The closed-loop dynamics of a linear optimal controller are simple but have interesting properties. From a starting state $\mathbf{v}(0)$ the dynamic can be iterated to reach a final state $\mathbf{v}(N)$ as in

$$ v(n) = (A - BC)^N\mathbf{v}(0). $$

I would like to optimize the transition parameter matrix $A$ through gradient descent with rate $\eta$:

$$ A' = A - \eta\frac{df}{dA} $$

with loss function $f$ is the error between a generic resultant state $(A - BC)^N\mathbf{v}$ and a target state $\mathbf{w}$. $\newcommand\norm[1]{\lVert#1\rVert}$I'm struggling to find the derivative $\frac{\partial{f(A)}}{\partial{A}}$ of this scalar loss function (L2 norm) $$ f(A) = \norm{(A - BC)^N\mathbf{v} - \mathbf{w}}_2^2 $$ where $A\in\mathbb{R}^{n\times n}$, $B\in\mathbb{R}^{n\times m}$, $C\in\mathbb{R}^{m\times n}$, and $\mathbf{v},\mathbf{w}\in\mathbb{R}^{n}$.

Using chain rule, I can write this as $$ \frac{\partial{f(X)}}{\partial{X}} = \frac{\partial{f(X)}}{\partial{g(X)}} \frac{\partial{g(X)}}{\partial{h(X)}} \frac{\partial{h(X)}}{\partial{X}} $$ where \begin{align*} f(X) & {}= g(h(X)) \\ g(X) & {}= \norm{X\mathbf{v} - \mathbf{w}}_2^2 \\ h(X) & {}= (X-BC)^N. \end{align*}

The problem is, I'm not clear on how to take these these intermediate derivatives.

The closed-loop dynamics of a linear optimal controller are simple but have interesting properties. From a starting state $\mathbf{v}(0)$ the dynamic can be iterated to reach a final state $\mathbf{v}(N)$ as in

$$ \mathbf{v}(N) = (A - BC)^N\mathbf{v}(0). $$

I would like to optimize the transition parameter matrix $A$ through gradient descent with rate $\eta$:

$$ A' = A - \eta\frac{df}{dA} $$

with loss function $f$ is the error between a generic resultant state $(A - BC)^N\mathbf{v}$ and a target state $\mathbf{w}$. $\newcommand\norm[1]{\lVert#1\rVert}$I'm struggling to find the derivative $\frac{\partial{f(A)}}{\partial{A}}$ of this scalar loss function (L2 norm) $$ f(A) = \norm{(A - BC)^N\mathbf{v} - \mathbf{w}}_2^2 $$ where $A\in\mathbb{R}^{n\times n}$, $B\in\mathbb{R}^{n\times m}$, $C\in\mathbb{R}^{m\times n}$, and $\mathbf{v},\mathbf{w}\in\mathbb{R}^{n}$.

Using chain rule, I can write this as $$ \frac{\partial{f(X)}}{\partial{X}} = \frac{\partial{f(X)}}{\partial{g(X)}} \frac{\partial{g(X)}}{\partial{h(X)}} \frac{\partial{h(X)}}{\partial{X}} $$ where \begin{align*} f(X) & {}= g(h(X)) \\ g(X) & {}= \norm{X\mathbf{v} - \mathbf{w}}_2^2 \\ h(X) & {}= (X-BC)^N. \end{align*}

The problem is, I'm not clear on how to take these these intermediate derivatives.

$\newcommand\norm[1]{\lVert#1\rVert}$I'm struggling to find the derivative $\frac{\partial{f(A)}}{\partial{A}}$ of a scalar function (L2 norm) $$ f(A) = \norm{(A - BC)^N\mathbf{v} - \mathbf{w}}_2^2 $$ where $A\in\mathbb{R}^{n\times n}$, $B\in\mathbb{R}^{n\times m}$, $C\in\mathbb{R}^{m\times n}$, and $\mathbf{v},\mathbf{w}\in\mathbb{R}^{n}$.

Using chain rule, I can write this as $$ \frac{\partial{f(X)}}{\partial{X}} = \frac{\partial{f(X)}}{\partial{g(X)}} \frac{\partial{g(X)}}{\partial{h(X)}} \frac{\partial{h(X)}}{\partial{X}} $$ where \begin{align*} f(X) & {}= g(h(X)) \\ g(X) & {}= \norm{X\mathbf{v} - \mathbf{w}}_2^2 \\ h(X) & {}= (X-BC)^N. \end{align*}

The problem is, I'm not clear on how to take these these intermediate derivatives. Since the function is scalar, the scalar-by-matrix derivative rules should be clear, but I'm struggling to get started on the matrix-by-matrix, chain rule derivatives.

