- Research
- Open access
- Published:
Inequalities involving hypergeometric and related functions
Journal of Inequalities and Applications volume 2018, Article number: 253 (2018)
Abstract
An inequality is being proved which is connected to cost-effective numerical density estimation of the hyper-gamma probability distribution. The left-hand side of the inequality is a combination of two in the third parameter distinct versions of the hypergeometric function at the point one. All three parameters are functions of the distribution’s terminal shape. The first and second are equal. The distinct third parameters of the two hypergeometric functions depend on terminal and initial shape. The other side of the inequality is determined by the quotient of two infinite series, which are related to the first derivatives with respect to terminal shape of the hypergeometric functions which appear in its left-hand side.
1 Introduction
Certain inequalities shall be considered, which involve combinations of gamma and psi functions of one positive variable β and one parameter x greater than unity. The basic functions involved are particular values of the hypergeometric function \(F(a,b,c;s)\) [1, 9.122.1], namely:
where \(\Gamma (s)\) is the gamma function, \(y_{\nu }= \nu \beta + x\), \(\nu = 0,1,2,\dots \), with \(\beta \in (0, \infty)\), \(x \in (1, \infty)\) throughout. The product representations of \(A(\beta; x)\) and \(B(\beta; x)\) are given is [1, 8.325.1]. Associated with these functions is their product
Since \(a+b-c < 0\) for each of the hypergeometric functions in (1.1), (1.2), and (1.3), each of their series expansions converges throughout the entire closed unit circle of the complex plane [1, 9.102.2], and the gamma function expressions are justified [1, 9.122.1].
For the sake of simplicity we shall from now on omit the arguments β and x whenever there is no chance for confusion, keeping in mind, however, that β in the actual independent variable, and that x is a parameter. If we attach an argument to a symbol of a dependent variable, it will be a particular value of β. For example, \(A(1) = A(1;x)\). Furthermore, derivatives will always be with respect to β and will be denoted by a prime.
The functions A, B, and C satisfy the inequalities
The infinite product representations of A, B, and C show that
and
as \(\beta \uparrow \infty \). (The down-arrow means that the function at its left decreases toward the value at its right as its argument decreases. The up-arrow indicates the opposite.) We note the particular values
The β-derivatives of A, B, and C are
with
\(\psi_{\nu }= \psi ((x+\nu -2)/\beta)\), \(\nu = 1, \dots,4, \psi (s)\) being the ψ function, \(\psi (s) = d\log \Gamma (s)/ds\) [1, 8.360].
If we consider instead of (1.1), (1.2), and (1.3) the more general function
with differentiable functions \(\varphi_{\nu }= \varphi_{\nu }(\beta)\), \(\nu = 1,2,3\), then
provided that Re\((\varphi_{1} + \varphi_{2} - \varphi_{3}) < 0\) in some region R of the complex β-plane. For \(\varphi_{1} = 1/\beta\), \(\varphi_{2} = -1/\beta\), \(\varphi_{3} = x/\beta \) this formula reduces to \(A' = 2AS\) with S given in (1.8).
Related to S and T is the function
The series expansion of the psi function [1, 8.362, 1],
where γ is Euler’s constant, leads to the series representations of T, Q, and S,
(Cancellation of the respective factors \(y_{\nu }+ \alpha\) (\(\alpha = -1, 0, 1, 2\)) in (1.11) would not bring any advantage. In fact, it would hamper comparison of equally numbered terms of these series and related expressions.)
Under the restrictions on β and x, each of the series in (1.11) is positive, and
Therefore, the derivatives \(A'\), \(B'\), and \(C'\) in (1.7) are positive, i.e., the functions A, B, and C are strictly monotonically increasing for \(\beta > 0\) and, by (1.5a), (1.5b), bounded. (This belatedly justifies the direction of the arrows in the limit relations (1.5a), (1.5b).)
