we ask for all nonnegative integers x. zl,

...,

with with (4.17).

Do-x1 w

- y.

k2

-kl-DO

Ue proceed with either (4.16) or the one with the smaller

whichever is the most convenient (e.g.

Ik21 1-

First we note that (mod pi) for for i i

zi

whenever

-k

is a quadratic nonresidue

If

- 1 . then (4.16) and (4.17) are trivial. So assume

1

. Thus we 1, . , ., t
1,

...,

i Let pi I k and p I k for i t+l, , s . Let ord (k) be odd i . pi -- -r and even for i r+l, ..., t Dividing by large enough

assume that this is not the case for all

powers of equations

pi. for

1, ... , t ,

--

;..

the smallest unit in

(D-2

with of v2 r th

> 1

known that the solutions v, w choose for

that

- D - v2

and k2

N(c) fl . It is well fall apart into a finite

D > 1 . Let

be

number of- a s s e s of associated solutions.

(4.2) reduces to a finite nunber of

v2

- D - w2 - k2

+t ---there be-

T such classes, (md

1.

..., T
+
wr.0

in the

v r .O

-/D > 1
v

are given by

class the solution v 7 . 0 ' w r,O such is minimal. Then a11 solutions - of

hr,n. w

, with

with

p kl for i 1. ..., s and Do composed of pl, ..., pr only. i and squarefree. Ue distinguish between the 2s-r combinations of zi odd or i

even for

r+l, i

and even for

...,

. Suppose

that

z ;

is odd for

u+l.

..., s . Put

r+l,

.. . , u

for

n E Z

where

7;

vrs0

w r.0 ./D

That is.

)* (vrSn n__-

and

are linear binary recurrence sequences. Now. (4.16) and (4.17) (wr,n1* n--reduce to T equations of M e (4.1). If k2 1 , then T 1. rl c. -1 If k2 1 2 - 0 . k 2 * 1 thenit is easy toprove that 7 : Ik21-c. 7i- c -1 , SO that 7 k c

with In both cases. (4.16) and (4.17) can be solved by elementary means (see Section 4.5, of related interest are Stsrmer [1897], Mahler [1935], Lehmer [1964], Runsey and Posner (19641 and Hignotte [1985]). If k2 1 2 - 0 apply the reduction algorithm to one of the equations
w
r ,n

c ( 2 - D , and either

or

composed of factors

and

11

only.

We have :

v
?*"

- fl .

- n
r

. Note that n is allowed to be i-r+l so we can use the modified algorithm of Remark 3, Section 4.7.

m. pi1

i-r+l negative. since

(i)
(ii) (iii) (iv)

z1 even, z2 even, y z1 odd, z2 even. y odd, y odd, y

then we
m-1 pi ,
-

fi

zl even, z2 zl odd, z2

z1/2 z2/2 2 -11 z x/7, (z1+1)/2 z2/2 2 -11 z x/7. z1/2 (z2-1)/2 x, z 2 -11 (zl-1)/2 (z2-1)/2 , x, z 2 11

. -

c c c

. -

1. D 2. D -7, D -7, D

7 ; 14 ; 77 ; 154

In the first example of Section 4.5 we have worked out case (i). We leave the other cases to the reader. Equation (4.19) can be solved by the reduction algorithm. Again we have four cases, each leading to an equation of the type

Thus we have a procedure for solving (4.2) completely. It is well known how
r ,O' r ,o be computed by the continued fraction algorithm for

the unit

and the minimal solutions

for

/D

T can We conclude this

1,

...,

section with an example. It extends the result of Nagell [I9481 (also proved by many others) on the original Ramanujan-Nagell equation x2 4.14, The only nonnegative integers 20 are the 16 x such that

+
x2

2z 7

.
has no

with either (i) (ii) (iii) (iv)

or

composed of factors

and

11 only. We have c--7. D 1 ; 2 ; 11 ; 22

z1 even, z2 even, y zl odd, z2 even, y odd, y

x, z

prime divisors larger than

in the following table.

- x. z
x, z

zl even, z2 zl odd, z2

odd, y - x , 2

z1/2 z2/2 2 -11 , 1 - / 2 z2/2 2 -11 z1/2 (z2-1)/2 2 -11 (z2-12/2 , - 2 -11

. .

c--7.D-

c
c

-7. D -7. D

Case (i) is trivial. The other three cases each lead to one equation of type (4.1). In the example in Section 4.7 we have worke? out case (ii) . With the roiEming data the & r is a correct upper bound. Proof, x2 Since -7 is a quadratic nonresidue modulo 2. 7 and 11 3. 5, 13, 17 and 19 should be able to perform Algorithm P by h a o r

the cases (iii) and (iv), thus coaplering the proof. In these eases

N < l o 3 '

Case (iii):

10

3-dl1

. X

( 2

+ dl1

)/2-(11

we have only the primes

left, Only one factor

7 can occur in

thus we have to solve the two equations

Equation (4.20) can be solved in an elementary way. We distinguish four cases, each leading to an equation of the type Case (iv):
a

197 + 42./22

( 9

2-422 )/2-/22

4(X,Y)

n v-q

is equivalent to finitely many ideal equations

with

(0)-(z) -

n (a-v-q

Hence we have the equations in algebraic numbers

I
pemarks, 1. The computation time for the above proof was less than 2 sec. 2. Let 4(X,Y)

a-X

" 2 - a - 3 - Y- y-n

'

a-X

"2 - a-3-Y y-r

where a-X

is composed of

o , units, and common divisors of finitely many

a-X

- a-@-Y

and
y

a.x2
Ci

coefficients, and integer, and p I,

b - X - Y + c-y2 be a quadratic form with integral

a-3-Y

Note

that there are only

choices for

b2 Ps

4 - a - c positive or negative. Let

be a nonzero

possible. Thus. (4.21) is equivalent to a finite number of equations

....

distinct prime numbers. Then we note that

4-a-O(X,Y)

2 ( 2 e a - ~ + b --~ A )~Y , or, if we put

A

so that the diophantine equations

y/a- (3-B) and

Gn 2

A-/*

in integers X z 0

and

zl.

.... zs

E No

, can be solved by our method.

Here.

(G In is a recurrence sequence with negative discriminant. SO " 2 2 (4.22) is of type (4.1). and can thus be solved by the reduction algorithm of

4.10.

A mired quadratic-exponential equation.

Section 4.8. Before giving an example we remark that (4.21) with with D > 0 is not solvable

In this section we give an application of Algorithm C to the following diophantine equation. Let

with the methods of this chapter. This is due to the fact that in

o(/D)
is to

D > 0 qn by

there are- infinitely many units, hence

-

possibilities for
- -

be a quadratic form with integral coefficients, such that D negative. Let q, v, w be nonzero integers, and pl, numbers. Consider the equation

..., ps

t b2

- 4-a-c

is

distinct prime

the method of 1-1 this chapter. since it does not lead to a binary recurrence sequence if t 2 2

replace

. Another generalization of- eqbation n q,i . This-problem is also not solvable by

--.

infinitely many (4.21)

. -These problems

can however be dealt with using l u l t i - d i ~ e n s i o ~ l in other chapters of this

approximation techniques. that are presented thesis. See Chapter 7.

in integers X , and
Let
@.

n.

~ y ..., ,

ms N o

. . Let
h-n2 h be the class number of
0

We finally present an example.

be the roots of (r)-(?)

e(x.1) n

Q(h)

There exists a

equation

r E 0(ID) such that we have the principal ideal

(qh) . P u t

-%+

with

a nl < h

Then

mEOREM 4.15,

The equation

in X E Z, n, ml, m2

No

has only the following 24

-

--

solutions:

The results of this chapter have been published in de Ueger [1987~]. 5.1. Let Introduction. S be the set of all positive integers composed of primes from a fixed
(

finite set

pl.

. . .,

ps

, vhere

s 1 2 ,

and let

6 E (0.1)

In this

chapter we study the diophantine inequality Proof Put

( 1

/-7 )/2

. Then
in x, y E S

x2 - X - Y +
Note that a(/-7)

2.2

- (x-8.~)-(x-3-Y) .
1 , and that 1' - 7m2 (P-/~)-Y - - 4-7-3 X

. We

give explicit upper bounds for the solutions. and ve show diophantine P1.

how the algorithms for homogeneous, one- and multi-dimensional has class number for finding all solutions of (5.1) for any set of parameters
6

approximation in the real case, that were presented in Chapter 3, can be used

For

Suppose

1 X

8 - Y and
y

I X

3.y

. Then
y

On the o t h c hand,

- fl-Y are coprime. Thus we

I 11-2"

It follows that

- 21 . hence

s r 3

the continued fraction method (cf. Section 3.2) is used. For 3 we use the L -algorithm for reducing upper bounds (cf. Section 3.7).

-2

. - . . Ps.

8 - Y and Tijdeman [1973] (see also Shorey and Tijdeman [l986]. Theorem 1.1) shoued that there exists a computable number c , depending on that for all x, y E S with x >y r 3

have tvo possibilities:

max(pi)

only, such

in each equation the 2nd- and 3rd solve

being Independent. Hence we have to

Thus. for any solution of (5.1) a bound for showed how to solve the equation x

, ( J ) n
with : :

A(J).Bn+~(J).gn

C ; j )

so that)':6 i

- ci2) $ ( ' )

- 2.~2:

for

1. Gil)

1. 2 , and

- -+(2)

- 3y.7 - , m
2 j 1. 2 3, Gi2) For

for j and

-k
.

x. y

follows. Stermer [I8971

with

1. 2

vith an

1, 2 ,

elementary method (see also Xahler [1935], tehmer [1964]). Our method can

~(l)

- I(~) -

(24-7)//-7

,
for

solve this equation for arbitrary k E Z For the one-dimensional case b 2 , Ellison (1971 ] has proved the following result: for all but finitely

-1

. Note

that ) ' : 6

-1

many explicitly given exceptions, all x. y

E

I 2X

3Y 1 > exp(x- (log 2

1/10))

for

Cijsouw, Korlaar and Tijdeman (appendix to Stroeker and of the inequality

we have solved (4.22) in the

example of Section 4.8. It is left to the reader to solve (4.22) for

This can be done with the numerical data given for the case j

.
0

Tijdeman [1982]) have found all the solutions x, y E N

for all primes

Rcpark. The computation time for the above proof was less than 3 sec.

p, q

with

p < q

< 20

and with

-1.

We shall extend

these results for many more values of pl,

p, q

and with 6 - L .

0.9

Further, we t

determine all the solutions of (5.1) for the multi-dimensional case

(

.... p6

- 1 2 . 3, 5. 7. 11. 13

-6

,
A

with

> exp ( -(log X + log(e-log ps)) -C1.(1-6) -log p1 )

Combining this with (5.3) we find In Section 5.2 we derive upper bounds for the solutions of (5.1). In Sections 5.3 and 5.4 we give a method for reducing such upper bounds in the one- and multi-dimensional cases respectively, and work them out explicitly for some examples. Section 5.5 contains tables with numerical data. The result now follows from Lemma 2.1, since s , With Ex-~l e 19 C2 < 1.97~10 . With We assume that the primes are ordered as x, y of (5.1). the finitely many
z E N

X < C1-log(e-log ps) + log 2/log pl

+
C

C1-log X

> e2

.
17 C1 < 2.30~10

2, 2

pi

199, 6

0.9

we have

5.2.

Upper bounds for the solutions. p1 <

- 6 , 2 m pi a 13. 6 - 1

we find

36 Cl < 8 . 3 7 ~ 1 0 ~ C2 ~. < 1.35~10

. ..

< p ,

For a solution is also a 5.3. Reducing the upper bounds in the one-dimensional case. s

for which

z-x, z-y

solution of (5.1) can be found without any difficulty. Therefore we may assume that X Put (x.y)

- 1 . Put

In this section we work out the examples max ord (x-y) . ULcs pi

run pl. p2 through either the set of primes below 200. or the set of non-powers below

2. 6

- 0.9 . and
p2, 6

50. We note that for any other set of parameters p similarly. We prove the following result.

the method works

THEOREHL

(a)

The diophantine inequality

with
THEOREM 5 - 1 - The solutions of (5.1) satisfy X < C2

p2

primes such that

< p2 < 200

.
xl, x2 E 2 . x1 z 2, x2 co%Gdicts--y-r

and

. .
---

proofL

]if y

A-x

y >A -x 2

. put

2. and either

then

log(x/y)

f >x -y . Then

a y

; uhrcx

. So

or 6 has only the

-A
(5 -5) non-powers such that

- 0.9,
.

- --

min

[ pl

-XI

X2 p2

-> 1015-

77 50

solutions listed in Table I. with pl, p2 and conditions (5.5). has only the 74 solutions listed in

By

(b) max(x.y)

The diophantine inequality (5.4) p1 < p2

5

r pl , we obtain

Table XI.

<

< 21-6.p;(l-6)-X
result. Lemma 2.4. to
[0(/PI,

We apply Waldschmidt's p

that the * independence conditione

r 3 we have Vi i

...,/Pn) :O ]

with

log pi

for

r 2 . Thus

2 "

s. q

Beeprks. Note "delta"

The Tables are given in Section 5.5. gives the real number with

In Tables 1. I1

the column

holds. Since

X1 P ,

X2 - P2

" n

I P1

x1

x2

' P2

delta

Note that in Theorem 5.2 we do not demand

(x1.x2)

- .
1

and in Theorem

5.2(b) w e do not demand chosen, min since the

pl.

p2

t o be primes. The conditions (5.5) of (5.4) with

6

are and

W e f i n d from (3.7)

that

k I 92.996

, hence

k I 92

Lemma 3 . 1 f u r t h e

numerous

solutions

y i e l d s : i f (5.3) holds then qk/10 1 1%P2 , ak+l > -2 + p1 qk 20-1

0.9

pyl, p 2 : Write

) + 10''

can be found without much e f f o r t . and i f

.---

Proof,

w e assume t h a t

then (5.3) holds f o r

p ;

>

.
log pl/log p2 have

(x1,x2)

(qk.rk)

.W e

computed the continued f r a c t i o r

s i n c e i t i s easy t o f i n d the remaining solutions. Let the simple continued f r a c t i o n expansion ( c f . Section 3.2)

i n the l o g pl/log p2 mentioned ranges f o r pl, p2 exactly up t o the index n such that 19 ( c f . Section 2.5 f o r d e t a i l s of the computational qn-1 S 1 . 9 7 ~ 1 0 < qn e checked a l l c o w e r g e n t s f o r (5.9), and method). Note t h a t n S 93 . W subsequently f o r (5.10). I t is possible, though u n l i k e l y , t h a t t h e r e is a convergent t h a t s a t i s f i e s (5.9) but f a i l s (5.10). log 15/log 23 22 W e met only one such a

expansions and

the

convergents of

a11 numbers

and l e t the convergents be (x1,x2) then

A

. First

rn/%

for

1. 2.

. .. .

w e show t h a t

x1 2 x2

, hence X

W e may assume t h a t

case: so that

pl

xl

. For
)

if

x1

< x2 ,

a5

15, p2

51. r4

23 , with 19, q4

. Now.

[ 0 , 1 , 6, 2, 1 , 51,

...I ,

(5.9) holds b u t (5.10) f a i l s , s i n c e

x2.10g p2

- xl-log

p1 > X.[

log p2

- log p1

2 X.log 197

-,

199

W e have i n t h i s case
is true. But

and from (5.3) and (5.6) w e then i n f e r 0.0101

. < 0.002771.. . 1 0 ~ ( 1 5 ~ ~ - 2 3 ~ ~ ) / 1 0 ~ ( 20.9008. 3 ~ ~ ) .. > 6

A

0.002714..

20.1.15-2.2 s o (5.1)

, s o (5.3) is not true.

a 0.0101-X < X-log

<

< 2.1-10-w<

0.0034

This example i l l u s t r a t e s tha-t (5.3) is weaker than (5.1). Therefire a l l found solutions of (5.3) have been checked f o r (5.1) as v e l l . The proof i s now completed by the d e t a i l s of the computations, which w e do not give here. 0

which i s contradictory. Next w e prove t h a t

The computations f o r the proof of Theorem 5 3 t o o k 3 5 sec.

Namely, suppose the contrary.

--

Then
21 0 ' p :

2X/10

i 3.1-X

, , and

it

follow^

that
5.4.

X a 80

. This

contradicts

3.1-X

>
-X/lO.

. From

(5.3) we i n f e r

Reducing the upper bounds i n the multi-dimensioarl case.

X2

-: : : :', I

<-. 20.1

log p

1 x

I t follows from (5.7) t h a t

Hence

x2/X

i s , by L e u 3.1. a comergent of

log pl/log

p2
S

Note t h a t (5.3) is of t h e , form (3.1). say

Hence by Theorem 5.1 ve can use the

Ve s h a l l do s o f o r the

' 1 ( 9 k 19

'

From t h e example a t the end of Section 5.2 w e see t h a t

Xo < 1 . 9 7 ~ 1 0

method described i n Section 3.7 f o r solving (5.3). example

t

6, 1 pl.

..., p6

2. 3. 5. 7, 1 1 , 13 1

(the f i r s t s i x

re

uJe

s m a ~ rexrnemencs ~

ox Lemmas

3. I

and 3.8, devised s p e c i a l l y f o r t h i s The remaining Remark. 34

a p p l i c a t i o n , a s follows. Let notation be a s i n Section 3.7.

solutions are listed in Table 111.

LEMMA 5.3,

Let
L(r)

X1

be a positive number such that -y2)

----

a / (4 -n2+(n-1)

-5 .
i

ord (xy) given f o r the 571 s o l u t i o n s not pi l i s t e d i n Table 111 a r e chosen such t h a t i t takes a reasonable amount of

The upper bounds f o r

computer time t o f i n d them a l l by a brute force method. The l i s t of a l l 605

Then (5.3) has no solutions with for

1.

..., s

s o l u t i o n s is too extensive t o be reproduced here.

