[isabelle] Goedel's paradox, canonical definitions vs. instances/models (dependent type theory and the language of species), expressiveness vs. automation, re-evaluating natural deduction



Dear Members of the Research Community,

Considering the results of Shinichi Mochizuki (see section 2 below) at
	http://www.kurims.kyoto-u.ac.jp/~motizuki/Inter-universal%20Teichmuller%20Theory%20IV.pdf

as well as offering a further response to Ramana Kumar's comment (see section 3 
below) at
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2015-July/msg00149.html

and answering Andrei Popescu's comment (see section 4 below) at
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2015-December/msg00059.html

is a good opportunity to re-evaluate natural deduction, and to summarize and 
conclude the discussion.

Most of the statements about the Isabelle proof assistant software should also 
apply to the HOL and the Coq proof assistants.


1. Goedel's paradox (Goedel's First Incompleteness Theorem)

As shown previously at
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2016-January/msg00029.html
because of the non-definability of a Goedel numbering (or encoding) function 
within Q0 (or more precisely: the non-expressibility, since it is not a 
well-formed formula) due to the violation of type restrictions on lambda 
application (case (b) in the definition of wffs (well-formed formulae) [cf. 
Andrews, 2002, p. 211]), Goedel's First Incompleteness Theorem is not a 
mathematical theorem (or metatheorem), and the same is also true of Goedel's 
Second Incompleteness Theorem, since both theorems would require a wff 
representing the Goedel numbering (or encoding) function.

This violation of type restrictions in formal logic or mathematics corresponds 
to the philosophical insight that Goedel's construction of a proposition 
involving self-referencing negativity ("I am not provable") has to be 
considered as an antinomy, as both Russell's paradox ("the set of all sets that 
are not members of themselves") and Goedel's paradox known as the proposition 
in Goedel's First Incompleteness Theorem ("I am not provable", or originally: 
"We therefore have before us a proposition that says about itself that it is 
not provable" [GÃdel, 1931, p. 598]) have the two constitutive properties of 
antinomies: self-reference and negativity (negation).

In the presentation of Goedel's First Incompleteness Theorem with the Isabelle 
proof assistant software, the definition of the Goedel encoding function was 
obtained by using the metamathematical notion "term" (level 1) as a concrete 
mathematical type symbol (level 2) [cf. Andrews, 2002, p. 210], which is a 
confusion of two different logical levels (or language levels), and hence a 
violation of type restrictions.

Generally speaking, due to certain decisions in the design of Isabelle the 
stratification of the three different logical levels (as, for example, in the 
R0 implementation) collapse into a single one, which requires justification and 
certain measures to prevent the confusion of logical levels (or language 
levels) within the single sphere obtained.


2. Canonical definitions vs. instances/models (dependent type theory and the 
language of species)

In his development of a "language of species", in which a "'species' is a 'type 
of mathematical object', such as a 'group', a 'ring', a 'scheme', etc." 
[Mochizuki, 2015, p. 66], one of Professor Mochizuki's interests is to 
distinguish canonical definitions from arbitrary choices: "The fact that the 
data involved in a species is given by abstract set-theoretic formulas imparts 
a certain canonicality to the mathematical notion constituted by the species, a 
canonicality that is not shared, for instance, by mathematical objects whose 
construction depends on an invocation of the axiom of choice in some particular 
ZFC-model [...]. Moreover, by furnishing a stock of such 'canonical notions', 
the theory of species allows one, in effect, to compute the extent of deviation 
of various 'non-canonical objects' [i.e., whose construction depends upon the 
invocation of the axiom of choice!] from a sort of 'canonical norm'." 
[Mochizuki, 2015, p. 70]

This is also one of the motivations for dependent type theory, in which the 
same can be expressed even more elegantly. The creator of R0 initially intended 
to prove that the definition of natural numbers by John von Neumann is a 
specific instance (model) of the canonical definition with the Peano axioms. 
For practical reasons, a simpler, but conceptually equivalent proof was carried 
out: The group property of the exclusive disjunction (XOR), which in dependent 
type theory can be directly expressed by the theorem
	Grp o XOR
as proven at
	http://www.kenkubota.de/files/R0_draft_excerpt_with_definitions_and_groups.pdf (p. 416)

using the canonical definitions of groups (cf. p. 359) and of XOR (cf. p. 358). 
Furthermore, for the canonical definition of groups the proof is given that 
every group has a unique identity element (cf. p. 359 ff.), and this property 
is then transferred to the XOR group by performing basically only a few 
substitutions (cf. pp. 417 ff.).