We now introduce the gamma function combination
with A and B defined in (1.1) and (1.2), respectively. The main objective of this paper is to establish the inequality
It is a crucial prerequisite to an efficient numerical solution routine of the four-parameter hyper-gamma density estimation problem for a given statistical data set (observations) [2, Chap. 9.3]. The hyper-gamma distribution has important applications in chemical, biological, and physical processes. The four parameters of the distribution are shift, scale, initial shape \(x>1\) (in the statistical four-parameter case \(x > 2\)), and terminal shape \(\beta > 0\). The statistical parameters β and x are to be determined from a set of two simultaneous equations [2, (9.33), (9.34)]. These are derived from the first four moments of the four-parameter hyper-gamma probability density function [2, (9.3.1)], which are defined by combinations of gamma functions as they appear in (1.1) and (1.2). A considerable computational cost advantage is achieved if it is known that for every fixed value of x, say, one of these equations has exactly one solution β. This will be the case if (1.15) holds. (See [2, Chaps. 9.2, 9.3], equation (9.3.34) and the discussion preceding (9.3.36).)
2 Approximating sequences for \(Q/S\) and \(T/S\)
The series (1.11) and their β-derivatives converge (absolutely and) uniformly as functions of β on every closed subinterval \([a, b] \subset (0, \infty)\) for any fixed \(x > 1\). To show this, we look at Q, for example,
(Here the common factor \(y_{\nu }\) in numerator and denominator has been canceled.) We have \(0 < q_{\nu }< 1 /\nu^{2} \beta^{3} < 1 / \nu^{2} a ^{3}\). Thus, by the Weierstraßcriterion,
where \(\zeta (s)\) is Riemann’s zeta function [1, 9.522.1]. The terms of the β-derivative \(Q'\) of Q are
with
Now, \(0 < \gamma_{\mu \nu } < 1 / \nu^{4} \beta^{4}\), \(\mu = 1, 2, 3\), \(\nu = 1, 2, \dots\), and, hence, \(|q'_{\nu }| < 3 / a^{4} \nu^{2}\). Therefore, \(|Q'| < 3 \zeta (2) / a^{4}\), i.e., the series \(Q' = \sum q'_{\nu }\) is absolutely and uniformly convergent on \([a, b]\). Since \(Q' < 0\), Q is strictly monotonically decreasing as β increases. The second derivative of Q is positive. This follows immediately from (2.1). Thus, with \(Q'' > 0\), the function Q is concave from above, and \(|Q'|\) is strictly monotonically decreasing. Corresponding facts are true for T and S. Note that T, Q, and \(S \uparrow \infty \) (monotonically) as \(\beta \downarrow 0\), and that T, Q, and \(S \downarrow 0\) (monotonically) as \(\beta \uparrow \infty \).
The following particular values of these functions for \(\beta = 1\) are of interest:
They are obtained from (1.9), (1.10a), (1.10b), and (1.8), respectively, and from the functional relation [1, 8.365.3] of the psi function.
We now define the sequence \(\{ F_{n} (\beta; x)\}\) with elements
where \(Q_{n}\) and \(S_{n}\) are the partial sums of Q and S,
with
We observe that \(r_{n}\) and \(p_{n}\) (both polynomials in β of degree \(4(n-1)\)) can be expressed as
and that the positive series
converge uniformly (for any \(x \in (1, \infty)\)) on every interval \([a, b] \subset (0, \infty)\). This follows immediately by the Weierstraßcriterion since
so that
Furthermore,
so that
and, consequently,
The series
converge absolutely and uniformly on every interval \([a, b] \subset (0, \infty)\). Note that the series (2.7) are strictly monotonically decreasing functions of \(\beta \in (0, \infty)\).
We return to the functions \(F_{n}\) defined by (2.3) and show that
i.e., that
By (1.11)
\(u_{\nu }\) given in (2.8), so that (2.12) may be replaced by
Here we replace \(S_{n} - Q_{n}\) by \(2 \sum_{\nu =1}^{n} u_{ \nu }\) and obtain, after dropping the common factor 2,
Comparing equally numbered terms of the two infinite series, we get
or, with \(q_{n+k} = (n+k) y_{n+k} / \rho_{n+k}\) and \(u_{n+k} = (n+k) / \rho_{n+k}\),
This difference is positive since \(y_{n+k} > y_{\nu }\) for \(\nu = 1, \dots, n\). Therefore, (2.13) is correct, and, consequently, (2.11) holds. With this result, and
we see that
(the upward arrow will be justified in Sect. 3), and that
Next, we show that
By means of (2.4a), (2.4b), the rational functions \(F_{n}\) defined in (2.3) can be written as
With this expression for \(F_{n}\), inequality (2.16) changes into \(r_{n} p_{n+1} > r_{n+1} p_{n}\). The recurrence relations for \(r_{n}\) and \(p_{n}\) in (2.5) lead to \((n+1) r_{n} > p_{n}\), which is correct. Thus (2.16) holds.
