Cite

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

/courses/ramsey-theorem/jockusch
@misc{ramsey-theorem-jockusch,
  title = {{Ramsey's theorem in the arithmetic hierarchy}},
  howpublished = {Lesson 1 of \emph{Ramsey's theorem for pairs}, Computability Theory course notes},
  url = {/courses/ramsey-theorem/jockusch},
  note = {Unreleased working version}
}

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

Lesson 1 of 3

Ramsey's theorem in the arithmetic hierarchy

Ramsey's theorem for n-tuples; pre-homogeneous sets and the Erdős–Radó tree; every coloring of n-tuples has a homogeneous set computable in any degree PA over its (n−1)-st jump; a computable coloring of pairs without Δ⁰₂ homogeneous set; Ramsey's theorem for triples is equivalent to ACA₀.

45 min4 questions
Before this lesson:Weak König's lemma (The Big Five)
Ramsey's theorem for pairs · lesson 1 of 3
  1. Ramsey's theorem in the arithmetic hierarchy
  2. Cohesiveness and stable colorings
  3. Seetapun's theorem

Ramsey’s theorem states that every finite coloring of the nn-element subsets of N\nat has an infinite set on which the coloring is constant. Jockusch (1972) determined how complex such sets must be for computable colorings: homogeneous sets can be found computably in any degree PA over 0(n−1)\obf^{(n-1)}, but not, in general and for n≥2n \geq 2, at the level Σn0\Sigma^0_n. For n≥3n \geq 3, this makes Ramsey’s theorem equivalent to ACA0\mathsf{ACA}_0. For pairs, the analysis leaves a gap, which the rest of the course explores.

1.1 Statements

For a set XX and n≥1n \geq 1, [X]n[X]^n is the set of nn-element subsets of XX, identified with increasing tuples x1<⋯<xnx_1 < \cdots < x_n. A kk-coloring of [X]n[X]^n is a function f:[X]n→k={0,…,k−1}f : [X]^n \to k = \{ 0, \dots, k-1 \}. A set H⊆XH \subseteq X is homogeneous for ff if ff is constant on [H]n[H]^n.

Definition 1.1.#

RTkn\mathsf{RT}^n_k is the statement: every coloring f:[N]n→kf : [\nat]^n \to k has an infinite homogeneous set.

In reverse mathematics, colorings and homogeneous sets are sets, and RTkn\mathsf{RT}^n_k is a Π21\Pi^1_2 statement. In computability theory, one asks how complex homogeneous sets must be relative to ff. The case n=1n = 1 is the infinite pigeonhole principle.

Proposition 1.2.#
  1. RCA0\mathsf{RCA}_0 proves RTk1\mathsf{RT}^1_k for every standard kk.
  2. Over RCA0\mathsf{RCA}_0, RT2n\mathsf{RT}^n_2 implies its version for colorings of [X]n[X]^n, for every infinite set XX.
  3. For every standard n≥1n \geq 1 and k≥2k \geq 2, RCA0\mathsf{RCA}_0 proves RT2n→RTkn\mathsf{RT}^n_2 \rightarrow \mathsf{RT}^n_k.
Proof of Proposition 1.2.
  1. For f:N→kf : \nat \to k, the sets Ci={x:f(x)=i}C_i = \{ x : f(x) = i \} exist. If all were finite, bounded by bib_i, then max⁡ibi\max_i b_i would bound N\nat. So some CiC_i is infinite, and it is homogeneous.
  2. Let pp enumerate XX in increasing order; it exists in RCA0\mathsf{RCA}_0 when XX is infinite. Apply RT2n\mathsf{RT}^n_2 to g({x1,…,xn})=f({p(x1),…,p(xn)})g(\{ x_1, \dots, x_n \}) = f(\{ p(x_1), \dots, p(x_n) \}), and take the image of a homogeneous set under pp.
  3. By induction on kk in the metatheory. Given f:[N]n→k+1f : [\nat]^n \to k + 1, merge the colors k−1k - 1 and kk, and apply RTkn\mathsf{RT}^n_k to obtain an infinite HH. If its color is less than k−1k - 1, HH is homogeneous for ff; otherwise ff takes only the values k−1k - 1 and kk on [H]n[H]^n, and (2) gives a homogeneous subset.

The number of colors is irrelevant for provability, as long as it is a fixed standard number; we focus on two colors.

1.2 Pre-homogeneous sets and the upper bound

A coloring of (n+1)(n+1)-tuples reduces to a coloring of nn-tuples on a well-chosen set.