~ O ~ [ ~ - C - / ~ / S - X ~ pi ) / ~ d - lxil ~ O ~J X d

x1

Proof
Xo
C

LPMA 5.4,

Suppose that

for B y the example a t the end of Section 5.2 w e know t h a t X < Xo 36 1.35~10 . W e apply the method described i n Section 3.7. Take 6 (which is chosen s o t h a t i t is somewhat l a r g e r than Xo ). and

V e applied the ~ ~ - a l ~ o r i t h to m the corresponding l a t t i c e cl,

rl

found a reduced b a s i s yields

.... s6

with

lcll >

9 . 4 0 x 1 0 ~ ~ SO L e u 3.4

and

Then
This is l a r g e r than
XI

2) .x0 /(4- 62+5-1

1.64.

.. x 1 0 ~ ~ s o

(5.11) holds with

Xo

. By

Lemma 5.3 w e find

Lemmas 5.3

and 5.4

a r e refinements

of

L e m m a 3.8,

i n t h a t they so

d i f f e r e n t - i a t e between the d i f f e r e n t

xi . Moreover, Lemma23.3 has slightly sharper condition and conclusion than L e m m a 3.7.

X d 1350
of

. Next

w e choose

---

1 , and
--

Xo

1350

. The ...

reduced

basis

the

corresponding -

lattice

r,

was

1 s 1 > 2 . 7 1 ~ 1 0 ~. Hence C(I',) - /1;9-1350 1.64.. .x104 Hence

2 -

> 4.;9x104

c6mputed, which
i

and
is

w e 'found

5.3 y i e l d s f o r a11

larger

1.

than

and it follows that

THEOREH 5.5,

The diophantine inequality

Next ve choose

in
(x.y)

x,

Y E S - ( 2 l -

...-

13X6

xieN0

for

i-1.

....

1012. 7

lo4 .

W e use

Lemma 5.4

as follows.

If

6 1

(A(

>

lo6

then (5.13) holds by (5. IS). and Lemma 5.4 y i e l d s

1 has exactly

605

solutions. Among those.

571 satisfy

All vectors

in the corresponding

lattice

r3

sa~isfying (5.15)

and

1 1 1 < lo6
Section 3.6. vectors, but

have been computed with the Fincke and Pohst algorithm. sf. We omit details. We found that there exist only two such they do not correspond to solutions of Next, we choose C then Lemma 5 -4 yields
Table 11, (Theorem 5.2(b)) : see p . 116-117.

solutions of (5.1) satisfy (5.16).

- . -

(5.1). 8 10 7

Hence all lo4

If

m l e 1- (Iheoreo 5.2(a)) : see p . 114-115.

- -.

I X I > 5x10'

There are 143 vectors in and with

the corresponding lattice

r4

satisfying (5.16)

IX I

I5 x 1 0 ~

. Of

them. 2 correspond to solutions of (4.1). namely those

Both also satisfy (5.17). a bound for

Hence all solutions of (5.1) satisfy (5.17).

At

this point it seems inefficient to choose appropriate parameters

C, 7 , and
0

[XI

to repeat the procedure with. But the bounds of (5.17) are

small enough to admit enumeration. Doing so, we found the result.

RelPark. Theorems 5.2 and 5.5 find applications in solving other exponential diophantine equati;as, see Stroeker and Tijdeman [I9821, Alex [ 1 9 8 ~ ~ ] ,
b [I985 1, Tijdenan and Wang [1987]. and Section 6.4 of this thesis. RemarkL The computation of the reduced basis of rl took 113 sec, where we - - -3 applied the L -algo'ritinas~describ&f-it In Section 3.5, i%-1Tsteps. The direct search for the solutions of (5.17) computations (computation of the Hence in total we used 349 sec. took 228 sec. The remaining
--

log pi up to 250 decimal digits, of the reduced basis of I'2 , and of the short vectors in r3 and r , ) took 8 sec.

5.5.

Tables.

(Throrrm 5.2(.)).
1
1

PI

XI

I
I

P:'
8 27

P2

x2

P;:
9 25 27 1 2 1 1 2 1 125 243 361 529 2197 2209 2209 6889 29929 32761 627 42241 1 11 91 3 04731 02767 1 63043 64614 03549 2 25229 22321 39041 2 21331 49190 66161 2 25229 22321 39041 2 35124 32775 37493 3 93658 88057 02081 9 97473 03260 05057 I 1 51499 04768 98413 1 1 51499 04768 98413 1 1 69414 60928 34141 12 20050 97657 05829 I2 35866 42791 61399 2 1 61 148 23132 84249 45 84850 07184 49031 58 871 58 67082 67913 174 88747 03655 13049 350 35640 37074 85209 498 31141 43181 21121 511116753300641401 498311414318121121

delta

C
C P

2 3 2 5 2 2 2 7 2 3 3 1 3 1 9 31 2 1 3 2 7 1 9 2 2 5 1 3 17 7 61

5
5 29 23

II
5 37 29 23 23

3 3 5 3 7 7 8 3 9 7 7 3 3 3 1 5 7 50 1 8 12 5 1 51 22 14 1 3 1 9 9 23 23 11 12 16 24

II

II

12 1 3 1 3 17

;1 128 I 1 1 2 8 1 256 343 51 2 2187 ' 2187 ' 2197 6859 29791 32768 627 48517 1 12589 99068 42624 I 62841 35979 10449 2 21331 49190 66161 2 25179 98136 85248 2 25179 98136 85248 2 38418 57910 15625 3 93737 63856 99289 9 90457 80329 05937 11 39889 51853 73143 1 1 69414 60928 34141 1 1 92092 89550 78125 1 I 92092 89550 78125 12 20050 97657 05829 2 1 91462 44320 20321 45 94972 98635 72161 59 60464 47753 90625 177 91762 17794 60413 353 81478 32054 69041 5 0 4 03636 19364 67383 504 03636 19364 67383 505 44702 84992 93771
1

3 5 3 11

II
5 3 1 9 23 1 3 47 47 83 173 1 8 1 89 47 149 83 1 9 83 157 89 193 1 ' 9 7 197
4!

199 43 7 1 73 53 8 8 9 163 59 163

1

2 2 3 2 2 3 5 2 2 3 2 2 2 2 2 4 9 7 8 1 2 8 7 8 7 7 7 9 11 7 10 9 9 10 9 8 10 8

0.00000 0.21534 0.48832 0.28906 0.40575 0.22754 0.46694 0.49512 0.45416 0.29941 0.40194 0.32293 0.38504 0.47828 0.18716 0.48703 0.85259 0.80898 0.88568 0.88532 0.76159 0.87942 0.76282 0.86560 0.87594 0.88743 0.89343 0.89862 0.88268 0.88656 0.84059 0.88642 0.89785 0.88568 0.89040 0.89536 0.89580

I I

a(Theorem 5.2(b)).
I

PI

XI

P;'

'

p2

s2

2 3 2 2 5 2 2 6 2 7 2

2
6 12 2 3 3 1 3 I 5 6

3 3 5 5 3 7 7 3 8 3 9 10 4 3

''

pi:

d e l t a

1'
I
I

I1
7 7 3 4 7

'

8 27 32 32 125 128 128 216 256 343 512 1024 1296 1728 2048 2187 2187 2197 50625 2 79936

3 5 3 6

II II
5 I 5 3 1 9 23 10

2 2 3 2 2 2 3 2 5 2 2 3 3 2 2 3 2 2 3 4

9 25 27 36 1 2 1 I 2 1 125 225 243 361 529 lo00 1331 1764 2025 2197 2209 2209 50653 2 79841

II

42 45 1 3 47 47 37 23

.00000 0.21534 0.48832 0.40000 0.289M 0.40575 0.22754 0.40876 0.46694 0.49512 0.45416 0.46007 0.49607 0.48070 0.41184 0.29941 0.40194 0.32293 0.30762 0.36309

'

'
t

gave a complete solution of

(

(6.1)

for the case

pl.

. .. ,

P4

2, 3, 5. 7

by elementary arguments. See also Rumsey and Posner [I9641

and Brenner and Foster [I982I . We conclude in Section 6.6 with some remarks on the Oesterle-Hasser conjecture, also known as the is related to equation (6.1). leads to a method of finding e&&les The results of this chapter have been published in de Weger [1987~].

' abc'-conjecture , which

In particular, our method of solving (6.1) that are of interest with respect to

the abc-conjecture. Finally, we give tables in Section 6.7.

6.2. 6.1. Let Introduction. S be the set of all positive integers composed of primes from a fixed

Upper bounds.

finite set ( pl. ... ps diophantine equation

. where

s 2 3

We give in this section an upper bound for the solutions of (6.1). based on lemma 2.6 (cf. Yu [1987a)). Note that in our paper de Weger [1987a] we used the result of van der Poorten (1977) instead of Yu's. We introduce a lot of notation. Assume that smallest prlme with qi 2 pi-(pi-1) for i pl < 1,

This chapter is devoted to the

in x, y, z E S

. Without loss of

. .. < ps . ..., s . Put

Let

qi

be the

generality we may assume that x, y, z

are

relatively prime. For any m(a)

a E S we define C1(2.t) and al as in leima 2.6 with n

- 1max ord (a) . Sus pi

i.e-. an effectively x, y, z of (6.1). m(x-y-z) for the solutions

-t,

It was proved by Hahler [I9331 that (6.1) has only finitely many solutions, but hi* proof is ineffective. An effective version. computable upper bound for

can be derived from the results of Coates [1969], (1970) and sprin&uk [I9691, since (6.1) can be reduced to a finite number of Thue equations. See also Chapof-Shorey -and -Tijdeman (19861

- --

----

We derive an explicit upper bound in Section 6.2. -%<tion

6.3 is devoted to

vi

-'=(l.log

pi)

- for

S-t+i,

.. .,

some details of the p-adic approximation lattices on vhich the reduction method of Sections 6.4 and 6.5 are based. In Section 6.4 we give a method of solving (6.1) in the one-dimensional case s

..

- ii

i-s-t+l

vi ,

This method is based on

the reduction procedure given in Section 3.10. As an example we find all the y w - z , where x. y, z solutions of the slightly more general equation x

are powers of 2. 3 or 5 , and w E Z , (w( S 1000000 , (w,z) 1 In Section 6.5 we give a procedure for solving (6.1) in the multi-dimensional case

s r 4

based on the reduction procedure described in Section 3.11. We

(

work out the example

pl,

..., p6

2, 3, 5. 7, 11. 13

, and actually

determine all the solutions. This generalizes the result of Alex [1976), who

Next suppose t h a t i n (6.2) the minus s i g n holds. Then w e apply Lemma 2 - 4 t o

I

N o w we s t a t e t h e main r e s u l t .
-

prove (6.4) f o r t h i s case, a s follows. Suppose (6.4) is f a l s e . Then m(x-y.z)

5 C8

The solutions of (6.1) satisfy

mEOREH 6.1, Proof,

If w e consider i n s t e a d of (6.1) t h e equivalent equation x f y - z (6.2) x-y has a t most

5

which is l e s s than

by the d e f i n i t i o n of

C4

and

C6

Hence

then we may assume t h a t say. Suppose f i r s t t h a t

prime d i v i s o r s ,

m(x-y)

ps

. Then

pi , 1

...,
Pit

O n the o t h e r hand, Lemma 2.4 y i e l d s

Thus w e obtain

Next suppose t h a t m(z)

m(x-y) 2 ps ord (z) P

and
(

m(z) 2 2

. Then

f o r some

- pi

- ordP

+E

1)

ordp(lo

El) 5( Y
- -

.
This c o n t r a d i c t s our assumption t h a t (6.4) (Yu's and t r u e i n a11 cases. Now, by (6.3) then y i e l d s EXlLmPfes - If -sC4

Put

Xi pi j . Then m(x-y) 1-1 j lemma) with n t. B O Bn B'

x/y

fl

- - - B

-W max lxi I t j

. -

m(x-y)

.W e apply Lemma 2.6 . Since m(x-y) r ps

> e2

i f f a l s e . Consequently (6.4) m(x-y) < C,

Lemma 2 . 1 y i e l d s

is

and
0

m(x.y-z) < C8 3.

.
3,

t r 2

we have

pl. p2. p j 7 ' 7 - 2 .

> r

then then

C8

< f;9al0l7
27

..

Note t h a t Cl(p,n) is maximal f o r p

If

s-6,

pl,

..., p6

I -

2, 3, 5 , 7 , 1 1 . 1 3 )
-

Cg<5.60x10

.W e

obtain 6.3.

The p-adic approxination l a t t i c e s .

i n s t e a d of (6.1). p x-y

Obviously (6.3) is a l s o t r u e i f then

m(z)

<2

. If

i n (6.2)

t h e p l u s s i g n holds.

As i n t h e proof of Theorem 6.1 w e consider (6.2)

any of t h e primes o t h e r primes a s i

pl,

1.

.... s-2

such t h a t po. put ( c f . Spction 3.11)

... ps .. . ps-2

Let

be the

.W e

may assume t h a t

. Rename

ord (lo$(po)) P

is minimal. For

B y (6.3)

and

C3

>0

it then follows t h a t

where

uiSC E

0. 1,

..., p-1
r
P for

The

ei

take the place of the

6;

of
6.4.

Section 3.11. Then it is c l e a r from Section 3.11 how t o define the p-adic approximation l a t t i c e s p E No

Reducing t h e upper bounds i n the one-dfrPensioarl case.

. Put

I n Section 3.10 w e have described how an upper bound f o r the s o l u t i o n s of (6.1) i n the case
s

-3

can be reduced. W e s h a l l apply t h a t method i n t h i s

s e c t i o n t o the following problem. Then Lemma 3.13 y i e l d s

THEOREM 6.2,

The diophantine equation

where
where p0

ordp(logp(po))

. In

Section 3.13 w e studied the s e t

(5.2.3) 291

solutions for
p u 5

for with
which is a s u b l a t t i c e of i f for p r 5 p

. -

po

Xo

- plX1 ,

xo. xl.

u E No p

, w E Z 2 , 412 solutions for p

u r 3 p
5

- . .
pU p

(p.pO.pl) (2.3.5). (3.2.5) or 6 I w l 5 10 and p w , has exactly

3 , and xo
S

570

solutions for

In Table I all solutions with

xo 14, xl i 9 15 for xo i 25. xl
S

I n Lemma 3.17 w e showed how a b a s i s of po

can be found from a b a s i s of ordp(logp(po)) basis), choose for

i

rP .

P I n practice t h i s i s very easy, e s p e c i a l l y such t h a t not only (mod p) po

is a primitive root

r*

- 3 . and

s 2 satisfy

for
5

- .
2

are given. The solutions 23. xl i 10

it happens t o be possible t o choose

Remark,

I t i s easy t o f i n d a l l solutions of (6.7) v i t h

I2

The Tables

is minimal, but a l s o

a r e presented i n Section 6.7. Boo,

max ord ( x - y - z ) The example a t the end of Section 6.2 p-2.3.5 1 w e have X < 3 . 9 8 ~ 1 0 ~ '. I t can be checked shows t h a t i n the case I w l

Then, using the notation of L e u 3.17 (with

hi

- hi
pi

- 1, .... S-2

I = po ( w d p)

. Then
pi

B y

k(Bg) (0.

. .. .1, ... ,0 ,t3:'))T

a s the l a s t element of the

, and

it follows t h a t
we

Put

have

withour d i f ~ i c u l t i ~ th sa t - t h e e f f e c t of the; w f o r the s o l u t i o n s of (6.7) a l s o X < Xo

with

JwJ s

lo6

i n the proof

of Theorem 6.1 can be neclected ( i t disappears i n the rounding o f f ) , s o t h a t

--

If

3.98~10~ holds. ~ Put

- -

po

_(mod p)

, then it-follows t h a t
--

-- -

-_

--

x/y Note t h a t 6.3, s o

6 I' P

- p2-$

,- 6-0

- logp(pl)/lo$(*--

_ L -

is a p-adic integer. Define t h e l a t t i c e s

is generated by

rp , rP

as i n Section

Lemma 3.14 (with

Ur;)

cl

0. c 2

1 ) now yields: i f
For p-2, 3 wehave

> 4(s-l)

-5

r*-I' , andfor P P

p - 5

r h s i s o f

rC

is

t h e n (6.2) has no s o l u t i o n s v i t h where

y

if

6")

- is odd.

if

b ( ~ '

i s even

. Using
cl.
g2 p

algorithm given i n Section 3.10.

Fig. 3, we can compute a b a s i s

that is reduced i n the sense t h a t

lcll

- C ( 5*)

the of as

.W e

d i d so. with

i n the following t a b l e .

B y Theorem 5.2(a), the i n e q u a l i t y with infer min(x,y)

I <

has m i n ( ~ , ~ ) ~ *no ~ solutions min(x.y)

5

> W ,

since

> lo4'

Hence

, and we

The values of

6")

can be found i n Table 111. Making an exception t o our

policy, w e give the reduced bases of the

r*
P

below.

I f i n (6.7) the plussign holds. then t h i s inequality follows a t once. So now the bounds given i n the above table f o r IyOI, lyll follow from

W e repeat the procedure with

a s i n the following t a b l e .

The numbers a r e now so small t h a t the computations can be performed by hand. For example, f o r p

- 5 . the l a t t i c e

F ;

is generated by

and a reduced b a s i s is

. -

Cl

-[

205

'y.

G-z
u

[ Yg4 ] 408 and

.-

W e f i n d upper bounds f o r three cases.

42

w ~ < ~ lo1' ' ~. On

supposing

a s given i n the above t a b l e . we i n f e r min(x.y) > 1 0 '

In a l l

From t h i s w e found the lower bounds f o r l a r g e r than /2-3.98x10~~ . Hence u

5 p

1 sI
1

given above. They a r e a l l

i n f e r from (6.6) t h a t

(6.5) h o l d s f o r X1 Xo , and then we 1 , and Iwl-z a V a s shown i n the

t a b l e above. V e now f i n d the new upper bounds f o r IyOI. lyll a s follows. I f i n (6.7) the minus s i g n holds, then. on supposing t h a t min(x,y) > V 10/9 , we infer

I x y I < m i n ( ~ . ~ ) has ~ . ~ By Theorem S.2(a) w e see t h a t the inequality (265 28 (281 , 3 53 ) . However. both have only two s o l u t i o m : (x,y) *s 1. '-0.9 I x y I - > lo1' So w e infer min(x,y) 5 lo1' , hence by J ) , a lo1' + V we obtain the bounds f o r I yo I 1 yl 1 a s given above. ~ ( x

These bounds a r e small enough t o admit enumereation of the remaining cases. 0

Renark. The computer c a l c u l a t i o n s f o r the above proof took l e s s than 1 sec.

