On interpretations of arithmetic and set theory∗

Richard Kaye and Tin Lok Wong

School of Mathematics
University of Birmingham
Birmingham
B15 2TT
U.K.
15 July, 2006

1 Introduction
The work described in this article starts with a piece of mathematical ‘folklore’ that is
‘well known’ but for which we know no satisfactory reference.1
Folklore Result. The first-order theories Peano arithmetic and ZF set theory with the
axiom of infinity negated are equivalent, in the sense that each is interpretable in the
other and the interpretations are inverse to each other.
This would make an excellent starting point for any beginning research student
working in the area of models of Peano arithmetic, since an understanding of how
these interpretations work enables all coding techniques from set theory to be em-
ployed in arithmetic in a uniform way, and also places PA ‘on the map’ (in the sense of
consistency and interpretation strength) relative to set theory. This, combined with the
lack of suitable references, permits us, we trust, to omit any apology for investigating
such folklore. In fact, when the details were finally uncovered, there were surprises for
both authors—this is the reason for the quotation marks when we said that the folkloric
result above is ‘well known’.
We should be more precise and highlight exactly where the imprecision in the folk-
loric result lies. There are two places: firstly, the notion of ‘ZF set theory with the
axiom of infinity negated’ turns out to be dependent in quite an important way on the
preferred choice of the initial presentation of ZF; secondly, the idea of interpretations
∗ This is a slightly expanded version of a paper to appear in the Notre Dame Journal of Formal Logic

under the same title. We have taken the opportunity in this web version of the paper to add a few additional
comments and references to some additional material.
1 Further references. Świerczkowski [20] wrote a detailed exposition of the equivalence of PA and “the

theory of hereditarily finite sets”, which uses an additional adduction operator. This theory of sets is said be
adapted from Tarski and Givant [25] with slight variations. The adduction operator C, defined by
x C y = x ∪ {y},
is also studied by Kirby [13] in the context of ZF without the axiom of infinity.
Another theory of sets that was proved to be “equivalent” to PA is PS introduced by Previale [17]. How-
ever, the interpretations there seem to be the Ackermann one and the ordinal one (cf. Kirby [13, §2] and
Sections 3, 4).

1
being ‘inverse to each other’ admits important variations. For example, Chang and
Keisler [5, §A.31] specify one particular choice of axiomatisation of ZF; for this ax-
iomatisation a weak form of interpretation-equivalence of ‘ZF with infinity negated’
and PA can be proved, but for stronger notions of interpretation-equivalence a different
axiomatisation of ZF seems to be required.
Our notation and terminology is standard. For a text on set theory we have used
Drake [7], though many others would be suitable. For background on Peano arith-
metic we refer the reader to Kaye [12], and especially Chapter 10 for background on
subsystems of PA.

2 Interpretations
The first issue to address here is what we shall mean by ‘interpretation’ and when
interpretations are ‘inverse’ to one another. In fact, for the folkloric result we are ad-
dressing, the notion of interpretation required is very straightforward and concrete, but
in general the word ‘interpretation’ is used in many contexts and with many different
meanings in logic and model theory.
As pointed out elsewhere, notably by Visser [26], notions of interpretation can be
best brought together using a category theoretic framework, though in this paper we use
category theory simply as a notational device. For us, an interpretation i of a theory T2
in a theory T1 is a morphism i : T2 → T1 in a particular category being studied.
To define the particular category we will use in this paper, let us agree that an
L -theory is a consistent set of L -sentences (not necessarily complete or closed un-
der deduction) and we make the convenient assumptions that all languages are purely
relational, equality (=) being one of the relation symbols, and a further unary rela-
tion Dom() for the domain is always present. Full first-order logic is assumed, includ-
ing all the rules for equality and the additional logical axiom ∀x Dom(x). For us here,
an interpretation i : T2 → T1 of an L2 -theory T2 in an L1 -theory T1 is given by a map-
ping of atomic formulas R(x1 , x2 , . . . , xn ) of L2 to formulas R(x1 , x2 , . . . , xn )i of L1 in
the same free variables. In particular, the domain and equality on this domain are de-
fined by Dom(x)i and (x = y)i . The mapping extends to all first-order formulas in the
natural way by setting (¬θ (x̄))i to be ¬ θ (x̄)i , (φ (x̄) → ψ(x̄))i to be φ (x̄)i → ψ(x̄)i ,
and (∀y θ (x̄, y))i to be ∀y (Dom(y)i → θ (x̄, y)i ). For this to define an interpretation, we
shall insist that T1 ` ∃x Dom(x)i and T1 ` σ i for every axiom σ ∈ T2 . Visser calls such
interpretations relative interpretations.
The category of theories and interpretations is obtained by taking theories as ob-
jects and such mappings i as morphisms, where we choose to say that f : T2 → T1 and
g : T2 → T1 are equal if T1 ` ∀x̄ (R(x̄)f ↔ R(x̄)g ) for all R in L2 . This makes a cate-
gory, where the identity interpretation 1 = 1T : T → T is the morphism given up to this
equality by the mapping R(x̄)1 = R(x̄) for all R. The composition gf : T3 → T1 of mor-
phisms f : T3 → T2 and g : T2 → T1 is given by (R(x̄))(gf) = (R(x̄)f )g , and associativity
holds as can easily be checked. (It is somewhat annoying that the convenient notation
of writing interpretation-applications as superscripts is at odds with the usual maps-
on-left convention for morphisms.) As usual, two morphisms f : T → S and g : S → T
are said to be inverse to each other if fg = 1S and gf = 1T . An interpretation is then
a morphism in this category, i.e., a mapping as described in the last paragraph modulo
the equivalence given by this notion of equality.
As for many kinds of interpretations, ours respects derivability in first-order logic
because of the way our mappings were defined on non-atomic formulas. Thus we have,

