Fixed Point and Steffensen's Acceleration
Forfatter:
zekeriya
Sidst opdateret:
9 år siden
Licens:
Creative Commons CC BY 4.0
Resumé:
Math336 Project
\begin
Opdag hvorfor 18 millioner mennesker verden rundt stoler på Overleaf med deres arbejde.
Math336 Project
\begin
Opdag hvorfor 18 millioner mennesker verden rundt stoler på Overleaf med deres arbejde.
\documentclass{beamer}
\mode<presentation>
{
\usetheme{Madrid} % or try Darmstadt, Madrid, Warsaw, ...
\usecolortheme{rose} % or try albatross, beaver, crane, ...
\usefonttheme{structurebold} % or try serif, structurebold, ...
\setbeamertemplate{navigation symbols}{}
\setbeamertemplate{caption}[numbered]
}
\usepackage[english]{babel}
\usepackage[utf8x]{inputenc}
\title[Fixed Point and Steffensen's Acceleration]{MATH 336 Presentation \\Fixed Point and Steffensen's Acceleration Method}
\author{Zekeriya Ünal}
\institute{Boğaziçi University}
\date{27/05/2015}
\begin{document}
\begin{frame}
\titlepage
\end{frame}
\begin{frame}[plain]
\tableofcontents
\end{frame}
% Uncomment these lines for an automatically generated outline.
%\begin{frame}{Outline}
% \tableofcontents
%\end{frame}
\section{Fixed Point Iteration Method}
\begin{frame}{Fixed Point Iteration Method}
\begin{block}{Definition}<1->
A point p is a \textbf{fixed point} for a given function g if g(p)= p.
\end{block}
\begin{block}{Remark}<2->
Fixed point problems and root finding problems are infact equivalent.
\begin{itemize}
\item<1->if p is a fixed point of the function g, then p is a root of the function
$$f(x)=[g(x)-x]h(x)$$ [as long as $h(x)\in \mathbb{R}$]
\item<2->if p is a root of the function of f, then p is a fixed point of the function
$$g(x)=x-h(x)f(x)$$ [as long as $h(x)\in \mathbb{R}$]
\end{itemize}
\end{block}
\end{frame}
\begin{frame}{Fixed Point Iteration Method}
\begin{block}{Definition}<1->
Let U be a subset of a metric space X.
\\A function g:U $\rightarrow$X called \textbf{Lipschitz continuous} provided there exists a constant $\lambda\ge$ 0 (called Lipschitz constant)
\\such that $\forall$ x,y$\in$U d(g(x),g(y))$\le$d(x,y)
\\if $\lambda\in$[0,1], then g is called \textbf{contraction} (with contraction constant $\lambda$).
\end{block}
\begin{block}{Theorem (A Fixed Point Theorem)}<2->
Suppose $g:[a,b]\rightarrow[a,b]$ is continuous. Then g has a fixed point.
\end{block}
\end{frame}
\begin{frame}{Fixed Point Iteration Method}
\begin{block}{Lemma}<1->
A contraction has at most one fixed point.
\end{block}
\begin{block}{Corollary}<2->
Suppose $g:[a,b]\rightarrow[a,b]$ is continuous and $\lambda:=sup\mid g'(x)\mid<1$ for $x\in (a,b)$
\\Then g is a contraction with contraction constant $\lambda.$
\end{block}
\vskip 1cm
\end{frame}
\begin{frame}{Graphical determination of the existence of a fixed point for the function $g(x)= \frac{x^2-3}{2}$}
\includegraphics[width=355px, height= 200px]{Figure2.jpg}
\end{frame}
\subsection{Banach Fixed Point Theorem}
\begin{frame}{Banach Fixed Point Theorem}
\begin{theorem}[Banach Fixed Point Theorem]
\begin{itemize}
\item[]<1-> Let U be a complete subset of a metric space X, and let g:U$\rightarrow$U be a contraction with contraction constant $\lambda$.
\\Then \begin{itemize}
\item<2-> g has a unique fixed point, say p.
\item<2->For any sequence $\left\{x_{n}\right\}$ defined by $x_n$=g($x_{n-1}$), n=1,2,,...
\\converges to this unique fixed point p.
