1. Introduction. The intention of this note is to provide a complete proof of the following formula: 1
0
2 log2 (2 cos 1 ) d = 2
11 4 11 = (4). 180 2
1+
1 2
n=1
++ n
1 n
17 4 17 = (4). 360 4
This identity was surprising and new to us when Enrico Au-Yeung (an undergraduate student in the Faculty of Mathematics in Waterloo) conjectured it on the basis of a computation of 500,000 terms (5 digits accuracy!). Our rst impulse was to perform a higher-order computation to show it to be false. It is not easy to naively compute the value of the sum to more than about 8 places. It occurred to us that we might be able to (i) obtain a function whose Fourier sine series was 1 1 1 1+ + + sin nt; n 2 n n=1 and (ii) numerically determine the value of the desired sum via Parsevals equation. This lead to equation (12) below and an evaluation of the sum to 25 digits in a
*This research was supported in part by the Natural Sciences and Engineering Research Council of Canada. 2000 Mathematics Subject Classication. Primary 42A24, 11M99; Secondary 33B15, 42A16, 65B10, 65B15. Key words and phrases. Riemann zeta function, Parsevals identity, generating functions, logcosine integrals, polylogarithms. Typeset by AMS-TEX
couple of minutes in Maple on a work station. Now armed with the assurance that the result was true, we were prepared to look for the closed form evaluation we present below. We did not know at the time that P.J. De Doelder[4,(9)] had established (2) below from which the above sum is an immediate consequence. In 5 below we present complete, self-contained and shorter versions of De Doelders proofs of this and related sums. We believe our derivations makes a very pleasant tour through diverse topics in Classical Analysis, almost all of which are well described in [1] or [6]. We will use the following standard denitions and identities involving the Riemann zeta function , the beta function , the gamma function , and the digamma function : 2 4 1 , (4) = , where (z) := for Re z > 1; (2) = 6 90 nz n=1
1
(x, y) :=
0
(1 u)x1 uy1 du =
(see [6, Theorem 7.69]); (z)(1 z) sin z = (see [6, Theorem 7.64]); d (1 + z) := log (1 + z), dz (see [6, Theorem 7.71]). 2. The main results. Theorem.
(1 + z) =
n=2
2 log2 (2 cos 1 ) d = 2
11 4 11 = (4). 180 2
(1)
Corollary. Let Hn :=
k=1
1 . Then k (2)
(3)
and
Hn 4 5 = = (4). 3 n 72 4 n=1
(4)
AN INTRIGUING INTEGRAL
3. Proofs of the main results. Proof of the Theorem. Suppose initially that x 0, 0. Integrating (1 + w) x w in the w = u + iv plane around the contour bounded by the u-axis from 1 to 1 and the upper semicircle of unit radius yields
1
(1 + w)x w dw = i
1 0
(1 + ei )x ei(+1) d,
since (1 + w)x w is holomorphic within and continuous on and within the contour (see [3]). The imaginary part of the right-hand term is
1 (2 cos 2 )x cos y d
(1 + u)x u du + ei
0 0
(1 u)x u du
(1 +
1 2x
(1 + x) . 1 + y)(1 + 2 x y)
0 1 (2 cos 2 )x cos y d =