2
Proposition 1. Let f : T2 → T1 and suppose T2 ` σ . Then T1 ` σ f .
The next proposition follows from the definitions just given and an induction on
formulas.
Proposition 2. Let f : T2 → T1 and g : T2 → T1 be equal as interpretations. Then for
each formula θ (x̄) of L2 we have T1 ` ∀x̄ (θ (x̄)f ↔ θ (x̄)g ). In particular, for each
sentence σ of L2 , T1 ` σ f if and only if T1 ` σ g .

3 The Ackermann interpretation

By ZF−inf we mean the theory in the first-order language L∈ of set theory with all the
usual axioms of ZF except infinity, which is negated. More specifically, and following
Baratella and Ferro [2], let the set theory EST have the usual axioms or axiom schemes
of extensionality, existence of the empty set, pair set, sum set, separation and replace-
ment2 ; denote the axiom of power set by Pow, the axiom of foundation, expressed
as ∀x (x 6= ∅ → ∃y ∈ x ∀z ∈ x z 6∈ y), by Found; and the usual axiom of infinity, i.e.,
∃w (∅ ∈ w ∧ ∀x ∈ w (x ∪ {x} ∈ w)), by Inf. Then ZF−inf is EST together with Pow,
Found, ¬Inf.
It was observed in 1937 by Wilhelm Ackermann [1] that N with the membership
relation defined by
n ∈ m iff the nth digit in the binary representation of m is 1
satisfies ZF−inf. This interpretation, formalised in ZF with ω in place of N yields a
bijection between ω and the collection Vω of hereditarily finite sets.
This interpretation has been given extensive treatment in the literature. Working in
PA (or a suitable fragment of PA) one may define the interpretation by setting Dom(x)a
to be ‘Dom(x)’, (x = y)a to be ‘x = y’, and (x ∈ y)a to be
∃w < y ∃p 6 y ∃r < p ( p = 2x ∧ y = (2w + 1)p + r ).
We will refer to this interpretation throughout the rest of the paper as the Ackermann
interpretation.
All that is required for this to work is a suitable formula p = 2x that represents ‘p
is the xth power of 2’. In fact many such formulas are known which have the necessary
inductive definition
1 = 20 ∧ ∀x ∀p (p = 2x → 2p = 2x+1 ) (1)
and partial-function nature
∀x ∀p ∀q (p = 2x ∧ q = 2x → p = q) (2)
provable in PA or even much weaker systems, including ∆0 formulas p = 2x for which
the above statements are provable in the subsystem of PA with induction on ∆0 formulas
only (see Gaifman and Dimitracopoulos [9] or Hájek and Pudlák [11, §I.1(b)]).
Making extensive use of the axiom scheme of induction, it is straightforward to
check that every model of PA with this binary relation defined satisfies ZF−inf.
Theorem 3. a : ZF−inf → PA.
More information on this interpretation in subsystems of PA is given in Section 7
below.
2 The axiom scheme of separation is redundant in the presence of replacement.

3
4 The ordinal interpretation
The finite ordinals in the standard model of ZFC resemble the usual natural numbers.
In a world without infinite numbers, one would expect the class of all ordinals to satisfy
the ordinary rules of arithmetic.
Definition 4. Trans(x) (‘x is transitive’) is the formula

∀y, z (z ∈ y ∧ y ∈ x → z ∈ x),

and x ∈ On (‘x is an ordinal’) is

Trans(x) ∧ ∀y, z ∈ x (y ∈ z ∨ y = z ∨ z ∈ y).

Most ordinal theory can be done inside our theory of sets just as it is in ZF. See for
example Chapter 2 of Drake [7] for details. However, as an easy consequence of ¬Inf,
every nonzero ordinal is a successor ordinal—there are no limit ordinals in ZF−inf.


Theorem 5. ZF−inf ` ∀x x ∈ On ↔ x = ∅ ∨ ∃y ∈ On (x = y ∪ {y}) .
As a corollary, every set of ordinals in ZF−inf has a maximum element.
Corollary 6. ZF−inf ` ∀x ⊆ On (x 6= ∅ → x ∈ x).
S

Arithmetic of ordinals can be defined in the usual way, and there is no problem at
all in showing that the ordinals with usual arithmetic yields an interpretation of PA in
ZF−inf.
Definition 7. Let Dom(x)o be ‘x ∈ On’, (x = y)o be ‘x = y’, (x < y)o be ‘x ∈ y’,
(x + y = z)o be ‘x + y = z’ (ordinal addition), and (x · y = z)o be ‘x × y = z’ (ordinal
multiplication).
Theorem 8. o : PA → ZF−inf.
Thus PA and ZF−inf have the same consistency strength, but this interpretation is
clearly not inverse to the Ackermann interpretation.

