Cite

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

/courses/k-trivials/lowness
@misc{k-trivials-lowness,
  title = {{Lowness for randomness}},
  howpublished = {Lesson 3 of \emph{K-trivial sets}, Computability Theory course notes},
  url = {/courses/k-trivials/lowness},
  note = {Unreleased working version}
}

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

Lesson 3 of 3

Lowness for randomness

Sets that are low for K, low for Martin-Löf randomness, and bases for randomness; the three notions coincide by the hungry sets theorem; they also coincide with K-triviality, by a theorem of Nies and Hirschfeldt, admitted here; consequences for the K-trivial sets.

45 min3 questions
Before this lesson:Cost functions and Solovay's construction · Randomness and Turing degrees (Martin-Löf randomness)
K-trivial sets · lesson 3 of 3
  1. K-trivial sets
  2. Cost functions and Solovay's construction
  3. Lowness for randomness

An oracle AA can be used to compress strings and to test sequences for randomness. A set is low for a relativizable notion if, used as an oracle, it does not change the notion. For prefix-free complexity and Martin-Löf randomness, the low sets are exactly the K-trivial sets. This lesson proves the equivalence of three lowness notions, by the hungry sets argument of Hirschfeldt, Nies and Stephan, and states the theorem of Nies and Hirschfeldt that connects them to K-triviality.

3.1 Relativized complexity and randomness

An oracle prefix-free machine is an oracle machine MM such that, for every oracle XX, the domain of MXM^X is prefix-free. The constructions of Prefix-free complexity relativize uniformly: there is an oracle prefix-free machine UU such that, for every oracle prefix-free machine MM, there is cc with KUX≤KMX+cK_{U^X} \leq K_{M^X} + c for every oracle XX. We write KA=KUAK^A = K_{U^A}, and MLRA\mathrm{MLR}^A for the class of sequences passing every Martin-Löf test relative to AA, the tests whose components are uniformly Σ10(A)\Sigma^0_1(A). The results of the course on Martin-Löf randomness relativize, by the same proofs; in particular Z∈MLRAZ \in \mathrm{MLR}^A if and only if KA(Z ⁣↾n)≥n−O(1)K^A(Z \uh n) \geq n - O(1), and ∑τ∈dom⁡UA2−∣τ∣≤1\sum_{\tau \in \dom U^A} 2^{-|\tau|} \leq 1.

If B≤TAB \leq_T A, then KA≤+KBK^A \leq^+ K^B, since a machine with oracle AA can simulate UBU^B; in particular KA≤+KK^A \leq^+ K. So oracles can only make complexity smaller and randomness harder.

For a string ρ\rho, we write Uρ(τ) ⁣ ⁣↓=xU^\rho(\tau)\converge = x when the computation of UU on input τ\tau with oracle ρ\rho halts with output xx and its use is exactly ∣ρ∣|\rho|, in the sense of Definition 1.6 of “Computation relative to an oracle”. The set of triples (ρ,τ,x)(\rho, \tau, x) with Uρ(τ) ⁣ ⁣↓=xU^\rho(\tau)\converge = x is c.e., and for every oracle XX and every τ∈dom⁡UX\tau \in \dom U^X, there is exactly one ρ≺X\rho \prec X with Uρ(τ) ⁣ ⁣↓U^\rho(\tau)\converge, namely X ⁣↾uX \uh u for the use uu of the computation.

3.2 Three lowness notions

Definition 3.1.#

Let AA be a set.

  1. AA is low for K if K(σ)≤+KA(σ)K(\sigma) \leq^+ K^A(\sigma) for every string σ\sigma.
  2. AA is low for Martin-Löf randomness if MLRA=MLR\mathrm{MLR}^A = \mathrm{MLR}.
  3. AA is a base for randomness if A≤TZA \leq_T Z for some Z∈MLRAZ \in \mathrm{MLR}^A.

Every computable set has the three properties. For a non-computable set AA, the sequences computing AA form a null class (Sacks); a base for randomness is a set computed by some sequence that is nevertheless random relative to AA.

