- Research
- Open access
- Published:
The infimum values of two probability functions for the Gamma distribution
Journal of Inequalities and Applications volume 2024, Article number: 5 (2024)
Abstract
Let α, β be positive real numbers and let \(X_{\alpha ,\beta}\) be a Gamma random variable with shape parameter α and scale parameter β. We study infimum values of the function \((\alpha ,\beta )\mapsto P\{X_{\alpha ,\beta}\le \kappa E[X_{\alpha ,\beta}] \}\) for any fixed \(\kappa >0\) and the function \((\alpha ,\beta )\mapsto P\{|X_{\alpha ,\beta}-E[X_{\alpha ,\beta}]| \le \sqrt{\operatorname{Var}(X_{\alpha ,\beta})}\}\). We show that \(\inf_{\alpha ,\beta}P\{X_{\alpha ,\beta}\le E[X_{\alpha ,\beta}]\}= \frac{1}{2}\) and \(\inf_{\alpha ,\beta}P\{|X_{\alpha ,\beta}-E[X_{\alpha ,\beta}]|\le \sqrt{\operatorname{Var}(X_{\alpha ,\beta})}\}=P\{|Z|\le 1\}\approx 0.6827\), where Z is a standard normal random variable.
1 Introduction
Special probability distributions play a fundamental role in probability theory, statistics, optimization, and different research fields of science, including physics, chemistry, bioscience, economy, and management science. Although they have been studied for a long time, our understanding of them is far from complete. This paper is motivated by Chvátal’s conjecture for the binomial distribution and Tomaszewaki’s conjecture for the Rademacher sequence, both of which were completely solved very recently.
Let \(B(n,p)\) denote a binomial random variable with parameters n and p. Janson [5] introduced the following conjecture suggested by Vašk Chvátal.
Conjecture 1
(Chvátal)
For any fixed \(n\geq 2\), as m ranges over \(\{0,\ldots ,n\}\), the probability \(q_{m}:=P\{B(n,\frac{m}{n})\leq m\}\) is the smallest when m is the integer closest to \(\frac{2n}{3}\).
Chvátal’s conjecture has interesting applications in machine learning. Janson [5] proved Chvátal’s conjecture for sufficiently large n. Barabesi et al. [1] and Sun [12] showed that Chvátal’s conjecture is true for any \(n\geq 2\).
The second motivation of this paper is the following problem, attributed to Boguslav Tomaszewski.
Conjecture 2
(Tomaszewski)
Let \(X=\sum_{i=1}^{n}a_{i}X_{i}\), where \(\sum_{i=1}^{n} a_{i}^{2}=1\) and \(\{X_{i}, i\ge 1\}\) is a sequence of independent \(\{-1,1\}\)-valued symmetric random variables. Then \(P\{|X|\leq 1\}\geq 1/2\).
Tomaszewski’s conjecture has many applications in probability theory, geometric analysis, and computer science. Recently, Keller and Klein [6] completely solved Tomaszewski’s conjecture. We refer the reader to Keller and Klein [6] for the details and to Dvorak and Klein [3] and Hu et al. [4] for some related problems.
In this paper, we will focus on the Gamma distribution. It is well known that the Gamma distribution, including the exponential distribution and the \(\chi ^{2}\)-distribution as two important special cases, is one of the most basic probability distributions (cf. [7, Chap. 16]). It is frequently applied to describe the time between independent events that occur at a constant average rate. The Gamma distribution has many significant applications. For example, it has been used to model the size of insurance claims, rainfall, failure times of repairable systems, load levels for telecommunication services, and the distribution of asset prices.
Motivated by Chvátal’s conjecture, Li et al. [9] initiated the study of the infimum value problem for special probability distributions. Let \(\{Y_{\lambda},\lambda >0\}\) be a family of random variables with the same distribution F but different parameters λ. Define
Li et al. [9] discussed the infimum value of the function r and gave a complete answer if F is the Poisson distribution or the geometric distribution. Further, Li et al. [8] considered the infimum value problem for the Weibull and Pareto distributions.