6.5.

Reducing t h e upper bounds i n the multi-dimension81 case.

t(rP) I n Section 3.11 w e have described how an upper bound f o r the solutions of (6.1) i n the case
s L 3

Is11/4 >

ord ( x - y - z ) P 606 385 275 220 165 165

can be reduced. W e s h a l l apply t h a t method i n t h i s

35 4.70~10 1 . 1 5 ~ 136 0 6.27~10~~ 3 . 1 7 ~ 1 36 0 33 5.74~10

s e c t i o n t o the following problem.

THEOREM 6.3,

The diophantine equation

(6.8)
(

x + y - z

1 . 7 3 ~ 1 36 0 These lower bounds f o r c(P*)

Ir

in

x. y, Z E S -

(x,y)

2X1-....13

X6

x i ~ H o for

i-1.

....
514
S

6 1

with

a r e a l l l a r g e r than

/ 5 - 5 . 6 0 ~ 1 0 ~(note. ~ that

1 and

has exactly

545

solutions. Of them.
S

satisfy

w e have a very large margin here, w e could have taken the

ord2(x-y-z) 5 12 ord7(x-y-z)
5

ord3( x - y - z ) ordll(x-y-z)

7 ,

ord5(x.y-z)

.
.

n .3.14 f o r about 20% smaller). So w e apply L

every

XI

4 ,

3 , ord13(x-y-z) 5 4

e thus f i n d ord (z) S p + p0 Since equation (6.2) i s i n v a r i a n t p w P under permutations of x , y, z , w e even have ord ( x - y - 2 ) S p + po , a s P shown i n the above t a b l e . Hence m(x-y-z) S 606 W e repeated the procedure with Xo

- Xo

~ ' s probably
5 . 6 0 ~ 1 0. ~ For ~

The remaining
Remark, (6.8). Proof

31

solutions are given in Table 11.

606

and

a s i n the following t a b l e .

After computing the reduced bases of the s i x l a t t i c e s following data. Note t h a t i n a l l cases t(r*) r /5-606 P

From Theorem 6.3 it i s not much e f f o r t t o f i n d a l l 545 solutions of

r* P

we found the

< Xo

- 5 . 6 0 ~ 1 0.~With ~ the
-

I n the example a t the end of Section 6.2 w e have seen t h a t

m(x-y-z)

notation of Section 6.3 we choose the following

parameters;

5 6
. .

0 0

-2

1
2
-

1443

18 18
-

3196
--

--

Hence

m(x-y-z)

s 67

. Next,
v2

we repeated the procedure with

Xo

- .
67

and

e found a s - fn the following t a b l e . W

yo

y1

v3

y4

C(S) >

ordP ( x - y - z ) d

W e computed t h e s i x values of the i n Table 111).

3

i9p)

for

- 1. 2.

3. 4

(and g i r . them

and the reduced bases of the s i x l a t t i c e s

r*
P

by the

~ ~ - r l ~ o r i t h Thus m . w obtained:

Hence

m(x-y-z)

56

To f i n d the solutions of (6.2) with

ord ( x - y - z ) below the bounds given i n P the above t a b l e , w e followed the following procedure. Suppose t h a t w e are a t f(p)
z

a c e r t a i n moment i n t e r e s t e d i n finding the solutions with where x,

y,

is given f o r

and consider the l a t t i c e

2.

ord ( x - y - z ) I f ( p ) P p and < f(p) p0 f o r these values of p, p I f a solution

...

13

Choose

; .

x1.

...

of

(6.2)

e x i s t s with with xi

x4. xO

IT

ord

ord7z) P (x/y)

r p + po
for
i

, then the vector

0.

. ..

4 , is i n the

l a t t i c e . Its length is bounded by

Pi 2+. ..+f(p4) 2) . / [f(po)

A l l vectors i n

r*
I '

with length below t h i s bound can be computed by the algorithm of Fincke and Pohst, a s given i n Section 3.6. Then a l l solutions of (6.2) corresponding t o l a t t i c e points can be selected. Then w e replace w e r e p e a t the procedure f o r newly chosen i n t h e above t a b l e f o r page. Here,
#

f(p)

by

p0

p,

and

W e performed t h i s procedure, s t a r t i n g with the bounds f o r f(p)

and with

p, m

ord ( x - y - z ) given P as i n the t a b l e on the next (5.2) p for

stands f o r the number of solutions of f (2)

e have stage. A t the end w

-4

, f (p)

-1

found a t t h a t

3,

.. . ,

13

The
0

remaining solutions can be found by hand, Remark,

(1976)).

Theorems 6.2 and 6.3 have applications i n group theory ( c f . Alex W e use Theorem 6.3 i n Section 7.2. The computer calculations f o r the proof of Theorem 6.3 took 438

Remark.

sec., o f which 412 were used f o r the f i r s t reduction s t e p . In t h i s f i r s t s t e p w e a p p l i e d the L3-algorithm in 1 1 s t e p s ( c f . Section 3.5). which c o s t on average about 60 see. per l a t t i c e . The remaining 50 sec. were mainly used f o r the computation of the 24 6") i *s

.
absolute constant
C

6.6.
Let

Examples r e l a t e d t o t h e abc-conjecture.
x, y. z

be p o s i t i v e integers. Put

such t h a t
c

c(x,y.z) all

<

for

all

conjectured the stronger a s s e r t i o n t h a t some bound depending on For a l l x, y, z c(x,y,z) Recently, with (x,y) only. f o r

c(x.y.z) < 1 + r

. when

x. y. z z

ndsser exceeds

r > 0 . For a survey of reLated

1 and

r e s u l t s and conjectures. see Stewart and Tijdeman [I9861 and Vojta (19871. z w e define I t might be i n t e r e s t i n g t o have some empirical r e s u l t s on search f o r
X,

log z / log G

.
t o decide whether there exists an

c(x.y,z)

a d to

y, z

f o r which it is large. From the preceding sectioms it x. y , z correspond t o r e l a t i v e l y s h o r t v e c t o r s i n

Oesterle posed the problem

may be c l e a r t h a t such

appropriate p-adic approximation l a t t i c e s .

As a byproduct of the w o o f s of Theorems 5.5 and 6.3 we computed the value of

c(x.y,z) c(x.y,z) for

, corresponding to many short vectors that we came across in

2 1.4

performing the algorithm of Fincke and Pohst. All examples that we found with are listed below. Our search was rather unsystematic, so we that occured is 1.626 , which was reached by
fable I
(Theorem 6 - 2 . )

do not guarantee that this list is complete in any sense. The largest value c(x,y.z)

p=2.p,,=3,p,=5

This example was found on September 20. 1985. and has not yet been beaten. to the author's knowledge.
X

c(x.y,z)

These results & not seem to yield any heuristical evidence for the truth or
- -. .

falsity of the abc-conjecture

6.7.

Tables.

(cont.)

Tabl+ (cont.)

x0

P2

"I

p;l

sign

W'

OOOCOO 040040 O d O O 4 4 040044 r t 4 0 0 0 0

(VNN NNN C J 4 0

4dO OON

NdO MOO MNO

044 O d d

444

44N
404

0 U Y 4 )

NNO
000

-. -04NN
ONO OAN -ON
044

uu m u NO
ou
4-

U d

uu m 4 . 4 y
mu 4m UO ON 4u UO m o 40 do
uu mN cum mu umo NNU moo
UNr)
N-4 NU
dr)

um

ON u4
OY)

Nv)

mul

ulul NN

N.4 NU

N4

NU

ddON ONNN 0604 d0ON NNOO NONO NONN

O O N 4

O C J H N

OONN 4004 NAN0

N.444

NdNO r.NO-4 NNON 00,-0 OONO

O O N 4
- 4 0 - 4

NNNO
0400 0400

40

m u m.4 m 4 NO
40

N4 0-

mu Mw w4 N w u mY) do N w O 00 N

om Nm mo w.4 dm m(Y wN om *7 tn m u
00 NU

w om w mu dm
04
04

AN m o

aro

u i n

ue

u4

OOON NNd4 rrONO

uo '
440

NNU

040d

NO40 OCOO

NNON N4ON
-0tN-4

Nor)

ONN

n m 4 OmM UUO NCO

W O N
-40

NUN

omv u m

4NU

ncu - 0 .

om u m m o

--

NUN

UUO Dm4 mm.4

000 000

,-u 0 . 3 MY) w*7 w u w In U Y 4 ) VO UO roo

O N

. \ -

rom N.4 wm ulN 0.4 ou

u Inl . 4

dU

roo we wul ON md

. N
N

I n n mu)
mm N d ou

WYJ

NN

om d u
O U

\ O N
40

N400

N4tJN

eNdN O W N 4

d0N 04N

N
-7

Y)

.-cw U mul O
UY) Nul

u4 Nm

0-

OOON CJAN4 -400 ONNN OOMO

dNNd

OdON O4NN NOPI0 OdCN -.ON0

2 A - 0

4000

04m mcuu OWN uuu

N44

40u Nm4 UUO

40-

NNNO dOOM NdON ddON NONN

0 h ' N . a
44Nd

ONNd

ONdN

dNd

ONON NNON Odm uou 04u

O W N
OUN

Nnm m o d 0 m udm ou4

. . v)d

umm NON N N W mmo omu

u-40

VIV v)N

wo

mN rod
uo

NN

ON dm d No N mul
N Y )

mmr) OOd

-7 uu AN4

mhn ON
AIn NN
In*

ON N *-J mN ar) WY) mul m u u l u *u w 4 Nm m u

4-

M dm U
UN
Oul
Y ) u

N-0

u w u o mul NU w u u u l VN -3 v roo N wO e
Wul

CdON -NNO OOCJO

.4ANd

dNdN
Nd4.4

NNdN dNOd 4NNN NdWO OONN dO4N

Nodo
NNNN

mmN VrtO mom

NNd
UNN d d U

OACN ON00 WONN

NNNO
-00-

OONN OONO

. -7u.4

UNN
0-0
N O 4

m7m4 UNO -u4 dm*

-7 -10 d
CJ I n

ON

m.0

N *N O om mu Vul MO
N4

NN

N . ? N

OMm

004

ONU O d d
. r
N

ON0

NdNN
NN-N

. miw
N

u4 uu OY)
-3 4

Nr) -7 N

mr) mul uo mm u m
In0

eul r7 In

NOCJN
dNNN
N4-4 C U - 4 0

OdON OONN NO4.4

A400

OOON

OANN
C J c ( N 4

OOdN NNON ONNN

NdNO

UNO NmN
duo
+Or)
ddv)

N O 4

w-4
Ud

em mr) wNH

InN NY)

Oul UO N.3 UIN

0 .r)

m-n NO \fu

coo NNm
-04

u u ) U* 3

&.NN @Om
oou
+N4

NUN
dUY)

-.-,sod NCNN -NNN ON04

44dN
h'N00

OddN

OONN OdON
dONN
NO00 40-0

em4 Odhl -mu ON4 4u4

040

ous
ANN

-d mo N

O N

mu) mm UN uu OW

Nu044

roul mu mr)
40

Inmmw u.4 m u u m o m m.m W 4 uu

UN
"-30

B0-

m u row

04NO

OOON

OOON 0040 dOOd

h.0h.N

O d d 0 ONNN -N-O
-00.4 04-4

OUN uum

Y)wm
00N4-4

NUY)
% - d

ul4

wul MN
OY)

4000

4-4

Y)UN

OY) w u '

OONO d4lWO
4N4d

CNN4 OOON 4OONOOdN NOON dOON 4OclN

00-4

NNN4 NdNd

ONNN

C;ONd CUOOO ONN4 NONN (VNOO NCNO NNNO ON4N

ON04

N4NN

NNU 4uu ON0 -00 UOU OOA NmN 4NU w-4 (vmw N#ON NNNO
4NU
444 -40

Y ) 4 u Y ) u 4 ONN
UUY)
-44

UY) y -7

'Y)

.
'

4 0 .

L ? C

UCJul

004

OVm
UInul

404

4dU

N4U

NUU

UNO A d 4 m o d vom

-7oul h * 0N Inulul

am0

\tdN

-00

u uL9 ro-4 WOO

U W d NNU
4dU

W U N NNN

d4N4 ddN4

NNNN

u u m
r)dN
0Y)Y)

00-

mow d m 4

n0

4-3N

OON

4uul

en-3 *am NO*

ul-lr,

0-u UmN
N-iW

uor)

NUN

ON0 NON ouu NOOd NdNO

NU0

OUN
004

ONNN

Y ) u u

mmN

UON udm omm mum

004

ON4 w4r) NmN Y)4N ouo U N . 4 u4r)

d 0 Y )

d4U

Nul4 I n u l u

4wm

UmN

d u l c )

NU0 o m 4 OOYI ul-u nulw

0Y)Y)

m4.a OmN uY)n 004 NCU

0

Inou
(Vw)Y)

NN4
W U N
0

UIUIm 0-m Y)WN &nu NUN

0

137

m
solution

computable constant x, y, z max (x.lyl.z) < c PIEOREM 7.2, Let

pl, - . - . ps be given. There exists an effectively pl, . . ., ps only, such that any C , depending on of equation (7.1) with conditions (7.2) satisfies

where y yields

ylVy2 .:and

y1 E S ,

ky2 E S , and

y1 > ly21

Subtraction

.
(

pl.

....

pS

where now all variables

(

2, 3. 5. 7

Equation (7.1) with

in or

By

(u,yl) ( a )

conditions (7.2) has exactly the Remarks,

388

solutions given in Table I.

2-a + 2-b

- 2 - c , where

hence u, yl. y2 (apart from the sign) are in S c , (u.y2) 1 , equation (7.5) is of the form a + b

only. and 1. The Tables are given in Section 7.9. We stress that the aim of this chapter is not only to prove these theorems, but to show as well that pl. .... ps ) a result similar to Theorem 7.2 can be proved along the same lines, in a more-or-less algorithmic way. for any given set of primes
(

a. b , c

are composed of primes

2 . pl.

....

ps we

1 ,

such an equation

a + b

-c

a r b > 0

In Chapter 6 it was shown how to solve

(

. For
)

our

pl.

have the following result . LEE(HA 7.3, with Proof. (a.b) Let

..-. ps
)

2. 3 . 5 . 7 1

I pl. D

2. Equation (7.1) with conditions (7.2) can be seen as a further generalization of the generalized Ramanujan-Nagell equation

conditions (7.2) and

-1.

.. .
1

. ps -

2 , 3. 5, 7

Equation (7.1) with

has exactly the

95

solutions given in Table I

(cf. Chapter 4). namely by taking

Ikl

arbitrary instead of

k E Z

-1 .

From Theorem 6.3 it follows that a

with

a, b, c E S ,

r b has exactly 63 solutions, that are easy to compute. Each

of these gives rise to three possibilities for (7.5): if if 2 1 a 2 1 b 2 1 c then then then (u.yl.y2) (u.yl,y2) (u,yl.y2)

fixed. The method of this chapter to solve (7.1) is also a generalization of the method of Chapter 4 to solve (7.3). Equation (7.1) can be transformed into a number of Pell-like equations. Put if

(ta.b.c).

(b.2c.2a).

(c.2a.-2b). (~.2a.-2b). (+.a,-b)

(a.2b.2~). (3.c.a). (a.2b.2~). (b.2c.2a).

where for- D
-

D. u E S

Of the thus found 189 possibilities. the 95 ones given in Table I with and D is squarefree. There are only
-

2S
-

possibilities

satisfy x L y

and

z > 0

. whereas the others don't.

-1
0

. Now.

(7.1) is equivalent to a finite number of equations

in

z2

2 D-u
+ -y

Th=ks-completes our t r e a t m ~ i i i t 3 i ~ c a S De (7-4)

--

u E S ,

E S

with z

z > 0 and u

and

(u,y)

equation (7.4) by factorizing its both sides in the field dealing with equation (7 -4) we allow

We treat

Q(/D)

. When

7.3.

T w a r d s generalized recurrences. D > 1 . Put K O(/D) . Let o : K + K be the K with o(/D) --/D -.-- For any number or ideal X in K we for u(X) , for convenience.

to be negative.

From now on, let

7.2.

The case

automorphism of 1

.
D

write

X'
pi

First we consider the special case

. Then

(7.4) is equivalent to

for

1,

....

be the prime ideal in

such that

Z + u z - u -

"1 y2

ord (pi) > 0 . If pi splits in OK , this is well defined if a choice has pi been made from the two possibilities for /D (mod pi) . Put for a solution

(ai-bil

max(ai,bi)

if pi

and

b , j

1 ). Hence

is a principal

ideal, say ord (y) pi

max(ai.bi) + 1

min(ai.bi)

-1
0

if

Pi

for an a E OK

. Up

to multXplication by a unit, there are only finitely many

a

- min(ai,bi) if - 2 occurs, and bo - otherwise. (Note that min(ai,bi) - 1 may occur only if p + pi , hence only if pi - 2 splits). Let us assume that the splitting primes of pl, . . . . Ps are P1. - . - ' Pt
Put bo
pi

possibilities for

. Let

be the fundamental unit of

with

> 1

Now, (7.7) leads to the system of equations

for some

s t s s . Put

For

where 1.

Put for

Z ,

...,

t , let

hi be the smallest positive integer such that

hi pi

ml-

, mt
a

No

, and for each possible a

is a principal ideal. say

If

denotes the class number of

. then

hi I h . Now, ri

determined up to multiplication by a unit. Thus we may choose

ri E K is such that Then (7.8) is equivalent to

For

1.

...,

. put . and
0 5 di I hi

The f e t i o n s sequences.

