4.

Second Order Arithmetic and Reverse Mathematics

4.1. The Language of Second Order Arithmetic.
We’ve mentioned that Peano arithmetic is sufficient to carry out large
portions of ordinary mathematics, but with a qualifier, namely that Peano
arithmetic suffices to carry out most proofs regarding those statements it can
express. Up to now, we’ve emphasized that this is less of a limitation than it
appears at first—while Peano arithmetic nominally talks about numbers, it
can also encode other notions, like sequences, finite groups, and proof theory
itself, and prove things about those.
But this encoding has a limit. The only objects we can discuss using
Peano arithmetic are those which are in some sense finite, or at least finitely
describable. This excludes large swaths of mathematics—almost all of topol-
ogy and analysis, just to start with. We would therefore like to extend our
system of logic to describe these notions as well.
For now, we will only move one step up, to include second order objects:
we will add sets to our langauge.
Definition 4.1. The language of second order arithmetic is a two-sorted
language: there are two kinds of terms, numeric terms and set terms.
• 0 is a numeric term,
• There are infinitely many numeric variables, x0 , x1 , . . ., each of which
is a numeric term
• If s is a numeric term then Ss is a numeric term,
• If s, t are numeric terms then +st and ·st are numeric terms (abbre-
viated s + t and s · t),
• There are infinitely many set variables, X0 , X1 , . . ., each of which is
a set term
• If t is a numeric term and S then ∈ tS is an atomic formula (abbre-
viated t ∈ S),
• If s and t are numeric terms then = st and < st are atomic formulas
(abbreviated s < t).
The formulas are built from the atomic formulas in the usual way.
As the examples in the definition suggest, we use upper case letters for
set variables and lower case letters for numeric terms. (Note that the only
set terms are the variables.) Later, it will be more convenient to work with
functions instead of sets, but within arithmetic, these are equivalent: we can
use our pairing operation, and say that X represents a function if for each
n there is exactly one m such that the pair (n, m) belongs to X.
We have to consider what we intend the semantics of this language to
be. One possibility is the semantics of “true” second order logic: a model
consists of a set M , representing the numeric objects, and interpretations
of the various functions and relations (probably with the requirement that
equality be the genuine equality relation), and a statement
∀Xφ(X)
1
2

is satisfied by the model if for every possible subset of M , the corresponding

statement holds.
“True” second order logic has no corresponding proof system. An easy
way to see this is to observe that it has no compactness theorem. For
example, the only model (up to isomorphism) of Peano arithmetic together
with the second order induction axiom
∀X(0 ∈ X ∧ ∀x(x ∈ X → Sx ∈ X) → ∀x x ∈ X).
is the standard model N. This is easily seen: any model of Peano arithmetic
has an initial segment isomorphic to N; applying the induction axiom to this
set, we see that it must be the whole of the model.
Consequently, this is not a very useful approach. An alternative semantics
is to treat this as an ordinary two-sorted first order language: a model M
consists of two sets, M and M 0 , where the elements of M are interpreted
as “numbers” and the elements of M 0 are interpreted as “sets”. There
is no obligation for M 0 to actually consist of subsets of M , but we have
an interpretation of ∈, and we may associate each element S of M 0 with
{m ∈ M | M  m ∈ S}. Importantly, M 0 need not include all possible
subsets of M . In particular, we can obtain nonstandard models of the second
order induction axiom by making sure to only include those sets which do
satisfy induction—and just leaving out the ones which don’t.
In terms of proof theory, this means we simply treat the language of
second order arithmetic as an ordinary first order language. In order to get
any mileage out of the new content, though, we need axioms which state
that there are any sets at all. We will only work in theories which have
comprehension axioms, which typically have the form
∃X∀x(x ∈ X ↔ φ(x))
where X does not appear free in the formula φ.
From here on, this will be our convention. In particular, whenever we use
terms like “second order” or “higher order”, we should remember that we
are still within the confines of first order logic, and are referring only to the
intended interpretation.

4.2. Z2 .
Definition 4.2. Z2 , the theory of full second order arithmetic consists of:
• The axioms P − ,
• The second order induction axiom
∀X(0 ∈ X ∧ ∀x(x ∈ X → Sx ∈ X) → ∀x x ∈ X),
• The comprehension axiom
∃X∀x(x ∈ X ↔ φ(x))
for each formula φ where X does not appear free.
 3