The closed-loop dynamics of a linear optimal controller are simple but have interesting properties. From a starting state $\mathbf{v}(0)$ the dynamic can be iterated to reach a final state $\mathbf{v}(N)$ as in

$$ v(n) = (A - BC)^N\mathbf{v}(0). $$

I would like to optimize the transition parameter matrix $A$ through gradient descent with rate $\eta$:

$$ A' = A - \eta\frac{df}{dA} $$

with loss function $f$ is the error between a generic resultant state $(A - BC)^N\mathbf{v}$ and a target state $\mathbf{w}$. $\newcommand\norm[1]{\lVert#1\rVert}$I'm struggling to find the derivative $\frac{\partial{f(A)}}{\partial{A}}$ of this scalar loss function (L2 norm) $$ f(A) = \norm{(A - BC)^N\mathbf{v} - \mathbf{w}}_2^2 $$ where $A\in\mathbb{R}^{n\times n}$, $B\in\mathbb{R}^{n\times m}$, $C\in\mathbb{R}^{m\times n}$, and $\mathbf{v},\mathbf{w}\in\mathbb{R}^{n}$.

Using chain rule, I can write this as $$ \frac{\partial{f(X)}}{\partial{X}} = \frac{\partial{f(X)}}{\partial{g(X)}} \frac{\partial{g(X)}}{\partial{h(X)}} \frac{\partial{h(X)}}{\partial{X}} $$ where \begin{align*} f(X) & {}= g(h(X)) \\ g(X) & {}= \norm{X\mathbf{v} - \mathbf{w}}_2^2 \\ h(X) & {}= (X-BC)^N. \end{align*}

The problem is, I'm not clear on how to take these these intermediate derivatives.

$\newcommand\norm[1]{\lVert#1\rVert}$I'm struggling to find the derivative $\frac{\partial{f(A)}}{\partial{A}}$ of a scalar function (L2 norm) $$ f(A) = \norm{(A - BC)^N\mathbf{v} - \mathbf{w}}_2^2 $$ where $A\in\mathbb{R}^{n\times n}$, $B\in\mathbb{R}^{n\times m}$, $C\in\mathbb{R}^{m\times n}$, and $\mathbf{v},\mathbf{w}\in\mathbb{R}^{n}$.

Using chain rule, I can write this as $$ \frac{\partial{f(X)}}{\partial{X}} = \frac{\partial{f(X)}}{\partial{g(X)}} \frac{\partial{g(X)}}{\partial{h(X)}} \frac{\partial{h(X)}}{\partial{X}} $$ where \begin{align*} f(X) & {}= g(h(X)) \\ g(X) & {}= \norm{X\mathbf{v} - \mathbf{w}}_2^2 \\ q(X) & {}= (X-BC)^N. \end{align*}

The problem is, I'm not clear on how to take these these intermediate derivatives. Since the function is scalar, the scalar-by-matrix derivative rules should be clear, but I'm struggling to get started on the matrix-by-matrix, chain rule derivatives.

$\newcommand\norm[1]{\lVert#1\rVert}$I'm struggling to find the derivative $\frac{\partial{f(A)}}{\partial{A}}$ of a scalar function (L2 norm) $$ f(A) = \norm{(A - BC)^N\mathbf{v} - \mathbf{w}}_2^2 $$ where $A\in\mathbb{R}^{n\times n}$, $B\in\mathbb{R}^{n\times m}$, $C\in\mathbb{R}^{m\times n}$, and $\mathbf{v},\mathbf{w}\in\mathbb{R}^{n}$.

Using chain rule, I can write this as $$ \frac{\partial{f(X)}}{\partial{X}} = \frac{\partial{f(X)}}{\partial{g(X)}} \frac{\partial{g(X)}}{\partial{h(X)}} \frac{\partial{h(X)}}{\partial{X}} $$ where \begin{align*} f(X) & {}= g(h(X)) \\ g(X) & {}= \norm{X\mathbf{v} - \mathbf{w}}_2^2 \\ h(X) & {}= (X-BC)^N. \end{align*}

The problem is, I'm not clear on how to take these these intermediate derivatives. Since the function is scalar, the scalar-by-matrix derivative rules should be clear, but I'm struggling to get started on the matrix-by-matrix, chain rule derivatives.