MATHEMATICAL COMMUNICATIONS Math. Commun. 16(2011), 105114.

105

Improving Stirlings formula

Department of Mathematics, Faculty of Arts and Sciences, Nev sehir University, Nev sehir, Turkey

Received December 19, 2009; accepted June 9, 2010

Abstract. We calculated the optimal values of the real parameters a and b in such a way that the asymptotic formula n! ea n+a e
n

2 (n + b) (as n )

gives the best accurate values for n!. Our estimations improve the classical Stirling and Burnsides formulas and their several recent improvements due to the author and C. Mortici. Apart from their simplicities and beauties our formulas give very accurate values for factorial n. Also, our results lead to new upper and lower bounds for the gamma function and recover some published inequalities for the gamma function. AMS subject classications: 33B15, 40A25 Key words: Stirling formula, Burnsides formula, gamma function, digamma function, inequalities

1. Introduction
The gamma function dened by the improper integral

(z ) =
0

uz1 eu du (z > 0)

has numerous applications in mathematics and sciences and it occurs e.g. in the expression of various mathematical constants. For instance, the formula for the volume of the unit ball in Rn involves the gamma function, see the recent papers [2, 18, 7] and its monotonicity properties have been studies in [7, 9]. It is also notable that the author used the gamma function to evaluate the sum of some series involving reciprocals of binomial coecients, see [5, 6]. The gamma function and the factorials are related with (n + 1) = n!, n N. The logarithmic derivative of the gamma function is called the digamma (or psi) function and denoted by . The derivatives , , ... are called polygamma functions in the literature. The Stirlings formula (1) n! nn en 2n = n
Corresponding

author. Email address: nbatir@hotmail.com (N. Batir) c 2011 Department of Mathematics, University of Osijek

http://www.mathos.hr/mc

106

N. Batir

is used to estimate factorial n and has many applications in statistical physics, probability theory and number theory. Actually it was discovered by A. De Moivre (1667-1754) in the form n! C nn en n, and Stirling (1692-1770) identied the constant C precisely 2. After the Stirlings formula has been published, many formulas to approximate n! have appeared in the literature but most of them are complicated and do not have a simple form. E. A. Karatsuba [11] answered a question posed by S. Ponnusamy and M. Vuorinen [15, Conjecture 2.35, p.293] motivated by a notebook of S. Ramanujan (1887-1920). She proved in [11] the following asymptotic expansion (for additional terms see [11, Eq. (5.5)] 11 1 6 (x + 1) xx ex 8x3 + 4x2 + x + + = x . 30 240x The accuracy of this and some other asymptotic expansions for the gamma function have been recently investigated by G. Nemes [14]. Undoubtedly in the literature there exist much better approximation formulas to approximate the gamma function or n! like this than (1), but in this work we are interested only in formulas having a simple form. The most well known simple estimation for n! after the Stirlings formula is 1 n+ 2 n + 1/2 n! 2 = n , (2) e which was published by Burnside [8] in 1917. It is known from [8] that this formula is better than (1). Very recently the author [3] and C. Mortici [12] published the following simple estimations for n!: n! n! where = (3 2e e 2 e

n+1 e n+ e

1 n+ 2

= n ,
1 n+ 2

(C. Mortici [12]), (C. Mortici [12]),

(3) (4)

= n ,

3)/6, 2 nn en n+ 1 = n , 6 (N. Batir [3]). (5)

n!

The aim of this work is to improve all these known estimations keeping their simplicities. If we carefully look at these formulas, we see that all are special cases of the following asymptotic formula: For real numbers a, b 0 n! ea n+a e
n

2 (n + b), (as n ).

(6)

Indeed, (1) is obtained from (6) for (a, b) = (0, 0); (2) is obtained for (a, b) = (1/2, 1/2); (3) is obtained for (a, b) = (1, 1), (4) is obtained for (a, b) = (, ) ( = (3 3)/6), and nally (5) is obtained from (6) for (a, b) = (0, 1/6). Hence

Improving Stirlings formula

107

it is natural to look for the best possible constant pairs (a, b) such that formula (6) gives the best accurate values for factorial n. We prove that the best approximations are obtained for the constant pairs (a, b) = (a1 , b1 ) and (a, b) = (a2 , b2 ), where a1 = and a2 = 1 1 + + 3 6 6 6 2 + 4/ = 0.950108831378...,
1 12

1 1 + 3 6 6

6 2 + 4/ = 0.540319070367...,

(7)

