Cite

This course has no released version yet: the text at this address may change. Cite a released version when one exists.

/courses/computable-ordinals/trees-and-pi11
@misc{computable-ordinals-trees-and-pi11,
  title = {{Well-founded trees and Π¹₁ sets}},
  howpublished = {Lesson 3 of \emph{Computable ordinals}, Computability Theory course notes},
  url = {/courses/computable-ordinals/trees-and-pi11},
  note = {Unreleased working version}
}

To cite a single statement, use the # link that appears next to its heading.

Lesson 3 of 3

Well-founded trees and Π¹₁ sets

Well-founded trees and their ranks; the Kleene–Brouwer ordering; the normal form of Π¹₁ sets as sets of indices of well-founded trees; the Π¹₁-completeness of well-foundedness and well-ordering; Π¹₁ is not Σ¹₁; Spector's boundedness theorem.

45 min3 questions
Computable ordinals · lesson 3 of 3
  1. Well-orderings and ordinals
  2. Computable ordinals and Kleene's O
  3. Well-founded trees and Π¹₁ sets

The arithmetic hierarchy classifies sets defined by quantifiers over numbers. Quantifying over sets or functions gives the analytic hierarchy, whose first level, the Π11\Pi^1_1 sets, is the central object of higher computability. This lesson shows that the Π11\Pi^1_1 sets are exactly the sets {n:Tn is well-founded}\{ n : T_n \text{ is well-founded} \} for uniformly computable sequences of trees (Tn)(T_n), that being a well-ordering is a complete Π11\Pi^1_1 property, and that the Σ11\Sigma^1_1 sets of well-orderings have bounded order types. The last result, Spector’s boundedness theorem, is the main tool of the next course.

3.1 Trees on ℕ and their ranks

A tree T⊆N<NT \subseteq \nat^{<\nat} is a set of finite sequences of numbers closed under prefixes. A path of TT is a function f∈NNf \in \nat^\nat all of whose prefixes belong to TT. The tree is well-founded if it has no path. For σ∈T\sigma \in T, write Tσ={τ:στ∈T}T_\sigma = \{ \tau : \sigma\tau \in T \}.

Definition 3.1.#

The rank of a well-founded tree is defined by transfinite recursion: rk(T)=sup⁡{rk(T⟨n⟩)+1:⟨n⟩∈T}\mathrm{rk}(T) = \sup \{ \mathrm{rk}(T_{\langle n \rangle}) + 1 : \langle n \rangle \in T \}, and rk(∅)=0\mathrm{rk}(\emptyset) = 0.

The recursion is legitimate because the relation ”τ\tau properly extends σ\sigma”, restricted to a well-founded tree, has no infinite increasing chains. The rank of a well-founded tree is a countable ordinal, and every countable ordinal is the rank of some well-founded tree. A tree with a path has no rank.

The Kleene–Brouwer ordering on N<N\nat^{<\nat} is defined by σ<KBτ\sigma <_{\mathrm{KB}} \tau if σ\sigma properly extends τ\tau, or if σ(i)<τ(i)\sigma(i) < \tau(i) at the first position ii where they differ. It is a computable linear order: extensions come before their prefixes, and incomparable sequences are compared lexicographically.

Theorem 3.2.#

A tree TT is well-founded if and only if <KB<_{\mathrm{KB}} restricted to TT is a well-ordering. In that case, its order type is at least rk(T)\mathrm{rk}(T).

Proof of Theorem 3.2.

If TT has a path ff, then f ⁣↾0>KBf ⁣↾1>KB⋯f \uh 0 >_{\mathrm{KB}} f \uh 1 >_{\mathrm{KB}} \cdots is a descending sequence.

