notes

uni notes
git clone git://popovic.xyz/notes.git
Log | Files | Refs

prb3.tex (9236B)


      1 \include{preamble.tex}
      2 
      3 \usepackage[final]{pdfpages}
      4 
      5 \begin{document}
      6 \maketitle
      7 \tableofcontents
      8 \section{Sheet 3}
      9 \subsection{Problem 1}
     10 Take a linear system $Ax = b$, where $A \in \mathbb{R}^{n\times n}$ and $b
     11 \in \mathbb{R}^{n}$. We want to solve it using the Gradient decent method, an
     12 iteration
     13 \begin{align}
     14     x^{k+1} = x^{k} + \alpha_k r^{k},
     15 \end{align}
     16 where $r^{k} = b - Ax^{k}$ and the residual $\alpha_k= \frac{(r^{k})^T
     17 r^k}{(r^k)^T Ar^k}$.
     18 \subsubsection{}
     19 We compute $x^1$ for
     20 \begin{align}
     21     A =
     22     \begin{pmatrix}
     23         2 & -1 \\
     24         -1 & 2
     25     \end{pmatrix},
     26     \qquad
     27         b = \begin{pmatrix}1 \\ 1 \end{pmatrix},
     28 \end{align}
     29 with an initial guess $x^0 = 0$.
     30 \begin{align}
     31     r^0 = \begin{pmatrix} 1 & 1 \end{pmatrix} \qquad
     32     Ar^0 = \begin{pmatrix} 1 & 1 \end{pmatrix} \quad \Rightarrow \quad
     33     \alpha_0 = 1.
     34 \end{align}
     35 Then for $x^1$ we have
     36 \begin{align}
     37     x^1 = \begin{pmatrix} 0 \\ 0 \end{pmatrix}  + \begin{pmatrix}1 \\
     38 1 \end{pmatrix}  = b
     39 \end{align}
     40 \subsubsection{}
     41 Suppose the $k$-th error $e^k = x - x^k$ is an eigenvector of $A$ to the
     42 eigenvalue $\lambda$, then
     43 \begin{align}
     44     A e^{k}=\lambda e^{k} = \lambda (x^{k}-x) = \lambda x^{k} - \lambda x.
     45 \end{align}
     46 For the next iteration step we need $r^{k}$ which is
     47 \begin{align}
     48     r^{k} &= b - Ax^{k} + Ax - Ax = (b - Ax) - A(x^{k} - x) = -\lambda e^{k}\\
     49     (r^k)^{T} r^k &= \lambda^2 (e^k)^{T}e^{k}\\
     50     (r^k)^{T} A r^k &= \lambda^3 (e^k)^{T}e^{k}\\
     51     \Rightarrow \alpha_k &= \frac{1}{\lambda}
     52 \end{align}
     53 Then the next step $x^{k+1}$ is
     54 \begin{align}
     55     x^{k+1} = x^{k}+\alpha_k r^{k} = x^{k}-\frac{\lambda}{\lambda}e^{k} =
     56     x^{k} - e^{k} = x^{k}-x^{k}+ x = x,
     57 \end{align}
     58 i.e. $x^{k+1}$ is then the solution.
     59 \subsection{Exercise 2}
     60 We show the norm equivalence of the vector norms $ \|\cdot\|_\infty$ and
     61 $\|\cdot\|_2$ and the optimality of the constants $C, C' > 0$and the
     62 optimality of the constants $C, C' > 0$
     63 \subsubsection{}
     64 We show
     65 \begin{align}
     66     \|v\|_\infty = \max_i |v_i| \leq \sqrt{\sum_{i}^{} v_i^2}  \quad
     67     \Rightarrow \quad C = 1.
     68 \end{align}
     69 Then
     70 \begin{align}
     71     \|v\|_2 = \sqrt{\sum_i v_i^2} \le \sqrt{\sum_i \max_i |v_i|^2}  =
     72     \sqrt{\sum_i  \|v\|^2_\infty} = \sqrt{n} \|v\|_\infty.
