Slide 1:

Some non-standard methods

to perform calculations
with Riemanns zeta function
Yuri Matiyasevich
Steklov Institute of Mathematics
at St.Petersbourg
http://logic.pdmi.ras.ru/~yumat/personaljournal/artlessmethod
These are annotated slides of the talk given by the author on November 22, 2012 at
join session of international meeting Zeta functions (http://www.mccme.ru/poncelet/
2012zeta/) and seminar Globus (http://www.mccme.ru/ium/globus.html). The lat-
ter seminar is devoted to mathematics in general, and by this reason the talk included a
short introduction to the history of the study of Riemanns zeta function giving necessary
(and, hopefully, sucient) background for understanding the rest of the talk.
Experts may skip the rst 21 slides and jump to Slide 22.
Other materials related to this talk, including video, can be found at http://logic.
pdmi.ras.ru/~yumat/personaljournal/artlessmethod/talks/poncelet_2012.
ATTENTION
The talk was based on intensive computations. After the talk was presented, Gleb
Beliakov from Deakin University, Australia, independently performed calculation of
the same determinants. The values obtained by him are in good agreement with the
values obtained by the author as long as the sizes of matrices are below 38003800.
However, for larger sizes there is great discrepancy. By this reasons some original
slides were modied or removed.
Slide 2:
Blowing the Trumpet
The physicist George Darwin used to
say that every once in a while one
should do a completely crazy exper-
iment, like blowing the trumpet to
the tulips every morning for a month.
Probably nothing will happen, but if
something did happen, that would be
a stupendous discovery.
Ian Hacking
Representing and Intervening
Cambridge University Press
p.154, 1983
This research started as a kind of mathematical blowing the trumpet to the tulips
but the outcome seems to be striking and deserving further investigation.
George Darwin was a son of famous naturalist Charles Darwin.
Slide 3:
Riemanns zeta function
Dirichlet series:
(s) =

n=1
n
s
=
1
1
s
+
1
2
s
+
1
3
s
+ +
1
n
s
+ . . . (1)
The series converges for s > 1 and diverges at s = 1:
1
1
+
1
2
+
1
3
+
1
4
+ . . . (2)
The zeta function is named after Georg Friedrich Bernhard Riemann but it was pre-
viously studied by Leonhard Euler.
Slide 4:
Basel Problem (Pietro Mengoli, 1644)
1
1
2
+
1
2
2
+
1
3
2
+ +
1
n
2
+ = ?
Leonhard Euler:
1
1
2
+
1
2
2
+
1
3
2
+ +
1
n
2
+ = (2) = 1.64493406684822644 . . .(3)

2
6
= 1.64493406684822644 . . .(4)
(2) =

2
6
(5)
Pietro Mengoli asked in 1644: What is the value of this sum?
Euler calculated more than a dozen of decimal digits of the sum, which wasnt an easy
exercise because the series in (3) converges very slowly. Luckily, Euler invented what is
nowadays called EulerMaclaurin summation.
Knowing the value (4), Euler conjectured the equality (5). His rst proof given in
1735 was not rigorous by todays standards but later he returned to this problem several
times and gave a number of quite rigorous proofs.
Slide 5:
Other values of (s) found by Euler
(2) =
1
6

2
(6)
(4) =
1
90

4
(7)
(6) =
1
945

6
(8)
(8) =
1
9450

8
(9)
(10) =
691
638512875

10
(10)
(12) =
2
18243225

12
(11)
(14) =
3617
325641566250

14
(12)
Euler didnt stop by merely answering Mengolis question and found many other values
of (s). But why such strange numerators of (10) and (14)?
Slide 6:
Bernoulli numbers
x
e
x
1
=

k=0
B
k
x
k
k!
(13)
B
0
= 1, B
2
=
1
6
, B
4
=
1
30
, B
6
=
1
42
, B
8
=
1
30
, (14)
B
10
=
5
66
, B
12
=
691
2730
, B
14
=
7
6
, B
16
=
3617
510
(15)
B
1
=
1
2
, B
3
= B
5
= B
7
= B
9
= B
11
= = 0 (16)
No doubts that Euler immediately recognized 691 and 3617 as numerators of the so-
called Bernoulli numbers. Named after Jacob Bernoulli, these numbers can be dened in
many ways, in particular, from the coecients in the Taylor expansion (13).
Notice that starting from 3, all Bernoulli numbers with odd indices are equal to zero.
Slide 7:
Other values of (s) found by Euler
(6) =
1
945

