The Bulletin of Symbolic Logic

Volume 20, Number 1, March 2014

CIRCULARITY IN SOUNDNESS AND COMPLETENESS

RICHARD KAYE

The strings, my lord, are false.

Shakespeare, The Tragedy of Julius Caesar
Abstract. We raise an issue of circularity in the argument for the completeness of first-
order logic. An analysis of the problem sheds light on the development of mathematics, and
suggests other possible directions for foundational research.

§1. Introduction. In a recent discussion I took part in with a number of

philosophers engaged in investigations into the philosophy of mathematics
over dinner one evening, the question arose whether the argument for the
Soundness Theorem of mathematical logic is circular or not.1 This was
fairly rapidly dispatched by a number of people present, but I was surprised
by the reactions I received when I expressed concerns that the Completeness
Theorem might have circularity issues. My colleagues obviously thought
this a much weaker position to hold. This article attempts to express those
doubts and why they might matter.

§2. The Soundness Theorem. The notation Σ   will be used to express

the assertion that there is a formal proof of the statement  (typically a
string of symbols from a finite alphabet) from assumptions in Σ (typically a
set of strings from the same finite alphabet), where the proof obeys some for-
mally defined and precise rules for classical first-order logic. The conclusion
Σ   expresses that every interpretation making Σ true also makes  true,
where usually a particular restricted family of mathematical interpretations
is under consideration, and ‘true’ means in terms of Tarski’s definition of
Truth. The Soundness Theorem is the familiar theorem stating
Σ implies Σ  
for all such Σ, .
Received April 8, 2013.
1
Throughout this article I am interested in classical first-order logic, and I use Soundness
and Completeness, capitalized, to denote the standard theorems for this logic. In contrast,
soundness and completeness, without capitals, will denote the desirable properties of formal
systems that first-order logic or any other system of logic may have. It is possible that similar
points I make might be valid for other logics such as intuitionistic, though this has not been
investigated.


c 2014, Association for Symbolic Logic
0022-4812/14/2001-0003/$2.50
DOI:10.1017/bsl.2013.4

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern,24 on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 CIRCULARITY IN SOUNDNESS AND COMPLETENESS 25

It seems that the argument for this is circular, since knowledge of semantics
and whether or not Σ   must be founded in some system of argumentation,
which, if it is the system of  itself, quickly leads to infinite regress:

Σ implies  ‘Σ  ’
 ‘Σ  ’ implies  ‘ ‘Σ  ’’
 ‘ ‘Σ  ’’ implies  ‘ ‘ ‘Σ  ’’’

and so on. This is not unlike the point in Lewis Carroll’s What the Tortoise
Said to Achilles [2] and whilst not incorrect is not usually helpful.
The main issue in this article is with Completeness, but the matter of
soundness of mathematical theories is one that has received a considerable
amount of attention. Taken in the form given above, the Soundness Theorem
requires a metatheory that is able to express details of a formal system and
its notion of proof, and as usual we must have some confidence that the
theory described by the metatheory is the one we want to study. The proof
of Soundness goes by induction on the length of proof, so this must be
available. Most importantly we also need a (semantic) interpretation of the
theory and a notion of semantics for it. In the case when for example our
theory Σ is first-order Peano Arithmetic (PA) the metatheory must have
access to a model or interpretation of PA and access to a notion of truth
over this. It might be, but need not be, the standard natural numbers N with
the usual addition and multiplication operations.
As just described, it is preferable to regard the Soundness Theorem as
a ‘relative consistency result’ and see it as stating that provability in the
formal system is sound relative to the system required to set up the notion
of semantic entailment, . A certain amount of induction and recursion is
required for this argument: most obviously enough induction and recursion
must be available in the metatheory to carry out the syntactical operations
required and to perform the induction on the length of formal derivation.
Less obviously, Tarski’s definition of truth is a recursive definition, where
the recursion over formulas is necessary to enable a single  to apply to
all formulas of arbitrary quantifier complexity in the language. To be sure,
Tarski’s definition can be given to any finite stage without induction. Thus
without induction on quantifier complexity we can define truth for Σn for-
mulas for any fixed n, and prove in PA the consistency of IΣn (the subtheory
of PA formed by restricting the induction axiom to Σn formulas [10]).
There is a possibility that certain applications of Soundness may indeed
be circular in a different way to the Lewis Carroll type of problem. In view
of the implicit induction required for Tarskian semantics, there would be
a problem should Soundness be used as a means to justify a theory of
induction. It would seem to be a highly worthwhile project to identify the
inductive content (as opposed to the semantic content) of Tarski’s definition.
Note that both the semantic and inductive aspects of Tarski’s definition are
allied to the well-known increase in complexity as one looks at semantics for
an increasing number of quantifier blocks—an increase in complexity that

