§ 2. Eulerian expansions of the trigonometric functions and Bernoulli numbers
1. Eulerian expansion of cotz
By formula (20) of VI, p. 275, the numbers bn/n! are the coefficients in the expansion of S/(eS−1) as a formal series; we shall show in this section that the function z/(ez−1) is equal to the sum of an absolutely convergent entire series on a neighbourhood of 0 in C; it will follow from the lemma of VI, p. 280, that the coefficients of this series are the numbers bn/n!, from which we shall deduce estimates for the Bernoulli numbers bn.
In the first place we note that
ez−1z=−2z+2zez−1ez+1=−2z+2izcot2iz.
(1)
We shall obtain below a series expansion for cotz, valid for every z not an integer multiple of π.
Proposition 1
For every complex number* zand every integern *one has
It suffices to divide both sides of (2) by sinz and let z tend to 0.
Corollary 2
For every odd integer n=2m+1, and every complex number z such that nz is not an integral multiple of π, one has*
cotnz=(−1)mcotzcot(z+nπ)…cot(z+n(n−1)π).
(4)
Indeed, sinn(z+2π)=sin(nz+2π+mπ)=(−1)mcosnz, whence, replacing z by z+2π in (2),
cosnz=(−1)m2n−1coszcos(z+nπ)…cos(z+n(n−1)π)
(5)
and the formulae (2) and (5) give (4) on division term-by-term when sinnz=0.
In all that follows we shall always assume that n=2m+1 is an odd integer; the formula (4) can also be written
cotnz=(−1)mk=−m∏mcot(z−nkπ).
or, one has
cot(z−nkπ)=tanz−tannkπ1+tanztannkπ
for tan z finite; cot nz is thus a rational function in u=tanz, whose numerator is of degree n−1, and whose denominator, of degree n, has the n simple roots tankπ/n; decomposing this into simple fractions yields
valid for every complex number z not an integral multiple of π/2. One can write this formula in the form
cotz=ntannz1+k=1∑∞vk(n,z)
with vk(n,z)=0 for k>m and
vk(n,z)=cos2nkπ(ntannz)2−(nsinnkπ)22ntannz
for 1≤k≤m. We shall see that for every z contained in a compact subset K of C, not containing any integral multiple of π, and for every sufficiently large odd n, the series with general term vk(n,z) is normally convergent. Indeed, as n tends to +∞, tannz tends to nz uniformly on K, so there exists a number M>0 such that ntannz≤M for every sufficiently large m and every z∈K. On the other hand, for 0≤x≤π/2 one has sinx/x≥1−6x2≥21, so for 1≤k≤m one has nsinnkπ≥kπ/2; consequently, when m is sufficiently large, for every integer k such that kπ/2>M one has ∣vk(n,z)∣≤k2π2−4M28M, which proves our assertion. For k odd, vk(n,z) tends (uniformly on K) to z2−k2π22z as n tends to +∞. Consequently:
Theorem 1
For every complex number z not an integral multiple of π one has
(the series on the right being normally convergent on every compact subset K⊂C not containing any integer multiple of π (Eulerian expansion of cotz).)
2. Eulerian expansion of sinz
For every odd integer n=2m+1 and complex z the formula (2) of VI, p. 283, can be written
sinnz=(−1)m2n−1k=−m∏msin(z−nkπ)
=(−1)m2n−1sinzk=1∏msin(z−nkπ)sin(z+nkπ).
Now, one has sin(z−nkπ)sin(z+nkπ)=sin2z−sin2nkπ, and, by (3) (VI, 284), ∏k=1msin2nkπ=2n−1n, whence, on replacing z by z/n,
sinz=nsinnzk=1∏m(1−sin2nkπsin2nz).
We can write this formula as sinz=nsinnz∏k=1m(1−wk(n,z)), with wk(n,z)=0 if k>m and wk(n,z)=sin2nkπsin2nz for 1≤k≤m. We shall see that for every z contained in a compact subset K of C, and for n odd, the series with general term wk(n,z) is normally convergent. Indeed, as n tends to +∞, nsinnz tends uniformly to z on K, so there exists a number M>0 such that nsinnz≤M for every integer n and every z∈K. We saw, moreover, in the proof of th. 1 of VI, p. 286, that for 1≤k≤m one has nsinnkπ≥2kπ; thus for every integer k such that kπ/2≥M one has ∣wk(n,z)∣≤4M2/k2π2, which proves our assertion. Since, for every fixed k, wk(n,z) tends (uniformly on K) to z2/k2π2 as n tends to +∞, one sees that:
Theorem 2
For every complex number z one has
sinz=zn=1∏∞(1−n2π2z2)
the infinite product on the right being absolutely and uniformly convergent on every compact subset of C (Eulerian expansion of sinz).
converges. On the other hand, one can write, for every complex number z such that ∣z∣<π,
n2π2−z22z=n2π22zk=0∑∞n2kπ2kz2k
the series on the right being absolutely convergent. We shall deduce from this that the "double" series
n=1∑∞k=1∑∞n2kπ2k−2z2k−1
is absolutely convergent in the open disc ∣z∣<π, normally convergent on every compact set contained in this disc, and has sum cotz−z1. Indeed, for ∣z∣≤a<π the absolute value of the general term of (11) is at most equal to 2a2k−1/n2kπ2k, and the sum of any finite number of the terms 2a2k−1/n2kπ2k is less than the finite number ∑n=1∞n2π2−a22a; on summing first over k, then over n, one sees that the sum of the series (11) is equal to ∑n=1∞z2−n2π22z, which proves our statement.
If now one sums the series (11) first over n and then over k one obtains the identity (for ∣z∣<π)
cotz−z1=−2k=1∑∞π2kS2kz2k−1
where we have put Sk=∑n=1∞nk1. By (1) (VI, p. 283) one thus has, for ∣z∣<2π whence we obtain the formula
b2n=(−1)n−1(2n)!(2π)2n2S2nfor n≥1,
a formula which shows in particular that the numbers S2n/π2n are rational. Clearly Sk+1≤Sk so, for every integer k≥2, we have Sk≤S2=π2/6≤2; from (14) one deduces the following inequalities for the Bernoulli numbers
(2π)2n2(2n)!≤∣b2n∣≤4(2π)2n(2n)!for n≥1.
From these inequalities one can deduce an estimate for the Bernoulli polynomial
Bn(x)=k=0∑n(kn)bkxn−k; in particular, for 0≤x≤1 one has