6
=
2
5
6!
B
6

6
(17)
(8) =
1
9450

8
=
2
7
8!
B
8

8
(18)
(10) =
691
638512875

10
=
2
9
10!
B
10

10
(19)
(12) =
2
18243225

12
=
2
11
12!
B
12

12
(20)
(14) =
3617
325641566250

14
=
2
13
14!
B
14

14
(21)
(2k) = (1)
k+1
2
2k1
(2k)!
B
2k

2k
k = 1, 2, . . . (22)
After substituting (14)(15) into (6)(12), it is easy to guess the form of the other
factors, and Euler gave general formula (22).
We still know very little about the values of (s) for odd positive values of s.
Slide 8:
One more value of (s) given by Euler
(0) = 1
0
+ 2
0
+ 3
0
+ = 1 + 1 + 1 + . . . =
1
2
(23)
(s) = (1 2 2
s
)(s) (24)
= (1 2 2
s
)(1
s
+ 2
s
+ 3
s
+ 4
s
+ . . . ) (25)
= 1
s
+ 2
s
+ 3
s
+ 4
s
+ . . .
2 2
s
2 4
s
. . . (26)
= 1
s
2
s
+ 3
s
4
s
+ . . . (27)
The alternating Dirichlet series
1
s
2
s
+ 3
s
4
s
+ . . . (28)
converges for s > 0.
1 1 + 1 1 + 1 =
1
2
(29)
Euler proceeded by giving a striking value (23) for (0). His argumentation is very
important for understanding the main part of this talk.
Namely, Euler considered function (24) dened also by alternating Dirichlet series (27).
For s = 0 the partial sums of this series oscillate around
1
2
and plugging this number as
the value of (0) in (24), one comes to (23).
Slide 9:
Other values of (s) given by Euler
(0) = 1
0
+ 2
0
+ 3
0
+ = 1 + 1 + 1 + =
1
2
(30)
(1) = 1
1
+ 2
1
+ 3
1
+ = 1 + 2 + 3 + . . . =
1
12
(31)
(2) = 1
2
+ 2
2
+ 3
2
+ = 1 + 4 + 9 + = 0 (32)
(3) = 1
3
+ 2
3
+ 3
3
+ = 1 + 8 + 27 + . . . =
1
120
(33)
.
.
.
.
.
.
.
.
.
(11) = 1
11
+ 2
11
+ = 1 + 2048 + . . . =
691
32760
(34)
(m) =
B
m+1
m + 1
m = 0, 1, . . . (35)
(2) = (4) = (6) = = 0 (36)
In similar style Euler obtain other values for negative integer argument. The ap-
pearance of 691 in the numerator of (11) suggests that Bernoulli numbers are again
involved, and Euler gave the general formula (35).
In particular, recalling (16) he got (36).
Slide 10:
The functional equation
(2k) = (1)
k+1
2
2k1
(2k)!
B
2k

2k
k = 1, 2, . . . (37)
(m) =
B
m+1
m + 1
m = 0, 1, . . . (38)
m = 2k 1 (39)
(1 2k) = (1)
k
2
12k

2k
(2k 1)!(2k) k = 1, 2, . . . (40)
Selecting m according to (39), one can eliminate Bernoulli numbers from (37) and (38)
and get (40).
Slide 11:
Euler Identity Fundamental Theorem of Arithmetic
Theorem [Euler].
(s) =
1
1
s
+
1
2
s
+
1
3
s
+ +
1
n
s
+ . . . (41)
=

p is a prime
1
1 p
s
(42)
Proof.

p is a prime
1
1 p
s
=

p is a prime
(1 +
1
p
s
+
1
p
2s
+
1
p
3s
+ . . . ) (43)
=
1
1
s
+
1
2
s
+
1
3
s
+ +
1
n
s
+ . . . (44)
= (s) (45)
In order to obtain (43) from the right-hand side, one needs just to sum up the geo-
metrical progression. Obtaining (43) from (44) is based on the Fundamental Theorem of
Arithmetic saying that every natural numbers can be represented as product of powers of
dierent primes, and this can be done in a unique way (up to the order of factors).
Slide 12:
Riemanns zeta function
Dirichlet series:
(s) =

n=1
n
s
s = + it
5 10 15 20
6
4
2
2
4
6
The series converges in the semiplane Re(s) > 1 and denes a function
that can be analytically extended to the entire complex plane except for
the point s = 1, its only (and simple) pole.
Riemann began to study zeta-function for complex argument and established the tra-
dition to denote it as s = + it.
Slide 13:
Riemanns zeta function
Eulers values of (s) for negative integer s were correct:
(m) =
B
m+1
m + 1
m = 0, 1, . . . (46)
(2) = (4) = (6) = = 0 (47)
Points s
1
= 2, s
2
= 4, . . . , s
k
= 2k, . . . are called the trivial zeroes
of the zeta function, and it has no other real zeroes.
Great Euler was right!
Slide 14:
The functional equation
Euler:
(1 2k) = (1)
k
2
12k

2k
(2k 1)!(2k) k = 1, 2, . . . (48)
Riemann:
(1 s) = cos

s
2

2
1s

s
(s)(s) s = + it (49)
To pass from (48) to (49) one had to guess correct counterpart of (1)
k
.
Slide 15:
Function (s)
(1 s) = cos

s
2

2
1s

s
(s)(s) (50)
cos

s
2

=

(
s
2
+
1
2
)(
s
2
+
1
2
)
(s) =
2
s1

s
2

s
2
+
1
2

(51)

1s
2
(s)

1s
2
+ 1

(1 s)

(1s)
=

s
2
(s 1)

s
2
+ 1

(s)

(s)
(52)
(1 s) = (s) (53)
Function (s) is entire, its zeroes are exactly non-trivial (i.e., non-real)
zeroes of (s).
In the right-hand side of (52), the pole of zeta-function is cancelled by the factor s1,
and the poles of the -factor are cancelled by the trivial zeroes of the zeta-function.
Slide 16:
Function (t)
(s) =

s
2
(s 1)

s
2
+ 1

(s)

g(s)
(54)
s =
1
2
+ it (55)
(t) =

1
2
+ it

1
2
+ it

= (t) (56)
(t) is an even entire function with real value for real values of its
argument t.
The Riemann Hypothesis. All zeroes of (t) are real numbers.
Making the change of the variable (55) is again a tradition due to Riemann. However,
for functions we use modern notation introduced by E. Landau, Riemann himself called
our (t) as (t).
There is no standard notation for the important factor in (54) that well be denot-
ing g(s).
The Riemann Hypothesis was originally stated in terms of function (t), not in terms
of function (s).
Slide 17:
Chebyshev function (x)
(x) = ln(LCM(1, 2, ..., x)) (57)
=
qx

q is a power
of prime p
ln(p) (58)
=
qx

q=1
(q) (59)
Von Mangoldt function
(q) =

ln(p), q = p
k
for some prime p
0, otherwise
(60)
Riemann established a relationship between prime numbers and zeroes of the zeta
function by giving a (somewhat complicated) expression for (x) the number of primes
below x via certain sums over these zeroes. Chebyshev function (x) turned out to be
more convenient for studding the distribution of prime numbers.
Slide 18:
Logarithmic derivative of (s)
(s) =

p is a prime
1
1 p
s
(61)

