Search Results

Carmosino, Impagliazzo, Kabanets, and Kolokolova (CCC, 2016) showed that the existence of natural properties in the sense of Razborov and Rudich (JCSS, 1997) implies PAC learning algorithms in the sense of Valiant (Comm. ACM, 1984), for boolean functions in P/poly, under the uniform distribution and with membership queries. It is still an open problem to get from natural properties learning algorithms that do not rely on membership queries but rather use randomly drawn labeled examples. Natural properties may be understood as an average-case version of MCSP, the problem of deciding the minimum size of a circuit computing a given truth-table. Problems related to MCSP include those concerning time-bounded Kolmogorov complexity. MKTP, for example, asks for the KT-complexity of a given string. KT-complexity is a relaxation of circuit size, as it does away with the requirement that a short description of a string be interpreted as a boolean circuit. In this work, under assumptions of MKTP and the related problem MK^tP being easy on average, we get learning algorithms for boolean functions in P/poly that - work over any distribution D samplable by a family of polynomial-size circuits (given explicitly in the case of MKTP), - only use randomly drawn labeled examples from D, and - are agnostic (do not require the target function to belong to the hypothesis class). Our results build upon the recent work of Hirahara and Nanashima (FOCS, 2021) who showed similar learning consequences but under a stronger assumption that NP is easy on average.

Understanding the relationship between the worst-case and average-case complexities of NP and of other subclasses of PH is a long-standing problem in complexity theory. Over the last few years, much progress has been achieved in this front through the investigation of meta-complexity: the complexity of problems that refer to the complexity of the input string x (e.g., given a string x, estimate its time-bounded Kolmogorov complexity). In particular, [Shuichi Hirahara, 2021] employed techniques from meta-complexity to show that if DistNP ⊆ AvgP then UP ⊆ DTIME[2^{O(n/log n)}]. While this and related results [Shuichi Hirahara and Mikito Nanashima, 2021; Lijie Chen et al., 2022] offer exciting progress after a long gap, they do not survive in the setting of randomized computations: roughly speaking, "randomness" is the opposite of "structure", and upper bounding the amount of structure (time-bounded Kolmogorov complexity) of different objects is crucial in recent applications of meta-complexity. This limitation is significant, since randomized computations are ubiquitous in algorithm design and give rise to a more robust theory of average-case complexity [Russell Impagliazzo and Leonid A. Levin, 1990]. In this work, we develop a probabilistic theory of meta-complexity, by incorporating randomness into the notion of complexity of a string x. This is achieved through a new probabilistic variant of time-bounded Kolmogorov complexity that we call pK^t complexity. Informally, pK^t(x) measures the complexity of x when shared randomness is available to all parties involved in a computation. By porting key results from meta-complexity to the probabilistic domain of pK^t complexity and its variants, we are able to establish new connections between worst-case and average-case complexity in the important setting of probabilistic computations: - If DistNP ⊆ AvgBPP, then UP ⊆ RTIME[2^O(n/log n)]. - If DistΣ^P_2 ⊆ AvgBPP, then AM ⊆ BPTIME[2^O(n/log n)]. - In the fine-grained setting [Lijie Chen et al., 2022], we get UTIME[2^O(√{nlog n})] ⊆ RTIME[2^O(√{nlog n})] and AMTIME[2^O(√{nlog n})] ⊆ BPTIME[2^O(√{nlog n})] from stronger average-case assumptions. - If DistPH ⊆ AvgBPP, then PH ⊆ BPTIME[2^O(n/log n)]. Specifically, for any 𝓁 ≥ 0, if DistΣ_{𝓁+2}^P ⊆ AvgBPP then Σ_𝓁^{P} ⊆ BPTIME[2^O(n/log n)]. - Strengthening a result from [Shuichi Hirahara and Mikito Nanashima, 2021], we show that if DistNP ⊆ AvgBPP then polynomial size Boolean circuits can be agnostically PAC learned under any unknown 𝖯/poly-samplable distribution in polynomial time. In some cases, our framework allows us to significantly simplify existing proofs, or to extend results to the more challenging probabilistic setting with little to no extra effort.