7
$\begingroup$

Consider the equation

$x^5-2x^2+z=0$

How do you derive the Lagrange inversion theorem series solution for it? I know it exists because the answer is here for any trinomial https://arxiv.org/pdf/0910.2957.pdf

I am trying to figure out how derive the series solution.

There is this well known result, which I think is needed:

$$ h(f^{-1}(z))=h(0)+\sum_{n\geq 1}\frac{z^n}{n}\cdot [z^{n-1}]\left(h'(z)\cdot\left(\frac{z}{f(z)}\right)^n\right).$$

What functions should I use for f(z), h(z)

There is no information about this online. Most sources just give the series solution for $x^5-x+a=0$ which has $f(x)=x^5-x$ and then the bring radical solution is as follows by applying the above formula.

the best I could come up with is:

$z=2y-y^{5/2}$ taking the square root of the solution of this gives the solution to the original equation

so $h(z)$ would be a square root

$$ h(f^{-1}(z))=\sum_{n\geq 1}\frac{z^n}{n}\cdot [y^{n-1}]\left(1/2y^{-1/2}\cdot\left(\frac{1}{2-y^{3/2}}\right)^n\right).$$

This gets messy and does not generate integer y's , so it does not work. Any ideas.

$\endgroup$
1

3 Answers 3

11
$\begingroup$

One pair of solutions will be Puiseux series $ax^{1/2}+bx+cx^{3/2}+\cdots$ (where $x^{1/2}$ can have one of two signs). Thus set $u=x^{1/2}$ in the solution, giving a series $F(u)$ satisfying $F(u)^5-2F(u)^2+u^2=0$. Hence $\sqrt{-F(u)^5+2F(u)^2}=u$. By ordinary Lagrange inversion, $$ [u^n]F(u) =[u^{n-1}]\frac 1n\left(\frac{1}{\sqrt{2-u^3}}\right)^n. $$

Addendum. The series $F(x^{1/2})$ and $F(-x^{1/2})$ give two solutions to $x^5-2x^2+z=0$. The other three solutions $G(x)$ are given by $$ [x^n]G(x) = \frac 1n[x^{n-1}]\left(\frac{x}{2(x+\alpha)^2 -(x+\alpha)^5}\right)^n, $$ for $n\geq 1$, and $G(0)=\alpha$, where $\alpha=2^{1/3}$ (three different values).

Addendum 2. The solution above can easily be generalized. Let $P(t)\in\mathbb{C}[t]$. The fractional power series (Puiseux series) $y$ satisfying $P(y)=x$ are given as follows: let $\alpha$ be a zero of $P(t)$ of multiplicity $m$. Then $y=\alpha + \sum_{n\geq 1} a_n x^{n/m}$, where $$ a_n = \frac 1n[u^{n-1}]\left( \frac{u}{P(u+\alpha)^{1/m}}\right)^n. $$ There are $m$ values of $P(u+\alpha)^{1/m}$ (differing by multiplication by an $m$th root of unity), so we get $m$ solutions corresponding to $\alpha$.

$\endgroup$
1
  • 3
    $\begingroup$ What happened to $z$? $\endgroup$ Mar 18, 2022 at 15:04
4
$\begingroup$

I.- Euler / Lambert trinomial equations. Proof based on

Wang, Fei, Proof of a series solution for Euler’s trinomial equation, ACM Commun. Comput. Algebra 50, No. 4, 136-144 (2016). ZBL1370.33022.,

using Lagrange-Bürmann Theorem (LBT) to derive a series solution. (Technique is used below to solve the equation asked for)