(s)
(s)
=

p is a prime

ln

1
1 p
s

(62)
=

p is a prime
ln(p)
1 p
s
(63)
=

p is a prime
ln(p)

k=1
p
ks
(64)
=

n=1
(n)n
s
(65)
Von Mangoldt function plays important role in the theory of the zeta-functions because
the values of the former functions are nothing else but (negations of) the coecients in
the Dirichlet series of the logarithmic derivative of the latter function.
Slide 19:
Chebyshev function (x)
0 5 10 15 20
5
10
15
20
The jumps occur at primes and at their powers.
Slide 20:
Von Mangoldt Theorem
(x) = ln(LCM(1, 2, ..., x)) =
x

q=1
(q) (66)
Theorem (Hans Carl Fridrich von Mangoldt [1895]). For any
non-integer x > 1
(x) = x

()=0
x

n
x
2n
2n
ln(2) (67)
Theorem (Jacques Salomon Hadamard [1893]).
(s) = (0)

()=0

1
s

(68)
(s) = g(s)(s) = g(s)

p is a prime
1
1 p
s
(69)
The proof of von Mangoldt theorem is based on the appearance of the two products
in (68) and (69).
Slide 21:
Von Mangoldt Theorem
(x) x
||<400

()=0
x

n
x
2n
2n
ln(2) (73)
5 10 15 20
5
10
15
20
Traditionally, (67) is viewed as an identity between well-dened left and right hand
sides, but it can be interpreted in a dierent way. Imagine that we know nothing about
prime numbers, even their denition, but have at our disposal suciently many initial
zeroes of the zeta function. In such a case we could discover prime numbers just by
looking at the plot of the truncated right hand side of (67) as in (73). The prime numbers
can be revealed from such a plot either by looking at integers in the vicinity of which the
function has a big jump (they will be powers of primes) or by looking at the sizes of jumps
(they will be close to natural logarithms of primes). To reveal more and more primes we
would need to use more and more zeroes of the zeta function.
Slide 22:
Blowing the Trumpet
Assuming that all zeroes of (t) are real and simple, let them be denoted

1
,
2
, . . . with 0 <
1
<
2
< . . . , thus the non-trivial zeroes of (s)
are
1
2
i
1
,
1
2
i
2
, . . .

1

2

3

5

6

6

7

1
= 14.1347 . . .
2
= 21.0220 . . . (74)

3
= 25.0109 . . .
4
= 30.4249 . . . (75)
Suppose that we have found
1
,
2
, . . . ,
N1
; how could these numbers
be used for calculating an (approximate) value of the next zero
N
?
It was a rather strange idea to seek an answer to this question because known initial
zeroes are distributed rather irregularly.
Slide 23:
Interpolating determinant
Let us try to approximate (t) by some simpler function also having zeroes
at the points
1
, . . . ,
N1
.
Consider an interpolating determinant with even functions f
1
, f
2
, . . .

f
1
(
1
) . . . f
1
(
N1
) f
1
(t)
.
.
.
.
.
.
.
.
.
.
.
.
f
N
(
1
) . . . f
N
(
N1
) f
N
(t)

(76)
Selecting f
n
(t) = t
2(n1)
we would obtain just an interpolating polynomial
C
N1

n=1
(t
2

n
2
) (77)
having no other zeros.
Clearly, the determinant (76) vanishes for t =
1
, . . . ,
N1
because for such values
of t the matrix contains two equal columns.
Slide 24:
Modelling situation
If

1
,

2
, . . . are zeros of the function

(t) =
N

k=1
f
k
(t) (78)
then the determinant

f
1
(

1
) . . . f
1
(

N1
) f
1
(t)
.
.
.
.
.
.
.
.
.
.
.
.
f
N
(

1
) . . . f
N
(

N1
) f
N
(t)

(79)
vanishes at every zero of

(t).
The determinant (79) vanishes as soon as t is equal to any zero of

(t) because for

such a t the rows of the matrix sum up to the zero row.
Slide 25:
Selecting f
1
, f
2
, . . .
(t) =

n=1

n
(t)
n
(t) = g

1
2
+it

n
(
1
2
+it)
(80)
(t) =

n=1

n
(t) for Im(t) <
1
2
(81)
(t) = (t) =

n=1

n
(t) for Im(t) >
1
2
(82)
(t) =

n=1

n
(t) +
n
(t)
2
=

n=1

n
(t) (83)
Our case diers from the modelling one in two respects. First, the sum in (80) is
innite. Second, its summands arent even functions.
To remove the rst dierence we introduce even functions
n
(t). However, the equal-
ity (83) is purely symbolic because the half planes of convergence of the two series (81)
and (82) dont intersect.
Slide 26:
Our interpolating determinant

N
(t) =

1
(
1
) . . .
1
(
N1
)
1
(t)
.
.
.
.
.
.
.
.
.
.
.
.

N
(
1
) . . .
N
(
N1
)
N
(t)

(84)

n
(t) =
g

1
2
it

n
(
1
2
it)
+g

1
2
+it

n
(
1
2
+it)
2
(85)
g(s) =

s
2
(s 1)

s
2
+1

(86)
(t) =

n=1

n
(t) (87)
For determinant we need only nitely many functions, so for the moment we can ignore
the divergence of (87).
Slide 27:
Blowing the trumpet: the outcome

47
( 138.11604)=2.18497 . . . 10
1216
< 0

47
= 138.1160420545334. . .

