notes

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

sesh2.tex (18030B)


      1 \include{./preamble.tex}
      2 
      3 \begin{document}
      4 \maketitle
      5 \tableofcontents
      6 \section{Sheet 2}
      7 \subsection{Exercise 7}
      8 For the functions $g:\mathbb{R}^{2}\to \mathbb{R}^{2}$, find $X =
      9 \{(x,y)\in\mathbb{R}^{2}: g(x, y)\leqq 0\} $, the tangent cone and the
     10 linearized tangent cone at $x_0 \in X$ and find out if $x_0$ fulfills
     11 (ADABIDE-CQ), i.e. $T_\text{lin}(x_0) = T_X(x_0)$.
     12 \begin{enumerate}
     13     \item $g(x,y) = (y-x^{3}, -y)^{T},\quad x_0=(0,0)^{T}$
     14     \item $g(x,y) = (y^{2}-x+1, 1-x-y)^{T},\quad x_0=(1,0)^{T}$
     15 \end{enumerate}
     16 For 1. we have that $g(x,y)\leqq 0$ means that
     17 \begin{align}
     18     y-x^{3}\le_0\\
     19     -y \le 0\\
     20     \Rightarrow 0\le y\le x^{3}.
     21 \end{align}
     22 So is defined as $X = \{\left( x,y \right) \in \mathbb{R}^{2}: 0\le y, y\le
     23 x^{3}\}$. Graphically represented $X$ looks like the following
     24 \begin{figure}[H]
     25     \centering
     26     \begin{tikzpicture}[yscale=1, xscale=1]
     27         \begin{axis}[
     28             xmin=-1, xmax=2,
     29             ymin=-1, ymax=2,
     30             axis lines = middle,
     31         ]
     32             \addplot[domain=0:2, samples=100, color=gray, name path=A]{x^3};
     33             \addplot[domain=0:2, samples=100, name path=B]{0};
     34             \addplot[fill=gray,fill opacity=0.5] fill between[of=A and B,soft
     35                 clip={domain=0:2},];
     36             \node[] at (axis cs:2,2) [below left] {X};
     37             \node[draw, circle, inner sep=1pt, fill=red, label=above right:{$x_0$}] at
     38                 (axis cs:0, 0) {};
     39         \end{axis}
     40     \end{tikzpicture}
     41     \label{fig: ex7.1}
     42 \end{figure}
     43 Then the tangent cone is
     44 \begin{align}
     45     T_X(x_0) = \left\{\begin{pmatrix} \lambda\\ 0 \end{pmatrix} : \lambda
     46 \ge 0\right\}.
     47 \end{align}
     48 Now for the linearized tangent cone we calculate, $g_1(x_0) = 0$ and
     49 $g_2(x_0) = 0$ meaning that $\mathcal{A}(x_0)=\{1, 2\}$ thereby
     50 \begin{align}
     51     T_\text{lin}(x_0)
     52     &= \{d\in \mathbb{R}^{2}: \nabla g_1(x_0)^{T}d\le 0,\; \nabla g_2(x_0)^{T}d
     53     \le 0\}\\
     54     &= \left\{d\in \mathbb{R}^{2}: \begin{pmatrix}0\\1\end{pmatrix}^{T}d\le 0,\;
     55         \begin{pmatrix}0\\-1\end{pmatrix}^{T}d
     56     \le 0\right\}\\
     57     &= \left\{\begin{pmatrix} \lambda \\ 0 \end{pmatrix}, \begin{pmatrix} -\lambda
     58 \\ 0 \end{pmatrix}:\; \lambda \ge 0 \right\}.
     59 \end{align}
     60 We conclude that $x_0 = (0, 0)^{T}$ does not satisfy the ADABIE-CQ
     61 condition for this optimization problem.