The facts established so far show that \(\{ F_{n} \}\) is a positive, increasing, bounded above sequence that converges for every \(\beta \in (0, \infty)\) and any fixed \(x \in (1, \infty)\) as \(n \uparrow \infty \). We want to show now that it converges to \(Q/S\) uniformly on every subinterval \([a, b]\) of \((0, \infty)\).
Observing (2.12), we have to show that for every \(\varepsilon > 0\) there exists \(n_{0} = n_{0} (\varepsilon)\) such that
Now,
Since \(q_{\nu }< s_{\nu }= \nu (y_{\nu }+ 2) / \rho_{\nu }\), we have
Here
The right-hand side of this inequality is strictly monotonically decreasing as stated earlier in connection with (2.8) and (2.10). Thus
Therefore, we may continue inequality (2.19) as follows:
Uniform convergence of S on \([a, b]\) implies that, given \(\varepsilon /K\), there exists \(n_{0} = n_{0}(\varepsilon)\) such that
This proves (2.18). So \(\{ F_{n} \}\) is a uniformly convergent approximating sequence for \(Q/S\) from below, and
This shows that \((Q/S) (0) = F_{1}(0) = x/(x+2)\), so that (2.15) may be sharpened to
(The downward arrow will be justified in Sect. 3.)
Along the same lines analogous results can be established for the function \(T/S\) and its approximating from below sequence
Note that
and
3 Monotonicity of \(Q/S\) and \(T/S\)
The functions \(Q/S\) and \(T/S\) are strictly monotonically increasing for \(\beta \in (0, \infty)\). It is sufficient to show this for \(Q/S\) since, by (1.10a), \(T/S = (3Q/S-1)/2\).
We introduce the rational functions
\(r_{n}\) and \(p_{n}\) defined by (2.6). Note that \(f_{n+1} < f_{n}\) because \(\{ u_{\nu }/ v_{\nu }\} = \{ 1 / \nu \}\) is a decreasing sequence [3, p. 10, Problem 28]. The approximating functions \(F_{n}\) in (2.17) can now be written as
The β-derivative of \(F_{n}\) is
and
This immediately shows that \(F'_{n} > 0\) for small positive values of β. We want to show that \(F'_{n} > 0\) for \(\beta \in (0, \infty)\). But first we must establish certain facts about the rational functions \(f_{n} = r_{n} / p_{n}\). By (2.5) and with \(\rho_{\nu }= ( \nu \beta + x - 1)(\nu \beta + x)(\nu \beta + x + 1) (\nu \beta + x +2)\) the constituent terms of \(r_{n}\) and \(p_{n}\) are products of \(n-1\) Hurwitz polynomials, each of degree 4, so that each of those terms is a Huwitz polynomial of degree \(4(n-1)\). The β-derivative of \(f_{n}\) is
Using (2.8) and (2.9) in (3.4b), we arrive at
Here
the positive functions \(\alpha_{\nu }\) (\(\nu = 1,\dots, 4\)) being defined in connection with (1.12). Thus, by (3.5), \(f'_{n} > 0\), \(\beta > 0\), \(x>1\). This establishes strict monotonicity of the functions \(f_{n}\) defined in (3.1). Note that \(0 < f_{n} < 1\), \(n \geq 2\), \(\beta > 0\) and
We also note the following facts about \(f'_{n}\), which follow from (3.4a),
Furthermore, (2.5) together with (1.12) and with the notation used in (3.6) shows that
so that
Consequently,
where \(c_{n}\) is a positive constant. Since
we see that \(f'_{n}\) as given in (3.4a) behaves like \(\beta^{-2}\) as \(\beta \uparrow \infty \). In other words, \(f'_{n}\) and \(\beta^{-2}\) are asymptotically proportional. Consequently,
