Proc. Nat. Acad. Sci.

USA
Vol. 71, No. 4, pp. 1013-1015, April 1974

At Least One-Third of Zeros of Riemann's Zeta-Function are on O* = 1/2

NORMAN LEVINSON
Department of Mathematics, Room 2-365, Massachusetts Institute of Technology, Cambridge, Mass. 02139
Contributed by Norman Levinson, December 6, 1973

ABSTRACT Starting from the derivative of the func- arg(h(l - s) { [f'(s) + f'(1 - s) ]r(1 - s) + r'(1 - s) })
tional equation of the zeta-function, it is shown that at
least one-third of the zeros of the zeta-function are on the = T/2(mod ir)
line a = 1/2.
on T = 1/2 or, what is the same thing, where
It has been shown by Selberg (1) that there exists a computable
number c such that the number of zeros of the Riemann zeta- arg(h(){ [f'(s) + f'(1 - s)](s) + D'(s) })
function on the interval (1/2 + iT, 1/2 + iT + iU) is at least = /2 (mod [5] Tr)
c times the total number of zeros in the critical strip in T < t < on the A/rline. Since arg h(8) is available from Stirling's
T + U. By a different method it will be shown here that more formula, it suffices to find the change in the argument of
than 1/3 of the zeros are on the '/r2line. The account here is a
sketch. A detailed account will appear later. The account is G(s) = ¢(s) + r'(s)/[f'(s) +f'(1 - 5)] [6]
self-contained. It could also begin from the fact that ¢'(s) and on the '/rline. By the formula of Riemann-Siegel (2, ref. 3
r(s) have the same number of complex zeros in a < 1/2. This section 2.10)
latter is a theorem of Levinson and Montgomery, which will
appear in a paper entitled "Zeros of the derivatives of the c(s) = fi(s) + X(8)f2(8) [7]
Riemann zeta-function." It has also been pointed out to me by in an obvious notation. From the derivative of Eq. 7
Montgomery that a by-product is that, counting multiplicity,
less than 66% of the zeros in the critical strip can be multiple t'(S) = fi'(S) + x(s)f2'(8) - ['(s) + f'(1 - 8)]x(s)f2(8).
zeros on the '/rline.
Let h(s) = 1-8/2r(8/2). Then the functional equation of Using this and Eq. 7 in Eq. 6
the Riemann zeta-function is G(s) = H(s) + Hi(s) [81
h(s)r(s) h(1 - s) (1 - s).
= [1] where
By Stirling's formula, if h(s) exp (f(s)) then
= H(s) = f(s) + [f'(s) + x(s)f2'(s)]/log (t/2wr), [91
f(s) = [(8 - 1) log (s/2x) - s + Co + 0(1/s)]/2 Hi(s) = [f'(s) + x(s)f2'(s)]O(t-' log-' t/2r). [10]
and To get the change in arg G(s) on the l/2-line, the principle of
h'/h(s) = f'(s) = log (s/2,r)/2 + 0(1/s). [2] the argument can be used. The determination of the number
of zeros of G(s) in a rectangle D with vertices ('/2 + iT, 3 +
Differentiation of Eq. 1 yields iT, '/2 + i(T + U), 3 + i(T + U)) leads in a familiar way to
h(s)r(s) [f'(s) + f'(1 -s)] = -h(s) '(s) the change in arg G on a = '/2, T < t < T + U. Here U = TI
log'0 T. To get the number of zeros of G(s) in D, Littlewood's
-h( - 8)¢'(l - ). [31 lemma (ref. 3, section 9.9) is used in familiar way (ref. 3, sec-
On 8 = '/2 + it, the right side of Eq. 3 is the sum of two tion 9.15) and the crucial term for the number of zeros in D is,
complex conjugates and the zeros of the right side occur where for a < 1/2 and l/2- a very small,
T+ U
arg[h('/2 + it)r'('/2 + it)] T/2(mod Tr). [4] I log G(a + it) Idt/2Tr('/2-'a). [11]
T
Since in Iao <10 for large t
Actually in a well-known way (1), G is first multiplied with a
f'(s) + f'(1 - s) = log t/2Tr + 0(1/t), modifier which can only increase the number of zeros in D but
it follows that the zeros due to Eq. 4 are those of hr('/2 + it) makes the analysis tractable. Let 2 here be for 1 < j < y and
on the left of Eq. 3 and hence are the zeros of t('/2 + it) itself. let t(s) = 2b,/j', y = T1'2/log4O T,bj = At(j) log y/j/(j'/2-a log
By Eq. 1, if x(s) = h(l - s)/h(s), then D(s) = x(s)r(l - s) y). Instead of the numerator in Eq. 1,
and so 'T+U
- iT log IGV(a + it)Idt
¢'(s) = -x(s){[f'(s) + f'(1 -s)Jr(1 s) +- '(1 8).
-

