- Research
- Open Access
- Published:
On the distribution-tail behaviour of the product of normal random variables
Journal of Inequalities and Applications volume 2023, Article number: 32 (2023)
Abstract
In this paper we consider the product \(\Pi _{n}=\prod_{k=1}^{n}\xi _{k}\) of n independent normally distributed zero mean random variables \(\xi _{1},\dots ,\xi _{n}\). We derive an asymptotic formula for the survival probability \({\mathbb{P}}(\Pi _{n}>x)\), as \(x\to \infty \), with the first remaining term.
1 Introduction
The normal distribution is the most popular continuous distribution in probability theory due to its wide range of applications in various fields of science and practice. We say that a random variable (r.v.) ξ has normal distribution if its density function has the form
where the parameter \(\mu \in{\mathbb{R}}\) is the expectation and \(\sigma >0\) is the standard deviation of ξ. The distribution function (d.f.) of the normal r.v. ξ is
In the case where \(\mu =0\) and \(\sigma =1\), the distribution above is called the standard normal distribution. In such a case,
In this paper we deal with the products of normal random variables. In general, the theory for the products of independent random variables is far less developed than the mature theory for the sums of independent random variables (we mention the books [5, 12], and [24, Chap. 4]).
In the case of two standard normal r.v.s., the exact expression for the distribution of their product was obtained almost a century ago. In [7, 28, 30], it was observed that for independent standard normal r.v.s ξ and η, the density function of the product ξη is
where \(K_{0}(\cdot )\) is the modified Bessel function of the second kind of order zero, i.e.,
Craig [7] also derived a closed-form expression for the exact density function \(f_{\xi \eta}\) in the case of dependent normally distributed r.v.s ξ and η, showing that essentially \(f_{\xi \eta}\) is of that form, reached when correlation \({\mathrm{Corr}}(\xi ,\eta )=0\), multiplied by an exponential function (see also [8, 13, 22]). Further developments of Craig’s result can be found in numerous subsequent papers. Haldane [15] derived the cumulant generating function, the first four moments, and the first four cumulants of ξη in the general case. Aroian [2] showed that the distribution of ξη can be approximated by a normal distribution when the ratios of the means to standard deviations of the normally distributed multipliers are large. Aroian et al. [3] and Simon [23] provided expressions for the distribution of ξη with normally distributed multipliers in the special cases of parameters and dependence structures. Recently, Gaunt [14] (see also the references therein) successfully used the Stein method to obtain the exact expression of the density function not only for the product of zero-mean normal multipliers but also for sum of such independent summands.
Concerning applications of the distribution of the product of two (correlated or not) normal variables, we mention [4], where the authors showed that a Lagrangian power distribution in two-dimensional turbulence is well described by Craig’s ξη distribution, and [19, 20] for testing of the indirect effect based on the distribution of the product of two normal variables. Along with the exact formula, the asymptotic behaviour of the tail distribution of the product of n normal r.v.s. is often of interest. An obvious application comes from the financial portfolio analysis, when calculating the compound return of the portfolio with normally distributed one-period returns and estimating tails of its distribution, as this could be beneficial for the portfolio decision making.
In Sect. 2 we formulate our main result, which is proved in Sect. 3.
2 Asymptotic behaviour of the product distribution tail
The asymptotic behaviour of the tail of the product of normal r.v.s. can be derived from exact formulas or using the saddle point method described in details by Butler [6], Fedoryuk [10], Jensen [16], for instance. Using the latter method described in [10], Arendarczyk and Dȩbicki [1] obtained the following result (see their Lemma 2.1) on the Weibull tail distributions.