Go

and

Ha

are generalized recurrences in the sense that if recurrence

all w i a b l e s but one -are-f ixed ,-then they become -integral bin-

with

ci, di E No

- .- -

. Consider

the ideal
7.4.

Towards linear forms in logarithms.

pl, ..., ps From the above considerations it follows that, for given K there are only finitely many possibilities for a . By (7.7) it follows that

Let us write ui for

ord (u) pi for each

1,

.... s . Put

(namely,

lai-bil

max(ai.bi)

if

pi + 2

since then min(a i .bi)

0 ; and

for at least one (n.ml

....,m t)

E Z x: N

Note t h a t s i n c e

(u.y)

1 the s e t s

are disjunct. W e proceed

with t h e f i r s t equation of system (7.9). Written out i n f u l l d e t a i l it reads

From equation (7.10) w e now derive Now,

I . 1'. IU depend on

pi-adic
i E I. I'

l i n e a r forms i n logarithms, i n or

three d i f f e r e n t ways, according t o which depends on the p a r t i c u l a r s o l u t i o n of

a

IU

. Put
if then p p 5 .

equation (7.4) t h a t w e presupposed. However, w e know t h a t f i n i t e s e t , which can be computed e x p l i c i t l y . completely f o r each hence of (7.1). The s e t of the If
D

belongs t o a Then yi

Ti-+

if

pi-2.

yi-1

if

pi-3.
y1

Ti-;

So i f w e can solve (7.10)

of t h i s s e t , then w e can find a l l solutions of (7.9).

> l/(pi-1)

, hence i f

ord (<) pi ord (Q) pi

z
.

for a

E K

ord (logp ( I f ( ) ) pi i a's may be reduced, vithout l o s s of g e n e r a l i t y , a s follows. bo a a.

1 (mod 8) then

- 0.

may both occur, with 2-aO u

respectively. U e only have t o consider a s o l u t i o n of (7.9) f o r (7.9) f o r

a

2-aO is

. Hence

- .
being a

then

. because i f u - uo. z - zO i s - 2-uO. z - 2-zO a s o l u t i o n of

is

- ao,

2-a0

W e now have the following r e s u l t .

( i I U I

Let
(7.10). (i).

n, ci

ui

( i E IU )

be a solution of

it i s not necessary t o consider

aO i f a l s o

2-a0

already

ord (a ) 0 also a 2-aO 2 0 may occur, s o t h a t w e only have t o consider the l a t t e r . Note t h a t i t may now occur t h a t
IU and

D = 5 (mod 8)

then with

considered. such t h a t (u,y)

B y

the

aO

same

argument,

if

For

i E IU put
A,

(u,y)
I u I'

- 2 . The condition
(a0)

- ordp i ( 2 ( ~ / a ' ) .

1 is used only t o ensure t h a t

e only have t o a O , then w
--

(u,y)

-2

a r e d i s j u n c t . This remains true i n the above cases with

'.
'

Further. i f a
I . I'

(a6)

f o r some

consider one denote the to

a'

of the p a i r belonging t o

a o , ai)

Namely, by

- -

c-c#-

- fl

w e have (we belonging

aO by

IO. Ii) , then the

I, I'

are

16,

IO )

i+AiLyi

then

--

(ii).

For

put

ord .Pi a

and analogously

If

hivei

' c ~ L yi

then

(ii*).

For

i E I * put

Ai-h

*- A i ,

- h

*- n + n O , *

3 + nj

Then it follows t h a t

Ki

a* i

logpi (55) a

n.lOgpi(e)

jEIU

uj - l o gP i (pj

Further, note t h a t the prime divisors of

If

a r e j u s t the ramifying primes.

h -c i i

So, by (7.12).

r yi

then

where Remark.
IU, I
a

vi

Note t h a t a l l the above pi-adic p and

logarithms a r e well-defined,

since splits,

- --h * Sordp i (4D)

1

for pi

t h e i r arguments have

I'

adic order zero. This follows from the f a c t t h a t ia r e disjunct. and i f D 1 (mod 8) from the choice

vo

otherwise. I f

- t+l.

. ..

s , and

vo

1 bo

if

splits w e have assumed t h a t

-1.

Hence the l a s t f a c t o r vanishes. So put

2-a 0

Proof,

For ( i ) , divide (7.10) by i t s second term. For ( i i ) . divide (7.10) by i n a l l three cases, take the pi-adic order. and apply Then i t follows t h a t

i t s second term. and add 1. For ( i i * ) . divide (7.10) by i t s f i r s t term, and

s u b t r a c t 1. Then. (7.11).

The l i n e a r forms i n logarithms Ai,

K i m K;

.
-

as they appear i n L e m m a 7 .S.

seem t o be inhomogeneous. since the f i r s t term has c o e f f i c i e n t 1. However. it can be made homogeneous by incorporating t h i s f i r s t term i n the other ones, as_foll~ws. Put
.

--

--

Note t h a t , by the d e f i n i t i o n of

a , This leads t o the following reformulation of Lemma 7.5. 7.6& Let

n, c i
aim a ;

for

1E I

u I*

where the exponents

ni

for

0 s i s s

a r e i n t e g r a l . I t follows t h a t

ui

for

i E

(7.10). l e t Ai,
ui* (i).

* I,,

be as i n L e m m a 7.5. and l e t
ui +-.Xi L yi

h , At.

be a solution o f
Kt,

K ; , ni, c i ,

be as above.
i E

Let

IU
A
i

. If
+

&en

ui

Put

ord (h ) pi

ord (Ai) pi

* .

(ii).

Let

i E I

. If
K

h -c. i 1

+ ii 2 yi

then

h -c i i (ii').

ord (h*) Pi h -c i i

- ordP i (K:)
+
s ;

.
then
)

and f o r

Ki,

* K i*

we f i n d , i n view of Lemma 7 . 6 ( i i ) , ( i i ' ) ,

Let

i E I'

.
i

If

r yi

h -c i i

K'

ord (h ) Pi

- ordP i (K;*

.
hi.

Here, that for

C1.

C2.

C3,

C ,

a r e constants t h a t can be w r i t t e n down e x p l i c i t l y . I n

B

order t o f i n d an upper bound f o r Remark. parameter each

a

we t r y t o f i n d constants

ClO, Cll

such

W e w i l l study the l i n e a r forms i n logarithms n , ci,

D

a r b i t r a r y i n t e g r a l values of the v a r i a b l e s
a

ui

Ki,

K;

. Notice t h a t the
I n view of (7.13) we may i n s e r t and d e l e t e a s t e r i s k s any time w e like, as In[ by a constant times log B

has disappeared completely from these l i n e a r forms. This means only, instead of f o r

t h a t we have t o consider the l i n e a r forms f o r each

long a s we don't specify the constants. I n order t o prove (7.16) it remains. i n view of (7.14) and (7.15). t o bound w i l l introduce c e r t a i n constants
C5.

.W e

C6, C7 , and d i s t i n g u i s h t h r e e cases:

7.5.

Upper bounds f o r the solutions: o u t l i n e .

(a). (b). (c).

-(C6+C7-H)5nsC5, n > C5 n <

Iat us f i r s t give a global explanation of our application of the theory of

- ( C6
C8, C9

(7.17)

C7-H )

.
Ga
dominates. W e s h a l l show t h a t there e x i s t

p-adic

l i n e a r forms i n logarithms, t h a t gives e x p l i c i t upper bounds f o r the

I n case ( a ) it i s , by (7.15). obvious t h a t (7.16) holds. I n cases (b) and ( c ) one of the two t e r n s of constants such t h a t

v a r i a b l e s occurring i n the l i n e a r forms

. Then we give arguments hi, Ki, K; w h y we choose t h i s way t o apply the theory, and not other possible ways. I n
t h e next s e c t i o n w e give f u l l d e t a i l s of the derivation of the upper bounds. I n t h e sequel, by the 'constants' C1,

...,

C12

w e mean numbers t h a t depend n , c i , ui

only on the parameters of (7.10). not on the unknowns

Then (7.16) follows from (7.14). F r o r (7.16) we derive immediately an x x p l i c i t upper bound hen= f o r a l l the v a ~ i a b l e s ~ i n v o l ~ Since d. the c k t a n t s also
C12
.

C12

C 1 . ,

..

for
C4

will

be very l a r g e .

w i l l be very large.

proceed by reducing t h i s upper bound, by applying the computational p-adic diophantine approximation technique described i n Section 3-11. t o t h e p-odic l i n e a r f o r a s i n logarithms t h a t t h e constants C5, hi, Ki.
C9

To f i n d a l l s o l u t i o n s we

* *

Ki

* . Crucial

i n t h a t l i n e of argument is Cl.

. .. ,

a r e very small compared t o

....

C 4

This method leads t o reduced bounds f o r the p-adic

orders of

the l i n e a r

Then i t follows t h a t

forms. Then w e can replace (7.14) and (7.15) by much sharper i n e q u a l i t i e s , and repeat t h e above argument, t o f i n d a much sharper i n e q u a l i t y f o r (7.16). I n general we expect t h a t it i s i n t h i s way possible t o reduce i n one s t e p

for

H,

U. B
A :

the upper bound W e apply

Le~mra

C12

for

2.6

t o the p-adic

linear

forms i n

t o a reduced bound of s i z e

log C

12

logarithms. For

we f i n d , i n view o f L e u 7 . 6 ( i ) .

Before going i n t o d e t a i l w e explain b r i e f l y t h a t i t is possible t o t r e a t

(7.10) partly by the theory of real (instead of p-adic) logarithms, and subsequently Ki and by a real approximation technique (cf. Section 3.7). First, note that

linear forms in diophantine

computational

and why we prefer not to do so.

Ki

have generically more terms than hi , and are

occur only in case (a), Ki, K; this is the most difficult case. Equation (7.10) consist of three terms, each therefore more complicated to handle. Since of which is purely exponential, i .e. the bases are fixed and the exponents are variable. If one of these three terms is essentially smaller than the other two (more specifically, smaller than the other terms raised to the power 6

. for a fixed

6 E (0.1) ) . then we can apply the real method. There

are two ways of doing this. Write (7.10) as

In1 cannot be very large, and we are essentially -(i.e. apart from a finite domain) in case (a). Unfortunately. the region for In1 that w can cover in this way becomes smaller as (see the example below). Second, suppose that

First, suppose that

Ix-X'

I < lx*1 6 . Then

1x1 > lx*

. or

1x1 <
*

1x' 1 6 . Then we are essentially in case (b) or (c). But this area can be dealt with easier p-adically, since here we use the linear forms
whereas the real linear forms in logarithms used in this case will The really hard part is C. It can be reduced to and Ix*( + IX'16 < x < c - ( x ' ~ for any c > 1 generically have more terms. The areas sketched above, in which we can apply the real theory. will not cover the whole domain corresponding to case (a) (cf . the white regions in Fig. 4- bel&) the p-adic linear forms Ki. K; use of real linear forms. --

~ence -we cannot avoid working 'with

1 z.lx'l < 64s (0.1)

. b u t w i l l never

<

Ix'l

Ix'l

But then it is more convenient to avoid the

disappear. So we cannot avoid the p-adic linear form in case (a), vhich then works in regions C and D together.

Let us illustrate the above reasoning with an example. Let

t--

1 + f2

(1+f2)n-(1+2-/2)x Fig. 4 below gives in the (n.X)-plane 2 the curves 2IX* * which are boundaries of the four regions A. C, We have the following y possibilities.

Then we have

- - , . - + . - . . - . I* - . -i . . x - X' . 1 x ' I . lxaI4Ix* I. Ix* 1. I-/Ix* I. f-lx*I ~ I X 1' . B. D .

--

--

s1

2./2

I - - (11

pl

and

7.6.

Upper bounds for the solutions: details.

-

--

We now proceed with filling in the details of the procedure outlined in the previous section. We apply Yu's lemma (Lemma 2 . 6 ) - - r o - -follows We have ai we have r / c ' . r /r'

number of terms in linear form region p-adic method real method

-2.

- K - a(&) .

so

For the

, or

t.

r*. pj. rj.

s j . Wehave

to compute the heights of these numbers. We have at once

1.
h

...,

a prime i d e a l in
; i

l y i n g above

pi

Let

hi

be a d i v i s o r ui

of

such t h a t all pi

i s p r i n c i p a l , and denote a generator by

Let

e i t h e r : (1)

s p l i t , and then fj

to
or: ( 2 )

ca

'

-3
"j

for

1.

...,

h(>)

- log(.) .
ai
1,

h
j

-1 0 ( l j l j 1 ]
ord (b ) is P " a r e the unknowns, w e should
Vi

Let

be an odd prime, n o t d i v i d i n g

. Then

The order of the

Vi

for

.. .
I?;

i s important i n two respects: it is required t h a t the

n-1

a r e i n increasing order, and t h a t

minimal among the assume t h a t product

Vl-

...Vn

ord (b ) . Since the bi P i V1 s . 5 Vn-l . I n the f i n a l bound however. only the

ErooL

Let

KO
i

..
v ,:

- K((:'~)

, and
qi*l

Ki

Ki-1(F:t21

induction on Suppose t h a t

t o prove t h a t

and

appear. So the ordering of the follows w e can take

only

[Ki:K]

matters f o r defining

. It

[Ks:K]

it s u f f i c e s t o prove t h a t

. I t remains el+, t Z Ki
K-vector

for

1.

. .. . s . W e
[KO:K]

.
q

Note t h a t [Ki+l:Ki]
is prime.

t o prove t h a t

- q . hence
, with as

use q

, since

Suppose the

contrary is t r u e .

Ki

is a

space of dimension

qi+l

b a s i s a l l the elements with t h e

ai

i n any order. i f w e define

-

- -

for Further, we take

--

k + ~ ( O , l ,
J .

...,

q-1 )

for

0,

...

,
I t follows t h a t there

exist

.ki

E K

such t h a t

B Then

- Bo

BnP-7

- max-{dbll,
. By

....
B 2 2

4 $Id lbnl.-3. ~ - n . ( p -11-1

<B

3 log(1~-Br ) f - ( l o g p)/d

i t follows t h a t

3 1 + z-B

The group of
j

Hence w e can take

w
There -are

log B

.
conditions to be checked. The first one

K-embeddings of defined by

Ki

into

i s generated by the

0.

..., i

'3

..
not a l l

-2
bi

two

more

is

that

r 1

. This

a r e zero. The second one is

i s immediate. i f r e assume the obvious condition t h a t n [ K ( C X : ~. . .a;Iq) :K] q which

..

- .
h
pi

where

i s a primitive

q th

root of unity. Hence a l l the embeddings a r e

is l e s s obvious. For our s i t u a t i o n it follows from the following lemma.

given by

7.7
number. tat

Let pl,

.... p,

- Q(/D) .

with

a s fundamental uni t . and

as class

be d l s t l n c t p r l w n d e r s , and let

be f o r

for C E I 0, 1. j

.... q-1

. It

follows that

.
and in case (2) to

The minimal polynomial of conjugates of j

over j
(

is

( : $ are

p j ( : $

for mj

- 0. 1.
O*

xq .... q-1

.
)

Hence the

(where that

p ' * '

(to)

, all with equal such that for

stands for q

or

p* ) for some fractional ideal mj-hj for j

(note

(1) ) . Because of unique factorization for ideals it follows in divides all q h

both cases that

multiplicity. There exist numbers

- 0, 1, .... q-1

we have

.--*

q-1

- 1.

...,

and

hi+l

This
0

contradicts the assumption bl o r ( a

.
1 > 1 - 1 then

Remarks, Hence i

1. If

- . ..a n-

1 ord (a P 1

.. . .-o "-1) n

ord (b -10 (al)+. . .+bn-log (a ) ) . P I $ P "

We prefer to work with the logarithmic version. since that is the one we use in the computational method of reducing the upper bounds. Now apply 2. In order to apply Yu's that does not divide lemma we can take for q the smallest odd prime

cpC0' ' ' . .ti to (7.19). Then for each tuple (to,.. . ,ti) we find

h-p-(pfp-1)

.
C1 to C12

We now proceed to comp-iite the constants we apply Lemma 2.6 to constants Here we have a system of q i+l linear equations in the q i+l unknowns i+l (where ti Cl.i* C2.i

- Ai *

To find

C1

and

, for all

, , i E I

. Then

=2

we find for each such

such that. under the conditions

% ( , .

- .- .ti .' ThC&CeiSi~nt'of-this

Ki over

system is exactlj-the-square root of the

discriminant of

K , hence nonzero. Consequently-there Is in cq

denotes the number of terns in ord (Ai) < C l e i + C2,i-log B Pi By L e ~ n a7.6(1) conditions and the relation ord P

just one solution of the system. But we know that solution:

hi ) , we obtain

*.

e -ord P P

we see that. assuming the

The latter equation now yields an equation over

K :

max (li-Xi)

Iu
it suffices to take In case (1) this leads to the i&al equation

.- *

4 PiB Z max [ 2. ~ - t ~ - ( p -1) ~ iIu

c1

max ( -(Ai+ord (h 1) + ClSi/ep iIu pi i

. C2 for a11

iIu

( C2,i/epi

which expresses that the first tern of Ga dominates. Put

Then (7.14) holds. Next we apply L e u 2.6 to respectively, to obtain and X* if i

E

K ;

and C4

Kif

C3

and

. By

x(' '

I if

and i
and
E

I* I

we denote constants

X C3.1

I*

There exist by

L e ~ a a a 2.6

Then we infer

C4,i

such that under the conditions

(where again

ti denotes the number of terns of

K:

'

'* ) . it follows that

hence n <

ord ( K : ' '*) < C3, Pi Again, by b m m a 7.6(ii),(ii*)

C4,

- log B* .

l o g ( )4 4 ~ + U -log(P) )/log
1 0 1

it follows that, under the conditions Next suppose that

n < min ( -(C6+C7-kt), 0 ) it suffices to take Then we find that the second term of

.
Ga dominates, namely

c ,

.:'.'+or@
. a x ( iIUI '

(h 1 . pi

+-

3.1

) , C4-

hi

hi.ePi

C 'Pax iIUI* i Pi

~" -~

Then (7.15) holds.

