Bounds for the Gamma function¶
1. Bounds for \(\Gamma(x)\) for real \(x\)¶
The classical bounds (Stirling formula) that results from the Euler-Maclaurin formula for positive integers \(n\) and \(m \geq 0\) are
These inequalities were described in words in more general form in [Bendersky, 1933] and subsequently rediscovered several times [Choi, 2012], [Camargo, 2024].
[Alzer, 1997] presents an extension of the inequalities above for real arguments in lighlty different form (see also remark 2.1 of [Chen, 2016]).
Theorem (1997)
For real \(x > 0\) and \(m \geq 0\), we have \(\displaystyle \sum\limits_{j =1}^{2m} \frac{B_{2j}}{2j(2j-1)x^{2j-1}} \ \leq \ \log \Gamma(x) - \left(x-\frac{1}{2}\right)\log(x) + x - \log(\sqrt{2\pi}) \ \leq \ \sum\limits_{j =1}^{2m+1} \frac{B_{2j}}{2j(2j-1)x^{2j-1}}\).
There are also several bounds for \(\Gamma(x)\) with a fixed number of terms available in the literature. For instance, in p. 118 of Ramanujan’s lost notebook [Andrews and Berndt, 2013], one finds
Bounds of similar shape was obtaineb by [Mortici, 2011].
Theorem (2011)
For \(\alpha = \frac{128}{1215} \approx 0.105349\), \(\beta = \frac{218 336}{135} − \frac{256e^{24}}{43 046 721 \pi^4} \approx 0.087944\) and \(x \geq 3\), we have \(\displaystyle \sqrt[8]{16x^4 + \frac{32}{3}x^3+\frac{32}{9}x^2 - \frac{176}{405}x - \alpha} \ \leq \ \frac{\Gamma(x+1)}{\sqrt{\pi}\left(\frac{x}{e}\right)^x} \ \leq \ \sqrt[8]{16x^4 + \frac{32}{3}x^3+\frac{32}{9}x^2 - \frac{176}{405}x - \beta}\).
Improving on Ramanujan’s and some other estimates, [Chen, 2016] obtained the following result.
Theorem (2016)
For \(x \geq 2\), we have \(\displaystyle 1 - \frac{2117}{5080320x^7}\leq \ \frac{\Gamma(x+1)}{\sqrt{2\pi x}\left(\frac{x}{e}\right)^x \left( 1 + \frac{1}{12x^3+\frac{24}{7}x - \frac{1}{2}} \right)^{x^2+\frac{53}{210}}} \ \leq \ 1 - \frac{2117}{5080320x^7} + \frac{1892069}{2347107840 x^9}\).
Meanwhile, [Batir, 2008] presented some estimates for \(\Gamma(x)\) of certain shapes with optimal constants.
Theorem (2008)
For \(x > 0\), \(a = \sqrt{2e} = 2.33164\), and \(b = \sqrt{2\pi} = 2.50662...\) we have \(\displaystyle a\left( \frac{x+1/2}{e}\right)^{x+1/2} \leq \Gamma(x+1) < b\left( \frac{x+1/2}{e}\right)^{x+1/2}\). The constants \(a\) and \(b\) are the best possible.
Theorem (2008)
For \(x \geq 1\), \(a = 1/6\), and \(b = \frac{e^2}{2\pi} - 1 = 0.176005...\) we have \(\displaystyle x^xe^{-x} \sqrt{2\pi(x+a)} < \Gamma(x+1) < x^xe^{-x} \sqrt{2\pi(x+b)}\). The constants \(a\) and \(b\) are the best possible.
2. Bounds for the Digamma function \(\psi(x) = \Gamma'(x)/\Gamma(x)\) for real \(x\)¶
Let
In [Diamond and Straub, 2016], the authors prove the following result.
Theorem (2016)
Let \(N\ge 1\) and let \(\lambda_0\) be the unique root of \(B_N(\lambda)\) in \([0,1/2]\) for \(N\) even or the unique root of \(B_{N+1}(\lambda)\) for \(N\) odd. For \(x > \lambda\), we have \(\left\{ \begin{array}{cl} \psi(x) > F_N(\lambda,x), & N \equiv 1 \mod 4, \ \lambda \ \in \ [\lambda_0, 1/2]; \\ \psi(x) < F_N(\lambda,x), & N \equiv 3 \mod 4, \ \lambda \ \in \ [\lambda_0, 1/2]; \\ \psi(x) > F_N(\lambda,x), & N \equiv 2 \mod 4, \ \lambda \ \in \ [0, \lambda_0]; \\ \psi(x) < F_N(\lambda,x), & N \equiv 0 \mod 4, \ \lambda \ \in \ [0, \lambda_0]. \end{array}\right.\)
The first two cases of the theorem above for \(\lambda = 0\) yields the following result which was previously published in [Gordon, 1994].
