SELF-R EF ER ENCE

Volker Halbach
Birmingham
7th January 2013

Some observations in this presentation are folklore and have been

known for some time, even though I came to learn about most of them
just recently from a email exchanges with Albert Visser.

Example
(A) propositionally implies (A).
Wir haben also einen Satz vor uns, der seine eigene
Unbeweisbarkeit behauptet.
We thus have a sentence before us that states its own
unprovability.
Gdel (1931, p. 175)

What do we mean when we say that a sentence says about itself that it is
provable (1 -true, not provable, Rosser provable, provable without cut
etc.)?

(A)

Does (A) say (about itself) that it propositionally implies itself?

Does (A) say that it propositionally implies (A)?

Does (A) say that it is propositionally implied by (A)?

Does (A) ascribe to itself the property of being propositionally

self-implying?

Does (A) ascribe to itself the property of propositionally implying

(A)?

Does (A) ascribe to itself the property of being propositionally

implied by (A)?

(B) propositionally implies (A).

(B)

There are three steps in the construction of a self-referential formulae,

that is, a formula saying about itself that it has property P:
1. Fix a Gdel coding.

Do (A) and (B) both say about themselves that they propositionally
imply (A)?

2. Pick a formula expressing the property P (under the chosen

coding).

Do (A) and (B) both ascribe to themselves the property of

propositionally implying (A)?

3. Construct a formula ascribing to itself property P via the chosen

formula.

(A) is true; (B) is false.

Corresponding to these three steps there are three dimensions of

intensionality.

observation
A fixed point of a formula (x) (relative to a system ) is a formula
such that () obtains.

definition
A fixed-point operator is a function f from the set of formulae with the
variable v free into the set of formulae such that
f () (f ()).
Any sentence saying about itself that it has the property expressed by
(x) is a fixed point of (x).

For any given formula (x) there is no formula (x) that defines the
set of fixed points of (x), that is, there is no (x) satisfying the
following condition:
N () (() )
Moreover, for any given (x) the set of its provable fixed points, that is,
the set of all sentences with
()
is not recursive but only recursively enumerable.

Henkin (1952) asked:

If is any standard formal system adequate for recursive
number theory, a formula (having a certain integer q as its
Gdel number) can be constructed which expresses the
proposition that the formula with Gdel number q is provable
in . Is this formula provable or independent in ?
Kreisel (1953) replied:

kreisels criterion for the expression of

provability
A formula Bew(x) is said to express provability in if it satisfies the
following condition: for numerals n, Bew(n) can be proved in if and
only if the formula with number n can be proved in .1

We shall show below that the answer to Henkins question

depends on which formula is used to express the notion of
provability in .

A Henkin sentence is a sentence that says of itself that its provable.

To be a Henkin sentence, a sentence has to be a provable fixed point of
the provability predicate at least. So if is a Henkin sentence we have:
Bew()

This is the third paragraph of Kreisels 1953 paper with the notation adapted.

I suspect that Kreisel and Henkin implicitly agreed on a criterion for

self-reference along the following lines:

kreiselhenkin criterion for self-reference

Let a formula (x) expressing a certain property P in be given. Then
a formula says about itself that it has property P iff it is of the form
(t) for some closed term t that has (the code of) (t) as its value.
If the usual Gdel sentence is constructed in a language with suitable
function symbols, it will satisfy this condition.

kreisels observation
There is a formula Bew1 (x) and a term t1 such that the following three
conditions are satisfied:

Proof
Fix some predicate Bew(x) that weakly represents -provability in .
By Gdels diagonal lemma there is a term t1 such that
t1 = t1 = t1 Bew(t1 )

(i) Bew1 weakly represents provability in .

(ii) t1 = Bew1 (t1 )

(1)

Now define Bew1 (x) as

(iii) Bew1 (t1 )

x = t1 Bew(x).
Similarly, there is a provability predicate Bew2 (x) and a term t2 such
that
(i) Bew2 weakly represents provability in .

(Bew1 (x))

Clearly t1 = t1 = t1 Bew(t1 ) and hence (ii) holds by (1). Since

t1 = t1 Bew(t1 )

(ii) t2 = Bew2 (t2 )

is provable in pure logic (and thus in ), Bew1 (t1 ) is provable and (iii) is
satisfied.

(iii) Bew2 (t2 )

Proof
Fix some predicate Bew(x) that weakly represents -provability in .
By Gdels diagonal lemma there is a term t2 such that
t2 = t2 =/ t2 Bew(t2 )

(2)

Now define Bew2 (x) as

x =/ t2 Bew(x)
Clearly t2 = t2 =/ t2 Bew(t2 ) and hence (ii) holds by (2). Since
t2 =/ t2 Bew(t2 )
is refutable in pure logic (and thus in ), Bew2 (t2 ) and (iii) is
satisfied.

Henkin and other people have complained ever since that Kreisel hadnt
used the canonical provability predicate.
But nobody (except for Smoryski 1991 and recently Visser) has
complained about the way Kreisel obtained the terms t1 and t2 .