Definition 1.3.#

A set XX is pre-homogeneous for f:[N]n+1→kf : [\nat]^{n+1} \to k if, for all x1<⋯<xnx_1 < \cdots < x_n in XX, the color f({x1,…,xn,y})f(\{ x_1, \dots, x_n, y \}) is the same for all y∈Xy \in X with y>xny > x_n.

On an infinite pre-homogeneous set XX, the coloring g({x1,…,xn})=f({x1,…,xn,y})g(\{ x_1, \dots, x_n \}) = f(\{ x_1, \dots, x_n, y \}), for any y∈Xy \in X above xnx_n, is well defined and computable in f⊕Xf \oplus X, and every set H⊆XH \subseteq X homogeneous for gg is homogeneous for ff.

Proposition 1.4 (Erdős, Radó).#

For every coloring f:[N]n+1→kf : [\nat]^{n+1} \to k, there is an infinite, finitely branching, ff-computable tree T⊆N<NT \subseteq \nat^{<\nat} of strictly increasing sequences whose infinite paths are pre-homogeneous for ff. Every set PA relative to f′f' computes an infinite pre-homogeneous set for ff.

Proof of Proposition 1.4.

We build TT by inserting the numbers 0,1,2,…0, 1, 2, \dots one at a time. The first nn numbers form a single branch ⟨0,1,…,n−1⟩\langle 0, 1, \dots, n - 1 \rangle. To insert s≥ns \geq n, start at this branch and descend: at a node σ\sigma, call pattern of a number aa over σ\sigma the tuple of colors f({σ(m1),…,σ(mn),a})f(\{ \sigma(m_1), \dots, \sigma(m_n), a \}) for m1<⋯<mn<∣σ∣m_1 < \cdots < m_n < |\sigma|. If σ\sigma has an immediate successor σ a\sigma\, a in the current tree such that aa and ss have the same pattern over σ\sigma, move to σ a\sigma\, a; otherwise add σ s\sigma\, s to the tree and stop.

TT is ff-computable and infinite. To decide whether a sequence σ s\sigma\, s belongs to TT, simulate the insertion of 0,…,s0, \dots, s. Every number is inserted, so TT is infinite.

TT is finitely branching. Two immediate successors σa\sigma a and σb\sigma b of a node, with a<ba < b, have different patterns over σ\sigma: otherwise bb would have moved to σa\sigma a instead of creating σb\sigma b. So σ\sigma has at most k(∣σ∣n)k^{\binom{|\sigma|}{n}} successors.

Paths are pre-homogeneous. Let PP be an infinite path, j<j′j < j', a=P(j)a = P(j) and b=P(j′)b = P(j'). When bb was inserted, it descended along P ⁣↾j′P \uh j', hence moved from P ⁣↾jP \uh j to P ⁣↾(j+1)=(P ⁣↾j) aP \uh (j + 1) = (P \uh j)\, a, so aa and bb have the same pattern over P ⁣↾jP \uh j. For x1<⋯<xnx_1 < \cdots < x_n in the range of PP, at positions m1<⋯<mnm_1 < \cdots < m_n, and y=P(j)y = P(j), y′=P(j′)y' = P(j') with mn<j<j′m_n < j < j', this gives f({x1,…,xn,y})=f({x1,…,xn,y′})f(\{ x_1, \dots, x_n, y \}) = f(\{ x_1, \dots, x_n, y' \}).

Computing a path. The number of successors of each node is bounded, but their values are not: whether σ\sigma has a successor above a given number is a Σ10(f)\Sigma^0_1(f) question. With f′f', one computes, for each node, the finite list of its successors. Coding each successor by its rank in this list, the paths of TT correspond, computably in f′f', to the paths of an f′f'-computable tree in ∏σk(∣σ∣n)\prod_\sigma k^{\binom{|\sigma|}{n}}, a Π10(f′)\Pi^0_1(f') class after coding into Cantor space. It is non-empty by König’s lemma. A set PA relative to f′f' computes a member of it (Theorem 4.2 of “PA degrees”, relativized), and it computes f′f' itself (second exercise of PA degrees, relativized to f′f'), so it computes a path of TT.

Theorem 1.5 (Jockusch).#

For every n≥1n \geq 1, every coloring f:[N]n→kf : [\nat]^n \to k has an infinite homogeneous set computable in every set that is PA relative to f(n−1)f^{(n-1)}. In particular, every computable coloring of [N]n[\nat]^n has an infinite homogeneous set computable in ∅(n)\emptyset^{(n)}.

Proof of Theorem 1.5.

