Zeros of Ramanujan Polynomials
Zeros of Ramanujan Polynomials
Zeros of Ramanujan Polynomials
net/publication/265065315
CITATIONS READS
32 185
3 authors, including:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Christopher James Smyth on 16 October 2014.
Abstract. In this paper, we investigate the properties of Ramanujan polynomials, a family of reciprocal
polynomials with real coefficients originating from Ramanujan’s work. We begin by finding their number
of real zeros, establishing a bound on their sizes, and determining their limiting values. Next, we prove
that all nonreal zeros of Ramanujan polynomials lie on the unit circle, and are asymptotically uniformly
distributed there. Finally, for each Ramunujan polynomial, we find all its zeros that are roots of unity.
1. Introduction
This paper investigates the properties of Ramanujan polynomials, which, for each k ≥ 0, the
authors of [2] define to be
k+1
X B2j B2k+2−2j
R2k+1 (z) = z 2j ,
(2j)!(2k + 2 − 2j)!
j=0
where Bj denotes the jth Bernoulli number. Of particular interest is the location of the zeros
of these polynomials, whose knowledge will give rise to explicit formulas for the Riemann zeta
function at odd arguments in terms of Eichler integrals.
Ramanujan polynomials are reciprocal polynomials with real coefficients, meaning that they
satisfy the functional equation
2k+2 1
R2k+1 (z) = z R2k+1 ,
z
where 2k + 2 = deg(R2k+1 ). This elegant property greatly simplifies the analysis of their zeros,
the details of which will be unveiled in later sections.
To begin, this paper will derive certain basic properties of Ramanujan polynomials, including
a bound on the sizes of their real zeros. Furthermore, we will show that the largest real zero of
R2k+1 tends to 2 from above as k approaches infinity.
The subsequent section will give a proof that all nonreal zeros of Ramanujan polynomials lie
on the unit circle. In particular, we prove that for each k, these zeros (which take the form eiθ )
are interlaced between angles θ for which sin kθ assumes the values ±1. Hence, as k tends to
infinity, the nonreal zeros of R2k+1 become uniformly distributed on the unit circle.
The final section of the paper will determine which zeros of R2k+1 are 2k-th roots of unity.
Specifically, the roots of unity that are zeros of R2k+1 are
• Both ±i if k is even;
• All four of ±ρ, ±ρ̄ if k is a multiple of 3,
and no others. Here ρ is a primitive cube root of unity.
1
2 M. RAM MURTY, CHRIS SMYTH, AND ROB J. WANG
2. Motivation
In Ramanujan’s notebooks, we find the following remarkable formula involving the odd values
of the Riemann-Zeta function (see [3]):
∞ ∞
( ) ( )
1 X n −2k−1 1 X n −2k−1
α−k ζ(2k + 1) + = (−β)−k ζ(2k + 1) +
2 e2αn − 1 2 e2βn − 1
n=1 n=1
(1) k+1
2k
X B2j B2k+2−2j
−2 (−1)j αk+1−j β j ,
(2j)!(2k + 2 − 2j)!
j=0
where α, β > 0 with αβ = π, and k is any positive integer. We recognize immediately that the
sum involving the Bernoulli numbers is
r !
k+1 β
α R2k+1 i .
α
A rigorous proof of this formula together with a generalization was obtained by Grosswald. He
proved the following (see [1]):
and set
∞
X σk (n)
Fk (z) = e2πinz
nk
n=1
(2πi)2k+1
2k 1
(2) F2k+1 (z) − z F2k+1 − = 12 ζ(2k + 1)(z 2k − 1) + R2k+1 (z).
z 2z
The function Fk (z) is an example of an Eichler integral, and the above formula relates the
values of two Eichler integrals to ζ(2k + 1) through the Ramanujan polynomial. In particular,
zeros of R2k+1 (z) that lie in the upper half plane and that are not 2k-th roots of unity give us
a formula for ζ(2k + 1) in terms of Eichler integrals. Indeed, the results of this paper tell us
that, for each k ≥ 4, there exists at least one algebraic number α with |α| = 1, α2k 6= 1 lying in
the upper half plane such that R2k+1 (α) = 0 and hence
In other words, there exists an explicit formula for the Riemann zeta function at odd arguments
9, 11, 13, . . . in terms of the difference of two Eichler integrals.
