Brent - 1994 - Ramanujan and Euler's Constant

Ramanujan and Euler’s Constant
Richard P. Brent
MSI, ANU
8 July 2010
In memory of Ed McMillan
1907 – 1991
Presented at the CARMA Workshop on Exploratory

Experimentation and Computation in Number Theory,
Newcastle, Australia, 7–9 July 2010.
Copyright
c 2010, R. P. Brent
Richard Brent Ramanujan and Euler’s Constant
Motivation
Ramanujan gave many beautiful formulas for π and 1/π. See,

for example, J. M. Borwein and P. B. Borwein, Pi and the AGM,
John Wiley and Sons, New York, 1987; also (same authors)
“Ramanujan and Pi”, Scientific American, February 1988,
66–73.
Euler’s constant
γ = −Γ0 (1) ' 0.577
is more mysterious than π. For example, unlike π, we do not
know any quadratically convergent iteration for γ. We do not
know if γ is transcendental. We do not even know if γ is
irrational, though this seems likely. All we know is that if
γ = p/q is rational, then q is large. This follows from a
computation of the regular continued fraction expansion for γ.

Analogy with ζ(3)
Apéry proved ζ(3) irrational using the series

∞
5 X (−1)k −1 k !k !
ζ(3) =
2 (2k )!k 3
k =1
and, in Chapter 9 of his Notebooks, Ramanujan gives several

similar series, some involving ζ(3).

Ramanujan rediscovered Euler’s formula
∞
X Hk
ζ(3) = ,
(k + 1)2
k =1
where
k
X 1
Hk =
j
j=1
is a Harmonic number. Harmonic numbers also occur in

formulas involving γ (examples later).
Thus, it is natural to look in the work of Ramanujan for formulas
involving γ, in the hope that some of these might be useful for
computing accurate approximations to γ, or even for proving
that γ is irrational.

Ramanujan’s Papers and Notebooks
Ramanujan published one paper specifically on γ: “A series for
Euler’s constant γ”, Messenger of Mathematics 46 (1917),
73–80. In this paper he generalizes an interesting series which
was first discovered by Glaisher:
∞
X ζ(2k + 1)
γ =1− .
(k + 1)(2k + 1)
k =1
This family of series all involve the Riemann zeta function or

related functions, so they are not convenient for computational
purposes.
Much of Ramanujan’s work was not published during his
lifetime, but was summarized in his Notebooks. Edited editions
have been published by Berndt [1]. In the following, page
numbers refer to Berndt’s edition (Part I for Chapters 1–9, Part
II for Chapters 10–15).
γ in Ramanujan’s Notebooks
Scanning Berndt, we find many occurrences of γ. Some involve

the logarithmic derivative ψ(x) of the gamma function, or the
sum
Xx
Hx = 1/k ,
k =1
which we can interpret as ψ(x + 1) + γ if x is not necessarily a

positive integer (Ch. 8, pg. 181). There are also applications of
the result
Hn = ln n + γ + O(1/n)
as n → ∞.
Other interesting formulas involving γ occur in Chapters 14–15,
e.g. Ch. 15, Entry 1, examples (i–ii), pp. 303–304.

Chapter 4, Entry 9
We shall concentrate on Chapter 4, Entry 9, Corollaries 1–2
(pg. 98), because these are potentially useful for computing γ.
Corollary 1 is
∞
X (−1)k −1 x k
∼ ln x + γ (1)
k !k
k =1
as x → ∞. In fact, Euler showed that

∞ ∞
(−1)k −1 x k e−t e−x
X Z
− ln x − γ = dt = O
k !k x t x
k =1
and this has been used by Sweeney and others to compute

Euler’s constant (one has to be careful because of cancellation
in the series). In Ch. 12, Entry 44(ii), Ramanujan states Euler’s
result that the error is between e−x /(1 + x) and e−x /x.