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 26 RICHARD KAYE

seems unavoidable and intractable (in the technical sense of the word) even
when one restricts interpretations to finite domains.
In the form given above, and even with a reasonably strong metatheory
accepted, it seems to me that the Soundness Theorem on its own does
not give credence to the commonly held assertion that if one accepts the
consistency of PA (for example) then one is entitled to accept the consistency
of PA + Con(PA). Even if one knows Con(PA) through knowledge of an
interpretation of PA it does not follow that that interpretation is (or is known
to be) -standard, and so it does not follow that it satisfies the arithmetized
sentence Con(PA). The precise details depend on our metatheory itself.
To take this further and to infer soundness of a system such as PA or
its extensions by semi-formal or informal means—and this is surely one of
the main aims of such foundational studies—is more difficult still. In an
excellent article, Dummett [3] discusses the issue of soundness thoroughly.
He concludes that there may be an idealist (or constructivist) view by which
progress can be made. An alternative but not necessarily contradictory posi-
tion is that by examination and checking of a great deal of cases, and also
perhaps by measurement in the physical world and grounding ones intu-
itions on these measurements, one might conclude (in the sense perhaps of
Popper’s Scientific Discovery [14]) that the best and most useful explanation
is indeed the existence of an interpretation of PA satisfying the reasonable
mathematical properties we expect.
Thus the notion of semantic entailment itself might be given informally
or else might be given in some more formal or semi-formal sense. As it
happens, mathematicians are, rightly or wrongly, typically very comfortable
with a semantic notion of Truth based on mathematical structures. The
process presented by the proof of the Soundness Theorem (and this process
is a straightforward recursion on the length of a proof) can be mimicked in
any such situation, and its validity checked by other informal means. Indeed
this is often done in mathematical practice, and as such provides supporting
evidence of the correctness of the theorem.
It should be added that it would seem that, from a mathematical perspec-
tive, the Soundness Theorem is not of much practical use, since the existence
of a formal proof of  from Σ may be more compelling than the manipu-
lation of this proof into semantic arguments why  should follow from Σ.
Put another way, Soundness is actually comparatively weak as a piece of
mathematics. Nevertheless, despite the apparent limitations of Soundness
when seen in this way, it remains one of the most important and useful
results in mathematical logic. Given statements Σ and  it may be difficult
to see why there is no formal argument of  from Σ from the point of view
of direct arguments with syntax, even though this is the subject matter of
the assertion ‘Σ  ’ and the search for such an argument is the natural
and direct approach. But instead, constructing an interpretation of Σ in
which  fails, and invoking the contrapositive of the Soundness Theorem is
commonly the more successful approach. For example, although there are
now methods of ordinal analysis of proofs that allow us to show in a direct
way from arguing about proofs that PA  PH, where PA is first-order Peano

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 CIRCULARITY IN SOUNDNESS AND COMPLETENESS 27

arithmetic and PH is the Paris–Harrington statement [13], the original argu-

ment for ‘PA  PH’ was conceived in a semantic way, and the original proof
was an indirect one via Soundness.2 Thus there remains a puzzle relating
to Soundness that, although it is rather weak mathematically and presents
little extra information, it is in practice rather useful.

§3. The Completeness Theorem. It is my contention that the converse

result, the Completeness Theorem
Σ implies Σ  ,
is much more problematic in its correct interpretation, where as before  is
provability in normal first-order logic.
It is not a problem to set up the notions of first-order provability
and semantic entailment via Tarski’s definition of Truth in one’s favorite
metatheory—set theory, say—and prove the above result in that system.
Moreover, since set theory itself is not at the outset intended to be a meta-
mathematical theory, but rather a theory of mathematical objects, the proof
of Completeness in set theory gives genuine new information. The issue
is, as it was for Soundness, whether some of the familiar foundational
consequences normally drawn from Completeness are reasonable. First
and foremost of these is the question whether the Completeness Theorem
supports the assertion,
First-order logic is complete.
By this I mean that it is complete for ordinary mathematics as is commonly
taught in a typical pure mathematics department in a University—in the
U.K. or U.S.A. perhaps. Mathematicians in such departments do their work
with whatever means are available to get the results they need. In partic-
ular, the law of excluded middle is always assumed where needed, as are
nonconstructive arguments using the Axiom of Choice. Additional assump-
tions, such as the continuum hypothesis (or its negation) are very rarely
assumed. In some sense there is (in mathematics departments) a consensus
that mathematics is precisely what is described by ZFC.
My focus is not the usual one of concern that first-order logic is too strong
(and some other logic, possibly a constructive logic based on intuitionistic
logic should be preferred) but rather on whether first-order logic is strong
enough for practical mathematics, and whether there are not in fact other
new logical principles that do not follow from first order logic that should
be accepted.
The argument for first-order logic in ‘ordinary’ mathematics is rather com-
pelling. If I am a mathematician working in, say, group theory, interested
in properties of groups that can be expressed in the language of first order
logic, I am comfortable with the idea that I need to look for theorems (with
2
It takes us off topic slightly, but it might be argued that the statement PH itself arose from
semantic considerations, so that the Paris–Harrington Theorem may not have even been
discovered at the time except by semantic means.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 28 RICHARD KAYE