Though Ramanujan polynomials have appeared in the work of Grosswald and others, they
were never studied for their own sake. It turns out that they are of tremendous interest in their
own right, and serve as motivation for further applications. Indeed, the authors of [2] study the
function
2
G2k+1 (z) = 2k (F2k+1 (z) − z 2k F2k+1 (−1/z))
z −1
and show that the set
{G2k+1 (z) | =(z) > 0, z ∈ Q, z 2k 6= 1}
contains at most one algebraic number.
ZEROS OF RAMANUJAN POLYNOMIALS 3
P (z) = p0 + p1 z 2 + · · · + pd z 2d
Proof. We have
d d d
2d 1 X X X
z P = pj z 2(d−j) = pd−j z 2j = pj z 2j = P (z).
z
j=0 j=0 j=0
−d 1
z P (z) = z P = z̄ −d P (z̄).
d
z
Proof. This follows from the fact that the coefficient of z 2j of R2k+2 (z) is the same as that of
z 2k+2−2j .
From this corollary we see that replacing z by −1/z in the identity (2) gives the same identity
again.
Before moving on, let us take a moment to list the first few Ramanujan polynomials and
their zeros (the values given in parentheses are approximations to exact solutions by radicals).
Notice that, for 1 ≤ k ≤ 8, R2k+1 (z) has exactly 4 real zeros. Furthermore, the largest of the
real zeros is always between 2 and 2.2 (and it seems to be approaching 2 as k increases). On
the other hand, the nonreal zeros seem to lie exactly on the unit circle.
1 2
R1 (z) = 2·3! (z + 1) (this is the trivial case) Zeros: ± i
q √
4
R3 (z) = 1
6! (−z + 5z 2 − 1) Zeros: ± 5±2 21 (±2.1889, ±0.4569)
q √
6
R5 (z) = 1
12·7! (2z − 7z 4 − 7z 2 + 2) Zeros: ± i, ± 9±4 65 (±2.0653, ±0.4842)
q √
8 6 4 2
R7 (z) = 1
10! (−3z + 10z + 7z + 10z − 3) Zeros: ± ρ, ±ρ̄, ± 13±6 133 (±2.0221, ±0.4945)
1 10 8 6 4 2
R9 (z) = 12! (10z − 33z − 22z − 22z − 33z + 10)
r r
√ q √ √ q √
43 3 201 1 1029 129 201 43 3 201 i −1029 129 201
Zeros: ± i, ± 40 + 40 ± 2 200 + 200 , ± 40 − 40 ± 2 200 + 200
(±2.0071, ±0.4982, ±0.7112 ± 0.7030i)
4 M. RAM MURTY, CHRIS SMYTH, AND ROB J. WANG
And a few more cases (all zeros other than ±i, ±ρ, ±ρ̄ are approximations):
1
R11 (z) = 2·15! (−1382z 12 + 4550z 10 + 3003z 8 + 2860z 6 + 3003z 4 + 4550z 2 − 1382)
Zeros: ± 2.0022, ±0.4995, ±0.3081 ± 0.9513i, ±0.8146 ± 0.5800i
1
R13 (z) = 12·15! (210z 14 − 691z 12 − 455z 10 − 429z 8 − 429z 6 − 455z 4 − 691z 2 + 210)
Zeros: ± i, ±ρ, ±ρ̄, ±2.0006, ±0.4998, ±0.8715 ± 0.4904i
16
R15 (z) = 1
5·18! (−10851z + 35700z 14 + 23494z 12 + 22100z 10 + 21879z 8
+ 22100z + 23494z 4 + 35700z 2 − 10851)
6
t 2
Hence we can eliminate (et −1) 2 to obtain
t2
2
2t t 2t d t t 2t
= − − − · 2t
e2t − 1 et − 1 et − 1 et − 1 dt et − 1 2 e −1
∞
X Bn n n+1 n t n
= 2t − t − (n + 1)t − (2t)
n! 2
n=0
∞
X Bn
(n − 1)tn + (1 + 2n−1 )tn+1 .
