Skip to article frontmatterSkip to article content

Recursion and the Von-Neumann Hierarchy

Penn State University

Transfinite induction

While the class Ord\Ord of all ordinals is not a set, it is still transitive and well-ordered by \in. Regarding the associated order \leq, every set of ordinals aa has a supremum a=ξaξ\bigcup a = \bigcup_{\xi \in a} \xi and (if aa \ne \emptyset) an infimum a=ξaξ\bigcap a = \bigcap_{\xi \in a} \xi, which is the smallest element of aa. Such a smallest element exists actually for every (non-empty) class AA (since if ξA\xi \in A, we only need to find the infimum of the set of ordinals ξ\le \xi.) This allows us to prove properties about all ordinals by induction.

We have repeatedly used induction already for ordinals <ω1< \omega_1, the first uncountable ordinal.

To prove this principle simply observe that if αφ(α)\forall \alpha \, \varphi(\alpha) failed there would have to be a smallest α{}\alpha with ¬φ(α)\neg \varphi(\alpha), contradicting the induction hypothesis.

Since every ordinal is either 0, a successor, or a limit ordinal, we have the following variant of induction.

(i) and (ii) coincide with the usual induction scheme for natural numbers. To cover all ordinals we need to add (iii).

Ordinal recursion

The induction principle can be used to define functions by recursion. For example, addition on the natural numbers is given by

x+0=xx+(y+1)=(x+y)+1.\begin{align*} x+0 \quad & = x\\ x+ (y+1) & = (x + y)+1. \end{align*}

In the case of ordinals, we have to consider the limit case, too.

Proof

The uniqueness of the function FF follows by induction.

To show the existence of FF, we define the following:

  • Call hh tame if
α(h:αVξα  h(ξ)=G(ξ,hα))\exists \alpha \, (h: \alpha \to \V \wedge \forall \xi \in \alpha \; h(\xi) = G(\xi, h \Rest{\alpha}))
  • Say hh is compatible with gg if
xDom(h)Dom(g)  h(x)=g(x)\forall x \in \Op{Dom}(h) \cap \Op{Dom}(g) \; h(x) = g(x)

It follows by induction that any two tame functions are compatible.

This lets us define the desired FF as

F:={h ⁣:h tame}F := \bigcup \{h \colon h\, \text{ tame}\}

Then FF is a function (otherwise there would be two incompatible tame functions), its domain is transitive, and satisfies the recursion condition (since it is the union of tame functions).

It remains to show that FF is defined on all of Ord\Ord. If D=Dom(F)OrdD = \Op{Dom}(F) \neq \Ord, then we would have D=αD = \alpha for some ordinal α{}\alpha. In particular BB is a set therefore F=fF = f is a set, for some tame ff. This ff could be extended to a tame h=f{(α,G(α,fα))}h = f \cup \{(\alpha,G(\alpha,f \Rest{\alpha}))\}, contradiction.

Note that we defined FF explicitly as a union of all partial solutions to the recursion equation.

As with induction, we have the following variant of the recursion principle.

We can establish a similar principle for a well-ordering << on a class AA. In case of a proper class, though, we have to require that for every aAa \in A, the class of all predecessors of aa

S(a,<):={xA ⁣:x<a},S(a,<): = \{x \in A\colon x < a \},

is a set (if AA is a set, this follows automatically by Separation). If this is the case, the recursion principle yields a function F:AVF: A \to \V such that

F(a)=G(a,FS(a,<)).F(a) = G(a,F\Rest{S(a,<)}).

Recursion for well-founded relations

More generally, we can define induction and recursion on well-founded relations. We already encountered those in a previous chapter.

If AA is a set, the minimality condition is again automatically satisfied by Separation.

The set condition allows for taking the RR-transitive closure of a set aAa \in A: the smallest superset TCR(a)\Op{TC}_R(a) of aa that is RR-transitive:

xTCR(a)  S(x,R)TCR(a)\forall x \in \Op{TC}_R(a)\; S(x,R) \subseteq \Op{TC}_R(a)

This is done by recursion over the natural numbers. The following is an important example.

We can use the existence of TCR\Op{TC}_R as a set to strengthen the minimality condition to subclasses, similar to the case of the well-ordering of Ord\Ord:

To prove this lemma, simply pick any xBx \in B, take its transitive RR-closure, and intersect it with BB:

C=TCR(x)B.C = \Op{TC}_R(x) \cap B.

CC is a set, and by the minimality condition (MinR)(\Op{Min}_R) has an RR-minimal element aa. aa has to be minimal for BB, too, since otherwise there exists bBb \in B with bRab R a. Since aTCR(x)a \in \Op{TC}_R(x), bTCR(x)b \in \Op{TC}_R(x), and therefore bCb \in C, contradicting the minimality of aa.

The lemma implies a corresponding induction principle for well-founded relations:

(IndR)xA[y(yRxφ(y))φ(x)]xAφ(x)).(\Op{Ind}_R) \qquad \forall x \in A [ \forall y ( yRx \, \to \varphi(y)) \to \varphi(x)] \to \forall x \in A \, \varphi(x)).

This in turn yields the following.

The Von Neumann hierarchy

Is there a way to systematically build the V\V, the universe of all sets, “from below”?

We start with the empty set:

V0=V_0 = \emptyset

Given VαV_\alpha, the Power Set axiom requires the set of all subsets to exist, so we set

Vα+1=P(Vα).V_{\alpha+1} = \mathcal{P}(V_{\alpha}).

Finally, at limit stages we simply collect all sets we have obtained so far:

Vλ=ξ<λVξfor limit λV_ \lambda = \bigcup_{\xi < \lambda} V_\xi \quad \text{for limit } \lambda

What we really are doing here is to construct a function V:OrdVV: \Ord \to \V by ordinal recursion. Think Vα=V(α)V_\alpha = V(\alpha).

The following facts about the hierarchy are easily verified by transfinite induction:

Remarkably, if we assume the axiom of Foundation, we reach all sets this way.

Proof

Let CC be the class of all sets not in any VαV_\alpha. Since \in is well-founded, if CC is non-empty, it has a \in-minimal element xx. This implies that for all zxz \in x, zαOrdVαz \in \bigcup_{\alpha \in \Ord} V_\alpha. Define a function hh by mapping each zxz\in x to the least α{}\alpha so that zVαz \in V_\alpha. Since xx is a set, h[x]h[x] is a set of ordinals, by Replacement. This set of ordinals has a supremum, say γ{}\gamma. Then xVγx \subseteq V_\gamma and therefore,

xP(Vγ)=Vγ+1.x \in \mathcal{P}(V_\gamma) = V_{\gamma+1}.

Hence CC must be empty, and the theorem follows.

We can split the question of “how large” V\V is into two sub-questions:

  • How “long” is V\V, that is, how many ordinals are there? Axioms for large cardinals attempt to extend this “length” as far as possible.
  • How “wide” is V\V, that is, how large is the power set of a set? A rather “slim” universe is given by the constructible sets, which we will encounter soon.