We now remember the fact that the polynomials \(\rho_{\nu }\) are Hurwitzian and that, consequently, the constituent terms of the polynomials \(r_{n}\) and \(p_{n}\) in (2.5) are Hurwitzian. By Theorem IV of [4] (in conjunction with the specification of terminology concerning circular regions and circles on p. 164 of [4]) the sum of any two of these constituent polynomials of degree \(4(n-1)\) is Hurwitzian. Thus, \(r_{n}\) and \(p_{n}\) are Hurwitz polynomials. Their zeros are located in the open left-hand half of the complex β-plane, which we denote by L. By another theorem [5, p. 115], all zeros of \(f'_{n}\) (and all its poles) are located in L. In other words, \(f'_{n} > 0\) for real \(\beta > 0\), a fact which has been established earlier already by direct means. Applying the theorem of [5] again, this time to the rational function \(f'_{n}\), we arrive at the result that \(f''_{n}\) has all its zeros (and poles) in L, i.e., \(f''_{n} \neq 0\) for real \(\beta > 0\). Limit relation (3.7) shows that \(f'_{n}\) decreases somewhere in the interval \((0, \infty)\). Thus, the β-derivative \(f''_{n}\) of \(f'_{n}\) must be negative somewhere. Since \(f''_{n} \neq 0\) for \(\beta > 0\), if follows that \(f''_{n} < 0\) for all \(\beta > 0\). In other words, \(f_{n}\), defined by (3.1) is concave from below on \(0 < \beta < \infty \), i.e., the tangent at any point \((\beta_{0}, f_{n}(\beta_{0}))\), \(\beta_{0} > 0\), lies above the graph of \(f_{n}\) for every \(\beta > 0\), \(\beta \neq \beta_{0}\).
Since, by (3.2), \(F'_{n} (0) > 0\), if follows that \(F' > 0\) at least for small positive values of β. This means that \(f_{n} - \beta f'_{n} > 0\) for small positive values of β as can be seen from (3.3). This can also be verified by means of (3.6), (3.7), and (3.8), respectively. By (3.6) and (3.8) we see that \(f_{n} - \beta f'_{n} > 0\) also for large values of β, so that \(F'_{n} > 0\) for large β. Suppose now that \(f_{n} - \beta f'_{n} < 0\) at some point \(\beta > 0\). Then there exist points \(\beta_{1}\) and \(\beta_{2}\), \(0 < \beta_{1} < \beta_{2}\), such that
\(f_{n} - \beta f'_{n} > 0\) for \(0 < \beta < \beta_{1}\) and \(\beta_{2} < \beta < \infty \), and \(f_{n} - \beta f'_{n} < 0\) for \(\beta_{1} < \beta < \beta_{2}\). Consequently, there would exist a continuous function \(\alpha (\beta)\) such that \(\alpha (\beta_{\nu }) = 0\), \(\nu = 1, 2\), \(\alpha (\beta) > 0\), \(\beta_{1} < \beta < \beta_{2}\), and
This inhomogeneous linear differential equation has the unique solution
with initial condition \((\beta_{1}, f_{n} (\beta_{1}))\). Its derivative is
Observing (3.9) for \(\nu =2\) and noting that \(\alpha (\beta_{2}) =0\), we have at \(\beta_{2}\)
since the integral is positive. Now, the tangent to the integral curve defined by (3.10) at the point \((\beta_{1}, f_{n} (\beta_{1}))\) is given by \(y(\beta) = f_{n} (\beta_{1}) + (\beta - \beta_{1}) f'_{n} (\beta _{1})\) with \(f'_{n} (\beta_{1}) = f_{n}(\beta_{1}) / \beta_{1}\) by (3.9).
Thus,
Since \(f_{\nu }\) is concave from below it follows that
or
in contradiction to (3.11). Consequently, \(f_{n} - \beta f'_{n} > 0\) for all \(\beta \in (0, \infty)\), and, hence, as (2.3) shows, \(F'_{n} > 0\) for \(0 < \beta < \infty \), which means that \(F_{n} = Q_{n} / S_{n}\), or \(F_{n} = (xf_{n} + \beta) / [ (x+2)f_{n} + \beta ]^{-1}\), is a strictly monotonically increasing function of \(\beta \in (0, \infty)\). A corresponding result holds for \(G_{n} = T_{n} / S_{n}\).