Theorem 2.1
Let \(\xi _{1}\) and \(\xi _{2}\) be two independent, nonnegative r.v.s such that
where \(A_{i}>0\), \(\gamma _{i}\in {\mathbb{R}}\), \(\beta _{i}>0\), \(\alpha _{i}>0\), \(i=1,2\). Then
where
In particular, this theorem implies that if \(\xi _{1}\), \(\xi _{2}\), \(\xi _{3}\) are i.i.d. standard normal r.v.s, then
Let us describe the derivation of the above formulas. At first, suppose \(n=2\). Random variables \(\xi _{1}\), \(\xi _{2}\) are distributed according to a symmetric law Φ. Therefore,
where here and everywhere below \(\eta ^{+}\) denotes the positive part of an r.v. η.
For \(x>0\), by partial integration, we obtain (see also [11, Sect. 7.1])
To derive an asymptotic formula for \(\overline{F}_{\xi ^{+}_{1}\xi ^{+}_{2}}(x)\), we apply Theorem 2.1 with
After some calculations we get
The asymptotics in (5), together with equality (3), implies the first formula in (2).
In the case of \(n=3\), similarly to the formula (3), it holds that
To derive the asymptotics for \(\overline{F}_{\xi _{1}^{+}\xi _{2}^{+}\xi _{3}^{+}}(x)\), we can apply Theorem 2.1 again by using the asymptotic formula (5) and the estimate (4). According to these relations, the parameters in Theorem 2.1 are the following:
Using the formulas of Theorem 2.1, we derive
with \(\alpha =2/3\), \(\beta =3/2\), \(\gamma =-1/3\), and \(A={1}/{(2\sqrt{6\pi})} \). To obtain the second formula in (2), it is enough to use formula (6).
Note that the product of normally distributed r.v.s (or Weibull r.v.s in general), which are light-tailed, can produce heavy-tailed distributions, as seen from (2). We refer to [9, 17, 18, 21, 27] and the references therein for various extensions related to this phenomenon.
In our paper, we derive a more precise formula for the tail of the product of normal r.v.s. The following two statements are the main results of the paper.
Theorem 2.2
Let \(\xi _{1},\xi _{2},\dots \) be i.i.d. r.v.s such that, for each k, \(\xi _{k}\) is distributed according to Φ and let \(\Pi _{n}:=\prod_{k=1}^{n}\xi _{k}\). Then
where the constant in \(O_{n}\) depends on n.
Corollary 2.1
Let \(\widehat{\xi}_{1},\widehat{\xi}_{2},\dots \) be independent r.v.s such that, for each k, \(\widehat{\xi}_{k}\) is distributed according to \({\Phi}_{0,\sigma _{k}}\) and let \(\widehat{\Pi}_{n}:=\prod_{k=1}^{n}\widehat{\xi}_{k}\). Then
where \(\sigma ^{(n)}:=\prod_{k=1}^{n}\sigma _{k}\) and the constant in \(O_{n}\) depends on n and \(\sigma ^{(n)}\).
Remark 2.1
The assertion of Corollary 2.1 follows from Theorem 2.2 immediately, because the r.v. \(\xi _{k}=\widehat{\xi}_{k}/\sigma _{k}\) is distributed according to Φ for each k and
Remark 2.2
Note that the density of product \(\widehat{\Pi}_{n}\), as in Corollary 2.1, can be written in terms of the Meijer G-function (see [25, 26]),
In the case \(n=2\), this formula gives the modified Bessel function of order zero,
cf. (1). In our proof, we use a different approach based on the saddle point method which does not require knowledge of G-functions and their properties. Moreover, it allows obtaining higher-order terms in the result (7).
3 Proof of the main theorem
We start with the formulation of an auxiliary lemma, which we use repeatedly in our proof. The lemma below can be obtained by applying the saddle point method to a real integral of a special form. The proof of the lemma can be found in [29] (see Theorem 1, Chap. II).