In the first part of this paper, we will consider the following more general infimum value problem for the Gamma distribution. Let \(\alpha ,\beta ,\kappa >0\) and let \(X_{\alpha ,\beta}\) be a Gamma random variable with shape parameter α and scale parameter β. Define
For fixed κ, what is the infimum value of the function \(g_{\kappa}(\alpha ,\beta )\)? In Sect. 2, we will give a complete answer to this question. Interestingly, we discover an unnoticed phase transition phenomenon (cf. Figs. 1–4 and Remark 2.2) and obtain the following result.
Theorem 1.1
Let α, β be arbitrary positive real numbers and let \(X_{\alpha ,\beta}\) be a Gamma random variable with shape parameter α and scale parameter β. Then
and
In the second part of this paper, we will prove a more interesting and deeper result.
Theorem 1.2
Let α, β be arbitrary positive real numbers, let \(X_{\alpha ,\beta}\) be a Gamma random variable with shape parameter α and scale parameter β, and let Z be a standard normal random variable. Then
and
We would like to point out that if \(0<\kappa \neq1\), then the following more general inequality might not hold for all Gamma random variables:
For example, we have
and
Note that the Gamma distribution is unimodal (cf. [10, Example 23.3]). The Vysochanskiĭ–Petunin inequality for unimodal distributions (cf. [2, Theorem 1.12] and [15]) tells us that
The estimate (1.4) is non-trivial if \(\kappa >1\); however, it yields a lower bound of 0 for the probabilities \(P\{|X_{\alpha ,\beta}-E[X_{\alpha ,\beta}]|\le \sqrt{\operatorname{Var}(X_{\alpha ,\beta})} \}\), \(\alpha ,\beta >0\).
The remainder of this paper is organized as follows. In Sect. 2, we investigate the infimum value problem for the function \(g_{\kappa}(\alpha ,\beta )\) and give the proof of Theorem 1.1. In Sects. 3 and 4, we discuss the variation comparison between the Gamma distribution and the normal distribution and give the proof of Theorem 1.2. In Sect. 5, we make concluding remarks. The software Mathematica has been used in Sect. 4. Some Mathematica calculations are given in the Appendix.
2 The infimum value problem for the function \(g_{\kappa}(\alpha ,\beta )\)
Let \(\alpha ,\beta >0\) and let \(X_{\alpha ,\beta}\) be a Gamma random variable with probability density function
For \(\kappa >0\), we consider the infimum value of the probability function \(g_{\kappa}(\alpha ,\beta )\) defined by (1.1).
We have
Then we may assume without loss of generality that \(\beta =1\) and focus on the infimum value of the following function:
By Euler’s reflection formula
we get
It follows that for any \(\kappa >0\), \(\sup_{\alpha >0}h_{\kappa}(\alpha )=1\).
2.1 Case \(\kappa \le 1\)
In this subsection, we assume that \(\kappa \le 1\). We will show that \(h_{\kappa}(\alpha +1)< h_{\kappa}(\alpha )\) for any \(\alpha >0\). In fact, we have
Note that \((1+\frac{1}{x})^{x}\) is strictly increasing with respect to \(x>0\) and the limit equals e. Then we have
Thus,
Hence, by (2.2) and (2.3), we obtain that \(\{h_{\kappa}(\alpha +n)\}_{n\in \mathbb{N}}\) is a strictly decreasing sequence.
Let \(Y_{0},Y_{1},Y_{2},\dots \) be independent random variables such that \(Y_{0}\sim \operatorname{Gamma} (\alpha ,1)\) and \(Y_{i}\sim \operatorname{Gamma} (1,1)\), \(i\ge 1\). Then, for \(\kappa \in (0,1)\), by the strong law of large numbers, we get
and for \(\kappa =1\), by the central limit theorem, we get
Therefore,
which implies that inequality (1.2) holds. The proof of Theorem 1.1 is complete.