5 Epsilon induction and transitive containment

In order to define an inverse to the Ackermann interpretation we shall need to consider
the principle of ∈-induction.3
Definition 9. For an L∈ -formula φ (x, ȳ), I∈ φ is the sentence


∀ȳ ∀x (∀w ∈ x φ (w, ȳ) → φ (x, ȳ)) → ∀x φ (x, ȳ) .

∈-Ind denotes the scheme {I∈ ψ : ψ(x, ȳ) is an L∈ -formula}.

It turns out that not every model of ZF−inf admits ∈-induction. This is proved
using the closely related notion of the transitive closure of a set.
3 Further references. The adduction operator previously mentioned gives rise to another scheme of

induction

φ (∅) ∧ ∀x∀y φ (x) ∧ φ (y) → φ (x C y) → ∀x φ (x),
which is said to have originated from Givant and Tarski [10] (cf. Kirby [13, §2]).

4
Definition 10. Define y = TC(x) (‘y is the transitive closure of x’) to be
y ⊇ x ∧ Trans(y) ∧ ∀y0 (y0 ⊇ x ∧ Trans(y0 ) → y0 ⊇ y).
TC is the axiom ∀x ∃u ⊇ x Trans(u).
It is easy to check (using extensionality) that the transitive closure y = TC(x) of a
set x, if it exists, is unique. The L∈ -sentence TC says that every set is contained in a
transitive set. This turns out to be equivalent to the apparently stronger statement that
every set has a transitive closure over a weak fragment of ZF−inf, because the transitive
closure of a set x can alternatively be defined as the intersection of all transitive sets
containing x.
Lemma 11. EST ` ∀x (∃y (y ⊇ x ∧ Trans(y)) → ∃y (TC(x) = y)).
Over EST + Found, the axiom of transitive containment or closure is equivalent to
∈-induction. In fact, a single instance of ∈-induction is enough to prove the whole
schema ∈-Ind.
Proposition 12. For all V  EST + Found,
V  ∈-Ind ⇔ V  TC.
Proof. (Sketch.) Let V  EST ∪ ∈-Ind. We prove the equivalent statement
V  ∀x ∃y (y = TC(x))
by ∈-induction on x in V .
Let x ∈ V such that every element of x has a transitive closure. By the axioms of
EST and the induction hypothesis, the set
{z : ∃x0 ∈ x (z = TC(x0 ))} ∪ x
[

exists and is a transitive superset of x. So Lemma 11 implies x has a transitive closure,

completing the induction.
The proof of the converse requires the axiom of foundation but is the same as the
standard proof of ∈-induction in ZF. See, for example, Chapter 2 of Drake [7].
The statement TC is not however provable from ZF−inf. We sketch a proof here
due to Mancini. (See Mancini and Zambella’s article [15] for more details.)
Theorem 13. ZF−inf ∪ {¬TC} is consistent.
Proof. (Sketch.) Consider the hereditarily finite sets (Vω , ∈), and the set of ordinals ω
in Vω . Let
ω ∗ = {{x ∪ {x}} ∈ Vω : x ∈ ω},
the set of singletons of nonzero elements of ω. Define F : Vω → Vω by
F(x) = {x ∪ {x}} and F({x ∪ {x}}) = x
for x ∈ ω and F(x) = x for x 6∈ ω ∪ ω ∗ . Thus, F is an involution, i.e., a permutation of
Vω which is a disjoint product of 2-cycles. Now define the binary relation ∈F by
∀x, y ∈ Vω (x ∈F y ⇔ x ∈ F(y)).
It can be verified that (Vω , ∈F )  ZF−inf. In addition, (Vω , ∈F )  ¬TC. In particular,
(Vω , ∈F )  ¬∃u (u ⊇ ∅ ∧ Trans(u))
since any such set is necessarily infinite.

5
 Thus, one really needs to add the extra hypothesis TC for the ∈-induction argument
to work.
It seems that this result, or ones like it, were known before Mancini and Zambella’s
paper. Kunen [14, p149(29)] attributes a result like this to Barwise though this model
fails to satisfy the power set axiom, it seems. And Baratella and Ferro [2, p345] point
out that the independence of transitive containment relative to other axioms of finite set
theory was proved in the context of the Alternative Set Theory by Sochor. They also
construct a related model (which they attribute to Kunen) in which transitive contain-
ment fails, but one in which the axiom of foundation also fails.
The following easy but important proposition shows that any inverse to the Ack-
ermann interpretation must use TC in some form, even for rather weak systems of
arithmetic.
Proposition 14. Let a : ZF−inf → PA be the Ackermann interpretation. Then I∆0 `
TCa .
Proof. For x ∈ M  I∆0 , let y ∈ M be one less than the smallest power of 2 that is bigger
than x. (Such a y exists by ∆0 induction as there must be a greatest u such that the ∆0
formula ∃z 6 x 2u = z holds.) If M  (u ∈ v ∧ v ∈ y)a then u < v < y so the uth binary
digit of y is 1.
In other words, a : ZF−inf + TC → PA. If b : PA → ZF−inf is inverse to a then we
would have ZF−inf ` (TCa )b hence ZF−inf ` TC by Proposition 2, which is impos-
sible. Thus we must consider the theory ZF−inf∗ = ZF−inf + TC rather than ZF−inf
itself.

