[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Axiom musings...
From: |
Tim Daly |
Subject: |
Re: Axiom musings... |
Date: |
Wed, 9 Dec 2020 22:36:38 -0500 |
Robert Harper's "Holy Trinity" of Computer Science
(aka "Computational Trinitarianism")
Proof Theory
(Logic)
/ \
/ \
Category Theory ---- Type Theory
(Mathematics) (Programming)
There is a correspondence between Proof Theory,
Category Theory, and Type Theory or equivalently
a correspondence of Logic, Mathematics, and
Programming.
A concept may arise in, say, programming or
the same concept can arise in mathematics or
logic.
"You don't know what you are talking about until
you understand it all three ways."
(https://www.cs.uoregon.edu/research/summerschool/summer12/videos/Harper1_0.mp4)
This is essentially the research goal of the SANE version
of Axiom. The goal is to unify computational mathematics
in one system. The current effort is cleaning up things like
proper functors (category theory), adding proofs for
programs (proof theory) and implementing fully-first
class dependent types (type theory).
Harper's Holy Trinity gives a sound way of thinking about
various design decisions in the SANE version of Axiom.
Progress, albeit glacial, is being made.
Tim
On 12/3/20, Tim Daly <axiomcas@gmail.com> wrote:
> HILL CLIMBING
>
> I really encourage you to read the slides:
> Vladimir Voevodsky. Univalent foundations, March 2014. Presentation at
> IAS,http://www.math.ias.edu/~vladimir/Site3/Univalent_Foundations_files/2014_IAS.pdf
>
> The mathematical references likely won't make sense but the
> most important point of the article will. He found mistakes in
> work he and others did. Since these people were famous their
> work was never really checked (fortunately that's not my issue).
>
> quoting from the slide 28:
>
> I now do my mathematics with a proof assistant and do not have to worry
> all the time about mistakes in my arguments or about how to convince
> others that my arguments are correct.
>
> But I think that the sense of urgency that pushed me to hurry with the
> program remains. Sooner or later computer proof assistants will become
> the norm, but the longer this process takes the more misery associated
> with mistakes and with unnecessary self-verification the practitioners of
> the
> field will have to endure.
>
>
>
>
> Of all of the software created, computer algebra seems like the most
> reasonable place to apply proof techniques. My feeling is that computer
> algebra is still 19th century, pre-formal mathematics. We can do better.
> We must. And as a consequence we elevate "computer algebra" to
> "computational mathematics", surely a worthwhile research goal.
>
> Tim
>
>
> On 12/3/20, Tim Daly <axiomcas@gmail.com> wrote:
>> I just ran across an interesting example about the difference
>> between testing and proving (not that I need to be convinced).
>> http://www.lix.polytechnique.fr/Labo/Samuel.Mimram/teaching/INF551/course.pdf
>> (page 10):
>>
>>
>> On day one, our mathematician finds out using a formal computation
>> software
>> that
>>
>> \int{\sin(t)/t}{dt} = \pt/2
>>
>> On day two, he tries to play around a bit with such formulas and finds
>> out
>> that
>>
>> \int{(\sin(t)/t) (\sin(t/101)/(t/101)}{dt} = \p2/2
>>
>> On day three, he thinks maybe a pattern could emerge and discovers that
>>
>> \int{(\sin(t)/t) (\sin(t/101)/(t/101) (\sin(t/201)/(t/201))}{dt} = \p2/2
>>
>> On day four, he gets quite confident and conjectures that, for every n∈N,
>>
>> \int{\prod{ /sin(t/(100n+1)) / (t/(100n+1))}{dt} = \pi/2
>>
>>
>> In fact, the conjecture breaks starting at
>>
>> n= 15 341 178 777 673 149 429 167 740 440 969 249 338 310 889
>>
>> and none of the usual tests would have found this out.
>>
>> Tim
>>
>>
>>
>> On 11/29/20, Tim Daly <axiomcas@gmail.com> wrote:
>>> Axiom, as it was released to me long ago, is being
>>> overtaken by current technology.
>>>
>>> I'm taking an MIT course on "Underactuated Robots" [0]
>>> (I spent a large portion of my life doing robotics).
>>>
>>> The instructor is using Jupyter notebooks to illustrate
>>> control algorithms. He is doing computation and
>>> showing the graphics in a single notebook. Axiom
>>> could "almost" do that in the 1980s.
>>>
>>> It is clear that, with a sufficiently rich set of algorithms,
>>> such as one finds in CoCalc, Axiom is showing its age.
>>>
>>> This is the primary reason why I consider the SANE
>>> research effort vital. An Axiom system that has
>>> proven algorithms, follows knowledge from leading
>>> mathematics (e.g. Category Theory), and pushes the
>>> state of the art will keep Axiom alive and relevant.
>>>
>>> Axiom was, and is, research software. Just doing
>>> "more of the same" goes nowhere slowly. Following
>>> that path leads to the dustbin of history.
>>>
>>> Tim
>>>
>>> [0]
>>> https://www.youtube.com/watch?v=GPvw92IKO44&list=UUhfUOAhz7ynELF-s_1LPpWg&index=28
>>>
>>>
>>>
>>> On 11/27/20, Tim Daly <axiomcas@gmail.com> wrote:
>>>> The current mental struggle involves properly re-constructing
>>>> Axiom so Category Theory (e.g. objects, morphisms, identity,
>>>> and composition) is respected between the domains. Done
>>>> properly, this should make resolving type coercion and
>>>> conversion less ad-hoc. This is especially important when
>>>> dealing with first-class dependent types.