47
( 138.11605)=+4.68242 . . . 10
1216
> 0

321
( 572.419984132452764045276927)=1.75211 . . . 10
34969
< 0

321
= 572.41998413245276404527692710734686. . .

321
( 572.419984132452764045276928)=+5.72682 . . . 10
34968
> 0

600
(939.02430089921838218421841870917956106917\
858909398785769363406563371 = +3.90403 . . . 10
108715
> 0

600
= 939.02430089921838218421841870917956106917\
85890939878576936340656337113493189 . . .

600
(939.02430089921838218421841870917956106917\
858909398785769363406563372) = 9.67471 . . . 10
108716
< 0
Here are just a few numerical examples.
A zero of
47
(t) has 8 decimal digits coinciding with digits of
47
.
For N = 321 there are already 15 coinciding decimal digits.
For N = 600 the number of common digits increases to 38.
Slide 28:
Blowing the trumpet: the outcome

47
( 138.11604)=2.18497 . . . 10
1216
< 0

47
= 138.1160420545334. . .

47
( 138.11605)=+4.68242 . . . 10
1216
> 0

47
( 139.7362)=+1.27744 . . . 10
1216
> 0

48
= 139.736208952121. . .

47
( 139.7363)=9.88309 . . . 10
1216
< 0

47
( 141.12370)=1.85988 . . . 10
1217
< 0

49
= 141.1237074040211. . .

47
( 141.12371)=+2.40777 . . . 10
1217
> 0

47
( 143.11184)=+8.00594 . . . 10
1218
> 0

50
= 143.1118458076206. . .

47
( 143.11185)=8.98353 . . . 10
1218
< 0
It turned out that
N
(t) allows us to calculate good approximations not only to the
next not yet used zero
N
but to
N+k
as well for values of k that are not too large.
Slide 29:
Blowing the trumpet: the outcome

321
( 572.419984132452764045276927)=1.75211 . . . 10
34969
< 0

321
= 572.41998413245276404527692710734686. . .

321
( 572.419984132452764045276928)=+5.72682 . . . 10
34968
> 0

321
( 573.61461052675812998724603)=+2.27756 . . . 10
34967
> 0

322
= 573.6146105267581299872460321096433. . .

321
( 573.61461052675812998724604)=9.27805 . . . 10
34967
< 0

321
( 575.09388601449488560745717)=5.83758 . . . 10
34968
< 0

323
= 575.0938860144948856074571722821982. . .

321
( 575.09388601449488560745718)=+2.78750 . . . 10
34967
> 0
...............................................................................................................

321
( 585.98456320498830057418)=5.42229 . . . 10
34968
< 0

331
= 585.9845632049883005741872589855. . .

321
( 585.98456320498830057419)=+8.57284 . . . 10
34969
> 0
These are just a few examples, tables showing the number of digits common to
(N + k)th zero of (t) and a zero of
N
(t) for diverse N and k can be found at
http://logic.pdmi.ras.ru/~yumat/personaljournal/artlessmethod.
Slide 30:
Slide removed
I dont have a full explanation of the high accuracy of zeroes of determinants
N
(t)
as approximations to zeroes of (t). Two heuristic hints will be presented now.
Slide 31:
Partial explanation. I
(t) =

n=1

n
(t) (88)

N
(t) =

1
(
1
) . . .
1
(
N1
)
1
(t)
.
.
.
.
.
.
.
.
.
.
.
.

N
(
1
) . . .
N
(
N1
)
N
(t)

=
N

n=1

N,n

n
(t) (89)

N,n
= (1)
N+n

1
(
1
) . . .
1
(
N1
)
.
.
.
.
.
.
.
.
.

n1
(
1
) . . .
n1
(
N1
)

n+1
(
1
) . . .
n+1
(
N1
)
.
.
.
.
.
.
.
.
.

N
(
1
) . . .
N
(
N1
)

(90)
The

s are particular numbers dened via the zeroes of the zeta-function.
Slide 32:
Normalization
(t) =

n=1

n
(t) (91)

N
(t) =

1
(
1
) . . .
1
(
N1
)
1
(t)
.
.
.
.
.
.
.
.
.
.
.
.

N
(
1
) . . .
N
(
N1
)
N
(t)

=
N

n=1

N,n

n
(t) (92)

N,n
=

N,n

N,1

N,1
= 1 (93)

N
(t) =
N

n=1

N,1

N,n

n
(t) =

N,1
N

n=1

N,n

n
(t) (94)
The normalization doesnt inuence zeroes, that is,
N
(t) has the same zeroes as

N
n=1

N,n

n
(t).
Slide 33:
Normalized coecients
47,n
47

n=1

47,n

n
(t) (95)
We the sum (95) is not a sharp but a smooth truncation of the divergent series (87).
It is known that smooth truncation can accelerate convergence of a series and it can even
transform a divergent series into a convergent one.
The smoothness of the truncation in (95) might be the rst reason why the summands
of the divergence series (87) are useful for calculation of the zeroes.
Slide 34:
Normalized coecients
47,n
with logarithmic scale

47,n
1 +
47
log(n) (96)
The curve from Slide 33 on which the coecients
47,n
lie looks like a logarithmic
curve.
This is better seen on the plot of the same coecients but with logarithmic scale.
Why do the initial coecients lie approximately on a straight line?
Slide 35:
Normalized coecients
48,n
with logarithmic scale

48,n
1 +
48
log(n) (97)
The graphs for N = 48, 49, 50, 51 look very similar to the graph for N = 47 from
Slide 34.
Slide 36:
Normalized coecients
52,n
with logarithmic scale

52,n
1 +
52
log(n) (101)
Normalized coecients
52,n
show slightly dierent behaviour of trailing coecients.
Slide 37:
Normalized coecients
54,n
with logarithmic scale

