On the accuracy of the Chakrabarti-Hudson

approximation to π
Mark B. Villarino∗
arXiv:1202.2783v2 [math.CA] 1 Dec 2015

Escuela de Matemática, Universidad de Costa Rica,

10101 San José, Costa Rica
December 2, 2015

Abstract
We obtain rigorous upper and lower bounds for the error in the recent approxima-
tion for π proposed by Chakrabarti & Hudson.

AMS Subject Classification: 11A03, 11A04, 01A08

Key Words: Archimedes’ method of approximating π

1 Introduction
In 2003, Chakrabarti and Hudson [1] proposed a new formula in the spirit of Archi-
medes for the computation of π, namely
1
π≈ {32πn + 4Π2n − 6an }, (1.1)
30
where πn is the perimeter of an inscribed regular polygon of n sides in a circle of perimeter
π, Πn is the perimeter of a circumscribed regular polygon of n sides in a circle of perimeter
π, and an is the area of an inscribed regular polygon of n sides in a circle of area π.
Archimedes, in The Measurement of the Circle (MC), see [2], obtains his own approxi-
mation by starting with the (evident) inequality

πn < π < Πn , (1.2)

and takes n = 6, 12, 24, 48, 96, respectively. He performs a brilliant tour de force of ma-
nipulating and rounding the inequalities resulting from continued fraction expansions of the
∗
corresponding author; email: [email protected]

1
square roots which arise in the computations of πn and Πn , whence he obtains his justly
famous bounds
10 1
3+ <π < 3+ . (1.3)
71 7
It is generally recognized, see [6], that the extant MC is a post-Archimedean revision
of Archimedes’ original and far more comprehensive treatment of the circle. For example,
Heron, Metrica I, 32 [3], quotes a theorem of Archimedes from (MC), which is not present
in the extant version. The theorem states that the area of a circular sector exceeds four-thirds
the area of the greatest inscribable triangle. In terms of the notation above, this inequality
affirms that:
π > 13 (4π2n − πn ). (1.4)
The right hand side of (1.4) is a (generalized) convex combination of π2n and πn since
4
3
− 31 = 1. (We used the adjetive “generalized” since usually the coefficients in a convex
combination are taken to be positive. We will supress the word “generalized” from here on,
hoping that the reader will keep this comment in mind.) Moreover, its form argues for the
fact that the original treatise of Archimedes studied convergence-improvement inequalities
derived by means of geometric theory. Unfortunately, such theoretical aspects apparently
proved too subtle for the uses of later commentators such as Heron.
Observe that the formula of Chakrabarti & Hudson (1.1), too, is a convex combina-
tion of its terms.
Such convex combinations did not reappear until the sixteenth and seventeenth centuries
in the work of Snell [7] and Huygens [5]. The former stated and the latter proved the
following upper bound inequality:

π < 23 πn + 31 Πn , (1.5)

which is a beautiful generalization of Archimedes’ original inequality. Both Snell and Huy-
gens use it in the form
π ≈ 23 πn + 13 Πn (1.6)
and they justifiably extol the convergence-rate improvement produced by this convex com-
binations of quantities already computed.
Now, Chakrabarti and Hudson ask what happens if we use the corresponding areal convex
combination
π ≈ 32 an + 31 An (1.7)
to approximate π, where An is the area of a circumscribed regular polygon of n sides in a
circle of area π. They prove the very interesting result that
2
π
3 n
+ 31 Πn − π 3
lim 2 1 = , (1.8)
n→∞ an
3
+ 3 An − π 8

which shows that the areas converge much more slowly than the perimeters to π.

