Defining a Neighborhood Around a Point Whose Elements Have Different Units

61 Views Asked by At

Suppose I have some n-dimensional point

$ \mathbf{x} = [x_1,...,x_n] $

where some of the $x_i$ have different units (for example, $x_1$ in Ohms and $x_5$ in seconds). Now if all the elements had the same units I could define a ball $B$ of radius $r$ around $\mathbf{x}$ by

$ B_r(\mathbf{x}) = \{ \mathbf{y} : ||\mathbf{x}-\mathbf{y}|| \leq r\}$

But because the elements of the point do not have the same units, the norm $||\mathbf{x}-\mathbf{y}||$ is not meaningful if it is taken to be the Euclidean norm or something similar.

How could one define a meaningful distance in this space in order to define a ball of a particular radius around a point? My first approach is to consider that each element $x_i$ is bounded by some interval with a minimum $m_i$ and a maximum $M_i$ such that

$x_i \in [m_i, M_i]$

and then define a radius for each element by

$r^*_i = \frac{1}{(M_i-m_i)}r$

Then I could define a "ball" of points around the point in question by

$ B_r(\mathbf{x}) = \{ \mathbf{y} : |x_i - y_i| \leq r^*_i\}$

But I am not sure that this is meaningful. Ultimately I am trying to define a ball of parameters for a dynamical model around a particular parameter, and the model is to be evaluated at each point of this ball (a "parameter hypersphere" instead of a parameter grid). Any help here is appreciated.

1

There are 1 best solutions below

1
On BEST ANSWER

Any metric formed in the way you speak of will be topologically equivalent to the standard Euclidean one whose coordinates will be your 'meaningless' Ohms-seconds-etc. This is because in finite dimensions over $\mathbb{R}$ the normal Euclidean metric is equivalent to the 'taxi-cab metric' which considers the sums of the distances in each variable. Scaling each one-dimensional distance by the (presumed non-zero) factors $(M_i - m_i)^{-1}$ will again give you an equivalent metric. This is a basic exercise in topology: You can easily show that for any point, in either metric you can find neighborhoods (e.g. n-dimensional rectangular prisms centered around the point with side lengths converging to zero) that are bases for the different topologies but which are nested in each other.