# What are the gamma-Pareto convolutions and how have they been used?

The Pareto distributions, i.e., density functions (pdf), are types I through IV and the type II variant; the Lomax distribution. This makes for a number of possible gamma-Pareto convolutions (GPC; also density functions). Of these the GPC type I and GPC Lomax have been used in the literature, at least in so far as I know. Let us start by defining the parameterizations used for GPC type I.

The gamma distribution (GD) can be given as

$$\begin{equation}\label{eq:GD} \text{GD}(a,b;t) = \,\dfrac{1}{t}\;\dfrac{e^{-b \, t}(b \, t)^{\,a} }{\Gamma (a)} \;\theta(t)\;, \end{equation}$$

where the gamma function satisfies $$\Gamma (a)=\int _0^{\infty } e^{-t} t^{a-1}\,dt$$, and $$\theta(t)$$ is the unit step function, i.e., 0 for $$t<0$$ and 1 for $$t\geq0$$. The GD rate parameter is $$b$$, whereas $$\frac{1}{b}$$ is the scale parameter. The shape parameter for the GD is $$a$$. The shape parameter aids in fitting data and data shapes. There is no location parameter.

Next, we define a type I Pareto distribution ($$\text{PD}_{\text{I}}$$ a pdf),

$$\begin{equation} \text{PD}_{\text{I}}(\alpha, \beta;t)= \dfrac{\alpha}{t} \left(\dfrac{\beta}{t}\right) ^{\alpha } \;\theta(t-\beta)\;. \end{equation}$$

Note the shift to the right of the unit step function; $$\theta(t-\beta)$$, i.e., the PD function is zero for $$t<\beta$$. $$\beta$$ is the scale parameter. The shape parameter is $$\alpha$$. There is no location parameter.

The PD type II pdf is

$$\text{PD}_{\text{II}}(\alpha ,\beta ,\mu;t)=\frac{\beta }{\alpha }\left(\frac{\alpha -\mu +t}{\alpha }\right)^{-\beta -1}\theta (t-\mu ).$$

The PD type II Lomax pdf is

$$\text{PD}_{\text{Lomax}}(\alpha ,\beta ,0;t)=\frac{\beta }{\alpha }\left(\frac{\alpha +t}{\alpha }\right)^{-\beta -1}\theta (t ) .$$

The PD type III pdf is

$$\text{PD}_{\text{III}}(\alpha ,1,\gamma ,\mu;t)=\frac{\alpha ^{-1/\gamma } (t-\mu )^{\frac{1}{\gamma }-1}}{\gamma \left[\left(\frac{\alpha }{t-\mu }\right)^{-1/\gamma }+1\right]^2}\theta (t-\mu ) .$$

The PD type IV pdf is

$$\text{PD}_{\text{IV}}(\alpha ,\beta,\gamma ,\mu;t)=\frac{\beta}{\gamma } \alpha ^{-1/\gamma } (t-\mu )^{\frac{1}{\gamma }-1} \left[\left(\frac{\alpha }{t-\mu }\right)^{-1/\gamma }+1\right]^{-\beta -1}\theta (t-\mu ) .$$

The infinite series describing the GPC type I convolution is given in the answer below. There is no closed form solution. Such functions are computer implemented as truncated infinite series.

I will award a bounty to anyone who shows a solution$$^\star$$ for a GPC distribution type whose solution is not referenced here, i.e., GPC type II (non-Lomax), GPC type III or GPC type IV.

$$^\star$$EDIT: Any exact and calculable reformatting of any convolution integral of the form: \begin{align}\text{GPC}_j\left(\begin{array}{cc} a & b \\ \alpha &\beta \end{array} \Big|\,t\right)&=\text{GD}(a,b;x)\ast\text{PD}_j(\alpha,\beta;x)\,(t)\\&=\int_{-\infty}^{\infty} \text{GD}(a,b;x-t)\,\text{PD}_{j}(\alpha,\beta,\,,\,;t) \;dt,\end{align} where $$j=\text{II, III, or IV}$$. I do not consider limiting discrete convolutions for $$\Delta t\to0$$, i.e., a limiting or Riemann sum of the convolution integral, to be exact solutions unless the limits sign can be removed in such a fashion as to provide a calculation of any desired accuracy.

