You are on page 1of 28

MAHLER, K.

Math. Annalen 168, 200---227 (1967)

Applications of Some Formulae by Hermite


to the Approximation of Exponentials and Logarithms
To C. L. SIEGELon his 70th birthday

K. MAHLER

While LIOUVILLE gave the first examples of transcendental numbers, the


modern theory of proofs of transcendency started with Hermite's beautiful
paper "Sur la fonction exponentielle" (HERMITE, 1873). In this paper, for a
given system of distinct complex numbers e~o, co1..... oJm and of positive
integers Qo, Q1, .--, Qm with the sum a, H ~ I T ~ constructed a set of m + 1
polynomials
~o(Z), 9a~(z).... ,gain(z)
of degrees not exceeding a - Qo, a - Q~. . . . . a - era, respectively, such that all
the functions
9.Ik(z) e '~' z _ 9~l(z) eO,k~ (0 < k < l < m)
vanish at z = 0 at least to the order a + 1. O n putting z = 1, these formulae
produce simultaneous rational approximations of the numbers 1, e, e 2. . . . . e m
that are so good that they imply the linear independence of these numbers and
hence the transcendency of e.
In a later paper (HERMITE, 1893), Hermite introduced a second system of
polynomials
Ao(z), Adz) ..... A.(z)
of degrees at most eo - 1, Q, - 1, ..., Qm- 1, respectively, for which the sum

A k ( z ) e °'~"
k=O

vanishes at z = 0 at least to the order a - 1. O n putting again z -- 1, one obtains


now a linear form
a o + a t e + "'" + a , , , e "
of small absolute value and with small integral coefficients, from which again
the transcendency of e m a y be deduced. Surprisingly, HERMrrE himself never
t o o k this step, and I was seemingly the first to use the polynomials A k ( z ) for
this purpose (MAHLER, 1931).
In the present paper I once m o r e wish to exhibit the usefulness of Hermite's
polynomials A k ( z ) for the study of transcendental numbers. I shall prove a
n u m b e r of explicit estimates, free from any unknown constants, for the simul-
taneous rational approximations of powers of e or of the natural logarithms
of sets of rational numbers.
Some Formulae by Hermite 201

1. Let coo, col..... corn,~ be m + 2 integers satisfying


O=coo<col <CO2 < "'" < c o , . = ~2,
and let

M k = ]-[(cok--coZ), M= lcm Mk, N= lcm (cok--col),


l~--O k=O,l,...,m l#k
l~k k,l=O, I , . . . , , .
where Icm denotes the least common multiple. Let z be any complex number,
0 a positive integer, and
10 if h = k ,
~hk= if h~=k,
the Kronecker sign. Denote by C O and Coo two circles in the complex ~-plar~e,
both with centres at ~ = 0, and of radii less than 1, and greater than I2, respective-
ly. Then put

1 ,. e ~ad~ , Rh(z ) = ~ cot) °+6h'


Ahk(Z)= ~ H (8 + COk_ co,)o+ah, l~I(8 -
Co 1 = 0 Coo t=O

These definitions imply (see, e.g. MAHLER, 1931) that Ahk(Z ) is a poly-
nomial in z at most of degree 0; that

Rh(z) = ~ Ahk(Z) r °'k~ (h = O, 1,..., m) ,


k=O

and that the determinant


Ao.o(Z) . . . . . A o " ( z )
D(z) = • • = Cz(,.+ I)o,
Ih~'o(Z)..... A , . , ( z )
where C 4:0 does not depend on z.
2. By the paper quoted, Rh(z) may also be written as
I tl tm - I
Rh(Z ) = z(m+ 1)0 S d t l ~ d t , ... ~ dt,.@(t) e "~'(0,
o o o

where the expressions ~P and ~' are defined by


(1 - t~) o+~'°-'(t~ - re) ~+~''- ~ ... (t._, - tm)O + dlh,ra- 1-- 1 it): 6 h m - I
• (t) = m
l-I(0 + ~ h t - I)!
I=0
and
~(t) = COo(1- tl) + col(t1 - t2) + "'" + co,._ 1(tin- 1 - t,.) + comtm,
respectively. Here the quantities
1 - tl, tl - t2, ..., trn- 1 -- t", tm
202 K . MAHLER :

are non-negative and have the s u m 1. Therefore, by the t h e o r e m on the arith-


metic and geometric means,
0=< (1 -- tt) (t 1 -- t2) ... (tm-a -- tm)tm<=(m+1) -(m+ 1) ,
so that
0 < ~(t) ~ (m + 1) -(m+ 1)(0 !(0 - 1)!m)- 1
Further
o=< ~(t)__< ~
and
1 tl tin- 1

Idt~fdt2... Idt,= m!
1
0 0 0

It follows then from the first m e a n value t h e o r e m that


izl(-+,o eal~l
IRn(z)l < m[ (m + 1) (m+ t)<0- x)O! (O - 1)! m

3. F r o m the integral, Ahk(Z) is the p o l y n o m i a l


¢ gJ
&k(z) = j •~= ° A h(s)
k ~.

where the general coefficient "~hka(J)is given by

at1)_ 1 I m ffd3 q = o , 1, ..., o).


-~hk-- ~ 1-I(3+°k-°~t) °+~'
Co / = O

If we choose for Co the circle


1
131 m + 1 '
then on this circle,
m
1 + ~______~__ > 1 _ 1 3 1 = 1
o94 - oh - -
1
m + 1
= - -
m+l
for k # l.
T h e f o r m u l a for AhOk
) m a y also be written as

A02 = f i (Ok -- cot)- Q- 0h,.


