Fast multipole method

Fast multipole method
Name	Fast multipole method
Introduced	1987
Inventor	Leslie Greengard; Vladimir Rokhlin
Field	Numerical analysis; Computational physics
Related	Boundary element method; Particle mesh Ewald; Barnes–Hut algorithm

Fast multipole method

The fast multipole method is a numerical algorithm for accelerating long-range interaction computations in many-body problems, developed to reduce the cost of evaluating pairwise potentials in classical Laplace-like kernels and other Green's functions. It transforms direct N^2 interaction calculations into hierarchical approximations that permit near-linear scaling, making large-scale simulations feasible for problems encountered by researchers at institutions such as Massachusetts Institute of Technology, Princeton University, Stanford University and laboratories like Lawrence Berkeley National Laboratory and Los Alamos National Laboratory. The method has influenced computational practice in disciplines represented by organizations including American Mathematical Society, Society for Industrial and Applied Mathematics, Royal Society and has connections to algorithms used in projects by companies such as IBM, Intel Corporation and NVIDIA.

Overview

The method exploits multipole expansions and local expansions derived from classical potential theory pioneered by figures like Isaac Newton and Pierre-Simon Laplace and formalized in modern analysis by researchers at universities such as Yale University and University of Cambridge. It partitions space into a hierarchical tree (octree or quadtree), grouping sources and targets within cells associated with academic centers like California Institute of Technology and University of Chicago. At each level of the tree, interactions are replaced by translations of multipole moments and converted into local fields, a strategy that complements earlier N-body methods such as the Barnes–Hut algorithm. Influential awardees including recipients of the Turing Award have highlighted the algorithm's impact on computational science.

Mathematical foundations

The FMM rests on expansions of kernels—often the fundamental solutions of operators named after mathematicians like Pierre-Simon Laplace and George Green—into series of spherical harmonics and equivalent representations studied by scholars from École Normale Supérieure to École Polytechnique. It leverages addition theorems like those associated with Friedrichs-style analysis and translation operators closely related to work by researchers at Institute for Advanced Study. Convergence, error bounds and stability results draw on classical estimates from analysts linked to institutions such as Princeton University and Harvard University, and use orthogonal function theory connected to figures like Carl Gustav Jacobi and Sofia Kovalevskaya. The method also employs matrix compression concepts akin to low-rank approximations studied at Courant Institute and in the literature of numerical linear algebra connected to names such as Gene H. Golub and Lloyd N. Trefethen.

Algorithm and implementation

Implementation builds a hierarchical spatial decomposition—quadtree in two dimensions and octree in three—using data structures utilized in computational geometry work at places like University of Illinois Urbana-Champaign and ETH Zurich. Key steps include upward pass (multipole aggregation), translation operators (multipole-to-local), and downward pass (local evaluation), operations implemented in high-performance libraries developed by teams at Lawrence Livermore National Laboratory and Argonne National Laboratory. Efficient implementations exploit vectorization and parallelism on architectures provided by Intel Corporation and NVIDIA, and parallel programming models from OpenMP and MPI often used in codebases affiliated with Sandia National Laboratories. Practical software packages incorporate fast transforms influenced by research at Courant Institute and integrate with meshing tools from Gmsh-related communities.

Variants and extensions

Numerous variants adapt the core idea: kernel-independent formulations developed in collaborations across University of California, Berkeley and University of Texas at Austin remove analytic kernel dependence; plane-wave and black-box FMMs align with spectral techniques advanced at California Institute of Technology and Imperial College London; and hybrid methods combine FMM with fast Fourier transform strategies popularized by groups at Princeton University and Stanford University. Extensions incorporate adaptive refinement used in weather and climate centers like National Center for Atmospheric Research and couple with boundary element methods pursued at Max Planck Institute laboratories. High-order and low-frequency adaptations connect to electromagnetics research at MIT Lincoln Laboratory and antenna design groups at Fermilab.

Complexity and performance

The canonical complexity of the FMM is O(N) or O(N log N) depending on dimensionality and implementation choices, claims analyzed in theoretical work from departments at University of California, Los Angeles and Columbia University. Performance depends on expansion order, translation cost, and tree balancing, with benchmarks run on supercomputers such as Summit (supercomputer) and programs funded by National Science Foundation and Department of Energy. Robust scalability has been demonstrated in multicore and GPU-accelerated implementations reported by research teams at Oak Ridge National Laboratory and industrial deployments within Google-scale data centers for certain kernel computations.

Applications

Applications span computational physics, chemistry, and engineering: N-body gravitational simulations in astrophysics groups at European Southern Observatory and Space Telescope Science Institute; electrostatics and molecular dynamics in research at National Institutes of Health-funded centers and pharmaceutical collaborations; acoustic and electromagnetic scattering problems tackled at Naval Research Laboratory and Air Force Research Laboratory; and inverse problems handled by geophysics departments at University of Oxford and ETH Zurich. The method underpins fast solvers used in computational design at Boeing and Airbus and in medical imaging workflows in institutions like Mayo Clinic.

History and development

Invented in 1987 by scholars at New York University and Yale University, the method was introduced during a period of rapid growth in computational capacity that involved collaborations across labs such as Los Alamos National Laboratory and Lawrence Livermore National Laboratory. Subsequent developments were propelled by research groups at Courant Institute, Princeton University and Stanford University, and by funding and recognition from agencies like National Science Foundation and awards from societies such as SIAM and the Royal Society. The FMM has since evolved into a rich ecosystem of theory, software, and applications spanning many of the organizations and institutions above.

Category:Numerical analysis