Remark 2.1
It is interesting to compare (1.2) with the well-known fact that the Gamma distribution is skewed to the right and its skewness is given by
2.2 Case \(\kappa > 1\)
In this subsection, we assume that \(\kappa > 1\). This case is especially interesting. By (1.2), for any \(\alpha >0\), we have
Let \(Y_{1},Y_{2},\dots \) be independent \(\operatorname{Gamma} (1,1)\) random variables. By the central limit theorem, we get
Hence, by (2.1), (2.4), and (2.5), we obtain
Further, by virtue of Mathematica, we obtain the following numerical results:
Graphs of the function \(h_{\kappa}(\alpha )\) for different values of κ are shown in Figs. 1–4.
Remark 2.2
The above analysis shows that there is an interesting phase transition phenomenon in the infimum value problem for the Gamma distribution. The critical point is \(\kappa =1\) and the behaviors for the three cases \(\kappa <1\), \(\kappa =1\), and \(\kappa >1\) are totally different.
3 Variation comparison between the Gamma distribution and the normal distribution
The empirical rule tells us that in an approximately normal distribution, about 68% of the values fall within one standard deviation of the mean, about 95% of the values fall within two standard deviations of the mean, and almost all values fall within three standard deviations of the mean. A question worth thinking about is what we can say about other distributions. Theorem 1.2 compares the variation between the Gamma distribution and the normal distribution. In this section, we will prove Theorem 1.2.
Let \(\alpha ,\beta >0\). Define
We have
Then we may assume without loss of generality that \(\beta =1\) and focus on the following function:
This section is devoted to proving the following result.
Theorem 3.1
For any \(\alpha >0\),
Note that by the central limit theorem, \(\lim_{\alpha \rightarrow \infty}t(\alpha )=P\{|Z|\leq 1\}\), where Z is a standard normal random variable. Once (3.1) is proved, we conclude that
Thus, inequality (1.3) holds and hence the proof of Theorem 1.2 is complete.
Proof of Theorem 3.1
First, we consider the case \(0<\alpha \le 1\). We have
Set
Then
Hence,
Define
Set
Then we have
which together with (3.2) and (3.3) implies that (3.1) holds for the case \(0<\alpha \le 1\).
Next we consider the case \(\alpha >1\). We have
The remainder of this section is devoted to proving the following inequality:
which implies (3.1). It is a bit surprising that inequality (3.4) is very delicate, which seems to be unknown in the literature. □
3.1 Proof of the “<1” part of inequality (3.4)
Note that
Then we have
Hence, to prove the “<1” part of inequality (3.4), it suffices to show that
We have
and
Then, for fixed α, \(1- (1+\frac{y}{\alpha +\alpha ^{\frac{1}{2}}} )^{\alpha}e^{-y}\) is an increasing and concave function of y on \([0,1]\). Hence, we have
Thus, to complete the proof of the “<1” part of inequality (3.4), it suffices to show that
which is equivalent to
The proof of (3.5) will be given in Sect. 4.
3.2 Proof of the “>1” part of inequality (3.4)
Note that
We consider the three cases \(1<\alpha \le 2\), \(2<\alpha < (\frac{15}{8} )^{2}\), and \(\alpha \ge (\frac{15}{8} )^{2}\) separately.
Case 1: \(1<\alpha \le 2\). By Taylor’s formula, we have
Set
Then
and
We have
and
Then
Thus, the “>1” part of inequality (3.4) holds for \(1<\alpha \le 2\).
Case 2: \(2<\alpha < (\frac{15}{8} )^{2}\). By Taylor’s formula, we have
Define
Set
Then
Hence,
Thus, the “>1” part of inequality (3.4) holds for \(2<\alpha < (\frac{15}{8} )^{2}\).