54,n
1 +
54
log(n) (103)
What a dierence with previous graphs! Now most of the coecients are negative and
greater than 1 in absolute value. Nevertheless, the initial coecients do lie on a straight
line and
54
(t) gives good predictions for a few of the next zeroes
54
,
55
, . . .
Slide 38:
Normalized coecients
55,n
with logarithmic scale

55,n
1 +
55
log(n) (104)
PictuGrapfs for N = 55, 56, 57 look similar to the graph for N = 47. Case N = 54
demonstrates sporadic behavour. Why number 54 is special?
Slide 39:
Normalized coecients
130,n
with logarithmic scale

130,n
1 +
130,2
dom
2
(n) +
130
log(n) (107)
dom
m
(k) =

1, if m | k
0, otherwise
(108)
We skip many values of N (the skipped and many other graphs can be found at http:
//logic.pdmi.ras.ru/~yumat/personaljournal/artlessmethod) and examine a new
phenomenon in the case N = 130.
Now the initial coecients lie on two parallel lines rather than on a single line, so
instead of an analog of (96), (97), (101), (103) and (104), in the case N = 130 we should
use approximation (107).
Slide 40:
Normalized coecients
131,n
with logarithmic scale

131,n
1 +
131,2
dom
2
(n) +
131
log(n) (109)
Graph for N = 131 shows that the case N = 130 is sporadic.
Slide 41:
Normalized coecients
169,n
with logarithmic scale

169,n
1 +
169,2
dom
2
(n) +
169
log(n) (111)
But with increasing N splitting into two lines becomes typical.
Slide 42:
Normalized coecients
321,n

321,n
(1)
n1
(113)
1 +
321,2
dom
2
(n) with
321,2
2 (114)
Skipping again many pictugraphs, we consider new pattern exhibited for N = 321
now initial coecients alternate between approximately +1 and 1.
Slide 43:
Partial explanation. II
The function
321
(t)/

321,n
=

321
n=1

321,n

n
(t) is a smooth truncation,
not of the divergent Diriclet series
(t) =

n=1

n
(t) (115)
but of the convergent (for real t) alternating series

n=1
(1)
n1

n
(t) =

n=1
(1)
n1

n
(t) +
n
(t)
2
(116)

n=1
(1)
n1

n
(t) =

n=1
(1)
n1
g
_
1
2
+
it
2
_
n
(
1
2
+it)
(117)
= g
_
1
2
+
it
2
_

n=1
(1)
n1
n
(
1
2
+it)
. .
(
1
2
+it)
= g
_
1
2
+
it
2
_
_
1 2 2
(
1
2
it)
_
(
1
2
+it)
Euler used function (s) to assign a value to the zeta function outside the area of con-
vergency of the Dirichlet series, and that convergence gives a second reason (besides the
smoothness of the truncation) why
321
is so good for prediction the values
321
,
322
, . . .
It is very remarkable that here the function (s) emerges by itself, just from our
calculation of the determinants
N
(t) as though they were as clever as Euler was.
Still some natural questions remain open. The switching from divergent (115) to
convergent (116) partly explains why
321
(t) is able to predict the values of
321
and
further zeroes of (t) but why does this also happen for, say, N = 47? The series (117)
converges on the critical line rather slowly, so why are the zeroes of
321
(t) so close to
those of (t)?
Slide 44:
Functions
N
(s)
(s) =

n=1
(1)
n1
n
s
(119)

N
(s) =
N

n=1

N,n
n
s
(120)

N
(t) =
1
2

N,1

g

1
2
it

1
2
it

+ g

1
2
+ it

1
2
+ it

(121)

N
(s)
?
(s) = (1 2 2
s
)(s) (122)
(s)
?

N
(s)
1 2 2
s
(123)
Based on (113), we introduce functions
N
(s) mimicking Eulers function (s).
Shouldnt (122) be true? If yes, than well be able to calculate (s) via
N
(s).
Slide 45:
Calculation of (s)
For s =
1
2
+ 100i

(s)

321
(s)
122
s
1

= +1.34... 10
12
(124)

(s)

600
(s)
122
s
1

= 2.95... 10
24
(125)
For s =
1
4
+ 100i

(s)

600
(s)
122
s
1

= 2.64... 10
24
(126)

(s)

3000
(s)
122
s
1

= +1.19... 10
125
(127)
For s =
1
4
+ 1000i

(s)

3000
(s)
122
s
1

= +4.07... 10
126
(128)
Indeed, such calculations turned out to be possible. For N = 321 we get 12 correct
decimal digits of (s), and 24 correct decimal digits in the case N = 600. However, earlier
(Slides 27 and Slide 29) we got for the same values of N much more correct decimal digits
of the s. How is it possible?
Slide 46:
Normalized coecients
321,n
Let us have a closer look at the areas marked in yellow.
Slide 47:
Normalized coecients
321,n

321,n
1 +
321,2
dom
2
(n) +
321,3
dom
3
(n) +
321
log(n) (129)

321,2
= 2 4.98 . . . 10
13

321,3
= 7.47 . . . 10
13
(130)

321
= 3.33 . . . 10
18
(131)
We observe that actually the plot splits into 4 parallel lines so the initial normalized
coecients are better approximated by (129).
Slide 48:
Normalized coecients: general case

N,n

M
m=1

N,m
dom
m
(n) +
N
log(n) (132)

N,n
=

n
m=1

N,m
dom
m
(n) =

m|n

N,m
n = 1, 2, . . . (133)