Lemma 3.1
Let h and g be two real functions defined on an interval \([a,b)\) (b can be finite or infinite) such that:
as \(z\searrow a\) for all \(N\geqslant 1\), where \(a_{0}\neq 0\), \(b_{0}\neq 0\), \(\mu >0\), and \(\alpha >0\);
If the integral \(\int _{a}^{b} g(z){\mathrm{e}}^{-xh(z)}{\,\mathrm{d}}z\) converges absolutely for all sufficiently large x, then
for all \(N\in {\mathbb{N}}\), where \({\Gamma}(\cdot )\) denotes the Gamma function and coefficients \(d_{k}\) are expressible in terms of \(a_{k}\) and \(b_{k}\).
Wong [29] also provided the explicit forms of the first three coefficients:
Proof of Theorem 2.2
We use induction on n. The estimate (4) implies the assertion of Theorem 2.2 in the case \(n=1\).
Suppose \(n=2\). By the symmetry of the distribution Φ, for \(x>0\), we have
Obviously,
for some positive constant \(C_{11}\).
For integral \(I_{1}\), due to the first inequality in (4), we have
where the variable change \(y=\sqrt{xu}\) was used in the second step.
Using Lemma 3.1 for integral \(I_{12}\) (with \(\mu =2\), \(\alpha =1\), \(a_{0}=b_{0}=1\), \(a_{1}=-1\), and \(b_{1}=0\)), we get
because \(u+\frac{1}{u}| _{u=1} = 2\), and
for all \(N\geqslant 1\) due to the Taylor formula.
In a similar way, using Lemma 3.1 (with \(\mu =2\), \(\alpha =1\), \(a_{0}=b_{0}=1\), \(a_{1}=-1\), and \(b_{1}=-2\)), we derive
because of (12) and the following decomposition:
according to the Taylor formula.
Estimates (10), (11), and (13) imply that for some \(C_{12}>0\),
Substituting the derived estimates (9) and (14) into (8), we obtain
with a positive constant \(C_{2}\).
Now let us consider the lower bound for \(\overline{F}_{\xi _{1}\xi _{2}}(x)\). Equality (8) and the lower bound in (4) imply
Thus, due to the estimates (11) and (13),
for a positive constant \(C_{3}\). Inequalities (15) and (16) imply the assertion of the theorem in the case \(n=2\).
Let us suppose now that formula (7) holds for \(n=m\), \(m\geqslant 2\), i.e., let
Obviously,
According to (4),
with a positive quantity \(C_{4m}\) depending only on m.
Using the induction hypothesis (17) and the variable change \(y=x^{1/(m+1)}u^{1/2}\), we get
where \(C_{5m}\) and \(C_{6m}\) are positive quantities depending only on m, and
Using Lemma 3.1 (with \(\mu =2\), \(\alpha =1\), \(a_{0}=(m+1)/(2m)\), \(a_{1}=-(m+1)(2m+1)/(6m^{2})\), \(b_{0}=1\), and \(b_{1}=-(m-1)/(2m)\)), we obtain
because \(m u^{-1/m}+u | _{u=1}=m+1\), and
for each \(N\geqslant 1\) due to the Taylor formula.
Concerning the term \(J_{11}\), after the variable change \(u=1/v\), we get
Therefore, using Lemma 3.1 again (with \(\mu =2\), \(\alpha =1\), \(a_{0}=(m+1)/(2m)\), \(a_{1}=-(4m-1)(m+1)/(6m^{2})\), \(b_{0}=1\), and \(b_{1}=-(3m+1)/(2m)\)), similarly as in (21), we get that
because \({m}{v^{1/m}}+v^{-1} | _{v=1}=m+1\) and
for each \(N\geqslant 1\). Equations (18)–(22) imply that
with a positive quantity \(C_{7m}\) depending only on m.
On the other hand, by hypothesis (17), similarly as in the derivation of (20), we get
with a positive quantity \(C_{8m}\). Using estimates (21) and (22), we obtain
with a positive quantity \(C_{9m}\) depending only on m, which together with (23) implies formula (7) in the case \(n=m+1\). The induction principle implies now the validity of (7) for all \(n\in {\mathbb{N}}\). Theorem 2.2 is proved. □
4 A small simulation study
In this section, we consider two examples of products of normally distributed random variables. In both cases, we will compare the obtained theoretical results with the tail probabilities estimated by the Monte Carlo method.