(1 + x) . 1 (1 + 1 x + y)(1 + 2 x y) 2
(5)
By analytic continuation we can conclude that (5) in fact holds for x > 1 and all real y. Dierentiating (5) twice with respect to y and then putting y = 0 yields 1
0
2 (2 cos 1 )x d = 2c(x), 2
(6)
where c(x) is the coecient of y 2 in the power series expansion of the right-hand expression in (5). Let f (y) := Then
1 f (y) = f (y) (1 + 2 x y) (1 + 1 x + y) , 2
1 . 1 (1 + 1 x + y)(1 + 2 x y) 2
1 (1 + x) (1 + 2 x) 1 = g(x) (1 + 2 x), 1 2 (1 + 2 x)
g(x) :=
(1 + x) . 2 (1 + 1 x) 2
(7)
where c is the coecient of x2 in the power series expansion of c(x). Now g(0) = 1 and g (x) = g(x) (1 + x) (1 + 1 x) , 2 so that g (0) = 0, and hence
1 g (0) = g(0) (1) 2 (1) = 1 (1) = 2
2 . 12
It follows that
1 c(x) = g(x) (1 + 2 x) = 1 +
2 2 x + 24
2 4 2 (3)x + x , 6 120
4 4 11 4 + = . 144 120 180 Identity (1) follows from (7) and (8). 4c = 4
(8)
Proof of the Corollary. First we observe that the ordinary generating function of Hn is given by log(1 z) = Hn z n 1z n=1 from which it follows that 1 Hn n+1 log2 (1 z) = z , 2 n+1 n=1
(9)
with both series converging for |z| < 1. For z = reit , 0 < r < 1, 0 < t , the imaginary part of the right-hand expression in (9) becomes Hn n+1 r sin (n + 1)t n+1 n=1 which, by Abels theorem, tends to Hn sin (n + 1)t n+1 n=1
AN INTRIGUING INTEGRAL
as r 1, the convergence of the latter series being assured because Hn /(n + 1) tends monotonically to 0 (see [6, Corollary 7.38]). On the other hand, the imaginary 1 1 1 part of 2 log2 (1 reit ) tends to 2 (t ) log(2 sin 2 t) as r 1, and hence, for 0 < t , 1 Hn 1 (t ) log(2 sin 2 t) = sin (n + 1)t . (10) 2 n+1 n=1 Applying Parsevals theorem to the Fourier series (10), we derive 1 2
0
2 Hn . (n + 1)2 n=1
(11)
On replacing t by in (11) and using (1) we obtain (2). To derive (3) we observe that it follows from (9) that, for |z| < 1, zn 1 Hn n + log2 (1 z) = z . n2 2 n n=1 n=1 Equating imaginary parts with z = eit , we obtain as in the derivation of (10) that, for 0 < t , Hn sin nt 1 1 + (t ) log(2 sin 2 t) = sin nt. 2 n 2 n n=1 n=1 Now, for 0 < r < 1, 0 < t ,
t
(12)
log(1 reiu ) du = i
rn (eint 1) , n2 n=1
a(t) :=
0
1 log(2 sin 2 u) du =
sin nt . n2 n=1
This identity, which appears as (4.4) in [3], was numerically investigated by Clausen in 1832, and the function a(t) is Clausens function Cl2 (t). It follows, by Parsevals theorem, that 2 a(t)2 dt = (4), 0 and hence, by partial integration, 2
2a(t)a (t)( t) dt =
0
a(t)2 dt = (4).
0
Applying Parsevals theorem to (12) and using the above identities together with (1), we get
2 2 Hn = n2 n=1 0 1 a(t) + 2 a (t)( t) 2
dt
which establishes (3). Finally, we can deduce (4) from (2) and (3) by observing that
2 Hn1 = Hn
1 n
2 = Hn +
1 2Hn , 2 n n
so that
Identity (4) was discovered by Euler in 1775, as also was the identity Hn = 2(3). n2 n=1 In fact, Euler obtained the more general formula Hn 2 = (m + 2)(m + 1) nm n=1
m2
(m n)(n + 1) for m = 2, 3, .