Proposition 3.2.#
  1. If AA is low for K, then AA is K-trivial.
  2. If AA is low for K, then AA is low for Martin-Löf randomness.
  3. If AA is low for Martin-Löf randomness, then AA is a base for randomness.
  4. The sets that are low for K are closed downward under ≤T\leq_T.
Proof of Proposition 3.2.
  1. With oracle AA, the string A ⁣↾nA \uh n is computed from nn, so KA(A ⁣↾n)≤+KA(n)≤+K(n)K^A(A \uh n) \leq^+ K^A(n) \leq^+ K(n). If AA is low for K, then K(A ⁣↾n)≤+KA(A ⁣↾n)≤+K(n)K(A \uh n) \leq^+ K^A(A \uh n) \leq^+ K(n).
  2. MLRA⊆MLR\mathrm{MLR}^A \subseteq \mathrm{MLR} for every AA, since a test without oracle is a test relative to AA. Conversely, let Z∈MLRZ \in \mathrm{MLR}, so that K(Z ⁣↾n)≥n−cK(Z \uh n) \geq n - c. If AA is low for K, then KA(Z ⁣↾n)≥K(Z ⁣↾n)−d≥n−c−dK^A(Z \uh n) \geq K(Z \uh n) - d \geq n - c - d, and Z∈MLRAZ \in \mathrm{MLR}^A by the relativized Levin–Schnorr theorem.
  3. By the Kučera–Gács theorem (Kučera, Gács), some Z∈MLRZ \in \mathrm{MLR} computes AA, and MLR=MLRA\mathrm{MLR} = \mathrm{MLR}^A.
  4. If B≤TAB \leq_T A and AA is low for K, then K≤+KA≤+KBK \leq^+ K^A \leq^+ K^B.

3.3 The hungry sets theorem

Theorem 3.3 (Hirschfeldt, Nies, Stephan).#

Every base for randomness is low for K.

Proof of Theorem 3.3.

Let Z∈MLRAZ \in \mathrm{MLR}^A and Φ\Phi a Turing functional with ΦZ=A\Phi^Z = A. For a string η\eta and a string ρ\rho, write Φη[s]⪰ρ\Phi^\eta[s] \succeq \rho if Φη(i)[s] ⁣ ⁣↓=ρ(i)\Phi^\eta(i)[s]\converge = \rho(i) for every i<∣ρ∣i < |\rho|.

The construction. Fix a parameter d∈Nd \in \nat. For each triple (ρ,τ,x)(\rho, \tau, x) with Uρ(τ) ⁣ ⁣↓=xU^\rho(\tau)\converge = x we enumerate a finite or infinite set of strings C=Cτ,xρC = C^\rho_{\tau, x}, the hungry set of the triple, which aims at a measure 2−∣τ∣−d2^{-|\tau| - d}. Each triple is first considered at the stage s0s_0 at which the computation Uρ(τ) ⁣ ⁣↓=xU^\rho(\tau)\converge = x is found. At each stage s≥s0s \geq s_0, the triples considered so far are processed in turn. For the triple (ρ,τ,x)(\rho, \tau, x), each string η\eta of length ss such that

  1. Φη[s]⪰ρ\Phi^\eta[s] \succeq \rho,
  2. [η][\eta] is disjoint from the classes [C′]≺[C']^{\prec} of all hungry sets C′C', in their current state, and
  3. λ([C]≺)+2−s≤2−∣τ∣−d\lambda([C]^{\prec}) + 2^{-s} \leq 2^{-|\tau| - d},

is added to CC, the strings being examined in lexicographic order. At the end of the processing of the triple, if λ([C]≺)≥2−∣τ∣−d−1\lambda([C]^{\prec}) \geq 2^{-|\tau| - d - 1} and no request has yet been made for this triple, the request (x,∣τ∣+d+1)(x, |\tau| + d + 1) is enumerated into a set RdR_d.

At stage ss, every string in a hungry set has length at most ss, so for ∣η∣=s|\eta| = s the cylinder [η][\eta] is either contained in or disjoint from each class [C′]≺[C']^{\prec}; condition 2 is decidable. By condition 2, the classes [C]≺[C]^{\prec} of distinct hungry sets are pairwise disjoint, and by condition 3, λ([Cτ,xρ]≺)≤2−∣τ∣−d\lambda([C^\rho_{\tau, x}]^{\prec}) \leq 2^{-|\tau| - d}.

RdR_d is a bounded request set. A request (x,∣τ∣+d+1)(x, |\tau| + d + 1) is made for a triple only when its hungry set has measure at least 2−∣τ∣−d−12^{-|\tau| - d - 1}, the weight of the request. Since the hungry sets are disjoint, the total weight of RdR_d is at most λ(⋃[Cτ,xρ]≺)≤1\lambda\big(\bigcup [C^\rho_{\tau, x}]^{\prec}\big) \leq 1.

A test relative to AA. Let Vd\mathcal V_d be the union of the classes [Cτ,xρ]≺[C^\rho_{\tau, x}]^{\prec} over the triples with ρ≺A\rho \prec A, for the construction with parameter dd. It is Σ10(A)\Sigma^0_1(A), uniformly in dd. For each τ∈dom⁡UA\tau \in \dom U^A there is exactly one triple (ρ,τ,x)(\rho, \tau, x) with ρ≺A\rho \prec A, so

λ(Vd)≤∑τ∈dom⁡UA2−∣τ∣−d≤2−d.\lambda(\mathcal V_d) \leq \sum_{\tau \in \dom U^A} 2^{-|\tau| - d} \leq 2^{-d}.