A Generalization
Ramanujan’s Corollary 2, Entry 9, Chapter 4 (page 98) is that,

for positive integer n,
∞ n
X (−1)k −1 x k
∼ ln x + γ (2)
nk k!
k =1
so (1) is just the case n = 1.

Berndt shows that (2) is false for n ≥ 3. In fact, the function
defined by the left side of (2) changes sign infinitely often, and
grows exponentially large as x → ∞. However, Berndt leaves
the case n = 2 open.

We shall sketch a proof that (2) is true in the case n = 2. In
fact, we shall obtain an exact expression for the error in (2) as
an integral involving the Bessel function J0 (x), and deduce an
asymptotic expansion.
The exact expression for n = 2 is a special case of a formula
given on page 48 of Y. L. Luke, Integrals of Bessel Functions,
1962. However, Luke does not comment on the connection with
Ramanujan.

Avoiding Cancellation
In Chapter 3, Entry 2, Cor. 2, page 46, Ramanujan states that
the sum
∞
X (−1)k −1 x k
k !k
k =1
occurring in (1) can be written as
∞
X xk
e−x Hk .
k!
k =0
This is easy to prove (Berndt, page 47). Thus (1) gives

∞ ∞
X xk X xk
Hk ∼ ln x + γ. (3)
k! k!
k =0 k =0
This is more convenient than (1) for computation, because

there is no cancellation in the series when x > 0. Later we
indicate how Ramanujan might have generalized (3) in much
the same way that he attempted to generalize (1).
Ramanujan’s Corollary for n = 2
The following result from [3] shows that (2) is valid for n = 2.
Recall that
∞
X (−1)k (x/2)2k
J0 (x) =
k !k !
k =0
is a Bessel function of the first kind and order zero.

Theorem
Let
∞ 2
X (−1)k −1 x k
e(x) = − ln x − γ.
2k k!
k =1
Then, for real positive x,

Z ∞
J0 (t)
e(x) = dt.
2x t

Sketch of Proof
Proceed as on pg. 99 of Berndt, and use the fact that
Z ∞ −t
e − J0 (2t)
dt = 0. (4)
0 t
A slightly more general result than (4) is given in equation

6.622.1 of Gradshteyn and Ryzhik, and is attributed to Nielsen.
An independent proof is given in [3].
Corollary
Let e(x) be as in Theorem 1. Then, for large positive x, e(x)
has an asymptotic expansion
π 13 sin 2x + π4

1 1
e(x) = cos 2x + + +O .
1/2
2π x 3/2 4 16x x2

Comparison of n = 1 and n = 2
We see that, for computational purposes, it is much better to

take n = 1 than n = 2 in (2), because the error for n = 1 is
O(e−x /x), but for n = 2 it is Ω± (x −3/2 ).

A Different Generalization
We obtained (2) from (1) by replacing x k /k ! by (x k /k !)n /n.
A similar generalization of (3) is
∞ k n X ∞ k n
X x x
Hk ∼ ln x + γ (5)
k! k!
k =0 k =0
as x → ∞. (3) is just the case n = 1.

It is easy to show that (5) is valid for all positive integer n. An
essential difference between (2) and (5) is that there is a large
amount of cancellation between terms on the left side of (2), but
there is no cancellation in the numerator and denominator on
the left side of (5). The function (x k /k !)n acts as a smoothing
kernel with a peak at k ' x. Since
Hk = ln k + γ + O(1/k ),
the result (5) is not surprising. What may be surprising is the

speed of convergence.
Speed of Convergence
Brent and McMillan [2] show that
∞ ∞
n X n
xk xk
X
Hk = ln x + γ + O(e−cn x ) (6)
k! k!
k =0 k =0
as x → ∞, where
(
1, if n = 1;
cn = 2
2n sin (π/n), if n ≥ 2.
In the case n = 2, (6) has error O(e−4x ). Brent and McMillan

used this case with x ' 17,400 to compute γ to high precision.
They deduced that, if γ = p/q is rational, then q > 1015000 .
From Corollary 1, the same value of x in (2) would give less
than 8-decimal place accuracy.