(8) = 0, and (9)

which are the real roots of the quartic equation 3a4 4a3 + a2 + b1 = a2 1+

1 1 = 0.458611364468... and b2 = a2 = 1.069373458129..., 2+ 6 6

where = 2 + 22/3 + 24/3 . The following lemmas are key in our proofs. The rst lemma was proved in [13]. Lemma 1. If (n )n1 is convergent to zero and there exists the limit
n

lim nk (n n+1 ) = c R,

with k > 1, then there exists the limit

lim nk1 n =

c . k1

This lemma, despite its simple appearance, is a strong tool to accelerate and measure the speed of convergence of some sequences having limit equal to zero. It is evident from this lemma that the speed of convergence of the sequence (n ) is as higher as the value of k is greater. The next lemma, as far as we know, was rst used in [10] (without proof) to establish some monotonicity results for the gamma function and later used by the author [4] and F. Qi and B-N Guo [16, 17] to prove some monotonicity and complete monotonicity properties of the polygamma functions. Lemma 2. Let f be a function dened on an interval I and lim f (x) = 0. If f (x + 1) f (x) > 0 for all x I , then f (x) < 0 . If f (x + 1) f (x) < 0, then f (x) > 0. Proof. Let f (x + 1) f (x) > 0 for all x I . By mathematical induction we have f (x) < f (x + n) for all n N. Letting n , we get f (x) < 0. The proof of the second part of the lemma follows from the same argument. The numerical and algebraic computations have been carried out with the computer program MATHEMATICA 5.
x

108

N. Batir

2. Main results
For real numbers a, b 0 and n N, we dene n (a, b) = log(n!) 1 1 log(2 ) + n + a n log(n + a) log(n + b). 2 2 (10)

Using the Stirlings formula (1) we see that lim n = 0. (11)

Now we are ready to state and prove our main results. Theorem 1.
1 (i) If b = a2 + 6 , then the speed of convergence of (n ) is n1 . 2 (ii) If a = a1 , a2 and b = a2 + 1 . 6 , then the speed of convergence of (n ) is n

(iii) If (a, b) = (a1 , b1 ) or (a, b) = (a2 , b2 ), then the speed of convergence of (n ) is n3 . Here ai , bi (i = 1, 2) are as given in (7), (8) and (9).

Proof. For real numbers x 0, we dene Ha,b (x) = log (x + 1) Dierentiation gives a 1 , x + a 2(x + b) a 1 1 + , Ha,b (x) = (x + 1) x + a (x + a)2 2(x + b)2 Ha,b (x) = (x + 1) log(x + a) + and Ha,b (x + 1) Ha,b (x) = Pa,b (x) , Qa,b (x) (12) 1 1 log(2 ) + x + a x log(x + a) log(x + b). 2 2

Improving Stirlings formula

109

where we have used the functional relation (x + 1) (x) = 1/x2 , and Pa,b (x) = ( 2 12a2 + 12b)x6 + (8 8a 48a2 16a3 + 40b +48ab 48a2 b + 24b2 )x5 + (12 28a 88a2 56a3 4a4 +48b + 136ab 96a2 b 64a3 b + 72b2 + 96ab2 72a2 b2 +16b3 )x4 + (8 36a 96a2 80a3 12a4 + 24b + 136ab 64a2 b 144a3 b 16a4 b + 76b2 + 240ab2 72a2 b2 96a3 b2 +40b3 + 64ab3 48a2 b3 + 4b4 )x3 + (2 20a 62a2 60a3 14a4 + 4b + 56ab 28a2 b 120a3 b 28a4 b + 32b2 + 208ab2 (13) +56ab + 36a2 b2 144a3 b2 24a4 b2 + 32b3 + 128ab3 24a2 b3 64a3 b3 + 8b4 + 16ab4 + 12a2 b4 )x2 + (4a 20a2 24a3 8a4 +8ab 16a2 b 48a3 b 16a4 b + 4b2 + 72ab2 + 48a2 b2 64a3 b2 24a4 b2 + 8b3 + 24ab4 16a3 b4 )x 2a2 4a3 2a4 4a2 b 8a3 b 4a4 b + 8ab2 + 8a2 b2 8a3 b2 4a4 b2 + 16ab3 +16a2 b3 16a3 b3 8a4 b3 + 8ab4 + 8a2 b4 8a3 b4 4a4 b4 , and Q(x) = 2(x + 1)2 (x + a)2 (x + b)2 . Utilizing Stirlings formula and the the following asymptotic formulas 1 1 + . (as x ), 2x 12x2 1 1 1 (x) + 2 + 3 + (as x ), x 2x 6x (x) log x (see [1, p.259]), we obtain
x

lim Ha,b (x) = lim Ha,b (x) = lim Ha,b (x) = 0.

