The logarithm constant: log 2

(Click here for a Postscript version of this page).

1  Introduction

The story of logarithms really began with the Scot John Napier (1550-1617) in a work published in 1614. This treatise was in Latin and entitled Mirifici logarithmorum canonis descriptio [7]. However it should be pointed out that the Swiss clockmaker Jost B�rgi (1552-1632) independently invented logarithms but his work remained unpublished until 1620.

Thanks to the possibility to replace painful multiplications and divisions by additions and subtractions respectively, this invention received an extraordinary welcome and spread rapidly on the Continent (in particular thanks to the enthusiasm of astronomers like Kepler). George Gibson wrote during Napier Tercentenary Exhibition: ''the invention of logarithms marks an epoch in the history of science''.

Very soon tables of logarithms were published. One of the first was due to the English mathematician Henry Briggs (1561-1631) who, in 1624, published his work in Arithmetica logarithmica [4]; the tables were to an accuracy of fourteen digits and containing the common logarithm of integers up to 20,000 and from 90,000 up to 101,000. The remaining gap was completed [12] by the Dutchman Adrian Vlacq (1600-1667) in 1628.

Mathematicians prefer to use the so-called natural or hyperbolic logarithm of a number (denoted log or ln, that is logarithms having base e=2.7182818284...) and the following definition allows to derive easily the main properties of logarithms.

Definition 1 Let x > 0

log x =   � � x 1  dt t  =   � � x-1 0  dt 1 + t .

It may be interpreted as the area under the hyperbola y=1/t with t going from 1 to x. (this geometric interpretation was showed by the Jesuit Gr�goire de Saint-Vincent (1584-1667) in 1647).

Therefore, log 2 will be defined by

log 2 = � � 2 1  dt t = � � 1 0  dt 1 + t  =  � � 1/2 0  dt 1 - t .

Theorem 2 log 2 is an irrational and transcendental number.

2  Algorithms based on integral representation

From the integral representation of log 2 we may deduce some formulae. The first one is deduced form the rectangular approximation of the integral (here h=(b-a)/n)

� � b a  f(t) dt = h n � k = 1  f(a + kh) + O(h),

which, applied to f(t)=1/(1+t) on [0,1] gives

log 2 = lim n� �  � � 1 n + 1  +   1 n + 2  + ... +   1 2n � � .

This sequence is increasing. Here are the numerical values of some partial sum for different value of n.

x10 = 0.6(687...), x100 = 0.69(065...), x1,000 = 0.69(289...).

It is a very slow convergence, but it can be improved if we use the trapezoidal rule

� � b a  f(t)dt =   h 2 � � f(a) + f(b) + 2 n-1 � k = 1  f(a + kh) � �  + O(h2),

yielding

log 2 = lim n� �  � � 3 4n  +   � � 1 n + 1  +   1 n + 2  + � +   1 2n - 1 � � � �

and some iterates are

x10 = 0.693(771...), x100 = 0.6931(534...), x1,000 = 0.693147(243...).

The rate of convergence has been improved but it remains logarithmic. We now give a last improvement on the computation of the integral, known as the Simpson's rule

� � b a  f(t)dt =   h 6 n � k = 0  � � f(a + kh) + f(a + (k-1)h) + 4f � � a + (k- 1 2 )h � � � �  + O(h4),

giving

log 2 =   lim n� �  n � k = 0  � � 1 n + k  +   1 n +k - 1  +   4 n + k - 1/2 � �

and the new iterates

x10 = 0.693147(374...), x100 = 0.6931471805(794...), x1,000 = 0.69314718055994(726...).

From those examples we can notice that such formulae may only be used to compute a few digits of log 2, the rate of convergence is too slow to find constants with high precision.

3  A simple limit

It is of interest to observe that logarithms may be computed by some very simple formulas that are not directly related to geometrical area and involving only square roots extractions.

We know that the constant e may be defined by the famous limit

e =   lim n� �  � � 1 +   1 n � � n   ,

and it is interesting to note that a similar formula exists for the logarithm constant:

log 2 = lim n� �  n (21/n - 1),

it is obviously deduced from the expansion

21/n = elog 2/n = 1 +   log 2 n  + O � � 1 n2 � � .

With n=1,000, we find

x1,000 = 0.693(387...).

A better formula is given by

log 2 = lim n� �  n 2 (21/n - 2-1/n),

again, with n=1,000

x1,000 = 0.693147(236...),

the error is O(1/n²). To conclude this section, we give the unusual infinite product published by Seidel in 1871

log x x - 1  =   k = � �  k = 1  2 x1/2k + 1 ,

we apply it for x=2

log 2 =   2 21/2 + 1 2 21/4 + 1 2 21/8 + 1 ...