Another view
If you are looking for a good way to compute Euler’s constant γ,
you might scan Abramowitz and Stegun (or the online Digital
Library of Mathematical Functions) looking for formulas in
which γ occurs.
For example, in the chapter on Bessel functions, we find
(9.6.13):
x2 1 x4
K0 (2x) = −(ln(x) + γ)I0 (2x) + + (1 + ) + ···
(1!)2 2 (2!)2
(where I replaced z by 2x). Here I0 (z) and K0 (z) are modified

Bessel functions (sometimes called Bessel functions of
imaginary argument because we obtain them by z 7→ iz in the
usual Bessel functions J0 (z) etc).

Another view
If you are looking for a good way to compute Euler’s constant γ,
you might scan Abramowitz and Stegun (or the online Digital
Library of Mathematical Functions) looking for formulas in
which γ occurs.
For example, in the chapter on Bessel functions, we find
(9.6.13):
x2 1 x4
K0 (2x) = −(ln(x) + γ)I0 (2x) + + (1 + ) + ···
(1!)2 2 (2!)2
(where I replaced z by 2x). Here I0 (z) and K0 (z) are modified

Bessel functions (sometimes called Bessel functions of
imaginary argument because we obtain them by z 7→ iz in the
usual Bessel functions J0 (z) etc).
Formula (9.6.13) might be useful for evaluating γ if we had an
independent way of evaluating K0 (2x) and I0 (2x).

Differential equation (ODE)
I0 (z) and K0 (z) are independent solutions of the modified

Bessel equation
zw 00 + w 0 − zw = 0.
This is the special case ν = 0 of
z 2 w 00 + zw 0 − (z 2 + ν 2 )w = 0.

Power series for I0
Abramowitz and Stegun (9.6.12) gives the nice series

∞
X x 2k
I0 (2x) = ,
(k !)2
k =0
so there is no difficulty in computing I0 (2x).

(As usual, we replaced z by 2x.)

Asymptotic series for I0 and K0
In the same chapter of Abramowitz and Stegun, we find the

asymptotic expansions:
ez 12 12 · 32 12 · 32 · 52

I0 (z) ∼ √ 1+ + + + ··· ,
2πz 1!(8z) 2!(8z)2 3!(8z)3
12 12 · 32 12 · 32 · 52
r
−z π
K0 (z) ∼ e 1− + − + · · · .
2z 1!(8z) 2!(8z)2 3!(8z)3
These expansions give a way of computing I0 (z) and K0 (z)
accurately if z is sufficiently large (z is always real and positive
in our applications).

Asymptotic series for I0 and K0
In the same chapter of Abramowitz and Stegun, we find the

asymptotic expansions:
ez 12 12 · 32 12 · 32 · 52

I0 (z) ∼ √ 1+ + + + ··· ,
2πz 1!(8z) 2!(8z)2 3!(8z)3
12 12 · 32 12 · 32 · 52
r
−z π
K0 (z) ∼ e 1− + − + · · · .
2z 1!(8z) 2!(8z)2 3!(8z)3
These expansions give a way of computing I0 (z) and K0 (z)
accurately if z is sufficiently large (z is always real and positive
in our applications).
The leading terms show that K0 (z)/I0 (z) = O(e−2z ) is
exponentially small if z is large and positive.

Asymptotic series for I0 K0
While on the subject of asymptotic expansions, note that if we
multiply the asymptotic expansions for I0 (z) and K0 (z), then
half the terms vanish, and we obtain (at least formally)
13 13 · 33 13 · 33 · 53

1
I0 (z)K0 (z) ∼ 1+ + + + ··· .
2z 1!(8z 2 ) 2!(8z 2 )2 3!(8z 2 )3
This turns out to be in Abramowitz and Stegun (9.7.5), though

no error bound or reference is given there.