     62 \newline
     63 For number 2. first the domain $X$, $g(x, y) \le 0$
     64 \begin{align}
     65     y^{2}-x +1 \le 0 \quad &\text{and} \quad 1-x-y\le 0\\
     66     y^2 - 1 \le x \quad &\text{and} \quad y-1\le x
     67 \end{align}
     68 so $X$ has the following form
     69 \begin{align}
     70     X = \left\{ \left( x,y \right) \in \mathbb{R}^{2}:
     71         \begin{cases}
     72             -\sqrt{x+1}\le y\le x+1
     73             \quad  \text{for}\; x\in(-1, 0]\\
     74             -\sqrt{x+1}\le y \le\sqrt{x+1}\quad \text{for}\; x > 0
     75         \end{cases}
     76     \right\}
     77 \end{align}
     78 and graphically
     79 \begin{figure}[H]
     80     \centering
     81     \begin{tikzpicture}[yscale=1, xscale=1]
     82         \begin{axis}[
     83             xmin=-1.5, xmax=2,
     84             ymin=-1.5, ymax=1.5,
     85             axis lines = middle,
     86         ]
     87             \addplot[domain=-1:0, samples=100, color=gray, name path=A]{x+1};
     88             \addplot[domain=-1:2, samples=100, color=gray, name
     89                 path=B]{-sqrt(x+1)};
     90             \addplot[domain=0:2, samples=100, color=gray, name
     91                 path=C]{sqrt(x+1)};
     92             \addplot[domain=-1:0, samples=100, color=gray, name
     93                 path=L]{0};
     94             \addplot[fill=gray,fill opacity=0.5] fill between[of=A and B,soft
     95                 clip={domain=-1:0},];
     96             \addplot[fill=gray,fill opacity=0.5] fill between[of=C and B,soft
     97                 clip={domain=0:2},];
     98             \node[] at (axis cs:2,1.5) [below left] {X};
     99             \node[draw, circle, inner sep=1pt, fill=red, label={$x_0$}] at (axis cs:1, 0) {};
    100         \end{axis}
    101     \end{tikzpicture}
    102     \label{fig: ex7.2}
    103 \end{figure}
    104 Then the tangent cone is obviously
    105 \begin{align}
    106     T_X(x_0) = \left\{ d: d \in \mathbb{R}^{2} \right\}
    107 \end{align}
    108 For the linearized tangent cone we calculate  $g_1(x_0) = 0$ and $g_2(x_0)=0$,
    109 thereby $\mathcal{A}(x_0)=\{1, 2\}$ and
    110 \begin{align}
    111     T_\text{lin}(x_0)&=
    112     \left\{ d\in\mathbb{R}^{2}: \begin{pmatrix} -1\\0 \end{pmatrix}^{T}d\le 0,\;
    113     \begin{pmatrix} -1 \\ -1 \end{pmatrix}^{T}d\le 0 \right\} \\
    114          &= \left\{ \begin{pmatrix} \lambda \\ \lambda \end{pmatrix},
    115          \begin{pmatrix} \lambda \\ -\lambda \end{pmatrix}:\; \lambda \ge
    116      0\right\}.
    117 \end{align}
    118 In this case $x_0$ also does not satisfy the ABADIE-CQ.
    119 \subsection{Exercise 8}
    120 Let $(x^{*}, \lambda^{*}, \mu^{*})$ be a KKT point of the optimization
    121 problem
    122 \begin{align}
    123     \text{min}\quad & f(x),\\
    124     \text{s.t.}\quad & g_i(x) \le 0, i=1,\ldots,m\nonumber\\
    125     &h_j(x) = 0, j=1,\ldots,p\nonumber\\
    126     & x \in \mathbb{R}^{n}\nonumber
    127 \end{align}
    128 for $f, g_i, h_i:\mathbb{R}^{n}\to \mathbb{R}$ continuously differentiable
    129 functions. Prove that $x^{*}$ is a critical point of the optimization point,
    130 namely that it holds
    131 \begin{align}
    132     \nabla f(x^{*})^{T}\ge 0 \;\; \forall d\in T_X(x^{*}),
    133 \end{align}
    134 where $X = \left\{ x \in R^{n}: g_i(x) \le 0, i=1,\ldots,m, h_j(x) = 0,
    135 j=1,\ldots,p\right\}$. Given a critical point $x^{*}$ when do Lagrange
    136 multipliers $\lambda^{*}, \mu^{*}$ exist such that $(x^{*}, \lambda^{*},
    137 \mu^{*})$ is a KKT point?