The expression of the relationship between a canonical definition and a 
specific instance in dependent type theory has a number of advantages:
1. It does not require a meta-theory (an additional theory layer on top of the 
formal/object language) such as, for example, model theory, since the 
relationship can be expressed _within_ the formal language (object language) 
itself (in philosophy this would be called "immanent").
2. It does not use (axiomatic) set theory.
3. It does not require the Axiom of Choice (which exists neither in Q0 nor in 
R0).
For the critique of the concept of meta-theory and of model theory, cf. 
[Kubota, 2013, pp. 13 ff.].
For the critique of the concept of meta-theory and of (axiomatic) set theory, 
see also
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2015-December/msg00042.html

Please note that seeking an "explicit description of certain tasks typically 
executed at an implicit, intuitive level by mathematicians" [Mochizuki, 2015, 
p. 66] is the aim of type theory - there are nearly identical formulations in 
the work of Andrews - and that "the emphasis in the theory of species lies in 
the programs â i.e., 'software' â that yield the desired output data, not on 
the output data itself" [Mochizuki, 2015, p. 69] is also one of the key 
features of lambda-calculus. For comparison, the corresponding quote in 
Andrews' work is: "[M]ost mathematicians do make mental distinctions between 
different types of mathematical objects, and use different types of letters to 
denote them, so it might be claimed that type theory provides a more natural 
formalization of mathematics as it is actually practiced. In addition, explicit 
syntactic distinctions between expressions denoting intuitively different types 
of mathematical entities are very useful in computerized systems for exploring 
and applying mathematics [...]." [Andrews, 2002, p. 205]


3. Expressiveness vs. automation

The main notion of proof verification, as done by the R0 implementation, is 
"expressiveness", focusing on principle (philosophical) questions such as 
expressibility, definability or provability without regard to practical 
criteria such as efficiency or automation in carrying out proofs. Hence proof 
verification is very close to establishing a logistic system like Q0 or R0, 
"finding a formulation of type theory which is as expressive as possible, 
allowing mathematical ideas to be expressed precisely with a minimum of 
circumlocutions, and which is as simple and economical as is possible without 
sacrificing expressiveness." [Andrews, 2002, p. 205]

Ramana Kumar's statement that "there are already several variations of theorem 
provers implementing a roughly equivalent logic", e.g., "Isabelle/HOL, HOL4, 
HOL Light", is correct if one considers logical equivalence only. But the aim 
here is to find the most adequate or natural formulation. In most cases this 
coincides with "The Principle of Reductionism" [Kubota, 2015, p. 11]. For 
example, "HOL Light [...] has 'ten primitive rules of inference' [Harrison, 
2009, p. 61]" [Kubota, 2015, p. 14], but Q0 has only a single one.

For this reason, I emphasized the definability of any well-formed formula 
(including all connectives and quantifiers) and derivability of any rule 
(including the rule of modus ponens) on the basis of only a very small set of 
primitive symbols and rules, namely
- a single rule of inference,
- a single variable binder,
- only two primitive constants, and
- only two primitive types
in the presentation of Q0 at
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2015-July/msg00147.html

Furthermore, since natural deduction requires more rules of inference than 
axiomatic (Hilbert-style) deductive systems, Hilbert-style logistic systems 
like Q0 are the preferred choice for finding a formulation of type theory as 
discussed above.

Like Church's type theory [cf. Church, 1940] known for its "precise formulation 
of the syntax" [Paulson, 1989, p. 5], Andrews' Q0 is known for its precision 
and accuracy due to its simple, clean and elegant design. "Church formalizes 
syntax, including quantifiers, in the typed [lambda]-calculus. His technique is 
now standard in generic theorem proving. See [...] Andrews [1] for the formal 
development." [Paulson, 1989, p. 5]

Andrews' definition of "the universal quantifier [...] defined in terms of 
truth" [Paulson, 1989, p. 14] has gained Paulson's attention, and in dependent 
type theory one can also add type abstraction. For comparison, see:
	http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-175.pdf (p. 14)
	http://www.kenkubota.de/files/R0_draft_excerpt_with_definitions_and_groups.pdf (p. 356)

Another example for the precision of the formal language is the 
non-definability (non-expressibility) of a Goedel numbering (or encoding) 
function within Q0 or R0. Isabelle/HOL currently lacks this precision as shown 
at
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2016-January/msg00029.html


4. Natural deduction and logical framework (Isabelle et al.)

However, for the practical purpose of automation, natural deduction provides 
the huge advantage of making metatheorems symbolically representable instead 
of, for example, only applying proof templates via recursive file inclusion as 
in the R0 implementation. So if Paulson writes: "Natural deduction is far 
superior for automated proof." [Paulson, 1989, p. 3], this is obviously true, 
but only half of the truth; the other half of the truth is that the precision 
of Hilbert-style systems (i.e., Church and Andrews) is lost when opting for 
natural deduction. Hence the decision between Hilbert-style systems and natural 
deduction is a question of weighting the advantages and disadvantages depending 
on the purpose (proof verification vs. automated deduction), and furthermore, 
in the case of natural deduction, additional measures have to be taken in order 
to preserve the strict distinctions between the logical levels. So for proof 
verification I would advise axiomatic (Hilbert-style) deductive systems, and 
for automated deduction natural deduction, but in the latter case the 
additional measures are important.