1=0 2hi " 0+~, "
l*k
Co 1 = 0 O k - - (DI /
l#k
It follows therefore that
1 27r
[A°~I~-Mk~ 2r~ m+l
Some Formulae by Hermite 203

and so, by 0 < 6hk _<--1, that


Ao) < M ; O m - m Q ( m + 1)(m+l)Q
hk ~

4. F r o m the original integral,

Ahk(Z) = l
0o( k--O~t+ -~Z /
l*k
(e+6h~- 1)!

This formula may also be written as


ra "-¢-a"'fi(1 1 df °-."' z *+'h~-'
Ahk(Z) = 1-I (Ok -- Ol) " +
t=o t=o o k - o I dz (q q- 5hk-- 1)! '
l#k l*k
or, what is the same,

- e - 6ht ( o k - o z ) - ~
Ahk(Z) = f i (O)k -- 0 l ) - ° - 6 ~ ' "l ~" (e + 6hk -- 1)! "
/=o
l*k l#k

Here the binomial coefficients are integers; the differences o k - o~ are divisors
of N ; and hence the operator has the form

fi
t=otz=o z ~ dz,~
l#k
where
go, gl, g2, .'. (go = 1)
are certain integers that also depend on h and k. It follows that
m O+&tk-- i Zq+6hk--).-- 1
Ahk(Z)= 1-I(~k--Ol) -°+~"'" Z gz N - ~
l=o z=o (q + dih,-- 2 - 1)! "
l*k

Here, from the definitions of M and N, the factor


m

MQ N "1-[ (°k - °l) -°-~h~


1=0
/*k

is an integer. Therefore the product


Q

ahk(z) _---M q N q + 1 Q!Ahk(Z) , = E t*hk~


MJ).,J
j=O

say, is a polynomial in z with integral coefficients ahk-


o)
Since
0 zJ
ahk(Z) = MQN~+ 10[ ~ '~(J)- -
j~=o - h ~ j ! '

these integral coefficients can be written in the form

aU) _ Axe ~re + 1 ,~ t Ah°)k


hk -- x~.* x • ~ •
204 K. MAHLER:

and so satisfy the inequality


M~NQ+ 1O! (m + 1)tin+ I~Q
la~l <-
M~ msQ
It is further obvious that
MQNQ+ 1 Q! (m + 1)(m+I~Qe N
lahk(z)l < M~ m rn~
because
~ tzlj <elzl
j=07
In analogy to ahk(Z) put also
rh(Z ) = M~N¢+ I Q! Rh(Z ) (h = O, 1..... m).
Then

rh(z) = ~ ahk(Z) e °'kz (h = O, 1.... , m).


k=0
From the identity for D(z), the new determinant
aoo(Z), ..., ao.m(z)
d(z) . . . . cz~,+ 1~o
I%o(Z) .... , a,,(z)[
where again c ~: 0 is independent of z.
We note that, by the estimate for Rh(z),
MQNQ+ 1 izl~m+ 1)Qeal~l
Irh(z)] < m! (m+ 1)~m+l)~Q-1)(e- 1)! m "
5. The inequalities just proved can be simplified by means of some simple
lower and upper bounds for Mk, M, and N.
First, the factors of Mk are integers distinct from one another and from
zero, and of these factors k are positive and m - k are negative. It follows there-
fore at once that
M~>k[(m-k)!=m! (7) >2-ram!

Secondly, N is the least common multiple of certain positive integers not


greater than [2, and hence
N < lcm(1, 2, ..., [2) ~ e 1"°4a
where the numerical inequality is taken from the paper ( R o s s ~ and
SCHOEa,~FELD,1962).
Thirdly, an upper bound for M may be obtained by the following method
due to B. H. NEUMANN.
For each suffix k and for each prime p let #~(p) denote the largest integer
for which
P~k~P)IM~.
Some Formulae by Hermite 205

Heno5
Mk = l-If2'k~p~ .
P
Since [o9k - oht __6<~, a power ff of p cannot be a divisor of some factor Ok -- W~
of M k unless
ff=<~ and therefore p~.
The largest possible value of t is then
[log 7
= L-i-0~] '
because 2~+ 1 > ~.
One counts as usual how many of the factors
~ok - o h, where 0_<l_<m, 14=k,
are successively divisible by pl, by p2, by pa, etc., and finally by p'; the sum of
all these numbers is equal to Pk(P). Now M k has just m factors a~k - wl, and so
none of these numbers can exceed m. Also these factors of M k lie in the interval
from Wk-- ~ to Wk of length ~, and this interval contains the multiple 0 of pt
which is not a factor of M k. Therefore at most

min (m' [ - ~ t ] )

factors of M k are divisible by if, whence

Pk(P)<--~.min(m[--~,])
--t=l \ ~ "
We replace this inequality by the weaker but more convenient one,

t~,(P)--<min('n,~~]~ + }] say.
\ LPJ/ ,=2LPJ
Let
M* = IJ ?~'.
p6~
Then all products Mk and so also their least common multiple M are divisors of
M*, and hence it follows that
M~M*.
6. Put now

i [7]'
t=!
so that, by a well known formula,
t2! = I I p,tp~.
pSD
It follows that
t2!
M* =
A
206 K. MAHLER:

where A denotes the product


A = [-I pV<p)-#~p).
p_-<~
From the definitions of #(p) and v(p),
f2
-m if p = < - - ,
m

g2
if p > - - ,
m
so that

A = H p[~]-m
<f/
P=~-
and therefore also

logA > ~ (f2 logp - (m + 1) logp).