By induction on nn, for all colorings ff simultaneously. For n=1n = 1, ff itself computes a homogeneous set, and a set PA relative to ff computes ff.

For n=2n = 2, let PP be PA relative to f′f'. By Proposition 1.4, PP computes an infinite pre-homogeneous set XX, the induced coloring gg of [X]1[X]^1 is computable in PP, and so is a homogeneous set for gg, which is homogeneous for ff.

For n+1n + 1 with n≥2n \geq 2, let P1P_1 be PA relative to f(n)f^{(n)}. By the low basis theorem relativized to f′f', there is P0P_0 PA relative to f′f' with P0′≤Tf′′P_0' \leq_T f''; then P0(n−1)≤Tf(n)P_0^{(n-1)} \leq_T f^{(n)}, since n−1≥1n - 1 \geq 1. P0P_0 computes an infinite pre-homogeneous set XX and the induced coloring gg of [X]n[X]^n, so g(n−1)≤TP0(n−1)≤Tf(n)g^{(n-1)} \leq_T P_0^{(n-1)} \leq_T f^{(n)}, and P1P_1 is PA relative to g(n−1)g^{(n-1)}. By the induction hypothesis, applied to gg transported to [N]n[\nat]^n through the enumeration of XX, P1P_1 computes an infinite homogeneous set for the transported coloring; its image under the enumeration of XX is computable in P1P_1, since X≤TP0≤Tf′′≤TP1X \leq_T P_0 \leq_T f'' \leq_T P_1, and it is homogeneous for gg, hence for ff.

For the last statement, ∅(n)\emptyset^{(n)} is PA relative to ∅(n−1)\emptyset^{(n-1)}: it computes a {0,1}\{0, 1\}-valued function DNC relative to ∅(n−1)\emptyset^{(n-1)}.

Jockusch proved more precisely that every computable coloring of [N]n[\nat]^n has a Πn0\Pi^0_n infinite homogeneous set. For pairs, one builds an infinite pre-homogeneous sequence and the set of its elements of a given limit color so that both are Π20\Pi^0_2; we do not reproduce the construction (Jockusch 1972; Hirschfeldt, Slicing the Truth, Chapter 6).

1.3 The lower bound

Theorem 1.6 (Jockusch).#

There is a computable coloring f:[N]2→2f : [\nat]^2 \to 2 with no infinite Δ20\Delta^0_2 homogeneous set, hence no infinite Σ20\Sigma^0_2 homogeneous set.

Proof of Theorem 1.6.

For each ee and ss, let Ae,s={x<s:Φe∅s′(x)[s] ⁣ ⁣↓=1}A_{e,s} = \{ x < s : \Phi^{\emptyset'_s}_e(x)[s]\converge = 1 \}, where ∅s′\emptyset'_s is the stage ss approximation of ∅′\emptyset'. If Φe∅′\Phi^{\emptyset'}_e is the characteristic function of a set AA, then for every xx, x∈Ae,sx \in A_{e,s} for all large ss if x∈Ax \in A, and x∉Ae,sx \notin A_{e,s} for all large ss otherwise. Every Δ20\Delta^0_2 set is of this form, by the limit lemma.

Define f({x,s})f(\{ x, s \}) for x<sx < s at stage ss: for e=0,1,…,s−1e = 0, 1, \dots, s - 1 in turn, if Ae,sA_{e,s} has at least 2e+22e + 2 elements, choose the two least elements a<ba < b of Ae,sA_{e,s} for which f({a,s})f(\{ a, s \}) and f({b,s})f(\{ b, s \}) are not yet defined, and set f({a,s})=0f(\{ a, s \}) = 0 and f({b,s})=1f(\{ b, s \}) = 1. Before ee acts, at most 2e2e values f({x,s})f(\{ x, s \}) have been defined, so the two elements exist. All remaining values f({x,s})f(\{ x, s \}) are set to 00. The coloring is computable.

Let A=Φe∅′A = \Phi^{\emptyset'}_e be an infinite Δ20\Delta^0_2 set, and nn such that A∩{0,…,n−1}A \cap \{0, \dots, n-1\} has exactly 2e+22e + 2 elements. For large enough s∈As \in A, Ae,s∩{0,…,n−1}=A∩{0,…,n−1}A_{e,s} \cap \{0, \dots, n-1\} = A \cap \{0, \dots, n-1\}. At stage ss, the 2e+22e + 2 least elements of Ae,sA_{e,s} are those of AA below nn, and ee chooses two of them, aa and bb, with f({a,s})≠f({b,s})f(\{ a, s \}) \neq f(\{ b, s \}). Since a,b,s∈Aa, b, s \in A, AA is not homogeneous.

