The least squared solution of a problem is one such that minimizes the least square error of the problem.
For a given [[Set|set]] of [[Point|points]] $\set{(x_{1},y_{1}),\dots,(x_{n},y_{n})}$ some function $f(x)$ is trying to interpolate through, the least squared [[Numerical Analysis Error Bounds|error]] is calculated as:
$\huge
\epsilon\set{f} =
\sum_{i=1}^{n}{
(f(x_{i}) - y_{i})^{2}
}
$
>[!note]
>The most common type of function used with this method is a [[Linear Transformation|Linear]] [[Function]] $f(x)=a+bx$
>
>$\huge
>\epsilon(a,b) =
>\sum_{i=1}^{n}{
>(f(a,b;x_{i}) - y_{i})^{2}
>}
>$
> More information about [[Linear Transformation|Linear]] systems here: [[Overdetermined Linear System]].
This form of solution is desirable for many reasons - one such is that the points we are trying to minimize error around don't need to constitute a [[Function|Function]] (which would break [[Interpolation]] methods), and we can have multiple points with equal $x$-values.
To find a solution from this is a type of [[Optimization]] problem where we want to minimize the [[Numerical Analysis Error Bounds|error]] [[Function|function]].
One possible way one can do this is through [[Gradient Descent]] using the [[Gradient]] of $\epsilon$.
However in the case of when we opearte [[Least Squared Solutions]] on [[Linear Combination|linear combinations]], we can explicitly solve this problem:
$\huge \begin{align}
\epsilon &= \sum_{i=1}^{n}(a+bx_{i}-y_{i})^{2} \\
\pderiv{\epsilon}{a} &= \sum_{i=1}^{n}{
2(a+bx_{i}-y_{i}) \\
} \\
&= 2(na + b \sigma_{x}- \sigma_{y}) \\
\pderiv{\epsilon}{b} &= \sum_{i=1}^{n}2x_{i}(a+bx_{i}-y_{i}) \\
&= 2(a \sigma_{x} + b\sigma_{x x} - \sigma_{xy})
\end{align}$
$\huge
\begin{cases}
na &+ b\sigma_{x} &- \sigma_{y} &= 0 \\
a \sigma_{x} &+ b \sigma_{x x} &- \sigma_{x y} &= 0
\end{cases}
$
$\huge \begin{align}
\mat{n & \sigma_{x}\\\sigma_{x} & \sigma_{x x}}\mat{a \\b} &= \mat{
\sigma_{y} \\ \sigma _{x y}
} \\
\mat{a\\b}
&= \mat{n & \sigma_{x}\\\sigma_{x} & \sigma_{x x}}^{-1}
\mat{\sigma_{y}\\\sigma_{x y}} \\
&=
\frac{1}{n\sigma_{x x}- \sigma_{x}^{2}}\mat{
\sigma_{x x}\sigma_{y} - \sigma_{x} \sigma_{x y } \\
- \sigma_{x} \sigma_{y} + n\sigma_{x y }
}
\end{align}$
Note that the coefficient is actually the [[Variance]] of $x$. $\op{Var}(x_{i})= \frac{1}{n \sigma_{x x}- \sigma_{x}^{2}}$