N,n
=

m|n

n
m

N,n
(134)
Mobius function
(k) =

(1)
m
, if k is the product of m dierent primes
0, otherwise
(135)
1
(s)
=
1

n=1
n
s
=

n=1
(n)n
s

n=1
n
s

n=1
(n)n
s
= 1 (136)
In fact, there are similar even smaller splitting depending on the values of n modulo
4, 5, . . . , so we need to use (132).
To give formal denition of
N,n
we ignore the logarithmic term in (132) and demand
that it to be exact equality when M = n. Resulting triangular linear system (133) has
solution (134).
The M obius function also play important role in the theory of the zeta-functions due
to (136).
Slide 49:
Function
N,M
(s)
M
3000,M
1 1
2 2 4.9... 10
126
3 7.40565... 10
126
4 2.85890... 10
284
5 1.85782... 10
411
6 2.56167... 10
509
7 5.02202... 10
585
8 4.39701... 10
641
9 1.08444... 10
681
10 1.90599... 10
716

N
(s)
(s)
=

N
n=1

N,n
n
s

n=1
n
s
(137)
=

n=1

N,n
n
s
(138)

N
(s) =

n=1

N,n
n
s

(s) (139)

N
(s) (1 2 2
s
)(s)
N
(s)

n=1

N,n
n
s

(s)

N,M
(s)
(s) (140)
Equivalently, numbers
N,n
can be dened as the coecients in the ratio (137).
The small numerical values of
3000,n
imply that between the initial
3000,n
there is a
number of almost linear relations of the form (134).
To what extent can function
N,M
(s) play the role of Eulers factor 1 2 2
s
?
Slide 50:
Improved calculation of (s) for s =
1
4
+ 1000i

N
(s)
N,M
(s)(s)
N,M
(s) =

M
n=1

N,n
n
s
(141)
M
3000,M

N
(s)

3000,M
(s)(s)
1

2 2 4.9... 10
126
2.46329... 10
126
3 7.40565... 10
126
8.84947... 10
285
4 2.85890... 10
284
5.43870... 10
412
5 1.85782... 10
411
7.16503... 10

510
6 2.56167... 10
509
1.35156... 10
585
7 5.02202... 10
585
1.14450... 10
641
8 4.39701... 10
641
2.13729... 10
681
9 1.08444... 10
681
2.18430... 10
681
10 1.90599... 10
716
2.18430... 10
681
11 2.37291... 10
681
2.56047... 10
681
For M = 2 we have almost Euler-like approximation (123).
Increasing M up to 8 greatly improves the accuracy.
But further increasing the value of M doesnt lead to better approximations. The
reason is as follows: decreasing of (absolute values of)
3000,M
stops at M = 10.
Slide 51:
log
10
(|
3000,n
|), magenta, if
3000,n
> 0, green otherwise
For n = 2..10 the absolute values of
3000,n
decrease, and their logarithms lie (why?)
on a smooth curve in spite of the fact that the values themselves have dierent signs.
Starting from n = 11 the absolute values of
3000,n
oscillate between approximately
10
681
and 10
752
. The values of n corresponding to the rst of these two cases are all
primes or powers of primes, and corresponding values of
3000,n
are all positive.
Slide 52:

3000,n
, magenta, if
3000,n
> 0, green otherwise
Closer look at values of
3000,n
for prime and prime power values of n reveals ner
structure.
Slide 53:

3000,n
ln(n)
, magenta, if
3000,n
> 0, green otherwise
This structure becomes more transparent when we normalize the s by ln(n).
Slide 54:

3000,n
(n)

3000
=

3000,13
ln(13)
= 9.895811... 10
682
(142)

3000,p
k / ln(p)

3000
1

< 3.85... 10
73
for 13 p
k
419, p is a prime

3000,p
k
ln(p)

3000

< 3.81... 10
754
for 13 p
k
419, p is a prime
It is even better to normalize by von Mangoldts (n) (dened on Slide 17).
For a long range of prime and prime power values of n the value of the ratio

3000,n
(n)
remains almost constant. For non-prime-power values of n in the same range the val-
ues of |
3000,n
| are by more than 70 decimal orders smaller than
3000
, that is,
3000,n
approximates very well (a multiple of) von Mangoldt function (n).
The connections between zeta-function and prime numbers is usually attributed to
Euler product (42), however, in this case it isnt clear how to establish connection between
Euler product and the values of
3000,n
in this range.
Another intriguing question is: why the values of
3000,n
for a few initial values of n
are so dierent? Their structure was analyzed to some extent in
http://logic.pdmi.ras.ru/~yumat/personaljournal/artlessmethod/
talks/leicester2012/leicester_2012_full.pdf.
Slide 55:

3000
(s)
(s)
=

n=1

3000,n
n
s
(143)
This is just a denition of numbers
3000,n
(case N = 3000 of (138)).
Slide 56:

3000
(s)
(s)

8

n=1

3000,n
n
s

3000

n=1
(n)n
s
(145)
?

3000

(s)
(s)
Filling in the table on Slide 50, we dropped all summands in (143) but a few initials,
say, the rst eight.
Slide 57:

3000
(s)
(s)

11

n=1

3000,n
n
s

3000

n=1
(n)n
s
(145)
?

3000

(s)
(s)
Let us now do the opposite drop initial 11 summands from (143)
Slide 58:

3000
(s)
(s)

11

n=1

3000,n

3000
(n)


3000,n

n
s

3000

n=1
(n)n
s
(145)
?

3000

(s)
(s)
and then add a few missing summands for n = 2, 3, 4, 5, 7, 8, 9, 11. Taking into account
the fact that for non-prime-power values of n the values of |
3000,n
| are by more than
70 decimal orders smaller than
3000
, we see that the initial summands in the left-hand
side of (145) (with

3000
(s)
(s)
replaced by the right-hand side of (143)) are very close to the
initial summands in the right-hand side. Shouldnt we expect approximate equality of the
numerical values of the left- and right-hand sides of (145)?
Recall that according to (65), the right-hand side of (145) is just a multiple of the
logarithmic derivative of the zeta function.
Slide 59:
Calculating zeta derivative at zeros

3000
(s)
(s)

11

n=1

3000,n
n
s
?