\item[]<3->Moreover, we have the \textbf{a priori} error estimate
$$ \mid x_n-p\mid\le\frac{\lambda^n}{1-\lambda}\mid x_1-x_0\mid$$
\item[]<4->and the \textbf{a posteriori} error estimate
$$\mid x_n-p\mid\le\frac{\lambda}{1-\lambda}\mid x_n-x_{n-1}\mid$$
\end{itemize}
\end{itemize}
\end{theorem}
\vskip 1cm
\end{frame}
\begin{frame}{Banach Fixed Point Theorem}
\begin{block}{Proof}
For $n > m$ , we have
\begin{align*}
\mid x_n-x_m\mid&=\mid x_n-x_{n-1}+x_{n-1}-x_{n-2}+...+x_{m+1}-x_{m}\mid\\
&\le\mid x_n-x_{n-1}\mid + \mid x_{n-1}-x_{n-2}\mid+...+\mid x_{m+1}-x_{m}\mid\\
&by *\\
&\le(\lambda^{n-1}+\lambda^{n-2}+...+\lambda^{m})\mid x_1-x_0\mid\\
&=\lambda^{m}(\lambda^{n-m-1}+\lambda^{n-m-2}+...+1)\mid x_1-x_0\mid\\
&=\lambda^{m}\frac{1-\lambda^{n-m}}{1-\lambda}\mid x_1-x_0\mid\le \frac{\lambda^m}{1-\lambda}\mid x_1-x_0\mid
\end{align*}
so that ${x_n}$ is Cauchy sequence in U.
\\Since U is complete, ${x_n}$ converges to a point p $\in$ U
$$*\color{red}\mid x_k-x_{k-1}\mid=\mid g(x_{k-1})-g(x_{k-2})\mid\le\lambda\mid x_{k-1}-x_{k-2}\mid\le..\le\lambda^{k-1}\mid x_1-x_0\mid\color{black}$$
\end{block}
\vskip 1cm
\end{frame}
\begin{frame}
\begin{proof}[Continue]
\begin{itemize}
\item[]<1->Now,since g being contraction is continuous, we have
$$p=\lim_{n\to\infty}x_{n}=\lim_{n\to\infty}g(x_{n-1})=g(lim_{n\to\infty}x_{n-1})=g(p)$$
\\so that p is fixed point of g.
\\By the lemma p is the unique fixed point of g.
\item[]<2->Since $$\mid x_n-x_m\mid\le \frac{\lambda^m}{1-\lambda}\mid x_1-x_0\mid,$$
\\letting n$\rightarrow\infty$
\item[]<3->we get $$\mid p-x_m\mid\le \frac{\lambda^m}{1-\lambda}\mid x_1-x_0\mid$$
\\for $y_0=x_{n-1},$ $y_1= x_n$
$$\mid y_1- p\mid \le \frac{\lambda}{1-\lambda}\mid y_1-y_0\mid$$
\end{itemize}
\end{proof}
\vskip 1cm
\end{frame}
\subsection{The Fixed Point Algorithm}
\begin{frame}{The Fixed Point Algorithm}
\begin{block}{The Fixed Point Algorithm}<1->
If g has a fixed point p, then the fixed point algorithm generates a sequence $\left\{x_n\right\} $ defined as
\\$x_0$: arbitrary but fixed,
\\$x_n=g(x_{n-1})$, n=1,2,3,... to approximate p.
\end{block}
\end{frame}
\subsection{Fixed Point The Case Where Multiple Derivatives Are Zero at The Fixed Point}
\begin{frame}{Fixed Point The Case Where Multiple Derivatives Are Zero at The Fixed Point}
\begin{block}{Theorem}
\begin{itemize}
\item[]<1->Let g be a continuous function on the closed internal $[a,b]$ with $\alpha>1$ continuous derivatives on the internal (a,b).
\\Further, Let p $\in $(a,b) be a fixed point of g.
\item[]<2->if $$g'(p)=g''(p)=...=g^{(\alpha-1)}(p)=0$$ but $g^{(\alpha)}(p)\neq 0$,
\item[]<3->then there exist a $\delta>0$ such that for any $p_0\in[p-\delta,p+\delta]$, the sequence $p_n=g(p_{n-1})$ converges to the fixed point p of order $\alpha$ with asymtotic error constant $$\lim_{n\to\infty}\frac{\mid e_{n+1}\mid}{\mid e_n\mid^\alpha}=\frac{\mid g^{(\alpha)}(p)\mid}{\alpha!}$$.