Asymptotic series for I0 K0
While on the subject of asymptotic expansions, note that if we
multiply the asymptotic expansions for I0 (z) and K0 (z), then
half the terms vanish, and we obtain (at least formally)
13 13 · 33 13 · 33 · 53

1
I0 (z)K0 (z) ∼ 1+ + + + ··· .
2z 1!(8z 2 ) 2!(8z 2 )2 3!(8z 2 )3
This turns out to be in Abramowitz and Stegun (9.7.5), though

no error bound or reference is given there.
To prove the formula for the general term, we could use the
Wilf-Zeilberger (WZ) method. Easier is to use the ODE
z 3 f 000 + z(1 − 4z 2 )f 0 − f = 0
satisfied by f (z) = zI0 (z)K0 (z). It is straightforward to deduce a

recurrence relation for the coefficients in the asymptotic
expansion from this ODE.
Error bounds for the asymptotic expansions
Suppose the asymptotic expansions for I0 , K0 or I0 K0 are
written as
F (z) ∼ a0 (z) + a1 (z) + a2 (z) + · · ·
(where the aj (z) are not identically zero), and the error En (z) is
defined by
F (z) = a0 (z) + a1 (z) + · · · + an−1 (z) + En (z).
Then, provided z is real, z ≥ 1, and n > 0, we can show that

√
|En (z)| = O( n|an (z)|),
and even give an explicit constant in the “O” result (e.g. 4). In
the case of K0 , the errors alternate in sign and there is a
sharper bound
|En (z)| < |an (z)|.

Proofs of Error Bounds
The proofs for K0 and I0 are discussed in Olver’s book

(Chapter 7, especially Ex. 13.2).

Proofs of Error Bounds
The proofs for K0 and I0 are discussed in Olver’s book

(Chapter 7, especially Ex. 13.2).
The proof for I0 K0 does not seem to have been published, and
was stated as a conjecture in Brent and McMillan (1980). It is
possible to deduce it from bounds for the K0 and I0 expansions.

Deducing γ
Rearranging (9.6.13) and using the power series for I0 gives

∞ ∞
2 X 2
xk xk

X K0 (2x)
γ + ln(x) = Hk − ,
k! k! I0 (2x)
k =0 k =0
but the last term is O(e−4x ) so can be neglected if x is large.

This is essentially Algorithm B1 of Brent & McMillan.

Deducing γ
Rearranging (9.6.13) and using the power series for I0 gives

∞ ∞
2 X 2
xk xk

X K0 (2x)
γ + ln(x) = Hk − ,
k! k! I0 (2x)
k =0 k =0
but the last term is O(e−4x ) so can be neglected if x is large.

This is essentially Algorithm B1 of Brent & McMillan.
We have just recovered (6) in the case n = 2, with an explicit
error term K0 (2x)/I0 (2x).

Approximating the error term
To get a more accurate algorithm (with the same x) we can try

to approximate the error term K0 (2x)/I0 (2x). Since I0 (2x) has
already been computed (denominator of the main term), we
only need to approximate K0 (2x). This can be done with
relative error O(e−4x ) by taking z = 2x and d4xe terms in the
asymptotic expansion for K0 (z).


A faster way is to take d2xe terms in the asymptotic expansion
for I0 (2x)K0 (2x), and divide the result by I0 (2x)2 .


A faster way is to take d2xe terms in the asymptotic expansion
for I0 (2x)K0 (2x), and divide the result by I0 (2x)2 .
In this way we get an algorithm for γ with error O(e−8x )
(Algorithm B3 of Brent & McMillan).

Binary Splitting
When summing series of rational numbers, the complexity can
be reduced by the technique of binary splitting.