We now show that the sequence \(\{ F'_{n} \}\) converges uniformly on every closed interval \([a, b] \subset (0, \infty)\). Differentiating \(F_{n} = Q_{n} / S_{n}\), we form
With \(Q = \sum q_{\nu }\), \(S = \sum s_{\nu }\), equation (3.12), after some manipulations, can be brought into the form
Here, we replace all negative terms between the absolute value bars by their absolute values. Since Q, \(|Q'|\), S, \(|S'|\) are series with positive terms, we increase the right-hand side of (3.13) by replacing their partial sums by the entire series. Then we remember the fact that these series are monotonically decreasing functions of β. Therefore, we increase the right-hand side of (3.13) further by setting \(Q_{n} < Q < Q(a)\), \(|Q'_{n}| < |Q'| < |Q'(a)|\), \(S_{n} < S(a)\), \(|S'_{n}| < |S'(a)|\) for \(\beta \in [a,b]\). We may also replace \(\sum q_{\nu }\) by \(\sum s_{\nu }\) since \(0 < q_{\nu }< s_{\nu }\). Thus, there exist positive constants \(K_{\nu }\) (\(\nu = 1,\dots,5\)) such that equality (3.13) may be replaced by the inequality
Furthermore, by (1.11),
so that
Now, since S, \(S'\), and \(Q'\) converge uniformly on \([a, b] \in (0, \infty)\), given \(\varepsilon > 0\) there exists a number \(n_{0}( \varepsilon)\) such that each of the five finite sums in (3.14) is less than \(B^{-4} (\sum_{\nu =1}^{5} K_{\nu })^{-1} \varepsilon \) for every \(n \geq n_{0}\) and for all \(k \geq 1\). Thus (3.14) together with (3.15) leads to \(|F'_{n+k} - F'_{n}| < \varepsilon \) for every \(n \geq n_{0}\) and for all \(k \geq 1\) on every \([a, b] \subset (0, \infty)\). The final result is that \(\{ F'_{n} \}\) converges uniformly for \(\beta \in [a,b]\) to the function \((Q/S)' > 0\), and this means that \(Q/S\) is strictly monotonically increasing. (This result justified the direction of the arrows in (2.14) and (2.20), and in (2.21) and (2.22) for \(T/S\).)
4 Properties of σ
The function
together with (1.1) and (1.2), evidently satisfies the inequality \(0 < \sigma < A\), \(\beta > 0\) and by (1.5a), (1.5b), the limit relations
(The downward arrows in (4.2) will be justified momentarily.) Furthermore, observing (1.6), we see that at \(\beta = 1\)
The β-derivative of σ is given by
This shows that \(\sigma < Q/S\) whenever \(\sigma '\) is nonpositive. For, if \(\sigma ' \le 0\), then \(\sigma \le CT/S < T/S < Q/S\), if we observe (1.3) and (1.4). The function \(\sigma '\) is negative somewhere. To see this, we note that for \(\beta = 1\), (4.4) becomes
Here (1.6) and (2.2) have been used. Continuity of σ implies that \(\sigma (1 + \varepsilon) > \sigma (1) = \sigma_{\infty }\) at least for sufficiently small values of \(\varepsilon > 0\). Inequality (4.5) and the second limit relation (4.2) show that σ has a local maximum at some point \(\beta^{\star }> 1\). Suppose σ had more than one maximum in \((0, \infty)\). Let \(0 < \beta_{1} < \beta _{2}\) be points at which maxima are attained. Then σ must have a minimum at some point \(\beta_{0}\), \(\beta_{1} < \beta_{0} < \beta _{2}\). We have
and \(\sigma ' (\beta_{1}) = \sigma ' (\beta_{0}) = 0\), so that, according to (4.4)
Since C and \(T/S\) are both monotonically increasing, it would follow that \(\sigma (\beta_{1}) < \sigma (\beta_{0})\), in contradiction to (4.6). Consequently, there exists one and only one point \(\beta^{ \star }\) at which \(\sigma ' = 0\) and at which σ takes its maximum. We have \(\sigma ' \geq 0\) if \(\beta \in (0, \beta^{\star }]\), \(\sigma ' < 0\) if \(\beta \in (\beta^{\star }, \infty)\). In other words, σ is strictly monotonically decreasing on \((\beta^{\star }, \infty)\). (This result belatedly justifies the direction of the arrows in the limit relations (4.2).)
