Higher order alternate directions methods (Q1892464)

From MaRDI portal
scientific article
Language Label Description Also known as
English
Higher order alternate directions methods
scientific article

    Statements

    Higher order alternate directions methods (English)
    0 references
    0 references
    13 November 1995
    0 references
    Let \(f\) be a continuous function from a neighborhood of real \(t = 0\) to the vector space \(\mathcal M\) of square matrices, \(f(t,a;b) = \text{exp} (a_ 1 tA) \cdot \text{exp} (b_ 1 tB) \cdot \dots \cdot \text{exp} (a_ k tA) \cdot \text{exp} (b_ k tB)\), \(A \in {\mathcal M}\), \(B \in {\mathcal M}\), \(a_ j\), \(b_ j\) real parameters. Alternate direction methods are based on the construction of a function \(f\) with \(f(t/n)^ n - \text{exp} (t (A+B)) = O(n^{-p})\). \(k = 1\), \(a_ 1 = b_ 1 = 1\) is the well-known Lie formula with the order \(p = 1\). \textit{G. Strang} [Arch. Rat. Mech. Anal. 12, 392-402 (1963; 113, 323), SIAM J. Numer. Anal. 5, 506-517 (1968; Zbl 0184.385)] has given a formula with order \(p = 2 : k = 2\), \(a_ 1 = a_ 2 = 1/2\), \(b_ 1 = 1\), \(b_ 2 = 0\). But \textit{Q. Sheng} [IMA J. Numer. Anal. 9, No. 2, 199- 212 (1989; Zbl 0676.65116)] has pointed out that (in the non-commutative case) no \(k\) exists with \(a_ j \geq 0\), \(b_ j \geq 0\) and order \(p \geq 3\). Hence, the search of more accurate product formulas is negative. However, there exist formulas of third order with (partially) negative coefficients, e.g. \(k = 3\) with \(a_ 1 = 1\), \(a_ 2 = -a_ 3 = -2/3\), \(b_ 1 = -1/24\), \(b_ 2 = 3/4\), \(b_ 3 = 7/24\). The author studies linear combinations of product formulas which yield order \(p = 4\), e.g. the result of \textit{S. Z. Burstein, A. A. Mirin} [J. Comput. Phys. 5, 547- 571 (1970; Zbl 0223.65053)] with the use of \(2/3 f(t, 1/2, 1/2; 1, 0) + 2/3 f(t,1, 0; 1/2, 1/2) - 1/6 f(t, 1, 0; 1, 0) - 1/6 f(t,0, 1; 0, 1)\) (this means \(k = 2\)), or a new result with \(k = 3\) obtained by Richardson's extrapolation: \(4/3 f(t, 1/4, 1/2, 1/4; 1/2, 1/2, 0) - 1/3 f(t, 1/2, 1/2, 0; 1, 0, 0).\) There are only few numerical experiments. But the method (based on a Campbell-Baker-Dynkin-Hausdorff formal series) is a promising one, of course, it requires care and, probably, the use of computer algebra.
    0 references
    iterative methods
    0 references
    finite difference methods
    0 references
    alternate direction methods
    0 references
    Lie formula
    0 references
    113, 323
    0 references
    product formulas
    0 references
    Richardson's extrapolation
    0 references
    numerical experiments
    0 references
    Campbell-Baker-Dynkin-Hausdorff formal series
    0 references
    computer algebra
    0 references
    0 references

    Identifiers