3000

(s)
(s)
(147)

3000
(s)

11

n=1

3000,n
n
s

(s)
?

3000

(s) (148)

3000

1
2
+ i
k

?

3000

1
2
+ i
k

(149)

3000

1
2
+ i
100

3000

1
2
+ i
100
1

= 1.024... 10
36
(150)

3000

1
2
+ i
500

3000

1
2
+ i
500
1

= 2.786... 10
74
(151)
Let us check our guess at rst at a pole of (s) when (148) simplies to (149). Indeed,
we get a lot of correct digits. This is rather surprising by two reasons. First, we calculate
outside the semiplane Re(s) > 1 of the convergence of (65). Second, passing from (147)
to (148) we multiplied both sides by (s) but in our case its value is zero.
Slide 60:
Calculating zeta derivative

3000
(s)

11

n=1

3000,n
n
s

(s)
?

3000

(s) (152)
s =
1
4
+ 1000i (153)

3000
(s)

11
n=1

3000,n
n
s

(s)

3000

(s)
1

= 6.44... 10
73
(154)
We can get many digits of

(s) for s dierent from a pole as well. However, for this

we need to know the value of (s) with much higher accuracy than that from table on
Slide 50.
Slide 61:
Calculating both zeta and its derivative. I

3000
(s)

11

n=1

3000,n
n
s

(s)
3000

(s) (155)

3500
(s)

11

n=1

3500,n
n
s

(s)
3500

(s) (156)
Solving this system for s =
1
4
+ 1000i produces 908 correct decimal digits
for (s) and 72 correct decimal digits for

(s).
Nevertherles, we can calculate both (s) and

(s) via the s and s. One way is to

use two values of N.
Slide 62:
Calculating both zeta and its derivative. II

3000
(s)

11

n=1

3000,n
n
s

(s)
3000

(s) (157)

3000
(1 s)

11

n=1

3000,n
n
s1

(1 s)
3000

(1 s) (158)
g(s)(s) = g(1 s)(1 s) (159)
g

(s)(s) + g(s)

(s) = g

(1 s)(1 s) g(1 s)

(1 s) (160)
Solving this system for s =
1
4
+ 1000i produces 752 correct decimal digits
for (s) and 72 correct decimal digits for

(s).
Another way is to use the functional equation relating the values of (s) and

(s)
with values of (1 s) and

(1 s).
Slide 63:
Near the pole
(s) = (1 2 2
s
)(s) (161)

N
(s)
N,M
(s)(s)
N,M
(s) =
M

n=1

N,n
n
s
(162)
M
3000,M

3000,M
(1)
2 2 4.9... 10
126
2.46855... 10
126
3 7.40566... 10
126
7.14726... 10
285
4 2.85891... 10
284
3.71565... 10
412
5 1.85782... 10
411
4.26945... 10
510
6 2.56167... 10
509
7.17431... 10
586
7 5.02202... 10
585
5.49626... 10
642
8 4.39701... 10
641
5.73467... 10
681
9 1.08444... 10
681
5.61417... 10
681
10 1.90599... 10
716
5.61417... 10
681
11 2.37291... 10
681
5.39845... 10
681
The role played by Eulers factor 1 2 2
s
in (161) was to cancel the pole of the zeta
function. To what extent our factor
N,M
(s) plays this role?
The small values of the s imply almost linear relation between the s, the coecients
being rational numbers with small denominators.
Slide 64:
Near the pole
(s) = (1 2 2
s
)(s) (161)

N
(s)
N,M
(s)(s)
N,M
(s) =
M

n=1

N,n
n
s
(162)
(1) = lim
s1
(1 2 2
s
)(s) (163)
= lim
s1
(1 2 2
s
)

1
s 1
+ . . .

(164)
= (1 2 2
s
)

s=1
(165)

N
(1)

N,M
(1) (166)

3000
(1)

3000,9
(1)
= 1 4.1515... 10
680
(167)
In spite of the fact that function
3000,M
(s) doesnt vanish as s = 1, its derivative
approximates
3000
(1) very well.
Slide 65:
log
10
(|
3000,n
|) (repeated)
We have analyzed the ne structure of numbers
N,n
for prime and prime power values
of n. Is there any interesting pattern for other values of n? For example, for such n the
value of
N,n
can be both positive and negative, and the sign depends on both n and N.
Is there any simple rule for determining the signs?
Clearly,
3000,n
cannot distinguish prime powers from the rest of integers for all n,
Slide 66:
log
10
(|
3000,n
|), magenta, if
3000,n
> 0, green otherwise
so let us look at larger values of n.
Slide 67:
log
10
(|
3000,n
|), magenta, if
3000,n
> 0, green otherwise
While after n = 420 the absolute values of
3000,n
for non-prime-power values of n
begin to grow up, for prime power values of n the values of
3000,n
continue to stay close
to 10
681
. We can clearly see points for n = 512, 625, 729.
Why the relatively slow growth of |
3000,n
| suddenly changes to much faster one after
n = 791?
Why after n = 797 there is no distinction between prime power and non-powers?
Slide 68:
log
10
(|
3000,n
|), magenta, if
3000,n
> 0, green otherwise
Slide 69:
log
10
(|
3000,n
|) (repeated)

3000,n
(3000)(n) (x) =
x

n=1
(n) (168)
Could we use the s for studding the growth of (x)? To this end we would need to
estimate certain sums with
N,n
.
Slide 70:

N,M
(s)

N,M
(s) =
M

n=1

N,n
n
s

N,n
=

m|n

n
m

N,n
(169)

N,n
=

N,n

N,1

N,n
= (1)
N+n

1
(
1
) . . .
1
(
N1
)
.
.
.
.
.
.
.
.
.

n1
(
1
) . . .
n1
(
N1
)

n+1
(
1
) . . .
n+1
(
N1
)
.
.
.
.
.
.
.
.
.