-

We take

C5

to

C ,

as follows:

c7
Note that C , take C8

-(
C9

1 log iI
C6

1 4

+
1'

log

I4
-

I'

Put )/2.log
c

.
-C6

C5 and

or

may be negative. but that always I

is always strictly positive, unless

. Suppose first

- I*

< C5

. firther.

Next we show how to

Then we infer

that

Then, from

r-c*

fl

and the choice of

ri we find by (7.8) that

RemarksL

1. Theorem 7.1 is an immediate corollary of Lehma 7.8. N of c ; ? will in practice disappear in the rounding C10 and Cll , the dominating factors

2. In practice, almost alvays the first term in the --definition dominates. Moreover, the tern are in practice Hence C1 to off. Similarly, in the definitions of

C4

7.7.
The remaining possibilities in cases (b) and (c) are 0
4

The reduction technique. C12 for B (or Clp

C5 < n

and

We nov want to reduce the upper bound

for

. which
i . We

n < -(C +C -M) < -C6 6 7

. So we

may take. noting that

1 ,

is equivalent), to a much smaller upper bound. We do so using the p-adic computational diophantine approximation technique described in Section 3.11. We perform this procedure for A

hi. Ki. Ki

, for the relevant

work in the p-adic approximation lattices Then (7.18) holds in the cases (b) and (c)

C1

themselves. and not in the

sublattices described in Section 3.13. The computational bottlenecks are the

. Now

take

computation of the p-adic logarithms to the desired precision. and the 3 application of the L -Algorithm. We refer to Chapter 3 for details. Once ve have found reduced bounds for ord (A) for the above mentioned A , we P combine these bounds with Lemma 7.6 and with estimates (7.13). (7.17) and (7.18) to find reduced bounds for B and B

Cl1

max

( C2, C4, C4.C7. C2 - C 9 ) .

* .

Then it follows that (7.16) is true, if conditions (7.20) and (7.21) hold. Hence,-by L e m 2.1, ve infer the following result.

When reduced upper bounds for

B.

are found in this vay, we may try the

LgMA 7.8- In the above notation,

replaced by their reduced analogons. above procedure again, with C 1 2 c;~ Ue nay repeat the argument as long as improvement is still being made. But at a certain stage. usually near to the actual largest solution. the procedure will-not yield any Curther iaprxwement, Then we have to find -all solu%otls-by

hold unconditionally. where

c;~

max

* [ 2 -( ~ +* h -ClO+h -Cll-log(h*-~ll)) max

--

some other method. One technique that may be useful is the algorithm of

* (h (ri-~i)+~]

iIu

Fincke and Pohst, described in Section 3.6. Another way is to search directly for solutions of the original diophantine equation below the reduced bounds. In our present equation this may well be done by below the obtained bounds. employing congruence arguments for finding all solutions of the second equation of system (7.9)

7.8.
RroofL Clear.

The standard example.

In this section we shall bork out the procedure outlined above for our standard example
(

pl,

.. ., ps

2, 3, 5 , 7

, thus proving Theorem

7.2. I n Tables I1 and 1 1 1w e give the necessary data on the f i e l d s f o r the 15 values of
D , and on the f a c t o r i z a t i o n of

Q(/D)

Sketch of each
Cl

roof,

I n Table V w e give the necessary d a t a f o r these 54 cases. w e give C 1 , nl, a l , h2, . . . h7 C1+l p 1 Ga(n) f o r a l l n E Z and

2, 3, 5, 7

in

W e explain t h i s t a b l e , and leave many d e t a i l s t o the reader t o check. For p

Explanation of Tables I 1 and 111. For generator of the i d e a l i d e a l . and w e give "Pia 2 2 , s o pi (si) hi Pi with

e give i n Table I1 a 2, 3, 5, 7 w pi ord (Pi) > 0 i f pi is a p r i n c i p a l pi i f it is not principal. I n a l l the l a t t e r cases, a t most one of the primes 2, 3, 5, 7 splits,

2, 3 . 5. 7

is given, then

I f for a

only for a t

Ga(n)

l e a s t one

n E Z

If

nl, al

a r e given, then

is p r i n c i p a l . A n a s t e r i s k (*) denotes a s p l i t t i n g D

prime. Note t h a t f o r each so pi

t 5 1

I n the f i n a l column of Table I1 w e give f o r the s p l i t t i n g prime

r

Define

n2

al

if

nl

0 , and

n2 cl+l p n

n1

if

nl

Then

n2

is the

a generator

of the i d e a l i : P .P
i

I n Table 111, when

pi

smallest p o s i t i v e index such t h a t

and
pj

not p r i n c i p a l , but

J
0

is, w e give a generator of i t .

are

C a (n2) I Ga(n) I , I * and

whenever symmetry

Ca(n2)

N o w it is true that

nl (mod a l l
of

.
the recurrence sequence

From Tables I1 and 1 1 1 it is easy t o f i n d a l l p o s s i b i l i t i e s f o r

a

W e may assume pi

I*

This

is
Q

related

I n Table I V w e give a l l possible

Ga (mod pi)

give primes ( a ) rr (a')

instead of indices set

I a (we i ). An a s t e r i s k (*) appears when

(Ga(n) In-

.
h 9

For q

to

properties

2. 3 . 5. 7

w e have defined

. The

XU

i s found by checking

f o r a11 pi

.
Hence

ord (C (n2)) 9 0

.
whenever C1+l p

There a r e 54 cases with

I + 0

-0
D

( t h e "symmetric" c a s e s ) , and 54 cases with W e s t a r t with the symmetric cases. This

t

2h2

(the "asymmetric" cases).

- P3. gh5-7h7

Ga(n)

I Ca(n)

W e have taken

incorporates a l l cases with 2, 3 , 5. 7 splits in Q(/D)

s o large t h a t always

3, 5, 35, 42, 210

. Now,

, when none of the primes

-- -

0 , hence equation (7.10) becomes

-

- --

-- -

--

With

(a)

s ' E
--

Z , B

- N c
-

c-r*

- .
fl
-

Consequently, there e x i s t s some prime r 2 1 1 t h a t divides Ga(n2) , hence C, +1 I Cp(n) I t follows t h a t f o r a s o l u t i o n r divides a 11 Ga(n) with p A

- - -- -

. -

Ga(n+2)
- .

A-Ga(n+l)

- B-Ga(n) .
no E 2

--

w e have f o r a l l

n E Z

of equation (7.22) w e must have

Since

(a')

t h e r e i s an

such t h a t

a'

fenO-a

. Hence

I n t h i s way w e f i n d with ease a l l solutions of (7.22).

L e t u s i l l u s t r a t e t h i s with the example

3 , a

/3

. Then

h y ve c a l l these cases w s y m e t r i c w . I n t h i s n E Z -,--which explains w s i t u a t i o n we can apply elementary congruence arguments, a s explained i n

for a l l

Section 4.5. W e have the following r e s u l t . =7.gL

and

Ga(-n)

Ga(n)

.W e

have f o r

Ga(n) :

Let

pl,
I

conditions (7.2) and

...,
(*).

pq

2, 3, 5, 7 ) 91

Equation (7.1) with

0 has e x a c t l y

solucions. that appear fn Table

I marked w i t h an a s t e r i s k

the n Ga(n) mod 4 mod mod 3 5

a , only on the

pi ) . The values

"nc, nx, n 2 , n3, n5, n7.

a r e the

integers such t h a t n n n "7 a 2 - k t c-nn-22.....7

I t follows t h a t i n a l l cases w e have

.
C12

< 3.23~10

30

mod 49 W e see that if n odd 2'. 3. 5 ( Ga(n) f o r a11 n

E

The next s t e p is t o define the l a t t i c e s ,

and f i n d lower bounds f o r the

hi

So n

and only i f

is the only i n t e r e s t i n g case. W e have 2 (mod 4) , 72 1 Ga(n) i f and only i f n

s h o r t e s t nonzero vectors i n the l a t t i c e s . W e start with t r e a t i n g the and 2

1 Ga(n)

i f and only 7 1 Ga(n) if 14 (mod 28) ,

which there a r e 3 f o r each of the 10 D ' s

.W e

of

have computed the 30 values of

(and i n general k 7 1 Ga(n) for

C1

- n

2-7k-1 ( a d 4.7*-l) such t h a t it i s a pi-adic took

p

integer, t o the desired precision of

digits. W e

r 1 , and a similar r e l a t i o n holds f o r any symmetric recurrence and

p f o r which a r b i t r a r y high powers of p occur i n Ga(n) ). Now,

a s follows:

- 0 does not lead t o (7.23). since then n2 - 2 . and Ga(2) - 7 . so t h a t no s u i t a b l e r e x i s t s . But with C 1 - 1 w e have n2 - 14 , and h2 - h3 - h5 - 0 , h7 - 2 . and (7.23) holds. since Ga(14) > 72 . Hence
t h e r e e x i s t s a prime whenever 72 1 Ga(n)

any prime

1 1

such t h a t

r I Ga(14)

and thus (7.22)

r I Ga(n)
w e have i n order t o have
p :

Ga(n) a 21-3O-5O-7l Ga(n2)

- .
14

I t follows t h a t f o r solutions of

somewhat l a r g e r t-+n

the maximal

so t h a t a11 solutions can be read from the above

t a b l e . Note t h a t it i s not necessary t h a t

is l a r g e enough. In our example,

e computed the 30 values of 1 . 0 5 ~ 1 0.~W ~ here. The l a t t i c e s

I'
C

is known e x p l i c i t l y , only t h a t

tG

b(')*s

c ; :

, being

b i t do not give them

337

or

- 3079

a r e generated by the column vectors of the matrices

satisfy.
-.
-- -

Finally w e t r e a t tXe remaining 54 cam,--*eye

f rr -0 ; Then we need tRee

-

non-elementary reduction technique described i n Sections 7.5 t o 7.7. I n a l l our instances, the s e t and ve v r i t e
m

W e performed the p-adic continued f r a c t i o n aigorithm of s e c t i o n 3.10 f o r each

contains only one element, since there is

of these 30 l a t t i c e s . I n the t a b l e below w e give f o r each

only one s p l i t t i n g prime. W e denote by for ci

the

ni

belonging t o t h i s prime,

. Equation

(7.10) now reads

C12 one f o r each

(there is one f o r each

a ). and the minimal bound f o r

C(I'C)

the maximal ( t h e r e is In a l l 1 yields

cases. L(I;)

* IU ) > (2-C12 . Hence

i
E

that w e found. W e o m i t further details.

L e u 3.14 with n

2. cl

0. c2

W e computed the constants

C 1

to

C12

C12

, according t o Section 7.6, f o r

D the pi E IU do not depend on

each o f t h e 54 cases. W e omit the d e t a i l s of these computations, and simply give the data i n Table VI. I n t h i s t a b l e w e give f o r each together with the vi and
Xi

( i t t u r n out t h a t the

Xi

It follows t h a t ord 's pi

ord (10% ( r e ) ) i s always the l e a s t doe of pi i i n the above t a b l e . So w e define:

the f i v e

and w e computed these numbers up t o

d i g i t s , with

a s follows.

a s given above. B y bounds f o r t a b l e above. Next. w e t r e a t the ui

e obtain from L e u 7 . 6 ( i ) upper ord (h*) 2 0 w pi a s given i n the i E IU , hence the upper bounds f o r U
Ai

so that

p :

i s somewhat l a r g e r than the maximal

: c :

W e computed the 40

values of the

6") , but do not give them here. The l a t t i c e s 1.2.3.4 generated by the columns of the following matrices:

cc

are

Ki

one f o r each

D , having 5 terms. namely

where i

j'i

, so

pi

is t h e s p l i t t i n g prime. W e have the following data;

From t h i s t a b l e our choice f o r

( D (mod pi)

becomes c l e a r .

W e computed the reduced bases of the 10 l a t t i c e s by the ~ ~ - a l ~ o r i t h Again. m. w e omit the computational d e t a i l s . W e f o d d a t a a s follows.

In all instances, : C ( I ' ri

ord (h ) 2 0 and hi 2 1 we have M d ord ( K ; ) PI I pi upper bound for M as given in the table above. Finally, we compute the new. reduced bounds for we do by (nl < max In1

"

> /5-C12 ,

SO

that by Lemmas 3.14 and 7.6(ii)

<

+ po

and

hence an

and thus for

. This

( C5, C6 + C7-X. C8 + C g - U )

Hence ve find data as in the following table. Here we used the bound

s h -B

and

h*

. So in one step we have reduced

B* < 3 . 2 3 ~ 1 0 ~ to ~

B* I 1627

The total computation time was

1715 sec. on average

0.7 sec for each 2-dimensional lattice. and 170 sec for

each 5-dimensional lattice.

The computation time was 1s sec. We made a third step, with for
.
--

Ai.

K ;

We ma&

a further reduction step, now using the reduced bound for C12

given above in stead of below. For

p

* .

B*

as

we took

pl-p2

. with

We give the data for the p1 as above, and p2

in the table

as below:

We found with p

L ( I .

; r e found. ) and bounds for U as given above. For the K P pl-p2 with p2 as above, and p1 as in the first table below.

the results given in the second table below.

~ O O I I I I I I ~

r r n n n m n n n n

n n m n n n n n n n

n n n o w w w o c e

c c e r c c r r r c

and finally for

In1 , and in more detail for

ord pi

for

I"

h l

L.

n n r m w e c a 0 ~ n n r m w e c m o ~ n n r m w ~ e m4 o eeerereeec w w w w w w w w w e mmmmmmmmmw

d n n r n w r - o a o c o r o c c a a r m

c . n n r m w e r a o m m a a a m s a m o
d

Now we will not find any further improvement if we proceed in the same way. But the upper bounds are now small enough to admit enumeration of the p remaining possibilities, making use of mod p arithmetic for only 3 sec computer time for this last step.

- 2.

3, 5 , 7

We did so. and found the remaining solutions, presented in Table I. We used

This completes the proof of Theorem 7.2.

7.9.

Tables.

o n n r o r m d m o
n
O I N

o - - o o c w n m n
N N

d W 1 I

wl-

e-+ncora
8 - W d r . 4
4

r a w - o o + o o m r n o ~ n o o
N

r I
I

Inn
n
I

c r

r n - r a w n o o n m o 4 r m n c o r I w o o n 1 I d N I
0

n r w m o r r - n n n r m n c c r n -on ~ n o mw
I d

I 0
-4

m n r n a m w o w r r r n n n o o a o a n n n r r n m m r
- 4 -

d - d

dooooooo+o c o m o n m n o o r w w r o d n m n r d d - r d - d - w n
r.

m o u n

o w ~ n

m o r o o o o - d o o n m o + m n o w r n o - o n - m o ~ n n m o m n o n ow n w a r r n r c o n o o n n o o o a w a r n - w o o w r r r w r n m0 - 0 - n o r r c o n n d n n n m n ~ n n n n r ~ ~ n ~ ~ nmr r r n n m r r
d

d N n r m w r c a 0 w w w w w w w w w r
-.ddr."-.lddd

WNOOOWlUb-d I c m o o m ~ r w m n I 1 1 1

w m w r o m o w a n m o a e m r w m r r
Nd
0

m m o o w o w m a o C I ~ I ~ P Q ~ P N ~ * m 0 w m r w o o l d

Table 11,

o d a m n w A w d m
d

o a w o r r o r o o o n o r m a d r nod - 1
I

m w o o w o o ~ n m r o d o a w r w n d w o o m o o ~ r l n aw a a r n r m
I

a ma w w m m m o r t c r n r a l o Pda m r r ~ m l n d
I
d

n on

m o ~ w m m o w ~ ~ w a o n n r o a c rr a n o r w r a p 1 N I d I I

a o o ~ a o o a m m r m ~ m o a o o o ro m m o o r m m d o r o o a r u m r n r r n a r n r r o o a o r o o w r n o r o o o o n o a ~ n o na w C m a o o n n w o a d o r d - m a 0 a n t c a a m n o m ~ c r w w r a m ~r r d n a m n o a a o m w w r m w a a r o a n n r a d w r a o o d m w r o a c m r 4 - 4 4d-d-n n n n n ~ n n w o r
d-

n n w o

w n r r d o m n d t r m a n w a d d n r m

r d c ~ r r r a

r a a m o o m c w o n o o r ~ m a w w o n r n r o m m o

Table IV,
NN NN
00
00

NN
00

MM
40
00

N
0

N
0

IU

40

0
4

0
0

0 4

0 0

0 0

0 0

4
0

N
0
4

0 0 0

0 0 0

00

6 - 4 0

N
0

O
0

0 0

0 0

do

MO 0000

00

MO 0 0 0

---

uu
04

NN

NN
OU
4 4400

ww
-In
00

0
0 -

Csf.
0 -

NN
04

N O U O M
0

N
0
4

M
0

A N
0
4

C s r .
0 -

N
44

N
44

NNOO

0000

00

0000

4040

0000

0040

0000

NO

-400

UN

MOO0

UMOO

-0

-004

MNOO

MdNd

4000

W - 0 0

Table VI,
4 4 4
4

4
0

0 0 0

i '

0
4

0
0

0
4

w l 4

( c b

( c b

(ccc
O w l
0000

b
0

(c

O w l
0000

O w l
0000

Y)

0000

O w l

u
4

U
0

N
4

N
0

U r n
4

4 .

wl00Y)

M o d 0

oouo

0000

NOON NQ

NNNN NNNN UUUU

wwwo

NNNN
0000

NNNN
0000

mmmm
0000 W h b W

0000 mnmm
0000
4.4.4.4

mV)YIUl
0000
4.4.4.4

NNNN

Table VI, (con

Acknovledgements.

The research for this chapter has been done in cooperation

with N. Tzanakis from Iraklion. The results have been published in Tzanakis b and de Weger 119871. See also Tzanakis [I9871 and de Veger [I987 1.