\end{itemize}
\end{block}
\end{frame}
\begin{frame}
\begin{block}{Proof}
\begin{itemize}
\item[]<1->Let$'$s start by establishing the existence of $\delta>0$ such that for any $p_0\in [p-\delta,p+\delta]$, the sequence $p_n=g(p_{n-1})$ converges to the fixed point p.
\item[]<2->Let $\lambda<1$. Since $g'(p)=0$ and $g'$ is continuous, it follows that there exists a $\delta>0$ such that$\mid g'(x)\mid\le \lambda<1$ for all $x\in I \equiv[p-\delta,p+\delta]$ From this, it follows that g:I$\rightarrow$I; for if x$\in$ I then,
\item[]<3->
\begin{align*}
\mid g(x)-p\mid&=\mid g(x)-g(p)\mid\\
&=\mid g'(\xi)\mid\mid x-p\mid\\
&\le \lambda\mid x-p\mid<\mid x-p\mid\\
&\le\delta
\end{align*}
\item[]<4->Therefore by the a fixed point theorem established earlier, the sequence $p_n=g(p_{n-1})$ converges to the fixed point p for any $p_0\in [p-\delta,p+\delta]$.
\end{itemize}
\end{block}
\vskip 1cm
\end{frame}
\begin{frame}
\begin{block}{Continue}
\begin{itemize}
\item[]<1->To establish the order of convergence, let x$\in$ I and expand the iteration function g into a Taylor series about x=p:
$$g(x)=g(p)+g'(p)(x-p)+...+\frac{g^{\alpha-1}(p)}{(\alpha-1)!}(x-p)^{\alpha-1}+\frac{g^{\alpha}(\xi)}{(\alpha)!}(x-p)^{\alpha}$$
\\where $\xi$ is between x and p.
\item[]<2->Using the hypotheses regarding the value of $g^{(k)}(p)$ for $1\le k\le \alpha-1$ and letting $x=p_n$, the Taylor series expansion simplifies to $$p_n+1-p=\frac{g^(\alpha)(\xi)}{\alpha!}(p_n-p)^\alpha$$
\\where $\xi$ is now between $p_n$ and p.
\end{itemize}
\end{block}
\vskip 1cm
\end{frame}
\begin{frame}
\begin{proof}[Continue]
\begin{itemize}
\item[]<1->The definitions of fixed point iteration scheme and of a fixed point have been used to replace $g(p_n)$ with $p_{n+1}$ and g(p) with p.
\item[]<2->Finally, let $n\rightarrow\infty$.Then $p_n\rightarrow p$, forcing $\xi\rightarrow$ p also.
Hence $$\lim_{n\to\infty}\frac{\mid e_{n+1}\mid}{\mid e_n\mid^\alpha}=\frac{\mid g^{(\alpha)}(p)\mid}{\alpha!}$$
\\or $p_n\rightarrow$ p of order $\alpha$.
\end{itemize}
\end{proof}
\end{frame}
\section{Steffensen's Acceleration Method}
\subsection{Aitken's $\Delta^2$ Method}
\begin{frame}
\begin{theorem}[Aitken's $\Delta^2$ method]
\begin{itemize}
\item[] <1-> Suppose that
\begin{itemize}
\item $\left\{x_{n}\right\}$ is a sequence with $x_{n}\neq$ p for all n $\in \mathbb{N}$
\item there is a constant c $\in \Re\setminus \left\{ \mp 1 \right\}$ and a sequece $\left\{ \delta_{n} \right\}$
such that
\begin{itemize}
\item $\lim_{n\rightarrow\infty}\delta_{n}=0$
\item $x_{n+1}-p=(c+\delta_{n})(x_{n}-p)$ for all n $\in \mathbb{N}$
\end{itemize}
\end{itemize}
\item[] <2->Then
\begin{itemize}
\item $\left\{x_{n}\right\}$ converges to p iff $\left|c\right|<1$
\item if $\left|c\right|<1$ , then
\[
y_{n}=\frac{x_{n+2}x_{n}-x_{n+1}^2}{x_{n+2} - 2x_{n+1}+x_{n}}= x_n-\frac{(x_{n+1}-x_n)^2}{x_{n+2}-2x_{n+1}+x_n}
\]
\end{itemize}
\ is well-defined for all sufficiently large n.