Finally, every infinite Σ20\Sigma^0_2 set B={x:∃y ∀z R(x,y,z)}B = \{ x : \exists y\, \forall z\, R(x, y, z) \}, with RR computable, has an infinite Δ20\Delta^0_2 subset: with ∅′\emptyset', search for pairs (x,y)(x, y) with ∀z R(x,y,z)\forall z\, R(x, y, z), and keep an increasing sequence of the first coordinates found. A subset of a homogeneous set is homogeneous.

The two bounds leave a gap at pairs: homogeneous sets for computable colorings can always be found below ∅′′\emptyset'', and sometimes not below ∅′\emptyset'. What computational power is necessarily carried by every homogeneous set is the subject of the next two lessons.

1.4 Coding the jump into triples

The upper bound for triples is not only a bound on complexity: some computable colorings of triples force all their homogeneous sets to compute ∅′\emptyset'.

Lemma 1.7.#

For every ∅′\emptyset'-computable coloring h:[N]n→kh : [\nat]^n \to k, there is a computable coloring f:[N]n+1→kf : [\nat]^{n+1} \to k such that every infinite homogeneous set for ff is homogeneous for hh.

Proof of Lemma 1.7.

By the limit lemma, h({x1,…,xn})=lim⁡tf({x1,…,xn,t})h(\{ x_1, \dots, x_n \}) = \lim_t f(\{ x_1, \dots, x_n, t \}) for a computable function ff defined for x1<⋯<xn<tx_1 < \cdots < x_n < t; replacing the values outside kk by 00, which does not change the limit, ff is a coloring [N]n+1→k[\nat]^{n+1} \to k. Let HH be infinite and homogeneous for ff with color ii. For x1<⋯<xnx_1 < \cdots < x_n in HH, take t∈Ht \in H large enough that f({x1,…,xn,t})=h({x1,…,xn})f(\{ x_1, \dots, x_n, t \}) = h(\{ x_1, \dots, x_n \}); this value is ii.

Proposition 1.8 (Jockusch).#

There is a computable coloring of [N]3[\nat]^3 with two colors all of whose infinite homogeneous sets compute ∅′\emptyset'.

Proof of Proposition 1.8.

Let m(x)m(x) be the least ss with ∅s′∩{0,…,x−1}=∅′∩{0,…,x−1}\emptyset'_s \cap \{0, \dots, x - 1\} = \emptyset' \cap \{0, \dots, x - 1\}, a ∅′\emptyset'-computable function. Let h({x,y})=1h(\{ x, y \}) = 1 if m(x)<ym(x) < y and 00 otherwise, for x<yx < y. An infinite homogeneous set HH for hh has color 11: color 00 would give m(x)≥ym(x) \geq y for all y>xy > x in HH. So for consecutive elements x<yx < y of HH, ∅′∩{0,…,x−1}=∅y′∩{0,…,x−1}\emptyset' \cap \{0, \dots, x-1\} = \emptyset'_y \cap \{0, \dots, x-1\}, and HH computes ∅′\emptyset'. Apply Lemma 1.7.

1.5 Reverse mathematics

Theorem 1.9 (Jockusch, Simpson).#

For every standard n≥3n \geq 3, RT2n\mathsf{RT}^n_2 is equivalent to ACA0\mathsf{ACA}_0 over RCA0\mathsf{RCA}_0. Moreover ACA0\mathsf{ACA}_0 proves RTkn\mathsf{RT}^n_k for all standard nn and kk.

Proof of Theorem 1.9.

ACA0\mathsf{ACA}_0 proves RTkn\mathsf{RT}^n_k, by induction on nn in the metatheory. The case n=1n = 1 is Proposition 1.2. Given f:[N]n+1→kf : [\nat]^{n+1} \to k, the Erdős–Radó tree exists by Δ10\Delta^0_1 comprehension, and RCA0\mathsf{RCA}_0 proves that it is infinite and finitely branching. By König’s lemma (Theorem 1.6 of “Arithmetic comprehension”), it has a path, which is pre-homogeneous. The induction hypothesis, with Proposition 1.2 (2), whose proof applies to any number of colors, gives a homogeneous set for the induced coloring.