     73 \end{align}
     74 We get
     75 \begin{align}
     76     \|v\|_\infty \le \|v\|_2 \le \sqrt{n} \|v\|_\infty.
     77 \end{align}
     78 For $u := \frac{v}{\|v\|_\infty} \; \Rightarrow \; \|v\|_\infty = 1$ we get
     79 \begin{align}
     80     1 \le  \|u\|_2 \le \sqrt{n}
     81 \end{align}
     82 which states optimality of the constants.
     83 \subsection{Problem 3}
     84 Now we do the same as in Problem 2 but with matrix norms induced by vector
     85 norms, specifically for $\|\cdot\|_2$ and $\|\cdot\|_\infty$ matrix norms
     86 induced by vector norms.
     87 \subsubsection{}
     88 We know that
     89 \begin{align}
     90     \|A\|_2 = \max_{v\neq 0} \frac{\|Ax\|_2}{\|x\|_2} \qquad \|A\|_2 &=
     91     \rho(A^*A)\\
     92     \|A\|_\infty = \max_{i,j}  |a_{ij}|
     93 \end{align}
     94 Together with the norm inequalities
     95 \begin{align}
     96     \|Av\|_\infty &\le \|A\|_\infty \|v\|_\infty\\
     97     \|A\|_2^2 = \rho(A^TA) \le \|A^TA\|_\infty\\
     98     \Rightarrow \|\rho(A^TA)\|_\infty = \|A^TA x\|_\infty \le \|A^T
     99     A\|_\infty \|x\|_\infty,
    100 \end{align}
    101 thereby
    102 \begin{align}
    103     \|A^T A\|_\infty &= \max_{i,j} \left|(A^T A)_{ij})\right| = \max_{i,j}
    104     \left| \sum_l A_{il} A_{lj} \right|  \\
    105                     &=\max_{i,j} \sum_l \left| A_{il} A_{lj} \right| \leq n
    106                     \|A\|_\infty^2\\
    107                     \nonumber\\
    108     \Rightarrow  \|A\|_2 &\leq \sqrt{n} \|A\|_\infty
    109 \end{align}
    110 \subsubsection{}
    111 Let $A \in \mathbb{C}^{n\times n}$, $b \in \mathbb{C}^n$ defined as
    112 \begin{align}
    113     A =
    114     \begin{pmatrix}
    115         1 & \cdots  &  1 \\
    116         0 & \cdots  &  0 \\
    117         \vdots & \vdots  &\vdots   \\
    118         0 & \cdots & 0
    119         \end{pmatrix} \qquad b = \begin{pmatrix} 1 \\ \vdots \\ 1
    120     \end{pmatrix}
    121 \end{align}
    122 Then $Ab = \begin{pmatrix} n & 0 & \cdots & 0 \end{pmatrix}^T$ and the
    123 $\|\cdot\|_2$ of $A$ is
    124 \begin{align}
    125     \|A\|_2 = \max_{v\neq 0} \frac{\|Av\|_2}{\|v\|} = \frac{\|Ab\|_2}{\|b\|}
    126     = \frac{n}{\sqrt{n}} = \sqrt{n}
    127 \end{align}
    128 \subsubsection{}
    129 Let $A$ be defined as above, we show that $\|A\|_\infty = \sqrt{n} \|A\|_2$,
    130 where $C = \sqrt{n} $ is optimal
    131 \begin{align}
    132     \|A\|_\infty = \max_i \sum_j |A_{ij}| = n = \sqrt{n} \sqrt{n}  = \sqrt{n}
    133      \|A\|_2,
    134 \end{align}
    135 thereby $C$ is optimal.