>>>>
>>>> Axiom categories and domains, in the SANE model, compile
>>>> to lisp classes. These classes are "objects" in the Category
>>>> Theory sense. Within a given category or domain, the
>>>> representation (aka, the "carrier" in type theory) is an
>>>> "object" (in Category Theory) and most functions are
>>>> morphisms from Rep to Rep. Coercions are functors,
>>>> which need to conform to the Category Theory properties.
>>>>
>>>> It also raises the interesting question of making Rep
>>>> be its own class. This allows, for example, constructing
>>>> the polynomial domain with the Rep as a parameter.
>>>> Thus you get sparse, recursive, or dense polynomials
>>>> by specifying different Rep classes. This is even more
>>>> interesting when playing with Rep for things like
>>>> Clifford algebras.
>>>>
>>>> It is quite a struggle to unify Category Theory, Type
>>>> Theory, Programming, Computer Algebra, and Proof
>>>> Theory so "it all just works as expected". Progress is
>>>> being made, although not quickly.
>>>>
>>>> For those who want to play along I can recommend the
>>>> MIT course in Category Theory [0] and the paper on
>>>> The Type Theory of Lean [1]. For the programming
>>>> aspects, look at The Art of the Metaobject Protocol [2].
>>>>
>>>> [0] https://www.youtube.com/user/youdsp/playlists
>>>> [1]
>>>> https://github.com/digama0/lean-type-theory/releases/download/v1.0/main.pdf
>>>> [2] https://mitpress.mit.edu/books/art-metaobject-protocol
>>>>
>>>> Tim
>>>>
>>>>
>>>> On 11/10/20, Tim Daly <axiomcas@gmail.com> wrote:
>>>>> Another area that is taking a great deal of time and effort
>>>>> is choosing a code architecture that has a solid background
>>>>> in research.
>>>>>
>>>>> Axiom used CLU and Group Theory as two scaffolds to
>>>>> guide design choices, making it possible to argue whether
>>>>> and where a domain should be structured and where it
>>>>> should occur in the system.
>>>>>
>>>>> Axiom uses a Pratt-parser scheme. This allows the
>>>>> ability to define and change syntax by playing with the
>>>>> LED and NUD numeric values. It works but it is not
>>>>> very clear or easy to maintain.
>>>>>
>>>>> An alternative under consideration is Hutton and Meijer's
>>>>> Monadic Parser Combinators. This constructs parser
>>>>> functions and combines them in higher order functions.
>>>>> It is strongly typed. It provides a hierarchy of functions
>>>>> that would allow the parser to be abstracted at many
>>>>> levels, making explanations clearer.
>>>>>
>>>>> Ideally the structure of the new parser should be clear,
>>>>> easy to maintain, and robust under changes. Since SANE
>>>>> is a research effort, maximum flexibility is ideal.
>>>>>
>>>>> Since SANE is intended to be easily maintained, it is
>>>>> important that the compiler / interpreter structure be
>>>>> clear and consistent. This is especially important as
>>>>> questions of which proof language syntax and a
>>>>> specification language syntax is not yet decided.
>>>>>
>>>>> Tim
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On 10/9/20, Tim Daly <axiomcas@gmail.com> wrote:
>>>>>> A great deal of thought has gone into the representation
>>>>>> of functions in the SANE version.
>>>>>>
>>>>>> It is important that a function lives within an environment.
>>>>>> There are no "stand alone" functions. Given a function
>>>>>> its environment contains the representation which itself
>>>>>> is a function type with accessor functions. Wrapped
>>>>>> around these is the domain type containing other
>>>>>> functions. The domain type lives within several
>>>>>> preorders of environments, some dictated by the
>>>>>> structure of group theory, some dictated by the structure
>>>>>> of the logic.
>>>>>>
>>>>>> Lisp has the ability to construct arbitrary structures
>>>>>> which can be nested or non-nested, and even potentially
>>>>>> circular.
>>>>>>
>>>>>> Even more interesting is that these structures can be
>>>>>> "self modified" so that one could have a domain (e.g.
>>>>>> genetic algorithms) that self-adapt, based on input.
>>>>>> Or "AI" domains, representated as weight matrices,
>>>>>> that self-adapt to input by changing weights.
>>>>>>
>>>>>> Ultimately the goal of the representation should be that,
>>>>>> given a function, it should be possible to traverse the
>>>>>> whole of the environment using a small, well-defined
>>>>>> set of well-typed structure walkers.
>>>>>>
>>>>>> All of this is easy to write down in Lisp.
>>>>>> The puzzle is how to reflect these ideas in Spad.
>>>>>>
>>>>>> Currently the compiler translates all of the Spad
>>>>>> code to Lisp objects so the Spad->Lisp conversion
>>>>>> is easy. Lisp->Spad is also easy for things that have
>>>>>> a surface representation in Spad. But, in general,
>>>>>> Lisp->Spad is only a partial function, and somewhat
>>>>>> limited at that.
>>>>>>
>>>>>> I suspect that the solution will allow some domains
>>>>>> to be pure Lisp and others will allow in-line Lisp.
>>>>>>
>>>>>> Most of these ideas are nearly impossible to even
>>>>>> think about in other languages or, if attempted,
>>>>>> fall afoul of Greenspun's Tenth Rule, to wit:
>>>>>>
>>>>>> "Any sufficiently complicated C or Fortran program
>>>>>> contains an ad hoc, informally-specified, bug-ridden,
>>>>>> slow implementation of half of Common Lisp."
>>>>>>
>>>>>> Fortunately Spad is only a domain specific language
>>>>>> on top of Lisp. Once past the syntax level it is Lisp
>>>>>> all the way down.
>>>>>>
>>>>>> Tim
>>>>>>
>>>>>
>>>>
>>>
>>
>