2
 Then they suppress the limit notation, treat (1.8) as an equation for π and when they
reduce it algebraically they obtain the approximative formula (1.1). Passing to trigonometric
functions, (1.1) becomes:
  2π 
n π  π 
π≈ 32 sin + 4 tan − 3 sin ≡ Π(n). (1.9)
30 n n n
It is worthwhile to point out that the procedures producing the improved convex com-
binations cited above are today instances of “Richardson extrapolation” (see [4]). In the
case of the Chakrabarti-Hudson formula, we can obtain it in the following way. Put
x := a sin x + b tan x + c sin(2x),
with unknown coeficients and match the first three terms of the Taylor series, which leads
to the linear system:
a b 8c a 2b 32c
a + 2b + c = 1 + − =0 + + = 0.
6 3 6 120 15 120
1
The error term then has initial term 105 x6 , which, indeed was obtained by Chakrabarti-
Hudson.
Nevertheless, it is of interest to see how the seemingly ad hoc method of the authors
produced the same result as the modern method.
Finally, the authors offer some numerical studies of the accuracy of (1.9).
The investigation of the authors needs to be completed in several areas.
• They do not state whether the approximation (1.9) is in excess or in defect. . . in fact,
we will prove that it is in excess.
• They do not develop any rigorous error analysis. . . no upper bounds for the error nor
lower bounds; we will present such bounds.
• They offer a non-standard definition of accuracy, which they call “precision”. This
makes it difficult to compare their numerical results with standard error studies. We
will present the standard definition of “correct significant digits” and reformulate the
discussion of the accuracy of (1.9) in light of our error bounds.
Recently, M. Szyszkowicz (see [8]) exploited the Richardson method to present eighteen
different approximations (!) for π of which the most accurate is (in his notation)
sin x · (187 + 24 cos x − cos 2x)) x9 x11
M12 := = x− − −··· . (1.10)
10 + 90 cos x 17640 226385
which in in excess. (The formula given in his paper has a misprint...an extra factor “x” in
the numerator which should not be there.) We point out that the third convergent of the
continued fraction expansion of sinx x given later on (see (4.2)) gives the approximation
sin x · (51 + 48 cos x + 6 cos2 x) x9 x11
= x− − −··· (1.11)
80 + 25 cos x 44100 226380

3
also in excess, and which is of comparable complexity, but is 2 12 times more accurate asymp-
totically. It would be interesting to develop rigorous error bounds for his eighteen approxi-
mation formulas.

2 Error Analysis
We will prove the following theorem:

Theorem 1. If n > 32 then the following inequality is valid:

1  π 6 Π(n) − π 1  π 6
< < 7 (2.1)
105 n π 104 10 n

1
where the lower-bound constant is the best possible.
105
An immediate consequence is:

Corollary 1. The approximation π ≈ Π(n) is in excess.

The inequality (2.1) bounds the relative error in the approximation Π(n) ≈ π. It is well
known that an approximation has n correct significant digits iff the relative error does not
( 21 )
exceed n , see [4]. Therefore, we can say:
10
Corollary 2. The approximation π ≈ Π(n) has about (6 log10 n − 1.27) correct significant
digits.

Now we turn to the proof of (2.1). We will use the MacLaurin expansions of the
functions involved. Define:
π
f (x) := (32 sin x + 4 tan x − 3 sin 2x). (2.2)
30x
Then we see that π
Π(n) = f . (2.3)
n
Since the sum of a convergent alternating series is bracketed by two consecutive partial
sums if the absolute values of the terms decrease monotonically (which here occurs), we
obtain:

Lemma 1. The following inequality is valid for all 32

π
> x > 0:

x3 x5 x7 x3 x5 x7 x9
   
32 x − + − < 32 sin x < 32 x − + − + . (2.4)
3! 5! 7! 3! 5! 7! 9!

4
 The MacLaurin expansion of the tangent function is not an alternating series. But we
can still use the Lagrange form of the remainder to obtain:

Lemma 2. The following inequality is valid for 0 < x 6 π

32
:
 
1 3 2 5 17 7 62 9
4 x+ x + x + x + x < 4 tan x
3 15 315 2835
 
1 3 2 5 17 7 62 9 1 11
<4 x+ x + x + x + x + x . (2.5)
3 15 315 2835 85

Proof. The MacLaurin expansion of order 11 of the tangent function is

1 2 17 7 62 9 1
tan x = x + x3 + x5 + x + x + R(θ11 )x11 , (2.6)
3 15 315 2835 11!
where
d11
R(x) := 11
(tan x) = 256(tan2 x + 1)(155925 tan10 x + 467775 tan8 x + 509355 tan6 x
dx
+ 238425 tan4 x + 42306 tan2 x + 1382)

and 0 6 θ11 6 π
32
. The function R(x) monotonically increases in the interval, whence
π
353792 = R(0) < R(θ11 ) 6 R = 469223.9941 . . . (2.7)
32
and we conclude (since x > 0) that
1 1 1 1
06 R(θ11 )x11 6 · 469223.9941 . . . x11 = x11 < x11 . (2.8)
11! 11! 85.06 . . . 85
This completes the proof.
Finally, as in the inequality (2.4) we obtain:

Lemma 3. The following inequality is valid for all 0 < x 6 π

32
:

(2x)3 (2x)5 (2x)7 (2x)9

 
− 3 2x − + − + < −3 sin 2x
3! 5! 7! 9!
(2x)3 (2x)5 (2x)7 (2x)9 (2x)11
 
< −3 2x − + − + − . (2.9)
3! 5! 7! 9! 11!

Now that we have set up the technical inequalities necessary in our main proof, we enter
into its details.