4  Taylor's expansion

In 1668, Nicolas Mercator (1620-1687) published in his Logarithmotechnia [6] the well known series expansion for the logarithmic function:

Theorem 3 (Mercator)

log(1 + x) = x -    x2 2  +    x3 3 - ... =   � k � 1   (-1)k-1xk k .       -1 < x � 1

The Mercator series is very similar to another series studied by James Gregory (1638-1675) at the same time:

arctan x = x -    x3  3  +    x5 5 -�

One of the first mathematician to use Mercator's series is Isaac Newton (1643-1727) who made various computation of logarithms in the celebrated ''De Methodis Serierum et Fluxionum'' written in 1671 but only published and translated from Latin to English in 1736, by John Colson. He made several accurate computations for small values of x, like �0.1,�0.2 and was then able to compute logarithms of larger numbers using relations like log 2=2log(1.2)-log(0.9)-log(0.8).

Applying the Mercator series for x=1 leads to the famous and very slowly convergent sequence

log 2 = 1 -   1 2  +   1 3  -   1 4  +   1 5  - ...

which first values, taking respectively 10, 100, 1000 and 10000 terms are:

x10 = 0.6(456...) x100 = 0.6(881...) x1,000 = 0.69(264...) x10,000 = 0.693(097...).

A much better idea is to apply the series with x=-1/2, this produces a formula that was often used for numerical computation:

log 2 =   � k � 1  1 k2k

(1)

The rate of convergence is now geometric (p iterations are needed to obtain one more digits).

x1 = 0.(5) x10 = 0.693(064...) x100 = 0.6931471805599453094172321214581(688...)

Applying the formula up to k=1000 gives more than 300 digits for log 2. The number of iterations k needed to obtain d decimal digits is given by (with N=2 in the previous formula)

1 10d = 1 kNk

which gives for large values of k

k � d log10(N)

If we use the trivial decomposition 2=(4/3)(3/2) and take the logarithm of both side of the equality we find a new formula

log 2 =   � k � 1  1 k � � 1 3k  +   1 4k � �

The partial sum of this series up to k=1000 will give more than 470 digits.

From the relation log 2 = -1/2log(1-1/4)+arctanh(1/2), (see next paragraph for the definition of  arctanh) we deduce the relations

log 2 = � k � 0  � � 1 8k + 8  +   1 4k + 2 � � 1 4k , log 2 = 2 3  +   � k � 1  � � 1 2k  +   1 4k + 1  +   1 8k + 4  +   1 16k + 12 � � 1 16k .

Those formulae can also be used to compute directly the n-th binary digit of log 2 without computing the previous ones.

5  Machin like formulae

Using Mercator series has improved a lot our ability to compute many digits for log 2, but like for the number p it is possible to go further. We recall the definition of the inverse hyperbolic tangent.

Definition 4 Let |x| < 1:

arctanh x =   1 2 log � � 1 + x 1 - x � � = � k � 0  x2k + 1 2k + 1 .

For x=1/3, this leads easily to the basic formula log 2=2 arctanh(1/3) which is similar to the formula p = arctan 1. We rewrite it as:

log 2 =   2 3 � k � 0   1 (2k + 1)9k .

Two iterates are then:

x1 = 0.6(666...), x10 = 0.6931471805(498...).

Applying the formula up to k=1000 gives more than 950 digits for log 2. Is it possible to do better? There is a famous formula for p, this formula is known as Machin's formula and is given by the celebrated relation:

p 4  = 4arctan � � 1 5 � �  - arctan � � 1 239 � � .

The idea is to look for relations like this one for log 2. If we note (a₁, a₂, ..., a_n) a set of rational numbers and (p₁, p₂, ..., p_n) a set of increasing integers we are looking for relations like

log 2 =   n � i = 1  ai arctanh � � 1 pi � � .

We also want the formula to be efficient. A good relation is a compromise between a small number of terms (n must be small) and large values for the p_i. According to Lehmer, we define the efficiency E by

Definition 5 (Lehmer's measure). Let p_i > 0 a set of n numbers, then:

E =   n � i=1  1 log10(pi2) .

For example the efficiency of log 2=2 arctanh(1/3) (n=1, p₁=3) is 1.05. With the same definition the efficiency of Machin's formula is 0.926 (n=2, p₁=5, p₂=239), so it is a little bit more efficient than the formula for log 2.

