Re: [isabelle] Non-idempotence of datatype constructors

Unfortunately, the simproc breaks a few things in the AFP.

One of them is the proof of viewL_correct in Finger-Trees.FingerTree,
where a call to "auto" loops (it also loops when changing it to "simp").

It's not an implementation problem of my simproc either. When I disable
the simproc and add the theorems

"Node3 a b c d ≠ b" "Node3 a b c d ≠ c" "Node3 a b c d ≠ d"

to the simpset instead, it also loops.

These rules look perfectly harmless to me, so my suspicion would be that
the simpset in this theory was fragile all this time and these new rules
simply exposed that. However, I couldn't find any problematic-looking
simp rules either, so perhaps it's some premise that the simplifier adds
to the simpset itself.

After inspecting the trace, it seems that the simplifier tries to apply
the following rule over and over again:

[1]Adding rewrite rule "??.unknown":
nodeToList nd2 = [] ∧ nodeToList nd3 = [] ⟹
False ⟹
undefined = n_unwrap nd1 ⟹
is_leveln_node 0 nd1 ⟹ is_measured_node nd1 ⟹ nodeToList nd1 ≡ [undefined]

(plus a few other variants of this with other permutations of nd1, nd2,
nd3, so I guess it loops when trying to rewrite the first premise)

I find this a bit odd. Why does the simplifier add something that has
"False" in its premises to the simpset in the first place? Shouldn't
this entire thing get rewritten to "True" and discarded before that happens?


On 02/05/2020 18:04, Manuel Eberl wrote:
> I attached a proof of concept (works with Isabelle 2020) using the
> simple size-based approach, including some example applications.
> It works well, although I'm not sure what the proper way to get the
> datatype information is (e.g. the list of all the constructors of the
> datatype and the associated other datatypes in case of mutual recursion).
> Is the ML interface of the BNF package documented anywhere (in
> particular this aspect)?
> Manuel
> On 02/05/2020 16:19, Manuel Eberl wrote:
>> True, but after your suggestion, I realised that the solution with the
>> "proper subexpression" relation (or, alternatively, the size function)
>> combined with a simproc that produces these theorems on the spot is
>> actually the superior approach in every respect. It's simpler, more
>> general, and probably more performant.
>> I can try to come up with a proof-of-concept implementation using the
>> size function approach, since that needs no additional new features from
>> the BNF package.
>> Manuel
>> On 02/05/2020 16:16, Tobias Nipkow wrote:
>>> A first version which only deals with depth 1 would already cover most
>>> of the practical cases.
>>> Tobias
>>> On 02/05/2020 15:50, Manuel Eberl wrote:
>>>> That sounds like a good idea.
>>>> However, if such a simproc is to work for any nesting of
>>>> constructors,having pre-proven theorems for every constructor will not
>>>> be enough.Instead, I suppose one would have to introduce a
>>>> "proper-subexpression"relation for datatypes (e.g. xs < Cons x xs) along
>>>> with a proof thatthis relation has the obvious properties (irreflexive,
>>>> asymmetric,transitive).
>>>> I guess that is something that only a datatype package plugin similar
>>>> tothe one for the "size" function could provide. Having looked at the
>>>> codebriefly, I think only the people who wrote the BNF package could (or
>>>> atleast should) implement that.
>>>> Alternatively, one could just use the size function (as someone
>>>> alreadysuggested in this thread) to get pretty much the same thing,
>>>> except thatit won't work for all datatypes (e.g. infinitely branching
>>>> ones).
>>>> Manuel
>>>> On 02/05/2020 15:36, Tobias Nipkow wrote:
>>>>> I do think such rules are useful, esp if they are there by default. I
>>>>> suggest they are best handled by a simproc that is triggered by any
>>>>> "(=)" but that checks immediately if the two argumenst are of the
>>>>> appropriate type and form. That can be done very quickly (there are
>>>>> similar simprocs already). The simproc should work for any datatype and
>>>>> any degree of nesting of the constructors.
>>>>> Tobias
>>>>> On 01/05/2020 22:51, Manuel Eberl wrote:
>>>>>>> Firstly, I don't think these theorem is especially useful. You might
>>>>>>> have planned to add this to the simplifier, but its term net
>>>>>>> doesn't do
>>>>>>> any magic here. It will end up checking every term that matches
>>>>>>> "Cons x
>>>>>>> xs = ys" for whether "xs" can match "ys". I'm not sure if that
>>>>>>> matching
>>>>>>> is equality, alpha-equivalent or unifiable.
>>>>>> I honestly never think /that/ much about the performance
>>>>>> implications of
>>>>>> simp rules (as long as they're unconditional). At least for lists, by
>>>>>> the way, this is already a simp rule by default though, and lists are
>>>>>> probably by far the most prevalent data type in the Isabelle universe.
>>>>>> But you're certainly right that it would make sense to keep a look at
>>>>>> this performance impact if one wanted to add these to the simp set for
>>>>>> all datatypes by default, and I agree that the rules are probably not
>>>>>> helpful /that/ often. Still, it might be nice to have them available
>>>>>> nonetheless.
>>>>>>> Secondly, you can prove these theorems by using this handy trivial
>>>>>>> theorem : "size x ~= size y ==> x ~= y". Apparently that theorem
>>>>>>> has the
>>>>>>> name  Sledgehammer.size_ne_size_imp_ne - presumably the sledgehammer
>>>>>>> uses it to prove such inequalities.
>>>>>> It's even easier to prove it by induction. Plus, in fact, the "size"
>>>>>> thing only works if the data type even has a sensible size function.
>>>>>> That is not always the case, e.g. when you nest the datatype through a
>>>>>> function.
>>>>>> My main point however is that when you have a datatype with a dozen
>>>>>> binary constructors, there's quite a bit of copying & pasting involved
>>>>>> before you've proven all those theorems. Since it can (probably) be
>>>>>> automated very easily, why not do that? Whether or not these should be
>>>>>> simp lemmas by default is another question.
>>>>>> Manuel

This archive was generated by a fusion of Pipermail (Mailman edition) and MHonArc.