n=1
See [2, p. 252, and the references there given] and also, for related material, [7]. Identities such as these involving Hn can be derived more easily than those involving 2 Hn . They can be dealt with by integrating appropriate generating functions without recourse to Parsevals theorem. 4. Two related examples. One may attempt to sum (a1 + a2 + + an )2 /n2 for more general sequences. Success depends on being able to get tractable forms of both the generating function corresponding to (9) and the Fourier-Parseval integral corresponding to (11) . If one commences with arctanh instead of log, one is lead to 3 1 log2 (tan 4 t) dt = , 4 0 and hence to De Doelders [4,(22)] formula
2
1+
1 3
++ n
1 2n1
n=1
4 45 = (4). 32 16
(13)
AN INTRIGUING INTEGRAL
We also attempted the same process for the sum involving alternating harmonic terms 2 n+1 1 1 2 + 1 + (1) 3 n . n n=1 In this case, however, we initially were only able to evaluate the corresponding integral numerically. Moreover, numerical experimentation by David Bailey at NASA Ames Research Center showed that the sum did not have as simple a closed form as in the previous examples, but involved more exotic polylogarithmic constants. In fact his heuristic evaluation of the sum (with vanishingly small probability of error) was 13 5 1 (4) + (2) log2 2 + log4 2 + 2Li4 ( 1 ), 2 8 2 12 n x where Li4 (x) = . Using arguments involving polylogarithmic and integral n4 n=1 identities appearing in [5], we eventually were able to establish Baileys evaluation rigorously. 5. Modied versions of De Doelders derivations of (2) and (13). In comparing De Doelders formulae with ours it should be observed that, for n = 1, 2, , Hn = (n + 1) (1), and On :=
k=1 n
1 1 1 = (n + 1 ) ( 2 ) . 2 2k 1 2
Because De Doelder did not choose to highlight these relations, it was not immediately evident from the title of his paper or his abstract that he was dealing with the same series as we were. Proof of (2). Observe rst that
1 0
xn logm x dx = (1)m
1 0
(14)
log3 x dx 1x
1 0
n=0
(15)
Next, for n = 0, 1, ,
1 1
xn log(1 x) dx =
0 1
xn dx
0
1 dt 1t
1
=
0
1 dt 1t
1 0
xn dx =
t
1 n+1
1 0
1 tn+1 dt 1t
(16)
1 n+1
(1 + t + + tn ) dt =
Hn+1 . n+1
xn log(1 x) dx =
0
1 4 so that
d 1+x log2 dx 1x
(17)
1 x2n 1 dx = 2 1x 2
u 0
1 0
dx 1 x2
1 x 1 0
1 2
1 0
1 1+u log2 du u 1u
dx 1 = 2 1x 4
Substituting (1 u)/(1 + u) = t in the nal integral and then applying (14), we obtain 1 2 On 1 1 log3 t 1 = dt = t2n log3 t dt 2 2 n 2 0 1t 2 n=0 0 n=1 = 3 3 4 = 3(4) (4) = . (2n + 1)4 16 32 n=0
AN INTRIGUING INTEGRAL
This establishes (13), which is the corrected version of identity (22) in [4]. As a bonus we also obtain from (17), on substituting x = 1 and (1 u)/(1 + u) = t, and then applying (14), that On = n2 n=1 = which is identity (15) in [4]. It should be noted that De Doelders derivation of (2) in conjunction with (11) provides an alternative proof of (1).
1 0
log2 t dt = 1 t2 n=0
1 0
t2n log2 t dt
References
1. T.M.A. Apostol, Mathematical Analysis, Second edition, Addison-Wesley, 1975. 2. B.C. Berndt, Ramanujans Notebooks, Part I, Springer-Verlag, New York, 1985. 3. R.B. Burckel, An Introduction to Classical Complex Analysis, Volume 1, Academic Press, 1979. 4. P.J. De Doelder, On some series containing (x) (y) and ((x) (y))2 for certain values of x and y, J. Comp. and Applied Math. 37 (1991), 125141. 5. L. Lewin, Polylogarithms and Associated Functions, North Holland, New York, 1981. 6. K.R. Stromberg, An Introduction to Classical Real Analysis, Wadsworth, 1981. 7. A.J. van der Poorten, Some wonderful formulae...An introduction to polylogarithms, Proceedings of Number Theory Conference, Kingston, Ontario (1979), 269286.
Department of Mathematics University of Western Ontario London, Ontario, Canada N6A 5B7
Department of Mathematics and Statistics Simon Fraser University Burnaby, British Columbia, Canada V5A 1S6