6 The inverse Ackermann interpretation

Equipped with ∈-induction, we shall obtain an inverse interpretation b : PA → ZF−inf∗ .
The plan is to define a natural bijection p : V → On between the whole universe and
the ordinals. The required interpretation can then be obtained by composing this map
with the map o defined on On.4
At first, it appears difficult to see how to use ∈-induction at all, since the required
inductive definition of p is p(x) = ∑y∈x 2p(y) and this seems to need a separate induction
on the cardinality of x—just the sort of induction we don’t yet have and are trying to
justify. However there is a way round this problem using ordinal summation.
Definition 15. Working in V  ZF−inf, let P(On) denote the class of sets of ordinals,
b : On × P(On) → On be the function defined recursively by
and let Σ
b(0, x) = 0
Σ
for all x ∈ P(On), and
(
b(c ∪ {c}, x) = Σ(c, x),
b if c ∪ {c} 6∈ x,
Σ b(c, x) + (c ∪ {c}), if c ∪ {c} ∈ x
Σ
for all c ∈ On and x ∈ P(On). Also, let Σ : P(On) → On be the function defined by
[
∀x ⊆ On Σ(x) = Σ b x, x
.
4 Further references. The inverse Ackermann interpretation is said to have been conjectured to exist in

1964 by Beth [4], and it was found by Mycielski [16] in the same year (cf. Kirby [13, §2]).

6
 Informally, this defines
Σ(x) = ∑ y
y∈x
for a set of ordinals x, where the summation on the right hand side of the equation
refers to ordinal addition. Using induction on the ordinals, it can be proved that both Σ
b
and Σ are class functions definable over ZF−inf. The following definition is the place
where TC comes in.
Definition 16. In V  ZF−inf∗ , define p : V → On recursively by
p(x) = Σ {2p(y) ∈ On : y ∈ x} .

As indicated, p is again a class function that can be defined by ∈-recursion on all

sets over ZF−inf∗ . Informally, in ZF−inf∗ we have
p(x) = ∑ 2p(y) .
y∈x

This is precisely the bijection we are looking for.

Proposition 17. ZF−inf∗ proves that p is a bijective class function V → On.
Proof. The injectivity part is proved by ∈-induction on all sets. The surjectivity part is
proved by ∈-induction on the ordinals.
Definition 18. Define a mapping b of atomic sentences of arithmetic into set theory by
saying Dom(x)b is ‘Dom(x)’, (x = y)b is ‘x = y’, (x < y)b is ‘p(x) < p(y)’, (x + y = z)b
is ‘p(x)+p(y) = p(z)’, and (x ·y = z)b is ‘p(x)×p(y) = p(z)’, where the target relations
and operations of <, + and × are the usual operations on the ordinals.
Theorem 19. b : PA → ZF−inf∗ .
Proof. This follows from Theorem 8 and Propositions 14 and 17.
It remains to prove the following.
Theorem 20. The interpretations a : ZF−inf∗ → PA and b : PA → ZF−inf∗ are inverse
to each other.
Proof. We must show that ab = 1PA and ba = 1ZF−inf∗ , i.e., that
PA ` ∀x̄ ((σ (x̄)b )a ↔ σ (x̄))
and
ZF−inf∗ ` ∀x̄ ((τ(x̄)a )b ↔ τ(x̄))
for all atomic formulas σ (x̄) of LA and τ(x̄) of L∈ . The details are routine; we make
only a few comments on them here.
Note first that both interpretations preserve the logical symbols (they map domains
to domains and do not alter equality), thus we need only look at non-logical symbols.
To see ab = 1PA , work in PA and write x <0 y for (x < y)(ab) , 00 for the <0 -least
number, 10 for the <0 -least number not equal to 00 , x +0 y for the unique z such that
(x + y = z)(ab) , and x ·0 y for the unique w such that (x · y = w)(ab) . It can be checked
that 00 = 0, 10 = 1, and ∀x x + 1 = x +0 10 , and by an induction on y this implies that
∀x, y x + y = x +0 y. Since x < y ↔ ∃z y = x + z + 1 and x <0 y ↔ ∃z x = y +0 z +0 10 , it
follows that ∀x, y (x < y ↔ x <0 y). Similarly x · (y + 1) = (x · y) + x and x ·0 (y +0 10 ) =
(x ·0 y) +0 x, so by induction on y we can show that ∀x, y x · y = x ·0 y, as required.
To see ba = 1ZF−inf∗ , work in ZF−inf∗ and write x ∈0 y for (x ∈ y)(ba) . Then by
∈-induction we may show ∀x ∀y (x ∈ y ↔ x ∈0 y) and the result then follows.