    138 \newline
    139 Firs of all if $(x^{*}, \lambda^{*}, \mu^{*}) $ is a KKT point then
    140 \begin{align}
    141     &\nabla_x L(x^{*},\lambda^{*}, \mu^{*}) = 0\\
    142     &\nabla f(x^{*}) + \sum_{i=1}^{m} \lambda_i^{*} \nabla g_i(x^{*})
    143     + \sum_{j=1}^{p} \mu_j^{*} \nabla h_j(x^{*}) = 0
    144 \end{align}
    145 is satisfied for the Lagrangian. Then we can take the scalar product with $d
    146 \in T_X(x^{*})$. We know that $\nabla g_i(x^{*})^{T} d \le 0$ and $\nabla h_j
    147 (x^{*})^{T}d =0$ for all $i = 1,\ldots,m$ and $j=1,\ldots,p$ and
    148 $\lambda_i^{*}\ge 0$ which means
    149 \begin{align}
    150     0 &= \nabla f(x^{*})^{T}d + \sum_{i=1}^{m} \lambda_i^{*} \nabla g_i(x^{*})^{T}d
    151     + \sum_{j=1}^{p} \mu^{*}_j \nabla h_j(x^{*})^{T}d \\
    152     &= \nabla f(x^{*})^{T}d + \sum_{i=1}^{m} \lambda_i^{*} \nabla
    153     g_i(x^{*})^{T}d\\
    154     &\le \nabla f(x^{*})^{T}d.
    155 \end{align}
    156 This concludes
    157 \begin{align}
    158     \nabla f(x^{*})^{T}d \ge 0.
    159 \end{align}
    160 Now if $x^{*}$ is a critical point then it is a local minimum. If it
    161 fulfills the ABADIE-CQ condition then there exist $\lambda^{*} \in
    162 \mathbb{R}^{m}$ and $\mu^{*}\in\mathbb{R}^{p}$ such that $(x^{*},
    163 \lambda^{*}, \mu^{*})$ is a KKT point. We know that $X$ is convex and $x^{*}$
    164 fulfills the ABADIE-CQ then $\nabla f(x^{*}) \in \left( T_X(x^{*})^{*}
    165 \right)$ and $\left( T_X(x^{(*)} \right)^{*} = (T_\text{lin}(x^{*}))^{*}$.
    166 This means that $\nabla f(x^{*}) \in \left( T_\text{lin}(x^{*} \right)^{*}$.
    167 By Farkas Lemma there exist $\lambda_i^{*} \ge 0$ and $\mu_j^{*}$,
    168 $i=1,\ldots,m$, $j=1,\ldots,p$ such that $\nabla_x L(x^{*}, \lambda^{*},
    169 \mu^{*}) = 0$, then $(x^{*}, \lambda^{*}, \mu^{*})$ is a KKT point.
    170 \subsection{Exercise 9}
    171 Consider the optimization problem
    172 \begin{align}
    173     \text{min}\quad & x_1^{2}\left( x_2 + 1 \right)^{2} ,\\
    174     \text{s.t.}\quad &x_1^{3} - x_2 \le 0\\
    175     & x_2 \le 0.