proofs of some kind) or counter-examples. As mentioned, most mathemati-

cians are perfectly happy with the idea of counter-examples, i.e., structures
such as groups and a notion of semantics attached to them, corresponding
to Tarski’s definition of Truth. The Completeness Theorem states that if
I cannot find a counter-example then there is a proof of the relevant theo-
rem, not just in some general informal mathematical terms, but in the formal
first-order language following the formal rules. If there are no such proofs
it even provides a means for constructing a counter-example. Thus for most
mathematicians one can reduce one’s activity to finding counter-examples
or finding first-order proofs.3
It is this conclusion that is normally drawn from Completeness, and it
is precisely this argument that I suggest is circular. The issue is important
because, by the reduction to first order logic just provided, mathematicians
in other disciplines (such as in set theory, or some areas of real analysis
where there are problematic questions such as the continuum problem, or
more generally what subsets of the reals are available, and whether they are
all measurable, etc.) need only look for first-order principles that settle these
questions, and there is the expectation that this is a helpful reduction. But
it may be that this is based on erroneous reasoning.
My counter-argument here is based on the fact that the Completeness The-
orem is not proved in the form given above, but rather in its contrapositive
form,
Σ  ⊥ implies Σ  ⊥.
In other words, one proves (using set theory or whatever foundational system
one has as the metatheory) that if a set of sentences Σ is consistent, i.e. Σ  ⊥,
then there is a structure or interpretation of Σ, that is Σ  ⊥ holds.4
There are two problems. The first is a mathematical one that complicates
the issue but must be given careful examination, and is that the interpretation
of Σ is not given constructively but typically requires a weak form of the
Axiom of Choice (more precisely, the Boolean Prime Ideal Theorem, BPIT,
or a statement equivalent to this) to do its work. The second is more pertinent
and is the observation that the metatheory (some flavor of set theory or
arithmetic, including the BPIT) begs the question of what interpretations
or structures should be available, by presenting them all in advance. If one
did not know the rules for first-order logic one would be able to discover
them by working through the proof of Completeness and at each step, when
some formal proof is required, writing down the simplest formally Sound
rule that justifies the step in question.
Thus the proof of the Completeness Theorem simply fits some logic to
the (known, or implied) class of interpretations or structures given by the
metatheory, and checks that the rules of that logic are adequate for this class
3
To be sure there are important ways of ‘speeding up’ first-order logic to ensure that proofs
are quicker to find, shorter to write down and/or clearer to understand, but the logic with
these extra speed-up principles added is still equivalent to first-order logic.
4
I am assuming of course that the target logic has standard rules for reductio ad absurdum
and for the false statement ⊥.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 CIRCULARITY IN SOUNDNESS AND COMPLETENESS 29

of interpretations. Our metatheory is typically a set theory or arithmetic with

BPIT in some form, though with additional assumptions on our language
and axioms Σ may be taken to be weaker than this—see below. This metathe-
ory is of course based on first-order logic (though the first-order nature of
this logic is sometimes hidden by appealing to multi-sorted languages such
as that of second-order arithmetic as if some higher order logic is being
employed). Our conclusion that first-order logic is complete for ordinary
mathematics is then seen to be based on a circularity, that first-order logic
is complete for the class of interpretations in a possibly artificially devised
class based on a metatheory based on first-order logic.
To illustrate the problem, suppose a mathematician rejected the idea of
uncountable structures and insisted that all countable structures should
have a recursive (i.e. computable) presentation.5 Such a mathematician
would have a reasonable and self-consistent notion of semantic entailment,
Σ Rec , meaning that  holds in all recursively presented models of Σ. He
would presumably reject ZFC as a metatheory, as it leads to unacceptable
consequences such as the existence of nonrecursive interpretations, but the
logic based on Σ Rec  is not unreasonable. For example it is consistent,
i.e. Σ Rec ⊥ for a great number of Σ including of course the empty set,
since there are recursive models available. This mathematician could accept
all rules for first-order logic as being sound, since the Soundness Theorem
still holds in the form Σ   implies Σ Rec , where  means for the
rules of first-order logic. He would not however accept that first-order logic
is complete, since for example Gödel’s Second Incompleteness Theorem
PA  Con(PA) for PA in first-order logic can be given adequate syntactic
arguments, but PA Rec Con(PA) since no model of PA + ¬Con(PA) is
recursive by Tennenbaum’s result [20].6
Recursive models have been a theme of mathematical logic since at least the
1950s, and a number of results aim at highlighting the distinctions between
 and Rec . For a discussion that brings many of these ideas together and
starts to explore the model theory of Rec in a more systematic way, see
Stolboushkin [19]. There are a number of interesting technical questions
related to the entailment notion Rec and I will return to this later.
As an even more extreme example, but one that is certainly of use in
computer science, finite model theory is the model theory of finite structures
with associated semantics Fin . For this theory we even have PA Fin ⊥.
I do not hold that finite models, nor even recursive models, are the only
interesting ones, and certainly I would not myself restrict the mathematical
universe in such a way: the examples are given to illustrate the consistency
of the view that first-order logic is not complete and thus demonstrate the

