An explicit Taylor series of the inverse error function?

259 Views Asked by At

$\def\erf{\text{erf}}$

Browsing Wolfram Functions, we found the $n$th derivative of the inverse error function and of inverse gamma regularized. Notice the regularized$\,_2\text F_2$ hypergeometric function and the Kronecker delta function $\delta_{n},\delta_{n_1,n_2}$:

$$\frac{d^n\erf^{-1}(z)}{dz^n}=\erf^{-1}(z)\delta_n+\left(\frac{\sqrt\pi}2e^{\erf^{-1}(z)^2}\right)^n\sum_{j_2=0}^n\cdots\sum_{j_n=0}^n (-1)^{\sum\limits_{k=2}^n j_k}\delta_{\sum\limits_{k=2}^n(k-1)j_k,n-1}\left(n-1+\sum_{k=2}^n j_k\right)!\prod_{k=2}^n \frac1{j_k!}\left(\frac{\sqrt\pi2^{k-1}e^{\erf^{-1}(z)^2}}{\erf^{-1}(z)^{k-1}k!}\right)^{j_k}\,_2\tilde{\text F}_2\left(\frac12,1;1-\frac k2,\frac{3-k}2;-\erf^{-1}(z)^2\right)^{j_k}\tag1$$ with the Mathematica code: D[InverseErf[z], {z, n}] == KroneckerDelta[n] InverseErf[z] + (Pi^(n/2)/2^n) E^(n InverseErf[z]^2) Sum[\[Ellipsis] Sum[KroneckerDelta[Sum[(i - 1) Subscript[j, i], {i, 2, n}], n - 1] (-1)^Sum[Subscript[j, i], {i, 2, n}] (n - 1 + Sum[Subscript[j, i], {i, 2, n}])! Product[(1/Subscript[j, i]!) ((2^(-1 + i) E^InverseErf[z]^2 Sqrt[Pi] InverseErf[z]^(1 - i))/i!)^Subscript[j, i] HypergeometricPFQRegularizedq[{1/2, 1}, {1 - i/2, (3 - i)/2}, -InverseErf[z]^2]^Subscript[j, i], {i, 2, n}], {Subscript[j, n], 0, n}], {Subscript[j, 2], 0, n}] /; Element[n, Integers] && n >= 0

No reference, outside of Wolfram functions, for the formula was found, but the top FAQ question implies that the formula has been verified with Mathematica. Therefore, the analytic inverse function’s expansion about $z=\erf(1)=0.8427…$ should be:

$$\erf^{-1}(z)\mathop=^?\sum_{n=0}^\infty\frac{(z-\erf(1))^n}{n!}\left(\delta_n+\left(\frac{\sqrt\pi}2e\right)^n\sum_{j_2=0}^n\cdots\sum_{j_n=0}^n (-1)^{\sum\limits_{k=2}^\infty j_k}\delta_{\sum\limits_{k=2}^n(k-1)j_k,n-1}\left(n-1+\sum_{k=2}^n j_k\right)!\prod_{k=2}^n \frac1{j_k!}\left(\frac{e\sqrt\pi2^{k-1}}{k!}\right)^{j_k}\,_2\tilde{\text F}_2\left(\frac12,1;1-\frac k2,\frac{3-k}2;-1\right)^{j_k}\right)= 1+\frac{e\sqrt\pi}2(z-\erf(1))+\sum_{n=2}^\infty\frac{(z-\erf(1))^n}{n!}\left(\delta_n+\left(\frac{\sqrt\pi}2e\right)^n\sum_{j_2=0}^n\cdots\sum_{j_n=0}^n (-1)^{\sum\limits_{k=2}^nj_k}\delta_{\sum\limits_{k=2}^n(k-1)j_k,n-1}\left(n-1+\sum_{k=2}^n j_k\right)!\prod_{k=2}^n \frac1{j_k!}\left(\frac{e\sqrt\pi2^{k-1}}{k!}\right)^{j_k}\,_2\tilde{\text F}_2\left(\frac12,1;1-\frac k2,\frac{3-k}2;-1\right)^{j_k}\right) $$

There is a Mathematica formula download for $\erf^{-1}(z)$ to test the series expansion, but I do not have Mathematica. Also, the series about $z\to0$ may diverge since $\lim\limits_{z\to0}\frac{e^{\erf^{-1}(z)^2}}{\erf^{-1}(z)^{k-1}}=\infty,k>1$ and the $\sum\limits_{j_2=0}^n\cdots\sum\limits_{j_n=0}^n (-1)^{\sum\limits\limits_{k=2}^n j_k}\delta_{\sum\limits\limits_{k=2}^n(k-1)j_k,n-1}\left(n-1+\sum\limits_{k=2}^n j_k\right)! $ reminds one of the explicit series reversion formula in formula $(11)$. If this series works, then we have an explicit series representation of the inverse error function and possibly an analytic continuation of the function.

Does this $n$th derivative formula give the Taylor series coefficients $a_n$ for $\erf^{-1}(z)=\sum\limits_{n=0}^\infty a_n\frac{(z-z_0)^n}{n!}$? If so, how does one derive an explicit series about $z=0$?

This final question is optional, but @Steven Clark suggested the radius of convergence of $\erf^{-1}(z)=\sum\limits_{n=0}^\infty\frac{(z-z_0)^n}{n!}\frac{d^n}{dz^n}\erf^{-1}(z)\big|_{z=z_0}\text{ using }(1)$ is needed to know convergence and where an analytic continuation is possible.

1

There are 1 best solutions below

9
On

This is not an answer.

Since we faced the problem of the inverse of the error function quite many years ago in my former research group, I give you the results for the inversion of $$z=\text{erf}(x) \qquad \text{around} \qquad x=a$$

It write $$\color{red}{x=\sum_{n=0}^\infty \frac{\pi ^{\frac n2}\, e^{a^2 n} }{n!\,2^{\left\lfloor \frac{n+1}{2}\right\rfloor } }\,P_n(a)\,\big[z-\text{erf}(a)\big]^n}$$ the first polynomials being $$\left( \begin{array}{cc} n & P_n(a) \\ 0 & a \\ 1 & 1 \\ 2 & a \\ 3 & 4 a^2+1 \\ 4 & a \left(12 a^2+7\right) \\ 5 & 96 a^4+92 a^2+7 \\ 6 & a \left(480 a^4+652 a^2+127\right) \\ 7 & 5760 a^6+10224 a^4+3480 a^2+127 \\ 8 & a \left(40320 a^6+88848 a^4+44808 a^2+4369\right) \\ 9 & 645120 a^8+1703808 a^6+1161168 a^4+204328 a^2+4369 \\ 10 & a \left(5806080 a^8+17914752 a^6+15561936 a^4+4161288 a^2+243649\right) \end{array} \right)$$ which show interesting patterns. After reading you post, it is almost obvious that these polynomials correspond to some hypergeometric functions (which I did not identify)

For $a=1$, the $P_n(1)$ generate the sequence $$\{1,1,1,5,19,195,1259,19591,178345,3718793,43687705,1141949277,16433602763\}$$ not found in $OEIS$.