Let $\mathbb G(q,l)$ to be the set of all floating-point numbers in the base $q$ and with a mantissa of a length $l$. In the lecture notes I am trying to learn from, the following is said to be obvious (and consequently has no proof):

$$ \forall x \in \mathbb G(q,l), \forall y \in \mathbb G(q,l)\setminus\{x\}: |x-y| \ge |x|q^{-l} $$

What I have tried so far:

Let $x=(-1)^m a \cdot q^\alpha$ and $y=(-1)^n b \cdot q^\beta \in \mathbb G(q,l)$ such that $x \ne y$, then

\begin{align} |x-y| &\ge |x|-|y| \\ &=|(-1)^m a \cdot q^\alpha|-|(-1)^n b \cdot q^\beta| \\ &=a \cdot q^\alpha - b \cdot q^\beta \\ &\ge ...? \end{align}

  • 1
    $\begingroup$ It's certainly not true if $y=x$. Presumably they mean $y \ne x$. $\endgroup$ – Robert Israel Dec 24 '12 at 0:24
  • 1
    $\begingroup$ The first line $|x-y|\geq |x|+|y|$ is incorrect. The inequality goes the other way. $\endgroup$ – Alex Becker Dec 24 '12 at 0:30
  • $\begingroup$ Thank you, Robert and Alex, I've fixed the above-mentioned. $\endgroup$ – Leo Dec 24 '12 at 0:35

Since it's true for $x=0$, we may assume $x \ne 0$. By symmetry, assume $x > 0$. Let $x = a \cdot q^\beta$ with $a$ an integer, $q^{\ell-1} \le a \le q^{\ell} - 1$. Let $y_1$ and $y_2$ be the greatest floating-point number $< x$ and the least floating-point number $>x$ respectively. If $q^{\ell-1} < a$, then $y_1 = (a-1) \cdot q^\beta$ and $|x - y_1| = q^\beta$ while $x q^{-\ell} = a \cdot q^{\beta - \ell} < q^{\beta}$. Similarly, if $a < q^{\ell}-1$, $|x - y_2| = q^\beta$. I'll leave it to you to try $|x - y_1|$ with $a = q^{\ell-1}$ and $|x - y_2|$ with $a = q^{\ell}-1$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.