= −
n!
n=0
The second lemma presents two known results concerning series involving the zeta-function.
Lemma 4.5 (see [4, equations (45) p. 163 and (193) p.178]). We have
∞
X ζ(2j)
= 12
4j
j=1
and
∞
X
(ζ(2j) − 1) = 43 .
j=1
Proof. Now
∞ ∞ ∞
X ζ(2j) X
−j
X 1
= 4
4j k 2j
j=1 j=1 k=1
∞ ∞
XX 1
=
(2k)2j
k=1 j=1
∞
X 1
=
(2k)2 − 1
k=1
∞
1
X 1 1
= 2 −
2k − 1 2k + 1
k=1
1
= 2 ,
where the second equality follows from changing the order of summation, and last equality
follows from telescoping series.
The second result is proved in a similar manner.
Next, we need the following estimate.
Lemma 4.6. For k ≥ 1 and j = 1, . . . , k, we have
ζ(2k + 2 − 2j)
− 1 < 3 · 4j−(k+1) .
ζ(2k + 2)
Proof.
ζ(2k + 2 − 2j)
− 1 < ζ(2k + 2 − 2j) − 1
ζ(2k + 2)
2k + 3 − 2j j−(k+1)
< 4
2k + 1 − 2j
≤ 3 · 4j−(k+1) ,
using Lemma 4.4, and the fact that j ≤ k.
Equipped with these lemmas, we are ready to prove Theorem 4.3.
Proof of Theorem 4.3. We have
√ k
M2k+1 ( 4 + t) X ζ(2j) ζ(2k + 2 − 2j) 1
k+1
= 1 − 2 j
+ ,
(4 + t) (4 + t) ζ(2k + 2) (4 + t)k+1
j=1
∞
X ζ(2j)
which, on replacing 1 by 2 using Lemma 4.5, gives
4j
j=1
∞ k
X ζ(2j) 1 X 1 1 ζ(2k + 2 − 2j)
2 + +2 ζ(2j) − .
4j (4 + t)k+1 4j (4 + t)j ζ(2k + 2)
j=k+1 j=1
ZEROS OF RAMANUJAN POLYNOMIALS 7
√
We now claim that M2k+1 ( 4 + t) is positive for some small t > 0 that goes to 0 as k → ∞.
For this to hold, we see from the above expression that a sufficient condition is
ζ(2k + 2 − 2j)
4−j > (4 + t)−j for j = 1, . . . , k.
ζ(2k + 2)
Using the upper bound in Lemma 4.6, it is therefore sufficient that
t j
1+ > 1 + 3 · 4j−(k+1) ,
4
or equivalently
t
> (1 + 3 · 4j−(k+1) )1/j − 1.
4
Since for a ≥ 0 and 0 < δ ≤ 1 we have (1 + a)δ ≤ 1 + aδ, we replace this condition by
3 · 4j−k
t> .
j
This lower bound √ attains its maximum at j = k, and hence we obtain our final sufficient
condition for M2k+1 ( 4 + t) to be positive, namely that
3
t> .
k
q
Hence for k ≥ 5 the zero z0 of M2k+1 lies in the open interval 2, 4 + k3 . It follows that
z0 < 2.15 for k ≥ 5 and z0 → 2 as k → ∞.
Lemma 5.3. For k ≥ 5, the polynomial ∆(z) satisfies |∆(z)| < 1.3 for z on the unit circle.
Furthermore, writing
k−1
X
∆(z) = (ε1 − 1)(z 2 + z 2k ) + εj z 2j ,
j=2
we have, for 1 ≤ j ≤ k,
ζ(2j)ζ(2k + 2 − 2j)
(6) εj = 2 −1 .