7
7 Fragments of arithmetic and set theory
As with many results for models of Peano arithmetic, the results presented above have
hierarchical variations for subsystems of PA and ZF. It seems worthwhile to briefly
survey what one can say about this here. Our results along these lines are not really
new since they are essentially contained in Ressayre’s Sous-systèmes remarquables de
ZF [19].
First, we fix a ∆0 formula for p = 2x which defines a partial-function (formula (2)
above), and for which the recursive definition (formula (1) above) and its ‘downward’
version,
∀x ∀p (p = 2x+1 → ∃q (p = 2q ∧ q = 2x )) (3)
are all provable in I∆0 . This latter sentence appears necessary and more than just a
convenience. But we cannot think of a natural example where the upward recursion
holds but the downward one fails. However, the additional assumption is quite mild,
for if a formula θ (x, y) for y = 2x satisfies (1) and (2) provably in I∆0 , then θ (x, y) ∧
∀u<x ∃v<y θ (u, v) satisfies (1), (2) and (3) provably in I∆0 , as the reader may check.
With such a definition of exponentiation chosen, the domain of definition of the
exponentiation operation y 7→ 2y is an initial segment of the model of I∆0 , necessarily
closed under successor and addition, but not in general closed under multiplication. We
can prove in I∆0 that for each x there is a power of two, 2y , that is greater than x, and in
fact there is a least such power of two. With a slight misuse of notation we denote the
least y such that 2y > x by log x.
It is fairly straightforward to see that I∆0 proves the Ackermann interpretation of
many set theory axioms, including: extensionality; empty set; sum set; foundation;
transitive closure; and the negation of the axiom of infinity.
Adding the further arithmetic axiom exp, ∀x ∃y y = 2x , stating the totality of the
function x 7→ 2x we may prove the Ackermann interpretation of the pair set axiom and
the power set axiom. Moreover, exp is necessary as well as sufficient for these: to show
I∆0 ` Paira → exp and I∆0 ` Powa → exp, take a number y and 2x , the smallest power
of two greater than y. Then as 2x codes the set {x}, by pair set or power set the set
x x
{{x}, {}} must also be coded, and this code can only be 22 + 1. Thus 22 exists and so
therefore does 2 . y 5

To study the remaining axioms for set theory, we introduce a hierarchy of formulas
related to Lévy’s but more convenient for weak systems where the power set axiom
may fail. Say a formula of the language of set theory is ∆P 0 if it is formed from atomic
formulas by boolean operations and bounded quantifiers of the form Qx ∈ y . . . and
Qx ⊆ y . . .. A formula is ΣP P
n (respectively Πn ) if it is ∃x̄1 ∀x̄2 . . . Qx̄n θ (respec-
∗ P
tively ∀x̄1 ∃x̄2 . . . Q x̄n θ ) where θ is ∆0 . With a symbol for power set, the quantifier
Qx ⊆ y . . . can be replaced by Qx ∈ P(y) . . . and as the power set of a set x is Π1 -
definable (in the sense of Lévy) this means that the ΣP P
n / Πn hierarchy agrees with
the Lévy hierarchy for n > 2 in the presence of the power set axiom.
With these definitions in place, we can readily prove by induction that I∆0 proves
the Ackermann interpretation of the ∆P P
0 -separation axiom scheme, ∆0 –Sep, and for
P a P a
n > 1, IΣn proves both Σn –Sep and Πn –Sep . Also, using the least number principle
we have that I∆0 proves ∆P a P
0 –Ind , the Ackermann interpretation of the ∆0 -epsilon
5 Further references. As mentioned by Enayat [personal communication], two theories of sets were

proved to be equivalent to I∆0 + exp by Gaifman and Dimitracopoulos [9], one of which is the theory of
exponentially finite sets.

8
induction axiom scheme

∀ā (∀y (∀x ∈ y θ (x, ā) → θ (y, ā)) → ∀y θ (y, ā))

for θ in ∆P P a P a
0 , and similarly IΣn proves Σn –Ind and Πn –Ind .
The collection and replacement axiom schemes are a little more delicate. Let
ΣPn –Coll be the scheme

∀a, c̄ (∀x ∈ a ∃y θ (x, y, c̄) → ∃b ∀x ∈ a ∃y ∈ b θ (x, y, c̄))

for θ in ΣPn , etc. To justify the Ackermann interpretation of axioms like this, we need
fast growing functions. For example, taking a ⊇ {l − 1} where l = log a and letting
θ (x, y) be the ∆0 formula y = {x} we trivially have (∀x ∈ a ∃y θ (x, y))a in all models
of I∆0 , but the code for any b with (∃b ∀x ∈ a ∃y ∈ b θ (x, y))a must be at least as large
l−1
as 22 ≈ 2a so exponentiation is required. On the other hand, it is straightforward
to see that (ΣP a
n –Coll) is provable in I∆0 + exp +BΣn . Similar remarks apply to the
replacement axioms, which we might as well take here to be

∀a, c̄ (∀x ∈ a ∃y θ (x, y, c̄) → ∃ f ∀x ∈ a θ (x, f (x), c̄)).