v<~_\ P

In the paper (RoSSER and SCHOENFELD, 1962), it is proved that


~, logp 1
-- > logx + E for x > l
p<_x P 21ogx
and
logp < 1.02x for x_>l,
p<x
where E is a certain constant satisfying
E > - 1.34.
Assume for the moment that
I2>=e2m>m,
and therefore
1
2log -~-~ >4, < 0.25
m t2
2 log - -
m
while trivially
m+l
--<2.
m
It follows then that

logA > f2Iog-~- - 1.34 1 .~ (m + 1) 1.02


2 log -~- /
Some Formulae by Hermite 207

or

logA > f 2 ~ o g - ~ - - ( 1 . 3 4 + --21og._~_


Q 1 + m+im 1"02)t'
m
and here
1 + -m-m
+l 11
1.34 + 1.02< 1.34 + 0.25 + 2.04 = 3.63 < 3
f2
21og-
re
Hence, finally,

3)'
that is,

This inequality trivially is valid also for


[2 < e2rn,

because then
A ~ I >e(2-~--)a> ( ~ - ) ° e-110
3 .

7. Thus it has been proved that always


f2, (_~_)-t~ 11
M <_M* <__--A- < f2! e ~ - t~

Here
f2[ < e V-~-s'2~ e -~
and therefore
8
M < e V-~-m ~ e -y o
But O > 1, hence
e l / ~ = e 1 +~o~ (1 + (~- 1)) < e½(~ + t)+ ½(a- 1) = e a ,

and so finally
11
M < m a e -T-~.
On combining this inequality with the earlier one for N,

MQNe+ 1 <=MQN2~ < (m ~ el---~o)Q (el.O4O)ze


and hence
MQNO+ ~ < m t ~ e6~e.
208 K. MAaLER:

8. F o r the moment put


MONO+ 10 ! (m + l) (m+ 1)~ MQNa+ 1
a= M[mm Q , r= m!(m+ 1)('*+l)(Q-l)(0-- 1)!m '
by what has been proved in § 4,
maxla~l-< a, maxlah~(z)[ < ae I't, maxlrh(z)lh < rtzl('+l)°e~l~l"

Thus upper bounds for a and r imply upper bounds for .~(i)
'*hk, tahk(z)l, and
Irh(z)l. Such upper bounds are obtained as follows.
To begin with a, we apply in addition to
Mk > 2-mm! and MeN Q+l <m#Qe 6~Q
the formulae
2V-~oQe-O<ot<eV-ooOe -Q, m!>~mme -m.
We find then that
ma° e 6c~°- e V f f o ° e - ° . (m + 1) ~m+ 1)Q
a< (2- "" ~ ms e-")° mmQ
__( e2o ~l/2( 2mem-l(m+ l) m+l e6a) °
-- \ ~-~-~-~/ \ m2m+½ "Q m° .
Here
e2 < 7.5, 2n>6
and hence
e20 7.50 7.50 3
-(2r0
- ~ < -(1 -+ 5=) ° < - 1 +- 5 o < Y < 4 .
Further the function
2m era- l ( m + 1)m+l
/,n2m + ½

of m assumes it maximum when m = 2, and this maximum has the value


27e
- - <13.

The final result is therefore


a < 2(13om~e6t~ t
and it follows that

f maxla~lh
~,~ < 2(13Qm~e6°)~' maxlah~'(z)ls,
k < 2(13Qm%6a)~el:l" 1

. Since
1 2/T~ . _,
(o-1)!> V To ,
Some Formulae by Hermite 209

we similarly find that


mOOe6OO
r<
V.~mmme_m.(m_t_l)(m+l)(e_l) 21/~Oee -°)"
ttl
= em Q2 mmO (mOe6ae2~)o
m+ 1 \ m m 0 rn "
(270 2 "m'(m + 1)tin+ 1){o- 1). emQ
Here m > 1 and O > 1. Further

mm(m + l)(m+l)(e-1)=mSO+O-l(l + __l)(m+l)(o-1)>mmo+o-l eO-t >mmo

because
( 1 + 1--)m+l > e ;

and also
m+l
(2r0 2 > 1 .
It follows that
m

e m Q2 mraO
m+l
< <1
(2re) 2 m"(m+ 1)('+a)(°-l)e ''°
since
eo-1 > 1 + ( 0 - 1)= O> Q½.
The final result is then that
( m~ e6O e2m_']o ( m° eS"'] Q "

here we have used that


re<t2 and hence e'~<e ° .
Thus it has been established that

mo e8a)o
maxlrdz)l< m,•0•, Izl(m+l)°em~l

9. As a first application, denote by r9 a positive integer and put


1
(D

Let further q >=1, qt, q2, ..., qm be m + 1 arbitrary integers, and let

k=l,2 ..... m

14 Math. Ann. 168


210 K, MAHLER:

and

e,k=2mq(e-~---
-~) (k = 0 , 1. . . . . m),

where we have put


qo=q>l

Since COo= 0, trivially


no=O,
and hence
e= max lekl = max I~kl.
k = O, 1, , . . , m k = 1,2,...,m

The powers
tO 1 tO2 tom

e °', e'~, ..., e ~'


are irrational numbers, and hence
e>O.
We shall now establish a positive lower estimate for e.
F o r this purpose we note that the (m + 1) 2 numbers
Q 1
(h, k = O, 1, ..., m)

are integers, with the determinant

[Aoo, ..-, A o , l
. : =~

O n putting

coarh ( 1 ) = Rh (h=O, 1. . . . , m ) ,

we have

Rh = ~ Ahk e " (h = O, 1, ..., m).