So (Vd)(\mathcal V_d) is a Martin-Löf test relative to AA, and since Z∈MLRAZ \in \mathrm{MLR}^A, there is dd with Z∉VdZ \notin \mathcal V_d. Fix this dd.

Every hungry set of a true computation becomes half full. Let ρ≺A\rho \prec A with Uρ(τ) ⁣ ⁣↓=xU^\rho(\tau)\converge = x, and suppose that the hungry set C=Cτ,xρC = C^\rho_{\tau, x} never reaches measure 2−∣τ∣−d−12^{-|\tau| - d - 1}. Since ΦZ=A≻ρ\Phi^Z = A \succ \rho, there is a stage ss, after the triple is first considered, with ΦZ ⁣↾s[s]⪰ρ\Phi^{Z \uh s}[s] \succeq \rho and 2−s≤2−∣τ∣−d−12^{-s} \leq 2^{-|\tau| - d - 1}. Let η=Z ⁣↾s\eta = Z \uh s, and consider the processing of the triple at stage ss. Condition 1 holds for η\eta. If [η][\eta] were contained in a hungry set C′=Cτ′,x′ρ′C' = C^{\rho'}_{\tau', x'}, then some η′⪯η\eta' \preceq \eta was added to C′C', so Φη′⪰ρ′\Phi^{\eta'} \succeq \rho' and, since η′≺Z\eta' \prec Z, ρ′≺ΦZ=A\rho' \prec \Phi^Z = A; then Z∈[C′]≺⊆VdZ \in [C']^{\prec} \subseteq \mathcal V_d, a contradiction. So condition 2 holds. Condition 3 holds since λ([C]≺)+2−s<2−∣τ∣−d−1+2−∣τ∣−d−1\lambda([C]^{\prec}) + 2^{-s} < 2^{-|\tau| - d - 1} + 2^{-|\tau| - d - 1}. So η\eta is added to CC, and Z∈[C]≺⊆VdZ \in [C]^{\prec} \subseteq \mathcal V_d, again a contradiction.

Conclusion. For every τ∈dom⁡UA\tau \in \dom U^A with UA(τ)=xU^A(\tau) = x, the triple with ρ≺A\rho \prec A has its hungry set half full, and the request (x,∣τ∣+d+1)(x, |\tau| + d + 1) belongs to RdR_d. By the KC theorem applied to the c.e. bounded request set RdR_d, K(x)≤∣τ∣+d+O(1)K(x) \leq |\tau| + d + O(1). Taking τ\tau a shortest description of xx relative to AA gives K(x)≤+KA(x)K(x) \leq^+ K^A(x).

The hungry sets convert descriptions relative to AA into descriptions without oracle. A description τ\tau of xx with oracle prefix ρ\rho is paid for by measure: the hungry set eats the oracles η\eta that compute ρ\rho through Φ\Phi, until it has eaten 2−∣τ∣−d−12^{-|\tau| - d - 1}. If it starves, ZZ is in a hungry set of a true computation, which is impossible when dd is chosen so that ZZ avoids the dd-th component of a test relative to AA. The parameter dd is not known effectively, but this is harmless: for each dd, RdR_d is a c.e. bounded request set, and only the existence of a good dd is needed.

Corollary 3.4.#

For every set AA, the following are equivalent: AA is low for K; AA is low for Martin-Löf randomness; AA is a base for randomness.

Proof of Corollary 3.4.

By Proposition 3.2 (2) and (3), and Theorem 3.3.

In particular, by Proposition 3.2 (1), all these sets are K-trivial, hence Δ20\Delta^0_2.

3.4 K-triviality and lowness

The converse of Proposition 3.2 (1) is the central theorem on K-triviality.

Theorem 3.5 (Nies; Hirschfeldt, Nies).#

Every K-trivial set is low for K.

The proof, known as the golden run construction, is a priority argument of considerable length, and is admitted here; see Nies (2005), or Nies, Computability and Randomness, Section 5.4, and Downey and Hirschfeldt, Algorithmic Randomness and Complexity, Chapter 11. Another proof, which avoids the golden run, goes through the theorem of Bienvenu, Day, Greenberg, Kučera, Miller, Nies and Turetsky (2014) that every c.e. K-trivial set is computable from a Martin-Löf random sequence that does not compute ∅′\emptyset'; it also uses Theorem 3.3 and the c.e. trees of K-trivial sets. Granting the theorem, the following consequences are immediate.