Binary Splitting
Divide the sum into two, take out common factors, and
recursively sum each half. This way we work with “small”
rational numbers most of the time. Use “fast” integer
multiplication for the rational arithmetic. If the numerators and
denominators of the rationals grow too large, they can be
“pruned”.

Binary Splitting
“pruned”.
It is easy to use this technique to sum the power series for
exp(z) or for I0 (z) when z is rational. The complexity for d
digits is reduced from O(d 2 ) to O(d(log d)c ) for some (small)
constant c.

Binary Splitting
“pruned”.
It is easy to use this technique to sum the power series for
exp(z) or for I0 (z) when z is rational. The complexity for d
digits is reduced from O(d 2 ) to O(d(log d)c ) for some (small)
constant c.
It is trickier to implement binary splitting for the series involving
Harmonic numbers Hk , but it can be done.

Binary Splitting – 1D case
The recursive procedure

0 if ` ≤ 0,

S1 (f , j, `) = fj if ` = 1,

S1 (f , j, b`/2c) + S1 (f , j + b`/2c, d`/2e) otherwise

returns the sum X

fj+k .
0≤k <`
It is easy to modify S1 to compute the polynomial

X
fj+k x j
0≤k <`

The recursive procedure

0 if ` ≤ 0,

S1 (f , j, `) = fj if ` = 1,

S1 (f , j, b`/2c) + S1 (f , j + b`/2c, d`/2e) otherwise

returns the sum X

fj+k .
0≤k <`
It is easy to modify S1 to compute the polynomial

X
fj+k x j
0≤k <`
(details left as an exercise).

Similarly, to compute the sum

X
fj+p gk +q
0≤p+q<`
we can use the recursive procedure




0 if ` ≤ 0,

f g if ` = 1,
j k
S2 (j, k , `) =
S2 (j + b`/2c, k , d`/2e) + S2 (j, k + d`/2e, b`/2c)+



 S (f , j, b`/2c)S (g, k , d`/2e) otherwise.
1 1

Similarly, to compute the sum

X
fj+p gk +q
0≤p+q<`
we can use the recursive procedure




0 if ` ≤ 0,

f g if ` = 1,
j k
S2 (j, k , `) =
S2 (j + b`/2c, k , d`/2e) + S2 (j, k + d`/2e, b`/2c)+



 S (f , j, b`/2c)S (g, k , d`/2e) otherwise.
1 1
This is essentially the “short product” algorithm of

Mulders (2000).

Binary Splitting cont.
We can use the recursive procedure S2 to compute sums such

as
X X bk
Hk bk =
j
0<k <n 0<j≤k <n

Binary Splitting cont.
We can use the recursive procedure S2 to compute sums such

as
X X bk
Hk bk =
j
0<k <n 0<j≤k <n
(details left as an exercise).

Record Computation of γ
Alexander Yee seems to hold the world record for the
computation of γ (though records do not last long, so this may
soon be obsolete). In March 2009 he computed the first
29 844 489 545 ≈ 236 / ln(10) decimal digits.

The computation was performed using Algorithm B3 with binary
splitting and x = 233 (taking 205 hours on a dual-processor
workstation “Nagisa”).

Verification used Algorithm B1 with x = 234 (taking 269 hours).

The time to compute ln(2) (about 40 = 16 + 24 hours) is not
included.

included.
Note that B1 gives 34 ln(2) + γ and B3 gives 33 ln(2) + γ, so it
is possible to deduce both ln(2) and γ, but if we do this then we
don’t get an independent confirmation of the γ value.

included.
Note that B1 gives 34 ln(2) + γ and B3 gives 33 ln(2) + γ, so it
is possible to deduce both ln(2) and γ, but if we do this then we
don’t get an independent confirmation of the γ value.
For more details, history, and other constants, see http://
numberworld.org/nagisa_runs/computations.html.