The trinomial equation $$y = q + y^m$$ was solved in series form by Lambert, in 1758. Few years later, 1779, Euler transformed Lambert’s equation into the more symmetrical form with $y=x^{-\beta}$, $m=\alpha/\beta$ and $q=(\alpha-\beta)\cdot v$ as $$x^\alpha-x^\beta = (\alpha-\beta)\cdot v\cdot x^{\alpha+\beta}$$ so that $$v=f(x)=\frac{x^{-\beta}-x^{-\alpha}}{\alpha-\beta}$$ where $f(1)=0$ and $f'(1)=1$, therefore applying LBT $$x = 1+\sum_{n=1}^{\infty}\frac{v^n}{n!}\cdot[ \lim_{x\rightarrow1}\frac{d^{n-1}}{dx^{n-1}}\frac{(x-1)^n}{f(x)^n}] $$ $$x = 1+\sum_{n=1}^{\infty}\frac{v^n(\alpha-\beta)^n}{n}\cdot \mathrm{res}_{x=1}\lbrack(x^{-\beta}-x^{-\alpha})^{-n}\rbrack$$ residues can be computed using some CAS (Wolfram or Maple). This gives $$x = 1+\sum_{n=1}^{\infty}\frac{v^n(\alpha-\beta)^n}{n}\cdot \frac{\prod_{j=1}^{n-1}[1+j\alpha+(n-j)\beta]}{(n-1)!(\alpha-\beta)^n}$$ finally, this expression can be written using Gamma functions or Binomials $$x = 1+\sum_{n=1}^{\infty}\frac{v^n(\alpha-\beta)^n}{n\alpha+1}\cdot \binom{(n\alpha+1)/(\alpha-\beta)}{n}$$ Lambert series is recovered with $\beta = -1$, $\alpha=-m$ and $x=y$, giving $$y = 1-\sum_{n=1}^{\infty}\frac{q^n}{nm-1}\cdot \binom{(nm-1)/(m-1)}{n}$$This series can be obtained directly in the same way using $q=g(y)=y-y^m$ and

computing the summand residues $\mathrm{res}_{y=1}[(y-y^m)^{-n}]$.

Both series can be represented as a special case of the generalized Fox-Wright Function $_k\Psi_ℓ^*(A;B;ζ)$ normalized such that $_k\Psi_ℓ^*(A;B;ζ)|_{ζ=0}=1$ (See this MO Answer here for details). $y$ series for $m\in \mathbb{R} \land m>1$ can be put as $$y = -\frac{1}{m-1}\sum_{n=0}^{\infty}\frac{q^n}{n!}\cdot \frac{\Gamma(-\frac{1}{m-1}+n\cdot\frac{m}{m-1})}{\Gamma(1-\frac{1}{m-1}+n\cdot\frac{1}{m-1})}$$ which converges for $|q|<(m-1)m^{-\frac{m}{m-1}}$ giving $$y=\ _1\Psi_1^*([\frac{-1}{m-1},\frac{m}{m-1}];[1-\frac{1}{m-1},\frac{1}{m-1}];q)$$ A similar solution can be found in Section 4 of

Miller, A. R.; Moskowitz, I. S., Reduction of a class of Fox-Wright Psi functions for certain rational parameters, Comput. Math. Appl. 30, No. 11, 73-82 (1995). ZBL0839.33003.

and

Belkić, Dževad, All the trinomial roots, their powers and logarithms from the Lambert series, Bell polynomials and Fox-Wright function: illustration for genome multiplicity in survival of irradiated cells, J. Math. Chem. 57, No. 1, 59-106 (2019). ZBL1406.92284.

as well.

II.- A slight variation of this procedure allows to find the solution series for the specific problem asked for ($\gamma=2,\ n=2,\ m=5$) assuming $\gamma>0,\ m,n \in \mathbb{N},\ m>n$ $$z=\gamma\cdot\zeta^n-\zeta^m$$ We start from Ch 3 Bürmann's Theorem in

Sofo, Anthony, Computational techniques for the summation of series, New York, NY: Kluwer Academic/Plenum Publishers (ISBN 0-306-47805-6/hbk). xvi, 189 p. (2003). ZBL1059.65002.