Case 3: \(\alpha \ge (\frac{15}{8} )^{2}\). Note that for \(0\leq y\leq 1-(\alpha +1)^{\frac{1}{2}}+\alpha ^{\frac{1}{2}}\), we have
Then
Hence, to prove the “>1” part of inequality (3.4), it suffices to show that
For \(y\in [0,1]\), we have
and
Then, for fixed α, \((1+\frac{y}{\alpha -\alpha ^{\frac{1}{2}}} )^{\alpha}e^{-y}-1\) is an increasing and concave function of y on \([0,1]\). Hence, we have
Thus, to complete the proof of the “>1” part of inequality (3.4), it suffices to show that
which is equivalent to
4 Proofs of inequalities (3.5) and (3.6)
4.1 Proof of (3.5)
Define
We have
Set
Then, by the condition \(\alpha >1\), we get
and
Define
By virtue of Mathematica, we get
Set
We have
which implies that \(P_{+}\) is negative. By virtue of Mathematica, we get
Further, by the transformation (4.2), we obtain
which implies that \(Q_{+}\) is negative. Thus, by (4.1), we get
Define
and
By virtue of Mathematica, we obtain the expansion of the polynomial \(V_{+}(q)\) (see Sect. 5). Note that all terms in the expansion are positive. Then \(R_{+}\) is negative. Therefore, the proof is complete by (4.3).
4.2 Proof of (3.6)
Define
We have
Set
Then, by the condition \(\alpha \geq (\frac{15}{8} )^{2}\), we have
and
Define
We have
Set
We get
which implies that \(P_{-}\) is positive. We have
By the transformation (4.5), we get
which implies that \(Q_{-}\) is positive. Thus, by (4.4), we get
Define
and
By virtue of Mathematica, we obtain the expansion of the polynomial \(V_{-}(q)\) (see Sect. 5). Note that all terms in the expansion are positive. Then \(R_{-}\) is positive. Therefore, the proof is complete by (4.6).
5 Concluding remarks
It is worth pointing out that the software Mathematica plays an important role in our work, which provides us with deep insight into how to handle delicate inequalities. Although it is possible to figure out clever, traditional methods to establish the positiveness of polynomials \(V_{+}(q)\) and \(V_{-}(q)\) in Sect. 4, it seems more natural to use the computer to quickly present expansions so as to solve the problem. Actually, the computer-assisted proof given for inequality (1.3) also explains why it was not discovered before. We would like to emphasize that all proofs contained in this paper remain rigorous and easily verifiable.
A distribution on \(\mathbb{R}\) is infinitely divisible if it can be expressed as the distribution of the sum of an arbitrary number of i.i.d. random variables. We know that the Gamma distribution is infinitely divisible and each Lévy process can be associated with an infinitely divisible distribution (cf. [10, 11]). Motivated by Theorem 1.2, it is natural to ask if any infinitely divisible random variable L satisfies the following inequality:
In the forthcoming papers [13, 14], we prove that inequality (5.1) holds for many familiar infinitely divisible continuous distributions including the Laplace, Gumbel, logistic, Pareto, infinitely divisible Weibull, log-normal, Student’s t, inverse Gaussian, and F distributions.
Inequality (5.1) has the potential to be used in fitting probability distributions to data. Given a data set, we compute the proportion of values that fall within one sample standard deviation of the sample mean. If this proportion is bigger than 0.6827, then we consider using a Gamma distribution or another infinitely divisible distribution that was mentioned above to fit the data. Otherwise, if the proportion is much smaller than 0.6827, then we should avoid using those familiar distributions to describe the data.
Data availability
This manuscript has no associated data.