Suppose TT well-founded; we prove by induction on rk(T)\mathrm{rk}(T) that (T,<KB)(T, <_{\mathrm{KB}}) is a well-ordering of type at least rk(T)\mathrm{rk}(T). If TT is empty, both are 00. Otherwise, the nodes of TT other than the root are the sequences ⟨n⟩τ\langle n \rangle \tau with τ∈T⟨n⟩\tau \in T_{\langle n \rangle}, and the ordering compares first the initial numbers nn, then the sequences τ\tau by the Kleene–Brouwer ordering; the root comes last. So (T,<KB)(T, <_{\mathrm{KB}}) is the ordered sum of the orderings (T⟨n⟩,<KB)(T_{\langle n \rangle}, <_{\mathrm{KB}}) for n=0,1,2,…n = 0, 1, 2, \dots with ⟨n⟩∈T\langle n \rangle \in T, followed by one element. Each summand is a well-ordering of type at least rk(T⟨n⟩)\mathrm{rk}(T_{\langle n \rangle}), by the induction hypothesis, and an ordered sum of well-orderings indexed by a well-ordering is a well-ordering. Its type is at least rk(T⟨n⟩)+1\mathrm{rk}(T_{\langle n \rangle}) + 1 for every nn, hence at least rk(T)\mathrm{rk}(T).

3.2 Π¹₁ sets and their normal form

Definition 3.3.#

A set A⊆NA \subseteq \nat is Π11\Pi^1_1 if there is an arithmetic formula φ(n,X)\varphi(n, X) such that n∈A  ⟺  ∀X⊆N φ(n,X)n \in A \iff \forall X \subseteq \nat\ \varphi(n, X). It is Σ11\Sigma^1_1 if its complement is Π11\Pi^1_1, and Δ11\Delta^1_1 if it is both.

Quantifiers over functions f∈NNf \in \nat^\nat can be used instead of quantifiers over sets, by coding a function by its graph; the class is the same. Every arithmetic set is Δ11\Delta^1_1. Relativized to an oracle YY, one obtains the Π11(Y)\Pi^1_1(Y) sets.

Theorem 3.4 (Kleene normal form).#

A set AA is Π11\Pi^1_1 if and only if there is a uniformly computable sequence of trees (Tn)(T_n) such that n∈An \in A if and only if TnT_n is well-founded.

Proof of Theorem 3.4.

(⇐\Leftarrow) TnT_n is well-founded if and only if ∀f ∃m f ⁣↾m∉Tn\forall f\ \exists m\ f \uh m \notin T_n, a Π11\Pi^1_1 formula.

(⇒\Rightarrow) Write n∈A  ⟺  ∀f φ(n,f)n \in A \iff \forall f\, \varphi(n, f) with φ\varphi arithmetic in prenex form. We first show, by induction on the number of number quantifiers, that every prenex arithmetic formula ψ(f)\psi(f) is equivalent to a formula ∀h ∃y θ(f,h,y)\forall h\ \exists y\ \theta(f, h, y) with θ\theta bounded.

  • A bounded formula θ\theta is equivalent to ∀h ∃y θ\forall h\, \exists y\, \theta, with dummy variables.
  • If ψ≡∀x χ(x)\psi \equiv \forall x\, \chi(x) and χ(x)≡∀h ∃y θ(x,h,y)\chi(x) \equiv \forall h\, \exists y\, \theta(x, h, y), then ψ≡∀h′ ∃y θ(h′(0),h′′,y)\psi \equiv \forall h'\, \exists y\, \theta(h'(0), h'', y), where h′′(i)=h′(i+1)h''(i) = h'(i+1): the single function h′h' codes both xx and hh.
  • If ψ≡∃x χ(x)\psi \equiv \exists x\, \chi(x) and χ(x)≡∀h ∃y θ(x,h,y)\chi(x) \equiv \forall h\, \exists y\, \theta(x, h, y), then ψ≡∀H ∃x ∃y θ(x,H[x],y)\psi \equiv \forall H\, \exists x\, \exists y\, \theta(x, H^{[x]}, y), where H[x](i)=H(⟨x,i⟩)H^{[x]}(i) = H(\langle x, i \rangle). Indeed, if x0x_0 witnesses ψ\psi, then χ(x0)\chi(x_0) applies to H[x0]H^{[x_0]}. Conversely, if ψ\psi fails, choose for each xx a function hxh_x with ¬∃y θ(x,hx,y)\neg \exists y\, \theta(x, h_x, y), and let H[x]=hxH^{[x]} = h_x; then ∃x ∃y θ(x,H[x],y)\exists x\, \exists y\, \theta(x, H^{[x]}, y) fails. The two existential quantifiers merge into one.