Who was McMillan?
I met Ed McMillan when I was on sabbatical leave in Berkeley

in 1977/8. At that time he had recently retired from Lawrence
Berkeley Laboratory but still had an office there. He had seen
my (first) paper on Euler’s constant in Math. Comp. 31 (1977)
and wanted to talk to me about possible improvements. Thus, I
walked up the hill to LBL to talk to him, and our collaboration
started.

Who was McMillan?
I met Ed McMillan when I was on sabbatical leave in Berkeley

in 1977/8. At that time he had recently retired from Lawrence
Berkeley Laboratory but still had an office there. He had seen
my (first) paper on Euler’s constant in Math. Comp. 31 (1977)
and wanted to talk to me about possible improvements. Thus, I
walked up the hill to LBL to talk to him, and our collaboration
started.
Luckily I knew nothing about him at the time, or I might have
been intimidated. He seemed to be just a scholarly old
gentleman who was interested in Bessel functions.

McMillan’s notes 17 Nov 1977, page 1




McMillan’s other life
Later I discovered that McMillan had worked at Los Alamos
during the second world war, and was famous for discovering
the first trans-uranium element (neptunium) in 1940, soon
followed by the discovery of plutonium with Seaborg.

He shared the 1951 Nobel prize in Chemistry with Seaborg for
this work.

this work.
In 1945 he discovered (independently of Veksler) the theory of
the synchrotron, making possible the construction of modern
particle accelerators.

this work.
particle accelerators. Modified Bessel functions of fractional
order occur in this theory.

this work.
particle accelerators. Modified Bessel functions of fractional
order occur in this theory.
According to the obituary by Jackson and Panofsky, McMillan’s
last published paper was the one that he wrote with me on the
computation of Euler’s constant. It was published in Math.
Comp. 34 (1980).

Edwin Mattison McMillan 1907–1991 (about 1950)

Another photo (much as I remember him in 1977)

Prehistory - Riemann 1855

Conclusion
We did not succeed in proving that γ is irrational, but the quest

was worthwhile because it provided the motivation:
I to read Ramanujan’s papers and Notebooks;
I to meet and collaborate with Ed McMillan;
I to learn more about Bessel functions (a topic of classical
mathematics that should be better known).

Conclusion

Finally, here is a nice integral for K0 (x), x > 0:
Z ∞
cos(xt)
K0 (x) = √ dt.
0 1 + t2

Conclusion

Finally, here is a nice integral for K0 (x), x > 0:
Z ∞
cos(xt)
K0 (x) = √ dt.
0 1 + t2
We could use this with numerical quadrature to compute γ, but

it would be unlikely to give a fast algorithm.

References
1. Bruce C. Berndt, Ramanujan’s Notebooks, Parts I–III,

Springer-Verlag, New York, 1985–1991.
2. Richard P. Brent and Edwin M. McMillan, Some new
algorithms for high-precision computation of Euler’s constant,
Mathematics of Computation 34 (1980), 305–312.
3. Richard P. Brent, An asymptotic expansion inspired by
Ramanujan, Australian Mathematical Society Gazette 20
(December 1993), 149–155. arXiv:1004.5506v1.
4. Richard P. Brent, Ramanujan and Euler’s constant,
Proceedings of Symposia in Applied Mathematics, Vol. 48
AMS, Providence, Rhode Island, 1994, 541–545.

References cont.
5. J. D. Jackson and W. K. H. Panofsky, Edwin Mattison
McMillan 1907-1991, Biographical Memoirs of the National
Academy of Sciences (USA), 69 (1996), 214–241.
6. Thom Mulders, On short multiplications and divisions,
Applicable Algebra in Engineering, Communication and
Computing, 11, 1 (2000), 69–88.
7. Frank W. J. Olver, Asymptotics and Special Functions,
Academic Press, New York, 1974.
8. Srinivasa Ramanujan, A series for Euler’s constant γ,
Messenger of Mathematics 46 (1917), 73–80.
9. G. F. B. Riemann, Zur Theorie der Nobili’schen Farbenringe,
Poggendorff’s Annalen der Physik und Chemie, 95 (1855),
130–139. Reprinted in Bernhard Riemann’s Gesammelte
Mathematische Werke, Dover, NY, 1953, 55–62.