    176 \end{align}
    177 Show that $x^{*} = (0, 0)^{T}$ fulfills ABADIE-CQ but not MFCQ.
    178 \newline
    179 The domain $X$ is defined by $x_1^2 \ge x^2$ and $x_2 \ge 0$,
    180 \begin{align}
    181     X = \left\{ (x_1, x_2) \in \mathbb{R}^{2}: x_1^{2}\ge x_2 \ge 0 \right\},
    182 \end{align}
    183 graphically
    184 \begin{figure}[H]
    185     \centering
    186     \begin{tikzpicture}[yscale=1, xscale=1]
    187         \begin{axis}[
    188             xmin=-2, xmax=2,
    189             ymin=-0.5, ymax=2,
    190             axis lines = middle,
    191             xlabel=$x_1$,
    192             ylabel=$x_2$,
    193         ]
    194             \addplot[domain=-2:2, samples=100, color=gray, name path=A]{x^2};
    195             \addplot[domain=-2:2, samples=100, name path=B]{0};
    196             \addplot[fill=gray,fill opacity=0.5] fill between[of=A and B,soft
    197                 clip={domain=-2:2},];
    198             \node[] at (axis cs:2,2) [below left] {X};
    199             \node[draw, circle, inner sep=1pt, fill=red, label=above
    200                 right:{$x^{*}$}] at
    201                 (axis cs:0, 0) {};
    202         \end{axis}
    203     \end{tikzpicture}
    204     \label{fig: ex9}
    205 \end{figure}
    206 meaning that
    207 \begin{align}
    208     T_X(x^{*}) = \left\{\begin{pmatrix} -\lambda \\ 0 \end{pmatrix},
    209     \begin{pmatrix} \lambda \\ 0 \end{pmatrix}:\; \lambda \ge 0   \right\},
    210 \end{align}
    211 Then
    212 \begin{align}
    213     T_\text{lin}(x^{*})
    214     &= \left\{d\in\mathbb{R}^{2}: \begin{pmatrix} 0\\1
    215     \end{pmatrix}^{T}d \le 0,\; \begin{pmatrix} 0 \\ -1
    216 \end{pmatrix}^{T}d\le 0  \right\} \\
    217     &= \left\{ \begin{pmatrix} -\lambda \\ 0 \end{pmatrix}, \begin{pmatrix}
    218 \lambda , 0 \end{pmatrix}:\; \lambda \ge 0   \right\}.
    219 \end{align}
    220 This means that $x^{*}$ fulfills the ABADIE-CQ condition. On the other hand
    221 MFCQ is fulfilled only if there exists $d\in\mathbb{R}^{2}$ such that $\nabla
    222 g_i(x^{*})^{T}d < 0$, for all $i\in \mathcal{A}(x^{*})$ but the problem is the
    223 strict constraint
    224 \begin{align}
    225     \nabla g_1 (x^{*}) = \begin{pmatrix} 0 \\ 1 \end{pmatrix}, \quad
    226     \nabla g_2 (x^{*}) = \begin{pmatrix} 0 \\ -1 \end{pmatrix}.
    227 \end{align}
    228 Any feasible solutions are of the form $(\pm \lambda , 0)^{T}$, $\lambda \ge
    229 0$. Both cases always equal to $0$.
    230 \subsection{Exercise 10}
    231 Consider the optimization problem
    232 \begin{align}
    233     \text{min}\quad & x_1^{2}\left( x_2 + 1 \right)^{2} ,\\
    234     \text{s.t.}\quad &-x_1^{3} - x_2 \le 0\\
    235     & -x_2 \le 0.
    236 \end{align}
    237 Show that $x^{*} = (0, 0)^{T}$ fulfills MFCQ but not LICQ.
    238 \newline
    239 The domain $X$ is defined by $x_1^2 \ge -x^2$ and $x_2 \ge 0$,
    240 \begin{align}
    241     X = \left\{ (x_1, x_2) \in \mathbb{R}^{2}: x_2 \ge 0 \right\},
    242 \end{align}
    243 and $g_1(x^{*}) = 0$ and $g_2(x^{*}) = 0$ so $\mathcal{A}(x^{*}) = \{1,2\}$.
    244 \begin{align}
    245     \nabla g_1(x^{*}) = \begin{pmatrix} 0\\-1 \end{pmatrix},\quad \nabla
    246     g_2(x^{*}) = \begin{pmatrix} 0 \\ -1 \end{pmatrix}
    247 \end{align}
    248 For strict inequality $\nabla g_i(x^{*})^{T}d < =$ for all $i \in
    249 \mathcal{A}(x^{*})$ we have that $d = (0, \lambda)$ with $\lambda > 0$. This
    250 means $x_0$ fulfills MFCQ. On the other hand LICQ is fulfilled if
    251 \begin{align}
    252     \{\nabla g_i(x^{*})\}_{i\in\mathcal{A}(x^{*})}
    253 \end{align}
    254 are linearly independent. But in our case $\nabla g_1(x^{*}) = \nabla
    255 g_2(x^{*})$, meaning that $x_0$ does not fulfill LICQ.