References
Bababesi, L., Pratelli, L., Rigo, P.: On the Chvátal-Janson conjecture. Stat. Probab. Lett. 194, 109744 (2023)
Dharmadhikari, S., Joag-Dev, K.: Unimodality, Convexity, and Applications. Academic Press, San Diego (1988)
Dvorak, V., Klein, O.: Probability mass of Rademacher sums beyond one standard deviation. SIAM J. Discrete Math. 36, 2393–2410 (2022)
Hu, Z.-C., Lan, G.-L., Sun, W.: Some explorations on two conjectures about Rademacher sequences. Acta Math. Appl. Sin. Engl. Ser. 37, 1–16 (2021)
Janson, S.: On the probability that a binomial variable is at most its expectation. Stat. Probab. Lett. 171, 109020 (2021)
Keller, N., Klein, O.: Proof of Tomaszewski’s conjecture on randomly signed sums. Adv. Math. 407, 108558 (2022)
Krishnamoorthy, K.: Handbook of Statistical Distributions with Applications, 2nd edn. CRC Press, Boca Raton (2016)
Li, C., Hu, Z.-C., Zhou, Q.-Q.: A study on the Weibull and Pareto distributions motivated by Chvátal’s theorem (2023). arXiv:2305.02114
Li, F.-B., Xu, K., Hu, Z.-C.: A study on the Poisson, geometric and Pascal distributions motivated by Chvátal’s conjecture. Stat. Probab. Lett. 200, 109871 (2023)
Sato, K.: Lévy Processes and Infinitely Divisible Distributions. Cambridge University Press, Cambridge (1999)
Steutel, F.W., Van Harn, K.: Infinite Divisibility of Probability Distributions on the Real Line. Dekker, New York (2003)
Sun, P.: Strictly unimodality of the probability that the binomial distribution is more than its expectation. Discrete Appl. Math. 301, 1–5 (2021)
Sun, P., Hu, Z.-C., Sun, W.: Variation comparison between infinitely divisible distributions and the normal distribution (2023). arXiv:2304.11459
Sun, P., Hu, Z.-C., Sun, W.: Variation comparison between the F-distribution and the normal distribution (2023). arXiv:2305.13615
Vysochanskiĭ, D.F., Petunin, Y.I.: Justification of the 3σ rule for unimodal distributions. Theory Probab. Math. Stat. 21, 25–36 (1980)
Acknowledgements
We thank the anonymous reviewers and Professor Jewgeni Dshalalow, the Editor, for their valuable suggestions and comments, which lead to improvement of the paper. This work was supported by the National Natural Science Foundation of China (No. 12171335), the Science Development Project of Sichuan University (No. 2020SCUNL201), and the Natural Sciences and Engineering Research Council of Canada (No. 4394-2018).
Funding
This work was supported by the National Natural Science Foundation of China (No. 12171335), the Science Development Project of Sichuan University (No. 2020SCUNL201), and the Natural Sciences and Engineering Research Council of Canada (No. 4394-2018).