Z2 is sometimes called the theory of analysis, in contrast to PA as the

theory of arithmetic, because Z2 can talk about real numbers in a coherent
way.
First, we need to talk about integers and rational numbers. For the
moment, let us simply consider coding in N, ignoring any issues of provability
or first order logic. We can encode the integers in a standard way: by an
element of Z, we mean an element of N viewed as a pair (n, m), taken to
represent n − m. Each integer has many representatives under this scheme,
so we write (n, m) ≡Z (n0 , m0 ) if n + m0 = n0 + m. We can define the
standard operations on the integers in a natural way: −(n, m) = (m, n),
(n, m)+(n0 , m0 ) = (n+n0 , m+m0 ), (n, m)·(n0 , m0 ) = (nn0 +mm0 , nm0 +mn0 ).
All these operations are of course well-defined on Z—they map different
representations of the same integer to representations of the same integer.
We can then represent the rationals Q by pairs of integers (a, b) with
b > 0: (a, b) represents a/b. Then (a, b) ≡Q (c, d) if ad = bc. (Of course,
equality here means ≡Z .) Again, addition, multiplication, and so on can be
defined on these representations in a natural way.
Now we observe that all the relevant definitions can be written down
using first order arithmetic, and their properties can all be proven in PA
(and even, as will be convenient later, IΣ1 ). Formally, what we mean by
this is that ordinary statements about the integers and the rationals can be
unfolded into (much more complicated) statements in the language of PA.
We give one example; fix formulas ρ1 (z, x) and ρ2 (z, y) which will both hold
exactly when z is the pair (x, y); formally, we should be able to deduce the
formulas:
• ∀x, y∃zρ1 (z, x) ∧ ρ2 (z, y),
• ∀z∀x, x0 (ρ1 (z, x) ∧ ρ1 (z, x0 ) → x = x0 ),
• ∀z∀y, y 0 (ρ2 (z, y) ∧ ρ2 (z, y 0 ) → y = y 0 ).
Then the statement that addition on integers is commutative,
∀s ∈ Z∀t ∈ Z(s +Z t =Z t +Z s)
becomes

∀s∀t∀n, m, m0 , n0 (ρ1 (s, n) ∧ ρ2 (s, m) ∧ ρ1 (t, n0 ) ∧ ρ2 (t, m0 ) →

(n + n0 ) + (m0 + m) = (n0 + n) + (m + m0 )).
Here the first part—the conjunction on the ρ—gives us a representation
of s, t as pairs, and checks that s, t really do count as integers (that is, are
actually pairs of natural numbers); formally, this is the translation of the
restriction of the quantification to be over the integers. The term s+Z t then
translates to (n + n0 , m + m0 ) while t +Z s translates to (n0 + n, m0 + m),
and then the formula stating that the terms are equal is the right side of
the implication.
To talk about real numbers, we’ll need to use sets. There are two standard
ways of defining the real numbers from the rationals: as Dedekind cuts, and
4

as Cauchy sequences. For technical reasons, the latter is more convenient

for us.
Definition 4.3. In second order arithmetic, we say a set R is a function
on a set D if for every x ∈ D there is exactly one y such that (x, y) ∈ R.
We say a set R is a infinite sequence if for every x there is exactly one y
such that (x, y) ∈ R. We often write R = hmn i where for each n, mn is the
number with (n, mn ) ∈ R.
In second order arithmetic, a real number is a sequence of rational num-
bers hqn i such that
∀q ∈ Q(q > 0 → ∃m∀n(m < n → |qm − qn | < q)),
It requires some unfolding, but we emphasize that this is a formula in
second order arithmetic. As with integers and rationals, two different sets
can encode the same real number, but we can write down formulas indicating
when two real numbers are equal, when one is larger than the other, how to
add real numbers, and so on.
Note that Z2 does not prove that “arbitrary” real numbers exist (whatever
that means). Rather, if we want to work with a real number, we have to
prove, within Z2 , that a corresponding representation exists. It is easy to
show that each rational number gives the corresponding real number—take
hqi. Further,√ all real numbers with “easy” descriptions exist. For instance,
to encode 2 we can take qn to be the largest rational of the form m/10n
such that (m/10n )2 <√2 (in other words, qn is the first n + 1 digits of the
decimal expansion of 2).
Similar coding efforts suffice to code most countable objects—countable
groups, for instance—and further, are sufficient to discuss most separable do-
mains (that is, situations where a countable subset is “dense” in some way).
Furthermore, the theory Z2 is very strong; it suffices to prove almost any
theorem it can state which has been proven in the mathematics literature.
However second order arithmetic cannot state everything—for instance, it
still can’t express statements about genuinely uncountable mathematics, in-
cluding things like set theory and portions of abstract analysis—and there
are statements (all known examples coming from logic) which can be stated
in second order arithmetic but are independent of Z2 .
There is a natural stratification of formulas in second order arithmetic
similar to the one in first order arithmetic.
Definition 4.4. The ∆00 formulas are those in which all quantifiers over
numeric variables are bounded and there are no quantifiers over set variables.
Σ00 and Π00 are alternate names for ∆00 .
The Σ0n+1 formulas are formulas of the form
∃xφ
(possibly with a block of several existential quantifiers) where φ is Π0n .
 5

