Homotopy Continuation is a numerical method for solving systems of polynomial equations.
It is based on the idea of ”deforming” a given system of equations into a simpler one, whose
solutions are known, and then tracking the solutions of the original system as the deformation
is undone.
In this project, the method will be implemented in the Julia programming language, making use
of parallel computing in order to speed multiple root finding. The method is described in detail
in \cite{BertiniBook}, which was the primary source for this report.
\section{Homotopy Continuation}
We will only consider \textit{square} systems of polynomial equations, i.e. systems of $n$ polynomial equations in $n$ variables, although or over- or under-determined systems can
often be solved by reducing them to square systems, by respectively choosing a suitable square subsystem or adding equations. Morever, we will restrict ourselves to systems with
isolated solutions, i.e. zero-dimensional varieties.
There are many ways to choose the "simpler" system, from now on called a \textit{start system}, but in general we can observe that, by Bezout's theorem, a system
$F=(f_1,\ldots,f_n)$ has at most $D:=d_1\ldots d_n$ solutions, where $d_i$ is the degre of $f_i(x_1,\ldots,x_n)$. So, we could build a start system of the same size and whose
polynomials have the same degrees, but whose solutions are easy to find, and thus can be used as starting points for the method.
For instance, the system $G=(g_1,\ldots g_n)$, where
$$ g_i(x_1,\ldots x_n)=x_i^{d_i}-1 ,$$
is such a system, since it has exactly the $D$ solutions
$$\left\{\left(e^{\frac{k_1}{d_1}2\pi i},\ldots,e^{\frac{k_n}{d_n}2\pi i}\right),\text{ for }0\leq k_i\leq d_i-1\,\text{ and }i=1,\ldots,n\right\} .$$
where $x:=(x_1,\ldots,x_n)$ and $t\in[0,1].$ This is such that the roots of $H(x,0)=G(x)$ are known, and the roots of $H(x,1)=F(x)$ are the solutions of the original system. Therefore, we can implicitly
define a curve $z(t)$ in $\C^n$ by the equation \begin{equation}\label{eq:h2} H(z(t),t)=0,\end{equation} so that in order to approximate the roots of $F$ it is enough to numerically track $z(t)$.
To do so, we derive the expression \eqref{eq:h2} with respect to $t$, and get the \textit{Davidenko Differential Equation}
This is a system of $n$ first-order differential equations, which can be solved numerically for $z(t)$ as an initial value problem, and is called \textit{path tracking}.
While \eqref{eq:h1} is a fine choice of a homotopy, it's not what it's called a \textit{good homotopy}: in order to ensure that the solution paths $z(t)$ for different roots
we can employ the \textit{Gamma trick}: this consists in modifying the linear homotopy
\eqref{eq:h1} by susbtituting the parameter $t\in[0,1]$ with a complex curve $q(t)$ connecting $0$ and $1$:
$$ q(t)=\frac{\gamma t}{\gamma t+(1-t)} ,$$
where $\gamma\in(0,1)$ is a random complex parameter.This "probability one" procedure, i.e. for any particular system choosing $\gamma$ outside of a finite amount of lines through
the origin ensures that we get a good homotopy, basically because of the finiteness of the branch locus of the homotopy.
We now want to track down individual roots, following the solution paths from
a root $z_0$ of the start system by solving the initial value problem associated to the Davidenko differential equation \eqref{eq:dav} with starting value $z_0$ and
$t$ ranging from $1$ to $0$.
This will be done numerically, using a first-order predictor-corrector tracking method, which consists in first using Euler's method to get an approximation
$\widetilde{z}_i$, and then using Newton's method to correct it
using equation \eqref{eq:h2} so that it becomes a good approximation $z_i$ of the next value of the solution path.
Recall that Euler's method consists in approximating the solution of the initial value problem associated to a first-order ordinary differential equations
% Braced system of equations below
\begin{equation*}
\left\{
\begin{aligned}
&\dot{z}=f(z,t)\\
&z(t_0)=z_0
\end{aligned}
\right.
\end{equation*}
by the sequence of points $(z_i)_{i\in\N}$ defined by the recurrence relation
$$ z_{i+1}=z_i+h\cdot f(z_i,t_i) ,$$
where $h$ is the step size.
In our case, we have $$f(z,t)=-\left(\frac{\partial H}{\partial z}(z,t)\right)^{-1}\frac{\partial H}{\partial t}(z,t)$$ and $t_0=1$, since we track from $1$ to $0$. For the same
Since we want to solve $$H(z,t)=0,$$ we can use Newton's method to improve the approximation $\widetilde{z_i}$ obtained by Euler's method to a solution of such equation.
This is done by moving towards the root of the tangent line of $H$ at the current approximation, or in other words through the iteration
In order to improve the efficiency of the method, we will use an adaptive step size, which will be based on the norm of the residual of the Newton iteration.
then we halve the step size; if instead we have 5 "successful" iterations in a row, we double the step size, as implemented in Appendix \hyperref[code:adapt]{A}.
When testing the method, we tried to use multithreading to speed up the computation.
This was done in Julia by using the \texttt{Threads.@threads} macro, which automatically distributes the work of a \texttt{for} loop among the available threads.
However, in the case of looping over multiple roots, this didn't improve the performance, as the overhead of the multithreading was too big compared to the actual computation time,
as the systems were too small to benefit from this kind of parallelization, as can be seen by the results in Appendix \hyperref[sec:mt]{B}.
\bibitem{BertiniBook} Bates, Daniel J. \textit{Numerically solving polynomial systems with Bertini}. SIAM, Society for Industrial Applied Mathematics, 2013.