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: Mathematica calculations
Appendix: Mathematica calculations
The expansion of the polynomial \(V_{+}(q)\):
V+=23565171557938261664962395 + 1985238765536369188253388462 q^2 + 76017937191609745093093565184 q^4 + 1815476155917282265018752272232 q^6 + 30868042081839055982554050213660 q^8 + 401897536051918258546845673711320 q^10 + 4195397709111549929883773768957292 q^12 + 36238699732610615067411794056699104 q^14 + 265002286089679374723172860122766982 q^16 + 1669237124849349342077586449716389470 q^18 + 9179934813394932229977676436328785920 q^20 + 44555295354320392501114611345123622400 q^22 + 192537160208281140648975165919934835200 q^24 + 746181252269526741637909507751082171520 q^26 + 2609372572626683435719917787491018652160 q^28 + 8276209631283583168755734561689661224960 q^30 + 23913569456882144063241575623509484876800 q^32 + 63185851825755484161960668172292699909120 q^34 + 153171842040744452342444666253152790732800 q^36 + 341628452529444844018632398179833131991040 q^38 + 702775058219816773204544225960017412751360 q^40 + 1336281286191807830241756821296507838955520 q^42 + 2352931028757956298911312671496544634142720 q^44 + 3842851078067257537573706091171559356497920 q^46 + 5829597689354288278514821031866786159656960 q^48 + 8224048629268397888867021111129844469596160 q^50 + 10800344227796355322915422778734559920914432 q^52 + 13214925874596962589048294078754839901241344 q^54 + 15075437985869745487585690312752934894436352 q^56 + 16043253396277674458218536937392302561689600 q^58 + 15933483495160554733717977505944446676500480 q^60 + 14772181225346687498328707734409833005187072 q^62 + 12786642631638024827680853914736629691449344 q^64 + 10333607587858511627607426462208954269171712 q^66 + 7796135691442288431829216828566360534548480 q^68 + 5489455045169343972022956242977322445045760 q^70 + 3606053976460179205452292516224406577479680 q^72 + 2208802483012122361676530694278736018145280 q^74 + 1260683716848402925787749700503070572544000 q^76 + 669904634456217504368236284579198848204800 q^78 + 331081492020125941589702200450146757509120 q^80 + 152001230583526972614803279225239581491200 q^82 + 64734238129983061042604032362158017740800 q^84 + 25531661312772564369272230408940525977600 q^86 + 9307900274880934185285303838052660019200 q^88 + 3129622227458365558314112917099983667200 q^90 + 968033669852811173795309928049750835200 q^92 + 274640462267821497605095290057418342400 q^94 + 71224038857339584104332613373237657600 q^96 + 16816854347488682979216179348688076800 q^98 + 3598246400042953802386878316412928000 q^100 + 693860503839280523254707460767744000 q^102 + 119794916777653504670468143054848000 q^104 + 18371891299642536871251828277248000 q^106 + 2478647665316721166509051740160000 q^108 + 290656583441325139861690122240000 q^110 + 29171448597603811259616067584000 q^112 + 2455470675466333890778497024000 q^114 + 168582129968383522599075840000 q^116 + 9065459012974557989437440000 q^118 + 358068461185282678456320000 q^120 + 9236522547766697656320000 q^122 + 116733302341443256320000 q^124.
The expansion of the polynomial \(V_{-}(q)\):
V-=1058023271132626023 + 51541890229923566472 q^2 + 1213009372688989850064 q^4 + 18352820646596071930240 q^6 + 200442482186879766344000 q^8 + 1682464063207304317242816 q^10 + 11285809233594557704985856 q^12 + 62123650712872430361438720 q^14 + 286006349074965960756670464 q^16 + 1116988330180696358380290048 q^18 + 3741070988530167056939876352 q^20 + 10836728622922107883411734528 q^22 + 27330768999389436608140804096 q^24 + 60330629581678789398471114752 q^26 + 117040259220868123540341129216 q^28 + 200169684615441568277429485568 q^30 + 302487634652646366318853881856 q^32 + 404485176548048584076188188672 q^34 + 478958931700928292722105647104 q^36 + 502214198209105947113507782656 q^38 + 465950173145939141611449483264 q^40 + 381911302204202972478305206272 q^42 + 275855094787796630236532047872 q^44 + 174972859491619945161380855808 q^46 + 97001064005371803174963249152 q^48 + 46708058892337905269349548032 q^50 + 19376798019028218231165812736 q^52 + 6852048396846611541188935680 q^54 + 2036474622748306983572471808 q^56 + 499070059195604547617685504 q^58 + 98184545971566239935365120 q^60 + 14905936080354118622773248 q^62 + 1639091209276985243074560 q^64 + 116172982490204328689664 q^66 + 3984496719921263149056 q^68.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Sun, P., Hu, ZC. & Sun, W. The infimum values of two probability functions for the Gamma distribution. J Inequal Appl 2024, 5 (2024). https://doi.org/10.1186/s13660-024-03081-w
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s13660-024-03081-w