Then (ΣP a
n –Rep) is provable in IΣn , for each n > 1, but for n = 0 the best result seems
to be I∆0 + exp +BΣ1 ` (∆P a
0 –Rep) .
The inverse b to the Ackermann interpretation is a somewhat more complicated
affair, and we make no effort here to even to attempt to say what happens here in the
absence of the axioms corresponding to exponentiation, namely pair set and power
set. We shall simply state our results, which are based on a straightforward but rather
tedious analysis of the proofs given in Sections 4, 5 and 6, and leave the verification
to the reader. (The presentation of Kripke–Platek set theory given in Chapter I of
Barwise [3] may be found helpful.) It turns out that for the results that follow, the usual
Lévy hierarchy (where bounded quantifiers take the form Qx ∈ y . . . only) gives the
sharpest result. We denote the levels in this hierarchy by ∆Ln , ΣLn and ΠLn .
The first stage is to define the ordinals, in particular the arithmetic structure on the
ordinals, and to define the bijection p : V → On between the universe and the ordinals.
We take as our base theory, the set theory consisting of the axioms of extensionality,
empty set, sum set, and the negation of the axiom of infinity, and the axiom schemes of:
∆L0 -separation, ∆L0 -collection, and both ΣL1 -epsilon induction and ΠL1 -epsilon induction.
With this base theory, the class of ordinals is defined (as the transitive sets linearly
ordered by ∈) by a ∆L0 formula, and the arithmetic operations on the ordinals including
+, ×, exp are all functions with ∆L1 graph. Moreover the scheme of ΣL1 -induction proves
the axiom of foundation and the axiom of transitive containment, and hence we may
define the bijection p : V → On, also with ∆L1 graph, and that this map is inverse to the
Ackermann interpretation of sets from the ordinals. The base axioms for arithmetic,
i.e., the axioms for nonnegative parts of discretely ordered rings, can then be shown to
hold in the ordinals.
To prove the induction scheme we need somewhat stronger axioms showing that
the ordinals satisfy IΣn , etc. Perhaps the simplest option is to take stronger instances
of the separation scheme. Given a Σn -definable subclass A of the ordinals containing 0
and closed under successor, and given n ∈ On, let a be the set of ordinals in A less than
n, by ΣLn -separation. Then by ordinal induction a = n so, as n was arbitrary, A = On.
Putting all this together, we have

9
Theorem 21. Let n > 1 and let ΣLn –Sep denote the set theory consisting of ΣLn -separation
together with base theory Ext, Emp, Sum, ¬Inf, ∆L0 –Coll, ΣL1 –Ind, and ΠL1 –Ind. Then
the Ackermann interpretation a and its inverse b are defined for IΣn and ΣLn –Sep, with
a : ΣLn –Sep → IΣn
and
b : IΣn → ΣLn –Sep
and these are inverse to each other.
Thus, the Ackermann interpretation behaves well for the ‘traditional’ subtheories
IΣn of PA. For weaker theories the pair set and power set axioms behave like the axiom
exp expressing the totality of exponentiation.
To close this section, we should remark that for systems of arithmetic without exp,
there are other choices of interpretations of set theory. The Ackermann interpretation is
of course valid, and in some sense represents the smallest collection of sets or smallest
model of set theory one might take, but we could transform a model M of I∆0 to a
different model of set theory by taking as sets all bounded ∆0 -definable subsets of
M, and identifying an a ∈ M with the set of x ∈ M such that M  (x ∈ a)a . In the
(apparent) absence of a formalised notion of truth for ∆0 formulas in I∆0 (this itself
an open problem) this does not appear to be an interpretation in the formal sense used
above. It is essentially the approach taken by Diaconescu and Kirby [6] and, as they
point out, counting principles and pigeonhole principles become highly relevant. (It is
still not known if I∆0 alone can prove the pigeonhole principle for ∆0 functions.) Of
course if M  exp this interpretation and the Ackermann one coincide since
I∆0 + exp ` ∀a ∃y ∀x<a ((x ∈ y)a ↔ θ (x))
for each ∆0 formula θ (x), possibly with further parameters.

8 Conclusions for finite set theories and arithmetic

Perhaps the first and most obvious conclusion is that statements concerning the equiv-
alence of ‘Peano Arithmetic’ and ‘ZF with the axiom of infinity negated’ require some
care to formulate and prove. It is certainly true that PA and ‘ZF with the axiom of infin-
ity negated’ are equiconsistent for just about any sensible axiomatisation of the latter,
in the sense that interpretations exist in both directions.6 Probably this is the ‘folklore
result’ that most people remember. But for the finer result with interpretations inverse
to each other, careful axiomatisation of the set theory is required. A category theoretic
framework for interpretations is useful to direct attention to these refinements.
The reader will have noticed that the axiom of choice (AC) has been completely
absent from the discussion here. It wasn’t necessary at all for the inverse Acker-
mann interpretation, though it comes for free with the Ackermann interpretation itself:
a : ZF−inf∗ + AC → PA. This shows ZF−inf∗ ` AC. In fact AC can be proved in much
less than this, as Baratella and Ferro [2] point out.
The universe of hereditarily finite sets has been an important source of analogies,
inspiration and axioms for set theory with infinite sets, from the time of Cantor on-
wards. In the case of the work here we have learnt that models of ZF−inf∗ are cat-
egorical with respect to their ordinals, in the sense that the model of set theory can
6 Further references. It is mentioned in Kirby’s article [13, §5] that PS (which was proved to be “equiv-

alent” to PA by Previale [17] as previously mentioned) is interpretable in ZF\{Inf} (attributed to Mycielski).

10
be reconstructed from the ordinals as a model of arithmetic. The analogous case for
models of full ZF is a result due to Harvey Friedman [8] that characterises ZF +V =L
as the first-order theory extending ZF in which any two models with the same ordinals
are isomorphic. (See Friedman’s paper for a precise statement of this result.)7
Results like this provide evidence that models of ZF in which V =L holds are actu-
ally the ‘uninteresting’ models of ZF, and a similar remark holds for models of finite
set theory too. Arguably, the uninteresting models of ZF−inf are the ones in which the
axiom of transitive containment holds, and the interesting ones are those for which TC
fails. It would be nice to know more about such models, but very few have appeared
in the literature. Note that every model M of ZF−inf has a transitive submodel V ⊆ M,
V  ZF−inf + TC, with the same ordinals, defined to be the image of the function
f : On → V given by the recursion on ordinals

f (x) = { f (y) : On  (y ∈ x)a }.