Example 4.1
Consider the product \(\Pi _{4}\) of four independent random variables \(\xi _{1}\), \(\xi _{2}\), \(\xi _{3}\), \(\xi _{4}\) distributed according to the normal law Φ. According to Theorem 2.2, there exist positive constants \(D_{1}\) and \(D_{2}\) such that for \(x\geqslant D_{2}\),
The simulated values of \(\mathbb{P} (\Pi _{4}>x )\), together with the upper and lower estimates (blue lines), are presented in Fig. 1. Several values of constants \(D_{1}\) and \(D_{2}\) can be selected from the obtained graphs. For instance, Fig. 1 shows that it is possible to take \(D_{1}= 0.8 \) and \(D_{2}= 6\) in the case under consideration.
Example 4.2
Consider the product \(\widehat{\Pi}_{3}\) of three independent random variables \(\widehat{\xi}_{1}\), \(\widehat{\xi}_{2}\), and \(\widehat{\xi}_{3}\) distributed according to the normal laws \(\Phi _{0,1}\), \(\Phi _{0,2}\), and \(\Phi _{0,3}\), respectively. Due to Corollary 2.1, there exist two positive constants \(\widehat{D}_{1}\) and \(\widehat{D}_{2}\) such that for \(x\geqslant \widehat{D}_{2}\),
The simulated values of \(\mathbb{P} (\widehat{\Pi}_{3}>x )\), together with the upper and lower bounds (blue lines), are presented in Fig. 2. In particular, Fig. 2 shows that it is possible to take \(\widehat{D}_{1}=2 \) and \(\widehat{D}_{2}=15 \) in this example.
5 Conclusions
In this paper, we obtained an asymptotic formula with the first leading-order term for the tail of the distribution of the product of independent normal laws. The main results of the work are formulated in Theorem 2.2 and Corollary 2.1. We considered normal random variables with zero means only, but in Corollary 2.1 we allowed the variances of these laws to be unconstrained. From the assertion of Theorem 2.2, it follows that for the product of n independent standard normal distributions \(\Pi _{n}\):
if \(x\geqslant C_{1n}\), where \(C_{1n}\), \(C_{2n}\), and \(C_{3n}\) are positive quantities depending only on n. It follows from the theorem that such three quantities exist. However, their structure remains unclear. Apparently, this requires a deeper analysis of the problem.
Availability of data and materials
Not applicable.
References
Arendarczyk, M., Dȩbicki, K.: Asymptotics of supremum distribution of a Gaussian process over a Weibullian time. Bernoulli 17, 194–210 (2011)
Aroian, L.A.: The probability function of the product of two normally distributed variables. Ann. Math. Stat. 18, 265–271 (1947)
Aroian, L.A., Taneja, V.S., Cornwell, L.W.: Mathematical forms of the distribution of the product of two normal variables. Commun. Stat., Theory Methods 7, 165–172 (1978)
Bandi, M.M., Connaughton, C.: Craig’s xy distribution and the statistics of Lagrangian power in two-dimensional turbulence. Phys. Rev. E 77, 036318 (2008)
Bareikis, G., Šiaulys, J.: In: Products of Independent Random Variables. Vilniaus universiteto leidykla, Vilnius (1998). (in Lithuanian)
Butler, R.W.: Saddlepoint Approximations with Applications. Cambridge University Press, Cambridge (2007)
Craig, C.C.: On the frequency function of xy. Ann. Math. Stat. 7, 1–15 (1936)
Cui, G., Yu, X., Iommelli, S., Kong, L.: Exact distribution for the product of two correlated Gaussian random variables. IEEE Signal Process. Lett. 23, 1662–1666 (2016)
Cui, Z., Wang, Y.: On the long tail property of product convolution. Lith. Math. J. 60, 315–329 (2020)
Fedoryuk, M.: The Saddle Point Method. Nauka, Moscow (1977). (in Russian)
Feller, W.: An Introduction to Probability Theory and Its Applications, vol. 1, 3rd edn. Wiley, Singapore (1968)
