6
$\begingroup$

In my research I have run across the hypergeometric function $${}_3F_2(d,d,d;d+1,d+1;z)$$ where d is a positive integer and 0≤z≤1. When I plot this as a function of d on a semilog plot, it appears to be exponential (or close to it) for large d, but I haven't been able to prove it or find a formula for the exponential as a function of z. I've poked around on the web, but I'm a physicist rather than a mathematician and I'm a bit overwhelmed by the hypergeometric literature. Can anyone help? Thanks.

$\endgroup$
  • 1
    $\begingroup$ Looks like this is a polynomial with unimodal terms, so you can probably do this by bare hands by locating the maximal term(s) and approximating the sum by a Gaussian integral. $\endgroup$ – Noam D. Elkies Jul 22 '16 at 16:29
5
$\begingroup$

Indeed the hypergeometric series is unimodal (the terms for fixed $z$ and $d$ first grow with $k$ then decrease again). The hypergeometric series is given by $$ {}_{3}F_{2}(d,d,d;d+1,d+1;z)=\sum_{k=0}^{\infty} \frac{(d)_{k}(d)_{k}(d)_{k}}{(d+1)_{k}(d+1)_{k} k!} z^{k}=\sum_{k=0}^{\infty} b(k,d) z^{k}, $$ with the Pochhammer symbol $(a)_{k}=\frac{\Gamma(a+k)}{\Gamma(a)}$.

Therefore the ratio of two consecutive terms gives the nicely compact $$ \frac{b(k+1,d) z}{b(k,d)}=\frac{(d+k)^{3}}{(d+k+1)^{2}(k+1)} z $$ which should be equal to 1 for $k=k_{\max}$. Expanding the ratio for large $d$ we find for the highest order in $d$ $$ k_{\max}=\frac{z}{1-z} \ d + O(1). $$ This expansion is valid for $d\ z$ and $d \ (1-z)$ well larger than 1, as can be seen by the following terms in the expansion. Now we do as Noam Elkies suggested in his comment: make the index $k$ continuous and expand the terms of the hypergeometric sum to second order around $k_{\max}$. Actually, it is the logarithm of the terms that gets expanded to afterwards have a nice Gaussian integral, when we switch from summation over $k$ to integral. We get $$ \sum_{k=0}^{\infty} b(k,d) z^{k}\approx \int_{-\infty}^{\infty} \exp\{\ln(b(k_{\max} + \tau,d) z^{k_{\max}+\tau})\}\ d\tau. $$ The exponent is expanded to second order in $\tau$ giving a Gaussian integral that can be calculated easily. (The factor in front of the $\tau^2$ term is indeed negative, the integral is converging.) The extension of the lower integration limit to $-\infty$ is okay, since at this stage the integrand is rapidly vanishing in that region as function of $\tau$.

All this can be done nicely with your favourite computer algebra system. After tweaking the result a bit and expanding it for large $d$, the final result is $$ {}_{3}F_{2}(d,d,d;d+1,d+1;z)\approx\exp\left\{\frac{(5 z +1)^2}{8 \ d \ z}\right\}\ (1-z)^{-d+2}. $$ This approximation becomes bad if $d \ z \approx 1$ and smaller (i.e., for very small $z$). But in this regime the even simpler $$ {}_{3}F_{2}(d,d,d;d+1,d+1;z)\approx (1-z)^{-d+2} $$ can be used. For $d\ (z-1)$ smaller than one we get also invalid results.

As a byproduct I mention below the asymptotic expansion for hypergeometric functions with higher indices $$ {}_{p+1}F_{p}(\overbrace{d,\dots,d}^{p+1};\underbrace{d+1,\dots,d+1}_{p};z)\approx\exp\left\{\frac{(z (2 p +1) +1)^2}{8 \ d \ z}\right\}\ (1-z)^{-d+p}. $$ Again for small $d\ z$ the exponential prefactor should be avoided for a good approximation. For $z$ approaching 1 the approximation breaks down.

All computations were conducted with Mathematica 10.

$\endgroup$
  • $\begingroup$ @Scott Hill - There's a very detailed example of this type of computation in Bender and Orszag, "Advanced Mathematical Methods for Scientists and Engineers," pages 304-305. (Nice answer, Johannes.) $\endgroup$ – Tom Dickens Jul 28 '16 at 22:28
  • $\begingroup$ @Tom Dickens Thanks, Tom. For the computation in Bender&Orszag, see my answer here: mathoverflow.net/questions/185822/…. $\endgroup$ – Johannes Trost Jul 29 '16 at 18:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.