(14)

(i) If b = a2 + 1 6 , using (12), (14) and lHospital rule, we obtain

lim n2 (n n+1 ) = lim = = = =

Ha,b (n) Ha,b (n + 1) 1/n2 Ha,b (n) Ha,b (n + 1) lim n 2/n3 1 lim n4 [Ha,b (n + 1) Ha,b (n)] 6 n 1 n4 Pa,b (n) lim 6 n Qa,b (n) 1 a2 + b. 6
n

(15)

Since (11) holds, an application of Lemma 1 yields

lim n n = a2 +

1 b=0 6

(16)

110

N. Batir

6 This proves (i). (ii) Let b = a2 + 1 6 and a = a1 , a2 . Then, the coecient of x in (13) vanishes, so using Stirlings formula we get

lim n3 (n n+1 ) = lim = = = =

Ha,b (n) Ha,b (n + 1) 1/n3 Ha,b (n) Ha,b (n + 1) lim n 3/n4 1 lim n5 [Ha,b (n + 1) Ha,b (n)] 12 n n5 Pa,b (n) 1 lim 12 n Qa,b (n) 1 (1 + a + 6a2 + 2a3 5b 6ab + 6a2 b 3b2 ). 3
n

(17)

If we set b = a2 +

1 6

in the last line, we get 1 3 3a4 4a3 + a2 + 1 12 .

lim n3 (n n+1 ) =

By virtue of Lemma 1, we nd that

lim n2 n =

1 6

3a4 4a3 + a2

1 12

= 0,

(18)

since a = a1 , a2 , this proves (ii). Now we shall prove (iii). Let (a, b) = (a1 , b1 ). In this case the coecients of x6 and x5 in (13) vanish, thus by Stirlings formula we obtain
n

lim n4 (n n+1 ) = lim = = = =

Ha,b (n) Ha1 , b1 (n + 1) 1/n4 Ha1 , b1 (n) Ha1 , b1 (n + 1) lim n 4/n5 1 lim n5 [Ha1 , b1 (n + 1) Ha1 ,b1 (n)] 20 n 1 n6 Pa,b (n) lim 20 n Qa,b (n) 0.014723740642427497...,
n

(19)

which implies by Lemma 1

lim n3 n (a1 , b1 ) = 0.004907913547475832.

(20)

In the same way we nd that

lim n3 n (a2 , b2 ) = 0.005746538709721112.

(21)

Improving Stirlings formula

111

If we set (a, b) = (a1 , b1 ) and (a, b) = (a2 , b2 ) in (13), respectively, we obtain Pa1 , b1 (x) = 0.393886 1.93247x 3.31813x2 2.35329x3 0.58895x4 < 0 and Pa2 , b2 (x) = 1.47415 + 5.06985x + 6.40706x2 + 3.50099x3 +0.689585x4 > 0,

(22)

(23)

and Ha1 , b1 (x + 1) Ha1 , b1 (x) < 0 and Ha2 , b2 (x + 1) Ha2 , b2 (x) > 0 by (12) for all x 0. Utilizing (14) and Lemma 2 we see that Ha1 , b1 (x) > 0 and Ha2 , b2 (x) < 0 for all x 0. From (14) it results that Ha1 , b1 is strictly decreasing and Ha2 , b2 is strictly increasing on [0, ), thus we have 1 1 0 = lim Ha1 , b1 (x) < Ha1 , b1 (x) Ha1 , b1 (0) = log(2 ) + a1 log b1 x 2 2 and 1 1 log(2 ) + a2 log b2 = Ha2 , b2 (0) Ha2 , b2 (x) < lim Ha2 , b2 (x) = 0. x 2 2 These can be written equivalently as follows: Theorem 2. For all reals x 0 the following double inequalities hold: exa1 (x + a1 )x where the constants = 2 = 2.5066282746 . . . and e a1 = = 2.5347503081133245 . . . b1 x + b1 < (x + 1) exa1 (x + a1 )x x + b1 (24)

are the best possible constants; and exa2 (x + a2 )x where the constants e a2 = = 2.5007041931433194 . . . b2 and = 2 = 2.506628274 . . . x + b2 < (x + 1) (x + a2 )x exa2 x + b2 (25)

are the best possible constants, and ai , bi (i=1,2) are as given in (7), (8) and (9). Remark 1. If we set (a, b) = (, )( = (13), respectively, we get
3 3 6 )

and (a, b) = (, )( =

3+ 3 6 )

in

P, (x) = 0.05772 0.69205x 2.79458x2 4.45859x3 3.07122x4 0.7698x5 < 0

112 and

N. Batir

P, (x) = 1.31081 + 6.32168x + 11.8316x2 + 10.6808x3 +4.62678x4 + 0.7698x5 > 0, that is, H, (x + 1) H, (x) < 0 and H, (x + 1) H, (x) > 0 by (12) for all x 0. Hence by Lemma 2 and (14), H, is strictly decreasing and H, is strictly increasing on [0, ). It results that
x

1 lim H, (x) = 0 H, (x) < log(2 ) + = H, (0), 2

and

1 H, (0) = log(2 ) + < H, (x) 0 = lim H, (x). x 2 These lead to the following double inequalities: 2e e
e , 2

x+ e and

1 x+ 2

< (x + 1) <

2e e

x+ e

x+ 1 2

where = where =

2e e

x+ e

x+ 1 2

< (x + 1) <

2e e

x+ e

x+ 1 2

e . 2

Precisely, these are the main results obtained in [12].

