We present a systematic study of the nested sampling algorithm based on theexample of the Potts model. This model, which exhibits a first order phasetransition for $q>4$, exemplifies a generic numerical challenge in statisticalphysics: The evaluation of the partition function and thermodynamicobservables, which involve high dimensional sums of sharply structuredmulti-modal density functions. It poses a major challenge to most standardnumerical techniques, such as Markov Chain Monte Carlo. In this paper we willdemonstrate that nested sampling is particularly suited for such problems andit has a couple of advantages. For calculating the partition function of thePotts model with $N$ sites: a) one run stops after $O(N)$ moves, so it takes$O(N^{2})$ operations for the run, b) only a single run is required to computethe partition function along with the assignment of confidence intervals, c)the confidence intervals of the logarithmic partition function decrease with$1/\sqrt{N}$ and d) a single run allows to compute quantities for alltemperatures while the autocorrelation time is very small, irrespective oftemperature. Thermodynamic expectation values of observables, which arecompletely determined by the bond configuration in the representation ofFortuin and Kasteleyn, like the Helmholtz free energy, the internal energy aswell as the entropy and heat capacity, can be calculated in the same single runneeded for the partition function along with their confidence intervals. Incontrast, thermodynamic expectation values of magnetic properties like themagnetization and the magnetic susceptibility require sampling the additionalspin degree of freedom. Results and performance are studied in detail andcompared with those obtained with multi-canonical sampling. Eventually theimplications of the findings on a parallel implementation of nested samplingare outlined.
展开▼
机译:我们基于Potts模型的示例对嵌套采样算法进行了系统的研究。该模型表现出$ q> 4 $的一阶相变,它例证了统计物理学中的一个通用数值挑战:分配函数和热力学可观测值的评估,其中涉及结构清晰的多峰密度函数的高维和。它对大多数标准数值技术(例如Markov Chain Monte Carlo)提出了重大挑战。在本文中,我们将证明嵌套采样特别适用于此类问题,并且具有许多优点。要计算具有$ N $个站点的Potts模型的分区函数:a)在$ O(N)$移动之后,一次运行停止,因此运行需要$ O(N ^ {2})$个操作,b)仅一个需要单次运行来计算分区函数以及置信区间的分配,c)对数分区函数的置信区间随着$ 1 / \ sqrt {N} $减小,并且d)单次运行允许计算所有温度的数量,而与温度无关,自相关时间非常短。观测值的热力学期望值完全由Fortuin和Kasteleyn表示中的键构型确定,例如亥姆霍兹自由能,内能以及熵和热容量,可以在分配函数的同一单一运行中沿与他们的置信区间。相反,诸如磁化和磁化率之类的磁性的热力学期望值需要对额外的自旋自由度进行采样。对结果和性能进行了详细研究,并与通过多规范采样获得的结果和性能进行了比较。最终概述了发现结果对并行采样嵌套实现的意义。
展开▼