[Paper Review] Beyond Natural Proofs: Hardness Magnification and Locality

Lijie Chen, Shuichi Hirahara+4 more|arXiv (Cornell University)|Nov 19, 2019

Complexity and Algorithms in Graphs4 citations

TL;DR

This paper investigates hardness magnification as a strategy to overcome the natural proofs barrier in circuit complexity, showing that slightly superlinear lower bounds for minimum circuit size problems (MCSP) imply the non-existence of natural proofs. It identifies a 'locality barrier' explaining why known lower bound techniques often fail when adapted to magnification, and demonstrates that certain magnification theorems can rule out efficient learning algorithms by implying strong circuit lower bounds.

ABSTRACT

Hardness magnification reduces major complexity separations (such as $\mathsf{\mathsf{EXP}} subseteq \mathsf{NC}^1$) to proving lower bounds for some natural problem $Q$ against weak circuit models. Several recent works [OS18, MMW19, CT19, OPS19, CMMW19, Oli19, CJW19a] have established results of this form. In the most intriguing cases, the required lower bound is known for problems that appear to be significantly easier than $Q$, while $Q$ itself is susceptible to lower bounds but these are not yet sufficient for magnification. In this work, we provide more examples of this phenomenon, and investigate the prospects of proving new lower bounds using this approach. In particular, we consider the following essential questions associated with the hardness magnification program: Does hardness magnification avoid the natural proofs barrier of Razborov and Rudich [RR97]? Can we adapt known lower bound techniques to establish the desired lower bound for $Q$?

Motivation & Objective

• To investigate whether hardness magnification can circumvent the natural proofs barrier of Razborov and Rudich.
• To assess the feasibility of adapting existing lower bound techniques to achieve strong complexity separations via magnification.
• To identify and formalize a 'locality barrier' that limits the applicability of known techniques in magnification-based proofs.
• To establish connections between hardness magnification, the non-existence of natural proofs, and the impossibility of efficient learning algorithms.
• To provide a systematic analysis of magnification frontiers across different circuit classes and problems like MCSP and MKtP.

Proposed method

• Reduces major complexity separations (e.g., EXP ⊈ NC¹) to proving lower bounds for natural problems Q against weak circuit models.
• Introduces a potential function Φ to track circuit size and oracle activity during random restrictions, enabling analysis of formula size reduction.
• Applies random restrictions (ρ) to oracle circuits, showing that the potential function drops by a factor of k/cF per round, leading to a tiny formula with high probability.
• Uses Chebyshev’s inequality to bound the probability of retaining a large number of active variables after t rounds of restrictions.
• Constructs a hard input x* by fixing unrestricted variables to 0, showing that the resulting formula cannot compute MCSP on inputs with high circuit complexity.
• Establishes that if MCSP has slightly superlinear-size circuits, then natural proofs do not exist, and efficient learning algorithms are ruled out.

Experimental results

Research questions

• RQ1 Can hardness magnification theorems avoid the natural proofs barrier?
• RQ2 To what extent can known lower bound techniques (e.g., polynomial approximation, random restrictions) be adapted to prove strong lower bounds via magnification?
• RQ3 What is the role of oracles and locality in undermining the effectiveness of standard lower bound methods in magnification frameworks?
• RQ4 Do magnification theorems that imply strong circuit lower bounds also rule out the existence of efficient learning algorithms?
• RQ5 What structural limitations (e.g., locality barriers) prevent direct application of existing techniques to magnification-based lower bounds?

Key findings

• Slightly superlinear-size circuit lower bounds for MCSP imply the non-existence of natural proofs, meaning magnification can bypass the natural proofs barrier.
• Certain magnification theorems not only imply strong worst-case circuit lower bounds but also rule out the existence of efficient learning algorithms for the problems involved.
• A 'locality barrier' is identified: existing magnification theorems show that problems Q admit highly efficient circuits with small fan-in oracle gates, while standard lower bound techniques often extend to such models.
• The Razborov-Smolensky polynomial approximation method and other known techniques fail to yield strong lower bounds via magnification due to their invariance under oracle addition.
• For oracle formulas with size N²⁻ε and adaptivity τ = o(log N / log log N), the potential function Φ drops by a factor of k/cF per restriction, leading to a formula depending on O(log³N) variables with high probability.
• After t = O(log N / log log N) rounds of random restrictions, the remaining formula depends on O(log³N) variables and has at least N^{ε/2−o(1)} unrestricted variables, enabling a hard input construction that violates MCSP computability.