Remark 2. Equations (15) and (18) enable us to measure the speed of convergence of the approximations n! n , n! n , n! n , n! n and n! n from (1)-(5): we obtain from (16)
n

lim n log(n!/n ) = lim n n (0, 0) =

1 , 6 1 , 12 1 , 6

lim n log(n!/n ) = lim n n (1/2, 1/2) = lim n log(n!/n ) = lim n n (1, 1) =

and from (19) for a = 0

lim n2 log(n!/n ) = lim n2 n (0, 1/6) =

1 = 0.0138888 . . . , 72

and nally
n

lim n2 log(n!/n ) = lim n2 n (, ) = 0.01603750 . . . .

We let n = ea1

n + a1 e

2 (n + b1 )

Improving Stirlings formula

113

and n = ea2 Then from (20) and (21) we obtain

n + a2 e

2 (n + b2 ).

lim n3 log(n!/n ) = 0.0049079 . . . and

lim n3 log(n!/n ) = 0.005746538 . . . .

These prove theoretically the great superiority of our estimations n! n and n! n over n! n , which is the best of (1)-(5). We conclude our paper with a numerical comparison between our estimations n! n , n! n and n! n , n! n , where n is Karatsubas asymptotic formula for (n + 1) given on the second page. n 1 2 10 25 50 100 |n n!| 0.00397819 0.0026369 239.175 1.6881020 8.362041058 4.3310151 |n n!| 0.0003378 0.00020675 7.81549 2.5341018 6.5871056 2.60210149 |n n!| 0.0006483 0.00029994 7.67069 2.2941018 5.7941056 2.25610149 |n n!| 0.0003037 0.000026 0.017993 7.8411014 4.7891052 4.58510144

References
[1] M. Abramowitz, I. A. Stegun, Handbook of Mathematical Functions with Formulas, Graphs and Mathematical Tables, Dover, New York, 1972. [2] H. Alzer, Inequalities for the volume of the unit ball in Rn II, Mediterr. J. Math. 5(2008), 395413. [3] N. Batir, Inequalities for the gamma function, Arch. Math. (Basel) 91(2008), 554 563. [4] N. Batir, Some new inequalities for gamma and polygamma functions, JIPAM. J. Inequal. Pure Appl. 6(2005), Article 103. k 1 n [5] N. Batir, Integral representations of some series involving 2 k and some rek lated series, Appl. Math. Comput. 147(2004), 645667. 3k 1 n k [6] N. Batir, On the series k x , Proc. Indian Acd. Sci.Math. Sciences, k=1 k 115(2005), 371381. [7] C. Berg, H. L. Pedersen, A Pick function related to the sequence of volumes of the unit ball in n-space, Proc. Amer. Math. Soc., to appear. [8] W. Burnside, A rapidly convergent series for log N!, Messenger Math. 46(1917), 157159. [9] CP. Chen, Complete monotonicity and logarithmically complete monotonicity properties of function for the gamma and psi functions, J. Math. Anal. Appl. 336(2007), 812822. [10] A. Elbert, A. Laforgia, On some properties of the gamma function, Proc. Amer. Math. Soc. 128(2000), 26672673. [11] E. A. Karatsuba, On the asymptotic representations of the Euler gamma function by Ramanujan, J. Comput. Appl. Math. 135(2001), 225240. [12] C. Mortici, An ultimate extremely accurate formula for approximation of the factorial function, Arch. Math (Basel) 93(2009), 3745.

114

N. Batir

[13] C. Mortici, New approximations of the gamma function in terms of the digamma function, Appl. Math. Lett. 23(2010), 97100. [14] G. Nemes, More accurate approximations for the gamma function, Thai J. Math. 9(2011), 21-28. [15] S. Ponnusamy, M. Vuorinen, Asymptotic expansions and inqualities for hypergeometric functions, Mathematica 44(1997), 278301. [16] F. Qi, BN. Guo, Completely monotonic functions involving divided dierences of the di and trigamma functions and some applications, Commun. Pure Appl. Anal. 8(2009), 19751989. [17] F. Qi, BN. Guo, Necessary and sucient conditions for functions involving the triand tetra-gamma functions to be completely monotonic, Adv. Appl. Math. 44(2010), 7183. [18] SL. Qiu, M. Vuorinen, Some properties of the gamma and psi functions with applications, Math. Comp. 74(2005), 723742.