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

Re: Motion 42: Decoration system, re-revised text



Michel

As usual, thanks for these comments. 
On 2 Jan 2013, at 17:20, Michel Hack wrote:
> Comments on
>   20121231DecoSystemCirculatedA.pdf   with §5 Flavors, §6 Decorations overview,
> 
> I like the notion of "included flavors".
Good.

> Page 12, last bullet in examples for 5.2:  The definition that gives
>    a smaller set of common evaluations is *stronger*, not weaker.
>    At least that's how I understand strong and weak conditions.
Are you sure? Take, say,
   sign([0,3])=[0,1]
which is true in the set-based flavor but at present is not common. If I insist that it be common, that makes makes sign's common evaluation set CE(sign) bigger, and is *more* restrictive, which is what I would call stronger -- e.g. csets can no longer be a flavor, whereas at present I believe they can be. 

> Page 12, Note at the end of 5.3:  "tightest" is NOT the same as the "tight"
>    defined in this context, because "tightest" is a level 2 concept.  Take
>    for example sqrt([1,2]).
You are right. Best to remove the parenthesised comment, how about "Note. Different qualities of Level 2 enclosure are distinguished by the terms tightest, accurate and valid..."?

> Comments on
>   20121231DecoSystemCirculatedB.pdf   with set-based decoration system details.
> 
> Page 34, 2nd line of 2nd para of 8.8.4:  is "and y_emp" supposed to be there?
Typo. Should be y_ill.

> Page 36, 8.8.8:  I would like to add:
> 
>         "com is (should be?) a valid 2nd argument to setDec() even
>         in implementations that do not support the com decoration;
>         the result shall be as if "dac" had been given."
Ah yes, you said that before. I agreed but forgot to include it. Done, saying "com shall be ..."

>     Perhaps clause 6.3 (in part A) should already mention this, e.g.:
> 
>        "All flavors shall support the com decoration, but implementations
>        supporting a single flavor may treat "com" as equivalent to that
>        flavor's strongest decoration when trying to set it, and would
>        never report "com" when extracting the decoration part."
> 
>     If this is done my parenthetical "should be?" would not be needed.

I should like something like this, but it struck me that "com" is not necessarily a flavor's *strongest* (best) decoration. E.g. a flavor might reasonably have a decoration "anl" which says f is real-analytic everywhere on the input box. If you can think of suitable wording I'll consider putting it in, and changing 8.8.8 to match.

> Page 37, 8.8.10:  Assuming my suggestion for 6.3 above is accepted, this
>     paragraph could become:
> 
>         "A single-flavor set-based implementation shall treat "com"
>         as if it were "dac", in accordance with #6.3 in Chapter 1.

Well, not necessarily, since a single-flavor implementation may like to detect overflow, see your next comments.

>         A multi-flavor implementation shall define the "com"
>         decoration as follows:"
> 
> Page 38, 3rd Note:  When com is provided  ->  when com is fully supported
>     (Again, subject to my 6.3. suggestion.)
> 
>     A similar correction would be needed in the penultimate paragraph on
>     this page.
> 
>     It may also be worth pointing out that as soon as a decoration is
>     produced the evaluation ceases to be common (because the implied
>     interval is then either Empty or Entire, neither of which is common).
When the threshold is com, events A = (the evaluation ceases to be common) and B = (a decoration is produced) occur at the same time. When the threshold is < com, A can happen before B. But that seems self-evident from the definitions: is it worth saying?

> Page 39, 2nd half:  Looks good.  (It reflects a private discussion I had
>     had with John a few days ago.)
Good. If anyone sees a flaw, please say so.

John Pryce