We now turn to the relation between σ given by (4.1) and \(C = AB\) defined by (1.3). By (4.3), \(\sigma (1) = C(1)\). Furthermore, by (4.2) and (1.5b) \(\sigma_{\infty }< C_{\infty }\). Thus, σ and C are not identical, and \(\sigma < C\) for large values of β since σ is strictly monotonically decreasing after it reaches its single maximum at \(\beta^{\star }> 1\), and C is strictly monotonically increasing. More can be obtained by comparing the derivatives of σ and C at \(\beta = 1\). Using the particular value of C at \(\beta = 1\) given in (1.6) and those of S and T for \(\beta = 1\) given in (2.2) and the definition of \(C'\) given in (1.7), we find that
Comparing this with \(\sigma ' (1)\) in (4.5), we see that
This and \(\sigma (1) = C(1)\) imply that for sufficiently small \(\varepsilon > 0\),
and
We show first that (4.7) holds for \(\beta \in (1, \infty)\). Observing (4.1) and setting \(C = AB\), we transform the desired inequality (4.7) into
Here, we set \(A = \kappa B\), \(0 < \kappa = \kappa (\beta; x) < 1\), \(\beta \in [1, \infty)\), \(x \in (1, \infty)\), noting that κ has a positive β-derivative (by (1.7) and (1.13)),
so that κ is strictly monotonically increasing, and by (1.6) and (1.5b),
Furthermore, since by (1.1) and (1.2), the factors of the infinite product for \(\kappa = A/B\) go to \((x+1)^{3} (x-1) / x^{3} (x+2) < 1\) as \(\beta \downarrow 0\), κ diverges toward 0 as \(\beta \downarrow 0\).
With \(A = \kappa B\), inequality (4.9) takes the form
The equation \(g(B) = 0\) has the two roots,
\(0 < B_{1} < 1 < B_{2}\). Since \(B < B_{\infty }= x(x+2) / (x+1)^{2} < B_{2}\) by (1.5b), the root \(B_{2}\) is outside the range of B. At \(B_{\infty }\), g takes the negative value \(g(B_{\infty }) = -2 [x(x+1)]^{-1}\). Consequently, as stipulated by (4.10), \(g(B) < 0\) for \(B_{1} < B < B_{\infty }\), i.e., for \(1 < \beta < \infty \), and \(g(B) = 0\) if and only if \(\beta = 1\), i.e., for \(B(1) = B_{1} ( \kappa (1))\). Thus (4.7) holds for \(\beta \in (1, \infty)\).
For \(\beta \in (0,1)\) we have \(0 < B < B(1) = B_{1} (\kappa (1))\), and the function \(g(B)\) in (4.10) is positive, i.e., the inequality sign in (4.10) is reversed. This proves (4.8) for the entire interval \(0 < \beta < 1\).
Our results are these:
and
Finally, for \(\beta = 1\), \(g(B(1)) = 0\), i.e.,
5 The inequality \(\sigma < Q/S\)
We turn now to our main objective and prove the inequality \(\sigma < Q/S\), \(\beta \in (0, \infty)\), fixed \(x \in (1, \infty)\). We know that \(\sigma < Q/S\) holds for small positive values of β since, by (4.2), \(\sigma \downarrow 0\) as \(\beta \downarrow 0\), and \(Q/S > x/(x+2)\) as \(\beta \downarrow 0\) by (2.15). Furthermore, \(\sigma < Q/S\) as \(\beta \uparrow \infty \) since σ decreases monotonically toward \(\sigma_{\infty }= (x-1)/(x+1) < 1\) as \(\beta \uparrow \infty \) by (4.2), whereas \(Q/S \uparrow 1\) by (2.14). (The fact that \(\sigma_{\infty }< 1\) also follows directly from \(\sigma < A < A_{ \infty }< 1\) for \(\beta \in (0, \infty)\).) Therefore, if the desired inequality \(\sigma < Q/S\) should not hold throughout, it must be violated somewhere in the interval \((0, \infty)\).
