We present a fast and approximate multifrontal solver for large-scale sparse linear systems arising from finite-difference, finite-volume or finite-element discretization of high-frequency wave equations. The proposed solver leverages the butterfly algorithm and its hierarchical matrix extension for compressing and factorizing large frontal matrices via graph-distance guided entry evaluation or randomized matrix-vector multiplication-based schemes. Complexity analysis and numerical experiments demonstrate $\mathcal{O}(N\log^2 N)$ computation and $\mathcal{O}(N)$ memory complexity when applied to an $N\times N$ sparse system arising from 3D high-frequency Helmholtz and Maxwell problems.
翻译:我们为高频波方程式的有限差异、有限体积或有限分解产生的大规模稀薄线性系统提出了一个快速和近似多前方求解器。提议的求解器利用蝴蝶算法及其等级矩阵扩展,通过图形远程引导输入评价或随机化矩阵-矢量倍增计划压缩和计算大型前方矩阵。复杂度分析和数字实验显示,在对3D高频Helmholtz和Maxwell问题产生的稀薄系统应用时,计算值为$\mathcal{O}(Nlog2N)$和$\mathcal{O}(N)$的内存复杂性。