    136 \subsubsection{}
    137 We show that $\|A\|_2 \le \sqrt{n} \|A\|_\infty$ for all $A \in
    138 \mathbb{C}^{n\times n}$, and equality holds for $A$ whose columns are all
    139 zero, accept the first one filled with ones. We know the norms
    140 $\|\cdot\|_2$ and $\|\cdot\|_\infty$ are induced by vector norms which are
    141 consistent. Then for all $v \in \mathbb{C}^{n}$ we have
    142 \begin{align}
    143     \|v\|_\infty &\le\sqrt{n} \|v\|_2\\
    144     \Leftrightarrow \|A\|_\infty &\le \sqrt{n} \|A\|_2 \qquad\ \forall \ A
    145     \in \mathbb{C}^{n\times n}.
    146 \end{align}
    147 \subsection{Problem 4}
    148 Let $\langle x, y \rangle$ be the standard Euclidean scalar product on
    149 $\mathbb{R}^n$ and $\|\cdot\|_2$ the Euclidean norm. Let $B\in
    150 \mathbb{R}^{n\times n}$ be an antisymmetric matrix, i.e. $B^T = -B$. And let
    151 $A:= I - B$
    152 \subsubsection{}
    153 We show that for all $x \in \mathbb{R}^{n}$ we have $\langle Bx, x\rangle =0$
    154 and $\langle Ax, x\rangle = \|x\|_2^2$.
    155 \begin{align}
    156     \langle Bx, x\rangle = (Bx)^T x = x^T B^T x = -x^T B x\\
    157     (-x^TBx)^T &= x^TBx\; \Rightarrow\; x^T Bx = jx^TBx\\
    158     \Rightarrow 2x^T B x &= 2 \langle Bx, x\rangle = 0.
    159 \end{align}
    160 And the second statement follows from the previous equation
    161 \begin{align}
    162     \langle Ax, x\rangle &= \rangle\left(I-B\right)x,x\rangle \\
    163                          &=\langle x, x\rangle - \underbrace{\langle Bx,
    164                          x\rangle}_{=0} = \langle x, x\rangle =
    165     \|x\|_2^2
    166 \end{align}
    167 \subsubsection{}
    168 We show that $\|Ax\|_2^2 = \|x\|_2^2 + \|Bx\|_2^2$ and that $\|A\|_2 =
    169 \sqrt{1+\|B\|_2^2}$. We start with the vector norm
    170 \begin{align}
    171     \|Ax\|^2_2
    172     &= \|x - Bx\|_2^2 \\
    173     &= \langle x -Bx, x-Bx\rangle \\
    174     &= (x^T - x^T B^T)(x - Bx)\\
    175     &= x^T x - x^T B^T x - x^T Bx + x^T B^T Bx\\
    176     &= x^T x + x^TB^TBx = \|x\|_2^2 + \|Bx\|_2^2
    177 \end{align}
    178 where $i, j$ run from $1$ to $n$. Then the vector induced norm follows
    179 \begin{align}
    180     \|A\|_2^2
    181     &= \|I-B\|_2^2 = \sup_{x\neq 0} \frac{\|Ax\|_2^2}{\|x\|^2_2}\\
    182     &= \sup_{x\neq 0} \frac{\|x\|^2_2 +\|Bx\|_2^2 }{\|x\|_2^2} = 1+
    183     \sup_{x\neq 0}\frac{\|Bx\|_2^2}{\|x\|^2_2}= 1+\|B\|_2^2,
    184 \end{align}
    185 pulling the square root on both sides gives us the answer.
    186 \subsubsection{}
    187 We show that $A$ is invertible and the inverse matrix norm is given
    188 \begin{align}
    189     \|A^{-1}\|^2 = \max_{x\neq}\frac{\|x\|^2}{\|Ax\|^2}
    190 \end{align}
    191 To show that $A$ is invertible we use the Rank-Nullity Theorem
    192 \begin{align}
    193     n &=\text{rg}(A) + \dim\text{Im}(A)\\
    194     \Rightarrow n - \dim\text{Im}(A) = rg(A).