    256 \subsection{Exercise 11}
    257 Let $U \subseteq \mathbb{R}^{n}$ be a nonempty, open convex set and $f \in U
    258 \to \mathbb{R}$ a differentiable function on $U$. Prove that the following
    259 statements are equivalent.
    260 \begin{enumerate}
    261     \item $f$ is convex on $U$
    262     \item $\langle\nabla f(x),y-x\rangle \le f(y) - f(x) \quad \forall x, y \in U$
    263     \item $\langle\nabla f(x)-\nabla f(y),y-x\rangle \le 0 \quad \forall x, y \in U$
    264     \item if f is twice differentiable on $U$, then $\nabla^{2}f(x)$ is
    265         positively semi definite for every $x \in U$.
    266 \end{enumerate}
    267 We start with (1) $\Leftrightarrow$ (2).\newline
    268 Ad $\Rightarrow$: $f$ is convex, then for all $x, y \in U$, $\lambda \in [0,
    269 1]$ we have
    270 \begin{align}
    271     f\left( (1-\lambda)x + \lambda y \right)
    272     &\le (1-\lambda)f(x) + \lambda(y)\\
    273     &=f(x) + \lambda\left( f(y) - f(x) \right)\\
    274     \frac{f\left( (1-\lambda)x + \lambda y \right)  -f(x)}{\lambda}
    275     &\le f(y) - f(x).
    276 \end{align}
    277 Letting $\lambda \downarrow 0 $ we get
    278 \begin{align}
    279     \nabla f(x)^{T}(y-x) \le f(x) - f(y)
    280 \end{align}
    281 Ad $\Leftarrow$: we have that $\forall x, y \in U$:
    282 \begin{align}
    283     \nabla f(x)^{T}(y-x) \le f(x) - f(y).
    284 \end{align}
    285 Since $U$ is convex then the above also holds for $z \in U$ where $z =
    286 (1-\lambda)x + \lambda y$, then
    287 \begin{align}
    288     &f(x) \ge f(z) + \nabla f(z)^{T}(x-z) \quad | \cdot (1-\lambda)\\
    289     &f(y) \ge f(z) + \nabla f(z)^{T}(y-z) \quad | \cdot \lambda
    290 \end{align}
    291 adding both of them together we get
    292 \begin{align}
    293     (1-\lambda)f(x) + \lambda f(y)
    294     &\ge f(z) + \nabla f(z)^{T}((1-\lambda) x + \lambda y - z) \\
    295     &= f(z)\\
    296     &= f((1-\lambda)x + \lambda y).
    297 \end{align}
    298 This shows that $f$ is convex on $U$.
    299 \newline
    300 Next we show (2) $\Leftrightarrow$ (3).\newline
    301 Ad $\Rightarrow$:  We start with
    302 \begin{align}
    303     &f(y) \ge f(x) + \nabla f(x)^{T}(y-x)\\
    304     &f(x) \ge f(y) + \nabla f(y)^{T}(x-y).
    305 \end{align}
    306 Adding them together we get
    307 \begin{align}
    308     \nabla f(y)^{T}(y-x) - \nabla f(x)^{T}(y-x) \ge 0\\
    309     \left( \nabla f(y)^{T} - \nabla f(x)^{T} \right) (y-x) \ge 0.
    310 \end{align}
    311 Ad $\Leftarrow$: We can just do the same operations as in $\Rightarrow$ in
    312 reverse.
    313 \newline
    314 Now we prove (2) $\Leftrightarrow$ (4).First we consider in one dimension and
    315 then generalize
    316 \newline
    317 Ad $\Rightarrow$: . In
    318 $U \subseteq \mathbb{R}$ we have that $\forall x ,y \in U$
    319 \begin{align}
    320     &f(y) \ge f(x) + f(x)'(y-x)\\
    321     &f(x) \ge f(y) + f(y)'(x-y).