The Π0n+1 formulas are formulas of the form

∀xφ
(possibly with a block of several universal quantifiers) where φ is Σ0n .
The arithmetic formulas, sometimes written Π0∞ , ∆10 , Σ10 , or Π10 , are those
formulas with no quantifiers over set variables.
The Σ1n+1 formuals are formulas of the form
∃Xφ
(possibly with a block of several existential quantifiers) where φ is Π1n .
The Π1n+1 formulas are formulas of the form
∀Xφ
(possibly with a block of several universal quantifiers) where φ is Σ1n .
The Π0n and Σ0n formulas are essentially our old Πn and Σn formulas
(except that they can contain set variables, but not set quantifiers). The
Σ1n and Π1n hierarchies are analagous, but with set quantifiers rather than
numeric ones. Note the (largely accurate) implication that even a single set
quantifier “trumps” any number of numeric quantifiers.
4.3. ACA0 .
It turns out that Z2 is very strong, and we would like to consider some
weaker theories.
Definition 4.5. The axioms of ACA0 consist of:
• The axioms of P − ,
• The second order induction axiom,
• The comprehension scheme restricted to arithmetic formulas.
ACA0 stands for “arithmetic comprehension axiom”. The subscript 0
means that instead of the induction axiom for arbitrary formulas, we have
only the second order induction axiom. In Z2 , this didn’t matter—every
formula defined a set, so it didn’t matter whether induction was over for-
mulas or sets. In ACA0 , the difference is significant—there are plenty of
formulas we can write down using set quantifiers which are perfectly sensi-
ble formulas, but since the comprehension axiom doesn’t apply, they don’t
define sets, and therefore we do not have a corresponding induction axiom.
Unlike PA, the system ACA0 can express statements about countable
sets, and objects encoded by countable sets. For an elementary example,
consider the following basic statement from real analysis:
Theorem 4.6 (Bolzano-Weierstraß Theorem). If hrn | n ∈ Ni is a bounded
infinite sequence of real numbers then there is a subsequence n0 < · · · <
nk < · · · such that limk→∞ rnk exists.
This statement is outside the domain of first order arithmetic, but it can
be expressed fairly naturally in the language of second order arithmetic. To
talk about infinite sequences of real numbers, we can just use pairing: S is
6

an infinite sequence of real numbers if for each n, Sn = {m | (n, m) ∈ S} is

a real number.
Theorem 4.7. ACA0 proves the Bolzano-Weierstraß Theorem.
Proof. Let S be a sequence of real numbers, so that for each n, Sn is a real
number. Let b be an integer such that for each n, |Sn | ≤ b. In general a
sequence of real numbers could have many limits, and it will be easiest to
pick out a natural one, namely the lim sup.
We will define a set T to consist of those pairs (n, k) such that Sn is
within 2−k of the lim sup of the sequence. More precisely, (n, k) ∈ T if there
are only finitely many m such that Sn + 2−k < Sm ; this is an arithmetic
formula, so ACA0 proves that the set T exists.
We claim that for each k, there are infinitely many n such that (n, k) ∈ T .
Suppose not, and pick some k with only finitely many n so (n, k) ∈ T . We
construct a sequence as follows: take r1 to be any element so (r1 , k) 6∈ T ,
and given ri with (ri , k) 6∈ T , there are infinitely many m with Sri + 2−k <
Sm , only finitely many of which satisfy (m, k) ∈ T , so take ri+1 so that
Sri + 2−k < Sri+1 and (ri+1 , k) 6∈ T . For each i we have
−b ≤ Sr1 < Sr2 − 2−k < Sr3 − 2 · 2−k < · · · Sri − i · 2−k ,
so by choosing i > b2k+1 , we have Sri − i · 2−k > −b, which means Sri > b,
contradicting the assumption that |Sn | ≤ b for all b.
We should check that ACA0 can actually prove that the sequence ri , . . . , ri
exists. In PA, we would prove the existence using the induction axiom—we
would take the statement “there exists a sequence ~r of length i so that for
each j ≤ i, (rj , k) 6∈ T and for each j < i, Srj + 2−k < Srj+1 ”, observe that
it holds trivially for length 0 (the empty sequence), and that if there is a
sequence of length i, there is also a sequence of length i + 1, which the argu-
ment above shows must exist. Similarly, in ACA0 we can consider the set
I of i such that there is a sequence of length i with the desired properties,
and argue that 0 ∈ I and whenever i ∈ I, i + 1 ∈ I; therefore every number
is in I, so in particular b2k+1 + 1 ∈ I.
Having shown that for each k, there are infinitely many n with (n, k) ∈ T ,
we define a sequence by setting rk to be least such that (rk , k) ∈ T and
rk > rk0 for k 0 < k. Again, to formalize this in ACA0 , we consider the
sequences ~r of length i such that for each k ≤ i, rk is least with (rk , k) ∈ T
and rk0 < rk for each k 0 < k. The statement that there exists such a sequence
is arithmetic, so we take I to be the set of i such that there is such a sequence
of length i; then 0 ∈ I since the empty sequence has the desired property,
and if there is a sequence of length i, there must also be a sequence of length
i + 1, so there are such sequences of every length. By a similar inductive
argument, we can show that there is a unique such sequence of any given
length. Finally, observe that ACA0 also proves the existence of the set R
of (k, rk ) such that rk is the k-th element in the unique such sequence of
length k.
 7