8.1.
I

Introduction. F(X.Y)
E

Let

Z[X,Y]

be a binary form with integral coefficients, of m be a nonzero integer. The

degree at least three, and irreducible. Let diophantine equation

in

X. Y E Z

is called a Thue equation. It plays a central role in the

theory of diophantine equations. In 1909 Thue proved that it has only finitely many solutions (cf. Thue [1909]). His proof was ineffective. An effective proof was given by Baker Tijdeman [ 19861 for. a s u . = e J r [1968]. See Chapter 5 of Shorey and

of -results on-Thue equations. By using;Lemaa 2 . 4

in Baker's argument, we derive a fully explicit upper bound for the solutions of the Thue equation. Then we show how the methods developed in Chapter 3 can
-

- -!
I

be used to actually find all the solutions of a Thue equation. Our method works- in pr-incipb for any Thue equation, and - iw-practice for a n * -

equation of not too large degree, provided that some algebraic data on the form

are available.

Variants of the method we use here have been used in practice to solve Thue equations by Ellison. Ellison. Pesek, Stahl and Stall (19751, Steiner [1986], Peth6 and Schulenberg [1987], and Blass. Glass, Heronk and Steiner 1 1 9 8 7 ~ 1 . b [I987 1 . When determining all cubes in the Fibonacci sequence, Peth6 [I9831 solved a Thue equation by the Gelfond-Baker method. but with a completely different vay to find all the solutions belov the upper bound.

8.2.

From the Thue equation to

linear form in logaiithms.

try to prove that no 'large* solutions exist. The value of Lemma 8.1 below. This lemma shows that if 1YI 'extremely close* to one of the real roots may be as large as 10los0 , Y2 <(i)

follows from Y1 is large enough then X/Y is

In this section we show how the solution of the general Thue equation implies an inequality involving a linear form in the logarithms of algebraic numbers with rational integral coefficients (unknowns). Let

In a typical example Y1 as small as 10.

Y3

as large as 10l0, and

.
. A

LEMMA 8.1,

Let

X. Y E Z

s a t i s f y (8.1). Put

- X - <-Y ,
if t . 1

2n-1. lml
be a binary form of degree the Thue equation n 2 3 and let m be a nonzero integer. Consider

yo

min l g * (t(s+i)) uist

min 1 I* ((s+i) B u t

in the unknowns

X. Y E Z

If

is reducible over

, then (8.1) can be

reduced to a system of finitely many equations of type (8.1) with irreducible binary forms. For such equations of degree 1 or 2 it is well known how to determine the solutions. Therefore we may assume from now on that irreducible over 9 and of degree
Z

is (i). I f IYI > Y o then there exf s t s an i0 E

(

is irreducible over 9 (X,Y)

. Let
g(x)

g(x)

. Then we
F(x.1)

may assume from now on that If g(x)

1,

has no real for

..., s

such t h a t

roots then one can trivially find small upper bounds for the solutions that the algebraic equation its roots as follows:

max(lXl. lY 1 )

of (8.1). Therefore. throughout this chapter we suppose [(l)

C(s+l)
have

F(s+t+l);,

.. . ,

. .... [ ' "

C(s*2t)

has at least one real root. We number

( s

IB(~)I 2
(ii). I f IYI > Y1

C ~ - I Y tor I
then2/Y

i E I 1.

.... n

i r io

r 1 )

are the real roots and i s a convergent from t h e continued f r a c t i o n

F(s*t)

are the non-real roots, so -that we s + 2-t

t ( r 0 )

pairs of complex-conjugate roots. and

-n

expansion of

<

( i , )

.
-

Consider the field

--

positive real nr$bers solutions

(X,Y)

, where g(f) 0 . We will define three Y1 < Y2 < .~;--,-that will divide the set of possible
. -

- a(()

--- -

Xio) n I be such that

?roofk

-Let

i0 E ( 1,

-18

__I

---min Uisn

of (8.1) into four classes: IYI

S

- .

have from (8.1)

I) the 'very small* solutions, with enumeration of all possibilities, 11) the 'small* solutions. with 111) the *largeq solutions, with

Y1

They will be found by They will be found by

c(~)*s ,

Y1 < IYI I Y2 Y p < IYI

S, . Y

evaluating the continued fraction expansions of the real

By the minirality of

. They
.

1 8

(Io)

we have for all

will be proved-nrto

exist by a computational diophantine approximation technique. IV) the 'very large* solutions. with IYI > Y3 They will be proved not to Hence exist by the theory of linear forms in logarithms.

le(i)l r C2-IYI

. Further.

The value of

Yj

follovs from the Gelfond-Baker theory of linear forms in follows from the restrictions that we use as we

logarithms. The value of Y2

Now, if

i > s 0

(and hence

t 2 1 )

then, by the definition of

Yo

where, in general, for logarithm of A E R and

r

C ,

Log(z)

denotes the principal value of the

5

(hence
n

-n

< Im L o g ( . )

r ). By

<(k)

- <(j)

we have

IA( I

.
(A1 is.

The following lemma shows how small

k I
n which is impossible if once. Moreover, if JYI > Yl and thus

min
-s+l.iIs+t

IImt(i)~

LEMMA 8.2. Put

1YI > Yo

. Hence

i0 I s

and now (i) follows at

then

C3

rnax i1ci*+i3+i1

I x F

- (io) ~1

< f - l ~ l - 2since ,

(io) (

is irrational. Now (ii)

0

If

( Y J> Y2

then

follows from a well known result on continued fractions. cf. (3.6). Now let j. k E
(

IYI > Y1 1.

and
)

i0

1,

...,

as in Lemma 8.1.

Choose Eroof. Consider first the real case. Fron IY( > Y ; and L e u 8.1 it a * -

....

such that

iO. j, k

are pairwise distinct and either ((k)

j. k e { 1. .. . s 1 or j + t choice of j. k is free. on eliminating the X and Y

-k

-X-

(so that

- ((j)
i

for

).

but further the

iO. j

.i

we get.

follows that the right hand side of (8.2) is absolutely less than consequently,

or, equivalently, It follows that -the-left hand side of (8.2) is equal to (i 0 - p -

eA-1 , and now (8.2)

B(k)

(k)-<(j (i0) we

( i , )

.L
@(j)

implies, in view of Lemma 8.1 and the definition of

C3

C
j

(io) -< (k)

,
s
)

8 " )

By Lemma 8.1, the right hand side of (8.2) is 'extremely small*. Put, if k E 1

- .

log

(let us call it *the real case') r(i~'-r(j)

On the other hand,

eA-1 1 <

implies (cf. L e -

2.2)

C
and if j, k E
(

(iO)

(k)

@(j) which proves our claim in the real case. (let us call it 'the complex case*) In the complex case the 1 e R hand side of (8.2) is equal to in the real case. we derive eu-1

s+l.

.... s+2-t )

and. as

Since by Lemma 2.3

1I

- 2- 1
5

sin A/2 1 , it follows that

I sin A/2 1 < f

. and therefore
and in the complex case
.

/ A1

2 . sin 1/4

Isin ~ / 2 1

- 2leiA-lls 1.02-/eiA-I/ , sin 1/4

0

---

which proves the lemma in the complex case. In the ring of integers of the field of r +1

with . a

.
.

and

-r < Arg(z) s r

for every

z E C

. Note
ai

that

in the

(as well as in any other order

cl.

- s + t - 1 (Dirichlet's Unit Theorem).

.

there exists a system of fundamental units

...
F

. where
K
are

real case, and

i-A

in the complex case, is a linear form in (principal) are integers.

logarithms of algebraic numbers, where the coefficients of maxla

Note that since

is irreducible

The Gelfond-Baker theory provides an explicit lower bound for

and we have supposed

s > 0 , the only roots of unity belonging to

I i

in terms A Using this in combination with Lerma 8.2 we can find an

We shall not discuss here the problem of finding such a system (for

efficient methods see e. g. Berwick [I9321, ~illevix [I9561 , [I9641, Pohst and Zassenhaus [1982]. Buchmann [1986], [1987]). We simply assume that a system of fundamental units is known. On the other hand, there exist only finitely many non-associates i pl.

explicit upper bound for maxla / i

. This

is what we do in the next section.

---.

pv

in

K
pi's
(

such that

fo-N(pi)

-m
K/a

8.3.

Upper bounds. A

for
.) We

1.

..., v .

(Ve use
$

N(-)

to &note

the norm of the extension is known. Let

Let

rnax (ail

. First we

find an upper bound for A in terns of

/YI

also assume that a complete set of such of all IfO/

be the set

lsisr

- - .
1 0 1

capi , where 1 (X,Y)

is a root of unity in

it is clear that

(In the important case Then, for any integral al.

-1. 1

) ).

solution such that

of (8.1) there exist & e

p E M

and

.. .

E 2 -,

(vhere
-

i n d i c a t e s a row and

a coluam of t h e matrix)

--

--

..

Thus, the initial problem of solving (8.1) is reduced to that of finding all a r for some p E M b e . integral r-tuples (a1.. ar) such tha; p:e tr of the special shape X

.. - Y-e

with

X, Y E Z

. As

. ..-

Put also

we have seen, X

and

can be eliminated, so that we obtain (8.2). solving finitely many equations of the type

Thus the problem reduces to

(the so-called 'unit equation').

In the real case we have

Then, for

i
(
\

Next we show that

w e have

~l~~l$ll

< (n-l)-log(~~1~

1 1 for

- 1.

...

A < C~-~O~(C~-. IY()

Indeed, in view of (8.6). Suppose now that

a stronger inequality is true if

(i)/p (ill (B

l/3(i)/p(i)l

< 1

By

m o o f L By

/ j

a a 1 - pee1 -...c

we have it follows that

On the other hand, for every of Lemma 8.1.

h E (1 1,

....

, using the end of the pra

in view of (8.6). Now the inequal

follows from (8.5). N[u;~]

(8.7).

the definition of

N [ u ; ' ]

and the fact that. as

we have not put so far any restriction on be minimal. It remains to show that

I , this could be chosen so that

A < max N [ u ~ ~ ] . ~ o ~ ( c IYI) ,+I

.
h E I

i0
Q

Choose

such that

Then, by Lemma 8.1. for every

lfl(h)/p(h)

> C2. IYI/p+ > 1 and now, in view of (8.6).

Note that

C4-lYI > 1

Indeed. by which implies our assertion:

i-1 it follows that min Uiln p-

Lemmas

8.2 and 8.3 immediately yield

a 1mI

Therefore
i

LgQu 8.4, - .

Put

c4- lY I
Then.

c;

.+

max 1 s i l < i 2 , 1 (

(ill

-t

(19I

I - 1YI.ImI -11"

>

I Y I> 1
21m11/n

If

IY I > Y ;

then

Next w e apply L e m m a 2.4 (Waldschmidt). I t yields i n the r e a l case (assuming that

A rr 0 )

8.4.

Reducing the upper bound. Let be given r e a l

W e a r e now l e f t with a problem of the following type.

and i n the-c-omplex case t h i s holds when

i s replaced by

A'

numbers max lai( 0s-r

6 . pl,

..., pq

( q L 2

, the case

1 is t r i v i a l ) . Write

where the max lail . I f K1. K2. K3 be Bilq given p o s i t i v e numbers. then f i n d a11 q-tuples (a1. .. . a ) E Zq s a t i s f y i n g 9 ai's belong t o

The precise values f o r

C7

and

C8

a r e given i n Section 2.3.

It should be

Z , and put

noted t h a t i n the complex case

makes now its appearance. while it was a . n o t present i n Lemmas 8.3 and 8.4. I n order t o obtain an upper bound f o r A
A'

w e .must f i n d an upper bound f o r relation

i n terms of

Indeed, using the I n our case, it follows from (8.3) o r (8.4) how t o define pi's q, 6 and the Put

, and from Lemmas 8.4 and 8.5

and
K

how t o define
K3

K1,

K2, K3

I n general,

w e f i n d from (8.4) and the proof of lemma 8.2 t h a t

(aO1 <

;-r-A

K1

are

'small'

constants, whereas

is 'very l a r g e ' .

IA1/2r
A

< 1+

r - A Ir - A C8

if by

A L 2

Thus w e may apply (8.8) i n both cases with

i f we replace

C i , where

so that

+ A. . W e

apply the methods of Chapter 3 t o problem (8.9).

Ci Ci

C8 C8

i n the r e a l case,

log r

i n the complex case.

A

Below w e d i s t i n g u i s h three cases. I n the f i r s t two w e suppose t h a t the are Q-independent. ( i ) . Let (iiX L e t (iii).
--

pi's

W e can now give an upper bound f o r

.
-

- .
0

so t h a t

A .

. Then

the l i n e a r form i s homogeneous. and

w e apply the method of Section 3.7.

8 . 5 ,
Cg

Put

6 z 0

Then the l i n e a r form is inhomogeneous, and w e apply the the pi's a r e a-dependent.
A

- -.(
2

5 log C6
-

method of Section 3.8.

C -C'

8
-

C5-C7 C7.10g -) n
-

Suppose now t h a t

Let

be

the

If

(YI >Y+

then-A<-Cg

approximation l a t t i c e f o r the l i n e a r form

-

a s defined i n Section 3.7. i n general t o b e

Theri we-expez-he 'very small',

lowerbound f o r

1x1

LI 7 z z Q )
. .-

Proof

As w e have -seen in-the proof of L e u 8.2,

1e

h l1

<

since the vector having a s coordinates the c o e f f i c i e n t s of the

i n the r e a l

dependence r e l a t i o n w i l l give r i s e t o a very s h o r t vector i n the l a t t i c e . So t h e reduction process, a s applied i n the two previous c a s e s , w i l l not work. I n such a case w e work as follows. Let

c a s e , and

leu-11

<

i n the complex case. Note t h a t Lemma 8.4 and (8.8) y i e l d

fi

(io)

. Hence

(8.2) 'implies -A rr 0
A

. Therefore
C6

be

maximal

subset of

C5 ( log < -n

C -C'

C7-log A )

.
0

(pl, ...,p ) q of s u b s c r i p t s w e may assume t h a t find integers d

consisting of a-independent numbers. With an appropriate choice

Fl

> 0 and d

iJ

for

) --p -<- q Then we can { pl, - - - * "P 1 S i 5 p , p+l S j S q such t h a t

The r e s u l t now follows from L e m m a 2.1.

Rtolark.
that

From t h i s upper bound f o r

an upper bound f o r

(YI

can be (These numbers d. dij can be found a s coordinates o f extremely s h o r t vectors i n reduced bases). On the other hand, (8.9) is equivalent t o

d e r i v e d , thus a value f o r

Yi

Y3 (cf. L e m m a 8.4) is not necessarily equal t o

( c f . Section 8.2). W e s h a l l not do t h i s . Note

Y2

( c f . Section 8.2).

small enough to admit enumeration of the remaining possibilities in a reasonable time. then it might be necessary, or at least advisable. to use where
A'

- d.A
ai

and d-a.

Ki

- d-K1 . Now. with

9 diJ .aj

6'

- d-6

and

the technique of Fincke and Pohst, cf. Section 3.6. However, when solving a Thue equation, and not only an inequality for a linear form in logarithms. it may be better to avoid this method, and to use continued fractions of the roots C

j-p+l

--

((i)

In practice we can search for the solutions C

(X.Y)

of (8.1)

we obtain

satisfying

- Y2 . and we can imagine

f

Y1 < IYI I C

as follows, referring to Lemma 8.1. Here e.g. here as being a 'large' constant compared to Y1. Y2 in Section 8.2).

Y1 , but not 'very large' (cf. the introduction of

Put

- max

( ldl, ldijl

: 1 1 i 5 p. p+l 5 j I q )

Then

Let

be a rational approximation of

<

( i , )

, such that

l a ' l s (q-p+l)-D-A for i Therefore, if we put implies A*

- 1.

.... p .
A ' I (q-p+l)-D-A , and (8.10) Since IYI > Y1

max la! 1 lsisp

'

then

X/Y

fraction expansion of

must be a convergent. pk/qk say, from the continued (id) . Denote by ao, al, a2, ... the partial

<

quotients in this expansion. First we claim that where by (3 5 )

a,+l

2 3

Indeed, we have

Ki

K2/(q-l+p).D

.
-0

Kj

- (q-p+l)-Kg .
A'

If since

-1
-

or

. then we would have

.
Thus.
- -

l ~ l n -< 2 4-C1
--

. which is absurd.

Now. to solve (8.11) we apply the reduction process described in (i) or (ii), depending on whether 6' or 6 ' rr 0 , and maybe more than once. if necessary, until we find a very small upper bound for found all solu~ons--(a'a for make lA'

IYI > Yl > (4-C1) 1/(n-2)

a,+l r 3 , and by (3.5) we-have

After having
-?

1 .

) of ( 8 3 ) . we have a l o w e r - b o u n d 3 > 0 P It is reasonable to expect that L is not 'extremely small'. ai.

Therefore.

because the -integers Ih'

...

. a'P

being 'small'

in absolute value cannot

'extremely small'.

Now combine

I A ' l

z L

with the first

inequality of (8.10) to get and this means that fraction expansion of Since for A Pk/Qk is in fact a convergent from the continued

too. Moreover. in view of the inequalities

is not 'very small', as argued heuristically, the above upper bound is 'small'.

Returning now to the general case. we point out that if the reduced upper bound for A (found after some reduction steps as described above) is not

must be sufficientli -large compared to

qk

namely

8.5.

An application: integral points on the elliptic curvc

y2 = x3

4-x

This inequality can be checked easily for all k

such that

qk

IC

.
( i , ) for

In this section we prove, as an application of the general theory described in the previous sections, the following result.
---.--

To sum up, we propose the following process for every real root i0

1.

...,

THEOREM 8.6,

The elliptic curve

(note that of

i0

is a priori not known).

(1) Compute a

rational approximation decimal expansion will do). partial quotients i

(io) satisfying (8.12) (a truncation of its

$ .

(2) Expand

into its continued fraction with and convergents

has only the following 22 integral points:

ao, al, a2, .. . , ak+l qk

5

for all for the that if

1.

