Bridging Computational Notions of Depth: Variants of Strong Depth | HackerNoon

Table of Links

Abstract and 1 Introduction

2 Background

3 On the slow growth law

4 Members of Deep Π0 1 classes

5 Strong depth is Negligible

6 Variants of Strong Depth

References

Appendix A. Proof of Lemma 3

6. Variants of Strong Depth

Proof. Suppose that X is not weakly deep. Then there is some computable measure µ such that X is µ-Martin-Lof random. By the Levin-Schnorr theorem for a priori complexity with respect to the measure µ (implicit in [Lev73]), there is some c such that

KA(X↾n) ≥ − log µ(X↾n) − c

for all n ∈ ω. Equivalently, we have

Since every computable measure is a computable semimeasure, the conclusion follows.

For the other direction, suppose that there is some computable, continuous semimeasure and some c ∈ ω such that for all n ∈ ω,

It thus follows from the Levin-Schnorr theorem for a priori complexity that X is µ-Martin-Lof random and hence is not weakly deep.

We define a sequence to be KA-deep if

(see the proof of [BDM23, Lemma 2.6] for discrete semimeasures which directly translates to the case of continuous semimeasures). Thus we can conclude:

6.2. Depth and monotone complexity. We can obtain a similar characterization of weak depth in terms of monotone complexity. Define a sequence to be Km-deep if

This notion was studied by Schnorr and Fuchs [SF77], who used the term super-learnable to refer to the failure of being Km-deep. In particular, Schnorr and Fuchs proved that a sequence is super-learnable if and only if it is Martin-Lof random with respect to a computable measure. Given that a sequence is weakly deep if and only if it is not Martin-Lof random with respect to a computable measure, we have the following.