[论文解读] Theoretical analysis of optimization problems - Some properties of random k-SAT and k-XORSAT

Fabrizio Altarelli|ArXiv.org|Jan 18, 2008

Constraint Satisfaction and Optimization参考文献 5被引用 4

一句话总结

本文使用统计力学方法对随机k-SAT和k-XORSAT进行理论分析，重点研究相变及解的性质。通过一阶/二阶矩界和消叶法推导出k-XORSAT的相图，并利用副本和腔方法表征高子句-变量比下k-SAT的解结构，表明在解中大多数变量被约束为唯一取值。

ABSTRACT

This thesis is divided in two parts. The first presents an overview of known results in statistical mechanics of disordered systems and its approach to random combinatorial optimization problems. The second part is a discussion of two original results. The first result concerns DPLL heuristics for random k-XORSAT, which is equivalent to the diluted Ising p-spin model. It is well known that DPLL is unable to find the ground states in the clustered phase of the problem, i.e. that it leads to contradictions with probability 1. However, no solid argument supports this is general. A class of heuristics, which includes the well known UC and GUC, is introduced and studied. It is shown that any heuristic in this class must fail if the clause to variable ratio is larger than some constant, which depends on the heuristic but is always smaller than the clustering threshold. The second result concerns the properties of random k-SAT at large clause to variable ratios. In this regime, it is well known that the uniform distribution of random instances is dominated by unsatisfiable instances. A general technique (based on the Replica method) to restrict the distribution to satisfiable instances with uniform weight is introduced, and is used to characterize their solutions. It is found that in the limit of large clause to variable ratios, the uniform distribution of satisfiable random k-SAT formulas is asymptotically equal to the much studied Planted distribution. Both results are already published and available as arXiv:0709.0367 and arXiv:cs/0609101 . A more detailed and self-contained derivation is presented here.

研究动机与目标

• 使用统计力学工具理解随机k-SAT和k-XORSAT中的相变。
• 表征高子句-变量比（α）下随机k-SAT中解的结构。
• 通过泊松启发式和消叶过程分析DPLL型算法的性能与局限性。
• 利用副本和腔方法比较均匀随机可满足公式的解空间与植株分布公式的解空间。
• 确定副本对称解的稳定性，并识别可满足性和聚集性的临界阈值。

提出的方法

• 使用副本方法和腔方法计算随机k-SAT中的自由能和场分布。
• 应用一阶和二阶矩不等式，推导出k-XORSAT可满足性阈值αs的严格界。
• 采用消叶过程分析k-XORSAT公式的核，推导相图。
• 引入DPLL算法的泊松启发式，以建模变量消除并预测算法成功性。
• 使用势函数V(b)表征混合k-XORSAT公式的相态，并分析该启发式的轨迹。
• 引入化学势以条件化可满足公式，并通过鞍点方程计算场的分布。

实验结果

研究问题

• RQ1 随机k-XORSAT中的相变是什么？它们与可满足性阈值有何关系？
• RQ2 在高α（子句-变量比）下，随机k-SAT的解空间与植株分布的解空间有何不同？
• RQ3 场分布和场强度在决定k-SAT中解的结构方面起什么作用？
• RQ4 泊松启发式能否准确预测DPLL算法在k-XORSAT上的成功性？其局限性是什么？
• RQ5 在高α下，k-SAT的副本对称解是否稳定？这对算法可解性有何含义？

主要发现

• 对于k-XORSAT，使用一阶和二阶矩方法严格界定了可满足性阈值αs，且消叶过程识别出公式的核。
• 在高α下，k-SAT中约1−e−O(α)的变量在所有解中被约束为相同取值，表明解空间具有高度结构化。
• k-SAT中作用于变量的场的尺度为O(α)，其分布非平凡，且有相当一部分变量具有较大的场强度。
• 在高α下，k-SAT的副本对称解不稳定，稳定性矩阵中出现负特征值，表明解空间中存在大量簇。
• DPLL的泊松启发式在超过临界αh后失败，而αh严格小于可满足性阈值αs，表明存在算法困难。
• 在k-XORSAT中，由于场符号无偏差（与k-SAT不同），标准消息传递算法（如警告传播）无法收敛，表明在高α下需要新算法。