k=O

T h e estimates in ~ 7 - - 8 now take the form


I
max lAbel < 2(130com%6°)~e ~
h,/t

and

max]R,]<(m~e'° "~e~-.
h \ mU 0" co,, I
Some Formulae by Hermite 211

10. Since the determinant of the integers Ahk is distinct f r o m zero, a n d


since the integers
qo > 1, ql . . . . ,qm
do not all vanish, there exists a suffix h such that
m
~, Ahk qk # 0
k=0

and that therefore


-
k~_oAhkqk ~ 1.
j
W i t h this value o f h, put
1 m m

Q= q k~=OAhkqk' E -2mq
- k~_lAhk£k.
F r o m the definition o f ek,
m to_.~k m £ x

Rh-- EAhke°--- EA k( + =e+E.


k=o k=O \ q
Here
1
IQI => - -
q
and
8
[El < - - m a x [Ahk[ •
= 2q *,k
Assume n o w that
1
m a x [Rh[ < - - .
h = 2q
It follows then that
1
IEI >
= 2q
and hence that
e max [Ahk[>=1.
h,k
Thus the following result is obtained.

If
maeSa ~, o l

then
1
> {2e~(13otomUe6r~}- 1.
14"
212 K . MAHLER:

This result can be slightly simplified. Since all three integers co, f2, 0 are at
least 1,
f~
2e o, < 2 e ~ < e2O< e2t~Q,
so that

2( mrn 0 rn COrn ) e ~ m m Om corn ,] "

Further
1
__ > {4e~ m(13ocomOe6O)Q}- 1q- 1 > (52e¢com~+ 1e6e)-O q- 1
2mq
Here
52 < e 4 ,
and so
_ _ > (QcomO+ 1 e6~+ 5)-Qq- 1.
2mq
Thus the following result holds.
Lemma t. If Q is chosen such that
( rn____
a e~°°_~ ~Q< 1__
mm Qrnoom/ = q '
then
max e~----~-I2>(otomST+le6ST+5)-Oq-1.
k = l , 2 ..... ml

11. When applying this lemma, one naturally will choose the integer ¢
as small as possible because this improves the estimate. It is now convenient
to distinguish between the two cases ¢ = 1 and Q > 1.
The case 0 = 1 holds exactly when
1
09~ (m~-rnel°a q)~,
and then, by the lemma,

max e~ - ~ t/
>(coma+le6a+5)-lq -I
k= 1,2 ..... m I

Next, excluding this case, let


1
co < (m°-m elOa q)~,
so that the smallest possible value for Q is at least 2. This value Q satisfies the
inequality
ma e 1°~ \ ~ < 1 < ( m~ e l o a ~0-i
( mm cm co" ) = q m ~ Z l-~-w.i/
Some F o r m u l a e by Hermite 213

It follows that
t 1
2
eco ~ 2(e -- 1)co < (m~ el°a) m q,(e- i),
m

and that therefore


t 1
2
Qo)m~+1 e 6~+5 < - - (m~ elO~)~ ran+ 1 e6O+ 5 qm(O- 1) <
m
1 1
< (m~ eXOO)~mt~e6r~+ 6 q m ( e - 1).

The lemma implies then in this case that

max e.,~ - -~- m+ 10q-1 o


k= 1,2 ..... m I > e(4~-6)e(m ~etot~) m rote- 1)

Here we once more use that Q > 2, hence that


0 1 1 1 2
q- - 1 - -m(e-1)=q
-- --1 m mto-1)>=q- - t ,,,.q me,

where, by the choice of Q,


2 (m~e,O~2
q ~n'°>\mmpmo)m/
Evidently/2 > m, and so, by this inequality,
2 e2O
q '~Q-> Q2o92 "
Assume, in particular, that also 12 > 2. Then
2 e20+20 e20
412-6>2, e(4U-6)eq me>_ Q2co____._____~>_
~ because eQ>Q.

Thus, in this second case, we arrive at the estimate


eO@ - ~ e2o m+l 1
__(nl~elOD] m o --1-~
max - > ~°2 x , q
k= t,2,...,m

Our result may be expressed as follows.


Theorem t. Let o9, 091. . . . . e)~, q, ql . . . . . qm, and 12 be 2m + 3 integers satis-
fying the conditions
oJ -> l , q~l, 0<e)i<co2<'"<Ogm=12, 12->2.
If
1
O~>. (m O-m 2 00 q)m,
then
max e~ _ q~ > (o~m~+ 1 e6t7+ 5 ) - 1 2 .
k=l,2 ..... m I q
214 K. M A n t £ g :

If, however,
!

co < (m
and if {? denotes the smallest integer satisfying
. m° e~°° )°<= 1

then
max , ~ _ qk > e2O (mUelOO) ,,+1
m ¢ q - 1 - ~1
k=I,2 .....
role---q- --~ .

