Cite

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

/courses/randomness-notions/weak-2-randomness
@misc{randomness-notions-weak-2-randomness,
  title = {{Weak 2-randomness}},
  howpublished = {Lesson 2 of \emph{Notions of randomness}, Computability Theory course notes},
  url = {/courses/randomness-notions/weak-2-randomness},
  note = {Unreleased working version}
}

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

Lesson 2 of 3

Weak 2-randomness

Sequences avoiding every null Π⁰₂ class; 2-randomness implies weak 2-randomness, which implies Martin-Löf randomness, both strictly; no universal test; the theorem of Downey, Nies, Weber and Yu that a random sequence is weakly 2-random exactly when it computes no non-computable c.e. set.

30 min2 questions
Notions of randomness · lesson 2 of 3
  1. Relative randomness and van Lambalgen's theorem
  2. Weak 2-randomness
  3. Schnorr and Kurtz randomness

A Martin-Löf test describes a null Π20\Pi^0_2 class ⋂nUn\bigcap_n \mathcal U_n whose measures converge to 00 at a computable rate. Dropping the condition on the rate gives a stronger notion: a sequence is weakly 2-random if it avoids every null Π20\Pi^0_2 class. The notion lies strictly between Martin-Löf randomness and 2-randomness, and it has a characterization in terms of computational power: a random sequence is weakly 2-random exactly when it computes no non-computable c.e. set.

2.1 Definition and position

Definition 2.1.#

A sequence is weakly 2-random if it belongs to no null Π20\Pi^0_2 class, that is, to no class ⋂nUn\bigcap_n \mathcal U_n with (Un)(\mathcal U_n) uniformly Σ10\Sigma^0_1 and λ(⋂nUn)=0\lambda(\bigcap_n \mathcal U_n) = 0.

Replacing Un\mathcal U_n by U0∩⋯∩Un\mathcal U_0 \cap \cdots \cap \mathcal U_n, one may assume the classes decreasing, so that λ(Un)\lambda(\mathcal U_n) tends to 00.

Proposition 2.2.#
  1. Every 2-random sequence is weakly 2-random, and every weakly 2-random sequence is Martin-Löf random.
  2. No Δ20\Delta^0_2 sequence is weakly 2-random. In particular Ω\Omega is random but not weakly 2-random.
  3. There is no universal test for weak 2-randomness: no single null Π20\Pi^0_2 class contains every sequence that is not weakly 2-random.
Proof of Proposition 2.2.
  1. A Martin-Löf test describes a null Π20\Pi^0_2 class. Let ⋂nUn\bigcap_n \mathcal U_n be null, with decreasing Un\mathcal U_n. The relation λ(Un)≤2−k\lambda(\mathcal U_n) \leq 2^{-k} is Π10\Pi^0_1, since λ(Un)\lambda(\mathcal U_n) is left-c.e. uniformly in nn; so ∅′\emptyset' computes, for each kk, the least nkn_k with λ(Unk)≤2−k\lambda(\mathcal U_{n_k}) \leq 2^{-k}, which exists since λ(Un)\lambda(\mathcal U_n) tends to 00. The classes Unk\mathcal U_{n_k} form a Martin-Löf test relative to ∅′\emptyset' containing ⋂nUn\bigcap_n \mathcal U_n.
  2. Let A=lim⁡sAsA = \lim_s A_s with (As)(A_s) computable. The classes U⟨n,t⟩=⋃s>t[As ⁣↾n]\mathcal U_{\langle n, t \rangle} = \bigcup_{s > t} [A_s \uh n] are uniformly Σ10\Sigma^0_1, and their intersection is {A}\{ A \}: a sequence XX in all of them agrees, for every nn, with As ⁣↾nA_s \uh n for infinitely many ss, hence with A ⁣↾nA \uh n. This is a null Π20\Pi^0_2 class containing AA.
  3. Let ⋂nUn\bigcap_n \mathcal U_n be null. Some Un\mathcal U_n has measure less than 11, so its complement is a non-empty Π10\Pi^0_1 class, whose leftmost member is Δ20\Delta^0_2. By (2), it is not weakly 2-random, yet it is outside ⋂nUn\bigcap_n \mathcal U_n.

2.2 The Downey–Nies–Weber–Yu theorem

Theorem 2.3 (Downey, Nies, Weber, Yu).#

For a Martin-Löf random sequence ZZ, the following are equivalent:

  1. ZZ is not weakly 2-random;
  2. ZZ computes a non-computable c.e. set;
  3. ZZ computes a non-computable Δ20\Delta^0_2 set.
Proof of Theorem 2.3.

(2 ⇒ 3) Every c.e. set is Δ20\Delta^0_2.

(3 ⇒ 1) Let A=lim⁡sAsA = \lim_s A_s be non-computable and Δ20\Delta^0_2, with ΦZ=A\Phi^Z = A. The classes U⟨n,t⟩=⋃s>t{X:ΦX ⁣↾n=As ⁣↾n}\mathcal U_{\langle n, t \rangle} = \bigcup_{s > t} \{ X : \Phi^X \uh n = A_s \uh n \} are uniformly Σ10\Sigma^0_1, and, as in Proposition 2.2 (2), their intersection is {X:ΦX=A}\{ X : \Phi^X = A \}, which is null by Sacks’s theorem (Sacks). It contains ZZ.