It is possible to show the following decomposition theorem:

Theorem 6 Let p > 1:

arctanh � � 1 p � � = arctanh � � 1 2p - 1 � � + arctanh � � 1 2p + 1 � � arctanh � � 1 p � � = 2 arctanh � � 1 2p - 1 � � - arctanh � � 1 2p2 - 1 � � arctanh � � 1 p � � = 2 arctanh � � 1 2p + 1 � � + arctanh � � 1 2p2 - 1 � � arctanh � � 1 p � � = 2 arctanh � � 1 2p � � + arctanh � � 1 4p3 - 3p � �

Applying this theorem for p=3 gives relations with n=2:

Corollary 7

log 2 = 2 arctanh � � 1 5 � � +2 arctanh � � 1 7 � �        Euler 1748 (2) log 2 = 4 arctanh � � 1 5 � � -2 arctanh � � 1 17 � � (3) log 2 = 4 arctanh � � 1 7 � � +2 arctanh � � 1 17 � � (4) log 2 = 4 arctanh � � 1 6 � � +2 arctanh � � 1 99 � � (5)

The efficiencies are respectively E=1.307, E=1.121, E=0.998 and E=0.893. Using other decomposition formulas and computer calculation it is possible to find many other relations of this nature (given in [8]). To illustrate this, we select another formula with 3 terms and two others with 4 terms.

Theorem 8 (1997). The constant log 2 may be obtained by one of the following fast converging series:

18 arctanh � � 1 26 � � -2 arctanh � � 1 4801 � � +8 arctanh � � 1 8749 � � , (6) 144 arctanh � � 1 251 � � +54 arctanh � � 1 449 � � -38 arctanh � � 1 4801 � � +62 arctanh � � 1 8749 � � , (7) 72 arctanh � � 1 127 � � +54 arctanh � � 1 449 � � +34 arctanh � � 1 4801 � � -10 arctanh � � 1 8749 � � . (8)

The efficiency are respectively (E=0.616, E=0.659, E=0.689). The two last formulae were used to compute more than 10⁸ digits for log 2. One was used for the computation and the other for the verification and, as you can see, only the computation of 5 arctanh were necessary because of the similitude of the 2 formulae. Thanks to relation (6), the record was increased up to more than 5.10⁸ a few years later.

Note that it is also possible to perform separately the computation of each term, making those relations parallelisable.

We give, for example, the first iterates of the second formula (the one with 251):

x1 = 0.69314(394...) x2 = 0.6931471805(304...) x3 = 0.69314718055994(497...) x4 = 0.69314718055994530941(317...)

and each iteration will add about 4.8 digits.

5.1  Formula with rational numbers

It may be convenient to look for relations with rational numbers as arguments of the arctanh function, that is identities of the form:

log 2 =   n � i = 1  ai arctanh � � mi pi � �

with m_i being integers eventually different of 1. Lehmer's measure must be adapted to take this in account

E =   n � i = 1  1 2log10(pi/mi) ,

but it only gives a rough estimation.

Theorem 9 (1997).

log 2 = 6 arctanh � � 1 9 � � +2 arctanh � � 3 253 � � , (9) log 2 = 10 arctanh � � 1 17 � � +4 arctanh � � 13 499 � � , (10)

with respective efficiency E=0.784 and E=0.722.

Note that relation (10) was used for several high precision computation of the constant.

6  Variation with hypergeometric sequences

Gauss studied the following family of series

F(a, b; c; z) = 1 +    a.b c z 1!  +    a(a+1).b(b+1) c(c+1)  z2 2!  + ...

where a, b, c are real numbers. The series converges in |z| < 1. On the circle |z|=1, the series converges when c > a+b. Those functions are called hypergeometric functions [1]. Many usual functions can be represented as hypergeometric functions with suitable values for a, b, c.

Example 10

ln(1+x) = xF(1, 1; 2; -x),  arctanh x = xF( 1 2 , 1;   3 2 ; x2), arctan x = xF( 1 2 , 1;   3 2 ; -x2),  1 1-x = F(1, 1; 1; x).

In 1797, Johann Friederich Pfaff (1765-1825), a friend of Gauss, gave the interesting relation

F(a, b; c; z) = (1 - z)-bF(c-a, b; c;   z z - 1 ).