T h e interest of this theorem lies in the fact that co, co~, ...,corn, q, q, ...,q~
m a y all be variable and are subject only to trivial restrictions. The assertion
is particularly strong when co, col, ..-, cos are fixed, while q, ql, ..., qm are
allowed to tend to infinity. F o r then the parameter 0 likewise tends to infinity
and is given asymptotically by
logq
Q~ loglogq "
Hence a positive constant c depending only on o9, col . . . . . corn exists so that

to__~ --1 ~n logToilq


max e o, _ > q
k= 1,2,...,m l
for large q.
If also co, co~. . . . , co, are variable, the theorem is much less strong. However,
some consequences seem still worth of being mentioned.
12. T h e o r e m I implies an analogous theorem on the simultaneous approxi-
mations of logarithms. Its p r o o f is based on the following elementary lemma.
Lemma 2. If x and y > 0 are real numbers such that
I x - logyJ < 1,
then
Ix - logyl ~ e - X - 2 je~ _ y}.
Proof. By the mean value theorem,
d-1
- - = e at where 0<0<1.
t
Hence, on putting t = x - logy,

0< e~--Y =yea¢~-t°sY)<-ey.


x -- l o g y
Here
logy-<x+l, y ~ e ~+1,
whence the assertion.
Some Formulae by Hermite 215

This lemma we apply to each of the m pairs of numbers


cok qk
x=--, y=-- ( k = 1, 2, .... m),
co q
for which, evidently,

X ~ - - Q <=gQ, e-X-2>--e - a - 2
CO
We next note that Theorem 1 remains valid if the conditions
0<O1<0)2< "-" <COrn= ~'~
are replaced by the weaker hypothesis that the integers col, ..., cos are all
distinct and have the maximum f2. By combining the theorem with the lcmma
we obtain therefore the following result.
Theorem 2. Let co, col .... , co=, q, ql ..... q,,, [2 be 2m + 3 positive integers
satisfying the conditions
COk~CO~ for k#l; ~2= max COk-->2.
k= 1,2,...,m
If co satisfies the inequality
1
CO> (ma-= e 1°~ q)m,
then

max log qk -- _COk[


_ > (COm~9+1 e7~+7 q ) - I
k=l,2 ..... = q
If, however,
1
co < (mt~-= e 1°° q)= ,

and if Q denotes the smallest integer satisfying


(mael°a ~'<1
m m Qmcon,/ q '

then
> e18-£~ m+l 1
max log q k cok ~ (m~ e xoo~ = . q t-~.
k= 1,2,...m q tO CO2 P

13. W e deal in detail with one special application of Theorem 2. For this
purpose denote by
P l = 2, P2 = 3,..., p .
the first m primes in their natural order. We apply the theorem with
q = 1, q l = P i , "",qm=Pm
and choose for co, col, ..., co= any m + 1 positive integers for which the fractions
216 K. MAHLER:

(2)1 ogra
--
"''7 are approximations of logpl, ..., logpm, respectively, that are
o9 o9
already so close that
_~_ 1 Pm
(A) max
k= 1,2,...,m 1ogpk -- < ~- log Pro- 1

Further put again


f~ = max(ogi, ..., ogre)
and assume that
m > 10.
From the hypothesis (A),

[1ogpk -- Iogpll > log Pm for k # l,


Pro- 1
and

logpk~log2>log Pm for all k.


P,n- 1
Hence

O)k+lo9 (Dk(D -- ( ' ( 0 ~ 1 lOgPk+l) + (logp~+ l -- lOgPk)+ QogPk-- --~-)


1 ~ Pm Pm 1 pm
> - T ~og ~ + log =0
Pro- 1 2 log Pro- 1
and
091 1 Pr, 1 3
--
co > l o g 2 - -~log Pro-1 • ~ l o g 2 - -~- log ~- > 0 .
The hypothesis (A) implies therefore that
0<o91 <092< -.. <og,n = ~ .
It also implies that
o9~2,
because, if co were equal to 1, it would follow that

logpl - ~llog2- 11> ~ - l o g ~ - > ~-log P , - t '

for all choices of the integer coI, contrary to (A).


Next we have co, ~ m ~ 10 and therefore
t2>2.
Thus all conditions of Theorem 2 are satisfied, and this theorem may be
applied.
Some Formulae by Hermite 217

From (A),

g2<co ogpm+ -~-log =-flog .

Here, by Bertrand's law on prime numbers,


1
P,,-1 > ~-Pm,

and by the paper (RossER and SCHOENFELD,1962),


p,, < V~m log m.
Therefore the quantity ~2 allows the upper estimate
g2 < co log(2m logm).
It follows that

(me-r" el°e) g < - - exp (10 + logm) • co log(2m logm) .


m

Here the right-hand side does not exceed 2 if


m log(2m)
(B) co _<
- (10 + logm) log(2m logm) '
and so, for such values of co, the second case 0 > 2 of Theorem 2 cannot hold.
Therefore, by this theorem,

max 1Ogpk ~-I > (COma+1 e7O+7)- 1


k = 1,2, ,..,m

In this estimate,
coma+ 1 eVC~+7 < eVmco exp {(7 + logm) • co log(2m logm)}
m log(2m) I m(7 + logm)log(2m) t
eTm (10 + logm) log(2m logm) exp ~ 10 + logm 3
m2
< e7~ exp{m log(2m)},

where, by m > 10,


m2 m2
e7 - < 2,000~--- <(2m) 5 .
logm
Hence it follows from (B) that

k=m2a,x.. logpk -- ---~I > (2m) - ' - s .