I.e. should have domain N with all predicates and functions recursive.
5
6
This argument can easily be adapted to show that, working in one of the usual metatheories
such as ZFC or a fragment of second-order arithmetic, there is no recursive system of proof
Rec that is Sound and Complete for Rec . One needs only modify the proof of the Gödel
Incompleteness Theorem for a Sound and recursive Rec to show that PA  ConRec (PA)
where ConRec means in terms of Rec . But of course PA Rec ConRec (PA) holds for the same
reason.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 30 RICHARD KAYE

inadequacy of the argument purporting to show that it is. Nevertheless, the

collection of recursively presented interpretations and its semantics is very
natural. It might conceivably be the case that some flavor of mathematics
with a different class of interpretations for our logical languages is more
appropriate for scientific work in understanding our physical universe, and
in particular then some stronger logical principles should be adopted. But
I have no sensible suggestions in this regard.
Looking at the argument given above, what we see is that it is more accu-
rate to see the Completeness Theorem as having presented mathematicians
with a self-consistent class of examples and counter-examples, rather than
to see it as justifying a system of proof. It seems highly unlikely that mathe-
maticians before 1930 (with the possible exception of Cantor) had any clear
conception of what sets or structures or interpretations might be admissible
to mathematics, and therefore the more accurate view of the historical devel-
opment of the subject is that first-order logic and the Completeness Theorem
in particular and the set theory needed to prove it in its full generality were
developed alongside each other, and together they provide mathematicians
with a clear conception of what might or might not constitute a mathematical
interpretation or structure. Furthermore, without a thorough investigation
into the foundational issues of the axioms of set theory and the Axiom of
Choice particularly, general techniques for transfinite induction and recur-
sion, Zorn’s lemma and other nonconstructive methods may never have been
accepted into the mathematical mainstream.

§4. Towards an analysis of completeness. It is well known that the Com-

pleteness Theorem for propositional logic in finitely many propositional
letters,
Σ  ⊥ implies there is a valuation making each  ∈ Σ true
(where Σ is a finite set of propositional formulas) is provable directly by con-
structive means. Perhaps the most straightforward argument uses tableaux.
If Σ  ⊥ then there is a completed tableau from Σ in which every formula
is subdivided into its atomic subformulas, and this is possible because Σ
is finite and so there are only finitely many subformulas to consider in the
propositional logic. Then one of the paths yields an appropriate valuation,
or else the whole tableau can be converted to a proof of ⊥ in Σ. The same
argument also constructively shows
no valuation making each  ∈ Σ true makes  true implies Σ  ¬,
for any completed tableau for Σ,  has all its paths closed, and this tableau can
be constructively converted to a proof showing Σ  ¬. Hilbert [9], Post [15],
and Bernays [1] all had proofs of the completeness of propositional logic in
this sense. See particularly Zach [21, p. 340] who traces Hilbert’s formulation
of completeness from
‘whether the axioms suffice to prove all “facts” of the theory in
question’ (1905)

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 CIRCULARITY IN SOUNDNESS AND COMPLETENESS 31

to

‘[whether] we always obtain an inconsistent system of axioms by

adding a formula which is so far not derivable,’ (1917–18)7

statements that are clearly ‘absolute’, i.e., not relative to some particular
choice of metatheory. In particular, Hilbert does not (and nor would we
expect it of him in 1905) suggest that the collection of “facts” to be captured
is dependent on a choice of position or metatheory from which one views
those “facts”.8
There is an issue of nonconstructivity in the Completeness Theorem for
propositional logic with infinitely many propositional letters, however. Our
mathematician who accepts only recursive objects will believe that propo-
sitional logic is not complete. To see this one needs only recall that there
are recursive trees of infinite 0, 1-branching trees with no recursive infinite
path. It is then easy to use such a tree to set up a consistent recursive set of
propositional formulas Σ with no recursive valuation. More generally, the
status of the Completeness Theorem for propositional logic (in the sense of
Reverse Mathematics) is that it is equivalent to WKL0 over RCA0 , where
RCA0 is the base system of Reverse Mathematics, the acronym standing for
‘Recursive Comprehension Axiom’, and WKL0 is the system obtained by
adding to RCA0 an axiom for Weak König’s Lemma—see e.g., Simpson [18].
Gödel’s 1930 proof [8] of the Completeness Theorem for first-order logic
follows Skolem’s technique of using what are today called Skolem functions
to reduce formulas to a simple Prenex Form, and then apply what is essen-
tially Completeness for propositional logic in its form with infinitely many
letters. The more familiar Henkin argument proceeds in a similar way, using
a family of (Henkin-)constants in place of Skolem functions. In the case
when the set of sentences Σ one wishes to provide an interpretation for is
presented in a recursive way suitable for investigation in a metatheory based
on arithmetic, the system WKL0 based on the König tree argument suffices
to prove Completeness. In more general cases (for example when there is
no arithmetization or Gödel numbering of syntax) the BPIT is required.
I should also mention that the issue of nonconstructivity for Completeness
in WKL0 is no longer one concerning the Axiom of Choice: merely that
non-recursive sets are required.
To take the discussion further we will need to be more precise as to what the
issues are and what reasonable starting assumptions could be. We will look
at logical statements in various languages or signatures, and to prevent the
7
Zach’s translations.
8
The improved statement of 1917–18 follows Hilbert’s thinking of the time and the devel-
opment of his Programme, but is slightly misleading here. My understanding of Hilbert’s
proof of the completeness of propositional logic is that a statement φ(p, q, . . .) in proposi-
tional logic should be interpreted as a universal, ∀p, q, . . . φ(p, q, . . .), with p, q, . . . thought
of as ranging over the two truth values, and this only allows Hilbert to give a proof of what
we would now call Completeness for propositional logic in the case of finitely many propo-
sitional letters. However, with propositions interpreted in this way, the new formulation of
completeness for propositional logic is correct.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 32 RICHARD KAYE