Brent - 1994 - Ramanujan and Euler's Constant

Uploaded by

Copyright:

Available Formats

Brent - 1994 - Ramanujan and Euler's Constant

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Brent - 1994 - Ramanujan and Euler's Constant

Uploaded by

Copyright:

Available Formats

Ramanujan and Euler’s Constant

Presented at the CARMA Workshop on Exploratory

Ramanujan gave many beautiful formulas for π and 1/π. See,

Richard Brent Ramanujan and Euler’s Constant

Apéry proved ζ(3) irrational using the series

and, in Chapter 9 of his Notebooks, Ramanujan gives several

Richard Brent Ramanujan and Euler’s Constant

is a Harmonic number. Harmonic numbers also occur in

Richard Brent Ramanujan and Euler’s Constant

This family of series all involve the Riemann zeta function or

Scanning Berndt, we find many occurrences of γ. Some involve

which we can interpret as ψ(x + 1) + γ if x is not necessarily a

Richard Brent Ramanujan and Euler’s Constant

as x → ∞. In fact, Euler showed that

and this has been used by Sweeney and others to compute

Richard Brent Ramanujan and Euler’s Constant

Ramanujan’s Corollary 2, Entry 9, Chapter 4 (page 98) is that,

so (1) is just the case n = 1.

Richard Brent Ramanujan and Euler’s Constant

Richard Brent Ramanujan and Euler’s Constant

This is easy to prove (Berndt, page 47). Thus (1) gives

This is more convenient than (1) for computation, because

is a Bessel function of the first kind and order zero.

Then, for real positive x,

Richard Brent Ramanujan and Euler’s Constant

A slightly more general result than (4) is given in equation

Richard Brent Ramanujan and Euler’s Constant

We see that, for computational purposes, it is much better to

Richard Brent Ramanujan and Euler’s Constant

as x → ∞. (3) is just the case n = 1.

the result (5) is not surprising. What may be surprising is the

In the case n = 2, (6) has error O(e−4x ). Brent and McMillan

Richard Brent Ramanujan and Euler’s Constant

(where I replaced z by 2x). Here I0 (z) and K0 (z) are modified

Richard Brent Ramanujan and Euler’s Constant

(where I replaced z by 2x). Here I0 (z) and K0 (z) are modified

Richard Brent Ramanujan and Euler’s Constant

I0 (z) and K0 (z) are independent solutions of the modified

Richard Brent Ramanujan and Euler’s Constant

Abramowitz and Stegun (9.6.12) gives the nice series

so there is no difficulty in computing I0 (2x).

Richard Brent Ramanujan and Euler’s Constant

In the same chapter of Abramowitz and Stegun, we find the

Richard Brent Ramanujan and Euler’s Constant

In the same chapter of Abramowitz and Stegun, we find the

Richard Brent Ramanujan and Euler’s Constant

This turns out to be in Abramowitz and Stegun (9.7.5), though

Richard Brent Ramanujan and Euler’s Constant

This turns out to be in Abramowitz and Stegun (9.7.5), though

satisfied by f (z) = zI0 (z)K0 (z). It is straightforward to deduce a

F (z) = a0 (z) + a1 (z) + · · · + an−1 (z) + En (z).

Then, provided z is real, z ≥ 1, and n > 0, we can show that

Richard Brent Ramanujan and Euler’s Constant

The proofs for K0 and I0 are discussed in Olver’s book

Richard Brent Ramanujan and Euler’s Constant

The proofs for K0 and I0 are discussed in Olver’s book

Richard Brent Ramanujan and Euler’s Constant

Rearranging (9.6.13) and using the power series for I0 gives

but the last term is O(e−4x ) so can be neglected if x is large.