    195 \end{align}
    196 If $A$ is invertible it has maximal rank that means $\dim \text{Im}(A) = 0$.
    197 Let's look what the image of A is
    198 \begin{align}
    199     Ax = 0 \Rightarrow \|Ax\|_2^2 = \|x\|_2^2 + \|Bx\|_2^2 = 0
    200 \end{align}
    201 holds only for $x = 0$, thereby $\dim \text{Im}(A) = 0$ and the rank is
    202 maximal. Now let $A^{-1}y =x$ then
    203 \begin{align}
    204     \frac{\|A^{-1}y\|_2}{\|y\|_2} &= \frac{\|x\|_2}{\|Ax\|_2}\\
    205     \Rightarrow \|A^{-1}\|_2 &= \max_{x\neq 0} \frac{\|x\|_2}{\|Ax\|_2}.
    206 \end{align}
    207 \subsubsection{}
    208 Next we show that $\|A^{-1}\|_2 \le 1$.
    209 \begin{align}
    210     \|A^{-1}\|_2 = \max_{x\neq 0} \frac{\|x\|_2}{\|Ax\|_2} =
    211     \frac{1}{\sqrt{1+\|B\|_2^2}} \le 1
    212 \end{align}
    213 \subsection{Problem 5}
    214 We take $A$ and $B$ from last exercise.
    215 \subsubsection{}
    216 We show that for $k \in \{1,\ldots,m\}$ and $\mathcal{W} \subset
    217 \mathbb{R}^n$ with $\dim \mathcal{W} = k$ spanned by $w_1,\ldots,w_k \in
    218 \mathbb{R}^n$. We show that if $x \in \mathcal{W}$ is such that
    219 \begin{align}
    220     \langle Ax, w\rangle = \langle b , w \rangle \qquad \forall w\in
    221     \mathcal{W}.
    222 \end{align}
    223 then $\|x\|_2 \le \|b\|_2$. We can choose $b = x$ then
    224 \begin{align}
    225     \langle Ax, x\rangle &= \|x\|_2^2 = \langle b, x\rangle \leq
    226     \|b\|_2\|x\|_2.\\
    227     \Rightarrow \|x\|_2 \le \|b\|_2
    228 \end{align}
    229 \subsubsection{}
    230 Next we show that for $x :=\sum_j c_j w_j$ and
    231 \begin{align}
    232     \sum_j c_j \langle Aw_j, w_i\rangle = \langle b , w_i\rangle
    233 \end{align}
    234 for $i = 1, \ldots , k$, we have a unique solution for $x \in \mathcal{W}$.
    235 We do this by showing that every solution is the 0 solution $b=0$
    236 \begin{align}
    237     \langle 0 , w\rangle = 0 = \langle Ax ,x \rangle = \|x\|_2^2.
    238 \end{align}
    239 \subsubsection{}
    240 Lastly for $x^* := A^{-1}b$ we show and inequality relation
    241 \begin{align}
    242     \|x^* - x\|_2 \le \|A\|_2 \min_{x \in \mathcal{W}}\|x^* -w\|_2.
    243 \end{align}
    244 (We take the minimum $w = x$ and calculate)
    245 \begin{align}
    246     \|x^* - x\|_2^2
    247     &= \langle A(x^* -x), x^* - x\rangle\\
    248     &= \langle A(x^* -x), x^* - x +w -w\rangle\\
    249     &= \langle A(x^* -x), x^* - w\rangle + \langle A(x^* -x), w - x\rangle\\
    250     &= \langle A(x^* -x), x^* - w\rangle +\underbrace{\langle A(x^* -x), w -
    251     x\rangle}_{=0}\\
    252     \le \|A\|_2 \|x^* - x\|_2\|x^* - w\|.
    253 \end{align}
    254 When we take the minimum and divide by $\|x^* - x\|_2$ we get
    255 \begin{align}
    256     \|x^* - x\| \le \|A\|_2\min_{x\in\mathcal{W}} \|x^* - w\|_2.
    257 \end{align}
    258 
    259 %\printbibliography
    260 \end{document}