A logical framework provides the means to implement several object logics, 
which is not necessary if one regards R0 as the natural and ideal logic. 
Practical reasons, such as re-using existing source code, may be relevant, 
however, in general "I would advise against using some kind of meta-logic or 
meta-language, as additional features may weaken the rigor of (or even create 
an inconsistency in) the logical kernel" as stated earlier at
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2015-December/msg00018.html
But there is no principle objection if additional measures are taken in order 
to preserve the strict distinctions between the logical levels as specified at
	https://lists.cam.ac.uk/pipermail/cl-isabelle-users/2016-January/msg00029.html

Additional measures have to be implemented in both a concrete and a general way.

a) At the concrete level, for the choice of
- natural deduction (with the danger of not distinguishing levels 2 and 3) such 
an implementation seems to be, at the first glance, provided with schematic 
(syntactical) variables, and with "Meta-logic" [Nipkow, 2015, p. 11] operators 
having a lower precedence than the "Logic" [Nipkow, 2015, p. 11] operators, but 
for the choice of
- a logical framework (with the danger of not distinguishing levels 1 and 2) 
not, since the metamathematical notion "term" (level 1) is treated as a 
concrete mathematical type symbol (level 2) in Paulson's presentation of 
Goedel's First Incompleteness Theorem.

b) At the general level, the implementation of automated deduction can receive 
its philosophical legitimation only through the ideal formulation of type 
theory as expressed by Andrews quoted above. This can be achieved by 
establishing a one-to-one correspondence between the theorems/metatheorems of 
the implementations of both automated deduction and proof verification, ideally 
by a metamathematical proof arithmetizing both systems, or at least by 
comparison of canonical definitions and proofs. My preference would be the 
generation of R0 code by proofs obtained in Isabelle/HOL, such that the 
Isabelle/HOL proofs could be verified by the R0 implementation.


If there are any further e-mail comments, please allow me to postpone answers 
where appropriate until my forthcoming publication of R0 [cf. Kubota, 2015].


Kind regards,

Ken Kubota


References

Andrews, Peter B. (2002), An Introduction to Mathematical Logic and Type 
Theory: To Truth Through Proof. Second edition. Dordrecht / Boston / London: 
Kluwer Academic Publishers. ISBN 1-4020-0763-9. DOI: 10.1007/978-94-015-9934-4.

Church, Alonzo (1940), "A Formulation of the Simple Theory of Types". In: 
Journal of Symbolic Logic 5, pp. 56-68.

GÃdel, Kurt (1931), "On formally undecidable propositions of Principia 
mathematica and related systems I". In: Heijenoort, Jean van, ed. (1967), From 
Frege to GÃdel: A Source Book in Mathematical Logic, 1879-1931. Cambridge, 
Massachusetts: Harvard University Press, pp. 596-616.

Kubota, Ken (2013), On Some Doubts Concerning the Formal Correctness of GÃdel's 
Incompleteness Theorem. Berlin: Owl of Minerva Press. ISBN 978-3-943334-04-3. 
DOI: 10.4444/100.101. See: http://dx.doi.org/10.4444/100.101

Kubota, Ken (2015), On the Theory of Mathematical Forms (Draft of May 18, 
2015). Unpublished manuscript. SHA-512: a0dfe205eb1a2cb29efaa579d68fa2e5 
45af74d8cd6c270cf4c95ed1ba6f7944 fdcffaef2e761c8215945a9dcd535a50 
011d8303fd59f2c8a4e6f64125867dc4. DOI: 10.4444/100.10. See: 
http://dx.doi.org/10.4444/100.10

Kubota, Ken (2015a), Excerpt from [Kubota, 2015] (pp. 356-364, pp. 411-420, and 
pp. 754-755). Available online at 
http://www.kenkubota.de/files/R0_draft_excerpt_with_definitions_and_groups.pdf 
(January 24, 2016). SHA-512: 67a48ee12a61583bd82c286136459c0c 
9fd77f5fd7bd820e8b6fe6384d977eda 4d92705dac5b3f97e7981d7b709fc783 
3d1047d1831bc357eb57b263b44c885b.

Mochizuki, Shinichi (2015), "Inter-universal Teichmuller Theory IV: Log-volume 
Computations and Set-theoretic Foundations" (December 2015). Available online 
at 
http://www.kurims.kyoto-u.ac.jp/~motizuki/Inter-universal%20Teichmuller%20Theory%20IV.pdf 
(January 24, 2016). SHA-512: 93372b4acba90195fc6ffc2e16830464 
b3bf27c4bdea156a915d3e7e5fdaad2a 4c85d0e1917064a8afd59e1a3af29a70 
a438252bf2cfcdb90b12cd20f272c179.

Nipkow, Tobias (2015), "What's in Main" (May 25, 2015). Available online at 
https://isabelle.in.tum.de/dist/Isabelle2015/doc/main.pdf (January 24, 2016).

Paulson, Lawrence C. (1989), "A Formulation of the Simple Theory of Types (for 
Isabelle)". Available online at 
http://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-175.pdf (January 24, 2016).

____________________

Ken Kubota
doi: 10.4444/100
http://dx.doi.org/10.4444/100





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