So n∈A  ⟺  ∀g ∃y θ(n,g,y)n \in A \iff \forall g\, \exists y\, \theta(n, g, y), where gg codes ff and the auxiliary function, and θ\theta is bounded. The evaluation of the bounded formula θ(n,g,y)\theta(n, g, y) queries gg at finitely many arguments only. Let R(n,σ)R(n, \sigma) hold if for some y<∣σ∣y < |\sigma|, the evaluation of θ(n,g,y)\theta(n, g, y) with σ\sigma in place of gg queries only arguments below ∣σ∣|\sigma| and returns true. Then RR is computable, and ∃y θ(n,g,y)  ⟺  ∃m R(n,g ⁣↾m)\exists y\, \theta(n, g, y) \iff \exists m\, R(n, g \uh m): from left to right, take mm larger than yy and than every argument queried; from right to left, gg and g ⁣↾mg \uh m agree on the queried arguments. Let Tn={σ:¬R(n,τ) for every τ⪯σ}T_n = \{ \sigma : \neg R(n, \tau) \text{ for every } \tau \preceq \sigma \}. Then TnT_n is a tree, uniformly computable, and it has a path gg if and only if ∃y θ(n,g,y)\exists y\, \theta(n, g, y) fails for gg, that is, if and only if n∉An \notin A.

The proof uses the axiom of choice in the third case, a countable choice of functions; this is harmless in the metatheory.

Corollary 3.5.#

Let WF\mathrm{WF} be the set of indices of computable trees that are well-founded, and WO\mathrm{WO} the set of indices of computable linear orders that are well-orderings. Both are Π11\Pi^1_1, and every Π11\Pi^1_1 set is many-one reducible to each of them.

Proof of Corollary 3.5.

Both definitions are Π11\Pi^1_1: “for every ff, ff is not a path”, and “for every ff, ff is not a descending sequence”, together with the arithmetic conditions of being a tree or a linear order. If AA is Π11\Pi^1_1, the normal form gives a computable function n↦n \mapsto an index of TnT_n, reducing AA to WF\mathrm{WF}. Composing with the computable map sending an index of a tree to an index of its Kleene–Brouwer ordering reduces AA to WO\mathrm{WO}, by Theorem 3.2.

The set O\mathcal O of notations is also Π11\Pi^1_1: it is the least set closed under the three rules of Definition 2.4, so a∈Oa \in \mathcal O if and only if aa belongs to every set that is closed under the rules, together with a relation playing the role of <O<_{\mathcal O}; this is a Π11\Pi^1_1 condition. It is also Π11\Pi^1_1-complete, a result we admit (Sacks, Higher Recursion Theory, Section I.5).

3.3 Π¹₁ is not Σ¹₁

Proposition 3.6.#

There is a Π11\Pi^1_1 set that is not Σ11\Sigma^1_1. In particular WF\mathrm{WF} and WO\mathrm{WO} are not Σ11\Sigma^1_1.

Proof of Proposition 3.6.

For each ee and nn, let Te,nT_{e,n} be the set of sequences σ\sigma such that no prefix τ⪯σ\tau \preceq \sigma satisfies Φe(⟨n,τ⟩)[∣σ∣] ⁣ ⁣↓=0\Phi_e(\langle n, \tau \rangle)[|\sigma|]\converge = 0. It is a tree, computable uniformly in ee and nn, even when Φe\Phi_e is partial. Let (Tn)(T_n) be given by Theorem 3.4 for a Π11\Pi^1_1 set AA, with Tn={σ:¬R(n,τ) for all τ⪯σ}T_n = \{ \sigma : \neg R(n, \tau) \text{ for all } \tau \preceq \sigma \}, and let Φe\Phi_e compute the characteristic function of RR, with value 00 when RR holds. A path of TnT_n is a path of Te,nT_{e,n}. Conversely, if ff is a path of Te,nT_{e,n}, then for each kk the computation Φe(⟨n,f ⁣↾k⟩)\Phi_e(\langle n, f \uh k \rangle) converges at some stage ss, and f ⁣↾m∈Te,nf \uh m \in T_{e,n} for m≥max⁡(s,k)m \geq \max(s, k) shows that R(n,f ⁣↾k)R(n, f \uh k) fails; so ff is a path of TnT_n. Hence Te,nT_{e,n} is well-founded if and only if TnT_n is, and U={⟨e,n⟩:Te,n is well-founded}U = \{ \langle e, n \rangle : T_{e,n} \text{ is well-founded} \} is a Π11\Pi^1_1 set whose sections Ue={n:⟨e,n⟩∈U}U_e = \{ n : \langle e, n \rangle \in U \} include all the Π11\Pi^1_1 sets.