discussion becoming too broad I continue to restrict these logical statements

to the syntax and scope of first-order logic, and allow that truth for these
logical statements is defined in a way very similar to (if not exactly the same
as) Tarski’s definition of Truth. I take it that there is a potential collection
of sets, models, and interpretations for these logical statements. I want to
be a little more flexible about these: they may come from a Platonistic
account of mathematics, and thus correspond to ‘existing abstract objects’;
they may come from a physical description of the universe (or part of it)
and thus correspond to ‘existing physical objects’; or they might be virtual
or ideal objects posited by a (formal or informal) theory. The universe of
interpretations and our arguments about them will be referred to as the
metatheory. This, together with the definition of Truth, gives rise to a notion
of semantic entailment , called true semantic entailment. The problem is to
identify the nature of . We do have a system of proof for first-order logic,
FO , and the statement that first-order logic is complete is
Σ implies Σ FO .
The Completeness Theorem (formulated in ZFC) is the statement
Σ ZFC  implies Σ FO 
of ZFC, where I have placed a subscript on  to indicate that this notion
of semantic entailment is the one internal to ZFC and may not correspond
to true semantic entailment. Other weaker theories (such as ACA0 , WKL0 ,
PA) also prove versions of the Completeness Theorem, sometimes restricted
to arithmetized theories and sets of sentences.
At this point, it is worth looking briefly at Kreisel’s squeezing argu-
ment [11]. As summarized by Field [5, Chapter 2], it involves two separate
notions of ‘validity’ or semantic entailment: an intuitive notion which we
can identify with our true semantic entailment  and a formalized notion
ZFC . The difficulty is to bring these into focus and at least show they are
extensionally identical,
Σ ZFC  ⇔ Σ  .
The point is that we may use the Completeness Theorem in the form ‘Σ ZFC
 implies Σ FO ’, and the Intuitive Soundness Theorem in the form
‘Σ FO  implies Σ  ’ (both of which being available and justifiable) to
obtain the full equivalence of ZFC and  from the direction
Σ   ⇒ Σ ZFC . (1)
Although there may be intuitive arguments for (1) from what is now tra-
ditional mathematical practice, this implication is precisely the one that is
being disputed and investigated in this article.
The model theory and proof theory of the suggested metatheories, ZFC,
ACA0 , WKL0 , PA, etc., should and does play and important role in explor-
ing and clarifying the foundational questions here, and it is quite impossible
to survey all the many positive contributions made in this direction in a
short article. I will point to two examples that highlight the nature of some
of these contributions as well as raising questions of their own.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 CIRCULARITY IN SOUNDNESS AND COMPLETENESS 33

My first example concerns whether there is a Whitehead group9 which is

not free [17]. This is the Whitehead problem, and I heard about it first (as an
undergraduate) from an algebraic topologist who had been working on an
equivalent problem for quite a few days without realizing the connections
or the contribution due to Shelah. Specifically, Shelah showed that no such
groups exist in models of ZFC + V = L but the existence of such groups
follows from Martin’s axiom plus the negation of CH.
The question this (and many other similar results in set theory) raises, is
what results like this say for our notion of true semantic entailment? We tend
to think of  and mathematical truth in absolute terms. Is this reasonable?
Results of the type just given are not enough in themselves to allow any
such groups to be allowable structures in our universe, and even though this
is a question that is often raised as part of general mathematical research
experts in Foundational Studies do not (currently) offer any further guid-
ance other than the observation that additional means going beyond ZFC
will be required to solve the problem. Many important contributions from
set theory are often similarly complicated and potentially confusing to the
non-expert. Most working mathematicians (including my topologist friend)
therefore duck such issues, in this case by accepting that ‘the Whitehead
problem is independent of the axioms of mathematics10 .’ This seems a great
pity, and a preferable state of affairs would be to present the matter in a way
that is palatable to the nonlogician and in such a way that she or he can and
might want to contribute to the problems instead of avoiding them.
In terms of our true semantics and what structures should play a part in it,
set theoretic research, dealing with questions of absoluteness for example, is
highly relevant. There are questions such as: which interpretations internal
to some model of set theory can be regarded as interpretations in their own
right?
My second example is that of the metamathematics of WKL0 itself. When
one looks at minimal theories required for the Completeness Theorem one
can restrict attention to first-order theories which are given in a recursive
way via some arithmetical notions such as Gödel numbering. I will assume
that these notions of Gödel numbering and the particular presentation of
the non-logical axioms are clear, unambiguous, and correct for the intended
application, though of course one could explore these ideas further and
no doubt find exceptional cases which are problematic. As we have seen,
when such a first-order theory is so-described and is known or assumed
to be consistent, then models of this theory can be given, working in the
metatheory WKL0 .
Various question then arise. In particular, what is the constructive content
of the Completeness Theorem as proved from WKL0 ? This is precisely the
‘unwinding’ programme of Kreisel in the particular case of a specialization

