Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: Balance



Dear Ulrich and P1788 members

Ulrich Kulisch wrote:
> 4. Can anybody tell me whether the Hausdorff metric can > be defined for extended intervals?

Good question and I would like other people's views on this.

For subsets A,B of a general metric space X with distance d(x,y), the definition I know is
  D(A,B) = larger of (
    sup over a in A of d(a,B)
    sup over b in B of d(b,A)
    ),
where for x in X and a subset Y of X
  d(x,Y) = inf over y in Y of d(x,y).

So for _arbitrary_ nonempty A,B, D(A,B) is a non-negative real, or +oo.
The usual conventions inf(Empty) = +oo, sup(Empty) = -oo give unambiguous values also when one or both of A,B are empty.

So in a real-based system (taking the usual metric |x-y|)the answer is Yes, and for instance I make D([5,+oo], [8,+oo]) = 3.

In an extended-real-based system, the usual metric does _not extend_ to infinite points, so that to my mind it makes no sense to say, for instance, d(+oo,+oo) = 0. Hence, in my BIAS implementation, I have set D(A,B) = NaN if A and B have a common infinite endpoint, eg.
  D([5,+oo], [8,+oo]) = NaN.

The set R* of extended reals is of course metrisable: one can map it to a finite interval by, say, x -> tanh(x) and thus take Hausdorff distance w.r.t. d(x,y) = |tanh(x) - tanh(y)|, but that is not "natural".

Do people agree with this approach?

John Pryce

-----------------------------------------
Email sent from www.virginmedia.com/email
Virus-checked using McAfee(R) Software and scanned for spam