ζ(2k + 2)
Proof. The formula for εj follows from the easily-verified fact that
A(z) = z 2k+2 − 3z 2k − 2z 2k−2 − · · · − 2z 4 − 3z 2 + 1.
To bound ∆(z), we first note that, since ε1 − 1 and all the εj for 2 ≤ j ≤ k − 1 are positive,
k−1
X
|∆(z)| < 2(ε1 − 1) + εj
j=2
for z on the unit circle. We therefore need to bound this sum from above.
Recalling that εj = εk+1−j , invoking Lemma 4.4 gives
π 2 2ζ(2k)
ε1 = εk = −2
6 ζ(2k + 2)
π2
2k + 1 −k
< 1+ 4 −2
3 2k − 1
< 1.3
for k ≥ 5, since the right hand side of the inequality in Lemma 4.4 is strictly decreasing.
Now, once again invoking Lemma 4.4, we have
2j 0 + 1 −j 0
2j + 1 −j
εj < 2 1+ 4 1+ 0 4 −1
2j − 1 2j − 1
2j + 1 −j 2j 0 + 1 −j 0 (2j + 1)(2j 0 + 1) −(k+1)
= 2 4 + 0 4 + 4 .
2j − 1 2j − 1 (2j − 1)(2j 0 − 1)
Summing both sides over j and using Lemma 5.2 gives us
k−1 ∞
X X 2j + 1 5
εj < 4 4−j + (k − 2)4−k
2j − 1 4
j=2 j=2
∞
X 2 5
= 4 1+ 4−j + (k − 2)4−k
2j − 1 4
j=2
5 5
= 2 log 3 − + (k − 2)4−k
3 4
< 0.7
for k ≥ 5, using the fact that
∞
x2j−1
1+x X
log =2 ,
1−x 2j − 1
j=1
and so
∞
X 1
log 3 = 1 + 4 4−j ,
2j − 1
j=2
ZEROS OF RAMANUJAN POLYNOMIALS 9
Proof of Theorem 5.1. We first note that the polynomial A(z), which approximates M2k+1 (z),
can, using (3), be written as
−k
k
k+1 z − z
A(z) = z ((z − z −1 )2 − 2),
z − z −1
Restricting z to the unit circle (in other words, putting z = eiθ ) gives us
sin kθ
z −(k+1) A(z) = −(2 + 4 sin2 θ) = f (θ),
sin θ
say. In particular, f is a real-valued function of θ. Next, consider
g(θ) = e−(k+1)iθ ∆(eiθ ).
Note that, by Lemma 3.1, ∆(z) is reciprocal, and g(θ) is real-valued; also, by Lemma 5.3, it is
bounded above by 1.3.
Finally, by confining our attention to θ ∈ (0, π), we see that f (θ) > 2 whenever sin kθ = −1
and f (θ) < −2 whenever sin kθ = 1. This is because
2 + 4 sin2 θ
> 2 + 4 sin2 θ
sin θ
> 2
on the interval (0, π).
Therefore the function
e(k+1)iθ M2k+1 (eiθ ) = f (θ) + g(θ)
has a zero between every consecutive pairs of values of θ ∈ (0, π) with sin kθ = ±1. There are
precisely k values for which sin kθ = ±1, namely
(2l + 1)π
, l = 1, . . . , k,
2k
giving k − 1 zeros of M2k+1 on the upper half of the unit circle. Taking complex conjugates
yields another k − 1 zeros on the lower half of the unit circle. Hence M2k+1 has exactly 2k − 2
zeros on the unit circle, which is what we wanted to show.
Observe that the above theorem not only gives the modulus of the nonreal zeros, but also
gives some restriction on their distribution. In particular, these zeros are interlaced between
angles θ for which sin kθ assumes the values ±1, which means that, asymptotically, they are
uniformly distributed on the unit circle.
We also observe another useful result concerning zeros of Ramanujan polynomials:
Corollary 5.4. Ramanujan polynomials have no repeated zeros.
