Cite

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

/courses/big-five/wkl0
@misc{big-five-wkl0,
  title = {{Weak König's lemma}},
  howpublished = {Lesson 2 of \emph{The Big Five}, Computability Theory course notes},
  url = {/courses/big-five/wkl0},
  note = {Unreleased working version}
}

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

Lesson 2 of 3

Weak König's lemma

The system WKL₀; its ω-models are the Scott ideals; there is a Scott ideal of low sets, so WKL₀ does not prove ACA₀; WKL₀ is equivalent to Σ⁰₁ separation and to the Heine–Borel covering theorem for [0, 1].

40 min4 questions
Before this lesson:Arithmetic comprehension · PA degrees (Π⁰₁ classes)
The Big Five · lesson 2 of 3
  1. Arithmetic comprehension
  2. Weak König's lemma
  3. Transfinite recursion and Π¹₁ comprehension

Weak König’s lemma, the statement that every infinite binary tree has an infinite path, is the set-existence principle behind compactness arguments. The system WKL0\mathsf{WKL}_0 lies strictly between RCA0\mathsf{RCA}_0 and ACA0\mathsf{ACA}_0, and its computability-theoretic counterpart is the notion of PA degree. It is equivalent to many theorems of analysis and algebra that are proved by compactness: the Heine–Borel theorem, the boundedness of continuous functions on [0,1][0, 1], the existence of prime ideals in countable commutative rings, Gödel’s completeness theorem for countable languages.

2.1 The system and its ω-models

Definition 2.1.#

Weak König’s lemma (WKL\mathsf{WKL}) is the statement: every infinite tree T⊆2<NT \subseteq 2^{<\nat} has an infinite path. WKL0\mathsf{WKL}_0 is RCA0+WKL\mathsf{RCA}_0 + \mathsf{WKL}.

Being infinite and being a path are arithmetic properties, so WKL\mathsf{WKL} is a Π21\Pi^1_2 statement: for every tree, if it is infinite, there exists a path.

Definition 2.2.#

A Scott ideal is a Turing ideal I\mathcal I such that every infinite binary tree T∈IT \in \mathcal I has an infinite path in I\mathcal I.

Proposition 2.3.#
  1. An ω-model satisfies WKL0\mathsf{WKL}_0 if and only if it is a Scott ideal.
  2. A Turing ideal I\mathcal I is a Scott ideal if and only if for every X∈IX \in \mathcal I there is Y∈IY \in \mathcal I that is PA relative to XX.
Proof of Proposition 2.3.
  1. By Friedman and absoluteness: a tree in the model is infinite in the model if and only if it is infinite, and a set of the model is a path in the model if and only if it is a path.
  2. (⇒\Rightarrow) Given X∈IX \in \mathcal I, the XX-computable tree whose paths are the {0,1}\{0, 1\}-valued functions DNC relative to XX belongs to I\mathcal I and is infinite, so it has a path Y∈IY \in \mathcal I, which is PA relative to XX. (⇐\Leftarrow) Let T∈IT \in \mathcal I be an infinite binary tree, and Y∈IY \in \mathcal I PA relative to TT. By the relativization of Theorem 4.2 of “PA degrees”, YY computes a path of TT, which belongs to I\mathcal I.

REC\mathrm{REC} is not a Scott ideal (Corollary 2.5 of “The base theory RCA₀”), so RCA0\mathsf{RCA}_0 does not prove WKL0\mathsf{WKL}_0. ACA0\mathsf{ACA}_0 proves WKL0\mathsf{WKL}_0, since WKL\mathsf{WKL} is a special case of König’s lemma (Theorem 1.6). Correspondingly, a jump ideal is a Scott ideal, since X′X' computes a path through every infinite XX-computable binary tree, the leftmost one. The converse fails.

Theorem 2.4.#

There is a Scott ideal all of whose members are low. Consequently WKL0\mathsf{WKL}_0 does not prove ACA0\mathsf{ACA}_0.