Suppose that UU were Σ11\Sigma^1_1. Then its complement would be Π11\Pi^1_1, and so would be D={n:⟨n,n⟩∉U}D = \{ n : \langle n, n \rangle \notin U \}, its preimage under the computable map n↦⟨n,n⟩n \mapsto \langle n, n \rangle; hence D=UeD = U_e for some ee; then e∈D  ⟺  ⟨e,e⟩∉U  ⟺  e∉Ue=De \in D \iff \langle e, e \rangle \notin U \iff e \notin U_e = D, a contradiction. So UU is not Σ11\Sigma^1_1. Since UU reduces to WF\mathrm{WF} and WO\mathrm{WO}, these are not Σ11\Sigma^1_1 either: Σ11\Sigma^1_1 sets are closed under computable preimages.

3.4 Spector’s boundedness theorem

For an index ee of a computable linear order, write LeL_e for this order; for e∈WOe \in \mathrm{WO}, its order type ∣Le∣|L_e| is a computable ordinal.

Theorem 3.7 (Spector).#

If A⊆WOA \subseteq \mathrm{WO} is Σ11\Sigma^1_1, then sup⁡{∣Le∣:e∈A}<ω1CK\sup \{ |L_e| : e \in A \} < \omega_1^{\mathrm{CK}}.

Proof of Theorem 3.7.

The supremum is at most ω1CK\omega_1^{\mathrm{CK}}; suppose that it is equal to ω1CK\omega_1^{\mathrm{CK}}. Then for every index ee,

e∈WO  ⟺  Le is a linear order and ∃d (d∈A∧∃h (h is an order-preserving injection from Le into Ld)).\begin{aligned} e \in \mathrm{WO} \iff{} & L_e \text{ is a linear order and } \exists d\, \big(d \in A \wedge \exists h\ \\ & (h \text{ is an order-preserving injection from } L_e \text{ into } L_d)\big). \end{aligned}

If e∈WOe \in \mathrm{WO}, then ∣Le∣<ω1CK|L_e| < \omega_1^{\mathrm{CK}}, so some d∈Ad \in A has ∣Ld∣≥∣Le∣|L_d| \geq |L_e|, and LeL_e embeds into LdL_d by the comparison theorem (Theorem 1.3). Conversely, a linear order that embeds into a well-ordering is a well-ordering, since a descending sequence would be mapped to a descending sequence. The right-hand side is Σ11\Sigma^1_1: "d∈Ad \in A" is Σ11\Sigma^1_1, ”hh is an order-preserving injection” is arithmetic, and existential quantifiers combine. So WO\mathrm{WO} would be Σ11\Sigma^1_1, contradicting Proposition 3.6.

Corollary 3.8.#
  1. For every computable ordinal α\alpha, the set WO<α\mathrm{WO}_{< \alpha} of indices e∈WOe \in \mathrm{WO} with ∣Le∣<α|L_e| < \alpha is Δ11\Delta^1_1.
  2. The set O\mathcal O is not Σ11\Sigma^1_1.
Proof of Corollary 3.8.
  1. Let dd be an index of a computable well-ordering of type α\alpha. Then e∈WO<αe \in \mathrm{WO}_{<\alpha} if and only if LeL_e is a linear order that embeds into a proper initial segment of LdL_d, a Σ11\Sigma^1_1 condition; and if and only if e∈WOe \in \mathrm{WO} and LdL_d does not embed into LeL_e, a Π11\Pi^1_1 condition, since ”LdL_d embeds into LeL_e” is Σ11\Sigma^1_1.
  2. Let ff be the total computable function of Proposition 2.6, which maps every a∈Oa \in \mathcal O to an index of a computable well-ordering of type ∣a∣|a|. If O\mathcal O were Σ11\Sigma^1_1, then {e:∃a (a∈O∧f(a)=e)}\{ e : \exists a\, (a \in \mathcal O \wedge f(a) = e) \} would be a Σ11\Sigma^1_1 subset of WO\mathrm{WO} containing well-orderings of every computable type, by Corollary 2.10. This contradicts Theorem 3.7.