RT23\mathsf{RT}^3_2 implies ACA0\mathsf{ACA}_0. We prove that X′X' exists, using Theorem 1.2 of “Arithmetic comprehension”. Define f({x,s,t})=1f(\{ x, s, t \}) = 1 for x<s<tx < s < t if every e<xe < x satisfies ΦeX(e)[s] ⁣ ⁣↓↔ΦeX(e)[t] ⁣ ⁣↓\Phi^X_e(e)[s]\converge \leftrightarrow \Phi^X_e(e)[t]\converge, and 00 otherwise. Let HH be infinite and homogeneous. Its color is 11: for x∈Hx \in H, among 2x+12^x + 1 elements of HH above xx, two have the same pattern of convergence for the indices e<xe < x, by the finite pigeonhole principle, provable in RCA0\mathsf{RCA}_0. For each ee, let xex_e be the least element of HH above ee and ses_e the least element of HH above xex_e. If ΦeX(e)[t0] ⁣ ⁣↓\Phi^X_e(e)[t_0]\converge for some t0t_0, take t∈Ht \in H above ses_e and t0t_0, so that ΦeX(e)[t] ⁣ ⁣↓\Phi^X_e(e)[t]\converge; then f({xe,se,t})=1f(\{ x_e, s_e, t \}) = 1 gives ΦeX(e)[se] ⁣ ⁣↓\Phi^X_e(e)[s_e]\converge. So X′={e:ΦeX(e)[se] ⁣ ⁣↓}X' = \{ e : \Phi^X_e(e)[s_e]\converge \}, which exists by Δ10\Delta^0_1 comprehension relative to X⊕HX \oplus H.

For n≥3n \geq 3, RT2n\mathsf{RT}^n_2 implies RT23\mathsf{RT}^3_2: a coloring ff of triples induces the coloring {x1,…,xn}↦f({x1,x2,x3})\{ x_1, \dots, x_n \} \mapsto f(\{ x_1, x_2, x_3 \}) of nn-tuples, with the same infinite homogeneous sets.

Corollary 1.10.#

WKL0\mathsf{WKL}_0 does not prove RT22\mathsf{RT}^2_2.

Proof of Corollary 1.10.

By Theorem 2.4 of “Weak König's lemma”, there is an ω-model of WKL0\mathsf{WKL}_0 consisting of low sets, hence of Δ20\Delta^0_2 sets. The computable coloring of Theorem 1.6 belongs to it and has no homogeneous set in it.

The coloring of Proposition 1.8 has no analogue for pairs: by Seetapun’s theorem, proved in the last lesson, every computable coloring of pairs has an infinite homogeneous set that does not compute ∅′\emptyset'.

1.6 Exercises

Exercise 1.11 ★.

Show that for every n≥2n \geq 2 there is a computable coloring f:[N]n→2f : [\nat]^n \to 2 with no infinite Σn0\Sigma^0_n homogeneous set.

Solution

By induction on nn, relativized to an arbitrary oracle ZZ: there is a ZZ-computable coloring of [N]n[\nat]^n with no infinite Σn0(Z)\Sigma^0_n(Z) homogeneous set. For n=2n = 2, this is Theorem 1.6 relativized. Given the statement for nn and every oracle, apply it to Z′Z': there is a Z′Z'-computable coloring hh of [N]n[\nat]^n with no infinite Σn0(Z′)\Sigma^0_n(Z'), that is Σn+10(Z)\Sigma^0_{n+1}(Z), homogeneous set. By Lemma 1.7 relativized to ZZ, there is a ZZ-computable coloring of [N]n+1[\nat]^{n+1} whose infinite homogeneous sets are homogeneous for hh.

Exercise 1.12 ★.

Show that every computable coloring f:[N]2→2f : [\nat]^2 \to 2 has an infinite homogeneous set HH with H′≤T∅′′H' \leq_T \emptyset''.

Solution

By the low basis theorem relativized to ∅′\emptyset', there is a set PP PA relative to ∅′\emptyset' with P′≤T∅′′P' \leq_T \emptyset''. By Theorem 1.5, PP computes an infinite homogeneous set HH, and H′≤TP′≤T∅′′H' \leq_T P' \leq_T \emptyset''.

1.7 Notes and references

Ramsey (1930) proved his theorem as a lemma in a decision procedure for a fragment of first-order logic. Specker (1971) built a computable coloring of pairs without computable infinite homogeneous set, and Jockusch (1972) obtained the bounds of this lesson. Simpson formalized the equivalence with ACA0\mathsf{ACA}_0. Hirschfeldt’s book is the standard reference for the reverse mathematics of combinatorial principles.

  • C. G. Jockusch, Jr., Ramsey’s theorem and recursion theory, Journal of Symbolic Logic 37 (1972), 268–280.
  • D. R. Hirschfeldt, Slicing the Truth, World Scientific, 2014.