• Could you explain what you mean by "documents those ... types"? What would constitute a "documentation" (listing of mathematical propertiess? potential applications? references?) and what specifically are the "types not listed here"? – whuber Jan 18 at 15:46
• @whuber I can, and did. Surprised I am that it was not clear from the context. – Carl Jan 18 at 16:31
• I cannot speak to the downvote, Carl, because I did not apply it, so please do not insinuate I did. (I would be happy to prove I did not by applying one :-) In general, please do not complain about downvotes: it's not constructive and can only antagonize people. Returning to my previous comment: please explain what you mean by "documenting a solution to" one of these distributions. – whuber Jan 18 at 19:06
• @whuber I knew you did not do the downvote, I know you fairly well. I did not insunuate anything but wrote to you as a moderator saying that I do not like downvoting on well-formatted questions or even on poorly formatted ones, I do not think they serve any purpose that is not better accomplished else-wise. I removed the comment so no-one else can misinterpret it either. I do not understand your comment as to "what do I mean?" I am asking for convolutions written as formulas not containing a convolution operator. What is unclear to you? – Carl Jan 18 at 19:55
• For starters, I have yet to see any convolution operators in this thread! I truly do not understand what you are going after in this question. Are you trying to obtain formulas for densities (or CDFs or CFs or CGs or something) of sums of Gamma and Pareto variables? And if so, what kind of formulas are wanted? I notice the one in your answer is an infinite sum, which is less than convenient and in many ways less useful than the original convolution integral. – whuber Jan 18 at 20:26

The GPC type I has been open access published recently and used to model metformin drug serum concentrations following intravenous bolus injection. Derived from series expansion of the GD exponential above, followed by convolution with a PD, as above and summation of the expanded parts. The GPC type I model inherits two shape parameters, $$a$$ and $$\alpha$$, and two scale related parameters, $$b$$ and $$\beta$$ from its convolved pdf's,
$$\begin{equation} \begin{split} \text{GPC}_{\text{I}}\left(\begin{array}{cc} a & b \\ \alpha &\beta \end{array} \Big|\,t\right) &=\text{GD}( a,b;x)\ast \text{PD}_{\text{I}}(\alpha , \beta;x) \;(t)\\ &=\theta (t-\beta) \frac{\alpha b^a\beta^\alpha}{\Gamma (a)}t^{a-\alpha-1}\sum _{n=0}^{\infty } \frac{(-b\,t)^n}{n!} B_{1-\frac{\beta}{t}}(a+n,-\alpha) \end{split}\;\;, \end{equation}$$
where the incomplete beta function is $$B_z(A,B)=\int _0^z u^{A-1} (1-u)^{B-1}d u$$. Note that this is exact but is calculable by removing the limiting $$\infty$$ for $$n$$ as follows. Because the series is alternating, for $$n>m$$, where $$m$$ is some positive integer, no matter how large, the series sum thereafter has decreasing absolute error, as it then satisfies the Leibniz test, where the absolute error is less than the absolute magnitude of the $$n>m^{\text{th}}$$ term.
Proof: First note that $$\lim_{t\to\infty}B_{1-\frac{\beta}{t}}(a+n,-\alpha)=B(a+n,-\alpha)$$, where $$B(\cdot,\cdot)$$ is the beta function. Next, $$\lim_{n\to\infty}B(a+n,-\alpha)=0$$. Then $$\lim_{n\to\infty}t^{a-\alpha-1} \frac{(-b\,t)^n}{n!}=0$$, no matter how large $$t$$. Finally, $$0\cdot0=0$$, thus the $$n^{\text{th}}$$ term absolute value can be made smaller than $$\epsilon$$, no matter how small for $$n$$ sufficiently large.||