If we apply this identity to the arctanh function

arctanh x = xF( 1 2 , 1;   3 2 ; x2),

we find

arctanh x =   x (1 - x2) F(1, 1;   3 2 ;   x2 x2 - 1 ),

giving a new sequence for this function

arctanh x =   x (1 - x2) � � 1 -   2 3 � � x2 1 - x2 � �  +   2.4 3.5 � � x2 1 - x2 � � 2    -   2.4.6 3.5.7 � � x2 1 - x2 � � 3   +... � �

and with log 2=2 arctanh(1/3); after some easy manipulations comes the series:

Corollary 11

log 2 =   3 4 � � 1 -   1 12  +   1.2 12.20  -   1.2.3 12.20.28  + ... � � .

Each new term of this series is multiplied by -k/(8k+4) where k=1,2,... The same kind of formula was given by Euler to compute p. One nice application of this corollary is to make it feasible to write a tiny code to compute log 2 just like it was done for p.

7  log 2 and the AGM

The rate of convergence of all the previous series were logarithmic or geometric. It is possible to find for log 2 just as for p, a sequence that will have quadratic convergence, based on the AGM (Arithmetic-Geometric Mean) (see [2]).

The most classical among the AGM based quadratic convergence algorithms for log 2 is the following. Starting from a₀ and b₀ > 0, we consider the AGM iteration

an+1= an + bn 2 ,       bn+1= �   anbn   ,

and we use the notation

R(a0, b0)= � � � 1 1- � n � 0  2n-1(an2 - bn2) � � � .

Theorem 12 Let N � 3 and 1/2 � x � 1,

| log x - R(1, 10-N) + R(1, 10-Nx)| � N 102(N-2) .

(11)

By choosing x=1/2, this algorithm gives about 2N decimal digits of log 2. The convergence is quadratic and should be stopped when 2^n-1(a_n²-b_n²) is less than 10^-2N, which occurs for n � log₂(N).

Notice that this algorithm is not specific to the computation of log 2 and can be used to evaluate log x for any real number x. Using FFT multiplication, its complexity is O(nlog²n) to obtain n digits of x.

8  Computation of log p

Starting with a good approximation of log 2, it is easy to find the logarithm of all the integers thanks to the relation

2 arctanh � � 1 2p + 1 � �  = log � � � � � � 1+ 1 2p + 1 1- 1 2p + 1 � � � � � �  = log � � p + 1 p � � ,

thus we deduce that:

Theorem 13 Let p > 0

log(p + 1) = log p + 2 � � 1 2p + 1  +   1 3 1 (2p + 1)3  +   1 5 1 (2p + 1)5  +... � � .

This may be useful when you need to compute the logarithms of many successive integers.

Example 14 With p=2, p=4 and p=7

log 3 = log 2 +    2 5 � � 1 +   1 3 1 25  +   1 5 1 252  + ... � � , log 5 = 2log 2 +    2 9 � � 1 +   1 3 1 81  +   1 5 1 812  + ... � � , log 7 = 3log 2 -   2 15 � � 1 +   1 3 1 225  +   1 5 1 2252  + ... � � .

9 Collection of formulae for log 2

Integral and series formulae for log 2 can be found in Collection of formulae for log 2.

10  Records of computation for log 2

References

[1]

M. Abramowitz and I. Stegun, Handbook of Mathematical Functions, Dover, New York, (1964)

[2]

J.M. Borwein and P.B. Borwein, ''Pi and the AGM - A study in Analytic Number Theory and Computational Complexity'', A Wiley-Interscience Publication, New York, (1987)

[3]

R.P. Brent, Fast multiple-Precision evaluation of elementary functions, J. Assoc. Comput. Mach., (1976), vol. 23, p.242-251

[4]

H.Briggs, Arithmetica logarithmica sive logarithmorum Chiliades Triginta, London, (1624)

[5]

L. Euler, Introduction � l'analyse infinit�simale (french traduction by Labey), Barrois, a�n�, Librairie, (original 1748, traduction 1796), vol. 1, p. 89-90

[6]

N. Mercator, Logarithmotechnia, London, (1668)

[7]

J. Napier, Mirifici logarithmorum canonis descriptio, Edimboug, (1614)

[8]

P. Sebah, Machin like formulae for logarithm, Unpublished, (1997).

[9]

W. Shanks, Contributions to Mathematics Comprising Chiefly the Rectification of the Circle to 607 Places of Decimals, G. Bell, London, (1853)

[10]

D.W. Sweeney, On the Computation of Euler's Constant, Mathematics of Computation, (1963), p. 170-178

[11]

H.S. Uhler, Recalculation and extension of the modulus and of the logarithms of 2, 3, 5, 7 and 17, Proc. Nat. Acad. Sci., (1940), vol. 26, p. 205-212

[12]

A. Vlacq, Arithmetica logarithmica, Gouda, (1628)

Back to Numbers, Constants and Computation