.. . , k

with

C < qk+l

conditions (8.13) and

X/Y X/Y

pi/qi , then

F(pi.qi) Z-pi m

. Y

. (3) Test all these convergents . Concerning this last test, note
Z.qi for some Z E Z with

Zn

.
We prove this theorem in two main steps. First, we reduce the problem to the solution of two quartic Thue equations. Then we solve these equations using the general theory developed in the previous sections. Let

This simple observation excludes in general most of the reducible quotients

. and all of them if

.

is an

n-th-powerfree integer. (YI

5

Having tested for all solutions in the range IYI > C For such solutions (X,Y)

we may suppose that

we can obtain a lower bound for the

corresponding A as follows (the idea is due to A. PethB. cf. also Section 1 b of Blass, Glass. Heronk and Steiner [I987 I). For every (i.j) E (1.. . .,rl x (1..

L be the totally real field Q($) , where

.. .nl

let

p.,

be the number

+1

or

-1

for which

c : j )

1 , Let. the conjugates of

$(3)

be

I'($

1.860.

.. . From

-7

0.254.4,

$(2)

a table of Delone and Faddcev ( il964]

i2.114..
Z[$]

.,

p. 141) we see

that the class number of

. -

is 1, its ring of integers is

--Table
2-$

discriminant is 229, and a pair of independent uniis (9. 2 I of Buchmann [ l m we-s& hat -7 + 2-$ 3 - $ + $* fundamental units

its From and

and hence for any pair

j l. j

with

z j

we have

+ d2

in

Z[$]

Since
$, 2

( 2 $ ) '

we see that

-$

is a - p a G 0 3 ( 2 )

is also'a pair of funun&ental

units in Z[$]

and from this we can find a lower bound for Of course, for an other pair j , j and therefore ve can take the larger one.

A , if we know that

IYI > C

The equation (8.14) of the elliptic curve can be written as

--

we may find a different lower bound,

y2

( x

- $ ) - ( x2 + x - $ +

($2-4) )

(8.15)
r

and the factors on the right hand side are relatively prime. Indeed, if were a common prime divisorof them, then n would divide

which is prime. since its norm is -229. Therefore we would have that r is a 2 unit times this prime. and then by (8.15). x - $ unitx(3 - $ -4)xsquare . f229xsquare which is clearly impossible. Take norms, then we get y2

,
I

Since the sign of assume w

- -(m+n12

z , and thus that of

. After

. is of no

importance, we may

substitution in the second equation of (8.18) we

Now (8.15) implies

obtain the Thue equation

Since (8.14) is trivial to solve for Since +(I) 0.254. . .

x 5 0 (the only solutions with x S O are the first three pairs stated in the theorem), we may assume that x r 1

Then. by

$(2)

. we see that the minus sign in (8 .l6) is impossible.

.
i
L

The left hand side can be factored as

-2.114. . .

. We

conclude therefore that and ?(m.n)

(x.3)

therefore

- (1.0).
(20.89).

it

can

be

solved very

easily.

(0.1)

They lead to x

then by (8.18) we find a r s t case:

1

-0

- 20. 12
ul

?(u,v,w)

Its

only

solutions (4.-2.-1)

are

(4.2,-1).

, and

respectively, which furnish the solutions

(12.41)

for (8.14). and

Then

(8.17)

implies.

on

equating

corresponding Secondly. we suppose that ul r 0 there exist

coefficients in both sides.

are odd. Then

vl

is even, so by

m, n E Z with

Note that u

- 2.ul . v - 2 - v l . The last equation of (8.18) now reads

w

is odd and

is even. hence

4 ( 2 - u - w , so

is even. Put It follows that u

?-(a2+n2) , v
3

2.m.n , w

- -m 2

or

- -n2 .

Consider this as a quadratic equation in 2 square, z say. Then

Its discriminant must be a

We may assume that

- -m2

Substituting this in the second equacion of

(8.18) we find the Thue equation

Note that ul and

The left hand side is again reducible. The only solutions; as-is easily seen,

z have the same parity. We may assume u r 0

are

f(m.n)

(1.0).

(1.1).

(1.-1)

Since

and

cannot have the same

(u.v.~)

First suppose that is odd, we find

2 zl

. Then
m, n

u2
E

U1 -- 2 - z1 .
ul
-

and

are even. Since

w2

2 u l - w + vl

2 vl

(mod 4) , and v1 is odd. Put ul 2 where u2 and vl are odd. By

and 2-u2 u2
2

.
1
1

parity, only the first pair is accepted. It leads to and hence to (x.fy) (4.7) for (8.14).

(2.0.-1)

case:
x

Then. equating the coefficients in (8.17) we get 2

there exist

Z such that

2-u2 + v

4.w2

2.u.w

- 4 - v - w,

(8.19)

It follovs that

The first relation of (8.20) can be replaced by

THEOREH 8.7,

(i) . The Thue equation 3 4.X .Y

x4 Note that yields w2 u is odd. Put z

2 2 12-X .Y + 4.y4 f(X,Y)

-1
.

- v - 2.w

Then the second equation of (8.20)

has o n l y t h e s o l u t i o n s

(1.0)

(ii). The Thue equation

- 2-z-(u-z) .
has o n l y t h e s o l u t i o n s

f(X,Y)

(1.0), (1.-l),

(1.3). (3.-1)

First we suppose that

is odd. Then there exist m, n

Z such that
Proof. We use the notation and results of Sections 8.2 and 8.3.
6 and
9 be defined by

where we use that

u 2 0

and

(u,w)

. Thus, choosing signs properly.

Let the algebraic numbers

Substituting this in (8.21) we obtain the Thue equation

Since
9

cp

- 2/6

, it follows that 6 and

cp

generate the same field K n

In the notation of Section 8.2 we have

(

or In Theorem 8.7 below we prove that this equation has only the solutions +(m,n) (x.3)

cp

. Simple computations show that for

< - 6, (p

4, s

4, t

over

0. and

<

we can take

- (1.0) . leading to
(3.4)

(u,v,w)

(1,1,0) , and finally for (8.14) to

Secondly we suppose that 2 z-2-m

is even. Then there exist 2

n, n

&

with

C2 from ; C2, -Cy _C4 from above and below, making use of the following approximations for the conjugates of 6
-

In these computations we estimate . cp :

u - z - n

and

Now, substituting into (8.21), we obtain the Thue equation

Now we work in the order In Theorem 8.7 below we prove that this equation has only the solutions ( n ) (u,v,w) (note that

of

with

Z-basis

I 1, 6. f ~ 6 ~ f -d3 . 1

- (1.0.0).

f-e2

is an algebraic integer). Note that

(01)

(1-1) (3.-2.-2),

(3.1),

(-13)

They

lead

respectively

to

q , -2 z-4+6-e-~.6 3 E R .

to the solutions (8.22),

(x,+) (2,1), (10,31), (1274,45673). (116.1217) Thus this result completes the proof of theorem 8.6. provided the Thue equations

(19.30,6),

(11,-10,-6) , which lead for (8.14)

On

the

other hand,

Nonk/p(X-Y-6)

(8.24)

and

(8.25)

and

Ilor4~q(X-Y-cp) then X

(8.23) have as their only solutions the pairs

(m,n)

mentioned

a solution of (8.24) or (8.25),

above. We now proceed to prove this.

, respectively, is a unit of the order R . A system of fundamental units of R is given by

or X

- Y-6

- .
1

are

respectively equivalent

to is

which means that if

- Y.9

(X,Y)

t l - 1 + 8 .

c 2 - 3 + 8 .

1 2 t3---8

the theorem. Therefore w e may assume t h a t

IYI > 3

so that

W e do not prove t h i s f a c t here. For a proof, see Tzanakis and de Ueger I19871 Section 111.2 and Appendix I . Before Thus the solution
E

w e

apply

the

reduction
A

method of S e c t i o n

3.8

w e

show

that

the
V 's

of

(8.24)

and

(8.25) kc1- c

is

reduced
t 3

to

finding

all

application of Lemma 2.4 y i e l d s the above constants r e s u l t i n the case of f o r the various for
i

C7, C8

W e apply t h i s

(a1,a2,a3)
X

z3
X

such t h a t the u n i t

Ye8

or

Y.9

has the s p e c i a l shape

respectively. I n the notation of L e m m a 8.3 w e have.

a f t e r some numerical computations. t h a t w e leave t o the reader t o check. t h a t enin N[u;'] I (here. of course.

- 1. 2,

ai*s 3

given by (8.26). I n t h i s case, w e compute the (k) appearing i n A a s follows. I f ai I ci

ifi

thl

then

ai

is a u n i t and hence

a .

(appearing i n the ai
is i n

- 0.634950.. . .
I

max N [ d l ] I I

1.210070..

computation of h(o ) ) is equal t o 1. Clearly. every conjugate of i absolute value l e s s than

1. 2. 3. 4

) ).

Therefore w e can take i n Lemma 8.4

and Also,

Hi

. Therefore.

h(oi)

Hi

, and we can take

Since the l a t t e r t e n equals the logarithm of e i t h e r (The values of

C5

1c

or its

and

C6

a r e estimated from above.)

inverse, it follows t h a t
Vi

Now, r e l a t i o n (8.3) becomes i n our case (i,) (k) - log C -p + 11 ai-log 1

log Hi

'

I:+

absolute value l e s s than where

i is -_a
-

C3

Therefore,

h(ai)

(log ao)/d + log C3 for

i

where choose

j. k

o,r cp As mentioned i n Section 2. once . -=-.- a r b i t r a r i l y . Thus w e can choose

a .

and

a r e a s i n the d e f i n i t i o n of
-

h(a)

a -

.
.

ai

. An

upper

fixed * w e can

boynd - fo _ r a

xih
Now, x(0)

can be computed a s f o l l o r s . X o q ider-the - a l g e b r a i c .numbers for

i , h E ( 1.

f-(((iP-((h')

....

with

.
8

I t can be or
cp

checked t h a t the numbers

.xih
o

f o r each permutation

(0 o o o ) E S4 w e consider the number 1 2 3 4 (independent of o4 1, and the polynomial

a r e algebraic i n t e g e r s f o r

(-0

Therefore. f o r each

<

E ( 8, cp )

w e have four p o s s i b i l i t i e s f o r

For

each of these e i g h t cases w e have, a s w i l l be shown below, Consider a l s o the number and therefore. by Lemw 8.5, i f t h e upper bound
C9

max lail w e have U5s3 3 . 2 6 ~ 1 0 ~ ' A s i s e a s i l y checked, the only solutions of

IYI > 3

then f o r

e i t h e r (8.24) o r (8.25) with

I Y I s 3 a r e those l i s t e d i n t h e statement of

s i t u a t i o n we have t o solve (8.9) with

'
1

K1

where

is the
A2

therefore

discriminant of

the defining polynomial

~ ( o )

of P(X)

229

. On

the other hand, the c o e f f i c i e n t s of for

, and a r e up t o

<

4 - - 6.38771~10 .
C6

K2

- - 1.211 >
n 4

3'303 , , . K

- 3 . 2 6 ~ 140 0

( K2

i s estimated from below), and

A

t h e s i g n the elementary symmetric functions of they a r e symmetrical expressions of the This means t h a t P(X) E Q[X]

S4 , and so
A

*s
d4

with r a t i o n a l c o e f f i c i e n t s . ,yih*s

6 + alepl
and the

On the other hand, by the d e f i n i t i o n of

a2-p2 + a - p 3 3 ' w e have the following p o s s i b i l i t i e s . i n view of

where f o r

pi's

any c o e f f i c i e n t of with c o e f f i c i e n t s i n since 22g2 that ai

P(X)

multiplied by

is a polynomial of the

(8.26) and (8.27):

and therefore it is an-algebraic integer. Combine P(X) E Q[X] t o see t h a t h(ai) Since 22g2-p(x) E Z[X] a .

t h i s with the f a c t t h a t
is a root of

. Hence.
(8.28)

P(X) , i t s leading c o e f f i c i e n t

i s a t most

. To

conclude. w e have

s 2 - ( l o g 229)/d + log C3
ai @ 9 w e have d r 2

!log ail/d 5 log C3

Vi

and i t is c l e a r so w e can take

log 229

log C3

Simple computations now show t h a t log H1 log H3 l o g C3 log C3

4.074586 ... , 4.821584 ... 1.262065

log H2

5.667432

...

.
Numerical cO

I
(

(3) (1) l o g ( 3 ) ( 2 )

or

:- log

.
if if

(8.29)

...

1.893823 ... log C3

5

- . <- ,
cp

details

are

given

i n Tzanakis

and

de

Ueger

[1987].

U e

take

and w e work with the l a t t i c e with associated matrix

l o g 229

7.327545...
-

.
n e have the same Note t h a t i n each of the four cases of (8.28) (resp. (8 .-29-)) w lattice,

Therefore we apply-Lemma-2 4 (Waldschmidt) with

rl

(resp.

r2

),

say. I n each case a r e 9-dependent.

. and
-

w e had no

for
V1

l o g H1 , Thus we f i n d t h a t
V2

8 'or

, - and bl
log H3

- . - . b3-a3* , - . - al b2 a3
V3 V ;

numerical evidence t h a t the

pi's

Therefore w e worked a s i n

B - A .

log H2

V4

V4

b4-1 * log 229 + log C3

case ( i i ) of Section 8.4.

-

For each

ri

3 w e have applied the i n t e g r a l version of the L -algorithm,

8 , U , U-'

and

each time we have computed the i n t e g r a l 3x3-trices

, a s defined

with

C7

5.71~10~ and ~

C8

- 6.17

i n Section 3.7.

I n our cases, the coordinates of the v e c t o r s of the reduced b & i s vectors a r e of the s i r e of
.

.
I n our

bases ( i . e . t h e elements o f o S ) turned out t o have 46 t o 48 d i g i t s , i . e . lengths of &'reduced c;l3

the

a s expected.

W e have now t o apply the reduction process described i n Section 3.7.

I n each of the e i g h t cases we computed the coordinates

i

s l , s2. s3 of

8.6.
Let bl, h p ,

The Thue-Xahler equation. an outline. F(X,Y) be as in Section 8.1. Let pl,

...,

ps

be fixed distinct prime

numbers. The diophantine equation with respect to the reduced basis computations we found

b3

of the lattice. --- -- From our

lhll > lbll >

3.247~10'~

in the case of lattice

T1 ,

in the variables

X. Y

Z , nl,

...,

NO , with

(X,Y)

-1
an

, is known

4 . 8 4 6 ~ 1 0 ~in ~ the case of lattice in all 8 cases.

r2

.
3 , and

as a Thue-Hahler equation. It was proved by Hahler I19331 that this equation has only finitely many solutions, and by Coates (19701 that they can, at least in principle, be determined effectively. since effectively computable upper bound for the variables can be derived from the p-adic theory of linear forms in logarithms. For the history of this equation we refer to Shorey and Tijdeman [1986], Chapter 7.

Is31 > 0.029

This means that in view of Lemma 3.5, in all cases L(ri.x) > 0 . 0 2 9 . ; .

i0

3 . 2 4 7 ~ 1 0> ~~ 4 . 7 0 8 ~ 1 0.~ ~

Then the assumptions of Lemma 3.10 are fulfilled with c

K1. 6

K2. Xo

- X1 A
I

K3 , since

427.~ < ~1 . 1 1 2 ~ 1 0 ~which ~ implies

3.

1. C

- c0'

We'believe that it is possible to solve Thue-hhler equations, not only in principle. but in practice. This can be done by reducing the above mentioned upper bounds, using a combination of real and p-adic computational h m reducing diophantine approximation techniques, based on the ~ ~ - a l ~ o r i t for

A <

-3.303 I

10g(l~140 6. 38771x104/3. 2 6 x 1 0 ~ 4 < 72.8 72. We repeat the procedure with

.
K3

It follows that c
0

- 1012 . We found from our computations

lPll > IP1l >
1 .092xlo4

72

and

bases of lattices (cf. Sections 3.7. 3.8, 3.11 and 3.12). which we report in the preceding sections.

The method can be

considered as a p-adic analogue of the method for solving Thue equations, on

1 . 2 9 3 ~ 1 0 ~in the case of laaice

-

in the case of laitice

A

. r2 .
T1 i0

- , 3 Such an idea (but without using the L -algorithm) was used by Agrawal.
-

Coates, Hunt and van der Poorten [1980], who determined all solutions of the

Is3! > 0.143

-

in all 8 cases.
--

This means that in view of Lemna 3.5. in all cases

- .
3

--and

equation
-.

x 3

- X 2 .Y + X - Y2 + y3 - +1lX .
-

--

This is one of the only 'two examples in the literature where a Thue-Hahler equation has been solved completely, the other one being ~ 3.742~10 , Then the assumptions of L e m a 3.10 are fulfilled, since 4 2 7 : ~ < which implies X
2

3.y3

2 " ,

which was solved by Tzanakis (19841 by a different method. Both examples are of the simplest kind, in view of the fact that the cubic field It follows that

10

We enumerated all remaining possibilities, and

O

is a root of

F(x.1)

Q(8)

, where
real

0 , has only one fundamental unit, and there occurs

two-dimensional

found no other solutions of (8.24) and (8.25) than mentioned in the theorem. This completes the proof of Theorem 8.7. hence also that of Theorem 8.6. The computations for the proof of Theorem 8.7 took 35 sec.

only one prime. Therefore it is sufficient to use

continued fractions and ane-dimensional p-adic continued fractions, instead of the more complicated ~ ~ - a l ~ o r i t h (which m was not yet available in 1980). 3 With the use of the L -algorithm the method can in principle be extended to

204

the general situation. where there are more than~onefundamental units. and more than one primes. In a forthcoming publication, Tzanakis and the present author plan to give details and worked-out examples.

After each reference we mention in brackets the section(s) reference occurs.

in which the

Agrawal, M.K., Coates, J.H., Hunt, D.C. and van der Poorten. A.J. [1980]. Elliptic curves of conductor 11, Wath. Comp. 35, 991-1002. (3.8;3.10; 8.6) Alex, L . J . [1976], Diophantine equations related to finite groups, Como. Algebra 4, 77-100. (6.1;6.5) Alex. L . J . [198sa]. On the diophantine equation Math. Comp. 44, 267-278. (1.1;5.4) b Alex, L . J . [I985 1. On the diophantine equation Arch. Math. 45, 538-545. (1.1;5.4) Babai, L. (19861, On LovAsz lattice reduction and the nearest lattice point problem. Combinatorica 6. 1-13. (3.4) Bachman, G. [1964], I n t r o d u c t f o n t o p-adic Numbers and V a l u a t i o n Theory. Academic Press, New York and London. (2.3) Baker, A. [1966]. Linea~formsin the - logarithms of algebraic numbers, Mathematika 13, 2042216. (2.4) Baker, A. [1968], Contributions to the theory of diophantine equations, I. rqwtion 1 1

+ +

2a 2a

C s ' 3

+ 2 d 3 e ~ f,

3b7C + 2d3e7f ,

On the representation of integers by binary forms, I_I. The diophantine 3 y2 u k , P M I . ~ r u uR-. . ond don^-^ 263. 173-208.

