\documentclass[12pt]{article}
\def\rit{ \hbox{\it I\hskip -2pt R} }
\topmargin=-0.5in
\textheight=8.5in 

\begin{document}

{\bf Math 273}: {\bf Homework \#2, due on Monday, November 2nd} 

\medbreak 

\noindent{\bf [1]} Consider the minimization problem 
$$\inf_{u}F(u)=\int_{x_0}^{x_1} L(x,u(x),u'(x),u''(x))dx,$$
with $u(x_0)=u_0$, $u(x_1)=u_1$, $u'(x_0)=U_0$, $u'(x_1)=U_1$ given, and $L$ is a sufficiently smooth 
function. Obtain the Euler-Lagrange equation of the minimization problem that is satisfied by a smooth optimal $u$. Choose test functions $v$ in $C^{\infty}[x_0,x_1]$ that satisfy $v(x_0)=v(x_1)=v'(x_0)=v'(x_1)=0$, and proceed as in HW1, problem [6] (you should obtain a fourth-order differential equation). 

\medbreak 

\noindent{\bf [2]} Consider the 1D length functional minimization problem 
$$\mbox{Min}_u F(u)=\int_{0}^1 L(u'(x))dx, \mbox{ or }\mbox{Min}_{u}\int_0^1 \sqrt{1+(u'(x))^2}dx,
$$
over functions $u:[0,1]\rightarrow \rit$ with boundary conditions $u(0)=0$, $u(1)=1$. 

(a) Find the exact solution of the problem. 

(b) Show that the functional $u\mapsto F(u)$ is convex. 

(c) Consider a discrete version of the problem: let 
$$x_0=0 <x_1< ... <x_n< x_{n+1}=1$$ 
be equidistant points, with $x_{i+1}-x_i=h$. For $\vec{u}=(u_1,...,u_n)$, 
consider $f(\vec{u})=\sum_{i=0}^{n}\sqrt{1+\Big(\frac{u_{i+1}-u_i}{h}\Big)^2}$, with the additional condition that $u_0=0$ and $u_{n+1}=1$. 

Choose an appropriate discretization integer $n$. Then numerically and experimentally analyze the behavior of the gradient descent method with backtracking line search. Choose the initial starting point $u^0$ as a curve joining the points $(0,0)$ and $(1,1)$. 
Record the number of iterations and plot the error $u^k-u^*$, where $u^*$ is the exact minimizer. You could also plot the curve given by $\vec{u}^k$ at some iterations. 

(d) Repeat question (c), using now Newton's method. 

(e) Discuss the results obtained in (c) and (d). 

\medbreak 

\noindent{\bf [3]} Let $A:\rit^n\rightarrow\rit^n$ be a (linear) self-adjoint operator, $b\in \rit^n$, and consider the quadratic function for $x\in \rit^n$
$$x\mapsto q(x):= \langle Ax,x\rangle -2\langle b,x\rangle. $$

Show that the three statements 

(i) $\inf\{q(x):\ x\in \rit^n\} > -\infty $

(ii) $A\geq O$ and $b\in \mbox{Im} A$. 

(iii) the problem $\inf\{q(x):\ x\in \rit^n\} > -\infty $ has a solution

\noindent are equivalent. When they hold, characterize the set of minimum points of $q$, in terms of the pseudo-inverse of $A$. 

\medbreak 

{\bf Notes:} 

\medbreak 

$\bullet$ If $A$ is a symmetric (or self-adjoint) linear operator on $X$, then 
$\mbox{Im} A^{\perp}=\mbox{Ker}A$. Let $p_{\mbox{Im}A}$ be the operator of orthogonal projection onto $\mbox{Im}A$. For given $y\in X$, there is a unique $x=x(y)$ in $\mbox{Im}A$ such that $Ax=p_{\mbox{Im}A}y$. Forthermore, the mapping $y\mapsto x(y)$ is linear. This mapping is called the pseudo-inverse, or generalized inverse of $A$. 

\medbreak 

$\bullet$ Let $\Omega$ be an open and bounded subset of $R^d$, with Lipschitz-continuous (or sufficiently smooth) boundary $\partial\Omega$. Let $\vec{n}=(n_1,n_2,...,n_d)$ be the exterior unit normal to $\partial\Omega$. 
 Recall the following fundamental Green's formula, or integration by parts formula: given two functions $u,v$ (with $u$, $v$, and all their 1st order partial derivatives belonging to $L^2(\Omega)$, or $u,v\in H^1(\Omega)$), then 
$$\int_{\Omega}u v_{x_i} dx=-\int_{\Omega} u_{x_i} v dx +\int_{\partial\Omega} u v n_i dS.$$




\end{document}