Let $\phi(\zeta)$ be a simple function in a domain $\mathbb{D}$, zero at a point $a$ of $\mathbb{D}$, and let $$\vartheta(\zeta)=\frac{\zeta-a}{\phi(\zeta)},\ \ \vartheta(a)=\frac{1}{\phi'(a)}$$ If $f(\zeta)$ is analytic in $\mathbb{D}$ then $\forall \zeta \in \mathbb{D}$

$$f(\zeta)=f(a)+\sum_{k=1}^n \frac{[\phi(\zeta)]^k}{k!}\cdot \frac{d^{k-1}}{dt^{k-1}}[f'(t)\cdot[\vartheta(t)]^k]_{t=a}+R_{n+1}$$ $$f(\zeta)=f(a)+\sum_{k=1}^n \frac{[\phi(\zeta)]^k}{k}\cdot\mathrm{res}_{\,t=a}[f'(t)\cdot[\phi(t)]^{-k}]+R_{n+1}$$ This provides an expansion of a function in terms of series of powers of another function. Conditions for convergence can be seen in the given references. For now we can take it as a formal series. Using these relationships $$z=\phi(\zeta)=\gamma\cdot\zeta^n-\zeta^m,\ \ \ f(\zeta)=\zeta$$ The zeroes of $\phi(\zeta)$ are given by $m-n$ different roots $\zeta_\ell=\gamma^{\frac{1}{m-n}}\cdot\omega_{m-n}^\ell,\ \ \ell=0,1,..,m-n-1$ being $\omega_{k}$ the $k$-th root of unity and $\zeta_{m-n}=0$ with multiplicity $n$. Thus $$\zeta=\zeta_j+\sum_{k=1}^\infty \frac{z^k}{k}\cdot \mathrm{res}_{t=\zeta_j}[(\gamma\cdot t^n-t^m)^{-k}]$$ For example by taking $\gamma=2,\ n=2,\ m=5$. The root $\zeta_\mathrm{o}=2^{1/3}$ provides $$\zeta=2^\frac{1}{3}+\sum_{k=1}^\infty c_k z^k$$ whose first 10 terms $c_k$ are given by

enter image description here

The same procedure can be followed with every zero of $\phi(\zeta)$, giving different expansions.

III.- Another approach is to consider a solution based on Fox Wright series function. This MO link provides the inversion series that corresponds to a $_1\Psi_1^*$ function. For $\omega\in\mathbb{C},\ \alpha,\beta\in\mathbb{R}$ the general (standard) trinomial equation $$u-\omega\cdot u^\alpha-1=0$$ has the principal solution for $\alpha>1$ $$u =\,_{1}\Psi_{1}^{*}([1,\alpha];[2,\alpha-1];\omega)$$ whose powers satisfy $$u^\beta=\,_{1}\Psi_{1}^{*}([\beta,\alpha];[\beta+1,\alpha-1];\omega)$$ giving these series converging for $|w|<(\alpha-1)^{\alpha-1}\alpha^{-\alpha}$ $$u=\sum_{k=0}^{\infty}\frac{\omega^k}{k!}\cdot\frac{\Gamma(1+\alpha\cdot k)}{\Gamma(2+(\alpha-1)\cdot k)}$$ $$u^\beta=\beta\cdot\sum_{k=0}^{\infty}\frac{\omega^k}{k!}\cdot\frac{\Gamma(\beta+\alpha\cdot k)}{\Gamma(\beta+1+(\alpha-1)\cdot k)}$$ respectively. Since $\log{u}=\lim_{\beta\rightarrow 0}\frac{u^\beta-1}{\beta}$, logarithm series is obtained from this last series as well$$\log{u}=\sum_{k=1}^{\infty}\frac{\omega^k}{k!}\cdot\frac{\Gamma(\alpha\cdot k)}{\Gamma(1+(\alpha-1)\cdot k)}$$