Proof. The real zeros of M2k+1 were already shown to be distinct. The nonreal zeros on the
unit circle are strictly interlaced between angles θ for which sin kθ assumes the values ±1, and
hence never coincide with each other.
10 M. RAM MURTY, CHRIS SMYTH, AND ROB J. WANG
Hence, choosing ω to be such a z where |B(z)| is largest, we must have |B(ω)| ≥ 1. Since
|z 4 − 4z 2 + 1| ≥ 2
for z on the unit circle, we see from (3) that |A(ω)| ≥ 2, a contradiction.
Next, we need the following lemma.
Lemma 6.3. Every cyclotomic polynomial φ` except for φ3 , φ4 and φ6 has a zero z = eiθ with
14 14
θ ∈ [0, 45 π] ∪ [π − 45 π, π].
Proof. If ` = 1, ` = 2 or ` ≥ 8, then the zero e2πi/` of φ` has its argument θ in the required
range. Also φ5 has the zero e4πi/5 and φ7 has the zero e6πi/7 , with both these zeros also having
their arguments in the required range.
Proof of Theorem 6.1. The theorem holds for k < 8 by the computations of Section 2; we can
therefore assume that k ≥ 8. Suppose that ω is a root of unity lying in the upper half plane
such that M2k+1 (ω) = 0. By Proposition 6.2, ω must be a 2k-th root of unity. Then, since
A(ω) = 0, we have ∆(ω) = 0. If ω is an `-th root of unity then, since ∆ has rational coefficients,
ZEROS OF RAMANUJAN POLYNOMIALS 11
where the εj are given by (6), and where the ∗ indicates that the final term is halved for k odd.
So, on putting ω = eiθ we obtain
b k+1 c
π2 2
ζ(2k) X
∗ 2ζ(2j)ζ(2k + 2 − 2j)
· − 3 2 cos θ + − 2 2 cos((2j − 1)θ) = 0.
3 ζ(2k + 2) ζ(2k + 2)
j=2
We now introduce an integer parameter r, to be chosen later, lying in the range 1 ≤ r < b k+1
2 c.
Then we have
r
π2 π2
ζ(2k) X
− 3 cos θ + · − 1 cos θ + 2 (ζ(2j) − 1) cos((2j − 1)θ)
3 3 ζ(2k + 2)
j=2
r
X ζ(2k + 2 − 2j)
+ 2ζ(2j) − 1 cos((2j − 1)θ)
ζ(2k + 2)
j=2
b k+1
2
c
X
∗ ζ(2j)ζ(2k + 2 − 2j)
+ 2 − 1 cos((2j − 1)θ) = 0.
ζ(2k + 2)
j=r+1
Hence, defining
r
π2
X
hr (θ) = − 3 cos θ + 2 (ζ(2j) − 1) cos((2j − 1)θ),
3
j=2
we have
r
π2
ζ(2k) X ζ(2k + 2 − 2j)
|hr (θ)| < · −1 +2 ζ(2j) −1
3 ζ(2k + 2) ζ(2k + 2)
j=2
b k+1
2
c
X ζ(2j)ζ(2k + 2 − 2j)
+2 −1
ζ(2k + 2)
j=r+1
r
π2 X
< · (ζ(2k) − 1) + 2 ζ(2j)(ζ(2k + 2 − 2r) − 1)
3
j=2
b k+1
2
c
X
+2 (ζ(2j)ζ(k + 1) − 1)
j=r+1
r
π 2 2k + 1 −k X 2k + 3 − 2r −(k+1−r)
< · 4 +2 ζ(2j) 4
3 2k − 1 2k + 1 − 2r
j=2
b k+1
2
c
X
k + 2 −(k+1)
+2 ζ(2j) 1 + 2 −1 ,
k
j=r+1
b k+1
2
c
X
k + 2 −(k+1)
2 ζ(2j) 1 + 2 −1
k
j=r+1
r
3 X k+2 k+1
< −2 (ζ(2j) − 1) + ζ(2r + 2) −r ,
2 k2k 2
j=1
r
π 2 2k + 1 −k X 2k + 3 − 2r −(k+1−r)
|hr (θ)| < · 4 +2 ζ(2j) 4
3 2k − 1 2k + 1 − 2r
j=2
r
3 X k+2 k+1
+ −2 (ζ(2j) − 1) + ζ(2r + 2) − r .