The details are unimportant for this discussion, but an abelian group A is Whitehead if
9

whenever B is abelian and f : B → A is a surjective group homomorphism with ker f ∼ = Z,

there is g : A → B with fg = idA .
10
By which is meant ZFC, of course.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 34 RICHARD KAYE

of the Completeness Theorem—see Feferman [4]. Quite a lot is known

about this, including the initial work by Scott on what are now called Scott
sets [16] (which are the second-order part of -models of WKL0 ), and
their rediscovery and application by Friedman [6] to models of arithmetic.
(See Simpson [18], especially Chapter IV, for aspects of this in second-
order arithmetic and Kaye [10], especially Chapter 13, for applications to
first-order arithmetic.)
It may help to observe that in the proof (in WKL0 , say, or by a direct
tree argument) of the Completeness Theorem for a recursive theory, when
presented carefully, only one recursive tree needs to be traversed to find
an infinite path in it. By additional coding, one may present the proof of
Completeness for all recursive first order theories using a single infinite
path from a particular recursive binary tree encoding all such theories.11
A recursive binary tree that has no recursive infinite path necessarily has
several infinite paths.12 It is precisely this that makes the WKL0 axioms
analogous to Choice, though they are of course provable directly. What
might our recursive model theorist make of this?
If he could be persuaded to accept countable models with domain N and
finitely many relations in some Scott set X , and this is a rather large ‘if’
since there are always nonrecursive models in such an X , he would no doubt
be willing to accept a model (N, R) where R is recursive relative to one of
the models in X . In other words, his collection of models is X itself.
The theory WKL0 and its models, the Scott sets, are precisely the universes
where Completeness and Soundness hold. Thus if X is a Scott set and X
is semantic entailment relative to models in X we have, for each recursive Σ,
the set Σ is defined by a Δ01 formula of WKL0 and

Σ X  ⇔ (N, X )  ‘Σ  ’ ⇔ (N, X )  ‘Σ  ’ ⇔ Σ  .

The first bi-implication here is a restatement of X , the second is by

Completeness and Soundness in WKL0 , and the third follows from the
fact that (N, X ) is an -model, so a proof in this model has a standard
Gödel-number; then (as observed by Gödel) the usual syntactic notions of
Gödel-number, formula and proof, etc., are all described by primitive recur-
sive functions and so mean the same thing in (N, X ) as in the metatheory
(i.e., the notion of first-order proof is absolute) hence the third bi-implication
above holds. This is the key observation by Scott and the central reason for
the importance of Scott sets: that they do indeed describe precisely the
constructive content of the Completeness Theorem.
11
The related result due to Scott that Scott sets are completion closed [10, Theorem 13.3]
explains the idea.
12
This is because if the tree is recursive and has a unique infinite path then this path can be
found by the algorithm which operates as follows. At node n, the algorithm simultaneously
searches the tree below both daughter nodes to check which of them is finite. One is finite,
by our assumption that there is precisely one infinite path through the tree, and so can be
traversed completely, so in one case the algorithm returns the answer. When this happens,
the algorithm takes the non-finite daughter node and repeats the process.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 CIRCULARITY IN SOUNDNESS AND COMPLETENESS 35

