Core Concepts
We describe a quantum algorithm based on an interior point method that can solve a linear program with n inequality constraints on d variables in time √n · poly(d, log(1/ε)), which is sublinear for tall linear programs (n ≫ d). The algorithm explicitly returns a feasible solution that is ε-close to optimal.
Abstract
The authors describe a quantum algorithm for solving linear programs (LPs) using an interior point method (IPM). The key contributions are:
A quantum algorithm for spectral approximation of tall matrices ATA, which allows for efficient approximation of the Hessian in the IPM. This algorithm uses leverage score sampling and Grover search.
A quantum algorithm for approximating the ℓp-Lewis weights of a matrix, which are used to define an improved self-concordant barrier function for the IPM.
A quantum algorithm for approximate matrix-vector multiplication, which allows for efficient approximation of the gradient in the IPM.
By combining these subroutines, the authors obtain a quantum IPM that can solve LPs with n constraints and d variables in time √n · poly(d, log(1/ε)), which is sublinear in the input size when n ≫ d. This improves upon previous quantum algorithms for LP solving, which typically had a poly(1/ε) dependence on the error parameter ε.
The authors also discuss lower bounds and compare their results to a quantum speedup using cutting plane methods. They identify several open directions for further improving the complexity of their quantum IPM.
Stats
The algorithm runs in time √n · poly(d, log(1/ε)).
The algorithm makes e
O(√nd/ε) row queries to the constraint matrix A.
The algorithm explicitly returns a feasible solution that is ε-close to optimal.