Now that we have our convergent subsequence, constructing the limit is

easy: a real number is a convergent sequence of rationals. Recall that each
Sn = hSn,i i is a sequence of rational numbers converging to Sn , so choose
qk so that qk = Srk ,ik where ik is chosen so that |Srk ,ik − Srk | < 2−k . Then
each qk is within 2−(k−1) of the lim sup of the sequence hSn i, so in particular
hqk i converges to the lim sup. 
Provability in ACA0 has been very thoroughly studied. ACA0 can ex-
press the notion of a continuous function (but not the notion of an arbitrary
function on the real numbers), and therefore much of elementary real analy-
sis. ACA0 can also describe notions like countable groups, fields, and vector
spaces, and prove many standard theorems on these objects.
We only discuss one more family of results in detail, coming from infinitary
combinatorics.
Theorem 4.8 (Infinite Pigeonhole Principle). Let S be an infinite set, let
f be a function on S and let T be a finite set such that for every s ∈ S,
f (s) ∈ T . Then there is an infinite set S 0 ⊆ S and a t ∈ T so that for every
s ∈ S 0 , f (s) = t.
Note that the notion of a set being infinite can be expressed in second
order arithmetic by saying that for every n, there is an m > n with m ∈ S.
Conversely, to say that a set is finite is to say that there is some n such that
every m ∈ S is ≤ n.
Proof. Suppose not; in particular, suppose that for each t ∈ T , St = {s ∈
S | f (s) = t} is finite.
Then for each t, let us define Ut ⊆ S to be the set of n ∈ S such that for
every m > n with m ∈ S, f (m) ≥ t. Let U be the set of t such that Ut is
infinite. Since U0 = S, certainly U0 is infinite, so 0 ∈ U . Suppose t ∈ U ,
so Ut is infinite. Ut = St ∪ Ut+1 ; since St is finite, there is an n so that
every element of S is ≤ n. So for any n0 , we may choose an m > max{n0 , n}
with m ∈ Ut , and since m 6∈ St , we have m ∈ Ut+1 . So Ut+1 is infinite, and
therefore t + 1 ∈ U .
By the second order induction axiom, every element is in U . In particular,
take n large enough that every element of T is < n. Then Un is infinite, so
there is an m ∈ Un ; but then f (m) is larger than every element of T . This
is a contradiction. 
Definition 4.9. Let σ = hs0 , s1 , . . . , sn i be a finite sequence. An initial
segment of σ is the sequence hs0 , . . . , sm i for some m ≤ n (including m = n
and the empty sequence, where m = −1). An immediate extension of σ is a
sequence hs0 , . . . , sn , ti for any t.
If hsn i is an infinite sequence, an initial segment is a finite sequence
hs0 , . . . , sm i.
A tree is a set S of finite sequences with the property that whenever
σ ∈ S, every initial segment of σ belongs to S.
8

Theorem 4.10 (König’s Lemma). Suppose that S is a tree of sequences of

integers such that S contains infinitely many elements, but each sequence in
S has only finitely many immediate extensions in S. Then there is an infinite
sequence hsn i such that all its initial segments are in S. Furthermore, this
can be proven in ACA0 .

Proof. First, we consider the proof in general, without worrying about ACA0 .
The idea is that we take the infinite sequence hsn i so that each of its initial
segments is also an initial segment of infinitely many elements of S. For
definiteness, we simply take the sn least such that this is the case.
Formally, suppose we have defined s0 , . . . , sm (possibly with m = −1 for
the base case) such that hs0 , . . . , sm i is an initial segment of infinitely many
elements of S. (This is true by assumption in the base case, since the empty
sequence is an initial segment of all sequences, and so in particular of all
the infinitely many elements of S.) By assumption, there are finitely many
t1 , . . . , tr such that hs0 , . . . , sm , ti i ∈ S. Consider those elements τ of S
such that hs0 , . . . , sm i is an initial segment of τ , and which are not equal to
hs0 , . . . , sm i. This is an infinite set S 0 ⊆ S. For each τ ∈ S 0 , there is exactly
one i ≤ r such that hs0 , . . . , sm , ti i is an initial segment of τ , so we can define
a function f : S 0 → [1, r]. By the infinite pigeonhole principle, there is an
i ≤ r so that f (τ ) = i for infinitely many τ ∈ S 0 ; we take sm+1 = ti .
To formalize this argument in ACA0 , we first show that we can write
down an arithmetic formula (using S as a parameter) defining an example
of the desired sequence. It is convenient to do this in two steps. First, we
simply define S ∗ to be the set of sequences hs0 , . . . , sm i ∈ S such that there
are infinitely many τ ∈ S extending hs0 , . . . , sm i; this is easily expressed with
a few numeric quantifiers (“for every r there is a t > r coding a sequence
extending hs0 , . . . , sm i”).
We wish to show that for every σ ∈ S ∗ , there is an immediate extension
of σ ∈ S ∗ . To see this, let σ ∈ S ∗ be given; then σ ∈ S and there are
infinitely many extensions of σ ∈ S. We carry out the argument above:
let Sσ = {τ | σ is an initial segment of τ }, let Sσ0 be the set of immediate
extensions of σ in S, and define the coloring f : Sσ → Sσ0 . Since Sσ0 is finite,
there is an element of Sσ0 which is the value of f infinitely often, so this value
of Sσ0 belongs to S ∗ .
Now we define our actual infinite sequence, T , to consist of those pairs
(m, sm ) such that there is a hs0 , . . . , sm i ∈ S ∗ such that there is no i ≤ m
and t < si such that hs0 , . . . , si−1 , ti ∈ S ∗ . The definition ensures that for
each m, there is at most one sm such that (m, sm ) ∈ T . To see that there
is such an sm , let T ∗ = {m | ∃sm (m, sm ) ∈ T }; by the induction axiom, it
suffices to show that 0 ∈ T ∗ and whenever m ∈ T ∗ , also m + 1 ∈ T ∗ . The
arguments are essentially the same; suppose (0, s0 ), . . . , (m, sm ) ∈ T . Then
hs0 , . . . , sm i = σ ∈ S ∗ , so there is some t with hs0 , . . . , sm , ti ∈ S ∗ . Let U be
the set of t such that for all t0 < t, hs0 , . . . , sm , t0 i 6∈ S ∗ ; clearly 0 ∈ U , but
there must be some t 6∈ U —take t + 1 where hs0 , . . . , sm , t + 1i ∈ S ∗ , which
 9