Observing definition (4.1) of σ and setting \(Q/S = z\) for simplicity, the proposed inequality \(\sigma < z\) can be transformed into
Here we set \(B = \lambda A\), \(\lambda = \lambda (\beta; x) > 1\), \(\beta \in (0, \infty)\), \(x \in (1, \infty)\). λ has a negative β-derivative,
so that λ is strictly monotonically decreasing. By (1.6) and (1.5b),
and, by (1.1) and (1.2), the factors of the infinite product \(\lambda = B/A =\kappa^{-1}\) go to \(x^{3} (x+2) / (x+1)^{3} (x-1) > 1\) as \(\beta \downarrow 0\), i.e., λ diverges ↑∞, as \(\beta \downarrow 0\).
With \(B = \lambda A\), inequality (5.1) takes the form
Suppose \(h(A)\) as function of β were negative somewhere in \((0, \infty)\). Then h would have to be zero somewhere in that interval. The roots of the equation \(h = 0\) are
Here \((1+z)/2\lambda < 1\) because \(1 + z < 2\) and \(\lambda > 1\), and the radicant in (5.3)
is negative as \(\beta \downarrow 0\) since \(\lambda \uparrow \infty \), and, as \(\beta \uparrow \infty \), by (1.5b) and with \(z \uparrow 1\),
Thus, for small and large values of β, \(A_{1,2}\) would be nonreal.
Suppose j, given in (5.4), were positive somewhere in \((0, \infty)\). Then it must be zero somewhere in that interval. The equation \(j=0\) has the two roots
The second one is greater than unity since \(\lambda > 1\). It is outside the range of \(z = Q/S < 1\). Then, using the root \(z_{1}\) for z in (5.3) and squaring the equality to get rid of the square root, we arrive at the equality \(\lambda A^{2} - 2\lambda A + 1 = 0\), or, if we replace λA by B and AB by C, at the equality
which we have encountered before in (4.11). We know that it holds if and only if \(\beta = 1\). But for \(\beta = 1\),
This contradiction shows that \(h(A)\) cannot be negative or zero on the interval \(0 < \beta < \infty \). Thus, (5.2) holds, and \(\sigma < Q/S\) for \(\beta \in (0, \infty)\), fixed \(x \in (1, \infty)\).
6 Declarations
6.1 Results and discussions
Inequalities have been proved which involve various combinations of psi- and hypergeometric functions. They add to the wealth of knowledge in the theory of these special function classes of higher analysis.
6.2 Conclusions
The main inequality of this paper guarantees uniqueness of the hyper-gamma parameter estimation and its application. Usefulness of this approach has been demonstrated in [6].
6.3 Methods/experimental
The aim of the study is to prove an inequality made up of functions of higher mathematical analysis. This inequality guarantees monotonicity of the first moment equation function of the four-parameter hyper-gamma probability density estimation problem. Monotonicity guarantees uniqueness of the numerical solution process. Standard analytical methods of higher analysis have been employed to accomplish the proof.
References
Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series, and Products. Academic Press, New York (1965)
Lehnigk, S.H.: The Generalized Feller Equation and Related Topics. Longman, Harlow (1993)
Knopp, K.: Theorie und Anwendungen der Unendlichen Reihen. Springer, Berlin (1964)
Walsh, J.L.: On the location of the roots of certain types of polynomials. Trans. Am. Math. Soc. 12, 163–180 (1922)
Walsh, J.L.: On the location of the roots of the Jacobian of two binary forms, and of the derivative of a rational function. Trans. Am. Math. Soc. 22, 101–116 (1921)
Dudel, H.P., Hall, C.E., Lehnigk, S.H.: Estimation algorithms for the hyper-gamma distribution class. Commun. Stat., Simul. Comput. 18(3), 1135–1153 (1989)
Acknowledgements
Not applicable.
Availability of data and materials
Not applicable.
Funding
Not applicable.
Author information
Authors and Affiliations
Contributions
Not applicable. Author read and approved the final manuscript.
Corresponding author
Ethics declarations
Competing interests
The author declares that he has no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Lehnigk, S.H. Inequalities involving hypergeometric and related functions. J Inequal Appl 2018, 253 (2018). https://doi.org/10.1186/s13660-018-1842-4
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s13660-018-1842-4