This analysis may not be wholly satisfactory to our recursive mathemati-

cian, however. He might be sufficiently uncomfortable with the need for
nonrecursive sets to ask for further justification. Given a recursive Σ, he asks
is there some canonical model that interprets Σ? And here our answer will
disappoint.
Given a nonrecursive set A ⊆ N there is always a Scott set X such that
A ∈ X [10, Lemma 11.2] (see also Simpson [18, Chapter VIII]). Thus
semantic entailment for recursive models, Rec has some claim to being the
weakest logic for which every model is canonical in some sense. It would
be of particular interest to explore these directions further and explore the
relative merits of Rec and X for Scott sets X in more detail.
Before I leave WKL0 there is one other remarkable result concerning it of
foundational interest that should be mentioned, and that is the Harrington–
Friedman result that WKL0 and RCA0 share the same Π11 consequences
[18, Chapter IX]. This result can be restated and proved using techniques
from PA too: The theories WKL0 , RCA0 and their first-order counter-
part IΣ1 [10, Section 7.1 and Chapter 10] all have the same first-order
consequences, and are all Π02 conservative over primitive recursive arith-
metic, PRA. It is arguable that PRA is a reasonable base for mathematics
(i.e. for finitistic methods, in the sense of Hilbert).13 The conservation result
relating WKL0 to PRA can be proved by direct proof-theoretic means in
PRA itself: one can show (in PRA) that if p is a proof from WKL0 of 
where  is Π02 , then there is a way of manipulating this proof p into a proof
of  from PRA itself. In the special case where  is an absurdity ⊥, this gives
an equiconsistency result Con(PRA) ⇒ Con(WKL0 ) in PRA. This is one
of the very few places where Hilbert’s programme can be realized success-
fully, and this in itself gives good grounds to accept WKL0 as consistent and
perhaps appropriate. If we are able in the future to justify completeness of
first-order logic based on the proof of Completeness in WKL0 and certain
intuitive properties of the true semantic entailment relation, it seems to me
that these properties of WKL0 will most likely be highly relevant.
Of course all the discussion here has been based on the notion of true
semantic entailment and the class of structures we admit in the universe.
The other aspect of circularity in the inference of completeness from the
Completeness Theorem is the logic that we use in our metatheory. Unpicking
this is likely to be a much more delicate task: our informal arguments in the
metatheory are (rightly or wrongly) based on conditioning derived from
mathematical practice, and from first-order logic in particular, so justifying
first-order logic by logical arguments in the metatheory is fraught with
difficulty. The theories we have been examining (WKL0 , RCA0 , IΣ1 , and
PRA) are all based on first-order logic and the conclusion we wish to draw
is an absolute one concerning first-order logic. Whether this is a problem
13
But there really is an issue here: PRA provides functions that grow at rates corresponding
to the individual levels of the Ackermann function, and so certainly defines functions that
are intractable in the sense of complexity theory and much more besides. It is not a priori
obvious that our intuitions for such functions are sufficient to guarantee the consistency and
relevance of PRA.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 36 RICHARD KAYE

will depend on one’s views of how mathematics is ultimately grounded, and

quite possibly also what mathematics is for. Nevertheless, even if one were
to accept usual practice in such informal arguments, one should accept that
there is an assumption here, or an argument to be made, and that others
with different starting points may have a different opinion.
I close this section with a brief mention of another well-known result
‘characterizing’ first-order logic, and offer its analysis (in the context of the
discussion here) as a worthy programme of study. Lindström’s Theorem [12]
is (rightly) regarded as a highlight of mathematical logic characterizing
first-order logic. Essentially, it says that first-order logic is unique amongst
abstract logics as being the only one enjoying certain basic properties, chief
amongst them being countable compactness and the downward Löwenheim
Theorem. Lindström’s Theorem ‘suffers’ (and I do not mean this in any
pejorative sense) the same issues when it is regarded as a ‘justification’ for
first-order logic. The processes used in the proof of Lindström’s Theorem
are very much in the realm of first order logic with a class of allowed models
(implicitly) given. And it is not so surprising, perhaps, that the theorem says
that this class of models corresponds to first-order logic itself in some strong
way. Once again, the theorem can, and does, give significant foundational
insight, through its mixture of internal and external arguments, including in
particular (externally) its view of all interpretations and the true semantic
entailment, and (internally) its application of back-and-forth systems and
isomorphisms.

§5. Conclusions. First-order logic, with its Completeness and Soundness

Theorems, is a remarkable achievement. It is certainly true that the argu-
ments about it, Completeness and Soundness in particular, show it to be
robust and natural and deserving of our attention. It is not the case that the
Completeness of first-order logic shows it to be necessarily strong enough
for any particular application, however, except where the metatheory can
be taken to be a particular flavor of set theory or arithmetic which proves
the theorem. Thus, some of the claims for first-order logic (including many
I myself have made in the classroom) have to be treated with considerable
care, to say the least.
As a foundational device, first-order logic (alongside other logics) provides
a framework for discussing our foundational issues in mathematics, and
again it is invaluable in this respect. But it may be wise not to allow the
Completeness and Soundness Theorems for first order logic to mislead us
into thinking that first-order logic necessarily captures usual mathematical
practice in every respect. In Section 4 particularly I have suggested further
avenues for investigation into just these matters.
Having said this, the circularity of the argument that takes us from Com-
pleteness for first-order logic to the completeness of first-order logic has its
beneficial side too. Whatever structures one might have originally accepted
in the mathematical universe, for example whether one originally might have
accepted the existence of nonstandard models of arithmetic which are neces-
sarily nonrecursive, techniques from the proof of the Completeness Theorem

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 CIRCULARITY IN SOUNDNESS AND COMPLETENESS 37

and more generally the use of set theory as a standard all-encompassing

theory for mathematics force us to extend what might have been a rather
limited mathematical view, to examination of structures built by noncon-
structive means. First-order logic, even if it was not the ‘original intended
logic of mathematics’ (if that phrase means anything at all), becomes the
practical logic of modern mathematics through its interrelations with set
theory. Those interrelations, I hope to have convinced the reader, beg ques-
tions, in a way that should properly be considered as a species of circulus in
probando. But the resulting shift in emphasis that arises is highly positive,
and asks how one can profitably use the new structures and tools at our
disposal.
These issues also shed some light on the development of mathematical
thinking in the twentieth century, and the acceptance of set theory as a
suitable framework for mathematics, as new existence principles became
available and more widely accepted. Hand in hand with this was the need to
understand these principles and their limitations, and first-order logic played
an intimate role in both aspects of this development, as did set theory in the
development of first-order logic.
REFERENCES