\item[]<3-> Moreover $\left\{y_{n}\right\}$ converges to p faster than $\left\{x_{n}\right\}$ in the sense that
\[
\lim_{n\rightarrow\infty} \frac{y_{n}-p}{x_{n}-p}=0
\]
\end{itemize}
\end{theorem}
\vskip 1cm
\end{frame}
\begin{frame}
\begin{proof}
\begin{itemize}
\item[]<1-> Let $e_{n} = x_{n}-p$
\item[]<2-> $y_{n}-p=\frac{x_{n+2}x_{n}-x_{n+1}^2}{x_{n+2} - 2x_{n+1}+x_{n}}-p$
\item[]<3-> $y_{n}-p$=$\frac{(e_{n+2}+p)(e_{n}+p)-(e_{n+1}+p)^2}{(e_{n+2}+p)-2(e_{n+1}+p)+(e_{n}+p)}$
\item[]<4-> $y_{n}-p$=$\frac{e_{n+2}e_{n}-e_{n}^2}{e_{n+2}-2e_{n+1}+e_{n}}$
\item[]<5-> since we have
\item[]<5-> $x_{n+1}-p= (c+\delta_{n})(x_{n}-p)$
\ and $e_{n+1}=(c+\delta_{n}e_{n})$
\item[]<6-> $y_{n}-p$=$\frac{(c+\delta_{n+1})(c+\delta_{n})e_{n}e_{n}-(c+\delta_{n})^2e_{n}^2}{(c+\delta_{n+1})(c+\delta_{n})e_{n}-2(c+\delta_{n})e_{n}+e_{n}}$
\item[]<7-> $y_{n}-p$=$\frac{(c+\delta_{n+1})(c+\delta_{n})-(c+\delta_{n})^2}{(c+\delta_{n+1})(c+\delta_{n})-2(c+\delta_{n})+1}$ $(x_{n}-p)$
\item[]<8-> $y_{n}-p$=$\frac{(c+\delta_{n})(\delta_{n+1}-\delta_{n})}{(c-1)^+c(\delta_{n+1}+\delta_{n})+\delta_{n}(\delta_{n+1}-2)}$
\item[]<9-> Therefore $\lim_{n\rightarrow\infty} \frac{y_{n}-p}{x_{n}-p}=0$
\end{itemize}
\end{proof}
\end{frame}
\subsection{Steffensen's Acceleration Method}
\begin{frame}
\begin{block}{Steffensen's Acceleration Method}
\begin{itemize}
\item[]<1->Steffensen’s Method is a combination of fixed-point iteration and the Aitken’s $\Delta^2$ method:
\item[]<2->Suppose we have a fixed point iteration: $$x_0,x_1=g(x_0),x_2=g(x_1),...$$
\item[]<3->Once we have $x_0, x_1,$ and $x_2,$ we can compute $$ y_0=x_0-\frac{(x_1-x_0)^2}{x_2-2x_1+x_0}$$
At this point we "restart" the fixed point iteration with $x_0=y_0$
\item[]<4->e.g.\color{brown} $$x_3=y_0, x_4=g(x_3), x_5=g(x_4),$$\color{black}
and compute \color{brown} $$y_3=x_3-\frac{(x_4-x_3)^2}{x_5-2x_4+x_3}$$ \color{black}
\end{itemize}
\end{block}
\end{frame}
\section{Comparison With Fixed point iteration and Steffensen's Acceleration Method}
\begin{frame}{Comparison with Fixed Point Iteration and Steffensen's Acceleration Method}
\begin{block}{EXAMPLE}
\begin{itemize}
\item[]<1-> Use the Fixed Point iteration method to find a solution to $f(x) = x^2-2x-3$ using $x_0=0$, tolerance =$10^{-1}$ and compare the approximations with those given by Steffensen's Acceleration method with $x_0=0$, tolerance = $10^{-2}$.
\item<2-> We can see that my MATLAB code while Fixed Point iteration method reaches the root by 788 iteration, Steffensen's Acceleration method reaches the root by only 3 iterations.
\end{itemize}
\end{block}
\end{frame}
\end{document}