A stronger result is obtained if co is restricted to the smaller range
m
(c) co _<
- (7 + logm) log(2m logm)
218 K. MAHLER :

Now
m e 7 m 2 em
coma+ 1 e 7 ~ + 7 < e7m. . em <
(7 + logm) log(2m logm) (log m)2 '
where, by m_~ 10,
e7m2e m 2000m2 em
< < m s e".
(logm) 2 22

It follows thus from (C) that

k=max., logp.----~--l>m-Se-'.

The two right-hand sides


(2m) -m-5 and m - S e -m
in the estimates just established are smaller than the right-hand side

l log P,. = 2 say,


Pm-- J.
of the hypothesis (A). For
Pm< l/~-m logm < m2,

because
1 m
logm <log2 + -~-(m- 2) < ~ - .

Therefore
1 p,. 1 m2 1
2-_> -~log p - ~ f _1 > -~-log -m-~ l > -~- log(1 + m- 2),

where, by m_~ 10,

llog(l+m-2)> 1
~-(m- 2 - m - 4 _ m _ 6 . . . . ) > - ~1r n _ 2.

Hence
I
2 > --(m-2 > m -3 '
giving the assertion easily.
We may then omit again the hypothesis (A), and we are also allowed in
including the trivial denominator to = 1. Then, on combining the preceeding
results, we obtain the following theorem.
Theorem 3. Let m>=lO; let p1=2, p2=3, ...,Pro be the first m primes;
and let to, to! .... , to,, be m + 1 positive integers. Then
m log(2m)
~max
1,2.....,, logp~ - - ~ >(2m) -m-s /f l__<to< (lO+logra)log(2mlogm)'
Some Formulae by Hermite 219

and
m
max lOgpk r~k > m -5 e -m if 1<o9<
k= 1,2,...,m o) - - (7 + l o g m ) log(2m logm) '
These two inequalities are rather weak, but it does not seem to be easy to
obtain m u c h better ones. F o r larger values of co the position is worse.
14. Next put
o~1 = 1, 092 = 2, ..., co= = m, hence f2 = m.

"hk, ahk(Z), and rh(z) can in this special case be a little


The general estimates for -tY)
improved. F o r now evidently

Mk=k!(m-k)[=m! >2-ram!, M=m!,

and by the paper (Ross~R and SCHOENFELD, 1962),


N < e 1'°4m.
The formulae in § 4 become therefore
aO)l < 2,,,Q el.O4,,,{Q+ 1) Q ! m-,.Q(m + 1)~,,,+ x)Q,
h k l ---~

lahk(Z)l<2mQ el.04m(Q+1) Q! m-mQ(m+ 1)(,,+l)QelZ[,


Irh(z)l _-<(m !)~- 1 el.O4mtQ+ i)(m + 1)-(m+ *){~- 1){(Q _ 1)!}-=lz[ (m+ 1)Q eml~l.

These estimates can be further simplified if we assume from now on that m is


already sufficiently large, but that Q may be any positive integer, small or
large. F o r

0[<eV~-QQe -°, Q-~<3 -$,


(1).1+ _<e,

while
1
(m+l)m>l
becomes arbitrarily close to 1. Since 2e T M < e 1"74, it follows that
1 \too
"¢J) = < 2 " Q e l ° ' m ¢ ~ + l ) ' e V ~ o Q e - ° ( m + l ) ¢ l + m )
"h, -<

< - 1 0 4 0 ~ -1 1 1 1
= {2e " eme Q27~a(m + 1)~}m~ QQ< eL75.{Q+ 1) QQ
and hence

(1) [ la~t < e 1"7sm~¢+ 1) QQ, lahdZ)l < e l'Tsm{~+ 1) QQel~l.


[
15. Next, the estimate for rh(Z) m a y be written as

Irh(z)t _~ Rlzl ~m+'° e-I.I,


220 K. MAHLER:

where R denotes the expression


R=(m!)~ -1 eI.O4,~(e+l)(m+ 1)-(m+i)(Q- 1){(Q_ 1)!} -m
which does not depend on z. Since

mt<el/--mm"e-'~' ( e - 1 ) ! = >l / --d- Qe e-e , +-- >= e ,


we find that

R < ee- l m 2 mm(O-1)e-m(o-1).el.O4m(o+l)(m + 1)-(,.+l)(o-1). 2 O-raCeme.

Here

mm(o-t)(m+ l)-(m+l)(a-1)=m -(a-i)


(1+ ~_)-(m+ 1)(0- 1) <=m-~e-1) e-(e-1),
so that after a trivial simplification,
0-I
R N e(°- l)-m(o- 1)+ l.O,:l-m(o+ 1 ) + m o - ( O - I) m 2 2 Q-me<
- Q _m
<_ _ e m+l'°4m(°+i) m
o l( ~2O-mo

On omitting the factors that are smaller than 1,


R < e t'04rn(°+ 2) Q-m(o-$),

whence

(2) f irh(z)l < e l.o*,.(Q+ 2) 0-m~Q-~>tzt(,,+ 1)o e~tZt.