2 k2k 2
j=1
We now proceed to prove that M2k+1 (i) = 0 if and only if k is even, while M2k+1 (ρ) = 0
if and only if 3 | k. To begin, we see from (3) that for k odd, A(i) = 6, so that |∆(i)| < 1.3
implies that M2k+1 (i) = A(i) − ∆(i) 6= 0. Similarly, if k is not a multiple of 3, then |A(ρ)| = 5,
which again shows that M2k+1 (ρ) 6= 0.
On the other hand, for k even we have 12 deg(M2k+1 ) = k + 1 is odd, so by the evenness and
the functional equation for M2k+1 we have
(2πi)2k+1
1 1
F2k+1 (ρ) − ρ2k F2k+1 − = ζ(2k + 1)(ρ2k − 1) + R2k+1 (ρ),
ρ 2 2ρ
or equivalently
(2πi)2k+1
F2k+1 (ρ) − F2k+1 (1 + ρ) = R2k+1 (ρ).
2ρ
ZEROS OF RAMANUJAN POLYNOMIALS 13
But since
∞
X σ2k+1 (n)
F2k+1 (1 + ρ) = e2πin(1+ρ)
n2k+1
n=1
∞
X σ2k+1 (n) 2πinρ
= e
n2k+1
n=1
= F2k+1 (ρ),
the left-hand side of the above equality is 0, giving R2k+1 (ρ) = 0.
Now, from M2k+1 (−i) = M2k+1 (i) we see that M2k+1 (i) = 0 if and only if M2k+1 (i) =
0. Similarly, we have M2k+1 (ρ̄) = M2k+1 (ρ) and M2k+1 (−ρ) = M2k+1 (ρ), since M2k+1 is a
polynomial in z 2 . So all four of ±ρ, ±ρ̄ are zeros of M2k+1 if any one of them is. Hence both of
±i are zeros of M2k+1 if and only if k is even, and all four of ±ρ, ±ρ̄ are zeros of M2k+1 if and
only if k is a multiple of 3. These are the only zeros of M2k+1 that are roots of unity.
As a final observation, for k even, the fact that R2k+1 (i) = 0 allows us to deduce that,
k+1
X B2j B2k+2−2j
(−1)j = 0,
(2j)!(2k + 2 − 2j)!
j=0
√
which agrees with claim (2) in [2]. This also follows (1), on putting α = β = π.
Acknowledgement. We thank the referee for helpful comments on an earlier version of this
paper.
References
[1] E. Grosswald, Die Werte der Riemannschen Zetafunktion an ungeraden Argumentstellen, Nachr. Akad. Wiss.
Göttingen (1970).
[2] S. Gun, M. R. Murty, and P. Rath, Transcendental Values of Certain Eichler Integrals (to appear).
[3] S. Ramanujan, Notebooks - Vols. 1, 2, Tata Institute of Fundamental Research, Bombay (1957).
[4] H. M. Srivastava and J. Choi, Series associated with the zeta and related functions, Kluwer Academic Publishers,
Dordrecht, 2001.
[5] B. L. van der Waerden, Algebra, Volume 1, Ungar, New York (1970).
Department of Mathematics and Statistics, Queen’s University, Kingston, Ontario K7L 3N6
E-mail address: [email protected]
School of Mathematics and Maxwell Institute for Mathematical Sciences, University of Edinburgh,
Edinburgh, Scotland, United Kingdom EH9 3JZ
E-mail address: [email protected]
Department of Mathematics and Statistics, Queen’s University, Kingston, Ontario K7L 3N6
E-mail address: [email protected]