Proof of Theorem 2.4.

Define sets Z0=∅Z_0 = \emptyset and Zn+1=Zn⊕PnZ_{n+1} = Z_n \oplus P_n, where PnP_n is PA relative to ZnZ_n and satisfies (Zn⊕Pn)′≤TZn′(Z_n \oplus P_n)' \leq_T Z_n'. Such a PnP_n exists by the low basis theorem (Low basis theorem) relativized to ZnZ_n, applied to the non-empty Π10(Zn)\Pi^0_1(Z_n) class of {0,1}\{0,1\}-valued functions DNC relative to ZnZ_n. By induction, Zn′≤T∅′Z_n' \leq_T \emptyset' for every nn.

Let I={X:∃n X≤TZn}\mathcal I = \{ X : \exists n\ X \leq_T Z_n \}. It is a Turing ideal, since the sequence ZnZ_n is increasing. It is a Scott ideal: if X≤TZnX \leq_T Z_n, then Pn≤TZn+1P_n \leq_T Z_{n+1} is in I\mathcal I and PA relative to ZnZ_n, hence relative to XX, and Proposition 2.3 (2) applies. Every X∈IX \in \mathcal I satisfies X′≤TZn′≤T∅′X' \leq_T Z_n' \leq_T \emptyset', so XX is low. In particular ∅′∉I\emptyset' \notin \mathcal I, and I\mathcal I is not a jump ideal: it is an ω-model of WKL0\mathsf{WKL}_0 in which ACA0\mathsf{ACA}_0 fails.

The proof is typical of separations in reverse mathematics: one closes a model under the solutions of a problem while preserving a computability-theoretic weakness, here lowness, that excludes the stronger principle.

2.2 Σ⁰₁ separation

The Σ10\Sigma^0_1 separation scheme consists of the universal closures of

¬∃n (φ(n)∧ψ(n))→∃X ∀n ((φ(n)→n∈X)∧(ψ(n)→n∉X)),\neg \exists n\, (\varphi(n) \wedge \psi(n)) \rightarrow \exists X\, \forall n\, \big( (\varphi(n) \rightarrow n \in X) \wedge (\psi(n) \rightarrow n \notin X) \big),

for Σ10\Sigma^0_1 formulas φ,ψ\varphi, \psi. Its computability-theoretic form is the existence of separating sets for disjoint c.e. sets, one of the characterizations of PA degrees.

Theorem 2.5.#

Over RCA0\mathsf{RCA}_0, WKL\mathsf{WKL} is equivalent to Σ10\Sigma^0_1 separation.

Proof of Theorem 2.5.

(⇒\Rightarrow) Write φ(n)≡∃k θ(n,k)\varphi(n) \equiv \exists k\, \theta(n, k) and ψ(n)≡∃k η(n,k)\psi(n) \equiv \exists k\, \eta(n, k), with θ,η\theta, \eta bounded. Let TT be the set of strings σ\sigma such that for all n,k<∣σ∣n, k < |\sigma|, θ(n,k)→σ(n)=1\theta(n, k) \rightarrow \sigma(n) = 1 and η(n,k)→σ(n)=0\eta(n, k) \rightarrow \sigma(n) = 0. It exists by Δ10\Delta^0_1 comprehension and is a tree. It is infinite: for each mm, the string σ\sigma of length mm with σ(n)=1\sigma(n) = 1 if and only if θ(n,k)\theta(n, k) for some k<mk < m belongs to TT, since no nn has witnesses for both formulas. A path XX separates: if φ(n)\varphi(n) with witness kk, then X(n)=1X(n) = 1, by considering a prefix of length larger than nn and kk; similarly for ψ\psi.