    322 \end{align}
    323 Let $x < y$, then
    324 \begin{align}
    325    &f'(x)(y-x) \le f(y) - f(x) \le f'(y) (y- x) \quad | \frac{1}{(y-x)^{2}}\\
    326    &\frac{f'(y) - f'(x)}{y-x} \ge 0 \quad | y\to x\\
    327    &f''(x) \ge 0 \quad \forall x \in U.
    328 \end{align}
    329 Ad $\Leftarrow$: We use Taylors expansion formula for $f(y)$ in $x \in U$
    330 \begin{align}
    331     f(y) = f(x) + f'(x)(y-x) + \frac{1}{2}f''(\xi) (y-x)\quad \xi \in [x,y]\\
    332     f(y) \ge f(x) + f'(x)(y-x).
    333 \end{align}
    334 In general dimensions convexivity means convexivity along all directions,
    335 i.e. $f : U\subseteq \mathbb{R}^{n} \to \mathbb{R}$ is convex if
    336 \begin{align}
    337     g(\alpha) = f(x + \alpha d)
    338 \end{align}
    339 is convex $\forall x \in U$ and $\forall d \in \mathbb{R}^{n}$. This is
    340 exactly the case if
    341 \begin{align}
    342     g''(\alpha) = d^{T}\nabla^{2}f(x+\alpha d) d \ge 0 \quad \forall x \in
    343     U\; \forall d \in \mathbb{R}^{n} \; \forall \alpha \in \mathbb{R}
    344 \end{align}
    345 such that $x + \alpha d \in U$ so $f$ is convex if and only if
    346 \begin{align}
    347     \nabla f(x) \ge 0 \quad \forall x \in U \qed
    348 \end{align}
    349 \subsection{Exercise 12}
    350 Let $c: \mathbb{R}\to \mathbb{R}$ be defined as
    351 \begin{align}
    352     c(y) =
    353     \begin{cases}
    354         (y+1)^{2} \qquad &y < -1\\
    355         0 \qquad &-1 \le y \le 1\\
    356         (y-1)^{2} \qquad &y > 1
    357     \end{cases}
    358 \end{align}
    359 Let $g_1, g_2: \mathbb{R}^{2}\to \mathbb{R}$
    360 \begin{align}
    361     g_1(x_1, x_2) = c(x_1) - x_2\\
    362     g_2(x_1, x_2) = c(x_1) + x_2\\
    363 \end{align}
    364 Let $f: \mathbb{R}^{2} \to \mathbb{R}$ be a convex function and continuously
    365 differentiable. Show that for the convex optimization problem
    366 \begin{align}
    367     \text{min}\quad & f(x),\\
    368     \text{s.t.}\quad & g_i(x) \le 0, i=1,2\nonumber\\
    369     & x \in \mathbb{R}^{2}\nonumber
    370 \end{align}
    371 ABADIE-CQ holds at $x^{*}= (0, 0)^{T}$ SLATER-CQ is not satisfied.
    372 \newline
    373 Bellow is a graphical representation of, $c(x_1)$, $g_1(x) \le 0$ and
    374 $g_2(x)$
    375 \begin{figure}[H]
    376     \centering
    377     \begin{tikzpicture}[yscale=1, xscale=1]
    378         \begin{axis}[
    379             xmin=-5, xmax=5,
    380             ymin=-1, ymax=5,
    381             axis lines = middle,
    382             xlabel=$x_1$,
    383             ylabel=$x_2$,
    384         ]
    385             \addplot[domain=1:5, samples=100, color=red, name
    386                 path=A]{(x-1)^2};
    387             \addplot[domain=-1:-5, samples=100, name path=B, color=red]{(x+1)^2};
    388 
    389             \addplot[domain=-5:5, samples=100, name path=D,, color=gray,
    390                 opacity=0]{0};
    391             \addplot[domain=-5:5, samples=100, name path=E,, color=gray,
    392                 opacity=0]{-1};
    393 
    394             \addplot[domain=-5:5, samples=100, name path=F,, color=blue,
    395                 opacity=0]{5};