-*.-

(8.1) Baker, A. [1972], A shatpening of the bounds for linear folms in logarithms
I . Acta A r i t h . 21, 117-129. (1.2)

Baker, A. [1977], The theory of linear forms in logarithms, Transcendence Theory: Advances and A p p l f c a t f o n s . Academic Press. London. pp. 1-27. (2 -4) Baker, A. and Davenport. H. [1969]. The equations 8x2

- 7 - z2

2 3x

- 2 - y2

and

Q . J 1 . Math. O x f o r d (2) 20. 129-137. (1.1;1.4;3.3)

Berwick, U.E.H. [1932]. Algebraic number fields with two independent units. Proc. London Uath. SOC: 34, 360-378. (8.2) h ; generalized Raaanujan-Nagell equation, Acta k i t h . Beukers. F. [1981], On t

I: 38, 389-410; 11: 39, 113-123. (1.1;4.1)

~illevig,K.K. [1956], On the units of algebraic fields of third and fourth degree (Russian), Hat. Sbornik Vol. 40, 82, 123-137. (8.2) Billevic, K.K. (19641, A theorem on the units of algebraic fields of n-th degree (Russian). Hat. Sbornik Vol. 64, 106, 145-152. (8.2) Practical Blass, J., Glass, A.H.U.. Heronk, D.B. and Steiner. R.P. [1987~],
-- -

C h e ~ b i n i ,J.M. and Walliser, R.V. [19871, On the computation of all imaginary quadratic fields of class number one. Wath. Comp. 49. 295-300. (3.2) Coates, J. [1969], An effective p-adic analogue of a theorem of Thue. Acta Arith. 1 5 , 279-305. (2.4;6.1) Coates, J. (19701. An effective p-adic analogue of a theorem of Thue 11: The greatest prime factor of a binary form, Acta k i t h . 16, 399-412. (2.4; 6.1;8.6) Delone, B.N. and Faddeev, D.K. [1964]. The theory of irrationalities of the third degree, Transl. of Math. Monogr.. Vol 10. A.M.S., Providence R.I.

solutions to Thue equations of degree 4 over the rational integers, Preprint. Bowling Green State University. (3.8;8.1) b Blass, J., Glass, A.M.W.. Heronk, D.B. and Steiner, R.P. (1987 1. Practical solutions to Thue equations over the rational integers, Preprint. Bowling Green State University. (3.8;8.1;8.4) Blass, J., Glass, A.M.W., Meronk. D.B. and Steiner, R.P. [1987C]. Constants for lower bounds for linear forms in the logarithms of algebraic numbers I: The general case. Preprint. Bowling Green State University. (2.4) d Blass. J., Glass. A.M.W., Meronk. D.B. and Steiner, R.P. [I987 1 . Constants for lower bounds for linear forms in the logarithms of algebraic numbers 11: The homogeneous rational case. Preprint, Bowling Green State University. (2.4) Borevich. 2.1. and Shafarevich, I.R. (19661, Number Theory, Academic Press. New York. (2.1;7.3) Bremner, A.. Calderbank. R., Hanlon, P.. Morton, P. and Wolfskill, J. [1983], Two-weight ternary codes and the equation y2 Theory 16, 212-234. (4.1) Bren~er,J.L. and Foster, L.1. [1982], Exponential diophantine equations,
- -

(8.5) Ellison, W.J. (197la] Recipes for solving diophantine problems by Baker's

method, SBm. Thdorie des Nombres, Universitd de Bordeaux I. 1970-1. Lab.

Th. Nombr. C.N.R.S., Exp. 11. 10 pp. (1.4;3.8) b Ellison. W.J. [I971 1. On a theorem of S. Sivasankaranarayana Pillai. SCm.
Thdorie des Nombres, Universitd de Bordeaux I, 1970-1, Lab. Th. Nombr. C.N.R.S., Exp. 12. 10 pp. (3.2;S.l) Ellison, U . J . . Ellison, F., Pesek. J., Stahl. C.E. and Stall. D.S. [1972]. The diophantine equation y2 + k

- x3 . J. Number Theory 4. 107-117.

(3.3;B.l) Evertse, J.-H. (1983). Upper Bounds for the Numbers of Solutions of Diophantine Equations, HC Tract 168. Centr. Hath. Comput. Sci.,
,
1

+ 313 ~, J.

Number

Amsterdam. (1.1) GyBry, K., Stewart, C.L. and Tijdeman. R. [1987]. S-unit equations and their applications, Proceedings of the Durham Symposium on Transcendental Number Theory, July 1986, to appear. (1.1)

Pacific J. Hath. 101, 263-301. (6.1) Brent, R.P. [1978], A Fortran multiple-precision arithmetic package, ACH Trans. Math. Software 4, 57-70; and: Algorithm 524. MP. a Fortran
pultiplz-precisioyarithme t ic package,_@ Trans. Hath . Sof m a r e 4, 71-81. (2.5)
4

Evertse, J.-H.,

__

--

Faltlngs, G. 119831, Endlichkeitssdltze far abe-lsche V d e t s t e a aber Zahlkarpern, Invent. Hath. 73, 349-366. (1.1)

-.

BrentJes, A.J. (19811. HuLti-dimensional Continued Fraction Algorithms, . MC Tract 145. Centr. Math. Comput. Sci.. Amsterdam. (1.3;3.4) Brown, E. [1985]. Sets in which xy + k is always a square, Uath. Comp. 45.

Fincke. U. and Pohst..M. [198S], Improved methodsfor calculating vectors of

'
I

short length in a lattice, including a complexity analysis, Math. Comp. 44, 463471. (3.6) *Grinstead. C.M. [1978]. On a method of solving a class of diophantine equations, Hath. Comp. 32, 936-940. (1.1) Hardy, G.H. and Wright, E.M. [1979], An Introduction to the Theory of Numbers, (5th ed.), Oxford University Press. Oxford. (1.3;3.2) Hasse. H. [1966]. b e r eine diophantische Gleichung von Ramanujan-Nagell und ihre Verallgemeinerung, Nagoya Hath. J. 27, 77-102. (4.1) Hunt, D.C. and van der Poorten, A.J., Solving diophantine equations u x2 + d a unpublished. (3.2i4.1)

B u c b n n , J. [1986]. A generalization of Voronoi's unit algorithm I & 11. J. Number Theory 20, 177-191 & 192-209. (8.2;a.S) Buchmann, J. (1987). The generalized Voronoi algorithm in totally real algebraic number fields, to appear. (8.2) Cassels, J.W.S. [1957]. An Introduction to Diophantine Approximation, Cambridge University Press. Cambridge. (1.3)

- .

Kiss, P. (19791, Zero.terms in second order linear recurrences. Math. Sem. Notes Kobe Univ. (Japan) 7 , 145-152. (4.3) Knuth, D.E. (19811. The Art of Computer Prograuming, Vol. 2: Seminumerical Algorithms. (2nd ed.), Addison-Wesley. Reading Mass. (2.5) Koblitz. N. (19771, p-adic Numbers, p-adic Analysis, and Zeta-functions. Springer Verlag, New York. (2.3) Koblitz. N. [1980], p a d i c Analysis: a Short Course on Recent Work, Cambridge University Press, Cambridge. (2.3) Koksma. J.F. [1937]. Diophantische Approximationen. Ergebnisse der Mathematik und ihrer Crenzgebiete, Vol. 4, Springer Verlag, pp. 407-571. (1.3) Lagarias. J.C. and Odlyzko. A.M. (19851, Solving low-density subset sum problems, J. Assoc. Comput. Mach. 32, 229-246. (3.4) Langevin. H. [1976]. Quelques applications de noweaux rdsultats de van der Poorten, Sbm. Delange-Pisot-Poitou 1975/76, Paris, Exp. G12. 11 pp. (1.1) Lehmer, D.H. (4.9;S.l) Lenstra. A.K. (19841, Polynomial-time Algorithms for the Factorization of Polynomials, Dissertation. University of Amsterdam. (3.5) Lenstra, A.K., Lenstra, H.W. Jr. and Lovlsz, L. [1982], Factoring polynomials with rational coefficients. Math. Ann. 261, 515-534. (1.4; 3.4;3.5) Loxton. J.H.,
-

Hignotte, M. [1984a]. On the automatic resolution of certain diophantine equations, Proceedings of EUROSM 84, Lecture Notes in Comput. Sci. 174, Springer Verlag. Berlin, pp. 378-38s. (4.1) b Mignotte, M. (1984 I, Une nowelle rbsolution de l'kquation x2 + 7

- 2" .

Rend. Sem. Fac. Sci. Univ. Cagliari 54, Fasc. 2. 41-43. (4.1) 2 Mignotte, M. (19851, P ( x +1) 2 17 si x r 240 , C. R. Acad. Sci. Paris 301, Sbrie I, No. 13, 661-664. (4.9) Mignotte, M. and Waldschmidt, H. (19781. Linear forms in two logarithms and Schneider's method. Math. Ann 231. 241-267. (2.4) Nagell. T. [1948], Lesning Oppg. 2. 1943. s.29 (Norwegian), Norsk Mat. Tidsskr. 30, 62-64. (1.1;4.1;4.9) Odlyzko. A.M. and te Riele, H.J.J. (19851. Disproof of the Mertens conjecture, J. reine angew. Math. 357, 138-160. (3.4) Peth6, A. (19831, Full cubes in the Fibonacci sequence. Publ. Math. Debrecen 30, 117-127. (1.1;8.1). Petha, A. (19851, On the solution of the diophantine equation Gn Sci. 204, Springer Verlag. Berlin, pp. 503-512. (4.1;4.7) Petha, A. and Schulenberg, R. [1987], Effektives E s e n von Thue Gleichungen. Publ. Math. Debrecen, to appear. (3.8;a.l) Peth6, A. and de Weger. B.H.H. [1986], Products of prime powers in binary recurrence sequences I: The hyperbolic case, with an application to the

(1964). On a problem of Starmer. Illinois J. Math. 8. 57-79.

- .
pz

Proceedings EUROCAL ' 8 5 . Linz. Austria, Vol 2. Lecture Notes in Comput.

Hignotte. H.. van der Poorten. A.J. and Wa_ldschmidt, M, [1987],

A lover bound for linear forms in the logarithms

of algebraic

generalized Ramanujan-Nagell equation. Math. Comp. 47, 713-727.; (1.1; . - 2.2;4) Philippon, P. and Waldschmidt, H. (19871, Lower bounds for linear forms in logarithms, preprint. (2.4)

numbers,

C.R. Math. rep. Acad. Sci. Canada 11, 119-124. (2.4) Lutz. $. [1951]. Sur les approximations diophantie~eslindaires P-adiques.
-

These. UniversitC de Strasbourg. - -. . . (1.3) MacUilliams. F.J. and Sloane, N.J.A. [1977], The Theory of Error-Correcting Codes. North-Holland, Amsterdam. (4.1) Mahler, K . [1933]. Zur Approximation algebraischer Zahlen I : iiber den gr6ssten Primteiler binErer Formen. Math. Ann. 107, 691-730. (6.1;8.6)

Pinch, R. (19-87h&imultaneous Cambridge. (1.1)

Pellian equations, prxiprint. University o f - -

Pohst, M. and Zassenhaus, H. (19821, On effective computation of fundamental units I & 11, Math. Comp. 38, 275-291 & 293-329. (8.2) van der Poorten, A.J. (19771, Linear forms in logarithms in the p a d i c case, Transcendence Theory: Advances and Applications, Academic press, London, pp. 29-57. (1.2;2.4;6.2) Rumsey, H. Jr. and Posner, E.C. (1964). On a class of exponential equations, Proc. Am. Math. SOC. 15, 974-978. (4.9;6.1) Schinzel. A. (19671, On two theorens of Gelfond and some of their applications, Acta k i t h . 13. 177-236. (2.4;4. 1)

Xahler. K. [1934]. Eine arithmetische Eigenschaft der rekurrierenden Reihen. Mathematika B (Leiden) 3, 153-156. (4.1) Mahler,l(. (19351, h e r den gr6ssten Primteilar spezieller Polynome zweiten Grades. Arch. Math. Naturvid. B 41, 3-26. (4.9;S.l) Mahler, K . (19611. Lectures on Diophancine Approxfmutions I: g-adic Numbers and Roth's Theorem, University of Notre Dame Press, Notre Dame. (3.10)

Shorey. T.N.. van der Poorten, A . J . .

Tijdeman, R. and Szhinzel. A. [1977].

Tzanakis, N. and de Weger, B.H.H.

[1987], On the practical solution of the

Applications of the Gelefond-Baker method to diophantine equations. Transcendence Theory: Advances and Applications, Academic Press. London, pp. 59-77. (1.1) Shorey, T.N. and Tijdeman, R. [1986]. Exponential Diophantine Equations, Cambridge University Press, Cambridge. (1.1;1.2;4.2;5.1;6.1;8.1;886) v Sprindzuk. V.G. [1969], Effective estimates in 'ternary' exponential diophantine equations (Russian), Dokl. Akad. Nauk BSSR 12. 293-297. (6.1) Steiner, R.P. [1977]. A theorem on the Syracuse problem. Proc. Seventh Hanicoba Conf. Numer. Hath. Comp., pp. 553-559. (3.2) Steiner. R.P. (19861, On Hordell's equation y2

Thue equation, Hemorandum No. 668, Faculty of Applied Mathematics, University of Twente, to appear. (1.1;8) Tzanakis, N. and Wolfskill, J. [1986], On the diophantine equation

- 4qn + 4q + 1 . Number Theory 23. 219-237. (4.1) Tzanakis, N. and Wolfskill, J. [1987], The diophantine equation x2 - 4qaI2 + 4q + 1 with an application in coding theory. to appear.
y2
3.

(4.1) Vojta, P. (19871, Diophantine Approximations and Value Distribution Theory, Lecture Notes in Mathematics 1239. Springer Verlag. Berlin. (6.6) Wagstaff, S.S. Jr. [1979], Solution of Nathanson's exponential congruence. Hath. Comp. 33. 1097-1100. (3.9) Wagstaff, S.S. Jr. [1981]. The computational complexity of solving exponential congruences, Congressus Numerantium. University of Winnipeg. Canada. Vol. 31. pp. 275-286. (3.9) Waldschmidt, H. [1980]. A lower bound for linear forms in logarithms. Acta Arith. 37, 257-283. (2.4) de Weger, B.H.H. [1986a], Approximation lattices of p-adic numbers, J. Number Theory 24, 70-88. (1.4;3.10) b de Weger. B.H.H. [I986 1, Products of prime powers in binary recurrence sequences 11: The elliptic case, with an application to a mixed quadratic-exponential equation. Math. Comp. 47, 729-739. (1.1;4) de Weger , B.H.H. [ 1 9 8 7 ~ 1 , Solving exponential diophantine equations using lattice basis reduction-algorithms. J. Number Theory 26. 325-367. (1.1; 5;6)
de Weger. B.H.H.
_-

-k

x3 : a problem of

Stolarsky, Hath Comp. 46, 703-714. (3.3;g.l) Stewart, C.L. and Tijdeman, R. [1986], On the Oesterld-Hasser conjecture, Honatsh. Hath. 102, 251-257. (6.6) Sterwr. C. [1897]. Quelques thLor&mes sur l'kquation de Pel1 No. 2. 48 pp. (6.9;S.l) Stroeker, R.J. and Tijdeman, R. [1982], Diophantine equations. (with an Appendix by P.L. Cijsouw, A. Korlaar and R. Tijdeman). Computational Hechods in Number Theory. HC Tract 155. Centr. Hath. Comp. Sci.. Amsterdam, pp. 321-369. (1.1;3.2;5.1;5.4) Thue. A. [1909], lher Anaherungswerten algebraischer Zahlen. J. reine angev. Hath. i35. 284-305. (8.1) Tijdeman, R. [1973]. On integers with many small prime factors. Compositio Hath. 26, 319-330. (5.1) the equation of Catalan, Acca Arith, 29, 197-209. Tijdeman. R. [1976], On . ----(1.1) Tijdeman, R. [1985], On the Fermat-Catalan equation. Jahresber,Deucsche
)

- D~~

- fl

et leurs applications. Vid. Skr. I Hath. Natur. K1. (Christiana), 1897,

b [I987 1. On the practical solution of Thue-Hahler equations, Faculty of -ApplieQU a t h e ~ ~ t i c sUniversity ,

an_ou_tline,He-rand-umA649, of Twente. (1.1;8.6) Yu, K.R.

[1987~], Linear forms in the p-adic logarithms, Report UPI/87-20.

Mach. Verein. 87, 1-18. (1.1) Tijdeman, R. and Wang. L. [1987). Sums of products of powers of given prime numbers. Preprinc, University of Leiden. (1.1;5.4) Tzanakis, N. [1983]. On the diophantine equation y2 Theory 17. 144-164. (4.1) Tzanakis. N. [1984]. The complete solution in integers of

Max Planck Institut M r Mathematik, Bonn. (1.2;2.4;6.2) b Yu, K.R. [I987 1, Linear forms in logarithms in the p-adic case, Proceedings of the Durham Symposium on Transcendental Number Theory, July 1986. to appear. (2.4)

- .
2k

J. Number 2y3

x3 +

- .
2"

J. Number Theory 19, 203-208. (8.6)

Tzanakis, N. [1987]. On the practical solution of the Thue equation, an outline. Proceedings Colloquium on Number Theory, Budapest, July 1987. to appear. (8)