N
(
1
) . . .
N
(
N1
)

(170)

n
(t) = g

1
2
it

n
(
1
2
it)
+ g

1
2
+ it

n
(
1
2
+it)
(171)
g(s) =

s
2
(s 1)

s
2
+ 1

0 <
1
<
2
< . . . (172)
Can we expect interesting properties of individual values of numbers
N,M
(s) dened
in this involved way?
Slide 71:

N,N
(0)

N,M
(s) =
M

n=1

N,n
n
s
(173)

N,M
(0) =
M

n=1

N,n
(174)

3000,3000
(0) =
1
2
+ 1.23... 10
126

3001,3001
(0) =
3
2
+ 6.67... 10
126
(175)

6000,6000
(0) =
1
2
5.48... 10
166

6001,6001
(0) =
3
2
3.43... 10
166
(176)
Numbers
N,N
(0) seem to be close to 1 +
(1)
N
2
.
Slide 72:

N,N
(1)

N,M
(s) =
M

n=1

N,n
n
s

N,M
(1) =
M

n=1

N,n
n
(177)
N = 2520 = LCM(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) (178)
2N
N,N
(1) = 0.9998015873172093... (179)
=
1
1 +
1
5039+
1
2520+
1
1680+
1
1260+
1
1008+
1
840+
1
720+
1
630+
1
560+
1
504+
1
.
.
.
(180)
5039 = 2N 1, 2520 =
2N
2
, 1680 =
2N
3
, 1260 =
2N
4
, 1008 =
2N
5
,
840 =
2N
6
, 720 =
2N
7
, 630 =
2N
8
, 560 =
2N
9
, 504 =
2N
10
In all other examples the choice of N wasnt crucial, neighbouring values would suit
equally well. Here neither N = 2519 nor N = 2521 would do, it is important that 2 2520
is divisible by 2,3,. . . ,10. Similar, the pattern doesnt continue because 2 2520 isnt
divisible by 11.
Also, this is the only place where the normalization (93) is crucial.
Slide 73:

N,N
(1)
(N) =
1
2N

1
1 +
1
2N 1 +
1
2N
2
+
1
2N
3
+
1
2N
4
+
1
2N
5
+
1
2N
6
+
.
.
.
(181)
=
1
2

N
2
+ 1

1
2

N + 1
2

(182)
(z) =

(z)
(z)
(183)
The equivalence of the two denitions of (N) was established by comparing a few
initial terms in the asymptotic expansions, I dont have a rigorous proof.
Slide 74:

N,N
(1)

N,M
(s) =
M

n=1

N,n
n
s

N,M
(1) =
M

n=1

N,n
n
(184)
(N) =
1
2

N
2
+ 1

1
2

N + 1
2

(z) =

(z)
(z)
(185)

3000,3000
(1)
(3000)
= 1 2.46827839... 10
126
(186)

6000,6000
(1)
(6000)
= 1 + 1.09736771... 10
165
(187)
Slide 75:
Normalized coecients
500,n
This graph looks similar to the case N = 321 on Slide 42.
Slide 76:
Normalized coecients
1200,n
And this pictugraph looks similar to the case N = 321 on Slide 42.
Slide 77:
Normalized coecients
3000,n
Again similar graph.
Slide 78:
Normalized coecients
505,n
This is a sporadic graph.
Slide 79:
Normalized coecients
621,n
And this graph is sporadic.
Slide 80:
Normalized coecients
810,n
Notice multiples change of sign.
Slides 81-87 were removed.
Slide 88:
log
10
(|
47,n
|), magenta, if
47,n
> 0, green otherwise
The s show two typical behaviour:
logarithmic for small" N like on Slide 33,
parallel for medium N like on Slide 42,
The s behave more uniformly. Already for N as small as 47 we can distinguish primes
and their powers.
Slide 89:
log
10
(|
321,n
|), magenta, if
321,n
> 0, green otherwise
Here we can see n = 25, 27, 32, 49, 64, 81, 121, 125, 128.
Slide 90:
log
10
(|
3000,n
|) (repeated)
Slides 9195 were removed.
I found very interesting continuation of calculations for N > 12000 but this exceeds
the computational resources I have at my disposal at present, international cooperation
seems to be necessary.
Also it is very interesting to see if there are similar phenomena for other L-functions.
We saw that for Riemann zeta function many things are connected with its pole, so,
probably, we should deal not with individual Dirichlet L-functions but with the product
of all of them corresponding to xed prime p.
The determinants
N
(t) were dened under the assumptions that all zeroes satisfy
Riemanns Hypothesis and are simple. What would happen if Riemanns Hypothesis is
valid but there are multiple zeroes? It might completely spoil considered above methods
for calculating the values of the zeta function, its derivative and zeroes in the case of
large N. More likely, it will only be necessary to modify the denitions (84) and (90) by
replacing columns with repeated values of zeroes by columns with derivatives of
n
. But
maybe there is no need in such modication and proving it would give a proof (under
RH) of the simplicity of all zeroes.
Possible non-real zero

of (t) wouldnt spoil denitions (84) and (90) but would

require making a decision which of the two zeroes,

or

, should be used rst. When

N is such that odd number of non-real zero of (t) are used in (90), the values of
N,n
and
N,n
will become in general non-real making (n)-like behaviour of
N,n
doubtful.
Does this observation open a way to proving RH?
Slide 96:
Welcome to
http://logic.pdmi.ras.ru/~yumat/personaljournal/artlessmethod
More graphs, tables, texts can be found on this site devoted to the ongoing research.