    396 
    397             \addplot[fill=gray,fill opacity=0.3] fill between[of=D and E,soft
    398                 clip={domain=-5:5},];
    399             \addplot[fill=gray,fill opacity=0.3] fill between[of=B and D,soft
    400                 clip={domain=-5:-1},];
    401             \addplot[fill=gray,fill opacity=0.3] fill between[of=A and D,soft
    402                 clip={domain=1:5},];
    403 
    404             \addplot[fill=blue,fill opacity=0.3] fill between[of=A and F,soft
    405                 clip={domain=1:5},];
    406             \addplot[fill=blue,fill opacity=0.3] fill between[of=B and F,soft
    407                 clip={domain=-1:-5},];
    408             \addplot[fill=blue,fill opacity=0.3] fill between[of=D and F,soft
    409                 clip={domain=-1:1},];
    410 
    411             \addplot[domain=-1:1, samples=100, name path=C, color=red]{0};
    412 
    413             \node[color=red] at (axis cs:3.5,4) [above right] {$c(x_1)$};
    414             \node[color=gray] at (axis cs:2.5,2) [below right] {$g_2(x) \le 0$};
    415             \node[color=blue] at (axis cs:-1.3,3.5) [] {$g_1(x) \le 0$};
    416             \node[draw, circle, inner sep=1pt, fill=red, label=above
    417                 right:{$x^{*}$}] at
    418                 (axis cs:0, 0) {};
    419         \end{axis}
    420     \end{tikzpicture}
    421     \label{fig: ex12}
    422 \end{figure}
    423 So $X$ has only elements on the curve $c(x)$, i.e. $X = \{x \in
    424 \mathbb{R}^{2}: g_1(x) \le 0, g_2(x) \le 0\}  = \{(x_1, c(x_1))^{T} : x_1 \in
    425 \mathbb{R}\}$ and thereby the tangent cone of $X$ at $x^{*}$ consists of
    426 tangent vectors of $c(x)$ at $x^{*}$
    427 \begin{align}
    428     T_X(x^{*}) = \left\{\begin{pmatrix} \lambda\\ 0 \end{pmatrix}, \begin{pmatrix}
    429 -\lambda\\ 0\end{pmatrix}   : \lambda \ge 0 \right\}.
    430 \end{align}
    431 For the linearized tangent cone we have that $g_1(x^{*}) = c(0) = 0 $ and
    432 $g_2(x^{*})= c(0) = 0$, then the gradients at $x^{*}$ are
    433 \begin{align}
    434     \nabla g_1(x^{*}) = \begin{pmatrix} 0\\1 \end{pmatrix}, \qquad
    435     \nabla g_2(x^{*}) = \begin{pmatrix} 0\\-1 \end{pmatrix}.
    436 \end{align}
    437 Thereby
    438 \begin{align}
    439     T_\text{lin}(x^{*})
    440     &= \left\{ d \in \mathbb{R}^{2}: \nabla g_1(x)^{T}d \le 0, \nabla
    441     g_2(x)^{T}d\le 0 \right\} \\
    442     &= \left\{ \begin{pmatrix} \lambda\\0 \end{pmatrix}, \begin{pmatrix}
    443 -\lambda \\ 0 \end{pmatrix} : \lambda \ge 0  \right\}.
    444 \end{align}
    445 We have that $x^{*}$ satisfies ABADIE-CQ.
    446 \newline
    447 In our case SLATER-CQ is fulfilled if there exists an $x' \in \mathbb{R}^{2}$
    448 such that $g_i(x') < 0$ for all $i = 1,2$. The problem arises because in case
    449 of strict inequality the domains defined by $g_1(x) < 0$ and $g_2(x) < 0$ do
    450 not match for any $x$ as seen the figure above. In the relaxed case they
    451 match exactly at the line $c(x_1)$. But $c(x_1) \ge 0$. Meaning that there
    452 exists no $x'$ such that SLATER-CQ is satisfied (in our case).
    453 \end{document}