The original problem $z=\gamma\cdot\zeta^n-\zeta^m$ can be worked now using simple algebra. Set $$u\cdot z =\gamma\cdot\zeta^n,\ \alpha=m/n,\ \omega=z^{-1}(z/\gamma)^\alpha\ \mathrm{and}\ \beta=1/n$$ This produces the Puiseux series (although it is also valid for $m,n\in\mathbb{R},\ \mathrm{s.t.}\ m>n>0$)$$\zeta=\frac{1}{n}(z/\gamma)^\frac{1}{n}\sum_{k=0}^{\infty}\frac{[z^{-1}(z/\gamma)^{\frac{m}{n}}]^k}{k!}\cdot\frac{\Gamma(\frac{1}{n}+\frac{m}{n}\cdot k)}{\Gamma(1+\frac{1}{n}+(\frac{m}{n}-1)\cdot k)}$$ Case $[m,n,\gamma]=[5,2,2]$, gives $$\zeta=\frac{1}{2}\sqrt{z/2}\sum_{k=0}^{\infty}\frac{(z/2)^{\frac{3}{2}k}}{2^k\,k!}\cdot\frac{\Gamma(\frac{1}{2}+\frac{5}{2}\cdot k)}{\Gamma(\frac{3}{2}+\frac{3}{2}\cdot k)}$$ converges for $0\le z\le (\frac{m}{n}-1)(\frac{\gamma\,n}{m})^\frac{m}{m-n}=1.034128651...$

$\endgroup$
4
$\begingroup$

Addendum, for trinomials of the form $x^n+bx^m+c=0$ for integers $n>m$, we can get $n-m$ roots using a simple modification Glasser formula, as cited in Wikipedia .

https://en.wikipedia.org/wiki/Bring_radical

What if instead of $ x = \zeta^{-\frac{1}{N-1}}\,$ we make the transformation more general, like this:

$x = \zeta^{-\frac{k}{N-1}}$

This means that the resulting root is raised to some $k$ power as a series.

So the idea behind the modification is we:

Start by letting N be a fraction and then apply the general transformation to raise this root to the 1/N th power

so solving $x^5-2x^2-.1=0$ would be

$y^{5/2}-2y-.1=0 $

$N=5/2$

the root of original equation is $y^{1/2}$ which uses the modification

Putting it all together, we have:

$q=(-1/b)^{1/(m-n)}$

$p_2=e^\frac{2\pi iv(m(1+w)-1}{n-m}$

$p_1=e^\frac{-2\pi iv}{n-m}$

$v$ runs through $0,1,...n-m-1$

$q\left[ p_1+\frac{1}{m-n}\sum_{w=0} \frac{ p_2q^{-n(w+1)} c^{w+1} }{w+1} {\frac{1-m-nw}{m-n} \choose w} \right] $

are n-m roots for the trinomial. this bypasses the cumbersome resolvent method of having to sum $n-1$ number of hypergeometric functions, by using a single series. It would seem like something that should have been discovered by now.

Example: find a series solution for three roots of $x^5-2x^2+.1=0$

https://www.wolframalpha.com/input?i=x%5E5-2x%5E2%2B.1%3D0

one of the complex roots

first complex root

second complex root

real root

This question is pretty popular. The explicit formulas for reference purposes are as follows:

Series solution for powers $k$ of the large roots for the trinomial: $x^v-jx^u+z=0$:

\begin{equation} R_I^k(z) = j^{\frac{k}{v-u}}\left[e^{\frac{2kI\pi i}{u-v}}+\frac{k}{u-v}\sum_{n=0} \frac{z^{1+n}}{n+1} j^{\frac{v(n+1)}{u-v}} e^{\frac{2I\pi i(u(1+n)-k)}{v-u}} {\frac{vn+u-k}{v-u} \choose n}\right] \end{equation}

for $I=0,1,..v-u-1$

Powers $k$ of the $u$ quantity of small roots of the trinomial $x^v-jx^u+z=0$:

\begin{equation} f^{k}(z)=r_I^k=k\sum_{n=0} \frac{e^{2i{\pi}I( (v-u)n+k)/u} z^{(n(v-u)+k)/u}}{j^{(nv+k)/u} (n(v-u)+k)} {nv/u +k/u-1 \choose n} \end{equation}

for $I=0,1,...u-1$

Source: https://arxiv.org/pdf/2212.09919.pdf

$\endgroup$
2
  • $\begingroup$ Is there any way to find to find the other two real roots using this method? $\endgroup$ Apr 23, 2023 at 14:43
  • $\begingroup$ Richard Stanley above gave a method for finding the large roots. This requires a different substitution owing to convergence issues. I updated my answer to include explicit formulas for large and small roots. $\endgroup$
    – CarP24
    Apr 24, 2023 at 18:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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