Corollary 3.6.#

For a set AA, the following are equivalent:

  1. AA is K-trivial;
  2. AA is low for K;
  3. AA is low for Martin-Löf randomness;
  4. AA is a base for randomness.

The K-trivial sets form an ideal of the Turing degrees: they are closed downward under ≤T\leq_T, and under ⊕\oplus.

Proof of Corollary 3.6.

(1 ⇔ 2) by Theorem 3.5 and Proposition 3.2 (1); the other equivalences by Corollary 3.4. Closure downward holds for low for K sets by Proposition 3.2 (4), and closure under join for K-trivial sets by Proposition 1.5.

Downward closure is not apparent from the definition of K-triviality, and closure under join is not apparent from the definition of lowness for K; the equivalence yields both.

3.5 Exercises

Exercise 3.7 ★★.

Show that every set that is low for K is low, that is, A′≤T∅′A' \leq_T \emptyset'. Deduce, using Theorem 3.5, that no K-trivial set is Turing complete.

Hint

For e∈A′e \in A', the use of the computation ΦeA(e)\Phi^A_e(e) gives a prefix of AA with a short description relative to AA.

Solution

Let AA be low for K; it is K-trivial, hence Δ20\Delta^0_2. Let gg be the partial AA-computable function with g(e)=A ⁣↾ug(e) = A \uh u when ΦeA(e) ⁣ ⁣↓\Phi^A_e(e)\converge with use uu. Then KA(g(e))≤+KA(e)≤+K(e)K^A(g(e)) \leq^+ K^A(e) \leq^+ K(e) for e∈A′e \in A', since gg is computable in AA. By Proposition 3.4 of “Prefix-free complexity” (2), K(e)≤+2⌊log⁡(e+1)⌋K(e) \leq^+ 2 \lfloor \log(e+1) \rfloor. Since AA is low for K, there is cc with K(g(e))≤h(e)K(g(e)) \leq h(e) for every e∈A′e \in A', where h(e)=2⌊log⁡(e+1)⌋+ch(e) = 2 \lfloor \log(e+1) \rfloor + c is computable.

With oracle ∅′\emptyset', decide e∈A′e \in A' as follows: for each string σ\sigma of length at most h(e)h(e), decide whether U(σ)U(\sigma) converges; if it does, with output ρ\rho, decide whether ρ≺A\rho \prec A, which is possible since A≤T∅′A \leq_T \emptyset', and whether Φeρ(e) ⁣ ⁣↓\Phi^\rho_e(e)\converge, a Σ10\Sigma^0_1 question. Answer yes if some σ\sigma passes all three tests. If e∈A′e \in A', then a shortest description of g(e)g(e) passes them. Conversely, if ρ≺A\rho \prec A and Φeρ(e) ⁣ ⁣↓\Phi^\rho_e(e)\converge, then ΦeA(e) ⁣ ⁣↓\Phi^A_e(e)\converge by the use principle. So A′≤T∅′A' \leq_T \emptyset'.

A complete set AA has A′≡T∅′′A' \equiv_T \emptyset'', which is not computable in ∅′\emptyset'; so a K-trivial set, which is low for K by Theorem 3.5, is not complete.

Exercise 3.8 ★★.

Show without Theorem 3.5 that a base for randomness AA is not Turing complete.

Solution

By Theorem 3.3, AA is low for K, and by the first exercise it is low, hence not complete.

3.6 Notes and references

Lowness for Martin-Löf randomness was studied by Zambella (1990) and by Kučera and Terwijn (1999); the latter built a non-computable c.e. set low for randomness. Lowness for K is due to Muchnik, who built a non-computable c.e. set low for K. Nies (2005) proved that low for K and low for randomness coincide, and, with Hirschfeldt, that both coincide with K-triviality. The notion of base for randomness and the hungry sets argument are due respectively to Kučera (1993) and to Hirschfeldt, Nies and Stephan (2007).

  • A. Nies, Lowness properties and randomness, Advances in Mathematics 197 (2005), 274–305.
  • D. R. Hirschfeldt, A. Nies and F. Stephan, Using random sets as oracles, Journal of the London Mathematical Society 75 (2007), 610–622.
  • A. Kučera, On relative randomness, Annals of Pure and Applied Logic 63 (1993), 61–67.
  • A. Kučera and S. A. Terwijn, Lowness for the class of random sets, Journal of Symbolic Logic 64 (1999), 1396–1402.
  • L. Bienvenu, A. R. Day, N. Greenberg, A. Kučera, J. S. Miller, A. Nies and D. Turetsky, Computing K-trivial sets by incomplete random sets, Bulletin of Symbolic Logic 20 (2014), 80–90.