Thus, by Eqs. 3 and 4 the zeros of W('/2 + it) occur where will be used and this is in turn dominated by
1013
1014 Mathematics: Levinson Proc. Nat. Acad. Sci. USA 71 (1974)
~ T+ U #T+ U
(/2) Ulog U-f IG4,(a + it)12dt . [12] .IT IH t'(a + it) U2dt'-. U2bklbktk2a kl-2a k2 -2a[-(1 -2a) -1
Because of Eq. 10 it is a triviality that G can be replaced in -(1 - 2a) -IL-2 log ki/k log k2/k + 2(1 - 2a) -2L-2
Eq. 12 by H and, hence, the crucial calculation involves X log kl/k - 2(1 - 2a)L -2 + 2(1 - 2a) -L-' log kl/k
Tr+U - 2(1 - 2a)-2 L-1] + Ur2-4a~bkbkk2-2a ki-1 k2-'
f+| IJHP(a it)12dt.
+ [13]
X [(1 - 2a)-'L-2 log ki/k log k2/k + 2(1 - 2a)-2L-2
The following are well known: X log ki/k + 2(1 - 2a)-L-2]. [14]
x = (t/2Tr) '1 exp (ri/4 - it log t/2re)(1 + O(t-1)) Because of the occurrence of r2-4a, the last terms are the more
important. Let
fi(8) = E j' + 0(t '2)
j< (t/2,,) 112
Kl = 2bklbtk2-2a ki-1 k2-'.
f2(8) = i 8'-1 + 0(t(o-1l)/2 Then from the definition of bj
j< (t/2-r)1/2
From Eq. 9, Kl = log-2 y 2u(k,),u(k2) k2-2a k1-'/2+a k2-"/2+a
rT+ U X log y/k, log y/k2
fTT+ Hik(a + it)12dt = In + I22 + In + 2(112 + I13 + I23)
since
where E. s(r)(m/r)2-2a = m2-2a" 2_(m)
rIm
J+
T P(a + it)(2dt where

I12 = Re
fT+U
f U fill' ',V(a + it)dt/log t/2,r
92-2a(m) = II
plm
(i - l/p2-2a) [15]

fT+U
Kii = log-2 y 2j2-2a O2-(j) (E j
(k1) ki -aI2+a log y/k1)2
IN 8= Ixf2
I 2' /i(a + it)12dt/log2 t/2ir etc.
Alk
2j-l 2O2a(j)j&2(j)(Z'(I(n) n-1I2+a log y/nj)2
= log-2 y

Let T1 = max (T, 2irji2, 2rj22). Then, after some calculation, where 2' is for n < y/j and (n,j) = 1. Let x = y/j for the
moment and let b be real and near 1. Let
r+ U
IIII- Dai2-af
jbkb2kl exp (it log jA12/jlkl)dt S = 2',i(n) n-b log x/n. [16]
where the outer sum is for kl, k2 < y and the inner one is
Then
for jl j2 < T where r = (T/2Tr)'/2. If k = (k1,k2), ki = A1k, 2 riS = fxW-b(w - b)-2/(r(w)l'(l - l/p'))dw
k2= A2kj1 = Aj, j2 = Aij, and AM = max (A1,A2) then
where I' is for plj. The line of integration is on Re w = 2.
Ill '~' U~biclbik2a ki-" k2-a Zj-2a Then by residue theory and appraisals on a shifted con-
where the inner sum is for j < TIAM. Let kM = max (k1,k2) tour
and km = min (k1,k2). From the above follows 2,riS 1) log x + 1]/Ob(j)
[(b -