exists since every element of S ∗ has an immediate extension in S ∗ . Since the

induction axiom applies to U , there must be a t ∈ U such that t + 1 6∈ U .
Therefore hs0 , . . . , sm , ti ∈ S ∗ , but for every t0 < t, hs0 , . . . , sm , t0 i 6∈ S ∗ , so
(m + 1, t) ∈ T . 
4.4. The Strength of ACA0 .
ACA0 corresponds very closely to PA. In fact, the first order conse-
quences of ACA0 are exactly the theorems provable in PA.
Theorem 4.11. If PA ` φ then ACA0 ` φ.
Proof. Easily follows from the fact that every axiom of PA is deducible in
ACA0 . 
Theorem 4.12. If ACA0 ` φ where φ is a formula in the language of
Peano arithmetic then already PA ` φ.
In order to prove this, we need some lemmas.
Definition 4.13. Let X be a second order variable, and let φ(x) be a
formula not containing X with a distinguished free variable x. Then we
define ψ[φ/X] recursively by:
• (t ∈ X)[φ/X] is φ(t),
• If ψ is atomic and not t ∈ X then ψ[φ/X] is ψ,
• (ψ ~ θ)[φ/X] is ψ[φ/X] ~ θ[φ/X],
• (Qyψ)[φ/X] is Qyψ[φ/X],
• (QXψ)[φ/X] is QXψ,
• (QY ψ)[φ/X] is QY ψ[φ/X] if Y 6= X.
Lemma 4.14. Suppose there is a cut-free deduction of ΓACA0 , Γ ⇒ Σ where
Γ, Σ are arithmetic. Suppose also that φ is substitutable for X in Γ, Σ. Then
there is a cut-free deduction of ΓACA0 , ΓP A , Γ[φ/X] ⇒ Σ[φ/X] of size no
larger than the original deduction.
Proof. By induction on deductions. The only cases which do not follow
immediately from the inductive hypothesis are the quantifiers rules L∃, R∀,
and L∀.
We first consider the R∀ case. This must be a first order variable, since
there are no set quantifiers on the right-hand side of the sequent. If the
eigenvariable appears in φ we we apply the substitution lemma to replace
the eigenvariable with a fresh variable, and apply the same R∀ inference. In
the L∃ case, if we are considering a first order variable, the same argument
applies. (In both cases, we also need to consider the possibility that the
quantifier binds a variable in φ; this is prevented by the substitutability
assumption.)
If we are substituting for a second order variable, either X is not the
eigenvariable, in which case the same argument applies, or X is the eigen-
variable, in which case X does not appear in Γ or Σ, so no changes are
required.
10

In the L∀ case, if the main variable is not X, the claim follows immediately
from the inductive hypothesis. So consider the difficult case:

ΓACA0 , Γ, ψ[X/Y ] ⇒ Σ
ΓACA0 , Γ, ∀Y ψ ⇒ Σ

In this case, the inductive hypothesis gives us a deduction of

ΓACA0 , Γ[φ/X], ψ[φ/Y ] ⇒ Σ[φ/X].
We must have ∀Y ψ ∈ ΓACA0 , so ∀Y ψ is an induction axiom. But then
ψ[φ/Y ] is an axiom in ΓP A , so we have
ΓACA0 , ΓP A , Γ[φ/X] ⇒ Σ[φ/X].

Lemma 4.15. Suppose we have a cut-free deduction of ΓACA0 , Γ ⇒ Σ where
Γ, Σ are arithmetic. Let Γ0 , Σ0 be the result of replacing all free second order
variables in Γ, Σ with ⊥. Then there is a deduction of ΓP A , Γ0 ⇒ Σ0 where
all cuts are over arithmetic formulas.
Proof. By induction on the size of the deduction. The only non-trivial cases
are quantifier rules with second order variables, and since Σ is arithmetic,
all such formulas appear on the left-hand side. Suppose we are in the second
order L∃ case:

ΓACA0 , Γ, ψ ⇒ Σ
ΓACA0 , Γ0 , ∃Xψ ⇒ Σ0

Since the formulas in Γ are arithmetic, ∃Xψ must be an axiom of ACA0 ,

and so must be a comprehension axiom
∃X∀x(x ∈ X ↔ φ(x))
for some arithmetic formula φ. We may apply the previous lemma to sub-
stitute φ for X, obtaining a deduction of
ΓACA0 , ΓP A , Γ, ψ[φ/X] ⇒ Σ.
(Since X is the eigenvariable, it does not appear in ΓΣ.) This deduction is
no larger than the original, so we may apply IH, obtaining a deduction of
ΓP A , Γ0 , ψ[φ/X]0 ⇒ Σ0 .
Since ψ[φ/X]0 is derivable, we may apply a cut to obtain a deduction of
ΓP A , Γ0 ⇒ Σ0 .
Suppose we are in the second order L∀ case:

ΓACA0 , Γ, ψ ⇒ Σ
ΓACA0 , Γ0 , ∀Xψ ⇒ Σ0
 11

Then ∀Xψ must be an induction axiom, so we may substitute X with

any formula—say, ⊥—and apply the previous lemma to obtain a deduction
of
ΓACA0 , ΓP A , Γ[⊥/X], ψ[⊥/X] ⇒ Σ[⊥/X].
SInce ψ[⊥/X] is an instance of the first order induction axiom, this is already
ΓACA0 , ΓP A , Γ[⊥/X] ⇒ Σ[⊥/X].
We may then apply IH to obtain a deduction of
ΓP A , Γ0 ⇒ Σ0 .


Proof of 4.12. Suppose ACA0 ` Γ ⇒ Σ where Γ, Σ do not contain any

second order variables. By the cut-elimination theorem, there must be a
cut-free deduction
` ΓACA0 , Γ ⇒ Σ.
We simply apply the previous lemma, obtaining a deduction of ΓP A , Γ0 ⇒ Σ0 .
Since Γ, Σ do not contain any second order variables, Γ0 = Γ, Σ0 = Σ, so we
are done. 

4.5. Reverse Mathematics.

Is ACA0 necessary to prove the examples above? Put another way,
we know from our work on cut-elimination that there is a natural notion
of “strength” of a mathematical theory, namely the proof-theoretic ordinal.
Second order arithmetic turns out to be a good place to quantify the strength
of mathematical statements.
To make this precise, we will fix a weak “base theory”, and ask the what
the consequences of some interesting mathematical statement σ are when
using the base theory. (Some base theory of arithmetic is needed; for in-
stance, in the absence of enough arithmetic to prove that the rationals and
reals have their intended properties, it’s hard to argue that the statement
even means what it is supposed to mean.)
Definition 4.16. The theory RCA0 consists of:
• The axioms of P − ,
• IΣ1 ,
• The second order induction axiom
∀X(0 ∈ X ∧ ∀x(x ∈ X → Sx ∈ X) → ∀x x ∈ X).
• The recursive comprehension scheme: if φ(x) is a Σ1 formula and
ψ(x) is a Π1 formula then
∀x(φ(x) ↔ ψ(x)) → ∃X∀x(x ∈ X ↔ φ(x))
is an axiom.
12

Note that RCA0 has slightly more induction than its comprehension
scheme suggests; we have to add induction over Σ1 formulas. It turns out
that without the additional induction axiom, the system is too weak even
to use as a base theory for many purposes.
One might wonder why, instead, we don’t add comprehension for Σ1 for-
mulas. The answer is that this theory is equivalent to ACA0 ! The reason
is that we can use multiple steps: if we want {x | ∀y∃zφ(x, y, z)}, we can
first define S1 to be {(x, y) | ∃zφ(x, y)}, define S2 to be {x | ∃y (x, y) 6∈ S1 },
and finally our desired set is {x | x 6∈ S2 }.
In particular, this means that to show that an extension of RCA0 includes
ACA0 , it suffices to show that it has comprehension for Σ1 formulas with
parameters.
In RCA0 , the definition of a real number we gave above does not work.
We should define a real number to by a Cauchy sequence which converges
with a specified rate of convergence.
Definition 4.17. In RCA0 , a real number is a sequence of rational numbers
hqn i such that
∀k ∈ N∀m, n(k < m < n → |qm − qn | < 2−k ).
In ACA0 , this is equivalent to the original definition, in the sense that
ACA0 proves that every number in the original sense is equal to a real
number in this stronger sense. On the other hand, RCA0 cannot prove
useful properties of real numbers in the sense of ACA0 .
The next two theorems are typical results in the area known as Reverse
Mathematics.
Theorem 4.18. Over RCA0 , the Bolzano-Weierstraß Theorem implies
ACA0 .
We mean the Bolzano-Weierstraß Theorem for real numbers in the sense
of RCA0 , of course.

Proof. Fix some Σ1 formula ∃xφ(x, y). We wish to show that {y | ∃xφ(x, y)}
actually exists. The idea is that we’ll define a sequence of rational numbers
which is already convergent, and whose limit encodes exactly this set. More
precisely, we will arrange to converge to
X
2−y .
y|∃xφ(x,y)

We define qn = y≤n,∃x≤nφ(x,y) 2−y . Observe that the sequence hqn i does

P

exist in RCA0 , since it is given by a formula with only bounded quantifiers.