Galambos, J., Simonelli, I.: Products of Random Variables: Applications to Problems of Physics and to Arithmetical Functions. Taylor and Francis, Boca Raton (2004)
Gaunt, R.E.: A note on the distribution of the product of zero-mean correlated normal random variables. Stat. Neerl. 73, 176–179 (2019)
Gaunt, R.E.: Stein’s method and the distribution of the product of zero mean correlated normal random variables. Commun. Stat., Theory Methods 50, 280–285 (2021)
Haldane, J.B.S.: Moments of the distributions of powers and products of normal variates. Biometrica 32, 226–242 (1942)
Jensen, J.L.: Saddlepoint Approximations. Oxford University Press, Oxford (1995)
Konstantinides, D., Leipus, R., Šiaulys, J.: A note on product-convolution for generalized subexponential distributions. Nonlinear Anal., Model. Control 27, 1024–1067 (2022)
Liu, Y., Tang, Q.: The subexponential product convolution of two Weibull-type distributions. J. Aust. Math. Soc. 89, 277–288 (2010)
MacKinnon, D.P.: Introduction to Statistical Mediation Analysis. Routledge, New York (2008)
MacKinnon, D.P., Fritz, M.S., Williams, J., Lockwood, C.M.: Distribution of the product confidence limits for the indirect effect: program PRODCLIN. Behav. Res. Methods 39, 384–389 (2007)
Mikutavičius, G., Šiaulys, J.: Product convolution of generalized subexponential distributions. Mathematics 11, 248 (2023)
Nadarajah, S., Pogány, T.K.: On the distribution of the product of correlated normal random variables. C. R. Math. 354, 201–204 (2016)
Simon, M.K.: Probability Distributions Involving Gaussian Random Variables: A Handbook for Engineers, Scientists and Mathematicians, 2nd printing edn. Springer, New York (2006)
Springer, M.D.: The Algebra of Random Variables. Wiley, New York (1979)
Springer, M.D., Thompson, W.E.: The distribution of products of independent random variables. SIAM J. Appl. Math. 14, 511–526 (1966)
Springer, M.D., Thompson, W.E.: The distribution of products of beta, gamma and Gaussian random variables. SIAM J. Appl. Math. 18, 721–737 (1970)
Tang, Q.: From light tails to heavy tails through multiplier. Extremes 11, 379–391 (2008)
Wishart, J., Bartlett, M.S.: The distribution of second order moment statistics in a normal system. Proc. Camb. Philol. Soc. 28, 455–459 (1932)
Wong, R.: Asymptotic Approximations of Integrals. Academic Press, Boston (1989)
Yuan, P.T.: On the logarithmic frequency distribution on the semi-logarithmic frequency surface. Ann. Math. Stat. 4, 46–47 (1933)
Acknowledgements
The authors would like to thank editor in chief and the anonymous referees for their valuable comments.
Funding
The research is supported by a grant No. S-MIP-20-16 from the Research Council of Lithuania.
Author information
Authors and Affiliations
Contributions
R.L. did a major contribution in conceptualization, supervision of the project and funding acquisition. J.Š. dealt with the methodology, investigation and theoretical part of the project. M.D. participated in the design of the manuscript and performed the simulation study. R.Z. performed the validation and writing the initial version of the manuscript. All authors read and approved the final manuscript.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
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
Leipus, R., Šiaulys, J., Dirma, M. et al. On the distribution-tail behaviour of the product of normal random variables. J Inequal Appl 2023, 32 (2023). https://doi.org/10.1186/s13660-023-02941-1
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s13660-023-02941-1
MSC
- 60E05
- 62G32
Keywords
- Distribution tail
- Product of random variables
- Normal distribution
- Saddle point method
- Asymptotic formulas