(⇐\Leftarrow) Let TT be an infinite binary tree. For a string σ\sigma, say that σ\sigma is extendible to length nn if some τ⪰σ\tau \succeq \sigma of length nn belongs to TT, a bounded property. Let

φ(σ)≡∃n (σ0 is not extendible to length n∧σ1 is extendible to length n),\varphi(\sigma) \equiv \exists n\, (\sigma 0 \text{ is not extendible to length } n \wedge \sigma 1 \text{ is extendible to length } n),

and ψ(σ)\psi(\sigma) the same with 00 and 11 exchanged, the lengths nn being larger than ∣σ∣|\sigma|. They are disjoint: if φ(σ)\varphi(\sigma) with nn and ψ(σ)\psi(\sigma) with n′n', say n≤n′n \leq n', then σ0\sigma 0 is extendible to length n′n', hence to length nn, a contradiction; the case n′≤nn' \leq n is symmetric. Let SS separate them, with φ(σ)→σ∈S\varphi(\sigma) \rightarrow \sigma \in S and ψ(σ)→σ∉S\psi(\sigma) \rightarrow \sigma \notin S. Define PP by P ⁣↾(k+1)=(P ⁣↾k) iP \uh (k+1) = (P \uh k)\, i with i=1i = 1 if P ⁣↾k∈SP \uh k \in S and i=0i = 0 otherwise; it exists by Δ10\Delta^0_1 comprehension.

Let E(σ)E(\sigma) be the Π10\Pi^0_1 property ”σ\sigma is extendible to every length”. We show E(P ⁣↾k)E(P \uh k) for every kk by Π10\Pi^0_1 induction (Proposition 2.2 of “The base theory RCA₀”). E(ε)E(\varepsilon) holds since TT is infinite. Suppose E(σ)E(\sigma) with σ=P ⁣↾k\sigma = P \uh k. If neither σ0\sigma 0 nor σ1\sigma 1 had EE, there would be lengths beyond which neither is extendible, and σ\sigma would not be extendible beyond their maximum. If only σ1\sigma 1 has EE, then φ(σ)\varphi(\sigma) holds, so σ∈S\sigma \in S and PP chooses σ1\sigma 1; if only σ0\sigma 0 has EE, then ψ(σ)\psi(\sigma), σ∉S\sigma \notin S, and PP chooses σ0\sigma 0. If both have EE, either choice works. So E(P ⁣↾(k+1))E(P \uh (k+1)). In particular every P ⁣↾kP \uh k belongs to TT, and PP is a path.

2.3 The Heine–Borel theorem

The Heine–Borel theorem for [0,1][0, 1] states: if a sequence of open intervals (ai,bi)(a_i, b_i) with rational endpoints covers [0,1][0, 1], that is, every real x∈[0,1]x \in [0, 1] lies in some (ai,bi)(a_i, b_i), then finitely many of them cover [0,1][0, 1].

Theorem 2.6 (Friedman).#

Over RCA0\mathsf{RCA}_0, WKL\mathsf{WKL} is equivalent to the Heine–Borel theorem for [0,1][0, 1].

Proof of Theorem 2.6.

For a binary string σ\sigma, let IσI_\sigma be the closed dyadic interval [0.σ,0.σ+2−∣σ∣][0.\sigma, 0.\sigma + 2^{-|\sigma|}].

(⇒\Rightarrow) Let (ai,bi)(a_i, b_i) cover [0,1][0, 1], and suppose that no finite subfamily does. Let TT be the set of strings σ\sigma such that Iσ⊈⋃i<∣σ∣(ai,bi)I_\sigma \not\subseteq \bigcup_{i < |\sigma|} (a_i, b_i), a decidable condition on rational intervals. TT is a tree, since Iτ⊇IσI_\tau \supseteq I_\sigma and fewer intervals are used for a prefix τ\tau of σ\sigma. It is infinite: for each nn, the finite union ⋃i<n(ai,bi)\bigcup_{i < n} (a_i, b_i) misses some point of [0,1][0, 1], which can be taken rational, as an endpoint of the uncovered set, and this point lies in some IσI_\sigma with ∣σ∣=n|\sigma| = n, so σ∈T\sigma \in T. By WKL\mathsf{WKL}, TT has a path XX. The real x=∑iX(i)2−i−1x = \sum_i X(i) 2^{-i-1} lies in every IX ⁣↾nI_{X \uh n}. It is covered by some (ai,bi)(a_i, b_i), so (x−δ,x+δ)⊆(ai,bi)(x - \delta, x + \delta) \subseteq (a_i, b_i) for some δ>0\delta > 0. For n>in > i with 2−n<δ2^{-n} < \delta, IX ⁣↾n⊆(ai,bi)I_{X \uh n} \subseteq (a_i, b_i), contradicting X ⁣↾n∈TX \uh n \in T.