Therefore the corresponding sequence of real numbers Sn = qn exists as well.
If the Bolzano-Weierstraß Theorem holds, this sequence has a limit S =
hqn0 i. If we want to know whether ∃xφ(x, y), we can simply look at the
0
binary expansion of qy+1 0
; qy+1 is within 2−(y+1) of the final value of the
 13

sequence, so there is any x such that ∃xφ(x, y) exactly if the y-th digit in
0
the expansion of qy+1 is 1. 
This means that the Bolzano-Weierstraß Theorem is equivalent to ACA0
(given that we are working in at least RCA0 ).
Theorem 4.19. Over RCA0 , König’s Lemma implies ACA0 .
Proof. Again, fix some Σ1 formula ∃xφ(x, y). We define a tree of sequences
by saying that σ = (s0 , . . . , sn ) ∈ T if for each i ≤ n, either:
• si = 0 and there is no x ≤ n such that φ(x, i) holds, or
• si > 0 and si − 1 is least such that φ(si − 1, i).
Again, notice that this is given by a formula with only bounded quantifiers.
This is clearly finitely branching; actually, each σ has at most two extensions—
one by 0 and one by the least si −1 with φ(si −1, i). To see that this is infinite,
it suffices to show that for each n, there is a sequence hs0 , . . . , sn i ∈ T ; we
may take this to be the sequence where si is x + 1 where x ≤ n is least such
that φ(x, i), if there is any, and 0 otherwise. This sequence always exists,
and always belongs to T .
By König’s Lemma, there is an infinite sequence hsn i through this. We
claim that ∃xφ(x, y) iff sy 6= 0. Suppose ∃xφ(x, y); then there is a least such
x. The sequence hs0 , . . . , sy−1 , 0, . . . , sx i 6∈ T , so we cannot have sy = 0. On
the other hand, suppose ∀xφ(x, y). Then hs0 , . . . , sy i ∈ T implies sy = 0
since the second clause can never be satisfied. 
Just as the first order part of ACA0 is exactly PA, it can be shown that
the first order part of RCA0 is exactly IΣ1 .
4.6. WKL0 and other theories weaker than ACA0 .
One of the surprising discoveries in the early investigation of reverse math-
ematics was that statements from analysis and algebra tended to group into
a small number of categories. In fact, almost all the early theorems investi-
gated turned out to be equivalent to one of five theories, the so-called Big 5
theories of reverse mathematics. RCA0 and ACA0 are two of them.
Definition 4.20. The theory WKL0 consists of the axioms of RCA0 to-
gether with the following axiom:
Suppose S is a tree of sequences of 0’s and 1’s containing
infinitely many elements. Then there is an infinite sequence
hsn i such that all its initial segments are in S.
This axiom is called “weak König’s lemma”. It is clearly a special case
of König’s lemma, so WKL0 is between RCA0 and ACA0 . To see that
WKL0 is actually stronger than RCA0 is typically an exercise in com-
putability: it is easy to show that there is a model of RCA0 where the sets
are exactly the computable sets, and one then shows that is not a model
of WKL0 . On the other hand, it can be shown that not only is WKL0
weaker than ACA0 , it’s actually quite close to RCA0 .
14

Theorem 4.21. If WKL0 ` φ where φ is arithmetic then RCA0 ` φ.

There are several proofs of this theorem; the simplest one, due to Har-
rington, uses forcing to show that any model of RCA0 can be extended
to a model of WKL0 by adding additional sets. An alternate proof, due
to Ferreira and Ferreira, uses cut-elimination. It immediately follows that
the same holds for Π11 sentences, since deducing ∀Xφ(X) is equivalent to
deducing φ(X) with X free.
One consequence is that the proof-theoretic ordinal of WKL0 is the same
as that of RCA0 (ω ω ); this is easily seen since stating that an ordinal is
well-founded is a Π11 sentence. Nonetheless, WKL0 proves many theorems
about the existence of sets that RCA0 does not. For example:

Theorem 4.22. The statement that every consistent theory has a complete
consistent extension is equivalent to WKL0 (using RCA0 as a base theory).

Proof. Recall that we can formalize all our proof theory, interpreting for-
mulas by sequences of symbols, which are in turn coded by numbers, and
deductions as sequences of formulas which can again be coded by numbers.
Let S be a set of formulas with the property that there is no deduction of
S 0 ⇒ ⊥ where S 0 ⊆ S.
For each sequence σ of 0’s and 1’s, we will define a finite set of formulas
Xσ : if n < |σ| and n codes a formula φ then n ∈ Xσ if the n-th position of σ
is 1, and the number coding ¬φ is in Xσ if the n-th position of σ is 0. Now
we define a tree T consisting of those σ such that there is no deduction of
S 0 ⇒ ⊥ where S 0 ⊆ S ∪ Xσ and the code of the deduction is ≤ |σ|.
Note that we really need the bound on the code of the deduction in
order to prove in RCA0 that the tree exists: the statement that S ∪ Xσ is
consistent—that is, that there is no deduction at all—is Σ1 .
We claim that for every n, there is a σ ∈ T with |σ| = n: if not, for all the
finitely many σ of length n we have a deduction of ⊥ from some subset of
S ∪Xσ . But given a deduction of ⊥ from S0 ∪{φ} and a deduction of ⊥ from
S1 ∪ {¬φ}, we may easily derive a deduction of ⊥ from S0 ∪ S1 . Iterating
this argument finitely many times, we obtain a deduction of ⊥ from some
subset of S, contradicting our assumption. S
By WKL0 , there is an infinite branch Λ through T . Take X = σ@Λ Xσ .
Clearly S ⊆ X and X is consistent: if there were any deduction of a contra-
diction, it would be coded by some number, and use only a finite subset of
X, and therefore at some finite length an initial segment would be rejected
from T . For any formula φ, φ is coded by some n, so when |σ| = n + 1,
we have ensured either φ is in Xσ or the negation of φ is. Therefore X is a
complete, consistent extension of S, as desired.
For the converse, suppose every consistent theory has a complete exten-
sion, and let T be a tree of 0’s and 1’s with infinitely many elements. We
define a language with countably many zero-ary relation symbols Pσ , one
for each sequence σ. Define a theory S to consist of:
 15
W
• For each n, the formula σ,|σ|=n Pσ ,
• If |σ| = |τ |, Pσ → ¬Pτ ,
• If σ 6∈ T , ¬Pσ .
Observe that this theory is consistent: a deduction of a contradiction would
use finitely many sentences from S, and in particular, only contains symbols
Pσ with |σ| ≤ n for some n. But the formulas with such propositional
variables have a model, namely take any σ ∈ T and make Pτ true iff τ v σ.
By
S assumption, this theory has a complete extension X, and we take Λ =
Pσ ∈X σ. 