One way to think about the submodel V ⊆ M is that this is the collection of sets in M
with a rank, i.e., they are the sets in the cumulative hierarchy. (Hence the name V we
have used for this submodel.)
Whether the theory of models with M 6= V will have any consequence for arithmetic
itself is not clear, as the ordinals in models of ZF−inf + ¬TC are of course just the usual
models of PA.
Finite set theories in which the axiom of power set fails are an altogether more diffi-
cult proposition—apparently for some very good reasons possibly related to the coding
problems and complexity theory issues one gets with models of arithmetic without ex-
ponentiation.
It is not quite clear to us exactly how this might arise from the point of view of the
Ackermann interpretation considered here. However, one view of the inverse Acker-
mann interpretation in a model M of some fragment of finite set theory is that it sets
out to define suitable arithmetic structure for + and · on the whole universe of sets, in a
way that is compatible with the Ackermann interpretation that maps ordinals to sets in
the cumulative hierarchy V ⊆ M. (V will in general be a subclass of the class of all sets
because the transitive containment axiom may fail in M.) But it is clear that if one did
have a general process for achieving this, or even for defining + and · on a significant
part of the whole universe, this arithmetic model would have to be an end-extension of
the image of the ordinals given by the Ackermann interpretation.8
Such a general method for producing end-extensions of models of arithmetic might
be very significant for models of I∆0 + ¬ exp. There are several open questions con-
cerning when models of I∆0 + ¬ exp have end-extensions [28]. More significantly,
success in the naı̈ve attempt at producing an inverse Ackermann interpretation would
give end-extensions M ⊆e K of M  I∆0 + ¬ exp where each element x ∈ M has an
exponential y = 2x ∈ K. Such a construction might be applied in the case when the
model of finite set theory is of the type studied by Diaconescu and Kirby [6], that is
formed from ∆0 -definable bounded subsets. This would in turn have consequences for
the Π1 -theory of the base model M. In particular this would have consequences for the
Π1 -theory provable in I∆0 with a single application of exponentiation. This Π1 -theory
contains some statements not known in I∆0 itself, such as the ∆0 pigeonhole principle,
7 Further references. The first-order theory of ordinals in models of ZF was studied by Takeuti [21, 22,

23, 24] in a language extending LA .

8 Further references. Kirby’s article [13] gives a definition of addition and multiplication on all sets in a

model of ZF\{Inf}, which is a apparently not isomorphic to the Ackermann one.

11
and also contains consistency statements such as the tableau consistency of Q, known
not to be provable in I∆0 alone. (See for example papers by Pudlák [18], and Paris and
Wilkie [27] for much more information.)
Finally, leaving arithmetic aside, the Axiom of Choice becomes interesting for sys-
tems of finite set theory that do not have the power set axiom. Apparently power set
is required for the equivalence of the well-ordering principle (WO) and AC, though
WO → AC can be proved by the usual proof without power set. However, we were
unable to find a model of AC + ¬WO.9 Using the notation of Baratella and Ferro
introduced above, we can ask

Question 22. Does EST + ¬Inf + Found ` AC?

Note that, as Baratella and Ferro point out, EST + ¬Inf + Pow ` WO by a straight-
forward argument mapping sets bijectively onto ordinals. They also provide a model
of EST + ¬Inf + ¬WO.

References
[1] ACKERMANN , W ILHELM (1937). Die Widerspruchsfreiheit der allgemeinen
Mengenlehre. Mathematische Annalen, 114:305–315.

[2] BARATELLA , S TEFANO and F ERRO , RUGGERO (1993). A theory of sets with the
negation of the axiom of infinity. Mathematical Logic Quarterly, 39(3):338–352.

[3] BARWISE , J ON (1975). Admissible Sets and Structures: An Approach to Defin-

ability Theory. Perspectives in Mathematical Logic. Springer-Verlag, Berlin.

[4] B ETH , E. W. (1964). Axiomatique de la théorie des ensembles sans axiome de

l’infini. Bulletin de la Société Mathématique de Belgique, 16:127–136. In French.
Inaccessible.

[5] C HANG , C HEN C HUNG and K EISLER , H. J EROME (1977). Model Theory, Stud-
ies in Logic and the Foundations of Mathematics, vol. 73. North-Holland Pub-
lishing Company, Amsterdam, second edition.

[6] D IACONESCU , R ADU and K IRBY, L AURENCE A. S. (1987). Models of arith-

metic and categories with finiteness conditions. Annals of Pure and Applied
Logic, 35(2):123–148.
[7] D RAKE , F RANK R. (1974). Set Theory: An Introduction to Large Cardinals,
Studies in Logic and the Foundations of Mathematics, vol. 76. North-Holland
Publishing Company, Amsterdam.

[8] F RIEDMAN , H ARVEY (1978). Categoricity with respect to ordinals. In Higher

Set Theory: Proceedings, Oberwolfach, Germany, April 13-23, 1977 (M ÜLLER ,
G ERT H. and S COTT, DANA S., eds.), Lecture Notes in Mathematics, vol. 669,
pp. 17–20. Springer-Verlag, Berlin.
9 Further references. Zarach [29, §4] proved that if ZF\{Pow} is consistent, then so is (ZF\{Pow}) +

AC + ¬WO. He attributed this to Zbigniew Szczepaniak.