(⇐\Leftarrow) Let TT be an infinite binary tree, and suppose that it has no path. We use the Cantor middle-thirds set: for a string σ\sigma, let CσC_\sigma be the closed interval of length 3−∣σ∣3^{-|\sigma|} with left endpoint ∑i<∣σ∣2σ(i)3−i−1\sum_{i < |\sigma|} 2\sigma(i) 3^{-i-1}. Consider the following intervals:

  • for each string σ∉T\sigma \notin T, the open interval UσU_\sigma obtained by enlarging CσC_\sigma by 3−∣σ∣−13^{-|\sigma|-1} on each side;
  • the open middle thirds removed in the construction of the Cantor set, at every level.

The sequence of these intervals exists by Δ10\Delta^0_1 comprehension. It covers [0,1][0, 1]: a real xx outside the removed middle thirds lies in ⋂nCX ⁣↾n\bigcap_n C_{X \uh n} for some XX, which can be found by deciding at each level whether xx is in the left or the right third; since XX is not a path of TT, some X ⁣↾n∉TX \uh n \notin T, and x∈CX ⁣↾n⊆UX ⁣↾nx \in C_{X \uh n} \subseteq U_{X \uh n}. By Heine–Borel, finitely many intervals cover [0,1][0, 1]; let nn exceed the lengths of the strings σ\sigma whose intervals UσU_\sigma are used. For a string τ\tau of length nn, the left endpoint cc of CτC_\tau is a point of the Cantor set, so it lies in none of the middle thirds, and it is covered by some UσU_\sigma in the finite family. The interval UσU_\sigma meets the Cantor set only in CσC_\sigma, because the neighboring intervals of level ∣σ∣|\sigma| are at distance at least 3−∣σ∣3^{-|\sigma|}. So c∈Cσc \in C_\sigma, which means that σ\sigma is a prefix of τ00⋯\tau 0 0 \cdots; since ∣σ∣<n|\sigma| < n, σ⪯τ\sigma \preceq \tau. As σ∉T\sigma \notin T and TT is a tree, τ∉T\tau \notin T. So TT has no string of length nn, contradicting its infiniteness.

The ternary coding avoids a difficulty of binary expansions: in [0,1][0, 1], the dyadic intervals Iσ0I_{\sigma 0} and Iσ1I_{\sigma 1} share an endpoint, and a real may have two expansions, while the Cantor set separates the pieces of each level by gaps. Many verifications in this proof, such as the existence of the uncovered rational point and the use of the case distinction on xx, require care in RCA0\mathsf{RCA}_0; they are carried out in Simpson, Section IV.1.

2.4 Conservation and other equivalents

WKL0\mathsf{WKL}_0 proves no new arithmetic sentence: by a theorem of Harrington, every Π11\Pi^1_1 sentence provable in WKL0\mathsf{WKL}_0 is provable in RCA0\mathsf{RCA}_0. The proof uses a forcing argument over non-standard models; it is given in the specialized course on induction and conservation (Harrington's conservation theorem). So compactness arguments, though non-constructive, do not increase the arithmetic consequences of the base theory; WKL0\mathsf{WKL}_0 is moreover Π20\Pi^0_2-conservative over primitive recursive arithmetic, a partial realization of Hilbert’s program.