4.7. Theories stronger than ACA0 .

The other two theories of the Big 5 are stronger than ACA0 .
The usual definition of the theory ATR0 is a bit technical, so we use the
following equivalent definition:
Definition 4.23. ATR0 consists of:
• RCA0 ,
• Whenever ≺α and ≺β are well-orders, either there is a function f
mapping the domain of ≺α onto an initial segment of the domain of
≺β , or vice-versa.
ATR0 stands for “arithmetic transfinite recursion”. The standard for-
mulation uses an axiom which states that any arithmetic operation can be
iterated along any well-ordering.
We briefly describe the proof-theoretic ordinal of ATR0 . We define a two
place function on ordinals, ϕ(α, β), as follows:
• φ(0, β) = ω β ,
• φ(α, β) is the least ordinal γ such that:
– If β 0 < β then φ(α + 1, β 0 ) < γ,
– If δ < γ and α0 < α then φ(α0 , δ) < γ.
ϕ is called the Veblen fucntion. (Sadly, there are some slight variations in
the precise definition, though this version is fairly typical.) Observe that
φ(1, β) = β , the β-th ordinal closed under the operation γ 7→ ω γ . More
generally, φ(α, β) is closed under γ 7→ φ(α0 , γ) for each α < α0 .
The ordinal Γ0 is the least ordinal such that whenever α, β < Γ0 , φ(α, β) <
Γ0 . The ordinal Γ0 has been associated with the philosophical position of
predicativism: the idea that the well-defined mathematical notions are those
which can be constructed from objects which have been constructed at a
previous stage.
The final, strongest, theory in the Big 5 is:
Definition 4.24. Π11 -CA0 consists of:
• P −,
• The second order induction axiom,
• The comprehension scheme for Π11 formulas.
16

Π11 -CA0 obviously includes ACA0 . The proof-theoretic ordinal of Π11 -CA0
is too complicated to describe here, but we briefly explain why. Suppose we
wanted to replace the Π11 comprehension axiom the same way we replaced
the arithmetic comprehension axiom. The main case is where we introduce
some formula on the left-hand side:
ΓACA0 , Γ, ∀x(x ∈ X ↔ ∀Y φ(x, Y )) ⇒ Σ.
In ACA0 , we could replace the variable X with ∀Y φ(x, Y ) everywhere it
appeared in the deduction. Here, if we can try to do the same, we discover
that the variable X could have been used to deduce instances of ∃Y ¬φ(x, Y )!
In other words, the set X consisting of those x where ∀Y φ(x, Y ) was true
might have used the set X itself to demonstrate that ∀Y φ(x, y) was false for
some particular x!
This is not a minor issue. Π11 -CA0 introduces a deep new obstacle to
cut-elimination: which elements belong to a set defined by a Π11 formula
depends on a quantification over all sets, including the set which is in the
process of being defined. This is a genuine circularity, and raises genuine
new difficulties, both mathematically and philosophically. (We should con-
sider that it is not obvious, when stated like this, that Π11 comprehension is
actually well-defined at all: how can we be sure that we won’t accidentally
write down some set X with the property that x ∈ X iff x 6∈ X?) This the-
ory is impredicative: the meaning of a set defined by a Π11 formula depends
essentially on a family of objects—the sets of numbers—which has not yet
been completely constructed.
For the sake of comparison with the literature, we mention that the proof
theoretic ordinal of Π11 -CA0 is given as the limit of a sequence of ordinals,
with most (though not all) of the technical work needed to produce the first
one (the later ones iterate the same idea relative to the first one). The first
of these ordinals is known as the Howard-Bachmann ordinal, and is usually
written ψΩ+1 . Here Ω is (a name for) some very large ordinal, larger than
any other in the system—for example, ℵ1 —Ω+1 is the first fixed point of
the α 7→ ω α which is larger than Ω, and ψ is a collapsing function which
takes a very large ordinal and “collapses” it to a smaller one. (Alternatively,
one can view Ω+1 as representing the function α 7→ α+1 and ψ as being
an operation which chooses an ordinal which is both a fixed point of this
function and also which is larger than the chosen fixed point of all “easily
defined” functions which grow more slowly.) Note that this definition is
itself impredicative—we define the ordinal in terms of larger ordinals (or
the class of functions on ordinals).