(1 ⇒ 2) Let Z∈⋂nUnZ \in \bigcap_n \mathcal U_n, a null Π20\Pi^0_2 class with decreasing Un\mathcal U_n, and let Un,t\mathcal U_{n,t} be the clopen approximation of Un\mathcal U_n at stage tt, generated by the strings enumerated by stage tt; it increases with tt, and Un=⋃tUn,t\mathcal U_n = \bigcup_t \mathcal U_{n,t}. We enumerate a set AA, together with classes Ve\mathcal V_e. At stage tt, for each e≤te \leq t such that no element has been enumerated for ee and We,t∩At=∅W_{e,t} \cap A_t = \emptyset, if there is n∈We,tn \in W_{e,t} with n>2en > 2e and λ(Un,t)<2−e\lambda(\mathcal U_{n,t}) < 2^{-e}, enumerate the least such nn into AA, and let Ve=Un,t\mathcal V_e = \mathcal U_{n,t}. If ee never acts, Ve=∅\mathcal V_e = \emptyset.

AA is simple, hence not computable. Each ee enumerates at most one element, larger than 2e2e, so AA has infinite complement. If WeW_e is infinite, it contains n>2en > 2e with λ(Un)<2−e\lambda(\mathcal U_n) < 2^{-e}, since these measures tend to 00, and then λ(Un,t)<2−e\lambda(\mathcal U_{n,t}) < 2^{-e} for every tt. At every stage t≥et \geq e with n∈We,tn \in W_{e,t}, either We,tW_{e,t} already meets AtA_t, or ee has already acted, or ee acts; in each case AA meets WeW_e (when ee acts, it enumerates an element of We,tW_{e,t}).

ZZ computes AA. The classes Ve\mathcal V_e are clopen, uniformly c.e. in ee, with λ(Ve)<2−e\lambda(\mathcal V_e) < 2^{-e}: a Solovay test. Since ZZ is random, ZZ belongs to Ve\mathcal V_e for finitely many ee only (Solovay), and these ee enumerate finitely many elements. Let mm exceed them. For n≥mn \geq m, if nn enters AA at stage tt, then Z∉Un,tZ \notin \mathcal U_{n,t}. To decide whether n≥mn \geq m belongs to AA, search for the least stage tt with Z∈Un,tZ \in \mathcal U_{n,t}, which exists since Z∈UnZ \in \mathcal U_n; then n∈An \in A if and only if n∈Atn \in A_t, since Un,t\mathcal U_{n,t} increases with tt.

The construction makes the enumeration of nn into AA costly for ZZ: it can only happen while ZZ is still outside Un\mathcal U_n, and the Solovay test charges each such enumeration with the measure of the current approximation. The same idea, charging each enumeration with a measure, underlies the cost functions used to build K-trivial sets.

Item (3) says that, for random sequences, weak 2-randomness amounts to forming a minimal pair with ∅′\emptyset': every set computable both in ZZ and in ∅′\emptyset' is computable. This formulation is due to Hirschfeldt and Miller.

2.3 Exercises

Exercise 2.4 ★.

Show that a weakly 2-random sequence is not computable in ∅′\emptyset', and that a 2-random sequence computes no non-computable c.e. set.

Solution

A set computable in ∅′\emptyset' is Δ20\Delta^0_2, hence not weakly 2-random by Proposition 2.2 (2). A 2-random sequence is weakly 2-random and random, so by Theorem 2.3 it computes no non-computable c.e. set.

Exercise 2.5 ★★.

Show that there is a weakly 2-random sequence that is not 2-random.

Solution

By Corollary 2.5 of “Martin-Löf tests” (3), there is a random sequence ZZ that is computably dominated. It computes no non-computable c.e. set AA: the settling function m(x)=min⁡{s:As∩{0,…,x}=A∩{0,…,x}}m(x) = \min \{ s : A_s \cap \{0, \dots, x\} = A \cap \{0, \dots, x\} \} is computable in AA, hence in ZZ, so there are a computable function gg and a number x0x_0 with m(x)≤g(x)m(x) \leq g(x) for x≥x0x \geq x_0. Then x∈A  ⟺  x∈Ag(x)x \in A \iff x \in A_{g(x)} for x≥x0x \geq x_0, which makes AA computable. By Theorem 2.3, ZZ is weakly 2-random. By Kurtz’s theorem, every 2-random sequence has hyperimmune degree, that is, is not computably dominated; so ZZ is not 2-random. Kurtz’s theorem is admitted, as stated in the previous lesson.

2.4 Notes and references

Weak 2-randomness was studied by Kurtz (1981) and by Gaifman and Snir (1982). Downey, Nies, Weber and Yu (2006) proved the characterization of this lesson, and Hirschfeldt and Miller the equivalence with forming a minimal pair with ∅′\emptyset'.

  • H. Gaifman and M. Snir, Probabilities over rich languages, testing and randomness, Journal of Symbolic Logic 47 (1982), 495–548.
  • S. A. Kurtz, Randomness and genericity in the degrees of unsolvability, PhD thesis, University of Illinois, 1981.
  • R. Downey, A. Nies, R. Weber and L. Yu, Lowness and Π⁰₂ nullsets, Journal of Symbolic Logic 71 (2006), 1044–1052.
  • R. G. Downey and D. R. Hirschfeldt, Algorithmic Randomness and Complexity, Springer, 2010, Section 7.2.