If also O is sufficiently large, this inequality can be further simplified to

(3) Ir~(z)t < e i'°Smo 0 -"~ Izl(m+ "Q e m l z l •

16. As a first application of the last estimates, let # be a very large positive
integer, and let V be the integer defined by
1 1
@=~+fi, where --- ~d~< + - - .
2 2
In the identity
rh(z)= ~ahk(z)e kz
k=O

substitute
z=g, e~=7+~.
Then
rh(g) = ahk(g)(7 + fi)k= ~ k ~ ah~(g)(kl) 7k-1 31,
k=O k=O 1=0
Some Formulae by Hermite 221

or, say,
m
rh(g) = ~ bhz jl
/=0
where bh~ denotes the expression

bht = ~m )ahk(g
(kl),k-,.
k=!
In particular,

bho = ~ ahk(#) yk.


k=0
Here, by § 4, the determinant d(g) with the elements ahk(g) does not vanish.
Hence a suffix h exists for which
bho ~ 0 .
Let h from now on be chosen in this manner,
17. Since
0
/,"*hk" ,
j=0
we have

k--t~=o hku \ 1 / ~ ,
so that bn~ is an integer. By the estimate (I),
tan~l < el.75m(e+ I) Oo.
Further
o /,.\ . .

~g j--<(g+l)°' ~ / ) N 2k,
E (k~< E 2k<2'~+'"
k=l\//=k=O
j=O
Hence, for all suffices l,
Ibhz[ < e 1"75m(~+I) 0o(g + t)o 2,.+ 1 y,..
On the other hand, bho is a non-vanishing integer, and hence
Ibhol > 1
Let us assume for the moment that
1 I 75re(Q+ 1) 1
ItSI < T {e • O~(g+ 1)Q2.+ x ~.}-
and hence that
1
161 < 5-"
222 K , MAHLER :

We find then that

Irh(o)l _~ Ibhot - 161 ~ Ibhtl l$l t- 1 >


1=1

>1- {eL75m~a+l)OO(o+l)e2"`+lym}-l" ei'75"~°+lto°(g+l)e2"`+~7 m.


/=l

Here

5-__E1 =
and so it follows that
1
Irh(0)t > y .
However, if both m and 0 are sufficiently large, then, by (3),
lrh(o)t < e 1"°sine 0 -'`~ gt'`+ ~)~ emg.
If now m and 0 are chosen so as to satisfy the inequality
1
(D) ei.O5"`o e-,,Q O(m+1)o e'`g < __
"--- 2 '
a contradiction arises. The assumed upper bound for 5 was therefore false,
and so (D) implies instead the lower bound
1 {el.75"`(o+1 ) 1)o2,,+1 y'`}-i
(E) 181 ~ T O°(O + "

Denote by ct and/8 two positive absolute constants to be selected imme-


diately, and take for m and 0 the integers
m=[0tlog0]+l, 0=[fl#]+l,
where, as usual, [x] is the integral part of x. Then m and Q will exceed any
given bounds as soon as g is sufficiently large, and so, under this hypothesis,
we were justified in applying the formula (3).
The inequality (D) is equivalent to
I g 1
0 > e 1"°5 0 m e Q2,,0.
Here, by our choice of m and 0,
m > ~ log0, O > fig,
and therefore
2 ± ~_ 1
ym < e ~ , eQ< e ~ .
The remaining factor
1
2"`o
is arbitrarily close to i as soon as O is sufficiently large. Thus, for such g,
1 g 1 1 I
e i.o5 O1÷ ,7, e [ 2,,-~ < e 1.°6+ ¥+'# g.
Some Formulae by Hermite 223

Assume now that


1 1
1.o6+-+---
(F) fl > e ~ ~.
The condition (D) is then satisfied because
I I I g I
Q > f i g ~ el.O6 + ~ - + ~ g > el.O s gl + m e ~ 2m~.
Also, for all sufficiently large g,
el.75rn(~+ l) <7,el.76=#alogo, Q# < (flg)l.oospg < el.Olpolof# '

(g + 1)~ < el.Olpg'o,,, ,m < (eO + .~ )~'°"+ ' < eZ.OZ=,lo,, '

3 x 2 m+ 1 < e=lOgg < eO.01=glogo.

The lower bound (E) for 6 takes therefore the form


161 > e - (1.76otflglogg+ 1.Ol #glogg+ 1.Ol#glogg+o.olotglogg+ 1.ol aglogg)

that is,
161 > g - (1.76aa+ 2.02/~ + 1.02at)g

We finally fix the constants ct and fl so that (F) is satisfied, while at the
same time the sum
tr = t.76~fl + 2.02fl + 1.02~
becomes small. After some numerical work one is led to the values
fl=7, 0~=1.35,
when
1 i
el'°6 + ¥ + # < e1'945 < 7, a < 32.4.

We arrive thus at the following result.


Theorem 4. Let g be any sufficiently large positive integer, and let 7 be the
integer closest to eg. Then
Ie ~ - ~'1 > g - 33g

By means of more careful estimates, the constant 33 in this theorem can be a


little decreased. However, it does not seem to be easy to obtain any essential
improvement of the theorem. Previously, by means of a different method, I had
proved the analogous estimate with 40 instead of 33 for the constant (MAHL~,
1953).
18. We finally apply the formulae (1) and (2) to the study of the rational
approximations of m
Denote by p and q any two positive integers such that
p 1 1
~t=--+6, where ----5<+--
q 2q - 2q
224 K . MAHLER :

It is trivial that there exists arbitrarily large integers of this kind.