12
 [9] G AIFMAN , H AIM and D IMITRACOPOULOS , C ONSTANTINE (1982). Fragments
of Peano’s arithmetic and the MRDP theorem. In Logic and Algorithmic: Pa-
pers from an International Symposium in honor of Ernst Specker held at Ei-
dgenössische Technische Hochschule, Zurich, February 5-11, 1980, no. 30 in
Monographies de L’Enseignement Mathématique, pp. 187–206. Université de
Genève, Geneva.

[10] G IVANT, S. and TARSKI , A. (1977). Peano arithmetic and the Zermelo-like
theory of sets with finite ranks. Notices of the American Mathematical Society,
77T–E51:A–437. Inaccessible.

[11] H ÁJEK , P ETR and P UDL ÁK , PAVEL (1993). Metamathematics of First-Order
Arithmetic. Perspectives in Mathematical Logic. Springer-Verlag, Berlin.

[12] K AYE , R ICHARD (1991). Models of Peano Arithmetic, Oxford Logic Guides,
vol. 15. Clarendon Press, Oxford.

[13] K IRBY, L AURENCE (2007). Addition and multiplication of sets. Mathematical

Logic Quarterly, 53(1):52–65.

[14] K UNEN , K ENNETH (1980). Set Theory: An Introduction to Independence Proofs,

Studies in Logic and the Foundations of Mathematics, vol. 102. North-Holland
Publishing Company, Amsterdam.

[15] M ANCINI , A NTONELLA and Z AMBELLA , D OMENICO (2001). A note on recur-

sive models of set theories. Notre Dame Journal of Formal Logic, 42(2):109–115.

[16] M YCIELSKI , JAN (1964). The definition of arithmetic operations in the Acker-
mann model. Algebra i Logika Seminar, 3:64–65. In Russian. Inaccessible.

[17] P REVIALE , F LAVIO (1994). Induction and foundation in the theory of the hered-
itarily finite sets. Archive for Mathematical Logic, 33:213–241. Inaccessible.

[18] P UDL ÁK , PAVEL (1985). Cuts, consistency statements and interpretations. The
Journal of Symbolic Logic, 50(2):423–441.

[19] R ESSAYRE , J EAN -P IERRE (1987). Modèles non standard et sous-systèmes re-
marquables de ZF. In Modèles non Standard en Arithmétique et théorie des
Ensembles, Publications Mathématiques de l’Université Paris VII, vol. 22, pp.
47–147. Université de Paris VII U.E.R. de Mathématiques, Paris.

[20] Ś WIERCZKOWSKI , S. (2003). Finite sets and Gödel’s incompleteness theorems.

Dissertationes Mathematicae (Rozprawy Matematyczne), 422:1–58.

[21] TAKEUTI , G AISI (1954). Construction of the Set Theory from the Theory of
Ordinal Numbers. Journal of the Mathematical Society of Japan, 6(2):196–220.

[22] TAKEUTI , G AISI (1957). On the Theory of Ordinal Numbers. Journal of the
Mathematical Society of Japan, 9(1):93–113.

[23] TAKEUTI , G AISI (1958). On the Theory of Ordinal Numbers, II. Journal of the
Mathematical Society of Japan, 10(1):106–120.

[24] TAKEUTI , G AISI (1965). A Formalization of the Theory of Ordinal Numbers.

The Journal of Symbolic Logic, 30(3):295–317.

13
[25] TARSKI , A. and G IVANT, S. (1987). A Formalization of Set Theory Without Vari-
ables, American Mathematical Society Colloquium Publications, vol. 41. Amer-
ican Mathematical Society, Providence. Inaccessible.

[26] V ISSER , A LBERT (2004). Categories of theories and interpretations. Logic

Group Preprint Series 228, Faculteit Wijsbegeerte, Universiteit Utrecht,
http://www.phil.uu.nl/preprints/preprints/PREPRINTS/preprint228.pdf.
Published in Logic in Tehran: Proceedings of the Workshop and Conference
on Logic, Algebra, and Arithmetic, held October 18-22, 2003 (E NAYAT, A LI,
K ALANTARI , I RAJ and M ONIRI , M OJTABA, eds.), Lecture Notes in Logic, vol.
26, pp. 284–341. Association for Symbolic Logic, La Jolla, CA, 2006.

[27] W ILKIE , A LEX J. and PARIS , J EFF B. (1987). On the scheme of induction for
bounded arithmetic formulas. Annals of Pure and Applied Logic, 35(3):261–302.

[28] W ILKIE , A LEX J. and PARIS , J EFF B. (1989). On the existence of end exten-
sions of models of bounded induction. In Logic, Methodology and Philosophy
of Science, VIII (Moscow, 1987) (F ENSTAD , J ENS E RIK, F ROLOV, I VAN T. and
H ILPINEN , R ISTO, eds.), Studies in Logic and the Foundations of Mathematics,
vol. 126, pp. 143–161. North-Holland Publishing Company, Amsterdam.

[29] Z ARACH , A NDRZEJ (1982). Unions of ZF− -models which are themselves ZF− -
models. In Logic Colloquium ’80 (VAN DALEN , D., L ASCAR , D. and S MILEY,
J., eds.), Studies in Logic and the Foundations of Mathematics, pp. 315–342.
North-Holland Publishing Company, Amsterdam.

14