The following are also equivalent to WKL0\mathsf{WKL}_0 over RCA0\mathsf{RCA}_0 (Simpson, Chapter IV): every continuous function on [0,1][0, 1] is bounded; every continuous function on [0,1][0, 1] attains its maximum; Gödel’s completeness theorem for countable languages; every countable commutative ring has a prime ideal; Brouwer’s fixed point theorem; the Hahn–Banach theorem for separable Banach spaces.

2.5 Exercises

Exercise 2.7 ★.

Show that there is no least ω-model of WKL0\mathsf{WKL}_0, but that the intersection of all ω-models of WKL0\mathsf{WKL}_0 is REC\mathrm{REC}.

Solution

For every non-computable set CC there is a Scott ideal not containing CC: build the sequence ZnZ_n of Theorem 2.4 with PnP_n PA relative to ZnZ_n and C̸≤TZn⊕PnC \not\leq_T Z_n \oplus P_n, using the cone avoidance basis theorem (Cone avoidance basis theorem) relativized to ZnZ_n; by induction C̸≤TZnC \not\leq_T Z_n for every nn. So every non-computable set is missing from some ω-model of WKL0\mathsf{WKL}_0, and the intersection of all of them is REC\mathrm{REC}, which is not a model. A least ω-model would be contained in this intersection, hence equal to REC\mathrm{REC}, a contradiction.

Exercise 2.8 ★★.

Show that WKL\mathsf{WKL} is equivalent over RCA0\mathsf{RCA}_0 to the statement: for every pair of injective functions f,gf, g with disjoint ranges, there is a set XX containing the range of ff and disjoint from the range of gg.

Solution

The statement follows from Σ10\Sigma^0_1 separation with φ(n)≡∃i f(i)=n\varphi(n) \equiv \exists i\, f(i) = n and ψ(n)≡∃i g(i)=n\psi(n) \equiv \exists i\, g(i) = n. Conversely, let ∃k θ(n,k)\exists k\, \theta(n, k) and ∃k η(n,k)\exists k\, \eta(n, k) be disjoint Σ10\Sigma^0_1 formulas. Let f(⟨n,k⟩)=2n+1f(\langle n, k \rangle) = 2n + 1 if kk is the least witness of θ(n,k)\theta(n, k), and f(⟨n,k⟩)=4⟨n,k⟩f(\langle n, k \rangle) = 4 \langle n, k \rangle otherwise; let g(⟨n,k⟩)=2n+1g(\langle n, k \rangle) = 2n + 1 if kk is the least witness of η(n,k)\eta(n, k), and g(⟨n,k⟩)=4⟨n,k⟩+2g(\langle n, k \rangle) = 4 \langle n, k \rangle + 2 otherwise. Both are injective, as in Theorem 1.2. Their ranges are disjoint: odd values 2n+12n + 1 come from ff only if θ\theta has a witness for nn, and from gg only if η\eta has one, which never happen together; the even values of ff are multiples of 44 and those of gg are not. If XX contains the range of ff and avoids the range of gg, the set {n:2n+1∈X}\{ n : 2n + 1 \in X \} separates the two formulas.

2.6 Notes and references

Friedman (1976) identified WKL0\mathsf{WKL}_0 and proved the equivalence with the Heine–Borel theorem. The characterization of its ω-models by Scott ideals goes back to Scott (1962). Harrington’s conservation theorem is unpublished; Simpson’s monograph presents it. The Π20\Pi^0_2 conservation over primitive recursive arithmetic is due to Friedman; Sieg gave a proof-theoretic proof.

  • S. G. Simpson, Subsystems of Second Order Arithmetic, 2nd edition, Cambridge University Press, 2009, Chapter IV.
  • H. Friedman, Systems of second order arithmetic with restricted induction, abstracts, Journal of Symbolic Logic 41 (1976), 557–559.
  • W. Sieg, Fragments of arithmetic, Annals of Pure and Applied Logic 28 (1985), 33–71.