Some variants of these inequalities for \(\psi(x)\) can be found in the literature. An example is the following.
Theorem ([Mortici, 2011])
For \(x \geq 1\), we have \(\displaystyle -\frac{1}{24x^2} + \frac{1}{12x^3} - \frac{337}{2280x^4} < \psi(x+1) + \log\left( e^{1/(x+1)} - 1 \right) < -\frac{1}{24x^2} + \frac{1}{12x^3} - \frac{337}{2280x^4} + \frac{97}{720x^5}\).
3. Bounds for the Polygamma functions \(\psi(x)', \psi(x)'', \psi(x)''' ...\) for real \(x\)¶
Let
[], proves the following result.
Theorem (1997)
For \(x > 0, k \geq 1\) and \(n \geq 0\), we have \(S_k(2n,x) < (-1)^{k+1} \psi^{(k)}(x) < S_k(2n+1,x)\).
The first two cases of the theorem above are
Similar inequalities were previously obtained by [Gordon, 1994]. For \(x > 0\),
Similar inequalities to those of [] were obtained in equation (3.4) of []
Theorem (2002)
For \(x \geq 1/2\), we have \(\displaystyle \frac{(n-1)!}{(x-\frac{1}{2})^n} + \sum\limits_{k = 1}^{2N+1} \frac{B_{2k}(1/2)}{(2k)!}\frac{(n+2k+1)!}{(x-\frac{1}{2})^{n+2k}} < (-1)^{(n+1)}\psi^{(n)}(x) < \sum\limits_{k = 1}^{2N} \frac{B_{2k}(1/2)}{(2k)!}\frac{(n+2k+1)!}{(x-\frac{1}{2})^{n+2k}}\).
The values \(B_{2k}(1/2)\) in the theorem above are known in explicit form ([Allasia et al., 2002])
Several other inequalities for \(\psi^{(k)}(x)\) can be obtained exploring the relations between \(\psi^{(k)}(x)\) and \(\psi(x)\). For instance, in [Guo and Qi, 2013], one finds (among other things)
4. Complex Stirling formula¶
The complex version of the Euler-maclaurin formula for the Gamma function (see, e.g., [Lang, 1999], p. 422) is
which holds for all nonzero complex numbers \(s\) which does not have negative real part. In the relation above, \(\log\) means the principal branch of the logarithm and \(B_1(x) = x - \frac{1}{2}\) is the Bernoulli polynomial of degree one. Using similar expressions with more terms of the Euler-Maclaurin formula, one might be able to deduce the complex version of Stirling’s formula (see [Gradshteyn and Ryzhik, n.d.], 8.344):
with
A few other estimates can be found in Section 20.2 of [Ramaré, 2021]. For \(0 < \delta < \pi\) and \(s = |s|e^{i\phi}\) with \(|\phi| \leq \pi - \delta\), the following holds
For \(s = \sigma + it\) with \(\sigma > 0\), we also have
and
4. Complex Stirling formula¶
The complex version of the Euler-maclaurin formula for the Gamma function (see, e.g., [Lang, 1999], p. 422) is
which holds for all nonzero complex numbers \(s\) which does not have negative real part. In the relation above, \(\log\) means the principal branch of the logarithm and \(B_1(x) = x - \frac{1}{2}\) is the Bernoulli polynomial of degree one. Using similar expressions with more terms of the Euler-Maclaurin formula, one might be able to deduce the complex version of Stirling’s formula (see [Gradshteyn and Ryzhik, n.d.], 8.344):
with
A few other estimates can be found in Section 20.2 of [Ramaré, 2021]. For \(0 < \delta < \pi\) and \(s = |s|e^{i\phi}\) with \(|\phi| \leq \pi - \delta\), the following holds
For \(s = \sigma + it\) with \(\sigma > 0\), we also have
and