[1] Paul Bernays, Beiträge zur axiomatischen Behandlung des Logik-Kalküls. Habilitattionsschrift,
Universität Göttingen, 1918.
[2] Lewis Carroll, What the tortoise said to Achilles. Mind, vol. 104 (1995), no. 416, pp. 691–693.
[3] Michael A. E. Dummett, The justification of deduction, Truth and Other Enigmas, British
Academy Lecture, 1973, Duckworth, 1978, pp. 290–318.
[4] Solomon Feferman, Kreisel’s “unwinding” program, Kreiseliana (Piergiorgio Odifreddi, editor),
A K Peters, Wellesley, MA, 1996, pp. 247–273.
[5] Hartry Field, Saving truth from paradox. Oxford University Press, Oxford, 2008.
[6] Harvey Friedman, Countable models of set theories, Cambridge Summer School in Mathematical
Logic, Cambridge, 1971, Lecture Notes in Mathematics, vol. 337. Springer, Berlin, 1973, pp. 539–573.
[7] Kurt Gödel, Über die Vollständigkeit der Axiome des logischen Funktionenkalküs. Monatshefte
für Mathematik und Physik, vol. 37 (1930), pp. 349–360.
[8] Kurt Gödel, The completeness of the axioms of the functional calculus of logic, From Frege to
Gödel. A source book in mathematical logic, 1879–1931 (Jean van Heijenoort, editor), Harvard University
Press, Cambridge, MA, 1967, pp. 582–591. Translation of Gödel [7].
[9] David Hilbert, Prinzipien der mathematik, Lecture notes by Paul Bernays. Bibliotek, Mathema-
tisches Institut, Universität Göttingen, 1917-8.
[10] Richard Kaye, Roman Kossak, and Tin Lok Wong, Adding standardness to nonstandard
arithmetic, New studies in weak arithmetics (P. Cégielski, Ch. Cornaros, and C. Dimitracopoulos, editors),
CSLI Lecture Notes number 211, CSLI Publications, Stanford, 2014, pp. 179–197.
[11] Georg Kreisel, Informal rigour and completeness proofs, Problems in the Philosophy of
Mathematics (Imre Lakatos, editor), Problems in the Philosophy of Mathematics. North-Holland, 1967.
[12] Per Lindström, On extensions of elementary logic. Theoria, vol. 35 (1969), pp. 1–11.
[13] Jeff Paris and Leo Harrington, A mathematical incompleteness in Peano arithmetic, Handbook
of Mathematical Logic, North-Holland Publishing Co., Amsterdam, 1977, pp. 1133–1142. Edited by
Jon Barwise, With the cooperation of H. J. Keisler, K. Kunen, Y. N. Moschovakis and A. S. Troelstra,
Studies in Logic and the Foundations of Mathematics, Vol. 90.
[14] Karl R. Popper, The logic of scientific discovery, Hutchinson and Co., Ltd., London, 1959.
[15] Emil Post, Introduction to a general theory of elementary propositions. American Journal of
Mathematics, vol. 43 (1921), pp. 163–85.
[16] Dana Scott, Algebras of sets binumerable in complete extensions of arithmetic, Proceedings of
Symposia in Pure Mathematics, vol. V, American Mathematical Society, Providence, RI, pp. 117–121,
1962.

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4
 38 RICHARD KAYE

[17] Saharon Shelah, Infinite abelian groups, Whitehead problem and some constructions. Israel
Journal of Mathematics , vol. 18 (1974), pp. 243–256.
[18] Stephen G. Simpson, Subsystems of second order arithmetic, second ed., Perspectives in Logic.
Cambridge University Press, Cambridge, 2009.
[19] Alexei P. Stolboushkin, Towards recursive model theory, Logic Colloquium ’95 (Haifa), vol. 11
Lecture Notes Logic, Springer, Berlin, 1998, pp. 325–338.
[20] Stanley Tennenbaum, Non-archimedian models for arithmetic. Notices of the American
Mathematical Society, vol. 270 (1959), p. 270.
[21] Richard Zach, Completeness before Post: Bernays, Hilbert, and the development of propositional
logic, this Journal, vol. 5 (1999), no. 3, pp. 331–366.

SCHOOL OF MATHEMATICS
UNIVERSITY OF BIRMINGHAM
BIRMINGHAM B15 2TT, UK
E-mail: [email protected]

Downloaded from https://www.cambridge.org/core. Universitätsbibliothek Bern, on 29 Mar 2020 at 11:06:22, subject to the Cambridge Core terms of use,
available at https://www.cambridge.org/core/terms. https://doi.org/10.1017/bsl.2013.4