In the identity

rh(z) = ~ ahk(Z)e kz
k=0

put now
rci
z= --2-, eZ=i'
so that

= oa ¢T) .
Here

ahk(n_~_~)= j ~2 0, a. /[i~\ j2 ) (P +~)J--~ ~=O(1)ahk(2)(q)


j • i j j-t 6'"
j=O

or, say,
°

ahk : E Chkl ~1
1=0

where
Chkl=~(~ ) / i \ J / p \ J - I
j=l~,/a (j) --
Therefore

rh = ~, Chklik ~t = ~ Chl ~l
k=O 1=0 1=0

where we have put

CM~ E Chklik"

In particular,
m Q - j j
Cho = ik= S~ ~,~a(hJ2(i~ (p~ ik= Em ahk( ip
i ~ ik.
k=oChko k ~ o ~ o " \ 2 7 \ q 7 k=O \ 2 q /

Here, similarly as before, the determinant d ip with the elements ahk


does not vanish. Therefore, from now on, we may again assume that h is
chosen so as to satisfy the inequality
Cho 4=0.
19. The expressions
Ql\j
(2q)' Chk,= ~, [t)aW 2Q-J q~-O-0 i'/r/-`
j=l~./
Some Formulae by Hermite 225

and

(2q) QCht = ~ (2q) QChkzik


k=0
are integers in the Gaussian field Q(i). In particular, (2q)QCho is a Gaussian
integer different from zero, and hence its absolute value is not less than 1.
Therefore
IChol ~ (2q) -Q.
Assume now that m is already very large, while no such restriction need
be imposed on Q. We are thus allowed to make use of the estimates (t) and (2).
Since
(~)____<2i, (0j-i)~l for l<=j<=o,
it follows from (1) that

IChktl<~=,(o--_ll)2'eL7'"Q+~'Q'(2)'(P)'-'.
Here
1
P-P-< n + 6 < r t + T <4 ,
q
and therefore
IChkt[ < e l . 7 5 m ( e + 1) Qe. 5 0 - t

and
m+ 1 el.75m(q+ 1) Qo. 5 e "
IChtl<
Let now, for the moment, 5 be so small that
161 < {(m+ 1)e l'vSm(°+ 1) e ° 5~(2q)Q}-1 ,
hence that
161 < 1.
Then

l rh(~) > [Chol--16[ ~ lCh,[[6l'- l >


/=1

>(2q)-Q-- {(m+ 1)el.7sm(o +1' oe(lOq)O}-I ~ m~---~l et'Ts"(°+l'e°5~,


/=1
and since
,=1~,3- ! - ~-,
it follows that

15 Math.Ann. 168
226 K. MAHLER:

On the other hand, by (2),

(-~-)1
rh \ z./, . / n ",('+ 1)~ e -~
<ei.O,tm(,+ 2)e-mto-½)~_~_) 2

Hence, if m and O satisfy the inequality


/ 7~ \tin+ I)# mlt 3 o
(G) e 1"04m(#+2) 0-=(*-½)~2 ) e 2 __<-4-(2q)- ,

a contradiction arises, showing that the assumed lower bound for 3 cannot be
valid. The inequality (G) implies therefore that, on the contrary,
(H) 161=> {(m + 1) e ' ' v s - ( ~ + i) O~(10q)Q}-,.
20. From now on let q be very large. If m is then defined by
m = [21ogq]+ 1
where 2 > 0 is a constant to be selected immediately, also m will be arbitrarily
large, as required in the preceeding proof.
The inequality (G) is equivalent to
/d.\ 1 . ^. O+2/_\(m+l)Q rt Q

Q> ~ 3 ) ~ eT M Q - - - ~ - ( 2 ) ~ e2d--, (2q),,(Q--½).

Here the factors

(X)' "(g-½) and 2 m(°-½)

are arbitrarily close to 1 and so may be assumed to have a product


4 1 o.o1~_~
(~_)-(Q--½) 2m(Qg--½)< e - .

We may similarly demand that


/ x \ (m+ l)#

Next
1 1
<
m 2 logq '

and hence
Q Q
qm(0-½) < e~te-½).
Thus (G) is for large q certainly satisfied if
l.os Q+.2/Tt \t.ol Q
0 ~ e ~-, ~.~_) ~-½ e2O: t e x(Q-½) ,
Some Formulae by Hermite 227

that is, if
Q--½
Q~e
Here
7[
~- < 1.571, log ~-- < 0.452, 1.01 log -~- < 0.457.

The condition for e is therefore satisfied if


(3,671+ 1,soT~+~)Q~
~e - ,
or equivalent to this, if

~ < Q- 1og0-(3.671+1.5070).
2 =
This inequality again is easily seen to hold if
0=14, 2=1.35.
On substituting the values
Q=14, m=[1.351ogq]+l
in (H), we finally obtain for large q the following result.
Theorem 5. I f p and q are positive integers and q is sufficiently large, then

~_P > q-45.

This result is not as strong as the estimate

P >q-42 for q>2

which I have previously obtained by means of a different method (MAHLER,


1953a).

References
CH. HERMITE1873, Oeuvres, t. III, 151--181.
CH. HERMIT~1893, Oeuvres, t. IV, 357--377.
K. MArlLER 1931, J. reine angew. Math., t66, 118--137.
K. MAHLER1953, Phil. Trans. Royal Soc., A, 245, 371--398.
K. MAHLER1953a, Proc. Acad. Amsterdam, A, 56, 30---42.
ROSSERand SCHO~NFELD1962, Illinois J. of Math., 6, 64--94.
K. M^HLEg
Institute of Advanced Studies
Australian National University
Canberra, A.C.T., Australia
(Received May 28, 1965)

15"

You might also like