Prove Optimal Line Derivation

31 Views Asked by At

In class we learned how to derive the optimal line approximated by given points.

There are different ways to approach. We looked at it in a more algebraic way, using matrices and I'm having trouble doing a derivation:

Given some points $\{x_1, \cdots, x_n\} \subset \mathbb{R}^{m}$.

$$ \tag 1 (\mathbf{u},\boldsymbol{\mu}) \leftarrow arg min \left[ \frac{1}{n} \sum_{i=1}^{n} || \underbrace{ \boldsymbol{\mu} + \langle \mathbf{x}_i - \boldsymbol{\mu}, \mathbf{u} \rangle \mathbf{u}}_{\mathbf{\widehat{x}}_i} - \mathbf{x}_i ||^2 \right] $$ $$ \tag 2 = \left[ \frac{1}{n} \sum_{i=1}^{n} || \left( \mathbf{I} - \mathbf{u} \mathbf{u}^{T} \right) (\mathbf{x}_i - \boldsymbol{\mu})||^2 \right] $$

Where $\boldsymbol{\mu}$ and $\mathbf{u}$ are line variables and $\mathbf{\widehat{x}}_i$ is the projected point of $\mathbf{x}_i$ onto the line.

So we are trying to find the line that minimizes the distances from all the points. What I don't understand is how we can manage to derive $(1)$ into $(2)$.

The hint is to use the identity $\langle \mathbf{v}, \mathbf{u} \rangle \mathbf{u} = (\mathbf{u}\mathbf{u}^{T})\mathbf{v}$.

Help is appreciated.

1

There are 1 best solutions below

0
On

$$ \boldsymbol{\mu} + \langle \mathbf{x}_i - \boldsymbol{\mu}, \mathbf{u} \rangle \mathbf{u} - \mathbf{x}_i = \langle \mathbf{x}_i - \boldsymbol{\mu}, \mathbf{u} \rangle \mathbf{u} - (\mathbf{x}_i-\boldsymbol{\mu})= (\mathbf{u}\mathbf{u}^T)(\mathbf{x}_i - \boldsymbol{\mu}) - \mathbf{I} (\mathbf{x}_i-\boldsymbol{\mu}) =(\mathbf{u}\mathbf{u}^T-\mathbf{I})(\mathbf{x}_i - \boldsymbol{\mu}). $$