Re: [office-formula] Semantics

[Daniel Carrera <daniel.carrera@zmsl.com>]

David A. Wheeler wrote:
> Okay, here's an email specifically about semantics; please reply to this 
> email to discuss semantics.  That way, those of us with threading email
> clients can follow the discussion. (Those of you without a threading
> client... good luck :-) ).

Well, you started your email by replying to a previous one, so 
Thunderbird shows it as a sub-thread of an existing, large, thread.

> The real test should be, "will typical spreadsheets with formulas 
> interchange correctly?"

I think that the real test should be, "will *compliant* spreadsheets 
interchange correctly?".

If I give you a compliant spreadsheet and you open it on your compliant 
program and look at cell D7 you should see the same value that I see on 
my compliant program. If the spec can't guarantee this, it's broken.

(except for an allowable margin of error)

Hence, I disagree with being fuzzy at low levels. If you give me a level 
1 spreadsheet and I see a sheet full of errors we do not have 
interoperability and the spec failed.

What should happen is that if you type "3"+3 in Gnumeric, the formula 
gets saved as VALUE("3")+3 so that when I open it in OOo I get the same 
result you do.


> Oh, why that particular semantic at level 3?
> One advantage of this particular semantic is that naive users either get
> what they were probably expecting, OR they get a warning message
> that there's a problem (instead of silently getting the wrong answer).

I don't see this. How does leaving scemantics to level 3 ensure that 
level 1 users get what they were expecting?


>> There should also be formal definitions of what the functions compute,
>> and a reference algorithm should be provided, if possible.
> 
> A formal definition in terms of a mathematical expression would be good.
> I don't think we should mandate a particular algorithm, but mandating 
> particular
> test case results are fair game,

I wouldn't /mandate/ an algorithm, but a /reference/ algorithm seems 
alright. In other words, "your function must give the same values as 
this function within the allowed margin of error". At least in theory, a 
reference algorithm is the strictest test case, because it allows 
testing against any value permitted by the spec.

I don't have strong feelings either way. If we did have reference 
algorithms, it should only be for functions that really merit them. It's 
stupid to have a reference algorithm for SUM.

Cheers,
Daniel.
-- 
      /\/`) http://opendocumentfellowship.org
     /\/_/
    /\/_/ I'm not saying there should be a capital punishment for
    \/_/  stupidity, but why don't we just take the safety labels
    /     off of everything and let the problem solve itself?