3.5 Exercises

Exercise 3.9 ★.

Show that the set of indices ee such that WeW_e, viewed as a set of sequences, is a well-founded tree is Π11\Pi^1_1, and that every Σ11\Sigma^1_1 set of such indices has ranks bounded below ω1CK\omega_1^{\mathrm{CK}}.

Solution

”WeW_e is closed under prefixes” is arithmetic and “no function is a path of WeW_e” is Π11\Pi^1_1. For the bound, the Kleene–Brouwer ordering restricted to WeW_e is a c.e. linear order, uniformly in ee; preceded by a copy of ω\omega, it has an infinite field, and Proposition 2.2 (3) gives a total computable function ff such that f(e)f(e) is an index of a computable linear order isomorphic to ω+(We,<KB)\omega + (W_e, <_{\mathrm{KB}}). If WeW_e is a well-founded tree, then f(e)∈WOf(e) \in \mathrm{WO} and ∣Lf(e)∣≥rk(We)|L_{f(e)}| \geq \mathrm{rk}(W_e), by Theorem 3.2. The image under ff of a Σ11\Sigma^1_1 set BB of such indices, {d:∃e (e∈B∧f(e)=d)}\{ d : \exists e\, (e \in B \wedge f(e) = d) \}, is a Σ11\Sigma^1_1 subset of WO\mathrm{WO}, so its types, and hence the ranks of the trees WeW_e, e∈Be \in B, are bounded below ω1CK\omega_1^{\mathrm{CK}} by Theorem 3.7.

Exercise 3.10 ★★.

Show that ω1CK\omega_1^{\mathrm{CK}} is the supremum of the ranks of the computable well-founded trees, and that some computable tree has a path but no computable path.

Solution

A computable well-founded tree has computable Kleene–Brouwer ordering, of type at least its rank, so its rank is computable. Conversely, a computable well-ordering LL of type α\alpha gives the computable tree of finite LL-descending sequences, which is well-founded, of rank at least α\alpha: by induction, the subtree above a one-element sequence ⟨x⟩\langle x \rangle has rank at least the type of the initial segment below xx.

For the second statement, let TT be the set of σ∈2<N\sigma \in 2^{<\nat} such that σ(e)≠Φe(e)\sigma(e) \neq \Phi_e(e) for every e<∣σ∣e < |\sigma| with Φe(e)[∣σ∣] ⁣ ⁣↓\Phi_e(e)[|\sigma|]\converge. It is a computable tree, and its paths are the {0,1}\{0,1\}-valued functions ff with f(e)≠Φe(e)f(e) \neq \Phi_e(e) whenever Φe(e) ⁣ ⁣↓\Phi_e(e)\converge. It has a path, for instance f(e)=1f(e) = 1 if Φe(e) ⁣ ⁣↓=0\Phi_e(e)\converge = 0 and f(e)=0f(e) = 0 otherwise. No path ff is computable: if f=Φef = \Phi_e, then Φe(e) ⁣ ⁣↓=f(e)\Phi_e(e)\converge = f(e), a contradiction. (This is the tree of Theorem 2.6 of “Computable members”.)

3.6 Notes and references

Kleene (1955) introduced the analytic hierarchy and proved the normal form theorem; the use of well-founded trees is due to Lusin and Sierpiński in descriptive set theory, and to Brouwer for the ordering. Spector (1955) proved the boundedness theorem. Moschovakis’s Descriptive Set Theory presents the effective theory in the general setting of Polish spaces.

  • S. C. Kleene, Arithmetical predicates and function quantifiers, Transactions of the American Mathematical Society 79 (1955), 312–340.
  • C. Spector, Recursive well-orderings, Journal of Symbolic Logic 20 (1955), 151–163.
  • Y. N. Moschovakis, Descriptive Set Theory, 2nd edition, American Mathematical Society, 2009.
  • G. E. Sacks, Higher Recursion Theory, Springer, 1990.