Let Bew2 (x) be the provability predicate from above, that is,
x =/ t2 Bew(x) with Bew(x) as the canonical provability predicate. It
weakly represents -provability and thus expresses provability
according to Kreisels criterion for the expression of provability.
From Kreisels Observation we have:

Lbs theorem is the answer to Henkins problem if the provability

predicate is kept canonical. Assume Bew(x) satisfies the derivability
conditions.

(i) t2 = Bew2 (t2 )

(ii) Bew2 (t2 )

lemma
Any two fixed points of Bew(v) are -provably equivalent.

Now apply the standard diagonal method to Bew2 (x) to obtain a term
t3

More formally:
1 Bew(1) and 2 Bew(2) imply 1 2 .

(i) t3 = Bew2 (t3 )

(ii) Bew2 (t3 )

lemma
Any two fixed points of Bew(v) are -provably equivalent.

Both t2 and t3 satisfy Kreisels criterion for self-reference.

Conclusion: Whether a Henkin sentence is provable or refutable does
not depend contra Kreisel only on the provability predicate; it also
depends on how self-reference is obtained.

But can one obtain a refutable Henkin sentence with canonical

diagonalisation but a nonstandard provability predicate?

theorem (Visser)
There is a provability predicate BewV (x) weakly representing
provability in such that its fixed point obtained by the usual diagonal
construction is refutable.

observation (Picollo)
There is a provability predicate BewP (x) weakly representing
provability in such that its fixed point obtained by the usual diagonal
construction is neither provable nor refutable.

Let d be the canonical fixed point operator that maps any formula (x)
to its Gdel fixed point and d. its representation in .
Let Bew(x) be some formula representing provability and construct a
formula BewV (x) using some fixed point construction:
BewV (x) x =/ d. (BewV (x)) Bew(x)
Now apply the canonical d to the predicate BewV (x).
(i) d(BewV )
(ii) BewV (x) weakly represents provability.

(3)

Henkin sentences: summary

If a canonical provability predicate (at least one satisfying the Lb

conditions) is chosen, all fixed points of this predicate are
equivalent.

(x) is a truth predicate for n iff for all sentences n :

()

There are provability predicates that have refutable and provable

Henkin sentences (that are self-referential in the sense of the
KreiselHenkin criterion).

In addition we may requite that the compositional axioms hold for n .

There is a refutable Henkin sentence obtained via the canonical

Gdel diagonalisation method.

0=0 and 0 =/ 0 are fixed points of each of these partial truth predicates.

For each n 1 there is a n -truth predicate, which is n .

Conclusions and questions

We can play the same tricks as in the case of provability:

theorem (Visser)
There is a truth predicate n (x) for the set of n -sentences so that the
truth teller formulated with n (x) using the standard diagonal function
d is provable in PA. There is also a truth predicate n (x) for the set of
n -sentences so that the truth teller formulated with n (x) and
standard diagonalisation d is refutable in PA.
The natural 1 -truth teller is refutable.

What happens to the Henkin sentences for canonical Rosser provability?

The properties of sentences saying about themselves that they are
provable, truth etc. does not only depend on how provability is
expressed but also on how these sentences given a representing
formula ascribe these properties to themselves.

Is the KreiselHenkin criterion for self-reference sufficient and/or

necessary for self-reference?
Can all the results (like refutable Henkin sentences) always be obtained
by tweaking the representing formula while canonical diagonalisation is
retained?
How serious should we be in metamathematics about the phrase say
about itself?
Is there any natural predicate with two fixed points that both satisfy
the KreiselHenkin criterion but differ in their properties?

Here is one very tentative proposal for an additional condition:

definition
A fixed-point operator is uniform iff the following condition is satisfied
for each (v):
f () is of the form ( f. ), where f. represents the function f .

Are there any reasonable additional or alternative conditions on top and

above the KreiselHenkin condition?

Sceptical doubts:
Perhaps saying that a sentence says about itself that its 1 -truth (Rosser
provable etc) is just a loose way of talking, which is imported from
natural language where we have personal pronouns for expressing de
se-claims.

Solomon Feferman. Arithmetization of metamathematics in a general

setting. Fundamenta Mathematicae, 49:3591, 1960.
Kurt Gdel. ber formal unentscheidbare Stze der Principia
Mathematica und verwandter Systeme I. Monatshefte fr
Mathematik, 38-38:173198, 1931.
Richard Heck. Self-reference and the languages of arithmetic.
Philosophia Mathematica, 15:129, 2007.
Leon Henkin. A problem concerning provability. Journal of Symbolic
Logic, 17:160, 1952.
Leon Henkin. Review of G. Kreisel: On a problem of Henkins. Journal
of Symbolic Logic, 19:219220, 1954.
Georg Kreisel. On a problem of Henkins. Indagationes Mathematicae,
15:405406, 1953.
Peter Milne. On Gdel sentences and what they say. Philosophia
Mathematica, 15:193226, 2007.
Craig Smoryski. The development of self-reference: Lbs theorem. In
Thoms Drucker, editor, Perspectives on the History of Mathematical
Logic. Birkhuser, Boston, 1991.