In this work we revisit the arithmetic and bit complexity of Hermitian eigenproblems. We first provide an analysis for the divide-and-conquer tridiagonal eigensolver of Gu and Eisenstat [GE95] in the Real RAM model, when accelerated with the Fast Multipole Method. The analysis asserts the claimed nearly-$O(n^2)$ complexity to compute a full diagonalization of a symmetric tridiagonal matrix. Combined with the tridiagonal reduction algorithm of Sch\"onhage [Sch72], it implies that a Hermitian matrix can be diagonalized deterministically in $O(n^{\omega}\log(n)+n^2\mathrm{polylog}(n/\epsilon))$ arithmetic operations, where $\omega\lesssim 2.371$ is the square matrix multiplication exponent. This improves the classic deterministic $O(n^3)$ diagonalization algorithms, and derandomizes the $ O(n^{\omega}\log^2(n/\epsilon))$ algorithm of [BGVKS, FOCS '20]. Ultimately, this has a direct application to the SVD, which is widely used as a subroutine in advanced algorithms, but its complexity and approximation guarantees are often unspecified. In finite precision, we show that Sch\"onhage's algorithm is stable in floating point using $O(\log(n/\epsilon))$ bits. Combined with the (rational arithmetic) algorithm of Bini and Pan [BP91], it provides a deterministic algorithm to compute all the eigenvalues of a Hermitian matrix in $O\left(n^{\omega}F\left(\log(n/\epsilon)\right)+n^2\mathrm{polylog}(n/\epsilon)\right)$ bit operations, where $F(b)\in\widetilde{O}(b)$ is the bit complexity of a single floating point operation on $b$ bits. This improves the best known $\widetilde{O}(n^3)$ deterministic and $O\left( n^{\omega}\log^2(n/\epsilon)F\left(\log^4(n/\epsilon)\log(n)\right)\right)$ randomized complexities. We conclude with some other useful subroutines such as computing spectral gaps, condition numbers, and spectral projectors, and few open problems.
翻译:暂无翻译