lizfcm/notes/Sep-11.org

Errors

Errors

$x,y \in \mathds{R}$, using y as a way to approximate x. Then the absolute error of in approximating x w/ y is $e_{abs}(x, y) = |x-y|$.

and the relative error is $e_{rel}(x, y) = \frac{|x-y|}{|x|}$

Table of Errors

  (load "../cl/lizfcm.asd")
  (ql:quickload 'lizfcm)

  (defun eabs (x y) (abs (- x y)))
  (defun erel (x y) (/ (abs (- x y)) (abs x)))

  (defparameter *u-v* '(
                        (1    0.99)
                        (1    1.01)
                        (-1.5 -1.2)
                        (100  99.9)
                        (100  99)
                        ))

  (lizfcm.utils:table (:headers '("u" "v" "e_{abs}" "e_{rel}")
                       :domain-order (u v)
                       :domain-values *u-v*)
    (fround (eabs u v) 4)
    (fround (erel u v) 4))

u	v	eabs	erel
1	0.99	0.0	0.0
1	1.01	0.0	0.0
-1.5	-1.2	0.0	0.0
100	99.9	0.0	0.0
100	99	0.0	0.0

Look at $u \approx 0$ then $v \approx 0$, $e_{abs}$ is better error since $e_{rel}$ is high.

Vector spaces & measures

Suppose we want solutions fo a linear system of the form $Ax = b$, and we want to approximate $x$, we need to find a form of "distance" between vectors in $\mathds{R}^n$

Vector Distances

A norm on a vector space $|| v ||$ is a function from $\mathds{R}^n$ such that:

1. $||v|| \geq 0$ for all $v \in \mathds{R}^n$ and $||v|| = \Leftrightarrow v = 0$
2. $||cv|| = |c| ||v||$ for all $c \in \mathds{R}, v \in \mathds{R}^n$
3. $||x + y|| \leq ||x|| + ||y|| \forall x,y \in \mathds{R}^n$

Example norms:

$||v||_2 = || [v_1, v_2, \dots v_n] || = (v_1^2 + v_2^2 + \dots + v_n^2)^{}^{\frac{1}{2}}$

$||v||_1 = |v_1| + |v_2| + \dots + |v_n|$

$||v||_{\infty} = \text{max}(|v_i|)$ (most restriction)

p-norm: $||v||_p = \sum_{i=1}^{h} (|v_i|^p)^{\frac{1}{p}}$

Length

The length of a vector in a given norm is $||v|| \forall v \in \mathds{R}^n$

All norms on finite dimensional vectors are equivalent. Then exist constants $\alpha, \beta > 0 \ni \alpha ||v||_p \leq ||v||_q \leq \beta||v||_p$

Distance

Let $u,v$ be vectors in $\mathds{R}^n$ then the distance is $||u - v||$ by some norm: $e_{abs} = d(v, u) = ||u - v||$

The relative errors is:

$e_{rel} = \frac{||u - v||}{||v||}$

Approxmiating Solutions to $Ax = b$

We define the residual vector $r(x) = b - Ax$

If $x$ is the exact solution, then $r(x) = 0$.

Then we can measure the "correctness" of the approximated solution on the norm of the residual. We want to minimize the norm.

But, $r(y) = b - Ay \approx 0 \nRightarrow y \equiv x$, if $A$ is not invertible.