Ill ' UT1-2a/(l - 2a)Zbklbk2kkM-l km-2a very crudely stated, which does yield
-U/(1 - 2a)Zbklbk2k2a k,-2a k2-2a. Kj log-2 y 2j-1 02_2a (j),s2(j) [(1/2- a) log y/j + 1 ]2/
_

Let L denote log T/2wr. Similar computations lead to

2I3/2-a(j) '--' (1/2- a)2(log y)/3 + (1/2- a) + (log y)'.
I22 ' Url-2a(2 - 4a) -1 L-1bklbkkkm-1 km-2a log Tkm/km
- U-rl2a(1 - 2a)-2 L-22bklbk2kkM-' km-2a(L/2
The other sums K12 and K22 are treated similarly where K12 in-
volves the additional factor log k/ik and K22 contains log k/ik
+ log Tkm/kM) log k2/k.
+ 2Url-2a(l - 2a) L-22bklbk2kkM-1 km-2a
-
Let R (1 - 2a) log T and recall log y - log T. Then the
=
- U(1 - 2a) -1 L-22bklbk2k2a k,-2a k2-2a log ki/k log k2/k
above computation leads to
+ 2U(1 - 2a)-2 L-22bbkC2k2a k,-2a k2-2a log kl/k
- 2U(1 - 2a)- L-2zbklbkk2a i-2a k2-2a
Ue2R(R-2/2 + R-1/24) [17]

and similarly for the other I's. The results for Il3 and I23 are for the terms involving r2-4". Similar treatment of the other
sums in Eq. 14 leads finally to
much more complicated to compute and involve the use of the
method of stationary phase (ref. 1, Lemma 2). For Il and In, OT+ ET
the use of polar coordinates at an appropriate point in the U-1 JHV,(a + it)I2dt _ e2R(R-3/2 + R-1/24)
calculation also is convenient. Fortunately, when all the terms
are added the asymmetric terms cancel and there results - R/12 + 7/12 - 25/(24R) - R-2 - 1/(2R3). [18]
Proc. Nat. Acad. Sci. U-SA 71 (1974) Zeros of Riemann's Zeta-Function on a- = 1/2 1015

If R = 1.3, an elementary computation shows that the right change in arg h('/2 + it) is essentially '/2U Log T/2-r and so
side of Eq. 18 is dominated by 2.3502. Since H can replace G the change in arg hG('/ ±+ it) is at least .17U log T/2T. Since
in Eq. 12, it follows that the zeros occur mod Tr, the number of them is at least .34U
r+ U (log T/2T)/2r, which does in fact exceed 1/3 the number of
fT log IGVI(a + it) Idt < U(log 2.3502)/2 < .4280U zeros of the zeta-function in the critical strip for T < t <
T + U and proves the result.
and so This work was supported in part by the National Science
Foundation Grant P22928.
('/2- a)-' f log IGV(a + it) Idt <
1. Selberg, A. (1942) "On the zeros of Riemann's zeta function,"
Skr. Norske Vid. Akad. Oslo 10, 1-59.
.4280U(Log T/2T)/R < .3293U Log T/2r. 2. Siegel, C. L. (1932) "Uber Riemann's Nachlass zur analy-
tischen Zahlentheorie," Quelen und Studien zur Geschichte
Accordingly, the change in arg(G('/, + it) on (T,T + U) is der Math. Astr. und Physik, Abstr. B 2, 45-80.
at most .33U Log T/2x. (The case of zeros of G on the l/2-line 3. Titchmarsh, E. C. (1951) The Theory of the Riemann Zeta.
requires a little added discussion.) By Stirling's formula, the Function (Oxford Univ. Press, London).