5
Proof of the main theorem. If we substitute the inequalities (2.4), (2.5) and (2.9) into the
formula for f (x), (2.2), we obtain
1 6 1 8 2776 10 f (x) 1 6 1 8 10531 10
x + x + x < −1< x + x + x . (2.10)
105 360 2338875 π 105 360 26507250
Now, the left-hand side
 
1 6 1 8 2776 10 1 6 7 2 2776 4 1 6
x + x + x = x 1+ x + x > x (2.11)
105 360 2338875 105 24 22275 105
1
for all positive x. So, we have proven the lower bound (2.1) and that the constant
105
cannot be replaced by a bigger one, i.e., it is the best possible.
For the upper bound, we observe that for 0 < x 6 32 π
,
 
1 6 1 8 10531 10 1 6 7 2 10531 4
x + x + x = x 1+ x + x
105 360 26507250 105 24 252450
 
1 6 7  π 2 10531  π 4
< x 1+ +
105 24 32 252450 32
1 1
= x6 < 6
7 x ,
104.705 . . . 104 10
which is the upper bound presented in (2.1).
This, together with (2.3) complete the proof of the theorem.

3 Numerical Error Studies

Chakrabarti and Hudson present a table of numerical studies of the error in their approxi-
mative formula (1.8). They define: a number α has precision n if
1
|α − π| < n . (3.1)
10
We already pointed out the standard definition (see [4]): an approximation N approximates
the true value N with n correct significant digits if the positive relative error
|N − N | (1)
< 2n . (3.2)
N 10
Of course the two definitions will coincide sometimes, and sometimes not. However, we note
that
1 |α − π| 1 ( 21 )
|α − π| < n ⇐⇒ < <
10 π π10n 10n
so that an approximation with precision n is always correct to n significant digits. It is
sufficient, but not necessary, since if
1 |α − π| ( 12 )
< < ,
π10n π 10n
then α will have precision n − 1 but still have n correct significant digits.

6
4 Earlier approximations
f (x)
We can rewrite the approximation 1 ≈ as follows:
π
sin x 15 cos x
≈ , (4.1)
x 2 + 16 cos x − 3 cos2 x
x6
where we know that the approximate value is smaller than the true value by about .
105
sin x
Now, has the following continued fraction expansion:
x
sin x 1·2
1·3
sin2 x2
=1− (4.2)
x 1·2
3·5
sin2 x2
1− 3·4
5·7
sin2 x2
1 − 3·4
7·9
sin2 x2
1−
1−···
The first few convergents are:
p1 2 + cos x 1 4
= ; error = − x + · · · (Snell)
q1 3 180
p2 9 + 6 cos x 1 6
= ; error = − x + · · · (Newton)
q2 14 + cos x 2100
p3 51 + 48 cos x + 6 cos2 x 1
= ; error = − x8 + · · · (4.3)
q3 80 + 25 cos x 44100
where we have used sin2 x2 = 21 (1 − cos x). (For all these results, see [9].) Every one of these
approximations is larger than the true value and is the best possible.
The Chakrabarti–Hudson approximation (4.1) has an error about twenty times greater
than that of Newton’s formula, although both are of order o(x6 ), nor is it as simple. Indeed, it
has the formal complexity of the third convergent without the latter’s extraordinary accuracy.
Nevertheless, it is the only approximation in defect, its accuracy is still quite good, and it is
interesting that such an ad hoc derivation produced such an intriguing approximation.

Acknowledgment
Support from the Vicerrectorı́a de Investigación of the University of Costa Rica is acknowl-
edged.

References
[1] G. Chakrabarti and R. Hudson, “An improvement of Archimedes’ method of approxi-
mating π”, Int. J. Pure Appl. Math. 7 (2003), 207–212.

7
[2] T. L. Heath. The Works of Archimedes, Cambridge University Press, Cambridge, 1897.

[3] Heron, Opera, 5 volumes (III: Metrica, ed. H. Schöne), Teubner, Leipzig, 1903.

[4] F. B. Hildebrandt. Introduction to Numerical Analysis, Dover, New York, 1987.

[5] Ch. Huygens, De circuli magnitudine inventa, Elzevier, Leiden, 1654.

[6] W. R. Knorr, “Archimedes and the Measurement of the Circle: A New Interpretation”,
Arch. Hist. Exact Sci. 15 (1976), 115–140.

[7] W. Snell, Cyclometricus, Leiden, 1621.

[8] M. Szyszkowicz, “Approximations of pi and squaring the circle”, JMEST 2 (2015), 330–
332.

[9] K. T. Vahlen, Konstruktionen und Approximationen, Teubner, Leipzig, 1911.